xref: /openbmc/linux/drivers/net/ifb.c (revision 4beec1d7)
1 /* drivers/net/ifb.c:
2 
3 	The purpose of this driver is to provide a device that allows
4 	for sharing of resources:
5 
6 	1) qdiscs/policies that are per device as opposed to system wide.
7 	ifb allows for a device which can be redirected to thus providing
8 	an impression of sharing.
9 
10 	2) Allows for queueing incoming traffic for shaping instead of
11 	dropping.
12 
13 	The original concept is based on what is known as the IMQ
14 	driver initially written by Martin Devera, later rewritten
15 	by Patrick McHardy and then maintained by Andre Correa.
16 
17 	You need the tc action  mirror or redirect to feed this device
18        	packets.
19 
20 	This program is free software; you can redistribute it and/or
21 	modify it under the terms of the GNU General Public License
22 	as published by the Free Software Foundation; either version
23 	2 of the License, or (at your option) any later version.
24 
25   	Authors:	Jamal Hadi Salim (2005)
26 
27 */
28 
29 
30 #include <linux/module.h>
31 #include <linux/kernel.h>
32 #include <linux/netdevice.h>
33 #include <linux/etherdevice.h>
34 #include <linux/init.h>
35 #include <linux/interrupt.h>
36 #include <linux/moduleparam.h>
37 #include <net/pkt_sched.h>
38 #include <net/net_namespace.h>
39 
40 #define TX_Q_LIMIT    32
41 struct ifb_q_private {
42 	struct net_device	*dev;
43 	struct tasklet_struct   ifb_tasklet;
44 	int			tasklet_pending;
45 	int			txqnum;
46 	struct sk_buff_head     rq;
47 	u64			rx_packets;
48 	u64			rx_bytes;
49 	struct u64_stats_sync	rsync;
50 
51 	struct u64_stats_sync	tsync;
52 	u64			tx_packets;
53 	u64			tx_bytes;
54 	struct sk_buff_head     tq;
55 } ____cacheline_aligned_in_smp;
56 
57 struct ifb_dev_private {
58 	struct ifb_q_private *tx_private;
59 };
60 
61 static netdev_tx_t ifb_xmit(struct sk_buff *skb, struct net_device *dev);
62 static int ifb_open(struct net_device *dev);
63 static int ifb_close(struct net_device *dev);
64 
65 static void ifb_ri_tasklet(unsigned long _txp)
66 {
67 	struct ifb_q_private *txp = (struct ifb_q_private *)_txp;
68 	struct netdev_queue *txq;
69 	struct sk_buff *skb;
70 
71 	txq = netdev_get_tx_queue(txp->dev, txp->txqnum);
72 	skb = skb_peek(&txp->tq);
73 	if (!skb) {
74 		if (!__netif_tx_trylock(txq))
75 			goto resched;
76 		skb_queue_splice_tail_init(&txp->rq, &txp->tq);
77 		__netif_tx_unlock(txq);
78 	}
79 
80 	while ((skb = __skb_dequeue(&txp->tq)) != NULL) {
81 		u32 from = G_TC_FROM(skb->tc_verd);
82 
83 		skb->tc_verd = 0;
84 		skb->tc_verd = SET_TC_NCLS(skb->tc_verd);
85 
86 		u64_stats_update_begin(&txp->tsync);
87 		txp->tx_packets++;
88 		txp->tx_bytes += skb->len;
89 		u64_stats_update_end(&txp->tsync);
90 
91 		rcu_read_lock();
92 		skb->dev = dev_get_by_index_rcu(dev_net(txp->dev), skb->skb_iif);
93 		if (!skb->dev) {
94 			rcu_read_unlock();
95 			dev_kfree_skb(skb);
96 			txp->dev->stats.tx_dropped++;
97 			if (skb_queue_len(&txp->tq) != 0)
98 				goto resched;
99 			break;
100 		}
101 		rcu_read_unlock();
102 		skb->skb_iif = txp->dev->ifindex;
103 
104 		if (from & AT_EGRESS) {
105 			dev_queue_xmit(skb);
106 		} else if (from & AT_INGRESS) {
107 			skb_pull(skb, skb->mac_len);
108 			netif_receive_skb(skb);
109 		} else
110 			BUG();
111 	}
112 
113 	if (__netif_tx_trylock(txq)) {
114 		skb = skb_peek(&txp->rq);
115 		if (!skb) {
116 			txp->tasklet_pending = 0;
