Linux Audio

Check our new training course

Loading...
v6.13.7
   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
  69struct disttable {
  70	u32  size;
  71	s16 table[] __counted_by(size);
  72};
  73
  74struct 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 */
 166struct netem_skb_cb {
 167	u64	        time_to_send;
 168};
 169
 170static 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 */
 180static 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 */
 190static 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 */
 210static 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 */
 275static 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
 297static 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 */
 329static 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
 357static 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
 372static 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
 391static 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 */
 426static 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	segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
 433
 434	if (IS_ERR_OR_NULL(segs)) {
 435		qdisc_drop(skb, sch, to_free);
 436		return NULL;
 437	}
 438	consume_skb(skb);
 439	return segs;
 440}
 441
 442/*
 443 * Insert one skb into qdisc.
 444 * Note: parent depends on return value to account for queue length.
 445 * 	NET_XMIT_DROP: queue length didn't change.
 446 *      NET_XMIT_SUCCESS: one skb was queued.
 447 */
 448static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch,
 449			 struct sk_buff **to_free)
 450{
 451	struct netem_sched_data *q = qdisc_priv(sch);
 452	/* We don't fill cb now as skb_unshare() may invalidate it */
 453	struct netem_skb_cb *cb;
 454	struct sk_buff *skb2 = NULL;
 455	struct sk_buff *segs = NULL;
 456	unsigned int prev_len = qdisc_pkt_len(skb);
 457	int count = 1;
 458
 459	/* Do not fool qdisc_drop_all() */
 460	skb->prev = NULL;
 461
 462	/* Random duplication */
 463	if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor, &q->prng))
 464		++count;
 465
 466	/* Drop packet? */
 467	if (loss_event(q)) {
 468		if (q->ecn && INET_ECN_set_ce(skb))
 469			qdisc_qstats_drop(sch); /* mark packet */
 470		else
 471			--count;
 472	}
 473	if (count == 0) {
 474		qdisc_qstats_drop(sch);
 475		__qdisc_drop(skb, to_free);
 476		return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
 477	}
 478
 479	/* If a delay is expected, orphan the skb. (orphaning usually takes
 480	 * place at TX completion time, so _before_ the link transit delay)
 481	 */
 482	if (q->latency || q->jitter || q->rate)
 483		skb_orphan_partial(skb);
 484
 485	/*
 486	 * If we need to duplicate packet, then clone it before
 487	 * original is modified.
 
 488	 */
 489	if (count > 1)
 490		skb2 = skb_clone(skb, GFP_ATOMIC);
 
 
 
 
 
 
 491
 492	/*
 493	 * Randomized packet corruption.
 494	 * Make copy if needed since we are modifying
 495	 * If packet is going to be hardware checksummed, then
 496	 * do it now in software before we mangle it.
 497	 */
 498	if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor, &q->prng)) {
 499		if (skb_is_gso(skb)) {
 500			skb = netem_segment(skb, sch, to_free);
 501			if (!skb)
 502				goto finish_segs;
 503
 504			segs = skb->next;
 505			skb_mark_not_on_list(skb);
 506			qdisc_skb_cb(skb)->pkt_len = skb->len;
 507		}
 508
 509		skb = skb_unshare(skb, GFP_ATOMIC);
 510		if (unlikely(!skb)) {
 511			qdisc_qstats_drop(sch);
 512			goto finish_segs;
 513		}
 514		if (skb->ip_summed == CHECKSUM_PARTIAL &&
 515		    skb_checksum_help(skb)) {
 516			qdisc_drop(skb, sch, to_free);
 517			skb = NULL;
 518			goto finish_segs;
 519		}
 520
 521		skb->data[get_random_u32_below(skb_headlen(skb))] ^=
 522			1<<get_random_u32_below(8);
 523	}
 524
 525	if (unlikely(q->t_len >= sch->limit)) {
 526		/* re-link segs, so that qdisc_drop_all() frees them all */
 527		skb->next = segs;
 528		qdisc_drop_all(skb, sch, to_free);
 529		if (skb2)
 530			__qdisc_drop(skb2, to_free);
 531		return NET_XMIT_DROP;
 532	}
 533
 534	/*
 535	 * If doing duplication then re-insert at top of the
 536	 * qdisc tree, since parent queuer expects that only one
 537	 * skb will be queued.
 538	 */
 539	if (skb2) {
 540		struct Qdisc *rootq = qdisc_root_bh(sch);
 541		u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
 542
 543		q->duplicate = 0;
 544		rootq->enqueue(skb2, rootq, to_free);
 545		q->duplicate = dupsave;
 546		skb2 = NULL;
 547	}
 548
 549	qdisc_qstats_backlog_inc(sch, skb);
 550
 551	cb = netem_skb_cb(skb);
 552	if (q->gap == 0 ||		/* not doing reordering */
 553	    q->counter < q->gap - 1 ||	/* inside last reordering gap */
 554	    q->reorder < get_crandom(&q->reorder_cor, &q->prng)) {
 555		u64 now;
 556		s64 delay;
 557
 558		delay = tabledist(q->latency, q->jitter,
 559				  &q->delay_cor, &q->prng, q->delay_dist);
 560
 561		now = ktime_get_ns();
 562
 563		if (q->rate) {
 564			struct netem_skb_cb *last = NULL;
 565
 566			if (sch->q.tail)
 567				last = netem_skb_cb(sch->q.tail);
 568			if (q->t_root.rb_node) {
 569				struct sk_buff *t_skb;
 570				struct netem_skb_cb *t_last;
 571
 572				t_skb = skb_rb_last(&q->t_root);
 573				t_last = netem_skb_cb(t_skb);
 574				if (!last ||
 575				    t_last->time_to_send > last->time_to_send)
 576					last = t_last;
 577			}
 578			if (q->t_tail) {
 579				struct netem_skb_cb *t_last =
 580					netem_skb_cb(q->t_tail);
 581
 582				if (!last ||
 583				    t_last->time_to_send > last->time_to_send)
 584					last = t_last;
 585			}
 586
 587			if (last) {
 588				/*
 589				 * Last packet in queue is reference point (now),
 590				 * calculate this time bonus and subtract
 591				 * from delay.
 592				 */
 593				delay -= last->time_to_send - now;
 594				delay = max_t(s64, 0, delay);
 595				now = last->time_to_send;
 596			}
 597
 598			delay += packet_time_ns(qdisc_pkt_len(skb), q);
 599		}
 600
 
 601		cb->time_to_send = now + delay;
 602		++q->counter;
 603		tfifo_enqueue(skb, sch);
 604	} else {
 605		/*
 606		 * Do re-ordering by putting one out of N packets at the front
 607		 * of the queue.
 608		 */
 609		cb->time_to_send = ktime_get_ns();
 610		q->counter = 0;
 611
 612		__qdisc_enqueue_head(skb, &sch->q);
 613		sch->qstats.requeues++;
 
