xref: /openbmc/linux/net/sched/sch_netem.c (revision ebd09753)
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 disttable {
72 	u32  size;
73 	s16 table[0];
74 };
75 
76 struct netem_sched_data {
77 	/* internal t(ime)fifo qdisc uses t_root and sch->limit */
78 	struct rb_root t_root;
79 
80 	/* optional qdisc for classful handling (NULL at netem init) */
81 	struct Qdisc	*qdisc;
82 
83 	struct qdisc_watchdog watchdog;
84 
85 	s64 latency;
86 	s64 jitter;
87 
88 	u32 loss;
89 	u32 ecn;
90 	u32 limit;
91 	u32 counter;
92 	u32 gap;
93 	u32 duplicate;
94 	u32 reorder;
95 	u32 corrupt;
96 	u64 rate;
97 	s32 packet_overhead;
98 	u32 cell_size;
99 	struct reciprocal_value cell_size_reciprocal;
100 	s32 cell_overhead;
101 
102 	struct crndstate {
103 		u32 last;
104 		u32 rho;
105 	} delay_cor, loss_cor, dup_cor, reorder_cor, corrupt_cor;
106 
107 	struct disttable *delay_dist;
108 
109 	enum  {
110 		CLG_RANDOM,
111 		CLG_4_STATES,
112 		CLG_GILB_ELL,
113 	} loss_model;
114 
115 	enum {
116 		TX_IN_GAP_PERIOD = 1,
117 		TX_IN_BURST_PERIOD,
118 		LOST_IN_GAP_PERIOD,
119 		LOST_IN_BURST_PERIOD,
120 	} _4_state_model;
121 
122 	enum {
123 		GOOD_STATE = 1,
124 		BAD_STATE,
125 	} GE_state_model;
126 
127 	/* Correlated Loss Generation models */
128 	struct clgstate {
129 		/* state of the Markov chain */
130 		u8 state;
131 
132 		/* 4-states and Gilbert-Elliot models */
133 		u32 a1;	/* p13 for 4-states or p for GE */
134 		u32 a2;	/* p31 for 4-states or r for GE */
135 		u32 a3;	/* p32 for 4-states or h for GE */
136 		u32 a4;	/* p14 for 4-states or 1-k for GE */
137 		u32 a5; /* p23 used only in 4-states */
138 	} clg;
139 
140 	struct tc_netem_slot slot_config;
141 	struct slotstate {
142 		u64 slot_next;
143 		s32 packets_left;
144 		s32 bytes_left;
145 	} slot;
146 
147 	struct disttable *slot_dist;
148 };
149 
150 /* Time stamp put into socket buffer control block
151  * Only valid when skbs are in our internal t(ime)fifo queue.
152  *
153  * As skb->rbnode uses same storage than skb->next, skb->prev and skb->tstamp,
154  * and skb->next & skb->prev are scratch space for a qdisc,
155  * we save skb->tstamp value in skb->cb[] before destroying it.
156  */
157 struct netem_skb_cb {
158 	u64	        time_to_send;
159 };
160 
161 static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb)
162 {
163 	/* we assume we can use skb next/prev/tstamp as storage for rb_node */
164 	qdisc_cb_private_validate(skb, sizeof(struct netem_skb_cb));
165 	return (struct netem_skb_cb *)qdisc_skb_cb(skb)->data;
166 }
167 
168 /* init_crandom - initialize correlated random number generator
169  * Use entropy source for initial seed.
170  */
171 static void init_crandom(struct crndstate *state, unsigned long rho)
172 {
173 	state->rho = rho;
174 	state->last = prandom_u32();
175 }
176 
177 /* get_crandom - correlated random number generator
178  * Next number depends on last value.
179  * rho is scaled to avoid floating point.
180  */
181 static u32 get_crandom(struct crndstate *state)
182 {
183 	u64 value, rho;
184 	unsigned long answer;
185 
186 	if (!state || state->rho == 0)	/* no correlation */
187 		return prandom_u32();
188 
189 	value = prandom_u32();
190 	rho = (u64)state->rho + 1;
191 	answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32;
192 	state->last = answer;
193 	return answer;
194 }
195 
196 /* loss_4state - 4-state model loss generator
197  * Generates losses according to the 4-state Markov chain adopted in
198  * the GI (General and Intuitive) loss model.
199  */
200 static bool loss_4state(struct netem_sched_data *q)
201 {
202 	struct clgstate *clg = &q->clg;
203 	u32 rnd = prandom_u32();
204 
205 	/*
206 	 * Makes a comparison between rnd and the transition
207 	 * probabilities outgoing from the current state, then decides the
208 	 * next state and if the next packet has to be transmitted or lost.
