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 int 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 char version[] __initdata = 163 "Equalizer2002: Simon Janes (simon@ncm.com) and David S. Miller (davem@redhat.com)\n"; 164 165 static void __init eql_setup(struct net_device *dev) 166 { 167 equalizer_t *eql = netdev_priv(dev); 168 169 init_timer(&eql->timer); 170 eql->timer.data = (unsigned long) eql; 171 eql->timer.expires = jiffies + EQL_DEFAULT_RESCHED_IVAL; 172 eql->timer.function = eql_timer; 173 174 spin_lock_init(&eql->queue.lock); 175 INIT_LIST_HEAD(&eql->queue.all_slaves); 176 eql->queue.master_dev = dev; 177 178 dev->open = eql_open; 179 dev->stop = eql_close; 180 dev->do_ioctl = eql_ioctl; 181 dev->hard_start_xmit = eql_slave_xmit; 182 183 /* 184 * Now we undo some of the things that eth_setup does 185 * that we don't like 186 */ 187 188 dev->mtu = EQL_DEFAULT_MTU; /* set to 576 in if_eql.h */ 189 dev->flags = IFF_MASTER; 190 191 dev->type = ARPHRD_SLIP; 192 dev->tx_queue_len = 5; /* Hands them off fast */ 193 } 194 195 static int eql_open(struct net_device *dev) 196 { 197 equalizer_t *eql = netdev_priv(dev); 198 199 /* XXX We should force this off automatically for the user. */ 200 printk(KERN_INFO "%s: remember to turn off Van-Jacobson compression on " 201 "your slave devices.\n", dev->name); 202 203 BUG_ON(!list_empty(&eql->queue.all_slaves)); 204 205 eql->min_slaves = 1; 206 eql->max_slaves = EQL_DEFAULT_MAX_SLAVES; /* 4 usually... */ 207 208 add_timer(&eql->timer); 209 210 return 0; 211 } 212 213 static void eql_kill_one_slave(slave_queue_t *queue, slave_t *slave) 214 { 215 list_del(&slave->list); 216 queue->num_slaves--; 217 slave->dev->flags &= ~IFF_SLAVE; 218 dev_put(slave->dev); 219 kfree(slave); 220 } 221 222 static void eql_kill_slave_queue(slave_queue_t *queue) 223 { 224 struct list_head *head, *tmp, *this; 225 226 spin_lock_bh(&queue->lock); 227 228 head = &queue->all_slaves; 229 list_for_each_safe(this, tmp, head) { 230 slave_t *s = list_entry(this, slave_t, list); 231 232 eql_kill_one_slave(queue, s); 233 } 234 235 spin_unlock_bh(&queue->lock); 236 } 237 238 static int eql_close(struct net_device *dev) 239 { 240 equalizer_t *eql = netdev_priv(dev); 241 242 /* 243 * The timer has to be stopped first before we start hacking away 244 * at the data structure it scans every so often... 245 */ 246 247 del_timer_sync(&eql->timer); 248 249 eql_kill_slave_queue(&eql->queue); 250 251 return 0; 252 } 253 254 static int eql_enslave(struct net_device *dev, slaving_request_t __user *srq); 255 static int eql_emancipate(struct net_device *dev, slaving_request_t __user *srq); 256 257 static int eql_g_slave_cfg(struct net_device *dev, slave_config_t __user *sc); 258 static int eql_s_slave_cfg(struct net_device *dev, slave_config_t __user *sc); 259 260 static int eql_g_master_cfg(struct net_device *dev, master_config_t __user *mc); 261 static int eql_s_master_cfg(struct net_device *dev, master_config_t __user *mc); 262 263 static int eql_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd) 264 { 265 if (cmd != EQL_GETMASTRCFG && cmd != EQL_GETSLAVECFG && 266 !capable(CAP_NET_ADMIN)) 267 return -EPERM; 268 269 switch (cmd) { 270 case EQL_ENSLAVE: 271 return eql_enslave(dev, ifr->ifr_data); 272 case EQL_EMANCIPATE: 273 return eql_emancipate(dev, ifr->ifr_data); 274 case EQL_GETSLAVECFG: 275 return eql_g_slave_cfg(dev, ifr->ifr_data); 276 case EQL_SETSLAVECFG: 277 return eql_s_slave_cfg(dev, ifr->ifr_data); 278 case EQL_GETMASTRCFG: 279 return eql_g_master_cfg(dev, ifr->ifr_data); 280 case EQL_SETMASTRCFG: 281 return eql_s_master_cfg(dev, ifr->ifr_data); 282 default: 283 return -EOPNOTSUPP; 284 }; 285 } 286 287 /* queue->lock must be held */ 288 static slave_t *__eql_schedule_slaves(slave_queue_t *queue) 289 { 290 unsigned long best_load = ~0UL; 291 struct list_head *this, *tmp, *head; 292 slave_t *best_slave; 293 294 best_slave = NULL; 295 296 /* Make a pass to set the best slave. */ 297 head = &queue->all_slaves; 298 list_for_each_safe(this, tmp, head) { 299 slave_t *slave = list_entry(this, slave_t, list); 300 unsigned long slave_load, bytes_queued, priority_Bps; 301 302 /* Go through the slave list once, updating best_slave 303 * whenever a new best_load is found. 