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: <c7839d65-1f07-02bf-19b4-aea3a6640227@redhat.com>
Date:   Tue, 18 Oct 2022 11:35:15 +0200
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101
 Thunderbird/102.3.1
Subject: Re: [PATCH] mm: hugetlb: support get/set_policy for hugetlb_vm_ops
Content-Language: en-US
To:     =?UTF-8?B?6buE5p2w?= <huangjie.albert@bytedance.com>
Cc:     Muchun Song <songmuchun@bytedance.com>,
        Mike Kravetz <mike.kravetz@oracle.com>,
        Andrew Morton <akpm@linux-foundation.org>, linux-mm@kvack.org,
        linux-kernel <linux-kernel@vger.kernel.org>
References: <20221012081526.73067-1-huangjie.albert@bytedance.com>
 <2aaf2c3a-6e49-abb9-b9c8-19ce87404982@redhat.com>
 <CABKxMyOnxiS6RsJtQEQPrnzNn-cO0Z+OGJRCUpw1M4=WANx0Dw@mail.gmail.com>
 <2f41fc4c-68eb-ab7d-970b-fcb10f474fd4@redhat.com>
 <CABKxMyO4RvEx===iYr_m2dfB0mtUTqt8A3WEuJn4Dq7r_NNM3w@mail.gmail.com>
 <03b90a2f-2854-6e19-6ccd-41f9933d8813@redhat.com>
 <CABKxMyO21rF+f2vpS+t++DAFHiy_MeWDBjB-AvupysKnDHRJfA@mail.gmail.com>
From:   David Hildenbrand <david@redhat.com>
Organization: Red Hat
In-Reply-To: <CABKxMyO21rF+f2vpS+t++DAFHiy_MeWDBjB-AvupysKnDHRJfA@mail.gmail.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 8bit
Precedence: bulk

On 18.10.22 11:27, 黄杰 wrote:
> David Hildenbrand <david@redhat.com> 于2022年10月17日周一 20:00写道：
>>
>> On 17.10.22 13:46, 黄杰 wrote:
>>> David Hildenbrand <david@redhat.com> 于2022年10月17日周一 19:33写道：
>>>>
>>>> On 17.10.22 11:48, 黄杰 wrote:
>>>>> David Hildenbrand <david@redhat.com> 于2022年10月17日周一 16:44写道：
>>>>>>
>>>>>> On 12.10.22 10:15, Albert Huang wrote:
>>>>>>> From: "huangjie.albert" <huangjie.albert@bytedance.com>
>>>>>>>
>>>>>>> implement these two functions so that we can set the mempolicy to
>>>>>>> the inode of the hugetlb file. This ensures that the mempolicy of
>>>>>>> all processes sharing this huge page file is consistent.
>>>>>>>
>>>>>>> In some scenarios where huge pages are shared:
>>>>>>> if we need to limit the memory usage of vm within node0, so I set qemu's
>>>>>>> mempilciy bind to node0, but if there is a process (such as virtiofsd)
>>>>>>> shared memory with the vm, in this case. If the page fault is triggered
>>>>>>> by virtiofsd, the allocated memory may go to node1 which  depends on
>>>>>>> virtiofsd.
>>>>>>>
>>>>>>
>>>>>> Any VM that uses hugetlb should be preallocating memory. For example,
>>>>>> this is the expected default under QEMU when using huge pages.
>>>>>>
>>>>>> Once preallocation does the right thing regarding NUMA policy, there is
>>>>>> no need to worry about it in other sub-processes.
>>>>>>
>>>>>
>>>>> Hi, David
>>>>> thanks for your reminder
>>>>>
>>>>> Yes, you are absolutely right, However, the pre-allocation mechanism
>>>>> does solve this problem.
>>>>> However, some scenarios do not like to use the pre-allocation mechanism, such as
>>>>> scenarios that are sensitive to virtual machine startup time, or
>>>>> scenarios that require
>>>>> high memory utilization. The on-demand allocation mechanism may be better,
>>>>> so the key point is to find a way support for shared policy。
>>>>
>>>> Using hugetlb -- with a fixed pool size -- without preallocation is like
>>>> playing with fire. Hugetlb reservation makes one believe that on-demand
>>>> allocation is going to work, but there are various scenarios where that
>>>> can go seriously wrong, and you can run out of huge pages.
>>>>
>>>> If you're using hugetlb as memory backend for a VM without
>>>> preallocation, you really have to be very careful. I can only advise
>>>> against doing that.
>>>>
>>>>
>>>> Also: why does another process read/write *first* to a guest physical
>>>> memory location before the OS running inside the VM even initialized
>>>> that memory? That sounds very wrong. What am I missing?
>>>>
>>>
>>> for example : virtio ring buffer.
>>> For the avial descriptor, the guest kernel only gives an address to
>>> the backend,
>>> and does not actually access the memory.
>>
>> Okay, thanks. So we're essentially providing uninitialized memory to a
>> device? Hm, that implies that the device might have access to memory
>> that was previously used by someone else ... not sure how to feel about
>> that, but maybe this is just the way of doing things.
>>
>> The "easy" user-space fix would be to simply similarly mbind() in  the
>> other processes where we mmap(). Has that option been explored?
> 
> This can also solve the problem temporarily, but we need to change all
> processes that share memory with it, so it can't be done once and for
> all

How many process that really care are we walking about? Usually, 
extending the vhost-user protocol and relevant libraries should be good 
enough.

-- 
Thanks,

David / dhildenb