Intel-GFX Archive on lore.kernel.org
 help / color / Atom feed
From: Chris Wilson <chris@chris-wilson.co.uk>
To: intel-gfx@lists.freedesktop.org
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Subject: [Intel-gfx] [PATCH 07/42] drm/i915/gt: Split the breadcrumb spinlock between global and contexts
Date: Sun,  2 Aug 2020 17:43:37 +0100
Message-ID: <20200802164412.2738-8-chris@chris-wilson.co.uk> (raw)
In-Reply-To: <20200802164412.2738-1-chris@chris-wilson.co.uk>

As we funnel more and more contexts into the breadcrumbs on an engine,
the hold time of b->irq_lock grows. As we may then contend with the
b->irq_lock during request submission, this increases the burden upon
the engine->active.lock and so directly impacts both our execution
latency and client latency. If we split the b->irq_lock by introducing a
per-context spinlock to manage the signalers within a context, we then
only need the b->irq_lock for enabling/disabling the interrupt and can
avoid taking the lock for walking the list of contexts within the signal
worker. Even with the current setup, this greatly reduces the number of
times we have to take and fight for b->irq_lock.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
---
 drivers/gpu/drm/i915/gt/intel_breadcrumbs.c   | 154 ++++++++++--------
 drivers/gpu/drm/i915/gt/intel_context.c       |   1 +
 drivers/gpu/drm/i915/gt/intel_context_types.h |   1 +
 3 files changed, 86 insertions(+), 70 deletions(-)

diff --git a/drivers/gpu/drm/i915/gt/intel_breadcrumbs.c b/drivers/gpu/drm/i915/gt/intel_breadcrumbs.c
index 9e7ac612fabb..98d323354082 100644
--- a/drivers/gpu/drm/i915/gt/intel_breadcrumbs.c
+++ b/drivers/gpu/drm/i915/gt/intel_breadcrumbs.c
@@ -103,17 +103,19 @@ static void __intel_breadcrumbs_disarm_irq(struct intel_breadcrumbs *b)
 static void add_signaling_context(struct intel_breadcrumbs *b,
 				  struct intel_context *ce)
 {
-	intel_context_get(ce);
-	list_add_tail(&ce->signal_link, &b->signalers);
-	if (list_is_first(&ce->signal_link, &b->signalers))
+	lockdep_assert_held(&b->irq_lock);
+
+	list_add_rcu(&ce->signal_link, &b->signalers);
+	if (list_is_last(&ce->signal_link, &b->signalers))
 		__intel_breadcrumbs_arm_irq(b);
 }
 
 static void remove_signaling_context(struct intel_breadcrumbs *b,
 				     struct intel_context *ce)
 {
-	list_del(&ce->signal_link);
-	intel_context_put(ce);
+	spin_lock(&b->irq_lock);
+	list_del_rcu(&ce->signal_link);
+	spin_unlock(&b->irq_lock);
 }
 
 static inline bool __request_completed(const struct i915_request *rq)
@@ -189,15 +191,12 @@ static void signal_irq_work(struct irq_work *work)
 	struct intel_breadcrumbs *b = container_of(work, typeof(*b), irq_work);
 	const ktime_t timestamp = ktime_get();
 	struct llist_node *signal, *sn;
-	struct intel_context *ce, *cn;
-	struct list_head *pos, *next;
+	struct intel_context *ce;
 
 	signal = NULL;
 	if (unlikely(!llist_empty(&b->signaled_requests)))
 		signal = llist_del_all(&b->signaled_requests);
 
