projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
netdev: Create netdev_queue abstraction.
[deliverable/linux.git] / net / sched / sch_netem.c
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/module.h>
17 #include <linux/types.h>
18 #include <linux/kernel.h>
19 #include <linux/errno.h>
20 #include <linux/skbuff.h>
21 #include <linux/rtnetlink.h>
22
23 #include <net/netlink.h>
24 #include <net/pkt_sched.h>
25
26 #define VERSION "1.2"
27
28 /* Network Emulation Queuing algorithm.
29 ====================================
30
31 Sources: [1] Mark Carson, Darrin Santay, "NIST Net - A Linux-based
32 Network Emulation Tool
33 [2] Luigi Rizzo, DummyNet for FreeBSD
34
35 ----------------------------------------------------------------
36
37 This started out as a simple way to delay outgoing packets to
38 test TCP but has grown to include most of the functionality
39 of a full blown network emulator like NISTnet. It can delay
40 packets and add random jitter (and correlation). The random
41 distribution can be loaded from a table as well to provide
42 normal, Pareto, or experimental curves. Packet loss,
43 duplication, and reordering can also be emulated.
44
45 This qdisc does not do classification that can be handled in
46 layering other disciplines. It does not need to do bandwidth
47 control either since that can be handled by using token
48 bucket or other rate control.
49
50 The simulator is limited by the Linux timer resolution
51 and will create packet bursts on the HZ boundary (1ms).
52 */
53
54 struct netem_sched_data {
55 struct Qdisc *qdisc;
56 struct qdisc_watchdog watchdog;
57
58 psched_tdiff_t latency;
59 psched_tdiff_t jitter;
60
61 u32 loss;
62 u32 limit;
63 u32 counter;
64 u32 gap;
65 u32 duplicate;
66 u32 reorder;
67 u32 corrupt;
68
69 struct crndstate {
70 u32 last;
71 u32 rho;
72 } delay_cor, loss_cor, dup_cor, reorder_cor, corrupt_cor;
73
74 struct disttable {
75 u32 size;
76 s16 table[0];
77 } *delay_dist;
78 };
79
80 /* Time stamp put into socket buffer control block */
81 struct netem_skb_cb {
82 psched_time_t time_to_send;
83 };
84
85 /* init_crandom - initialize correlated random number generator
86 * Use entropy source for initial seed.
87 */
88 static void init_crandom(struct crndstate *state, unsigned long rho)
89 {
90 state->rho = rho;
91 state->last = net_random();
92 }
93
94 /* get_crandom - correlated random number generator
95 * Next number depends on last value.
96 * rho is scaled to avoid floating point.
97 */
98 static u32 get_crandom(struct crndstate *state)
99 {
100 u64 value, rho;
101 unsigned long answer;
102
103 if (state->rho == 0) /* no correlation */
104 return net_random();
105
106 value = net_random();
107 rho = (u64)state->rho + 1;
108 answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32;
109 state->last = answer;
110 return answer;
111 }
112
113 /* tabledist - return a pseudo-randomly distributed value with mean mu and
114 * std deviation sigma. Uses table lookup to approximate the desired
115 * distribution, and a uniformly-distributed pseudo-random source.
116 */
117 static psched_tdiff_t tabledist(psched_tdiff_t mu, psched_tdiff_t sigma,
118 struct crndstate *state,
119 const struct disttable *dist)
120 {
121 psched_tdiff_t x;
122 long t;
123 u32 rnd;
124
125 if (sigma == 0)
126 return mu;
127
128 rnd = get_crandom(state);
129
130 /* default uniform distribution */
131 if (dist == NULL)
132 return (rnd % (2*sigma)) - sigma + mu;
133
134 t = dist->table[rnd % dist->size];
135 x = (sigma % NETEM_DIST_SCALE) * t;
136 if (x >= 0)
137 x += NETEM_DIST_SCALE/2;
138 else
139 x -= NETEM_DIST_SCALE/2;
140
141 return x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu;
142 }
143
144 /*
145 * Insert one skb into qdisc.
146 * Note: parent depends on return value to account for queue length.
147 * NET_XMIT_DROP: queue length didn't change.
148 * NET_XMIT_SUCCESS: one skb was queued.
