Loading...
1/*
2 * net/sched/sch_netem.c Network emulator
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License.
8 *
9 * Many of the algorithms and ideas for this came from
10 * NIST Net which is not copyrighted.
11 *
12 * Authors: Stephen Hemminger <shemminger@osdl.org>
13 * Catalin(ux aka Dino) BOIE <catab at umbrella dot ro>
14 */
15
16#include <linux/mm.h>
17#include <linux/module.h>
18#include <linux/slab.h>
19#include <linux/types.h>
20#include <linux/kernel.h>
21#include <linux/errno.h>
22#include <linux/skbuff.h>
23#include <linux/vmalloc.h>
24#include <linux/rtnetlink.h>
25#include <linux/reciprocal_div.h>
26
27#include <net/netlink.h>
28#include <net/pkt_sched.h>
29#include <net/inet_ecn.h>
30
31#define VERSION "1.3"
32
33/* Network Emulation Queuing algorithm.
34 ====================================
35
36 Sources: [1] Mark Carson, Darrin Santay, "NIST Net - A Linux-based
37 Network Emulation Tool
38 [2] Luigi Rizzo, DummyNet for FreeBSD
39
40 ----------------------------------------------------------------
41
42 This started out as a simple way to delay outgoing packets to
43 test TCP but has grown to include most of the functionality
44 of a full blown network emulator like NISTnet. It can delay
45 packets and add random jitter (and correlation). The random
46 distribution can be loaded from a table as well to provide
47 normal, Pareto, or experimental curves. Packet loss,
48 duplication, and reordering can also be emulated.
49
50 This qdisc does not do classification that can be handled in
51 layering other disciplines. It does not need to do bandwidth
52 control either since that can be handled by using token
53 bucket or other rate control.
54
55 Correlated Loss Generator models
56
57 Added generation of correlated loss according to the
58 "Gilbert-Elliot" model, a 4-state markov model.
59
60 References:
61 [1] NetemCLG Home http://netgroup.uniroma2.it/NetemCLG
62 [2] S. Salsano, F. Ludovici, A. Ordine, "Definition of a general
63 and intuitive loss model for packet networks and its implementation
64 in the Netem module in the Linux kernel", available in [1]
65
66 Authors: Stefano Salsano <stefano.salsano at uniroma2.it
67 Fabio Ludovici <fabio.ludovici at yahoo.it>
68*/
69
70struct netem_sched_data {
71 /* internal t(ime)fifo qdisc uses sch->q and sch->limit */
72
73 /* optional qdisc for classful handling (NULL at netem init) */
74 struct Qdisc *qdisc;
75
76 struct qdisc_watchdog watchdog;
77
78 psched_tdiff_t latency;
79 psched_tdiff_t jitter;
80
81 u32 loss;
82 u32 ecn;
83 u32 limit;
84 u32 counter;
85 u32 gap;
86 u32 duplicate;
87 u32 reorder;
88 u32 corrupt;
89 u32 rate;
90 s32 packet_overhead;
91 u32 cell_size;
92 u32 cell_size_reciprocal;
93 s32 cell_overhead;
94
95 struct crndstate {
96 u32 last;
97 u32 rho;
98 } delay_cor, loss_cor, dup_cor, reorder_cor, corrupt_cor;
99
100 struct disttable {
101 u32 size;
102 s16 table[0];
103 } *delay_dist;
104
105 enum {
106 CLG_RANDOM,
107 CLG_4_STATES,
108 CLG_GILB_ELL,
109 } loss_model;
110
111 /* Correlated Loss Generation models */
112 struct clgstate {
113 /* state of the Markov chain */
114 u8 state;
115
116 /* 4-states and Gilbert-Elliot models */
117 u32 a1; /* p13 for 4-states or p for GE */
118 u32 a2; /* p31 for 4-states or r for GE */
119 u32 a3; /* p32 for 4-states or h for GE */
120 u32 a4; /* p14 for 4-states or 1-k for GE */
121 u32 a5; /* p23 used only in 4-states */
122 } clg;
123
124};
125
126/* Time stamp put into socket buffer control block
127 * Only valid when skbs are in our internal t(ime)fifo queue.
