drm/i915: Interrupt driven fences

The intended usage model for struct fence is that the signalled status
should be set on demand rather than polled. That is, there should not
be a need for a 'signaled' function to be called everytime the status
is queried. Instead, 'something' should be done to enable a signal
callback from the hardware which will update the state directly. In
the case of requests, this is the seqno update interrupt. The idea is
that this callback will only be enabled on demand when something
actually tries to wait on the fence.

This change removes the polling test and replaces it with the callback
scheme. Each fence is added to a 'please poke me' list at the start of
i915_add_request(). The interrupt handler then scans through the 'poke
me' list when a new seqno pops out and signals any matching
fence/request. The fence is then removed from the list so the entire
request stack does not need to be scanned every time. Note that the
fence is added to the list before the commands to generate the seqno
interrupt are added to the ring. Thus the sequence is guaranteed to be
race free if the interrupt is already enabled.

Note that the interrupt is only enabled on demand (i.e. when
__wait_request() is called). Thus there is still a potential race when
enabling the interrupt as the request may already have completed.
However, this is simply solved by calling the interrupt processing
code immediately after enabling the interrupt and thereby checking for
already completed requests.

Lastly, the ring clean up code has the possibility to cancel
outstanding requests (e.g. because TDR has reset the ring). These
requests will never get signalled and so must be removed from the
signal list manually. This is done by setting a 'cancelled' flag and
then calling the regular notify/retire code path rather than
attempting to duplicate the list manipulatation and clean up code in
multiple places. This also avoid any race condition where the
cancellation request might occur after/during the completion interrupt
actually arriving.

v2: Updated to take advantage of the request unreference no longer
requiring the mutex lock.

v3: Move the signal list processing around to prevent unsubmitted
requests being added to the list. This was occurring on Android
because the native sync implementation calls the
fence->enable_signalling API immediately on fence creation.

Updated after review comments by Tvrtko Ursulin. Renamed list nodes to
'link' instead of 'list'. Added support for returning an error code on
a cancelled fence. Update list processing to be more efficient/safer
with respect to spinlocks.

v5: Made i915_gem_request_submit a static as it is only ever called
from one place.

Fixed up the low latency wait optimisation. The time delay between the
seqno value being to memory and the drive's ISR running can be
significant, at least for the wait request micro-benchmark. This can
be greatly improved by explicitly checking for seqno updates in the
pre-wait busy poll loop. Also added some documentation comments to the
busy poll code.

Fixed up support for the faking of lost interrupts
(test_irq_rings/missed_irq_rings). That is, there is an IGT test that
tells the driver to loose interrupts deliberately and then check that
everything still works as expected (albeit much slower).

Updates from review comments: use non IRQ-save spinlocking, early exit
on WARN and improved comments (Tvrtko Ursulin).

v6: Updated to newer nigthly and resolved conflicts around the
wait_request busy spin optimisation. Also fixed a race condition
between this early exit path and the regular completion path.

v7: Updated to newer nightly - lots of ring -> engine renaming plus an
interface change on get_seqno(). Also added a list_empty() check
before acquring spinlocks and doing list processing.

For: VIZ-5190
Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>

6 files changed, 241 insertions, 24 deletions

diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index 81324ba2ebf7..9519b11c5fb7 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -2255,7 +2255,12 @@ void i915_gem_track_fb(struct drm_i915_gem_object *old,
 struct drm_i915_gem_request {
 	/** Underlying object for implementing the signal/wait stuff. */
 	struct fence fence;
+	struct list_head signal_link;
+	struct list_head unsignal_link;
 	struct list_head delayed_free_link;
+	bool cancelled;
+	bool irq_enabled;
+	bool signal_requested;
 
 	/** On Which ring this request was generated */
 	struct drm_i915_private *i915;
@@ -2341,6 +2346,9 @@ struct drm_i915_gem_request * __must_check
 i915_gem_request_alloc(struct intel_engine_cs *engine,
 		       struct intel_context *ctx);
 void i915_gem_request_cancel(struct drm_i915_gem_request *req);
+void i915_gem_request_enable_interrupt(struct drm_i915_gem_request *req,
+				       bool fence_locked);
+void i915_gem_request_notify(struct intel_engine_cs *ring, bool fence_locked);
 
 int i915_create_fence_timeline(struct drm_device *dev,
 			       struct intel_context *ctx,
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index 67eeefbefa7d..48fbd02480a9 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -40,6 +40,8 @@
 
