xref: /openbmc/linux/net/sched/sch_netem.c (revision 6fa24b41)
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/gso.h>
25 #include <net/netlink.h>
26 #include <net/pkt_sched.h>
27 #include <net/inet_ecn.h>
28 
29 #define VERSION "1.3"
30 
31 /*	Network Emulation Queuing algorithm.
32 	====================================
33 
34 	Sources: [1] Mark Carson, Darrin Santay, "NIST Net - A Linux-based
35 		 Network Emulation Tool
36 		 [2] Luigi Rizzo, DummyNet for FreeBSD
37 
38 	 ----------------------------------------------------------------
39 
40 	 This started out as a simple way to delay outgoing packets to
41 	 test TCP but has grown to include most of the functionality
42 	 of a full blown network emulator like NISTnet. It can delay
43 	 packets and add random jitter (and correlation). The random
44 	 distribution can be loaded from a table as well to provide
45 	 normal, Pareto, or experimental curves. Packet loss,
46 	 duplication, and reordering can also be emulated.
47 
48 	 This qdisc does not do classification that can be handled in
49 	 layering other disciplines.  It does not need to do bandwidth
50 	 control either since that can be handled by using token
51 	 bucket or other rate control.
52 
53      Correlated Loss Generator models
54 
55 	Added generation of correlated loss according to the
56 	"Gilbert-Elliot" model, a 4-state markov model.
57 
58 	References:
59 	[1] NetemCLG Home http://netgroup.uniroma2.it/NetemCLG
60 	[2] S. Salsano, F. Ludovici, A. Ordine, "Definition of a general
61 	and intuitive loss model for packet networks and its implementation
62 	in the Netem module in the Linux kernel", available in [1]
63 
64 	Authors: Stefano Salsano <stefano.salsano at uniroma2.it
65 		 Fabio Ludovici <fabio.ludovici at yahoo.it>
66 */
67 
68 struct disttable {
69 	u32  size;
70 	s16 table[];
71 };
72 
73 struct netem_sched_data {
74 	/* internal t(ime)fifo qdisc uses t_root and sch->limit */
75 	struct rb_root t_root;
76 
77 	/* a linear queue; reduces rbtree rebalancing when jitter is low */
78 	struct sk_buff	*t_head;
79 	struct sk_buff	*t_tail;
80 
81 	/* optional qdisc for classful handling (NULL at netem init) */
82 	struct Qdisc	*qdisc;
83 
84 	struct qdisc_watchdog watchdog;
85 
86 	s64 latency;
87 	s64 jitter;
88 
89 	u32 loss;
90 	u32 ecn;
91 	u32 limit;
92 	u32 counter;
93 	u32 gap;
94 	u32 duplicate;
95 	u32 reorder;
96 	u32 corrupt;
97 	u64 rate;
98 	s32 packet_overhead;
99 	u32 cell_size;
100 	struct reciprocal_value cell_size_reciprocal;
101 	s32 cell_overhead;
102 
103 	struct crndstate {
104 		u32 last;
105 		u32 rho;
106 	} delay_cor, loss_cor, dup_cor, reorder_cor, corrupt_cor;
107 
108 	struct prng  {
109 		u64 seed;
110 		struct rnd_state prng_state;
111 	} prng;
112 
113 	struct disttable *delay_dist;
114 
115 	enum  {
116 		CLG_RANDOM,
117 		CLG_4_STATES,
118 		CLG_GILB_ELL,
119 	} loss_model;
120 
121 	enum {
122 		TX_IN_GAP_PERIOD = 1,
123 		TX_IN_BURST_PERIOD,
124 		LOST_IN_GAP_PERIOD,
125 		LOST_IN_BURST_PERIOD,
126 	} _4_state_model;
127 
128 	enum {
129 		GOOD_STATE = 1,
130 		BAD_STATE,
131 	} GE_state_model;
132 
133 	/* Correlated Loss Generation models */
134 	struct clgstate {
135 		/* state of the Markov chain */
136 		u8 state;
137 
138 		/* 4-states and Gilbert-Elliot models */
139 		u32 a1;	/* p13 for 4-states or p for GE */
140 		u32 a2;	/* p31 for 4-states or r for GE */
141 		u32 a3;	/* p32 for 4-states or h for GE */
142 		u32 a4;	/* p14 for 4-states or 1-k for GE */
143 		u32 a5; /* p23 used only in 4-states */
144 	} clg;
145 
146 	struct tc_netem_slot slot_config;
147 	struct slotstate {
148 		u64 slot_next;
149 		s32 packets_left;
150 		s32 bytes_left;
151 	} slot;
152 
153 	struct disttable *slot_dist;
154 };
155 
156 /* Time stamp put into socket buffer control block
157  * Only valid when skbs are in our internal t(ime)fifo queue.
158  *
159  * As skb->rbnode uses same storage than skb->next, skb->prev and skb->tstamp,
160  * and skb->next & skb->prev are scratch space for a qdisc,
161  * we save skb->tstamp value in skb->cb[] before destroying it.
162  */
163 struct netem_skb_cb {
164 	u64	        time_to_send;
165 };
166 
167 static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb)
168 {
169 	/* we assume we can use skb next/prev/tstamp as storage for rb_node */
170 	qdisc_cb_private_validate(skb, sizeof(struct netem_skb_cb));
171 	return (struct netem_skb_cb *)qdisc_skb_cb(skb)->data;
172 }
173 
174 /* init_crandom - initialize correlated random number generator
175  * Use entropy source for initial seed.
176  */
177 static void init_crandom(struct crndstate *state, unsigned long rho)
178 {
179 	state->rho = rho;
180 	state->last = get_random_u32();
181 }
182 
183 /* get_crandom - correlated random number generator
184  * Next number depends on last value.
185  * rho is scaled to avoid floating point.
