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/netlink.h> 25 #include <net/pkt_sched.h> 26 #include <net/inet_ecn.h> 27 28 #define VERSION "1.3" 29 30 /* Network Emulation Queuing algorithm. 31 ==================================== 32 33 Sources: [1] Mark Carson, Darrin Santay, "NIST Net - A Linux-based 34 Network Emulation Tool 35 [2] Luigi Rizzo, DummyNet for FreeBSD 36 37 ---------------------------------------------------------------- 38 39 This started out as a simple way to delay outgoing packets to 40 test TCP but has grown to include most of the functionality 41 of a full blown network emulator like NISTnet. It can delay 42 packets and add random jitter (and correlation). The random 43 distribution can be loaded from a table as well to provide 44 normal, Pareto, or experimental curves. Packet loss, 45 duplication, and reordering can also be emulated. 46 47 This qdisc does not do classification that can be handled in 48 layering other disciplines. It does not need to do bandwidth 49 control either since that can be handled by using token 50 bucket or other rate control. 51 52 Correlated Loss Generator models 53 54 Added generation of correlated loss according to the 55 "Gilbert-Elliot" model, a 4-state markov model. 56 57 References: 58 [1] NetemCLG Home http://netgroup.uniroma2.it/NetemCLG 59 [2] S. Salsano, F. Ludovici, A. Ordine, "Definition of a general 60 and intuitive loss model for packet networks and its implementation 61 in the Netem module in the Linux kernel", available in [1] 62 63 Authors: Stefano Salsano <stefano.salsano at uniroma2.it 64 Fabio Ludovici <fabio.ludovici at yahoo.it> 65 */ 66 67 struct disttable { 68 u32 size; 69 s16 table[0]; 70 }; 71 72 struct netem_sched_data { 73 /* internal t(ime)fifo qdisc uses t_root and sch->limit */ 74 struct rb_root t_root; 75 76 /* a linear queue; reduces rbtree rebalancing when jitter is low */ 77 struct sk_buff *t_head; 78 struct sk_buff *t_tail; 79 80 /* optional qdisc for classful handling (NULL at netem init) */ 81 struct Qdisc *qdisc; 82 83 struct qdisc_watchdog watchdog; 84 85 s64 latency; 86 s64 jitter; 87 88 u32 loss; 89 u32 ecn; 90 u32 limit; 91 u32 counter; 92 u32 gap; 93 u32 duplicate; 94 u32 reorder; 95 u32 corrupt; 96 u64 rate; 97 s32 packet_overhead; 98 u32 cell_size; 99 struct reciprocal_value cell_size_reciprocal; 100 s32 cell_overhead; 101 102 struct crndstate { 103 u32 last; 104 u32 rho; 105 } delay_cor, loss_cor, dup_cor, reorder_cor, corrupt_cor; 106 107 struct disttable *delay_dist; 108 109 enum { 110 CLG_RANDOM, 111 CLG_4_STATES, 112 CLG_GILB_ELL, 113 } loss_model; 114 115 enum { 116 TX_IN_GAP_PERIOD = 1, 117 TX_IN_BURST_PERIOD, 118 LOST_IN_GAP_PERIOD, 119 LOST_IN_BURST_PERIOD, 120 } _4_state_model; 121 122 enum { 123 GOOD_STATE = 1, 124 BAD_STATE, 125 } GE_state_model; 126 127 /* Correlated Loss Generation models */ 128 struct clgstate { 129 /* state of the Markov chain */ 130 u8 state; 131 132 /* 4-states and Gilbert-Elliot models */ 133 u32 a1; /* p13 for 4-states or p for GE */ 134 u32 a2; /* p31 for 4-states or r for GE */ 135 u32 a3; /* p32 for 4-states or h for GE */ 136 u32 a4; /* p14 for 4-states or 1-k for GE */ 137 u32 a5; /* p23 used only in 4-states */ 138 } clg; 139 140 struct tc_netem_slot slot_config; 141 struct slotstate { 142 u64 slot_next; 143 s32 packets_left; 144 s32 bytes_left; 145 } slot; 146 147 struct disttable *slot_dist; 148 }; 149 150 /* Time stamp put into socket buffer control block 151 * Only valid when skbs are in our internal t(ime)fifo queue. 152 * 153 * As skb->rbnode uses same storage than skb->next, skb->prev and skb->tstamp, 154 * and skb->next & skb->prev are scratch space for a qdisc, 155 * we save skb->tstamp value in skb->cb[] before destroying it. 156 */ 157 struct netem_skb_cb { 158 u64 time_to_send; 159 }; 160 161 static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb) 162 { 163 /* we assume we can use skb next/prev/tstamp as storage for rb_node */ 164 qdisc_cb_private_validate(skb, sizeof(struct netem_skb_cb)); 165 return (struct netem_skb_cb *)qdisc_skb_cb(skb)->data; 166 } 167 168 /* init_crandom - initialize correlated random number generator 169 * Use entropy source for initial seed. 170 */ 171 static void init_crandom(struct crndstate *state, unsigned long rho) 172 { 173 state->rho = rho; 174 state->last = prandom_u32(); 175 } 176 177 /* get_crandom - correlated random number generator 178 * Next number depends on last value. 