On 7/23/23 3:39?AM, Oleksandr Natalenko wrote: > Hello. > > On ned?le 16. ?ervence 2023 21:50:53 CEST Greg Kroah-Hartman wrote: >> From: Andres Freund <andres@xxxxxxxxxxx> >> >> commit 8a796565cec3601071cbbd27d6304e202019d014 upstream. >> >> I observed poor performance of io_uring compared to synchronous IO. That >> turns out to be caused by deeper CPU idle states entered with io_uring, >> due to io_uring using plain schedule(), whereas synchronous IO uses >> io_schedule(). >> >> The losses due to this are substantial. On my cascade lake workstation, >> t/io_uring from the fio repository e.g. yields regressions between 20% >> and 40% with the following command: >> ./t/io_uring -r 5 -X0 -d 1 -s 1 -c 1 -p 0 -S$use_sync -R 0 /mnt/t2/fio/write.0.0 >> >> This is repeatable with different filesystems, using raw block devices >> and using different block devices. >> >> Use io_schedule_prepare() / io_schedule_finish() in >> io_cqring_wait_schedule() to address the difference. >> >> After that using io_uring is on par or surpassing synchronous IO (using >> registered files etc makes it reliably win, but arguably is a less fair >> comparison). >> >> There are other calls to schedule() in io_uring/, but none immediately >> jump out to be similarly situated, so I did not touch them. Similarly, >> it's possible that mutex_lock_io() should be used, but it's not clear if >> there are cases where that matters. >> >> Cc: stable@xxxxxxxxxxxxxxx # 5.10+ >> Cc: Pavel Begunkov <asml.silence@xxxxxxxxx> >> Cc: io-uring@xxxxxxxxxxxxxxx >> Cc: linux-kernel@xxxxxxxxxxxxxxx >> Signed-off-by: Andres Freund <andres@xxxxxxxxxxx> >> Link: https://lore.kernel.org/r/20230707162007.194068-1-andres@xxxxxxxxxxx >> [axboe: minor style fixup] >> Signed-off-by: Jens Axboe <axboe@xxxxxxxxx> >> Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx> >> --- >> io_uring/io_uring.c | 15 +++++++++++++-- >> 1 file changed, 13 insertions(+), 2 deletions(-) >> >> --- a/io_uring/io_uring.c >> +++ b/io_uring/io_uring.c >> @@ -2575,6 +2575,8 @@ int io_run_task_work_sig(struct io_ring_ >> static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx, >> struct io_wait_queue *iowq) >> { >> + int token, ret; >> + >> if (unlikely(READ_ONCE(ctx->check_cq))) >> return 1; >> if (unlikely(!llist_empty(&ctx->work_llist))) >> @@ -2585,11 +2587,20 @@ static inline int io_cqring_wait_schedul >> return -EINTR; >> if (unlikely(io_should_wake(iowq))) >> return 0; >> + >> + /* >> + * Use io_schedule_prepare/finish, so cpufreq can take into account >> + * that the task is waiting for IO - turns out to be important for low >> + * QD IO. >> + */ >> + token = io_schedule_prepare(); >> + ret = 0; >> if (iowq->timeout == KTIME_MAX) >> schedule(); >> else if (!schedule_hrtimeout(&iowq->timeout, HRTIMER_MODE_ABS)) >> - return -ETIME; >> - return 0; >> + ret = -ETIME; >> + io_schedule_finish(token); >> + return ret; >> } >> >> /* > > Reportedly, this caused a regression as reported in [1] [2] [3]. Not only v6.4.4 is affected, v6.1.39 is affected too. > > Reverting this commit fixes the issue. > > Please check. > > Thanks. > > [1] https://bbs.archlinux.org/viewtopic.php?id=287343 > [2] https://bugzilla.kernel.org/show_bug.cgi?id=217700 > [3] https://bugzilla.kernel.org/show_bug.cgi?id=217699 Just read the first one, but this is very much expected. It's now just correctly reflecting that one thread is waiting on IO. IO wait being 100% doesn't mean that one core is running 100% of the time, it just means it's WAITING on IO 100% of the time. -- Jens Axboe
