Re: [bug] 3.9-rc7+next: NULL deref in btrfs_next_old_leaf/btrfs_search_slot

[David Sterba <dsterba@suse.cz>]

On Thu, Apr 18, 2013 at 04:42:18PM +0200, David Sterba wrote:
> [64394.422743] BUG: unable to handle kernel NULL pointer dereference at 0000000000000078
> [64394.426716] RIP: 0010:[<ffffffffa0010e0f>]  [<ffffffffa0010e0f>] btrfs_search_slot+0xbf/0x9e0 [btrfs]
> [64394.426716]  [<ffffffffa0014977>] btrfs_next_old_leaf+0x247/0x4e0 [btrfs]
> [64394.426716]  [<ffffffffa0014c20>] btrfs_next_leaf+0x10/0x20 [btrfs]

The bisection set was reduced to these patches (on top of
cmason/for-linus 4bc4bee45):

1  Btrfs: cleanup unused function				Liu Bo
2  btrfs: enhance superblock checks				David Sterba
3  Btrfs: add some free space cache tests			Josef Bacik
4  btrfs: merge save_error_info helpers into one		David Sterba
5  btrfs: clean up transaction abort messages			David Sterba
6  Btrfs: cleanup unused arguments of btrfs_csum_data		Liu Bo
7  Btrfs: use helper to cleanup tree roots			Liu Bo
8  Btrfs: add a incompatible format change for smaller metadata extent ref  Josef Bacik

The bisecting process points to patch 8, ie the 'first bad', but the
reproducer has proved to be unreliable and I think it's very sensitive
to scheduling timing.

Reproducer is to simply run 273 in a loop on a single/single filesystem,
the null deref happens during umount. The tricky part is that it does not
happen every time, I must not touch the testbox nor let any process run
except 'dstat'. It usually crashed on first or second test run, but I've
left it up to 10 to be sure.

I don't think it's caused by the skinny metadata, because it does not
touch the affected functions, but somehow helps to make it visible.

Although the test never crashed when patch 7
  Btrfs: use helper to cleanup tree roots
  http://git.kernel.org/cgit/linux/kernel/git/josef/btrfs-next.git/commit/?id=3fa215686c574d26f43f1bcf6c9f69658e02908f
was on top (I once left it running overnight, all fine), that's my main
suspect:

--- a/fs/btrfs/disk-io.c
+++ b/fs/btrfs/disk-io.c
@@ -3483,20 +3483,7 @@ int close_ctree(struct btrfs_root *root)
 		       percpu_counter_sum(&fs_info->delalloc_bytes));
 	}
 
-	free_extent_buffer(fs_info->extent_root->node);
-	free_extent_buffer(fs_info->extent_root->commit_root);
-	free_extent_buffer(fs_info->tree_root->node);
-	free_extent_buffer(fs_info->tree_root->commit_root);
-	free_extent_buffer(fs_info->chunk_root->node);
-	free_extent_buffer(fs_info->chunk_root->commit_root);
-	free_extent_buffer(fs_info->dev_root->node);
-	free_extent_buffer(fs_info->dev_root->commit_root);
-	free_extent_buffer(fs_info->csum_root->node);
-	free_extent_buffer(fs_info->csum_root->commit_root);
-	if (fs_info->quota_root) {
-		free_extent_buffer(fs_info->quota_root->node);
-		free_extent_buffer(fs_info->quota_root->commit_root);
-	}
+	free_root_pointers(fs_info, 1);

and if you look what free_root_pointers does, sets all the pointers to NULL.

The crash site:

gdb) l *(btrfs_search_slot+0xb6)
0x11076 is in btrfs_search_slot (/home/dsterba/linux-2.6/arch/x86/include/asm/atomic.h:95).
90       *
91       * Atomically increments @v by 1.
92       */
93      static inline void atomic_inc(atomic_t *v)
94      {
95              asm volatile(LOCK_PREFIX "incl %0"
96                           : "+m" (v->counter));
97      }

(gdb) l *(btrfs_search_slot+0xb5)
0x11075 is in btrfs_search_slot (fs/btrfs/ctree.c:2513).
2508            if (p->search_commit_root) {
2509                    /*
2510                     * the commit roots are read only
2511                     * so we always do read locks
2512                     */
2513                    b = root->commit_root;
2514                    extent_buffer_get(b);
2515                    level = btrfs_header_level(b);
2516                    if (!p->skip_locking)
2517                            btrfs_tree_read_lock(b);

so it's the extent_buffer_get() call, the offset of ->refs field is 120 = 0x78,
matches "NULL pointer dereference at 0000000000000078". So root->commit_root
is NULL.

If we look what happens in disk-io.c::close_ctree:

3476         fs_info->closing = 2;
3477         smp_mb();
...
3486         free_root_pointers(fs_info, 1);
3487
3488         btrfs_free_block_groups(fs_info);
3489
3490         del_fs_roots(fs_info);
3491
3492         iput(fs_info->btree_inode);
3493
3494         btrfs_stop_workers(&fs_info->generic_worker);
...
3507         btrfs_stop_workers(&fs_info->caching_workers);

Line 3477 assures that caching thread will exit the main loop when it sees
fs_closing > 1, but if it's past this check yet has to do some work, and
close_tree() calls free_root_pointers(), bang. As we can see the caching thread
goes down later.

What I see here is a use-after-free that has been undetected so far and only
exposed by patch 7.

This diff looks like the fix (on top of patch 8):

--- a/fs/btrfs/disk-io.c
+++ b/fs/btrfs/disk-io.c
@@ -3486,14 +3486,14 @@ int close_ctree(struct btrfs_root *root)
                       percpu_counter_sum(&fs_info->delalloc_bytes));
        }

-       free_root_pointers(fs_info, 1);
-
        btrfs_free_block_groups(fs_info);

        del_fs_roots(fs_info);

        iput(fs_info->btree_inode);

+       free_root_pointers(fs_info, 1);
+
        btrfs_stop_workers(&fs_info->generic_worker);
        btrfs_stop_workers(&fs_info->fixup_workers);
        btrfs_stop_workers(&fs_info->delalloc_workers);
---


david