1// SPDX-License-Identifier: GPL-2.0-only 2/* 3 * net/sched/sch_netem.c Network emulator 4 * 5 * Many of the algorithms and ideas for this came from 6 * NIST Net which is not copyrighted. 7 * 8 * Authors: Stephen Hemminger <shemminger@osdl.org> 9 * Catalin(ux aka Dino) BOIE <catab at umbrella dot ro> 10 */ 11 12#include <linux/mm.h> 13#include <linux/module.h> 14#include <linux/slab.h> 15#include <linux/types.h> 16#include <linux/kernel.h> 17#include <linux/errno.h> 18#include <linux/skbuff.h> 19#include <linux/vmalloc.h> 20#include <linux/rtnetlink.h> 21#include <linux/reciprocal_div.h> 22#include <linux/rbtree.h> 23 24#include <net/gso.h> 25#include <net/netlink.h> 26#include <net/pkt_sched.h> 27#include <net/inet_ecn.h> 28 29#define VERSION "1.3" 30 31/* Network Emulation Queuing algorithm. 32 ==================================== 33 34 Sources: [1] Mark Carson, Darrin Santay, "NIST Net - A Linux-based 35 Network Emulation Tool 36 [2] Luigi Rizzo, DummyNet for FreeBSD 37 38 ---------------------------------------------------------------- 39 40 This started out as a simple way to delay outgoing packets to 41 test TCP but has grown to include most of the functionality 42 of a full blown network emulator like NISTnet. It can delay 43 packets and add random jitter (and correlation). The random 44 distribution can be loaded from a table as well to provide 45 normal, Pareto, or experimental curves. Packet loss, 46 duplication, and reordering can also be emulated. 47 48 This qdisc does not do classification that can be handled in 49 layering other disciplines. It does not need to do bandwidth 50 control either since that can be handled by using token 51 bucket or other rate control. 52 53 Correlated Loss Generator models 54 55 Added generation of correlated loss according to the 56 "Gilbert-Elliot" model, a 4-state markov model. 57 58 References: 59 [1] NetemCLG Home http://netgroup.uniroma2.it/NetemCLG 60 [2] S. Salsano, F. Ludovici, A. Ordine, "Definition of a general 61 and intuitive loss model for packet networks and its implementation 62 in the Netem module in the Linux kernel", available in [1] 63 64 Authors: Stefano Salsano <stefano.salsano at uniroma2.it 65 Fabio Ludovici <fabio.ludovici at yahoo.it> 66*/ 67 68struct disttable { 69 u32 size; 70 s16 table[] __counted_by(size); 71}; 72 73struct netem_sched_data { 74 /* internal t(ime)fifo qdisc uses t_root and sch->limit */ 75 struct rb_root t_root; 76 77 /* a linear queue; reduces rbtree rebalancing when jitter is low */ 78 struct sk_buff *t_head; 79 struct sk_buff *t_tail; 80 81 /* optional qdisc for classful handling (NULL at netem init) */ 82 struct Qdisc *qdisc; 83 84 struct qdisc_watchdog watchdog; 85 86 s64 latency; 87 s64 jitter; 88 89 u32 loss; 90 u32 ecn; 91 u32 limit; 92 u32 counter; 93 u32 gap; 94 u32 duplicate; 95 u32 reorder; 96 u32 corrupt; 97 u64 rate; 98 s32 packet_overhead; 99 u32 cell_size; 100 struct reciprocal_value cell_size_reciprocal; 101 s32 cell_overhead; 102 103 struct crndstate { 104 u32 last; 105 u32 rho; 106 } delay_cor, loss_cor, dup_cor, reorder_cor, corrupt_cor; 107 108 struct prng { 109 u64 seed; 110 struct rnd_state prng_state; 111 } prng; 112 113 struct disttable *delay_dist; 114 115 enum { 116 CLG_RANDOM, 117 CLG_4_STATES, 118 CLG_GILB_ELL, 119 } loss_model; 120 121 enum { 122 TX_IN_GAP_PERIOD = 1, 123 TX_IN_BURST_PERIOD, 124 LOST_IN_GAP_PERIOD, 125 LOST_IN_BURST_PERIOD, 126 } _4_state_model; 127 128 enum { 129 GOOD_STATE = 1, 130 BAD_STATE, 131 } GE_state_model; 132 133 /* Correlated Loss Generation models */ 134 struct clgstate { 135 /* state of the Markov chain */ 136 u8 state; 137 138 /* 4-states and Gilbert-Elliot models */ 139 u32 a1; /* p13 for 4-states or p for GE */ 140 u32 a2; /* p31 for 4-states or r for GE */ 141 u32 a3; /* p32 for 4-states or h for GE */ 142 u32 a4; /* p14 for 4-states or 1-k for GE */ 143 u32 a5; /* p23 used only in 4-states */ 144 } clg; 145 146 struct tc_netem_slot slot_config; 147 struct slotstate { 148 u64 slot_next; 149 s32 packets_left; 150 s32 bytes_left; 151 } slot; 152 153 struct disttable *slot_dist; 154}; 155 156/* Time stamp put into socket buffer control block 157 * Only valid when skbs are in our internal t(ime)fifo queue. 