[04/15] drm/i915: Cache the reset_counter for the request

Submitted by Chris Wilson on Nov. 29, 2015, 8:48 a.m.

Details

Message ID 1448786893-2522-5-git-send-email-chris@chris-wilson.co.uk
State New
Headers show
Series "Series without cover letter" ( rev: 3 2 1 ) in Intel GFX

Not browsing as part of any series.

Commit Message

Chris Wilson Nov. 29, 2015, 8:48 a.m.
Instead of querying the reset counter before every access to the ring,
query it the first time we touch the ring, and do a final compare when
submitting the request. For correctness, we need to then sanitize how
the reset_counter is incremented to prevent broken submission and
waiting across resets, in the process fixing the persistent -EIO we
still see today on failed waits.

v2: Rebase
v3: Now with added testcase
v4: Rebase

Testcase: igt/gem_eio
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
---
 drivers/gpu/drm/i915/i915_debugfs.c     |  2 +-
 drivers/gpu/drm/i915/i915_drv.c         | 32 +++++++-----
 drivers/gpu/drm/i915/i915_drv.h         | 39 ++++++++++----
 drivers/gpu/drm/i915/i915_gem.c         | 90 +++++++++++++++------------------
 drivers/gpu/drm/i915/i915_irq.c         | 21 +-------
 drivers/gpu/drm/i915/intel_display.c    | 13 ++---
 drivers/gpu/drm/i915/intel_lrc.c        | 13 +++--
 drivers/gpu/drm/i915/intel_ringbuffer.c | 14 ++---
 8 files changed, 111 insertions(+), 113 deletions(-)

Patch hide | download patch | download mbox

diff --git a/drivers/gpu/drm/i915/i915_debugfs.c b/drivers/gpu/drm/i915/i915_debugfs.c
index a728ff11e389..8458447ddc17 100644
--- a/drivers/gpu/drm/i915/i915_debugfs.c
+++ b/drivers/gpu/drm/i915/i915_debugfs.c
@@ -4660,7 +4660,7 @@  i915_wedged_get(void *data, u64 *val)
 	struct drm_device *dev = data;
 	struct drm_i915_private *dev_priv = dev->dev_private;
 
-	*val = atomic_read(&dev_priv->gpu_error.reset_counter);
+	*val = i915_terminally_wedged(&dev_priv->gpu_error);
 
 	return 0;
 }
diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c
index 90faa8e03fca..eb893a5e00b1 100644
--- a/drivers/gpu/drm/i915/i915_drv.c
+++ b/drivers/gpu/drm/i915/i915_drv.c
@@ -923,23 +923,31 @@  int i915_resume_switcheroo(struct drm_device *dev)
 int i915_reset(struct drm_device *dev)
 {
 	struct drm_i915_private *dev_priv = dev->dev_private;
+	struct i915_gpu_error *error = &dev_priv->gpu_error;
+	unsigned reset_counter;
 	bool simulated;
 	int ret;
 
 	intel_reset_gt_powersave(dev);
 
 	mutex_lock(&dev->struct_mutex);
+	atomic_andnot(I915_WEDGED, &error->reset_counter);
+	reset_counter = atomic_inc_return(&error->reset_counter);
+	if (WARN_ON(__i915_reset_in_progress(reset_counter))) {
+		ret = -EIO;
+		goto error;
+	}
 
 	i915_gem_reset(dev);
 
-	simulated = dev_priv->gpu_error.stop_rings != 0;
+	simulated = error->stop_rings != 0;
 
 	ret = intel_gpu_reset(dev);
 
 	/* Also reset the gpu hangman. */
 	if (simulated) {
 		DRM_INFO("Simulated gpu hang, resetting stop_rings\n");
-		dev_priv->gpu_error.stop_rings = 0;
+		error->stop_rings = 0;
 		if (ret == -ENODEV) {
 			DRM_INFO("Reset not implemented, but ignoring "
 				 "error for simulated gpu hangs\n");
@@ -952,8 +960,7 @@  int i915_reset(struct drm_device *dev)
 
 	if (ret) {
 		DRM_ERROR("Failed to reset chip: %i\n", ret);
-		mutex_unlock(&dev->struct_mutex);
-		return ret;
+		goto error;
 	}
 
 	intel_overlay_reset(dev_priv);
@@ -972,20 +979,14 @@  int i915_reset(struct drm_device *dev)
 	 * was running at the time of the reset (i.e. we weren't VT
 	 * switched away).
 	 */
-
-	/* Used to prevent gem_check_wedged returning -EAGAIN during gpu reset */
-	dev_priv->gpu_error.reload_in_reset = true;
-
 	ret = i915_gem_init_hw(dev);
-
-	dev_priv->gpu_error.reload_in_reset = false;
-
-	mutex_unlock(&dev->struct_mutex);
 	if (ret) {
 		DRM_ERROR("Failed hw init on reset %d\n", ret);
-		return ret;
+		goto error;
 	}
 
+	mutex_unlock(&dev->struct_mutex);
+
 	/*
 	 * rps/rc6 re-init is necessary to restore state lost after the
 	 * reset and the re-install of gt irqs. Skip for ironlake per
@@ -996,6 +997,11 @@  int i915_reset(struct drm_device *dev)
 		intel_enable_gt_powersave(dev);
 
 	return 0;
+
+error:
+	atomic_or(I915_WEDGED, &error->reset_counter);
+	mutex_unlock(&dev->struct_mutex);
+	return ret;
 }
 
 static int i915_pci_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index d07041c1729d..c24c23d8a0c0 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -1384,9 +1384,6 @@  struct i915_gpu_error {
 
 	/* For missed irq/seqno simulation. */
 	unsigned int test_irq_rings;
-
-	/* Used to prevent gem_check_wedged returning -EAGAIN during gpu reset   */
-	bool reload_in_reset;
 };
 
 enum modeset_restore {
@@ -2181,6 +2178,7 @@  struct drm_i915_gem_request {
 	/** On Which ring this request was generated */
 	struct drm_i915_private *i915;
 	struct intel_engine_cs *ring;
+	unsigned reset_counter;
 
 	 /** GEM sequence number associated with the previous request,
 	  * when the HWS breadcrumb is equal to this the GPU is processing
@@ -2976,23 +2974,47 @@  i915_gem_find_active_request(struct intel_engine_cs *ring);
 
 bool i915_gem_retire_requests(struct drm_device *dev);
 void i915_gem_retire_requests_ring(struct intel_engine_cs *ring);
-int __must_check i915_gem_check_wedge(struct i915_gpu_error *error,
+int __must_check i915_gem_check_wedge(unsigned reset_counter,
 				      bool interruptible);
 
