xref: /openbmc/linux/net/sched/sch_netem.c (revision 3e26a691)
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 	u64 rate;
92 	s32 packet_overhead;
93 	u32 cell_size;
94 	struct reciprocal_value 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 	enum {
114 		TX_IN_GAP_PERIOD = 1,
115 		TX_IN_BURST_PERIOD,
116 		LOST_IN_GAP_PERIOD,
117 		LOST_IN_BURST_PERIOD,
118 	} _4_state_model;
119 
120 	enum {
121 		GOOD_STATE = 1,
122 		BAD_STATE,
123 	} GE_state_model;
124 
125 	/* Correlated Loss Generation models */
126 	struct clgstate {
127 		/* state of the Markov chain */
128 		u8 state;
129 
130 		/* 4-states and Gilbert-Elliot models */
131 		u32 a1;	/* p13 for 4-states or p for GE */
132 		u32 a2;	/* p31 for 4-states or r for GE */
133 		u32 a3;	/* p32 for 4-states or h for GE */
134 		u32 a4;	/* p14 for 4-states or 1-k for GE */
135 		u32 a5; /* p23 used only in 4-states */
136 	} clg;
137 
138 };
139 
140 /* Time stamp put into socket buffer control block
141  * Only valid when skbs are in our internal t(ime)fifo queue.
142  *
143  * As skb->rbnode uses same storage than skb->next, skb->prev and skb->tstamp,
144  * and skb->next & skb->prev are scratch space for a qdisc,
145  * we save skb->tstamp value in skb->cb[] before destroying it.
146  */
147 struct netem_skb_cb {
148 	psched_time_t	time_to_send;
149 	ktime_t		tstamp_save;
150 };
151 
152 
153 static struct sk_buff *netem_rb_to_skb(struct rb_node *rb)
154 {
155 	return container_of(rb, struct sk_buff, rbnode);
156 }
157 
158 static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb)
159 {
160 	/* we assume we can use skb next/prev/tstamp as storage for rb_node */
161 	qdisc_cb_private_validate(skb, sizeof(struct netem_skb_cb));
162 	return (struct netem_skb_cb *)qdisc_skb_cb(skb)->data;
163 }
164 
165 /* init_crandom - initialize correlated random number generator
166  * Use entropy source for initial seed.
167  */
168 static void init_crandom(struct crndstate *state, unsigned long rho)
169 {
170 	state->rho = rho;
171 	state->last = prandom_u32();
172 }
173 
174 /* get_crandom - correlated random number generator
175  * Next number depends on last value.
176  * rho is scaled to avoid floating point.
177  */
178 static u32 get_crandom(struct crndstate *state)
179 {
180 	u64 value, rho;
181 	unsigned long answer;
182 
183 	if (state->rho == 0)	/* no correlation */
184 		return prandom_u32();
185 
186 	value = prandom_u32();
187 	rho = (u64)state->rho + 1;
188 	answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32;
189 	state->last = answer;
190 	return answer;
191 }
192 
193 /* loss_4state - 4-state model loss generator
194  * Generates losses according to the 4-state Markov chain adopted in
195  * the GI (General and Intuitive) loss model.
196  */
197 static bool loss_4state(struct netem_sched_data *q)
198 {
199 	struct clgstate *clg = &q->clg;
200 	u32 rnd = prandom_u32();
201 
202 	/*
203 	 * Makes a comparison between rnd and the transition
204 	 * probabilities outgoing from the current state, then decides the
205 	 * next state and if the next packet has to be transmitted or lost.