149 */
150 static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch)
151 {
152 struct netem_sched_data *q = qdisc_priv(sch);
153 /* We don't fill cb now as skb_unshare() may invalidate it */
154 struct netem_skb_cb *cb;
155 struct sk_buff *skb2;
156 int ret;
157 int count = 1;
158
159 pr_debug("netem_enqueue skb=%p\n", skb);
160
161 /* Random duplication */
162 if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor))
163 ++count;
164
165 /* Random packet drop 0 => none, ~0 => all */
166 if (q->loss && q->loss >= get_crandom(&q->loss_cor))
167 --count;
168
169 if (count == 0) {
170 sch->qstats.drops++;
171 kfree_skb(skb);
172 return NET_XMIT_BYPASS;
173 }
174
175 skb_orphan(skb);
176
177 /*
178 * If we need to duplicate packet, then re-insert at top of the
179 * qdisc tree, since parent queuer expects that only one
180 * skb will be queued.
181 */
182 if (count > 1 && (skb2 = skb_clone(skb, GFP_ATOMIC)) != NULL) {
183 struct Qdisc *rootq = sch->dev->qdisc;
184 u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
185 q->duplicate = 0;
186
187 rootq->enqueue(skb2, rootq);
188 q->duplicate = dupsave;
189 }
190
191 /*
192 * Randomized packet corruption.
193 * Make copy if needed since we are modifying
194 * If packet is going to be hardware checksummed, then
195 * do it now in software before we mangle it.
196 */
197 if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor)) {
198 if (!(skb = skb_unshare(skb, GFP_ATOMIC))
199 || (skb->ip_summed == CHECKSUM_PARTIAL
200 && skb_checksum_help(skb))) {
201 sch->qstats.drops++;
202 return NET_XMIT_DROP;
203 }
204
205 skb->data[net_random() % skb_headlen(skb)] ^= 1<<(net_random() % 8);
206 }
207
208 cb = (struct netem_skb_cb *)skb->cb;
209 if (q->gap == 0 /* not doing reordering */
210 || q->counter < q->gap /* inside last reordering gap */
211 || q->reorder < get_crandom(&q->reorder_cor)) {
212 psched_time_t now;
213 psched_tdiff_t delay;
214
215 delay = tabledist(q->latency, q->jitter,
216 &q->delay_cor, q->delay_dist);
217
218 now = psched_get_time();
219 cb->time_to_send = now + delay;
220 ++q->counter;
221 ret = q->qdisc->enqueue(skb, q->qdisc);
222 } else {
223 /*
224 * Do re-ordering by putting one out of N packets at the front
225 * of the queue.
226 */
227 cb->time_to_send = psched_get_time();
228 q->counter = 0;
229 ret = q->qdisc->ops->requeue(skb, q->qdisc);
230 }
231
232 if (likely(ret == NET_XMIT_SUCCESS)) {
233 sch->q.qlen++;
234 sch->bstats.bytes += skb->len;
235 sch->bstats.packets++;
236 } else
237 sch->qstats.drops++;
238
239 pr_debug("netem: enqueue ret %d\n", ret);
240 return ret;
241 }
242
243 /* Requeue packets but don't change time stamp */
244 static int netem_requeue(struct sk_buff *skb, struct Qdisc *sch)
245 {
246 struct netem_sched_data *q = qdisc_priv(sch);
247 int ret;
248
249 if ((ret = q->qdisc->ops->requeue(skb, q->qdisc)) == 0) {
250 sch->q.qlen++;
251 sch->qstats.requeues++;
252 }
253
254 return ret;
255 }
256
257 static unsigned int netem_drop(struct Qdisc* sch)
258 {
259 struct netem_sched_data *q = qdisc_priv(sch);
260 unsigned int len = 0;
261
262 if (q->qdisc->ops->drop && (len = q->qdisc->ops->drop(q->qdisc)) != 0) {
263 sch->q.qlen--;
264 sch->qstats.drops++;
265 }
266 return len;
267 }
268
269 static struct sk_buff *netem_dequeue(struct Qdisc *sch)
270 {
271 struct netem_sched_data *q = qdisc_priv(sch);
272 struct sk_buff *skb;
273
274 smp_mb();
275 if (sch->flags & TCQ_F_THROTTLED)
276 return NULL;
277
278 skb = q->qdisc->dequeue(q->qdisc);
279 if (skb) {
280 const struct netem_skb_cb *cb
281 = (const struct netem_skb_cb *)skb->cb;
282 psched_time_t now = psched_get_time();
283
284 /* if more time remaining? */
285 if (cb->time_to_send <= now) {
286 pr_debug("netem_dequeue: return skb=%p\n", skb);
287 sch->q.qlen--;
288 return skb;
289 }
290
291 if (unlikely(q->qdisc->ops->requeue(skb, q->qdisc) != NET_XMIT_SUCCESS)) {
292 qdisc_tree_decrease_qlen(q->qdisc, 1);
293 sch->qstats.drops++;
294 printk(KERN_ERR "netem: %s could not requeue\n",
295 q->qdisc->ops->id);
296 }
297
298 qdisc_watchdog_schedule(&q->watchdog, cb->time_to_send);
299 }
300
301 return NULL;
302 }
303
304 static void netem_reset(struct Qdisc *sch)
305 {
306 struct netem_sched_data *q = qdisc_priv(sch);
307
308 qdisc_reset(q->qdisc);
309 sch->q.qlen = 0;
310 qdisc_watchdog_cancel(&q->watchdog);
311 }
312
313 /*
314 * Distribution data is a variable size payload containing
315 * signed 16 bit values.
