drm/i915: Replace global breadcrumbs with per-context interrupt tracking

A few years ago, see commit 688e6c725816 ("drm/i915: Slaughter the
thundering i915_wait_request herd"), the issue of handling multiple
clients waiting in parallel was brought to our attention. The
requirement was that every client should be woken immediately upon its
request being signaled, without incurring any cpu overhead.

To handle certain fragility of our hw meant that we could not do a
simple check inside the irq handler (some generations required almost
unbounded delays before we could be sure of seqno coherency) and so
request completion checking required delegation.

Before commit 688e6c725816, the solution was simple. Every client
waiting on a request would be woken on every interrupt and each would do
a heavyweight check to see if their request was complete. Commit
688e6c725816 introduced an rbtree so that only the earliest waiter on
the global timeline would woken, and would wake the next and so on.
(Along with various complications to handle requests being reordered
along the global timeline, and also a requirement for kthread to provide
a delegate for fence signaling that had no process context.)

The global rbtree depends on knowing the execution timeline (and global
seqno). Without knowing that order, we must instead check all contexts
queued to the HW to see which may have advanced. We trim that list by
only checking queued contexts that are being waited on, but still we
keep a list of all active contexts and their active signalers that we
inspect from inside the irq handler. By moving the waiters onto the fence
signal list, we can combine the client wakeup with the dma_fence
signaling (a dramatic reduction in complexity, but does require the HW
being coherent, the seqno must be visible from the cpu before the
interrupt is raised - we keep a timer backup just in case).

Having previously fixed all the issues with irq-seqno serialisation (by
inserting delays onto the GPU after each request instead of random delays
on the CPU after each interrupt), we can rely on the seqno state to
perfom direct wakeups from the interrupt handler. This allows us to
preserve our single context switch behaviour of the current routine,
with the only downside that we lose the RT priority sorting of wakeups.
In general, direct wakeup latency of multiple clients is about the same
(about 10% better in most cases) with a reduction in total CPU time spent
in the waiter (about 20-50% depending on gen). Average herd behaviour is
improved, but at the cost of not delegating wakeups on task_prio.

v2: Capture fence signaling state for error state and add comments to
warm even the most cold of hearts.
v3: Check if the request is still active before busywaiting
v4: Reduce the amount of pointer misdirection with list_for_each_safe
and using a local i915_request variable inside the loops
v5: Add a missing pluralisation to a purely informative selftest message.

References: 688e6c725816 ("drm/i915: Slaughter the thundering i915_wait_request herd")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190129205230.19056-2-chris@chris-wilson.co.uk

9 files changed, 494 insertions, 489 deletions

diff --git a/drivers/gpu/drm/i915/selftests/i915_mock_selftests.h b/drivers/gpu/drm/i915/selftests/i915_mock_selftests.h
index 4a83a1c6c406..88e5ab586337 100644
--- a/drivers/gpu/drm/i915/selftests/i915_mock_selftests.h
+++ b/drivers/gpu/drm/i915/selftests/i915_mock_selftests.h
@@ -15,7 +15,6 @@ selftest(scatterlist, scatterlist_mock_selftests)
 selftest(syncmap, i915_syncmap_mock_selftests)
 selftest(uncore, intel_uncore_mock_selftests)
 selftest(engine, intel_engine_cs_mock_selftests)
-selftest(breadcrumbs, intel_breadcrumbs_mock_selftests)
 selftest(timelines, i915_timeline_mock_selftests)
 selftest(requests, i915_request_mock_selftests)
 selftest(objects, i915_gem_object_mock_selftests)
diff --git a/drivers/gpu/drm/i915/selftests/i915_request.c b/drivers/gpu/drm/i915/selftests/i915_request.c
index 4d4b86b5fa11..6733dc5b6b4c 100644
--- a/drivers/gpu/drm/i915/selftests/i915_request.c
+++ b/drivers/gpu/drm/i915/selftests/i915_request.c
@@ -25,9 +25,12 @@
 #include <linux/prime_numbers.h>
 
 #include "../i915_selftest.h"
+#include "i915_random.h"
 #include "igt_live_test.h"
+#include "lib_sw_fence.h"
 
 #include "mock_context.h"
+#include "mock_drm.h"
 #include "mock_gem_device.h"
 
 static int igt_add_request(void *arg)
@@ -247,6 +250,254 @@ err_context_0:
 	return err;
 }
 
