1/*
2 * Equalizer Load-balancer for serial network interfaces.
3 *
4 * (c) Copyright 1995 Simon "Guru Aleph-Null" Janes
5 * NCM: Network and Communications Management, Inc.
6 *
7 * (c) Copyright 2002 David S. Miller (davem@redhat.com)
8 *
9 * This software may be used and distributed according to the terms
10 * of the GNU General Public License, incorporated herein by reference.
11 *
12 * The author may be reached as simon@ncm.com, or C/O
13 * NCM
14 * Attn: Simon Janes
15 * 6803 Whittier Ave
16 * McLean VA 22101
17 * Phone: 1-703-847-0040 ext 103
18 */
19
20/*
21 * Sources:
22 * skeleton.c by Donald Becker.
23 * Inspirations:
24 * The Harried and Overworked Alan Cox
25 * Conspiracies:
26 * The Alan Cox and Mike McLagan plot to get someone else to do the code,
27 * which turned out to be me.
28 */
29
30/*
31 * $Log: eql.c,v $
32 * Revision 1.2 1996/04/11 17:51:52 guru
33 * Added one-line eql_remove_slave patch.
34 *
35 * Revision 1.1 1996/04/11 17:44:17 guru
36 * Initial revision
37 *
38 * Revision 3.13 1996/01/21 15:17:18 alan
39 * tx_queue_len changes.
40 * reformatted.
41 *
42 * Revision 3.12 1995/03/22 21:07:51 anarchy
43 * Added capable() checks on configuration.
44 * Moved header file.
45 *
46 * Revision 3.11 1995/01/19 23:14:31 guru
47 * slave_load = (ULONG_MAX - (ULONG_MAX / 2)) -
48 * (priority_Bps) + bytes_queued * 8;
49 *
50 * Revision 3.10 1995/01/19 23:07:53 guru
51 * back to
52 * slave_load = (ULONG_MAX - (ULONG_MAX / 2)) -
53 * (priority_Bps) + bytes_queued;
54 *
55 * Revision 3.9 1995/01/19 22:38:20 guru
56 * slave_load = (ULONG_MAX - (ULONG_MAX / 2)) -
57 * (priority_Bps) + bytes_queued * 4;
58 *
59 * Revision 3.8 1995/01/19 22:30:55 guru
60 * slave_load = (ULONG_MAX - (ULONG_MAX / 2)) -
61 * (priority_Bps) + bytes_queued * 2;
62 *
63 * Revision 3.7 1995/01/19 21:52:35 guru
64 * printk's trimmed out.
65 *
66 * Revision 3.6 1995/01/19 21:49:56 guru
67 * This is working pretty well. I gained 1 K/s in speed.. now it's just
68 * robustness and printk's to be diked out.
69 *
70 * Revision 3.5 1995/01/18 22:29:59 guru
71 * still crashes the kernel when the lock_wait thing is woken up.
72 *
73 * Revision 3.4 1995/01/18 21:59:47 guru
74 * Broken set-bit locking snapshot
75 *
76 * Revision 3.3 1995/01/17 22:09:18 guru
77 * infinite sleep in a lock somewhere..
78 *
79 * Revision 3.2 1995/01/15 16:46:06 guru
80 * Log trimmed of non-pertinent 1.x branch messages
81 *
82 * Revision 3.1 1995/01/15 14:41:45 guru
83 * New Scheduler and timer stuff...
84 *
85 * Revision 1.15 1995/01/15 14:29:02 guru
86 * Will make 1.14 (now 1.15) the 3.0 branch, and the 1.12 the 2.0 branch, the one
87 * with the dumber scheduler
88 *
89 * Revision 1.14 1995/01/15 02:37:08 guru
90 * shock.. the kept-new-versions could have zonked working
91 * stuff.. shudder
92 *
93 * Revision 1.13 1995/01/15 02:36:31 guru
94 * big changes
95 *
96 * scheduler was torn out and replaced with something smarter
97 *
98 * global names not prefixed with eql_ were renamed to protect
99 * against namespace collisions
100 *
101 * a few more abstract interfaces were added to facilitate any
102 * potential change of datastructure. the driver is still using
103 * a linked list of slaves. going to a heap would be a bit of
104 * an overkill.