186  */
187 static u32 get_crandom(struct crndstate *state, struct prng *p)
188 {
189 	u64 value, rho;
190 	unsigned long answer;
191 	struct rnd_state *s = &p->prng_state;
192 
193 	if (!state || state->rho == 0)	/* no correlation */
194 		return prandom_u32_state(s);
195 
196 	value = prandom_u32_state(s);
197 	rho = (u64)state->rho + 1;
198 	answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32;
199 	state->last = answer;
200 	return answer;
201 }
202 
203 /* loss_4state - 4-state model loss generator
204  * Generates losses according to the 4-state Markov chain adopted in
205  * the GI (General and Intuitive) loss model.
206  */
207 static bool loss_4state(struct netem_sched_data *q)
208 {
209 	struct clgstate *clg = &q->clg;
210 	u32 rnd = prandom_u32_state(&q->prng.prng_state);
211 
212 	/*
213 	 * Makes a comparison between rnd and the transition
214 	 * probabilities outgoing from the current state, then decides the
215 	 * next state and if the next packet has to be transmitted or lost.
216 	 * The four states correspond to:
217 	 *   TX_IN_GAP_PERIOD => successfully transmitted packets within a gap period
218 	 *   LOST_IN_GAP_PERIOD => isolated losses within a gap period
219 	 *   LOST_IN_BURST_PERIOD => lost packets within a burst period
220 	 *   TX_IN_BURST_PERIOD => successfully transmitted packets within a burst period
221 	 */
222 	switch (clg->state) {
223 	case TX_IN_GAP_PERIOD:
224 		if (rnd < clg->a4) {
225 			clg->state = LOST_IN_GAP_PERIOD;
226 			return true;
227 		} else if (clg->a4 < rnd && rnd < clg->a1 + clg->a4) {
228 			clg->state = LOST_IN_BURST_PERIOD;
229 			return true;
230 		} else if (clg->a1 + clg->a4 < rnd) {
231 			clg->state = TX_IN_GAP_PERIOD;
232 		}
233 
234 		break;
235 	case TX_IN_BURST_PERIOD:
236 		if (rnd < clg->a5) {
237 			clg->state = LOST_IN_BURST_PERIOD;
238 			return true;
239 		} else {
240 			clg->state = TX_IN_BURST_PERIOD;
241 		}
242 
243 		break;
244 	case LOST_IN_BURST_PERIOD:
245 		if (rnd < clg->a3)
246 			clg->state = TX_IN_BURST_PERIOD;
247 		else if (clg->a3 < rnd && rnd < clg->a2 + clg->a3) {
248 			clg->state = TX_IN_GAP_PERIOD;
249 		} else if (clg->a2 + clg->a3 < rnd) {
250 			clg->state = LOST_IN_BURST_PERIOD;
251 			return true;
252 		}
253 		break;
254 	case LOST_IN_GAP_PERIOD:
255 		clg->state = TX_IN_GAP_PERIOD;
256 		break;
257 	}
258 
259 	return false;
260 }
261 
262 /* loss_gilb_ell - Gilbert-Elliot model loss generator
263  * Generates losses according to the Gilbert-Elliot loss model or
264  * its special cases  (Gilbert or Simple Gilbert)
265  *
266  * Makes a comparison between random number and the transition
267  * probabilities outgoing from the current state, then decides the
268  * next state. A second random number is extracted and the comparison
269  * with the loss probability of the current state decides if the next
270  * packet will be transmitted or lost.
271  */
272 static bool loss_gilb_ell(struct netem_sched_data *q)
273 {
274 	struct clgstate *clg = &q->clg;
275 	struct rnd_state *s = &q->prng.prng_state;
276 
277 	switch (clg->state) {
278 	case GOOD_STATE:
279 		if (prandom_u32_state(s) < clg->a1)
280 			clg->state = BAD_STATE;
281 		if (prandom_u32_state(s) < clg->a4)
282 			return true;
283 		break;
284 	case BAD_STATE:
285 		if (prandom_u32_state(s) < clg->a2)
286 			clg->state = GOOD_STATE;
287 		if (prandom_u32_state(s) > clg->a3)
288 			return true;
289 	}
290 
291 	return false;
292 }
293 
294 static bool loss_event(struct netem_sched_data *q)
295 {
296 	switch (q->loss_model) {
297 	case CLG_RANDOM:
298 		/* Random packet drop 0 => none, ~0 => all */
299 		return q->loss && q->loss >= get_crandom(&q->loss_cor, &q->prng);
300 
301 	case CLG_4_STATES:
302 		/* 4state loss model algorithm (used also for GI model)
303 		* Extracts a value from the markov 4 state 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_4state(q);
308 
309 	case CLG_GILB_ELL:
310 		/* Gilbert-Elliot loss model algorithm
311 		* Extracts a value from the Gilbert-Elliot loss generator,
312 		* if it is 1 drops a packet and if needed writes the event in
313 		* the kernel logs
314 		*/
315 		return loss_gilb_ell(q);
316 	}
317 
318 	return false;	/* not reached */
319 }
320 
321 
322 /* tabledist - return a pseudo-randomly distributed value with mean mu and
323  * std deviation sigma.  Uses table lookup to approximate the desired
324  * distribution, and a uniformly-distributed pseudo-random source.
