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 1/2] KVM: arm/arm64: Skip updating PMD entry if no
 change
To:     Punit Agrawal <punit.agrawal@arm.com>, kvmarm@lists.cs.columbia.edu
Cc:     linux-arm-kernel@lists.infradead.org, linux-kernel@vger.kernel.org,
        suzuki.poulose@arm.com, christoffer.dall@arm.com,
        stable@vger.kernel.org
References: <20180813094049.3726-1-punit.agrawal@arm.com>
 <20180813094049.3726-2-punit.agrawal@arm.com>
From:   Marc Zyngier <marc.zyngier@arm.com>
Organization: ARM Ltd
Message-ID: <e265fe94-aa14-00c5-b144-4eff77413f5a@arm.com>
Date:   Mon, 13 Aug 2018 11:11:30 +0100
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:52.0) Gecko/20100101
 Thunderbird/52.9.1
MIME-Version: 1.0
In-Reply-To: <20180813094049.3726-2-punit.agrawal@arm.com>
Content-Type: text/plain; charset=utf-8
Content-Language: en-GB
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

Hi Punit,

On 13/08/18 10:40, Punit Agrawal wrote:
> Contention on updating a PMD entry by a large number of vcpus can lead
> to duplicate work when handling stage 2 page faults. As the page table
> update follows the break-before-make requirement of the architecture,
> it can lead to repeated refaults due to clearing the entry and
> flushing the tlbs.
> 
> This problem is more likely when -
> 
> * there are large number of vcpus
> * the mapping is large block mapping
> 
> such as when using PMD hugepages (512MB) with 64k pages.
> 
> Fix this by skipping the page table update if there is no change in
> the entry being updated.
> 
> Fixes: ad361f093c1e ("KVM: ARM: Support hugetlbfs backed huge pages")
> Change-Id: Ib417957c842ef67a6f4b786f68df62048d202c24
> Signed-off-by: Punit Agrawal <punit.agrawal@arm.com>
> Cc: Marc Zyngier <marc.zyngier@arm.com>
> Cc: Christoffer Dall <christoffer.dall@arm.com>
> Cc: Suzuki Poulose <suzuki.poulose@arm.com>
> Cc: stable@vger.kernel.org
> ---
>  virt/kvm/arm/mmu.c | 40 +++++++++++++++++++++++++++++-----------
>  1 file changed, 29 insertions(+), 11 deletions(-)
> 
> diff --git a/virt/kvm/arm/mmu.c b/virt/kvm/arm/mmu.c
> index 1d90d79706bd..2ab977edc63c 100644
> --- a/virt/kvm/arm/mmu.c
> +++ b/virt/kvm/arm/mmu.c
> @@ -1015,19 +1015,36 @@ static int stage2_set_pmd_huge(struct kvm *kvm, struct kvm_mmu_memory_cache
>  	pmd = stage2_get_pmd(kvm, cache, addr);
>  	VM_BUG_ON(!pmd);
>  
> -	/*
> -	 * Mapping in huge pages should only happen through a fault.  If a
> -	 * page is merged into a transparent huge page, the individual
> -	 * subpages of that huge page should be unmapped through MMU
> -	 * notifiers before we get here.
> -	 *
> -	 * Merging of CompoundPages is not supported; they should become
> -	 * splitting first, unmapped, merged, and mapped back in on-demand.
> -	 */
> -	VM_BUG_ON(pmd_present(*pmd) && pmd_pfn(*pmd) != pmd_pfn(*new_pmd));
> -
>  	old_pmd = *pmd;
> +
>  	if (pmd_present(old_pmd)) {
> +		/*
> +		 * Mapping in huge pages should only happen through a
> +		 * fault.  If a page is merged into a transparent huge
> +		 * page, the individual subpages of that huge page
> +		 * should be unmapped through MMU notifiers before we
> +		 * get here.
> +		 *
> +		 * Merging of CompoundPages is not supported; they
> +		 * should become splitting first, unmapped, merged,
> +		 * and mapped back in on-demand.
> +		 */
> +		VM_BUG_ON(pmd_pfn(old_pmd) != pmd_pfn(*new_pmd));
> +
> +		/*
> +		 * Multiple vcpus faulting on the same PMD entry, can
> +		 * lead to them sequentially updating the PMD with the
> +		 * same value. Following the break-before-make
> +		 * (pmd_clear() followed by tlb_flush()) process can
> +		 * hinder forward progress due to refaults generated
> +		 * on missing translations.
> +		 *
> +		 * Skip updating the page table if the entry is
> +		 * unchanged.
> +		 */
> +		if (pmd_val(old_pmd) == pmd_val(*new_pmd))
> +			goto out;

I think the order of these two checks should be reversed: the first one
is clearly a subset of the second one, so it'd make sense to have the
global comparison before having the more specific one. Not that it
matter much in practice, but I just find it easier to reason about.

> +
>  		pmd_clear(pmd);
>  		kvm_tlb_flush_vmid_ipa(kvm, addr);
>  	} else {
> @@ -1035,6 +1052,7 @@ static int stage2_set_pmd_huge(struct kvm *kvm, struct kvm_mmu_memory_cache
>  	}
>  
>  	kvm_set_pmd(pmd, *new_pmd);
> +out:
>  	return 0;
>  }
>  
> 

Thanks,

	M.
-- 
Jazz is not dead. It just smells funny...