[PATCH v2 0/3] btrfs: zoned: reduce lock contention in zoned extent allocator

[Johannes Thumshirn <jth@kernel.org>]

From: Johannes Thumshirn <johannes.thumshirn@wdc.com>

While investigating performance issues on zoned btrfs, I've analyzed
lockstat output while running fio workloads.

Three locks that have always been on top are the space_info's ->lock, as
well as fs_info's ->treelog_bg_lock and ->relocation_bg_lock.

The first one is not as trivial to get rid and most likely legit in most
cases (see patch #3). The other two are only protecting a single value in
fs_info and can be replaced by REAC_ONCE() and WRITE_ONCE().

The series has been run through fstests on tcmu-runner emulated SMR drives
as well as Damien's fio benchmarks on real ZNS hardware and no regressions
have been observed.

Changes in v2:
- Rewording of the commit message (Damien)
- No more local variables caching the values (Damien)

Johannes Thumshirn (3):
  btrfs: zoned: get rid of relocation_bg_lock
  btrfs: zoned: get rid of treelog_bg_lock
  btrfs: zoned: don't hold space_info lock on zoned allocation

 fs/btrfs/disk-io.c     |  2 -
 fs/btrfs/extent-tree.c | 86 +++++++++++-------------------------------
 fs/btrfs/fs.h          |  7 +---
 fs/btrfs/zoned.c       | 12 +++---
 fs/btrfs/zoned.h       |  6 +--
 5 files changed, 31 insertions(+), 82 deletions(-)

-- 
2.49.0