206 	 * The four states correspond to:
207 	 *   TX_IN_GAP_PERIOD => successfully transmitted packets within a gap period
208 	 *   LOST_IN_BURST_PERIOD => isolated losses within a gap period
209 	 *   LOST_IN_GAP_PERIOD => lost packets within a burst period
210 	 *   TX_IN_GAP_PERIOD => successfully transmitted packets within a burst period
211 	 */
212 	switch (clg->state) {
213 	case TX_IN_GAP_PERIOD:
214 		if (rnd < clg->a4) {
215 			clg->state = LOST_IN_BURST_PERIOD;
216 			return true;
217 		} else if (clg->a4 < rnd && rnd < clg->a1 + clg->a4) {
218 			clg->state = LOST_IN_GAP_PERIOD;
219 			return true;
220 		} else if (clg->a1 + clg->a4 < rnd) {
221 			clg->state = TX_IN_GAP_PERIOD;
222 		}
223 
224 		break;
225 	case TX_IN_BURST_PERIOD:
226 		if (rnd < clg->a5) {
227 			clg->state = LOST_IN_GAP_PERIOD;
228 			return true;
229 		} else {
230 			clg->state = TX_IN_BURST_PERIOD;
231 		}
232 
233 		break;
234 	case LOST_IN_GAP_PERIOD:
235 		if (rnd < clg->a3)
236 			clg->state = TX_IN_BURST_PERIOD;
237 		else if (clg->a3 < rnd && rnd < clg->a2 + clg->a3) {
238 			clg->state = TX_IN_GAP_PERIOD;
239 		} else if (clg->a2 + clg->a3 < rnd) {
240 			clg->state = LOST_IN_GAP_PERIOD;
241 			return true;
242 		}
243 		break;
244 	case LOST_IN_BURST_PERIOD:
245 		clg->state = TX_IN_GAP_PERIOD;
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 GOOD_STATE:
268 		if (prandom_u32() < clg->a1)
269 			clg->state = BAD_STATE;
270 		if (prandom_u32() < clg->a4)
271 			return true;
272 		break;
273 	case BAD_STATE:
274 		if (prandom_u32() < clg->a2)
275 			clg->state = GOOD_STATE;
276 		if (prandom_u32() > clg->a3)
277 			return true;
278 	}
279 
280 	return false;
281 }
282 
283 static bool loss_event(struct netem_sched_data *q)
284 {
285 	switch (q->loss_model) {
286 	case CLG_RANDOM:
287 		/* Random packet drop 0 => none, ~0 => all */
288 		return q->loss && q->loss >= get_crandom(&q->loss_cor);
289 
290 	case CLG_4_STATES:
291 		/* 4state loss model algorithm (used also for GI model)
292 		* Extracts a value from the markov 4 state loss generator,
293 		* if it is 1 drops a packet and if needed writes the event in
294 		* the kernel logs
295 		*/
296 		return loss_4state(q);
297 
298 	case CLG_GILB_ELL:
299 		/* Gilbert-Elliot loss model algorithm
300 		* Extracts a value from the Gilbert-Elliot loss generator,
301 		* if it is 1 drops a packet and if needed writes the event in
302 		* the kernel logs
303 		*/
304 		return loss_gilb_ell(q);
305 	}
306 
307 	return false;	/* not reached */
308 }
309 
310 
311 /* tabledist - return a pseudo-randomly distributed value with mean mu and
312  * std deviation sigma.  Uses table lookup to approximate the desired
313  * distribution, and a uniformly-distributed pseudo-random source.
314  */
315 static psched_tdiff_t tabledist(psched_tdiff_t mu, psched_tdiff_t sigma,
316 				struct crndstate *state,
317 				const struct disttable *dist)
318 {
319 	psched_tdiff_t x;
320 	long t;
321 	u32 rnd;
322 
323 	if (sigma == 0)
324 		return mu;
325 
326 	rnd = get_crandom(state);
327 
328 	/* default uniform distribution */
329 	if (dist == NULL)
330 		return (rnd % (2*sigma)) - sigma + mu;
331 
332 	t = dist->table[rnd % dist->size];
333 	x = (sigma % NETEM_DIST_SCALE) * t;
334 	if (x >= 0)
335 		x += NETEM_DIST_SCALE/2;
336 	else
337 		x -= NETEM_DIST_SCALE/2;
338 
339 	return  x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu;
340 }
341 
342 static psched_time_t packet_len_2_sched_time(unsigned int len, struct netem_sched_data *q)
343 {
344 	u64 ticks;
345 
346 	len += q->packet_overhead;
347 
348 	if (q->cell_size) {
349 		u32 cells = reciprocal_divide(len, q->cell_size_reciprocal);
350 
351 		if (len > cells * q->cell_size)	/* extra cell needed for remainder */
352 			cells++;
353 		len = cells * (q->cell_size + q->cell_overhead);
354 	}
355 
356 	ticks = (u64)len * NSEC_PER_SEC;
357 
358 	do_div(ticks, q->rate);
359 	return PSCHED_NS2TICKS(ticks);
360 }
361 
362 static void tfifo_reset(struct Qdisc *sch)
363 {
364 	struct netem_sched_data *q = qdisc_priv(sch);
365 	struct rb_node *p;
366 
367 	while ((p = rb_first(&q->t_root))) {
368 		struct sk_buff *skb = netem_rb_to_skb(p);
369 
370 		rb_erase(p, &q->t_root);
371 		skb->next = NULL;
372 		skb->prev = NULL;
373 		kfree_skb(skb);
374 	}
375 }
376 
377 static void tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
378 {
379 	struct netem_sched_data *q = qdisc_priv(sch);
380 	psched_time_t tnext = netem_skb_cb(nskb)->time_to_send;
381 	struct rb_node **p = &q->t_root.rb_node, *parent = NULL;
382 
383 	while (*p) {
384 		struct sk_buff *skb;
385 
386 		parent = *p;
387 		skb = netem_rb_to_skb(parent);
388 		if (tnext >= netem_skb_cb(skb)->time_to_send)
389 			p = &parent->rb_right;
390 		else
391 			p = &parent->rb_left;
392 	}
393 	rb_link_node(&nskb->rbnode, parent, p);
394 	rb_insert_color(&nskb->rbnode, &q->t_root);
395 	sch->q.qlen++;
396 }
397 
398 /*
399  * Insert one skb into qdisc.
