?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
pkt_sched: Add multiqueue handling to qdisc_graft().
[linux-2.6/linux-loongson.git] / net / sched / sch_netem.c
blobc5ea40c9eb21645ec5cef67c64f24b78e81de52d
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 = qdisc_root(sch);
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 spinlock_t *root_lock;
323 struct disttable *d;
324 int i;
325
326 if (n > 65536)
327 return -EINVAL;
328
329 d = kmalloc(sizeof(*d) + n*sizeof(d->table[0]), GFP_KERNEL);
330 if (!d)
331 return -ENOMEM;
332
333 d->size = n;
334 for (i = 0; i < n; i++)
335 d->table[i] = data[i];
336
337 root_lock = qdisc_root_lock(sch);
338
339 spin_lock_bh(root_lock);
340 d = xchg(&q->delay_dist, d);
341 spin_unlock_bh(root_lock);
342
343 kfree(d);
344 return 0;
345 }
346
347 static int get_correlation(struct Qdisc *sch, const struct nlattr *attr)
348 {
349 struct netem_sched_data *q = qdisc_priv(sch);
350 const struct tc_netem_corr *c = nla_data(attr);
351
352 init_crandom(&q->delay_cor, c->delay_corr);
353 init_crandom(&q->loss_cor, c->loss_corr);
354 init_crandom(&q->dup_cor, c->dup_corr);
355 return 0;
356 }
357
358 static int get_reorder(struct Qdisc *sch, const struct nlattr *attr)
359 {
360 struct netem_sched_data *q = qdisc_priv(sch);
361 const struct tc_netem_reorder *r = nla_data(attr);
362
363 q->reorder = r->probability;
364 init_crandom(&q->reorder_cor, r->correlation);
365 return 0;
366 }
367
368 static int get_corrupt(struct Qdisc *sch, const struct nlattr *attr)
369 {
370 struct netem_sched_data *q = qdisc_priv(sch);
371 const struct tc_netem_corrupt *r = nla_data(attr);
372
373 q->corrupt = r->probability;
374 init_crandom(&q->corrupt_cor, r->correlation);
375 return 0;
376 }
377
378 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
379 [TCA_NETEM_CORR] = { .len = sizeof(struct tc_netem_corr) },
380 [TCA_NETEM_REORDER] = { .len = sizeof(struct tc_netem_reorder) },
381 [TCA_NETEM_CORRUPT] = { .len = sizeof(struct tc_netem_corrupt) },
382 };
383
384 /* Parse netlink message to set options */
385 static int netem_change(struct Qdisc *sch, struct nlattr *opt)
386 {
387 struct netem_sched_data *q = qdisc_priv(sch);
388 struct nlattr *tb[TCA_NETEM_MAX + 1];
389 struct tc_netem_qopt *qopt;
390 int ret;
391
392 if (opt == NULL)
393 return -EINVAL;
394
395 ret = nla_parse_nested_compat(tb, TCA_NETEM_MAX, opt, netem_policy,
396 qopt, sizeof(*qopt));
397 if (ret < 0)
398 return ret;
399
400 ret = fifo_set_limit(q->qdisc, qopt->limit);
401 if (ret) {
402 pr_debug("netem: can't set fifo limit\n");
403 return ret;
404 }
405
406 q->latency = qopt->latency;
407 q->jitter = qopt->jitter;
408 q->limit = qopt->limit;
409 q->gap = qopt->gap;
410 q->counter = 0;
411 q->loss = qopt->loss;
412 q->duplicate = qopt->duplicate;
413
414 /* for compatibility with earlier versions.
415 * if gap is set, need to assume 100% probability
416 */
417 if (q->gap)
418 q->reorder = ~0;
419
420 if (tb[TCA_NETEM_CORR]) {
421 ret = get_correlation(sch, tb[TCA_NETEM_CORR]);
422 if (ret)
423 return ret;
424 }
425
426 if (tb[TCA_NETEM_DELAY_DIST]) {
427 ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST]);
428 if (ret)
429 return ret;
430 }
431
432 if (tb[TCA_NETEM_REORDER]) {
433 ret = get_reorder(sch, tb[TCA_NETEM_REORDER]);
434 if (ret)
435 return ret;
436 }
437
438 if (tb[TCA_NETEM_CORRUPT]) {
439 ret = get_corrupt(sch, tb[TCA_NETEM_CORRUPT]);
440 if (ret)
441 return ret;
442 }
443
444 return 0;
445 }
446
447 /*
448 * Special case version of FIFO queue for use by netem.
