Re: [PATCH] mm,page_alloc: Split stall warning and failure warning.

[Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>]

Michal Hocko wrote:
> On Mon 10-04-17 15:03:08, Andrew Morton wrote:
> > On Mon, 10 Apr 2017 20:58:13 +0900 Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> wrote:
> > 
> > > Patch "mm: page_alloc: __GFP_NOWARN shouldn't suppress stall warnings"
> > > changed to drop __GFP_NOWARN when calling warn_alloc() for stall warning.
> > > Although I suggested for two times to drop __GFP_NOWARN when warn_alloc()
> > > for stall warning was proposed, Michal Hocko does not want to print stall
> > > warnings when __GFP_NOWARN is given [1][2].
> > > 
> > >  "I am not going to allow defining a weird __GFP_NOWARN semantic which
> > >   allows warnings but only sometimes. At least not without having a proper
> > >   way to silence both failures _and_ stalls or just stalls. I do not
> > >   really thing this is worth the additional gfp flag."
> > 
> > I interpret __GFP_NOWARN to mean "don't warn about this allocation
> > attempt failing", not "don't warn about anything at all".  It's a very
> > minor issue but yes, methinks that stall warning should still come out.
> 
> This is what the patch from Johannes already does and you have it in the
> mmotm tree.
> 
> > Unless it's known to cause a problem for the stall warning to come out
> > for __GFP_NOWARN attempts?  If so then perhaps a
> > __GFP_NOWARN_ABOUT_STALLS is needed?
> 
> And this is one of the reason why I didn't like it. But whatever it
> doesn't make much sense to spend too much time discussing this again.
> This patch doesn't really fix anything important IMHO and it just
> generates more churn.

This patch does not fix anything important for Michal Hocko, but
this patch does find something important (e.g. GFP_NOFS | __GFP_NOWARN
allocations) for administrators and troubleshooting staffs at support
centers. As a troubleshooting staff, giving administrators some clue to
start troubleshooting is critically important.

Speak from my experience, hardcoded 10 seconds is really useless.
Some cluster system has only 10 seconds timeout for failover. Failing
to report allocations stalls longer than a few seconds can make this
warn_alloc() pointless. On the other hand, some administrators do not
want to receive this warn_alloc(). If we had tunable interface like
/proc/sys/kernel/memalloc_task_warning_secs , we can handle both cases
(assuming that stalling allocations can reach this warn_alloc() within
a few seconds; if this assumption does not hold, only allocation watchdog
can handle it).

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>