| 1 | /* net/sched/sch_teql.c "True" (or "trivial") link equalizer. |
| 2 | * |
| 3 | * This program is free software; you can redistribute it and/or |
| 4 | * modify it under the terms of the GNU General Public License |
| 5 | * as published by the Free Software Foundation; either version |
| 6 | * 2 of the License, or (at your option) any later version. |
| 7 | * |
| 8 | * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru> |
| 9 | */ |
| 10 | |
| 11 | #include <linux/module.h> |
| 12 | #include <linux/types.h> |
| 13 | #include <linux/kernel.h> |
| 14 | #include <linux/slab.h> |
| 15 | #include <linux/string.h> |
| 16 | #include <linux/errno.h> |
| 17 | #include <linux/if_arp.h> |
| 18 | #include <linux/netdevice.h> |
| 19 | #include <linux/init.h> |
| 20 | #include <linux/skbuff.h> |
| 21 | #include <linux/moduleparam.h> |
| 22 | #include <net/dst.h> |
| 23 | #include <net/neighbour.h> |
| 24 | #include <net/pkt_sched.h> |
| 25 | |
| 26 | /* |
| 27 | How to setup it. |
| 28 | ---------------- |
| 29 | |
| 30 | After loading this module you will find a new device teqlN |
| 31 | and new qdisc with the same name. To join a slave to the equalizer |
| 32 | you should just set this qdisc on a device f.e. |
| 33 | |
| 34 | # tc qdisc add dev eth0 root teql0 |
| 35 | # tc qdisc add dev eth1 root teql0 |
| 36 | |
| 37 | That's all. Full PnP 8) |
| 38 | |
| 39 | Applicability. |
| 40 | -------------- |
| 41 | |
| 42 | 1. Slave devices MUST be active devices, i.e., they must raise the tbusy |
| 43 | signal and generate EOI events. If you want to equalize virtual devices |
| 44 | like tunnels, use a normal eql device. |
| 45 | 2. This device puts no limitations on physical slave characteristics |
| 46 | f.e. it will equalize 9600baud line and 100Mb ethernet perfectly :-) |
| 47 | Certainly, large difference in link speeds will make the resulting |
| 48 | eqalized link unusable, because of huge packet reordering. |
| 49 | I estimate an upper useful difference as ~10 times. |
| 50 | 3. If the slave requires address resolution, only protocols using |
| 51 | neighbour cache (IPv4/IPv6) will work over the equalized link. |
| 52 | Other protocols are still allowed to use the slave device directly, |
| 53 | which will not break load balancing, though native slave |
| 54 | traffic will have the highest priority. */ |
| 55 | |
| 56 | struct teql_master { |
| 57 | struct Qdisc_ops qops; |
| 58 | struct net_device *dev; |
| 59 | struct Qdisc *slaves; |
| 60 | struct list_head master_list; |
| 61 | unsigned long tx_bytes; |
| 62 | unsigned long tx_packets; |
| 63 | unsigned long tx_errors; |
| 64 | unsigned long tx_dropped; |
| 65 | }; |
| 66 | |
| 67 | struct teql_sched_data { |
| 68 | struct Qdisc *next; |
| 69 | struct teql_master *m; |
| 70 | struct neighbour *ncache; |
| 71 | struct sk_buff_head q; |
