xref: /openbmc/linux/drivers/net/macvtap.c (revision 9a393b5d5988ea4eaa3e0da138321abe0dc03a68)
1*9a393b5dSSainath Grandhi #include <linux/etherdevice.h>
2*9a393b5dSSainath Grandhi #include <linux/if_macvlan.h>
3*9a393b5dSSainath Grandhi #include <linux/if_tap.h>
4*9a393b5dSSainath Grandhi #include <linux/if_vlan.h>
5*9a393b5dSSainath Grandhi #include <linux/interrupt.h>
6*9a393b5dSSainath Grandhi #include <linux/nsproxy.h>
7*9a393b5dSSainath Grandhi #include <linux/compat.h>
8*9a393b5dSSainath Grandhi #include <linux/if_tun.h>
9*9a393b5dSSainath Grandhi #include <linux/module.h>
10*9a393b5dSSainath Grandhi #include <linux/skbuff.h>
11*9a393b5dSSainath Grandhi #include <linux/cache.h>
12*9a393b5dSSainath Grandhi #include <linux/sched.h>
13*9a393b5dSSainath Grandhi #include <linux/types.h>
14*9a393b5dSSainath Grandhi #include <linux/slab.h>
15*9a393b5dSSainath Grandhi #include <linux/wait.h>
16*9a393b5dSSainath Grandhi #include <linux/cdev.h>
17*9a393b5dSSainath Grandhi #include <linux/idr.h>
18*9a393b5dSSainath Grandhi #include <linux/fs.h>
19*9a393b5dSSainath Grandhi #include <linux/uio.h>
20*9a393b5dSSainath Grandhi 
21*9a393b5dSSainath Grandhi #include <net/net_namespace.h>
22*9a393b5dSSainath Grandhi #include <net/rtnetlink.h>
23*9a393b5dSSainath Grandhi #include <net/sock.h>
24*9a393b5dSSainath Grandhi #include <linux/virtio_net.h>
25*9a393b5dSSainath Grandhi #include <linux/skb_array.h>
26*9a393b5dSSainath Grandhi 
27*9a393b5dSSainath Grandhi struct macvtap_dev {
28*9a393b5dSSainath Grandhi 	struct macvlan_dev vlan;
29*9a393b5dSSainath Grandhi 	struct tap_dev    tap;
30*9a393b5dSSainath Grandhi };
31*9a393b5dSSainath Grandhi 
32*9a393b5dSSainath Grandhi /*
33*9a393b5dSSainath Grandhi  * Variables for dealing with macvtaps device numbers.
34*9a393b5dSSainath Grandhi  */
35*9a393b5dSSainath Grandhi static dev_t macvtap_major;
36*9a393b5dSSainath Grandhi 
37*9a393b5dSSainath Grandhi static const void *macvtap_net_namespace(struct device *d)
38*9a393b5dSSainath Grandhi {
39*9a393b5dSSainath Grandhi 	struct net_device *dev = to_net_dev(d->parent);
40*9a393b5dSSainath Grandhi 	return dev_net(dev);
41*9a393b5dSSainath Grandhi }
42*9a393b5dSSainath Grandhi 
43*9a393b5dSSainath Grandhi static struct class macvtap_class = {
44*9a393b5dSSainath Grandhi 	.name = "macvtap",
45*9a393b5dSSainath Grandhi 	.owner = THIS_MODULE,
46*9a393b5dSSainath Grandhi 	.ns_type = &net_ns_type_operations,
47*9a393b5dSSainath Grandhi 	.namespace = macvtap_net_namespace,
48*9a393b5dSSainath Grandhi };
49*9a393b5dSSainath Grandhi static struct cdev macvtap_cdev;
50*9a393b5dSSainath Grandhi 
51*9a393b5dSSainath Grandhi #define TUN_OFFLOADS (NETIF_F_HW_CSUM | NETIF_F_TSO_ECN | NETIF_F_TSO | \
52*9a393b5dSSainath Grandhi 		      NETIF_F_TSO6 | NETIF_F_UFO)
53*9a393b5dSSainath Grandhi 
54*9a393b5dSSainath Grandhi static void macvtap_count_tx_dropped(struct tap_dev *tap)
55*9a393b5dSSainath Grandhi {