+struct smoketest {
+	struct intel_engine_cs *engine;
+	struct i915_gem_context **contexts;
+	atomic_long_t num_waits, num_fences;
+	int ncontexts, max_batch;
+	struct i915_request *(*request_alloc)(struct i915_gem_context *,
+					      struct intel_engine_cs *);
+};
+
+static struct i915_request *
+__mock_request_alloc(struct i915_gem_context *ctx,
+		     struct intel_engine_cs *engine)
+{
+	return mock_request(engine, ctx, 0);
+}
+
+static struct i915_request *
+__live_request_alloc(struct i915_gem_context *ctx,
+		     struct intel_engine_cs *engine)
+{
+	return i915_request_alloc(engine, ctx);
+}
+
+static int __igt_breadcrumbs_smoketest(void *arg)
+{
+	struct smoketest *t = arg;
+	struct mutex * const BKL = &t->engine->i915->drm.struct_mutex;
+	const unsigned int max_batch = min(t->ncontexts, t->max_batch) - 1;
+	const unsigned int total = 4 * t->ncontexts + 1;
+	unsigned int num_waits = 0, num_fences = 0;
+	struct i915_request **requests;
+	I915_RND_STATE(prng);
+	unsigned int *order;
+	int err = 0;
+
+	/*
+	 * A very simple test to catch the most egregious of list handling bugs.
+	 *
+	 * At its heart, we simply create oodles of requests running across
+	 * multiple kthreads and enable signaling on them, for the sole purpose
+	 * of stressing our breadcrumb handling. The only inspection we do is
+	 * that the fences were marked as signaled.
+	 */
+
+	requests = kmalloc_array(total, sizeof(*requests), GFP_KERNEL);
+	if (!requests)
+		return -ENOMEM;
+
+	order = i915_random_order(total, &prng);
+	if (!order) {
+		err = -ENOMEM;
+		goto out_requests;
+	}
+
+	while (!kthread_should_stop()) {
+		struct i915_sw_fence *submit, *wait;
+		unsigned int n, count;
+
+		submit = heap_fence_create(GFP_KERNEL);
+		if (!submit) {
+			err = -ENOMEM;
+			break;
+		}
+
+		wait = heap_fence_create(GFP_KERNEL);
+		if (!wait) {
+			i915_sw_fence_commit(submit);
+			heap_fence_put(submit);
+			err = ENOMEM;
+			break;
+		}
+
+		i915_random_reorder(order, total, &prng);
+		count = 1 + i915_prandom_u32_max_state(max_batch, &prng);
+
+		for (n = 0; n < count; n++) {
+			struct i915_gem_context *ctx =
+				t->contexts[order[n] % t->ncontexts];
+			struct i915_request *rq;
+
+			mutex_lock(BKL);
+
+			rq = t->request_alloc(ctx, t->engine);
+			if (IS_ERR(rq)) {
+				mutex_unlock(BKL);
+				err = PTR_ERR(rq);
+				count = n;
+				break;
+			}
+
+			err = i915_sw_fence_await_sw_fence_gfp(&rq->submit,
+							       submit,
+							       GFP_KERNEL);
+
+			requests[n] = i915_request_get(rq);
+			i915_request_add(rq);
+
+			mutex_unlock(BKL);
+
+			if (err >= 0)
+				err = i915_sw_fence_await_dma_fence(wait,
+								    &rq->fence,
+								    0,
+								    GFP_KERNEL);
+
+			if (err < 0) {
+				i915_request_put(rq);
+				count = n;
+				break;
+			}
+		}
+
+		i915_sw_fence_commit(submit);
+		i915_sw_fence_commit(wait);
+
+		if (!wait_event_timeout(wait->wait,
+					i915_sw_fence_done(wait),
+					HZ / 2)) {
+			struct i915_request *rq = requests[count - 1];
+
+			pr_err("waiting for %d fences (last %llx:%lld) on %s timed out!\n",
+			       count,
+			       rq->fence.context, rq->fence.seqno,
+			       t->engine->name);
+			i915_gem_set_wedged(t->engine->i915);
+			GEM_BUG_ON(!i915_request_completed(rq));
+			i915_sw_fence_wait(wait);
+			err = -EIO;
+		}
+
+		for (n = 0; n < count; n++) {
+			struct i915_request *rq = requests[n];
+
+			if (!test_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
+				      &rq->fence.flags)) {
+				pr_err("%llu:%llu was not signaled!\n",
+				       rq->fence.context, rq->fence.seqno);
+				err = -EINVAL;
+			}
+
+			i915_request_put(rq);
+		}
+
+		heap_fence_put(wait);
+		heap_fence_put(submit);
+
+		if (err < 0)
+			break;
+
+		num_fences += count;
+		num_waits++;
+
+		cond_resched();
+	}
+
+	atomic_long_add(num_fences, &t->num_fences);
+	atomic_long_add(num_waits, &t->num_waits);
+
+	kfree(order);
+out_requests:
+	kfree(requests);
+	return err;
+}
+
+static int mock_breadcrumbs_smoketest(void *arg)
+{
+	struct drm_i915_private *i915 = arg;
+	struct smoketest t = {
+		.engine = i915->engine[RCS],
+		.ncontexts = 1024,
+		.max_batch = 1024,
+		.request_alloc = __mock_request_alloc
+	};
+	unsigned int ncpus = num_online_cpus();
+	struct task_struct **threads;
+	unsigned int n;
+	int ret = 0;
+
+	/*
+	 * Smoketest our breadcrumb/signal handling for requests across multiple
+	 * threads. A very simple test to only catch the most egregious of bugs.
+	 * See __igt_breadcrumbs_smoketest();
+	 */
+
+	threads = kmalloc_array(ncpus, sizeof(*threads), GFP_KERNEL);
+	if (!threads)
+		return -ENOMEM;
+
+	t.contexts =
+		kmalloc_array(t.ncontexts, sizeof(*t.contexts), GFP_KERNEL);
+	if (!t.contexts) {
+		ret = -ENOMEM;
+		goto out_threads;
+	}
+
+	mutex_lock(&t.engine->i915->drm.struct_mutex);
+	for (n = 0; n < t.ncontexts; n++) {
+		t.contexts[n] = mock_context(t.engine->i915, "mock");
+		if (!t.contexts[n]) {
+			ret = -ENOMEM;
+			goto out_contexts;
+		}
+	}
+	mutex_unlock(&t.engine->i915->drm.struct_mutex);
+
+	for (n = 0; n < ncpus; n++) {
+		threads[n] = kthread_run(__igt_breadcrumbs_smoketest,
+					 &t, "igt/%d", n);
+		if (IS_ERR(threads[n])) {
+			ret = PTR_ERR(threads[n]);
+			ncpus = n;
+			break;
+		}
+
+		get_task_struct(threads[n]);
+	}
+
+	msleep(jiffies_to_msecs(i915_selftest.timeout_jiffies));
+
+	for (n = 0; n < ncpus; n++) {
+		int err;
+
+		err = kthread_stop(threads[n]);
+		if (err < 0 && !ret)
+			ret = err;
+
+		put_task_struct(threads[n]);
+	}
+	pr_info("Completed %lu waits for %lu fence across %d cpus\n",
+		atomic_long_read(&t.num_waits),
+		atomic_long_read(&t.num_fences),
+		ncpus);
+
+	mutex_lock(&t.engine->i915->drm.struct_mutex);
+out_contexts:
+	for (n = 0; n < t.ncontexts; n++) {
+		if (!t.contexts[n])
+			break;
+		mock_context_close(t.contexts[n]);
+	}
+	mutex_unlock(&t.engine->i915->drm.struct_mutex);
+	kfree(t.contexts);
+out_threads:
+	kfree(threads);
+
+	return ret;
+}
+
 int i915_request_mock_selftests(void)
 {
 	static const struct i915_subtest tests[] = {
@@ -254,6 +505,7 @@ int i915_request_mock_selftests(void)
 		SUBTEST(igt_wait_request),
 		SUBTEST(igt_fence_wait),
 		SUBTEST(igt_request_rewind),
+		SUBTEST(mock_breadcrumbs_smoketest),
 	};
 	struct drm_i915_private *i915;
 	intel_wakeref_t wakeref;
@@ -812,6 +1064,178 @@ out_unlock:
 	return err;
 }
 
