Re: [PATCH v3] mm/vmalloc: Use dedicated unbound workqueues for vmap drain

[Baoquan He]

On 03/31/26 at 10:23pm, Uladzislau Rezki (Sony) wrote:
> drain_vmap_area_work() function can take >10ms to complete
> when there are many accumulated vmap areas in a system with
> high CPU count, causing workqueue watchdog warnings when run
> via schedule_work():
> 
>   workqueue: drain_vmap_area_work hogged CPU for >10000us
> 
> Move the top-level drain work to a dedicated WQ_UNBOUND
> workqueue so the scheduler can run this background work
> on any available CPU, improving responsiveness. Use the
> WQ_MEM_RECLAIM to ensure forward progress under memory
> pressure.
> 
> Move purge helpers to separate WQ_UNBOUND | WQ_MEM_RECLAIM
> workqueue. This allows drain_vmap_work to wait for helpers
> completion without creating dependency on the same rescuer
> thread and avoid a potential parent/child deadlock.
...snip...  
> @@ -2385,29 +2390,31 @@ static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end,
>  		nr_purge_helpers = atomic_long_read(&vmap_lazy_nr) / lazy_max_pages();
>  		nr_purge_helpers = clamp(nr_purge_helpers, 1U, nr_purge_nodes) - 1;
>  
> -		for_each_cpu(i, &purge_nodes) {
> -			vn = &vmap_nodes[i];
> +		for_each_vmap_node(vn) {
> +			vn->work_queued = false;
> +
> +			if (list_empty(&vn->purge_list))
> +				continue;
>  
>  			if (nr_purge_helpers > 0) {
>  				INIT_WORK(&vn->purge_work, purge_vmap_node);
> +				vn->work_queued = schedule_drain_vmap_work(
> +					READ_ONCE(drain_vmap_helpers_wq), &vn->purge_work);

The new schedule_drain_vmap_work() could submit all purge_work on one
CPU, do we need use queue_work_on(cpu, wq, work) instead?

>  
> -				if (cpumask_test_cpu(i, cpu_online_mask))
> -					schedule_work_on(i, &vn->purge_work);
> -				else
> -					schedule_work(&vn->purge_work);
> -
> -				nr_purge_helpers--;
> -			} else {
> -				vn->purge_work.func = NULL;
> -				purge_vmap_node(&vn->purge_work);
> -				nr_purged_areas += vn->nr_purged;
> +				if (vn->work_queued) {
> +					nr_purge_helpers--;
> +					continue;
> +				}
>  			}
> -		}
>  
> -		for_each_cpu(i, &purge_nodes) {
> -			vn = &vmap_nodes[i];
> +			/* Sync path. Process locally. */
> +			purge_vmap_node(&vn->purge_work);
> +			nr_purged_areas += vn->nr_purged;
> +		}
>  
> -			if (vn->purge_work.func) {
> +		/* Wait for completion if queued any. */
> +		for_each_vmap_node(vn) {
> +			if (vn->work_queued) {
>  				flush_work(&vn->purge_work);
>  				nr_purged_areas += vn->nr_purged;
>  			}
...snip...
> +
> +static int __init vmalloc_init_workqueue(void)
> +{
> +	struct workqueue_struct *drain_wq, *helpers_wq;

Maybe there's one local variable is enough like below:

	struct workqueue_struct *wq;
	unsigned int flags = WQ_UNBOUND | WQ_MEM_RECLAIM;

	wq = alloc_workqueue("vmap_drain", flags, 0);
	WARN_ON_ONCE(wq == NULL);
	WRITE_ONCE(drain_vmap_wq, wq);

	wq = alloc_workqueue("vmap_drain_helpers", flags, 0);
	WARN_ON_ONCE(wq == NULL);
	WRITE_ONCE(drain_vmap_helpers_wq, wq);

	return 0;
}

Just personal preference on nitpick, not strong opionion.