On 11/4/19 4:46 AM, yangerkun wrote: > > > On 2019/11/4 18:09, Bob Liu wrote: >> On 11/4/19 4:56 PM, yangerkun wrote: >>> There is no problem to support other type request for the ctx with >>> IORING_SETUP_IOPOLL. >> >> Could you describe the benefit of doing this? > > Hi, > > I am trying to replace libaio with io_uring in InnoDB/MariaDB(which > build on xfs/nvme). And in order to simulate the timeout mechanism > like io_getevents, firstly, to use the poll function of io_uring's fd > has been selected and it really did work. But while trying to enable > IORING_SETUP_IOPOLL since xfs has iopoll function interface, the > mechanism will fail since io_uring_poll does check the cq.head between > cached_cq_tail, which will not update until we call io_uring_enter and > do the poll. So, instead, I decide to use timeout requests in > io_uring but will return -EINVAL since we enable IORING_SETUP_IOPOLL > at the same time. I think this combination is a normal scene so as > the other combination descibed in this patch. I am not sure does it a > good solution for this problem, and maybe there exists some better way. I think we can support timeouts pretty easily with SETUP_IOPOLL, but we can't mix and match everything. Pretty sure I've written at length about that before, but the tldr is that for purely polled commands, we won't have an IRQ event. That's the case for nvme if it's misconfigured, but for an optimal setup where nvme is loaded with poll queues, there will never be an interrupt for the command. This means that we can never wait in io_cqring_wait(), we must always call the iopoll poller, because if we wait we might very well be waiting for events that will never happen unless we actively poll for them. This could be supported if we accounted requests, but I don't want to add that kind of overhead. Same with the lock+irqdisable you had to add for this, it's not acceptable overhead. Sounds like you just need timeout support for polling? If so, then that is supportable as we know that these events will trigger an async event when they happen. Either that, or it triggers when we poll for completions. So it's safe to support, and we can definitely do that. But don't mix polled IO with "normal" IO, it just won't work without extra additions that I'm not willing to pay the cost of. -- Jens Axboe