179 * rho is scaled to avoid floating point. 180 */ 181 static u32 get_crandom(struct crndstate *state) 182 { 183 u64 value, rho; 184 unsigned long answer; 185 186 if (!state || state->rho == 0) /* no correlation */ 187 return prandom_u32(); 188 189 value = prandom_u32(); 190 rho = (u64)state->rho + 1; 191 answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32; 192 state->last = answer; 193 return answer; 194 } 195 196 /* loss_4state - 4-state model loss generator 197 * Generates losses according to the 4-state Markov chain adopted in 198 * the GI (General and Intuitive) loss model. 199 */ 200 static bool loss_4state(struct netem_sched_data *q) 201 { 202 struct clgstate *clg = &q->clg; 203 u32 rnd = prandom_u32(); 204 205 /* 206 * Makes a comparison between rnd and the transition 207 * probabilities outgoing from the current state, then decides the 208 * next state and if the next packet has to be transmitted or lost. 209 * The four states correspond to: 210 * TX_IN_GAP_PERIOD => successfully transmitted packets within a gap period 211 * LOST_IN_BURST_PERIOD => isolated losses within a gap period 212 * LOST_IN_GAP_PERIOD => lost packets within a burst period 213 * TX_IN_GAP_PERIOD => successfully transmitted packets within a burst period 214 */ 215 switch (clg->state) { 216 case TX_IN_GAP_PERIOD: 217 if (rnd < clg->a4) { 218 clg->state = LOST_IN_BURST_PERIOD; 219 return true; 220 } else if (clg->a4 < rnd && rnd < clg->a1 + clg->a4) { 221 clg->state = LOST_IN_GAP_PERIOD; 222 return true; 223 } else if (clg->a1 + clg->a4 < rnd) { 224 clg->state = TX_IN_GAP_PERIOD; 225 } 226 227 break; 228 case TX_IN_BURST_PERIOD: 229 if (rnd < clg->a5) { 230 clg->state = LOST_IN_GAP_PERIOD; 231 return true; 232 } else { 233 clg->state = TX_IN_BURST_PERIOD; 234 } 235 236 break; 237 case LOST_IN_GAP_PERIOD: 238 if (rnd < clg->a3) 239 clg->state = TX_IN_BURST_PERIOD; 240 else if (clg->a3 < rnd && rnd < clg->a2 + clg->a3) { 241 clg->state = TX_IN_GAP_PERIOD; 242 } else if (clg->a2 + clg->a3 < rnd) { 243 clg->state = LOST_IN_GAP_PERIOD; 244 return true; 245 } 246 break; 247 case LOST_IN_BURST_PERIOD: 248 clg->state = TX_IN_GAP_PERIOD; 249 break; 250 } 251 252 return false; 253 } 254 255 /* loss_gilb_ell - Gilbert-Elliot model loss generator 256 * Generates losses according to the Gilbert-Elliot loss model or 257 * its special cases (Gilbert or Simple Gilbert) 258 * 259 * Makes a comparison between random number and the transition 260 * probabilities outgoing from the current state, then decides the 261 * next state. A second random number is extracted and the comparison 262 * with the loss probability of the current state decides if the next 263 * packet will be transmitted or lost. 264 */ 265 static bool loss_gilb_ell(struct netem_sched_data *q) 266 { 267 struct clgstate *clg = &q->clg; 268 269 switch (clg->state) { 270 case GOOD_STATE: 271 if (prandom_u32() < clg->a1) 272 clg->state = BAD_STATE; 273 if (prandom_u32() < clg->a4) 274 return true; 275 break; 276 case BAD_STATE: 277 if (prandom_u32() < clg->a2) 278 clg->state = GOOD_STATE; 279 if (prandom_u32() > clg->a3) 280 return true; 281 } 282 283 return false; 284 } 285 286 static bool loss_event(struct netem_sched_data *q) 287 { 288 switch (q->loss_model) { 289 case CLG_RANDOM: 290 /* Random packet drop 0 => none, ~0 => all */ 291 return q->loss && q->loss >= get_crandom(&q->loss_cor); 292 293 case CLG_4_STATES: 294 /* 4state loss model algorithm (used also for GI model) 295 * Extracts a value from the markov 4 state loss generator, 296 * if it is 1 drops a packet and if needed writes the event in 297 * the kernel logs 298 */ 299 return loss_4state(q); 300 301 case CLG_GILB_ELL: 302 /* Gilbert-Elliot loss model algorithm 303 * Extracts a value from the Gilbert-Elliot 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_gilb_ell(q); 308 } 309 310 return false; /* not reached */ 311 } 312 313 314 /* tabledist - return a pseudo-randomly distributed value with mean mu and 315 * std deviation sigma. Uses table lookup to approximate the desired 316 * distribution, and a uniformly-distributed pseudo-random source. 