xref: /linux/net/openvswitch/vport-internal_dev.c (revision 088e88be5a380cc4e81963a9a02815da465d144f)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (c) 2007-2012 Nicira, Inc.
4  */
5 
6 #include <linux/if_vlan.h>
7 #include <linux/kernel.h>
8 #include <linux/netdevice.h>
9 #include <linux/etherdevice.h>
10 #include <linux/ethtool.h>
11 #include <linux/skbuff.h>
12 
13 #include <net/dst.h>
14 #include <net/xfrm.h>
15 #include <net/rtnetlink.h>
16 
17 #include "datapath.h"
18 #include "vport-internal_dev.h"
19 #include "vport-netdev.h"
20 
21 struct internal_dev {
22 	struct vport *vport;
23 };
24 
25 static struct vport_ops ovs_internal_vport_ops;
26 
27 static struct internal_dev *internal_dev_priv(struct net_device *netdev)
28 {
29 	return netdev_priv(netdev);
30 }
31 
32 /* Called with rcu_read_lock_bh. */
33 static netdev_tx_t
34 internal_dev_xmit(struct sk_buff *skb, struct net_device *netdev)
35 {
36 	int len, err;
37 
38 	len = skb->len;
39 	rcu_read_lock();
40 	err = ovs_vport_receive(internal_dev_priv(netdev)->vport, skb, NULL);
41 	rcu_read_unlock();
42 
43 	if (likely(!err)) {
44 		struct pcpu_sw_netstats *tstats = this_cpu_ptr(netdev->tstats);
45 
46 		u64_stats_update_begin(&tstats->syncp);
47 		tstats->tx_bytes += len;
48 		tstats->tx_packets++;
49 		u64_stats_update_end(&tstats->syncp);
50 	} else {
51 		netdev->stats.tx_errors++;
52 	}
53 	return NETDEV_TX_OK;
54 }
55 
56 static int internal_dev_open(struct net_device *netdev)
57 {
58 	netif_start_queue(netdev);
59 	return 0;
60 }
61 
62 static int internal_dev_stop(struct net_device *netdev)
63 {
64 	netif_stop_queue(netdev);
65 	return 0;
66 }
67 
68 static void internal_dev_getinfo(struct net_device *netdev,
69 				 struct ethtool_drvinfo *info)
70 {
71 	strlcpy(info->driver, "openvswitch", sizeof(info->driver));
72 }
73 
74 static const struct ethtool_ops internal_dev_ethtool_ops = {
75 	.get_drvinfo	= internal_dev_getinfo,
76 	.get_link	= ethtool_op_get_link,
77 };
78 
79 static void internal_dev_destructor(struct net_device *dev)
80 {
81 	struct vport *vport = ovs_internal_dev_get_vport(dev);
82 
83 	ovs_vport_free(vport);
84 }
85 
86 static void
87 internal_get_stats(struct net_device *dev, struct rtnl_link_stats64 *stats)
88 {
89 	int i;
90 
91 	memset(stats, 0, sizeof(*stats));
92 	stats->rx_errors  = dev->stats.rx_errors;
93 	stats->tx_errors  = dev->stats.tx_errors;
94 	stats->tx_dropped = dev->stats.tx_dropped;
95 	stats->rx_dropped = dev->stats.rx_dropped;
96 
97 	for_each_possible_cpu(i) {
98 		const struct pcpu_sw_netstats *percpu_stats;
99 		struct pcpu_sw_netstats local_stats;
100 		unsigned int start;
101 
102 		percpu_stats = per_cpu_ptr(dev->tstats, i);
103 
104 		do {
105 			start = u64_stats_fetch_begin_irq(&percpu_stats->syncp);
106 			local_stats = *percpu_stats;
107 		} while (u64_stats_fetch_retry_irq(&percpu_stats->syncp, start));
108 
109 		stats->rx_bytes         += local_stats.rx_bytes;
110 		stats->rx_packets       += local_stats.rx_packets;
111 		stats->tx_bytes         += local_stats.tx_bytes;
112 		stats->tx_packets       += local_stats.tx_packets;
113 	}
114 }
115 
116 static const struct net_device_ops internal_dev_netdev_ops = {
117 	.ndo_open = internal_dev_open,
118 	.ndo_stop = internal_dev_stop,
119 	.ndo_start_xmit = internal_dev_xmit,
120 	.ndo_set_mac_address = eth_mac_addr,
121 	.ndo_get_stats64 = internal_get_stats,
122 };
123 
124 static struct rtnl_link_ops internal_dev_link_ops __read_mostly = {
125 	.kind = "openvswitch",
126 };
127 
128 static void do_setup(struct net_device *netdev)
129 {
130 	ether_setup(netdev);
131 
132 	netdev->max_mtu = ETH_MAX_MTU;
133 
134 	netdev->netdev_ops = &internal_dev_netdev_ops;
135 
136 	netdev->priv_flags &= ~IFF_TX_SKB_SHARING;
137 	netdev->priv_flags |= IFF_LIVE_ADDR_CHANGE | IFF_OPENVSWITCH |
138 			      IFF_NO_QUEUE;
139 	netdev->needs_free_netdev = true;
140 	netdev->priv_destructor = internal_dev_destructor;
141 	netdev->ethtool_ops = &internal_dev_ethtool_ops;
142 	netdev->rtnl_link_ops = &internal_dev_link_ops;
143 
144 	netdev->features = NETIF_F_LLTX | NETIF_F_SG | NETIF_F_FRAGLIST |
