diff mbox series

[v2] drm/i915: Be more gentle with exiting non-persistent context

Message ID 20210430141051.1138439-1-tvrtko.ursulin@linux.intel.com (mailing list archive)
State New, archived
Headers show
Series [v2] drm/i915: Be more gentle with exiting non-persistent context | expand

Commit Message

Tvrtko Ursulin April 30, 2021, 2:10 p.m. UTC 
From: Tvrtko Ursulin <tvrtko.ursulin@intel.com>

When a non-persistent context exits we currently mark it as banned in
order to trigger fast termination of any outstanding GPU jobs it may have
left running.

In doing so we apply a very strict 1ms limit in which the left over job
has to preempt before we issues an engine resets.

Some workloads are not able to cleanly preempt in that time window and it
can be argued that it would instead be better to give them a bit more
grace since avoiding engine resets is generally preferrable.

To achieve this the patch splits handling of banned contexts from simply
exited non-persistent ones and then applies different timeouts for both
and also extends the criteria which determines if a request should be
scheduled back in after preemption or not.

15ms preempt timeout grace is given to exited non-persistent contexts
which have been empirically tested to satisfy customers requirements
and still provides reasonably quick cleanup post exit.

v2:
 * Streamline fast path checks.

Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Zhen Han <zhen.han@intel.com>
---
 drivers/gpu/drm/i915/gem/i915_gem_context.c   | 16 +++++++++------
 drivers/gpu/drm/i915/gt/intel_context.c       |  2 ++
 drivers/gpu/drm/i915/gt/intel_context.h       | 20 +++++++++++++++++++
 drivers/gpu/drm/i915/gt/intel_context_types.h |  2 ++
 .../drm/i915/gt/intel_execlists_submission.c  | 11 ++++++++--
 drivers/gpu/drm/i915/i915_request.c           |  2 +-
 6 files changed, 44 insertions(+), 9 deletions(-)
diff mbox series

Patch

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_context.c b/drivers/gpu/drm/i915/gem/i915_gem_context.c
index fd8ee52e17a4..090c891f029b 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_context.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_context.c
@@ -426,7 +426,8 @@  static struct intel_engine_cs *active_engine(struct intel_context *ce)
 	return engine;
 }
 
-static void kill_engines(struct i915_gem_engines *engines, bool ban)
+static void
+kill_engines(struct i915_gem_engines *engines, bool ban, bool persistent)
 {
 	struct i915_gem_engines_iter it;
 	struct intel_context *ce;
@@ -443,6 +444,9 @@  static void kill_engines(struct i915_gem_engines *engines, bool ban)
 
 		if (ban && intel_context_set_banned(ce))
 			continue;
+		else if (!persistent &&
+			 intel_context_set_closed_non_persistent(ce))
+			continue;
 
 		/*
 		 * Check the current active state of this context; if we
@@ -454,7 +458,7 @@  static void kill_engines(struct i915_gem_engines *engines, bool ban)
 		engine = active_engine(ce);
 
 		/* First attempt to gracefully cancel the context */
-		if (engine && !__cancel_engine(engine) && ban)
+		if (engine && !__cancel_engine(engine) && (ban || !persistent))
 			/*
 			 * If we are unable to send a preemptive pulse to bump
 			 * the context from the GPU, we have to resort to a full
@@ -466,8 +470,6 @@  static void kill_engines(struct i915_gem_engines *engines, bool ban)
 
 static void kill_context(struct i915_gem_context *ctx)
 {
-	bool ban = (!i915_gem_context_is_persistent(ctx) ||
-		    !ctx->i915->params.enable_hangcheck);
 	struct i915_gem_engines *pos, *next;
 
 	spin_lock_irq(&ctx->stale.lock);
@@ -480,7 +482,8 @@  static void kill_context(struct i915_gem_context *ctx)
 
 		spin_unlock_irq(&ctx->stale.lock);
 
-		kill_engines(pos, ban);
+		kill_engines(pos, !ctx->i915->params.enable_hangcheck,
+			     i915_gem_context_is_persistent(ctx));
 
 		spin_lock_irq(&ctx->stale.lock);
 		GEM_BUG_ON(i915_sw_fence_signaled(&pos->fence));
@@ -526,7 +529,8 @@  static void engines_idle_release(struct i915_gem_context *ctx,
 
 kill:
 	if (list_empty(&engines->link)) /* raced, already closed */
-		kill_engines(engines, true);
+		kill_engines(engines, true,
+			     i915_gem_context_is_persistent(ctx));
 
 	i915_sw_fence_commit(&engines->fence);
 }
diff --git a/drivers/gpu/drm/i915/gt/intel_context.c b/drivers/gpu/drm/i915/gt/intel_context.c
index 17cf2640b082..efbda677a0a8 100644
--- a/drivers/gpu/drm/i915/gt/intel_context.c
+++ b/drivers/gpu/drm/i915/gt/intel_context.c
@@ -374,6 +374,8 @@  intel_context_init(struct intel_context *ce, struct intel_engine_cs *engine)
 	ce->sseu = engine->sseu;
 	ce->ring = __intel_context_ring_size(SZ_4K);
 
