projects / deliverable / linux.git / blob
? search:


f69698ff88d900cbbea7054098f72750e265177d
[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 static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb)
86 {
87 BUILD_BUG_ON(sizeof(skb->cb) <
88 sizeof(struct qdisc_skb_cb) + sizeof(struct netem_skb_cb));
89 return (struct netem_skb_cb *)qdisc_skb_cb(skb)->data;
90 }
91
92 /* init_crandom - initialize correlated random number generator
93 * Use entropy source for initial seed.
94 */
95 static void init_crandom(struct crndstate *state, unsigned long rho)
96 {
97 state->rho = rho;
98 state->last = net_random();
99 }
100
101 /* get_crandom - correlated random number generator
102 * Next number depends on last value.
103 * rho is scaled to avoid floating point.
104 */
105 static u32 get_crandom(struct crndstate *state)
106 {
107 u64 value, rho;
108 unsigned long answer;
109
110 if (state->rho == 0) /* no correlation */
111 return net_random();
112
113 value = net_random();
114 rho = (u64)state->rho + 1;
115 answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32;
116 state->last = answer;
117 return answer;
118 }
119
120 /* tabledist - return a pseudo-randomly distributed value with mean mu and
121 * std deviation sigma. Uses table lookup to approximate the desired
122 * distribution, and a uniformly-distributed pseudo-random source.
123 */
124 static psched_tdiff_t tabledist(psched_tdiff_t mu, psched_tdiff_t sigma,
125 struct crndstate *state,
126 const struct disttable *dist)
127 {
128 psched_tdiff_t x;
129 long t;
130 u32 rnd;
131
132 if (sigma == 0)
133 return mu;
134
135 rnd = get_crandom(state);
136
137 /* default uniform distribution */
138 if (dist == NULL)
139 return (rnd % (2*sigma)) - sigma + mu;
140
141 t = dist->table[rnd % dist->size];
142 x = (sigma % NETEM_DIST_SCALE) * t;
143 if (x >= 0)
144 x += NETEM_DIST_SCALE/2;
145 else
146 x -= NETEM_DIST_SCALE/2;
147
148 return x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu;
149 }
150
151 /*
152 * Insert one skb into qdisc.
153 * Note: parent depends on return value to account for queue length.
154 * NET_XMIT_DROP: queue length didn't change.
155 * NET_XMIT_SUCCESS: one skb was queued.
156 */
157 static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch)
158 {
159 struct netem_sched_data *q = qdisc_priv(sch);
160 /* We don't fill cb now as skb_unshare() may invalidate it */
161 struct netem_skb_cb *cb;
162 struct sk_buff *skb2;
163 int ret;
164 int count = 1;
165
166 pr_debug("netem_enqueue skb=%p\n", skb);
167
168 /* Random duplication */
169 if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor))
170 ++count;
171
172 /* Random packet drop 0 => none, ~0 => all */
173 if (q->loss && q->loss >= get_crandom(&q->loss_cor))
174 --count;
175
176 if (count == 0) {
177 sch->qstats.drops++;
178 kfree_skb(skb);
179 return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
180 }
181
182 skb_orphan(skb);
183
184 /*
185 * If we need to duplicate packet, then re-insert at top of the
186 * qdisc tree, since parent queuer expects that only one
187 * skb will be queued.
188 */
189 if (count > 1 && (skb2 = skb_clone(skb, GFP_ATOMIC)) != NULL) {
190 struct Qdisc *rootq = qdisc_root(sch);
191 u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
192 q->duplicate = 0;
193
194 qdisc_enqueue_root(skb2, rootq);
195 q->duplicate = dupsave;
196 }
197
198 /*
199 * Randomized packet corruption.
200 * Make copy if needed since we are modifying
201 * If packet is going to be hardware checksummed, then
202 * do it now in software before we mangle it.
203 */
204 if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor)) {
205 if (!(skb = skb_unshare(skb, GFP_ATOMIC))
206 || (skb->ip_summed == CHECKSUM_PARTIAL
207 && skb_checksum_help(skb))) {
208 sch->qstats.drops++;
209 return NET_XMIT_DROP;
210 }
211
212 skb->data[net_random() % skb_headlen(skb)] ^= 1<<(net_random() % 8);
213 }
214
215 cb = netem_skb_cb(skb);
216 if (q->gap == 0 /* not doing reordering */
217 || q->counter < q->gap /* inside last reordering gap */
218 || q->reorder < get_crandom(&q->reorder_cor)) {
219 psched_time_t now;
220 psched_tdiff_t delay;
221
222 delay = tabledist(q->latency, q->jitter,
223 &q->delay_cor, q->delay_dist);
224
225 now = psched_get_time();
226 cb->time_to_send = now + delay;
227 ++q->counter;
228 ret = qdisc_enqueue(skb, q->qdisc);
229 } else {
230 /*
231 * Do re-ordering by putting one out of N packets at the front
232 * of the queue.
