Re: [PATCH v2] mm/page_isolation: fix a deadlock with printk()
From: Michal Hocko
Date: Tue Oct 08 2019 - 09:37:30 EST
On Tue 08-10-19 09:06:29, Qian Cai wrote:
> On Tue, 2019-10-08 at 14:39 +0200, Michal Hocko wrote:
> > On Tue 08-10-19 08:00:43, Qian Cai wrote:
> > >
> > >
> > > > On Oct 8, 2019, at 6:39 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > > >
> > > > Have you actually triggered any real deadlock? With a zone->lock in
> > > > place it would be pretty clear with hard lockups detected.
> > >
> > > Yes, I did trigger here and there, and those lockdep splats are
> > > especially useful to figure out why.
> > Can you provide a lockdep splat from an actual deadlock please? I am
> > sorry but your responses tend to be really cryptic and I never know when
> > you are talking about actual deadlocks and lockdep splats. I have asked
> > about the former several times never receiving a specific answer.
> It is very time-consuming to confirm a lockdep splat is 100% matching a deadlock
> giving that it is not able to reproduce on will yet, so when I did encounter a
> memory offline deadlock where "echo offline > memory/state" just hang, but there
> is no hard lockup probably because the hard lockup detector did not work
> properly for some reasons or it keep trying to acquire a spin lock that only
> keep the CPU 100%.
If there is a real deadlock due to zone->lock then you would certainly
get a hard lockup splat. So I strongly suspect that you are seeing a
completely different problem. Most likely some pages cannot be migrated
and the offlining code will retry for ever. You can terminate that from
the userspace by a fatal signal of course.