?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
pkt_sched: remove unnecessary xchg() in packet schedulers
[linux-2.6/mini2440.git] / net / sched / sch_netem.c
blobf840d6b27c657ab1ca85d2190aaf216a15724b34
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 static unsigned int netem_drop(struct Qdisc* sch)
256 {
257 struct netem_sched_data *q = qdisc_priv(sch);
258 unsigned int len = 0;
259
260 if (q->qdisc->ops->drop && (len = q->qdisc->ops->drop(q->qdisc)) != 0) {
261 sch->q.qlen--;
262 sch->qstats.drops++;
263 }
264 return len;
265 }
266
267 static struct sk_buff *netem_dequeue(struct Qdisc *sch)
268 {
269 struct netem_sched_data *q = qdisc_priv(sch);
270 struct sk_buff *skb;
271
272 smp_mb();
273 if (sch->flags & TCQ_F_THROTTLED)
274 return NULL;
275
276 skb = q->qdisc->ops->peek(q->qdisc);
277 if (skb) {
278 const struct netem_skb_cb *cb = netem_skb_cb(skb);
279 psched_time_t now = psched_get_time();
280
281 /* if more time remaining? */
282 if (cb->time_to_send <= now) {
283 skb = qdisc_dequeue_peeked(q->qdisc);
284 if (unlikely(!skb))
285 return NULL;
286
287 pr_debug("netem_dequeue: return skb=%p\n", skb);
288 sch->q.qlen--;
289 return skb;
290 }
291
292 qdisc_watchdog_schedule(&q->watchdog, cb->time_to_send);
293 }
294
295 return NULL;
296 }
297
298 static void netem_reset(struct Qdisc *sch)
299 {
300 struct netem_sched_data *q = qdisc_priv(sch);
301
302 qdisc_reset(q->qdisc);
303 sch->q.qlen = 0;
304 qdisc_watchdog_cancel(&q->watchdog);
305 }
306
307 /*
308 * Distribution data is a variable size payload containing
309 * signed 16 bit values.
310 */
311 static int get_dist_table(struct Qdisc *sch, const struct nlattr *attr)
312 {
313 struct netem_sched_data *q = qdisc_priv(sch);
314 unsigned long n = nla_len(attr)/sizeof(__s16);
315 const __s16 *data = nla_data(attr);
316 spinlock_t *root_lock;
317 struct disttable *d;
318 int i;
319
320 if (n > 65536)
321 return -EINVAL;
322
323 d = kmalloc(sizeof(*d) + n*sizeof(d->table[0]), GFP_KERNEL);
324 if (!d)
325 return -ENOMEM;
326
327 d->size = n;
328 for (i = 0; i < n; i++)
329 d->table[i] = data[i];
330
331 root_lock = qdisc_root_sleeping_lock(sch);
332
333 spin_lock_bh(root_lock);
334 kfree(q->delay_dist);
335 q->delay_dist = d;
336 spin_unlock_bh(root_lock);
337 return 0;
338 }
339
340 static void get_correlation(struct Qdisc *sch, const struct nlattr *attr)
341 {
342 struct netem_sched_data *q = qdisc_priv(sch);
343 const struct tc_netem_corr *c = nla_data(attr);
344
345 init_crandom(&q->delay_cor, c->delay_corr);
346 init_crandom(&q->loss_cor, c->loss_corr);
347 init_crandom(&q->dup_cor, c->dup_corr);
348 }
349
350 static void get_reorder(struct Qdisc *sch, const struct nlattr *attr)
351 {
352 struct netem_sched_data *q = qdisc_priv(sch);
353 const struct tc_netem_reorder *r = nla_data(attr);
354
355 q->reorder = r->probability;
356 init_crandom(&q->reorder_cor, r->correlation);
357 }
358
359 static void get_corrupt(struct Qdisc *sch, const struct nlattr *attr)
360 {
361 struct netem_sched_data *q = qdisc_priv(sch);
362 const struct tc_netem_corrupt *r = nla_data(attr);
363
364 q->corrupt = r->probability;
365 init_crandom(&q->corrupt_cor, r->correlation);
366 }
367
368 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
369 [TCA_NETEM_CORR] = { .len = sizeof(struct tc_netem_corr) },
370 [TCA_NETEM_REORDER] = { .len = sizeof(struct tc_netem_reorder) },
