xref: /openbmc/linux/net/sched/sch_netem.c (revision e0f6d1a5)
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 	s64 latency;
81 	s64 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 	struct tc_netem_slot slot_config;
139 	struct slotstate {
140 		u64 slot_next;
141 		s32 packets_left;
142 		s32 bytes_left;
143 	} slot;
144 
145 };
146 
147 /* Time stamp put into socket buffer control block
148  * Only valid when skbs are in our internal t(ime)fifo queue.
149  *
150  * As skb->rbnode uses same storage than skb->next, skb->prev and skb->tstamp,
151  * and skb->next & skb->prev are scratch space for a qdisc,
152  * we save skb->tstamp value in skb->cb[] before destroying it.
153  */
154 struct netem_skb_cb {
155 	u64	        time_to_send;
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 s64 tabledist(s64 mu, s32 sigma,
316 		     struct crndstate *state,
317 		     const struct disttable *dist)
318 {
319 	s64 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)) + mu) - sigma;
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 u64 packet_time_ns(u64 len, const struct netem_sched_data *q)
343 {
344 	len += q->packet_overhead;
345 
346 	if (q->cell_size) {
347 		u32 cells = reciprocal_divide(len, q->cell_size_reciprocal);
348 
349 		if (len > cells * q->cell_size)	/* extra cell needed for remainder */
350 			cells++;
351 		len = cells * (q->cell_size + q->cell_overhead);
352 	}
353 
354 	return div64_u64(len * NSEC_PER_SEC, q->rate);
355 }
356 
357 static void tfifo_reset(struct Qdisc *sch)
358 {
359 	struct netem_sched_data *q = qdisc_priv(sch);
360 	struct rb_node *p = rb_first(&q->t_root);
361 
362 	while (p) {
363 		struct sk_buff *skb = rb_to_skb(p);
364 
365 		p = rb_next(p);
366 		rb_erase(&skb->rbnode, &q->t_root);
367 		rtnl_kfree_skbs(skb, skb);
368 	}
369 }
370 
371 static void tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
372 {
373 	struct netem_sched_data *q = qdisc_priv(sch);
374 	u64 tnext = netem_skb_cb(nskb)->time_to_send;
375 	struct rb_node **p = &q->t_root.rb_node, *parent = NULL;
376 
377 	while (*p) {
378 		struct sk_buff *skb;
379 
380 		parent = *p;
381 		skb = rb_to_skb(parent);
382 		if (tnext >= netem_skb_cb(skb)->time_to_send)
383 			p = &parent->rb_right;
384 		else
385 			p = &parent->rb_left;
386 	}
387 	rb_link_node(&nskb->rbnode, parent, p);
388 	rb_insert_color(&nskb->rbnode, &q->t_root);
389 	sch->q.qlen++;
390 }
391 
392 /* netem can't properly corrupt a megapacket (like we get from GSO), so instead
393  * when we statistically choose to corrupt one, we instead segment it, returning
394  * the first packet to be corrupted, and re-enqueue the remaining frames
395  */
396 static struct sk_buff *netem_segment(struct sk_buff *skb, struct Qdisc *sch,
397 				     struct sk_buff **to_free)
398 {
399 	struct sk_buff *segs;
400 	netdev_features_t features = netif_skb_features(skb);
401 
402 	segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
403 
404 	if (IS_ERR_OR_NULL(segs)) {
405 		qdisc_drop(skb, sch, to_free);
406 		return NULL;
407 	}
408 	consume_skb(skb);
409 	return segs;
410 }
411 
412 static void netem_enqueue_skb_head(struct qdisc_skb_head *qh, struct sk_buff *skb)
413 {
414 	skb->next = qh->head;
415 
416 	if (!qh->head)
417 		qh->tail = skb;
418 	qh->head = skb;
419 	qh->qlen++;
420 }
421 
422 /*
423  * Insert one skb into qdisc.
424  * Note: parent depends on return value to account for queue length.
425  * 	NET_XMIT_DROP: queue length didn't change.
426  *      NET_XMIT_SUCCESS: one skb was queued.
