Index: head/sys/kern/subr_witness.c =================================================================== --- head/sys/kern/subr_witness.c (revision 335355) +++ head/sys/kern/subr_witness.c (revision 335356) @@ -1,3084 +1,3084 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 2008 Isilon Systems, Inc. * Copyright (c) 2008 Ilya Maykov * Copyright (c) 1998 Berkeley Software Design, Inc. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Berkeley Software Design Inc's name may not be used to endorse or * promote products derived from this software without specific prior * written permission. * * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $ * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $ */ /* * Implementation of the `witness' lock verifier. Originally implemented for * mutexes in BSD/OS. Extended to handle generic lock objects and lock * classes in FreeBSD. */ /* * Main Entry: witness * Pronunciation: 'wit-n&s * Function: noun * Etymology: Middle English witnesse, from Old English witnes knowledge, * testimony, witness, from 2wit * Date: before 12th century * 1 : attestation of a fact or event : TESTIMONY * 2 : one that gives evidence; specifically : one who testifies in * a cause or before a judicial tribunal * 3 : one asked to be present at a transaction so as to be able to * testify to its having taken place * 4 : one who has personal knowledge of something * 5 a : something serving as evidence or proof : SIGN * b : public affirmation by word or example of usually * religious faith or conviction * 6 capitalized : a member of the Jehovah's Witnesses */ /* * Special rules concerning Giant and lock orders: * * 1) Giant must be acquired before any other mutexes. Stated another way, * no other mutex may be held when Giant is acquired. * * 2) Giant must be released when blocking on a sleepable lock. * * This rule is less obvious, but is a result of Giant providing the same * semantics as spl(). Basically, when a thread sleeps, it must release * Giant. When a thread blocks on a sleepable lock, it sleeps. Hence rule * 2). * * 3) Giant may be acquired before or after sleepable locks. * * This rule is also not quite as obvious. Giant may be acquired after * a sleepable lock because it is a non-sleepable lock and non-sleepable * locks may always be acquired while holding a sleepable lock. The second * case, Giant before a sleepable lock, follows from rule 2) above. Suppose * you have two threads T1 and T2 and a sleepable lock X. Suppose that T1 * acquires X and blocks on Giant. Then suppose that T2 acquires Giant and * blocks on X. When T2 blocks on X, T2 will release Giant allowing T1 to * execute. Thus, acquiring Giant both before and after a sleepable lock * will not result in a lock order reversal. */ #include __FBSDID("$FreeBSD$"); #include "opt_ddb.h" #include "opt_hwpmc_hooks.h" #include "opt_stack.h" #include "opt_witness.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef DDB #include #endif #include #if !defined(DDB) && !defined(STACK) #error "DDB or STACK options are required for WITNESS" #endif /* Note that these traces do not work with KTR_ALQ. */ #if 0 #define KTR_WITNESS KTR_SUBSYS #else #define KTR_WITNESS 0 #endif #define LI_RECURSEMASK 0x0000ffff /* Recursion depth of lock instance. */ #define LI_EXCLUSIVE 0x00010000 /* Exclusive lock instance. */ #define LI_NORELEASE 0x00020000 /* Lock not allowed to be released. */ /* Define this to check for blessed mutexes */ #undef BLESSING #ifndef WITNESS_COUNT #define WITNESS_COUNT 1536 #endif #define WITNESS_HASH_SIZE 251 /* Prime, gives load factor < 2 */ #define WITNESS_PENDLIST (512 + (MAXCPU * 4)) /* Allocate 256 KB of stack data space */ #define WITNESS_LO_DATA_COUNT 2048 /* Prime, gives load factor of ~2 at full load */ #define WITNESS_LO_HASH_SIZE 1021 /* * XXX: This is somewhat bogus, as we assume here that at most 2048 threads * will hold LOCK_NCHILDREN locks. We handle failure ok, and we should * probably be safe for the most part, but it's still a SWAG. */ #define LOCK_NCHILDREN 5 #define LOCK_CHILDCOUNT 2048 #define MAX_W_NAME 64 #define FULLGRAPH_SBUF_SIZE 512 /* * These flags go in the witness relationship matrix and describe the * relationship between any two struct witness objects. */ #define WITNESS_UNRELATED 0x00 /* No lock order relation. */ #define WITNESS_PARENT 0x01 /* Parent, aka direct ancestor. */ #define WITNESS_ANCESTOR 0x02 /* Direct or indirect ancestor. */ #define WITNESS_CHILD 0x04 /* Child, aka direct descendant. */ #define WITNESS_DESCENDANT 0x08 /* Direct or indirect descendant. */ #define WITNESS_ANCESTOR_MASK (WITNESS_PARENT | WITNESS_ANCESTOR) #define WITNESS_DESCENDANT_MASK (WITNESS_CHILD | WITNESS_DESCENDANT) #define WITNESS_RELATED_MASK \ (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK) #define WITNESS_REVERSAL 0x10 /* A lock order reversal has been * observed. */ #define WITNESS_RESERVED1 0x20 /* Unused flag, reserved. */ #define WITNESS_RESERVED2 0x40 /* Unused flag, reserved. */ #define WITNESS_LOCK_ORDER_KNOWN 0x80 /* This lock order is known. */ /* Descendant to ancestor flags */ #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2) /* Ancestor to descendant flags */ #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2) #define WITNESS_INDEX_ASSERT(i) \ MPASS((i) > 0 && (i) <= w_max_used_index && (i) < witness_count) static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness"); /* * Lock instances. A lock instance is the data associated with a lock while * it is held by witness. For example, a lock instance will hold the * recursion count of a lock. Lock instances are held in lists. Spin locks * are held in a per-cpu list while sleep locks are held in per-thread list. */ struct lock_instance { struct lock_object *li_lock; const char *li_file; int li_line; u_int li_flags; }; /* * A simple list type used to build the list of locks held by a thread * or CPU. We can't simply embed the list in struct lock_object since a * lock may be held by more than one thread if it is a shared lock. Locks * are added to the head of the list, so we fill up each list entry from * "the back" logically. To ease some of the arithmetic, we actually fill * in each list entry the normal way (children[0] then children[1], etc.) but * when we traverse the list we read children[count-1] as the first entry * down to children[0] as the final entry. */ struct lock_list_entry { struct lock_list_entry *ll_next; struct lock_instance ll_children[LOCK_NCHILDREN]; u_int ll_count; }; /* * The main witness structure. One of these per named lock type in the system * (for example, "vnode interlock"). */ struct witness { char w_name[MAX_W_NAME]; uint32_t w_index; /* Index in the relationship matrix */ struct lock_class *w_class; STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */ STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */ struct witness *w_hash_next; /* Linked list in hash buckets. */ const char *w_file; /* File where last acquired */ uint32_t w_line; /* Line where last acquired */ uint32_t w_refcount; uint16_t w_num_ancestors; /* direct/indirect * ancestor count */ uint16_t w_num_descendants; /* direct/indirect * descendant count */ int16_t w_ddb_level; unsigned w_displayed:1; unsigned w_reversed:1; }; STAILQ_HEAD(witness_list, witness); /* * The witness hash table. Keys are witness names (const char *), elements are * witness objects (struct witness *). */ struct witness_hash { struct witness *wh_array[WITNESS_HASH_SIZE]; uint32_t wh_size; uint32_t wh_count; }; /* * Key type for the lock order data hash table. */ struct witness_lock_order_key { uint16_t from; uint16_t to; }; struct witness_lock_order_data { struct stack wlod_stack; struct witness_lock_order_key wlod_key; struct witness_lock_order_data *wlod_next; }; /* * The witness lock order data hash table. Keys are witness index tuples * (struct witness_lock_order_key), elements are lock order data objects * (struct witness_lock_order_data). */ struct witness_lock_order_hash { struct witness_lock_order_data *wloh_array[WITNESS_LO_HASH_SIZE]; u_int wloh_size; u_int wloh_count; }; #ifdef BLESSING struct witness_blessed { const char *b_lock1; const char *b_lock2; }; #endif struct witness_pendhelp { const char *wh_type; struct lock_object *wh_lock; }; struct witness_order_list_entry { const char *w_name; struct lock_class *w_class; }; /* * Returns 0 if one of the locks is a spin lock and the other is not. * Returns 1 otherwise. */ static __inline int witness_lock_type_equal(struct witness *w1, struct witness *w2) { return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) == (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK))); } static __inline int witness_lock_order_key_equal(const struct witness_lock_order_key *a, const struct witness_lock_order_key *b) { return (a->from == b->from && a->to == b->to); } static int _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname); static void adopt(struct witness *parent, struct witness *child); #ifdef BLESSING static int blessed(struct witness *, struct witness *); #endif static void depart(struct witness *w); static struct witness *enroll(const char *description, struct lock_class *lock_class); static struct lock_instance *find_instance(struct lock_list_entry *list, const struct lock_object *lock); static int isitmychild(struct witness *parent, struct witness *child); static int isitmydescendant(struct witness *parent, struct witness *child); static void itismychild(struct witness *parent, struct witness *child); static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS); static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS); static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS); static int sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS); static void witness_add_fullgraph(struct sbuf *sb, struct witness *parent); #ifdef DDB static void witness_ddb_compute_levels(void); static void witness_ddb_display(int(*)(const char *fmt, ...)); static void witness_ddb_display_descendants(int(*)(const char *fmt, ...), struct witness *, int indent); static void witness_ddb_display_list(int(*prnt)(const char *fmt, ...), struct witness_list *list); static void witness_ddb_level_descendants(struct witness *parent, int l); static void witness_ddb_list(struct thread *td); #endif static void witness_debugger(int cond, const char *msg); static void witness_free(struct witness *m); static struct witness *witness_get(void); static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size); static struct witness *witness_hash_get(const char *key); static void witness_hash_put(struct witness *w); static void witness_init_hash_tables(void); static void witness_increment_graph_generation(void); static void witness_lock_list_free(struct lock_list_entry *lle); static struct lock_list_entry *witness_lock_list_get(void); static int witness_lock_order_add(struct witness *parent, struct witness *child); static int witness_lock_order_check(struct witness *parent, struct witness *child); static struct witness_lock_order_data *witness_lock_order_get( struct witness *parent, struct witness *child); static void witness_list_lock(struct lock_instance *instance, int (*prnt)(const char *fmt, ...)); static int witness_output(const char *fmt, ...) __printflike(1, 2); static int witness_voutput(const char *fmt, va_list ap) __printflike(1, 0); static void witness_setflag(struct lock_object *lock, int flag, int set); static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL, "Witness Locking"); /* * If set to 0, lock order checking is disabled. If set to -1, * witness is completely disabled. Otherwise witness performs full * lock order checking for all locks. At runtime, lock order checking * may be toggled. However, witness cannot be reenabled once it is * completely disabled. */ static int witness_watch = 1; SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RWTUN | CTLTYPE_INT, NULL, 0, sysctl_debug_witness_watch, "I", "witness is watching lock operations"); #ifdef KDB /* * When KDB is enabled and witness_kdb is 1, it will cause the system * to drop into kdebug() when: * - a lock hierarchy violation occurs * - locks are held when going to sleep. */ #ifdef WITNESS_KDB int witness_kdb = 1; #else int witness_kdb = 0; #endif SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, ""); #endif /* KDB */ #if defined(DDB) || defined(KDB) /* * When DDB or KDB is enabled and witness_trace is 1, it will cause the system * to print a stack trace: * - a lock hierarchy violation occurs * - locks are held when going to sleep. */ int witness_trace = 1; SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, ""); #endif /* DDB || KDB */ #ifdef WITNESS_SKIPSPIN int witness_skipspin = 1; #else int witness_skipspin = 0; #endif SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, ""); int badstack_sbuf_size; int witness_count = WITNESS_COUNT; SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN, &witness_count, 0, ""); /* * Output channel for witness messages. By default we print to the console. */ enum witness_channel { WITNESS_CONSOLE, WITNESS_LOG, WITNESS_NONE, }; static enum witness_channel witness_channel = WITNESS_CONSOLE; SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel, CTLTYPE_STRING | CTLFLAG_RWTUN, NULL, 0, sysctl_debug_witness_channel, "A", "Output channel for warnings"); /* * Call this to print out the relations between locks. */ SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs"); /* * Call this to print out the witness faulty stacks. */ SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks"); static struct mtx w_mtx; /* w_list */ static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free); static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all); /* w_typelist */ static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin); static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep); /* lock list */ static struct lock_list_entry *w_lock_list_free = NULL; static struct witness_pendhelp pending_locks[WITNESS_PENDLIST]; static u_int pending_cnt; static int w_free_cnt, w_spin_cnt, w_sleep_cnt; SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, ""); SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, ""); SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0, ""); static struct witness *w_data; static uint8_t **w_rmatrix; static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT]; static struct witness_hash w_hash; /* The witness hash table. */ /* The lock order data hash */ static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT]; static struct witness_lock_order_data *w_lofree = NULL; static struct witness_lock_order_hash w_lohash; static int w_max_used_index = 0; static unsigned int w_generation = 0; static const char w_notrunning[] = "Witness not running\n"; static const char w_stillcold[] = "Witness is still cold\n"; #ifdef __i386__ static const char w_notallowed[] = "The sysctl is disabled on the arch\n"; #endif static struct witness_order_list_entry order_lists[] = { /* * sx locks */ { "proctree", &lock_class_sx }, { "allproc", &lock_class_sx }, { "allprison", &lock_class_sx }, { NULL, NULL }, /* * Various mutexes */ { "Giant", &lock_class_mtx_sleep }, { "pipe mutex", &lock_class_mtx_sleep }, { "sigio lock", &lock_class_mtx_sleep }, { "process group", &lock_class_mtx_sleep }, #ifdef HWPMC_HOOKS { "pmc-sleep", &lock_class_mtx_sleep }, #endif { "process lock", &lock_class_mtx_sleep }, { "session", &lock_class_mtx_sleep }, { "uidinfo hash", &lock_class_rw }, { "time lock", &lock_class_mtx_sleep }, { NULL, NULL }, /* * umtx */ { "umtx lock", &lock_class_mtx_sleep }, { NULL, NULL }, /* * Sockets */ { "accept", &lock_class_mtx_sleep }, { "so_snd", &lock_class_mtx_sleep }, { "so_rcv", &lock_class_mtx_sleep }, { "sellck", &lock_class_mtx_sleep }, { NULL, NULL }, /* * Routing */ { "so_rcv", &lock_class_mtx_sleep }, { "radix node head", &lock_class_rm }, { "rtentry", &lock_class_mtx_sleep }, { "ifaddr", &lock_class_mtx_sleep }, { NULL, NULL }, /* * IPv4 multicast: * protocol locks before interface locks, after UDP locks. */ { "in_multi_sx", &lock_class_sx }, { "udpinp", &lock_class_rw }, { "in_multi_list_mtx", &lock_class_mtx_sleep }, { "igmp_mtx", &lock_class_mtx_sleep }, { "ifnet_rw", &lock_class_rw }, { "if_addr_lock", &lock_class_mtx_sleep }, { NULL, NULL }, /* * IPv6 multicast: * protocol locks before interface locks, after UDP locks. */ { "in6_multi_sx", &lock_class_sx }, { "udpinp", &lock_class_rw }, { "in6_multi_list_mtx", &lock_class_mtx_sleep }, { "mld_mtx", &lock_class_mtx_sleep }, { "ifnet_rw", &lock_class_rw }, { "if_addr_lock", &lock_class_mtx_sleep }, { NULL, NULL }, /* * UNIX Domain Sockets */ { "unp_link_rwlock", &lock_class_rw }, { "unp_list_lock", &lock_class_mtx_sleep }, { "unp", &lock_class_mtx_sleep }, { "so_snd", &lock_class_mtx_sleep }, { NULL, NULL }, /* * UDP/IP */ - { "udp", &lock_class_rw }, + { "udp", &lock_class_mtx_sleep }, { "udpinp", &lock_class_rw }, { "so_snd", &lock_class_mtx_sleep }, { NULL, NULL }, /* * TCP/IP */ - { "tcp", &lock_class_rw }, + { "tcp", &lock_class_mtx_sleep }, { "tcpinp", &lock_class_rw }, { "so_snd", &lock_class_mtx_sleep }, { NULL, NULL }, /* * BPF */ { "bpf global lock", &lock_class_sx }, { "bpf interface lock", &lock_class_rw }, { "bpf cdev lock", &lock_class_mtx_sleep }, { NULL, NULL }, /* * NFS server */ { "nfsd_mtx", &lock_class_mtx_sleep }, { "so_snd", &lock_class_mtx_sleep }, { NULL, NULL }, /* * IEEE 802.11 */ { "802.11 com lock", &lock_class_mtx_sleep}, { NULL, NULL }, /* * Network drivers */ { "network driver", &lock_class_mtx_sleep}, { NULL, NULL }, /* * Netgraph */ { "ng_node", &lock_class_mtx_sleep }, { "ng_worklist", &lock_class_mtx_sleep }, { NULL, NULL }, /* * CDEV */ { "vm map (system)", &lock_class_mtx_sleep }, { "vnode interlock", &lock_class_mtx_sleep }, { "cdev", &lock_class_mtx_sleep }, { NULL, NULL }, /* * VM */ { "vm map (user)", &lock_class_sx }, { "vm object", &lock_class_rw }, { "vm page", &lock_class_mtx_sleep }, { "pmap pv global", &lock_class_rw }, { "pmap", &lock_class_mtx_sleep }, { "pmap pv list", &lock_class_rw }, { "vm page free queue", &lock_class_mtx_sleep }, { "vm pagequeue", &lock_class_mtx_sleep }, { NULL, NULL }, /* * kqueue/VFS interaction */ { "kqueue", &lock_class_mtx_sleep }, { "struct mount mtx", &lock_class_mtx_sleep }, { "vnode interlock", &lock_class_mtx_sleep }, { NULL, NULL }, /* * VFS namecache */ { "ncvn", &lock_class_mtx_sleep }, { "ncbuc", &lock_class_rw }, { "vnode interlock", &lock_class_mtx_sleep }, { "ncneg", &lock_class_mtx_sleep }, { NULL, NULL }, /* * ZFS locking */ { "dn->dn_mtx", &lock_class_sx }, { "dr->dt.di.dr_mtx", &lock_class_sx }, { "db->db_mtx", &lock_class_sx }, { NULL, NULL }, /* * TCP log locks */ { "TCP ID tree", &lock_class_rw }, { "tcp log id bucket", &lock_class_mtx_sleep }, { "tcpinp", &lock_class_rw }, { "TCP log expireq", &lock_class_mtx_sleep }, { NULL, NULL }, /* * spin locks */ #ifdef SMP { "ap boot", &lock_class_mtx_spin }, #endif { "rm.mutex_mtx", &lock_class_mtx_spin }, { "sio", &lock_class_mtx_spin }, #ifdef __i386__ { "cy", &lock_class_mtx_spin }, #endif #ifdef __sparc64__ { "pcib_mtx", &lock_class_mtx_spin }, { "rtc_mtx", &lock_class_mtx_spin }, #endif { "scc_hwmtx", &lock_class_mtx_spin }, { "uart_hwmtx", &lock_class_mtx_spin }, { "fast_taskqueue", &lock_class_mtx_spin }, { "intr table", &lock_class_mtx_spin }, { "process slock", &lock_class_mtx_spin }, { "syscons video lock", &lock_class_mtx_spin }, { "sleepq chain", &lock_class_mtx_spin }, { "rm_spinlock", &lock_class_mtx_spin }, { "turnstile chain", &lock_class_mtx_spin }, { "turnstile lock", &lock_class_mtx_spin }, { "sched lock", &lock_class_mtx_spin }, { "td_contested", &lock_class_mtx_spin }, { "callout", &lock_class_mtx_spin }, { "entropy harvest mutex", &lock_class_mtx_spin }, #ifdef SMP { "smp rendezvous", &lock_class_mtx_spin }, #endif #ifdef __powerpc__ { "tlb0", &lock_class_mtx_spin }, #endif { NULL, NULL }, { "sched lock", &lock_class_mtx_spin }, #ifdef HWPMC_HOOKS { "pmc-per-proc", &lock_class_mtx_spin }, #endif { NULL, NULL }, /* * leaf locks */ { "intrcnt", &lock_class_mtx_spin }, { "icu", &lock_class_mtx_spin }, #if defined(SMP) && defined(__sparc64__) { "ipi", &lock_class_mtx_spin }, #endif #ifdef __i386__ { "allpmaps", &lock_class_mtx_spin }, { "descriptor tables", &lock_class_mtx_spin }, #endif { "clk", &lock_class_mtx_spin }, { "cpuset", &lock_class_mtx_spin }, { "mprof lock", &lock_class_mtx_spin }, { "zombie lock", &lock_class_mtx_spin }, { "ALD Queue", &lock_class_mtx_spin }, #if defined(__i386__) || defined(__amd64__) { "pcicfg", &lock_class_mtx_spin }, { "NDIS thread lock", &lock_class_mtx_spin }, #endif { "tw_osl_io_lock", &lock_class_mtx_spin }, { "tw_osl_q_lock", &lock_class_mtx_spin }, { "tw_cl_io_lock", &lock_class_mtx_spin }, { "tw_cl_intr_lock", &lock_class_mtx_spin }, { "tw_cl_gen_lock", &lock_class_mtx_spin }, #ifdef HWPMC_HOOKS { "pmc-leaf", &lock_class_mtx_spin }, #endif { "blocked lock", &lock_class_mtx_spin }, { NULL, NULL }, { NULL, NULL } }; #ifdef BLESSING /* * Pairs of locks which have been blessed * Don't complain about order problems with blessed locks */ static struct witness_blessed blessed_list[] = { }; #endif /* * This global is set to 0 once it becomes safe to use the witness code. */ static int witness_cold = 1; /* * This global is set to 1 once the static lock orders have been enrolled * so that a warning can be issued for any spin locks enrolled later. */ static int witness_spin_warn = 0; /* Trim useless garbage from filenames. */ static const char * fixup_filename(const char *file) { if (file == NULL) return (NULL); while (strncmp(file, "../", 3) == 0) file += 3; return (file); } /* * Calculate the size of early witness structures. */ int witness_startup_count(void) { int sz; sz = sizeof(struct witness) * witness_count; sz += sizeof(*w_rmatrix) * (witness_count + 1); sz += sizeof(*w_rmatrix[0]) * (witness_count + 1) * (witness_count + 1); return (sz); } /* * The WITNESS-enabled diagnostic code. Note that the witness code does * assume that the early boot is single-threaded at least until after this * routine is completed. */ void witness_startup(void *mem) { struct lock_object *lock; struct witness_order_list_entry *order; struct witness *w, *w1; uintptr_t p; int i; p = (uintptr_t)mem; w_data = (void *)p; p += sizeof(struct witness) * witness_count; w_rmatrix = (void *)p; p += sizeof(*w_rmatrix) * (witness_count + 1); for (i = 0; i < witness_count + 1; i++) { w_rmatrix[i] = (void *)p; p += sizeof(*w_rmatrix[i]) * (witness_count + 1); } badstack_sbuf_size = witness_count * 256; /* * We have to release Giant before initializing its witness * structure so that WITNESS doesn't get confused. */ mtx_unlock(&Giant); mtx_assert(&Giant, MA_NOTOWNED); CTR1(KTR_WITNESS, "%s: initializing witness", __func__); mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET | MTX_NOWITNESS | MTX_NOPROFILE); for (i = witness_count - 1; i >= 0; i--) { w = &w_data[i]; memset(w, 0, sizeof(*w)); w_data[i].w_index = i; /* Witness index never changes. */ witness_free(w); } KASSERT(STAILQ_FIRST(&w_free)->w_index == 0, ("%s: Invalid list of free witness objects", __func__)); /* Witness with index 0 is not used to aid in debugging. */ STAILQ_REMOVE_HEAD(&w_free, w_list); w_free_cnt--; for (i = 0; i < witness_count; i++) { memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) * (witness_count + 1)); } for (i = 0; i < LOCK_CHILDCOUNT; i++) witness_lock_list_free(&w_locklistdata[i]); witness_init_hash_tables(); /* First add in all the specified order lists. */ for (order = order_lists; order->w_name != NULL; order++) { w = enroll(order->w_name, order->w_class); if (w == NULL) continue; w->w_file = "order list"; for (order++; order->w_name != NULL; order++) { w1 = enroll(order->w_name, order->w_class); if (w1 == NULL) continue; w1->w_file = "order list"; itismychild(w, w1); w = w1; } } witness_spin_warn = 1; /* Iterate through all locks and add them to witness. */ for (i = 0; pending_locks[i].wh_lock != NULL; i++) { lock = pending_locks[i].wh_lock; KASSERT(lock->lo_flags & LO_WITNESS, ("%s: lock %s is on pending list but not LO_WITNESS", __func__, lock->lo_name)); lock->lo_witness = enroll(pending_locks[i].wh_type, LOCK_CLASS(lock)); } /* Mark the witness code as being ready for use. */ witness_cold = 0; mtx_lock(&Giant); } void witness_init(struct lock_object *lock, const char *type) { struct lock_class *class; /* Various sanity checks. */ class = LOCK_CLASS(lock); if ((lock->lo_flags & LO_RECURSABLE) != 0 && (class->lc_flags & LC_RECURSABLE) == 0) kassert_panic("%s: lock (%s) %s can not be recursable", __func__, class->lc_name, lock->lo_name); if ((lock->lo_flags & LO_SLEEPABLE) != 0 && (class->lc_flags & LC_SLEEPABLE) == 0) kassert_panic("%s: lock (%s) %s can not be sleepable", __func__, class->lc_name, lock->lo_name); if ((lock->lo_flags & LO_UPGRADABLE) != 0 && (class->lc_flags & LC_UPGRADABLE) == 0) kassert_panic("%s: lock (%s) %s can not be upgradable", __func__, class->lc_name, lock->lo_name); /* * If we shouldn't watch this lock, then just clear lo_witness. * Otherwise, if witness_cold is set, then it is too early to * enroll this lock, so defer it to witness_initialize() by adding * it to the pending_locks list. If it is not too early, then enroll * the lock now. */ if (witness_watch < 1 || panicstr != NULL || (lock->lo_flags & LO_WITNESS) == 0) lock->lo_witness = NULL; else if (witness_cold) { pending_locks[pending_cnt].wh_lock = lock; pending_locks[pending_cnt++].wh_type = type; if (pending_cnt > WITNESS_PENDLIST) panic("%s: pending locks list is too small, " "increase WITNESS_PENDLIST\n", __func__); } else lock->lo_witness = enroll(type, class); } void witness_destroy(struct lock_object *lock) { struct lock_class *class; struct witness *w; class = LOCK_CLASS(lock); if (witness_cold) panic("lock (%s) %s destroyed while witness_cold", class->lc_name, lock->lo_name); /* XXX: need to verify that no one holds the lock */ if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL) return; w = lock->lo_witness; mtx_lock_spin(&w_mtx); MPASS(w->w_refcount > 0); w->w_refcount--; if (w->w_refcount == 0) depart(w); mtx_unlock_spin(&w_mtx); } #ifdef DDB static void witness_ddb_compute_levels(void) { struct witness *w; /* * First clear all levels. */ STAILQ_FOREACH(w, &w_all, w_list) w->w_ddb_level = -1; /* * Look for locks with no parents and level all their descendants. */ STAILQ_FOREACH(w, &w_all, w_list) { /* If the witness has ancestors (is not a root), skip it. */ if (w->w_num_ancestors > 0) continue; witness_ddb_level_descendants(w, 0); } } static void witness_ddb_level_descendants(struct witness *w, int l) { int i; if (w->w_ddb_level >= l) return; w->w_ddb_level = l; l++; for (i = 1; i <= w_max_used_index; i++) { if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) witness_ddb_level_descendants(&w_data[i], l); } } static void witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...), struct witness *w, int indent) { int i; for (i = 0; i < indent; i++) prnt(" "); prnt("%s (type: %s, depth: %d, active refs: %d)", w->w_name, w->w_class->lc_name, w->w_ddb_level, w->w_refcount); if (w->w_displayed) { prnt(" -- (already displayed)\n"); return; } w->w_displayed = 1; if (w->w_file != NULL && w->w_line != 0) prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file), w->w_line); else prnt(" -- never acquired\n"); indent++; WITNESS_INDEX_ASSERT(w->w_index); for (i = 1; i <= w_max_used_index; i++) { if (db_pager_quit) return; if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) witness_ddb_display_descendants(prnt, &w_data[i], indent); } } static void witness_ddb_display_list(int(*prnt)(const char *fmt, ...), struct witness_list *list) { struct witness *w; STAILQ_FOREACH(w, list, w_typelist) { if (w->w_file == NULL || w->w_ddb_level > 0) continue; /* This lock has no anscestors - display its descendants. */ witness_ddb_display_descendants(prnt, w, 0); if (db_pager_quit) return; } } static void witness_ddb_display(int(*prnt)(const char *fmt, ...)) { struct witness *w; KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); witness_ddb_compute_levels(); /* Clear all the displayed flags. */ STAILQ_FOREACH(w, &w_all, w_list) w->w_displayed = 0; /* * First, handle sleep locks which have been acquired at least * once. */ prnt("Sleep locks:\n"); witness_ddb_display_list(prnt, &w_sleep); if (db_pager_quit) return; /* * Now do spin locks which have been acquired at least once. */ prnt("\nSpin locks:\n"); witness_ddb_display_list(prnt, &w_spin); if (db_pager_quit) return; /* * Finally, any locks which have not been acquired yet. */ prnt("\nLocks which were never acquired:\n"); STAILQ_FOREACH(w, &w_all, w_list) { if (w->w_file != NULL || w->w_refcount == 0) continue; prnt("%s (type: %s, depth: %d)\n", w->w_name, w->w_class->lc_name, w->w_ddb_level); if (db_pager_quit) return; } } #endif /* DDB */ int witness_defineorder(struct lock_object *lock1, struct lock_object *lock2) { if (witness_watch == -1 || panicstr != NULL) return (0); /* Require locks that witness knows about. */ if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL || lock2->lo_witness == NULL) return (EINVAL); mtx_assert(&w_mtx, MA_NOTOWNED); mtx_lock_spin(&w_mtx); /* * If we already have either an explicit or implied lock order that * is the other way around, then return an error. */ if (witness_watch && isitmydescendant(lock2->lo_witness, lock1->lo_witness)) { mtx_unlock_spin(&w_mtx); return (EDOOFUS); } /* Try to add the new order. */ CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__, lock2->lo_witness->w_name, lock1->lo_witness->w_name); itismychild(lock1->lo_witness, lock2->lo_witness); mtx_unlock_spin(&w_mtx); return (0); } void witness_checkorder(struct lock_object *lock, int flags, const char *file, int line, struct lock_object *interlock) { struct lock_list_entry *lock_list, *lle; struct lock_instance *lock1, *lock2, *plock; struct lock_class *class, *iclass; struct witness *w, *w1; struct thread *td; int i, j; if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL || panicstr != NULL) return; w = lock->lo_witness; class = LOCK_CLASS(lock); td = curthread; if (class->lc_flags & LC_SLEEPLOCK) { /* * Since spin locks include a critical section, this check * implicitly enforces a lock order of all sleep locks before * all spin locks. */ if (td->td_critnest != 0 && !kdb_active) kassert_panic("acquiring blockable sleep lock with " "spinlock or critical section held (%s) %s @ %s:%d", class->lc_name, lock->lo_name, fixup_filename(file), line); /* * If this is the first lock acquired then just return as * no order checking is needed. */ lock_list = td->td_sleeplocks; if (lock_list == NULL || lock_list->ll_count == 0) return; } else { /* * If this is the first lock, just return as no order * checking is needed. Avoid problems with thread * migration pinning the thread while checking if * spinlocks are held. If at least one spinlock is held * the thread is in a safe path and it is allowed to * unpin it. */ sched_pin(); lock_list = PCPU_GET(spinlocks); if (lock_list == NULL || lock_list->ll_count == 0) { sched_unpin(); return; } sched_unpin(); } /* * Check to see if we are recursing on a lock we already own. If * so, make sure that we don't mismatch exclusive and shared lock * acquires. */ lock1 = find_instance(lock_list, lock); if (lock1 != NULL) { if ((lock1->li_flags & LI_EXCLUSIVE) != 0 && (flags & LOP_EXCLUSIVE) == 0) { witness_output("shared lock of (%s) %s @ %s:%d\n", class->lc_name, lock->lo_name, fixup_filename(file), line); witness_output("while exclusively locked from %s:%d\n", fixup_filename(lock1->li_file), lock1->li_line); kassert_panic("excl->share"); } if ((lock1->li_flags & LI_EXCLUSIVE) == 0 && (flags & LOP_EXCLUSIVE) != 0) { witness_output("exclusive lock of (%s) %s @ %s:%d\n", class->lc_name, lock->lo_name, fixup_filename(file), line); witness_output("while share locked from %s:%d\n", fixup_filename(lock1->li_file), lock1->li_line); kassert_panic("share->excl"); } return; } /* Warn if the interlock is not locked exactly once. */ if (interlock != NULL) { iclass = LOCK_CLASS(interlock); lock1 = find_instance(lock_list, interlock); if (lock1 == NULL) kassert_panic("interlock (%s) %s not locked @ %s:%d", iclass->lc_name, interlock->lo_name, fixup_filename(file), line); else if ((lock1->li_flags & LI_RECURSEMASK) != 0) kassert_panic("interlock (%s) %s recursed @ %s:%d", iclass->lc_name, interlock->lo_name, fixup_filename(file), line); } /* * Find the previously acquired lock, but ignore interlocks. */ plock = &lock_list->ll_children[lock_list->ll_count - 1]; if (interlock != NULL && plock->li_lock == interlock) { if (lock_list->ll_count > 1) plock = &lock_list->ll_children[lock_list->ll_count - 2]; else { lle = lock_list->ll_next; /* * The interlock is the only lock we hold, so * simply return. */ if (lle == NULL) return; plock = &lle->ll_children[lle->ll_count - 1]; } } /* * Try to perform most checks without a lock. If this succeeds we * can skip acquiring the lock and return success. Otherwise we redo * the check with the lock held to handle races with concurrent updates. */ w1 = plock->li_lock->lo_witness; if (witness_lock_order_check(w1, w)) return; mtx_lock_spin(&w_mtx); if (witness_lock_order_check(w1, w)) { mtx_unlock_spin(&w_mtx); return; } witness_lock_order_add(w1, w); /* * Check for duplicate locks of the same type. Note that we only * have to check for this on the last lock we just acquired. Any * other cases will be caught as lock order violations. */ if (w1 == w) { i = w->w_index; if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) && !(w_rmatrix[i][i] & WITNESS_REVERSAL)) { w_rmatrix[i][i] |= WITNESS_REVERSAL; w->w_reversed = 1; mtx_unlock_spin(&w_mtx); witness_output( "acquiring duplicate lock of same type: \"%s\"\n", w->w_name); witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name, fixup_filename(plock->li_file), plock->li_line); witness_output(" 2nd %s @ %s:%d\n", lock->lo_name, fixup_filename(file), line); witness_debugger(1, __func__); } else mtx_unlock_spin(&w_mtx); return; } mtx_assert(&w_mtx, MA_OWNED); /* * If we know that the lock we are acquiring comes after * the lock we most recently acquired in the lock order tree, * then there is no need for any further checks. */ if (isitmychild(w1, w)) goto out; for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) { for (i = lle->ll_count - 1; i >= 0; i--, j++) { MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN); lock1 = &lle->ll_children[i]; /* * Ignore the interlock. */ if (interlock == lock1->li_lock) continue; /* * If this lock doesn't undergo witness checking, * then skip it. */ w1 = lock1->li_lock->lo_witness; if (w1 == NULL) { KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0, ("lock missing witness structure")); continue; } /* * If we are locking Giant and this is a sleepable * lock, then skip it. */ if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 && lock == &Giant.lock_object) continue; /* * If we are locking a sleepable lock and this lock * is Giant, then skip it. */ if ((lock->lo_flags & LO_SLEEPABLE) != 0 && lock1->li_lock == &Giant.lock_object) continue; /* * If we are locking a sleepable lock and this lock * isn't sleepable, we want to treat it as a lock * order violation to enfore a general lock order of * sleepable locks before non-sleepable locks. */ if (((lock->lo_flags & LO_SLEEPABLE) != 0 && (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0)) goto reversal; /* * If we are locking Giant and this is a non-sleepable * lock, then treat it as a reversal. */ if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 && lock == &Giant.lock_object) goto reversal; /* * Check the lock order hierarchy for a reveresal. */ if (!isitmydescendant(w, w1)) continue; reversal: /* * We have a lock order violation, check to see if it * is allowed or has already been yelled about. */ #ifdef BLESSING /* * If the lock order is blessed, just bail. We don't * look for other lock order violations though, which * may be a bug. */ if (blessed(w, w1)) goto out; #endif /* Bail if this violation is known */ if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL) goto out; /* Record this as a violation */ w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL; w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL; w->w_reversed = w1->w_reversed = 1; witness_increment_graph_generation(); mtx_unlock_spin(&w_mtx); #ifdef WITNESS_NO_VNODE /* * There are known LORs between VNODE locks. They are * not an indication of a bug. VNODE locks are flagged * as such (LO_IS_VNODE) and we don't yell if the LOR * is between 2 VNODE locks. */ if ((lock->lo_flags & LO_IS_VNODE) != 0 && (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0) return; #endif /* * Ok, yell about it. */ if (((lock->lo_flags & LO_SLEEPABLE) != 0 && (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0)) witness_output( "lock order reversal: (sleepable after non-sleepable)\n"); else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 && lock == &Giant.lock_object) witness_output( "lock order reversal: (Giant after non-sleepable)\n"); else witness_output("lock order reversal:\n"); /* * Try to locate an earlier lock with * witness w in our list. */ do { lock2 = &lle->ll_children[i]; MPASS(lock2->li_lock != NULL); if (lock2->li_lock->lo_witness == w) break; if (i == 0 && lle->ll_next != NULL) { lle = lle->ll_next; i = lle->ll_count - 1; MPASS(i >= 0 && i < LOCK_NCHILDREN); } else i--; } while (i >= 0); if (i < 0) { witness_output(" 1st %p %s (%s) @ %s:%d\n", lock1->li_lock, lock1->li_lock->lo_name, w1->w_name, fixup_filename(lock1->li_file), lock1->li_line); witness_output(" 2nd %p %s (%s) @ %s:%d\n", lock, lock->lo_name, w->w_name, fixup_filename(file), line); } else { witness_output(" 1st %p %s (%s) @ %s:%d\n", lock2->li_lock, lock2->li_lock->lo_name, lock2->li_lock->lo_witness->w_name, fixup_filename(lock2->li_file), lock2->li_line); witness_output(" 2nd %p %s (%s) @ %s:%d\n", lock1->li_lock, lock1->li_lock->lo_name, w1->w_name, fixup_filename(lock1->li_file), lock1->li_line); witness_output(" 3rd %p %s (%s) @ %s:%d\n", lock, lock->lo_name, w->w_name, fixup_filename(file), line); } witness_debugger(1, __func__); return; } } /* * If requested, build a new lock order. However, don't build a new * relationship between a sleepable lock and Giant if it is in the * wrong direction. The correct lock order is that sleepable locks * always come before Giant. */ if (flags & LOP_NEWORDER && !(plock->li_lock == &Giant.lock_object && (lock->lo_flags & LO_SLEEPABLE) != 0)) { CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__, w->w_name, plock->li_lock->lo_witness->w_name); itismychild(plock->li_lock->lo_witness, w); } out: mtx_unlock_spin(&w_mtx); } void witness_lock(struct lock_object *lock, int flags, const char *file, int line) { struct lock_list_entry **lock_list, *lle; struct lock_instance *instance; struct witness *w; struct thread *td; if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL || panicstr != NULL) return; w = lock->lo_witness; td = curthread; /* Determine lock list for this lock. */ if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK) lock_list = &td->td_sleeplocks; else lock_list = PCPU_PTR(spinlocks); /* Check to see if we are recursing on a lock we already own. */ instance = find_instance(*lock_list, lock); if (instance != NULL) { instance->li_flags++; CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__, td->td_proc->p_pid, lock->lo_name, instance->li_flags & LI_RECURSEMASK); instance->li_file = file; instance->li_line = line; return; } /* Update per-witness last file and line acquire. */ w->w_file = file; w->w_line = line; /* Find the next open lock instance in the list and fill it. */ lle = *lock_list; if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) { lle = witness_lock_list_get(); if (lle == NULL) return; lle->ll_next = *lock_list; CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__, td->td_proc->p_pid, lle); *lock_list = lle; } instance = &lle->ll_children[lle->ll_count++]; instance->li_lock = lock; instance->li_line = line; instance->li_file = file; if ((flags & LOP_EXCLUSIVE) != 0) instance->li_flags = LI_EXCLUSIVE; else instance->li_flags = 0; CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__, td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1); } void witness_upgrade(struct lock_object *lock, int flags, const char *file, int line) { struct lock_instance *instance; struct lock_class *class; KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) return; class = LOCK_CLASS(lock); if (witness_watch) { if ((lock->lo_flags & LO_UPGRADABLE) == 0) kassert_panic( "upgrade of non-upgradable lock (%s) %s @ %s:%d", class->lc_name, lock->lo_name, fixup_filename(file), line); if ((class->lc_flags & LC_SLEEPLOCK) == 0) kassert_panic( "upgrade of non-sleep lock (%s) %s @ %s:%d", class->lc_name, lock->lo_name, fixup_filename(file), line); } instance = find_instance(curthread->td_sleeplocks, lock); if (instance == NULL) { kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d", class->lc_name, lock->lo_name, fixup_filename(file), line); return; } if (witness_watch) { if ((instance->li_flags & LI_EXCLUSIVE) != 0) kassert_panic( "upgrade of exclusive lock (%s) %s @ %s:%d", class->lc_name, lock->lo_name, fixup_filename(file), line); if ((instance->li_flags & LI_RECURSEMASK) != 0) kassert_panic( "upgrade of recursed lock (%s) %s r=%d @ %s:%d", class->lc_name, lock->lo_name, instance->li_flags & LI_RECURSEMASK, fixup_filename(file), line); } instance->li_flags |= LI_EXCLUSIVE; } void witness_downgrade(struct lock_object *lock, int flags, const char *file, int line) { struct lock_instance *instance; struct lock_class *class; KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) return; class = LOCK_CLASS(lock); if (witness_watch) { if ((lock->lo_flags & LO_UPGRADABLE) == 0) kassert_panic( "downgrade of non-upgradable lock (%s) %s @ %s:%d", class->lc_name, lock->lo_name, fixup_filename(file), line); if ((class->lc_flags & LC_SLEEPLOCK) == 0) kassert_panic( "downgrade of non-sleep lock (%s) %s @ %s:%d", class->lc_name, lock->lo_name, fixup_filename(file), line); } instance = find_instance(curthread->td_sleeplocks, lock); if (instance == NULL) { kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d", class->lc_name, lock->lo_name, fixup_filename(file), line); return; } if (witness_watch) { if ((instance->li_flags & LI_EXCLUSIVE) == 0) kassert_panic( "downgrade of shared lock (%s) %s @ %s:%d", class->lc_name, lock->lo_name, fixup_filename(file), line); if ((instance->li_flags & LI_RECURSEMASK) != 0) kassert_panic( "downgrade of recursed lock (%s) %s r=%d @ %s:%d", class->lc_name, lock->lo_name, instance->li_flags & LI_RECURSEMASK, fixup_filename(file), line); } instance->li_flags &= ~LI_EXCLUSIVE; } void witness_unlock(struct lock_object *lock, int flags, const char *file, int line) { struct lock_list_entry **lock_list, *lle; struct lock_instance *instance; struct lock_class *class; struct thread *td; register_t s; int i, j; if (witness_cold || lock->lo_witness == NULL || panicstr != NULL) return; td = curthread; class = LOCK_CLASS(lock); /* Find lock instance associated with this lock. */ if (class->lc_flags & LC_SLEEPLOCK) lock_list = &td->td_sleeplocks; else lock_list = PCPU_PTR(spinlocks); lle = *lock_list; for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next) for (i = 0; i < (*lock_list)->ll_count; i++) { instance = &(*lock_list)->ll_children[i]; if (instance->li_lock == lock) goto found; } /* * When disabling WITNESS through witness_watch we could end up in * having registered locks in the td_sleeplocks queue. * We have to make sure we flush these queues, so just search for * eventual register locks and remove them. */ if (witness_watch > 0) { kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name, fixup_filename(file), line); return; } else { return; } found: /* First, check for shared/exclusive mismatches. */ if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 && (flags & LOP_EXCLUSIVE) == 0) { witness_output("shared unlock of (%s) %s @ %s:%d\n", class->lc_name, lock->lo_name, fixup_filename(file), line); witness_output("while exclusively locked from %s:%d\n", fixup_filename(instance->li_file), instance->li_line); kassert_panic("excl->ushare"); } if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 && (flags & LOP_EXCLUSIVE) != 0) { witness_output("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name, lock->lo_name, fixup_filename(file), line); witness_output("while share locked from %s:%d\n", fixup_filename(instance->li_file), instance->li_line); kassert_panic("share->uexcl"); } /* If we are recursed, unrecurse. */ if ((instance->li_flags & LI_RECURSEMASK) > 0) { CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__, td->td_proc->p_pid, instance->li_lock->lo_name, instance->li_flags); instance->li_flags--; return; } /* The lock is now being dropped, check for NORELEASE flag */ if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) { witness_output("forbidden unlock of (%s) %s @ %s:%d\n", class->lc_name, lock->lo_name, fixup_filename(file), line); kassert_panic("lock marked norelease"); } /* Otherwise, remove this item from the list. */ s = intr_disable(); CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__, td->td_proc->p_pid, instance->li_lock->lo_name, (*lock_list)->ll_count - 1); for (j = i; j < (*lock_list)->ll_count - 1; j++) (*lock_list)->ll_children[j] = (*lock_list)->ll_children[j + 1]; (*lock_list)->ll_count--; intr_restore(s); /* * In order to reduce contention on w_mtx, we want to keep always an * head object into lists so that frequent allocation from the * free witness pool (and subsequent locking) is avoided. * In order to maintain the current code simple, when the head * object is totally unloaded it means also that we do not have * further objects in the list, so the list ownership needs to be * hand over to another object if the current head needs to be freed. */ if ((*lock_list)->ll_count == 0) { if (*lock_list == lle) { if (lle->ll_next == NULL) return; } else lle = *lock_list; *lock_list = lle->ll_next; CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__, td->td_proc->p_pid, lle); witness_lock_list_free(lle); } } void witness_thread_exit(struct thread *td) { struct lock_list_entry *lle; int i, n; lle = td->td_sleeplocks; if (lle == NULL || panicstr != NULL) return; if (lle->ll_count != 0) { for (n = 0; lle != NULL; lle = lle->ll_next) for (i = lle->ll_count - 1; i >= 0; i--) { if (n == 0) witness_output( "Thread %p exiting with the following locks held:\n", td); n++; witness_list_lock(&lle->ll_children[i], witness_output); } kassert_panic( "Thread %p cannot exit while holding sleeplocks\n", td); } witness_lock_list_free(lle); } /* * Warn if any locks other than 'lock' are held. Flags can be passed in to * exempt Giant and sleepable locks from the checks as well. If any * non-exempt locks are held, then a supplied message is printed to the * output channel along with a list of the offending locks. If indicated in the * flags then a failure results in a panic as well. */ int witness_warn(int flags, struct lock_object *lock, const char *fmt, ...) { struct lock_list_entry *lock_list, *lle; struct lock_instance *lock1; struct thread *td; va_list ap; int i, n; if (witness_cold || witness_watch < 1 || panicstr != NULL) return (0); n = 0; td = curthread; for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next) for (i = lle->ll_count - 1; i >= 0; i--) { lock1 = &lle->ll_children[i]; if (lock1->li_lock == lock) continue; if (flags & WARN_GIANTOK && lock1->li_lock == &Giant.lock_object) continue; if (flags & WARN_SLEEPOK && (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) continue; if (n == 0) { va_start(ap, fmt); vprintf(fmt, ap); va_end(ap); printf(" with the following %slocks held:\n", (flags & WARN_SLEEPOK) != 0 ? "non-sleepable " : ""); } n++; witness_list_lock(lock1, printf); } /* * Pin the thread in order to avoid problems with thread migration. * Once that all verifies are passed about spinlocks ownership, * the thread is in a safe path and it can be unpinned. */ sched_pin(); lock_list = PCPU_GET(spinlocks); if (lock_list != NULL && lock_list->ll_count != 0) { sched_unpin(); /* * We should only have one spinlock and as long as * the flags cannot match for this locks class, * check if the first spinlock is the one curthread * should hold. */ lock1 = &lock_list->ll_children[lock_list->ll_count - 1]; if (lock_list->ll_count == 1 && lock_list->ll_next == NULL && lock1->li_lock == lock && n == 0) return (0); va_start(ap, fmt); vprintf(fmt, ap); va_end(ap); printf(" with the following %slocks held:\n", (flags & WARN_SLEEPOK) != 0 ? "non-sleepable " : ""); n += witness_list_locks(&lock_list, printf); } else sched_unpin(); if (flags & WARN_PANIC && n) kassert_panic("%s", __func__); else witness_debugger(n, __func__); return (n); } const char * witness_file(struct lock_object *lock) { struct witness *w; if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL) return ("?"); w = lock->lo_witness; return (w->w_file); } int witness_line(struct lock_object *lock) { struct witness *w; if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL) return (0); w = lock->lo_witness; return (w->w_line); } static struct witness * enroll(const char *description, struct lock_class *lock_class) { struct witness *w; MPASS(description != NULL); if (witness_watch == -1 || panicstr != NULL) return (NULL); if ((lock_class->lc_flags & LC_SPINLOCK)) { if (witness_skipspin) return (NULL); } else if ((lock_class->lc_flags & LC_SLEEPLOCK) == 0) { kassert_panic("lock class %s is not sleep or spin", lock_class->lc_name); return (NULL); } mtx_lock_spin(&w_mtx); w = witness_hash_get(description); if (w) goto found; if ((w = witness_get()) == NULL) return (NULL); MPASS(strlen(description) < MAX_W_NAME); strcpy(w->w_name, description); w->w_class = lock_class; w->w_refcount = 1; STAILQ_INSERT_HEAD(&w_all, w, w_list); if (lock_class->lc_flags & LC_SPINLOCK) { STAILQ_INSERT_HEAD(&w_spin, w, w_typelist); w_spin_cnt++; } else if (lock_class->lc_flags & LC_SLEEPLOCK) { STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist); w_sleep_cnt++; } /* Insert new witness into the hash */ witness_hash_put(w); witness_increment_graph_generation(); mtx_unlock_spin(&w_mtx); return (w); found: w->w_refcount++; if (w->w_refcount == 1) w->w_class = lock_class; mtx_unlock_spin(&w_mtx); if (lock_class != w->w_class) kassert_panic( "lock (%s) %s does not match earlier (%s) lock", description, lock_class->lc_name, w->w_class->lc_name); return (w); } static void depart(struct witness *w) { MPASS(w->w_refcount == 0); if (w->w_class->lc_flags & LC_SLEEPLOCK) { w_sleep_cnt--; } else { w_spin_cnt--; } /* * Set file to NULL as it may point into a loadable module. */ w->w_file = NULL; w->w_line = 0; witness_increment_graph_generation(); } static void adopt(struct witness *parent, struct witness *child) { int pi, ci, i, j; if (witness_cold == 0) mtx_assert(&w_mtx, MA_OWNED); /* If the relationship is already known, there's no work to be done. */ if (isitmychild(parent, child)) return; /* When the structure of the graph changes, bump up the generation. */ witness_increment_graph_generation(); /* * The hard part ... create the direct relationship, then propagate all * indirect relationships. */ pi = parent->w_index; ci = child->w_index; WITNESS_INDEX_ASSERT(pi); WITNESS_INDEX_ASSERT(ci); MPASS(pi != ci); w_rmatrix[pi][ci] |= WITNESS_PARENT; w_rmatrix[ci][pi] |= WITNESS_CHILD; /* * If parent was not already an ancestor of child, * then we increment the descendant and ancestor counters. */ if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) { parent->w_num_descendants++; child->w_num_ancestors++; } /* * Find each ancestor of 'pi'. Note that 'pi' itself is counted as * an ancestor of 'pi' during this loop. */ for (i = 1; i <= w_max_used_index; i++) { if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 && (i != pi)) continue; /* Find each descendant of 'i' and mark it as a descendant. */ for (j = 1; j <= w_max_used_index; j++) { /* * Skip children that are already marked as * descendants of 'i'. */ if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) continue; /* * We are only interested in descendants of 'ci'. Note * that 'ci' itself is counted as a descendant of 'ci'. */ if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 && (j != ci)) continue; w_rmatrix[i][j] |= WITNESS_ANCESTOR; w_rmatrix[j][i] |= WITNESS_DESCENDANT; w_data[i].w_num_descendants++; w_data[j].w_num_ancestors++; /* * Make sure we aren't marking a node as both an * ancestor and descendant. We should have caught * this as a lock order reversal earlier. */ if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) && (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) { printf("witness rmatrix paradox! [%d][%d]=%d " "both ancestor and descendant\n", i, j, w_rmatrix[i][j]); kdb_backtrace(); printf("Witness disabled.\n"); witness_watch = -1; } if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) && (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) { printf("witness rmatrix paradox! [%d][%d]=%d " "both ancestor and descendant\n", j, i, w_rmatrix[j][i]); kdb_backtrace(); printf("Witness disabled.\n"); witness_watch = -1; } } } } static void itismychild(struct witness *parent, struct witness *child) { int unlocked; MPASS(child != NULL && parent != NULL); if (witness_cold == 0) mtx_assert(&w_mtx, MA_OWNED); if (!witness_lock_type_equal(parent, child)) { if (witness_cold == 0) { unlocked = 1; mtx_unlock_spin(&w_mtx); } else { unlocked = 0; } kassert_panic( "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not " "the same lock type", __func__, parent->w_name, parent->w_class->lc_name, child->w_name, child->w_class->lc_name); if (unlocked) mtx_lock_spin(&w_mtx); } adopt(parent, child); } /* * Generic code for the isitmy*() functions. The rmask parameter is the * expected relationship of w1 to w2. */ static int _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname) { unsigned char r1, r2; int i1, i2; i1 = w1->w_index; i2 = w2->w_index; WITNESS_INDEX_ASSERT(i1); WITNESS_INDEX_ASSERT(i2); r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK; r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK; /* The flags on one better be the inverse of the flags on the other */ if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) || (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) { /* Don't squawk if we're potentially racing with an update. */ if (!mtx_owned(&w_mtx)) return (0); printf("%s: rmatrix mismatch between %s (index %d) and %s " "(index %d): w_rmatrix[%d][%d] == %hhx but " "w_rmatrix[%d][%d] == %hhx\n", fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1, i2, i1, r2); kdb_backtrace(); printf("Witness disabled.\n"); witness_watch = -1; } return (r1 & rmask); } /* * Checks if @child is a direct child of @parent. */ static int isitmychild(struct witness *parent, struct witness *child) { return (_isitmyx(parent, child, WITNESS_PARENT, __func__)); } /* * Checks if @descendant is a direct or inderect descendant of @ancestor. */ static int isitmydescendant(struct witness *ancestor, struct witness *descendant) { return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK, __func__)); } #ifdef BLESSING static int blessed(struct witness *w1, struct witness *w2) { int i; struct witness_blessed *b; for (i = 0; i < nitems(blessed_list); i++) { b = &blessed_list[i]; if (strcmp(w1->w_name, b->b_lock1) == 0) { if (strcmp(w2->w_name, b->b_lock2) == 0) return (1); continue; } if (strcmp(w1->w_name, b->b_lock2) == 0) if (strcmp(w2->w_name, b->b_lock1) == 0) return (1); } return (0); } #endif static struct witness * witness_get(void) { struct witness *w; int index; if (witness_cold == 0) mtx_assert(&w_mtx, MA_OWNED); if (witness_watch == -1) { mtx_unlock_spin(&w_mtx); return (NULL); } if (STAILQ_EMPTY(&w_free)) { witness_watch = -1; mtx_unlock_spin(&w_mtx); printf("WITNESS: unable to allocate a new witness object\n"); return (NULL); } w = STAILQ_FIRST(&w_free); STAILQ_REMOVE_HEAD(&w_free, w_list); w_free_cnt--; index = w->w_index; MPASS(index > 0 && index == w_max_used_index+1 && index < witness_count); bzero(w, sizeof(*w)); w->w_index = index; if (index > w_max_used_index) w_max_used_index = index; return (w); } static void witness_free(struct witness *w) { STAILQ_INSERT_HEAD(&w_free, w, w_list); w_free_cnt++; } static struct lock_list_entry * witness_lock_list_get(void) { struct lock_list_entry *lle; if (witness_watch == -1) return (NULL); mtx_lock_spin(&w_mtx); lle = w_lock_list_free; if (lle == NULL) { witness_watch = -1; mtx_unlock_spin(&w_mtx); printf("%s: witness exhausted\n", __func__); return (NULL); } w_lock_list_free = lle->ll_next; mtx_unlock_spin(&w_mtx); bzero(lle, sizeof(*lle)); return (lle); } static void witness_lock_list_free(struct lock_list_entry *lle) { mtx_lock_spin(&w_mtx); lle->ll_next = w_lock_list_free; w_lock_list_free = lle; mtx_unlock_spin(&w_mtx); } static struct lock_instance * find_instance(struct lock_list_entry *list, const struct lock_object *lock) { struct lock_list_entry *lle; struct lock_instance *instance; int i; for (lle = list; lle != NULL; lle = lle->ll_next) for (i = lle->ll_count - 1; i >= 0; i--) { instance = &lle->ll_children[i]; if (instance->li_lock == lock) return (instance); } return (NULL); } static void witness_list_lock(struct lock_instance *instance, int (*prnt)(const char *fmt, ...)) { struct lock_object *lock; lock = instance->li_lock; prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ? "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name); if (lock->lo_witness->w_name != lock->lo_name) prnt(" (%s)", lock->lo_witness->w_name); prnt(" r = %d (%p) locked @ %s:%d\n", instance->li_flags & LI_RECURSEMASK, lock, fixup_filename(instance->li_file), instance->li_line); } static int witness_output(const char *fmt, ...) { va_list ap; int ret; va_start(ap, fmt); ret = witness_voutput(fmt, ap); va_end(ap); return (ret); } static int witness_voutput(const char *fmt, va_list ap) { int ret; ret = 0; switch (witness_channel) { case WITNESS_CONSOLE: ret = vprintf(fmt, ap); break; case WITNESS_LOG: vlog(LOG_NOTICE, fmt, ap); break; case WITNESS_NONE: break; } return (ret); } #ifdef DDB static int witness_thread_has_locks(struct thread *td) { if (td->td_sleeplocks == NULL) return (0); return (td->td_sleeplocks->ll_count != 0); } static int witness_proc_has_locks(struct proc *p) { struct thread *td; FOREACH_THREAD_IN_PROC(p, td) { if (witness_thread_has_locks(td)) return (1); } return (0); } #endif int witness_list_locks(struct lock_list_entry **lock_list, int (*prnt)(const char *fmt, ...)) { struct lock_list_entry *lle; int i, nheld; nheld = 0; for (lle = *lock_list; lle != NULL; lle = lle->ll_next) for (i = lle->ll_count - 1; i >= 0; i--) { witness_list_lock(&lle->ll_children[i], prnt); nheld++; } return (nheld); } /* * This is a bit risky at best. We call this function when we have timed * out acquiring a spin lock, and we assume that the other CPU is stuck * with this lock held. So, we go groveling around in the other CPU's * per-cpu data to try to find the lock instance for this spin lock to * see when it was last acquired. */ void witness_display_spinlock(struct lock_object *lock, struct thread *owner, int (*prnt)(const char *fmt, ...)) { struct lock_instance *instance; struct pcpu *pc; if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU) return; pc = pcpu_find(owner->td_oncpu); instance = find_instance(pc->pc_spinlocks, lock); if (instance != NULL) witness_list_lock(instance, prnt); } void witness_save(struct lock_object *lock, const char **filep, int *linep) { struct lock_list_entry *lock_list; struct lock_instance *instance; struct lock_class *class; /* * This function is used independently in locking code to deal with * Giant, SCHEDULER_STOPPED() check can be removed here after Giant * is gone. */ if (SCHEDULER_STOPPED()) return; KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) return; class = LOCK_CLASS(lock); if (class->lc_flags & LC_SLEEPLOCK) lock_list = curthread->td_sleeplocks; else { if (witness_skipspin) return; lock_list = PCPU_GET(spinlocks); } instance = find_instance(lock_list, lock); if (instance == NULL) { kassert_panic("%s: lock (%s) %s not locked", __func__, class->lc_name, lock->lo_name); return; } *filep = instance->li_file; *linep = instance->li_line; } void witness_restore(struct lock_object *lock, const char *file, int line) { struct lock_list_entry *lock_list; struct lock_instance *instance; struct lock_class *class; /* * This function is used independently in locking code to deal with * Giant, SCHEDULER_STOPPED() check can be removed here after Giant * is gone. */ if (SCHEDULER_STOPPED()) return; KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) return; class = LOCK_CLASS(lock); if (class->lc_flags & LC_SLEEPLOCK) lock_list = curthread->td_sleeplocks; else { if (witness_skipspin) return; lock_list = PCPU_GET(spinlocks); } instance = find_instance(lock_list, lock); if (instance == NULL) kassert_panic("%s: lock (%s) %s not locked", __func__, class->lc_name, lock->lo_name); lock->lo_witness->w_file = file; lock->lo_witness->w_line = line; if (instance == NULL) return; instance->li_file = file; instance->li_line = line; } void witness_assert(const struct lock_object *lock, int flags, const char *file, int line) { #ifdef INVARIANT_SUPPORT struct lock_instance *instance; struct lock_class *class; if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL) return; class = LOCK_CLASS(lock); if ((class->lc_flags & LC_SLEEPLOCK) != 0) instance = find_instance(curthread->td_sleeplocks, lock); else if ((class->lc_flags & LC_SPINLOCK) != 0) instance = find_instance(PCPU_GET(spinlocks), lock); else { kassert_panic("Lock (%s) %s is not sleep or spin!", class->lc_name, lock->lo_name); return; } switch (flags) { case LA_UNLOCKED: if (instance != NULL) kassert_panic("Lock (%s) %s locked @ %s:%d.", class->lc_name, lock->lo_name, fixup_filename(file), line); break; case LA_LOCKED: case LA_LOCKED | LA_RECURSED: case LA_LOCKED | LA_NOTRECURSED: case LA_SLOCKED: case LA_SLOCKED | LA_RECURSED: case LA_SLOCKED | LA_NOTRECURSED: case LA_XLOCKED: case LA_XLOCKED | LA_RECURSED: case LA_XLOCKED | LA_NOTRECURSED: if (instance == NULL) { kassert_panic("Lock (%s) %s not locked @ %s:%d.", class->lc_name, lock->lo_name, fixup_filename(file), line); break; } if ((flags & LA_XLOCKED) != 0 && (instance->li_flags & LI_EXCLUSIVE) == 0) kassert_panic( "Lock (%s) %s not exclusively locked @ %s:%d.", class->lc_name, lock->lo_name, fixup_filename(file), line); if ((flags & LA_SLOCKED) != 0 && (instance->li_flags & LI_EXCLUSIVE) != 0) kassert_panic( "Lock (%s) %s exclusively locked @ %s:%d.", class->lc_name, lock->lo_name, fixup_filename(file), line); if ((flags & LA_RECURSED) != 0 && (instance->li_flags & LI_RECURSEMASK) == 0) kassert_panic("Lock (%s) %s not recursed @ %s:%d.", class->lc_name, lock->lo_name, fixup_filename(file), line); if ((flags & LA_NOTRECURSED) != 0 && (instance->li_flags & LI_RECURSEMASK) != 0) kassert_panic("Lock (%s) %s recursed @ %s:%d.", class->lc_name, lock->lo_name, fixup_filename(file), line); break; default: kassert_panic("Invalid lock assertion at %s:%d.", fixup_filename(file), line); } #endif /* INVARIANT_SUPPORT */ } static void witness_setflag(struct lock_object *lock, int flag, int set) { struct lock_list_entry *lock_list; struct lock_instance *instance; struct lock_class *class; if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) return; class = LOCK_CLASS(lock); if (class->lc_flags & LC_SLEEPLOCK) lock_list = curthread->td_sleeplocks; else { if (witness_skipspin) return; lock_list = PCPU_GET(spinlocks); } instance = find_instance(lock_list, lock); if (instance == NULL) { kassert_panic("%s: lock (%s) %s not locked", __func__, class->lc_name, lock->lo_name); return; } if (set) instance->li_flags |= flag; else instance->li_flags &= ~flag; } void witness_norelease(struct lock_object *lock) { witness_setflag(lock, LI_NORELEASE, 1); } void witness_releaseok(struct lock_object *lock) { witness_setflag(lock, LI_NORELEASE, 0); } #ifdef DDB static void witness_ddb_list(struct thread *td) { KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); KASSERT(kdb_active, ("%s: not in the debugger", __func__)); if (witness_watch < 1) return; witness_list_locks(&td->td_sleeplocks, db_printf); /* * We only handle spinlocks if td == curthread. This is somewhat broken * if td is currently executing on some other CPU and holds spin locks * as we won't display those locks. If we had a MI way of getting * the per-cpu data for a given cpu then we could use * td->td_oncpu to get the list of spinlocks for this thread * and "fix" this. * * That still wouldn't really fix this unless we locked the scheduler * lock or stopped the other CPU to make sure it wasn't changing the * list out from under us. It is probably best to just not try to * handle threads on other CPU's for now. */ if (td == curthread && PCPU_GET(spinlocks) != NULL) witness_list_locks(PCPU_PTR(spinlocks), db_printf); } DB_SHOW_COMMAND(locks, db_witness_list) { struct thread *td; if (have_addr) td = db_lookup_thread(addr, true); else td = kdb_thread; witness_ddb_list(td); } DB_SHOW_ALL_COMMAND(locks, db_witness_list_all) { struct thread *td; struct proc *p; /* * It would be nice to list only threads and processes that actually * held sleep locks, but that information is currently not exported * by WITNESS. */ FOREACH_PROC_IN_SYSTEM(p) { if (!witness_proc_has_locks(p)) continue; FOREACH_THREAD_IN_PROC(p, td) { if (!witness_thread_has_locks(td)) continue; db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid, p->p_comm, td, td->td_tid); witness_ddb_list(td); if (db_pager_quit) return; } } } DB_SHOW_ALIAS(alllocks, db_witness_list_all) DB_SHOW_COMMAND(witness, db_witness_display) { witness_ddb_display(db_printf); } #endif static void sbuf_print_witness_badstacks(struct sbuf *sb, size_t *oldidx) { struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2; struct witness *tmp_w1, *tmp_w2, *w1, *w2; int generation, i, j; tmp_data1 = NULL; tmp_data2 = NULL; tmp_w1 = NULL; tmp_w2 = NULL; /* Allocate and init temporary storage space. */ tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO); tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO); tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP, M_WAITOK | M_ZERO); tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP, M_WAITOK | M_ZERO); stack_zero(&tmp_data1->wlod_stack); stack_zero(&tmp_data2->wlod_stack); restart: mtx_lock_spin(&w_mtx); generation = w_generation; mtx_unlock_spin(&w_mtx); sbuf_printf(sb, "Number of known direct relationships is %d\n", w_lohash.wloh_count); for (i = 1; i < w_max_used_index; i++) { mtx_lock_spin(&w_mtx); if (generation != w_generation) { mtx_unlock_spin(&w_mtx); /* The graph has changed, try again. */ *oldidx = 0; sbuf_clear(sb); goto restart; } w1 = &w_data[i]; if (w1->w_reversed == 0) { mtx_unlock_spin(&w_mtx); continue; } /* Copy w1 locally so we can release the spin lock. */ *tmp_w1 = *w1; mtx_unlock_spin(&w_mtx); if (tmp_w1->w_reversed == 0) continue; for (j = 1; j < w_max_used_index; j++) { if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j) continue; mtx_lock_spin(&w_mtx); if (generation != w_generation) { mtx_unlock_spin(&w_mtx); /* The graph has changed, try again. */ *oldidx = 0; sbuf_clear(sb); goto restart; } w2 = &w_data[j]; data1 = witness_lock_order_get(w1, w2); data2 = witness_lock_order_get(w2, w1); /* * Copy information locally so we can release the * spin lock. */ *tmp_w2 = *w2; if (data1) { stack_zero(&tmp_data1->wlod_stack); stack_copy(&data1->wlod_stack, &tmp_data1->wlod_stack); } if (data2 && data2 != data1) { stack_zero(&tmp_data2->wlod_stack); stack_copy(&data2->wlod_stack, &tmp_data2->wlod_stack); } mtx_unlock_spin(&w_mtx); sbuf_printf(sb, "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n", tmp_w1->w_name, tmp_w1->w_class->lc_name, tmp_w2->w_name, tmp_w2->w_class->lc_name); if (data1) { sbuf_printf(sb, "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n", tmp_w1->w_name, tmp_w1->w_class->lc_name, tmp_w2->w_name, tmp_w2->w_class->lc_name); stack_sbuf_print(sb, &tmp_data1->wlod_stack); sbuf_printf(sb, "\n"); } if (data2 && data2 != data1) { sbuf_printf(sb, "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n", tmp_w2->w_name, tmp_w2->w_class->lc_name, tmp_w1->w_name, tmp_w1->w_class->lc_name); stack_sbuf_print(sb, &tmp_data2->wlod_stack); sbuf_printf(sb, "\n"); } } } mtx_lock_spin(&w_mtx); if (generation != w_generation) { mtx_unlock_spin(&w_mtx); /* * The graph changed while we were printing stack data, * try again. */ *oldidx = 0; sbuf_clear(sb); goto restart; } mtx_unlock_spin(&w_mtx); /* Free temporary storage space. */ free(tmp_data1, M_TEMP); free(tmp_data2, M_TEMP); free(tmp_w1, M_TEMP); free(tmp_w2, M_TEMP); } static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS) { struct sbuf *sb; int error; if (witness_watch < 1) { error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning)); return (error); } if (witness_cold) { error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold)); return (error); } error = 0; sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND); if (sb == NULL) return (ENOMEM); sbuf_print_witness_badstacks(sb, &req->oldidx); sbuf_finish(sb); error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); sbuf_delete(sb); return (error); } #ifdef DDB static int sbuf_db_printf_drain(void *arg __unused, const char *data, int len) { return (db_printf("%.