Received-SPF: pass (google.com: domain of linux-kernel+bounces-163996-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.48.161 as permitted sender) client-ip=147.75.48.161;
Date: Tue, 30 Apr 2024 09:25:55 -0400
From: Lucas Karpinski <lkarpins@redhat.com>
To: Al Viro <viro@zeniv.linux.org.uk>
Cc: brauner@kernel.org, jack@suse.cz, linux-fsdevel@vger.kernel.org, 
	linux-kernel@vger.kernel.org, alexl@redhat.com, echanude@redhat.com, ikent@redhat.com, 
	ahalaney@redhat.com
Subject: Re: [RFC v2 1/1] fs/namespace: defer RCU sync for MNT_DETACH umount
Message-ID: <6rp73lih7g2b7i5rhsztwc66quq6fi3mesel52uavvt7uhfzlf@6rytjc7gb2tj>
References: <20240426195429.28547-1-lkarpins@redhat.com>
 <20240426195429.28547-2-lkarpins@redhat.com>
 <20240426200941.GP2118490@ZenIV>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20240426200941.GP2118490@ZenIV>
User-Agent: NeoMutt/20240201

On Fri, Apr 26, 2024 at 09:09:41PM +0100, Al Viro wrote:
> > +		call_rcu(&drelease->rcu, delayed_mount_release);
> 
> ... which is a bad idea, since call_rcu() callbacks are run
> from interrupt context.  Which makes blocking in them a problem.
>

Thanks for the quick review.

Documentation/RCU/checklist.rst suggests switching to queue_rcu_work()
function in scenarios where the callback function can block. This seems
like it would fix the issue you found, while still providing similar
performance improvements.

workqueue:
perf stat -r 10 --null --pre 'mount -t tmpfs tmpfs mnt' -- umount -l mnt
        0.003066 +- 0.000307 seconds time elapsed  ( +- 10.02% )

callrcu:
perf stat -r 10 --null --pre 'mount -t tmpfs tmpfs mnt' -- umount -l mnt
        0.0030812 +- 0.0000524 seconds time elapsed  ( +-  1.70% )

Regards,
Lucas