Date:   Mon, 10 Jan 2022 19:45:15 +0100
From:   Peter Zijlstra <peterz@infradead.org>
To:     Waiman Long <longman@redhat.com>
Cc:     Christoph Hellwig <hch@infradead.org>,
        Jaegeuk Kim <jaegeuk@kernel.org>, linux-kernel@vger.kernel.org,
        linux-f2fs-devel@lists.sourceforge.net,
        Tim Murray <timmurray@google.com>,
        Ingo Molnar <mingo@redhat.com>, Will Deacon <will@kernel.org>,
        Boqun Feng <boqun.feng@gmail.com>
Subject: Re: [PATCH] f2fs: move f2fs to use reader-unfair rwsems
Message-ID: <Ydx+u9YpzS8AZHrl@hirez.programming.kicks-ass.net>
References: <20220108164617.3130175-1-jaegeuk@kernel.org>
 <YdvoxkAAquI17UbX@infradead.org>
 <a23a3226-95d9-9835-c1c7-2d13f4a1ee16@redhat.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <a23a3226-95d9-9835-c1c7-2d13f4a1ee16@redhat.com>
Precedence: bulk

On Mon, Jan 10, 2022 at 11:18:27AM -0500, Waiman Long wrote:
> 
> On 1/10/22 03:05, Christoph Hellwig wrote:
> > Adding the locking primitive maintainers to this patch adding open coded
> > locking primitives..
> > 
> > On Sat, Jan 08, 2022 at 08:46:17AM -0800, Jaegeuk Kim wrote:
> > > From: Tim Murray <timmurray@google.com>
> > > 
> > > f2fs rw_semaphores work better if writers can starve readers,
> > > especially for the checkpoint thread, because writers are strictly
> > > more important than reader threads. This prevents significant priority
> > > inversion between low-priority readers that blocked while trying to
> > > acquire the read lock and a second acquisition of the write lock that
> > > might be blocking high priority work.
> > > 
> > > Signed-off-by: Tim Murray <timmurray@google.com>
> > > Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
> 
> We could certainly implement a down_read() variant (e.g.
> down_read_lowprio()) with its own slowpath function to do this within the
> rwsem code as long as there is a good use-case for this kind of
> functionality.

I think _unfair() or something along those lines is a *much* better
naming that _lowprio(). Consider a RT task ending up calling _lowprio().
That just doesn't make conceptual sense.

And then there's the lockdep angle; the thing being unfair will lead to
scenarios where lockdep will give a false positive because it expects
the r-w-r order to block things, which won't happen. A position needs to
be taken a-prioriy.

But before all that, a sane problem description. Can't propose solutions
without having a problem.