400  * Note: parent depends on return value to account for queue length.
401  * 	NET_XMIT_DROP: queue length didn't change.
402  *      NET_XMIT_SUCCESS: one skb was queued.
403  */
404 static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch)
405 {
406 	struct netem_sched_data *q = qdisc_priv(sch);
407 	/* We don't fill cb now as skb_unshare() may invalidate it */
408 	struct netem_skb_cb *cb;
409 	struct sk_buff *skb2;
410 	int count = 1;
411 
412 	/* Random duplication */
413 	if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor))
414 		++count;
415 
416 	/* Drop packet? */
417 	if (loss_event(q)) {
418 		if (q->ecn && INET_ECN_set_ce(skb))
419 			qdisc_qstats_drop(sch); /* mark packet */
420 		else
421 			--count;
422 	}
423 	if (count == 0) {
424 		qdisc_qstats_drop(sch);
425 		kfree_skb(skb);
426 		return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
427 	}
428 
429 	/* If a delay is expected, orphan the skb. (orphaning usually takes
430 	 * place at TX completion time, so _before_ the link transit delay)
431 	 */
432 	if (q->latency || q->jitter)
433 		skb_orphan_partial(skb);
434 
435 	/*
436 	 * If we need to duplicate packet, then re-insert at top of the
437 	 * qdisc tree, since parent queuer expects that only one
438 	 * skb will be queued.
439 	 */
440 	if (count > 1 && (skb2 = skb_clone(skb, GFP_ATOMIC)) != NULL) {
441 		struct Qdisc *rootq = qdisc_root(sch);
442 		u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
443 
444 		q->duplicate = 0;
445 		rootq->enqueue(skb2, rootq);
446 		q->duplicate = dupsave;
447 	}
448 
449 	/*
450 	 * Randomized packet corruption.
451 	 * Make copy if needed since we are modifying
452 	 * If packet is going to be hardware checksummed, then
453 	 * do it now in software before we mangle it.
454 	 */
455 	if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor)) {
456 		if (!(skb = skb_unshare(skb, GFP_ATOMIC)) ||
457 		    (skb->ip_summed == CHECKSUM_PARTIAL &&
458 		     skb_checksum_help(skb)))
459 			return qdisc_drop(skb, sch);
460 
461 		skb->data[prandom_u32() % skb_headlen(skb)] ^=
462 			1<<(prandom_u32() % 8);
463 	}
464 
465 	if (unlikely(skb_queue_len(&sch->q) >= sch->limit))
466 		return qdisc_reshape_fail(skb, sch);
467 
468 	qdisc_qstats_backlog_inc(sch, skb);
469 
470 	cb = netem_skb_cb(skb);
471 	if (q->gap == 0 ||		/* not doing reordering */
472 	    q->counter < q->gap - 1 ||	/* inside last reordering gap */
473 	    q->reorder < get_crandom(&q->reorder_cor)) {
474 		psched_time_t now;
475 		psched_tdiff_t delay;
476 
477 		delay = tabledist(q->latency, q->jitter,
478 				  &q->delay_cor, q->delay_dist);
479 
480 		now = psched_get_time();
481 
482 		if (q->rate) {
483 			struct sk_buff *last;
484 
485 			if (!skb_queue_empty(&sch->q))
486 				last = skb_peek_tail(&sch->q);
487 			else
488 				last = netem_rb_to_skb(rb_last(&q->t_root));
489 			if (last) {
490 				/*
491 				 * Last packet in queue is reference point (now),
492 				 * calculate this time bonus and subtract
493 				 * from delay.
