Btrfs: don't attach unnecessary extents to transaction on fsync
authorFilipe Manana <fdmanana@suse.com>
Fri, 17 Apr 2015 16:08:37 +0000 (17:08 +0100)
committerChris Mason <clm@fb.com>
Wed, 10 Jun 2015 14:02:44 +0000 (07:02 -0700)
We don't need to attach ordered extents that have completed to the current
transaction. Doing so only makes us hold memory for longer than necessary
and delaying the iput of the inode until the transaction is committed (for
each created ordered extent we do an igrab and then schedule an asynchronous
iput when the ordered extent's reference count drops to 0), preventing the
inode from being evictable until the transaction commits.

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/ordered-data.c

index ceccd078c93dd24480ae547f68148af4ea405993..89656d799ff6fa3772b52cb17d535ebdc1cbb910 100644 (file)
@@ -502,7 +502,21 @@ void btrfs_wait_logged_extents(struct btrfs_trans_handle *trans,
                wait_event(ordered->wait, test_bit(BTRFS_ORDERED_IO_DONE,
                                                   &ordered->flags));
 
-               list_add_tail(&ordered->trans_list, &trans->ordered);
+               /*
+                * If our ordered extent completed it means it updated the
+                * fs/subvol and csum trees already, so no need to make the
+                * current transaction's commit wait for it, as we end up
+                * holding memory unnecessarily and delaying the inode's iput
+                * until the transaction commit (we schedule an iput for the
+                * inode when the ordered extent's refcount drops to 0), which
+                * prevents it from being evictable until the transaction
+                * commits.
+                */
+               if (test_bit(BTRFS_ORDERED_COMPLETE, &ordered->flags))
+                       btrfs_put_ordered_extent(ordered);
+               else
+                       list_add_tail(&ordered->trans_list, &trans->ordered);
+
                spin_lock_irq(&log->log_extents_lock[index]);
        }
        spin_unlock_irq(&log->log_extents_lock[index]);