317 */ 318 static s64 tabledist(s64 mu, s32 sigma, 319 struct crndstate *state, 320 const struct disttable *dist) 321 { 322 s64 x; 323 long t; 324 u32 rnd; 325 326 if (sigma == 0) 327 return mu; 328 329 rnd = get_crandom(state); 330 331 /* default uniform distribution */ 332 if (dist == NULL) 333 return ((rnd % (2 * sigma)) + mu) - sigma; 334 335 t = dist->table[rnd % dist->size]; 336 x = (sigma % NETEM_DIST_SCALE) * t; 337 if (x >= 0) 338 x += NETEM_DIST_SCALE/2; 339 else 340 x -= NETEM_DIST_SCALE/2; 341 342 return x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu; 343 } 344 345 static u64 packet_time_ns(u64 len, const struct netem_sched_data *q) 346 { 347 len += q->packet_overhead; 348 349 if (q->cell_size) { 350 u32 cells = reciprocal_divide(len, q->cell_size_reciprocal); 351 352 if (len > cells * q->cell_size) /* extra cell needed for remainder */ 353 cells++; 354 len = cells * (q->cell_size + q->cell_overhead); 355 } 356 357 return div64_u64(len * NSEC_PER_SEC, q->rate); 358 } 359 360 static void tfifo_reset(struct Qdisc *sch) 361 { 362 struct netem_sched_data *q = qdisc_priv(sch); 363 struct rb_node *p = rb_first(&q->t_root); 364 365 while (p) { 366 struct sk_buff *skb = rb_to_skb(p); 367 368 p = rb_next(p); 369 rb_erase(&skb->rbnode, &q->t_root); 370 rtnl_kfree_skbs(skb, skb); 371 } 372 373 rtnl_kfree_skbs(q->t_head, q->t_tail); 374 q->t_head = NULL; 375 q->t_tail = NULL; 376 } 377 378 static void tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch) 379 { 380 struct netem_sched_data *q = qdisc_priv(sch); 381 u64 tnext = netem_skb_cb(nskb)->time_to_send; 382 383 if (!q->t_tail || tnext >= netem_skb_cb(q->t_tail)->time_to_send) { 384 if (q->t_tail) 385 q->t_tail->next = nskb; 386 else 387 q->t_head = nskb; 388 q->t_tail = nskb; 389 } else { 390 struct rb_node **p = &q->t_root.rb_node, *parent = NULL; 391 392 while (*p) { 393 struct sk_buff *skb; 394 395 parent = *p; 396 skb = rb_to_skb(parent); 397 if (tnext >= netem_skb_cb(skb)->time_to_send) 398 p = &parent->rb_right; 399 else 400 p = &parent->rb_left; 401 } 402 rb_link_node(&nskb->rbnode, parent, p); 403 rb_insert_color(&nskb->rbnode, &q->t_root); 404 } 405 sch->q.qlen++; 406 } 407 408 /* netem can't properly corrupt a megapacket (like we get from GSO), so instead 409 * when we statistically choose to corrupt one, we instead segment it, returning 410 * the first packet to be corrupted, and re-enqueue the remaining frames 411 */ 412 static struct sk_buff *netem_segment(struct sk_buff *skb, struct Qdisc *sch, 413 struct sk_buff **to_free) 414 { 415 struct sk_buff *segs; 416 netdev_features_t features = netif_skb_features(skb); 417 418 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK); 419 420 if (IS_ERR_OR_NULL(segs)) { 421 qdisc_drop(skb, sch, to_free); 422 return NULL; 423 } 424 consume_skb(skb); 425 return segs; 426 } 427 428 /* 429 * Insert one skb into qdisc. 430 * Note: parent depends on return value to account for queue length. 431 * NET_XMIT_DROP: queue length didn't change. 432 * NET_XMIT_SUCCESS: one skb was queued. 433 */ 434 static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch, 435 struct sk_buff **to_free) 436 { 437 struct netem_sched_data *q = qdisc_priv(sch); 438 /* We don't fill cb now as skb_unshare() may invalidate it */ 439 struct netem_skb_cb *cb; 440 struct sk_buff *skb2; 441 struct sk_buff *segs = NULL; 442 unsigned int prev_len = qdisc_pkt_len(skb); 443 int count = 1; 444 int rc = NET_XMIT_SUCCESS; 445 int rc_drop = NET_XMIT_DROP; 446 447 /* Do not fool qdisc_drop_all() */ 448 skb->prev = NULL; 449 450 /* Random duplication */ 451 if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor)) 452 ++count; 453 454 /* Drop packet? */ 455 if (loss_event(q)) { 456 if (q->ecn && INET_ECN_set_ce(skb)) 457 qdisc_qstats_drop(sch); /* mark packet */ 458 else 459 --count; 460 } 461 if (count == 0) { 462 qdisc_qstats_drop(sch); 463 __qdisc_drop(skb, to_free); 464 return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS; 465 } 466 467 /* If a delay is expected, orphan the skb. (orphaning usually takes 468 * place at TX completion time, so _before_ the link transit delay) 469 */ 470 if (q->latency || q->jitter || q->rate) 471 skb_orphan_partial(skb); 472 473 /* 474 * If we need to duplicate packet, then re-insert at top of the 475 * qdisc tree, since parent queuer expects that only one 476 * skb will be queued. 