xref: /openbmc/linux/drivers/net/ifb.c (revision f9834f18)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* drivers/net/ifb.c:
3 
4 	The purpose of this driver is to provide a device that allows
5 	for sharing of resources:
6 
7 	1) qdiscs/policies that are per device as opposed to system wide.
8 	ifb allows for a device which can be redirected to thus providing
9 	an impression of sharing.
10 
11 	2) Allows for queueing incoming traffic for shaping instead of
12 	dropping.
13 
14 	The original concept is based on what is known as the IMQ
15 	driver initially written by Martin Devera, later rewritten
16 	by Patrick McHardy and then maintained by Andre Correa.
17 
18 	You need the tc action  mirror or redirect to feed this device
19        	packets.
20 
21 
22   	Authors:	Jamal Hadi Salim (2005)
23 
24 */
25 
26 
27 #include <linux/module.h>
28 #include <linux/kernel.h>
29 #include <linux/netdevice.h>
30 #include <linux/etherdevice.h>
31 #include <linux/init.h>
32 #include <linux/interrupt.h>
33 #include <linux/moduleparam.h>
34 #include <net/pkt_sched.h>
35 #include <net/net_namespace.h>
36 
37 #define TX_Q_LIMIT    32
38 struct ifb_q_private {
39 	struct net_device	*dev;
40 	struct tasklet_struct   ifb_tasklet;
41 	int			tasklet_pending;
42 	int			txqnum;
43 	struct sk_buff_head     rq;
44 	u64			rx_packets;
45 	u64			rx_bytes;
46 	struct u64_stats_sync	rsync;
47 
48 	struct u64_stats_sync	tsync;
49 	u64			tx_packets;
50 	u64			tx_bytes;
51 	struct sk_buff_head     tq;
52 } ____cacheline_aligned_in_smp;
53 
54 struct ifb_dev_private {
55 	struct ifb_q_private *tx_private;
56 };
57 
58 static netdev_tx_t ifb_xmit(struct sk_buff *skb, struct net_device *dev);
59 static int ifb_open(struct net_device *dev);
60 static int ifb_close(struct net_device *dev);
61 
62 static void ifb_ri_tasklet(unsigned long _txp)
63 {
64 	struct ifb_q_private *txp = (struct ifb_q_private *)_txp;
65 	struct netdev_queue *txq;
66 	struct sk_buff *skb;
67 
68 	txq = netdev_get_tx_queue(txp->dev, txp->txqnum);
69 	skb = skb_peek(&txp->tq);
70 	if (!skb) {
71 		if (!__netif_tx_trylock(txq))
72 			goto resched;
73 		skb_queue_splice_tail_init(&txp->rq, &txp->tq);
74 		__netif_tx_unlock(txq);
75 	}
76 
77 	while ((skb = __skb_dequeue(&txp->tq)) != NULL) {
78 		skb->tc_redirected = 0;
79 		skb->tc_skip_classify = 1;
80 
81 		u64_stats_update_begin(&txp->tsync);
82 		txp->tx_packets++;
83 		txp->tx_bytes += skb->len;
84 		u64_stats_update_end(&txp->tsync);
85 
86 		rcu_read_lock();
87 		skb->dev = dev_get_by_index_rcu(dev_net(txp->dev), skb->skb_iif);
88 		if (!skb->dev) {
89 			rcu_read_unlock();
90 			dev_kfree_skb(skb);
91 			txp->dev->stats.tx_dropped++;
92 			if (skb_queue_len(&txp->tq) != 0)
93 				goto resched;
94 			break;
95 		}
96 		rcu_read_unlock();
97 		skb->skb_iif = txp->dev->ifindex;
98 
99 		if (!skb->tc_from_ingress) {
100 			dev_queue_xmit(skb);
101 		} else {
102 			skb_pull_rcsum(skb, skb->mac_len);
103 			netif_receive_skb(skb);
104 		}
105 	}
106 
107 	if (__netif_tx_trylock(txq)) {
108 		skb = skb_peek(&txp->rq);
109 		if (!skb) {
110 			txp->tasklet_pending = 0;
111 			if (netif_tx_queue_stopped(txq))
112 				netif_tx_wake_queue(txq);
113 		} else {
114 			__netif_tx_unlock(txq);
115 			goto resched;
116 		}
117 		__netif_tx_unlock(txq);
118 	} else {
119 resched:
120 		txp->tasklet_pending = 1;
