This is an RFC because I'm still trying to grok the correct behavior.
Consider a dma_fence_array created two two fence and signal_on_any is true. A reference to dma_fence_array is taken for each waiting fence.
When the client calls dma_fence_wait() only one of the fences is signaled. The client returns successfully from the wait and puts it's reference to the array fence but the array fence still remains because of the remaining un-signaled fence.
Now consider that the unsignaled fence is signaled while the timeline is being destroyed much later. The timeline destroy calls dma_fence_signal_locked(). The following sequence occurs:
1) dma_fence_array_cb_func is called
2) array->num_pending is 0 (because it was set to 1 due to signal_on_any) so the callback function calls dma_fence_put() instead of triggering the irq work
3) The array fence is released which in turn puts the lingering fence which is then released
4) deadlock with the timeline
I think that we can fix this with the attached patch. Once the fence is signaled signaling it again in the irq worker shouldn't hurt anything. The only gotcha might be how the error is propagated - I wasn't quite sure the intent of clearing it only after getting to the irq worker.
Signed-off-by: Jordan Crouse jcrouse@codeaurora.org ---
drivers/dma-buf/dma-fence-array.c | 10 ++++------ 1 file changed, 4 insertions(+), 6 deletions(-)
diff --git a/drivers/dma-buf/dma-fence-array.c b/drivers/dma-buf/dma-fence-array.c index d3fbd950be94..b8829b024255 100644 --- a/drivers/dma-buf/dma-fence-array.c +++ b/drivers/dma-buf/dma-fence-array.c @@ -46,8 +46,6 @@ static void irq_dma_fence_array_work(struct irq_work *wrk) { struct dma_fence_array *array = container_of(wrk, typeof(*array), work);
- dma_fence_array_clear_pending_error(array); - dma_fence_signal(&array->base); dma_fence_put(&array->base); } @@ -61,10 +59,10 @@ static void dma_fence_array_cb_func(struct dma_fence *f,
dma_fence_array_set_pending_error(array, f->error);
- if (atomic_dec_and_test(&array->num_pending)) - irq_work_queue(&array->work); - else - dma_fence_put(&array->base); + if (!atomic_dec_and_test(&array->num_pending)) + dma_fence_array_set_pending_error(array, f->error); + + irq_work_queue(&array->work); }
static bool dma_fence_array_enable_signaling(struct dma_fence *fence)
Quoting Jordan Crouse (2020-08-13 00:55:44)
This is an RFC because I'm still trying to grok the correct behavior.
Consider a dma_fence_array created two two fence and signal_on_any is true. A reference to dma_fence_array is taken for each waiting fence.
When the client calls dma_fence_wait() only one of the fences is signaled. The client returns successfully from the wait and puts it's reference to the array fence but the array fence still remains because of the remaining un-signaled fence.
Now consider that the unsignaled fence is signaled while the timeline is being destroyed much later. The timeline destroy calls dma_fence_signal_locked(). The following sequence occurs:
dma_fence_array_cb_func is called
array->num_pending is 0 (because it was set to 1 due to signal_on_any) so the
callback function calls dma_fence_put() instead of triggering the irq work
- The array fence is released which in turn puts the lingering fence which is
then released
- deadlock with the timeline
It's the same recursive lock as we previously resolved in sw_sync.c by removing the locking from timeline_fence_release(). -Chris
On Thu, Aug 13, 2020 at 07:49:24AM +0100, Chris Wilson wrote:
Quoting Jordan Crouse (2020-08-13 00:55:44)
This is an RFC because I'm still trying to grok the correct behavior.
Consider a dma_fence_array created two two fence and signal_on_any is true. A reference to dma_fence_array is taken for each waiting fence.
When the client calls dma_fence_wait() only one of the fences is signaled. The client returns successfully from the wait and puts it's reference to the array fence but the array fence still remains because of the remaining un-signaled fence.
Now consider that the unsignaled fence is signaled while the timeline is being destroyed much later. The timeline destroy calls dma_fence_signal_locked(). The following sequence occurs:
dma_fence_array_cb_func is called
array->num_pending is 0 (because it was set to 1 due to signal_on_any) so the
callback function calls dma_fence_put() instead of triggering the irq work
- The array fence is released which in turn puts the lingering fence which is
then released
- deadlock with the timeline
It's the same recursive lock as we previously resolved in sw_sync.c by removing the locking from timeline_fence_release().
Ah, yep. I'm working on a not-quite-ready-for-primetime version of a vulkan timeline implementation for drm/msm and I was doing something similar to how sw_sync used to work in the release function. Getting rid of the recursive lock in the timeline seems a better solution than this. Thanks for taking the time to respond.
Jordan
-Chris
Am 13.08.20 um 01:55 schrieb Jordan Crouse:
This is an RFC because I'm still trying to grok the correct behavior.
Consider a dma_fence_array created two two fence and signal_on_any is true. A reference to dma_fence_array is taken for each waiting fence.
Ok, that sounds like you seem to mix a couple of things up here.
A dma_fence_array takes the reference to the fences it contains on creation. There is only one reference to the dma_fence_array even if it contains N unsignaled fences.
What we do is to grab a reference to the array in dma_fence_array_enable_signaling(), but this is because we are registering the callback here.
When the client calls dma_fence_wait() only one of the fences is signaled. The client returns successfully from the wait and puts it's reference to the array fence but the array fence still remains because of the remaining un-signaled fence.
If signaling was enabled then this is correct, because otherwise we would crash when the other callbacks are called.
Now consider that the unsignaled fence is signaled while the timeline is being destroyed much later. The timeline destroy calls dma_fence_signal_locked(). The following sequence occurs:
dma_fence_array_cb_func is called
array->num_pending is 0 (because it was set to 1 due to signal_on_any) so the
callback function calls dma_fence_put() instead of triggering the irq work
- The array fence is released which in turn puts the lingering fence which is
then released
- deadlock with the timeline
Why do we have a deadlock here? That doesn't seems to add up.
Christian.
I think that we can fix this with the attached patch. Once the fence is signaled signaling it again in the irq worker shouldn't hurt anything. The only gotcha might be how the error is propagated - I wasn't quite sure the intent of clearing it only after getting to the irq worker.
Signed-off-by: Jordan Crouse jcrouse@codeaurora.org
drivers/dma-buf/dma-fence-array.c | 10 ++++------ 1 file changed, 4 insertions(+), 6 deletions(-)
diff --git a/drivers/dma-buf/dma-fence-array.c b/drivers/dma-buf/dma-fence-array.c index d3fbd950be94..b8829b024255 100644 --- a/drivers/dma-buf/dma-fence-array.c +++ b/drivers/dma-buf/dma-fence-array.c @@ -46,8 +46,6 @@ static void irq_dma_fence_array_work(struct irq_work *wrk) { struct dma_fence_array *array = container_of(wrk, typeof(*array), work);
- dma_fence_array_clear_pending_error(array);
- dma_fence_signal(&array->base); dma_fence_put(&array->base); }
@@ -61,10 +59,10 @@ static void dma_fence_array_cb_func(struct dma_fence *f, dma_fence_array_set_pending_error(array, f->error);
- if (atomic_dec_and_test(&array->num_pending))
irq_work_queue(&array->work);
- else
dma_fence_put(&array->base);
- if (!atomic_dec_and_test(&array->num_pending))
dma_fence_array_set_pending_error(array, f->error);
- irq_work_queue(&array->work); }
static bool dma_fence_array_enable_signaling(struct dma_fence *fence)
linaro-mm-sig@lists.linaro.org