datapath: Pad OVS_PACKET_ATTR_PACKET if linear copy was performed
[cascardo/ovs.git] / datapath / datapath.c
index 1a5bffb..9e9c73a 100644 (file)
@@ -110,10 +110,9 @@ int lockdep_ovsl_is_held(void)
 #endif
 
 static struct vport *new_vport(const struct vport_parms *);
-static int queue_gso_packets(struct net *, int dp_ifindex, struct sk_buff *,
+static int queue_gso_packets(struct datapath *dp, struct sk_buff *,
                             const struct dp_upcall_info *);
-static int queue_userspace_packet(struct net *, int dp_ifindex,
-                                 struct sk_buff *,
+static int queue_userspace_packet(struct datapath *dp, struct sk_buff *,
                                  const struct dp_upcall_info *);
 
 /* Must be called with rcu_read_lock or ovs_mutex. */
@@ -279,7 +278,6 @@ int ovs_dp_upcall(struct datapath *dp, struct sk_buff *skb,
                  const struct dp_upcall_info *upcall_info)
 {
        struct dp_stats_percpu *stats;
-       int dp_ifindex;
        int err;
 
        if (upcall_info->portid == 0) {
@@ -287,16 +285,10 @@ int ovs_dp_upcall(struct datapath *dp, struct sk_buff *skb,
                goto err;
        }
 
-       dp_ifindex = get_dpifindex(dp);
-       if (!dp_ifindex) {
-               err = -ENODEV;
-               goto err;
-       }
-
        if (!skb_is_gso(skb))
-               err = queue_userspace_packet(ovs_dp_get_net(dp), dp_ifindex, skb, upcall_info);
+               err = queue_userspace_packet(dp, skb, upcall_info);
        else
-               err = queue_gso_packets(ovs_dp_get_net(dp), dp_ifindex, skb, upcall_info);
+               err = queue_gso_packets(dp, skb, upcall_info);
        if (err)
                goto err;
 
@@ -312,8 +304,7 @@ err:
        return err;
 }
 
-static int queue_gso_packets(struct net *net, int dp_ifindex,
-                            struct sk_buff *skb,
+static int queue_gso_packets(struct datapath *dp, struct sk_buff *skb,
                             const struct dp_upcall_info *upcall_info)
 {
        unsigned short gso_type = skb_shinfo(skb)->gso_type;
@@ -322,14 +313,14 @@ static int queue_gso_packets(struct net *net, int dp_ifindex,
        struct sk_buff *segs, *nskb;
        int err;
 
-       segs = __skb_gso_segment(skb, NETIF_F_SG | NETIF_F_HW_CSUM, false);
+       segs = __skb_gso_segment(skb, NETIF_F_SG, false);
        if (IS_ERR(segs))
                return PTR_ERR(segs);
 
        /* Queue all of the segments. */
        skb = segs;
        do {
-               err = queue_userspace_packet(net, dp_ifindex, skb, upcall_info);
+               err = queue_userspace_packet(dp, skb, upcall_info);
                if (err)
                        break;
 
@@ -382,11 +373,11 @@ static size_t key_attr_size(void)
                + nla_total_size(28); /* OVS_KEY_ATTR_ND */
 }
 
-static size_t upcall_msg_size(const struct sk_buff *skb,
-                             const struct nlattr *userdata)
+static size_t upcall_msg_size(const struct nlattr *userdata,
+                             unsigned int hdrlen)
 {
        size_t size = NLMSG_ALIGN(sizeof(struct ovs_header))
-               + nla_total_size(skb->len) /* OVS_PACKET_ATTR_PACKET */
+               + nla_total_size(hdrlen) /* OVS_PACKET_ATTR_PACKET */
                + nla_total_size(key_attr_size()); /* OVS_PACKET_ATTR_KEY */
 
        /* OVS_PACKET_ATTR_USERDATA */
@@ -396,8 +387,7 @@ static size_t upcall_msg_size(const struct sk_buff *skb,
        return size;
 }
 
