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 ([email protected])
  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 [email protected], 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. #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  112. #include <linux/compat.h>
  113. #include <linux/capability.h>
  114. #include <linux/module.h>
  115. #include <linux/kernel.h>
  116. #include <linux/init.h>
  117. #include <linux/slab.h>
  118. #include <linux/timer.h>
  119. #include <linux/netdevice.h>
  120. #include <net/net_namespace.h>
  121. #include <linux/if.h>
  122. #include <linux/if_arp.h>
  123. #include <linux/if_eql.h>
  124. #include <linux/pkt_sched.h>
  125. #include <linux/uaccess.h>
  126. static int eql_open(struct net_device *dev);
  127. static int eql_close(struct net_device *dev);
  128. static int eql_siocdevprivate(struct net_device *dev, struct ifreq *ifr,
  129. void __user *data, int cmd);
  130. static netdev_tx_t eql_slave_xmit(struct sk_buff *skb, struct net_device *dev);
  131. #define eql_is_slave(dev) ((dev->flags & IFF_SLAVE) == IFF_SLAVE)
  132. #define eql_is_master(dev) ((dev->flags & IFF_MASTER) == IFF_MASTER)
  133. static void eql_kill_one_slave(slave_queue_t *queue, slave_t *slave);
  134. static void eql_timer(struct timer_list *t)
  135. {
  136. equalizer_t *eql = from_timer(eql, t, timer);
  137. struct list_head *this, *tmp, *head;
  138. spin_lock(&eql->queue.lock);
  139. head = &eql->queue.all_slaves;
  140. list_for_each_safe(this, tmp, head) {
  141. slave_t *slave = list_entry(this, slave_t, list);
  142. if ((slave->dev->flags & IFF_UP) == IFF_UP) {
  143. slave->bytes_queued -= slave->priority_Bps;
  144. if (slave->bytes_queued < 0)
  145. slave->bytes_queued = 0;
  146. } else {
  147. eql_kill_one_slave(&eql->queue, slave);
  148. }
  149. }
  150. spin_unlock(&eql->queue.lock);
  151. eql->timer.expires = jiffies + EQL_DEFAULT_RESCHED_IVAL;
  152. add_timer(&eql->timer);
  153. }
  154. static const char version[] __initconst =
  155. "Equalizer2002: Simon Janes ([email protected]) and David S. Miller ([email protected])";
  156. static const struct net_device_ops eql_netdev_ops = {
  157. .ndo_open = eql_open,
  158. .ndo_stop = eql_close,
  159. .ndo_siocdevprivate = eql_siocdevprivate,
  160. .ndo_start_xmit = eql_slave_xmit,
  161. };
  162. static void __init eql_setup(struct net_device *dev)
  163. {
  164. equalizer_t *eql = netdev_priv(dev);
  165. timer_setup(&eql->timer, eql_timer, 0);
  166. eql->timer.expires = jiffies + EQL_DEFAULT_RESCHED_IVAL;
  167. spin_lock_init(&eql->queue.lock);
  168. INIT_LIST_HEAD(&eql->queue.all_slaves);
  169. eql->queue.master_dev = dev;
  170. dev->netdev_ops = &eql_netdev_ops;
  171. /*
  172. * Now we undo some of the things that eth_setup does
  173. * that we don't like
  174. */
  175. dev->mtu = EQL_DEFAULT_MTU; /* set to 576 in if_eql.h */
  176. dev->flags = IFF_MASTER;
  177. dev->type = ARPHRD_SLIP;
  178. dev->tx_queue_len = 5; /* Hands them off fast */
  179. netif_keep_dst(dev);
  180. }
  181. static int eql_open(struct net_device *dev)
  182. {
  183. equalizer_t *eql = netdev_priv(dev);
  184. /* XXX We should force this off automatically for the user. */
  185. netdev_info(dev,
  186. "remember to turn off Van-Jacobson compression on your slave devices\n");
  187. BUG_ON(!list_empty(&eql->queue.all_slaves));
  188. eql->min_slaves = 1;
  189. eql->max_slaves = EQL_DEFAULT_MAX_SLAVES; /* 4 usually... */
  190. add_timer(&eql->timer);
  191. return 0;
  192. }
  193. static void eql_kill_one_slave(slave_queue_t *queue, slave_t *slave)
  194. {
  195. list_del(&slave->list);
  196. queue->num_slaves--;
  197. slave->dev->flags &= ~IFF_SLAVE;
  198. netdev_put(slave->dev, &slave->dev_tracker);
  199. kfree(slave);
  200. }
  201. static void eql_kill_slave_queue(slave_queue_t *queue)
  202. {
  203. struct list_head *head, *tmp, *this;
  204. spin_lock_bh(&queue->lock);
  205. head = &queue->all_slaves;
  206. list_for_each_safe(this, tmp, head) {
  207. slave_t *s = list_entry(this, slave_t, list);
  208. eql_kill_one_slave(queue, s);
  209. }
  210. spin_unlock_bh(&queue->lock);
  211. }
  212. static int eql_close(struct net_device *dev)
  213. {
  214. equalizer_t *eql = netdev_priv(dev);
  215. /*
  216. * The timer has to be stopped first before we start hacking away
  217. * at the data structure it scans every so often...