325  */
326 static s64 tabledist(s64 mu, s32 sigma,
327 		     struct crndstate *state,
328 		     struct prng *prng,
329 		     const struct disttable *dist)
330 {
331 	s64 x;
332 	long t;
333 	u32 rnd;
334 
335 	if (sigma == 0)
336 		return mu;
337 
338 	rnd = get_crandom(state, prng);
339 
340 	/* default uniform distribution */
341 	if (dist == NULL)
342 		return ((rnd % (2 * (u32)sigma)) + mu) - sigma;
343 
344 	t = dist->table[rnd % dist->size];
345 	x = (sigma % NETEM_DIST_SCALE) * t;
346 	if (x >= 0)
347 		x += NETEM_DIST_SCALE/2;
348 	else
349 		x -= NETEM_DIST_SCALE/2;
350 
351 	return  x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu;
352 }
353 
354 static u64 packet_time_ns(u64 len, const struct netem_sched_data *q)
355 {
356 	len += q->packet_overhead;
357 
358 	if (q->cell_size) {
359 		u32 cells = reciprocal_divide(len, q->cell_size_reciprocal);
360 
361 		if (len > cells * q->cell_size)	/* extra cell needed for remainder */
362 			cells++;
363 		len = cells * (q->cell_size + q->cell_overhead);
364 	}
365 
366 	return div64_u64(len * NSEC_PER_SEC, q->rate);
367 }
368 
369 static void tfifo_reset(struct Qdisc *sch)
370 {
371 	struct netem_sched_data *q = qdisc_priv(sch);
372 	struct rb_node *p = rb_first(&q->t_root);
373 
374 	while (p) {
375 		struct sk_buff *skb = rb_to_skb(p);
376 
377 		p = rb_next(p);
378 		rb_erase(&skb->rbnode, &q->t_root);
379 		rtnl_kfree_skbs(skb, skb);
380 	}
381 
382 	rtnl_kfree_skbs(q->t_head, q->t_tail);
383 	q->t_head = NULL;
384 	q->t_tail = NULL;
385 }
386 
387 static void tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
388 {
389 	struct netem_sched_data *q = qdisc_priv(sch);
390 	u64 tnext = netem_skb_cb(nskb)->time_to_send;
391 
392 	if (!q->t_tail || tnext >= netem_skb_cb(q->t_tail)->time_to_send) {
393 		if (q->t_tail)
394 			q->t_tail->next = nskb;
395 		else
396 			q->t_head = nskb;
397 		q->t_tail = nskb;
398 	} else {
399 		struct rb_node **p = &q->t_root.rb_node, *parent = NULL;
400 
401 		while (*p) {
402 			struct sk_buff *skb;
403 
404 			parent = *p;
405 			skb = rb_to_skb(parent);
406 			if (tnext >= netem_skb_cb(skb)->time_to_send)
407 				p = &parent->rb_right;
408 			else
409 				p = &parent->rb_left;
410 		}
411 		rb_link_node(&nskb->rbnode, parent, p);
412 		rb_insert_color(&nskb->rbnode, &q->t_root);
413 	}
414 	sch->q.qlen++;
415 }
416 
417 /* netem can't properly corrupt a megapacket (like we get from GSO), so instead
418  * when we statistically choose to corrupt one, we instead segment it, returning
419  * the first packet to be corrupted, and re-enqueue the remaining frames
420  */
421 static struct sk_buff *netem_segment(struct sk_buff *skb, struct Qdisc *sch,
422 				     struct sk_buff **to_free)
423 {
424 	struct sk_buff *segs;
425 	netdev_features_t features = netif_skb_features(skb);
426 
427 	segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
428 
429 	if (IS_ERR_OR_NULL(segs)) {
430 		qdisc_drop(skb, sch, to_free);
431 		return NULL;
432 	}
433 	consume_skb(skb);
434 	return segs;
435 }
436 
437 /*
438  * Insert one skb into qdisc.
439  * Note: parent depends on return value to account for queue length.
440  * 	NET_XMIT_DROP: queue length didn't change.
441  *      NET_XMIT_SUCCESS: one skb was queued.
442  */
443 static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch,
444 			 struct sk_buff **to_free)
445 {
446 	struct netem_sched_data *q = qdisc_priv(sch);
447 	/* We don't fill cb now as skb_unshare() may invalidate it */
448 	struct netem_skb_cb *cb;
449 	struct sk_buff *skb2 = NULL;
450 	struct sk_buff *segs = NULL;
451 	unsigned int prev_len = qdisc_pkt_len(skb);
452 	int count = 1;
453 
454 	/* Do not fool qdisc_drop_all() */
455 	skb->prev = NULL;
456 
457 	/* Random duplication */
458 	if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor, &q->prng))
459 		++count;
460 
461 	/* Drop packet? */
462 	if (loss_event(q)) {
463 		if (q->ecn && INET_ECN_set_ce(skb))
464 			qdisc_qstats_drop(sch); /* mark packet */
465 		else
466 			--count;
467 	}
468 	if (count == 0) {
469 		qdisc_qstats_drop(sch);
470 		__qdisc_drop(skb, to_free);
471 		return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
472 	}
473 
474 	/* If a delay is expected, orphan the skb. (orphaning usually takes
475 	 * place at TX completion time, so _before_ the link transit delay)
476 	 */
477 	if (q->latency || q->jitter || q->rate)
478 		skb_orphan_partial(skb);
479 
480 	/*
481 	 * If we need to duplicate packet, then clone it before
482 	 * original is modified.
483 	 */
484 	if (count > 1)
485 		skb2 = skb_clone(skb, GFP_ATOMIC);
486 
487 	/*
488 	 * Randomized packet corruption.
489 	 * Make copy if needed since we are modifying
490 	 * If packet is going to be hardware checksummed, then
491 	 * do it now in software before we mangle it.
492 	 */
493 	if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor, &q->prng)) {
494 		if (skb_is_gso(skb)) {
495 			skb = netem_segment(skb, sch, to_free);
496 			if (!skb)
497 				goto finish_segs;
498 
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[get_random_u32_below(skb_headlen(skb))] ^=
517 			1<<get_random_u32_below(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 		if (skb2)
525 			__qdisc_drop(skb2, to_free);
526 		return NET_XMIT_DROP;
527 	}
528 
529 	/*
530 	 * If doing duplication then re-insert at top of the
531 	 * qdisc tree, since parent queuer expects that only one
532 	 * skb will be queued.