+static int
+max_batches(struct i915_gem_context *ctx, struct intel_engine_cs *engine)
+{
+	struct i915_request *rq;
+	int ret;
+
+	/*
+	 * Before execlists, all contexts share the same ringbuffer. With
+	 * execlists, each context/engine has a separate ringbuffer and
+	 * for the purposes of this test, inexhaustible.
+	 *
+	 * For the global ringbuffer though, we have to be very careful
+	 * that we do not wrap while preventing the execution of requests
+	 * with a unsignaled fence.
+	 */
+	if (HAS_EXECLISTS(ctx->i915))
+		return INT_MAX;
+
+	rq = i915_request_alloc(engine, ctx);
+	if (IS_ERR(rq)) {
+		ret = PTR_ERR(rq);
+	} else {
+		int sz;
+
+		ret = rq->ring->size - rq->reserved_space;
+		i915_request_add(rq);
+
+		sz = rq->ring->emit - rq->head;
+		if (sz < 0)
+			sz += rq->ring->size;
+		ret /= sz;
+		ret /= 2; /* leave half spare, in case of emergency! */
+	}
+
+	return ret;
+}
+
+static int live_breadcrumbs_smoketest(void *arg)
+{
+	struct drm_i915_private *i915 = arg;
+	struct smoketest t[I915_NUM_ENGINES];
+	unsigned int ncpus = num_online_cpus();
+	unsigned long num_waits, num_fences;
+	struct intel_engine_cs *engine;
+	struct task_struct **threads;
+	struct igt_live_test live;
+	enum intel_engine_id id;
+	intel_wakeref_t wakeref;
+	struct drm_file *file;
+	unsigned int n;
+	int ret = 0;
+
+	/*
+	 * Smoketest our breadcrumb/signal handling for requests across multiple
+	 * threads. A very simple test to only catch the most egregious of bugs.
+	 * See __igt_breadcrumbs_smoketest();
+	 *
+	 * On real hardware this time.
+	 */
+
+	wakeref = intel_runtime_pm_get(i915);
+
+	file = mock_file(i915);
+	if (IS_ERR(file)) {
+		ret = PTR_ERR(file);
+		goto out_rpm;
+	}
+
+	threads = kcalloc(ncpus * I915_NUM_ENGINES,
+			  sizeof(*threads),
+			  GFP_KERNEL);
+	if (!threads) {
+		ret = -ENOMEM;
+		goto out_file;
+	}
+
+	memset(&t[0], 0, sizeof(t[0]));
+	t[0].request_alloc = __live_request_alloc;
+	t[0].ncontexts = 64;
+	t[0].contexts = kmalloc_array(t[0].ncontexts,
+				      sizeof(*t[0].contexts),
+				      GFP_KERNEL);
+	if (!t[0].contexts) {
+		ret = -ENOMEM;
+		goto out_threads;
+	}
+
+	mutex_lock(&i915->drm.struct_mutex);
+	for (n = 0; n < t[0].ncontexts; n++) {
+		t[0].contexts[n] = live_context(i915, file);
+		if (!t[0].contexts[n]) {
+			ret = -ENOMEM;
+			goto out_contexts;
+		}
+	}
+
+	ret = igt_live_test_begin(&live, i915, __func__, "");
+	if (ret)
+		goto out_contexts;
+
+	for_each_engine(engine, i915, id) {
+		t[id] = t[0];
+		t[id].engine = engine;
+		t[id].max_batch = max_batches(t[0].contexts[0], engine);
+		if (t[id].max_batch < 0) {
+			ret = t[id].max_batch;
+			mutex_unlock(&i915->drm.struct_mutex);
+			goto out_flush;
+		}
+		/* One ring interleaved between requests from all cpus */
+		t[id].max_batch /= num_online_cpus() + 1;
+		pr_debug("Limiting batches to %d requests on %s\n",
+			 t[id].max_batch, engine->name);
+
+		for (n = 0; n < ncpus; n++) {
+			struct task_struct *tsk;
+
+			tsk = kthread_run(__igt_breadcrumbs_smoketest,
+					  &t[id], "igt/%d.%d", id, n);
+			if (IS_ERR(tsk)) {
+				ret = PTR_ERR(tsk);
+				mutex_unlock(&i915->drm.struct_mutex);
+				goto out_flush;
+			}
+
+			get_task_struct(tsk);
+			threads[id * ncpus + n] = tsk;
+		}
+	}
+	mutex_unlock(&i915->drm.struct_mutex);
+
+	msleep(jiffies_to_msecs(i915_selftest.timeout_jiffies));
+
+out_flush:
+	num_waits = 0;
+	num_fences = 0;
+	for_each_engine(engine, i915, id) {
+		for (n = 0; n < ncpus; n++) {
+			struct task_struct *tsk = threads[id * ncpus + n];
+			int err;
+
+			if (!tsk)
+				continue;
+
+			err = kthread_stop(tsk);
+			if (err < 0 && !ret)
+				ret = err;
+
+			put_task_struct(tsk);
+		}
+
+		num_waits += atomic_long_read(&t[id].num_waits);
+		num_fences += atomic_long_read(&t[id].num_fences);
+	}
+	pr_info("Completed %lu waits for %lu fences across %d engines and %d cpus\n",
+		num_waits, num_fences, RUNTIME_INFO(i915)->num_rings, ncpus);
+
+	mutex_lock(&i915->drm.struct_mutex);
+	ret = igt_live_test_end(&live) ?: ret;
+out_contexts:
+	mutex_unlock(&i915->drm.struct_mutex);
+	kfree(t[0].contexts);
+out_threads:
+	kfree(threads);
+out_file:
+	mock_file_free(i915, file);
+out_rpm:
+	intel_runtime_pm_put(i915, wakeref);
+
+	return ret;
+}
+
 int i915_request_live_selftests(struct drm_i915_private *i915)
 {
 	static const struct i915_subtest tests[] = {
@@ -819,6 +1243,7 @@ int i915_request_live_selftests(struct drm_i915_private *i915)
 		SUBTEST(live_all_engines),
 		SUBTEST(live_sequential_engines),
 		SUBTEST(live_empty_request),
+		SUBTEST(live_breadcrumbs_smoketest),
 	};
 
