| Message ID | 1421318323-2547-1-git-send-email-sriharsha.basavapatna@emulex.com |
|---|---|
| State | Accepted, archived |
| Delegated to: | David Miller |
| Headers | show |
On Thu, Jan 15, 2015 at 2:38 AM, Sriharsha Basavapatna <sriharsha.basavapatna@emulex.com> wrote: > Other tunnels like GRE break while VxLAN offloads are enabled in Skyhawk-R. To > avoid this, we should restrict offload features on a per-packet basis in such > conditions. > > Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@emulex.com> > --- > v2 changes: fixed minor nits pointed out by Sergei Shtylyov > --- > drivers/net/ethernet/emulex/benet/be_main.c | 41 +++++++++++++++++++++++++-- > 1 file changed, 38 insertions(+), 3 deletions(-) > > diff --git a/drivers/net/ethernet/emulex/benet/be_main.c b/drivers/net/ethernet/emulex/benet/be_main.c > index 41a0a54..d48806b 100644 > --- a/drivers/net/ethernet/emulex/benet/be_main.c > +++ b/drivers/net/ethernet/emulex/benet/be_main.c > @@ -4383,8 +4383,9 @@ static int be_ndo_bridge_getlink(struct sk_buff *skb, u32 pid, u32 seq, > * distinguish various types of transports (VxLAN, GRE, NVGRE ..). So, offload > * is expected to work across all types of IP tunnels once exported. Skyhawk > * supports offloads for either VxLAN or NVGRE, exclusively. So we export VxLAN > - * offloads in hw_enc_features only when a VxLAN port is added. Note this only > - * ensures that other tunnels work fine while VxLAN offloads are not enabled. > + * offloads in hw_enc_features only when a VxLAN port is added. If other (non > + * VxLAN) tunnels are configured while VxLAN offloads are enabled, offloads for > + * those other tunnels are unexported on the fly through ndo_features_check(). > * > * Skyhawk supports VxLAN offloads only for one UDP dport. So, if the stack > * adds more than one port, disable offloads and don't re-enable them again > @@ -4463,7 +4464,41 @@ static netdev_features_t be_features_check(struct sk_buff *skb, > struct net_device *dev, > netdev_features_t features) > { > - return vxlan_features_check(skb, features); > + struct be_adapter *adapter = netdev_priv(dev); > + u8 l4_hdr = 0; > + > + /* The code below restricts offload features for some tunneled packets. > + * Offload features for normal (non tunnel) packets are unchanged. > + */ > + if (!skb->encapsulation || > + !(adapter->flags & BE_FLAGS_VXLAN_OFFLOADS)) > + return features; > + > + /* It's an encapsulated packet and VxLAN offloads are enabled. We > + * should disable tunnel offload features if it's not a VxLAN packet, > + * as tunnel offloads have been enabled only for VxLAN. This is done to > + * allow other tunneled traffic like GRE work fine while VxLAN > + * offloads are configured in Skyhawk-R. > + */ > + switch (vlan_get_protocol(skb)) { > + case htons(ETH_P_IP): > + l4_hdr = ip_hdr(skb)->protocol; > + break; > + case htons(ETH_P_IPV6): > + l4_hdr = ipv6_hdr(skb)->nexthdr; > + break; > + default: > + return features; > + } > + > + if (l4_hdr != IPPROTO_UDP || I don't understand why this is needed. The only GSO type with with encapsulation allowed by device features SKB_GSO_UDP_TUNNEL. This should not be GRE for instance. Do you see cases where protocol is not UDP at this point? > + skb->inner_protocol_type != ENCAP_TYPE_ETHER || > + skb->inner_protocol != htons(ETH_P_TEB) || > + skb_inner_mac_header(skb) - skb_transport_header(skb) != > + sizeof(struct udphdr) + sizeof(struct vxlanhdr)) > + return features & ~(NETIF_F_ALL_CSUM | NETIF_F_GSO_MASK); > + > + return features; > } > #endif > > -- > 1.7.9.5 > > -- > 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 -- 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
