[PATCH 13/33] drm/i915: Improve DFS for priority inheritance

[Chris Wilson]

The core of the scheduling algorithm is that we compute the topological
order of the fence DAG. Knowing that we have a DAG, we should be able to
use a DFS to compute the topological sort in linear time. However,
during the conversion of the recursive algorithm into an iterative one,
the memoization of how far we had progressed down a branch was
forgotten. The result was that instead of running in linear time, it was
running in geometric time and could easily run for a few hundred
milliseconds given a wide enough graph, not the microseconds as required.

Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
Reviewed-by: Andi Shyti <andi.shyti at intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
---
 drivers/gpu/drm/i915/i915_scheduler.c       | 58 ++++++++++++---------
 drivers/gpu/drm/i915/i915_scheduler_types.h |  6 ++-
 2 files changed, 39 insertions(+), 25 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_scheduler.c b/drivers/gpu/drm/i915/i915_scheduler.c
index a56a812cbf29..cea5129560a5 100644
--- a/drivers/gpu/drm/i915/i915_scheduler.c
+++ b/drivers/gpu/drm/i915/i915_scheduler.c
@@ -242,6 +242,26 @@ void __i915_priolist_free(struct i915_priolist *p)
 	kmem_cache_free(global.slab_priorities, p);
 }
 
+static struct i915_request *
+stack_push(struct i915_request *rq,
+	   struct i915_request *prev,
+	   struct list_head *pos)
+{
+	prev->sched.dfs.pos = pos;
+	rq->sched.dfs.prev = prev;
+	return rq;
+}
+
+static struct i915_request *
+stack_pop(struct i915_request *rq,
+	  struct list_head **pos)
+{
+	rq = rq->sched.dfs.prev;
+	if (rq)
+		*pos = rq->sched.dfs.pos;
+	return rq;
+}
+
 static inline bool need_preempt(int prio, int active)
 {
 	/*
@@ -306,11 +326,10 @@ static void ipi_priority(struct i915_request *rq, int prio)
 static void __i915_request_set_priority(struct i915_request *rq, int prio)
 {
 	struct intel_engine_cs *engine = rq->engine;
-	struct i915_request *rn;
+	struct list_head *pos = &rq->sched.signalers_list;
 	struct list_head *plist;
-	LIST_HEAD(dfs);
 
-	list_add(&rq->sched.dfs, &dfs);
+	plist = i915_sched_lookup_priolist(engine, prio);
 
 	/*
 	 * Recursively bump all dependent priorities to match the new request.
@@ -330,40 +349,31 @@ static void __i915_request_set_priority(struct i915_request *rq, int prio)
 	 * end result is a topological list of requests in reverse order, the
 	 * last element in the list is the request we must execute first.
 	 */
-	list_for_each_entry(rq, &dfs, sched.dfs) {
-		struct i915_dependency *p;
-
-		/* Also release any children on this engine that are ready */
-		GEM_BUG_ON(rq->engine != engine);
-
-		for_each_signaler(p, rq) {
+	rq->sched.dfs.prev = NULL;
+	do {
+		list_for_each_continue(pos, &rq->sched.signalers_list) {
+			struct i915_dependency *p =
+				list_entry(pos, typeof(*p), signal_link);
 			struct i915_request *s =
 				container_of(p->signaler, typeof(*s), sched);
 
-			GEM_BUG_ON(s == rq);
-
 			if (rq_prio(s) >= prio)
 				continue;
 
 			if (__i915_request_is_complete(s))
 				continue;
 
-			if (s->engine != rq->engine) {
+			if (s->engine != engine) {
 				ipi_priority(s, prio);
 				continue;
 			}
 
-			list_move_tail(&s->sched.dfs, &dfs);
+			/* Remember our position along this branch */
+			rq = stack_push(s, rq, pos);
+			pos = &rq->sched.signalers_list;
 		}
-	}
 
-	plist = i915_sched_lookup_priolist(engine, prio);
-
-	/* Fifo and depth-first replacement ensure our deps execute first */
-	list_for_each_entry_safe_reverse(rq, rn, &dfs, sched.dfs) {
-		GEM_BUG_ON(rq->engine != engine);
-
-		INIT_LIST_HEAD(&rq->sched.dfs);
+		RQ_TRACE(rq, "set-priority:%d\n", prio);
 		WRITE_ONCE(rq->sched.attr.priority, prio);
 
 		/*
@@ -377,12 +387,13 @@ static void __i915_request_set_priority(struct i915_request *rq, int prio)
 		if (!i915_request_is_ready(rq))
 			continue;
 
+		GEM_BUG_ON(rq->engine != engine);
 		if (i915_request_in_priority_queue(rq))
 			list_move_tail(&rq->sched.link, plist);
 
 		/* Defer (tasklet) submission until after all updates. */
 		kick_submission(engine, rq, prio);
-	}
+	} while ((rq = stack_pop(rq, &pos)));
 }
 
 #define all_signalers_checked(p, rq) \
@@ -456,7 +467,6 @@ void i915_sched_node_init(struct i915_sched_node *node)
 	INIT_LIST_HEAD(&node->signalers_list);
 	INIT_LIST_HEAD(&node->waiters_list);
 	INIT_LIST_HEAD(&node->link);
-	INIT_LIST_HEAD(&node->dfs);
 
 	node->ipi_link = NULL;
 
diff --git a/drivers/gpu/drm/i915/i915_scheduler_types.h b/drivers/gpu/drm/i915/i915_scheduler_types.h
index e5eabff6f20b..60a3aa743ab8 100644
--- a/drivers/gpu/drm/i915/i915_scheduler_types.h
+++ b/drivers/gpu/drm/i915/i915_scheduler_types.h
@@ -69,7 +69,11 @@ struct i915_sched_node {
 	struct list_head signalers_list; /* those before us, we depend upon */
 	struct list_head waiters_list; /* those after us, they depend upon us */
 	struct list_head link; /* guarded by engine->active.lock */
-	struct list_head dfs; /* guarded by engine->active.lock */
+	struct i915_sched_stack {
+		/* Branch memoization used during depth-first search */
+		struct i915_request *prev;
+		struct list_head *pos;
+	} dfs; /* guarded by engine->active.lock */
 	struct i915_sched_attr attr;
 	unsigned long flags;
 #define I915_SCHED_HAS_EXTERNAL_CHAIN	BIT(0)
-- 
2.20.1