158 * 159 * As skb->rbnode uses same storage than skb->next, skb->prev and skb->tstamp, 160 * and skb->next & skb->prev are scratch space for a qdisc, 161 * we save skb->tstamp value in skb->cb[] before destroying it. 162 */ 163struct netem_skb_cb { 164 u64 time_to_send; 165}; 166 167static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb) 168{ 169 /* we assume we can use skb next/prev/tstamp as storage for rb_node */ 170 qdisc_cb_private_validate(skb, sizeof(struct netem_skb_cb)); 171 return (struct netem_skb_cb *)qdisc_skb_cb(skb)->data; 172} 173 174/* init_crandom - initialize correlated random number generator 175 * Use entropy source for initial seed. 176 */ 177static void init_crandom(struct crndstate *state, unsigned long rho) 178{ 179 state->rho = rho; 180 state->last = get_random_u32(); 181} 182 183/* get_crandom - correlated random number generator 184 * Next number depends on last value. 185 * rho is scaled to avoid floating point. 186 */ 187static u32 get_crandom(struct crndstate *state, struct prng *p) 188{ 189 u64 value, rho; 190 unsigned long answer; 191 struct rnd_state *s = &p->prng_state; 192 193 if (!state || state->rho == 0) /* no correlation */ 194 return prandom_u32_state(s); 195 196 value = prandom_u32_state(s); 197 rho = (u64)state->rho + 1; 198 answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32; 199 state->last = answer; 200 return answer; 201} 202 203/* loss_4state - 4-state model loss generator 204 * Generates losses according to the 4-state Markov chain adopted in 205 * the GI (General and Intuitive) loss model. 206 */ 207static bool loss_4state(struct netem_sched_data *q) 208{ 209 struct clgstate *clg = &q->clg; 210 u32 rnd = prandom_u32_state(&q->prng.prng_state); 211 212 /* 213 * Makes a comparison between rnd and the transition 214 * probabilities outgoing from the current state, then decides the 215 * next state and if the next packet has to be transmitted or lost. 216 * The four states correspond to: 217 * TX_IN_GAP_PERIOD => successfully transmitted packets within a gap period 218 * LOST_IN_GAP_PERIOD => isolated losses within a gap period 219 * LOST_IN_BURST_PERIOD => lost packets within a burst period 220 * TX_IN_BURST_PERIOD => successfully transmitted packets within a burst period 221 */ 222 switch (clg->state) { 223 case TX_IN_GAP_PERIOD: 224 if (rnd < clg->a4) { 225 clg->state = LOST_IN_GAP_PERIOD; 226 return true; 227 } else if (clg->a4 < rnd && rnd < clg->a1 + clg->a4) { 228 clg->state = LOST_IN_BURST_PERIOD; 229 return true; 230 } else if (clg->a1 + clg->a4 < rnd) { 231 clg->state = TX_IN_GAP_PERIOD; 232 } 233 234 break; 235 case TX_IN_BURST_PERIOD: 236 if (rnd < clg->a5) { 237 clg->state = LOST_IN_BURST_PERIOD; 238 return true; 239 } else { 240 clg->state = TX_IN_BURST_PERIOD; 241 } 242 243 break; 244 case LOST_IN_BURST_PERIOD: 245 if (rnd < clg->a3) 246 clg->state = TX_IN_BURST_PERIOD; 247 else if (clg->a3 < rnd && rnd < clg->a2 + clg->a3) { 248 clg->state = TX_IN_GAP_PERIOD; 249 } else if (clg->a2 + clg->a3 < rnd) { 250 clg->state = LOST_IN_BURST_PERIOD; 251 return true; 252 } 253 break; 254 case LOST_IN_GAP_PERIOD: 255 clg->state = TX_IN_GAP_PERIOD; 256 break; 257 } 258 259 return false; 260} 261 262/* loss_gilb_ell - Gilbert-Elliot model loss generator 263 * Generates losses according to the Gilbert-Elliot loss model or 264 * its special cases (Gilbert or Simple Gilbert) 265 * 266 * Makes a comparison between random number and the transition 267 * probabilities outgoing from the current state, then decides the 268 * next state. A second random number is extracted and the comparison 269 * with the loss probability of the current state decides if the next 270 * packet will be transmitted or lost. 271 */ 272static bool loss_gilb_ell(struct netem_sched_data *q) 273{ 274 struct clgstate *clg = &q->clg; 275 struct rnd_state *s = &q->prng.prng_state; 276 277 switch (clg->state) { 278 case GOOD_STATE: 279 if (prandom_u32_state(s) < clg->a1) 280 clg->state = BAD_STATE; 281 if (prandom_u32_state(s) < clg->a4) 282 return true; 283 break; 284 case BAD_STATE: 285 if (prandom_u32_state(s) < clg->a2) 286 clg->state = GOOD_STATE; 287 if (prandom_u32_state(s) > clg->a3) 288 return true; 289 } 290 291 return false; 292} 293 294static bool loss_event(struct netem_sched_data *q) 295{ 296 switch (q->loss_model) { 297 case CLG_RANDOM: 298 /* Random packet drop 0 => none, ~0 => all */ 299 return q->loss && q->loss >= get_crandom(&q->loss_cor, &q->prng); 300 301 case CLG_4_STATES: 302 /* 4state loss model algorithm (used also for GI model) 303 * Extracts a value from the markov 4 state loss generator, 304 * if it is 1 drops a packet and if needed writes the event in 305 * the kernel logs 306 */ 307 return loss_4state(q); 308 309 case CLG_GILB_ELL: 310 /* Gilbert-Elliot loss model algorithm 311 * Extracts a value from the Gilbert-Elliot loss generator, 312 * if it is 1 drops a packet and if needed writes the event in 313 * the kernel logs 314 */ 315 return loss_gilb_ell(q); 316 } 317 318 return false; /* not reached */ 319} 320 321 322/* tabledist - return a pseudo-randomly distributed value with mean mu and 323 * std deviation sigma. Uses table lookup to approximate the desired 324 * distribution, and a uniformly-distributed pseudo-random source. 