[PATCH v2 0/3] scsi/block: NUMA-local scan allocations, shared-tag path cleanup, and SCSI I/O counters

[Sumit Saxena <sumit.saxena@broadcom.com>]

This series contains three performance improvements targeting the SCSI
and block layers on multi-socket NUMA and heavily loaded SMP systems.

On multi-socket NUMA systems we observed extreme I/O throughput variance
of 50-60% between runs.  This series identifies and fixes two root causes:
cross-node memory accesses due to NUMA-unaware allocations in the scan
path, and false sharing between hot atomic counters in struct request_queue
and struct scsi_device.

Performance notes:

Tested on a dual-socket NUMA system (2x 32-core, 256 GB/socket) with
an mpi3mr HBA, running fio (random read, 4K, QD 64, 16 jobs, 60 s,
direct I/O).  IOPS figures are in KIOPS (thousands of IOPS):

  Configuration                    Avg KIOPS   Range (KIOPS)   Spread
  Baseline                         6,255       4,200 - 6,700   ~37%
  Baseline + all patches           7,350       7,000 - 7,700    ~10%

Key findings:

These patches combinedly reduces the observed 50-60% run-to-run variance
to under 10%, significantly improving workload predictability and
improves IOPs by 16-18%.

No functional regressions observed.

This patch series is based on Martin's for-next tree.

Changes in v2
--------------

  Patch 1 — Same functional goal as v1 patch 1: NUMA-local scsi_device /
  scsi_target allocations in the scan path so steady-state I/O does not
  habitually touch remote memory when the host has a fixed DMA/NUMA
  affinity.

  Patch 2 — Replaces v1’s ____cacheline_aligned_in_smp on
  nr_active_requests_shared_tags with shared tags removal work done by
  Bart Van Assche [1], rebased for the current tree; it removes the
  atomic counter- nr_active_request_shared_tags that motivated the v1
  false-sharing workaround and, in our testing, improves IOPS on the order
  of roughly 16–18% for the shared-tag workload exercised.
  This patch touches include/linux/blkdev.h, so needs review from
  linux-block@vger.kernel.org; an Acked-by from the block maintainer is
  requested before merging via the SCSI tree.

  Patch 3 — Replaces v1’s cache-line padding of iodone_cnt with
  percpu_counter for both iorequest_cnt and iodone_cnt, so submission and
  completion paths mostly update CPU-local state instead of bouncing a
  single cache line, without inflating struct scsi_device for SMP
  alignment.

[1]: https://lore.kernel.org/linux-block/20240529213921.3166462-1-bvanassche@acm.org/

James Rizzo (1):
  scsi: scan: allocate sdev and starget on the NUMA node of the host adapter

Bart Van Assche (1):
  block: drop shared-tag fairness throttling

Sumit Saxena (1):
  scsi: use percpu counters for iorequest_cnt and iodone_cnt

 12 files changed, 68 insertions(+), 145 deletions(-)

---
2.43.7