xref: /openbmc/linux/net/sched/sch_netem.c (revision 5104d265)
1 /*
2  * net/sched/sch_netem.c	Network emulator
3  *
4  * 		This program is free software; you can redistribute it and/or
5  * 		modify it under the terms of the GNU General Public License
6  * 		as published by the Free Software Foundation; either version
7  * 		2 of the License.
8  *
9  *  		Many of the algorithms and ideas for this came from
10  *		NIST Net which is not copyrighted.
11  *
12  * Authors:	Stephen Hemminger <shemminger@osdl.org>
13  *		Catalin(ux aka Dino) BOIE <catab at umbrella dot ro>
14  */
15 
16 #include <linux/mm.h>
17 #include <linux/module.h>
18 #include <linux/slab.h>
19 #include <linux/types.h>
20 #include <linux/kernel.h>
21 #include <linux/errno.h>
22 #include <linux/skbuff.h>
23 #include <linux/vmalloc.h>
24 #include <linux/rtnetlink.h>
25 #include <linux/reciprocal_div.h>
26 #include <linux/rbtree.h>
27 
28 #include <net/netlink.h>
29 #include <net/pkt_sched.h>
30 #include <net/inet_ecn.h>
31 
32 #define VERSION "1.3"
33 
34 /*	Network Emulation Queuing algorithm.
35 	====================================
36 
37 	Sources: [1] Mark Carson, Darrin Santay, "NIST Net - A Linux-based
38 		 Network Emulation Tool
39 		 [2] Luigi Rizzo, DummyNet for FreeBSD
40 
41 	 ----------------------------------------------------------------
42 
43 	 This started out as a simple way to delay outgoing packets to
44 	 test TCP but has grown to include most of the functionality
45 	 of a full blown network emulator like NISTnet. It can delay
46 	 packets and add random jitter (and correlation). The random
47 	 distribution can be loaded from a table as well to provide
48 	 normal, Pareto, or experimental curves. Packet loss,
49 	 duplication, and reordering can also be emulated.
50 
51 	 This qdisc does not do classification that can be handled in
52 	 layering other disciplines.  It does not need to do bandwidth
53 	 control either since that can be handled by using token
54 	 bucket or other rate control.
55 
56      Correlated Loss Generator models
57 
58 	Added generation of correlated loss according to the
59 	"Gilbert-Elliot" model, a 4-state markov model.
60 
61 	References:
62 	[1] NetemCLG Home http://netgroup.uniroma2.it/NetemCLG
63 	[2] S. Salsano, F. Ludovici, A. Ordine, "Definition of a general
64 	and intuitive loss model for packet networks and its implementation
65 	in the Netem module in the Linux kernel", available in [1]
66 
67 	Authors: Stefano Salsano <stefano.salsano at uniroma2.it
68 		 Fabio Ludovici <fabio.ludovici at yahoo.it>
69 */
70 
71 struct netem_sched_data {
72 	/* internal t(ime)fifo qdisc uses t_root and sch->limit */
73 	struct rb_root t_root;
74 
75 	/* optional qdisc for classful handling (NULL at netem init) */
76 	struct Qdisc	*qdisc;
77 
78 	struct qdisc_watchdog watchdog;
79 
80 	psched_tdiff_t latency;
81 	psched_tdiff_t jitter;
82 
83 	u32 loss;
84 	u32 ecn;
85 	u32 limit;
86 	u32 counter;
87 	u32 gap;
88 	u32 duplicate;
89 	u32 reorder;
90 	u32 corrupt;
91 	u32 rate;
92 	s32 packet_overhead;
93 	u32 cell_size;
94 	u32 cell_size_reciprocal;
95 	s32 cell_overhead;
96 
97 	struct crndstate {
98 		u32 last;
99 		u32 rho;
100 	} delay_cor, loss_cor, dup_cor, reorder_cor, corrupt_cor;
101 
102 	struct disttable {
103 		u32  size;
104 		s16 table[0];
105 	} *delay_dist;
106 
107 	enum  {
108 		CLG_RANDOM,
109 		CLG_4_STATES,
110 		CLG_GILB_ELL,
111 	} loss_model;
112 
113 	/* Correlated Loss Generation models */
114 	struct clgstate {
115 		/* state of the Markov chain */
116 		u8 state;
117 
118 		/* 4-states and Gilbert-Elliot models */
119 		u32 a1;	/* p13 for 4-states or p for GE */
120 		u32 a2;	/* p31 for 4-states or r for GE */
121 		u32 a3;	/* p32 for 4-states or h for GE */
122 		u32 a4;	/* p14 for 4-states or 1-k for GE */
123 		u32 a5; /* p23 used only in 4-states */
124 	} clg;
125 
126 };
127 
128 /* Time stamp put into socket buffer control block
129  * Only valid when skbs are in our internal t(ime)fifo queue.
