1 /* 2 * net/sched/sch_netem.c Network emulator 3 * 4 * This program is free software; you can redistribute it and/or 5 * modify it under the terms of the GNU General Public License 6 * as published by the Free Software Foundation; either version 7 * 2 of the License. 8 * 9 * Many of the algorithms and ideas for this came from 10 * NIST Net which is not copyrighted. 11 * 12 * Authors: Stephen Hemminger <shemminger@osdl.org> 13 * Catalin(ux aka Dino) BOIE <catab at umbrella dot ro> 14 */ 15 16 #include <linux/mm.h> 17 #include <linux/module.h> 18 #include <linux/slab.h> 19 #include <linux/types.h> 20 #include <linux/kernel.h> 21 #include <linux/errno.h> 22 #include <linux/skbuff.h> 23 #include <linux/vmalloc.h> 24 #include <linux/rtnetlink.h> 25 #include <linux/reciprocal_div.h> 26 #include <linux/rbtree.h> 27 28 #include <net/netlink.h> 29 #include <net/pkt_sched.h> 30 #include <net/inet_ecn.h> 31 32 #define VERSION "1.3" 33 34 /* Network Emulation Queuing algorithm. 35 ==================================== 36 37 Sources: [1] Mark Carson, Darrin Santay, "NIST Net - A Linux-based 38 Network Emulation Tool 39 [2] Luigi Rizzo, DummyNet for FreeBSD 40 41 ---------------------------------------------------------------- 42 43 This started out as a simple way to delay outgoing packets to 44 test TCP but has grown to include most of the functionality 45 of a full blown network emulator like NISTnet. It can delay 46 packets and add random jitter (and correlation). The random 47 distribution can be loaded from a table as well to provide 48 normal, Pareto, or experimental curves. Packet loss, 49 duplication, and reordering can also be emulated. 50 51 This qdisc does not do classification that can be handled in 52 layering other disciplines. It does not need to do bandwidth 53 control either since that can be handled by using token 54 bucket or other rate control. 55 56 Correlated Loss Generator models 57 58 Added generation of correlated loss according to the 59 "Gilbert-Elliot" model, a 4-state markov model. 60 61 References: 62 [1] NetemCLG Home http://netgroup.uniroma2.it/NetemCLG 63 [2] S. Salsano, F. Ludovici, A. Ordine, "Definition of a general 64 and intuitive loss model for packet networks and its implementation 65 in the Netem module in the Linux kernel", available in [1] 66 67 Authors: Stefano Salsano <stefano.salsano at uniroma2.it 68 Fabio Ludovici <fabio.ludovici at yahoo.it> 69 */ 70 71 struct netem_sched_data { 72 /* internal t(ime)fifo qdisc uses t_root and sch->limit */ 73 struct rb_root t_root; 74 75 /* optional qdisc for classful handling (NULL at netem init) */ 76 struct Qdisc *qdisc; 77 78 struct qdisc_watchdog watchdog; 79 80 psched_tdiff_t latency; 81 psched_tdiff_t jitter; 82 83 u32 loss; 84 u32 ecn; 85 u32 limit; 86 u32 counter; 87 u32 gap; 88 u32 duplicate; 89 u32 reorder; 90 u32 corrupt; 91 u64 rate; 92 s32 packet_overhead; 93 u32 cell_size; 94 u32 cell_size_reciprocal; 95 s32 cell_overhead; 96 97 struct crndstate { 98 u32 last; 99 u32 rho; 100 } delay_cor, loss_cor, dup_cor, reorder_cor, corrupt_cor; 101 102 struct disttable { 103 u32 size; 104 s16 table[0]; 105 } *delay_dist; 106 107 enum { 108 CLG_RANDOM, 109 CLG_4_STATES, 110 CLG_GILB_ELL, 111 } loss_model; 112 113 /* Correlated Loss Generation models */ 114 struct clgstate { 115 /* state of the Markov chain */ 116 u8 state; 117 118 /* 4-states and Gilbert-Elliot models */ 119 u32 a1; /* p13 for 4-states or p for GE */ 120 u32 a2; /* p31 for 4-states or r for GE */ 121 u32 a3; /* p32 for 4-states or h for GE */ 122 u32 a4; /* p14 for 4-states or 1-k for GE */ 123 u32 a5; /* p23 used only in 4-states */ 124 } clg; 125 126 }; 127 128 /* Time stamp put into socket buffer control block 129 * Only valid when skbs are in our internal t(ime)fifo queue. 