371 [TCA_NETEM_CORRUPT] = { .len = sizeof(struct tc_netem_corrupt) },
372 };
373
374 static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
375 const struct nla_policy *policy, int len)
376 {
377 int nested_len = nla_len(nla) - NLA_ALIGN(len);
378
379 if (nested_len < 0)
380 return -EINVAL;
381 if (nested_len >= nla_attr_size(0))
382 return nla_parse(tb, maxtype, nla_data(nla) + NLA_ALIGN(len),
383 nested_len, policy);
384 memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
385 return 0;
386 }
387
388 /* Parse netlink message to set options */
389 static int netem_change(struct Qdisc *sch, struct nlattr *opt)
390 {
391 struct netem_sched_data *q = qdisc_priv(sch);
392 struct nlattr *tb[TCA_NETEM_MAX + 1];
393 struct tc_netem_qopt *qopt;
394 int ret;
395
396 if (opt == NULL)
397 return -EINVAL;
398
399 qopt = nla_data(opt);
400 ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
401 if (ret < 0)
402 return ret;
403
404 ret = fifo_set_limit(q->qdisc, qopt->limit);
405 if (ret) {
406 pr_debug("netem: can't set fifo limit\n");
407 return ret;
408 }
409
410 q->latency = qopt->latency;
411 q->jitter = qopt->jitter;
412 q->limit = qopt->limit;
413 q->gap = qopt->gap;
414 q->counter = 0;
415 q->loss = qopt->loss;
416 q->duplicate = qopt->duplicate;
417
418 /* for compatibility with earlier versions.
419 * if gap is set, need to assume 100% probability
420 */
421 if (q->gap)
422 q->reorder = ~0;
423
424 if (tb[TCA_NETEM_CORR])
425 get_correlation(sch, tb[TCA_NETEM_CORR]);
426
427 if (tb[TCA_NETEM_DELAY_DIST]) {
428 ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST]);
429 if (ret)
430 return ret;
431 }
432
433 if (tb[TCA_NETEM_REORDER])
434 get_reorder(sch, tb[TCA_NETEM_REORDER]);
435
436 if (tb[TCA_NETEM_CORRUPT])
437 get_corrupt(sch, tb[TCA_NETEM_CORRUPT]);
438
439 return 0;
440 }
441
442 /*
443 * Special case version of FIFO queue for use by netem.
444 * It queues in order based on timestamps in skb's
445 */
446 struct fifo_sched_data {
447 u32 limit;
448 psched_time_t oldest;
449 };
450
451 static int tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
452 {
453 struct fifo_sched_data *q = qdisc_priv(sch);
454 struct sk_buff_head *list = &sch->q;
455 psched_time_t tnext = netem_skb_cb(nskb)->time_to_send;
456 struct sk_buff *skb;
457
458 if (likely(skb_queue_len(list) < q->limit)) {
459 /* Optimize for add at tail */
460 if (likely(skb_queue_empty(list) || tnext >= q->oldest)) {
461 q->oldest = tnext;
462 return qdisc_enqueue_tail(nskb, sch);
463 }
464
465 skb_queue_reverse_walk(list, skb) {
466 const struct netem_skb_cb *cb = netem_skb_cb(skb);
467
468 if (tnext >= cb->time_to_send)
469 break;
470 }
471
472 __skb_queue_after(list, skb, nskb);
473
474 sch->qstats.backlog += qdisc_pkt_len(nskb);
475 sch->bstats.bytes += qdisc_pkt_len(nskb);
476 sch->bstats.packets++;
477
478 return NET_XMIT_SUCCESS;
479 }
480
481 return qdisc_reshape_fail(nskb, sch);
482 }
483
484 static int tfifo_init(struct Qdisc *sch, struct nlattr *opt)
485 {
486 struct fifo_sched_data *q = qdisc_priv(sch);
487
488 if (opt) {
489 struct tc_fifo_qopt *ctl = nla_data(opt);
490 if (nla_len(opt) < sizeof(*ctl))
491 return -EINVAL;
492
493 q->limit = ctl->limit;
494 } else
495 q->limit = max_t(u32, qdisc_dev(sch)->tx_queue_len, 1);
496
497 q->oldest = PSCHED_PASTPERFECT;
498 return 0;
499 }
500
501 static int tfifo_dump(struct Qdisc *sch, struct sk_buff *skb)
502 {
503 struct fifo_sched_data *q = qdisc_priv(sch);
