eql.c 15 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610
  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. * Sources:
  21. * skeleton.c by Donald Becker.
  22. * Inspirations:
  23. * The Harried and Overworked Alan Cox
  24. * Conspiracies:
  25. * The Alan Cox and Mike McLagan plot to get someone else to do the code,
  26. * which turned out to be me.
  27. */
  28. /*
  29. * $Log: eql.c,v $
  30. * Revision 1.2 1996/04/11 17:51:52 guru
  31. * Added one-line eql_remove_slave patch.
  32. *
  33. * Revision 1.1 1996/04/11 17:44:17 guru
  34. * Initial revision
  35. *
  36. * Revision 3.13 1996/01/21 15:17:18 alan
  37. * tx_queue_len changes.
  38. * reformatted.
  39. *
  40. * Revision 3.12 1995/03/22 21:07:51 anarchy
  41. * Added capable() checks on configuration.
  42. * Moved header file.
  43. *
  44. * Revision 3.11 1995/01/19 23:14:31 guru
  45. * slave_load = (ULONG_MAX - (ULONG_MAX / 2)) -
  46. * (priority_Bps) + bytes_queued * 8;
  47. *
  48. * Revision 3.10 1995/01/19 23:07:53 guru
  49. * back to
  50. * slave_load = (ULONG_MAX - (ULONG_MAX / 2)) -
  51. * (priority_Bps) + bytes_queued;
  52. *
  53. * Revision 3.9 1995/01/19 22:38:20 guru
  54. * slave_load = (ULONG_MAX - (ULONG_MAX / 2)) -
  55. * (priority_Bps) + bytes_queued * 4;
  56. *
  57. * Revision 3.8 1995/01/19 22:30:55 guru
  58. * slave_load = (ULONG_MAX - (ULONG_MAX / 2)) -
  59. * (priority_Bps) + bytes_queued * 2;
  60. *
  61. * Revision 3.7 1995/01/19 21:52:35 guru
  62. * printk's trimmed out.
  63. *
  64. * Revision 3.6 1995/01/19 21:49:56 guru
  65. * This is working pretty well. I gained 1 K/s in speed.. now it's just
  66. * robustness and printk's to be diked out.
  67. *
  68. * Revision 3.5 1995/01/18 22:29:59 guru
  69. * still crashes the kernel when the lock_wait thing is woken up.
  70. *
  71. * Revision 3.4 1995/01/18 21:59:47 guru
  72. * Broken set-bit locking snapshot
  73. *
  74. * Revision 3.3 1995/01/17 22:09:18 guru
  75. * infinite sleep in a lock somewhere..
  76. *
  77. * Revision 3.2 1995/01/15 16:46:06 guru
  78. * Log trimmed of non-pertinent 1.x branch messages
  79. *
  80. * Revision 3.1 1995/01/15 14:41:45 guru
  81. * New Scheduler and timer stuff...
  82. *
  83. * Revision 1.15 1995/01/15 14:29:02 guru
  84. * Will make 1.14 (now 1.15) the 3.0 branch, and the 1.12 the 2.0 branch, the one
  85. * with the dumber scheduler
  86. *
  87. * Revision 1.14 1995/01/15 02:37:08 guru
  88. * shock.. the kept-new-versions could have zonked working
  89. * stuff.. shudder
  90. *
  91. * Revision 1.13 1995/01/15 02:36:31 guru
  92. * big changes
  93. *
  94. * scheduler was torn out and replaced with something smarter
  95. *
  96. * global names not prefixed with eql_ were renamed to protect
  97. * against namespace collisions
  98. *
  99. * a few more abstract interfaces were added to facilitate any
  100. * potential change of datastructure. the driver is still using
  101. * a linked list of slaves. going to a heap would be a bit of
  102. * an overkill.
  103. *
  104. * this compiles fine with no warnings.
