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;
Subject: Re: [PATCH v3 0/8] Lazily allocate memslot rmaps
To:     Ben Gardon <bgardon@google.com>, linux-kernel@vger.kernel.org,
        kvm@vger.kernel.org
Cc:     Paolo Bonzini <pbonzini@redhat.com>, Peter Xu <peterx@redhat.com>,
        Sean Christopherson <seanjc@google.com>,
        Peter Shier <pshier@google.com>,
        Yulei Zhang <yulei.kernel@gmail.com>,
        Wanpeng Li <kernellwp@gmail.com>,
        Xiao Guangrong <xiaoguangrong.eric@gmail.com>,
        Kai Huang <kai.huang@intel.com>,
        Keqian Zhu <zhukeqian1@huawei.com>
References: <20210506184241.618958-1-bgardon@google.com>
From:   David Hildenbrand <david@redhat.com>
Organization: Red Hat
Message-ID: <05848766-b13c-2a58-81da-0f1e839a6cd0@redhat.com>
Date:   Fri, 7 May 2021 09:40:23 +0200
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:78.0) Gecko/20100101
 Thunderbird/78.8.1
MIME-Version: 1.0
In-Reply-To: <20210506184241.618958-1-bgardon@google.com>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Language: en-US
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 06.05.21 20:42, Ben Gardon wrote:
> This series enables KVM to save memory when using the TDP MMU by waiting
> to allocate memslot rmaps until they are needed. To do this, KVM tracks
> whether or not a shadow root has been allocated. In order to get away
> with not allocating the rmaps, KVM must also be sure to skip operations
> which iterate over the rmaps. If the TDP MMU is in use and we have not
> allocated a shadow root, these operations would essentially be op-ops
> anyway. Skipping the rmap operations has a secondary benefit of avoiding
> acquiring the MMU lock in write mode in many cases, substantially
> reducing MMU lock contention.
> 
> This series was tested on an Intel Skylake machine. With the TDP MMU off
> and on, this introduced no new failures on kvm-unit-tests or KVM selftests.
> 

Happy to see this change pop up, I remember discussing this with Paolo 
recently.

Another step to reduce the rmap overhead could be looking into using a 
dynamic datastructure to manage the rmap, instead of allocating a 
fixed-sized array. That could also significantly reduce memory overhead 
in some setups and give us more flexibility, for example, for resizing 
or splitting slots atomically.

-- 
Thanks,

David / dhildenb