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
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");
1/*
2 * net/sched/sch_netem.c Network emulator
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License.
8 *
9 * Many of the algorithms and ideas for this came from
10 * NIST Net which is not copyrighted.
11 *
12 * Authors: Stephen Hemminger <shemminger@osdl.org>
13 * Catalin(ux aka Dino) BOIE <catab at umbrella dot ro>
14 */
15
16#include <linux/mm.h>
17#include <linux/module.h>
18#include <linux/slab.h>
19#include <linux/types.h>
20#include <linux/kernel.h>
21#include <linux/errno.h>
22#include <linux/skbuff.h>
23#include <linux/vmalloc.h>
24#include <linux/rtnetlink.h>
25#include <linux/reciprocal_div.h>
26#include <linux/rbtree.h>
27
28#include <net/netlink.h>
29#include <net/pkt_sched.h>
30#include <net/inet_ecn.h>
31
32#define VERSION "1.3"
33
34/* Network Emulation Queuing algorithm.
35 ====================================
36
37 Sources: [1] Mark Carson, Darrin Santay, "NIST Net - A Linux-based
38 Network Emulation Tool
39 [2] Luigi Rizzo, DummyNet for FreeBSD
40
41 ----------------------------------------------------------------
42
43 This started out as a simple way to delay outgoing packets to
44 test TCP but has grown to include most of the functionality
45 of a full blown network emulator like NISTnet. It can delay
46 packets and add random jitter (and correlation). The random
47 distribution can be loaded from a table as well to provide
48 normal, Pareto, or experimental curves. Packet loss,
49 duplication, and reordering can also be emulated.
50
51 This qdisc does not do classification that can be handled in
52 layering other disciplines. It does not need to do bandwidth
53 control either since that can be handled by using token
54 bucket or other rate control.
55
56 Correlated Loss Generator models
57
58 Added generation of correlated loss according to the
59 "Gilbert-Elliot" model, a 4-state markov model.
60
61 References:
62 [1] NetemCLG Home http://netgroup.uniroma2.it/NetemCLG
63 [2] S. Salsano, F. Ludovici, A. Ordine, "Definition of a general
64 and intuitive loss model for packet networks and its implementation
65 in the Netem module in the Linux kernel", available in [1]
66
67 Authors: Stefano Salsano <stefano.salsano at uniroma2.it
68 Fabio Ludovici <fabio.ludovici at yahoo.it>
69*/
70
71struct netem_sched_data {
72 /* internal t(ime)fifo qdisc uses t_root and sch->limit */
73 struct rb_root t_root;
74
75 /* optional qdisc for classful handling (NULL at netem init) */
76 struct Qdisc *qdisc;
77
78 struct qdisc_watchdog watchdog;
79
80 psched_tdiff_t latency;
81 psched_tdiff_t jitter;
82
83 u32 loss;
84 u32 ecn;
85 u32 limit;
86 u32 counter;
87 u32 gap;
88 u32 duplicate;
89 u32 reorder;
90 u32 corrupt;
91 u64 rate;
92 s32 packet_overhead;
93 u32 cell_size;
94 struct reciprocal_value cell_size_reciprocal;
95 s32 cell_overhead;
96
97 struct crndstate {
98 u32 last;
99 u32 rho;
100 } delay_cor, loss_cor, dup_cor, reorder_cor, corrupt_cor;
101
102 struct disttable {
103 u32 size;
104 s16 table[0];
105 } *delay_dist;
106
107 enum {
108 CLG_RANDOM,
109 CLG_4_STATES,
110 CLG_GILB_ELL,
111 } loss_model;
112
113 enum {
114 TX_IN_GAP_PERIOD = 1,
115 TX_IN_BURST_PERIOD,
116 LOST_IN_GAP_PERIOD,
117 LOST_IN_BURST_PERIOD,
118 } _4_state_model;
119
120 enum {
121 GOOD_STATE = 1,
122 BAD_STATE,
123 } GE_state_model;
124
125 /* Correlated Loss Generation models */
126 struct clgstate {
127 /* state of the Markov chain */
128 u8 state;
129
130 /* 4-states and Gilbert-Elliot models */
131 u32 a1; /* p13 for 4-states or p for GE */
132 u32 a2; /* p31 for 4-states or r for GE */
133 u32 a3; /* p32 for 4-states or h for GE */
134 u32 a4; /* p14 for 4-states or 1-k for GE */
135 u32 a5; /* p23 used only in 4-states */
136 } clg;
137
138};
139
140/* Time stamp put into socket buffer control block
141 * Only valid when skbs are in our internal t(ime)fifo queue.