233 */
234 cb->time_to_send = psched_get_time();
235 q->counter = 0;
236
237 __skb_queue_head(&q->qdisc->q, skb);
238 q->qdisc->qstats.backlog += qdisc_pkt_len(skb);
239 q->qdisc->qstats.requeues++;
240 ret = NET_XMIT_SUCCESS;
241 }
242
243 if (likely(ret == NET_XMIT_SUCCESS)) {
244 sch->q.qlen++;
245 sch->bstats.bytes += qdisc_pkt_len(skb);
246 sch->bstats.packets++;
247 } else if (net_xmit_drop_count(ret)) {
248 sch->qstats.drops++;
249 }
250
251 pr_debug("netem: enqueue ret %d\n", ret);
252 return ret;
253 }
254
255 /* Requeue packets but don't change time stamp */
256 static int netem_requeue(struct sk_buff *skb, struct Qdisc *sch)
257 {
258 struct netem_sched_data *q = qdisc_priv(sch);
259 int ret;
260
261 if ((ret = q->qdisc->ops->requeue(skb, q->qdisc)) == 0) {
262 sch->q.qlen++;
263 sch->qstats.requeues++;
264 }
265
266 return ret;
267 }
268
269 static unsigned int netem_drop(struct Qdisc* sch)
270 {
271 struct netem_sched_data *q = qdisc_priv(sch);
272 unsigned int len = 0;
273
274 if (q->qdisc->ops->drop && (len = q->qdisc->ops->drop(q->qdisc)) != 0) {
275 sch->q.qlen--;
276 sch->qstats.drops++;
277 }
278 return len;
279 }
280
281 static struct sk_buff *netem_dequeue(struct Qdisc *sch)
282 {
283 struct netem_sched_data *q = qdisc_priv(sch);
284 struct sk_buff *skb;
285
286 smp_mb();
287 if (sch->flags & TCQ_F_THROTTLED)
288 return NULL;
289
290 skb = q->qdisc->ops->peek(q->qdisc);
291 if (skb) {
292 const struct netem_skb_cb *cb = netem_skb_cb(skb);
293 psched_time_t now = psched_get_time();
294
295 /* if more time remaining? */
296 if (cb->time_to_send <= now) {
297 skb = qdisc_dequeue_peeked(q->qdisc);
298 if (unlikely(!skb))
299 return NULL;
300
301 pr_debug("netem_dequeue: return skb=%p\n", skb);
302 sch->q.qlen--;
303 return skb;
304 }
305
306 qdisc_watchdog_schedule(&q->watchdog, cb->time_to_send);
307 }
308
309 return NULL;
310 }
311
312 static void netem_reset(struct Qdisc *sch)
313 {
314 struct netem_sched_data *q = qdisc_priv(sch);
315
316 qdisc_reset(q->qdisc);
317 sch->q.qlen = 0;
318 qdisc_watchdog_cancel(&q->watchdog);
319 }
320
321 /*
322 * Distribution data is a variable size payload containing
323 * signed 16 bit values.
324 */
325 static int get_dist_table(struct Qdisc *sch, const struct nlattr *attr)
326 {
327 struct netem_sched_data *q = qdisc_priv(sch);
328 unsigned long n = nla_len(attr)/sizeof(__s16);
329 const __s16 *data = nla_data(attr);
330 spinlock_t *root_lock;
331 struct disttable *d;
332 int i;
333
334 if (n > 65536)
335 return -EINVAL;
336
337 d = kmalloc(sizeof(*d) + n*sizeof(d->table[0]), GFP_KERNEL);
338 if (!d)
339 return -ENOMEM;
340
341 d->size = n;
342 for (i = 0; i < n; i++)
343 d->table[i] = data[i];
344
345 root_lock = qdisc_root_sleeping_lock(sch);
346
347 spin_lock_bh(root_lock);
348 d = xchg(&q->delay_dist, d);
349 spin_unlock_bh(root_lock);
350
351 kfree(d);
352 return 0;
353 }
354
355 static void get_correlation(struct Qdisc *sch, const struct nlattr *attr)
356 {
357 struct netem_sched_data *q = qdisc_priv(sch);
358 const struct tc_netem_corr *c = nla_data(attr);
359
360 init_crandom(&q->delay_cor, c->delay_corr);
361 init_crandom(&q->loss_cor, c->loss_corr);
362 init_crandom(&q->dup_cor, c->dup_corr);
363 }
364
365 static void get_reorder(struct Qdisc *sch, const struct nlattr *attr)
366 {
367 struct netem_sched_data *q = qdisc_priv(sch);
368 const struct tc_netem_reorder *r = nla_data(attr);
369
370 q->reorder = r->probability;
371 init_crandom(&q->reorder_cor, r->correlation);
372 }
373
374 static void get_corrupt(struct Qdisc *sch, const struct nlattr *attr)
375 {
376 struct netem_sched_data *q = qdisc_priv(sch);
