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// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * net/sched/sch_netem.c Network emulator
4 *
5 * Many of the algorithms and ideas for this came from
6 * NIST Net which is not copyrighted.
7 *
8 * Authors: Stephen Hemminger <shemminger@osdl.org>
9 * Catalin(ux aka Dino) BOIE <catab at umbrella dot ro>
10 */
11
12#include <linux/mm.h>
13#include <linux/module.h>
14#include <linux/slab.h>
15#include <linux/types.h>
16#include <linux/kernel.h>
17#include <linux/errno.h>
18#include <linux/skbuff.h>
19#include <linux/vmalloc.h>
20#include <linux/rtnetlink.h>
21#include <linux/reciprocal_div.h>
22#include <linux/rbtree.h>
23
24#include <net/gso.h>
25#include <net/netlink.h>
26#include <net/pkt_sched.h>
27#include <net/inet_ecn.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 disttable {
69 u32 size;
70 s16 table[] __counted_by(size);
71};
72
73struct netem_sched_data {
74 /* internal t(ime)fifo qdisc uses t_root and sch->limit */
75 struct rb_root t_root;
76
77 /* a linear queue; reduces rbtree rebalancing when jitter is low */
78 struct sk_buff *t_head;
79 struct sk_buff *t_tail;
80
81 /* optional qdisc for classful handling (NULL at netem init) */
82 struct Qdisc *qdisc;
83
84 struct qdisc_watchdog watchdog;
85
86 s64 latency;
87 s64 jitter;
88
89 u32 loss;
90 u32 ecn;
91 u32 limit;
92 u32 counter;
93 u32 gap;
94 u32 duplicate;
95 u32 reorder;
96 u32 corrupt;
97 u64 rate;
98 s32 packet_overhead;
99 u32 cell_size;
100 struct reciprocal_value cell_size_reciprocal;
101 s32 cell_overhead;
102
103 struct crndstate {
104 u32 last;
105 u32 rho;
106 } delay_cor, loss_cor, dup_cor, reorder_cor, corrupt_cor;
107
108 struct prng {
109 u64 seed;
110 struct rnd_state prng_state;
111 } prng;
112
113 struct disttable *delay_dist;
114
115 enum {
116 CLG_RANDOM,
117 CLG_4_STATES,
118 CLG_GILB_ELL,
119 } loss_model;
120
121 enum {
122 TX_IN_GAP_PERIOD = 1,
123 TX_IN_BURST_PERIOD,
124 LOST_IN_GAP_PERIOD,
125 LOST_IN_BURST_PERIOD,
126 } _4_state_model;
127
128 enum {
129 GOOD_STATE = 1,
130 BAD_STATE,
131 } GE_state_model;
132
133 /* Correlated Loss Generation models */
134 struct clgstate {
135 /* state of the Markov chain */
136 u8 state;
137
138 /* 4-states and Gilbert-Elliot models */
139 u32 a1; /* p13 for 4-states or p for GE */
140 u32 a2; /* p31 for 4-states or r for GE */
141 u32 a3; /* p32 for 4-states or h for GE */
142 u32 a4; /* p14 for 4-states or 1-k for GE */
143 u32 a5; /* p23 used only in 4-states */
144 } clg;
145
146 struct tc_netem_slot slot_config;
147 struct slotstate {
148 u64 slot_next;
149 s32 packets_left;
150 s32 bytes_left;
151 } slot;
152
153 struct disttable *slot_dist;
154};
155
156/* Time stamp put into socket buffer control block
157 * Only valid when skbs are in our internal t(ime)fifo queue.
158 *
159 * As skb->rbnode uses same storage than skb->next, skb->prev and skb->tstamp,
160 * and skb->next & skb->prev are scratch space for a qdisc,
161 * we save skb->tstamp value in skb->cb[] before destroying it.
162 */
163struct netem_skb_cb {
164 u64 time_to_send;
165};
166
167static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb)
168{
169 /* we assume we can use skb next/prev/tstamp as storage for rb_node */
170 qdisc_cb_private_validate(skb, sizeof(struct netem_skb_cb));
171 return (struct netem_skb_cb *)qdisc_skb_cb(skb)->data;
172}
173
174/* init_crandom - initialize correlated random number generator
175 * Use entropy source for initial seed.
176 */
177static void init_crandom(struct crndstate *state, unsigned long rho)
178{
179 state->rho = rho;
180 state->last = get_random_u32();
181}
182
183/* get_crandom - correlated random number generator
184 * Next number depends on last value.
185 * rho is scaled to avoid floating point.
186 */
187static u32 get_crandom(struct crndstate *state, struct prng *p)
188{
189 u64 value, rho;
190 unsigned long answer;
191 struct rnd_state *s = &p->prng_state;
192
193 if (!state || state->rho == 0) /* no correlation */
194 return prandom_u32_state(s);
195
196 value = prandom_u32_state(s);
197 rho = (u64)state->rho + 1;
198 answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32;
199 state->last = answer;
200 return answer;
201}
202
203/* loss_4state - 4-state model loss generator
204 * Generates losses according to the 4-state Markov chain adopted in
205 * the GI (General and Intuitive) loss model.
206 */
207static bool loss_4state(struct netem_sched_data *q)
208{
209 struct clgstate *clg = &q->clg;
210 u32 rnd = prandom_u32_state(&q->prng.prng_state);
211
212 /*
213 * Makes a comparison between rnd and the transition
214 * probabilities outgoing from the current state, then decides the
215 * next state and if the next packet has to be transmitted or lost.