427  */
428 static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch,
429 			 struct sk_buff **to_free)
430 {
431 	struct netem_sched_data *q = qdisc_priv(sch);
432 	/* We don't fill cb now as skb_unshare() may invalidate it */
433 	struct netem_skb_cb *cb;
434 	struct sk_buff *skb2;
435 	struct sk_buff *segs = NULL;
436 	unsigned int len = 0, last_len, prev_len = qdisc_pkt_len(skb);
437 	int nb = 0;
438 	int count = 1;
439 	int rc = NET_XMIT_SUCCESS;
440 
441 	/* Random duplication */
442 	if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor))
443 		++count;
444 
445 	/* Drop packet? */
446 	if (loss_event(q)) {
447 		if (q->ecn && INET_ECN_set_ce(skb))
448 			qdisc_qstats_drop(sch); /* mark packet */
449 		else
450 			--count;
451 	}
452 	if (count == 0) {
453 		qdisc_qstats_drop(sch);
454 		__qdisc_drop(skb, to_free);
455 		return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
456 	}
457 
458 	/* If a delay is expected, orphan the skb. (orphaning usually takes
459 	 * place at TX completion time, so _before_ the link transit delay)
460 	 */
461 	if (q->latency || q->jitter || q->rate)
462 		skb_orphan_partial(skb);
463 
464 	/*
465 	 * If we need to duplicate packet, then re-insert at top of the
466 	 * qdisc tree, since parent queuer expects that only one
467 	 * skb will be queued.
468 	 */
469 	if (count > 1 && (skb2 = skb_clone(skb, GFP_ATOMIC)) != NULL) {
470 		struct Qdisc *rootq = qdisc_root(sch);
471 		u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
472 
473 		q->duplicate = 0;
474 		rootq->enqueue(skb2, rootq, to_free);
475 		q->duplicate = dupsave;
476 	}
477 
478 	/*
479 	 * Randomized packet corruption.
480 	 * Make copy if needed since we are modifying
481 	 * If packet is going to be hardware checksummed, then
482 	 * do it now in software before we mangle it.
483 	 */
484 	if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor)) {
485 		if (skb_is_gso(skb)) {
486 			segs = netem_segment(skb, sch, to_free);
487 			if (!segs)
488 				return NET_XMIT_DROP;
489 		} else {
490 			segs = skb;
491 		}
492 
493 		skb = segs;
494 		segs = segs->next;
495 
496 		skb = skb_unshare(skb, GFP_ATOMIC);
497 		if (unlikely(!skb)) {
498 			qdisc_qstats_drop(sch);
499 			goto finish_segs;
500 		}
501 		if (skb->ip_summed == CHECKSUM_PARTIAL &&
502 		    skb_checksum_help(skb)) {
503 			qdisc_drop(skb, sch, to_free);
504 			goto finish_segs;
505 		}
506 
507 		skb->data[prandom_u32() % skb_headlen(skb)] ^=
508 			1<<(prandom_u32() % 8);
509 	}
510 
511 	if (unlikely(sch->q.qlen >= sch->limit))
512 		return qdisc_drop_all(skb, sch, to_free);
513 
514 	qdisc_qstats_backlog_inc(sch, skb);
515 
516 	cb = netem_skb_cb(skb);
517 	if (q->gap == 0 ||		/* not doing reordering */
518 	    q->counter < q->gap - 1 ||	/* inside last reordering gap */
519 	    q->reorder < get_crandom(&q->reorder_cor)) {
520 		u64 now;
521 		s64 delay;
522 
523 		delay = tabledist(q->latency, q->jitter,
524 				  &q->delay_cor, q->delay_dist);
525 
526 		now = ktime_get_ns();
527 
528 		if (q->rate) {
529 			struct netem_skb_cb *last = NULL;
530 
531 			if (sch->q.tail)
532 				last = netem_skb_cb(sch->q.tail);
533 			if (q->t_root.rb_node) {
534 				struct sk_buff *t_skb;
535 				struct netem_skb_cb *t_last;
536 
537 				t_skb = skb_rb_last(&q->t_root);
538 				t_last = netem_skb_cb(t_skb);
539 				if (!last ||
540 				    t_last->time_to_send > last->time_to_send) {
541 					last = t_last;
542 				}
543 			}
544 
545 			if (last) {
546 				/*
547 				 * Last packet in queue is reference point (now),
548 				 * calculate this time bonus and subtract
549 				 * from delay.
