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:   Fri, 21 Apr 2023 11:24:48 -0400
From:   Johannes Weiner <hannes@cmpxchg.org>
To:     Mel Gorman <mgorman@techsingularity.net>
Cc:     linux-mm@kvack.org, Kaiyang Zhao <kaiyang2@cs.cmu.edu>,
        Vlastimil Babka <vbabka@suse.cz>,
        David Rientjes <rientjes@google.com>,
        linux-kernel@vger.kernel.org, kernel-team@fb.com
Subject: Re: [RFC PATCH 25/26] mm: page_alloc: disallow fallbacks when 2M
 defrag is enabled
Message-ID: <20230421152448.GD320347@cmpxchg.org>
References: <20230418191313.268131-1-hannes@cmpxchg.org>
 <20230418191313.268131-26-hannes@cmpxchg.org>
 <20230421145657.fnpjqkuyquy3z24t@techsingularity.net>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20230421145657.fnpjqkuyquy3z24t@techsingularity.net>
Precedence: bulk

On Fri, Apr 21, 2023 at 03:56:57PM +0100, Mel Gorman wrote:
> On Tue, Apr 18, 2023 at 03:13:12PM -0400, Johannes Weiner wrote:
> > Fallbacks are already unlikely due to watermarks being enforced
> > against MIGRATE_FREE blocks. Eliminate them altogether. This allows
> > compaction to look exclusively at movable blocks, reducing the number
> > of pageblocks it needs to scan on an ongoing basis.
> > 
> > Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
> 
> Conceptually this could be fun if a GFP_NOFS allocation cannot migrate
> enough memory to free one pageblock and there are no pageblocks
> available of the correct migratetype. Fallbacks might be unlikely but
> never being able to fallback is a livelock risk, no?

The reserves below the watermarks are maintained in neutral
MIGRATE_FREE blocks. So just like today, critical/limited allocation
contexts are ensured forward progress as long as there are reserves.

An argument could be made that because smaller orders type-claim the
entire neutral block on allocation now, the reserves can deplete
faster than they do today given comparable watermarks. I haven't run
into this issue during testing that would have me raise the reserves
by a factor of NR_MIGRATETYPES. But something to keep an eye out for.