Re: [PATCH] md/raid5: fix write stalls when a write is queued during a drain
From: Miroslav Cimerman
Date: Sun Nov 15 2026 - 07:55:47 EST
Anton Gavriliuk tested the patch on the same machine as the raid6
store patch: HPE DL345 Gen11, AMD EPYC 9455, 6x PCIe 5.0 NVMe, RAID-6
4+2, chunk 512K, no bitmap, rmw_level=2, v7.4-rc2 with and without
the patch, fio 4k randwrite, O_DIRECT, io_uring, 60 s, 2 runs per
kernel. Slowest write and IOPS:
md defaults, 1 job x QD64:
without max 59.3 s / 45.1 s 128k / 124k IOPS
with max 2.9 ms / 2.9 ms 129k / 126k IOPS
group_thread_cnt=8, stripe_cache_size=32768, 16 jobs x QD64:
without max 56.4 s / 57.5 s 472k / 474k IOPS
with max 9.2 ms / 11.2 ms 483k / 479k IOPS
Without the patch the slowest write of every run took 45-59 s of the
60 s. With the patch a few milliseconds. IOPS stayed the same or went
up a bit.
Miroslav