Received-SPF: pass (google.com: domain of linux-kernel+bounces-56168-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:45d1:ec00::1 as permitted sender) client-ip=2604:1380:45d1:ec00::1;
Date: Wed, 7 Feb 2024 18:08:59 +0900
From: Sergey Senozhatsky <senozhatsky@chromium.org>
To: Minchan Kim <minchan@kernel.org>
Cc: Andrew Morton <akpm@linux-foundation.org>, linux-block@vger.kernel.org,
	linux-kernel@vger.kernel.org,
	Sergey Senozhatsky <senozhatsky@chromium.org>
Subject: Re: [RFC][PATCH 0/2] zram: decouple comp stream and comp buffer
Message-ID: <20240207090859.GE489524@google.com>
References: <20240207065751.1908939-1-senozhatsky@chromium.org>
 <20240207084720.GD489524@google.com>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20240207084720.GD489524@google.com>

On (24/02/07 17:47), Sergey Senozhatsky wrote:
> On (24/02/07 15:57), Sergey Senozhatsky wrote:
> > 	RFC
> > 
> > 	We keep compression work memory buffer per-comp stream (which
> > is per-CPU), but we don't need that many buffers, because on each
> > CPU only one compression backend can access work memory at any given
> > time. Hence the patch series moves compression work memory to a
> > dedicated per-CPU area, reducing the amount of memory we allocate
> > for those buffers.
> > 
> > For instance, assume a 12 CPUs system, 2 compression streams
> > per-CPU (a default and one for deferred recompression). Before
> > we'd allocate 12 * 2 * 2 pages, after we'll allocate 12 * 2 pages.
> > 
> > NOTE:
> > The series stops short of moving comp buffers to a global per-CPU
> > area, which all zram devices can share. Compression backends use
> > CPUs exclusively (disable migration and CPU hotplug), so in theory
> > comp work memory can be in global per-CPU data. This can reduce
> > memory usage on systems that init numerous zram devices.
> > E.g. instead of num-zram-devices * num-cpus buffers we'll allocate
> > only num-cpus buffers.
> 
> And this is the patch that moves comp buffers to global per-CPU
> area, so that *all* zram devices and *all* comp backends share
> them. It also moves local_lock (which disables migration/preemption
> and CPU-hotplug once taken) to global comp mem per-CPU buffer.

That's a squashed patch. I also have an incremental one, atop of 0002
in the series. So first move comp buffers to device per-CPU, then to
global per-CPU.