533 	 */
534 	if (skb2) {
535 		struct Qdisc *rootq = qdisc_root_bh(sch);
536 		u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
537 
538 		q->duplicate = 0;
539 		rootq->enqueue(skb2, rootq, to_free);
540 		q->duplicate = dupsave;
541 		skb2 = NULL;
542 	}
543 
544 	qdisc_qstats_backlog_inc(sch, skb);
545 
546 	cb = netem_skb_cb(skb);
547 	if (q->gap == 0 ||		/* not doing reordering */
548 	    q->counter < q->gap - 1 ||	/* inside last reordering gap */
549 	    q->reorder < get_crandom(&q->reorder_cor, &q->prng)) {
550 		u64 now;
551 		s64 delay;
552 
553 		delay = tabledist(q->latency, q->jitter,
554 				  &q->delay_cor, &q->prng, q->delay_dist);
555 
556 		now = ktime_get_ns();
557 
558 		if (q->rate) {
559 			struct netem_skb_cb *last = NULL;
560 
561 			if (sch->q.tail)
562 				last = netem_skb_cb(sch->q.tail);
563 			if (q->t_root.rb_node) {
564 				struct sk_buff *t_skb;
565 				struct netem_skb_cb *t_last;
566 
567 				t_skb = skb_rb_last(&q->t_root);
568 				t_last = netem_skb_cb(t_skb);
569 				if (!last ||
570 				    t_last->time_to_send > last->time_to_send)
571 					last = t_last;
572 			}
573 			if (q->t_tail) {
574 				struct netem_skb_cb *t_last =
575 					netem_skb_cb(q->t_tail);
576 
577 				if (!last ||
578 				    t_last->time_to_send > last->time_to_send)
579 					last = t_last;
580 			}
581 
582 			if (last) {
583 				/*
584 				 * Last packet in queue is reference point (now),
585 				 * calculate this time bonus and subtract
586 				 * from delay.
587 				 */
588 				delay -= last->time_to_send - now;
589 				delay = max_t(s64, 0, delay);
590 				now = last->time_to_send;
591 			}
592 
593 			delay += packet_time_ns(qdisc_pkt_len(skb), q);
594 		}
595 
596 		cb->time_to_send = now + delay;
597 		++q->counter;
598 		tfifo_enqueue(skb, sch);
599 	} else {
600 		/*
601 		 * Do re-ordering by putting one out of N packets at the front
602 		 * of the queue.
603 		 */
604 		cb->time_to_send = ktime_get_ns();
605 		q->counter = 0;
606 
607 		__qdisc_enqueue_head(skb, &sch->q);
608 		sch->qstats.requeues++;
609 	}
610 
611 finish_segs:
612 	if (skb2)
613 		__qdisc_drop(skb2, to_free);
614 
615 	if (segs) {
616 		unsigned int len, last_len;
617 		int rc, nb;
618 
619 		len = skb ? skb->len : 0;
620 		nb = skb ? 1 : 0;
621 
622 		while (segs) {
623 			skb2 = segs->next;
624 			skb_mark_not_on_list(segs);
625 			qdisc_skb_cb(segs)->pkt_len = segs->len;
626 			last_len = segs->len;
627 			rc = qdisc_enqueue(segs, sch, to_free);
628 			if (rc != NET_XMIT_SUCCESS) {
629 				if (net_xmit_drop_count(rc))
630 					qdisc_qstats_drop(sch);
631 			} else {
632 				nb++;
633 				len += last_len;
634 			}
635 			segs = skb2;
636 		}
637 		/* Parent qdiscs accounted for 1 skb of size @prev_len */
638 		qdisc_tree_reduce_backlog(sch, -(nb - 1), -(len - prev_len));
639 	} else if (!skb) {
640 		return NET_XMIT_DROP;
641 	}
642 	return NET_XMIT_SUCCESS;
643 }
644 
645 /* Delay the next round with a new future slot with a
646  * correct number of bytes and packets.
647  */
648 
649 static void get_slot_next(struct netem_sched_data *q, u64 now)
650 {
651 	s64 next_delay;
652 
653 	if (!q->slot_dist)
654 		next_delay = q->slot_config.min_delay +
655 				(get_random_u32() *
656 				 (q->slot_config.max_delay -
657 				  q->slot_config.min_delay) >> 32);
658 	else
659 		next_delay = tabledist(q->slot_config.dist_delay,
660 				       (s32)(q->slot_config.dist_jitter),
661 				       NULL, &q->prng, q->slot_dist);
662 
663 	q->slot.slot_next = now + next_delay;
664 	q->slot.packets_left = q->slot_config.max_packets;
665 	q->slot.bytes_left = q->slot_config.max_bytes;
666 }
667 
668 static struct sk_buff *netem_peek(struct netem_sched_data *q)
669 {
670 	struct sk_buff *skb = skb_rb_first(&q->t_root);
671 	u64 t1, t2;
672 
673 	if (!skb)
674 		return q->t_head;
675 	if (!q->t_head)
676 		return skb;
677 
678 	t1 = netem_skb_cb(skb)->time_to_send;
679 	t2 = netem_skb_cb(q->t_head)->time_to_send;
680 	if (t1 < t2)
681 		return skb;
682 	return q->t_head;
683 }
684 
685 static void netem_erase_head(struct netem_sched_data *q, struct sk_buff *skb)
686 {
687 	if (skb == q->t_head) {
688 		q->t_head = skb->next;
689 		if (!q->t_head)
690 			q->t_tail = NULL;
691 	} else {
692 		rb_erase(&skb->rbnode, &q->t_root);
693 	}
694 }
695 
696 static struct sk_buff *netem_dequeue(struct Qdisc *sch)
697 {
698 	struct netem_sched_data *q = qdisc_priv(sch);
699 	struct sk_buff *skb;
700 
701 tfifo_dequeue:
702 	skb = __qdisc_dequeue_head(&sch->q);
703 	if (skb) {
704 		qdisc_qstats_backlog_dec(sch, skb);
705 deliver:
706 		qdisc_bstats_update(sch, skb);
707 		return skb;
708 	}
709 	skb = netem_peek(q);
710 	if (skb) {
711 		u64 time_to_send;
712 		u64 now = ktime_get_ns();
713 
714 		/* if more time remaining? */
715 		time_to_send = netem_skb_cb(skb)->time_to_send;
716 		if (q->slot.slot_next && q->slot.slot_next < time_to_send)
717 			get_slot_next(q, now);
718 
719 		if (time_to_send <= now && q->slot.slot_next <= now) {
720 			netem_erase_head(q, skb);
721 			sch->q.qlen--;
722 			qdisc_qstats_backlog_dec(sch, skb);
723 			skb->next = NULL;
724 			skb->prev = NULL;
725 			/* skb->dev shares skb->rbnode area,
726 			 * we need to restore its value.
