io-uring.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: Glauber Costa <glauber@scylladb.com>
To: Avi Kivity <avi@scylladb.com>
Cc: Pavel Begunkov <asml.silence@gmail.com>,
	io-uring@vger.kernel.org, Jens Axboe <axboe@kernel.dk>
Subject: Re: shutdown not affecting connection?
Date: Sat, 8 Feb 2020 15:43:37 -0500	[thread overview]
Message-ID: <CAD-J=zbxqDD_=Q-Y6T5DPycdKY=aDmvrjP08QSiuWao851UGUA@mail.gmail.com> (raw)
In-Reply-To: <d651f706-68eb-0f15-6e5d-3919eb90f3da@scylladb.com>

On Sat, Feb 8, 2020 at 3:29 PM Avi Kivity <avi@scylladb.com> wrote:
>
> On 2/8/20 10:20 PM, Glauber Costa wrote:
> >>
> >>> Perhaps you can reduce the
> >>> problem to a small C reproducer?
> >>>
> >> That was my intended next step, yes
> > s***, I didn't resist and I had to explain to my wife that no, I don't
> > like io_uring more than I like her.
> >
> > But here it is.
> >
> > This is a modification of test/connect.c.
> > I added a pthread comparison example that should achieve the same
> > sequence of events:
> > - try to sync connect
> > - wait a bit
> > - shutdown
> >
> > I added a fixed wait for pthread to make sure that shutdown is not
> > called before connect.
> >
> > For io_uring, the shutdown is configurable with the program argument.
> > This works just fine if I sleep before shutdown (as I would expect from a race).
> > This hangs every time if I don't.
> >
> > Unless I am missing something I don't think this is the expected behavior
>
>
> I think it is understandable. Since the socket is blocking uring moves
> the work to a workqueue, and the shutdown() happens before the workqueue
> has had a chance to process the connection attempt. So we'll have to
> cancel the sqe.

It does seem to me that this implies that every shutdown must imply a cancel
to a connection.

From the user's perspective, this still feels like a bug to me:
the fact that we had to move this to a work queue is an implementation detail:
1) we asked the kernel to do something
2) the kernel returned
3) we called shutdown() to expecting that cancel to go away and never returned.

If cancel-after-connect to avoid these races is the intended behavior,
it would be nice to
get this documented somehow in the io_uring fantastic documentation.

In hindsight, cancel-on-shutdown is quite obvious and natural.
But I just spent two days to make this obvious and natural.

>
>
> Jens, does the blocking connect doesn't consume a kernel thread while
> it's waiting for a connection? Or does it just set things up and move on?
>

  reply	other threads:[~2020-02-08 20:43 UTC|newest]

Thread overview: 10+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2020-02-08 13:55 shutdown not affecting connection? Glauber Costa
2020-02-08 14:26 ` Pavel Begunkov
2020-02-08 18:42   ` Glauber Costa
2020-02-08 18:48     ` Avi Kivity
2020-02-08 18:57       ` Glauber Costa
2020-02-08 20:20         ` Glauber Costa
2020-02-08 20:28           ` Avi Kivity
2020-02-08 20:43             ` Glauber Costa [this message]
2020-02-08 18:48 ` Andres Freund
2020-02-08 18:54   ` Glauber Costa

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to='CAD-J=zbxqDD_=Q-Y6T5DPycdKY=aDmvrjP08QSiuWao851UGUA@mail.gmail.com' \
    --to=glauber@scylladb.com \
    --cc=asml.silence@gmail.com \
    --cc=avi@scylladb.com \
    --cc=axboe@kernel.dk \
    --cc=io-uring@vger.kernel.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).