On Tue, Jun 22, 2021 at 07:24:56PM +0200, Oleksandr Natalenko wrote:
Hello.
On úterý 22. června 2021 18:47:59 CEST Greg KH wrote:
On Tue, Jun 22, 2021 at 06:30:46PM +0200, Oleksandr Natalenko wrote:
I'd like to nominate d583d360a6 ("psi: Fix psi state corruption when schedule() races with cgroup move") for 5.12 stable tree.
Recently, I've hit this:
kernel: psi: inconsistent task state! task=2667:clementine cpu=21 psi_flags=0 clear=1 set=0and after that PSI IO went crazy high. That seems to match the symptoms described in the commit message.
But this says it fixes 4117cebf1a9f ("psi: Optimize task switch inside shared cgroups") which did not show up until 5.13-rc1, so how are you hitting this issue?
I'm not positive 4117cebf1a9f was a root cause of the race. To me it looks like 4117cebf1a9f just made an older issue more likely to be hit.
Peter, Johannes, am I correct saying that it is still possible to hit a corruption described in d583d360a6 on 5.12?
I'm not aware of a previous issue, but it's possible you discovered one that was incidentally fixed by this change.
That said, there haven't been many changes in this area prior to 5.12, and I stared at the old code quite a bit to see if there are other possible scenarios, so this gives me pause.
Did you try this patch on 5.12.y and see that it solved your problem?
Yes, I've built the kernel with this patch, and so far it runs fine. It can take a while until the condition is hit though since it seems to be very unlikely on 5.12.
Is your task moving / being moved between cgroups while it's doing work?
How long does it usually take to trigger it?
Would it be possible to share a simpler reproducer, or is this part of a more complex application?