130 */ 131 struct netem_skb_cb { 132 psched_time_t time_to_send; 133 ktime_t tstamp_save; 134 }; 135 136 /* Because space in skb->cb[] is tight, netem overloads skb->next/prev/tstamp 137 * to hold a rb_node structure. 138 * 139 * If struct sk_buff layout is changed, the following checks will complain. 140 */ 141 static struct rb_node *netem_rb_node(struct sk_buff *skb) 142 { 143 BUILD_BUG_ON(offsetof(struct sk_buff, next) != 0); 144 BUILD_BUG_ON(offsetof(struct sk_buff, prev) != 145 offsetof(struct sk_buff, next) + sizeof(skb->next)); 146 BUILD_BUG_ON(offsetof(struct sk_buff, tstamp) != 147 offsetof(struct sk_buff, prev) + sizeof(skb->prev)); 148 BUILD_BUG_ON(sizeof(struct rb_node) > sizeof(skb->next) + 149 sizeof(skb->prev) + 150 sizeof(skb->tstamp)); 151 return (struct rb_node *)&skb->next; 152 } 153 154 static struct sk_buff *netem_rb_to_skb(struct rb_node *rb) 155 { 156 return (struct sk_buff *)rb; 157 } 158 159 static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb) 160 { 161 /* we assume we can use skb next/prev/tstamp as storage for rb_node */ 162 qdisc_cb_private_validate(skb, sizeof(struct netem_skb_cb)); 163 return (struct netem_skb_cb *)qdisc_skb_cb(skb)->data; 164 } 165 166 /* init_crandom - initialize correlated random number generator 167 * Use entropy source for initial seed. 168 */ 169 static void init_crandom(struct crndstate *state, unsigned long rho) 170 { 171 state->rho = rho; 172 state->last = net_random(); 173 } 174 175 /* get_crandom - correlated random number generator 176 * Next number depends on last value. 177 * rho is scaled to avoid floating point. 178 */ 179 static u32 get_crandom(struct crndstate *state) 180 { 181 u64 value, rho; 182 unsigned long answer; 183 184 if (state->rho == 0) /* no correlation */ 185 return net_random(); 186 187 value = net_random(); 188 rho = (u64)state->rho + 1; 189 answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32; 190 state->last = answer; 191 return answer; 192 } 193 194 /* loss_4state - 4-state model loss generator 195 * Generates losses according to the 4-state Markov chain adopted in 196 * the GI (General and Intuitive) loss model. 197 */ 198 static bool loss_4state(struct netem_sched_data *q) 199 { 200 struct clgstate *clg = &q->clg; 201 u32 rnd = net_random(); 202 203 /* 204 * Makes a comparison between rnd and the transition 205 * probabilities outgoing from the current state, then decides the 206 * next state and if the next packet has to be transmitted or lost. 207 * The four states correspond to: 208 * 1 => successfully transmitted packets within a gap period 209 * 4 => isolated losses within a gap period 210 * 3 => lost packets within a burst period 211 * 2 => successfully transmitted packets within a burst period 212 */ 213 switch (clg->state) { 214 case 1: 215 if (rnd < clg->a4) { 216 clg->state = 4; 217 return true; 218 } else if (clg->a4 < rnd && rnd < clg->a1 + clg->a4) { 219 clg->state = 3; 220 return true; 221 } else if (clg->a1 + clg->a4 < rnd) 222 clg->state = 1; 223 224 break; 225 case 2: 226 if (rnd < clg->a5) { 227 clg->state = 3; 228 return true; 229 } else 230 clg->state = 2; 231 232 break; 233 case 3: 234 if (rnd < clg->a3) 235 clg->state = 2; 236 else if (clg->a3 < rnd && rnd < clg->a2 + clg->a3) { 237 clg->state = 1; 238 } else if (clg->a2 + clg->a3 < rnd) { 239 clg->state = 3; 240 return true; 241 } 242 break; 243 case 4: 244 clg->state = 1; 245 break; 246 } 247 248 return false; 249 } 250 251 /* loss_gilb_ell - Gilbert-Elliot model loss generator 252 * Generates losses according to the Gilbert-Elliot