377 const struct tc_netem_corrupt *r = nla_data(attr);
378
379 q->corrupt = r->probability;
380 init_crandom(&q->corrupt_cor, r->correlation);
381 }
382
383 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
384 [TCA_NETEM_CORR] = { .len = sizeof(struct tc_netem_corr) },
385 [TCA_NETEM_REORDER] = { .len = sizeof(struct tc_netem_reorder) },
386 [TCA_NETEM_CORRUPT] = { .len = sizeof(struct tc_netem_corrupt) },
387 };
388
389 static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
390 const struct nla_policy *policy, int len)
391 {
392 int nested_len = nla_len(nla) - NLA_ALIGN(len);
393
394 if (nested_len < 0)
395 return -EINVAL;
396 if (nested_len >= nla_attr_size(0))
397 return nla_parse(tb, maxtype, nla_data(nla) + NLA_ALIGN(len),
398 nested_len, policy);
399 memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
400 return 0;
401 }
402
403 /* Parse netlink message to set options */
404 static int netem_change(struct Qdisc *sch, struct nlattr *opt)
405 {
406 struct netem_sched_data *q = qdisc_priv(sch);
407 struct nlattr *tb[TCA_NETEM_MAX + 1];
408 struct tc_netem_qopt *qopt;
409 int ret;
410
411 if (opt == NULL)
412 return -EINVAL;
413
414 qopt = nla_data(opt);
415 ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
416 if (ret < 0)
417 return ret;
418
419 ret = fifo_set_limit(q->qdisc, qopt->limit);
420 if (ret) {
421 pr_debug("netem: can't set fifo limit\n");
422 return ret;
423 }
424
425 q->latency = qopt->latency;
426 q->jitter = qopt->jitter;
427 q->limit = qopt->limit;
428 q->gap = qopt->gap;
429 q->counter = 0;
430 q->loss = qopt->loss;
431 q->duplicate = qopt->duplicate;
432
433 /* for compatibility with earlier versions.
434 * if gap is set, need to assume 100% probability
435 */
436 if (q->gap)
437 q->reorder = ~0;
438
439 if (tb[TCA_NETEM_CORR])
440 get_correlation(sch, tb[TCA_NETEM_CORR]);
441
442 if (tb[TCA_NETEM_DELAY_DIST]) {
443 ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST]);
444 if (ret)
445 return ret;
446 }
447
448 if (tb[TCA_NETEM_REORDER])
449 get_reorder(sch, tb[TCA_NETEM_REORDER]);
450
451 if (tb[TCA_NETEM_CORRUPT])
452 get_corrupt(sch, tb[TCA_NETEM_CORRUPT]);
453
454 return 0;
455 }
456
457 /*
458 * Special case version of FIFO queue for use by netem.
459 * It queues in order based on timestamps in skb's
460 */
461 struct fifo_sched_data {
462 u32 limit;
463 psched_time_t oldest;
464 };
465
466 static int tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
467 {
468 struct fifo_sched_data *q = qdisc_priv(sch);
469 struct sk_buff_head *list = &sch->q;
470 psched_time_t tnext = netem_skb_cb(nskb)->time_to_send;
471 struct sk_buff *skb;
472
473 if (likely(skb_queue_len(list) < q->limit)) {
474 /* Optimize for add at tail */
475 if (likely(skb_queue_empty(list) || tnext >= q->oldest)) {
476 q->oldest = tnext;
477 return qdisc_enqueue_tail(nskb, sch);
478 }
479
480 skb_queue_reverse_walk(list, skb) {
481 const struct netem_skb_cb *cb = netem_skb_cb(skb);
482
483 if (tnext >= cb->time_to_send)
484 break;
485 }
486
487 __skb_queue_after(list, skb, nskb);
488
489 sch->qstats.backlog += qdisc_pkt_len(nskb);
490 sch->bstats.bytes += qdisc_pkt_len(nskb);
491 sch->bstats.packets++;
492
493 return NET_XMIT_SUCCESS;
494 }
495
496 return qdisc_reshape_fail(nskb, sch);
497 }
498
499 static int tfifo_init(struct Qdisc *sch, struct nlattr *opt)
500 {
501 struct fifo_sched_data *q = qdisc_priv(sch);
502
503 if (opt) {
504 struct tc_fifo_qopt *ctl = nla_data(opt);
505 if (nla_len(opt) < sizeof(*ctl))
506 return -EINVAL;
507
508 q->limit = ctl->limit;
509 } else
510 q->limit = max_t(u32, qdisc_dev(sch)->tx_queue_len, 1);
511
512 q->oldest = PSCHED_PASTPERFECT;