+static inline u32 i915_reset_counter(struct i915_gpu_error *error)
+{
+	return atomic_read(&error->reset_counter);
+}
+
+static inline bool __i915_reset_in_progress(u32 reset)
+{
+	return unlikely(reset & I915_RESET_IN_PROGRESS_FLAG);
+}
+
+static inline bool __i915_reset_in_progress_or_wedged(u32 reset)
+{
+	return unlikely(reset & (I915_RESET_IN_PROGRESS_FLAG | I915_WEDGED));
+}
+
+static inline bool __i915_terminally_wedged(u32 reset)
+{
+	return unlikely(reset & I915_WEDGED);
+}
+
 static inline bool i915_reset_in_progress(struct i915_gpu_error *error)
 {
-	return unlikely(atomic_read(&error->reset_counter)
-			& (I915_RESET_IN_PROGRESS_FLAG | I915_WEDGED));
+	return __i915_reset_in_progress(i915_reset_counter(error));
+}
+
+static inline bool i915_reset_in_progress_or_wedged(struct i915_gpu_error *error)
+{
+	return __i915_reset_in_progress_or_wedged(i915_reset_counter(error));
 }
 
 static inline bool i915_terminally_wedged(struct i915_gpu_error *error)
 {
-	return atomic_read(&error->reset_counter) & I915_WEDGED;
+	return __i915_terminally_wedged(i915_reset_counter(error));
 }
 
 static inline u32 i915_reset_count(struct i915_gpu_error *error)
 {
-	return ((atomic_read(&error->reset_counter) & ~I915_WEDGED) + 1) / 2;
+	return ((i915_reset_counter(error) & ~I915_WEDGED) + 1) / 2;
 }
 
 static inline bool i915_stop_ring_allow_ban(struct drm_i915_private *dev_priv)
@@ -3025,7 +3047,6 @@  void __i915_add_request(struct drm_i915_gem_request *req,
 #define i915_add_request_no_flush(req) \
 	__i915_add_request(req, NULL, false)
 int __i915_wait_request(struct drm_i915_gem_request *req,
-			unsigned reset_counter,
 			bool interruptible,
 			s64 *timeout,
 			struct intel_rps_client *rps);
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index ada461e02718..646d189e23a1 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -85,9 +85,7 @@  i915_gem_wait_for_error(struct i915_gpu_error *error)
 {
 	int ret;
 
-#define EXIT_COND (!i915_reset_in_progress(error) || \
-		   i915_terminally_wedged(error))
-	if (EXIT_COND)
+	if (!i915_reset_in_progress(error))
 		return 0;
 
 	/*
@@ -96,17 +94,16 @@  i915_gem_wait_for_error(struct i915_gpu_error *error)
 	 * we should simply try to bail out and fail as gracefully as possible.
 	 */
 	ret = wait_event_interruptible_timeout(error->reset_queue,
-					       EXIT_COND,
+					       !i915_reset_in_progress(error),
 					       10*HZ);
 	if (ret == 0) {
 		DRM_ERROR("Timed out waiting for the gpu reset to complete\n");
 		return -EIO;
 	} else if (ret < 0) {
 		return ret;
+	} else {
+		return 0;
 	}
-#undef EXIT_COND
-
-	return 0;
 }
 
 int i915_mutex_lock_interruptible(struct drm_device *dev)
@@ -1110,26 +1107,20 @@  put_rpm:
 }
 
 int
-i915_gem_check_wedge(struct i915_gpu_error *error,
+i915_gem_check_wedge(unsigned reset_counter,
 		     bool interruptible)
 {
-	if (i915_reset_in_progress(error)) {
+	if (__i915_reset_in_progress_or_wedged(reset_counter)) {
 		/* Non-interruptible callers can't handle -EAGAIN, hence return
 		 * -EIO unconditionally for these. */
 		if (!interruptible)
 			return -EIO;
 
 		/* Recovery complete, but the reset failed ... */
-		if (i915_terminally_wedged(error))
+		if (__i915_terminally_wedged(reset_counter))
 			return -EIO;
 
-		/*
-		 * Check if GPU Reset is in progress - we need intel_ring_begin
-		 * to work properly to reinit the hw state while the gpu is
-		 * still marked as reset-in-progress. Handle this with a flag.
-		 */
-		if (!error->reload_in_reset)
-			return -EAGAIN;
+		return -EAGAIN;
 	}
 
 	return 0;
@@ -1223,7 +1214,6 @@  static int __i915_spin_request(struct drm_i915_gem_request *req, int state)
 /**
  * __i915_wait_request - wait until execution of request has finished
  * @req: duh!
- * @reset_counter: reset sequence associated with the given request
  * @interruptible: do an interruptible wait (normally yes)
  * @timeout: in - how long to wait (NULL forever); out - how much time remaining
  *
@@ -1238,7 +1228,6 @@  static int __i915_spin_request(struct drm_i915_gem_request *req, int state)
  * errno with remaining time filled in timeout argument.
  */
 int __i915_wait_request(struct drm_i915_gem_request *req,
-			unsigned reset_counter,
 			bool interruptible,
 			s64 *timeout,
 			struct intel_rps_client *rps)
@@ -1289,12 +1278,12 @@  int __i915_wait_request(struct drm_i915_gem_request *req,
 
 		/* We need to check whether any gpu reset happened in between
 		 * the caller grabbing the seqno and now ... */
-		if (reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter)) {
-			/* ... but upgrade the -EAGAIN to an -EIO if the gpu
-			 * is truely gone. */
-			ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
-			if (ret == 0)
-				ret = -EAGAIN;
+		if (req->reset_counter != i915_reset_counter(&dev_priv->gpu_error)) {
+			/* As we do not requeue the request over a GPU reset,
+			 * if one does occur we know that the request is
+			 * effectively complete.
+			 */
+			ret = 0;
 			break;
 		}
 
@@ -1462,13 +1451,7 @@  i915_wait_request(struct drm_i915_gem_request *req)
 
 	BUG_ON(!mutex_is_locked(&dev->struct_mutex));
 
-	ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
-	if (ret)
-		return ret;
-
-	ret = __i915_wait_request(req,
-				  atomic_read(&dev_priv->gpu_error.reset_counter),
-				  interruptible, NULL, NULL);
+	ret = __i915_wait_request(req, interruptible, NULL, NULL);
 	if (ret)
 		return ret;
 
@@ -1543,7 +1526,6 @@  i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj,
 	struct drm_device *dev = obj->base.dev;
 	struct drm_i915_private *dev_priv = dev->dev_private;
 	struct drm_i915_gem_request *requests[I915_NUM_RINGS];
-	unsigned reset_counter;
 	int ret, i, n = 0;
 
 	BUG_ON(!mutex_is_locked(&dev->struct_mutex));
@@ -1552,12 +1534,6 @@  i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj,
 	if (!obj->active)
 		return 0;
 
-	ret = i915_gem_check_wedge(&dev_priv->gpu_error, true);
-	if (ret)
-		return ret;
-
-	reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
-
 	if (readonly) {
 		struct drm_i915_gem_request *req;
 
@@ -1579,9 +1555,9 @@  i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj,
 	}
 
 	mutex_unlock(&dev->struct_mutex);
+	ret = 0;
 	for (i = 0; ret == 0 && i < n; i++)
-		ret = __i915_wait_request(requests[i], reset_counter, true,
-					  NULL, rps);
+		ret = __i915_wait_request(requests[i], true, NULL, rps);
 	mutex_lock(&dev->struct_mutex);
 
 	for (i = 0; i < n; i++) {
@@ -2685,6 +2661,7 @@  int i915_gem_request_alloc(struct intel_engine_cs *ring,
 			   struct drm_i915_gem_request **req_out)
 {
 	struct drm_i915_private *dev_priv = to_i915(ring->dev);
+	unsigned reset_counter = i915_reset_counter(&dev_priv->gpu_error);
 	struct drm_i915_gem_request *req;
 	int ret;
 
