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/prandom.h> 21 #include <linux/rtnetlink.h> 22 #include <linux/reciprocal_div.h> 23 #include <linux/rbtree.h> 24 25 #include <net/gso.h> 26 #include <net/netlink.h> 27 #include <net/pkt_sched.h> 28 #include <net/inet_ecn.h> 29 30 #define VERSION "1.3" 31 32 /* Network Emulation Queuing algorithm. 33 ==================================== 34 35 Sources: [1] Mark Carson, Darrin Santay, "NIST Net - A Linux-based 36 Network Emulation Tool 37 [2] Luigi Rizzo, DummyNet for FreeBSD 38 39 ---------------------------------------------------------------- 40 41 This started out as a simple way to delay outgoing packets to 42 test TCP but has grown to include most of the functionality 43 of a full blown network emulator like NISTnet. It can delay 44 packets and add random jitter (and correlation). The random 45 distribution can be loaded from a table as well to provide 46 normal, Pareto, or experimental curves. Packet loss, 47 duplication, and reordering can also be emulated. 48 49 This qdisc does not do classification that can be handled in 50 layering other disciplines. It does not need to do bandwidth 51 control either since that can be handled by using token 52 bucket or other rate control. 53 54 Correlated Loss Generator models 55 56 Added generation of correlated loss according to the 57 "Gilbert-Elliot" model, a 4-state markov model. 58 59 References: 60 [1] NetemCLG Home http://netgroup.uniroma2.it/NetemCLG 61 [2] S. Salsano, F. Ludovici, A. Ordine, "Definition of a general 62 and intuitive loss model for packet networks and its implementation 63 in the Netem module in the Linux kernel", available in [1] 64 65 Authors: Stefano Salsano <stefano.salsano at uniroma2.it 66 Fabio Ludovici <fabio.ludovici at yahoo.it> 67 */ 68 69 struct disttable { 70 u32 size; 71 s16 table[] __counted_by(size); 72 }; 73 74 struct netem_sched_data { 75 /* internal t(ime)fifo qdisc uses t_root and sch->limit */ 76 struct rb_root t_root; 77 78 /* a linear queue; reduces rbtree rebalancing when jitter is low */ 79 struct sk_buff *t_head; 80 struct sk_buff *t_tail; 81 82 u32 t_len; 83 84 /* optional qdisc for classful handling (NULL at netem init) */ 85 struct Qdisc *qdisc; 86 87 struct qdisc_watchdog watchdog; 88 89 s64 latency; 90 s64 jitter; 91 92 u32 loss; 93 u32 ecn; 94 u32 limit; 95 u32 counter; 96 u32 gap; 97 u32 duplicate; 98 u32 reorder; 99 u32 corrupt; 100 u64 rate; 101 s32 packet_overhead; 102 u32 cell_size; 103 struct reciprocal_value cell_size_reciprocal; 104 s32 cell_overhead; 105 106 struct crndstate { 107 u32 last; 108 u32 rho; 109 } delay_cor, loss_cor, dup_cor, reorder_cor, corrupt_cor; 110 111 struct prng { 112 u64 seed; 113 struct rnd_state prng_state; 114 } prng; 115 116 struct disttable *delay_dist; 117 118 enum { 119 CLG_RANDOM, 120 CLG_4_STATES, 121 CLG_GILB_ELL, 122 } loss_model; 123 124 enum { 125 TX_IN_GAP_PERIOD = 1, 126 TX_IN_BURST_PERIOD, 127 LOST_IN_GAP_PERIOD, 128 LOST_IN_BURST_PERIOD, 129 } _4_state_model; 130 131 enum { 132 GOOD_STATE = 1, 133 BAD_STATE, 134 } GE_state_model; 135 136 /* Correlated Loss Generation models */ 137 struct clgstate { 138 /* state of the Markov chain */ 139 u8 state; 140 141 /* 4-states and Gilbert-Elliot models */ 142 u32 a1; /* p13 for 4-states or p for GE */ 143 u32 a2; /* p31 for 4-states or r for GE */ 144 u32 a3; /* p32 for 4-states or h for GE */ 145 u32 a4; /* p14 for 4-states or 1-k for GE */ 146 u32 a5; /* p23 used only in 4-states */ 147 } clg; 148 149 struct tc_netem_slot slot_config; 150 struct slotstate { 151 u64 slot_next; 152 s32 packets_left; 153 s32 bytes_left; 154 } slot; 155 156 struct disttable *slot_dist; 157 }; 158 159 /* Time stamp put into socket buffer control block 160 * Only valid when skbs are in our internal t(ime)fifo queue. 161 * 162 * As skb->rbnode uses same storage than skb->next, skb->prev and skb->tstamp, 163 * and skb->next & skb->prev are scratch space for a qdisc, 164 * we save skb->tstamp value in skb->cb[] before destroying it. 165 */ 166 struct netem_skb_cb { 167 u64 time_to_send; 168 }; 169 170 static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb) 171 { 172 /* we assume we can use skb next/prev/tstamp as storage for rb_node */ 173 qdisc_cb_private_validate(skb, sizeof(struct netem_skb_cb)); 174 return (struct netem_skb_cb *)qdisc_skb_cb(skb)->data; 175 } 176 177 /* init_crandom - initialize correlated random number generator 178 * Use entropy source for initial seed. 179 */ 180 static void init_crandom(struct crndstate *state, unsigned long rho) 181 { 182 state->rho = rho; 183 state->last = get_random_u32(); 184 } 185 186 /* get_crandom - correlated random number generator 187 * Next number depends on last value. 