Re: [PATCH] net: core: page_pool: add user refcnt and reintroduce page_pool_destroy

[Ivan Khoronzhuk]

On Wed, Jul 03, 2019 at 12:15:36AM +0300, Ilias Apalodimas wrote:
> Hi Jesper,
> Getting late here, i'll respond in detail tomorrow. One point though
>
> [...]
> > This special use-case, seems confined to your driver. And Ilias told me
> > that XDP is not really a performance benefit for this driver as the HW
> > PPS-limit is hit before the XDP and netstack limit.  I ask, does it
> > make sense to add XDP to this driver, if it complicates the code for
> > everybody else?
> I think yes. This is a widely used driver on TI embedded devices so having XDP
> to play along is a nice feature. It's also the first and only armv7 we have
> supporting this. Ivan already found a couple of issues due to the 32-bit
> architecture he is trying to fix, i think there's real benefit in having that,
> performance aside.
> I fully agree we should not impact the performance of the API to support a
> special hardware though. I'll have a look on the 2 solutions tomorrow, but the
> general approach on this one should be 'the simpler the better'
>
> Cheers
> /Ilias

BTW even w/o optimization it has close to 300kpps (but with increased number of
descs) on drop which is very close to netsec measurements Ilias sent. But from
what I know there is no h/w limit on cpsw at all that this CPU can serve, so
my assumption it's rather s/w limit. But that's not main here and XDP usage
has not been estimated enough yet in embedded, where hi speed not only benefit
that can be taken from XDP.

I need more clear circumstances to send v6 ...

--
Regards,
Ivan Khoronzhuk