Re: btrfs thread_pool_size logic out of sync with workqueue

[Tejun Heo <tj@kernel.org>]

On Mon, Apr 28, 2025 at 06:23:39PM -0500, Junxuan Liao wrote:
> Hi all,
> 
> Commit 636b927eba5bc63375 (workqueue: Make unbound workqueues to use
> per-cpu pool_workqueues) makes max_active per CPU for unbounded
> workqueues as well, but thread_pool_size in btrfs_fs_info and the mount
> option thread_pool still assume the limit is global.
> 
> e.g. The default value of 8 allows a total of 64 workers on an 8-CPU
> machine.
> 
> As far as I know, this means that on the Btrfs side we can no longer
> control the concurrency level at the same granularity as before. We
> should rewrite the auto-scaling logic, change the default
> thread_pool_size value, and update the documentation. 
> 
> Am I missing something?

5797b1c18919 ("workqueue: Implement system-wide nr_active enforcement for
unbound workqueues") turned max_active system-wide. The count is now split
across nodes proportional to the number of cpus each node has. This is still
a different behavior from before where max_active was applied per node, but
no behavior change on single node machines and the new behavior is easier to
work with.

Thanks.

-- 
tejun