xref: /openbmc/linux/net/sched/sch_netem.c (revision 161f4089)
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_reset(struct Qdisc *sch)
362 {
363 	struct netem_sched_data *q = qdisc_priv(sch);
364 	struct rb_node *p;
365 
366 	while ((p = rb_first(&q->t_root))) {
367 		struct sk_buff *skb = netem_rb_to_skb(p);
368 
369 		rb_erase(p, &q->t_root);
370 		skb->next = NULL;
371 		skb->prev = NULL;
372 		kfree_skb(skb);
373 	}
374 }
375 
376 static void tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
377 {
378 	struct netem_sched_data *q = qdisc_priv(sch);
379 	psched_time_t tnext = netem_skb_cb(nskb)->time_to_send;
380 	struct rb_node **p = &q->t_root.rb_node, *parent = NULL;
381 
382 	while (*p) {
383 		struct sk_buff *skb;
384 
385 		parent = *p;
386 		skb = netem_rb_to_skb(parent);
387 		if (tnext >= netem_skb_cb(skb)->time_to_send)
388 			p = &parent->rb_right;
389 		else
390 			p = &parent->rb_left;
391 	}
392 	rb_link_node(netem_rb_node(nskb), parent, p);
393 	rb_insert_color(netem_rb_node(nskb), &q->t_root);
394 	sch->q.qlen++;
395 }
396 
397 /*
398  * Insert one skb into qdisc.
399  * Note: parent depends on return value to account for queue length.
400  * 	NET_XMIT_DROP: queue length didn't change.
401  *      NET_XMIT_SUCCESS: one skb was queued.
402  */
403 static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch)
404 {
405 	struct netem_sched_data *q = qdisc_priv(sch);
406 	/* We don't fill cb now as skb_unshare() may invalidate it */
407 	struct netem_skb_cb *cb;
408 	struct sk_buff *skb2;
409 	int count = 1;
410 
411 	/* Random duplication */
412 	if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor))
413 		++count;
414 
415 	/* Drop packet? */
416 	if (loss_event(q)) {
417 		if (q->ecn && INET_ECN_set_ce(skb))
418 			sch->qstats.drops++; /* mark packet */
419 		else
420 			--count;
421 	}
422 	if (count == 0) {
423 		sch->qstats.drops++;
424 		kfree_skb(skb);
425 		return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
426 	}
427 
428 	/* If a delay is expected, orphan the skb. (orphaning usually takes
429 	 * place at TX completion time, so _before_ the link transit delay)
430 	 */
431 	if (q->latency || q->jitter)
432 		skb_orphan_partial(skb);
433 
434 	/*
435 	 * If we need to duplicate packet, then re-insert at top of the
436 	 * qdisc tree, since parent queuer expects that only one
437 	 * skb will be queued.
438 	 */
439 	if (count > 1 && (skb2 = skb_clone(skb, GFP_ATOMIC)) != NULL) {
440 		struct Qdisc *rootq = qdisc_root(sch);
441 		u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
442 		q->duplicate = 0;
443 
444 		qdisc_enqueue_root(skb2, rootq);
445 		q->duplicate = dupsave;
446 	}
447 
448 	/*
449 	 * Randomized packet corruption.
450 	 * Make copy if needed since we are modifying
451 	 * If packet is going to be hardware checksummed, then
452 	 * do it now in software before we mangle it.
453 	 */
454 	if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor)) {
455 		if (!(skb = skb_unshare(skb, GFP_ATOMIC)) ||
456 		    (skb->ip_summed == CHECKSUM_PARTIAL &&
457 		     skb_checksum_help(skb)))
458 			return qdisc_drop(skb, sch);
459 
460 		skb->data[net_random() % skb_headlen(skb)] ^= 1<<(net_random() % 8);
461 	}
462 
463 	if (unlikely(skb_queue_len(&sch->q) >= sch->limit))
464 		return qdisc_reshape_fail(skb, sch);
465 
466 	sch->qstats.backlog += qdisc_pkt_len(skb);
467 
468 	cb = netem_skb_cb(skb);
469 	if (q->gap == 0 ||		/* not doing reordering */
470 	    q->counter < q->gap - 1 ||	/* inside last reordering gap */
471 	    q->reorder < get_crandom(&q->reorder_cor)) {
472 		psched_time_t now;
473 		psched_tdiff_t delay;
474 
475 		delay = tabledist(q->latency, q->jitter,
476 				  &q->delay_cor, q->delay_dist);
477 
478 		now = psched_get_time();
479 
480 		if (q->rate) {
481 			struct sk_buff *last;
482 
483 			if (!skb_queue_empty(&sch->q))
484 				last = skb_peek_tail(&sch->q);
485 			else
486 				last = netem_rb_to_skb(rb_last(&q->t_root));
487 			if (last) {
488 				/*
489 				 * Last packet in queue is reference point (now),
490 				 * calculate this time bonus and subtract
491 				 * from delay.