@@ -2693,6 +2670,10 @@  int i915_gem_request_alloc(struct intel_engine_cs *ring,
 
 	*req_out = NULL;
 
+	ret = i915_gem_check_wedge(reset_counter, dev_priv->mm.interruptible);
+	if (ret)
+		return ret;
+
 	req = kmem_cache_zalloc(dev_priv->requests, GFP_KERNEL);
 	if (req == NULL)
 		return -ENOMEM;
@@ -2704,6 +2685,7 @@  int i915_gem_request_alloc(struct intel_engine_cs *ring,
 	kref_init(&req->ref);
 	req->i915 = dev_priv;
 	req->ring = ring;
+	req->reset_counter = reset_counter;
 	req->ctx  = ctx;
 	i915_gem_context_reference(req->ctx);
 
@@ -3064,11 +3046,9 @@  retire:
 int
 i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
 {
-	struct drm_i915_private *dev_priv = dev->dev_private;
 	struct drm_i915_gem_wait *args = data;
 	struct drm_i915_gem_object *obj;
 	struct drm_i915_gem_request *req[I915_NUM_RINGS];
-	unsigned reset_counter;
 	int i, n = 0;
 	int ret;
 
@@ -3102,7 +3082,6 @@  i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
 	}
 
 	drm_gem_object_unreference(&obj->base);
-	reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
 
 	for (i = 0; i < I915_NUM_RINGS; i++) {
 		if (obj->last_read_req[i] == NULL)
@@ -3115,11 +3094,23 @@  i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
 
 	for (i = 0; i < n; i++) {
 		if (ret == 0)
-			ret = __i915_wait_request(req[i], reset_counter, true,
+			ret = __i915_wait_request(req[i], true,
 						  args->timeout_ns > 0 ? &args->timeout_ns : NULL,
 						  file->driver_priv);
+
+		/* If the GPU hung before this, report it. Ideally we only
+		 * report if this request cannot be completed. Currently
+		 * when we don't mark the guilty party and abort all
+		 * requests on reset, so just mark all as EIO.
+		 */
+		if (ret == 0 &&
+		    req[i]->reset_counter != i915_reset_counter(&req[i]->i915->gpu_error))
+			ret = -EIO;
+
 		i915_gem_request_unreference__unlocked(req[i]);
 	}
+
+
 	return ret;
 
 out:
@@ -3147,7 +3138,6 @@  __i915_gem_object_sync(struct drm_i915_gem_object *obj,
 	if (!i915_semaphore_is_enabled(obj->base.dev)) {
 		struct drm_i915_private *i915 = to_i915(obj->base.dev);
 		ret = __i915_wait_request(from_req,
-					  atomic_read(&i915->gpu_error.reset_counter),
 					  i915->mm.interruptible,
 					  NULL,
 					  &i915->rps.semaphores);
@@ -4076,14 +4066,15 @@  i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
 	struct drm_i915_file_private *file_priv = file->driver_priv;
 	unsigned long recent_enough = jiffies - DRM_I915_THROTTLE_JIFFIES;
 	struct drm_i915_gem_request *request, *target = NULL;
-	unsigned reset_counter;
 	int ret;
 
 	ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
 	if (ret)
 		return ret;
 
-	ret = i915_gem_check_wedge(&dev_priv->gpu_error, false);
+	/* ABI: return -EIO if wedged */
+	ret = i915_gem_check_wedge(i915_reset_counter(&dev_priv->gpu_error),
+				   false);
 	if (ret)
 		return ret;
 
@@ -4101,7 +4092,6 @@  i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
 
 		target = request;
 	}
-	reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
 	if (target)
 		i915_gem_request_reference(target);
 	spin_unlock(&file_priv->mm.lock);
@@ -4109,7 +4099,7 @@  i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
 	if (target == NULL)
 		return 0;
 
-	ret = __i915_wait_request(target, reset_counter, true, NULL, NULL);
+	ret = __i915_wait_request(target, true, NULL, NULL);
 	if (ret == 0)
 		queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0);
 
diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c
index e88d692583a5..3b2a8fbe0392 100644
--- a/drivers/gpu/drm/i915/i915_irq.c
+++ b/drivers/gpu/drm/i915/i915_irq.c
@@ -2434,7 +2434,6 @@  static void i915_error_wake_up(struct drm_i915_private *dev_priv,
 static void i915_reset_and_wakeup(struct drm_device *dev)
 {
 	struct drm_i915_private *dev_priv = to_i915(dev);
-	struct i915_gpu_error *error = &dev_priv->gpu_error;
 	char *error_event[] = { I915_ERROR_UEVENT "=1", NULL };
 	char *reset_event[] = { I915_RESET_UEVENT "=1", NULL };
 	char *reset_done_event[] = { I915_ERROR_UEVENT "=0", NULL };
@@ -2452,7 +2451,7 @@  static void i915_reset_and_wakeup(struct drm_device *dev)
 	 * the reset in-progress bit is only ever set by code outside of this
 	 * work we don't need to worry about any other races.
 	 */
-	if (i915_reset_in_progress(error) && !i915_terminally_wedged(error)) {
+	if (i915_reset_in_progress(&dev_priv->gpu_error)) {
 		DRM_DEBUG_DRIVER("resetting chip\n");
 		kobject_uevent_env(&dev->primary->kdev->kobj, KOBJ_CHANGE,
 				   reset_event);
@@ -2480,25 +2479,9 @@  static void i915_reset_and_wakeup(struct drm_device *dev)
 
 		intel_runtime_pm_put(dev_priv);
 
-		if (ret == 0) {
-			/*
-			 * After all the gem state is reset, increment the reset
-			 * counter and wake up everyone waiting for the reset to
-			 * complete.
-			 *
-			 * Since unlock operations are a one-sided barrier only,
-			 * we need to insert a barrier here to order any seqno
-			 * updates before
-			 * the counter increment.
-			 */
-			smp_mb__before_atomic();
-			atomic_inc(&dev_priv->gpu_error.reset_counter);
-
+		if (ret == 0)
 			kobject_uevent_env(&dev->primary->kdev->kobj,
 					   KOBJ_CHANGE, reset_done_event);
-		} else {
-			atomic_or(I915_WEDGED, &error->reset_counter);
-		}
 
 		/*
 		 * Note: The wake_up also serves as a memory barrier so that
diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c
index 0743337ac9e8..d77a17bb94d0 100644
--- a/drivers/gpu/drm/i915/intel_display.c
+++ b/drivers/gpu/drm/i915/intel_display.c
@@ -3289,8 +3289,7 @@  static bool intel_crtc_has_pending_flip(struct drm_crtc *crtc)
 	struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
 	bool pending;
 