loss model or 253 * its special cases (Gilbert or Simple Gilbert) 254 * 255 * Makes a comparison between random number and the transition 256 * probabilities outgoing from the current state, then decides the 257 * next state. A second random number is extracted and the comparison 258 * with the loss probability of the current state decides if the next 259 * packet will be transmitted or lost. 260 */ 261 static bool loss_gilb_ell(struct netem_sched_data *q) 262 { 263 struct clgstate *clg = &q->clg; 264 265 switch (clg->state) { 266 case 1: 267 if (net_random() < clg->a1) 268 clg->state = 2; 269 if (net_random() < clg->a4) 270 return true; 271 break; 272 case 2: 273 if (net_random() < clg->a2) 274 clg->state = 1; 275 if (net_random() > clg->a3) 276 return true; 277 } 278 279 return false; 280 } 281 282 static bool loss_event(struct netem_sched_data *q) 283 { 284 switch (q->loss_model) { 285 case CLG_RANDOM: 286 /* Random packet drop 0 => none, ~0 => all */ 287 return q->loss && q->loss >= get_crandom(&q->loss_cor); 288 289 case CLG_4_STATES: 290 /* 4state loss model algorithm (used also for GI model) 291 * Extracts a value from the markov 4 state loss generator, 292 * if it is 1 drops a packet and if needed writes the event in 293 * the kernel logs 294 */ 295 return loss_4state(q); 296 297 case CLG_GILB_ELL: 298 /* Gilbert-Elliot loss model algorithm 299 * Extracts a value from the Gilbert-Elliot loss generator, 300 * if it is 1 drops a packet and if needed writes the event in 301 * the kernel logs 302 */ 303 return loss_gilb_ell(q); 304 } 305 306 return false; /* not reached */ 307 } 308 309 310 /* tabledist - return a pseudo-randomly distributed value with mean mu and 311 * std deviation sigma. Uses table lookup to approximate the desired 312 * distribution, and a uniformly-distributed pseudo-random source. 313 */ 314 static psched_tdiff_t tabledist(psched_tdiff_t mu, psched_tdiff_t sigma, 315 struct crndstate *state, 316 const struct disttable *dist) 317 { 318 psched_tdiff_t x; 319 long t; 320 u32 rnd; 321 322 if (sigma == 0) 323 return mu; 324 325 rnd = get_crandom(state); 326 327 /* default uniform distribution */ 328 if (dist == NULL) 329 return (rnd % (2*sigma)) - sigma + mu; 330 331 t = dist->table[rnd % dist->size]; 332 x = (sigma % NETEM_DIST_SCALE) * t; 333 if (x >= 0) 334 x += NETEM_DIST_SCALE/2; 335 else 336 x -= NETEM_DIST_SCALE/2; 337 338 return x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu; 339 } 340 341 static psched_time_t packet_len_2_sched_time(unsigned int len, struct netem_sched_data *q) 342 { 343 u64 ticks; 344 345 len += q->packet_overhead; 346 347 if (q->cell_size) { 348 u32 cells = reciprocal_divide(len, q->cell_size_reciprocal); 349 350 if (len > cells * q->cell_size) /* extra cell needed for remainder */ 351 cells++; 352 len = cells * (q->cell_size + q->cell_overhead); 353 } 354 355 ticks = (u64)len * NSEC_PER_SEC; 356 357 do_div(ticks, q->rate); 358 return PSCHED_NS2TICKS(ticks); 359 } 360 361 static void tfifo_reset(struct Qdisc *sch) 362 { 363 struct netem_sched_data *q = qdisc_priv(sch); 364 struct rb_node *p; 365 366 while ((p = rb_first(&q->t_root))) { 367 struct sk_buff *skb = netem_rb_to_skb(p); 368 369 rb_erase(p, &q->t_root); 370 skb->next = NULL; 371 skb->prev = NULL; 372 kfree_skb(skb); 373 } 374 } 375 376 static void tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch) 377 { 378 struct netem_sched_data *q = qdisc_priv(sch); 379 psched_time_t tnext = netem_skb_cb(nskb)->time_to_send; 380 struct rb_node **p = &q->t_root.rb_node, *parent = NULL; 381 382 while (*p) { 383 struct sk_buff *skb; 384 385 parent = *p; 386 skb = netem_rb_to_skb(parent); 387 if (tnext >= netem_skb_cb(skb)->time_to_send) 388 p = &parent->rb_right; 389 else 390 p = &parent->rb_left; 391 } 392 rb_link_node(netem_rb_node(nskb), parent, p); 393 rb_insert_color(netem_rb_node(nskb), &q->t_root); 394 sch->q.qlen++; 395 } 396 397 /* 398 * Insert one skb into qdisc. 