727 			 */
728 			skb->dev = qdisc_dev(sch);
729 
730 			if (q->slot.slot_next) {
731 				q->slot.packets_left--;
732 				q->slot.bytes_left -= qdisc_pkt_len(skb);
733 				if (q->slot.packets_left <= 0 ||
734 				    q->slot.bytes_left <= 0)
735 					get_slot_next(q, now);
736 			}
737 
738 			if (q->qdisc) {
739 				unsigned int pkt_len = qdisc_pkt_len(skb);
740 				struct sk_buff *to_free = NULL;
741 				int err;
742 
743 				err = qdisc_enqueue(skb, q->qdisc, &to_free);
744 				kfree_skb_list(to_free);
745 				if (err != NET_XMIT_SUCCESS) {
746 					if (net_xmit_drop_count(err))
747 						qdisc_qstats_drop(sch);
748 					qdisc_tree_reduce_backlog(sch, 1, pkt_len);
749 				}
750 				goto tfifo_dequeue;
751 			}
752 			goto deliver;
753 		}
754 
755 		if (q->qdisc) {
756 			skb = q->qdisc->ops->dequeue(q->qdisc);
757 			if (skb)
758 				goto deliver;
759 		}
760 
761 		qdisc_watchdog_schedule_ns(&q->watchdog,
762 					   max(time_to_send,
763 					       q->slot.slot_next));
764 	}
765 
766 	if (q->qdisc) {
767 		skb = q->qdisc->ops->dequeue(q->qdisc);
768 		if (skb)
769 			goto deliver;
770 	}
771 	return NULL;
772 }
773 
774 static void netem_reset(struct Qdisc *sch)
775 {
776 	struct netem_sched_data *q = qdisc_priv(sch);
777 
778 	qdisc_reset_queue(sch);
779 	tfifo_reset(sch);
780 	if (q->qdisc)
781 		qdisc_reset(q->qdisc);
782 	qdisc_watchdog_cancel(&q->watchdog);
783 }
784 
785 static void dist_free(struct disttable *d)
786 {
787 	kvfree(d);
788 }
789 
790 /*
791  * Distribution data is a variable size payload containing
792  * signed 16 bit values.
793  */
794 
795 static int get_dist_table(struct disttable **tbl, const struct nlattr *attr)
796 {
797 	size_t n = nla_len(attr)/sizeof(__s16);
798 	const __s16 *data = nla_data(attr);
799 	struct disttable *d;
800 	int i;
801 
802 	if (!n || n > NETEM_DIST_MAX)
803 		return -EINVAL;
804 
805 	d = kvmalloc(struct_size(d, table, n), GFP_KERNEL);
806 	if (!d)
807 		return -ENOMEM;
808 
809 	d->size = n;
810 	for (i = 0; i < n; i++)
811 		d->table[i] = data[i];
812 
813 	*tbl = d;
814 	return 0;
815 }
816 
817 static void get_slot(struct netem_sched_data *q, const struct nlattr *attr)
818 {
819 	const struct tc_netem_slot *c = nla_data(attr);
820 
821 	q->slot_config = *c;
822 	if (q->slot_config.max_packets == 0)
823 		q->slot_config.max_packets = INT_MAX;
824 	if (q->slot_config.max_bytes == 0)
825 		q->slot_config.max_bytes = INT_MAX;
826 
827 	/* capping dist_jitter to the range acceptable by tabledist() */
828 	q->slot_config.dist_jitter = min_t(__s64, INT_MAX, abs(q->slot_config.dist_jitter));
829 
830 	q->slot.packets_left = q->slot_config.max_packets;
831 	q->slot.bytes_left = q->slot_config.max_bytes;
832 	if (q->slot_config.min_delay | q->slot_config.max_delay |
833 	    q->slot_config.dist_jitter)
834 		q->slot.slot_next = ktime_get_ns();
835 	else
836 		q->slot.slot_next = 0;
837 }
838 
839 static void get_correlation(struct netem_sched_data *q, const struct nlattr *attr)
840 {
841 	const struct tc_netem_corr *c = nla_data(attr);
842 
843 	init_crandom(&q->delay_cor, c->delay_corr);
844 	init_crandom(&q->loss_cor, c->loss_corr);
845 	init_crandom(&q->dup_cor, c->dup_corr);
846 }
847 
848 static void get_reorder(struct netem_sched_data *q, const struct nlattr *attr)
849 {
850 	const struct tc_netem_reorder *r = nla_data(attr);
851 
852 	q->reorder = r->probability;
853 	init_crandom(&q->reorder_cor, r->correlation);
854 }
855 
856 static void get_corrupt(struct netem_sched_data *q, const struct nlattr *attr)
857 {
858 	const struct tc_netem_corrupt *r = nla_data(attr);
859 
860 	q->corrupt = r->probability;
861 	init_crandom(&q->corrupt_cor, r->correlation);
862 }
863 
864 static void get_rate(struct netem_sched_data *q, const struct nlattr *attr)
865 {
866 	const struct tc_netem_rate *r = nla_data(attr);
867 
868 	q->rate = r->rate;
869 	q->packet_overhead = r->packet_overhead;
870 	q->cell_size = r->cell_size;
871 	q->cell_overhead = r->cell_overhead;
872 	if (q->cell_size)
873 		q->cell_size_reciprocal = reciprocal_value(q->cell_size);
874 	else
875 		q->cell_size_reciprocal = (struct reciprocal_value) { 0 };
876 }
877 
878 static int get_loss_clg(struct netem_sched_data *q, const struct nlattr *attr)
