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 #include <linux/module.h> 115 #include <linux/kernel.h> 116 #include <linux/init.h> 117 #include <linux/timer.h> 118 #include <linux/netdevice.h> 119 #include <net/net_namespace.h> 120 121 #include <linux/if.h> 122 #include <linux/if_arp.h> 123 #include <linux/if_eql.h> 124 125 #include <asm/uaccess.h> 126 127 static int eql_open(struct net_device *dev); 128 static int eql_close(struct net_device *dev); 129 static int eql_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd); 130 static netdev_tx_t eql_slave_xmit(struct sk_buff *skb, struct net_device *dev); 131 132 #define eql_is_slave(dev) ((dev->flags & IFF_SLAVE) == IFF_SLAVE) 133 #define eql_is_master(dev) ((dev->flags & IFF_MASTER) == IFF_MASTER) 134 135 static void eql_kill_one_slave(slave_queue_t *queue, slave_t *slave); 136 137 static void eql_timer(unsigned long param) 138 { 139 equalizer_t *eql = (equalizer_t *) param; 140 struct list_head *this, *tmp, *head; 141 142 spin_lock_bh(&eql->queue.lock); 143 head = &eql->queue.all_slaves; 144 list_for_each_safe(this, tmp, head) { 145 slave_t *slave = list_entry(this, slave_t, list); 146 147 if ((slave->dev->flags & IFF_UP) == IFF_UP) { 148 slave->bytes_queued -= slave->priority_Bps; 149 if (slave->bytes_queued < 0) 150 slave->bytes_queued = 0; 151 } else { 152 eql_kill_one_slave(&eql->queue, slave); 153 } 154 155 } 156 spin_unlock_bh(&eql->queue.lock); 157 158 eql->timer.expires = jiffies + EQL_DEFAULT_RESCHED_IVAL; 159 add_timer(&eql->timer); 160 } 161 162 static const char version[] __initconst = 163 "Equalizer2002: Simon Janes (simon@ncm.com) and David S. Miller (davem@redhat.com)\n"; 164 165 static const struct net_device_ops eql_netdev_ops = { 166 .ndo_open = eql_open, 167 .ndo_stop = eql_close, 168 .ndo_do_ioctl = eql_ioctl, 169 .ndo_start_xmit = eql_slave_xmit, 170 }; 171 172 static void __init eql_setup(struct net_device *dev) 173 { 174 equalizer_t *eql = netdev_priv(dev); 175 176 init_timer(&eql->timer); 177 eql->timer.data = (unsigned long) eql; 178 eql->timer.expires = jiffies + EQL_DEFAULT_RESCHED_IVAL; 179 eql->timer.function = eql_timer; 180 181 spin_lock_init(&eql->queue.lock); 182 INIT_LIST_HEAD(&eql->queue.all_slaves); 183 eql->queue.master_dev = dev; 184 185 dev->netdev_ops = &eql_netdev_ops; 186 187 /* 188 * Now we undo some of the things that eth_setup does 189 * that we don't like 190 */ 191 192 dev->mtu = EQL_DEFAULT_MTU; /* set to 576 in if_eql.h */ 193 dev->flags = IFF_MASTER; 194 195 dev->type = ARPHRD_SLIP; 196 dev->tx_queue_len = 5; /* Hands them off fast */ 197 dev->priv_flags &= ~IFF_XMIT_DST_RELEASE; 198 } 199 200 static int eql_open(struct net_device *dev) 201 { 202 equalizer_t *eql = netdev_priv(dev); 203 204 /* XXX We should force this off automatically for the user. */ 205 printk(KERN_INFO "%s: remember to turn off Van-Jacobson compression on " 206 "your slave devices.