 	if (i915_terminally_wedged(&i915->gpu_error))
diff --git a/drivers/gpu/drm/i915/selftests/igt_spinner.c b/drivers/gpu/drm/i915/selftests/igt_spinner.c
index 0e70df0230b8..9ebd9225684e 100644
--- a/drivers/gpu/drm/i915/selftests/igt_spinner.c
+++ b/drivers/gpu/drm/i915/selftests/igt_spinner.c
@@ -185,11 +185,6 @@ void igt_spinner_fini(struct igt_spinner *spin)
 
 bool igt_wait_for_spinner(struct igt_spinner *spin, struct i915_request *rq)
 {
-	if (!wait_event_timeout(rq->execute,
-				READ_ONCE(rq->global_seqno),
-				msecs_to_jiffies(10)))
-		return false;
-
 	return !(wait_for_us(i915_seqno_passed(hws_seqno(spin, rq),
 					       rq->fence.seqno),
 			     10) &&
diff --git a/drivers/gpu/drm/i915/selftests/intel_breadcrumbs.c b/drivers/gpu/drm/i915/selftests/intel_breadcrumbs.c
deleted file mode 100644
index f03b407fdbe2..000000000000
--- a/drivers/gpu/drm/i915/selftests/intel_breadcrumbs.c
+++ /dev/null
@@ -1,470 +0,0 @@
-/*
- * Copyright © 2016 Intel Corporation
- *
- * Permission is hereby granted, free of charge, to any person obtaining a
- * copy of this software and associated documentation files (the "Software"),
- * to deal in the Software without restriction, including without limitation
- * the rights to use, copy, modify, merge, publish, distribute, sublicense,
- * and/or sell copies of the Software, and to permit persons to whom the
- * Software is furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice (including the next
- * paragraph) shall be included in all copies or substantial portions of the
- * Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
- * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
- * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
- * IN THE SOFTWARE.
- *
- */
-
-#include "../i915_selftest.h"
-#include "i915_random.h"
-
-#include "mock_gem_device.h"
-#include "mock_engine.h"
-
-static int check_rbtree(struct intel_engine_cs *engine,
-			const unsigned long *bitmap,
-			const struct intel_wait *waiters,
-			const int count)
-{
-	struct intel_breadcrumbs *b = &engine->breadcrumbs;
-	struct rb_node *rb;
-	int n;
-
-	if (&b->irq_wait->node != rb_first(&b->waiters)) {
-		pr_err("First waiter does not match first element of wait-tree\n");
-		return -EINVAL;
-	}
-
-	n = find_first_bit(bitmap, count);
-	for (rb = rb_first(&b->waiters); rb; rb = rb_next(rb)) {
-		struct intel_wait *w = container_of(rb, typeof(*w), node);
-		int idx = w - waiters;
-
-		if (!test_bit(idx, bitmap)) {
-			pr_err("waiter[%d, seqno=%d] removed but still in wait-tree\n",
-			       idx, w->seqno);
-			return -EINVAL;
-		}
-
-		if (n != idx) {
-			pr_err("waiter[%d, seqno=%d] does not match expected next element in tree [%d]\n",
-			       idx, w->seqno, n);
-			return -EINVAL;
-		}
-
-		n = find_next_bit(bitmap, count, n + 1);
-	}
-
-	return 0;
-}
-
-static int check_completion(struct intel_engine_cs *engine,
-			    const unsigned long *bitmap,
-			    const struct intel_wait *waiters,
-			    const int count)
-{
-	int n;
-
-	for (n = 0; n < count; n++) {
-		if (intel_wait_complete(&waiters[n]) != !!test_bit(n, bitmap))
-			continue;
-
-		pr_err("waiter[%d, seqno=%d] is %s, but expected %s\n",
-		       n, waiters[n].seqno,
-		       intel_wait_complete(&waiters[n]) ? "complete" : "active",
-		       test_bit(n, bitmap) ? "active" : "complete");
-		return -EINVAL;
-	}
-
-	return 0;
-}
-
-static int check_rbtree_empty(struct intel_engine_cs *engine)
-{
-	struct intel_breadcrumbs *b = &engine->breadcrumbs;
-
-	if (b->irq_wait) {
-		pr_err("Empty breadcrumbs still has a waiter\n");
-		return -EINVAL;
-	}
-
-	if (!RB_EMPTY_ROOT(&b->waiters)) {
-		pr_err("Empty breadcrumbs, but wait-tree not empty\n");
-		return -EINVAL;
-	}
-
-	return 0;
-}
-
-static int igt_random_insert_remove(void *arg)
-{
-	const u32 seqno_bias = 0x1000;
-	I915_RND_STATE(prng);
-	struct intel_engine_cs *engine = arg;
-	struct intel_wait *waiters;
-	const int count = 4096;
-	unsigned int *order;
-	unsigned long *bitmap;
-	int err = -ENOMEM;
-	int n;
-
-	mock_engine_reset(engine);
-
-	waiters = kvmalloc_array(count, sizeof(*waiters), GFP_KERNEL);
-	if (!waiters)
-		goto out_engines;
-
-	bitmap = kcalloc(DIV_ROUND_UP(count, BITS_PER_LONG), sizeof(*bitmap),
-			 GFP_KERNEL);
-	if (!bitmap)
-		goto out_waiters;
-
-	order = i915_random_order(count, &prng);
