xref: /openbmc/linux/net/sched/sch_netem.c (revision 78bb17f7)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * net/sched/sch_netem.c	Network emulator
4  *
5  *  		Many of the algorithms and ideas for this came from
6  *		NIST Net which is not copyrighted.
7  *
8  * Authors:	Stephen Hemminger <shemminger@osdl.org>
9  *		Catalin(ux aka Dino) BOIE <catab at umbrella dot ro>
10  */
11 
12 #include <linux/mm.h>
13 #include <linux/module.h>
14 #include <linux/slab.h>
15 #include <linux/types.h>
16 #include <linux/kernel.h>
17 #include <linux/errno.h>
18 #include <linux/skbuff.h>
19 #include <linux/vmalloc.h>
20 #include <linux/rtnetlink.h>
21 #include <linux/reciprocal_div.h>
22 #include <linux/rbtree.h>
23 
24 #include <net/netlink.h>
25 #include <net/pkt_sched.h>
26 #include <net/inet_ecn.h>
27 
28 #define VERSION "1.3"
29 
30 /*	Network Emulation Queuing algorithm.
31 	====================================
32 
33 	Sources: [1] Mark Carson, Darrin Santay, "NIST Net - A Linux-based
34 		 Network Emulation Tool
35 		 [2] Luigi Rizzo, DummyNet for FreeBSD
36 
37 	 ----------------------------------------------------------------
38 
39 	 This started out as a simple way to delay outgoing packets to
40 	 test TCP but has grown to include most of the functionality
41 	 of a full blown network emulator like NISTnet. It can delay
42 	 packets and add random jitter (and correlation). The random
43 	 distribution can be loaded from a table as well to provide
44 	 normal, Pareto, or experimental curves. Packet loss,
45 	 duplication, and reordering can also be emulated.
46 
47 	 This qdisc does not do classification that can be handled in
48 	 layering other disciplines.  It does not need to do bandwidth
49 	 control either since that can be handled by using token
50 	 bucket or other rate control.
51 
52      Correlated Loss Generator models
53 
54 	Added generation of correlated loss according to the
55 	"Gilbert-Elliot" model, a 4-state markov model.
56 
57 	References:
58 	[1] NetemCLG Home http://netgroup.uniroma2.it/NetemCLG
59 	[2] S. Salsano, F. Ludovici, A. Ordine, "Definition of a general
60 	and intuitive loss model for packet networks and its implementation
61 	in the Netem module in the Linux kernel", available in [1]
62 
63 	Authors: Stefano Salsano <stefano.salsano at uniroma2.it
64 		 Fabio Ludovici <fabio.ludovici at yahoo.it>
65 */
66 
67 struct disttable {
68 	u32  size;
69 	s16 table[];
70 };
71 
72 struct netem_sched_data {
73 	/* internal t(ime)fifo qdisc uses t_root and sch->limit */
74 	struct rb_root t_root;
75 
76 	/* a linear queue; reduces rbtree rebalancing when jitter is low */
77 	struct sk_buff	*t_head;
78 	struct sk_buff	*t_tail;
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 	rtnl_kfree_skbs(q->t_head, q->t_tail);
374 	q->t_head = NULL;
375 	q->t_tail = NULL;
376 }
377 
378 static void tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
379 {
380 	struct netem_sched_data *q = qdisc_priv(sch);
381 	u64 tnext = netem_skb_cb(nskb)->time_to_send;
382 
383 	if (!q->t_tail || tnext >= netem_skb_cb(q->t_tail)->time_to_send) {
384 		if (q->t_tail)
385 			q->t_tail->next = nskb;
386 		else
387 			q->t_head = nskb;
388 		q->t_tail = nskb;
389 	} else {
390 		struct rb_node **p = &q->t_root.rb_node, *parent = NULL;
391 
392 		while (*p) {
393 			struct sk_buff *skb;
394 
395 			parent = *p;
396 			skb = rb_to_skb(parent);
397 			if (tnext >= netem_skb_cb(skb)->time_to_send)
398 				p = &parent->rb_right;
399 			else
400 				p = &parent->rb_left;
401 		}
402 		rb_link_node(&nskb->rbnode, parent, p);
403 		rb_insert_color(&nskb->rbnode, &q->t_root);
404 	}
405 	sch->q.qlen++;
406 }
407 
408 /* netem can't properly corrupt a megapacket (like we get from GSO), so instead
409  * when we statistically choose to corrupt one, we instead segment it, returning
410  * the first packet to be corrupted, and re-enqueue the remaining frames
411  */
412 static struct sk_buff *netem_segment(struct sk_buff *skb, struct Qdisc *sch,
413 				     struct sk_buff **to_free)
414 {
415 	struct sk_buff *segs;
416 	netdev_features_t features = netif_skb_features(skb);
417 
418 	segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
419 
420 	if (IS_ERR_OR_NULL(segs)) {
421 		qdisc_drop(skb, sch, to_free);
422 		return NULL;
423 	}
424 	consume_skb(skb);
425 	return segs;
426 }
427 
428 /*
429  * Insert one skb into qdisc.