-----Original Message----- From: Tom Herbert [mailto:therbert@google.com] Sent: Thursday, January 15, 2015 8:00 AM To: Sriharsha Basavapatna Cc: Linux Netdev List Subject: Re: [PATCH v2 net] be2net: Allow GRE to work concurrently while a VxLAN tunnel is configured On Thu, Jan 15, 2015 at 2:38 AM, Sriharsha Basavapatna <sriharsha.basavapatna@emulex.com> wrote: > Other tunnels like GRE break while VxLAN offloads are enabled in > Skyhawk-R. To avoid this, we should restrict offload features on a > per-packet basis in such conditions. > > Signed-off-by: Sriharsha Basavapatna > <sriharsha.basavapatna@emulex.com> > --- > v2 changes: fixed minor nits pointed out by Sergei Shtylyov > --- > drivers/net/ethernet/emulex/benet/be_main.c | 41 +++++++++++++++++++++++++-- > 1 file changed, 38 insertions(+), 3 deletions(-) > > diff --git a/drivers/net/ethernet/emulex/benet/be_main.c > b/drivers/net/ethernet/emulex/benet/be_main.c > index 41a0a54..d48806b 100644 > --- a/drivers/net/ethernet/emulex/benet/be_main.c > +++ b/drivers/net/ethernet/emulex/benet/be_main.c > @@ -4383,8 +4383,9 @@ static int be_ndo_bridge_getlink(struct sk_buff *skb, u32 pid, u32 seq, > * distinguish various types of transports (VxLAN, GRE, NVGRE ..). So, offload > * is expected to work across all types of IP tunnels once exported. Skyhawk > * supports offloads for either VxLAN or NVGRE, exclusively. So we > export VxLAN > - * offloads in hw_enc_features only when a VxLAN port is added. Note > this only > - * ensures that other tunnels work fine while VxLAN offloads are not enabled. > + * offloads in hw_enc_features only when a VxLAN port is added. If > + other (non > + * VxLAN) tunnels are configured while VxLAN offloads are enabled, > + offloads for > + * those other tunnels are unexported on the fly through ndo_features_check(). > * > * Skyhawk supports VxLAN offloads only for one UDP dport. So, if the stack > * adds more than one port, disable offloads and don't re-enable them > again @@ -4463,7 +4464,41 @@ static netdev_features_t be_features_check(struct sk_buff *skb, > struct net_device *dev, > netdev_features_t features) > { > - return vxlan_features_check(skb, features); > + struct be_adapter *adapter = netdev_priv(dev); > + u8 l4_hdr = 0; > + > + /* The code below restricts offload features for some tunneled packets. > + * Offload features for normal (non tunnel) packets are unchanged. > + */ > + if (!skb->encapsulation || > + !(adapter->flags & BE_FLAGS_VXLAN_OFFLOADS)) > + return features; > + > + /* It's an encapsulated packet and VxLAN offloads are enabled. We > + * should disable tunnel offload features if it's not a VxLAN packet, > + * as tunnel offloads have been enabled only for VxLAN. This is done to > + * allow other tunneled traffic like GRE work fine while VxLAN > + * offloads are configured in Skyhawk-R. > + */ > + switch (vlan_get_protocol(skb)) { > + case htons(ETH_P_IP): > + l4_hdr = ip_hdr(skb)->protocol; > + break; > + case htons(ETH_P_IPV6): > + l4_hdr = ipv6_hdr(skb)->nexthdr; > + break; > + default: > + return features; > + } > + > + if (l4_hdr != IPPROTO_UDP || I don't understand why this is needed. The only GSO type with with encapsulation allowed by device features SKB_GSO_UDP_TUNNEL. This should not be GRE for instance. Do you see cases where protocol is not UDP at this point? [Harsha] It's needed for GRE checksum case; without this, GRE pkts are sent down without checksum by the stack, but HW can only offload checksum for VxLAN pkts. > + skb->inner_protocol_type != ENCAP_TYPE_ETHER || > + skb->inner_protocol != htons(ETH_P_TEB) || > + skb_inner_mac_header(skb) - skb_transport_header(skb) != > + sizeof(struct udphdr) + sizeof(struct vxlanhdr)) > + return features & ~(NETIF_F_ALL_CSUM | > + NETIF_F_GSO_MASK); > + > + return features; > } > #endif > > -- > 1.7.9.5 > > -- > 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