 #define RQ_BUG_ON(expr)
 
+static void i915_gem_request_submit(struct drm_i915_gem_request *req);
+
 static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj);
 static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj);
 static void
@@ -1253,9 +1255,8 @@ int __i915_wait_request(struct drm_i915_gem_request *req,
 	struct intel_engine_cs *engine = i915_gem_request_get_engine(req);
 	struct drm_device *dev = engine->dev;
 	struct drm_i915_private *dev_priv = dev->dev_private;
-	const bool irq_test_in_progress =
-		ACCESS_ONCE(dev_priv->gpu_error.test_irq_rings) & intel_engine_flag(engine);
 	int state = interruptible ? TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE;
+	uint32_t seqno;
 	DEFINE_WAIT(wait);
 	unsigned long timeout_expire;
 	s64 before = 0; /* Only to silence a compiler warning. */
@@ -1263,9 +1264,6 @@ int __i915_wait_request(struct drm_i915_gem_request *req,
 
 	WARN(!intel_irqs_enabled(dev_priv), "IRQs disabled");
 
-	if (list_empty(&req->list))
-		return 0;
-
 	if (i915_gem_request_completed(req))
 		return 0;
 
@@ -1291,15 +1289,17 @@ int __i915_wait_request(struct drm_i915_gem_request *req,
 	trace_i915_gem_request_wait_begin(req);
 
 	/* Optimistic spin for the next jiffie before touching IRQs */
-	ret = __i915_spin_request(req, state);
-	if (ret == 0)
-		goto out;
-
-	if (!irq_test_in_progress && WARN_ON(!engine->irq_get(engine))) {
-		ret = -ENODEV;
-		goto out;
+	if (req->seqno) {
+		ret = __i915_spin_request(req, state);
+		if (ret == 0)
+			goto out;
 	}
 
+	/*
+	 * Enable interrupt completion of the request.
+	 */
+	fence_enable_sw_signaling(&req->fence);
+
 	for (;;) {
 		struct timer_list timer;
 
@@ -1321,6 +1321,21 @@ int __i915_wait_request(struct drm_i915_gem_request *req,
 			break;
 		}
 
+		if (req->seqno) {
+			/*
+			 * There is quite a lot of latency in the user interrupt
+			 * path. So do an explicit seqno check and potentially
+			 * remove all that delay.
+			 */
+			if (req->engine->irq_seqno_barrier)
+				req->engine->irq_seqno_barrier(req->engine);
+			seqno = engine->get_seqno(engine);
+			if (i915_seqno_passed(seqno, req->seqno)) {
+				ret = 0;
+				break;
+			}
+		}
+
 		if (signal_pending_state(state, current)) {
 			ret = -ERESTARTSYS;
 			break;
@@ -1347,14 +1362,32 @@ int __i915_wait_request(struct drm_i915_gem_request *req,
 			destroy_timer_on_stack(&timer);
 		}
 	}
-	if (!irq_test_in_progress)
-		engine->irq_put(engine);
 
 	finish_wait(&engine->irq_queue, &wait);
 
 out:
 	trace_i915_gem_request_wait_end(req);
 
+	if ((ret == 0) && (req->seqno)) {
+		if (req->engine->irq_seqno_barrier)
+			req->engine->irq_seqno_barrier(req->engine);
+		seqno = engine->get_seqno(engine);
+		if (i915_seqno_passed(seqno, req->seqno) &&
+		    !i915_gem_request_completed(req)) {
+			/*
+			 * Make sure the request is marked as completed before
+			 * returning. NB: Need to acquire the spinlock around
+			 * the whole call to avoid a race condition with the
+			 * interrupt handler is running concurrently and could
+			 * cause this invocation to early exit even though the
+			 * request has not actually been fully processed yet.
+			 */
+			spin_lock_irq(&req->engine->fence_lock);
+			i915_gem_request_notify(req->engine, true);
+			spin_unlock_irq(&req->engine->fence_lock);
+		}
+	}
+
 	if (timeout) {
 		s64 tres = *timeout - (ktime_get_raw_ns() - before);
 
@@ -1432,6 +1465,22 @@ static void i915_gem_request_retire(struct drm_i915_gem_request *request)
 	list_del_init(&request->list);
 	i915_gem_request_remove_from_client(request);
 
