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/compat.h> 117 #include <linux/capability.h> 118 #include <linux/module.h> 119 #include <linux/kernel.h> 120 #include <linux/init.h> 121 #include <linux/slab.h> 122 #include <linux/timer.h> 123 #include <linux/netdevice.h> 124 #include <net/net_namespace.h> 125 126 #include <linux/if.h> 127 #include <linux/if_arp.h> 128 #include <linux/if_eql.h> 129 #include <linux/pkt_sched.h> 130 131 #include <linux/uaccess.h> 132 133 static int eql_open(struct net_device *dev); 134 static int eql_close(struct net_device *dev); 135 static int eql_siocdevprivate(struct net_device *dev, struct ifreq *ifr, 136 void __user *data, int cmd); 137 static netdev_tx_t eql_slave_xmit(struct sk_buff *skb, struct net_device *dev); 138 139 #define eql_is_slave(dev) ((dev->flags & IFF_SLAVE) == IFF_SLAVE) 140 #define eql_is_master(dev) ((dev->flags & IFF_MASTER) == IFF_MASTER) 141 142 static void eql_kill_one_slave(slave_queue_t *queue, slave_t *slave); 143 144 static void eql_timer(struct timer_list *t) 145 { 146 equalizer_t *eql = from_timer(eql, t, timer); 147 struct list_head *this, *tmp, *head; 148 149 spin_lock(&eql->queue.lock); 150 head = &eql->queue.all_slaves; 151 list_for_each_safe(this, tmp, head) { 152 slave_t *slave = list_entry(this, slave_t, list); 153 154 if ((slave->dev->flags & IFF_UP) == IFF_UP) { 155 slave->bytes_queued -= slave->priority_Bps; 156 if (slave->bytes_queued < 0) 157 slave->bytes_queued = 0; 158 } else { 159 eql_kill_one_slave(&eql->queue, slave); 160 } 161 162 } 163 spin_unlock(&eql->queue.lock); 164 165 eql->timer.expires = jiffies + EQL_DEFAULT_RESCHED_IVAL; 166 add_timer(&eql->timer); 167 } 168 169 static const char version[] __initconst = 170 "Equalizer2002: Simon Janes (simon@ncm.com) and David S. Miller (davem@redhat.com)"; 171 172 static const struct net_device_ops eql_netdev_ops = { 173 .ndo_open = eql_open, 174 .ndo_stop = eql_close, 175 .ndo_siocdevprivate = eql_siocdevprivate, 176 .ndo_start_xmit = eql_slave_xmit, 177 }; 178 179 static void __init eql_setup(struct net_device *dev) 180 { 181 equalizer_t *eql = netdev_priv(dev); 182 183 timer_setup(&eql->timer, eql_timer, 0); 184 eql->timer.expires = jiffies + EQL_DEFAULT_RESCHED_IVAL; 185 186 spin_lock_init(&eql->queue.lock); 187 INIT_LIST_HEAD(&eql->queue.all_slaves); 188 eql->queue.master_dev = dev; 189 190 dev->netdev_ops = &eql_netdev_ops; 191 192 /* 193 * Now we undo some of the things that eth_setup does 194 * that we don't like 195 */ 196 197 dev->mtu = EQL_DEFAULT_MTU; /* set to 576 in if_eql.h */ 198 dev->flags = IFF_MASTER; 199 200 dev->type = ARPHRD_SLIP; 201 dev->tx_queue_len = 5; /* Hands them off fast */ 202 netif_keep_dst(dev); 203 } 204 205 static int eql_open(struct net_device *dev) 206 { 207 equalizer_t *eql = netdev_priv(dev); 208 209 /* XXX We should force this off automatically for the user. */ 210 netdev_info(dev, 211 "remember to turn off Van-Jacobson compression on your slave devices\n"); 212 213 BUG_ON(!list_empty(&eql->queue.all_slaves)); 214 215 eql->min_slaves = 1; 216 eql->max_slaves = EQL_DEFAULT_MAX_SLAVES; /* 4 usually... */ 217 218 add_timer(&eql->timer); 219 220 return 0; 221 } 222 223 static void eql_kill_one_slave(slave_queue_t *queue, slave_t *slave) 224 { 225 list_del(&slave->list); 226 queue->num_slaves--; 227 slave->dev->flags &= ~IFF_SLAVE; 228 dev_put_track(slave->dev, &slave->dev_tracker); 229 kfree(slave); 230 } 231 232 static void eql_kill_slave_queue(slave_queue_t *queue) 233 { 234 struct list_head *head, *tmp, *this; 235 236 spin_lock_bh(&queue->lock); 237 238 head = &queue->all_slaves; 239 list_for_each_safe(this, tmp, head) { 240 slave_t *s = list_entry(this, slave_t, list); 241 242 eql_kill_one_slave(queue, s); 243 } 244 245 spin_unlock_bh(&queue->lock); 246 } 247 248 static int eql_close(struct net_device *dev) 249 { 250 equalizer_t *eql = netdev_priv(dev); 251 252 /* 253 * The timer has to be stopped first before we start hacking away 254 * at the data structure it scans every so often... 