psci: Teach the arm SMCCC code about KMSAN

[andrew]

Summary

We write to a data structure in assembly. Because of this we need to
annotate the function manually for KMSAN to know the contents have been
written otherwise it picks up later reads as loading uninitialised data.

Sponsored by: Arm Ltd

[markj]

This feels over-engineered: why not just make always-on C wrappers for the assembly routines and handle marking there? I don't see a need for a separate file and the extra indirection.

sys/dev/psci/smccc_kmsan.c

28

The purpose of these guards is to prevent recursion in the bits of the KMSAN runtime which actually need to use intercepted symbols, but that's not a danger for the SMCCC interfaces.

[andrew]

• Move kmsan calls to asm
• Add kmsan_check and kmsan_mark to arm_smccc_1_2_*

[markj]

Why can't this be done in C? e.g., have asm _arm_smccc_1_2_hvc() which is called by the C function arm_smccc_1_2_hvc().

[andrew]

It seems like an unneeded level of indirection in the non-KMSAN case

sys/dev/psci/smccc_kmsan.c

28

I don't understand, these use the intercepted symbols, e.g. kmsan_arm_smccc_hvc calls arm_smccc_hvc.

[markj]

In D47317#1079638, @andrew wrote:

It seems like an unneeded level of indirection in the non-KMSAN case

These aren't performance-critical in any way, so I'm not sure why it matters.

I don't object to handling everything in asm, but it really seems like it'd be simpler and cleaner to handle KMSAN marking in C.

sys/dev/psci/smccc_kmsan.c

28

Sorry, my comment wasn't right. I really just meant that this machinery isn't needed for routines which are only called from one file.