+
 static void
 _update_default_route (NMDevice *self, int addr_family, gboolean has, gboolean is_assumed)

spurious newline


+              g_dbus_method_invocation_return_value (data->context,
+                                         g_variant_new ("(u)", data->state));

indention.


+         priv->concheck_periodic_id = g_signal_connect (nm_connectivity_get (),
+                                                     NM_CONNECTIVITY_PERIODIC_CHECK,

indention.



I think the global connectivity check should not consider all devices. In device_connectivity_changed(), the connectivity is not what is the connectivity state of the best-device (the one with the default route).

For the same reason, as the global metered state is not the best state of all devices, but the state of the best-device.

On the other hand, check_connectivity_auth_done_cb() should trigger a recheck for all devices, but it should only return when all pending requests are complete (not the shortcut via "if (data->state != NM_CONNECTIVITY_FULL)".




+         g_signal_handler_disconnect (nm_connectivity_get (), priv->concheck_periodic_id);

I would prefer to take a reference to singletons for which we subscribe. But arguably, NMDevice doesn't do that for NMConfig or NMPlatform either... which is fine, as long as *somebody* keeps the singleton alive at least as long as the NMDevice is alive. In short: have NMManager own a reference to NMConnectivity so that the lifetime is NMDevice < NMManager < NMConnectivity.

Alternatively,

  NM_UTILS_KEEP_ALIVE (manager, nm_connectivity_get (), "NMManager-depends-on-NMConnectivity");

in main().



+ const gchar *connectivity_state_string

I think char,int,float,double is always preferred over gchar,gint,gfloat,gdouble.

Pushed an updated branch.

Added a commit that adjusts metrics according to the connectivity check results.

> core: make connectivity checking per-device

+#if WITH_CONCHECK
+static void
+concheck_cb (GObject *source_object,
+                       GAsyncResult *result,
+                       gpointer user_data)

Indentation.


+static void
+concheck_periodic_update (NMDevice *self)
+ [...]
+       } else if (!check_enable && priv->concheck_periodic_id) {
+               /* The default route has gone off, and so has connectivity. */
+               update_connectivity_state (self, NM_CONNECTIVITY_NONE);
+               g_signal_handler_disconnect (nm_connectivity_get (), priv->concheck_periodic_id);
+               priv->concheck_periodic_id = 0;
+       }

update_connectivity_state() triggers a reconfiguration of IPv4/IPv6,
which in turn calls again into concheck_periodic_update(). I think we
should move the clearing of priv->concheck_periodic_id before
update_connectivity_state() here, so that no action is taken in the
nested execution of concheck_periodic_update().


The rest LGTM.

29af644f9ad92e7ee7b602875a3b1afbbd4f3e8e