430  * Note: parent depends on return value to account for queue length.
431  * 	NET_XMIT_DROP: queue length didn't change.
432  *      NET_XMIT_SUCCESS: one skb was queued.
433  */
434 static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch,
435 			 struct sk_buff **to_free)
436 {
437 	struct netem_sched_data *q = qdisc_priv(sch);
438 	/* We don't fill cb now as skb_unshare() may invalidate it */
439 	struct netem_skb_cb *cb;
440 	struct sk_buff *skb2;
441 	struct sk_buff *segs = NULL;
442 	unsigned int prev_len = qdisc_pkt_len(skb);
443 	int count = 1;
444 	int rc = NET_XMIT_SUCCESS;
445 	int rc_drop = NET_XMIT_DROP;
446 
447 	/* Do not fool qdisc_drop_all() */
448 	skb->prev = NULL;
449 
450 	/* Random duplication */
451 	if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor))
452 		++count;
453 
454 	/* Drop packet? */
455 	if (loss_event(q)) {
456 		if (q->ecn && INET_ECN_set_ce(skb))
457 			qdisc_qstats_drop(sch); /* mark packet */
458 		else
459 			--count;
460 	}
461 	if (count == 0) {
462 		qdisc_qstats_drop(sch);
463 		__qdisc_drop(skb, to_free);
464 		return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
465 	}
466 
467 	/* If a delay is expected, orphan the skb. (orphaning usually takes
468 	 * place at TX completion time, so _before_ the link transit delay)
469 	 */
470 	if (q->latency || q->jitter || q->rate)
471 		skb_orphan_partial(skb);
472 
473 	/*
474 	 * If we need to duplicate packet, then re-insert at top of the
475 	 * qdisc tree, since parent queuer expects that only one
476 	 * skb will be queued.
477 	 */
478 	if (count > 1 && (skb2 = skb_clone(skb, GFP_ATOMIC)) != NULL) {
479 		struct Qdisc *rootq = qdisc_root_bh(sch);
480 		u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
481 
482 		q->duplicate = 0;
483 		rootq->enqueue(skb2, rootq, to_free);
484 		q->duplicate = dupsave;
485 		rc_drop = NET_XMIT_SUCCESS;
486 	}
487 
488 	/*
489 	 * Randomized packet corruption.
490 	 * Make copy if needed since we are modifying
491 	 * If packet is going to be hardware checksummed, then
492 	 * do it now in software before we mangle it.
493 	 */
494 	if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor)) {
495 		if (skb_is_gso(skb)) {
496 			skb = netem_segment(skb, sch, to_free);
497 			if (!skb)
498 				return rc_drop;
499 			segs = skb->next;
500 			skb_mark_not_on_list(skb);
501 			qdisc_skb_cb(skb)->pkt_len = skb->len;
502 		}
503 
504 		skb = skb_unshare(skb, GFP_ATOMIC);
505 		if (unlikely(!skb)) {
506 			qdisc_qstats_drop(sch);
507 			goto finish_segs;
508 		}
509 		if (skb->ip_summed == CHECKSUM_PARTIAL &&
510 		    skb_checksum_help(skb)) {
511 			qdisc_drop(skb, sch, to_free);
512 			skb = NULL;
513 			goto finish_segs;
514 		}
515 
516 		skb->data[prandom_u32() % skb_headlen(skb)] ^=
517 			1<<(prandom_u32() % 8);
518 	}
519 
520 	if (unlikely(sch->q.qlen >= sch->limit)) {
521 		/* re-link segs, so that qdisc_drop_all() frees them all */
522 		skb->next = segs;
523 		qdisc_drop_all(skb, sch, to_free);
524 		return rc_drop;
525 	}
526 
527 	qdisc_qstats_backlog_inc(sch, skb);
528 
529 	cb = netem_skb_cb(skb);
530 	if (q->gap == 0 ||		/* not doing reordering */
531 	    q->counter < q->gap - 1 ||	/* inside last reordering gap */
532 	    q->reorder < get_crandom(&q->reorder_cor)) {
533 		u64 now;
534 		s64 delay;
535 
536 		delay = tabledist(q->latency, q->jitter,
537 				  &q->delay_cor, q->delay_dist);
538 
539 		now = ktime_get_ns();
540 
541 		if (q->rate) {
542 			struct netem_skb_cb *last = NULL;
543 
544 			if (sch->q.tail)
545 				last = netem_skb_cb(sch->q.tail);
546 			if (q->t_root.rb_node) {
547 				struct sk_buff *t_skb;
548 				struct netem_skb_cb *t_last;
549 
550 				t_skb = skb_rb_last(&q->t_root);
551 				t_last = netem_skb_cb(t_skb);
552 				if (!last ||
553 				    t_last->time_to_send > last->time_to_send)
554 					last = t_last;
555 			}
556 			if (q->t_tail) {
557 				struct netem_skb_cb *t_last =
558 					netem_skb_cb(q->t_tail);
559 
560 				if (!last ||
561 				    t_last->time_to_send > last->time_to_send)
562 					last = t_last;
563 			}
564 
565 			if (last) {
566 				/*
567 				 * Last packet in queue is reference point (now),
568 				 * calculate this time bonus and subtract
569 				 * from delay.
