drm/i915: Advance ring->head fully when idle
authorChris Wilson <chris@chris-wilson.co.uk>
Thu, 6 Apr 2017 17:00:28 +0000 (18:00 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Fri, 7 Apr 2017 09:10:58 +0000 (10:10 +0100)
When we retire the last request on the ring, before we ever access that
ring again we know it will be completely idle and so we can advance the
ring->head fully to the end (i.e. ring->tail) and not just to the start
of the breadcrumb. This allows us to skip re-emitting the breadcrumb
after resetting the GPU if the ring was entirely idle. This prevents us
from overwriting a seqno wraparound by re-executing a stale breadcrumb,
i.e.
submit_request(1)
intel_engine_init_global_seqno(0)
i915_reset()
would then leave 1 in the HWS, but the next request to execute would
also be with seqno 1. The sanity checks upon submission detect this as a
timewarp and explode. By setting the ring as empty, upon reset the HWS
is left as 0, leaving it consistent with the timeline.

v2: Fix check for deleting last element of list. We know that this
request is always the first element of the ring, so only if next
points back to the start will this be the only request in flight.
v3: Remove opencoding of list_is_last()
v4: Move the block to its own function for some clarity.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=100144
Testcase: igt/gem_exec_whisper/hang-*
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170406170028.26871-1-chris@chris-wilson.co.uk
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
drivers/gpu/drm/i915/i915_gem_request.c

index 2f8c5132b54e5adad8fe1977d15814c0d7dbb5db..313cdff7c6dd6e92d62179f6c751378afe8ca7d0 100644 (file)
@@ -271,6 +271,27 @@ void i915_gem_retire_noop(struct i915_gem_active *active,
        /* Space left intentionally blank */
 }
 
+static void advance_ring(struct drm_i915_gem_request *request)
+{
+       unsigned int tail;
+
+       /* We know the GPU must have read the request to have
+        * sent us the seqno + interrupt, so use the position
+        * of tail of the request to update the last known position
+        * of the GPU head.
+        *
+        * Note this requires that we are always called in request
+        * completion order.
+        */
+       if (list_is_last(&request->ring_link, &request->ring->request_list))
+               tail = request->ring->tail;
+       else
+               tail = request->postfix;
+       list_del(&request->ring_link);
+
+       request->ring->head = tail;
+}
+
 static void i915_gem_request_retire(struct drm_i915_gem_request *request)
 {
        struct intel_engine_cs *engine = request->engine;
@@ -287,16 +308,6 @@ static void i915_gem_request_retire(struct drm_i915_gem_request *request)
        list_del_init(&request->link);
        spin_unlock_irq(&engine->timeline->lock);
 
-       /* We know the GPU must have read the request to have
-        * sent us the seqno + interrupt, so use the position
-        * of tail of the request to update the last known position
-        * of the GPU head.
-        *
-        * Note this requires that we are always called in request
-        * completion order.
-        */
-       list_del(&request->ring_link);
-       request->ring->head = request->postfix;
        if (!--request->i915->gt.active_requests) {
                GEM_BUG_ON(!request->i915->gt.awake);
                mod_delayed_work(request->i915->wq,
@@ -304,6 +315,7 @@ static void i915_gem_request_retire(struct drm_i915_gem_request *request)
                                 msecs_to_jiffies(100));
        }
        unreserve_seqno(request->engine);
+       advance_ring(request);
 
        /* Walk through the active list, calling retire on each. This allows
         * objects to track their GPU activity and mark themselves as idle