105 *
106 * this compiles fine with no warnings.
107 *
108 * the locking mechanism and timer stuff must be written however,
109 * this version will not work otherwise
110 *
111 * Sorry, I had to rewrite most of this for 2.5.x -DaveM
112 */
113
114#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
115
116#include <linux/capability.h>
117#include <linux/module.h>
118#include <linux/kernel.h>
119#include <linux/init.h>
120#include <linux/slab.h>
121#include <linux/timer.h>
122#include <linux/netdevice.h>
123#include <net/net_namespace.h>
124
125#include <linux/if.h>
126#include <linux/if_arp.h>
127#include <linux/if_eql.h>
128#include <linux/pkt_sched.h>
129
130#include <linux/uaccess.h>
131
132static int eql_open(struct net_device *dev);
133static int eql_close(struct net_device *dev);
134static int eql_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd);
135static netdev_tx_t eql_slave_xmit(struct sk_buff *skb, struct net_device *dev);
136
137#define eql_is_slave(dev) ((dev->flags & IFF_SLAVE) == IFF_SLAVE)
138#define eql_is_master(dev) ((dev->flags & IFF_MASTER) == IFF_MASTER)
139
140static void eql_kill_one_slave(slave_queue_t *queue, slave_t *slave);
141
142static void eql_timer(struct timer_list *t)
143{
144 equalizer_t *eql = from_timer(eql, t, timer);
145 struct list_head *this, *tmp, *head;
146
147 spin_lock(&eql->queue.lock);
148 head = &eql->queue.all_slaves;
149 list_for_each_safe(this, tmp, head) {
150 slave_t *slave = list_entry(this, slave_t, list);
151
152 if ((slave->dev->flags & IFF_UP) == IFF_UP) {
153 slave->bytes_queued -= slave->priority_Bps;
154 if (slave->bytes_queued < 0)
155 slave->bytes_queued = 0;
156 } else {
157 eql_kill_one_slave(&eql->queue, slave);
158 }
159
160 }
161 spin_unlock(&eql->queue.lock);
162
163 eql->timer.expires = jiffies + EQL_DEFAULT_RESCHED_IVAL;
164 add_timer(&eql->timer);
165}
166
167static const char version[] __initconst =
168 "Equalizer2002: Simon Janes (simon@ncm.com) and David S. Miller (davem@redhat.com)";
169
170static const struct net_device_ops eql_netdev_ops = {
171 .ndo_open = eql_open,
172 .ndo_stop = eql_close,
173 .ndo_do_ioctl = eql_ioctl,
174 .ndo_start_xmit = eql_slave_xmit,
175};
176
177static void __init eql_setup(struct net_device *dev)
178{
179 equalizer_t *eql = netdev_priv(dev);
180
181 timer_setup(&eql->timer, eql_timer, 0);
182 eql->timer.expires = jiffies + EQL_DEFAULT_RESCHED_IVAL;
183
184 spin_lock_init(&eql->queue.lock);
185 INIT_LIST_HEAD(&eql->queue.all_slaves);
186 eql->queue.master_dev = dev;
187
188 dev->netdev_ops = &eql_netdev_ops;
189
190 /*
191 * Now we undo some of the things that eth_setup does
192 * that we don't like
193 */
194
195 dev->mtu = EQL_DEFAULT_MTU; /* set to 576 in if_eql.h */
196 dev->flags = IFF_MASTER;
197
198 dev->type = ARPHRD_SLIP;
199 dev->tx_queue_len = 5; /* Hands them off fast */
200 netif_keep_dst(dev);
201}
202
203static int eql_open(struct net_device *dev)
204{
205 equalizer_t *eql = netdev_priv(dev);
206
207 /* XXX We should force this off automatically for the user. */
208 netdev_info(dev,
209 "remember to turn off Van-Jacobson compression on your slave devices\n");
210
211 BUG_ON(!list_empty(&eql->queue.all_slaves));
212
213 eql->min_slaves = 1;
214 eql->max_slaves = EQL_DEFAULT_MAX_SLAVES; /* 4 usually... */
215
216 add_timer(&eql->timer);
217
218 return 0;
219}
220
221static void eql_kill_one_slave(slave_queue_t *queue, slave_t *slave)
222{
223 list_del(&slave->list);
224 queue->num_slaves--;
225 slave->dev->flags &= ~IFF_SLAVE;
226 dev_put(slave->dev);
227 kfree(slave);
228}
229
230static void eql_kill_slave_queue(slave_queue_t *queue)
231{
232 struct list_head *head, *tmp, *this;
233
234 spin_lock_bh(&queue->lock);
235
236 head = &queue->all_slaves;
237 list_for_each_safe(this, tmp, head) {
238 slave_t *s = list_entry(this, slave_t, list);
239
240 eql_kill_one_slave(queue, s);
241 }
242
243 spin_unlock_bh(&queue->lock);
244}
245
246static int eql_close(struct net_device *dev)
247{
248 equalizer_t *eql = netdev_priv(dev);
249
250 /*
251 * The timer has to be stopped first before we start hacking away
252 * at the data structure it scans every so often...
