btrfs: fix race that makes inode logging fallback to transaction commit

btrfs: fix race that makes inode logging fallback to transaction commit

When logging an inode and the previous transaction is still committing, we
have a time window where we can end up incorrectly think an inode has its
last_unlink_trans field with a value greater than the last transaction
committed, which results in the logging to fallback to a full transaction
commit, which is usually much more expensive than doing a log commit.

The race is described by the following steps:

1) We are at transaction 1000;

2) We modify an inode X (a directory) using transaction 1000 and set its
   last_unlink_trans field to 1000, because for example we removed one
   of its subdirectories;

3) We create a new inode Y with a dentry in inode X using transaction 1000,
   so its generation field is set to 1000;

4) The commit for transaction 1000 is started by task A;

5) The task committing transaction 1000 sets the transaction state to
   unblocked, writes the dirty extent buffers and the super blocks, then
   unlocks tree_log_mutex;

6) Some task starts a new transaction with a generation of 1001;

7) We do some modification to inode Y (using transaction 1001);

8) The transaction 1000 commit starts unpinning extents. At this point
   fs_info->last_trans_committed still has a value of 999;

9) Task B starts an fsync on inode Y, and gets a handle for transaction
   1001. When it gets to check_parent_dirs_for_sync() it does the checking
   of the ancestor dentries because the following check does not evaluate
   to true:

       if (S_ISREG(inode->vfs_inode.i_mode) &&
           inode->generation <= last_committed &&
           inode->last_unlink_trans <= last_committed)
               goto out;

   The generation value for inode Y is 1000 and last_committed, which has
   the value read from fs_info->last_trans_committed, has a value of 999,
   so that check evaluates to false and we proceed to check the ancestor
   inodes.

   Once we get to the first ancestor, inode X, we call
   btrfs_must_commit_transaction() on it, which evaluates to true:

   static bool btrfs_must_commit_transaction(...)
   {
       struct btrfs_fs_info *fs_info = inode->root->fs_info;
       bool ret = false;

       mutex_lock(&inode->log_mutex);
       if (inode->last_unlink_trans > fs_info->last_trans_committed) {
           /*
            * Make sure any commits to the log are forced to be full
            * commits.
            */
            btrfs_set_log_full_commit(trans);
            ret = true;
       }
    (...)

    because inode's X last_unlink_trans has a value of 1000 and
    fs_info->last_trans_committed still has a value of 999, it returns
    true to check_parent_dirs_for_sync(), making it return 1 which is
    propagated up to btrfs_sync_file(), causing it to fallback to a full
    transaction commit of transaction 1001.

    We should have not fallen back to commit transaction 1001, since inode
    X had last_unlink_trans set to 1000 and the super blocks for
    transaction 1000 were already written. So while not resulting in a
    functional problem, it leads to a lot more work and higher latencies
    for a fsync since committing a transaction is usually more expensive
    than committing a log (if other filesystem changes happened under that
    transaction).

Similar problem happens when logging directories, for the same reason as
btrfs_must_commit_transaction() returns true on an inode with its
last_unlink_trans having the generation of the previous transaction and
that transaction is still committing, unpinning its freed extents.

So fix this by comparing last_unlink_trans with the id of the current
transaction instead of fs_info->last_trans_committed.

This case is often hit when running dbench for a long enough duration, as
it does lots of rename and rmdir operations (both update the field
last_unlink_trans of an inode) and fsyncs of files and directories.

This patch belongs to a patch set that is comprised of the following
patches:

  btrfs: fix race causing unnecessary inode logging during link and rename
  btrfs: fix race that results in logging old extents during a fast fsync
  btrfs: fix race that causes unnecessary logging of ancestor inodes
  btrfs: fix race that makes inode logging fallback to transaction commit
  btrfs: fix race leading to unnecessary transaction commit when logging inode
  btrfs: do not block inode logging for so long during transaction commit

Performance results are mentioned in the change log of the last patch.

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>