*s", len, data)); } DB_SHOW_COMMAND(badstacks, db_witness_badstacks) { struct sbuf sb; char buffer[128]; size_t dummy; sbuf_new(&sb, buffer, sizeof(buffer), SBUF_FIXEDLEN); sbuf_set_drain(&sb, sbuf_db_printf_drain, NULL); sbuf_print_witness_badstacks(&sb, &dummy); sbuf_finish(&sb); } #endif static int sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS) { static const struct { enum witness_channel channel; const char *name; } channels[] = { { WITNESS_CONSOLE, "console" }, { WITNESS_LOG, "log" }, { WITNESS_NONE, "none" }, }; char buf[16]; u_int i; int error; buf[0] = '\0'; for (i = 0; i < nitems(channels); i++) if (witness_channel == channels[i].channel) { snprintf(buf, sizeof(buf), "%s", channels[i].name); break; } error = sysctl_handle_string(oidp, buf, sizeof(buf), req); if (error != 0 || req->newptr == NULL) return (error); error = EINVAL; for (i = 0; i < nitems(channels); i++) if (strcmp(channels[i].name, buf) == 0) { witness_channel = channels[i].channel; error = 0; break; } return (error); } static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS) { struct witness *w; struct sbuf *sb; int error; #ifdef __i386__ error = SYSCTL_OUT(req, w_notallowed, sizeof(w_notallowed)); return (error); #endif if (witness_watch < 1) { error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning)); return (error); } if (witness_cold) { error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold)); return (error); } error = 0; error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req); if (sb == NULL) return (ENOMEM); sbuf_printf(sb, "\n"); mtx_lock_spin(&w_mtx); STAILQ_FOREACH(w, &w_all, w_list) w->w_displayed = 0; STAILQ_FOREACH(w, &w_all, w_list) witness_add_fullgraph(sb, w); mtx_unlock_spin(&w_mtx); /* * Close the sbuf and return to userland. */ error = sbuf_finish(sb); sbuf_delete(sb); return (error); } static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS) { int error, value; value = witness_watch; error = sysctl_handle_int(oidp, &value, 0, req); if (error != 0 || req->newptr == NULL) return (error); if (value > 1 || value < -1 || (witness_watch == -1 && value != witness_watch)) return (EINVAL); witness_watch = value; return (0); } static void witness_add_fullgraph(struct sbuf *sb, struct witness *w) { int i; if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0)) return; w->w_displayed = 1; WITNESS_INDEX_ASSERT(w->w_index); for (i = 1; i <= w_max_used_index; i++) { if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) { sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name, w_data[i].w_name); witness_add_fullgraph(sb, &w_data[i]); } } } /* * A simple hash function. Takes a key pointer and a key size. If size == 0, * interprets the key as a string and reads until the null * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit * hash value computed from the key. */ static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size) { unsigned int hash = 5381; int i; /* hash = hash * 33 + key[i] */ if (size) for (i = 0; i < size; i++) hash = ((hash << 5) + hash) + (unsigned int)key[i]; else for (i = 0; key[i] != 0; i++) hash = ((hash << 5) + hash) + (unsigned int)key[i]; return (hash); } /* * Initializes the two witness hash tables. Called exactly once from * witness_initialize(). */ static void witness_init_hash_tables(void) { int i; MPASS(witness_cold); /* Initialize the hash tables. */ for (i = 0; i < WITNESS_HASH_SIZE; i++) w_hash.wh_array[i] = NULL; w_hash.wh_size = WITNESS_HASH_SIZE; w_hash.wh_count = 0; /* Initialize the lock order data hash. */ w_lofree = NULL; for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) { memset(&w_lodata[i], 0, sizeof(w_lodata[i])); w_lodata[i].wlod_next = w_lofree; w_lofree = &w_lodata[i]; } w_lohash.wloh_size = WITNESS_LO_HASH_SIZE; w_lohash.wloh_count = 0; for (i = 0; i < WITNESS_LO_HASH_SIZE; i++) w_lohash.wloh_array[i] = NULL; } static struct witness * witness_hash_get(const char *key) { struct witness *w; uint32_t hash; MPASS(key != NULL); if (witness_cold == 0) mtx_assert(&w_mtx, MA_OWNED); hash = witness_hash_djb2(key, 0) % w_hash.wh_size; w = w_hash.wh_array[hash]; while (w != NULL) { if (strcmp(w->w_name, key) == 0) goto out; w = w->w_hash_next; } out: return (w); } static void witness_hash_put(struct witness *w) { uint32_t hash; MPASS(w != NULL); MPASS(w->w_name != NULL); if (witness_cold == 0) mtx_assert(&w_mtx, MA_OWNED); KASSERT(witness_hash_get(w->w_name) == NULL, ("%s: trying to add a hash entry that already exists!", __func__)); KASSERT(w->w_hash_next == NULL, ("%s: w->w_hash_next != NULL", __func__)); hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size; w->w_hash_next = w_hash.wh_array[hash]; w_hash.wh_array[hash] = w; w_hash.wh_count++; } static struct witness_lock_order_data * witness_lock_order_get(struct witness *parent, struct witness *child) { struct witness_lock_order_data *data = NULL; struct witness_lock_order_key key; unsigned int hash; MPASS(parent != NULL && child != NULL); key.from = parent->w_index; key.to = child->w_index; WITNESS_INDEX_ASSERT(key.from); WITNESS_INDEX_ASSERT(key.to); if ((w_rmatrix[parent->w_index][child->w_index] & WITNESS_LOCK_ORDER_KNOWN) == 0) goto out; hash = witness_hash_djb2((const char*)&key, sizeof(key)) % w_lohash.wloh_size; data = w_lohash.wloh_array[hash]; while (data != NULL) { if (witness_lock_order_key_equal(&data->wlod_key, &key)) break; data = data->wlod_next; } out: return (data); } /* * Verify that parent and child have a known relationship, are not the same, * and child is actually a child of parent. This is done without w_mtx * to avoid contention in the common case. */ static int witness_lock_order_check(struct witness *parent, struct witness *child) { if (parent != child && w_rmatrix[parent->w_index][child->w_index] & WITNESS_LOCK_ORDER_KNOWN && isitmychild(parent, child)) return (1); return (0); } static int witness_lock_order_add(struct witness *parent, struct witness *child) { struct witness_lock_order_data *data = NULL; struct witness_lock_order_key key; unsigned int hash; MPASS(parent != NULL && child != NULL); key.from = parent->w_index; key.to = child->w_index; WITNESS_INDEX_ASSERT(key.from); WITNESS_INDEX_ASSERT(key.to); if (w_rmatrix[parent->w_index][child->w_index] & WITNESS_LOCK_ORDER_KNOWN) return (1); hash = witness_hash_djb2((const char*)&key, sizeof(key)) % w_lohash.wloh_size; w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN; data = w_lofree; if (data == NULL) return (0); w_lofree = data->wlod_next; data->wlod_next = w_lohash.wloh_array[hash]; data->wlod_key = key; w_lohash.wloh_array[hash] = data; w_lohash.wloh_count++; stack_zero(&data->wlod_stack); stack_save(&data->wlod_stack); return (1); } /* Call this whenever the structure of the witness graph changes. */ static void witness_increment_graph_generation(void) { if (witness_cold == 0) mtx_assert(&w_mtx, MA_OWNED); w_generation++; } static int witness_output_drain(void *arg __unused, const char *data, int len) { witness_output("%.*s", len, data); return (len); } static void witness_debugger(int cond, const char *msg) { char buf[32]; struct sbuf sb; struct stack st; if (!cond) return; if (witness_trace) { sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN); sbuf_set_drain(&sb, witness_output_drain, NULL); stack_zero(&st); stack_save(&st); witness_output("stack backtrace:\n"); stack_sbuf_print_ddb(&sb, &st); sbuf_finish(&sb); } #ifdef KDB if (witness_kdb) kdb_enter(KDB_WHY_WITNESS, msg); #endif } Index: head/sys/netinet/in_pcb.h =================================================================== --- head/sys/netinet/in_pcb.h (revision 335355) +++ head/sys/netinet/in_pcb.h (revision 335356) @@ -1,892 +1,891 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1990, 1993 * The Regents of the University of California. * Copyright (c) 2010-2011 Juniper Networks, Inc. * All rights reserved. * * Portions of this software were developed by Robert N. M. Watson under * contract to Juniper Networks, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)in_pcb.h 8.1 (Berkeley) 6/10/93 * $FreeBSD$ */ #ifndef _NETINET_IN_PCB_H_ #define _NETINET_IN_PCB_H_ #include #include #include #include #include #include #ifdef _KERNEL #include #include #include +#include +#include #include #endif #include #define in6pcb inpcb /* for KAME src sync over BSD*'s */ #define in6p_sp inp_sp /* for KAME src sync over BSD*'s */ /* * struct inpcb is the common protocol control block structure used in most * IP transport protocols. * * Pointers to local and foreign host table entries, local and foreign socket * numbers, and pointers up (to a socket structure) and down (to a * protocol-specific control block) are stored here. */ CK_LIST_HEAD(inpcbhead, inpcb); CK_LIST_HEAD(inpcbporthead, inpcbport); typedef uint64_t inp_gen_t; /* * PCB with AF_INET6 null bind'ed laddr can receive AF_INET input packet. * So, AF_INET6 null laddr is also used as AF_INET null laddr, by utilizing * the following structure. */ struct in_addr_4in6 { u_int32_t ia46_pad32[3]; struct in_addr ia46_addr4; }; union in_dependaddr { struct in_addr_4in6 id46_addr; struct in6_addr id6_addr; }; /* * NOTE: ipv6 addrs should be 64-bit aligned, per RFC 2553. in_conninfo has * some extra padding to accomplish this. * NOTE 2: tcp_syncache.c uses first 5 32-bit words, which identify fport, * lport, faddr to generate hash, so these fields shouldn't be moved. */ struct in_endpoints { u_int16_t ie_fport; /* foreign port */ u_int16_t ie_lport; /* local port */ /* protocol dependent part, local and foreign addr */ union in_dependaddr ie_dependfaddr; /* foreign host table entry */ union in_dependaddr ie_dependladdr; /* local host table entry */ #define ie_faddr ie_dependfaddr.id46_addr.ia46_addr4 #define ie_laddr ie_dependladdr.id46_addr.ia46_addr4 #define ie6_faddr ie_dependfaddr.id6_addr #define ie6_laddr ie_dependladdr.id6_addr u_int32_t ie6_zoneid; /* scope zone id */ }; /* * XXX The defines for inc_* are hacks and should be changed to direct * references. */ struct in_conninfo { u_int8_t inc_flags; u_int8_t inc_len; u_int16_t inc_fibnum; /* XXX was pad, 16 bits is plenty */ /* protocol dependent part */ struct in_endpoints inc_ie; }; /* * Flags for inc_flags. */ #define INC_ISIPV6 0x01 #define inc_isipv6 inc_flags /* temp compatibility */ #define inc_fport inc_ie.ie_fport #define inc_lport inc_ie.ie_lport #define inc_faddr inc_ie.ie_faddr #define inc_laddr inc_ie.ie_laddr #define inc6_faddr inc_ie.ie6_faddr #define inc6_laddr inc_ie.ie6_laddr #define inc6_zoneid inc_ie.ie6_zoneid #if defined(_KERNEL) || defined(_WANT_INPCB) /* * struct inpcb captures the network layer state for TCP, UDP, and raw IPv4 and * IPv6 sockets. In the case of TCP and UDP, further per-connection state is * hung off of inp_ppcb most of the time. Almost all fields of struct inpcb * are static after creation or protected by a per-inpcb rwlock, inp_lock. A * few fields are protected by multiple locks as indicated in the locking notes * below. For these fields, all of the listed locks must be write-locked for * any modifications. However, these fields can be safely read while any one of * the listed locks are read-locked. This model can permit greater concurrency * for read operations. For example, connections can be looked up while only * holding a read lock on the global pcblist lock. This is important for * performance when attempting to find the connection for a packet given its IP * and port tuple. * * One noteworthy exception is that the global pcbinfo lock follows a different * set of rules in relation to the inp_list field. Rather than being * write-locked for modifications and read-locked for list iterations, it must * be read-locked during modifications and write-locked during list iterations. * This ensures that the relatively rare global list iterations safely walk a * stable snapshot of connections while allowing more common list modifications * to safely grab the pcblist lock just while adding or removing a connection * from the global list. * * Key: * (b) - Protected by the hpts lock. * (c) - Constant after initialization + * (e) - Protected by the net_epoch_prempt epoch * (g) - Protected by the pcbgroup lock * (i) - Protected by the inpcb lock * (p) - Protected by the pcbinfo lock for the inpcb * (l) - Protected by the pcblist lock for the inpcb * (h) - Protected by the pcbhash lock for the inpcb * (s) - Protected by another subsystem's locks * (x) - Undefined locking * * Notes on the tcp_hpts: * * First Hpts lock order is * 1) INP_WLOCK() * 2) HPTS_LOCK() i.e. hpts->pmtx * * To insert a TCB on the hpts you *must* be holding the INP_WLOCK(). * You may check the inp->inp_in_hpts flag without the hpts lock. * The hpts is the only one that will clear this flag holding * only the hpts lock. This means that in your tcp_output() * routine when you test for the inp_in_hpts flag to be 1 * it may be transitioning to 0 (by the hpts). * That's ok since that will just mean an extra call to tcp_output * that most likely will find the call you executed * (when the mis-match occured) will have put the TCB back * on the hpts and it will return. If your * call did not add the inp back to the hpts then you will either * over-send or the cwnd will block you from sending more. * * Note you should also be holding the INP_WLOCK() when you * call the remove from the hpts as well. Though usually * you are either doing this from a timer, where you need and have * the INP_WLOCK() or from destroying your TCB where again * you should already have the INP_WLOCK(). * * The inp_hpts_cpu, inp_hpts_cpu_set, inp_input_cpu and * inp_input_cpu_set fields are controlled completely by * the hpts. Do not ever set these. The inp_hpts_cpu_set * and inp_input_cpu_set fields indicate if the hpts has * setup the respective cpu field. It is advised if this * field is 0, to enqueue the packet with the appropriate * hpts_immediate() call. If the _set field is 1, then * you may compare the inp_*_cpu field to the curcpu and * may want to again insert onto the hpts if these fields * are not equal (i.e. you are not on the expected CPU). * * A note on inp_hpts_calls and inp_input_calls, these * flags are set when the hpts calls either the output * or do_segment routines respectively. If the routine * being called wants to use this, then it needs to * clear the flag before returning. The hpts will not * clear the flag. The flags can be used to tell if * the hpts is the function calling the respective * routine. * * A few other notes: * * When a read lock is held, stability of the field is guaranteed; to write * to a field, a write lock must generally be held. * * netinet/netinet6-layer code should not assume that the inp_socket pointer * is safe to dereference without inp_lock being held, even for protocols * other than TCP (where the inpcb persists during TIMEWAIT even after the * socket has been freed), or there may be close(2)-related races. * * The inp_vflag field is overloaded, and would otherwise ideally be (c). * * TODO: Currently only the TCP stack is leveraging the global pcbinfo lock * read-lock usage during modification, this model can be applied to other * protocols (especially SCTP). */ struct icmp6_filter; struct inpcbpolicy; struct m_snd_tag; struct inpcb { /* Cache line #1 (amd64) */ - CK_LIST_ENTRY(inpcb) inp_hash; /* (h/i) hash list */ + CK_LIST_ENTRY(inpcb) inp_hash; /* [w](h/i) [r](e/i) hash list */ CK_LIST_ENTRY(inpcb) inp_pcbgrouphash; /* (g/i) hash list */ struct rwlock inp_lock; /* Cache line #2 (amd64) */ #define inp_start_zero inp_hpts #define inp_zero_size (sizeof(struct inpcb) - \ offsetof(struct inpcb, inp_start_zero)) TAILQ_ENTRY(inpcb) inp_hpts; /* pacing out queue next lock(b) */ uint32_t inp_hpts_request; /* Current hpts request, zero if * fits in the pacing window (i&b). */ /* * Note the next fields are protected by a * different lock (hpts-lock). This means that * they must correspond in size to the smallest * protectable bit field (uint8_t on x86, and * other platfomrs potentially uint32_t?). Also * since CPU switches can occur at different times the two * fields can *not* be collapsed into a signal bit field. */ #if defined(__amd64__) || defined(__i386__) volatile uint8_t inp_in_hpts; /* on output hpts (lock b) */ volatile uint8_t inp_in_input; /* on input hpts (lock b) */ #else volatile uint32_t inp_in_hpts; /* on output hpts (lock b) */ volatile uint32_t inp_in_input; /* on input hpts (lock b) */ #endif volatile uint16_t inp_hpts_cpu; /* Lock (i) */ u_int inp_refcount; /* (i) refcount */ int inp_flags; /* (i) generic IP/datagram flags */ int inp_flags2; /* (i) generic IP/datagram flags #2*/ volatile uint16_t inp_input_cpu; /* Lock (i) */ volatile uint8_t inp_hpts_cpu_set :1, /* on output hpts (i) */ inp_input_cpu_set : 1, /* on input hpts (i) */ inp_hpts_calls :1, /* (i) from output hpts */ inp_input_calls :1, /* (i) from input hpts */ inp_spare_bits2 : 4; uint8_t inp_spare_byte; /* Compiler hole */ void *inp_ppcb; /* (i) pointer to per-protocol pcb */ struct socket *inp_socket; /* (i) back pointer to socket */ uint32_t inp_hptsslot; /* Hpts wheel slot this tcb is Lock(i&b) */ uint32_t inp_hpts_drop_reas; /* reason we are dropping the PCB (lock i&b) */ TAILQ_ENTRY(inpcb) inp_input; /* pacing in queue next lock(b) */ struct inpcbinfo *inp_pcbinfo; /* (c) PCB list info */ struct inpcbgroup *inp_pcbgroup; /* (g/i) PCB group list */ CK_LIST_ENTRY(inpcb) inp_pcbgroup_wild; /* (g/i/h) group wildcard entry */ struct ucred *inp_cred; /* (c) cache of socket cred */ u_int32_t inp_flow; /* (i) IPv6 flow information */ u_char inp_vflag; /* (i) IP version flag (v4/v6) */ u_char inp_ip_ttl; /* (i) time to live proto */ u_char inp_ip_p; /* (c) protocol proto */ u_char inp_ip_minttl; /* (i) minimum TTL or drop */ uint32_t inp_flowid; /* (x) flow id / queue id */ struct m_snd_tag *inp_snd_tag; /* (i) send tag for outgoing mbufs */ uint32_t inp_flowtype; /* (x) M_HASHTYPE value */ uint32_t inp_rss_listen_bucket; /* (x) overridden RSS listen bucket */ /* Local and foreign ports, local and foreign addr. */ struct in_conninfo inp_inc; /* (i) list for PCB's local port */ /* MAC and IPSEC policy information. */ struct label *inp_label; /* (i) MAC label */ struct inpcbpolicy *inp_sp; /* (s) for IPSEC */ /* Protocol-dependent part; options. */ struct { u_char inp_ip_tos; /* (i) type of service proto */ struct mbuf *inp_options; /* (i) IP options */ struct ip_moptions *inp_moptions; /* (i) mcast options */ }; struct { /* (i) IP options */ struct mbuf *in6p_options; /* (i) IP6 options for outgoing packets */ struct ip6_pktopts *in6p_outputopts; /* (i) IP multicast options */ struct ip6_moptions *in6p_moptions; /* (i) ICMPv6 code type filter */ struct icmp6_filter *in6p_icmp6filt; /* (i) IPV6_CHECKSUM setsockopt */ int in6p_cksum; short in6p_hops; }; CK_LIST_ENTRY(inpcb) inp_portlist; /* (i/h) */ struct inpcbport *inp_phd; /* (i/h) head of this list */ inp_gen_t inp_gencnt; /* (c) generation count */ struct llentry *inp_lle; /* cached L2 information */ rt_gen_t inp_rt_cookie; /* generation for route entry */ union { /* cached L3 information */ struct route inp_route; struct route_in6 inp_route6; }; CK_LIST_ENTRY(inpcb) inp_list; /* (p/l) list for all PCBs for proto */ - /* (p[w]) for list iteration */ - /* (p[r]/l) for addition/removal */ + /* (e[r]) for list iteration */ + /* (p[w]/l) for addition/removal */ struct epoch_context inp_epoch_ctx; }; #endif /* _KERNEL */ #define inp_fport inp_inc.inc_fport #define inp_lport inp_inc.inc_lport #define inp_faddr inp_inc.inc_faddr #define inp_laddr inp_inc.inc_laddr #define in6p_faddr inp_inc.inc6_faddr #define in6p_laddr inp_inc.inc6_laddr #define in6p_zoneid inp_inc.inc6_zoneid #define in6p_flowinfo inp_flow #define inp_vnet inp_pcbinfo->ipi_vnet /* * The range of the generation count, as used in this implementation, is 9e19. * We would have to create 300 billion connections per second for this number * to roll over in a year. This seems sufficiently unlikely that we simply * don't concern ourselves with that possibility. */ /* * Interface exported to userland by various protocols which use inpcbs. Hack * alert -- only define if struct xsocket is in scope. * Fields prefixed with "xi_" are unique to this structure, and the rest * match fields in the struct inpcb, to ease coding and porting. * * Legend: * (s) - used by userland utilities in src * (p) - used by utilities in ports * (3) - is known to be used by third party software not in ports * (n) - no known usage */ #ifdef _SYS_SOCKETVAR_H_ struct xinpcb { size_t xi_len; /* length of this structure */ struct xsocket xi_socket; /* (s,p) */ struct in_conninfo inp_inc; /* (s,p) */ uint64_t inp_gencnt; /* (s,p) */ union { void *inp_ppcb; /* (s) netstat(1) */ int64_t ph_ppcb; }; int64_t inp_spare64[4]; uint32_t inp_flow; /* (s) */ uint32_t inp_flowid; /* (s) */ uint32_t inp_flowtype; /* (s) */ int32_t inp_flags; /* (s,p) */ int32_t inp_flags2; /* (s) */ int32_t inp_rss_listen_bucket; /* (n) */ int32_t in6p_cksum; /* (n) */ int32_t inp_spare32[4]; uint16_t in6p_hops; /* (n) */ uint8_t inp_ip_tos; /* (n) */ int8_t pad8; uint8_t inp_vflag; /* (s,p) */ uint8_t inp_ip_ttl; /* (n) */ uint8_t inp_ip_p; /* (n) */ uint8_t inp_ip_minttl; /* (n) */ int8_t inp_spare8[4]; } __aligned(8); struct xinpgen { size_t xig_len; /* length of this structure */ u_int xig_count; /* number of PCBs at this time */ inp_gen_t xig_gen; /* generation count at this time */ so_gen_t xig_sogen; /* socket generation count this time */ } __aligned(8); #ifdef _KERNEL void in_pcbtoxinpcb(const struct inpcb *, struct xinpcb *); #endif #endif /* _SYS_SOCKETVAR_H_ */ struct inpcbport { struct epoch_context phd_epoch_ctx; CK_LIST_ENTRY(inpcbport) phd_hash; struct inpcbhead phd_pcblist; u_short phd_port; }; struct in_pcblist { int il_count; struct epoch_context il_epoch_ctx; struct inpcbinfo *il_pcbinfo; struct inpcb *il_inp_list[0]; }; /*- * Global data structure for each high-level protocol (UDP, TCP, ...) in both * IPv4 and IPv6. Holds inpcb lists and information for managing them. * * Each pcbinfo is protected by three locks: ipi_lock, ipi_hash_lock and * ipi_list_lock: * - ipi_lock covering the global pcb list stability during loop iteration, * - ipi_hash_lock covering the hashed lookup tables, * - ipi_list_lock covering mutable global fields (such as the global * pcb list) * * The lock order is: * * ipi_lock (before) * inpcb locks (before) * ipi_list locks (before) * {ipi_hash_lock, pcbgroup locks} * * Locking key: * * (c) Constant or nearly constant after initialisation + * (e) - Protected by the net_epoch_prempt epoch * (g) Locked by ipi_lock * (l) Locked by ipi_list_lock - * (h) Read using either ipi_hash_lock or inpcb lock; write requires both + * (h) Read using either net_epoch_preempt or inpcb lock; write requires both ipi_hash_lock and inpcb lock * (p) Protected by one or more pcbgroup locks * (x) Synchronisation properties poorly defined */ struct inpcbinfo { /* - * Global lock protecting full inpcb list traversal + * Global lock protecting inpcb list modification */ - struct rwlock ipi_lock; + struct mtx ipi_lock; /* * Global list of inpcbs on the protocol. */ - struct inpcbhead *ipi_listhead; /* (g/l) */ + struct inpcbhead *ipi_listhead; /* [r](e) [w](g/l) */ u_int ipi_count; /* (l) */ /* * Generation count -- incremented each time a connection is allocated * or freed. */ u_quad_t ipi_gencnt; /* (l) */ /* * Fields associated with port lookup and allocation. */ u_short ipi_lastport; /* (x) */ u_short ipi_lastlow; /* (x) */ u_short ipi_lasthi; /* (x) */ /* * UMA zone from which inpcbs are allocated for this protocol. */ struct uma_zone *ipi_zone; /* (c) */ /* * Connection groups associated with this protocol. These fields are * constant, but pcbgroup structures themselves are protected by * per-pcbgroup locks. */ struct inpcbgroup *ipi_pcbgroups; /* (c) */ u_int ipi_npcbgroups; /* (c) */ u_int ipi_hashfields; /* (c) */ /* - * Global lock protecting non-pcbgroup hash lookup tables. + * Global lock protecting modification non-pcbgroup hash lookup tables. */ - struct rwlock ipi_hash_lock; + struct mtx ipi_hash_lock; /* * Global hash of inpcbs, hashed by local and foreign addresses and * port numbers. */ struct inpcbhead *ipi_hashbase; /* (h) */ u_long ipi_hashmask; /* (h) */ /* * Global hash of inpcbs, hashed by only local port number. */ struct inpcbporthead *ipi_porthashbase; /* (h) */ u_long ipi_porthashmask; /* (h) */ /* * List of wildcard inpcbs for use with pcbgroups. In the past, was * per-pcbgroup but is now global. All pcbgroup locks must be held * to modify the list, so any is sufficient to read it. */ struct inpcbhead *ipi_wildbase; /* (p) */ u_long ipi_wildmask; /* (p) */ /* * Load balance groups used for the SO_REUSEPORT_LB option, * hashed by local port. */ struct inpcblbgrouphead *ipi_lbgrouphashbase; /* (h) */ u_long ipi_lbgrouphashmask; /* (h) */ /* * Pointer to network stack instance */ struct vnet *ipi_vnet; /* (c) */ /* * general use 2 */ void *ipi_pspare[2]; /* * Global lock protecting global inpcb list, inpcb count, etc. */ struct rwlock ipi_list_lock; }; #ifdef _KERNEL /* * Connection groups hold sets of connections that have similar CPU/thread * affinity. Each connection belongs to exactly one connection group. */ struct inpcbgroup { /* * Per-connection group hash of inpcbs, hashed by local and foreign * addresses and port numbers. */ struct inpcbhead *ipg_hashbase; /* (c) */ u_long ipg_hashmask; /* (c) */ /* * Notional affinity of this pcbgroup. */ u_int ipg_cpu; /* (p) */ /* * Per-connection group lock, not to be confused with ipi_lock. * Protects the hash table hung off the group, but also the global * wildcard list in inpcbinfo. */ struct mtx ipg_lock; } __aligned(CACHE_LINE_SIZE); /* * Load balance groups used for the SO_REUSEPORT_LB socket option. Each group * (or unique address:port combination) can be re-used at most * INPCBLBGROUP_SIZMAX (256) times. The inpcbs are stored in il_inp which * is dynamically resized as processes bind/unbind to that specific group. */ struct inpcblbgroup { LIST_ENTRY(inpcblbgroup) il_list; uint16_t il_lport; /* (c) */ u_char il_vflag; /* (c) */ u_char il_pad; uint32_t il_pad2; union in_dependaddr il_dependladdr; /* (c) */ #define il_laddr il_dependladdr.id46_addr.ia46_addr4 #define il6_laddr il_dependladdr.id6_addr uint32_t il_inpsiz; /* max count in il_inp[] (h) */ uint32_t il_inpcnt; /* cur count in il_inp[] (h) */ struct inpcb *il_inp[]; /* (h) */ }; LIST_HEAD(inpcblbgrouphead, inpcblbgroup); #define INP_LOCK_INIT(inp, d, t) \ rw_init_flags(&(inp)->inp_lock, (t), RW_RECURSE | RW_DUPOK) #define INP_LOCK_DESTROY(inp) rw_destroy(&(inp)->inp_lock) #define INP_RLOCK(inp) rw_rlock(&(inp)->inp_lock) #define INP_WLOCK(inp) rw_wlock(&(inp)->inp_lock) #define INP_TRY_RLOCK(inp) rw_try_rlock(&(inp)->inp_lock) #define INP_TRY_WLOCK(inp) rw_try_wlock(&(inp)->inp_lock) #define INP_RUNLOCK(inp) rw_runlock(&(inp)->inp_lock) #define INP_WUNLOCK(inp) rw_wunlock(&(inp)->inp_lock) #define INP_TRY_UPGRADE(inp) rw_try_upgrade(&(inp)->inp_lock) #define INP_DOWNGRADE(inp) rw_downgrade(&(inp)->inp_lock) #define INP_WLOCKED(inp) rw_wowned(&(inp)->inp_lock) #define INP_LOCK_ASSERT(inp) rw_assert(&(inp)->inp_lock, RA_LOCKED) #define INP_RLOCK_ASSERT(inp) rw_assert(&(inp)->inp_lock, RA_RLOCKED) #define INP_WLOCK_ASSERT(inp) rw_assert(&(inp)->inp_lock, RA_WLOCKED) #define INP_UNLOCK_ASSERT(inp) rw_assert(&(inp)->inp_lock, RA_UNLOCKED) /* * These locking functions are for inpcb consumers outside of sys/netinet, * more specifically, they were added for the benefit of TOE drivers. The * macros are reserved for use by the stack. */ void inp_wlock(struct inpcb *); void inp_wunlock(struct inpcb *); void inp_rlock(struct inpcb *); void inp_runlock(struct inpcb *); #ifdef INVARIANT_SUPPORT void inp_lock_assert(struct inpcb *); void inp_unlock_assert(struct inpcb *); #else #define inp_lock_assert(inp) do {} while (0) #define inp_unlock_assert(inp) do {} while (0) #endif void inp_apply_all(void (*func)(struct inpcb *, void *), void *arg); int inp_ip_tos_get(const struct inpcb *inp); void inp_ip_tos_set(struct inpcb *inp, int val); struct socket * inp_inpcbtosocket(struct inpcb *inp); struct tcpcb * inp_inpcbtotcpcb(struct inpcb *inp); void inp_4tuple_get(struct inpcb *inp, uint32_t *laddr, uint16_t *lp, uint32_t *faddr, uint16_t *fp); int inp_so_options(const struct inpcb *inp); #endif /* _KERNEL */ #define INP_INFO_LOCK_INIT(ipi, d) \ - rw_init_flags(&(ipi)->ipi_lock, (d), RW_RECURSE) -#define INP_INFO_LOCK_DESTROY(ipi) rw_destroy(&(ipi)->ipi_lock) -#define INP_INFO_RLOCK(ipi) rw_rlock(&(ipi)->ipi_lock) -#define INP_INFO_WLOCK(ipi) rw_wlock(&(ipi)->ipi_lock) -#define INP_INFO_TRY_RLOCK(ipi) rw_try_rlock(&(ipi)->ipi_lock) -#define INP_INFO_TRY_WLOCK(ipi) rw_try_wlock(&(ipi)->ipi_lock) -#define INP_INFO_TRY_UPGRADE(ipi) rw_try_upgrade(&(ipi)->ipi_lock) -#define INP_INFO_WLOCKED(ipi) rw_wowned(&(ipi)->ipi_lock) -#define INP_INFO_RUNLOCK(ipi) rw_runlock(&(ipi)->ipi_lock) -#define INP_INFO_WUNLOCK(ipi) rw_wunlock(&(ipi)->ipi_lock) -#define INP_INFO_LOCK_ASSERT(ipi) rw_assert(&(ipi)->ipi_lock, RA_LOCKED) -#define INP_INFO_RLOCK_ASSERT(ipi) rw_assert(&(ipi)->ipi_lock, RA_RLOCKED) -#define INP_INFO_WLOCK_ASSERT(ipi) rw_assert(&(ipi)->ipi_lock, RA_WLOCKED) -#define INP_INFO_UNLOCK_ASSERT(ipi) rw_assert(&(ipi)->ipi_lock, RA_UNLOCKED) + mtx_init(&(ipi)->ipi_lock, (d), NULL, MTX_DEF| MTX_RECURSE) +#define INP_INFO_LOCK_DESTROY(ipi) mtx_destroy(&(ipi)->ipi_lock) +#define INP_INFO_RLOCK(ipi) NET_EPOCH_ENTER() +#define INP_INFO_WLOCK(ipi) mtx_lock(&(ipi)->ipi_lock) +#define INP_INFO_TRY_WLOCK(ipi) mtx_trylock(&(ipi)->ipi_lock) +#define INP_INFO_WLOCKED(ipi) mtx_owned(&(ipi)->ipi_lock) +#define INP_INFO_RUNLOCK(ipi) NET_EPOCH_EXIT() +#define INP_INFO_WUNLOCK(ipi) mtx_unlock(&(ipi)->ipi_lock) +#define INP_INFO_LOCK_ASSERT(ipi) MPASS(in_epoch() || mtx_owned(&(ipi)->ipi_lock)) +#define INP_INFO_RLOCK_ASSERT(ipi) MPASS(in_epoch()) +#define INP_INFO_WLOCK_ASSERT(ipi) mtx_assert(&(ipi)->ipi_lock, MA_OWNED) +#define INP_INFO_UNLOCK_ASSERT(ipi) MPASS(!in_epoch() && !mtx_owned(&(ipi)->ipi_lock)) #define INP_LIST_LOCK_INIT(ipi, d) \ rw_init_flags(&(ipi)->ipi_list_lock, (d), 0) #define INP_LIST_LOCK_DESTROY(ipi) rw_destroy(&(ipi)->ipi_list_lock) #define INP_LIST_RLOCK(ipi) rw_rlock(&(ipi)->ipi_list_lock) #define INP_LIST_WLOCK(ipi) rw_wlock(&(ipi)->ipi_list_lock) #define INP_LIST_TRY_RLOCK(ipi) rw_try_rlock(&(ipi)->ipi_list_lock) #define INP_LIST_TRY_WLOCK(ipi) rw_try_wlock(&(ipi)->ipi_list_lock) #define INP_LIST_TRY_UPGRADE(ipi) rw_try_upgrade(&(ipi)->ipi_list_lock) #define INP_LIST_RUNLOCK(ipi) rw_runlock(&(ipi)->ipi_list_lock) #define INP_LIST_WUNLOCK(ipi) rw_wunlock(&(ipi)->ipi_list_lock) #define INP_LIST_LOCK_ASSERT(ipi) \ rw_assert(&(ipi)->ipi_list_lock, RA_LOCKED) #define INP_LIST_RLOCK_ASSERT(ipi) \ rw_assert(&(ipi)->ipi_list_lock, RA_RLOCKED) #define INP_LIST_WLOCK_ASSERT(ipi) \ rw_assert(&(ipi)->ipi_list_lock, RA_WLOCKED) #define INP_LIST_UNLOCK_ASSERT(ipi) \ rw_assert(&(ipi)->ipi_list_lock, RA_UNLOCKED) -#define INP_HASH_LOCK_INIT(ipi, d) \ - rw_init_flags(&(ipi)->ipi_hash_lock, (d), 0) -#define INP_HASH_LOCK_DESTROY(ipi) rw_destroy(&(ipi)->ipi_hash_lock) -#define INP_HASH_RLOCK(ipi) rw_rlock(&(ipi)->ipi_hash_lock) -#define INP_HASH_WLOCK(ipi) rw_wlock(&(ipi)->ipi_hash_lock) -#define INP_HASH_RUNLOCK(ipi) rw_runlock(&(ipi)->ipi_hash_lock) -#define INP_HASH_WUNLOCK(ipi) rw_wunlock(&(ipi)->ipi_hash_lock) -#define INP_HASH_LOCK_ASSERT(ipi) rw_assert(&(ipi)->ipi_hash_lock, \ - RA_LOCKED) -#define INP_HASH_WLOCK_ASSERT(ipi) rw_assert(&(ipi)->ipi_hash_lock, \ - RA_WLOCKED) +#define INP_HASH_LOCK_INIT(ipi, d) mtx_init(&(ipi)->ipi_hash_lock, (d), NULL, MTX_DEF) +#define INP_HASH_LOCK_DESTROY(ipi) mtx_destroy(&(ipi)->ipi_hash_lock) +#define INP_HASH_RLOCK(ipi) NET_EPOCH_ENTER() +#define INP_HASH_WLOCK(ipi) mtx_lock(&(ipi)->ipi_hash_lock) +#define INP_HASH_RUNLOCK(ipi) NET_EPOCH_EXIT() +#define INP_HASH_WUNLOCK(ipi) mtx_unlock(&(ipi)->ipi_hash_lock) +#define INP_HASH_LOCK_ASSERT(ipi) MPASS(in_epoch() || mtx_owned(&(ipi)->ipi_hash_lock)) +#define INP_HASH_WLOCK_ASSERT(ipi) mtx_assert(&(ipi)->ipi_hash_lock, MA_OWNED); #define INP_GROUP_LOCK_INIT(ipg, d) mtx_init(&(ipg)->ipg_lock, (d), NULL, \ MTX_DEF | MTX_DUPOK) #define INP_GROUP_LOCK_DESTROY(ipg) mtx_destroy(&(ipg)->ipg_lock) #define INP_GROUP_LOCK(ipg) mtx_lock(&(ipg)->ipg_lock) #define INP_GROUP_LOCK_ASSERT(ipg) mtx_assert(&(ipg)->ipg_lock, MA_OWNED) #define INP_GROUP_UNLOCK(ipg) mtx_unlock(&(ipg)->ipg_lock) #define INP_PCBHASH(faddr, lport, fport, mask) \ (((faddr) ^ ((faddr) >> 16) ^ ntohs((lport) ^ (fport))) & (mask)) #define INP_PCBPORTHASH(lport, mask) \ (ntohs((lport)) & (mask)) #define INP_PCBLBGROUP_PORTHASH(lport, mask) \ (ntohs((lport)) & (mask)) #define INP_PCBLBGROUP_PKTHASH(faddr, lport, fport) \ ((faddr) ^ ((faddr) >> 16) ^ ntohs((lport) ^ (fport))) #define INP6_PCBHASHKEY(faddr) ((faddr)->s6_addr32[3]) /* * Flags for inp_vflags -- historically version flags only */ #define INP_IPV4 0x1 #define INP_IPV6 0x2 #define INP_IPV6PROTO 0x4 /* opened under IPv6 protocol */ /* * Flags for inp_flags. */ #define INP_RECVOPTS 0x00000001 /* receive incoming IP options */ #define INP_RECVRETOPTS 0x00000002 /* receive IP options for reply */ #define INP_RECVDSTADDR 0x00000004 /* receive IP dst address */ #define INP_HDRINCL 0x00000008 /* user supplies entire IP header */ #define INP_HIGHPORT 0x00000010 /* user wants "high" port binding */ #define INP_LOWPORT 0x00000020 /* user wants "low" port binding */ #define INP_ANONPORT 0x00000040 /* port chosen for user */ #define INP_RECVIF 0x00000080 /* receive incoming interface */ #define INP_MTUDISC 0x00000100 /* user can do MTU discovery */ /* 0x000200 unused: was INP_FAITH */ #define INP_RECVTTL 0x00000400 /* receive incoming IP TTL */ #define INP_DONTFRAG 0x00000800 /* don't fragment packet */ #define INP_BINDANY 0x00001000 /* allow bind to any address */ #define INP_INHASHLIST 0x00002000 /* in_pcbinshash() has been called */ #define INP_RECVTOS 0x00004000 /* receive incoming IP TOS */ #define IN6P_IPV6_V6ONLY 0x00008000 /* restrict AF_INET6 socket for v6 */ #define IN6P_PKTINFO 0x00010000 /* receive IP6 dst and I/F */ #define IN6P_HOPLIMIT 0x00020000 /* receive hoplimit */ #define IN6P_HOPOPTS 0x00040000 /* receive hop-by-hop options */ #define IN6P_DSTOPTS 0x00080000 /* receive dst options after rthdr */ #define IN6P_RTHDR 0x00100000 /* receive routing header */ #define IN6P_RTHDRDSTOPTS 0x00200000 /* receive dstoptions before rthdr */ #define IN6P_TCLASS 0x00400000 /* receive traffic class value */ #define IN6P_AUTOFLOWLABEL 0x00800000 /* attach flowlabel automatically */ #define INP_TIMEWAIT 0x01000000 /* in TIMEWAIT, ppcb is tcptw */ #define INP_ONESBCAST 0x02000000 /* send all-ones broadcast */ #define INP_DROPPED 0x04000000 /* protocol drop flag */ #define INP_SOCKREF 0x08000000 /* strong socket reference */ #define INP_RESERVED_0 0x10000000 /* reserved field */ #define INP_RESERVED_1 0x20000000 /* reserved field */ #define IN6P_RFC2292 0x40000000 /* used RFC2292 API on the socket */ #define IN6P_MTU 0x80000000 /* receive path MTU */ #define INP_CONTROLOPTS (INP_RECVOPTS|INP_RECVRETOPTS|INP_RECVDSTADDR|\ INP_RECVIF|INP_RECVTTL|INP_RECVTOS|\ IN6P_PKTINFO|IN6P_HOPLIMIT|IN6P_HOPOPTS|\ IN6P_DSTOPTS|IN6P_RTHDR|IN6P_RTHDRDSTOPTS|\ IN6P_TCLASS|IN6P_AUTOFLOWLABEL|IN6P_RFC2292|\ IN6P_MTU) /* * Flags for inp_flags2. */ #define INP_LLE_VALID 0x00000001 /* cached lle is valid */ #define INP_RT_VALID 0x00000002 /* cached rtentry is valid */ #define INP_PCBGROUPWILD 0x00000004 /* in pcbgroup wildcard list */ #define INP_REUSEPORT 0x00000008 /* SO_REUSEPORT option is set */ #define INP_FREED 0x00000010 /* inp itself is not valid */ #define INP_REUSEADDR 0x00000020 /* SO_REUSEADDR option is set */ #define INP_BINDMULTI 0x00000040 /* IP_BINDMULTI option is set */ #define INP_RSS_BUCKET_SET 0x00000080 /* IP_RSS_LISTEN_BUCKET is set */ #define INP_RECVFLOWID 0x00000100 /* populate recv datagram with flow info */ #define INP_RECVRSSBUCKETID 0x00000200 /* populate recv datagram with bucket id */ #define INP_RATE_LIMIT_CHANGED 0x00000400 /* rate limit needs attention */ #define INP_ORIGDSTADDR 0x00000800 /* receive IP dst address/port */ #define INP_CANNOT_DO_ECN 0x00001000 /* The stack does not do ECN */ #define INP_REUSEPORT_LB 0x00002000 /* SO_REUSEPORT_LB option is set */ /* * Flags passed to in_pcblookup*() functions. */ #define INPLOOKUP_WILDCARD 0x00000001 /* Allow wildcard sockets. */ #define INPLOOKUP_RLOCKPCB 0x00000002 /* Return inpcb read-locked. */ #define INPLOOKUP_WLOCKPCB 0x00000004 /* Return inpcb write-locked. */ #define INPLOOKUP_MASK (INPLOOKUP_WILDCARD | INPLOOKUP_RLOCKPCB | \ INPLOOKUP_WLOCKPCB) #define sotoinpcb(so) ((struct inpcb *)(so)->so_pcb) #define sotoin6pcb(so) sotoinpcb(so) /* for KAME src sync over BSD*'s */ #define INP_SOCKAF(so) so->so_proto->pr_domain->dom_family #define INP_CHECK_SOCKAF(so, af) (INP_SOCKAF(so) == af) /* * Constants for pcbinfo.ipi_hashfields. */ #define IPI_HASHFIELDS_NONE 0 #define IPI_HASHFIELDS_2TUPLE 1 #define IPI_HASHFIELDS_4TUPLE 2 #ifdef _KERNEL VNET_DECLARE(int, ipport_reservedhigh); VNET_DECLARE(int, ipport_reservedlow); VNET_DECLARE(int, ipport_lowfirstauto); VNET_DECLARE(int, ipport_lowlastauto); VNET_DECLARE(int, ipport_firstauto); VNET_DECLARE(int, ipport_lastauto); VNET_DECLARE(int, ipport_hifirstauto); VNET_DECLARE(int, ipport_hilastauto); VNET_DECLARE(int, ipport_randomized); VNET_DECLARE(int, ipport_randomcps); VNET_DECLARE(int, ipport_randomtime); VNET_DECLARE(int, ipport_stoprandom); VNET_DECLARE(int, ipport_tcpallocs); #define V_ipport_reservedhigh VNET(ipport_reservedhigh) #define V_ipport_reservedlow VNET(ipport_reservedlow) #define V_ipport_lowfirstauto VNET(ipport_lowfirstauto) #define V_ipport_lowlastauto VNET(ipport_lowlastauto) #define V_ipport_firstauto VNET(ipport_firstauto) #define V_ipport_lastauto VNET(ipport_lastauto) #define V_ipport_hifirstauto VNET(ipport_hifirstauto) #define V_ipport_hilastauto VNET(ipport_hilastauto) #define V_ipport_randomized VNET(ipport_randomized) #define V_ipport_randomcps VNET(ipport_randomcps) #define V_ipport_randomtime VNET(ipport_randomtime) #define V_ipport_stoprandom VNET(ipport_stoprandom) #define V_ipport_tcpallocs VNET(ipport_tcpallocs) void in_pcbinfo_destroy(struct inpcbinfo *); void in_pcbinfo_init(struct inpcbinfo *, const char *, struct inpcbhead *, int, int, char *, uma_init, u_int); int in_pcbbind_check_bindmulti(const struct inpcb *ni, const struct inpcb *oi); struct inpcbgroup * in_pcbgroup_byhash(struct inpcbinfo *, u_int, uint32_t); struct inpcbgroup * in_pcbgroup_byinpcb(struct inpcb *); struct inpcbgroup * in_pcbgroup_bytuple(struct inpcbinfo *, struct in_addr, u_short, struct in_addr, u_short); void in_pcbgroup_destroy(struct inpcbinfo *); int in_pcbgroup_enabled(struct inpcbinfo *); void in_pcbgroup_init(struct inpcbinfo *, u_int, int); void in_pcbgroup_remove(struct inpcb *); void in_pcbgroup_update(struct inpcb *); void in_pcbgroup_update_mbuf(struct inpcb *, struct mbuf *); void in_pcbpurgeif0(struct inpcbinfo *, struct ifnet *); int in_pcballoc(struct socket *, struct inpcbinfo *); int in_pcbbind(struct inpcb *, struct sockaddr *, struct ucred *); int in_pcb_lport(struct inpcb *, struct in_addr *, u_short *, struct ucred *, int); int in_pcbbind_setup(struct inpcb *, struct sockaddr *, in_addr_t *, u_short *, struct ucred *); int in_pcbconnect(struct inpcb *, struct sockaddr *, struct ucred *); int in_pcbconnect_mbuf(struct inpcb *, struct sockaddr *, struct ucred *, struct mbuf *); int in_pcbconnect_setup(struct inpcb *, struct sockaddr *, in_addr_t *, u_short *, in_addr_t *, u_short *, struct inpcb **, struct ucred *); void in_pcbdetach(struct inpcb *); void in_pcbdisconnect(struct inpcb *); void in_pcbdrop(struct inpcb *); void in_pcbfree(struct inpcb *); int in_pcbinshash(struct inpcb *); int in_pcbinshash_nopcbgroup(struct inpcb *); int in_pcbladdr(struct inpcb *, struct in_addr *, struct in_addr *, struct ucred *); struct inpcb * in_pcblookup_local(struct inpcbinfo *, struct in_addr, u_short, int, struct ucred *); struct inpcb * in_pcblookup(struct inpcbinfo *, struct in_addr, u_int, struct in_addr, u_int, int, struct ifnet *); struct inpcb * in_pcblookup_mbuf(struct inpcbinfo *, struct in_addr, u_int, struct in_addr, u_int, int, struct ifnet *, struct mbuf *); void in_pcbnotifyall(struct inpcbinfo *pcbinfo, struct in_addr, int, struct inpcb *(*)(struct inpcb *, int)); void in_pcbref(struct inpcb *); void in_pcbrehash(struct inpcb *); void in_pcbrehash_mbuf(struct inpcb *, struct mbuf *); int in_pcbrele(struct inpcb *); int in_pcbrele_rlocked(struct inpcb *); int in_pcbrele_wlocked(struct inpcb *); void in_pcblist_rele_rlocked(epoch_context_t ctx); void in_losing(struct inpcb *); void in_pcbsetsolabel(struct socket *so); int in_getpeeraddr(struct socket *so, struct sockaddr **nam); int in_getsockaddr(struct socket *so, struct sockaddr **nam); struct sockaddr * in_sockaddr(in_port_t port, struct in_addr *addr); void in_pcbsosetlabel(struct socket *so); #ifdef RATELIMIT int in_pcbattach_txrtlmt(struct inpcb *, struct ifnet *, uint32_t, uint32_t, uint32_t); void in_pcbdetach_txrtlmt(struct inpcb *); int in_pcbmodify_txrtlmt(struct inpcb *, uint32_t); int in_pcbquery_txrtlmt(struct inpcb *, uint32_t *); int in_pcbquery_txrlevel(struct inpcb *, uint32_t *); void in_pcboutput_txrtlmt(struct inpcb *, struct ifnet *, struct mbuf *); void in_pcboutput_eagain(struct inpcb *); #endif #endif /* _KERNEL */ #endif /* !_NETINET_IN_PCB_H_ */ Index: head/sys/netinet/tcp_hpts.c =================================================================== --- head/sys/netinet/tcp_hpts.c (revision 335355) +++ head/sys/netinet/tcp_hpts.c (revision 335356) @@ -1,1963 +1,1902 @@ /*- * Copyright (c) 2016-2018 Netflix Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_tcpdebug.h" /** * Some notes about usage. * * The tcp_hpts system is designed to provide a high precision timer * system for tcp. Its main purpose is to provide a mechanism for * pacing packets out onto the wire. It can be used in two ways * by a given TCP stack (and those two methods can be used simultaneously). * * First, and probably the main thing its used by Rack and BBR for, it can * be used to call tcp_output() of a transport stack at some time in the future. * The normal way this is done is that tcp_output() of the stack schedules * itself to be called again by calling tcp_hpts_insert(tcpcb, slot). The * slot is the time from now that the stack wants to be called but it * must be converted to tcp_hpts's notion of slot. This is done with * one of the macros HPTS_MS_TO_SLOTS or HPTS_USEC_TO_SLOTS. So a typical * call from the tcp_output() routine might look like: * * tcp_hpts_insert(tp, HPTS_USEC_TO_SLOTS(550)); * * The above would schedule tcp_ouput() to be called in 550 useconds. * Note that if using this mechanism the stack will want to add near * its top a check to prevent unwanted calls (from user land or the * arrival of incoming ack's). So it would add something like: * * if (inp->inp_in_hpts) * return; * * to prevent output processing until the time alotted has gone by. * Of course this is a bare bones example and the stack will probably * have more consideration then just the above. * * Now the tcp_hpts system will call tcp_output in one of two forms, * it will first check to see if the stack as defined a * tfb_tcp_output_wtime() function, if so that is the routine it * will call, if that function is not defined then it will call the * tfb_tcp_output() function. The only difference between these * two calls is that the former passes the time in to the function * so the function does not have to access the time (which tcp_hpts * already has). What these functions do is of course totally up * to the individual tcp stack. * * Now the second function (actually two functions I guess :D) * the tcp_hpts system provides is the ability to either abort * a connection (later) or process input on a connection. * Why would you want to do this? To keep processor locality. * * So in order to use the input redirection function the * stack changes its tcp_do_segment() routine to instead * of process the data call the function: * * tcp_queue_pkt_to_input() * * You will note that the arguments to this function look * a lot like tcp_do_segments's arguments. This function * will assure that the tcp_hpts system will * call the functions tfb_tcp_hpts_do_segment() from the * correct CPU. Note that multiple calls can get pushed * into the tcp_hpts system this will be indicated by * the next to last argument to tfb_tcp_hpts_do_segment() * (nxt_pkt). If nxt_pkt is a 1 then another packet is * coming. If nxt_pkt is a 0 then this is the last call * that the tcp_hpts system has available for the tcp stack. * * The other point of the input system is to be able to safely * drop a tcp connection without worrying about the recursive * locking that may be occuring on the INP_WLOCK. So if * a stack wants to drop a connection it calls: * * tcp_set_inp_to_drop(tp, ETIMEDOUT) * * To schedule the tcp_hpts system to call * * tcp_drop(tp, drop_reason) * * at a future point. This is quite handy to prevent locking * issues when dropping connections. * */ #include #include #include #include #include #include #include #include #include /* for proc0 declaration */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define TCPSTATES /* for logging */ #include #include #include #include #include /* required for icmp_var.h */ #include /* for ICMP_BANDLIM */ #include #include #include #include #include #include #include #include #include #include #include #include #ifdef tcpdebug #include #endif /* tcpdebug */ #ifdef tcp_offload #include #endif #ifdef ipsec #include #include #endif /* ipsec */ #include "opt_rss.h" MALLOC_DEFINE(M_TCPHPTS, "tcp_hpts", "TCP hpts"); #ifdef RSS static int tcp_bind_threads = 1; #else static int tcp_bind_threads = 0; #endif TUNABLE_INT("net.inet.tcp.bind_hptss", &tcp_bind_threads); static uint32_t tcp_hpts_logging_size = DEFAULT_HPTS_LOG; TUNABLE_INT("net.inet.tcp.hpts_logging_sz", &tcp_hpts_logging_size); static struct tcp_hptsi tcp_pace; -static int -tcp_hptsi_lock_inpinfo(struct inpcb *inp, - struct tcpcb **tp); static void tcp_wakehpts(struct tcp_hpts_entry *p); static void tcp_wakeinput(struct tcp_hpts_entry *p); static void tcp_input_data(struct tcp_hpts_entry *hpts, struct timeval *tv); static void tcp_hptsi(struct tcp_hpts_entry *hpts, struct timeval *ctick); static void tcp_hpts_thread(void *ctx); static void tcp_init_hptsi(void *st); int32_t tcp_min_hptsi_time = DEFAULT_MIN_SLEEP; static int32_t tcp_hpts_callout_skip_swi = 0; SYSCTL_NODE(_net_inet_tcp, OID_AUTO, hpts, CTLFLAG_RW, 0, "TCP Hpts controls"); #define timersub(tvp, uvp, vvp) \ do { \ (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec; \ (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec; \ if ((vvp)->tv_usec < 0) { \ (vvp)->tv_sec--; \ (vvp)->tv_usec += 1000000; \ } \ } while (0) static int32_t logging_on = 0; static int32_t hpts_sleep_max = (NUM_OF_HPTSI_SLOTS - 2); static int32_t tcp_hpts_precision = 120; SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, precision, CTLFLAG_RW, &tcp_hpts_precision, 120, "Value for PRE() precision of callout"); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, logging, CTLFLAG_RW, &logging_on, 0, "Turn on logging if compiled in"); counter_u64_t hpts_loops; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, loops, CTLFLAG_RD, &hpts_loops, "Number of times hpts had to loop to catch up"); counter_u64_t back_tosleep; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, no_tcbsfound, CTLFLAG_RD, &back_tosleep, "Number of times hpts found no tcbs"); static int32_t in_newts_every_tcb = 0; SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, in_tsperpcb, CTLFLAG_RW, &in_newts_every_tcb, 0, "Do we have a new cts every tcb we process for input"); static int32_t in_ts_percision = 0; SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, in_tspercision, CTLFLAG_RW, &in_ts_percision, 0, "Do we use percise timestamp for clients on input"); static int32_t out_newts_every_tcb = 0; SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, out_tsperpcb, CTLFLAG_RW, &out_newts_every_tcb, 0, "Do we have a new cts every tcb we process for output"); static int32_t out_ts_percision = 0; SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, out_tspercision, CTLFLAG_RW, &out_ts_percision, 0, "Do we use a percise timestamp for every output cts"); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, maxsleep, CTLFLAG_RW, &hpts_sleep_max, 0, "The maximum time the hpts will sleep <1 - 254>"); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, minsleep, CTLFLAG_RW, &tcp_min_hptsi_time, 0, "The minimum time the hpts must sleep before processing more slots"); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, skip_swi, CTLFLAG_RW, &tcp_hpts_callout_skip_swi, 0, "Do we have the callout call directly to the hpts?"); static void __tcp_hpts_log_it(struct tcp_hpts_entry *hpts, struct inpcb *inp, int event, uint32_t slot, uint32_t ticknow, int32_t line) { struct hpts_log *pl; HPTS_MTX_ASSERT(hpts); if (hpts->p_log == NULL) return; pl = &hpts->p_log[hpts->p_log_at]; hpts->p_log_at++; if (hpts->p_log_at >= hpts->p_logsize) { hpts->p_log_at = 0; hpts->p_log_wrapped = 1; } pl->inp = inp; if (inp) { pl->t_paceslot = inp->inp_hptsslot; pl->t_hptsreq = inp->inp_hpts_request; pl->p_onhpts = inp->inp_in_hpts; pl->p_oninput = inp->inp_in_input; } else { pl->t_paceslot = 0; pl->t_hptsreq = 0; pl->p_onhpts = 0; pl->p_oninput = 0; } pl->is_notempty = 1; pl->event = event; pl->line = line; pl->cts = tcp_get_usecs(NULL); pl->p_curtick = hpts->p_curtick; pl->p_prevtick = hpts->p_prevtick; pl->p_on_queue_cnt = hpts->p_on_queue_cnt; pl->ticknow = ticknow; pl->slot_req = slot; pl->p_nxt_slot = hpts->p_nxt_slot; pl->p_cur_slot = hpts->p_cur_slot; pl->p_hpts_sleep_time = hpts->p_hpts_sleep_time; pl->p_flags = (hpts->p_cpu & 0x7f); pl->p_flags <<= 7; pl->p_flags |= (hpts->p_num & 0x7f); pl->p_flags <<= 2; if (hpts->p_hpts_active) { pl->p_flags |= HPTS_HPTS_ACTIVE; } } #define tcp_hpts_log_it(a, b, c, d, e) __tcp_hpts_log_it(a, b, c, d, e, __LINE__) static void hpts_timeout_swi(void *arg) { struct tcp_hpts_entry *hpts; hpts = (struct tcp_hpts_entry *)arg; swi_sched(hpts->ie_cookie, 0); } static void hpts_timeout_dir(void *arg) { tcp_hpts_thread(arg); } static inline void hpts_sane_pace_remove(struct tcp_hpts_entry *hpts, struct inpcb *inp, struct hptsh *head, int clear) { #ifdef INVARIANTS if (mtx_owned(&hpts->p_mtx) == 0) { /* We don't own the mutex? */ panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp); } if (hpts->p_cpu != inp->inp_hpts_cpu) { /* It is not the right cpu/mutex? */ panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp); } if (inp->inp_in_hpts == 0) { /* We are not on the hpts? */ panic("%s: hpts:%p inp:%p not on the hpts?", __FUNCTION__, hpts, inp); } if (TAILQ_EMPTY(head) && (hpts->p_on_queue_cnt != 0)) { /* We should not be empty with a queue count */ panic("%s hpts:%p hpts bucket empty but cnt:%d", __FUNCTION__, hpts, hpts->p_on_queue_cnt); } #endif TAILQ_REMOVE(head, inp, inp_hpts); hpts->p_on_queue_cnt--; if (hpts->p_on_queue_cnt < 0) { /* Count should not go negative .. */ #ifdef INVARIANTS panic("Hpts goes negative inp:%p hpts:%p", inp, hpts); #endif hpts->p_on_queue_cnt = 0; } if (clear) { inp->inp_hpts_request = 0; inp->inp_in_hpts = 0; } } static inline void hpts_sane_pace_insert(struct tcp_hpts_entry *hpts, struct inpcb *inp, struct hptsh *head, int line, int noref) { #ifdef INVARIANTS if (mtx_owned(&hpts->p_mtx) == 0) { /* We don't own the mutex? */ panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp); } if (hpts->p_cpu != inp->inp_hpts_cpu) { /* It is not the right cpu/mutex? */ panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp); } if ((noref == 0) && (inp->inp_in_hpts == 1)) { /* We are already on the hpts? */ panic("%s: hpts:%p inp:%p already on the hpts?", __FUNCTION__, hpts, inp); } #endif TAILQ_INSERT_TAIL(head, inp, inp_hpts); inp->inp_in_hpts = 1; hpts->p_on_queue_cnt++; if (noref == 0) { in_pcbref(inp); } } static inline void hpts_sane_input_remove(struct tcp_hpts_entry *hpts, struct inpcb *inp, int clear) { #ifdef INVARIANTS if (mtx_owned(&hpts->p_mtx) == 0) { /* We don't own the mutex? */ panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp); } if (hpts->p_cpu != inp->inp_input_cpu) { /* It is not the right cpu/mutex? */ panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp); } if (inp->inp_in_input == 0) { /* We are not on the input hpts? */ panic("%s: hpts:%p inp:%p not on the input hpts?", __FUNCTION__, hpts, inp); } #endif TAILQ_REMOVE(&hpts->p_input, inp, inp_input); hpts->p_on_inqueue_cnt--; if (hpts->p_on_inqueue_cnt < 0) { #ifdef INVARIANTS panic("Hpts in goes negative inp:%p hpts:%p", inp, hpts); #endif hpts->p_on_inqueue_cnt = 0; } #ifdef INVARIANTS if (TAILQ_EMPTY(&hpts->p_input) && (hpts->p_on_inqueue_cnt != 0)) { /* We should not be empty with a queue count */ panic("%s hpts:%p in_hpts input empty but cnt:%d", __FUNCTION__, hpts, hpts->p_on_inqueue_cnt); } #endif if (clear) inp->inp_in_input = 0; } static inline void hpts_sane_input_insert(struct tcp_hpts_entry *hpts, struct inpcb *inp, int line) { #ifdef INVARIANTS if (mtx_owned(&hpts->p_mtx) == 0) { /* We don't own the mutex? */ panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp); } if (hpts->p_cpu != inp->inp_input_cpu) { /* It is not the right cpu/mutex? */ panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp); } if (inp->inp_in_input == 1) { /* We are already on the input hpts? */ panic("%s: hpts:%p inp:%p already on the input hpts?", __FUNCTION__, hpts, inp); } #endif TAILQ_INSERT_TAIL(&hpts->p_input, inp, inp_input); inp->inp_in_input = 1; hpts->p_on_inqueue_cnt++; in_pcbref(inp); } static int sysctl_tcp_hpts_log(SYSCTL_HANDLER_ARGS) { struct tcp_hpts_entry *hpts; size_t sz; int32_t logging_was, i; int32_t error = 0; /* * HACK: Turn off logging so no locks are required this really needs * a memory barrier :) */ logging_was = logging_on; logging_on = 0; if (!req->oldptr) { /* How much? */ sz = 0; for (i = 0; i < tcp_pace.rp_num_hptss; i++) { hpts = tcp_pace.rp_ent[i]; if (hpts->p_log == NULL) continue; sz += (sizeof(struct hpts_log) * hpts->p_logsize); } error = SYSCTL_OUT(req, 0, sz); } else { for (i = 0; i < tcp_pace.rp_num_hptss; i++) { hpts = tcp_pace.rp_ent[i]; if (hpts->p_log == NULL) continue; if (hpts->p_log_wrapped) sz = (sizeof(struct hpts_log) * hpts->p_logsize); else sz = (sizeof(struct hpts_log) * hpts->p_log_at); error = SYSCTL_OUT(req, hpts->p_log, sz); } } logging_on = logging_was; return error; } SYSCTL_PROC(_net_inet_tcp_hpts, OID_AUTO, log, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0, sysctl_tcp_hpts_log, "A", "tcp hptsi log"); -/* - * Try to get the INP_INFO lock. - * - * This function always succeeds in getting the lock. It will clear - * *tpp and return (1) if something critical changed while the inpcb - * was unlocked. Otherwise, it will leave *tpp unchanged and return (0). - * - * This function relies on the fact that the hpts always holds a - * reference on the inpcb while the segment is on the hptsi wheel and - * in the input queue. - * - */ -static int -tcp_hptsi_lock_inpinfo(struct inpcb *inp, struct tcpcb **tpp) -{ - struct tcp_function_block *tfb; - struct tcpcb *tp; - void *ptr; - - /* Try the easy way. */ - if (INP_INFO_TRY_RLOCK(&V_tcbinfo)) - return (0); - - /* - * OK, let's try the hard way. We'll save the function pointer block - * to make sure that doesn't change while we aren't holding the - * lock. - */ - tp = *tpp; - tfb = tp->t_fb; - ptr = tp->t_fb_ptr; - INP_WUNLOCK(inp); - INP_INFO_RLOCK(&V_tcbinfo); - INP_WLOCK(inp); - /* If the session went away, return an error. */ - if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) || - (inp->inp_flags2 & INP_FREED)) { - *tpp = NULL; - return (1); - } - /* - * If the function block or stack-specific data block changed, - * report an error. - */ - tp = intotcpcb(inp); - if ((tp->t_fb != tfb) && (tp->t_fb_ptr != ptr)) { - *tpp = NULL; - return (1); - } - return (0); -} - - static void tcp_wakehpts(struct tcp_hpts_entry *hpts) { HPTS_MTX_ASSERT(hpts); swi_sched(hpts->ie_cookie, 0); if (hpts->p_hpts_active == 2) { /* Rare sleeping on a ENOBUF */ wakeup_one(hpts); } } static void tcp_wakeinput(struct tcp_hpts_entry *hpts) { HPTS_MTX_ASSERT(hpts); swi_sched(hpts->ie_cookie, 0); if (hpts->p_hpts_active == 2) { /* Rare sleeping on a ENOBUF */ wakeup_one(hpts); } } struct tcp_hpts_entry * tcp_cur_hpts(struct inpcb *inp) { int32_t hpts_num; struct tcp_hpts_entry *hpts; hpts_num = inp->inp_hpts_cpu; hpts = tcp_pace.rp_ent[hpts_num]; return (hpts); } struct tcp_hpts_entry * tcp_hpts_lock(struct inpcb *inp) { struct tcp_hpts_entry *hpts; int32_t hpts_num; again: hpts_num = inp->inp_hpts_cpu; hpts = tcp_pace.rp_ent[hpts_num]; #ifdef INVARIANTS if (mtx_owned(&hpts->p_mtx)) { panic("Hpts:%p owns mtx prior-to lock line:%d", hpts, __LINE__); } #endif mtx_lock(&hpts->p_mtx); if (hpts_num != inp->inp_hpts_cpu) { mtx_unlock(&hpts->p_mtx); goto again; } return (hpts); } struct tcp_hpts_entry * tcp_input_lock(struct inpcb *inp) { struct tcp_hpts_entry *hpts; int32_t hpts_num; again: hpts_num = inp->inp_input_cpu; hpts = tcp_pace.rp_ent[hpts_num]; #ifdef INVARIANTS