24 MIOPS vs 31.5, or ~30% win for fio/t/io_uring nops batching=32
Jens mentioned that with his standard test against Optane it gave
yet another +3% to throughput.
1-14 are about optimising the completion path:
- replaces lists with single linked lists
- kills 64 * 8B of caches in ctx
- adds some shuffling of iopoll bits
- list splice instead of per-req list_add in one place
- inlines io_req_free_batch() and other helpers
15-22: inlines __io_queue_sqe() so all the submission path
up to io_issue_sqe() is inlined + little tweaks
v2: rebase for-5.16/io_uring
multicqe_drain was hanging because it's a bit buggy, i.e. doesn't
consider that requests may get punted, but still add 24th patch to
avoid it.
Pavel Begunkov (24):
io_uring: mark having different creds unlikely
io_uring: force_nonspin
io_uring: make io_do_iopoll return number of reqs
io_uring: use slist for completion batching
io_uring: remove allocation cache array
io-wq: add io_wq_work_node based stack
io_uring: replace list with stack for req caches
io_uring: split iopoll loop
io_uring: use single linked list for iopoll
io_uring: add a helper for batch free
io_uring: convert iopoll_completed to store_release
io_uring: optimise batch completion
io_uring: inline completion batching helpers
io_uring: don't pass tail into io_free_batch_list
io_uring: don't pass state to io_submit_state_end
io_uring: deduplicate io_queue_sqe() call sites
io_uring: remove drain_active check from hot path
io_uring: split slow path from io_queue_sqe
io_uring: inline hot path of __io_queue_sqe()
io_uring: reshuffle queue_sqe completion handling
io_uring: restructure submit sqes to_submit checks
io_uring: kill off ->inflight_entry field
io_uring: comment why inline complete calls io_clean_op()
io_uring: disable draining earlier
fs/io-wq.h | 60 +++++-
fs/io_uring.c | 508 +++++++++++++++++++++++---------------------------
2 files changed, 287 insertions(+), 281 deletions(-)
--
2.33.0