117 			if (netif_tx_queue_stopped(txq))
118 				netif_tx_wake_queue(txq);
119 		} else {
120 			__netif_tx_unlock(txq);
121 			goto resched;
122 		}
123 		__netif_tx_unlock(txq);
124 	} else {
125 resched:
126 		txp->tasklet_pending = 1;
127 		tasklet_schedule(&txp->ifb_tasklet);
128 	}
129 
130 }
131 
132 static struct rtnl_link_stats64 *ifb_stats64(struct net_device *dev,
133 					     struct rtnl_link_stats64 *stats)
134 {
135 	struct ifb_dev_private *dp = netdev_priv(dev);
136 	struct ifb_q_private *txp = dp->tx_private;
137 	unsigned int start;
138 	u64 packets, bytes;
139 	int i;
140 
141 	for (i = 0; i < dev->num_tx_queues; i++,txp++) {
142 		do {
143 			start = u64_stats_fetch_begin_irq(&txp->rsync);
144 			packets = txp->rx_packets;
145 			bytes = txp->rx_bytes;
146 		} while (u64_stats_fetch_retry_irq(&txp->rsync, start));
147 		stats->rx_packets += packets;
148 		stats->rx_bytes += bytes;
149 
150 		do {
151 			start = u64_stats_fetch_begin_irq(&txp->tsync);
152 			packets = txp->tx_packets;
153 			bytes = txp->tx_bytes;
154 		} while (u64_stats_fetch_retry_irq(&txp->tsync, start));
155 		stats->tx_packets += packets;
156 		stats->tx_bytes += bytes;
157 	}
158 	stats->rx_dropped = dev->stats.rx_dropped;
159 	stats->tx_dropped = dev->stats.tx_dropped;
160 
161 	return stats;
162 }
163 
164 static int ifb_dev_init(struct net_device *dev)
165 {
166 	struct ifb_dev_private *dp = netdev_priv(dev);
167 	struct ifb_q_private *txp;
168 	int i;
169 
170 	txp = kcalloc(dev->num_tx_queues, sizeof(*txp), GFP_KERNEL);
171 	if (!txp)
172 		return -ENOMEM;
173 	dp->tx_private = txp;
174 	for (i = 0; i < dev->num_tx_queues; i++,txp++) {
175 		txp->txqnum = i;
176 		txp->dev = dev;
177 		__skb_queue_head_init(&txp->rq);
178 		__skb_queue_head_init(&txp->tq);
179 		u64_stats_init(&txp->rsync);
180 		u64_stats_init(&txp->tsync);
181 		tasklet_init(&txp->ifb_tasklet, ifb_ri_tasklet,
182 			     (unsigned long)txp);
183 		netif_tx_start_queue(netdev_get_tx_queue(dev, i));
184 	}
185 	return 0;
186 }
187 
188 static const struct net_device_ops ifb_netdev_ops = {
189 	.ndo_open	= ifb_open,
190 	.ndo_stop	= ifb_close,
191 	.ndo_get_stats64 = ifb_stats64,
192 	.ndo_start_xmit	= ifb_xmit,
193 	.ndo_validate_addr = eth_validate_addr,
194 	.ndo_init	= ifb_dev_init,
195 };
196 
197 #define IFB_FEATURES (NETIF_F_HW_CSUM | NETIF_F_SG  | NETIF_F_FRAGLIST	| \
198 		      NETIF_F_TSO_ECN | NETIF_F_TSO | NETIF_F_TSO6	| \
199 		      NETIF_F_GSO_ENCAP_ALL 				| \
200 		      NETIF_F_HIGHDMA | NETIF_F_HW_VLAN_CTAG_TX		| \
201 		      NETIF_F_HW_VLAN_STAG_TX)
202 
203 static void ifb_dev_free(struct net_device *dev)
204 {
205 	struct ifb_dev_private *dp = netdev_priv(dev);
206 	struct ifb_q_private *txp = dp->tx_private;
207 	int i;
208 
209 	for (i = 0; i < dev->num_tx_queues; i++,txp++) {
210 		tasklet_kill(&txp->ifb_tasklet);
211 		__skb_queue_purge(&txp->rq);
212 		__skb_queue_purge(&txp->tq);
213 	}
214 	kfree(dp->tx_private);
215 	free_netdev(dev);
216 }
217 
218 static void ifb_setup(struct net_device *dev)
219 {
220 	/* Initialize the device structure. */
221 	dev->netdev_ops = &ifb_netdev_ops;
222 
223 	/* Fill in device structure with ethernet-generic values. */
224 	ether_setup(dev);
225 	dev->tx_queue_len = TX_Q_LIMIT;
226 
227 	dev->features |= IFB_FEATURES;
228 	dev->hw_features |= dev->features;
229 	dev->hw_enc_features |= dev->features;