128 */
129struct netem_skb_cb {
130 psched_time_t time_to_send;
131};
132
133static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb)
134{
135 qdisc_cb_private_validate(skb, sizeof(struct netem_skb_cb));
136 return (struct netem_skb_cb *)qdisc_skb_cb(skb)->data;
137}
138
139/* init_crandom - initialize correlated random number generator
140 * Use entropy source for initial seed.
141 */
142static void init_crandom(struct crndstate *state, unsigned long rho)
143{
144 state->rho = rho;
145 state->last = net_random();
146}
147
148/* get_crandom - correlated random number generator
149 * Next number depends on last value.
150 * rho is scaled to avoid floating point.
151 */
152static u32 get_crandom(struct crndstate *state)
153{
154 u64 value, rho;
155 unsigned long answer;
156
157 if (state->rho == 0) /* no correlation */
158 return net_random();
159
160 value = net_random();
161 rho = (u64)state->rho + 1;
162 answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32;
163 state->last = answer;
164 return answer;
165}
166
167/* loss_4state - 4-state model loss generator
168 * Generates losses according to the 4-state Markov chain adopted in
169 * the GI (General and Intuitive) loss model.
170 */
171static bool loss_4state(struct netem_sched_data *q)
172{
173 struct clgstate *clg = &q->clg;
174 u32 rnd = net_random();
175
176 /*
177 * Makes a comparison between rnd and the transition
178 * probabilities outgoing from the current state, then decides the
179 * next state and if the next packet has to be transmitted or lost.
180 * The four states correspond to:
181 * 1 => successfully transmitted packets within a gap period
182 * 4 => isolated losses within a gap period
183 * 3 => lost packets within a burst period
184 * 2 => successfully transmitted packets within a burst period
185 */
186 switch (clg->state) {
187 case 1:
188 if (rnd < clg->a4) {
189 clg->state = 4;
190 return true;
191 } else if (clg->a4 < rnd && rnd < clg->a1) {
192 clg->state = 3;
193 return true;
194 } else if (clg->a1 < rnd)
195 clg->state = 1;
196
197 break;
198 case 2:
199 if (rnd < clg->a5) {
200 clg->state = 3;
201 return true;
202 } else
203 clg->state = 2;
204
205 break;
206 case 3:
207 if (rnd < clg->a3)
208 clg->state = 2;
209 else if (clg->a3 < rnd && rnd < clg->a2 + clg->a3) {
210 clg->state = 1;
211 return true;
212 } else if (clg->a2 + clg->a3 < rnd) {
213 clg->state = 3;
214 return true;
215 }
216 break;
217 case 4:
218 clg->state = 1;
219 break;
220 }
221
222 return false;
223}
224
225/* loss_gilb_ell - Gilbert-Elliot model loss generator
226 * Generates losses according to the Gilbert-Elliot loss model or
227 * its special cases (Gilbert or Simple Gilbert)
228 *
229 * Makes a comparison between random number and the transition
230 * probabilities outgoing from the current state, then decides the
231 * next state. A second random number is extracted and the comparison
232 * with the loss probability of the current state decides if the next
233 * packet will be transmitted or lost.
