Date:   Thu, 11 Nov 2021 21:39:42 +0100
From:   Peter Zijlstra <peterz@infradead.org>
To:     Waiman Long <longman@redhat.com>
Cc:     Hillf Danton <hdanton@sina.com>,
        =?utf-8?B?6ams5oyv5Y2O?= <mazhenhua@xiaomi.com>,
        mingo <mingo@redhat.com>, will <will@kernel.org>,
        "boqun.feng" <boqun.feng@gmail.com>,
        linux-kernel <linux-kernel@vger.kernel.org>
Subject: Re: [BUG]locking/rwsem: only clean RWSEM_FLAG_HANDOFF when already
 set
Message-ID: <20211111203942.GJ174730@worktop.programming.kicks-ass.net>
References: <4fafad133b074f279dbab1aa3642e23f@xiaomi.com>
 <20211107090131.1535-1-hdanton@sina.com>
 <13d683ed-793c-b502-44ff-f28114d9386b@redhat.com>
 <02e118c0-2116-b806-2b48-b9c91dc847dd@redhat.com>
 <20211110213854.GE174703@worktop.programming.kicks-ass.net>
 <YY0x55wxO2v5HCOW@hirez.programming.kicks-ass.net>
 <61735528-141c-8d77-592d-b6b8fb75ebaa@redhat.com>
 <YY1s6v9b/tYtNnGv@hirez.programming.kicks-ass.net>
 <e16f9fc2-ce01-192b-065d-460c2ad9b317@redhat.com>
 <20211111203500.GI174703@worktop.programming.kicks-ass.net>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20211111203500.GI174703@worktop.programming.kicks-ass.net>
Precedence: bulk

On Thu, Nov 11, 2021 at 09:35:00PM +0100, Peter Zijlstra wrote:
> On Thu, Nov 11, 2021 at 02:36:52PM -0500, Waiman Long wrote:
> > 
> > On 11/11/21 14:20, Peter Zijlstra wrote:
> > > On Thu, Nov 11, 2021 at 02:14:48PM -0500, Waiman Long wrote:
> > > > As for the PHASE_CHANGE name, we have to be consistent in both rwsem and
> > > > mutex. Maybe a follow up patch if you think we should change the
> > > > terminology.
> > > Well, that's exactly the point, they do radically different things.
> > > Having the same name for two different things is confusing.
> > > 
> > > Anyway, let me go read that patch you sent.
> > 
> > My understanding of handoff is to disable optimistic spinning to let waiters
> > in the wait queue have an opportunity to acquire the lock. There are
> > difference in details on how to do that in mutex and rwsem, though.
> 
> Ah, but the mutex does an actual hand-off, it hands the lock to a
> specific waiting task. That is, unlock() sets owner, as opposed to
> trylock().
> 
> The rwsem code doesn't, it just forces a phase change. Once a waiter has
> been blocked too long, the handoff bit is set, causing new readers to be
> blocked. Then we wait for existing readers to complete. At that point,
> any next waiter (most likely a writer) should really get the lock (and
> in that regards the rwsem code is a bit funny).

And this is I think the thing you tried in your earlier inherit patch.
Keep the quescent state and simply let whatever next waiter is in line
have a go.

I suspect that change is easier now. But I've not tried.