> I don't understand why this is needed. The only GSO type with with encapsulation allowed by device features SKB_GSO_UDP_TUNNEL. This should not be GRE for instance. Do you see cases where protocol is not UDP at this point? > [Harsha] It's needed for GRE checksum case; without this, GRE pkts are sent down without checksum by the stack, but HW > can only offload checksum for VxLAN pkts. > Okay, I see that this is a problem with checksum not GSO. Please ask your hardware guys to provide NETIF_F_HW_CSUM to avoid any more of this unpleasantness in the future :-) >> + skb->inner_protocol_type != ENCAP_TYPE_ETHER || >> + skb->inner_protocol != htons(ETH_P_TEB) || >> + skb_inner_mac_header(skb) - skb_transport_header(skb) != >> + sizeof(struct udphdr) + sizeof(struct vxlanhdr)) >> + return features & ~(NETIF_F_ALL_CSUM | >> + NETIF_F_GSO_MASK); >> + >> + return features; >> } >> #endif >> >> -- >> 1.7.9.5 >> >> -- >> 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 -- 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
From: Sriharsha Basavapatna <sriharsha.basavapatna@emulex.com> Date: Thu, 15 Jan 2015 16:08:43 +0530 > Other tunnels like GRE break while VxLAN offloads are enabled in Skyhawk-R. To > avoid this, we should restrict offload features on a per-packet basis in such > conditions. > > Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@emulex.com> Applied, thanks. -- 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
diff --git a/drivers/net/ethernet/emulex/benet/be_main.c b/drivers/net/ethernet/emulex/benet/be_main.c index 41a0a54..d48806b 100644 --- a/drivers/net/ethernet/emulex/benet/be_main.c +++ b/drivers/net/ethernet/emulex/benet/be_main.c @@ -4383,8 +4383,9 @@ static int be_ndo_bridge_getlink(struct sk_buff *skb, u32 pid, u32 seq, * distinguish various types of transports (VxLAN, GRE, NVGRE ..). So, offload * is expected to work across all types of IP tunnels once exported. Skyhawk * supports offloads for either VxLAN or NVGRE, exclusively. So we export VxLAN - * offloads in hw_enc_features only when a VxLAN port is added. Note this only - * ensures that other tunnels work fine while VxLAN offloads are not enabled. + * offloads in hw_enc_features only when a VxLAN port is added. If other (non + * VxLAN) tunnels are configured while VxLAN offloads are enabled, offloads for + * those other tunnels are unexported on the fly through ndo_features_check(). * * Skyhawk supports VxLAN offloads only for one UDP dport. So, if the stack * adds more than one port, disable offloads and don't re-enable them again @@ -4463,7 +4464,41 @@ static netdev_features_t be_features_check(struct sk_buff *skb, struct net_device *dev, netdev_features_t features) { - return vxlan_features_check(skb, features); + struct be_adapter *adapter = netdev_priv(dev); + u8 l4_hdr = 0; + + /* The code below restricts offload features for some tunneled packets. + * Offload features for normal (non tunnel) packets are unchanged. + */ + if (!skb->encapsulation || + !(adapter->flags & BE_FLAGS_VXLAN_OFFLOADS)) + return features; + + /* It's an encapsulated packet and VxLAN offloads are enabled. We + * should disable tunnel offload features if it's not a VxLAN packet, + * as tunnel offloads have been enabled only for VxLAN. This is done to + * allow other tunneled traffic like GRE work fine while VxLAN + * offloads are configured in Skyhawk-R. + */ + switch (vlan_get_protocol(skb)) { + case htons(ETH_P_IP): + l4_hdr = ip_hdr(skb)->protocol; + break; + case htons(ETH_P_IPV6): + l4_hdr = ipv6_hdr(skb)->nexthdr; + break; + default: + return features; + } + + if (l4_hdr != IPPROTO_UDP || + skb->inner_protocol_type != ENCAP_TYPE_ETHER || + skb->inner_protocol != htons(ETH_P_TEB) || + skb_inner_mac_header(skb) - skb_transport_header(skb) != + sizeof(struct udphdr) + sizeof(struct vxlanhdr)) + return features & ~(NETIF_F_ALL_CSUM | NETIF_F_GSO_MASK); + + return features; } #endif
Other tunnels like GRE break while VxLAN offloads are enabled in Skyhawk-R. To avoid this, we should restrict offload features on a per-packet basis in such conditions. Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@emulex.com> --- v2 changes: fixed minor nits pointed out by Sergei Shtylyov --- drivers/net/ethernet/emulex/benet/be_main.c | 41 +++++++++++++++++++++++++-- 1 file changed, 38 insertions(+), 3 deletions(-)