drm/i915/execlists: Refactor out can_merge_rq()
authorChris Wilson <chris@chris-wilson.co.uk>
Fri, 8 Feb 2019 23:51:08 +0000 (23:51 +0000)
committerChris Wilson <chris@chris-wilson.co.uk>
Sat, 9 Feb 2019 00:31:41 +0000 (00:31 +0000)
In the next patch, we add another user that wants to check whether
requests can be merge into a single HW execution, and in the future we
want to add more conditions under which requests from the same context
cannot be merge. In preparation, extract out can_merge_rq().

v2: Reorder tests to decide if we can continue filling ELSP and bonus
comments.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190208235108.23127-1-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/intel_lrc.c

index 1b567a3f006a14fc418001a71735e5baac52f4dc..a62791a30c7eefe634b0bb31cf086471fb423710 100644 (file)
@@ -254,12 +254,11 @@ static inline bool need_preempt(const struct intel_engine_cs *engine,
 }
 
 __maybe_unused static inline bool
-assert_priority_queue(const struct intel_engine_execlists *execlists,
-                     const struct i915_request *prev,
+assert_priority_queue(const struct i915_request *prev,
                      const struct i915_request *next)
 {
-       if (!prev)
-               return true;
+       const struct intel_engine_execlists *execlists =
+               &prev->engine->execlists;
 
        /*
         * Without preemption, the prev may refer to the still active element
@@ -564,6 +563,17 @@ static bool can_merge_ctx(const struct intel_context *prev,
        return true;
 }
 
+static bool can_merge_rq(const struct i915_request *prev,
+                        const struct i915_request *next)
+{
+       GEM_BUG_ON(!assert_priority_queue(prev, next));
+
+       if (!can_merge_ctx(prev->hw_context, next->hw_context))
+               return false;
+
+       return true;
+}
+
 static void port_assign(struct execlist_port *port, struct i915_request *rq)
 {
        GEM_BUG_ON(rq == port_request(port));
@@ -716,8 +726,6 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
                int i;
 
                priolist_for_each_request_consume(rq, rn, p, i) {
-                       GEM_BUG_ON(!assert_priority_queue(execlists, last, rq));
-
                        /*
                         * Can we combine this request with the current port?
                         * It has to be the same context/ringbuffer and not
@@ -729,8 +737,7 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
                         * second request, and so we never need to tell the
                         * hardware about the first.
                         */
-                       if (last &&
-                           !can_merge_ctx(rq->hw_context, last->hw_context)) {
+                       if (last && !can_merge_rq(last, rq)) {
                                /*
                                 * If we are on the second port and cannot
                                 * combine this request with the last, then we
@@ -739,6 +746,14 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
                                if (port == last_port)
                                        goto done;
 
+                               /*
+                                * We must not populate both ELSP[] with the
+                                * same LRCA, i.e. we must submit 2 different
+                                * contexts if we submit 2 ELSP.
+                                */
+                               if (last->hw_context == rq->hw_context)
+                                       goto done;
+
                                /*
                                 * If GVT overrides us we only ever submit
                                 * port[0], leaving port[1] empty. Note that we
@@ -750,7 +765,6 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
                                    ctx_single_port_submission(rq->hw_context))
                                        goto done;
 
-                               GEM_BUG_ON(last->hw_context == rq->hw_context);
 
                                if (submit)
                                        port_assign(port, last);
@@ -790,8 +804,7 @@ done:
         * request triggering preemption on the next dequeue (or subsequent
         * interrupt for secondary ports).
         */
-       execlists->queue_priority_hint =
-               port != execlists->port ? rq_prio(last) : INT_MIN;
+       execlists->queue_priority_hint = queue_prio(execlists);
 
        if (submit) {
                port_assign(port, last);