mbox series

[PATCHv2,net-next,00/10] net: support ipv4 big tcp

Message ID cover.1674526718.git.lucien.xin@gmail.com (mailing list archive)
Headers show
Series net: support ipv4 big tcp | expand

Message

Xin Long Jan. 24, 2023, 2:19 a.m. UTC
This is similar to the BIG TCP patchset added by Eric for IPv6:

  https://lwn.net/Articles/895398/

Different from IPv6, IPv4 tot_len is 16-bit long only, and IPv4 header
doesn't have exthdrs(options) for the BIG TCP packets' length. To make
it simple, as David and Paolo suggested, we set IPv4 tot_len to 0 to
indicate this might be a BIG TCP packet and use skb->len as the real
IPv4 total length.

This will work safely, as all BIG TCP packets are GSO/GRO packets and
processed on the same host as they were created; There is no padding
in GSO/GRO packets, and skb->len - network_offset is exactly the IPv4
packet total length; Also, before implementing the feature, all those
places that may get iph tot_len from BIG TCP packets are taken care
with some new APIs:

Patch 1 adds some APIs for iph tot_len setting and getting, which are
used in all these places where IPv4 BIG TCP packets may reach in Patch
2-8, and Patch 9 implements this feature and Patch 10 adds a selftest
for it.

Note that the similar change as in Patch 2-6 are also needed for IPv6
BIG TCP packets, and will be addressed in another patchset.

The similar performance test is done for IPv4 BIG TCP with 25Gbit NIC
and 1.5K MTU:

No BIG TCP:
for i in {1..10}; do netperf -t TCP_RR -H 192.168.100.1 -- -r80000,80000 -O MIN_LATENCY,P90_LATENCY,P99_LATENCY,THROUGHPUT|tail -1; done
168          322          337          3776.49
143          236          277          4654.67
128          258          288          4772.83
171          229          278          4645.77
175          228          243          4678.93
149          239          279          4599.86
164          234          268          4606.94
155          276          289          4235.82
180          255          268          4418.95
168          241          249          4417.82

Enable BIG TCP:
ip link set dev ens1f0np0 gro_max_size 128000 gso_max_size 128000
for i in {1..10}; do netperf -t TCP_RR -H 192.168.100.1 -- -r80000,80000 -O MIN_LATENCY,P90_LATENCY,P99_LATENCY,THROUGHPUT|tail -1; done
161          241          252          4821.73
174          205          217          5098.28
167          208          220          5001.43
164          228          249          4883.98
150          233          249          4914.90
180          233          244          4819.66
154          208          219          5004.92
157          209          247          4999.78
160          218          246          4842.31
174          206          217          5080.99

Thanks for the feedback from Eric and David Ahern.

v1->v2:
  - add GSO_TCP for tp_status in packet sockets to tell the af_packet
    users that this is a TCP GSO packet in Patch 8.
  - remove the fixes and the selftest for IPv6 BIG TCP, will do it in
    another patchset.
  - also check skb_is_gso() when checking if it's a GSO TCP packet in
    Patch 1.

Xin Long (10):
  net: add a couple of helpers for iph tot_len
  bridge: use skb_ip_totlen in br netfilter
  openvswitch: use skb_ip_totlen in conntrack
  net: sched: use skb_ip_totlen and iph_totlen
  netfilter: use skb_ip_totlen and iph_totlen
  cipso_ipv4: use iph_set_totlen in skbuff_setattr
  ipvlan: use skb_ip_totlen in ipvlan_get_L3_hdr
  packet: add TP_STATUS_GSO_TCP for tp_status
  net: add support for ipv4 big tcp
  selftests: add a selftest for ipv4 big tcp