399 * Note: parent depends on return value to account for queue length. 400 * NET_XMIT_DROP: queue length didn't change. 401 * NET_XMIT_SUCCESS: one skb was queued. 402 */ 403 static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch) 404 { 405 struct netem_sched_data *q = qdisc_priv(sch); 406 /* We don't fill cb now as skb_unshare() may invalidate it */ 407 struct netem_skb_cb *cb; 408 struct sk_buff *skb2; 409 int count = 1; 410 411 /* Random duplication */ 412 if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor)) 413 ++count; 414 415 /* Drop packet? */ 416 if (loss_event(q)) { 417 if (q->ecn && INET_ECN_set_ce(skb)) 418 sch->qstats.drops++; /* mark packet */ 419 else 420 --count; 421 } 422 if (count == 0) { 423 sch->qstats.drops++; 424 kfree_skb(skb); 425 return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS; 426 } 427 428 /* If a delay is expected, orphan the skb. (orphaning usually takes 429 * place at TX completion time, so _before_ the link transit delay) 430 */ 431 if (q->latency || q->jitter) 432 skb_orphan_partial(skb); 433 434 /* 435 * If we need to duplicate packet, then re-insert at top of the 436 * qdisc tree, since parent queuer expects that only one 437 * skb will be queued. 438 */ 439 if (count > 1 && (skb2 = skb_clone(skb, GFP_ATOMIC)) != NULL) { 440 struct Qdisc *rootq = qdisc_root(sch); 441 u32 dupsave = q->duplicate; /* prevent duplicating a dup... */ 442 q->duplicate = 0; 443 444 qdisc_enqueue_root(skb2, rootq); 445 q->duplicate = dupsave; 446 } 447 448 /* 449 * Randomized packet corruption. 450 * Make copy if needed since we are modifying 451 * If packet is going to be hardware checksummed, then 452 * do it now in software before we mangle it. 453 */ 454 if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor)) { 455 if (!(skb = skb_unshare(skb, GFP_ATOMIC)) || 456 (skb->ip_summed == CHECKSUM_PARTIAL && 457 skb_checksum_help(skb))) 458 return qdisc_drop(skb, sch); 459 460 skb->data[net_random() % skb_headlen(skb)] ^= 1<<(net_random() % 8); 461 } 462 463 if (unlikely(skb_queue_len(&sch->q) >= sch->limit)) 464 return qdisc_reshape_fail(skb, sch); 465 466 sch->qstats.backlog += qdisc_pkt_len(skb); 467 468 cb = netem_skb_cb(skb); 469 if (q->gap == 0 || /* not doing reordering */ 470 q->counter < q->gap - 1 || /* inside last reordering gap */ 471 q->reorder < get_crandom(&q->reorder_cor)) { 472 psched_time_t now; 473 psched_tdiff_t delay; 474 475 delay = tabledist(q->latency, q->jitter, 476 &q->delay_cor, q->delay_dist); 477 478 now = psched_get_time(); 479 480 if (q->rate) { 481 struct sk_buff *last; 482 483 if (!skb_queue_empty(&sch->q)) 484 last = skb_peek_tail(&sch->q); 485 else 486 last = netem_rb_to_skb(rb_last(&q->t_root)); 487 if (last) { 488 /* 489 * Last packet in queue is reference point (now), 490 * calculate this time bonus and subtract 491 * from delay. 492 */ 493 delay -= netem_skb_cb(last)->time_to_send - now; 494 delay = max_t(psched_tdiff_t, 0, delay); 495 now = netem_skb_cb(last)->time_to_send; 496 } 497 498 delay += packet_len_2_sched_time(qdisc_pkt_len(skb), q); 499 } 500 501 cb->time_to_send = now + delay; 502 cb->tstamp_save = skb->tstamp; 503 ++q->counter; 504 tfifo_enqueue(skb, sch); 505 } else { 506 /* 507 * Do re-ordering by putting one out of N packets at the front 508 * of the queue. 