-static int queue_userspace_packet(struct net *net, int dp_ifindex,
-                                 struct sk_buff *skb,
+static int queue_userspace_packet(struct datapath *dp, struct sk_buff *skb,
                                  const struct dp_upcall_info *upcall_info)
 {
        struct ovs_header *upcall;
@@ -406,12 +396,17 @@ static int queue_userspace_packet(struct net *net, int dp_ifindex,
        struct nlattr *nla;
        struct genl_info info = {
 #if LINUX_VERSION_CODE >= KERNEL_VERSION(3,14,0)
-               .dst_sk = net->genl_sock,
+               .dst_sk = ovs_dp_get_net(dp)->genl_sock,
 #endif
                .snd_portid = upcall_info->portid,
        };
-       size_t len;
-       int err;
+       size_t len, plen;
+       unsigned int hlen;
+       int err, dp_ifindex;
+
+       dp_ifindex = get_dpifindex(dp);
+       if (!dp_ifindex)
+               return -ENODEV;
 
        if (vlan_tx_tag_present(skb)) {
                nskb = skb_clone(skb, GFP_ATOMIC);
@@ -432,7 +427,21 @@ static int queue_userspace_packet(struct net *net, int dp_ifindex,
                goto out;
        }
 
-       len = upcall_msg_size(skb, upcall_info->userdata);
+       /* Complete checksum if needed */
+       if (skb->ip_summed == CHECKSUM_PARTIAL &&
+           (err = skb_checksum_help(skb)))
+               goto out;
+
+       /* Older versions of OVS user space enforce alignment of the last
+        * Netlink attribute to NLA_ALIGNTO which would require extensive
+        * padding logic. Only perform zerocopy if padding is not required.
+        */
+       if (dp->user_features & OVS_DP_F_UNALIGNED)
+               hlen = skb_zerocopy_headlen(skb);
+       else
+               hlen = skb->len;
+
+       len = upcall_msg_size(upcall_info->userdata, hlen);
        user_skb = genlmsg_new_unicast(len, &info, GFP_ATOMIC);
        if (!user_skb) {
                err = -ENOMEM;
@@ -452,13 +461,24 @@ static int queue_userspace_packet(struct net *net, int dp_ifindex,
                          nla_len(upcall_info->userdata),
                          nla_data(upcall_info->userdata));
 
-       nla = __nla_reserve(user_skb, OVS_PACKET_ATTR_PACKET, skb->len);
+       /* Only reserve room for attribute header, packet data is added
+        * in skb_zerocopy() */
+       if (!(nla = nla_reserve(user_skb, OVS_PACKET_ATTR_PACKET, 0))) {
+               err = -ENOBUFS;
+               goto out;
+       }
+       nla->nla_len = nla_attr_size(skb->len);
+
+       skb_zerocopy(user_skb, skb, skb->len, hlen);
 
-       skb_copy_and_csum_dev(skb, nla_data(nla));
+       /* Pad OVS_PACKET_ATTR_PACKET if linear copy was performed */
+       if (!(dp->user_features & OVS_DP_F_UNALIGNED) &&
+           (plen = (ALIGN(user_skb->len, NLA_ALIGNTO) - user_skb->len)) > 0)
+               memset(skb_put(user_skb, plen), 0, plen);
 
-       genlmsg_end(user_skb, upcall);
-       err = genlmsg_unicast(net, user_skb, upcall_info->portid);
+       ((struct nlmsghdr *) user_skb->data)->nlmsg_len = user_skb->len;
 
+       err = genlmsg_unicast(ovs_dp_get_net(dp), user_skb, upcall_info->portid);
 out:
        kfree_skb(nskb);
        return err;
@@ -502,7 +522,7 @@ static int ovs_packet_cmd_execute(struct sk_buff *skb, struct genl_info *info)
                packet->protocol = htons(ETH_P_802_2);
 