216 * The four states correspond to:
217 * TX_IN_GAP_PERIOD => successfully transmitted packets within a gap period
218 * LOST_IN_GAP_PERIOD => isolated losses within a gap period
219 * LOST_IN_BURST_PERIOD => lost packets within a burst period
220 * TX_IN_BURST_PERIOD => successfully transmitted packets within a burst period
221 */
222 switch (clg->state) {
223 case TX_IN_GAP_PERIOD:
224 if (rnd < clg->a4) {
225 clg->state = LOST_IN_GAP_PERIOD;
226 return true;
227 } else if (clg->a4 < rnd && rnd < clg->a1 + clg->a4) {
228 clg->state = LOST_IN_BURST_PERIOD;
229 return true;
230 } else if (clg->a1 + clg->a4 < rnd) {
231 clg->state = TX_IN_GAP_PERIOD;
232 }
233
234 break;
235 case TX_IN_BURST_PERIOD:
236 if (rnd < clg->a5) {
237 clg->state = LOST_IN_BURST_PERIOD;
238 return true;
239 } else {
240 clg->state = TX_IN_BURST_PERIOD;
241 }
242
243 break;
244 case LOST_IN_BURST_PERIOD:
245 if (rnd < clg->a3)
246 clg->state = TX_IN_BURST_PERIOD;
247 else if (clg->a3 < rnd && rnd < clg->a2 + clg->a3) {
248 clg->state = TX_IN_GAP_PERIOD;
249 } else if (clg->a2 + clg->a3 < rnd) {
250 clg->state = LOST_IN_BURST_PERIOD;
251 return true;
252 }
253 break;
254 case LOST_IN_GAP_PERIOD:
255 clg->state = TX_IN_GAP_PERIOD;
256 break;
257 }
258
259 return false;
260}
261
262/* loss_gilb_ell - Gilbert-Elliot model loss generator
263 * Generates losses according to the Gilbert-Elliot loss model or
264 * its special cases (Gilbert or Simple Gilbert)
265 *
266 * Makes a comparison between random number and the transition
267 * probabilities outgoing from the current state, then decides the
268 * next state. A second random number is extracted and the comparison
269 * with the loss probability of the current state decides if the next
270 * packet will be transmitted or lost.
271 */
272static bool loss_gilb_ell(struct netem_sched_data *q)
273{
274 struct clgstate *clg = &q->clg;
275 struct rnd_state *s = &q->prng.prng_state;
276
277 switch (clg->state) {
278 case GOOD_STATE:
279 if (prandom_u32_state(s) < clg->a1)
280 clg->state = BAD_STATE;
281 if (prandom_u32_state(s) < clg->a4)
282 return true;
283 break;
284 case BAD_STATE:
285 if (prandom_u32_state(s) < clg->a2)
286 clg->state = GOOD_STATE;
287 if (prandom_u32_state(s) > clg->a3)
288 return true;
289 }
290
291 return false;
292}
293
294static bool loss_event(struct netem_sched_data *q)
295{
296 switch (q->loss_model) {
297 case CLG_RANDOM:
298 /* Random packet drop 0 => none, ~0 => all */
299 return q->loss && q->loss >= get_crandom(&q->loss_cor, &q->prng);
300
301 case CLG_4_STATES:
302 /* 4state loss model algorithm (used also for GI model)
303 * Extracts a value from the markov 4 state loss generator,
304 * if it is 1 drops a packet and if needed writes the event in
305 * the kernel logs
306 */
307 return loss_4state(q);
308
309 case CLG_GILB_ELL:
310 /* Gilbert-Elliot loss model algorithm
311 * Extracts a value from the Gilbert-Elliot loss generator,
312 * if it is 1 drops a packet and if needed writes the event in
313 * the kernel logs
314 */
315 return loss_gilb_ell(q);
316 }
317
318 return false; /* not reached */
319}
320
321
322/* tabledist - return a pseudo-randomly distributed value with mean mu and
323 * std deviation sigma. Uses table lookup to approximate the desired
324 * distribution, and a uniformly-distributed pseudo-random source.
