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 *
143 * As skb->rbnode uses same storage than skb->next, skb->prev and skb->tstamp,
144 * and skb->next & skb->prev are scratch space for a qdisc,
145 * we save skb->tstamp value in skb->cb[] before destroying it.
146 */
147struct netem_skb_cb {
148 psched_time_t time_to_send;
149 ktime_t tstamp_save;
150};
151
152
153static struct sk_buff *netem_rb_to_skb(struct rb_node *rb)
154{
155 return container_of(rb, struct sk_buff, rbnode);
156}
157
158static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb)
159{
160 /* we assume we can use skb next/prev/tstamp as storage for rb_node */
161 qdisc_cb_private_validate(skb, sizeof(struct netem_skb_cb));
162 return (struct netem_skb_cb *)qdisc_skb_cb(skb)->data;
163}
164
165/* init_crandom - initialize correlated random number generator
166 * Use entropy source for initial seed.
167 */
168static void init_crandom(struct crndstate *state, unsigned long rho)
169{
170 state->rho = rho;
171 state->last = prandom_u32();
172}
173
174/* get_crandom - correlated random number generator
175 * Next number depends on last value.
176 * rho is scaled to avoid floating point.
177 */
178static u32 get_crandom(struct crndstate *state)
179{
180 u64 value, rho;
181 unsigned long answer;
182
183 if (state->rho == 0) /* no correlation */
184 return prandom_u32();
185
186 value = prandom_u32();
187 rho = (u64)state->rho + 1;
188 answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32;
189 state->last = answer;
190 return answer;
191}
192
193/* loss_4state - 4-state model loss generator
194 * Generates losses according to the 4-state Markov chain adopted in
195 * the GI (General and Intuitive) loss model.
196 */
197static bool loss_4state(struct netem_sched_data *q)
198{
199 struct clgstate *clg = &q->clg;
200 u32 rnd = prandom_u32();
201
202 /*
203 * Makes a comparison between rnd and the transition
204 * probabilities outgoing from the current state, then decides the
205 * next state and if the next packet has to be transmitted or lost.
206 * The four states correspond to:
207 * TX_IN_GAP_PERIOD => successfully transmitted packets within a gap period
208 * LOST_IN_BURST_PERIOD => isolated losses within a gap period
209 * LOST_IN_GAP_PERIOD => lost packets within a burst period
210 * TX_IN_GAP_PERIOD => successfully transmitted packets within a burst period
211 */
212 switch (clg->state) {
213 case TX_IN_GAP_PERIOD:
214 if (rnd < clg->a4) {
215 clg->state = LOST_IN_BURST_PERIOD;
216 return true;
217 } else if (clg->a4 < rnd && rnd < clg->a1 + clg->a4) {
218 clg->state = LOST_IN_GAP_PERIOD;
219 return true;
220 } else if (clg->a1 + clg->a4 < rnd) {
221 clg->state = TX_IN_GAP_PERIOD;
222 }
223
224 break;
225 case TX_IN_BURST_PERIOD:
226 if (rnd < clg->a5) {
227 clg->state = LOST_IN_GAP_PERIOD;
228 return true;
229 } else {
230 clg->state = TX_IN_BURST_PERIOD;
231 }
232
233 break;
234 case LOST_IN_GAP_PERIOD:
235 if (rnd < clg->a3)
236 clg->state = TX_IN_BURST_PERIOD;
237 else if (clg->a3 < rnd && rnd < clg->a2 + clg->a3) {
238 clg->state = TX_IN_GAP_PERIOD;
239 } else if (clg->a2 + clg->a3 < rnd) {
240 clg->state = LOST_IN_GAP_PERIOD;
241 return true;
242 }
243 break;
244 case LOST_IN_BURST_PERIOD:
245 clg->state = TX_IN_GAP_PERIOD;
246 break;
247 }
248
249 return false;
250}
251
252/* loss_gilb_ell - Gilbert-Elliot model loss generator
253 * Generates losses according to the Gilbert-Elliot loss model or
254 * its special cases (Gilbert or Simple Gilbert)
255 *
256 * Makes a comparison between random number and the transition
257 * probabilities outgoing from the current state, then decides the
258 * next state. A second random number is extracted and the comparison
259 * with the loss probability of the current state decides if the next
260 * packet will be transmitted or lost.