253 */
254
255 del_timer_sync(&eql->timer);
256
257 eql_kill_slave_queue(&eql->queue);
258
259 return 0;
260}
261
262static int eql_enslave(struct net_device *dev, slaving_request_t __user *srq);
263static int eql_emancipate(struct net_device *dev, slaving_request_t __user *srq);
264
265static int eql_g_slave_cfg(struct net_device *dev, slave_config_t __user *sc);
266static int eql_s_slave_cfg(struct net_device *dev, slave_config_t __user *sc);
267
268static int eql_g_master_cfg(struct net_device *dev, master_config_t __user *mc);
269static int eql_s_master_cfg(struct net_device *dev, master_config_t __user *mc);
270
271static int eql_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd)
272{
273 if (cmd != EQL_GETMASTRCFG && cmd != EQL_GETSLAVECFG &&
274 !capable(CAP_NET_ADMIN))
275 return -EPERM;
276
277 switch (cmd) {
278 case EQL_ENSLAVE:
279 return eql_enslave(dev, ifr->ifr_data);
280 case EQL_EMANCIPATE:
281 return eql_emancipate(dev, ifr->ifr_data);
282 case EQL_GETSLAVECFG:
283 return eql_g_slave_cfg(dev, ifr->ifr_data);
284 case EQL_SETSLAVECFG:
285 return eql_s_slave_cfg(dev, ifr->ifr_data);
286 case EQL_GETMASTRCFG:
287 return eql_g_master_cfg(dev, ifr->ifr_data);
288 case EQL_SETMASTRCFG:
289 return eql_s_master_cfg(dev, ifr->ifr_data);
290 default:
291 return -EOPNOTSUPP;
292 }
293}
294
295/* queue->lock must be held */
296static slave_t *__eql_schedule_slaves(slave_queue_t *queue)
297{
298 unsigned long best_load = ~0UL;
299 struct list_head *this, *tmp, *head;
300 slave_t *best_slave;
301
302 best_slave = NULL;
303
304 /* Make a pass to set the best slave. */
305 head = &queue->all_slaves;
306 list_for_each_safe(this, tmp, head) {
307 slave_t *slave = list_entry(this, slave_t, list);
308 unsigned long slave_load, bytes_queued, priority_Bps;
309
310 /* Go through the slave list once, updating best_slave
311 * whenever a new best_load is found.
