[PATCH 9/9] posix: Fix glob with GLOB_NOCHECK returning modified patterns (BZ#10246)

[Paul Eggert]

Although this is a definite bug and the patch fixes this instance of it, I'm 
afraid other instances remain unfixed. For example:

    glob_t g; glob ("//a*b", 0, NULL, &g)

can do the wrong thing, since glob calls opendir on "/" instead of "//", and on 
some platforms "/" and "//" are different directories (POSIX allows this as a 
special exception).

A more serious example. If you do this:

   ln -s /no-such-file globlink1
   ln -s . globlink2

then:

       glob_t g;
       int res = glob ("globlink[12]/", 0, NULL, &g);
       assert (res == 0 && g.gl_pathc == 1);
       assert (strcmp (g.gl_pathv[0], "globlink2/") == 0);

fails, since glob gets confused about directories and slashes and mistakenly 
returns two results. Although this bug is seemingly unrelated, the underlying 
cause is the same: glob gets confused about whether to include or exclude 
slashes when doing its tests.

I'll take a look at it, though the fix won't be trivial.

PS. This finishes my review of this patchset. Patches 1-8 are OK to be 
installed, with the trivial changes I suggested earlier. This patch (patch 9) 
I'd like to hold off on, until we've had a chance to work out a 
more-comprehensive fix.

PPS. I'm still slowly wending my way through your original patchset. Most 
recently I looked at "[PATCH 07/18] posix: User LOGIN_NAME_MAX for all user 
names in glob" <https://sourceware.org/ml/libc-alpha/2017-08/msg00447.html>. I'm 
afraid a good fix needs to be hairier there too, as POSIX does not require 
LOGIN_NAME_MAX to be suitable for a stack-based buffer, or even to be defined. I 
have a partly-drafted patch which I hope to finish in the not-too-distant future.