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

[Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>]

(2014/12/15 12:34), Lai Jiangshan wrote:
> On 12/15/2014 10:55 AM, Kamezawa Hiroyuki wrote:
>> (2014/12/15 11:48), Lai Jiangshan wrote:
>>> On 12/15/2014 10:20 AM, Kamezawa Hiroyuki wrote:
>>>> (2014/12/15 11:12), Lai Jiangshan wrote:
>>>>> 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.
>>>>
>>>> because the result of wq_calc_node_cpumask() changes ?
>>>
>>> Yes.
>>>
>>>> Do you mean some comment should be added here ? or explaination for your reply for [3/4] ?
>>>
>>> this fix [4/4] breaks the original design.
>>>
>>
>> I'm sorry that I can't understand what this patch breaks.
>
> the pwqs/pools should be kept if the node still have cpu online.

So, the fix's grand design should be

  1. drop old pwq/pools only at node offline.
  2. set proper pool->node based on online node info.
  3. update pool->node of per-cpu-pool at cpu ONLINE.

Hm. (1) is  done because cpumask_of_node() turns to be zero-filled
after all cpus on a node offlined.

But, cpu-to-node relationship cannot be available until a cpu get onlined.
It changes at every cpu onlining. So, at node online, refleshing cpumasks
of workqueues only after _all_ cpus on node are onlined seems to be the only way
but I'm not sure how to get a mask of possible cpus on a node.

Possible another way may be using auto numa balancing's numa node hint for worker scheduling.

Do you have any idea ?

Thanks,
-Kame