if (mtx_owned(&hpts->p_mtx)) { panic("Hpts:%p owns mtx prior-to lock line:%d", hpts, __LINE__); } #endif mtx_lock(&hpts->p_mtx); if (hpts_num != inp->inp_input_cpu) { mtx_unlock(&hpts->p_mtx); goto again; } return (hpts); } static void tcp_remove_hpts_ref(struct inpcb *inp, struct tcp_hpts_entry *hpts, int line) { int32_t add_freed; if (inp->inp_flags2 & INP_FREED) { /* * Need to play a special trick so that in_pcbrele_wlocked * does not return 1 when it really should have returned 0. */ add_freed = 1; inp->inp_flags2 &= ~INP_FREED; } else { add_freed = 0; } #ifndef INP_REF_DEBUG if (in_pcbrele_wlocked(inp)) { /* * This should not happen. We have the inpcb referred to by * the main socket (why we are called) and the hpts. It * should always return 0. */ panic("inpcb:%p release ret 1", inp); } #else if (__in_pcbrele_wlocked(inp, line)) { /* * This should not happen. We have the inpcb referred to by * the main socket (why we are called) and the hpts. It * should always return 0. */ panic("inpcb:%p release ret 1", inp); } #endif if (add_freed) { inp->inp_flags2 |= INP_FREED; } } static void tcp_hpts_remove_locked_output(struct tcp_hpts_entry *hpts, struct inpcb *inp, int32_t flags, int32_t line) { if (inp->inp_in_hpts) { hpts_sane_pace_remove(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], 1); tcp_remove_hpts_ref(inp, hpts, line); } } static void tcp_hpts_remove_locked_input(struct tcp_hpts_entry *hpts, struct inpcb *inp, int32_t flags, int32_t line) { HPTS_MTX_ASSERT(hpts); if (inp->inp_in_input) { hpts_sane_input_remove(hpts, inp, 1); tcp_remove_hpts_ref(inp, hpts, line); } } /* * Called normally with the INP_LOCKED but it * does not matter, the hpts lock is the key * but the lock order allows us to hold the * INP lock and then get the hpts lock. * * Valid values in the flags are * HPTS_REMOVE_OUTPUT - remove from the output of the hpts. * HPTS_REMOVE_INPUT - remove from the input of the hpts. * Note that you can or both values together and get two * actions. */ void __tcp_hpts_remove(struct inpcb *inp, int32_t flags, int32_t line) { struct tcp_hpts_entry *hpts; INP_WLOCK_ASSERT(inp); if (flags & HPTS_REMOVE_OUTPUT) { hpts = tcp_hpts_lock(inp); tcp_hpts_remove_locked_output(hpts, inp, flags, line); mtx_unlock(&hpts->p_mtx); } if (flags & HPTS_REMOVE_INPUT) { hpts = tcp_input_lock(inp); tcp_hpts_remove_locked_input(hpts, inp, flags, line); mtx_unlock(&hpts->p_mtx); } } static inline int hpts_tick(struct tcp_hpts_entry *hpts, int32_t plus) { return ((hpts->p_prevtick + plus) % NUM_OF_HPTSI_SLOTS); } static int tcp_queue_to_hpts_immediate_locked(struct inpcb *inp, struct tcp_hpts_entry *hpts, int32_t line, int32_t noref) { int32_t need_wake = 0; uint32_t ticknow = 0; HPTS_MTX_ASSERT(hpts); if (inp->inp_in_hpts == 0) { /* Ok we need to set it on the hpts in the current slot */ if (hpts->p_hpts_active == 0) { /* A sleeping hpts we want in next slot to run */ if (logging_on) { tcp_hpts_log_it(hpts, inp, HPTSLOG_INSERT_SLEEPER, 0, hpts_tick(hpts, 1)); } inp->inp_hptsslot = hpts_tick(hpts, 1); inp->inp_hpts_request = 0; if (logging_on) { tcp_hpts_log_it(hpts, inp, HPTSLOG_SLEEP_BEFORE, 1, ticknow); } need_wake = 1; } else if ((void *)inp == hpts->p_inp) { /* * We can't allow you to go into the same slot we * are in. We must put you out. */ inp->inp_hptsslot = hpts->p_nxt_slot; } else inp->inp_hptsslot = hpts->p_cur_slot; hpts_sane_pace_insert(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], line, noref); inp->inp_hpts_request = 0; if (logging_on) { tcp_hpts_log_it(hpts, inp, HPTSLOG_IMMEDIATE, 0, 0); } if (need_wake) { /* * Activate the hpts if it is sleeping and its * timeout is not 1. */ if (logging_on) { tcp_hpts_log_it(hpts, inp, HPTSLOG_WAKEUP_HPTS, 0, ticknow); } hpts->p_direct_wake = 1; tcp_wakehpts(hpts); } } return (need_wake); } int __tcp_queue_to_hpts_immediate(struct inpcb *inp, int32_t line) { int32_t ret; struct tcp_hpts_entry *hpts; INP_WLOCK_ASSERT(inp); hpts = tcp_hpts_lock(inp); ret = tcp_queue_to_hpts_immediate_locked(inp, hpts, line, 0); mtx_unlock(&hpts->p_mtx); return (ret); } static void tcp_hpts_insert_locked(struct tcp_hpts_entry *hpts, struct inpcb *inp, uint32_t slot, uint32_t cts, int32_t line, struct hpts_diag *diag, int32_t noref) { int32_t need_new_to = 0; int32_t need_wakeup = 0; uint32_t largest_slot; uint32_t ticknow = 0; uint32_t slot_calc; HPTS_MTX_ASSERT(hpts); if (diag) { memset(diag, 0, sizeof(struct hpts_diag)); diag->p_hpts_active = hpts->p_hpts_active; diag->p_nxt_slot = hpts->p_nxt_slot; diag->p_cur_slot = hpts->p_cur_slot; diag->slot_req = slot; } if ((inp->inp_in_hpts == 0) || noref) { inp->inp_hpts_request = slot; if (slot == 0) { /* Immediate */ tcp_queue_to_hpts_immediate_locked(inp, hpts, line, noref); return; } if (hpts->p_hpts_active) { /* * Its slot - 1 since nxt_slot is the next tick that * will go off since the hpts is awake */ if (logging_on) { tcp_hpts_log_it(hpts, inp, HPTSLOG_INSERT_NORMAL, slot, 0); } /* * We want to make sure that we don't place a inp in * the range of p_cur_slot <-> p_nxt_slot. If we * take from p_nxt_slot to the end, plus p_cur_slot * and then take away 2, we will know how many is * the max slots we can use. */ if (hpts->p_nxt_slot > hpts->p_cur_slot) { /* * Non-wrap case nxt_slot <-> cur_slot we * don't want to land in. So the diff gives * us what is taken away from the number of * slots. */ largest_slot = NUM_OF_HPTSI_SLOTS - (hpts->p_nxt_slot - hpts->p_cur_slot); } else if (hpts->p_nxt_slot == hpts->p_cur_slot) { largest_slot = NUM_OF_HPTSI_SLOTS - 2; } else { /* * Wrap case so the diff gives us the number * of slots that we can land in. */ largest_slot = hpts->p_cur_slot - hpts->p_nxt_slot; } /* * We take away two so we never have a problem (20 * usec's) out of 1024000 usecs */ largest_slot -= 2; if (inp->inp_hpts_request > largest_slot) { /* * Restrict max jump of slots and remember * leftover */ slot = largest_slot; inp->inp_hpts_request -= largest_slot; } else { /* This one will run when we hit it */ inp->inp_hpts_request = 0; } if (hpts->p_nxt_slot == hpts->p_cur_slot) slot_calc = (hpts->p_nxt_slot + slot) % NUM_OF_HPTSI_SLOTS; else slot_calc = (hpts->p_nxt_slot + slot - 1) % NUM_OF_HPTSI_SLOTS; if (slot_calc == hpts->p_cur_slot) { #ifdef INVARIANTS /* TSNH */ panic("Hpts:%p impossible slot calculation slot_calc:%u slot:%u largest:%u\n", hpts, slot_calc, slot, largest_slot); #endif if (slot_calc) slot_calc--; else slot_calc = NUM_OF_HPTSI_SLOTS - 1; } inp->inp_hptsslot = slot_calc; if (diag) { diag->inp_hptsslot = inp->inp_hptsslot; } } else { /* * The hpts is sleeping, we need to figure out where * it will wake up at and if we need to reschedule * its time-out. */ uint32_t have_slept, yet_to_sleep; uint32_t slot_now; struct timeval tv; ticknow = tcp_gethptstick(&tv); slot_now = ticknow % NUM_OF_HPTSI_SLOTS; /* * The user wants to be inserted at (slot_now + * slot) % NUM_OF_HPTSI_SLOTS, so lets set that up. */ largest_slot = NUM_OF_HPTSI_SLOTS - 2; if (inp->inp_hpts_request > largest_slot) { /* Adjust the residual in inp_hpts_request */ slot = largest_slot; inp->inp_hpts_request -= largest_slot; } else { /* No residual it all fits */ inp->inp_hpts_request = 0; } inp->inp_hptsslot = (slot_now + slot) % NUM_OF_HPTSI_SLOTS; if (diag) { diag->slot_now = slot_now; diag->inp_hptsslot = inp->inp_hptsslot; diag->p_on_min_sleep = hpts->p_on_min_sleep; } if (logging_on) { tcp_hpts_log_it(hpts, inp, HPTSLOG_INSERT_SLEEPER, slot, ticknow); } /* Now do we need to restart the hpts's timer? */ if (TSTMP_GT(ticknow, hpts->p_curtick)) have_slept = ticknow - hpts->p_curtick; else have_slept = 0; if (have_slept < hpts->p_hpts_sleep_time) { /* This should be what happens */ yet_to_sleep = hpts->p_hpts_sleep_time - have_slept; } else { /* We are over-due */ yet_to_sleep = 0; need_wakeup = 1; } if (diag) { diag->have_slept = have_slept; diag->yet_to_sleep = yet_to_sleep; diag->hpts_sleep_time = hpts->p_hpts_sleep_time; } if ((hpts->p_on_min_sleep == 0) && (yet_to_sleep > slot)) { /* * We need to reschedule the hptss time-out. */ hpts->p_hpts_sleep_time = slot; need_new_to = slot * HPTS_TICKS_PER_USEC; } } hpts_sane_pace_insert(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], line, noref); if (logging_on) { tcp_hpts_log_it(hpts, inp, HPTSLOG_INSERTED, slot, ticknow); } /* * Now how far is the hpts sleeping to? if active is 1, its * up and ticking we do nothing, otherwise we may need to * reschedule its callout if need_new_to is set from above. */ if (need_wakeup) { if (logging_on) { tcp_hpts_log_it(hpts, inp, HPTSLOG_RESCHEDULE, 1, 0); } hpts->p_direct_wake = 1; tcp_wakehpts(hpts); if (diag) { diag->need_new_to = 0; diag->co_ret = 0xffff0000; } } else if (need_new_to) { int32_t co_ret; struct timeval tv; sbintime_t sb; tv.tv_sec = 0; tv.tv_usec = 0; while (need_new_to > HPTS_USEC_IN_SEC) { tv.tv_sec++; need_new_to -= HPTS_USEC_IN_SEC; } tv.tv_usec = need_new_to; sb = tvtosbt(tv); if (tcp_hpts_callout_skip_swi == 0) { co_ret = callout_reset_sbt_on(&hpts->co, sb, 0, hpts_timeout_swi, hpts, hpts->p_cpu, (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision))); } else { co_ret = callout_reset_sbt_on(&hpts->co, sb, 0, hpts_timeout_dir, hpts, hpts->p_cpu, C_PREL(tcp_hpts_precision)); } if (diag) { diag->need_new_to = need_new_to; diag->co_ret = co_ret; } } } else { #ifdef INVARIANTS panic("Hpts:%p tp:%p already on hpts and add?", hpts, inp); #endif } } uint32_t tcp_hpts_insert_diag(struct inpcb *inp, uint32_t slot, int32_t line, struct hpts_diag *diag){ struct tcp_hpts_entry *hpts; uint32_t slot_on, cts; struct timeval tv; /* * We now return the next-slot the hpts will be on, beyond its * current run (if up) or where it was when it stopped if it is * sleeping. */ INP_WLOCK_ASSERT(inp); hpts = tcp_hpts_lock(inp); if (in_ts_percision) microuptime(&tv); else getmicrouptime(&tv); cts = tcp_tv_to_usectick(&tv); tcp_hpts_insert_locked(hpts, inp, slot, cts, line, diag, 0); slot_on = hpts->p_nxt_slot; mtx_unlock(&hpts->p_mtx); return (slot_on); } uint32_t __tcp_hpts_insert(struct inpcb *inp, uint32_t slot, int32_t line){ return (tcp_hpts_insert_diag(inp, slot, line, NULL)); } int __tcp_queue_to_input_locked(struct inpcb *inp, struct tcp_hpts_entry *hpts, int32_t line) { int32_t retval = 0; HPTS_MTX_ASSERT(hpts); if (inp->inp_in_input == 0) { /* Ok we need to set it on the hpts in the current slot */ hpts_sane_input_insert(hpts, inp, line); retval = 1; if (hpts->p_hpts_active == 0) { /* * Activate the hpts if it is sleeping. */ if (logging_on) { tcp_hpts_log_it(hpts, inp, HPTSLOG_WAKEUP_INPUT, 0, 0); } retval = 2; hpts->p_direct_wake = 1; tcp_wakeinput(hpts); } } else if (hpts->p_hpts_active == 0) { retval = 4; hpts->p_direct_wake = 1; tcp_wakeinput(hpts); } return (retval); } void tcp_queue_pkt_to_input(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th, int32_t tlen, int32_t drop_hdrlen, uint8_t iptos, uint8_t ti_locked) { /* Setup packet for input first */ INP_WLOCK_ASSERT(tp->t_inpcb); m->m_pkthdr.pace_thoff = (uint16_t) ((caddr_t)th - mtod(m, caddr_t)); m->m_pkthdr.pace_tlen = (uint16_t) tlen; m->m_pkthdr.pace_drphdrlen = drop_hdrlen; m->m_pkthdr.pace_tos = iptos; m->m_pkthdr.pace_lock = (uint8_t) ti_locked; if (tp->t_in_pkt == NULL) { tp->t_in_pkt = m; tp->t_tail_pkt = m; } else { tp->t_tail_pkt->m_nextpkt = m; tp->t_tail_pkt = m; } } int32_t __tcp_queue_to_input(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th, int32_t tlen, int32_t drop_hdrlen, uint8_t iptos, uint8_t ti_locked, int32_t line){ struct tcp_hpts_entry *hpts; int32_t ret; tcp_queue_pkt_to_input(tp, m, th, tlen, drop_hdrlen, iptos, ti_locked); hpts = tcp_input_lock(tp->t_inpcb); ret = __tcp_queue_to_input_locked(tp->t_inpcb, hpts, line); mtx_unlock(&hpts->p_mtx); return (ret); } void __tcp_set_inp_to_drop(struct inpcb *inp, uint16_t reason, int32_t line) { struct tcp_hpts_entry *hpts; struct tcpcb *tp; tp = intotcpcb(inp); hpts = tcp_input_lock(tp->t_inpcb); if (inp->inp_in_input == 0) { /* Ok we need to set it on the hpts in the current slot */ hpts_sane_input_insert(hpts, inp, line); if (hpts->p_hpts_active == 0) { /* * Activate the hpts if it is sleeping. */ hpts->p_direct_wake = 1; tcp_wakeinput(hpts); } } else if (hpts->p_hpts_active == 0) { hpts->p_direct_wake = 1; tcp_wakeinput(hpts); } inp->inp_hpts_drop_reas = reason; mtx_unlock(&hpts->p_mtx); } static uint16_t hpts_random_cpu(struct inpcb *inp){ /* * No flow type set distribute the load randomly. */ uint16_t cpuid; uint32_t ran; /* * If one has been set use it i.e. we want both in and out on the * same hpts. */ if (inp->inp_input_cpu_set) { return (inp->inp_input_cpu); } else if (inp->inp_hpts_cpu_set) { return (inp->inp_hpts_cpu); } /* Nothing set use a random number */ ran = arc4random(); cpuid = (ran & 0xffff) % mp_ncpus; return (cpuid); } static uint16_t hpts_cpuid(struct inpcb *inp){ uint16_t cpuid; /* * If one has been set use it i.e. we want both in and out on the * same hpts. */ if (inp->inp_input_cpu_set) { return (inp->inp_input_cpu); } else if (inp->inp_hpts_cpu_set) { return (inp->inp_hpts_cpu); } /* If one is set the other must be the same */ #ifdef RSS cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype); if (cpuid == NETISR_CPUID_NONE) return (hpts_random_cpu(inp)); else return (cpuid); #else /* * We don't have a flowid -> cpuid mapping, so cheat and just map * unknown cpuids to curcpu. Not the best, but apparently better * than defaulting to swi 0. */ if (inp->inp_flowtype != M_HASHTYPE_NONE) { cpuid = inp->inp_flowid % mp_ncpus; return (cpuid); } cpuid = hpts_random_cpu(inp); return (cpuid); #endif } /* * Do NOT try to optimize the processing of inp's * by first pulling off all the inp's into a temporary * list (e.g. TAILQ_CONCAT). If you do that the subtle * interactions of switching CPU's will kill because of * problems in the linked list manipulation. Basically * you would switch cpu's with the hpts mutex locked * but then while you were processing one of the inp's * some other one that you switch will get a new * packet on the different CPU. It will insert it * on the new hptss input list. Creating a temporary * link in the inp will not fix it either, since * the other hpts will be doing the same thing and * you will both end up using the temporary link. * * You will die in an ASSERT for tailq corruption if you * run INVARIANTS or you will die horribly without * INVARIANTS in some unknown way with a corrupt linked * list. */ static void tcp_input_data(struct tcp_hpts_entry *hpts, struct timeval *tv) { struct mbuf *m, *n; struct tcpcb *tp; struct inpcb *inp; uint16_t drop_reason; int16_t set_cpu; uint32_t did_prefetch = 0; int32_t ti_locked = TI_UNLOCKED; HPTS_MTX_ASSERT(hpts); while ((inp = TAILQ_FIRST(&hpts->p_input)) != NULL) { HPTS_MTX_ASSERT(hpts); hpts_sane_input_remove(hpts, inp, 0); if (inp->inp_input_cpu_set == 0) { set_cpu = 1; } else { set_cpu = 0; } hpts->p_inp = inp; drop_reason = inp->inp_hpts_drop_reas; inp->inp_in_input = 0; tp = intotcpcb(inp); mtx_unlock(&hpts->p_mtx); CURVNET_SET(inp->inp_vnet); if (drop_reason) { INP_INFO_RLOCK(&V_tcbinfo); ti_locked = TI_RLOCKED; } else { ti_locked = TI_UNLOCKED; } INP_WLOCK(inp); if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) || (inp->inp_flags2 & INP_FREED)) { out: hpts->p_inp = NULL; if (ti_locked == TI_RLOCKED) { INP_INFO_RUNLOCK(&V_tcbinfo); } if (in_pcbrele_wlocked(inp) == 0) { INP_WUNLOCK(inp); } ti_locked = TI_UNLOCKED; CURVNET_RESTORE(); mtx_lock(&hpts->p_mtx); continue; } if ((tp == NULL) || (tp->t_inpcb == NULL)) { goto out; } if (drop_reason) { /* This tcb is being destroyed for drop_reason */ m = tp->t_in_pkt; if (m) n = m->m_nextpkt; else n = NULL; tp->t_in_pkt = NULL; while (m) { m_freem(m); m = n; if (m) n = m->m_nextpkt; } tp = tcp_drop(tp, drop_reason); INP_INFO_RUNLOCK(&V_tcbinfo); if (tp == NULL) { INP_WLOCK(inp); } if (in_pcbrele_wlocked(inp) == 0) INP_WUNLOCK(inp); CURVNET_RESTORE(); mtx_lock(&hpts->p_mtx); continue; } if (set_cpu) { /* * Setup so the next time we will move to the right * CPU. This should be a rare event. It will * sometimes happens when we are the client side * (usually not the server). Somehow tcp_output() * gets called before the tcp_do_segment() sets the * intial state. This means the r_cpu and r_hpts_cpu * is 0. We get on the hpts, and then tcp_input() * gets called setting up the r_cpu to the correct * value. The hpts goes off and sees the mis-match. * We simply correct it here and the CPU will switch * to the new hpts nextime the tcb gets added to the * the hpts (not this time) :-) */ tcp_set_hpts(inp); } m = tp->t_in_pkt; n = NULL; if (m != NULL && (m->m_pkthdr.pace_lock == TI_RLOCKED || tp->t_state != TCPS_ESTABLISHED)) { ti_locked = TI_RLOCKED; - if (tcp_hptsi_lock_inpinfo(inp, &tp)) { - CURVNET_RESTORE(); - goto out; - } + INP_INFO_RLOCK(&V_tcbinfo); m = tp->t_in_pkt; } if (in_newts_every_tcb) { if (in_ts_percision) microuptime(tv); else getmicrouptime(tv); } if (tp->t_fb_ptr != NULL) { kern_prefetch(tp->t_fb_ptr, &did_prefetch); did_prefetch = 1; } /* Any input work to do, if so do it first */ if ((m != NULL) && (m == tp->t_in_pkt)) { struct tcphdr *th; int32_t tlen, drop_hdrlen, nxt_pkt; uint8_t iptos; n = m->m_nextpkt; tp->t_in_pkt = tp->t_tail_pkt = NULL; while (m) { th = (struct tcphdr *)(mtod(m, caddr_t)+m->m_pkthdr.pace_thoff); tlen = m->m_pkthdr.pace_tlen; drop_hdrlen = m->m_pkthdr.pace_drphdrlen; iptos = m->m_pkthdr.pace_tos; m->m_nextpkt = NULL; if (n) nxt_pkt = 1; else nxt_pkt = 0; inp->inp_input_calls = 1; if (tp->t_fb->tfb_tcp_hpts_do_segment) { /* Use the hpts specific do_segment */ (*tp->t_fb->tfb_tcp_hpts_do_segment) (m, th, inp->inp_socket, tp, drop_hdrlen, tlen, iptos, ti_locked, nxt_pkt, tv); } else { /* Use the default do_segment */ (*tp->t_fb->tfb_tcp_do_segment) (m, th, inp->inp_socket, tp, drop_hdrlen, tlen, iptos, ti_locked); } /* * Do segment returns unlocked we need the * lock again but we also need some kasserts * here. */ INP_INFO_UNLOCK_ASSERT(&V_tcbinfo); INP_UNLOCK_ASSERT(inp); m = n; if (m) n = m->m_nextpkt; if (m != NULL && m->m_pkthdr.pace_lock == TI_RLOCKED) { INP_INFO_RLOCK(&V_tcbinfo); ti_locked = TI_RLOCKED; } else ti_locked = TI_UNLOCKED; INP_WLOCK(inp); /* * Since we have an opening here we must * re-check if the tcb went away while we * were getting the lock(s). */ if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) || (inp->inp_flags2 & INP_FREED)) { - out_free: while (m) { m_freem(m); m = n; if (m) n = m->m_nextpkt; } goto out; } /* * Now that we hold the INP lock, check if * we need to upgrade our lock. */ if (ti_locked == TI_UNLOCKED && (tp->t_state != TCPS_ESTABLISHED)) { ti_locked = TI_RLOCKED; - if (tcp_hptsi_lock_inpinfo(inp, &tp)) - goto out_free; + INP_INFO_RLOCK(&V_tcbinfo); } } /** end while(m) */ } /** end if ((m != NULL) && (m == tp->t_in_pkt)) */ if (in_pcbrele_wlocked(inp) == 0) INP_WUNLOCK(inp); if (ti_locked == TI_RLOCKED) INP_INFO_RUNLOCK(&V_tcbinfo); INP_INFO_UNLOCK_ASSERT(&V_tcbinfo); INP_UNLOCK_ASSERT(inp); ti_locked = TI_UNLOCKED; mtx_lock(&hpts->p_mtx); hpts->p_inp = NULL; CURVNET_RESTORE(); } } static int tcp_hpts_est_run(struct tcp_hpts_entry *hpts) { int32_t ticks_to_run; if (hpts->p_prevtick && (SEQ_GT(hpts->p_curtick, hpts->p_prevtick))) { ticks_to_run = hpts->p_curtick - hpts->p_prevtick; if (ticks_to_run >= (NUM_OF_HPTSI_SLOTS - 1)) { ticks_to_run = NUM_OF_HPTSI_SLOTS - 2; } } else { if (hpts->p_prevtick == hpts->p_curtick) { /* This happens when we get woken up right away */ return (-1); } ticks_to_run = 1; } /* Set in where we will be when we catch up */ hpts->p_nxt_slot = (hpts->p_cur_slot + ticks_to_run) % NUM_OF_HPTSI_SLOTS; if (hpts->p_nxt_slot == hpts->p_cur_slot) { panic("Impossible math -- hpts:%p p_nxt_slot:%d p_cur_slot:%d ticks_to_run:%d", hpts, hpts->p_nxt_slot, hpts->p_cur_slot, ticks_to_run); } return (ticks_to_run); } static void tcp_hptsi(struct tcp_hpts_entry *hpts, struct timeval *ctick) { struct tcpcb *tp; struct inpcb *inp = NULL, *ninp; struct timeval tv; int32_t ticks_to_run, i, error, tick_now, interum_tick; int32_t paced_cnt = 0; int32_t did_prefetch = 0; int32_t prefetch_ninp = 0; int32_t prefetch_tp = 0; uint32_t cts; int16_t set_cpu; HPTS_MTX_ASSERT(hpts); hpts->p_curtick = tcp_tv_to_hptstick(ctick); cts = tcp_tv_to_usectick(ctick); memcpy(&tv, ctick, sizeof(struct timeval)); hpts->p_cur_slot = hpts_tick(hpts, 1); /* Figure out if we had missed ticks */ again: HPTS_MTX_ASSERT(hpts); ticks_to_run = tcp_hpts_est_run(hpts); if (!TAILQ_EMPTY(&hpts->p_input)) { tcp_input_data(hpts, &tv); } #ifdef INVARIANTS if (TAILQ_EMPTY(&hpts->p_input) && (hpts->p_on_inqueue_cnt != 0)) { panic("tp:%p in_hpts input empty but cnt:%d", hpts, hpts->p_on_inqueue_cnt); } #endif HPTS_MTX_ASSERT(hpts); /* Reset the ticks to run and time if we need too */ interum_tick = tcp_gethptstick(&tv); if (interum_tick != hpts->p_curtick) { /* Save off the new time we execute to */ *ctick = tv; hpts->p_curtick = interum_tick; cts = tcp_tv_to_usectick(&tv); hpts->p_cur_slot = hpts_tick(hpts, 1); ticks_to_run = tcp_hpts_est_run(hpts); } if (ticks_to_run == -1) { goto no_run; } if (logging_on) { tcp_hpts_log_it(hpts, inp, HPTSLOG_SETTORUN, ticks_to_run, 0); } if (hpts->p_on_queue_cnt == 0) { goto no_one; } HPTS_MTX_ASSERT(hpts); for (i = 0; i < ticks_to_run; i++) { /* * Calculate our delay, if there are no extra ticks there * was not any */ hpts->p_delayed_by = (ticks_to_run - (i + 1)) * HPTS_TICKS_PER_USEC; HPTS_MTX_ASSERT(hpts); while ((inp = TAILQ_FIRST(&hpts->p_hptss[hpts->p_cur_slot])) != NULL) { /* For debugging */ if (logging_on) { tcp_hpts_log_it(hpts, inp, HPTSLOG_HPTSI, ticks_to_run, i); } hpts->p_inp = inp; paced_cnt++; if (hpts->p_cur_slot != inp->inp_hptsslot) { panic("Hpts:%p inp:%p slot mis-aligned %u vs %u", hpts, inp, hpts->p_cur_slot, inp->inp_hptsslot); } /* Now pull it */ if (inp->inp_hpts_cpu_set == 0) { set_cpu = 1; } else { set_cpu = 0; } hpts_sane_pace_remove(hpts, inp, &hpts->p_hptss[hpts->p_cur_slot], 0); if ((ninp = TAILQ_FIRST(&hpts->p_hptss[hpts->p_cur_slot])) != NULL) { /* We prefetch the next inp if possible */ kern_prefetch(ninp, &prefetch_ninp); prefetch_ninp = 1; } if (inp->inp_hpts_request) { /* * This guy is deferred out further in time * then our wheel had on it. Push him back * on the wheel. */ int32_t remaining_slots; remaining_slots = ticks_to_run - (i + 1); if (inp->inp_hpts_request > remaining_slots) { /* * Keep INVARIANTS happy by clearing * the flag */ tcp_hpts_insert_locked(hpts, inp, inp->inp_hpts_request, cts, __LINE__, NULL, 1); hpts->p_inp = NULL; continue; } inp->inp_hpts_request = 0; } /* * We clear the hpts flag here after dealing with * remaining slots. This way anyone looking with the * TCB lock will see its on the hpts until just * before we unlock. */ inp->inp_in_hpts = 0; mtx_unlock(&hpts->p_mtx); INP_WLOCK(inp); if (in_pcbrele_wlocked(inp)) { mtx_lock(&hpts->p_mtx); if (logging_on) tcp_hpts_log_it(hpts, hpts->p_inp, HPTSLOG_INP_DONE, 0, 1); hpts->p_inp = NULL; continue; } if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { out_now: #ifdef INVARIANTS if (mtx_owned(&hpts->p_mtx)) { panic("Hpts:%p owns mtx prior-to lock line:%d", hpts, __LINE__); } #endif INP_WUNLOCK(inp); mtx_lock(&hpts->p_mtx); if (logging_on) tcp_hpts_log_it(hpts, hpts->p_inp, HPTSLOG_INP_DONE, 0, 3); hpts->p_inp = NULL; continue; } tp = intotcpcb(inp); if ((tp == NULL) || (tp->t_inpcb == NULL)) { goto out_now; } if (set_cpu) { /* * Setup so the next time we will move to * the right CPU. This should be a rare * event. It will sometimes happens when we * are the client side (usually not the * server). Somehow tcp_output() gets called * before the tcp_do_segment() sets the * intial state. This means the r_cpu and * r_hpts_cpu is 0. We get on the hpts, and * then tcp_input() gets called setting up * the r_cpu to the correct value. The hpts * goes off and sees the mis-match. We * simply correct it here and the CPU will * switch to the new hpts nextime the tcb * gets added to the the hpts (not this one) * :-) */ tcp_set_hpts(inp); } if (out_newts_every_tcb) { struct timeval sv; if (out_ts_percision) microuptime(&sv); else getmicrouptime(&sv); cts = tcp_tv_to_usectick(&sv); } CURVNET_SET(inp->inp_vnet); /* * There is a hole here, we get the refcnt on the * inp so it will still be preserved but to make * sure we can get the INP we need to hold the p_mtx * above while we pull out the tp/inp, as long as * fini gets the lock first we are assured of having * a sane INP we can lock and test. */ #ifdef INVARIANTS if (mtx_owned(&hpts->p_mtx)) { panic("Hpts:%p owns mtx before tcp-output:%d", hpts, __LINE__); } #endif if (tp->t_fb_ptr != NULL) { kern_prefetch(tp->t_fb_ptr, &did_prefetch); did_prefetch = 1; } inp->inp_hpts_calls = 1; if (tp->t_fb->tfb_tcp_output_wtime != NULL) { error = (*tp->t_fb->tfb_tcp_output_wtime) (tp, &tv); } else { error = tp->t_fb->tfb_tcp_output(tp); } if (ninp && ninp->inp_ppcb) { /* * If we have a nxt inp, see if we can * prefetch its ppcb. Note this may seem * "risky" since we have no locks (other * than the previous inp) and there no * assurance that ninp was not pulled while * we were processing inp and freed. If this * occured it could mean that either: * * a) Its NULL (which is fine we won't go * here) b) Its valid (which is cool we * will prefetch it) c) The inp got * freed back to the slab which was * reallocated. Then the piece of memory was * re-used and something else (not an * address) is in inp_ppcb. If that occurs * we don't crash, but take a TLB shootdown * performance hit (same as if it was NULL * and we tried to pre-fetch it). * * Considering that the likelyhood of is * quite rare we will take a risk on doing * this. If performance drops after testing * we can always take this out. NB: the * kern_prefetch on amd64 actually has * protection against a bad address now via * the DMAP_() tests. This will prevent the * TLB hit, and instead if occurs just * cause us to load cache with a useless * address (to us). */ kern_prefetch(ninp->inp_ppcb, &prefetch_tp); prefetch_tp = 1; } INP_WUNLOCK(inp); INP_UNLOCK_ASSERT(inp); CURVNET_RESTORE(); #ifdef INVARIANTS if (mtx_owned(&hpts->p_mtx)) { panic("Hpts:%p owns mtx prior-to lock line:%d", hpts, __LINE__); } #endif mtx_lock(&hpts->p_mtx); if (logging_on) tcp_hpts_log_it(hpts, hpts->p_inp, HPTSLOG_INP_DONE, 0, 4); hpts->p_inp = NULL; } HPTS_MTX_ASSERT(hpts); hpts->p_inp = NULL; hpts->p_cur_slot++; if (hpts->p_cur_slot >= NUM_OF_HPTSI_SLOTS) { hpts->p_cur_slot = 0; } } no_one: HPTS_MTX_ASSERT(hpts); hpts->p_prevtick = hpts->p_curtick; hpts->p_delayed_by = 0; /* * Check to see if we took an excess amount of time and need to run * more ticks (if we did not hit eno-bufs). */ /* Re-run any input that may be there */ (void)tcp_gethptstick(&tv); if (!TAILQ_EMPTY(&hpts->p_input)) { tcp_input_data(hpts, &tv); } #ifdef INVARIANTS if (TAILQ_EMPTY(&hpts->p_input) && (hpts->p_on_inqueue_cnt != 0)) { panic("tp:%p in_hpts input empty but cnt:%d", hpts, hpts->p_on_inqueue_cnt); } #endif tick_now = tcp_gethptstick(&tv); if (SEQ_GT(tick_now, hpts->p_prevtick)) { struct timeval res; /* Did we really spend a full tick or more in here? */ timersub(&tv, ctick, &res); if (res.tv_sec || (res.tv_usec >= HPTS_TICKS_PER_USEC)) { counter_u64_add(hpts_loops, 1); if (logging_on) { tcp_hpts_log_it(hpts, inp, HPTSLOG_TOLONG, (uint32_t) res.tv_usec, tick_now); } *ctick = res; hpts->p_curtick = tick_now; goto again; } } no_run: { uint32_t t = 0, i, fnd = 0; if (hpts->p_on_queue_cnt) { /* * Find next slot that is occupied and use that to * be the sleep time. */ for (i = 1, t = hpts->p_nxt_slot; i < NUM_OF_HPTSI_SLOTS; i++) { if (TAILQ_EMPTY(&hpts->p_hptss[t]) == 0) { fnd = 1; break; } t = (t + 1) % NUM_OF_HPTSI_SLOTS; } if (fnd) { hpts->p_hpts_sleep_time = i; } else { counter_u64_add(back_tosleep, 1); #ifdef INVARIANTS panic("Hpts:%p cnt:%d but non found", hpts, hpts->p_on_queue_cnt); #endif hpts->p_on_queue_cnt = 0; goto non_found; } t++; } else { /* No one on the wheel sleep for all but 2 slots */ non_found: if (hpts_sleep_max == 0) hpts_sleep_max = 1; hpts->p_hpts_sleep_time = min((NUM_OF_HPTSI_SLOTS - 2), hpts_sleep_max); t = 0; } if (logging_on) { tcp_hpts_log_it(hpts, inp, HPTSLOG_SLEEPSET, t, (hpts->p_hpts_sleep_time * HPTS_TICKS_PER_USEC)); } } } void __tcp_set_hpts(struct inpcb *inp, int32_t line) { struct tcp_hpts_entry *hpts; INP_WLOCK_ASSERT(inp); hpts = tcp_hpts_lock(inp); if ((inp->inp_in_hpts == 0) && (inp->inp_hpts_cpu_set == 0)) { inp->inp_hpts_cpu = hpts_cpuid(inp); inp->inp_hpts_cpu_set = 1; } mtx_unlock(&hpts->p_mtx); hpts = tcp_input_lock(inp); if ((inp->inp_input_cpu_set == 0) && (inp->inp_in_input == 0)) { inp->inp_input_cpu = hpts_cpuid(inp); inp->inp_input_cpu_set = 1; } mtx_unlock(&hpts->p_mtx); } uint16_t tcp_hpts_delayedby(struct inpcb *inp){ return (tcp_pace.rp_ent[inp->inp_hpts_cpu]->p_delayed_by); } static void tcp_hpts_thread(void *ctx) { struct tcp_hpts_entry *hpts; struct timeval tv; sbintime_t sb; hpts = (struct tcp_hpts_entry *)ctx; mtx_lock(&hpts->p_mtx); if (hpts->p_direct_wake) { /* Signaled by input */ if (logging_on) tcp_hpts_log_it(hpts, NULL, HPTSLOG_AWAKE, 1, 1); callout_stop(&hpts->co); } else { /* Timed out */ if (callout_pending(&hpts->co) || !callout_active(&hpts->co)) { if (logging_on) tcp_hpts_log_it(hpts, NULL, HPTSLOG_AWAKE, 2, 2); mtx_unlock(&hpts->p_mtx); return; } callout_deactivate(&hpts->co); if (logging_on) tcp_hpts_log_it(hpts, NULL, HPTSLOG_AWAKE, 3, 3); } hpts->p_hpts_active = 1; (void)tcp_gethptstick(&tv); tcp_hptsi(hpts, &tv); HPTS_MTX_ASSERT(hpts); tv.tv_sec = 0; tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_USEC; if (tcp_min_hptsi_time && (tv.tv_usec < tcp_min_hptsi_time)) { tv.tv_usec = tcp_min_hptsi_time; hpts->p_on_min_sleep = 1; } else { /* Clear the min sleep flag */ hpts->p_on_min_sleep = 0; } hpts->p_hpts_active = 0; sb = tvtosbt(tv); if (tcp_hpts_callout_skip_swi == 0) { callout_reset_sbt_on(&hpts->co, sb, 0, hpts_timeout_swi, hpts, hpts->p_cpu, (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision))); } else { callout_reset_sbt_on(&hpts->co, sb, 0, hpts_timeout_dir, hpts, hpts->p_cpu, C_PREL(tcp_hpts_precision)); } hpts->p_direct_wake = 0; mtx_unlock(&hpts->p_mtx); } #undef timersub static void tcp_init_hptsi(void *st) { int32_t i, j, error, bound = 0, created = 0; size_t sz, asz; struct timeval tv; sbintime_t sb; struct tcp_hpts_entry *hpts; char unit[16]; uint32_t ncpus = mp_ncpus ? mp_ncpus : MAXCPU; tcp_pace.rp_proc = NULL; tcp_pace.rp_num_hptss = ncpus; hpts_loops = counter_u64_alloc(M_WAITOK); back_tosleep = counter_u64_alloc(M_WAITOK); sz = (tcp_pace.rp_num_hptss * sizeof(struct tcp_hpts_entry *)); tcp_pace.rp_ent = malloc(sz, M_TCPHPTS, M_WAITOK | M_ZERO); asz = sizeof(struct hptsh) * NUM_OF_HPTSI_SLOTS; for (i = 0; i < tcp_pace.rp_num_hptss; i++) { tcp_pace.rp_ent[i] = malloc(sizeof(struct tcp_hpts_entry), M_TCPHPTS, M_WAITOK | M_ZERO); tcp_pace.rp_ent[i]->p_hptss = malloc(asz, M_TCPHPTS, M_WAITOK); hpts = tcp_pace.rp_ent[i]; /* * Init all the hpts structures that are not specifically * zero'd by the allocations. Also lets attach them to the * appropriate sysctl block as well. */ mtx_init(&hpts->p_mtx, "tcp_hpts_lck", "hpts", MTX_DEF | MTX_DUPOK); TAILQ_INIT(&hpts->p_input); for (j = 0; j < NUM_OF_HPTSI_SLOTS; j++) { TAILQ_INIT(&hpts->p_hptss[j]); } sysctl_ctx_init(&hpts->hpts_ctx); sprintf(unit, "%d", i); hpts->hpts_root = SYSCTL_ADD_NODE(&hpts->hpts_ctx, SYSCTL_STATIC_CHILDREN(_net_inet_tcp_hpts), OID_AUTO, unit, CTLFLAG_RW, 0, ""); SYSCTL_ADD_INT(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "in_qcnt", CTLFLAG_RD, &hpts->p_on_inqueue_cnt, 0, "Count TCB's awaiting input processing"); SYSCTL_ADD_INT(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "out_qcnt", CTLFLAG_RD, &hpts->p_on_queue_cnt, 0, "Count TCB's awaiting output processing"); SYSCTL_ADD_UINT(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "active", CTLFLAG_RD, &hpts->p_hpts_active, 0, "Is the hpts active"); SYSCTL_ADD_UINT(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "curslot", CTLFLAG_RD, &hpts->p_cur_slot, 0, "What the current slot is if active"); SYSCTL_ADD_UINT(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "curtick", CTLFLAG_RD, &hpts->p_curtick, 0, "What the current tick on if active"); SYSCTL_ADD_UINT(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "logsize", CTLFLAG_RD, &hpts->p_logsize, 0, "Hpts logging buffer size"); hpts->p_hpts_sleep_time = NUM_OF_HPTSI_SLOTS - 2; hpts->p_num = i; hpts->p_prevtick = hpts->p_curtick = tcp_gethptstick(&tv); hpts->p_prevtick -= 1; hpts->p_prevtick %= NUM_OF_HPTSI_SLOTS; hpts->p_cpu = 0xffff; hpts->p_nxt_slot = 1; hpts->p_logsize = tcp_hpts_logging_size; if (hpts->p_logsize) { sz = (sizeof(struct hpts_log) * hpts->p_logsize); hpts->p_log = malloc(sz, M_TCPHPTS, M_WAITOK | M_ZERO); } callout_init(&hpts->co, 1); } /* * Now lets start ithreads to handle the hptss. */ CPU_FOREACH(i) { hpts = tcp_pace.rp_ent[i]; hpts->p_cpu = i; error = swi_add(&hpts->ie, "hpts", tcp_hpts_thread, (void *)hpts, SWI_NET, INTR_MPSAFE, &hpts->ie_cookie); if (error) { panic("Can't add hpts:%p i:%d err:%d", hpts, i, error); } created++; if (tcp_bind_threads) { if (intr_event_bind(hpts->ie, i) == 0) bound++; } tv.tv_sec = 0; tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_USEC; sb = tvtosbt(tv); if (tcp_hpts_callout_skip_swi == 0) { callout_reset_sbt_on(&hpts->co, sb, 0, hpts_timeout_swi, hpts, hpts->p_cpu, (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision))); } else { callout_reset_sbt_on(&hpts->co, sb, 0, hpts_timeout_dir, hpts, hpts->p_cpu, C_PREL(tcp_hpts_precision)); } } printf("TCP Hpts created %d swi interrupt thread and bound %d\n", created, bound); return; } SYSINIT(tcphptsi, SI_SUB_KTHREAD_IDLE, SI_ORDER_ANY, tcp_init_hptsi, NULL); MODULE_VERSION(tcphpts, 1); Index: head/sys/netinet/tcp_input.c =================================================================== --- head/sys/netinet/tcp_input.c (revision 335355) +++ head/sys/netinet/tcp_input.c (revision 335356) @@ -1,3895 +1,3865 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995 * The Regents of the University of California. All rights reserved. * Copyright (c) 2007-2008,2010 * Swinburne University of Technology, Melbourne, Australia. * Copyright (c) 2009-2010 Lawrence Stewart * Copyright (c) 2010 The FreeBSD Foundation * Copyright (c) 2010-2011 Juniper Networks, Inc. * All rights reserved. * * Portions of this software were developed at the Centre for Advanced Internet * Architectures, Swinburne University of Technology, by Lawrence Stewart, * James Healy and David Hayes, made possible in part by a grant from the Cisco * University Research Program Fund at Community Foundation Silicon Valley. * * Portions of this software were developed at the Centre for Advanced * Internet Architectures, Swinburne University of Technology, Melbourne, * Australia by David Hayes under sponsorship from the FreeBSD Foundation. * * Portions of this software were developed by Robert N. M. Watson under * contract to Juniper Networks, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)tcp_input.c 8.12 (Berkeley) 5/24/95 */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_tcpdebug.h" #include #include #ifdef TCP_HHOOK #include #endif #include #include #include /* for proc0 declaration */ #include #include #include #include #include #include #include #include #include /* before tcp_seq.h, for tcp_random18() */ #include #include #include #include #include #define TCPSTATES /* for logging */ #include #include #include #include #include #include /* required for icmp_var.h */ #include /* for ICMP_BANDLIM */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef TCPPCAP #include #endif #include #ifdef TCPDEBUG #include #endif /* TCPDEBUG */ #ifdef TCP_OFFLOAD #include #endif #include #include #include const int tcprexmtthresh = 3; int tcp_log_in_vain = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, log_in_vain, CTLFLAG_RW, &tcp_log_in_vain, 0, "Log all incoming TCP segments to closed ports"); VNET_DEFINE(int, blackhole) = 0; #define V_blackhole VNET(blackhole) SYSCTL_INT(_net_inet_tcp, OID_AUTO, blackhole, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(blackhole), 0, "Do not send RST on segments to closed ports"); VNET_DEFINE(int, tcp_delack_enabled) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, delayed_ack, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_delack_enabled), 0, "Delay ACK to try and piggyback it onto a data packet"); VNET_DEFINE(int, drop_synfin) = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, drop_synfin, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(drop_synfin), 0, "Drop TCP packets with SYN+FIN set"); VNET_DEFINE(int, tcp_do_rfc6675_pipe) = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc6675_pipe, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_rfc6675_pipe), 0, "Use calculated pipe/in-flight bytes per RFC 6675"); VNET_DEFINE(int, tcp_do_rfc3042) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3042, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_rfc3042), 0, "Enable RFC 3042 (Limited Transmit)"); VNET_DEFINE(int, tcp_do_rfc3390) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_rfc3390), 0, "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)"); VNET_DEFINE(int, tcp_initcwnd_segments) = 10; SYSCTL_INT(_net_inet_tcp, OID_AUTO, initcwnd_segments, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_initcwnd_segments), 0, "Slow-start flight size (initial congestion window) in number of segments"); VNET_DEFINE(int, tcp_do_rfc3465) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3465, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_rfc3465), 0, "Enable RFC 3465 (Appropriate Byte Counting)"); VNET_DEFINE(int, tcp_abc_l_var) = 2; SYSCTL_INT(_net_inet_tcp, OID_AUTO, abc_l_var, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_abc_l_var), 2, "Cap the max cwnd increment during slow-start to this number of segments"); static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, ecn, CTLFLAG_RW, 0, "TCP ECN"); VNET_DEFINE(int, tcp_do_ecn) = 2; SYSCTL_INT(_net_inet_tcp_ecn, OID_AUTO, enable, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_ecn), 0, "TCP ECN support"); VNET_DEFINE(int, tcp_ecn_maxretries) = 1; SYSCTL_INT(_net_inet_tcp_ecn, OID_AUTO, maxretries, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_ecn_maxretries), 0, "Max retries before giving up on ECN"); VNET_DEFINE(int, tcp_insecure_syn) = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, insecure_syn, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_insecure_syn), 0, "Follow RFC793 instead of RFC5961 criteria for accepting SYN packets"); VNET_DEFINE(int, tcp_insecure_rst) = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, insecure_rst, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_insecure_rst), 0, "Follow RFC793 instead of RFC5961 criteria for accepting RST packets"); VNET_DEFINE(int, tcp_recvspace) = 1024*64; #define V_tcp_recvspace VNET(tcp_recvspace) SYSCTL_INT(_net_inet_tcp, TCPCTL_RECVSPACE, recvspace, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_recvspace), 0, "Initial receive socket buffer size"); VNET_DEFINE(int, tcp_do_autorcvbuf) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_auto, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_autorcvbuf), 0, "Enable automatic receive buffer sizing"); VNET_DEFINE(int, tcp_autorcvbuf_inc) = 16*1024; SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_inc, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_autorcvbuf_inc), 0, "Incrementor step size of automatic receive buffer"); VNET_DEFINE(int, tcp_autorcvbuf_max) = 2*1024*1024; SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_max, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_autorcvbuf_max), 0, "Max size of automatic receive buffer"); VNET_DEFINE(struct inpcbhead, tcb); #define tcb6 tcb /* for KAME src sync over BSD*'s */ VNET_DEFINE(struct inpcbinfo, tcbinfo); /* * TCP statistics are stored in an array of counter(9)s, which size matches * size of struct tcpstat. TCP running connection count is a regular array. */ VNET_PCPUSTAT_DEFINE(struct tcpstat, tcpstat); SYSCTL_VNET_PCPUSTAT(_net_inet_tcp, TCPCTL_STATS, stats, struct tcpstat, tcpstat, "TCP statistics (struct tcpstat, netinet/tcp_var.h)"); VNET_DEFINE(counter_u64_t, tcps_states[TCP_NSTATES]); SYSCTL_COUNTER_U64_ARRAY(_net_inet_tcp, TCPCTL_STATES, states, CTLFLAG_RD | CTLFLAG_VNET, &VNET_NAME(tcps_states)[0], TCP_NSTATES, "TCP connection counts by TCP state"); static void tcp_vnet_init(const void *unused) { COUNTER_ARRAY_ALLOC(V_tcps_states, TCP_NSTATES, M_WAITOK); VNET_PCPUSTAT_ALLOC(tcpstat, M_WAITOK); } VNET_SYSINIT(tcp_vnet_init, SI_SUB_PROTO_IFATTACHDOMAIN, SI_ORDER_ANY, tcp_vnet_init, NULL); #ifdef VIMAGE static void tcp_vnet_uninit(const void *unused) { COUNTER_ARRAY_FREE(V_tcps_states, TCP_NSTATES); VNET_PCPUSTAT_FREE(tcpstat); } VNET_SYSUNINIT(tcp_vnet_uninit, SI_SUB_PROTO_IFATTACHDOMAIN, SI_ORDER_ANY, tcp_vnet_uninit, NULL); #endif /* VIMAGE */ /* * Kernel module interface for updating tcpstat. The argument is an index * into tcpstat treated as an array. */ void kmod_tcpstat_inc(int statnum) { counter_u64_add(VNET(tcpstat)[statnum], 1); } #ifdef TCP_HHOOK /* * Wrapper for the TCP established input helper hook. */ void hhook_run_tcp_est_in(struct tcpcb *tp, struct tcphdr *th, struct tcpopt *to) { struct tcp_hhook_data hhook_data; if (V_tcp_hhh[HHOOK_TCP_EST_IN]->hhh_nhooks > 0) { hhook_data.tp = tp; hhook_data.th = th; hhook_data.to = to; hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_IN], &hhook_data, tp->osd); } } #endif /* * CC wrapper hook functions */ void cc_ack_received(struct tcpcb *tp, struct tcphdr *th, uint16_t nsegs, uint16_t type) { INP_WLOCK_ASSERT(tp->t_inpcb); tp->ccv->nsegs = nsegs; tp->ccv->bytes_this_ack = BYTES_THIS_ACK(tp, th); if (tp->snd_cwnd <= tp->snd_wnd) tp->ccv->flags |= CCF_CWND_LIMITED; else tp->ccv->flags &= ~CCF_CWND_LIMITED; if (type == CC_ACK) { if (tp->snd_cwnd > tp->snd_ssthresh) { tp->t_bytes_acked += min(tp->ccv->bytes_this_ack, nsegs * V_tcp_abc_l_var * tcp_maxseg(tp)); if (tp->t_bytes_acked >= tp->snd_cwnd) { tp->t_bytes_acked -= tp->snd_cwnd; tp->ccv->flags |= CCF_ABC_SENTAWND; } } else { tp->ccv->flags &= ~CCF_ABC_SENTAWND; tp->t_bytes_acked = 0; } } if (CC_ALGO(tp)->ack_received != NULL) { /* XXXLAS: Find a way to live without this */ tp->ccv->curack = th->th_ack; CC_ALGO(tp)->ack_received(tp->ccv, type); } } void cc_conn_init(struct tcpcb *tp) { struct hc_metrics_lite metrics; struct inpcb *inp = tp->t_inpcb; u_int maxseg; int rtt; INP_WLOCK_ASSERT(tp->t_inpcb); tcp_hc_get(&inp->inp_inc, &metrics); maxseg = tcp_maxseg(tp); if (tp->t_srtt == 0 && (rtt = metrics.rmx_rtt)) { tp->t_srtt = rtt; tp->t_rttbest = tp->t_srtt + TCP_RTT_SCALE; TCPSTAT_INC(tcps_usedrtt); if (metrics.rmx_rttvar) { tp->t_rttvar = metrics.rmx_rttvar; TCPSTAT_INC(tcps_usedrttvar); } else { /* default variation is +- 1 rtt */ tp->t_rttvar = tp->t_srtt * TCP_RTTVAR_SCALE / TCP_RTT_SCALE; } TCPT_RANGESET(tp->t_rxtcur, ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1, tp->t_rttmin, TCPTV_REXMTMAX); } if (metrics.rmx_ssthresh) { /* * There's some sort of gateway or interface * buffer limit on the path. Use this to set * the slow start threshold, but set the * threshold to no less than 2*mss. */ tp->snd_ssthresh = max(2 * maxseg, metrics.rmx_ssthresh); TCPSTAT_INC(tcps_usedssthresh); } /* * Set the initial slow-start flight size. * * RFC5681 Section 3.1 specifies the default conservative values. * RFC3390 specifies slightly more aggressive values. * RFC6928 increases it to ten segments. * Support for user specified value for initial flight size. * * If a SYN or SYN/ACK was lost and retransmitted, we have to * reduce the initial CWND to one segment as congestion is likely * requiring us to be cautious. */ if (tp->snd_cwnd == 1) tp->snd_cwnd = maxseg; /* SYN(-ACK) lost */ else if (V_tcp_initcwnd_segments) tp->snd_cwnd = min(V_tcp_initcwnd_segments * maxseg, max(2 * maxseg, V_tcp_initcwnd_segments * 1460)); else if (V_tcp_do_rfc3390) tp->snd_cwnd = min(4 * maxseg, max(2 * maxseg, 4380)); else { /* Per RFC5681 Section 3.1 */ if (maxseg > 2190) tp->snd_cwnd = 2 * maxseg; else if (maxseg > 1095) tp->snd_cwnd = 3 * maxseg; else tp->snd_cwnd = 4 * maxseg; } if (CC_ALGO(tp)->conn_init != NULL) CC_ALGO(tp)->conn_init(tp->ccv); } void inline cc_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type) { u_int maxseg; INP_WLOCK_ASSERT(tp->t_inpcb); switch(type) { case CC_NDUPACK: if (!IN_FASTRECOVERY(tp->t_flags)) { tp->snd_recover = tp->snd_max; if (tp->t_flags & TF_ECN_PERMIT) tp->t_flags |= TF_ECN_SND_CWR; } break; case CC_ECN: if (!IN_CONGRECOVERY(tp->t_flags)) { TCPSTAT_INC(tcps_ecn_rcwnd); tp->snd_recover = tp->snd_max; if (tp->t_flags & TF_ECN_PERMIT) tp->t_flags |= TF_ECN_SND_CWR; } break; case CC_RTO: maxseg = tcp_maxseg(tp); tp->t_dupacks = 0; tp->t_bytes_acked = 0; EXIT_RECOVERY(tp->t_flags); tp->snd_ssthresh = max(2, min(tp->snd_wnd, tp->snd_cwnd) / 2 / maxseg) * maxseg; tp->snd_cwnd = maxseg; break; case CC_RTO_ERR: TCPSTAT_INC(tcps_sndrexmitbad); /* RTO was unnecessary, so reset everything. */ tp->snd_cwnd = tp->snd_cwnd_prev; tp->snd_ssthresh = tp->snd_ssthresh_prev; tp->snd_recover = tp->snd_recover_prev; if (tp->t_flags & TF_WASFRECOVERY) ENTER_FASTRECOVERY(tp->t_flags); if (tp->t_flags & TF_WASCRECOVERY) ENTER_CONGRECOVERY(tp->t_flags); tp->snd_nxt = tp->snd_max; tp->t_flags &= ~TF_PREVVALID; tp->t_badrxtwin = 0; break; } if (CC_ALGO(tp)->cong_signal != NULL) { if (th != NULL) tp->ccv->curack = th->th_ack; CC_ALGO(tp)->cong_signal(tp->ccv, type); } } void inline cc_post_recovery(struct tcpcb *tp, struct tcphdr *th) { INP_WLOCK_ASSERT(tp->t_inpcb); /* XXXLAS: KASSERT that we're in recovery? */ if (CC_ALGO(tp)->post_recovery != NULL) { tp->ccv->curack = th->th_ack; CC_ALGO(tp)->post_recovery(tp->ccv); } /* XXXLAS: EXIT_RECOVERY ? */ tp->t_bytes_acked = 0; } /* * Indicate whether this ack should be delayed. We can delay the ack if * following conditions are met: * - There is no delayed ack timer in progress. * - Our last ack wasn't a 0-sized window. We never want to delay * the ack that opens up a 0-sized window. * - LRO wasn't used for this segment. We make sure by checking that the * segment size is not larger than the MSS. */ #define DELAY_ACK(tp, tlen) \ ((!tcp_timer_active(tp, TT_DELACK) && \ (tp->t_flags & TF_RXWIN0SENT) == 0) && \ (tlen <= tp->t_maxseg) && \ (V_tcp_delack_enabled || (tp->t_flags & TF_NEEDSYN))) static void inline cc_ecnpkt_handler(struct tcpcb *tp, struct tcphdr *th, uint8_t iptos) { INP_WLOCK_ASSERT(tp->t_inpcb); if (CC_ALGO(tp)->ecnpkt_handler != NULL) { switch (iptos & IPTOS_ECN_MASK) { case IPTOS_ECN_CE: tp->ccv->flags |= CCF_IPHDR_CE; break; case IPTOS_ECN_ECT0: tp->ccv->flags &= ~CCF_IPHDR_CE; break; case IPTOS_ECN_ECT1: tp->ccv->flags &= ~CCF_IPHDR_CE; break; } if (th->th_flags & TH_CWR) tp->ccv->flags |= CCF_TCPHDR_CWR; else tp->ccv->flags &= ~CCF_TCPHDR_CWR; if (tp->t_flags & TF_DELACK) tp->ccv->flags |= CCF_DELACK; else tp->ccv->flags &= ~CCF_DELACK; CC_ALGO(tp)->ecnpkt_handler(tp->ccv); if (tp->ccv->flags & CCF_ACKNOW) tcp_timer_activate(tp, TT_DELACK, tcp_delacktime); } } /* * TCP input handling is split into multiple parts: * tcp6_input is a thin wrapper around tcp_input for the extended * ip6_protox[] call format in ip6_input * tcp_input handles primary segment validation, inpcb lookup and * SYN processing on listen sockets * tcp_do_segment processes the ACK and text of the segment for * establishing, established and closing connections */ #ifdef INET6 int tcp6_input(struct mbuf **mp, int *offp, int proto) { struct mbuf *m = *mp; struct in6_ifaddr *ia6; struct ip6_hdr *ip6; IP6_EXTHDR_CHECK(m, *offp, sizeof(struct tcphdr), IPPROTO_DONE); /* * draft-itojun-ipv6-tcp-to-anycast * better place to put this in? */ ip6 = mtod(m, struct ip6_hdr *); ia6 = in6ifa_ifwithaddr(&ip6->ip6_dst, 0 /* XXX */); if (ia6 && (ia6->ia6_flags & IN6_IFF_ANYCAST)) { struct ip6_hdr *ip6; ifa_free(&ia6->ia_ifa); ip6 = mtod(m, struct ip6_hdr *); icmp6_error(m, ICMP6_DST_UNREACH, ICMP6_DST_UNREACH_ADDR, (caddr_t)&ip6->ip6_dst - (caddr_t)ip6); return (IPPROTO_DONE); } if (ia6) ifa_free(&ia6->ia_ifa); return (tcp_input(mp, offp, proto)); } #endif /* INET6 */ int tcp_input(struct mbuf **mp, int *offp, int proto) { struct mbuf *m = *mp; struct tcphdr *th = NULL; struct ip *ip = NULL; struct inpcb *inp = NULL; struct tcpcb *tp = NULL; struct socket *so = NULL; u_char *optp = NULL; int off0; int optlen = 0; #ifdef INET int len; #endif int tlen = 0, off; int drop_hdrlen; int thflags; int rstreason = 0; /* For badport_bandlim accounting purposes */ uint8_t iptos; struct m_tag *fwd_tag = NULL; #ifdef INET6 struct ip6_hdr *ip6 = NULL; int isipv6; #else const void *ip6 = NULL; #endif /* INET6 */ struct tcpopt to; /* options in this segment */ char *s = NULL; /* address and port logging */ int ti_locked; #ifdef TCPDEBUG /* * The size of tcp_saveipgen must be the size of the max ip header, * now IPv6. */ u_char tcp_saveipgen[IP6_HDR_LEN]; struct tcphdr tcp_savetcp; short ostate = 0; #endif #ifdef INET6 isipv6 = (mtod(m, struct ip *)->ip_v == 6) ? 1 : 0; #endif off0 = *offp; m = *mp; *mp = NULL; to.to_flags = 0; TCPSTAT_INC(tcps_rcvtotal); #ifdef INET6 if (isipv6) { /* IP6_EXTHDR_CHECK() is already done at tcp6_input(). */ if (m->m_len < (sizeof(*ip6) + sizeof(*th))) { m = m_pullup(m, sizeof(*ip6) + sizeof(*th)); if (m == NULL) { TCPSTAT_INC(tcps_rcvshort); return (IPPROTO_DONE); } } ip6 = mtod(m, struct ip6_hdr *); th = (struct tcphdr *)((caddr_t)ip6 + off0); tlen = sizeof(*ip6) + ntohs(ip6->ip6_plen) - off0; if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID_IPV6) { if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) th->th_sum = m->m_pkthdr.csum_data; else th->th_sum = in6_cksum_pseudo(ip6, tlen, IPPROTO_TCP, m->m_pkthdr.csum_data); th->th_sum ^= 0xffff; } else th->th_sum = in6_cksum(m, IPPROTO_TCP, off0, tlen); if (th->th_sum) { TCPSTAT_INC(tcps_rcvbadsum); goto drop; } /* * Be proactive about unspecified IPv6 address in source. * As we use all-zero to indicate unbounded/unconnected pcb, * unspecified IPv6 address can be used to confuse us. * * Note that packets with unspecified IPv6 destination is * already dropped in ip6_input. */ if (IN6_IS_ADDR_UNSPECIFIED(&ip6->ip6_src)) { /* XXX stat */ goto drop; } iptos = (ntohl(ip6->ip6_flow) >> 20) & 0xff; } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { /* * Get IP and TCP header together in first mbuf. * Note: IP leaves IP header in first mbuf. */ if (off0 > sizeof (struct ip)) { ip_stripoptions(m); off0 = sizeof(struct ip); } if (m->m_len < sizeof (struct tcpiphdr)) { if ((m = m_pullup(m, sizeof (struct tcpiphdr))) == NULL) { TCPSTAT_INC(tcps_rcvshort); return (IPPROTO_DONE); } } ip = mtod(m, struct ip *); th = (struct tcphdr *)((caddr_t)ip + off0); tlen = ntohs(ip->ip_len) - off0; iptos = ip->ip_tos; if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) { if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) th->th_sum = m->m_pkthdr.csum_data; else th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htonl(m->m_pkthdr.csum_data + tlen + IPPROTO_TCP)); th->th_sum ^= 0xffff; } else { struct ipovly *ipov = (struct ipovly *)ip; /* * Checksum extended TCP header and data. */ len = off0 + tlen; bzero(ipov->ih_x1, sizeof(ipov->ih_x1)); ipov->ih_len = htons(tlen); th->th_sum = in_cksum(m, len); /* Reset length for SDT probes. */ ip->ip_len = htons(len); /* Reset TOS bits */ ip->ip_tos = iptos; /* Re-initialization for later version check */ ip->ip_v = IPVERSION; ip->ip_hl = off0 >> 2; } if (th->th_sum) { TCPSTAT_INC(tcps_rcvbadsum); goto drop; } } #endif /* INET */ /* * Check that TCP offset makes sense, * pull out TCP options and adjust length. XXX */ off = th->th_off << 2; if (off < sizeof (struct tcphdr) || off > tlen) { TCPSTAT_INC(tcps_rcvbadoff); goto drop; } tlen -= off; /* tlen is used instead of ti->ti_len */ if (off > sizeof (struct tcphdr)) { #ifdef INET6 if (isipv6) { IP6_EXTHDR_CHECK(m, off0, off, IPPROTO_DONE); ip6 = mtod(m, struct ip6_hdr *); th = (struct tcphdr *)((caddr_t)ip6 + off0); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { if (m->m_len < sizeof(struct ip) + off) { if ((m = m_pullup(m, sizeof (struct ip) + off)) == NULL) { TCPSTAT_INC(tcps_rcvshort); return (IPPROTO_DONE); } ip = mtod(m, struct ip *); th = (struct tcphdr *)((caddr_t)ip + off0); } } #endif optlen = off - sizeof (struct tcphdr); optp = (u_char *)(th + 1); } thflags = th->th_flags; /* * Convert TCP protocol specific fields to host format. */ tcp_fields_to_host(th); /* * Delay dropping TCP, IP headers, IPv6 ext headers, and TCP options. */ drop_hdrlen = off0 + off; /* * Locate pcb for segment; if we're likely to add or remove a * connection then first acquire pcbinfo lock. There are three cases * where we might discover later we need a write lock despite the * flags: ACKs moving a connection out of the syncache, ACKs for a * connection in TIMEWAIT and SYNs not targeting a listening socket. */ if ((thflags & (TH_FIN | TH_RST)) != 0) { INP_INFO_RLOCK(&V_tcbinfo); ti_locked = TI_RLOCKED; } else ti_locked = TI_UNLOCKED; /* * Grab info from PACKET_TAG_IPFORWARD tag prepended to the chain. */ if ( #ifdef INET6 (isipv6 && (m->m_flags & M_IP6_NEXTHOP)) #ifdef INET || (!isipv6 && (m->m_flags & M_IP_NEXTHOP)) #endif #endif #if defined(INET) && !defined(INET6) (m->m_flags & M_IP_NEXTHOP) #endif ) fwd_tag = m_tag_find(m, PACKET_TAG_IPFORWARD, NULL); findpcb: #ifdef INVARIANTS if (ti_locked == TI_RLOCKED) { INP_INFO_RLOCK_ASSERT(&V_tcbinfo); } else { INP_INFO_UNLOCK_ASSERT(&V_tcbinfo); } #endif #ifdef INET6 if (isipv6 && fwd_tag != NULL) { struct sockaddr_in6 *next_hop6; next_hop6 = (struct sockaddr_in6 *)(fwd_tag + 1); /* * Transparently forwarded. Pretend to be the destination. * Already got one like this? */ inp = in6_pcblookup_mbuf(&V_tcbinfo, &ip6->ip6_src, th->th_sport, &ip6->ip6_dst, th->th_dport, INPLOOKUP_WLOCKPCB, m->m_pkthdr.rcvif, m); if (!inp) { /* * It's new. Try to find the ambushing socket. * Because we've rewritten the destination address, * any hardware-generated hash is ignored. */ inp = in6_pcblookup(&V_tcbinfo, &ip6->ip6_src, th->th_sport, &next_hop6->sin6_addr, next_hop6->sin6_port ? ntohs(next_hop6->sin6_port) : th->th_dport, INPLOOKUP_WILDCARD | INPLOOKUP_WLOCKPCB, m->m_pkthdr.rcvif); } } else if (isipv6) { inp = in6_pcblookup_mbuf(&V_tcbinfo, &ip6->ip6_src, th->th_sport, &ip6->ip6_dst, th->th_dport, INPLOOKUP_WILDCARD | INPLOOKUP_WLOCKPCB, m->m_pkthdr.rcvif, m); } #endif /* INET6 */ #if defined(INET6) && defined(INET) else #endif #ifdef INET if (fwd_tag != NULL) { struct sockaddr_in *next_hop; next_hop = (struct sockaddr_in *)(fwd_tag+1); /* * Transparently forwarded. Pretend to be the destination. * already got one like this? */ inp = in_pcblookup_mbuf(&V_tcbinfo, ip->ip_src, th->th_sport, ip->ip_dst, th->th_dport, INPLOOKUP_WLOCKPCB, m->m_pkthdr.rcvif, m); if (!inp) { /* * It's new. Try to find the ambushing socket. * Because we've rewritten the destination address, * any hardware-generated hash is ignored. */ inp = in_pcblookup(&V_tcbinfo, ip->ip_src, th->th_sport, next_hop->sin_addr, next_hop->sin_port ? ntohs(next_hop->sin_port) : th->th_dport, INPLOOKUP_WILDCARD | INPLOOKUP_WLOCKPCB, m->m_pkthdr.rcvif); } } else inp = in_pcblookup_mbuf(&V_tcbinfo, ip->ip_src, th->th_sport, ip->ip_dst, th->th_dport, INPLOOKUP_WILDCARD | INPLOOKUP_WLOCKPCB, m->m_pkthdr.rcvif, m); #endif /* INET */ /* * If the INPCB does not exist then all data in the incoming * segment is discarded and an appropriate RST is sent back. * XXX MRT Send RST using which routing table? */ if (inp == NULL) { /* * Log communication attempts to ports that are not * in use. */ if ((tcp_log_in_vain == 1 && (thflags & TH_SYN)) || tcp_log_in_vain == 2) { if ((s = tcp_log_vain(NULL, th, (void *)ip, ip6))) log(LOG_INFO, "%s; %s: Connection attempt " "to closed port\n", s, __func__); } /* * When blackholing do not respond with a RST but * completely ignore the segment and drop it. */ if ((V_blackhole == 1 && (thflags & TH_SYN)) || V_blackhole == 2) goto dropunlock; rstreason = BANDLIM_RST_CLOSEDPORT; goto dropwithreset; } INP_WLOCK_ASSERT(inp); /* * While waiting for inp lock during the lookup, another thread * can have dropped the inpcb, in which case we need to loop back * and try to find a new inpcb to deliver to. */ if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); inp = NULL; goto findpcb; } if ((inp->inp_flowtype == M_HASHTYPE_NONE) && (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) && ((inp->inp_socket == NULL) || (inp->inp_socket->so_options & SO_ACCEPTCONN) == 0)) { inp->inp_flowid = m->m_pkthdr.flowid; inp->inp_flowtype = M_HASHTYPE_GET(m); } #if defined(IPSEC) || defined(IPSEC_SUPPORT) #ifdef INET6 if (isipv6 && IPSEC_ENABLED(ipv6) && IPSEC_CHECK_POLICY(ipv6, m, inp) != 0) { goto dropunlock; } #ifdef INET else #endif #endif /* INET6 */ #ifdef INET if (IPSEC_ENABLED(ipv4) && IPSEC_CHECK_POLICY(ipv4, m, inp) != 0) { goto dropunlock; } #endif /* INET */ #endif /* IPSEC */ /* * Check the minimum TTL for socket. */ if (inp->inp_ip_minttl != 0) { #ifdef INET6 if (isipv6) { if (inp->inp_ip_minttl > ip6->ip6_hlim) goto dropunlock; } else #endif if (inp->inp_ip_minttl > ip->ip_ttl) goto dropunlock; } /* * A previous connection in TIMEWAIT state is supposed to catch stray * or duplicate segments arriving late. If this segment was a * legitimate new connection attempt, the old INPCB gets removed and * we can try again to find a listening socket. * * At this point, due to earlier optimism, we may hold only an inpcb * lock, and not the inpcbinfo write lock. If so, we need to try to * acquire it, or if that fails, acquire a reference on the inpcb, * drop all locks, acquire a global write lock, and then re-acquire * the inpcb lock. We may at that point discover that another thread * has tried to free the inpcb, in which case we need to loop back * and try to find a new inpcb to deliver to. * * XXXRW: It may be time to rethink timewait locking. */ -relocked: if (inp->inp_flags & INP_TIMEWAIT) { if (ti_locked == TI_UNLOCKED) { - if (INP_INFO_TRY_RLOCK(&V_tcbinfo) == 0) { - in_pcbref(inp); - INP_WUNLOCK(inp); - INP_INFO_RLOCK(&V_tcbinfo); - ti_locked = TI_RLOCKED; - INP_WLOCK(inp); - if (in_pcbrele_wlocked(inp)) { - inp = NULL; - goto findpcb; - } else if (inp->inp_flags & INP_DROPPED) { - INP_WUNLOCK(inp); - inp = NULL; - goto findpcb; - } - } else - ti_locked = TI_RLOCKED; + INP_INFO_RLOCK(); + ti_locked = TI_RLOCKED; } INP_INFO_RLOCK_ASSERT(&V_tcbinfo); if (thflags & TH_SYN) tcp_dooptions(&to, optp, optlen, TO_SYN); /* * NB: tcp_twcheck unlocks the INP and frees the mbuf. */ if (tcp_twcheck(inp, &to, th, m, tlen)) goto findpcb; INP_INFO_RUNLOCK(&V_tcbinfo); return (IPPROTO_DONE); } /* * The TCPCB may no longer exist if the connection is winding * down or it is in the CLOSED state. Either way we drop the * segment and send an appropriate response. */ tp = intotcpcb(inp); if (tp == NULL || tp->t_state == TCPS_CLOSED) { rstreason = BANDLIM_RST_CLOSEDPORT; goto dropwithreset; } #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) { tcp_offload_input(tp, m); m = NULL; /* consumed by the TOE driver */ goto dropunlock; } #endif /* * We've identified a valid inpcb, but it could be that we need an * inpcbinfo write lock but don't hold it. In this case, attempt to * acquire using the same strategy as the TIMEWAIT case above. If we * relock, we have to jump back to 'relocked' as the connection might * now be in TIMEWAIT. */ #ifdef INVARIANTS if ((thflags & (TH_FIN | TH_RST)) != 0) INP_INFO_RLOCK_ASSERT(&V_tcbinfo); #endif if (!((tp->t_state == TCPS_ESTABLISHED && (thflags & TH_SYN) == 0) || (tp->t_state == TCPS_LISTEN && (thflags & TH_SYN) && !IS_FASTOPEN(tp->t_flags)))) { if (ti_locked == TI_UNLOCKED) { - if (INP_INFO_TRY_RLOCK(&V_tcbinfo) == 0) { - in_pcbref(inp); - INP_WUNLOCK(inp); - INP_INFO_RLOCK(&V_tcbinfo); - ti_locked = TI_RLOCKED; - INP_WLOCK(inp); - if (in_pcbrele_wlocked(inp)) { - inp = NULL; - goto findpcb; - } else if (inp->inp_flags & INP_DROPPED) { - INP_WUNLOCK(inp); - inp = NULL; - goto findpcb; - } - goto relocked; - } else - ti_locked = TI_RLOCKED; + INP_INFO_RLOCK(); + ti_locked = TI_RLOCKED; } INP_INFO_RLOCK_ASSERT(&V_tcbinfo); } #ifdef MAC INP_WLOCK_ASSERT(inp); if (mac_inpcb_check_deliver(inp, m)) goto dropunlock; #endif so = inp->inp_socket; KASSERT(so != NULL, ("%s: so == NULL", __func__)); #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) { ostate = tp->t_state; #ifdef INET6 if (isipv6) { bcopy((char *)ip6, (char *)tcp_saveipgen, sizeof(*ip6)); } else #endif bcopy((char *)ip, (char *)tcp_saveipgen, sizeof(*ip)); tcp_savetcp = *th; } #endif /* TCPDEBUG */ /* * When the socket is accepting connections (the INPCB is in LISTEN * state) we look into the SYN cache if this is a new connection * attempt or the completion of a previous one. */ KASSERT(tp->t_state == TCPS_LISTEN || !