Re: NMI reason 2d when running perf

[Ravi Bangoria]

On 17-Mar-23 7:55 PM, Jens Axboe wrote:
> Hi,
> 
> When running perf on my Dell R7525 on a running process, I get a ton of:
> 
> [  504.234782] Dazed and confused, but trying to continue
> [  504.267843] Uhhuh. NMI received for unknown reason 2d on CPU 48.
> [  504.267846] Dazed and confused, but trying to continue
> [  504.335975] Uhhuh. NMI received for unknown reason 2d on CPU 48.
> [  504.335977] Dazed and confused, but trying to continue
> [  504.368031] Uhhuh. NMI received for unknown reason 2d on CPU 48.
> [  504.368033] Dazed and confused, but trying to continue
> [  504.371037] Uhhuh. NMI received for unknown reason 2d on CPU 48.
> [  504.371038] Dazed and confused, but trying to continue
> [  504.439165] Uhhuh. NMI received for unknown reason 2d on CPU 48.
> [  504.439167] Dazed and confused, but trying to continue
> 
> spew in dmesg. The box has 2x7763 CPUS.

Are you using IBS directly or indirectly (perf record, if run as root
without explicit -e event, internally invokes IBS)? fwiw, Zen2 had
similar hw issue where IBS can raise NMI without sample valid bit set
while exiting from CC6 state[1]. But as per my knowledge, it's already
fixed in Zen3.

[1]: https://lore.kernel.org/all/YFDSSxftYw9tCGC6@krava/

> This seems to be a recent
> regression, been using this box for a while and haven't seen this
> before. The test being traced is pinned to CPU 48. The box is currently
> running:
> 
> commit 6015b1aca1a233379625385feb01dd014aca60b5 (origin/master, origin/HEAD)
> Author: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
> Date:   Tue Mar 14 19:32:38 2023 -0700
> 
>     sched_getaffinity: don't assume 'cpumask_size()' is fully initialized
> 
> with the pending block/io_uring branches merged in for testing.
>