145 			   NETIF_F_HIGHDMA | NETIF_F_HW_CSUM |
146 			   NETIF_F_GSO_SOFTWARE | NETIF_F_GSO_ENCAP_ALL;
147 
148 	netdev->vlan_features = netdev->features;
149 	netdev->hw_enc_features = netdev->features;
150 	netdev->features |= NETIF_F_HW_VLAN_CTAG_TX | NETIF_F_HW_VLAN_STAG_TX;
151 	netdev->hw_features = netdev->features & ~NETIF_F_LLTX;
152 
153 	eth_hw_addr_random(netdev);
154 }
155 
156 static struct vport *internal_dev_create(const struct vport_parms *parms)
157 {
158 	struct vport *vport;
159 	struct internal_dev *internal_dev;
160 	struct net_device *dev;
161 	int err;
162 	bool free_vport = true;
163 
164 	vport = ovs_vport_alloc(0, &ovs_internal_vport_ops, parms);
165 	if (IS_ERR(vport)) {
166 		err = PTR_ERR(vport);
167 		goto error;
168 	}
169 
170 	dev = alloc_netdev(sizeof(struct internal_dev),
171 			   parms->name, NET_NAME_USER, do_setup);
172 	vport->dev = dev;
173 	if (!vport->dev) {
174 		err = -ENOMEM;
175 		goto error_free_vport;
176 	}
177 	vport->dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats);
178 	if (!vport->dev->tstats) {
179 		err = -ENOMEM;
180 		goto error_free_netdev;
181 	}
182 
183 	dev_net_set(vport->dev, ovs_dp_get_net(vport->dp));
184 	internal_dev = internal_dev_priv(vport->dev);
185 	internal_dev->vport = vport;
186 
187 	/* Restrict bridge port to current netns. */
188 	if (vport->port_no == OVSP_LOCAL)
189 		vport->dev->features |= NETIF_F_NETNS_LOCAL;
190 
191 	rtnl_lock();
192 	err = register_netdevice(vport->dev);
193 	if (err) {
194 		free_vport = false;
195 		goto error_unlock;
196 	}
197 
198 	dev_set_promiscuity(vport->dev, 1);
199 	rtnl_unlock();
200 	netif_start_queue(vport->dev);
201 
202 	return vport;
203 
204 error_unlock:
205 	rtnl_unlock();
206 	free_percpu(dev->tstats);
207 error_free_netdev:
208 	free_netdev(dev);
209 error_free_vport:
210 	if (free_vport)
211 		ovs_vport_free(vport);
212 error:
213 	return ERR_PTR(err);
214 }
215 
216 static void internal_dev_destroy(struct vport *vport)
217 {
218 	netif_stop_queue(vport->dev);
219 	rtnl_lock();
220 	dev_set_promiscuity(vport->dev, -1);
221 
222 	/* unregister_netdevice() waits for an RCU grace period. */
223 	unregister_netdevice(vport->dev);
224 	free_percpu(vport->dev->tstats);
225 	rtnl_unlock();
226 }
227 
228 static netdev_tx_t internal_dev_recv(struct sk_buff *skb)
229 {
230 	struct net_device *netdev = skb->dev;
231 	struct pcpu_sw_netstats *stats;
232 
233 	if (unlikely(!(netdev->flags & IFF_UP))) {
234 		kfree_skb(skb);
235 		netdev->stats.rx_dropped++;
236 		return NETDEV_TX_OK;
237 	}
238 
239 	skb_dst_drop(skb);
240 	nf_reset(skb);
241 	secpath_reset(skb);
242 
243 	skb->pkt_type = PACKET_HOST;
244 	skb->protocol = eth_type_trans(skb, netdev);
245 	skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
246 
247 	stats = this_cpu_ptr(netdev->tstats);
248 	u64_stats_update_begin(&stats->syncp);
249 	stats->rx_packets++;
250 	stats->rx_bytes += skb->len;
251 	u64_stats_update_end(&stats->syncp);
252 
253 	netif_rx(skb);
254 	return NETDEV_TX_OK;
255 }
256 
257 static struct vport_ops ovs_internal_vport_ops = {
258 	.type		= OVS_VPORT_TYPE_INTERNAL,
259 	.create		= internal_dev_create,
260 	.destroy	= internal_dev_destroy,
261 	.send		= internal_dev_recv,
262 };
263 
264 int ovs_is_internal_dev(const struct net_device *netdev)
265 {
266 	return netdev->netdev_ops == &internal_dev_netdev_ops;
267 }
268 
269 struct vport *ovs_internal_dev_get_vport(struct net_device *netdev)
270 {
271 	if (!ovs_is_internal_dev(netdev))
272 		return NULL;
273 
274 	return internal_dev_priv(netdev)->vport;
275 }
276 
277 int ovs_internal_dev_rtnl_link_register(void)
278 {
279 	int err;
280 
281 	err = rtnl_link_register(&internal_dev_link_ops);
282 	if (err < 0)
283 		return err;
284 
285 	err = ovs_vport_ops_register(&ovs_internal_vport_ops);
286 	if (err < 0)
287 		rtnl_link_unregister(&internal_dev_link_ops);
288 
289 	return err;
290 }
291 
292 void ovs_internal_dev_rtnl_link_unregister(void)
293 {
294 	ovs_vport_ops_unregister(&ovs_internal_vport_ops);
295 	rtnl_link_unregister(&internal_dev_link_ops);
296 }
297