509 */ 510 cb->time_to_send = psched_get_time(); 511 q->counter = 0; 512 513 __skb_queue_head(&sch->q, skb); 514 sch->qstats.requeues++; 515 } 516 517 return NET_XMIT_SUCCESS; 518 } 519 520 static unsigned int netem_drop(struct Qdisc *sch) 521 { 522 struct netem_sched_data *q = qdisc_priv(sch); 523 unsigned int len; 524 525 len = qdisc_queue_drop(sch); 526 527 if (!len) { 528 struct rb_node *p = rb_first(&q->t_root); 529 530 if (p) { 531 struct sk_buff *skb = netem_rb_to_skb(p); 532 533 rb_erase(p, &q->t_root); 534 sch->q.qlen--; 535 skb->next = NULL; 536 skb->prev = NULL; 537 len = qdisc_pkt_len(skb); 538 sch->qstats.backlog -= len; 539 kfree_skb(skb); 540 } 541 } 542 if (!len && q->qdisc && q->qdisc->ops->drop) 543 len = q->qdisc->ops->drop(q->qdisc); 544 if (len) 545 sch->qstats.drops++; 546 547 return len; 548 } 549 550 static struct sk_buff *netem_dequeue(struct Qdisc *sch) 551 { 552 struct netem_sched_data *q = qdisc_priv(sch); 553 struct sk_buff *skb; 554 struct rb_node *p; 555 556 if (qdisc_is_throttled(sch)) 557 return NULL; 558 559 tfifo_dequeue: 560 skb = __skb_dequeue(&sch->q); 561 if (skb) { 562 deliver: 563 sch->qstats.backlog -= qdisc_pkt_len(skb); 564 qdisc_unthrottled(sch); 565 qdisc_bstats_update(sch, skb); 566 return skb; 567 } 568 p = rb_first(&q->t_root); 569 if (p) { 570 psched_time_t time_to_send; 571 572 skb = netem_rb_to_skb(p); 573 574 /* if more time remaining? */ 575 time_to_send = netem_skb_cb(skb)->time_to_send; 576 if (time_to_send <= psched_get_time()) { 577 rb_erase(p, &q->t_root); 578 579 sch->q.qlen--; 580 skb->next = NULL; 581 skb->prev = NULL; 582 skb->tstamp = netem_skb_cb(skb)->tstamp_save; 583 584 #ifdef CONFIG_NET_CLS_ACT 585 /* 586 * If it's at ingress let's pretend the delay is 587 * from the network (tstamp will be updated). 588 */ 589 if (G_TC_FROM(skb->tc_verd) & AT_INGRESS) 590 skb->tstamp.tv64 = 0; 591 #endif 592 593 if (q->qdisc) { 594 int err = qdisc_enqueue(skb, q->qdisc); 595 596 if (unlikely(err != NET_XMIT_SUCCESS)) { 597 if (net_xmit_drop_count(err)) { 598 sch->qstats.drops++; 599 qdisc_tree_decrease_qlen(sch, 1); 600 } 601 } 602 goto tfifo_dequeue; 603 } 604 goto deliver; 605 } 606 607 if (q->qdisc) { 608 skb = q->qdisc->ops->dequeue(q->qdisc); 609 if (skb) 610 goto deliver; 611 } 612 qdisc_watchdog_schedule(&q->watchdog, time_to_send); 613 } 614 615 if (q->qdisc) { 616 skb = q->qdisc->ops->dequeue(q->qdisc); 617 if (skb) 618 goto deliver; 619 } 620 return NULL; 621 } 622 623 static void netem_reset(struct Qdisc *sch) 624 { 625 struct netem_sched_data *q = qdisc_priv(sch); 626 627 qdisc_reset_queue(sch); 628 tfifo_reset(sch); 629 if (q->qdisc) 630 qdisc_reset(q->qdisc); 631 qdisc_watchdog_cancel(&q->watchdog); 632 } 633 634 static void dist_free(struct disttable *d) 635 { 636 if (d) { 637 if (is_vmalloc_addr(d)) 638 vfree(d); 639 else 640 kfree(d); 641 } 642 } 643 644 /* 645 * Distribution data is a variable size payload containing 646 * signed 16 bit values. 647 */ 648 static int get_dist_table(struct Qdisc *sch, const struct nlattr *attr) 649 { 650 struct netem_sched_data *q = qdisc_priv(sch); 651 size_t n = nla_len(attr)/sizeof(__s16); 652 const __s16 *data = nla_data(attr); 653 spinlock_t *root_lock; 654 struct disttable *d; 655 int i; 656 size_t s; 657 658 if (n > NETEM_DIST_MAX) 659 return -EINVAL; 660 661 s = sizeof(struct disttable) + n * sizeof(s16); 662 d = kmalloc(s, GFP_KERNEL | __GFP_NOWARN); 663 if (!d) 664 d = vmalloc(s); 665 if (!d) 666 return -ENOMEM; 667 668 d->size = n; 669 for (i = 0; i < n; i++) 670 d->table[i] = data[i]; 671 672 root_lock = qdisc_root_sleeping_lock(sch); 673 674 spin_lock_bh(root_lock); 675 swap(q->delay_dist, d); 676 spin_unlock_bh(root_lock); 677 678 dist_free(d); 679 return 0; 680 } 681 682 static