Question on io-wq

Question on io-wq

[Zhang, Qiang]

Hi Jens Axboe

There are some problem in 'io_wqe_worker' thread, when the 
'io_wqe_worker' be create and  Setting the affinity of CPUs in NUMA 
nodes, due to CPU hotplug, When the last CPU going down, the 
'io_wqe_worker' thread will run anywhere. when the CPU in the node goes 
online again, we should restore their cpu bindings?

Thanks
Qiang

Question on io-wq

[Zhang,Qiang]

Hi Jens Axboe

There are some problem in 'io_wqe_worker' thread, when the 
'io_wqe_worker' be create and  Setting the affinity of CPUs in NUMA 
nodes, due to CPU hotplug, When the last CPU going down, the 
'io_wqe_worker' thread will run anywhere. when the CPU in the node goes 
online again, we should restore their cpu bindings?

Thanks
Qiang

Re: Question on io-wq

[Jens Axboe]

On 10/22/20 3:02 AM, Zhang,Qiang wrote:
> 
> Hi Jens Axboe
> 
> There are some problem in 'io_wqe_worker' thread, when the 
> 'io_wqe_worker' be create and  Setting the affinity of CPUs in NUMA 
> nodes, due to CPU hotplug, When the last CPU going down, the 
> 'io_wqe_worker' thread will run anywhere. when the CPU in the node goes 
> online again, we should restore their cpu bindings?

Something like the below should help in ensuring affinities are
always correct - trigger an affinity set for an online CPU event. We
should not need to do it for offlining. Can you test it?


diff --git a/fs/io-wq.c b/fs/io-wq.c
index 4012ff541b7b..3bf029d1170e 100644
--- a/fs/io-wq.c
+++ b/fs/io-wq.c
@@ -19,6 +19,7 @@
 #include <linux/task_work.h>
 #include <linux/blk-cgroup.h>
 #include <linux/audit.h>
+#include <linux/cpu.h>
 
 #include "io-wq.h"
 
@@ -123,9 +124,13 @@ struct io_wq {
 	refcount_t refs;
 	struct completion done;
 
+	struct hlist_node cpuhp_node;
+
 	refcount_t use_refs;
 };
 
+static enum cpuhp_state io_wq_online;
+
 static bool io_worker_get(struct io_worker *worker)
 {
 	return refcount_inc_not_zero(&worker->ref);
@@ -1096,6 +1101,13 @@ struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
 		return ERR_PTR(-ENOMEM);
 	}
 
+	ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
+	if (ret) {
+		kfree(wq->wqes);
+		kfree(wq);
+		return ERR_PTR(ret);
+	}
+
 	wq->free_work = data->free_work;
 	wq->do_work = data->do_work;
 
@@ -1145,6 +1157,7 @@ struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
 	ret = PTR_ERR(wq->manager);
 	complete(&wq->done);
 err:
+	cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
 	for_each_node(node)
 		kfree(wq->wqes[node]);
 	kfree(wq->wqes);
@@ -1164,6 +1177,8 @@ static void __io_wq_destroy(struct io_wq *wq)
 {
 	int node;
 
+	cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
+
 	set_bit(IO_WQ_BIT_EXIT, &wq->state);
 	if (wq->manager)
 		kthread_stop(wq->manager);
@@ -1191,3 +1206,40 @@ struct task_struct *io_wq_get_task(struct io_wq *wq)
 {
 	return wq->manager;
 }
+
+static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
+{
+	struct task_struct *task = worker->task;
+	unsigned long flags;
+
+	raw_spin_lock_irqsave(&task->pi_lock, flags);
+	do_set_cpus_allowed(task, cpumask_of_node(worker->wqe->node));
+	task->flags |= PF_NO_SETAFFINITY;
+	raw_spin_unlock_irqrestore(&task->pi_lock, flags);
+	return false;
+}
+
+static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
+{
+	struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
+	int i;
+
+	rcu_read_lock();
+	for_each_node(i)
+		io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, NULL);
+	rcu_read_unlock();
+	return 0;
+}
+
+static __init int io_wq_init(void)
+{
+	int ret;
+
+	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
+					io_wq_cpu_online, NULL);
+	if (ret < 0)
+		return ret;
+	io_wq_online = ret;
+	return 0;
+}
+subsys_initcall(io_wq_init);

-- 
Jens Axboe

Re: Question on io-wq

[Jens Axboe]

On 10/22/20 8:05 PM, Hillf Danton wrote:
> On Thu, 22 Oct 2020 08:08:09 -0600 Jens Axboe wrote:
>> On 10/22/20 3:02 AM, Zhang,Qiang wrote:
>>>
>>> Hi Jens Axboe
>>>
>>> There are some problem in 'io_wqe_worker' thread, when the 
>>> 'io_wqe_worker' be create and  Setting the affinity of CPUs in NUMA 
>>> nodes, due to CPU hotplug, When the last CPU going down, the 
>>> 'io_wqe_worker' thread will run anywhere. when the CPU in the node goes 
>>> online again, we should restore their cpu bindings?
>>
>> Something like the below should help in ensuring affinities are
>> always correct - trigger an affinity set for an online CPU event. We
>> should not need to do it for offlining. Can you test it?
> 
> CPU affinity is intact because of nothing to do on offline, and scheduler
> will move the stray workers on to the correct NUMA node if any CPU goes
> online, so it's a bit hard to see what is going to be tested.

Test it yourself:

- Boot with > 1 NUMA node
- Start an io_uring, you now get 2 workers, each affinitized to a node
- Now offline all CPUs in one node
- Online one or more of the CPU in that same node

The end result is that the worker on the node that was offlined now
has a mask of the other node, plus the newly added CPU.

So your last statement isn't correct, which is what the original
reporter stated.

-- 
Jens Axboe