Add L2 bridge interface support

[rjarry]

Implement L2 bridging functionality with MAC learning, flooding, and per-bridge configuration options.

This series introduces a new bridge interface type that can attach port and bond interfaces as members. Packets received on member interfaces are forwarded based on learned MAC addresses, with unknown unicast, broadcast and multicast traffic flooded to all other members.

The first commits prepare the infrastructure by moving the control queue to avoid circular dependencies and allowing nexthop creation from datapath workers. This enables the bridge input node to learn MAC addresses on the fly without blocking the control plane.

The L2 nexthop type stores (bridge_id, vlan_id, mac) tuples with automatic ageing based on per-bridge configurable timeout. Static entries can be created via CLI to override learned ones.

Bridge configuration supports disabling flooding (NO_FLOOD) and MAC learning (NO_LEARN) flags, custom MAC address, and up to 64 member interfaces. The frr plugin exposes bridges to zebra as ZEBRA_IF_BRIDGE with proper slave type on members.

Summary by CodeRabbit

Release Notes

• New Features

  • Added L2 bridge interface type with member management, supporting PORT, VLAN, and BOND interfaces as members.
  • Introduced Forwarding Database (FDB) for bridge learning and MAC address management.
  • Added bridge configuration: MAC address, ageing time, and flood/learn behavior flags.
  • Added CLI commands for bridge and FDB management: create, configure, and query operations.
  • Enhanced interface lifecycle with improved VRF binding and cleanup mechanisms.
  • Improved IPv6 address initialization and lifecycle management on interface events.

• Bug Fixes

  • Fixed null-pointer dereference in cross-connect processing.
  • Enhanced RCU synchronization for proper resource cleanup across dataplane operations.

• Tests

  • Added comprehensive bridge and L3 routing test coverage.

[coderabbitai]

Actionable comments posted: 3

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

main/control_queue.c (1)

50-52: ⚠️ Potential issue | 🟠 Major

drained accumulates ARRAY_DIM(items) instead of count — may exit the loop prematurely.

Line 52 adds the array capacity (ARRAY_DIM(items)) to drained rather than the actual number of items dequeued (count). If the ring has more than ARRAY_DIM(items) pending items but a burst returns fewer, drained inflates faster than actual progress, causing the loop to hit the CONTROL_QUEUE_SIZE cap before all items are processed.

Proposed fix

-	drained += ARRAY_DIM(items);
+	drained += count;

🤖 Fix all issues with AI agents

In `@modules/l2/control/bridge.c`:
- Around line 31-52: bridge_attach_member currently allows adding the same iface
multiple times; add a duplicate check similar to bond_attach_member by iterating
br->members[0..br->n_members-1] and if any entry equals the incoming member
return 0 (no-op) before appending. Use the same comparison semantics as
bond_attach_member, keep the check before the capacity check (br->n_members ==
ARRAY_DIM(br->members)), and do not change member state when returning early.

In `@modules/l2/control/fdb.c`:
- Around line 378-381: The ageing calculation uses CLOCKS_PER_SEC to convert the
microsecond value returned by gr_clock_us() which is misleading; update the
conversion in the age calculation (the code using fdb->last_seen, gr_clock_us(),
and age) to divide by an explicit microsecond-per-second constant (e.g. 1000000
or a US_PER_SEC constant) instead of CLOCKS_PER_SEC so the unit conversion is
explicit and robust; add a named constant US_PER_SEC and use (now -
fdb->last_seen) / US_PER_SEC when computing age.
- Around line 417-422: fdb_pool allocated in fdb_reconfig is never freed in
fdb_fini causing a mempool leak; update fdb_fini to check fdb_pool and call the
appropriate mempool free (e.g. rte_mempool_free or the project-specific free
function) and set fdb_pool to NULL, in addition to the existing cleanup of
ageing_timer and fdb_hash so both the hash and the mempool are released on
shutdown.

[coderabbitai]

⚠️ Potential issue | 🟠 Major

🧩 Analysis chain

🏁 Script executed:

# First, let's check the bond_attach_member function to see if it has a duplicate check
cat modules/infra/control/bond.c | sed -n '170,185p'

Repository: DPDK/grout

Length of output: 529

---

🏁 Script executed:

# Let's also find the caller of bridge_attach_member to understand the attach flow
rg -n 'bridge_attach_member' --type c

Repository: DPDK/grout

Length of output: 234

---

🏁 Script executed:

# Check how iface attach_domain is called to understand if there's higher-level protection
rg -n 'attach_domain|detach_domain' modules/infra/control/iface.c -A 5 -B 5

Repository: DPDK/grout

Length of output: 2246

---

🏁 Script executed:

# Look for any domain attachment safeguards in the iface module
rg -n 'struct iface' modules/infra/control/iface.h -A 3 | head -40

Repository: DPDK/grout

Length of output: 123

---

Missing duplicate member check in bridge_attach_member.

If the same interface is attached to a bridge multiple times, it will occupy multiple slots in br->members and cause inconsistencies on detach (only the first slot is removed). The analogous bond_attach_member in modules/infra/control/bond.c explicitly checks for and returns early on duplicates—bridge should do the same.

