Re: hardlink(1) has buffer overflows, is unsafe on changing trees

[Solar Designer <solar () openwall com>]

On Sat, Oct 22, 2011 at 05:19:03AM +0400, Solar Designer wrote:
> On Sat, Oct 22, 2011 at 04:56:21AM +0400, Solar Designer wrote:
> >       strcpy (p, di->d_name);
> >
> > where "p" points somewhere inside nambuf1.
> >
> > These will just need different reproducers.
>
> Actually, I think my proposed reproducer (many nested 250-char dirs)
> triggers this one and not the strcat().  On one build, hardlink then
> crashes after dereferencing the "dirs" pointer, which happens to be
> overwritten with a directory name.  On another build (different gcc
> version and arch), hardlink does not crash (although I think it would on
> even more nested directories), but reports a ridiculous directory count
> (so "ndirs" is overwritten).  -D_FORTIFY_SOURCE=2 didn't make a
> difference here (different program binary, same observed behavior).

I investigated the non-crashing build further.  No, adding more
directories did not cause a crash either.  What happens is that lstat()
starts failing with ENAMETOOLONG shortly _after_ the overflow occurs.
This happens to limit the largest overflow size.  If "dirs" is not yet
overwritten by this point (was not reached by the overflow), then the
program may proceed without crashing and without descending to deeper
directories (thus not overflowing the buffer even further).  So
different builds may be affected to a different extent, depending on
relative placement of variables in .bss.  The behavior may also vary by
kernel version, though (when lstat() starts to fail is a property of the
kernel, whereas NAMELEN in hardlink.c is fixed).  I am able to make this
build crash with "*** buffer overflow detected ***" on the strcat(),
though, by carefully adjusting the directory name lengths (but that's
relatively uninteresting).

Alexander