btrfs: fix directory logging due to race with concurrent index key deletion

btrfs: fix directory logging due to race with concurrent index key deletion

Sometimes we log a directory without holding its VFS lock, so while we
logging it, dir index entries may be added or removed. This typically
happens when logging a dentry from a parent directory that points to a
new directory, through log_new_dir_dentries(), or when while logging
some other inode we also need to log its parent directories (through
btrfs_log_all_parents()).

This means that while we are at log_dir_items(), we may not find a dir
index key we found before, because it was deleted in the meanwhile, so
a call to btrfs_search_slot() may return 1 (key not found). In that case
we return from log_dir_items() with a success value (the variable 'err'
has a value of 0). This can lead to a few problems, specially in the case
where the variable 'last_offset' has a value of (u64)-1 (and it's
initialized to that when it was declared):

1) By returning from log_dir_items() with success (0) and a value of
   (u64)-1 for '*last_offset_ret', we end up not logging any other dir
   index keys that follow the missing, just deleted, index key. The
   (u64)-1 value makes log_directory_changes() not call log_dir_items()
   again;

2) Before returning with success (0), log_dir_items(), will log a dir
   index range item covering a range from the last old dentry index
   (stored in the variable 'last_old_dentry_offset') to the value of
   'last_offset'. If 'last_offset' has a value of (u64)-1, then it means
   if the log is persisted and replayed after a power failure, it will
   cause deletion of all the directory entries that have an index number
   between last_old_dentry_offset + 1 and (u64)-1;

3) We can end up returning from log_dir_items() with
   ctx->last_dir_item_offset having a lower value than
   inode->last_dir_index_offset, because the former is set to the current
   key we are processing at process_dir_items_leaf(), and at the end of
   log_directory_changes() we set inode->last_dir_index_offset to the
   current value of ctx->last_dir_item_offset. So if for example a
   deletion of a lower dir index key happened, we set
   ctx->last_dir_item_offset to that index value, then if we return from
   log_dir_items() because btrfs_search_slot() returned 1, we end up
   returning from log_dir_items() with success (0) and then
   log_directory_changes() sets inode->last_dir_index_offset to a lower
   value than it had before.
   This can result in unpredictable and unexpected behaviour when we
   need to log again the directory in the same transaction, and can result
   in ending up with a log tree leaf that has duplicated keys, as we do
   batch insertions of dir index keys into a log tree.

So fix this by making log_dir_items() move on to the next dir index key
if it does not find the one it was looking for.

Reported-by: David Arendt <admin@prnet.org>
Link: https://lore.kernel.org/linux-btrfs/ae169fc6-f504-28f0-a098-6fa6a4dfb612@leemhuis.info/
CC: stable@vger.kernel.org # 4.14+
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>