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
Cc: <[email protected]> # 3.9+
Cc: Ben Skeggs <[email protected]>
Signed-off-by: Jeff Mahoney <[email protected]>
---
drivers/gpu/drm/nouveau/nouveau_fence.c | 11 +++++------
1 file changed, 5 insertions(+), 6 deletions(-)
--- a/drivers/gpu/drm/nouveau/nouveau_fence.c
+++ b/drivers/gpu/drm/nouveau/nouveau_fence.c
@@ -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_fence
int 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;
}
--
Jeff Mahoney
SUSE Labs