255 */ 256 257 del_timer_sync(&eql->timer); 258 259 eql_kill_slave_queue(&eql->queue); 260 261 return 0; 262 } 263 264 static int eql_enslave(struct net_device *dev, slaving_request_t __user *srq); 265 static int eql_emancipate(struct net_device *dev, slaving_request_t __user *srq); 266 267 static int eql_g_slave_cfg(struct net_device *dev, slave_config_t __user *sc); 268 static int eql_s_slave_cfg(struct net_device *dev, slave_config_t __user *sc); 269 270 static int eql_g_master_cfg(struct net_device *dev, master_config_t __user *mc); 271 static int eql_s_master_cfg(struct net_device *dev, master_config_t __user *mc); 272 273 static int eql_siocdevprivate(struct net_device *dev, struct ifreq *ifr, 274 void __user *data, int cmd) 275 { 276 if (cmd != EQL_GETMASTRCFG && cmd != EQL_GETSLAVECFG && 277 !capable(CAP_NET_ADMIN)) 278 return -EPERM; 279 280 if (in_compat_syscall()) /* to be implemented */ 281 return -EOPNOTSUPP; 282 283 switch (cmd) { 284 case EQL_ENSLAVE: 285 return eql_enslave(dev, data); 286 case EQL_EMANCIPATE: 287 return eql_emancipate(dev, data); 288 case EQL_GETSLAVECFG: 289 return eql_g_slave_cfg(dev, data); 290 case EQL_SETSLAVECFG: 291 return eql_s_slave_cfg(dev, data); 292 case EQL_GETMASTRCFG: 293 return eql_g_master_cfg(dev, data); 294 case EQL_SETMASTRCFG: 295 return eql_s_master_cfg(dev, data); 296 default: 297 return -EOPNOTSUPP; 298 } 299 } 300 301 /* queue->lock must be held */ 302 static slave_t *__eql_schedule_slaves(slave_queue_t *queue) 303 { 304 unsigned long best_load = ~0UL; 305 struct list_head *this, *tmp, *head; 306 slave_t *best_slave; 307 308 best_slave = NULL; 309 310 /* Make a pass to set the best slave. */ 311 head = &queue->all_slaves; 312 list_for_each_safe(this, tmp, head) { 313 slave_t *slave = list_entry(this, slave_t, list); 314 unsigned long slave_load, bytes_queued, priority_Bps; 315 316 /* Go through the slave list once, updating best_slave 317 * whenever a new best_load is found. 318 */ 319 bytes_queued = slave->bytes_queued; 320 priority_Bps = slave->priority_Bps; 321 if ((slave->dev->flags & IFF_UP) == IFF_UP) { 322 slave_load = (~0UL - (~0UL / 2)) - 323 (priority_Bps) + bytes_queued * 8; 324 325 if (slave_load < best_load) { 326 best_load = slave_load; 327 best_slave = slave; 328 } 329 } else { 330 /* We found a dead slave, kill it. */ 331 eql_kill_one_slave(queue, slave); 332 } 333 } 334 return best_slave; 335 } 336 337 static netdev_tx_t eql_slave_xmit(struct sk_buff *skb, struct net_device *dev) 338 { 339 equalizer_t *eql = netdev_priv(dev); 340 slave_t *slave; 341 342 spin_lock(&eql->queue.lock); 343 344 slave = __eql_schedule_slaves(&eql->queue); 345 if (slave) { 346 struct net_device *slave_dev = slave->dev; 347 348 skb->dev = slave_dev; 349 skb->priority = TC_PRIO_FILLER; 350 slave->bytes_queued += skb->len; 351 dev_queue_xmit(skb); 352 dev->stats.tx_packets++; 353 } else { 354 dev->stats.tx_dropped++; 355 dev_kfree_skb(skb); 356 } 357 358 spin_unlock(&eql->queue.lock); 359 360 return NETDEV_TX_OK; 361 } 362 363 /* 364 * Private ioctl functions 365 */ 366 367 /* queue->lock must be held */ 368 static slave_t *__eql_find_slave_dev(slave_queue_t *queue, struct net_device *dev) 369 { 370 struct list_head *this, *head; 371 372 head = &queue->all_slaves; 373 list_for_each(this, head) { 374 slave_t *slave = list_entry(this, slave_t, list); 375 376 if (slave->dev == dev) 377 return slave; 378 } 379 380 return NULL; 381 } 382 383 static inline int eql_is_full(slave_queue_t *queue) 384 { 385 equalizer_t *eql = netdev_priv(queue->master_dev); 386 387 if (queue->num_slaves >= eql->max_slaves) 388 return 1; 389 return 0; 