Re: [PATCH net-next V2 1/3] tap: use build_skb() for small packet

[Jason Wang]

This is a multi-part message in MIME format.

On 2017å08æ16æ 12:07, Jason Wang wrote:
> On 2017å08æ16æ 11:59, Michael S. Tsirkin wrote:
> > On Wed, Aug 16, 2017 at 11:57:51AM +0800, Jason Wang wrote:
> > > On 2017å08æ16æ 11:55, Michael S. Tsirkin wrote:
> > > > On Tue, Aug 15, 2017 at 08:45:20PM -0700, Eric Dumazet wrote:
> > > > > On Fri, 2017-08-11 at 19:41 +0800, Jason Wang wrote:
> > > > > > We use tun_alloc_skb() which calls sock_alloc_send_pskb() to 
> > > > > > allocate
> > > > > > skb in the past. This socket based method is not suitable for high
> > > > > > speed userspace like virtualization which usually:
> > > > > >
> > > > > > - ignore sk_sndbuf (INT_MAX) and expect to receive the packet as 
> > > > > > fast as
> > > > > >     possible
> > > > > > - don't want to be block at sendmsg()
> > > > > >
> > > > > > To eliminate the above overheads, this patch tries to use 
> > > > > > build_skb()
> > > > > > for small packet. We will do this only when the following conditions
> > > > > > are all met:
> > > > > >
> > > > > > - TAP instead of TUN
> > > > > > - sk_sndbuf is INT_MAX
> > > > > > - caller don't want to be blocked
> > > > > > - zerocopy is not used
> > > > > > - packet size is smaller enough to use build_skb()
> > > > > >
> > > > > > Pktgen from guest to host shows ~11% improvement for rx pps of tap:
> > > > > >
> > > > > > Before: ~1.70Mpps
> > > > > > After : ~1.88Mpps
> > > > > >
> > > > > > What's more important, this makes it possible to implement XDP 
> > > > > > for tap
> > > > > > before creating skbs.
> > > > > Well well well.
> > > > >
> > > > > You do realize that tun_build_skb() is not thread safe ?
> > > > The issue is alloc frag, isn't it?
> > > > I guess for now we can limit this to XDP mode only, and
> > > > just allocate full pages in that mode.
> > > Limit this to XDP mode only does not prevent user from sending 
> > > packets to
> > > same queue in parallel I think?
> > >
> > > Thanks
> > Yes but then you can just drop the page frag allocator since
> > XDP is assumed not to care about truesize for most packets.
>
> Ok, let me do some test to see the numbers between the two methods first.
>
> Thanks

It looks like full page allocation just produce too much stress on the 
page allocator.

I get 1.58Mpps (full page) vs 1.95Mpps (page frag) with the patches 
attached.

Since non-XDP case can also benefit from build_skb(), I tend to use 
spinlock instead of full page in this case.

Thanks