drm/ttm: Fix up a theoretical deadlock
Thomas Hellstrom [Thu, 11 Nov 2010 09:04:53 +0000 (10:04 +0100)]
A process suspended waiting for a higher sequence or no sequence to unreserve,
a bo may be beaten to the reservation by a process with a lower sequence.
In that case the first process should give up trying to reserve and
return -EAGAIN. In order for that to happen, we must wake waiting processes
when we change sequence, so that they have a chance to detect the new
sequence.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drivers/gpu/drm/ttm/ttm_bo.c

index 3ca77dc..148a322 100644 (file)
@@ -224,6 +224,9 @@ int ttm_bo_reserve_locked(struct ttm_buffer_object *bo,
        int ret;
 
        while (unlikely(atomic_cmpxchg(&bo->reserved, 0, 1) != 0)) {
+               /**
+                * Deadlock avoidance for multi-bo reserving.
+                */
                if (use_sequence && bo->seq_valid &&
                        (sequence - bo->val_seq < (1 << 31))) {
                        return -EAGAIN;
@@ -241,6 +244,14 @@ int ttm_bo_reserve_locked(struct ttm_buffer_object *bo,
        }
 
        if (use_sequence) {
+               /**
+                * Wake up waiters that may need to recheck for deadlock,
+                * if we decreased the sequence number.
+                */
+               if (unlikely((bo->val_seq - sequence < (1 << 31))
+                            || !bo->seq_valid))
+                       wake_up_all(&bo->event_queue);
+
                bo->val_seq = sequence;
                bo->seq_valid = true;
        } else {