[PATCHSET,RFC,0/4] Coalesce provided buffer segments

Message

Jens Axboe Aug. 12, 2024, 4:55 p.m. UTC

Hi,

When selecting provided buffers for a send/recv for bundles, there's
no reason why the number of buffers selected is the same as the mapped
segments that will be passed to send/recv. If some (or all) of these
buffers are virtually contigious, then they can get collapsed into much
fewer segments. Sometimes even just a single segment. This avoids costly
iteration on the send/recv processing side.

The return value is the number of bytes sent/received, and the starting
buffer ID where the operation begun. This is again identical to how
bundles work, from the application point of view this doesn't change
anything in terms of how send/recv bundles are handled, hence this is
a transparent feature.

Patch 1-3 are just basic prep patches, and patch 4 allows for actual
coalescing of segments. This is only enabled for bundles, as those are
the types of requests that process multiple buffers in a single
operation.

Patches are on top of 6.11-rc3 with pending io_uring patches, as well
as the incremental buffer consumption patches [1] posted earlier today.

 io_uring/kbuf.c | 71 ++++++++++++++++++++++++++++++++++++++++++-------
 io_uring/kbuf.h |  7 +++--
 io_uring/net.c  | 55 +++++++++++++++++---------------------
 io_uring/net.h  |  1 +
 4 files changed, 91 insertions(+), 43 deletions(-)