325 */
326static s64 tabledist(s64 mu, s32 sigma,
327 struct crndstate *state,
328 struct prng *prng,
329 const struct disttable *dist)
330{
331 s64 x;
332 long t;
333 u32 rnd;
334
335 if (sigma == 0)
336 return mu;
337
338 rnd = get_crandom(state, prng);
339
340 /* default uniform distribution */
341 if (dist == NULL)
342 return ((rnd % (2 * (u32)sigma)) + mu) - sigma;
343
344 t = dist->table[rnd % dist->size];
345 x = (sigma % NETEM_DIST_SCALE) * t;
346 if (x >= 0)
347 x += NETEM_DIST_SCALE/2;
348 else
349 x -= NETEM_DIST_SCALE/2;
350
351 return x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu;
352}
353
354static u64 packet_time_ns(u64 len, const struct netem_sched_data *q)
355{
356 len += q->packet_overhead;
357
358 if (q->cell_size) {
359 u32 cells = reciprocal_divide(len, q->cell_size_reciprocal);
360
361 if (len > cells * q->cell_size) /* extra cell needed for remainder */
362 cells++;
363 len = cells * (q->cell_size + q->cell_overhead);
364 }
365
366 return div64_u64(len * NSEC_PER_SEC, q->rate);
367}
368
369static void tfifo_reset(struct Qdisc *sch)
370{
371 struct netem_sched_data *q = qdisc_priv(sch);
372 struct rb_node *p = rb_first(&q->t_root);
373
374 while (p) {
375 struct sk_buff *skb = rb_to_skb(p);
376
377 p = rb_next(p);
378 rb_erase(&skb->rbnode, &q->t_root);
379 rtnl_kfree_skbs(skb, skb);
380 }
381
382 rtnl_kfree_skbs(q->t_head, q->t_tail);
383 q->t_head = NULL;
384 q->t_tail = NULL;
385}
386
387static void tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
388{
389 struct netem_sched_data *q = qdisc_priv(sch);
390 u64 tnext = netem_skb_cb(nskb)->time_to_send;
391
392 if (!q->t_tail || tnext >= netem_skb_cb(q->t_tail)->time_to_send) {
393 if (q->t_tail)
394 q->t_tail->next = nskb;
395 else
396 q->t_head = nskb;
397 q->t_tail = nskb;
398 } else {
399 struct rb_node **p = &q->t_root.rb_node, *parent = NULL;
400
401 while (*p) {
402 struct sk_buff *skb;
403
404 parent = *p;
405 skb = rb_to_skb(parent);
406 if (tnext >= netem_skb_cb(skb)->time_to_send)
407 p = &parent->rb_right;
408 else
409 p = &parent->rb_left;
410 }
411 rb_link_node(&nskb->rbnode, parent, p);
412 rb_insert_color(&nskb->rbnode, &q->t_root);
413 }
414 sch->q.qlen++;
415}
416
417/* netem can't properly corrupt a megapacket (like we get from GSO), so instead
418 * when we statistically choose to corrupt one, we instead segment it, returning
419 * the first packet to be corrupted, and re-enqueue the remaining frames
420 */
421static struct sk_buff *netem_segment(struct sk_buff *skb, struct Qdisc *sch,
422 struct sk_buff **to_free)
423{
424 struct sk_buff *segs;
425 netdev_features_t features = netif_skb_features(skb);
426
427 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
428
429 if (IS_ERR_OR_NULL(segs)) {
430 qdisc_drop(skb, sch, to_free);
431 return NULL;
432 }
433 consume_skb(skb);
434 return segs;
435}
436
437/*
438 * Insert one skb into qdisc.
439 * Note: parent depends on return value to account for queue length.
440 * NET_XMIT_DROP: queue length didn't change.
441 * NET_XMIT_SUCCESS: one skb was queued.
442 */
443static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch,
444 struct sk_buff **to_free)
445{
446 struct netem_sched_data *q = qdisc_priv(sch);
447 /* We don't fill cb now as skb_unshare() may invalidate it */
448 struct netem_skb_cb *cb;
449 struct sk_buff *skb2;
450 struct sk_buff *segs = NULL;
451 unsigned int prev_len = qdisc_pkt_len(skb);
452 int count = 1;
453 int rc = NET_XMIT_SUCCESS;
454 int rc_drop = NET_XMIT_DROP;
455
456 /* Do not fool qdisc_drop_all() */
457 skb->prev = NULL;
458
459 /* Random duplication */
460 if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor, &q->prng))
461 ++count;
462
463 /* Drop packet? */
464 if (loss_event(q)) {
465 if (q->ecn && INET_ECN_set_ce(skb))
466 qdisc_qstats_drop(sch); /* mark packet */
467 else
468 --count;
469 }
470 if (count == 0) {
471 qdisc_qstats_drop(sch);
472 __qdisc_drop(skb, to_free);
473 return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
474 }
475
476 /* If a delay is expected, orphan the skb. (orphaning usually takes
477 * place at TX completion time, so _before_ the link transit delay)
478 */
479 if (q->latency || q->jitter || q->rate)
480 skb_orphan_partial(skb);
481
482 /*
483 * If we need to duplicate packet, then re-insert at top of the
484 * qdisc tree, since parent queuer expects that only one
485 * skb will be queued.
486 */
487 if (count > 1 && (skb2 = skb_clone(skb, GFP_ATOMIC)) != NULL) {
488 struct Qdisc *rootq = qdisc_root_bh(sch);
489 u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
490
491 q->duplicate = 0;
492 rootq->enqueue(skb2, rootq, to_free);
493 q->duplicate = dupsave;
494 rc_drop = NET_XMIT_SUCCESS;
495 }
496
497 /*
498 * Randomized packet corruption.
499 * Make copy if needed since we are modifying
500 * If packet is going to be hardware checksummed, then
501 * do it now in software before we mangle it.