550 				 */
551 				delay -= last->time_to_send - now;
552 				delay = max_t(s64, 0, delay);
553 				now = last->time_to_send;
554 			}
555 
556 			delay += packet_time_ns(qdisc_pkt_len(skb), q);
557 		}
558 
559 		cb->time_to_send = now + delay;
560 		++q->counter;
561 		tfifo_enqueue(skb, sch);
562 	} else {
563 		/*
564 		 * Do re-ordering by putting one out of N packets at the front
565 		 * of the queue.
566 		 */
567 		cb->time_to_send = ktime_get_ns();
568 		q->counter = 0;
569 
570 		netem_enqueue_skb_head(&sch->q, skb);
571 		sch->qstats.requeues++;
572 	}
573 
574 finish_segs:
575 	if (segs) {
576 		while (segs) {
577 			skb2 = segs->next;
578 			segs->next = NULL;
579 			qdisc_skb_cb(segs)->pkt_len = segs->len;
580 			last_len = segs->len;
581 			rc = qdisc_enqueue(segs, sch, to_free);
582 			if (rc != NET_XMIT_SUCCESS) {
583 				if (net_xmit_drop_count(rc))
584 					qdisc_qstats_drop(sch);
585 			} else {
586 				nb++;
587 				len += last_len;
588 			}
589 			segs = skb2;
590 		}
591 		sch->q.qlen += nb;
592 		if (nb > 1)
593 			qdisc_tree_reduce_backlog(sch, 1 - nb, prev_len - len);
594 	}
595 	return NET_XMIT_SUCCESS;
596 }
597 
598 /* Delay the next round with a new future slot with a
599  * correct number of bytes and packets.
600  */
601 
602 static void get_slot_next(struct netem_sched_data *q, u64 now)
603 {
604 	q->slot.slot_next = now + q->slot_config.min_delay +
605 		(prandom_u32() *
606 			(q->slot_config.max_delay -
607 				q->slot_config.min_delay) >> 32);
608 	q->slot.packets_left = q->slot_config.max_packets;
609 	q->slot.bytes_left = q->slot_config.max_bytes;
610 }
611 
612 static struct sk_buff *netem_dequeue(struct Qdisc *sch)
613 {
614 	struct netem_sched_data *q = qdisc_priv(sch);
615 	struct sk_buff *skb;
616 	struct rb_node *p;
617 
618 tfifo_dequeue:
619 	skb = __qdisc_dequeue_head(&sch->q);
620 	if (skb) {
621 		qdisc_qstats_backlog_dec(sch, skb);
622 deliver:
623 		qdisc_bstats_update(sch, skb);
624 		return skb;
625 	}
626 	p = rb_first(&q->t_root);
627 	if (p) {
628 		u64 time_to_send;
629 		u64 now = ktime_get_ns();
630 
631 		skb = rb_to_skb(p);
632 
633 		/* if more time remaining? */
634 		time_to_send = netem_skb_cb(skb)->time_to_send;
635 		if (q->slot.slot_next && q->slot.slot_next < time_to_send)
636 			get_slot_next(q, now);
637 
638 		if (time_to_send <= now &&  q->slot.slot_next <= now) {
639 			rb_erase(p, &q->t_root);
640 			sch->q.qlen--;
641 			qdisc_qstats_backlog_dec(sch, skb);
642 			skb->next = NULL;
643 			skb->prev = NULL;
644 			/* skb->dev shares skb->rbnode area,
645 			 * we need to restore its value.
646 			 */
647 			skb->dev = qdisc_dev(sch);
648 
649 #ifdef CONFIG_NET_CLS_ACT
650 			/*
651 			 * If it's at ingress let's pretend the delay is
652 			 * from the network (tstamp will be updated).