234 */
235static bool loss_gilb_ell(struct netem_sched_data *q)
236{
237 struct clgstate *clg = &q->clg;
238
239 switch (clg->state) {
240 case 1:
241 if (net_random() < clg->a1)
242 clg->state = 2;
243 if (net_random() < clg->a4)
244 return true;
245 case 2:
246 if (net_random() < clg->a2)
247 clg->state = 1;
248 if (clg->a3 > net_random())
249 return true;
250 }
251
252 return false;
253}
254
255static bool loss_event(struct netem_sched_data *q)
256{
257 switch (q->loss_model) {
258 case CLG_RANDOM:
259 /* Random packet drop 0 => none, ~0 => all */
260 return q->loss && q->loss >= get_crandom(&q->loss_cor);
261
262 case CLG_4_STATES:
263 /* 4state loss model algorithm (used also for GI model)
264 * Extracts a value from the markov 4 state loss generator,
265 * if it is 1 drops a packet and if needed writes the event in
266 * the kernel logs
267 */
268 return loss_4state(q);
269
270 case CLG_GILB_ELL:
271 /* Gilbert-Elliot loss model algorithm
272 * Extracts a value from the Gilbert-Elliot loss generator,
273 * if it is 1 drops a packet and if needed writes the event in
274 * the kernel logs
275 */
276 return loss_gilb_ell(q);
277 }
278
279 return false; /* not reached */
280}
281
282
283/* tabledist - return a pseudo-randomly distributed value with mean mu and
284 * std deviation sigma. Uses table lookup to approximate the desired
285 * distribution, and a uniformly-distributed pseudo-random source.
286 */
287static psched_tdiff_t tabledist(psched_tdiff_t mu, psched_tdiff_t sigma,
288 struct crndstate *state,
289 const struct disttable *dist)
290{
291 psched_tdiff_t x;
292 long t;
293 u32 rnd;
294
295 if (sigma == 0)
296 return mu;
297
298 rnd = get_crandom(state);
299
300 /* default uniform distribution */
301 if (dist == NULL)
302 return (rnd % (2*sigma)) - sigma + mu;
303
304 t = dist->table[rnd % dist->size];
305 x = (sigma % NETEM_DIST_SCALE) * t;
306 if (x >= 0)
307 x += NETEM_DIST_SCALE/2;
308 else
309 x -= NETEM_DIST_SCALE/2;
310
311 return x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu;
312}
313
314static psched_time_t packet_len_2_sched_time(unsigned int len, struct netem_sched_data *q)
315{
316 u64 ticks;
317
318 len += q->packet_overhead;
319
320 if (q->cell_size) {
321 u32 cells = reciprocal_divide(len, q->cell_size_reciprocal);
322
323 if (len > cells * q->cell_size) /* extra cell needed for remainder */
324 cells++;
325 len = cells * (q->cell_size + q->cell_overhead);
326 }
327
328 ticks = (u64)len * NSEC_PER_SEC;
329
330 do_div(ticks, q->rate);
331 return PSCHED_NS2TICKS(ticks);
332}
333
334static void tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
335{
336 struct sk_buff_head *list = &sch->q;
337 psched_time_t tnext = netem_skb_cb(nskb)->time_to_send;
338 struct sk_buff *skb = skb_peek_tail(list);
339
340 /* Optimize for add at tail */
341 if (likely(!skb || tnext >= netem_skb_cb(skb)->time_to_send))
342 return __skb_queue_tail(list, nskb);
343
344 skb_queue_reverse_walk(list, skb) {
345 if (tnext >= netem_skb_cb(skb)->time_to_send)
346 break;
347 }
348
349 __skb_queue_after(list, skb, nskb);
350}
351
352/*
353 * Insert one skb into qdisc.
354 * Note: parent depends on return value to account for queue length.
355 * NET_XMIT_DROP: queue length didn't change.
356 * NET_XMIT_SUCCESS: one skb was queued.
357 */
358static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch)
359{
360 struct netem_sched_data *q = qdisc_priv(sch);
361 /* We don't fill cb now as skb_unshare() may invalidate it */
362 struct netem_skb_cb *cb;
363 struct sk_buff *skb2;
364 int count = 1;
365
366 /* Random duplication */
367 if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor))
368 ++count;
369
370 /* Drop packet? */
371 if (loss_event(q)) {
372 if (q->ecn && INET_ECN_set_ce(skb))
373 sch->qstats.drops++; /* mark packet */
374 else
375 --count;
376 }
377 if (count == 0) {
378 sch->qstats.drops++;
379 kfree_skb(skb);
380 return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
381 }
382
383 skb_orphan(skb);
384
385 /*
386 * If we need to duplicate packet, then re-insert at top of the
387 * qdisc tree, since parent queuer expects that only one
388 * skb will be queued.