 drivers/net/ipvlan/ipvlan_core.c           |   2 +-
 include/linux/ip.h                         |  21 ++++
 include/net/netfilter/nf_tables_ipv4.h     |   4 +-
 include/net/route.h                        |   3 -
 include/uapi/linux/if_packet.h             |   1 +
 net/bridge/br_netfilter_hooks.c            |   2 +-
 net/bridge/netfilter/nf_conntrack_bridge.c |   4 +-
 net/core/gro.c                             |   6 +-
 net/core/sock.c                            |  11 +-
 net/ipv4/af_inet.c                         |   7 +-
 net/ipv4/cipso_ipv4.c                      |   2 +-
 net/ipv4/ip_input.c                        |   2 +-
 net/ipv4/ip_output.c                       |   2 +-
 net/netfilter/ipvs/ip_vs_xmit.c            |   2 +-
 net/netfilter/nf_log_syslog.c              |   2 +-
 net/netfilter/xt_length.c                  |   2 +-
 net/openvswitch/conntrack.c                |   2 +-
 net/packet/af_packet.c                     |   4 +
 net/sched/act_ct.c                         |   2 +-
 net/sched/sch_cake.c                       |   2 +-
 tools/testing/selftests/net/Makefile       |   1 +
 tools/testing/selftests/net/big_tcp.sh     | 140 +++++++++++++++++++++
 22 files changed, 191 insertions(+), 33 deletions(-)
 create mode 100644 tools/testing/selftests/net/big_tcp.sh

Comments

Eric Dumazet Jan. 24, 2023, 8:27 a.m. UTC | #1
On Tue, Jan 24, 2023 at 3:20 AM Xin Long <lucien.xin@gmail.com> wrote:
>
> This is similar to the BIG TCP patchset added by Eric for IPv6:
>
>   https://lwn.net/Articles/895398/
>
> Different from IPv6, IPv4 tot_len is 16-bit long only, and IPv4 header
> doesn't have exthdrs(options) for the BIG TCP packets' length. To make
> it simple, as David and Paolo suggested, we set IPv4 tot_len to 0 to
> indicate this might be a BIG TCP packet and use skb->len as the real
> IPv4 total length.
>
> This will work safely, as all BIG TCP packets are GSO/GRO packets and
> processed on the same host as they were created; There is no padding
> in GSO/GRO packets, and skb->len - network_offset is exactly the IPv4
> packet total length; Also, before implementing the feature, all those
> places that may get iph tot_len from BIG TCP packets are taken care
> with some new APIs:
>
> Patch 1 adds some APIs for iph tot_len setting and getting, which are
> used in all these places where IPv4 BIG TCP packets may reach in Patch
> 2-8, and Patch 9 implements this feature and Patch 10 adds a selftest
> for it.
>
> Note that the similar change as in Patch 2-6 are also needed for IPv6
> BIG TCP packets, and will be addressed in another patchset.
>
> The similar performance test is done for IPv4 BIG TCP with 25Gbit NIC
> and 1.5K MTU:
>
> No BIG TCP:
> for i in {1..10}; do netperf -t TCP_RR -H 192.168.100.1 -- -r80000,80000 -O MIN_LATENCY,P90_LATENCY,P99_LATENCY,THROUGHPUT|tail -1; done
> 168          322          337          3776.49
> 143          236          277          4654.67
> 128          258          288          4772.83
> 171          229          278          4645.77
> 175          228          243          4678.93
> 149          239          279          4599.86
> 164          234          268          4606.94
> 155          276          289          4235.82
> 180          255          268          4418.95
> 168          241          249          4417.82
>

NACK again

You have not addressed my feedback.

Given the experimental nature of BIG TCP, we need separate netlink attributes,
so that we can selectively enable BIG TCP for IPV6, and not for IPV4.