502 */
503 if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor, &q->prng)) {
504 if (skb_is_gso(skb)) {
505 skb = netem_segment(skb, sch, to_free);
506 if (!skb)
507 return rc_drop;
508 segs = skb->next;
509 skb_mark_not_on_list(skb);
510 qdisc_skb_cb(skb)->pkt_len = skb->len;
511 }
512
513 skb = skb_unshare(skb, GFP_ATOMIC);
514 if (unlikely(!skb)) {
515 qdisc_qstats_drop(sch);
516 goto finish_segs;
517 }
518 if (skb->ip_summed == CHECKSUM_PARTIAL &&
519 skb_checksum_help(skb)) {
520 qdisc_drop(skb, sch, to_free);
521 skb = NULL;
522 goto finish_segs;
523 }
524
525 skb->data[get_random_u32_below(skb_headlen(skb))] ^=
526 1<<get_random_u32_below(8);
527 }
528
529 if (unlikely(sch->q.qlen >= sch->limit)) {
530 /* re-link segs, so that qdisc_drop_all() frees them all */
531 skb->next = segs;
532 qdisc_drop_all(skb, sch, to_free);
533 return rc_drop;
534 }
535
536 qdisc_qstats_backlog_inc(sch, skb);
537
538 cb = netem_skb_cb(skb);
539 if (q->gap == 0 || /* not doing reordering */
540 q->counter < q->gap - 1 || /* inside last reordering gap */
541 q->reorder < get_crandom(&q->reorder_cor, &q->prng)) {
542 u64 now;
543 s64 delay;
544
545 delay = tabledist(q->latency, q->jitter,
546 &q->delay_cor, &q->prng, q->delay_dist);
547
548 now = ktime_get_ns();
549
550 if (q->rate) {
551 struct netem_skb_cb *last = NULL;
552
553 if (sch->q.tail)
554 last = netem_skb_cb(sch->q.tail);
555 if (q->t_root.rb_node) {
556 struct sk_buff *t_skb;
557 struct netem_skb_cb *t_last;
558
559 t_skb = skb_rb_last(&q->t_root);
560 t_last = netem_skb_cb(t_skb);
561 if (!last ||
562 t_last->time_to_send > last->time_to_send)
563 last = t_last;
564 }
565 if (q->t_tail) {
566 struct netem_skb_cb *t_last =
567 netem_skb_cb(q->t_tail);
568
569 if (!last ||
570 t_last->time_to_send > last->time_to_send)
571 last = t_last;
572 }
573
574 if (last) {
575 /*
576 * Last packet in queue is reference point (now),
577 * calculate this time bonus and subtract
578 * from delay.
579 */
580 delay -= last->time_to_send - now;
581 delay = max_t(s64, 0, delay);
582 now = last->time_to_send;
583 }
584
585 delay += packet_time_ns(qdisc_pkt_len(skb), q);
586 }
587
588 cb->time_to_send = now + delay;
589 ++q->counter;
590 tfifo_enqueue(skb, sch);
591 } else {
592 /*
593 * Do re-ordering by putting one out of N packets at the front
594 * of the queue.
595 */
596 cb->time_to_send = ktime_get_ns();
597 q->counter = 0;
598
599 __qdisc_enqueue_head(skb, &sch->q);
600 sch->qstats.requeues++;
601 }
602
603finish_segs:
604 if (segs) {
605 unsigned int len, last_len;
606 int nb;
607
608 len = skb ? skb->len : 0;
609 nb = skb ? 1 : 0;
610
611 while (segs) {
612 skb2 = segs->next;
613 skb_mark_not_on_list(segs);
614 qdisc_skb_cb(segs)->pkt_len = segs->len;
615 last_len = segs->len;
616 rc = qdisc_enqueue(segs, sch, to_free);
617 if (rc != NET_XMIT_SUCCESS) {
618 if (net_xmit_drop_count(rc))
619 qdisc_qstats_drop(sch);
620 } else {
621 nb++;
622 len += last_len;
623 }
624 segs = skb2;
625 }
626 /* Parent qdiscs accounted for 1 skb of size @prev_len */
627 qdisc_tree_reduce_backlog(sch, -(nb - 1), -(len - prev_len));
628 } else if (!skb) {
629 return NET_XMIT_DROP;
630 }
631 return NET_XMIT_SUCCESS;
632}
633
634/* Delay the next round with a new future slot with a
635 * correct number of bytes and packets.
636 */
637
638static void get_slot_next(struct netem_sched_data *q, u64 now)
639{
640 s64 next_delay;
641
642 if (!q->slot_dist)
643 next_delay = q->slot_config.min_delay +
644 (get_random_u32() *
645 (q->slot_config.max_delay -
646 q->slot_config.min_delay) >> 32);
647 else
648 next_delay = tabledist(q->slot_config.dist_delay,
649 (s32)(q->slot_config.dist_jitter),
650 NULL, &q->prng, q->slot_dist);
651
652 q->slot.slot_next = now + next_delay;
653 q->slot.packets_left = q->slot_config.max_packets;
654 q->slot.bytes_left = q->slot_config.max_bytes;
655}
656
657static struct sk_buff *netem_peek(struct netem_sched_data *q)
658{
659 struct sk_buff *skb = skb_rb_first(&q->t_root);
660 u64 t1, t2;
661
662 if (!skb)
663 return q->t_head;
664 if (!q->t_head)
665 return skb;
666
667 t1 = netem_skb_cb(skb)->time_to_send;
668 t2 = netem_skb_cb(q->t_head)->time_to_send;
669 if (t1 < t2)
670 return skb;
671 return q->t_head;
672}
673
674static void netem_erase_head(struct netem_sched_data *q, struct sk_buff *skb)
675{
676 if (skb == q->t_head) {
677 q->t_head = skb->next;
678 if (!q->t_head)
679 q->t_tail = NULL;
680 } else {
681 rb_erase(&skb->rbnode, &q->t_root);
682 }
683}
684
685static struct sk_buff *netem_dequeue(struct Qdisc *sch)
686{
687 struct netem_sched_data *q = qdisc_priv(sch);
688 struct sk_buff *skb;
689
690tfifo_dequeue:
691 skb = __qdisc_dequeue_head(&sch->q);
692 if (skb) {
693 qdisc_qstats_backlog_dec(sch, skb);
694deliver:
695 qdisc_bstats_update(sch, skb);
696 return skb;
697 }
698 skb = netem_peek(q);
699 if (skb) {
700 u64 time_to_send;
701 u64 now = ktime_get_ns();
702
703 /* if more time remaining? */
704 time_to_send = netem_skb_cb(skb)->time_to_send;
705 if (q->slot.slot_next && q->slot.slot_next < time_to_send)
706 get_slot_next(q, now);
707
708 if (time_to_send <= now && q->slot.slot_next <= now) {
709 netem_erase_head(q, skb);
710 sch->q.qlen--;
711 qdisc_qstats_backlog_dec(sch, skb);
712 skb->next = NULL;
713 skb->prev = NULL;
714 /* skb->dev shares skb->rbnode area,
715 * we need to restore its value.