209 	 * The four states correspond to:
210 	 *   TX_IN_GAP_PERIOD => successfully transmitted packets within a gap period
211 	 *   LOST_IN_BURST_PERIOD => isolated losses within a gap period
212 	 *   LOST_IN_GAP_PERIOD => lost packets within a burst period
213 	 *   TX_IN_GAP_PERIOD => successfully transmitted packets within a burst period
214 	 */
215 	switch (clg->state) {
216 	case TX_IN_GAP_PERIOD:
217 		if (rnd < clg->a4) {
218 			clg->state = LOST_IN_BURST_PERIOD;
219 			return true;
220 		} else if (clg->a4 < rnd && rnd < clg->a1 + clg->a4) {
221 			clg->state = LOST_IN_GAP_PERIOD;
222 			return true;
223 		} else if (clg->a1 + clg->a4 < rnd) {
224 			clg->state = TX_IN_GAP_PERIOD;
225 		}
226 
227 		break;
228 	case TX_IN_BURST_PERIOD:
229 		if (rnd < clg->a5) {
230 			clg->state = LOST_IN_GAP_PERIOD;
231 			return true;
232 		} else {
233 			clg->state = TX_IN_BURST_PERIOD;
234 		}
235 
236 		break;
237 	case LOST_IN_GAP_PERIOD:
238 		if (rnd < clg->a3)
239 			clg->state = TX_IN_BURST_PERIOD;
240 		else if (clg->a3 < rnd && rnd < clg->a2 + clg->a3) {
241 			clg->state = TX_IN_GAP_PERIOD;
242 		} else if (clg->a2 + clg->a3 < rnd) {
243 			clg->state = LOST_IN_GAP_PERIOD;
244 			return true;
245 		}
246 		break;
247 	case LOST_IN_BURST_PERIOD:
248 		clg->state = TX_IN_GAP_PERIOD;
249 		break;
250 	}
251 
252 	return false;
253 }
254 
255 /* loss_gilb_ell - Gilbert-Elliot model loss generator
256  * Generates losses according to the Gilbert-Elliot loss model or
257  * its special cases  (Gilbert or Simple Gilbert)
258  *
259  * Makes a comparison between random number and the transition
260  * probabilities outgoing from the current state, then decides the
261  * next state. A second random number is extracted and the comparison
262  * with the loss probability of the current state decides if the next
263  * packet will be transmitted or lost.
264  */
265 static bool loss_gilb_ell(struct netem_sched_data *q)
266 {
267 	struct clgstate *clg = &q->clg;
268 
269 	switch (clg->state) {
270 	case GOOD_STATE:
271 		if (prandom_u32() < clg->a1)
272 			clg->state = BAD_STATE;
273 		if (prandom_u32() < clg->a4)
274 			return true;
275 		break;
276 	case BAD_STATE:
277 		if (prandom_u32() < clg->a2)
278 			clg->state = GOOD_STATE;
279 		if (prandom_u32() > clg->a3)
280 			return true;
281 	}
282 
283 	return false;
284 }
285 
286 static bool loss_event(struct netem_sched_data *q)
287 {
288 	switch (q->loss_model) {
289 	case CLG_RANDOM:
290 		/* Random packet drop 0 => none, ~0 => all */
291 		return q->loss && q->loss >= get_crandom(&q->loss_cor);
292 
293 	case CLG_4_STATES:
294 		/* 4state loss model algorithm (used also for GI model)
295 		* Extracts a value from the markov 4 state loss generator,
296 		* if it is 1 drops a packet and if needed writes the event in
297 		* the kernel logs
298 		*/
299 		return loss_4state(q);
300 
301 	case CLG_GILB_ELL:
302 		/* Gilbert-Elliot loss model algorithm
303 		* Extracts a value from the Gilbert-Elliot loss generator,
304 		* if it is 1 drops a packet and if needed writes the event in
305 		* the kernel logs
306 		*/
307 		return loss_gilb_ell(q);
308 	}
309 
310 	return false;	/* not reached */
311 }
312 
313 
314 /* tabledist - return a pseudo-randomly distributed value with mean mu and
315  * std deviation sigma.  Uses table lookup to approximate the desired
316  * distribution, and a uniformly-distributed pseudo-random source.
317  */
318 static s64 tabledist(s64 mu, s32 sigma,
319 		     struct crndstate *state,
320 		     const struct disttable *dist)
321 {
322 	s64 x;
323 	long t;
324 	u32 rnd;
325 
326 	if (sigma == 0)
327 		return mu;
328 
329 	rnd = get_crandom(state);
330 
331 	/* default uniform distribution */
332 	if (dist == NULL)
333 		return ((rnd % (2 * sigma)) + mu) - sigma;
334 
335 	t = dist->table[rnd % dist->size];
336 	x = (sigma % NETEM_DIST_SCALE) * t;
337 	if (x >= 0)
338 		x += NETEM_DIST_SCALE/2;
339 	else
340 		x -= NETEM_DIST_SCALE/2;
341 
342 	return  x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu;
343 }
344 
345 static u64 packet_time_ns(u64 len, const struct netem_sched_data *q)
346 {
347 	len += q->packet_overhead;
348 
349 	if (q->cell_size) {
350 		u32 cells = reciprocal_divide(len, q->cell_size_reciprocal);
351 
352 		if (len > cells * q->cell_size)	/* extra cell needed for remainder */
353 			cells++;
354 		len = cells * (q->cell_size + q->cell_overhead);
355 	}
356 
357 	return div64_u64(len * NSEC_PER_SEC, q->rate);
358 }
359 
360 static void tfifo_reset(struct Qdisc *sch)
361 {
362 	struct netem_sched_data *q = qdisc_priv(sch);
363 	struct rb_node *p = rb_first(&q->t_root);
364 
365 	while (p) {
366 		struct sk_buff *skb = rb_to_skb(p);
367 
368 		p = rb_next(p);
369 		rb_erase(&skb->rbnode, &q->t_root);
370 		rtnl_kfree_skbs(skb, skb);
371 	}
372 }
373 
374 static void tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
375 {
376 	struct netem_sched_data *q = qdisc_priv(sch);
377 	u64 tnext = netem_skb_cb(nskb)->time_to_send;
378 	struct rb_node **p = &q->t_root.rb_node, *parent = NULL;
379 
380 	while (*p) {
381 		struct sk_buff *skb;
382 
383 		parent = *p;
384 		skb = rb_to_skb(parent);
385 		if (tnext >= netem_skb_cb(skb)->time_to_send)
386 			p = &parent->rb_right;
387 		else
388 			p = &parent->rb_left;
389 	}
390 	rb_link_node(&nskb->rbnode, parent, p);
391 	rb_insert_color(&nskb->rbnode, &q->t_root);
392 	sch->q.qlen++;
393 }
394 
395 /* netem can't properly corrupt a megapacket (like we get from GSO), so instead
396  * when we statistically choose to corrupt one, we instead segment it, returning
397  * the first packet to be corrupted, and re-enqueue the remaining frames
398  */
399 static struct sk_buff *netem_segment(struct sk_buff *skb, struct Qdisc *sch,
400 				     struct sk_buff **to_free)
401 {
402 	struct sk_buff *segs;
403 	netdev_features_t features = netif_skb_features(skb);
404 
405 	segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
406 
407 	if (IS_ERR_OR_NULL(segs)) {
408 		qdisc_drop(skb, sch, to_free);
409 		return NULL;
410 	}
411 	consume_skb(skb);
412 	return segs;
413 }
414 
415 /*
416  * Insert one skb into qdisc.