389 */
390 if (count > 1 && (skb2 = skb_clone(skb, GFP_ATOMIC)) != NULL) {
391 struct Qdisc *rootq = qdisc_root(sch);
392 u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
393 q->duplicate = 0;
394
395 qdisc_enqueue_root(skb2, rootq);
396 q->duplicate = dupsave;
397 }
398
399 /*
400 * Randomized packet corruption.
401 * Make copy if needed since we are modifying
402 * If packet is going to be hardware checksummed, then
403 * do it now in software before we mangle it.
404 */
405 if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor)) {
406 if (!(skb = skb_unshare(skb, GFP_ATOMIC)) ||
407 (skb->ip_summed == CHECKSUM_PARTIAL &&
408 skb_checksum_help(skb)))
409 return qdisc_drop(skb, sch);
410
411 skb->data[net_random() % skb_headlen(skb)] ^= 1<<(net_random() % 8);
412 }
413
414 if (unlikely(skb_queue_len(&sch->q) >= sch->limit))
415 return qdisc_reshape_fail(skb, sch);
416
417 sch->qstats.backlog += qdisc_pkt_len(skb);
418
419 cb = netem_skb_cb(skb);
420 if (q->gap == 0 || /* not doing reordering */
421 q->counter < q->gap - 1 || /* inside last reordering gap */
422 q->reorder < get_crandom(&q->reorder_cor)) {
423 psched_time_t now;
424 psched_tdiff_t delay;
425
426 delay = tabledist(q->latency, q->jitter,
427 &q->delay_cor, q->delay_dist);
428
429 now = psched_get_time();
430
431 if (q->rate) {
432 struct sk_buff_head *list = &sch->q;
433
434 delay += packet_len_2_sched_time(skb->len, q);
435
436 if (!skb_queue_empty(list)) {
437 /*
438 * Last packet in queue is reference point (now).
439 * First packet in queue is already in flight,
440 * calculate this time bonus and substract
441 * from delay.
442 */
443 delay -= now - netem_skb_cb(skb_peek(list))->time_to_send;
444 now = netem_skb_cb(skb_peek_tail(list))->time_to_send;
445 }
446 }
447
448 cb->time_to_send = now + delay;
449 ++q->counter;
450 tfifo_enqueue(skb, sch);
451 } else {
452 /*
453 * Do re-ordering by putting one out of N packets at the front
454 * of the queue.
455 */
456 cb->time_to_send = psched_get_time();
457 q->counter = 0;
458
459 __skb_queue_head(&sch->q, skb);
460 sch->qstats.requeues++;
461 }
462
463 return NET_XMIT_SUCCESS;
464}
465
466static unsigned int netem_drop(struct Qdisc *sch)
467{
468 struct netem_sched_data *q = qdisc_priv(sch);
469 unsigned int len;
470
471 len = qdisc_queue_drop(sch);
472 if (!len && q->qdisc && q->qdisc->ops->drop)
473 len = q->qdisc->ops->drop(q->qdisc);
474 if (len)
475 sch->qstats.drops++;
476
477 return len;
478}
479
480static struct sk_buff *netem_dequeue(struct Qdisc *sch)
481{
482 struct netem_sched_data *q = qdisc_priv(sch);
483 struct sk_buff *skb;
484
485 if (qdisc_is_throttled(sch))
486 return NULL;
487
488tfifo_dequeue:
489 skb = qdisc_peek_head(sch);
490 if (skb) {
491 const struct netem_skb_cb *cb = netem_skb_cb(skb);
492
493 /* if more time remaining? */
494 if (cb->time_to_send <= psched_get_time()) {
495 __skb_unlink(skb, &sch->q);
496 sch->qstats.backlog -= qdisc_pkt_len(skb);
497
498#ifdef CONFIG_NET_CLS_ACT
499 /*
500 * If it's at ingress let's pretend the delay is
501 * from the network (tstamp will be updated).
