Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Mon, 31 Aug 2020 15:49:11 -0700
From:   "Paul E. McKenney" <paulmck@kernel.org>
To:     peterz@infradead.org, mingo@redhat.com, will@kernel.org
Cc:     linux-kernel@vger.kernel.org
Subject: Question on task_blocks_on_rt_mutex()
Message-ID: <20200831224911.GA13114@paulmck-ThinkPad-P72>
Reply-To: paulmck@kernel.org
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
User-Agent: Mutt/1.9.4 (2018-02-28)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

Hello!

The task_blocks_on_rt_mutex() function uses rt_mutex_owner() to
take a snapshot of the lock owner right up front.  At this point,
the ->wait_lock is held, which at first glance prevents the owner
from leaving.  Except that if there are not yet any waiters (that is,
the low-order bit of ->owner is zero), rt_mutex_fastunlock() might
locklessly clear the ->owner field.  And in that case, it looks like
task_blocks_on_rt_mutex() will blithely continue using the ex-owner's
task_struct structure, without anything that I can see that prevents
the ex-owner from exiting.

What am I missing here?

The reason that I am looking into this is that locktorture scenario LOCK05
hangs, and does so leaving the torture_rtmutex.waiters field equal to 0x1.
This is of course a legal transitional state, but I would not expect it
to persist for more than three minutes.  Yet it often does.

This leads me to believe that there is a way for an unlock to fail to wake
up a task concurrently acquiring the lock.  This seems to be repaired
by later lock acquisitions, and in fact setting the locktorture.stutter
module parameter to zero avoids the hang.  Except that I first found the
above apparently unprotected access to what was recently the owner task.

Thoughts?

							Thanx, Paul