-	if (!order)
-		goto out_bitmap;
-
-	for (n = 0; n < count; n++)
-		intel_wait_init_for_seqno(&waiters[n], seqno_bias + n);
-
-	err = check_rbtree(engine, bitmap, waiters, count);
-	if (err)
-		goto out_order;
-
-	/* Add and remove waiters into the rbtree in random order. At each
-	 * step, we verify that the rbtree is correctly ordered.
-	 */
-	for (n = 0; n < count; n++) {
-		int i = order[n];
-
-		intel_engine_add_wait(engine, &waiters[i]);
-		__set_bit(i, bitmap);
-
-		err = check_rbtree(engine, bitmap, waiters, count);
-		if (err)
-			goto out_order;
-	}
-
-	i915_random_reorder(order, count, &prng);
-	for (n = 0; n < count; n++) {
-		int i = order[n];
-
-		intel_engine_remove_wait(engine, &waiters[i]);
-		__clear_bit(i, bitmap);
-
-		err = check_rbtree(engine, bitmap, waiters, count);
-		if (err)
-			goto out_order;
-	}
-
-	err = check_rbtree_empty(engine);
-out_order:
-	kfree(order);
-out_bitmap:
-	kfree(bitmap);
-out_waiters:
-	kvfree(waiters);
-out_engines:
-	mock_engine_flush(engine);
-	return err;
-}
-
-static int igt_insert_complete(void *arg)
-{
-	const u32 seqno_bias = 0x1000;
-	struct intel_engine_cs *engine = arg;
-	struct intel_wait *waiters;
-	const int count = 4096;
-	unsigned long *bitmap;
-	int err = -ENOMEM;
-	int n, m;
-
-	mock_engine_reset(engine);
-
-	waiters = kvmalloc_array(count, sizeof(*waiters), GFP_KERNEL);
-	if (!waiters)
-		goto out_engines;
-
-	bitmap = kcalloc(DIV_ROUND_UP(count, BITS_PER_LONG), sizeof(*bitmap),
-			 GFP_KERNEL);
-	if (!bitmap)
-		goto out_waiters;
-
-	for (n = 0; n < count; n++) {
-		intel_wait_init_for_seqno(&waiters[n], n + seqno_bias);
-		intel_engine_add_wait(engine, &waiters[n]);
-		__set_bit(n, bitmap);
-	}
-	err = check_rbtree(engine, bitmap, waiters, count);
-	if (err)
-		goto out_bitmap;
-
-	/* On each step, we advance the seqno so that several waiters are then
-	 * complete (we increase the seqno by increasingly larger values to
-	 * retire more and more waiters at once). All retired waiters should
-	 * be woken and removed from the rbtree, and so that we check.
-	 */
-	for (n = 0; n < count; n = m) {
-		int seqno = 2 * n;
-
-		GEM_BUG_ON(find_first_bit(bitmap, count) != n);
-
-		if (intel_wait_complete(&waiters[n])) {
-			pr_err("waiter[%d, seqno=%d] completed too early\n",
-			       n, waiters[n].seqno);
-			err = -EINVAL;
-			goto out_bitmap;
-		}
-
-		/* complete the following waiters */
-		mock_seqno_advance(engine, seqno + seqno_bias);
-		for (m = n; m <= seqno; m++) {
-			if (m == count)
-				break;
-
-			GEM_BUG_ON(!test_bit(m, bitmap));
-			__clear_bit(m, bitmap);
-		}
-
-		intel_engine_remove_wait(engine, &waiters[n]);
-		RB_CLEAR_NODE(&waiters[n].node);
-
-		err = check_rbtree(engine, bitmap, waiters, count);
-		if (err) {
-			pr_err("rbtree corrupt after seqno advance to %d\n",
-			       seqno + seqno_bias);
-			goto out_bitmap;
-		}
-
-		err = check_completion(engine, bitmap, waiters, count);
-		if (err) {
-			pr_err("completions after seqno advance to %d failed\n",
-			       seqno + seqno_bias);
-			goto out_bitmap;
-		}
-	}
-
-	err = check_rbtree_empty(engine);
-out_bitmap:
-	kfree(bitmap);
-out_waiters:
-	kvfree(waiters);
-out_engines:
-	mock_engine_flush(engine);
-	return err;
-}
-
-struct igt_wakeup {
-	struct task_struct *tsk;
-	atomic_t *ready, *set, *done;
-	struct intel_engine_cs *engine;
-	unsigned long flags;
-#define STOP 0
-#define IDLE 1
-	wait_queue_head_t *wq;
-	u32 seqno;
-};
-
-static bool wait_for_ready(struct igt_wakeup *w)
-{
-	DEFINE_WAIT(ready);
-
-	set_bit(IDLE, &w->flags);
-	if (atomic_dec_and_test(w->done))
-		wake_up_var(w->done);
-
-	if (test_bit(STOP, &w->flags))
-		goto out;
-
-	for (;;) {
-		prepare_to_wait(w->wq, &ready, TASK_INTERRUPTIBLE);
-		if (atomic_read(w->ready) == 0)
-			break;
-
-		schedule();
-	}
-	finish_wait(w->wq, &ready);
-
-out:
-	clear_bit(IDLE, &w->flags);
-	if (atomic_dec_and_test(w->set))
-		wake_up_var(w->set);
-
-	return !test_bit(STOP, &w->flags);
-}
-
-static int igt_wakeup_thread(void *arg)
-{
-	struct igt_wakeup *w = arg;
-	struct intel_wait wait;
-
-	while (wait_for_ready(w)) {
-		GEM_BUG_ON(kthread_should_stop());
-
-		intel_wait_init_for_seqno(&wait, w->seqno);
-		intel_engine_add_wait(w->engine, &wait);
-		for (;;) {
-			set_current_state(TASK_UNINTERRUPTIBLE);
-			if (i915_seqno_passed(intel_engine_get_seqno(w->engine),