417  * Note: parent depends on return value to account for queue length.
418  * 	NET_XMIT_DROP: queue length didn't change.
419  *      NET_XMIT_SUCCESS: one skb was queued.
420  */
421 static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch,
422 			 struct sk_buff **to_free)
423 {
424 	struct netem_sched_data *q = qdisc_priv(sch);
425 	/* We don't fill cb now as skb_unshare() may invalidate it */
426 	struct netem_skb_cb *cb;
427 	struct sk_buff *skb2;
428 	struct sk_buff *segs = NULL;
429 	unsigned int len = 0, last_len, prev_len = qdisc_pkt_len(skb);
430 	int nb = 0;
431 	int count = 1;
432 	int rc = NET_XMIT_SUCCESS;
433 
434 	/* Random duplication */
435 	if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor))
436 		++count;
437 
438 	/* Drop packet? */
439 	if (loss_event(q)) {
440 		if (q->ecn && INET_ECN_set_ce(skb))
441 			qdisc_qstats_drop(sch); /* mark packet */
442 		else
443 			--count;
444 	}
445 	if (count == 0) {
446 		qdisc_qstats_drop(sch);
447 		__qdisc_drop(skb, to_free);
448 		return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
449 	}
450 
451 	/* If a delay is expected, orphan the skb. (orphaning usually takes
452 	 * place at TX completion time, so _before_ the link transit delay)
453 	 */
454 	if (q->latency || q->jitter || q->rate)
455 		skb_orphan_partial(skb);
456 
457 	/*
458 	 * If we need to duplicate packet, then re-insert at top of the
459 	 * qdisc tree, since parent queuer expects that only one
460 	 * skb will be queued.
461 	 */
462 	if (count > 1 && (skb2 = skb_clone(skb, GFP_ATOMIC)) != NULL) {
463 		struct Qdisc *rootq = qdisc_root(sch);
464 		u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
465 
466 		q->duplicate = 0;
467 		rootq->enqueue(skb2, rootq, to_free);
468 		q->duplicate = dupsave;
469 	}
470 
471 	/*
472 	 * Randomized packet corruption.
473 	 * Make copy if needed since we are modifying
474 	 * If packet is going to be hardware checksummed, then
475 	 * do it now in software before we mangle it.
476 	 */
477 	if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor)) {
478 		if (skb_is_gso(skb)) {
479 			segs = netem_segment(skb, sch, to_free);
480 			if (!segs)
481 				return NET_XMIT_DROP;
482 		} else {
483 			segs = skb;
484 		}
485 
486 		skb = segs;
487 		segs = segs->next;
488 
489 		skb = skb_unshare(skb, GFP_ATOMIC);
490 		if (unlikely(!skb)) {
491 			qdisc_qstats_drop(sch);
492 			goto finish_segs;
493 		}
494 		if (skb->ip_summed == CHECKSUM_PARTIAL &&
495 		    skb_checksum_help(skb)) {
496 			qdisc_drop(skb, sch, to_free);
497 			goto finish_segs;
498 		}
499 
500 		skb->data[prandom_u32() % skb_headlen(skb)] ^=
501 			1<<(prandom_u32() % 8);
502 	}
503 
504 	if (unlikely(sch->q.qlen >= sch->limit))
505 		return qdisc_drop_all(skb, sch, to_free);
506 
507 	qdisc_qstats_backlog_inc(sch, skb);
508 
509 	cb = netem_skb_cb(skb);
510 	if (q->gap == 0 ||		/* not doing reordering */
511 	    q->counter < q->gap - 1 ||	/* inside last reordering gap */
512 	    q->reorder < get_crandom(&q->reorder_cor)) {
513 		u64 now;
514 		s64 delay;
515 
516 		delay = tabledist(q->latency, q->jitter,
517 				  &q->delay_cor, q->delay_dist);
518 
519 		now = ktime_get_ns();
520 
521 		if (q->rate) {
522 			struct netem_skb_cb *last = NULL;
523 
524 			if (sch->q.tail)
525 				last = netem_skb_cb(sch->q.tail);
526 			if (q->t_root.rb_node) {
527 				struct sk_buff *t_skb;
528 				struct netem_skb_cb *t_last;
529 
530 				t_skb = skb_rb_last(&q->t_root);
531 				t_last = netem_skb_cb(t_skb);
532 				if (!last ||
533 				    t_last->time_to_send > last->time_to_send) {
534 					last = t_last;
535 				}
536 			}
537 
538 			if (last) {
539 				/*
540 				 * Last packet in queue is reference point (now),
541 				 * calculate this time bonus and subtract
542 				 * from delay.
