linux-kernel - Re: [PATCH] mm: memcontrol: zap task_struct->memcg_oom_{gfp

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20160311143031.GS27701@dhcp22.suse.cz>
Date:	Fri, 11 Mar 2016 15:30:31 +0100
From:	Michal Hocko <mhocko@...nel.org>
To:	Vladimir Davydov <vdavydov@...tuozzo.com>
Cc:	Andrew Morton <akpm@...ux-foundation.org>,
	Johannes Weiner <hannes@...xchg.org>, linux-mm@...ck.org,
	linux-kernel@...r.kernel.org
Subject: Re: [PATCH] mm: memcontrol: zap
 task_struct->memcg_oom_{gfp_mask,order}

On Fri 11-03-16 16:45:34, Vladimir Davydov wrote:
> On Fri, Mar 11, 2016 at 01:51:05PM +0100, Michal Hocko wrote:
> > On Fri 11-03-16 15:39:00, Vladimir Davydov wrote:
> > > On Fri, Mar 11, 2016 at 12:54:50PM +0100, Michal Hocko wrote:
> > > > On Fri 11-03-16 13:12:47, Vladimir Davydov wrote:
> > > > > These fields are used for dumping info about allocation that triggered
> > > > > OOM. For cgroup this information doesn't make much sense, because OOM
> > > > > killer is always invoked from page fault handler.
> > > > 
> > > > The oom killer is indeed invoked in a different context but why printing
> > > > the original mask and order doesn't make any sense? Doesn't it help to
> > > > see that the reclaim has failed because of GFP_NOFS?
> > > 
> > > I don't see how this can be helpful. How would you use it?
> > 
> > If we start seeing GFP_NOFS triggered OOMs we might be enforced to
> > rethink our current strategy to ignore this charge context for OOM.
> 
> IMO the fact that a lot of OOMs are triggered by GFP_NOFS allocations
> can't be a good enough reason to reconsider OOM strategy.

What I meant was that the global OOM doesn't trigger OOM got !__GFP_FS
while we do in the memcg charge path.

> We need to
> know what kind of allocation fails anyway, and the current OOM dump
> gives us no clue about that.

We do print gfp_mask now so we know what was the charging context.

> Besides, what if OOM was triggered by GFP_NOFS by pure chance, i.e. it
> would have been triggered by GFP_KERNEL if it had happened at that time?

Not really. GFP_KERNEL would allow to invoke some shrinkers which are
GFP_NOFS incopatible.

> IMO it's just confusing.
> 
> >  
> > > Wouldn't it be better to print err msg in try_charge anyway?
> > 
> > Wouldn't that lead to excessive amount of logged messages?
> 
> We could ratelimit these messages. Slab charge failures are already
> reported to dmesg (see ___slab_alloc -> slab_out_of_memory) and nobody's
> complained so far. Are there any non-slab GFP_NOFS allocations charged
> to memcg?

I believe there might be some coming from FS via add_to_page_cache_lru.
Especially when their mapping gfp_mask clears __GFP_FS. I haven't
checked the code deeper but some of those might be called from the page
fault path and trigger memcg OOM. I would have to look closer.
-- 
Michal Hocko
SUSE Labs