Re: [PATCH 1/4] workqueue:Fix unbound workqueue's node affinity detection

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

(2014/12/16 14:30), Lai Jiangshan wrote:
> On 12/15/2014 07:14 PM, Kamezawa Hiroyuki wrote:
>> Unbound wq pool's node attribute is calculated at its allocation.
>> But it's now calculated based on possible cpu<->node information
>> which can be wrong after cpu hotplug/unplug.
>>
>> If wrong pool->node is set, following allocation error will happen.
>> ==
>>   SLUB: Unable to allocate memory on node 2 (gfp=0x80d0)
>>    cache: kmalloc-192, object size: 192, buffer size: 192, default order:
>> 1, min order: 0
>>    node 0: slabs: 6172, objs: 259224, free: 245741
>>    node 1: slabs: 3261, objs: 136962, free: 127656
>> ==
>>
>> This patch fixes the node detection by making use of online cpu info.
>> Unlike cpumask, the best node can be calculated by degree of overlap
>> between attr->cpumask and numanode->online_cpumask.
>> This change doesn't corrupt original purpose of the old calculation.
>>
>> Note: it's expected that this function is called as
>>        pool_detect_best_node
>>        get_unbound_pool
>>        alloc_unbound_pwq
>>        wq_update_unbound_numa
>>        called at CPU_ONLINE/CPU_DOWN_PREPARE
>> and the latest online cpu info can be applied to a new wq pool,
>> which replaces old one.
>>
>> Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
>> ---
>>   kernel/workqueue.c | 38 ++++++++++++++++++++++++++------------
>>   1 file changed, 26 insertions(+), 12 deletions(-)
>>
>> diff --git a/kernel/workqueue.c b/kernel/workqueue.c
>> index 09b685d..7809154 100644
>> --- a/kernel/workqueue.c
>> +++ b/kernel/workqueue.c
>> @@ -3440,6 +3440,31 @@ static void put_unbound_pool(struct worker_pool *pool)
>>   }
>>
>>   /**
>> + * pool_detect_best_node - detect a node which contains specified cpumask.
>> + * Should be called with wq_pool_mutex held.
>> + * Returns a online node where the most of given cpus are tied to.
>> + */
>> +static int pool_detect_best_node(const struct cpumask *cpumask)
>> +{
>> +	int node, best, match, selected = NUMA_NO_NODE;
>> +	static struct cpumask andmask; /* under wq_pool_mutex */
>> +
>> +	if (!wq_numa_enabled ||
>> +		cpumask_subset(cpu_online_mask, cpumask))
>> +		goto out;
>> +	best = 0;
>> +	/* select a node which contains the most number of cpu */
>> +	for_each_node_state(node, N_ONLINE) {
>> +		cpumask_and(&andmask, cpumask, cpumask_of_node(node));
>> +		match = cpumask_weight(&andmask);
>> +		if (match > best)
>> +			selected = best;
>> +	}
>> +out:
>> +	return selected;
>> +}
>
>
> This is a mixture of fix and development.  Why not just keep the original calculation?
>
Just because wq_numa_possible_mask is broken, it shouldn't be used if possible.
In this patch series, the mask is updated only when a node is coming up.
Is it better to clear it at node offline ?

> if the mask cover multiple nodes, NUMA_NO_NODE is the best for pool->node
> after the pool was created. The memory allocation will select the best node
> for manage_workers(), from which CPU that the worker actually is running on.
>

I'll drop this and try to keep original code as much as possible.

Thanks,
-Kame