Re: [bug-gawk] Gawk 4.0.0 Redirection Bug

[Aharon Robbins]

Applied this, with slight modifications to take Eli's suggestion
into account. It'll be pushed shortly.

Thanks,

Arnold

> Date: Fri, 30 Sep 2011 14:26:12 +0300
> From: Eli Zaretskii <address@hidden>
> Subject: Re: Gawk 4.0.0 Redirection Bug
> To: David Millis <address@hidden>, address@hidden
> Cc: address@hidden, Jim Meyering <address@hidden>
>
> > Date: Sun, 25 Sep 2011 19:38:58 -0700 (PDT)
> > From: David Millis <address@hidden>
> > 
> > # Exhausting stdin seems to break redirection in general.
> > 
> > # echo hello|gawk -f redirect.txt
> > # ... fatal: print to "tmp.txt" failed (reason unknown)
> > 
> > BEGIN {
> >   while ((getline tmpLine < "/dev/stdin") > 0);
> >   print "hi" > "tmp.txt";
> >   close("tmp.txt");
> > }
> > 
> > # 4.0.0 (Windows, http://www.klabaster.com/freeware.htm#dl) has the issue.
> > # 4.0.0 (Debian) is fine.
> > # 3.1.6 (GnuWin32) was fine.
>
> This happens because Gawk 4.0.0 introduced non-portable code in the
> new function remap_std_file, that has no chance of working on any
> non-Posix platform that doesn't emulate /dev/null in its C library.
>
> The patch to fix this (relative to stock 4.0.0) is below.
>
> Btw, I think the code of this function, even after the patch, is still
> dangerously wrong, in that it closes oldfd right away.  The code
> obviously assumes that nothing can go wrong with opening /dev/null.
> But if something does go wrong, the effect is that we forcibly close
> Gawk's standard input, which will cause the next file to be opened on
> file descriptor of zero, and all hell will break lose on it.  IOW, the
> exact problem that pre-4.0.0 Gawk was trying to avoid by not closing
> the standard handles to begin with.  This will definitely bite some
> platform that doesn't support /dev/null, but can also happen on Posix
> platforms, I think, at least under some extreme conditions, like
> shortage of available file handles.
>
> I therefore think we should restructure the code of this function so
> that oldfd is closed only as a side effect of calling `dup2'.
>
>
> 2011-09-30  Eli Zaretskii  <address@hidden>
>
>       * io.c (remap_std_file): Fix non-portable code that caused
>       redirected "print" to fail if a previous read from standard input
>       returned EOF.  Reported by David Millis <address@hidden>.
>
>
> --- io.c~0    2011-05-19 00:47:29.000000000 +0300
> +++ io.c      2011-09-30 14:10:00.031250000 +0300
> @@ -449,9 +449,14 @@ remap_std_file(int oldfd)
>       int ret = -1;
>  
>       close(oldfd);
> -     newfd = open("/dev/null", O_RDWR);
> +     /*
> +      * Give OS-specific routines in gawkmisc.c chance to interpret
> +      * "/dev/null" as appropriate for their platforms.
> +      */
> +     newfd = os_devopen("/dev/null", O_RDWR);
> +     if (newfd == INVALID_HANDLE)
> +             newfd = open("/dev/null", O_RDWR);
>       if (newfd >= 0 && newfd != oldfd) {
> -             /* dup2() will close oldfd for us first. */
>               ret = dup2(newfd, oldfd);
>               if (ret == 0)
>                       close(newfd);