Message-ID: <CAKgT0Ue_-2KTMY6LsJjonTC7A16jB3hJZj+VU=4Fcp_9ZoNf+g@mail.gmail.com>
Date:	Mon, 13 Jun 2016 19:50:02 -0700
From:	Alexander Duyck <alexander.duyck@...il.com>
To:	Tom Herbert <tom@...bertland.com>
Cc:	Alexander Duyck <aduyck@...antis.com>,
	Linux Kernel Network Developers <netdev@...r.kernel.org>,
	intel-wired-lan <intel-wired-lan@...ts.osuosl.org>,
	Hannes Frederic Sowa <hannes@...hat.com>,
	Jesse Gross <jesse@...nel.org>, Jiri Benc <jbenc@...hat.com>,
	Saeed Mahameed <saeedm@...lanox.com>,
	Ariel Elior <ariel.elior@...gic.com>,
	Dept-GELinuxNICDev@...gic.com,
	"David S. Miller" <davem@...emloft.net>,
	Eugenia Emantayev <eugenia@...lanox.com>,
	"Brandeburg, Jesse" <jesse.brandeburg@...el.com>
Subject: Re: [net-next PATCH 01/15] net: Combine GENEVE and VXLAN port offload
 notifiers into single functions

On Mon, Jun 13, 2016 at 5:28 PM, Tom Herbert <tom@...bertland.com> wrote:
> On Mon, Jun 13, 2016 at 4:12 PM, Alexander Duyck
> <alexander.duyck@...il.com> wrote:
>> On Mon, Jun 13, 2016 at 3:17 PM, Tom Herbert <tom@...bertland.com> wrote:
>>> On Mon, Jun 13, 2016 at 2:51 PM, Alexander Duyck
>>> <alexander.duyck@...il.com> wrote:
>>>> On Mon, Jun 13, 2016 at 1:36 PM, Tom Herbert <tom@...bertland.com> wrote:
>>>>> On Mon, Jun 13, 2016 at 1:24 PM, Alexander Duyck
>>>>> <alexander.duyck@...il.com> wrote:
>>>>>> On Mon, Jun 13, 2016 at 12:55 PM, Tom Herbert <tom@...bertland.com> wrote:
>>>>>>> On Mon, Jun 13, 2016 at 10:47 AM, Alexander Duyck <aduyck@...antis.com> wrote:
>>>>>>>> This patch merges the GENEVE and VXLAN code so that both functions pass
>>>>>>>> through a shared code path.  This way we can start the effort of using a
>>>>>>>> single function on the network device drivers to handle both of these
>>>>>>>> tunnel offload types.
>>>>>>>>
>>>>>>>> Signed-off-by: Alexander Duyck <aduyck@...antis.com>
>>>>>>>> ---
>>>>>>>>  drivers/net/geneve.c     |   48 ++++-------------------------
>>>>>>>>  drivers/net/vxlan.c      |   46 ++++-----------------------
>>>>>>>>  include/net/udp_tunnel.h |   12 +++++++
>>>>>>>>  net/ipv4/udp_tunnel.c    |   77 ++++++++++++++++++++++++++++++++++++++++++++++
>>>>>>>>  4 files changed, 103 insertions(+), 80 deletions(-)
>>>>>>>>
>>>>>>>> diff --git a/drivers/net/geneve.c b/drivers/net/geneve.c
>>>>>>>> index cadefe4fdaa2..f5ce41532cf4 100644
>>>>>>>> --- a/drivers/net/geneve.c
>>>>>>>> +++ b/drivers/net/geneve.c
>>>>>>>> @@ -399,19 +399,7 @@ static struct socket *geneve_create_sock(struct net *net, bool ipv6,
>>>>>>>>
>>>>>>>>  static void geneve_notify_add_rx_port(struct geneve_sock *gs)
>>>>>>>>  {
>>>>>>>> -       struct net_device *dev;
>>>>>>>> -       struct sock *sk = gs->sock->sk;
>>>>>>>> -       struct net *net = sock_net(sk);
>>>>>>>> -       sa_family_t sa_family = geneve_get_sk_family(gs);
>>>>>>>> -       __be16 port = inet_sk(sk)->inet_sport;
>>>>>>>> -
>>>>>>>> -       rcu_read_lock();
>>>>>>>> -       for_each_netdev_rcu(net, dev) {
>>>>>>>> -               if (dev->netdev_ops->ndo_add_geneve_port)
>>>>>>>> -                       dev->netdev_ops->ndo_add_geneve_port(dev, sa_family,
>>>>>>>> -                                                            port);
>>>>>>>> -       }
>>>>>>>> -       rcu_read_unlock();
>>>>>>>> +       udp_tunnel_notify_add_rx_port(gs->sock, UDP_ENC_OFFLOAD_TYPE_GENEVE);
>>>>>>>>  }
>>>>>>>>
>>>>>>>>  static int geneve_hlen(struct genevehdr *gh)
>>>>>>>> @@ -550,20 +538,7 @@ static struct geneve_sock *geneve_socket_create(struct net *net, __be16 port,
>>>>>>>>
>>>>>>>>  static void geneve_notify_del_rx_port(struct geneve_sock *gs)
>>>>>>>>  {
>>>>>>>> -       struct net_device *dev;
>>>>>>>> -       struct sock *sk = gs->sock->sk;
