Revert "drm/i915: Skip execlists_dequeue() early if the list is empty"
authorChris Wilson <chris@chris-wilson.co.uk>
Wed, 29 Mar 2017 09:47:46 +0000 (10:47 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Wed, 29 Mar 2017 12:02:24 +0000 (13:02 +0100)
This reverts commit 6c943de6686f ("drm/i915: Skip execlists_dequeue()
early if the list is empty").

The validity of using READ_ONCE there depends upon having a mb to
coordinate the assignment of engine->execlist_first inside
submit_request() and checking prior to taking the spinlock in
execlists_dequeue(). We wrote "the update to TASKLET_SCHED incurs a
memory barrier making this cross-cpu checking safe", but failed to
notice that this mb was *conditional* on the execlists being ready, i.e.
there wasn't the required mb when it was most necessary!

We could install an unconditional memory barrier to fixup the
READ_ONCE():

diff --git a/drivers/gpu/drm/i915/intel_lrc.c
b/drivers/gpu/drm/i915/intel_lrc.c
index 7dd732cb9f57..1ed164b16d44 100644
--- a/drivers/gpu/drm/i915/intel_lrc.c
+++ b/drivers/gpu/drm/i915/intel_lrc.c
@@ -616,6 +616,7 @@ static void execlists_submit_request(struct
drm_i915_gem_request *request)

        if (insert_request(&request->priotree, &engine->execlist_queue))
{
                engine->execlist_first = &request->priotree.node;
+               smp_wmb();
                if (execlists_elsp_ready(engine))

But we have opted to remove the race as it should be rarely effective,
and saves us having to explain the necessary memory barriers which we
quite clearly failed at.

Reported-and-tested-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Fixes: 6c943de6686f ("drm/i915: Skip execlists_dequeue() early if the list is empty")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: MichaƂ Winiarski <michal.winiarski@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170329100052.29505-1-chris@chris-wilson.co.uk
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
drivers/gpu/drm/i915/i915_guc_submission.c
drivers/gpu/drm/i915/intel_lrc.c

index 6193ad7edcf47d81c0ec17b117654837d82b2db6..1642fff9cf135d5edbe85864d1b327d59002c026 100644 (file)
@@ -662,18 +662,6 @@ static bool i915_guc_dequeue(struct intel_engine_cs *engine)
        struct rb_node *rb;
        bool submit = false;
 
-       /* After execlist_first is updated, the tasklet will be rescheduled.
-        *
-        * If we are currently running (inside the tasklet) and a third
-        * party queues a request and so updates engine->execlist_first under
-        * the spinlock (which we have elided), it will atomically set the
-        * TASKLET_SCHED flag causing the us to be re-executed and pick up
-        * the change in state (the update to TASKLET_SCHED incurs a memory
-        * barrier making this cross-cpu checking safe).
-        */
-       if (!READ_ONCE(engine->execlist_first))
-               return false;
-
        spin_lock_irq(&engine->timeline->lock);
        rb = engine->execlist_first;
        while (rb) {
index 12cf645793ed06d0cbb14120521d923e1942a053..c8f7c631fc1f8e354cac0038c80aa35d0a1dd0d2 100644 (file)
@@ -402,18 +402,6 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
        struct rb_node *rb;
        bool submit = false;
 
-       /* After execlist_first is updated, the tasklet will be rescheduled.
-        *
-        * If we are currently running (inside the tasklet) and a third
-        * party queues a request and so updates engine->execlist_first under
-        * the spinlock (which we have elided), it will atomically set the
-        * TASKLET_SCHED flag causing the us to be re-executed and pick up
-        * the change in state (the update to TASKLET_SCHED incurs a memory
-        * barrier making this cross-cpu checking safe).
-        */
-       if (!READ_ONCE(engine->execlist_first))
-               return;
-
        last = port->request;
        if (last)
                /* WaIdleLiteRestore:bdw,skl