Re: [PATCH RFC 3/8] memory-provider: dmabuf devmem memory provider

[Yunsheng Lin <linyunsheng@huawei.com>]

On 2023/11/16 23:58, David Ahern wrote:
> On 11/16/23 4:12 AM, Yunsheng Lin wrote:
>> On 2023/11/16 1:57, David Ahern wrote:
>>> On 11/15/23 2:21 AM, Yunsheng Lin wrote:
>>>> On 2023/11/14 21:16, Jason Gunthorpe wrote:
>>>>> On Tue, Nov 14, 2023 at 04:21:26AM -0800, Mina Almasry wrote:
>>>>>
>>>>>> Actually because you put the 'strtuct page for devmem' in
>>>>>> skb->bv_frag, the net stack will grab the 'struct page' for devmem
>>>>>> using skb_frag_page() then call things like page_address(), kmap,
>>>>>> get_page, put_page, etc, etc, etc.
>>>>>
>>>>> Yikes, please no. If net has its own struct page look alike it has to
>>>>> stay entirely inside net. A non-mm owned struct page should not be
>>>>> passed into mm calls. It is just way too hacky to be seriously
>>>>> considered :(
>>>>
>>>> Yes, that is something this patchset is trying to do, defining its own
>>>> struct page look alike for page pool to support devmem.
>>>>
>>>
>>> Networking needs to be able to move away from struct page references.
>>> The devmem and host memory for Rx use cases do not need to be page based.
>>
>> Yes, I am agreed the ultimate goal is to move away from struct page
>> references. But I am not sure if we can do it right away as there
>> still are different types of existing 'struct page' in the netstack,
>> see:
>>
>> https://lore.kernel.org/all/8b7d25eb-1f10-3e37-8753-92b42da3fb34@huawei.com/
> 
> yes, that is the point of a blended approach -- pages and buffers (or
> iov) -- leveraging the LSB of the address. That proposal is the right

I am not sure leveraging the LSB of the address is necessary yet, as it
does not seems to provide the type check protection, it seems to just
provide a way to demux between pages(including page pool owned page and
non-page pool owned page) and page pool owned buffer.
That info is avaliable through the page->pp_magic and page->pp->mp_*
too if we mirror the page pool specific union in 'struct page'.

> direction to be moving for co-existence. Adding fake struct page
> instances is the wrong direction.

Perhaps a fake struct page with type check protection is the right
direction?

Intergrating devmem to page pool without a unified metadata between
pages and buffers or without a proper abstract layer does not seems
like the good direction either.

> .
>