[PATCH 00/11] NFSv4/flexfiles: scale layoutstats for striped layouts

[Mike Snitzer <snitzer@kernel.org>]

Hi,

Since striped flexfiles layouts were introduced, each stripe of a mirror
carries its own read_stat and write_stat, but the lock protecting them
stayed in the mirror.  For a RAID0-style layout (one mirror, many
stripes) that lock is effectively per-file: every DS RPC takes it twice,
once at prepare and once at completion, so striping multiplies RPC
parallelism without multiplying lock parallelism.

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

This series removes that bottleneck and, along the way, fixes what the
server is told in LAYOUTSTATS for a striped mirror.

Patches 1-2 (Jonathan) move the layoutstats lock into the stripe, next
to the statistics it protects, and remove the now-unused mirror lock.

Patches 3-8 keep the per-stripe hot path to the stripe's own cachelines:
  3: cacheline align the per-stripe statistics so neighbouring stripes
     do not false-share
  4: make dss_info->start_time write-once (try_cmpxchg64), removing its
     dependence on any lock
  5: drop the redundant atomic from the busy timer
  6: pair the in-flight count with the layoutstats updates, under the
     same lock, with lockdep assertions
  7: derive ffil_ops_requested rather than storing it
  8: give each direction its own lock and cacheline, so an I/O dirties
     one line (the stripe shrinks from 384 to 320 bytes)

Patches 9-10 fix LAYOUTSTATS reporting for striped mirrors:
  9: drop NFS4_FF_MIRROR_STAT_AVAIL.  The per-mirror bit cost two atomic
     RMWs per RPC on the mirror's shared cacheline, and since the first
     stripe reported cleared it, a report carried only one
     ff_layoutupdate4 per mirror: always the lowest usable dss_id.
 10: rotate reporting across a mirror's stripes.  Callers cap a report
     at four entries; resume each mirror's scan where the previous
     report stopped, so every busy stripe is reported within
     ceil(busy_stripes / 4) report intervals instead of only the first
     four ever being reported.

Patch 11 takes the report decision (pointer chases into the mirror and
layout, ktime_to_ms(), the last_report_time cmpxchg) out of the critical
section, leaving the lock to cover two writes on one cacheline.

The wire encoding is unchanged; the values reported remain cumulative
since the layout was granted.

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 (9):
  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

 fs/nfs/flexfilelayout/flexfilelayout.c | 251 ++++++++++++++++---------
 fs/nfs/flexfilelayout/flexfilelayout.h |  74 ++++++--
 2 files changed, 218 insertions(+), 107 deletions(-)


base-commit: ef30b3efb7e8bae94d790a6503a595e1051323e7
-- 
2.52.0