-					      w->seqno))
-				break;
-
-			if (test_bit(STOP, &w->flags)) /* emergency escape */
-				break;
-
-			schedule();
-		}
-		intel_engine_remove_wait(w->engine, &wait);
-		__set_current_state(TASK_RUNNING);
-	}
-
-	return 0;
-}
-
-static void igt_wake_all_sync(atomic_t *ready,
-			      atomic_t *set,
-			      atomic_t *done,
-			      wait_queue_head_t *wq,
-			      int count)
-{
-	atomic_set(set, count);
-	atomic_set(ready, 0);
-	wake_up_all(wq);
-
-	wait_var_event(set, !atomic_read(set));
-	atomic_set(ready, count);
-	atomic_set(done, count);
-}
-
-static int igt_wakeup(void *arg)
-{
-	I915_RND_STATE(prng);
-	struct intel_engine_cs *engine = arg;
-	struct igt_wakeup *waiters;
-	DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
-	const int count = 4096;
-	const u32 max_seqno = count / 4;
-	atomic_t ready, set, done;
-	int err = -ENOMEM;
-	int n, step;
-
-	mock_engine_reset(engine);
-
-	waiters = kvmalloc_array(count, sizeof(*waiters), GFP_KERNEL);
-	if (!waiters)
-		goto out_engines;
-
-	/* Create a large number of threads, each waiting on a random seqno.
-	 * Multiple waiters will be waiting for the same seqno.
-	 */
-	atomic_set(&ready, count);
-	for (n = 0; n < count; n++) {
-		waiters[n].wq = &wq;
-		waiters[n].ready = &ready;
-		waiters[n].set = &set;
-		waiters[n].done = &done;
-		waiters[n].engine = engine;
-		waiters[n].flags = BIT(IDLE);
-
-		waiters[n].tsk = kthread_run(igt_wakeup_thread, &waiters[n],
-					     "i915/igt:%d", n);
-		if (IS_ERR(waiters[n].tsk))
-			goto out_waiters;
-
-		get_task_struct(waiters[n].tsk);
-	}
-
-	for (step = 1; step <= max_seqno; step <<= 1) {
-		u32 seqno;
-
-		/* The waiter threads start paused as we assign them a random
-		 * seqno and reset the engine. Once the engine is reset,
-		 * we signal that the threads may begin their wait upon their
-		 * seqno.
-		 */
-		for (n = 0; n < count; n++) {
-			GEM_BUG_ON(!test_bit(IDLE, &waiters[n].flags));
-			waiters[n].seqno =
-				1 + prandom_u32_state(&prng) % max_seqno;
-		}
-		mock_seqno_advance(engine, 0);
-		igt_wake_all_sync(&ready, &set, &done, &wq, count);
-
-		/* Simulate the GPU doing chunks of work, with one or more
-		 * seqno appearing to finish at the same time. A random number
-		 * of threads will be waiting upon the update and hopefully be
-		 * woken.
-		 */
-		for (seqno = 1; seqno <= max_seqno + step; seqno += step) {
-			usleep_range(50, 500);
-			mock_seqno_advance(engine, seqno);
-		}
-		GEM_BUG_ON(intel_engine_get_seqno(engine) < 1 + max_seqno);
-
-		/* With the seqno now beyond any of the waiting threads, they
-		 * should all be woken, see that they are complete and signal
-		 * that they are ready for the next test. We wait until all
-		 * threads are complete and waiting for us (i.e. not a seqno).
-		 */
-		if (!wait_var_event_timeout(&done,
-					    !atomic_read(&done), 10 * HZ)) {
-			pr_err("Timed out waiting for %d remaining waiters\n",
-			       atomic_read(&done));
-			err = -ETIMEDOUT;
-			break;
-		}
-
-		err = check_rbtree_empty(engine);
-		if (err)
-			break;
-	}
-
-out_waiters:
-	for (n = 0; n < count; n++) {
-		if (IS_ERR(waiters[n].tsk))
-			break;
-
-		set_bit(STOP, &waiters[n].flags);
-	}
-	mock_seqno_advance(engine, INT_MAX); /* wakeup any broken waiters */
-	igt_wake_all_sync(&ready, &set, &done, &wq, n);
-
-	for (n = 0; n < count; n++) {
-		if (IS_ERR(waiters[n].tsk))
-			break;
-
-		kthread_stop(waiters[n].tsk);
-		put_task_struct(waiters[n].tsk);
-	}
-
-	kvfree(waiters);
-out_engines:
-	mock_engine_flush(engine);
-	return err;
-}
-
-int intel_breadcrumbs_mock_selftests(void)
-{
-	static const struct i915_subtest tests[] = {
-		SUBTEST(igt_random_insert_remove),
-		SUBTEST(igt_insert_complete),
-		SUBTEST(igt_wakeup),
-	};
-	struct drm_i915_private *i915;
-	int err;
-
-	i915 = mock_gem_device();
-	if (!i915)
-		return -ENOMEM;
-
-	err = i915_subtests(tests, i915->engine[RCS]);
-	drm_dev_put(&i915->drm);
-
-	return err;
-}
diff --git a/drivers/gpu/drm/i915/selftests/intel_hangcheck.c b/drivers/gpu/drm/i915/selftests/intel_hangcheck.c
index 2c38ea5892d9..7b6f3bea9ef8 100644
--- a/drivers/gpu/drm/i915/selftests/intel_hangcheck.c
+++ b/drivers/gpu/drm/i915/selftests/intel_hangcheck.c
@@ -1127,7 +1127,7 @@ static int __igt_reset_evict_vma(struct drm_i915_private *i915,
 