(so->so_options & SO_ACCEPTCONN), ("%s: so accepting but tp %p not listening", __func__, tp)); if (tp->t_state == TCPS_LISTEN && (so->so_options & SO_ACCEPTCONN)) { struct in_conninfo inc; bzero(&inc, sizeof(inc)); #ifdef INET6 if (isipv6) { inc.inc_flags |= INC_ISIPV6; inc.inc6_faddr = ip6->ip6_src; inc.inc6_laddr = ip6->ip6_dst; } else #endif { inc.inc_faddr = ip->ip_src; inc.inc_laddr = ip->ip_dst; } inc.inc_fport = th->th_sport; inc.inc_lport = th->th_dport; inc.inc_fibnum = so->so_fibnum; /* * Check for an existing connection attempt in syncache if * the flag is only ACK. A successful lookup creates a new * socket appended to the listen queue in SYN_RECEIVED state. */ if ((thflags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK) { INP_INFO_RLOCK_ASSERT(&V_tcbinfo); /* * Parse the TCP options here because * syncookies need access to the reflected * timestamp. */ tcp_dooptions(&to, optp, optlen, 0); /* * NB: syncache_expand() doesn't unlock * inp and tcpinfo locks. */ rstreason = syncache_expand(&inc, &to, th, &so, m); if (rstreason < 0) { /* * A failing TCP MD5 signature comparison * must result in the segment being dropped * and must not produce any response back * to the sender. */ goto dropunlock; } else if (rstreason == 0) { /* * No syncache entry or ACK was not * for our SYN/ACK. Send a RST. * NB: syncache did its own logging * of the failure cause. */ rstreason = BANDLIM_RST_OPENPORT; goto dropwithreset; } tfo_socket_result: if (so == NULL) { /* * We completed the 3-way handshake * but could not allocate a socket * either due to memory shortage, * listen queue length limits or * global socket limits. Send RST * or wait and have the remote end * retransmit the ACK for another * try. */ if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "Socket allocation failed due to " "limits or memory shortage, %s\n", s, __func__, V_tcp_sc_rst_sock_fail ? "sending RST" : "try again"); if (V_tcp_sc_rst_sock_fail) { rstreason = BANDLIM_UNLIMITED; goto dropwithreset; } else goto dropunlock; } /* * Socket is created in state SYN_RECEIVED. * Unlock the listen socket, lock the newly * created socket and update the tp variable. */ INP_WUNLOCK(inp); /* listen socket */ inp = sotoinpcb(so); /* * New connection inpcb is already locked by * syncache_expand(). */ INP_WLOCK_ASSERT(inp); tp = intotcpcb(inp); KASSERT(tp->t_state == TCPS_SYN_RECEIVED, ("%s: ", __func__)); /* * Process the segment and the data it * contains. tcp_do_segment() consumes * the mbuf chain and unlocks the inpcb. */ tp->t_fb->tfb_tcp_do_segment(m, th, so, tp, drop_hdrlen, tlen, iptos, ti_locked); INP_INFO_UNLOCK_ASSERT(&V_tcbinfo); return (IPPROTO_DONE); } /* * Segment flag validation for new connection attempts: * * Our (SYN|ACK) response was rejected. * Check with syncache and remove entry to prevent * retransmits. * * NB: syncache_chkrst does its own logging of failure * causes. */ if (thflags & TH_RST) { syncache_chkrst(&inc, th); goto dropunlock; } /* * We can't do anything without SYN. */ if ((thflags & TH_SYN) == 0) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "SYN is missing, segment ignored\n", s, __func__); TCPSTAT_INC(tcps_badsyn); goto dropunlock; } /* * (SYN|ACK) is bogus on a listen socket. */ if (thflags & TH_ACK) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "SYN|ACK invalid, segment rejected\n", s, __func__); syncache_badack(&inc); /* XXX: Not needed! */ TCPSTAT_INC(tcps_badsyn); rstreason = BANDLIM_RST_OPENPORT; goto dropwithreset; } /* * If the drop_synfin option is enabled, drop all * segments with both the SYN and FIN bits set. * This prevents e.g. nmap from identifying the * TCP/IP stack. * XXX: Poor reasoning. nmap has other methods * and is constantly refining its stack detection * strategies. * XXX: This is a violation of the TCP specification * and was used by RFC1644. */ if ((thflags & TH_FIN) && V_drop_synfin) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "SYN|FIN segment ignored (based on " "sysctl setting)\n", s, __func__); TCPSTAT_INC(tcps_badsyn); goto dropunlock; } /* * Segment's flags are (SYN) or (SYN|FIN). * * TH_PUSH, TH_URG, TH_ECE, TH_CWR are ignored * as they do not affect the state of the TCP FSM. * The data pointed to by TH_URG and th_urp is ignored. */ KASSERT((thflags & (TH_RST|TH_ACK)) == 0, ("%s: Listen socket: TH_RST or TH_ACK set", __func__)); KASSERT(thflags & (TH_SYN), ("%s: Listen socket: TH_SYN not set", __func__)); #ifdef INET6 /* * If deprecated address is forbidden, * we do not accept SYN to deprecated interface * address to prevent any new inbound connection from * getting established. * When we do not accept SYN, we send a TCP RST, * with deprecated source address (instead of dropping * it). We compromise it as it is much better for peer * to send a RST, and RST will be the final packet * for the exchange. * * If we do not forbid deprecated addresses, we accept * the SYN packet. RFC2462 does not suggest dropping * SYN in this case. * If we decipher RFC2462 5.5.4, it says like this: * 1. use of deprecated addr with existing * communication is okay - "SHOULD continue to be * used" * 2. use of it with new communication: * (2a) "SHOULD NOT be used if alternate address * with sufficient scope is available" * (2b) nothing mentioned otherwise. * Here we fall into (2b) case as we have no choice in * our source address selection - we must obey the peer. * * The wording in RFC2462 is confusing, and there are * multiple description text for deprecated address * handling - worse, they are not exactly the same. * I believe 5.5.4 is the best one, so we follow 5.5.4. */ if (isipv6 && !V_ip6_use_deprecated) { struct in6_ifaddr *ia6; ia6 = in6ifa_ifwithaddr(&ip6->ip6_dst, 0 /* XXX */); if (ia6 != NULL && (ia6->ia6_flags & IN6_IFF_DEPRECATED)) { ifa_free(&ia6->ia_ifa); if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "Connection attempt to deprecated " "IPv6 address rejected\n", s, __func__); rstreason = BANDLIM_RST_OPENPORT; goto dropwithreset; } if (ia6) ifa_free(&ia6->ia_ifa); } #endif /* INET6 */ /* * Basic sanity checks on incoming SYN requests: * Don't respond if the destination is a link layer * broadcast according to RFC1122 4.2.3.10, p. 104. * If it is from this socket it must be forged. * Don't respond if the source or destination is a * global or subnet broad- or multicast address. * Note that it is quite possible to receive unicast * link-layer packets with a broadcast IP address. Use * in_broadcast() to find them. */ if (m->m_flags & (M_BCAST|M_MCAST)) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "Connection attempt from broad- or multicast " "link layer address ignored\n", s, __func__); goto dropunlock; } #ifdef INET6 if (isipv6) { if (th->th_dport == th->th_sport && IN6_ARE_ADDR_EQUAL(&ip6->ip6_dst, &ip6->ip6_src)) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "Connection attempt to/from self " "ignored\n", s, __func__); goto dropunlock; } if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) || IN6_IS_ADDR_MULTICAST(&ip6->ip6_src)) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "Connection attempt from/to multicast " "address ignored\n", s, __func__); goto dropunlock; } } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { if (th->th_dport == th->th_sport && ip->ip_dst.s_addr == ip->ip_src.s_addr) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "Connection attempt from/to self " "ignored\n", s, __func__); goto dropunlock; } if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) || IN_MULTICAST(ntohl(ip->ip_src.s_addr)) || ip->ip_src.s_addr == htonl(INADDR_BROADCAST) || in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif)) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "Connection attempt from/to broad- " "or multicast address ignored\n", s, __func__); goto dropunlock; } } #endif /* * SYN appears to be valid. Create compressed TCP state * for syncache. */ #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif TCP_PROBE3(debug__input, tp, th, m); tcp_dooptions(&to, optp, optlen, TO_SYN); if (syncache_add(&inc, &to, th, inp, &so, m, NULL, NULL)) goto tfo_socket_result; /* * Entry added to syncache and mbuf consumed. * Only the listen socket is unlocked by syncache_add(). */ if (ti_locked == TI_RLOCKED) { INP_INFO_RUNLOCK(&V_tcbinfo); ti_locked = TI_UNLOCKED; } INP_INFO_UNLOCK_ASSERT(&V_tcbinfo); return (IPPROTO_DONE); } else if (tp->t_state == TCPS_LISTEN) { /* * When a listen socket is torn down the SO_ACCEPTCONN * flag is removed first while connections are drained * from the accept queue in a unlock/lock cycle of the * ACCEPT_LOCK, opening a race condition allowing a SYN * attempt go through unhandled. */ goto dropunlock; } #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (tp->t_flags & TF_SIGNATURE) { tcp_dooptions(&to, optp, optlen, thflags); if ((to.to_flags & TOF_SIGNATURE) == 0) { TCPSTAT_INC(tcps_sig_err_nosigopt); goto dropunlock; } if (!TCPMD5_ENABLED() || TCPMD5_INPUT(m, th, to.to_signature) != 0) goto dropunlock; } #endif TCP_PROBE5(receive, NULL, tp, m, tp, th); /* * Segment belongs to a connection in SYN_SENT, ESTABLISHED or later * state. tcp_do_segment() always consumes the mbuf chain, unlocks * the inpcb, and unlocks pcbinfo. */ tp->t_fb->tfb_tcp_do_segment(m, th, so, tp, drop_hdrlen, tlen, iptos, ti_locked); INP_INFO_UNLOCK_ASSERT(&V_tcbinfo); return (IPPROTO_DONE); dropwithreset: TCP_PROBE5(receive, NULL, tp, m, tp, th); if (ti_locked == TI_RLOCKED) { INP_INFO_RUNLOCK(&V_tcbinfo); ti_locked = TI_UNLOCKED; } #ifdef INVARIANTS else { KASSERT(ti_locked == TI_UNLOCKED, ("%s: dropwithreset " "ti_locked: %d", __func__, ti_locked)); INP_INFO_UNLOCK_ASSERT(&V_tcbinfo); } #endif if (inp != NULL) { tcp_dropwithreset(m, th, tp, tlen, rstreason); INP_WUNLOCK(inp); } else tcp_dropwithreset(m, th, NULL, tlen, rstreason); m = NULL; /* mbuf chain got consumed. */ goto drop; dropunlock: if (m != NULL) TCP_PROBE5(receive, NULL, tp, m, tp, th); if (ti_locked == TI_RLOCKED) { INP_INFO_RUNLOCK(&V_tcbinfo); ti_locked = TI_UNLOCKED; } #ifdef INVARIANTS else { KASSERT(ti_locked == TI_UNLOCKED, ("%s: dropunlock " "ti_locked: %d", __func__, ti_locked)); INP_INFO_UNLOCK_ASSERT(&V_tcbinfo); } #endif if (inp != NULL) INP_WUNLOCK(inp); drop: INP_INFO_UNLOCK_ASSERT(&V_tcbinfo); if (s != NULL) free(s, M_TCPLOG); if (m != NULL) m_freem(m); return (IPPROTO_DONE); } /* * Automatic sizing of receive socket buffer. Often the send * buffer size is not optimally adjusted to the actual network * conditions at hand (delay bandwidth product). Setting the * buffer size too small limits throughput on links with high * bandwidth and high delay (eg. trans-continental/oceanic links). * * On the receive side the socket buffer memory is only rarely * used to any significant extent. This allows us to be much * more aggressive in scaling the receive socket buffer. For * the case that the buffer space is actually used to a large * extent and we run out of kernel memory we can simply drop * the new segments; TCP on the sender will just retransmit it * later. Setting the buffer size too big may only consume too * much kernel memory if the application doesn't read() from * the socket or packet loss or reordering makes use of the * reassembly queue. * * The criteria to step up the receive buffer one notch are: * 1. Application has not set receive buffer size with * SO_RCVBUF. Setting SO_RCVBUF clears SB_AUTOSIZE. * 2. the number of bytes received during the time it takes * one timestamp to be reflected back to us (the RTT); * 3. received bytes per RTT is within seven eighth of the * current socket buffer size; * 4. receive buffer size has not hit maximal automatic size; * * This algorithm does one step per RTT at most and only if * we receive a bulk stream w/o packet losses or reorderings. * Shrinking the buffer during idle times is not necessary as * it doesn't consume any memory when idle. * * TODO: Only step up if the application is actually serving * the buffer to better manage the socket buffer resources. */ int tcp_autorcvbuf(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int tlen) { int newsize = 0; if (V_tcp_do_autorcvbuf && (so->so_rcv.sb_flags & SB_AUTOSIZE) && tp->t_srtt != 0 && tp->rfbuf_ts != 0 && TCP_TS_TO_TICKS(tcp_ts_getticks() - tp->rfbuf_ts) > (tp->t_srtt >> TCP_RTT_SHIFT)) { if (tp->rfbuf_cnt > (so->so_rcv.sb_hiwat / 8 * 7) && so->so_rcv.sb_hiwat < V_tcp_autorcvbuf_max) { newsize = min(so->so_rcv.sb_hiwat + V_tcp_autorcvbuf_inc, V_tcp_autorcvbuf_max); } TCP_PROBE6(receive__autoresize, NULL, tp, m, tp, th, newsize); /* Start over with next RTT. */ tp->rfbuf_ts = 0; tp->rfbuf_cnt = 0; } else { tp->rfbuf_cnt += tlen; /* add up */ } return (newsize); } void tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int drop_hdrlen, int tlen, uint8_t iptos, int ti_locked) { int thflags, acked, ourfinisacked, needoutput = 0, sack_changed; int rstreason, todrop, win; uint32_t tiwin; uint16_t nsegs; char *s; struct in_conninfo *inc; struct mbuf *mfree; struct tcpopt to; int tfo_syn; #ifdef TCPDEBUG /* * The size of tcp_saveipgen must be the size of the max ip header, * now IPv6. */ u_char tcp_saveipgen[IP6_HDR_LEN]; struct tcphdr tcp_savetcp; short ostate = 0; #endif thflags = th->th_flags; inc = &tp->t_inpcb->inp_inc; tp->sackhint.last_sack_ack = 0; sack_changed = 0; nsegs = max(1, m->m_pkthdr.lro_nsegs); /* * If this is either a state-changing packet or current state isn't * established, we require a write lock on tcbinfo. Otherwise, we * allow the tcbinfo to be in either alocked or unlocked, as the * caller may have unnecessarily acquired a write lock due to a race. */ if ((thflags & (TH_SYN | TH_FIN | TH_RST)) != 0 || tp->t_state != TCPS_ESTABLISHED) { KASSERT(ti_locked == TI_RLOCKED, ("%s ti_locked %d for " "SYN/FIN/RST/!EST", __func__, ti_locked)); INP_INFO_RLOCK_ASSERT(&V_tcbinfo); } else { #ifdef INVARIANTS if (ti_locked == TI_RLOCKED) INP_INFO_RLOCK_ASSERT(&V_tcbinfo); else { KASSERT(ti_locked == TI_UNLOCKED, ("%s: EST " "ti_locked: %d", __func__, ti_locked)); INP_INFO_UNLOCK_ASSERT(&V_tcbinfo); } #endif } INP_WLOCK_ASSERT(tp->t_inpcb); KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", __func__)); KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT", __func__)); #ifdef TCPPCAP /* Save segment, if requested. */ tcp_pcap_add(th, m, &(tp->t_inpkts)); #endif TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0, tlen, NULL, true); if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) { if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: " "SYN|FIN segment ignored (based on " "sysctl setting)\n", s, __func__); free(s, M_TCPLOG); } goto drop; } /* * If a segment with the ACK-bit set arrives in the SYN-SENT state * check SEQ.ACK first. */ if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { rstreason = BANDLIM_UNLIMITED; goto dropwithreset; } /* * Segment received on connection. * Reset idle time and keep-alive timer. * XXX: This should be done after segment * validation to ignore broken/spoofed segs. */ tp->t_rcvtime = ticks; /* * Scale up the window into a 32-bit value. * For the SYN_SENT state the scale is zero. */ tiwin = th->th_win << tp->snd_scale; /* * TCP ECN processing. */ if (tp->t_flags & TF_ECN_PERMIT) { if (thflags & TH_CWR) tp->t_flags &= ~TF_ECN_SND_ECE; switch (iptos & IPTOS_ECN_MASK) { case IPTOS_ECN_CE: tp->t_flags |= TF_ECN_SND_ECE; TCPSTAT_INC(tcps_ecn_ce); break; case IPTOS_ECN_ECT0: TCPSTAT_INC(tcps_ecn_ect0); break; case IPTOS_ECN_ECT1: TCPSTAT_INC(tcps_ecn_ect1); break; } /* Process a packet differently from RFC3168. */ cc_ecnpkt_handler(tp, th, iptos); /* Congestion experienced. */ if (thflags & TH_ECE) { cc_cong_signal(tp, th, CC_ECN); } } /* * Parse options on any incoming segment. */ tcp_dooptions(&to, (u_char *)(th + 1), (th->th_off << 2) - sizeof(struct tcphdr), (thflags & TH_SYN) ? TO_SYN : 0); #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if ((tp->t_flags & TF_SIGNATURE) != 0 && (to.to_flags & TOF_SIGNATURE) == 0) { TCPSTAT_INC(tcps_sig_err_sigopt); /* XXX: should drop? */ } #endif /* * If echoed timestamp is later than the current time, * fall back to non RFC1323 RTT calculation. Normalize * timestamp if syncookies were used when this connection * was established. */ if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) { to.to_tsecr -= tp->ts_offset; if (TSTMP_GT(to.to_tsecr, tcp_ts_getticks())) to.to_tsecr = 0; else if (tp->t_flags & TF_PREVVALID && tp->t_badrxtwin != 0 && SEQ_LT(to.to_tsecr, tp->t_badrxtwin)) cc_cong_signal(tp, th, CC_RTO_ERR); } /* * Process options only when we get SYN/ACK back. The SYN case * for incoming connections is handled in tcp_syncache. * According to RFC1323 the window field in a SYN (i.e., a * or ) segment itself is never scaled. * XXX this is traditional behavior, may need to be cleaned up. */ if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { if ((to.to_flags & TOF_SCALE) && (tp->t_flags & TF_REQ_SCALE)) { tp->t_flags |= TF_RCVD_SCALE; tp->snd_scale = to.to_wscale; } /* * Initial send window. It will be updated with * the next incoming segment to the scaled value. */ tp->snd_wnd = th->th_win; if (to.to_flags & TOF_TS) { tp->t_flags |= TF_RCVD_TSTMP; tp->ts_recent = to.to_tsval; tp->ts_recent_age = tcp_ts_getticks(); } if (to.to_flags & TOF_MSS) tcp_mss(tp, to.to_mss); if ((tp->t_flags & TF_SACK_PERMIT) && (to.to_flags & TOF_SACKPERM) == 0) tp->t_flags &= ~TF_SACK_PERMIT; if (IS_FASTOPEN(tp->t_flags)) { if (to.to_flags & TOF_FASTOPEN) tcp_fastopen_update_cache(tp, to.to_mss, to.to_tfo_len, to.to_tfo_cookie); else tcp_fastopen_disable_path(tp); } } /* * If timestamps were negotiated during SYN/ACK they should * appear on every segment during this session and vice versa. */ if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS)) { if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: Timestamp missing, " "no action\n", s, __func__); free(s, M_TCPLOG); } } if (!(tp->t_flags & TF_RCVD_TSTMP) && (to.to_flags & TOF_TS)) { if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: Timestamp not expected, " "no action\n", s, __func__); free(s, M_TCPLOG); } } /* * Header prediction: check for the two common cases * of a uni-directional data xfer. If the packet has * no control flags, is in-sequence, the window didn't * change and we're not retransmitting, it's a * candidate. If the length is zero and the ack moved * forward, we're the sender side of the xfer. Just * free the data acked & wake any higher level process * that was blocked waiting for space. If the length * is non-zero and the ack didn't move, we're the * receiver side. If we're getting packets in-order * (the reassembly queue is empty), add the data to * the socket buffer and note that we need a delayed ack. * Make sure that the hidden state-flags are also off. * Since we check for TCPS_ESTABLISHED first, it can only * be TH_NEEDSYN. */ if (tp->t_state == TCPS_ESTABLISHED && th->th_seq == tp->rcv_nxt && (thflags & (TH_SYN|TH_FIN|TH_RST|TH_URG|TH_ACK)) == TH_ACK && tp->snd_nxt == tp->snd_max && tiwin && tiwin == tp->snd_wnd && ((tp->t_flags & (TF_NEEDSYN|TF_NEEDFIN)) == 0) && LIST_EMPTY(&tp->t_segq) && ((to.to_flags & TOF_TS) == 0 || TSTMP_GEQ(to.to_tsval, tp->ts_recent)) ) { /* * If last ACK falls within this segment's sequence numbers, * record the timestamp. * NOTE that the test is modified according to the latest * proposal of the tcplw@cray.com list (Braden 1993/04/26). */ if ((to.to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to.to_tsval; } if (tlen == 0) { if (SEQ_GT(th->th_ack, tp->snd_una) && SEQ_LEQ(th->th_ack, tp->snd_max) && !IN_RECOVERY(tp->t_flags) && (to.to_flags & TOF_SACK) == 0 && TAILQ_EMPTY(&tp->snd_holes)) { /* * This is a pure ack for outstanding data. */ if (ti_locked == TI_RLOCKED) INP_INFO_RUNLOCK(&V_tcbinfo); ti_locked = TI_UNLOCKED; TCPSTAT_INC(tcps_predack); /* * "bad retransmit" recovery without timestamps. */ if ((to.to_flags & TOF_TS) == 0 && tp->t_rxtshift == 1 && tp->t_flags & TF_PREVVALID && (int)(ticks - tp->t_badrxtwin) < 0) { cc_cong_signal(tp, th, CC_RTO_ERR); } /* * Recalculate the transmit timer / rtt. * * Some boxes send broken timestamp replies * during the SYN+ACK phase, ignore * timestamps of 0 or we could calculate a * huge RTT and blow up the retransmit timer. */ if ((to.to_flags & TOF_TS) != 0 && to.to_tsecr) { uint32_t t; t = tcp_ts_getticks() - to.to_tsecr; if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; tcp_xmit_timer(tp, TCP_TS_TO_TICKS(t) + 1); } else if (tp->t_rtttime && SEQ_GT(th->th_ack, tp->t_rtseq)) { if (!tp->t_rttlow || tp->t_rttlow > ticks - tp->t_rtttime) tp->t_rttlow = ticks - tp->t_rtttime; tcp_xmit_timer(tp, ticks - tp->t_rtttime); } acked = BYTES_THIS_ACK(tp, th); #ifdef TCP_HHOOK /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ hhook_run_tcp_est_in(tp, th, &to); #endif TCPSTAT_ADD(tcps_rcvackpack, nsegs); TCPSTAT_ADD(tcps_rcvackbyte, acked); sbdrop(&so->so_snd, acked); if (SEQ_GT(tp->snd_una, tp->snd_recover) && SEQ_LEQ(th->th_ack, tp->snd_recover)) tp->snd_recover = th->th_ack - 1; /* * Let the congestion control algorithm update * congestion control related information. This * typically means increasing the congestion * window. */ cc_ack_received(tp, th, nsegs, CC_ACK); tp->snd_una = th->th_ack; /* * Pull snd_wl2 up to prevent seq wrap relative * to th_ack. */ tp->snd_wl2 = th->th_ack; tp->t_dupacks = 0; m_freem(m); /* * If all outstanding data are acked, stop * retransmit timer, otherwise restart timer * using current (possibly backed-off) value. * If process is waiting for space, * wakeup/selwakeup/signal. If data * are ready to send, let tcp_output * decide between more output or persist. */ #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif TCP_PROBE3(debug__input, tp, th, m); if (tp->snd_una == tp->snd_max) tcp_timer_activate(tp, TT_REXMT, 0); else if (!tcp_timer_active(tp, TT_PERSIST)) tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); sowwakeup(so); if (sbavail(&so->so_snd)) (void) tp->t_fb->tfb_tcp_output(tp); goto check_delack; } } else if (th->th_ack == tp->snd_una && tlen <= sbspace(&so->so_rcv)) { int newsize = 0; /* automatic sockbuf scaling */ /* * This is a pure, in-sequence data packet with * nothing on the reassembly queue and we have enough * buffer space to take it. */ if (ti_locked == TI_RLOCKED) INP_INFO_RUNLOCK(&V_tcbinfo); ti_locked = TI_UNLOCKED; /* Clean receiver SACK report if present */ if ((tp->t_flags & TF_SACK_PERMIT) && tp->rcv_numsacks) tcp_clean_sackreport(tp); TCPSTAT_INC(tcps_preddat); tp->rcv_nxt += tlen; /* * Pull snd_wl1 up to prevent seq wrap relative to * th_seq. */ tp->snd_wl1 = th->th_seq; /* * Pull rcv_up up to prevent seq wrap relative to * rcv_nxt. */ tp->rcv_up = tp->rcv_nxt; TCPSTAT_ADD(tcps_rcvpack, nsegs); TCPSTAT_ADD(tcps_rcvbyte, tlen); #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif TCP_PROBE3(debug__input, tp, th, m); newsize = tcp_autorcvbuf(m, th, so, tp, tlen); /* Add data to socket buffer. */ SOCKBUF_LOCK(&so->so_rcv); if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { m_freem(m); } else { /* * Set new socket buffer size. * Give up when limit is reached. */ if (newsize) if (!sbreserve_locked(&so->so_rcv, newsize, so, NULL)) so->so_rcv.sb_flags &= ~SB_AUTOSIZE; m_adj(m, drop_hdrlen); /* delayed header drop */ sbappendstream_locked(&so->so_rcv, m, 0); } /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); if (DELAY_ACK(tp, tlen)) { tp->t_flags |= TF_DELACK; } else { tp->t_flags |= TF_ACKNOW; tp->t_fb->tfb_tcp_output(tp); } goto check_delack; } } /* * Calculate amount of space in receive window, * and then do TCP input processing. * Receive window is amount of space in rcv queue, * but not less than advertised window. */ win = sbspace(&so->so_rcv); if (win < 0) win = 0; tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt)); switch (tp->t_state) { /* * If the state is SYN_RECEIVED: * if seg contains an ACK, but not for our SYN/ACK, send a RST. */ case TCPS_SYN_RECEIVED: if ((thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->snd_una) || SEQ_GT(th->th_ack, tp->snd_max))) { rstreason = BANDLIM_RST_OPENPORT; goto dropwithreset; } if (IS_FASTOPEN(tp->t_flags)) { /* * When a TFO connection is in SYN_RECEIVED, the * only valid packets are the initial SYN, a * retransmit/copy of the initial SYN (possibly with * a subset of the original data), a valid ACK, a * FIN, or a RST. */ if ((thflags & (TH_SYN|TH_ACK)) == (TH_SYN|TH_ACK)) { rstreason = BANDLIM_RST_OPENPORT; goto dropwithreset; } else if (thflags & TH_SYN) { /* non-initial SYN is ignored */ if ((tcp_timer_active(tp, TT_DELACK) || tcp_timer_active(tp, TT_REXMT))) goto drop; } else if (!(thflags & (TH_ACK|TH_FIN|TH_RST))) { goto drop; } } break; /* * If the state is SYN_SENT: * if seg contains a RST with valid ACK (SEQ.ACK has already * been verified), then drop the connection. * if seg contains a RST without an ACK, drop the seg. * if seg does not contain SYN, then drop the seg. * Otherwise this is an acceptable SYN segment * initialize tp->rcv_nxt and tp->irs * if seg contains ack then advance tp->snd_una * if seg contains an ECE and ECN support is enabled, the stream * is ECN capable. * if SYN has been acked change to ESTABLISHED else SYN_RCVD state * arrange for segment to be acked (eventually) * continue processing rest of data/controls, beginning with URG */ case TCPS_SYN_SENT: if ((thflags & (TH_ACK|TH_RST)) == (TH_ACK|TH_RST)) { TCP_PROBE5(connect__refused, NULL, tp, m, tp, th); tp = tcp_drop(tp, ECONNREFUSED); } if (thflags & TH_RST) goto drop; if (!(thflags & TH_SYN)) goto drop; tp->irs = th->th_seq; tcp_rcvseqinit(tp); if (thflags & TH_ACK) { int tfo_partial_ack = 0; TCPSTAT_INC(tcps_connects); soisconnected(so); #ifdef MAC mac_socketpeer_set_from_mbuf(m, so); #endif /* Do window scaling on this connection? */ if ((tp->t_flags & (TF_RCVD_SCALE|TF_REQ_SCALE)) == (TF_RCVD_SCALE|TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; } tp->rcv_adv += min(tp->rcv_wnd, TCP_MAXWIN << tp->rcv_scale); tp->snd_una++; /* SYN is acked */ /* * If not all the data that was sent in the TFO SYN * has been acked, resend the remainder right away. */ if (IS_FASTOPEN(tp->t_flags) && (tp->snd_una != tp->snd_max)) { tp->snd_nxt = th->th_ack; tfo_partial_ack = 1; } /* * If there's data, delay ACK; if there's also a FIN * ACKNOW will be turned on later. */ if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial_ack) tcp_timer_activate(tp, TT_DELACK, tcp_delacktime); else tp->t_flags |= TF_ACKNOW; if ((thflags & TH_ECE) && V_tcp_do_ecn) { tp->t_flags |= TF_ECN_PERMIT; TCPSTAT_INC(tcps_ecn_shs); } /* * Received in SYN_SENT[*] state. * Transitions: * SYN_SENT --> ESTABLISHED * SYN_SENT* --> FIN_WAIT_1 */ tp->t_starttime = ticks; if (tp->t_flags & TF_NEEDFIN) { tcp_state_change(tp, TCPS_FIN_WAIT_1); tp->t_flags &= ~TF_NEEDFIN; thflags &= ~TH_SYN; } else { tcp_state_change(tp, TCPS_ESTABLISHED); TCP_PROBE5(connect__established, NULL, tp, m, tp, th); cc_conn_init(tp); tcp_timer_activate(tp, TT_KEEP, TP_KEEPIDLE(tp)); } } else { /* * Received initial SYN in SYN-SENT[*] state => * simultaneous open. * If it succeeds, connection is * half-synchronized. * Otherwise, do 3-way handshake: * SYN-SENT -> SYN-RECEIVED * SYN-SENT* -> SYN-RECEIVED* */ tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN); tcp_timer_activate(tp, TT_REXMT, 0); tcp_state_change(tp, TCPS_SYN_RECEIVED); } KASSERT(ti_locked == TI_RLOCKED, ("%s: trimthenstep6: " "ti_locked %d", __func__, ti_locked)); INP_INFO_RLOCK_ASSERT(&V_tcbinfo); INP_WLOCK_ASSERT(tp->t_inpcb); /* * Advance th->th_seq to correspond to first data byte. * If data, trim to stay within window, * dropping FIN if necessary. */ th->th_seq++; if (tlen > tp->rcv_wnd) { todrop = tlen - tp->rcv_wnd; m_adj(m, -todrop); tlen = tp->rcv_wnd; thflags &= ~TH_FIN; TCPSTAT_INC(tcps_rcvpackafterwin); TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); } tp->snd_wl1 = th->th_seq - 1; tp->rcv_up = th->th_seq; /* * Client side of transaction: already sent SYN and data. * If the remote host used T/TCP to validate the SYN, * our data will be ACK'd; if so, enter normal data segment * processing in the middle of step 5, ack processing. * Otherwise, goto step 6. */ if (thflags & TH_ACK) goto process_ACK; goto step6; /* * If the state is LAST_ACK or CLOSING or TIME_WAIT: * do normal processing. * * NB: Leftover from RFC1644 T/TCP. Cases to be reused later. */ case TCPS_LAST_ACK: case TCPS_CLOSING: break; /* continue normal processing */ } /* * States other than LISTEN or SYN_SENT. * First check the RST flag and sequence number since reset segments * are exempt from the timestamp and connection count tests. This * fixes a bug introduced by the Stevens, vol. 2, p. 960 bugfix * below which allowed reset segments in half the sequence space * to fall though and be processed (which gives forged reset * segments with a random sequence number a 50 percent chance of * killing a connection). * Then check timestamp, if present. * Then check the connection count, if present. * Then check that at least some bytes of segment are within * receive window. If segment begins before rcv_nxt, * drop leading data (and SYN); if nothing left, just ack. */ if (thflags & TH_RST) { /* * RFC5961 Section 3.2 * * - RST drops connection only if SEG.SEQ == RCV.NXT. * - If RST is in window, we send challenge ACK. * * Note: to take into account delayed ACKs, we should * test against last_ack_sent instead of rcv_nxt. * Note 2: we handle special case of closed window, not * covered by the RFC. */ if ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) && SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) || (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq)) { INP_INFO_RLOCK_ASSERT(&V_tcbinfo); KASSERT(ti_locked == TI_RLOCKED, ("%s: TH_RST ti_locked %d, th %p tp %p", __func__, ti_locked, th, tp)); KASSERT(tp->t_state != TCPS_SYN_SENT, ("%s: TH_RST for TCPS_SYN_SENT th %p tp %p", __func__, th, tp)); if (V_tcp_insecure_rst || tp->last_ack_sent == th->th_seq) { TCPSTAT_INC(tcps_drops); /* Drop the connection. */ switch (tp->t_state) { case TCPS_SYN_RECEIVED: so->so_error = ECONNREFUSED; goto close; case TCPS_ESTABLISHED: case TCPS_FIN_WAIT_1: case TCPS_FIN_WAIT_2: case TCPS_CLOSE_WAIT: case TCPS_CLOSING: case TCPS_LAST_ACK: so->so_error = ECONNRESET; close: /* FALLTHROUGH */ default: tp = tcp_close(tp); } } else { TCPSTAT_INC(tcps_badrst); /* Send challenge ACK. */ tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt, tp->snd_nxt, TH_ACK); tp->last_ack_sent = tp->rcv_nxt; m = NULL; } } goto drop; } /* * RFC5961 Section 4.2 * Send challenge ACK for any SYN in synchronized state. */ if ((thflags & TH_SYN) && tp->t_state != TCPS_SYN_SENT && tp->t_state != TCPS_SYN_RECEIVED) { KASSERT(ti_locked == TI_RLOCKED, ("tcp_do_segment: TH_SYN ti_locked %d", ti_locked)); INP_INFO_RLOCK_ASSERT(&V_tcbinfo); TCPSTAT_INC(tcps_badsyn); if (V_tcp_insecure_syn && SEQ_GEQ(th->th_seq, tp->last_ack_sent) && SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) { tp = tcp_drop(tp, ECONNRESET); rstreason = BANDLIM_UNLIMITED; } else { /* Send challenge ACK. */ tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt, tp->snd_nxt, TH_ACK); tp->last_ack_sent = tp->rcv_nxt; m = NULL; } goto drop; } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment * and it's less than ts_recent, drop it. */ if ((to.to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to.to_tsval, tp->ts_recent)) { /* Check to see if ts_recent is over 24 days old. */ if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) { /* * Invalidate ts_recent. If this segment updates * ts_recent, the age will be reset later and ts_recent * will get a valid value. If it does not, setting * ts_recent to zero will at least satisfy the * requirement that zero be placed in the timestamp * echo reply when ts_recent isn't valid. The * age isn't reset until we get a valid ts_recent * because we don't want out-of-order segments to be * dropped when ts_recent is old. */ tp->ts_recent = 0; } else { TCPSTAT_INC(tcps_rcvduppack); TCPSTAT_ADD(tcps_rcvdupbyte, tlen); TCPSTAT_INC(tcps_pawsdrop); if (tlen) goto dropafterack; goto drop; } } /* * In the SYN-RECEIVED state, validate that the packet belongs to * this connection before trimming the data to fit the receive * window. Check the sequence number versus IRS since we know * the sequence numbers haven't wrapped. This is a partial fix * for the "LAND" DoS attack. */ if (tp->t_state == TCPS_SYN_RECEIVED && SEQ_LT(th->th_seq, tp->irs)) { rstreason = BANDLIM_RST_OPENPORT; goto dropwithreset; } todrop = tp->rcv_nxt - th->th_seq; if (todrop > 0) { if (thflags & TH_SYN) { thflags &= ~TH_SYN; th->th_seq++; if (th->th_urp > 1) th->th_urp--; else thflags &= ~TH_URG; todrop--; } /* * Following if statement from Stevens, vol. 2, p. 960. */ if (todrop > tlen || (todrop == tlen && (thflags & TH_FIN) == 0)) { /* * Any valid FIN must be to the left of the window. * At this point the FIN must be a duplicate or out * of sequence; drop it. */ thflags &= ~TH_FIN; /* * Send an ACK to resynchronize and drop any data. * But keep on processing for RST or ACK. */ tp->t_flags |= TF_ACKNOW; todrop = tlen; TCPSTAT_INC(tcps_rcvduppack); TCPSTAT_ADD(tcps_rcvdupbyte, todrop); } else { TCPSTAT_INC(tcps_rcvpartduppack); TCPSTAT_ADD(tcps_rcvpartdupbyte, todrop); } drop_hdrlen += todrop; /* drop from the top afterwards */ th->th_seq += todrop; tlen -= todrop; if (th->th_urp > todrop) th->th_urp -= todrop; else { thflags &= ~TH_URG; th->th_urp = 0; } } /* * If new data are received on a connection after the * user processes are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tp->t_state > TCPS_CLOSE_WAIT && tlen) { KASSERT(ti_locked == TI_RLOCKED, ("%s: SS_NOFDEREF && " "CLOSE_WAIT && tlen ti_locked %d", __func__, ti_locked)); INP_INFO_RLOCK_ASSERT(&V_tcbinfo); if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: %s: Received %d bytes of data " "after socket was closed, " "sending RST and removing tcpcb\n", s, __func__, tcpstates[tp->t_state], tlen); free(s, M_TCPLOG); } tp = tcp_close(tp); TCPSTAT_INC(tcps_rcvafterclose); rstreason = BANDLIM_UNLIMITED; goto dropwithreset; } /* * If segment ends after window, drop trailing data * (and PUSH and FIN); if nothing left, just ACK. */ todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd); if (todrop > 0) { TCPSTAT_INC(tcps_rcvpackafterwin); if (todrop >= tlen) { TCPSTAT_ADD(tcps_rcvbyteafterwin, tlen); /* * If window is closed can only take segments at * window edge, and have to drop data and PUSH from * incoming segments. Continue processing, but * remember to ack. Otherwise, drop segment * and ack. */ if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) { tp->t_flags |= TF_ACKNOW; TCPSTAT_INC(tcps_rcvwinprobe); } else goto dropafterack; } else TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); m_adj(m, -todrop); tlen -= todrop; thflags &= ~(TH_PUSH|TH_FIN); } /* * If last ACK falls within this segment's sequence numbers, * record its timestamp. * NOTE: * 1) That the test incorporates suggestions from the latest * proposal of the tcplw@cray.com list (Braden 1993/04/26). * 2) That updating only on newer timestamps interferes with * our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. * 3) That we modify the segment boundary check to be * Last.ACK.Sent <= SEG.SEQ + SEG.Len * instead of RFC1323's * Last.ACK.Sent < SEG.SEQ + SEG.Len, * This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated * Vol. 2 p.869. In such cases, we can still calculate the * RTT correctly when RCV.NXT == Last.ACK.Sent. */ if ((to.to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN|TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to.to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN * flag is on (half-synchronized state), then queue data for * later processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_state == TCPS_SYN_RECEIVED || (tp->t_flags & TF_NEEDSYN)) { if (tp->t_state == TCPS_SYN_RECEIVED && IS_FASTOPEN(tp->t_flags)) { tp->snd_wnd = tiwin; cc_conn_init(tp); } goto step6; } else if (tp->t_flags & TF_ACKNOW) goto dropafterack; else goto drop; } /* * Ack processing. */ switch (tp->t_state) { /* * In SYN_RECEIVED state, the ack ACKs our SYN, so enter * ESTABLISHED state and continue processing. * The ACK was checked above. */ case TCPS_SYN_RECEIVED: TCPSTAT_INC(tcps_connects); soisconnected(so); /* Do window scaling? */ if ((tp->t_flags & (TF_RCVD_SCALE|TF_REQ_SCALE)) == (TF_RCVD_SCALE|TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; tp->snd_wnd = tiwin; } /* * Make transitions: * SYN-RECEIVED -> ESTABLISHED * SYN-RECEIVED* -> FIN-WAIT-1 */ tp->t_starttime = ticks; if (tp->t_flags & TF_NEEDFIN) { tcp_state_change(tp, TCPS_FIN_WAIT_1); tp->t_flags &= ~TF_NEEDFIN; } else { tcp_state_change(tp, TCPS_ESTABLISHED); TCP_PROBE5(accept__established, NULL, tp, m, tp, th); if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) { tcp_fastopen_decrement_counter(tp->t_tfo_pending); tp->t_tfo_pending = NULL; /* * Account for the ACK of our SYN prior to * regular ACK processing below. */ tp->snd_una++; } /* * TFO connections call cc_conn_init() during SYN * processing. Calling it again here for such * connections is not harmless as it would undo the * snd_cwnd reduction that occurs when a TFO SYN|ACK * is retransmitted. */ if (!IS_FASTOPEN(tp->t_flags)) cc_conn_init(tp); tcp_timer_activate(tp, TT_KEEP, TP_KEEPIDLE(tp)); } /* * If segment contains data or ACK, will call tcp_reass() * later; if not, do so now to pass queued data to user. */ if (tlen == 0 && (thflags & TH_FIN) == 0) (void) tcp_reass(tp, (struct tcphdr *)0, 0, (struct mbuf *)0); tp->snd_wl1 = th->th_seq - 1; /* FALLTHROUGH */ /* * In ESTABLISHED state: drop duplicate ACKs; ACK out of range * ACKs. If the ack is in the range * tp->snd_una < th->th_ack <= tp->snd_max * then advance tp->snd_una to th->th_ack and drop * data from the retransmission queue. If this ACK reflects * more up to date window information we update our window information. */ case TCPS_ESTABLISHED: case TCPS_FIN_WAIT_1: case TCPS_FIN_WAIT_2: case TCPS_CLOSE_WAIT: case TCPS_CLOSING: case TCPS_LAST_ACK: if (SEQ_GT(th->th_ack, tp->snd_max)) { TCPSTAT_INC(tcps_rcvacktoomuch); goto dropafterack; } if ((tp->t_flags & TF_SACK_PERMIT) && ((to.to_flags & TOF_SACK) || !TAILQ_EMPTY(&tp->snd_holes))) sack_changed = tcp_sack_doack(tp, &to, th->th_ack); else /* * Reset the value so that previous (valid) value * from the last ack with SACK doesn't get used. */ tp->sackhint.sacked_bytes = 0; #ifdef TCP_HHOOK /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ hhook_run_tcp_est_in(tp, th, &to); #endif if (SEQ_LEQ(th->th_ack, tp->snd_una)) { u_int maxseg; maxseg = tcp_maxseg(tp); if (tlen == 0 && (tiwin == tp->snd_wnd || (tp->t_flags & TF_SACK_PERMIT))) { /* * If this is the first time we've seen a * FIN from the remote, this is not a * duplicate and it needs to be processed * normally. This happens during a * simultaneous close. */ if ((thflags & TH_FIN) && (TCPS_HAVERCVDFIN(tp->t_state) == 0)) { tp->t_dupacks = 0; break; } TCPSTAT_INC(tcps_rcvdupack); /* * If we have outstanding data (other than * a window probe), this is a completely * duplicate ack (ie, window info didn't * change and FIN isn't set), * the ack is the biggest we've * seen and we've seen exactly our rexmt * threshold of them, assume a packet * has been dropped and retransmit it. * Kludge snd_nxt & the congestion * window so we send only this one * packet. * * We know we're losing at the current * window size so do congestion avoidance * (set ssthresh to half the current window * and pull our congestion window back to * the new ssthresh). * * Dup acks mean that packets have left the * network (they're now cached at the receiver) * so bump cwnd by the amount in the receiver * to keep a constant cwnd packets in the * network. * * When using TCP ECN, notify the peer that * we reduced the cwnd. */ /* * Following 2 kinds of acks should not affect * dupack counting: * 1) Old acks * 2) Acks with SACK but without any new SACK * information in them. These could result from * any anomaly in the network like a switch * duplicating packets or a possible DoS attack. */ if (th->th_ack != tp->snd_una || ((tp->t_flags & TF_SACK_PERMIT) && !sack_changed)) break; else if (!tcp_timer_active(tp, TT_REXMT)) tp->t_dupacks = 0; else if (++tp->t_dupacks > tcprexmtthresh || IN_FASTRECOVERY(tp->t_flags)) { cc_ack_received(tp, th, nsegs, CC_DUPACK); if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags)) { int awnd; /* * Compute the amount of data in flight first. * We can inject new data into the pipe iff * we have less than 1/2 the original window's * worth of data in flight. */ if (V_tcp_do_rfc6675_pipe) awnd = tcp_compute_pipe(tp); else awnd = (tp->snd_nxt - tp->snd_fack) + tp->sackhint.sack_bytes_rexmit; if (awnd < tp->snd_ssthresh) { tp->snd_cwnd += maxseg; if (tp->snd_cwnd > tp->snd_ssthresh) tp->snd_cwnd = tp->snd_ssthresh; } } else tp->snd_cwnd += maxseg; (void) tp->t_fb->tfb_tcp_output(tp); goto drop; } else if (tp->t_dupacks == tcprexmtthresh) { tcp_seq onxt = tp->snd_nxt; /* * If we're doing sack, check to * see if we're already in sack * recovery. If we're not doing sack, * check to see if we're in newreno * recovery. */ if (tp->t_flags & TF_SACK_PERMIT) { if (IN_FASTRECOVERY(tp->t_flags)) { tp->t_dupacks = 0; break; } } else { if (SEQ_LEQ(th->th_ack, tp->snd_recover)) { tp->t_dupacks = 0; break; } } /* Congestion signal before ack. */ cc_cong_signal(tp, th, CC_NDUPACK); cc_ack_received(tp, th, nsegs, CC_DUPACK); tcp_timer_activate(tp, TT_REXMT, 0); tp->t_rtttime = 0; if (tp->t_flags & TF_SACK_PERMIT) { TCPSTAT_INC( tcps_sack_recovery_episode); tp->sack_newdata = tp->snd_nxt; tp->snd_cwnd = maxseg; (void) tp->t_fb->tfb_tcp_output(tp); goto drop; } tp->snd_nxt = th->th_ack; tp->snd_cwnd = maxseg; (void) tp->t_fb->tfb_tcp_output(tp); KASSERT(tp->snd_limited <= 2, ("%s: tp->snd_limited too big", __func__)); tp->snd_cwnd = tp->snd_ssthresh + maxseg * (tp->t_dupacks - tp->snd_limited); if (SEQ_GT(onxt, tp->snd_nxt)) tp->snd_nxt = onxt; goto drop; } else if (V_tcp_do_rfc3042) { /* * Process first and second duplicate * ACKs. Each indicates a segment * leaving the network, creating room * for more. Make sure we can send a * packet on reception of each duplicate * ACK by increasing snd_cwnd by one * segment. Restore the original * snd_cwnd after packet transmission. */ cc_ack_received(tp, th, nsegs, CC_DUPACK); uint32_t oldcwnd = tp->snd_cwnd; tcp_seq oldsndmax = tp->snd_max; u_int sent; int avail; KASSERT(tp->t_dupacks == 1 || tp->t_dupacks == 2, ("%s: dupacks not 1 or 2", __func__)); if (tp->t_dupacks == 1) tp->snd_limited = 0; tp->snd_cwnd = (tp->snd_nxt - tp->snd_una) + (tp->t_dupacks - tp->snd_limited) * maxseg; /* * Only call tcp_output when there * is new data available to be sent. * Otherwise we would send pure ACKs. */ SOCKBUF_LOCK(&so->so_snd); avail = sbavail(&so->so_snd) - (tp->snd_nxt - tp->snd_una); SOCKBUF_UNLOCK(&so->so_snd); if (avail > 0) (void) tp->t_fb->tfb_tcp_output(tp); sent = tp->snd_max - oldsndmax; if (sent > maxseg) { KASSERT((tp->t_dupacks == 2 && tp->snd_limited == 0) || (sent == maxseg + 1 && tp->t_flags & TF_SENTFIN), ("%s: sent too much", __func__)); tp->snd_limited = 2; } else if (sent > 0) ++tp->snd_limited; tp->snd_cwnd = oldcwnd; goto drop; } } break; } else { /* * This ack is advancing the left edge, reset the * counter. */ tp->t_dupacks = 0; /* * If this ack also has new SACK info, increment the * counter as per rfc6675. */ if ((tp->t_flags & TF_SACK_PERMIT) && sack_changed) tp->t_dupacks++; } KASSERT(SEQ_GT(th->th_ack, tp->snd_una), ("%s: th_ack <= snd_una", __func__)); /* * If the congestion window was inflated to account * for the other side's cached packets, retract it. */ if (IN_FASTRECOVERY(tp->t_flags)) { if (SEQ_LT(th->th_ack, tp->snd_recover)) { if (tp->t_flags & TF_SACK_PERMIT) tcp_sack_partialack(tp, th); else tcp_newreno_partial_ack(tp, th); } else cc_post_recovery(tp, th); } /* * If we reach this point, ACK is not a duplicate, * i.e., it ACKs something we sent. */ if (tp->t_flags & TF_NEEDSYN) { /* * T/TCP: Connection was half-synchronized, and our * SYN has been ACK'd (so connection is now fully * synchronized). Go to non-starred state, * increment snd_una for ACK of SYN, and check if * we can do window scaling. */ tp->t_flags &= ~TF_NEEDSYN; tp->snd_una++; /* Do window scaling? */ if ((tp->t_flags & (TF_RCVD_SCALE|TF_REQ_SCALE)) == (TF_RCVD_SCALE|TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; /* Send window already scaled. */ } } process_ACK: INP_WLOCK_ASSERT(tp->t_inpcb); acked = BYTES_THIS_ACK(tp, th); KASSERT(acked >= 0, ("%s: acked unexepectedly negative " "(tp->snd_una=%u, th->th_ack=%u, tp=%p, m=%p)", __func__, tp->snd_una, th->th_ack, tp, m)); TCPSTAT_ADD(tcps_rcvackpack, nsegs); TCPSTAT_ADD(tcps_rcvackbyte, acked); /* * If we just performed our first retransmit, and the ACK * arrives within our recovery window, then it was a mistake * to do the retransmit in the first place. Recover our * original cwnd and ssthresh, and proceed to transmit where * we left off. */ if (tp->t_rxtshift == 1 && tp->t_flags & TF_PREVVALID && tp->t_badrxtwin && SEQ_LT(to.to_tsecr, tp->t_badrxtwin)) cc_cong_signal(tp, th, CC_RTO_ERR); /* * If we have a timestamp reply, update smoothed * round trip time. If no timestamp is present but * transmit timer is running and timed sequence * number was acked, update smoothed round trip time. * Since we now have an rtt measurement, cancel the * timer backoff (cf., Phil Karn's retransmit alg.). * Recompute the initial retransmit timer. * * Some boxes send broken timestamp replies * during the SYN+ACK phase, ignore * timestamps of 0 or we could calculate a * huge RTT and blow up the retransmit timer. */ if ((to.to_flags & TOF_TS) != 0 && to.to_tsecr) { uint32_t t; t = tcp_ts_getticks() - to.to_tsecr; if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; tcp_xmit_timer(tp, TCP_TS_TO_TICKS(t) + 1); } else if (tp->t_rtttime && SEQ_GT(th->th_ack, tp->t_rtseq)) { if (!tp->t_rttlow || tp->t_rttlow > ticks - tp->t_rtttime) tp->t_rttlow = ticks - tp->t_rtttime; tcp_xmit_timer(tp, ticks - tp->t_rtttime); } /* * If all outstanding data is acked, stop retransmit * timer and remember to restart (more output or persist). * If there is more data to be acked, restart retransmit * timer, using current (possibly backed-off) value. */ if (th->th_ack == tp->snd_max) { tcp_timer_activate(tp, TT_REXMT, 0); needoutput = 1; } else if (!tcp_timer_active(tp, TT_PERSIST)) tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); /* * If no data (only SYN) was ACK'd, * skip rest of ACK processing. */ if (acked == 0) goto step6; /* * Let the congestion control algorithm update congestion * control related information. This typically means increasing * the congestion window. */ cc_ack_received(tp, th, nsegs, CC_ACK); SOCKBUF_LOCK(&so->so_snd); if (acked > sbavail(&so->so_snd)) { if (tp->snd_wnd >= sbavail(&so->so_snd)) tp->snd_wnd -= sbavail(&so->so_snd); else tp->snd_wnd = 0; mfree = sbcut_locked(&so->so_snd, (int)sbavail(&so->so_snd)); ourfinisacked = 1; } else { mfree = sbcut_locked(&so->so_snd, acked); if (tp->snd_wnd >= (uint32_t) acked) tp->snd_wnd -= acked; else tp->snd_wnd = 0; ourfinisacked = 0; } /* NB: sowwakeup_locked() does an implicit unlock. */ sowwakeup_locked(so); m_freem(mfree); /* Detect una wraparound. */ if (!IN_RECOVERY(tp->t_flags) && SEQ_GT(tp->snd_una, tp->snd_recover) && SEQ_LEQ(th->th_ack, tp->snd_recover)) tp->snd_recover = th->th_ack - 1; /* XXXLAS: Can this be moved up into cc_post_recovery? */ if (IN_RECOVERY(tp->t_flags) && SEQ_GEQ(th->th_ack, tp->snd_recover)) { EXIT_RECOVERY(tp->t_flags); } tp->snd_una = th->th_ack; if (tp->t_flags & TF_SACK_PERMIT) { if (SEQ_GT(tp->snd_una, tp->snd_recover)) tp->snd_recover = tp->snd_una; } if (SEQ_LT(tp->snd_nxt, tp->snd_una)) tp->snd_nxt = tp->snd_una; switch (tp->t_state) { /* * In FIN_WAIT_1 STATE in addition to the processing * for the ESTABLISHED state if our FIN is now acknowledged * then enter FIN_WAIT_2. */ case TCPS_FIN_WAIT_1: if (ourfinisacked) { /* * If we can't receive any more * data, then closing user can proceed. * Starting the timer is contrary to the * specification, but if we don't get a FIN * we'll hang forever. * * XXXjl: * we should release the tp also, and use a * compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } break; /* * In CLOSING STATE in addition to the processing for * the ESTABLISHED state if the ACK acknowledges our FIN * then enter the TIME-WAIT state, otherwise ignore * the segment. */ case TCPS_CLOSING: if (ourfinisacked) { INP_INFO_RLOCK_ASSERT(&V_tcbinfo); tcp_twstart(tp); INP_INFO_RUNLOCK(&V_tcbinfo); m_freem(m); return; } break; /* * In LAST_ACK, we may still be waiting for data to drain * and/or to be acked, as well as for the ack of our FIN. * If our FIN is now acknowledged, delete the TCB, * enter the closed state and return. */ case TCPS_LAST_ACK: if (ourfinisacked) { INP_INFO_RLOCK_ASSERT(&V_tcbinfo); tp = tcp_close(tp); goto drop; } break; } } step6: INP_WLOCK_ASSERT(tp->t_inpcb); /* * Update window information. * Don't look at window if no ACK: TAC's send garbage on first SYN. */ if ((thflags & TH_ACK) && (SEQ_LT(tp->snd_wl1, th->th_seq) || (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) || (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) { /* keep track of pure window updates */ if (tlen == 0 && tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd) TCPSTAT_INC(tcps_rcvwinupd); tp->snd_wnd = tiwin; tp->snd_wl1 = th->th_seq; tp->snd_wl2 = th->th_ack; if (tp->snd_wnd > tp->max_sndwnd) tp->max_sndwnd = tp->snd_wnd; needoutput = 1; } /* * Process segments with URG. */ if ((thflags & TH_URG) && th->th_urp && TCPS_HAVERCVDFIN(tp->t_state) == 0) { /* * This is a kludge, but if we receive and accept * random urgent pointers, we'll crash in * soreceive. It's hard to imagine someone * actually wanting to send this much urgent data. */ SOCKBUF_LOCK(&so->so_rcv); if (th->th_urp + sbavail(&so->so_rcv) > sb_max) { th->th_urp = 0; /* XXX */ thflags &= ~TH_URG; /* XXX */ SOCKBUF_UNLOCK(&so->so_rcv); /* XXX */ goto dodata; /* XXX */ } /* * If this segment advances the known urgent pointer, * then mark the data stream. This should not happen * in CLOSE_WAIT, CLOSING, LAST_ACK or TIME_WAIT STATES since * a FIN has been received from the remote side. * In these states we ignore the URG. * * According to RFC961 (Assigned Protocols), * the urgent pointer points to the last octet * of urgent data. We continue, however, * to consider it to indicate the first octet * of data past the urgent section as the original * spec states (in one of two places). */ if (SEQ_GT(th->th_seq+th->th_urp, tp->rcv_up)) { tp->rcv_up = th->th_seq + th->th_urp; so->so_oobmark = sbavail(&so->so_rcv) + (tp->rcv_up - tp->rcv_nxt) - 1; if (so->so_oobmark == 0) so->so_rcv.sb_state |= SBS_RCVATMARK; sohasoutofband(so); tp->t_oobflags &= ~(TCPOOB_HAVEDATA | TCPOOB_HADDATA); } SOCKBUF_UNLOCK(&so->so_rcv); /* * Remove out of band data so doesn't get presented to user. * This can happen independent of advancing the URG pointer, * but if two URG's are pending at once, some out-of-band * data may creep in... ick. */ if (th->th_urp <= (uint32_t)tlen && !(so->so_options & SO_OOBINLINE)) { /* hdr drop is delayed */ tcp_pulloutofband(so, th, m, drop_hdrlen); } } else { /* * If no out of band data is expected, * pull receive urgent pointer along * with the receive window. */ if (SEQ_GT(tp->rcv_nxt, tp->rcv_up)) tp->rcv_up = tp->rcv_nxt; } dodata: /* XXX */ INP_WLOCK_ASSERT(tp->t_inpcb); /* * Process the segment text, merging it into the TCP sequencing queue, * and arranging for acknowledgment of receipt if necessary. * This process logically involves adjusting tp->rcv_wnd as data * is presented to the user (this happens in tcp_usrreq.c, * case PRU_RCVD). If a FIN has already been received on this * connection then we just ignore the text. */ tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) && IS_FASTOPEN(tp->t_flags)); if ((tlen || (thflags & TH_FIN) || tfo_syn) && TCPS_HAVERCVDFIN(tp->t_state) == 0) { tcp_seq save_start = th->th_seq; m_adj(m, drop_hdrlen); /* delayed header drop */ /* * Insert segment which includes th into TCP reassembly queue * with control block tp. Set thflags to whether reassembly now * includes a segment with FIN. This handles the common case * inline (segment is the next to be received on an established * connection, and the queue is empty), avoiding linkage into * and removal from the queue and repetition of various * conversions. * Set DELACK for segments received in order, but ack * immediately when segments are out of order (so * fast retransmit can work). */ if (th->th_seq == tp->rcv_nxt && LIST_EMPTY(&tp->t_segq) && (TCPS_HAVEESTABLISHED(tp->t_state) || tfo_syn)) { if (DELAY_ACK(tp, tlen) || tfo_syn) tp->t_flags |= TF_DELACK; else tp->t_flags |= TF_ACKNOW; tp->rcv_nxt += tlen; thflags = th->th_flags & TH_FIN; TCPSTAT_INC(tcps_rcvpack); TCPSTAT_ADD(tcps_rcvbyte, tlen); SOCKBUF_LOCK(&so->so_rcv); if (so->so_rcv.sb_state & SBS_CANTRCVMORE) m_freem(m); else sbappendstream_locked(&so->so_rcv, m, 0); /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); } else { /* * XXX: Due to the header drop above "th" is * theoretically invalid by now. Fortunately * m_adj() doesn't actually frees any mbufs * when trimming from the head. */ thflags = tcp_reass(tp, th, &tlen, m); tp->t_flags |= TF_ACKNOW; } if (tlen > 0 && (tp->t_flags & TF_SACK_PERMIT)) tcp_update_sack_list(tp, save_start, save_start + tlen); #if 0 /* * Note the amount of data that peer has sent into * our window, in order to estimate the sender's * buffer size. * XXX: Unused. */ if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) len = so->so_rcv.sb_hiwat - (tp->rcv_adv - tp->rcv_nxt); else len = so->so_rcv.sb_hiwat; #endif } else { m_freem(m); thflags &= ~TH_FIN; } /* * If FIN is received ACK the FIN and let the user know * that the connection is closing. */ if (thflags & TH_FIN) { if (TCPS_HAVERCVDFIN(tp->t_state) == 0) { socantrcvmore(so); /* * If connection is half-synchronized * (ie NEEDSYN flag on) then delay ACK, * so it may be piggybacked when SYN is sent. * Otherwise, since we received a FIN then no * more input can be expected, send ACK now. */ if (tp->t_flags & TF_NEEDSYN) tp->t_flags |= TF_DELACK; else tp->t_flags |= TF_ACKNOW; tp->rcv_nxt++; } switch (tp->t_state) { /* * In SYN_RECEIVED and ESTABLISHED STATES * enter the CLOSE_WAIT state. */ case TCPS_SYN_RECEIVED: tp->t_starttime = ticks; /* FALLTHROUGH */ case TCPS_ESTABLISHED: tcp_state_change(tp, TCPS_CLOSE_WAIT); break; /* * If still in FIN_WAIT_1 STATE FIN has not been acked so * enter the CLOSING state. */ case TCPS_FIN_WAIT_1: tcp_state_change(tp, TCPS_CLOSING); break; /* * In FIN_WAIT_2 state enter the TIME_WAIT state, * starting the time-wait timer, turning off the other * standard timers. */ case TCPS_FIN_WAIT_2: INP_INFO_RLOCK_ASSERT(&V_tcbinfo); KASSERT(ti_locked == TI_RLOCKED, ("%s: dodata " "TCP_FIN_WAIT_2 ti_locked: %d", __func__, ti_locked)); tcp_twstart(tp); INP_INFO_RUNLOCK(&V_tcbinfo); return; } } if (ti_locked == TI_RLOCKED) INP_INFO_RUNLOCK(&V_tcbinfo); ti_locked = TI_UNLOCKED; #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif TCP_PROBE3(debug__input, tp, th, m); /* * Return any desired output. */ if (needoutput || (tp->t_flags & TF_ACKNOW)) (void) tp->t_fb->tfb_tcp_output(tp); check_delack: KASSERT(ti_locked == TI_UNLOCKED, ("%s: check_delack ti_locked %d", __func__, ti_locked)); INP_INFO_UNLOCK_ASSERT(&V_tcbinfo); INP_WLOCK_ASSERT(tp->t_inpcb); if (tp->t_flags & TF_DELACK) { tp->t_flags &= ~TF_DELACK; tcp_timer_activate(tp, TT_DELACK, tcp_delacktime); } INP_WUNLOCK(tp->t_inpcb); return; dropafterack: /* * Generate an ACK dropping incoming segment if it occupies * sequence space, where the ACK reflects our state. * * We can now skip the test for the RST flag since all * paths to this code happen after packets containing * RST have been dropped. * * In the SYN-RECEIVED state, don't send an ACK unless the * segment we received passes the SYN-RECEIVED ACK test. * If it fails send a RST. This breaks the loop in the * "LAND" DoS attack, and also prevents an ACK storm * between two listening ports that have been sent forged * SYN segments, each with the source address of the other. */ if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) && (SEQ_GT(tp->snd_una, th->th_ack) || SEQ_GT(th->th_ack, tp->snd_max)) ) { rstreason = BANDLIM_RST_OPENPORT; goto dropwithreset; } #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_DROP, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif TCP_PROBE3(debug__input, tp, th, m); if (ti_locked == TI_RLOCKED) INP_INFO_RUNLOCK(&V_tcbinfo); ti_locked = TI_UNLOCKED; tp->t_flags |= TF_ACKNOW; (void) tp->t_fb->tfb_tcp_output(tp); INP_WUNLOCK(tp->t_inpcb); m_freem(m); return; dropwithreset: if (ti_locked == TI_RLOCKED) INP_INFO_RUNLOCK(&V_tcbinfo); ti_locked = TI_UNLOCKED; if (tp != NULL) { tcp_dropwithreset(m, th, tp, tlen, rstreason); INP_WUNLOCK(tp->t_inpcb); } else tcp_dropwithreset(m, th, NULL, tlen, rstreason); return; drop: if (ti_locked == TI_RLOCKED) { INP_INFO_RUNLOCK(&V_tcbinfo); ti_locked = TI_UNLOCKED; } #ifdef INVARIANTS else INP_INFO_UNLOCK_ASSERT(&V_tcbinfo); #endif /* * Drop space held by incoming segment and return. */ #ifdef TCPDEBUG if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) tcp_trace(TA_DROP, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif TCP_PROBE3(debug__input, tp, th, m); if (tp != NULL) INP_WUNLOCK(tp->t_inpcb); m_freem(m); } /* * Issue RST and make ACK acceptable to originator of segment. * The mbuf must still include the original packet header. * tp may be NULL. */ void tcp_dropwithreset(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int tlen, int rstreason) { #ifdef INET struct ip *ip; #endif #ifdef INET6 struct ip6_hdr *ip6; #endif if (tp != NULL) { INP_WLOCK_ASSERT(tp->t_inpcb); } /* Don't bother if destination was broadcast/multicast. */ if ((th->th_flags & TH_RST) || m->m_flags & (M_BCAST|M_MCAST)) goto drop; #ifdef INET6 if (mtod(m, struct ip *)->ip_v == 6) { ip6 = mtod(m, struct ip6_hdr *); if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) || IN6_IS_ADDR_MULTICAST(&ip6->ip6_src)) goto drop; /* IPv6 anycast check is done at tcp6_input() */ } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { ip = mtod(m, struct ip *); if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) || IN_MULTICAST(ntohl(ip->ip_src.s_addr)) || ip->ip_src.s_addr == htonl(INADDR_BROADCAST) || in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif)) goto drop; } #endif /* Perform bandwidth limiting. */ if (badport_bandlim(rstreason) < 0) goto drop; /* tcp_respond consumes the mbuf chain. */ if (th->th_flags & TH_ACK) { tcp_respond(tp, mtod(m, void *), th, m, (tcp_seq)0, th->th_ack, TH_RST); } else { if (th->th_flags & TH_SYN) tlen++; if (th->th_flags & TH_FIN) tlen++; tcp_respond(tp, mtod(m, void *), th, m, th->th_seq+tlen, (tcp_seq)0, TH_RST|TH_ACK); } return; drop: m_freem(m); } /* * Parse TCP options and place in tcpopt. */ void tcp_dooptions(struct tcpopt *to, u_char *cp, int cnt, int flags) { int opt, optlen; to->to_flags = 0; for (; cnt > 0; cnt -= optlen, cp += optlen) { opt = cp[0]; if (opt == TCPOPT_EOL) break; if (opt == TCPOPT_NOP) optlen = 1; else { if (cnt < 2) break; optlen = cp[1]; if (optlen < 2 || optlen > cnt) break; } switch (opt) { case TCPOPT_MAXSEG: if (optlen != TCPOLEN_MAXSEG) continue; if (!(flags & TO_SYN)) continue; to->to_flags |= TOF_MSS; bcopy((char *)cp + 2, (char *)&to->to_mss, sizeof(to->to_mss)); to->to_mss = ntohs(to->to_mss); break; case TCPOPT_WINDOW: if (optlen != TCPOLEN_WINDOW) continue; if (!(flags & TO_SYN)) continue; to->to_flags |= TOF_SCALE; to->to_wscale = min(cp[2], TCP_MAX_WINSHIFT); break; case TCPOPT_TIMESTAMP: if (optlen != TCPOLEN_TIMESTAMP) continue; to->to_flags |= TOF_TS; bcopy((char *)cp + 2, (char *)&to->to_tsval, sizeof(to->to_tsval)); to->to_tsval = ntohl(to->to_tsval); bcopy((char *)cp + 6, (char *)&to->to_tsecr, sizeof(to->to_tsecr)); to->to_tsecr = ntohl(to->to_tsecr); break; case TCPOPT_SIGNATURE: /* * In order to reply to a host which has set the * TCP_SIGNATURE option in its initial SYN, we have * to record the fact that the option was observed * here for the syncache code to perform the correct * response. */ if (optlen != TCPOLEN_SIGNATURE) continue; to->to_flags |= TOF_SIGNATURE; to->to_signature = cp + 2; break; case TCPOPT_SACK_PERMITTED: if (optlen != TCPOLEN_SACK_PERMITTED) continue; if (!(flags & TO_SYN)) continue; if (!V_tcp_do_sack) continue; to->to_flags |= TOF_SACKPERM; break; case TCPOPT_SACK: if (optlen <= 2 || (optlen - 2) % TCPOLEN_SACK != 0) continue; if (flags & TO_SYN) continue; to->to_flags |= TOF_SACK; to->to_nsacks = (optlen - 2) / TCPOLEN_SACK; to->to_sacks = cp + 2; TCPSTAT_INC(tcps_sack_rcv_blocks); break; case TCPOPT_FAST_OPEN: /* * Cookie length validation is performed by the * server side cookie checking code or the client * side cookie cache update code. */ if (!(flags & TO_SYN)) continue; if (!V_tcp_fastopen_client_enable && !V_tcp_fastopen_server_enable) continue; to->to_flags |= TOF_FASTOPEN; to->to_tfo_len = optlen - 2; to->to_tfo_cookie = to->to_tfo_len ? cp + 2 : NULL; break; default: continue; } } } /* * Pull out of band byte out of a segment so * it doesn't appear in the user's data queue. * It is still reflected in the segment length for * sequencing purposes. */ void tcp_pulloutofband(struct socket *so, struct tcphdr *th, struct mbuf *m, int off) { int cnt = off + th->th_urp - 1; while (cnt >= 0) { if (m->m_len > cnt) { char *cp = mtod(m, caddr_t) + cnt; struct tcpcb *tp = sototcpcb(so); INP_WLOCK_ASSERT(tp->t_inpcb); tp->t_iobc = *cp; tp->t_oobflags |= TCPOOB_HAVEDATA; bcopy(cp+1, cp, (unsigned)(m->m_len - cnt - 1)); m->m_len--; if (m->m_flags & M_PKTHDR) m->m_pkthdr.len--; return; } cnt -= m->m_len; m = m->m_next; if (m == NULL) break; } panic("tcp_pulloutofband"); } /* * Collect new round-trip time estimate * and update averages and current timeout. */ void tcp_xmit_timer(struct tcpcb *tp, int rtt) { int delta; INP_WLOCK_ASSERT(tp->t_inpcb); TCPSTAT_INC(tcps_rttupdated); tp->t_rttupdated++; if ((tp->t_srtt != 0) && (tp->t_rxtshift <= TCP_RTT_INVALIDATE)) { /* * srtt is stored as fixed point with 5 bits after the * binary point (i.e., scaled by 8). The following magic * is equivalent to the smoothing algorithm in rfc793 with * an alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed * point). Adjust rtt to origin 0. */ delta = ((rtt - 1) << TCP_DELTA_SHIFT) - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)); if ((tp->t_srtt += delta) <= 0) tp->t_srtt = 1; /* * We accumulate a smoothed rtt variance (actually, a * smoothed mean difference), then set the retransmit * timer to smoothed rtt + 4 times the smoothed variance. * rttvar is stored as fixed point with 4 bits after the * binary point (scaled by 16). The following is * equivalent to rfc793 smoothing with an alpha of .75 * (rttvar = rttvar*3/4 + |delta| / 4). This replaces * rfc793's wired-in beta. */ if (delta < 0) delta = -delta; delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT); if ((tp->t_rttvar += delta) <= 0) tp->t_rttvar = 1; if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar) tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } else { /* * No rtt measurement yet - use the unsmoothed rtt. * Set the variance to half the rtt (so our first * retransmit happens at 3*rtt). */ tp->t_srtt = rtt << TCP_RTT_SHIFT; tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1); tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } tp->t_rtttime = 0; tp->t_rxtshift = 0; /* * the retransmit should happen at rtt + 4 * rttvar. * Because of the way we do the smoothing, srtt and rttvar * will each average +1/2 tick of bias. When we compute * the retransmit timer, we want 1/2 tick of rounding and * 1 extra tick because of +-1/2 tick uncertainty in the * firing of the timer. The bias will give us exactly the * 1.5 tick we need. But, because the bias is * statistical, we have to test that we don't drop below * the minimum feasible timer (which is 2 ticks). */ TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp), max(tp->t_rttmin, rtt + 2), TCPTV_REXMTMAX); /* * We received an ack for a packet that wasn't retransmitted; * it is probably safe to discard any error indications we've * received recently. This isn't quite right, but close enough * for now (a route might have failed after we sent a segment, * and the return path might not be symmetrical). */ tp->t_softerror = 0; } /* * Determine a reasonable value for maxseg size. * If the route is known, check route for mtu. * If none, use an mss that can be handled on the outgoing interface * without forcing IP to fragment. If no route is found, route has no mtu, * or the destination isn't local, use a default, hopefully conservative * size (usually 512 or the default IP max size, but no more than the mtu * of the interface), as we can't discover anything about intervening * gateways or networks. We also initialize the congestion/slow start * window to be a single segment if the destination isn't local. * While looking at the routing entry, we also initialize other path-dependent * parameters from pre-set or cached values in the routing entry. * * NOTE that resulting t_maxseg doesn't include space for TCP options or * IP options, e.g. IPSEC data, since length of this data may vary, and * thus it is calculated for every segment separately in tcp_output(). * * NOTE that this routine is only called when we process an incoming * segment, or an ICMP need fragmentation datagram. Outgoing SYN/ACK MSS * settings are handled in tcp_mssopt(). */ void tcp_mss_update(struct tcpcb *tp, int offer, int mtuoffer, struct hc_metrics_lite *metricptr, struct tcp_ifcap *cap) { int mss = 0; uint32_t maxmtu = 0; struct inpcb *inp = tp->t_inpcb; struct hc_metrics_lite metrics; #ifdef INET6 int isipv6 = ((inp->inp_vflag & INP_IPV6) != 0) ? 