Re: [PATCH RFC v1] ksmbd: add use_unbound_wq module parameter for the ksmbd-io workqueue
From: Namjae Jeon
Date: Mon Nov 16 2026 - 20:34:31 EST
On Mon, Nov 16, 2026 at 9:52 PM Stefan Chulski
<stefan.chulski@xxxxxxxxxxxx> wrote:
>
> On Mon, Nov 16, 2026 at 2:10 PM Namjae Jeon <linkinjeon@xxxxxxxxxx> wrote:
> >
> > On Mon, Nov 16, 2026 at 4:49 PM Stefan Chulski
> > <stefan.chulski@xxxxxxxxxxxx> wrote:
> > >
> > > The "ksmbd-io" workqueue has always been per-CPU. Commit 69635d7f4b34
> > > ("fs: WQ_PERCPU added to alloc_workqueue users") only made that
> > > default explicit as part of the treewide transition to unbound-by-default;
> > > it was not a ksmbd-specific decision.
> > >
> > > Per-CPU work keeps request processing on the CPU that queued it, which
> > > is good for cache locality, but it also runs work on CPUs that were
> > > isolated from workqueue noise (isolcpus=, nohz_full=) and prevents the
> > > scheduler from placing blocked or long-running workers elsewhere.
> > > Which behaviour is better depends on the workload and platform.
> > >
> > > Add a boolean use_unbound_wq module parameter. It defaults to false, so
> > > behaviour is unchanged unless it is set; when true, "ksmbd-io" is
> > > allocated with WQ_UNBOUND instead of WQ_PERCPU. The "ksmbd-break"
> > > workqueue is already WQ_UNBOUND and is not affected.
> > >
> > > Document the parameter in ksmbd.rst.
> > >
> > > Signed-off-by: Stefan Chulski <stefan.chulski@xxxxxxxxxxxx>
> > Hi Stefan,
> >
> > Do you have numbers comparing the per-CPU and unbound queues,
>
> Yes, on our systems we observed up to an 18% performance improvement
> when using the unbound workqueue.
>
> The primary reason for this improvement is that with the default
> per-CPU workqueue, ksmbd-io work can sometimes start on a CPU that is
> already heavily loaded. Because it is bound to that specific CPU, the
> scheduler cannot move it to a less loaded or idle CPU, leading to
> bottlenecks. By using WQ_UNBOUND, the scheduler is free to migrate the
> work to less loaded CPUs, which significantly improves our CPU cycle
> efficiency.
>
> > How does this compare with the earlier following RFC ?
> >
> > https://lore.kernel.org/linux-cifs/20261022092514.2590010-1-huiwen.he@xxxxxxxxx/
> >
> > Thanks.
>
> While both address the underlying scheduling limitations of per-CPU
> workqueues for ksmbd, my patch introduces the `use_unbound_wq` module
> parameter rather than changing the behavior unconditionally. Because
> the optimal behavior (cache locality vs. load balancing) heavily
> depends on the specific workload and platform, this approach keeps the
> existing per-CPU behavior as the default while allowing platforms to
> easily opt-in to the unbound workqueue when it yields better
> performance.
Could you provide the benchmark setup and actual results? If
WQ_UNBOUND generally performs better across representative workloads,
We could make it the default and provide a module parameter to opt
into WQ_PERCPU instead.
Thanks!
>
> Thanks,
> Stefan