[Intel-xe] [PATCH] drm/xe: Fix an invalid locking wait context bug

[Rodrigo Vivi]

On Fri, Jul 21, 2023 at 12:17:47PM +0100, Matthew Auld wrote:
> On 20/07/2023 16:42, Rodrigo Vivi wrote:
> > On Thu, Jul 20, 2023 at 01:38:18PM +0100, Matthew Auld wrote:
> > > On 20/07/2023 13:01, Rodrigo Vivi wrote:
> > > > On Thu, Jul 20, 2023 at 10:11:00AM +0100, Matthew Auld wrote:
> > > > > On 19/07/2023 20:27, Rodrigo Vivi wrote:
> > > > > > xe_irq_{suspend,resume} were incorrectly using the xe->irq.lock.
> > > > > > 
> > > > > > The lock was created to protect the gt irq handlers, and not
> > > > > > the irq.enabled. Since suspend/resume and other places touching
> > > > > > irq.enabled are already serialized they don't need protection.
> > > > > > (see other irq.enabled accesses).
> > > > > > 
> > > > > > Then with this spin lock around xe_irq_reset, we will end up
> > > > > > calling the intel_display_power_is_enabled() function, and
> > > > > > that needs a mutex lock. Hence causing the undesired
> > > > > > "[ BUG: Invalid wait context ]"
> > > > > > 
> > > > > > Cc: Matthew Auld <matthew.auld at intel.com>
> > > > > > Signed-off-by: Rodrigo Vivi <rodrigo.vivi at intel.com>
> > > > > > ---
> > > > > >     drivers/gpu/drm/xe/xe_irq.c | 5 -----
> > > > > >     1 file changed, 5 deletions(-)
> > > > > > 
> > > > > > diff --git a/drivers/gpu/drm/xe/xe_irq.c b/drivers/gpu/drm/xe/xe_irq.c
> > > > > > index eae190cb0969..df01af780a57 100644
> > > > > > --- a/drivers/gpu/drm/xe/xe_irq.c
> > > > > > +++ b/drivers/gpu/drm/xe/xe_irq.c
> > > > > > @@ -574,10 +574,8 @@ void xe_irq_shutdown(struct xe_device *xe)
> > > > > >     void xe_irq_suspend(struct xe_device *xe)
> > > > > >     {
> > > > > > -	spin_lock_irq(&xe->irq.lock);
> > > > > >     	xe->irq.enabled = false;
> > > > > >     	xe_irq_reset(xe);
> > > > > > -	spin_unlock_irq(&xe->irq.lock);
> > > > > 
> > > > > Do we not need something like:
> > > > > 
> > > > > spin_lock_irq(&xe->irq.lock);
> > > > > xe->irq.enabled = false; /* no new irqs */
> > > > > spin_unlock_irq(&xe->irq.lock);
> > > > > 
> > > > > synchronize_irq(...); /* flush irqs */
> > > > > xe_irq_reset(); /* turn off irqs */
> > > > > ....
> > > > > 
> > > > > And then at the start of the irq handler:
> > > > > 
> > > > > spin_lock_irq(&xe->irq.lock);
> > > > > if (!xe->irq.enabled) {
> > > > >       spin_unlock_irq(&xe->irq.lock);
> > > > >       return ....;
> > > > > }
> > > > > 
> > > > > Or did something happen prior to xe_irq_suspend() to ensure proper
> > > > > serialisation with irq and the above steps are not really needed?
> > > > 
> > > > the suspend and resume calls should be serialized by itself, no?!
> > > 
> > > Is it not possible for IRQs to still be firing or potentially be in-progress
> > > here as we are preparing to suspend?
> > 
> > yes, it is. We are letting the rpm to run with irq enabled otherwise
> > we will face the same invalid wait bug that this patch is trying to solve.
> > 
> > But I don't believe the right way is to use the lock to protect the irq.enabled.
> > 
> > Taking a look around I believe that what we are missing is the
> > synchronize_irq() call right after the reset. So we ensure that all the
> > racy handlers were properly processed before we allow the suspend.
> > 
> > So I believe we need something like i915 that would be:
> > 
> > xe_irq_suspend()
> > {
> > 	xe_irq_reset(xe);
> > 	xe->irq.enable = false;
> > 	synchronize_irq(pdev->irq);
> > }
> > 
> > xe_irq_resume()
> > {
> > 	xe->irq.enabled = true;
> > 	xe_irq_reset(xe);
> > 	xe_irq_postinstall(xe);
> > 
> > 	for_each_gt(gt, xe, id)
> > 			xe_irq_enable_hwe(gt);
> > }
> 
> Ok, but tbh I'm not completely sure what is going on with irq.enabled.
> AFAICT it looks to also be exposed into i915/display with
> intel_irqs_enabled() and that seems to have various callers. Are they not
> holding the lock?

Oh, they are really holding the lock.
I had missed the
#define irq_lock irq.lock
at xe/compat-i915-headers/i915_drv.h

In i915 I noticed something strange. instead of checking the irq_enable
they are using a secondary runtime_pm.irqs_enable and the only difference
is that this is set at the end of the irq reset function. while irq_enable
is set at the beginning.... it looks a hack to workaround races?!
at least very suspicious.

So, probably the right solution here lays in i915 side:

1. convert power_domains.lock from mutex to spinlock
(sent to trybot to get some feedback:
https://patchwork.freedesktop.org/series/121155/
)

2. then, add the spinlocks around the irq reset (or install and uninstall)

3. then get rid of this runtime_pm.irqs_enabled in favor
of the irq.enabled.

thoughts?

> Also the write here is to shared memory so without the
> lock it's unclear what is going on here wrt barriers and unmarked accesses
> (missing READ_ONCE/WRITE_ONCE?). I don't want to block you here since
> "Invalid wait context" is breaking CI, so I guess just go with the simplest
> thing for now.
> 
> > 
> > > 
> > > > 
> > > > no other place touching or inspecting irq.enable uses this lock
> > > > anyway, since it was created to serialize the gt_handler.
> > > > 
> > > > > 
> > > > > >     }
> > > > > >     void xe_irq_resume(struct xe_device *xe)
> > > > > > @@ -585,13 +583,10 @@ void xe_irq_resume(struct xe_device *xe)
> > > > > >     	struct xe_gt *gt;
> > > > > >     	int id;
> > > > > > -	spin_lock_irq(&xe->irq.lock);
> > > > > >     	xe->irq.enabled = true;
> > > > > >     	xe_irq_reset(xe);
> > > > > >     	xe_irq_postinstall(xe);
> > > > > >     	for_each_gt(gt, xe, id)
> > > > > >     		xe_irq_enable_hwe(gt);
> > > > > > -
> > > > > > -	spin_unlock_irq(&xe->irq.lock);
> > > > > >     }