drm/i915: Cancel all ready but queued requests when wedging

When wedging the hw, we want to mark all in-flight requests as -EIO.
This is made slightly more complex by execlists who store the ready but
not yet submitted-to-hw requests on a private queue (an rbtree
priolist). Call into execlists to cancel not only the ELSP tracking for
the submitted requests, but also the queue of unsubmitted requests.

v2: Move the majority of engine_set_wedged to the backends (both legacy
ringbuffer and execlists handling their own lists).
Reported-by: Michał Winiarski <michal.winiarski@intel.com>
Testcase: igt/gem_eio/in-flight-contexts
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Michał Winiarski <michal.winiarski@intel.com>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20170915173100.26470-1-chris@chris-wilson.co.ukReviewed-by: Michał Winiarski <michal.winiarski@intel.com>

drivers/gpu/drm/i915/i915_gem.c

@@ -3022,9 +3022,6 @@ static void nop_submit_request(struct drm_i915_gem_request *request)
 
 static void engine_set_wedged(struct intel_engine_cs *engine)
 {
-	struct drm_i915_gem_request *request;
-	unsigned long flags;
-
 	/* We need to be sure that no thread is running the old callback as
 	 * we install the nop handler (otherwise we would submit a request
 	 * to hardware that will never complete). In order to prevent this
@@ -3034,40 +3031,7 @@ static void engine_set_wedged(struct intel_engine_cs *engine)
 	engine->submit_request = nop_submit_request;
 
 	/* Mark all executing requests as skipped */
-	spin_lock_irqsave(&engine->timeline->lock, flags);
-	list_for_each_entry(request, &engine->timeline->requests, link)
-		if (!i915_gem_request_completed(request))
-			dma_fence_set_error(&request->fence, -EIO);
-	spin_unlock_irqrestore(&engine->timeline->lock, flags);
-
-	/*
-	 * Clear the execlists queue up before freeing the requests, as those
-	 * are the ones that keep the context and ringbuffer backing objects
-	 * pinned in place.
-	 */
-
-	if (i915.enable_execlists) {
-		struct execlist_port *port = engine->execlist_port;
-		unsigned long flags;
-		unsigned int n;
-
-		spin_lock_irqsave(&engine->timeline->lock, flags);
-
-		for (n = 0; n < ARRAY_SIZE(engine->execlist_port); n++)
-			i915_gem_request_put(port_request(&port[n]));
-		memset(engine->execlist_port, 0, sizeof(engine->execlist_port));
-		engine->execlist_queue = RB_ROOT;
-		engine->execlist_first = NULL;
-
-		spin_unlock_irqrestore(&engine->timeline->lock, flags);
-
-		/* The port is checked prior to scheduling a tasklet, but
-		 * just in case we have suspended the tasklet to do the
-		 * wedging make sure that when it wakes, it decides there
-		 * is no work to do by clearing the irq_posted bit.
-		 */
-		clear_bit(ENGINE_IRQ_EXECLIST, &engine->irq_posted);
-	}
+	engine->cancel_requests(engine);
 
 	/* Mark all pending requests as complete so that any concurrent
 	 * (lockless) lookup doesn't try and wait upon the request as we

drivers/gpu/drm/i915/intel_lrc.c

@@ -506,6 +506,65 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
 		execlists_submit_ports(engine);
 }
 
