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: <c3577a83-9889-c741-bb74-051a6d9a0f61@suse.com>
Date:   Wed, 13 Apr 2022 22:08:49 +0300
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101
 Thunderbird/91.7.0
Subject: Re: [PATCH v2] btrfs: zstd: use spin_lock in timer callback
Content-Language: en-US
To:     Schspa Shi <schspa@gmail.com>
Cc:     dsterba@suse.cz, clm@fb.com, dsterba@suse.com,
        josef@toxicpanda.com, linux-btrfs@vger.kernel.org,
        linux-kernel@vger.kernel.org, terrelln@fb.com
References: <20220411135136.GG15609@suse.cz>
 <20220411155540.36853-1-schspa@gmail.com>
 <09c2a9ce-3b04-ed94-1d62-0e5a072b9dac@suse.com>
 <CAMA88TpjDczKAGN3f+tcsa98rbM7EA0XgT3bHn8UjDqNJ_DeFQ@mail.gmail.com>
From:   Nikolay Borisov <nborisov@suse.com>
In-Reply-To: <CAMA88TpjDczKAGN3f+tcsa98rbM7EA0XgT3bHn8UjDqNJ_DeFQ@mail.gmail.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 8bit
Precedence: bulk


On 13.04.22 г. 19:03 ч., Schspa Shi wrote:
> Nikolay Borisov <nborisov@suse.com> writes:
> 
>> On 11.04.22 г. 18:55 ч., Schspa Shi wrote:
>>> This is an optimization for fix fee13fe96529 ("btrfs:
>>> correct zstd workspace manager lock to use spin_lock_bh()")
>>> The critical region for wsm.lock is only accessed by the process context and
>>> the softirq context.
>>> Because in the soft interrupt, the critical section will not be preempted by
>>> the
>>> soft interrupt again, there is no need to call spin_lock_bh(&wsm.lock) to turn
>>> off the soft interrupt, spin_lock(&wsm.lock) is enough for this situation.
>>> Changelog:
>>> v1 -> v2:
>>>       - Change the commit message to make it more readable.
>>> [1] https://lore.kernel.org/all/20220408181523.92322-1-schspa@gmail.com/
>>> Signed-off-by: Schspa Shi <schspa@gmail.com>
>>
>> Has there been any measurable impact by this change? While it's correct it does mean that
>>   someone looking at the code would see that in one call site we use plain spinlock and in
>> another a _bh version and this is somewhat inconsistent.
>>
> Yes, it may seem a little confused. but it's allowed to save some
> little peace of CPU times.
> and "static inline void red_adaptative_timer(struct timer_list *t) in
> net/sched/sch_red.c"
> have similar usage.
> 
>> What's more I believe this is a noop since when softirqs are executing preemptible() would
>> be false due to preempt_count() being non-0 and in the bh-disabling code
>> in the spinlock we have:
>>
>>   /* First entry of a task into a BH disabled section? */
>>      1         if (!current->softirq_disable_cnt) {
>>    167                 if (preemptible()) {
>>      1                         local_lock(&softirq_ctrl.lock);
>>      2                         /* Required to meet the RCU bottomhalf requirements. */
>>      3                         rcu_read_lock();
>>      4                 } else {
>>      5                         DEBUG_LOCKS_WARN_ON(this_cpu_read(softirq_ctrl.cnt));
>>      6                 }
>>      7         }
>>
>>
>> In this case we'd hit the else branch.
> 
> We won't hit the else branch. because current->softirq_disable_cnt
> won't be zero in the origin case.
> 
> __do_softirq(void)
>          softirq_handle_begin(void)
>          __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
>                          current->softirq_disable_cnt will be > 0 at this time.

That's only relevant on CONFIG_PREEMPT_RT though, on usual kernels 
softirq_handle_being is empty. Furthermore, in case of the non-preempt 
rt if preemptible() always returns false this means that even in the 
__do_softirq path we'll never increment softirq_disable_cnt. So if 
anything this change is only beneficial (theoretically at that in 
preempt_rt scenarios).

>      ......
>          zstd_reclaim_timer_fn(struct timer_list *timer)
>                          spin_lock_bh(&wsm.lock);
>                          __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);
>                          if (!current->softirq_disable_cnt) {
>                                                  // this if branch won't hit
>                                          }
> 
>          softirq_handle_end();
> 
> In this case, the "__local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET);"
> won't do anything useful it only
> increase softirq disable depth and decrease it in
> "__local_bh_enable_ip(_RET_IP_, SOFTIRQ_LOCK_OFFSET);".
> 
> So it's safe to replace spin_lock_bh with spin_lock in a timer
> callback function.
> 
> 
> For the ksoftirqd, it's all the same.
>