312 */
313 bytes_queued = slave->bytes_queued;
314 priority_Bps = slave->priority_Bps;
315 if ((slave->dev->flags & IFF_UP) == IFF_UP) {
316 slave_load = (~0UL - (~0UL / 2)) -
317 (priority_Bps) + bytes_queued * 8;
318
319 if (slave_load < best_load) {
320 best_load = slave_load;
321 best_slave = slave;
322 }
323 } else {
324 /* We found a dead slave, kill it. */
325 eql_kill_one_slave(queue, slave);
326 }
327 }
328 return best_slave;
329}
330
331static netdev_tx_t eql_slave_xmit(struct sk_buff *skb, struct net_device *dev)
332{
333 equalizer_t *eql = netdev_priv(dev);
334 slave_t *slave;
335
336 spin_lock(&eql->queue.lock);
337
338 slave = __eql_schedule_slaves(&eql->queue);
339 if (slave) {
340 struct net_device *slave_dev = slave->dev;
341
342 skb->dev = slave_dev;
343 skb->priority = TC_PRIO_FILLER;
344 slave->bytes_queued += skb->len;
345 dev_queue_xmit(skb);
346 dev->stats.tx_packets++;
347 } else {
348 dev->stats.tx_dropped++;
349 dev_kfree_skb(skb);
350 }
351
352 spin_unlock(&eql->queue.lock);
353
354 return NETDEV_TX_OK;
355}
356
357/*
358 * Private ioctl functions
359 */
360
361/* queue->lock must be held */
362static slave_t *__eql_find_slave_dev(slave_queue_t *queue, struct net_device *dev)
363{
364 struct list_head *this, *head;
365
366 head = &queue->all_slaves;
367 list_for_each(this, head) {
368 slave_t *slave = list_entry(this, slave_t, list);
369
370 if (slave->dev == dev)
371 return slave;
372 }
373
374 return NULL;
375}
376
377static inline int eql_is_full(slave_queue_t *queue)
378{
379 equalizer_t *eql = netdev_priv(queue->master_dev);
380
381 if (queue->num_slaves >= eql->max_slaves)
382 return 1;
383 return 0;
384}
385
386/* queue->lock must be held */
387static int __eql_insert_slave(slave_queue_t *queue, slave_t *slave)
388{
389 if (!eql_is_full(queue)) {
390 slave_t *duplicate_slave = NULL;
391
392 duplicate_slave = __eql_find_slave_dev(queue, slave->dev);
393 if (duplicate_slave)
394 eql_kill_one_slave(queue, duplicate_slave);
395
396 dev_hold(slave->dev);
397 list_add(&slave->list, &queue->all_slaves);
398 queue->num_slaves++;
399 slave->dev->flags |= IFF_SLAVE;
400
401 return 0;
402 }
403
404 return -ENOSPC;
405}
406
407static int eql_enslave(struct net_device *master_dev, slaving_request_t __user *srqp)
408{
409 struct net_device *slave_dev;
410 slaving_request_t srq;
411
412 if (copy_from_user(&srq, srqp, sizeof (slaving_request_t)))
413 return -EFAULT;
414
415 slave_dev = __dev_get_by_name(&init_net, srq.slave_name);
416 if (!slave_dev)
417 return -ENODEV;
418
419 if ((master_dev->flags & IFF_UP) == IFF_UP) {
420 /* slave is not a master & not already a slave: */
421 if (!eql_is_master(slave_dev) && !eql_is_slave(slave_dev)) {
422 slave_t *s = kmalloc(sizeof(*s), GFP_KERNEL);
423 equalizer_t *eql = netdev_priv(master_dev);
424 int ret;
425
426 if (!s)
427 return -ENOMEM;
428
429 memset(s, 0, sizeof(*s));
430 s->dev = slave_dev;
431 s->priority = srq.priority;
432 s->priority_bps = srq.priority;
433 s->priority_Bps = srq.priority / 8;
434
435 spin_lock_bh(&eql->queue.lock);
436 ret = __eql_insert_slave(&eql->queue, s);
437 if (ret)
438 kfree(s);
439
440 spin_unlock_bh(&eql->queue.lock);
441
442 return ret;
443 }
444 }
445
446 return -EINVAL;
447}
448
449static int eql_emancipate(struct net_device *master_dev, slaving_request_t __user *srqp)
450{
451 equalizer_t *eql = netdev_priv(master_dev);
452 struct net_device *slave_dev;
453 slaving_request_t srq;
454 int ret;
455
456 if (copy_from_user(&srq, srqp, sizeof (slaving_request_t)))