Some of us do not have time to risk breaking IPV4, even if IPV4 for
our networks is less than 0.01 % of the traffic.
Xin Long Jan. 24, 2023, 3:51 p.m. UTC | #2
On Tue, Jan 24, 2023 at 3:27 AM Eric Dumazet <edumazet@google.com> wrote:
>
> On Tue, Jan 24, 2023 at 3:20 AM Xin Long <lucien.xin@gmail.com> wrote:
> >
> > This is similar to the BIG TCP patchset added by Eric for IPv6:
> >
> >   https://lwn.net/Articles/895398/
> >
> > Different from IPv6, IPv4 tot_len is 16-bit long only, and IPv4 header
> > doesn't have exthdrs(options) for the BIG TCP packets' length. To make
> > it simple, as David and Paolo suggested, we set IPv4 tot_len to 0 to
> > indicate this might be a BIG TCP packet and use skb->len as the real
> > IPv4 total length.
> >
> > This will work safely, as all BIG TCP packets are GSO/GRO packets and
> > processed on the same host as they were created; There is no padding
> > in GSO/GRO packets, and skb->len - network_offset is exactly the IPv4
> > packet total length; Also, before implementing the feature, all those
> > places that may get iph tot_len from BIG TCP packets are taken care
> > with some new APIs:
> >
> > Patch 1 adds some APIs for iph tot_len setting and getting, which are
> > used in all these places where IPv4 BIG TCP packets may reach in Patch
> > 2-8, and Patch 9 implements this feature and Patch 10 adds a selftest
> > for it.
> >
> > Note that the similar change as in Patch 2-6 are also needed for IPv6
> > BIG TCP packets, and will be addressed in another patchset.
> >
> > The similar performance test is done for IPv4 BIG TCP with 25Gbit NIC
> > and 1.5K MTU:
> >
> > No BIG TCP:
> > for i in {1..10}; do netperf -t TCP_RR -H 192.168.100.1 -- -r80000,80000 -O MIN_LATENCY,P90_LATENCY,P99_LATENCY,THROUGHPUT|tail -1; done
> > 168          322          337          3776.49
> > 143          236          277          4654.67
> > 128          258          288          4772.83
> > 171          229          278          4645.77
> > 175          228          243          4678.93
> > 149          239          279          4599.86
> > 164          234          268          4606.94
> > 155          276          289          4235.82
> > 180          255          268          4418.95
> > 168          241          249          4417.82
> >
>
> NACK again
>
> You have not addressed my feedback.
>
> Given the experimental nature of BIG TCP, we need separate netlink attributes,
> so that we can selectively enable BIG TCP for IPV6, and not for IPV4.
>
That will be some change, and I will try to work on it.

While at it, just try to be clearer, about the fixes for IPv6 BIG TCP
I mentioned in this patchset. Since skb->len is trustable for GSO TCP
packets. Are you still not okay with checking the skb_ipv6_pktlen()
API added to fix them in netfilter/tc/bridge/openvswitch?

Thanks.
Eric Dumazet Jan. 24, 2023, 4:34 p.m. UTC | #3
On Tue, Jan 24, 2023 at 4:51 PM Xin Long <lucien.xin@gmail.com> wrote:
>
> On Tue, Jan 24, 2023 at 3:27 AM Eric Dumazet <edumazet@google.com> wrote:
> >
> > On Tue, Jan 24, 2023 at 3:20 AM Xin Long <lucien.xin@gmail.com> wrote:
> > >
> > > This is similar to the BIG TCP patchset added by Eric for IPv6:
> > >
> > >   https://lwn.net/Articles/895398/
> > >
> > > Different from IPv6, IPv4 tot_len is 16-bit long only, and IPv4 header
> > > doesn't have exthdrs(options) for the BIG TCP packets' length. To make
> > > it simple, as David and Paolo suggested, we set IPv4 tot_len to 0 to
> > > indicate this might be a BIG TCP packet and use skb->len as the real
> > > IPv4 total length.
> > >
> > > This will work safely, as all BIG TCP packets are GSO/GRO packets and
> > > processed on the same host as they were created; There is no padding
> > > in GSO/GRO packets, and skb->len - network_offset is exactly the IPv4
> > > packet total length; Also, before implementing the feature, all those
> > > places that may get iph tot_len from BIG TCP packets are taken care
> > > with some new APIs:
> > >
> > > Patch 1 adds some APIs for iph tot_len setting and getting, which are
> > > used in all these places where IPv4 BIG TCP packets may reach in Patch
> > > 2-8, and Patch 9 implements this feature and Patch 10 adds a selftest
> > > for it.
> > >
> > > Note that the similar change as in Patch 2-6 are also needed for IPv6
> > > BIG TCP packets, and will be addressed in another patchset.
> > >
> > > The similar performance test is done for IPv4 BIG TCP with 25Gbit NIC
> > > and 1.5K MTU:
> > >
> > > No BIG TCP:
> > > for i in {1..10}; do netperf -t TCP_RR -H 192.168.100.1 -- -r80000,80000 -O MIN_LATENCY,P90_LATENCY,P99_LATENCY,THROUGHPUT|tail -1; done
> > > 168          322          337          3776.49
> > > 143          236          277          4654.67
> > > 128          258          288          4772.83
> > > 171          229          278          4645.77
> > > 175          228          243          4678.93
> > > 149          239          279          4599.86
> > > 164          234          268          4606.94
> > > 155          276          289          4235.82
> > > 180          255          268          4418.95
> > > 168          241          249          4417.82
> > >
> >
> > NACK again
> >
> > You have not addressed my feedback.
> >
> > Given the experimental nature of BIG TCP, we need separate netlink attributes,
> > so that we can selectively enable BIG TCP for IPV6, and not for IPV4.
> >
> That will be some change, and I will try to work on it.
>
> While at it, just try to be clearer, about the fixes for IPv6 BIG TCP
> I mentioned in this patchset. Since skb->len is trustable for GSO TCP
> packets. Are you still not okay with checking the skb_ipv6_pktlen()
> API added to fix them in netfilter/tc/bridge/openvswitch?
>

