Re: [PATCH RFC 6/8] SUNRPC: Reduce rpciod workqueue contention

[Tim Menninger]

I took the SLUB side of this to linux-mm as you suggested:

    https://lore.kernel.org/linux-mm/20260916232227.4098143-1-tmenninger@xxxxxxxxxxxxxxxx/

The more careful reruns also resolved the inconsistency you pointed
out between the perf-lock wait time and the idle CPU measurement.
When mpstat is collected during the same 10-second perf-lock window,
the machine is almost fully utilized and throughput drops
substantially; without that instrumentation there is significant idle
CPU remaining.

I also collected cmpxchg_double_fail over uninstrumented 10-second
windows in several NUMA/CQ placements. Those numbers are included in
the linux-mm report.

Thanks again for helping separate the allocator behavior from the NFS
regression.