]> asedeno.scripts.mit.edu Git - linux.git/commitdiff
enic: add udp_tunnel ndo for vxlan offload
authorGovindarajulu Varadarajan <gvaradar@cisco.com>
Thu, 9 Feb 2017 00:43:08 +0000 (16:43 -0800)
committerDavid S. Miller <davem@davemloft.net>
Thu, 9 Feb 2017 22:24:29 +0000 (17:24 -0500)
Defines enic_udp_tunnel_add/del for configuring vxlan tunnel offload.
enic supports offload of only one ipv4/udp port.

There are two modes that fw supports for vxlan offload.

mode 0: fcoe bit is set for encapsulated packet. fcoe_fc_crc_ok is set
if checksum of csum is ok. This bit is or of ip_csum_ok and
tcp_udp_csum_ok

mode 2: BIT(0) in rss_hash is set if it is encapsulated packet.
BIT(1) is set if outer_ip_csum_ok/
BIT(2) is set if outer_tcp_csum_ok

tcp_udp_csum_ok/ipv4_csum_ok is set if inner csum is OK.

Signed-off-by: Govindarajulu Varadarajan <gvaradar@cisco.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/cisco/enic/enic.h
drivers/net/ethernet/cisco/enic/enic_main.c

index 9023c858715d5759064081eee54b120ddaace673..2b23f46b34d3a8dff54bb93ce76d26834fe7a8a8 100644 (file)
@@ -135,6 +135,11 @@ struct enic_rfs_flw_tbl {
        struct timer_list rfs_may_expire;
 };
 
+struct vxlan_offload {
+       u16 vxlan_udp_port_number;
+       u8 patch_level;
+};
+
 /* Per-instance private data structure */
 struct enic {
        struct net_device *netdev;
@@ -175,6 +180,7 @@ struct enic {
        /* receive queue cache line section */
        ____cacheline_aligned struct vnic_rq rq[ENIC_RQ_MAX];
        unsigned int rq_count;
+       struct vxlan_offload vxlan;
        u64 rq_truncated_pkts;
        u64 rq_bad_fcs;
        struct napi_struct napi[ENIC_RQ_MAX + ENIC_WQ_MAX];
index c009f6ddabf73f49b2e133651e6ee45d9f5950c5..7e56bf95cfc765939e299b5d290fbd6c3de4f536 100644 (file)
@@ -45,6 +45,7 @@
 #endif
 #include <linux/crash_dump.h>
 #include <net/busy_poll.h>
+#include <net/vxlan.h>
 
 #include "cq_enet_desc.h"
 #include "vnic_dev.h"
@@ -176,6 +177,92 @@ static void enic_unset_affinity_hint(struct enic *enic)
                irq_set_affinity_hint(enic->msix_entry[i].vector, NULL);
 }
 