716 */
717 skb->dev = qdisc_dev(sch);
718
719 if (q->slot.slot_next) {
720 q->slot.packets_left--;
721 q->slot.bytes_left -= qdisc_pkt_len(skb);
722 if (q->slot.packets_left <= 0 ||
723 q->slot.bytes_left <= 0)
724 get_slot_next(q, now);
725 }
726
727 if (q->qdisc) {
728 unsigned int pkt_len = qdisc_pkt_len(skb);
729 struct sk_buff *to_free = NULL;
730 int err;
731
732 err = qdisc_enqueue(skb, q->qdisc, &to_free);
733 kfree_skb_list(to_free);
734 if (err != NET_XMIT_SUCCESS &&
735 net_xmit_drop_count(err)) {
736 qdisc_qstats_drop(sch);
737 qdisc_tree_reduce_backlog(sch, 1,
738 pkt_len);
739 }
740 goto tfifo_dequeue;
741 }
742 goto deliver;
743 }
744
745 if (q->qdisc) {
746 skb = q->qdisc->ops->dequeue(q->qdisc);
747 if (skb)
748 goto deliver;
749 }
750
751 qdisc_watchdog_schedule_ns(&q->watchdog,
752 max(time_to_send,
753 q->slot.slot_next));
754 }
755
756 if (q->qdisc) {
757 skb = q->qdisc->ops->dequeue(q->qdisc);
758 if (skb)
759 goto deliver;
760 }
761 return NULL;
762}
763
764static void netem_reset(struct Qdisc *sch)
765{
766 struct netem_sched_data *q = qdisc_priv(sch);
767
768 qdisc_reset_queue(sch);
769 tfifo_reset(sch);
770 if (q->qdisc)
771 qdisc_reset(q->qdisc);
772 qdisc_watchdog_cancel(&q->watchdog);
773}
774
775static void dist_free(struct disttable *d)
776{
777 kvfree(d);
778}
779
780/*
781 * Distribution data is a variable size payload containing
782 * signed 16 bit values.
783 */
784
785static int get_dist_table(struct disttable **tbl, const struct nlattr *attr)
786{
787 size_t n = nla_len(attr)/sizeof(__s16);
788 const __s16 *data = nla_data(attr);
789 struct disttable *d;
790 int i;
791
792 if (!n || n > NETEM_DIST_MAX)
793 return -EINVAL;
794
795 d = kvmalloc(struct_size(d, table, n), GFP_KERNEL);
796 if (!d)
797 return -ENOMEM;
798
799 d->size = n;
800 for (i = 0; i < n; i++)
801 d->table[i] = data[i];
802
803 *tbl = d;
804 return 0;
805}
806
807static void get_slot(struct netem_sched_data *q, const struct nlattr *attr)
808{
809 const struct tc_netem_slot *c = nla_data(attr);
810
811 q->slot_config = *c;
812 if (q->slot_config.max_packets == 0)
813 q->slot_config.max_packets = INT_MAX;
814 if (q->slot_config.max_bytes == 0)
815 q->slot_config.max_bytes = INT_MAX;
816
817 /* capping dist_jitter to the range acceptable by tabledist() */
818 q->slot_config.dist_jitter = min_t(__s64, INT_MAX, abs(q->slot_config.dist_jitter));
819
820 q->slot.packets_left = q->slot_config.max_packets;
821 q->slot.bytes_left = q->slot_config.max_bytes;
822 if (q->slot_config.min_delay | q->slot_config.max_delay |
823 q->slot_config.dist_jitter)
824 q->slot.slot_next = ktime_get_ns();
825 else
826 q->slot.slot_next = 0;
827}
828
829static void get_correlation(struct netem_sched_data *q, const struct nlattr *attr)
830{
831 const struct tc_netem_corr *c = nla_data(attr);
832
833 init_crandom(&q->delay_cor, c->delay_corr);
834 init_crandom(&q->loss_cor, c->loss_corr);
835 init_crandom(&q->dup_cor, c->dup_corr);
836}
837
838static void get_reorder(struct netem_sched_data *q, const struct nlattr *attr)
839{
840 const struct tc_netem_reorder *r = nla_data(attr);
841
842 q->reorder = r->probability;
843 init_crandom(&q->reorder_cor, r->correlation);
844}
845
846static void get_corrupt(struct netem_sched_data *q, const struct nlattr *attr)
847{
848 const struct tc_netem_corrupt *r = nla_data(attr);
849
850 q->corrupt = r->probability;
851 init_crandom(&q->corrupt_cor, r->correlation);
852}
853
854static void get_rate(struct netem_sched_data *q, const struct nlattr *attr)
855{
856 const struct tc_netem_rate *r = nla_data(attr);
857
858 q->rate = r->rate;
859 q->packet_overhead = r->packet_overhead;
860 q->cell_size = r->cell_size;
861 q->cell_overhead = r->cell_overhead;
862 if (q->cell_size)
863 q->cell_size_reciprocal = reciprocal_value(q->cell_size);
864 else
865 q->cell_size_reciprocal = (struct reciprocal_value) { 0 };
866}
867
868static int get_loss_clg(struct netem_sched_data *q, const struct nlattr *attr)
