WRR support for BPF load balancer.

Go side changes

1. Extend types.LBBackEnd, Service4Value, Service6Value to take Weight.
2. Generate ServiceRRSeq{}, a sequence of backends to pick based on weights.
3. Create cilium_lb4_rr_seq BPF map mapping master Service4Key to ServiceRRSeq
4. Create cilium_lb6_rr_seq BPF map mapping master Service6Key to ServiceRRSeq

BPF data path changes

1. split lb_select_slave into lb6_select_slave and lb4_select_slave
2. If master has slaves with weights configured do weight based lb
3. Else do hash based lb.

Weight based load balancing.

1. Compute hash like before
2. Lookup lb_sequence and index into (hash % sequence length) for slave idx.

Addressed @aanm, @daniel & @tgraf code review comments.

Signed-off-by: Madhu Challa madhu@cilium.io

api/v1/models/backend_address.go

@@ -20,6 +20,9 @@ type BackendAddress struct {
 
 	// Layer 4 port number
 	Port uint16 `json:"port,omitempty"`
+
+	// Weight for Round Robin
+	Weight uint16 `json:"weight,omitempty"`
 }
 
 // Validate validates this backend address

api/v1/openapi.yaml

@@ -992,6 +992,10 @@ definitions:
         description: Layer 4 port number
         type: integer
         format: uint16
+      weight:
+        description: Weight for Round Robin
+        type: integer
+        format: uint16
   Service:
     description: Collection of endpoints to be served
     type: object

api/v1/server/embedded_spec.go

@@ -851,6 +851,11 @@ func init() {
           "description": "Layer 4 port number",
           "type": "integer",
           "format": "uint16"
+        },
+        "weight": {
+          "description": "Weight for Round Robin",
+          "type": "integer",
+          "format": "uint16"
         }
       }
     },

bpf/bpf_lb.c

@@ -35,6 +35,7 @@
 
 #define DISABLE_LOOPBACK_LB
 
+#include <node_config.h>
 #include <netdev_config.h>
 
 #include <bpf/api.h>
@@ -95,7 +96,7 @@ static inline int handle_ipv6(struct __sk_buff *skb)
 		return TC_ACT_OK;
 	}
 
-	slave = lb_select_slave(skb, svc->count);
+	slave = lb6_select_slave(skb, &key, svc->count, svc->weight);
 	if (!(svc = lb6_lookup_slave(skb, &key, slave)))
 		return DROP_NO_SERVICE;
 
@@ -153,7 +154,7 @@ static inline int handle_ipv4(struct __sk_buff *skb)
 		return TC_ACT_OK;
 	}
 
-	slave = lb_select_slave(skb, svc->count);
+	slave = lb4_select_slave(skb, &key, svc->count, svc->weight);
 	if (!(svc = lb4_lookup_slave(skb, &key, slave)))
 		return DROP_NO_SERVICE;
 

bpf/lib/common.h

@@ -202,6 +202,7 @@ struct lb6_service {
 	__u16 port;
 	__u16 count;
 	__u16 rev_nat_index;
+	__u16 weight;
 } __attribute__((packed));
 
 struct lb6_reverse_nat {
@@ -220,13 +221,20 @@ struct lb4_service {
 	__u16 port;
 	__u16 count;
 	__u16 rev_nat_index;
+	__u16 weight;
 } __attribute__((packed));
 
 struct lb4_reverse_nat {
 	__be32 address;
 	__u16 port;
 } __attribute__((packed));
 
+// LB_RR_MAX_SEQ generated by daemon in node_config.h
+struct lb_sequence {
+	__u16 count;
+	__u16 idx[LB_RR_MAX_SEQ];
+};
+
 struct ct_state {
 	__u16 rev_nat_index;
 	__u16 loopback:1,

bpf/lib/dbg.h

@@ -56,6 +56,7 @@ enum {
 	DBG_LB4_LOOPBACK_SNAT,
 	DBG_LB4_LOOPBACK_SNAT_REV,
 	DBG_CT_LOOKUP4,
+	DBG_RR_SLAVE_SEL,
 };
 
 /* Capture types */

bpf/lib/lb.h

@@ -34,6 +34,7 @@
 
 /* FIXME: Make configurable */
 #define CILIUM_LB_MAP_MAX_ENTRIES	65536
+#define CILIUM_LB_MAP_MAX_FE		256
 
 struct bpf_elf_map __section_maps cilium_lb6_reverse_nat = {
 	.type		= BPF_MAP_TYPE_HASH,
@@ -51,6 +52,14 @@ struct bpf_elf_map __section_maps cilium_lb6_services = {
 	.max_elem	= CILIUM_LB_MAP_MAX_ENTRIES,
 };
 
