tree 173c889545c881ce14c5fa3519a37b13f41544cc
parent ef07b74e1be56f9eafda6aadebb9ebba0743c9f0
author Filipe Manana <fdmanana@suse.com> 1754478691 +0100
committer David Sterba <dsterba@suse.com> 1755817127 +0200

btrfs: fix race between setting last_dir_index_offset and inode logging

At inode_logged() if we find that the inode was not logged before we
update its ->last_dir_index_offset to (u64)-1 with the goal that the
next directory log operation will see the (u64)-1 and then figure out
it must check what was the index of the last logged dir index key and
update ->last_dir_index_offset to that key's offset (this is done in
update_last_dir_index_offset()).

This however has a possibility for a time window where a race can happen
and lead to directory logging skipping dir index keys that should be
logged. The race happens like this:

1) Task A calls inode_logged(), sees ->logged_trans as 0 and then checks
   that the inode item was logged before, but before it sets the inode's
   ->last_dir_index_offset to (u64)-1...

2) Task B is at btrfs_log_inode() which calls inode_logged() early, and
   that has set ->last_dir_index_offset to (u64)-1;

3) Task B then enters log_directory_changes() which calls
   update_last_dir_index_offset(). There it sees ->last_dir_index_offset
   is (u64)-1 and that the inode was logged before (ctx->logged_before is
   true), and so it searches for the last logged dir index key in the log
   tree and it finds that it has an offset (index) value of N, so it sets
   ->last_dir_index_offset to N, so that we can skip index keys that are
   less than or equal to N (later at process_dir_items_leaf());

4) Task A now sets ->last_dir_index_offset to (u64)-1, undoing the update
   that task B just did;

5) Task B will now skip every index key when it enters
   process_dir_items_leaf(), since ->last_dir_index_offset is (u64)-1.

Fix this by making inode_logged() not touch ->last_dir_index_offset and
initializing it to 0 when an inode is loaded (at btrfs_alloc_inode()) and
then having update_last_dir_index_offset() treat a value of 0 as meaning
we must check the log tree and update with the index of the last logged
index key. This is fine since the minimum possible value for
->last_dir_index_offset is 1 (BTRFS_DIR_START_INDEX - 1 = 2 - 1 = 1).
This also simplifies the management of ->last_dir_index_offset and now
all accesses to it are done under the inode's log_mutex.

Fixes: 0f8ce49821de ("btrfs: avoid inode logging during rename and link when possible")
Reviewed-by: Boris Burkov <boris@bur.io>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>