 
 614	}
 615
 616finish_segs:
 617	if (skb2)
 618		__qdisc_drop(skb2, to_free);
 619
 620	if (segs) {
 621		unsigned int len, last_len;
 622		int rc, nb;
 623
 624		len = skb ? skb->len : 0;
 625		nb = skb ? 1 : 0;
 626
 627		while (segs) {
 628			skb2 = segs->next;
 629			skb_mark_not_on_list(segs);
 630			qdisc_skb_cb(segs)->pkt_len = segs->len;
 631			last_len = segs->len;
 632			rc = qdisc_enqueue(segs, sch, to_free);
 633			if (rc != NET_XMIT_SUCCESS) {
 634				if (net_xmit_drop_count(rc))
 635					qdisc_qstats_drop(sch);
 636			} else {
 637				nb++;
 638				len += last_len;
 639			}
 640			segs = skb2;
 641		}
 642		/* Parent qdiscs accounted for 1 skb of size @prev_len */
 643		qdisc_tree_reduce_backlog(sch, -(nb - 1), -(len - prev_len));
 644	} else if (!skb) {
 645		return NET_XMIT_DROP;
 646	}
 647	return NET_XMIT_SUCCESS;
 648}
 649
 650/* Delay the next round with a new future slot with a
 651 * correct number of bytes and packets.
 652 */
 653
 654static void get_slot_next(struct netem_sched_data *q, u64 now)
 655{
 656	s64 next_delay;
 657
 658	if (!q->slot_dist)
 659		next_delay = q->slot_config.min_delay +
 660				(get_random_u32() *
 661				 (q->slot_config.max_delay -
 662				  q->slot_config.min_delay) >> 32);
 663	else
 664		next_delay = tabledist(q->slot_config.dist_delay,
 665				       (s32)(q->slot_config.dist_jitter),
 666				       NULL, &q->prng, q->slot_dist);
 667
 668	q->slot.slot_next = now + next_delay;
 669	q->slot.packets_left = q->slot_config.max_packets;
 670	q->slot.bytes_left = q->slot_config.max_bytes;
 671}
 672
 673static struct sk_buff *netem_peek(struct netem_sched_data *q)
 674{
 675	struct sk_buff *skb = skb_rb_first(&q->t_root);
 676	u64 t1, t2;
 677
 678	if (!skb)
 679		return q->t_head;
 680	if (!q->t_head)
 681		return skb;
 682
 683	t1 = netem_skb_cb(skb)->time_to_send;
 684	t2 = netem_skb_cb(q->t_head)->time_to_send;
 685	if (t1 < t2)
 686		return skb;
 687	return q->t_head;
 688}
 689
 690static void netem_erase_head(struct netem_sched_data *q, struct sk_buff *skb)
 691{
 692	if (skb == q->t_head) {
 693		q->t_head = skb->next;
 694		if (!q->t_head)
 695			q->t_tail = NULL;
 696	} else {
 697		rb_erase(&skb->rbnode, &q->t_root);
 698	}
 
 699}
 700
 701static struct sk_buff *netem_dequeue(struct Qdisc *sch)
 702{
 703	struct netem_sched_data *q = qdisc_priv(sch);
 704	struct sk_buff *skb;
 705
 706tfifo_dequeue:
 707	skb = __qdisc_dequeue_head(&sch->q);
 708	if (skb) {
 709deliver:
 710		qdisc_qstats_backlog_dec(sch, skb);
 711		qdisc_bstats_update(sch, skb);
 712		return skb;
 713	}
 714	skb = netem_peek(q);
 715	if (skb) {
 716		u64 time_to_send;
 717		u64 now = ktime_get_ns();
 718
 719		/* if more time remaining? */
 720		time_to_send = netem_skb_cb(skb)->time_to_send;
 721		if (q->slot.slot_next && q->slot.slot_next < time_to_send)
 722			get_slot_next(q, now);
 723
 724		if (time_to_send <= now && q->slot.slot_next <= now) {
 725			netem_erase_head(q, skb);
 726			q->t_len--;
 727			skb->next = NULL;
 728			skb->prev = NULL;
 729			/* skb->dev shares skb->rbnode area,
 730			 * we need to restore its value.
 731			 */
 732			skb->dev = qdisc_dev(sch);
 
 
 733
 734			if (q->slot.slot_next) {
 735				q->slot.packets_left--;
 736				q->slot.bytes_left -= qdisc_pkt_len(skb);
 737				if (q->slot.packets_left <= 0 ||
 738				    q->slot.bytes_left <= 0)
 739					get_slot_next(q, now);
 740			}
 741
 742			if (q->qdisc) {
 743				unsigned int pkt_len = qdisc_pkt_len(skb);
 744				struct sk_buff *to_free = NULL;
 745				int err;
 746
 747				err = qdisc_enqueue(skb, q->qdisc, &to_free);
 748				kfree_skb_list(to_free);
 749				if (err != NET_XMIT_SUCCESS) {
 750					if (net_xmit_drop_count(err))
 751						qdisc_qstats_drop(sch);
 752					sch->qstats.backlog -= pkt_len;
 753					sch->q.qlen--;
 754					qdisc_tree_reduce_backlog(sch, 1, pkt_len);
 755				}
 756				goto tfifo_dequeue;
 757			}
 758			sch->q.qlen--;
 759			goto deliver;
 
 
 760		}
 761
 762		if (q->qdisc) {
 763			skb = q->qdisc->ops->dequeue(q->qdisc);
 764			if (skb) {
 765				sch->q.qlen--;
 766				goto deliver;
 767			}
 768		}
 769
 770		qdisc_watchdog_schedule_ns(&q->watchdog,
 771					   max(time_to_send,
 772					       q->slot.slot_next));
 773	}
 774
 775	if (q->qdisc) {
 776		skb = q->qdisc->ops->dequeue(q->qdisc);
 777		if (skb) {
 778			sch->q.qlen--;
 779			goto deliver;
 780		}
 781	}
 782	return NULL;
 783}
 784
 785static void netem_reset(struct Qdisc *sch)
 786{
 787	struct netem_sched_data *q = qdisc_priv(sch);
 788
 789	qdisc_reset_queue(sch);
 790	tfifo_reset(sch);
 791	if (q->qdisc)
 792		qdisc_reset(q->qdisc);
 793	qdisc_watchdog_cancel(&q->watchdog);
 794}
 795
 796static void dist_free(struct disttable *d)
 797{
 798	kvfree(d);
 
 
 
 
 
 799}
 800
 801/*
 802 * Distribution data is a variable size payload containing
 803 * signed 16 bit values.
 804 */
 805
 806static int get_dist_table(struct disttable **tbl, const struct nlattr *attr)
 807{
 
 808	size_t n = nla_len(attr)/sizeof(__s16);
 809	const __s16 *data = nla_data(attr);
 
 810	struct disttable *d;
 811	int i;
 
 812
 813	if (!n || n > NETEM_DIST_MAX)
 814		return -EINVAL;
 815
 816	d = kvmalloc(struct_size(d, table, n), GFP_KERNEL);
 
 
 
 817	if (!d)
 818		return -ENOMEM;
 819
 820	d->size = n;
 821	for (i = 0; i < n; i++)
 822		d->table[i] = data[i];
 823
 824	*tbl = d;
 825	return 0;
 826}
 827
 828static void get_slot(struct netem_sched_data *q, const struct nlattr *attr)
 829{
 830	const struct tc_netem_slot *c = nla_data(attr);
 831
 832	q->slot_config = *c;
 833	if (q->slot_config.max_packets == 0)
 834		q->slot_config.max_packets = INT_MAX;
 835	if (q->slot_config.max_bytes == 0)
 836		q->slot_config.max_bytes = INT_MAX;
 837
 838	/* capping dist_jitter to the range acceptable by tabledist() */
 839	q->slot_config.dist_jitter = min_t(__s64, INT_MAX, abs(q->slot_config.dist_jitter));
 840
 841	q->slot.packets_left = q->slot_config.max_packets;
 842	q->slot.bytes_left = q->slot_config.max_bytes;
 843	if (q->slot_config.min_delay | q->slot_config.max_delay |
 844	    q->slot_config.dist_jitter)
 845		q->slot.slot_next = ktime_get_ns();
 846	else
 847		q->slot.slot_next = 0;
 848}
 849
 850static void get_correlation(struct netem_sched_data *q, const struct nlattr *attr)
 851{
 
 852	const struct tc_netem_corr *c = nla_data(attr);
 853
 854	init_crandom(&q->delay_cor, c->delay_corr);
 855	init_crandom(&q->loss_cor, c->loss_corr);
 856	init_crandom(&q->dup_cor, c->dup_corr);
 857}
 858
 859static void get_reorder(struct netem_sched_data *q, const struct nlattr *attr)
 860{
 
 861	const struct tc_netem_reorder *r = nla_data(attr);
 862
 863	q->reorder = r->probability;
 864	init_crandom(&q->reorder_cor, r->correlation);
 865}
 866
 867static void get_corrupt(struct netem_sched_data *q, const struct nlattr *attr)
 868{
 