570 				 */
571 				delay -= last->time_to_send - now;
572 				delay = max_t(s64, 0, delay);
573 				now = last->time_to_send;
574 			}
575 
576 			delay += packet_time_ns(qdisc_pkt_len(skb), q);
577 		}
578 
579 		cb->time_to_send = now + delay;
580 		++q->counter;
581 		tfifo_enqueue(skb, sch);
582 	} else {
583 		/*
584 		 * Do re-ordering by putting one out of N packets at the front
585 		 * of the queue.
586 		 */
587 		cb->time_to_send = ktime_get_ns();
588 		q->counter = 0;
589 
590 		__qdisc_enqueue_head(skb, &sch->q);
591 		sch->qstats.requeues++;
592 	}
593 
594 finish_segs:
595 	if (segs) {
596 		unsigned int len, last_len;
597 		int nb;
598 
599 		len = skb ? skb->len : 0;
600 		nb = skb ? 1 : 0;
601 
602 		while (segs) {
603 			skb2 = segs->next;
604 			skb_mark_not_on_list(segs);
605 			qdisc_skb_cb(segs)->pkt_len = segs->len;
606 			last_len = segs->len;
607 			rc = qdisc_enqueue(segs, sch, to_free);
608 			if (rc != NET_XMIT_SUCCESS) {
609 				if (net_xmit_drop_count(rc))
610 					qdisc_qstats_drop(sch);
611 			} else {
612 				nb++;
613 				len += last_len;
614 			}
615 			segs = skb2;
616 		}
617 		/* Parent qdiscs accounted for 1 skb of size @prev_len */
618 		qdisc_tree_reduce_backlog(sch, -(nb - 1), -(len - prev_len));
619 	} else if (!skb) {
620 		return NET_XMIT_DROP;
621 	}
622 	return NET_XMIT_SUCCESS;
623 }
624 
625 /* Delay the next round with a new future slot with a
626  * correct number of bytes and packets.
627  */
628 
629 static void get_slot_next(struct netem_sched_data *q, u64 now)
630 {
631 	s64 next_delay;
632 
633 	if (!q->slot_dist)
634 		next_delay = q->slot_config.min_delay +
635 				(prandom_u32() *
636 				 (q->slot_config.max_delay -
637 				  q->slot_config.min_delay) >> 32);
638 	else
639 		next_delay = tabledist(q->slot_config.dist_delay,
640 				       (s32)(q->slot_config.dist_jitter),
641 				       NULL, q->slot_dist);
642 
643 	q->slot.slot_next = now + next_delay;
644 	q->slot.packets_left = q->slot_config.max_packets;
645 	q->slot.bytes_left = q->slot_config.max_bytes;
646 }
647 
648 static struct sk_buff *netem_peek(struct netem_sched_data *q)
649 {
650 	struct sk_buff *skb = skb_rb_first(&q->t_root);
651 	u64 t1, t2;
652 
653 	if (!skb)
654 		return q->t_head;
655 	if (!q->t_head)
656 		return skb;
657 
658 	t1 = netem_skb_cb(skb)->time_to_send;
659 	t2 = netem_skb_cb(q->t_head)->time_to_send;
660 	if (t1 < t2)
661 		return skb;
662 	return q->t_head;
663 }
664 
665 static void netem_erase_head(struct netem_sched_data *q, struct sk_buff *skb)
666 {
667 	if (skb == q->t_head) {
668 		q->t_head = skb->next;
669 		if (!q->t_head)
670 			q->t_tail = NULL;
671 	} else {
672 		rb_erase(&skb->rbnode, &q->t_root);
673 	}
674 }
675 
676 static struct sk_buff *netem_dequeue(struct Qdisc *sch)
677 {
678 	struct netem_sched_data *q = qdisc_priv(sch);
679 	struct sk_buff *skb;
680 
681 tfifo_dequeue:
682 	skb = __qdisc_dequeue_head(&sch->q);
683 	if (skb) {
684 		qdisc_qstats_backlog_dec(sch, skb);
685 deliver:
686 		qdisc_bstats_update(sch, skb);
687 		return skb;
688 	}
689 	skb = netem_peek(q);
690 	if (skb) {
691 		u64 time_to_send;
692 		u64 now = ktime_get_ns();
693 
694 		/* if more time remaining? */
695 		time_to_send = netem_skb_cb(skb)->time_to_send;
696 		if (q->slot.slot_next && q->slot.slot_next < time_to_send)
697 			get_slot_next(q, now);
698 
699 		if (time_to_send <= now && q->slot.slot_next <= now) {
700 			netem_erase_head(q, skb);
701 			sch->q.qlen--;
702 			qdisc_qstats_backlog_dec(sch, skb);
703 			skb->next = NULL;
704 			skb->prev = NULL;
705 			/* skb->dev shares skb->rbnode area,
706 			 * we need to restore its value.
