kernel/linux
0
Fork 0
mirror of https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/ synced 2026-11-10 20:48:12 +00:00
Code Issues Releases Packages Activity

Merge branch 'vsock-receive-queue-hint-for-io_uring'

Browse source 
David Carlier says:

====================
vsock: receive queue hint for io_uring

AF_VSOCK stream receives never fill msghdr.msg_inq. io_uring initialises the
hint to -1 and treats that as "unknown", so it cannot set
IORING_CQE_F_SOCK_NONEMPTY, and it speculatively retries a multishot receive
after every completion, including when the queue is already empty.

Patch 1 fills the hint at the common exit of __vsock_connectible_recvmsg(),
reusing vsock_stream_has_data(), the transport callback that SIOCINQ already
uses. Following tcp_inq_hint(), it reports a terminal hint of 1 once the
connection is finished, so the caller still performs the receive that
observes EOF. Patch 2 adds selftest coverage.

Numbers come from a ping-pong over vsock loopback using io_uring multishot
receive, 4 KiB messages, the queue drained between messages, with kernel
receives counted by a kprobe. The baseline is the same tree with patch 1
reverted:

  entries per delivered message   1.97 -> 1.00       (5000 messages)
  receiver CPU time, median       1.640s -> 1.587s   (-3.2%)
  receiver wall time, median      3.260s -> 3.183s   (-2.4%)

Timing is 25 runs of 50000 messages per kernel; CPU gives Mann-Whitney
p=0.006, wall time does not separate from noise (p=0.12), as expected for a
latency-bound workload.

The hint is gated to SOCK_STREAM; seqpacket is untouched. SO_INQ and SCM_INQ
are deliberately left out: they need their own per-socket state and
ancillary-message semantics, and a byte-count contract would have to account
for transports reporting less than the whole queue. That can follow.

Testing: built and run under virtme-ng on loopback, where the new cases pass
and vsock_test still passes 41 of 41 with a clean dmesg. Other transports
were reviewed but not tested. Hyper-V reports the current packet rather than
the whole queue and can return -EIO; an undercount only understates what is
readable, and a negative result reports -1, the same "unknown" value io_uring
presets.
====================

Link: https://patch.msgid.link/20260923184008.153541-1-devnexen@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This commit is contained in:
Jakub Kicinski 2026-09-28 17:27:11 -07:00
commit bd8e2c16f1
4 changed files with 436 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files