  218. */
  219. del_timer_sync(&eql->timer);
  220. eql_kill_slave_queue(&eql->queue);
  221. return 0;
  222. }
  223. static int eql_enslave(struct net_device *dev, slaving_request_t __user *srq);
  224. static int eql_emancipate(struct net_device *dev, slaving_request_t __user *srq);
  225. static int eql_g_slave_cfg(struct net_device *dev, slave_config_t __user *sc);
  226. static int eql_s_slave_cfg(struct net_device *dev, slave_config_t __user *sc);
  227. static int eql_g_master_cfg(struct net_device *dev, master_config_t __user *mc);
  228. static int eql_s_master_cfg(struct net_device *dev, master_config_t __user *mc);
  229. static int eql_siocdevprivate(struct net_device *dev, struct ifreq *ifr,
  230. void __user *data, int cmd)
  231. {
  232. if (cmd != EQL_GETMASTRCFG && cmd != EQL_GETSLAVECFG &&
  233. !capable(CAP_NET_ADMIN))
  234. return -EPERM;
  235. if (in_compat_syscall()) /* to be implemented */
  236. return -EOPNOTSUPP;
  237. switch (cmd) {
  238. case EQL_ENSLAVE:
  239. return eql_enslave(dev, data);
  240. case EQL_EMANCIPATE:
  241. return eql_emancipate(dev, data);
  242. case EQL_GETSLAVECFG:
  243. return eql_g_slave_cfg(dev, data);
  244. case EQL_SETSLAVECFG:
  245. return eql_s_slave_cfg(dev, data);
  246. case EQL_GETMASTRCFG:
  247. return eql_g_master_cfg(dev, data);
  248. case EQL_SETMASTRCFG:
  249. return eql_s_master_cfg(dev, data);
  250. default:
  251. return -EOPNOTSUPP;
  252. }
  253. }
  254. /* queue->lock must be held */
  255. static slave_t *__eql_schedule_slaves(slave_queue_t *queue)
  256. {
  257. unsigned long best_load = ~0UL;
  258. struct list_head *this, *tmp, *head;
  259. slave_t *best_slave;
  260. best_slave = NULL;
  261. /* Make a pass to set the best slave. */
  262. head = &queue->all_slaves;
  263. list_for_each_safe(this, tmp, head) {
  264. slave_t *slave = list_entry(this, slave_t, list);
  265. unsigned long slave_load, bytes_queued, priority_Bps;
  266. /* Go through the slave list once, updating best_slave
  267. * whenever a new best_load is found.
