Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Subject: Re: [RFC PATCH] blk-mq: avoid housekeeping CPUs scheduling a worker
 on a non-housekeeping CPU
To:     Ming Lei <ming.lei@redhat.com>
CC:     <axboe@kernel.dk>, <hch@lst.de>, <linux-block@vger.kernel.org>,
        <linux-kernel@vger.kernel.org>, <yuyufen@huawei.com>,
        <guohanjun@huawei.com>
References: <20220210093532.182818-1-wangxiongfeng2@huawei.com>
 <881ae7a8-5dff-ff50-9bc2-a983b6a53c30@huawei.com> <Ygst7R+X7u2OBgUW@T590>
From:   Xiongfeng Wang <wangxiongfeng2@huawei.com>
Message-ID: <ccaf5c9b-bede-a3d9-fbc2-b26ab1c94143@huawei.com>
Date:   Tue, 15 Feb 2022 17:32:54 +0800
User-Agent: Mozilla/5.0 (Windows NT 10.0; WOW64; rv:68.0) Gecko/20100101
 Thunderbird/68.4.1
MIME-Version: 1.0
In-Reply-To: <Ygst7R+X7u2OBgUW@T590>
Content-Type: text/plain; charset="utf-8"
Content-Transfer-Encoding: 7bit
Precedence: bulk

Hi Ming,

Thanks for your reply !

On 2022/2/15 12:37, Ming Lei wrote:
> Hello Xiongfeng,
> 
> On Tue, Feb 15, 2022 at 10:29:51AM +0800, Xiongfeng Wang wrote:
>> Hi Ming,
>>
>> Sorry to disturb you. It's just that I think you may be interested at this
>> patch. I found the following commit written by you.
>>   commit 11ea68f553e244851d15793a7fa33a97c46d8271
>>   genirq, sched/isolation: Isolate from handling managed interrupts
>> It removed the managed_irq interruption from non-housekeeping CPUs as long as
>> the non-housekeeping CPUs do not request IO. But the the work thread
>> blk_mq_run_work_fn() may still run on the non-housekeeping CPUs.
>> Appreciate it a lot if you can give it a look.
> 
> Yeah, commit 11ea68f553e24 touches irq subsystem to try not assign
> isolated cpus for managed irq's effective affinity.
> 
> Here blk-mq just selects one cpu and calls mod_delayed_work_on()
> to execute the run queue handler on specified cpu. There are lots of
> such bound wq usage in tree, so I guess it might belong to one wq or
> scheduler generic problem instead of blk-mq specific issue. Not sure
> if it is good to address it in block layer.

Yes, I also find some other worker thread running on the non-housekeeping CPUs.
Some of them need to read the per-cpu data, such as drain_local_pages_wq(). But
workqueue subsystem doesn't know if the work threads read any per-cpu data and
can be migrated to another CPU.

For the workqueue marked as WQ_UNBOUND, the following commit can move the worker
threads to the housekeeping CPUs.
    commit 1bda3f8087fce9063da0b8aef87f17a3fe541aca
    sched/isolation: Isolate workqueues when "nohz_full=" is set
But for the workqueue without flag WQ_UNBOUND, workqueue subsystem doesn't know
if the worker threads can be migrated to another CPU.

So I think maybe the subsystem who create the workqueue can decide whether the
worker threads can be migrated.

Thanks,
Xiongfeng

> 
> thanks,
> Ming
> 
>>
>> Thanks,
>> Xiongfeng
>>
>> On 2022/2/10 17:35, Xiongfeng Wang wrote:
>>> When NOHZ_FULL is enabled, such as in HPC situation, CPUs are divided
>>> into housekeeping CPUs and non-housekeeping CPUs. Non-housekeeping CPUs
>>> are NOHZ_FULL CPUs and are often monopolized by the userspace process,
>>> such HPC application process. Any sort of interruption is not expected.
>>>
>>> blk_mq_hctx_next_cpu() selects each cpu in 'hctx->cpumask' alternately
>>> to schedule the work thread blk_mq_run_work_fn(). When 'hctx->cpumask'
>>> contains housekeeping CPU and non-housekeeping CPU at the same time, a
>>> housekeeping CPU, which want to request a IO, may schedule a worker on a
>>> non-housekeeping CPU. This may affect the performance of the userspace
>>> application running on non-housekeeping CPUs.
>>>
>>> So let's just schedule the worker thread on the current CPU when the
>>> current CPU is housekeeping CPU.
>>>
>>> Signed-off-by: Xiongfeng Wang <wangxiongfeng2@huawei.com>
>>> ---
>>>  block/blk-mq.c | 15 ++++++++++++++-
>>>  1 file changed, 14 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/block/blk-mq.c b/block/blk-mq.c
>>> index 1adfe4824ef5..ff9a4bf16858 100644
>>> --- a/block/blk-mq.c
>>> +++ b/block/blk-mq.c
>>> @@ -24,6 +24,7 @@
>>>  #include <linux/sched/sysctl.h>
>>>  #include <linux/sched/topology.h>
>>>  #include <linux/sched/signal.h>
>>> +#include <linux/sched/isolation.h>
>>>  #include <linux/delay.h>
>>>  #include <linux/crash_dump.h>
>>>  #include <linux/prefetch.h>
>>> @@ -2036,6 +2037,8 @@ static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
>>>  static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
>>>  					unsigned long msecs)
>>>  {
>>> +	int work_cpu;
>>> +
>>>  	if (unlikely(blk_mq_hctx_stopped(hctx)))
>>>  		return;
>>>  
>>> @@ -2050,7 +2053,17 @@ static void __blk_mq_delay_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async,
>>>  		put_cpu();
>>>  	}
>>>  
>>> -	kblockd_mod_delayed_work_on(blk_mq_hctx_next_cpu(hctx), &hctx->run_work,
>>> +	/*
>>> +	 * Avoid housekeeping CPUs scheduling a worker on a non-housekeeping
>>> +	 * CPU
>>> +	 */
>>> +	if (tick_nohz_full_enabled() && housekeeping_cpu(smp_processor_id(),
>>> +							 HK_FLAG_WQ))
>>> +		work_cpu = smp_processor_id();
>>> +	else
>>> +		work_cpu = blk_mq_hctx_next_cpu(hctx);
>>> +
>>> +	kblockd_mod_delayed_work_on(work_cpu, &hctx->run_work,
>>>  				    msecs_to_jiffies(msecs));
>>>  }
>>>  
>>>
>>
>