lkml.org 
[lkml]   [2019]   [Oct]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    SubjectRe: [PATCH] io_uring: remove wait loop spurious wakeups
    From
    Date
    On 10/8/19 4:05 PM, Pavel Begunkov wrote:
    > On 09/10/2019 00:22, Jens Axboe wrote:
    >> On 10/8/19 2:58 PM, Pavel Begunkov wrote:
    >>> On 08/10/2019 20:00, Jens Axboe wrote:
    >>>> On 10/8/19 10:43 AM, Pavel Begunkov wrote:
    >>>>> On 08/10/2019 06:16, Jens Axboe wrote:
    >>>>>> On 10/7/19 5:18 PM, Pavel Begunkov (Silence) wrote:
    >>>>>>> From: Pavel Begunkov <asml.silence@gmail.com>
    >>>>>>>
    >>>>>>> Any changes interesting to tasks waiting in io_cqring_wait() are
    >>>>>>> commited with io_cqring_ev_posted(). However, io_ring_drop_ctx_refs()
    >>>>>>> also tries to do that but with no reason, that means spurious wakeups
    >>>>>>> every io_free_req() and io_uring_enter().
    >>>>>>>
    >>>>>>> Just use percpu_ref_put() instead.
    >>>>>>
    >>>>>> Looks good, this is a leftover from when the ctx teardown used
    >>>>>> the waitqueue as well.
    >>>>>>
    >>>>> BTW, is there a reason for ref-counting in struct io_kiocb? I understand
    >>>>> the idea behind submission reference, but don't see any actual part
    >>>>> needing it.
    >>>>
    >>>> In short, it's to prevent the completion running before we're done with
    >>>> the iocb on the submission side.
    >>>
    >>> Yep, that's what I expected. Perhaps I missed something, but what I've
    >>> seen following code paths all the way down, it either
    >>> 1. gets error / completes synchronously and then frees req locally
    >>> 2. or passes it further (e.g. async list) and never accesses it after
    >>
    >> As soon as the IO is passed on, it can complete. In fact, it can complete
    >> even _before_ that call returns. That's the issue. Obviously this isn't
    >> true for purely polled IO, but it is true for IRQ based IO.
    >
    > And the idea was to not use io_kiocb after submission. Except when we know,
    > that it won't complete asynchronously (e.g. error), that could be checked
    > with return code, I guess.

    I think you're still missing the point. During the submission it can go
    away, it can be deep in a call chain. So it's not enough to say "we
    won't touch it after completion returns", we need to hold a reference to
    ensure it doesn't go away WHILE being submitted.

    Hope that helps!

    --
    Jens Axboe

    \
     
     \ /
      Last update: 2019-10-09 04:55    [W:8.333 / U:0.204 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site