494 				 */
495 				delay -= netem_skb_cb(last)->time_to_send - now;
496 				delay = max_t(psched_tdiff_t, 0, delay);
497 				now = netem_skb_cb(last)->time_to_send;
498 			}
499 
500 			delay += packet_len_2_sched_time(qdisc_pkt_len(skb), q);
501 		}
502 
503 		cb->time_to_send = now + delay;
504 		cb->tstamp_save = skb->tstamp;
505 		++q->counter;
506 		tfifo_enqueue(skb, sch);
507 	} else {
508 		/*
509 		 * Do re-ordering by putting one out of N packets at the front
510 		 * of the queue.
511 		 */
512 		cb->time_to_send = psched_get_time();
513 		q->counter = 0;
514 
515 		__skb_queue_head(&sch->q, skb);
516 		sch->qstats.requeues++;
517 	}
518 
519 	return NET_XMIT_SUCCESS;
520 }
521 
522 static unsigned int netem_drop(struct Qdisc *sch)
523 {
524 	struct netem_sched_data *q = qdisc_priv(sch);
525 	unsigned int len;
526 
527 	len = qdisc_queue_drop(sch);
528 
529 	if (!len) {
530 		struct rb_node *p = rb_first(&q->t_root);
531 
532 		if (p) {
533 			struct sk_buff *skb = netem_rb_to_skb(p);
534 
535 			rb_erase(p, &q->t_root);
536 			sch->q.qlen--;
537 			skb->next = NULL;
538 			skb->prev = NULL;
539 			qdisc_qstats_backlog_dec(sch, skb);
540 			kfree_skb(skb);
541 		}
542 	}
543 	if (!len && q->qdisc && q->qdisc->ops->drop)
544 	    len = q->qdisc->ops->drop(q->qdisc);
545 	if (len)
546 		qdisc_qstats_drop(sch);
547 
548 	return len;
549 }
550 
551 static struct sk_buff *netem_dequeue(struct Qdisc *sch)
552 {
553 	struct netem_sched_data *q = qdisc_priv(sch);
554 	struct sk_buff *skb;
555 	struct rb_node *p;
556 
557 	if (qdisc_is_throttled(sch))
558 		return NULL;
559 
560 tfifo_dequeue:
561 	skb = __skb_dequeue(&sch->q);
562 	if (skb) {
563 		qdisc_qstats_backlog_dec(sch, skb);
564 deliver:
565 		qdisc_unthrottled(sch);
566 		qdisc_bstats_update(sch, skb);
567 		return skb;
568 	}
569 	p = rb_first(&q->t_root);
570 	if (p) {
571 		psched_time_t time_to_send;
572 
573 		skb = netem_rb_to_skb(p);
574 
575 		/* if more time remaining? */
576 		time_to_send = netem_skb_cb(skb)->time_to_send;
577 		if (time_to_send <= psched_get_time()) {
578 			rb_erase(p, &q->t_root);
579 
580 			sch->q.qlen--;
581 			qdisc_qstats_backlog_dec(sch, skb);
582 			skb->next = NULL;
583 			skb->prev = NULL;
584 			skb->tstamp = netem_skb_cb(skb)->tstamp_save;
585 
586 #ifdef CONFIG_NET_CLS_ACT
587 			/*
588 			 * If it's at ingress let's pretend the delay is
589 			 * from the network (tstamp will be updated).
