Loading...
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
67struct disttable {
68 u32 size;
69 s16 table[0];
70};
71
72struct 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 */
157struct netem_skb_cb {
158 u64 time_to_send;
159};
160
161static 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 */
171static 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 */
181static 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 */
200static 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 */
265static 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
286static 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 */
318static 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
345static 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
360static 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
378static 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 */
412static 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 */
434static 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
594finish_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
629static 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
648static 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
665static 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
676static struct sk_buff *netem_dequeue(struct Qdisc *sch)
677{
678 struct netem_sched_data *q = qdisc_priv(sch);
679 struct sk_buff *skb;
680
681tfifo_dequeue:
682 skb = __qdisc_dequeue_head(&sch->q);
683 if (skb) {
684 qdisc_qstats_backlog_dec(sch, skb);
685deliver:
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
755static 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
766static 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
776static 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
806static 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
824static 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
833static 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
841static 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
849static 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
863static 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
917static 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
930static 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 */
950static 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
1042get_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
1052static 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
1070static 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
1081static 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
1126nla_put_failure:
1127 nla_nest_cancel(skb, nest);
1128 return -1;
1129}
1130
1131static 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
1208nla_put_failure:
1209 nlmsg_trim(skb, nla);
1210 return -1;
1211}
1212
1213static 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
1227static 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
1236static 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
1242static unsigned long netem_find(struct Qdisc *sch, u32 classid)
1243{
1244 return 1;
1245}
1246
1247static 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
1259static 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
1267static 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
1283static int __init netem_module_init(void)
1284{
1285 pr_info("netem: version " VERSION "\n");
1286 return register_qdisc(&netem_qdisc_ops);
1287}
1288static void __exit netem_module_exit(void)
1289{
1290 unregister_qdisc(&netem_qdisc_ops);
1291}
1292module_init(netem_module_init)
1293module_exit(netem_module_exit)
1294MODULE_LICENSE("GPL");
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
26#include <net/netlink.h>
27#include <net/pkt_sched.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
68struct netem_sched_data {
69 struct Qdisc *qdisc;
70 struct qdisc_watchdog watchdog;
71
72 psched_tdiff_t latency;
73 psched_tdiff_t jitter;
74
75 u32 loss;
76 u32 limit;
77 u32 counter;
78 u32 gap;
79 u32 duplicate;
80 u32 reorder;
81 u32 corrupt;
82
83 struct crndstate {
84 u32 last;
85 u32 rho;
86 } delay_cor, loss_cor, dup_cor, reorder_cor, corrupt_cor;
87
88 struct disttable {
89 u32 size;
90 s16 table[0];
91 } *delay_dist;
92
93 enum {
94 CLG_RANDOM,
95 CLG_4_STATES,
96 CLG_GILB_ELL,
97 } loss_model;
98
99 /* Correlated Loss Generation models */
100 struct clgstate {
101 /* state of the Markov chain */
102 u8 state;
103
104 /* 4-states and Gilbert-Elliot models */
105 u32 a1; /* p13 for 4-states or p for GE */
106 u32 a2; /* p31 for 4-states or r for GE */
107 u32 a3; /* p32 for 4-states or h for GE */
108 u32 a4; /* p14 for 4-states or 1-k for GE */
109 u32 a5; /* p23 used only in 4-states */
110 } clg;
111
112};
113
114/* Time stamp put into socket buffer control block */
115struct netem_skb_cb {
116 psched_time_t time_to_send;
117};
118
119static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb)
120{
121 BUILD_BUG_ON(sizeof(skb->cb) <
122 sizeof(struct qdisc_skb_cb) + sizeof(struct netem_skb_cb));
123 return (struct netem_skb_cb *)qdisc_skb_cb(skb)->data;
124}
125
126/* init_crandom - initialize correlated random number generator
127 * Use entropy source for initial seed.
128 */
129static void init_crandom(struct crndstate *state, unsigned long rho)
130{
131 state->rho = rho;
132 state->last = net_random();
133}
134
135/* get_crandom - correlated random number generator
136 * Next number depends on last value.
137 * rho is scaled to avoid floating point.
138 */
139static u32 get_crandom(struct crndstate *state)
140{
141 u64 value, rho;
142 unsigned long answer;
143
144 if (state->rho == 0) /* no correlation */
145 return net_random();
146
147 value = net_random();
148 rho = (u64)state->rho + 1;
149 answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32;
150 state->last = answer;
151 return answer;
152}
153
154/* loss_4state - 4-state model loss generator
155 * Generates losses according to the 4-state Markov chain adopted in
156 * the GI (General and Intuitive) loss model.
