Received-SPF: pass (google.com: domain of linux-kernel+bounces-132349-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.199.223 as permitted sender) client-ip=147.75.199.223;
Date: Thu, 4 Apr 2024 22:54:07 -0400
From: Johannes Weiner <hannes@cmpxchg.org>
To: Yosry Ahmed <yosryahmed@google.com>
Cc: Andrew Morton <akpm@linux-foundation.org>,
	Nhat Pham <nphamcs@gmail.com>,
	Chengming Zhou <chengming.zhou@linux.dev>, linux-mm@kvack.org,
	linux-kernel@vger.kernel.org
Subject: Re: [PATCH v1 5/5] mm: zswap: do not check the global limit for
 same-filled pages
Message-ID: <20240405025407.GF641486@cmpxchg.org>
References: <20240405013547.1859126-1-yosryahmed@google.com>
 <20240405013547.1859126-6-yosryahmed@google.com>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20240405013547.1859126-6-yosryahmed@google.com>

On Fri, Apr 05, 2024 at 01:35:47AM +0000, Yosry Ahmed wrote:
> When storing same-filled pages, there is no point of checking the global
> zswap limit as storing them does not contribute toward the limit Move
> the limit checking after same-filled pages are handled.
> 
> This avoids having same-filled pages skip zswap and go to disk swap if
> the limit is hit. It also avoids queueing the shrink worker, which may
> end up being unnecessary if the zswap usage goes down on its own before
> another store is attempted.
> 
> Ignoring the memcg limits as well for same-filled pages is more
> controversial. Those limits are more a matter of per-workload policy.
> Some workloads disable zswap completely by setting memory.zswap.max = 0,
> and those workloads could start observing some zswap activity even after
> disabling zswap. Although harmless, this could cause confusion to
> userspace. Remain conservative and keep respecting those limits.
> 
> Signed-off-by: Yosry Ahmed <yosryahmed@google.com>

I'm not sure this buys us enough in practice to justify special-casing
those entries even further. Especially with the quirk of checking
cgroup limits but not the global ones; that would definitely need a
code comment similar to what you have in the changelog; and once you
add that, the real estate this special treatment takes up really
doesn't seem reasonable anymore.

In most cases we'd expect a mix of pages to hit swap. Waking up the
shrinker on a zero-filled entry is not strictly necessary of course,
but the zswap limit has been reached and the system is swapping - a
wakeup seems imminent anyway.