diff mbox

GRE with GRO very slow when forwarding starting with 3.14.24

Message ID CA+mtBx9SOr5RH7S63S28YbgxR8EwjgfS_dUkS2A4D1W3NzazHQ@mail.gmail.com
State RFC, archived
Delegated to: David Miller
Headers show

Commit Message

Tom Herbert Nov. 24, 2014, 6:16 p.m. UTC 
On Mon, Nov 24, 2014 at 8:48 AM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
>
> On Mon, 2014-11-24 at 17:13 +0100, Wolfgang Walter wrote:
> > Hello,
> >
> > starting with 3.14.24 GRE with GRO on is very slow. To be more specific:
> >
> > yyyy <--> GRO_endpoint <-_> .... <--> |eth0<->GRO-endpoint | eth1 |<-> xxxx
> >
> > routing (IPv4) between xxxx and yyyy is very slow when GRO is enabled on eth0
> > and/or eth1 starting with stable kernel 3.14.24
> >
> > Regards,
>
>
> tcpdump might help, but I presume GSO is no longer working properly on
> egress.
>
Inner mac header is probably not being set in GRO->GSO GRE path.
Please try this also:


> Can you try to revert :
>
> commit abe640984aa492652232b65d3579361cf6d461f5
> Author: Tom Herbert <therbert@google.com>
> Date:   Thu Oct 30 08:40:56 2014 -0700
>
>     gre: Use inner mac length when computing tunnel length
>
>     [ Upstream commit 14051f0452a2c26a3f4791e6ad6a435e8f1945ff ]
>
>     Currently, skb_inner_network_header is used but this does not account
>     for Ethernet header for ETH_P_TEB. Use skb_inner_mac_header which
>     handles TEB and also should work with IP encapsulation in which case
>     inner mac and inner network headers are the same.
>
>     Tested: Ran TCP_STREAM over GRE, worked as expected.
>
>     Signed-off-by: Tom Herbert <therbert@google.com>
>     Acked-by: Alexander Duyck <alexander.h.duyck@redhat.com>
>     Signed-off-by: David S. Miller <davem@davemloft.net>
>     Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
>
> diff --git a/net/ipv4/gre_offload.c b/net/ipv4/gre_offload.c
> index 2d24f293f977..8c8493ea6b1c 100644
> --- a/net/ipv4/gre_offload.c
> +++ b/net/ipv4/gre_offload.c
> @@ -50,7 +50,7 @@ static struct sk_buff *gre_gso_segment(struct sk_buff *skb,
>
>         greh = (struct gre_base_hdr *)skb_transport_header(skb);
>
> -       ghl = skb_inner_network_header(skb) - skb_transport_header(skb);
> +       ghl = skb_inner_mac_header(skb) - skb_transport_header(skb);
>         if (unlikely(ghl < sizeof(*greh)))
>                 goto out;
>
>
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Comments

Wolfgang Walter Nov. 25, 2014, 4:55 p.m. UTC | #1 
Am Montag, 24. November 2014, 10:16:26 schrieb Tom Herbert:
> On Mon, Nov 24, 2014 at 8:48 AM, Eric Dumazet <eric.dumazet@gmail.com> 
wrote:
> > On Mon, 2014-11-24 at 17:13 +0100, Wolfgang Walter wrote:
> > > Hello,
> > > 
> > > starting with 3.14.24 GRE with GRO on is very slow. To be more specific:
> > > 
> > > yyyy <--> GRO_endpoint <-_> .... <--> |eth0<->GRO-endpoint | eth1 |<->
> > > xxxx
> > > 
> > > routing (IPv4) between xxxx and yyyy is very slow when GRO is enabled on
> > > eth0 and/or eth1 starting with stable kernel 3.14.24
> > > 
> > > Regards,
> > 
> > tcpdump might help, but I presume GSO is no longer working properly on
> > egress.
> 
> Inner mac header is probably not being set in GRO->GSO GRE path.
> Please try this also:


Thanks, this patch fixes the problem.


> 
> diff --git a/net/ipv4/gre_offload.c b/net/ipv4/gre_offload.c
> index bb5947b..51973dd 100644
> --- a/net/ipv4/gre_offload.c
> +++ b/net/ipv4/gre_offload.c
> @@ -247,6 +247,9 @@ static int gre_gro_complete(struct sk_buff *skb, int
> nhoff) err = ptype->callbacks.gro_complete(skb, nhoff + grehlen);
> 
>         rcu_read_unlock();
> +
> +       skb_set_inner_mac_header(skb, nhoff + grehlen);
> +
>         return err;
>  }
> 
> > Can you try to revert :
> > 
> > commit abe640984aa492652232b65d3579361cf6d461f5
> > Author: Tom Herbert <therbert@google.com>
> > Date:   Thu Oct 30 08:40:56 2014 -0700
> > 
> >     gre: Use inner mac length when computing tunnel length
> >     
> >     [ Upstream commit 14051f0452a2c26a3f4791e6ad6a435e8f1945ff ]
> >     
> >     Currently, skb_inner_network_header is used but this does not account
> >     for Ethernet header for ETH_P_TEB. Use skb_inner_mac_header which
> >     handles TEB and also should work with IP encapsulation in which case
> >     inner mac and inner network headers are the same.
> >     
> >     Tested: Ran TCP_STREAM over GRE, worked as expected.
> >     
> >     Signed-off-by: Tom Herbert <therbert@google.com>
> >     Acked-by: Alexander Duyck <alexander.h.duyck@redhat.com>
> >     Signed-off-by: David S. Miller <davem@davemloft.net>
> >     Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
> > 
> > diff --git a/net/ipv4/gre_offload.c b/net/ipv4/gre_offload.c
> > index 2d24f293f977..8c8493ea6b1c 100644
> > --- a/net/ipv4/gre_offload.c
> > +++ b/net/ipv4/gre_offload.c
> > @@ -50,7 +50,7 @@ static struct sk_buff *gre_gso_segment(struct sk_buff
> > *skb,> 
> >         greh = (struct gre_base_hdr *)skb_transport_header(skb);
> > 
> > -       ghl = skb_inner_network_header(skb) - skb_transport_header(skb);
> > +       ghl = skb_inner_mac_header(skb) - skb_transport_header(skb);
> > 
> >         if (unlikely(ghl < sizeof(*greh)))
> >         
> >                 goto out;

Regards,
diff mbox

Patch

diff --git a/net/ipv4/gre_offload.c b/net/ipv4/gre_offload.c
index bb5947b..51973dd 100644
--- a/net/ipv4/gre_offload.c
+++ b/net/ipv4/gre_offload.c
@@ -247,6 +247,9 @@  static int gre_gro_complete(struct sk_buff *skb, int nhoff)
                err = ptype->callbacks.gro_complete(skb, nhoff + grehlen);

        rcu_read_unlock();
+
+       skb_set_inner_mac_header(skb, nhoff + grehlen);
+
        return err;
 }