157 */
158static bool loss_4state(struct netem_sched_data *q)
159{
160 struct clgstate *clg = &q->clg;
161 u32 rnd = net_random();
162
163 /*
164 * Makes a comparison between rnd and the transition
165 * probabilities outgoing from the current state, then decides the
166 * next state and if the next packet has to be transmitted or lost.
167 * The four states correspond to:
168 * 1 => successfully transmitted packets within a gap period
169 * 4 => isolated losses within a gap period
170 * 3 => lost packets within a burst period
171 * 2 => successfully transmitted packets within a burst period
172 */
173 switch (clg->state) {
174 case 1:
175 if (rnd < clg->a4) {
176 clg->state = 4;
177 return true;
178 } else if (clg->a4 < rnd && rnd < clg->a1) {
179 clg->state = 3;
180 return true;
181 } else if (clg->a1 < rnd)
182 clg->state = 1;
183
184 break;
185 case 2:
186 if (rnd < clg->a5) {
187 clg->state = 3;
188 return true;
189 } else
190 clg->state = 2;
191
192 break;
193 case 3:
194 if (rnd < clg->a3)
195 clg->state = 2;
196 else if (clg->a3 < rnd && rnd < clg->a2 + clg->a3) {
197 clg->state = 1;
198 return true;
199 } else if (clg->a2 + clg->a3 < rnd) {
200 clg->state = 3;
201 return true;
202 }
203 break;
204 case 4:
205 clg->state = 1;
206 break;
207 }
208
209 return false;
210}
211
212/* loss_gilb_ell - Gilbert-Elliot model loss generator
213 * Generates losses according to the Gilbert-Elliot loss model or
214 * its special cases (Gilbert or Simple Gilbert)
215 *
216 * Makes a comparison between random number and the transition
217 * probabilities outgoing from the current state, then decides the
218 * next state. A second random number is extracted and the comparison
219 * with the loss probability of the current state decides if the next
220 * packet will be transmitted or lost.
221 */
222static bool loss_gilb_ell(struct netem_sched_data *q)
223{
224 struct clgstate *clg = &q->clg;
225
226 switch (clg->state) {
227 case 1:
228 if (net_random() < clg->a1)
229 clg->state = 2;
230 if (net_random() < clg->a4)
231 return true;
232 case 2:
233 if (net_random() < clg->a2)
234 clg->state = 1;
235 if (clg->a3 > net_random())
236 return true;
237 }
238
239 return false;
240}
241
242static bool loss_event(struct netem_sched_data *q)
243{
244 switch (q->loss_model) {
245 case CLG_RANDOM:
246 /* Random packet drop 0 => none, ~0 => all */
247 return q->loss && q->loss >= get_crandom(&q->loss_cor);
248
249 case CLG_4_STATES:
250 /* 4state loss model algorithm (used also for GI model)
251 * Extracts a value from the markov 4 state loss generator,
252 * if it is 1 drops a packet and if needed writes the event in
253 * the kernel logs
254 */
255 return loss_4state(q);
256
257 case CLG_GILB_ELL:
258 /* Gilbert-Elliot loss model algorithm
259 * Extracts a value from the Gilbert-Elliot loss generator,
260 * if it is 1 drops a packet and if needed writes the event in
261 * the kernel logs
262 */
263 return loss_gilb_ell(q);
264 }
265
266 return false; /* not reached */
267}
268
269
270/* tabledist - return a pseudo-randomly distributed value with mean mu and
271 * std deviation sigma. Uses table lookup to approximate the desired
272 * distribution, and a uniformly-distributed pseudo-random source.
273 */
274static psched_tdiff_t tabledist(psched_tdiff_t mu, psched_tdiff_t sigma,
275 struct crndstate *state,
276 const struct disttable *dist)
277{
278 psched_tdiff_t x;
279 long t;
280 u32 rnd;
281
282 if (sigma == 0)
283 return mu;
284
285 rnd = get_crandom(state);
286
287 /* default uniform distribution */
288 if (dist == NULL)
289 return (rnd % (2*sigma)) - sigma + mu;
290
291 t = dist->table[rnd % dist->size];
292 x = (sigma % NETEM_DIST_SCALE) * t;
293 if (x >= 0)
294 x += NETEM_DIST_SCALE/2;
295 else
296 x -= NETEM_DIST_SCALE/2;
297
298 return x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu;
299}
300
301/*
302 * Insert one skb into qdisc.