        /* Build an sw_flow for sending this packet. */
-       flow = ovs_flow_alloc();
+       flow = ovs_flow_alloc(false);
        err = PTR_ERR(flow);
        if (IS_ERR(flow))
                goto err_kfree_skb;
@@ -637,7 +657,9 @@ static int ovs_flow_cmd_fill_info(struct sw_flow *flow, struct datapath *dp,
 {
        const int skb_orig_len = skb->len;
        struct nlattr *start;
-       struct sw_flow_stats flow_stats;
+       struct ovs_flow_stats stats;
+       __be16 tcp_flags;
+       unsigned long used;
        struct ovs_header *ovs_header;
        struct nlattr *nla;
        int err;
@@ -668,25 +690,17 @@ static int ovs_flow_cmd_fill_info(struct sw_flow *flow, struct datapath *dp,
 
        nla_nest_end(skb, nla);
 
-       ovs_flow_stats_get(flow, &flow_stats);
-       if (flow_stats.used &&
-           nla_put_u64(skb, OVS_FLOW_ATTR_USED, ovs_flow_used_time(flow_stats.used)))
+       ovs_flow_stats_get(flow, &stats, &used, &tcp_flags);
+       if (used &&
+           nla_put_u64(skb, OVS_FLOW_ATTR_USED, ovs_flow_used_time(used)))
                goto nla_put_failure;
 
-       if (flow_stats.packet_count) {
-               struct ovs_flow_stats stats = {
-                       .n_packets = flow_stats.packet_count,
-                       .n_bytes = flow_stats.byte_count,
-               };
-
-               if (nla_put(skb, OVS_FLOW_ATTR_STATS,
-                           sizeof(struct ovs_flow_stats), &stats))
-                       goto nla_put_failure;
-       }
+       if (stats.n_packets &&
+           nla_put(skb, OVS_FLOW_ATTR_STATS, sizeof(struct ovs_flow_stats), &stats))
+               goto nla_put_failure;
 
-       if ((u8)ntohs(flow_stats.tcp_flags) &&
-           nla_put_u8(skb, OVS_FLOW_ATTR_TCP_FLAGS,
-                      (u8)ntohs(flow_stats.tcp_flags)))
+       if ((u8)ntohs(tcp_flags) &&
+            nla_put_u8(skb, OVS_FLOW_ATTR_TCP_FLAGS, (u8)ntohs(tcp_flags)))
                goto nla_put_failure;
 
        /* If OVS_FLOW_ATTR_ACTIONS doesn't fit, skip dumping the actions if
@@ -766,6 +780,7 @@ static int ovs_flow_cmd_new_or_set(struct sk_buff *skb, struct genl_info *info)
        struct datapath *dp;
        struct sw_flow_actions *acts = NULL;
        struct sw_flow_match match;
+       bool exact_5tuple;
        int error;
 
        /* Extract key. */
@@ -774,7 +789,7 @@ static int ovs_flow_cmd_new_or_set(struct sk_buff *skb, struct genl_info *info)
                goto error;
 
        ovs_match_init(&match, &key, &mask);
-       error = ovs_nla_get_match(&match,
+       error = ovs_nla_get_match(&match, &exact_5tuple,
                                  a[OVS_FLOW_ATTR_KEY], a[OVS_FLOW_ATTR_MASK]);
        if (error)
                goto error;
@@ -813,7 +828,7 @@ static int ovs_flow_cmd_new_or_set(struct sk_buff *skb, struct genl_info *info)
                        goto err_unlock_ovs;
 
                /* Allocate flow. */
-               flow = ovs_flow_alloc();
+               flow = ovs_flow_alloc(!exact_5tuple);
                if (IS_ERR(flow)) {
                        error = PTR_ERR(flow);
                        goto err_unlock_ovs;
@@ -900,7 +915,7 @@ static int ovs_flow_cmd_get(struct sk_buff *skb, struct genl_info *info)
        }
 
        ovs_match_init(&match, &key, NULL);
-       err = ovs_nla_get_match(&match, a[OVS_FLOW_ATTR_KEY], NULL);
+       err = ovs_nla_get_match(&match, NULL, a[OVS_FLOW_ATTR_KEY], NULL);
        if (err)
                return err;
 