\n", dev->name); 207 208 BUG_ON(!list_empty(&eql->queue.all_slaves)); 209 210 eql->min_slaves = 1; 211 eql->max_slaves = EQL_DEFAULT_MAX_SLAVES; /* 4 usually... */ 212 213 add_timer(&eql->timer); 214 215 return 0; 216 } 217 218 static void eql_kill_one_slave(slave_queue_t *queue, slave_t *slave) 219 { 220 list_del(&slave->list); 221 queue->num_slaves--; 222 slave->dev->flags &= ~IFF_SLAVE; 223 dev_put(slave->dev); 224 kfree(slave); 225 } 226 227 static void eql_kill_slave_queue(slave_queue_t *queue) 228 { 229 struct list_head *head, *tmp, *this; 230 231 spin_lock_bh(&queue->lock); 232 233 head = &queue->all_slaves; 234 list_for_each_safe(this, tmp, head) { 235 slave_t *s = list_entry(this, slave_t, list); 236 237 eql_kill_one_slave(queue, s); 238 } 239 240 spin_unlock_bh(&queue->lock); 241 } 242 243 static int eql_close(struct net_device *dev) 244 { 245 equalizer_t *eql = netdev_priv(dev); 246 247 /* 248 * The timer has to be stopped first before we start hacking away 249 * at the data structure it scans every so often... 250 */ 251 252 del_timer_sync(&eql->timer); 253 254 eql_kill_slave_queue(&eql->queue); 255 256 return 0; 257 } 258 259 static int eql_enslave(struct net_device *dev, slaving_request_t __user *srq); 260 static int eql_emancipate(struct net_device *dev, slaving_request_t __user *srq); 261 262 static int eql_g_slave_cfg(struct net_device *dev, slave_config_t __user *sc); 263 static int eql_s_slave_cfg(struct net_device *dev, slave_config_t __user *sc); 264 265 static int eql_g_master_cfg(struct net_device *dev, master_config_t __user *mc); 266 static int eql_s_master_cfg(struct net_device *dev, master_config_t __user *mc); 267 268 static int eql_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd) 269 { 270 if (cmd != EQL_GETMASTRCFG && cmd != EQL_GETSLAVECFG && 271 !capable(CAP_NET_ADMIN)) 272 return -EPERM; 273 274 switch (cmd) { 275 case EQL_ENSLAVE: 276 return eql_enslave(dev, ifr->ifr_data); 277 case EQL_EMANCIPATE: 278 return eql_emancipate(dev, ifr->ifr_data); 279 case EQL_GETSLAVECFG: 280 return eql_g_slave_cfg(dev, ifr->ifr_data); 281 case EQL_SETSLAVECFG: 282 return eql_s_slave_cfg(dev, ifr->ifr_data); 283 case EQL_GETMASTRCFG: 284 return eql_g_master_cfg(dev, ifr->ifr_data); 285 case EQL_SETMASTRCFG: 286 return eql_s_master_cfg(dev, ifr->ifr_data); 287 default: 288 return -EOPNOTSUPP; 289 }; 290 } 291 292 /* queue->lock must be held */ 293 static slave_t *__eql_schedule_slaves(slave_queue_t *queue) 294 { 295 unsigned long best_load = ~0UL; 296 struct list_head *this, *tmp, *head; 297 slave_t *best_slave; 298 299 best_slave = NULL; 300 301 /* Make a pass to set the best slave. */ 302 head = &queue->all_slaves; 303 list_for_each_safe(this, tmp, head) { 304 slave_t *slave = list_entry(this, slave_t, list); 305 unsigned long slave_load, bytes_queued, priority_Bps; 306 307 /* Go through the slave list once, updating best_slave 308 * whenever a new best_load is found. 