707 			 */
708 			skb->dev = qdisc_dev(sch);
709 
710 			if (q->slot.slot_next) {
711 				q->slot.packets_left--;
712 				q->slot.bytes_left -= qdisc_pkt_len(skb);
713 				if (q->slot.packets_left <= 0 ||
714 				    q->slot.bytes_left <= 0)
715 					get_slot_next(q, now);
716 			}
717 
718 			if (q->qdisc) {
719 				unsigned int pkt_len = qdisc_pkt_len(skb);
720 				struct sk_buff *to_free = NULL;
721 				int err;
722 
723 				err = qdisc_enqueue(skb, q->qdisc, &to_free);
724 				kfree_skb_list(to_free);
725 				if (err != NET_XMIT_SUCCESS &&
726 				    net_xmit_drop_count(err)) {
727 					qdisc_qstats_drop(sch);
728 					qdisc_tree_reduce_backlog(sch, 1,
729 								  pkt_len);
730 				}
731 				goto tfifo_dequeue;
732 			}
733 			goto deliver;
734 		}
735 
736 		if (q->qdisc) {
737 			skb = q->qdisc->ops->dequeue(q->qdisc);
738 			if (skb)
739 				goto deliver;
740 		}
741 
742 		qdisc_watchdog_schedule_ns(&q->watchdog,
743 					   max(time_to_send,
744 					       q->slot.slot_next));
745 	}
746 
747 	if (q->qdisc) {
748 		skb = q->qdisc->ops->dequeue(q->qdisc);
749 		if (skb)
750 			goto deliver;
751 	}
752 	return NULL;
753 }
754 
755 static void netem_reset(struct Qdisc *sch)
756 {
757 	struct netem_sched_data *q = qdisc_priv(sch);
758 
759 	qdisc_reset_queue(sch);
760 	tfifo_reset(sch);
761 	if (q->qdisc)
762 		qdisc_reset(q->qdisc);
763 	qdisc_watchdog_cancel(&q->watchdog);
764 }
765 
766 static void dist_free(struct disttable *d)
767 {
768 	kvfree(d);
769 }
770 
771 /*
772  * Distribution data is a variable size payload containing
773  * signed 16 bit values.
774  */
775 
776 static int get_dist_table(struct Qdisc *sch, struct disttable **tbl,
777 			  const struct nlattr *attr)
778 {
779 	size_t n = nla_len(attr)/sizeof(__s16);
780 	const __s16 *data = nla_data(attr);
781 	spinlock_t *root_lock;
782 	struct disttable *d;
783 	int i;
784 
785 	if (!n || n > NETEM_DIST_MAX)
786 		return -EINVAL;
787 
788 	d = kvmalloc(sizeof(struct disttable) + n * sizeof(s16), GFP_KERNEL);
789 	if (!d)
790 		return -ENOMEM;
791 
792 	d->size = n;
793 	for (i = 0; i < n; i++)
794 		d->table[i] = data[i];
795 
796 	root_lock = qdisc_root_sleeping_lock(sch);
797 
798 	spin_lock_bh(root_lock);
799 	swap(*tbl, d);
800 	spin_unlock_bh(root_lock);
801 
802 	dist_free(d);
803 	return 0;
804 }
805 
806 static void get_slot(struct netem_sched_data *q, const struct nlattr *attr)
807 {
808 	const struct tc_netem_slot *c = nla_data(attr);
809 
810 	q->slot_config = *c;
811 	if (q->slot_config.max_packets == 0)
812 		q->slot_config.max_packets = INT_MAX;
813 	if (q->slot_config.max_bytes == 0)
814 		q->slot_config.max_bytes = INT_MAX;
815 	q->slot.packets_left = q->slot_config.max_packets;
816 	q->slot.bytes_left = q->slot_config.max_bytes;
817 	if (q->slot_config.min_delay | q->slot_config.max_delay |
818 	    q->slot_config.dist_jitter)
819 		q->slot.slot_next = ktime_get_ns();
820 	else
821 		q->slot.slot_next = 0;
822 }
823 
824 static void get_correlation(struct netem_sched_data *q, const struct nlattr *attr)
825 {
826 	const struct tc_netem_corr *c = nla_data(attr);
827 
828 	init_crandom(&q->delay_cor, c->delay_corr);
829 	init_crandom(&q->loss_cor, c->loss_corr);
830 	init_crandom(&q->dup_cor, c->dup_corr);
831 }
832 