325 */ 326static s64 tabledist(s64 mu, s32 sigma, 327 struct crndstate *state, 328 struct prng *prng, 329 const struct disttable *dist) 330{ 331 s64 x; 332 long t; 333 u32 rnd; 334 335 if (sigma == 0) 336 return mu; 337 338 rnd = get_crandom(state, prng); 339 340 /* default uniform distribution */ 341 if (dist == NULL) 342 return ((rnd % (2 * (u32)sigma)) + mu) - sigma; 343 344 t = dist->table[rnd % dist->size]; 345 x = (sigma % NETEM_DIST_SCALE) * t; 346 if (x >= 0) 347 x += NETEM_DIST_SCALE/2; 348 else 349 x -= NETEM_DIST_SCALE/2; 350 351 return x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu; 352} 353 354static u64 packet_time_ns(u64 len, const struct netem_sched_data *q) 355{ 356 len += q->packet_overhead; 357 358 if (q->cell_size) { 359 u32 cells = reciprocal_divide(len, q->cell_size_reciprocal); 360 361 if (len > cells * q->cell_size) /* extra cell needed for remainder */ 362 cells++; 363 len = cells * (q->cell_size + q->cell_overhead); 364 } 365 366 return div64_u64(len * NSEC_PER_SEC, q->rate); 367} 368 369static void tfifo_reset(struct Qdisc *sch) 370{ 371 struct netem_sched_data *q = qdisc_priv(sch); 372 struct rb_node *p = rb_first(&q->t_root); 373 374 while (p) { 375 struct sk_buff *skb = rb_to_skb(p); 376 377 p = rb_next(p); 378 rb_erase(&skb->rbnode, &q->t_root); 379 rtnl_kfree_skbs(skb, skb); 380 } 381 382 rtnl_kfree_skbs(q->t_head, q->t_tail); 383 q->t_head = NULL; 384 q->t_tail = NULL; 385} 386 387static void tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch) 388{ 389 struct netem_sched_data *q = qdisc_priv(sch); 390 u64 tnext = netem_skb_cb(nskb)->time_to_send; 391 392 if (!q->t_tail || tnext >= netem_skb_cb(q->t_tail)->time_to_send) { 393 if (q->t_tail) 394 q->t_tail->next = nskb; 395 else 396 q->t_head = nskb; 397 q->t_tail = nskb; 398 } else { 399 struct rb_node **p = &q->t_root.rb_node, *parent = NULL; 400 401 while (*p) { 402 struct sk_buff *skb; 403 404 parent = *p; 405 skb = rb_to_skb(parent); 406 if (tnext >= netem_skb_cb(skb)->time_to_send) 407 p = &parent->rb_right; 408 else 409 p = &parent->rb_left; 410 } 411 rb_link_node(&nskb->rbnode, parent, p); 412 rb_insert_color(&nskb->rbnode, &q->t_root); 413 } 414 sch->q.qlen++; 415} 416 417/* netem can't properly corrupt a megapacket (like we get from GSO), so instead 418 * when we statistically choose to corrupt one, we instead segment it, returning 419 * the first packet to be corrupted, and re-enqueue the remaining frames 420 */ 421static struct sk_buff *netem_segment(struct sk_buff *skb, struct Qdisc *sch, 422 struct sk_buff **to_free) 423{ 424 struct sk_buff *segs; 425 netdev_features_t features = netif_skb_features(skb); 426 427 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK); 428 429 if (IS_ERR_OR_NULL(segs)) { 430 qdisc_drop(skb, sch, to_free); 431 return NULL; 432 } 433 consume_skb(skb); 434 return segs; 435} 436 437/* 438 * Insert one skb into qdisc. 439 * Note: parent depends on return value to account for queue length. 440 * NET_XMIT_DROP: queue length didn't change. 441 * NET_XMIT_SUCCESS: one skb was queued. 442 */ 443static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch, 444 struct sk_buff **to_free) 445{ 446 struct netem_sched_data *q = qdisc_priv(sch); 447 /* We don't fill cb now as skb_unshare() may invalidate it */ 448 struct netem_skb_cb *cb; 449 struct sk_buff *skb2; 450 struct sk_buff *segs = NULL; 451 unsigned int prev_len = qdisc_pkt_len(skb); 452 int count = 1; 453 int rc = NET_XMIT_SUCCESS; 454 int rc_drop = NET_XMIT_DROP; 455 456 /* Do not fool qdisc_drop_all() */ 457 skb->prev = NULL; 458 459 /* Random duplication */ 460 if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor, &q->prng)) 461 ++count; 462 463 /* Drop packet? */ 464 if (loss_event(q)) { 465 if (q->ecn && INET_ECN_set_ce(skb)) 466 qdisc_qstats_drop(sch); /* mark packet */ 467 else 468 --count; 469 } 470 if (count == 0) { 471 qdisc_qstats_drop(sch); 472 __qdisc_drop(skb, to_free); 473 return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS; 474 } 475 476 /* If a delay is expected, orphan the skb. (orphaning usually takes 477 * place at TX completion time, so _before_ the link transit delay) 478 */ 479 if (q->latency || q->jitter || q->rate) 480 skb_orphan_partial(skb); 481 482 /* 483 * If we need to duplicate packet, then re-insert at top of the 484 * qdisc tree, since parent queuer expects that only one 485 * skb will be queued. 