Are you speaking of length_mt6() ?

Quite frankly I do not think its implementation should care of GSO or anything.

Considering the definition of this thing clearly never thought of
having big packets,
and an overflow was already possible, I do not see how you can fix it
without some hack...

struct xt_length_info {
    __u16 min, max;
    __u8 invert;
};

Something like:

diff --git a/net/netfilter/xt_length.c b/net/netfilter/xt_length.c
index 1873da3a945abbc6e8849e4555b42acdd34cff2d..90eba619cbe1d11f0fdd394f6dfda2b03fa573cd
100644
--- a/net/netfilter/xt_length.c
+++ b/net/netfilter/xt_length.c
@@ -30,8 +30,7 @@ static bool
 length_mt6(const struct sk_buff *skb, struct xt_action_param *par)
 {
        const struct xt_length_info *info = par->matchinfo;
-       const u_int16_t pktlen = ntohs(ipv6_hdr(skb)->payload_len) +
-                                sizeof(struct ipv6hdr);
+       u32 pktlen = min_t(u32, skb->len, 65535);

        return (pktlen >= info->min && pktlen <= info->max) ^ info->invert;
 }
Xin Long Jan. 24, 2023, 5:14 p.m. UTC | #4
On Tue, Jan 24, 2023 at 11:35 AM Eric Dumazet <edumazet@google.com> wrote:
>
> On Tue, Jan 24, 2023 at 4:51 PM Xin Long <lucien.xin@gmail.com> wrote:
> >
> > On Tue, Jan 24, 2023 at 3:27 AM Eric Dumazet <edumazet@google.com> wrote:
> > >
> > > On Tue, Jan 24, 2023 at 3:20 AM Xin Long <lucien.xin@gmail.com> wrote:
> > > >
> > > > This is similar to the BIG TCP patchset added by Eric for IPv6:
> > > >
> > > >   https://lwn.net/Articles/895398/
> > > >
> > > > Different from IPv6, IPv4 tot_len is 16-bit long only, and IPv4 header
> > > > doesn't have exthdrs(options) for the BIG TCP packets' length. To make
> > > > it simple, as David and Paolo suggested, we set IPv4 tot_len to 0 to
> > > > indicate this might be a BIG TCP packet and use skb->len as the real
> > > > IPv4 total length.
> > > >
> > > > This will work safely, as all BIG TCP packets are GSO/GRO packets and
> > > > processed on the same host as they were created; There is no padding
> > > > in GSO/GRO packets, and skb->len - network_offset is exactly the IPv4
> > > > packet total length; Also, before implementing the feature, all those
> > > > places that may get iph tot_len from BIG TCP packets are taken care
> > > > with some new APIs:
> > > >
> > > > Patch 1 adds some APIs for iph tot_len setting and getting, which are
> > > > used in all these places where IPv4 BIG TCP packets may reach in Patch
> > > > 2-8, and Patch 9 implements this feature and Patch 10 adds a selftest
> > > > for it.
> > > >
> > > > Note that the similar change as in Patch 2-6 are also needed for IPv6
> > > > BIG TCP packets, and will be addressed in another patchset.
> > > >
> > > > The similar performance test is done for IPv4 BIG TCP with 25Gbit NIC
> > > > and 1.5K MTU:
> > > >
> > > > No BIG TCP:
> > > > for i in {1..10}; do netperf -t TCP_RR -H 192.168.100.1 -- -r80000,80000 -O MIN_LATENCY,P90_LATENCY,P99_LATENCY,THROUGHPUT|tail -1; done
> > > > 168          322          337          3776.49
> > > > 143          236          277          4654.67
> > > > 128          258          288          4772.83
> > > > 171          229          278          4645.77
> > > > 175          228          243          4678.93
> > > > 149          239          279          4599.86
> > > > 164          234          268          4606.94
> > > > 155          276          289          4235.82
> > > > 180          255          268          4418.95
> > > > 168          241          249          4417.82
> > > >
> > >
> > > NACK again
> > >
> > > You have not addressed my feedback.
> > >
> > > Given the experimental nature of BIG TCP, we need separate netlink attributes,
> > > so that we can selectively enable BIG TCP for IPV6, and not for IPV4.
> > >
> > That will be some change, and I will try to work on it.
> >
> > While at it, just try to be clearer, about the fixes for IPv6 BIG TCP
> > I mentioned in this patchset. Since skb->len is trustable for GSO TCP
> > packets. Are you still not okay with checking the skb_ipv6_pktlen()
> > API added to fix them in netfilter/tc/bridge/openvswitch?
> >
>
> Are you speaking of length_mt6() ?
Yes, but not only there, see also:

[1]:
dump_ipv6_packet()
nf_ct_bridge_pre()
ovs_skb_network_trim()/tcf_ct_skb_network_trim()
cake_ack_filter()

These places get pktlen directly from ipv6_hdr(skb)->payload_len.

>
> Quite frankly I do not think its implementation should care of GSO or anything.
Agree, and IPv6 jumbo packets don't only include IPv6 BIG TCP.

But,
>
> Considering the definition of this thing clearly never thought of
> having big packets,
> and an overflow was already possible, I do not see how you can fix it
> without some hack...
>
> struct xt_length_info {
>     __u16 min, max;
>     __u8 invert;
> };
>
> Something like:
>
> diff --git a/net/netfilter/xt_length.c b/net/netfilter/xt_length.c
> index 1873da3a945abbc6e8849e4555b42acdd34cff2d..90eba619cbe1d11f0fdd394f6dfda2b03fa573cd
> 100644
> --- a/net/netfilter/xt_length.c
> +++ b/net/netfilter/xt_length.c
> @@ -30,8 +30,7 @@ static bool
>  length_mt6(const struct sk_buff *skb, struct xt_action_param *par)
>  {
>         const struct xt_length_info *info = par->matchinfo;
> -       const u_int16_t pktlen = ntohs(ipv6_hdr(skb)->payload_len) +
> -                                sizeof(struct ipv6hdr);
> +       u32 pktlen = min_t(u32, skb->len, 65535);
do you also expect this to be applied to other places in [1] above?

I was just thinking since skb->len is trustable for GSO TCP (not all jumbo
packets), we should use skb->len as the pktlen only when it's a GSO TCP
packet in such places.