  268. */
  269. bytes_queued = slave->bytes_queued;
  270. priority_Bps = slave->priority_Bps;
  271. if ((slave->dev->flags & IFF_UP) == IFF_UP) {
  272. slave_load = (~0UL - (~0UL / 2)) -
  273. (priority_Bps) + bytes_queued * 8;
  274. if (slave_load < best_load) {
  275. best_load = slave_load;
  276. best_slave = slave;
  277. }
  278. } else {
  279. /* We found a dead slave, kill it. */
  280. eql_kill_one_slave(queue, slave);
  281. }
  282. }
  283. return best_slave;
  284. }
  285. static netdev_tx_t eql_slave_xmit(struct sk_buff *skb, struct net_device *dev)
  286. {
  287. equalizer_t *eql = netdev_priv(dev);
  288. slave_t *slave;
  289. spin_lock(&eql->queue.lock);
  290. slave = __eql_schedule_slaves(&eql->queue);
  291. if (slave) {
  292. struct net_device *slave_dev = slave->dev;
  293. skb->dev = slave_dev;
  294. skb->priority = TC_PRIO_FILLER;
  295. slave->bytes_queued += skb->len;
  296. dev_queue_xmit(skb);
  297. dev->stats.tx_packets++;
  298. } else {
  299. dev->stats.tx_dropped++;
  300. dev_kfree_skb(skb);
  301. }
  302. spin_unlock(&eql->queue.lock);
  303. return NETDEV_TX_OK;
  304. }
  305. /*
  306. * Private ioctl functions
  307. */
  308. /* queue->lock must be held */
  309. static slave_t *__eql_find_slave_dev(slave_queue_t *queue, struct net_device *dev)
  310. {
  311. struct list_head *this, *head;
  312. head = &queue->all_slaves;
  313. list_for_each(this, head) {
  314. slave_t *slave = list_entry(this, slave_t, list);
  315. if (slave->dev == dev)
  316. return slave;
  317. }
  318. return NULL;
  319. }
  320. static inline int eql_is_full(slave_queue_t *queue)
  321. {
  322. equalizer_t *eql = netdev_priv(queue->master_dev);
  323. if (queue->num_slaves >= eql->max_slaves)
  324. return 1;
  325. return 0;
  326. }
  327. /* queue->lock must be held */
  328. static int __eql_insert_slave(slave_queue_t *queue, slave_t *slave)
  329. {
  330. if (!eql_is_full(queue)) {
  331. slave_t *duplicate_slave = NULL;
  332. duplicate_slave = __eql_find_slave_dev(queue, slave->dev);
  333. if (duplicate_slave)
  334. eql_kill_one_slave(queue, duplicate_slave);
  335. netdev_hold(slave->dev, &slave->dev_tracker, GFP_ATOMIC);
  336. list_add(&slave->list, &queue->all_slaves);
  337. queue->num_slaves++;
  338. slave->dev->flags |= IFF_SLAVE;
  339. return 0;
  340. }
  341. return -ENOSPC;
  342. }
  343. static int eql_enslave(struct net_device *master_dev, slaving_request_t __user *srqp)
  344. {
  345. struct net_device *slave_dev;
  346. slaving_request_t srq;
  347. if (copy_from_user(&srq, srqp, sizeof (slaving_request_t)))
  348. return -EFAULT;
  349. slave_dev = __dev_get_by_name(&init_net, srq.slave_name);
  350. if (!slave_dev)
  351. return -ENODEV;
  352. if ((master_dev->flags & IFF_UP) == IFF_UP) {
  353. /* slave is not a master & not already a slave: */
  354. if (!eql_is_master(slave_dev) && !eql_is_slave(slave_dev)) {
  355. slave_t *s = kzalloc(sizeof(*s), GFP_KERNEL);
  356. equalizer_t *eql = netdev_priv(master_dev);
  357. int ret;
  358. if (!s)
  359. return -ENOMEM;
  360. s->dev = slave_dev;
  361. s->priority = srq.priority;
  362. s->priority_bps = srq.priority;
  363. s->priority_Bps = srq.priority / 8;
  364. spin_lock_bh(&eql->queue.lock);
  365. ret = __eql_insert_slave(&eql->queue, s);
  366. if (ret)
  367. kfree(s);
  368. spin_unlock_bh(&eql->queue.lock);
  369. return ret;
  370. }
  371. }
  372. return -EINVAL;
  373. }
  374. static int eql_emancipate(struct net_device *master_dev, slaving_request_t __user *srqp)
  375. {
  376. equalizer_t *eql = netdev_priv(master_dev);
  377. struct net_device *slave_dev;
  378. slaving_request_t srq;
  379. int ret;
  380. if (copy_from_user(&srq, srqp, sizeof (slaving_request_t)))