-	spin_lock(&b->irq_lock);
-
 	/*
 	 * Keep the irq armed until the interrupt after all listeners are gone.
 	 *
@@ -221,11 +220,24 @@ static void signal_irq_work(struct irq_work *work)
 	 * interrupt draw less ire from other users of the system and tools
 	 * like powertop.
 	 */
-	if (!signal && list_empty(&b->signalers))
-		__intel_breadcrumbs_disarm_irq(b);
+	if (!signal && READ_ONCE(b->irq_armed) && list_empty(&b->signalers)) {
+		if (spin_trylock(&b->irq_lock)) {
+			if (list_empty(&b->signalers))
+				__intel_breadcrumbs_disarm_irq(b);
+			spin_unlock(&b->irq_lock);
+		}
+	}
+
+	rcu_read_lock();
+	list_for_each_entry_rcu(ce, &b->signalers, signal_link) {
+		struct list_head *pos, *next;
+		bool release = false;
 
-	list_for_each_entry_safe(ce, cn, &b->signalers, signal_link) {
-		GEM_BUG_ON(list_empty(&ce->signals));
+		if (!spin_trylock(&ce->signal_lock))
+			continue;
+
+		if (list_empty(&ce->signals))
+			goto unlock;
 
 		list_for_each_safe(pos, next, &ce->signals) {
 			struct i915_request *rq =
@@ -258,11 +270,16 @@ static void signal_irq_work(struct irq_work *work)
 			if (&ce->signals == pos) { /* now empty */
 				add_retire(b, ce->timeline);
 				remove_signaling_context(b, ce);
+				release = true;
 			}
 		}
-	}
 
-	spin_unlock(&b->irq_lock);
+unlock:
+		spin_unlock(&ce->signal_lock);
+		if (release)
+			intel_context_put(ce);
+	}
+	rcu_read_unlock();
 
 	llist_for_each_safe(signal, sn, signal) {
 		struct i915_request *rq =
@@ -336,9 +353,9 @@ void intel_breadcrumbs_free(struct intel_breadcrumbs *b)
 	kfree(b);
 }
 
-static void insert_breadcrumb(struct i915_request *rq,
-			      struct intel_breadcrumbs *b)
+static void insert_breadcrumb(struct i915_request *rq)
 {
+	struct intel_breadcrumbs *b = READ_ONCE(rq->engine)->breadcrumbs;
 	struct intel_context *ce = rq->context;
 	struct list_head *pos;
 
@@ -360,7 +377,33 @@ static void insert_breadcrumb(struct i915_request *rq,
 	}
 
 	if (list_empty(&ce->signals)) {
+		/*
+		 * rq->engine is locked by rq->engine->active.lock. That
+		 * however is not known until after rq->engine has been
+		 * dereferenced and the lock acquired. Hence we acquire the
+		 * lock and then validate that rq->engine still matches the
+		 * lock we hold for it.
+		 *
+		 * Here, we are using the breadcrumb lock as a proxy for the
+		 * rq->engine->active.lock, and we know that since the
+		 * breadcrumb will be serialised within i915_request_submit
+		 * the engine cannot change while active as long as we hold
+		 * the breadcrumb lock on that engine.
+		 *
+		 * From the dma_fence_enable_signaling() path, we are outside
+		 * of the request submit/unsubmit path, and so we must be more
+		 * careful to acquire the right lock.
+		 */
+		intel_context_get(ce);
+		spin_lock(&b->irq_lock);
+		while (unlikely(b != READ_ONCE(rq->engine)->breadcrumbs)) {
+			spin_unlock(&b->irq_lock);
+			b = READ_ONCE(rq->engine)->breadcrumbs;
+			spin_lock(&b->irq_lock);
+		}
 		add_signaling_context(b, ce);
+		spin_unlock(&b->irq_lock);
+
 		pos = &ce->signals;
 	} else {
 		/*
@@ -396,7 +439,7 @@ static void insert_breadcrumb(struct i915_request *rq,
 
 bool i915_request_enable_breadcrumb(struct i915_request *rq)
 {
-	struct intel_breadcrumbs *b;
+	struct intel_context *ce = rq->context;
 
 	/* Serialises with i915_request_retire() using rq->lock */
 	if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &rq->fence.flags))
@@ -411,67 +454,37 @@ bool i915_request_enable_breadcrumb(struct i915_request *rq)
 	if (!test_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags))
 		return true;
 
