Add L2 bridge interface support

docs/meson.build

@@ -74,9 +74,10 @@ custom_target(
 # Individual command man pages
 # The list is hardcoded since we can't run grcli during meson configuration.
 grcli_commands = [
-  'address', 'affinity', 'conntrack', 'dnat44', 'events', 'graph', 'interface',
-  'logging', 'nexthop', 'ping', 'ping6', 'route', 'router-advert', 'snat44',
-  'stats', 'trace', 'traceroute', 'traceroute6', 'tunsrc',
+  'address', 'affinity', 'conntrack', 'dnat44', 'events', 'fdb', 'graph',
+  'interface', 'logging', 'nexthop', 'ping', 'ping6', 'route',
+  'router-advert', 'snat44', 'stats', 'trace', 'traceroute', 'traceroute6',
+  'tunsrc',
 ]
 
 foreach cmd : grcli_commands

modules/infra/control/control_queue.c → main/control_queue.c

@@ -2,14 +2,12 @@
 // Copyright (c) 2024 Christophe Fontaine
 
 #include <gr_control_queue.h>
-#include <gr_event.h>
-#include <gr_iface.h>
 #include <gr_log.h>
 #include <gr_macro.h>
 #include <gr_module.h>
-#include <gr_nexthop.h>
 
 #include <event2/event.h>
+#include <rte_lcore.h>
 #include <rte_ring.h>
 
 #include <pthread.h>
@@ -88,23 +86,11 @@ static void *sem_wait_to_event(void *) {
 	return NULL;
 }
 
-// When interfaces or nexthops are deleted, drain the control queue
-// to free any packets that reference the deleted object. This prevents
-// callbacks from being invoked with dangling pointers.
-static void event_handler(uint32_t event, const void *obj) {
+void control_queue_drain(uint32_t event, const void *obj) {
 	struct control_queue_drain drain = {event, obj};
 	control_queue_poll(0, 0, &drain);
 }
 
-static struct gr_event_subscription event_sub = {
-	.callback = event_handler,
-	.ev_count = 2,
-	.ev_types = {
-		GR_EVENT_IFACE_REMOVE,
-		GR_EVENT_NEXTHOP_DELETE,
-	},
-};
-
 static void control_queue_init(struct event_base *ev_base) {
 	atomic_init(&thread_shutdown, false);
 
@@ -146,5 +132,4 @@ static struct gr_module module = {
 
 RTE_INIT(control_queue_module_init) {
 	gr_register_module(&module);
-	gr_event_subscribe(&event_sub);
 }

main/event.c

@@ -4,10 +4,13 @@
 #include "api.h"
 
 #include <gr_api.h>
+#include <gr_control_queue.h>
 #include <gr_event.h>
 #include <gr_log.h>
 #include <gr_queue.h>
 
+#include <rte_lcore.h>
+
 #include <string.h>
 
 STAILQ_HEAD(subscribers, gr_event_subscription);
@@ -17,7 +20,7 @@ void gr_event_subscribe(struct gr_event_subscription *sub) {
 	STAILQ_INSERT_TAIL(&subscribers, sub, next);
 }
 
-void gr_event_push(uint32_t ev_type, const void *obj) {
+static void notify_subscribers(void *obj, uintptr_t ev_type, const struct control_queue_drain *) {
 	const struct gr_event_subscription *sub;
 
 	STAILQ_FOREACH (sub, &subscribers, next) {
@@ -28,9 +31,24 @@ void gr_event_push(uint32_t ev_type, const void *obj) {
 			}
 		}
 	}
+
 	api_send_notifications(ev_type, obj);
 }
 
+void gr_event_push(uint32_t ev_type, const void *obj) {
+	if (rte_lcore_has_role(rte_lcore_id(), ROLE_NON_EAL)) {
+		// Called from a dataplane worker thread.
+		// Defer the notification to the control plane thread.
+		if (control_queue_push(notify_subscribers, (void *)obj, ev_type) < 0) {
+			// XXX: add error stat if push fails?
+		}
+	} else {
+		// Called from the control plane thread.
+		// Notify subscribers immediately.
+		notify_subscribers((void *)obj, ev_type, NULL);
+	}
+}
+
 STAILQ_HEAD(serializers, gr_event_serializer);
 static struct serializers serializers = STAILQ_HEAD_INITIALIZER(serializers);
 

modules/infra/control/gr_control_queue.h → main/gr_control_queue.h

@@ -13,6 +13,10 @@ struct control_queue_drain {
 	const void *obj; // Object being deleted
 };
 