869{
870 const struct nlattr *la;
871 int rem;
872
873 nla_for_each_nested(la, attr, rem) {
874 u16 type = nla_type(la);
875
876 switch (type) {
877 case NETEM_LOSS_GI: {
878 const struct tc_netem_gimodel *gi = nla_data(la);
879
880 if (nla_len(la) < sizeof(struct tc_netem_gimodel)) {
881 pr_info("netem: incorrect gi model size\n");
882 return -EINVAL;
883 }
884
885 q->loss_model = CLG_4_STATES;
886
887 q->clg.state = TX_IN_GAP_PERIOD;
888 q->clg.a1 = gi->p13;
889 q->clg.a2 = gi->p31;
890 q->clg.a3 = gi->p32;
891 q->clg.a4 = gi->p14;
892 q->clg.a5 = gi->p23;
893 break;
894 }
895
896 case NETEM_LOSS_GE: {
897 const struct tc_netem_gemodel *ge = nla_data(la);
898
899 if (nla_len(la) < sizeof(struct tc_netem_gemodel)) {
900 pr_info("netem: incorrect ge model size\n");
901 return -EINVAL;
902 }
903
904 q->loss_model = CLG_GILB_ELL;
905 q->clg.state = GOOD_STATE;
906 q->clg.a1 = ge->p;
907 q->clg.a2 = ge->r;
908 q->clg.a3 = ge->h;
909 q->clg.a4 = ge->k1;
910 break;
911 }
912
913 default:
914 pr_info("netem: unknown loss type %u\n", type);
915 return -EINVAL;
916 }
917 }
918
919 return 0;
920}
921
922static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
923 [TCA_NETEM_CORR] = { .len = sizeof(struct tc_netem_corr) },
924 [TCA_NETEM_REORDER] = { .len = sizeof(struct tc_netem_reorder) },
925 [TCA_NETEM_CORRUPT] = { .len = sizeof(struct tc_netem_corrupt) },
926 [TCA_NETEM_RATE] = { .len = sizeof(struct tc_netem_rate) },
927 [TCA_NETEM_LOSS] = { .type = NLA_NESTED },
928 [TCA_NETEM_ECN] = { .type = NLA_U32 },
929 [TCA_NETEM_RATE64] = { .type = NLA_U64 },
930 [TCA_NETEM_LATENCY64] = { .type = NLA_S64 },
931 [TCA_NETEM_JITTER64] = { .type = NLA_S64 },
932 [TCA_NETEM_SLOT] = { .len = sizeof(struct tc_netem_slot) },
933 [TCA_NETEM_PRNG_SEED] = { .type = NLA_U64 },
934};
935
936static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
937 const struct nla_policy *policy, int len)
938{
939 int nested_len = nla_len(nla) - NLA_ALIGN(len);
940
941 if (nested_len < 0) {
942 pr_info("netem: invalid attributes len %d\n", nested_len);
943 return -EINVAL;
944 }
945
946 if (nested_len >= nla_attr_size(0))
947 return nla_parse_deprecated(tb, maxtype,
948 nla_data(nla) + NLA_ALIGN(len),
949 nested_len, policy, NULL);
950
951 memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
952 return 0;
953}
954
955/* Parse netlink message to set options */
956static int netem_change(struct Qdisc *sch, struct nlattr *opt,
957 struct netlink_ext_ack *extack)
958{
959 struct netem_sched_data *q = qdisc_priv(sch);
960 struct nlattr *tb[TCA_NETEM_MAX + 1];
961 struct disttable *delay_dist = NULL;
962 struct disttable *slot_dist = NULL;
963 struct tc_netem_qopt *qopt;
964 struct clgstate old_clg;
965 int old_loss_model = CLG_RANDOM;
966 int ret;
967
968 qopt = nla_data(opt);
969 ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
970 if (ret < 0)
971 return ret;
972
973 if (tb[TCA_NETEM_DELAY_DIST]) {
974 ret = get_dist_table(&delay_dist, tb[TCA_NETEM_DELAY_DIST]);
975 if (ret)
976 goto table_free;
977 }
978
979 if (tb[TCA_NETEM_SLOT_DIST]) {
980 ret = get_dist_table(&slot_dist, tb[TCA_NETEM_SLOT_DIST]);
981 if (ret)
982 goto table_free;
983 }
984
985 sch_tree_lock(sch);
986 /* backup q->clg and q->loss_model */
987 old_clg = q->clg;
988 old_loss_model = q->loss_model;
989
990 if (tb[TCA_NETEM_LOSS]) {
991 ret = get_loss_clg(q, tb[TCA_NETEM_LOSS]);
992 if (ret) {
993 q->loss_model = old_loss_model;
994 q->clg = old_clg;
995 goto unlock;
996 }
997 } else {
998 q->loss_model = CLG_RANDOM;
999 }
1000
1001 if (delay_dist)
1002 swap(q->delay_dist, delay_dist);
1003 if (slot_dist)
1004 swap(q->slot_dist, slot_dist);
1005 sch->limit = qopt->limit;
1006
1007 q->latency = PSCHED_TICKS2NS(qopt->latency);
1008 q->jitter = PSCHED_TICKS2NS(qopt->jitter);
1009 q->limit = qopt->limit;
1010 q->gap = qopt->gap;
1011 q->counter = 0;
1012 q->loss = qopt->loss;
1013 q->duplicate = qopt->duplicate;
1014
1015 /* for compatibility with earlier versions.