492 				 */
493 				delay -= netem_skb_cb(last)->time_to_send - now;
494 				delay = max_t(psched_tdiff_t, 0, delay);
495 				now = netem_skb_cb(last)->time_to_send;
496 			}
497 
498 			delay += packet_len_2_sched_time(skb->len, q);
499 		}
500 
501 		cb->time_to_send = now + delay;
502 		cb->tstamp_save = skb->tstamp;
503 		++q->counter;
504 		tfifo_enqueue(skb, sch);
505 	} else {
506 		/*
507 		 * Do re-ordering by putting one out of N packets at the front
508 		 * of the queue.
509 		 */
510 		cb->time_to_send = psched_get_time();
511 		q->counter = 0;
512 
513 		__skb_queue_head(&sch->q, skb);
514 		sch->qstats.requeues++;
515 	}
516 
517 	return NET_XMIT_SUCCESS;
518 }
519 
520 static unsigned int netem_drop(struct Qdisc *sch)
521 {
522 	struct netem_sched_data *q = qdisc_priv(sch);
523 	unsigned int len;
524 
525 	len = qdisc_queue_drop(sch);
526 
527 	if (!len) {
528 		struct rb_node *p = rb_first(&q->t_root);
529 
530 		if (p) {
531 			struct sk_buff *skb = netem_rb_to_skb(p);
532 
533 			rb_erase(p, &q->t_root);
534 			sch->q.qlen--;
535 			skb->next = NULL;
536 			skb->prev = NULL;
537 			len = qdisc_pkt_len(skb);
538 			sch->qstats.backlog -= len;
539 			kfree_skb(skb);
540 		}
541 	}
542 	if (!len && q->qdisc && q->qdisc->ops->drop)
543 	    len = q->qdisc->ops->drop(q->qdisc);
544 	if (len)
545 		sch->qstats.drops++;
546 
547 	return len;
548 }
549 
550 static struct sk_buff *netem_dequeue(struct Qdisc *sch)
551 {
552 	struct netem_sched_data *q = qdisc_priv(sch);
553 	struct sk_buff *skb;
554 	struct rb_node *p;
555 
556 	if (qdisc_is_throttled(sch))
557 		return NULL;
558 
559 tfifo_dequeue:
560 	skb = __skb_dequeue(&sch->q);
561 	if (skb) {
562 deliver:
563 		sch->qstats.backlog -= qdisc_pkt_len(skb);
564 		qdisc_unthrottled(sch);
565 		qdisc_bstats_update(sch, skb);
566 		return skb;
567 	}
568 	p = rb_first(&q->t_root);
569 	if (p) {
570 		psched_time_t time_to_send;
571 
572 		skb = netem_rb_to_skb(p);
573 
574 		/* if more time remaining? */
575 		time_to_send = netem_skb_cb(skb)->time_to_send;
576 		if (time_to_send <= psched_get_time()) {
577 			rb_erase(p, &q->t_root);
578 
579 			sch->q.qlen--;
580 			skb->next = NULL;
581 			skb->prev = NULL;
582 			skb->tstamp = netem_skb_cb(skb)->tstamp_save;
583 
584 #ifdef CONFIG_NET_CLS_ACT
585 			/*
586 			 * If it's at ingress let's pretend the delay is
587 			 * from the network (tstamp will be updated).
