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;
Message-ID: <7dba12ce-f981-6017-0613-542472e3ec5c@bytedance.com>
Date:   Fri, 14 Oct 2022 10:03:58 +0800
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10.15; rv:102.0)
 Gecko/20100101 Thunderbird/102.3.2
Subject: Re: [PATCH] sched/psi: Fix avgs_work re-arm in psi_avgs_work()
Content-Language: en-US
To:     Suren Baghdasaryan <surenb@google.com>,
        Johannes Weiner <hannes@cmpxchg.org>
Cc:     quic_pkondeti@quicinc.com, peterz@infradead.org,
        quic_charante@quicinc.com, linux-kernel@vger.kernel.org
References: <20220913140817.GA9091@hu-pkondeti-hyd.qualcomm.com>
 <20221010104206.12184-1-zhouchengming@bytedance.com>
 <CAJuCfpF7Z+CYhk-f_aaDTE232+m9z_n-QfjGfdLje7QrX9bFtw@mail.gmail.com>
 <a73f58a3-9f96-2ce5-38a0-8abab27a2260@bytedance.com>
 <CAJuCfpET+B3X-uX2vDp-2yH-+OVxOu3YXL7JWZrPuoh22P+5SQ@mail.gmail.com>
 <dea56c22-ab5b-25e2-9819-cc598f9aad80@bytedance.com>
 <CAJuCfpFTDyR1V+JYOY_uN6Xg1Nip5b=9dzkwm-CNd8vMWaQQFQ@mail.gmail.com>
 <46c6e1cc-77d3-eac1-fa18-deed2bac4a0e@bytedance.com>
 <Y0g0UAAJMhPczNm/@cmpxchg.org>
 <CAJuCfpFo5EEEg24VLSmcPo=VDMszB9Q2a4L_Eq6E9VWnbsBJDQ@mail.gmail.com>
From:   Chengming Zhou <zhouchengming@bytedance.com>
In-Reply-To: <CAJuCfpFo5EEEg24VLSmcPo=VDMszB9Q2a4L_Eq6E9VWnbsBJDQ@mail.gmail.com>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 2022/10/14 00:10, Suren Baghdasaryan wrote:
> On Thu, Oct 13, 2022 at 8:52 AM Johannes Weiner <hannes@cmpxchg.org> wrote:
>>
>> On Thu, Oct 13, 2022 at 07:06:55PM +0800, Chengming Zhou wrote:
>>> Should I still need to copy groupc->tasks[] out for the current_cpu as you
>>> suggested before?
>>
>> It'd be my preference as well. This way the resched logic can be
>> consolidated into a single block of comment + code at the end of the
>> function.
> 
> Sounds good to me. If we are copying times in the retry loop then
> let's move the `reschedule =` decision out of that loop completely. At
> the end of get_recent_times we can do:
> 
> if (cpu == current_cpu)
>     reschedule = tasks[NR_RUNNING] +
>                             tasks[NR_IOWAIT] +
>                             tasks[NR_MEMSTALL] > 1;
> else
>     reschedule = *pchanged_states & (1 << PSI_NONIDLE);
> 

Ok, I will send an updated patch later.

Thanks!

> 
>>
>>> @@ -242,6 +242,8 @@ static void get_recent_times(struct psi_group *group, int cpu,
>>>                              u32 *pchanged_states)
>>>  {
>>>         struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu);
>>> +       int current_cpu = raw_smp_processor_id();
>>> +       bool reschedule;
>>>         u64 now, state_start;
>>>         enum psi_states s;
>>>         unsigned int seq;
>>> @@ -256,6 +258,10 @@ static void get_recent_times(struct psi_group *group, int cpu,
>>>                 memcpy(times, groupc->times, sizeof(groupc->times));
>>>                 state_mask = groupc->state_mask;
>>>                 state_start = groupc->state_start;
>>> +               if (cpu == current_cpu)
>>> +                       reschedule = groupc->tasks[NR_RUNNING] +
>>> +                               groupc->tasks[NR_IOWAIT] +
>>> +                               groupc->tasks[NR_MEMSTALL] > 1;
>>>         } while (read_seqcount_retry(&groupc->seq, seq));
>>
>> This also matches psi_show() and the poll worker. They don't currently
>> use the flag, but it's somewhat fragile and confusing. Add a test for
>> current_work() == &group->avgs_work?
> 
> Good point. (tasks[NR_RUNNING] + tasks[NR_IOWAIT] + tasks[NR_MEMSTALL]
>> 1) condition should also contain this check.