Merge pull request #14553 from donaldsharp/zebra_weighted_ecmp

Zebra weighted ecmp

bgpd/bgp_zebra.c

@@ -1207,26 +1207,20 @@ static bool update_ipv6nh_for_route_install(int nh_othervrf, struct bgp *nh_bgp,
 }
 
 static bool bgp_zebra_use_nhop_weighted(struct bgp *bgp, struct attr *attr,
-					uint64_t tot_bw, uint32_t *nh_weight)
+					uint32_t *nh_weight)
 {
-	uint32_t bw;
-	uint64_t tmp;
-
-	bw = attr->link_bw;
 	/* zero link-bandwidth and link-bandwidth not present are treated
 	 * as the same situation.
 	 */
-	if (!bw) {
+	if (!attr->link_bw) {
 		/* the only situations should be if we're either told
 		 * to skip or use default weight.
 		 */
 		if (bgp->lb_handling == BGP_LINK_BW_SKIP_MISSING)
 			return false;
 		*nh_weight = BGP_ZEBRA_DEFAULT_NHOP_WEIGHT;
-	} else {
-		tmp = (uint64_t)bw * 100;
-		*nh_weight = ((uint32_t)(tmp / tot_bw));
-	}
+	} else
+		*nh_weight = attr->link_bw;
 
 	return true;
 }
@@ -1255,7 +1249,6 @@ void bgp_zebra_announce(struct bgp_dest *dest, const struct prefix *p,
 	int nh_othervrf = 0;
 	bool nh_updated = false;
 	bool do_wt_ecmp;
-	uint64_t cum_bw = 0;
 	uint32_t nhg_id = 0;
 	bool is_add;
 	uint32_t ttl = 0;
@@ -1339,8 +1332,6 @@ void bgp_zebra_announce(struct bgp_dest *dest, const struct prefix *p,
 
 	/* Determine if we're doing weighted ECMP or not */
 	do_wt_ecmp = bgp_path_info_mpath_chkwtd(bgp, info);
-	if (do_wt_ecmp)
-		cum_bw = bgp_path_info_mpath_cumbw(info);
 
 	/* EVPN MAC-IP routes are installed with a L3 NHG id */
 	if (bgp_evpn_path_es_use_nhg(bgp, info, &nhg_id)) {
@@ -1382,7 +1373,7 @@ void bgp_zebra_announce(struct bgp_dest *dest, const struct prefix *p,
 		 */
 		if (do_wt_ecmp) {
 			if (!bgp_zebra_use_nhop_weighted(bgp, mpinfo->attr,
-							 cum_bw, &nh_weight))
+							 &nh_weight))
 				continue;
 		}
 		api_nh = &api.nexthops[valid_nh_count];

tests/topotests/bgp_link_bw_ip/r1/ip-route-1.json

@@ -7,12 +7,12 @@
         {
           "fib":true,
           "ip":"11.1.1.6",
-          "weight":25
+          "weight":85
         },
         {
           "fib":true,
           "ip":"11.1.1.2",
-          "weight":75
+          "weight":255
         }
       ]
     }

tests/topotests/bgp_link_bw_ip/r1/ip-route-2.json

@@ -7,12 +7,12 @@
         {
           "fib":true,
           "ip":"11.1.1.6",
-          "weight":33
+          "weight":127
         },
         {
           "fib":true,
           "ip":"11.1.1.2",
-          "weight":66
+          "weight":255
         }
       ]
     }

tests/topotests/bgp_link_bw_ip/r1/ip-route-3.json

@@ -7,12 +7,12 @@
         {
           "fib":true,
           "ip":"11.1.1.6",
-          "weight":33
+          "weight":127
         },
         {
           "fib":true,
           "ip":"11.1.1.2",
-          "weight":66
+          "weight":255
         }
       ]
     }

tests/topotests/bgp_link_bw_ip/r1/ip-route-6.json

@@ -7,7 +7,7 @@
         {
           "fib":true,
           "ip":"11.1.1.2",
-          "weight":100
+          "weight":255
         }
       ]
     }

