Hi, I went back to the drawing board a bit on the send multishot, and this is what came out. First support was added for provided buffers for send. This works like provided buffers for recv/recvmsg, and the intent here to use the buffer ring queue as an outgoing sequence for sending. But the real meat is adding support for picking multiple buffers at the time, what I dubbed "bundles" here. Rather than just pick a single buffer for send, it can pick a bunch of them and send them in one go. The idea here is that the expensive part of a request is not the sqe issue, it's the fact that we have to do each buffer separately. That entails calling all the way down into the networking stack, locking the socket, checking what needs doing afterwards (like flushing the backlog), unlocking the socket, etc. If we have an outgoing send queue, then pick what buffers we have (up to a certain cap), and pass them to the networking stack in one go. Bundles must be used with provided buffers, obviously. At completion time, they pass the starting buffer ID in cqe->flags, like any other provided buffer completion. cqe->res is the TOTAL number of bytes sent, so it's up to the application to iterate buffers to figure out how many completed. This part is trivial. I'll push the proxy changes out soon, just need to cleanup them up as I did the sendmsg bundling too and would love to compare. With that in place, I added support for recv for bundles as well. Exactly the same as the send side - if we have a known amount of data pending, pick enough buffers to satisfy the receive and post a single completion for that round. Buffer ID in cqe->flags, cqe->res is the total number of buffers sent. Receive can be used with multishot as well - fire off one multishot recv, and keep getting big completions. Unfortunately, recvmsg multishot is just not as efficient as recv, as it carries additional data that needs copying. recv multishot with bundles provide a good alternative to recvmsg, if all you need is more than one range of data. I'll compare these too soon as well. This is obviously a bigger win for smaller packets than for large ones, as the overall cost of entering sys_sendmsg/sys_recvmsg() in terms of throughput decreases as the packet size increases. For the extreme end, using 32b packets, performance increases substantially. Runtime for proxying 32b packets between three machines on a 10G link for the test: Send ring: 3462 msec 1183Mbit Send ring + bundles 844 msec 4853Mbit and bundles reach 100% bandwidth at 80b of packet size, compared to send ring alone needing 320b to reach 95% of bandwidth (I didn't redo that test so don't have the 100% number). Patches are on top of my for-6.9/io_uring branch and can also be found here: https://git.kernel.dk/cgit/linux/log/?h=io_uring-recvsend-bundle include/linux/io_uring_types.h | 3 + include/uapi/linux/io_uring.h | 10 + io_uring/io_uring.c | 3 +- io_uring/kbuf.c | 203 ++++++++++++- io_uring/kbuf.h | 39 ++- io_uring/net.c | 528 +++++++++++++++++++++++---------- io_uring/net.h | 2 +- io_uring/opdef.c | 9 +- 8 files changed, 609 insertions(+), 188 deletions(-) -- Jens Axboe