142 */
143struct netem_skb_cb {
144 psched_time_t time_to_send;
145 ktime_t tstamp_save;
146};
147
148/* Because space in skb->cb[] is tight, netem overloads skb->next/prev/tstamp
149 * to hold a rb_node structure.
150 *
151 * If struct sk_buff layout is changed, the following checks will complain.
152 */
153static struct rb_node *netem_rb_node(struct sk_buff *skb)
154{
155 BUILD_BUG_ON(offsetof(struct sk_buff, next) != 0);
156 BUILD_BUG_ON(offsetof(struct sk_buff, prev) !=
157 offsetof(struct sk_buff, next) + sizeof(skb->next));
158 BUILD_BUG_ON(offsetof(struct sk_buff, tstamp) !=
159 offsetof(struct sk_buff, prev) + sizeof(skb->prev));
160 BUILD_BUG_ON(sizeof(struct rb_node) > sizeof(skb->next) +
161 sizeof(skb->prev) +
162 sizeof(skb->tstamp));
163 return (struct rb_node *)&skb->next;
164}
165
166static struct sk_buff *netem_rb_to_skb(struct rb_node *rb)
167{
168 return (struct sk_buff *)rb;
169}
170
171static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb)
172{
173 /* we assume we can use skb next/prev/tstamp as storage for rb_node */
174 qdisc_cb_private_validate(skb, sizeof(struct netem_skb_cb));
175 return (struct netem_skb_cb *)qdisc_skb_cb(skb)->data;
176}
177
178/* init_crandom - initialize correlated random number generator
179 * Use entropy source for initial seed.
180 */
181static void init_crandom(struct crndstate *state, unsigned long rho)
182{
183 state->rho = rho;
184 state->last = prandom_u32();
185}
186
187/* get_crandom - correlated random number generator
188 * Next number depends on last value.
189 * rho is scaled to avoid floating point.
190 */
191static u32 get_crandom(struct crndstate *state)
192{
193 u64 value, rho;
194 unsigned long answer;
195
196 if (state->rho == 0) /* no correlation */
197 return prandom_u32();
198
199 value = prandom_u32();
200 rho = (u64)state->rho + 1;
201 answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32;
202 state->last = answer;
203 return answer;
204}
205
206/* loss_4state - 4-state model loss generator
207 * Generates losses according to the 4-state Markov chain adopted in
208 * the GI (General and Intuitive) loss model.
209 */
210static bool loss_4state(struct netem_sched_data *q)
211{
212 struct clgstate *clg = &q->clg;
213 u32 rnd = prandom_u32();
214
215 /*
216 * Makes a comparison between rnd and the transition
217 * probabilities outgoing from the current state, then decides the
218 * next state and if the next packet has to be transmitted or lost.
219 * The four states correspond to:
220 * TX_IN_GAP_PERIOD => successfully transmitted packets within a gap period
221 * LOST_IN_BURST_PERIOD => isolated losses within a gap period
222 * LOST_IN_GAP_PERIOD => lost packets within a burst period
223 * TX_IN_GAP_PERIOD => successfully transmitted packets within a burst period
224 */
225 switch (clg->state) {
226 case TX_IN_GAP_PERIOD:
227 if (rnd < clg->a4) {
228 clg->state = LOST_IN_BURST_PERIOD;
229 return true;
230 } else if (clg->a4 < rnd && rnd < clg->a1 + clg->a4) {
231 clg->state = LOST_IN_GAP_PERIOD;
232 return true;
233 } else if (clg->a1 + clg->a4 < rnd) {
234 clg->state = TX_IN_GAP_PERIOD;
235 }
236
237 break;
238 case TX_IN_BURST_PERIOD:
239 if (rnd < clg->a5) {
240 clg->state = LOST_IN_GAP_PERIOD;
241 return true;
242 } else {
243 clg->state = TX_IN_BURST_PERIOD;
244 }
245
246 break;
247 case LOST_IN_GAP_PERIOD:
248 if (rnd < clg->a3)
249 clg->state = TX_IN_BURST_PERIOD;
250 else if (clg->a3 < rnd && rnd < clg->a2 + clg->a3) {
251 clg->state = TX_IN_GAP_PERIOD;
252 } else if (clg->a2 + clg->a3 < rnd) {
253 clg->state = LOST_IN_GAP_PERIOD;
254 return true;
255 }
256 break;
257 case LOST_IN_BURST_PERIOD:
258 clg->state = TX_IN_GAP_PERIOD;
259 break;
260 }
261
262 return false;
263}
264
265/* loss_gilb_ell - Gilbert-Elliot model loss generator
266 * Generates losses according to the Gilbert-Elliot loss model or
267 * its special cases (Gilbert or Simple Gilbert)
268 *
269 * Makes a comparison between random number and the transition
270 * probabilities outgoing from the current state, then decides the
271 * next state. A second random number is extracted and the comparison
272 * with the loss probability of the current state decides if the next
273 * packet will be transmitted or lost.
