Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Wed, 24 May 2023 11:58:11 -0400
From:   Johannes Weiner <hannes@cmpxchg.org>
To:     Vlastimil Babka <vbabka@suse.cz>
Cc:     Andrew Morton <akpm@linux-foundation.org>,
        Mel Gorman <mgorman@techsingularity.net>,
        Michal Hocko <mhocko@suse.com>, linux-mm@kvack.org,
        linux-kernel@vger.kernel.org, kernel-team@fb.com
Subject: Re: [PATCH] mm: compaction: avoid GFP_NOFS ABBA deadlock
Message-ID: <20230524155811.GA14306@cmpxchg.org>
References: <20230519111359.40475-1-hannes@cmpxchg.org>
 <8fd1a56d-5a22-4bde-59a5-169a4696219e@suse.cz>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <8fd1a56d-5a22-4bde-59a5-169a4696219e@suse.cz>
Precedence: bulk

On Wed, May 24, 2023 at 11:21:43AM +0200, Vlastimil Babka wrote:
> On 5/19/23 13:13, Johannes Weiner wrote:
> > During stress testing with higher-order allocations, a deadlock
> > scenario was observed in compaction: One GFP_NOFS allocation was
> > sleeping on mm/compaction.c::too_many_isolated(), while all CPUs in
> > the system were busy with compactors spinning on buffer locks held by
> > the sleeping GFP_NOFS allocation.
> > 
> > Reclaim is susceptible to this same deadlock; we fixed it by granting
> > GFP_NOFS allocations additional LRU isolation headroom, to ensure it
> > makes forward progress while holding fs locks that other reclaimers
> > might acquire. Do the same here.
> > 
> > This code has been like this since compaction was initially merged,
> > and I only managed to trigger this with out-of-tree patches that
> > dramatically increase the contexts that do GFP_NOFS compaction. While
> > the issue is real, it seems theoretical in nature given existing
> > allocation sites. Worth fixing now, but no Fixes tag or stable CC.
> 
> > Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
> 
> So IIUC the change is done by not giving GFP_NOFS extra headroom, but
> instead restricting the headroom of __GFP_FS allocations. But the original
> one was probably too generous anyway so it should be fine?

Yes, the original limit is generally half the LRU, which is quite high.

The new limit is 1/16th of the LRU for regular compactors and half for
GFP_NOFS ones. Note that I didn't make these up; they're stolen from
too_many_isolated() in vmscan.c. I figured those are proven values and
no sense in deviating from them until we have a reason to do so.

> Acked-by: Vlastimil Babka <vbabka@suse.cz>

Thanks!