121 		tasklet_schedule(&txp->ifb_tasklet);
122 	}
123 
124 }
125 
126 static void ifb_stats64(struct net_device *dev,
127 			struct rtnl_link_stats64 *stats)
128 {
129 	struct ifb_dev_private *dp = netdev_priv(dev);
130 	struct ifb_q_private *txp = dp->tx_private;
131 	unsigned int start;
132 	u64 packets, bytes;
133 	int i;
134 
135 	for (i = 0; i < dev->num_tx_queues; i++,txp++) {
136 		do {
137 			start = u64_stats_fetch_begin_irq(&txp->rsync);
138 			packets = txp->rx_packets;
139 			bytes = txp->rx_bytes;
140 		} while (u64_stats_fetch_retry_irq(&txp->rsync, start));
141 		stats->rx_packets += packets;
142 		stats->rx_bytes += bytes;
143 
144 		do {
145 			start = u64_stats_fetch_begin_irq(&txp->tsync);
146 			packets = txp->tx_packets;
147 			bytes = txp->tx_bytes;
148 		} while (u64_stats_fetch_retry_irq(&txp->tsync, start));
149 		stats->tx_packets += packets;
150 		stats->tx_bytes += bytes;
151 	}
152 	stats->rx_dropped = dev->stats.rx_dropped;
153 	stats->tx_dropped = dev->stats.tx_dropped;
154 }
155 
156 static int ifb_dev_init(struct net_device *dev)
157 {
158 	struct ifb_dev_private *dp = netdev_priv(dev);
159 	struct ifb_q_private *txp;
160 	int i;
161 
162 	txp = kcalloc(dev->num_tx_queues, sizeof(*txp), GFP_KERNEL);
163 	if (!txp)
164 		return -ENOMEM;
165 	dp->tx_private = txp;
166 	for (i = 0; i < dev->num_tx_queues; i++,txp++) {
167 		txp->txqnum = i;
168 		txp->dev = dev;
169 		__skb_queue_head_init(&txp->rq);
170 		__skb_queue_head_init(&txp->tq);
171 		u64_stats_init(&txp->rsync);
172 		u64_stats_init(&txp->tsync);
173 		tasklet_init(&txp->ifb_tasklet, ifb_ri_tasklet,
174 			     (unsigned long)txp);
175 		netif_tx_start_queue(netdev_get_tx_queue(dev, i));
176 	}
177 	return 0;
178 }
179 
180 static const struct net_device_ops ifb_netdev_ops = {
181 	.ndo_open	= ifb_open,
182 	.ndo_stop	= ifb_close,
183 	.ndo_get_stats64 = ifb_stats64,
184 	.ndo_start_xmit	= ifb_xmit,
185 	.ndo_validate_addr = eth_validate_addr,
186 	.ndo_init	= ifb_dev_init,
187 };
188 
189 #define IFB_FEATURES (NETIF_F_HW_CSUM | NETIF_F_SG  | NETIF_F_FRAGLIST	| \
190 		      NETIF_F_TSO_ECN | NETIF_F_TSO | NETIF_F_TSO6	| \
191 		      NETIF_F_GSO_ENCAP_ALL 				| \
192 		      NETIF_F_HIGHDMA | NETIF_F_HW_VLAN_CTAG_TX		| \
193 		      NETIF_F_HW_VLAN_STAG_TX)
194 
195 static void ifb_dev_free(struct net_device *dev)
196 {
197 	struct ifb_dev_private *dp = netdev_priv(dev);
198 	struct ifb_q_private *txp = dp->tx_private;
199 	int i;
200 
201 	for (i = 0; i < dev->num_tx_queues; i++,txp++) {
202 		tasklet_kill(&txp->ifb_tasklet);
203 		__skb_queue_purge(&txp->rq);
204 		__skb_queue_purge(&txp->tq);
205 	}
206 	kfree(dp->tx_private);
207 }
208 
209 static void ifb_setup(struct net_device *dev)
210 {
211 	/* Initialize the device structure. */
212 	dev->netdev_ops = &ifb_netdev_ops;
213 
214 	/* Fill in device structure with ethernet-generic values. */
215 	ether_setup(dev);
216 	dev->tx_queue_len = TX_Q_LIMIT;
217 
218 	dev->features |= IFB_FEATURES;
219 	dev->hw_features |= dev->features;
220 	dev->hw_enc_features |= dev->features;
221 	dev->vlan_features |= IFB_FEATURES & ~(NETIF_F_HW_VLAN_CTAG_TX |
222 					       NETIF_F_HW_VLAN_STAG_TX);
223 
224 	dev->flags |= IFF_NOARP;
225 	dev->flags &= ~IFF_MULTICAST;
226 	dev->priv_flags &= ~IFF_TX_SKB_SHARING;
227 	netif_keep_dst(dev);