| 72 | }; |
| 73 | |
| 74 | #define NEXT_SLAVE(q) (((struct teql_sched_data *)qdisc_priv(q))->next) |
| 75 | |
| 76 | #define FMASK (IFF_BROADCAST | IFF_POINTOPOINT) |
| 77 | |
| 78 | /* "teql*" qdisc routines */ |
| 79 | |
| 80 | static int |
| 81 | teql_enqueue(struct sk_buff *skb, struct Qdisc *sch) |
| 82 | { |
| 83 | struct net_device *dev = qdisc_dev(sch); |
| 84 | struct teql_sched_data *q = qdisc_priv(sch); |
| 85 | |
| 86 | if (q->q.qlen < dev->tx_queue_len) { |
| 87 | __skb_queue_tail(&q->q, skb); |
| 88 | return NET_XMIT_SUCCESS; |
| 89 | } |
| 90 | |
| 91 | return qdisc_drop(skb, sch); |
| 92 | } |
| 93 | |
| 94 | static struct sk_buff * |
| 95 | teql_dequeue(struct Qdisc *sch) |
| 96 | { |
| 97 | struct teql_sched_data *dat = qdisc_priv(sch); |
| 98 | struct netdev_queue *dat_queue; |
| 99 | struct sk_buff *skb; |
| 100 | |
| 101 | skb = __skb_dequeue(&dat->q); |
| 102 | dat_queue = netdev_get_tx_queue(dat->m->dev, 0); |
| 103 | if (skb == NULL) { |
| 104 | struct net_device *m = qdisc_dev(dat_queue->qdisc); |
| 105 | if (m) { |
| 106 | dat->m->slaves = sch; |
| 107 | netif_wake_queue(m); |
| 108 | } |
| 109 | } else { |
| 110 | qdisc_bstats_update(sch, skb); |
| 111 | } |
| 112 | sch->q.qlen = dat->q.qlen + dat_queue->qdisc->q.qlen; |
| 113 | return skb; |
| 114 | } |
| 115 | |
| 116 | static struct sk_buff * |
| 117 | teql_peek(struct Qdisc *sch) |
| 118 | { |
| 119 | /* teql is meant to be used as root qdisc */ |
| 120 | return NULL; |
| 121 | } |
| 122 | |
| 123 | static inline void |
| 124 | teql_neigh_release(struct neighbour *n) |
| 125 | { |
| 126 | if (n) |
| 127 | neigh_release(n); |
| 128 | } |
| 129 | |
| 130 | static void |
| 131 | teql_reset(struct Qdisc *sch) |
| 132 | { |
| 133 | struct teql_sched_data *dat = qdisc_priv(sch); |
| 134 | |
| 135 | skb_queue_purge(&dat->q); |
| 136 | sch->q.qlen = 0; |
| 137 | teql_neigh_release(xchg(&dat->ncache, NULL)); |
| 138 | } |
| 139 | |
| 140 | static void |
| 141 | teql_destroy(struct Qdisc *sch) |
| 142 | { |
| 143 | struct Qdisc *q, *prev; |
| 144 | struct teql_sched_data *dat = qdisc_priv(sch); |
| 145 | struct teql_master *master = dat->m; |
| 146 | |
| 147 | prev = master->slaves; |
| 148 | if (prev) { |
| 149 | do { |
| 150 | q = NEXT_SLAVE(prev); |
| 151 | if (q == sch) { |
| 152 | NEXT_SLAVE(prev) = NEXT_SLAVE(q); |
| 153 | if (q == master->slaves) { |
| 154 | master->slaves = NEXT_SLAVE(q); |
| 155 | if (q == master->slaves) { |
| 156 | struct netdev_queue *txq; |
| 157 | spinlock_t *root_lock; |
| 158 | |
| 159 | txq = netdev_get_tx_queue(master->dev, 0); |
| 160 | master->slaves = NULL; |
| 161 | |
| 162 | root_lock = qdisc_root_sleeping_lock(txq->qdisc); |
| 163 | spin_lock_bh(root_lock); |
| 164 | qdisc_reset(txq->qdisc); |
| 165 | spin_unlock_bh(root_lock); |
| 166 | } |
| 167 | } |
