drm/i915: Make dev_priv->mm.wedged an atomic_t

There is a very real possibility that multiple CPUs will notice that the
GPU is wedged. This introduces all sorts of potential race conditions.
Make the wedged flag atomic to mitigate this risk.
Signed-off-by: Ben Gamari <bgamari.foss@gmail.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>

drivers/gpu/drm/i915/i915_drv.h

@@ -450,7 +450,7 @@ typedef struct drm_i915_private {
 		 * It prevents command submission from occuring and makes
 		 * every pending request fail
 		 */
-		int wedged;
+		atomic_t wedged;
 
 		/** Bit 6 swizzling required for X tiling */
 		uint32_t bit_6_swizzle_x;

drivers/gpu/drm/i915/i915_gem.c

@@ -1712,7 +1712,7 @@ i915_gem_retire_requests(struct drm_device *dev)
 		retiring_seqno = request->seqno;
 
 		if (i915_seqno_passed(seqno, retiring_seqno) ||
-		    dev_priv->mm.wedged) {
+		    atomic_read(&dev_priv->mm.wedged)) {
 			i915_gem_retire_request(dev, request);
 
 			list_del(&request->list);
@@ -1754,7 +1754,7 @@ i915_wait_request(struct drm_device *dev, uint32_t seqno)
 
 	BUG_ON(seqno == 0);
 
-	if (dev_priv->mm.wedged)
+	if (atomic_read(&dev_priv->mm.wedged))
 		return -EIO;
 
 	if (!i915_seqno_passed(i915_get_gem_seqno(dev), seqno)) {
@@ -1774,11 +1774,11 @@ i915_wait_request(struct drm_device *dev, uint32_t seqno)
 		ret = wait_event_interruptible(dev_priv->irq_queue,
 					       i915_seqno_passed(i915_get_gem_seqno(dev),
 								 seqno) ||
-					       dev_priv->mm.wedged);
+					       atomic_read(&dev_priv->mm.wedged));
 		i915_user_irq_put(dev);
 		dev_priv->mm.waiting_gem_seqno = 0;
 	}
-	if (dev_priv->mm.wedged)
+	if (atomic_read(&dev_priv->mm.wedged))
 		ret = -EIO;
 
 	if (ret && ret != -ERESTARTSYS)
@@ -3359,7 +3359,7 @@ i915_gem_execbuffer(struct drm_device *dev, void *data,
 
 	i915_verify_inactive(dev, __FILE__, __LINE__);
 
-	if (dev_priv->mm.wedged) {
+	if (atomic_read(&dev_priv->mm.wedged)) {
 		DRM_ERROR("Execbuf while wedged\n");
 		mutex_unlock(&dev->struct_mutex);
 		ret = -EIO;
@@ -3929,7 +3929,7 @@ i915_gem_idle(struct drm_device *dev)
 		if (last_seqno == cur_seqno) {
 			if (stuck++ > 100) {
 				DRM_ERROR("hardware wedged\n");
-				dev_priv->mm.wedged = 1;
+				atomic_set(&dev_priv->mm.wedged, 1);
 				DRM_WAKEUP(&dev_priv->irq_queue);
 				break;
 			}
@@ -3942,7 +3942,7 @@ i915_gem_idle(struct drm_device *dev)
 	i915_gem_retire_requests(dev);
 
 	spin_lock(&dev_priv->mm.active_list_lock);
-	if (!dev_priv->mm.wedged) {
+	if (!atomic_read(&dev_priv->mm.wedged)) {
 		/* Active and flushing should now be empty as we've
 		 * waited for a sequence higher than any pending execbuffer
 		 */
@@ -4204,9 +4204,9 @@ i915_gem_entervt_ioctl(struct drm_device *dev, void *data,
 	if (drm_core_check_feature(dev, DRIVER_MODESET))
 		return 0;
 
-	if (dev_priv->mm.wedged) {
+	if (atomic_read(&dev_priv->mm.wedged)) {
 		DRM_ERROR("Reenabling wedged hardware, good luck\n");
-		dev_priv->mm.wedged = 0;
+		atomic_set(&dev_priv->mm.wedged, 0);
 	}
 
 	mutex_lock(&dev->struct_mutex);

drivers/gpu/drm/i915/i915_irq.c

@@ -309,12 +309,12 @@ static void i915_error_work_func(struct work_struct *work)
 	DRM_DEBUG("generating error event\n");
 	kobject_uevent_env(&dev->primary->kdev.kobj, KOBJ_CHANGE, error_event);
 
-	if (dev_priv->mm.wedged) {
+	if (atomic_read(&dev_priv->mm.wedged)) {
 		if (IS_I965G(dev)) {
 			DRM_DEBUG("resetting chip\n");
 			kobject_uevent_env(&dev->primary->kdev.kobj, KOBJ_CHANGE, reset_event);
 			if (!i965_reset(dev, GDRST_RENDER)) {
-				dev_priv->mm.wedged = 0;
+				atomic_set(&dev_priv->mm.wedged, 0);
 				kobject_uevent_env(&dev->primary->kdev.kobj, KOBJ_CHANGE, reset_done_event);
 			}
 		} else {
@@ -385,7 +385,7 @@ static void i915_capture_error_state(struct drm_device *dev)
  * so userspace knows something bad happened (should trigger collection
  * of a ring dump etc.).
  */
-static void i915_handle_error(struct drm_device *dev)
+static void i915_handle_error(struct drm_device *dev, bool wedged)
 {
 	struct drm_i915_private *dev_priv = dev->dev_private;
 	u32 eir = I915_READ(EIR);
@@ -495,7 +495,9 @@ static void i915_handle_error(struct drm_device *dev)
 		I915_WRITE(IIR, I915_RENDER_COMMAND_PARSER_ERROR_INTERRUPT);
 	}
 
-	if (dev_priv->mm.wedged) {
+	if (wedged) {
+		atomic_set(&dev_priv->mm.wedged, 1);
+
 		/*
 		 * Wakeup waiting processes so they don't hang
 		 */
@@ -548,7 +550,7 @@ irqreturn_t i915_driver_irq_handler(DRM_IRQ_ARGS)
 		pipeb_stats = I915_READ(PIPEBSTAT);
 
 		if (iir & I915_RENDER_COMMAND_PARSER_ERROR_INTERRUPT)
-			i915_handle_error(dev);
+			i915_handle_error(dev, false);
 
 		/*
 		 * Clear the PIPE(A|B)STAT regs before the IIR
@@ -934,8 +936,7 @@ void i915_hangcheck_elapsed(unsigned long data)
 
 	if (dev_priv->last_acthd == acthd && dev_priv->hangcheck_count > 0) {
 		DRM_ERROR("Hangcheck timer elapsed... GPU hung\n");
-		dev_priv->mm.wedged = true; /* Hopefully this is atomic */
-		i915_handle_error(dev);
+		i915_handle_error(dev, true);
 		return;
 	}