Re: [PATCH v2 2/3] drm/panthor: Revisit reqs_lock handling in flush/reset paths

[Boris Brezillon]

On Mon, 03 Aug 2026 15:13:25 +0200
Nicolas Frattaroli <nicolas.frattaroli@xxxxxxxxxxxxx> wrote:

> On Monday, 3 August 2026 10:53:24 Central European Summer Time Boris Brezillon wrote:
> > Hello Nicolas,
> > 
> > On Thu, 30 Jul 2026 13:45:15 +0200
> > Nicolas Frattaroli <nicolas.frattaroli@xxxxxxxxxxxxx> wrote:
> >   
> > > panthor_gpu_flush_caches() and panthor_gpu_soft_reset() would read (and
> > > even reset) the contents of the pending_reqs register outside of holding
> > > the reqs_lock.  
> > 
> > Can you elaborate a bit on the race being fixed here? If pending_reqs bits
> > are truly cleared before the wake_up_all() call (which would require a
> > WRITE_ONCE() to be enforced, admittedly), there's no risk for the
> > wait_event() call to do a test before the bits have been updated,
> > and this holds even if the test is done without the lock held.
> > 
> > The other race I could think of is two threads calling
> > panthor_gpu_flush_caches() concurrently, and the second one stealing
> > the FLUSH_COMPLETED event the first thread waits on and re-issuing a
> > second flush on top, thus delaying the completion for the first thread.
> > But that should be covered by the cache_flush_lock.  
> 
> panthor_gpu_flush_caches() is not the only thing that sets/gets
> pending_reqs. Notably, the threaded interrupt handler does, as
> well as any other functionality using the same member for reqs
> tracking (e.g. the soft reset).
> 
> Consider the following serialisation of events:
> 1. T1 asks to flush caches by writing GPU_CMD and setting pending_reqs
> 2. T1 drops reqs_lock.
> 3. T2 enters IRQ handler for flush complete, spins lock waiting for
>    reqs_lock
> 4. T1 sleeps at wait_event_timeout
> 5. T2 updates pending_reqs and wakes up the waiter in any order, since
>    the effects of those two can't consistently be observed as sequential
>    logic without the outer reqs_lock being held by the observer
> 6. T1 wakes up, checks pending_reqs, but since pending_reqs is checked
>    without holding any lock, so we implictly depend on the synchronisation
>    point that is the waitqueue's lock rather than the reqs_lock spinlock,
>    which says nothing about whether the pending_reqs change materialised
>    on T1's side yet as far as I can tell?
> 7. T1 sees that pending_reqs & GPU_IRQ_CLEAN_CACHES_COMPLETED is still != 0,
>    so goes back to sleep for some future wake-up of reqs_acked or a timeout.
> 
> I'm not 100% sure, but I think 6. means that the memory model would permit
> T1 to re-use the pending_reqs it previously set, rather than the updated
> one set by T2, since there's nothing stopping us from being woken up before
> the pending_reqs change has made itself known to observers not serialising
> with the reqs_lock being released by T2 in panthor_gpu_irq_handler after.

If I read the "SLEEP AND WAKE-UP FUNCTIONS" section in [1] correctly, I
think we're covered by the "general memory barrier" here.

> 
> If you check lock_stat before the change, you see that the reqs_lock is
> actually never contended. This isn't a good sign because it means
> whatever situation it exists to protect against never occurs, so either
> the lock is pointless or the lock is non-functional.

Or we're just in a situation where there's almost never two threads updating
the pending_reqs concurrently, which is not the same as saying this can't
happen at all.