261 */
262static bool loss_gilb_ell(struct netem_sched_data *q)
263{
264 struct clgstate *clg = &q->clg;
265
266 switch (clg->state) {
267 case GOOD_STATE:
268 if (prandom_u32() < clg->a1)
269 clg->state = BAD_STATE;
270 if (prandom_u32() < clg->a4)
271 return true;
272 break;
273 case BAD_STATE:
274 if (prandom_u32() < clg->a2)
275 clg->state = GOOD_STATE;
276 if (prandom_u32() > clg->a3)
277 return true;
278 }
279
280 return false;
281}
282
283static bool loss_event(struct netem_sched_data *q)
284{
285 switch (q->loss_model) {
286 case CLG_RANDOM:
287 /* Random packet drop 0 => none, ~0 => all */
288 return q->loss && q->loss >= get_crandom(&q->loss_cor);
289
290 case CLG_4_STATES:
291 /* 4state loss model algorithm (used also for GI model)
292 * Extracts a value from the markov 4 state loss generator,
293 * if it is 1 drops a packet and if needed writes the event in
294 * the kernel logs
295 */
296 return loss_4state(q);
297
298 case CLG_GILB_ELL:
299 /* Gilbert-Elliot loss model algorithm
300 * Extracts a value from the Gilbert-Elliot loss generator,
301 * if it is 1 drops a packet and if needed writes the event in
302 * the kernel logs
303 */
304 return loss_gilb_ell(q);
305 }
306
307 return false; /* not reached */
308}
309
310
311/* tabledist - return a pseudo-randomly distributed value with mean mu and
312 * std deviation sigma. Uses table lookup to approximate the desired
313 * distribution, and a uniformly-distributed pseudo-random source.
314 */
315static psched_tdiff_t tabledist(psched_tdiff_t mu, psched_tdiff_t sigma,
316 struct crndstate *state,
317 const struct disttable *dist)
318{
319 psched_tdiff_t x;
320 long t;
321 u32 rnd;
322
323 if (sigma == 0)
324 return mu;
325
326 rnd = get_crandom(state);
327
328 /* default uniform distribution */
329 if (dist == NULL)
330 return (rnd % (2*sigma)) - sigma + mu;
331
332 t = dist->table[rnd % dist->size];
333 x = (sigma % NETEM_DIST_SCALE) * t;
334 if (x >= 0)
335 x += NETEM_DIST_SCALE/2;
336 else
337 x -= NETEM_DIST_SCALE/2;
338
339 return x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu;
340}
341
342static psched_time_t packet_len_2_sched_time(unsigned int len, struct netem_sched_data *q)
343{
344 u64 ticks;
345
346 len += q->packet_overhead;
347
348 if (q->cell_size) {
349 u32 cells = reciprocal_divide(len, q->cell_size_reciprocal);
350
351 if (len > cells * q->cell_size) /* extra cell needed for remainder */
352 cells++;
353 len = cells * (q->cell_size + q->cell_overhead);
354 }
355
356 ticks = (u64)len * NSEC_PER_SEC;
357
358 do_div(ticks, q->rate);
359 return PSCHED_NS2TICKS(ticks);
360}
361
362static void tfifo_reset(struct Qdisc *sch)
363{
364 struct netem_sched_data *q = qdisc_priv(sch);
365 struct rb_node *p;
366
367 while ((p = rb_first(&q->t_root))) {
368 struct sk_buff *skb = netem_rb_to_skb(p);
369
370 rb_erase(p, &q->t_root);
371 skb->next = NULL;
372 skb->prev = NULL;
373 kfree_skb(skb);
374 }
375}
376
377static void tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
378{
379 struct netem_sched_data *q = qdisc_priv(sch);
380 psched_time_t tnext = netem_skb_cb(nskb)->time_to_send;
381 struct rb_node **p = &q->t_root.rb_node, *parent = NULL;
382
383 while (*p) {
384 struct sk_buff *skb;
385
386 parent = *p;
387 skb = netem_rb_to_skb(parent);
388 if (tnext >= netem_skb_cb(skb)->time_to_send)
389 p = &parent->rb_right;
390 else
391 p = &parent->rb_left;
392 }
393 rb_link_node(&nskb->rbnode, parent, p);
394 rb_insert_color(&nskb->rbnode, &q->t_root);
395 sch->q.qlen++;
396}
397
398/* netem can't properly corrupt a megapacket (like we get from GSO), so instead
399 * when we statistically choose to corrupt one, we instead segment it, returning
400 * the first packet to be corrupted, and re-enqueue the remaining frames
401 */
402static struct sk_buff *netem_segment(struct sk_buff *skb, struct Qdisc *sch)
403{
404 struct sk_buff *segs;
405 netdev_features_t features = netif_skb_features(skb);
406
407 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
408
409 if (IS_ERR_OR_NULL(segs)) {
410 qdisc_reshape_fail(skb, sch);
411 return NULL;
412 }
413 consume_skb(skb);
414 return segs;
415}
416
417/*
418 * Insert one skb into qdisc.
