nouveau: Fix race with fence signaling
Subsystems (3)
PATCH 1/1 nouveau: Fix race with fence signaling
There exists a tight race between the call to nouveau_fence_done from nouveau_fence_wait and the call to nouveau_fence_wait_uevent.
nouveau_fence_done checks to see if fence->channel is NULL before calling nouveau_fence_wait_uevent, but it's not good enough since the dereference in nouveau_fence_wait_uevent is done outside the lock. Another thread may have signaled the fence in that tight window and then we Oops while dereferencing fence->channel->drm at the beginning of nouveau_fence_wait_uevent.
The good news is that nouveau_fence_wait_uevent only uses fence->channel directly to grab the chan->drm pointer. If we pass that in directly as a known good pointer, we can avoid the race. Passing the nouveau_fence_done check in the caller ensures that the chan pointer is valid.
Original bug report at: https://bugzilla.novell.com/show_bug.cgi?id=844177
--- drivers/gpu/drm/nouveau/nouveau_fence.c | 11 +++++------ 1 file changed, 5 insertions(+), 6 deletions(-)
drivers/gpu/drm/nouveau/nouveau_fence.c +5 -6
@@ -180,12 +180,11 @@ nouveau_fence_wait_uevent_handler(struct }static int -nouveau_fence_wait_uevent (struct nouveau_fence *fence,bool intr) - +nouveau_fence_wait_uevent (struct nouveau_fence *fence, +struct nouveau_drm *drm,bool intr) { -struct nouveau_channel *chan = fence->channel; -struct nouveau_fifo *pfifo =nouveau_fifo (chan->drm->device); -struct nouveau_fence_priv *priv = chan->drm->fence; +struct nouveau_fifo *pfifo =nouveau_fifo (drm->device); +struct nouveau_fence_priv *priv = drm->fence;struct nouveau_fence_uevent uevent = { .handler.func = nouveau_fence_wait_uevent_handler, .priv = priv,
@@ -241,7 +240,7 @@ nouveau_fence_wait(struct nouveau_fenceint ret =0 ;while (priv && priv->uevent && lazy && !nouveau_fence_done (fence)) { - ret =nouveau_fence_wait_uevent (fence, intr); + ret =nouveau_fence_wait_uevent (fence, chan->drm, intr);if (ret <0 )return ret; }
Signature · 2 lines
Jeff Mahoney SUSE Labs