void get_correlation(struct Qdisc *sch, const struct nlattr *attr) 683 { 684 struct netem_sched_data *q = qdisc_priv(sch); 685 const struct tc_netem_corr *c = nla_data(attr); 686 687 init_crandom(&q->delay_cor, c->delay_corr); 688 init_crandom(&q->loss_cor, c->loss_corr); 689 init_crandom(&q->dup_cor, c->dup_corr); 690 } 691 692 static void get_reorder(struct Qdisc *sch, const struct nlattr *attr) 693 { 694 struct netem_sched_data *q = qdisc_priv(sch); 695 const struct tc_netem_reorder *r = nla_data(attr); 696 697 q->reorder = r->probability; 698 init_crandom(&q->reorder_cor, r->correlation); 699 } 700 701 static void get_corrupt(struct Qdisc *sch, const struct nlattr *attr) 702 { 703 struct netem_sched_data *q = qdisc_priv(sch); 704 const struct tc_netem_corrupt *r = nla_data(attr); 705 706 q->corrupt = r->probability; 707 init_crandom(&q->corrupt_cor, r->correlation); 708 } 709 710 static void get_rate(struct Qdisc *sch, const struct nlattr *attr) 711 { 712 struct netem_sched_data *q = qdisc_priv(sch); 713 const struct tc_netem_rate *r = nla_data(attr); 714 715 q->rate = r->rate; 716 q->packet_overhead = r->packet_overhead; 717 q->cell_size = r->cell_size; 718 if (q->cell_size) 719 q->cell_size_reciprocal = reciprocal_value(q->cell_size); 720 q->cell_overhead = r->cell_overhead; 721 } 722 723 static int get_loss_clg(struct Qdisc *sch, const struct nlattr *attr) 724 { 725 struct netem_sched_data *q = qdisc_priv(sch); 726 const struct nlattr *la; 727 int rem; 728 729 nla_for_each_nested(la, attr, rem) { 730 u16 type = nla_type(la); 731 732 switch (type) { 733 case NETEM_LOSS_GI: { 734 const struct tc_netem_gimodel *gi = nla_data(la); 735 736 if (nla_len(la) < sizeof(struct tc_netem_gimodel)) { 737 pr_info("netem: incorrect gi model size\n"); 738 return -EINVAL; 739 } 740 741 q->loss_model = CLG_4_STATES; 742 743 q->clg.state = 1; 744 q->clg.a1 = gi->p13; 745 q->clg.a2 = gi->p31; 746 q->clg.a3 = gi->p32; 747 q->clg.a4 = gi->p14; 748 q->clg.a5 = gi->p23; 749 break; 750 } 751 752 case NETEM_LOSS_GE: { 753 const struct tc_netem_gemodel *ge = nla_data(la); 754 755 if (nla_len(la) < sizeof(struct tc_netem_gemodel)) { 756 pr_info("netem: incorrect ge model size\n"); 757 return -EINVAL; 758 } 759 760 q->loss_model = CLG_GILB_ELL; 761 q->clg.state = 1; 762 q->clg.a1 = ge->p; 763 q->clg.a2 = ge->r; 764 q->clg.a3 = ge->h; 765 q->clg.a4 = ge->k1; 766 break; 767 } 768 769 default: 770 pr_info("netem: unknown loss type %u\n", type); 771 return -EINVAL; 772 } 773 } 774 775 return 0; 776 } 777 778 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = { 779 [TCA_NETEM_CORR] = { .len = sizeof(struct tc_netem_corr) }, 780 [TCA_NETEM_REORDER] = { .len = sizeof(struct tc_netem_reorder) }, 781 [TCA_NETEM_CORRUPT] = { .len = sizeof(struct tc_netem_corrupt) }, 782 [TCA_NETEM_RATE] = { .len = sizeof(struct tc_netem_rate) }, 783 [TCA_NETEM_LOSS] = { .type = NLA_NESTED }, 784 [TCA_NETEM_ECN] = { .type = NLA_U32 }, 785 [TCA_NETEM_RATE64] = { .type = NLA_U64 }, 786 }; 787 788 static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla, 789 const struct nla_policy *policy, int len) 790 { 791 int nested_len = nla_len(nla) - NLA_ALIGN(len); 792 793 if (nested_len < 0) { 794 pr_info("netem: invalid attributes len %d\n", nested_len); 795 return -EINVAL; 796 } 797 798 if (nested_len >= nla_attr_size(0)) 799 return nla_parse(tb, maxtype, nla_data(nla) + NLA_ALIGN(len), 800 nested_len, policy); 801 802 memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1)); 803 return 0; 804 } 805 806 /* Parse netlink message to set options */ 807 static int netem_change(struct Qdisc *sch, struct nlattr *opt) 808 { 809 struct netem_sched_data *q = qdisc_priv(sch); 810 struct nlattr *tb[TCA_NETEM_MAX + 1]; 811 struct tc_netem_qopt *qopt; 812 int ret; 813 814 if (opt == NULL) 815 return -EINVAL; 816 817 qopt = nla_data(opt); 818 ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt)); 819 if (ret < 0) 820 return ret; 821 822 sch->limit = qopt->limit; 823 824 q->latency = qopt->latency; 825 q->jitter = qopt->jitter; 826 q->limit = qopt->limit; 827 q->gap = qopt->gap; 828 q->counter = 0; 829 q->loss = qopt->loss; 830 q->duplicate = qopt->duplicate; 831 832 /* for compatibility with earlier versions. 833 * if gap is set, need to assume 100% probability 834 */ 835 if (q->gap) 836 q->reorder = ~0; 837 838 if (tb[TCA_NETEM_CORR]) 839 get_correlation(sch, tb[TCA_NETEM_CORR]); 840 841 if (tb[TCA_NETEM_DELAY_DIST]) { 842 ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST]); 843 if (ret) 844 return ret; 845 } 846 847 if (tb[TCA_NETEM_REORDER]) 848 get_reorder(sch, tb[TCA_NETEM_REORDER]); 849 850 if (tb[TCA_NETEM_CORRUPT]) 851 get_corrupt(sch, tb[TCA_NETEM_CORRUPT]); 852 853 if (tb[TCA_NETEM_RATE]) 854 get_rate(sch, tb[TCA_NETEM_RATE]); 855 856 if (tb[TCA_NETEM_RATE64]) 857 q->rate = max_t(u64, q->rate, 858 nla_get_u64(tb[TCA_NETEM_RATE64])); 859 860 if (tb[TCA_NETEM_ECN]) 861 q->ecn = nla_get_u32(tb[TCA_NETEM_ECN]); 862 863 q->loss_model = CLG_RANDOM; 864 if (tb[TCA_NETEM_LOSS]) 865 ret = get_loss_clg(sch, tb[TCA_NETEM_LOSS]); 866 867 return ret; 868 } 869 870 static int netem_init(struct Qdisc *sch, struct nlattr *opt) 871 { 872 struct netem_sched_data *q = qdisc_priv(sch); 873 int ret; 874 875 if (!opt) 876 return -EINVAL; 877 878 qdisc_watchdog_init(&q->watchdog, sch); 879 880 q->loss_model = CLG_RANDOM; 881 ret = netem_change(sch, opt); 882 if (ret) 883 pr_info("netem: change failed\n"); 884 return ret; 885 } 886 887 static void netem_destroy(struct Qdisc *sch) 888 { 889 struct netem_sched_data *q = qdisc_priv(sch); 890 891 qdisc_watchdog_cancel(&q->watchdog); 892 if (q->qdisc) 893 qdisc_destroy(q->qdisc); 894 dist_free(q->delay_dist); 895 } 896 897 static int dump_loss_model(const struct netem_sched_data *q, 898 struct sk_buff *skb) 899 { 900 struct nlattr *nest; 901 902 nest = nla_nest_start(skb, TCA_NETEM_LOSS); 903 if (nest == NULL) 904 goto nla_put_failure; 905 906 switch (q->loss_model) { 907 case CLG_RANDOM: 908 /* legacy loss model */ 909 nla_nest_cancel(skb, nest); 910 return 0; /* no data */ 911 912 case CLG_4_STATES: { 913 struct tc_netem_gimodel gi = { 914 .p13 = q->clg.a1, 915 .p31 = q->clg.a2, 916 .p32 = q->clg.a3, 917 .p14 = q->clg.a4, 918 .p23 = q->clg.a5, 919 }; 920 921 if (nla_put(skb, NETEM_LOSS_GI, sizeof(gi), &gi)) 922 goto nla_put_failure; 923 break; 924 } 925 case CLG_GILB_ELL: { 926 struct tc_netem_gemodel ge = { 927 .p = q->clg.a1, 928 .r = q->clg.a2, 929 .h = q->clg.a3, 930 .k1 = q->clg.a4, 931 }; 932 933 if (nla_put(skb, NETEM_LOSS_GE, sizeof(ge), &ge)) 934 goto nla_put_failure; 935 break; 936 } 937 } 938 939 nla_nest_end(skb, nest); 940 return 0; 941 942 nla_put_failure: 943 nla_nest_cancel(skb, nest); 944 return -1; 945 } 946 947 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb) 948 { 949 const struct netem_sched_data *q = qdisc_priv(sch); 950 struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb); 951 struct tc_netem_qopt qopt; 952 struct tc_netem_corr cor; 953 struct tc_netem_reorder reorder; 954 struct tc_netem_corrupt corrupt; 955 struct tc_netem_rate rate; 956 957 qopt.latency = q->latency; 958 qopt.jitter = q->jitter; 959 qopt.limit = q->limit; 960 qopt.loss = q->loss; 961 qopt.gap = q->gap; 962 qopt.duplicate = q->duplicate; 963 if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt)) 964 goto nla_put_failure; 