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

[Mina Almasry <almasrymina@google.com>]

On Tue, Nov 14, 2023 at 4:49 AM Yunsheng Lin <linyunsheng@huawei.com> wrote:
>
> On 2023/11/14 20:21, Mina Almasry wrote:
> > On Tue, Nov 14, 2023 at 12:23 AM Yunsheng Lin <linyunsheng@huawei.com> wrote:
> >>
> >> +cc Christian, Jason and Willy
> >>
> >> On 2023/11/14 7:05, Jakub Kicinski wrote:
> >>> On Mon, 13 Nov 2023 05:42:16 -0800 Mina Almasry wrote:
> >>>> You're doing exactly what I think you're doing, and what was nacked in RFC v1.
> >>>>
> >>>> You've converted 'struct page_pool_iov' to essentially become a
> >>>> duplicate of 'struct page'. Then, you're casting page_pool_iov* into
> >>>> struct page* in mp_dmabuf_devmem_alloc_pages(), then, you're calling
> >>>> mm APIs like page_ref_*() on the page_pool_iov* because you've fooled
> >>>> the mm stack into thinking dma-buf memory is a struct page.
> >>
> >> Yes, something like above, but I am not sure about the 'fooled the mm
> >> stack into thinking dma-buf memory is a struct page' part, because:
> >> 1. We never let the 'struct page' for devmem leaking out of net stacking
> >>    through the 'not kmap()able and not readable' checking in your patchset.
> >
> > RFC never used dma-buf pages outside the net stack, so that is the same.
> >
> > You are not able to get rid of the 'net kmap()able and not readable'
> > checking with this approach, because dma-buf memory is fundamentally
> > unkmapable and unreadable. This approach would still need
> > skb_frags_not_readable checks in net stack, so that is also the same.
>
> Yes, I am agreed that checking is still needed whatever the proposal is.
>
> >
> >> 2. We inititiate page->_refcount for devmem to one and it remains as one,
> >>    we will never call page_ref_inc()/page_ref_dec()/get_page()/put_page(),
> >>    instead, we use page pool's pp_frag_count to do reference counting for
> >>    devmem page in patch 6.
> >>
> >
> > I'm not sure that moves the needle in terms of allowing dma-buf
> > memory to look like struct pages.
> >
> >>>>
> >>>> RFC v1 was almost exactly the same, except instead of creating a
> >>>> duplicate definition of struct page, it just allocated 'struct page'
> >>>> instead of allocating another struct that is identical to struct page
> >>>> and casting it into struct page.
> >>
> >> Perhaps it is more accurate to say this is something between RFC v1 and
> >> RFC v3, in order to decouple 'struct page' for devmem from mm subsystem,
> >> but still have most unified handling for both normal memory and devmem
> >> in page pool and net stack.
> >>
> >> The main difference between this patchset and RFC v1:
> >> 1. The mm subsystem is not supposed to see the 'struct page' for devmem
> >>    in this patchset, I guess we could say it is decoupled from the mm
> >>    subsystem even though we still call PageTail()/page_ref_count()/
> >>    page_is_pfmemalloc() on 'struct page' for devmem.
> >>
> >
> > In this patchset you pretty much allocate a struct page for your
> > dma-buf memory, and then cast it into a struct page, so all the mm
> > calls in page_pool.c are seeing a struct page when it's really dma-buf
> > memory.
> >
> > 'even though we still call
> > PageTail()/page_ref_count()/page_is_pfmemalloc() on 'struct page' for
> > devmem' is basically making dma-buf memory look like struct pages.
> >
> > 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.
>
> Yes, as above, skb_frags_not_readable() checking is still needed for
> kmap() and page_address().
>
> get_page, put_page related calling is avoided in page_pool_frag_ref()
> and napi_pp_put_page() for devmem page as the above checking is true
> for devmem page:
> (pp_iov->pp_magic & ~0x3UL) == PP_SIGNATURE
>

So, devmem needs special handling with if statement for refcounting,
even after using struct pages for devmem, which is not allowed (IIUC
the dma-buf maintainer).

> >
> >> The main difference between this patchset and RFC v3:
> >> 1. It reuses the 'struct page' to have more unified handling between
> >>    normal page and devmem page for net stack.
> >
> > This is what was nacked in RFC v1.
> >
> >> 2. It relies on the page->pp_frag_count to do reference counting.
> >>
> >
> > I don't see you change any of the page_ref_* calls in page_pool.c, for
> > example this one:
> >
> > https://elixir.bootlin.com/linux/latest/source/net/core/page_pool.c#L601
> >
> > So the reference the page_pool is seeing is actually page->_refcount,
> > not page->pp_frag_count? I'm confused here. Is this a bug in the
> > patchset?
>
> page->_refcount is the same as page_pool_iov->_refcount for devmem, which
> is ensured by the 'PAGE_POOL_MATCH(_refcount, _refcount);', and
> page_pool_iov->_refcount is set to one in mp_dmabuf_devmem_alloc_pages()
> by calling 'refcount_set(&ppiov->_refcount, 1)' and always remains as one.
>
> So the 'page_ref_count(page) == 1' checking is always true for devmem page.

Which, of course, is a bug in the patchset, and it only works because
it's a POC for you. devmem pages (which shouldn't exist according to
the dma-buf maintainer, IIUC) can't be recycled all the time. See
SO_DEVMEM_DONTNEED patch in my RFC and refcounting needed for devmem.

-- 
Thanks,
Mina