>>>>>>>> -       struct net *net = sock_net(sk);
>>>>>>>> -       sa_family_t sa_family = geneve_get_sk_family(gs);
>>>>>>>> -       __be16 port = inet_sk(sk)->inet_sport;
>>>>>>>> -
>>>>>>>> -       rcu_read_lock();
>>>>>>>> -       for_each_netdev_rcu(net, dev) {
>>>>>>>> -               if (dev->netdev_ops->ndo_del_geneve_port)
>>>>>>>> -                       dev->netdev_ops->ndo_del_geneve_port(dev, sa_family,
>>>>>>>> -                                                            port);
>>>>>>>> -       }
>>>>>>>> -
>>>>>>>> -       rcu_read_unlock();
>>>>>>>> +       udp_tunnel_notify_add_rx_port(gs->sock, UDP_ENC_OFFLOAD_TYPE_GENEVE);
>>>>>>>>  }
>>>>>>>>
>>>>>>>>  static void __geneve_sock_release(struct geneve_sock *gs)
>>>>>>>> @@ -1165,29 +1140,20 @@ static struct device_type geneve_type = {
>>>>>>>>         .name = "geneve",
>>>>>>>>  };
>>>>>>>>
>>>>>>>> -/* Calls the ndo_add_geneve_port of the caller in order to
>>>>>>>> +/* Calls the ndo_add_udp_enc_port of the caller in order to
>>>>>>>>   * supply the listening GENEVE udp ports. Callers are expected
>>>>>>>> - * to implement the ndo_add_geneve_port.
>>>>>>>> + * to implement the ndo_add_udp_enc_port.
>>>>>>>>   */
>>>>>>>>  static void geneve_push_rx_ports(struct net_device *dev)
>>>>>>>>  {
>>>>>>>>         struct net *net = dev_net(dev);
>>>>>>>>         struct geneve_net *gn = net_generic(net, geneve_net_id);
>>>>>>>>         struct geneve_sock *gs;
>>>>>>>> -       sa_family_t sa_family;
>>>>>>>> -       struct sock *sk;
>>>>>>>> -       __be16 port;
>>>>>>>> -
>>>>>>>> -       if (!dev->netdev_ops->ndo_add_geneve_port)
>>>>>>>> -               return;
>>>>>>>>
>>>>>>>>         rcu_read_lock();
>>>>>>>> -       list_for_each_entry_rcu(gs, &gn->sock_list, list) {
>>>>>>>> -               sk = gs->sock->sk;
>>>>>>>> -               sa_family = sk->sk_family;
>>>>>>>> -               port = inet_sk(sk)->inet_sport;
>>>>>>>> -               dev->netdev_ops->ndo_add_geneve_port(dev, sa_family, port);
>>>>>>>> -       }
>>>>>>>> +       list_for_each_entry_rcu(gs, &gn->sock_list, list)
>>>>>>>> +               udp_tunnel_push_rx_port(dev, gs->sock,
>>>>>>>> +                                       UDP_ENC_OFFLOAD_TYPE_GENEVE);
>>>>>>>>         rcu_read_unlock();
>>>>>>>>  }
>>>>>>>>
>>>>>>>> diff --git a/drivers/net/vxlan.c b/drivers/net/vxlan.c
>>>>>>>> index f999db2f97b4..43f634282726 100644
>>>>>>>> --- a/drivers/net/vxlan.c
>>>>>>>> +++ b/drivers/net/vxlan.c
>>>>>>>> @@ -622,37 +622,13 @@ static int vxlan_gro_complete(struct sock *sk, struct sk_buff *skb, int nhoff)
>>>>>>>>  /* Notify netdevs that UDP port started listening */
>>>>>>>>  static void vxlan_notify_add_rx_port(struct vxlan_sock *vs)
>>>>>>>>  {
>>>>>>>> -       struct net_device *dev;
>>>>>>>> -       struct sock *sk = vs->sock->sk;
>>>>>>>> -       struct net *net = sock_net(sk);
>>>>>>>> -       sa_family_t sa_family = vxlan_get_sk_family(vs);
>>>>>>>> -       __be16 port = inet_sk(sk)->inet_sport;
>>>>>>>> -
>>>>>>>> -       rcu_read_lock();
>>>>>>>> -       for_each_netdev_rcu(net, dev) {
>>>>>>>> -               if (dev->netdev_ops->ndo_add_vxlan_port)
>>>>>>>> -                       dev->netdev_ops->ndo_add_vxlan_port(dev, sa_family,
>>>>>>>> -                                                           port);
>>>>>>>> -       }
>>>>>>>> -       rcu_read_unlock();
>>>>>>>> +       udp_tunnel_notify_add_rx_port(vs->sock, UDP_ENC_OFFLOAD_TYPE_VXLAN);
>>>>>>>>  }
>>>>>>>>
>>>>>>>>  /* Notify netdevs that UDP port is no more listening */
>>>>>>>>  static void vxlan_notify_del_rx_port(struct vxlan_sock *vs)
>>>>>>>>  {
>>>>>>>> -       struct net_device *dev;