188 * rho is scaled to avoid floating point. 189 */ 190 static u32 get_crandom(struct crndstate *state, struct prng *p) 191 { 192 u64 value, rho; 193 unsigned long answer; 194 struct rnd_state *s = &p->prng_state; 195 196 if (!state || state->rho == 0) /* no correlation */ 197 return prandom_u32_state(s); 198 199 value = prandom_u32_state(s); 200 rho = (u64)state->rho + 1; 201 answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32; 202 state->last = answer; 203 return answer; 204 } 205 206 /* loss_4state - 4-state model loss generator 207 * Generates losses according to the 4-state Markov chain adopted in 208 * the GI (General and Intuitive) loss model. 209 */ 210 static bool loss_4state(struct netem_sched_data *q) 211 { 212 struct clgstate *clg = &q->clg; 213 u32 rnd = prandom_u32_state(&q->prng.prng_state); 214 215 /* 216 * Makes a comparison between rnd and the transition 217 * probabilities outgoing from the current state, then decides the 218 * next state and if the next packet has to be transmitted or lost. 219 * The four states correspond to: 220 * TX_IN_GAP_PERIOD => successfully transmitted packets within a gap period 221 * LOST_IN_GAP_PERIOD => isolated losses within a gap period 222 * LOST_IN_BURST_PERIOD => lost packets within a burst period 223 * TX_IN_BURST_PERIOD => successfully transmitted packets within a burst period 224 */ 225 switch (clg->state) { 226 case TX_IN_GAP_PERIOD: 227 if (rnd < clg->a4) { 228 clg->state = LOST_IN_GAP_PERIOD; 229 return true; 230 } else if (clg->a4 < rnd && rnd < clg->a1 + clg->a4) { 231 clg->state = LOST_IN_BURST_PERIOD; 232 return true; 233 } else if (clg->a1 + clg->a4 < rnd) { 234 clg->state = TX_IN_GAP_PERIOD; 235 } 236 237 break; 238 case TX_IN_BURST_PERIOD: 239 if (rnd < clg->a5) { 240 clg->state = LOST_IN_BURST_PERIOD; 241 return true; 242 } else { 243 clg->state = TX_IN_BURST_PERIOD; 244 } 245 246 break; 247 case LOST_IN_BURST_PERIOD: 248 if (rnd < clg->a3) 249 clg->state = TX_IN_BURST_PERIOD; 250 else if (clg->a3 < rnd && rnd < clg->a2 + clg->a3) { 251 clg->state = TX_IN_GAP_PERIOD; 252 } else if (clg->a2 + clg->a3 < rnd) { 253 clg->state = LOST_IN_BURST_PERIOD; 254 return true; 255 } 256 break; 257 case LOST_IN_GAP_PERIOD: 258 clg->state = TX_IN_GAP_PERIOD; 259 break; 260 } 261 262 return false; 263 } 264 265 /* loss_gilb_ell - Gilbert-Elliot model loss generator 266 * Generates losses according to the Gilbert-Elliot loss model or 267 * its special cases (Gilbert or Simple Gilbert) 268 * 269 * Makes a comparison between random number and the transition 270 * probabilities outgoing from the current state, then decides the 271 * next state. A second random number is extracted and the comparison 272 * with the loss probability of the current state decides if the next 273 * packet will be transmitted or lost. 274 */ 275 static bool loss_gilb_ell(struct netem_sched_data *q) 276 { 277 struct clgstate *clg = &q->clg; 278 struct rnd_state *s = &q->prng.prng_state; 279 280 switch (clg->state) { 281 case GOOD_STATE: 282 if (prandom_u32_state(s) < clg->a1) 283 clg->state = BAD_STATE; 284 if (prandom_u32_state(s) < clg->a4) 285 return true; 286 break; 287 case BAD_STATE: 288 if (prandom_u32_state(s) < clg->a2) 289 clg->state = GOOD_STATE; 290 if (prandom_u32_state(s) > clg->a3) 291 return true; 292 } 293 294 return false; 295 } 296 297 static bool loss_event(struct netem_sched_data *q) 298 { 299 switch (q->loss_model) { 300 case CLG_RANDOM: 301 /* Random packet drop 0 => none, ~0 => all */ 302 return q->loss && q->loss >= get_crandom(&q->loss_cor, &q->prng); 303 304 case CLG_4_STATES: 305 /* 4state loss model algorithm (used also for GI model) 306 * Extracts a value from the markov 4 state loss generator, 307 * if it is 1 drops a packet and if needed writes the event in 308 * the kernel logs 309 */ 310 return loss_4state(q); 311 312 case CLG_GILB_ELL: 313 /* Gilbert-Elliot loss model algorithm 314 * Extracts a value from the Gilbert-Elliot loss generator, 315 * if it is 1 drops a packet and if needed writes the event in 316 * the kernel logs 317 */ 318 return loss_gilb_ell(q); 319 } 320 321 return false; /* not reached */ 322 } 323 324 325 /* tabledist - return a pseudo-randomly distributed value with mean mu and 326 * std deviation sigma. Uses table lookup to approximate the desired 327 * distribution, and a uniformly-distributed pseudo-random source. 