Protocol: wl_buffer.release is racy
Submitted by Pekka Paalanen
Assigned to Wayland bug list
Description
Consider a client doing this:
- surface1.attach(buffer1)
- surface1.commit
- surface2.attach(buffer1)
- surface2.commit
Then the client receives buffer1.release.
It is ambiguous, whether the release corresponds to step 2 or step 4. That is, the client cannot know if the latter commit still holds the buffer reserved in the server. The server may have had time to process and release the buffer before step 4, in which case the buffer would be reserved again after step 4.
The problem is the same also, if surface1 and surface2 were both just surface1. If the compositor has time to repaint in between the commits, the client may get a release it most likely interprets as the release corresponding to step 4, which would be wrong.
How should this be resolved?
A suggestion from Jason Ekstrand was to modify the protocol to guarantee one release event for each commit that makes the buffer reserved. This would allow clients to use simple reference counting.
Another possibility would be to require, that clients must not commit a wl_buffer that is already reserved by the server. However, this seems quite limiting, especially with the coming presentation extension that allows queueing an arbitrary number of updates which could reference the same buffer more than once (for e.g. simple cyclic animations like spinners or cursors).