879 {
880 	const struct nlattr *la;
881 	int rem;
882 
883 	nla_for_each_nested(la, attr, rem) {
884 		u16 type = nla_type(la);
885 
886 		switch (type) {
887 		case NETEM_LOSS_GI: {
888 			const struct tc_netem_gimodel *gi = nla_data(la);
889 
890 			if (nla_len(la) < sizeof(struct tc_netem_gimodel)) {
891 				pr_info("netem: incorrect gi model size\n");
892 				return -EINVAL;
893 			}
894 
895 			q->loss_model = CLG_4_STATES;
896 
897 			q->clg.state = TX_IN_GAP_PERIOD;
898 			q->clg.a1 = gi->p13;
899 			q->clg.a2 = gi->p31;
900 			q->clg.a3 = gi->p32;
901 			q->clg.a4 = gi->p14;
902 			q->clg.a5 = gi->p23;
903 			break;
904 		}
905 
906 		case NETEM_LOSS_GE: {
907 			const struct tc_netem_gemodel *ge = nla_data(la);
908 
909 			if (nla_len(la) < sizeof(struct tc_netem_gemodel)) {
910 				pr_info("netem: incorrect ge model size\n");
911 				return -EINVAL;
912 			}
913 
914 			q->loss_model = CLG_GILB_ELL;
915 			q->clg.state = GOOD_STATE;
916 			q->clg.a1 = ge->p;
917 			q->clg.a2 = ge->r;
918 			q->clg.a3 = ge->h;
919 			q->clg.a4 = ge->k1;
920 			break;
921 		}
922 
923 		default:
924 			pr_info("netem: unknown loss type %u\n", type);
925 			return -EINVAL;
926 		}
927 	}
928 
929 	return 0;
930 }
931 
932 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
933 	[TCA_NETEM_CORR]	= { .len = sizeof(struct tc_netem_corr) },
934 	[TCA_NETEM_REORDER]	= { .len = sizeof(struct tc_netem_reorder) },
935 	[TCA_NETEM_CORRUPT]	= { .len = sizeof(struct tc_netem_corrupt) },
936 	[TCA_NETEM_RATE]	= { .len = sizeof(struct tc_netem_rate) },
937 	[TCA_NETEM_LOSS]	= { .type = NLA_NESTED },
938 	[TCA_NETEM_ECN]		= { .type = NLA_U32 },
939 	[TCA_NETEM_RATE64]	= { .type = NLA_U64 },
940 	[TCA_NETEM_LATENCY64]	= { .type = NLA_S64 },
941 	[TCA_NETEM_JITTER64]	= { .type = NLA_S64 },
942 	[TCA_NETEM_SLOT]	= { .len = sizeof(struct tc_netem_slot) },
943 	[TCA_NETEM_PRNG_SEED]	= { .type = NLA_U64 },
944 };
945 
946 static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
947 		      const struct nla_policy *policy, int len)
948 {
949 	int nested_len = nla_len(nla) - NLA_ALIGN(len);
950 
951 	if (nested_len < 0) {
952 		pr_info("netem: invalid attributes len %d\n", nested_len);
953 		return -EINVAL;
954 	}
955 
956 	if (nested_len >= nla_attr_size(0))
957 		return nla_parse_deprecated(tb, maxtype,
958 					    nla_data(nla) + NLA_ALIGN(len),
959 					    nested_len, policy, NULL);
960 
961 	memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
962 	return 0;
963 }
964 
965 /* Parse netlink message to set options */
966 static int netem_change(struct Qdisc *sch, struct nlattr *opt,
967 			struct netlink_ext_ack *extack)
968 {
969 	struct netem_sched_data *q = qdisc_priv(sch);
970 	struct nlattr *tb[TCA_NETEM_MAX + 1];
971 	struct disttable *delay_dist = NULL;
972 	struct disttable *slot_dist = NULL;
973 	struct tc_netem_qopt *qopt;
974 	struct clgstate old_clg;
975 	int old_loss_model = CLG_RANDOM;
976 	int ret;
977 
978 	qopt = nla_data(opt);
979 	ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
980 	if (ret < 0)
981 		return ret;
982 
983 	if (tb[TCA_NETEM_DELAY_DIST]) {
984 		ret = get_dist_table(&delay_dist, tb[TCA_NETEM_DELAY_DIST]);
985 		if (ret)
986 			goto table_free;
987 	}
988 
989 	if (tb[TCA_NETEM_SLOT_DIST]) {
990 		ret = get_dist_table(&slot_dist, tb[TCA_NETEM_SLOT_DIST]);
991 		if (ret)
992 			goto table_free;
993 	}
994 
995 	sch_tree_lock(sch);
996 	/* backup q->clg and q->loss_model */
997 	old_clg = q->clg;
998 	old_loss_model = q->loss_model;
999 
1000 	if (tb[TCA_NETEM_LOSS]) {
1001 		ret = get_loss_clg(q, tb[TCA_NETEM_LOSS]);
1002 		if (ret) {
1003 			q->loss_model = old_loss_model;
1004 			q->clg = old_clg;
1005 			goto unlock;
1006 		}
1007 	} else {
1008 		q->loss_model = CLG_RANDOM;
1009 	}
1010 
1011 	if (delay_dist)
1012 		swap(q->delay_dist, delay_dist);
1013 	if (slot_dist)
1014 		swap(q->slot_dist, slot_dist);
1015 	sch->limit = qopt->limit;
1016 
1017 	q->latency = PSCHED_TICKS2NS(qopt->latency);
1018 	q->jitter = PSCHED_TICKS2NS(qopt->jitter);
1019 	q->limit = qopt->limit;
1020 	q->gap = qopt->gap;
1021 	q->counter = 0;
1022 	q->loss = qopt->loss;
1023 	q->duplicate = qopt->duplicate;
1024 
1025 	/* for compatibility with earlier versions.
