Received-SPF: pass (google.com: domain of linux-nfs-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Thu, 13 May 2021 14:34:35 +0100
From:   Matthew Wilcox <willy@infradead.org>
To:     Steve French <smfrench@gmail.com>
Cc:     linux-nfs <linux-nfs@vger.kernel.org>,
        CIFS <linux-cifs@vger.kernel.org>,
        v9fs-developer@lists.sourceforge.net,
        linux-fsdevel <linux-fsdevel@vger.kernel.org>
Subject: Re: Removing readpages aop
Message-ID: <YJ0q6/Oe5yJ+H+Tn@casper.infradead.org>
References: <YJvwVq3Gl35RQrIe@casper.infradead.org>
 <CAH2r5msOQsdeknBdTsfMXYzrb5=NuKEBPc4WD1CkYp10t19Guw@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <CAH2r5msOQsdeknBdTsfMXYzrb5=NuKEBPc4WD1CkYp10t19Guw@mail.gmail.com>
Precedence: bulk

On Wed, May 12, 2021 at 02:28:41PM -0500, Steve French wrote:
> I don't have any objections as long as:
> - we see at least mild performance benefit (or at least we are
> confident that no performance loss)

Nobody's complained of a performance loss in the other ~30 filesystems
which have already been converted (some almost a year ago).  And CIFS
has one of the more convoluted readpages implementation, so I'd expect
a higher likelihood of a performance gain from CIFS.

> - it passes regression tests (the usual xfstest bucket)
> - it doesn't complicate the code too much (sounds like it actually
> might simplify it, but needs a little more work)
> - make sure that the usual tuning parms still work (e.g. "rsize" and
> "rasize" mount options) or we can figure out a sane way to autotune
> readhead so those wouldn't be needed for any workload

One of the enhancements added as part of the recent netfs merge
was readahead_expand().  Take a look at it and see if it works for you.

> But currently since we get the most benefit from multichannel (as that
> allows even better parallelization of i/o) ... I have been focused on
> various multichannel issues (low credit situations, reconnect, fall
> back to different channels when weird errors, adjusting channels
> dynamically when server adds or removes adapters on the fly) for the
> short term

Understood.  Only so many hours in the day.

I think
https://lore.kernel.org/linux-fsdevel/1794123.1605713481@warthog.procyon.org.uk/
is the most recent version, but as Dave notes, it needs attention from
somebody who knows the CIFS code better.