328 */ 329 static s64 tabledist(s64 mu, s32 sigma, 330 struct crndstate *state, 331 struct prng *prng, 332 const struct disttable *dist) 333 { 334 s64 x; 335 long t; 336 u32 rnd; 337 338 if (sigma == 0) 339 return mu; 340 341 rnd = get_crandom(state, prng); 342 343 /* default uniform distribution */ 344 if (dist == NULL) 345 return ((rnd % (2 * (u32)sigma)) + mu) - sigma; 346 347 t = dist->table[rnd % dist->size]; 348 x = (sigma % NETEM_DIST_SCALE) * t; 349 if (x >= 0) 350 x += NETEM_DIST_SCALE/2; 351 else 352 x -= NETEM_DIST_SCALE/2; 353 354 return x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu; 355 } 356 357 static u64 packet_time_ns(u64 len, const struct netem_sched_data *q) 358 { 359 len += q->packet_overhead; 360 361 if (q->cell_size) { 362 u32 cells = reciprocal_divide(len, q->cell_size_reciprocal); 363 364 if (len > cells * q->cell_size) /* extra cell needed for remainder */ 365 cells++; 366 len = cells * (q->cell_size + q->cell_overhead); 367 } 368 369 return div64_u64(len * NSEC_PER_SEC, q->rate); 370 } 371 372 static void tfifo_reset(struct Qdisc *sch) 373 { 374 struct netem_sched_data *q = qdisc_priv(sch); 375 struct rb_node *p = rb_first(&q->t_root); 376 377 while (p) { 378 struct sk_buff *skb = rb_to_skb(p); 379 380 p = rb_next(p); 381 rb_erase(&skb->rbnode, &q->t_root); 382 rtnl_kfree_skbs(skb, skb); 383 } 384 385 rtnl_kfree_skbs(q->t_head, q->t_tail); 386 q->t_head = NULL; 387 q->t_tail = NULL; 388 q->t_len = 0; 389 } 390 391 static void tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch) 392 { 393 struct netem_sched_data *q = qdisc_priv(sch); 394 u64 tnext = netem_skb_cb(nskb)->time_to_send; 395 396 if (!q->t_tail || tnext >= netem_skb_cb(q->t_tail)->time_to_send) { 397 if (q->t_tail) 398 q->t_tail->next = nskb; 399 else 400 q->t_head = nskb; 401 q->t_tail = nskb; 402 } else { 403 struct rb_node **p = &q->t_root.rb_node, *parent = NULL; 404 405 while (*p) { 406 struct sk_buff *skb; 407 408 parent = *p; 409 skb = rb_to_skb(parent); 410 if (tnext >= netem_skb_cb(skb)->time_to_send) 411 p = &parent->rb_right; 412 else 413 p = &parent->rb_left; 414 } 415 rb_link_node(&nskb->rbnode, parent, p); 416 rb_insert_color(&nskb->rbnode, &q->t_root); 417 } 418 q->t_len++; 419 sch->q.qlen++; 420 } 421 422 /* netem can't properly corrupt a megapacket (like we get from GSO), so instead 423 * when we statistically choose to corrupt one, we instead segment it, returning 424 * the first packet to be corrupted, and re-enqueue the remaining frames 425 */ 426 static struct sk_buff *netem_segment(struct sk_buff *skb, struct Qdisc *sch, 427 struct sk_buff **to_free) 428 { 429 struct sk_buff *segs; 430 netdev_features_t features = netif_skb_features(skb); 431 432 qdisc_skb_cb(skb)->pkt_segs = 1; 433 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK); 434 435 if (IS_ERR_OR_NULL(segs)) { 436 qdisc_drop(skb, sch, to_free); 437 return NULL; 438 } 439 consume_skb(skb); 440 return segs; 441 } 442 443 /* 444 * Insert one skb into qdisc. 445 * Note: parent depends on return value to account for queue length. 446 * NET_XMIT_DROP: queue length didn't change. 447 * NET_XMIT_SUCCESS: one skb was queued. 448 */ 449 static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch, 450 struct sk_buff **to_free) 451 { 452 struct netem_sched_data *q = qdisc_priv(sch); 453 /* We don't fill cb now as skb_unshare() may invalidate it */ 454 struct netem_skb_cb *cb; 455 struct sk_buff *skb2 = NULL; 456 struct sk_buff *segs = NULL; 457 unsigned int prev_len = qdisc_pkt_len(skb); 458 int count = 1; 459 460 /* Do not fool qdisc_drop_all() */ 461 skb->prev = NULL; 462 463 /* Random duplication */ 464 if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor, &q->prng)) 465 ++count; 466 467 /* Drop packet? */ 468 if (loss_event(q)) { 469 if (q->ecn && INET_ECN_set_ce(skb)) 470 qdisc_qstats_drop(sch); /* mark packet */ 471 else 472 --count; 473 } 474 if (count == 0) { 475 qdisc_qstats_drop(sch); 476 __qdisc_drop(skb, to_free); 477 return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS; 478 } 479 480 /* If a delay is expected, orphan the skb. (orphaning usually takes 481 * place at TX completion time, so _before_ the link transit delay) 482 */ 483 if (q->latency || q->jitter || q->rate) 484 skb_orphan_partial(skb); 485 486 /* 487 * If we need to duplicate packet, then clone it before 488 * original is modified. 489 */ 490 if (count > 1) 491 skb2 = skb_clone(skb, GFP_ATOMIC); 492 493 /* 494 * Randomized packet corruption. 