477 */ 478 if (count > 1 && (skb2 = skb_clone(skb, GFP_ATOMIC)) != NULL) { 479 struct Qdisc *rootq = qdisc_root_bh(sch); 480 u32 dupsave = q->duplicate; /* prevent duplicating a dup... */ 481 482 q->duplicate = 0; 483 rootq->enqueue(skb2, rootq, to_free); 484 q->duplicate = dupsave; 485 rc_drop = NET_XMIT_SUCCESS; 486 } 487 488 /* 489 * Randomized packet corruption. 490 * Make copy if needed since we are modifying 491 * If packet is going to be hardware checksummed, then 492 * do it now in software before we mangle it. 493 */ 494 if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor)) { 495 if (skb_is_gso(skb)) { 496 skb = netem_segment(skb, sch, to_free); 497 if (!skb) 498 return rc_drop; 499 segs = skb->next; 500 skb_mark_not_on_list(skb); 501 qdisc_skb_cb(skb)->pkt_len = skb->len; 502 } 503 504 skb = skb_unshare(skb, GFP_ATOMIC); 505 if (unlikely(!skb)) { 506 qdisc_qstats_drop(sch); 507 goto finish_segs; 508 } 509 if (skb->ip_summed == CHECKSUM_PARTIAL && 510 skb_checksum_help(skb)) { 511 qdisc_drop(skb, sch, to_free); 512 goto finish_segs; 513 } 514 515 skb->data[prandom_u32() % skb_headlen(skb)] ^= 516 1<<(prandom_u32() % 8); 517 } 518 519 if (unlikely(sch->q.qlen >= sch->limit)) { 520 /* re-link segs, so that qdisc_drop_all() frees them all */ 521 skb->next = segs; 522 qdisc_drop_all(skb, sch, to_free); 523 return rc_drop; 524 } 525 526 qdisc_qstats_backlog_inc(sch, skb); 527 528 cb = netem_skb_cb(skb); 529 if (q->gap == 0 || /* not doing reordering */ 530 q->counter < q->gap - 1 || /* inside last reordering gap */ 531 q->reorder < get_crandom(&q->reorder_cor)) { 532 u64 now; 533 s64 delay; 534 535 delay = tabledist(q->latency, q->jitter, 536 &q->delay_cor, q->delay_dist); 537 538 now = ktime_get_ns(); 539 540 if (q->rate) { 541 struct netem_skb_cb *last = NULL; 542 543 if (sch->q.tail) 544 last = netem_skb_cb(sch->q.tail); 545 if (q->t_root.rb_node) { 546 struct sk_buff *t_skb; 547 struct netem_skb_cb *t_last; 548 549 t_skb = skb_rb_last(&q->t_root); 550 t_last = netem_skb_cb(t_skb); 551 if (!last || 552 t_last->time_to_send > last->time_to_send) 553 last = t_last; 554 } 555 if (q->t_tail) { 556 struct netem_skb_cb *t_last = 557 netem_skb_cb(q->t_tail); 558 559 if (!last || 560 t_last->time_to_send > last->time_to_send) 561 last = t_last; 562 } 563 564 if (last) { 565 /* 566 * Last packet in queue is reference point (now), 567 * calculate this time bonus and subtract 568 * from delay. 569 */ 570 delay -= last->time_to_send - now; 571 delay = max_t(s64, 0, delay); 572 now = last->time_to_send; 573 } 574 575 delay += packet_time_ns(qdisc_pkt_len(skb), q); 576 } 577 578 cb->time_to_send = now + delay; 579 ++q->counter; 580 tfifo_enqueue(skb, sch); 581 } else { 582 /* 583 * Do re-ordering by putting one out of N packets at the front 584 * of the queue. 585 */ 586 cb->time_to_send = ktime_get_ns(); 587 q->counter = 0; 588 589 __qdisc_enqueue_head(skb, &sch->q); 590 sch->qstats.requeues++; 591 } 592 593 finish_segs: 594 if (segs) { 595 unsigned int len, last_len; 596 int nb = 0; 597 598 len = skb->len; 599 600 while (segs) { 601 skb2 = segs->next; 602 skb_mark_not_on_list(segs); 603 qdisc_skb_cb(segs)->pkt_len = segs->len; 604 last_len = segs->len; 605 rc = qdisc_enqueue(segs, sch, to_free); 606 if (rc != NET_XMIT_SUCCESS) { 607 if (net_xmit_drop_count(rc)) 608 qdisc_qstats_drop(sch); 609 } else { 610 nb++; 611 len += last_len; 612 } 613 segs = skb2; 614 } 615 qdisc_tree_reduce_backlog(sch, -nb, prev_len - len); 616 } 617 return NET_XMIT_SUCCESS; 618 } 619 620 /* Delay the next round with a new future slot with a 621 * correct number of bytes and packets. 