| 168 | skb_queue_purge(&dat->q); |
| 169 | teql_neigh_release(xchg(&dat->ncache, NULL)); |
| 170 | break; |
| 171 | } |
| 172 | |
| 173 | } while ((prev = q) != master->slaves); |
| 174 | } |
| 175 | } |
| 176 | |
| 177 | static int teql_qdisc_init(struct Qdisc *sch, struct nlattr *opt) |
| 178 | { |
| 179 | struct net_device *dev = qdisc_dev(sch); |
| 180 | struct teql_master *m = (struct teql_master *)sch->ops; |
| 181 | struct teql_sched_data *q = qdisc_priv(sch); |
| 182 | |
| 183 | if (dev->hard_header_len > m->dev->hard_header_len) |
| 184 | return -EINVAL; |
| 185 | |
| 186 | if (m->dev == dev) |
| 187 | return -ELOOP; |
| 188 | |
| 189 | q->m = m; |
| 190 | |
| 191 | skb_queue_head_init(&q->q); |
| 192 | |
| 193 | if (m->slaves) { |
| 194 | if (m->dev->flags & IFF_UP) { |
| 195 | if ((m->dev->flags & IFF_POINTOPOINT && |
| 196 | !(dev->flags & IFF_POINTOPOINT)) || |
| 197 | (m->dev->flags & IFF_BROADCAST && |
| 198 | !(dev->flags & IFF_BROADCAST)) || |
| 199 | (m->dev->flags & IFF_MULTICAST && |
| 200 | !(dev->flags & IFF_MULTICAST)) || |
| 201 | dev->mtu < m->dev->mtu) |
| 202 | return -EINVAL; |
| 203 | } else { |
| 204 | if (!(dev->flags&IFF_POINTOPOINT)) |
| 205 | m->dev->flags &= ~IFF_POINTOPOINT; |
| 206 | if (!(dev->flags&IFF_BROADCAST)) |
| 207 | m->dev->flags &= ~IFF_BROADCAST; |
| 208 | if (!(dev->flags&IFF_MULTICAST)) |
| 209 | m->dev->flags &= ~IFF_MULTICAST; |
| 210 | if (dev->mtu < m->dev->mtu) |
| 211 | m->dev->mtu = dev->mtu; |
| 212 | } |
| 213 | q->next = NEXT_SLAVE(m->slaves); |
| 214 | NEXT_SLAVE(m->slaves) = sch; |
| 215 | } else { |
| 216 | q->next = sch; |
| 217 | m->slaves = sch; |
| 218 | m->dev->mtu = dev->mtu; |
| 219 | m->dev->flags = (m->dev->flags&~FMASK)|(dev->flags&FMASK); |
| 220 | } |
| 221 | return 0; |
| 222 | } |
| 223 | |
| 224 | |
| 225 | static int |
| 226 | __teql_resolve(struct sk_buff *skb, struct sk_buff *skb_res, |
| 227 | struct net_device *dev, struct netdev_queue *txq, |
| 228 | struct neighbour *mn) |
| 229 | { |
| 230 | struct teql_sched_data *q = qdisc_priv(txq->qdisc); |
| 231 | struct neighbour *n = q->ncache; |
| 232 | |
| 233 | if (mn->tbl == NULL) |
| 234 | return -EINVAL; |
| 235 | if (n && n->tbl == mn->tbl && |
| 236 | memcmp(n->primary_key, mn->primary_key, mn->tbl->key_len) == 0) { |
| 237 | atomic_inc(&n->refcnt); |
| 238 | } else { |
| 239 | n = __neigh_lookup_errno(mn->tbl, mn->primary_key, dev); |
| 240 | if (IS_ERR(n)) |
| 241 | return PTR_ERR(n); |
| 242 | } |
| 243 | if (neigh_event_send(n, skb_res) == 0) { |
| 244 | int err; |
| 245 | char haddr[MAX_ADDR_LEN]; |
| 246 | |
| 247 | neigh_ha_snapshot(haddr, n, dev); |
| 248 | err = dev_hard_header(skb, dev, ntohs(skb->protocol), haddr, |
| 249 | NULL, skb->len); |
| 250 | |
| 251 | if (err < 0) { |
| 252 | neigh_release(n); |
| 253 | return -EINVAL; |