449 * It queues in order based on timestamps in skb's
450 */
451 struct fifo_sched_data {
452 u32 limit;
453 psched_time_t oldest;
454 };
455
456 static int tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
457 {
458 struct fifo_sched_data *q = qdisc_priv(sch);
459 struct sk_buff_head *list = &sch->q;
460 psched_time_t tnext = ((struct netem_skb_cb *)nskb->cb)->time_to_send;
461 struct sk_buff *skb;
462
463 if (likely(skb_queue_len(list) < q->limit)) {
464 /* Optimize for add at tail */
465 if (likely(skb_queue_empty(list) || tnext >= q->oldest)) {
466 q->oldest = tnext;
467 return qdisc_enqueue_tail(nskb, sch);
468 }
469
470 skb_queue_reverse_walk(list, skb) {
471 const struct netem_skb_cb *cb
472 = (const struct netem_skb_cb *)skb->cb;
473
474 if (tnext >= cb->time_to_send)
475 break;
476 }
477
478 __skb_queue_after(list, skb, nskb);
479
480 sch->qstats.backlog += nskb->len;
481 sch->bstats.bytes += nskb->len;
482 sch->bstats.packets++;
483
484 return NET_XMIT_SUCCESS;
485 }
486
487 return qdisc_reshape_fail(nskb, sch);
488 }
489
490 static int tfifo_init(struct Qdisc *sch, struct nlattr *opt)
491 {
492 struct fifo_sched_data *q = qdisc_priv(sch);
493
494 if (opt) {
495 struct tc_fifo_qopt *ctl = nla_data(opt);
496 if (nla_len(opt) < sizeof(*ctl))
497 return -EINVAL;
498
499 q->limit = ctl->limit;
500 } else
501 q->limit = max_t(u32, qdisc_dev(sch)->tx_queue_len, 1);
502
503 q->oldest = PSCHED_PASTPERFECT;
504 return 0;
505 }
506
507 static int tfifo_dump(struct Qdisc *sch, struct sk_buff *skb)
508 {
509 struct fifo_sched_data *q = qdisc_priv(sch);
510 struct tc_fifo_qopt opt = { .limit = q->limit };
511
512 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
513 return skb->len;
514
515 nla_put_failure:
516 return -1;
517 }
518
519 static struct Qdisc_ops tfifo_qdisc_ops __read_mostly = {
520 .id = "tfifo",
521 .priv_size = sizeof(struct fifo_sched_data),
522 .enqueue = tfifo_enqueue,
523 .dequeue = qdisc_dequeue_head,
524 .requeue = qdisc_requeue,
525 .drop = qdisc_queue_drop,
526 .init = tfifo_init,
527 .reset = qdisc_reset_queue,
528 .change = tfifo_init,
529 .dump = tfifo_dump,
530 };
531
532 static int netem_init(struct Qdisc *sch, struct nlattr *opt)
533 {
534 struct netem_sched_data *q = qdisc_priv(sch);
535 int ret;
536
537 if (!opt)
538 return -EINVAL;
539
540 qdisc_watchdog_init(&q->watchdog, sch);
541
542 q->qdisc = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
543 &tfifo_qdisc_ops,
544 TC_H_MAKE(sch->handle, 1));
545 if (!q->qdisc) {
546 pr_debug("netem: qdisc create failed\n");
547 return -ENOMEM;
548 }
549
550 ret = netem_change(sch, opt);
551 if (ret) {
552 pr_debug("netem: change failed\n");
553 qdisc_destroy(q->qdisc);
554 }
555 return ret;
556 }
557
558 static void netem_destroy(struct Qdisc *sch)
559 {
560 struct netem_sched_data *q = qdisc_priv(sch);
561
562 qdisc_watchdog_cancel(&q->watchdog);
563 qdisc_destroy(q->qdisc);
564 kfree(q->delay_dist);
565 }
566
567 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
568 {
569 const struct netem_sched_data *q = qdisc_priv(sch);
570 unsigned char *b = skb_tail_pointer(skb);
571 struct nlattr *nla = (struct nlattr *) b;
572 struct tc_netem_qopt qopt;
573 struct tc_netem_corr cor;
574 struct tc_netem_reorder reorder;
575 struct tc_netem_corrupt corrupt;
576
577 qopt.latency = q->latency;
578 qopt.jitter = q->jitter;
579 qopt.limit = q->limit;
580 qopt.loss = q->loss;
581 qopt.gap = q->gap;
582 qopt.duplicate = q->duplicate;