-	if (i915_reset_in_progress(&dev_priv->gpu_error) ||
-	    intel_crtc->reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter))
+	if (intel_crtc->reset_counter != i915_reset_counter(&dev_priv->gpu_error))
 		return false;
 
 	spin_lock_irq(&dev->event_lock);
@@ -10879,8 +10878,7 @@  static bool page_flip_finished(struct intel_crtc *crtc)
 	struct drm_device *dev = crtc->base.dev;
 	struct drm_i915_private *dev_priv = dev->dev_private;
 
-	if (i915_reset_in_progress(&dev_priv->gpu_error) ||
-	    crtc->reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter))
+	if (crtc->reset_counter != i915_reset_counter(&dev_priv->gpu_error))
 		return true;
 
 	/*
@@ -11322,7 +11320,6 @@  static void intel_mmio_flip_work_func(struct work_struct *work)
 
 	if (mmio_flip->req) {
 		WARN_ON(__i915_wait_request(mmio_flip->req,
-					    mmio_flip->crtc->reset_counter,
 					    false, NULL,
 					    &mmio_flip->i915->rps.mmioflips));
 		i915_gem_request_unreference__unlocked(mmio_flip->req);
@@ -13312,9 +13309,6 @@  static int intel_atomic_prepare_commit(struct drm_device *dev,
 
 	ret = drm_atomic_helper_prepare_planes(dev, state);
 	if (!ret && !async && !i915_reset_in_progress(&dev_priv->gpu_error)) {
-		u32 reset_counter;
-
-		reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
 		mutex_unlock(&dev->struct_mutex);
 
 		for_each_plane_in_state(state, plane, plane_state, i) {
@@ -13325,8 +13319,7 @@  static int intel_atomic_prepare_commit(struct drm_device *dev,
 				continue;
 
 			ret = __i915_wait_request(intel_plane_state->wait_req,
-						  reset_counter, true,
-						  NULL, NULL);
+						  true, NULL, NULL);
 
 			/* Swallow -EIO errors to allow updates during hw lockup. */
 			if (ret == -EIO)
diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
index 4ebafab53f30..b40ffc3607e7 100644
--- a/drivers/gpu/drm/i915/intel_lrc.c
+++ b/drivers/gpu/drm/i915/intel_lrc.c
@@ -711,6 +711,14 @@  static int logical_ring_wait_for_space(struct drm_i915_gem_request *req,
 	if (ret)
 		return ret;
 
+	/* If the request was completed due to a GPU hang, we want to
+	 * error out before we continue to emit more commands to the GPU.
+	 */
+	ret = i915_gem_check_wedge(i915_reset_counter(&req->i915->gpu_error),
+				   req->i915->mm.interruptible);
+	if (ret)
+		return ret;
+
 	ringbuf->space = space;
 	return 0;
 }
@@ -825,11 +833,6 @@  int intel_logical_ring_begin(struct drm_i915_gem_request *req, int num_dwords)
 	WARN_ON(req == NULL);
 	dev_priv = req->ring->dev->dev_private;
 
-	ret = i915_gem_check_wedge(&dev_priv->gpu_error,
-				   dev_priv->mm.interruptible);
-	if (ret)
-		return ret;
-
 	ret = logical_ring_prepare(req, num_dwords * sizeof(uint32_t));
 	if (ret)
 		return ret;
diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.c b/drivers/gpu/drm/i915/intel_ringbuffer.c
index 57d78f264b53..511efe556d73 100644
--- a/drivers/gpu/drm/i915/intel_ringbuffer.c
+++ b/drivers/gpu/drm/i915/intel_ringbuffer.c
@@ -2254,6 +2254,14 @@  static int ring_wait_for_space(struct intel_engine_cs *ring, int n)
 	if (ret)
 		return ret;
 
+	/* If the request was completed due to a GPU hang, we want to
+	 * error out before we continue to emit more commands to the GPU.
+	 */
+	ret = i915_gem_check_wedge(i915_reset_counter(&to_i915(ring->dev)->gpu_error),
+				   to_i915(ring->dev)->mm.interruptible);
+	if (ret)
+		return ret;
+
 	ringbuf->space = space;
 	return 0;
 }
@@ -2286,7 +2294,6 @@  int intel_ring_idle(struct intel_engine_cs *ring)
 
 	/* Make sure we do not trigger any retires */
 	return __i915_wait_request(req,
-				   atomic_read(&to_i915(ring->dev)->gpu_error.reset_counter),
 				   to_i915(ring->dev)->mm.interruptible,
 				   NULL, NULL);
 }
@@ -2417,11 +2424,6 @@  int intel_ring_begin(struct drm_i915_gem_request *req,
 	ring = req->ring;
 	dev_priv = ring->dev->dev_private;
 
-	ret = i915_gem_check_wedge(&dev_priv->gpu_error,
-				   dev_priv->mm.interruptible);
-	if (ret)
-		return ret;
-
 	ret = __intel_ring_prepare(ring, num_dwords * sizeof(uint32_t));
 	if (ret)
 		return ret;

Comments

On Sun, Nov 29, 2015 at 08:48:02AM +0000, Chris Wilson wrote:
> Instead of querying the reset counter before every access to the ring,
> query it the first time we touch the ring, and do a final compare when
> submitting the request. For correctness, we need to then sanitize how
> the reset_counter is incremented to prevent broken submission and
> waiting across resets, in the process fixing the persistent -EIO we
> still see today on failed waits.
> 
> v2: Rebase
> v3: Now with added testcase
> v4: Rebase
> 
> Testcase: igt/gem_eio
> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

Imo you really can't hide an ABI change in a refactor/optimize patch like
this ... I really think moving check_wedge out of wait_request must be
it's own little patch.
-Daniel