@@ -954,7 +969,7 @@ static int ovs_flow_cmd_del(struct sk_buff *skb, struct genl_info *info)
        }
 
        ovs_match_init(&match, &key, NULL);
-       err = ovs_nla_get_match(&match, a[OVS_FLOW_ATTR_KEY], NULL);
+       err = ovs_nla_get_match(&match, NULL, a[OVS_FLOW_ATTR_KEY], NULL);
        if (err)
                goto unlock;
 
@@ -1050,6 +1065,7 @@ static struct genl_ops dp_flow_genl_ops[] = {
 static const struct nla_policy datapath_policy[OVS_DP_ATTR_MAX + 1] = {
        [OVS_DP_ATTR_NAME] = { .type = NLA_NUL_STRING, .len = IFNAMSIZ - 1 },
        [OVS_DP_ATTR_UPCALL_PID] = { .type = NLA_U32 },
+       [OVS_DP_ATTR_USER_FEATURES] = { .type = NLA_U32 },
 };
 
 static struct genl_family dp_datapath_genl_family = {
@@ -1108,6 +1124,9 @@ static int ovs_dp_cmd_fill_info(struct datapath *dp, struct sk_buff *skb,
                        &dp_megaflow_stats))
                goto nla_put_failure;
 
+       if (nla_put_u32(skb, OVS_DP_ATTR_USER_FEATURES, dp->user_features))
+               goto nla_put_failure;
+
        return genlmsg_end(skb, ovs_header);
 
 nla_put_failure:
@@ -1154,6 +1173,24 @@ static struct datapath *lookup_datapath(struct net *net,
        return dp ? dp : ERR_PTR(-ENODEV);
 }
 
+static void ovs_dp_reset_user_features(struct sk_buff *skb, struct genl_info *info)
+{
+       struct datapath *dp;
+
+       dp = lookup_datapath(sock_net(skb->sk), info->userhdr, info->attrs);
+       if (!dp)
+               return;
+
+       WARN(dp->user_features, "Dropping previously announced user features\n");
+       dp->user_features = 0;
+}
+
+static void ovs_dp_change(struct datapath *dp, struct nlattr **a)
+{
+       if (a[OVS_DP_ATTR_USER_FEATURES])
+               dp->user_features = nla_get_u32(a[OVS_DP_ATTR_USER_FEATURES]);
+}
+
 static int ovs_dp_cmd_new(struct sk_buff *skb, struct genl_info *info)
 {
        struct nlattr **a = info->attrs;
@@ -1206,12 +1243,23 @@ static int ovs_dp_cmd_new(struct sk_buff *skb, struct genl_info *info)
        parms.port_no = OVSP_LOCAL;
        parms.upcall_portid = nla_get_u32(a[OVS_DP_ATTR_UPCALL_PID]);
 
+       ovs_dp_change(dp, a);
+
        vport = new_vport(&parms);
        if (IS_ERR(vport)) {
                err = PTR_ERR(vport);
                if (err == -EBUSY)
                        err = -EEXIST;
 
+               if (err == -EEXIST) {
+                       /* An outdated user space instance that does not understand
+                        * the concept of user_features has attempted to create a new
+                        * datapath and is likely to reuse it. Drop all user features.
+                        */
+                       if (info->genlhdr->version < OVS_DP_VER_FEATURES)
+                               ovs_dp_reset_user_features(skb, info);
+               }
+
                goto err_destroy_ports_array;
        }
 
@@ -1309,6 +1357,8 @@ static int ovs_dp_cmd_set(struct sk_buff *skb, struct genl_info *info)
        if (IS_ERR(dp))
                goto unlock;
 
+       ovs_dp_change(dp, info->attrs);
+
        reply = ovs_dp_cmd_build_info(dp, info, OVS_DP_CMD_NEW);
        if (IS_ERR(reply)) {
                err = PTR_ERR(reply);