303 * Note: parent depends on return value to account for queue length.
304 * NET_XMIT_DROP: queue length didn't change.
305 * NET_XMIT_SUCCESS: one skb was queued.
306 */
307static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch)
308{
309 struct netem_sched_data *q = qdisc_priv(sch);
310 /* We don't fill cb now as skb_unshare() may invalidate it */
311 struct netem_skb_cb *cb;
312 struct sk_buff *skb2;
313 int ret;
314 int count = 1;
315
316 /* Random duplication */
317 if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor))
318 ++count;
319
320 /* Drop packet? */
321 if (loss_event(q))
322 --count;
323
324 if (count == 0) {
325 sch->qstats.drops++;
326 kfree_skb(skb);
327 return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
328 }
329
330 skb_orphan(skb);
331
332 /*
333 * If we need to duplicate packet, then re-insert at top of the
334 * qdisc tree, since parent queuer expects that only one
335 * skb will be queued.
336 */
337 if (count > 1 && (skb2 = skb_clone(skb, GFP_ATOMIC)) != NULL) {
338 struct Qdisc *rootq = qdisc_root(sch);
339 u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
340 q->duplicate = 0;
341
342 qdisc_enqueue_root(skb2, rootq);
343 q->duplicate = dupsave;
344 }
345
346 /*
347 * Randomized packet corruption.
348 * Make copy if needed since we are modifying
349 * If packet is going to be hardware checksummed, then
350 * do it now in software before we mangle it.
351 */
352 if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor)) {
353 if (!(skb = skb_unshare(skb, GFP_ATOMIC)) ||
354 (skb->ip_summed == CHECKSUM_PARTIAL &&
355 skb_checksum_help(skb))) {
356 sch->qstats.drops++;
357 return NET_XMIT_DROP;
358 }
359
360 skb->data[net_random() % skb_headlen(skb)] ^= 1<<(net_random() % 8);
361 }
362
363 cb = netem_skb_cb(skb);
364 if (q->gap == 0 || /* not doing reordering */
365 q->counter < q->gap || /* inside last reordering gap */
366 q->reorder < get_crandom(&q->reorder_cor)) {
367 psched_time_t now;
368 psched_tdiff_t delay;
369
370 delay = tabledist(q->latency, q->jitter,
371 &q->delay_cor, q->delay_dist);
372
373 now = psched_get_time();
374 cb->time_to_send = now + delay;
375 ++q->counter;
376 ret = qdisc_enqueue(skb, q->qdisc);
377 } else {
378 /*
379 * Do re-ordering by putting one out of N packets at the front
380 * of the queue.
381 */
382 cb->time_to_send = psched_get_time();
383 q->counter = 0;
384
385 __skb_queue_head(&q->qdisc->q, skb);
386 q->qdisc->qstats.backlog += qdisc_pkt_len(skb);
387 q->qdisc->qstats.requeues++;
388 ret = NET_XMIT_SUCCESS;
389 }
390
391 if (ret != NET_XMIT_SUCCESS) {
392 if (net_xmit_drop_count(ret)) {
393 sch->qstats.drops++;
394 return ret;
395 }
396 }
397
398 sch->q.qlen++;
399 return NET_XMIT_SUCCESS;
400}
401
402static unsigned int netem_drop(struct Qdisc *sch)
403{
404 struct netem_sched_data *q = qdisc_priv(sch);
405 unsigned int len = 0;
406
407 if (q->qdisc->ops->drop && (len = q->qdisc->ops->drop(q->qdisc)) != 0) {
408 sch->q.qlen--;
409 sch->qstats.drops++;
410 }
411 return len;
412}
413
414static struct sk_buff *netem_dequeue(struct Qdisc *sch)
415{
416 struct netem_sched_data *q = qdisc_priv(sch);
417 struct sk_buff *skb;
418
419 if (qdisc_is_throttled(sch))
420 return NULL;
421
422 skb = q->qdisc->ops->peek(q->qdisc);
423 if (skb) {
424 const struct netem_skb_cb *cb = netem_skb_cb(skb);
425 psched_time_t now = psched_get_time();
426
427 /* if more time remaining? */
428 if (cb->time_to_send <= now) {
429 skb = qdisc_dequeue_peeked(q->qdisc);
430 if (unlikely(!skb))
431 return NULL;
432
433#ifdef CONFIG_NET_CLS_ACT
434 /*
435 * If it's at ingress let's pretend the delay is
436 * from the network (tstamp will be updated).