+	/*
+	 * In case the request is still in the signal pending list,
+	 * e.g. due to being cancelled by TDR, preemption, etc.
+	 */
+	if (!list_empty(&request->signal_link)) {
+		/*
+		 * The request must be marked as cancelled and the underlying
+		 * fence as failed. NB: There is no explicit fence fail API,
+		 * there is only a manual poke and signal.
+		 */
+		request->cancelled = true;
+		/* How to propagate to any associated sync_fence??? */
+		request->fence.status = -EIO;
+		fence_signal_locked(&request->fence);
+	}
+
 	i915_gem_request_unreference(request);
 }
 
@@ -2660,6 +2709,12 @@ void __i915_add_request(struct drm_i915_gem_request *request,
 	 */
 	request->postfix = intel_ring_get_tail(ringbuf);
 
+	/*
+	 * Add the fence to the pending list before emitting the commands to
+	 * generate a seqno notification interrupt.
+	 */
+	i915_gem_request_submit(request);
+
 	if (i915.enable_execlists)
 		ret = engine->emit_request(request);
 	else {
@@ -2760,28 +2815,142 @@ static void i915_gem_request_free(struct drm_i915_gem_request *req)
 		i915_gem_context_unreference(ctx);
 	}
 
+	if (req->irq_enabled)
+		req->engine->irq_put(req->engine);
+
 	kmem_cache_free(req->i915->requests, req);
 }
 
-static bool i915_gem_request_enable_signaling(struct fence *req_fence)
+/*
+ * The request is about to be submitted to the hardware so add the fence to
+ * the list of signalable fences.
+ *
+ * NB: This does not necessarily enable interrupts yet. That only occurs on
+ * demand when the request is actually waited on. However, adding it to the
+ * list early ensures that there is no race condition where the interrupt
+ * could pop out prematurely and thus be completely lost. The race is merely
+ * that the interrupt must be manually checked for after being enabled.
+ */
+static void i915_gem_request_submit(struct drm_i915_gem_request *req)
 {
-	/* Interrupt driven fences are not implemented yet.*/
-	WARN(true, "This should not be called!");
-	return true;
+	/*
+	 * Always enable signal processing for the request's fence object
+	 * before that request is submitted to the hardware. Thus there is no
+	 * race condition whereby the interrupt could pop out before the
+	 * request has been added to the signal list. Hence no need to check
+	 * for completion, undo the list add and return false.
+	 */
+	i915_gem_request_reference(req);
+	spin_lock_irq(&req->engine->fence_lock);
+	WARN_ON(!list_empty(&req->signal_link));
+	list_add_tail(&req->signal_link, &req->engine->fence_signal_list);
+	spin_unlock_irq(&req->engine->fence_lock);
+
+	/*
+	 * NB: Interrupts are only enabled on demand. Thus there is still a
+	 * race where the request could complete before the interrupt has
+	 * been enabled. Thus care must be taken at that point.
+	 */
+
+	/* Have interrupts already been requested? */
+	if (req->signal_requested)
+		i915_gem_request_enable_interrupt(req, false);
+}
+
+/*
+ * The request is being actively waited on, so enable interrupt based
+ * completion signalling.
+ */
+void i915_gem_request_enable_interrupt(struct drm_i915_gem_request *req,
+				       bool fence_locked)
+{
+	struct drm_i915_private *dev_priv = to_i915(req->engine->dev);
+	const bool irq_test_in_progress =
+		ACCESS_ONCE(dev_priv->gpu_error.test_irq_rings) &
+						intel_engine_flag(req->engine);
+
+	if (req->irq_enabled)
+		return;
+
+	if (irq_test_in_progress)
+		return;
+
+	if (!WARN_ON(!req->engine->irq_get(req->engine)))
+		req->irq_enabled = true;
+
+	/*
+	 * Because the interrupt is only enabled on demand, there is a race
+	 * where the interrupt can fire before anyone is looking for it. So
+	 * do an explicit check for missed interrupts.
+	 */
+	i915_gem_request_notify(req->engine, fence_locked);
 }
 
-static bool i915_gem_request_is_completed(struct fence *req_fence)
+static bool i915_gem_request_enable_signaling(struct fence *req_fence)
 {
 	struct drm_i915_gem_request *req = container_of(req_fence,
 						 typeof(*req), fence);
+
+	/*
+	 * No need to actually enable interrupt based processing until the
+	 * request has been submitted to the hardware. At which point
+	 * 'i915_gem_request_submit()' is called. So only really enable
+	 * signalling in there. Just set a flag to say that interrupts are
+	 * wanted when the request is eventually submitted. On the other hand
+	 * if the request has already been submitted then interrupts do need
+	 * to be enabled now.
+	 */
+
+	req->signal_requested = true;
+
+	if (!list_empty(&req->signal_link))
+		i915_gem_request_enable_interrupt(req, true);
+
+	return true;
+}
+
+void i915_gem_request_notify(struct intel_engine_cs *engine, bool fence_locked)
+{
+	struct drm_i915_gem_request *req, *req_next;
+	unsigned long flags;
 	u32 seqno;
 