+struct bpf_elf_map __section_maps cilium_lb6_rr_seq = {
+	.type           = BPF_MAP_TYPE_HASH,
+	.size_key       = sizeof(struct lb6_key),
+	.size_value     = sizeof(struct lb_sequence),
+	.pinning        = PIN_GLOBAL_NS,
+	.max_elem       = CILIUM_LB_MAP_MAX_FE,
+};
+
 struct bpf_elf_map __section_maps cilium_lb4_reverse_nat = {
 	.type		= BPF_MAP_TYPE_HASH,
 	.size_key	= sizeof(__u16),
@@ -67,16 +76,78 @@ struct bpf_elf_map __section_maps cilium_lb4_services = {
 	.max_elem	= CILIUM_LB_MAP_MAX_ENTRIES,
 };
 
+struct bpf_elf_map __section_maps cilium_lb4_rr_seq = {
+	.type           = BPF_MAP_TYPE_HASH,
+	.size_key       = sizeof(struct lb4_key),
+	.size_value     = sizeof(struct lb_sequence),
+	.pinning        = PIN_GLOBAL_NS,
+	.max_elem       = CILIUM_LB_MAP_MAX_FE,
+};
 #define REV_NAT_F_TUPLE_SADDR 1
+#ifdef LB_DEBUG
+#define cilium_trace_lb cilium_trace
+#else
+#define cilium_trace_lb(a, b, c, d)
+#endif
 
-static inline int lb_select_slave(struct __sk_buff *skb, __u16 count)
+static inline int lb_next_rr(struct __sk_buff *skb,
+			     struct lb_sequence *seq,
+			     __be16 hash)
+{
+	int slave = 0;
+	__u8 offset = hash % seq->count;
+
+	if (offset < LB_RR_MAX_SEQ) {
+		/* Slave 0 is reserved for the master slot */
+		slave = seq->idx[offset] + 1;
+		cilium_trace(skb, DBG_RR_SLAVE_SEL, hash, slave);
+	}
+
+	return slave;
+}
+
+static inline int lb6_select_slave(struct __sk_buff *skb,
+				   struct lb6_key *key,
+				   __u16 count, __u16 weight)
 {
 	__be16 hash = get_hash_recalc(skb);
-	int slave;
+	struct lb_sequence *seq;
+	int slave = 0;
 
-	/* Slave 0 is reserved for the master slot */
-	slave = (hash % count) + 1;
-	cilium_trace(skb, DBG_PKT_HASH, hash, slave);
+	if (weight) {
+		seq = map_lookup_elem(&cilium_lb6_rr_seq, key);
+		if (seq && seq->count != 0)
+			slave = lb_next_rr(skb, seq, hash);
+	}
+
+	if (slave == 0) {
+		/* Slave 0 is reserved for the master slot */
+		slave = (hash % count) + 1;
+		cilium_trace(skb, DBG_PKT_HASH, hash, slave);
+	}
+
+	return slave;
+}
+
+static inline int lb4_select_slave(struct __sk_buff *skb,
+				   struct lb4_key *key,
+				   __u16 count, __u16 weight)
+{
+	__be16 hash = get_hash_recalc(skb);
+	struct lb_sequence *seq;
+	int slave = 0;
+
+	if (weight) {
+		seq = map_lookup_elem(&cilium_lb4_rr_seq, key);
+		if (seq && seq->count != 0)
+			slave = lb_next_rr(skb, seq, hash);
+	}
+
+	if (slave == 0) {
+		/* Slave 0 is reserved for the master slot */
+		slave = (hash % count) + 1;
+		cilium_trace_lb(skb, DBG_PKT_HASH, hash, slave);
+	}
 
 	return slave;
 }
@@ -154,7 +225,7 @@ static inline int __inline__ __lb6_rev_nat(struct __sk_buff *skb, int l4_off,
 	__be32 sum;
 	int ret;
 
-	cilium_trace(skb, DBG_LB6_REVERSE_NAT, nat->address.p4, nat->port);
+	cilium_trace_lb(skb, DBG_LB6_REVERSE_NAT, nat->address.p4, nat->port);
 
 	if (nat->port) {
 		ret = reverse_map_l4_port(skb, tuple->nexthdr, nat->port, l4_off, csum_off);
@@ -200,7 +271,7 @@ static inline int __inline__ lb6_rev_nat(struct __sk_buff *skb, int l4_off,
 {
 	struct lb6_reverse_nat *nat;
 