590 			 */
591 			if (G_TC_FROM(skb->tc_verd) & AT_INGRESS)
592 				skb->tstamp.tv64 = 0;
593 #endif
594 
595 			if (q->qdisc) {
596 				int err = qdisc_enqueue(skb, q->qdisc);
597 
598 				if (unlikely(err != NET_XMIT_SUCCESS)) {
599 					if (net_xmit_drop_count(err)) {
600 						qdisc_qstats_drop(sch);
601 						qdisc_tree_reduce_backlog(sch, 1,
602 									  qdisc_pkt_len(skb));
603 					}
604 				}
605 				goto tfifo_dequeue;
606 			}
607 			goto deliver;
608 		}
609 
610 		if (q->qdisc) {
611 			skb = q->qdisc->ops->dequeue(q->qdisc);
612 			if (skb)
613 				goto deliver;
614 		}
615 		qdisc_watchdog_schedule(&q->watchdog, time_to_send);
616 	}
617 
618 	if (q->qdisc) {
619 		skb = q->qdisc->ops->dequeue(q->qdisc);
620 		if (skb)
621 			goto deliver;
622 	}
623 	return NULL;
624 }
625 
626 static void netem_reset(struct Qdisc *sch)
627 {
628 	struct netem_sched_data *q = qdisc_priv(sch);
629 
630 	qdisc_reset_queue(sch);
631 	tfifo_reset(sch);
632 	if (q->qdisc)
633 		qdisc_reset(q->qdisc);
634 	qdisc_watchdog_cancel(&q->watchdog);
635 }
636 
637 static void dist_free(struct disttable *d)
638 {
639 	kvfree(d);
640 }
641 
642 /*
643  * Distribution data is a variable size payload containing
644  * signed 16 bit values.
645  */
646 static int get_dist_table(struct Qdisc *sch, const struct nlattr *attr)
647 {
648 	struct netem_sched_data *q = qdisc_priv(sch);
649 	size_t n = nla_len(attr)/sizeof(__s16);
650 	const __s16 *data = nla_data(attr);
651 	spinlock_t *root_lock;
652 	struct disttable *d;
653 	int i;
654 	size_t s;
655 
656 	if (n > NETEM_DIST_MAX)
657 		return -EINVAL;
658 
659 	s = sizeof(struct disttable) + n * sizeof(s16);
660 	d = kmalloc(s, GFP_KERNEL | __GFP_NOWARN);
661 	if (!d)
662 		d = vmalloc(s);
663 	if (!d)
664 		return -ENOMEM;
665 
666 	d->size = n;
667 	for (i = 0; i < n; i++)
668 		d->table[i] = data[i];
669 
670 	root_lock = qdisc_root_sleeping_lock(sch);
671 
672 	spin_lock_bh(root_lock);
673 	swap(q->delay_dist, d);
674 	spin_unlock_bh(root_lock);
675 
676 	dist_free(d);
677 	return 0;
678 }
679 
680 static void get_correlation(struct netem_sched_data *q, const struct nlattr *attr)
681 {
682 	const struct tc_netem_corr *c = nla_data(attr);
683 
684 	init_crandom(&q->delay_cor, c->delay_corr);
685 	init_crandom(&q->loss_cor, c->loss_corr);
686 	init_crandom(&q->dup_cor, c->dup_corr);
687 }
688 
689 static void get_reorder(struct netem_sched_data *q, const struct nlattr *attr)
690 {
691 	const struct tc_netem_reorder *r = nla_data(attr);
692 
693 	q->reorder = r->probability;
694 	init_crandom(&q->reorder_cor, r->correlation);
695 }
696 
697 static void get_corrupt(struct netem_sched_data *q, const struct nlattr *attr)
698 {
699 	const struct tc_netem_corrupt *r = nla_data(attr);
700 
701 	q->corrupt = r->probability;
702 	init_crandom(&q->corrupt_cor, r->correlation);
703 }
704 
705 static void get_rate(struct netem_sched_data *q, const struct nlattr *attr)
706 {
707 	const struct tc_netem_rate *r = nla_data(attr);
708 
709 	q->rate = r->rate;
710 	q->packet_overhead = r->packet_overhead;
711 	q->cell_size = r->cell_size;
712 	q->cell_overhead = r->cell_overhead;
713 	if (q->cell_size)
714 		q->cell_size_reciprocal = reciprocal_value(q->cell_size);
715 	else
716 		q->cell_size_reciprocal = (struct reciprocal_value) { 0 };
717 }
718 
719 static int get_loss_clg(struct netem_sched_data *q, const struct nlattr *attr)
720 {
721 	const struct nlattr *la;
722 	int rem;
723 
724 	nla_for_each_nested(la, attr, rem) {
725 		u16 type = nla_type(la);
726 
727 		switch (type) {