130  */
131 struct netem_skb_cb {
132 	psched_time_t	time_to_send;
133 	ktime_t		tstamp_save;
134 };
135 
136 /* Because space in skb->cb[] is tight, netem overloads skb->next/prev/tstamp
137  * to hold a rb_node structure.
138  *
139  * If struct sk_buff layout is changed, the following checks will complain.
140  */
141 static struct rb_node *netem_rb_node(struct sk_buff *skb)
142 {
143 	BUILD_BUG_ON(offsetof(struct sk_buff, next) != 0);
144 	BUILD_BUG_ON(offsetof(struct sk_buff, prev) !=
145 		     offsetof(struct sk_buff, next) + sizeof(skb->next));
146 	BUILD_BUG_ON(offsetof(struct sk_buff, tstamp) !=
147 		     offsetof(struct sk_buff, prev) + sizeof(skb->prev));
148 	BUILD_BUG_ON(sizeof(struct rb_node) > sizeof(skb->next) +
149 					      sizeof(skb->prev) +
150 					      sizeof(skb->tstamp));
151 	return (struct rb_node *)&skb->next;
152 }
153 
154 static struct sk_buff *netem_rb_to_skb(struct rb_node *rb)
155 {
156 	return (struct sk_buff *)rb;
157 }
158 
159 static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb)
160 {
161 	/* we assume we can use skb next/prev/tstamp as storage for rb_node */
162 	qdisc_cb_private_validate(skb, sizeof(struct netem_skb_cb));
163 	return (struct netem_skb_cb *)qdisc_skb_cb(skb)->data;
164 }
165 
166 /* init_crandom - initialize correlated random number generator
167  * Use entropy source for initial seed.
168  */
169 static void init_crandom(struct crndstate *state, unsigned long rho)
170 {
171 	state->rho = rho;
172 	state->last = net_random();
173 }
174 
175 /* get_crandom - correlated random number generator
176  * Next number depends on last value.
177  * rho is scaled to avoid floating point.
178  */
179 static u32 get_crandom(struct crndstate *state)
180 {
181 	u64 value, rho;
182 	unsigned long answer;
183 
184 	if (state->rho == 0)	/* no correlation */
185 		return net_random();
186 
187 	value = net_random();
188 	rho = (u64)state->rho + 1;
189 	answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32;
190 	state->last = answer;
191 	return answer;
192 }
193 
194 /* loss_4state - 4-state model loss generator
195  * Generates losses according to the 4-state Markov chain adopted in
196  * the GI (General and Intuitive) loss model.
197  */
198 static bool loss_4state(struct netem_sched_data *q)
199 {
200 	struct clgstate *clg = &q->clg;
201 	u32 rnd = net_random();
202 
203 	/*
204 	 * Makes a comparison between rnd and the transition
205 	 * probabilities outgoing from the current state, then decides the
206 	 * next state and if the next packet has to be transmitted or lost.
207 	 * The four states correspond to:
208 	 *   1 => successfully transmitted packets within a gap period
209 	 *   4 => isolated losses within a gap period
210 	 *   3 => lost packets within a burst period
211 	 *   2 => successfully transmitted packets within a burst period
212 	 */
213 	switch (clg->state) {
214 	case 1:
215 		if (rnd < clg->a4) {
216 			clg->state = 4;
217 			return true;
218 		} else if (clg->a4 < rnd && rnd < clg->a1) {
219 			clg->state = 3;
220 			return true;
221 		} else if (clg->a1 < rnd)
222 			clg->state = 1;
223 
224 		break;
225 	case 2:
226 		if (rnd < clg->a5) {
227 			clg->state = 3;
228 			return true;
229 		} else
230 			clg->state = 2;
231 
232 		break;
233 	case 3:
234 		if (rnd < clg->a3)
235 			clg->state = 2;
236 		else if (clg->a3 < rnd && rnd < clg->a2 + clg->a3) {
237 			clg->state = 1;
238 			return true;
239 		} else if (clg->a2 + clg->a3 < rnd) {
240 			clg->state = 3;
241 			return true;
242 		}
243 		break;
244 	case 4:
245 		clg->state = 1;
246 		break;
247 	}
248 
249 	return false;
250 }
251 
252 /* loss_gilb_ell - Gilbert-Elliot model loss generator
253  * Generates losses according to the Gilbert-Elliot loss model or
254  * its special cases  (Gilbert or Simple Gilbert)
255  *
256  * Makes a comparison between random number and the transition
257  * probabilities outgoing from the current state, then decides the
258  * next state. A second random number is extracted and the comparison
259  * with the loss probability of the current state decides if the next
260  * packet will be transmitted or lost.
