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

From: Lai Jiangshan
Date: Sun Dec 14 2014 - 22:26:17 EST


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@xxxxxxxxxxxxxx>
>>>>> ---
>>>>> 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.
> Do you mean it's better to work with broken wq_numa_possible_cpumask ?
>
> I guess removing wq_numa_possible_mask entirely may be the best way
> byusing online_mask of the node.


You patch achieves the same result of "removing wq_numa_possible_mask entirely
and using cpumask_of_node()", but it breaks the original design which we don't want to do.

>
> Thanks,
> -Kame
>
>
>
>
> .
>

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/