316 */
317 static int get_dist_table(struct Qdisc *sch, const struct nlattr *attr)
318 {
319 struct netem_sched_data *q = qdisc_priv(sch);
320 unsigned long n = nla_len(attr)/sizeof(__s16);
321 const __s16 *data = nla_data(attr);
322 struct disttable *d;
323 int i;
324
325 if (n > 65536)
326 return -EINVAL;
327
328 d = kmalloc(sizeof(*d) + n*sizeof(d->table[0]), GFP_KERNEL);
329 if (!d)
330 return -ENOMEM;
331
332 d->size = n;
333 for (i = 0; i < n; i++)
334 d->table[i] = data[i];
335
336 spin_lock_bh(&sch->dev->queue_lock);
337 d = xchg(&q->delay_dist, d);
338 spin_unlock_bh(&sch->dev->queue_lock);
339
340 kfree(d);
341 return 0;
342 }
343
344 static int get_correlation(struct Qdisc *sch, const struct nlattr *attr)
345 {
346 struct netem_sched_data *q = qdisc_priv(sch);
347 const struct tc_netem_corr *c = nla_data(attr);
348
349 init_crandom(&q->delay_cor, c->delay_corr);
350 init_crandom(&q->loss_cor, c->loss_corr);
351 init_crandom(&q->dup_cor, c->dup_corr);
352 return 0;
353 }
354
355 static int get_reorder(struct Qdisc *sch, const struct nlattr *attr)
356 {
357 struct netem_sched_data *q = qdisc_priv(sch);
358 const struct tc_netem_reorder *r = nla_data(attr);
359
360 q->reorder = r->probability;
361 init_crandom(&q->reorder_cor, r->correlation);
362 return 0;
363 }
364
365 static int get_corrupt(struct Qdisc *sch, const struct nlattr *attr)
366 {
367 struct netem_sched_data *q = qdisc_priv(sch);
368 const struct tc_netem_corrupt *r = nla_data(attr);
369
370 q->corrupt = r->probability;
371 init_crandom(&q->corrupt_cor, r->correlation);
372 return 0;
373 }
374
375 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
376 [TCA_NETEM_CORR] = { .len = sizeof(struct tc_netem_corr) },
377 [TCA_NETEM_REORDER] = { .len = sizeof(struct tc_netem_reorder) },
378 [TCA_NETEM_CORRUPT] = { .len = sizeof(struct tc_netem_corrupt) },
379 };
380
381 /* Parse netlink message to set options */
382 static int netem_change(struct Qdisc *sch, struct nlattr *opt)
383 {
384 struct netem_sched_data *q = qdisc_priv(sch);
385 struct nlattr *tb[TCA_NETEM_MAX + 1];
386 struct tc_netem_qopt *qopt;
387 int ret;
388
389 if (opt == NULL)
390 return -EINVAL;
391
392 ret = nla_parse_nested_compat(tb, TCA_NETEM_MAX, opt, netem_policy,
393 qopt, sizeof(*qopt));
394 if (ret < 0)
395 return ret;
396
397 ret = fifo_set_limit(q->qdisc, qopt->limit);
398 if (ret) {
399 pr_debug("netem: can't set fifo limit\n");
400 return ret;
401 }
402
403 q->latency = qopt->latency;
404 q->jitter = qopt->jitter;
405 q->limit = qopt->limit;
406 q->gap = qopt->gap;
407 q->counter = 0;
408 q->loss = qopt->loss;
409 q->duplicate = qopt->duplicate;
410
411 /* for compatibility with earlier versions.