+// Force drain the control queue from all items.
+// Pass ev_type and deleted_obj to item callbacks so that they can ignore/free references.
+void control_queue_drain(uint32_t ev_type, const void *deleted_obj);
+
 // Callback definition to pass arbitrary data to be processed by the control plane event loop.
 // It is up to the function to free any data referenced by the pointer if necessary.
 //

main/meson.build

@@ -3,6 +3,7 @@
 
 src += files(
   'api.c',
+  'control_queue.c',
   'dpdk.c',
   'event.c',
   'main.c',

modules/infra/api/gr_infra.h

@@ -21,6 +21,7 @@ typedef enum : uint8_t {
 	GR_IFACE_TYPE_VLAN,
 	GR_IFACE_TYPE_IPIP,
 	GR_IFACE_TYPE_BOND,
+	GR_IFACE_TYPE_BRIDGE,
 	GR_IFACE_TYPE_COUNT
 } gr_iface_type_t;
 
@@ -56,6 +57,7 @@ typedef enum : uint8_t {
 	GR_IFACE_MODE_VRF = 0,
 	GR_IFACE_MODE_XC,
 	GR_IFACE_MODE_BOND,
+	GR_IFACE_MODE_BRIDGE,
 	GR_IFACE_MODE_COUNT
 } gr_iface_mode_t;
 
@@ -433,8 +435,6 @@ struct gr_infra_cpu_affinity_set_req {
 // Helper function to convert iface type enum to string
 static inline const char *gr_iface_type_name(gr_iface_type_t type) {
 	switch (type) {
-	case GR_IFACE_TYPE_UNDEF:
-		return "undef";
 	case GR_IFACE_TYPE_VRF:
 		return "vrf";
 	case GR_IFACE_TYPE_PORT:
@@ -445,6 +445,9 @@ static inline const char *gr_iface_type_name(gr_iface_type_t type) {
 		return "ipip";
 	case GR_IFACE_TYPE_BOND:
 		return "bond";
+	case GR_IFACE_TYPE_BRIDGE:
+		return "bridge";
+	case GR_IFACE_TYPE_UNDEF:
 	case GR_IFACE_TYPE_COUNT:
 		break;
 	}
@@ -460,6 +463,8 @@ static inline const char *gr_iface_mode_name(gr_iface_mode_t mode) {
 		return "XC";
 	case GR_IFACE_MODE_BOND:
 		return "bond";
+	case GR_IFACE_MODE_BRIDGE:
+		return "bridge";
 	case GR_IFACE_MODE_COUNT:
 		break;
 	}

modules/infra/control/ctlplane.c

@@ -206,7 +206,8 @@ static void iface_cp_poll(evutil_socket_t, short reason, void *ev_iface) {
 		}
 	}
 
-	mbuf_data(mbuf)->iface = iface;
+	iface_mbuf_data(mbuf)->iface = iface;
+	iface_mbuf_data(mbuf)->vlan_id = 0;
 
 	if (post_to_stack(iface_output, mbuf) < 0) {
 		LOG(ERR, "post_to_stack: %s", strerror(errno));
@@ -396,6 +397,7 @@ static void iface_event(uint32_t event, const void *obj) {
 	case GR_IFACE_TYPE_PORT:
 	case GR_IFACE_TYPE_VLAN:
 	case GR_IFACE_TYPE_BOND:
+	case GR_IFACE_TYPE_BRIDGE:
 		break;
 	default:
 		return;

modules/infra/control/iface.c

@@ -2,6 +2,7 @@
 // Copyright (c) 2024 Robin Jarry
 
 #include <gr_config.h>
+#include <gr_control_queue.h>
 #include <gr_event.h>
 #include <gr_iface.h>
 #include <gr_log.h>
@@ -33,6 +34,7 @@ static bool iface_type_valid(gr_iface_type_t type) {
 	case GR_IFACE_TYPE_VLAN:
 	case GR_IFACE_TYPE_IPIP:
 	case GR_IFACE_TYPE_BOND:
+	case GR_IFACE_TYPE_BRIDGE:
 		return true;
 	case GR_IFACE_TYPE_UNDEF:
 	case GR_IFACE_TYPE_COUNT:
@@ -586,10 +588,11 @@ int iface_destroy(struct iface *iface) {
 
 	rte_rcu_qsbr_synchronize(gr_datapath_rcu(), RTE_QSBR_THRID_INVALID);
 
