Re: [RFC PATCH v7 01/19] Add a new structure for skb buffer from external.

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

On Fri, Jun 18, 2010 at 03:14:18PM +0800, Xin, Xiaohui wrote:
> >-----Original Message-----
> >From: Herbert Xu [mailto:herbert@gondor.apana.org.au]
> >Sent: Friday, June 18, 2010 1:59 PM
> >To: Xin, Xiaohui
> >Cc: Stephen Hemminger; netdev@vger.kernel.org; kvm@vger.kernel.org;
> >linux-kernel@vger.kernel.org; mst@redhat.com; mingo@elte.hu; davem@davemloft.net;
> >jdike@linux.intel.com; Rusty Russell
> >Subject: Re: [RFC PATCH v7 01/19] Add a new structure for skb buffer from external.
> >
> >On Fri, Jun 18, 2010 at 01:26:49PM +0800, Xin, Xiaohui wrote:
> >>
> >> Herbert,
> >> I have questions about the idea above:
> >> 1) Since netdev_alloc_skb() is still there, and we only modify alloc_page(),
> >> then we don't need napi_gro_frags() any more, the driver's original receiving
> >> function is ok. Right?
> >
> >Well I was actually thinking about converting all drivers that
> >need this to napi_gro_frags.  But now that you mention it, yes
> >we could still keep the old interface to minimise the work.
> >
> >> 2) Is napi_gro_frags() only suitable for TCP protocol packet?
> >> I have done a small test for ixgbe driver to let it only allocate paged buffers
> >> and found kernel hangs when napi_gro_frags() receives an ARP packet.
> >
> >It should work with any packet.  In fact, I'm pretty sure the
> >other drivers (e.g., cxgb3) use that interface for all packets.
> >
> Thanks for the verification. By the way, does that mean that nearly all drivers can use the 
> same napi_gro_frags() to receive buffers though currently each driver has it's own receiving 
> function?
> 
> >> 3) As I have mentioned above, with this idea, netdev_alloc_skb() will allocate
> >> as usual, the data pointed by skb->data will be copied into the first guest buffer.
> >> That means we should reserve sufficient room in guest buffer. For PS mode
> >> supported driver (for example ixgbe), the room will be more than 128. After 128bytes,
> >> we will put the first frag data. Look into virtio-net.c the function page_to_skb()
> >> and receive_mergeable(), that means we should modify guest virtio-net driver to
> >> compute the offset as the parameter for skb_set_frag().
> >>
> >> How do you think about this? Attached is a patch to how to modify the guest driver.
> >> I reserve 512 bytes as an example, and transfer the header len of the skb in hdr->hdr_len.
> >
> >Expanding the buffer size to 512 bytes to accomodate PS mode
> >looks reasonable to me.
> >
> >However, I don't think we should increase the copy threshold to
> >512 bytes at the same time.  I don't have any figures myself but
> >I think if we are to make such a change it should be a separate
> >one and come with supporting numbers.
> >
> Let me have a look to see if I can retain the copy threshold as 128 bytes 
> and copy the header data safely.

Changing the guest virtio to match the backend is a problem,
this breaks migration etc.


> >Cheers,
> >--
> >Visit Openswan at http://www.openswan.org/
> >Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
> >Home Page: http://gondor.apana.org.au/~herbert/
> >PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt