> 2020年2月14日 下午6:34，Pavel Begunkov <asml.silence@xxxxxxxxx> 写道： > > On 2/14/2020 11:29 AM, Carter Li 李通洲 wrote: >> To implement io_uring_wait_cqe_timeout, we introduce a magic number >> called `LIBURING_UDATA_TIMEOUT`. The problem is that not only we >> must make sure that users should never set sqe->user_data to >> LIBURING_UDATA_TIMEOUT, but also introduce extra complexity to >> filter out TIMEOUT cqes. >> >> Former discussion: https://github.com/axboe/liburing/issues/53 >> >> I’m suggesting introducing a new SQE flag called IOSQE_IGNORE_CQE >> to solve this problem. >> >> For a sqe tagged with IOSQE_IGNORE_CQE flag, it won’t generate a cqe >> on completion. So that IORING_OP_TIMEOUT can be filtered on kernel >> side. >> >> In addition, `IOSQE_IGNORE_CQE` can be used to save cq size. >> >> For example `POLL_ADD(POLLIN)->READ/RECV` link chain, people usually >> don’t care the result of `POLL_ADD` is ( since it will always be >> POLLIN ), `IOSQE_IGNORE_CQE` can be set on `POLL_ADD` to save lots >> of cq size. >> >> Besides POLL_ADD, people usually don’t care the result of POLL_REMOVE >> /TIMEOUT_REMOVE/ASYNC_CANCEL/CLOSE. These operations can also be tagged >> with IOSQE_IGNORE_CQE. >> >> Thoughts? >> > > I like the idea! And that's one of my TODOs for the eBPF plans. > Let me list my use cases, so we can think how to extend it a bit. > > 1. In case of link fail, we need to reap all -ECANCELLED, analise it and > resubmit the rest. It's quite inconvenient. We may want to have CQE only > for not cancelled requests. > > 2. When chain succeeded, you in the most cases already know the result > of all intermediate CQEs, but you still need to reap and match them. > I'd prefer to have only 1 CQE per link, that is either for the first > failed or for the last request in the chain. > > These 2 may shed much processing overhead from the userspace. I couldn't agree more! Another problem is that io_uring_enter will be awaked for completion of every operation in a link, which results in unnecessary context switch. When awaked, users have nothing to do but issue another io_uring_enter syscall to wait for completion of the entire link chain. > > 3. If we generate requests by eBPF even the notion of per-request event > may broke. > - eBPF creating new requests would also need to specify user-data, and > this may be problematic from the user perspective. > - may want to not generate CQEs automatically, but let eBPF do it. > > -- > Pavel Begunkov