833 static void get_reorder(struct netem_sched_data *q, const struct nlattr *attr)
834 {
835 	const struct tc_netem_reorder *r = nla_data(attr);
836 
837 	q->reorder = r->probability;
838 	init_crandom(&q->reorder_cor, r->correlation);
839 }
840 
841 static void get_corrupt(struct netem_sched_data *q, const struct nlattr *attr)
842 {
843 	const struct tc_netem_corrupt *r = nla_data(attr);
844 
845 	q->corrupt = r->probability;
846 	init_crandom(&q->corrupt_cor, r->correlation);
847 }
848 
849 static void get_rate(struct netem_sched_data *q, const struct nlattr *attr)
850 {
851 	const struct tc_netem_rate *r = nla_data(attr);
852 
853 	q->rate = r->rate;
854 	q->packet_overhead = r->packet_overhead;
855 	q->cell_size = r->cell_size;
856 	q->cell_overhead = r->cell_overhead;
857 	if (q->cell_size)
858 		q->cell_size_reciprocal = reciprocal_value(q->cell_size);
859 	else
860 		q->cell_size_reciprocal = (struct reciprocal_value) { 0 };
861 }
862 
863 static int get_loss_clg(struct netem_sched_data *q, const struct nlattr *attr)
864 {
865 	const struct nlattr *la;
866 	int rem;
867 
868 	nla_for_each_nested(la, attr, rem) {
869 		u16 type = nla_type(la);
870 
871 		switch (type) {
872 		case NETEM_LOSS_GI: {
873 			const struct tc_netem_gimodel *gi = nla_data(la);
874 
875 			if (nla_len(la) < sizeof(struct tc_netem_gimodel)) {
876 				pr_info("netem: incorrect gi model size\n");
877 				return -EINVAL;
878 			}
879 
880 			q->loss_model = CLG_4_STATES;
881 
882 			q->clg.state = TX_IN_GAP_PERIOD;
883 			q->clg.a1 = gi->p13;
884 			q->clg.a2 = gi->p31;
885 			q->clg.a3 = gi->p32;
886 			q->clg.a4 = gi->p14;
887 			q->clg.a5 = gi->p23;
888 			break;
889 		}
890 
891 		case NETEM_LOSS_GE: {
892 			const struct tc_netem_gemodel *ge = nla_data(la);
893 
894 			if (nla_len(la) < sizeof(struct tc_netem_gemodel)) {
895 				pr_info("netem: incorrect ge model size\n");
896 				return -EINVAL;
897 			}
898 
899 			q->loss_model = CLG_GILB_ELL;
900 			q->clg.state = GOOD_STATE;
901 			q->clg.a1 = ge->p;
902 			q->clg.a2 = ge->r;
903 			q->clg.a3 = ge->h;
904 			q->clg.a4 = ge->k1;
905 			break;
906 		}
907 
908 		default:
909 			pr_info("netem: unknown loss type %u\n", type);
910 			return -EINVAL;
911 		}
912 	}
913 
914 	return 0;
915 }
916 
917 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
918 	[TCA_NETEM_CORR]	= { .len = sizeof(struct tc_netem_corr) },
919 	[TCA_NETEM_REORDER]	= { .len = sizeof(struct tc_netem_reorder) },
920 	[TCA_NETEM_CORRUPT]	= { .len = sizeof(struct tc_netem_corrupt) },
921 	[TCA_NETEM_RATE]	= { .len = sizeof(struct tc_netem_rate) },
922 	[TCA_NETEM_LOSS]	= { .type = NLA_NESTED },
923 	[TCA_NETEM_ECN]		= { .type = NLA_U32 },
924 	[TCA_NETEM_RATE64]	= { .type = NLA_U64 },
925 	[TCA_NETEM_LATENCY64]	= { .type = NLA_S64 },
926 	[TCA_NETEM_JITTER64]	= { .type = NLA_S64 },
927 	[TCA_NETEM_SLOT]	= { .len = sizeof(struct tc_netem_slot) },
928 };
929 
930 static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
931 		      const struct nla_policy *policy, int len)
932 {
933 	int nested_len = nla_len(nla) - NLA_ALIGN(len);
934 
935 	if (nested_len < 0) {
936 		pr_info("netem: invalid attributes len %d\n", nested_len);
937 		return -EINVAL;
938 	}
939 
940 	if (nested_len >= nla_attr_size(0))
941 		return nla_parse_deprecated(tb, maxtype,