502 */
503 if (G_TC_FROM(skb->tc_verd) & AT_INGRESS)
504 skb->tstamp.tv64 = 0;
505#endif
506
507 if (q->qdisc) {
508 int err = qdisc_enqueue(skb, q->qdisc);
509
510 if (unlikely(err != NET_XMIT_SUCCESS)) {
511 if (net_xmit_drop_count(err)) {
512 sch->qstats.drops++;
513 qdisc_tree_decrease_qlen(sch, 1);
514 }
515 }
516 goto tfifo_dequeue;
517 }
518deliver:
519 qdisc_unthrottled(sch);
520 qdisc_bstats_update(sch, skb);
521 return skb;
522 }
523
524 if (q->qdisc) {
525 skb = q->qdisc->ops->dequeue(q->qdisc);
526 if (skb)
527 goto deliver;
528 }
529 qdisc_watchdog_schedule(&q->watchdog, cb->time_to_send);
530 }
531
532 if (q->qdisc) {
533 skb = q->qdisc->ops->dequeue(q->qdisc);
534 if (skb)
535 goto deliver;
536 }
537 return NULL;
538}
539
540static void netem_reset(struct Qdisc *sch)
541{
542 struct netem_sched_data *q = qdisc_priv(sch);
543
544 qdisc_reset_queue(sch);
545 if (q->qdisc)
546 qdisc_reset(q->qdisc);
547 qdisc_watchdog_cancel(&q->watchdog);
548}
549
550static void dist_free(struct disttable *d)
551{
552 if (d) {
553 if (is_vmalloc_addr(d))
554 vfree(d);
555 else
556 kfree(d);
557 }
558}
559
560/*
561 * Distribution data is a variable size payload containing
562 * signed 16 bit values.
563 */
564static int get_dist_table(struct Qdisc *sch, const struct nlattr *attr)
565{
566 struct netem_sched_data *q = qdisc_priv(sch);
567 size_t n = nla_len(attr)/sizeof(__s16);
568 const __s16 *data = nla_data(attr);
569 spinlock_t *root_lock;
570 struct disttable *d;
571 int i;
572 size_t s;
573
574 if (n > NETEM_DIST_MAX)
575 return -EINVAL;
576
577 s = sizeof(struct disttable) + n * sizeof(s16);
578 d = kmalloc(s, GFP_KERNEL | __GFP_NOWARN);
579 if (!d)
580 d = vmalloc(s);
581 if (!d)
582 return -ENOMEM;
583
584 d->size = n;
585 for (i = 0; i < n; i++)
586 d->table[i] = data[i];
587
588 root_lock = qdisc_root_sleeping_lock(sch);
589
590 spin_lock_bh(root_lock);
591 swap(q->delay_dist, d);
592 spin_unlock_bh(root_lock);
593
594 dist_free(d);
595 return 0;
596}
597
598static void get_correlation(struct Qdisc *sch, const struct nlattr *attr)
599{
600 struct netem_sched_data *q = qdisc_priv(sch);
601 const struct tc_netem_corr *c = nla_data(attr);
602
603 init_crandom(&q->delay_cor, c->delay_corr);
604 init_crandom(&q->loss_cor, c->loss_corr);
605 init_crandom(&q->dup_cor, c->dup_corr);
606}
607
608static void get_reorder(struct Qdisc *sch, const struct nlattr *attr)
609{
610 struct netem_sched_data *q = qdisc_priv(sch);
611 const struct tc_netem_reorder *r = nla_data(attr);
612
613 q->reorder = r->probability;
614 init_crandom(&q->reorder_cor, r->correlation);
615}
616
617static void get_corrupt(struct Qdisc *sch, const struct nlattr *attr)
618{
619 struct netem_sched_data *q = qdisc_priv(sch);
620 const struct tc_netem_corrupt *r = nla_data(attr);
621
622 q->corrupt = r->probability;
623 init_crandom(&q->corrupt_cor, r->correlation);
624}
625
626static void get_rate(struct Qdisc *sch, const struct nlattr *attr)
627{
628 struct netem_sched_data *q = qdisc_priv(sch);
629 const struct tc_netem_rate *r = nla_data(attr);
630
