[PATCH v3 00/12] NFSv4/flexfiles: scale layoutstats for striped layouts

[Mike Snitzer <snitzer@kernel.org>]

Hi,

Each stripe of a flexfiles mirror has its own read_stat and write_stat,
but the lock that protects them is in the mirror.  For a RAID0-style
layout (one mirror, many stripes) that lock is per file, and every DS
RPC takes it twice.

A 64-core client reads one file from a 32-stripe layout with a 128KiB
stripe unit (~36.7 GB/s, ~275k NFSv3 READ RPCs/s).  perf puts 45% of
its CPU in native_queued_spin_lock_slowpath, and 73% of the lock
contention events in the two layoutstats paths.

This series removes that bottleneck.  It also fixes which stripes a
LAYOUTSTATS report contains for a striped mirror.

Patches 1-2 (Jonathan) move the lock into the stripe and remove the
mirror lock.

Patches 3-8 keep the per-stripe hot path in the stripe's own
cachelines.  After patch 8, an I/O dirties one cacheline, and reads and
writes to one stripe share no lock.

Patches 9-10 fix LAYOUTSTATS for striped mirrors.  A report contained
only the lowest usable stripe of each mirror.  It now contains each
stripe with new completions, and rotates through the stripes when there
are more than four.

Patch 11 moves the report decision out of the critical section.

Patch 12 combines the read and write start helpers.

The wire encoding does not change.

Changes since v2:
- last_report_time (patch 1) and start_time (patch 4) are now
  atomic64_t, updated with atomic64_try_cmpxchg().  try_cmpxchg64() on a
  plain 64-bit field does not build where the CPU has no 64-bit
  cmpxchg: the kernel test robot reported it for ARMv7-M.  This is what
  Anna asked about on v1.  Checked with an ARMv7-M build of each patch.

Changes since v1, from Anna's review:
- Shorter commit messages throughout.  Removed the paragraphs and the
  "what the server sees" passages that Anna named.
- Code comments: the cacheline explanations are now only in the commit
  messages.  Patch 3 keeps the original lock comment.  Patch 7 cuts the
  ops_requested comment to its locking note.  Patch 9 cuts the
  last_reported_ops comment to its first sentence.  Patch 10 removes the
  scan comment.
- Patch 10: renamed dss_report_start to dss_report_index and removed its
  comment.
- New patch 12: one start helper for reads and writes.
- American spelling throughout.

v2: https://lore.kernel.org/linux-nfs/20261106001251.73472-1-snitzer@kernel.org/
v1: https://lore.kernel.org/linux-nfs/20261102191109.13470-1-snitzer@kernel.org/

The series is based on Anna's linux-next (nfs-for-7.4-1).

Thanks,
Mike

Jonathan Flynn (2):
  NFSv4/flexfiles: move layoutstats accounting to a per-stripe lock
  NFSv4/flexfiles: drop the now-unused per-mirror lock

Mike Snitzer (10):
  NFSv4/flexfiles: cacheline align the per-stripe layoutstats
  NFSv4/flexfiles: make dss_info->start_time write-once
  NFSv4/flexfiles: drop the redundant atomic from the busy timer
  NFSv4/flexfiles: pair the in-flight count with the layoutstats updates
  NFSv4/flexfiles: derive ffil_ops_requested rather than storing it
  NFSv4/flexfiles: give each direction its own lock and cacheline
  NFSv4/flexfiles: drop NFS4_FF_MIRROR_STAT_AVAIL
  NFSv4/flexfiles: rotate LAYOUTSTATS reporting across a mirror's stripes
  NFSv4/flexfiles: take the report decision out of the critical section
  NFSv4/flexfiles: share one helper for the start of a read and a write

 fs/nfs/flexfilelayout/flexfilelayout.c | 259 +++++++++++++------------
 fs/nfs/flexfilelayout/flexfilelayout.h |  37 ++--
 2 files changed, 153 insertions(+), 143 deletions(-)


base-commit: 8cdff0c1549d3fce6bb652e6bc5dd7546e1cd1e3
-- 
2.52.0