nouveau: fix fence hang

If there is not enough space in pushbuffer for fence emission
(nouveau_fence_emit -> nv50_screen_fence_emit -> MARK_RING),
the pushbuffer is flushed, which through flush_notify ->
nv50_default_flush_notify -> nouveau_fence_update marks currently
emitting fence as flushed. But actual emission is done after this mark.
So later when there is a need to wait on this fence and pushbuffer
was not flushed in between, fence wait will never finish causing
application to hang.

To fix this, introduce new fence state between AVAILABLE and EMITTED,
set it before emission and handle it everywhere.

Additionally obtain fence sequence numbers after possible flush in
MARK_RING, because we want to emit fences in correct order.

Reviewed-by: Christoph Bumiller <[email protected]>

Note: This is a candidate for the 7.11 branch.

1 files changed, 6 insertions, 2 deletions

diff --git a/src/gallium/drivers/nvc0/nvc0_screen.c b/src/gallium/drivers/nvc0/nvc0_screen.c
index cc6e6830af2..292cbddcc0c 100644
--- a/src/gallium/drivers/nvc0/nvc0_screen.c
+++ b/src/gallium/drivers/nvc0/nvc0_screen.c
@@ -334,16 +334,20 @@ nvc0_magic_3d_init(struct nouveau_channel *chan)
 }
 
 static void
-nvc0_screen_fence_emit(struct pipe_screen *pscreen, u32 sequence)
+nvc0_screen_fence_emit(struct pipe_screen *pscreen, u32 *sequence)
 {
    struct nvc0_screen *screen = nvc0_screen(pscreen);
    struct nouveau_channel *chan = screen->base.channel;
 
    MARK_RING (chan, 5, 2);
+
+   /* we need to do it after possible flush in MARK_RING */
+   *sequence = ++screen->base.fence.sequence;
+
    BEGIN_RING(chan, RING_3D(QUERY_ADDRESS_HIGH), 4);
    OUT_RELOCh(chan, screen->fence.bo, 0, NOUVEAU_BO_WR);
    OUT_RELOCl(chan, screen->fence.bo, 0, NOUVEAU_BO_WR);
-   OUT_RING  (chan, sequence);
+   OUT_RING  (chan, *sequence);
    OUT_RING  (chan, NVC0_3D_QUERY_GET_FENCE | NVC0_3D_QUERY_GET_SHORT |
               (0xf << NVC0_3D_QUERY_GET_UNIT__SHIFT));
 }