[PATCH 0/2] ext4: reduce fast-commit write amplification for scattered writes

[Daejun Park <daejun7.park@samsung.com>]

ext4 fast commit tracks a single coalesced logical range per inode.  When
an inode is dirtied at several disjoint offsets between two commits (e.g.
sparse/scattered random writes), that range is widened to span [min, max]
of all the touched offsets, and ext4_fc_write_inode_data() then re-logs
every extent inside that span -- including the unmodified ones.  On sparse
allocation this inflates fast-commit traffic and frequently overflows the
fast-commit area, forcing a fallback to a full jbd2 commit.

This series replaces the single range with a small, bounded set of disjoint
ranges so that only the actually-modified regions are logged, while keeping
the per-inode memory cost negligible:

  1/2 tracks up to EXT4_FC_MAX_RANGES (16) disjoint ranges, merging the two
      closest ranges when the set would overflow -- so the worst case
      degrades gracefully to the old single-span behaviour.  The on-disk
      fast-commit (TLV) format is unchanged.

  2/2 allocates that array lazily: the first range is kept inline, the array
      is allocated only when a second disjoint range appears, and on an
      allocation failure we fall back to the inline single range.  The
      per-inode fast-commit footprint drops from ~140 to 20 bytes.

Measured on a sparse random-write workload (1 GiB span, R disjoint dirty
regions per fsync, 300 fsyncs, bare-metal NVMe):

  - fast-commit blocks per commit (R=16):  18.6 -> 1.0
  - full-commit fallback rate     (R=16):  22%  -> 2%   (on a small fs)
  - mean fsync latency:  R=16  -10%,  R=64  -14%
  - p99  fsync latency:  R=16  -31%

The p99 improvement comes from eliminating the full-commit fallback spikes.

Testing: crash recovery (power loss -> fast-commit replay -> verify every
fsync'd block, then e2fsck) is clean; the ext4/generic fast-commit xfstests
show no regression; the unchanged on-disk format means e2fsprogs needs no
update.  Both patches are checkpatch --strict clean.

Based on v6.17-rc3.

Daejun Park (2):
  ext4: track multiple disjoint fast-commit ranges per inode
  ext4: allocate the fast-commit range array lazily

 fs/ext4/ext4.h        |  40 +++++++--
 fs/ext4/fast_commit.c | 196 +++++++++++++++++++++++++++++++++++-------
 fs/ext4/super.c       |   1 +
 3 files changed, 199 insertions(+), 38 deletions(-)

-- 
2.43.0