261  */
262 static bool loss_gilb_ell(struct netem_sched_data *q)
263 {
264 	struct clgstate *clg = &q->clg;
265 
266 	switch (clg->state) {
267 	case 1:
268 		if (net_random() < clg->a1)
269 			clg->state = 2;
270 		if (net_random() < clg->a4)
271 			return true;
272 	case 2:
273 		if (net_random() < clg->a2)
274 			clg->state = 1;
275 		if (clg->a3 > net_random())
276 			return true;
277 	}
278 
279 	return false;
280 }
281 
282 static bool loss_event(struct netem_sched_data *q)
283 {
284 	switch (q->loss_model) {
285 	case CLG_RANDOM:
286 		/* Random packet drop 0 => none, ~0 => all */
287 		return q->loss && q->loss >= get_crandom(&q->loss_cor);
288 
289 	case CLG_4_STATES:
290 		/* 4state loss model algorithm (used also for GI model)
291 		* Extracts a value from the markov 4 state loss generator,
292 		* if it is 1 drops a packet and if needed writes the event in
293 		* the kernel logs
294 		*/
295 		return loss_4state(q);
296 
297 	case CLG_GILB_ELL:
298 		/* Gilbert-Elliot loss model algorithm
299 		* Extracts a value from the Gilbert-Elliot loss generator,
300 		* if it is 1 drops a packet and if needed writes the event in
301 		* the kernel logs
302 		*/
303 		return loss_gilb_ell(q);
304 	}
305 
306 	return false;	/* not reached */
307 }
308 
309 
310 /* tabledist - return a pseudo-randomly distributed value with mean mu and
311  * std deviation sigma.  Uses table lookup to approximate the desired
312  * distribution, and a uniformly-distributed pseudo-random source.
313  */
314 static psched_tdiff_t tabledist(psched_tdiff_t mu, psched_tdiff_t sigma,
315 				struct crndstate *state,
316 				const struct disttable *dist)
317 {
318 	psched_tdiff_t x;
319 	long t;
320 	u32 rnd;
321 
322 	if (sigma == 0)
323 		return mu;
324 
325 	rnd = get_crandom(state);
326 
327 	/* default uniform distribution */
328 	if (dist == NULL)
329 		return (rnd % (2*sigma)) - sigma + mu;
330 
331 	t = dist->table[rnd % dist->size];
332 	x = (sigma % NETEM_DIST_SCALE) * t;
333 	if (x >= 0)
334 		x += NETEM_DIST_SCALE/2;
335 	else
336 		x -= NETEM_DIST_SCALE/2;
337 
338 	return  x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu;
339 }
340 
341 static psched_time_t packet_len_2_sched_time(unsigned int len, struct netem_sched_data *q)
342 {
343 	u64 ticks;
344 
345 	len += q->packet_overhead;
346 
347 	if (q->cell_size) {
348 		u32 cells = reciprocal_divide(len, q->cell_size_reciprocal);
349 
350 		if (len > cells * q->cell_size)	/* extra cell needed for remainder */
351 			cells++;
352 		len = cells * (q->cell_size + q->cell_overhead);
353 	}
354 
355 	ticks = (u64)len * NSEC_PER_SEC;
356 
357 	do_div(ticks, q->rate);
358 	return PSCHED_NS2TICKS(ticks);
359 }
360 
361 static void tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
362 {
363 	struct netem_sched_data *q = qdisc_priv(sch);
364 	psched_time_t tnext = netem_skb_cb(nskb)->time_to_send;
365 	struct rb_node **p = &q->t_root.rb_node, *parent = NULL;
366 
367 	while (*p) {
368 		struct sk_buff *skb;
369 
370 		parent = *p;
371 		skb = netem_rb_to_skb(parent);
372 		if (tnext >= netem_skb_cb(skb)->time_to_send)
373 			p = &parent->rb_right;
374 		else
375 			p = &parent->rb_left;
376 	}
377 	rb_link_node(netem_rb_node(nskb), parent, p);
378 	rb_insert_color(netem_rb_node(nskb), &q->t_root);
379 	sch->q.qlen++;
380 }
381 
382 /*
383  * Insert one skb into qdisc.
384  * Note: parent depends on return value to account for queue length.
385  * 	NET_XMIT_DROP: queue length didn't change.
386  *      NET_XMIT_SUCCESS: one skb was queued.