486 */ 487 if (count > 1 && (skb2 = skb_clone(skb, GFP_ATOMIC)) != NULL) { 488 struct Qdisc *rootq = qdisc_root_bh(sch); 489 u32 dupsave = q->duplicate; /* prevent duplicating a dup... */ 490 491 q->duplicate = 0; 492 rootq->enqueue(skb2, rootq, to_free); 493 q->duplicate = dupsave; 494 rc_drop = NET_XMIT_SUCCESS; 495 } 496 497 /* 498 * Randomized packet corruption. 499 * Make copy if needed since we are modifying 500 * If packet is going to be hardware checksummed, then 501 * do it now in software before we mangle it. 502 */ 503 if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor, &q->prng)) { 504 if (skb_is_gso(skb)) { 505 skb = netem_segment(skb, sch, to_free); 506 if (!skb) 507 return rc_drop; 508 segs = skb->next; 509 skb_mark_not_on_list(skb); 510 qdisc_skb_cb(skb)->pkt_len = skb->len; 511 } 512 513 skb = skb_unshare(skb, GFP_ATOMIC); 514 if (unlikely(!skb)) { 515 qdisc_qstats_drop(sch); 516 goto finish_segs; 517 } 518 if (skb->ip_summed == CHECKSUM_PARTIAL && 519 skb_checksum_help(skb)) { 520 qdisc_drop(skb, sch, to_free); 521 skb = NULL; 522 goto finish_segs; 523 } 524 525 skb->data[get_random_u32_below(skb_headlen(skb))] ^= 526 1<<get_random_u32_below(8); 527 } 528 529 if (unlikely(sch->q.qlen >= sch->limit)) { 530 /* re-link segs, so that qdisc_drop_all() frees them all */ 531 skb->next = segs; 532 qdisc_drop_all(skb, sch, to_free); 533 return rc_drop; 534 } 535 536 qdisc_qstats_backlog_inc(sch, skb); 537 538 cb = netem_skb_cb(skb); 539 if (q->gap == 0 || /* not doing reordering */ 540 q->counter < q->gap - 1 || /* inside last reordering gap */ 541 q->reorder < get_crandom(&q->reorder_cor, &q->prng)) { 542 u64 now; 543 s64 delay; 544 545 delay = tabledist(q->latency, q->jitter, 546 &q->delay_cor, &q->prng, q->delay_dist); 547 548 now = ktime_get_ns(); 549 550 if (q->rate) { 551 struct netem_skb_cb *last = NULL; 552 553 if (sch->q.tail) 554 last = netem_skb_cb(sch->q.tail); 555 if (q->t_root.rb_node) { 556 struct sk_buff *t_skb; 557 struct netem_skb_cb *t_last; 558 559 t_skb = skb_rb_last(&q->t_root); 560 t_last = netem_skb_cb(t_skb); 561 if (!last || 562 t_last->time_to_send > last->time_to_send) 563 last = t_last; 564 } 565 if (q->t_tail) { 566 struct netem_skb_cb *t_last = 567 netem_skb_cb(q->t_tail); 568 569 if (!last || 570 t_last->time_to_send > last->time_to_send) 571 last = t_last; 572 } 573 574 if (last) { 575 /* 576 * Last packet in queue is reference point (now), 577 * calculate this time bonus and subtract 578 * from delay. 579 */ 580 delay -= last->time_to_send - now; 581 delay = max_t(s64, 0, delay); 582 now = last->time_to_send; 583 } 584 585 delay += packet_time_ns(qdisc_pkt_len(skb), q); 586 } 587 588 cb->time_to_send = now + delay; 589 ++q->counter; 590 tfifo_enqueue(skb, sch); 591 } else { 592 /* 593 * Do re-ordering by putting one out of N packets at the front 594 * of the queue. 595 */ 596 cb->time_to_send = ktime_get_ns(); 597 q->counter = 0; 598 599 __qdisc_enqueue_head(skb, &sch->q); 600 sch->qstats.requeues++; 601 } 602 603finish_segs: 604 if (segs) { 605 unsigned int len, last_len; 606 int nb; 607 608 len = skb ? skb->len : 0; 609 nb = skb ? 1 : 0; 610 611 while (segs) { 612 skb2 = segs->next; 613 skb_mark_not_on_list(segs); 614 qdisc_skb_cb(segs)->pkt_len = segs->len; 615 last_len = segs->len; 616 rc = qdisc_enqueue(segs, sch, to_free); 617 if (rc != NET_XMIT_SUCCESS) { 618 if (net_xmit_drop_count(rc)) 619 qdisc_qstats_drop(sch); 620 } else { 621 nb++; 622 len += last_len; 623 } 624 segs = skb2; 625 } 626 /* Parent qdiscs accounted for 1 skb of size @prev_len */ 627 qdisc_tree_reduce_backlog(sch, -(nb - 1), -(len - prev_len)); 628 } else if (!skb) { 629 return NET_XMIT_DROP; 630 } 631 return NET_XMIT_SUCCESS; 632} 633 634/* Delay the next round with a new future slot with a 635 * correct number of bytes and packets. 