583 NLA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
584
585 cor.delay_corr = q->delay_cor.rho;
586 cor.loss_corr = q->loss_cor.rho;
587 cor.dup_corr = q->dup_cor.rho;
588 NLA_PUT(skb, TCA_NETEM_CORR, sizeof(cor), &cor);
589
590 reorder.probability = q->reorder;
591 reorder.correlation = q->reorder_cor.rho;
592 NLA_PUT(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder);
593
594 corrupt.probability = q->corrupt;
595 corrupt.correlation = q->corrupt_cor.rho;
596 NLA_PUT(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt);
597
598 nla->nla_len = skb_tail_pointer(skb) - b;
599
600 return skb->len;
601
602 nla_put_failure:
603 nlmsg_trim(skb, b);
604 return -1;
605 }
606
607 static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
608 struct sk_buff *skb, struct tcmsg *tcm)
609 {
610 struct netem_sched_data *q = qdisc_priv(sch);
611
612 if (cl != 1) /* only one class */
613 return -ENOENT;
614
615 tcm->tcm_handle |= TC_H_MIN(1);
616 tcm->tcm_info = q->qdisc->handle;
617
618 return 0;
619 }
620
621 static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
622 struct Qdisc **old)
623 {
624 struct netem_sched_data *q = qdisc_priv(sch);
625
626 if (new == NULL)
627 new = &noop_qdisc;
628
629 sch_tree_lock(sch);
630 *old = xchg(&q->qdisc, new);
631 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
632 qdisc_reset(*old);
633 sch_tree_unlock(sch);
634
635 return 0;
636 }
637
638 static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
639 {
640 struct netem_sched_data *q = qdisc_priv(sch);
641 return q->qdisc;
642 }
643
644 static unsigned long netem_get(struct Qdisc *sch, u32 classid)
645 {
646 return 1;
647 }
648
649 static void netem_put(struct Qdisc *sch, unsigned long arg)
650 {
651 }
652
653 static int netem_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
654 struct nlattr **tca, unsigned long *arg)
655 {
656 return -ENOSYS;
657 }
658
659 static int netem_delete(struct Qdisc *sch, unsigned long arg)
660 {
661 return -ENOSYS;
662 }
663
664 static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
665 {
666 if (!walker->stop) {
667 if (walker->count >= walker->skip)
668 if (walker->fn(sch, 1, walker) < 0) {
669 walker->stop = 1;
670 return;
671 }
672 walker->count++;
673 }
674 }
675
676 static struct tcf_proto **netem_find_tcf(struct Qdisc *sch, unsigned long cl)
677 {
678 return NULL;
679 }
680
681 static const struct Qdisc_class_ops netem_class_ops = {
682 .graft = netem_graft,
683 .leaf = netem_leaf,
684 .get = netem_get,
685 .put = netem_put,
686 .change = netem_change_class,
687 .delete = netem_delete,
688 .walk = netem_walk,
689 .tcf_chain = netem_find_tcf,
690 .dump = netem_dump_class,
691 };
692
693 static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
694 .id = "netem",
695 .cl_ops = &netem_class_ops,
696 .priv_size = sizeof(struct netem_sched_data),
697 .enqueue = netem_enqueue,
698 .dequeue = netem_dequeue,
699 .requeue = netem_requeue,
700 .drop = netem_drop,
701 .init = netem_init,
702 .reset = netem_reset,
703 .destroy = netem_destroy,
704 .change = netem_change,
705 .dump = netem_dump,
706 .owner = THIS_MODULE,
707 };
708
709
710 static int __init netem_module_init(void)
711 {
712 pr_info("netem: version " VERSION "\n");
713 return register_qdisc(&netem_qdisc_ops);
714 }
715 static void __exit netem_module_exit(void)
716 {
717 unregister_qdisc(&netem_qdisc_ops);
718 }
719 module_init(netem_module_init)
720 module_exit(netem_module_exit)
721 MODULE_LICENSE("GPL");
linux kernel for loongson