631 q->rate = r->rate;
632 q->packet_overhead = r->packet_overhead;
633 q->cell_size = r->cell_size;
634 if (q->cell_size)
635 q->cell_size_reciprocal = reciprocal_value(q->cell_size);
636 q->cell_overhead = r->cell_overhead;
637}
638
639static int get_loss_clg(struct Qdisc *sch, const struct nlattr *attr)
640{
641 struct netem_sched_data *q = qdisc_priv(sch);
642 const struct nlattr *la;
643 int rem;
644
645 nla_for_each_nested(la, attr, rem) {
646 u16 type = nla_type(la);
647
648 switch(type) {
649 case NETEM_LOSS_GI: {
650 const struct tc_netem_gimodel *gi = nla_data(la);
651
652 if (nla_len(la) < sizeof(struct tc_netem_gimodel)) {
653 pr_info("netem: incorrect gi model size\n");
654 return -EINVAL;
655 }
656
657 q->loss_model = CLG_4_STATES;
658
659 q->clg.state = 1;
660 q->clg.a1 = gi->p13;
661 q->clg.a2 = gi->p31;
662 q->clg.a3 = gi->p32;
663 q->clg.a4 = gi->p14;
664 q->clg.a5 = gi->p23;
665 break;
666 }
667
668 case NETEM_LOSS_GE: {
669 const struct tc_netem_gemodel *ge = nla_data(la);
670
671 if (nla_len(la) < sizeof(struct tc_netem_gemodel)) {
672 pr_info("netem: incorrect ge model size\n");
673 return -EINVAL;
674 }
675
676 q->loss_model = CLG_GILB_ELL;
677 q->clg.state = 1;
678 q->clg.a1 = ge->p;
679 q->clg.a2 = ge->r;
680 q->clg.a3 = ge->h;
681 q->clg.a4 = ge->k1;
682 break;
683 }
684
685 default:
686 pr_info("netem: unknown loss type %u\n", type);
687 return -EINVAL;
688 }
689 }
690
691 return 0;
692}
693
694static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
695 [TCA_NETEM_CORR] = { .len = sizeof(struct tc_netem_corr) },
696 [TCA_NETEM_REORDER] = { .len = sizeof(struct tc_netem_reorder) },
697 [TCA_NETEM_CORRUPT] = { .len = sizeof(struct tc_netem_corrupt) },
698 [TCA_NETEM_RATE] = { .len = sizeof(struct tc_netem_rate) },
699 [TCA_NETEM_LOSS] = { .type = NLA_NESTED },
700 [TCA_NETEM_ECN] = { .type = NLA_U32 },
701};
702
703static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
704 const struct nla_policy *policy, int len)
705{
706 int nested_len = nla_len(nla) - NLA_ALIGN(len);
707
708 if (nested_len < 0) {
709 pr_info("netem: invalid attributes len %d\n", nested_len);
710 return -EINVAL;
711 }
712
713 if (nested_len >= nla_attr_size(0))
714 return nla_parse(tb, maxtype, nla_data(nla) + NLA_ALIGN(len),
715 nested_len, policy);
716
717 memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
718 return 0;
719}
720
721/* Parse netlink message to set options */
722static int netem_change(struct Qdisc *sch, struct nlattr *opt)
723{
724 struct netem_sched_data *q = qdisc_priv(sch);
725 struct nlattr *tb[TCA_NETEM_MAX + 1];
726 struct tc_netem_qopt *qopt;
727 int ret;
728
729 if (opt == NULL)
730 return -EINVAL;
731
732 qopt = nla_data(opt);
733 ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
734 if (ret < 0)
735 return ret;
736
737 sch->limit = qopt->limit;
738
739 q->latency = qopt->latency;
740 q->jitter = qopt->jitter;
741 q->limit = qopt->limit;
742 q->gap = qopt->gap;
743 q->counter = 0;
744 q->loss = qopt->loss;
745 q->duplicate = qopt->duplicate;
746
747 /* for compatibility with earlier versions.