636 */ 637 638static void get_slot_next(struct netem_sched_data *q, u64 now) 639{ 640 s64 next_delay; 641 642 if (!q->slot_dist) 643 next_delay = q->slot_config.min_delay + 644 (get_random_u32() * 645 (q->slot_config.max_delay - 646 q->slot_config.min_delay) >> 32); 647 else 648 next_delay = tabledist(q->slot_config.dist_delay, 649 (s32)(q->slot_config.dist_jitter), 650 NULL, &q->prng, q->slot_dist); 651 652 q->slot.slot_next = now + next_delay; 653 q->slot.packets_left = q->slot_config.max_packets; 654 q->slot.bytes_left = q->slot_config.max_bytes; 655} 656 657static struct sk_buff *netem_peek(struct netem_sched_data *q) 658{ 659 struct sk_buff *skb = skb_rb_first(&q->t_root); 660 u64 t1, t2; 661 662 if (!skb) 663 return q->t_head; 664 if (!q->t_head) 665 return skb; 666 667 t1 = netem_skb_cb(skb)->time_to_send; 668 t2 = netem_skb_cb(q->t_head)->time_to_send; 669 if (t1 < t2) 670 return skb; 671 return q->t_head; 672} 673 674static void netem_erase_head(struct netem_sched_data *q, struct sk_buff *skb) 675{ 676 if (skb == q->t_head) { 677 q->t_head = skb->next; 678 if (!q->t_head) 679 q->t_tail = NULL; 680 } else { 681 rb_erase(&skb->rbnode, &q->t_root); 682 } 683} 684 685static struct sk_buff *netem_dequeue(struct Qdisc *sch) 686{ 687 struct netem_sched_data *q = qdisc_priv(sch); 688 struct sk_buff *skb; 689 690tfifo_dequeue: 691 skb = __qdisc_dequeue_head(&sch->q); 692 if (skb) { 693 qdisc_qstats_backlog_dec(sch, skb); 694deliver: 695 qdisc_bstats_update(sch, skb); 696 return skb; 697 } 698 skb = netem_peek(q); 699 if (skb) { 700 u64 time_to_send; 701 u64 now = ktime_get_ns(); 702 703 /* if more time remaining? */ 704 time_to_send = netem_skb_cb(skb)->time_to_send; 705 if (q->slot.slot_next && q->slot.slot_next < time_to_send) 706 get_slot_next(q, now); 707 708 if (time_to_send <= now && q->slot.slot_next <= now) { 709 netem_erase_head(q, skb); 710 sch->q.qlen--; 711 qdisc_qstats_backlog_dec(sch, skb); 712 skb->next = NULL; 713 skb->prev = NULL; 714 /* skb->dev shares skb->rbnode area, 715 * we need to restore its value. 716 */ 717 skb->dev = qdisc_dev(sch); 718 719 if (q->slot.slot_next) { 720 q->slot.packets_left--; 721 q->slot.bytes_left -= qdisc_pkt_len(skb); 722 if (q->slot.packets_left <= 0 || 723 q->slot.bytes_left <= 0) 724 get_slot_next(q, now); 725 } 726 727 if (q->qdisc) { 728 unsigned int pkt_len = qdisc_pkt_len(skb); 729 struct sk_buff *to_free = NULL; 730 int err; 731 732 err = qdisc_enqueue(skb, q->qdisc, &to_free); 733 kfree_skb_list(to_free); 734 if (err != NET_XMIT_SUCCESS && 735 net_xmit_drop_count(err)) { 736 qdisc_qstats_drop(sch); 737 qdisc_tree_reduce_backlog(sch, 1, 738 pkt_len); 739 } 740 goto tfifo_dequeue; 741 } 742 goto deliver; 743 } 744 745 if (q->qdisc) { 746 skb = q->qdisc->ops->dequeue(q->qdisc); 747 if (skb) 748 goto deliver; 749 } 750 751 qdisc_watchdog_schedule_ns(&q->watchdog, 752 max(time_to_send, 753 q->slot.slot_next)); 754 } 755 756 if (q->qdisc) { 757 skb = q->qdisc->ops->dequeue(q->qdisc); 758 if (skb) 759 goto deliver; 760 } 761 return NULL; 762} 763 764static void netem_reset(struct Qdisc *sch) 765{ 766 struct netem_sched_data *q = qdisc_priv(sch); 767 768 qdisc_reset_queue(sch); 769 tfifo_reset(sch); 770 if (q->qdisc) 771 qdisc_reset(q->qdisc); 772 qdisc_watchdog_cancel(&q->watchdog); 773} 774 775static void dist_free(struct disttable *d) 776{ 777 kvfree(d); 778} 779 780/* 781 * Distribution data is a variable size payload containing 782 * signed 16 bit values. 783 */ 784 785static int get_dist_table(struct disttable **tbl, const struct nlattr *attr) 786{ 787 size_t n = nla_len(attr)/sizeof(__s16); 788 const __s16 *data = nla_data(attr); 789 struct disttable *d; 790 int i; 791 792 if (!n || n > NETEM_DIST_MAX) 793 return -EINVAL; 794 795 d = kvmalloc(struct_size(d, table, n), GFP_KERNEL); 796 if (!d) 797 return -ENOMEM; 798 799 d->size = n; 800 for (i = 0; i < n; i++) 801 d->table[i] = data[i]; 802 803 *tbl = d; 804 return 0; 805} 806 807static void get_slot(struct netem_sched_data *q, const struct nlattr *attr) 808{ 809 const struct tc_netem_slot *c = nla_data(attr); 810 811 q->slot_config = *c; 812 if (q->slot_config.max_packets == 0) 813 q->slot_config.max_packets = INT_MAX; 814 if (q->slot_config.max_bytes == 0) 815 q->slot_config.max_bytes = INT_MAX; 816 817 /* capping dist_jitter to the range acceptable by tabledist() */ 818 q->slot_config.dist_jitter = min_t(__s64, INT_MAX, abs(q->slot_config.dist_jitter)); 819 820 q->slot.packets_left = q->slot_config.max_packets; 821 q->slot.bytes_left = q->slot_config.max_bytes; 822 if (q->slot_config.min_delay | q->slot_config.max_delay | 823 