56*9a393b5dSSainath Grandhi 	struct macvtap_dev *vlantap = container_of(tap, struct macvtap_dev, tap);
57*9a393b5dSSainath Grandhi 	struct macvlan_dev *vlan = &vlantap->vlan;
58*9a393b5dSSainath Grandhi 
59*9a393b5dSSainath Grandhi 	this_cpu_inc(vlan->pcpu_stats->tx_dropped);
60*9a393b5dSSainath Grandhi }
61*9a393b5dSSainath Grandhi 
62*9a393b5dSSainath Grandhi static void macvtap_count_rx_dropped(struct tap_dev *tap)
63*9a393b5dSSainath Grandhi {
64*9a393b5dSSainath Grandhi 	struct macvtap_dev *vlantap = container_of(tap, struct macvtap_dev, tap);
65*9a393b5dSSainath Grandhi 	struct macvlan_dev *vlan = &vlantap->vlan;
66*9a393b5dSSainath Grandhi 
67*9a393b5dSSainath Grandhi 	macvlan_count_rx(vlan, 0, 0, 0);
68*9a393b5dSSainath Grandhi }
69*9a393b5dSSainath Grandhi 
70*9a393b5dSSainath Grandhi static void macvtap_update_features(struct tap_dev *tap,
71*9a393b5dSSainath Grandhi 				    netdev_features_t features)
72*9a393b5dSSainath Grandhi {
73*9a393b5dSSainath Grandhi 	struct macvtap_dev *vlantap = container_of(tap, struct macvtap_dev, tap);
74*9a393b5dSSainath Grandhi 	struct macvlan_dev *vlan = &vlantap->vlan;
75*9a393b5dSSainath Grandhi 
76*9a393b5dSSainath Grandhi 	vlan->set_features = features;
77*9a393b5dSSainath Grandhi 	netdev_update_features(vlan->dev);
78*9a393b5dSSainath Grandhi }
79*9a393b5dSSainath Grandhi 
80*9a393b5dSSainath Grandhi static int macvtap_newlink(struct net *src_net,
81*9a393b5dSSainath Grandhi 			   struct net_device *dev,
82*9a393b5dSSainath Grandhi 			   struct nlattr *tb[],
83*9a393b5dSSainath Grandhi 			   struct nlattr *data[])
84*9a393b5dSSainath Grandhi {
85*9a393b5dSSainath Grandhi 	struct macvtap_dev *vlantap = netdev_priv(dev);
86*9a393b5dSSainath Grandhi 	int err;
87*9a393b5dSSainath Grandhi 
88*9a393b5dSSainath Grandhi 	INIT_LIST_HEAD(&vlantap->tap.queue_list);
89*9a393b5dSSainath Grandhi 
90*9a393b5dSSainath Grandhi 	/* Since macvlan supports all offloads by default, make
91*9a393b5dSSainath Grandhi 	 * tap support all offloads also.
92*9a393b5dSSainath Grandhi 	 */
93*9a393b5dSSainath Grandhi 	vlantap->tap.tap_features = TUN_OFFLOADS;
94*9a393b5dSSainath Grandhi 
95*9a393b5dSSainath Grandhi 	/* Register callbacks for rx/tx drops accounting and updating
96*9a393b5dSSainath Grandhi 	 * net_device features
97*9a393b5dSSainath Grandhi 	 */
98*9a393b5dSSainath Grandhi 	vlantap->tap.count_tx_dropped = macvtap_count_tx_dropped;
99*9a393b5dSSainath Grandhi 	vlantap->tap.count_rx_dropped = macvtap_count_rx_dropped;
100*9a393b5dSSainath Grandhi 	vlantap->tap.update_features  = macvtap_update_features;
101*9a393b5dSSainath Grandhi 
102*9a393b5dSSainath Grandhi 	err = netdev_rx_handler_register(dev, tap_handle_frame, &vlantap->tap);
103*9a393b5dSSainath Grandhi 	if (err)
104*9a393b5dSSainath Grandhi 		return err;
105*9a393b5dSSainath Grandhi 
106*9a393b5dSSainath Grandhi 	/* Don't put anything that may fail after macvlan_common_newlink
107*9a393b5dSSainath Grandhi 	 * because we can't undo what it does.
