Re: [PATCH v2] mm/vmalloc: Use dedicated unbound workqueue for vmap purge/drain

[Uladzislau Rezki <urezki@gmail.com>]

On Mon, Mar 30, 2026 at 11:56:18AM -0700, Andrew Morton wrote:
> On Mon, 30 Mar 2026 19:58:24 +0200 "Uladzislau Rezki (Sony)" <urezki@gmail.com> wrote:
> 
> > The drain_vmap_area_work() function can take >10ms to complete
> > when there are many accumulated vmap areas in a system with a
> > high CPU count, causing workqueue watchdog warnings when run
> > via schedule_work():
> > 
> > [ 2069.796205] workqueue: drain_vmap_area_work hogged CPU for >10000us 4 times, consider switching to WQ_UNBOUND
> > [ 2192.823225] workqueue: drain_vmap_area_work hogged CPU for >10000us 5 times, consider switching to WQ_UNBOUND
> > 
> > Switch to a dedicated WQ_UNBOUND workqueue to allow the scheduler to
> > run this background task on any available CPU, improving responsiveness.
> > Use WQ_MEM_RECLAIM to ensure forward progress under memory pressure.
> > 
> > If queuing work to the dedicated workqueue is not possible(during
> > early boot), fall back to processing locally to avoid losing progress.
> > 
> > Also simplify purge helper scheduling by removing cpumask-based
> > iteration in favour to iterating directly over vmap nodes with
> > pending work.
> 
> Thanks, both.
> 
> > Cc: lirongqing <lirongqing@baidu.com>
> > Link: https://lore.kernel.org/all/20260319074307.2325-1-lirongqing@baidu.com/
> > Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
> 
> We don't want to be scaring our users with kernel warnings.  Do you
> think a Fixes: or cc:stable are justified?
> 
Probably we can CC stable.

> And I wonder if that workqueue warning should be WARN_ON_ONCE.  That
> would mean that other, later call sites wouldn't get the report, but
> we'll still get to hear about those callsites from someone else.
> 
I can switch to ONCE version. Below code:

static int __init vmalloc_init_workqueue(void)
{
	struct workqueue_struct *wq;

	wq = alloc_workqueue("vmap_drain", WQ_UNBOUND | WQ_MEM_RECLAIM, 0);
	WARN_ON(wq == NULL);
	WRITE_ONCE(drain_vmap_wq, wq);

	return 0;
}
early_initcall(vmalloc_init_workqueue);


implies to be called/initialized only once.

--
Uladzislau Rezki