[openib-general] Dropping NETIF_F_SG since no checksum feature.

[Michael S. Tsirkin]

Quoting r. Stephen Hemminger <shemminger at osdl.org>:
> Subject: Re: Dropping NETIF_F_SG since no checksum feature.
> 
> On Wed, 11 Oct 2006 21:11:38 +0100
> Steven Whitehouse <steve at chygwyn.com> wrote:
> 
> > Hi,
> > 
> > On Wed, Oct 11, 2006 at 05:01:03PM +0200, Michael S. Tsirkin wrote:
> > > Quoting Steven Whitehouse <steve at chygwyn.com>:
> > > > > ssize_t tcp_sendpage(struct socket *sock, struct page *page, int offset,
> > > > >                      size_t size, int flags)
> > > > > {
> > > > >         ssize_t res;
> > > > >         struct sock *sk = sock->sk;
> > > > > 
> > > > >         if (!(sk->sk_route_caps & NETIF_F_SG) ||
> > > > >             !(sk->sk_route_caps & NETIF_F_ALL_CSUM))
> > > > >                 return sock_no_sendpage(sock, page, offset, size, flags);
> > > > > 
> > > > > 
> > > > > So, it seems that if I set NETIF_F_SG but clear NETIF_F_ALL_CSUM,
> > > > > data will be copied over rather than sent directly.
> > > > > So why does dev.c have to force set NETIF_F_SG to off then?
> > > > >
> > > > I agree with that analysis,
> > > 
> > > So, would you Ack something like the following then?
> > >
> > 
> > In so far as I'm able to ack it, then yes, but with the following
> > caveats: that you also need to look at the tcp code's checks for
> > NETIF_F_SG (aside from the interface to tcp_sendpage which I think
> > we've agreed is ok) and ensure that this patch will not change their
> > behaviour, and here I'm thinking of the test in net/ipv4/tcp.c:select_size()
> > in particular - there may be others but thats the only one I can think
> > of off the top of my head. I think this is what davem was getting at
> > with his comment about copy & sum for smaller packets.
> > 
> > Also all subject to approval by davem and shemminger of course :-)
> > 
> > My general feeling is that devices should advertise the features that
> > they actually have and that the protocols should make the decision
> > as to which ones to use or not depending on the combinations available
> > (which I think is pretty much your argument).
> > 
> > Steve.
> > 
> 
> You might want to try ignoring the check in dev.c and testing
> to see if there is a performance gain.  It wouldn't be hard to test
> a modified version and validate the performance change.

Yes. With my patch, there is a huge performance gain by increasing MTU to 64K.
And it seems the only way to do this is by S/G.

> You could even do what I suggested and use skb_checksum_help()
> to do inplace checksumming, as a performance test.

I can. But as network algorithmics says (chapter 5)
"Since such bus reads are expensive, the CPU might as well piggyback
the checksum computation with the copy process".

It speaks about onboard the adapter buffers, but memory bus reads are also much slower
than CPU nowdays.  So I think even if this works well in benchmark in real life
single copy should better.

-- 
MST