653 			 */
654 			if (skb->tc_redirected && skb->tc_from_ingress)
655 				skb->tstamp = 0;
656 #endif
657 
658 			if (q->slot.slot_next) {
659 				q->slot.packets_left--;
660 				q->slot.bytes_left -= qdisc_pkt_len(skb);
661 				if (q->slot.packets_left <= 0 ||
662 				    q->slot.bytes_left <= 0)
663 					get_slot_next(q, now);
664 			}
665 
666 			if (q->qdisc) {
667 				unsigned int pkt_len = qdisc_pkt_len(skb);
668 				struct sk_buff *to_free = NULL;
669 				int err;
670 
671 				err = qdisc_enqueue(skb, q->qdisc, &to_free);
672 				kfree_skb_list(to_free);
673 				if (err != NET_XMIT_SUCCESS &&
674 				    net_xmit_drop_count(err)) {
675 					qdisc_qstats_drop(sch);
676 					qdisc_tree_reduce_backlog(sch, 1,
677 								  pkt_len);
678 				}
679 				goto tfifo_dequeue;
680 			}
681 			goto deliver;
682 		}
683 
684 		if (q->qdisc) {
685 			skb = q->qdisc->ops->dequeue(q->qdisc);
686 			if (skb)
687 				goto deliver;
688 		}
689 
690 		qdisc_watchdog_schedule_ns(&q->watchdog,
691 					   max(time_to_send,
692 					       q->slot.slot_next));
693 	}
694 
695 	if (q->qdisc) {
696 		skb = q->qdisc->ops->dequeue(q->qdisc);
697 		if (skb)
698 			goto deliver;
699 	}
700 	return NULL;
701 }
702 
703 static void netem_reset(struct Qdisc *sch)
704 {
705 	struct netem_sched_data *q = qdisc_priv(sch);
706 
707 	qdisc_reset_queue(sch);
708 	tfifo_reset(sch);
709 	if (q->qdisc)
710 		qdisc_reset(q->qdisc);
711 	qdisc_watchdog_cancel(&q->watchdog);
712 }
713 
714 static void dist_free(struct disttable *d)
715 {
716 	kvfree(d);
717 }
718 
719 /*
720  * Distribution data is a variable size payload containing
721  * signed 16 bit values.
722  */
723 
724 static int get_dist_table(struct Qdisc *sch, const struct nlattr *attr)
725 {
726 	struct netem_sched_data *q = qdisc_priv(sch);
727 	size_t n = nla_len(attr)/sizeof(__s16);
728 	const __s16 *data = nla_data(attr);
729 	spinlock_t *root_lock;
730 	struct disttable *d;
731 	int i;
732 
733 	if (n > NETEM_DIST_MAX)
734 		return -EINVAL;
735 
736 	d = kvmalloc(sizeof(struct disttable) + n * sizeof(s16), GFP_KERNEL);
737 	if (!d)
738 		return -ENOMEM;
739 
740 	d->size = n;
741 	for (i = 0; i < n; i++)
742 		d->table[i] = data[i];
743 
744 	root_lock = qdisc_root_sleeping_lock(sch);
745 
746 	spin_lock_bh(root_lock);
747 	swap(q->delay_dist, d);
748 	spin_unlock_bh(root_lock);
749 
750 	dist_free(d);
751 	return 0;
752 }
753 
754 static void get_slot(struct netem_sched_data *q, const struct nlattr *attr)
755 {
756 	const struct tc_netem_slot *c = nla_data(attr);
757 
758 	q->slot_config = *c;
759 	if (q->slot_config.max_packets == 0)
760 		q->slot_config.max_packets = INT_MAX;
761 	if (q->slot_config.max_bytes == 0)
762 		q->slot_config.max_bytes = INT_MAX;
763 	q->slot.packets_left = q->slot_config.max_packets;
764 	q->slot.bytes_left = q->slot_config.max_bytes;
765 	if (q->slot_config.min_delay | q->slot_config.max_delay)
766 		q->slot.slot_next = ktime_get_ns();
767 	else
768 		q->slot.slot_next = 0;
769 }
770 
771 static void get_correlation(struct netem_sched_data *q, const struct nlattr *attr)
772 {
773 	const struct tc_netem_corr *c = nla_data(attr);
774 
775 	init_crandom(&q->delay_cor, c->delay_corr);