387  */
388 static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch)
389 {
390 	struct netem_sched_data *q = qdisc_priv(sch);
391 	/* We don't fill cb now as skb_unshare() may invalidate it */
392 	struct netem_skb_cb *cb;
393 	struct sk_buff *skb2;
394 	int count = 1;
395 
396 	/* Random duplication */
397 	if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor))
398 		++count;
399 
400 	/* Drop packet? */
401 	if (loss_event(q)) {
402 		if (q->ecn && INET_ECN_set_ce(skb))
403 			sch->qstats.drops++; /* mark packet */
404 		else
405 			--count;
406 	}
407 	if (count == 0) {
408 		sch->qstats.drops++;
409 		kfree_skb(skb);
410 		return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
411 	}
412 
413 	/* If a delay is expected, orphan the skb. (orphaning usually takes
414 	 * place at TX completion time, so _before_ the link transit delay)
415 	 * Ideally, this orphaning should be done after the rate limiting
416 	 * module, because this breaks TCP Small Queue, and other mechanisms
417 	 * based on socket sk_wmem_alloc.
418 	 */
419 	if (q->latency || q->jitter)
420 		skb_orphan(skb);
421 
422 	/*
423 	 * If we need to duplicate packet, then re-insert at top of the
424 	 * qdisc tree, since parent queuer expects that only one
425 	 * skb will be queued.
426 	 */
427 	if (count > 1 && (skb2 = skb_clone(skb, GFP_ATOMIC)) != NULL) {
428 		struct Qdisc *rootq = qdisc_root(sch);
429 		u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
430 		q->duplicate = 0;
431 
432 		qdisc_enqueue_root(skb2, rootq);
433 		q->duplicate = dupsave;
434 	}
435 
436 	/*
437 	 * Randomized packet corruption.
438 	 * Make copy if needed since we are modifying
439 	 * If packet is going to be hardware checksummed, then
440 	 * do it now in software before we mangle it.
441 	 */
442 	if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor)) {
443 		if (!(skb = skb_unshare(skb, GFP_ATOMIC)) ||
444 		    (skb->ip_summed == CHECKSUM_PARTIAL &&
445 		     skb_checksum_help(skb)))
446 			return qdisc_drop(skb, sch);
447 
448 		skb->data[net_random() % skb_headlen(skb)] ^= 1<<(net_random() % 8);
449 	}
450 
451 	if (unlikely(skb_queue_len(&sch->q) >= sch->limit))
452 		return qdisc_reshape_fail(skb, sch);
453 
454 	sch->qstats.backlog += qdisc_pkt_len(skb);
455 
456 	cb = netem_skb_cb(skb);
457 	if (q->gap == 0 ||		/* not doing reordering */
458 	    q->counter < q->gap - 1 ||	/* inside last reordering gap */
459 	    q->reorder < get_crandom(&q->reorder_cor)) {
460 		psched_time_t now;
461 		psched_tdiff_t delay;
462 
463 		delay = tabledist(q->latency, q->jitter,
464 				  &q->delay_cor, q->delay_dist);
465 
466 		now = psched_get_time();
467 
468 		if (q->rate) {
469 			struct sk_buff *last;
470 
471 			if (!skb_queue_empty(&sch->q))
472 				last = skb_peek_tail(&sch->q);
473 			else
474 				last = netem_rb_to_skb(rb_last(&q->t_root));
475 			if (last) {
476 				/*
477 				 * Last packet in queue is reference point (now),
478 				 * calculate this time bonus and subtract
479 				 * from delay.
480 				 */
481 				delay -= netem_skb_cb(last)->time_to_send - now;
482 				delay = max_t(psched_tdiff_t, 0, delay);
483 				now = netem_skb_cb(last)->time_to_send;
484 			}
485 
486 			delay += packet_len_2_sched_time(skb->len, q);
487 		}
488 
489 		cb->time_to_send = now + delay;
490 		cb->tstamp_save = skb->tstamp;
491 		++q->counter;
492 		tfifo_enqueue(skb, sch);
493 	} else {
494 		/*
495 		 * Do re-ordering by putting one out of N packets at the front
496 		 * of the queue.