543 				 */
544 				delay -= last->time_to_send - now;
545 				delay = max_t(s64, 0, delay);
546 				now = last->time_to_send;
547 			}
548 
549 			delay += packet_time_ns(qdisc_pkt_len(skb), q);
550 		}
551 
552 		cb->time_to_send = now + delay;
553 		++q->counter;
554 		tfifo_enqueue(skb, sch);
555 	} else {
556 		/*
557 		 * Do re-ordering by putting one out of N packets at the front
558 		 * of the queue.
559 		 */
560 		cb->time_to_send = ktime_get_ns();
561 		q->counter = 0;
562 
563 		__qdisc_enqueue_head(skb, &sch->q);
564 		sch->qstats.requeues++;
565 	}
566 
567 finish_segs:
568 	if (segs) {
569 		while (segs) {
570 			skb2 = segs->next;
571 			skb_mark_not_on_list(segs);
572 			qdisc_skb_cb(segs)->pkt_len = segs->len;
573 			last_len = segs->len;
574 			rc = qdisc_enqueue(segs, sch, to_free);
575 			if (rc != NET_XMIT_SUCCESS) {
576 				if (net_xmit_drop_count(rc))
577 					qdisc_qstats_drop(sch);
578 			} else {
579 				nb++;
580 				len += last_len;
581 			}
582 			segs = skb2;
583 		}
584 		sch->q.qlen += nb;
585 		if (nb > 1)
586 			qdisc_tree_reduce_backlog(sch, 1 - nb, prev_len - len);
587 	}
588 	return NET_XMIT_SUCCESS;
589 }
590 
591 /* Delay the next round with a new future slot with a
592  * correct number of bytes and packets.
593  */
594 
595 static void get_slot_next(struct netem_sched_data *q, u64 now)
596 {
597 	s64 next_delay;
598 
599 	if (!q->slot_dist)
600 		next_delay = q->slot_config.min_delay +
601 				(prandom_u32() *
602 				 (q->slot_config.max_delay -
603 				  q->slot_config.min_delay) >> 32);
604 	else
605 		next_delay = tabledist(q->slot_config.dist_delay,
606 				       (s32)(q->slot_config.dist_jitter),
607 				       NULL, q->slot_dist);
608 
609 	q->slot.slot_next = now + next_delay;
610 	q->slot.packets_left = q->slot_config.max_packets;
611 	q->slot.bytes_left = q->slot_config.max_bytes;
612 }
613 
614 static struct sk_buff *netem_dequeue(struct Qdisc *sch)
615 {
616 	struct netem_sched_data *q = qdisc_priv(sch);
617 	struct sk_buff *skb;
618 	struct rb_node *p;
619 
620 tfifo_dequeue:
621 	skb = __qdisc_dequeue_head(&sch->q);
622 	if (skb) {
623 		qdisc_qstats_backlog_dec(sch, skb);
624 deliver:
625 		qdisc_bstats_update(sch, skb);
626 		return skb;
627 	}
628 	p = rb_first(&q->t_root);
629 	if (p) {
630 		u64 time_to_send;
631 		u64 now = ktime_get_ns();
632 
633 		skb = rb_to_skb(p);
634 
635 		/* if more time remaining? */
636 		time_to_send = netem_skb_cb(skb)->time_to_send;
637 		if (q->slot.slot_next && q->slot.slot_next < time_to_send)
638 			get_slot_next(q, now);
639 
640 		if (time_to_send <= now &&  q->slot.slot_next <= now) {
641 			rb_erase(p, &q->t_root);
642 			sch->q.qlen--;
643 			qdisc_qstats_backlog_dec(sch, skb);
644 			skb->next = NULL;
645 			skb->prev = NULL;
646 			/* skb->dev shares skb->rbnode area,
647 			 * we need to restore its value.
648 			 */
649 			skb->dev = qdisc_dev(sch);
650 
651 #ifdef CONFIG_NET_CLS_ACT
652 			/*
653 			 * If it's at ingress let's pretend the delay is
654 			 * from the network (tstamp will be updated).