776 	init_crandom(&q->loss_cor, c->loss_corr);
777 	init_crandom(&q->dup_cor, c->dup_corr);
778 }
779 
780 static void get_reorder(struct netem_sched_data *q, const struct nlattr *attr)
781 {
782 	const struct tc_netem_reorder *r = nla_data(attr);
783 
784 	q->reorder = r->probability;
785 	init_crandom(&q->reorder_cor, r->correlation);
786 }
787 
788 static void get_corrupt(struct netem_sched_data *q, const struct nlattr *attr)
789 {
790 	const struct tc_netem_corrupt *r = nla_data(attr);
791 
792 	q->corrupt = r->probability;
793 	init_crandom(&q->corrupt_cor, r->correlation);
794 }
795 
796 static void get_rate(struct netem_sched_data *q, const struct nlattr *attr)
797 {
798 	const struct tc_netem_rate *r = nla_data(attr);
799 
800 	q->rate = r->rate;
801 	q->packet_overhead = r->packet_overhead;
802 	q->cell_size = r->cell_size;
803 	q->cell_overhead = r->cell_overhead;
804 	if (q->cell_size)
805 		q->cell_size_reciprocal = reciprocal_value(q->cell_size);
806 	else
807 		q->cell_size_reciprocal = (struct reciprocal_value) { 0 };
808 }
809 
810 static int get_loss_clg(struct netem_sched_data *q, const struct nlattr *attr)
811 {
812 	const struct nlattr *la;
813 	int rem;
814 
815 	nla_for_each_nested(la, attr, rem) {
816 		u16 type = nla_type(la);
817 
818 		switch (type) {
819 		case NETEM_LOSS_GI: {
820 			const struct tc_netem_gimodel *gi = nla_data(la);
821 
822 			if (nla_len(la) < sizeof(struct tc_netem_gimodel)) {
823 				pr_info("netem: incorrect gi model size\n");
824 				return -EINVAL;
825 			}
826 
827 			q->loss_model = CLG_4_STATES;
828 
829 			q->clg.state = TX_IN_GAP_PERIOD;
830 			q->clg.a1 = gi->p13;
831 			q->clg.a2 = gi->p31;
832 			q->clg.a3 = gi->p32;
833 			q->clg.a4 = gi->p14;
834 			q->clg.a5 = gi->p23;
835 			break;
836 		}
837 
838 		case NETEM_LOSS_GE: {
839 			const struct tc_netem_gemodel *ge = nla_data(la);
840 
841 			if (nla_len(la) < sizeof(struct tc_netem_gemodel)) {
842 				pr_info("netem: incorrect ge model size\n");
843 				return -EINVAL;
844 			}
845 
846 			q->loss_model = CLG_GILB_ELL;
847 			q->clg.state = GOOD_STATE;
848 			q->clg.a1 = ge->p;
849 			q->clg.a2 = ge->r;
850 			q->clg.a3 = ge->h;
851 			q->clg.a4 = ge->k1;
852 			break;
853 		}
854 
855 		default:
856 			pr_info("netem: unknown loss type %u\n", type);
857 			return -EINVAL;
858 		}
859 	}
860 
861 	return 0;
862 }
863 
864 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
865 	[TCA_NETEM_CORR]	= { .len = sizeof(struct tc_netem_corr) },
866 	[TCA_NETEM_REORDER]	= { .len = sizeof(struct tc_netem_reorder) },
867 	[TCA_NETEM_CORRUPT]	= { .len = sizeof(struct tc_netem_corrupt) },
868 	[TCA_NETEM_RATE]	= { .len = sizeof(struct tc_netem_rate) },
869 	[TCA_NETEM_LOSS]	= { .type = NLA_NESTED },
870 	[TCA_NETEM_ECN]		= { .type = NLA_U32 },
871 	[TCA_NETEM_RATE64]	= { .type = NLA_U64 },
872 	[TCA_NETEM_LATENCY64]	= { .type = NLA_S64 },
873 	[TCA_NETEM_JITTER64]	= { .type = NLA_S64 },
874 	[TCA_NETEM_SLOT]	= { .len = sizeof(struct tc_netem_slot) },
875 };
876 
877 static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
878 		      const struct nla_policy *policy, int len)
879 {
880 	int nested_len = nla_len(nla) - NLA_ALIGN(len);
881 
882 	if (nested_len < 0) {