tests/topotests/bgp_link_bw_ip/r1/ip-route-7.json

@@ -7,7 +7,7 @@
         {
           "fib":true,
           "ip":"11.1.1.2",
-          "weight":100
+          "weight":255
         }
       ]
     }

tests/topotests/bgp_link_bw_ip/r1/ip-route-8.json

@@ -12,7 +12,7 @@
         {
           "fib":true,
           "ip":"11.1.1.2",
-          "weight":100
+          "weight":255
         }
       ]
     }

tests/topotests/bgp_link_bw_ip/r1/ip-route-9.json

@@ -12,7 +12,7 @@
         {
           "fib":true,
           "ip":"11.1.1.2",
-          "weight":100
+          "weight":255
         }
       ]
     }

tests/topotests/bgp_link_bw_ip/r2/ip-route-2.json

@@ -7,12 +7,12 @@
         {
           "fib":true,
           "ip":"11.1.2.6",
-          "weight":33
+          "weight":127
         },
         {
           "fib":true,
           "ip":"11.1.2.2",
-          "weight":66
+          "weight":255
         }
       ]
     }

zebra/zapi_msg.c

@@ -1679,10 +1679,14 @@ static bool zapi_read_nexthops(struct zserv *client, struct prefix *p,
 			       struct nexthop_group **png,
 			       struct nhg_backup_info **pbnhg)
 {
+	struct zapi_nexthop *znh;
 	struct nexthop_group *ng = NULL;
 	struct nhg_backup_info *bnhg = NULL;
 	uint16_t i;
 	struct nexthop *last_nh = NULL;
+	bool same_weight = true;
+	uint64_t max_weight = 0;
+	uint64_t tmp;
 
 	assert(!(png && pbnhg));
 
@@ -1697,6 +1701,41 @@ static bool zapi_read_nexthops(struct zserv *client, struct prefix *p,
 		bnhg = zebra_nhg_backup_alloc();
 	}
 
+	for (i = 0; i < nexthop_num; i++) {
+		znh = &nhops[i];
+
+		if (max_weight < znh->weight) {
+			if (i != 0 || znh->weight != 1)
+				same_weight = false;
+
+			max_weight = znh->weight;
+		}
+	}
+
+	/*
+	 * Let's convert the weights to a scaled value
+	 * between 1 and zrouter.nexthop_weight_scale_value
+	 * This is a simple application of a ratio:
+	 * scaled_weight/zrouter.nexthop_weight_scale_value = 
+         * weight/max_weight
+	 * This translates to:
+	 * scaled_weight = weight * zrouter.nexthop_weight_scale_value
+	 *                 -------------------------------------------
+	 *                           max_weight
+	 *
+	 * This same formula is applied to both the nexthops
+	 * and the backup nexthops
+	 */
+	if (!same_weight) {
+		for (i = 0; i < nexthop_num; i++) {
+			znh = &nhops[i];
+
+			tmp = (uint64_t)znh->weight *
+				zrouter.nexthop_weight_scale_value;
+			znh->weight = MAX(1, ((uint32_t)(tmp / max_weight)));
+		}
+	}
+
 	/*
 	 * TBD should _all_ of the nexthop add operations use
 	 * api_nh->vrf_id instead of re->vrf_id ? I only changed

zebra/zebra_router.c

@@ -323,6 +323,8 @@ void zebra_router_init(bool asic_offload, bool notify_on_ack,
 #endif
 	zrouter.asic_notification_nexthop_control = false;
 
+	zrouter.nexthop_weight_scale_value = 255;
+
 #ifdef HAVE_SCRIPTING
 	zebra_script_init();
 #endif

zebra/zebra_router.h

@@ -231,6 +231,8 @@ struct zebra_router {
 	bool allow_delete;
 
 	uint8_t protodown_r_bit;
+
+	uint64_t nexthop_weight_scale_value;
 };
 
 #define GRACEFUL_RESTART_TIME 60