497 		 */
498 		cb->time_to_send = psched_get_time();
499 		q->counter = 0;
500 
501 		__skb_queue_head(&sch->q, skb);
502 		sch->qstats.requeues++;
503 	}
504 
505 	return NET_XMIT_SUCCESS;
506 }
507 
508 static unsigned int netem_drop(struct Qdisc *sch)
509 {
510 	struct netem_sched_data *q = qdisc_priv(sch);
511 	unsigned int len;
512 
513 	len = qdisc_queue_drop(sch);
514 
515 	if (!len) {
516 		struct rb_node *p = rb_first(&q->t_root);
517 
518 		if (p) {
519 			struct sk_buff *skb = netem_rb_to_skb(p);
520 
521 			rb_erase(p, &q->t_root);
522 			sch->q.qlen--;
523 			skb->next = NULL;
524 			skb->prev = NULL;
525 			len = qdisc_pkt_len(skb);
526 			kfree_skb(skb);
527 		}
528 	}
529 	if (!len && q->qdisc && q->qdisc->ops->drop)
530 	    len = q->qdisc->ops->drop(q->qdisc);
531 	if (len)
532 		sch->qstats.drops++;
533 
534 	return len;
535 }
536 
537 static struct sk_buff *netem_dequeue(struct Qdisc *sch)
538 {
539 	struct netem_sched_data *q = qdisc_priv(sch);
540 	struct sk_buff *skb;
541 	struct rb_node *p;
542 
543 	if (qdisc_is_throttled(sch))
544 		return NULL;
545 
546 tfifo_dequeue:
547 	skb = __skb_dequeue(&sch->q);
548 	if (skb) {
549 deliver:
550 		sch->qstats.backlog -= qdisc_pkt_len(skb);
551 		qdisc_unthrottled(sch);
552 		qdisc_bstats_update(sch, skb);
553 		return skb;
554 	}
555 	p = rb_first(&q->t_root);
556 	if (p) {
557 		psched_time_t time_to_send;
558 
559 		skb = netem_rb_to_skb(p);
560 
561 		/* if more time remaining? */
562 		time_to_send = netem_skb_cb(skb)->time_to_send;
563 		if (time_to_send <= psched_get_time()) {
564 			rb_erase(p, &q->t_root);
565 
566 			sch->q.qlen--;
567 			skb->next = NULL;
568 			skb->prev = NULL;
569 			skb->tstamp = netem_skb_cb(skb)->tstamp_save;
570 
571 #ifdef CONFIG_NET_CLS_ACT
572 			/*
573 			 * If it's at ingress let's pretend the delay is
574 			 * from the network (tstamp will be updated).
575 			 */
576 			if (G_TC_FROM(skb->tc_verd) & AT_INGRESS)
577 				skb->tstamp.tv64 = 0;
578 #endif
579 
580 			if (q->qdisc) {
581 				int err = qdisc_enqueue(skb, q->qdisc);
582 
583 				if (unlikely(err != NET_XMIT_SUCCESS)) {
584 					if (net_xmit_drop_count(err)) {
585 						sch->qstats.drops++;
586 						qdisc_tree_decrease_qlen(sch, 1);
587 					}
588 				}
589 				goto tfifo_dequeue;
590 			}
591 			goto deliver;
592 		}
593 
594 		if (q->qdisc) {
595 			skb = q->qdisc->ops->dequeue(q->qdisc);
596 			if (skb)
597 				goto deliver;
598 		}
599 		qdisc_watchdog_schedule(&q->watchdog, time_to_send);
600 	}
601 
602 	if (q->qdisc) {
603 		skb = q->qdisc->ops->dequeue(q->qdisc);
604 		if (skb)
605 			goto deliver;
606 	}
607 	return NULL;
608 }
609 
610 static void netem_reset(struct Qdisc *sch)
611 {
612 	struct netem_sched_data *q = qdisc_priv(sch);
613 
614 	qdisc_reset_queue(sch);
615 	if (q->qdisc)
616 		qdisc_reset(q->qdisc);
617 	qdisc_watchdog_cancel(&q->watchdog);
618 }
619 
620 static void dist_free(struct disttable *d)
621 {
622 	if (d) {
623 		if (is_vmalloc_addr(d))
624 			vfree(d);
625 		else
626 			kfree(d);
627 	}
628 }
629 
630 /*
631  * Distribution data is a variable size payload containing
632  * signed 16 bit values.
633  */
634 static int get_dist_table(struct Qdisc *sch, const struct nlattr *attr)
635 {
636 	struct netem_sched_data *q = qdisc_priv(sch);
637 	size_t n = nla_len(attr)/sizeof(__s16);
638 	const __s16 *data = nla_data(attr);
639 	spinlock_t *root_lock;
640 	struct disttable *d;
641 	int i;
642 	size_t s;
643 
644 	if (n > NETEM_DIST_MAX)
645 		return -EINVAL;
646 
647 	s = sizeof(struct disttable) + n * sizeof(s16);
648 	d = kmalloc(s, GFP_KERNEL | __GFP_NOWARN);
649 	if (!d)
650 		d = vmalloc(s);
651 	if (!d)
652 		return -ENOMEM;
653 
654 	d->size = n;
655 	for (i = 0; i < n; i++)
656 		d->table[i] = data[i];
657 
658 	root_lock = qdisc_root_sleeping_lock(sch);
659 
660 	spin_lock_bh(root_lock);
661 	swap(q->delay_dist, d);
662 	spin_unlock_bh(root_lock);
663 
664 	dist_free(d);
665 	return 0;
666 }
667 
668 static void get_correlation(struct Qdisc *sch, const struct nlattr *attr)
669 {
670 	struct netem_sched_data *q = qdisc_priv(sch);
671 	const struct tc_netem_corr *c = nla_data(attr);
672 