728 		case NETEM_LOSS_GI: {
729 			const struct tc_netem_gimodel *gi = nla_data(la);
730 
731 			if (nla_len(la) < sizeof(struct tc_netem_gimodel)) {
732 				pr_info("netem: incorrect gi model size\n");
733 				return -EINVAL;
734 			}
735 
736 			q->loss_model = CLG_4_STATES;
737 
738 			q->clg.state = TX_IN_GAP_PERIOD;
739 			q->clg.a1 = gi->p13;
740 			q->clg.a2 = gi->p31;
741 			q->clg.a3 = gi->p32;
742 			q->clg.a4 = gi->p14;
743 			q->clg.a5 = gi->p23;
744 			break;
745 		}
746 
747 		case NETEM_LOSS_GE: {
748 			const struct tc_netem_gemodel *ge = nla_data(la);
749 
750 			if (nla_len(la) < sizeof(struct tc_netem_gemodel)) {
751 				pr_info("netem: incorrect ge model size\n");
752 				return -EINVAL;
753 			}
754 
755 			q->loss_model = CLG_GILB_ELL;
756 			q->clg.state = GOOD_STATE;
757 			q->clg.a1 = ge->p;
758 			q->clg.a2 = ge->r;
759 			q->clg.a3 = ge->h;
760 			q->clg.a4 = ge->k1;
761 			break;
762 		}
763 
764 		default:
765 			pr_info("netem: unknown loss type %u\n", type);
766 			return -EINVAL;
767 		}
768 	}
769 
770 	return 0;
771 }
772 
773 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
774 	[TCA_NETEM_CORR]	= { .len = sizeof(struct tc_netem_corr) },
775 	[TCA_NETEM_REORDER]	= { .len = sizeof(struct tc_netem_reorder) },
776 	[TCA_NETEM_CORRUPT]	= { .len = sizeof(struct tc_netem_corrupt) },
777 	[TCA_NETEM_RATE]	= { .len = sizeof(struct tc_netem_rate) },
778 	[TCA_NETEM_LOSS]	= { .type = NLA_NESTED },
779 	[TCA_NETEM_ECN]		= { .type = NLA_U32 },
780 	[TCA_NETEM_RATE64]	= { .type = NLA_U64 },
781 };
782 
783 static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
784 		      const struct nla_policy *policy, int len)
785 {
786 	int nested_len = nla_len(nla) - NLA_ALIGN(len);
787 
788 	if (nested_len < 0) {
789 		pr_info("netem: invalid attributes len %d\n", nested_len);
790 		return -EINVAL;
791 	}
792 
793 	if (nested_len >= nla_attr_size(0))
794 		return nla_parse(tb, maxtype, nla_data(nla) + NLA_ALIGN(len),
795 				 nested_len, policy);
796 
797 	memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
798 	return 0;
799 }
800 
801 /* Parse netlink message to set options */
802 static int netem_change(struct Qdisc *sch, struct nlattr *opt)
803 {
804 	struct netem_sched_data *q = qdisc_priv(sch);
805 	struct nlattr *tb[TCA_NETEM_MAX + 1];
806 	struct tc_netem_qopt *qopt;
807 	struct clgstate old_clg;
808 	int old_loss_model = CLG_RANDOM;
809 	int ret;
810 
811 	if (opt == NULL)
812 		return -EINVAL;
813 
814 	qopt = nla_data(opt);
815 	ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
816 	if (ret < 0)
817 		return ret;
818 
819 	/* backup q->clg and q->loss_model */
820 	old_clg = q->clg;
821 	old_loss_model = q->loss_model;
822 
823 	if (tb[TCA_NETEM_LOSS]) {
824 		ret = get_loss_clg(q, tb[TCA_NETEM_LOSS]);
825 		if (ret) {
826 			q->loss_model = old_loss_model;
827 			return ret;
828 		}
829 	} else {
830 		q->loss_model = CLG_RANDOM;
831 	}
832 
833 	if (tb[TCA_NETEM_DELAY_DIST]) {
834 		ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST]);
835 		if (ret) {
836 			/* recover clg and loss_model, in case of
837 			 * q->clg and q->loss_model were modified
838 			 * in get_loss_clg()
839 			 */
840 			q->clg = old_clg;
841 			q->loss_model = old_loss_model;
842 			return ret;
843 		}
844 	}
845 
846 	sch->limit = qopt->limit;
847 
848 	q->latency = qopt->latency;
849 	q->jitter = qopt->jitter;
850 	q->limit = qopt->limit;
851 	q->gap = qopt->gap;
852 	q->counter = 0;
853 	q->loss = qopt->loss;
854 	q->duplicate = qopt->duplicate;
855 
856 	/* for compatibility with earlier versions.
