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

[Uladzislau Rezki]

On Mon, Mar 30, 2026 at 06:05:52PM +0200, Uladzislau Rezki (Sony) 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.
> 
> Also simplify purge helper scheduling by removing cpumask-based
> iteration in favour to iterating directly over vmap nodes with
> pending work.
> 
> Cc: lirongqing <lirongqing@xxxxxxxxx>
> Link: https://lore.kernel.org/all/20260319074307.2325-1-lirongqing@xxxxxxxxx/
> Signed-off-by: Uladzislau Rezki (Sony) <urezki@xxxxxxxxx>
> ---
>  mm/vmalloc.c | 63 ++++++++++++++++++++++++++++++++--------------------
>  1 file changed, 39 insertions(+), 24 deletions(-)
> 
> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> index 61caa55a4402..7c1ab4a57409 100644
> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -949,6 +949,7 @@ static struct vmap_node {
>  	struct list_head purge_list;
>  	struct work_struct purge_work;
>  	unsigned long nr_purged;
> +	bool work_queued;
>  } single;
>  
>  /*
> @@ -1067,6 +1068,7 @@ static void reclaim_and_purge_vmap_areas(void);
>  static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
>  static void drain_vmap_area_work(struct work_struct *work);
>  static DECLARE_WORK(drain_vmap_work, drain_vmap_area_work);
> +static struct workqueue_struct *drain_vmap_wq;
>  
>  static __cacheline_aligned_in_smp atomic_long_t nr_vmalloc_pages;
>  static __cacheline_aligned_in_smp atomic_long_t vmap_lazy_nr;
> @@ -2335,6 +2337,19 @@ static void purge_vmap_node(struct work_struct *work)
>  	reclaim_list_global(&local_list);
>  }
>  
> +static bool
> +schedule_drain_vmap_work(struct work_struct *work)
> +{
> +	struct workqueue_struct *wq = READ_ONCE(drain_vmap_wq);
> +
> +	if (wq) {
> +		queue_work(wq, work);
> +		return true;
> +	}
> +
> +	return false;
> +}
> +
>  /*
>   * Purges all lazily-freed vmap areas.
>   */
> @@ -2342,19 +2357,12 @@ static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end,
>  		bool full_pool_decay)
>  {
>  	unsigned long nr_purged_areas = 0;
> +	unsigned int nr_purge_nodes = 0;
>  	unsigned int nr_purge_helpers;
> -	static cpumask_t purge_nodes;
> -	unsigned int nr_purge_nodes;
>  	struct vmap_node *vn;
> -	int i;
>  
>  	lockdep_assert_held(&vmap_purge_lock);
>  
> -	/*
> -	 * Use cpumask to mark which node has to be processed.
> -	 */
> -	purge_nodes = CPU_MASK_NONE;
> -
>  	for_each_vmap_node(vn) {
>  		INIT_LIST_HEAD(&vn->purge_list);
>  		vn->skip_populate = full_pool_decay;
> @@ -2374,10 +2382,9 @@ static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end,
>  		end = max(end, list_last_entry(&vn->purge_list,
>  			struct vmap_area, list)->va_end);
>  
> -		cpumask_set_cpu(node_to_id(vn), &purge_nodes);
> +		nr_purge_nodes++;
>  	}
>  
> -	nr_purge_nodes = cpumask_weight(&purge_nodes);
>  	if (nr_purge_nodes > 0) {
>  		flush_tlb_kernel_range(start, end);
>  
> @@ -2385,29 +2392,25 @@ 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);
> -
> -				if (cpumask_test_cpu(i, cpu_online_mask))
> -					schedule_work_on(i, &vn->purge_work);
> -				else
> -					schedule_work(&vn->purge_work);
> -
> +				vn->work_queued = schedule_drain_vmap_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;
>  			}
>  		}
>  
> -		for_each_cpu(i, &purge_nodes) {
> -			vn = &vmap_nodes[i];
> -
> -			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;
>  			}
> @@ -2471,7 +2474,7 @@ static void free_vmap_area_noflush(struct vmap_area *va)
>  
>  	/* After this point, we may free va at any time */
>  	if (unlikely(nr_lazy > nr_lazy_max))
> -		schedule_work(&drain_vmap_work);
> +		schedule_drain_vmap_work(&drain_vmap_work);
>  }
>  
>  /*
> @@ -5483,3 +5486,15 @@ void __init vmalloc_init(void)
>  	vmap_node_shrinker->scan_objects = vmap_node_shrink_scan;
>  	shrinker_register(vmap_node_shrinker);
>  }
> +
> +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);
> -- 
> 2.47.3
> 
I will send v2 to prevent progress lose during boot.

--
Uladzislau Rezki