 869	const struct tc_netem_corrupt *r = nla_data(attr);
 870
 871	q->corrupt = r->probability;
 872	init_crandom(&q->corrupt_cor, r->correlation);
 873}
 874
 875static void get_rate(struct netem_sched_data *q, const struct nlattr *attr)
 876{
 877	const struct tc_netem_rate *r = nla_data(attr);
 878
 879	q->rate = r->rate;
 880	q->packet_overhead = r->packet_overhead;
 881	q->cell_size = r->cell_size;
 882	q->cell_overhead = r->cell_overhead;
 883	if (q->cell_size)
 884		q->cell_size_reciprocal = reciprocal_value(q->cell_size);
 885	else
 886		q->cell_size_reciprocal = (struct reciprocal_value) { 0 };
 887}
 888
 889static int get_loss_clg(struct netem_sched_data *q, const struct nlattr *attr)
 890{
 
 891	const struct nlattr *la;
 892	int rem;
 893
 894	nla_for_each_nested(la, attr, rem) {
 895		u16 type = nla_type(la);
 896
 897		switch (type) {
 898		case NETEM_LOSS_GI: {
 899			const struct tc_netem_gimodel *gi = nla_data(la);
 900
 901			if (nla_len(la) < sizeof(struct tc_netem_gimodel)) {
 902				pr_info("netem: incorrect gi model size\n");
 903				return -EINVAL;
 904			}
 905
 906			q->loss_model = CLG_4_STATES;
 907
 908			q->clg.state = TX_IN_GAP_PERIOD;
 909			q->clg.a1 = gi->p13;
 910			q->clg.a2 = gi->p31;
 911			q->clg.a3 = gi->p32;
 912			q->clg.a4 = gi->p14;
 913			q->clg.a5 = gi->p23;
 914			break;
 915		}
 916
 917		case NETEM_LOSS_GE: {
 918			const struct tc_netem_gemodel *ge = nla_data(la);
 919
 920			if (nla_len(la) < sizeof(struct tc_netem_gemodel)) {
 921				pr_info("netem: incorrect ge model size\n");
 922				return -EINVAL;
 923			}
 924
 925			q->loss_model = CLG_GILB_ELL;
 926			q->clg.state = GOOD_STATE;
 927			q->clg.a1 = ge->p;
 928			q->clg.a2 = ge->r;
 929			q->clg.a3 = ge->h;
 930			q->clg.a4 = ge->k1;
 931			break;
 932		}
 933
 934		default:
 935			pr_info("netem: unknown loss type %u\n", type);
 936			return -EINVAL;
 937		}
 938	}
 939
 940	return 0;
 941}
 942
 943static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
 944	[TCA_NETEM_CORR]	= { .len = sizeof(struct tc_netem_corr) },
 945	[TCA_NETEM_REORDER]	= { .len = sizeof(struct tc_netem_reorder) },
 946	[TCA_NETEM_CORRUPT]	= { .len = sizeof(struct tc_netem_corrupt) },
 947	[TCA_NETEM_RATE]	= { .len = sizeof(struct tc_netem_rate) },
 948	[TCA_NETEM_LOSS]	= { .type = NLA_NESTED },
 949	[TCA_NETEM_ECN]		= { .type = NLA_U32 },
 950	[TCA_NETEM_RATE64]	= { .type = NLA_U64 },
 951	[TCA_NETEM_LATENCY64]	= { .type = NLA_S64 },
 952	[TCA_NETEM_JITTER64]	= { .type = NLA_S64 },
 953	[TCA_NETEM_SLOT]	= { .len = sizeof(struct tc_netem_slot) },
 954	[TCA_NETEM_PRNG_SEED]	= { .type = NLA_U64 },
 955};
 956
 957static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
 958		      const struct nla_policy *policy, int len)
 959{
 960	int nested_len = nla_len(nla) - NLA_ALIGN(len);
 961
 962	if (nested_len < 0) {
 963		pr_info("netem: invalid attributes len %d\n", nested_len);
 964		return -EINVAL;
 965	}
 966
 967	if (nested_len >= nla_attr_size(0))
 968		return nla_parse_deprecated(tb, maxtype,
 969					    nla_data(nla) + NLA_ALIGN(len),
 970					    nested_len, policy, NULL);
 971
 972	memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
 973	return 0;
 974}
 975
 976/* Parse netlink message to set options */
 977static int netem_change(struct Qdisc *sch, struct nlattr *opt,
 978			struct netlink_ext_ack *extack)
 979{
 980	struct netem_sched_data *q = qdisc_priv(sch);
 981	struct nlattr *tb[TCA_NETEM_MAX + 1];
 982	struct disttable *delay_dist = NULL;
 983	struct disttable *slot_dist = NULL;
 984	struct tc_netem_qopt *qopt;
 985	struct clgstate old_clg;
 986	int old_loss_model = CLG_RANDOM;
 987	int ret;
 988
 
 
 
 989	qopt = nla_data(opt);
 990	ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
 991	if (ret < 0)
 992		return ret;
 993
 994	if (tb[TCA_NETEM_DELAY_DIST]) {
 995		ret = get_dist_table(&delay_dist, tb[TCA_NETEM_DELAY_DIST]);
 996		if (ret)
 997			goto table_free;
 998	}
 999
1000	if (tb[TCA_NETEM_SLOT_DIST]) {
1001		ret = get_dist_table(&slot_dist, tb[TCA_NETEM_SLOT_DIST]);
1002		if (ret)
1003			goto table_free;
1004	}
1005
1006	sch_tree_lock(sch);
1007	/* backup q->clg and q->loss_model */
1008	old_clg = q->clg;
1009	old_loss_model = q->loss_model;
1010
1011	if (tb[TCA_NETEM_LOSS]) {
1012		ret = get_loss_clg(q, tb[TCA_NETEM_LOSS]);
1013		if (ret) {
1014			q->loss_model = old_loss_model;
1015			q->clg = old_clg;
1016			goto unlock;
1017		}
1018	} else {
1019		q->loss_model = CLG_RANDOM;
1020	}
1021
1022	if (delay_dist)
1023		swap(q->delay_dist, delay_dist);
1024	if (slot_dist)
1025		swap(q->slot_dist, slot_dist);
1026	sch->limit = qopt->limit;
1027
1028	q->latency = PSCHED_TICKS2NS(qopt->latency);
1029	q->jitter = PSCHED_TICKS2NS(qopt->jitter);
1030	q->limit = qopt->limit;
1031	q->gap = qopt->gap;
1032	q->counter = 0;
1033	q->loss = qopt->loss;
1034	q->duplicate = qopt->duplicate;
1035
1036	/* for compatibility with earlier versions.
1037	 * if gap is set, need to assume 100% probability
1038	 */
1039	if (q->gap)
1040		q->reorder = ~0;
1041
1042	if (tb[TCA_NETEM_CORR])
1043		get_correlation(q, tb[TCA_NETEM_CORR]);
 
 
 
 
 
 
1044
1045	if (tb[TCA_NETEM_REORDER])
1046		get_reorder(q, tb[TCA_NETEM_REORDER]);
1047
1048	if (tb[TCA_NETEM_CORRUPT])
1049		get_corrupt(q, tb[TCA_NETEM_CORRUPT]);
1050
1051	if (tb[TCA_NETEM_RATE])
1052		get_rate(q, tb[TCA_NETEM_RATE]);
 
1053
1054	if (tb[TCA_NETEM_RATE64])
1055		q->rate = max_t(u64, q->rate,
1056				nla_get_u64(tb[TCA_NETEM_RATE64]));
1057
1058	if (tb[TCA_NETEM_LATENCY64])
1059		q->latency = nla_get_s64(tb[TCA_NETEM_LATENCY64]);
 
 
 
 
 
 
1060
1061	if (tb[TCA_NETEM_JITTER64])
1062		q->jitter = nla_get_s64(tb[TCA_NETEM_JITTER64]);
 
 
 
 
1063
1064	if (tb[TCA_NETEM_ECN])
1065		q->ecn = nla_get_u32(tb[TCA_NETEM_ECN]);
 
 
 