437 */
438 if (G_TC_FROM(skb->tc_verd) & AT_INGRESS)
439 skb->tstamp.tv64 = 0;
440#endif
441
442 sch->q.qlen--;
443 qdisc_unthrottled(sch);
444 qdisc_bstats_update(sch, skb);
445 return skb;
446 }
447
448 qdisc_watchdog_schedule(&q->watchdog, cb->time_to_send);
449 }
450
451 return NULL;
452}
453
454static void netem_reset(struct Qdisc *sch)
455{
456 struct netem_sched_data *q = qdisc_priv(sch);
457
458 qdisc_reset(q->qdisc);
459 sch->q.qlen = 0;
460 qdisc_watchdog_cancel(&q->watchdog);
461}
462
463static void dist_free(struct disttable *d)
464{
465 if (d) {
466 if (is_vmalloc_addr(d))
467 vfree(d);
468 else
469 kfree(d);
470 }
471}
472
473/*
474 * Distribution data is a variable size payload containing
475 * signed 16 bit values.
476 */
477static int get_dist_table(struct Qdisc *sch, const struct nlattr *attr)
478{
479 struct netem_sched_data *q = qdisc_priv(sch);
480 size_t n = nla_len(attr)/sizeof(__s16);
481 const __s16 *data = nla_data(attr);
482 spinlock_t *root_lock;
483 struct disttable *d;
484 int i;
485 size_t s;
486
487 if (n > NETEM_DIST_MAX)
488 return -EINVAL;
489
490 s = sizeof(struct disttable) + n * sizeof(s16);
491 d = kmalloc(s, GFP_KERNEL);
492 if (!d)
493 d = vmalloc(s);
494 if (!d)
495 return -ENOMEM;
496
497 d->size = n;
498 for (i = 0; i < n; i++)
499 d->table[i] = data[i];
500
501 root_lock = qdisc_root_sleeping_lock(sch);
502
503 spin_lock_bh(root_lock);
504 dist_free(q->delay_dist);
505 q->delay_dist = d;
506 spin_unlock_bh(root_lock);
507 return 0;
508}
509
510static void get_correlation(struct Qdisc *sch, const struct nlattr *attr)
511{
512 struct netem_sched_data *q = qdisc_priv(sch);
513 const struct tc_netem_corr *c = nla_data(attr);
514
515 init_crandom(&q->delay_cor, c->delay_corr);
516 init_crandom(&q->loss_cor, c->loss_corr);
517 init_crandom(&q->dup_cor, c->dup_corr);
518}
519
520static void get_reorder(struct Qdisc *sch, const struct nlattr *attr)
521{
522 struct netem_sched_data *q = qdisc_priv(sch);
523 const struct tc_netem_reorder *r = nla_data(attr);
524
525 q->reorder = r->probability;
526 init_crandom(&q->reorder_cor, r->correlation);
527}
528
529static void get_corrupt(struct Qdisc *sch, const struct nlattr *attr)
530{
531 struct netem_sched_data *q = qdisc_priv(sch);
532 const struct tc_netem_corrupt *r = nla_data(attr);
533
534 q->corrupt = r->probability;
535 init_crandom(&q->corrupt_cor, r->correlation);
536}
537
538static int get_loss_clg(struct Qdisc *sch, const struct nlattr *attr)
539{
540 struct netem_sched_data *q = qdisc_priv(sch);
541 const struct nlattr *la;
542 int rem;
543
544 nla_for_each_nested(la, attr, rem) {
545 u16 type = nla_type(la);
546
547 switch(type) {
548 case NETEM_LOSS_GI: {
549 const struct tc_netem_gimodel *gi = nla_data(la);
550
551 if (nla_len(la) != sizeof(struct tc_netem_gimodel)) {
552 pr_info("netem: incorrect gi model size\n");
553 return -EINVAL;
554 }
555
556 q->loss_model = CLG_4_STATES;
557
558 q->clg.state = 1;
559 q->clg.a1 = gi->p13;
