[Intel-gfx] Interrupt handling across leave/entervt

Jesse Barnes jbarnes at virtuousgeek.org
Wed Dec 31 22:56:15 CET 2008


On Wednesday, December 31, 2008 12:03 pm Keith Packard wrote:
> On Wed, 2008-12-31 at 11:29 -0800, Jesse Barnes wrote:
> > The problem with disabling interrupts is that the server may wait on a
> > vblank event before the DDX has had a chance to call into the entervt
> > ioctl, resulting in a hang if say compiz is running before the VT switch
> > (see bug 18879).
>
> How about waking up anyone blocked on vblank at vt switch time instead?
>
> I kinda like the idea of not touching the HW while switched away.

The real problem is that we're violating the assumption in the drm_irq.c code, 
which tries to keep interrupts on across mode switches (which includes VT 
switch).  So what's happening is this:
  1) drm code takes a vblank ref
  2) vt switch ioctl disables irqs
  <sitting at console>
  3) vt switch re-enables irqs
  4) drm code puts vblank ref
  5) X server hangs waiting for an event (getting interrupted a lot)

Even though the drm_wait_vblank code takes a ref, the vblank interrupt doesn't 
actually get re-enabled since when we get to drm_vblank_get dev-
>vblank_enabled[crtc] is still set.  So something like the below also works.

But really, having the kernel try to keep its hands off of the hardware while 
X is VT switched away is of arguable benefit, so I'd still rather just leave 
things enabled across VT switches.

-- 
Jesse Barnes, Intel Open Source Technology Center


diff --git a/drivers/gpu/drm/drm_irq.c b/drivers/gpu/drm/drm_irq.c
index 1e787f8..e2f5d23 100644
--- a/drivers/gpu/drm/drm_irq.c
+++ b/drivers/gpu/drm/drm_irq.c
@@ -259,7 +259,8 @@ EXPORT_SYMBOL(drm_irq_install);
  */
 int drm_irq_uninstall(struct drm_device * dev)
 {
-	int irq_enabled;
+	unsigned long irqflags;
+	int irq_enabled, i;
 
 	if (!drm_core_check_feature(dev, DRIVER_HAVE_IRQ))
 		return -EINVAL;
@@ -269,6 +270,16 @@ int drm_irq_uninstall(struct drm_device * dev)
 	dev->irq_enabled = 0;
 	mutex_unlock(&dev->struct_mutex);
 
+	/*
+	 * Wake up any waiters so they don't hang.
+	 */
+	spin_lock_irqsave(&dev->vbl_lock, irqflags);
+	for (i = 0; i < dev->num_crtcs; i++) {
+		DRM_WAKEUP(&dev->vbl_queue[i]);
+		dev->vblank_enabled[i] = 0;
+	}
+	spin_unlock_irqrestore(&dev->vbl_lock, irqflags);
+
 	if (!irq_enabled)
 		return -EINVAL;
 
@@ -617,8 +628,9 @@ int drm_wait_vblank(struct drm_device *dev, void *data,
 		DRM_DEBUG("waiting on vblank count %d, crtc %d\n",
 			  vblwait->request.sequence, crtc);
 		DRM_WAIT_ON(ret, dev->vbl_queue[crtc], 3 * DRM_HZ,
-			    ((drm_vblank_count(dev, crtc)
-			      - vblwait->request.sequence) <= (1 << 23)));
+			    (((drm_vblank_count(dev, crtc) -
+			       vblwait->request.sequence) <= (1 << 23)) ||
+			     !dev->irq_enabled));
 
 		if (ret != -EINTR) {
 			struct timeval now;
 



More information about the Intel-gfx mailing list