>>>>>>>> -       struct sock *sk = vs->sock->sk;
>>>>>>>> -       struct net *net = sock_net(sk);
>>>>>>>> -       sa_family_t sa_family = vxlan_get_sk_family(vs);
>>>>>>>> -       __be16 port = inet_sk(sk)->inet_sport;
>>>>>>>> -
>>>>>>>> -       rcu_read_lock();
>>>>>>>> -       for_each_netdev_rcu(net, dev) {
>>>>>>>> -               if (dev->netdev_ops->ndo_del_vxlan_port)
>>>>>>>> -                       dev->netdev_ops->ndo_del_vxlan_port(dev, sa_family,
>>>>>>>> -                                                           port);
>>>>>>>> -       }
>>>>>>>> -       rcu_read_unlock();
>>>>>>>> +       udp_tunnel_notify_del_rx_port(vs->sock, UDP_ENC_OFFLOAD_TYPE_VXLAN);
>>>>>>>>  }
>>>>>>>>
>>>>>>>>  /* Add new entry to forwarding table -- assumes lock held */
>>>>>>>> @@ -2525,30 +2501,22 @@ static struct device_type vxlan_type = {
>>>>>>>>         .name = "vxlan",
>>>>>>>>  };
>>>>>>>>
>>>>>>>> -/* Calls the ndo_add_vxlan_port of the caller in order to
>>>>>>>> +/* Calls the ndo_add_udp_enc_port of the caller in order to
>>>>>>>>   * supply the listening VXLAN udp ports. Callers are expected
>>>>>>>> - * to implement the ndo_add_vxlan_port.
>>>>>>>> + * to implement the ndo_add_udp_enc_port.
>>>>>>>>   */
>>>>>>>>  static void vxlan_push_rx_ports(struct net_device *dev)
>>>>>>>>  {
>>>>>>>>         struct vxlan_sock *vs;
>>>>>>>>         struct net *net = dev_net(dev);
>>>>>>>>         struct vxlan_net *vn = net_generic(net, vxlan_net_id);
>>>>>>>> -       sa_family_t sa_family;
>>>>>>>> -       __be16 port;
>>>>>>>>         unsigned int i;
>>>>>>>>
>>>>>>>> -       if (!dev->netdev_ops->ndo_add_vxlan_port)
>>>>>>>> -               return;
>>>>>>>> -
>>>>>>>>         spin_lock(&vn->sock_lock);
>>>>>>>>         for (i = 0; i < PORT_HASH_SIZE; ++i) {
>>>>>>>> -               hlist_for_each_entry_rcu(vs, &vn->sock_list[i], hlist) {
>>>>>>>> -                       port = inet_sk(vs->sock->sk)->inet_sport;
>>>>>>>> -                       sa_family = vxlan_get_sk_family(vs);
>>>>>>>> -                       dev->netdev_ops->ndo_add_vxlan_port(dev, sa_family,
>>>>>>>> -                                                           port);
>>>>>>>> -               }
>>>>>>>> +               hlist_for_each_entry_rcu(vs, &vn->sock_list[i], hlist)
>>>>>>>> +                       udp_tunnel_push_rx_port(dev, vs->sock,
>>>>>>>> +                                               UDP_ENC_OFFLOAD_TYPE_VXLAN);
>>>>>>>>         }
>>>>>>>>         spin_unlock(&vn->sock_lock);
>>>>>>>>  }
>>>>>>>> diff --git a/include/net/udp_tunnel.h b/include/net/udp_tunnel.h
>>>>>>>> index 9d14f707e534..704f931fd9ad 100644
>>>>>>>> --- a/include/net/udp_tunnel.h
>>>>>>>> +++ b/include/net/udp_tunnel.h
>>>>>>>> @@ -84,6 +84,18 @@ struct udp_tunnel_sock_cfg {
>>>>>>>>  void setup_udp_tunnel_sock(struct net *net, struct socket *sock,
>>>>>>>>                            struct udp_tunnel_sock_cfg *sock_cfg);
>>>>>>>>
>>>>>>>> +/* List of offloadable UDP tunnel types */
>>>>>>>> +enum udp_enc_offloads {
>>>>>>>> +       UDP_ENC_OFFLOAD_TYPE_VXLAN,     /* RFC 7348 */
>>>>>>>> +       UDP_ENC_OFFLOAD_TYPE_GENEVE,    /* draft-ietf-nvo3-geneve */
>>>>>>>> +};
>>>>>>>> +
>>>>>>> We've already had a lot of discussion on this. The clear outcome from
>>>>>>> netdev was that we need to support generic offloads and move away from
>>>>>>> protocol specific offload. Generalizing the interface to allow vendors
>>>>>>> to unnecessarily leak out protocol specific features undermines that
>>>>>>> effort.
>>>>>>
>>>>>> Then in turn we get dirty hacks like what we have right now where
>>>>>> VXLAN-GPE is attempting to reuse the VXLAN offload functions or
