Hi, Over the last 6 months, several people have asked for if it's possible to send a MSG_RING request to a ring without having a source ring to do it from. The answer is no, as you'd need a source ring to submit such a request in the first place. However, we can easily support this use case of allowing someone to send a message to a ring that their own, without needing to setup a source ring just for that alone. This adds support for "blind" register opcodes for io_uring_register(2), which simply means that there's no io_uring ring fd being passed in. The 'fd' must be set to -1. IORING_REGISTER_SEND_MSG_RING is added, which simply takes a pointer to an io_uring_sqe. That sqe must be setup just like an sqe that would have been otherwise prepared via sending over a normal ring. An sqe pointer is used to keep the app side trivial, as they can just put an sqe on the stack, initialize it to zeroes, and then call io_uring_prep_msg_ring() on it like they would for an async MSG_RING. Once setup, the app can call: io_uring_register(-1, IORING_REGISTER_SEND_MSG_RING, &sqe, 1); which would like like: io_uring_send_msg_ring_sync(&sqe); if using linuring. The return value of this syscall is what would have been in cqe->res using the async approach - 0 on success, or a negative error value in case of failure. Patches can also be found in a kernel branch here: https://git.kernel.dk/cgit/linux/log/?h=io_uring-sync-msg_ring and a liburing branch with support (and test cases) is here: https://git.kernel.dk/cgit/liburing/log/?h=sync-msg Since v1: - Cleanups - Sanity check sqe->flags and the actual opcode - Use fdget/fdput - Add a few comments include/uapi/linux/io_uring.h | 3 ++ io_uring/msg_ring.c | 60 ++++++++++++++++++++++++++++------- io_uring/msg_ring.h | 1 + io_uring/register.c | 30 ++++++++++++++++++ 4 files changed, 83 insertions(+), 11 deletions(-) -- Jens Axboe