-	cilium_trace(skb, DBG_LB6_REVERSE_NAT_LOOKUP, index, 0);
+	cilium_trace_lb(skb, DBG_LB6_REVERSE_NAT_LOOKUP, index, 0);
 	nat = map_lookup_elem(&cilium_lb6_reverse_nat, &index);
 	if (nat == NULL)
 		return 0;
@@ -244,7 +315,7 @@ static inline struct lb6_service *lb6_lookup_service(struct __sk_buff *skb,
 	if (key->dport) {
 		struct lb6_service *svc;
 
-		cilium_trace(skb, DBG_LB6_LOOKUP_MASTER, key->address.p4, key->dport);
+		cilium_trace_lb(skb, DBG_LB6_LOOKUP_MASTER, key->address.p4, key->dport);
 		svc = map_lookup_elem(&cilium_lb6_services, key);
 		if (svc && svc->count != 0)
 			return svc;
@@ -257,14 +328,14 @@ static inline struct lb6_service *lb6_lookup_service(struct __sk_buff *skb,
 	if (1) {
 		struct lb6_service *svc;
 
-		cilium_trace(skb, DBG_LB6_LOOKUP_MASTER, key->address.p4, key->dport);
+		cilium_trace_lb(skb, DBG_LB6_LOOKUP_MASTER, key->address.p4, key->dport);
 		svc = map_lookup_elem(&cilium_lb6_services, key);
 		if (svc && svc->count != 0)
 			return svc;
 	}
 #endif
 
-	cilium_trace(skb, DBG_LB6_LOOKUP_MASTER_FAIL, key->address.p2, key->address.p3);
+	cilium_trace_lb(skb, DBG_LB6_LOOKUP_MASTER_FAIL, key->address.p2, key->address.p3);
 	return NULL;
 }
 
@@ -274,10 +345,10 @@ static inline struct lb6_service *lb6_lookup_slave(struct __sk_buff *skb,
 	struct lb6_service *svc;
 
 	key->slave = slave;
-	cilium_trace(skb, DBG_LB6_LOOKUP_SLAVE, key->slave, key->dport);
+	cilium_trace_lb(skb, DBG_LB6_LOOKUP_SLAVE, key->slave, key->dport);
 	svc = map_lookup_elem(&cilium_lb6_services, key);
 	if (svc != NULL) {
-		cilium_trace(skb, DBG_LB6_LOOKUP_SLAVE_SUCCESS, svc->target.p4, svc->port);
+		cilium_trace_lb(skb, DBG_LB6_LOOKUP_SLAVE_SUCCESS, svc->target.p4, svc->port);
 		return svc;
 	}
 
@@ -319,7 +390,7 @@ static inline int __inline__ lb6_local(struct __sk_buff *skb, int l3_off, int l4
 {
 	__u16 slave;
 
-	slave = lb_select_slave(skb, svc->count);
+	slave = lb6_select_slave(skb, key, svc->count, svc->weight);
 	if (!(svc = lb6_lookup_slave(skb, key, slave)))
 		return DROP_NO_SERVICE;
 
@@ -341,7 +412,7 @@ static inline int __inline__ __lb4_rev_nat(struct __sk_buff *skb, int l3_off, in
 	__be32 old_sip, new_sip, sum = 0;
 	int ret;
 
-	cilium_trace(skb, DBG_LB4_REVERSE_NAT, nat->address, nat->port);
+	cilium_trace_lb(skb, DBG_LB4_REVERSE_NAT, nat->address, nat->port);
 
 	if (nat->port) {
 		ret = reverse_map_l4_port(skb, tuple->nexthdr, nat->port, l4_off, csum_off);
@@ -372,7 +443,7 @@ static inline int __inline__ __lb4_rev_nat(struct __sk_buff *skb, int l3_off, in
 		if (IS_ERR(ret))
 			return ret;
 
-		cilium_trace(skb, DBG_LB4_LOOPBACK_SNAT_REV, old_dip, old_sip);
+		cilium_trace_lb(skb, DBG_LB4_LOOPBACK_SNAT_REV, old_dip, old_sip);
 
 		ret = skb_store_bytes(skb, l3_off + offsetof(struct iphdr, daddr), &old_sip, 4, 0);
 		if (IS_ERR(ret))
@@ -416,7 +487,7 @@ static inline int __inline__ lb4_rev_nat(struct __sk_buff *skb, int l3_off, int
 {
 	struct lb4_reverse_nat *nat;
 
-	cilium_trace(skb, DBG_LB4_REVERSE_NAT_LOOKUP, ct_state->rev_nat_index, 0);
+	cilium_trace_lb(skb, DBG_LB4_REVERSE_NAT_LOOKUP, ct_state->rev_nat_index, 0);
 	nat = map_lookup_elem(&cilium_lb4_reverse_nat, &ct_state->rev_nat_index);
 	if (nat == NULL)
 		return 0;
@@ -459,7 +530,7 @@ static inline struct lb4_service *lb4_lookup_service(struct __sk_buff *skb,
 		struct lb4_service *svc;
 