965 966 cor.delay_corr = q->delay_cor.rho; 967 cor.loss_corr = q->loss_cor.rho; 968 cor.dup_corr = q->dup_cor.rho; 969 if (nla_put(skb, TCA_NETEM_CORR, sizeof(cor), &cor)) 970 goto nla_put_failure; 971 972 reorder.probability = q->reorder; 973 reorder.correlation = q->reorder_cor.rho; 974 if (nla_put(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder)) 975 goto nla_put_failure; 976 977 corrupt.probability = q->corrupt; 978 corrupt.correlation = q->corrupt_cor.rho; 979 if (nla_put(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt)) 980 goto nla_put_failure; 981 982 if (q->rate >= (1ULL << 32)) { 983 if (nla_put_u64(skb, TCA_NETEM_RATE64, q->rate)) 984 goto nla_put_failure; 985 rate.rate = ~0U; 986 } else { 987 rate.rate = q->rate; 988 } 989 rate.packet_overhead = q->packet_overhead; 990 rate.cell_size = q->cell_size; 991 rate.cell_overhead = q->cell_overhead; 992 if (nla_put(skb, TCA_NETEM_RATE, sizeof(rate), &rate)) 993 goto nla_put_failure; 994 995 if (q->ecn && nla_put_u32(skb, TCA_NETEM_ECN, q->ecn)) 996 goto nla_put_failure; 997 998 if (dump_loss_model(q, skb) != 0) 999 goto nla_put_failure; 1000 1001 return nla_nest_end(skb, nla); 1002 1003 nla_put_failure: 1004 nlmsg_trim(skb, nla); 1005 return -1; 1006 } 1007 1008 static int netem_dump_class(struct Qdisc *sch, unsigned long cl, 1009 struct sk_buff *skb, struct tcmsg *tcm) 1010 { 1011 struct netem_sched_data *q = qdisc_priv(sch); 1012 1013 if (cl != 1 || !q->qdisc) /* only one class */ 1014 return -ENOENT; 1015 1016 tcm->tcm_handle |= TC_H_MIN(1); 1017 tcm->tcm_info = q->qdisc->handle; 1018 1019 return 0; 1020 } 1021 1022 static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new, 1023 struct Qdisc **old) 1024 { 1025 struct netem_sched_data *q = qdisc_priv(sch); 1026 1027 sch_tree_lock(sch); 1028 *old = q->qdisc; 1029 q->qdisc = new; 1030 if (*old) { 1031 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen); 1032 qdisc_reset(*old); 1033 } 1034 sch_tree_unlock(sch); 1035 1036 return 0; 1037 } 1038 1039 static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg) 1040 { 1041 struct netem_sched_data *q = qdisc_priv(sch); 1042 return q->qdisc; 1043 } 1044 1045 static unsigned long netem_get(struct Qdisc *sch, u32 classid) 1046 { 1047 return 1; 1048 } 1049 1050 static void netem_put(struct Qdisc *sch, unsigned long arg) 1051 { 1052 } 1053 1054 static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker) 1055 { 1056 if (!walker->stop) { 1057 if (walker->count >= walker->skip) 1058 if (walker->fn(sch, 1, walker) < 0) { 1059 walker->stop = 1; 1060 return; 1061 } 1062 walker->count++; 1063 } 1064 } 1065 1066 static const struct Qdisc_class_ops netem_class_ops = { 1067 .graft = netem_graft, 1068 .leaf = netem_leaf, 1069 .get = netem_get, 1070 .put = netem_put, 1071 .walk = netem_walk, 1072 .dump = netem_dump_class, 1073 }; 1074 1075 static struct Qdisc_ops netem_qdisc_ops __read_mostly = { 1076 .id = "netem", 1077 .cl_ops = &netem_class_ops, 1078 .priv_size = sizeof(struct netem_sched_data), 1079 .enqueue = netem_enqueue, 1080 .dequeue = netem_dequeue, 1081 .peek = qdisc_peek_dequeued, 1082 .drop = netem_drop, 1083 .init = netem_init, 1084 .reset = netem_reset, 1085 .destroy = netem_destroy, 1086 .change = netem_change, 1087 .dump = netem_dump, 1088 .owner = THIS_MODULE, 1089 }; 1090 1091 1092 static int __init netem_module_init(void) 1093 { 1094 pr_info("netem: version " VERSION "\n"); 1095 return register_qdisc(&netem_qdisc_ops); 1096 } 1097 static void __exit netem_module_exit(void) 1098 { 1099 unregister_qdisc(&netem_qdisc_ops); 1100 } 1101 module_init(netem_module_init) 1102 module_exit(netem_module_exit) 1103 MODULE_LICENSE("GPL"); 1104