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: <4ccbafb5-9157-ec73-c751-ec71164f8688@redhat.com>
Date:   Tue, 2 Aug 2022 19:42:49 +0200
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101
 Thunderbird/91.11.0
Content-Language: en-US
To:     Sean Christopherson <seanjc@google.com>,
        David Matlack <dmatlack@google.com>
Cc:     Vipin Sharma <vipinsh@google.com>, kvm@vger.kernel.org,
        linux-kernel@vger.kernel.org
References: <20220801151928.270380-1-vipinsh@google.com>
 <YuhPT2drgqL+osLl@google.com> <YuhoJUoPBOu5eMz8@google.com>
 <YulRZ+uXFOE1y2dj@google.com> <YuldSf4T2j4rIrIo@google.com>
From:   Paolo Bonzini <pbonzini@redhat.com>
Subject: Re: [PATCH] KVM: x86/mmu: Make page tables for eager page splitting
 NUMA aware
In-Reply-To: <YuldSf4T2j4rIrIo@google.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 8/2/22 19:22, Sean Christopherson wrote:
> Userspace can already force the ideal setup for eager page splitting by configuring
> vNUMA-aware memslots and using a task with appropriate policy to toggle dirty
> logging.  And userspace really should be encouraged to do that, because otherwise
> walking the page tables in software to do the split is going to be constantly
> accessing remote memory.

Yes, it's possible to locate the page tables on the node that holds the 
memory they're mapping by enable dirty logging from different tasks for 
different memslots, but that seems a bit weird.

Walking the page tables in software is going to do several remote memory 
accesses, but it will do that in a thread that probably is devoted to 
background tasks anyway.  The relative impact of remote memory accesses 
in the thread that enables dirty logging vs. in the vCPU thread should 
also be visible in the overall performance of dirty_log_perf_test.

So I agree with Vipin's patch and would even extend it to all page table 
allocations, however dirty_log_perf_test should be run with fixed CPU 
mappings to measure accurately the impact of the remote memory accesses.

Paolo