Re: SIGPIPE echoing to fifo exits zsh

[Bart Schaefer <schaefer@brasslantern.com>]

On Jan 15,  4:43pm, Christian Neukirchen wrote:
}
} juno% repeat 1000 echo a >/tmp/fifo
} 
} Program received signal SIGPIPE, Broken pipe.
}
} Using /bin/echo (coreutils 8.20), the code works flawlessy (all lines
} get through the FIFO).

On Jan 16,  7:16am, Han Pingtian wrote:
} Subject: Re: SIGPIPE echoing to fifo exits zsh
}
} Looks like the crash occurs at this line in bin_print():
} 
}     4084         /* Testing EBADF special-cases >&- redirections */
}     4085         if ((fout != stdout) ? (fclose(fout) != 0) :
}     4086             (fflush(fout) != 0 && errno != EBADF)) {


Hrm.  If this were a script instead of an interactive command at the
prompt, the correct behavior would in fact be to exit on SIGPIPE.  I
don't know offhand if that's also (per POSIX spec for example) correct
in this instance.

You can prevent the shell from exiting by simply blocking the SIGPIPE
signal:

    trap '' PIPE

As for why the builtin echo gets a SIGPIPE and /bin/echo does not ...
there's a race condition involved.  With the trap above in place, if
you try

    a=0
    repeat 100 echo $((++a)) > /tmp/fifo

you'll probably see something like this on the reading end:

1
5
14
24
34
43
51
60
70
80
89
97
100

What's happening is that the the repeat loop writes a bunch of data to
the pipe (several executions of "echo") before "head -1" wakes up and
reads it.  Then "head" exits, closing the pipe, which SIGPIPE's one of
the echo commands.  A new "head" then re-opens the pipe so another few
passes of the loop go by before there's another exit+SIGPIPE event.

With /bin/echo it takes about the same amount of time for each of the
two loops to spawn their external processes, so the race is much less
likely to happen.

If you put a "sleep 1" or really almost any delay into the repeat loop
-- or use a single command like "cat" that doesn't repeatedly open/close
the read end of the pipe -- you'll see all the lines get through even
with the builtin echo.

You can see this even more glaringly by using

    while :; do head -1 ; done < /tmp/fifo

as the reading end; with that, each run of "head" will consume multiple
lines of the output from the writing end, discard all but the first line
each time, and produce output similar to

1
53
83

without ever triggering the SIGPIPE on the write side.

In short, when using FIFOs, synchronizing the amount of data written/read
by each side is the programmer's job; commands like "head" are not meant
to work that way.