883 		pr_info("netem: invalid attributes len %d\n", nested_len);
884 		return -EINVAL;
885 	}
886 
887 	if (nested_len >= nla_attr_size(0))
888 		return nla_parse(tb, maxtype, nla_data(nla) + NLA_ALIGN(len),
889 				 nested_len, policy, NULL);
890 
891 	memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
892 	return 0;
893 }
894 
895 /* Parse netlink message to set options */
896 static int netem_change(struct Qdisc *sch, struct nlattr *opt,
897 			struct netlink_ext_ack *extack)
898 {
899 	struct netem_sched_data *q = qdisc_priv(sch);
900 	struct nlattr *tb[TCA_NETEM_MAX + 1];
901 	struct tc_netem_qopt *qopt;
902 	struct clgstate old_clg;
903 	int old_loss_model = CLG_RANDOM;
904 	int ret;
905 
906 	if (opt == NULL)
907 		return -EINVAL;
908 
909 	qopt = nla_data(opt);
910 	ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
911 	if (ret < 0)
912 		return ret;
913 
914 	/* backup q->clg and q->loss_model */
915 	old_clg = q->clg;
916 	old_loss_model = q->loss_model;
917 
918 	if (tb[TCA_NETEM_LOSS]) {
919 		ret = get_loss_clg(q, tb[TCA_NETEM_LOSS]);
920 		if (ret) {
921 			q->loss_model = old_loss_model;
922 			return ret;
923 		}
924 	} else {
925 		q->loss_model = CLG_RANDOM;
926 	}
927 
928 	if (tb[TCA_NETEM_DELAY_DIST]) {
929 		ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST]);
930 		if (ret) {
931 			/* recover clg and loss_model, in case of
932 			 * q->clg and q->loss_model were modified
933 			 * in get_loss_clg()
934 			 */
935 			q->clg = old_clg;
936 			q->loss_model = old_loss_model;
937 			return ret;
938 		}
939 	}
940 
941 	sch->limit = qopt->limit;
942 
943 	q->latency = PSCHED_TICKS2NS(qopt->latency);
944 	q->jitter = PSCHED_TICKS2NS(qopt->jitter);
945 	q->limit = qopt->limit;
946 	q->gap = qopt->gap;
947 	q->counter = 0;
948 	q->loss = qopt->loss;
949 	q->duplicate = qopt->duplicate;
950 
951 	/* for compatibility with earlier versions.
952 	 * if gap is set, need to assume 100% probability
953 	 */
954 	if (q->gap)
955 		q->reorder = ~0;
956 
957 	if (tb[TCA_NETEM_CORR])
958 		get_correlation(q, tb[TCA_NETEM_CORR]);
959 
960 	if (tb[TCA_NETEM_REORDER])
961 		get_reorder(q, tb[TCA_NETEM_REORDER]);
962 
963 	if (tb[TCA_NETEM_CORRUPT])
964 		get_corrupt(q, tb[TCA_NETEM_CORRUPT]);
965 
966 	if (tb[TCA_NETEM_RATE])
967 		get_rate(q, tb[TCA_NETEM_RATE]);
968 
969 	if (tb[TCA_NETEM_RATE64])
970 		q->rate = max_t(u64, q->rate,
971 				nla_get_u64(tb[TCA_NETEM_RATE64]));
972 
973 	if (tb[TCA_NETEM_LATENCY64])
974 		q->latency = nla_get_s64(tb[TCA_NETEM_LATENCY64]);
975 
976 	if (tb[TCA_NETEM_JITTER64])
977 		q->jitter = nla_get_s64(tb[TCA_NETEM_JITTER64]);
978 
979 	if (tb[TCA_NETEM_ECN])
980 		q->ecn = nla_get_u32(tb[TCA_NETEM_ECN]);
981 
982 	if (tb[TCA_NETEM_SLOT])
983 		get_slot(q, tb[TCA_NETEM_SLOT]);
984 
985 	return ret;
986 }
987 
988 static int netem_init(struct Qdisc *sch, struct nlattr *opt,
989 		      struct netlink_ext_ack *extack)
990 {
991 	struct netem_sched_data *q = qdisc_priv(sch);
992 	int ret;
993 
994 	qdisc_watchdog_init(&q->watchdog, sch);
995 
996 	if (!opt)
997 		return -EINVAL;
998 
999 	q->loss_model = CLG_RANDOM;
1000 	ret = netem_change(sch, opt, extack);
1001 	if (ret)
1002 		pr_info("netem: change failed\n");