 
1066
1067	if (tb[TCA_NETEM_SLOT])
1068		get_slot(q, tb[TCA_NETEM_SLOT]);
1069
1070	/* capping jitter to the range acceptable by tabledist() */
1071	q->jitter = min_t(s64, abs(q->jitter), INT_MAX);
 
1072
1073	if (tb[TCA_NETEM_PRNG_SEED])
1074		q->prng.seed = nla_get_u64(tb[TCA_NETEM_PRNG_SEED]);
1075	else
1076		q->prng.seed = get_random_u64();
1077	prandom_seed_state(&q->prng.prng_state, q->prng.seed);
1078
1079unlock:
1080	sch_tree_unlock(sch);
1081
1082table_free:
1083	dist_free(delay_dist);
1084	dist_free(slot_dist);
1085	return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1086}
1087
1088static int netem_init(struct Qdisc *sch, struct nlattr *opt,
1089		      struct netlink_ext_ack *extack)
 
 
 
 
 
 
 
 
 
 
 
 
1090{
1091	struct netem_sched_data *q = qdisc_priv(sch);
1092	int ret;
1093
1094	qdisc_watchdog_init(&q->watchdog, sch);
1095
1096	if (!opt)
1097		return -EINVAL;
1098
 
 
1099	q->loss_model = CLG_RANDOM;
1100	ret = netem_change(sch, opt, extack);
1101	if (ret)
 
 
 
 
 
 
 
1102		pr_info("netem: change failed\n");
 
 
1103	return ret;
1104}
1105
1106static void netem_destroy(struct Qdisc *sch)
1107{
1108	struct netem_sched_data *q = qdisc_priv(sch);
1109
1110	qdisc_watchdog_cancel(&q->watchdog);
1111	if (q->qdisc)
1112		qdisc_put(q->qdisc);
1113	dist_free(q->delay_dist);
1114	dist_free(q->slot_dist);
1115}
1116
1117static int dump_loss_model(const struct netem_sched_data *q,
1118			   struct sk_buff *skb)
1119{
1120	struct nlattr *nest;
1121
1122	nest = nla_nest_start_noflag(skb, TCA_NETEM_LOSS);
1123	if (nest == NULL)
1124		goto nla_put_failure;
1125
1126	switch (q->loss_model) {
1127	case CLG_RANDOM:
1128		/* legacy loss model */
1129		nla_nest_cancel(skb, nest);
1130		return 0;	/* no data */
1131
1132	case CLG_4_STATES: {
1133		struct tc_netem_gimodel gi = {
1134			.p13 = q->clg.a1,
1135			.p31 = q->clg.a2,
1136			.p32 = q->clg.a3,
1137			.p14 = q->clg.a4,
1138			.p23 = q->clg.a5,
1139		};
1140
1141		if (nla_put(skb, NETEM_LOSS_GI, sizeof(gi), &gi))
1142			goto nla_put_failure;
1143		break;
1144	}
1145	case CLG_GILB_ELL: {
1146		struct tc_netem_gemodel ge = {
1147			.p = q->clg.a1,
1148			.r = q->clg.a2,
1149			.h = q->clg.a3,
1150			.k1 = q->clg.a4,
1151		};
1152
1153		if (nla_put(skb, NETEM_LOSS_GE, sizeof(ge), &ge))
1154			goto nla_put_failure;
1155		break;
1156	}
1157	}
1158
1159	nla_nest_end(skb, nest);
1160	return 0;
1161
1162nla_put_failure:
1163	nla_nest_cancel(skb, nest);
1164	return -1;
1165}
1166
1167static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
1168{
1169	const struct netem_sched_data *q = qdisc_priv(sch);
1170	struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb);
1171	struct tc_netem_qopt qopt;
1172	struct tc_netem_corr cor;
1173	struct tc_netem_reorder reorder;
1174	struct tc_netem_corrupt corrupt;
1175	struct tc_netem_rate rate;
1176	struct tc_netem_slot slot;
1177
1178	qopt.latency = min_t(psched_time_t, PSCHED_NS2TICKS(q->latency),
1179			     UINT_MAX);
1180	qopt.jitter = min_t(psched_time_t, PSCHED_NS2TICKS(q->jitter),
1181			    UINT_MAX);
1182	qopt.limit = q->limit;
1183	qopt.loss = q->loss;
1184	qopt.gap = q->gap;
1185	qopt.duplicate = q->duplicate;
1186	if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
1187		goto nla_put_failure;
1188
1189	if (nla_put(skb, TCA_NETEM_LATENCY64, sizeof(q->latency), &q->latency))
1190		goto nla_put_failure;
1191
1192	if (nla_put(skb, TCA_NETEM_JITTER64, sizeof(q->jitter), &q->jitter))
1193		goto nla_put_failure;
1194
1195	cor.delay_corr = q->delay_cor.rho;
1196	cor.loss_corr = q->loss_cor.rho;
1197	cor.dup_corr = q->dup_cor.rho;
1198	if (nla_put(skb, TCA_NETEM_CORR, sizeof(cor), &cor))
1199		goto nla_put_failure;
1200
1201	reorder.probability = q->reorder;
1202	reorder.correlation = q->reorder_cor.rho;
1203	if (nla_put(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder))
1204		goto nla_put_failure;
1205
1206	corrupt.probability = q->corrupt;
1207	corrupt.correlation = q->corrupt_cor.rho;
1208	if (nla_put(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt))
1209		goto nla_put_failure;
1210
1211	if (q->rate >= (1ULL << 32)) {
1212		if (nla_put_u64_64bit(skb, TCA_NETEM_RATE64, q->rate,
1213				      TCA_NETEM_PAD))
1214			goto nla_put_failure;
1215		rate.rate = ~0U;
1216	} else {
1217		rate.rate = q->rate;
1218	}
1219	rate.packet_overhead = q->packet_overhead;
1220	rate.cell_size = q->cell_size;
1221	rate.cell_overhead = q->cell_overhead;
1222	if (nla_put(skb, TCA_NETEM_RATE, sizeof(rate), &rate))
1223		goto nla_put_failure;
1224
1225	if (q->ecn && nla_put_u32(skb, TCA_NETEM_ECN, q->ecn))
1226		goto nla_put_failure;
1227
1228	if (dump_loss_model(q, skb) != 0)
1229		goto nla_put_failure;
1230
1231	if (q->slot_config.min_delay | q->slot_config.max_delay |
1232	    q->slot_config.dist_jitter) {
1233		slot = q->slot_config;
1234		if (slot.max_packets == INT_MAX)
1235			slot.max_packets = 0;
1236		if (slot.max_bytes == INT_MAX)
1237			slot.max_bytes = 0;
1238		if (nla_put(skb, TCA_NETEM_SLOT, sizeof(slot), &slot))
1239			goto nla_put_failure;
1240	}
1241
1242	if (nla_put_u64_64bit(skb, TCA_NETEM_PRNG_SEED, q->prng.seed,
1243			      TCA_NETEM_PAD))
1244		goto nla_put_failure;
1245
1246	return nla_nest_end(skb, nla);
1247
1248nla_put_failure:
1249	nlmsg_trim(skb, nla);
1250	return -1;
1251}
1252
1253static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
1254			  struct sk_buff *skb, struct tcmsg *tcm)
1255{
1256	struct netem_sched_data *q = qdisc_priv(sch);
1257
1258	if (cl != 1 || !q->qdisc) 	/* only one class */
1259		return -ENOENT;
1260
1261	tcm->tcm_handle |= TC_H_MIN(1);
1262	tcm->tcm_info = q->qdisc->handle;
1263
1264	return 0;
1265}
1266
1267static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1268		     struct Qdisc **old, struct netlink_ext_ack *extack)
1269{
1270	struct netem_sched_data *q = qdisc_priv(sch);
1271
1272	*old = qdisc_replace(sch, new, &q->qdisc);
 
 
 
 
 
 
 
 
 
1273	return 0;
1274}
1275
1276static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
1277{
1278	struct netem_sched_data *q = qdisc_priv(sch);
1279	return q->qdisc;
1280}
1281
1282static unsigned long netem_find(struct Qdisc *sch, u32 classid)
1283{
1284	return 1;
1285}
1286
 
