drm/i915: Stop tracking execlists retired requests
authorTvrtko Ursulin <tvrtko@ursulin.net>
Thu, 28 Apr 2016 08:56:58 +0000 (09:56 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Thu, 28 Apr 2016 11:17:32 +0000 (12:17 +0100)
With the previous patch having extended the pinned lifetime of
contexts by referencing the previous context from the current
request until the latter is retired (completed by the GPU),
we can now remove usage of execlist retired queue entirely.

This is because the above now guarantees that all execlist
object access requirements are satisfied by this new tracking,
and we can stop taking additional references and stop keeping
request on the execlists retired queue.

The latter was a source of significant scalability issues in
the driver causing performance hits on some tests. Most
dramatical of which was igt/gem_close_race which had run time
in tens of minutes which is now reduced to tens of seconds.

Signed-off-by: Tvrtko Ursulin <tvrtko@ursulin.net>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: http://patchwork.freedesktop.org/patch/msgid/1461833819-3991-24-git-send-email-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_gem.c
drivers/gpu/drm/i915/intel_lrc.c
drivers/gpu/drm/i915/intel_lrc.h
drivers/gpu/drm/i915/intel_ringbuffer.h

index 72cd08419655910b3e1dcb3dfce041c26b250ccb..411201f94b43b0900e79cb33a23b37bc40dace70 100644 (file)
@@ -2874,13 +2874,7 @@ static void i915_gem_reset_engine_cleanup(struct drm_i915_private *dev_priv,
                /* Ensure irq handler finishes or is cancelled. */
                tasklet_kill(&engine->irq_tasklet);
 
-               spin_lock_bh(&engine->execlist_lock);
-               /* list_splice_tail_init checks for empty lists */
-               list_splice_tail_init(&engine->execlist_queue,
-                                     &engine->execlist_retired_req_list);
-               spin_unlock_bh(&engine->execlist_lock);
-
-               intel_execlists_retire_requests(engine);
+               intel_execlists_cancel_requests(engine);
        }
 
        /*
@@ -3004,8 +2998,6 @@ i915_gem_retire_requests(struct drm_device *dev)
                        spin_lock_bh(&engine->execlist_lock);
                        idle &= list_empty(&engine->execlist_queue);
                        spin_unlock_bh(&engine->execlist_lock);
-
-                       intel_execlists_retire_requests(engine);
                }
        }
 
index dc25cfe7bfc06d74f02a9a21e4a85807ba40868e..874c2515f9d4f35d9a7fe327242c93175a95f4d4 100644 (file)
@@ -435,8 +435,8 @@ static void execlists_context_unqueue(struct intel_engine_cs *engine)
                        /* Same ctx: ignore first request, as second request
                         * will update tail past first request's workload */
                        cursor->elsp_submitted = req0->elsp_submitted;
-                       list_move_tail(&req0->execlist_link,
-                                      &engine->execlist_retired_req_list);
+                       list_del(&req0->execlist_link);
+                       i915_gem_request_unreference(req0);
                        req0 = cursor;
                } else {
                        req1 = cursor;
@@ -468,7 +468,7 @@ static void execlists_context_unqueue(struct intel_engine_cs *engine)
 }
 
 static unsigned int
-execlists_check_remove_request(struct intel_engine_cs *engine, u32 request_id)
+execlists_check_remove_request(struct intel_engine_cs *engine, u32 ctx_id)
 {
        struct drm_i915_gem_request *head_req;
 
@@ -478,19 +478,16 @@ execlists_check_remove_request(struct intel_engine_cs *engine, u32 request_id)
                                            struct drm_i915_gem_request,
                                            execlist_link);
 
-       if (!head_req)
-               return 0;
-
-       if (unlikely(head_req->ctx_hw_id != request_id))
-               return 0;
+       if (WARN_ON(!head_req || (head_req->ctx_hw_id != ctx_id)))
+               return 0;
 
        WARN(head_req->elsp_submitted == 0, "Never submitted head request\n");
 
        if (--head_req->elsp_submitted > 0)
                return 0;
 