419 * Note: parent depends on return value to account for queue length.
420 * NET_XMIT_DROP: queue length didn't change.
421 * NET_XMIT_SUCCESS: one skb was queued.
422 */
423static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch)
424{
425 struct netem_sched_data *q = qdisc_priv(sch);
426 /* We don't fill cb now as skb_unshare() may invalidate it */
427 struct netem_skb_cb *cb;
428 struct sk_buff *skb2;
429 struct sk_buff *segs = NULL;
430 unsigned int len = 0, last_len, prev_len = qdisc_pkt_len(skb);
431 int nb = 0;
432 int count = 1;
433 int rc = NET_XMIT_SUCCESS;
434
435 /* Random duplication */
436 if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor))
437 ++count;
438
439 /* Drop packet? */
440 if (loss_event(q)) {
441 if (q->ecn && INET_ECN_set_ce(skb))
442 qdisc_qstats_drop(sch); /* mark packet */
443 else
444 --count;
445 }
446 if (count == 0) {
447 qdisc_qstats_drop(sch);
448 kfree_skb(skb);
449 return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
450 }
451
452 /* If a delay is expected, orphan the skb. (orphaning usually takes
453 * place at TX completion time, so _before_ the link transit delay)
454 */
455 if (q->latency || q->jitter)
456 skb_orphan_partial(skb);
457
458 /*
459 * If we need to duplicate packet, then re-insert at top of the
460 * qdisc tree, since parent queuer expects that only one
461 * skb will be queued.
462 */
463 if (count > 1 && (skb2 = skb_clone(skb, GFP_ATOMIC)) != NULL) {
464 struct Qdisc *rootq = qdisc_root(sch);
465 u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
466
467 q->duplicate = 0;
468 rootq->enqueue(skb2, rootq);
469 q->duplicate = dupsave;
470 }
471
472 /*
473 * Randomized packet corruption.
474 * Make copy if needed since we are modifying
475 * If packet is going to be hardware checksummed, then
476 * do it now in software before we mangle it.
477 */
478 if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor)) {
479 if (skb_is_gso(skb)) {
480 segs = netem_segment(skb, sch);
481 if (!segs)
482 return NET_XMIT_DROP;
483 } else {
484 segs = skb;
485 }
486
487 skb = segs;
488 segs = segs->next;
489
490 if (!(skb = skb_unshare(skb, GFP_ATOMIC)) ||
491 (skb->ip_summed == CHECKSUM_PARTIAL &&
492 skb_checksum_help(skb))) {
493 rc = qdisc_drop(skb, sch);
494 goto finish_segs;
495 }
496
497 skb->data[prandom_u32() % skb_headlen(skb)] ^=
498 1<<(prandom_u32() % 8);
499 }
500
501 if (unlikely(skb_queue_len(&sch->q) >= sch->limit))
502 return qdisc_reshape_fail(skb, sch);
503
504 qdisc_qstats_backlog_inc(sch, skb);
505
506 cb = netem_skb_cb(skb);
507 if (q->gap == 0 || /* not doing reordering */
508 q->counter < q->gap - 1 || /* inside last reordering gap */
509 q->reorder < get_crandom(&q->reorder_cor)) {
510 psched_time_t now;
511 psched_tdiff_t delay;
512
513 delay = tabledist(q->latency, q->jitter,
514 &q->delay_cor, q->delay_dist);
515
516 now = psched_get_time();
517
518 if (q->rate) {
519 struct sk_buff *last;
520
521 if (!skb_queue_empty(&sch->q))
522 last = skb_peek_tail(&sch->q);
523 else
524 last = netem_rb_to_skb(rb_last(&q->t_root));
525 if (last) {
526 /*
527 * Last packet in queue is reference point (now),
528 * calculate this time bonus and subtract
529 * from delay.