412 * if gap is set, need to assume 100% probability
413 */
414 if (q->gap)
415 q->reorder = ~0;
416
417 if (tb[TCA_NETEM_CORR]) {
418 ret = get_correlation(sch, tb[TCA_NETEM_CORR]);
419 if (ret)
420 return ret;
421 }
422
423 if (tb[TCA_NETEM_DELAY_DIST]) {
424 ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST]);
425 if (ret)
426 return ret;
427 }
428
429 if (tb[TCA_NETEM_REORDER]) {
430 ret = get_reorder(sch, tb[TCA_NETEM_REORDER]);
431 if (ret)
432 return ret;
433 }
434
435 if (tb[TCA_NETEM_CORRUPT]) {
436 ret = get_corrupt(sch, tb[TCA_NETEM_CORRUPT]);
437 if (ret)
438 return ret;
439 }
440
441 return 0;
442 }
443
444 /*
445 * Special case version of FIFO queue for use by netem.
446 * It queues in order based on timestamps in skb's
447 */
448 struct fifo_sched_data {
449 u32 limit;
450 psched_time_t oldest;
451 };
452
453 static int tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
454 {
455 struct fifo_sched_data *q = qdisc_priv(sch);
456 struct sk_buff_head *list = &sch->q;
457 psched_time_t tnext = ((struct netem_skb_cb *)nskb->cb)->time_to_send;
458 struct sk_buff *skb;
459
460 if (likely(skb_queue_len(list) < q->limit)) {
461 /* Optimize for add at tail */
462 if (likely(skb_queue_empty(list) || tnext >= q->oldest)) {
463 q->oldest = tnext;
464 return qdisc_enqueue_tail(nskb, sch);
465 }
466
467 skb_queue_reverse_walk(list, skb) {
468 const struct netem_skb_cb *cb
469 = (const struct netem_skb_cb *)skb->cb;
470
471 if (tnext >= cb->time_to_send)
472 break;
473 }
474
475 __skb_queue_after(list, skb, nskb);
476
477 sch->qstats.backlog += nskb->len;
478 sch->bstats.bytes += nskb->len;
479 sch->bstats.packets++;
480
481 return NET_XMIT_SUCCESS;
482 }
483
484 return qdisc_reshape_fail(nskb, sch);
485 }
486
487 static int tfifo_init(struct Qdisc *sch, struct nlattr *opt)
488 {
489 struct fifo_sched_data *q = qdisc_priv(sch);
490
491 if (opt) {
492 struct tc_fifo_qopt *ctl = nla_data(opt);
493 if (nla_len(opt) < sizeof(*ctl))
494 return -EINVAL;
495
496 q->limit = ctl->limit;
497 } else
498 q->limit = max_t(u32, sch->dev->tx_queue_len, 1);
499
500 q->oldest = PSCHED_PASTPERFECT;
501 return 0;
502 }
503
504 static int tfifo_dump(struct Qdisc *sch, struct sk_buff *skb)
505 {
506 struct fifo_sched_data *q = qdisc_priv(sch);
507 struct tc_fifo_qopt opt = { .limit = q->limit };
508
509 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
510 return skb->len;
511
512 nla_put_failure:
513 return -1;
514 }
515
516 static struct Qdisc_ops tfifo_qdisc_ops __read_mostly = {
517 .id = "tfifo",
518 .priv_size = sizeof(struct fifo_sched_data),
519 .enqueue = tfifo_enqueue,
520 .dequeue = qdisc_dequeue_head,
521 .requeue = qdisc_requeue,
522 .drop = qdisc_queue_drop,
523 .init = tfifo_init,
524 .reset = qdisc_reset_queue,
525 .change = tfifo_init,
526 .dump = tfifo_dump,
527 };
528
529 static int netem_init(struct Qdisc *sch, struct nlattr *opt)
530 {
531 struct netem_sched_data *q = qdisc_priv(sch);
532 int ret;
533
534 if (!opt)
535 return -EINVAL;
536
537 qdisc_watchdog_init(&q->watchdog, sch);
538
539 q->qdisc = qdisc_create_dflt(sch->dev, sch->dev_queue,
540 &tfifo_qdisc_ops,
541 TC_H_MAKE(sch->handle, 1));
542 if (!q->qdisc) {
543 pr_debug("netem: qdisc create failed\n");
544 return -ENOMEM;
545 }
546
547 ret = netem_change(sch, opt);
548 if (ret) {
549 pr_debug("netem: change failed\n");
550 qdisc_destroy(q->qdisc);
551 }
552 return ret;
553 }
554
555 static void netem_destroy(struct Qdisc *sch)
556 {
557 struct netem_sched_data *q = qdisc_priv(sch);
558
559 qdisc_watchdog_cancel(&q->watchdog);
560 qdisc_destroy(q->qdisc);
561 kfree(q->delay_dist);
562 }
563
564 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
565 {
566 const struct netem_sched_data *q = qdisc_priv(sch);
567 unsigned char *b = skb_tail_pointer(skb);
568 struct nlattr *nla = (struct nlattr *) b;
569 struct tc_netem_qopt qopt;