274 */
275static bool loss_gilb_ell(struct netem_sched_data *q)
276{
277 struct clgstate *clg = &q->clg;
278
279 switch (clg->state) {
280 case GOOD_STATE:
281 if (prandom_u32() < clg->a1)
282 clg->state = BAD_STATE;
283 if (prandom_u32() < clg->a4)
284 return true;
285 break;
286 case BAD_STATE:
287 if (prandom_u32() < clg->a2)
288 clg->state = GOOD_STATE;
289 if (prandom_u32() > clg->a3)
290 return true;
291 }
292
293 return false;
294}
295
296static bool loss_event(struct netem_sched_data *q)
297{
298 switch (q->loss_model) {
299 case CLG_RANDOM:
300 /* Random packet drop 0 => none, ~0 => all */
301 return q->loss && q->loss >= get_crandom(&q->loss_cor);
302
303 case CLG_4_STATES:
304 /* 4state loss model algorithm (used also for GI model)
305 * Extracts a value from the markov 4 state loss generator,
306 * if it is 1 drops a packet and if needed writes the event in
307 * the kernel logs
308 */
309 return loss_4state(q);
310
311 case CLG_GILB_ELL:
312 /* Gilbert-Elliot loss model algorithm
313 * Extracts a value from the Gilbert-Elliot loss generator,
314 * if it is 1 drops a packet and if needed writes the event in
315 * the kernel logs
316 */
317 return loss_gilb_ell(q);
318 }
319
320 return false; /* not reached */
321}
322
323
324/* tabledist - return a pseudo-randomly distributed value with mean mu and
325 * std deviation sigma. Uses table lookup to approximate the desired
326 * distribution, and a uniformly-distributed pseudo-random source.
327 */
328static psched_tdiff_t tabledist(psched_tdiff_t mu, psched_tdiff_t sigma,
329 struct crndstate *state,
330 const struct disttable *dist)
331{
332 psched_tdiff_t x;
333 long t;
334 u32 rnd;
335
336 if (sigma == 0)
337 return mu;
338
339 rnd = get_crandom(state);
340
341 /* default uniform distribution */
342 if (dist == NULL)
343 return (rnd % (2*sigma)) - sigma + mu;
344
345 t = dist->table[rnd % dist->size];
346 x = (sigma % NETEM_DIST_SCALE) * t;
347 if (x >= 0)
348 x += NETEM_DIST_SCALE/2;
349 else
350 x -= NETEM_DIST_SCALE/2;
351
352 return x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu;
353}
354
355static psched_time_t packet_len_2_sched_time(unsigned int len, struct netem_sched_data *q)
356{
357 u64 ticks;
358
359 len += q->packet_overhead;
360
361 if (q->cell_size) {
362 u32 cells = reciprocal_divide(len, q->cell_size_reciprocal);
363
364 if (len > cells * q->cell_size) /* extra cell needed for remainder */
365 cells++;
366 len = cells * (q->cell_size + q->cell_overhead);
367 }
368
369 ticks = (u64)len * NSEC_PER_SEC;
370
371 do_div(ticks, q->rate);
372 return PSCHED_NS2TICKS(ticks);
373}
374
375static void tfifo_reset(struct Qdisc *sch)
376{
377 struct netem_sched_data *q = qdisc_priv(sch);
378 struct rb_node *p;
379
380 while ((p = rb_first(&q->t_root))) {
381 struct sk_buff *skb = netem_rb_to_skb(p);
382
383 rb_erase(p, &q->t_root);
384 skb->next = NULL;
385 skb->prev = NULL;
386 kfree_skb(skb);
387 }
388}
389
390static void tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
391{
392 struct netem_sched_data *q = qdisc_priv(sch);
393 psched_time_t tnext = netem_skb_cb(nskb)->time_to_send;
394 struct rb_node **p = &q->t_root.rb_node, *parent = NULL;
395
396 while (*p) {
397 struct sk_buff *skb;
398
399 parent = *p;
400 skb = netem_rb_to_skb(parent);
401 if (tnext >= netem_skb_cb(skb)->time_to_send)
402 p = &parent->rb_right;
403 else
404 p = &parent->rb_left;
405 }
406 rb_link_node(netem_rb_node(nskb), parent, p);
407 rb_insert_color(netem_rb_node(nskb), &q->t_root);
408 sch->q.qlen++;
409}
410
411/*
412 * Insert one skb into qdisc.
413 * Note: parent depends on return value to account for queue length.
414 * NET_XMIT_DROP: queue length didn't change.
