[PATCH 10/14] drm/i915/gt: Defer context release to after unlocking pin_mutex

[Chris Wilson]

Atomically acquire the barriers we need to release under the pin_mutex
when unpinning the context, but we can defer the actual work to after
dropping the mutex, reducing the context therein and hopefully
preventing some work when busy.

Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
---
 drivers/gpu/drm/i915/gt/intel_context.c | 10 ++++++++--
 drivers/gpu/drm/i915/i915_active.c      | 15 +++++++++++++--
 drivers/gpu/drm/i915/i915_active.h      |  5 +++++
 3 files changed, 26 insertions(+), 4 deletions(-)

diff --git a/drivers/gpu/drm/i915/gt/intel_context.c b/drivers/gpu/drm/i915/gt/intel_context.c
index c25d2be2ae0f..c9df7853dede 100644
--- a/drivers/gpu/drm/i915/gt/intel_context.c
+++ b/drivers/gpu/drm/i915/gt/intel_context.c
@@ -122,6 +122,8 @@ int __intel_context_do_pin(struct intel_context *ce)
 
 void intel_context_unpin(struct intel_context *ce)
 {
+	struct llist_node *release = ERR_PTR(-1);
+
 	if (likely(atomic_add_unless(&ce->pin_count, -1, 1)))
 		return;
 
@@ -133,11 +135,15 @@ void intel_context_unpin(struct intel_context *ce)
 		CE_TRACE(ce, "retire\n");
 
 		ce->ops->unpin(ce);
+		release = __i915_active_take_barriers(&ce->active);
+	}
+	mutex_unlock(&ce->pin_mutex);
 
-		intel_context_active_release(ce);
+	if (release != ERR_PTR(-1)) {
+		__i915_active_acquire_barriers(&ce->active, release);
+		i915_active_release(&ce->active);
 	}
 
-	mutex_unlock(&ce->pin_mutex);
 	intel_context_put(ce);
 }
 
diff --git a/drivers/gpu/drm/i915/i915_active.c b/drivers/gpu/drm/i915/i915_active.c
index cfe09964622b..4336102595f5 100644
--- a/drivers/gpu/drm/i915/i915_active.c
+++ b/drivers/gpu/drm/i915/i915_active.c
@@ -671,7 +671,8 @@ int i915_active_acquire_preallocate_barrier(struct i915_active *ref,
 	return err;
 }
 
-void i915_active_acquire_barrier(struct i915_active *ref)
+void __i915_active_acquire_barriers(struct i915_active *ref,
+				    struct llist_node *barriers)
 {
 	struct llist_node *pos, *next;
 	unsigned long flags;
@@ -684,7 +685,7 @@ void i915_active_acquire_barrier(struct i915_active *ref)
 	 * populated by i915_request_add_active_barriers() to point to the
 	 * request that will eventually release them.
 	 */
-	llist_for_each_safe(pos, next, take_preallocated_barriers(ref)) {
+	llist_for_each_safe(pos, next, barriers) {
 		struct active_node *node = barrier_from_ll(pos);
 		struct intel_engine_cs *engine = barrier_to_engine(node);
 		struct rb_node **p, *parent;
@@ -714,6 +715,16 @@ void i915_active_acquire_barrier(struct i915_active *ref)
 	}
 }
 
+struct llist_node *__i915_active_take_barriers(struct i915_active *ref)
+{
+	return take_preallocated_barriers(ref);
+}
+
+void i915_active_acquire_barrier(struct i915_active *ref)
+{
+	__i915_active_acquire_barriers(ref, __i915_active_take_barriers(ref));
+}
+
 static struct dma_fence **ll_to_fence_slot(struct llist_node *node)
 {
 	return __active_fence_slot(&barrier_from_ll(node)->base);
diff --git a/drivers/gpu/drm/i915/i915_active.h b/drivers/gpu/drm/i915/i915_active.h
index b571f675c795..2358d2762879 100644
--- a/drivers/gpu/drm/i915/i915_active.h
+++ b/drivers/gpu/drm/i915/i915_active.h
@@ -202,7 +202,12 @@ static inline void i915_active_fini(struct i915_active *ref) { }
 
 int i915_active_acquire_preallocate_barrier(struct i915_active *ref,
 					    struct intel_engine_cs *engine);
+
+struct llist_node *__i915_active_take_barriers(struct i915_active *ref);
+void __i915_active_acquire_barriers(struct i915_active *ref,
+				    struct llist_node *barriers);
 void i915_active_acquire_barrier(struct i915_active *ref);
+
 void i915_request_add_active_barriers(struct i915_request *rq);
 
 void i915_active_print(struct i915_active *ref, struct drm_printer *m);
-- 
2.25.0.rc0