-	/*
-	 * rq->engine is locked by rq->engine->active.lock. That however
-	 * is not known until after rq->engine has been dereferenced and
-	 * the lock acquired. Hence we acquire the lock and then validate
-	 * that rq->engine still matches the lock we hold for it.
-	 *
-	 * Here, we are using the breadcrumb lock as a proxy for the
-	 * rq->engine->active.lock, and we know that since the breadcrumb
-	 * will be serialised within i915_request_submit/i915_request_unsubmit,
-	 * the engine cannot change while active as long as we hold the
-	 * breadcrumb lock on that engine.
-	 *
-	 * From the dma_fence_enable_signaling() path, we are outside of the
-	 * request submit/unsubmit path, and so we must be more careful to
-	 * acquire the right lock.
-	 */
-	b = READ_ONCE(rq->engine)->breadcrumbs;
-	spin_lock(&b->irq_lock);
-	while (unlikely(b != READ_ONCE(rq->engine)->breadcrumbs)) {
-		spin_unlock(&b->irq_lock);
-		b = READ_ONCE(rq->engine)->breadcrumbs;
-		spin_lock(&b->irq_lock);
-	}
-
-	/*
-	 * Now that we are finally serialised with request submit/unsubmit,
-	 * [with b->irq_lock] and with i915_request_retire() [via checking
-	 * SIGNALED with rq->lock] confirm the request is indeed active. If
-	 * it is no longer active, the breadcrumb will be attached upon
-	 * i915_request_submit().
-	 */
+	spin_lock(&ce->signal_lock);
 	if (test_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags))
-		insert_breadcrumb(rq, b);
-
-	spin_unlock(&b->irq_lock);
+		insert_breadcrumb(rq);
+	spin_unlock(&ce->signal_lock);
 
 	return true;
 }
 
 void i915_request_cancel_breadcrumb(struct i915_request *rq)
 {
-	struct intel_breadcrumbs *b = rq->engine->breadcrumbs;
+	struct intel_context *ce = rq->context;
+	bool release = false;
 
-	/*
-	 * We must wait for b->irq_lock so that we know the interrupt handler
-	 * has released its reference to the intel_context and has completed
-	 * the DMA_FENCE_FLAG_SIGNALED_BIT/I915_FENCE_FLAG_SIGNAL dance (if
-	 * required).
-	 */
-	spin_lock(&b->irq_lock);
+	if (!test_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags))
+		return;
+
+	spin_lock(&ce->signal_lock);
 	if (test_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags)) {
-		struct intel_context *ce = rq->context;
+		clear_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags);
 
 		list_del(&rq->signal_link);
-		if (list_empty(&ce->signals))
-			remove_signaling_context(b, ce);
+		if (list_empty(&ce->signals)) {
+			remove_signaling_context(rq->engine->breadcrumbs, ce);
+			release = true;
+		}
 
-		clear_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags);
 		i915_request_put(rq);
 	}
-	spin_unlock(&b->irq_lock);
+	spin_unlock(&ce->signal_lock);
+	if (release)
+		intel_context_put(ce);
 }
 
 static void print_signals(struct intel_breadcrumbs *b, struct drm_printer *p)
@@ -481,18 +494,19 @@ static void print_signals(struct intel_breadcrumbs *b, struct drm_printer *p)
 
 	drm_printf(p, "Signals:\n");
 
