Re: gnulib portability issues

[Rich Felker]

On Sat, Jun 09, 2012 at 09:06:21PM -0700, Paul Eggert wrote:
> Patches for any of these problems would be welcome.
> As freadahead seems to be most pressing, it sounds
> like that's the best one to work on first.
> 
> > 1. freadahead is inherently non-portable
> 
> That sounds like a challenge!  How about using
> the following patch?  Does it work for you, if
> you define SLOW_BUT_NO_HACKS?  If so, perhaps
> we should just make SLOW_BUT_NO_HACKS the default.
> 
> diff --git a/lib/freadahead.c b/lib/freadahead.c
> index 2ba8b34..ad6eb13 100644
> --- a/lib/freadahead.c
> +++ b/lib/freadahead.c
> @@ -85,7 +85,7 @@ freadahead (FILE *fp)
>      return 0;
>    return fp->wp - fp->rp;
>  #elif defined SLOW_BUT_NO_HACKS     /* users can define this */
> -  abort ();
> +  fflush (fp);

Calling fflush on a stream that's reading invokes undefined behavior
per ISO C. POSIX defines the behavior for _seekable_ input, but leaves
it undefined for non-seekable. (On many systems it purges/discards the
unread input, but this is not required, and it's highly undesirable in
this usage case anyway.)

For seekable input, comparing ftello(f) and the underlying fd offset
from lseek(fileno(f),0,SEEK_CUR) would work, but that also will fail
for non-seekable input.

Here's a sketch of a possible solution, but it's horribly ugly:

tmp=dup(fileno(f));
tmp2=socket(AF_UNIX,SOCK_DGRAM,0); /* non-readable */
dup2(tmp2,fileno(f));
close(tmp2);
e = ferror(f);
c = getc(f);
if (!e) clearerr(f);
ungetc(c, f);
dup2(tmp,fileno(f));
close(tmp);
return c>=0;

It also may clear the eof flag; not sure how to best fix that yet.

While fixing the underlying freadahead function would be nice, it
might be better to make sense of WHY it's needed/wanted, if it's even
actually useful, and how to avoid the need for it in the first place.

Rich