Recent Power changes and stack_trace_save_tsk_reliable?

[Joe Lawrence]

Hi ppc-dev list,

We noticed that our kpatch integration tests started failing on ppc64le
when targeting the upstream v6.4 kernel, and then confirmed that the
in-tree livepatching kselftests similarly fail, too.  From the kselftest
results, it appears that livepatch transitions are no longer completing.

Looking at the commit logs for v6.4, there looks to be some churn in the
powerpc stack layout code -- I am suspicious that "reliable" stack
unwinding may be left untested/broken after those changes.  AFAICT, the
livepatching subsystem is the only user of this interface in
kernel/livepatch/transition.c :: klp_check_stack()'s call to
stack_trace_save_tsk_reliable().  As such, the livepatching kselftests
are probably the only way to test reliable unwinding.

Unfortunately, git bisect isn't cooperating (we keep falling into a long
span of non-bootable commits, despite efforts to `git bisect skip` over
them), so we don't have an offending commit or patchset to point to.

A few other details:

- Test machine is a Power 9 9009-42A (IBM Power System S924)
- Reproducable with v6.4, v6.5
- Minimal repro:
-- Build with CONFIG_TEST_LIVEPATCH=m
-- Run tools/testing/selftests/livepatch/test-livepatch.sh

If this has already been report or fixed, please send any pointers to
threads / commits.  If not, I can provide any other info to help
reproduce.

Thanks,

--
Joe