btrfs: qgroup: Don't scan leaf if we're modifying reloc tree
authorQu Wenruo <wqu@suse.com>
Thu, 4 Apr 2019 06:45:37 +0000 (14:45 +0800)
committerDavid Sterba <dsterba@suse.com>
Mon, 29 Apr 2019 17:02:49 +0000 (19:02 +0200)
commitc4140cbf35b90422be6589024f47e132eb2298b1
treed7630d95189d2cd5223305a8b61272df5f465343
parentffd4bb2a19cd29681f5b70a200654ab92619de8a
btrfs: qgroup: Don't scan leaf if we're modifying reloc tree

Since reloc tree doesn't contribute to qgroup numbers, just skip them.

This should catch the final cause of unnecessary data ref processing
when running balance of metadata with qgroups on.

The 4G data 16 snapshots test (*) should explain it pretty well:

             | delayed subtree | refactor delayed ref | this patch
---------------------------------------------------------------------
relocated    |           22653 |                22673 |         22744
qgroup dirty |          122792 |                48360 |            70
time         |          24.494 |               11.606 |         3.944

Finally, we're at the stage where qgroup + metadata balance cost no
obvious overhead.

Test environment:

Test VM:
- vRAM 8G
- vCPU 8
- block dev vitrio-blk, 'unsafe' cache mode
- host block 850evo

Test workload:
- Copy 4G data from /usr/ to one subvolume
- Create 16 snapshots of that subvolume, and modify 3 files in each
  snapshot
- Enable quota, rescan
- Time "btrfs balance start -m"

Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
fs/btrfs/extent-tree.c