304 */ 305 bytes_queued = slave->bytes_queued; 306 priority_Bps = slave->priority_Bps; 307 if ((slave->dev->flags & IFF_UP) == IFF_UP) { 308 slave_load = (~0UL - (~0UL / 2)) - 309 (priority_Bps) + bytes_queued * 8; 310 311 if (slave_load < best_load) { 312 best_load = slave_load; 313 best_slave = slave; 314 } 315 } else { 316 /* We found a dead slave, kill it. */ 317 eql_kill_one_slave(queue, slave); 318 } 319 } 320 return best_slave; 321 } 322 323 static int eql_slave_xmit(struct sk_buff *skb, struct net_device *dev) 324 { 325 equalizer_t *eql = netdev_priv(dev); 326 slave_t *slave; 327 328 spin_lock(&eql->queue.lock); 329 330 slave = __eql_schedule_slaves(&eql->queue); 331 if (slave) { 332 struct net_device *slave_dev = slave->dev; 333 334 skb->dev = slave_dev; 335 skb->priority = 1; 336 slave->bytes_queued += skb->len; 337 dev_queue_xmit(skb); 338 dev->stats.tx_packets++; 339 } else { 340 dev->stats.tx_dropped++; 341 dev_kfree_skb(skb); 342 } 343 344 spin_unlock(&eql->queue.lock); 345 346 return 0; 347 } 348 349 /* 350 * Private ioctl functions 351 */ 352 353 /* queue->lock must be held */ 354 static slave_t *__eql_find_slave_dev(slave_queue_t *queue, struct net_device *dev) 355 { 356 struct list_head *this, *head; 357 358 head = &queue->all_slaves; 359 list_for_each(this, head) { 360 slave_t *slave = list_entry(this, slave_t, list); 361 362 if (slave->dev == dev) 363 return slave; 364 } 365 366 return NULL; 367 } 368 369 static inline int eql_is_full(slave_queue_t *queue) 370 { 371 equalizer_t *eql = netdev_priv(queue->master_dev); 372 373 if (queue->num_slaves >= eql->max_slaves) 374 return 1; 375 return 0; 376 } 377 378 /* queue->lock must be held */ 379 static int __eql_insert_slave(slave_queue_t *queue, slave_t *slave) 380 { 381 if (!eql_is_full(queue)) { 382 slave_t *duplicate_slave = NULL; 383 384 duplicate_slave = __eql_find_slave_dev(queue, slave->dev); 385 if (duplicate_slave) 386 eql_kill_one_slave(queue, duplicate_slave); 387 388 list_add(&slave->list, &queue->all_slaves); 389 queue->num_slaves++; 390 slave->dev->flags |= IFF_SLAVE; 391 392 return 0; 393 } 394 395 return -ENOSPC; 396 } 397 398 static int eql_enslave(struct net_device *master_dev, slaving_request_t __user *srqp) 399 { 400 struct net_device *slave_dev; 401 slaving_request_t srq; 402 403 if (copy_from_user(&srq, srqp, sizeof (slaving_request_t))) 404 return -EFAULT; 405 406 slave_dev = dev_get_by_name(&init_net, srq.slave_name); 407 if (slave_dev) { 408 if ((master_dev->flags & IFF_UP) == IFF_UP) { 409 /* slave is not a master & not already a slave: */ 410 if (!eql_is_master(slave_dev) && 411 !eql_is_slave(slave_dev)) { 412 slave_t *s = kmalloc(sizeof(*s), GFP_KERNEL); 413 equalizer_t *eql = netdev_priv(master_dev); 414 int ret; 415 416 if (!s) { 417 dev_put(slave_dev); 418 return -ENOMEM; 419 } 420 421 memset(s, 0, sizeof(*s)); 422 s->dev = slave_dev; 423 s->priority = srq.priority; 424 s->priority_bps = srq.priority; 425 s->priority_Bps = srq.priority / 8; 426 427 spin_lock_bh(&eql->queue.lock); 428 ret = __eql_insert_slave(&eql->queue, s); 429 if (ret) { 430 dev_put(slave_dev); 431 kfree(s); 432 } 433 spin_unlock_bh(&eql->queue.lock); 434 435 return ret; 436 } 437 } 438 dev_put(slave_dev); 439 } 440 441 return -EINVAL; 