-	spin_lock_irq(&b->irq_lock);
-	list_for_each_entry(ce, &b->signalers, signal_link) {
-		list_for_each_entry(rq, &ce->signals, signal_link) {
+	rcu_read_lock();
+	list_for_each_entry_rcu(ce, &b->signalers, signal_link) {
+		spin_lock_irq(&ce->signal_lock);
+		list_for_each_entry(rq, &ce->signals, signal_link)
 			drm_printf(p, "\t[%llx:%llx%s] @ %dms\n",
 				   rq->fence.context, rq->fence.seqno,
 				   i915_request_completed(rq) ? "!" :
 				   i915_request_started(rq) ? "*" :
 				   "",
 				   jiffies_to_msecs(jiffies - rq->emitted_jiffies));
-		}
+		spin_unlock_irq(&ce->signal_lock);
 	}
-	spin_unlock_irq(&b->irq_lock);
+	rcu_read_unlock();
 }
 
 void intel_engine_print_breadcrumbs(struct intel_engine_cs *engine,
diff --git a/drivers/gpu/drm/i915/gt/intel_context.c b/drivers/gpu/drm/i915/gt/intel_context.c
index 4e7924640ffa..cde356c7754d 100644
--- a/drivers/gpu/drm/i915/gt/intel_context.c
+++ b/drivers/gpu/drm/i915/gt/intel_context.c
@@ -147,6 +147,7 @@ static void __intel_context_ctor(void *arg)
 {
 	struct intel_context *ce = arg;
 
+	spin_lock_init(&ce->signal_lock);
 	INIT_LIST_HEAD(&ce->signal_link);
 	INIT_LIST_HEAD(&ce->signals);
 
diff --git a/drivers/gpu/drm/i915/gt/intel_context_types.h b/drivers/gpu/drm/i915/gt/intel_context_types.h
index 4954b0df4864..a78c1c225ce3 100644
--- a/drivers/gpu/drm/i915/gt/intel_context_types.h
+++ b/drivers/gpu/drm/i915/gt/intel_context_types.h
@@ -51,6 +51,7 @@ struct intel_context {
 	struct i915_address_space *vm;
 	struct i915_gem_context __rcu *gem_context;
 
+	spinlock_t signal_lock;
 	struct list_head signal_link;
 	struct list_head signals;
 
