[Intel-gfx] [PATCH 15/18] drm/i915/execlists: Try preempt-reset from softirq context

[Chris Wilson]

When circumstances allow, trying resetting the engine directly from the
preemption timeout handler. As this is softirq context, we have to be
careful both not to sleep and not to spin on anything we may be
interrupting (e.g. the submission tasklet).

Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala at linux.intel.com>
Cc: Michał Winiarski <michal.winiarski at intel.com>
CC: Michel Thierry <michel.thierry at intel.com>
Cc: Jeff McGee <jeff.mcgee at intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
---
 drivers/gpu/drm/i915/intel_lrc.c           |  31 +++++++-
 drivers/gpu/drm/i915/selftests/intel_lrc.c | 121 +++++++++++++++++++++++++++++
 2 files changed, 151 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
index bee8a58f340f..fea275f718f5 100644
--- a/drivers/gpu/drm/i915/intel_lrc.c
+++ b/drivers/gpu/drm/i915/intel_lrc.c
@@ -549,6 +549,33 @@ static void inject_preempt_context(struct intel_engine_cs *engine)
 	execlists_set_active(execlists, EXECLISTS_ACTIVE_PREEMPT);
 }
 
+static int try_preempt_reset(struct intel_engine_execlists *execlists)
+{
+	struct intel_engine_cs *engine =
+		container_of(execlists, typeof(*engine), execlists);
+	int err = -EBUSY;
+
+	if (!test_bit(ENGINE_IRQ_EXECLIST, &engine->irq_posted) &&
+	    tasklet_trylock(&execlists->tasklet)) {
+		unsigned long *lock = &engine->i915->gpu_error.flags;
+		unsigned int bit = I915_RESET_ENGINE + engine->id;
+
+		if (!test_and_set_bit(bit, lock)) {
+			tasklet_disable_nosync(&engine->execlists.tasklet);
+			err = i915_reset_engine(engine,
+						"preemption time out");
+			tasklet_enable(&engine->execlists.tasklet);
+
+			clear_bit(bit, lock);
+			wake_up_bit(lock, bit);
+		}
+
+		tasklet_unlock(&execlists->tasklet);
+	}
+
+	return err;
+}
+
 static enum hrtimer_restart preempt_timeout(struct hrtimer *hrtimer)
 {
 	struct intel_engine_execlists *execlists =
@@ -562,7 +589,9 @@ static enum hrtimer_restart preempt_timeout(struct hrtimer *hrtimer)
 	if (!execlists_is_active(execlists, EXECLISTS_ACTIVE_PREEMPT_TIMEOUT))
 		return HRTIMER_NORESTART;
 
-	queue_work(system_highpri_wq, &execlists->preempt_reset);
+	if (try_preempt_reset(execlists))
+		queue_work(system_highpri_wq, &execlists->preempt_reset);
+
 	return HRTIMER_NORESTART;
 }
 
diff --git a/drivers/gpu/drm/i915/selftests/intel_lrc.c b/drivers/gpu/drm/i915/selftests/intel_lrc.c
index a8f2491eab1d..cd97a6461ec7 100644
--- a/drivers/gpu/drm/i915/selftests/intel_lrc.c
+++ b/drivers/gpu/drm/i915/selftests/intel_lrc.c
@@ -560,6 +560,126 @@ static int live_preempt_timeout(void *arg)
 	return err;
 }
 
+static void __softirq_begin(void)
+{
+	local_bh_disable();
+}
+
+static void __softirq_end(void)
+{
+	local_bh_enable();
+}
+
+static void __hardirq_begin(void)
+{
+	local_irq_disable();
+}
+
+static void __hardirq_end(void)
+{
+	local_irq_enable();
+}
+
+static int live_preempt_reset(void *arg)
+{
+	struct drm_i915_private *i915 = arg;
+	struct intel_engine_cs *engine;
+	struct i915_gem_context *ctx;
+	enum intel_engine_id id;
+	struct spinner spin;
+	int err = -ENOMEM;
+
+	if (!HAS_LOGICAL_RING_PREEMPTION(i915))
+		return 0;
+
+	mutex_lock(&i915->drm.struct_mutex);
+
+	if (spinner_init(&spin, i915))
+		goto err_unlock;
+
+	ctx= kernel_context(i915);
+	if (!ctx)
+		goto err_spin;
+
+	for_each_engine(engine, i915, id) {
+		static const struct {
+			const char *name;
+			void (*critical_section_begin)(void);
+			void (*critical_section_end)(void);
+		} phases[] = {
+			{ "softirq", __softirq_begin, __softirq_end },
+			{ "hardirq", __hardirq_begin, __hardirq_end },
+			{ }
+		};
+		const typeof(*phases) *p;
+
+		for (p = phases; p->name; p++) {
+			struct i915_request *rq;
+
+			rq = spinner_create_request(&spin, ctx, engine,
+						    MI_NOOP);
+			if (IS_ERR(rq)) {
+				err = PTR_ERR(rq);
+				goto err_ctx;
+			}
+
+			i915_request_add(rq);
+			if (!wait_for_spinner(&spin, rq)) {
+				i915_gem_set_wedged(i915);
+				err = -EIO;
+				goto err_ctx;
+			}
+
+			/* Flush to give try_preempt_reset a chance */
+			do {
+				tasklet_schedule(&engine->execlists.tasklet);
+				usleep_range(100, 1000);
+				tasklet_kill(&engine->execlists.tasklet);
+			} while (test_bit(ENGINE_IRQ_EXECLIST,
+					  &engine->irq_posted));
+			GEM_BUG_ON(i915_request_completed(rq));
+
+			GEM_TRACE("%s triggering %s reset\n",
+				  engine->name, p->name);
+			p->critical_section_begin();
+			err = try_preempt_reset(&engine->execlists);
+			p->critical_section_end();
+			if (err) {
+				pr_err("Preempt softirq reset failed on %s, irq_posted? %d, tasklet state %lx\n",
+				       engine->name,
+				       test_bit(ENGINE_IRQ_EXECLIST,
+						&engine->irq_posted),
+				       engine->execlists.tasklet.state);
+				i915_gem_set_wedged(i915);
+				goto err_ctx;
+			}
+
+			/*
+			 * In such a simple case as above; triggering a reset
+			 * allows us to save and restore the hog nearly
+			 * perfectly...
+			 */
+			GEM_BUG_ON(i915_request_completed(rq));
+			spinner_end(&spin);
+
+			if (flush_test(i915, I915_WAIT_LOCKED)) {
+				err = -EIO;
+				goto err_ctx;
+			}
+		}
+	}
+
+	err = 0;
+err_ctx:
+	kernel_context_close(ctx);
+err_spin:
+	spinner_fini(&spin);
+err_unlock:
+	flush_test(i915, I915_WAIT_LOCKED);
+	mutex_unlock(&i915->drm.struct_mutex);
+	return err;
+}
+
 int intel_execlists_live_selftests(struct drm_i915_private *i915)
 {
 	static const struct i915_subtest tests[] = {
@@ -567,6 +687,7 @@ int intel_execlists_live_selftests(struct drm_i915_private *i915)
 		SUBTEST(live_preempt),
 		SUBTEST(live_late_preempt),
 		SUBTEST(live_preempt_timeout),
+		SUBTEST(live_preempt_reset),
 	};
 	return i915_subtests(tests, i915);
 }
-- 
2.16.3