 	wait_for_completion(&arg.completion);
 
-	if (wait_for(waitqueue_active(&rq->execute), 10)) {
+	if (wait_for(!list_empty(&rq->fence.cb_list), 10)) {
 		struct drm_printer p = drm_info_printer(i915->drm.dev);
 
 		pr_err("igt/evict_vma kthread did not wait\n");
diff --git a/drivers/gpu/drm/i915/selftests/lib_sw_fence.c b/drivers/gpu/drm/i915/selftests/lib_sw_fence.c
index b26f07b55d86..2bfa72c1654b 100644
--- a/drivers/gpu/drm/i915/selftests/lib_sw_fence.c
+++ b/drivers/gpu/drm/i915/selftests/lib_sw_fence.c
@@ -76,3 +76,57 @@ void timed_fence_fini(struct timed_fence *tf)
 	destroy_timer_on_stack(&tf->timer);
 	i915_sw_fence_fini(&tf->fence);
 }
+
+struct heap_fence {
+	struct i915_sw_fence fence;
+	union {
+		struct kref ref;
+		struct rcu_head rcu;
+	};
+};
+
+static int __i915_sw_fence_call
+heap_fence_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
+{
+	struct heap_fence *h = container_of(fence, typeof(*h), fence);
+
+	switch (state) {
+	case FENCE_COMPLETE:
+		break;
+
+	case FENCE_FREE:
+		heap_fence_put(&h->fence);
+	}
+
+	return NOTIFY_DONE;
+}
+
+struct i915_sw_fence *heap_fence_create(gfp_t gfp)
+{
+	struct heap_fence *h;
+
+	h = kmalloc(sizeof(*h), gfp);
+	if (!h)
+		return NULL;
+
+	i915_sw_fence_init(&h->fence, heap_fence_notify);
+	refcount_set(&h->ref.refcount, 2);
+
+	return &h->fence;
+}
+
+static void heap_fence_release(struct kref *ref)
+{
+	struct heap_fence *h = container_of(ref, typeof(*h), ref);
+
+	i915_sw_fence_fini(&h->fence);
+
+	kfree_rcu(h, rcu);
+}
+
+void heap_fence_put(struct i915_sw_fence *fence)
+{
+	struct heap_fence *h = container_of(fence, typeof(*h), fence);
+
+	kref_put(&h->ref, heap_fence_release);
+}
diff --git a/drivers/gpu/drm/i915/selftests/lib_sw_fence.h b/drivers/gpu/drm/i915/selftests/lib_sw_fence.h
index 474aafb92ae1..1f9927e10f3a 100644
--- a/drivers/gpu/drm/i915/selftests/lib_sw_fence.h
+++ b/drivers/gpu/drm/i915/selftests/lib_sw_fence.h
@@ -39,4 +39,7 @@ struct timed_fence {
 void timed_fence_init(struct timed_fence *tf, unsigned long expires);
 void timed_fence_fini(struct timed_fence *tf);
 