+static void enic_udp_tunnel_add(struct net_device *netdev,
+                               struct udp_tunnel_info *ti)
+{
+       struct enic *enic = netdev_priv(netdev);
+       __be16 port = ti->port;
+       int err;
+
+       spin_lock_bh(&enic->devcmd_lock);
+
+       if (ti->type != UDP_TUNNEL_TYPE_VXLAN) {
+               netdev_info(netdev, "udp_tnl: only vxlan tunnel offload supported");
+               goto error;
+       }
+
+       if (ti->sa_family != AF_INET) {
+               netdev_info(netdev, "vxlan: only IPv4 offload supported");
+               goto error;
+       }
+
+       if (enic->vxlan.vxlan_udp_port_number) {
+               if (ntohs(port) == enic->vxlan.vxlan_udp_port_number)
+                       netdev_warn(netdev, "vxlan: udp port already offloaded");
+               else
+                       netdev_info(netdev, "vxlan: offload supported for only one UDP port");
+
+               goto error;
+       }
+
+       err = vnic_dev_overlay_offload_cfg(enic->vdev,
+                                          OVERLAY_CFG_VXLAN_PORT_UPDATE,
+                                          ntohs(port));
+       if (err)
+               goto error;
+
+       err = vnic_dev_overlay_offload_ctrl(enic->vdev, OVERLAY_FEATURE_VXLAN,
+                                           enic->vxlan.patch_level);
+       if (err)
+               goto error;
+
+       enic->vxlan.vxlan_udp_port_number = ntohs(port);
+
+       netdev_info(netdev, "vxlan fw-vers-%d: offload enabled for udp port: %d, sa_family: %d ",
+                   (int)enic->vxlan.patch_level, ntohs(port), ti->sa_family);
+
+       goto unlock;
+
+error:
+       netdev_info(netdev, "failed to offload udp port: %d, sa_family: %d, type: %d",
+                   ntohs(port), ti->sa_family, ti->type);
+unlock:
+       spin_unlock_bh(&enic->devcmd_lock);
+}
+
+static void enic_udp_tunnel_del(struct net_device *netdev,
+                               struct udp_tunnel_info *ti)
+{
+       struct enic *enic = netdev_priv(netdev);
+       int err;
+
+       spin_lock_bh(&enic->devcmd_lock);
+
+       if ((ti->sa_family != AF_INET) ||
+           ((ntohs(ti->port) != enic->vxlan.vxlan_udp_port_number)) ||
+           (ti->type != UDP_TUNNEL_TYPE_VXLAN)) {
+               netdev_info(netdev, "udp_tnl: port:%d, sa_family: %d, type: %d not offloaded",
+                           ntohs(ti->port), ti->sa_family, ti->type);
+               goto unlock;
+       }
+
+       err = vnic_dev_overlay_offload_ctrl(enic->vdev, OVERLAY_FEATURE_VXLAN,
+                                           OVERLAY_OFFLOAD_DISABLE);
+       if (err) {
+               netdev_err(netdev, "vxlan: del offload udp port: %d failed",
+                          ntohs(ti->port));
+               goto unlock;
+       }
+
+       enic->vxlan.vxlan_udp_port_number = 0;
+
+       netdev_info(netdev, "vxlan: del offload udp port %d, family %d\n",
+                   ntohs(ti->port), ti->sa_family);
+
+unlock:
+       spin_unlock_bh(&enic->devcmd_lock);
+}
+
 int enic_is_dynamic(struct enic *enic)
 {
        return enic->pdev->device == PCI_DEVICE_ID_CISCO_VIC_ENET_DYN;
@@ -1113,6 +1200,7 @@ static void enic_rq_indicate_buf(struct vnic_rq *rq,
        u8 packet_error;
        u16 q_number, completed_index, bytes_written, vlan_tci, checksum;
        u32 rss_hash;
+       bool outer_csum_ok = true, encap = false;
 
        if (skipped)
                return;
@@ -1161,7 +1249,8 @@ static void enic_rq_indicate_buf(struct vnic_rq *rq,
                skb_put(skb, bytes_written);
                skb->protocol = eth_type_trans(skb, netdev);
                skb_record_rx_queue(skb, q_number);
-               if (netdev->features & NETIF_F_RXHASH) {
+               if ((netdev->features & NETIF_F_RXHASH) && rss_hash &&
+                   (type == 3)) {
                        switch (rss_type) {
                        case CQ_ENET_RQ_DESC_RSS_TYPE_TCP_IPv4:
                        case CQ_ENET_RQ_DESC_RSS_TYPE_TCP_IPv6:
@@ -1175,15 +1264,39 @@ static void enic_rq_indicate_buf(struct vnic_rq *rq,
                                break;
                        }
                }
+               if (enic->vxlan.vxlan_udp_port_number) {
+                       switch (enic->vxlan.patch_level) {
+                       case 0:
+                               if (fcoe) {
+                                       encap = true;
+                                       outer_csum_ok = fcoe_fc_crc_ok;
+                               }
+                               break;
+                       case 2:
+                               if ((type == 7) &&
+                                   (rss_hash & BIT(0))) {
+                                       encap = true;
+                                       outer_csum_ok = (rss_hash & BIT(1)) &&
+                                                       (rss_hash & BIT(2));
+                               }
+                               break;
+                       }
+               }
 