-/*	if (!lazy_coherency && req->engine->irq_seqno_barrier)
-		req->engine->irq_seqno_barrier(req->engine);*/
+	if (list_empty(&engine->fence_signal_list))
+		return;
 
-	seqno = req->engine->get_seqno(req->engine);
+	if (!fence_locked)
+		spin_lock_irqsave(&engine->fence_lock, flags);
+
+	if (engine->irq_seqno_barrier)
+		engine->irq_seqno_barrier(engine);
+	seqno = engine->get_seqno(engine);
+
+	list_for_each_entry_safe(req, req_next, &engine->fence_signal_list, signal_link) {
+		if (!req->cancelled) {
+			if (!i915_seqno_passed(seqno, req->seqno))
+				break;
+		}
+
+		/*
+		 * Start by removing the fence from the signal list otherwise
+		 * the retire code can run concurrently and get confused.
+		 */
+		list_del_init(&req->signal_link);
+
+		if (!req->cancelled)
+			fence_signal_locked(&req->fence);
+
+		if (req->irq_enabled) {
+			req->engine->irq_put(req->engine);
+			req->irq_enabled = false;
+		}
 
-	return i915_seqno_passed(seqno, req->seqno);
+		/* Can't unreference here because that might grab fence_lock */
+		list_add_tail(&req->unsignal_link, &engine->fence_unsignal_list);
+	}
+
+	if (!fence_locked)
+		spin_unlock_irqrestore(&engine->fence_lock, flags);
 }
 
 static const char *i915_gem_request_get_driver_name(struct fence *req_fence)
@@ -2824,7 +2993,6 @@ static void i915_gem_request_fence_value_str(struct fence *req_fence,
 
 static const struct fence_ops i915_gem_request_fops = {
 	.enable_signaling	= i915_gem_request_enable_signaling,
-	.signaled		= i915_gem_request_is_completed,
 	.wait			= fence_default_wait,
 	.release		= i915_gem_request_release,
 	.get_driver_name	= i915_gem_request_get_driver_name,
@@ -2909,6 +3077,7 @@ __i915_gem_request_alloc(struct intel_engine_cs *engine,
 		goto err;
 	}
 
+	INIT_LIST_HEAD(&req->signal_link);
 	fence_init(&req->fence, &i915_gem_request_fops, &engine->fence_lock,
 		   ctx->engine[engine->id].fence_timeline.fence_context,
 		   i915_fence_timeline_get_next_seqno(&ctx->engine[engine->id].fence_timeline));
@@ -2971,6 +3140,11 @@ void i915_gem_request_cancel(struct drm_i915_gem_request *req)
 {
 	intel_ring_reserved_space_cancel(req->ringbuf);
 
+	req->cancelled = true;
+	/* How to propagate to any associated sync_fence??? */
+	req->fence.status = -EINVAL;
+	fence_signal_locked(&req->fence);
+
 	i915_gem_request_unreference(req);
 }
 
@@ -3059,6 +3233,13 @@ static void i915_gem_reset_engine_cleanup(struct drm_i915_private *dev_priv,
 		i915_gem_request_retire(request);
 	}
 
+	/*
+	 * Tidy up anything left over. This includes a call to
+	 * i915_gem_request_notify() which will make sure that any requests
+	 * that were on the signal pending list get also cleaned up.
+	 */
+	i915_gem_retire_requests_ring(engine);
+
 	/* Having flushed all requests from all queues, we know that all
 	 * ringbuffers must now be empty. However, since we do not reclaim
 	 * all space when retiring the request (to prevent HEADs colliding
@@ -3108,6 +3289,13 @@ i915_gem_retire_requests_ring(struct intel_engine_cs *engine)
 
 	WARN_ON(i915_verify_lists(engine->dev));
 