1016 * if gap is set, need to assume 100% probability
1017 */
1018 if (q->gap)
1019 q->reorder = ~0;
1020
1021 if (tb[TCA_NETEM_CORR])
1022 get_correlation(q, tb[TCA_NETEM_CORR]);
1023
1024 if (tb[TCA_NETEM_REORDER])
1025 get_reorder(q, tb[TCA_NETEM_REORDER]);
1026
1027 if (tb[TCA_NETEM_CORRUPT])
1028 get_corrupt(q, tb[TCA_NETEM_CORRUPT]);
1029
1030 if (tb[TCA_NETEM_RATE])
1031 get_rate(q, tb[TCA_NETEM_RATE]);
1032
1033 if (tb[TCA_NETEM_RATE64])
1034 q->rate = max_t(u64, q->rate,
1035 nla_get_u64(tb[TCA_NETEM_RATE64]));
1036
1037 if (tb[TCA_NETEM_LATENCY64])
1038 q->latency = nla_get_s64(tb[TCA_NETEM_LATENCY64]);
1039
1040 if (tb[TCA_NETEM_JITTER64])
1041 q->jitter = nla_get_s64(tb[TCA_NETEM_JITTER64]);
1042
1043 if (tb[TCA_NETEM_ECN])
1044 q->ecn = nla_get_u32(tb[TCA_NETEM_ECN]);
1045
1046 if (tb[TCA_NETEM_SLOT])
1047 get_slot(q, tb[TCA_NETEM_SLOT]);
1048
1049 /* capping jitter to the range acceptable by tabledist() */
1050 q->jitter = min_t(s64, abs(q->jitter), INT_MAX);
1051
1052 if (tb[TCA_NETEM_PRNG_SEED])
1053 q->prng.seed = nla_get_u64(tb[TCA_NETEM_PRNG_SEED]);
1054 else
1055 q->prng.seed = get_random_u64();
1056 prandom_seed_state(&q->prng.prng_state, q->prng.seed);
1057
1058unlock:
1059 sch_tree_unlock(sch);
1060
1061table_free:
1062 dist_free(delay_dist);
1063 dist_free(slot_dist);
1064 return ret;
1065}
1066
1067static int netem_init(struct Qdisc *sch, struct nlattr *opt,
1068 struct netlink_ext_ack *extack)
1069{
1070 struct netem_sched_data *q = qdisc_priv(sch);
1071 int ret;
1072
1073 qdisc_watchdog_init(&q->watchdog, sch);
1074
1075 if (!opt)
1076 return -EINVAL;
1077
1078 q->loss_model = CLG_RANDOM;
1079 ret = netem_change(sch, opt, extack);
1080 if (ret)
1081 pr_info("netem: change failed\n");
1082 return ret;
1083}
1084
1085static void netem_destroy(struct Qdisc *sch)
1086{
1087 struct netem_sched_data *q = qdisc_priv(sch);
1088
1089 qdisc_watchdog_cancel(&q->watchdog);
1090 if (q->qdisc)
1091 qdisc_put(q->qdisc);
1092 dist_free(q->delay_dist);
1093 dist_free(q->slot_dist);
1094}
1095
1096static int dump_loss_model(const struct netem_sched_data *q,
1097 struct sk_buff *skb)
1098{
1099 struct nlattr *nest;
1100
1101 nest = nla_nest_start_noflag(skb, TCA_NETEM_LOSS);
1102 if (nest == NULL)
1103 goto nla_put_failure;
1104
1105 switch (q->loss_model) {
1106 case CLG_RANDOM:
1107 /* legacy loss model */
1108 nla_nest_cancel(skb, nest);
1109 return 0; /* no data */
1110
1111 case CLG_4_STATES: {
1112 struct tc_netem_gimodel gi = {
1113 .p13 = q->clg.a1,
1114 .p31 = q->clg.a2,
1115 .p32 = q->clg.a3,
1116 .p14 = q->clg.a4,
1117 .p23 = q->clg.a5,
1118 };
1119
1120 if (nla_put(skb, NETEM_LOSS_GI, sizeof(gi), &gi))
1121 goto nla_put_failure;
1122 break;
1123 }
1124 case CLG_GILB_ELL: {
1125 struct tc_netem_gemodel ge = {
1126 .p = q->clg.a1,
1127 .r = q->clg.a2,
1128 .h = q->clg.a3,
1129 .k1 = q->clg.a4,
1130 };
1131
1132 if (nla_put(skb, NETEM_LOSS_GE, sizeof(ge), &ge))
1133 goto nla_put_failure;
1134 break;
1135 }
1136 }
1137
1138 nla_nest_end(skb, nest);
1139 return 0;
1140
1141nla_put_failure:
1142 nla_nest_cancel(skb, nest);
1143 return -1;
1144}
1145
1146static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
1147{
1148 const struct netem_sched_data *q = qdisc_priv(sch);
1149 struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb);
1150 struct tc_netem_qopt qopt;
1151 struct tc_netem_corr cor;
1152 struct tc_netem_reorder reorder;
1153 struct tc_netem_corrupt corrupt;
1154 struct tc_netem_rate rate;
1155 struct tc_netem_slot slot;
1156
1157 qopt.latency = min_t(psched_time_t, PSCHED_NS2TICKS(q->latency),
1158 UINT_MAX);
1159 qopt.jitter = min_t(psched_time_t, PSCHED_NS2TICKS(q->jitter),
1160 UINT_MAX);
1161 qopt.limit = q->limit;
1162 qopt.loss = q->loss;
1163 qopt.gap = q->gap;
1164 qopt.duplicate = q->duplicate;
1165 if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
1166 goto nla_put_failure;
1167
1168 if (nla_put(skb, TCA_NETEM_LATENCY64, sizeof(q->latency), &q->latency))
1169 goto nla_put_failure;