530 */
531 delay -= netem_skb_cb(last)->time_to_send - now;
532 delay = max_t(psched_tdiff_t, 0, delay);
533 now = netem_skb_cb(last)->time_to_send;
534 }
535
536 delay += packet_len_2_sched_time(qdisc_pkt_len(skb), q);
537 }
538
539 cb->time_to_send = now + delay;
540 cb->tstamp_save = skb->tstamp;
541 ++q->counter;
542 tfifo_enqueue(skb, sch);
543 } else {
544 /*
545 * Do re-ordering by putting one out of N packets at the front
546 * of the queue.
547 */
548 cb->time_to_send = psched_get_time();
549 q->counter = 0;
550
551 __skb_queue_head(&sch->q, skb);
552 sch->qstats.requeues++;
553 }
554
555finish_segs:
556 if (segs) {
557 while (segs) {
558 skb2 = segs->next;
559 segs->next = NULL;
560 qdisc_skb_cb(segs)->pkt_len = segs->len;
561 last_len = segs->len;
562 rc = qdisc_enqueue(segs, sch);
563 if (rc != NET_XMIT_SUCCESS) {
564 if (net_xmit_drop_count(rc))
565 qdisc_qstats_drop(sch);
566 } else {
567 nb++;
568 len += last_len;
569 }
570 segs = skb2;
571 }
572 sch->q.qlen += nb;
573 if (nb > 1)
574 qdisc_tree_reduce_backlog(sch, 1 - nb, prev_len - len);
575 }
576 return NET_XMIT_SUCCESS;
577}
578
579static unsigned int netem_drop(struct Qdisc *sch)
580{
581 struct netem_sched_data *q = qdisc_priv(sch);
582 unsigned int len;
583
584 len = qdisc_queue_drop(sch);
585
586 if (!len) {
587 struct rb_node *p = rb_first(&q->t_root);
588
589 if (p) {
590 struct sk_buff *skb = netem_rb_to_skb(p);
591
592 rb_erase(p, &q->t_root);
593 sch->q.qlen--;
594 skb->next = NULL;
595 skb->prev = NULL;
596 qdisc_qstats_backlog_dec(sch, skb);
597 kfree_skb(skb);
598 }
599 }
600 if (!len && q->qdisc && q->qdisc->ops->drop)
601 len = q->qdisc->ops->drop(q->qdisc);
602 if (len)
603 qdisc_qstats_drop(sch);
604
605 return len;
606}
607
608static struct sk_buff *netem_dequeue(struct Qdisc *sch)
609{
610 struct netem_sched_data *q = qdisc_priv(sch);
611 struct sk_buff *skb;
612 struct rb_node *p;
613
614 if (qdisc_is_throttled(sch))
615 return NULL;
616
617tfifo_dequeue:
618 skb = __skb_dequeue(&sch->q);
619 if (skb) {
620 qdisc_qstats_backlog_dec(sch, skb);
621deliver:
622 qdisc_unthrottled(sch);
623 qdisc_bstats_update(sch, skb);
624 return skb;
625 }
626 p = rb_first(&q->t_root);
627 if (p) {
628 psched_time_t time_to_send;
629
630 skb = netem_rb_to_skb(p);
631
632 /* if more time remaining? */
633 time_to_send = netem_skb_cb(skb)->time_to_send;
634 if (time_to_send <= psched_get_time()) {
635 rb_erase(p, &q->t_root);
636
637 sch->q.qlen--;
638 qdisc_qstats_backlog_dec(sch, skb);
639 skb->next = NULL;
640 skb->prev = NULL;
641 skb->tstamp = netem_skb_cb(skb)->tstamp_save;
642
643#ifdef CONFIG_NET_CLS_ACT
644 /*
645 * If it's at ingress let's pretend the delay is
646 * from the network (tstamp will be updated).