 
 
 
1287static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
1288{
1289	if (!walker->stop) {
1290		if (!tc_qdisc_stats_dump(sch, 1, walker))
1291			return;
 
 
 
 
1292	}
1293}
1294
1295static const struct Qdisc_class_ops netem_class_ops = {
1296	.graft		=	netem_graft,
1297	.leaf		=	netem_leaf,
1298	.find		=	netem_find,
 
1299	.walk		=	netem_walk,
1300	.dump		=	netem_dump_class,
1301};
1302
1303static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
1304	.id		=	"netem",
1305	.cl_ops		=	&netem_class_ops,
1306	.priv_size	=	sizeof(struct netem_sched_data),
1307	.enqueue	=	netem_enqueue,
1308	.dequeue	=	netem_dequeue,
1309	.peek		=	qdisc_peek_dequeued,
 
1310	.init		=	netem_init,
1311	.reset		=	netem_reset,
1312	.destroy	=	netem_destroy,
1313	.change		=	netem_change,
1314	.dump		=	netem_dump,
1315	.owner		=	THIS_MODULE,
1316};
1317MODULE_ALIAS_NET_SCH("netem");
1318
1319
1320static int __init netem_module_init(void)
1321{
1322	pr_info("netem: version " VERSION "\n");
1323	return register_qdisc(&netem_qdisc_ops);
1324}
1325static void __exit netem_module_exit(void)
1326{
1327	unregister_qdisc(&netem_qdisc_ops);
1328}
1329module_init(netem_module_init)
1330module_exit(netem_module_exit)
1331MODULE_LICENSE("GPL");
1332MODULE_DESCRIPTION("Network characteristics emulator qdisc");
v3.1
 
  1/*
  2 * net/sched/sch_netem.c	Network emulator
  3 *
  4 * 		This program is free software; you can redistribute it and/or
  5 * 		modify it under the terms of the GNU General Public License
  6 * 		as published by the Free Software Foundation; either version
  7 * 		2 of the License.
  8 *
  9 *  		Many of the algorithms and ideas for this came from
 10 *		NIST Net which is not copyrighted.
 11 *
 12 * Authors:	Stephen Hemminger <shemminger@osdl.org>
 13 *		Catalin(ux aka Dino) BOIE <catab at umbrella dot ro>
 14 */
 15
 16#include <linux/mm.h>
 17#include <linux/module.h>
 18#include <linux/slab.h>
 19#include <linux/types.h>
 20#include <linux/kernel.h>
 21#include <linux/errno.h>
 22#include <linux/skbuff.h>
 23#include <linux/vmalloc.h>
 
 24#include <linux/rtnetlink.h>
 
 
 25
 
 26#include <net/netlink.h>
 27#include <net/pkt_sched.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 netem_sched_data {
 
 
 
 
 
 
 
 
 
 
 69	struct Qdisc	*qdisc;
 
 70	struct qdisc_watchdog watchdog;
 71
 72	psched_tdiff_t latency;
 73	psched_tdiff_t jitter;
 74
 75	u32 loss;
 
 76	u32 limit;
 77	u32 counter;
 78	u32 gap;
 79	u32 duplicate;
 80	u32 reorder;
 81	u32 corrupt;
 
 
 
 
 
 82
 83	struct crndstate {
 84		u32 last;
 85		u32 rho;
 86	} delay_cor, loss_cor, dup_cor, reorder_cor, corrupt_cor;
 87
 88	struct disttable {
 89		u32  size;
 90		s16 table[0];
 91	} *delay_dist;
 
 
 92
 93	enum  {
 94		CLG_RANDOM,
 95		CLG_4_STATES,
 96		CLG_GILB_ELL,
 97	} loss_model;
 98
 
 
 
 
 
 
 
 
 
 
 
 
 99	/* Correlated Loss Generation models */
100	struct clgstate {
101		/* state of the Markov chain */
102		u8 state;
103
104		/* 4-states and Gilbert-Elliot models */
105		u32 a1;	/* p13 for 4-states or p for GE */
106		u32 a2;	/* p31 for 4-states or r for GE */
107		u32 a3;	/* p32 for 4-states or h for GE */
108		u32 a4;	/* p14 for 4-states or 1-k for GE */
109		u32 a5; /* p23 used only in 4-states */
110	} clg;
111
 
 
 
 
 
 
 
 
112};
113
114/* Time stamp put into socket buffer control block */
 
 
 
 
 
 
115struct netem_skb_cb {
116	psched_time_t	time_to_send;
117};
118
119static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb)
120{
121	BUILD_BUG_ON(sizeof(skb->cb) <
122		sizeof(struct qdisc_skb_cb) + sizeof(struct netem_skb_cb));
123	return (struct netem_skb_cb *)qdisc_skb_cb(skb)->data;
124}
125
126/* init_crandom - initialize correlated random number generator
127 * Use entropy source for initial seed.
128 */
129static void init_crandom(struct crndstate *state, unsigned long rho)
130{
131	state->rho = rho;
132	state->last = net_random();
133}
134
135/* get_crandom - correlated random number generator
136 * Next number depends on last value.
137 * rho is scaled to avoid floating point.
138 */
139static u32 get_crandom(struct crndstate *state)
140{
141	u64 value, rho;
142	unsigned long answer;
 
143
144	if (state->rho == 0)	/* no correlation */
145		return net_random();
146
147	value = net_random();
148	rho = (u64)state->rho + 1;
149	answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32;
150	state->last = answer;
151	return answer;
152}
153
154/* loss_4state - 4-state model loss generator
155 * Generates losses according to the 4-state Markov chain adopted in
156 * the GI (General and Intuitive) loss model.
157 */
158static bool loss_4state(struct netem_sched_data *q)
159{
160	struct clgstate *clg = &q->clg;
161	u32 rnd = net_random();
162
163	/*
164	 * Makes a comparison between rnd and the transition
165	 * probabilities outgoing from the current state, then decides the
166	 * next state and if the next packet has to be transmitted or lost.
167	 * The four states correspond to:
168	 *   1 => successfully transmitted packets within a gap period
169	 *   4 => isolated losses within a gap period
170	 *   3 => lost packets within a burst period
171	 *   2 => successfully transmitted packets within a burst period
172	 */
173	switch (clg->state) {
174	case 1:
175		if (rnd < clg->a4) {
176			clg->state = 4;
177			return true;
178		} else if (clg->a4 < rnd && rnd < clg->a1) {
179			clg->state = 3;
180			return true;
181		} else if (clg->a1 < rnd)
182			clg->state = 1;
 
183
184		break;
185	case 2:
186		if (rnd < clg->a5) {
187			clg->state = 3;
188			return true;
189		} else
190			clg->state = 2;
 
191
192		break;
193	case 3:
194		if (rnd < clg->a3)
195			clg->state = 2;
196		else if (clg->a3 < rnd && rnd < clg->a2 + clg->a3) {
197			clg->state = 1;
198			return true;
199		} else if (clg->a2 + clg->a3 < rnd) {
200			clg->state = 3;
201			return true;
202		}
203		break;
204	case 4:
205		clg->state = 1;
206		break;
207	}
208
209	return false;
210}
211
212/* loss_gilb_ell - Gilbert-Elliot model loss generator
213 * Generates losses according to the Gilbert-Elliot loss model or
214 * its special cases  (Gilbert or Simple Gilbert)
215 *
216 * Makes a comparison between random number and the transition
217 * probabilities outgoing from the current state, then decides the
218 * next state. A second random number is extracted and the comparison
219 * with the loss probability of the current state decides if the next
220 * packet will be transmitted or lost.
221 */
222static bool loss_gilb_ell(struct netem_sched_data *q)
223{
224	struct clgstate *clg = &q->clg;
 
225
226	switch (clg->state) {
227	case 1:
228		if (net_random() < clg->a1)
229			clg->state = 2;
230		if (net_random() < clg->a4)
231			return true;
232	case 2:
233		if (net_random() < clg->a2)
234			clg->state = 1;
235		if (clg->a3 > net_random())
 