-	// Push IFACE_REMOVE event after RCU sync to ensure all datapath threads
+	// Drain the control queue after RCU sync to ensure all datapath threads
 	// have seen that this iface is gone. At this point, only packets already
-	// in the control queue may still reference it. The event triggers
-	// a drain that frees those packets before type->fini() frees the iface.
+	// in the control queue may still reference it.
+	control_queue_drain(GR_EVENT_IFACE_REMOVE, iface);
+
 	gr_event_push(GR_EVENT_IFACE_REMOVE, iface);
 
 	type = iface_type_get(iface->type);

modules/infra/control/meson.build

@@ -3,7 +3,6 @@
 
 src += files(
   'bond.c',
-  'control_queue.c',
   'ctlplane.c',
   'graph.c',
   'group_nexthop.c',

modules/infra/control/nexthop.c

@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: BSD-3-Clause
 // Copyright (c) 2024 Robin Jarry
 
+#include <gr_control_queue.h>
 #include <gr_event.h>
 #include <gr_id_pool.h>
 #include <gr_iface.h>
@@ -478,11 +479,11 @@ void nexthop_destroy(struct nexthop *nh) {
 
 	rte_rcu_qsbr_synchronize(gr_datapath_rcu(), RTE_QSBR_THRID_INVALID);
 
-	// Push NEXTHOP_DELETE event after RCU sync to ensure all datapath
+	// Drain the control queue after RCU sync to ensure all datapath
 	// threads have seen that this nexthop is gone. At this point, only
 	// packets already in the control queue may still reference it.
-	// The event triggers a drain that frees those packets before we free
-	// the nexthop memory.
+	control_queue_drain(GR_EVENT_NEXTHOP_DELETE, nh);
+
 	if (nh->origin != GR_NH_ORIGIN_INTERNAL)
 		gr_event_push(GR_EVENT_NEXTHOP_DELETE, nh);
 

modules/infra/datapath/eth_output.c

@@ -51,6 +51,8 @@ eth_output_process(struct rte_graph *graph, struct rte_node *node, void **objs,
 		eth->src_addr = src_mac;
 		eth->ether_type = priv->ether_type;
 
+		iface_mbuf_data(mbuf)->vlan_id = 0;
+
 		edge = OUTPUT;
 next:
 		if (gr_mbuf_is_traced(mbuf)) {

modules/infra/datapath/iface_output.c

@@ -56,45 +56,42 @@ static uint16_t iface_output_process(
 	void **objs,
 	uint16_t nb_objs
 ) {
-	uint16_t iface_id, vlan_id;
 	const struct iface *iface;
+	struct iface_mbuf_data *d;
 	struct rte_mbuf *m;
 	rte_edge_t edge;
 
 	IFACE_STATS_VARS(tx);
 
 	for (uint16_t i = 0; i < nb_objs; i++) {
 		m = objs[i];
-		iface = mbuf_data(m)->iface;
-		iface_id = iface->id;
+		d = iface_mbuf_data(m);
+		iface = d->iface;
 
 		if (iface->type == GR_IFACE_TYPE_VLAN) {
 			const struct iface_info_vlan *vlan = iface_info_vlan(iface);
-			vlan_id = vlan->vlan_id;
+			d->vlan_id = vlan->vlan_id;
 			iface = iface_from_id(vlan->parent_id);
-		} else {
-			vlan_id = 0;
 		}
 
 		if (gr_mbuf_is_traced(m)) {
 			struct iface_output_trace_data *t = gr_mbuf_trace_add(m, node, sizeof(*t));
-			t->iface_id = iface_id;
-			t->vlan_id = vlan_id;
+			t->iface_id = d->iface->id;
+			t->vlan_id = d->vlan_id;
 		}
 
 		if (iface == NULL) {
 			edge = NO_PARENT;
 			goto next;
 		}
-		if (!(iface->flags & GR_IFACE_F_UP)) {
+		if (!(d->iface->flags & GR_IFACE_F_UP)) {
 			edge = IFACE_DOWN;
 			goto next;
 		}
 
-		IFACE_STATS_INC(tx, m, iface);
+		IFACE_STATS_INC(tx, m, d->iface);
 
-		iface_mbuf_data(m)->iface = iface;
-		iface_mbuf_data(m)->vlan_id = vlan_id;
+		d->iface = iface;
 		edge = iface_type_edges[iface->type];
 next:
 		rte_node_enqueue_x1(graph, node, edge, m);