Re: [PATCH RFC v1] ksmbd: add use_unbound_wq module parameter for the ksmbd-io workqueue
From: Huiwen He
Date: Tue Nov 17 2026 - 09:39:26 EST
Hi Namjae, Stefan,
Simply routing WRITE requests to an unbound workqueue is not sufficient,
as I've observed performance regressions across a number of workloads. I've since developed a more complete approach and will post it soon.
Thanks,
Huiwen
在 2026/11/17 22:25, Namjae Jeon 写道:
On Tue, Nov 17, 2026 at 5:51 PM Stefan Chulski
<stefan.chulski@xxxxxxxxxxxx> wrote:
Okay.
On Tue, Nov 17, 2026 at 3:34 AM Namjae Jeon <linkinjeon@xxxxxxxxxx> wrote:
On Mon, Nov 16, 2026 at 9:52 PM Stefan Chulski
<stefan.chulski@xxxxxxxxxxxx> wrote:
Could you provide the benchmark setup and actual results? If
On Mon, Nov 16, 2026 at 2:10 PM Namjae Jeon <linkinjeon@xxxxxxxxxx> wrote:
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.
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.
These are the results from our embedded systems featuring dual-core ARM CPUs.
The benchmark setup is fairly simple: writing to a USB disk over the
network. Because the network traffic is a single 5-tuple TCP stream,
it is pinned to one of the CPUs (depending on the 5-tuple hash).
We tested this on two different platforms and saw the following
throughput improvements:
Platform 1: Improved from 178 MB/s to 210 MB/s
Platform 2: Improved from 215 MB/s to 263 MB/s
https://lore.kernel.org/linux-cifs/20261022092514.2590010-1-huiwen.he@xxxxxxxxx/
Huiwen’s RFC also provides a write_wq module parameter, disabled by
default, so both approaches keep per-CPU request processing as the
default. The main difference is that Huiwen’s patch routes eligible
single-channel WRITEs to a separate unbound queue, while your patch
switches the entire ksmbd-io queue to unbound.
So, Could you test Huiwen’s patch series with write_wq=1 on your
platform and compare the results with your patch ?
Thanks.