495 * Make copy if needed since we are modifying 496 * If packet is going to be hardware checksummed, then 497 * do it now in software before we mangle it. 498 */ 499 if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor, &q->prng)) { 500 if (skb_is_gso(skb)) { 501 skb = netem_segment(skb, sch, to_free); 502 if (!skb) 503 goto finish_segs; 504 505 segs = skb->next; 506 skb_mark_not_on_list(skb); 507 qdisc_skb_cb(skb)->pkt_len = skb->len; 508 } 509 510 skb = skb_unshare(skb, GFP_ATOMIC); 511 if (unlikely(!skb)) { 512 qdisc_qstats_drop(sch); 513 goto finish_segs; 514 } 515 if (skb->ip_summed == CHECKSUM_PARTIAL && 516 skb_checksum_help(skb)) { 517 qdisc_drop(skb, sch, to_free); 518 skb = NULL; 519 goto finish_segs; 520 } 521 522 if (skb_headlen(skb)) 523 skb->data[get_random_u32_below(skb_headlen(skb))] ^= 524 1 << get_random_u32_below(8); 525 } 526 527 if (unlikely(q->t_len >= sch->limit)) { 528 /* re-link segs, so that qdisc_drop_all() frees them all */ 529 skb->next = segs; 530 qdisc_drop_all(skb, sch, to_free); 531 if (skb2) 532 __qdisc_drop(skb2, to_free); 533 return NET_XMIT_DROP; 534 } 535 536 /* 537 * If doing duplication then re-insert at top of the 538 * qdisc tree, since parent queuer expects that only one 539 * skb will be queued. 540 */ 541 if (skb2) { 542 struct Qdisc *rootq = qdisc_root_bh(sch); 543 u32 dupsave = q->duplicate; /* prevent duplicating a dup... */ 544 545 q->duplicate = 0; 546 rootq->enqueue(skb2, rootq, to_free); 547 q->duplicate = dupsave; 548 skb2 = NULL; 549 } 550 551 qdisc_qstats_backlog_inc(sch, skb); 552 553 cb = netem_skb_cb(skb); 554 if (q->gap == 0 || /* not doing reordering */ 555 q->counter < q->gap - 1 || /* inside last reordering gap */ 556 q->reorder < get_crandom(&q->reorder_cor, &q->prng)) { 557 u64 now; 558 s64 delay; 559 560 delay = tabledist(q->latency, q->jitter, 561 &q->delay_cor, &q->prng, q->delay_dist); 562 563 now = ktime_get_ns(); 564 565 if (q->rate) { 566 struct netem_skb_cb *last = NULL; 567 568 if (sch->q.tail) 569 last = netem_skb_cb(sch->q.tail); 570 if (q->t_root.rb_node) { 571 struct sk_buff *t_skb; 572 struct netem_skb_cb *t_last; 573 574 t_skb = skb_rb_last(&q->t_root); 575 t_last = netem_skb_cb(t_skb); 576 if (!last || 577 t_last->time_to_send > last->time_to_send) 578 last = t_last; 579 } 580 if (q->t_tail) { 581 struct netem_skb_cb *t_last = 582 netem_skb_cb(q->t_tail); 583 584 if (!last || 585 t_last->time_to_send > last->time_to_send) 586 last = t_last; 587 } 588 589 if (last) { 590 /* 591 * Last packet in queue is reference point (now), 592 * calculate this time bonus and subtract 593 * from delay. 594 */ 595 delay -= last->time_to_send - now; 596 delay = max_t(s64, 0, delay); 597 now = last->time_to_send; 598 } 599 600 delay += packet_time_ns(qdisc_pkt_len(skb), q); 601 } 602 603 cb->time_to_send = now + delay; 604 ++q->counter; 605 tfifo_enqueue(skb, sch); 606 } else { 607 /* 608 * Do re-ordering by putting one out of N packets at the front 609 * of the queue. 610 */ 611 cb->time_to_send = ktime_get_ns(); 612 q->counter = 0; 613 614 __qdisc_enqueue_head(skb, &sch->q); 615 sch->qstats.requeues++; 616 } 617 618 finish_segs: 619 if (skb2) 620 __qdisc_drop(skb2, to_free); 621 622 if (segs) { 623 unsigned int len, last_len; 624 int rc, nb; 625 626 len = skb ? skb->len : 0; 627 nb = skb ? 1 : 0; 628 629 while (segs) { 630 skb2 = segs->next; 631 skb_mark_not_on_list(segs); 632 qdisc_skb_cb(segs)->pkt_len = segs->len; 633 last_len = segs->len; 634 rc = qdisc_enqueue(segs, sch, to_free); 635 if (rc != NET_XMIT_SUCCESS) { 636 if (net_xmit_drop_count(rc)) 637 qdisc_qstats_drop(sch); 638 } else { 639 nb++; 640 len += last_len; 641 } 642 segs = skb2; 643 } 644 /* Parent qdiscs accounted for 1 skb of size @prev_len */ 645 qdisc_tree_reduce_backlog(sch, -(nb - 1), -(len - prev_len)); 646 } else if (!skb) { 647 return NET_XMIT_DROP; 648 } 649 return NET_XMIT_SUCCESS; 650 } 651 652 /* Delay the next round with a new future slot with a 653 * correct number of bytes and packets. 654 */ 655 656 static void get_slot_next(struct netem_sched_data *q, u64 now) 657 { 658 s64 next_delay; 659 660 if (!q->slot_dist) 661 next_delay = q->slot_config.min_delay + 662 (get_random_u32() * 663 (q->slot_config.max_delay - 664 q->slot_config.min_delay) >> 32); 665 else 666 next_delay = tabledist(q->slot_config.dist_delay, 667 (s32)(q->slot_config.dist_jitter), 668 NULL, &q->prng, q->slot_dist); 669 670 q->slot.slot_next = now + next_delay; 671 q->slot.packets_left = q->slot_config.max_packets; 672 q->slot.bytes_left = q->slot_config.max_bytes; 673 } 674 675 static struct sk_buff *netem_peek(struct netem_sched_data *q) 676 { 677 struct sk_buff *skb = skb_rb_first(&q->t_root); 678 u64 t1, t2; 679 680 if (!skb) 681 return q->t_head; 682 if (!q->t_head) 683 return skb; 684 685 t1 = netem_skb_cb(skb)->time_to_send; 