Re: [PATCH] mm: memcontrol: zap task_struct->memcg_oom_{gfp_mask,order}

From: Vladimir Davydov
Date: Fri Mar 11 2016 - 10:02:42 EST


On Fri, Mar 11, 2016 at 03:30:31PM +0100, Michal Hocko wrote:
> 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.

OK, missed your point, sorry.

>
> > 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.

Can't a GFP_NOFS allocation happen when there is no shrinkable objects
to drop so that there's no real difference between GFP_KERNEL and
GFP_NOFS?

>
> > 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.

If you think this warning is really a must have, and you don't like to
warn about every charge failure, may be we could just print info about
allocation that triggered OOM right in mem_cgroup_oom, like the code
below does? I think it would be more-or-less equivalent to what we have
now except it wouldn't require storing gfp_mask on task_struct.

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index a217b1374c32..d8e130d14f5d 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -1604,6 +1604,8 @@ static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
*/
css_get(&memcg->css);
current->memcg_in_oom = memcg;
+
+ pr_warn("Process ... triggered OOM in memcg ... gfp ...\n");
}

/**