655 			 */
656 			if (skb->tc_redirected && skb->tc_from_ingress)
657 				skb->tstamp = 0;
658 #endif
659 
660 			if (q->slot.slot_next) {
661 				q->slot.packets_left--;
662 				q->slot.bytes_left -= qdisc_pkt_len(skb);
663 				if (q->slot.packets_left <= 0 ||
664 				    q->slot.bytes_left <= 0)
665 					get_slot_next(q, now);
666 			}
667 
668 			if (q->qdisc) {
669 				unsigned int pkt_len = qdisc_pkt_len(skb);
670 				struct sk_buff *to_free = NULL;
671 				int err;
672 
673 				err = qdisc_enqueue(skb, q->qdisc, &to_free);
674 				kfree_skb_list(to_free);
675 				if (err != NET_XMIT_SUCCESS &&
676 				    net_xmit_drop_count(err)) {
677 					qdisc_qstats_drop(sch);
678 					qdisc_tree_reduce_backlog(sch, 1,
679 								  pkt_len);
680 				}
681 				goto tfifo_dequeue;
682 			}
683 			goto deliver;
684 		}
685 
686 		if (q->qdisc) {
687 			skb = q->qdisc->ops->dequeue(q->qdisc);
688 			if (skb)
689 				goto deliver;
690 		}
691 
692 		qdisc_watchdog_schedule_ns(&q->watchdog,
693 					   max(time_to_send,
694 					       q->slot.slot_next));
695 	}
696 
697 	if (q->qdisc) {
698 		skb = q->qdisc->ops->dequeue(q->qdisc);
699 		if (skb)
700 			goto deliver;
701 	}
702 	return NULL;
703 }
704 
705 static void netem_reset(struct Qdisc *sch)
706 {
707 	struct netem_sched_data *q = qdisc_priv(sch);
708 
709 	qdisc_reset_queue(sch);
710 	tfifo_reset(sch);
711 	if (q->qdisc)
712 		qdisc_reset(q->qdisc);
713 	qdisc_watchdog_cancel(&q->watchdog);
714 }
715 
716 static void dist_free(struct disttable *d)
717 {
718 	kvfree(d);
719 }
720 
721 /*
722  * Distribution data is a variable size payload containing
723  * signed 16 bit values.
724  */
725 
726 static int get_dist_table(struct Qdisc *sch, struct disttable **tbl,
727 			  const struct nlattr *attr)
728 {
729 	size_t n = nla_len(attr)/sizeof(__s16);
730 	const __s16 *data = nla_data(attr);
731 	spinlock_t *root_lock;
732 	struct disttable *d;
733 	int i;
734 
735 	if (n > NETEM_DIST_MAX)
736 		return -EINVAL;
737 
738 	d = kvmalloc(sizeof(struct disttable) + n * sizeof(s16), GFP_KERNEL);
739 	if (!d)
740 		return -ENOMEM;
741 
742 	d->size = n;
743 	for (i = 0; i < n; i++)
744 		d->table[i] = data[i];
745 
746 	root_lock = qdisc_root_sleeping_lock(sch);
747 
748 	spin_lock_bh(root_lock);
749 	swap(*tbl, d);
750 	spin_unlock_bh(root_lock);
751 
752 	dist_free(d);
753 	return 0;
754 }
755 
756 static void get_slot(struct netem_sched_data *q, const struct nlattr *attr)
757 {
758 	const struct tc_netem_slot *c = nla_data(attr);
759 
760 	q->slot_config = *c;
761 	if (q->slot_config.max_packets == 0)
762 		q->slot_config.max_packets = INT_MAX;
763 	if (q->slot_config.max_bytes == 0)
764 		q->slot_config.max_bytes = INT_MAX;
765 	q->slot.packets_left = q->slot_config.max_packets;
766 	q->slot.bytes_left = q->slot_config.max_bytes;
767 	if (q->slot_config.min_delay | q->slot_config.max_delay |
768 	    q->slot_config.dist_jitter)
769 		q->slot.slot_next = ktime_get_ns();
770 	else
771 		q->slot.slot_next = 0;
772 }
773 
774 static void get_correlation(struct netem_sched_data *q, const struct nlattr *attr)
775 {
776 	const struct tc_netem_corr *c = nla_data(attr);
777 
778 	init_crandom(&q->delay_cor, c->delay_corr);
779 	init_crandom(&q->loss_cor, c->loss_corr);
780 	init_crandom(&q->dup_cor, c->dup_corr);
781 }
782 
783 static void get_reorder(struct netem_sched_data *q, const struct nlattr *attr)
784 {
785 	const struct tc_netem_reorder *r = nla_data(attr);
786 
787 	q->reorder = r->probability;
788 	init_crandom(&q->reorder_cor, r->correlation);
789 }
790 
791 static void get_corrupt(struct netem_sched_data *q, const struct nlattr *attr)
792 {
793 	const struct tc_netem_corrupt *r = nla_data(attr);
794 
795 	q->corrupt = r->probability;
796 	init_crandom(&q->corrupt_cor, r->correlation);
797 }
798 
799 static void get_rate(struct netem_sched_data *q, const struct nlattr *attr)
800 {
801 	const struct tc_netem_rate *r = nla_data(attr);
802 
803 	q->rate = r->rate;
804 	q->packet_overhead = r->packet_overhead;
805 	q->cell_size = r->cell_size;
806 	q->cell_overhead = r->cell_overhead;
807 	if (q->cell_size)
808 		q->cell_size_reciprocal = reciprocal_value(q->cell_size);