686 t2 = netem_skb_cb(q->t_head)->time_to_send; 687 if (t1 < t2) 688 return skb; 689 return q->t_head; 690 } 691 692 static void netem_erase_head(struct netem_sched_data *q, struct sk_buff *skb) 693 { 694 if (skb == q->t_head) { 695 q->t_head = skb->next; 696 if (!q->t_head) 697 q->t_tail = NULL; 698 } else { 699 rb_erase(&skb->rbnode, &q->t_root); 700 } 701 } 702 703 static struct sk_buff *netem_dequeue(struct Qdisc *sch) 704 { 705 struct netem_sched_data *q = qdisc_priv(sch); 706 struct sk_buff *skb; 707 708 tfifo_dequeue: 709 skb = __qdisc_dequeue_head(&sch->q); 710 if (skb) { 711 deliver: 712 qdisc_qstats_backlog_dec(sch, skb); 713 qdisc_bstats_update(sch, skb); 714 return skb; 715 } 716 skb = netem_peek(q); 717 if (skb) { 718 u64 time_to_send; 719 u64 now = ktime_get_ns(); 720 721 /* if more time remaining? */ 722 time_to_send = netem_skb_cb(skb)->time_to_send; 723 if (q->slot.slot_next && q->slot.slot_next < time_to_send) 724 get_slot_next(q, now); 725 726 if (time_to_send <= now && q->slot.slot_next <= now) { 727 netem_erase_head(q, skb); 728 q->t_len--; 729 skb->next = NULL; 730 skb->prev = NULL; 731 /* skb->dev shares skb->rbnode area, 732 * we need to restore its value. 733 */ 734 skb->dev = qdisc_dev(sch); 735 736 if (q->slot.slot_next) { 737 q->slot.packets_left--; 738 q->slot.bytes_left -= qdisc_pkt_len(skb); 739 if (q->slot.packets_left <= 0 || 740 q->slot.bytes_left <= 0) 741 get_slot_next(q, now); 742 } 743 744 if (q->qdisc) { 745 unsigned int pkt_len = qdisc_pkt_len(skb); 746 struct sk_buff *to_free = NULL; 747 int err; 748 749 err = qdisc_enqueue(skb, q->qdisc, &to_free); 750 kfree_skb_list(to_free); 751 if (err != NET_XMIT_SUCCESS) { 752 if (net_xmit_drop_count(err)) 753 qdisc_qstats_drop(sch); 754 sch->qstats.backlog -= pkt_len; 755 sch->q.qlen--; 756 qdisc_tree_reduce_backlog(sch, 1, pkt_len); 757 } 758 goto tfifo_dequeue; 759 } 760 sch->q.qlen--; 761 goto deliver; 762 } 763 764 if (q->qdisc) { 765 skb = q->qdisc->ops->dequeue(q->qdisc); 766 if (skb) { 767 sch->q.qlen--; 768 goto deliver; 769 } 770 } 771 772 qdisc_watchdog_schedule_ns(&q->watchdog, 773 max(time_to_send, 774 q->slot.slot_next)); 775 } 776 777 if (q->qdisc) { 778 skb = q->qdisc->ops->dequeue(q->qdisc); 779 if (skb) { 780 sch->q.qlen--; 781 goto deliver; 782 } 783 } 784 return NULL; 785 } 786 787 static void netem_reset(struct Qdisc *sch) 788 { 789 struct netem_sched_data *q = qdisc_priv(sch); 790 791 qdisc_reset_queue(sch); 792 tfifo_reset(sch); 793 if (q->qdisc) 794 qdisc_reset(q->qdisc); 795 qdisc_watchdog_cancel(&q->watchdog); 796 } 797 798 static void dist_free(struct disttable *d) 799 { 800 kvfree(d); 801 } 802 803 /* 804 * Distribution data is a variable size payload containing 805 * signed 16 bit values. 806 */ 807 808 static int get_dist_table(struct disttable **tbl, const struct nlattr *attr) 809 { 810 size_t n = nla_len(attr)/sizeof(__s16); 811 const __s16 *data = nla_data(attr); 812 struct disttable *d; 813 int i; 814 815 if (!n || n > NETEM_DIST_MAX) 816 return -EINVAL; 817 818 d = kvmalloc_flex(*d, table, n); 819 if (!d) 820 return -ENOMEM; 821 822 d->size = n; 823 for (i = 0; i < n; i++) 824 d->table[i] = data[i]; 825 826 *tbl = d; 827 return 0; 828 } 829 830 static void get_slot(struct netem_sched_data *q, const struct nlattr *attr) 831 { 832 const struct tc_netem_slot *c = nla_data(attr); 833 834 q->slot_config = *c; 835 if (q->slot_config.max_packets == 0) 836 q->slot_config.max_packets = INT_MAX; 837 if (q->slot_config.max_bytes == 0) 838 q->slot_config.max_bytes = INT_MAX; 839 840 /* capping dist_jitter to the range acceptable by tabledist() */ 841 q->slot_config.dist_jitter = min_t(__s64, INT_MAX, abs(q->slot_config.dist_jitter)); 842 843 q->slot.packets_left = q->slot_config.max_packets; 844 q->slot.bytes_left = q->slot_config.max_bytes; 845 if (q->slot_config.min_delay | q->slot_config.max_delay | 846 q->slot_config.dist_jitter) 847 q->slot.slot_next = ktime_get_ns(); 848 else 849 q->slot.slot_next = 0; 850 } 851 852 static void get_correlation(struct netem_sched_data *q, const struct nlattr *attr) 853 { 854 const struct tc_netem_corr *c = nla_data(attr); 855 856 init_crandom(&q->delay_cor, c->delay_corr); 857 init_crandom(&q->loss_cor, c->loss_corr); 858 init_crandom(&q->dup_cor, c->dup_corr); 859 } 860 861 static void get_reorder(struct netem_sched_data *q, const struct nlattr *attr) 862 { 863 const struct tc_netem_reorder *r = nla_data(attr); 864 865 q->reorder = r->probability; 866 init_crandom(&q->reorder_cor, r->correlation); 867 } 868 869 static void get_corrupt(struct netem_sched_data *q, const