857 	 * if gap is set, need to assume 100% probability
858 	 */
859 	if (q->gap)
860 		q->reorder = ~0;
861 
862 	if (tb[TCA_NETEM_CORR])
863 		get_correlation(q, tb[TCA_NETEM_CORR]);
864 
865 	if (tb[TCA_NETEM_REORDER])
866 		get_reorder(q, tb[TCA_NETEM_REORDER]);
867 
868 	if (tb[TCA_NETEM_CORRUPT])
869 		get_corrupt(q, tb[TCA_NETEM_CORRUPT]);
870 
871 	if (tb[TCA_NETEM_RATE])
872 		get_rate(q, tb[TCA_NETEM_RATE]);
873 
874 	if (tb[TCA_NETEM_RATE64])
875 		q->rate = max_t(u64, q->rate,
876 				nla_get_u64(tb[TCA_NETEM_RATE64]));
877 
878 	if (tb[TCA_NETEM_ECN])
879 		q->ecn = nla_get_u32(tb[TCA_NETEM_ECN]);
880 
881 	return ret;
882 }
883 
884 static int netem_init(struct Qdisc *sch, struct nlattr *opt)
885 {
886 	struct netem_sched_data *q = qdisc_priv(sch);
887 	int ret;
888 
889 	if (!opt)
890 		return -EINVAL;
891 
892 	qdisc_watchdog_init(&q->watchdog, sch);
893 
894 	q->loss_model = CLG_RANDOM;
895 	ret = netem_change(sch, opt);
896 	if (ret)
897 		pr_info("netem: change failed\n");
898 	return ret;
899 }
900 
901 static void netem_destroy(struct Qdisc *sch)
902 {
903 	struct netem_sched_data *q = qdisc_priv(sch);
904 
905 	qdisc_watchdog_cancel(&q->watchdog);
906 	if (q->qdisc)
907 		qdisc_destroy(q->qdisc);
908 	dist_free(q->delay_dist);
909 }
910 
911 static int dump_loss_model(const struct netem_sched_data *q,
912 			   struct sk_buff *skb)
913 {
914 	struct nlattr *nest;
915 
916 	nest = nla_nest_start(skb, TCA_NETEM_LOSS);
917 	if (nest == NULL)
918 		goto nla_put_failure;
919 
920 	switch (q->loss_model) {
921 	case CLG_RANDOM:
922 		/* legacy loss model */
923 		nla_nest_cancel(skb, nest);
924 		return 0;	/* no data */
925 
926 	case CLG_4_STATES: {
927 		struct tc_netem_gimodel gi = {
928 			.p13 = q->clg.a1,
929 			.p31 = q->clg.a2,
930 			.p32 = q->clg.a3,
931 			.p14 = q->clg.a4,
932 			.p23 = q->clg.a5,
933 		};
934 
935 		if (nla_put(skb, NETEM_LOSS_GI, sizeof(gi), &gi))
936 			goto nla_put_failure;
937 		break;
938 	}
939 	case CLG_GILB_ELL: {
940 		struct tc_netem_gemodel ge = {
941 			.p = q->clg.a1,
942 			.r = q->clg.a2,
943 			.h = q->clg.a3,
944 			.k1 = q->clg.a4,
945 		};
946 
947 		if (nla_put(skb, NETEM_LOSS_GE, sizeof(ge), &ge))
948 			goto nla_put_failure;
949 		break;
950 	}
951 	}
952 
953 	nla_nest_end(skb, nest);
954 	return 0;
955 
956 nla_put_failure:
957 	nla_nest_cancel(skb, nest);
958 	return -1;
959 }
960 
961 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
962 {
963 	const struct netem_sched_data *q = qdisc_priv(sch);
964 	struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb);
965 	struct tc_netem_qopt qopt;
966 	struct tc_netem_corr cor;
967 	struct tc_netem_reorder reorder;
968 	struct tc_netem_corrupt corrupt;
969 	struct tc_netem_rate rate;
970 
971 	qopt.latency = q->latency;
972 	qopt.jitter = q->jitter;
973 	qopt.limit = q->limit;
974 	qopt.loss = q->loss;
975 	qopt.gap = q->gap;
976 	qopt.duplicate = q->duplicate;
977 	if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
978 		goto nla_put_failure;
979 
980 	cor.delay_corr = q->delay_cor.rho;
981 	cor.loss_corr = q->loss_cor.rho;
982 	cor.dup_corr = q->dup_cor.rho;
983 	if (nla_put(skb, TCA_NETEM_CORR, sizeof(cor), &cor))
984 		goto nla_put_failure;
985 
986 	reorder.probability = q->reorder;
987 	reorder.correlation = q->reorder_cor.rho;