-- 
2.20.1

_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

  parent reply index

Thread overview: 49+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2020-08-02 16:43 [Intel-gfx] Time, where did it go? Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 01/42] drm/i915: Fix wrong return value Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 02/42] drm/i915/gem: Don't drop the timeline lock during execbuf Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 03/42] drm/i915/gem: Reduce context termination list iteration guard to RCU Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 04/42] drm/i915/gt: Protect context lifetime with RCU Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 05/42] drm/i915/gt: Free stale request on destroying the virtual engine Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 06/42] drm/i915/gt: Track signaled breadcrumbs outside of the breadcrumb spinlock Chris Wilson
2020-08-02 16:43 ` Chris Wilson [this message]
2020-08-02 16:43 ` [Intel-gfx] [PATCH 08/42] drm/i915: Drop i915_request.lock serialisation around await_start Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 09/42] drm/i915: Drop i915_request.lock requirement for intel_rps_boost() Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 10/42] drm/i915/gem: Reduce ctx->engine_mutex for reading the clone source Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 11/42] drm/i915/gem: Reduce ctx->engines_mutex for get_engines() Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 12/42] drm/i915: Reduce test_and_set_bit to set_bit in i915_request_submit() Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 13/42] drm/i915/gt: Decouple completed requests on unwind Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 14/42] drm/i915/gt: Check for a completed last request once Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 15/42] drm/i915/gt: Refactor heartbeat request construction and submission Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 16/42] drm/i915/gt: Replace direct submit with direct call to tasklet Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 17/42] drm/i915/gt: Use virtual_engine during execlists_dequeue Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 18/42] drm/i915/gt: Decouple inflight virtual engines Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 19/42] drm/i915/gt: Defer schedule_out until after the next dequeue Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 20/42] drm/i915/gt: Resubmit the virtual engine on schedule-out Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 21/42] drm/i915/gt: Simplify virtual engine handling for execlists_hold() Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 22/42] drm/i915/gt: ce->inflight updates are now serialised Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 23/42] drm/i915/gt: Drop atomic for engine->fw_active tracking Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 24/42] drm/i915/gt: Extract busy-stats for ring-scheduler Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 25/42] drm/i915/gt: Convert stats.active to plain unsigned int Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 26/42] drm/i915: Lift waiter/signaler iterators Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 27/42] drm/i915: Strip out internal priorities Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 28/42] drm/i915: Remove I915_USER_PRIORITY_SHIFT Chris Wilson
2020-08-02 16:43 ` [Intel-gfx] [PATCH 29/42] drm/i915/gt: Defer the kmem_cache_free() until after the HW submit Chris Wilson
2020-08-02 16:44 ` [Intel-gfx] [PATCH 30/42] drm/i915: Prune empty priolists Chris Wilson
2020-08-02 16:44 ` [Intel-gfx] [PATCH 31/42] drm/i915: Replace engine->schedule() with a known request operation Chris Wilson
2020-08-02 16:44 ` [Intel-gfx] [PATCH 32/42] drm/i915/gt: Do not suspend bonded requests if one hangs Chris Wilson
2020-08-02 16:44 ` [Intel-gfx] [PATCH 33/42] drm/i915: Teach the i915_dependency to use a double-lock Chris Wilson
2020-08-02 16:44 ` [Intel-gfx] [PATCH 34/42] drm/i915: Restructure priority inheritance Chris Wilson
2020-08-02 16:44 ` [Intel-gfx] [PATCH 35/42] drm/i915/selftests: Measure set-priority duration Chris Wilson
2020-08-02 16:44 ` [Intel-gfx] [PATCH 36/42] drm/i915: Improve DFS for priority inheritance Chris Wilson
2020-08-02 16:44 ` [Intel-gfx] [PATCH 37/42] drm/i915/gt: Remove timeslice suppression Chris Wilson
2020-08-02 16:44 ` [Intel-gfx] [PATCH 38/42] drm/i915: Fair low-latency scheduling Chris Wilson
2020-08-02 16:44 ` [Intel-gfx] [PATCH 39/42] drm/i915/gt: Specify a deadline for the heartbeat Chris Wilson
2020-08-02 16:44 ` [Intel-gfx] [PATCH 40/42] drm/i915: Replace the priority boosting for the display with a deadline Chris Wilson
2020-08-02 16:44 ` [Intel-gfx] [PATCH 41/42] drm/i915: Move saturated workload detection back to the context Chris Wilson
2020-08-02 16:44 ` [Intel-gfx] [PATCH 42/42] drm/i915/gt: Another tweak for flushing the tasklets Chris Wilson
2020-08-02 17:14 ` [Intel-gfx] ✗ Fi.CI.CHECKPATCH: warning for series starting with [01/42] drm/i915: Fix wrong return value Patchwork
2020-08-02 17:14 ` [Intel-gfx] ✗ Fi.CI.SPARSE: " Patchwork
2020-08-02 17:34 ` [Intel-gfx] ✓ Fi.CI.BAT: success " Patchwork
2020-08-02 17:56 ` [Intel-gfx] Time, where did it go? Dave Airlie
2020-08-02 19:36   ` Chris Wilson
2020-08-02 21:03 ` [Intel-gfx] ✗ Fi.CI.IGT: failure for series starting with [01/42] drm/i915: Fix wrong return value Patchwork

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20200802164412.2738-8-chris@chris-wilson.co.uk \
    --to=chris@chris-wilson.co.uk \
    --cc=intel-gfx@lists.freedesktop.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

Intel-GFX Archive on lore.kernel.org

Archives are clonable:
	git clone --mirror https://lore.kernel.org/intel-gfx/0 intel-gfx/git/0.git
	git clone --mirror https://lore.kernel.org/intel-gfx/1 intel-gfx/git/1.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 intel-gfx intel-gfx/ https://lore.kernel.org/intel-gfx \
		intel-gfx@lists.freedesktop.org
	public-inbox-index intel-gfx

Example config snippet for mirrors

Newsgroup available over NNTP:
	nntp://nntp.lore.kernel.org/org.freedesktop.lists.intel-gfx


AGPL code for this site: git clone https://public-inbox.org/public-inbox.git