622 */ 623 624 static void get_slot_next(struct netem_sched_data *q, u64 now) 625 { 626 s64 next_delay; 627 628 if (!q->slot_dist) 629 next_delay = q->slot_config.min_delay + 630 (prandom_u32() * 631 (q->slot_config.max_delay - 632 q->slot_config.min_delay) >> 32); 633 else 634 next_delay = tabledist(q->slot_config.dist_delay, 635 (s32)(q->slot_config.dist_jitter), 636 NULL, q->slot_dist); 637 638 q->slot.slot_next = now + next_delay; 639 q->slot.packets_left = q->slot_config.max_packets; 640 q->slot.bytes_left = q->slot_config.max_bytes; 641 } 642 643 static struct sk_buff *netem_peek(struct netem_sched_data *q) 644 { 645 struct sk_buff *skb = skb_rb_first(&q->t_root); 646 u64 t1, t2; 647 648 if (!skb) 649 return q->t_head; 650 if (!q->t_head) 651 return skb; 652 653 t1 = netem_skb_cb(skb)->time_to_send; 654 t2 = netem_skb_cb(q->t_head)->time_to_send; 655 if (t1 < t2) 656 return skb; 657 return q->t_head; 658 } 659 660 static void netem_erase_head(struct netem_sched_data *q, struct sk_buff *skb) 661 { 662 if (skb == q->t_head) { 663 q->t_head = skb->next; 664 if (!q->t_head) 665 q->t_tail = NULL; 666 } else { 667 rb_erase(&skb->rbnode, &q->t_root); 668 } 669 } 670 671 static struct sk_buff *netem_dequeue(struct Qdisc *sch) 672 { 673 struct netem_sched_data *q = qdisc_priv(sch); 674 struct sk_buff *skb; 675 676 tfifo_dequeue: 677 skb = __qdisc_dequeue_head(&sch->q); 678 if (skb) { 679 qdisc_qstats_backlog_dec(sch, skb); 680 deliver: 681 qdisc_bstats_update(sch, skb); 682 return skb; 683 } 684 skb = netem_peek(q); 685 if (skb) { 686 u64 time_to_send; 687 u64 now = ktime_get_ns(); 688 689 /* if more time remaining? */ 690 time_to_send = netem_skb_cb(skb)->time_to_send; 691 if (q->slot.slot_next && q->slot.slot_next < time_to_send) 692 get_slot_next(q, now); 693 694 if (time_to_send <= now && q->slot.slot_next <= now) { 695 netem_erase_head(q, skb); 696 sch->q.qlen--; 697 qdisc_qstats_backlog_dec(sch, skb); 698 skb->next = NULL; 699 skb->prev = NULL; 700 /* skb->dev shares skb->rbnode area, 701 * we need to restore its value. 702 */ 703 skb->dev = qdisc_dev(sch); 704 705 if (q->slot.slot_next) { 706 q->slot.packets_left--; 707 q->slot.bytes_left -= qdisc_pkt_len(skb); 708 if (q->slot.packets_left <= 0 || 709 q->slot.bytes_left <= 0) 710 get_slot_next(q, now); 711 } 712 713 if (q->qdisc) { 714 unsigned int pkt_len = qdisc_pkt_len(skb); 715 struct sk_buff *to_free = NULL; 716 int err; 717 718 err = qdisc_enqueue(skb, q->qdisc, &to_free); 719 kfree_skb_list(to_free); 720 if (err != NET_XMIT_SUCCESS && 721 net_xmit_drop_count(err)) { 722 qdisc_qstats_drop(sch); 723 qdisc_tree_reduce_backlog(sch, 1, 724 pkt_len); 725 } 726 goto tfifo_dequeue; 727 } 728 goto deliver; 729 } 730 731 if (q->qdisc) { 732 skb = q->qdisc->ops->dequeue(q->qdisc); 733 if (skb) 734 goto deliver; 735 } 736 737 qdisc_watchdog_schedule_ns(&q->watchdog, 738 max(time_to_send, 739 q->slot.slot_next)); 740 } 741 742 if (q->qdisc) { 743 skb = q->qdisc->ops->dequeue(q->qdisc); 744 if (skb) 745 goto deliver; 746 } 747 return NULL; 748 } 749 750 static void netem_reset(struct Qdisc *sch) 751 { 752 struct netem_sched_data *q = qdisc_priv(sch); 753 754 qdisc_reset_queue(sch); 755 tfifo_reset(sch); 756 if (q->qdisc) 757 qdisc_reset(q->qdisc); 758 qdisc_watchdog_cancel(&q->watchdog); 759 } 760 761 static void dist_free(struct disttable *d) 762 { 763 kvfree(d); 764 } 765 766 /* 767 * Distribution data is a variable size payload containing 768 * signed 16 bit values. 769 */ 770 771 static int get_dist_table(struct Qdisc *sch, struct disttable **tbl, 772 const struct nlattr *attr) 773 { 774 size_t n = nla_len(attr)/sizeof(__s16); 775 const __s16 *data = nla_data(attr); 776 spinlock_t *root_lock; 777 struct disttable *d; 778 int i; 779 780 if (!n || n > NETEM_DIST_MAX) 781 return -EINVAL; 782 783 d = kvmalloc(sizeof(struct disttable) + n * sizeof(s16), GFP_KERNEL); 784 if (!d) 785 return -ENOMEM; 786 787 d->size = n; 788 for (i = 0; i < n; i++) 789 d->table[i] = data[i]; 790 791 root_lock = qdisc_root_sleeping_lock(sch); 792 793 spin_lock_bh(root_lock); 794 swap(*tbl, d); 795 spin_unlock_bh(root_lock); 796 797 dist_free(d); 798 return 0; 799 } 800 801 static void get_slot(struct netem_sched_data *q, const struct nlattr *attr) 802 { 803 const struct tc_netem_slot *c = nla_data(attr); 804 805 q->slot_config = *c; 806 if (q->slot_config.max_packets == 0) 807 q->slot_config.max_packets = INT_MAX; 808 if (q->slot_config.max_bytes == 0) 809 q->slot_config.max_bytes = INT_MAX; 810 q->slot.packets_left = q->slot_config.max_packets; 811 q->slot.bytes_left = q->slot_config.max_bytes; 812 if (q->slot_config.min_delay | q->slot_config.max_delay | 813 q->slot_config.dist_jitter) 814 q->slot.slot_next = ktime_get_ns(); 815 else 816 q->slot.slot_next = 0; 817 } 818 