513 return 0;
514 }
515
516 static int tfifo_dump(struct Qdisc *sch, struct sk_buff *skb)
517 {
518 struct fifo_sched_data *q = qdisc_priv(sch);
519 struct tc_fifo_qopt opt = { .limit = q->limit };
520
521 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
522 return skb->len;
523
524 nla_put_failure:
525 return -1;
526 }
527
528 static struct Qdisc_ops tfifo_qdisc_ops __read_mostly = {
529 .id = "tfifo",
530 .priv_size = sizeof(struct fifo_sched_data),
531 .enqueue = tfifo_enqueue,
532 .dequeue = qdisc_dequeue_head,
533 .peek = qdisc_peek_head,
534 .requeue = qdisc_requeue,
535 .drop = qdisc_queue_drop,
536 .init = tfifo_init,
537 .reset = qdisc_reset_queue,
538 .change = tfifo_init,
539 .dump = tfifo_dump,
540 };
541
542 static int netem_init(struct Qdisc *sch, struct nlattr *opt)
543 {
544 struct netem_sched_data *q = qdisc_priv(sch);
545 int ret;
546
547 if (!opt)
548 return -EINVAL;
549
550 qdisc_watchdog_init(&q->watchdog, sch);
551
552 q->qdisc = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
553 &tfifo_qdisc_ops,
554 TC_H_MAKE(sch->handle, 1));
555 if (!q->qdisc) {
556 pr_debug("netem: qdisc create failed\n");
557 return -ENOMEM;
558 }
559
560 ret = netem_change(sch, opt);
561 if (ret) {
562 pr_debug("netem: change failed\n");
563 qdisc_destroy(q->qdisc);
564 }
565 return ret;
566 }
567
568 static void netem_destroy(struct Qdisc *sch)
569 {
570 struct netem_sched_data *q = qdisc_priv(sch);
571
572 qdisc_watchdog_cancel(&q->watchdog);
573 qdisc_destroy(q->qdisc);
574 kfree(q->delay_dist);
575 }
576
577 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
578 {
579 const struct netem_sched_data *q = qdisc_priv(sch);
580 unsigned char *b = skb_tail_pointer(skb);
581 struct nlattr *nla = (struct nlattr *) b;
582 struct tc_netem_qopt qopt;
583 struct tc_netem_corr cor;
584 struct tc_netem_reorder reorder;
585 struct tc_netem_corrupt corrupt;
586
587 qopt.latency = q->latency;
588 qopt.jitter = q->jitter;
589 qopt.limit = q->limit;
590 qopt.loss = q->loss;
591 qopt.gap = q->gap;
592 qopt.duplicate = q->duplicate;
593 NLA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
594
595 cor.delay_corr = q->delay_cor.rho;
596 cor.loss_corr = q->loss_cor.rho;
597 cor.dup_corr = q->dup_cor.rho;
598 NLA_PUT(skb, TCA_NETEM_CORR, sizeof(cor), &cor);
599
600 reorder.probability = q->reorder;
601 reorder.correlation = q->reorder_cor.rho;
602 NLA_PUT(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder);
603
604 corrupt.probability = q->corrupt;
605 corrupt.correlation = q->corrupt_cor.rho;
606 NLA_PUT(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt);
607
608 nla->nla_len = skb_tail_pointer(skb) - b;
609
610 return skb->len;
611
612 nla_put_failure:
613 nlmsg_trim(skb, b);
614 return -1;
615 }
616
617 static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
618 .id = "netem",
619 .priv_size = sizeof(struct netem_sched_data),
620 .enqueue = netem_enqueue,
621 .dequeue = netem_dequeue,
622 .peek = qdisc_peek_dequeued,
623 .requeue = netem_requeue,
624 .drop = netem_drop,
625 .init = netem_init,
626 .reset = netem_reset,
627 .destroy = netem_destroy,
628 .change = netem_change,
629 .dump = netem_dump,
630 .owner = THIS_MODULE,
631 };
632
633
634 static int __init netem_module_init(void)
635 {
636 pr_info("netem: version " VERSION "\n");
637 return register_qdisc(&netem_qdisc_ops);
638 }
639 static void __exit netem_module_exit(void)
640 {
641 unregister_qdisc(&netem_qdisc_ops);
642 }
643 module_init(netem_module_init)
644 module_exit(netem_module_exit)
645 MODULE_LICENSE("GPL");
Linux kernel - Deliverables
This page took 0.042597 seconds and 4 git commands to generate.