988 	if (nla_put(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder))
989 		goto nla_put_failure;
990 
991 	corrupt.probability = q->corrupt;
992 	corrupt.correlation = q->corrupt_cor.rho;
993 	if (nla_put(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt))
994 		goto nla_put_failure;
995 
996 	if (q->rate >= (1ULL << 32)) {
997 		if (nla_put_u64(skb, TCA_NETEM_RATE64, q->rate))
998 			goto nla_put_failure;
999 		rate.rate = ~0U;
1000 	} else {
1001 		rate.rate = q->rate;
1002 	}
1003 	rate.packet_overhead = q->packet_overhead;
1004 	rate.cell_size = q->cell_size;
1005 	rate.cell_overhead = q->cell_overhead;
1006 	if (nla_put(skb, TCA_NETEM_RATE, sizeof(rate), &rate))
1007 		goto nla_put_failure;
1008 
1009 	if (q->ecn && nla_put_u32(skb, TCA_NETEM_ECN, q->ecn))
1010 		goto nla_put_failure;
1011 
1012 	if (dump_loss_model(q, skb) != 0)
1013 		goto nla_put_failure;
1014 
1015 	return nla_nest_end(skb, nla);
1016 
1017 nla_put_failure:
1018 	nlmsg_trim(skb, nla);
1019 	return -1;
1020 }
1021 
1022 static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
1023 			  struct sk_buff *skb, struct tcmsg *tcm)
1024 {
1025 	struct netem_sched_data *q = qdisc_priv(sch);
1026 
1027 	if (cl != 1 || !q->qdisc) 	/* only one class */
1028 		return -ENOENT;
1029 
1030 	tcm->tcm_handle |= TC_H_MIN(1);
1031 	tcm->tcm_info = q->qdisc->handle;
1032 
1033 	return 0;
1034 }
1035 
1036 static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1037 		     struct Qdisc **old)
1038 {
1039 	struct netem_sched_data *q = qdisc_priv(sch);
1040 
1041 	*old = qdisc_replace(sch, new, &q->qdisc);
1042 	return 0;
1043 }
1044 
1045 static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
1046 {
1047 	struct netem_sched_data *q = qdisc_priv(sch);
1048 	return q->qdisc;
1049 }
1050 
1051 static unsigned long netem_get(struct Qdisc *sch, u32 classid)
1052 {
1053 	return 1;
1054 }
1055 
1056 static void netem_put(struct Qdisc *sch, unsigned long arg)
1057 {
1058 }
1059 
1060 static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
1061 {
1062 	if (!walker->stop) {
1063 		if (walker->count >= walker->skip)
1064 			if (walker->fn(sch, 1, walker) < 0) {
1065 				walker->stop = 1;
1066 				return;
1067 			}
1068 		walker->count++;
1069 	}
1070 }
1071 
1072 static const struct Qdisc_class_ops netem_class_ops = {
1073 	.graft		=	netem_graft,
1074 	.leaf		=	netem_leaf,
1075 	.get		=	netem_get,
1076 	.put		=	netem_put,
1077 	.walk		=	netem_walk,
1078 	.dump		=	netem_dump_class,
1079 };
1080 
1081 static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
1082 	.id		=	"netem",
1083 	.cl_ops		=	&netem_class_ops,
1084 	.priv_size	=	sizeof(struct netem_sched_data),
1085 	.enqueue	=	netem_enqueue,
1086 	.dequeue	=	netem_dequeue,
1087 	.peek		=	qdisc_peek_dequeued,
1088 	.drop		=	netem_drop,
1089 	.init		=	netem_init,
1090 	.reset		=	netem_reset,
1091 	.destroy	=	netem_destroy,
1092 	.change		=	netem_change,
1093 	.dump		=	netem_dump,
1094 	.owner		=	THIS_MODULE,
1095 };
1096 
1097 
1098 static int __init netem_module_init(void)
1099 {
1100 	pr_info("netem: version " VERSION "\n");
1101 	return register_qdisc(&netem_qdisc_ops);
1102 }
1103 static void __exit netem_module_exit(void)
1104 {
1105 	unregister_qdisc(&netem_qdisc_ops);
1106 }
1107 module_init(netem_module_init)
1108 module_exit(netem_module_exit)
1109 MODULE_LICENSE("GPL");
1110