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, 2 Nov 2020 10:09:27 +0100
From:   Andrea Righi <andrea.righi@canonical.com>
To:     Pavel Machek <pavel@ucw.cz>
Cc:     Boqun Feng <boqun.feng@gmail.com>,
        Peter Zijlstra <peterz@infradead.org>,
        Ingo Molnar <mingo@redhat.com>, Will Deacon <will@kernel.org>,
        linux-kernel@vger.kernel.org
Subject: Re: lockdep: possible irq lock inversion dependency detected
 (trig->leddev_list_lock)
Message-ID: <20201102090927.GC9930@xps-13-7390>
References: <20201101092614.GB3989@xps-13-7390>
 <20201031101740.GA1875@boqun-laptop.fareast.corp.microsoft.com>
 <20201102073328.GA9930@xps-13-7390>
 <20201102085658.GA5506@amd>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20201102085658.GA5506@amd>
Precedence: bulk

On Mon, Nov 02, 2020 at 09:56:58AM +0100, Pavel Machek wrote:
> Hi!
> 
> > > > I'm getting the following lockdep splat (see below).
> > > > 
> > > > Apparently this warning starts to be reported after applying:
> > > > 
> > > >  e918188611f0 ("locking: More accurate annotations for read_lock()")
> > > > 
> > > > It looks like a false positive to me, but it made me think a bit and
> > > > IIUC there can be still a potential deadlock, even if the deadlock
> > > > scenario is a bit different than what lockdep is showing.
> > > > 
> > > > In the assumption that read-locks are recursive only in_interrupt()
> > > > context (as stated in e918188611f0), the following scenario can still
> > > > happen:
> > > > 
> > > >  CPU0                                     CPU1
> > > >  ----                                     ----
> > > >  read_lock(&trig->leddev_list_lock);
> > > >                                           write_lock(&trig->leddev_list_lock);
> > > >  <soft-irq>
> > > >  kbd_bh()
> > > >    -> read_lock(&trig->leddev_list_lock);
> > > > 
> > > >  *** DEADLOCK ***
> > > > 
> > > > The write-lock is waiting on CPU1 and the second read_lock() on CPU0
> > > > would be blocked by the write-lock *waiter* on CPU1 => deadlock.
> > > > 
> > > 
> > > No, this is not a deadlock, as a write-lock waiter only blocks
> > > *non-recursive* readers, so since the read_lock() in kbd_bh() is called
> > > in soft-irq (which in_interrupt() returns true), so it's a recursive
> > > reader and won't get blocked by the write-lock waiter.
> > 
> > That's right, I was missing that in_interrupt() returns true also from
> > soft-irq context.
> > 
> > > > In that case we could prevent this deadlock condition using a workqueue
> > > > to call kbd_propagate_led_state() instead of calling it directly from
> > > > kbd_bh() (even if lockdep would still report the false positive).
> > > > 
> > > 
> > > The deadlock senario reported by the following splat is:
> > > 
> > > 	
> > > 	CPU 0:				CPU 1:					CPU 2:
> > > 	-----				-----					-----
> > > 	led_trigger_event():
> > > 	  read_lock(&trig->leddev_list_lock);
> > > 					<work queue processing>
> > > 	  				ata_hsm_qs_complete():
> > > 					  spin_lock_irqsave(&host->lock);
> > > 					  					write_lock(&trig->leddev_list_lock);
> > > 					  ata_port_freeze():
> > > 					    ata_do_link_abort():
> > > 					      ata_qc_complete():
> > > 					        ledtrig_disk_activity():
> > > 						  led_trigger_blink_oneshot():
> > > 						    read_lock(&trig->leddev_list_lock);
> > > 						    // ^ not in in_interrupt() context, so could get blocked by CPU 2
> > > 	<interrupt>
> > > 	  ata_bmdma_interrupt():
> > > 	    spin_lock_irqsave(&host->lock);
> > > 	  
> > > , where CPU 0 is blocked by CPU 1 because of the spin_lock_irqsave() in
> > > ata_bmdma_interrupt() and CPU 1 is blocked by CPU 2 because of the
> > > read_lock() in led_trigger_blink_oneshot() and CPU 2 is blocked by CPU 0
> > > because of an arbitrary writer on &trig->leddev_list_lock.
> > > 
> > > So I don't think it's false positive, but I might miss something
> > > obvious, because I don't know what the code here actually does ;-)
> > 
> > With the CPU2 part it all makes sense now and lockdep was right. :)
> > 
> > At this point I think we could just schedule a separate work to do the
> > led trigger and avoid calling it with host->lock held and that should
> > prevent the deadlock. I'll send a patch to do that.
> 
> Let's... not do that, unless we have no choice.
> 
> Would it help if leddev_list_lock used _irqsave() locking?

Using read_lock_irqsave/irqrestore() in led_trigger_event() would be
enough to prevent the deadlock. If it's an acceptable solution I can
send a patch (already tested it and lockdep doesn't complain :)).

Thanks,
-Andrea