673 	init_crandom(&q->delay_cor, c->delay_corr);
674 	init_crandom(&q->loss_cor, c->loss_corr);
675 	init_crandom(&q->dup_cor, c->dup_corr);
676 }
677 
678 static void get_reorder(struct Qdisc *sch, const struct nlattr *attr)
679 {
680 	struct netem_sched_data *q = qdisc_priv(sch);
681 	const struct tc_netem_reorder *r = nla_data(attr);
682 
683 	q->reorder = r->probability;
684 	init_crandom(&q->reorder_cor, r->correlation);
685 }
686 
687 static void get_corrupt(struct Qdisc *sch, const struct nlattr *attr)
688 {
689 	struct netem_sched_data *q = qdisc_priv(sch);
690 	const struct tc_netem_corrupt *r = nla_data(attr);
691 
692 	q->corrupt = r->probability;
693 	init_crandom(&q->corrupt_cor, r->correlation);
694 }
695 
696 static void get_rate(struct Qdisc *sch, const struct nlattr *attr)
697 {
698 	struct netem_sched_data *q = qdisc_priv(sch);
699 	const struct tc_netem_rate *r = nla_data(attr);
700 
701 	q->rate = r->rate;
702 	q->packet_overhead = r->packet_overhead;
703 	q->cell_size = r->cell_size;
704 	if (q->cell_size)
705 		q->cell_size_reciprocal = reciprocal_value(q->cell_size);
706 	q->cell_overhead = r->cell_overhead;
707 }
708 
709 static int get_loss_clg(struct Qdisc *sch, const struct nlattr *attr)
710 {
711 	struct netem_sched_data *q = qdisc_priv(sch);
712 	const struct nlattr *la;
713 	int rem;
714 
715 	nla_for_each_nested(la, attr, rem) {
716 		u16 type = nla_type(la);
717 
718 		switch(type) {
719 		case NETEM_LOSS_GI: {
720 			const struct tc_netem_gimodel *gi = nla_data(la);
721 
722 			if (nla_len(la) < sizeof(struct tc_netem_gimodel)) {
723 				pr_info("netem: incorrect gi model size\n");
724 				return -EINVAL;
725 			}
726 
727 			q->loss_model = CLG_4_STATES;
728 
729 			q->clg.state = 1;
730 			q->clg.a1 = gi->p13;
731 			q->clg.a2 = gi->p31;
732 			q->clg.a3 = gi->p32;
733 			q->clg.a4 = gi->p14;
734 			q->clg.a5 = gi->p23;
735 			break;
736 		}
737 
738 		case NETEM_LOSS_GE: {
739 			const struct tc_netem_gemodel *ge = nla_data(la);
740 
741 			if (nla_len(la) < sizeof(struct tc_netem_gemodel)) {
742 				pr_info("netem: incorrect ge model size\n");
743 				return -EINVAL;
744 			}
745 
746 			q->loss_model = CLG_GILB_ELL;
747 			q->clg.state = 1;
748 			q->clg.a1 = ge->p;
749 			q->clg.a2 = ge->r;
750 			q->clg.a3 = ge->h;
751 			q->clg.a4 = ge->k1;
752 			break;
753 		}
754 
755 		default:
756 			pr_info("netem: unknown loss type %u\n", type);
757 			return -EINVAL;
758 		}
759 	}
760 
761 	return 0;
762 }
763 
764 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
765 	[TCA_NETEM_CORR]	= { .len = sizeof(struct tc_netem_corr) },
766 	[TCA_NETEM_REORDER]	= { .len = sizeof(struct tc_netem_reorder) },
767 	[TCA_NETEM_CORRUPT]	= { .len = sizeof(struct tc_netem_corrupt) },
768 	[TCA_NETEM_RATE]	= { .len = sizeof(struct tc_netem_rate) },
769 	[TCA_NETEM_LOSS]	= { .type = NLA_NESTED },
770 	[TCA_NETEM_ECN]		= { .type = NLA_U32 },
771 };
772 
773 static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
774 		      const struct nla_policy *policy, int len)
775 {
776 	int nested_len = nla_len(nla) - NLA_ALIGN(len);
777 
778 	if (nested_len < 0) {
779 		pr_info("netem: invalid attributes len %d\n", nested_len);
780 		return -EINVAL;
781 	}
782 
783 	if (nested_len >= nla_attr_size(0))
784 		return nla_parse(tb, maxtype, nla_data(nla) + NLA_ALIGN(len),
785 				 nested_len, policy);
786 
787 	memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
788 	return 0;
789 }
790 
791 /* Parse netlink message to set options */
792 static int netem_change(struct Qdisc *sch, struct nlattr *opt)
793 {
794 	struct netem_sched_data *q = qdisc_priv(sch);
795 	struct nlattr *tb[TCA_NETEM_MAX + 1];
796 	struct tc_netem_qopt *qopt;
797 	int ret;
798 
799 	if (opt == NULL)
800 		return -EINVAL;
801 
802 	qopt = nla_data(opt);
803 	ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
804 	if (ret < 0)
805 		return ret;
806 
807 	sch->limit = qopt->limit;
808 
809 	q->latency = qopt->latency;
810 	q->jitter = qopt->jitter;
811 	q->limit = qopt->limit;
812 	q->gap = qopt->gap;
813 	q->counter = 0;
814 	q->loss = qopt->loss;
815 	q->duplicate = qopt->duplicate;
816 
817 	/* for compatibility with earlier versions.