struct nlattr *attr) 870 { 871 const struct tc_netem_corrupt *r = nla_data(attr); 872 873 q->corrupt = r->probability; 874 init_crandom(&q->corrupt_cor, r->correlation); 875 } 876 877 static void get_rate(struct netem_sched_data *q, const struct nlattr *attr) 878 { 879 const struct tc_netem_rate *r = nla_data(attr); 880 881 q->rate = r->rate; 882 q->packet_overhead = r->packet_overhead; 883 q->cell_size = r->cell_size; 884 q->cell_overhead = r->cell_overhead; 885 if (q->cell_size) 886 q->cell_size_reciprocal = reciprocal_value(q->cell_size); 887 else 888 q->cell_size_reciprocal = (struct reciprocal_value) { 0 }; 889 } 890 891 static int get_loss_clg(struct netem_sched_data *q, const struct nlattr *attr) 892 { 893 const struct nlattr *la; 894 int rem; 895 896 nla_for_each_nested(la, attr, rem) { 897 u16 type = nla_type(la); 898 899 switch (type) { 900 case NETEM_LOSS_GI: { 901 const struct tc_netem_gimodel *gi = nla_data(la); 902 903 if (nla_len(la) < sizeof(struct tc_netem_gimodel)) { 904 pr_info("netem: incorrect gi model size\n"); 905 return -EINVAL; 906 } 907 908 q->loss_model = CLG_4_STATES; 909 910 q->clg.state = TX_IN_GAP_PERIOD; 911 q->clg.a1 = gi->p13; 912 q->clg.a2 = gi->p31; 913 q->clg.a3 = gi->p32; 914 q->clg.a4 = gi->p14; 915 q->clg.a5 = gi->p23; 916 break; 917 } 918 919 case NETEM_LOSS_GE: { 920 const struct tc_netem_gemodel *ge = nla_data(la); 921 922 if (nla_len(la) < sizeof(struct tc_netem_gemodel)) { 923 pr_info("netem: incorrect ge model size\n"); 924 return -EINVAL; 925 } 926 927 q->loss_model = CLG_GILB_ELL; 928 q->clg.state = GOOD_STATE; 929 q->clg.a1 = ge->p; 930 q->clg.a2 = ge->r; 931 q->clg.a3 = ge->h; 932 q->clg.a4 = ge->k1; 933 break; 934 } 935 936 default: 937 pr_info("netem: unknown loss type %u\n", type); 938 return -EINVAL; 939 } 940 } 941 942 return 0; 943 } 944 945 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = { 946 [TCA_NETEM_CORR] = { .len = sizeof(struct tc_netem_corr) }, 947 [TCA_NETEM_REORDER] = { .len = sizeof(struct tc_netem_reorder) }, 948 [TCA_NETEM_CORRUPT] = { .len = sizeof(struct tc_netem_corrupt) }, 949 [TCA_NETEM_RATE] = { .len = sizeof(struct tc_netem_rate) }, 950 [TCA_NETEM_LOSS] = { .type = NLA_NESTED }, 951 [TCA_NETEM_ECN] = { .type = NLA_U32 }, 952 [TCA_NETEM_RATE64] = { .type = NLA_U64 }, 953 [TCA_NETEM_LATENCY64] = { .type = NLA_S64 }, 954 [TCA_NETEM_JITTER64] = { .type = NLA_S64 }, 955 [TCA_NETEM_SLOT] = { .len = sizeof(struct tc_netem_slot) }, 956 [TCA_NETEM_PRNG_SEED] = { .type = NLA_U64 }, 957 }; 958 959 static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla, 960 const struct nla_policy *policy, int len) 961 { 962 int nested_len = nla_len(nla) - NLA_ALIGN(len); 963 964 if (nested_len < 0) { 965 pr_info("netem: invalid attributes len %d\n", nested_len); 966 return -EINVAL; 967 } 968 969 if (nested_len >= nla_attr_size(0)) 970 return nla_parse_deprecated(tb, maxtype, 971 nla_data(nla) + NLA_ALIGN(len), 972 nested_len, policy, NULL); 973 974 memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1)); 975 return 0; 976 } 977 978 static const struct Qdisc_class_ops netem_class_ops; 979 980 static int check_netem_in_tree(struct Qdisc *sch, bool duplicates, 981 struct netlink_ext_ack *extack) 982 { 983 struct Qdisc *root, *q; 984 unsigned int i; 985 986 root = qdisc_root_sleeping(sch); 987 988 if (sch != root && root->ops->cl_ops == &netem_class_ops) { 989 if (duplicates || 990 ((struct netem_sched_data *)qdisc_priv(root))->duplicate) 991 goto err; 992 } 993 994 if (!qdisc_dev(root)) 995 return 0; 996 997 hash_for_each(qdisc_dev(root)->qdisc_hash, i, q, hash) { 998 if (sch != q && q->ops->cl_ops == &netem_class_ops) { 999 if (duplicates || 1000 ((struct netem_sched_data *)qdisc_priv(q))->duplicate) 1001 goto err; 1002 } 1003 } 1004 1005 return 0; 1006 1007 err: 1008 NL_SET_ERR_MSG(extack, 1009 "netem: cannot mix duplicating netems with other netems in tree"); 1010 return -EINVAL; 1011 } 1012 1013 /* Parse netlink message to set options */ 1014 static int netem_change(struct Qdisc *sch, struct nlattr *opt, 1015 struct netlink_ext_ack *extack) 1016 { 1017 struct netem_sched_data *q = qdisc_priv(sch); 1018 struct nlattr *tb[TCA_NETEM_MAX + 1]; 1019 struct disttable *delay_dist = NULL; 1020 struct disttable *slot_dist = NULL; 1021 struct tc_netem_qopt *qopt; 1022 struct clgstate old_clg; 1023 int old_loss_model = CLG_RANDOM; 1024 int ret; 1025 1026 qopt = nla_data(opt); 1027 ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt)); 1028 if (ret < 0) 1029 return ret; 1030 1031 if (tb[TCA_NETEM_DELAY_DIST]) { 1032 ret = get_dist_table(&delay_dist, tb[TCA_NETEM_DELAY_DIST]); 