-       list_move_tail(&head_req->execlist_link,
-                      &engine->execlist_retired_req_list);
+       list_del(&head_req->execlist_link);
+       i915_gem_request_unreference(head_req);
 
        return 1;
 }
@@ -594,9 +591,6 @@ static void execlists_context_queue(struct drm_i915_gem_request *request)
        struct drm_i915_gem_request *cursor;
        int num_elements = 0;
 
-       intel_lr_context_pin(request->ctx, request->engine);
-       i915_gem_request_reference(request);
-
        spin_lock_bh(&engine->execlist_lock);
 
        list_for_each_entry(cursor, &engine->execlist_queue, execlist_link)
@@ -613,11 +607,12 @@ static void execlists_context_queue(struct drm_i915_gem_request *request)
                if (request->ctx == tail_req->ctx) {
                        WARN(tail_req->elsp_submitted != 0,
                                "More than 2 already-submitted reqs queued\n");
-                       list_move_tail(&tail_req->execlist_link,
-                                      &engine->execlist_retired_req_list);
+                       list_del(&tail_req->execlist_link);
+                       i915_gem_request_unreference(tail_req);
                }
        }
 
+       i915_gem_request_reference(request);
        list_add_tail(&request->execlist_link, &engine->execlist_queue);
        request->ctx_hw_id = request->ctx->hw_id;
        if (num_elements == 0)
@@ -883,23 +878,18 @@ int intel_execlists_submission(struct i915_execbuffer_params *params,
        return 0;
 }
 
-void intel_execlists_retire_requests(struct intel_engine_cs *engine)
+void intel_execlists_cancel_requests(struct intel_engine_cs *engine)
 {
        struct drm_i915_gem_request *req, *tmp;
-       struct list_head retired_list;
+       LIST_HEAD(cancel_list);
 
        WARN_ON(!mutex_is_locked(&engine->dev->struct_mutex));
-       if (list_empty(&engine->execlist_retired_req_list))
-               return;
 
-       INIT_LIST_HEAD(&retired_list);
        spin_lock_bh(&engine->execlist_lock);
-       list_replace_init(&engine->execlist_retired_req_list, &retired_list);
+       list_replace_init(&engine->execlist_queue, &cancel_list);
        spin_unlock_bh(&engine->execlist_lock);
 
-       list_for_each_entry_safe(req, tmp, &retired_list, execlist_link) {
-               intel_lr_context_unpin(req->ctx, engine);
-
+       list_for_each_entry_safe(req, tmp, &cancel_list, execlist_link) {
                list_del(&req->execlist_link);
                i915_gem_request_unreference(req);
        }
@@ -1993,7 +1983,6 @@ logical_ring_init(struct drm_device *dev, struct intel_engine_cs *engine)
 
        INIT_LIST_HEAD(&engine->buffers);
        INIT_LIST_HEAD(&engine->execlist_queue);
-       INIT_LIST_HEAD(&engine->execlist_retired_req_list);
        spin_lock_init(&engine->execlist_lock);
 
        tasklet_init(&engine->irq_tasklet,
index 28ff324ba5dc8ac3b79621c55b8b2d8f8f8019b5..229b8a9742620e856fa06cfecadbf26e874c201b 100644 (file)
@@ -118,6 +118,6 @@ int intel_execlists_submission(struct i915_execbuffer_params *params,
                               struct drm_i915_gem_execbuffer2 *args,
                               struct list_head *vmas);
 
-void intel_execlists_retire_requests(struct intel_engine_cs *engine);
+void intel_execlists_cancel_requests(struct intel_engine_cs *engine);
 
 #endif /* _INTEL_LRC_H_ */
index 038914ccc6fde82abe5a4f84e2dd3ca133b954f5..7023e88531b5bd4e86cff9cca5b2079f7ba92a38 100644 (file)
@@ -266,7 +266,6 @@ struct  intel_engine_cs {
        struct tasklet_struct irq_tasklet;
        spinlock_t execlist_lock; /* used inside tasklet, use spin_lock_bh */
        struct list_head execlist_queue;
-       struct list_head execlist_retired_req_list;
        unsigned int fw_domains;
        unsigned int next_context_status_buffer;
        unsigned int idle_lite_restore_wa;