Proposed fix

 static int bridge_attach_member(struct iface *bridge, struct iface *member) {
 	struct iface_info_bridge *br = iface_info_bridge(bridge);
 
 	switch (member->type) {
 	case GR_IFACE_TYPE_PORT:
 	case GR_IFACE_TYPE_VLAN:
 	case GR_IFACE_TYPE_BOND:
 		break;
 	default:
 		return errno_set(EMEDIUMTYPE);
 	}
 
+	for (unsigned i = 0; i < br->n_members; i++) {
+		if (br->members[i] == member)
+			return 0; // already a member
+	}
+
 	if (br->n_members == ARRAY_DIM(br->members))
 		return errno_set(EUSERS);

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	static int bridge_attach_member(struct iface *bridge, struct iface *member) {
	struct iface_info_bridge *br = iface_info_bridge(bridge);
	switch (member->type) {
	case GR_IFACE_TYPE_PORT:
	case GR_IFACE_TYPE_VLAN:
	case GR_IFACE_TYPE_BOND:
	break;
	default:
	return errno_set(EMEDIUMTYPE);
	}
	if (br->n_members == ARRAY_DIM(br->members))
	return errno_set(EUSERS);
	br->members[br->n_members++] = member;
	member->domain_id = bridge->id;
	member->vrf_id = GR_VRF_ID_UNDEF;
	member->mode = GR_IFACE_MODE_BRIDGE;
	return 0;
	}
	static int bridge_attach_member(struct iface *bridge, struct iface *member) {
	struct iface_info_bridge *br = iface_info_bridge(bridge);
	switch (member->type) {
	case GR_IFACE_TYPE_PORT:
	case GR_IFACE_TYPE_VLAN:
	case GR_IFACE_TYPE_BOND:
	break;
	default:
	return errno_set(EMEDIUMTYPE);
	}
	for (unsigned i = 0; i < br->n_members; i++) {
	if (br->members[i] == member)
	return 0; // already a member
	}
	if (br->n_members == ARRAY_DIM(br->members))
	return errno_set(EUSERS);
	br->members[br->n_members++] = member;
	member->domain_id = bridge->id;
	member->vrf_id = GR_VRF_ID_UNDEF;
	member->mode = GR_IFACE_MODE_BRIDGE;
	return 0;
	}

🤖 Prompt for AI Agents

In `@modules/l2/control/bridge.c` around lines 31 - 52, bridge_attach_member
currently allows adding the same iface multiple times; add a duplicate check
similar to bond_attach_member by iterating br->members[0..br->n_members-1] and
if any entry equals the incoming member return 0 (no-op) before appending. Use
the same comparison semantics as bond_attach_member, keep the check before the
capacity check (br->n_members == ARRAY_DIM(br->members)), and do not change
member state when returning early.

[coderabbitai]

⚠️ Potential issue | 🟡 Minor

Ageing arithmetic uses CLOCKS_PER_SEC to convert gr_clock_us() microseconds — semantically misleading.

gr_clock_us() returns microseconds. Dividing by CLOCKS_PER_SEC (which is 1,000,000 on POSIX) happens to produce seconds, but only by coincidence — CLOCKS_PER_SEC is defined for clock() ticks, not for microsecond values. If gr_clock_us() ever changes unit, or if this runs on a non-POSIX platform, this silently breaks.

Consider using 1000000 or a US_PER_SEC constant explicitly.

🤖 Prompt for AI Agents

In `@modules/l2/control/fdb.c` around lines 378 - 381, The ageing calculation uses
CLOCKS_PER_SEC to convert the microsecond value returned by gr_clock_us() which
is misleading; update the conversion in the age calculation (the code using
fdb->last_seen, gr_clock_us(), and age) to divide by an explicit
microsecond-per-second constant (e.g. 1000000 or a US_PER_SEC constant) instead
of CLOCKS_PER_SEC so the unit conversion is explicit and robust; add a named
constant US_PER_SEC and use (now - fdb->last_seen) / US_PER_SEC when computing
age.

[coderabbitai]

⚠️ Potential issue | 🟠 Major

fdb_pool is never freed in fdb_fini — resource leak.

fdb_reconfig allocates both fdb_hash and fdb_pool, but fdb_fini only frees the hash. The mempool leaks on shutdown.

Proposed fix

 static void fdb_fini(struct event_base *) {
 	if (ageing_timer != NULL)
 		event_free(ageing_timer);
 
 	rte_hash_free(fdb_hash);
+	rte_mempool_free(fdb_pool);
 }

🤖 Prompt for AI Agents

In `@modules/l2/control/fdb.c` around lines 417 - 422, fdb_pool allocated in
fdb_reconfig is never freed in fdb_fini causing a mempool leak; update fdb_fini
to check fdb_pool and call the appropriate mempool free (e.g. rte_mempool_free
or the project-specific free function) and set fdb_pool to NULL, in addition to
the existing cleanup of ageing_timer and fdb_hash so both the hash and the
mempool are released on shutdown.