There is a couple of tricks we can do with io_uring to improve ubuf_info refcounting. First, we ammortise reference grabbing and then give them away to the network layer, which is implemented in 8 and 11. Also, we don't need need additional pinning for TCP, which is removed by 7. 1-4 are needed because otherwise we're out of space in io_notif_data and using ->desc or some other field of ubuf_info would be ugly. It'll also facilitate further ideas like adding a simpler notification model for UDP. liburing/examples/io_uring-sendzc benchmark using a branch containing the patchset and some more [1] showed ~1.6% qps improvement for UDP (dummy dev), and ~1% for TCP (localhost + hacks enabling zc). I didn't specifically test xen and vhost and not sure how, would love some help with that. [1] https://github.com/isilence/linux/tree/net/zc-ref-optimisation Pavel Begunkov (11): net: introduce struct ubuf_info_msgzc xen/netback: use struct ubuf_info_msgzc vhost/net: use struct ubuf_info_msgzc net: shrink struct ubuf_info net: rename ubuf_info's flags net: add flags for controlling ubuf_info net/tcp: optimise tcp ubuf refcounting net: let callers provide ->msg_ubuf refs io_uring/notif: add helper for flushing refs io_uring/notif: mark notifs with UARGFL_CALLER_PINNED io_uring/notif: add ubuf_info ref caching drivers/net/xen-netback/common.h | 2 +- drivers/net/xen-netback/interface.c | 4 +-- drivers/net/xen-netback/netback.c | 7 +++--- drivers/vhost/net.c | 17 +++++++------ include/linux/skbuff.h | 35 +++++++++++++++++++++++--- io_uring/net.c | 8 +++++- io_uring/notif.c | 21 ++++++++++------ io_uring/notif.h | 22 +++++++++++++++- net/core/skbuff.c | 39 ++++++++++++++++------------- net/ipv4/ip_output.c | 3 ++- net/ipv4/tcp.c | 11 +++++--- net/ipv6/ip6_output.c | 3 ++- 12 files changed, 123 insertions(+), 49 deletions(-) -- 2.37.0