1003 	return ret;
1004 }
1005 
1006 static void netem_destroy(struct Qdisc *sch)
1007 {
1008 	struct netem_sched_data *q = qdisc_priv(sch);
1009 
1010 	qdisc_watchdog_cancel(&q->watchdog);
1011 	if (q->qdisc)
1012 		qdisc_destroy(q->qdisc);
1013 	dist_free(q->delay_dist);
1014 }
1015 
1016 static int dump_loss_model(const struct netem_sched_data *q,
1017 			   struct sk_buff *skb)
1018 {
1019 	struct nlattr *nest;
1020 
1021 	nest = nla_nest_start(skb, TCA_NETEM_LOSS);
1022 	if (nest == NULL)
1023 		goto nla_put_failure;
1024 
1025 	switch (q->loss_model) {
1026 	case CLG_RANDOM:
1027 		/* legacy loss model */
1028 		nla_nest_cancel(skb, nest);
1029 		return 0;	/* no data */
1030 
1031 	case CLG_4_STATES: {
1032 		struct tc_netem_gimodel gi = {
1033 			.p13 = q->clg.a1,
1034 			.p31 = q->clg.a2,
1035 			.p32 = q->clg.a3,
1036 			.p14 = q->clg.a4,
1037 			.p23 = q->clg.a5,
1038 		};
1039 
1040 		if (nla_put(skb, NETEM_LOSS_GI, sizeof(gi), &gi))
1041 			goto nla_put_failure;
1042 		break;
1043 	}
1044 	case CLG_GILB_ELL: {
1045 		struct tc_netem_gemodel ge = {
1046 			.p = q->clg.a1,
1047 			.r = q->clg.a2,
1048 			.h = q->clg.a3,
1049 			.k1 = q->clg.a4,
1050 		};
1051 
1052 		if (nla_put(skb, NETEM_LOSS_GE, sizeof(ge), &ge))
1053 			goto nla_put_failure;
1054 		break;
1055 	}
1056 	}
1057 
1058 	nla_nest_end(skb, nest);
1059 	return 0;
1060 
1061 nla_put_failure:
1062 	nla_nest_cancel(skb, nest);
1063 	return -1;
1064 }
1065 
1066 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
1067 {
1068 	const struct netem_sched_data *q = qdisc_priv(sch);
1069 	struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb);
1070 	struct tc_netem_qopt qopt;
1071 	struct tc_netem_corr cor;
1072 	struct tc_netem_reorder reorder;
1073 	struct tc_netem_corrupt corrupt;
1074 	struct tc_netem_rate rate;
1075 	struct tc_netem_slot slot;
1076 
1077 	qopt.latency = min_t(psched_tdiff_t, PSCHED_NS2TICKS(q->latency),
1078 			     UINT_MAX);
1079 	qopt.jitter = min_t(psched_tdiff_t, PSCHED_NS2TICKS(q->jitter),
1080 			    UINT_MAX);
1081 	qopt.limit = q->limit;
1082 	qopt.loss = q->loss;
1083 	qopt.gap = q->gap;
1084 	qopt.duplicate = q->duplicate;
1085 	if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
1086 		goto nla_put_failure;
1087 
1088 	if (nla_put(skb, TCA_NETEM_LATENCY64, sizeof(q->latency), &q->latency))
1089 		goto nla_put_failure;
1090 
1091 	if (nla_put(skb, TCA_NETEM_JITTER64, sizeof(q->jitter), &q->jitter))
1092 		goto nla_put_failure;
1093 
1094 	cor.delay_corr = q->delay_cor.rho;
1095 	cor.loss_corr = q->loss_cor.rho;
1096 	cor.dup_corr = q->dup_cor.rho;
1097 	if (nla_put(skb, TCA_NETEM_CORR, sizeof(cor), &cor))
1098 		goto nla_put_failure;
1099 
1100 	reorder.probability = q->reorder;
1101 	reorder.correlation = q->reorder_cor.rho;
1102 	if (nla_put(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder))
1103 		goto nla_put_failure;
1104 
1105 	corrupt.probability = q->corrupt;
1106 	corrupt.correlation = q->corrupt_cor.rho;
1107 	if (nla_put(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt))
1108 		goto nla_put_failure;
1109 
1110 	if (q->rate >= (1ULL << 32)) {
1111 		if (nla_put_u64_64bit(skb, TCA_NETEM_RATE64, q->rate,
1112 				      TCA_NETEM_PAD))