+struct i915_sw_fence *heap_fence_create(gfp_t gfp);
+void heap_fence_put(struct i915_sw_fence *fence);
+
 #endif /* _LIB_SW_FENCE_H_ */
diff --git a/drivers/gpu/drm/i915/selftests/mock_engine.c b/drivers/gpu/drm/i915/selftests/mock_engine.c
index 3b226ebc6bc4..08f0cab02e0f 100644
--- a/drivers/gpu/drm/i915/selftests/mock_engine.c
+++ b/drivers/gpu/drm/i915/selftests/mock_engine.c
@@ -86,17 +86,21 @@ static struct mock_request *first_request(struct mock_engine *engine)
 static void advance(struct mock_request *request)
 {
 	list_del_init(&request->link);
-	mock_seqno_advance(request->base.engine, request->base.global_seqno);
+	intel_engine_write_global_seqno(request->base.engine,
+					request->base.global_seqno);
 	i915_request_mark_complete(&request->base);
 	GEM_BUG_ON(!i915_request_completed(&request->base));
+
+	intel_engine_queue_breadcrumbs(request->base.engine);
 }
 
 static void hw_delay_complete(struct timer_list *t)
 {
 	struct mock_engine *engine = from_timer(engine, t, hw_delay);
 	struct mock_request *request;
+	unsigned long flags;
 
-	spin_lock(&engine->hw_lock);
+	spin_lock_irqsave(&engine->hw_lock, flags);
 
 	/* Timer fired, first request is complete */
 	request = first_request(engine);
@@ -116,7 +120,7 @@ static void hw_delay_complete(struct timer_list *t)
 		advance(request);
 	}
 
-	spin_unlock(&engine->hw_lock);
+	spin_unlock_irqrestore(&engine->hw_lock, flags);
 }
 
 static void mock_context_unpin(struct intel_context *ce)
@@ -191,11 +195,12 @@ static void mock_submit_request(struct i915_request *request)
 	struct mock_request *mock = container_of(request, typeof(*mock), base);
 	struct mock_engine *engine =
 		container_of(request->engine, typeof(*engine), base);
+	unsigned long flags;
 
 	i915_request_submit(request);
 	GEM_BUG_ON(!request->global_seqno);
 
-	spin_lock_irq(&engine->hw_lock);
+	spin_lock_irqsave(&engine->hw_lock, flags);
 	list_add_tail(&mock->link, &engine->hw_queue);
 	if (mock->link.prev == &engine->hw_queue) {
 		if (mock->delay)
@@ -203,7 +208,7 @@ static void mock_submit_request(struct i915_request *request)
 		else
 			advance(mock);
 	}
-	spin_unlock_irq(&engine->hw_lock);
+	spin_unlock_irqrestore(&engine->hw_lock, flags);
 }
 
 struct intel_engine_cs *mock_engine(struct drm_i915_private *i915,
@@ -273,7 +278,7 @@ void mock_engine_flush(struct intel_engine_cs *engine)
 
 void mock_engine_reset(struct intel_engine_cs *engine)
 {
-	intel_write_status_page(engine, I915_GEM_HWS_INDEX, 0);
+	intel_engine_write_global_seqno(engine, 0);
 }
 
 void mock_engine_free(struct intel_engine_cs *engine)
diff --git a/drivers/gpu/drm/i915/selftests/mock_engine.h b/drivers/gpu/drm/i915/selftests/mock_engine.h
index 133d0c21790d..b9cc3a245f16 100644
--- a/drivers/gpu/drm/i915/selftests/mock_engine.h
+++ b/drivers/gpu/drm/i915/selftests/mock_engine.h
@@ -46,10 +46,4 @@ void mock_engine_flush(struct intel_engine_cs *engine);
 void mock_engine_reset(struct intel_engine_cs *engine);
 void mock_engine_free(struct intel_engine_cs *engine);
 
-static inline void mock_seqno_advance(struct intel_engine_cs *engine, u32 seqno)
-{
-	intel_write_status_page(engine, I915_GEM_HWS_INDEX, seqno);
-	intel_engine_wakeup(engine);
-}
-
 #endif /* !__MOCK_ENGINE_H__ */