On 4/21/22 14:44, Pavel Begunkov wrote:
For experiments only. If proves to be useful would need to make it
nicer on the non-io_uring side.
0-10 save 1 spinlock/unlock_irq pair and 2 cmpxchg per batch. 11/11 in
general trades 1 per tw add spin_lock/unlock_irq and 2 per batch spinlocking
with 2 cmpxchg to 1 per tw add cmpxchg and 1 per batch cmpxchg.
null_blk irqmode=1 completion_nsec=0 submit_queues=32 poll_queues=32
echo -n 0 > /sys/block/nullb0/queue/iostats
echo -n 2 > /sys/block/nullb0/queue/nomerges
io_uring -d<QD> -s<QD> -c<QD> -p0 -B1 -F1 -b512 /dev/nullb0
| base | 1-10 | 1-11
___________________________________________
QD1 | 1.88 | 2.15 (+14%) | 2.19 (+16.4%)
QD4 | 2.8 | 3.06 (+9.2%) | 3.11 (+11%)
QD32 | 3.61 | 3.81 (+5.5%) | 3.96 (+9.6%)
The numbers are in MIOPS, (%) is relative diff with the baseline.
It gives more than I expected, but the testing is not super
consistent, so a part of it might be due to variance.
Pavel Begunkov (11):
io_uring: optimise io_req_task_work_add
io_uringg: add io_should_fail_tw() helper
io_uring: ban tw queue for exiting processes
io_uring: don't take ctx refs in tctx_task_work()
io_uring: add dummy io_uring_task_work_run()
task_work: add helper for signalling a task
io_uring: run io_uring task_works on TIF_NOTIFY_SIGNAL
io_uring: wire io_uring specific task work
io_uring: refactor io_run_task_work()
io_uring: remove priority tw list
io_uring: lock-free task_work stack
fs/io-wq.c | 1 +
fs/io_uring.c | 213 +++++++++++++++-----------------------
include/linux/io_uring.h | 4 +
include/linux/task_work.h | 4 +
kernel/entry/kvm.c | 1 +
kernel/signal.c | 2 +
kernel/task_work.c | 33 +++---
7 files changed, 115 insertions(+), 143 deletions(-)
--
Pavel Begunkov