818 	 * if gap is set, need to assume 100% probability
819 	 */
820 	if (q->gap)
821 		q->reorder = ~0;
822 
823 	if (tb[TCA_NETEM_CORR])
824 		get_correlation(sch, tb[TCA_NETEM_CORR]);
825 
826 	if (tb[TCA_NETEM_DELAY_DIST]) {
827 		ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST]);
828 		if (ret)
829 			return ret;
830 	}
831 
832 	if (tb[TCA_NETEM_REORDER])
833 		get_reorder(sch, tb[TCA_NETEM_REORDER]);
834 
835 	if (tb[TCA_NETEM_CORRUPT])
836 		get_corrupt(sch, tb[TCA_NETEM_CORRUPT]);
837 
838 	if (tb[TCA_NETEM_RATE])
839 		get_rate(sch, tb[TCA_NETEM_RATE]);
840 
841 	if (tb[TCA_NETEM_ECN])
842 		q->ecn = nla_get_u32(tb[TCA_NETEM_ECN]);
843 
844 	q->loss_model = CLG_RANDOM;
845 	if (tb[TCA_NETEM_LOSS])
846 		ret = get_loss_clg(sch, tb[TCA_NETEM_LOSS]);
847 
848 	return ret;
849 }
850 
851 static int netem_init(struct Qdisc *sch, struct nlattr *opt)
852 {
853 	struct netem_sched_data *q = qdisc_priv(sch);
854 	int ret;
855 
856 	if (!opt)
857 		return -EINVAL;
858 
859 	qdisc_watchdog_init(&q->watchdog, sch);
860 
861 	q->loss_model = CLG_RANDOM;
862 	ret = netem_change(sch, opt);
863 	if (ret)
864 		pr_info("netem: change failed\n");
865 	return ret;
866 }
867 
868 static void netem_destroy(struct Qdisc *sch)
869 {
870 	struct netem_sched_data *q = qdisc_priv(sch);
871 
872 	qdisc_watchdog_cancel(&q->watchdog);
873 	if (q->qdisc)
874 		qdisc_destroy(q->qdisc);
875 	dist_free(q->delay_dist);
876 }
877 
878 static int dump_loss_model(const struct netem_sched_data *q,
879 			   struct sk_buff *skb)
880 {
881 	struct nlattr *nest;
882 
883 	nest = nla_nest_start(skb, TCA_NETEM_LOSS);
884 	if (nest == NULL)
885 		goto nla_put_failure;
886 
887 	switch (q->loss_model) {
888 	case CLG_RANDOM:
889 		/* legacy loss model */
890 		nla_nest_cancel(skb, nest);
891 		return 0;	/* no data */
892 
893 	case CLG_4_STATES: {
894 		struct tc_netem_gimodel gi = {
895 			.p13 = q->clg.a1,
896 			.p31 = q->clg.a2,
897 			.p32 = q->clg.a3,
898 			.p14 = q->clg.a4,
899 			.p23 = q->clg.a5,
900 		};
901 
902 		if (nla_put(skb, NETEM_LOSS_GI, sizeof(gi), &gi))
903 			goto nla_put_failure;
904 		break;
905 	}
906 	case CLG_GILB_ELL: {
907 		struct tc_netem_gemodel ge = {
908 			.p = q->clg.a1,
909 			.r = q->clg.a2,
910 			.h = q->clg.a3,
911 			.k1 = q->clg.a4,
912 		};
913 
914 		if (nla_put(skb, NETEM_LOSS_GE, sizeof(ge), &ge))
915 			goto nla_put_failure;
916 		break;
917 	}
918 	}
919 
920 	nla_nest_end(skb, nest);
921 	return 0;
922 
923 nla_put_failure:
924 	nla_nest_cancel(skb, nest);
925 	return -1;
926 }
927 
928 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
929 {
930 	const struct netem_sched_data *q = qdisc_priv(sch);
931 	struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb);
932 	struct tc_netem_qopt qopt;
933 	struct tc_netem_corr cor;
934 	struct tc_netem_reorder reorder;
935 	struct tc_netem_corrupt corrupt;
936 	struct tc_netem_rate rate;
937 
938 	qopt.latency = q->latency;
939 	qopt.jitter = q->jitter;
940 	qopt.limit = q->limit;
941 	qopt.loss = q->loss;
942 	qopt.gap = q->gap;
943 	qopt.duplicate = q->duplicate;
944 	if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
945 		goto nla_put_failure;
946 
947 	cor.delay_corr = q->delay_cor.rho;
948 	cor.loss_corr = q->loss_cor.rho;
949 	cor.dup_corr = q->dup_cor.rho;
950 	if (nla_put(skb, TCA_NETEM_CORR, sizeof(cor), &cor))
951 		goto nla_put_failure;
952 