> ---
>  drivers/gpu/drm/i915/i915_debugfs.c     |  2 +-
>  drivers/gpu/drm/i915/i915_drv.c         | 32 +++++++-----
>  drivers/gpu/drm/i915/i915_drv.h         | 39 ++++++++++----
>  drivers/gpu/drm/i915/i915_gem.c         | 90 +++++++++++++++------------------
>  drivers/gpu/drm/i915/i915_irq.c         | 21 +-------
>  drivers/gpu/drm/i915/intel_display.c    | 13 ++---
>  drivers/gpu/drm/i915/intel_lrc.c        | 13 +++--
>  drivers/gpu/drm/i915/intel_ringbuffer.c | 14 ++---
>  8 files changed, 111 insertions(+), 113 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/i915_debugfs.c b/drivers/gpu/drm/i915/i915_debugfs.c
> index a728ff11e389..8458447ddc17 100644
> --- a/drivers/gpu/drm/i915/i915_debugfs.c
> +++ b/drivers/gpu/drm/i915/i915_debugfs.c
> @@ -4660,7 +4660,7 @@ i915_wedged_get(void *data, u64 *val)
>  	struct drm_device *dev = data;
>  	struct drm_i915_private *dev_priv = dev->dev_private;
>  
> -	*val = atomic_read(&dev_priv->gpu_error.reset_counter);
> +	*val = i915_terminally_wedged(&dev_priv->gpu_error);
>  
>  	return 0;
>  }
> diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c
> index 90faa8e03fca..eb893a5e00b1 100644
> --- a/drivers/gpu/drm/i915/i915_drv.c
> +++ b/drivers/gpu/drm/i915/i915_drv.c
> @@ -923,23 +923,31 @@ int i915_resume_switcheroo(struct drm_device *dev)
>  int i915_reset(struct drm_device *dev)
>  {
>  	struct drm_i915_private *dev_priv = dev->dev_private;
> +	struct i915_gpu_error *error = &dev_priv->gpu_error;
> +	unsigned reset_counter;
>  	bool simulated;
>  	int ret;
>  
>  	intel_reset_gt_powersave(dev);
>  
>  	mutex_lock(&dev->struct_mutex);
> +	atomic_andnot(I915_WEDGED, &error->reset_counter);
> +	reset_counter = atomic_inc_return(&error->reset_counter);
> +	if (WARN_ON(__i915_reset_in_progress(reset_counter))) {
> +		ret = -EIO;
> +		goto error;
> +	}
>  
>  	i915_gem_reset(dev);
>  
> -	simulated = dev_priv->gpu_error.stop_rings != 0;
> +	simulated = error->stop_rings != 0;
>  
>  	ret = intel_gpu_reset(dev);
>  
>  	/* Also reset the gpu hangman. */
>  	if (simulated) {
>  		DRM_INFO("Simulated gpu hang, resetting stop_rings\n");
> -		dev_priv->gpu_error.stop_rings = 0;
> +		error->stop_rings = 0;
>  		if (ret == -ENODEV) {
>  			DRM_INFO("Reset not implemented, but ignoring "
>  				 "error for simulated gpu hangs\n");
> @@ -952,8 +960,7 @@ int i915_reset(struct drm_device *dev)
>  
>  	if (ret) {
>  		DRM_ERROR("Failed to reset chip: %i\n", ret);
> -		mutex_unlock(&dev->struct_mutex);
> -		return ret;
> +		goto error;
>  	}
>  
>  	intel_overlay_reset(dev_priv);
> @@ -972,20 +979,14 @@ int i915_reset(struct drm_device *dev)
>  	 * was running at the time of the reset (i.e. we weren't VT
>  	 * switched away).
>  	 */
> -
> -	/* Used to prevent gem_check_wedged returning -EAGAIN during gpu reset */
> -	dev_priv->gpu_error.reload_in_reset = true;
> -
>  	ret = i915_gem_init_hw(dev);
> -
> -	dev_priv->gpu_error.reload_in_reset = false;
> -
> -	mutex_unlock(&dev->struct_mutex);
>  	if (ret) {
>  		DRM_ERROR("Failed hw init on reset %d\n", ret);
> -		return ret;
> +		goto error;
>  	}
>  
> +	mutex_unlock(&dev->struct_mutex);
> +
>  	/*
>  	 * rps/rc6 re-init is necessary to restore state lost after the
>  	 * reset and the re-install of gt irqs. Skip for ironlake per
> @@ -996,6 +997,11 @@ int i915_reset(struct drm_device *dev)
>  		intel_enable_gt_powersave(dev);
>  
>  	return 0;
> +
> +error:
> +	atomic_or(I915_WEDGED, &error->reset_counter);
> +	mutex_unlock(&dev->struct_mutex);
> +	return ret;
>  }
>  
>  static int i915_pci_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
> diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> index d07041c1729d..c24c23d8a0c0 100644
> --- a/drivers/gpu/drm/i915/i915_drv.h
> +++ b/drivers/gpu/drm/i915/i915_drv.h
> @@ -1384,9 +1384,6 @@ struct i915_gpu_error {
>  
>  	/* For missed irq/seqno simulation. */
>  	unsigned int test_irq_rings;
> -
> -	/* Used to prevent gem_check_wedged returning -EAGAIN during gpu reset   */
> -	bool reload_in_reset;
>  };
>  
>  enum modeset_restore {
> @@ -2181,6 +2178,7 @@ struct drm_i915_gem_request {
>  	/** On Which ring this request was generated */
>  	struct drm_i915_private *i915;
>  	struct intel_engine_cs *ring;
> +	unsigned reset_counter;
>  
>  	 /** GEM sequence number associated with the previous request,
>  	  * when the HWS breadcrumb is equal to this the GPU is processing
> @@ -2976,23 +2974,47 @@ i915_gem_find_active_request(struct intel_engine_cs *ring);
>  
>  bool i915_gem_retire_requests(struct drm_device *dev);
>  void i915_gem_retire_requests_ring(struct intel_engine_cs *ring);
> -int __must_check i915_gem_check_wedge(struct i915_gpu_error *error,
> +int __must_check i915_gem_check_wedge(unsigned reset_counter,
>  				      bool interruptible);
>  
> +static inline u32 i915_reset_counter(struct i915_gpu_error *error)
> +{
> +	return atomic_read(&error->reset_counter);
> +}
> +
> +static inline bool __i915_reset_in_progress(u32 reset)
> +{
> +	return unlikely(reset & I915_RESET_IN_PROGRESS_FLAG);
> +}
> +
> +static inline bool __i915_reset_in_progress_or_wedged(u32 reset)
> +{
> +	return unlikely(reset & (I915_RESET_IN_PROGRESS_FLAG | I915_WEDGED));
> +}
> +
> +static inline bool __i915_terminally_wedged(u32 reset)
> +{
> +	return unlikely(reset & I915_WEDGED);
> +}
> +
>  static inline bool i915_reset_in_progress(struct i915_gpu_error *error)
>  {
> -	return unlikely(atomic_read(&error->reset_counter)
> -			& (I915_RESET_IN_PROGRESS_FLAG | I915_WEDGED));
> +	return __i915_reset_in_progress(i915_reset_counter(error));
> +}
> +
> +static inline bool i915_reset_in_progress_or_wedged(struct i915_gpu_error *error)
> +{
> +	return __i915_reset_in_progress_or_wedged(i915_reset_counter(error));
>  }
>  
>  static inline bool i915_terminally_wedged(struct i915_gpu_error *error)
>  {
> -	return atomic_read(&error->reset_counter) & I915_WEDGED;
> +	return __i915_terminally_wedged(i915_reset_counter(error));
>  }
>  
>  static inline u32 i915_reset_count(struct i915_gpu_error *error)
>  {
> -	return ((atomic_read(&error->reset_counter) & ~I915_WEDGED) + 1) / 2;
> +	return ((i915_reset_counter(error) & ~I915_WEDGED) + 1) / 2;
>  }
>  
>  static inline bool i915_stop_ring_allow_ban(struct drm_i915_private *dev_priv)
> @@ -3025,7 +3047,6 @@ void __i915_add_request(struct drm_i915_gem_request *req,
>  #define i915_add_request_no_flush(req) \
>  	__i915_add_request(req, NULL, false)
>  int __i915_wait_request(struct drm_i915_gem_request *req,
> -			unsigned reset_counter,
>  			bool interruptible,
>  			s64 *timeout,
>  			struct intel_rps_client *rps);
> diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
> index ada461e02718..646d189e23a1 100644
> --- a/drivers/gpu/drm/i915/i915_gem.c
> +++ b/drivers/gpu/drm/i915/i915_gem.c
> @@ -85,9 +85,7 @@ i915_gem_wait_for_error(struct i915_gpu_error *error)
>  {
>  	int ret;
>  
> -#define EXIT_COND (!i915_reset_in_progress(error) || \
> -		   i915_terminally_wedged(error))
> -	if (EXIT_COND)
> +	if (!i915_reset_in_progress(error))
>  		return 0;
>  
>  	/*
> @@ -96,17 +94,16 @@ i915_gem_wait_for_error(struct i915_gpu_error *error)
>  	 * we should simply try to bail out and fail as gracefully as possible.
>  	 */
>  	ret = wait_event_interruptible_timeout(error->reset_queue,
> -					       EXIT_COND,
> +					       !i915_reset_in_progress(error),
>  					       10*HZ);
>  	if (ret == 0) {
>  		DRM_ERROR("Timed out waiting for the gpu reset to complete\n");
>  		return -EIO;
>  	} else if (ret < 0) {
>  		return ret;
> +	} else {
> +		return 0;
>  	}
> -#undef EXIT_COND
> -
> -	return 0;
>  }
>  
>  int i915_mutex_lock_interruptible(struct drm_device *dev)
> @@ -1110,26 +1107,20 @@ put_rpm:
>  }
>  
>  int