309 */ 310 bytes_queued = slave->bytes_queued; 311 priority_Bps = slave->priority_Bps; 312 if ((slave->dev->flags & IFF_UP) == IFF_UP) { 313 slave_load = (~0UL - (~0UL / 2)) - 314 (priority_Bps) + bytes_queued * 8; 315 316 if (slave_load < best_load) { 317 best_load = slave_load; 318 best_slave = slave; 319 } 320 } else { 321 /* We found a dead slave, kill it. */ 322 eql_kill_one_slave(queue, slave); 323 } 324 } 325 return best_slave; 326 } 327 328 static netdev_tx_t eql_slave_xmit(struct sk_buff *skb, struct net_device *dev) 329 { 330 equalizer_t *eql = netdev_priv(dev); 331 slave_t *slave; 332 333 spin_lock(&eql->queue.lock); 334 335 slave = __eql_schedule_slaves(&eql->queue); 336 if (slave) { 337 struct net_device *slave_dev = slave->dev; 338 339 skb->dev = slave_dev; 340 skb->priority = 1; 341 slave->bytes_queued += skb->len; 342 dev_queue_xmit(skb); 343 dev->stats.tx_packets++; 344 } else { 345 dev->stats.tx_dropped++; 346 dev_kfree_skb(skb); 347 } 348 349 spin_unlock(&eql->queue.lock); 350 351 return NETDEV_TX_OK; 352 } 353 354 /* 355 * Private ioctl functions 356 */ 357 358 /* queue->lock must be held */ 359 static slave_t *__eql_find_slave_dev(slave_queue_t *queue, struct net_device *dev) 360 { 361 struct list_head *this, *head; 362 363 head = &queue->all_slaves; 364 list_for_each(this, head) { 365 slave_t *slave = list_entry(this, slave_t, list); 366 367 if (slave->dev == dev) 368 return slave; 369 } 370 371 return NULL; 372 } 373 374 static inline int eql_is_full(slave_queue_t *queue) 375 { 376 equalizer_t *eql = netdev_priv(queue->master_dev); 377 378 if (queue->num_slaves >= eql->max_slaves) 379 return 1; 380 return 0; 381 } 382 383 /* queue->lock must be held */ 384 static int __eql_insert_slave(slave_queue_t *queue, slave_t *slave) 385 { 386 if (!eql_is_full(queue)) { 387 slave_t *duplicate_slave = NULL; 388 389 duplicate_slave = __eql_find_slave_dev(queue, slave->dev); 390 if (duplicate_slave) 391 eql_kill_one_slave(queue, duplicate_slave); 392 393 list_add(&slave->list, &queue->all_slaves); 394 queue->num_slaves++; 395 slave->dev->flags |= IFF_SLAVE; 396 397 return 0; 398 } 399 400 return -ENOSPC; 401 } 402 403 static int eql_enslave(struct net_device *master_dev, slaving_request_t __user *srqp) 404 { 405 struct net_device *slave_dev; 406 slaving_request_t srq; 407 408 if (copy_from_user(&srq, srqp, sizeof (slaving_request_t))) 409 return -EFAULT; 410 411 slave_dev = dev_get_by_name(&init_net, srq.slave_name); 412 if (slave_dev) { 413 if ((master_dev->flags & IFF_UP) == IFF_UP) { 414 /* slave is not a master & not already a slave: */ 415 if (!eql_is_master(slave_dev) && 416 !eql_is_slave(slave_dev)) { 417 slave_t *s = kmalloc(sizeof(*s), GFP_KERNEL); 418 equalizer_t *eql = netdev_priv(master_dev); 419 int ret; 420 421 if (!s) { 422 dev_put(slave_dev); 423 return -ENOMEM; 424 } 425 426 memset(s, 0, sizeof(*s)); 427 s->dev = slave_dev; 428 s->priority = srq.priority; 429 s->priority_bps = srq.priority; 430 s->priority_Bps = srq.priority / 8; 431 432 spin_lock_bh(&eql->queue.lock); 433 ret = __eql_insert_slave(&eql->queue, s); 434 if (ret) { 435 dev_put(slave_dev); 436 kfree(s); 437 } 438 spin_unlock_bh(&eql->queue.lock); 439 440 return ret; 441 } 442 } 443 dev_put(slave_dev); 444 } 445 446 return -EINVAL; 447 } 448 449 static 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 ret = -EINVAL; 461 if (slave_dev) { 462 spin_lock_bh(&eql->queue.lock); 463 464 if (eql_is_slave(slave_dev)) { 465 slave_t *slave = __eql_find_slave_dev(&eql->queue, 466 slave_dev); 467 468 if (slave) { 469 eql_kill_one_slave(&eql->queue, slave); 