647 */
648 if (G_TC_FROM(skb->tc_verd) & AT_INGRESS)
649 skb->tstamp.tv64 = 0;
650#endif
651
652 if (q->qdisc) {
653 int err = qdisc_enqueue(skb, q->qdisc);
654
655 if (unlikely(err != NET_XMIT_SUCCESS)) {
656 if (net_xmit_drop_count(err)) {
657 qdisc_qstats_drop(sch);
658 qdisc_tree_reduce_backlog(sch, 1,
659 qdisc_pkt_len(skb));
660 }
661 }
662 goto tfifo_dequeue;
663 }
664 goto deliver;
665 }
666
667 if (q->qdisc) {
668 skb = q->qdisc->ops->dequeue(q->qdisc);
669 if (skb)
670 goto deliver;
671 }
672 qdisc_watchdog_schedule(&q->watchdog, time_to_send);
673 }
674
675 if (q->qdisc) {
676 skb = q->qdisc->ops->dequeue(q->qdisc);
677 if (skb)
678 goto deliver;
679 }
680 return NULL;
681}
682
683static void netem_reset(struct Qdisc *sch)
684{
685 struct netem_sched_data *q = qdisc_priv(sch);
686
687 qdisc_reset_queue(sch);
688 tfifo_reset(sch);
689 if (q->qdisc)
690 qdisc_reset(q->qdisc);
691 qdisc_watchdog_cancel(&q->watchdog);
692}
693
694static void dist_free(struct disttable *d)
695{
696 kvfree(d);
697}
698
699/*
700 * Distribution data is a variable size payload containing
701 * signed 16 bit values.
702 */
703static int get_dist_table(struct Qdisc *sch, const struct nlattr *attr)
704{
705 struct netem_sched_data *q = qdisc_priv(sch);
706 size_t n = nla_len(attr)/sizeof(__s16);
707 const __s16 *data = nla_data(attr);
708 spinlock_t *root_lock;
709 struct disttable *d;
710 int i;
711 size_t s;
712
713 if (n > NETEM_DIST_MAX)
714 return -EINVAL;
715
716 s = sizeof(struct disttable) + n * sizeof(s16);
717 d = kmalloc(s, GFP_KERNEL | __GFP_NOWARN);
718 if (!d)
719 d = vmalloc(s);
720 if (!d)
721 return -ENOMEM;
722
723 d->size = n;
724 for (i = 0; i < n; i++)
725 d->table[i] = data[i];
726
727 root_lock = qdisc_root_sleeping_lock(sch);
728
729 spin_lock_bh(root_lock);
730 swap(q->delay_dist, d);
731 spin_unlock_bh(root_lock);
732
733 dist_free(d);
734 return 0;
735}
736
737static void get_correlation(struct netem_sched_data *q, const struct nlattr *attr)
738{
739 const struct tc_netem_corr *c = nla_data(attr);
740
741 init_crandom(&q->delay_cor, c->delay_corr);
742 init_crandom(&q->loss_cor, c->loss_corr);
743 init_crandom(&q->dup_cor, c->dup_corr);
744}
745
746static void get_reorder(struct netem_sched_data *q, const struct nlattr *attr)
747{
748 const struct tc_netem_reorder *r = nla_data(attr);
749
750 q->reorder = r->probability;
751 init_crandom(&q->reorder_cor, r->correlation);
752}
753
754static void get_corrupt(struct netem_sched_data *q, const struct nlattr *attr)
755{
756 const struct tc_netem_corrupt *r = nla_data(attr);
757
758 q->corrupt = r->probability;
759 init_crandom(&q->corrupt_cor, r->correlation);
760}
761
762static void get_rate(struct netem_sched_data *q, const struct nlattr *attr)
763{
764 const struct tc_netem_rate *r = nla_data(attr);
765
766 q->rate = r->rate;
767 q->packet_overhead = r->packet_overhead;
768 q->cell_size = r->cell_size;
769 q->cell_overhead = r->cell_overhead;
770 if (q->cell_size)
771 q->cell_size_reciprocal = reciprocal_value(q->cell_size);
772 else
773 q->cell_size_reciprocal = (struct reciprocal_value) { 0 };
774}
775
776static int get_loss_clg(struct netem_sched_data *q, const struct nlattr *attr)
777{
778 const struct nlattr *la;
779 int rem;
780
781 nla_for_each_nested(la, attr, rem) {
782 u16 type = nla_type(la);
783
784 switch (type) {
