Re: [PATCH v7] mm: Add memory allocation watchdog kernel thread.
From: Andrew Morton
Date: Thu Mar 09 2017 - 17:38:04 EST
On Thu, 9 Mar 2017 19:46:14 +0900 Tetsuo Handa <penguin-kernel@xxxxxxxxxxxxxxxxxxx> wrote:
> Tetsuo Handa wrote:
> > This patch adds a watchdog which periodically reports number of memory
> > allocating tasks, dying tasks and OOM victim tasks when some task is
> > spending too long time inside __alloc_pages_slowpath(). This patch also
> > serves as a hook for obtaining additional information using SystemTap
> > (e.g. examine other variables using printk(), capture a crash dump by
> > calling panic()) by triggering a callback only when a stall is detected.
> > Ability to take administrator-controlled actions based on some threshold
> > is a big advantage gained by introducing a state tracking.
> > Commit 63f53dea0c9866e9 ("mm: warn about allocations which stall for
> > too long") was a great step for reducing possibility of silent hang up
> > problem caused by memory allocation stalls . However, there are
> > reports of long stalls (e.g.  is over 30 minutes!) and lockups (e.g.
> >  is an "unable to invoke the OOM killer due to !__GFP_FS allocation"
> > lockup problem) where this patch is more useful than that commit, for
> > this patch can report possibly related tasks even if allocating tasks
> > are unexpectedly blocked for so long. Regarding premature OOM killer
> > invocation, tracepoints which can accumulate samples in short interval
> > would be useful. But regarding too late to report allocation stalls,
> > this patch which can capture all tasks (for reporting overall situation)
> > in longer interval and act as a trigger (for accumulating short interval
> > samples) would be useful.
> Andrew, do you have any questions on this patch?
> I really need this patch for finding bugs which MM people overlook.
(top-posting repaired - please don't do that)
Undecided. I can see the need but it is indeed quite a large lump of
code. Perhaps some additional examples of how this new code was used
to understand and improve real-world kernel problems would be persuasive.