Add support for re-resolving arm64 ifuncs when all CPUs are online

[andrew]

Summary

As arm64 uses rela relocations the addend is stored in in the relocaation.
As nothing writes to these we can re-run ifunc resolvers late in the
boot process to allow them to take into account the new CPUs that have
come online.

[andrew]

I'm working on a patch to generate a common kernel view of the ID registers that will be available when link_elf_late_ireloc is called. It will also store the ID registers of CPU 0 when the resolvers are initially called so the resolvers will be identical early in the boot and after all CPUs have started.

[kib]

sys/arm64/arm64/mp_machdep.c

171 ↗

(On Diff #73648)

Is it fine to re-reloc only on SMP ?

sys/kern/link_elf.c

1929

This should be not arm64 specific, rather it should depend on RELA vs REL.

[andrew]

I have created D25463 to create the common kernel view that will be used by ifunc resolvers.

[andrew]

• Call link_elf_late_ireloc from a SYSINIT function
• Pass a flag into elf_reloc_internal rather than more arguments

[andrew]

sys/kern/link_elf.c

1929

It needs an implementation of elf_reloc_late in the per-arch code. I've only implemented it on arm64 as I'm unsure what other architectures use RELA relocations.

[kib]

sys/kern/link_elf.c

1929

I think for now it is enough to have aarch64 and amd64.

[andrew]

Add untested amd64 support

[kib]

It is fine, except one detail that I do not like. You do the late pass after the APs were released.

I mean, despite the fact that relocations are idempotent, there are some dependencies between the set of functions returned by resolver. For instance, on amd64 pmap_activate() and pmap_invalidate_*() must be aligned.

I believe that it is needed to park other CPUs into safe loops that cannot be preempted while the resolution is performed.

[andrew]

Aren't the CPUs parked until release_aps is called? This will be in SI_SUB_SMP which is later than SI_SUB_CPU where the ifunc resolvers are called.

[kib]

In D25455#564949, @andrew wrote:

Aren't the CPUs parked until release_aps is called? This will be in SI_SUB_SMP which is later than SI_SUB_CPU where the ifunc resolvers are called.

I see, I mixed _CPU/_SMP.