Can you send ma a git patch with this change? e.g. the output of git format-patch

xregs_sig doesn't follow the ABI so it is valid to trash these registers.

I tested a patch to move vfp_save_state & haven't hit the KASSERT, even when loading the system down to try cause more context switching.

After the first call to vfp_save_state from get_arm64_sve with buf == NULL the PCB_FP_SVEVALID flag will be set correctly based on if pcb_svesaved contains valid data or not (it may be clear if called while in a syscall). If we then context switch then vfp_save_state_switch will exit early as VFP is disabled so PCB_FP_SVEVALID isn't changed & later calls to vfp_save_state will also not change the state.

Can this be more specific to GCE? There are systems where the kernel may enter in EL1, but not be under a hypervisor, e.g. the. Windows Dev Kit 2023. Setting vm_guest = VM_GUEST_VM also changes things like hz to 100.

In D54559#1246536, @sarah.walker2_arm.com wrote:

In D54559#1246501, @andrew wrote:

There should be a followup change to sys/arm64/include/ifunc.h to update DEFINE_UIFUNC. We can also use this to simplify the ifunc handlers in sys/crypto/sha2/sha512c.c and sys/crypto/sha2/sha256c.c

Can you give more detail on what you want here please?

There should be a followup change to sys/arm64/include/ifunc.h to update DEFINE_UIFUNC. We can also use this to simplify the ifunc handlers in sys/crypto/sha2/sha512c.c and sys/crypto/sha2/sha256c.c

Just a minor style suggestion

Would it be enough to move the td == curthread .. check & call to vfp_save_state before checking for PCB_FP_SVEVALID in line 942?