[Intel-gfx] [PATCH 07/15] drm/i915: Hold a reference on the request for its fence chain

[Chris Wilson]

Currently, we have an active reference for the request until it is
retired. Though it cannot be retired before it has been executed by
hardware, the request may be completed before we have finished
processing the execute fence, i.e. we may continue to process that fence
as we free the request.

Fixes: 5590af3e115a ("drm/i915: Drive request submission through fence callbacks")
Fixes: 23902e49c999 ("drm/i915: Split request submit/execute phase into two")
Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
---
 drivers/gpu/drm/i915/i915_gem_request.c | 34 ++++++++++++++++++++++++++-------
 drivers/gpu/drm/i915/i915_sw_fence.h    |  5 +++++
 2 files changed, 32 insertions(+), 7 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_gem_request.c b/drivers/gpu/drm/i915/i915_gem_request.c
index ed6cead22a86..94d71b639f12 100644
--- a/drivers/gpu/drm/i915/i915_gem_request.c
+++ b/drivers/gpu/drm/i915/i915_gem_request.c
@@ -200,8 +200,8 @@ static void i915_gem_request_retire(struct drm_i915_gem_request *request)
 	struct i915_gem_active *active, *next;
 
 	lockdep_assert_held(&request->i915->drm.struct_mutex);
-	GEM_BUG_ON(!i915_sw_fence_done(&request->submit));
-	GEM_BUG_ON(!i915_sw_fence_done(&request->execute));
+	GEM_BUG_ON(!i915_sw_fence_signaled(&request->submit));
+	GEM_BUG_ON(!i915_sw_fence_signaled(&request->execute));
 	GEM_BUG_ON(!i915_gem_request_completed(request));
 	GEM_BUG_ON(!request->i915->gt.active_requests);
 
@@ -451,11 +451,17 @@ void i915_gem_request_submit(struct drm_i915_gem_request *request)
 static int __i915_sw_fence_call
 submit_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
 {
-	if (state == FENCE_COMPLETE) {
-		struct drm_i915_gem_request *request =
-			container_of(fence, typeof(*request), submit);
+	struct drm_i915_gem_request *request =
+		container_of(fence, typeof(*request), submit);
 
+	switch (state) {
+	case FENCE_COMPLETE:
 		request->engine->submit_request(request);
+		break;
+
+	case FENCE_FREE:
+		i915_gem_request_put(request);
+		break;
 	}
 
 	return NOTIFY_DONE;
@@ -464,6 +470,18 @@ submit_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
 static int __i915_sw_fence_call
 execute_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
 {
+	struct drm_i915_gem_request *request =
+		container_of(fence, typeof(*request), execute);
+
+	switch (state) {
+	case FENCE_COMPLETE:
+		break;
+
+	case FENCE_FREE:
+		i915_gem_request_put(request);
+		break;
+	}
+
 	return NOTIFY_DONE;
 }
 
@@ -551,8 +569,10 @@ i915_gem_request_alloc(struct intel_engine_cs *engine,
 		       req->timeline->fence_context,
 		       __timeline_get_seqno(req->timeline->common));
 
-	i915_sw_fence_init(&req->submit, submit_notify);
-	i915_sw_fence_init(&req->execute, execute_notify);
+	/* We bump the ref for the fence chain */
+	i915_sw_fence_init(&i915_gem_request_get(req)->submit, submit_notify);
+	i915_sw_fence_init(&i915_gem_request_get(req)->execute, execute_notify);
+
 	/* Ensure that the execute fence completes after the submit fence -
 	 * as we complete the execute fence from within the submit fence
 	 * callback, its completion would otherwise be visible first.
diff --git a/drivers/gpu/drm/i915/i915_sw_fence.h b/drivers/gpu/drm/i915/i915_sw_fence.h
index 7508d23f823b..0f3185ef7f4e 100644
--- a/drivers/gpu/drm/i915/i915_sw_fence.h
+++ b/drivers/gpu/drm/i915/i915_sw_fence.h
@@ -75,6 +75,11 @@ int i915_sw_fence_await_reservation(struct i915_sw_fence *fence,
 				    unsigned long timeout,
 				    gfp_t gfp);
 
+static inline bool i915_sw_fence_signaled(const struct i915_sw_fence *fence)
+{
+	return atomic_read(&fence->pending) <= 0;
+}
+
 static inline bool i915_sw_fence_done(const struct i915_sw_fence *fence)
 {
 	return atomic_read(&fence->pending) < 0;
-- 
2.10.2