The original implementation of sw_sync in 3.14 kernel signaled all
active fences with error status on timeline release. However, after the
conversion to use DMA-buf fence code under the hood, the behavior was
(most likely by mistake) changed and the fences are being left active,
with their users possibly getting stuck in sync_wait().
There is indeed a sign of this not being an intentional behavior change,
as timeline retained its ->destroyed field, which is being set in
sync_timeline_destroy(). However there is no code checking it anymore.
Let's fix this by adding a check for timeline->destroyed in
android_fence_signaled(), so if the fence is neither signaled nor in
error state and timeline was destroyed, it will end up in -ENOENT error
state.
Change-Id: I0313b12b37e9d391d5caf218f381fa4b07a2a5e5
Signed-off-by: Tomasz Figa <tfiga@chromium.org>
Git-commit:
85839014db07d3c675d5a8f2f9f94f2aeb33fc5e
Git-repo: https://chromium.googlesource.com/chromiumos/third_party/kernel
Signed-off-by: Firoz Khan <firozk@codeaurora.org>
int ret;
ret = parent->ops->has_signaled(pt);
+ if (!ret && parent->destroyed)
+ ret = -ENOENT;
if (ret < 0)
fence->status = ret;
return ret;