q->slot_config.dist_jitter) 824 q->slot.slot_next = ktime_get_ns(); 825 else 826 q->slot.slot_next = 0; 827} 828 829static void get_correlation(struct netem_sched_data *q, const struct nlattr *attr) 830{ 831 const struct tc_netem_corr *c = nla_data(attr); 832 833 init_crandom(&q->delay_cor, c->delay_corr); 834 init_crandom(&q->loss_cor, c->loss_corr); 835 init_crandom(&q->dup_cor, c->dup_corr); 836} 837 838static void get_reorder(struct netem_sched_data *q, const struct nlattr *attr) 839{ 840 const struct tc_netem_reorder *r = nla_data(attr); 841 842 q->reorder = r->probability; 843 init_crandom(&q->reorder_cor, r->correlation); 844} 845 846static void get_corrupt(struct netem_sched_data *q, const struct nlattr *attr) 847{ 848 const struct tc_netem_corrupt *r = nla_data(attr); 849 850 q->corrupt = r->probability; 851 init_crandom(&q->corrupt_cor, r->correlation); 852} 853 854static void get_rate(struct netem_sched_data *q, const struct nlattr *attr) 855{ 856 const struct tc_netem_rate *r = nla_data(attr); 857 858 q->rate = r->rate; 859 q->packet_overhead = r->packet_overhead; 860 q->cell_size = r->cell_size; 861 q->cell_overhead = r->cell_overhead; 862 if (q->cell_size) 863 q->cell_size_reciprocal = reciprocal_value(q->cell_size); 864 else 865 q->cell_size_reciprocal = (struct reciprocal_value) { 0 }; 866} 867 868static int get_loss_clg(struct netem_sched_data *q, const struct nlattr *attr) 869{ 870 const struct nlattr *la; 871 int rem; 872 873 nla_for_each_nested(la, attr, rem) { 874 u16 type = nla_type(la); 875 876 switch (type) { 877 case NETEM_LOSS_GI: { 878 const struct tc_netem_gimodel *gi = nla_data(la); 879 880 if (nla_len(la) < sizeof(struct tc_netem_gimodel)) { 881 pr_info("netem: incorrect gi model size\n"); 882 return -EINVAL; 883 } 884 885 q->loss_model = CLG_4_STATES; 886 887 q->clg.state = TX_IN_GAP_PERIOD; 888 q->clg.a1 = gi->p13; 889 q->clg.a2 = gi->p31; 890 q->clg.a3 = gi->p32; 891 q->clg.a4 = gi->p14; 892 q->clg.a5 = gi->p23; 893 break; 894 } 895 896 case NETEM_LOSS_GE: { 897 const struct tc_netem_gemodel *ge = nla_data(la); 898 899 if (nla_len(la) < sizeof(struct tc_netem_gemodel)) { 900 pr_info("netem: incorrect ge model size\n"); 901 return -EINVAL; 902 } 903 904 q->loss_model = CLG_GILB_ELL; 905 q->clg.state = GOOD_STATE; 906 q->clg.a1 = ge->p; 907 q->clg.a2 = ge->r; 908 q->clg.a3 = ge->h; 909 q->clg.a4 = ge->k1; 910 break; 911 } 912 913 default: 914 pr_info("netem: unknown loss type %u\n", type); 915 return -EINVAL; 916 } 917 } 918 919 return 0; 920} 921 922static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = { 923 [TCA_NETEM_CORR] = { .len = sizeof(struct tc_netem_corr) }, 924 [TCA_NETEM_REORDER] = { .len = sizeof(struct tc_netem_reorder) }, 925 [TCA_NETEM_CORRUPT] = { .len = sizeof(struct tc_netem_corrupt) }, 926 [TCA_NETEM_RATE] = { .len = sizeof(struct tc_netem_rate) }, 927 [TCA_NETEM_LOSS] = { .type = NLA_NESTED }, 928 [TCA_NETEM_ECN] = { .type = NLA_U32 }, 929 [TCA_NETEM_RATE64] = { .type = NLA_U64 }, 930 [TCA_NETEM_LATENCY64] = { .type = NLA_S64 }, 931 [TCA_NETEM_JITTER64] = { .type = NLA_S64 }, 932 [TCA_NETEM_SLOT] = { .len = sizeof(struct tc_netem_slot) }, 933 [TCA_NETEM_PRNG_SEED] = { .type = NLA_U64 }, 934}; 935 936static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla, 937 const struct nla_policy *policy, int len) 938{ 939 int nested_len = nla_len(nla) - NLA_ALIGN(len); 940 941 if (nested_len < 0) { 942 pr_info("netem: invalid attributes len %d\n", nested_len); 943 return -EINVAL; 944 } 945 946 if (nested_len >= nla_attr_size(0)) 947 return nla_parse_deprecated(tb, maxtype, 948 nla_data(nla) + NLA_ALIGN(len), 949 nested_len, policy, NULL); 950 951 memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1)); 952 return 0; 953} 954 955/* Parse netlink message to set options */ 956static int netem_change(struct Qdisc *sch, struct nlattr *opt, 957 struct netlink_ext_ack *extack) 958{ 959 struct netem_sched_data *q = qdisc_priv(sch); 960 struct nlattr *tb[TCA_NETEM_MAX + 1]; 961 struct disttable *delay_dist = NULL; 962 struct disttable *slot_dist = NULL; 963 struct tc_netem_qopt *qopt; 964 struct clgstate old_clg; 965 int old_loss_model = CLG_RANDOM; 966 int ret; 967 968 qopt = nla_data(opt); 969 ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt)); 970 if (ret < 0) 971 return ret; 972 973 if (tb[TCA_NETEM_DELAY_DIST]) { 974 ret = get_dist_table(&delay_dist, tb[TCA_NETEM_DELAY_DIST]); 975 if (ret) 976 goto table_free; 977 } 978 979 if (tb[TCA_NETEM_SLOT_DIST]) { 