Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Message-ID: <98a72081-6a2b-b644-d029-edd03da84135@redhat.com>
Date:   Wed, 20 Oct 2021 13:52:54 +0200
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101
 Thunderbird/91.1.0
Subject: Re: [PATCH] rcuwait: do not enter RCU protection unless a wakeup is
 needed
Content-Language: en-US
To:     Peter Zijlstra <peterz@infradead.org>
Cc:     linux-kernel@vger.kernel.org, kvm@vger.kernel.org,
        Davidlohr Bueso <dave@stgolabs.net>,
        Oleg Nesterov <oleg@redhat.com>,
        Ingo Molnar <mingo@kernel.org>,
        "Paul E . McKenney" <paulmck@linux.vnet.ibm.com>,
        Wanpeng Li <wanpengli@tencent.com>
References: <20211020110638.797389-1-pbonzini@redhat.com>
 <YW/61zpycsD8/z4g@hirez.programming.kicks-ass.net>
From:   Paolo Bonzini <pbonzini@redhat.com>
In-Reply-To: <YW/61zpycsD8/z4g@hirez.programming.kicks-ass.net>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 20/10/21 13:17, Peter Zijlstra wrote:
> On Wed, Oct 20, 2021 at 07:06:38AM -0400, Paolo Bonzini wrote:
>> In some cases, rcuwait_wake_up can be called even if the actual likelihood
>> of a wakeup is very low.  If CONFIG_PREEMPT_RCU is active, the resulting
>> rcu_read_lock/rcu_read_unlock pair can be relatively expensive, and in
>> fact it is unnecessary when there is no w->task to keep alive: the
>> memory barrier before the read is what matters in order to avoid missed
>> wakeups.
>>
>> Therefore, do an early check of w->task right after the barrier, and skip
>> rcu_read_lock/rcu_read_unlock unless there is someone waiting for a wakeup.
>>
>> Running kvm-unit-test/vmexit.flat with APICv disabled, most interrupt
>> injection tests (tscdeadline*, self_ipi*, x2apic_self_ipi*) improve
>> by around 600 cpu cycles.
> 
> *how* ?!?
> 
> AFAICT, rcu_read_lock() for PREEMPT_RCU is:
> 
>    WRITE_ONCE(current->rcu_read_lock_nesting, READ_ONCE(current->rcu_read_lock_nesting) + 1);
>    barrier();
> 
> Paul?

Wanpeng, can you share your full .config?

Paolo