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, 7 Jun 2023 08:02:33 +1000
From:   Dave Chinner <david@fromorbit.com>
To:     Kirill Tkhai <tkhai@ya.ru>
Cc:     akpm@linux-foundation.org, roman.gushchin@linux.dev,
        vbabka@suse.cz, viro@zeniv.linux.org.uk, brauner@kernel.org,
        djwong@kernel.org, hughd@google.com, paulmck@kernel.org,
        muchun.song@linux.dev, linux-mm@kvack.org,
        linux-fsdevel@vger.kernel.org, linux-xfs@vger.kernel.org,
        linux-kernel@vger.kernel.org, zhengqi.arch@bytedance.com
Subject: Re: [PATCH v2 0/3] mm: Make unregistration of super_block shrinker
 more faster
Message-ID: <ZH+s+XOI2HlLTDzs@dread.disaster.area>
References: <168599103578.70911.9402374667983518835.stgit@pro.pro>
 <ZH5ig590WleaH1Ed@dread.disaster.area>
 <ef1b0ecd-5a03-4256-2a7a-3e22b755aa53@ya.ru>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <ef1b0ecd-5a03-4256-2a7a-3e22b755aa53@ya.ru>
Precedence: bulk

On Wed, Jun 07, 2023 at 12:06:03AM +0300, Kirill Tkhai wrote:
> On 06.06.2023 01:32, Dave Chinner wrote:
> > On Mon, Jun 05, 2023 at 10:02:46PM +0300, Kirill Tkhai wrote:
> >> This patch set introduces a new scheme of shrinker unregistration. It allows to split
> >> the unregistration in two parts: fast and slow. This allows to hide slow part from
> >> a user, so user-visible unregistration becomes fast.
> >>
> >> This fixes the -88.8% regression of stress-ng.ramfs.ops_per_sec noticed
> >> by kernel test robot:
> >>
> >> https://lore.kernel.org/lkml/202305230837.db2c233f-yujie.liu@intel.com/
> >>
> >> ---
> >>
> >> Kirill Tkhai (2):
> >>       mm: Split unregister_shrinker() in fast and slow part
> >>       fs: Use delayed shrinker unregistration
> > 
> > Did you test any filesystem other than ramfs?
> > 
> > Filesystems more complex than ramfs have internal shrinkers, and so
> > they will still be running the slow synchronize_srcu() - potentially
> > multiple times! - in every unmount. Both XFS and ext4 have 3
> > internal shrinker instances per mount, so they will still call
> > synchronize_srcu() at least 3 times per unmount after this change.
> > 
> > What about any other subsystem that runs a shrinker - do they have
> > context depedent shrinker instances that get frequently created and
> > destroyed? They'll need the same treatment.
> 
> Of course, all of shrinkers should be fixed. This patch set just aims to describe
> the idea more wider, because I'm not sure most people read replys to kernel robot reports.
> 
> This is my suggestion of way to go. Probably, Qi is right person to ask whether
> we're going to extend this and to maintain f95bdb700bc6 in tree.
> 
> There is not much time. Unfortunately, kernel test robot reported this significantly late.

And that's why it should be reverted rather than trying to rush to
try to fix it.

I'm kind of tired of finding out about mm reclaim regressions only
when I see patches making naive and/or broken changes to subsystem
shrinker implementations without any real clue about what they are
doing.  If people/subsystems who maintain shrinker implementations
were cc'd on the changes to the shrinker implementation, this would
have all been resolved before merging occurred....

Lockless shrinker lists need a heap of supporting changes to be done
first so that they aren't reliant on synchronise_srcu() *at all*. If
the code was properly designed in the first place (i.e. dynamic
shrinker structures freed via call_rcu()), we wouldn't be in rushing
to fix weird regressions right now. 

Can we please revert this and start again with a properly throught
out and reveiwed design?

-Dave.
-- 
Dave Chinner
david@fromorbit.com