+	/*
+	 * If no-one has waited on a request recently then interrupts will
+	 * not have been enabled and thus no requests will ever be marked as
+	 * completed. So do an interrupt check now.
+	 */
+	i915_gem_request_notify(engine, false);
+
 	/* Retire requests first as we use it above for the early return.
 	 * If we retire requests last, we may use a later seqno and so clear
 	 * the requests lists without clearing the active list, leading to
@@ -3149,6 +3337,17 @@ i915_gem_retire_requests_ring(struct intel_engine_cs *engine)
 		i915_gem_request_assign(&engine->trace_irq_req, NULL);
 	}
 
+	/* Tidy up any requests that were recently signalled */
+	if (!list_empty(&engine->fence_unsignal_list)) {
+		spin_lock_irq(&engine->fence_lock);
+		list_splice_init(&engine->fence_unsignal_list, &list_head);
+		spin_unlock_irq(&engine->fence_lock);
+		list_for_each_entry_safe(req, req_next, &list_head, unsignal_link) {
+			list_del(&req->unsignal_link);
+			i915_gem_request_unreference(req);
+		}
+	}
+
 	/* Really free any requests that were recently unreferenced */
 	if (!list_empty(&engine->delayed_free_list)) {
 		spin_lock(&engine->delayed_free_lock);
@@ -5231,6 +5430,8 @@ init_engine_lists(struct intel_engine_cs *engine)
 {
 	INIT_LIST_HEAD(&engine->active_list);
 	INIT_LIST_HEAD(&engine->request_list);
+	INIT_LIST_HEAD(&engine->fence_signal_list);
+	INIT_LIST_HEAD(&engine->fence_unsignal_list);
 	INIT_LIST_HEAD(&engine->delayed_free_list);
 }
 
diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c
index 247d962afabb..d9390a2da896 100644
--- a/drivers/gpu/drm/i915/i915_irq.c
+++ b/drivers/gpu/drm/i915/i915_irq.c
@@ -1002,6 +1002,8 @@ static void notify_ring(struct intel_engine_cs *engine)
 	trace_i915_gem_request_notify(engine);
 	engine->user_interrupts++;
 
+	i915_gem_request_notify(engine, false);
+
 	wake_up_all(&engine->irq_queue);
 }
 
diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
index 8eb0c5ed5f12..a351a0e7a9cf 100644
--- a/drivers/gpu/drm/i915/intel_lrc.c
+++ b/drivers/gpu/drm/i915/intel_lrc.c
@@ -2133,6 +2133,8 @@ logical_ring_init(struct drm_device *dev, struct intel_engine_cs *engine)
 	engine->dev = dev;
 	INIT_LIST_HEAD(&engine->active_list);
 	INIT_LIST_HEAD(&engine->request_list);
+	INIT_LIST_HEAD(&engine->fence_signal_list);
+	INIT_LIST_HEAD(&engine->fence_unsignal_list);
 	INIT_LIST_HEAD(&engine->delayed_free_list);
 	spin_lock_init(&engine->fence_lock);
 	spin_lock_init(&engine->delayed_free_lock);
diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.c b/drivers/gpu/drm/i915/intel_ringbuffer.c
index ac922a53e6b7..5f209ba71822 100644
--- a/drivers/gpu/drm/i915/intel_ringbuffer.c
+++ b/drivers/gpu/drm/i915/intel_ringbuffer.c
@@ -2229,6 +2229,8 @@ static int intel_init_ring_buffer(struct drm_device *dev,
 	INIT_LIST_HEAD(&engine->request_list);
 	INIT_LIST_HEAD(&engine->execlist_queue);
 	INIT_LIST_HEAD(&engine->buffers);
+	INIT_LIST_HEAD(&engine->fence_signal_list);
+	INIT_LIST_HEAD(&engine->fence_unsignal_list);
 	INIT_LIST_HEAD(&engine->delayed_free_list);
 	spin_lock_init(&engine->fence_lock);
 	spin_lock_init(&engine->delayed_free_lock);
diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.h b/drivers/gpu/drm/i915/intel_ringbuffer.h
index fa07ea9afe6c..160e71e89925 100644
--- a/drivers/gpu/drm/i915/intel_ringbuffer.h
+++ b/drivers/gpu/drm/i915/intel_ringbuffer.h
@@ -357,6 +357,8 @@ struct  intel_engine_cs {
 	u32 (*get_cmd_length_mask)(u32 cmd_header);
 
 	spinlock_t fence_lock;
+	struct list_head fence_signal_list;
+	struct list_head fence_unsignal_list;
 };
 
 static inline bool