1033 if (ret) 1034 goto table_free; 1035 } 1036 1037 if (tb[TCA_NETEM_SLOT_DIST]) { 1038 ret = get_dist_table(&slot_dist, tb[TCA_NETEM_SLOT_DIST]); 1039 if (ret) 1040 goto table_free; 1041 } 1042 1043 sch_tree_lock(sch); 1044 /* backup q->clg and q->loss_model */ 1045 old_clg = q->clg; 1046 old_loss_model = q->loss_model; 1047 1048 if (tb[TCA_NETEM_LOSS]) { 1049 ret = get_loss_clg(q, tb[TCA_NETEM_LOSS]); 1050 if (ret) { 1051 q->loss_model = old_loss_model; 1052 q->clg = old_clg; 1053 goto unlock; 1054 } 1055 } else { 1056 q->loss_model = CLG_RANDOM; 1057 } 1058 1059 if (delay_dist) 1060 swap(q->delay_dist, delay_dist); 1061 if (slot_dist) 1062 swap(q->slot_dist, slot_dist); 1063 sch->limit = qopt->limit; 1064 1065 q->latency = PSCHED_TICKS2NS(qopt->latency); 1066 q->jitter = PSCHED_TICKS2NS(qopt->jitter); 1067 q->limit = qopt->limit; 1068 q->gap = qopt->gap; 1069 q->counter = 0; 1070 q->loss = qopt->loss; 1071 1072 ret = check_netem_in_tree(sch, qopt->duplicate, extack); 1073 if (ret) 1074 goto unlock; 1075 1076 q->duplicate = qopt->duplicate; 1077 1078 /* for compatibility with earlier versions. 1079 * if gap is set, need to assume 100% probability 1080 */ 1081 if (q->gap) 1082 q->reorder = ~0; 1083 1084 if (tb[TCA_NETEM_CORR]) 1085 get_correlation(q, tb[TCA_NETEM_CORR]); 1086 1087 if (tb[TCA_NETEM_REORDER]) 1088 get_reorder(q, tb[TCA_NETEM_REORDER]); 1089 1090 if (tb[TCA_NETEM_CORRUPT]) 1091 get_corrupt(q, tb[TCA_NETEM_CORRUPT]); 1092 1093 if (tb[TCA_NETEM_RATE]) 1094 get_rate(q, tb[TCA_NETEM_RATE]); 1095 1096 if (tb[TCA_NETEM_RATE64]) 1097 q->rate = max_t(u64, q->rate, 1098 nla_get_u64(tb[TCA_NETEM_RATE64])); 1099 1100 if (tb[TCA_NETEM_LATENCY64]) 1101 q->latency = nla_get_s64(tb[TCA_NETEM_LATENCY64]); 1102 1103 if (tb[TCA_NETEM_JITTER64]) 1104 q->jitter = nla_get_s64(tb[TCA_NETEM_JITTER64]); 1105 1106 if (tb[TCA_NETEM_ECN]) 1107 q->ecn = nla_get_u32(tb[TCA_NETEM_ECN]); 1108 1109 if (tb[TCA_NETEM_SLOT]) 1110 get_slot(q, tb[TCA_NETEM_SLOT]); 1111 1112 /* capping jitter to the range acceptable by tabledist() */ 1113 q->jitter = min_t(s64, abs(q->jitter), INT_MAX); 1114 1115 if (tb[TCA_NETEM_PRNG_SEED]) 1116 q->prng.seed = nla_get_u64(tb[TCA_NETEM_PRNG_SEED]); 1117 else 1118 q->prng.seed = get_random_u64(); 1119 prandom_seed_state(&q->prng.prng_state, q->prng.seed); 1120 1121 unlock: 1122 sch_tree_unlock(sch); 1123 1124 table_free: 1125 dist_free(delay_dist); 1126 dist_free(slot_dist); 1127 return ret; 1128 } 1129 1130 static int netem_init(struct Qdisc *sch, struct nlattr *opt, 1131 struct netlink_ext_ack *extack) 1132 { 1133 struct netem_sched_data *q = qdisc_priv(sch); 1134 int ret; 1135 1136 qdisc_watchdog_init(&q->watchdog, sch); 1137 1138 if (!opt) 1139 return -EINVAL; 1140 1141 q->loss_model = CLG_RANDOM; 1142 ret = netem_change(sch, opt, extack); 1143 if (ret) 1144 pr_info("netem: change failed\n"); 1145 return ret; 1146 } 1147 1148 static void netem_destroy(struct Qdisc *sch) 1149 { 1150 struct netem_sched_data *q = qdisc_priv(sch); 1151 1152 qdisc_watchdog_cancel(&q->watchdog); 1153 if (q->qdisc) 1154 qdisc_put(q->qdisc); 1155 dist_free(q->delay_dist); 1156 dist_free(q->slot_dist); 1157 } 1158 1159 static int dump_loss_model(const struct netem_sched_data *q, 1160 struct sk_buff *skb) 1161 { 1162 struct nlattr *nest; 1163 1164 nest = nla_nest_start_noflag(skb, TCA_NETEM_LOSS); 1165 if (nest == NULL) 1166 goto nla_put_failure; 1167 1168 switch (q->loss_model) { 1169 case CLG_RANDOM: 1170 /* legacy loss model */ 1171 nla_nest_cancel(skb, nest); 1172 return 0; /* no data */ 1173 1174 case CLG_4_STATES: { 1175 struct tc_netem_gimodel gi = { 1176 .p13 = q->clg.a1, 1177 .p31 = q->clg.a2, 1178 .p32 = q->clg.a3, 1179 .p14 = q->clg.a4, 1180 .p23 = q->clg.a5, 1181 }; 1182 1183 if (nla_put(skb, NETEM_LOSS_GI, sizeof(gi), &gi)) 1184 goto nla_put_failure; 1185 break; 1186 } 1187 case CLG_GILB_ELL: { 1188 struct tc_netem_gemodel ge = { 1189 .p = q->clg.a1, 1190 .r = q->clg.a2, 1191 .h = q->clg.a3, 1192 .k1 = q->clg.a4, 1193 }; 1194 1195 if (nla_put(skb, NETEM_LOSS_GE, sizeof(ge), &ge)) 1196 goto nla_put_failure; 1197 break; 1198 } 1199 } 1200 1201 nla_nest_end(skb, nest); 1202 return 0; 1203 1204 nla_put_failure: 1205 nla_nest_cancel(skb, nest); 1206 return -1; 1207 } 1208 1209 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb) 1210 { 1211 const struct netem_sched_data *q = qdisc_priv(sch); 1212 struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb); 1213 struct tc_netem_qopt qopt; 1214 struct tc_netem_corr cor; 1215 struct tc_netem_reorder reorder; 1216 struct tc_netem_corrupt corrupt; 1217 struct