  105. *
  106. * the locking mechanism and timer stuff must be written however,
  107. * this version will not work otherwise
  108. *
  109. * Sorry, I had to rewrite most of this for 2.5.x -DaveM
  110. */
  111. #include <linux/module.h>
  112. #include <linux/kernel.h>
  113. #include <linux/init.h>
  114. #include <linux/timer.h>
  115. #include <linux/netdevice.h>
  116. #include <net/net_namespace.h>
  117. #include <linux/if.h>
  118. #include <linux/if_arp.h>
  119. #include <linux/if_eql.h>
  120. #include <asm/uaccess.h>
  121. static int eql_open(struct net_device *dev);
  122. static int eql_close(struct net_device *dev);
  123. static int eql_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd);
  124. static int eql_slave_xmit(struct sk_buff *skb, struct net_device *dev);
  125. #define eql_is_slave(dev) ((dev->flags & IFF_SLAVE) == IFF_SLAVE)
  126. #define eql_is_master(dev) ((dev->flags & IFF_MASTER) == IFF_MASTER)
  127. static void eql_kill_one_slave(slave_queue_t *queue, slave_t *slave);
  128. static void eql_timer(unsigned long param)
  129. {
  130. equalizer_t *eql = (equalizer_t *) param;
  131. struct list_head *this, *tmp, *head;
  132. spin_lock_bh(&eql->queue.lock);
  133. head = &eql->queue.all_slaves;
  134. list_for_each_safe(this, tmp, head) {
  135. slave_t *slave = list_entry(this, slave_t, list);
  136. if ((slave->dev->flags & IFF_UP) == IFF_UP) {
  137. slave->bytes_queued -= slave->priority_Bps;
  138. if (slave->bytes_queued < 0)
  139. slave->bytes_queued = 0;
  140. } else {
  141. eql_kill_one_slave(&eql->queue, slave);
  142. }
  143. }
  144. spin_unlock_bh(&eql->queue.lock);
  145. eql->timer.expires = jiffies + EQL_DEFAULT_RESCHED_IVAL;
  146. add_timer(&eql->timer);
  147. }
  148. static const char version[] __initconst =
  149. "Equalizer2002: Simon Janes (simon@ncm.com) and David S. Miller (davem@redhat.com)\n";
  150. static const struct net_device_ops eql_netdev_ops = {
  151. .ndo_open = eql_open,
  152. .ndo_stop = eql_close,
  153. .ndo_do_ioctl = eql_ioctl,
  154. .ndo_start_xmit = eql_slave_xmit,
  155. };
  156. static void __init eql_setup(struct net_device *dev)
  157. {
  158. equalizer_t *eql = netdev_priv(dev);
  159. init_timer(&eql->timer);
  160. eql->timer.data = (unsigned long) eql;
  161. eql->timer.expires = jiffies + EQL_DEFAULT_RESCHED_IVAL;
  162. eql->timer.function = eql_timer;
  163. spin_lock_init(&eql->queue.lock);
  164. INIT_LIST_HEAD(&eql->queue.all_slaves);
  165. eql->queue.master_dev = dev;
  166. dev->netdev_ops = &eql_netdev_ops;
  167. /*
  168. * Now we undo some of the things that eth_setup does
  169. * that we don't like
  170. */
  171. dev->mtu = EQL_DEFAULT_MTU; /* set to 576 in if_eql.h */
  172. dev->flags = IFF_MASTER;
  173. dev->type = ARPHRD_SLIP;
  174. dev->tx_queue_len = 5; /* Hands them off fast */
  175. dev->priv_flags &= ~IFF_XMIT_DST_RELEASE;
  176. }
  177. static int eql_open(struct net_device *dev)
  178. {
  179. equalizer_t *eql = netdev_priv(dev);
  180. /* XXX We should force this off automatically for the user. */
  181. printk(KERN_INFO "%s: remember to turn off Van-Jacobson compression on "
  182. "your slave devices.\n", dev->name);
  183. BUG_ON(!list_empty(&eql->queue.all_slaves));
  184. eql->min_slaves = 1;
  185. eql->max_slaves = EQL_DEFAULT_MAX_SLAVES; /* 4 usually... */
  186. add_timer(&eql->timer);
  187. return 0;
  188. }
  189. static void eql_kill_one_slave(slave_queue_t *queue, slave_t *slave)
  190. {
  191. list_del(&slave->list);
  192. queue->num_slaves--;
  193. slave->dev->flags &= ~IFF_SLAVE;
  194. dev_put(slave->dev);
  195. kfree(slave);
  196. }
  197. static void eql_kill_slave_queue(slave_queue_t *queue)
  198. {
  199. struct list_head *head, *tmp, *this;
  200. spin_lock_bh(&queue->lock);
  201. head = &queue->all_slaves;
  202. list_for_each_safe(this, tmp, head) {
  203. slave_t *s = list_entry(this, slave_t, list);
  204. eql_kill_one_slave(queue, s);
  205. }
  206. spin_unlock_bh(&queue->lock);
  207. }
  208. static int eql_close(struct net_device *dev)
  209. {
  210. equalizer_t *eql = netdev_priv(dev);
  211. /*
  212. * The timer has to be stopped first before we start hacking away
  213. * at the data structure it scans every so often...