819 static void get_correlation(struct netem_sched_data *q, const struct nlattr *attr) 820 { 821 const struct tc_netem_corr *c = nla_data(attr); 822 823 init_crandom(&q->delay_cor, c->delay_corr); 824 init_crandom(&q->loss_cor, c->loss_corr); 825 init_crandom(&q->dup_cor, c->dup_corr); 826 } 827 828 static void get_reorder(struct netem_sched_data *q, const struct nlattr *attr) 829 { 830 const struct tc_netem_reorder *r = nla_data(attr); 831 832 q->reorder = r->probability; 833 init_crandom(&q->reorder_cor, r->correlation); 834 } 835 836 static void get_corrupt(struct netem_sched_data *q, const struct nlattr *attr) 837 { 838 const struct tc_netem_corrupt *r = nla_data(attr); 839 840 q->corrupt = r->probability; 841 init_crandom(&q->corrupt_cor, r->correlation); 842 } 843 844 static void get_rate(struct netem_sched_data *q, const struct nlattr *attr) 845 { 846 const struct tc_netem_rate *r = nla_data(attr); 847 848 q->rate = r->rate; 849 q->packet_overhead = r->packet_overhead; 850 q->cell_size = r->cell_size; 851 q->cell_overhead = r->cell_overhead; 852 if (q->cell_size) 853 q->cell_size_reciprocal = reciprocal_value(q->cell_size); 854 else 855 q->cell_size_reciprocal = (struct reciprocal_value) { 0 }; 856 } 857 858 static int get_loss_clg(struct netem_sched_data *q, const struct nlattr *attr) 859 { 860 const struct nlattr *la; 861 int rem; 862 863 nla_for_each_nested(la, attr, rem) { 864 u16 type = nla_type(la); 865 866 switch (type) { 867 case NETEM_LOSS_GI: { 868 const struct tc_netem_gimodel *gi = nla_data(la); 869 870 if (nla_len(la) < sizeof(struct tc_netem_gimodel)) { 871 pr_info("netem: incorrect gi model size\n"); 872 return -EINVAL; 873 } 874 875 q->loss_model = CLG_4_STATES; 876 877 q->clg.state = TX_IN_GAP_PERIOD; 878 q->clg.a1 = gi->p13; 879 q->clg.a2 = gi->p31; 880 q->clg.a3 = gi->p32; 881 q->clg.a4 = gi->p14; 882 q->clg.a5 = gi->p23; 883 break; 884 } 885 886 case NETEM_LOSS_GE: { 887 const struct tc_netem_gemodel *ge = nla_data(la); 888 889 if (nla_len(la) < sizeof(struct tc_netem_gemodel)) { 890 pr_info("netem: incorrect ge model size\n"); 891 return -EINVAL; 892 } 893 894 q->loss_model = CLG_GILB_ELL; 895 q->clg.state = GOOD_STATE; 896 q->clg.a1 = ge->p; 897 q->clg.a2 = ge->r; 898 q->clg.a3 = ge->h; 899 q->clg.a4 = ge->k1; 900 break; 901 } 902 903 default: 904 pr_info("netem: unknown loss type %u\n", type); 905 return -EINVAL; 906 } 907 } 908 909 return 0; 910 } 911 912 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = { 913 [TCA_NETEM_CORR] = { .len = sizeof(struct tc_netem_corr) }, 914 [TCA_NETEM_REORDER] = { .len = sizeof(struct tc_netem_reorder) }, 915 [TCA_NETEM_CORRUPT] = { .len = sizeof(struct tc_netem_corrupt) }, 916 [TCA_NETEM_RATE] = { .len = sizeof(struct tc_netem_rate) }, 917 [TCA_NETEM_LOSS] = { .type = NLA_NESTED }, 918 [TCA_NETEM_ECN] = { .type = NLA_U32 }, 919 [TCA_NETEM_RATE64] = { .type = NLA_U64 }, 920 [TCA_NETEM_LATENCY64] = { .type = NLA_S64 }, 921 [TCA_NETEM_JITTER64] = { .type = NLA_S64 }, 922 [TCA_NETEM_SLOT] = { .len = sizeof(struct tc_netem_slot) }, 923 }; 924 925 static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla, 926 const struct nla_policy *policy, int len) 927 { 928 int nested_len = nla_len(nla) - NLA_ALIGN(len); 929 930 if (nested_len < 0) { 931 pr_info("netem: invalid attributes len %d\n", nested_len); 932 return -EINVAL; 933 } 934 935 if (nested_len >= nla_attr_size(0)) 936 return nla_parse_deprecated(tb, maxtype, 937 nla_data(nla) + NLA_ALIGN(len), 938 nested_len, policy, NULL); 939 940 memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1)); 941 return 0; 942 } 943 944 /* Parse netlink message to set options */ 945 static int netem_change(struct Qdisc *sch, struct nlattr *opt, 946 struct netlink_ext_ack *extack) 947 { 948 struct netem_sched_data *q = qdisc_priv(sch); 949 struct nlattr *tb[TCA_NETEM_MAX + 1]; 950 struct tc_netem_qopt *qopt; 951 struct clgstate old_clg; 952 int old_loss_model = CLG_RANDOM; 953 int ret; 954 955 if (opt == NULL) 956 return -EINVAL; 957 958 qopt = nla_data(opt); 959 ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt)); 960 if (ret < 0) 961 return ret; 962 963 /* backup q->clg and q->loss_model */ 964 old_clg = q->clg; 965 old_loss_model = q->loss_model; 966 967 if (tb[TCA_NETEM_LOSS]) { 968 ret = get_loss_clg(q, tb[TCA_NETEM_LOSS]); 969 if (ret) { 970 q->loss_model = old_loss_model; 971 return ret; 972 } 973 } else { 974 q->loss_model = CLG_RANDOM; 975 } 976 977 if (tb[TCA_NETEM_DELAY_DIST]) { 978 ret = get_dist_table(sch, &q->delay_dist, 979 tb[TCA_NETEM_DELAY_DIST]); 980 if (ret) 981 goto get_table_failure; 982 } 983 984 if (tb[TCA_NETEM_SLOT_DIST]) { 985 ret = get_dist_table(sch, &q->slot_dist, 986 tb[TCA_NETEM_SLOT_DIST]); 987 if (ret) 988 goto get_table_failure; 989 } 990 991 sch->limit = qopt->limit; 992 993 q->latency = PSCHED_TICKS2NS(qopt->latency); 994 q->jitter = PSCHED_TICKS2NS(qopt->jitter); 995 q->limit = qopt->limit; 996 q->gap = qopt->gap; 997 q->counter = 0; 998 q->loss = qopt->loss; 999 q->duplicate = qopt->duplicate; 1000 1001 /* for compatibility with earlier versions. 1002 * if gap is set, need to assume 100% probability 1003 */ 1004 if (q->gap) 1005 q->reorder = ~0; 1006 1007 if (tb[TCA_NETEM_CORR]) 1008 get_correlation(q, tb[TCA_NETEM_CORR]); 1009 1010 if (tb[TCA_NETEM_REORDER]) 1011 get_reorder(q, tb[TCA_NETEM_REORDER]); 1012 1013 if (tb[TCA_NETEM_CORRUPT]) 1014 get_corrupt(q, tb[TCA_NETEM_CORRUPT]); 1015 1016 if (tb[TCA_NETEM_RATE]) 1017 get_rate(q, tb[TCA_NETEM_RATE]); 1018 1019 if (tb[TCA_NETEM_RATE64]) 1020 q->rate = max_t(u64, q->rate, 1021 nla_get_u64(tb[TCA_NETEM_RATE64])); 1022 1023 if (tb[TCA_NETEM_LATENCY64]) 1024 q->latency = nla_get_s64(tb[TCA_NETEM_LATENCY64]); 1025 1026 if (tb[TCA_NETEM_JITTER64]) 1027 q->jitter = nla_get_s64(tb[TCA_NETEM_JITTER64]); 1028 1029 if (tb[TCA_NETEM_ECN]) 1030 q->ecn = nla_get_u32(tb[TCA_NETEM_ECN]); 1031 1032 if (tb[TCA_NETEM_SLOT]) 1033 get_slot(q, tb[TCA_NETEM_SLOT]); 1034 1035 return ret; 1036 1037 get_table_failure: 1038 /* recover clg and loss_model, in case of 1039 * q->clg and q->loss_model were modified 1040 * in get_loss_clg() 1041 */ 1042 q->clg = old_clg; 1043 q->loss_model = old_loss_model; 1044 return ret; 1045 } 1046 1047 static int netem_init(struct Qdisc *sch, struct nlattr *opt, 1048 struct netlink_ext_ack *extack) 1049 { 1050 struct netem_sched_data *q = qdisc_priv(sch); 1051 int ret; 1052 1053 qdisc_watchdog_init(&q->watchdog, sch); 1054 1055 if (!opt) 1056 return -EINVAL; 1057 1058 q->loss_model = CLG_RANDOM; 1059 ret = netem_change(sch, opt, extack); 1060 if (ret) 1061 pr_info("netem: change failed\n"); 1062 return ret; 1063 } 1064 1065 static void netem_destroy(struct Qdisc *sch) 1066 { 1067 struct netem_sched_data *q = qdisc_priv(sch); 1068 1069 qdisc_watchdog_cancel(&q->watchdog); 1070 if (q->qdisc) 1071 qdisc_put(q->qdisc); 1072 dist_free(q->delay_dist); 1073 dist_free(q->slot_dist); 1074 } 1075 1076 static int dump_loss_model(const struct netem_sched_data *q, 1077 struct sk_buff *skb) 1078 { 1079 struct nlattr *nest; 1080 1081 nest = nla_nest_start_noflag(skb, TCA_NETEM_LOSS); 1082 if (nest == NULL) 1083 goto nla_put_failure; 1084 1085 switch (q->loss_model) { 1086 case CLG_RANDOM: 1087 /* legacy loss model */ 1088 nla_nest_cancel(skb, nest); 1089 return 0; /* no data */ 1090 1091 case CLG_4_STATES: { 1092 struct tc_netem_gimodel gi = { 1093 .p13 = q->clg.a1, 1094 .p31 = q->clg.a2, 1095 .p32 = q->clg.a3, 1096 .p14 = q->clg.a4, 1097 .p23 = q->clg.a5, 1098 }; 1099 1100 if (nla_put(skb, NETEM_LOSS_GI, sizeof(gi), &gi)) 1101 goto nla_put_failure; 1102 break; 1103 } 1104 case CLG_GILB_ELL: { 1105 struct tc_netem_gemodel ge = { 1106 .p = q->clg.a1, 1107 .r = q->clg.a2, 1108 .h = q->clg.a3, 1109 .k1 = q->clg.a4, 1110 }; 1111 1112 if (nla_put(skb, NETEM_LOSS_GE, sizeof(ge), &ge)) 1113 goto nla_put_failure; 1114 break; 1115 } 1116 } 1117 1118 nla_nest_end(skb, nest); 1119 return 0; 1120 1121 nla_put_failure: 1122 nla_nest_cancel(skb, nest); 1123 return -1; 1124 } 1125 1126 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb) 1127 { 1128 const struct netem_sched_data *q = qdisc_priv(sch); 1129 struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb); 1130 struct tc_netem_qopt qopt; 1131 struct tc_netem_corr cor; 1132 struct tc_netem_reorder reorder; 1133 struct tc_netem_corrupt corrupt; 1134 struct tc_netem_rate rate; 1135 struct tc_netem_slot slot; 1136 1137 qopt.latency = min_t(psched_tdiff_t, PSCHED_NS2TICKS(q->latency), 1138 UINT_MAX); 1139 qopt.jitter = min_t(psched_tdiff_t, PSCHED_NS2TICKS(q->jitter), 