588 			 */
589 			if (G_TC_FROM(skb->tc_verd) & AT_INGRESS)
590 				skb->tstamp.tv64 = 0;
591 #endif
592 
593 			if (q->qdisc) {
594 				int err = qdisc_enqueue(skb, q->qdisc);
595 
596 				if (unlikely(err != NET_XMIT_SUCCESS)) {
597 					if (net_xmit_drop_count(err)) {
598 						sch->qstats.drops++;
599 						qdisc_tree_decrease_qlen(sch, 1);
600 					}
601 				}
602 				goto tfifo_dequeue;
603 			}
604 			goto deliver;
605 		}
606 
607 		if (q->qdisc) {
608 			skb = q->qdisc->ops->dequeue(q->qdisc);
609 			if (skb)
610 				goto deliver;
611 		}
612 		qdisc_watchdog_schedule(&q->watchdog, time_to_send);
613 	}
614 
615 	if (q->qdisc) {
616 		skb = q->qdisc->ops->dequeue(q->qdisc);
617 		if (skb)
618 			goto deliver;
619 	}
620 	return NULL;
621 }
622 
623 static void netem_reset(struct Qdisc *sch)
624 {
625 	struct netem_sched_data *q = qdisc_priv(sch);
626 
627 	qdisc_reset_queue(sch);
628 	tfifo_reset(sch);
629 	if (q->qdisc)
630 		qdisc_reset(q->qdisc);
631 	qdisc_watchdog_cancel(&q->watchdog);
632 }
633 
634 static void dist_free(struct disttable *d)
635 {
636 	if (d) {
637 		if (is_vmalloc_addr(d))
638 			vfree(d);
639 		else
640 			kfree(d);
641 	}
642 }
643 
644 /*
645  * Distribution data is a variable size payload containing
646  * signed 16 bit values.
647  */
648 static int get_dist_table(struct Qdisc *sch, const struct nlattr *attr)
649 {
650 	struct netem_sched_data *q = qdisc_priv(sch);
651 	size_t n = nla_len(attr)/sizeof(__s16);
652 	const __s16 *data = nla_data(attr);
653 	spinlock_t *root_lock;
654 	struct disttable *d;
655 	int i;
656 	size_t s;
657 
658 	if (n > NETEM_DIST_MAX)
659 		return -EINVAL;
660 
661 	s = sizeof(struct disttable) + n * sizeof(s16);
662 	d = kmalloc(s, GFP_KERNEL | __GFP_NOWARN);
663 	if (!d)
664 		d = vmalloc(s);
665 	if (!d)
666 		return -ENOMEM;
667 
668 	d->size = n;
669 	for (i = 0; i < n; i++)
670 		d->table[i] = data[i];
671 
672 	root_lock = qdisc_root_sleeping_lock(sch);
673 
674 	spin_lock_bh(root_lock);
675 	swap(q->delay_dist, d);
676 	spin_unlock_bh(root_lock);
677 
678 	dist_free(d);
679 	return 0;
680 }
681 
682 static void get_correlation(struct Qdisc *sch, const struct nlattr *attr)
683 {
684 	struct netem_sched_data *q = qdisc_priv(sch);
685 	const struct tc_netem_corr *c = nla_data(attr);
686 
687 	init_crandom(&q->delay_cor, c->delay_corr);
688 	init_crandom(&q->loss_cor, c->loss_corr);
689 	init_crandom(&q->dup_cor, c->dup_corr);
690 }
691 
692 static void get_reorder(struct Qdisc *sch, const struct nlattr *attr)
693 {
694 	struct netem_sched_data *q = qdisc_priv(sch);
695 	const struct tc_netem_reorder *r = nla_data(attr);
696 
697 	q->reorder = r->probability;
698 	init_crandom(&q->reorder_cor, r->correlation);
699 }
700 
701 static void get_corrupt(struct Qdisc *sch, const struct nlattr *attr)
702 {
703 	struct netem_sched_data *q = qdisc_priv(sch);
704 	const struct tc_netem_corrupt *r = nla_data(attr);
705 
706 	q->corrupt = r->probability;
707 	init_crandom(&q->corrupt_cor, r->correlation);
708 }
709 
710 static void get_rate(struct Qdisc *sch, const struct nlattr *attr)
711 {
712 	struct netem_sched_data *q = qdisc_priv(sch);
713 	const struct tc_netem_rate *r = nla_data(attr);
714 
715 	q->rate = r->rate;