415 * NET_XMIT_SUCCESS: one skb was queued.
416 */
417static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch)
418{
419 struct netem_sched_data *q = qdisc_priv(sch);
420 /* We don't fill cb now as skb_unshare() may invalidate it */
421 struct netem_skb_cb *cb;
422 struct sk_buff *skb2;
423 int count = 1;
424
425 /* Random duplication */
426 if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor))
427 ++count;
428
429 /* Drop packet? */
430 if (loss_event(q)) {
431 if (q->ecn && INET_ECN_set_ce(skb))
432 sch->qstats.drops++; /* mark packet */
433 else
434 --count;
435 }
436 if (count == 0) {
437 sch->qstats.drops++;
438 kfree_skb(skb);
439 return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
440 }
441
442 /* If a delay is expected, orphan the skb. (orphaning usually takes
443 * place at TX completion time, so _before_ the link transit delay)
444 */
445 if (q->latency || q->jitter)
446 skb_orphan_partial(skb);
447
448 /*
449 * If we need to duplicate packet, then re-insert at top of the
450 * qdisc tree, since parent queuer expects that only one
451 * skb will be queued.
452 */
453 if (count > 1 && (skb2 = skb_clone(skb, GFP_ATOMIC)) != NULL) {
454 struct Qdisc *rootq = qdisc_root(sch);
455 u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
456 q->duplicate = 0;
457
458 qdisc_enqueue_root(skb2, rootq);
459 q->duplicate = dupsave;
460 }
461
462 /*
463 * Randomized packet corruption.
464 * Make copy if needed since we are modifying
465 * If packet is going to be hardware checksummed, then
466 * do it now in software before we mangle it.
467 */
468 if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor)) {
469 if (!(skb = skb_unshare(skb, GFP_ATOMIC)) ||
470 (skb->ip_summed == CHECKSUM_PARTIAL &&
471 skb_checksum_help(skb)))
472 return qdisc_drop(skb, sch);
473
474 skb->data[prandom_u32() % skb_headlen(skb)] ^=
475 1<<(prandom_u32() % 8);
476 }
477
478 if (unlikely(skb_queue_len(&sch->q) >= sch->limit))
479 return qdisc_reshape_fail(skb, sch);
480
481 sch->qstats.backlog += qdisc_pkt_len(skb);
482
483 cb = netem_skb_cb(skb);
484 if (q->gap == 0 || /* not doing reordering */
485 q->counter < q->gap - 1 || /* inside last reordering gap */
486 q->reorder < get_crandom(&q->reorder_cor)) {
487 psched_time_t now;
488 psched_tdiff_t delay;
489
490 delay = tabledist(q->latency, q->jitter,
491 &q->delay_cor, q->delay_dist);
492
493 now = psched_get_time();
494
495 if (q->rate) {
496 struct sk_buff *last;
497
498 if (!skb_queue_empty(&sch->q))
499 last = skb_peek_tail(&sch->q);
500 else
501 last = netem_rb_to_skb(rb_last(&q->t_root));
502 if (last) {
503 /*
504 * Last packet in queue is reference point (now),
505 * calculate this time bonus and subtract
506 * from delay.
507 */
508 delay -= netem_skb_cb(last)->time_to_send - now;
509 delay = max_t(psched_tdiff_t, 0, delay);
510 now = netem_skb_cb(last)->time_to_send;
511 }
512
513 delay += packet_len_2_sched_time(qdisc_pkt_len(skb), q);
514 }
515
516 cb->time_to_send = now + delay;
517 cb->tstamp_save = skb->tstamp;
518 ++q->counter;
519 tfifo_enqueue(skb, sch);
520 } else {
521 /*
522 * Do re-ordering by putting one out of N packets at the front
523 * of the queue.