236			return true;
237	}
238
239	return false;
240}
241
242static bool loss_event(struct netem_sched_data *q)
243{
244	switch (q->loss_model) {
245	case CLG_RANDOM:
246		/* Random packet drop 0 => none, ~0 => all */
247		return q->loss && q->loss >= get_crandom(&q->loss_cor);
248
249	case CLG_4_STATES:
250		/* 4state loss model algorithm (used also for GI model)
251		* Extracts a value from the markov 4 state loss generator,
252		* if it is 1 drops a packet and if needed writes the event in
253		* the kernel logs
254		*/
255		return loss_4state(q);
256
257	case CLG_GILB_ELL:
258		/* Gilbert-Elliot loss model algorithm
259		* Extracts a value from the Gilbert-Elliot loss generator,
260		* if it is 1 drops a packet and if needed writes the event in
261		* the kernel logs
262		*/
263		return loss_gilb_ell(q);
264	}
265
266	return false;	/* not reached */
267}
268
269
270/* tabledist - return a pseudo-randomly distributed value with mean mu and
271 * std deviation sigma.  Uses table lookup to approximate the desired
272 * distribution, and a uniformly-distributed pseudo-random source.
273 */
274static psched_tdiff_t tabledist(psched_tdiff_t mu, psched_tdiff_t sigma,
275				struct crndstate *state,
276				const struct disttable *dist)
 
277{
278	psched_tdiff_t x;
279	long t;
280	u32 rnd;
281
282	if (sigma == 0)
283		return mu;
284
285	rnd = get_crandom(state);
286
287	/* default uniform distribution */
288	if (dist == NULL)
289		return (rnd % (2*sigma)) - sigma + mu;
290
291	t = dist->table[rnd % dist->size];
292	x = (sigma % NETEM_DIST_SCALE) * t;
293	if (x >= 0)
294		x += NETEM_DIST_SCALE/2;
295	else
296		x -= NETEM_DIST_SCALE/2;
297
298	return  x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu;
299}
300
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
301/*
302 * Insert one skb into qdisc.
303 * Note: parent depends on return value to account for queue length.
304 * 	NET_XMIT_DROP: queue length didn't change.
305 *      NET_XMIT_SUCCESS: one skb was queued.
306 */
307static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch)
 
308{
309	struct netem_sched_data *q = qdisc_priv(sch);
310	/* We don't fill cb now as skb_unshare() may invalidate it */
311	struct netem_skb_cb *cb;
312	struct sk_buff *skb2;
313	int ret;
 
314	int count = 1;
315
 
 
 
316	/* Random duplication */
317	if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor))
318		++count;
319
320	/* Drop packet? */
321	if (loss_event(q))
322		--count;
323
 
 
 
324	if (count == 0) {
325		sch->qstats.drops++;
326		kfree_skb(skb);
327		return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
328	}
329
330	skb_orphan(skb);
 
 
 
 
331
332	/*
333	 * If we need to duplicate packet, then re-insert at top of the
334	 * qdisc tree, since parent queuer expects that only one
335	 * skb will be queued.
336	 */
337	if (count > 1 && (skb2 = skb_clone(skb, GFP_ATOMIC)) != NULL) {
338		struct Qdisc *rootq = qdisc_root(sch);
339		u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
340		q->duplicate = 0;
341
342		qdisc_enqueue_root(skb2, rootq);
343		q->duplicate = dupsave;
344	}
345
346	/*
347	 * Randomized packet corruption.
348	 * Make copy if needed since we are modifying
349	 * If packet is going to be hardware checksummed, then
350	 * do it now in software before we mangle it.
351	 */
352	if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor)) {
353		if (!(skb = skb_unshare(skb, GFP_ATOMIC)) ||
354		    (skb->ip_summed == CHECKSUM_PARTIAL &&
355		     skb_checksum_help(skb))) {
356			sch->qstats.drops++;
357			return NET_XMIT_DROP;
 
 
 
358		}
359
360		skb->data[net_random() % skb_headlen(skb)] ^= 1<<(net_random() % 8);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
361	}
362
 
 
363	cb = netem_skb_cb(skb);
364	if (q->gap == 0 ||		/* not doing reordering */
365	    q->counter < q->gap ||	/* inside last reordering gap */
366	    q->reorder < get_crandom(&q->reorder_cor)) {
367		psched_time_t now;
368		psched_tdiff_t delay;
369
370		delay = tabledist(q->latency, q->jitter,
371				  &q->delay_cor, q->delay_dist);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
372
373		now = psched_get_time();
374		cb->time_to_send = now + delay;
375		++q->counter;
376		ret = qdisc_enqueue(skb, q->qdisc);
377	} else {
378		/*
379		 * Do re-ordering by putting one out of N packets at the front
380		 * of the queue.
381		 */
382		cb->time_to_send = psched_get_time();
383		q->counter = 0;
384
385		__skb_queue_head(&q->qdisc->q, skb);
386		q->qdisc->qstats.backlog += qdisc_pkt_len(skb);
387		q->qdisc->qstats.requeues++;
388		ret = NET_XMIT_SUCCESS;
389	}
390
391	if (ret != NET_XMIT_SUCCESS) {
392		if (net_xmit_drop_count(ret)) {
393			sch->qstats.drops++;
394			return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
395		}
 
 
 
 
396	}
 
 
397
398	sch->q.qlen++;
399	return NET_XMIT_SUCCESS;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
400}
401
402static unsigned int netem_drop(struct Qdisc *sch)
403{
404	struct netem_sched_data *q = qdisc_priv(sch);
405	unsigned int len = 0;
 
 
 
 
 
 
 
 
 
 
 
 
406
407	if (q->qdisc->ops->drop && (len = q->qdisc->ops->drop(q->qdisc)) != 0) {
408		sch->q.qlen--;
409		sch->qstats.drops++;
 
 
 
 
 
410	}
411	return len;
412}
413
414static struct sk_buff *netem_dequeue(struct Qdisc *sch)
415{
416	struct netem_sched_data *q = qdisc_priv(sch);
417	struct sk_buff *skb;
418
419	if (qdisc_is_throttled(sch))
420		return NULL;
421
422	skb = q->qdisc->ops->peek(q->qdisc);
 
 
 
 
 
423	if (skb) {
424		const struct netem_skb_cb *cb = netem_skb_cb(skb);
425		psched_time_t now = psched_get_time();
426
427		/* if more time remaining? */
428		if (cb->time_to_send <= now) {
429			skb = qdisc_dequeue_peeked(q->qdisc);
430			if (unlikely(!skb))
431				return NULL;
432
433#ifdef CONFIG_NET_CLS_ACT
434			/*
435			 * If it's at ingress let's pretend the delay is
436			 * from the network (tstamp will be updated).
 
 
437			 */
438			if (G_TC_FROM(skb->tc_verd) & AT_INGRESS)
439				skb->tstamp.tv64 = 0;
440#endif
441
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
442			sch->q.qlen--;
443			qdisc_unthrottled(sch);
444			qdisc_bstats_update(sch, skb);
445			return skb;
446		}
447
448		qdisc_watchdog_schedule(&q->watchdog, cb->time_to_send);
 
 
 
 
 
 
 
 
 
 
449	}
450
 
 
 
 
 
 
 
451	return NULL;
452}
453
454static void netem_reset(struct Qdisc *sch)
455{
456	struct netem_sched_data *q = qdisc_priv(sch);
457
458	qdisc_reset(q->qdisc);
459	sch->q.qlen = 0;
 
 
460	qdisc_watchdog_cancel(&q->watchdog);
461}
462
463static void dist_free(struct disttable *d)
464{
465	if (d) {
466		if (is_vmalloc_addr(d))
467			vfree(d);
468		else
469			kfree(d);
470	}
471}
472
473/*
474 * Distribution data is a variable size payload containing
475 * signed 16 bit values.
476 */
477static int get_dist_table(struct Qdisc *sch, const struct nlattr *attr)
 
