Re: [PATCH 2/3] Linux: Use rseq in sched_getcpu if available (v9)

----- On Jul 6, 2020, at 9:59 AM, Florian Weimer fweimer@redhat.com wrote:

* Mathieu Desnoyers:

quoted

When available, use the cpu_id field from __rseq_abi on Linux to
implement sched_getcpu().  Fall-back on the vgetcpu vDSO if
unavailable.

I've pushed this to glibc master, but unfortunately it looks like this
exposes a kernel bug related to affinity mask changes.

After building and testing glibc, this

 for x in {1..2000} ; do posix/tst-affinity-static  & done

produces some “error:” lines for me:

error: Unexpected CPU 2, expected 0
error: Unexpected CPU 2, expected 0
error: Unexpected CPU 2, expected 0
error: Unexpected CPU 2, expected 0
error: Unexpected CPU 138, expected 0
error: Unexpected CPU 138, expected 0
error: Unexpected CPU 138, expected 0
error: Unexpected CPU 138, expected 0

“expected 0” is a result of how the test has been written, it bails out
on the first failure, which happens with CPU ID 0.

Smaller systems can use a smaller count than 2000 to reproduce this.  It
also happens sporadically when running the glibc test suite itself
(which is why it took further testing to reveal this issue).

I can reproduce this with the Debian 4.19.118-2+deb10u1 kernel, the
Fedora 5.6.19-300.fc32 kernel, and the Red Hat Enterprise Linux kernel
4.18.0-193.el8 (all x86_64).

As to the cause, I'd guess that the exit path in the sched_setaffinity
system call fails to update the rseq area, so that userspace can observe
the outdated CPU ID there.

Hi Florian,

We have a similar test in Linux, see tools/testing/selftests/rseq/basic_test.c.
That test does not trigger this issue, even when executed repeatedly.

I'll investigate further what is happening within the glibc test.

Thanks,

Mathieu


-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com