>>>>>> drivers that just hard-code GENEVE ports.
>>>>>>
>>>>>> Going full obstructionist on this isn't going to work.  We need to be
>>>>>> able to support these type of offloads because the switch vendors are
>>>>>> going to force the NIC vendors to do so.  We will likely never be able
>>>>>> to convince Cisco to implement an outer transmit checksum on their
>>>>>> switches.  In order to make offloads work without the outer checksum
>>>>>> we will need to be able to parse the frames in order to be able to
>>>>>> validate the inner checksum values.
>>>>>>
>>>>> NIC vendors can support checksum-complete. This works with any form of
>>>>> UDP encapsulation, and IP protocol (like extension headers), and other
>>>>> form of tunneling we can dream up. That was the whole point of Dave's
>>>>> keynote at netdev.
>>>>
>>>> Right.  That covers one tiny piece of the whole problem, but you are
>>>> holding out for hardware that may not be introduced for another 3 to 5
>>>> years.  The fact is trying to getting NIC vendors to support
>>>> checksum-complete is all well and good, but you seem to have forgotten
>>>> that NIC vendors are incredibly slow when it comes to implementing
>>>> anything.  In the meantime we will have the stuff that was already in
>>>> the pipeline coming out over the next several years.
>>>>
>>>> How about the fact that we need to know that there is a tunnel there
>>>> if we want to do anything like try to parse the inner headers of a
>>>> given tunnel on Rx?  How do you propose to solve the RSS problem?
>>>
>>> Solved by doing RSS and ECMP hash over 3-tuple of IP addresses and
>>> IPv6 flow label (not ports). Non-zero flow labels will soon be widely
>>> used over the Internet. IOS already is already setting them, Android
>>> should pick up support in the next rebase, and the MS guys have
>>> assured me that they will add support to next version of Windows. Like
>>> a generic checksum offload, flow label works with an IP protocol,
>>> extension header, fragmentation, UDP encapsulation, etc. With this
>>> there is no reason for devices to parse L4 headers just to forward a
>>> packet. HW vendors (both switches and NICs) are strongly encouraged to
>>> support them.
>>
>> Right so with all this encouragement, where are we at on getting the
>> switches out there to support this?  The switches are what is pushing
>> most of the tunnel workloads and such, and it is the switches that are
>> going to be the painful interfaces for us to try to offload data from.
>>
>> Also have you looked into the effect on regular traffic for trying to
>> deal with an extra IPv4 options header?  The Intel NICs should be able
>> to deal with it, at least for Tx, but my concern would be the NICs
>> that perform parsing based offloads and such.
>>
>> It seems like we might be arguing past each other as I am focused on
>> what we can do right now, and it seems like you are arguing the point
>> that we need features that probably won't be enabled in the ecosystem
>> for 5 to 10 years depending on the hardware release cycles.  Hopefully
>> we can agree to disagree and can address this again when new features
>> start to become available.  At that time we could probably look at
>> deprecating the driver API for the UDP ports assuming we have other
>> means of resolving this available by then.
>>
> Alex,
>
> I looked though the previous thread on protocol specific vs. generic
> offloads in http://www.spinics.net/lists/netdev/msg354311.html and the
> outcome from that. To quote davem:
>
> "Doing anything other than providing 2's complement checksums in the
> RX descriptor doesn't work.  We know this.

