Message-ID: <CAHS8izMR-FrTtCty8v29atAMor5FmzV_Ogk85H=gqGaJNvJnuA@mail.gmail.com>
Date: Wed, 15 Nov 2023 09:44:37 -0800
From: Mina Almasry <almasrymina@...gle.com>
To: Yunsheng Lin <linyunsheng@...wei.com>
Cc: Jason Gunthorpe <jgg@...dia.com>, Jakub Kicinski <kuba@...nel.org>, davem@...emloft.net, 
	pabeni@...hat.com, netdev@...r.kernel.org, linux-kernel@...r.kernel.org, 
	Willem de Bruijn <willemb@...gle.com>, Kaiyuan Zhang <kaiyuanz@...gle.com>, 
	Jesper Dangaard Brouer <hawk@...nel.org>, Ilias Apalodimas <ilias.apalodimas@...aro.org>, 
	Eric Dumazet <edumazet@...gle.com>, Christian König <christian.koenig@....com>, 
	Matthew Wilcox <willy@...radead.org>, Linux-MM <linux-mm@...ck.org>
Subject: Re: [PATCH RFC 3/8] memory-provider: dmabuf devmem memory provider

On Wed, Nov 15, 2023 at 1:21 AM Yunsheng Lin <linyunsheng@...wei.com> 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.
>
> struct page for devmem will not be called into the mm subsystem, so most
> of the mm calls is avoided by calling into the devmem memory provider'
> ops instead of calling mm calls.
>
> As far as I see for now, only page_ref_count(), page_is_pfmemalloc() and
> PageTail() is called for devmem page, which should be easy to ensure that
> those call for devmem page is consistent with the struct page owned by mm.

I'm not sure this is true. These 3 calls are just the calls you're
aware of. In your proposal you're casting mirror pages into page* and
releasing them into the net stack. You need to scrub the entire net
stack for mm calls, i.e. all driver code and all skb_frag_page() call
sites. Of the top of my head, the driver is probably calling
page_address() and illegal_highdma() is calling PageHighMem(). TCP
zerocopy receive is calling vm_insert_pages().

> I am not sure if we can use some kind of compile/runtime checking to ensure
> those kinds of consistency?
>
> >
> >>> I would expect net stack, page pool, driver still see the 'struct page',
> >>> only memory provider see the specific struct for itself, for the above,
> >>> devmem memory provider sees the 'struct page_pool_iov'.
> >>>
> >>> The reason I still expect driver to see the 'struct page' is that driver
> >>> will still need to support normal memory besides devmem.
> >
> > I wouldn't say this approach is unreasonable, but it does have to be
> > done carefully to isolate the mm. Keeping the struct page in the API
> > is going to make this very hard.
>
> I would expect that most of the isolation is done in page pool, as far as
> I can see:
>
> 1. For control part: the driver may need to tell the page pool which memory
>                      provider it want to use. Or the administrator specifies
>                      which memory provider to use by some netlink-based cmd.
>
> 2. For data part: I am thinking that driver should only call page_pool_alloc(),
>                   page_pool_free() and page_pool_get_dma_addr related function.
>
> Of course the driver may need to be aware of that if it can call kmap() or
> page_address() on the page returned from page_pool_alloc(), and maybe tell
> net stack that those pages is not kmap()'able and page_address()'able.
>
> >
> > Jason
> > .
> >



-- 
Thanks,
Mina

Powered by blists - more mailing lists