504 struct tc_fifo_qopt opt = { .limit = q->limit };
505
506 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
507 return skb->len;
508
509 nla_put_failure:
510 return -1;
511 }
512
513 static struct Qdisc_ops tfifo_qdisc_ops __read_mostly = {
514 .id = "tfifo",
515 .priv_size = sizeof(struct fifo_sched_data),
516 .enqueue = tfifo_enqueue,
517 .dequeue = qdisc_dequeue_head,
518 .peek = qdisc_peek_head,
519 .drop = qdisc_queue_drop,
520 .init = tfifo_init,
521 .reset = qdisc_reset_queue,
522 .change = tfifo_init,
523 .dump = tfifo_dump,
524 };
525
526 static int netem_init(struct Qdisc *sch, struct nlattr *opt)
527 {
528 struct netem_sched_data *q = qdisc_priv(sch);
529 int ret;
530
531 if (!opt)
532 return -EINVAL;
533
534 qdisc_watchdog_init(&q->watchdog, sch);
535
536 q->qdisc = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
537 &tfifo_qdisc_ops,
538 TC_H_MAKE(sch->handle, 1));
539 if (!q->qdisc) {
540 pr_debug("netem: qdisc create failed\n");
541 return -ENOMEM;
542 }
543
544 ret = netem_change(sch, opt);
545 if (ret) {
546 pr_debug("netem: change failed\n");
547 qdisc_destroy(q->qdisc);
548 }
549 return ret;
550 }
551
552 static void netem_destroy(struct Qdisc *sch)
553 {
554 struct netem_sched_data *q = qdisc_priv(sch);
555
556 qdisc_watchdog_cancel(&q->watchdog);
557 qdisc_destroy(q->qdisc);
558 kfree(q->delay_dist);
559 }
560
561 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
562 {
563 const struct netem_sched_data *q = qdisc_priv(sch);
564 unsigned char *b = skb_tail_pointer(skb);
565 struct nlattr *nla = (struct nlattr *) b;
566 struct tc_netem_qopt qopt;
567 struct tc_netem_corr cor;
568 struct tc_netem_reorder reorder;
569 struct tc_netem_corrupt corrupt;
570
571 qopt.latency = q->latency;
572 qopt.jitter = q->jitter;
573 qopt.limit = q->limit;
574 qopt.loss = q->loss;
575 qopt.gap = q->gap;
576 qopt.duplicate = q->duplicate;
577 NLA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
578
579 cor.delay_corr = q->delay_cor.rho;
580 cor.loss_corr = q->loss_cor.rho;
581 cor.dup_corr = q->dup_cor.rho;
582 NLA_PUT(skb, TCA_NETEM_CORR, sizeof(cor), &cor);
583
584 reorder.probability = q->reorder;
585 reorder.correlation = q->reorder_cor.rho;
586 NLA_PUT(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder);
587
588 corrupt.probability = q->corrupt;
589 corrupt.correlation = q->corrupt_cor.rho;
590 NLA_PUT(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt);
591
592 nla->nla_len = skb_tail_pointer(skb) - b;
593
594 return skb->len;
595
596 nla_put_failure:
597 nlmsg_trim(skb, b);
598 return -1;
599 }
600
601 static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
602 .id = "netem",
603 .priv_size = sizeof(struct netem_sched_data),
604 .enqueue = netem_enqueue,
605 .dequeue = netem_dequeue,
606 .peek = qdisc_peek_dequeued,
607 .drop = netem_drop,
608 .init = netem_init,
609 .reset = netem_reset,
610 .destroy = netem_destroy,
611 .change = netem_change,
612 .dump = netem_dump,
613 .owner = THIS_MODULE,
614 };
615
616
617 static int __init netem_module_init(void)
618 {
619 pr_info("netem: version " VERSION "\n");
620 return register_qdisc(&netem_qdisc_ops);
621 }
622 static void __exit netem_module_exit(void)
623 {
624 unregister_qdisc(&netem_qdisc_ops);
625 }
626 module_init(netem_module_init)
627 module_exit(netem_module_exit)
628 MODULE_LICENSE("GPL");
Support for the Chinese Samsung S3C2440 based development boards