 		/* FIXME: The verifier barks on these calls right now for some reason */
-		/* cilium_trace(skb, DBG_LB4_LOOKUP_MASTER, key->address, key->dport); */
+		/* cilium_trace_lb(skb, DBG_LB4_LOOKUP_MASTER, key->address, key->dport); */
 		svc = map_lookup_elem(&cilium_lb4_services, key);
 		if (svc && svc->count != 0)
 			return svc;
@@ -473,14 +544,14 @@ static inline struct lb4_service *lb4_lookup_service(struct __sk_buff *skb,
 		struct lb4_service *svc;
 
 		/* FIXME: The verifier barks on these calls right now for some reason */
-		/* cilium_trace(skb, DBG_LB4_LOOKUP_MASTER, key->address, key->dport); */
+		/* cilium_trace_lb(skb, DBG_LB4_LOOKUP_MASTER, key->address, key->dport); */
 		svc = map_lookup_elem(&cilium_lb4_services, key);
 		if (svc && svc->count != 0)
 			return svc;
 	}
 #endif
 
-	cilium_trace(skb, DBG_LB4_LOOKUP_MASTER_FAIL, 0, 0);
+	cilium_trace_lb(skb, DBG_LB4_LOOKUP_MASTER_FAIL, 0, 0);
 	return NULL;
 }
 
@@ -490,10 +561,10 @@ static inline struct lb4_service *lb4_lookup_slave(struct __sk_buff *skb,
 	struct lb4_service *svc;
 
 	key->slave = slave;
-	cilium_trace(skb, DBG_LB4_LOOKUP_SLAVE, key->slave, key->dport);
+	cilium_trace_lb(skb, DBG_LB4_LOOKUP_SLAVE, key->slave, key->dport);
 	svc = map_lookup_elem(&cilium_lb4_services, key);
 	if (svc != NULL) {
-		cilium_trace(skb, DBG_LB4_LOOKUP_SLAVE_SUCCESS, svc->target, svc->port);
+		cilium_trace_lb(skb, DBG_LB4_LOOKUP_SLAVE_SUCCESS, svc->target, svc->port);
 		return svc;
 	}
 
@@ -516,7 +587,7 @@ lb4_xlate(struct __sk_buff *skb, __be32 *new_daddr, __be32 *new_saddr,
 	sum = csum_diff(&key->address, 4, new_daddr, 4, 0);
 
 	if (new_saddr && *new_saddr) {
-		cilium_trace(skb, DBG_LB4_LOOPBACK_SNAT, *old_saddr, *new_saddr);
+		cilium_trace_lb(skb, DBG_LB4_LOOPBACK_SNAT, *old_saddr, *new_saddr);
 		ret = skb_store_bytes(skb, l3_off + offsetof(struct iphdr, saddr), new_saddr, 4, 0);
 		if (ret < 0)
 			return DROP_WRITE_ERROR;
@@ -555,7 +626,7 @@ static inline int __inline__ lb4_local(struct __sk_buff *skb, int l3_off, int l4
 	__be32 new_saddr = 0, new_daddr;
 	__u16 slave;
 
-	slave = lb_select_slave(skb, svc->count);
+	slave = lb4_select_slave(skb, key, svc->count, svc->weight);
 	if (!(svc = lb4_lookup_slave(skb, key, slave)))
 		return DROP_NO_SERVICE;
 

bpf/node_config.h

@@ -41,3 +41,4 @@
 #define ENABLE_ARP_RESPONDER
 #define NODE_MAC { .addr = { 0xde, 0xad, 0xbe, 0xef, 0xc0, 0xde } }
 #define ENABLE_IPV4
+#define LB_RR_MAX_SEQ 31

cilium/cmd/service_list.go

@@ -53,7 +53,12 @@ func listServices() {
 				fmt.Fprintf(os.Stderr, "error parsing backend %+v", be)
 				continue
 			}
-			str := fmt.Sprintf("%d => %s (%d)", i+1, beA.String(), svc.ID)
+			var str string
+			if be.Weight != 0 {
+				str = fmt.Sprintf("%d => %s (W: %d, ID: %d)", i+1, beA.String(), be.Weight, svc.ID)
+			} else {
+				str = fmt.Sprintf("%d => %s (%d)", i+1, beA.String(), svc.ID)
+			}
 			besWithID = append(besWithID, str)
 		}