Re: [Quilt-dev] workaround "received broken pipe signal" bash bug (was [patch 2/8])

[Gary V. Vaughan]

Hi Jean, John,

Jean Delvare wrote:
> [Gary V. Vaughan]
>
> > This does (I think there needs to be a separate long running non-shell
> > process generating the output):
> >
> > $ cvs diff | (exit 1)
>
> Unfortunately this is not a portable test (you need to be in a CVS
> repository.)

Indeed, I was just looking for a quick way to see whether one is running
a shell that has the bug.

> This seems to work for me:
> cat /dev/zero | true

I don't see a pipe error when I do this.

> Note that your "cvs diff" example above is interesting by itself. I
> quite frequently run "cvs diff | less" and happened to get "Broken
> pipe" error messages when quitting less before hitting the bottom of
> the file. Obviously, now that I recompiled bash with the fix, it
> doesn't show anymore. This clearly demonstrates that the problem is
> with bash in the first place, and several other applications or use
> cases are affected, not just quilt, so there's nothing to be fixed on
> our side.

I agree whole-heartedly that the bug is the fault of bash.  And I
concede that distribution vendors can work around it for the sake
of quilt by compiling the bash they ship with appropriate options.
However, I think it hurts quilt to see a spurious error message or
two part way through some of the commands it is running... it's only
because I recognised the error that I didn't give up on quilt as
being low quality when it started doing this for me.

I think there are two good approaches to saving face with users
here:

  1. Workaround the known bash bug in applications built on bash.
  2. Have configure choke and give explicit instructions on where
     to get and how to compile and install a fresh bash that doesn't
     exhibit the bug.

Anything else is just passing the buck IMHO.

Unfortunately, I can't come up with a reliable test case for whether
bash has the problem :-(

> > There is a compile time option to bash (DONT_REPORT_SIGPIPE) which
> > turns off the error message, but it is not part of the default build,
> > so many vendor supplied and self built bash installations exhibit the
> > problem.
>
>
> At least Red Hat/Fedora and Debian enable that compile time option.
> Gentoo did not the last time I checked but I would be very surprised if
> they don't change their mind soon. Slackware doesn't have the option
> and (unfortunately) I don't expect them to change. Can't remember
> which of Mandriva and SuSE has it, I'll check again. The bottom line is
> that 3 out of 6 major Linux distributions already use the compile time
> option. I still need to check *BSD.

Bash is installed as a binary package in very many more places than just
a handful of Linux distributors who happen to have got it right. 
*Those* installations won't be kept up to date as well as even debian,
and it's where a big portion of quilt's potential user base live.

> To me the solution is quite simple. Any distribution providing a bash
> package and a quilt package have to enable the compile time option. If
> they don't, this is a distribution bug and they have to fix it. For
> other users, they have to install quilt by themselves already, so
> installing bash shouldn't be a problem.

Any we can help the distributions do that by having the quilt build
tell them what is wrong with bash and how to fix it.

> >  {
> >         @AWK@ '
> >         $1 == "***" || $1 == "---" \
> > -               { exit }
> > +               { stop=1 }
> > +       stop    { next }

> I don't buy this, sorry. This hurts readability and performance. And we
> would need to do this in several places. And we would then have a hard
> time remembering that some constructs should be avoided in later
> additions to quilt. So let's just not do it.

Okay.  Was just a thought :-D  I have compiled

> If you really want things to improve, you should:
> * Ask your distribution to enable the compile time option in their bash
> package (unless they already do it.)
> * Ask the bash people to change their mind and at least make it a runtime
> option, if not the default.
> If enough people ask for this, it might happen.

Maybe.  Eventually.  But it will be literally *years* before any 
significant proportion of bash users are running that hypothetical
build.  We should really try to work around the bug *now*.

Cheers,
        Gary.
--
Gary V. Vaughan      ())_.  address@hidden,gnu.org}
Research Scientist   ( '/   http://tkd.kicks-ass.net
GNU Hacker           / )=   http://www.gnu.org/software/libtool
Technical Author   `(_~)_   http://sources.redhat.com/autobook

signature.asc
Description: OpenPGP digital signature