524 */
525 cb->time_to_send = psched_get_time();
526 q->counter = 0;
527
528 __skb_queue_head(&sch->q, skb);
529 sch->qstats.requeues++;
530 }
531
532 return NET_XMIT_SUCCESS;
533}
534
535static unsigned int netem_drop(struct Qdisc *sch)
536{
537 struct netem_sched_data *q = qdisc_priv(sch);
538 unsigned int len;
539
540 len = qdisc_queue_drop(sch);
541
542 if (!len) {
543 struct rb_node *p = rb_first(&q->t_root);
544
545 if (p) {
546 struct sk_buff *skb = netem_rb_to_skb(p);
547
548 rb_erase(p, &q->t_root);
549 sch->q.qlen--;
550 skb->next = NULL;
551 skb->prev = NULL;
552 len = qdisc_pkt_len(skb);
553 sch->qstats.backlog -= len;
554 kfree_skb(skb);
555 }
556 }
557 if (!len && q->qdisc && q->qdisc->ops->drop)
558 len = q->qdisc->ops->drop(q->qdisc);
559 if (len)
560 sch->qstats.drops++;
561
562 return len;
563}
564
565static struct sk_buff *netem_dequeue(struct Qdisc *sch)
566{
567 struct netem_sched_data *q = qdisc_priv(sch);
568 struct sk_buff *skb;
569 struct rb_node *p;
570
571 if (qdisc_is_throttled(sch))
572 return NULL;
573
574tfifo_dequeue:
575 skb = __skb_dequeue(&sch->q);
576 if (skb) {
577deliver:
578 sch->qstats.backlog -= qdisc_pkt_len(skb);
579 qdisc_unthrottled(sch);
580 qdisc_bstats_update(sch, skb);
581 return skb;
582 }
583 p = rb_first(&q->t_root);
584 if (p) {
585 psched_time_t time_to_send;
586
587 skb = netem_rb_to_skb(p);
588
589 /* if more time remaining? */
590 time_to_send = netem_skb_cb(skb)->time_to_send;
591 if (time_to_send <= psched_get_time()) {
592 rb_erase(p, &q->t_root);
593
594 sch->q.qlen--;
595 skb->next = NULL;
596 skb->prev = NULL;
597 skb->tstamp = netem_skb_cb(skb)->tstamp_save;
598
599#ifdef CONFIG_NET_CLS_ACT
600 /*
601 * If it's at ingress let's pretend the delay is
602 * from the network (tstamp will be updated).
603 */
604 if (G_TC_FROM(skb->tc_verd) & AT_INGRESS)
605 skb->tstamp.tv64 = 0;
606#endif
607
608 if (q->qdisc) {
609 int err = qdisc_enqueue(skb, q->qdisc);
610
611 if (unlikely(err != NET_XMIT_SUCCESS)) {
612 if (net_xmit_drop_count(err)) {
613 sch->qstats.drops++;
614 qdisc_tree_decrease_qlen(sch, 1);
615 }
616 }
617 goto tfifo_dequeue;
618 }
619 goto deliver;
620 }
621
622 if (q->qdisc) {
623 skb = q->qdisc->ops->dequeue(q->qdisc);
624 if (skb)
625 goto deliver;
626 }
627 qdisc_watchdog_schedule(&q->watchdog, time_to_send);
628 }
629
630 if (q->qdisc) {
631 skb = q->qdisc->ops->dequeue(q->qdisc);
632 if (skb)
633 goto deliver;
634 }
635 return NULL;
636}
637
638static void netem_reset(struct Qdisc *sch)
639{
640 struct netem_sched_data *q = qdisc_priv(sch);
641
642 qdisc_reset_queue(sch);
643 tfifo_reset(sch);
644 if (q->qdisc)
645 qdisc_reset(q->qdisc);
646 qdisc_watchdog_cancel(&q->watchdog);
647}
648
649static void dist_free(struct disttable *d)
650{
651 if (d) {
652 if (is_vmalloc_addr(d))
653 vfree(d);
654 else
655 kfree(d);
656 }
657}
658
659/*
660 * Distribution data is a variable size payload containing
661 * signed 16 bit values.
662 */
663static int get_dist_table(struct Qdisc *sch, const struct nlattr *attr)
664{
665 struct netem_sched_data *q = qdisc_priv(sch);
666 size_t n = nla_len(attr)/sizeof(__s16);
667 const __s16 *data = nla_data(attr);
668 spinlock_t *root_lock;
669 struct disttable *d;
670 int i;
671 size_t s;
672
673 if (n > NETEM_DIST_MAX)
674 return -EINVAL;
675
676 s = sizeof(struct disttable) + n * sizeof(s16);
677 d = kmalloc(s, GFP_KERNEL | __GFP_NOWARN);
678 if (!d)
679 d = vmalloc(s);
680 if (!d)
681 return -ENOMEM;
682
683 d->size = n;
684 for (i = 0; i < n; i++)
685 d->table[i] = data[i];
686
687 root_lock = qdisc_root_sleeping_lock(sch);
688
689 spin_lock_bh(root_lock);
690 swap(q->delay_dist, d);
691 spin_unlock_bh(root_lock);
692
693 dist_free(d);
694 return 0;
695}
696
697static void get_correlation(struct netem_sched_data *q, const struct nlattr *attr)
698{
699 const struct tc_netem_corr *c = nla_data(attr);
700
701 init_crandom(&q->delay_cor, c->delay_corr);
702 init_crandom(&q->loss_cor, c->loss_corr);
703 init_crandom(&q->dup_cor, c->dup_corr);
704}
705
706static void get_reorder(struct netem_sched_data *q, const struct nlattr *attr)
707{
708 const struct tc_netem_reorder *r = nla_data(attr);
709
710 q->reorder = r->probability;
711 init_crandom(&q->reorder_cor, r->correlation);
712}
713
