network: register already enslaved interfaces when their master appears (#43557)

When networkd starts or restarts, links are enumerated in ifindex
order. A slave enumerated before its master cannot register itself in
link_append_to_master(), as the master Link object does not exist yet.
The master then wrongly believes it has no slaves until the next
RTM_NEWLINK message for the slave arrives.

The stale slaves set affects operational state propagation
(degraded-carrier detection) and any code that inspects the slaves of
a master.
For instance, on restart, when networkd thinks that the master has no
slaves yet, we try to set the bond fail-over policy, which the kernel
then refuses because in reality the bond does have slaves.

To fix the issue, this updates the list of slaves after enumeration.

Co-authored-by: Yu Watanabe <watanabe.yu+github@gmail.com>

src/network/networkd-link.c

@@ -1035,7 +1035,7 @@ static void link_free_bound_by_list(Link *link) {
                 link_dirty(link);
 }
 
-static int link_append_to_master(Link *link) {
+int link_append_to_master(Link *link) {
         Link *master;
         int r;
 

src/network/networkd-link.h

@@ -222,6 +222,7 @@ int link_get_master(Link *link, Link **ret);
 int link_getlink_handler_internal(sd_netlink *rtnl, sd_netlink_message *m, Link *link, const char *error_msg);
 int link_call_getlink(Link *link, link_netlink_message_handler_t callback);
 int link_handle_bound_to_list(Link *link);
+int link_append_to_master(Link *link);
 
 void link_enter_failed(Link *link);
 void link_set_state(Link *link, LinkState state);

src/network/networkd-manager.c

@@ -939,7 +939,18 @@ static int manager_enumerate_links(Manager *m) {
         if (r < 0)
                 return r;
 
-        return manager_enumerate_internal(m, m->rtnl, req, manager_rtnl_process_link);
+        r = manager_enumerate_internal(m, m->rtnl, req, manager_rtnl_process_link);
+        if (r < 0)
+                return r;
+
+        /* Slave interfaces enumerated before the master could not register themselves, as the master Link
+         * object did not exist yet at that time. Register them now, so that the slaves set correctly
+         * reflects the kernel state. */
+        Link *link;
+        HASHMAP_FOREACH(link, m->links_by_index)
+                RET_GATHER(r, link_append_to_master(link));
+
+        return r;
 }
 
 static int manager_enumerate_qdisc(Manager *m) {

test/test-network/conf/25-bond-fail-over-mac.netdev

@@ -0,0 +1,9 @@
+# SPDX-License-Identifier: LGPL-2.1-or-later
+[NetDev]
+Name=bond99
+Kind=bond
+
+[Bond]
+Mode=active-backup
+FailOverMACPolicy=active
+MIIMonitorSec=1s

test/test-network/conf/25-bond-fail-over-mac.network

@@ -0,0 +1,7 @@
+# SPDX-License-Identifier: LGPL-2.1-or-later
+[Match]
+Name=bond99
+
+[Network]
+IPv6AcceptRA=no
+Address=192.168.123.45/24

test/test-network/systemd-networkd-tests.py

@@ -6734,6 +6734,23 @@ class NetworkdBondTests(unittest.TestCase, Utilities):
             check_output('ip link set dummy98 master bond199')
             self.wait_online('dummy98:enslaved')
 
+    def test_bond_fail_over_mac_restart(self):
+        # The slave needs a lower ifindex than the bond, so that it is enumerated
+        # first and the race is deterministic. So create it before networkd starts.
+        check_output('ip link add dummy98 type dummy')
+        copy_network_unit(
+            '25-bond-fail-over-mac.netdev',
+            '25-bond-fail-over-mac.network',
+            '25-bond-slave.network',
+        )
+        start_networkd()
+        self.wait_online('dummy98:enslaved', 'bond99:routable')
+
+        self.check_link_attr('bond99', 'bonding', 'fail_over_mac', 'active 1')
+
+        restart_networkd()
+        self.wait_online('dummy98:enslaved', 'bond99:routable')
+
     def test_bond_active_slave(self):
         copy_network_unit(
             '23-active-slave.network',