tc_netem_rate rate; 1218 struct tc_netem_slot slot; 1219 1220 qopt.latency = min_t(psched_time_t, PSCHED_NS2TICKS(q->latency), 1221 UINT_MAX); 1222 qopt.jitter = min_t(psched_time_t, PSCHED_NS2TICKS(q->jitter), 1223 UINT_MAX); 1224 qopt.limit = q->limit; 1225 qopt.loss = q->loss; 1226 qopt.gap = q->gap; 1227 qopt.duplicate = q->duplicate; 1228 if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt)) 1229 goto nla_put_failure; 1230 1231 if (nla_put(skb, TCA_NETEM_LATENCY64, sizeof(q->latency), &q->latency)) 1232 goto nla_put_failure; 1233 1234 if (nla_put(skb, TCA_NETEM_JITTER64, sizeof(q->jitter), &q->jitter)) 1235 goto nla_put_failure; 1236 1237 cor.delay_corr = q->delay_cor.rho; 1238 cor.loss_corr = q->loss_cor.rho; 1239 cor.dup_corr = q->dup_cor.rho; 1240 if (nla_put(skb, TCA_NETEM_CORR, sizeof(cor), &cor)) 1241 goto nla_put_failure; 1242 1243 reorder.probability = q->reorder; 1244 reorder.correlation = q->reorder_cor.rho; 1245 if (nla_put(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder)) 1246 goto nla_put_failure; 1247 1248 corrupt.probability = q->corrupt; 1249 corrupt.correlation = q->corrupt_cor.rho; 1250 if (nla_put(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt)) 1251 goto nla_put_failure; 1252 1253 if (q->rate >= (1ULL << 32)) { 1254 if (nla_put_u64_64bit(skb, TCA_NETEM_RATE64, q->rate, 1255 TCA_NETEM_PAD)) 1256 goto nla_put_failure; 1257 rate.rate = ~0U; 1258 } else { 1259 rate.rate = q->rate; 1260 } 1261 rate.packet_overhead = q->packet_overhead; 1262 rate.cell_size = q->cell_size; 1263 rate.cell_overhead = q->cell_overhead; 1264 if (nla_put(skb, TCA_NETEM_RATE, sizeof(rate), &rate)) 1265 goto nla_put_failure; 1266 1267 if (q->ecn && nla_put_u32(skb, TCA_NETEM_ECN, q->ecn)) 1268 goto nla_put_failure; 1269 1270 if (dump_loss_model(q, skb) != 0) 1271 goto nla_put_failure; 1272 1273 if (q->slot_config.min_delay | q->slot_config.max_delay | 1274 q->slot_config.dist_jitter) { 1275 slot = q->slot_config; 1276 if (slot.max_packets == INT_MAX) 1277 slot.max_packets = 0; 1278 if (slot.max_bytes == INT_MAX) 1279 slot.max_bytes = 0; 1280 if (nla_put(skb, TCA_NETEM_SLOT, sizeof(slot), &slot)) 1281 goto nla_put_failure; 1282 } 1283 1284 if (nla_put_u64_64bit(skb, TCA_NETEM_PRNG_SEED, q->prng.seed, 1285 TCA_NETEM_PAD)) 1286 goto nla_put_failure; 1287 1288 return nla_nest_end(skb, nla); 1289 1290 nla_put_failure: 1291 nlmsg_trim(skb, nla); 1292 return -1; 1293 } 1294 1295 static int netem_dump_class(struct Qdisc *sch, unsigned long cl, 1296 struct sk_buff *skb, struct tcmsg *tcm) 1297 { 1298 struct netem_sched_data *q = qdisc_priv(sch); 1299 1300 if (cl != 1 || !q->qdisc) /* only one class */ 1301 return -ENOENT; 1302 1303 tcm->tcm_handle |= TC_H_MIN(1); 1304 tcm->tcm_info = q->qdisc->handle; 1305 1306 return 0; 1307 } 1308 1309 static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new, 1310 struct Qdisc **old, struct netlink_ext_ack *extack) 1311 { 1312 struct netem_sched_data *q = qdisc_priv(sch); 1313 1314 *old = qdisc_replace(sch, new, &q->qdisc); 1315 return 0; 1316 } 1317 1318 static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg) 1319 { 1320 struct netem_sched_data *q = qdisc_priv(sch); 1321 return q->qdisc; 1322 } 1323 1324 static unsigned long netem_find(struct Qdisc *sch, u32 classid) 1325 { 1326 return 1; 1327 } 1328 1329 static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker) 1330 { 1331 if (!walker->stop) { 1332 if (!tc_qdisc_stats_dump(sch, 1, walker)) 1333 return; 1334 } 1335 } 1336 1337 static const struct Qdisc_class_ops netem_class_ops = { 1338 .graft = netem_graft, 1339 .leaf = netem_leaf, 1340 .find = netem_find, 1341 .walk = netem_walk, 1342 .dump = netem_dump_class, 1343 }; 1344 1345 static struct Qdisc_ops netem_qdisc_ops __read_mostly = { 1346 .id = "netem", 1347 .cl_ops = &netem_class_ops, 1348 .priv_size = sizeof(struct netem_sched_data), 1349 .enqueue = netem_enqueue, 1350 .dequeue = netem_dequeue, 1351 .peek = qdisc_peek_dequeued, 1352 .init = netem_init, 1353 .reset = netem_reset, 1354 .destroy = netem_destroy, 1355 .change = netem_change, 1356 .dump = netem_dump, 1357 .owner = THIS_MODULE, 1358 }; 1359 MODULE_ALIAS_NET_SCH("netem"); 1360 1361 1362 static int __init netem_module_init(void) 1363 { 1364 pr_info("netem: version " VERSION "\n"); 1365 return register_qdisc(&netem_qdisc_ops); 1366 } 1367 static void __exit netem_module_exit(void) 1368 { 1369 unregister_qdisc(&netem_qdisc_ops); 1370 } 1371 module_init(netem_module_init) 1372 module_exit(netem_module_exit) 1373 MODULE_LICENSE("GPL"); 1374 MODULE_DESCRIPTION("Network characteristics emulator qdisc"); 1375