560 q->clg.a2 = gi->p31;
561 q->clg.a3 = gi->p32;
562 q->clg.a4 = gi->p14;
563 q->clg.a5 = gi->p23;
564 break;
565 }
566
567 case NETEM_LOSS_GE: {
568 const struct tc_netem_gemodel *ge = nla_data(la);
569
570 if (nla_len(la) != sizeof(struct tc_netem_gemodel)) {
571 pr_info("netem: incorrect gi model size\n");
572 return -EINVAL;
573 }
574
575 q->loss_model = CLG_GILB_ELL;
576 q->clg.state = 1;
577 q->clg.a1 = ge->p;
578 q->clg.a2 = ge->r;
579 q->clg.a3 = ge->h;
580 q->clg.a4 = ge->k1;
581 break;
582 }
583
584 default:
585 pr_info("netem: unknown loss type %u\n", type);
586 return -EINVAL;
587 }
588 }
589
590 return 0;
591}
592
593static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
594 [TCA_NETEM_CORR] = { .len = sizeof(struct tc_netem_corr) },
595 [TCA_NETEM_REORDER] = { .len = sizeof(struct tc_netem_reorder) },
596 [TCA_NETEM_CORRUPT] = { .len = sizeof(struct tc_netem_corrupt) },
597 [TCA_NETEM_LOSS] = { .type = NLA_NESTED },
598};
599
600static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
601 const struct nla_policy *policy, int len)
602{
603 int nested_len = nla_len(nla) - NLA_ALIGN(len);
604
605 if (nested_len < 0) {
606 pr_info("netem: invalid attributes len %d\n", nested_len);
607 return -EINVAL;
608 }
609
610 if (nested_len >= nla_attr_size(0))
611 return nla_parse(tb, maxtype, nla_data(nla) + NLA_ALIGN(len),
612 nested_len, policy);
613
614 memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
615 return 0;
616}
617
618/* Parse netlink message to set options */
619static int netem_change(struct Qdisc *sch, struct nlattr *opt)
620{
621 struct netem_sched_data *q = qdisc_priv(sch);
622 struct nlattr *tb[TCA_NETEM_MAX + 1];
623 struct tc_netem_qopt *qopt;
624 int ret;
625
626 if (opt == NULL)
627 return -EINVAL;
628
629 qopt = nla_data(opt);
630 ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
631 if (ret < 0)
632 return ret;
633
634 ret = fifo_set_limit(q->qdisc, qopt->limit);
635 if (ret) {
636 pr_info("netem: can't set fifo limit\n");
637 return ret;
638 }
639
640 q->latency = qopt->latency;
641 q->jitter = qopt->jitter;
642 q->limit = qopt->limit;
643 q->gap = qopt->gap;
644 q->counter = 0;
645 q->loss = qopt->loss;
646 q->duplicate = qopt->duplicate;
647
648 /* for compatibility with earlier versions.
649 * if gap is set, need to assume 100% probability
650 */
651 if (q->gap)
652 q->reorder = ~0;
653
654 if (tb[TCA_NETEM_CORR])
655 get_correlation(sch, tb[TCA_NETEM_CORR]);
656
657 if (tb[TCA_NETEM_DELAY_DIST]) {
658 ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST]);
659 if (ret)
660 return ret;
661 }
662
663 if (tb[TCA_NETEM_REORDER])
664 get_reorder(sch, tb[TCA_NETEM_REORDER]);
665
666 if (tb[TCA_NETEM_CORRUPT])
667 get_corrupt(sch, tb[TCA_NETEM_CORRUPT]);
668
669 q->loss_model = CLG_RANDOM;
670 if (tb[TCA_NETEM_LOSS])
671 ret = get_loss_clg(sch, tb[TCA_NETEM_LOSS]);
672
673 return ret;
674}
675
676/*
677 * Special case version of FIFO queue for use by netem.
