[PATCH 7/7] drm/i915/gem: Wait until the context is finally retired before releasing engines

[Chris Wilson]

If we want to percolate information back from the HW, up through the GEM
context, we need to wait until the intel_context is scheduled out for
the last time. This is handled by the retirement of the intel_context's
barrier, i.e. by listening to the pulse after the notional unpin.

To accommodate this, we need to be able to flush the i915_active's
barriers before awaiting on them. However, this also requires us to
ensure the context is unpinned *before* the barrier request can be
signaled, so mark it as a sentinel.

Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
---
 drivers/gpu/drm/i915/gem/i915_gem_context.c | 17 ++++-----
 drivers/gpu/drm/i915/i915_active.c          | 40 ++++++++++++++++-----
 drivers/gpu/drm/i915/i915_active.h          |  3 +-
 3 files changed, 40 insertions(+), 20 deletions(-)

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_context.c b/drivers/gpu/drm/i915/gem/i915_gem_context.c
index dffe0daee0d2..2a5c40691f0d 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_context.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_context.c
@@ -570,23 +570,20 @@ static void engines_idle_release(struct i915_gem_context *ctx,
 	engines->ctx = i915_gem_context_get(ctx);
 
 	for_each_gem_engine(ce, engines, it) {
-		struct dma_fence *fence;
-		int err = 0;
+		int err;
 
 		/* serialises with execbuf */
 		set_bit(CONTEXT_CLOSED_BIT, &ce->flags);
 		if (!intel_context_pin_if_active(ce))
 			continue;
 
-		fence = i915_active_fence_get(&ce->timeline->last_request);
-		if (fence) {
-			err = i915_sw_fence_await_dma_fence(&engines->fence,
-							    fence, 0,
-							    GFP_KERNEL);
-			dma_fence_put(fence);
-		}
+		/* Wait until context is finally scheduled out and retired */
+		err = i915_sw_fence_await_active(&engines->fence,
+						 &ce->active,
+						 I915_ACTIVE_AWAIT_ACTIVE |
+						 I915_ACTIVE_AWAIT_BARRIER);
 		intel_context_unpin(ce);
-		if (err < 0)
+		if (err)
 			goto kill;
 	}
 
diff --git a/drivers/gpu/drm/i915/i915_active.c b/drivers/gpu/drm/i915/i915_active.c
index 2572dab57d8b..1b48c037363e 100644
--- a/drivers/gpu/drm/i915/i915_active.c
+++ b/drivers/gpu/drm/i915/i915_active.c
@@ -520,19 +520,18 @@ int __i915_active_wait(struct i915_active *ref, int state)
 	return 0;
 }
 
-static int __await_active(struct i915_active_fence *active,
-			  int (*fn)(void *arg, struct dma_fence *fence),
-			  void *arg)
+static int __await_fence(struct i915_active_fence *active,
+			 int (*fn)(void *arg, struct dma_fence *fence),
+			 void *arg)
 {
 	struct dma_fence *fence;
+	int err;
 
-	if (is_barrier(active)) /* XXX flush the barrier? */
+	if (is_barrier(active))
 		return 0;
 
 	fence = i915_active_fence_get(active);
 	if (fence) {
-		int err;
-
 		err = fn(arg, fence);
 		dma_fence_put(fence);
 		if (err < 0)
@@ -542,6 +541,22 @@ static int __await_active(struct i915_active_fence *active,
 	return 0;
 }
 
+static int __await_active(struct active_node *it,
+			  unsigned int flags,
+			  int (*fn)(void *arg, struct dma_fence *fence),
+			  void *arg)
+{
+	int err;
+
+	if (flags & I915_ACTIVE_AWAIT_BARRIER) {
+		err = flush_barrier(it);
+		if (err)
+			return err;
+	}
+
+	return __await_fence(&it->base, fn, arg);
+}
+
 static int await_active(struct i915_active *ref,
 			unsigned int flags,
 			int (*fn)(void *arg, struct dma_fence *fence),
@@ -551,16 +566,20 @@ static int await_active(struct i915_active *ref,
 
 	/* We must always wait for the exclusive fence! */
 	if (rcu_access_pointer(ref->excl.fence)) {
-		err = __await_active(&ref->excl, fn, arg);
+		err = __await_fence(&ref->excl, fn, arg);
 		if (err)
 			return err;
 	}
 
-	if (flags & I915_ACTIVE_AWAIT_ALL && i915_active_acquire_if_busy(ref)) {
+	if (flags & I915_ACTIVE_AWAIT_ACTIVE &&
+	    i915_active_acquire_if_busy(ref)) {
 		struct active_node *it, *n;
 
+		if (flags & I915_ACTIVE_AWAIT_BARRIER)
+			i915_active_acquire_barrier(ref);
+
 		rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
-			err = __await_active(&it->base, fn, arg);
+			err = __await_active(it, flags, fn, arg);
 			if (err)
 				break;
 		}
@@ -854,6 +873,9 @@ void i915_request_add_active_barriers(struct i915_request *rq)
 		list_add_tail((struct list_head *)node, &rq->fence.cb_list);
 	}
 	spin_unlock_irqrestore(&rq->lock, flags);
+
+	/* Ensure that all who came before the barrier are flushed out */
+	__set_bit(I915_FENCE_FLAG_SENTINEL, &rq->fence.flags);
 }
 
 /*
diff --git a/drivers/gpu/drm/i915/i915_active.h b/drivers/gpu/drm/i915/i915_active.h
index 4f9aa7bab514..c7d54d2302a2 100644
--- a/drivers/gpu/drm/i915/i915_active.h
+++ b/drivers/gpu/drm/i915/i915_active.h
@@ -193,7 +193,8 @@ int i915_sw_fence_await_active(struct i915_sw_fence *fence,
 int i915_request_await_active(struct i915_request *rq,
 			      struct i915_active *ref,
 			      unsigned int flags);
-#define I915_ACTIVE_AWAIT_ALL BIT(0)
+#define I915_ACTIVE_AWAIT_ACTIVE BIT(0)
+#define I915_ACTIVE_AWAIT_BARRIER BIT(1)
 
 int i915_active_acquire(struct i915_active *ref);
 bool i915_active_acquire_if_busy(struct i915_active *ref);
-- 
2.20.1