> -i915_gem_check_wedge(struct i915_gpu_error *error,
> +i915_gem_check_wedge(unsigned reset_counter,
>  		     bool interruptible)
>  {
> -	if (i915_reset_in_progress(error)) {
> +	if (__i915_reset_in_progress_or_wedged(reset_counter)) {
>  		/* Non-interruptible callers can't handle -EAGAIN, hence return
>  		 * -EIO unconditionally for these. */
>  		if (!interruptible)
>  			return -EIO;
>  
>  		/* Recovery complete, but the reset failed ... */
> -		if (i915_terminally_wedged(error))
> +		if (__i915_terminally_wedged(reset_counter))
>  			return -EIO;
>  
> -		/*
> -		 * Check if GPU Reset is in progress - we need intel_ring_begin
> -		 * to work properly to reinit the hw state while the gpu is
> -		 * still marked as reset-in-progress. Handle this with a flag.
> -		 */
> -		if (!error->reload_in_reset)
> -			return -EAGAIN;
> +		return -EAGAIN;
>  	}
>  
>  	return 0;
> @@ -1223,7 +1214,6 @@ static int __i915_spin_request(struct drm_i915_gem_request *req, int state)
>  /**
>   * __i915_wait_request - wait until execution of request has finished
>   * @req: duh!
> - * @reset_counter: reset sequence associated with the given request
>   * @interruptible: do an interruptible wait (normally yes)
>   * @timeout: in - how long to wait (NULL forever); out - how much time remaining
>   *
> @@ -1238,7 +1228,6 @@ static int __i915_spin_request(struct drm_i915_gem_request *req, int state)
>   * errno with remaining time filled in timeout argument.
>   */
>  int __i915_wait_request(struct drm_i915_gem_request *req,
> -			unsigned reset_counter,
>  			bool interruptible,
>  			s64 *timeout,
>  			struct intel_rps_client *rps)
> @@ -1289,12 +1278,12 @@ int __i915_wait_request(struct drm_i915_gem_request *req,
>  
>  		/* We need to check whether any gpu reset happened in between
>  		 * the caller grabbing the seqno and now ... */
> -		if (reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter)) {
> -			/* ... but upgrade the -EAGAIN to an -EIO if the gpu
> -			 * is truely gone. */
> -			ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
> -			if (ret == 0)
> -				ret = -EAGAIN;
> +		if (req->reset_counter != i915_reset_counter(&dev_priv->gpu_error)) {
> +			/* As we do not requeue the request over a GPU reset,
> +			 * if one does occur we know that the request is
> +			 * effectively complete.
> +			 */
> +			ret = 0;
>  			break;
>  		}
>  
> @@ -1462,13 +1451,7 @@ i915_wait_request(struct drm_i915_gem_request *req)
>  
>  	BUG_ON(!mutex_is_locked(&dev->struct_mutex));
>  
> -	ret = i915_gem_check_wedge(&dev_priv->gpu_error, interruptible);
> -	if (ret)
> -		return ret;
> -
> -	ret = __i915_wait_request(req,
> -				  atomic_read(&dev_priv->gpu_error.reset_counter),
> -				  interruptible, NULL, NULL);
> +	ret = __i915_wait_request(req, interruptible, NULL, NULL);
>  	if (ret)
>  		return ret;
>  
> @@ -1543,7 +1526,6 @@ i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj,
>  	struct drm_device *dev = obj->base.dev;
>  	struct drm_i915_private *dev_priv = dev->dev_private;
>  	struct drm_i915_gem_request *requests[I915_NUM_RINGS];
> -	unsigned reset_counter;
>  	int ret, i, n = 0;
>  
>  	BUG_ON(!mutex_is_locked(&dev->struct_mutex));
> @@ -1552,12 +1534,6 @@ i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj,
>  	if (!obj->active)
>  		return 0;
>  
> -	ret = i915_gem_check_wedge(&dev_priv->gpu_error, true);
> -	if (ret)
> -		return ret;
> -
> -	reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
> -
>  	if (readonly) {
>  		struct drm_i915_gem_request *req;
>  
> @@ -1579,9 +1555,9 @@ i915_gem_object_wait_rendering__nonblocking(struct drm_i915_gem_object *obj,
>  	}
>  
>  	mutex_unlock(&dev->struct_mutex);
> +	ret = 0;
>  	for (i = 0; ret == 0 && i < n; i++)
> -		ret = __i915_wait_request(requests[i], reset_counter, true,
> -					  NULL, rps);
> +		ret = __i915_wait_request(requests[i], true, NULL, rps);
>  	mutex_lock(&dev->struct_mutex);
>  
>  	for (i = 0; i < n; i++) {
> @@ -2685,6 +2661,7 @@ int i915_gem_request_alloc(struct intel_engine_cs *ring,
>  			   struct drm_i915_gem_request **req_out)
>  {
>  	struct drm_i915_private *dev_priv = to_i915(ring->dev);
> +	unsigned reset_counter = i915_reset_counter(&dev_priv->gpu_error);
>  	struct drm_i915_gem_request *req;
>  	int ret;
>  
> @@ -2693,6 +2670,10 @@ int i915_gem_request_alloc(struct intel_engine_cs *ring,
>  
>  	*req_out = NULL;
>  
> +	ret = i915_gem_check_wedge(reset_counter, dev_priv->mm.interruptible);
> +	if (ret)
> +		return ret;
> +
>  	req = kmem_cache_zalloc(dev_priv->requests, GFP_KERNEL);
>  	if (req == NULL)
>  		return -ENOMEM;
> @@ -2704,6 +2685,7 @@ int i915_gem_request_alloc(struct intel_engine_cs *ring,
>  	kref_init(&req->ref);
>  	req->i915 = dev_priv;
>  	req->ring = ring;
> +	req->reset_counter = reset_counter;
>  	req->ctx  = ctx;
>  	i915_gem_context_reference(req->ctx);
>  
> @@ -3064,11 +3046,9 @@ retire:
>  int
>  i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
>  {
> -	struct drm_i915_private *dev_priv = dev->dev_private;
>  	struct drm_i915_gem_wait *args = data;
>  	struct drm_i915_gem_object *obj;
>  	struct drm_i915_gem_request *req[I915_NUM_RINGS];
> -	unsigned reset_counter;
>  	int i, n = 0;
>  	int ret;
>  
> @@ -3102,7 +3082,6 @@ i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
>  	}
>  
>  	drm_gem_object_unreference(&obj->base);
> -	reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
>  
>  	for (i = 0; i < I915_NUM_RINGS; i++) {
>  		if (obj->last_read_req[i] == NULL)
> @@ -3115,11 +3094,23 @@ i915_gem_wait_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
>  
>  	for (i = 0; i < n; i++) {
>  		if (ret == 0)
> -			ret = __i915_wait_request(req[i], reset_counter, true,
> +			ret = __i915_wait_request(req[i], true,
>  						  args->timeout_ns > 0 ? &args->timeout_ns : NULL,
>  						  file->driver_priv);
> +
> +		/* If the GPU hung before this, report it. Ideally we only
> +		 * report if this request cannot be completed. Currently
> +		 * when we don't mark the guilty party and abort all
> +		 * requests on reset, so just mark all as EIO.
> +		 */
> +		if (ret == 0 &&
> +		    req[i]->reset_counter != i915_reset_counter(&req[i]->i915->gpu_error))
> +			ret = -EIO;
> +
>  		i915_gem_request_unreference__unlocked(req[i]);
>  	}
> +
> +
>  	return ret;
>  
>  out:
> @@ -3147,7 +3138,6 @@ __i915_gem_object_sync(struct drm_i915_gem_object *obj,
>  	if (!i915_semaphore_is_enabled(obj->base.dev)) {
>  		struct drm_i915_private *i915 = to_i915(obj->base.dev);
>  		ret = __i915_wait_request(from_req,
> -					  atomic_read(&i915->gpu_error.reset_counter),
>  					  i915->mm.interruptible,
>  					  NULL,
>  					  &i915->rps.semaphores);
> @@ -4076,14 +4066,15 @@ i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
>  	struct drm_i915_file_private *file_priv = file->driver_priv;
>  	unsigned long recent_enough = jiffies - DRM_I915_THROTTLE_JIFFIES;
>  	struct drm_i915_gem_request *request, *target = NULL;
> -	unsigned reset_counter;
>  	int ret;
>  
>  	ret = i915_gem_wait_for_error(&dev_priv->gpu_error);
>  	if (ret)
>  		return ret;
>  
> -	ret = i915_gem_check_wedge(&dev_priv->gpu_error, false);
> +	/* ABI: return -EIO if wedged */
> +	ret = i915_gem_check_wedge(i915_reset_counter(&dev_priv->gpu_error),
> +				   false);
>  	if (ret)
>  		return ret;
>  
> @@ -4101,7 +4092,6 @@ i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
>  
>  		target = request;
>  	}
> -	reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
>  	if (target)
>  		i915_gem_request_reference(target);
>  	spin_unlock(&file_priv->mm.lock);
> @@ -4109,7 +4099,7 @@ i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file)
>  	if (target == NULL)
>  		return 0;
>  
> -	ret = __i915_wait_request(target, reset_counter, true, NULL, NULL);