  214. */
  215. del_timer_sync(&eql->timer);
  216. eql_kill_slave_queue(&eql->queue);
  217. return 0;
  218. }
  219. static int eql_enslave(struct net_device *dev, slaving_request_t __user *srq);
  220. static int eql_emancipate(struct net_device *dev, slaving_request_t __user *srq);
  221. static int eql_g_slave_cfg(struct net_device *dev, slave_config_t __user *sc);
  222. static int eql_s_slave_cfg(struct net_device *dev, slave_config_t __user *sc);
  223. static int eql_g_master_cfg(struct net_device *dev, master_config_t __user *mc);
  224. static int eql_s_master_cfg(struct net_device *dev, master_config_t __user *mc);
  225. static int eql_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd)
  226. {
  227. if (cmd != EQL_GETMASTRCFG && cmd != EQL_GETSLAVECFG &&
  228. !capable(CAP_NET_ADMIN))
  229. return -EPERM;
  230. switch (cmd) {
  231. case EQL_ENSLAVE:
  232. return eql_enslave(dev, ifr->ifr_data);
  233. case EQL_EMANCIPATE:
  234. return eql_emancipate(dev, ifr->ifr_data);
  235. case EQL_GETSLAVECFG:
  236. return eql_g_slave_cfg(dev, ifr->ifr_data);
  237. case EQL_SETSLAVECFG:
  238. return eql_s_slave_cfg(dev, ifr->ifr_data);
  239. case EQL_GETMASTRCFG:
  240. return eql_g_master_cfg(dev, ifr->ifr_data);
  241. case EQL_SETMASTRCFG:
  242. return eql_s_master_cfg(dev, ifr->ifr_data);
  243. default:
  244. return -EOPNOTSUPP;
  245. };
  246. }
  247. /* queue->lock must be held */
  248. static slave_t *__eql_schedule_slaves(slave_queue_t *queue)
  249. {
  250. unsigned long best_load = ~0UL;
  251. struct list_head *this, *tmp, *head;
  252. slave_t *best_slave;
  253. best_slave = NULL;
  254. /* Make a pass to set the best slave. */
  255. head = &queue->all_slaves;
  256. list_for_each_safe(this, tmp, head) {
  257. slave_t *slave = list_entry(this, slave_t, list);
  258. unsigned long slave_load, bytes_queued, priority_Bps;
  259. /* Go through the slave list once, updating best_slave
  260. * whenever a new best_load is found.