230 	dev->vlan_features |= IFB_FEATURES & ~(NETIF_F_HW_VLAN_CTAG_TX |
231 					       NETIF_F_HW_VLAN_STAG_TX);
232 
233 	dev->flags |= IFF_NOARP;
234 	dev->flags &= ~IFF_MULTICAST;
235 	dev->priv_flags &= ~IFF_TX_SKB_SHARING;
236 	netif_keep_dst(dev);
237 	eth_hw_addr_random(dev);
238 	dev->destructor = ifb_dev_free;
239 }
240 
241 static netdev_tx_t ifb_xmit(struct sk_buff *skb, struct net_device *dev)
242 {
243 	struct ifb_dev_private *dp = netdev_priv(dev);
244 	u32 from = G_TC_FROM(skb->tc_verd);
245 	struct ifb_q_private *txp = dp->tx_private + skb_get_queue_mapping(skb);
246 
247 	u64_stats_update_begin(&txp->rsync);
248 	txp->rx_packets++;
249 	txp->rx_bytes += skb->len;
250 	u64_stats_update_end(&txp->rsync);
251 
252 	if (!(from & (AT_INGRESS|AT_EGRESS)) || !skb->skb_iif) {
253 		dev_kfree_skb(skb);
254 		dev->stats.rx_dropped++;
255 		return NETDEV_TX_OK;
256 	}
257 
258 	if (skb_queue_len(&txp->rq) >= dev->tx_queue_len)
259 		netif_tx_stop_queue(netdev_get_tx_queue(dev, txp->txqnum));
260 
261 	__skb_queue_tail(&txp->rq, skb);
262 	if (!txp->tasklet_pending) {
263 		txp->tasklet_pending = 1;
264 		tasklet_schedule(&txp->ifb_tasklet);
265 	}
266 
267 	return NETDEV_TX_OK;
268 }
269 
270 static int ifb_close(struct net_device *dev)
271 {
272 	netif_tx_stop_all_queues(dev);
273 	return 0;
274 }
275 
276 static int ifb_open(struct net_device *dev)
277 {
278 	netif_tx_start_all_queues(dev);
279 	return 0;
280 }
281 
282 static int ifb_validate(struct nlattr *tb[], struct nlattr *data[])
283 {
284 	if (tb[IFLA_ADDRESS]) {
285 		if (nla_len(tb[IFLA_ADDRESS]) != ETH_ALEN)
286 			return -EINVAL;
287 		if (!is_valid_ether_addr(nla_data(tb[IFLA_ADDRESS])))
288 			return -EADDRNOTAVAIL;
289 	}
290 	return 0;
291 }
292 
293 static struct rtnl_link_ops ifb_link_ops __read_mostly = {
294 	.kind		= "ifb",
295 	.priv_size	= sizeof(struct ifb_dev_private),
296 	.setup		= ifb_setup,
297 	.validate	= ifb_validate,
298 };
299 
300 /* Number of ifb devices to be set up by this module.
301  * Note that these legacy devices have one queue.
302  * Prefer something like : ip link add ifb10 numtxqueues 8 type ifb
303  */
304 static int numifbs = 2;
305 module_param(numifbs, int, 0);
306 MODULE_PARM_DESC(numifbs, "Number of ifb devices");
307 
308 static int __init ifb_init_one(int index)
309 {
310 	struct net_device *dev_ifb;
311 	int err;
312 
313 	dev_ifb = alloc_netdev(sizeof(struct ifb_dev_private), "ifb%d",
314 			       NET_NAME_UNKNOWN, ifb_setup);
315 
316 	if (!dev_ifb)
317 		return -ENOMEM;
318 
319 	dev_ifb->rtnl_link_ops = &ifb_link_ops;
320 	err = register_netdevice(dev_ifb);
321 	if (err < 0)
322 		goto err;
323 
324 	return 0;
325 
326 err:
327 	free_netdev(dev_ifb);
328 	return err;
329 }
330 
331 static int __init ifb_init_module(void)
332 {
333 	int i, err;
334 
335 	rtnl_lock();
336 	err = __rtnl_link_register(&ifb_link_ops);
337 	if (err < 0)
338 		goto out;
339 
340 	for (i = 0; i < numifbs && !err; i++) {
341 		err = ifb_init_one(i);
342 		cond_resched();
343 	}
344 	if (err)
345 		__rtnl_link_unregister(&ifb_link_ops);
346 
347 out:
348 	rtnl_unlock();
349 
350 	return err;
351 }
352 
353 static void __exit ifb_cleanup_module(void)
354 {
355 	rtnl_link_unregister(&ifb_link_ops);
356 }
357 
358 module_init(ifb_init_module);
359 module_exit(ifb_cleanup_module);
360 MODULE_LICENSE("GPL");
361 MODULE_AUTHOR("Jamal Hadi Salim");
362 MODULE_ALIAS_RTNL_LINK("ifb");
363