> +	ret = __i915_wait_request(target, true, NULL, NULL);
>  	if (ret == 0)
>  		queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, 0);
>  
> diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c
> index e88d692583a5..3b2a8fbe0392 100644
> --- a/drivers/gpu/drm/i915/i915_irq.c
> +++ b/drivers/gpu/drm/i915/i915_irq.c
> @@ -2434,7 +2434,6 @@ static void i915_error_wake_up(struct drm_i915_private *dev_priv,
>  static void i915_reset_and_wakeup(struct drm_device *dev)
>  {
>  	struct drm_i915_private *dev_priv = to_i915(dev);
> -	struct i915_gpu_error *error = &dev_priv->gpu_error;
>  	char *error_event[] = { I915_ERROR_UEVENT "=1", NULL };
>  	char *reset_event[] = { I915_RESET_UEVENT "=1", NULL };
>  	char *reset_done_event[] = { I915_ERROR_UEVENT "=0", NULL };
> @@ -2452,7 +2451,7 @@ static void i915_reset_and_wakeup(struct drm_device *dev)
>  	 * the reset in-progress bit is only ever set by code outside of this
>  	 * work we don't need to worry about any other races.
>  	 */
> -	if (i915_reset_in_progress(error) && !i915_terminally_wedged(error)) {
> +	if (i915_reset_in_progress(&dev_priv->gpu_error)) {
>  		DRM_DEBUG_DRIVER("resetting chip\n");
>  		kobject_uevent_env(&dev->primary->kdev->kobj, KOBJ_CHANGE,
>  				   reset_event);
> @@ -2480,25 +2479,9 @@ static void i915_reset_and_wakeup(struct drm_device *dev)
>  
>  		intel_runtime_pm_put(dev_priv);
>  
> -		if (ret == 0) {
> -			/*
> -			 * After all the gem state is reset, increment the reset
> -			 * counter and wake up everyone waiting for the reset to
> -			 * complete.
> -			 *
> -			 * Since unlock operations are a one-sided barrier only,
> -			 * we need to insert a barrier here to order any seqno
> -			 * updates before
> -			 * the counter increment.
> -			 */
> -			smp_mb__before_atomic();
> -			atomic_inc(&dev_priv->gpu_error.reset_counter);
> -
> +		if (ret == 0)
>  			kobject_uevent_env(&dev->primary->kdev->kobj,
>  					   KOBJ_CHANGE, reset_done_event);
> -		} else {
> -			atomic_or(I915_WEDGED, &error->reset_counter);
> -		}
>  
>  		/*
>  		 * Note: The wake_up also serves as a memory barrier so that
> diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c
> index 0743337ac9e8..d77a17bb94d0 100644
> --- a/drivers/gpu/drm/i915/intel_display.c
> +++ b/drivers/gpu/drm/i915/intel_display.c
> @@ -3289,8 +3289,7 @@ static bool intel_crtc_has_pending_flip(struct drm_crtc *crtc)
>  	struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
>  	bool pending;
>  
> -	if (i915_reset_in_progress(&dev_priv->gpu_error) ||
> -	    intel_crtc->reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter))
> +	if (intel_crtc->reset_counter != i915_reset_counter(&dev_priv->gpu_error))
>  		return false;
>  
>  	spin_lock_irq(&dev->event_lock);
> @@ -10879,8 +10878,7 @@ static bool page_flip_finished(struct intel_crtc *crtc)
>  	struct drm_device *dev = crtc->base.dev;
>  	struct drm_i915_private *dev_priv = dev->dev_private;
>  
> -	if (i915_reset_in_progress(&dev_priv->gpu_error) ||
> -	    crtc->reset_counter != atomic_read(&dev_priv->gpu_error.reset_counter))
> +	if (crtc->reset_counter != i915_reset_counter(&dev_priv->gpu_error))
>  		return true;
>  
>  	/*
> @@ -11322,7 +11320,6 @@ static void intel_mmio_flip_work_func(struct work_struct *work)
>  
>  	if (mmio_flip->req) {
>  		WARN_ON(__i915_wait_request(mmio_flip->req,
> -					    mmio_flip->crtc->reset_counter,
>  					    false, NULL,
>  					    &mmio_flip->i915->rps.mmioflips));
>  		i915_gem_request_unreference__unlocked(mmio_flip->req);
> @@ -13312,9 +13309,6 @@ static int intel_atomic_prepare_commit(struct drm_device *dev,
>  
>  	ret = drm_atomic_helper_prepare_planes(dev, state);
>  	if (!ret && !async && !i915_reset_in_progress(&dev_priv->gpu_error)) {
> -		u32 reset_counter;
> -
> -		reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
>  		mutex_unlock(&dev->struct_mutex);
>  
>  		for_each_plane_in_state(state, plane, plane_state, i) {
> @@ -13325,8 +13319,7 @@ static int intel_atomic_prepare_commit(struct drm_device *dev,
>  				continue;
>  
>  			ret = __i915_wait_request(intel_plane_state->wait_req,
> -						  reset_counter, true,
> -						  NULL, NULL);
> +						  true, NULL, NULL);
>  
>  			/* Swallow -EIO errors to allow updates during hw lockup. */
>  			if (ret == -EIO)
> diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
> index 4ebafab53f30..b40ffc3607e7 100644
> --- a/drivers/gpu/drm/i915/intel_lrc.c
> +++ b/drivers/gpu/drm/i915/intel_lrc.c
> @@ -711,6 +711,14 @@ static int logical_ring_wait_for_space(struct drm_i915_gem_request *req,
>  	if (ret)
>  		return ret;
>  
> +	/* If the request was completed due to a GPU hang, we want to
> +	 * error out before we continue to emit more commands to the GPU.
> +	 */
> +	ret = i915_gem_check_wedge(i915_reset_counter(&req->i915->gpu_error),
> +				   req->i915->mm.interruptible);
> +	if (ret)
> +		return ret;
> +
>  	ringbuf->space = space;
>  	return 0;
>  }
> @@ -825,11 +833,6 @@ int intel_logical_ring_begin(struct drm_i915_gem_request *req, int num_dwords)
>  	WARN_ON(req == NULL);
>  	dev_priv = req->ring->dev->dev_private;
>  
> -	ret = i915_gem_check_wedge(&dev_priv->gpu_error,
> -				   dev_priv->mm.interruptible);
> -	if (ret)
> -		return ret;
> -
>  	ret = logical_ring_prepare(req, num_dwords * sizeof(uint32_t));
>  	if (ret)
>  		return ret;
> diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.c b/drivers/gpu/drm/i915/intel_ringbuffer.c
> index 57d78f264b53..511efe556d73 100644
> --- a/drivers/gpu/drm/i915/intel_ringbuffer.c
> +++ b/drivers/gpu/drm/i915/intel_ringbuffer.c
> @@ -2254,6 +2254,14 @@ static int ring_wait_for_space(struct intel_engine_cs *ring, int n)
>  	if (ret)
>  		return ret;
>  
> +	/* If the request was completed due to a GPU hang, we want to
> +	 * error out before we continue to emit more commands to the GPU.
> +	 */
> +	ret = i915_gem_check_wedge(i915_reset_counter(&to_i915(ring->dev)->gpu_error),
> +				   to_i915(ring->dev)->mm.interruptible);
> +	if (ret)
> +		return ret;
> +
>  	ringbuf->space = space;
>  	return 0;
>  }
> @@ -2286,7 +2294,6 @@ int intel_ring_idle(struct intel_engine_cs *ring)
>  
>  	/* Make sure we do not trigger any retires */
>  	return __i915_wait_request(req,
> -				   atomic_read(&to_i915(ring->dev)->gpu_error.reset_counter),
>  				   to_i915(ring->dev)->mm.interruptible,
>  				   NULL, NULL);
>  }
> @@ -2417,11 +2424,6 @@ int intel_ring_begin(struct drm_i915_gem_request *req,
>  	ring = req->ring;
>  	dev_priv = ring->dev->dev_private;
>  
> -	ret = i915_gem_check_wedge(&dev_priv->gpu_error,
> -				   dev_priv->mm.interruptible);
> -	if (ret)
> -		return ret;
> -
>  	ret = __intel_ring_prepare(ring, num_dwords * sizeof(uint32_t));
>  	if (ret)
>  		return ret;
> -- 
> 2.6.2
> 
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx@lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/intel-gfx
On Tue, Dec 01, 2015 at 09:31:31AM +0100, Daniel Vetter wrote:
> On Sun, Nov 29, 2015 at 08:48:02AM +0000, Chris Wilson wrote:
> > Instead of querying the reset counter before every access to the ring,
> > query it the first time we touch the ring, and do a final compare when
> > submitting the request. For correctness, we need to then sanitize how
> > the reset_counter is incremented to prevent broken submission and
> > waiting across resets, in the process fixing the persistent -EIO we
> > still see today on failed waits.
> > 
> > v2: Rebase
> > v3: Now with added testcase
> > v4: Rebase
> > 
> > Testcase: igt/gem_eio
> > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> 
> Imo you really can't hide an ABI change in a refactor/optimize patch like
> this ... I really think moving check_wedge out of wait_request must be
> it's own little patch.