| 254 | } |
| 255 | teql_neigh_release(xchg(&q->ncache, n)); |
| 256 | return 0; |
| 257 | } |
| 258 | neigh_release(n); |
| 259 | return (skb_res == NULL) ? -EAGAIN : 1; |
| 260 | } |
| 261 | |
| 262 | static inline int teql_resolve(struct sk_buff *skb, |
| 263 | struct sk_buff *skb_res, |
| 264 | struct net_device *dev, |
| 265 | struct netdev_queue *txq) |
| 266 | { |
| 267 | struct dst_entry *dst = skb_dst(skb); |
| 268 | struct neighbour *mn; |
| 269 | int res; |
| 270 | |
| 271 | if (txq->qdisc == &noop_qdisc) |
| 272 | return -ENODEV; |
| 273 | |
| 274 | if (!dev->header_ops || !dst) |
| 275 | return 0; |
| 276 | |
| 277 | rcu_read_lock(); |
| 278 | mn = dst_get_neighbour_noref(dst); |
| 279 | res = mn ? __teql_resolve(skb, skb_res, dev, txq, mn) : 0; |
| 280 | rcu_read_unlock(); |
| 281 | |
| 282 | return res; |
| 283 | } |
| 284 | |
| 285 | static netdev_tx_t teql_master_xmit(struct sk_buff *skb, struct net_device *dev) |
| 286 | { |
| 287 | struct teql_master *master = netdev_priv(dev); |
| 288 | struct Qdisc *start, *q; |
| 289 | int busy; |
| 290 | int nores; |
| 291 | int subq = skb_get_queue_mapping(skb); |
| 292 | struct sk_buff *skb_res = NULL; |
| 293 | |
| 294 | start = master->slaves; |
| 295 | |
| 296 | restart: |
| 297 | nores = 0; |
| 298 | busy = 0; |
| 299 | |
| 300 | q = start; |
| 301 | if (!q) |
| 302 | goto drop; |
| 303 | |
| 304 | do { |
| 305 | struct net_device *slave = qdisc_dev(q); |
| 306 | struct netdev_queue *slave_txq = netdev_get_tx_queue(slave, 0); |
| 307 | const struct net_device_ops *slave_ops = slave->netdev_ops; |
| 308 | |
| 309 | if (slave_txq->qdisc_sleeping != q) |
| 310 | continue; |
| 311 | if (netif_xmit_stopped(netdev_get_tx_queue(slave, subq)) || |
| 312 | !netif_running(slave)) { |
| 313 | busy = 1; |
| 314 | continue; |
| 315 | } |
| 316 | |
| 317 | switch (teql_resolve(skb, skb_res, slave, slave_txq)) { |
| 318 | case 0: |
| 319 | if (__netif_tx_trylock(slave_txq)) { |
| 320 | unsigned int length = qdisc_pkt_len(skb); |
| 321 | |
| 322 | if (!netif_xmit_frozen_or_stopped(slave_txq) && |
| 323 | slave_ops->ndo_start_xmit(skb, slave) == NETDEV_TX_OK) { |
| 324 | txq_trans_update(slave_txq); |
| 325 | __netif_tx_unlock(slave_txq); |
| 326 | master->slaves = NEXT_SLAVE(q); |
| 327 | netif_wake_queue(dev); |
| 328 | master->tx_packets++; |
| 329 | master->tx_bytes += length; |
| 330 | return NETDEV_TX_OK; |
| 331 | } |
| 332 | __netif_tx_unlock(slave_txq); |
| 333 | } |
| 334 | if (netif_xmit_stopped(netdev_get_tx_queue(dev, 0))) |
| 335 | busy = 1; |
| 336 | break; |
| 337 | case 1: |
| 338 | master->slaves = NEXT_SLAVE(q); |
| 339 | return NETDEV_TX_OK; |
| 340 | default: |
| 341 | nores = 1; |
| 342 | break; |
| 343 | } |
| 344 | __skb_pull(skb, skb_network_offset(skb)); |
| 345 | } while ((q = NEXT_SLAVE(q)) != start); |
| 346 | |
