[FPM]: Fix dplane_fpm_sonic for multiple SRv6 nexthops

Currently, there is a bug in the dplane_fpm_sonic to handle ECMP routes.
When FRR has an ECMP route with multiple SRv6 nexthops, the
dplane_fpm_sonic encodes the route with only the first SRv6 nextop in
the Netlink message, ignoring the other nexthops.

As a result, the Netlink message sent to fpmsyncd leads to the
installation of the route using only the first SRv6 nexthop, breaking
the EMCP nature of SRv6 routes.

This pull request fixes the issue in the dplane_fpm_sonic to handle all
SRv6 nexthops in the route.

Signed-off-by: Carmine Scarpitta <[email protected]>

Author: cscarpitta

src/sonic-frr/dplane_fpm_sonic/dplane_fpm_sonic.c

@@ -1051,7 +1051,42 @@ static struct zebra_vrf *vrf_lookup_by_table_id(uint32_t table_id)
  	}
 
  	return NULL;
- }
+}
+
+static bool has_srv6_nexthop(struct zebra_dplane_ctx *ctx)
+{
+	struct nexthop *nexthop;
+
+	for (ALL_NEXTHOPS_PTR(dplane_ctx_get_ng(ctx), nexthop))
+		if (nexthop->nh_srv6)
+			return true;
+
+	return false;
+}
+
+static bool has_srv6_sidlist_nexthop(struct zebra_dplane_ctx *ctx)
+{
+	struct nexthop *nexthop;
+
+	for (ALL_NEXTHOPS_PTR(dplane_ctx_get_ng(ctx), nexthop))
+		if (nexthop->nh_srv6 && nexthop->nh_srv6->seg6_segs &&
+		    !sid_zero(nexthop->nh_srv6->seg6_segs))
+			return true;
+
+	return false;
+}
+
+static bool has_srv6_localsid_nexthop(struct zebra_dplane_ctx *ctx)
+{
+	struct nexthop *nexthop;
+
+	for (ALL_NEXTHOPS_PTR(dplane_ctx_get_ng(ctx), nexthop))
+		if (nexthop->nh_srv6 &&
+		    nexthop->nh_srv6->seg6local_action != ZEBRA_SEG6_LOCAL_ACTION_UNSPEC)
+			return true;
+
+	return false;
+}
 
 /**
  * Resets the SRv6 routes FPM flags so we send all SRv6 routes again.
@@ -1474,6 +1509,70 @@ static ssize_t netlink_vpn_route_msg_encode(int cmd,
 	return NLMSG_ALIGN(req->n.nlmsg_len);
 }
 
+static bool netlink_srv6_vpn_route_msg_encode_multipath(int cmd, struct zebra_dplane_ctx *ctx,
+							uint8_t *data, size_t datalen,
+							const struct nexthop *nexthop,
+							struct nlmsghdr *nlmsg, size_t req_size,
+							bool fpm, bool force_nhg)
+{
+	struct rtattr *nest;
+	struct rtnexthop *rtnh;
+	struct interface *ifp;
+	struct in6_addr encap_src_addr = {};
+	struct connected *connected;
+	struct vrf *vrf;
+	struct prefix *cp;
+
+	rtnh = nl_attr_rtnh(nlmsg, req_size);
+	if (rtnh == NULL)
+		return false;
+
+	if (!nl_attr_put16(nlmsg, req_size, RTA_ENCAP_TYPE, FPM_ROUTE_ENCAP_SRV6))
+		return false;
+
+	nest = nl_attr_nest(nlmsg, req_size, RTA_ENCAP);
+	if (!nest)
+		return false;
+
+	/*
+     * by default, we use the loopback address as encap source address,
+     * if it is valid
+     */
+	ifp = if_lookup_by_name("lo", VRF_DEFAULT);
+
+	if (ifp) {
+		vrf = vrf_lookup_by_name(VRF_DEFAULT_NAME);
+		if (!vrf)
+			return false;
+
+		FOR_ALL_INTERFACES (vrf, ifp) {
+			frr_each (if_connected, ifp->connected, connected) {
+				cp = connected->address;
+				if (cp->family == AF_INET6 &&
+				    !IN6_IS_ADDR_LOOPBACK(&cp->u.prefix6) &&
+				    !IN6_IS_ADDR_LINKLOCAL(&cp->u.prefix6)) {
+					encap_src_addr = cp->u.prefix6;
+					break;
+				}
+			}
+		}
+	}
+
+	if (!nl_attr_put(nlmsg, req_size, FPM_ROUTE_ENCAP_SRV6_ENCAP_SRC_ADDR, &encap_src_addr,
+			 IPV6_MAX_BYTELEN))
+		return false;
+
+	if (!nl_attr_put(nlmsg, req_size, FPM_ROUTE_ENCAP_SRV6_VPN_SID,
+			 &nexthop->nh_srv6->seg6_segs->seg[0], IPV6_MAX_BYTELEN))
+		return false;
+
+	nl_attr_nest_end(nlmsg, nest);
+
+	nl_attr_rtnh_end(nlmsg, rtnh);
+
+	return true;
+}
+
 /*
  * SRv6 VPN route change via netlink interface, using a dataplane context object
  *
@@ -1497,17 +1596,14 @@ static ssize_t netlink_srv6_vpn_route_msg_encode(int cmd,
 	struct connected *connected;
 	struct vrf *vrf;
 	struct prefix *cp;
+	unsigned int nexthop_num;
 
