Re: Hang opening a pipe written to by a child, with SIGCHLD

Re: Hang opening a pipe written to by a child, with SIGCHLD

[Oleg Nesterov]

On 04/09, Andrew Morton wrote:
>
> Begin forwarded message:
>
> Date: Mon, 2 Apr 2012 16:46:11 +0200
> From: Lluís Batlle i Rossell <viric@viric.name>
> To: linux-kernel@vger.kernel.org
> Subject: Hang opening a pipe written to by a child, with SIGCHLD
>
>
> Hello,
>
> I had troubles running a simple bash script where a child wrote to a named pipe
> before dying, and the parent opened and read the pipe.
>
> On a computer, my script hangs always. Also under 'strace'. Not under 'strace
> -f'. Thinking of some race, I wrote a small script that hangs in the linuxes I
> could try (3.2.11 now):
>
> ----------
> #!/var/run/current-system/sw/bin/bash
>
> PIPE=/tmp/pipe
>
> rm -f $PIPE
> mkfifo $PIPE
>
> function spawn {
>     echo DONE > $PIPE
> }
>
> spawn sleep 1 &
>
> while true; do
>     echo reading
>     while read LINE < $PIPE; do
>         echo $LINE
>         spawn &
>     done
> done

Looks like, it hangs "correctly". At least from the kernel pov.

The parent sleeps in fifo_open()->wait_for_partner(), but there
are no pipe->writers.

Note that the parent (bash) re-opens /tmp/pipe after every "spawn".
Now,

	- the parent (bash) forks the new child

	- the new child sleeps waiting for reader

	- the parent does open(pipe), this wakes up the child

	- the child opens the pipe, writes, and exits

	- the open from the child wakes up the parent, but
	  since it exits quickly fifo_open(FMODE_READ) can
	  notice the pending SIGCHLD and return -EINTR

	- the parent restarts sys_open(), but nobody can
	  open it for writing

May be you can ask bash developers. Perhaps bash can move the
"restore stdin" logic into the child process, I dunno.

Or may be you can "fix" this script, just add "exec < $PIPE"
before the main "while true" loop.

Oleg.

Re: Hang opening a pipe written to by a child, with SIGCHLD

[Oleg Nesterov]

On 04/11, Oleg Nesterov wrote:
>
> Or may be you can "fix" this script, just add "exec < $PIPE"
> before the main "while true" loop.

Cough, no ;) this can equally race with the exiting child.

Oleg.

Hang opening a pipe written to by a child, with SIGCHLD

[Lluís Batlle i Rossell]

Hello,

I had troubles running a simple bash script where a child wrote to a named pipe
before dying, and the parent opened and read the pipe.

On a computer, my script hangs always. Also under 'strace'. Not under 'strace
-f'. Thinking of some race, I wrote a small script that hangs in the linuxes I
could try (3.2.11 now):

----------
#!/var/run/current-system/sw/bin/bash

PIPE=/tmp/pipe

rm -f $PIPE
mkfifo $PIPE

function spawn {
    echo DONE > $PIPE
}

spawn sleep 1 &

while true; do
    echo reading
    while read LINE < $PIPE; do
        echo $LINE
        spawn &
    done
done
-------------

bash uses SA_RESTART on SIGCHLD handler, and nevertheless the script halts at
printing some amount of 'DONE'. Strace shows it's inside the open() call, while the open
call has been interrupted by SIGCHLD. One of the computers I have hangs at the *first* DONE always, and that's why it's annoying to me.

Can anyone reproduce this?

Regards,
Lluís.

Re: Hang opening a pipe written to by a child, with SIGCHLD

[Shea Levy]

On 04/02/2012 10:46 AM, Lluís Batlle i Rossell wrote:
> Hello,
>
> I had troubles running a simple bash script where a child wrote to a named pipe
> before dying, and the parent opened and read the pipe.
>
> On a computer, my script hangs always. Also under 'strace'. Not under 'strace
> -f'. Thinking of some race, I wrote a small script that hangs in the linuxes I
> could try (3.2.11 now):
>
> ----------
> #!/var/run/current-system/sw/bin/bash
>
> PIPE=/tmp/pipe
>
> rm -f $PIPE
> mkfifo $PIPE
>
> function spawn {
>      echo DONE>  $PIPE
> }
>
> spawn sleep 1&
>
> while true; do
>      echo reading
>      while read LINE<  $PIPE; do
>          echo $LINE
>          spawn&
>      done
> done
> -------------
>
> bash uses SA_RESTART on SIGCHLD handler, and nevertheless the script halts at
> printing some amount of 'DONE'. Strace shows it's inside the open() call, while the open
> call has been interrupted by SIGCHLD. One of the computers I have hangs at the *first* DONE always, and that's why it's annoying to me.
>
> Can anyone reproduce this?

I've reproduced this on my system as well (3.3). No idea about cause though.

~Shea

Re: Hang opening a pipe written to by a child, with SIGCHLD

[Lluís Batlle i Rossell]

On Mon, Apr 02, 2012 at 10:49:10AM -0400, Shea Levy wrote:
> On 04/02/2012 10:46 AM, Lluís Batlle i Rossell wrote:
> >Can anyone reproduce this?
> 
> I've reproduced this on my system as well (3.3). No idea about cause though.
> 

Thank you Shea,

with a little change, adding "sleep 0.1" makes it hang early in all (few)
linuces I tried: Let me know if this helps you reproduce the problem.

As before, hangs under strace, does not hang under 'strace -f'. Sorry if I'm
misunderstanding something.

New script:
----------
#!/bin/sh
                                                                                                                   
PIPE=/tmp/pipe
                                                                                                                   
rm -f $PIPE
mkfifo $PIPE
set -x
                                                                                                                   
spawn() {
    sleep 0.1
    echo DONE > $PIPE
}

spawn &

while true; do
    while read LINE < $PIPE; do
        echo $LINE
        spawn &
    done
done
-----------

Regards,
Lluís.

Re: Hang opening a pipe written to by a child, with SIGCHLD

[Lluís Batlle i Rossell]

On Mon, Apr 02, 2012 at 08:50:57PM +0200, Lluís Batlle i Rossell wrote:
> On Mon, Apr 02, 2012 at 10:49:10AM -0400, Shea Levy wrote:
> > On 04/02/2012 10:46 AM, Lluís Batlle i Rossell wrote:
> > >Can anyone reproduce this?
> > 
> > I've reproduced this on my system as well (3.3). No idea about cause though.
> > 
> 
> Thank you Shea,

All clear. I had a race condition; sorry for the noise. Proper synchronization
through close() in the child and read() = 0 in the parent fixes all.