Re: [PATCH 4/4] workqueue: handle change in cpu-node relationship.

[Lai Jiangshan <laijs@cn.fujitsu.com>]

On 12/14/2014 12:38 AM, Kamezawa Hiroyuki wrote:
> Although workqueue detects relationship between cpu<->node at boot,
> it is finally determined in cpu_up().
> This patch tries to update pool->node using online status of cpus.
> 
> 1. When a node goes down, clear per-cpu pool's node attr.
> 2. When a cpu comes up, update per-cpu pool's node attr.
> 3. When a cpu comes up, update possinle node cpumask workqueue is using for sched.
> 4. Detect the best node for unbound pool's cpumask using the latest info.
> 
> Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
> ---
>  kernel/workqueue.c | 67 ++++++++++++++++++++++++++++++++++++++++++------------
>  1 file changed, 53 insertions(+), 14 deletions(-)
> 
> diff --git a/kernel/workqueue.c b/kernel/workqueue.c
> index 07b4eb5..259b3ba 100644
> --- a/kernel/workqueue.c
> +++ b/kernel/workqueue.c
> @@ -266,7 +266,8 @@ struct workqueue_struct {
>  static struct kmem_cache *pwq_cache;
>  
>  static cpumask_var_t *wq_numa_possible_cpumask;
> -					/* possible CPUs of each node */
> +	/* possible CPUs of each node initialized with possible info at boot.
> +           but modified at cpu hotplug to be adjusted to real info.  */
>  
>  static bool wq_disable_numa;
>  module_param_named(disable_numa, wq_disable_numa, bool, 0444);
> @@ -3449,6 +3450,31 @@ static void put_unbound_pool(struct worker_pool *pool)
>  	call_rcu_sched(&pool->rcu, rcu_free_pool);
>  }
>  
> +/*
> + * detect best node for given cpumask.
> + */
> +static int pool_detect_best_node(const struct cpumask *cpumask)
> +{
> +	int node, best, match, selected;
> +	static struct cpumask andmask; /* we're under mutex */
> +
> +	/* Is any node okay ? */
> +	if (!wq_numa_enabled ||
> +	    cpumask_subset(cpu_online_mask, cpumask))
> +		return NUMA_NO_NODE;
> +	best = 0;
> +	selected = NUMA_NO_NODE;
> +	/* select a node which contains the most cpu of cpumask */
> +	for_each_node_state(node, N_ONLINE) {
> +		cpumask_and(&andmask, cpumask, cpumask_of_node(node));
> +		match = cpumask_weight(&andmask);
> +		if (match > best)
> +			selected = node;
> +	}
> +	return selected;
> +}
> +
> +
>  /**
>   * get_unbound_pool - get a worker_pool with the specified attributes
>   * @attrs: the attributes of the worker_pool to get
> @@ -3467,7 +3493,6 @@ static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
>  {
>  	u32 hash = wqattrs_hash(attrs);
>  	struct worker_pool *pool;
> -	int node;
>  
>  	lockdep_assert_held(&wq_pool_mutex);
>  
> @@ -3492,17 +3517,7 @@ static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
>  	 * 'struct workqueue_attrs' comments for detail.
>  	 */
>  	pool->attrs->no_numa = false;
> -
> -	/* if cpumask is contained inside a NUMA node, we belong to that node */
> -	if (wq_numa_enabled) {
> -		for_each_node(node) {
> -			if (cpumask_subset(pool->attrs->cpumask,
> -					   wq_numa_possible_cpumask[node])) {
> -				pool->node = node;
> -				break;
> -			}
> -		}
> -	}
> +	pool->node = pool_detect_best_node(pool->attrs->cpumask);
>  
>  	if (worker_pool_assign_id(pool) < 0)
>  		goto fail;
> @@ -4567,7 +4582,7 @@ static int workqueue_cpu_up_callback(struct notifier_block *nfb,
>  	int cpu = (unsigned long)hcpu;
>  	struct worker_pool *pool;
>  	struct workqueue_struct *wq;
> -	int pi;
> +	int pi, node;
>  
>  	switch (action & ~CPU_TASKS_FROZEN) {
>  	case CPU_UP_PREPARE:
> @@ -4583,6 +4598,16 @@ static int workqueue_cpu_up_callback(struct notifier_block *nfb,
>  	case CPU_ONLINE:
>  		mutex_lock(&wq_pool_mutex);
>  
> +		/* now cpu <-> node info is established, update the info. */
> +		if (!wq_disable_numa) {



> +			for_each_node_state(node, N_POSSIBLE)
> +				cpumask_clear_cpu(cpu,
> +					wq_numa_possible_cpumask[node]);

The wq code try to reuse the origin pwqs/pools when the node still have cpu online.
these 3 lines of code will cause the origin pwqs/pools be on the road of dying, and
create a new set of pwqs/pools.

> +			node = cpu_to_node(cpu);
> +			cpumask_set_cpu(cpu, wq_numa_possible_cpumask[node]);
> +		}
> +		for_each_cpu_worker_pool(pool, cpu)
> +			pool->node = cpu_to_node(cpu);
>  		for_each_pool(pool, pi) {
>  			mutex_lock(&pool->attach_mutex);
>  
> @@ -4951,7 +4976,21 @@ void workqueue_register_numanode(int nid)
>  void workqueue_unregister_numanode(int nid)
>  {
>  	struct workqueue_struct *wq;
> +	const struct cpumask *nodecpumask;
> +	struct worker_pool *pool;
> +	int cpu;
>  
> +	/* at this point, cpu-to-node relationship is not lost */
> +	nodecpumask = cpumask_of_node(nid);
> +	for_each_cpu(cpu, nodecpumask) {
> +		/*
> +		 * pool is allcated at boot and assumed to be persistent,
> +		 * we cannot free this.
> +		 * Update to be NUMA_NO_NODE. This will be fixed at ONLINE
> +		 */
> +		for_each_cpu_worker_pool(pool, cpu)
> +			pool->node = NUMA_NO_NODE;
> +	}
>  	mutex_lock(&wq_pool_mutex);
>  	list_for_each_entry(wq, &workqueues, list)
>  		wq_release_unbound_numa(wq, nid);