714static void get_corrupt(struct netem_sched_data *q, const struct nlattr *attr)
715{
716 const struct tc_netem_corrupt *r = nla_data(attr);
717
718 q->corrupt = r->probability;
719 init_crandom(&q->corrupt_cor, r->correlation);
720}
721
722static void get_rate(struct netem_sched_data *q, const struct nlattr *attr)
723{
724 const struct tc_netem_rate *r = nla_data(attr);
725
726 q->rate = r->rate;
727 q->packet_overhead = r->packet_overhead;
728 q->cell_size = r->cell_size;
729 q->cell_overhead = r->cell_overhead;
730 if (q->cell_size)
731 q->cell_size_reciprocal = reciprocal_value(q->cell_size);
732 else
733 q->cell_size_reciprocal = (struct reciprocal_value) { 0 };
734}
735
736static int get_loss_clg(struct netem_sched_data *q, const struct nlattr *attr)
737{
738 const struct nlattr *la;
739 int rem;
740
741 nla_for_each_nested(la, attr, rem) {
742 u16 type = nla_type(la);
743
744 switch (type) {
745 case NETEM_LOSS_GI: {
746 const struct tc_netem_gimodel *gi = nla_data(la);
747
748 if (nla_len(la) < sizeof(struct tc_netem_gimodel)) {
749 pr_info("netem: incorrect gi model size\n");
750 return -EINVAL;
751 }
752
753 q->loss_model = CLG_4_STATES;
754
755 q->clg.state = TX_IN_GAP_PERIOD;
756 q->clg.a1 = gi->p13;
757 q->clg.a2 = gi->p31;
758 q->clg.a3 = gi->p32;
759 q->clg.a4 = gi->p14;
760 q->clg.a5 = gi->p23;
761 break;
762 }
763
764 case NETEM_LOSS_GE: {
765 const struct tc_netem_gemodel *ge = nla_data(la);
766
767 if (nla_len(la) < sizeof(struct tc_netem_gemodel)) {
768 pr_info("netem: incorrect ge model size\n");
769 return -EINVAL;
770 }
771
772 q->loss_model = CLG_GILB_ELL;
773 q->clg.state = GOOD_STATE;
774 q->clg.a1 = ge->p;
775 q->clg.a2 = ge->r;
776 q->clg.a3 = ge->h;
777 q->clg.a4 = ge->k1;
778 break;
779 }
780
781 default:
782 pr_info("netem: unknown loss type %u\n", type);
783 return -EINVAL;
784 }
785 }
786
787 return 0;
788}
789
790static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
791 [TCA_NETEM_CORR] = { .len = sizeof(struct tc_netem_corr) },
792 [TCA_NETEM_REORDER] = { .len = sizeof(struct tc_netem_reorder) },
793 [TCA_NETEM_CORRUPT] = { .len = sizeof(struct tc_netem_corrupt) },
794 [TCA_NETEM_RATE] = { .len = sizeof(struct tc_netem_rate) },
795 [TCA_NETEM_LOSS] = { .type = NLA_NESTED },
796 [TCA_NETEM_ECN] = { .type = NLA_U32 },
797 [TCA_NETEM_RATE64] = { .type = NLA_U64 },
798};
799
800static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
801 const struct nla_policy *policy, int len)
802{
803 int nested_len = nla_len(nla) - NLA_ALIGN(len);
804
805 if (nested_len < 0) {
806 pr_info("netem: invalid attributes len %d\n", nested_len);
807 return -EINVAL;
808 }
809
810 if (nested_len >= nla_attr_size(0))
811 return nla_parse(tb, maxtype, nla_data(nla) + NLA_ALIGN(len),
812 nested_len, policy);
813
814 memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
815 return 0;
816}
817
818/* Parse netlink message to set options */
819static int netem_change(struct Qdisc *sch, struct nlattr *opt)
820{
821 struct netem_sched_data *q = qdisc_priv(sch);
822 struct nlattr *tb[TCA_NETEM_MAX + 1];
823 struct tc_netem_qopt *qopt;
824 struct clgstate old_clg;
825 int old_loss_model = CLG_RANDOM;
826 int ret;
827
828 if (opt == NULL)
829 return -EINVAL;
830
831 qopt = nla_data(opt);
832 ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
833 if (ret < 0)
834 return ret;
835
836 /* backup q->clg and q->loss_model */
837 old_clg = q->clg;
838 old_loss_model = q->loss_model;
839
840 if (tb[TCA_NETEM_LOSS]) {
841 ret = get_loss_clg(q, tb[TCA_NETEM_LOSS]);
842 if (ret) {
843 q->loss_model = old_loss_model;
844 return ret;
845 }
846 } else {
847 q->loss_model = CLG_RANDOM;
848 }
849
850 if (tb[TCA_NETEM_DELAY_DIST]) {
851 ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST]);
852 if (ret) {
853 /* recover clg and loss_model, in case of
854 * q->clg and q->loss_model were modified
855 * in get_loss_clg()
856 */
857 q->clg = old_clg;
858 q->loss_model = old_loss_model;
859 return ret;
860 }
861 }
862
863 sch->limit = qopt->limit;
864
865 q->latency = qopt->latency;
866 q->jitter = qopt->jitter;
867 q->limit = qopt->limit;
868 q->gap = qopt->gap;
869 q->counter = 0;
870 q->loss = qopt->loss;
871 q->duplicate = qopt->duplicate;
872
873 /* for compatibility with earlier versions.