785 case NETEM_LOSS_GI: {
786 const struct tc_netem_gimodel *gi = nla_data(la);
787
788 if (nla_len(la) < sizeof(struct tc_netem_gimodel)) {
789 pr_info("netem: incorrect gi model size\n");
790 return -EINVAL;
791 }
792
793 q->loss_model = CLG_4_STATES;
794
795 q->clg.state = TX_IN_GAP_PERIOD;
796 q->clg.a1 = gi->p13;
797 q->clg.a2 = gi->p31;
798 q->clg.a3 = gi->p32;
799 q->clg.a4 = gi->p14;
800 q->clg.a5 = gi->p23;
801 break;
802 }
803
804 case NETEM_LOSS_GE: {
805 const struct tc_netem_gemodel *ge = nla_data(la);
806
807 if (nla_len(la) < sizeof(struct tc_netem_gemodel)) {
808 pr_info("netem: incorrect ge model size\n");
809 return -EINVAL;
810 }
811
812 q->loss_model = CLG_GILB_ELL;
813 q->clg.state = GOOD_STATE;
814 q->clg.a1 = ge->p;
815 q->clg.a2 = ge->r;
816 q->clg.a3 = ge->h;
817 q->clg.a4 = ge->k1;
818 break;
819 }
820
821 default:
822 pr_info("netem: unknown loss type %u\n", type);
823 return -EINVAL;
824 }
825 }
826
827 return 0;
828}
829
830static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
831 [TCA_NETEM_CORR] = { .len = sizeof(struct tc_netem_corr) },
832 [TCA_NETEM_REORDER] = { .len = sizeof(struct tc_netem_reorder) },
833 [TCA_NETEM_CORRUPT] = { .len = sizeof(struct tc_netem_corrupt) },
834 [TCA_NETEM_RATE] = { .len = sizeof(struct tc_netem_rate) },
835 [TCA_NETEM_LOSS] = { .type = NLA_NESTED },
836 [TCA_NETEM_ECN] = { .type = NLA_U32 },
837 [TCA_NETEM_RATE64] = { .type = NLA_U64 },
838};
839
840static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
841 const struct nla_policy *policy, int len)
842{
843 int nested_len = nla_len(nla) - NLA_ALIGN(len);
844
845 if (nested_len < 0) {
846 pr_info("netem: invalid attributes len %d\n", nested_len);
847 return -EINVAL;
848 }
849
850 if (nested_len >= nla_attr_size(0))
851 return nla_parse(tb, maxtype, nla_data(nla) + NLA_ALIGN(len),
852 nested_len, policy);
853
854 memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
855 return 0;
856}
857
858/* Parse netlink message to set options */
859static int netem_change(struct Qdisc *sch, struct nlattr *opt)
860{
861 struct netem_sched_data *q = qdisc_priv(sch);
862 struct nlattr *tb[TCA_NETEM_MAX + 1];
863 struct tc_netem_qopt *qopt;
864 struct clgstate old_clg;
865 int old_loss_model = CLG_RANDOM;
866 int ret;
867
868 if (opt == NULL)
869 return -EINVAL;
870
871 qopt = nla_data(opt);
872 ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
873 if (ret < 0)
874 return ret;
875
876 /* backup q->clg and q->loss_model */
877 old_clg = q->clg;
878 old_loss_model = q->loss_model;
879
880 if (tb[TCA_NETEM_LOSS]) {
881 ret = get_loss_clg(q, tb[TCA_NETEM_LOSS]);
882 if (ret) {
883 q->loss_model = old_loss_model;
884 return ret;
885 }
886 } else {
887 q->loss_model = CLG_RANDOM;
888 }
889
890 if (tb[TCA_NETEM_DELAY_DIST]) {
891 ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST]);
892 if (ret) {
893 /* recover clg and loss_model, in case of
894 * q->clg and q->loss_model were modified
895 * in get_loss_clg()
896 */
897 q->clg = old_clg;
898 q->loss_model = old_loss_model;
899 return ret;
900 }
901 }
902
903 sch->limit = qopt->limit;
904
905 q->latency = qopt->latency;
906 q->jitter = qopt->jitter;
907 q->limit = qopt->limit;
908 q->gap = qopt->gap;
909 q->counter = 0;
910 q->loss = qopt->loss;
911 q->duplicate = qopt->duplicate;
912
913 /* for compatibility with earlier versions.