| 347 | if (nores && skb_res == NULL) { |
| 348 | skb_res = skb; |
| 349 | goto restart; |
| 350 | } |
| 351 | |
| 352 | if (busy) { |
| 353 | netif_stop_queue(dev); |
| 354 | return NETDEV_TX_BUSY; |
| 355 | } |
| 356 | master->tx_errors++; |
| 357 | |
| 358 | drop: |
| 359 | master->tx_dropped++; |
| 360 | dev_kfree_skb(skb); |
| 361 | return NETDEV_TX_OK; |
| 362 | } |
| 363 | |
| 364 | static int teql_master_open(struct net_device *dev) |
| 365 | { |
| 366 | struct Qdisc *q; |
| 367 | struct teql_master *m = netdev_priv(dev); |
| 368 | int mtu = 0xFFFE; |
| 369 | unsigned int flags = IFF_NOARP | IFF_MULTICAST; |
| 370 | |
| 371 | if (m->slaves == NULL) |
| 372 | return -EUNATCH; |
| 373 | |
| 374 | flags = FMASK; |
| 375 | |
| 376 | q = m->slaves; |
| 377 | do { |
| 378 | struct net_device *slave = qdisc_dev(q); |
| 379 | |
| 380 | if (slave == NULL) |
| 381 | return -EUNATCH; |
| 382 | |
| 383 | if (slave->mtu < mtu) |
| 384 | mtu = slave->mtu; |
| 385 | if (slave->hard_header_len > LL_MAX_HEADER) |
| 386 | return -EINVAL; |
| 387 | |
| 388 | /* If all the slaves are BROADCAST, master is BROADCAST |
| 389 | If all the slaves are PtP, master is PtP |
| 390 | Otherwise, master is NBMA. |
| 391 | */ |
| 392 | if (!(slave->flags&IFF_POINTOPOINT)) |
| 393 | flags &= ~IFF_POINTOPOINT; |
| 394 | if (!(slave->flags&IFF_BROADCAST)) |
| 395 | flags &= ~IFF_BROADCAST; |
| 396 | if (!(slave->flags&IFF_MULTICAST)) |
| 397 | flags &= ~IFF_MULTICAST; |
| 398 | } while ((q = NEXT_SLAVE(q)) != m->slaves); |
| 399 | |
| 400 | m->dev->mtu = mtu; |
| 401 | m->dev->flags = (m->dev->flags&~FMASK) | flags; |
| 402 | netif_start_queue(m->dev); |
| 403 | return 0; |
| 404 | } |
| 405 | |
| 406 | static int teql_master_close(struct net_device *dev) |
| 407 | { |
| 408 | netif_stop_queue(dev); |
| 409 | return 0; |
| 410 | } |
| 411 | |
| 412 | static struct rtnl_link_stats64 *teql_master_stats64(struct net_device *dev, |
| 413 | struct rtnl_link_stats64 *stats) |
| 414 | { |
| 415 | struct teql_master *m = netdev_priv(dev); |
| 416 | |
| 417 | stats->tx_packets = m->tx_packets; |
| 418 | stats->tx_bytes = m->tx_bytes; |
| 419 | stats->tx_errors = m->tx_errors; |
| 420 | stats->tx_dropped = m->tx_dropped; |
| 421 | return stats; |
| 422 | } |
| 423 | |
| 424 | static int teql_master_mtu(struct net_device *dev, int new_mtu) |
| 425 | { |
| 426 | struct teql_master *m = netdev_priv(dev); |
| 427 | struct Qdisc *q; |
| 428 | |
| 429 | if (new_mtu < 68) |
| 430 | return -EINVAL; |
| 431 | |
| 432 | q = m->slaves; |
| 433 | if (q) { |
| 434 | do { |
| 435 | if (new_mtu > qdisc_dev(q)->mtu) |
| 436 | return -EINVAL; |
| 437 | } while ((q = NEXT_SLAVE(q)) != m->slaves); |
| 438 | } |
| 439 | |
| 440 | dev->mtu = new_mtu; |
| 441 | return 0; |
| 442 | } |
| 443 | |
| 444 | static const struct net_device_ops teql_netdev_ops = { |