1026 	 * if gap is set, need to assume 100% probability
1027 	 */
1028 	if (q->gap)
1029 		q->reorder = ~0;
1030 
1031 	if (tb[TCA_NETEM_CORR])
1032 		get_correlation(q, tb[TCA_NETEM_CORR]);
1033 
1034 	if (tb[TCA_NETEM_REORDER])
1035 		get_reorder(q, tb[TCA_NETEM_REORDER]);
1036 
1037 	if (tb[TCA_NETEM_CORRUPT])
1038 		get_corrupt(q, tb[TCA_NETEM_CORRUPT]);
1039 
1040 	if (tb[TCA_NETEM_RATE])
1041 		get_rate(q, tb[TCA_NETEM_RATE]);
1042 
1043 	if (tb[TCA_NETEM_RATE64])
1044 		q->rate = max_t(u64, q->rate,
1045 				nla_get_u64(tb[TCA_NETEM_RATE64]));
1046 
1047 	if (tb[TCA_NETEM_LATENCY64])
1048 		q->latency = nla_get_s64(tb[TCA_NETEM_LATENCY64]);
1049 
1050 	if (tb[TCA_NETEM_JITTER64])
1051 		q->jitter = nla_get_s64(tb[TCA_NETEM_JITTER64]);
1052 
1053 	if (tb[TCA_NETEM_ECN])
1054 		q->ecn = nla_get_u32(tb[TCA_NETEM_ECN]);
1055 
1056 	if (tb[TCA_NETEM_SLOT])
1057 		get_slot(q, tb[TCA_NETEM_SLOT]);
1058 
1059 	/* capping jitter to the range acceptable by tabledist() */
1060 	q->jitter = min_t(s64, abs(q->jitter), INT_MAX);
1061 
1062 	if (tb[TCA_NETEM_PRNG_SEED])
1063 		q->prng.seed = nla_get_u64(tb[TCA_NETEM_PRNG_SEED]);
1064 	else
1065 		q->prng.seed = get_random_u64();
1066 	prandom_seed_state(&q->prng.prng_state, q->prng.seed);
1067 
1068 unlock:
1069 	sch_tree_unlock(sch);
1070 
1071 table_free:
1072 	dist_free(delay_dist);
1073 	dist_free(slot_dist);
1074 	return ret;
1075 }
1076 
1077 static int netem_init(struct Qdisc *sch, struct nlattr *opt,
1078 		      struct netlink_ext_ack *extack)
1079 {
1080 	struct netem_sched_data *q = qdisc_priv(sch);
1081 	int ret;
1082 
1083 	qdisc_watchdog_init(&q->watchdog, sch);
1084 
1085 	if (!opt)
1086 		return -EINVAL;
1087 
1088 	q->loss_model = CLG_RANDOM;
1089 	ret = netem_change(sch, opt, extack);
1090 	if (ret)
1091 		pr_info("netem: change failed\n");
1092 	return ret;
1093 }
1094 
1095 static void netem_destroy(struct Qdisc *sch)
1096 {
1097 	struct netem_sched_data *q = qdisc_priv(sch);
1098 
1099 	qdisc_watchdog_cancel(&q->watchdog);
1100 	if (q->qdisc)
1101 		qdisc_put(q->qdisc);
1102 	dist_free(q->delay_dist);
1103 	dist_free(q->slot_dist);
1104 }
1105 
1106 static int dump_loss_model(const struct netem_sched_data *q,
1107 			   struct sk_buff *skb)
1108 {
1109 	struct nlattr *nest;
1110 
1111 	nest = nla_nest_start_noflag(skb, TCA_NETEM_LOSS);
1112 	if (nest == NULL)
1113 		goto nla_put_failure;
1114 
1115 	switch (q->loss_model) {
1116 	case CLG_RANDOM:
1117 		/* legacy loss model */
1118 		nla_nest_cancel(skb, nest);
1119 		return 0;	/* no data */
1120 
1121 	case CLG_4_STATES: {
1122 		struct tc_netem_gimodel gi = {
1123 			.p13 = q->clg.a1,
1124 			.p31 = q->clg.a2,
1125 			.p32 = q->clg.a3,
1126 			.p14 = q->clg.a4,
1127 			.p23 = q->clg.a5,
1128 		};
1129 
1130 		if (nla_put(skb, NETEM_LOSS_GI, sizeof(gi), &gi))
1131 			goto nla_put_failure;
1132 		break;
1133 	}
1134 	case CLG_GILB_ELL: {
1135 		struct tc_netem_gemodel ge = {
1136 			.p = q->clg.a1,
1137 			.r = q->clg.a2,
1138 			.h = q->clg.a3,
1139 			.k1 = q->clg.a4,
1140 		};
1141 
1142 		if (nla_put(skb, NETEM_LOSS_GE, sizeof(ge), &ge))
1143 			goto nla_put_failure;
1144 		break;
1145 	}
1146 	}
1147 
1148 	nla_nest_end(skb, nest);
1149 	return 0;
1150 
1151 nla_put_failure:
1152 	nla_nest_cancel(skb, nest);
1153 	return -1;
1154 }
1155 
1156 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
1157 {
1158 	const struct netem_sched_data *q = qdisc_priv(sch);
1159 	struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb);
1160 	struct tc_netem_qopt qopt;
1161 	struct tc_netem_corr cor;
1162 	struct tc_netem_reorder reorder;
1163 	struct tc_netem_corrupt corrupt;
1164 	struct tc_netem_rate rate;
1165 	struct tc_netem_slot slot;
1166 
1167 	qopt.latency = min_t(psched_time_t, PSCHED_NS2TICKS(q->latency),
1168 			     UINT_MAX);
1169 	qopt.jitter = min_t(psched_time_t, PSCHED_NS2TICKS(q->jitter),
1170 			    UINT_MAX);
1171 	qopt.limit = q->limit;
1172 	qopt.loss = q->loss;
1173 	qopt.gap = q->gap;
1174 	qopt.duplicate = q->duplicate;
1175 	if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
1176 		goto nla_put_failure;
1177 
1178 	if (nla_put(skb, TCA_NETEM_LATENCY64, sizeof(q->latency), &q->latency))
