btrfs: fix race on ENOMEM in alloc_extent_buffer

Consider the following interleaving of overlapping calls to
alloc_extent_buffer:

Call 1:

- Successfully allocates a few pages with find_or_create_page
- find_or_create_page fails, goto free_eb
- Unlocks the allocated pages

Call 2:
- Calls find_or_create_page and gets a page in call 1's extent_buffer
- Finds that the page is already associated with an extent_buffer
- Grabs a reference to the half-written extent_buffer and calls
  mark_extent_buffer_accessed on it

mark_extent_buffer_accessed will then try to call mark_page_accessed on
a null page and panic.

The fix is to decrement the reference count on the half-written
extent_buffer before unlocking the pages so call 2 won't use it. We
should also set exists = NULL in the case that we don't use exists to
avoid accidentally returning a freed extent_buffer in an error case.

Signed-off-by: Omar Sandoval <osandov@osandov.com>
Reviewed-by: David Sterba <dsterba@suse.cz>
Reviewed-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Chris Mason <clm@fb.com>

diff --git a/fs/btrfs/extent_io.c b/fs/btrfs/extent_io.c
index 782f3bc..ea100eb 100644 (file)
--- a/fs/btrfs/extent_io.c
+++ b/fs/btrfs/extent_io.c
@@ -4870,6 +4870,7 @@ struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info,
                                mark_extent_buffer_accessed(exists, p);
                                goto free_eb;
                        }
+                       exists = NULL;
 
                        /*
                         * Do this so attach doesn't complain and we need to
@@ -4933,12 +4934,12 @@ again:
        return eb;
 
 free_eb:
+       WARN_ON(!atomic_dec_and_test(&eb->refs));
        for (i = 0; i < num_pages; i++) {
                if (eb->pages[i])
                        unlock_page(eb->pages[i]);
        }
 
-       WARN_ON(!atomic_dec_and_test(&eb->refs));
        btrfs_release_extent_buffer(eb);
        return exists;
 }