390 } 391 392 /* queue->lock must be held */ 393 static int __eql_insert_slave(slave_queue_t *queue, slave_t *slave) 394 { 395 if (!eql_is_full(queue)) { 396 slave_t *duplicate_slave = NULL; 397 398 duplicate_slave = __eql_find_slave_dev(queue, slave->dev); 399 if (duplicate_slave) 400 eql_kill_one_slave(queue, duplicate_slave); 401 402 dev_hold_track(slave->dev, &slave->dev_tracker, GFP_ATOMIC); 403 list_add(&slave->list, &queue->all_slaves); 404 queue->num_slaves++; 405 slave->dev->flags |= IFF_SLAVE; 406 407 return 0; 408 } 409 410 return -ENOSPC; 411 } 412 413 static int eql_enslave(struct net_device *master_dev, slaving_request_t __user *srqp) 414 { 415 struct net_device *slave_dev; 416 slaving_request_t srq; 417 418 if (copy_from_user(&srq, srqp, sizeof (slaving_request_t))) 419 return -EFAULT; 420 421 slave_dev = __dev_get_by_name(&init_net, srq.slave_name); 422 if (!slave_dev) 423 return -ENODEV; 424 425 if ((master_dev->flags & IFF_UP) == IFF_UP) { 426 /* slave is not a master & not already a slave: */ 427 if (!eql_is_master(slave_dev) && !eql_is_slave(slave_dev)) { 428 slave_t *s = kmalloc(sizeof(*s), GFP_KERNEL); 429 equalizer_t *eql = netdev_priv(master_dev); 430 int ret; 431 432 if (!s) 433 return -ENOMEM; 434 435 memset(s, 0, sizeof(*s)); 436 s->dev = slave_dev; 437 s->priority = srq.priority; 438 s->priority_bps = srq.priority; 439 s->priority_Bps = srq.priority / 8; 440 441 spin_lock_bh(&eql->queue.lock); 442 ret = __eql_insert_slave(&eql->queue, s); 443 if (ret) 444 kfree(s); 445 446 spin_unlock_bh(&eql->queue.lock); 447 448 return ret; 449 } 450 } 451 452 return -EINVAL; 453 } 454 455 static int eql_emancipate(struct net_device *master_dev, slaving_request_t __user *srqp) 456 { 457 equalizer_t *eql = netdev_priv(master_dev); 458 struct net_device *slave_dev; 459 slaving_request_t srq; 460 int ret; 461 462 if (copy_from_user(&srq, srqp, sizeof (slaving_request_t))) 463 return -EFAULT; 464 465 slave_dev = __dev_get_by_name(&init_net, srq.slave_name); 466 if (!slave_dev) 467 return -ENODEV; 468 469 ret = -EINVAL; 470 spin_lock_bh(&eql->queue.lock); 471 if (eql_is_slave(slave_dev)) { 472 slave_t *slave = __eql_find_slave_dev(&eql->queue, slave_dev); 473 if (slave) { 474 eql_kill_one_slave(&eql->queue, slave); 475 ret = 0; 476 } 477 } 478 spin_unlock_bh(&eql->queue.lock); 479 480 return ret; 481 } 482 483 static int eql_g_slave_cfg(struct net_device *dev, slave_config_t __user *scp) 484 { 485 equalizer_t *eql = netdev_priv(dev); 486 slave_t *slave; 487 struct net_device *slave_dev; 488 slave_config_t sc; 489 int ret; 490 491 if (copy_from_user(&sc, scp, sizeof (slave_config_t))) 492 return -EFAULT; 493 494 slave_dev = __dev_get_by_name(&init_net, sc.slave_name); 495 if (!slave_dev) 496 return -ENODEV; 497 498 ret = -EINVAL; 499 500 spin_lock_bh(&eql->queue.lock); 501 if (eql_is_slave(slave_dev)) { 502 slave = __eql_find_slave_dev(&eql->queue, slave_dev); 503 if (slave) { 504 sc.priority = slave->priority; 505 ret = 0; 506 } 507 } 508 spin_unlock_bh(&eql->queue.lock); 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 return ret; 547 } 548 549 static int eql_g_master_cfg(struct net_device *dev, master_config_t __user *mcp) 550 { 551 equalizer_t *eql; 552 master_config_t mc; 553 554 memset(&mc, 0, sizeof(master_config_t)); 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 pr_info("%s\n", version); 591 592 dev_eql = alloc_netdev(sizeof(equalizer_t), "eql", NET_NAME_UNKNOWN, 593 eql_setup); 594 if (!dev_eql) 595 return -ENOMEM; 596 597 err = register_netdev(dev_eql); 598 if (err) 599 free_netdev(dev_eql); 600 return err; 601 } 602 603 static void __exit eql_cleanup_module(void) 604 { 605 unregister_netdev(dev_eql); 606 free_netdev(dev_eql); 607 } 608 609 module_init(eql_init_module); 610 module_exit(eql_cleanup_module); 611 MODULE_LICENSE("GPL"); 612