Hi Fritz, On Fri, Sep 15, 2023 at 03:07:38PM -0700, Fritz Koenig wrote: > Hi, > > Looking at the documentation[1] it states that the fd returned from > MEDIA_IOC_REQUEST_ALLOC can be close()ed and the kernel will take care > of making sure it is ref counted correctly. > > Because MEDIA_REQUEST_IOC_REINIT needs to be called on an fd before it > can be reused, is there a performance penalty to allocating a new fd > (with MEDIA_IOC_REQUEST_ALLOC) for every frame instead of having a > pool of fd's that are reused, and then immediately after queuing the > buffer, close the file descriptor? > > I have a proof of concept that works, but I haven't tested it for > performance yet. > > It would seem that maybe there is a latency penalty because the > MEDIA_REQUEST_IOC_REINIT is called at the end of a frame while > MEDIA_IOC_REQUEST_ALLOC would be called at the start of a frame. That's up to userspace to decide, you could also call MEDIA_IOC_REQUEST_ALLOC at the end of a frame. That would likely defeit the point though, as you would need to put that fd in a pool to be used later. Allocating a new fd isn't very expensive. The reason why the reinit mechanism was implemented is to avoid fd allocation errors at runtime. By pre-allocating a pool of requests and reusing them, you are guaranteed that you will always have a request available, even if the system runs out of fds. > [1]: https://docs.kernel.org/userspace-api/media/mediactl/request-api.html#recycling-and-destruction -- Regards, Laurent Pinchart