678 * It queues in order based on timestamps in skb's
679 */
680struct fifo_sched_data {
681 u32 limit;
682 psched_time_t oldest;
683};
684
685static int tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
686{
687 struct fifo_sched_data *q = qdisc_priv(sch);
688 struct sk_buff_head *list = &sch->q;
689 psched_time_t tnext = netem_skb_cb(nskb)->time_to_send;
690 struct sk_buff *skb;
691
692 if (likely(skb_queue_len(list) < q->limit)) {
693 /* Optimize for add at tail */
694 if (likely(skb_queue_empty(list) || tnext >= q->oldest)) {
695 q->oldest = tnext;
696 return qdisc_enqueue_tail(nskb, sch);
697 }
698
699 skb_queue_reverse_walk(list, skb) {
700 const struct netem_skb_cb *cb = netem_skb_cb(skb);
701
702 if (tnext >= cb->time_to_send)
703 break;
704 }
705
706 __skb_queue_after(list, skb, nskb);
707
708 sch->qstats.backlog += qdisc_pkt_len(nskb);
709
710 return NET_XMIT_SUCCESS;
711 }
712
713 return qdisc_reshape_fail(nskb, sch);
714}
715
716static int tfifo_init(struct Qdisc *sch, struct nlattr *opt)
717{
718 struct fifo_sched_data *q = qdisc_priv(sch);
719
720 if (opt) {
721 struct tc_fifo_qopt *ctl = nla_data(opt);
722 if (nla_len(opt) < sizeof(*ctl))
723 return -EINVAL;
724
725 q->limit = ctl->limit;
726 } else
727 q->limit = max_t(u32, qdisc_dev(sch)->tx_queue_len, 1);
728
729 q->oldest = PSCHED_PASTPERFECT;
730 return 0;
731}
732
733static int tfifo_dump(struct Qdisc *sch, struct sk_buff *skb)
734{
735 struct fifo_sched_data *q = qdisc_priv(sch);
736 struct tc_fifo_qopt opt = { .limit = q->limit };
737
738 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
739 return skb->len;
740
741nla_put_failure:
742 return -1;
743}
744
745static struct Qdisc_ops tfifo_qdisc_ops __read_mostly = {
746 .id = "tfifo",
747 .priv_size = sizeof(struct fifo_sched_data),
748 .enqueue = tfifo_enqueue,
749 .dequeue = qdisc_dequeue_head,
750 .peek = qdisc_peek_head,
751 .drop = qdisc_queue_drop,
752 .init = tfifo_init,
753 .reset = qdisc_reset_queue,
754 .change = tfifo_init,
755 .dump = tfifo_dump,
756};
757
758static int netem_init(struct Qdisc *sch, struct nlattr *opt)
759{
760 struct netem_sched_data *q = qdisc_priv(sch);
761 int ret;
762
763 if (!opt)
764 return -EINVAL;
765
766 qdisc_watchdog_init(&q->watchdog, sch);
767
768 q->loss_model = CLG_RANDOM;
769 q->qdisc = qdisc_create_dflt(sch->dev_queue, &tfifo_qdisc_ops,
770 TC_H_MAKE(sch->handle, 1));
771 if (!q->qdisc) {
772 pr_notice("netem: qdisc create tfifo qdisc failed\n");
773 return -ENOMEM;
774 }
775
776 ret = netem_change(sch, opt);
777 if (ret) {
778 pr_info("netem: change failed\n");
779 qdisc_destroy(q->qdisc);
780 }
781 return ret;
782}
783
784static void netem_destroy(struct Qdisc *sch)
785{
786 struct netem_sched_data *q = qdisc_priv(sch);
787
788 qdisc_watchdog_cancel(&q->watchdog);
789 qdisc_destroy(q->qdisc);
790 dist_free(q->delay_dist);
791}
792
793static int dump_loss_model(const struct netem_sched_data *q,
794 struct sk_buff *skb)
795{
796 struct nlattr *nest;
797
798 nest = nla_nest_start(skb, TCA_NETEM_LOSS);
799 if (nest == NULL)
800 goto nla_put_failure;
801
802 switch (q->loss_model) {
803 case CLG_RANDOM:
804 /* legacy loss model */
805 nla_nest_cancel(skb, nest);
806 return 0; /* no data */
807
808 case CLG_4_STATES: {
809 struct tc_netem_gimodel gi = {
810 .p13 = q->clg.a1,
811 .p31 = q->clg.a2,
812 .p32 = q->clg.a3,
813 .p14 = q->clg.a4,
814 .p23 = q->clg.a5,
815 };
816
817 NLA_PUT(skb, NETEM_LOSS_GI, sizeof(gi), &gi);
818 break;
819 }
820 case CLG_GILB_ELL: {
821 struct tc_netem_gemodel ge = {
822 .p = q->clg.a1,
823 .r = q->clg.a2,
824 .h = q->clg.a3,
825 .k1 = q->clg.a4,
826 };
827
828 NLA_PUT(skb, NETEM_LOSS_GE, sizeof(ge), &ge);