470 ret = 0; 471 } 472 } 473 dev_put(slave_dev); 474 475 spin_unlock_bh(&eql->queue.lock); 476 } 477 478 return ret; 479 } 480 481 static int eql_g_slave_cfg(struct net_device *dev, slave_config_t __user *scp) 482 { 483 equalizer_t *eql = netdev_priv(dev); 484 slave_t *slave; 485 struct net_device *slave_dev; 486 slave_config_t sc; 487 int ret; 488 489 if (copy_from_user(&sc, scp, sizeof (slave_config_t))) 490 return -EFAULT; 491 492 slave_dev = dev_get_by_name(&init_net, sc.slave_name); 493 if (!slave_dev) 494 return -ENODEV; 495 496 ret = -EINVAL; 497 498 spin_lock_bh(&eql->queue.lock); 499 if (eql_is_slave(slave_dev)) { 500 slave = __eql_find_slave_dev(&eql->queue, slave_dev); 501 if (slave) { 502 sc.priority = slave->priority; 503 ret = 0; 504 } 505 } 506 spin_unlock_bh(&eql->queue.lock); 507 508 dev_put(slave_dev); 509 510 if (!ret && copy_to_user(scp, &sc, sizeof (slave_config_t))) 511 ret = -EFAULT; 512 513 return ret; 514 } 515 516 static int eql_s_slave_cfg(struct net_device *dev, slave_config_t __user *scp) 517 { 518 slave_t *slave; 519 equalizer_t *eql; 520 struct net_device *slave_dev; 521 slave_config_t sc; 522 int ret; 523 524 if (copy_from_user(&sc, scp, sizeof (slave_config_t))) 525 return -EFAULT; 526 527 slave_dev = dev_get_by_name(&init_net, sc.slave_name); 528 if (!slave_dev) 529 return -ENODEV; 530 531 ret = -EINVAL; 532 533 eql = netdev_priv(dev); 534 spin_lock_bh(&eql->queue.lock); 535 if (eql_is_slave(slave_dev)) { 536 slave = __eql_find_slave_dev(&eql->queue, slave_dev); 537 if (slave) { 538 slave->priority = sc.priority; 539 slave->priority_bps = sc.priority; 540 slave->priority_Bps = sc.priority / 8; 541 ret = 0; 542 } 543 } 544 spin_unlock_bh(&eql->queue.lock); 545 546 dev_put(slave_dev); 547 548 return ret; 549 } 550 551 static int eql_g_master_cfg(struct net_device *dev, master_config_t __user *mcp) 552 { 553 equalizer_t *eql; 554 master_config_t mc; 555 556 if (eql_is_master(dev)) { 557 eql = netdev_priv(dev); 558 mc.max_slaves = eql->max_slaves; 559 mc.min_slaves = eql->min_slaves; 560 if (copy_to_user(mcp, &mc, sizeof (master_config_t))) 561 return -EFAULT; 562 return 0; 563 } 564 return -EINVAL; 565 } 566 567 static int eql_s_master_cfg(struct net_device *dev, master_config_t __user *mcp) 568 { 569 equalizer_t *eql; 570 master_config_t mc; 571 572 if (copy_from_user(&mc, mcp, sizeof (master_config_t))) 573 return -EFAULT; 574 575 if (eql_is_master(dev)) { 576 eql = netdev_priv(dev); 577 eql->max_slaves = mc.max_slaves; 578 eql->min_slaves = mc.min_slaves; 579 return 0; 580 } 581 return -EINVAL; 582 } 583 584 static struct net_device *dev_eql; 585 586 static int __init eql_init_module(void) 587 { 588 int err; 589 590 printk(version); 591 592 dev_eql = alloc_netdev(sizeof(equalizer_t), "eql", eql_setup); 593 if (!dev_eql) 594 return -ENOMEM; 595 596 err = register_netdev(dev_eql); 597 if (err) 598 free_netdev(dev_eql); 599 return err; 600 } 601 602 static void __exit eql_cleanup_module(void) 603 { 604 unregister_netdev(dev_eql); 605 free_netdev(dev_eql); 606 } 607 608 module_init(eql_init_module); 609 module_exit(eql_cleanup_module); 610 MODULE_LICENSE("GPL"); 611