Thanks.
Eric Dumazet Jan. 24, 2023, 5:25 p.m. UTC | #5
On Tue, Jan 24, 2023 at 6:14 PM Xin Long <lucien.xin@gmail.com> wrote:
>
> On Tue, Jan 24, 2023 at 11:35 AM Eric Dumazet <edumazet@google.com> wrote:
> >
> > On Tue, Jan 24, 2023 at 4:51 PM Xin Long <lucien.xin@gmail.com> wrote:
> > >
> > > On Tue, Jan 24, 2023 at 3:27 AM Eric Dumazet <edumazet@google.com> wrote:
> > > >
> > > > On Tue, Jan 24, 2023 at 3:20 AM Xin Long <lucien.xin@gmail.com> wrote:
> > > > >
> > > > > This is similar to the BIG TCP patchset added by Eric for IPv6:
> > > > >
> > > > >   https://lwn.net/Articles/895398/
> > > > >
> > > > > Different from IPv6, IPv4 tot_len is 16-bit long only, and IPv4 header
> > > > > doesn't have exthdrs(options) for the BIG TCP packets' length. To make
> > > > > it simple, as David and Paolo suggested, we set IPv4 tot_len to 0 to
> > > > > indicate this might be a BIG TCP packet and use skb->len as the real
> > > > > IPv4 total length.
> > > > >
> > > > > This will work safely, as all BIG TCP packets are GSO/GRO packets and
> > > > > processed on the same host as they were created; There is no padding
> > > > > in GSO/GRO packets, and skb->len - network_offset is exactly the IPv4
> > > > > packet total length; Also, before implementing the feature, all those
> > > > > places that may get iph tot_len from BIG TCP packets are taken care
> > > > > with some new APIs:
> > > > >
> > > > > Patch 1 adds some APIs for iph tot_len setting and getting, which are
> > > > > used in all these places where IPv4 BIG TCP packets may reach in Patch
> > > > > 2-8, and Patch 9 implements this feature and Patch 10 adds a selftest
> > > > > for it.
> > > > >
> > > > > Note that the similar change as in Patch 2-6 are also needed for IPv6
> > > > > BIG TCP packets, and will be addressed in another patchset.
> > > > >
> > > > > The similar performance test is done for IPv4 BIG TCP with 25Gbit NIC
> > > > > and 1.5K MTU:
> > > > >
> > > > > No BIG TCP:
> > > > > for i in {1..10}; do netperf -t TCP_RR -H 192.168.100.1 -- -r80000,80000 -O MIN_LATENCY,P90_LATENCY,P99_LATENCY,THROUGHPUT|tail -1; done
> > > > > 168          322          337          3776.49
> > > > > 143          236          277          4654.67
> > > > > 128          258          288          4772.83
> > > > > 171          229          278          4645.77
> > > > > 175          228          243          4678.93
> > > > > 149          239          279          4599.86
> > > > > 164          234          268          4606.94
> > > > > 155          276          289          4235.82
> > > > > 180          255          268          4418.95
> > > > > 168          241          249          4417.82
> > > > >
> > > >
> > > > NACK again
> > > >
> > > > You have not addressed my feedback.
> > > >
> > > > Given the experimental nature of BIG TCP, we need separate netlink attributes,
> > > > so that we can selectively enable BIG TCP for IPV6, and not for IPV4.
> > > >
> > > That will be some change, and I will try to work on it.
> > >
> > > While at it, just try to be clearer, about the fixes for IPv6 BIG TCP
> > > I mentioned in this patchset. Since skb->len is trustable for GSO TCP
> > > packets. Are you still not okay with checking the skb_ipv6_pktlen()
> > > API added to fix them in netfilter/tc/bridge/openvswitch?
> > >
> >
> > Are you speaking of length_mt6() ?
> Yes, but not only there, see also:
>
> [1]:
> dump_ipv6_packet()
> nf_ct_bridge_pre()
> ovs_skb_network_trim()/tcf_ct_skb_network_trim()
> cake_ack_filter()
>
> These places get pktlen directly from ipv6_hdr(skb)->payload_len.

dump_ipv6_packet() is right if its intent is to 'dump header values'
If jumbo is not yet parsed, this should be added.

cake_ack_filter() depends on receiving non malicious packets, which is
not guaranteed.

Special casing GSO seems the wrong way to fix buggy spots like these.