716 	q->packet_overhead = r->packet_overhead;
717 	q->cell_size = r->cell_size;
718 	if (q->cell_size)
719 		q->cell_size_reciprocal = reciprocal_value(q->cell_size);
720 	q->cell_overhead = r->cell_overhead;
721 }
722 
723 static int get_loss_clg(struct Qdisc *sch, const struct nlattr *attr)
724 {
725 	struct netem_sched_data *q = qdisc_priv(sch);
726 	const struct nlattr *la;
727 	int rem;
728 
729 	nla_for_each_nested(la, attr, rem) {
730 		u16 type = nla_type(la);
731 
732 		switch(type) {
733 		case NETEM_LOSS_GI: {
734 			const struct tc_netem_gimodel *gi = nla_data(la);
735 
736 			if (nla_len(la) < sizeof(struct tc_netem_gimodel)) {
737 				pr_info("netem: incorrect gi model size\n");
738 				return -EINVAL;
739 			}
740 
741 			q->loss_model = CLG_4_STATES;
742 
743 			q->clg.state = 1;
744 			q->clg.a1 = gi->p13;
745 			q->clg.a2 = gi->p31;
746 			q->clg.a3 = gi->p32;
747 			q->clg.a4 = gi->p14;
748 			q->clg.a5 = gi->p23;
749 			break;
750 		}
751 
752 		case NETEM_LOSS_GE: {
753 			const struct tc_netem_gemodel *ge = nla_data(la);
754 
755 			if (nla_len(la) < sizeof(struct tc_netem_gemodel)) {
756 				pr_info("netem: incorrect ge model size\n");
757 				return -EINVAL;
758 			}
759 
760 			q->loss_model = CLG_GILB_ELL;
761 			q->clg.state = 1;
762 			q->clg.a1 = ge->p;
763 			q->clg.a2 = ge->r;
764 			q->clg.a3 = ge->h;
765 			q->clg.a4 = ge->k1;
766 			break;
767 		}
768 
769 		default:
770 			pr_info("netem: unknown loss type %u\n", type);
771 			return -EINVAL;
772 		}
773 	}
774 
775 	return 0;
776 }
777 
778 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
779 	[TCA_NETEM_CORR]	= { .len = sizeof(struct tc_netem_corr) },
780 	[TCA_NETEM_REORDER]	= { .len = sizeof(struct tc_netem_reorder) },
781 	[TCA_NETEM_CORRUPT]	= { .len = sizeof(struct tc_netem_corrupt) },
782 	[TCA_NETEM_RATE]	= { .len = sizeof(struct tc_netem_rate) },
783 	[TCA_NETEM_LOSS]	= { .type = NLA_NESTED },
784 	[TCA_NETEM_ECN]		= { .type = NLA_U32 },
785 };
786 
787 static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
788 		      const struct nla_policy *policy, int len)
789 {
790 	int nested_len = nla_len(nla) - NLA_ALIGN(len);
791 
792 	if (nested_len < 0) {
793 		pr_info("netem: invalid attributes len %d\n", nested_len);
794 		return -EINVAL;
795 	}
796 
797 	if (nested_len >= nla_attr_size(0))
798 		return nla_parse(tb, maxtype, nla_data(nla) + NLA_ALIGN(len),
799 				 nested_len, policy);
800 
801 	memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
802 	return 0;
803 }
804 
805 /* Parse netlink message to set options */
806 static int netem_change(struct Qdisc *sch, struct nlattr *opt)
807 {
808 	struct netem_sched_data *q = qdisc_priv(sch);
809 	struct nlattr *tb[TCA_NETEM_MAX + 1];
810 	struct tc_netem_qopt *qopt;
811 	int ret;
812 
813 	if (opt == NULL)
814 		return -EINVAL;
815 
816 	qopt = nla_data(opt);
817 	ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
818 	if (ret < 0)
819 		return ret;
820 
821 	sch->limit = qopt->limit;
822 
823 	q->latency = qopt->latency;
824 	q->jitter = qopt->jitter;
825 	q->limit = qopt->limit;
826 	q->gap = qopt->gap;
827 	q->counter = 0;
828 	q->loss = qopt->loss;
829 	q->duplicate = qopt->duplicate;
830 
831 	/* for compatibility with earlier versions.