748 * if gap is set, need to assume 100% probability
749 */
750 if (q->gap)
751 q->reorder = ~0;
752
753 if (tb[TCA_NETEM_CORR])
754 get_correlation(sch, tb[TCA_NETEM_CORR]);
755
756 if (tb[TCA_NETEM_DELAY_DIST]) {
757 ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST]);
758 if (ret)
759 return ret;
760 }
761
762 if (tb[TCA_NETEM_REORDER])
763 get_reorder(sch, tb[TCA_NETEM_REORDER]);
764
765 if (tb[TCA_NETEM_CORRUPT])
766 get_corrupt(sch, tb[TCA_NETEM_CORRUPT]);
767
768 if (tb[TCA_NETEM_RATE])
769 get_rate(sch, tb[TCA_NETEM_RATE]);
770
771 if (tb[TCA_NETEM_ECN])
772 q->ecn = nla_get_u32(tb[TCA_NETEM_ECN]);
773
774 q->loss_model = CLG_RANDOM;
775 if (tb[TCA_NETEM_LOSS])
776 ret = get_loss_clg(sch, tb[TCA_NETEM_LOSS]);
777
778 return ret;
779}
780
781static int netem_init(struct Qdisc *sch, struct nlattr *opt)
782{
783 struct netem_sched_data *q = qdisc_priv(sch);
784 int ret;
785
786 if (!opt)
787 return -EINVAL;
788
789 qdisc_watchdog_init(&q->watchdog, sch);
790
791 q->loss_model = CLG_RANDOM;
792 ret = netem_change(sch, opt);
793 if (ret)
794 pr_info("netem: change failed\n");
795 return ret;
796}
797
798static void netem_destroy(struct Qdisc *sch)
799{
800 struct netem_sched_data *q = qdisc_priv(sch);
801
802 qdisc_watchdog_cancel(&q->watchdog);
803 if (q->qdisc)
804 qdisc_destroy(q->qdisc);
805 dist_free(q->delay_dist);
806}
807
808static int dump_loss_model(const struct netem_sched_data *q,
809 struct sk_buff *skb)
810{
811 struct nlattr *nest;
812
813 nest = nla_nest_start(skb, TCA_NETEM_LOSS);
814 if (nest == NULL)
815 goto nla_put_failure;
816
817 switch (q->loss_model) {
818 case CLG_RANDOM:
819 /* legacy loss model */
820 nla_nest_cancel(skb, nest);
821 return 0; /* no data */
822
823 case CLG_4_STATES: {
824 struct tc_netem_gimodel gi = {
825 .p13 = q->clg.a1,
826 .p31 = q->clg.a2,
827 .p32 = q->clg.a3,
828 .p14 = q->clg.a4,
829 .p23 = q->clg.a5,
830 };
831
832 if (nla_put(skb, NETEM_LOSS_GI, sizeof(gi), &gi))
833 goto nla_put_failure;
834 break;
835 }
836 case CLG_GILB_ELL: {
837 struct tc_netem_gemodel ge = {
838 .p = q->clg.a1,
839 .r = q->clg.a2,
840 .h = q->clg.a3,
841 .k1 = q->clg.a4,
842 };
843
844 if (nla_put(skb, NETEM_LOSS_GE, sizeof(ge), &ge))
845 goto nla_put_failure;
846 break;
847 }
848 }
849
850 nla_nest_end(skb, nest);
851 return 0;
852
853nla_put_failure:
854 nla_nest_cancel(skb, nest);
855 return -1;
856}
857
858static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
859{
860 const struct netem_sched_data *q = qdisc_priv(sch);
861 struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb);
862 struct tc_netem_qopt qopt;
863 struct tc_netem_corr cor;
864 struct tc_netem_reorder reorder;
865 struct tc_netem_corrupt corrupt;
866 struct tc_netem_rate rate;
867
868 qopt.latency = q->latency;
869 qopt.jitter = q->jitter;
870 qopt.limit = q->limit;
871 qopt.loss = q->loss;
872 qopt.gap = q->gap;
873 qopt.duplicate = q->duplicate;
874 if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
875 goto nla_put_failure;
876
877 cor.delay_corr = q->delay_cor.rho;
878 cor.loss_corr = q->loss_cor.rho;
879 cor.dup_corr = q->dup_cor.rho;
880 if (nla_put(skb, TCA_NETEM_CORR, sizeof(cor), &cor))
881 goto nla_put_failure;
882