478{
479	struct netem_sched_data *q = qdisc_priv(sch);
480	size_t n = nla_len(attr)/sizeof(__s16);
481	const __s16 *data = nla_data(attr);
482	spinlock_t *root_lock;
483	struct disttable *d;
484	int i;
485	size_t s;
486
487	if (n > NETEM_DIST_MAX)
488		return -EINVAL;
489
490	s = sizeof(struct disttable) + n * sizeof(s16);
491	d = kmalloc(s, GFP_KERNEL);
492	if (!d)
493		d = vmalloc(s);
494	if (!d)
495		return -ENOMEM;
496
497	d->size = n;
498	for (i = 0; i < n; i++)
499		d->table[i] = data[i];
500
501	root_lock = qdisc_root_sleeping_lock(sch);
 
 
 
 
 
 
502
503	spin_lock_bh(root_lock);
504	dist_free(q->delay_dist);
505	q->delay_dist = d;
506	spin_unlock_bh(root_lock);
507	return 0;
 
 
 
 
 
 
 
 
 
 
 
508}
509
510static void get_correlation(struct Qdisc *sch, const struct nlattr *attr)
511{
512	struct netem_sched_data *q = qdisc_priv(sch);
513	const struct tc_netem_corr *c = nla_data(attr);
514
515	init_crandom(&q->delay_cor, c->delay_corr);
516	init_crandom(&q->loss_cor, c->loss_corr);
517	init_crandom(&q->dup_cor, c->dup_corr);
518}
519
520static void get_reorder(struct Qdisc *sch, const struct nlattr *attr)
521{
522	struct netem_sched_data *q = qdisc_priv(sch);
523	const struct tc_netem_reorder *r = nla_data(attr);
524
525	q->reorder = r->probability;
526	init_crandom(&q->reorder_cor, r->correlation);
527}
528
529static void get_corrupt(struct Qdisc *sch, const struct nlattr *attr)
530{
531	struct netem_sched_data *q = qdisc_priv(sch);
532	const struct tc_netem_corrupt *r = nla_data(attr);
533
534	q->corrupt = r->probability;
535	init_crandom(&q->corrupt_cor, r->correlation);
536}
537
538static int get_loss_clg(struct Qdisc *sch, const struct nlattr *attr)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
539{
540	struct netem_sched_data *q = qdisc_priv(sch);
541	const struct nlattr *la;
542	int rem;
543
544	nla_for_each_nested(la, attr, rem) {
545		u16 type = nla_type(la);
546
547		switch(type) {
548		case NETEM_LOSS_GI: {
549			const struct tc_netem_gimodel *gi = nla_data(la);
550
551			if (nla_len(la) != sizeof(struct tc_netem_gimodel)) {
552				pr_info("netem: incorrect gi model size\n");
553				return -EINVAL;
554			}
555
556			q->loss_model = CLG_4_STATES;
557
558			q->clg.state = 1;
559			q->clg.a1 = gi->p13;
560			q->clg.a2 = gi->p31;
561			q->clg.a3 = gi->p32;
562			q->clg.a4 = gi->p14;
563			q->clg.a5 = gi->p23;
564			break;
565		}
566
567		case NETEM_LOSS_GE: {
568			const struct tc_netem_gemodel *ge = nla_data(la);
569
570			if (nla_len(la) != sizeof(struct tc_netem_gemodel)) {
571				pr_info("netem: incorrect gi model size\n");
572				return -EINVAL;
573			}
574
575			q->loss_model = CLG_GILB_ELL;
576			q->clg.state = 1;
577			q->clg.a1 = ge->p;
578			q->clg.a2 = ge->r;
579			q->clg.a3 = ge->h;
580			q->clg.a4 = ge->k1;
581			break;
582		}
583
584		default:
585			pr_info("netem: unknown loss type %u\n", type);
586			return -EINVAL;
587		}
588	}
589
590	return 0;
591}
592
593static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
594	[TCA_NETEM_CORR]	= { .len = sizeof(struct tc_netem_corr) },
595	[TCA_NETEM_REORDER]	= { .len = sizeof(struct tc_netem_reorder) },
596	[TCA_NETEM_CORRUPT]	= { .len = sizeof(struct tc_netem_corrupt) },
 
597	[TCA_NETEM_LOSS]	= { .type = NLA_NESTED },
 
 
 
 
 
 
598};
599
600static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
601		      const struct nla_policy *policy, int len)
602{
603	int nested_len = nla_len(nla) - NLA_ALIGN(len);
604
605	if (nested_len < 0) {
606		pr_info("netem: invalid attributes len %d\n", nested_len);
607		return -EINVAL;
608	}
609
610	if (nested_len >= nla_attr_size(0))
611		return nla_parse(tb, maxtype, nla_data(nla) + NLA_ALIGN(len),
612				 nested_len, policy);
 
613
614	memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
615	return 0;
616}
617
618/* Parse netlink message to set options */
619static int netem_change(struct Qdisc *sch, struct nlattr *opt)
 
620{
621	struct netem_sched_data *q = qdisc_priv(sch);
622	struct nlattr *tb[TCA_NETEM_MAX + 1];
 
 
623	struct tc_netem_qopt *qopt;
 
 
624	int ret;
625
626	if (opt == NULL)
627		return -EINVAL;
628
629	qopt = nla_data(opt);
630	ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
631	if (ret < 0)
632		return ret;
633
634	ret = fifo_set_limit(q->qdisc, qopt->limit);
635	if (ret) {
636		pr_info("netem: can't set fifo limit\n");
637		return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
638	}
639
640	q->latency = qopt->latency;
641	q->jitter = qopt->jitter;
 
 
 
 
 
 
642	q->limit = qopt->limit;
643	q->gap = qopt->gap;
644	q->counter = 0;
645	q->loss = qopt->loss;
646	q->duplicate = qopt->duplicate;
647
648	/* for compatibility with earlier versions.
649	 * if gap is set, need to assume 100% probability
650	 */
651	if (q->gap)
652		q->reorder = ~0;
653
654	if (tb[TCA_NETEM_CORR])
655		get_correlation(sch, tb[TCA_NETEM_CORR]);
656
657	if (tb[TCA_NETEM_DELAY_DIST]) {
658		ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST]);
659		if (ret)
660			return ret;
661	}
662
663	if (tb[TCA_NETEM_REORDER])
664		get_reorder(sch, tb[TCA_NETEM_REORDER]);
665
666	if (tb[TCA_NETEM_CORRUPT])
667		get_corrupt(sch, tb[TCA_NETEM_CORRUPT]);
668
669	q->loss_model = CLG_RANDOM;
670	if (tb[TCA_NETEM_LOSS])
671		ret = get_loss_clg(sch, tb[TCA_NETEM_LOSS]);
672
673	return ret;
674}
 
675
676/*
677 * Special case version of FIFO queue for use by netem.
678 * It queues in order based on timestamps in skb's
679 */
680struct fifo_sched_data {
681	u32 limit;
682	psched_time_t oldest;
683};
684
685static int tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
686{
687	struct fifo_sched_data *q = qdisc_priv(sch);
688	struct sk_buff_head *list = &sch->q;
689	psched_time_t tnext = netem_skb_cb(nskb)->time_to_send;
690	struct sk_buff *skb;
691
692	if (likely(skb_queue_len(list) < q->limit)) {
693		/* Optimize for add at tail */
694		if (likely(skb_queue_empty(list) || tnext >= q->oldest)) {
695			q->oldest = tnext;
696			return qdisc_enqueue_tail(nskb, sch);
697		}
698
699		skb_queue_reverse_walk(list, skb) {
700			const struct netem_skb_cb *cb = netem_skb_cb(skb);
701
702			if (tnext >= cb->time_to_send)
703				break;
704		}
705
706		__skb_queue_after(list, skb, nskb);
 
 
 
 
707
708		sch->qstats.backlog += qdisc_pkt_len(nskb);
 