108*9a393b5dSSainath Grandhi 	 */
109*9a393b5dSSainath Grandhi 	err = macvlan_common_newlink(src_net, dev, tb, data);
110*9a393b5dSSainath Grandhi 	if (err) {
111*9a393b5dSSainath Grandhi 		netdev_rx_handler_unregister(dev);
112*9a393b5dSSainath Grandhi 		return err;
113*9a393b5dSSainath Grandhi 	}
114*9a393b5dSSainath Grandhi 
115*9a393b5dSSainath Grandhi 	vlantap->tap.dev = vlantap->vlan.dev;
116*9a393b5dSSainath Grandhi 
117*9a393b5dSSainath Grandhi 	return 0;
118*9a393b5dSSainath Grandhi }
119*9a393b5dSSainath Grandhi 
120*9a393b5dSSainath Grandhi static void macvtap_dellink(struct net_device *dev,
121*9a393b5dSSainath Grandhi 			    struct list_head *head)
122*9a393b5dSSainath Grandhi {
123*9a393b5dSSainath Grandhi 	struct macvtap_dev *vlantap = netdev_priv(dev);
124*9a393b5dSSainath Grandhi 
125*9a393b5dSSainath Grandhi 	netdev_rx_handler_unregister(dev);
126*9a393b5dSSainath Grandhi 	tap_del_queues(&vlantap->tap);
127*9a393b5dSSainath Grandhi 	macvlan_dellink(dev, head);
128*9a393b5dSSainath Grandhi }
129*9a393b5dSSainath Grandhi 
130*9a393b5dSSainath Grandhi static void macvtap_setup(struct net_device *dev)
131*9a393b5dSSainath Grandhi {
132*9a393b5dSSainath Grandhi 	macvlan_common_setup(dev);
133*9a393b5dSSainath Grandhi 	dev->tx_queue_len = TUN_READQ_SIZE;
134*9a393b5dSSainath Grandhi }
135*9a393b5dSSainath Grandhi 
136*9a393b5dSSainath Grandhi static struct rtnl_link_ops macvtap_link_ops __read_mostly = {
137*9a393b5dSSainath Grandhi 	.kind		= "macvtap",
138*9a393b5dSSainath Grandhi 	.setup		= macvtap_setup,
139*9a393b5dSSainath Grandhi 	.newlink	= macvtap_newlink,
140*9a393b5dSSainath Grandhi 	.dellink	= macvtap_dellink,
141*9a393b5dSSainath Grandhi 	.priv_size      = sizeof(struct macvtap_dev),
142*9a393b5dSSainath Grandhi };
143*9a393b5dSSainath Grandhi 
144*9a393b5dSSainath Grandhi static int macvtap_device_event(struct notifier_block *unused,
145*9a393b5dSSainath Grandhi 				unsigned long event, void *ptr)
146*9a393b5dSSainath Grandhi {
147*9a393b5dSSainath Grandhi 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
148*9a393b5dSSainath Grandhi 	struct macvtap_dev *vlantap;
149*9a393b5dSSainath Grandhi 	struct device *classdev;
150*9a393b5dSSainath Grandhi 	dev_t devt;
151*9a393b5dSSainath Grandhi 	int err;
152*9a393b5dSSainath Grandhi 	char tap_name[IFNAMSIZ];
153*9a393b5dSSainath Grandhi 
154*9a393b5dSSainath Grandhi 	if (dev->rtnl_link_ops != &macvtap_link_ops)
155*9a393b5dSSainath Grandhi 		return NOTIFY_DONE;
156*9a393b5dSSainath Grandhi 
157*9a393b5dSSainath Grandhi 	snprintf(tap_name, IFNAMSIZ, "tap%d", dev->ifindex);
158*9a393b5dSSainath Grandhi 	vlantap = netdev_priv(dev);
159*9a393b5dSSainath Grandhi 
160*9a393b5dSSainath Grandhi 	switch (event) {
161*9a393b5dSSainath Grandhi 	case NETDEV_REGISTER:
162*9a393b5dSSainath Grandhi 		/* Create the device node here after the network device has
163*9a393b5dSSainath Grandhi 		 * been registered but before register_netdevice has
164*9a393b5dSSainath Grandhi 		 * finished running.
165*9a393b5dSSainath Grandhi 		 */
166*9a393b5dSSainath Grandhi 		err = tap_get_minor(macvtap_major, &vlantap->tap);
167*9a393b5dSSainath Grandhi 		if (err)
168*9a393b5dSSainath Grandhi 			return notifier_from_errno(err);
169*9a393b5dSSainath Grandhi 
170*9a393b5dSSainath Grandhi 		devt = MKDEV(MAJOR(macvtap_major), vlantap->tap.minor);
171*9a393b5dSSainath Grandhi 		classdev = device_create(&macvtap_class, &dev->dev, devt,
172*9a393b5dSSainath Grandhi 					 dev, tap_name);
173*9a393b5dSSainath Grandhi 		if (IS_ERR(classdev)) {
174*9a393b5dSSainath Grandhi 			tap_free_minor(macvtap_major, &vlantap->tap);
175*9a393b5dSSainath Grandhi 			return notifier_from_errno(PTR_ERR(classdev));
176*9a393b5dSSainath Grandhi 		}