228 	eth_hw_addr_random(dev);
229 	dev->needs_free_netdev = true;
230 	dev->priv_destructor = ifb_dev_free;
231 
232 	dev->min_mtu = 0;
233 	dev->max_mtu = 0;
234 }
235 
236 static netdev_tx_t ifb_xmit(struct sk_buff *skb, struct net_device *dev)
237 {
238 	struct ifb_dev_private *dp = netdev_priv(dev);
239 	struct ifb_q_private *txp = dp->tx_private + skb_get_queue_mapping(skb);
240 
241 	u64_stats_update_begin(&txp->rsync);
242 	txp->rx_packets++;
243 	txp->rx_bytes += skb->len;
244 	u64_stats_update_end(&txp->rsync);
245 
246 	if (!skb->tc_redirected || !skb->skb_iif) {
247 		dev_kfree_skb(skb);
248 		dev->stats.rx_dropped++;
249 		return NETDEV_TX_OK;
250 	}
251 
252 	if (skb_queue_len(&txp->rq) >= dev->tx_queue_len)
253 		netif_tx_stop_queue(netdev_get_tx_queue(dev, txp->txqnum));
254 
255 	__skb_queue_tail(&txp->rq, skb);
256 	if (!txp->tasklet_pending) {
257 		txp->tasklet_pending = 1;
258 		tasklet_schedule(&txp->ifb_tasklet);
259 	}
260 
261 	return NETDEV_TX_OK;
262 }
263 
264 static int ifb_close(struct net_device *dev)
265 {
266 	netif_tx_stop_all_queues(dev);
267 	return 0;
268 }
269 
270 static int ifb_open(struct net_device *dev)
271 {
272 	netif_tx_start_all_queues(dev);
273 	return 0;
274 }
275 
276 static int ifb_validate(struct nlattr *tb[], struct nlattr *data[],
277 			struct netlink_ext_ack *extack)
278 {
279 	if (tb[IFLA_ADDRESS]) {
280 		if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN)
281 			return -EINVAL;
282 		if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS])))
283 			return -EADDRNOTAVAIL;
284 	}
285 	return 0;
286 }
287 
288 static struct rtnl_link_ops ifb_link_ops __read_mostly = {
289 	.kind		= "ifb",
290 	.priv_size	= sizeof(struct ifb_dev_private),
291 	.setup		= ifb_setup,
292 	.validate	= ifb_validate,
293 };
294 
295 /* Number of ifb devices to be set up by this module.
296  * Note that these legacy devices have one queue.
297  * Prefer something like : ip link add ifb10 numtxqueues 8 type ifb
298  */
299 static int numifbs = 2;
300 module_param(numifbs, int, 0);
301 MODULE_PARM_DESC(numifbs, "Number of ifb devices");
302 
303 static int __init ifb_init_one(int index)
304 {
305 	struct net_device *dev_ifb;
306 	int err;
307 
308 	dev_ifb = alloc_netdev(sizeof(struct ifb_dev_private), "ifb%d",
309 			       NET_NAME_UNKNOWN, ifb_setup);
310 
311 	if (!dev_ifb)
312 		return -ENOMEM;
313 
314 	dev_ifb->rtnl_link_ops = &ifb_link_ops;
315 	err = register_netdevice(dev_ifb);
316 	if (err < 0)
317 		goto err;
318 
319 	return 0;
320 
321 err:
322 	free_netdev(dev_ifb);
323 	return err;
324 }
325 
326 static int __init ifb_init_module(void)
327 {
328 	int i, err;
329 
330 	down_write(&pernet_ops_rwsem);
331 	rtnl_lock();
332 	err = __rtnl_link_register(&ifb_link_ops);
333 	if (err < 0)
334 		goto out;
335 
336 	for (i = 0; i < numifbs && !err; i++) {
337 		err = ifb_init_one(i);
338 		cond_resched();
339 	}
340 	if (err)
341 		__rtnl_link_unregister(&ifb_link_ops);
342 
343 out:
344 	rtnl_unlock();
345 	up_write(&pernet_ops_rwsem);
346 
347 	return err;
348 }
349 
350 static void __exit ifb_cleanup_module(void)
351 {
352 	rtnl_link_unregister(&ifb_link_ops);
353 }
354 
355 module_init(ifb_init_module);
356 module_exit(ifb_cleanup_module);
357 MODULE_LICENSE("GPL");
358 MODULE_AUTHOR("Jamal Hadi Salim");
359 MODULE_ALIAS_RTNL_LINK("ifb");
360