Re: [PATCH 3/4] mm: Centralize & improve oom reporting in show_mem.c

From: Kent Overstreet <kent.overstreet@gmail.com>
To: Michal Hocko <mhocko@suse.com>
Cc: linux-kernel@vger.kernel.org, linux-mm@kvack.org,
	linux-fsdevel@vger.kernel.org, roman.gushchin@linux.dev,
	hannes@cmpxchg.org
Subject: Re: [PATCH 3/4] mm: Centralize & improve oom reporting in show_mem.c
Date: Fri, 22 Apr 2022 06:58:40 -0400	[thread overview]
Message-ID: <20220422105840.wsrlxt3emw4vagcm@moria.home.lan> (raw)
In-Reply-To: <YmKInWEihG+7mkU6@dhcp22.suse.cz>

On Fri, Apr 22, 2022 at 12:51:09PM +0200, Michal Hocko wrote:
> On Fri 22-04-22 05:44:13, Kent Overstreet wrote:
> > On Fri, Apr 22, 2022 at 11:27:05AM +0200, Michal Hocko wrote:
> > > We already do that in some form. We dump unreclaimable slabs if they
> > > consume more memory than user pages on LRUs. We also dump all slab
> > > caches with some objects. Why is this approach not good? Should we tweak
> > > the condition to dump or should we limit the dump? These are reasonable 
> > > questions to ask. Your patch has dropped those without explaining any
> > > of the motivation.
> > > 
> > > I am perfectly OK to modify should_dump_unreclaim_slab to dump even if
> > > the slab memory consumption is lower. Also dumping small caches with
> > > handful of objects can be excessive.
> > > 
> > > Wrt to shrinkers I really do not know what kind of shrinkers data would
> > > be useful to dump and when. Therefore I am asking about examples.
> > 
> > Look, I've given you the sample
> 
> That sample is of no use as it doesn't really show how the additional
> information is useful to analyze the allocation failure. I thought we
> have agreed on that. You still haven't given any example where the
> information is useful. So I do not really see any reason to change the
> existing output.
> 
> > output you asked for and explained repeatedly my
> > rationale and you haven't directly responded;
> 
> Your rationale is that we need more data and I do agree but it is not
> clear which data and under which conditions.

You're completely mischaractarizing and making this _way_ more complicated than
it has to be, but I'll repeat:

- For the slab changes, top 10 slabs in sorted order, with human readable units
  are _vastly_ easier on human eyes than pages of slab output, in the previous
  format

- Shrinkers weren't reported on before at all, and as shrinkers are part of
  memory reclaim they're pretty integral to OOM debugging.

> > if you have a reason you're
> > against the patches please say so, but please give your reasoning.
> 
> I have expressed that already, I believe, but let me repeat. I do not
> like altering the oom report without a justification on how this new
> output is useful. You have failed to explained that so far.

Uh huh.

Sounds like someone has some scripts he doesn't want to have to update.