980 ret = get_dist_table(&slot_dist, tb[TCA_NETEM_SLOT_DIST]); 981 if (ret) 982 goto table_free; 983 } 984 985 sch_tree_lock(sch); 986 /* backup q->clg and q->loss_model */ 987 old_clg = q->clg; 988 old_loss_model = q->loss_model; 989 990 if (tb[TCA_NETEM_LOSS]) { 991 ret = get_loss_clg(q, tb[TCA_NETEM_LOSS]); 992 if (ret) { 993 q->loss_model = old_loss_model; 994 q->clg = old_clg; 995 goto unlock; 996 } 997 } else { 998 q->loss_model = CLG_RANDOM; 999 } 1000 1001 if (delay_dist) 1002 swap(q->delay_dist, delay_dist); 1003 if (slot_dist) 1004 swap(q->slot_dist, slot_dist); 1005 sch->limit = qopt->limit; 1006 1007 q->latency = PSCHED_TICKS2NS(qopt->latency); 1008 q->jitter = PSCHED_TICKS2NS(qopt->jitter); 1009 q->limit = qopt->limit; 1010 q->gap = qopt->gap; 1011 q->counter = 0; 1012 q->loss = qopt->loss; 1013 q->duplicate = qopt->duplicate; 1014 1015 /* for compatibility with earlier versions. 1016 * if gap is set, need to assume 100% probability 1017 */ 1018 if (q->gap) 1019 q->reorder = ~0; 1020 1021 if (tb[TCA_NETEM_CORR]) 1022 get_correlation(q, tb[TCA_NETEM_CORR]); 1023 1024 if (tb[TCA_NETEM_REORDER]) 1025 get_reorder(q, tb[TCA_NETEM_REORDER]); 1026 1027 if (tb[TCA_NETEM_CORRUPT]) 1028 get_corrupt(q, tb[TCA_NETEM_CORRUPT]); 1029 1030 if (tb[TCA_NETEM_RATE]) 1031 get_rate(q, tb[TCA_NETEM_RATE]); 1032 1033 if (tb[TCA_NETEM_RATE64]) 1034 q->rate = max_t(u64, q->rate, 1035 nla_get_u64(tb[TCA_NETEM_RATE64])); 1036 1037 if (tb[TCA_NETEM_LATENCY64]) 1038 q->latency = nla_get_s64(tb[TCA_NETEM_LATENCY64]); 1039 1040 if (tb[TCA_NETEM_JITTER64]) 1041 q->jitter = nla_get_s64(tb[TCA_NETEM_JITTER64]); 1042 1043 if (tb[TCA_NETEM_ECN]) 1044 q->ecn = nla_get_u32(tb[TCA_NETEM_ECN]); 1045 1046 if (tb[TCA_NETEM_SLOT]) 1047 get_slot(q, tb[TCA_NETEM_SLOT]); 1048 1049 /* capping jitter to the range acceptable by tabledist() */ 1050 q->jitter = min_t(s64, abs(q->jitter), INT_MAX); 1051 1052 if (tb[TCA_NETEM_PRNG_SEED]) 1053 q->prng.seed = nla_get_u64(tb[TCA_NETEM_PRNG_SEED]); 1054 else 1055 q->prng.seed = get_random_u64(); 1056 prandom_seed_state(&q->prng.prng_state, q->prng.seed); 1057 1058unlock: 1059 sch_tree_unlock(sch); 1060 1061table_free: 1062 dist_free(delay_dist); 1063 dist_free(slot_dist); 1064 return ret; 1065} 1066 1067static int netem_init(struct Qdisc *sch, struct nlattr *opt, 1068 struct netlink_ext_ack *extack) 1069{ 1070 struct netem_sched_data *q = qdisc_priv(sch); 1071 int ret; 1072 1073 qdisc_watchdog_init(&q->watchdog, sch); 1074 1075 if (!opt) 1076 return -EINVAL; 1077 1078 q->loss_model = CLG_RANDOM; 1079 ret = netem_change(sch, opt, extack); 1080 if (ret) 1081 pr_info("netem: change failed\n"); 1082 return ret; 1083} 1084 1085static void netem_destroy(struct Qdisc *sch) 1086{ 1087 struct netem_sched_data *q = qdisc_priv(sch); 1088 1089 qdisc_watchdog_cancel(&q->watchdog); 1090 if (q->qdisc) 1091 qdisc_put(q->qdisc); 1092 dist_free(q->delay_dist); 1093 dist_free(q->slot_dist); 1094} 1095 1096static int dump_loss_model(const struct netem_sched_data *q, 1097 struct sk_buff *skb) 1098{ 1099 struct nlattr *nest; 1100 1101 nest = nla_nest_start_noflag(skb, TCA_NETEM_LOSS); 1102 if (nest == NULL) 1103 goto nla_put_failure; 1104 1105 switch (q->loss_model) { 1106 case CLG_RANDOM: 1107 /* legacy loss model */ 1108 nla_nest_cancel(skb, nest); 1109 return 0; /* no data */ 1110 1111 case CLG_4_STATES: { 1112 struct tc_netem_gimodel gi = { 1113 .p13 = q->clg.a1, 1114 .p31 = q->clg.a2, 1115 .p32 = q->clg.a3, 1116 .p14 = q->clg.a4, 1117 .p23 = q->clg.a5, 1118 }; 1119 1120 if (nla_put(skb, NETEM_LOSS_GI, sizeof(gi), &gi)) 1121 goto nla_put_failure; 1122 break; 1123 } 1124 case CLG_GILB_ELL: { 1125 struct tc_netem_gemodel ge = { 1126 .p = q->clg.a1, 1127 .r = q->clg.a2, 1128 .h = q->clg.a3, 1129 .k1 = q->clg.a4, 1130 }; 1131 1132 if (nla_put(skb, NETEM_LOSS_GE, sizeof(ge), &ge)) 1133 goto nla_put_failure; 1134 break; 1135 } 1136 } 1137 1138 nla_nest_end(skb, nest); 1139 return 0; 1140 1141nla_put_failure: 1142 nla_nest_cancel(skb, nest); 1143 return -1; 1144} 1145 1146static int netem_dump(struct Qdisc *sch, struct sk_buff *skb) 1147{ 1148 const struct netem_sched_data *q = qdisc_priv(sch); 1149 struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb); 1150 struct tc_netem_qopt qopt; 1151 struct tc_netem_corr cor; 1152 struct tc_netem_reorder reorder; 1153 struct tc_netem_corrupt corrupt; 1154 struct tc_netem_rate rate; 1155 struct tc_netem_slot slot; 1156 1157 qopt.latency = min_t(psched_time_t, PSCHED_NS2TICKS(q->latency), 