1 : 0; size_t min_protoh = isipv6 ? sizeof (struct ip6_hdr) + sizeof (struct tcphdr) : sizeof (struct tcpiphdr); #else const size_t min_protoh = sizeof(struct tcpiphdr); #endif INP_WLOCK_ASSERT(tp->t_inpcb); if (mtuoffer != -1) { KASSERT(offer == -1, ("%s: conflict", __func__)); offer = mtuoffer - min_protoh; } /* Initialize. */ #ifdef INET6 if (isipv6) { maxmtu = tcp_maxmtu6(&inp->inp_inc, cap); tp->t_maxseg = V_tcp_v6mssdflt; } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { maxmtu = tcp_maxmtu(&inp->inp_inc, cap); tp->t_maxseg = V_tcp_mssdflt; } #endif /* * No route to sender, stay with default mss and return. */ if (maxmtu == 0) { /* * In case we return early we need to initialize metrics * to a defined state as tcp_hc_get() would do for us * if there was no cache hit. */ if (metricptr != NULL) bzero(metricptr, sizeof(struct hc_metrics_lite)); return; } /* What have we got? */ switch (offer) { case 0: /* * Offer == 0 means that there was no MSS on the SYN * segment, in this case we use tcp_mssdflt as * already assigned to t_maxseg above. */ offer = tp->t_maxseg; break; case -1: /* * Offer == -1 means that we didn't receive SYN yet. */ /* FALLTHROUGH */ default: /* * Prevent DoS attack with too small MSS. Round up * to at least minmss. */ offer = max(offer, V_tcp_minmss); } /* * rmx information is now retrieved from tcp_hostcache. */ tcp_hc_get(&inp->inp_inc, &metrics); if (metricptr != NULL) bcopy(&metrics, metricptr, sizeof(struct hc_metrics_lite)); /* * If there's a discovered mtu in tcp hostcache, use it. * Else, use the link mtu. */ if (metrics.rmx_mtu) mss = min(metrics.rmx_mtu, maxmtu) - min_protoh; else { #ifdef INET6 if (isipv6) { mss = maxmtu - min_protoh; if (!V_path_mtu_discovery && !in6_localaddr(&inp->in6p_faddr)) mss = min(mss, V_tcp_v6mssdflt); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { mss = maxmtu - min_protoh; if (!V_path_mtu_discovery && !in_localaddr(inp->inp_faddr)) mss = min(mss, V_tcp_mssdflt); } #endif /* * XXX - The above conditional (mss = maxmtu - min_protoh) * probably violates the TCP spec. * The problem is that, since we don't know the * other end's MSS, we are supposed to use a conservative * default. But, if we do that, then MTU discovery will * never actually take place, because the conservative * default is much less than the MTUs typically seen * on the Internet today. For the moment, we'll sweep * this under the carpet. * * The conservative default might not actually be a problem * if the only case this occurs is when sending an initial * SYN with options and data to a host we've never talked * to before. Then, they will reply with an MSS value which * will get recorded and the new parameters should get * recomputed. For Further Study. */ } mss = min(mss, offer); /* * Sanity check: make sure that maxseg will be large * enough to allow some data on segments even if the * all the option space is used (40bytes). Otherwise * funny things may happen in tcp_output. * * XXXGL: shouldn't we reserve space for IP/IPv6 options? */ mss = max(mss, 64); tp->t_maxseg = mss; } void tcp_mss(struct tcpcb *tp, int offer) { int mss; uint32_t bufsize; struct inpcb *inp; struct socket *so; struct hc_metrics_lite metrics; struct tcp_ifcap cap; KASSERT(tp != NULL, ("%s: tp == NULL", __func__)); bzero(&cap, sizeof(cap)); tcp_mss_update(tp, offer, -1, &metrics, &cap); mss = tp->t_maxseg; inp = tp->t_inpcb; /* * If there's a pipesize, change the socket buffer to that size, * don't change if sb_hiwat is different than default (then it * has been changed on purpose with setsockopt). * Make the socket buffers an integral number of mss units; * if the mss is larger than the socket buffer, decrease the mss. */ so = inp->inp_socket; SOCKBUF_LOCK(&so->so_snd); if ((so->so_snd.sb_hiwat == V_tcp_sendspace) && metrics.rmx_sendpipe) bufsize = metrics.rmx_sendpipe; else bufsize = so->so_snd.sb_hiwat; if (bufsize < mss) mss = bufsize; else { bufsize = roundup(bufsize, mss); if (bufsize > sb_max) bufsize = sb_max; if (bufsize > so->so_snd.sb_hiwat) (void)sbreserve_locked(&so->so_snd, bufsize, so, NULL); } SOCKBUF_UNLOCK(&so->so_snd); /* * Sanity check: make sure that maxseg will be large * enough to allow some data on segments even if the * all the option space is used (40bytes). Otherwise * funny things may happen in tcp_output. * * XXXGL: shouldn't we reserve space for IP/IPv6 options? */ tp->t_maxseg = max(mss, 64); SOCKBUF_LOCK(&so->so_rcv); if ((so->so_rcv.sb_hiwat == V_tcp_recvspace) && metrics.rmx_recvpipe) bufsize = metrics.rmx_recvpipe; else bufsize = so->so_rcv.sb_hiwat; if (bufsize > mss) { bufsize = roundup(bufsize, mss); if (bufsize > sb_max) bufsize = sb_max; if (bufsize > so->so_rcv.sb_hiwat) (void)sbreserve_locked(&so->so_rcv, bufsize, so, NULL); } SOCKBUF_UNLOCK(&so->so_rcv); /* Check the interface for TSO capabilities. */ if (cap.ifcap & CSUM_TSO) { tp->t_flags |= TF_TSO; tp->t_tsomax = cap.tsomax; tp->t_tsomaxsegcount = cap.tsomaxsegcount; tp->t_tsomaxsegsize = cap.tsomaxsegsize; } } /* * Determine the MSS option to send on an outgoing SYN. */ int tcp_mssopt(struct in_conninfo *inc) { int mss = 0; uint32_t thcmtu = 0; uint32_t maxmtu = 0; size_t min_protoh; KASSERT(inc != NULL, ("tcp_mssopt with NULL in_conninfo pointer")); #ifdef INET6 if (inc->inc_flags & INC_ISIPV6) { mss = V_tcp_v6mssdflt; maxmtu = tcp_maxmtu6(inc, NULL); min_protoh = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { mss = V_tcp_mssdflt; maxmtu = tcp_maxmtu(inc, NULL); min_protoh = sizeof(struct tcpiphdr); } #endif #if defined(INET6) || defined(INET) thcmtu = tcp_hc_getmtu(inc); /* IPv4 and IPv6 */ #endif if (maxmtu && thcmtu) mss = min(maxmtu, thcmtu) - min_protoh; else if (maxmtu || thcmtu) mss = max(maxmtu, thcmtu) - min_protoh; return (mss); } /* * On a partial ack arrives, force the retransmission of the * next unacknowledged segment. Do not clear tp->t_dupacks. * By setting snd_nxt to ti_ack, this forces retransmission timer to * be started again. */ void tcp_newreno_partial_ack(struct tcpcb *tp, struct tcphdr *th) { tcp_seq onxt = tp->snd_nxt; uint32_t ocwnd = tp->snd_cwnd; u_int maxseg = tcp_maxseg(tp); INP_WLOCK_ASSERT(tp->t_inpcb); tcp_timer_activate(tp, TT_REXMT, 0); tp->t_rtttime = 0; tp->snd_nxt = th->th_ack; /* * Set snd_cwnd to one segment beyond acknowledged offset. * (tp->snd_una has not yet been updated when this function is called.) */ tp->snd_cwnd = maxseg + BYTES_THIS_ACK(tp, th); tp->t_flags |= TF_ACKNOW; (void) tp->t_fb->tfb_tcp_output(tp); tp->snd_cwnd = ocwnd; if (SEQ_GT(onxt, tp->snd_nxt)) tp->snd_nxt = onxt; /* * Partial window deflation. Relies on fact that tp->snd_una * not updated yet. */ if (tp->snd_cwnd > BYTES_THIS_ACK(tp, th)) tp->snd_cwnd -= BYTES_THIS_ACK(tp, th); else tp->snd_cwnd = 0; tp->snd_cwnd += maxseg; } int tcp_compute_pipe(struct tcpcb *tp) { return (tp->snd_max - tp->snd_una + tp->sackhint.sack_bytes_rexmit - tp->sackhint.sacked_bytes); } Index: head/sys/netinet/tcp_stacks/rack.c =================================================================== --- head/sys/netinet/tcp_stacks/rack.c (revision 335355) +++ head/sys/netinet/tcp_stacks/rack.c (revision 335356) @@ -1,9214 +1,9188 @@ /*- * Copyright (c) 2016-2018 * Netflix Inc. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_tcpdebug.h" #include #include #include #ifdef TCP_HHOOK #include #endif #include #include #include #include #include #include /* for proc0 declaration */ #include #include #include #include #ifdef NETFLIX_STATS #include #endif #include #include #include #include #include #include #include #include #define TCPSTATES /* for logging */ #include #include #include #include #include /* required for icmp_var.h */ #include /* for ICMP_BANDLIM */ #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef NETFLIX_CWV #include #endif #include #ifdef TCPDEBUG #include #endif /* TCPDEBUG */ #ifdef TCP_OFFLOAD #include #endif #ifdef INET6 #include #endif #include #if defined(IPSEC) || defined(IPSEC_SUPPORT) #include #include #endif /* IPSEC */ #include #include #include #ifdef MAC #include #endif #include "sack_filter.h" #include "tcp_rack.h" #include "rack_bbr_common.h" uma_zone_t rack_zone; uma_zone_t rack_pcb_zone; #ifndef TICKS2SBT #define TICKS2SBT(__t) (tick_sbt * ((sbintime_t)(__t))) #endif struct sysctl_ctx_list rack_sysctl_ctx; struct sysctl_oid *rack_sysctl_root; #define CUM_ACKED 1 #define SACKED 2 /* * The RACK module incorporates a number of * TCP ideas that have been put out into the IETF * over the last few years: * - Matt Mathis's Rate Halving which slowly drops * the congestion window so that the ack clock can * be maintained during a recovery. * - Yuchung Cheng's RACK TCP (for which its named) that * will stop us using the number of dup acks and instead * use time as the gage of when we retransmit. * - Reorder Detection of RFC4737 and the Tail-Loss probe draft * of Dukkipati et.al. * RACK depends on SACK, so if an endpoint arrives that * cannot do SACK the state machine below will shuttle the * connection back to using the "default" TCP stack that is * in FreeBSD. * * To implement RACK the original TCP stack was first decomposed * into a functional state machine with individual states * for each of the possible TCP connection states. The do_segement * functions role in life is to mandate the connection supports SACK * initially and then assure that the RACK state matches the conenction * state before calling the states do_segment function. Each * state is simplified due to the fact that the original do_segment * has been decomposed and we *know* what state we are in (no * switches on the state) and all tests for SACK are gone. This * greatly simplifies what each state does. * * TCP output is also over-written with a new version since it * must maintain the new rack scoreboard. * */ static int32_t rack_precache = 1; static int32_t rack_tlp_thresh = 1; static int32_t rack_reorder_thresh = 2; static int32_t rack_reorder_fade = 60000; /* 0 - never fade, def 60,000 * - 60 seconds */ static int32_t rack_pkt_delay = 1; static int32_t rack_inc_var = 0;/* For TLP */ static int32_t rack_reduce_largest_on_idle = 0; static int32_t rack_min_pace_time = 0; static int32_t rack_min_pace_time_seg_req=6; static int32_t rack_early_recovery = 1; static int32_t rack_early_recovery_max_seg = 6; static int32_t rack_send_a_lot_in_prr = 1; static int32_t rack_min_to = 1; /* Number of ms minimum timeout */ static int32_t rack_tlp_in_recovery = 1; /* Can we do TLP in recovery? */ static int32_t rack_verbose_logging = 0; static int32_t rack_ignore_data_after_close = 1; /* * Currently regular tcp has a rto_min of 30ms * the backoff goes 12 times so that ends up * being a total of 122.850 seconds before a * connection is killed. */ static int32_t rack_tlp_min = 10; static int32_t rack_rto_min = 30; /* 30ms same as main freebsd */ static int32_t rack_rto_max = 30000; /* 30 seconds */ static const int32_t rack_free_cache = 2; static int32_t rack_hptsi_segments = 40; static int32_t rack_rate_sample_method = USE_RTT_LOW; static int32_t rack_pace_every_seg = 1; static int32_t rack_delayed_ack_time = 200; /* 200ms */ static int32_t rack_slot_reduction = 4; static int32_t rack_lower_cwnd_at_tlp = 0; static int32_t rack_use_proportional_reduce = 0; static int32_t rack_proportional_rate = 10; static int32_t rack_tlp_max_resend = 2; static int32_t rack_limited_retran = 0; static int32_t rack_always_send_oldest = 0; static int32_t rack_sack_block_limit = 128; static int32_t rack_use_sack_filter = 1; static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE; /* Rack specific counters */ counter_u64_t rack_badfr; counter_u64_t rack_badfr_bytes; counter_u64_t rack_rtm_prr_retran; counter_u64_t rack_rtm_prr_newdata; counter_u64_t rack_timestamp_mismatch; counter_u64_t rack_reorder_seen; counter_u64_t rack_paced_segments; counter_u64_t rack_unpaced_segments; counter_u64_t rack_saw_enobuf; counter_u64_t rack_saw_enetunreach; /* Tail loss probe counters */ counter_u64_t rack_tlp_tot; counter_u64_t rack_tlp_newdata; counter_u64_t rack_tlp_retran; counter_u64_t rack_tlp_retran_bytes; counter_u64_t rack_tlp_retran_fail; counter_u64_t rack_to_tot; counter_u64_t rack_to_arm_rack; counter_u64_t rack_to_arm_tlp; counter_u64_t rack_to_alloc; counter_u64_t rack_to_alloc_hard; counter_u64_t rack_to_alloc_emerg; counter_u64_t rack_sack_proc_all; counter_u64_t rack_sack_proc_short; counter_u64_t rack_sack_proc_restart; counter_u64_t rack_runt_sacks; counter_u64_t rack_used_tlpmethod; counter_u64_t rack_used_tlpmethod2; counter_u64_t rack_enter_tlp_calc; counter_u64_t rack_input_idle_reduces; counter_u64_t rack_tlp_does_nada; /* Temp CPU counters */ counter_u64_t rack_find_high; counter_u64_t rack_progress_drops; counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE]; counter_u64_t rack_opts_arry[RACK_OPTS_SIZE]; static void rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line); static int rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t * ti_locked, uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val); static int rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, int32_t * ti_locked, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); static void rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, struct tcphdr *th, uint16_t nsegs, uint16_t type, int32_t recovery); static struct rack_sendmap *rack_alloc(struct tcp_rack *rack); static struct rack_sendmap * rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused); static void rack_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type); static void rack_counter_destroy(void); static int rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp); static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how); static void rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t ti_locked); static void rack_dtor(void *mem, int32_t size, void *arg); static void rack_earlier_retran(struct tcpcb *tp, struct rack_sendmap *rsm, uint32_t t, uint32_t cts); static struct rack_sendmap * rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm); static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack); static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm); static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged); static int rack_get_sockopt(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack); static int32_t rack_handoff_ok(struct tcpcb *tp); static int32_t rack_init(struct tcpcb *tp); static void rack_init_sysctls(void); static void rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th); static void rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len, uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t ts, uint8_t pass, struct rack_sendmap *hintrsm); static void rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm); static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num); static int32_t rack_output(struct tcpcb *tp); static void rack_hpts_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t ti_locked, int32_t nxt_pkt, struct timeval *tv); static uint32_t rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts); static void rack_post_recovery(struct tcpcb *tp, struct tcphdr *th); static void rack_remxt_tmr(struct tcpcb *tp); static int rack_set_sockopt(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack); static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack); static int32_t rack_stopall(struct tcpcb *tp); static void rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta); static int32_t rack_timer_active(struct tcpcb *tp, uint32_t timer_type); static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line); static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type); static uint32_t rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t ts, int32_t * lenp); static void rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t ts); static int rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type); static int32_t tcp_addrack(module_t mod, int32_t type, void *data); static void rack_challenge_ack(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t * ti_locked, int32_t * ret_val); static int rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, int32_t * ti_locked, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); static int rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, int32_t * ti_locked, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); static void rack_do_drop(struct mbuf *m, struct tcpcb *tp, int32_t * ti_locked); static void rack_do_dropafterack(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, int32_t * ti_locked, int32_t thflags, int32_t tlen, int32_t * ret_val); static void rack_do_dropwithreset(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, int32_t * ti_locked, int32_t rstreason, int32_t tlen); static int rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, int32_t * ti_locked, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); static int rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, int32_t * ti_locked, uint32_t tiwin, int32_t nxt_pkt); static int rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, int32_t * ti_locked, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); static int rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, int32_t * ti_locked, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); static int rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, int32_t * ti_locked, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); static int rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, int32_t * ti_locked, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); static int rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, int32_t * ti_locked, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); static int rack_drop_checks(struct tcpopt *to, struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t * tlenp, int32_t * ti_locked, int32_t * thf, int32_t * drop_hdrlen, int32_t * ret_val); static int rack_process_rst(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t * ti_locked); struct rack_sendmap * tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused); static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt); static void tcp_rack_partialack(struct tcpcb *tp, struct tcphdr *th); static int rack_ts_check(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t * ti_locked, int32_t tlen, int32_t thflags, int32_t * ret_val); int32_t rack_clear_counter=0; static int sysctl_rack_clear(SYSCTL_HANDLER_ARGS) { uint32_t stat; int32_t error; error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t)); if (error || req->newptr == NULL) return error; error = SYSCTL_IN(req, &stat, sizeof(uint32_t)); if (error) return (error); if (stat == 1) { #ifdef INVARIANTS printf("Clearing RACK counters\n"); #endif counter_u64_zero(rack_badfr); counter_u64_zero(rack_badfr_bytes); counter_u64_zero(rack_rtm_prr_retran); counter_u64_zero(rack_rtm_prr_newdata); counter_u64_zero(rack_timestamp_mismatch); counter_u64_zero(rack_reorder_seen); counter_u64_zero(rack_tlp_tot); counter_u64_zero(rack_tlp_newdata); counter_u64_zero(rack_tlp_retran); counter_u64_zero(rack_tlp_retran_bytes); counter_u64_zero(rack_tlp_retran_fail); counter_u64_zero(rack_to_tot); counter_u64_zero(rack_to_arm_rack); counter_u64_zero(rack_to_arm_tlp); counter_u64_zero(rack_paced_segments); counter_u64_zero(rack_unpaced_segments); counter_u64_zero(rack_saw_enobuf); counter_u64_zero(rack_saw_enetunreach); counter_u64_zero(rack_to_alloc_hard); counter_u64_zero(rack_to_alloc_emerg); counter_u64_zero(rack_sack_proc_all); counter_u64_zero(rack_sack_proc_short); counter_u64_zero(rack_sack_proc_restart); counter_u64_zero(rack_to_alloc); counter_u64_zero(rack_find_high); counter_u64_zero(rack_runt_sacks); counter_u64_zero(rack_used_tlpmethod); counter_u64_zero(rack_used_tlpmethod2); counter_u64_zero(rack_enter_tlp_calc); counter_u64_zero(rack_progress_drops); counter_u64_zero(rack_tlp_does_nada); } rack_clear_counter = 0; return (0); } static void rack_init_sysctls() { SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "rate_sample_method", CTLFLAG_RW, &rack_rate_sample_method , USE_RTT_LOW, "What method should we use for rate sampling 0=high, 1=low "); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "data_after_close", CTLFLAG_RW, &rack_ignore_data_after_close, 0, "Do we hold off sending a RST until all pending data is ack'd"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "tlpmethod", CTLFLAG_RW, &rack_tlp_threshold_use, TLP_USE_TWO_ONE, "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "min_pace_time", CTLFLAG_RW, &rack_min_pace_time, 0, "Should we enforce a minimum pace time of 1ms"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "min_pace_segs", CTLFLAG_RW, &rack_min_pace_time_seg_req, 6, "How many segments have to be in the len to enforce min-pace-time"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "idle_reduce_high", CTLFLAG_RW, &rack_reduce_largest_on_idle, 0, "Should we reduce the largest cwnd seen to IW on idle reduction"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "bb_verbose", CTLFLAG_RW, &rack_verbose_logging, 0, "Should RACK black box logging be verbose"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "sackfiltering", CTLFLAG_RW, &rack_use_sack_filter, 1, "Do we use sack filtering?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "delayed_ack", CTLFLAG_RW, &rack_delayed_ack_time, 200, "Delayed ack time (200ms)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "tlpminto", CTLFLAG_RW, &rack_tlp_min, 10, "TLP minimum timeout per the specification (10ms)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "precache", CTLFLAG_RW, &rack_precache, 0, "Where should we precache the mcopy (0 is not at all)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "sblklimit", CTLFLAG_RW, &rack_sack_block_limit, 128, "When do we start paying attention to small sack blocks"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "send_oldest", CTLFLAG_RW, &rack_always_send_oldest, 1, "Should we always send the oldest TLP and RACK-TLP"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "rack_tlp_in_recovery", CTLFLAG_RW, &rack_tlp_in_recovery, 1, "Can we do a TLP during recovery?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "rack_tlimit", CTLFLAG_RW, &rack_limited_retran, 0, "How many times can a rack timeout drive out sends"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "minrto", CTLFLAG_RW, &rack_rto_min, 0, "Minimum RTO in ms -- set with caution below 1000 due to TLP"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "maxrto", CTLFLAG_RW, &rack_rto_max, 0, "Maxiumum RTO in ms -- should be at least as large as min_rto"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "tlp_retry", CTLFLAG_RW, &rack_tlp_max_resend, 2, "How many times does TLP retry a single segment or multiple with no ACK"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "recovery_loss_prop", CTLFLAG_RW, &rack_use_proportional_reduce, 0, "Should we proportionaly reduce cwnd based on the number of losses "); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "recovery_prop", CTLFLAG_RW, &rack_proportional_rate, 10, "What percent reduction per loss"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW, &rack_lower_cwnd_at_tlp, 0, "When a TLP completes a retran should we enter recovery?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "hptsi_reduces", CTLFLAG_RW, &rack_slot_reduction, 4, "When setting a slot should we reduce by divisor"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "hptsi_every_seg", CTLFLAG_RW, &rack_pace_every_seg, 1, "Should we pace out every segment hptsi"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "hptsi_seg_max", CTLFLAG_RW, &rack_hptsi_segments, 6, "Should we pace out only a limited size of segments"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "prr_sendalot", CTLFLAG_RW, &rack_send_a_lot_in_prr, 1, "Send a lot in prr"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "minto", CTLFLAG_RW, &rack_min_to, 1, "Minimum rack timeout in milliseconds"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "earlyrecoveryseg", CTLFLAG_RW, &rack_early_recovery_max_seg, 6, "Max segments in early recovery"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "earlyrecovery", CTLFLAG_RW, &rack_early_recovery, 1, "Do we do early recovery with rack"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "reorder_thresh", CTLFLAG_RW, &rack_reorder_thresh, 2, "What factor for rack will be added when seeing reordering (shift right)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW, &rack_tlp_thresh, 1, "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "reorder_fade", CTLFLAG_RW, &rack_reorder_fade, 0, "Does reorder detection fade, if so how many ms (0 means never)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "pktdelay", CTLFLAG_RW, &rack_pkt_delay, 1, "Extra RACK time (in ms) besides reordering thresh"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "inc_var", CTLFLAG_RW, &rack_inc_var, 0, "Should rack add to the TLP timer the variance in rtt calculation"); rack_badfr = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "badfr", CTLFLAG_RD, &rack_badfr, "Total number of bad FRs"); rack_badfr_bytes = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "badfr_bytes", CTLFLAG_RD, &rack_badfr_bytes, "Total number of bad FRs"); rack_rtm_prr_retran = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "prrsndret", CTLFLAG_RD, &rack_rtm_prr_retran, "Total number of prr based retransmits"); rack_rtm_prr_newdata = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "prrsndnew", CTLFLAG_RD, &rack_rtm_prr_newdata, "Total number of prr based new transmits"); rack_timestamp_mismatch = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "tsnf", CTLFLAG_RD, &rack_timestamp_mismatch, "Total number of timestamps that we could not find the reported ts"); rack_find_high = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "findhigh", CTLFLAG_RD, &rack_find_high, "Total number of FIN causing find-high"); rack_reorder_seen = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "reordering", CTLFLAG_RD, &rack_reorder_seen, "Total number of times we added delay due to reordering"); rack_tlp_tot = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "tlp_to_total", CTLFLAG_RD, &rack_tlp_tot, "Total number of tail loss probe expirations"); rack_tlp_newdata = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "tlp_new", CTLFLAG_RD, &rack_tlp_newdata, "Total number of tail loss probe sending new data"); rack_tlp_retran = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "tlp_retran", CTLFLAG_RD, &rack_tlp_retran, "Total number of tail loss probe sending retransmitted data"); rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD, &rack_tlp_retran_bytes, "Total bytes of tail loss probe sending retransmitted data"); rack_tlp_retran_fail = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "tlp_retran_fail", CTLFLAG_RD, &rack_tlp_retran_fail, "Total number of tail loss probe sending retransmitted data that failed (wait for t3)"); rack_to_tot = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "rack_to_tot", CTLFLAG_RD, &rack_to_tot, "Total number of times the rack to expired?"); rack_to_arm_rack = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "arm_rack", CTLFLAG_RD, &rack_to_arm_rack, "Total number of times the rack timer armed?"); rack_to_arm_tlp = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "arm_tlp", CTLFLAG_RD, &rack_to_arm_tlp, "Total number of times the tlp timer armed?"); rack_paced_segments = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "paced", CTLFLAG_RD, &rack_paced_segments, "Total number of times a segment send caused hptsi"); rack_unpaced_segments = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "unpaced", CTLFLAG_RD, &rack_unpaced_segments, "Total number of times a segment did not cause hptsi"); rack_saw_enobuf = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "saw_enobufs", CTLFLAG_RD, &rack_saw_enobuf, "Total number of times a segment did not cause hptsi"); rack_saw_enetunreach = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "saw_enetunreach", CTLFLAG_RD, &rack_saw_enetunreach, "Total number of times a segment did not cause hptsi"); rack_to_alloc = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "allocs", CTLFLAG_RD, &rack_to_alloc, "Total allocations of tracking structures"); rack_to_alloc_hard = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "allochard", CTLFLAG_RD, &rack_to_alloc_hard, "Total allocations done with sleeping the hard way"); rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "allocemerg", CTLFLAG_RD, &rack_to_alloc_emerg, "Total alocations done from emergency cache"); rack_sack_proc_all = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "sack_long", CTLFLAG_RD, &rack_sack_proc_all, "Total times we had to walk whole list for sack processing"); rack_sack_proc_restart = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "sack_restart", CTLFLAG_RD, &rack_sack_proc_restart, "Total times we had to walk whole list due to a restart"); rack_sack_proc_short = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "sack_short", CTLFLAG_RD, &rack_sack_proc_short, "Total times we took shortcut for sack processing"); rack_enter_tlp_calc = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "tlp_calc_entered", CTLFLAG_RD, &rack_enter_tlp_calc, "Total times we called calc-tlp"); rack_used_tlpmethod = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "hit_tlp_method", CTLFLAG_RD, &rack_used_tlpmethod, "Total number of runt sacks"); rack_used_tlpmethod2 = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "hit_tlp_method2", CTLFLAG_RD, &rack_used_tlpmethod2, "Total number of runt sacks 2"); rack_runt_sacks = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "runtsacks", CTLFLAG_RD, &rack_runt_sacks, "Total number of runt sacks"); rack_progress_drops = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "prog_drops", CTLFLAG_RD, &rack_progress_drops, "Total number of progress drops"); rack_input_idle_reduces = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD, &rack_input_idle_reduces, "Total number of idle reductions on input"); rack_tlp_does_nada = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "tlp_nada", CTLFLAG_RD, &rack_tlp_does_nada, "Total number of nada tlp calls"); COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "outsize", CTLFLAG_RD, rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes"); COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "opts", CTLFLAG_RD, rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats"); SYSCTL_ADD_PROC(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters"); } static inline int32_t rack_progress_timeout_check(struct tcpcb *tp) { if (tp->t_maxunacktime && tp->t_acktime && TSTMP_GT(ticks, tp->t_acktime)) { if ((ticks - tp->t_acktime) >= tp->t_maxunacktime) { /* * There is an assumption that the caller * will drop the connection so we will * increment the counters here. */ struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; counter_u64_add(rack_progress_drops, 1); #ifdef NETFLIX_STATS TCPSTAT_INC(tcps_progdrops); #endif rack_log_progress_event(rack, tp, ticks, PROGRESS_DROP, __LINE__); return (1); } } return (0); } static void rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = TICKS_2_MSEC(rack->rc_tp->t_srtt >> TCP_RTT_SHIFT); log.u_bbr.flex2 = to; log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags; log.u_bbr.flex4 = slot; log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot; log.u_bbr.flex6 = rack->rc_tp->t_rxtcur; log.u_bbr.flex8 = which; log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; TCP_LOG_EVENT(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_TIMERSTAR, 0, 0, &log, false); } } static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex8 = to_num; log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt; log.u_bbr.flex2 = rack->rc_rack_rtt; TCP_LOG_EVENT(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_RTO, 0, 0, &log, false); } } static void rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, int32_t t, uint32_t o_srtt, uint32_t o_var) { if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = t; log.u_bbr.flex2 = o_srtt; log.u_bbr.flex3 = o_var; log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest; log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest; log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_rtt_cnt; log.u_bbr.rttProp = rack->r_ctl.rack_rs.rs_rtt_tot; log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method; TCP_LOG_EVENT(tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_BBRRTT, 0, 0, &log, false); } } static void rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt) { /* * Log the rtt sample we are * applying to the srtt algorithm in * useconds. */ if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; /* Convert our ms to a microsecond */ log.u_bbr.flex1 = rtt * 1000; log.u_bbr.timeStamp = tcp_get_usecs(&tv); TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, TCP_LOG_RTT, 0, 0, &log, false, &tv); } } static inline void rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line) { if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = line; log.u_bbr.flex2 = tick; log.u_bbr.flex3 = tp->t_maxunacktime; log.u_bbr.flex4 = tp->t_acktime; log.u_bbr.flex8 = event; TCP_LOG_EVENT(tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_PROGRESS, 0, 0, &log, false); } } static void rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = slot; log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags); log.u_bbr.flex8 = rack->rc_in_persist; TCP_LOG_EVENT(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_BBRSND, 0, 0, &log, false); } } static void rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; log.u_bbr.flex1 = did_out; log.u_bbr.flex2 = nxt_pkt; log.u_bbr.flex3 = way_out; log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; log.u_bbr.flex7 = rack->r_wanted_output; log.u_bbr.flex8 = rack->rc_in_persist; TCP_LOG_EVENT(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_DOSEG_DONE, 0, 0, &log, false); } } static void rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot, uint8_t hpts_calling) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = slot; log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags; log.u_bbr.flex7 = hpts_calling; log.u_bbr.flex8 = rack->rc_in_persist; TCP_LOG_EVENT(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_JUSTRET, 0, tlen, &log, false); } } static void rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = line; log.u_bbr.flex2 = 0; log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags; log.u_bbr.flex4 = 0; log.u_bbr.flex6 = rack->rc_tp->t_rxtcur; log.u_bbr.flex8 = hpts_removed; TCP_LOG_EVENT(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_TIMERCANC, 0, 0, &log, false); } } static void rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = timers; log.u_bbr.flex2 = ret; log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp; log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; log.u_bbr.flex5 = cts; TCP_LOG_EVENT(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_TO_PROCESS, 0, 0, &log, false); } } static void rack_counter_destroy() { counter_u64_free(rack_badfr); counter_u64_free(rack_badfr_bytes); counter_u64_free(rack_rtm_prr_retran); counter_u64_free(rack_rtm_prr_newdata); counter_u64_free(rack_timestamp_mismatch); counter_u64_free(rack_reorder_seen); counter_u64_free(rack_tlp_tot); counter_u64_free(rack_tlp_newdata); counter_u64_free(rack_tlp_retran); counter_u64_free(rack_tlp_retran_bytes); counter_u64_free(rack_tlp_retran_fail); counter_u64_free(rack_to_tot); counter_u64_free(rack_to_arm_rack); counter_u64_free(rack_to_arm_tlp); counter_u64_free(rack_paced_segments); counter_u64_free(rack_unpaced_segments); counter_u64_free(rack_saw_enobuf); counter_u64_free(rack_saw_enetunreach); counter_u64_free(rack_to_alloc_hard); counter_u64_free(rack_to_alloc_emerg); counter_u64_free(rack_sack_proc_all); counter_u64_free(rack_sack_proc_short); counter_u64_free(rack_sack_proc_restart); counter_u64_free(rack_to_alloc); counter_u64_free(rack_find_high); counter_u64_free(rack_runt_sacks); counter_u64_free(rack_enter_tlp_calc); counter_u64_free(rack_used_tlpmethod); counter_u64_free(rack_used_tlpmethod2); counter_u64_free(rack_progress_drops); counter_u64_free(rack_input_idle_reduces); counter_u64_free(rack_tlp_does_nada); COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE); COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE); } static struct rack_sendmap * rack_alloc(struct tcp_rack *rack) { struct rack_sendmap *rsm; counter_u64_add(rack_to_alloc, 1); rack->r_ctl.rc_num_maps_alloced++; rsm = uma_zalloc(rack_zone, M_NOWAIT); if (rsm) { return (rsm); } if (rack->rc_free_cnt) { counter_u64_add(rack_to_alloc_emerg, 1); rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_next); rack->rc_free_cnt--; return (rsm); } return (NULL); } static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm) { rack->r_ctl.rc_num_maps_alloced--; if (rack->r_ctl.rc_tlpsend == rsm) rack->r_ctl.rc_tlpsend = NULL; if (rack->r_ctl.rc_next == rsm) rack->r_ctl.rc_next = NULL; if (rack->r_ctl.rc_sacklast == rsm) rack->r_ctl.rc_sacklast = NULL; if (rack->rc_free_cnt < rack_free_cache) { memset(rsm, 0, sizeof(struct rack_sendmap)); TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_next); rack->rc_free_cnt++; return; } uma_zfree(rack_zone, rsm); } /* * CC wrapper hook functions */ static void rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, struct tcphdr *th, uint16_t nsegs, uint16_t type, int32_t recovery) { #ifdef NETFLIX_STATS int32_t gput; #endif #ifdef NETFLIX_CWV u_long old_cwnd = tp->snd_cwnd; #endif INP_WLOCK_ASSERT(tp->t_inpcb); tp->ccv->nsegs = nsegs; tp->ccv->bytes_this_ack = BYTES_THIS_ACK(tp, th); if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) { uint32_t max; max = rack->r_ctl.rc_early_recovery_segs * tp->t_maxseg; if (tp->ccv->bytes_this_ack > max) { tp->ccv->bytes_this_ack = max; } } if (tp->snd_cwnd <= tp->snd_wnd) tp->ccv->flags |= CCF_CWND_LIMITED; else tp->ccv->flags &= ~CCF_CWND_LIMITED; if (type == CC_ACK) { #ifdef NETFLIX_STATS stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF, ((int32_t) tp->snd_cwnd) - tp->snd_wnd); if ((tp->t_flags & TF_GPUTINPROG) && SEQ_GEQ(th->th_ack, tp->gput_ack)) { gput = (((int64_t) (th->th_ack - tp->gput_seq)) << 3) / max(1, tcp_ts_getticks() - tp->gput_ts); stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT, gput); /* * XXXLAS: This is a temporary hack, and should be * chained off VOI_TCP_GPUT when stats(9) grows an * API to deal with chained VOIs. */ if (tp->t_stats_gput_prev > 0) stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_GPUT_ND, ((gput - tp->t_stats_gput_prev) * 100) / tp->t_stats_gput_prev); tp->t_flags &= ~TF_GPUTINPROG; tp->t_stats_gput_prev = gput; if (tp->t_maxpeakrate) { /* * We update t_peakrate_thr. This gives us roughly * one update per round trip time. */ tcp_update_peakrate_thr(tp); } } #endif if (tp->snd_cwnd > tp->snd_ssthresh) { tp->t_bytes_acked += min(tp->ccv->bytes_this_ack, nsegs * V_tcp_abc_l_var * tp->t_maxseg); if (tp->t_bytes_acked >= tp->snd_cwnd) { tp->t_bytes_acked -= tp->snd_cwnd; tp->ccv->flags |= CCF_ABC_SENTAWND; } } else { tp->ccv->flags &= ~CCF_ABC_SENTAWND; tp->t_bytes_acked = 0; } } if (CC_ALGO(tp)->ack_received != NULL) { /* XXXLAS: Find a way to live without this */ tp->ccv->curack = th->th_ack; CC_ALGO(tp)->ack_received(tp->ccv, type); } #ifdef NETFLIX_STATS stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, tp->snd_cwnd); #endif if (rack->r_ctl.rc_rack_largest_cwnd < tp->snd_cwnd) { rack->r_ctl.rc_rack_largest_cwnd = tp->snd_cwnd; } #ifdef NETFLIX_CWV if (tp->cwv_enabled) { /* * Per RFC 7661: The behaviour in the non-validated phase is * specified as: o A sender determines whether to increase * the cwnd based upon whether it is cwnd-limited (see * Section 4.5.3): * A sender that is cwnd-limited MAY use * the standard TCP method to increase cwnd (i.e., the * standard method permits a TCP sender that fully utilises * the cwnd to increase the cwnd each time it receives an * ACK). * A sender that is not cwnd-limited MUST NOT * increase the cwnd when ACK packets are received in this * phase (i.e., needs to avoid growing the cwnd when it has * not recently sent using the current size of cwnd). */ if ((tp->snd_cwnd > old_cwnd) && (tp->cwv_cwnd_valid == 0) && (!(tp->ccv->flags & CCF_CWND_LIMITED))) { tp->snd_cwnd = old_cwnd; } /* Try to update pipeAck and NCWV state */ if (TCPS_HAVEESTABLISHED(tp->t_state) && !IN_RECOVERY(tp->t_flags)) { uint32_t data = sbavail(&(tp->t_inpcb->inp_socket->so_snd)); tcp_newcwv_update_pipeack(tp, data); } } #endif /* we enforce max peak rate if it is set. */ if (tp->t_peakrate_thr && tp->snd_cwnd > tp->t_peakrate_thr) { tp->snd_cwnd = tp->t_peakrate_thr; } } static void tcp_rack_partialack(struct tcpcb *tp, struct tcphdr *th) { struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; INP_WLOCK_ASSERT(tp->t_inpcb); if (rack->r_ctl.rc_prr_sndcnt > 0) rack->r_wanted_output++; } static void rack_post_recovery(struct tcpcb *tp, struct tcphdr *th) { struct tcp_rack *rack; INP_WLOCK_ASSERT(tp->t_inpcb); rack = (struct tcp_rack *)tp->t_fb_ptr; if (CC_ALGO(tp)->post_recovery != NULL) { tp->ccv->curack = th->th_ack; CC_ALGO(tp)->post_recovery(tp->ccv); } /* * Here we can in theory adjust cwnd to be based on the number of * losses in the window (rack->r_ctl.rc_loss_count). This is done * based on the rack_use_proportional flag. */ if (rack->r_ctl.rc_prop_reduce && rack->r_ctl.rc_prop_rate) { int32_t reduce; reduce = (rack->r_ctl.rc_loss_count * rack->r_ctl.rc_prop_rate); if (reduce > 50) { reduce = 50; } tp->snd_cwnd -= ((reduce * tp->snd_cwnd) / 100); } else { if (tp->snd_cwnd > tp->snd_ssthresh) { /* Drop us down to the ssthresh (1/2 cwnd at loss) */ tp->snd_cwnd = tp->snd_ssthresh; } } if (rack->r_ctl.rc_prr_sndcnt > 0) { /* Suck the next prr cnt back into cwnd */ tp->snd_cwnd += rack->r_ctl.rc_prr_sndcnt; rack->r_ctl.rc_prr_sndcnt = 0; } EXIT_RECOVERY(tp->t_flags); #ifdef NETFLIX_CWV if (tp->cwv_enabled) { if ((tp->cwv_cwnd_valid == 0) && (tp->snd_cwv.in_recovery)) tcp_newcwv_end_recovery(tp); } #endif } static void rack_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type) { struct tcp_rack *rack; INP_WLOCK_ASSERT(tp->t_inpcb); rack = (struct tcp_rack *)tp->t_fb_ptr; switch (type) { case CC_NDUPACK: /* rack->r_ctl.rc_ssthresh_set = 1;*/ if (!IN_FASTRECOVERY(tp->t_flags)) { rack->r_ctl.rc_tlp_rtx_out = 0; rack->r_ctl.rc_prr_delivered = 0; rack->r_ctl.rc_prr_out = 0; rack->r_ctl.rc_loss_count = 0; rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una; tp->snd_recover = tp->snd_max; if (tp->t_flags & TF_ECN_PERMIT) tp->t_flags |= TF_ECN_SND_CWR; } break; case CC_ECN: if (!IN_CONGRECOVERY(tp->t_flags)) { TCPSTAT_INC(tcps_ecn_rcwnd); tp->snd_recover = tp->snd_max; if (tp->t_flags & TF_ECN_PERMIT) tp->t_flags |= TF_ECN_SND_CWR; } break; case CC_RTO: tp->t_dupacks = 0; tp->t_bytes_acked = 0; EXIT_RECOVERY(tp->t_flags); tp->snd_ssthresh = max(2, min(tp->snd_wnd, tp->snd_cwnd) / 2 / tp->t_maxseg) * tp->t_maxseg; tp->snd_cwnd = tp->t_maxseg; break; case CC_RTO_ERR: TCPSTAT_INC(tcps_sndrexmitbad); /* RTO was unnecessary, so reset everything. */ tp->snd_cwnd = tp->snd_cwnd_prev; tp->snd_ssthresh = tp->snd_ssthresh_prev; tp->snd_recover = tp->snd_recover_prev; if (tp->t_flags & TF_WASFRECOVERY) ENTER_FASTRECOVERY(tp->t_flags); if (tp->t_flags & TF_WASCRECOVERY) ENTER_CONGRECOVERY(tp->t_flags); tp->snd_nxt = tp->snd_max; tp->t_badrxtwin = 0; break; } if (CC_ALGO(tp)->cong_signal != NULL) { if (th != NULL) tp->ccv->curack = th->th_ack; CC_ALGO(tp)->cong_signal(tp->ccv, type); } #ifdef NETFLIX_CWV if (tp->cwv_enabled) { if (tp->snd_cwv.in_recovery == 0 && IN_RECOVERY(tp->t_flags)) { tcp_newcwv_enter_recovery(tp); } if (type == CC_RTO) { tcp_newcwv_reset(tp); } } #endif } static inline void rack_cc_after_idle(struct tcpcb *tp, int reduce_largest) { uint32_t i_cwnd; INP_WLOCK_ASSERT(tp->t_inpcb); #ifdef NETFLIX_STATS TCPSTAT_INC(tcps_idle_restarts); if (tp->t_state == TCPS_ESTABLISHED) TCPSTAT_INC(tcps_idle_estrestarts); #endif if (CC_ALGO(tp)->after_idle != NULL) CC_ALGO(tp)->after_idle(tp->ccv); if (tp->snd_cwnd == 1) i_cwnd = tp->t_maxseg; /* SYN(-ACK) lost */ else if (V_tcp_initcwnd_segments) i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg), max(2 * tp->t_maxseg, V_tcp_initcwnd_segments * 1460)); else if (V_tcp_do_rfc3390) i_cwnd = min(4 * tp->t_maxseg, max(2 * tp->t_maxseg, 4380)); else { /* Per RFC5681 Section 3.1 */ if (tp->t_maxseg > 2190) i_cwnd = 2 * tp->t_maxseg; else if (tp->t_maxseg > 1095) i_cwnd = 3 * tp->t_maxseg; else i_cwnd = 4 * tp->t_maxseg; } if (reduce_largest) { /* * Do we reduce the largest cwnd to make * rack play nice on restart hptsi wise? */ if (((struct tcp_rack *)tp->t_fb_ptr)->r_ctl.rc_rack_largest_cwnd > i_cwnd) ((struct tcp_rack *)tp->t_fb_ptr)->r_ctl.rc_rack_largest_cwnd = i_cwnd; } /* * Being idle is no differnt than the initial window. If the cc * clamps it down below the initial window raise it to the initial * window. */ if (tp->snd_cwnd < i_cwnd) { tp->snd_cwnd = i_cwnd; } } /* * Indicate whether this ack should be delayed. We can delay the ack if * following conditions are met: * - There is no delayed ack timer in progress. * - Our last ack wasn't a 0-sized window. We never want to delay * the ack that opens up a 0-sized window. * - LRO wasn't used for this segment. We make sure by checking that the * segment size is not larger than the MSS. * - Delayed acks are enabled or this is a half-synchronized T/TCP * connection. */ #define DELAY_ACK(tp, tlen) \ (((tp->t_flags & TF_RXWIN0SENT) == 0) && \ ((tp->t_flags & TF_DELACK) == 0) && \ (tlen <= tp->t_maxseg) && \ (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN))) static inline void rack_calc_rwin(struct socket *so, struct tcpcb *tp) { int32_t win; /* * Calculate amount of space in receive window, and then do TCP * input processing. Receive window is amount of space in rcv queue, * but not less than advertised window. */ win = sbspace(&so->so_rcv); if (win < 0) win = 0; tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt)); } static void rack_do_drop(struct mbuf *m, struct tcpcb *tp, int32_t * ti_locked) { if (*ti_locked == TI_RLOCKED) { INP_INFO_RUNLOCK(&V_tcbinfo); *ti_locked = TI_UNLOCKED; } /* * Drop space held by incoming segment and return. */ if (tp != NULL) INP_WUNLOCK(tp->t_inpcb); if (m) m_freem(m); } static void rack_do_dropwithreset(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, int32_t * ti_locked, int32_t rstreason, int32_t tlen) { if (*ti_locked == TI_RLOCKED) { INP_INFO_RUNLOCK(&V_tcbinfo); *ti_locked = TI_UNLOCKED; } if (tp != NULL) { tcp_dropwithreset(m, th, tp, tlen, rstreason); INP_WUNLOCK(tp->t_inpcb); } else tcp_dropwithreset(m, th, NULL, tlen, rstreason); } /* * The value in ret_val informs the caller * if we dropped the tcb (and lock) or not. * 1 = we dropped it, 0 = the TCB is still locked * and valid. */ static void rack_do_dropafterack(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, int32_t * ti_locked, int32_t thflags, int32_t tlen, int32_t * ret_val) { /* * Generate an ACK dropping incoming segment if it occupies sequence * space, where the ACK reflects our state. * * We can now skip the test for the RST flag since all paths to this * code happen after packets containing RST have been dropped. * * In the SYN-RECEIVED state, don't send an ACK unless the segment * we received passes the SYN-RECEIVED ACK test. If it fails send a * RST. This breaks the loop in the "LAND" DoS attack, and also * prevents an ACK storm between two listening ports that have been * sent forged SYN segments, each with the source address of the * other. */ struct tcp_rack *rack; if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) && (SEQ_GT(tp->snd_una, th->th_ack) || SEQ_GT(th->th_ack, tp->snd_max))) { *ret_val = 1; rack_do_dropwithreset(m, tp, th, ti_locked, BANDLIM_RST_OPENPORT, tlen); return; } else *ret_val = 0; if (*ti_locked == TI_RLOCKED) { INP_INFO_RUNLOCK(&V_tcbinfo); *ti_locked = TI_UNLOCKED; } rack = (struct tcp_rack *)tp->t_fb_ptr; rack->r_wanted_output++; tp->t_flags |= TF_ACKNOW; if (m) m_freem(m); } static int rack_process_rst(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t * ti_locked) { /* * RFC5961 Section 3.2 * * - RST drops connection only if SEG.SEQ == RCV.NXT. - If RST is in * window, we send challenge ACK. * * Note: to take into account delayed ACKs, we should test against * last_ack_sent instead of rcv_nxt. Note 2: we handle special case * of closed window, not covered by the RFC. */ int dropped = 0; if ((SEQ_GEQ(th->th_seq, (tp->last_ack_sent - 1)) && SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) || (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq)) { INP_INFO_RLOCK_ASSERT(&V_tcbinfo); KASSERT(*ti_locked == TI_RLOCKED, ("%s: TH_RST ti_locked %d, th %p tp %p", __func__, *ti_locked, th, tp)); KASSERT(tp->t_state != TCPS_SYN_SENT, ("%s: TH_RST for TCPS_SYN_SENT th %p tp %p", __func__, th, tp)); if (V_tcp_insecure_rst || (tp->last_ack_sent == th->th_seq) || (tp->rcv_nxt == th->th_seq) || ((tp->last_ack_sent - 1) == th->th_seq)) { TCPSTAT_INC(tcps_drops); /* Drop the connection. */ switch (tp->t_state) { case TCPS_SYN_RECEIVED: so->so_error = ECONNREFUSED; goto close; case TCPS_ESTABLISHED: case TCPS_FIN_WAIT_1: case TCPS_FIN_WAIT_2: case TCPS_CLOSE_WAIT: case TCPS_CLOSING: case TCPS_LAST_ACK: so->so_error = ECONNRESET; close: tcp_state_change(tp, TCPS_CLOSED); /* FALLTHROUGH */ default: tp = tcp_close(tp); } dropped = 1; rack_do_drop(m, tp, ti_locked); } else { TCPSTAT_INC(tcps_badrst); /* Send challenge ACK. */ tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt, tp->snd_nxt, TH_ACK); tp->last_ack_sent = tp->rcv_nxt; } } else { m_freem(m); } return (dropped); } /* * The value in ret_val informs the caller * if we dropped the tcb (and lock) or not. * 1 = we dropped it, 0 = the TCB is still locked * and valid. */ static void rack_challenge_ack(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t * ti_locked, int32_t * ret_val) { KASSERT(*ti_locked == TI_RLOCKED, ("tcp_do_segment: TH_SYN ti_locked %d", *ti_locked)); INP_INFO_RLOCK_ASSERT(&V_tcbinfo); TCPSTAT_INC(tcps_badsyn); if (V_tcp_insecure_syn && SEQ_GEQ(th->th_seq, tp->last_ack_sent) && SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) { tp = tcp_drop(tp, ECONNRESET); *ret_val = 1; rack_do_drop(m, tp, ti_locked); } else { /* Send challenge ACK. */ tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt, tp->snd_nxt, TH_ACK); tp->last_ack_sent = tp->rcv_nxt; m = NULL; *ret_val = 0; rack_do_drop(m, NULL, ti_locked); } } /* * rack_ts_check returns 1 for you should not proceed. It places * in ret_val what should be returned 1/0 by the caller. The 1 indicates * that the TCB is unlocked and probably dropped. The 0 indicates the * TCB is still valid and locked. */ static int rack_ts_check(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t * ti_locked, int32_t tlen, int32_t thflags, int32_t * ret_val) { /* Check to see if ts_recent is over 24 days old. */ if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) { /* * Invalidate ts_recent. If this segment updates ts_recent, * the age will be reset later and ts_recent will get a * valid value. If it does not, setting ts_recent to zero * will at least satisfy the requirement that zero be placed * in the timestamp echo reply when ts_recent isn't valid. * The age isn't reset until we get a valid ts_recent * because we don't want out-of-order segments to be dropped * when ts_recent is old. */ tp->ts_recent = 0; } else { TCPSTAT_INC(tcps_rcvduppack); TCPSTAT_ADD(tcps_rcvdupbyte, tlen); TCPSTAT_INC(tcps_pawsdrop); *ret_val = 0; if (tlen) { rack_do_dropafterack(m, tp, th, ti_locked, thflags, tlen, ret_val); } else { rack_do_drop(m, NULL, ti_locked); } return (1); } return (0); } /* * rack_drop_checks returns 1 for you should not proceed. It places * in ret_val what should be returned 1/0 by the caller. The 1 indicates * that the TCB is unlocked and probably dropped. The 0 indicates the * TCB is still valid and locked. */ static int rack_drop_checks(struct tcpopt *to, struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t * tlenp, int32_t * ti_locked, int32_t * thf, int32_t * drop_hdrlen, int32_t * ret_val) { int32_t todrop; int32_t thflags; int32_t tlen; thflags = *thf; tlen = *tlenp; todrop = tp->rcv_nxt - th->th_seq; if (todrop > 0) { if (thflags & TH_SYN) { thflags &= ~TH_SYN; th->th_seq++; if (th->th_urp > 1) th->th_urp--; else thflags &= ~TH_URG; todrop--; } /* * Following if statement from Stevens, vol. 2, p. 960. */ if (todrop > tlen || (todrop == tlen && (thflags & TH_FIN) == 0)) { /* * Any valid FIN must be to the left of the window. * At this point the FIN must be a duplicate or out * of sequence; drop it. */ thflags &= ~TH_FIN; /* * Send an ACK to resynchronize and drop any data. * But keep on processing for RST or ACK. */ tp->t_flags |= TF_ACKNOW; todrop = tlen; TCPSTAT_INC(tcps_rcvduppack); TCPSTAT_ADD(tcps_rcvdupbyte, todrop); } else { TCPSTAT_INC(tcps_rcvpartduppack); TCPSTAT_ADD(tcps_rcvpartdupbyte, todrop); } *drop_hdrlen += todrop; /* drop from the top afterwards */ th->th_seq += todrop; tlen -= todrop; if (th->th_urp > todrop) th->th_urp -= todrop; else { thflags &= ~TH_URG; th->th_urp = 0; } } /* * If segment ends after window, drop trailing data (and PUSH and * FIN); if nothing left, just ACK. */ todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd); if (todrop > 0) { TCPSTAT_INC(tcps_rcvpackafterwin); if (todrop >= tlen) { TCPSTAT_ADD(tcps_rcvbyteafterwin, tlen); /* * If window is closed can only take segments at * window edge, and have to drop data and PUSH from * incoming segments. Continue processing, but * remember to ack. Otherwise, drop segment and * ack. */ if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) { tp->t_flags |= TF_ACKNOW; TCPSTAT_INC(tcps_rcvwinprobe); } else { rack_do_dropafterack(m, tp, th, ti_locked, thflags, tlen, ret_val); return (1); } } else TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); m_adj(m, -todrop); tlen -= todrop; thflags &= ~(TH_PUSH | TH_FIN); } *thf = thflags; *tlenp = tlen; return (0); } static struct rack_sendmap * rack_find_lowest_rsm(struct tcp_rack *rack) { struct rack_sendmap *rsm; /* * Walk the time-order transmitted list looking for an rsm that is * not acked. This will be the one that was sent the longest time * ago that is still outstanding. */ TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) { if (rsm->r_flags & RACK_ACKED) { continue; } goto finish; } finish: return (rsm); } static struct rack_sendmap * rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm) { struct rack_sendmap *prsm; /* * Walk the sequence order list backward until we hit and arrive at * the highest seq not acked. In theory when this is called it * should be the last segment (which it was not). */ counter_u64_add(rack_find_high, 1); prsm = rsm; TAILQ_FOREACH_REVERSE_FROM(prsm, &rack->r_ctl.rc_map, rack_head, r_next) { if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) { continue; } return (prsm); } return (NULL); } static uint32_t rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts) { int32_t lro; uint32_t thresh; /* * lro is the flag we use to determine if we have seen reordering. * If it gets set we have seen reordering. The reorder logic either * works in one of two ways: * * If reorder-fade is configured, then we track the last time we saw * re-ordering occur. If we reach the point where enough time as * passed we no longer consider reordering has occuring. * * Or if reorder-face is 0, then once we see reordering we consider * the connection to alway be subject to reordering and just set lro * to 1. * * In the end if lro is non-zero we add the extra time for * reordering in. */ if (srtt == 0) srtt = 1; if (rack->r_ctl.rc_reorder_ts) { if (rack->r_ctl.rc_reorder_fade) { if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) { lro = cts - rack->r_ctl.rc_reorder_ts; if (lro == 0) { /* * No time as passed since the last * reorder, mark it as reordering. */ lro = 1; } } else { /* Negative time? */ lro = 0; } if (lro > rack->r_ctl.rc_reorder_fade) { /* Turn off reordering seen too */ rack->r_ctl.rc_reorder_ts = 0; lro = 0; } } else { /* Reodering does not fade */ lro = 1; } } else { lro = 0; } thresh = srtt + rack->r_ctl.rc_pkt_delay; if (lro) { /* It must be set, if not you get 1/4 rtt */ if (rack->r_ctl.rc_reorder_shift) thresh += (srtt >> rack->r_ctl.rc_reorder_shift); else thresh += (srtt >> 2); } else { thresh += 1; } /* We don't let the rack timeout be above a RTO */ if (thresh > TICKS_2_MSEC(rack->rc_tp->t_rxtcur)) { thresh = TICKS_2_MSEC(rack->rc_tp->t_rxtcur); } /* And we don't want it above the RTO max either */ if (thresh > rack_rto_max) { thresh = rack_rto_max; } return (thresh); } static uint32_t rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t srtt) { struct rack_sendmap *prsm; uint32_t thresh, len; int maxseg; if (srtt == 0) srtt = 1; if (rack->r_ctl.rc_tlp_threshold) thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold); else thresh = (srtt * 2); /* Get the previous sent packet, if any */ maxseg = tcp_maxseg(tp); counter_u64_add(rack_enter_tlp_calc, 1); len = rsm->r_end - rsm->r_start; if (rack->rack_tlp_threshold_use == TLP_USE_ID) { /* Exactly like the ID */ if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= maxseg) { uint32_t alt_thresh; /* * Compensate for delayed-ack with the d-ack time. */ counter_u64_add(rack_used_tlpmethod, 1); alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; if (alt_thresh > thresh) thresh = alt_thresh; } } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) { /* 2.1 behavior */ prsm = TAILQ_PREV(rsm, rack_head, r_tnext); if (prsm && (len <= maxseg)) { /* * Two packets outstanding, thresh should be (2*srtt) + * possible inter-packet delay (if any). */ uint32_t inter_gap = 0; int idx, nidx; counter_u64_add(rack_used_tlpmethod, 1); idx = rsm->r_rtr_cnt - 1; nidx = prsm->r_rtr_cnt - 1; if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) { /* Yes it was sent later (or at the same time) */ inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx]; } thresh += inter_gap; } else if (len <= maxseg) { /* * Possibly compensate for delayed-ack. */ uint32_t alt_thresh; counter_u64_add(rack_used_tlpmethod2, 1); alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; if (alt_thresh > thresh) thresh = alt_thresh; } } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) { /* 2.2 behavior */ if (len <= maxseg) { uint32_t alt_thresh; /* * Compensate for delayed-ack with the d-ack time. */ counter_u64_add(rack_used_tlpmethod, 1); alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; if (alt_thresh > thresh) thresh = alt_thresh; } } /* Not above an RTO */ if (thresh > TICKS_2_MSEC(tp->t_rxtcur)) { thresh = TICKS_2_MSEC(tp->t_rxtcur); } /* Not above a RTO max */ if (thresh > rack_rto_max) { thresh = rack_rto_max; } /* Apply user supplied min TLP */ if (thresh < rack_tlp_min) { thresh = rack_tlp_min; } return (thresh); } static struct rack_sendmap * rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused) { /* * Check to see that we don't need to fall into recovery. We will * need to do so if our oldest transmit is past the time we should * have had an ack. */ struct tcp_rack *rack; struct rack_sendmap *rsm; int32_t idx; uint32_t srtt_cur, srtt, thresh; rack = (struct tcp_rack *)tp->t_fb_ptr; if (TAILQ_EMPTY(&rack->r_ctl.rc_map)) { return (NULL); } srtt_cur = tp->t_srtt >> TCP_RTT_SHIFT; srtt = TICKS_2_MSEC(srtt_cur); if (rack->rc_rack_rtt && (srtt > rack->rc_rack_rtt)) srtt = rack->rc_rack_rtt; rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (rsm == NULL) return (NULL); if (rsm->r_flags & RACK_ACKED) { rsm = rack_find_lowest_rsm(rack); if (rsm == NULL) return (NULL); } idx = rsm->r_rtr_cnt - 1; thresh = rack_calc_thresh_rack(rack, srtt, tsused); if (tsused < rsm->r_tim_lastsent[idx]) { return (NULL); } if ((tsused - rsm->r_tim_lastsent[idx]) < thresh) { return (NULL); } /* Ok if we reach here we are over-due */ rack->r_ctl.rc_rsm_start = rsm->r_start; rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; rack_cong_signal(tp, NULL, CC_NDUPACK); return (rsm); } static uint32_t rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack) { int32_t t; int32_t tt; uint32_t ret_val; t = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT) + ((tp->t_rttvar * 4) >> TCP_RTT_SHIFT)); TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], tcp_persmin, tcp_persmax); if (tp->t_rxtshift < TCP_MAXRXTSHIFT) tp->t_rxtshift++; rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT; ret_val = (uint32_t)tt; return (ret_val); } static uint32_t rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { /* * Start the FR timer, we do this based on getting the first one in * the rc_tmap. Note that if its NULL we must stop the timer. in all * events we need to stop the running timer (if its running) before * starting the new one. */ uint32_t thresh, exp, to, srtt, time_since_sent; uint32_t srtt_cur; int32_t idx; int32_t is_tlp_timer = 0; struct rack_sendmap *rsm; if (rack->t_timers_stopped) { /* All timers have been stopped none are to run */ return (0); } if (rack->rc_in_persist) { /* We can't start any timer in persists */ return (rack_get_persists_timer_val(tp, rack)); } if (tp->t_state < TCPS_ESTABLISHED) goto activate_rxt; rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (rsm == NULL) { /* Nothing on the send map */ activate_rxt: if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) { rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT; to = TICKS_2_MSEC(tp->t_rxtcur); if (to == 0) to = 1; return (to); } return (0); } if (rsm->r_flags & RACK_ACKED) { rsm = rack_find_lowest_rsm(rack); if (rsm == NULL) { /* No lowest? */ goto activate_rxt; } } /* Convert from ms to usecs */ if (rsm->r_flags & RACK_SACK_PASSED) { if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) == 1) && (rsm->r_flags & RACK_HAS_FIN)) { /* * We don't start a rack timer if all we have is a * FIN outstanding. */ goto activate_rxt; } if (tp->t_srtt) { srtt_cur = (tp->t_srtt >> TCP_RTT_SHIFT); srtt = TICKS_2_MSEC(srtt_cur); } else srtt = RACK_INITIAL_RTO; thresh = rack_calc_thresh_rack(rack, srtt, cts); idx = rsm->r_rtr_cnt - 1; exp = rsm->r_tim_lastsent[idx] + thresh; if (SEQ_GEQ(exp, cts)) { to = exp - cts; if (to < rack->r_ctl.rc_min_to) { to = rack->r_ctl.rc_min_to; } } else { to = rack->r_ctl.rc_min_to; } } else { /* Ok we need to do a TLP not RACK */ if ((rack->rc_tlp_in_progress != 0) || (rack->r_ctl.rc_tlp_rtx_out != 0)) { /* * The previous send was a TLP or a tlp_rtx is in * process. */ goto activate_rxt; } rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext); if (rsm == NULL) { /* We found no rsm to TLP with. */ goto activate_rxt; } if (rsm->r_flags & RACK_HAS_FIN) { /* If its a FIN we dont do TLP */ rsm = NULL; goto activate_rxt; } idx = rsm->r_rtr_cnt - 1; if (TSTMP_GT(cts, rsm->r_tim_lastsent[idx])) time_since_sent = cts - rsm->r_tim_lastsent[idx]; else time_since_sent = 0; is_tlp_timer = 1; if (tp->t_srtt) { srtt_cur = (tp->t_srtt >> TCP_RTT_SHIFT); srtt = TICKS_2_MSEC(srtt_cur); } else srtt = RACK_INITIAL_RTO; thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt); if (thresh > time_since_sent) to = thresh - time_since_sent; else to = rack->r_ctl.rc_min_to; if (to > TCPTV_REXMTMAX) { /* * If the TLP time works out to larger than the max * RTO lets not do TLP.. just RTO. */ goto