914 * if gap is set, need to assume 100% probability
915 */
916 if (q->gap)
917 q->reorder = ~0;
918
919 if (tb[TCA_NETEM_CORR])
920 get_correlation(q, tb[TCA_NETEM_CORR]);
921
922 if (tb[TCA_NETEM_REORDER])
923 get_reorder(q, tb[TCA_NETEM_REORDER]);
924
925 if (tb[TCA_NETEM_CORRUPT])
926 get_corrupt(q, tb[TCA_NETEM_CORRUPT]);
927
928 if (tb[TCA_NETEM_RATE])
929 get_rate(q, tb[TCA_NETEM_RATE]);
930
931 if (tb[TCA_NETEM_RATE64])
932 q->rate = max_t(u64, q->rate,
933 nla_get_u64(tb[TCA_NETEM_RATE64]));
934
935 if (tb[TCA_NETEM_ECN])
936 q->ecn = nla_get_u32(tb[TCA_NETEM_ECN]);
937
938 return ret;
939}
940
941static int netem_init(struct Qdisc *sch, struct nlattr *opt)
942{
943 struct netem_sched_data *q = qdisc_priv(sch);
944 int ret;
945
946 if (!opt)
947 return -EINVAL;
948
949 qdisc_watchdog_init(&q->watchdog, sch);
950
951 q->loss_model = CLG_RANDOM;
952 ret = netem_change(sch, opt);
953 if (ret)
954 pr_info("netem: change failed\n");
955 return ret;
956}
957
958static void netem_destroy(struct Qdisc *sch)
959{
960 struct netem_sched_data *q = qdisc_priv(sch);
961
962 qdisc_watchdog_cancel(&q->watchdog);
963 if (q->qdisc)
964 qdisc_destroy(q->qdisc);
965 dist_free(q->delay_dist);
966}
967
968static int dump_loss_model(const struct netem_sched_data *q,
969 struct sk_buff *skb)
970{
971 struct nlattr *nest;
972
973 nest = nla_nest_start(skb, TCA_NETEM_LOSS);
974 if (nest == NULL)
975 goto nla_put_failure;
976
977 switch (q->loss_model) {
978 case CLG_RANDOM:
979 /* legacy loss model */
980 nla_nest_cancel(skb, nest);
981 return 0; /* no data */
982
983 case CLG_4_STATES: {
984 struct tc_netem_gimodel gi = {
985 .p13 = q->clg.a1,
986 .p31 = q->clg.a2,
987 .p32 = q->clg.a3,
988 .p14 = q->clg.a4,
989 .p23 = q->clg.a5,
990 };
991
992 if (nla_put(skb, NETEM_LOSS_GI, sizeof(gi), &gi))
993 goto nla_put_failure;
994 break;
995 }
996 case CLG_GILB_ELL: {
997 struct tc_netem_gemodel ge = {
998 .p = q->clg.a1,
999 .r = q->clg.a2,
1000 .h = q->clg.a3,
1001 .k1 = q->clg.a4,
1002 };
1003
1004 if (nla_put(skb, NETEM_LOSS_GE, sizeof(ge), &ge))
1005 goto nla_put_failure;
1006 break;
1007 }
1008 }
1009
1010 nla_nest_end(skb, nest);
1011 return 0;
1012
1013nla_put_failure:
1014 nla_nest_cancel(skb, nest);
1015 return -1;
1016}
1017
1018static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
1019{
1020 const struct netem_sched_data *q = qdisc_priv(sch);
1021 struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb);
1022 struct tc_netem_qopt qopt;
1023 struct tc_netem_corr cor;
1024 struct tc_netem_reorder reorder;
1025 struct tc_netem_corrupt corrupt;
1026 struct tc_netem_rate rate;
1027
1028 qopt.latency = q->latency;
1029 qopt.jitter = q->jitter;
1030 qopt.limit = q->limit;
1031 qopt.loss = q->loss;
1032 qopt.gap = q->gap;
1033 qopt.duplicate = q->duplicate;
1034 if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
1035 goto nla_put_failure;
1036
1037 cor.delay_corr = q->delay_cor.rho;
1038 cor.loss_corr = q->loss_cor.rho;
1039 cor.dup_corr = q->dup_cor.rho;
1040 if (nla_put(skb, TCA_NETEM_CORR, sizeof(cor), &cor))
1041 goto nla_put_failure;
1042
1043 reorder.probability = q->reorder;
1044 reorder.correlation = q->reorder_cor.rho;
