Re: [PATCH net-next RFC 1/2] page_pool: add page recycling support based on elevated refcnt

From: Ilias Apalodimas
Date: Tue Jul 06 2021 - 00:55:04 EST


Hi Yunsheng,

Thanks for having a look!

On Fri, Jul 02, 2021 at 06:15:13PM +0800, Yunsheng Lin wrote:
> On 2021/7/2 17:42, Jesper Dangaard Brouer wrote:
> >
> > On 30/06/2021 11.17, Yunsheng Lin wrote:
> >> Currently page pool only support page recycling only when
> >> refcnt of page is one, which means it can not support the
> >> split page recycling implemented in the most ethernet driver.
> >
> > Cc. Alex Duyck as I consider him an expert in this area.
>
> Thanks.
>
> >
> >
> >> So add elevated refcnt support in page pool, and support
> >> allocating page frag to enable multi-frames-per-page based
> >> on the elevated refcnt support.
> >>
> >> As the elevated refcnt is per page, and there is no space
> >> for that in "struct page" now, so add a dynamically allocated
> >> "struct page_pool_info" to record page pool ptr and refcnt
> >> corrsponding to a page for now. Later, we can recycle the
> >> "struct page_pool_info" too, or use part of page memory to
> >> record pp_info.
> >
> > I'm not happy with allocating a memory (slab) object "struct page_pool_info" per page.
> >
> > This also gives us an extra level of indirection.
>
> I'm not happy with that either, if there is better way to
> avoid that, I will be happy to change it:)

I think what we have to answer here is, do we want and does it make sense
for page_pool to do the housekeeping of the buffer splitting or are we
better of having each driver do that. IIRC your previous patch on top of
the original recycling patchset was just 'atomic' refcnts on top of page pool.

I think I'd prefer each driver having it's own meta-data of how he splits
the page, mostly due to hardware diversity, but tbh I don't have any
strong preference atm.

>
> >
> >
> > You are also adding a page "frag" API inside page pool, which I'm not 100% convinced belongs inside page_pool APIs.
> >
> > Please notice the APIs that Alex Duyck added in mm/page_alloc.c:
>
> Actually, that is where the idea of using "page frag" come from.
>
> Aside from the performance improvement, there is memory usage
> decrease for 64K page size kernel, which means a 64K page can
> be used by 32 description with 2k buffer size, and that is a
> lot of memory saving for 64 page size kernel comparing to the
> current split page reusing implemented in the driver.
>

Whether the driver or page_pool itself keeps the meta-data, the outcome
here won't change. We'll still be able to use page frags.


Cheers
/Ilias
>
> >
> > __page_frag_cache_refill() + __page_frag_cache_drain() + page_frag_alloc_align()
> >
> >
>
> [...]