 	struct {
 		struct nlmsghdr n;
 		struct rtmsg r;
 		char buf[];
 	} *req = (void *)data;
 
-	nexthop = dplane_ctx_get_ng(ctx)->nexthop;
-	if (!nexthop || !nexthop->nh_srv6 || sid_zero((const struct seg6_seg_stack *)nexthop->nh_srv6->seg6_segs))
-		return -1;
-
 	p = dplane_ctx_get_dest(ctx);
 
 	if (datalen < sizeof(*req))
@@ -1572,6 +1668,79 @@ static ssize_t netlink_srv6_vpn_route_msg_encode(int cmd,
 			nl_msg_type_to_str(cmd), p, dplane_ctx_get_vrf(ctx),
 			table_id);
 
+	/*
+	 * Counts the number of nexthops to determine if the route is singlepath
+	 * (single nexthop) or multipath (multiple nexthops)
+	 */
+	nexthop_num = 0;
+	for (ALL_NEXTHOPS_PTR(dplane_ctx_get_ng(ctx), nexthop)) {
+		if (CHECK_FLAG(nexthop->flags, NEXTHOP_FLAG_RECURSIVE))
+			continue;
+		if (!NEXTHOP_IS_ACTIVE(nexthop->flags))
+			continue;
+
+		nexthop_num++;
+	}
+
+	/* Multipath case */
+	if (nexthop_num > 1) {
+		struct rtattr *nest;
+
+		nest = nl_attr_nest(&req->n, datalen, RTA_MULTIPATH);
+		if (nest == NULL)
+			return 0;
+
+		nexthop_num = 0;
+		for (ALL_NEXTHOPS_PTR(dplane_ctx_get_ng(ctx), nexthop)) {
+			if (CHECK_FLAG(nexthop->flags, NEXTHOP_FLAG_RECURSIVE))
+				continue;
+
+			/* Skip non-SRv6 nexthops */
+			if (!nexthop->nh_srv6 || sid_zero(nexthop->nh_srv6->seg6_segs))
+				continue;
+
+			if (NEXTHOP_IS_ACTIVE(nexthop->flags)) {
+				nexthop_num++;
+
+				if (!netlink_srv6_vpn_route_msg_encode_multipath(cmd, ctx, data,
+										 datalen, nexthop,
+										 &req->n, datalen,
+										 fpm, force_nhg))
+					return 0;
+			}
+		}
+
+		nl_attr_nest_end(&req->n, nest);
+
+		if (nexthop_num == 0) {
+			if (IS_ZEBRA_DEBUG_FPM)
+				zlog_debug("%s: No useful nexthop.", __func__);
+		}
+
+		return NLMSG_ALIGN(req->n.nlmsg_len);
+	}
+
+	/* Singlepath case */
+	nexthop_num = 0;
+	for (ALL_NEXTHOPS_PTR(dplane_ctx_get_ng(ctx), nexthop)) {
+		if (CHECK_FLAG(nexthop->flags, NEXTHOP_FLAG_RECURSIVE))
+			continue;
+		if (!NEXTHOP_IS_ACTIVE(nexthop->flags))
+			continue;
+		if (!nexthop->nh_srv6 || sid_zero(nexthop->nh_srv6->seg6_segs))
+			continue;
+
+		nexthop_num++;
+		break;
+	}
+
+	if (nexthop_num == 0) {
+		if (IS_ZEBRA_DEBUG_FPM)
+			zlog_debug("%s: No useful nexthop.", __func__);
+
+		return NLMSG_ALIGN(req->n.nlmsg_len);
+	}
+
 	if (!nl_attr_put16(&req->n, datalen, RTA_ENCAP_TYPE,
 				FPM_ROUTE_ENCAP_SRV6))
 		return false;
@@ -1625,20 +1794,16 @@ static ssize_t netlink_srv6_msg_encode(int cmd,
 					   uint8_t *data, size_t datalen,
 					   bool fpm, bool force_nhg)
 {
-	struct nexthop *nexthop = NULL;
-
 	struct {
 		struct nlmsghdr n;
 		struct rtmsg r;
 		char buf[];
 	} *req = (void *)data;
 