Urm, that is the essential part of the patch in this series.
-Chris
On Tue, Dec 01, 2015 at 08:47:44AM +0000, Chris Wilson wrote:
> On Tue, Dec 01, 2015 at 09:31:31AM +0100, Daniel Vetter wrote:
> > On Sun, Nov 29, 2015 at 08:48:02AM +0000, Chris Wilson wrote:
> > > Instead of querying the reset counter before every access to the ring,
> > > query it the first time we touch the ring, and do a final compare when
> > > submitting the request. For correctness, we need to then sanitize how
> > > the reset_counter is incremented to prevent broken submission and
> > > waiting across resets, in the process fixing the persistent -EIO we
> > > still see today on failed waits.
> > > 
> > > v2: Rebase
> > > v3: Now with added testcase
> > > v4: Rebase
> > > 
> > > Testcase: igt/gem_eio
> > > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> > 
> > Imo you really can't hide an ABI change in a refactor/optimize patch like
> > this ... I really think moving check_wedge out of wait_request must be
> > it's own little patch.
> 
> Urm, that is the essential part of the patch in this series.

Ok, no it's not. The key is having the request->reset_counter. I was
thinking that the wakeup logic had to be the same in both - which is
true, but that doesn't restrict how the wakeup is then propagated from
i915_wait_request.

Getting the wakeup robust though is the meat of the patch.
-Chris