  381. return -EFAULT;
  382. slave_dev = __dev_get_by_name(&init_net, srq.slave_name);
  383. if (!slave_dev)
  384. return -ENODEV;
  385. ret = -EINVAL;
  386. spin_lock_bh(&eql->queue.lock);
  387. if (eql_is_slave(slave_dev)) {
  388. slave_t *slave = __eql_find_slave_dev(&eql->queue, slave_dev);
  389. if (slave) {
  390. eql_kill_one_slave(&eql->queue, slave);
  391. ret = 0;
  392. }
  393. }
  394. spin_unlock_bh(&eql->queue.lock);
  395. return ret;
  396. }
  397. static int eql_g_slave_cfg(struct net_device *dev, slave_config_t __user *scp)
  398. {
  399. equalizer_t *eql = netdev_priv(dev);
  400. slave_t *slave;
  401. struct net_device *slave_dev;
  402. slave_config_t sc;
  403. int ret;
  404. if (copy_from_user(&sc, scp, sizeof (slave_config_t)))
  405. return -EFAULT;
  406. slave_dev = __dev_get_by_name(&init_net, sc.slave_name);
  407. if (!slave_dev)
  408. return -ENODEV;
  409. ret = -EINVAL;
  410. spin_lock_bh(&eql->queue.lock);
  411. if (eql_is_slave(slave_dev)) {
  412. slave = __eql_find_slave_dev(&eql->queue, slave_dev);
  413. if (slave) {
  414. sc.priority = slave->priority;
  415. ret = 0;
  416. }
  417. }
  418. spin_unlock_bh(&eql->queue.lock);
  419. if (!ret && copy_to_user(scp, &sc, sizeof (slave_config_t)))
  420. ret = -EFAULT;
  421. return ret;
  422. }
  423. static int eql_s_slave_cfg(struct net_device *dev, slave_config_t __user *scp)
  424. {
  425. slave_t *slave;
  426. equalizer_t *eql;
  427. struct net_device *slave_dev;
  428. slave_config_t sc;
  429. int ret;
  430. if (copy_from_user(&sc, scp, sizeof (slave_config_t)))
  431. return -EFAULT;
  432. slave_dev = __dev_get_by_name(&init_net, sc.slave_name);
  433. if (!slave_dev)
  434. return -ENODEV;
  435. ret = -EINVAL;
  436. eql = netdev_priv(dev);
  437. spin_lock_bh(&eql->queue.lock);
  438. if (eql_is_slave(slave_dev)) {
  439. slave = __eql_find_slave_dev(&eql->queue, slave_dev);
  440. if (slave) {
  441. slave->priority = sc.priority;
  442. slave->priority_bps = sc.priority;
  443. slave->priority_Bps = sc.priority / 8;
  444. ret = 0;
  445. }
  446. }
  447. spin_unlock_bh(&eql->queue.lock);
  448. return ret;
  449. }
  450. static int eql_g_master_cfg(struct net_device *dev, master_config_t __user *mcp)
  451. {
  452. equalizer_t *eql;
  453. master_config_t mc;
  454. memset(&mc, 0, sizeof(master_config_t));
  455. if (eql_is_master(dev)) {
  456. eql = netdev_priv(dev);
  457. mc.max_slaves = eql->max_slaves;
  458. mc.min_slaves = eql->min_slaves;
  459. if (copy_to_user(mcp, &mc, sizeof (master_config_t)))
  460. return -EFAULT;
  461. return 0;
  462. }
  463. return -EINVAL;
  464. }
  465. static int eql_s_master_cfg(struct net_device *dev, master_config_t __user *mcp)
  466. {
  467. equalizer_t *eql;
  468. master_config_t mc;
  469. if (copy_from_user(&mc, mcp, sizeof (master_config_t)))
  470. return -EFAULT;
  471. if (eql_is_master(dev)) {
  472. eql = netdev_priv(dev);
  473. eql->max_slaves = mc.max_slaves;
  474. eql->min_slaves = mc.min_slaves;
  475. return 0;
  476. }
  477. return -EINVAL;
  478. }
  479. static struct net_device *dev_eql;
  480. static int __init eql_init_module(void)
  481. {
  482. int err;
  483. pr_info("%s\n", version);
  484. dev_eql = alloc_netdev(sizeof(equalizer_t), "eql", NET_NAME_UNKNOWN,
  485. eql_setup);
  486. if (!dev_eql)
  487. return -ENOMEM;
  488. err = register_netdev(dev_eql);
  489. if (err)
  490. free_netdev(dev_eql);
  491. return err;
  492. }
  493. static void __exit eql_cleanup_module(void)
  494. {
  495. unregister_netdev(dev_eql);
  496. free_netdev(dev_eql);
  497. }
  498. module_init(eql_init_module);
  499. module_exit(eql_cleanup_module);
  500. MODULE_LICENSE("GPL");