+static void execlists_cancel_requests(struct intel_engine_cs *engine)
+{
+	struct execlist_port *port = engine->execlist_port;
+	struct drm_i915_gem_request *rq, *rn;
+	struct rb_node *rb;
+	unsigned long flags;
+	unsigned long n;
+
+	spin_lock_irqsave(&engine->timeline->lock, flags);
+
+	/* Cancel the requests on the HW and clear the ELSP tracker. */
+	for (n = 0; n < ARRAY_SIZE(engine->execlist_port); n++)
+		i915_gem_request_put(port_request(&port[n]));
+	memset(engine->execlist_port, 0, sizeof(engine->execlist_port));
+
+	/* Mark all executing requests as skipped. */
+	list_for_each_entry(rq, &engine->timeline->requests, link) {
+		GEM_BUG_ON(!rq->global_seqno);
+		if (!i915_gem_request_completed(rq))
+			dma_fence_set_error(&rq->fence, -EIO);
+	}
+
+	/* Flush the queued requests to the timeline list (for retiring). */
+	rb = engine->execlist_first;
+	while (rb) {
+		struct i915_priolist *p = rb_entry(rb, typeof(*p), node);
+
+		list_for_each_entry_safe(rq, rn, &p->requests, priotree.link) {
+			INIT_LIST_HEAD(&rq->priotree.link);
+			rq->priotree.priority = INT_MAX;
+
+			dma_fence_set_error(&rq->fence, -EIO);
+			__i915_gem_request_submit(rq);
+		}
+
+		rb = rb_next(rb);
+		rb_erase(&p->node, &engine->execlist_queue);
+		INIT_LIST_HEAD(&p->requests);
+		if (p->priority != I915_PRIORITY_NORMAL)
+			kmem_cache_free(engine->i915->priorities, p);
+	}
+
+	/* Remaining _unready_ requests will be nop'ed when submitted */
+
+	engine->execlist_queue = RB_ROOT;
+	engine->execlist_first = NULL;
+	GEM_BUG_ON(port_isset(&port[0]));
+
+	/*
+	 * The port is checked prior to scheduling a tasklet, but
+	 * just in case we have suspended the tasklet to do the
+	 * wedging make sure that when it wakes, it decides there
+	 * is no work to do by clearing the irq_posted bit.
+	 */
+	clear_bit(ENGINE_IRQ_EXECLIST, &engine->irq_posted);
+
+	spin_unlock_irqrestore(&engine->timeline->lock, flags);
+}
+
 static bool execlists_elsp_ready(const struct intel_engine_cs *engine)
 {
 	const struct execlist_port *port = engine->execlist_port;
@@ -1704,6 +1763,7 @@ void intel_logical_ring_cleanup(struct intel_engine_cs *engine)
 static void execlists_set_default_submission(struct intel_engine_cs *engine)
 {
 	engine->submit_request = execlists_submit_request;
+	engine->cancel_requests = execlists_cancel_requests;
 	engine->schedule = execlists_schedule;
 	engine->irq_tasklet.func = intel_lrc_irq_handler;
 }

drivers/gpu/drm/i915/intel_ringbuffer.c

@@ -782,6 +782,24 @@ static u32 *gen6_signal(struct drm_i915_gem_request *req, u32 *cs)
 	return cs;
 }
 
+static void cancel_requests(struct intel_engine_cs *engine)
+{
+	struct drm_i915_gem_request *request;
+	unsigned long flags;
+
+	spin_lock_irqsave(&engine->timeline->lock, flags);
+
+	/* Mark all submitted requests as skipped. */
+	list_for_each_entry(request, &engine->timeline->requests, link) {
+		GEM_BUG_ON(!request->global_seqno);
+		if (!i915_gem_request_completed(request))
+			dma_fence_set_error(&request->fence, -EIO);
+	}
+	/* Remaining _unready_ requests will be nop'ed when submitted */
+
+	spin_unlock_irqrestore(&engine->timeline->lock, flags);
+}
+
 static void i9xx_submit_request(struct drm_i915_gem_request *request)
 {
 	struct drm_i915_private *dev_priv = request->i915;
@@ -1996,11 +2014,13 @@ static void intel_ring_init_irq(struct drm_i915_private *dev_priv,
 static void i9xx_set_default_submission(struct intel_engine_cs *engine)
 {
 	engine->submit_request = i9xx_submit_request;
+	engine->cancel_requests = cancel_requests;
 }
 
 static void gen6_bsd_set_default_submission(struct intel_engine_cs *engine)
 {
 	engine->submit_request = gen6_bsd_submit_request;
+	engine->cancel_requests = cancel_requests;
 }
 
 static void intel_ring_default_vfuncs(struct drm_i915_private *dev_priv,

drivers/gpu/drm/i915/intel_ringbuffer.h

@@ -306,6 +306,14 @@ struct intel_engine_cs {
 	void		(*schedule)(struct drm_i915_gem_request *request,
 				    int priority);
 
+	/*
+	 * Cancel all requests on the hardware, or queued for execution.
+	 * This should only cancel the ready requests that have been
+	 * submitted to the engine (via the engine->submit_request callback).
+	 * This is called when marking the device as wedged.
+	 */
+	void		(*cancel_requests)(struct intel_engine_cs *engine);
+
 	/* Some chipsets are not quite as coherent as advertised and need
 	 * an expensive kick to force a true read of the up-to-date seqno.
 	 * However, the up-to-date seqno is not always required and the last