Re: [PATCH net-next RFC 0/2] add elevated refcnt support for page pool

From: Ilias Apalodimas
Date: Fri Jul 09 2021 - 02:42:46 EST


On Fri, Jul 09, 2021 at 02:40:02PM +0800, Yunsheng Lin wrote:
> On 2021/7/9 12:15, Matteo Croce wrote:
> > On Wed, Jul 7, 2021 at 6:50 PM Marcin Wojtas <mw@xxxxxxxxxxxx> wrote:
> >>
> >> Hi,
> >>
> >>
> >> ??r., 7 lip 2021 o 01:20 Matteo Croce <mcroce@xxxxxxxxxxxxxxxxxxx> napisa??(a):
> >>>
> >>> On Tue, Jul 6, 2021 at 5:51 PM Russell King (Oracle)
> >>> <linux@xxxxxxxxxxxxxxx> wrote:
> >>>>
> >>>> On Fri, Jul 02, 2021 at 03:39:47PM +0200, Matteo Croce wrote:
> >>>>> On Wed, 30 Jun 2021 17:17:54 +0800
> >>>>> Yunsheng Lin <linyunsheng@xxxxxxxxxx> wrote:
> >>>>>
> >>>>>> This patchset adds elevated refcnt support for page pool
> >>>>>> and enable skb's page frag recycling based on page pool
> >>>>>> in hns3 drvier.
> >>>>>>
> >>>>>> Yunsheng Lin (2):
> >>>>>> page_pool: add page recycling support based on elevated refcnt
> >>>>>> net: hns3: support skb's frag page recycling based on page pool
> >>>>>>
> >>>>>> drivers/net/ethernet/hisilicon/hns3/hns3_enet.c | 79 +++++++-
> >>>>>> drivers/net/ethernet/hisilicon/hns3/hns3_enet.h | 3 +
> >>>>>> drivers/net/ethernet/hisilicon/hns3/hns3_ethtool.c | 1 +
> >>>>>> drivers/net/ethernet/marvell/mvneta.c | 6 +-
> >>>>>> drivers/net/ethernet/marvell/mvpp2/mvpp2_main.c | 2 +-
> >>>>>> include/linux/mm_types.h | 2 +-
> >>>>>> include/linux/skbuff.h | 4 +-
> >>>>>> include/net/page_pool.h | 30 ++-
> >>>>>> net/core/page_pool.c | 215
> >>>>>> +++++++++++++++++---- 9 files changed, 285 insertions(+), 57
> >>>>>> deletions(-)
> >>>>>>
> >>>>>
> >>>>> Interesting!
> >>>>> Unfortunately I'll not have access to my macchiatobin anytime soon, can
> >>>>> someone test the impact, if any, on mvpp2?
> >>>>
> >>>> I'll try to test. Please let me know what kind of testing you're
> >>>> looking for (I haven't been following these patches, sorry.)
> >>>>
> >>>
> >>> A drop test or L2 routing will be enough.
> >>> BTW I should have the macchiatobin back on friday.
> >>
> >> I have a 10G packet generator connected to 10G ports of CN913x-DB - I
> >> will stress mvpp2 in l2 forwarding early next week (I'm mostly AFK
> >> this until Monday).
> >>
> >
> > I managed to to a drop test on mvpp2. Maybe there is a slowdown but
> > it's below the measurement uncertainty.
> >
> > Perf top before:
> >
> > Overhead Shared O Symbol
> > 8.48% [kernel] [k] page_pool_put_page
> > 2.57% [kernel] [k] page_pool_refill_alloc_cache
> > 1.58% [kernel] [k] page_pool_alloc_pages
> > 0.75% [kernel] [k] page_pool_return_skb_page
> >
> > after:
> >
> > Overhead Shared O Symbol
> > 8.34% [kernel] [k] page_pool_put_page
> > 4.52% [kernel] [k] page_pool_return_skb_page
> > 4.42% [kernel] [k] page_pool_sub_bias
> > 3.16% [kernel] [k] page_pool_alloc_pages
> > 2.43% [kernel] [k] page_pool_refill_alloc_cache
>
> Hi, Matteo
> Thanks for the testing.
> it seems you have adapted the mvpp2 driver to use the new frag
> API for page pool, There is one missing optimization for XDP case,
> the page is always returned to the pool->ring regardless of the
> context of page_pool_put_page() for elevated refcnt case.
>
> Maybe adding back that optimization will close some gap of the above
> performance difference if the drop is happening in softirq context.
>

I think what Matteo did was a pure netstack test. We'll need testing on
both XDP and normal network cases to be able to figure out the exact
impact.

Thanks
/Ilias
> >
> > Regards,
> >