activate_rxt; } if (rsm->r_start != rack->r_ctl.rc_last_tlp_seq) { /* * The tail is no longer the last one I did a probe * on */ rack->r_ctl.rc_tlp_seg_send_cnt = 0; rack->r_ctl.rc_last_tlp_seq = rsm->r_start; } } if (is_tlp_timer == 0) { rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK; } else { if ((rack->r_ctl.rc_tlp_send_cnt > rack_tlp_max_resend) || (rack->r_ctl.rc_tlp_seg_send_cnt > rack_tlp_max_resend)) { /* * We have exceeded how many times we can retran the * current TLP timer, switch to the RTO timer. */ goto activate_rxt; } else { rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP; } } if (to == 0) to = 1; return (to); } static void rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { if (rack->rc_in_persist == 0) { if (((tp->t_flags & TF_SENTFIN) == 0) && (tp->snd_max - tp->snd_una) >= sbavail(&rack->rc_inp->inp_socket->so_snd)) /* Must need to send more data to enter persist */ return; rack->r_ctl.rc_went_idle_time = cts; rack_timer_cancel(tp, rack, cts, __LINE__); tp->t_rxtshift = 0; rack->rc_in_persist = 1; } } static void rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack) { if (rack->rc_inp->inp_in_hpts) { tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); rack->r_ctl.rc_hpts_flags = 0; } rack->rc_in_persist = 0; rack->r_ctl.rc_went_idle_time = 0; tp->t_flags &= ~TF_FORCEDATA; tp->t_rxtshift = 0; } static void rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts, int32_t line, int32_t slot, uint32_t tot_len_this_send, int32_t frm_out_sbavail) { struct inpcb *inp; uint32_t delayed_ack = 0; uint32_t hpts_timeout; uint8_t stopped; uint32_t left = 0; inp = tp->t_inpcb; if (inp->inp_in_hpts) { /* A previous call is already set up */ return; } if (tp->t_state == TCPS_CLOSED) { return; } stopped = rack->rc_tmr_stopped; if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) { left = rack->r_ctl.rc_timer_exp - cts; } rack->r_ctl.rc_timer_exp = 0; if (rack->rc_inp->inp_in_hpts == 0) { rack->r_ctl.rc_hpts_flags = 0; } if (slot) { /* We are hptsi too */ rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT; } else if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { /* * We are still left on the hpts when the to goes * it will be for output. */ if (TSTMP_GT(cts, rack->r_ctl.rc_last_output_to)) slot = cts - rack->r_ctl.rc_last_output_to; else slot = 1; } if ((tp->snd_wnd == 0) && TCPS_HAVEESTABLISHED(tp->t_state)) { /* No send window.. we must enter persist */ rack_enter_persist(tp, rack, cts); } else if ((frm_out_sbavail && (frm_out_sbavail > (tp->snd_max - tp->snd_una)) && (tp->snd_wnd < tp->t_maxseg)) && TCPS_HAVEESTABLISHED(tp->t_state)) { /* * If we have no window or we can't send a segment (and have * data to send.. we cheat here and frm_out_sbavail is * passed in with the sbavail(sb) only from bbr_output) and * we are established, then we must enter persits (if not * already in persits). */ rack_enter_persist(tp, rack, cts); } hpts_timeout = rack_timer_start(tp, rack, cts); if (tp->t_flags & TF_DELACK) { delayed_ack = tcp_delacktime; rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK; } if (delayed_ack && ((hpts_timeout == 0) || (delayed_ack < hpts_timeout))) hpts_timeout = delayed_ack; else rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; /* * If no timers are going to run and we will fall off the hptsi * wheel, we resort to a keep-alive timer if its configured. */ if ((hpts_timeout == 0) && (slot == 0)) { if ((tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && (tp->t_state <= TCPS_CLOSING)) { /* * Ok we have no timer (persists, rack, tlp, rxt or * del-ack), we don't have segments being paced. So * all that is left is the keepalive timer. */ if (TCPS_HAVEESTABLISHED(tp->t_state)) { /* Get the established keep-alive time */ hpts_timeout = TP_KEEPIDLE(tp); } else { /* Get the initial setup keep-alive time */ hpts_timeout = TP_KEEPINIT(tp); } rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP; } } if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) == (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) { /* * RACK, TLP, persists and RXT timers all are restartable * based on actions input .. i.e we received a packet (ack * or sack) and that changes things (rw, or snd_una etc). * Thus we can restart them with a new value. For * keep-alive, delayed_ack we keep track of what was left * and restart the timer with a smaller value. */ if (left < hpts_timeout) hpts_timeout = left; } if (hpts_timeout) { /* * Hack alert for now we can't time-out over 2,147,483 * seconds (a bit more than 596 hours), which is probably ok * :). */ if (hpts_timeout > 0x7ffffffe) hpts_timeout = 0x7ffffffe; rack->r_ctl.rc_timer_exp = cts + hpts_timeout; } if (slot) { rack->r_ctl.rc_last_output_to = cts + slot; if ((hpts_timeout == 0) || (hpts_timeout > slot)) { if (rack->rc_inp->inp_in_hpts == 0) tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(slot)); rack_log_to_start(rack, cts, hpts_timeout, slot, 1); } else { /* * Arrange for the hpts to kick back in after the * t-o if the t-o does not cause a send. */ if (rack->rc_inp->inp_in_hpts == 0) tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout)); rack_log_to_start(rack, cts, hpts_timeout, slot, 0); } } else if (hpts_timeout) { if (rack->rc_inp->inp_in_hpts == 0) tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout)); rack_log_to_start(rack, cts, hpts_timeout, slot, 0); } else { /* No timer starting */ #ifdef INVARIANTS if (SEQ_GT(tp->snd_max, tp->snd_una)) { panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?", tp, rack, tot_len_this_send, cts, slot, hpts_timeout); } #endif } rack->rc_tmr_stopped = 0; if (slot) rack_log_type_bbrsnd(rack, tot_len_this_send, slot, cts); } /* * RACK Timer, here we simply do logging and house keeping. * the normal rack_output() function will call the * appropriate thing to check if we need to do a RACK retransmit. * We return 1, saying don't proceed with rack_output only * when all timers have been stopped (destroyed PCB?). */ static int rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { /* * This timer simply provides an internal trigger to send out data. * The check_recovery_mode call will see if there are needed * retransmissions, if so we will enter fast-recovery. The output * call may or may not do the same thing depending on sysctl * settings. */ struct rack_sendmap *rsm; int32_t recovery; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { /* Its not time yet */ return (0); } rack_log_to_event(rack, RACK_TO_FRM_RACK); recovery = IN_RECOVERY(tp->t_flags); counter_u64_add(rack_to_tot, 1); if (rack->r_state && (rack->r_state != tp->t_state)) rack_set_state(tp, rack); rsm = rack_check_recovery_mode(tp, cts); if (rsm) { uint32_t rtt; rtt = rack->rc_rack_rtt; if (rtt == 0) rtt = 1; if ((recovery == 0) && (rack->r_ctl.rc_prr_sndcnt < tp->t_maxseg)) { /* * The rack-timeout that enter's us into recovery * will force out one MSS and set us up so that we * can do one more send in 2*rtt (transitioning the * rack timeout into a rack-tlp). */ rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; } else if ((rack->r_ctl.rc_prr_sndcnt < tp->t_maxseg) && ((rsm->r_end - rsm->r_start) > rack->r_ctl.rc_prr_sndcnt)) { /* * When a rack timer goes, we have to send at * least one segment. They will be paced a min of 1ms * apart via the next rack timer (or further * if the rack timer dictates it). */ rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; } } else { /* This is a case that should happen rarely if ever */ counter_u64_add(rack_tlp_does_nada, 1); #ifdef TCP_BLACKBOX tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true); #endif rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_tmap); } rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK; return (0); } /* * TLP Timer, here we simply setup what segment we want to * have the TLP expire on, the normal rack_output() will then * send it out. * * We return 1, saying don't proceed with rack_output only * when all timers have been stopped (destroyed PCB?). */ static int rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { /* * Tail Loss Probe. */ struct rack_sendmap *rsm = NULL; struct socket *so; uint32_t amm, old_prr_snd = 0; uint32_t out, avail; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { /* Its not time yet */ return (0); } if (rack_progress_timeout_check(tp)) { tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); return (1); } /* * A TLP timer has expired. We have been idle for 2 rtts. So we now * need to figure out how to force a full MSS segment out. */ rack_log_to_event(rack, RACK_TO_FRM_TLP); counter_u64_add(rack_tlp_tot, 1); if (rack->r_state && (rack->r_state != tp->t_state)) rack_set_state(tp, rack); so = tp->t_inpcb->inp_socket; avail = sbavail(&so->so_snd); out = tp->snd_max - tp->snd_una; rack->rc_timer_up = 1; /* * If we are in recovery we can jazz out a segment if new data is * present simply by setting rc_prr_sndcnt to a segment. */ if ((avail > out) && ((rack_always_send_oldest == 0) || (TAILQ_EMPTY(&rack->r_ctl.rc_tmap)))) { /* New data is available */ amm = avail - out; if (amm > tp->t_maxseg) { amm = tp->t_maxseg; } else if ((amm < tp->t_maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) { /* not enough to fill a MTU and no-delay is off */ goto need_retran; } if (IN_RECOVERY(tp->t_flags)) { /* Unlikely */ old_prr_snd = rack->r_ctl.rc_prr_sndcnt; if (out + amm <= tp->snd_wnd) rack->r_ctl.rc_prr_sndcnt = amm; else goto need_retran; } else { /* Set the send-new override */ if (out + amm <= tp->snd_wnd) rack->r_ctl.rc_tlp_new_data = amm; else goto need_retran; } rack->r_ctl.rc_tlp_seg_send_cnt = 0; rack->r_ctl.rc_last_tlp_seq = tp->snd_max; rack->r_ctl.rc_tlpsend = NULL; counter_u64_add(rack_tlp_newdata, 1); goto send; } need_retran: /* * Ok we need to arrange the last un-acked segment to be re-sent, or * optionally the first un-acked segment. */ if (rack_always_send_oldest) rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); else { rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_map, rack_sendmap, r_next); if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) { rsm = rack_find_high_nonack(rack, rsm); } } if (rsm == NULL) { counter_u64_add(rack_tlp_does_nada, 1); #ifdef TCP_BLACKBOX tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true); #endif goto out; } if ((rsm->r_end - rsm->r_start) > tp->t_maxseg) { /* * We need to split this the last segment in two. */ int32_t idx; struct rack_sendmap *nrsm; nrsm = rack_alloc(rack); if (nrsm == NULL) { /* * No memory to split, we will just exit and punt * off to the RXT timer. */ counter_u64_add(rack_tlp_does_nada, 1); goto out; } nrsm->r_start = (rsm->r_end - tp->t_maxseg); nrsm->r_end = rsm->r_end; nrsm->r_rtr_cnt = rsm->r_rtr_cnt; nrsm->r_flags = rsm->r_flags; nrsm->r_sndcnt = rsm->r_sndcnt; nrsm->r_rtr_bytes = 0; rsm->r_end = nrsm->r_start; for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; } TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~RACK_HAS_FIN); rsm = nrsm; } rack->r_ctl.rc_tlpsend = rsm; rack->r_ctl.rc_tlp_rtx_out = 1; if (rsm->r_start == rack->r_ctl.rc_last_tlp_seq) { rack->r_ctl.rc_tlp_seg_send_cnt++; tp->t_rxtshift++; } else { rack->r_ctl.rc_last_tlp_seq = rsm->r_start; rack->r_ctl.rc_tlp_seg_send_cnt = 1; } send: rack->r_ctl.rc_tlp_send_cnt++; if (rack->r_ctl.rc_tlp_send_cnt > rack_tlp_max_resend) { /* * Can't [re]/transmit a segment we have not heard from the * peer in max times. We need the retransmit timer to take * over. */ restore: rack->r_ctl.rc_tlpsend = NULL; if (rsm) rsm->r_flags &= ~RACK_TLP; rack->r_ctl.rc_prr_sndcnt = old_prr_snd; counter_u64_add(rack_tlp_retran_fail, 1); goto out; } else if (rsm) { rsm->r_flags |= RACK_TLP; } if (rsm && (rsm->r_start == rack->r_ctl.rc_last_tlp_seq) && (rack->r_ctl.rc_tlp_seg_send_cnt > rack_tlp_max_resend)) { /* * We don't want to send a single segment more than the max * either. */ goto restore; } rack->r_timer_override = 1; rack->r_tlp_running = 1; rack->rc_tlp_in_progress = 1; rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; return (0); out: rack->rc_timer_up = 0; rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; return (0); } /* * Delayed ack Timer, here we simply need to setup the * ACK_NOW flag and remove the DELACK flag. From there * the output routine will send the ack out. * * We only return 1, saying don't proceed, if all timers * are stopped (destroyed PCB?). */ static int rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } rack_log_to_event(rack, RACK_TO_FRM_DELACK); tp->t_flags &= ~TF_DELACK; tp->t_flags |= TF_ACKNOW; TCPSTAT_INC(tcps_delack); rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; return (0); } /* * Persists timer, here we simply need to setup the * FORCE-DATA flag the output routine will send * the one byte send. * * We only return 1, saying don't proceed, if all timers * are stopped (destroyed PCB?). */ static int rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { struct inpcb *inp; int32_t retval = 0; inp = tp->t_inpcb; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } if (rack->rc_in_persist == 0) return (0); if (rack_progress_timeout_check(tp)) { tcp_set_inp_to_drop(inp, ETIMEDOUT); return (1); } KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp)); /* * Persistence timer into zero window. Force a byte to be output, if * possible. */ TCPSTAT_INC(tcps_persisttimeo); /* * Hack: if the peer is dead/unreachable, we do not time out if the * window is closed. After a full backoff, drop the connection if * the idle time (no responses to probes) reaches the maximum * backoff that we would use if retransmitting. */ if (tp->t_rxtshift == TCP_MAXRXTSHIFT && (ticks - tp->t_rcvtime >= tcp_maxpersistidle || ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) { TCPSTAT_INC(tcps_persistdrop); retval = 1; tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); goto out; } if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) && tp->snd_una == tp->snd_max) rack_exit_persist(tp, rack); rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT; /* * If the user has closed the socket then drop a persisting * connection after a much reduced timeout. */ if (tp->t_state > TCPS_CLOSE_WAIT && (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) { retval = 1; TCPSTAT_INC(tcps_persistdrop); tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); goto out; } tp->t_flags |= TF_FORCEDATA; out: rack_log_to_event(rack, RACK_TO_FRM_PERSIST); return (retval); } /* * If a keepalive goes off, we had no other timers * happening. We always return 1 here since this * routine either drops the connection or sends * out a segment with respond. */ static int rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { struct tcptemp *t_template; struct inpcb *inp; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP; inp = tp->t_inpcb; rack_log_to_event(rack, RACK_TO_FRM_KEEP); /* * Keep-alive timer went off; send something or drop connection if * idle for too long. */ TCPSTAT_INC(tcps_keeptimeo); if (tp->t_state < TCPS_ESTABLISHED) goto dropit; if ((tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && tp->t_state <= TCPS_CLOSING) { if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp)) goto dropit; /* * Send a packet designed to force a response if the peer is * up and reachable: either an ACK if the connection is * still alive, or an RST if the peer has closed the * connection due to timeout or reboot. Using sequence * number tp->snd_una-1 causes the transmitted zero-length * segment to lie outside the receive window; by the * protocol spec, this requires the correspondent TCP to * respond. */ TCPSTAT_INC(tcps_keepprobe); t_template = tcpip_maketemplate(inp); if (t_template) { tcp_respond(tp, t_template->tt_ipgen, &t_template->tt_t, (struct mbuf *)NULL, tp->rcv_nxt, tp->snd_una - 1, 0); free(t_template, M_TEMP); } } rack_start_hpts_timer(rack, tp, cts, __LINE__, 0, 0, 0); return (1); dropit: TCPSTAT_INC(tcps_keepdrops); tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); return (1); } /* * Retransmit helper function, clear up all the ack * flags and take care of important book keeping. */ static void rack_remxt_tmr(struct tcpcb *tp) { /* * The retransmit timer went off, all sack'd blocks must be * un-acked. */ struct rack_sendmap *rsm, *trsm = NULL; struct tcp_rack *rack; int32_t cnt = 0; rack = (struct tcp_rack *)tp->t_fb_ptr; rack_timer_cancel(tp, rack, tcp_ts_getticks(), __LINE__); rack_log_to_event(rack, RACK_TO_FRM_TMR); if (rack->r_state && (rack->r_state != tp->t_state)) rack_set_state(tp, rack); /* * Ideally we would like to be able to * mark SACK-PASS on anything not acked here. * However, if we do that we would burst out * all that data 1ms apart. This would be unwise, * so for now we will just let the normal rxt timer * and tlp timer take care of it. */ TAILQ_FOREACH(rsm, &rack->r_ctl.rc_map, r_next) { if (rsm->r_flags & RACK_ACKED) { cnt++; rsm->r_sndcnt = 0; if (rsm->r_in_tmap == 0) { /* We must re-add it back to the tlist */ if (trsm == NULL) { TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext); } else { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext); } rsm->r_in_tmap = 1; trsm = rsm; } } rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS); } /* Clear the count (we just un-acked them) */ rack->r_ctl.rc_sacked = 0; /* Clear the tlp rtx mark */ rack->r_ctl.rc_tlp_rtx_out = 0; rack->r_ctl.rc_tlp_seg_send_cnt = 0; rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_map); /* Setup so we send one segment */ if (rack->r_ctl.rc_prr_sndcnt < tp->t_maxseg) rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; rack->r_timer_override = 1; } /* * Re-transmit timeout! If we drop the PCB we will return 1, otherwise * we will setup to retransmit the lowest seq number outstanding. */ static int rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { int32_t rexmt; struct inpcb *inp; int32_t retval = 0; inp = tp->t_inpcb; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } if (rack_progress_timeout_check(tp)) { tcp_set_inp_to_drop(inp, ETIMEDOUT); return (1); } rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT; if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_una == tp->snd_max)) { /* Nothing outstanding .. nothing to do */ return (0); } /* * Retransmission timer went off. Message has not been acked within * retransmit interval. Back off to a longer retransmit interval * and retransmit one segment. */ if (++tp->t_rxtshift > TCP_MAXRXTSHIFT) { tp->t_rxtshift = TCP_MAXRXTSHIFT; TCPSTAT_INC(tcps_timeoutdrop); retval = 1; tcp_set_inp_to_drop(rack->rc_inp, (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT)); goto out; } rack_remxt_tmr(tp); if (tp->t_state == TCPS_SYN_SENT) { /* * If the SYN was retransmitted, indicate CWND to be limited * to 1 segment in cc_conn_init(). */ tp->snd_cwnd = 1; } else if (tp->t_rxtshift == 1) { /* * first retransmit; record ssthresh and cwnd so they can be * recovered if this turns out to be a "bad" retransmit. A * retransmit is considered "bad" if an ACK for this segment * is received within RTT/2 interval; the assumption here is * that the ACK was already in flight. See "On Estimating * End-to-End Network Path Properties" by Allman and Paxson * for more details. */ tp->snd_cwnd_prev = tp->snd_cwnd; tp->snd_ssthresh_prev = tp->snd_ssthresh; tp->snd_recover_prev = tp->snd_recover; if (IN_FASTRECOVERY(tp->t_flags)) tp->t_flags |= TF_WASFRECOVERY; else tp->t_flags &= ~TF_WASFRECOVERY; if (IN_CONGRECOVERY(tp->t_flags)) tp->t_flags |= TF_WASCRECOVERY; else tp->t_flags &= ~TF_WASCRECOVERY; tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1)); tp->t_flags |= TF_PREVVALID; } else tp->t_flags &= ~TF_PREVVALID; TCPSTAT_INC(tcps_rexmttimeo); if ((tp->t_state == TCPS_SYN_SENT) || (tp->t_state == TCPS_SYN_RECEIVED)) rexmt = MSEC_2_TICKS(RACK_INITIAL_RTO * tcp_syn_backoff[tp->t_rxtshift]); else rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift]; TCPT_RANGESET(tp->t_rxtcur, rexmt, max(MSEC_2_TICKS(rack_rto_min), rexmt), MSEC_2_TICKS(rack_rto_max)); /* * We enter the path for PLMTUD if connection is established or, if * connection is FIN_WAIT_1 status, reason for the last is that if * amount of data we send is very small, we could send it in couple * of packets and process straight to FIN. In that case we won't * catch ESTABLISHED state. */ if (V_tcp_pmtud_blackhole_detect && (((tp->t_state == TCPS_ESTABLISHED)) || (tp->t_state == TCPS_FIN_WAIT_1))) { #ifdef INET6 int32_t isipv6; #endif /* * Idea here is that at each stage of mtu probe (usually, * 1448 -> 1188 -> 524) should be given 2 chances to recover * before further clamping down. 'tp->t_rxtshift % 2 == 0' * should take care of that. */ if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) == (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) && (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 && tp->t_rxtshift % 2 == 0)) { /* * Enter Path MTU Black-hole Detection mechanism: - * Disable Path MTU Discovery (IP "DF" bit). - * Reduce MTU to lower value than what we negotiated * with peer. */ if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) { /* Record that we may have found a black hole. */ tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE; /* Keep track of previous MSS. */ tp->t_pmtud_saved_maxseg = tp->t_maxseg; } /* * Reduce the MSS to blackhole value or to the * default in an attempt to retransmit. */ #ifdef INET6 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? 1 : 0; if (isipv6 && tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) { /* Use the sysctl tuneable blackhole MSS. */ tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss; TCPSTAT_INC(tcps_pmtud_blackhole_activated); } else if (isipv6) { /* Use the default MSS. */ tp->t_maxseg = V_tcp_v6mssdflt; /* * Disable Path MTU Discovery when we switch * to minmss. */ tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) { /* Use the sysctl tuneable blackhole MSS. */ tp->t_maxseg = V_tcp_pmtud_blackhole_mss; TCPSTAT_INC(tcps_pmtud_blackhole_activated); } else { /* Use the default MSS. */ tp->t_maxseg = V_tcp_mssdflt; /* * Disable Path MTU Discovery when we switch * to minmss. */ tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); } #endif } else { /* * If further retransmissions are still unsuccessful * with a lowered MTU, maybe this isn't a blackhole * and we restore the previous MSS and blackhole * detection flags. The limit '6' is determined by * giving each probe stage (1448, 1188, 524) 2 * chances to recover. */ if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) && (tp->t_rxtshift >= 6)) { tp->t_flags2 |= TF2_PLPMTU_PMTUD; tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE; tp->t_maxseg = tp->t_pmtud_saved_maxseg; TCPSTAT_INC(tcps_pmtud_blackhole_failed); } } } /* * Disable RFC1323 and SACK if we haven't got any response to our * third SYN to work-around some broken terminal servers (most of * which have hopefully been retired) that have bad VJ header * compression code which trashes TCP segments containing * unknown-to-them TCP options. */ if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) && (tp->t_rxtshift == 3)) tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT); /* * If we backed off this far, our srtt estimate is probably bogus. * Clobber it so we'll take the next rtt measurement as our srtt; * move the current srtt into rttvar to keep the current retransmit * times until then. */ if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) { #ifdef INET6 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) in6_losing(tp->t_inpcb); else #endif in_losing(tp->t_inpcb); tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT); tp->t_srtt = 0; } if (rack_use_sack_filter) sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); tp->snd_recover = tp->snd_max; tp->t_flags |= TF_ACKNOW; tp->t_rtttime = 0; rack_cong_signal(tp, NULL, CC_RTO); out: return (retval); } static int rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling) { int32_t ret = 0; int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK); if (timers == 0) { return (0); } if (tp->t_state == TCPS_LISTEN) { /* no timers on listen sockets */ if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) return (0); return (1); } if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { uint32_t left; if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { ret = -1; rack_log_to_processing(rack, cts, ret, 0); return (0); } if (hpts_calling == 0) { ret = -2; rack_log_to_processing(rack, cts, ret, 0); return (0); } /* * Ok our timer went off early and we are not paced false * alarm, go back to sleep. */ ret = -3; left = rack->r_ctl.rc_timer_exp - cts; tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left)); rack_log_to_processing(rack, cts, ret, left); rack->rc_last_pto_set = 0; return (1); } rack->rc_tmr_stopped = 0; rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK; if (timers & PACE_TMR_DELACK) { ret = rack_timeout_delack(tp, rack, cts); } else if (timers & PACE_TMR_RACK) { ret = rack_timeout_rack(tp, rack, cts); } else if (timers & PACE_TMR_TLP) { ret = rack_timeout_tlp(tp, rack, cts); } else if (timers & PACE_TMR_RXT) { ret = rack_timeout_rxt(tp, rack, cts); } else if (timers & PACE_TMR_PERSIT) { ret = rack_timeout_persist(tp, rack, cts); } else if (timers & PACE_TMR_KEEP) { ret = rack_timeout_keepalive(tp, rack, cts); } rack_log_to_processing(rack, cts, ret, timers); return (ret); } static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line) { uint8_t hpts_removed = 0; if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) { tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); hpts_removed = 1; } if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK; if (rack->rc_inp->inp_in_hpts && ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) { /* * Canceling timer's when we have no output being * paced. We also must remove ourselves from the * hpts. */ tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); hpts_removed = 1; } rack_log_to_cancel(rack, hpts_removed, line); rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK); } } static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type) { return; } static int rack_stopall(struct tcpcb *tp) { struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; rack->t_timers_stopped = 1; return (0); } static void rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta) { return; } static int rack_timer_active(struct tcpcb *tp, uint32_t timer_type) { return (0); } static void rack_stop_all_timers(struct tcpcb *tp) { struct tcp_rack *rack; /* * Assure no timers are running. */ if (tcp_timer_active(tp, TT_PERSIST)) { /* We enter in persists, set the flag appropriately */ rack = (struct tcp_rack *)tp->t_fb_ptr; rack->rc_in_persist = 1; } tcp_timer_suspend(tp, TT_PERSIST); tcp_timer_suspend(tp, TT_REXMT); tcp_timer_suspend(tp, TT_KEEP); tcp_timer_suspend(tp, TT_DELACK); } static void rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t ts) { int32_t idx; rsm->r_rtr_cnt++; rsm->r_sndcnt++; if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) { rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS; rsm->r_flags |= RACK_OVERMAX; } if ((rsm->r_rtr_cnt > 1) && (rack->r_tlp_running == 0)) { rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start); rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start); } idx = rsm->r_rtr_cnt - 1; rsm->r_tim_lastsent[idx] = ts; if (rsm->r_flags & RACK_ACKED) { /* Problably MTU discovery messing with us */ rsm->r_flags &= ~RACK_ACKED; rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); } if (rsm->r_in_tmap) { TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); } TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 1; if (rsm->r_flags & RACK_SACK_PASSED) { /* We have retransmitted due to the SACK pass */ rsm->r_flags &= ~RACK_SACK_PASSED; rsm->r_flags |= RACK_WAS_SACKPASS; } /* Update memory for next rtr */ rack->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); } static uint32_t rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t ts, int32_t * lenp) { /* * We (re-)transmitted starting at rsm->r_start for some length * (possibly less than r_end. */ struct rack_sendmap *nrsm; uint32_t c_end; int32_t len; int32_t idx; len = *lenp; c_end = rsm->r_start + len; if (SEQ_GEQ(c_end, rsm->r_end)) { /* * We retransmitted the whole piece or more than the whole * slopping into the next rsm. */ rack_update_rsm(tp, rack, rsm, ts); if (c_end == rsm->r_end) { *lenp = 0; return (0); } else { int32_t act_len; /* Hangs over the end return whats left */ act_len = rsm->r_end - rsm->r_start; *lenp = (len - act_len); return (rsm->r_end); } /* We don't get out of this block. */ } /* * Here we retransmitted less than the whole thing which means we * have to split this into what was transmitted and what was not. */ nrsm = rack_alloc(rack); if (nrsm == NULL) { /* * We can't get memory, so lets not proceed. */ *lenp = 0; return (0); } /* * So here we are going to take the original rsm and make it what we * retransmitted. nrsm will be the tail portion we did not * retransmit. For example say the chunk was 1, 11 (10 bytes). And * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to * 1, 6 and the new piece will be 6, 11. */ nrsm->r_start = c_end; nrsm->r_end = rsm->r_end; nrsm->r_rtr_cnt = rsm->r_rtr_cnt; nrsm->r_flags = rsm->r_flags; nrsm->r_sndcnt = rsm->r_sndcnt; nrsm->r_rtr_bytes = 0; rsm->r_end = c_end; for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; } TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~RACK_HAS_FIN); rack_update_rsm(tp, rack, rsm, ts); *lenp = 0; return (0); } static void rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len, uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t ts, uint8_t pass, struct rack_sendmap *hintrsm) { struct tcp_rack *rack; struct rack_sendmap *rsm, *nrsm; register uint32_t snd_max, snd_una; int32_t idx; /* * Add to the RACK log of packets in flight or retransmitted. If * there is a TS option we will use the TS echoed, if not we will * grab a TS. * * Retransmissions will increment the count and move the ts to its * proper place. Note that if options do not include TS's then we * won't be able to effectively use the ACK for an RTT on a retran. * * Notes about r_start and r_end. Lets consider a send starting at * sequence 1 for 10 bytes. In such an example the r_start would be * 1 (starting sequence) but the r_end would be r_start+len i.e. 11. * This means that r_end is actually the first sequence for the next * slot (11). * */ /* * If err is set what do we do XXXrrs? should we not add the thing? * -- i.e. return if err != 0 or should we pretend we sent it? -- * i.e. proceed with add ** do this for now. */ INP_WLOCK_ASSERT(tp->t_inpcb); if (err) /* * We don't log errors -- we could but snd_max does not * advance in this case either. */ return; if (th_flags & TH_RST) { /* * We don't log resets and we return immediately from * sending */ return; } rack = (struct tcp_rack *)tp->t_fb_ptr; snd_una = tp->snd_una; if (SEQ_LEQ((seq_out + len), snd_una)) { /* Are sending an old segment to induce an ack (keep-alive)? */ return; } if (SEQ_LT(seq_out, snd_una)) { /* huh? should we panic? */ uint32_t end; end = seq_out + len; seq_out = snd_una; len = end - seq_out; } snd_max = tp->snd_max; if (th_flags & (TH_SYN | TH_FIN)) { /* * The call to rack_log_output is made before bumping * snd_max. This means we can record one extra byte on a SYN * or FIN if seq_out is adding more on and a FIN is present * (and we are not resending). */ if (th_flags & TH_SYN) len++; if (th_flags & TH_FIN) len++; if (SEQ_LT(snd_max, tp->snd_nxt)) { /* * The add/update as not been done for the FIN/SYN * yet. */ snd_max = tp->snd_nxt; } } if (len == 0) { /* We don't log zero window probes */ return; } rack->r_ctl.rc_time_last_sent = ts; if (IN_RECOVERY(tp->t_flags)) { rack->r_ctl.rc_prr_out += len; } /* First question is it a retransmission? */ if (seq_out == snd_max) { again: rsm = rack_alloc(rack); if (rsm == NULL) { /* * Hmm out of memory and the tcb got destroyed while * we tried to wait. */ #ifdef INVARIANTS panic("Out of memory when we should not be rack:%p", rack); #endif return; } if (th_flags & TH_FIN) { rsm->r_flags = RACK_HAS_FIN; } else { rsm->r_flags = 0; } rsm->r_tim_lastsent[0] = ts; rsm->r_rtr_cnt = 1; rsm->r_rtr_bytes = 0; if (th_flags & TH_SYN) { /* The data space is one beyond snd_una */ rsm->r_start = seq_out + 1; rsm->r_end = rsm->r_start + (len - 1); } else { /* Normal case */ rsm->r_start = seq_out; rsm->r_end = rsm->r_start + len; } rsm->r_sndcnt = 0; TAILQ_INSERT_TAIL(&rack->r_ctl.rc_map, rsm, r_next); TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 1; return; } /* * If we reach here its a retransmission and we need to find it. */ more: if (hintrsm && (hintrsm->r_start == seq_out)) { rsm = hintrsm; hintrsm = NULL; } else if (rack->r_ctl.rc_next) { /* We have a hint from a previous run */ rsm = rack->r_ctl.rc_next; } else { /* No hints sorry */ rsm = NULL; } if ((rsm) && (rsm->r_start == seq_out)) { /* * We used rc_next or hintrsm to retransmit, hopefully the * likely case. */ seq_out = rack_update_entry(tp, rack, rsm, ts, &len); if (len == 0) { return; } else { goto more; } } /* Ok it was not the last pointer go through it the hard way. */ TAILQ_FOREACH(rsm, &rack->r_ctl.rc_map, r_next) { if (rsm->r_start == seq_out) { seq_out = rack_update_entry(tp, rack, rsm, ts, &len); rack->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); if (len == 0) { return; } else { continue; } } if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) { /* Transmitted within this piece */ /* * Ok we must split off the front and then let the * update do the rest */ nrsm = rack_alloc(rack); if (nrsm == NULL) { #ifdef INVARIANTS panic("Ran out of memory that was preallocated? rack:%p", rack); #endif rack_update_rsm(tp, rack, rsm, ts); return; } /* * copy rsm to nrsm and then trim the front of rsm * to not include this part. */ nrsm->r_start = seq_out; nrsm->r_end = rsm->r_end; nrsm->r_rtr_cnt = rsm->r_rtr_cnt; nrsm->r_flags = rsm->r_flags; nrsm->r_sndcnt = rsm->r_sndcnt; nrsm->r_rtr_bytes = 0; for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; } rsm->r_end = nrsm->r_start; TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~RACK_HAS_FIN); seq_out = rack_update_entry(tp, rack, nrsm, ts, &len); if (len == 0) { return; } } } /* * Hmm not found in map did they retransmit both old and on into the * new? */ if (seq_out == tp->snd_max) { goto again; } else if (SEQ_LT(seq_out, tp->snd_max)) { #ifdef INVARIANTS printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n", seq_out, len, tp->snd_una, tp->snd_max); printf("Starting Dump of all rack entries\n"); TAILQ_FOREACH(rsm, &rack->r_ctl.rc_map, r_next) { printf("rsm:%p start:%u end:%u\n", rsm, rsm->r_start, rsm->r_end); } printf("Dump complete\n"); panic("seq_out not found rack:%p tp:%p", rack, tp); #endif } else { #ifdef INVARIANTS /* * Hmm beyond sndmax? (only if we are using the new rtt-pack * flag) */ panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p", seq_out, len, tp->snd_max, tp); #endif } } /* * Record one of the RTT updates from an ack into * our sample structure. */ static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt) { if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) || (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) { rack->r_ctl.rack_rs.rs_rtt_lowest = rtt; } if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) || (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) { rack->r_ctl.rack_rs.rs_rtt_highest = rtt; } rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID; rack->r_ctl.rack_rs.rs_rtt_tot += rtt; rack->r_ctl.rack_rs.rs_rtt_cnt++; } /* * Collect new round-trip time estimate * and update averages and current timeout. */ static void tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp) { int32_t delta; uint32_t o_srtt, o_var; int32_t rtt; if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) /* No valid sample */ return; if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) { /* We are to use the lowest RTT seen in a single ack */ rtt = rack->r_ctl.rack_rs.rs_rtt_lowest; } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) { /* We are to use the highest RTT seen in a single ack */ rtt = rack->r_ctl.rack_rs.rs_rtt_highest; } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) { /* We are to use the average RTT seen in a single ack */ rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot / (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt); } else { #ifdef INVARIANTS panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method); #endif return; } if (rtt == 0) rtt = 1; rack_log_rtt_sample(rack, rtt); o_srtt = tp->t_srtt; o_var = tp->t_rttvar; rack = (struct tcp_rack *)tp->t_fb_ptr; if (tp->t_srtt != 0) { /* * srtt is stored as fixed point with 5 bits after the * binary point (i.e., scaled by 8). The following magic is * equivalent to the smoothing algorithm in rfc793 with an * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point). * Adjust rtt to origin 0. */ delta = ((rtt - 1) << TCP_DELTA_SHIFT) - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)); tp->t_srtt += delta; if (tp->t_srtt <= 0) tp->t_srtt = 1; /* * We accumulate a smoothed rtt variance (actually, a * smoothed mean difference), then set the retransmit timer * to smoothed rtt + 4 times the smoothed variance. rttvar * is stored as fixed point with 4 bits after the binary * point (scaled by 16). The following is equivalent to * rfc793 smoothing with an alpha of .75 (rttvar = * rttvar*3/4 + |delta| / 4). This replaces rfc793's * wired-in beta. */ if (delta < 0) delta = -delta; delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT); tp->t_rttvar += delta; if (tp->t_rttvar <= 0) tp->t_rttvar = 1; if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar) tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } else { /* * No rtt measurement yet - use the unsmoothed rtt. Set the * variance to half the rtt (so our first retransmit happens * at 3*rtt). */ tp->t_srtt = rtt << TCP_RTT_SHIFT; tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1); tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } TCPSTAT_INC(tcps_rttupdated); rack_log_rtt_upd(tp, rack, rtt, o_srtt, o_var); tp->t_rttupdated++; #ifdef NETFLIX_STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt)); #endif tp->t_rxtshift = 0; /* * the retransmit should happen at rtt + 4 * rttvar. Because of the * way we do the smoothing, srtt and rttvar will each average +1/2 * tick of bias. When we compute the retransmit timer, we want 1/2 * tick of rounding and 1 extra tick because of +-1/2 tick * uncertainty in the firing of the timer. The bias will give us * exactly the 1.5 tick we need. But, because the bias is * statistical, we have to test that we don't drop below the minimum * feasible timer (which is 2 ticks). */ TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp), max(MSEC_2_TICKS(rack_rto_min), rtt + 2), MSEC_2_TICKS(rack_rto_max)); tp->t_softerror = 0; } static void rack_earlier_retran(struct tcpcb *tp, struct rack_sendmap *rsm, uint32_t t, uint32_t cts) { /* * For this RSM, we acknowledged the data from a previous * transmission, not the last one we made. This means we did a false * retransmit. */ struct tcp_rack *rack; if (rsm->r_flags & RACK_HAS_FIN) { /* * The sending of the FIN often is multiple sent when we * have everything outstanding ack'd. We ignore this case * since its over now. */ return; } if (rsm->r_flags & RACK_TLP) { /* * We expect TLP's to have this occur. */ return; } rack = (struct tcp_rack *)tp->t_fb_ptr; /* should we undo cc changes and exit recovery? */ if (IN_RECOVERY(tp->t_flags)) { if (rack->r_ctl.rc_rsm_start == rsm->r_start) { /* * Undo what we ratched down and exit recovery if * possible */ EXIT_RECOVERY(tp->t_flags); tp->snd_recover = tp->snd_una; if (rack->r_ctl.rc_cwnd_at > tp->snd_cwnd) tp->snd_cwnd = rack->r_ctl.rc_cwnd_at; if (rack->r_ctl.rc_ssthresh_at > tp->snd_ssthresh) tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at; } } if (rsm->r_flags & RACK_WAS_SACKPASS) { /* * We retransmitted based on a sack and the earlier * retransmission ack'd it - re-ordering is occuring. */ counter_u64_add(rack_reorder_seen, 1); rack->r_ctl.rc_reorder_ts = cts; } counter_u64_add(rack_badfr, 1); counter_u64_add(rack_badfr_bytes, (rsm->r_end - rsm->r_start)); } static int rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type) { int32_t i; uint32_t t; if (rsm->r_flags & RACK_ACKED) /* Already done */ return (0); if ((rsm->r_rtr_cnt == 1) || ((ack_type == CUM_ACKED) && (to->to_flags & TOF_TS) && (to->to_tsecr) && (rsm->r_tim_lastsent[rsm->r_rtr_cnt - 1] == to->to_tsecr)) ) { /* * We will only find a matching timestamp if its cum-acked. * But if its only one retransmission its for-sure matching * :-) */ t = cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; if ((int)t <= 0) t = 1; if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { rack->r_ctl.rc_rack_min_rtt = t; if (rack->r_ctl.rc_rack_min_rtt == 0) { rack->r_ctl.rc_rack_min_rtt = 1; } } tcp_rack_xmit_timer(rack, TCP_TS_TO_TICKS(t) + 1); if ((rsm->r_flags & RACK_TLP) && (!IN_RECOVERY(tp->t_flags))) { /* Segment was a TLP and our retrans matched */ if (rack->r_ctl.rc_tlp_cwnd_reduce) { rack->r_ctl.rc_rsm_start = tp->snd_max; rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; rack_cong_signal(tp, NULL, CC_NDUPACK); /* * When we enter recovery we need to assure * we send one packet. */ rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; } else rack->r_ctl.rc_tlp_rtx_out = 0; } if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { /* New more recent rack_tmit_time */ rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; rack->rc_rack_rtt = t; } return (1); } /* * We clear the soft/rxtshift since we got an ack. * There is no assurance we will call the commit() function * so we need to clear these to avoid incorrect handling. */ tp->t_rxtshift = 0; tp->t_softerror = 0; if ((to->to_flags & TOF_TS) && (ack_type == CUM_ACKED) && (to->to_tsecr) && ((rsm->r_flags & (RACK_DEFERRED | RACK_OVERMAX)) == 0)) { /* * Now which timestamp does it match? In this block the ACK * must be coming from a previous transmission. */ for (i = 0; i < rsm->r_rtr_cnt; i++) { if (rsm->r_tim_lastsent[i] == to->to_tsecr) { t = cts - rsm->r_tim_lastsent[i]; if ((int)t <= 0) t = 1; if ((i + 1) < rsm->r_rtr_cnt) { /* Likely */ rack_earlier_retran(tp, rsm, t, cts); } if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { rack->r_ctl.rc_rack_min_rtt = t; if (rack->r_ctl.rc_rack_min_rtt == 0) { rack->r_ctl.rc_rack_min_rtt = 1; } } /* * Note the following calls to * tcp_rack_xmit_timer() are being commented * out for now. They give us no more accuracy * and often lead to a wrong choice. We have * enough samples that have not been * retransmitted. I leave the commented out * code in here in case in the future we * decide to add it back (though I can't forsee * doing that). That way we will easily see * where they need to be placed. */ if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { /* New more recent rack_tmit_time */ rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; rack->rc_rack_rtt = t; } return (1); } } goto ts_not_found; } else { /* * Ok its a SACK block that we retransmitted. or a windows * machine without timestamps. We can tell nothing from the * time-stamp since its not there or the time the peer last * recieved a segment that moved forward its cum-ack point. */ ts_not_found: i = rsm->r_rtr_cnt - 1; t = cts - rsm->r_tim_lastsent[i]; if ((int)t <= 0) t = 1; if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { /* * We retransmitted and the ack came back in less * than the smallest rtt we have observed. We most * likey did an improper retransmit as outlined in * 4.2 Step 3 point 2 in the rack-draft. */ i = rsm->r_rtr_cnt - 2; t = cts - rsm->r_tim_lastsent[i]; rack_earlier_retran(tp, rsm, t, cts); } else if (rack->r_ctl.rc_rack_min_rtt) { /* * We retransmitted it and the retransmit did the * job. */ if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { rack->r_ctl.rc_rack_min_rtt = t; if (rack->r_ctl.rc_rack_min_rtt == 0) { rack->r_ctl.rc_rack_min_rtt = 1; } } if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[i])) { /* New more recent rack_tmit_time */ rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[i]; rack->rc_rack_rtt = t; } return (1); } } return (0); } /* * Mark the SACK_PASSED flag on all entries prior to rsm send wise. */ static void rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm) { struct rack_sendmap *nrsm; uint32_t ts; int32_t idx; idx = rsm->r_rtr_cnt - 1; ts = rsm->r_tim_lastsent[idx]; nrsm = rsm; TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap, rack_head, r_tnext) { if (nrsm == rsm) { /* Skip orginal segment he is acked */ continue; } if (nrsm->r_flags & RACK_ACKED) { /* Skip ack'd segments */ continue; } idx = nrsm->r_rtr_cnt - 1; if (ts == nrsm->r_tim_lastsent[idx]) { /* * For this case lets use seq no, if we sent in a * big block (TSO) we would have a bunch of segments * sent at the same time. * * We would only get a report if its SEQ is earlier. * If we have done multiple retransmits the times * would not be equal. */ if (SEQ_LT(nrsm->r_start, rsm->r_start)) { nrsm->r_flags |= RACK_SACK_PASSED; nrsm->r_flags &= ~RACK_WAS_SACKPASS; } } else { /* * Here they were sent at different times, not a big * block. Since we transmitted this one later and * see it sack'd then this must also be missing (or * we would have gotten a sack block for it) */ nrsm->r_flags |= RACK_SACK_PASSED; nrsm->r_flags &= ~RACK_WAS_SACKPASS; } } } static uint32_t rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts) { int32_t idx; int32_t times = 0; uint32_t start, end, changed = 0; struct rack_sendmap *rsm, *nrsm; int32_t used_ref = 1; start = sack->start; end = sack->end; rsm = *prsm; if (rsm && SEQ_LT(start, rsm->r_start)) { TAILQ_FOREACH_REVERSE_FROM(rsm, &rack->r_ctl.rc_map, rack_head, r_next) { if (SEQ_GEQ(start, rsm->r_start) && SEQ_LT(start, rsm->r_end)) { goto do_rest_ofb; } } } if (rsm == NULL) { start_at_beginning: rsm = NULL; used_ref = 0; } /* First lets locate the block where this guy is */ TAILQ_FOREACH_FROM(rsm, &rack->r_ctl.rc_map, r_next) { if (SEQ_GEQ(start, rsm->r_start) && SEQ_LT(start, rsm->r_end)) { break; } } do_rest_ofb: if (rsm == NULL) { /* * This happens when we get duplicate sack blocks with the * same end. For example SACK 4: 100 SACK 3: 100 The sort * will not change there location so we would just start at * the end of the first one and get lost. */ if (tp->t_flags & TF_SENTFIN) { /* * Check to see if we have not logged the FIN that * went out. */ nrsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_map, rack_sendmap, r_next); if (nrsm && (nrsm->r_end + 1) == tp->snd_max) { /* * Ok we did not get the FIN logged. */ nrsm->r_end++; rsm = nrsm; goto do_rest_ofb; } } if (times == 1) { #ifdef INVARIANTS panic("tp:%p rack:%p sack:%p to:%p prsm:%p", tp, rack, sack, to, prsm); #else goto out; #endif } times++; counter_u64_add(rack_sack_proc_restart, 1); goto start_at_beginning; } /* Ok we have an ACK for some piece of rsm */ if (rsm->r_start != start) { /* * Need to split this in two pieces the before and after. */ nrsm = rack_alloc(rack); if (nrsm == NULL) { /* * failed XXXrrs what can we do but loose the sack * info? */ goto out; } nrsm->r_start = start; nrsm->r_rtr_bytes = 0; nrsm->r_end = rsm->r_end; nrsm->r_rtr_cnt = rsm->r_rtr_cnt; nrsm->r_flags = rsm->r_flags; nrsm->r_sndcnt = rsm->r_sndcnt; for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; } rsm->r_end = nrsm->r_start; TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~RACK_HAS_FIN); rsm = nrsm; } if (SEQ_GEQ(end, rsm->r_end)) { /* * The end of this block is either beyond this guy or right * at this guy. */ if ((rsm->r_flags & RACK_ACKED) == 0) { rack_update_rtt(tp, rack, rsm, to, cts, SACKED); changed += (rsm->r_end - rsm->r_start); rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); rack_log_sack_passed(tp, rack, rsm); /* Is Reordering occuring? */ if (rsm->r_flags & RACK_SACK_PASSED) { counter_u64_add(rack_reorder_seen, 1); rack->r_ctl.rc_reorder_ts = cts; } rsm->r_flags |= RACK_ACKED; rsm->r_flags &= ~RACK_TLP; if (rsm->r_in_tmap) { TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } } if (end == rsm->r_end) { /* This block only - done */ goto out; } /* There is more not coverend by this rsm move on */ start = rsm->r_end; nrsm = TAILQ_NEXT(rsm, r_next); rsm = nrsm; times = 0; goto do_rest_ofb; } /* Ok we need to split off this one at the tail */ nrsm = rack_alloc(rack); if (nrsm == NULL) { /* failed rrs what can we do but loose the sack info? */ goto out; } /* Clone it */ nrsm->r_start = end; nrsm->r_end = rsm->r_end; nrsm->r_rtr_bytes = 0; nrsm->r_rtr_cnt = rsm->r_rtr_cnt; nrsm->r_flags = rsm->r_flags; nrsm->r_sndcnt = rsm->r_sndcnt; for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; } /* The sack block does not cover this guy fully */ rsm->r_flags &= (~RACK_HAS_FIN); rsm->r_end = end; TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } if (rsm->r_flags & RACK_ACKED) { /* Been here done that */ goto out; } rack_update_rtt(tp, rack, rsm, to, cts, SACKED); changed += (rsm->r_end - rsm->r_start); rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); rack_log_sack_passed(tp, rack, rsm); /* Is Reordering occuring? */ if (rsm->r_flags & RACK_SACK_PASSED) { counter_u64_add(rack_reorder_seen, 1); rack->r_ctl.rc_reorder_ts = cts; } rsm->r_flags |= RACK_ACKED; rsm->r_flags &= ~RACK_TLP; if (rsm->r_in_tmap) { TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } out: if (used_ref == 0) { counter_u64_add(rack_sack_proc_all, 1); } else { counter_u64_add(rack_sack_proc_short, 1); } /* Save off where we last were */ if (rsm) rack->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next); else rack->r_ctl.rc_sacklast = NULL; *prsm = rsm; return (changed); } static void inline rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack) { struct rack_sendmap *tmap; tmap = NULL; while (rsm && (rsm->r_flags & RACK_ACKED)) { /* Its no longer sacked, mark it so */ rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); #ifdef INVARIANTS if (rsm->r_in_tmap) { panic("rack:%p rsm:%p flags:0x%x in tmap?", rack, rsm, rsm->r_flags); } #endif rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS); /* Rebuild it into our tmap */ if (tmap == NULL) { TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext); tmap = rsm; } else { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext); tmap = rsm; } tmap->r_in_tmap = 1; rsm = TAILQ_NEXT(rsm, r_next); } /* * Now lets possibly clear the sack filter so we start * recognizing sacks that cover this area. */ if (rack_use_sack_filter) sack_filter_clear(&rack->r_ctl.rack_sf, th_ack); } static void rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th) { uint32_t changed, last_seq, entered_recovery = 0; struct tcp_rack *rack; struct rack_sendmap *rsm; struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1]; register uint32_t th_ack; int32_t i, j, k, num_sack_blks = 0; uint32_t cts, acked, ack_point, sack_changed = 0; INP_WLOCK_ASSERT(tp->t_inpcb); if (th->th_flags & TH_RST) { /* We don't log resets */ return; } rack = (struct tcp_rack *)tp->t_fb_ptr; cts = tcp_ts_getticks(); rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); changed = 0; th_ack = th->th_ack; if (SEQ_GT(th_ack, tp->snd_una)) { rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__); tp->t_acktime = ticks; } if (rsm && SEQ_GT(th_ack, rsm->r_start)) changed = th_ack - rsm->r_start; if (changed) { /* * The ACK point is advancing to th_ack, we must drop off * the packets in the rack log and calculate any eligble * RTT's. */ rack->r_wanted_output++; more: rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); if (rsm == NULL) { if ((th_ack - 1) == tp->iss) { /* * For the SYN incoming case we will not * have called tcp_output for the sending of * the SYN, so there will be no map. All * other cases should probably be a panic. */ goto proc_sack; } if (tp->t_flags & TF_SENTFIN) { /* if we send a FIN we will not hav a map */ goto proc_sack; } #ifdef INVARIANTS panic("No rack map tp:%p for th:%p state:%d rack:%p snd_una:%u snd_max:%u snd_nxt:%u chg:%d\n", tp, th, tp->t_state, rack, tp->snd_una, tp->snd_max, tp->snd_nxt, changed); #endif goto proc_sack; } if (SEQ_LT(th_ack, rsm->r_start)) { /* Huh map is missing this */ #ifdef INVARIANTS printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n", rsm->r_start, th_ack, tp->t_state, rack->r_state); #endif goto proc_sack; } rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED); /* Now do we consume the whole thing? */ if (SEQ_GEQ(th_ack, rsm->r_end)) { /* Its all consumed. */ uint32_t left; rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; rsm->r_rtr_bytes = 0; TAILQ_REMOVE(&rack->r_ctl.rc_map, rsm, r_next); if (rsm->r_in_tmap) { TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } if (rack->r_ctl.rc_next == rsm) { /* scoot along the marker */ rack->r_ctl.rc_next = TAILQ_FIRST(&rack->r_ctl.rc_map); } if (rsm->r_flags & RACK_ACKED) { /* * It was acked on the scoreboard -- remove * it from total */ rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); } else if (rsm->r_flags & RACK_SACK_PASSED) { /* * There are acked segments ACKED on the * scoreboard further up. We are seeing * reordering. */ counter_u64_add(rack_reorder_seen, 1); rsm->r_flags |= RACK_ACKED; rack->r_ctl.rc_reorder_ts = cts; } left = th_ack - rsm->r_end; if (rsm->r_rtr_cnt > 1) { /* * Technically we should make r_rtr_cnt be * monotonicly increasing and just mod it to * the timestamp it is replacing.. that way * we would have the last 3 retransmits. Now * rc_loss_count will be wrong if we * retransmit something more than 2 times in * recovery :( */ rack->r_ctl.rc_loss_count += (rsm->r_rtr_cnt - 1); } /* Free back to zone */ rack_free(rack, rsm); if (left) { goto more; } goto proc_sack; } if (rsm->r_flags & RACK_ACKED) { /* * It was acked on the scoreboard -- remove it from * total for the part being cum-acked. */ rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start); } rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; rsm->r_rtr_bytes = 0; rsm->r_start = th_ack; } proc_sack: /* Check for reneging */ rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) { /* * The peer has moved snd_una up to * the edge of this send, i.e. one * that it had previously acked. The only * way that can be true if the peer threw * away data (space issues) that it had * previously sacked (else it would have * given us snd_una up to (rsm->r_end). * We need to undo the acked markings here. * * Note we have to look to make sure th_ack is * our rsm->r_start in case we get an old ack * where th_ack is behind snd_una. */ rack_peer_reneges(rack, rsm, th->th_ack); } if ((to->to_flags & TOF_SACK) == 0) { /* We are done nothing left to log */ goto out; } rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_map, rack_sendmap, r_next); if (rsm) { last_seq = rsm->r_end; } else { last_seq = tp->snd_max; } /* Sack block processing */ if (SEQ_GT(th_ack, tp->snd_una)) ack_point = th_ack; else ack_point = tp->snd_una; for (i = 0; i < to->to_nsacks; i++) { bcopy((to->to_sacks + i * TCPOLEN_SACK), &sack, sizeof(sack)); sack.start = ntohl(sack.start); sack.end = ntohl(sack.end); if (SEQ_GT(sack.end, sack.start) && SEQ_GT(sack.start, ack_point) && SEQ_LT(sack.start, tp->snd_max) && SEQ_GT(sack.end, ack_point) && SEQ_LEQ(sack.end, tp->snd_max)) { if ((rack->r_ctl.rc_num_maps_alloced > rack_sack_block_limit) && (SEQ_LT(sack.end, last_seq)) && ((sack.end - sack.start) < (tp->t_maxseg / 8))) { /* * Not the last piece and its smaller than * 1/8th of a MSS. We ignore this. */ counter_u64_add(rack_runt_sacks, 1); continue; } sack_blocks[num_sack_blks] = sack; num_sack_blks++; #ifdef NETFLIX_STATS } else if (SEQ_LEQ(sack.start, th_ack) && SEQ_LEQ(sack.end, th_ack)) { /* * Its a D-SACK block. */ tcp_record_dsack(sack.start, sack.end); #endif } } if (num_sack_blks == 0) goto out; /* * Sort the SACK blocks so we can update the rack scoreboard with * just one pass. */ if (rack_use_sack_filter) { num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks, num_sack_blks, th->th_ack); } if (num_sack_blks < 2) { goto do_sack_work; } /* Sort the sacks */ for (i = 0; i < num_sack_blks; i++) { for (j = i + 1; j < num_sack_blks; j++) { if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) { sack = sack_blocks[i]; sack_blocks[i] = sack_blocks[j]; sack_blocks[j] = sack; } } } /* * Now are any of the sack block ends the same (yes some * implememtations send these)? */ again: if (num_sack_blks > 1) { for (i = 0; i < num_sack_blks; i++) { for (j = i + 1; j < num_sack_blks; j++) { if (sack_blocks[i].end == sack_blocks[j].end) { /* * Ok these two have the same end we * want the smallest end and then * throw away the larger and start * again. */ if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) { /* * The second block covers * more area use that */ sack_blocks[i].start = sack_blocks[j].start; } /* * Now collapse out the dup-sack and * lower the count */ for (k = (j + 1); k < num_sack_blks; k++) { sack_blocks[j].start = sack_blocks[k].start; sack_blocks[j].end = sack_blocks[k].end; j++; } num_sack_blks--; goto again; } } } } do_sack_work: rsm = rack->r_ctl.rc_sacklast; for (i = 0; i < num_sack_blks; i++) { acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts); if (acked) { rack->r_wanted_output++; changed += acked; sack_changed += acked; } } out: if (changed) { /* Something changed cancel the rack timer */ rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); } if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) { /* * Ok we have a high probability that we need to go in to * recovery since we have data sack'd */ struct rack_sendmap *rsm; uint32_t tsused; tsused = tcp_ts_getticks(); rsm = tcp_rack_output(tp, rack, tsused); if (rsm) { /* Enter recovery */ rack->r_ctl.rc_rsm_start = rsm->r_start; rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; entered_recovery = 1; rack_cong_signal(tp, NULL, CC_NDUPACK); /* * When we enter recovery we need to assure we send * one packet. */ rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; rack->r_timer_override = 1; } } if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) { /* Deal with changed an PRR here (in recovery only) */ uint32_t pipe, snd_una; rack->r_ctl.rc_prr_delivered += changed; /* Compute prr_sndcnt */ if (SEQ_GT(tp->snd_una, th_ack)) { snd_una = tp->snd_una; } else { snd_una = th_ack; } pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt; if (pipe > tp->snd_ssthresh) { long sndcnt; sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh; if (rack->r_ctl.rc_prr_recovery_fs > 0) sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs; else { rack->r_ctl.rc_prr_sndcnt = 0; sndcnt = 0; } sndcnt++; if (sndcnt > (long)rack->r_ctl.rc_prr_out) sndcnt -= rack->r_ctl.rc_prr_out; else sndcnt = 0; rack->r_ctl.rc_prr_sndcnt = sndcnt; } else { uint32_t limit; if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out) limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out); else limit = 0; if (changed > limit) limit = changed; limit += tp->t_maxseg; if (tp->snd_ssthresh > pipe) { rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit); } else { rack->r_ctl.rc_prr_sndcnt = min(0, limit); } } if (rack->r_ctl.rc_prr_sndcnt >= tp->t_maxseg) { rack->r_timer_override = 1; } } } /* * Return value of 1, we do not need to call rack_process_data(). * return value of 0, rack_process_data can be called. * For ret_val if its 0 the TCP is locked, if its non-zero * its unlocked and probably unsafe to touch the TCB. */ static int rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t * ti_locked, uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val) { int32_t ourfinisacked = 0; int32_t nsegs, acked_amount; int32_t acked; struct mbuf *mfree; struct tcp_rack *rack; int32_t recovery = 0; rack = (struct tcp_rack *)tp->t_fb_ptr; if (SEQ_GT(th->th_ack, tp->snd_max)) { rack_do_dropafterack(m, tp, th, ti_locked, thflags, tlen, ret_val); return (1); } if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) { rack_log_ack(tp, to, th); } if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { /* * Old ack, behind (or duplicate to) the last one rcv'd * Note: Should mark reordering is occuring! We should also * look for sack blocks arriving e.g. ack 1, 4-4 then ack 1, * 3-3, 4-4 would be reording. As well as ack 1, 3-3 ack 3 */ return (0); } /* * If we reach this point, ACK is not a duplicate, i.e., it ACKs * something we sent. */ if (tp->t_flags & TF_NEEDSYN) { /* * T/TCP: Connection was half-synchronized, and our SYN has * been ACK'd (so connection is now fully synchronized). Go * to non-starred state, increment snd_una for ACK of SYN, * and check if we can do window scaling. */ tp->t_flags &= ~TF_NEEDSYN; tp->snd_una++; /* Do window scaling? