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