| 445 | .ndo_open = teql_master_open, |
| 446 | .ndo_stop = teql_master_close, |
| 447 | .ndo_start_xmit = teql_master_xmit, |
| 448 | .ndo_get_stats64 = teql_master_stats64, |
| 449 | .ndo_change_mtu = teql_master_mtu, |
| 450 | }; |
| 451 | |
| 452 | static __init void teql_master_setup(struct net_device *dev) |
| 453 | { |
| 454 | struct teql_master *master = netdev_priv(dev); |
| 455 | struct Qdisc_ops *ops = &master->qops; |
| 456 | |
| 457 | master->dev = dev; |
| 458 | ops->priv_size = sizeof(struct teql_sched_data); |
| 459 | |
| 460 | ops->enqueue = teql_enqueue; |
| 461 | ops->dequeue = teql_dequeue; |
| 462 | ops->peek = teql_peek; |
| 463 | ops->init = teql_qdisc_init; |
| 464 | ops->reset = teql_reset; |
| 465 | ops->destroy = teql_destroy; |
| 466 | ops->owner = THIS_MODULE; |
| 467 | |
| 468 | dev->netdev_ops = &teql_netdev_ops; |
| 469 | dev->type = ARPHRD_VOID; |
| 470 | dev->mtu = 1500; |
| 471 | dev->tx_queue_len = 100; |
| 472 | dev->flags = IFF_NOARP; |
| 473 | dev->hard_header_len = LL_MAX_HEADER; |
| 474 | dev->priv_flags &= ~IFF_XMIT_DST_RELEASE; |
| 475 | } |
| 476 | |
| 477 | static LIST_HEAD(master_dev_list); |
| 478 | static int max_equalizers = 1; |
| 479 | module_param(max_equalizers, int, 0); |
| 480 | MODULE_PARM_DESC(max_equalizers, "Max number of link equalizers"); |
| 481 | |
| 482 | static int __init teql_init(void) |
| 483 | { |
| 484 | int i; |
| 485 | int err = -ENODEV; |
| 486 | |
| 487 | for (i = 0; i < max_equalizers; i++) { |
| 488 | struct net_device *dev; |
| 489 | struct teql_master *master; |
| 490 | |
| 491 | dev = alloc_netdev(sizeof(struct teql_master), |
| 492 | "teql%d", teql_master_setup); |
| 493 | if (!dev) { |
| 494 | err = -ENOMEM; |
| 495 | break; |
| 496 | } |
| 497 | |
| 498 | if ((err = register_netdev(dev))) { |
| 499 | free_netdev(dev); |
| 500 | break; |
| 501 | } |
| 502 | |
| 503 | master = netdev_priv(dev); |
| 504 | |
| 505 | strlcpy(master->qops.id, dev->name, IFNAMSIZ); |
| 506 | err = register_qdisc(&master->qops); |
| 507 | |
| 508 | if (err) { |
| 509 | unregister_netdev(dev); |
| 510 | free_netdev(dev); |
| 511 | break; |
| 512 | } |
| 513 | |
| 514 | list_add_tail(&master->master_list, &master_dev_list); |
| 515 | } |
| 516 | return i ? 0 : err; |
| 517 | } |
| 518 | |
| 519 | static void __exit teql_exit(void) |
| 520 | { |
| 521 | struct teql_master *master, *nxt; |
| 522 | |
| 523 | list_for_each_entry_safe(master, nxt, &master_dev_list, master_list) { |
| 524 | |
| 525 | list_del(&master->master_list); |
| 526 | |
| 527 | unregister_qdisc(&master->qops); |
| 528 | unregister_netdev(master->dev); |
| 529 | free_netdev(master->dev); |
| 530 | } |
| 531 | } |
| 532 | |
| 533 | module_init(teql_init); |
| 534 | module_exit(teql_exit); |
| 535 | |
| 536 | MODULE_LICENSE("GPL"); |