832 	 * if gap is set, need to assume 100% probability
833 	 */
834 	if (q->gap)
835 		q->reorder = ~0;
836 
837 	if (tb[TCA_NETEM_CORR])
838 		get_correlation(sch, tb[TCA_NETEM_CORR]);
839 
840 	if (tb[TCA_NETEM_DELAY_DIST]) {
841 		ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST]);
842 		if (ret)
843 			return ret;
844 	}
845 
846 	if (tb[TCA_NETEM_REORDER])
847 		get_reorder(sch, tb[TCA_NETEM_REORDER]);
848 
849 	if (tb[TCA_NETEM_CORRUPT])
850 		get_corrupt(sch, tb[TCA_NETEM_CORRUPT]);
851 
852 	if (tb[TCA_NETEM_RATE])
853 		get_rate(sch, tb[TCA_NETEM_RATE]);
854 
855 	if (tb[TCA_NETEM_ECN])
856 		q->ecn = nla_get_u32(tb[TCA_NETEM_ECN]);
857 
858 	q->loss_model = CLG_RANDOM;
859 	if (tb[TCA_NETEM_LOSS])
860 		ret = get_loss_clg(sch, tb[TCA_NETEM_LOSS]);
861 
862 	return ret;
863 }
864 
865 static int netem_init(struct Qdisc *sch, struct nlattr *opt)
866 {
867 	struct netem_sched_data *q = qdisc_priv(sch);
868 	int ret;
869 
870 	if (!opt)
871 		return -EINVAL;
872 
873 	qdisc_watchdog_init(&q->watchdog, sch);
874 
875 	q->loss_model = CLG_RANDOM;
876 	ret = netem_change(sch, opt);
877 	if (ret)
878 		pr_info("netem: change failed\n");
879 	return ret;
880 }
881 
882 static void netem_destroy(struct Qdisc *sch)
883 {
884 	struct netem_sched_data *q = qdisc_priv(sch);
885 
886 	qdisc_watchdog_cancel(&q->watchdog);
887 	if (q->qdisc)
888 		qdisc_destroy(q->qdisc);
889 	dist_free(q->delay_dist);
890 }
891 
892 static int dump_loss_model(const struct netem_sched_data *q,
893 			   struct sk_buff *skb)
894 {
895 	struct nlattr *nest;
896 
897 	nest = nla_nest_start(skb, TCA_NETEM_LOSS);
898 	if (nest == NULL)
899 		goto nla_put_failure;
900 
901 	switch (q->loss_model) {
902 	case CLG_RANDOM:
903 		/* legacy loss model */
904 		nla_nest_cancel(skb, nest);
905 		return 0;	/* no data */
906 
907 	case CLG_4_STATES: {
908 		struct tc_netem_gimodel gi = {
909 			.p13 = q->clg.a1,
910 			.p31 = q->clg.a2,
911 			.p32 = q->clg.a3,
912 			.p14 = q->clg.a4,
913 			.p23 = q->clg.a5,
914 		};
915 
916 		if (nla_put(skb, NETEM_LOSS_GI, sizeof(gi), &gi))
917 			goto nla_put_failure;
918 		break;
919 	}
920 	case CLG_GILB_ELL: {
921 		struct tc_netem_gemodel ge = {
922 			.p = q->clg.a1,
923 			.r = q->clg.a2,
924 			.h = q->clg.a3,
925 			.k1 = q->clg.a4,
926 		};
927 
928 		if (nla_put(skb, NETEM_LOSS_GE, sizeof(ge), &ge))
929 			goto nla_put_failure;
930 		break;
931 	}
932 	}
933 
934 	nla_nest_end(skb, nest);
935 	return 0;
936 
937 nla_put_failure:
938 	nla_nest_cancel(skb, nest);
939 	return -1;
940 }
941 
942 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
943 {
944 	const struct netem_sched_data *q = qdisc_priv(sch);
945 	struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb);
946 	struct tc_netem_qopt qopt;
947 	struct tc_netem_corr cor;
948 	struct tc_netem_reorder reorder;
949 	struct tc_netem_corrupt corrupt;
950 	struct tc_netem_rate rate;
951 
952 	qopt.latency = q->latency;
953 	qopt.jitter = q->jitter;
954 	qopt.limit = q->limit;
955 	qopt.loss = q->loss;
956 	qopt.gap = q->gap;
957 	qopt.duplicate = q->duplicate;
958 	if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
959 		goto nla_put_failure;
960 
961 	cor.delay_corr = q->delay_cor.rho;
962 	cor.loss_corr = q->loss_cor.rho;
963 	cor.dup_corr = q->dup_cor.rho;
964 	if (nla_put(skb, TCA_NETEM_CORR, sizeof(cor), &cor))
965 		goto nla_put_failure;
966 
967 	reorder.probability = q->reorder;