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; /* Send window already scaled. */ } } nsegs = max(1, m->m_pkthdr.lro_nsegs); INP_WLOCK_ASSERT(tp->t_inpcb); acked = BYTES_THIS_ACK(tp, th); TCPSTAT_ADD(tcps_rcvackpack, nsegs); TCPSTAT_ADD(tcps_rcvackbyte, acked); /* * If we just performed our first retransmit, and the ACK arrives * within our recovery window, then it was a mistake to do the * retransmit in the first place. Recover our original cwnd and * ssthresh, and proceed to transmit where we left off. */ if (tp->t_flags & TF_PREVVALID) { tp->t_flags &= ~TF_PREVVALID; if (tp->t_rxtshift == 1 && (int)(ticks - tp->t_badrxtwin) < 0) rack_cong_signal(tp, th, CC_RTO_ERR); } /* * If we have a timestamp reply, update smoothed round trip time. If * no timestamp is present but transmit timer is running and timed * sequence number was acked, update smoothed round trip time. Since * we now have an rtt measurement, cancel the timer backoff (cf., * Phil Karn's retransmit alg.). Recompute the initial retransmit * timer. * * Some boxes send broken timestamp replies during the SYN+ACK * phase, ignore timestamps of 0 or we could calculate a huge RTT * and blow up the retransmit timer. */ /* * If all outstanding data is acked, stop retransmit timer and * remember to restart (more output or persist). If there is more * data to be acked, restart retransmit timer, using current * (possibly backed-off) value. */ if (th->th_ack == tp->snd_max) { rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); rack->r_wanted_output++; } /* * If no data (only SYN) was ACK'd, skip rest of ACK processing. */ if (acked == 0) { if (ofia) *ofia = ourfinisacked; return (0); } if (rack->r_ctl.rc_early_recovery) { if (IN_FASTRECOVERY(tp->t_flags)) { if (SEQ_LT(th->th_ack, tp->snd_recover)) { tcp_rack_partialack(tp, th); } else { rack_post_recovery(tp, th); recovery = 1; } } } /* * Let the congestion control algorithm update congestion control * related information. This typically means increasing the * congestion window. */ rack_ack_received(tp, rack, th, nsegs, CC_ACK, recovery); SOCKBUF_LOCK(&so->so_snd); acked_amount = min(acked, (int)sbavail(&so->so_snd)); tp->snd_wnd -= acked_amount; mfree = sbcut_locked(&so->so_snd, acked_amount); if ((sbused(&so->so_snd) == 0) && (acked > acked_amount) && (tp->t_state >= TCPS_FIN_WAIT_1)) { ourfinisacked = 1; } /* NB: sowwakeup_locked() does an implicit unlock. */ sowwakeup_locked(so); m_freem(mfree); if (rack->r_ctl.rc_early_recovery == 0) { if (IN_FASTRECOVERY(tp->t_flags)) { if (SEQ_LT(th->th_ack, tp->snd_recover)) { tcp_rack_partialack(tp, th); } else { rack_post_recovery(tp, th); } } } tp->snd_una = th->th_ack; if (SEQ_GT(tp->snd_una, tp->snd_recover)) tp->snd_recover = tp->snd_una; if (SEQ_LT(tp->snd_nxt, tp->snd_una)) { tp->snd_nxt = tp->snd_una; } if (tp->snd_una == tp->snd_max) { /* Nothing left outstanding */ rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__); tp->t_acktime = 0; rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); /* Set need output so persist might get set */ rack->r_wanted_output++; if (rack_use_sack_filter) sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); if ((tp->t_state >= TCPS_FIN_WAIT_1) && (sbavail(&so->so_snd) == 0) && (tp->t_flags2 & TF2_DROP_AF_DATA)) { /* * The socket was gone and the * peer sent data, time to * reset him. */ *ret_val = 1; tp = tcp_close(tp); rack_do_dropwithreset(m, tp, th, ti_locked, BANDLIM_UNLIMITED, tlen); return (1); } } if (ofia) *ofia = ourfinisacked; return (0); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, int32_t * ti_locked, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) { /* * Update window information. Don't look at window if no ACK: TAC's * send garbage on first SYN. */ int32_t nsegs; int32_t tfo_syn; struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; INP_WLOCK_ASSERT(tp->t_inpcb); nsegs = max(1, m->m_pkthdr.lro_nsegs); if ((thflags & TH_ACK) && (SEQ_LT(tp->snd_wl1, th->th_seq) || (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) || (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) { /* keep track of pure window updates */ if (tlen == 0 && tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd) TCPSTAT_INC(tcps_rcvwinupd); tp->snd_wnd = tiwin; tp->snd_wl1 = th->th_seq; tp->snd_wl2 = th->th_ack; if (tp->snd_wnd > tp->max_sndwnd) tp->max_sndwnd = tp->snd_wnd; rack->r_wanted_output++; } else if (thflags & TH_ACK) { if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) { tp->snd_wnd = tiwin; tp->snd_wl1 = th->th_seq; tp->snd_wl2 = th->th_ack; } } /* Was persist timer active and now we have window space? */ if ((rack->rc_in_persist != 0) && tp->snd_wnd) { rack_exit_persist(tp, rack); tp->snd_nxt = tp->snd_max; /* Make sure we output to start the timer */ rack->r_wanted_output++; } /* * Process segments with URG. */ if ((thflags & TH_URG) && th->th_urp && TCPS_HAVERCVDFIN(tp->t_state) == 0) { /* * This is a kludge, but if we receive and accept random * urgent pointers, we'll crash in soreceive. It's hard to * imagine someone actually wanting to send this much urgent * data. */ SOCKBUF_LOCK(&so->so_rcv); if (th->th_urp + sbavail(&so->so_rcv) > sb_max) { th->th_urp = 0; /* XXX */ thflags &= ~TH_URG; /* XXX */ SOCKBUF_UNLOCK(&so->so_rcv); /* XXX */ goto dodata; /* XXX */ } /* * If this segment advances the known urgent pointer, then * mark the data stream. This should not happen in * CLOSE_WAIT, CLOSING, LAST_ACK or TIME_WAIT STATES since a * FIN has been received from the remote side. In these * states we ignore the URG. * * According to RFC961 (Assigned Protocols), the urgent * pointer points to the last octet of urgent data. We * continue, however, to consider it to indicate the first * octet of data past the urgent section as the original * spec states (in one of two places). */ if (SEQ_GT(th->th_seq + th->th_urp, tp->rcv_up)) { tp->rcv_up = th->th_seq + th->th_urp; so->so_oobmark = sbavail(&so->so_rcv) + (tp->rcv_up - tp->rcv_nxt) - 1; if (so->so_oobmark == 0) so->so_rcv.sb_state |= SBS_RCVATMARK; sohasoutofband(so); tp->t_oobflags &= ~(TCPOOB_HAVEDATA | TCPOOB_HADDATA); } SOCKBUF_UNLOCK(&so->so_rcv); /* * Remove out of band data so doesn't get presented to user. * This can happen independent of advancing the URG pointer, * but if two URG's are pending at once, some out-of-band * data may creep in... ick. */ if (th->th_urp <= (uint32_t) tlen && !(so->so_options & SO_OOBINLINE)) { /* hdr drop is delayed */ tcp_pulloutofband(so, th, m, drop_hdrlen); } } else { /* * If no out of band data is expected, pull receive urgent * pointer along with the receive window. */ if (SEQ_GT(tp->rcv_nxt, tp->rcv_up)) tp->rcv_up = tp->rcv_nxt; } dodata: /* XXX */ INP_WLOCK_ASSERT(tp->t_inpcb); /* * Process the segment text, merging it into the TCP sequencing * queue, and arranging for acknowledgment of receipt if necessary. * This process logically involves adjusting tp->rcv_wnd as data is * presented to the user (this happens in tcp_usrreq.c, case * PRU_RCVD). If a FIN has already been received on this connection * then we just ignore the text. */ tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) && IS_FASTOPEN(tp->t_flags)); if ((tlen || (thflags & TH_FIN) || tfo_syn) && TCPS_HAVERCVDFIN(tp->t_state) == 0) { tcp_seq save_start = th->th_seq; m_adj(m, drop_hdrlen); /* delayed header drop */ /* * Insert segment which includes th into TCP reassembly * queue with control block tp. Set thflags to whether * reassembly now includes a segment with FIN. This handles * the common case inline (segment is the next to be * received on an established connection, and the queue is * empty), avoiding linkage into and removal from the queue * and repetition of various conversions. Set DELACK for * segments received in order, but ack immediately when * segments are out of order (so fast retransmit can work). */ if (th->th_seq == tp->rcv_nxt && LIST_EMPTY(&tp->t_segq) && (TCPS_HAVEESTABLISHED(tp->t_state) || tfo_syn)) { if (DELAY_ACK(tp, tlen) || tfo_syn) { rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tp->t_flags |= TF_DELACK; } else { rack->r_wanted_output++; tp->t_flags |= TF_ACKNOW; } tp->rcv_nxt += tlen; thflags = th->th_flags & TH_FIN; TCPSTAT_ADD(tcps_rcvpack, nsegs); TCPSTAT_ADD(tcps_rcvbyte, tlen); SOCKBUF_LOCK(&so->so_rcv); if (so->so_rcv.sb_state & SBS_CANTRCVMORE) m_freem(m); else sbappendstream_locked(&so->so_rcv, m, 0); /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); } else { /* * XXX: Due to the header drop above "th" is * theoretically invalid by now. Fortunately * m_adj() doesn't actually frees any mbufs when * trimming from the head. */ thflags = tcp_reass(tp, th, &tlen, m); tp->t_flags |= TF_ACKNOW; } if (tlen > 0) tcp_update_sack_list(tp, save_start, save_start + tlen); } else { m_freem(m); thflags &= ~TH_FIN; } /* * If FIN is received ACK the FIN and let the user know that the * connection is closing. */ if (thflags & TH_FIN) { if (TCPS_HAVERCVDFIN(tp->t_state) == 0) { socantrcvmore(so); /* * If connection is half-synchronized (ie NEEDSYN * flag on) then delay ACK, so it may be piggybacked * when SYN is sent. Otherwise, since we received a * FIN then no more input can be expected, send ACK * now. */ if (tp->t_flags & TF_NEEDSYN) { rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tp->t_flags |= TF_DELACK; } else { tp->t_flags |= TF_ACKNOW; } tp->rcv_nxt++; } switch (tp->t_state) { /* * In SYN_RECEIVED and ESTABLISHED STATES enter the * CLOSE_WAIT state. */ case TCPS_SYN_RECEIVED: tp->t_starttime = ticks; /* FALLTHROUGH */ case TCPS_ESTABLISHED: rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tcp_state_change(tp, TCPS_CLOSE_WAIT); break; /* * If still in FIN_WAIT_1 STATE FIN has not been * acked so enter the CLOSING state. */ case TCPS_FIN_WAIT_1: rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tcp_state_change(tp, TCPS_CLOSING); break; /* * In FIN_WAIT_2 state enter the TIME_WAIT state, * starting the time-wait timer, turning off the * other standard timers. */ case TCPS_FIN_WAIT_2: rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); INP_INFO_RLOCK_ASSERT(&V_tcbinfo); KASSERT(*ti_locked == TI_RLOCKED, ("%s: dodata " "TCP_FIN_WAIT_2 ti_locked: %d", __func__, *ti_locked)); tcp_twstart(tp); *ti_locked = TI_UNLOCKED; INP_INFO_RUNLOCK(&V_tcbinfo); return (1); } } if (*ti_locked == TI_RLOCKED) { INP_INFO_RUNLOCK(&V_tcbinfo); *ti_locked = TI_UNLOCKED; } /* * Return any desired output. */ if ((tp->t_flags & TF_ACKNOW) || (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) { rack->r_wanted_output++; } KASSERT(*ti_locked == TI_UNLOCKED, ("%s: check_delack ti_locked %d", __func__, *ti_locked)); INP_INFO_UNLOCK_ASSERT(&V_tcbinfo); INP_WLOCK_ASSERT(tp->t_inpcb); return (0); } /* * Here nothing is really faster, its just that we * have broken out the fast-data path also just like * the fast-ack. */ static int rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, int32_t * ti_locked, uint32_t tiwin, int32_t nxt_pkt) { int32_t nsegs; int32_t newsize = 0; /* automatic sockbuf scaling */ struct tcp_rack *rack; #ifdef TCPDEBUG /* * The size of tcp_saveipgen must be the size of the max ip header, * now IPv6. */ u_char tcp_saveipgen[IP6_HDR_LEN]; struct tcphdr tcp_savetcp; short ostate = 0; #endif /* * If last ACK falls within this segment's sequence numbers, record * the timestamp. NOTE that the test is modified according to the * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). */ if (__predict_false(th->th_seq != tp->rcv_nxt)) { return (0); } if (__predict_false(tp->snd_nxt != tp->snd_max)) { return (0); } if (tiwin && tiwin != tp->snd_wnd) { return (0); } if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) { return (0); } if (__predict_false((to->to_flags & TOF_TS) && (TSTMP_LT(to->to_tsval, tp->ts_recent)))) { return (0); } if (__predict_false((th->th_ack != tp->snd_una))) { return (0); } if (__predict_false(tlen > sbspace(&so->so_rcv))) { return (0); } if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } rack = (struct tcp_rack *)tp->t_fb_ptr; /* * This is a pure, in-sequence data packet with nothing on the * reassembly queue and we have enough buffer space to take it. */ if (*ti_locked == TI_RLOCKED) { INP_INFO_RUNLOCK(&V_tcbinfo); *ti_locked = TI_UNLOCKED; } nsegs = max(1, m->m_pkthdr.lro_nsegs); /* Clean receiver SACK report if present */ if (tp->rcv_numsacks) tcp_clean_sackreport(tp); TCPSTAT_INC(tcps_preddat); tp->rcv_nxt += tlen; /* * Pull snd_wl1 up to prevent seq wrap relative to th_seq. */ tp->snd_wl1 = th->th_seq; /* * Pull rcv_up up to prevent seq wrap relative to rcv_nxt. */ tp->rcv_up = tp->rcv_nxt; TCPSTAT_ADD(tcps_rcvpack, nsegs); TCPSTAT_ADD(tcps_rcvbyte, tlen); #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif newsize = tcp_autorcvbuf(m, th, so, tp, tlen); /* Add data to socket buffer. */ SOCKBUF_LOCK(&so->so_rcv); if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { m_freem(m); } else { /* * Set new socket buffer size. Give up when limit is * reached. */ if (newsize) if (!sbreserve_locked(&so->so_rcv, newsize, so, NULL)) so->so_rcv.sb_flags &= ~SB_AUTOSIZE; m_adj(m, drop_hdrlen); /* delayed header drop */ sbappendstream_locked(&so->so_rcv, m, 0); rack_calc_rwin(so, tp); } /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); if (DELAY_ACK(tp, tlen)) { rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tp->t_flags |= TF_DELACK; } else { tp->t_flags |= TF_ACKNOW; rack->r_wanted_output++; } if ((tp->snd_una == tp->snd_max) && rack_use_sack_filter) sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); return (1); } /* * This subfunction is used to try to highly optimize the * fast path. We again allow window updates that are * in sequence to remain in the fast-path. We also add * in the __predict's to attempt to help the compiler. * Note that if we return a 0, then we can *not* process * it and the caller should push the packet into the * slow-path. */ static int rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, int32_t * ti_locked, uint32_t tiwin, int32_t nxt_pkt, uint32_t cts) { int32_t acked; int32_t nsegs; #ifdef TCPDEBUG /* * The size of tcp_saveipgen must be the size of the max ip header, * now IPv6. */ u_char tcp_saveipgen[IP6_HDR_LEN]; struct tcphdr tcp_savetcp; short ostate = 0; #endif struct tcp_rack *rack; if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { /* Old ack, behind (or duplicate to) the last one rcv'd */ return (0); } if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) { /* Above what we have sent? */ return (0); } if (__predict_false(tp->snd_nxt != tp->snd_max)) { /* We are retransmitting */ return (0); } if (__predict_false(tiwin == 0)) { /* zero window */ return (0); } if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) { /* We need a SYN or a FIN, unlikely.. */ return (0); } if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) { /* Timestamp is behind .. old ack with seq wrap? */ return (0); } if (__predict_false(IN_RECOVERY(tp->t_flags))) { /* Still recovering */ return (0); } rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack->r_ctl.rc_sacked) { /* We have sack holes on our scoreboard */ return (0); } /* Ok if we reach here, we can process a fast-ack */ nsegs = max(1, m->m_pkthdr.lro_nsegs); rack_log_ack(tp, to, th); /* Did the window get updated? */ if (tiwin != tp->snd_wnd) { tp->snd_wnd = tiwin; tp->snd_wl1 = th->th_seq; if (tp->snd_wnd > tp->max_sndwnd) tp->max_sndwnd = tp->snd_wnd; } if ((rack->rc_in_persist != 0) && (tp->snd_wnd >= tp->t_maxseg)) { rack_exit_persist(tp, rack); } /* * If last ACK falls within this segment's sequence numbers, record * the timestamp. NOTE that the test is modified according to the * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * This is a pure ack for outstanding data. */ if (*ti_locked == TI_RLOCKED) { INP_INFO_RUNLOCK(&V_tcbinfo); *ti_locked = TI_UNLOCKED; } TCPSTAT_INC(tcps_predack); /* * "bad retransmit" recovery. */ if (tp->t_flags & TF_PREVVALID) { tp->t_flags &= ~TF_PREVVALID; if (tp->t_rxtshift == 1 && (int)(ticks - tp->t_badrxtwin) < 0) rack_cong_signal(tp, th, CC_RTO_ERR); } /* * Recalculate the transmit timer / rtt. * * Some boxes send broken timestamp replies during the SYN+ACK * phase, ignore timestamps of 0 or we could calculate a huge RTT * and blow up the retransmit timer. */ acked = BYTES_THIS_ACK(tp, th); #ifdef TCP_HHOOK /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ hhook_run_tcp_est_in(tp, th, to); #endif TCPSTAT_ADD(tcps_rcvackpack, nsegs); TCPSTAT_ADD(tcps_rcvackbyte, acked); sbdrop(&so->so_snd, acked); /* * Let the congestion control algorithm update congestion control * related information. This typically means increasing the * congestion window. */ rack_ack_received(tp, rack, th, nsegs, CC_ACK, 0); tp->snd_una = th->th_ack; /* * Pull snd_wl2 up to prevent seq wrap relative to th_ack. */ tp->snd_wl2 = th->th_ack; tp->t_dupacks = 0; m_freem(m); /* ND6_HINT(tp); *//* Some progress has been made. */ /* * If all outstanding data are acked, stop retransmit timer, * otherwise restart timer using current (possibly backed-off) * value. If process is waiting for space, wakeup/selwakeup/signal. * If data are ready to send, let tcp_output decide between more * output or persist. */ #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif if (tp->snd_una == tp->snd_max) { rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__); tp->t_acktime = 0; rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); } /* Wake up the socket if we have room to write more */ sowwakeup(so); if (sbavail(&so->so_snd)) { rack->r_wanted_output++; } return (1); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, int32_t * ti_locked, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) { int32_t ret_val = 0; int32_t todrop; int32_t ourfinisacked = 0; rack_calc_rwin(so, tp); /* * If the state is SYN_SENT: if seg contains an ACK, but not for our * SYN, drop the input. if seg contains a RST, then drop the * connection. if seg does not contain SYN, then drop it. Otherwise * this is an acceptable SYN segment initialize tp->rcv_nxt and * tp->irs if seg contains ack then advance tp->snd_una if seg * contains an ECE and ECN support is enabled, the stream is ECN * capable. if SYN has been acked change to ESTABLISHED else * SYN_RCVD state arrange for segment to be acked (eventually) * continue processing rest of data/controls, beginning with URG */ if ((thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { rack_do_dropwithreset(m, tp, th, ti_locked, BANDLIM_RST_OPENPORT, tlen); return (1); } if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) { TCP_PROBE5(connect__refused, NULL, tp, mtod(m, const char *), tp, th); tp = tcp_drop(tp, ECONNREFUSED); rack_do_drop(m, tp, ti_locked); return (1); } if (thflags & TH_RST) { rack_do_drop(m, tp, ti_locked); return (1); } if (!(thflags & TH_SYN)) { rack_do_drop(m, tp, ti_locked); return (1); } tp->irs = th->th_seq; tcp_rcvseqinit(tp); if (thflags & TH_ACK) { int tfo_partial = 0; TCPSTAT_INC(tcps_connects); soisconnected(so); #ifdef MAC mac_socketpeer_set_from_mbuf(m, so); #endif /* Do window scaling on this connection? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; } tp->rcv_adv += min(tp->rcv_wnd, TCP_MAXWIN << tp->rcv_scale); /* * If not all the data that was sent in the TFO SYN * has been acked, resend the remainder right away. */ if (IS_FASTOPEN(tp->t_flags) && (tp->snd_una != tp->snd_max)) { tp->snd_nxt = th->th_ack; tfo_partial = 1; } /* * If there's data, delay ACK; if there's also a FIN ACKNOW * will be turned on later. */ if (DELAY_ACK(tp, tlen) && tlen != 0 && (tfo_partial == 0)) { rack_timer_cancel(tp, (struct tcp_rack *)tp->t_fb_ptr, ((struct tcp_rack *)tp->t_fb_ptr)->r_ctl.rc_rcvtime, __LINE__); tp->t_flags |= TF_DELACK; } else { ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output++; tp->t_flags |= TF_ACKNOW; } if ((thflags & TH_ECE) && V_tcp_do_ecn) { tp->t_flags |= TF_ECN_PERMIT; TCPSTAT_INC(tcps_ecn_shs); } if (SEQ_GT(th->th_ack, tp->snd_una)) { /* * We advance snd_una for the * fast open case. If th_ack is * acknowledging data beyond * snd_una we can't just call * ack-processing since the * data stream in our send-map * will start at snd_una + 1 (one * beyond the SYN). If its just * equal we don't need to do that * and there is no send_map. */ tp->snd_una++; } /* * Received in SYN_SENT[*] state. Transitions: * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1 */ tp->t_starttime = ticks; if (tp->t_flags & TF_NEEDFIN) { tcp_state_change(tp, TCPS_FIN_WAIT_1); tp->t_flags &= ~TF_NEEDFIN; thflags &= ~TH_SYN; } else { tcp_state_change(tp, TCPS_ESTABLISHED); TCP_PROBE5(connect__established, NULL, tp, mtod(m, const char *), tp, th); cc_conn_init(tp); } } else { /* * Received initial SYN in SYN-SENT[*] state => simultaneous * open. If segment contains CC option and there is a * cached CC, apply TAO test. If it succeeds, connection is * * half-synchronized. Otherwise, do 3-way handshake: * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If * there was no CC option, clear cached CC value. */ tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN); tcp_state_change(tp, TCPS_SYN_RECEIVED); } KASSERT(*ti_locked == TI_RLOCKED, ("%s: trimthenstep6: " "ti_locked %d", __func__, *ti_locked)); INP_INFO_RLOCK_ASSERT(&V_tcbinfo); INP_WLOCK_ASSERT(tp->t_inpcb); /* * Advance th->th_seq to correspond to first data byte. If data, * trim to stay within window, dropping FIN if necessary. */ th->th_seq++; if (tlen > tp->rcv_wnd) { todrop = tlen - tp->rcv_wnd; m_adj(m, -todrop); tlen = tp->rcv_wnd; thflags &= ~TH_FIN; TCPSTAT_INC(tcps_rcvpackafterwin); TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); } tp->snd_wl1 = th->th_seq - 1; tp->rcv_up = th->th_seq; /* * Client side of transaction: already sent SYN and data. If the * remote host used T/TCP to validate the SYN, our data will be * ACK'd; if so, enter normal data segment processing in the middle * of step 5, ack processing. Otherwise, goto step 6. */ if (thflags & TH_ACK) { if (rack_process_ack(m, th, so, tp, to, ti_locked, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) return (ret_val); /* We may have changed to FIN_WAIT_1 above */ if (tp->t_state == TCPS_FIN_WAIT_1) { /* * In FIN_WAIT_1 STATE in addition to the processing * for the ESTABLISHED state if our FIN is now * acknowledged then enter FIN_WAIT_2. */ if (ourfinisacked) { /* * If we can't receive any more data, then * closing user can proceed. Starting the * timer is contrary to the specification, * but if we don't get a FIN we'll hang * forever. * * XXXjl: we should release the tp also, and * use a compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, ti_locked, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, int32_t * ti_locked, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) { int32_t ret_val = 0; int32_t ourfinisacked = 0; rack_calc_rwin(so, tp); if ((thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->snd_una) || SEQ_GT(th->th_ack, tp->snd_max))) { rack_do_dropwithreset(m, tp, th, ti_locked, BANDLIM_RST_OPENPORT, tlen); return (1); } if (IS_FASTOPEN(tp->t_flags)) { /* * When a TFO connection is in SYN_RECEIVED, the * only valid packets are the initial SYN, a * retransmit/copy of the initial SYN (possibly with * a subset of the original data), a valid ACK, a * FIN, or a RST. */ if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) { rack_do_dropwithreset(m, tp, th, ti_locked, BANDLIM_RST_OPENPORT, tlen); return (1); } else if (thflags & TH_SYN) { /* non-initial SYN is ignored */ struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) || (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) || (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) { rack_do_drop(m, NULL, ti_locked); return (0); } } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) { rack_do_drop(m, NULL, ti_locked); return (0); } } if (thflags & TH_RST) return (rack_process_rst(m, th, so, tp, ti_locked)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { rack_challenge_ack(m, th, tp, ti_locked, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (rack_ts_check(m, th, tp, ti_locked, tlen, thflags, &ret_val)) return (ret_val); } /* * In the SYN-RECEIVED state, validate that the packet belongs to * this connection before trimming the data to fit the receive * window. Check the sequence number versus IRS since we know the * sequence numbers haven't wrapped. This is a partial fix for the * "LAND" DoS attack. */ if (SEQ_LT(th->th_seq, tp->irs)) { rack_do_dropwithreset(m, tp, th, ti_locked, BANDLIM_RST_OPENPORT, tlen); return (1); } if (rack_drop_checks(to, m, th, tp, &tlen, ti_locked, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (IS_FASTOPEN(tp->t_flags)) { tp->snd_wnd = tiwin; cc_conn_init(tp); } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, ti_locked, tiwin, thflags, nxt_pkt)); } TCPSTAT_INC(tcps_connects); soisconnected(so); /* Do window scaling? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; tp->snd_wnd = tiwin; } /* * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* -> * FIN-WAIT-1 */ tp->t_starttime = ticks; if (tp->t_flags & TF_NEEDFIN) { tcp_state_change(tp, TCPS_FIN_WAIT_1); tp->t_flags &= ~TF_NEEDFIN; } else { tcp_state_change(tp, TCPS_ESTABLISHED); TCP_PROBE5(accept__established, NULL, tp, mtod(m, const char *), tp, th); if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) { tcp_fastopen_decrement_counter(tp->t_tfo_pending); tp->t_tfo_pending = NULL; /* * Account for the ACK of our SYN prior to regular * ACK processing below. */ tp->snd_una++; } /* * TFO connections call cc_conn_init() during SYN * processing. Calling it again here for such connections * is not harmless as it would undo the snd_cwnd reduction * that occurs when a TFO SYN|ACK is retransmitted. */ if (!IS_FASTOPEN(tp->t_flags)) cc_conn_init(tp); } /* * If segment contains data or ACK, will call tcp_reass() later; if * not, do so now to pass queued data to user. */ if (tlen == 0 && (thflags & TH_FIN) == 0) (void)tcp_reass(tp, (struct tcphdr *)0, 0, (struct mbuf *)0); tp->snd_wl1 = th->th_seq - 1; if (rack_process_ack(m, th, so, tp, to, ti_locked, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (tp->t_state == TCPS_FIN_WAIT_1) { /* We could have went to FIN_WAIT_1 (or EST) above */ /* * In FIN_WAIT_1 STATE in addition to the processing for the * ESTABLISHED state if our FIN is now acknowledged then * enter FIN_WAIT_2. */ if (ourfinisacked) { /* * If we can't receive any more data, then closing * user can proceed. Starting the timer is contrary * to the specification, but if we don't get a FIN * we'll hang forever. * * XXXjl: we should release the tp also, and use a * compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, ti_locked, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, int32_t * ti_locked, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) { int32_t ret_val = 0; /* * Header prediction: check for the two common cases of a * uni-directional data xfer. If the packet has no control flags, * is in-sequence, the window didn't change and we're not * retransmitting, it's a candidate. If the length is zero and the * ack moved forward, we're the sender side of the xfer. Just free * the data acked & wake any higher level process that was blocked * waiting for space. If the length is non-zero and the ack didn't * move, we're the receiver side. If we're getting packets in-order * (the reassembly queue is empty), add the data toc The socket * buffer and note that we need a delayed ack. Make sure that the * hidden state-flags are also off. Since we check for * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN. */ if (__predict_true(((to->to_flags & TOF_SACK) == 0)) && __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) && __predict_true(LIST_EMPTY(&tp->t_segq)) && __predict_true(th->th_seq == tp->rcv_nxt)) { struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; if (tlen == 0) { if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen, ti_locked, tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) { return (0); } } else { if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen, ti_locked, tiwin, nxt_pkt)) { return (0); } } } rack_calc_rwin(so, tp); if (thflags & TH_RST) return (rack_process_rst(m, th, so, tp, ti_locked)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { rack_challenge_ack(m, th, tp, ti_locked, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (rack_ts_check(m, th, tp, ti_locked, tlen, thflags, &ret_val)) return (ret_val); } if (rack_drop_checks(to, m, th, tp, &tlen, ti_locked, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, ti_locked, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { rack_do_dropafterack(m, tp, th, ti_locked, thflags, tlen, &ret_val); return (ret_val); } else { rack_do_drop(m, NULL, ti_locked); return (0); } } /* * Ack processing. */ if (rack_process_ack(m, th, so, tp, to, ti_locked, tiwin, tlen, NULL, thflags, &ret_val)) { return (ret_val); } if (sbavail(&so->so_snd)) { if (rack_progress_timeout_check(tp)) { tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); rack_do_dropwithreset(m, tp, th, ti_locked, BANDLIM_RST_OPENPORT, tlen); return (1); } } /* State changes only happen in rack_process_data() */ return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, ti_locked, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, int32_t * ti_locked, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) { int32_t ret_val = 0; rack_calc_rwin(so, tp); if (thflags & TH_RST) return (rack_process_rst(m, th, so, tp, ti_locked)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { rack_challenge_ack(m, th, tp, ti_locked, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (rack_ts_check(m, th, tp, ti_locked, tlen, thflags, &ret_val)) return (ret_val); } if (rack_drop_checks(to, m, th, tp, &tlen, ti_locked, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, ti_locked, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { rack_do_dropafterack(m, tp, th, ti_locked, thflags, tlen, &ret_val); return (ret_val); } else { rack_do_drop(m, NULL, ti_locked); return (0); } } /* * Ack processing. */ if (rack_process_ack(m, th, so, tp, to, ti_locked, tiwin, tlen, NULL, thflags, &ret_val)) { return (ret_val); } if (sbavail(&so->so_snd)) { if (rack_progress_timeout_check(tp)) { tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); rack_do_dropwithreset(m, tp, th, ti_locked, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, ti_locked, tiwin, thflags, nxt_pkt)); } static int rack_check_data_after_close(struct mbuf *m, struct tcpcb *tp, int32_t *ti_locked, int32_t *tlen, struct tcphdr *th, struct socket *so) { struct tcp_rack *rack; KASSERT(*ti_locked == TI_RLOCKED, ("%s: SS_NOFDEREF && " "CLOSE_WAIT && tlen ti_locked %d", __func__, *ti_locked)); INP_INFO_RLOCK_ASSERT(&V_tcbinfo); rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack->rc_allow_data_af_clo == 0) { close_now: tp = tcp_close(tp); TCPSTAT_INC(tcps_rcvafterclose); rack_do_dropwithreset(m, tp, th, ti_locked, BANDLIM_UNLIMITED, (*tlen)); return (1); } if (sbavail(&so->so_snd) == 0) goto close_now; /* Ok we allow data that is ignored and a followup reset */ tp->rcv_nxt = th->th_seq + *tlen; tp->t_flags2 |= TF2_DROP_AF_DATA; rack->r_wanted_output = 1; *tlen = 0; return (0); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, int32_t * ti_locked, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) { int32_t ret_val = 0; int32_t ourfinisacked = 0; rack_calc_rwin(so, tp); if (thflags & TH_RST) return (rack_process_rst(m, th, so, tp, ti_locked)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { rack_challenge_ack(m, th, tp, ti_locked, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (rack_ts_check(m, th, tp, ti_locked, tlen, thflags, &ret_val)) return (ret_val); } if (rack_drop_checks(to, m, th, tp, &tlen, ti_locked, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tlen) { if (rack_check_data_after_close(m, tp, ti_locked, &tlen, th, so)) return (1); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, ti_locked, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { rack_do_dropafterack(m, tp, th, ti_locked, thflags, tlen, &ret_val); return (ret_val); } else { rack_do_drop(m, NULL, ti_locked); return (0); } } /* * Ack processing. */ if (rack_process_ack(m, th, so, tp, to, ti_locked, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (ourfinisacked) { /* * If we can't receive any more data, then closing user can * proceed. Starting the timer is contrary to the * specification, but if we don't get a FIN we'll hang * forever. * * XXXjl: we should release the tp also, and use a * compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } if (sbavail(&so->so_snd)) { if (rack_progress_timeout_check(tp)) { tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); rack_do_dropwithreset(m, tp, th, ti_locked, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, ti_locked, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, int32_t * ti_locked, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) { int32_t ret_val = 0; int32_t ourfinisacked = 0; rack_calc_rwin(so, tp); if (thflags & TH_RST) return (rack_process_rst(m, th, so, tp, ti_locked)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { rack_challenge_ack(m, th, tp, ti_locked, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (rack_ts_check(m, th, tp, ti_locked, tlen, thflags, &ret_val)) return (ret_val); } if (rack_drop_checks(to, m, th, tp, &tlen, ti_locked, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tlen) { if (rack_check_data_after_close(m, tp, ti_locked, &tlen, th, so)) return (1); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, ti_locked, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { rack_do_dropafterack(m, tp, th, ti_locked, thflags, tlen, &ret_val); return (ret_val); } else { rack_do_drop(m, NULL, ti_locked); return (0); } } /* * Ack processing. */ if (rack_process_ack(m, th, so, tp, to, ti_locked, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (ourfinisacked) { INP_INFO_RLOCK_ASSERT(&V_tcbinfo); tcp_twstart(tp); INP_INFO_RUNLOCK(&V_tcbinfo); *ti_locked = TI_UNLOCKED; m_freem(m); return (1); } if (sbavail(&so->so_snd)) { if (rack_progress_timeout_check(tp)) { tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); rack_do_dropwithreset(m, tp, th, ti_locked, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, ti_locked, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, int32_t * ti_locked, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) { int32_t ret_val = 0; int32_t ourfinisacked = 0; rack_calc_rwin(so, tp); if (thflags & TH_RST) return (rack_process_rst(m, th, so, tp, ti_locked)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { rack_challenge_ack(m, th, tp, ti_locked, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (rack_ts_check(m, th, tp, ti_locked, tlen, thflags, &ret_val)) return (ret_val); } if (rack_drop_checks(to, m, th, tp, &tlen, ti_locked, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tlen) { if (rack_check_data_after_close(m, tp, ti_locked, &tlen, th, so)) return (1); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, ti_locked, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { rack_do_dropafterack(m, tp, th, ti_locked, thflags, tlen, &ret_val); return (ret_val); } else { rack_do_drop(m, NULL, ti_locked); return (0); } } /* * case TCPS_LAST_ACK: Ack processing. */ if (rack_process_ack(m, th, so, tp, to, ti_locked, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (ourfinisacked) { INP_INFO_RLOCK_ASSERT(&V_tcbinfo); tp = tcp_close(tp); rack_do_drop(m, tp, ti_locked); return (1); } if (sbavail(&so->so_snd)) { if (rack_progress_timeout_check(tp)) { tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); rack_do_dropwithreset(m, tp, th, ti_locked, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, ti_locked, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, int32_t * ti_locked, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) { int32_t ret_val = 0; int32_t ourfinisacked = 0; rack_calc_rwin(so, tp); /* Reset receive buffer auto scaling when not in bulk receive mode. */ if (thflags & TH_RST) return (rack_process_rst(m, th, so, tp, ti_locked)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { rack_challenge_ack(m, th, tp, ti_locked, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (rack_ts_check(m, th, tp, ti_locked, tlen, thflags, &ret_val)) return (ret_val); } if (rack_drop_checks(to, m, th, tp, &tlen, ti_locked, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tlen) { if (rack_check_data_after_close(m, tp, ti_locked, &tlen, th, so)) return (1); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, ti_locked, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { rack_do_dropafterack(m, tp, th, ti_locked, thflags, tlen, &ret_val); return (ret_val); } else { rack_do_drop(m, NULL, ti_locked); return (0); } } /* * Ack processing. */ if (rack_process_ack(m, th, so, tp, to, ti_locked, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (sbavail(&so->so_snd)) { if (rack_progress_timeout_check(tp)) { tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); rack_do_dropwithreset(m, tp, th, ti_locked, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, ti_locked, tiwin, thflags, nxt_pkt)); } static void inline rack_clear_rate_sample(struct tcp_rack *rack) { rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY; rack->r_ctl.rack_rs.rs_rtt_cnt = 0; rack->r_ctl.rack_rs.rs_rtt_tot = 0; } static int rack_init(struct tcpcb *tp) { struct tcp_rack *rack = NULL; tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT); if (tp->t_fb_ptr == NULL) { /* * We need to allocate memory but cant. The INP and INP_INFO * locks and they are recusive (happens during setup. So a * scheme to drop the locks fails :( * */ return (ENOMEM); } memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack)); rack = (struct tcp_rack *)tp->t_fb_ptr; TAILQ_INIT(&rack->r_ctl.rc_map); TAILQ_INIT(&rack->r_ctl.rc_free); TAILQ_INIT(&rack->r_ctl.rc_tmap); rack->rc_tp = tp; if (tp->t_inpcb) { rack->rc_inp = tp->t_inpcb; } /* Probably not needed but lets be sure */ rack_clear_rate_sample(rack); rack->r_cpu = 0; rack->r_ctl.rc_reorder_fade = rack_reorder_fade; rack->rc_allow_data_af_clo = rack_ignore_data_after_close; rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh; rack->rc_pace_reduce = rack_slot_reduction; if (V_tcp_delack_enabled) tp->t_delayed_ack = 1; else tp->t_delayed_ack = 0; rack->rc_pace_max_segs = rack_hptsi_segments; rack->r_ctl.rc_early_recovery_segs = rack_early_recovery_max_seg; rack->r_ctl.rc_reorder_shift = rack_reorder_thresh; rack->r_ctl.rc_pkt_delay = rack_pkt_delay; rack->r_ctl.rc_prop_reduce = rack_use_proportional_reduce; rack->r_idle_reduce_largest = rack_reduce_largest_on_idle; rack->r_enforce_min_pace = rack_min_pace_time; rack->r_min_pace_seg_thresh = rack_min_pace_time_seg_req; rack->r_ctl.rc_prop_rate = rack_proportional_rate; rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp; rack->r_ctl.rc_early_recovery = rack_early_recovery; rack->rc_always_pace = rack_pace_every_seg; rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method; rack->rack_tlp_threshold_use = rack_tlp_threshold_use; rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr; rack->r_ctl.rc_min_to = rack_min_to; rack->r_ctl.rc_prr_inc_var = rack_inc_var; rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); if (tp->snd_una != tp->snd_max) { /* Create a send map for the current outstanding data */ struct rack_sendmap *rsm; rsm = rack_alloc(rack); if (rsm == NULL) { uma_zfree(rack_pcb_zone, tp->t_fb_ptr); tp->t_fb_ptr = NULL; return (ENOMEM); } rsm->r_flags = RACK_OVERMAX; rsm->r_tim_lastsent[0] = tcp_ts_getticks(); rsm->r_rtr_cnt = 1; rsm->r_rtr_bytes = 0; rsm->r_start = tp->snd_una; rsm->r_end = tp->snd_max; rsm->r_sndcnt = 0; TAILQ_INSERT_TAIL(&rack->r_ctl.rc_map, rsm, r_next); TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 1; } return (0); } static int rack_handoff_ok(struct tcpcb *tp) { if ((tp->t_state == TCPS_CLOSED) || (tp->t_state == TCPS_LISTEN)) { /* Sure no problem though it may not stick */ return (0); } if ((tp->t_state == TCPS_SYN_SENT) || (tp->t_state == TCPS_SYN_RECEIVED)) { /* * We really don't know you have to get to ESTAB or beyond * to tell. */ return (EAGAIN); } if (tp->t_flags & TF_SACK_PERMIT) { return (0); } /* * If we reach here we don't do SACK on this connection so we can * never do rack. */ return (EINVAL); } static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged) { if (tp->t_fb_ptr) { struct tcp_rack *rack; struct rack_sendmap *rsm; rack = (struct tcp_rack *)tp->t_fb_ptr; #ifdef TCP_BLACKBOX tcp_log_flowend(tp); #endif rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); while (rsm) { TAILQ_REMOVE(&rack->r_ctl.rc_map, rsm, r_next); uma_zfree(rack_zone, rsm); rsm = TAILQ_FIRST(&rack->r_ctl.rc_map); } rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); while (rsm) { TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_next); uma_zfree(rack_zone, rsm); rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); } rack->rc_free_cnt = 0; uma_zfree(rack_pcb_zone, tp->t_fb_ptr); tp->t_fb_ptr = NULL; } } static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack) { switch (tp->t_state) { case TCPS_SYN_SENT: rack->r_state = TCPS_SYN_SENT; rack->r_substate = rack_do_syn_sent; break; case TCPS_SYN_RECEIVED: rack->r_state = TCPS_SYN_RECEIVED; rack->r_substate = rack_do_syn_recv; break; case TCPS_ESTABLISHED: rack->r_state = TCPS_ESTABLISHED; rack->r_substate = rack_do_established; break; case TCPS_CLOSE_WAIT: rack->r_state = TCPS_CLOSE_WAIT; rack->r_substate = rack_do_close_wait; break; case TCPS_FIN_WAIT_1: rack->r_state = TCPS_FIN_WAIT_1; rack->r_substate = rack_do_fin_wait_1; break; case TCPS_CLOSING: rack->r_state = TCPS_CLOSING; rack->r_substate = rack_do_closing; break; case TCPS_LAST_ACK: rack->r_state = TCPS_LAST_ACK; rack->r_substate = rack_do_lastack; break; case TCPS_FIN_WAIT_2: rack->r_state = TCPS_FIN_WAIT_2; rack->r_substate = rack_do_fin_wait_2; break; case TCPS_LISTEN: case TCPS_CLOSED: case TCPS_TIME_WAIT: default: #ifdef INVARIANTS panic("tcp tp:%p state:%d sees impossible state?", tp, tp->t_state); #endif break; }; } static void rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb) { /* * We received an ack, and then did not * call send or were bounced out due to the * hpts was running. Now a timer is up as well, is * it the right timer? */ struct rack_sendmap *rsm; int tmr_up; tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK; if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT)) return; rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) && (tmr_up == PACE_TMR_RXT)) { /* Should be an RXT */ return; } if (rsm == NULL) { /* Nothing outstanding? */ if (tp->t_flags & TF_DELACK) { if (tmr_up == PACE_TMR_DELACK) /* We are supposed to have delayed ack up and we do */ return; } else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) { /* * if we hit enobufs then we would expect the possiblity * of nothing outstanding and the RXT up (and the hptsi timer). */ return; } else if (((tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && (tp->t_state <= TCPS_CLOSING)) && (tmr_up == PACE_TMR_KEEP) && (tp->snd_max == tp->snd_una)) { /* We should have keep alive up and we do */ return; } } if (rsm && (rsm->r_flags & RACK_SACK_PASSED)) { if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) == 1) && (rsm->r_flags & RACK_HAS_FIN)) { /* needs to be a RXT */ if (tmr_up == PACE_TMR_RXT) return; } else if (tmr_up == PACE_TMR_RACK) return; } else if (SEQ_GT(tp->snd_max,tp->snd_una) && ((tmr_up == PACE_TMR_TLP) || (tmr_up == PACE_TMR_RXT))) { /* * Either a TLP or RXT is fine if no sack-passed * is in place and data is outstanding. */ return; } else if (tmr_up == PACE_TMR_DELACK) { /* * If the delayed ack was going to go off * before the rtx/tlp/rack timer were going to * expire, then that would be the timer in control. * Note we don't check the time here trusting the * code is correct. */ return; } /* * Ok the timer originally started is not what we want now. * We will force the hpts to be stopped if any, and restart * with the slot set to what was in the saved slot. */ rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); } static void rack_hpts_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t ti_locked, int32_t nxt_pkt, struct timeval *tv) { int32_t thflags, retval, did_out = 0; int32_t way_out = 0; uint32_t cts; uint32_t tiwin; struct tcpopt to; struct tcp_rack *rack; struct rack_sendmap *rsm; int32_t prev_state = 0; cts = tcp_tv_to_mssectick(tv); rack = (struct tcp_rack *)tp->t_fb_ptr; kern_prefetch(rack, &prev_state); prev_state = 0; thflags = th->th_flags; /* * If this is either a state-changing packet or current state isn't * established, we require a read lock on tcbinfo. Otherwise, we * allow the tcbinfo to be in either locked or unlocked, as the * caller may have unnecessarily acquired a lock due to a race. */ if ((thflags & (TH_SYN | TH_FIN | TH_RST)) != 0 || tp->t_state != TCPS_ESTABLISHED) { KASSERT(ti_locked == TI_RLOCKED, ("%s ti_locked %d for " "SYN/FIN/RST/!EST", __func__, ti_locked)); INP_INFO_RLOCK_ASSERT(&V_tcbinfo); } else { #ifdef INVARIANTS if (ti_locked == TI_RLOCKED) { INP_INFO_RLOCK_ASSERT(&V_tcbinfo); } else { KASSERT(ti_locked == TI_UNLOCKED, ("%s: EST " "ti_locked: %d", __func__, ti_locked)); INP_INFO_UNLOCK_ASSERT(&V_tcbinfo); } #endif } INP_WLOCK_ASSERT(tp->t_inpcb); KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", __func__)); KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT", __func__)); { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0, tlen, &log, true); } /* * Segment received on connection. Reset idle time and keep-alive * timer. XXX: This should be done after segment validation to * ignore broken/spoofed segs. */ if (tp->t_idle_reduce && (tp->snd_max == tp->snd_una)) { #ifdef NETFLIX_CWV if ((tp->cwv_enabled) && ((tp->cwv_cwnd_valid == 0) && TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_cwnd > tp->snd_cwv.init_cwnd))) { tcp_newcwv_nvp_closedown(tp); } else #endif if ((ticks - tp->t_rcvtime) >= tp->t_rxtcur) { counter_u64_add(rack_input_idle_reduces, 1); rack_cc_after_idle(tp, (rack->r_idle_reduce_largest ? 1 :0)); } } rack->r_ctl.rc_rcvtime = cts; tp->t_rcvtime = ticks; #ifdef NETFLIX_CWV if (tp->cwv_enabled) { if ((tp->cwv_cwnd_valid == 0) && TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_cwnd > tp->snd_cwv.init_cwnd)) tcp_newcwv_nvp_closedown(tp); } #endif /* * Unscale the window into a 32-bit value. For the SYN_SENT state * the scale is zero. */ tiwin = th->th_win << tp->snd_scale; #ifdef NETFLIX_STATS stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin); #endif /* * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move * this to occur after we've validated the segment. */ if (tp->t_flags & TF_ECN_PERMIT) { if (thflags & TH_CWR) tp->t_flags &= ~TF_ECN_SND_ECE; switch (iptos & IPTOS_ECN_MASK) { case IPTOS_ECN_CE: tp->t_flags |= TF_ECN_SND_ECE; TCPSTAT_INC(tcps_ecn_ce); break; case IPTOS_ECN_ECT0: TCPSTAT_INC(tcps_ecn_ect0); break; case IPTOS_ECN_ECT1: TCPSTAT_INC(tcps_ecn_ect1); break; } /* Congestion experienced. */ if (thflags & TH_ECE) { rack_cong_signal(tp, th, CC_ECN); } } /* * Parse options on any incoming segment. */ tcp_dooptions(&to, (u_char *)(th + 1), (th->th_off << 2) - sizeof(struct tcphdr), (thflags & TH_SYN) ? TO_SYN : 0); /* * If echoed timestamp is later than the current time, fall back to * non RFC1323 RTT calculation. Normalize timestamp if syncookies * were used when this connection was established. */ if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) { to.to_tsecr -= tp->ts_offset; if (TSTMP_GT(to.to_tsecr, cts)) to.to_tsecr = 0; } /* * If its the first time in we need to take care of options and * verify we can do SACK for rack! */ if (rack->r_state == 0) { /* Should be init'd by rack_init() */ KASSERT(rack->rc_inp != NULL, ("%s: rack->rc_inp unexpectedly NULL", __func__)); if (rack->rc_inp == NULL) { rack->rc_inp = tp->t_inpcb; } /* * Process options only when we get SYN/ACK back. The SYN * case for incoming connections is handled in tcp_syncache. * According to RFC1323 the window field in a SYN (i.e., a * or ) segment itself is never scaled. XXX * this is traditional behavior, may need to be cleaned up. */ rack->r_cpu = inp_to_cpuid(tp->t_inpcb); if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { if ((to.to_flags & TOF_SCALE) && (tp->t_flags & TF_REQ_SCALE)) { tp->t_flags |= TF_RCVD_SCALE; tp->snd_scale = to.to_wscale; } /* * Initial send window. It will be updated with the * next incoming segment to the scaled value. */ tp->snd_wnd = th->th_win; if (to.to_flags & TOF_TS) { tp->t_flags |= TF_RCVD_TSTMP; tp->ts_recent = to.to_tsval; tp->ts_recent_age = cts; } if (to.to_flags & TOF_MSS) tcp_mss(tp, to.to_mss); if ((tp->t_flags & TF_SACK_PERMIT) && (to.to_flags & TOF_SACKPERM) == 0) tp->t_flags &= ~TF_SACK_PERMIT; } /* * At this point we are at the initial call. Here we decide * if we are doing RACK or not. We do this by seeing if * TF_SACK_PERMIT is set, if not rack is *not* possible and * we switch to the default code. */ if ((tp->t_flags & TF_SACK_PERMIT) == 0) { tcp_switch_back_to_default(tp); (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen, tlen, iptos, ti_locked); return; } /* Set the flag */ rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; tcp_set_hpts(tp->t_inpcb); rack_stop_all_timers(tp); sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack); } /* * This is the one exception case where we set the rack state * always. All other times (timers etc) we must have a rack-state * set (so we assure we have done the checks above for SACK). */ if (rack->r_state != tp->t_state) rack_set_state(tp, rack); if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&rack->r_ctl.rc_map)) != NULL) kern_prefetch(rsm, &prev_state); prev_state = rack->r_state; rack->r_ctl.rc_tlp_send_cnt = 0; rack_clear_rate_sample(rack); retval = (*rack->r_substate) (m, th, so, tp, &to, drop_hdrlen, tlen, &ti_locked, tiwin, thflags, nxt_pkt); #ifdef INVARIANTS if ((retval == 0) && (tp->t_inpcb == NULL)) { panic("retval:%d tp:%p t_inpcb:NULL state:%d", retval, tp, prev_state); } #endif if (ti_locked != TI_UNLOCKED) { INP_INFO_RLOCK_ASSERT(&V_tcbinfo); INP_INFO_RUNLOCK(&V_tcbinfo); ti_locked = TI_UNLOCKED; } if (retval == 0) { /* * If retval is 1 the tcb is unlocked and most likely the tp * is gone. */ INP_WLOCK_ASSERT(tp->t_inpcb); tcp_rack_xmit_timer_commit(rack, tp); if (((tp->snd_max - tp->snd_una) > tp->snd_wnd) && (rack->rc_in_persist == 0)){ /* * The peer shrunk its window on us to the point * where we have sent too much. The only thing * we can do here is stop any timers and * enter persist. We most likely lost the last * bytes we sent but oh well, we will have to * retransmit them after the peer is caught up. */ if (rack->rc_inp->inp_in_hpts) tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); rack_timer_cancel(tp, rack, cts, __LINE__); rack_enter_persist(tp, rack, cts); rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); way_out = 3; goto done_with_input; } if (nxt_pkt == 0) { if (rack->r_wanted_output != 0) { did_out = 1; (void)tp->t_fb->tfb_tcp_output(tp); } rack_start_hpts_timer(rack, tp, cts, __LINE__, 0, 0, 0); } if (((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) && (SEQ_GT(tp->snd_max, tp->snd_una) || (tp->t_flags & TF_DELACK) || ((tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && (tp->t_state <= TCPS_CLOSING)))) { /* We could not send (probably in the hpts but stopped the timer earlier)? */ if ((tp->snd_max == tp->snd_una) && ((tp->t_flags & TF_DELACK) == 0) && (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { /* keep alive not needed if we are hptsi output yet */ ; } else { if (rack->rc_inp->inp_in_hpts) tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0); } way_out = 1; } else { /* Do we have the correct timer running? */ rack_timer_audit(tp, rack, &so->so_snd); way_out = 2; } done_with_input: rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out); if (did_out) rack->r_wanted_output = 0; #ifdef INVARIANTS if (tp->t_inpcb == NULL) { panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d", did_out, retval, tp, prev_state); } #endif INP_WUNLOCK(tp->t_inpcb); } } void rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t ti_locked) { struct timeval tv; #ifdef RSS struct tcp_function_block *tfb; struct tcp_rack *rack; struct inpcb *inp; rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack->r_state == 0) { /* * Initial input (ACK to SYN-ACK etc)lets go ahead and get * it processed */ - if (ti_locked != TI_RLOCKED && INP_INFO_TRY_RLOCK(&V_tcbinfo)) - ti_locked = TI_RLOCKED; - if (ti_locked != TI_RLOCKED) { - inp = tp->t_inpcb; - tfb = tp->t_fb; - in_pcbref(inp); - INP_WUNLOCK(inp); - INP_INFO_RLOCK(&V_tcbinfo); - ti_locked = TI_RLOCKED; - INP_WLOCK(inp); - if (in_pcbrele_wlocked(inp)) - inp = NULL; - if (inp == NULL || (inp->inp_flags2 & INP_FREED) || - (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED))) { - /* The TCPCB went away. Free the packet. */ - INP_INFO_RUNLOCK(&V_tcbinfo); - if (inp) - INP_WUNLOCK(inp); - m_freem(m); - return; - } - /* If the stack changed, call the correct stack. */ - if (tp->t_fb != tfb) { - tp->t_fb->tfb_tcp_do_segment(m, th, so, tp, - drop_hdrlen, tlen, iptos, ti_locked); - return; - } - } + INP_INFO_RLOCK(); + ti_locked = TI_RLOCKED; tcp_get_usecs(&tv); rack_hpts_do_segment(m, th, so, tp, drop_hdrlen, tlen, iptos, ti_locked, 0, &tv); return; } if (ti_locked == TI_RLOCKED) INP_INFO_RUNLOCK(&V_tcbinfo); tcp_queue_to_input(tp, m, th, tlen, drop_hdrlen, iptos, (uint8_t) ti_locked); INP_WUNLOCK(tp->t_inpcb); #else tcp_get_usecs(&tv); rack_hpts_do_segment(m, th, so, tp, drop_hdrlen, tlen, iptos, ti_locked, 0, &tv); #endif } struct rack_sendmap * tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused) { struct rack_sendmap *rsm = NULL; int32_t idx; uint32_t srtt_cur, srtt = 0, thresh = 0, ts_low = 0; /* Return the next guy to be re-transmitted */ if (TAILQ_EMPTY(&rack->r_ctl.rc_map)) { return (NULL); } if (tp->t_flags & TF_SENTFIN) { /* retran the end FIN? */ return (NULL); } /* ok lets look at this one */ rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) { goto check_it; } rsm = rack_find_lowest_rsm(rack); if (rsm == NULL) { return (NULL); } check_it: srtt_cur = tp->t_srtt >> TCP_RTT_SHIFT; srtt = TICKS_2_MSEC(srtt_cur); if (rack->rc_rack_rtt && (srtt > rack->rc_rack_rtt)) srtt = rack->rc_rack_rtt; if (rsm->r_flags & RACK_ACKED) { return (NULL); } if ((rsm->r_flags & RACK_SACK_PASSED) == 0) { /* Its not yet ready */ return (NULL); } idx = rsm->r_rtr_cnt - 1; ts_low = rsm->r_tim_lastsent[idx]; thresh = rack_calc_thresh_rack(rack, srtt, tsused); if (tsused <= ts_low) { return (NULL); } if ((tsused - ts_low) >= thresh) { return (rsm); } return (NULL); } static int rack_output(struct tcpcb *tp) { struct socket *so; uint32_t recwin, sendwin; uint32_t sb_offset; int32_t len, flags, error = 0; struct mbuf *m; struct mbuf *mb; uint32_t if_hw_tsomaxsegcount = 0; uint32_t if_hw_tsomaxsegsize; long tot_len_this_send = 0; struct ip *ip = NULL; #ifdef TCPDEBUG struct ipovly *ipov = NULL; #endif struct udphdr *udp = NULL; struct tcp_rack *rack; struct tcphdr *th; uint8_t pass = 0; uint8_t wanted_cookie = 0; u_char opt[TCP_MAXOLEN]; unsigned ipoptlen, optlen, hdrlen, ulen=0; uint32_t rack_seq; #if defined(IPSEC) || defined(IPSEC_SUPPORT) unsigned ipsec_optlen = 0; #endif int32_t idle, sendalot; int32_t sub_from_prr = 0; volatile int32_t sack_rxmit; struct rack_sendmap *rsm = NULL; int32_t tso, mtu, would_have_fin = 0; struct tcpopt to; int32_t slot = 0; uint32_t cts; uint8_t hpts_calling, doing_tlp = 0; int32_t do_a_prefetch; int32_t prefetch_rsm = 0; int32_t prefetch_so_done = 0; struct tcp_log_buffer *lgb = NULL; struct inpcb *inp; struct sockbuf *sb; #ifdef INET6 struct ip6_hdr *ip6 = NULL; int32_t isipv6; #endif /* setup and take the cache hits here */ rack = (struct tcp_rack *)tp->t_fb_ptr; inp = rack->rc_inp; so = inp->inp_socket; sb = &so->so_snd; kern_prefetch(sb, &do_a_prefetch); do_a_prefetch = 1; INP_WLOCK_ASSERT(inp); #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) return (tcp_offload_output(tp)); #endif /* * For TFO connections in SYN_RECEIVED, only allow the initial * SYN|ACK and those sent by the retransmit timer. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_state == TCPS_SYN_RECEIVED) && SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN|ACK sent */ (rack->r_ctl.rc_resend == NULL)) /* not a retransmit */ return (0); #ifdef INET6 if (rack->r_state) { /* Use the cache line loaded if possible */ isipv6 = rack->r_is_v6; } else { isipv6 = (inp->inp_vflag & INP_IPV6) != 0; } #endif cts = tcp_ts_getticks(); if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) && inp->inp_in_hpts) { /* * We are on the hpts for some timer but not hptsi output. * Remove from the hpts unconditionally. */ rack_timer_cancel(tp, rack, cts, __LINE__); } /* Mark that we have called rack_output(). */ if ((rack->r_timer_override) || (tp->t_flags & TF_FORCEDATA) || (tp->t_state < TCPS_ESTABLISHED)) { if (tp->t_inpcb->inp_in_hpts) tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT); } else if (tp->t_inpcb->inp_in_hpts) { /* * On the hpts you can't pass even if ACKNOW is on, we will * when the hpts fires. */ counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1); return (0); } hpts_calling = inp->inp_hpts_calls; inp->inp_hpts_calls = 0; if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { if (rack_process_timers(tp, rack, cts, hpts_calling)) { counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1); return (0); } } rack->r_wanted_output = 0; rack->r_timer_override = 0; /* * Determine length of data that should be transmitted, and flags * that will be used. If there is some data or critical controls * (SYN, RST) to send, then transmit; otherwise, investigate * further. */ idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); #ifdef NETFLIX_CWV if (tp->cwv_enabled) { if ((tp->cwv_cwnd_valid == 0) && TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_cwnd > tp->snd_cwv.init_cwnd)) tcp_newcwv_nvp_closedown(tp); } else #endif if (tp->t_idle_reduce) { if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) rack_cc_after_idle(tp, (rack->r_idle_reduce_largest ? 1 :0)); } tp->t_flags &= ~TF_LASTIDLE; if (idle) { if (tp->t_flags & TF_MORETOCOME) { tp->t_flags |= TF_LASTIDLE; idle = 0; } } again: /* * If we've recently taken a timeout, snd_max will be greater than * snd_nxt. There may be SACK information that allows us to avoid * resending already delivered data. Adjust snd_nxt accordingly. */ sendalot = 0; cts = tcp_ts_getticks(); tso = 0; mtu = 0; sb_offset = tp->snd_max - tp->snd_una; sendwin = min(tp->snd_wnd, tp->snd_cwnd); flags = tcp_outflags[tp->t_state]; /* * Send any SACK-generated retransmissions. If we're explicitly * trying to send out new data (when sendalot is 1), bypass this * function. If we retransmit in fast recovery mode, decrement * snd_cwnd, since we're replacing a (future) new transmission with * a retransmission now, and we previously incremented snd_cwnd in * tcp_input(). */ /* * Still in sack recovery , reset rxmit flag to zero. */ while (rack->rc_free_cnt < rack_free_cache) { rsm = rack_alloc(rack); if (rsm == NULL) { if (inp->inp_hpts_calls) /* Retry in a ms */ slot = 1; goto just_return_nolock; } TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_next); rack->rc_free_cnt++; rsm = NULL; } if (inp->inp_hpts_calls) inp->inp_hpts_calls = 0; sack_rxmit = 0; len = 0; rsm = NULL; if (flags & TH_RST) { SOCKBUF_LOCK(sb); goto send; } if (rack->r_ctl.rc_tlpsend) { /* Tail loss probe */ long cwin; long tlen; doing_tlp = 1; rsm = rack->r_ctl.rc_tlpsend; rack->r_ctl.rc_tlpsend = NULL; sack_rxmit = 1; tlen = rsm->r_end - rsm->r_start; if (tlen > tp->t_maxseg) tlen = tp->t_maxseg; #ifdef INVARIANTS if (SEQ_GT(tp->snd_una, rsm->r_start)) { panic("tp:%p rack:%p snd_una:%u rsm:%p r_start:%u", tp, rack, tp->snd_una, rsm, rsm->r_start); } #endif sb_offset = rsm->r_start - tp->snd_una; cwin = min(tp->snd_wnd, tlen); len = cwin; } else if (rack->r_ctl.rc_resend) { /* Retransmit timer */ rsm = rack->r_ctl.rc_resend; rack->r_ctl.rc_resend = NULL; len = rsm->r_end - rsm->r_start; sack_rxmit = 1; sendalot = 0; sb_offset = rsm->r_start - tp->snd_una; if (len >= tp->t_maxseg) { len = tp->t_maxseg; } KASSERT(sb_offset >= 0, ("%s: sack block to the left of una : %d", __func__, sb_offset)); } else if ((rack->rc_in_persist == 0) && ((rsm = tcp_rack_output(tp, rack, cts)) != NULL)) { long tlen; if ((!IN_RECOVERY(tp->t_flags)) && ((tp->t_flags & (TF_WASFRECOVERY | TF_WASCRECOVERY)) == 0)) { /* Enter recovery if not induced by a time-out */ rack->r_ctl.rc_rsm_start = rsm->r_start; rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; rack_cong_signal(tp, NULL, CC_NDUPACK); /* * When we enter recovery we need to assure we send * one packet. */ rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg; } #ifdef INVARIANTS if (SEQ_LT(rsm->r_start, tp->snd_una)) { panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n", tp, rack, rsm, rsm->r_start, tp->snd_una); } #endif tlen = rsm->r_end - rsm->r_start; sb_offset = rsm->r_start - tp->snd_una; if (tlen > rack->r_ctl.rc_prr_sndcnt) { len = rack->r_ctl.rc_prr_sndcnt; } else { len = tlen; } if (len >= tp->t_maxseg) { sendalot = 1; len = tp->t_maxseg; } else { sendalot = 0; if ((rack->rc_timer_up == 0) && (len < tlen)) { /* * If its not a timer don't send a partial * segment. */ len = 0; goto just_return_nolock; } } KASSERT(sb_offset >= 0, ("%s: sack block to the left of una : %d", __func__, sb_offset)); if (len > 0) { sub_from_prr = 1; sack_rxmit = 1; TCPSTAT_INC(tcps_sack_rexmits); TCPSTAT_ADD(tcps_sack_rexmit_bytes, min(len, tp->t_maxseg)); counter_u64_add(rack_rtm_prr_retran, 1); } } if (rsm && (rsm->r_flags & RACK_HAS_FIN)) { /* we are retransmitting the fin */ len--; if (len) { /* * When retransmitting data do *not* include the * FIN. This could happen from a TLP probe. */ flags &= ~TH_FIN; } } #ifdef INVARIANTS /* For debugging */ rack->r_ctl.rc_rsm_at_retran = rsm; #endif /* * Get standard flags, and add SYN or FIN if requested by 'hidden' * state flags. */ if (tp->t_flags & TF_NEEDFIN) flags |= TH_FIN; if (tp->t_flags & TF_NEEDSYN) flags |= TH_SYN; if ((sack_rxmit == 0) && (prefetch_rsm == 0)) { void *end_rsm; end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext); if (end_rsm) kern_prefetch(end_rsm, &prefetch_rsm); prefetch_rsm = 1; } SOCKBUF_LOCK(sb); /* * If in persist timeout with window of 0, send 1 byte. Otherwise, * if window is small but nonzero and time TF_SENTFIN expired, we * will send what we can and go to transmit state. */ if (tp->t_flags & TF_FORCEDATA) { if (sendwin == 0) { /* * If we still have some data to send, then clear * the FIN bit. Usually this would happen below * when it realizes that we aren't sending all the * data. However, if we have exactly 1 byte of * unsent data, then it won't clear the FIN bit * below, and if we are in persist state, we wind up * sending the packet without recording that we sent * the FIN bit. * * We can't just blindly clear the FIN bit, because * if we don't have any more data to send then the * probe will be the FIN itself. */ if (sb_offset < sbused(sb)) flags &= ~TH_FIN; sendwin = 1; } else { if (rack->rc_in_persist) rack_exit_persist(tp, rack); /* * If we are dropping persist mode then we need to * correct snd_nxt/snd_max and off. */ tp->snd_nxt = tp->snd_max; sb_offset = tp->snd_nxt - tp->snd_una; } } /* * If snd_nxt == snd_max and we have transmitted a FIN, the * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a * negative length. This can also occur when TCP opens up its * congestion window while receiving additional duplicate acks after * fast-retransmit because TCP will reset snd_nxt to snd_max after * the fast-retransmit. * * In the normal retransmit-FIN-only case, however, snd_nxt will be * set to snd_una, the sb_offset will be 0, and the length may wind * up 0. * * If sack_rxmit is true we are retransmitting from the scoreboard * in which case len is already set. */ if (sack_rxmit == 0) { uint32_t avail; avail = sbavail(sb); if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail) sb_offset = tp->snd_nxt - tp->snd_una; else sb_offset = 0; if (IN_RECOVERY(tp->t_flags) == 0) { if (rack->r_ctl.rc_tlp_new_data) { /* TLP is forcing out new data */ if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) { rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset); } if (rack->r_ctl.rc_tlp_new_data > tp->snd_wnd) len = tp->snd_wnd; else len = rack->r_ctl.rc_tlp_new_data; rack->r_ctl.rc_tlp_new_data = 0; doing_tlp = 1; } else { if (sendwin > avail) { /* use the available */ if (avail > sb_offset) { len = (int32_t)(avail - sb_offset); } else { len = 0; } } else { if (sendwin > sb_offset) { len = (int32_t)(sendwin - sb_offset); } else { len = 0; } } } } else { uint32_t outstanding; /* * We are inside of a SACK recovery episode and are * sending new data, having retransmitted all the * data possible so far in the scoreboard. */ outstanding = tp->snd_max - tp->snd_una; if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) len = 0; else if (avail > sb_offset) len = avail - sb_offset; else len = 0; if (len > 0) { if (len > rack->r_ctl.rc_prr_sndcnt) len = rack->r_ctl.rc_prr_sndcnt; if (len > 0) { sub_from_prr = 1; counter_u64_add(rack_rtm_prr_newdata, 1); } } if (len > tp->t_maxseg) { /* * We should never send more than a MSS when * retransmitting or sending new data in prr * mode unless the override flag is on. Most * likely the PRR algorithm is not going to * let us send a lot as well :-) */ if (rack->r_ctl.rc_prr_sendalot == 0) len = tp->t_maxseg; } else if (len < tp->t_maxseg) { /* * Do we send any? The idea here is if the * send empty's the socket buffer we want to * do it. However if not then lets just wait * for our prr_sndcnt to get bigger. */ long leftinsb; leftinsb = sbavail(sb) - sb_offset; if (leftinsb > len) { /* This send does not empty the sb */ len = 0; } } } } if (prefetch_so_done == 0) { kern_prefetch(so, &prefetch_so_done); prefetch_so_done = 1; } /* * Lop off SYN bit if it has already been sent. However, if this is * SYN-SENT state and if segment contains data and if we don't know * that foreign host supports TAO, suppress sending segment. */ if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) && ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) { if (tp->t_state != TCPS_SYN_RECEIVED) flags &= ~TH_SYN; /* * When sending additional segments following a TFO SYN|ACK, * do not include the SYN bit. