Re: [RFC 00/12] io_uring zerocopy send

From: Martin KaFai Lau <kafai@fb.com>
To: Pavel Begunkov <asml.silence@gmail.com>
Cc: David Ahern <dsahern@gmail.com>, <io-uring@vger.kernel.org>,
	<netdev@vger.kernel.org>, <linux-kernel@vger.kernel.org>,
	Jakub Kicinski <kuba@kernel.org>,
	Jonathan Lemon <jonathan.lemon@gmail.com>,
	"David S . Miller" <davem@davemloft.net>,
	Willem de Bruijn <willemb@google.com>,
	Eric Dumazet <edumazet@google.com>,
	Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>,
	David Ahern <dsahern@kernel.org>, Jens Axboe <axboe@kernel.dk>
Subject: Re: [RFC 00/12] io_uring zerocopy send
Date: Thu, 2 Dec 2021 09:40:29 -0800	[thread overview]
Message-ID: <20211202174029.qtwtw7e2je7v3chl@kafai-mbp.dhcp.thefacebook.com> (raw)
In-Reply-To: <ffd25188-aa92-2d69-a749-3058d1d33bc1@gmail.com>

On Thu, Dec 02, 2021 at 03:48:14PM +0000, Pavel Begunkov wrote:
> On 12/1/21 21:51, Martin KaFai Lau wrote:
> > On Wed, Dec 01, 2021 at 08:15:28PM +0000, Pavel Begunkov wrote:
> > > On 12/1/21 19:20, David Ahern wrote:
> > > > On 12/1/21 12:11 PM, Pavel Begunkov wrote:
> > > > > btw, why a dummy device would ever go through loopback? It doesn't
> > > > > seem to make sense, though may be missing something.
> > > > 
> > > > You are sending to a local ip address, so the fib_lookup returns
> > > > RTN_LOCAL. The code makes dev_out the loopback:
> > > > 
> > > > https://git.kernel.org/pub/scm/linux/kernel/git/netdev/net-next.git/tree/net/ipv4/route.c#n2773
> > > 
> > > I see, thanks. I still don't use the skb_orphan_frags_rx() hack
> > > and it doesn't go through the loopback (for my dummy tests), just
> > > dummy_xmit() and no mention of loopback in perf data, see the
> > > flamegraph. Don't know what is the catch.
> > > 
> > > I'm illiterate of the routing paths. Can it be related to
> > > the "ip route add"? How do you get an ipv4 address for the device?
> > I also bumped into the udp-connect() => ECONNREFUSED (111) error from send-zc.
> > because I assumed no server is needed by using dummy.  Then realized
> > the cover letter mentioned msg_zerocopy is used as the server.
> > Mentioning just in case someone hits it also.
> > 
> > To tx out dummy, I did:
> > #> ip a add 10.0.0.1/24 dev dummy0
> 
> Works well for me, IOW getting the same behaviour as with my
> ip route add <ip> dev dummy0
> 
> I'm curious what is the difference bw them?
No difference.  It should be the same.  The skb should still go out
of dummy (instead of lo) and then get drop/kfree.  I think
the confusion is probably from the name "<dummy_ip_addr>" which
points to the intention that the dummy0 has this ip addr
instead of dummy having a route to this ip address.
The need for running msg_zerocopy as the server also
adds to this confusion.  There should be no need for
server in dummy test.  No skb can reach the server anyway.

> 
> 
> > #> ip -4 r
> > 10.0.0.0/24 dev dummy0 proto kernel scope link src 10.0.0.1
> > 
> > #> ./send-zc -4 -D 10.0.0.(2) -t 10 udp
> > ip -s link show dev dummy0
> > 2: dummy0: <BROADCAST,NOARP,UP,LOWER_UP> mtu 65535 qdisc noqueue state UNKNOWN mode DEFAULT group default qlen 1000
> >     link/ether 82:0f:e0:dc:f7:e6 brd ff:ff:ff:ff:ff:ff
> >     RX:    bytes packets errors dropped  missed   mcast
> >                0       0      0       0       0       0
> >     TX:    bytes packets errors dropped carrier collsns
> >     140800890299 2150397      0       0       0       0
> > 
> 
> -- 
> Pavel Begunkov