809 	else
810 		q->cell_size_reciprocal = (struct reciprocal_value) { 0 };
811 }
812 
813 static int get_loss_clg(struct netem_sched_data *q, const struct nlattr *attr)
814 {
815 	const struct nlattr *la;
816 	int rem;
817 
818 	nla_for_each_nested(la, attr, rem) {
819 		u16 type = nla_type(la);
820 
821 		switch (type) {
822 		case NETEM_LOSS_GI: {
823 			const struct tc_netem_gimodel *gi = nla_data(la);
824 
825 			if (nla_len(la) < sizeof(struct tc_netem_gimodel)) {
826 				pr_info("netem: incorrect gi model size\n");
827 				return -EINVAL;
828 			}
829 
830 			q->loss_model = CLG_4_STATES;
831 
832 			q->clg.state = TX_IN_GAP_PERIOD;
833 			q->clg.a1 = gi->p13;
834 			q->clg.a2 = gi->p31;
835 			q->clg.a3 = gi->p32;
836 			q->clg.a4 = gi->p14;
837 			q->clg.a5 = gi->p23;
838 			break;
839 		}
840 
841 		case NETEM_LOSS_GE: {
842 			const struct tc_netem_gemodel *ge = nla_data(la);
843 
844 			if (nla_len(la) < sizeof(struct tc_netem_gemodel)) {
845 				pr_info("netem: incorrect ge model size\n");
846 				return -EINVAL;
847 			}
848 
849 			q->loss_model = CLG_GILB_ELL;
850 			q->clg.state = GOOD_STATE;
851 			q->clg.a1 = ge->p;
852 			q->clg.a2 = ge->r;
853 			q->clg.a3 = ge->h;
854 			q->clg.a4 = ge->k1;
855 			break;
856 		}
857 
858 		default:
859 			pr_info("netem: unknown loss type %u\n", type);
860 			return -EINVAL;
861 		}
862 	}
863 
864 	return 0;
865 }
866 
867 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
868 	[TCA_NETEM_CORR]	= { .len = sizeof(struct tc_netem_corr) },
869 	[TCA_NETEM_REORDER]	= { .len = sizeof(struct tc_netem_reorder) },
870 	[TCA_NETEM_CORRUPT]	= { .len = sizeof(struct tc_netem_corrupt) },
871 	[TCA_NETEM_RATE]	= { .len = sizeof(struct tc_netem_rate) },
872 	[TCA_NETEM_LOSS]	= { .type = NLA_NESTED },
873 	[TCA_NETEM_ECN]		= { .type = NLA_U32 },
874 	[TCA_NETEM_RATE64]	= { .type = NLA_U64 },
875 	[TCA_NETEM_LATENCY64]	= { .type = NLA_S64 },
876 	[TCA_NETEM_JITTER64]	= { .type = NLA_S64 },
877 	[TCA_NETEM_SLOT]	= { .len = sizeof(struct tc_netem_slot) },
878 };
879 
880 static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
881 		      const struct nla_policy *policy, int len)
882 {
883 	int nested_len = nla_len(nla) - NLA_ALIGN(len);
884 
885 	if (nested_len < 0) {
886 		pr_info("netem: invalid attributes len %d\n", nested_len);
887 		return -EINVAL;
888 	}
889 
890 	if (nested_len >= nla_attr_size(0))
891 		return nla_parse(tb, maxtype, nla_data(nla) + NLA_ALIGN(len),
892 				 nested_len, policy, NULL);
893 
894 	memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
895 	return 0;
896 }
897 
898 /* Parse netlink message to set options */
899 static int netem_change(struct Qdisc *sch, struct nlattr *opt,
900 			struct netlink_ext_ack *extack)
901 {
902 	struct netem_sched_data *q = qdisc_priv(sch);
903 	struct nlattr *tb[TCA_NETEM_MAX + 1];
904 	struct tc_netem_qopt *qopt;
905 	struct clgstate old_clg;
906 	int old_loss_model = CLG_RANDOM;
907 	int ret;
908 
909 	if (opt == NULL)
910 		return -EINVAL;
911 
912 	qopt = nla_data(opt);
913 	ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
914 	if (ret < 0)
915 		return ret;
916 
917 	/* backup q->clg and q->loss_model */
918 	old_clg = q->clg;
919 	old_loss_model = q->loss_model;
920 
921 	if (tb[TCA_NETEM_LOSS]) {
922 		ret = get_loss_clg(q, tb[TCA_NETEM_LOSS]);
923 		if (ret) {
924 			q->loss_model = old_loss_model;
925 			return ret;
926 		}
927 	} else {
928 		q->loss_model = CLG_RANDOM;
929 	}
930 
931 	if (tb[TCA_NETEM_DELAY_DIST]) {
932 		ret = get_dist_table(sch, &q->delay_dist,
933 				     tb[TCA_NETEM_DELAY_DIST]);
934 		if (ret)
935 			goto get_table_failure;
936 	}
937 
938 	if (tb[TCA_NETEM_SLOT_DIST]) {
939 		ret = get_dist_table(sch, &q->slot_dist,
940 				     tb[TCA_NETEM_SLOT_DIST]);
941 		if (ret)
942 			goto get_table_failure;
943 	}
944 
945 	sch->limit = qopt->limit;
946 
947 	q->latency = PSCHED_TICKS2NS(qopt->latency);
948 	q->jitter = PSCHED_TICKS2NS(qopt->jitter);
949 	q->limit = qopt->limit;
950 	q->gap = qopt->gap;
951 	q->counter = 0;
952 	q->loss = qopt->loss;
953 	q->duplicate = qopt->duplicate;
954 
955 	/* for compatibility with earlier versions.
