Btrfs: don't commit the transaction in the async space flushing
authorJosef Bacik <jbacik@fb.com>
Wed, 18 Feb 2015 21:58:15 +0000 (13:58 -0800)
committerChris Mason <clm@fb.com>
Fri, 10 Apr 2015 21:06:54 +0000 (14:06 -0700)
We're triggering a huge number of commits from
btrfs_async_reclaim_metadata_space.  These aren't really requried,
because everyone calling the async reclaim code is going to end up
triggering a commit on their own.

Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/extent-tree.c

index ae8db3ba467cf6eba74a31cce2d952d8e51a67f7..3d4b3d680feaa4819a8c215a4f2b0a2a45c9c8a7 100644 (file)
@@ -4329,8 +4329,13 @@ out:
 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
                                        struct btrfs_fs_info *fs_info, u64 used)
 {
-       return (used >= div_factor_fine(space_info->total_bytes, 98) &&
-               !btrfs_fs_closing(fs_info) &&
+       u64 thresh = div_factor_fine(space_info->total_bytes, 98);
+
+       /* If we're just plain full then async reclaim just slows us down. */
+       if (space_info->bytes_used >= thresh)
+               return 0;
+
+       return (used >= thresh && !btrfs_fs_closing(fs_info) &&
                !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
 }
 
@@ -4385,10 +4390,7 @@ static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
                if (!btrfs_need_do_async_reclaim(space_info, fs_info,
                                                 flush_state))
                        return;
-       } while (flush_state <= COMMIT_TRANS);
-
-       if (btrfs_need_do_async_reclaim(space_info, fs_info, flush_state))
-               queue_work(system_unbound_wq, work);
+       } while (flush_state < COMMIT_TRANS);
 }
 
 void btrfs_init_async_reclaim_work(struct work_struct *work)