953 	reorder.probability = q->reorder;
954 	reorder.correlation = q->reorder_cor.rho;
955 	if (nla_put(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder))
956 		goto nla_put_failure;
957 
958 	corrupt.probability = q->corrupt;
959 	corrupt.correlation = q->corrupt_cor.rho;
960 	if (nla_put(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt))
961 		goto nla_put_failure;
962 
963 	rate.rate = q->rate;
964 	rate.packet_overhead = q->packet_overhead;
965 	rate.cell_size = q->cell_size;
966 	rate.cell_overhead = q->cell_overhead;
967 	if (nla_put(skb, TCA_NETEM_RATE, sizeof(rate), &rate))
968 		goto nla_put_failure;
969 
970 	if (q->ecn && nla_put_u32(skb, TCA_NETEM_ECN, q->ecn))
971 		goto nla_put_failure;
972 
973 	if (dump_loss_model(q, skb) != 0)
974 		goto nla_put_failure;
975 
976 	return nla_nest_end(skb, nla);
977 
978 nla_put_failure:
979 	nlmsg_trim(skb, nla);
980 	return -1;
981 }
982 
983 static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
984 			  struct sk_buff *skb, struct tcmsg *tcm)
985 {
986 	struct netem_sched_data *q = qdisc_priv(sch);
987 
988 	if (cl != 1 || !q->qdisc) 	/* only one class */
989 		return -ENOENT;
990 
991 	tcm->tcm_handle |= TC_H_MIN(1);
992 	tcm->tcm_info = q->qdisc->handle;
993 
994 	return 0;
995 }
996 
997 static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
998 		     struct Qdisc **old)
999 {
1000 	struct netem_sched_data *q = qdisc_priv(sch);
1001 
1002 	sch_tree_lock(sch);
1003 	*old = q->qdisc;
1004 	q->qdisc = new;
1005 	if (*old) {
1006 		qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
1007 		qdisc_reset(*old);
1008 	}
1009 	sch_tree_unlock(sch);
1010 
1011 	return 0;
1012 }
1013 
1014 static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
1015 {
1016 	struct netem_sched_data *q = qdisc_priv(sch);
1017 	return q->qdisc;
1018 }
1019 
1020 static unsigned long netem_get(struct Qdisc *sch, u32 classid)
1021 {
1022 	return 1;
1023 }
1024 
1025 static void netem_put(struct Qdisc *sch, unsigned long arg)
1026 {
1027 }
1028 
1029 static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
1030 {
1031 	if (!walker->stop) {
1032 		if (walker->count >= walker->skip)
1033 			if (walker->fn(sch, 1, walker) < 0) {
1034 				walker->stop = 1;
1035 				return;
1036 			}
1037 		walker->count++;
1038 	}
1039 }
1040 
1041 static const struct Qdisc_class_ops netem_class_ops = {
1042 	.graft		=	netem_graft,
1043 	.leaf		=	netem_leaf,
1044 	.get		=	netem_get,
1045 	.put		=	netem_put,
1046 	.walk		=	netem_walk,
1047 	.dump		=	netem_dump_class,
1048 };
1049 
1050 static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
1051 	.id		=	"netem",
1052 	.cl_ops		=	&netem_class_ops,
1053 	.priv_size	=	sizeof(struct netem_sched_data),
1054 	.enqueue	=	netem_enqueue,
1055 	.dequeue	=	netem_dequeue,
1056 	.peek		=	qdisc_peek_dequeued,
1057 	.drop		=	netem_drop,
1058 	.init		=	netem_init,
1059 	.reset		=	netem_reset,
1060 	.destroy	=	netem_destroy,
1061 	.change		=	netem_change,
1062 	.dump		=	netem_dump,
1063 	.owner		=	THIS_MODULE,
1064 };
1065 
1066 
1067 static int __init netem_module_init(void)
1068 {
1069 	pr_info("netem: version " VERSION "\n");
1070 	return register_qdisc(&netem_qdisc_ops);
1071 }
1072 static void __exit netem_module_exit(void)
1073 {
1074 	unregister_qdisc(&netem_qdisc_ops);
1075 }
1076 module_init(netem_module_init)
1077 module_exit(netem_module_exit)
1078 MODULE_LICENSE("GPL");
1079