[PATCH v3] drm/panfrost: Move the GPU reset bits outside the timeout handler

[Boris Brezillon]

On Tue, 3 Nov 2020 12:08:47 +0100
Daniel Vetter <daniel at ffwll.ch> wrote:

> On Tue, Nov 03, 2020 at 12:03:26PM +0100, Boris Brezillon wrote:
> > On Tue, 3 Nov 2020 11:25:40 +0100
> > Daniel Vetter <daniel at ffwll.ch> wrote:
> >   
> > > On Tue, Nov 03, 2020 at 09:13:47AM +0100, Boris Brezillon wrote:  
> > > > We've fixed many races in panfrost_job_timedout() but some remain.
> > > > Instead of trying to fix it again, let's simplify the logic and move
> > > > the reset bits to a separate work scheduled when one of the queue
> > > > reports a timeout.
> > > > 
> > > > v3:
> > > > - Replace the atomic_cmpxchg() by an atomic_xchg() (Robin Murphy)
> > > > - Add Steven's R-b
> > > > 
> > > > v2:
> > > > - Use atomic_cmpxchg() to conditionally schedule the reset work (Steven Price)
> > > > 
> > > > Fixes: 1a11a88cfd9a ("drm/panfrost: Fix job timeout handling")
> > > > Cc: <stable at vger.kernel.org>
> > > > Signed-off-by: Boris Brezillon <boris.brezillon at collabora.com>
> > > > Reviewed-by: Steven Price <steven.price at arm.com>    
> > > 
> > > Sprinkling the dma_fence annotations over this would be really nice ...  
> > 
> > You mean something like that?  
> 
> That's just the irq annotations, i.e. the one that's already guaranteed by
> the irq vs. locks checks. So this does nothing.
> 
> What I mean is annotating your new reset work (it's part of the critical
> path to complete batches, since it's holding up other batches that are
> stuck in the scheduler still), and the drm/scheduler annotations I've
> floated a while ago. The drm/scheduler annotations are stuck somewhat for
> lack of feedback from any of the driver teams using it though :-/
> 
> The thing is pulling something out into a worker of it's own generally
> doesn't fix any deadlocks, it just hides them from lockdep.

Hm, except that's not exactly a deadlock we were trying to fix here (as
in, not a situation where 2 threads try to acquire locks in different
orders), just a situation where the scheduler stops dequeuing jobs
because it ends up in an inconsistent state (which is caused by a
bad/lack-of synchronization between timeout handlers). The problem here
is that we have 3 schedulers (one per HW queue) but when a timeout
occurs on one of them, we need to reset them all, thus requiring some
synchronization between the different timeout works. Moving the reset
logic to a separate work simplifies the synchronization.

> So it would be
> good to make sure lockdep can see through your maze again.

Okay, but it's not clear to me which part of the panfrost_reset()
function should be annotated. I mean, I probably call functions that
can signal fences, but I don't call dma_signal_fence() directly. Are
callers of the dma_sched_xxx() helpers expected to place such
annotations?