  261. */
  262. bytes_queued = slave->bytes_queued;
  263. priority_Bps = slave->priority_Bps;
  264. if ((slave->dev->flags & IFF_UP) == IFF_UP) {
  265. slave_load = (~0UL - (~0UL / 2)) -
  266. (priority_Bps) + bytes_queued * 8;
  267. if (slave_load < best_load) {
  268. best_load = slave_load;
  269. best_slave = slave;
  270. }
  271. } else {
  272. /* We found a dead slave, kill it. */
  273. eql_kill_one_slave(queue, slave);
  274. }
  275. }
  276. return best_slave;
  277. }
  278. static int eql_slave_xmit(struct sk_buff *skb, struct net_device *dev)
  279. {
  280. equalizer_t *eql = netdev_priv(dev);
  281. slave_t *slave;
  282. spin_lock(&eql->queue.lock);
  283. slave = __eql_schedule_slaves(&eql->queue);
  284. if (slave) {
  285. struct net_device *slave_dev = slave->dev;
  286. skb->dev = slave_dev;
  287. skb->priority = 1;
  288. slave->bytes_queued += skb->len;
  289. dev_queue_xmit(skb);
  290. dev->stats.tx_packets++;
  291. } else {
  292. dev->stats.tx_dropped++;
  293. dev_kfree_skb(skb);
  294. }
  295. spin_unlock(&eql->queue.lock);
  296. return 0;
  297. }
  298. /*
  299. * Private ioctl functions
  300. */
  301. /* queue->lock must be held */
  302. static slave_t *__eql_find_slave_dev(slave_queue_t *queue, struct net_device *dev)
  303. {
  304. struct list_head *this, *head;
  305. head = &queue->all_slaves;
  306. list_for_each(this, head) {
  307. slave_t *slave = list_entry(this, slave_t, list);
  308. if (slave->dev == dev)
  309. return slave;
  310. }
  311. return NULL;
  312. }
  313. static inline int eql_is_full(slave_queue_t *queue)
  314. {
  315. equalizer_t *eql = netdev_priv(queue->master_dev);
  316. if (queue->num_slaves >= eql->max_slaves)
  317. return 1;
  318. return 0;
  319. }
  320. /* queue->lock must be held */
  321. static int __eql_insert_slave(slave_queue_t *queue, slave_t *slave)
  322. {
  323. if (!eql_is_full(queue)) {
  324. slave_t *duplicate_slave = NULL;
  325. duplicate_slave = __eql_find_slave_dev(queue, slave->dev);
  326. if (duplicate_slave)
  327. eql_kill_one_slave(queue, duplicate_slave);
  328. list_add(&slave->list, &queue->all_slaves);
  329. queue->num_slaves++;
  330. slave->dev->flags |= IFF_SLAVE;
  331. return 0;
  332. }
  333. return -ENOSPC;
  334. }
  335. static int eql_enslave(struct net_device *master_dev, slaving_request_t __user *srqp)
  336. {
  337. struct net_device *slave_dev;
  338. slaving_request_t srq;
  339. if (copy_from_user(&srq, srqp, sizeof (slaving_request_t)))
  340. return -EFAULT;
  341. slave_dev = dev_get_by_name(&init_net, srq.slave_name);
  342. if (slave_dev) {
  343. if ((master_dev->flags & IFF_UP) == IFF_UP) {
  344. /* slave is not a master & not already a slave: */
  345. if (!eql_is_master(slave_dev) &&
  346. !eql_is_slave(slave_dev)) {
  347. slave_t *s = kmalloc(sizeof(*s), GFP_KERNEL);
  348. equalizer_t *eql = netdev_priv(master_dev);
  349. int ret;
  350. if (!s) {
  351. dev_put(slave_dev);
  352. return -ENOMEM;
  353. }
  354. memset(s, 0, sizeof(*s));
  355. s->dev = slave_dev;
  356. s->priority = srq.priority;
  357. s->priority_bps = srq.priority;
  358. s->priority_Bps = srq.priority / 8;
  359. spin_lock_bh(&eql->queue.lock);
  360. ret = __eql_insert_slave(&eql->queue, s);
  361. if (ret) {
  362. dev_put(slave_dev);
  363. kfree(s);
  364. }
  365. spin_unlock_bh(&eql->queue.lock);
  366. return ret;
  367. }
  368. }
  369. dev_put(slave_dev);
  370. }
  371. return -EINVAL;
  372. }
  373. static int eql_emancipate(struct net_device *master_dev, slaving_request_t __user *srqp)
  374. {
  375. equalizer_t *eql = netdev_priv(master_dev);
  376. struct net_device *slave_dev;
  377. slaving_request_t srq;
  378. int ret;