874 * if gap is set, need to assume 100% probability
875 */
876 if (q->gap)
877 q->reorder = ~0;
878
879 if (tb[TCA_NETEM_CORR])
880 get_correlation(q, tb[TCA_NETEM_CORR]);
881
882 if (tb[TCA_NETEM_REORDER])
883 get_reorder(q, tb[TCA_NETEM_REORDER]);
884
885 if (tb[TCA_NETEM_CORRUPT])
886 get_corrupt(q, tb[TCA_NETEM_CORRUPT]);
887
888 if (tb[TCA_NETEM_RATE])
889 get_rate(q, tb[TCA_NETEM_RATE]);
890
891 if (tb[TCA_NETEM_RATE64])
892 q->rate = max_t(u64, q->rate,
893 nla_get_u64(tb[TCA_NETEM_RATE64]));
894
895 if (tb[TCA_NETEM_ECN])
896 q->ecn = nla_get_u32(tb[TCA_NETEM_ECN]);
897
898 return ret;
899}
900
901static int netem_init(struct Qdisc *sch, struct nlattr *opt)
902{
903 struct netem_sched_data *q = qdisc_priv(sch);
904 int ret;
905
906 if (!opt)
907 return -EINVAL;
908
909 qdisc_watchdog_init(&q->watchdog, sch);
910
911 q->loss_model = CLG_RANDOM;
912 ret = netem_change(sch, opt);
913 if (ret)
914 pr_info("netem: change failed\n");
915 return ret;
916}
917
918static void netem_destroy(struct Qdisc *sch)
919{
920 struct netem_sched_data *q = qdisc_priv(sch);
921
922 qdisc_watchdog_cancel(&q->watchdog);
923 if (q->qdisc)
924 qdisc_destroy(q->qdisc);
925 dist_free(q->delay_dist);
926}
927
928static int dump_loss_model(const struct netem_sched_data *q,
929 struct sk_buff *skb)
930{
931 struct nlattr *nest;
932
933 nest = nla_nest_start(skb, TCA_NETEM_LOSS);
934 if (nest == NULL)
935 goto nla_put_failure;
936
937 switch (q->loss_model) {
938 case CLG_RANDOM:
939 /* legacy loss model */
940 nla_nest_cancel(skb, nest);
941 return 0; /* no data */
942
943 case CLG_4_STATES: {
944 struct tc_netem_gimodel gi = {
945 .p13 = q->clg.a1,
946 .p31 = q->clg.a2,
947 .p32 = q->clg.a3,
948 .p14 = q->clg.a4,
949 .p23 = q->clg.a5,
950 };
951
952 if (nla_put(skb, NETEM_LOSS_GI, sizeof(gi), &gi))
953 goto nla_put_failure;
954 break;
955 }
956 case CLG_GILB_ELL: {
957 struct tc_netem_gemodel ge = {
958 .p = q->clg.a1,
959 .r = q->clg.a2,
960 .h = q->clg.a3,
961 .k1 = q->clg.a4,
962 };
963
964 if (nla_put(skb, NETEM_LOSS_GE, sizeof(ge), &ge))
965 goto nla_put_failure;
966 break;
967 }
968 }
969
970 nla_nest_end(skb, nest);
971 return 0;
972
973nla_put_failure:
974 nla_nest_cancel(skb, nest);
975 return -1;
976}
977
978static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
979{
980 const struct netem_sched_data *q = qdisc_priv(sch);
981 struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb);
982 struct tc_netem_qopt qopt;
983 struct tc_netem_corr cor;
984 struct tc_netem_reorder reorder;
985 struct tc_netem_corrupt corrupt;
986 struct tc_netem_rate rate;
987
988 qopt.latency = q->latency;
989 qopt.jitter = q->jitter;
990 qopt.limit = q->limit;
991 qopt.loss = q->loss;
992 qopt.gap = q->gap;
993 qopt.duplicate = q->duplicate;
994 if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
995 goto nla_put_failure;
996
997 cor.delay_corr = q->delay_cor.rho;
998 cor.loss_corr = q->loss_cor.rho;
999 cor.dup_corr = q->dup_cor.rho;
1000 if (nla_put(skb, TCA_NETEM_CORR, sizeof(cor), &cor))
1001 goto nla_put_failure;
1002
1003 reorder.probability = q->reorder;
1004 reorder.correlation = q->reorder_cor.rho;
1005 if (nla_put(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder))
1006 goto nla_put_failure;
1007