942 					    nla_data(nla) + NLA_ALIGN(len),
943 					    nested_len, policy, NULL);
944 
945 	memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
946 	return 0;
947 }
948 
949 /* Parse netlink message to set options */
950 static int netem_change(struct Qdisc *sch, struct nlattr *opt,
951 			struct netlink_ext_ack *extack)
952 {
953 	struct netem_sched_data *q = qdisc_priv(sch);
954 	struct nlattr *tb[TCA_NETEM_MAX + 1];
955 	struct tc_netem_qopt *qopt;
956 	struct clgstate old_clg;
957 	int old_loss_model = CLG_RANDOM;
958 	int ret;
959 
960 	if (opt == NULL)
961 		return -EINVAL;
962 
963 	qopt = nla_data(opt);
964 	ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
965 	if (ret < 0)
966 		return ret;
967 
968 	/* backup q->clg and q->loss_model */
969 	old_clg = q->clg;
970 	old_loss_model = q->loss_model;
971 
972 	if (tb[TCA_NETEM_LOSS]) {
973 		ret = get_loss_clg(q, tb[TCA_NETEM_LOSS]);
974 		if (ret) {
975 			q->loss_model = old_loss_model;
976 			return ret;
977 		}
978 	} else {
979 		q->loss_model = CLG_RANDOM;
980 	}
981 
982 	if (tb[TCA_NETEM_DELAY_DIST]) {
983 		ret = get_dist_table(sch, &q->delay_dist,
984 				     tb[TCA_NETEM_DELAY_DIST]);
985 		if (ret)
986 			goto get_table_failure;
987 	}
988 
989 	if (tb[TCA_NETEM_SLOT_DIST]) {
990 		ret = get_dist_table(sch, &q->slot_dist,
991 				     tb[TCA_NETEM_SLOT_DIST]);
992 		if (ret)
993 			goto get_table_failure;
994 	}
995 
996 	sch->limit = qopt->limit;
997 
998 	q->latency = PSCHED_TICKS2NS(qopt->latency);
999 	q->jitter = PSCHED_TICKS2NS(qopt->jitter);
1000 	q->limit = qopt->limit;
1001 	q->gap = qopt->gap;
1002 	q->counter = 0;
1003 	q->loss = qopt->loss;
1004 	q->duplicate = qopt->duplicate;
1005 
1006 	/* for compatibility with earlier versions.
1007 	 * if gap is set, need to assume 100% probability
1008 	 */
1009 	if (q->gap)
1010 		q->reorder = ~0;
1011 
1012 	if (tb[TCA_NETEM_CORR])
1013 		get_correlation(q, tb[TCA_NETEM_CORR]);
1014 
1015 	if (tb[TCA_NETEM_REORDER])
1016 		get_reorder(q, tb[TCA_NETEM_REORDER]);
1017 
1018 	if (tb[TCA_NETEM_CORRUPT])
1019 		get_corrupt(q, tb[TCA_NETEM_CORRUPT]);
1020 
1021 	if (tb[TCA_NETEM_RATE])
1022 		get_rate(q, tb[TCA_NETEM_RATE]);
1023 
1024 	if (tb[TCA_NETEM_RATE64])
1025 		q->rate = max_t(u64, q->rate,
1026 				nla_get_u64(tb[TCA_NETEM_RATE64]));
1027 
1028 	if (tb[TCA_NETEM_LATENCY64])
1029 		q->latency = nla_get_s64(tb[TCA_NETEM_LATENCY64]);
1030 
1031 	if (tb[TCA_NETEM_JITTER64])
1032 		q->jitter = nla_get_s64(tb[TCA_NETEM_JITTER64]);
1033 
1034 	if (tb[TCA_NETEM_ECN])
1035 		q->ecn = nla_get_u32(tb[TCA_NETEM_ECN]);
1036 
1037 	if (tb[TCA_NETEM_SLOT])
1038 		get_slot(q, tb[TCA_NETEM_SLOT]);
1039 
1040 	return ret;
1041 
1042 get_table_failure:
1043 	/* recover clg and loss_model, in case of
1044 	 * q->clg and q->loss_model were modified
1045 	 * in get_loss_clg()
1046 	 */
1047 	q->clg = old_clg;
1048 	q->loss_model = old_loss_model;
1049 	return ret;
1050 }
1051 
1052 static int netem_init(struct Qdisc *sch, struct nlattr *opt,
1053 		      struct netlink_ext_ack *extack)
1054 {
1055 	struct netem_sched_data *q = qdisc_priv(sch);
1056 	int ret;
1057 
1058 	qdisc_watchdog_init(&q->watchdog, sch);
1059 
1060 	if (!opt)
1061 		return -EINVAL;
1062 
1063 	q->loss_model = CLG_RANDOM;
1064 	ret = netem_change(sch, opt, extack);