457 return -EFAULT;
458
459 slave_dev = __dev_get_by_name(&init_net, srq.slave_name);
460 if (!slave_dev)
461 return -ENODEV;
462
463 ret = -EINVAL;
464 spin_lock_bh(&eql->queue.lock);
465 if (eql_is_slave(slave_dev)) {
466 slave_t *slave = __eql_find_slave_dev(&eql->queue, slave_dev);
467 if (slave) {
468 eql_kill_one_slave(&eql->queue, slave);
469 ret = 0;
470 }
471 }
472 spin_unlock_bh(&eql->queue.lock);
473
474 return ret;
475}
476
477static int eql_g_slave_cfg(struct net_device *dev, slave_config_t __user *scp)
478{
479 equalizer_t *eql = netdev_priv(dev);
480 slave_t *slave;
481 struct net_device *slave_dev;
482 slave_config_t sc;
483 int ret;
484
485 if (copy_from_user(&sc, scp, sizeof (slave_config_t)))
486 return -EFAULT;
487
488 slave_dev = __dev_get_by_name(&init_net, sc.slave_name);
489 if (!slave_dev)
490 return -ENODEV;
491
492 ret = -EINVAL;
493
494 spin_lock_bh(&eql->queue.lock);
495 if (eql_is_slave(slave_dev)) {
496 slave = __eql_find_slave_dev(&eql->queue, slave_dev);
497 if (slave) {
498 sc.priority = slave->priority;
499 ret = 0;
500 }
501 }
502 spin_unlock_bh(&eql->queue.lock);
503
504 if (!ret && copy_to_user(scp, &sc, sizeof (slave_config_t)))
505 ret = -EFAULT;
506
507 return ret;
508}
509
510static int eql_s_slave_cfg(struct net_device *dev, slave_config_t __user *scp)
511{
512 slave_t *slave;
513 equalizer_t *eql;
514 struct net_device *slave_dev;
515 slave_config_t sc;
516 int ret;
517
518 if (copy_from_user(&sc, scp, sizeof (slave_config_t)))
519 return -EFAULT;
520
521 slave_dev = __dev_get_by_name(&init_net, sc.slave_name);
522 if (!slave_dev)
523 return -ENODEV;
524
525 ret = -EINVAL;
526
527 eql = netdev_priv(dev);
528 spin_lock_bh(&eql->queue.lock);
529 if (eql_is_slave(slave_dev)) {
530 slave = __eql_find_slave_dev(&eql->queue, slave_dev);
531 if (slave) {
532 slave->priority = sc.priority;
533 slave->priority_bps = sc.priority;
534 slave->priority_Bps = sc.priority / 8;
535 ret = 0;
536 }
537 }
538 spin_unlock_bh(&eql->queue.lock);
539
540 return ret;
541}
542
543static int eql_g_master_cfg(struct net_device *dev, master_config_t __user *mcp)
544{
545 equalizer_t *eql;
546 master_config_t mc;
547
548 memset(&mc, 0, sizeof(master_config_t));
549
550 if (eql_is_master(dev)) {
551 eql = netdev_priv(dev);
552 mc.max_slaves = eql->max_slaves;
553 mc.min_slaves = eql->min_slaves;
554 if (copy_to_user(mcp, &mc, sizeof (master_config_t)))
555 return -EFAULT;
556 return 0;
557 }
558 return -EINVAL;
559}
560
561static int eql_s_master_cfg(struct net_device *dev, master_config_t __user *mcp)
562{
563 equalizer_t *eql;
564 master_config_t mc;
565
566 if (copy_from_user(&mc, mcp, sizeof (master_config_t)))
567 return -EFAULT;
568
569 if (eql_is_master(dev)) {
570 eql = netdev_priv(dev);
571 eql->max_slaves = mc.max_slaves;
572 eql->min_slaves = mc.min_slaves;
573 return 0;
574 }
575 return -EINVAL;
576}
577
578static struct net_device *dev_eql;
579
580static int __init eql_init_module(void)
581{
582 int err;
583
584 pr_info("%s\n", version);
585
586 dev_eql = alloc_netdev(sizeof(equalizer_t), "eql", NET_NAME_UNKNOWN,
587 eql_setup);
588 if (!dev_eql)
589 return -ENOMEM;
590
591 err = register_netdev(dev_eql);
592 if (err)
593 free_netdev(dev_eql);
594 return err;
595}
596
597static void __exit eql_cleanup_module(void)
598{
599 unregister_netdev(dev_eql);
600 free_netdev(dev_eql);
601}
602
603module_init(eql_init_module);
604module_exit(eql_cleanup_module);
605MODULE_LICENSE("GPL");
606