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_state == TCPS_SYN_RECEIVED)) flags &= ~TH_SYN; sb_offset--, len++; } /* * Be careful not to send data and/or FIN on SYN segments. This * measure is needed to prevent interoperability problems with not * fully conformant TCP implementations. */ if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { len = 0; flags &= ~TH_FIN; } /* * On TFO sockets, ensure no data is sent in the following cases: * * - When retransmitting SYN|ACK on a passively-created socket * * - When retransmitting SYN on an actively created socket * * - When sending a zero-length cookie (cookie request) on an * actively created socket * * - When the socket is in the CLOSED state (RST is being sent) */ if (IS_FASTOPEN(tp->t_flags) && (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || ((tp->t_state == TCPS_SYN_SENT) && (tp->t_tfo_client_cookie_len == 0)) || (flags & TH_RST))) len = 0; /* Without fast-open there should never be data sent on a SYN */ if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) len = 0; if (len <= 0) { /* * If FIN has been sent but not acked, but we haven't been * called to retransmit, len will be < 0. Otherwise, window * shrank after we sent into it. If window shrank to 0, * cancel pending retransmit, pull snd_nxt back to (closed) * window, and set the persist timer if it isn't already * going. If the window didn't close completely, just wait * for an ACK. * * We also do a general check here to ensure that we will * set the persist timer when we have data to send, but a * 0-byte window. This makes sure the persist timer is set * even if the packet hits one of the "goto send" lines * below. */ len = 0; if ((tp->snd_wnd == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) && (sb_offset < (int)sbavail(sb))) { tp->snd_nxt = tp->snd_una; rack_enter_persist(tp, rack, cts); } } /* len will be >= 0 after this point. */ KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); tcp_sndbuf_autoscale(tp, so, sendwin); /* * Decide if we can use TCP Segmentation Offloading (if supported by * hardware). * * TSO may only be used if we are in a pure bulk sending state. The * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP * options prevent using TSO. With TSO the TCP header is the same * (except for the sequence number) for all generated packets. This * makes it impossible to transmit any options which vary per * generated segment or packet. * * IPv4 handling has a clear separation of ip options and ip header * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does * the right thing below to provide length of just ip options and thus * checking for ipoptlen is enough to decide if ip options are present. */ #ifdef INET6 if (isipv6) ipoptlen = ip6_optlen(tp->t_inpcb); else #endif if (tp->t_inpcb->inp_options) ipoptlen = tp->t_inpcb->inp_options->m_len - offsetof(struct ipoption, ipopt_list); else ipoptlen = 0; #if defined(IPSEC) || defined(IPSEC_SUPPORT) /* * Pre-calculate here as we save another lookup into the darknesses * of IPsec that way and can actually decide if TSO is ok. */ #ifdef INET6 if (isipv6 && IPSEC_ENABLED(ipv6)) ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb); #ifdef INET else #endif #endif /* INET6 */ #ifdef INET if (IPSEC_ENABLED(ipv4)) ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb); #endif /* INET */ #endif #if defined(IPSEC) || defined(IPSEC_SUPPORT) ipoptlen += ipsec_optlen; #endif if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg && (tp->t_port == 0) && ((tp->t_flags & TF_SIGNATURE) == 0) && tp->rcv_numsacks == 0 && sack_rxmit == 0 && ipoptlen == 0) tso = 1; { uint32_t outstanding; outstanding = tp->snd_max - tp->snd_una; if (tp->t_flags & TF_SENTFIN) { /* * If we sent a fin, snd_max is 1 higher than * snd_una */ outstanding--; } if (outstanding > 0) { /* * This is sub-optimal. We only send a stand alone * FIN on its own segment. */ if (flags & TH_FIN) { flags &= ~TH_FIN; would_have_fin = 1; } } else if (sack_rxmit) { if ((rsm->r_flags & RACK_HAS_FIN) == 0) flags &= ~TH_FIN; } else { if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + sbused(sb))) flags &= ~TH_FIN; } } recwin = sbspace(&so->so_rcv); /* * Sender silly window avoidance. We transmit under the following * conditions when len is non-zero: * * - We have a full segment (or more with TSO) - This is the last * buffer in a write()/send() and we are either idle or running * NODELAY - we've timed out (e.g. persist timer) - we have more * then 1/2 the maximum send window's worth of data (receiver may be * limited the window size) - we need to retransmit */ if (len) { if (len >= tp->t_maxseg) { pass = 1; goto send; } /* * NOTE! on localhost connections an 'ack' from the remote * end may occur synchronously with the output and cause us * to flush a buffer queued with moretocome. XXX * */ if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ (idle || (tp->t_flags & TF_NODELAY)) && ((uint32_t)len + (uint32_t)sb_offset >= sbavail(&so->so_snd)) && (tp->t_flags & TF_NOPUSH) == 0) { pass = 2; goto send; } if (tp->t_flags & TF_FORCEDATA) { /* typ. timeout case */ pass = 3; goto send; } if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */ goto send; } if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) { pass = 4; goto send; } if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */ pass = 5; goto send; } if (sack_rxmit) { pass = 6; goto send; } } /* * Sending of standalone window updates. * * Window updates are important when we close our window due to a * full socket buffer and are opening it again after the application * reads data from it. Once the window has opened again and the * remote end starts to send again the ACK clock takes over and * provides the most current window information. * * We must avoid the silly window syndrome whereas every read from * the receive buffer, no matter how small, causes a window update * to be sent. We also should avoid sending a flurry of window * updates when the socket buffer had queued a lot of data and the * application is doing small reads. * * Prevent a flurry of pointless window updates by only sending an * update when we can increase the advertized window by more than * 1/4th of the socket buffer capacity. When the buffer is getting * full or is very small be more aggressive and send an update * whenever we can increase by two mss sized segments. In all other * situations the ACK's to new incoming data will carry further * window increases. * * Don't send an independent window update if a delayed ACK is * pending (it will get piggy-backed on it) or the remote side * already has done a half-close and won't send more data. Skip * this if the connection is in T/TCP half-open state. */ if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && !(tp->t_flags & TF_DELACK) && !TCPS_HAVERCVDFIN(tp->t_state)) { /* * "adv" is the amount we could increase the window, taking * into account that we are limited by TCP_MAXWIN << * tp->rcv_scale. */ int32_t adv; int oldwin; adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale); if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { oldwin = (tp->rcv_adv - tp->rcv_nxt); adv -= oldwin; } else oldwin = 0; /* * If the new window size ends up being the same as the old * size when it is scaled, then don't force a window update. */ if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale) goto dontupdate; if (adv >= (int32_t)(2 * tp->t_maxseg) && (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) || recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) || so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg)) { pass = 7; goto send; } if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) goto send; } dontupdate: /* * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW * is also a catch-all for the retransmit timer timeout case. */ if (tp->t_flags & TF_ACKNOW) { pass = 8; goto send; } if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) { pass = 9; goto send; } if (SEQ_GT(tp->snd_up, tp->snd_una)) { pass = 10; goto send; } /* * If our state indicates that FIN should be sent and we have not * yet done so, then we need to send. */ if (flags & TH_FIN) { if ((tp->t_flags & TF_SENTFIN) || (((tp->t_flags & TF_SENTFIN) == 0) && (tp->snd_nxt == tp->snd_una))) { pass = 11; goto send; } } /* * No reason to send a segment, just return. */ just_return: SOCKBUF_UNLOCK(sb); just_return_nolock: if (tot_len_this_send == 0) counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1); rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, tot_len_this_send, 1); rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling); tp->t_flags &= ~TF_FORCEDATA; return (0); send: if (doing_tlp == 0) { /* * Data not a TLP, and its not the rxt firing. If it is the * rxt firing, we want to leave the tlp_in_progress flag on * so we don't send another TLP. It has to be a rack timer * or normal send (response to acked data) to clear the tlp * in progress flag. */ rack->rc_tlp_in_progress = 0; } SOCKBUF_LOCK_ASSERT(sb); if (len > 0) { if (len >= tp->t_maxseg) tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; else tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; } /* * Before ESTABLISHED, force sending of initial options unless TCP * set not to do any options. NOTE: we assume that the IP/TCP header * plus TCP options always fit in a single mbuf, leaving room for a * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr) * + optlen <= MCLBYTES */ optlen = 0; #ifdef INET6 if (isipv6) hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); else #endif hdrlen = sizeof(struct tcpiphdr); /* * Compute options for segment. We only have to care about SYN and * established connection segments. Options for SYN-ACK segments * are handled in TCP syncache. */ to.to_flags = 0; if ((tp->t_flags & TF_NOOPT) == 0) { /* Maximum segment size. */ if (flags & TH_SYN) { tp->snd_nxt = tp->iss; to.to_mss = tcp_mssopt(&inp->inp_inc); #ifdef NETFLIX_TCPOUDP if (tp->t_port) to.to_mss -= V_tcp_udp_tunneling_overhead; #endif to.to_flags |= TOF_MSS; /* * On SYN or SYN|ACK transmits on TFO connections, * only include the TFO option if it is not a * retransmit, as the presence of the TFO option may * have caused the original SYN or SYN|ACK to have * been dropped by a middlebox. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_rxtshift == 0)) { if (tp->t_state == TCPS_SYN_RECEIVED) { to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; to.to_tfo_cookie = (u_int8_t *)&tp->t_tfo_cookie.server; to.to_flags |= TOF_FASTOPEN; wanted_cookie = 1; } else if (tp->t_state == TCPS_SYN_SENT) { to.to_tfo_len = tp->t_tfo_client_cookie_len; to.to_tfo_cookie = tp->t_tfo_cookie.client; to.to_flags |= TOF_FASTOPEN; wanted_cookie = 1; /* * If we wind up having more data to * send with the SYN than can fit in * one segment, don't send any more * until the SYN|ACK comes back from * the other end. */ sendalot = 0; } } } /* Window scaling. */ if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { to.to_wscale = tp->request_r_scale; to.to_flags |= TOF_SCALE; } /* Timestamps. */ if ((tp->t_flags & TF_RCVD_TSTMP) || ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { to.to_tsval = cts + tp->ts_offset; to.to_tsecr = tp->ts_recent; to.to_flags |= TOF_TS; } /* Set receive buffer autosizing timestamp. */ if (tp->rfbuf_ts == 0 && (so->so_rcv.sb_flags & SB_AUTOSIZE)) tp->rfbuf_ts = tcp_ts_getticks(); /* Selective ACK's. */ if (flags & TH_SYN) to.to_flags |= TOF_SACKPERM; else if (TCPS_HAVEESTABLISHED(tp->t_state) && tp->rcv_numsacks > 0) { to.to_flags |= TOF_SACK; to.to_nsacks = tp->rcv_numsacks; to.to_sacks = (u_char *)tp->sackblks; } #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) /* TCP-MD5 (RFC2385). */ if (tp->t_flags & TF_SIGNATURE) to.to_flags |= TOF_SIGNATURE; #endif /* TCP_SIGNATURE */ /* Processing the options. */ hdrlen += optlen = tcp_addoptions(&to, opt); /* * If we wanted a TFO option to be added, but it was unable * to fit, ensure no data is sent. */ if (IS_FASTOPEN(tp->t_flags) && wanted_cookie && !(to.to_flags & TOF_FASTOPEN)) len = 0; } #ifdef NETFLIX_TCPOUDP if (tp->t_port) { if (V_tcp_udp_tunneling_port == 0) { /* The port was removed?? */ SOCKBUF_UNLOCK(&so->so_snd); return (EHOSTUNREACH); } hdrlen += sizeof(struct udphdr); } #endif ipoptlen = 0; #if defined(IPSEC) || defined(IPSEC_SUPPORT) ipoptlen += ipsec_optlen; #endif /* * Adjust data length if insertion of options will bump the packet * length beyond the t_maxseg length. Clear the FIN bit because we * cut off the tail of the segment. */ if (len + optlen + ipoptlen > tp->t_maxseg) { if (flags & TH_FIN) { would_have_fin = 1; flags &= ~TH_FIN; } if (tso) { uint32_t if_hw_tsomax; uint32_t moff; int32_t max_len; /* extract TSO information */ if_hw_tsomax = tp->t_tsomax; if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; KASSERT(ipoptlen == 0, ("%s: TSO can't do IP options", __func__)); /* * Check if we should limit by maximum payload * length: */ if (if_hw_tsomax != 0) { /* compute maximum TSO length */ max_len = (if_hw_tsomax - hdrlen - max_linkhdr); if (max_len <= 0) { len = 0; } else if (len > max_len) { sendalot = 1; len = max_len; } } /* * Prevent the last segment from being fractional * unless the send sockbuf can be emptied: */ max_len = (tp->t_maxseg - optlen); if ((sb_offset + len) < sbavail(sb)) { moff = len % (u_int)max_len; if (moff != 0) { len -= moff; sendalot = 1; } } /* * In case there are too many small fragments don't * use TSO: */ if (len <= max_len) { len = max_len; sendalot = 1; tso = 0; } /* * Send the FIN in a separate segment after the bulk * sending is done. We don't trust the TSO * implementations to clear the FIN flag on all but * the last segment. */ if (tp->t_flags & TF_NEEDFIN) sendalot = 1; } else { len = tp->t_maxseg - optlen - ipoptlen; sendalot = 1; } } else tso = 0; KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, ("%s: len > IP_MAXPACKET", __func__)); #ifdef DIAGNOSTIC #ifdef INET6 if (max_linkhdr + hdrlen > MCLBYTES) #else if (max_linkhdr + hdrlen > MHLEN) #endif panic("tcphdr too big"); #endif /* * This KASSERT is here to catch edge cases at a well defined place. * Before, those had triggered (random) panic conditions further * down. */ KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); if ((len == 0) && (flags & TH_FIN) && (sbused(sb))) { /* * We have outstanding data, don't send a fin by itself!. */ goto just_return; } /* * Grab a header mbuf, attaching a copy of data to be transmitted, * and initialize the header from the template for sends on this * connection. */ if (len) { uint32_t max_val; uint32_t moff; if (rack->rc_pace_max_segs) max_val = rack->rc_pace_max_segs * tp->t_maxseg; else max_val = len; /* * We allow a limit on sending with hptsi. */ if (len > max_val) { len = max_val; } #ifdef INET6 if (MHLEN < hdrlen + max_linkhdr) m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); else #endif m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) { SOCKBUF_UNLOCK(sb); error = ENOBUFS; sack_rxmit = 0; goto out; } m->m_data += max_linkhdr; m->m_len = hdrlen; /* * Start the m_copy functions from the closest mbuf to the * sb_offset in the socket buffer chain. */ mb = sbsndptr_noadv(sb, sb_offset, &moff); if (len <= MHLEN - hdrlen - max_linkhdr) { m_copydata(mb, moff, (int)len, mtod(m, caddr_t)+hdrlen); if (SEQ_LT(tp->snd_nxt, tp->snd_max)) sbsndptr_adv(sb, mb, len); m->m_len += len; } else { struct sockbuf *msb; if (SEQ_LT(tp->snd_nxt, tp->snd_max)) msb = NULL; else msb = sb; m->m_next = tcp_m_copym(mb, moff, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb); if (len <= (tp->t_maxseg - optlen)) { /* * Must have ran out of mbufs for the copy * shorten it to no longer need tso. Lets * not put on sendalot since we are low on * mbufs. */ tso = 0; } if (m->m_next == NULL) { SOCKBUF_UNLOCK(sb); (void)m_free(m); error = ENOBUFS; sack_rxmit = 0; goto out; } } if ((tp->t_flags & TF_FORCEDATA) && len == 1) { TCPSTAT_INC(tcps_sndprobe); #ifdef NETFLIX_STATS if (SEQ_LT(tp->snd_nxt, tp->snd_max)) stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, len); else stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, len); #endif } else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { if (rsm && (rsm->r_flags & RACK_TLP)) { /* * TLP should not count in retran count, but * in its own bin */ counter_u64_add(rack_tlp_retran, 1); counter_u64_add(rack_tlp_retran_bytes, len); } else { tp->t_sndrexmitpack++; TCPSTAT_INC(tcps_sndrexmitpack); TCPSTAT_ADD(tcps_sndrexmitbyte, len); } #ifdef NETFLIX_STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, len); #endif } else { TCPSTAT_INC(tcps_sndpack); TCPSTAT_ADD(tcps_sndbyte, len); #ifdef NETFLIX_STATS stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, len); #endif } /* * If we're sending everything we've got, set PUSH. (This * will keep happy those implementations which only give * data to the user when a buffer fills or a PUSH comes in.) */ if (sb_offset + len == sbused(sb) && sbused(sb) && !(flags & TH_SYN)) flags |= TH_PUSH; /* * Are we doing hptsi, if so we must calculate the slot. We * only do hptsi in ESTABLISHED and with no RESET being * sent where we have data to send. */ if (((tp->t_state == TCPS_ESTABLISHED) || (tp->t_state == TCPS_CLOSE_WAIT) || ((tp->t_state == TCPS_FIN_WAIT_1) && ((tp->t_flags & TF_SENTFIN) == 0) && ((flags & TH_FIN) == 0))) && ((flags & TH_RST) == 0) && (rack->rc_always_pace)) { /* * We use the most optimistic possible cwnd/srtt for * sending calculations. This will make our * calculation anticipate getting more through * quicker then possible. But thats ok we don't want * the peer to have a gap in data sending. */ uint32_t srtt, cwnd, tr_perms = 0; if (rack->r_ctl.rc_rack_min_rtt) srtt = rack->r_ctl.rc_rack_min_rtt; else srtt = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT)); if (rack->r_ctl.rc_rack_largest_cwnd) cwnd = rack->r_ctl.rc_rack_largest_cwnd; else cwnd = tp->snd_cwnd; tr_perms = cwnd / srtt; if (tr_perms == 0) { tr_perms = tp->t_maxseg; } tot_len_this_send += len; /* * Calculate how long this will take to drain, if * the calculation comes out to zero, thats ok we * will use send_a_lot to possibly spin around for * more increasing tot_len_this_send to the point * that its going to require a pace, or we hit the * cwnd. Which in that case we are just waiting for * a ACK. */ slot = tot_len_this_send / tr_perms; /* Now do we reduce the time so we don't run dry? */ if (slot && rack->rc_pace_reduce) { int32_t reduce; reduce = (slot / rack->rc_pace_reduce); if (reduce < slot) { slot -= reduce; } else slot = 0; } if (rack->r_enforce_min_pace && (slot == 0) && (tot_len_this_send >= (rack->r_min_pace_seg_thresh * tp->t_maxseg))) { /* We are enforcing a minimum pace time of 1ms */ slot = rack->r_enforce_min_pace; } } SOCKBUF_UNLOCK(sb); } else { SOCKBUF_UNLOCK(sb); if (tp->t_flags & TF_ACKNOW) TCPSTAT_INC(tcps_sndacks); else if (flags & (TH_SYN | TH_FIN | TH_RST)) TCPSTAT_INC(tcps_sndctrl); else if (SEQ_GT(tp->snd_up, tp->snd_una)) TCPSTAT_INC(tcps_sndurg); else TCPSTAT_INC(tcps_sndwinup); m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) { error = ENOBUFS; sack_rxmit = 0; goto out; } #ifdef INET6 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && MHLEN >= hdrlen) { M_ALIGN(m, hdrlen); } else #endif m->m_data += max_linkhdr; m->m_len = hdrlen; } SOCKBUF_UNLOCK_ASSERT(sb); m->m_pkthdr.rcvif = (struct ifnet *)0; #ifdef MAC mac_inpcb_create_mbuf(inp, m); #endif #ifdef INET6 if (isipv6) { ip6 = mtod(m, struct ip6_hdr *); #ifdef NETFLIX_TCPOUDP if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip6 + ipoptlen + sizeof(struct ip6_hdr)); udp->uh_sport = htons(V_tcp_udp_tunneling_port); udp->uh_dport = tp->t_port; ulen = hdrlen + len - sizeof(struct ip6_hdr); udp->uh_ulen = htons(ulen); th = (struct tcphdr *)(udp + 1); } else #endif th = (struct tcphdr *)(ip6 + 1); tcpip_fillheaders(inp, ip6, th); } else #endif /* INET6 */ { ip = mtod(m, struct ip *); #ifdef TCPDEBUG ipov = (struct ipovly *)ip; #endif #ifdef NETFLIX_TCPOUDP if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip + ipoptlen + sizeof(struct ip)); udp->uh_sport = htons(V_tcp_udp_tunneling_port); udp->uh_dport = tp->t_port; ulen = hdrlen + len - sizeof(struct ip); udp->uh_ulen = htons(ulen); th = (struct tcphdr *)(udp + 1); } else #endif th = (struct tcphdr *)(ip + 1); tcpip_fillheaders(inp, ip, th); } /* * Fill in fields, remembering maximum advertised window for use in * delaying messages about window sizes. If resending a FIN, be sure * not to use a new sequence number. */ if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && tp->snd_nxt == tp->snd_max) tp->snd_nxt--; /* * If we are starting a connection, send ECN setup SYN packet. If we * are on a retransmit, we may resend those bits a number of times * as per RFC 3168. */ if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) { if (tp->t_rxtshift >= 1) { if (tp->t_rxtshift <= V_tcp_ecn_maxretries) flags |= TH_ECE | TH_CWR; } else flags |= TH_ECE | TH_CWR; } if (tp->t_state == TCPS_ESTABLISHED && (tp->t_flags & TF_ECN_PERMIT)) { /* * If the peer has ECN, mark data packets with ECN capable * transmission (ECT). Ignore pure ack packets, * retransmissions and window probes. */ if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && !((tp->t_flags & TF_FORCEDATA) && len == 1)) { #ifdef INET6 if (isipv6) ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); else #endif ip->ip_tos |= IPTOS_ECN_ECT0; TCPSTAT_INC(tcps_ecn_ect0); } /* * Reply with proper ECN notifications. */ if (tp->t_flags & TF_ECN_SND_CWR) { flags |= TH_CWR; tp->t_flags &= ~TF_ECN_SND_CWR; } if (tp->t_flags & TF_ECN_SND_ECE) flags |= TH_ECE; } /* * If we are doing retransmissions, then snd_nxt will not reflect * the first unsent octet. For ACK only packets, we do not want the * sequence number of the retransmitted packet, we want the sequence * number of the next unsent octet. So, if there is no data (and no * SYN or FIN), use snd_max instead of snd_nxt when filling in * ti_seq. But if we are in persist state, snd_max might reflect * one byte beyond the right edge of the window, so use snd_nxt in * that case, since we know we aren't doing a retransmission. * (retransmit and persist are mutually exclusive...) */ if (sack_rxmit == 0) { if (len || (flags & (TH_SYN | TH_FIN)) || rack->rc_in_persist) { th->th_seq = htonl(tp->snd_nxt); rack_seq = tp->snd_nxt; } else if (flags & TH_RST) { /* * For a Reset send the last cum ack in sequence * (this like any other choice may still generate a * challenge ack, if a ack-update packet is in * flight). */ th->th_seq = htonl(tp->snd_una); rack_seq = tp->snd_una; } else { th->th_seq = htonl(tp->snd_max); rack_seq = tp->snd_max; } } else { th->th_seq = htonl(rsm->r_start); rack_seq = rsm->r_start; } th->th_ack = htonl(tp->rcv_nxt); if (optlen) { bcopy(opt, th + 1, optlen); th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; } th->th_flags = flags; /* * Calculate receive window. Don't shrink window, but avoid silly * window syndrome. */ if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && recwin < (long)tp->t_maxseg) recwin = 0; if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) recwin = (long)(tp->rcv_adv - tp->rcv_nxt); if (recwin > (long)TCP_MAXWIN << tp->rcv_scale) recwin = (long)TCP_MAXWIN << tp->rcv_scale; /* * According to RFC1323 the window field in a SYN (i.e., a or * ) segment itself is never scaled. The case is * handled in syncache. */ if (flags & TH_SYN) th->th_win = htons((u_short) (min(sbspace(&so->so_rcv), TCP_MAXWIN))); else th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); /* * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0 * window. This may cause the remote transmitter to stall. This * flag tells soreceive() to disable delayed acknowledgements when * draining the buffer. This can occur if the receiver is * attempting to read more data than can be buffered prior to * transmitting on the connection. */ if (th->th_win == 0) { tp->t_sndzerowin++; tp->t_flags |= TF_RXWIN0SENT; } else tp->t_flags &= ~TF_RXWIN0SENT; if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); th->th_flags |= TH_URG; } else /* * If no urgent pointer to send, then we pull the urgent * pointer to the left edge of the send window so that it * doesn't drift into the send window on sequence number * wraparound. */ tp->snd_up = tp->snd_una; /* drag it along */ #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (to.to_flags & TOF_SIGNATURE) { /* * Calculate MD5 signature and put it into the place * determined before. * NOTE: since TCP options buffer doesn't point into * mbuf's data, calculate offset and use it. */ if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th, (u_char *)(th + 1) + (to.to_signature - opt)) != 0) { /* * Do not send segment if the calculation of MD5 * digest has failed. */ goto out; } } #endif /* * Put TCP length in extended header, and then checksum extended * header and data. */ m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ #ifdef INET6 if (isipv6) { /* * ip6_plen is not need to be filled now, and will be filled * in ip6_output. */ if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); th->th_sum = htons(0); } else { m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP, 0); } } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET { if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); th->th_sum = htons(0); } else { m->m_pkthdr.csum_flags = CSUM_TCP; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); } /* IP version must be set here for ipv4/ipv6 checking later */ KASSERT(ip->ip_v == IPVERSION, ("%s: IP version incorrect: %d", __func__, ip->ip_v)); } #endif /* * Enable TSO and specify the size of the segments. The TCP pseudo * header checksum is always provided. XXX: Fixme: This is currently * not the case for IPv6. */ if (tso) { KASSERT(len > tp->t_maxseg - optlen, ("%s: len <= tso_segsz", __func__)); m->m_pkthdr.csum_flags |= CSUM_TSO; m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; } #if defined(IPSEC) || defined(IPSEC_SUPPORT) KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL), ("%s: mbuf chain shorter than expected: %d + %u + %u - %u != %u", __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL))); #else KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL), ("%s: mbuf chain shorter than expected: %d + %u + %u != %u", __func__, len, hdrlen, ipoptlen, m_length(m, NULL))); #endif #ifdef TCP_HHOOK /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ hhook_run_tcp_est_out(tp, th, &to, len, tso); #endif #ifdef TCPDEBUG /* * Trace. */ if (so->so_options & SO_DEBUG) { u_short save = 0; #ifdef INET6 if (!isipv6) #endif { save = ipov->ih_len; ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + * (th->th_off << 2) */ ); } tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); #ifdef INET6 if (!isipv6) #endif ipov->ih_len = save; } #endif /* TCPDEBUG */ /* We're getting ready to send; log now. */ if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt; if (rsm || sack_rxmit) { log.u_bbr.flex8 = 1; } else { log.u_bbr.flex8 = 0; } lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, len, &log, false, NULL, NULL, 0, NULL); } else lgb = NULL; /* * Fill in IP length and desired time to live and send to IP level. * There should be a better way to handle ttl and tos; we could keep * them in the template, but need a way to checksum without them. */ /* * m->m_pkthdr.len should have been set before cksum calcuration, * because in6_cksum() need it. */ #ifdef INET6 if (isipv6) { /* * we separately set hoplimit for every segment, since the * user might want to change the value via setsockopt. Also, * desired default hop limit might be changed via Neighbor * Discovery. */ ip6->ip6_hlim = in6_selecthlim(inp, NULL); /* * Set the packet size here for the benefit of DTrace * probes. ip6_output() will set it properly; it's supposed * to include the option header lengths as well. */ ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) tp->t_flags2 |= TF2_PLPMTU_PMTUD; else tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; if (tp->t_state == TCPS_SYN_SENT) TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); TCP_PROBE5(send, NULL, tp, ip6, tp, th); /* TODO: IPv6 IP6TOS_ECT bit on */ error = ip6_output(m, tp->t_inpcb->in6p_outputopts, &inp->inp_route6, ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), NULL, NULL, inp); if (error == EMSGSIZE && inp->inp_route6.ro_rt != NULL) mtu = inp->inp_route6.ro_rt->rt_mtu; } #endif /* INET6 */ #if defined(INET) && defined(INET6) else #endif #ifdef INET { ip->ip_len = htons(m->m_pkthdr.len); #ifdef INET6 if (inp->inp_vflag & INP_IPV6PROTO) ip->ip_ttl = in6_selecthlim(inp, NULL); #endif /* INET6 */ /* * If we do path MTU discovery, then we set DF on every * packet. This might not be the best thing to do according * to RFC3390 Section 2. However the tcp hostcache migitates * the problem so it affects only the first tcp connection * with a host. * * NB: Don't set DF on small MTU/MSS to have a safe * fallback. */ if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { tp->t_flags2 |= TF2_PLPMTU_PMTUD; if (tp->t_port == 0 || len < V_tcp_minmss) { ip->ip_off |= htons(IP_DF); } } else { tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; } if (tp->t_state == TCPS_SYN_SENT) TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); TCP_PROBE5(send, NULL, tp, ip, tp, th); error = ip_output(m, tp->t_inpcb->inp_options, &inp->inp_route, ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, inp); if (error == EMSGSIZE && inp->inp_route.ro_rt != NULL) mtu = inp->inp_route.ro_rt->rt_mtu; } #endif /* INET */ out: if (lgb) { lgb->tlb_errno = error; lgb = NULL; } /* * In transmit state, time the transmission and arrange for the * retransmit. In persist state, just set snd_max. */ if (error == 0) { if (len == 0) counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1); else if (len == 1) { counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1); } else if (len > 1) { int idx; idx = (len / tp->t_maxseg) + 3; if (idx >= TCP_MSS_ACCT_ATIMER) counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1); else counter_u64_add(rack_out_size[idx], 1); } } if (sub_from_prr && (error == 0)) { rack->r_ctl.rc_prr_sndcnt -= len; } sub_from_prr = 0; rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error, cts, pass, rsm); if ((tp->t_flags & TF_FORCEDATA) == 0 || (rack->rc_in_persist == 0)) { #ifdef NETFLIX_STATS tcp_seq startseq = tp->snd_nxt; #endif /* * Advance snd_nxt over sequence space of this segment. */ if (error) /* We don't log or do anything with errors */ goto timer; if (flags & (TH_SYN | TH_FIN)) { if (flags & TH_SYN) tp->snd_nxt++; if (flags & TH_FIN) { tp->snd_nxt++; tp->t_flags |= TF_SENTFIN; } } /* In the ENOBUFS case we do *not* update snd_max */ if (sack_rxmit) goto timer; tp->snd_nxt += len; if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { if (tp->snd_una == tp->snd_max) { /* * Update the time we just added data since * none was outstanding. */ rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__); tp->t_acktime = ticks; } tp->snd_max = tp->snd_nxt; #ifdef NETFLIX_STATS if (!(tp->t_flags & TF_GPUTINPROG) && len) { tp->t_flags |= TF_GPUTINPROG; tp->gput_seq = startseq; tp->gput_ack = startseq + ulmin(sbavail(sb) - sb_offset, sendwin); tp->gput_ts = tcp_ts_getticks(); } #endif } /* * Set retransmit timer if not currently set, and not doing * a pure ack or a keep-alive probe. Initial value for * retransmit timer is smoothed round-trip time + 2 * * round-trip time variance. Initialize shift counter which * is used for backoff of retransmit time. */ timer: if ((tp->snd_wnd == 0) && TCPS_HAVEESTABLISHED(tp->t_state)) { /* * If the persists timer was set above (right before * the goto send), and still needs to be on. Lets * make sure all is canceled. If the persist timer * is not running, we want to get it up. */ if (rack->rc_in_persist == 0) { rack_enter_persist(tp, rack, cts); } } } else { /* * Persist case, update snd_max but since we are in persist * mode (no window) we do not update snd_nxt. */ int32_t xlen = len; if (error) goto nomore; if (flags & TH_SYN) ++xlen; if (flags & TH_FIN) { ++xlen; tp->t_flags |= TF_SENTFIN; } /* In the ENOBUFS case we do *not* update snd_max */ if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) { if (tp->snd_una == tp->snd_max) { /* * Update the time we just added data since * none was outstanding. */ rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__); tp->t_acktime = ticks; } tp->snd_max = tp->snd_nxt + len; } } nomore: if (error) { SOCKBUF_UNLOCK_ASSERT(sb); /* Check gotos. */ /* * Failures do not advance the seq counter above. For the * case of ENOBUFS we will fall out and retry in 1ms with * the hpts. Everything else will just have to retransmit * with the timer. * * In any case, we do not want to loop around for another * send without a good reason. */ sendalot = 0; switch (error) { case EPERM: tp->t_flags &= ~TF_FORCEDATA; tp->t_softerror = error; return (error); case ENOBUFS: if (slot == 0) { /* * Pace us right away to retry in a some * time */ slot = 1 + rack->rc_enobuf; if (rack->rc_enobuf < 255) rack->rc_enobuf++; if (slot > (rack->rc_rack_rtt / 2)) { slot = rack->rc_rack_rtt / 2; } if (slot < 10) slot = 10; } counter_u64_add(rack_saw_enobuf, 1); error = 0; goto enobufs; case EMSGSIZE: /* * For some reason the interface we used initially * to send segments changed to another or lowered * its MTU. If TSO was active we either got an * interface without TSO capabilits or TSO was * turned off. If we obtained mtu from ip_output() * then update it and try again. */ if (tso) tp->t_flags &= ~TF_TSO; if (mtu != 0) { tcp_mss_update(tp, -1, mtu, NULL, NULL); goto again; } slot = 10; rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, 0, 1); tp->t_flags &= ~TF_FORCEDATA; return (error); case ENETUNREACH: counter_u64_add(rack_saw_enetunreach, 1); case EHOSTDOWN: case EHOSTUNREACH: case ENETDOWN: if (TCPS_HAVERCVDSYN(tp->t_state)) { tp->t_softerror = error; } /* FALLTHROUGH */ default: slot = 10; rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, 0, 1); tp->t_flags &= ~TF_FORCEDATA; return (error); } } else { rack->rc_enobuf = 0; } TCPSTAT_INC(tcps_sndtotal); /* * Data sent (as far as we can tell). If this advertises a larger * window than any other segment, then remember the size of the * advertised window. Any pending ACK has now been sent. */ if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) tp->rcv_adv = tp->rcv_nxt + recwin; tp->last_ack_sent = tp->rcv_nxt; tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); enobufs: rack->r_tlp_running = 0; if ((flags & TH_RST) || (would_have_fin == 1)) { /* * We don't send again after a RST. We also do *not* send * again if we would have had a find, but now have * outstanding data. */ slot = 0; sendalot = 0; } if (slot) { /* set the rack tcb into the slot N */ counter_u64_add(rack_paced_segments, 1); } else if (sendalot) { if (len) counter_u64_add(rack_unpaced_segments, 1); sack_rxmit = 0; tp->t_flags &= ~TF_FORCEDATA; goto again; } else if (len) { counter_u64_add(rack_unpaced_segments, 1); } tp->t_flags &= ~TF_FORCEDATA; rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, tot_len_this_send, 1); return (error); } /* * rack_ctloutput() must drop the inpcb lock before performing copyin on * socket option arguments. When it re-acquires the lock after the copy, it * has to revalidate that the connection is still valid for the socket * option. */ static int rack_set_sockopt(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack) { int32_t error = 0, optval; switch (sopt->sopt_name) { case TCP_RACK_PROP_RATE: case TCP_RACK_PROP: case TCP_RACK_TLP_REDUCE: case TCP_RACK_EARLY_RECOV: case TCP_RACK_PACE_ALWAYS: case TCP_DELACK: case TCP_RACK_PACE_REDUCE: case TCP_RACK_PACE_MAX_SEG: case TCP_RACK_PRR_SENDALOT: case TCP_RACK_MIN_TO: case TCP_RACK_EARLY_SEG: case TCP_RACK_REORD_THRESH: case TCP_RACK_REORD_FADE: case TCP_RACK_TLP_THRESH: case TCP_RACK_PKT_DELAY: case TCP_RACK_TLP_USE: case TCP_RACK_TLP_INC_VAR: case TCP_RACK_IDLE_REDUCE_HIGH: case TCP_RACK_MIN_PACE: case TCP_RACK_MIN_PACE_SEG: case TCP_BBR_RACK_RTT_USE: case TCP_DATA_AFTER_CLOSE: break; default: return (tcp_default_ctloutput(so, sopt, inp, tp)); break; } INP_WUNLOCK(inp); error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); if (error) return (error); INP_WLOCK(inp); if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { INP_WUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); rack = (struct tcp_rack *)tp->t_fb_ptr; switch (sopt->sopt_name) { case TCP_RACK_PROP_RATE: if ((optval <= 0) || (optval >= 100)) { error = EINVAL; break; } RACK_OPTS_INC(tcp_rack_prop_rate); rack->r_ctl.rc_prop_rate = optval; break; case TCP_RACK_TLP_USE: if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) { error = EINVAL; break; } RACK_OPTS_INC(tcp_tlp_use); rack->rack_tlp_threshold_use = optval; break; case TCP_RACK_PROP: /* RACK proportional rate reduction (bool) */ RACK_OPTS_INC(tcp_rack_prop); rack->r_ctl.rc_prop_reduce = optval; break; case TCP_RACK_TLP_REDUCE: /* RACK TLP cwnd reduction (bool) */ RACK_OPTS_INC(tcp_rack_tlp_reduce); rack->r_ctl.rc_tlp_cwnd_reduce = optval; break; case TCP_RACK_EARLY_RECOV: /* Should recovery happen early (bool) */ RACK_OPTS_INC(tcp_rack_early_recov); rack->r_ctl.rc_early_recovery = optval; break; case TCP_RACK_PACE_ALWAYS: /* Use the always pace method (bool) */ RACK_OPTS_INC(tcp_rack_pace_always); if (optval > 0) rack->rc_always_pace = 1; else rack->rc_always_pace = 0; break; case TCP_RACK_PACE_REDUCE: /* RACK Hptsi reduction factor (divisor) */ RACK_OPTS_INC(tcp_rack_pace_reduce); if (optval) /* Must be non-zero */ rack->rc_pace_reduce = optval; else error = EINVAL; break; case TCP_RACK_PACE_MAX_SEG: /* Max segments in a pace */ RACK_OPTS_INC(tcp_rack_max_seg); rack->rc_pace_max_segs = optval; break; case TCP_RACK_PRR_SENDALOT: /* Allow PRR to send more than one seg */ RACK_OPTS_INC(tcp_rack_prr_sendalot); rack->r_ctl.rc_prr_sendalot = optval; break; case TCP_RACK_MIN_TO: /* Minimum time between rack t-o's in ms */ RACK_OPTS_INC(tcp_rack_min_to); rack->r_ctl.rc_min_to = optval; break; case TCP_RACK_EARLY_SEG: /* If early recovery max segments */ RACK_OPTS_INC(tcp_rack_early_seg); rack->r_ctl.rc_early_recovery_segs = optval; break; case TCP_RACK_REORD_THRESH: /* RACK reorder threshold (shift amount) */ RACK_OPTS_INC(tcp_rack_reord_thresh); if ((optval > 0) && (optval < 31)) rack->r_ctl.rc_reorder_shift = optval; else error = EINVAL; break; case TCP_RACK_REORD_FADE: /* Does reordering fade after ms time */ RACK_OPTS_INC(tcp_rack_reord_fade); rack->r_ctl.rc_reorder_fade = optval; break; case TCP_RACK_TLP_THRESH: /* RACK TLP theshold i.e. srtt+(srtt/N) */ RACK_OPTS_INC(tcp_rack_tlp_thresh); if (optval) rack->r_ctl.rc_tlp_threshold = optval; else error = EINVAL; break; case TCP_RACK_PKT_DELAY: /* RACK added ms i.e. rack-rtt + reord + N */ RACK_OPTS_INC(tcp_rack_pkt_delay); rack->r_ctl.rc_pkt_delay = optval; break; case TCP_RACK_TLP_INC_VAR: /* Does TLP include rtt variance in t-o */ RACK_OPTS_INC(tcp_rack_tlp_inc_var); rack->r_ctl.rc_prr_inc_var = optval; break; case TCP_RACK_IDLE_REDUCE_HIGH: RACK_OPTS_INC(tcp_rack_idle_reduce_high); if (optval) rack->r_idle_reduce_largest = 1; else rack->r_idle_reduce_largest = 0; break; case TCP_DELACK: if (optval == 0) tp->t_delayed_ack = 0; else tp->t_delayed_ack = 1; if (tp->t_flags & TF_DELACK) { tp->t_flags &= ~TF_DELACK; tp->t_flags |= TF_ACKNOW; rack_output(tp); } break; case TCP_RACK_MIN_PACE: RACK_OPTS_INC(tcp_rack_min_pace); if (optval > 3) rack->r_enforce_min_pace = 3; else rack->r_enforce_min_pace = optval; break; case TCP_RACK_MIN_PACE_SEG: RACK_OPTS_INC(tcp_rack_min_pace_seg); if (optval >= 16) rack->r_min_pace_seg_thresh = 15; else rack->r_min_pace_seg_thresh = optval; break; case TCP_BBR_RACK_RTT_USE: if ((optval != USE_RTT_HIGH) && (optval != USE_RTT_LOW) && (optval != USE_RTT_AVG)) error = EINVAL; else rack->r_ctl.rc_rate_sample_method = optval; break; case TCP_DATA_AFTER_CLOSE: if (optval) rack->rc_allow_data_af_clo = 1; else rack->rc_allow_data_af_clo = 0; break; default: return (tcp_default_ctloutput(so, sopt, inp, tp)); break; } #ifdef NETFLIX_STATS tcp_log_socket_option(tp, sopt->sopt_name, optval, error); #endif INP_WUNLOCK(inp); return (error); } static int rack_get_sockopt(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack) { int32_t error, optval; /* * Because all our options are either boolean or an int, we can just * pull everything into optval and then unlock and copy. If we ever * add a option that is not a int, then this will have quite an * impact to this routine. */ switch (sopt->sopt_name) { case TCP_RACK_PROP_RATE: optval = rack->r_ctl.rc_prop_rate; break; case TCP_RACK_PROP: /* RACK proportional rate reduction (bool) */ optval = rack->r_ctl.rc_prop_reduce; break; case TCP_RACK_TLP_REDUCE: /* RACK TLP cwnd reduction (bool) */ optval = rack->r_ctl.rc_tlp_cwnd_reduce; break; case TCP_RACK_EARLY_RECOV: /* Should recovery happen early (bool) */ optval = rack->r_ctl.rc_early_recovery; break; case TCP_RACK_PACE_REDUCE: /* RACK Hptsi reduction factor (divisor) */ optval = rack->rc_pace_reduce; break; case TCP_RACK_PACE_MAX_SEG: /* Max segments in a pace */ optval = rack->rc_pace_max_segs; break; case TCP_RACK_PACE_ALWAYS: /* Use the always pace method */ optval = rack->rc_always_pace; break; case TCP_RACK_PRR_SENDALOT: /* Allow PRR to send more than one seg */ optval = rack->r_ctl.rc_prr_sendalot; break; case TCP_RACK_MIN_TO: /* Minimum time between rack t-o's in ms */ optval = rack->r_ctl.rc_min_to; break; case TCP_RACK_EARLY_SEG: /* If early recovery max segments */ optval = rack->r_ctl.rc_early_recovery_segs; break; case TCP_RACK_REORD_THRESH: /* RACK reorder threshold (shift amount) */ optval = rack->r_ctl.rc_reorder_shift; break; case TCP_RACK_REORD_FADE: /* Does reordering fade after ms time */ optval = rack->r_ctl.rc_reorder_fade; break; case TCP_RACK_TLP_THRESH: /* RACK TLP theshold i.e. srtt+(srtt/N) */ optval = rack->r_ctl.rc_tlp_threshold; break; case TCP_RACK_PKT_DELAY: /* RACK added ms i.e. rack-rtt + reord + N */ optval = rack->r_ctl.rc_pkt_delay; break; case TCP_RACK_TLP_USE: optval = rack->rack_tlp_threshold_use; break; case TCP_RACK_TLP_INC_VAR: /* Does TLP include rtt variance in t-o */ optval = rack->r_ctl.rc_prr_inc_var; break; case TCP_RACK_IDLE_REDUCE_HIGH: optval = rack->r_idle_reduce_largest; break; case TCP_RACK_MIN_PACE: optval = rack->r_enforce_min_pace; break; case TCP_RACK_MIN_PACE_SEG: optval = rack->r_min_pace_seg_thresh; break; case TCP_BBR_RACK_RTT_USE: optval = rack->r_ctl.rc_rate_sample_method; break; case TCP_DELACK: optval = tp->t_delayed_ack; break; case TCP_DATA_AFTER_CLOSE: optval = rack->rc_allow_data_af_clo; break; default: return (tcp_default_ctloutput(so, sopt, inp, tp)); break; } INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof optval); return (error); } static int rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp) { int32_t error = EINVAL; struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack == NULL) { /* Huh? */ goto out; } if (sopt->sopt_dir == SOPT_SET) { return (rack_set_sockopt(so, sopt, inp, tp, rack)); } else if (sopt->sopt_dir == SOPT_GET) { return (rack_get_sockopt(so, sopt, inp, tp, rack)); } out: INP_WUNLOCK(inp); return (error); } struct tcp_function_block __tcp_rack = { .tfb_tcp_block_name = __XSTRING(STACKNAME), .tfb_tcp_output = rack_output, .tfb_tcp_do_segment = rack_do_segment, .tfb_tcp_hpts_do_segment = rack_hpts_do_segment, .tfb_tcp_ctloutput = rack_ctloutput, .tfb_tcp_fb_init = rack_init, .tfb_tcp_fb_fini = rack_fini, .tfb_tcp_timer_stop_all = rack_stopall, .tfb_tcp_timer_activate = rack_timer_activate, .tfb_tcp_timer_active = rack_timer_active, .tfb_tcp_timer_stop = rack_timer_stop, .tfb_tcp_rexmit_tmr = rack_remxt_tmr, .tfb_tcp_handoff_ok = rack_handoff_ok }; static const char *rack_stack_names[] = { __XSTRING(STACKNAME), #ifdef STACKALIAS __XSTRING(STACKALIAS), #endif }; static int rack_ctor(void *mem, int32_t size, void *arg, int32_t how) { memset(mem, 0, size); return (0); } static void rack_dtor(void *mem, int32_t size, void *arg) { } static bool rack_mod_inited = false; static int tcp_addrack(module_t mod, int32_t type, void *data) { int32_t err = 0; int num_stacks; switch (type) { case MOD_LOAD: rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map", sizeof(struct rack_sendmap), rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0); rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb", sizeof(struct tcp_rack), rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); sysctl_ctx_init(&rack_sysctl_ctx); rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_STATIC_CHILDREN(_net_inet_tcp), OID_AUTO, __XSTRING(STACKNAME), CTLFLAG_RW, 0, ""); if (rack_sysctl_root == NULL) { printf("Failed to add sysctl node\n"); err = EFAULT; goto free_uma; } rack_init_sysctls(); num_stacks = nitems(rack_stack_names); err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK, rack_stack_names, &num_stacks); if (err) { printf("Failed to register %s stack name for " "%s module\n", rack_stack_names[num_stacks], __XSTRING(MODNAME)); sysctl_ctx_free(&rack_sysctl_ctx); free_uma: uma_zdestroy(rack_zone); uma_zdestroy(rack_pcb_zone); rack_counter_destroy(); printf("Failed to register rack module -- err:%d\n", err); return (err); } rack_mod_inited = true; break; case MOD_QUIESCE: err = deregister_tcp_functions(&__tcp_rack, true, false); break; case MOD_UNLOAD: err = deregister_tcp_functions(&__tcp_rack, false, true); if (err == EBUSY) break; if (rack_mod_inited) { uma_zdestroy(rack_zone); uma_zdestroy(rack_pcb_zone); sysctl_ctx_free(&rack_sysctl_ctx); rack_counter_destroy(); rack_mod_inited = false; } err = 0; break; default: return (EOPNOTSUPP); } return (err); } static moduledata_t tcp_rack = { .name = __XSTRING(MODNAME), .evhand = tcp_addrack, .priv = 0 }; MODULE_VERSION(MODNAME, 1); DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY); MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1); Index: head/sys/netinet/tcp_timewait.c =================================================================== --- head/sys/netinet/tcp_timewait.c (revision 335355) +++ head/sys/netinet/tcp_timewait.c (revision 335356) @@ -1,761 +1,753 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 * The Regents of the University of California. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95 */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_tcpdebug.h" #include #include #include #include #include #include #include #include #include #include #include #ifndef INVARIANTS #include #endif #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef INET6 #include #include #include #include #include #endif #include #include #include #include #include #ifdef INET6 #include #endif #include #ifdef TCPDEBUG #include #endif #ifdef INET6 #include #endif #include #include static VNET_DEFINE(uma_zone_t, tcptw_zone); #define V_tcptw_zone VNET(tcptw_zone) static int maxtcptw; /* * The timed wait queue contains references to each of the TCP sessions * currently in the TIME_WAIT state. The queue pointers, including the * queue pointers in each tcptw structure, are protected using the global * timewait lock, which must be held over queue iteration and modification. * * Rules on tcptw usage: * - a inpcb is always freed _after_ its tcptw * - a tcptw relies on its inpcb reference counting for memory stability * - a tcptw is dereferenceable only while its inpcb is locked */ static VNET_DEFINE(TAILQ_HEAD(, tcptw), twq_2msl); #define V_twq_2msl VNET(twq_2msl) /* Global timewait lock */ static VNET_DEFINE(struct rwlock, tw_lock); #define V_tw_lock VNET(tw_lock) #define TW_LOCK_INIT(tw, d) rw_init_flags(&(tw), (d), 0) #define TW_LOCK_DESTROY(tw) rw_destroy(&(tw)) #define TW_RLOCK(tw) rw_rlock(&(tw)) #define TW_WLOCK(tw) rw_wlock(&(tw)) #define TW_RUNLOCK(tw) rw_runlock(&(tw)) #define TW_WUNLOCK(tw) rw_wunlock(&(tw)) #define TW_LOCK_ASSERT(tw) rw_assert(&(tw), RA_LOCKED) #define TW_RLOCK_ASSERT(tw) rw_assert(&(tw), RA_RLOCKED) #define TW_WLOCK_ASSERT(tw) rw_assert(&(tw), RA_WLOCKED) #define TW_UNLOCK_ASSERT(tw) rw_assert(&(tw), RA_UNLOCKED) static void tcp_tw_2msl_reset(struct tcptw *, int); static void tcp_tw_2msl_stop(struct tcptw *, int); static int tcp_twrespond(struct tcptw *, int); static int tcptw_auto_size(void) { int halfrange; /* * Max out at half the ephemeral port range so that TIME_WAIT * sockets don't tie up too many ephemeral ports. */ if (V_ipport_lastauto > V_ipport_firstauto) halfrange = (V_ipport_lastauto - V_ipport_firstauto) / 2; else halfrange = (V_ipport_firstauto - V_ipport_lastauto) / 2; /* Protect against goofy port ranges smaller than 32. */ return (imin(imax(halfrange, 32), maxsockets / 5)); } static int sysctl_maxtcptw(SYSCTL_HANDLER_ARGS) { int error, new; if (maxtcptw == 0) new = tcptw_auto_size(); else new = maxtcptw; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr) if (new >= 32) { maxtcptw = new; uma_zone_set_max(V_tcptw_zone, maxtcptw); } return (error); } SYSCTL_PROC(_net_inet_tcp, OID_AUTO, maxtcptw, CTLTYPE_INT|CTLFLAG_RW, &maxtcptw, 0, sysctl_maxtcptw, "IU", "Maximum number of compressed TCP TIME_WAIT entries"); static VNET_DEFINE(int, nolocaltimewait) = 0; #define V_nolocaltimewait VNET(nolocaltimewait) SYSCTL_INT(_net_inet_tcp, OID_AUTO, nolocaltimewait, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(nolocaltimewait), 0, "Do not create compressed TCP TIME_WAIT entries for local connections"); void tcp_tw_zone_change(void) { if (maxtcptw == 0) uma_zone_set_max(V_tcptw_zone, tcptw_auto_size()); } void tcp_tw_init(void) { V_tcptw_zone = uma_zcreate("tcptw", sizeof(struct tcptw), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); TUNABLE_INT_FETCH("net.inet.tcp.maxtcptw", &maxtcptw); if (maxtcptw == 0) uma_zone_set_max(V_tcptw_zone, tcptw_auto_size()); else uma_zone_set_max(V_tcptw_zone, maxtcptw); TAILQ_INIT(&V_twq_2msl); TW_LOCK_INIT(V_tw_lock, "tcptw"); } #ifdef VIMAGE void tcp_tw_destroy(void) { struct tcptw *tw; INP_INFO_RLOCK(&V_tcbinfo); while ((tw = TAILQ_FIRST(&V_twq_2msl)) != NULL) tcp_twclose(tw, 0); INP_INFO_RUNLOCK(&V_tcbinfo); TW_LOCK_DESTROY(V_tw_lock); uma_zdestroy(V_tcptw_zone); } #endif /* * Move a TCP connection into TIME_WAIT state. * tcbinfo is locked. * inp is locked, and is unlocked before returning. */ void tcp_twstart(struct tcpcb *tp) { struct tcptw twlocal, *tw; struct inpcb *inp = tp->t_inpcb; struct socket *so; bool acknow, local; #ifdef INET6 bool isipv6 = inp->inp_inc.inc_flags & INC_ISIPV6; #endif INP_INFO_RLOCK_ASSERT(&V_tcbinfo); INP_WLOCK_ASSERT(inp); /* A dropped inp should never transition to TIME_WAIT state. */ KASSERT((inp->inp_flags & INP_DROPPED) == 0, ("tcp_twstart: " "(inp->inp_flags & INP_DROPPED) != 0")); if (V_nolocaltimewait) { #ifdef INET6 if (isipv6) local = in6_localaddr(&inp->in6p_faddr); else #endif #ifdef INET local = in_localip(inp->inp_faddr); #else local = false; #endif } else local = false; /* * For use only by DTrace. We do not reference the state * after this point so modifying it in place is not a problem. */ tcp_state_change(tp, TCPS_TIME_WAIT); if (local) tw = &twlocal; else tw = uma_zalloc(V_tcptw_zone, M_NOWAIT); if (tw == NULL) { /* * Reached limit on total number of TIMEWAIT connections * allowed. Remove a connection from TIMEWAIT queue in LRU * fashion to make room for this connection. * * XXX: Check if it possible to always have enough room * in advance based on guarantees provided by uma_zalloc(). */ tw = tcp_tw_2msl_scan(1); if (tw == NULL) { tp = tcp_close(tp); if (tp != NULL) INP_WUNLOCK(inp); return; } } /* * For !local case the tcptw will hold a reference on its inpcb * until tcp_twclose is called. */ tw->tw_inpcb = inp; /* * Recover last window size sent. */ if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) tw->last_win = (tp->rcv_adv - tp->rcv_nxt) >> tp->rcv_scale; else tw->last_win = 0; /* * Set t_recent if timestamps are used on the connection. */ if ((tp->t_flags & (TF_REQ_TSTMP|TF_RCVD_TSTMP|TF_NOOPT)) == (TF_REQ_TSTMP|TF_RCVD_TSTMP)) { tw->t_recent = tp->ts_recent; tw->ts_offset = tp->ts_offset; } else { tw->t_recent = 0; tw->ts_offset = 0; } tw->snd_nxt = tp->snd_nxt; tw->rcv_nxt = tp->rcv_nxt; tw->iss = tp->iss; tw->irs = tp->irs; tw->t_starttime = tp->t_starttime; tw->tw_time = 0; /* XXX * If this code will * be used for fin-wait-2 state also, then we may need * a ts_recent from the last segment. */ acknow = tp->t_flags & TF_ACKNOW; /* * First, discard tcpcb state, which includes stopping its timers and * freeing it. tcp_discardcb() used to also release the inpcb, but * that work is now done in the caller. * * Note: soisdisconnected() call used to be made in tcp_discardcb(), * and might not be needed here any longer. */ tcp_discardcb(tp); so = inp->inp_socket; soisdisconnected(so); tw->tw_so_options = so->so_options; inp->inp_flags |= INP_TIMEWAIT; if (acknow) tcp_twrespond(tw, TH_ACK); if (local) in_pcbdrop(inp); else { in_pcbref(inp); /* Reference from tw */ tw->tw_cred = crhold(so->so_cred); inp->inp_ppcb = tw; TCPSTATES_INC(TCPS_TIME_WAIT); tcp_tw_2msl_reset(tw, 0); } /* * If the inpcb owns the sole reference to the socket, then we can * detach and free the socket as it is not needed in time wait. */ if (inp->inp_flags & INP_SOCKREF) { KASSERT(so->so_state & SS_PROTOREF, ("tcp_twstart: !SS_PROTOREF")); inp->inp_flags &= ~INP_SOCKREF; INP_WUNLOCK(inp); SOCK_LOCK(so); so->so_state &= ~SS_PROTOREF; sofree(so); } else INP_WUNLOCK(inp); } /* * Returns 1 if the TIME_WAIT state was killed and we should start over, * looking for a pcb in the listen state. Returns 0 otherwise. */ int tcp_twcheck(struct inpcb *inp, struct tcpopt *to __unused, struct tcphdr *th, struct mbuf *m, int tlen) { struct tcptw *tw; int thflags; tcp_seq seq; INP_INFO_RLOCK_ASSERT(&V_tcbinfo); INP_WLOCK_ASSERT(inp); /* * XXXRW: Time wait state for inpcb has been recycled, but inpcb is * still present. This is undesirable, but temporarily necessary * until we work out how to handle inpcb's who's timewait state has * been removed. */ tw = intotw(inp); if (tw == NULL) goto drop; thflags = th->th_flags; /* * NOTE: for FIN_WAIT_2 (to be added later), * must validate sequence number before accepting RST */ /* * If the segment contains RST: * Drop the segment - see Stevens, vol. 2, p. 964 and * RFC 1337. */ if (thflags & TH_RST) goto drop; #if 0 /* PAWS not needed at the moment */ /* * RFC 1323 PAWS: If we have a timestamp reply on this segment * and it's less than ts_recent, drop it. */ if ((to.to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to.to_tsval, tp->ts_recent)) { if ((thflags & TH_ACK) == 0) goto drop; goto ack; } /* * ts_recent is never updated because we never accept new segments. */ #endif /* * If a new connection request is received * while in TIME_WAIT, drop the old connection * and start over if the sequence numbers * are above the previous ones. */ if ((thflags & TH_SYN) && SEQ_GT(th->th_seq, tw->rcv_nxt)) { tcp_twclose(tw, 0); return (1); } /* * Drop the segment if it does not contain an ACK. */ if ((thflags & TH_ACK) == 0) goto drop; /* * Reset the 2MSL timer if this is a duplicate FIN. */ if (thflags & TH_FIN) { seq = th->th_seq + tlen + (thflags & TH_SYN ? 1 : 0); if (seq + 1 == tw->rcv_nxt) tcp_tw_2msl_reset(tw, 1); } /* * Acknowledge the segment if it has data or is not a duplicate ACK. */ if (thflags != TH_ACK || tlen != 0 || th->th_seq != tw->rcv_nxt || th->th_ack != tw->snd_nxt) tcp_twrespond(tw, TH_ACK); drop: INP_WUNLOCK(inp); m_freem(m); return (0); } void tcp_twclose(struct tcptw *tw, int reuse) { struct socket *so; struct inpcb *inp; /* * At this point, we are in one of two situations: * * (1) We have no socket, just an inpcb<->twtcp pair. We can free * all state. * * (2) We have a socket -- if we own a reference, release it and * notify the socket layer. */ inp = tw->tw_inpcb; KASSERT((inp->inp_flags & INP_TIMEWAIT), ("tcp_twclose: !timewait")); KASSERT(intotw(inp) == tw, ("tcp_twclose: inp_ppcb != tw")); INP_INFO_RLOCK_ASSERT(&V_tcbinfo); /* in_pcbfree() */ INP_WLOCK_ASSERT(inp); tcp_tw_2msl_stop(tw, reuse); inp->inp_ppcb = NULL; in_pcbdrop(inp); so = inp->inp_socket; if (so != NULL) { /* * If there's a socket, handle two cases: first, we own a * strong reference, which we will now release, or we don't * in which case another reference exists (XXXRW: think * about this more), and we don't need to take action. */ if (inp->inp_flags & INP_SOCKREF) { inp->inp_flags &= ~INP_SOCKREF; INP_WUNLOCK(inp); SOCK_LOCK(so); KASSERT(so->so_state & SS_PROTOREF, ("tcp_twclose: INP_SOCKREF && !SS_PROTOREF")); so->so_state &= ~SS_PROTOREF; sofree(so); } else { /* * If we don't own the only reference, the socket and * inpcb need to be left around to be handled by * tcp_usr_detach() later. */ INP_WUNLOCK(inp); } } else { /* * The socket has been already cleaned-up for us, only free the * inpcb. */ in_pcbfree(inp); } TCPSTAT_INC(tcps_closed); } static int tcp_twrespond(struct tcptw *tw, int flags) { struct inpcb *inp = tw->tw_inpcb; #if defined(INET6) || defined(INET) struct tcphdr *th = NULL; #endif struct mbuf *m; #ifdef INET struct ip *ip = NULL; #endif u_int hdrlen, optlen; int error = 0; /* Keep compiler happy */ struct tcpopt to; #ifdef INET6 struct ip6_hdr *ip6 = NULL; int isipv6 = inp->inp_inc.inc_flags & INC_ISIPV6; #endif hdrlen = 0; /* Keep compiler happy */ INP_WLOCK_ASSERT(inp); m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) return (ENOBUFS); m->m_data += max_linkhdr; #ifdef MAC mac_inpcb_create_mbuf(inp, m); #endif #ifdef INET6 if (isipv6) { hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); ip6 = mtod(m, struct ip6_hdr *); th = (struct tcphdr *)(ip6 + 1); tcpip_fillheaders(inp, ip6, th); } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET { hdrlen = sizeof(struct tcpiphdr); ip = mtod(m, struct ip *); th = (struct tcphdr *)(ip + 1); tcpip_fillheaders(inp, ip, th); } #endif to.to_flags = 0; /* * Send a timestamp and echo-reply if both our side and our peer * have sent timestamps in our SYN's and this is not a RST. */ if (tw->t_recent && flags == TH_ACK) { to.to_flags |= TOF_TS; to.to_tsval = tcp_ts_getticks() + tw->ts_offset; to.to_tsecr = tw->t_recent; } optlen = tcp_addoptions(&to, (u_char *)(th + 1)); m->m_len = hdrlen + optlen; m->m_pkthdr.len = m->m_len; KASSERT(max_linkhdr + m->m_len <= MHLEN, ("tcptw: mbuf too small")); th->th_seq = htonl(tw->snd_nxt); th->th_ack = htonl(tw->rcv_nxt); th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; th->th_flags = flags; th->th_win = htons(tw->last_win); m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); #ifdef INET6 if (isipv6) { m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + optlen, IPPROTO_TCP, 0); ip6->ip6_hlim = in6_selecthlim(inp, NULL); error = ip6_output(m, inp->in6p_outputopts, NULL, (tw->tw_so_options & SO_DONTROUTE), NULL, NULL, inp); } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET { m->m_pkthdr.csum_flags = CSUM_TCP; th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + optlen + IPPROTO_TCP)); ip->ip_len = htons(m->m_pkthdr.len); if (V_path_mtu_discovery) ip->ip_off |= htons(IP_DF); error = ip_output(m, inp->inp_options, NULL, ((tw->tw_so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), NULL, inp); } #endif if (flags & TH_ACK) TCPSTAT_INC(tcps_sndacks); else TCPSTAT_INC(tcps_sndctrl); TCPSTAT_INC(tcps_sndtotal); return (error); } static void tcp_tw_2msl_reset(struct tcptw *tw, int rearm) { INP_INFO_RLOCK_ASSERT(&V_tcbinfo); INP_WLOCK_ASSERT(tw->tw_inpcb); TW_WLOCK(V_tw_lock); if (rearm) TAILQ_REMOVE(&V_twq_2msl, tw, tw_2msl); tw->tw_time = ticks + 2 * tcp_msl; TAILQ_INSERT_TAIL(&V_twq_2msl, tw, tw_2msl); TW_WUNLOCK(V_tw_lock); } static void tcp_tw_2msl_stop(struct tcptw *tw, int reuse) { struct ucred *cred; struct inpcb *inp; int released __unused; INP_INFO_RLOCK_ASSERT(&V_tcbinfo); TW_WLOCK(V_tw_lock); inp = tw->tw_inpcb; tw->tw_inpcb = NULL; TAILQ_REMOVE(&V_twq_2msl, tw, tw_2msl); cred = tw->tw_cred; tw->tw_cred = NULL; TW_WUNLOCK(V_tw_lock); if (cred != NULL) crfree(cred); released = in_pcbrele_wlocked(inp); KASSERT(!released, ("%s: inp should not be released here", __func__)); if (!reuse) uma_zfree(V_tcptw_zone, tw); TCPSTATES_DEC(TCPS_TIME_WAIT); } struct tcptw * tcp_tw_2msl_scan(int reuse) { struct tcptw *tw; struct inpcb *inp; #ifdef INVARIANTS if (reuse) { /* * Exclusive pcbinfo lock is not required in reuse case even if * two inpcb locks can be acquired simultaneously: * - the inpcb transitioning to TIME_WAIT state in * tcp_tw_start(), * - the inpcb closed by tcp_twclose(). * * It is because only inpcbs in FIN_WAIT2 or CLOSING states can * transition in TIME_WAIT state. Then a pcbcb cannot be in * TIME_WAIT list and transitioning to TIME_WAIT state at same * time. */ INP_INFO_RLOCK_ASSERT(&V_tcbinfo); } #endif for (;;) { TW_RLOCK(V_tw_lock); tw = TAILQ_FIRST(&V_twq_2msl); if (tw == NULL || (!reuse && (tw->tw_time - ticks) > 0)) { TW_RUNLOCK(V_tw_lock); break; } KASSERT(tw->tw_inpcb != NULL, ("%s: tw->tw_inpcb == NULL", __func__)); inp = tw->tw_inpcb; in_pcbref(inp); TW_RUNLOCK(V_tw_lock); - if (INP_INFO_TRY_RLOCK(&V_tcbinfo)) { - - INP_WLOCK(inp); - tw = intotw(inp); - if (in_pcbrele_wlocked(inp)) { - if (__predict_true(tw == NULL)) { - INP_INFO_RUNLOCK(&V_tcbinfo); - continue; - } else { - /* This should not happen as in TIMEWAIT - * state the inp should not be destroyed - * before its tcptw. If INVARIANTS is - * defined panic. - */ + INP_INFO_RLOCK(&V_tcbinfo); + INP_WLOCK(inp); + tw = intotw(inp); + if (in_pcbrele_wlocked(inp)) { + if (__predict_true(tw == NULL)) { + INP_INFO_RUNLOCK(&V_tcbinfo); + continue; + } else { + /* This should not happen as in TIMEWAIT + * state the inp should not be destroyed + * before its tcptw. If INVARIANTS is + * defined panic. + */ #ifdef INVARIANTS - panic("%s: Panic before an infinite " - "loop: INP_TIMEWAIT && (INP_FREED " - "|| inp last reference) && tw != " - "NULL", __func__); + panic("%s: Panic before an infinite " + "loop: INP_TIMEWAIT && (INP_FREED " + "|| inp last reference) && tw != " + "NULL", __func__); #else - log(LOG_ERR, "%s: Avoid an infinite " - "loop: INP_TIMEWAIT && (INP_FREED " - "|| inp last reference) && tw != " - "NULL", __func__); + log(LOG_ERR, "%s: Avoid an infinite " + "loop: INP_TIMEWAIT && (INP_FREED " + "|| inp last reference) && tw != " + "NULL", __func__); #endif - INP_INFO_RUNLOCK(&V_tcbinfo); - break; - } - } - - if (tw == NULL) { - /* tcp_twclose() has already been called */ - INP_WUNLOCK(inp); INP_INFO_RUNLOCK(&V_tcbinfo); - continue; + break; } + } - tcp_twclose(tw, reuse); + if (tw == NULL) { + /* tcp_twclose() has already been called */ + INP_WUNLOCK(inp); INP_INFO_RUNLOCK(&V_tcbinfo); - if (reuse) - return tw; - } else { - /* INP_INFO lock is busy, continue later. */ - INP_WLOCK(inp); - if (!in_pcbrele_wlocked(inp)) - INP_WUNLOCK(inp); - break; + continue; } + + tcp_twclose(tw, reuse); + INP_INFO_RUNLOCK(&V_tcbinfo); + if (reuse) + return tw; } return NULL; }