1179 		goto nla_put_failure;
1180 
1181 	if (nla_put(skb, TCA_NETEM_JITTER64, sizeof(q->jitter), &q->jitter))
1182 		goto nla_put_failure;
1183 
1184 	cor.delay_corr = q->delay_cor.rho;
1185 	cor.loss_corr = q->loss_cor.rho;
1186 	cor.dup_corr = q->dup_cor.rho;
1187 	if (nla_put(skb, TCA_NETEM_CORR, sizeof(cor), &cor))
1188 		goto nla_put_failure;
1189 
1190 	reorder.probability = q->reorder;
1191 	reorder.correlation = q->reorder_cor.rho;
1192 	if (nla_put(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder))
1193 		goto nla_put_failure;
1194 
1195 	corrupt.probability = q->corrupt;
1196 	corrupt.correlation = q->corrupt_cor.rho;
1197 	if (nla_put(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt))
1198 		goto nla_put_failure;
1199 
1200 	if (q->rate >= (1ULL << 32)) {
1201 		if (nla_put_u64_64bit(skb, TCA_NETEM_RATE64, q->rate,
1202 				      TCA_NETEM_PAD))
1203 			goto nla_put_failure;
1204 		rate.rate = ~0U;
1205 	} else {
1206 		rate.rate = q->rate;
1207 	}
1208 	rate.packet_overhead = q->packet_overhead;
1209 	rate.cell_size = q->cell_size;
1210 	rate.cell_overhead = q->cell_overhead;
1211 	if (nla_put(skb, TCA_NETEM_RATE, sizeof(rate), &rate))
1212 		goto nla_put_failure;
1213 
1214 	if (q->ecn && nla_put_u32(skb, TCA_NETEM_ECN, q->ecn))
1215 		goto nla_put_failure;
1216 
1217 	if (dump_loss_model(q, skb) != 0)
1218 		goto nla_put_failure;
1219 
1220 	if (q->slot_config.min_delay | q->slot_config.max_delay |
1221 	    q->slot_config.dist_jitter) {
1222 		slot = q->slot_config;
1223 		if (slot.max_packets == INT_MAX)
1224 			slot.max_packets = 0;
1225 		if (slot.max_bytes == INT_MAX)
1226 			slot.max_bytes = 0;
1227 		if (nla_put(skb, TCA_NETEM_SLOT, sizeof(slot), &slot))
1228 			goto nla_put_failure;
1229 	}
1230 
1231 	if (nla_put_u64_64bit(skb, TCA_NETEM_PRNG_SEED, q->prng.seed,
1232 			      TCA_NETEM_PAD))
1233 		goto nla_put_failure;
1234 
1235 	return nla_nest_end(skb, nla);
1236 
1237 nla_put_failure:
1238 	nlmsg_trim(skb, nla);
1239 	return -1;
1240 }
1241 
1242 static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
1243 			  struct sk_buff *skb, struct tcmsg *tcm)
1244 {
1245 	struct netem_sched_data *q = qdisc_priv(sch);
1246 
1247 	if (cl != 1 || !q->qdisc) 	/* only one class */
1248 		return -ENOENT;
1249 
1250 	tcm->tcm_handle |= TC_H_MIN(1);
1251 	tcm->tcm_info = q->qdisc->handle;
1252 
1253 	return 0;
1254 }
1255 
1256 static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1257 		     struct Qdisc **old, struct netlink_ext_ack *extack)
1258 {
1259 	struct netem_sched_data *q = qdisc_priv(sch);
1260 
1261 	*old = qdisc_replace(sch, new, &q->qdisc);
1262 	return 0;
1263 }
1264 
1265 static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
1266 {
1267 	struct netem_sched_data *q = qdisc_priv(sch);
1268 	return q->qdisc;
1269 }
1270 
1271 static unsigned long netem_find(struct Qdisc *sch, u32 classid)
1272 {
1273 	return 1;
1274 }
1275 
1276 static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
1277 {
1278 	if (!walker->stop) {
1279 		if (!tc_qdisc_stats_dump(sch, 1, walker))
1280 			return;
1281 	}
1282 }
1283 
1284 static const struct Qdisc_class_ops netem_class_ops = {
1285 	.graft		=	netem_graft,
1286 	.leaf		=	netem_leaf,
1287 	.find		=	netem_find,
1288 	.walk		=	netem_walk,
1289 	.dump		=	netem_dump_class,
1290 };
1291 
1292 static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
1293 	.id		=	"netem",
1294 	.cl_ops		=	&netem_class_ops,
1295 	.priv_size	=	sizeof(struct netem_sched_data),
1296 	.enqueue	=	netem_enqueue,
1297 	.dequeue	=	netem_dequeue,
1298 	.peek		=	qdisc_peek_dequeued,
1299 	.init		=	netem_init,
1300 	.reset		=	netem_reset,
1301 	.destroy	=	netem_destroy,
1302 	.change		=	netem_change,
1303 	.dump		=	netem_dump,
1304 	.owner		=	THIS_MODULE,
1305 };
1306 
1307 
1308 static int __init netem_module_init(void)
1309 {
1310 	pr_info("netem: version " VERSION "\n");
1311 	return register_qdisc(&netem_qdisc_ops);
1312 }
1313 static void __exit netem_module_exit(void)
1314 {
1315 	unregister_qdisc(&netem_qdisc_ops);
1316 }
1317 module_init(netem_module_init)
1318 module_exit(netem_module_exit)
1319 MODULE_LICENSE("GPL");
1320