956 	 * if gap is set, need to assume 100% probability
957 	 */
958 	if (q->gap)
959 		q->reorder = ~0;
960 
961 	if (tb[TCA_NETEM_CORR])
962 		get_correlation(q, tb[TCA_NETEM_CORR]);
963 
964 	if (tb[TCA_NETEM_REORDER])
965 		get_reorder(q, tb[TCA_NETEM_REORDER]);
966 
967 	if (tb[TCA_NETEM_CORRUPT])
968 		get_corrupt(q, tb[TCA_NETEM_CORRUPT]);
969 
970 	if (tb[TCA_NETEM_RATE])
971 		get_rate(q, tb[TCA_NETEM_RATE]);
972 
973 	if (tb[TCA_NETEM_RATE64])
974 		q->rate = max_t(u64, q->rate,
975 				nla_get_u64(tb[TCA_NETEM_RATE64]));
976 
977 	if (tb[TCA_NETEM_LATENCY64])
978 		q->latency = nla_get_s64(tb[TCA_NETEM_LATENCY64]);
979 
980 	if (tb[TCA_NETEM_JITTER64])
981 		q->jitter = nla_get_s64(tb[TCA_NETEM_JITTER64]);
982 
983 	if (tb[TCA_NETEM_ECN])
984 		q->ecn = nla_get_u32(tb[TCA_NETEM_ECN]);
985 
986 	if (tb[TCA_NETEM_SLOT])
987 		get_slot(q, tb[TCA_NETEM_SLOT]);
988 
989 	return ret;
990 
991 get_table_failure:
992 	/* recover clg and loss_model, in case of
993 	 * q->clg and q->loss_model were modified
994 	 * in get_loss_clg()
995 	 */
996 	q->clg = old_clg;
997 	q->loss_model = old_loss_model;
998 	return ret;
999 }
1000 
1001 static int netem_init(struct Qdisc *sch, struct nlattr *opt,
1002 		      struct netlink_ext_ack *extack)
1003 {
1004 	struct netem_sched_data *q = qdisc_priv(sch);
1005 	int ret;
1006 
1007 	qdisc_watchdog_init(&q->watchdog, sch);
1008 
1009 	if (!opt)
1010 		return -EINVAL;
1011 
1012 	q->loss_model = CLG_RANDOM;
1013 	ret = netem_change(sch, opt, extack);
1014 	if (ret)
1015 		pr_info("netem: change failed\n");
1016 	return ret;
1017 }
1018 
1019 static void netem_destroy(struct Qdisc *sch)
1020 {
1021 	struct netem_sched_data *q = qdisc_priv(sch);
1022 
1023 	qdisc_watchdog_cancel(&q->watchdog);
1024 	if (q->qdisc)
1025 		qdisc_put(q->qdisc);
1026 	dist_free(q->delay_dist);
1027 	dist_free(q->slot_dist);
1028 }
1029 
1030 static int dump_loss_model(const struct netem_sched_data *q,
1031 			   struct sk_buff *skb)
1032 {
1033 	struct nlattr *nest;
1034 
1035 	nest = nla_nest_start(skb, TCA_NETEM_LOSS);
1036 	if (nest == NULL)
1037 		goto nla_put_failure;
1038 
1039 	switch (q->loss_model) {
1040 	case CLG_RANDOM:
1041 		/* legacy loss model */
1042 		nla_nest_cancel(skb, nest);
1043 		return 0;	/* no data */
1044 
1045 	case CLG_4_STATES: {
1046 		struct tc_netem_gimodel gi = {
1047 			.p13 = q->clg.a1,
1048 			.p31 = q->clg.a2,
1049 			.p32 = q->clg.a3,
1050 			.p14 = q->clg.a4,
1051 			.p23 = q->clg.a5,
1052 		};
1053 
1054 		if (nla_put(skb, NETEM_LOSS_GI, sizeof(gi), &gi))
1055 			goto nla_put_failure;
1056 		break;
1057 	}
1058 	case CLG_GILB_ELL: {
1059 		struct tc_netem_gemodel ge = {
1060 			.p = q->clg.a1,
1061 			.r = q->clg.a2,
1062 			.h = q->clg.a3,
1063 			.k1 = q->clg.a4,
1064 		};
1065 
1066 		if (nla_put(skb, NETEM_LOSS_GE, sizeof(ge), &ge))
1067 			goto nla_put_failure;
1068 		break;
1069 	}
1070 	}
1071 
1072 	nla_nest_end(skb, nest);
1073 	return 0;
1074 
1075 nla_put_failure:
1076 	nla_nest_cancel(skb, nest);
1077 	return -1;
1078 }
1079 
1080 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
1081 {
1082 	const struct netem_sched_data *q = qdisc_priv(sch);
1083 	struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb);
1084 	struct tc_netem_qopt qopt;
1085 	struct tc_netem_corr cor;
1086 	struct tc_netem_reorder reorder;
1087 	struct tc_netem_corrupt corrupt;
1088 	struct tc_netem_rate rate;
1089 	struct tc_netem_slot slot;
1090 
1091 	qopt.latency = min_t(psched_tdiff_t, PSCHED_NS2TICKS(q->latency),
1092 			     UINT_MAX);
1093 	qopt.jitter = min_t(psched_tdiff_t, PSCHED_NS2TICKS(q->jitter),
1094 			    UINT_MAX);
1095 	qopt.limit = q->limit;
1096 	qopt.loss = q->loss;
1097 	qopt.gap = q->gap;
1098 	qopt.duplicate = q->duplicate;
1099 	if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
1100 		goto nla_put_failure;
1101 
1102 	if (nla_put(skb, TCA_NETEM_LATENCY64, sizeof(q->latency), &q->latency))
1103 		goto nla_put_failure;
1104 
1105 	if (nla_put(skb, TCA_NETEM_JITTER64, sizeof(q->jitter), &q->jitter))