709
710		return NET_XMIT_SUCCESS;
711	}
712
713	return qdisc_reshape_fail(nskb, sch);
714}
715
716static int tfifo_init(struct Qdisc *sch, struct nlattr *opt)
717{
718	struct fifo_sched_data *q = qdisc_priv(sch);
719
720	if (opt) {
721		struct tc_fifo_qopt *ctl = nla_data(opt);
722		if (nla_len(opt) < sizeof(*ctl))
723			return -EINVAL;
724
725		q->limit = ctl->limit;
726	} else
727		q->limit = max_t(u32, qdisc_dev(sch)->tx_queue_len, 1);
728
729	q->oldest = PSCHED_PASTPERFECT;
730	return 0;
731}
732
733static int tfifo_dump(struct Qdisc *sch, struct sk_buff *skb)
734{
735	struct fifo_sched_data *q = qdisc_priv(sch);
736	struct tc_fifo_qopt opt = { .limit = q->limit };
737
738	NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
739	return skb->len;
740
741nla_put_failure:
742	return -1;
743}
744
745static struct Qdisc_ops tfifo_qdisc_ops __read_mostly = {
746	.id		=	"tfifo",
747	.priv_size	=	sizeof(struct fifo_sched_data),
748	.enqueue	=	tfifo_enqueue,
749	.dequeue	=	qdisc_dequeue_head,
750	.peek		=	qdisc_peek_head,
751	.drop		=	qdisc_queue_drop,
752	.init		=	tfifo_init,
753	.reset		=	qdisc_reset_queue,
754	.change		=	tfifo_init,
755	.dump		=	tfifo_dump,
756};
757
758static int netem_init(struct Qdisc *sch, struct nlattr *opt)
759{
760	struct netem_sched_data *q = qdisc_priv(sch);
761	int ret;
762
 
 
763	if (!opt)
764		return -EINVAL;
765
766	qdisc_watchdog_init(&q->watchdog, sch);
767
768	q->loss_model = CLG_RANDOM;
769	q->qdisc = qdisc_create_dflt(sch->dev_queue, &tfifo_qdisc_ops,
770				     TC_H_MAKE(sch->handle, 1));
771	if (!q->qdisc) {
772		pr_notice("netem: qdisc create tfifo qdisc failed\n");
773		return -ENOMEM;
774	}
775
776	ret = netem_change(sch, opt);
777	if (ret) {
778		pr_info("netem: change failed\n");
779		qdisc_destroy(q->qdisc);
780	}
781	return ret;
782}
783
784static void netem_destroy(struct Qdisc *sch)
785{
786	struct netem_sched_data *q = qdisc_priv(sch);
787
788	qdisc_watchdog_cancel(&q->watchdog);
789	qdisc_destroy(q->qdisc);
 
790	dist_free(q->delay_dist);
 
791}
792
793static int dump_loss_model(const struct netem_sched_data *q,
794			   struct sk_buff *skb)
795{
796	struct nlattr *nest;
797
798	nest = nla_nest_start(skb, TCA_NETEM_LOSS);
799	if (nest == NULL)
800		goto nla_put_failure;
801
802	switch (q->loss_model) {
803	case CLG_RANDOM:
804		/* legacy loss model */
805		nla_nest_cancel(skb, nest);
806		return 0;	/* no data */
807
808	case CLG_4_STATES: {
809		struct tc_netem_gimodel gi = {
810			.p13 = q->clg.a1,
811			.p31 = q->clg.a2,
812			.p32 = q->clg.a3,
813			.p14 = q->clg.a4,
814			.p23 = q->clg.a5,
815		};
816
817		NLA_PUT(skb, NETEM_LOSS_GI, sizeof(gi), &gi);
 
818		break;
819	}
820	case CLG_GILB_ELL: {
821		struct tc_netem_gemodel ge = {
822			.p = q->clg.a1,
823			.r = q->clg.a2,
824			.h = q->clg.a3,
825			.k1 = q->clg.a4,
826		};
827
828		NLA_PUT(skb, NETEM_LOSS_GE, sizeof(ge), &ge);
 
829		break;
830	}
831	}
832
833	nla_nest_end(skb, nest);
834	return 0;
835
836nla_put_failure:
837	nla_nest_cancel(skb, nest);
838	return -1;
839}
840
841static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
842{
843	const struct netem_sched_data *q = qdisc_priv(sch);
844	struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb);
845	struct tc_netem_qopt qopt;
846	struct tc_netem_corr cor;
847	struct tc_netem_reorder reorder;
848	struct tc_netem_corrupt corrupt;
 
 
849
850	qopt.latency = q->latency;
851	qopt.jitter = q->jitter;
 
 
852	qopt.limit = q->limit;
853	qopt.loss = q->loss;
854	qopt.gap = q->gap;
855	qopt.duplicate = q->duplicate;
856	NLA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
 
 
 
 
 
 
 
857
858	cor.delay_corr = q->delay_cor.rho;
859	cor.loss_corr = q->loss_cor.rho;
860	cor.dup_corr = q->dup_cor.rho;
861	NLA_PUT(skb, TCA_NETEM_CORR, sizeof(cor), &cor);
 
862
863	reorder.probability = q->reorder;
864	reorder.correlation = q->reorder_cor.rho;
865	NLA_PUT(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder);
 
866
867	corrupt.probability = q->corrupt;
868	corrupt.correlation = q->corrupt_cor.rho;
869	NLA_PUT(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
870
871	if (dump_loss_model(q, skb) != 0)
872		goto nla_put_failure;
873
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
874	return nla_nest_end(skb, nla);
875
876nla_put_failure:
877	nlmsg_trim(skb, nla);
878	return -1;
879}
880
881static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
882			  struct sk_buff *skb, struct tcmsg *tcm)
883{
884	struct netem_sched_data *q = qdisc_priv(sch);
885
886	if (cl != 1) 	/* only one class */
887		return -ENOENT;
888
889	tcm->tcm_handle |= TC_H_MIN(1);
890	tcm->tcm_info = q->qdisc->handle;
891
892	return 0;
893}
894
895static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
896		     struct Qdisc **old)
897{
898	struct netem_sched_data *q = qdisc_priv(sch);
899
900	if (new == NULL)
901		new = &noop_qdisc;
902
903	sch_tree_lock(sch);
904	*old = q->qdisc;
905	q->qdisc = new;
906	qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
907	qdisc_reset(*old);
908	sch_tree_unlock(sch);
909
910	return 0;
911}
912
913static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
914{
915	struct netem_sched_data *q = qdisc_priv(sch);
916	return q->qdisc;
917}
918
919static unsigned long netem_get(struct Qdisc *sch, u32 classid)
920{
921	return 1;
922}
923
924static void netem_put(struct Qdisc *sch, unsigned long arg)
925{
926}
927
928static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
929{
930	if (!walker->stop) {
931		if (walker->count >= walker->skip)
932			if (walker->fn(sch, 1, walker) < 0) {
933				walker->stop = 1;
934				return;
935			}
936		walker->count++;
937	}
938}
939
940static const struct Qdisc_class_ops netem_class_ops = {
941	.graft		=	netem_graft,
942	.leaf		=	netem_leaf,
943	.get		=	netem_get,
944	.put		=	netem_put,
945	.walk		=	netem_walk,
946	.dump		=	netem_dump_class,
947};
948
949static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
950	.id		=	"netem",
951	.cl_ops		=	&netem_class_ops,
952	.priv_size	=	sizeof(struct netem_sched_data),
953	.enqueue	=	netem_enqueue,
954	.dequeue	=	netem_dequeue,
955	.peek		=	qdisc_peek_dequeued,
956	.drop		=	netem_drop,
957	.init		=	netem_init,
958	.reset		=	netem_reset,
959	.destroy	=	netem_destroy,
960	.change		=	netem_change,
961	.dump		=	netem_dump,
962	.owner		=	THIS_MODULE,
963};
 
964
965
966static int __init netem_module_init(void)
967{
968	pr_info("netem: version " VERSION "\n");
969	return register_qdisc(&netem_qdisc_ops);
970}
971static void __exit netem_module_exit(void)
972{
973	unregister_qdisc(&netem_qdisc_ops);
974}
975module_init(netem_module_init)
976module_exit(netem_module_exit)
977MODULE_LICENSE("GPL");