1065 	if (ret)
1066 		pr_info("netem: change failed\n");
1067 	return ret;
1068 }
1069 
1070 static void netem_destroy(struct Qdisc *sch)
1071 {
1072 	struct netem_sched_data *q = qdisc_priv(sch);
1073 
1074 	qdisc_watchdog_cancel(&q->watchdog);
1075 	if (q->qdisc)
1076 		qdisc_put(q->qdisc);
1077 	dist_free(q->delay_dist);
1078 	dist_free(q->slot_dist);
1079 }
1080 
1081 static int dump_loss_model(const struct netem_sched_data *q,
1082 			   struct sk_buff *skb)
1083 {
1084 	struct nlattr *nest;
1085 
1086 	nest = nla_nest_start_noflag(skb, TCA_NETEM_LOSS);
1087 	if (nest == NULL)
1088 		goto nla_put_failure;
1089 
1090 	switch (q->loss_model) {
1091 	case CLG_RANDOM:
1092 		/* legacy loss model */
1093 		nla_nest_cancel(skb, nest);
1094 		return 0;	/* no data */
1095 
1096 	case CLG_4_STATES: {
1097 		struct tc_netem_gimodel gi = {
1098 			.p13 = q->clg.a1,
1099 			.p31 = q->clg.a2,
1100 			.p32 = q->clg.a3,
1101 			.p14 = q->clg.a4,
1102 			.p23 = q->clg.a5,
1103 		};
1104 
1105 		if (nla_put(skb, NETEM_LOSS_GI, sizeof(gi), &gi))
1106 			goto nla_put_failure;
1107 		break;
1108 	}
1109 	case CLG_GILB_ELL: {
1110 		struct tc_netem_gemodel ge = {
1111 			.p = q->clg.a1,
1112 			.r = q->clg.a2,
1113 			.h = q->clg.a3,
1114 			.k1 = q->clg.a4,
1115 		};
1116 
1117 		if (nla_put(skb, NETEM_LOSS_GE, sizeof(ge), &ge))
1118 			goto nla_put_failure;
1119 		break;
1120 	}
1121 	}
1122 
1123 	nla_nest_end(skb, nest);
1124 	return 0;
1125 
1126 nla_put_failure:
1127 	nla_nest_cancel(skb, nest);
1128 	return -1;
1129 }
1130 
1131 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
1132 {
1133 	const struct netem_sched_data *q = qdisc_priv(sch);
1134 	struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb);
1135 	struct tc_netem_qopt qopt;
1136 	struct tc_netem_corr cor;
1137 	struct tc_netem_reorder reorder;
1138 	struct tc_netem_corrupt corrupt;
1139 	struct tc_netem_rate rate;
1140 	struct tc_netem_slot slot;
1141 
1142 	qopt.latency = min_t(psched_tdiff_t, PSCHED_NS2TICKS(q->latency),
1143 			     UINT_MAX);
1144 	qopt.jitter = min_t(psched_tdiff_t, PSCHED_NS2TICKS(q->jitter),
1145 			    UINT_MAX);
1146 	qopt.limit = q->limit;
1147 	qopt.loss = q->loss;
1148 	qopt.gap = q->gap;
1149 	qopt.duplicate = q->duplicate;
1150 	if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
1151 		goto nla_put_failure;
1152 
1153 	if (nla_put(skb, TCA_NETEM_LATENCY64, sizeof(q->latency), &q->latency))
1154 		goto nla_put_failure;
1155 
1156 	if (nla_put(skb, TCA_NETEM_JITTER64, sizeof(q->jitter), &q->jitter))
1157 		goto nla_put_failure;
1158 
1159 	cor.delay_corr = q->delay_cor.rho;
1160 	cor.loss_corr = q->loss_cor.rho;
1161 	cor.dup_corr = q->dup_cor.rho;
1162 	if (nla_put(skb, TCA_NETEM_CORR, sizeof(cor), &cor))
1163 		goto nla_put_failure;
1164 
1165 	reorder.probability = q->reorder;
1166 	reorder.correlation = q->reorder_cor.rho;
1167 	if (nla_put(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder))
1168 		goto nla_put_failure;
1169 
1170 	corrupt.probability = q->corrupt;
1171 	corrupt.correlation = q->corrupt_cor.rho;
1172 	if (nla_put(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt))
1173 		goto nla_put_failure;
1174 
1175 	if (q->rate >= (1ULL << 32)) {
1176 		if (nla_put_u64_64bit(skb, TCA_NETEM_RATE64, q->rate,