1106 		goto nla_put_failure;
1107 
1108 	cor.delay_corr = q->delay_cor.rho;
1109 	cor.loss_corr = q->loss_cor.rho;
1110 	cor.dup_corr = q->dup_cor.rho;
1111 	if (nla_put(skb, TCA_NETEM_CORR, sizeof(cor), &cor))
1112 		goto nla_put_failure;
1113 
1114 	reorder.probability = q->reorder;
1115 	reorder.correlation = q->reorder_cor.rho;
1116 	if (nla_put(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder))
1117 		goto nla_put_failure;
1118 
1119 	corrupt.probability = q->corrupt;
1120 	corrupt.correlation = q->corrupt_cor.rho;
1121 	if (nla_put(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt))
1122 		goto nla_put_failure;
1123 
1124 	if (q->rate >= (1ULL << 32)) {
1125 		if (nla_put_u64_64bit(skb, TCA_NETEM_RATE64, q->rate,
1126 				      TCA_NETEM_PAD))
1127 			goto nla_put_failure;
1128 		rate.rate = ~0U;
1129 	} else {
1130 		rate.rate = q->rate;
1131 	}
1132 	rate.packet_overhead = q->packet_overhead;
1133 	rate.cell_size = q->cell_size;
1134 	rate.cell_overhead = q->cell_overhead;
1135 	if (nla_put(skb, TCA_NETEM_RATE, sizeof(rate), &rate))
1136 		goto nla_put_failure;
1137 
1138 	if (q->ecn && nla_put_u32(skb, TCA_NETEM_ECN, q->ecn))
1139 		goto nla_put_failure;
1140 
1141 	if (dump_loss_model(q, skb) != 0)
1142 		goto nla_put_failure;
1143 
1144 	if (q->slot_config.min_delay | q->slot_config.max_delay |
1145 	    q->slot_config.dist_jitter) {
1146 		slot = q->slot_config;
1147 		if (slot.max_packets == INT_MAX)
1148 			slot.max_packets = 0;
1149 		if (slot.max_bytes == INT_MAX)
1150 			slot.max_bytes = 0;
1151 		if (nla_put(skb, TCA_NETEM_SLOT, sizeof(slot), &slot))
1152 			goto nla_put_failure;
1153 	}
1154 
1155 	return nla_nest_end(skb, nla);
1156 
1157 nla_put_failure:
1158 	nlmsg_trim(skb, nla);
1159 	return -1;
1160 }
1161 
1162 static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
1163 			  struct sk_buff *skb, struct tcmsg *tcm)
1164 {
1165 	struct netem_sched_data *q = qdisc_priv(sch);
1166 
1167 	if (cl != 1 || !q->qdisc) 	/* only one class */
1168 		return -ENOENT;
1169 
1170 	tcm->tcm_handle |= TC_H_MIN(1);
1171 	tcm->tcm_info = q->qdisc->handle;
1172 
1173 	return 0;
1174 }
1175 
1176 static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1177 		     struct Qdisc **old, struct netlink_ext_ack *extack)
1178 {
1179 	struct netem_sched_data *q = qdisc_priv(sch);
1180 
1181 	*old = qdisc_replace(sch, new, &q->qdisc);
1182 	return 0;
1183 }
1184 
1185 static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
1186 {
1187 	struct netem_sched_data *q = qdisc_priv(sch);
1188 	return q->qdisc;
1189 }
1190 
1191 static unsigned long netem_find(struct Qdisc *sch, u32 classid)
1192 {
1193 	return 1;
1194 }
1195 
1196 static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
1197 {
1198 	if (!walker->stop) {
1199 		if (walker->count >= walker->skip)
1200 			if (walker->fn(sch, 1, walker) < 0) {
1201 				walker->stop = 1;
1202 				return;
1203 			}
1204 		walker->count++;
1205 	}
1206 }
1207 
1208 static const struct Qdisc_class_ops netem_class_ops = {
1209 	.graft		=	netem_graft,
1210 	.leaf		=	netem_leaf,
1211 	.find		=	netem_find,
1212 	.walk		=	netem_walk,
1213 	.dump		=	netem_dump_class,
1214 };
1215 
1216 static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
1217 	.id		=	"netem",
1218 	.cl_ops		=	&netem_class_ops,
1219 	.priv_size	=	sizeof(struct netem_sched_data),
1220 	.enqueue	=	netem_enqueue,
1221 	.dequeue	=	netem_dequeue,
1222 	.peek		=	qdisc_peek_dequeued,
1223 	.init		=	netem_init,
1224 	.reset		=	netem_reset,
1225 	.destroy	=	netem_destroy,
1226 	.change		=	netem_change,
1227 	.dump		=	netem_dump,
1228 	.owner		=	THIS_MODULE,
1229 };
1230 
1231 
1232 static int __init netem_module_init(void)
1233 {
1234 	pr_info("netem: version " VERSION "\n");
1235 	return register_qdisc(&netem_qdisc_ops);
1236 }
1237 static void __exit netem_module_exit(void)
1238 {
1239 	unregister_qdisc(&netem_qdisc_ops);
1240 }
1241 module_init(netem_module_init)
1242 module_exit(netem_module_exit)
1243 MODULE_LICENSE("GPL");
1244