570 struct tc_netem_corr cor;
571 struct tc_netem_reorder reorder;
572 struct tc_netem_corrupt corrupt;
573
574 qopt.latency = q->latency;
575 qopt.jitter = q->jitter;
576 qopt.limit = q->limit;
577 qopt.loss = q->loss;
578 qopt.gap = q->gap;
579 qopt.duplicate = q->duplicate;
580 NLA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
581
582 cor.delay_corr = q->delay_cor.rho;
583 cor.loss_corr = q->loss_cor.rho;
584 cor.dup_corr = q->dup_cor.rho;
585 NLA_PUT(skb, TCA_NETEM_CORR, sizeof(cor), &cor);
586
587 reorder.probability = q->reorder;
588 reorder.correlation = q->reorder_cor.rho;
589 NLA_PUT(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder);
590
591 corrupt.probability = q->corrupt;
592 corrupt.correlation = q->corrupt_cor.rho;
593 NLA_PUT(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt);
594
595 nla->nla_len = skb_tail_pointer(skb) - b;
596
597 return skb->len;
598
599 nla_put_failure:
600 nlmsg_trim(skb, b);
601 return -1;
602 }
603
604 static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
605 struct sk_buff *skb, struct tcmsg *tcm)
606 {
607 struct netem_sched_data *q = qdisc_priv(sch);
608
609 if (cl != 1) /* only one class */
610 return -ENOENT;
611
612 tcm->tcm_handle |= TC_H_MIN(1);
613 tcm->tcm_info = q->qdisc->handle;
614
615 return 0;
616 }
617
618 static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
619 struct Qdisc **old)
620 {
621 struct netem_sched_data *q = qdisc_priv(sch);
622
623 if (new == NULL)
624 new = &noop_qdisc;
625
626 sch_tree_lock(sch);
627 *old = xchg(&q->qdisc, new);
628 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
629 qdisc_reset(*old);
630 sch_tree_unlock(sch);
631
632 return 0;
633 }
634
635 static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
636 {
637 struct netem_sched_data *q = qdisc_priv(sch);
638 return q->qdisc;
639 }
640
641 static unsigned long netem_get(struct Qdisc *sch, u32 classid)
642 {
643 return 1;
644 }
645
646 static void netem_put(struct Qdisc *sch, unsigned long arg)
647 {
648 }
649
650 static int netem_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
651 struct nlattr **tca, unsigned long *arg)
652 {
653 return -ENOSYS;
654 }
655
656 static int netem_delete(struct Qdisc *sch, unsigned long arg)
657 {
658 return -ENOSYS;
659 }
660
661 static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
662 {
663 if (!walker->stop) {
664 if (walker->count >= walker->skip)
665 if (walker->fn(sch, 1, walker) < 0) {
666 walker->stop = 1;
667 return;
668 }
669 walker->count++;
670 }
671 }
672
673 static struct tcf_proto **netem_find_tcf(struct Qdisc *sch, unsigned long cl)
674 {
675 return NULL;
676 }
677
678 static const struct Qdisc_class_ops netem_class_ops = {
679 .graft = netem_graft,
680 .leaf = netem_leaf,
681 .get = netem_get,
682 .put = netem_put,
683 .change = netem_change_class,
684 .delete = netem_delete,
685 .walk = netem_walk,
686 .tcf_chain = netem_find_tcf,
687 .dump = netem_dump_class,
688 };
689
690 static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
691 .id = "netem",
692 .cl_ops = &netem_class_ops,
693 .priv_size = sizeof(struct netem_sched_data),
694 .enqueue = netem_enqueue,
695 .dequeue = netem_dequeue,
696 .requeue = netem_requeue,
697 .drop = netem_drop,
698 .init = netem_init,
699 .reset = netem_reset,
700 .destroy = netem_destroy,
701 .change = netem_change,
702 .dump = netem_dump,
703 .owner = THIS_MODULE,
704 };
705
706
707 static int __init netem_module_init(void)
708 {
709 pr_info("netem: version " VERSION "\n");
710 return register_qdisc(&netem_qdisc_ops);
711 }
712 static void __exit netem_module_exit(void)
713 {
714 unregister_qdisc(&netem_qdisc_ops);
715 }
716 module_init(netem_module_init)
717 module_exit(netem_module_exit)
718 MODULE_LICENSE("GPL");
Linux kernel - Deliverables
This page took 0.062761 seconds and 6 git commands to generate.