177*9a393b5dSSainath Grandhi 		err = sysfs_create_link(&dev->dev.kobj, &classdev->kobj,
178*9a393b5dSSainath Grandhi 					tap_name);
179*9a393b5dSSainath Grandhi 		if (err)
180*9a393b5dSSainath Grandhi 			return notifier_from_errno(err);
181*9a393b5dSSainath Grandhi 		break;
182*9a393b5dSSainath Grandhi 	case NETDEV_UNREGISTER:
183*9a393b5dSSainath Grandhi 		/* vlan->minor == 0 if NETDEV_REGISTER above failed */
184*9a393b5dSSainath Grandhi 		if (vlantap->tap.minor == 0)
185*9a393b5dSSainath Grandhi 			break;
186*9a393b5dSSainath Grandhi 		sysfs_remove_link(&dev->dev.kobj, tap_name);
187*9a393b5dSSainath Grandhi 		devt = MKDEV(MAJOR(macvtap_major), vlantap->tap.minor);
188*9a393b5dSSainath Grandhi 		device_destroy(&macvtap_class, devt);
189*9a393b5dSSainath Grandhi 		tap_free_minor(macvtap_major, &vlantap->tap);
190*9a393b5dSSainath Grandhi 		break;
191*9a393b5dSSainath Grandhi 	case NETDEV_CHANGE_TX_QUEUE_LEN:
192*9a393b5dSSainath Grandhi 		if (tap_queue_resize(&vlantap->tap))
193*9a393b5dSSainath Grandhi 			return NOTIFY_BAD;
194*9a393b5dSSainath Grandhi 		break;
195*9a393b5dSSainath Grandhi 	}
196*9a393b5dSSainath Grandhi 
197*9a393b5dSSainath Grandhi 	return NOTIFY_DONE;
198*9a393b5dSSainath Grandhi }
199*9a393b5dSSainath Grandhi 
200*9a393b5dSSainath Grandhi static struct notifier_block macvtap_notifier_block __read_mostly = {
201*9a393b5dSSainath Grandhi 	.notifier_call	= macvtap_device_event,
202*9a393b5dSSainath Grandhi };
203*9a393b5dSSainath Grandhi 
204*9a393b5dSSainath Grandhi static int macvtap_init(void)
205*9a393b5dSSainath Grandhi {
206*9a393b5dSSainath Grandhi 	int err;
207*9a393b5dSSainath Grandhi 
208*9a393b5dSSainath Grandhi 	err = tap_create_cdev(&macvtap_cdev, &macvtap_major, "macvtap");
209*9a393b5dSSainath Grandhi 
210*9a393b5dSSainath Grandhi 	if (err)
211*9a393b5dSSainath Grandhi 		goto out1;
212*9a393b5dSSainath Grandhi 
213*9a393b5dSSainath Grandhi 	err = class_register(&macvtap_class);
214*9a393b5dSSainath Grandhi 	if (err)
215*9a393b5dSSainath Grandhi 		goto out2;
216*9a393b5dSSainath Grandhi 
217*9a393b5dSSainath Grandhi 	err = register_netdevice_notifier(&macvtap_notifier_block);
218*9a393b5dSSainath Grandhi 	if (err)
219*9a393b5dSSainath Grandhi 		goto out3;
220*9a393b5dSSainath Grandhi 
221*9a393b5dSSainath Grandhi 	err = macvlan_link_register(&macvtap_link_ops);
222*9a393b5dSSainath Grandhi 	if (err)
223*9a393b5dSSainath Grandhi 		goto out4;
224*9a393b5dSSainath Grandhi 
225*9a393b5dSSainath Grandhi 	return 0;
226*9a393b5dSSainath Grandhi 
227*9a393b5dSSainath Grandhi out4:
228*9a393b5dSSainath Grandhi 	unregister_netdevice_notifier(&macvtap_notifier_block);
229*9a393b5dSSainath Grandhi out3:
230*9a393b5dSSainath Grandhi 	class_unregister(&macvtap_class);
231*9a393b5dSSainath Grandhi out2:
232*9a393b5dSSainath Grandhi 	tap_destroy_cdev(macvtap_major, &macvtap_cdev);
233*9a393b5dSSainath Grandhi out1:
234*9a393b5dSSainath Grandhi 	return err;
235*9a393b5dSSainath Grandhi }
236*9a393b5dSSainath Grandhi module_init(macvtap_init);
237*9a393b5dSSainath Grandhi 
238*9a393b5dSSainath Grandhi static void macvtap_exit(void)
239*9a393b5dSSainath Grandhi {
240*9a393b5dSSainath Grandhi 	rtnl_link_unregister(&macvtap_link_ops);
241*9a393b5dSSainath Grandhi 	unregister_netdevice_notifier(&macvtap_notifier_block);
242*9a393b5dSSainath Grandhi 	class_unregister(&macvtap_class);
243*9a393b5dSSainath Grandhi 	tap_destroy_cdev(macvtap_major, &macvtap_cdev);
244*9a393b5dSSainath Grandhi }
245*9a393b5dSSainath Grandhi module_exit(macvtap_exit);
246*9a393b5dSSainath Grandhi 
247*9a393b5dSSainath Grandhi MODULE_ALIAS_RTNL_LINK("macvtap");
248*9a393b5dSSainath Grandhi MODULE_AUTHOR("Arnd Bergmann <arnd@arndb.de>");
249*9a393b5dSSainath Grandhi MODULE_LICENSE("GPL");
250