For Rx checksum I would agree with you there.  Although I would argue
that we can solve much of this by just providing an outer UDP checksum
on Tx anyway.  The one that annoys me the most is the ones that end up
using this for Tx, but that is a topic for another time.

The problem is this goes beyond just the Rx checksum.  There end up
being a number of Rx offloads and the like that also need to figure
out where the inner headers start in order to handle things like MAC
based filtering, RSS, and the like.

> So we will not add to our core architecture and frameworks anything
> that directly facilitates designs which we know are suboptimal.  And
> protocol specific support for tunnel offloading is suboptimal and not
> the way forward."

Yes, but then we still ended up with the GENEVE offloads being
accepted anyway.  My concern is that we didn't spend enough time
talking about the implementation details of it because everything got
drowned out by the arguments for not accepting the offloads at all.
So we ended up just cloning what was already there for VXLAN which
ended up setting a bad precedent for how we should handle this going
forward.  It would be just as easy for us to stop any future additions
after my changes as it would the current code.  The only big
difference is that with my changes applied we can strip out all the
stupid defines and would likely be arguing over 1 or 2 patches rather
than a set of patches to accomplish the same task.

> Unless Dave has changed his mind on this or you have new arguments or
> new data, I really don't see point in participating in further
> discussion on patches like this. Sorry.

I'm not going to speculate on what Dave's opinion on this is.  I'll
wait to hear it from him.

My concern at this point is that we have several issues.  Specifically
we have VXLAN-GPE trying to pass itself off as VXLAN when it clearly
is not, and I know we are going to end up with somebody eventually
trying to push this feature into the kernel.  I know for a fact there
is hardware out there that already supports it.  I'm trying to get
ahead of this and define what the interface is supposed to look like
myself so that we don't end up with somebody unfamiliar with all this
trying to push it.  This way we can avoid having some hardware vendor
on a timeline trying to push it through quick as in the case of i40e,
or somebody trying to get around it by just hard coding it into their
driver like occurred with bnxt.

While I appreciate the opinion, outright refusing to enable the
existing offloads is counterproductive.  There are customers out there
that already have this hardware.  There are driver writers out there
who are going to have to enable these features one way or another.  If
we want to be obstructionists then I am sure they can just work around
us and write up out-of-tree drivers and use something like module
parameters to enable offloads on a specific port.  Most of these
implementations only seem to support one port anyway.  I just thought
it might be better to have this figured out in the kernel so that we
didn't end up creating a bigger mess than needed with each vendor
going off and doing their own out-of-tree implementation.

- Alex

Powered by blists - more mailing lists