On 3/18/22 7:52 AM, Pavel Begunkov wrote: > When only one task submits requests, most of CQEs are expected to be > filled from that task context so we have natural serialisation. That > would mean that in those cases we don't need spinlocking around CQE > posting. One downside is that it also mean that io-wq workers can't emit > CQEs directly but should do it through the original task context using > task_works. That may hurt latency and performance and might matter much > to some workloads, but it's not a huge deal in general as io-wq is a > slow path and there is some additional merit from tw completion > batching. Not too worried about io-wq task_work for cq filling, it is the slower path after all. And I think we can get away with doing notifications as it's just for CQ filling. If the task is currently waiting in cqring_wait, then it'll get woken anyway and it will process task work. If it's in userspace, it doesn't need a notification. That should make it somewhat lighter than requiring using TIF_NOTIFY_SIGNAL for that. > The feature should be opted-in by the userspace by setting a new > IORING_SETUP_PRIVATE_CQ flag. It doesn't work with IOPOLL, and also for > now only the task that created a ring can submit requests to it. I know this is a WIP, but why do we need CQ_PRIVATE? And this needs to work with registered files (and ring fd) as that is probably a bigger win than skipping the completion_lock if you're not shared anyway. -- Jens Axboe
