netdev-offload-tc: Add don't fragment support to encap action.

This patch adds "Don't Fragment" (DF) flag support to the encap action,
if supported by the kernel. If the kernel does not support this,
it falls back to the previous behavior of ignoring the DF request.

Acked-by: Roi Dayan <roid@nvidia.com>
Acked-by: Simon Horman <horms@ovn.org>
Signed-off-by: Eelco Chaudron <echaudro@redhat.com>

NEWS

@@ -7,6 +7,8 @@ Post-v3.4.0
        the kernel datapath.
    - Linux TC offload:
      * Add support for matching tunnel flags if the kernel supports it.
+     * Add support for the "Don't Fragment" (DF) flag in the encap action,
+       if the kernel supports it.
    - Python:
      * Added tool called "ovs-flowviz" capable of parsing OpenFlow
        and datapath flow dumps and displaying them in several different

acinclude.m4

@@ -193,10 +193,10 @@ AC_DEFUN([OVS_CHECK_LINUX_TC], [
 
   AC_COMPILE_IFELSE([
     AC_LANG_PROGRAM([#include <linux/tc_act/tc_tunnel_key.h>], [
-        int x = TCA_TUNNEL_KEY_ENC_OPTS_VXLAN;
+        int x = TCA_TUNNEL_KEY_NO_FRAG;
     ])],
-    [AC_DEFINE([HAVE_TCA_TUNNEL_KEY_ENC_OPTS_VXLAN], [1],
+    [AC_DEFINE([HAVE_TCA_TUNNEL_KEY_NO_FRAG], [1],
-               [Define to 1 if TCA_TUNNEL_KEY_ENC_OPTS_VXLAN is available.])])
+               [Define to 1 if TCA_TUNNEL_KEY_NO_FRAG is available.])])
 
   AC_COMPILE_IFELSE([
     AC_LANG_PROGRAM([#include <linux/tc_act/tc_pedit.h>], [

include/linux/tc_act/tc_tunnel_key.h

@@ -1,7 +1,7 @@
 #ifndef __LINUX_TC_ACT_TC_TUNNEL_KEY_WRAPPER_H
 #define __LINUX_TC_ACT_TC_TUNNEL_KEY_WRAPPER_H 1
 
-#if defined(__KERNEL__) || defined(HAVE_TCA_TUNNEL_KEY_ENC_OPTS_VXLAN)
+#if defined(__KERNEL__) || defined(HAVE_TCA_TUNNEL_KEY_NO_FRAG)
 #include_next <linux/tc_act/tc_tunnel_key.h>
 #else
 
@@ -42,6 +42,7 @@ enum {
                                      */
 	TCA_TUNNEL_KEY_ENC_TOS,		    /* u8 */
 	TCA_TUNNEL_KEY_ENC_TTL,		    /* u8 */
+	TCA_TUNNEL_KEY_NO_FRAG,		    /* flag */
 	__TCA_TUNNEL_KEY_MAX,
 };
 

lib/netdev-offload-tc.c

@@ -919,6 +919,9 @@ parse_tc_flower_to_actions__(struct tc_flower *flower, struct ofpbuf *buf,
             if (!action->encap.no_csum) {
                 nl_msg_put_flag(buf, OVS_TUNNEL_KEY_ATTR_CSUM);
             }
+            if (action->encap.dont_fragment) {
+                nl_msg_put_flag(buf, OVS_TUNNEL_KEY_ATTR_DONT_FRAGMENT);
+            }
             ret = parse_tc_flower_vxlan_tun_opts(action, buf);
             if (ret) {
                 return ret;
@@ -1641,11 +1644,14 @@ parse_put_flow_set_action(struct tc_flower *flower, struct tc_action *action,
         }
         break;
         case OVS_TUNNEL_KEY_ATTR_DONT_FRAGMENT: {
-            /* XXX: This is wrong!  We're ignoring the DF flag configuration
+            if (enc_flags_support) {
-             * requested by the user.  However, TC for now has no way to pass
+                action->encap.dont_fragment = true;
-             * that flag and it is set by default, meaning tunnel offloading
+            } else {
-             * will not work if 'options:df_default=false' is not set.
+                /* For kernels not supporting the DF flag, we ignoring the
-             * Keeping incorrect behavior for now. */
+                 * configuration requested by the user.  This to keep the old,
+                 * incorrect behaviour, and allow tunnels to be offloaded by
+                 * TC with these kernels. */
+            }
         }
         break;
         case OVS_TUNNEL_KEY_ATTR_CSUM: {

lib/tc.c

@@ -1229,6 +1229,7 @@ static const struct nl_policy tunnel_key_policy[] = {
     [TCA_TUNNEL_KEY_ENC_TTL] = { .type = NL_A_U8, .optional = true, },
     [TCA_TUNNEL_KEY_ENC_OPTS] = { .type = NL_A_NESTED, .optional = true, },
     [TCA_TUNNEL_KEY_NO_CSUM] = { .type = NL_A_U8, .optional = true, },
+    [TCA_TUNNEL_KEY_NO_FRAG] = { .type = NL_A_FLAG, .optional = true, },
 };
 
 static int
@@ -1394,6 +1395,7 @@ nl_parse_act_tunnel_key(struct nlattr *options, struct tc_flower *flower)
         struct nlattr *ttl = tun_attrs[TCA_TUNNEL_KEY_ENC_TTL];
         struct nlattr *tun_opt = tun_attrs[TCA_TUNNEL_KEY_ENC_OPTS];
         struct nlattr *no_csum = tun_attrs[TCA_TUNNEL_KEY_NO_CSUM];
+        struct nlattr *no_frag = tun_attrs[TCA_TUNNEL_KEY_NO_FRAG];
 
         action = &flower->actions[flower->action_count++];
         action->type = TC_ACT_ENCAP;
@@ -1411,6 +1413,7 @@ nl_parse_act_tunnel_key(struct nlattr *options, struct tc_flower *flower)
         action->encap.tos = tos ? nl_attr_get_u8(tos) : 0;
         action->encap.ttl = ttl ? nl_attr_get_u8(ttl) : 0;
         action->encap.no_csum = no_csum ? nl_attr_get_u8(no_csum) : 0;
+        action->encap.dont_fragment = no_frag ? true : false;
 
         err = nl_parse_act_tunnel_opts(tun_opt, action);
         if (err) {
@@ -2747,6 +2750,9 @@ nl_msg_put_act_tunnel_key_set(struct ofpbuf *request,
             nl_msg_put_be16(request, TCA_TUNNEL_KEY_ENC_DST_PORT,
                             encap->tp_dst);
         }
+        if (encap->dont_fragment) {
+            nl_msg_put_flag(request, TCA_TUNNEL_KEY_NO_FRAG);
+        }
         nl_msg_put_act_tunnel_vxlan_opts(request, encap);
         nl_msg_put_act_tunnel_geneve_option(request, &encap->data);
         nl_msg_put_u8(request, TCA_TUNNEL_KEY_NO_CSUM, encap->no_csum);

lib/tc.h

@@ -228,6 +228,7 @@ struct tc_action_encap {
     uint8_t tos;
     uint8_t ttl;
     uint8_t no_csum;
+    bool dont_fragment;
     struct {
         ovs_be32 ipv4_src;
         ovs_be32 ipv4_dst;