Re: [Bug 11175] New: ext3 BUG in add_dirent_to_buf+0x6c/0x269

[Theodore Tso <tytso@mit.edu>]

On Tue, Jul 29, 2008 at 10:03:38PM -0500, Eric Sandeen wrote:
> Theodore Tso wrote:
> > Hmm... disassembling the code, it's pretty clear the problem is here
> > in do_split(), around line 1208:
> > 
> > 	map = (struct dx_map_entry *) (data2 + blocksize);
> > 	count = dx_make_map ((struct ext3_dir_entry_2 *) data1,
> > 			     blocksize, hinfo, map);
> > 	map -= count;
> > 	dx_sort_map (map, count);
> > 	/* Split the existing block in the middle, size-wise */
> > 	size = 0;
> > 	move = 0;
> > 	for (i = count-1; i >= 0; i--) {
> > 		/* is more than half of this entry in 2nd half of the block? */
> > 		if (size + map[i].size/2 > blocksize/2)  <====
> 
> You sure this isn't our old friend
> https://bugzilla.redhat.com/show_bug.cgi?id=451068 ?
> 
> which version of gcc compiled this?

As we discussed on IRC, I think you're theory is dead on.  %ecx is at
the very end of the page-2, which would correspond to
map[count-1].size.  And size (%esi) is zero, which rules out my scenario.

This very much looks like a GCC bug.  Asheesh, can you confirm which
version of GCC you used to build your kernel?

Longer term, do_split() was coded in a very non-robust fashion.
Looking at do_split(), it was pretty easy to imagine corrupted
directory blocks that might force count to be 0 (causing the for loop
to do something insane, since i is unsigned), and adding some checks
to make sure that the split variable is neither 0 nor equal to count
might also be a really good idea.

					- Ted