  379. if (copy_from_user(&srq, srqp, sizeof (slaving_request_t)))
  380. return -EFAULT;
  381. slave_dev = dev_get_by_name(&init_net, srq.slave_name);
  382. ret = -EINVAL;
  383. if (slave_dev) {
  384. spin_lock_bh(&eql->queue.lock);
  385. if (eql_is_slave(slave_dev)) {
  386. slave_t *slave = __eql_find_slave_dev(&eql->queue,
  387. slave_dev);
  388. if (slave) {
  389. eql_kill_one_slave(&eql->queue, slave);
  390. ret = 0;
  391. }
  392. }
  393. dev_put(slave_dev);
  394. spin_unlock_bh(&eql->queue.lock);
  395. }
  396. return ret;
  397. }
  398. static int eql_g_slave_cfg(struct net_device *dev, slave_config_t __user *scp)
  399. {
  400. equalizer_t *eql = netdev_priv(dev);
  401. slave_t *slave;
  402. struct net_device *slave_dev;
  403. slave_config_t sc;
  404. int ret;
  405. if (copy_from_user(&sc, scp, sizeof (slave_config_t)))
  406. return -EFAULT;
  407. slave_dev = dev_get_by_name(&init_net, sc.slave_name);
  408. if (!slave_dev)
  409. return -ENODEV;
  410. ret = -EINVAL;
  411. spin_lock_bh(&eql->queue.lock);
  412. if (eql_is_slave(slave_dev)) {
  413. slave = __eql_find_slave_dev(&eql->queue, slave_dev);
  414. if (slave) {
  415. sc.priority = slave->priority;
  416. ret = 0;
  417. }
  418. }
  419. spin_unlock_bh(&eql->queue.lock);
  420. dev_put(slave_dev);
  421. if (!ret && copy_to_user(scp, &sc, sizeof (slave_config_t)))
  422. ret = -EFAULT;
  423. return ret;
  424. }
  425. static int eql_s_slave_cfg(struct net_device *dev, slave_config_t __user *scp)
  426. {
  427. slave_t *slave;
  428. equalizer_t *eql;
  429. struct net_device *slave_dev;
  430. slave_config_t sc;
  431. int ret;
  432. if (copy_from_user(&sc, scp, sizeof (slave_config_t)))
  433. return -EFAULT;
  434. slave_dev = dev_get_by_name(&init_net, sc.slave_name);
  435. if (!slave_dev)
  436. return -ENODEV;
  437. ret = -EINVAL;
  438. eql = netdev_priv(dev);
  439. spin_lock_bh(&eql->queue.lock);
  440. if (eql_is_slave(slave_dev)) {
  441. slave = __eql_find_slave_dev(&eql->queue, slave_dev);
  442. if (slave) {
  443. slave->priority = sc.priority;
  444. slave->priority_bps = sc.priority;
  445. slave->priority_Bps = sc.priority / 8;
  446. ret = 0;
  447. }
  448. }
  449. spin_unlock_bh(&eql->queue.lock);
  450. dev_put(slave_dev);
  451. return ret;
  452. }
  453. static int eql_g_master_cfg(struct net_device *dev, master_config_t __user *mcp)
  454. {
  455. equalizer_t *eql;
  456. master_config_t mc;
  457. if (eql_is_master(dev)) {
  458. eql = netdev_priv(dev);
  459. mc.max_slaves = eql->max_slaves;
  460. mc.min_slaves = eql->min_slaves;
  461. if (copy_to_user(mcp, &mc, sizeof (master_config_t)))
  462. return -EFAULT;
  463. return 0;
  464. }
  465. return -EINVAL;
  466. }
  467. static int eql_s_master_cfg(struct net_device *dev, master_config_t __user *mcp)
  468. {
  469. equalizer_t *eql;
  470. master_config_t mc;
  471. if (copy_from_user(&mc, mcp, sizeof (master_config_t)))
  472. return -EFAULT;
  473. if (eql_is_master(dev)) {
  474. eql = netdev_priv(dev);
  475. eql->max_slaves = mc.max_slaves;
  476. eql->min_slaves = mc.min_slaves;
  477. return 0;
  478. }
  479. return -EINVAL;
  480. }
  481. static struct net_device *dev_eql;
  482. static int __init eql_init_module(void)
  483. {
  484. int err;
  485. printk(version);
  486. dev_eql = alloc_netdev(sizeof(equalizer_t), "eql", eql_setup);
  487. if (!dev_eql)
  488. return -ENOMEM;
  489. err = register_netdev(dev_eql);
  490. if (err)
  491. free_netdev(dev_eql);
  492. return err;
  493. }
  494. static void __exit eql_cleanup_module(void)
  495. {
  496. unregister_netdev(dev_eql);
  497. free_netdev(dev_eql);
  498. }
  499. module_init(eql_init_module);
  500. module_exit(eql_cleanup_module);
  501. MODULE_LICENSE("GPL");