1113 			goto nla_put_failure;
1114 		rate.rate = ~0U;
1115 	} else {
1116 		rate.rate = q->rate;
1117 	}
1118 	rate.packet_overhead = q->packet_overhead;
1119 	rate.cell_size = q->cell_size;
1120 	rate.cell_overhead = q->cell_overhead;
1121 	if (nla_put(skb, TCA_NETEM_RATE, sizeof(rate), &rate))
1122 		goto nla_put_failure;
1123 
1124 	if (q->ecn && nla_put_u32(skb, TCA_NETEM_ECN, q->ecn))
1125 		goto nla_put_failure;
1126 
1127 	if (dump_loss_model(q, skb) != 0)
1128 		goto nla_put_failure;
1129 
1130 	if (q->slot_config.min_delay | q->slot_config.max_delay) {
1131 		slot = q->slot_config;
1132 		if (slot.max_packets == INT_MAX)
1133 			slot.max_packets = 0;
1134 		if (slot.max_bytes == INT_MAX)
1135 			slot.max_bytes = 0;
1136 		if (nla_put(skb, TCA_NETEM_SLOT, sizeof(slot), &slot))
1137 			goto nla_put_failure;
1138 	}
1139 
1140 	return nla_nest_end(skb, nla);
1141 
1142 nla_put_failure:
1143 	nlmsg_trim(skb, nla);
1144 	return -1;
1145 }
1146 
1147 static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
1148 			  struct sk_buff *skb, struct tcmsg *tcm)
1149 {
1150 	struct netem_sched_data *q = qdisc_priv(sch);
1151 
1152 	if (cl != 1 || !q->qdisc) 	/* only one class */
1153 		return -ENOENT;
1154 
1155 	tcm->tcm_handle |= TC_H_MIN(1);
1156 	tcm->tcm_info = q->qdisc->handle;
1157 
1158 	return 0;
1159 }
1160 
1161 static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1162 		     struct Qdisc **old, struct netlink_ext_ack *extack)
1163 {
1164 	struct netem_sched_data *q = qdisc_priv(sch);
1165 
1166 	*old = qdisc_replace(sch, new, &q->qdisc);
1167 	return 0;
1168 }
1169 
1170 static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
1171 {
1172 	struct netem_sched_data *q = qdisc_priv(sch);
1173 	return q->qdisc;
1174 }
1175 
1176 static unsigned long netem_find(struct Qdisc *sch, u32 classid)
1177 {
1178 	return 1;
1179 }
1180 
1181 static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
1182 {
1183 	if (!walker->stop) {
1184 		if (walker->count >= walker->skip)
1185 			if (walker->fn(sch, 1, walker) < 0) {
1186 				walker->stop = 1;
1187 				return;
1188 			}
1189 		walker->count++;
1190 	}
1191 }
1192 
1193 static const struct Qdisc_class_ops netem_class_ops = {
1194 	.graft		=	netem_graft,
1195 	.leaf		=	netem_leaf,
1196 	.find		=	netem_find,
1197 	.walk		=	netem_walk,
1198 	.dump		=	netem_dump_class,
1199 };
1200 
1201 static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
1202 	.id		=	"netem",
1203 	.cl_ops		=	&netem_class_ops,
1204 	.priv_size	=	sizeof(struct netem_sched_data),
1205 	.enqueue	=	netem_enqueue,
1206 	.dequeue	=	netem_dequeue,
1207 	.peek		=	qdisc_peek_dequeued,
1208 	.init		=	netem_init,
1209 	.reset		=	netem_reset,
1210 	.destroy	=	netem_destroy,
1211 	.change		=	netem_change,
1212 	.dump		=	netem_dump,
1213 	.owner		=	THIS_MODULE,
1214 };
1215 
1216 
1217 static int __init netem_module_init(void)
1218 {
1219 	pr_info("netem: version " VERSION "\n");
1220 	return register_qdisc(&netem_qdisc_ops);
1221 }
1222 static void __exit netem_module_exit(void)
1223 {
1224 	unregister_qdisc(&netem_qdisc_ops);
1225 }
1226 module_init(netem_module_init)
1227 module_exit(netem_module_exit)
1228 MODULE_LICENSE("GPL");
1229