1140 UINT_MAX); 1141 qopt.limit = q->limit; 1142 qopt.loss = q->loss; 1143 qopt.gap = q->gap; 1144 qopt.duplicate = q->duplicate; 1145 if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt)) 1146 goto nla_put_failure; 1147 1148 if (nla_put(skb, TCA_NETEM_LATENCY64, sizeof(q->latency), &q->latency)) 1149 goto nla_put_failure; 1150 1151 if (nla_put(skb, TCA_NETEM_JITTER64, sizeof(q->jitter), &q->jitter)) 1152 goto nla_put_failure; 1153 1154 cor.delay_corr = q->delay_cor.rho; 1155 cor.loss_corr = q->loss_cor.rho; 1156 cor.dup_corr = q->dup_cor.rho; 1157 if (nla_put(skb, TCA_NETEM_CORR, sizeof(cor), &cor)) 1158 goto nla_put_failure; 1159 1160 reorder.probability = q->reorder; 1161 reorder.correlation = q->reorder_cor.rho; 1162 if (nla_put(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder)) 1163 goto nla_put_failure; 1164 1165 corrupt.probability = q->corrupt; 1166 corrupt.correlation = q->corrupt_cor.rho; 1167 if (nla_put(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt)) 1168 goto nla_put_failure; 1169 1170 if (q->rate >= (1ULL << 32)) { 1171 if (nla_put_u64_64bit(skb, TCA_NETEM_RATE64, q->rate, 1172 TCA_NETEM_PAD)) 1173 goto nla_put_failure; 1174 rate.rate = ~0U; 1175 } else { 1176 rate.rate = q->rate; 1177 } 1178 rate.packet_overhead = q->packet_overhead; 1179 rate.cell_size = q->cell_size; 1180 rate.cell_overhead = q->cell_overhead; 1181 if (nla_put(skb, TCA_NETEM_RATE, sizeof(rate), &rate)) 1182 goto nla_put_failure; 1183 1184 if (q->ecn && nla_put_u32(skb, TCA_NETEM_ECN, q->ecn)) 1185 goto nla_put_failure; 1186 1187 if (dump_loss_model(q, skb) != 0) 1188 goto nla_put_failure; 1189 1190 if (q->slot_config.min_delay | q->slot_config.max_delay | 1191 q->slot_config.dist_jitter) { 1192 slot = q->slot_config; 1193 if (slot.max_packets == INT_MAX) 1194 slot.max_packets = 0; 1195 if (slot.max_bytes == INT_MAX) 1196 slot.max_bytes = 0; 1197 if (nla_put(skb, TCA_NETEM_SLOT, sizeof(slot), &slot)) 1198 goto nla_put_failure; 1199 } 1200 1201 return nla_nest_end(skb, nla); 1202 1203 nla_put_failure: 1204 nlmsg_trim(skb, nla); 1205 return -1; 1206 } 1207 1208 static int netem_dump_class(struct Qdisc *sch, unsigned long cl, 1209 struct sk_buff *skb, struct tcmsg *tcm) 1210 { 1211 struct netem_sched_data *q = qdisc_priv(sch); 1212 1213 if (cl != 1 || !q->qdisc) /* only one class */ 1214 return -ENOENT; 1215 1216 tcm->tcm_handle |= TC_H_MIN(1); 1217 tcm->tcm_info = q->qdisc->handle; 1218 1219 return 0; 1220 } 1221 1222 static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new, 1223 struct Qdisc **old, struct netlink_ext_ack *extack) 1224 { 1225 struct netem_sched_data *q = qdisc_priv(sch); 1226 1227 *old = qdisc_replace(sch, new, &q->qdisc); 1228 return 0; 1229 } 1230 1231 static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg) 1232 { 1233 struct netem_sched_data *q = qdisc_priv(sch); 1234 return q->qdisc; 1235 } 1236 1237 static unsigned long netem_find(struct Qdisc *sch, u32 classid) 1238 { 1239 return 1; 1240 } 1241 1242 static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker) 1243 { 1244 if (!walker->stop) { 1245 if (walker->count >= walker->skip) 1246 if (walker->fn(sch, 1, walker) < 0) { 1247 walker->stop = 1; 1248 return; 1249 } 1250 walker->count++; 1251 } 1252 } 1253 1254 static const struct Qdisc_class_ops netem_class_ops = { 1255 .graft = netem_graft, 1256 .leaf = netem_leaf, 1257 .find = netem_find, 1258 .walk = netem_walk, 1259 .dump = netem_dump_class, 1260 }; 1261 1262 static struct Qdisc_ops netem_qdisc_ops __read_mostly = { 1263 .id = "netem", 1264 .cl_ops = &netem_class_ops, 1265 .priv_size = sizeof(struct netem_sched_data), 1266 .enqueue = netem_enqueue, 1267 .dequeue = netem_dequeue, 1268 .peek = qdisc_peek_dequeued, 1269 .init = netem_init, 1270 .reset = netem_reset, 1271 .destroy = netem_destroy, 1272 .change = netem_change, 1273 .dump = netem_dump, 1274 .owner = THIS_MODULE, 1275 }; 1276 1277 1278 static int __init netem_module_init(void) 1279 { 1280 pr_info("netem: version " VERSION "\n"); 1281 return register_qdisc(&netem_qdisc_ops); 1282 } 1283 static void __exit netem_module_exit(void) 1284 { 1285 unregister_qdisc(&netem_qdisc_ops); 1286 } 1287 module_init(netem_module_init) 1288 module_exit(netem_module_exit) 1289 MODULE_LICENSE("GPL"); 1290