dma-buf: Relax the write-seqlock for reallocating the shared fence list
authorChris Wilson <chris@chris-wilson.co.uk>
Fri, 12 Jul 2019 08:03:14 +0000 (09:03 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Tue, 16 Jul 2019 20:02:39 +0000 (21:02 +0100)
As the set of shared fences is not being changed during reallocation of
the reservation list, we can skip updating the write_seqlock.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: Christian König <christian.koenig@amd.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20190712080314.21018-2-chris@chris-wilson.co.uk
drivers/dma-buf/reservation.c

index 80ecc12..c71b85c 100644 (file)
@@ -157,15 +157,15 @@ int reservation_object_reserve_shared(struct reservation_object *obj,
                (ksize(new) - offsetof(typeof(*new), shared)) /
                sizeof(*new->shared);
 
-       preempt_disable();
-       write_seqcount_begin(&obj->seq);
        /*
-        * RCU_INIT_POINTER can be used here,
-        * seqcount provides the necessary barriers
+        * We are not changing the effective set of fences here so can
+        * merely update the pointer to the new array; both existing
+        * readers and new readers will see exactly the same set of
+        * active (unsignaled) shared fences. Individual fences and the
+        * old array are protected by RCU and so will not vanish under
+        * the gaze of the rcu_read_lock() readers.
         */
-       RCU_INIT_POINTER(obj->fence, new);
-       write_seqcount_end(&obj->seq);
-       preempt_enable();
+       rcu_assign_pointer(obj->fence, new);
 
        if (!old)
                return 0;