-	nexthop = dplane_ctx_get_ng(ctx)->nexthop;
-	if (!nexthop || !nexthop->nh_srv6)
+	if (!has_srv6_nexthop(ctx))
 		return -1;
 
-	if (nexthop->nh_srv6->seg6local_action !=
-			ZEBRA_SEG6_LOCAL_ACTION_UNSPEC) {
+	if (has_srv6_localsid_nexthop(ctx)) {
 		if (cmd == RTM_NEWROUTE)
 			cmd = RTM_NEWSRV6LOCALSID;
 		else if (cmd == RTM_DELROUTE)
@@ -1647,7 +1812,7 @@ static ssize_t netlink_srv6_msg_encode(int cmd,
 		if (!netlink_srv6_localsid_msg_encode(
 				cmd, ctx, data, datalen, fpm, force_nhg))
 			return 0;
-	} else if (!sid_zero(nexthop->nh_srv6->seg6_segs)) {
+	} else if (has_srv6_sidlist_nexthop(ctx)) {
 		if(force_nhg){
 			if (!netlink_vpn_route_msg_encode(
 				cmd, ctx, data, datalen, force_nhg))
@@ -2280,7 +2445,6 @@ static int fpm_nl_enqueue(struct fpm_nl_ctx *fnc, struct zebra_dplane_ctx *ctx)
 	ssize_t rv;
 	uint64_t obytes, obytes_peak;
 	enum dplane_op_e op = dplane_ctx_get_op(ctx);
-	struct nexthop *nexthop;
 
 	/*
 	 * If we were configured to not use next hop groups, then quit as soon
@@ -2321,8 +2485,7 @@ static int fpm_nl_enqueue(struct fpm_nl_ctx *fnc, struct zebra_dplane_ctx *ctx)
 	switch (op) {
 	case DPLANE_OP_ROUTE_UPDATE:
 	case DPLANE_OP_ROUTE_DELETE:
-		nexthop = dplane_ctx_get_ng(ctx)->nexthop;
-		if (nexthop && nexthop->nh_srv6) {
+		if (has_srv6_nexthop(ctx)) {
 			rv = netlink_srv6_msg_encode(RTM_DELROUTE, ctx,
 								nl_buf, sizeof(nl_buf),
 								true, fnc->use_nhg);
@@ -2352,8 +2515,7 @@ static int fpm_nl_enqueue(struct fpm_nl_ctx *fnc, struct zebra_dplane_ctx *ctx)
 
 		/* FALL THROUGH */
 	case DPLANE_OP_ROUTE_INSTALL:
-		nexthop = dplane_ctx_get_ng(ctx)->nexthop;
-		if (nexthop && nexthop->nh_srv6) {
+		if (has_srv6_nexthop(ctx)) {
 			rv = netlink_srv6_msg_encode(
 				RTM_NEWROUTE, ctx, &nl_buf[nl_buf_len],
 				sizeof(nl_buf) - nl_buf_len, true, fnc->use_nhg);