1008 corrupt.probability = q->corrupt;
1009 corrupt.correlation = q->corrupt_cor.rho;
1010 if (nla_put(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt))
1011 goto nla_put_failure;
1012
1013 if (q->rate >= (1ULL << 32)) {
1014 if (nla_put_u64(skb, TCA_NETEM_RATE64, q->rate))
1015 goto nla_put_failure;
1016 rate.rate = ~0U;
1017 } else {
1018 rate.rate = q->rate;
1019 }
1020 rate.packet_overhead = q->packet_overhead;
1021 rate.cell_size = q->cell_size;
1022 rate.cell_overhead = q->cell_overhead;
1023 if (nla_put(skb, TCA_NETEM_RATE, sizeof(rate), &rate))
1024 goto nla_put_failure;
1025
1026 if (q->ecn && nla_put_u32(skb, TCA_NETEM_ECN, q->ecn))
1027 goto nla_put_failure;
1028
1029 if (dump_loss_model(q, skb) != 0)
1030 goto nla_put_failure;
1031
1032 return nla_nest_end(skb, nla);
1033
1034nla_put_failure:
1035 nlmsg_trim(skb, nla);
1036 return -1;
1037}
1038
1039static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
1040 struct sk_buff *skb, struct tcmsg *tcm)
1041{
1042 struct netem_sched_data *q = qdisc_priv(sch);
1043
1044 if (cl != 1 || !q->qdisc) /* only one class */
1045 return -ENOENT;
1046
1047 tcm->tcm_handle |= TC_H_MIN(1);
1048 tcm->tcm_info = q->qdisc->handle;
1049
1050 return 0;
1051}
1052
1053static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1054 struct Qdisc **old)
1055{
1056 struct netem_sched_data *q = qdisc_priv(sch);
1057
1058 sch_tree_lock(sch);
1059 *old = q->qdisc;
1060 q->qdisc = new;
1061 if (*old) {
1062 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
1063 qdisc_reset(*old);
1064 }
1065 sch_tree_unlock(sch);
1066
1067 return 0;
1068}
1069
1070static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
1071{
1072 struct netem_sched_data *q = qdisc_priv(sch);
1073 return q->qdisc;
1074}
1075
1076static unsigned long netem_get(struct Qdisc *sch, u32 classid)
1077{
1078 return 1;
1079}
1080
1081static void netem_put(struct Qdisc *sch, unsigned long arg)
1082{
1083}
1084
1085static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
1086{
1087 if (!walker->stop) {
1088 if (walker->count >= walker->skip)
1089 if (walker->fn(sch, 1, walker) < 0) {
1090 walker->stop = 1;
1091 return;
1092 }
1093 walker->count++;
1094 }
1095}
1096
1097static const struct Qdisc_class_ops netem_class_ops = {
1098 .graft = netem_graft,
1099 .leaf = netem_leaf,
1100 .get = netem_get,
1101 .put = netem_put,
1102 .walk = netem_walk,
1103 .dump = netem_dump_class,
1104};
1105
1106static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
1107 .id = "netem",
1108 .cl_ops = &netem_class_ops,
1109 .priv_size = sizeof(struct netem_sched_data),
1110 .enqueue = netem_enqueue,
1111 .dequeue = netem_dequeue,
1112 .peek = qdisc_peek_dequeued,
1113 .drop = netem_drop,
1114 .init = netem_init,
1115 .reset = netem_reset,
1116 .destroy = netem_destroy,
1117 .change = netem_change,
1118 .dump = netem_dump,
1119 .owner = THIS_MODULE,
1120};
1121
1122
1123static int __init netem_module_init(void)
1124{
1125 pr_info("netem: version " VERSION "\n");
1126 return register_qdisc(&netem_qdisc_ops);
1127}
1128static void __exit netem_module_exit(void)
1129{
1130 unregister_qdisc(&netem_qdisc_ops);
1131}
1132module_init(netem_module_init)
1133module_exit(netem_module_exit)
1134MODULE_LICENSE("GPL");