Re: [PATCH net-next RFC WIP] Patch for XDP support for virtio_net

["Michael S. Tsirkin" <mst@redhat.com>]

On Thu, Oct 27, 2016 at 02:09:08PM -0700, John Fastabend wrote:
> On 16-10-27 01:55 AM, Jesper Dangaard Brouer wrote:
> > On Wed, 26 Oct 2016 13:11:22 -0400 (EDT)
> > David Miller <davem@davemloft.net> wrote:
> > 
> >> From: "Michael S. Tsirkin" <mst@redhat.com>
> >> Date: Wed, 26 Oct 2016 20:07:19 +0300
> >>
> >>> On Wed, Oct 26, 2016 at 12:52:45PM -0400, David Miller wrote:  
> >>>> From: "Michael S. Tsirkin" <mst@redhat.com>
> >>>> Date: Wed, 26 Oct 2016 19:36:45 +0300
> >>>>   
> >>>>> On Wed, Oct 26, 2016 at 03:52:02PM +0200, Jesper Dangaard Brouer wrote:  
> >>>>>> On Sat, 22 Oct 2016 04:07:23 +0000
> >>>>>> Shrijeet Mukherjee <shrijeet@gmail.com> wrote:
> >>>>>>   
> >>>>>>> This patch adds support for xdp ndo and also inserts the xdp program
> >>>>>>> call into the merged RX buffers and big buffers paths  
> >>>>>>
> >>>>>> I really appreciate you are doing this for virtio_net.
> >>>>>>
> >>>>>> My first question is: Is the (packet) page data writable?
> >>>>>> (MST might be able to answer?)
> >>>>>>
> >>>>>> As this is currently an XDP requirement[1].    
> >>>>>
> >>>>> I'm not sure I understand what does writable mean.
> >>>>> Could you explain a bit more pls?
> >>>>> We do copy data into skb ATM but I plan to change that.  
> >>>>
> >>>> The packet data area must be writable,  
> >>>
> >>> This is the part I don't fully understand.
> >>> It's in RAM so of course it's writeable.  
> >>
> >> Pages in SKB frag lists are not usually writable, because they share
> >> space with data from other packets the way drivers usually carve up
> >> pages to receive packets into.
> >>
> >> It is therefore illegal for the networking code to write into SKB frag
> >> pages.
> >>
> >> Pages used for XDP processed packets must not have this restriction.
> >>
> >>> We share pages between arbitrary multiple packets. I think that's
> >>> OK  
> >>
> >> That's exactly what is not allowed with XDP.
> >>
> >> Each packet must be the sole user of a page, otherwise the semantics
> >> required by XDP are not met.
> > 
> > Looking at the virtio_net.c code, the function call receive_big() might
> > actually be okay for XDP, unless the incoming packet is larger than
> > PAGE_SIZE and thus uses several pages (via a linked list in page->private).
> > 
> > The receive_mergeable() does not look compatible with XDP.
> > 
> 
> Looks to me the correct conditions can be met by getting the correct
> feature negotiation to happen, specifically no VIRTIO_NET_F_MRG_RXBUF
> and one of the TSO/ECN/GSO feature bits set the big_packets flag as
> Jesper notes.
> 
> Srijeet, are you going to give this a try? I'm trying to get the device
> side working by the way on the vhost interface.
> 
> Thanks,
> John

Something I'd like to understand is how does XDP address the
problem that 100Byte packets are consuming 4K of memory now.

-- 
MST