                /* Hardware does not provide whole packet checksum. It only
                 * provides pseudo checksum. Since hw validates the packet
                 * checksum but not provide us the checksum value. use
                 * CHECSUM_UNNECESSARY.
+                *
+                * In case of encap pkt tcp_udp_csum_ok/tcp_udp_csum_ok is
+                * inner csum_ok. outer_csum_ok is set by hw when outer udp
+                * csum is correct or is zero.
                 */
-               if ((netdev->features & NETIF_F_RXCSUM) && tcp_udp_csum_ok &&
-                   ipv4_csum_ok)
+               if ((netdev->features & NETIF_F_RXCSUM) && !csum_not_calc &&
+                   tcp_udp_csum_ok && ipv4_csum_ok && outer_csum_ok) {
                        skb->ip_summed = CHECKSUM_UNNECESSARY;
+                       skb->csum_level = encap;
+               }
 
                if (vlan_stripped)
                        __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tci);
@@ -2285,6 +2398,8 @@ static const struct net_device_ops enic_netdev_dynamic_ops = {
 #ifdef CONFIG_RFS_ACCEL
        .ndo_rx_flow_steer      = enic_rx_flow_steer,
 #endif
+       .ndo_udp_tunnel_add     = enic_udp_tunnel_add,
+       .ndo_udp_tunnel_del     = enic_udp_tunnel_del,
 };
 
 static const struct net_device_ops enic_netdev_ops = {
@@ -2308,6 +2423,8 @@ static const struct net_device_ops enic_netdev_ops = {
 #ifdef CONFIG_RFS_ACCEL
        .ndo_rx_flow_steer      = enic_rx_flow_steer,
 #endif
+       .ndo_udp_tunnel_add     = enic_udp_tunnel_add,
+       .ndo_udp_tunnel_del     = enic_udp_tunnel_del,
 };
 
 static void enic_dev_deinit(struct enic *enic)
@@ -2683,6 +2800,39 @@ static int enic_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
                netdev->hw_features |= NETIF_F_RXHASH;
        if (ENIC_SETTING(enic, RXCSUM))
                netdev->hw_features |= NETIF_F_RXCSUM;
+       if (ENIC_SETTING(enic, VXLAN)) {
+               u64 patch_level;
+
+               netdev->hw_enc_features |= NETIF_F_RXCSUM               |
+                                          NETIF_F_TSO                  |
+                                          NETIF_F_TSO_ECN              |
+                                          NETIF_F_GSO_UDP_TUNNEL       |
+                                          NETIF_F_HW_CSUM              |
+                                          NETIF_F_GSO_UDP_TUNNEL_CSUM;
+               netdev->hw_features |= netdev->hw_enc_features;
+               /* get bit mask from hw about supported offload bit level
+                * BIT(0) = fw supports patch_level 0
+                *          fcoe bit = encap
+                *          fcoe_fc_crc_ok = outer csum ok
+                * BIT(1) = always set by fw
+                * BIT(2) = fw supports patch_level 2
+                *          BIT(0) in rss_hash = encap
+                *          BIT(1,2) in rss_hash = outer_ip_csum_ok/
+                *                                 outer_tcp_csum_ok
+                * used in enic_rq_indicate_buf
+                */
+               err = vnic_dev_get_supported_feature_ver(enic->vdev,
+                                                        VIC_FEATURE_VXLAN,
+                                                        &patch_level);
+               if (err)
+                       patch_level = 0;
+               /* mask bits that are supported by driver
+                */
+               patch_level &= BIT_ULL(0) | BIT_ULL(2);
+               patch_level = fls(patch_level);
+               patch_level = patch_level ? patch_level - 1 : 0;
+               enic->vxlan.patch_level = patch_level;
+       }
 
        netdev->features |= netdev->hw_features;
        netdev->vlan_features |= netdev->features;