drm/i915: Drop i915_request.lock serialisation around await_start
authorChris Wilson <chris@chris-wilson.co.uk>
Thu, 14 Jan 2021 13:56:09 +0000 (13:56 +0000)
committerChris Wilson <chris@chris-wilson.co.uk>
Fri, 15 Jan 2021 08:00:03 +0000 (08:00 +0000)
Originally, we used the signal->lock as a means of following the
previous link in its timeline and peeking at the previous fence.
However, we have replaced the explicit serialisation with a series of
very careful probes that anticipate the links being deleted and the
fences recycled before we are able to acquire a strong reference to it.
We do not need the signal->lock crutch anymore, nor want the contention.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Andi Shyti <andi.shyti@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210114135612.13210-2-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_request.c

index 784c05a..973ecea 100644 (file)
@@ -969,9 +969,16 @@ i915_request_await_start(struct i915_request *rq, struct i915_request *signal)
        if (i915_request_started(signal))
                return 0;
 
+       /*
+        * The caller holds a reference on @signal, but we do not serialise
+        * against it being retired and removed from the lists.
+        *
+        * We do not hold a reference to the request before @signal, and
+        * so must be very careful to ensure that it is not _recycled_ as
+        * we follow the link backwards.
+        */
        fence = NULL;
        rcu_read_lock();
-       spin_lock_irq(&signal->lock);
        do {
                struct list_head *pos = READ_ONCE(signal->link.prev);
                struct i915_request *prev;
@@ -1002,7 +1009,6 @@ i915_request_await_start(struct i915_request *rq, struct i915_request *signal)
 
                fence = &prev->fence;
        } while (0);
-       spin_unlock_irq(&signal->lock);
        rcu_read_unlock();
        if (!fence)
                return 0;