968 	reorder.correlation = q->reorder_cor.rho;
969 	if (nla_put(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder))
970 		goto nla_put_failure;
971 
972 	corrupt.probability = q->corrupt;
973 	corrupt.correlation = q->corrupt_cor.rho;
974 	if (nla_put(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt))
975 		goto nla_put_failure;
976 
977 	rate.rate = q->rate;
978 	rate.packet_overhead = q->packet_overhead;
979 	rate.cell_size = q->cell_size;
980 	rate.cell_overhead = q->cell_overhead;
981 	if (nla_put(skb, TCA_NETEM_RATE, sizeof(rate), &rate))
982 		goto nla_put_failure;
983 
984 	if (q->ecn && nla_put_u32(skb, TCA_NETEM_ECN, q->ecn))
985 		goto nla_put_failure;
986 
987 	if (dump_loss_model(q, skb) != 0)
988 		goto nla_put_failure;
989 
990 	return nla_nest_end(skb, nla);
991 
992 nla_put_failure:
993 	nlmsg_trim(skb, nla);
994 	return -1;
995 }
996 
997 static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
998 			  struct sk_buff *skb, struct tcmsg *tcm)
999 {
1000 	struct netem_sched_data *q = qdisc_priv(sch);
1001 
1002 	if (cl != 1 || !q->qdisc) 	/* only one class */
1003 		return -ENOENT;
1004 
1005 	tcm->tcm_handle |= TC_H_MIN(1);
1006 	tcm->tcm_info = q->qdisc->handle;
1007 
1008 	return 0;
1009 }
1010 
1011 static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1012 		     struct Qdisc **old)
1013 {
1014 	struct netem_sched_data *q = qdisc_priv(sch);
1015 
1016 	sch_tree_lock(sch);
1017 	*old = q->qdisc;
1018 	q->qdisc = new;
1019 	if (*old) {
1020 		qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
1021 		qdisc_reset(*old);
1022 	}
1023 	sch_tree_unlock(sch);
1024 
1025 	return 0;
1026 }
1027 
1028 static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
1029 {
1030 	struct netem_sched_data *q = qdisc_priv(sch);
1031 	return q->qdisc;
1032 }
1033 
1034 static unsigned long netem_get(struct Qdisc *sch, u32 classid)
1035 {
1036 	return 1;
1037 }
1038 
1039 static void netem_put(struct Qdisc *sch, unsigned long arg)
1040 {
1041 }
1042 
1043 static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
1044 {
1045 	if (!walker->stop) {
1046 		if (walker->count >= walker->skip)
1047 			if (walker->fn(sch, 1, walker) < 0) {
1048 				walker->stop = 1;
1049 				return;
1050 			}
1051 		walker->count++;
1052 	}
1053 }
1054 
1055 static const struct Qdisc_class_ops netem_class_ops = {
1056 	.graft		=	netem_graft,
1057 	.leaf		=	netem_leaf,
1058 	.get		=	netem_get,
1059 	.put		=	netem_put,
1060 	.walk		=	netem_walk,
1061 	.dump		=	netem_dump_class,
1062 };
1063 
1064 static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
1065 	.id		=	"netem",
1066 	.cl_ops		=	&netem_class_ops,
1067 	.priv_size	=	sizeof(struct netem_sched_data),
1068 	.enqueue	=	netem_enqueue,
1069 	.dequeue	=	netem_dequeue,
1070 	.peek		=	qdisc_peek_dequeued,
1071 	.drop		=	netem_drop,
1072 	.init		=	netem_init,
1073 	.reset		=	netem_reset,
1074 	.destroy	=	netem_destroy,
1075 	.change		=	netem_change,
1076 	.dump		=	netem_dump,
1077 	.owner		=	THIS_MODULE,
1078 };
1079 
1080 
1081 static int __init netem_module_init(void)
1082 {
1083 	pr_info("netem: version " VERSION "\n");
1084 	return register_qdisc(&netem_qdisc_ops);
1085 }
1086 static void __exit netem_module_exit(void)
1087 {
1088 	unregister_qdisc(&netem_qdisc_ops);
1089 }
1090 module_init(netem_module_init)
1091 module_exit(netem_module_exit)
1092 MODULE_LICENSE("GPL");
1093