883 reorder.probability = q->reorder;
884 reorder.correlation = q->reorder_cor.rho;
885 if (nla_put(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder))
886 goto nla_put_failure;
887
888 corrupt.probability = q->corrupt;
889 corrupt.correlation = q->corrupt_cor.rho;
890 if (nla_put(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt))
891 goto nla_put_failure;
892
893 rate.rate = q->rate;
894 rate.packet_overhead = q->packet_overhead;
895 rate.cell_size = q->cell_size;
896 rate.cell_overhead = q->cell_overhead;
897 if (nla_put(skb, TCA_NETEM_RATE, sizeof(rate), &rate))
898 goto nla_put_failure;
899
900 if (q->ecn && nla_put_u32(skb, TCA_NETEM_ECN, q->ecn))
901 goto nla_put_failure;
902
903 if (dump_loss_model(q, skb) != 0)
904 goto nla_put_failure;
905
906 return nla_nest_end(skb, nla);
907
908nla_put_failure:
909 nlmsg_trim(skb, nla);
910 return -1;
911}
912
913static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
914 struct sk_buff *skb, struct tcmsg *tcm)
915{
916 struct netem_sched_data *q = qdisc_priv(sch);
917
918 if (cl != 1 || !q->qdisc) /* only one class */
919 return -ENOENT;
920
921 tcm->tcm_handle |= TC_H_MIN(1);
922 tcm->tcm_info = q->qdisc->handle;
923
924 return 0;
925}
926
927static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
928 struct Qdisc **old)
929{
930 struct netem_sched_data *q = qdisc_priv(sch);
931
932 sch_tree_lock(sch);
933 *old = q->qdisc;
934 q->qdisc = new;
935 if (*old) {
936 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
937 qdisc_reset(*old);
938 }
939 sch_tree_unlock(sch);
940
941 return 0;
942}
943
944static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
945{
946 struct netem_sched_data *q = qdisc_priv(sch);
947 return q->qdisc;
948}
949
950static unsigned long netem_get(struct Qdisc *sch, u32 classid)
951{
952 return 1;
953}
954
955static void netem_put(struct Qdisc *sch, unsigned long arg)
956{
957}
958
959static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
960{
961 if (!walker->stop) {
962 if (walker->count >= walker->skip)
963 if (walker->fn(sch, 1, walker) < 0) {
964 walker->stop = 1;
965 return;
966 }
967 walker->count++;
968 }
969}
970
971static const struct Qdisc_class_ops netem_class_ops = {
972 .graft = netem_graft,
973 .leaf = netem_leaf,
974 .get = netem_get,
975 .put = netem_put,
976 .walk = netem_walk,
977 .dump = netem_dump_class,
978};
979
980static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
981 .id = "netem",
982 .cl_ops = &netem_class_ops,
983 .priv_size = sizeof(struct netem_sched_data),
984 .enqueue = netem_enqueue,
985 .dequeue = netem_dequeue,
986 .peek = qdisc_peek_dequeued,
987 .drop = netem_drop,
988 .init = netem_init,
989 .reset = netem_reset,
990 .destroy = netem_destroy,
991 .change = netem_change,
992 .dump = netem_dump,
993 .owner = THIS_MODULE,
994};
995
996
997static int __init netem_module_init(void)
998{
999 pr_info("netem: version " VERSION "\n");
1000 return register_qdisc(&netem_qdisc_ops);
1001}
1002static void __exit netem_module_exit(void)
1003{
1004 unregister_qdisc(&netem_qdisc_ops);
1005}
1006module_init(netem_module_init)
1007module_exit(netem_module_exit)
1008MODULE_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");