+	__set_bit(CONTEXT_SCHEDULABLE, &ce->flags);
+
 	ewma_runtime_init(&ce->runtime.avg);
 
 	ce->vm = i915_vm_get(engine->gt->vm);
diff --git a/drivers/gpu/drm/i915/gt/intel_context.h b/drivers/gpu/drm/i915/gt/intel_context.h
index f83a73a2b39f..a6d4c2d7f2f5 100644
--- a/drivers/gpu/drm/i915/gt/intel_context.h
+++ b/drivers/gpu/drm/i915/gt/intel_context.h
@@ -217,9 +217,29 @@  static inline bool intel_context_is_banned(const struct intel_context *ce)
 
 static inline bool intel_context_set_banned(struct intel_context *ce)
 {
+	clear_bit(CONTEXT_SCHEDULABLE, &ce->flags);
 	return test_and_set_bit(CONTEXT_BANNED, &ce->flags);
 }
 
+static inline bool
+intel_context_is_closed_non_persistent(const struct intel_context *ce)
+{
+	return test_bit(CONTEXT_NON_PERSISTENT, &ce->flags);
+}
+
+static inline bool
+intel_context_set_closed_non_persistent(struct intel_context *ce)
+{
+	GEM_BUG_ON(!intel_context_is_closed(ce));
+	clear_bit(CONTEXT_SCHEDULABLE, &ce->flags);
+	return test_and_set_bit(CONTEXT_NON_PERSISTENT, &ce->flags);
+}
+
+static inline bool intel_context_is_schedulable(const struct intel_context *ce)
+{
+	return test_bit(CONTEXT_SCHEDULABLE, &ce->flags);
+}
+
 static inline bool
 intel_context_force_single_submission(const struct intel_context *ce)
 {
diff --git a/drivers/gpu/drm/i915/gt/intel_context_types.h b/drivers/gpu/drm/i915/gt/intel_context_types.h
index ed8c447a7346..2d07b821a20e 100644
--- a/drivers/gpu/drm/i915/gt/intel_context_types.h
+++ b/drivers/gpu/drm/i915/gt/intel_context_types.h
@@ -95,6 +95,8 @@  struct intel_context {
 #define CONTEXT_BANNED			6
 #define CONTEXT_FORCE_SINGLE_SUBMISSION	7
 #define CONTEXT_NOPREEMPT		8
+#define CONTEXT_SCHEDULABLE		9  /* Unless banned or non-persistent closed. */
+#define CONTEXT_NON_PERSISTENT		10 /* Only if also closed. */
 
 	struct {
 		u64 timeout_us;
diff --git a/drivers/gpu/drm/i915/gt/intel_execlists_submission.c b/drivers/gpu/drm/i915/gt/intel_execlists_submission.c
index de124870af44..3f3e2ee6a24d 100644
--- a/drivers/gpu/drm/i915/gt/intel_execlists_submission.c
+++ b/drivers/gpu/drm/i915/gt/intel_execlists_submission.c
@@ -478,7 +478,7 @@  __execlists_schedule_in(struct i915_request *rq)
 		     !intel_engine_has_heartbeat(engine)))
 		intel_context_set_banned(ce);
 
-	if (unlikely(intel_context_is_banned(ce) || bad_request(rq)))
+	if (unlikely(!intel_context_is_schedulable(ce) || bad_request(rq)))
 		reset_active(rq, engine);
 
 	if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
@@ -1204,12 +1204,19 @@  static void record_preemption(struct intel_engine_execlists *execlists)
 static unsigned long active_preempt_timeout(struct intel_engine_cs *engine,
 					    const struct i915_request *rq)
 {
+	struct intel_context *ce;
+
 	if (!rq)
 		return 0;
 
+	ce = rq->context;
+
 	/* Force a fast reset for terminated contexts (ignoring sysfs!) */
-	if (unlikely(intel_context_is_banned(rq->context) || bad_request(rq)))
+	if (unlikely(intel_context_is_banned(ce) || bad_request(rq)))
 		return 1;
+	/* Longer grace for closed non-persistent contexts to avoid resets. */
+	else if (unlikely(intel_context_is_closed_non_persistent(ce)))
+		return 15;
 
 	return READ_ONCE(engine->props.preempt_timeout_ms);
 }
diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c
index bec9c3652188..bbac87535923 100644
--- a/drivers/gpu/drm/i915/i915_request.c
+++ b/drivers/gpu/drm/i915/i915_request.c
@@ -660,7 +660,7 @@  bool __i915_request_submit(struct i915_request *request)
 		goto active;
 	}
 
-	if (unlikely(intel_context_is_banned(request->context)))
+	if (unlikely(!intel_context_is_schedulable(request->context)))
 		i915_request_set_error_once(request, -EIO);
 
 	if (unlikely(fatal_error(request->fence.error)))