442 } 443 444 static int eql_emancipate(struct net_device *master_dev, slaving_request_t __user *srqp) 445 { 446 equalizer_t *eql = netdev_priv(master_dev); 447 struct net_device *slave_dev; 448 slaving_request_t srq; 449 int ret; 450 451 if (copy_from_user(&srq, srqp, sizeof (slaving_request_t))) 452 return -EFAULT; 453 454 slave_dev = dev_get_by_name(&init_net, srq.slave_name); 455 ret = -EINVAL; 456 if (slave_dev) { 457 spin_lock_bh(&eql->queue.lock); 458 459 if (eql_is_slave(slave_dev)) { 460 slave_t *slave = __eql_find_slave_dev(&eql->queue, 461 slave_dev); 462 463 if (slave) { 464 eql_kill_one_slave(&eql->queue, slave); 465 ret = 0; 466 } 467 } 468 dev_put(slave_dev); 469 470 spin_unlock_bh(&eql->queue.lock); 471 } 472 473 return ret; 474 } 475 476 static int eql_g_slave_cfg(struct net_device *dev, slave_config_t __user *scp) 477 { 478 equalizer_t *eql = netdev_priv(dev); 479 slave_t *slave; 480 struct net_device *slave_dev; 481 slave_config_t sc; 482 int ret; 483 484 if (copy_from_user(&sc, scp, sizeof (slave_config_t))) 485 return -EFAULT; 486 487 slave_dev = dev_get_by_name(&init_net, sc.slave_name); 488 if (!slave_dev) 489 return -ENODEV; 490 491 ret = -EINVAL; 492 493 spin_lock_bh(&eql->queue.lock); 494 if (eql_is_slave(slave_dev)) { 495 slave = __eql_find_slave_dev(&eql->queue, slave_dev); 496 if (slave) { 497 sc.priority = slave->priority; 498 ret = 0; 499 } 500 } 501 spin_unlock_bh(&eql->queue.lock); 502 503 dev_put(slave_dev); 504 505 if (!ret && copy_to_user(scp, &sc, sizeof (slave_config_t))) 506 ret = -EFAULT; 507 508 return ret; 509 } 510 511 static int eql_s_slave_cfg(struct net_device *dev, slave_config_t __user *scp) 512 { 513 slave_t *slave; 514 equalizer_t *eql; 515 struct net_device *slave_dev; 516 slave_config_t sc; 517 int ret; 518 519 if (copy_from_user(&sc, scp, sizeof (slave_config_t))) 520 return -EFAULT; 521 522 slave_dev = dev_get_by_name(&init_net, sc.slave_name); 523 if (!slave_dev) 524 return -ENODEV; 525 526 ret = -EINVAL; 527 528 eql = netdev_priv(dev); 529 spin_lock_bh(&eql->queue.lock); 530 if (eql_is_slave(slave_dev)) { 531 slave = __eql_find_slave_dev(&eql->queue, slave_dev); 532 if (slave) { 533 slave->priority = sc.priority; 534 slave->priority_bps = sc.priority; 535 slave->priority_Bps = sc.priority / 8; 536 ret = 0; 537 } 538 } 539 spin_unlock_bh(&eql->queue.lock); 540 541 return ret; 542 } 543 544 static int eql_g_master_cfg(struct net_device *dev, master_config_t __user *mcp) 545 { 546 equalizer_t *eql; 547 master_config_t mc; 548 549 if (eql_is_master(dev)) { 550 eql = netdev_priv(dev); 551 mc.max_slaves = eql->max_slaves; 552 mc.min_slaves = eql->min_slaves; 553 if (copy_to_user(mcp, &mc, sizeof (master_config_t))) 554 return -EFAULT; 555 return 0; 556 } 557 return -EINVAL; 558 } 559 560 static int eql_s_master_cfg(struct net_device *dev, master_config_t __user *mcp) 561 { 562 equalizer_t *eql; 563 master_config_t mc; 564 565 if (copy_from_user(&mc, mcp, sizeof (master_config_t))) 566 return -EFAULT; 567 568 if (eql_is_master(dev)) { 569 eql = netdev_priv(dev); 570 eql->max_slaves = mc.max_slaves; 571 eql->min_slaves = mc.min_slaves; 572 return 0; 573 } 574 return -EINVAL; 575 } 576 577 static struct net_device *dev_eql; 578 579 static int __init eql_init_module(void) 580 { 581 int err; 582 583 printk(version); 584 585 dev_eql = alloc_netdev(sizeof(equalizer_t), "eql", eql_setup); 586 if (!dev_eql) 587 return -ENOMEM; 588 589 err = register_netdev(dev_eql); 590 if (err) 591 free_netdev(dev_eql); 592 return err; 593 } 594 595 static void __exit eql_cleanup_module(void) 596 { 597 unregister_netdev(dev_eql); 598 free_netdev(dev_eql); 599 } 600 601 module_init(eql_init_module); 602 module_exit(eql_cleanup_module); 603 MODULE_LICENSE("GPL"); 604