> 
> > > Additionally, when it did hold the lock, it did so with
> > > the irqsave/irqrestore variants, even though the spinlock was never
> > > acquired in an atomic context, just the threaded handler.
> > > 
> > > Use the new wait_event_lock_timeout() macro to check pending_reqs under
> > > the lock, and only do so without disabling interrupts.
> > > 
> > > Fixes: 5cd894e258c4 ("drm/panthor: Add the GPU logical block")
> > > Signed-off-by: Nicolas Frattaroli <nicolas.frattaroli@xxxxxxxxxxxxx>
> > > ---
> > >  drivers/gpu/drm/panthor/panthor_gpu.c | 25 +++++++++++--------------
> > >  1 file changed, 11 insertions(+), 14 deletions(-)
> > > 
> > > diff --git a/drivers/gpu/drm/panthor/panthor_gpu.c b/drivers/gpu/drm/panthor/panthor_gpu.c
> > > index c013d6bf9a59..f015bde80abf 100644
> > > --- a/drivers/gpu/drm/panthor/panthor_gpu.c
> > > +++ b/drivers/gpu/drm/panthor/panthor_gpu.c
> > > @@ -330,35 +330,34 @@ int panthor_gpu_flush_caches(struct panthor_device *ptdev,
> > >  			     u32 l2, u32 lsc, u32 other)
> > >  {
> > >  	struct panthor_gpu *gpu = ptdev->gpu;
> > > -	unsigned long flags;
> > >  	int ret = 0;
> > >  
> > >  	/* Serialize cache flush operations. */
> > >  	guard(mutex)(&ptdev->gpu->cache_flush_lock);
> > >  
> > > -	spin_lock_irqsave(&ptdev->gpu->reqs_lock, flags);
> > > +	spin_lock(&ptdev->gpu->reqs_lock);  
> > 
> > Can we make the _irq{save,restore}-drop its own patch?  
> 
> I'm not sure it's fine to drop the IRQ disabling without fixing the read of
> pending_reqs outside its lock.

Can be done after fixing the race, if it's really making the race more
likely. Although, as I said above, I'm not yet convinced the race we're
chasing is the one we've been discussing here.

> 
> >   
> > >  	if (!(ptdev->gpu->pending_reqs & GPU_IRQ_CLEAN_CACHES_COMPLETED)) {
> > >  		ptdev->gpu->pending_reqs |= GPU_IRQ_CLEAN_CACHES_COMPLETED;
> > >  		gpu_write(gpu->iomem, GPU_CMD, GPU_FLUSH_CACHES(l2, lsc, other));
> > >  	} else {
> > >  		ret = -EIO;
> > >  	}
> > > -	spin_unlock_irqrestore(&ptdev->gpu->reqs_lock, flags);
> > >  
> > > -	if (ret)
> > > +	if (ret) {
> > > +		spin_unlock(&ptdev->gpu->reqs_lock);
> > >  		return ret;
> > > +	}
> > >  
> > > -	if (!wait_event_timeout(ptdev->gpu->reqs_acked,
> > > +	if (!wait_event_lock_timeout(ptdev->gpu->reqs_acked,
> > >  				!(ptdev->gpu->pending_reqs & GPU_IRQ_CLEAN_CACHES_COMPLETED),  
> > 
> > Assuming we really need to do the test with the lock held, could we add
> > a patch at the beginning of the series that fixes the race without depending
> > on the new wait macro, so that we have a version that can easily be backported?  
> 
> It's either backporting the prerequisite new macro or still doing this all
> with IRQs disabled using the pre-existing wait_event_lock_irq_timeout macro,
> and the IRQ disabled thing is what caused problems.

Can't we have the following helper function?

static bool cache_flush_done(struct panthor_device *ptdev)
{
	guard(spinlock_irqsave)(&ptdev->gpu->reqs_lock);

	return ptdev->gpu->pending_reqs & GPU_IRQ_CLEAN_CACHES_COMPLETED;
}

and then:

	if (!wait_event_timeout(ptdev->gpu->reqs_acked,
			        cache_flush_done(ptdev),
				msecs_to_jiffies(100))) {
....

[1]https://docs.kernel.org/core-api/wrappers/memory-barriers.html#