Re: Seg fault in 1.12.5

[Steve McIntyre]

On Sat, Feb 21, 2004 at 04:58:50PM +0000, Steve McIntyre wrote:
>Guys,
>
>I've just had a bug reported in Debian about 1.12.5:
>
>===========================================================================
>
>$ cvs update -jupstream_version_2_12_0-CVS20031225 
>-jupstream_version_2_13_0-rc1-CVS20040221
>AUTHORS already contains the differences between 1.1.1.1 and 1.1.1.2
>ChangeLog already contains the differences between 1.1.1.3 and 1.1.1.5
>INSTALL already contains the differences between 1.1.1.1 and 1.1.1.3
>R INSTALL.jp
>Makefile.am already contains the differences between 1.1.1.3 and
>1.1.1.5
>cvs update: use `cvs add' to create an entry for `Makefile.in'
>Segmentation fault
>
>See attached ltrace.  Tarball of the repository is available upon
>request.
>
>Severity grave because:
> 1) This renders cvs unusable if you hit the bug
> 2) strcmp something to NULL makes me suspicious that it is part of
>    a security hole
>
>This bug was reproduced in the 1.12.2-2 version of the package.
>Package version 1.12.1-7 did not exibit the bug. (I love
>snapshot.debian.net
>:) ), and managed to apply the update.

Looks like I've maybe found it - it's simply the check around the new
optimisations in join_file():

+    /* If we know that the user file is up-to-date, then it becomes an
+     * optimization to skip the merge when rev2 is the same as the base
+     * revision.  i.e. we know that diff3(file2,file1,file2) will produce
+     * file2. Note: If rev1 did not exist in the file (rev1 == NULL),
+     * be silent.
+     */
+    if (vers->ts_user
+        && strcmp (vers->ts_user, vers->ts_rcs) == 0
+        && strcmp (rev2, vers->vn_user) == 0)
+    {
+       if (!really_quiet && rev1 != NULL)
+       {
+           cvs_output (finfo->fullname, 0);
+           cvs_output (" already contains the differences between ", 0);
+           cvs_output (rev1, 0);
+           cvs_output (" and ", 0);
+           cvs_output (rev2, 0);
+           cvs_output ("\n", 1);
+       }
+

Simply replacing the outer check with the old (1.12.1) code:

if (vers->vn_user != NULL && strcmp (rev2, vers->vn_user) == 0)

_seems_ to work, but I don't pretend to be able to follow this code at
1.30AM. Hopefully I've helped point somebody to the right area of the
code here...

-- 
Steve McIntyre, Cambridge, UK.                                steve@einval.com
We don't need no education.
We don't need no thought control.

signature.asc
Description: Digital signature