Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Subject: Re: [PATCH v2] KVM: LAPIC: Recalculate apic map in batch
To:     Wanpeng Li <kernellwp@gmail.com>, linux-kernel@vger.kernel.org,
        kvm@vger.kernel.org
Cc:     Sean Christopherson <sean.j.christopherson@intel.com>,
        Vitaly Kuznetsov <vkuznets@redhat.com>,
        Wanpeng Li <wanpengli@tencent.com>,
        Jim Mattson <jmattson@google.com>,
        Joerg Roedel <joro@8bytes.org>
References: <1582624061-5814-1-git-send-email-wanpengli@tencent.com>
From:   Paolo Bonzini <pbonzini@redhat.com>
Message-ID: <0af6b96a-16ac-5054-7754-6ab4a239a2d4@redhat.com>
Date:   Tue, 25 Feb 2020 15:20:43 +0100
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:68.0) Gecko/20100101
 Thunderbird/68.4.1
MIME-Version: 1.0
In-Reply-To: <1582624061-5814-1-git-send-email-wanpengli@tencent.com>
Content-Type: text/plain; charset=utf-8
Content-Language: en-US
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On 25/02/20 10:47, Wanpeng Li wrote:
> From: Wanpeng Li <wanpengli@tencent.com>
> 
> In the vCPU reset and set APIC_BASE MSR path, the apic map will be recalculated 
> several times, each time it will consume 10+ us observed by ftrace in my 
> non-overcommit environment since the expensive memory allocate/mutex/rcu etc 
> operations. This patch optimizes it by recaluating apic map in batch, I hope 
> this can benefit the serverless scenario which can frequently create/destroy 
> VMs.
> 
> Signed-off-by: Wanpeng Li <wanpengli@tencent.com>
> ---
> v1 -> v2:
>  * add apic_map_dirty to kvm_lapic
>  * error condition in kvm_apic_set_state, do recalcuate  unconditionally
> 
>  arch/x86/kvm/lapic.c | 29 +++++++++++++++++++----------
>  arch/x86/kvm/lapic.h |  2 ++
>  arch/x86/kvm/x86.c   |  2 ++
>  3 files changed, 23 insertions(+), 10 deletions(-)
> 
> diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
> index afcd30d..3476dbc 100644
> --- a/arch/x86/kvm/lapic.c
> +++ b/arch/x86/kvm/lapic.c
> @@ -164,7 +164,7 @@ static void kvm_apic_map_free(struct rcu_head *rcu)
>  	kvfree(map);
>  }
>  
> -static void recalculate_apic_map(struct kvm *kvm)
> +void kvm_recalculate_apic_map(struct kvm *kvm)
>  {

It's better to add an "if" here rather than in every caller.  It should
be like:

	if (!apic->apic_map_dirty) {
		/*
		 * Read apic->apic_map_dirty before
		 * kvm->arch.apic_map.
		 */
		smp_rmb();
		return;
	}

        mutex_lock(&kvm->arch.apic_map_lock);
	if (!apic->apic_map_dirty) {
		/* Someone else has updated the map.  */
		mutex_unlock(&kvm->arch.apic_map_lock);
		return;
	}
	...
out:
        old = rcu_dereference_protected(kvm->arch.apic_map,
                        lockdep_is_held(&kvm->arch.apic_map_lock));
        rcu_assign_pointer(kvm->arch.apic_map, new);
	/*
	 * Write kvm->arch.apic_map before
	 * clearing apic->apic_map_dirty.
	 */
	smp_wmb();
	apic->apic_map_dirty = false;
        mutex_unlock(&kvm->arch.apic_map_lock);
	...

But actually it seems to me that, given we're going through all this
pain, it's better to put the "dirty" flag in kvm->arch, next to the
mutex and the map itself.  This should also reduce the number of calls
to kvm_recalculate_apic_map that recompute the map.  A lot of them will
just wait on the mutex and exit.

Paolo