1158 UINT_MAX); 1159 qopt.jitter = min_t(psched_time_t, PSCHED_NS2TICKS(q->jitter), 1160 UINT_MAX); 1161 qopt.limit = q->limit; 1162 qopt.loss = q->loss; 1163 qopt.gap = q->gap; 1164 qopt.duplicate = q->duplicate; 1165 if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt)) 1166 goto nla_put_failure; 1167 1168 if (nla_put(skb, TCA_NETEM_LATENCY64, sizeof(q->latency), &q->latency)) 1169 goto nla_put_failure; 1170 1171 if (nla_put(skb, TCA_NETEM_JITTER64, sizeof(q->jitter), &q->jitter)) 1172 goto nla_put_failure; 1173 1174 cor.delay_corr = q->delay_cor.rho; 1175 cor.loss_corr = q->loss_cor.rho; 1176 cor.dup_corr = q->dup_cor.rho; 1177 if (nla_put(skb, TCA_NETEM_CORR, sizeof(cor), &cor)) 1178 goto nla_put_failure; 1179 1180 reorder.probability = q->reorder; 1181 reorder.correlation = q->reorder_cor.rho; 1182 if (nla_put(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder)) 1183 goto nla_put_failure; 1184 1185 corrupt.probability = q->corrupt; 1186 corrupt.correlation = q->corrupt_cor.rho; 1187 if (nla_put(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt)) 1188 goto nla_put_failure; 1189 1190 if (q->rate >= (1ULL << 32)) { 1191 if (nla_put_u64_64bit(skb, TCA_NETEM_RATE64, q->rate, 1192 TCA_NETEM_PAD)) 1193 goto nla_put_failure; 1194 rate.rate = ~0U; 1195 } else { 1196 rate.rate = q->rate; 1197 } 1198 rate.packet_overhead = q->packet_overhead; 1199 rate.cell_size = q->cell_size; 1200 rate.cell_overhead = q->cell_overhead; 1201 if (nla_put(skb, TCA_NETEM_RATE, sizeof(rate), &rate)) 1202 goto nla_put_failure; 1203 1204 if (q->ecn && nla_put_u32(skb, TCA_NETEM_ECN, q->ecn)) 1205 goto nla_put_failure; 1206 1207 if (dump_loss_model(q, skb) != 0) 1208 goto nla_put_failure; 1209 1210 if (q->slot_config.min_delay | q->slot_config.max_delay | 1211 q->slot_config.dist_jitter) { 1212 slot = q->slot_config; 1213 if (slot.max_packets == INT_MAX) 1214 slot.max_packets = 0; 1215 if (slot.max_bytes == INT_MAX) 1216 slot.max_bytes = 0; 1217 if (nla_put(skb, TCA_NETEM_SLOT, sizeof(slot), &slot)) 1218 goto nla_put_failure; 1219 } 1220 1221 if (nla_put_u64_64bit(skb, TCA_NETEM_PRNG_SEED, q->prng.seed, 1222 TCA_NETEM_PAD)) 1223 goto nla_put_failure; 1224 1225 return nla_nest_end(skb, nla); 1226 1227nla_put_failure: 1228 nlmsg_trim(skb, nla); 1229 return -1; 1230} 1231 1232static int netem_dump_class(struct Qdisc *sch, unsigned long cl, 1233 struct sk_buff *skb, struct tcmsg *tcm) 1234{ 1235 struct netem_sched_data *q = qdisc_priv(sch); 1236 1237 if (cl != 1 || !q->qdisc) /* only one class */ 1238 return -ENOENT; 1239 1240 tcm->tcm_handle |= TC_H_MIN(1); 1241 tcm->tcm_info = q->qdisc->handle; 1242 1243 return 0; 1244} 1245 1246static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new, 1247 struct Qdisc **old, struct netlink_ext_ack *extack) 1248{ 1249 struct netem_sched_data *q = qdisc_priv(sch); 1250 1251 *old = qdisc_replace(sch, new, &q->qdisc); 1252 return 0; 1253} 1254 1255static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg) 1256{ 1257 struct netem_sched_data *q = qdisc_priv(sch); 1258 return q->qdisc; 1259} 1260 1261static unsigned long netem_find(struct Qdisc *sch, u32 classid) 1262{ 1263 return 1; 1264} 1265 1266static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker) 1267{ 1268 if (!walker->stop) { 1269 if (!tc_qdisc_stats_dump(sch, 1, walker)) 1270 return; 1271 } 1272} 1273 1274static const struct Qdisc_class_ops netem_class_ops = { 1275 .graft = netem_graft, 1276 .leaf = netem_leaf, 1277 .find = netem_find, 1278 .walk = netem_walk, 1279 .dump = netem_dump_class, 1280}; 1281 1282static struct Qdisc_ops netem_qdisc_ops __read_mostly = { 1283 .id = "netem", 1284 .cl_ops = &netem_class_ops, 1285 .priv_size = sizeof(struct netem_sched_data), 1286 .enqueue = netem_enqueue, 1287 .dequeue = netem_dequeue, 1288 .peek = qdisc_peek_dequeued, 1289 .init = netem_init, 1290 .reset = netem_reset, 1291 .destroy = netem_destroy, 1292 .change = netem_change, 1293 .dump = netem_dump, 1294 .owner = THIS_MODULE, 1295}; 1296MODULE_ALIAS_NET_SCH("netem"); 1297 1298 1299static int __init netem_module_init(void) 1300{ 1301 pr_info("netem: version " VERSION "\n"); 1302 return register_qdisc(&netem_qdisc_ops); 1303} 1304static void __exit netem_module_exit(void) 1305{ 1306 unregister_qdisc(&netem_qdisc_ops); 1307} 1308module_init(netem_module_init) 1309module_exit(netem_module_exit) 1310MODULE_LICENSE("GPL"); 1311MODULE_DESCRIPTION("Network characteristics emulator qdisc"); 1312