1177 				      TCA_NETEM_PAD))
1178 			goto nla_put_failure;
1179 		rate.rate = ~0U;
1180 	} else {
1181 		rate.rate = q->rate;
1182 	}
1183 	rate.packet_overhead = q->packet_overhead;
1184 	rate.cell_size = q->cell_size;
1185 	rate.cell_overhead = q->cell_overhead;
1186 	if (nla_put(skb, TCA_NETEM_RATE, sizeof(rate), &rate))
1187 		goto nla_put_failure;
1188 
1189 	if (q->ecn && nla_put_u32(skb, TCA_NETEM_ECN, q->ecn))
1190 		goto nla_put_failure;
1191 
1192 	if (dump_loss_model(q, skb) != 0)
1193 		goto nla_put_failure;
1194 
1195 	if (q->slot_config.min_delay | q->slot_config.max_delay |
1196 	    q->slot_config.dist_jitter) {
1197 		slot = q->slot_config;
1198 		if (slot.max_packets == INT_MAX)
1199 			slot.max_packets = 0;
1200 		if (slot.max_bytes == INT_MAX)
1201 			slot.max_bytes = 0;
1202 		if (nla_put(skb, TCA_NETEM_SLOT, sizeof(slot), &slot))
1203 			goto nla_put_failure;
1204 	}
1205 
1206 	return nla_nest_end(skb, nla);
1207 
1208 nla_put_failure:
1209 	nlmsg_trim(skb, nla);
1210 	return -1;
1211 }
1212 
1213 static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
1214 			  struct sk_buff *skb, struct tcmsg *tcm)
1215 {
1216 	struct netem_sched_data *q = qdisc_priv(sch);
1217 
1218 	if (cl != 1 || !q->qdisc) 	/* only one class */
1219 		return -ENOENT;
1220 
1221 	tcm->tcm_handle |= TC_H_MIN(1);
1222 	tcm->tcm_info = q->qdisc->handle;
1223 
1224 	return 0;
1225 }
1226 
1227 static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1228 		     struct Qdisc **old, struct netlink_ext_ack *extack)
1229 {
1230 	struct netem_sched_data *q = qdisc_priv(sch);
1231 
1232 	*old = qdisc_replace(sch, new, &q->qdisc);
1233 	return 0;
1234 }
1235 
1236 static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
1237 {
1238 	struct netem_sched_data *q = qdisc_priv(sch);
1239 	return q->qdisc;
1240 }
1241 
1242 static unsigned long netem_find(struct Qdisc *sch, u32 classid)
1243 {
1244 	return 1;
1245 }
1246 
1247 static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
1248 {
1249 	if (!walker->stop) {
1250 		if (walker->count >= walker->skip)
1251 			if (walker->fn(sch, 1, walker) < 0) {
1252 				walker->stop = 1;
1253 				return;
1254 			}
1255 		walker->count++;
1256 	}
1257 }
1258 
1259 static const struct Qdisc_class_ops netem_class_ops = {
1260 	.graft		=	netem_graft,
1261 	.leaf		=	netem_leaf,
1262 	.find		=	netem_find,
1263 	.walk		=	netem_walk,
1264 	.dump		=	netem_dump_class,
1265 };
1266 
1267 static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
1268 	.id		=	"netem",
1269 	.cl_ops		=	&netem_class_ops,
1270 	.priv_size	=	sizeof(struct netem_sched_data),
1271 	.enqueue	=	netem_enqueue,
1272 	.dequeue	=	netem_dequeue,
1273 	.peek		=	qdisc_peek_dequeued,
1274 	.init		=	netem_init,
1275 	.reset		=	netem_reset,
1276 	.destroy	=	netem_destroy,
1277 	.change		=	netem_change,
1278 	.dump		=	netem_dump,
1279 	.owner		=	THIS_MODULE,
1280 };
1281 
1282 
1283 static int __init netem_module_init(void)
1284 {
1285 	pr_info("netem: version " VERSION "\n");
1286 	return register_qdisc(&netem_qdisc_ops);
1287 }
1288 static void __exit netem_module_exit(void)
1289 {
1290 	unregister_qdisc(&netem_qdisc_ops);
1291 }
1292 module_init(netem_module_init)
1293 module_exit(netem_module_exit)
1294 MODULE_LICENSE("GPL");
1295