829 break;
830 }
831 }
832
833 nla_nest_end(skb, nest);
834 return 0;
835
836nla_put_failure:
837 nla_nest_cancel(skb, nest);
838 return -1;
839}
840
841static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
842{
843 const struct netem_sched_data *q = qdisc_priv(sch);
844 struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb);
845 struct tc_netem_qopt qopt;
846 struct tc_netem_corr cor;
847 struct tc_netem_reorder reorder;
848 struct tc_netem_corrupt corrupt;
849
850 qopt.latency = q->latency;
851 qopt.jitter = q->jitter;
852 qopt.limit = q->limit;
853 qopt.loss = q->loss;
854 qopt.gap = q->gap;
855 qopt.duplicate = q->duplicate;
856 NLA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
857
858 cor.delay_corr = q->delay_cor.rho;
859 cor.loss_corr = q->loss_cor.rho;
860 cor.dup_corr = q->dup_cor.rho;
861 NLA_PUT(skb, TCA_NETEM_CORR, sizeof(cor), &cor);
862
863 reorder.probability = q->reorder;
864 reorder.correlation = q->reorder_cor.rho;
865 NLA_PUT(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder);
866
867 corrupt.probability = q->corrupt;
868 corrupt.correlation = q->corrupt_cor.rho;
869 NLA_PUT(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt);
870
871 if (dump_loss_model(q, skb) != 0)
872 goto nla_put_failure;
873
874 return nla_nest_end(skb, nla);
875
876nla_put_failure:
877 nlmsg_trim(skb, nla);
878 return -1;
879}
880
881static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
882 struct sk_buff *skb, struct tcmsg *tcm)
883{
884 struct netem_sched_data *q = qdisc_priv(sch);
885
886 if (cl != 1) /* only one class */
887 return -ENOENT;
888
889 tcm->tcm_handle |= TC_H_MIN(1);
890 tcm->tcm_info = q->qdisc->handle;
891
892 return 0;
893}
894
895static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
896 struct Qdisc **old)
897{
898 struct netem_sched_data *q = qdisc_priv(sch);
899
900 if (new == NULL)
901 new = &noop_qdisc;
902
903 sch_tree_lock(sch);
904 *old = q->qdisc;
905 q->qdisc = new;
906 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
907 qdisc_reset(*old);
908 sch_tree_unlock(sch);
909
910 return 0;
911}
912
913static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
914{
915 struct netem_sched_data *q = qdisc_priv(sch);
916 return q->qdisc;
917}
918
919static unsigned long netem_get(struct Qdisc *sch, u32 classid)
920{
921 return 1;
922}
923
924static void netem_put(struct Qdisc *sch, unsigned long arg)
925{
926}
927
928static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
929{
930 if (!walker->stop) {
931 if (walker->count >= walker->skip)
932 if (walker->fn(sch, 1, walker) < 0) {
933 walker->stop = 1;
934 return;
935 }
936 walker->count++;
937 }
938}
939
940static const struct Qdisc_class_ops netem_class_ops = {
941 .graft = netem_graft,
942 .leaf = netem_leaf,
943 .get = netem_get,
944 .put = netem_put,
945 .walk = netem_walk,
946 .dump = netem_dump_class,
947};
948
949static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
950 .id = "netem",
951 .cl_ops = &netem_class_ops,
952 .priv_size = sizeof(struct netem_sched_data),
953 .enqueue = netem_enqueue,
954 .dequeue = netem_dequeue,
955 .peek = qdisc_peek_dequeued,
956 .drop = netem_drop,
957 .init = netem_init,
958 .reset = netem_reset,
959 .destroy = netem_destroy,
960 .change = netem_change,
961 .dump = netem_dump,
962 .owner = THIS_MODULE,
963};
964
965
966static int __init netem_module_init(void)
967{
968 pr_info("netem: version " VERSION "\n");
969 return register_qdisc(&netem_qdisc_ops);
970}
971static void __exit netem_module_exit(void)
972{
973 unregister_qdisc(&netem_qdisc_ops);
974}
975module_init(netem_module_init)
976module_exit(netem_module_exit)
977MODULE_LICENSE("GPL");