1170
1171 if (nla_put(skb, TCA_NETEM_JITTER64, sizeof(q->jitter), &q->jitter))
1172 goto nla_put_failure;
1173
1174 cor.delay_corr = q->delay_cor.rho;
1175 cor.loss_corr = q->loss_cor.rho;
1176 cor.dup_corr = q->dup_cor.rho;
1177 if (nla_put(skb, TCA_NETEM_CORR, sizeof(cor), &cor))
1178 goto nla_put_failure;
1179
1180 reorder.probability = q->reorder;
1181 reorder.correlation = q->reorder_cor.rho;
1182 if (nla_put(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder))
1183 goto nla_put_failure;
1184
1185 corrupt.probability = q->corrupt;
1186 corrupt.correlation = q->corrupt_cor.rho;
1187 if (nla_put(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt))
1188 goto nla_put_failure;
1189
1190 if (q->rate >= (1ULL << 32)) {
1191 if (nla_put_u64_64bit(skb, TCA_NETEM_RATE64, q->rate,
1192 TCA_NETEM_PAD))
1193 goto nla_put_failure;
1194 rate.rate = ~0U;
1195 } else {
1196 rate.rate = q->rate;
1197 }
1198 rate.packet_overhead = q->packet_overhead;
1199 rate.cell_size = q->cell_size;
1200 rate.cell_overhead = q->cell_overhead;
1201 if (nla_put(skb, TCA_NETEM_RATE, sizeof(rate), &rate))
1202 goto nla_put_failure;
1203
1204 if (q->ecn && nla_put_u32(skb, TCA_NETEM_ECN, q->ecn))
1205 goto nla_put_failure;
1206
1207 if (dump_loss_model(q, skb) != 0)
1208 goto nla_put_failure;
1209
1210 if (q->slot_config.min_delay | q->slot_config.max_delay |
1211 q->slot_config.dist_jitter) {
1212 slot = q->slot_config;
1213 if (slot.max_packets == INT_MAX)
1214 slot.max_packets = 0;
1215 if (slot.max_bytes == INT_MAX)
1216 slot.max_bytes = 0;
1217 if (nla_put(skb, TCA_NETEM_SLOT, sizeof(slot), &slot))
1218 goto nla_put_failure;
1219 }
1220
1221 if (nla_put_u64_64bit(skb, TCA_NETEM_PRNG_SEED, q->prng.seed,
1222 TCA_NETEM_PAD))
1223 goto nla_put_failure;
1224
1225 return nla_nest_end(skb, nla);
1226
1227nla_put_failure:
1228 nlmsg_trim(skb, nla);
1229 return -1;
1230}
1231
1232static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
1233 struct sk_buff *skb, struct tcmsg *tcm)
1234{
1235 struct netem_sched_data *q = qdisc_priv(sch);
1236
1237 if (cl != 1 || !q->qdisc) /* only one class */
1238 return -ENOENT;
1239
1240 tcm->tcm_handle |= TC_H_MIN(1);
1241 tcm->tcm_info = q->qdisc->handle;
1242
1243 return 0;
1244}
1245
1246static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1247 struct Qdisc **old, struct netlink_ext_ack *extack)
1248{
1249 struct netem_sched_data *q = qdisc_priv(sch);
1250
1251 *old = qdisc_replace(sch, new, &q->qdisc);
1252 return 0;
1253}
1254
1255static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
1256{
1257 struct netem_sched_data *q = qdisc_priv(sch);
1258 return q->qdisc;
1259}
1260
1261static unsigned long netem_find(struct Qdisc *sch, u32 classid)
1262{
1263 return 1;
1264}
1265
1266static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
1267{
1268 if (!walker->stop) {
1269 if (!tc_qdisc_stats_dump(sch, 1, walker))
1270 return;
1271 }
1272}
1273
1274static const struct Qdisc_class_ops netem_class_ops = {
1275 .graft = netem_graft,
1276 .leaf = netem_leaf,
1277 .find = netem_find,
1278 .walk = netem_walk,
1279 .dump = netem_dump_class,
1280};
1281
1282static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
1283 .id = "netem",
1284 .cl_ops = &netem_class_ops,
1285 .priv_size = sizeof(struct netem_sched_data),
1286 .enqueue = netem_enqueue,
1287 .dequeue = netem_dequeue,
1288 .peek = qdisc_peek_dequeued,
1289 .init = netem_init,
1290 .reset = netem_reset,
1291 .destroy = netem_destroy,
1292 .change = netem_change,
1293 .dump = netem_dump,
1294 .owner = THIS_MODULE,
1295};
1296
1297
1298static int __init netem_module_init(void)
1299{
1300 pr_info("netem: version " VERSION "\n");
1301 return register_qdisc(&netem_qdisc_ops);
1302}
1303static void __exit netem_module_exit(void)
1304{
1305 unregister_qdisc(&netem_qdisc_ops);
1306}
1307module_init(netem_module_init)
1308module_exit(netem_module_exit)
1309MODULE_LICENSE("GPL");
1310MODULE_DESCRIPTION("Network characteristics emulator qdisc");