1045 if (nla_put(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder))
1046 goto nla_put_failure;
1047
1048 corrupt.probability = q->corrupt;
1049 corrupt.correlation = q->corrupt_cor.rho;
1050 if (nla_put(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt))
1051 goto nla_put_failure;
1052
1053 if (q->rate >= (1ULL << 32)) {
1054 if (nla_put_u64(skb, TCA_NETEM_RATE64, q->rate))
1055 goto nla_put_failure;
1056 rate.rate = ~0U;
1057 } else {
1058 rate.rate = q->rate;
1059 }
1060 rate.packet_overhead = q->packet_overhead;
1061 rate.cell_size = q->cell_size;
1062 rate.cell_overhead = q->cell_overhead;
1063 if (nla_put(skb, TCA_NETEM_RATE, sizeof(rate), &rate))
1064 goto nla_put_failure;
1065
1066 if (q->ecn && nla_put_u32(skb, TCA_NETEM_ECN, q->ecn))
1067 goto nla_put_failure;
1068
1069 if (dump_loss_model(q, skb) != 0)
1070 goto nla_put_failure;
1071
1072 return nla_nest_end(skb, nla);
1073
1074nla_put_failure:
1075 nlmsg_trim(skb, nla);
1076 return -1;
1077}
1078
1079static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
1080 struct sk_buff *skb, struct tcmsg *tcm)
1081{
1082 struct netem_sched_data *q = qdisc_priv(sch);
1083
1084 if (cl != 1 || !q->qdisc) /* only one class */
1085 return -ENOENT;
1086
1087 tcm->tcm_handle |= TC_H_MIN(1);
1088 tcm->tcm_info = q->qdisc->handle;
1089
1090 return 0;
1091}
1092
1093static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1094 struct Qdisc **old)
1095{
1096 struct netem_sched_data *q = qdisc_priv(sch);
1097
1098 *old = qdisc_replace(sch, new, &q->qdisc);
1099 return 0;
1100}
1101
1102static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
1103{
1104 struct netem_sched_data *q = qdisc_priv(sch);
1105 return q->qdisc;
1106}
1107
1108static unsigned long netem_get(struct Qdisc *sch, u32 classid)
1109{
1110 return 1;
1111}
1112
1113static void netem_put(struct Qdisc *sch, unsigned long arg)
1114{
1115}
1116
1117static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
1118{
1119 if (!walker->stop) {
1120 if (walker->count >= walker->skip)
1121 if (walker->fn(sch, 1, walker) < 0) {
1122 walker->stop = 1;
1123 return;
1124 }
1125 walker->count++;
1126 }
1127}
1128
1129static const struct Qdisc_class_ops netem_class_ops = {
1130 .graft = netem_graft,
1131 .leaf = netem_leaf,
1132 .get = netem_get,
1133 .put = netem_put,
1134 .walk = netem_walk,
1135 .dump = netem_dump_class,
1136};
1137
1138static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
1139 .id = "netem",
1140 .cl_ops = &netem_class_ops,
1141 .priv_size = sizeof(struct netem_sched_data),
1142 .enqueue = netem_enqueue,
1143 .dequeue = netem_dequeue,
1144 .peek = qdisc_peek_dequeued,
1145 .drop = netem_drop,
1146 .init = netem_init,
1147 .reset = netem_reset,
1148 .destroy = netem_destroy,
1149 .change = netem_change,
1150 .dump = netem_dump,
1151 .owner = THIS_MODULE,
1152};
1153
1154
1155static int __init netem_module_init(void)
1156{
1157 pr_info("netem: version " VERSION "\n");
1158 return register_qdisc(&netem_qdisc_ops);
1159}
1160static void __exit netem_module_exit(void)
1161{
1162 unregister_qdisc(&netem_qdisc_ops);
1163}
1164module_init(netem_module_init)
1165module_exit(netem_module_exit)
1166MODULE_LICENSE("GPL");