ia64/kvm: compilation fix. export account_system_vtime.
[pv_ops_mirror.git] / net / sched / sch_netem.c
blobc9c649b26eaa1d943b404ccbd4aea3be98fa661d
1 /*
2 * net/sched/sch_netem.c Network emulator
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.
9 * Many of the algorithms and ideas for this came from
10 * NIST Net which is not copyrighted.
12 * Authors: Stephen Hemminger <shemminger@osdl.org>
13 * Catalin(ux aka Dino) BOIE <catab at umbrella dot ro>
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>
23 #include <net/netlink.h>
24 #include <net/pkt_sched.h>
26 #define VERSION "1.2"
28 /* Network Emulation Queuing algorithm.
29 ====================================
31 Sources: [1] Mark Carson, Darrin Santay, "NIST Net - A Linux-based
32 Network Emulation Tool
33 [2] Luigi Rizzo, DummyNet for FreeBSD
35 ----------------------------------------------------------------
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.
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.
50 The simulator is limited by the Linux timer resolution
51 and will create packet bursts on the HZ boundary (1ms).
54 struct netem_sched_data {
55 struct Qdisc *qdisc;
56 struct qdisc_watchdog watchdog;
58 psched_tdiff_t latency;
59 psched_tdiff_t jitter;
61 u32 loss;
62 u32 limit;
63 u32 counter;
64 u32 gap;
65 u32 duplicate;
66 u32 reorder;
67 u32 corrupt;
69 struct crndstate {
70 u32 last;
71 u32 rho;
72 } delay_cor, loss_cor, dup_cor, reorder_cor, corrupt_cor;
74 struct disttable {
75 u32 size;
76 s16 table[0];
77 } *delay_dist;
80 /* Time stamp put into socket buffer control block */
81 struct netem_skb_cb {
82 psched_time_t time_to_send;
85 /* init_crandom - initialize correlated random number generator
86 * Use entropy source for initial seed.
88 static void init_crandom(struct crndstate *state, unsigned long rho)
90 state->rho = rho;
91 state->last = net_random();
94 /* get_crandom - correlated random number generator
95 * Next number depends on last value.
96 * rho is scaled to avoid floating point.
98 static u32 get_crandom(struct crndstate *state)
100 u64 value, rho;
101 unsigned long answer;
103 if (state->rho == 0) /* no correlation */
104 return net_random();
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;
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.
117 static psched_tdiff_t tabledist(psched_tdiff_t mu, psched_tdiff_t sigma,
118 struct crndstate *state,
119 const struct disttable *dist)
121 psched_tdiff_t x;
122 long t;
123 u32 rnd;
125 if (sigma == 0)
126 return mu;
128 rnd = get_crandom(state);
130 /* default uniform distribution */
131 if (dist == NULL)
132 return (rnd % (2*sigma)) - sigma + mu;
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;
141 return x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu;
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.
150 static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch)
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;
159 pr_debug("netem_enqueue skb=%p\n", skb);
161 /* Random duplication */
162 if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor))
163 ++count;
165 /* Random packet drop 0 => none, ~0 => all */
166 if (q->loss && q->loss >= get_crandom(&q->loss_cor))
167 --count;
169 if (count == 0) {
170 sch->qstats.drops++;
171 kfree_skb(skb);
172 return NET_XMIT_BYPASS;
175 skb_orphan(skb);
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.
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;
187 rootq->enqueue(skb2, rootq);
188 q->duplicate = dupsave;
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.
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;
205 skb->data[net_random() % skb_headlen(skb)] ^= 1<<(net_random() % 8);
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;
215 delay = tabledist(q->latency, q->jitter,
216 &q->delay_cor, q->delay_dist);
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 {
224 * Do re-ordering by putting one out of N packets at the front
225 * of the queue.
227 cb->time_to_send = psched_get_time();
228 q->counter = 0;
229 ret = q->qdisc->ops->requeue(skb, q->qdisc);
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++;
239 pr_debug("netem: enqueue ret %d\n", ret);
240 return ret;
243 /* Requeue packets but don't change time stamp */
244 static int netem_requeue(struct sk_buff *skb, struct Qdisc *sch)
246 struct netem_sched_data *q = qdisc_priv(sch);
247 int ret;
249 if ((ret = q->qdisc->ops->requeue(skb, q->qdisc)) == 0) {
250 sch->q.qlen++;
251 sch->qstats.requeues++;
254 return ret;
257 static unsigned int netem_drop(struct Qdisc* sch)
259 struct netem_sched_data *q = qdisc_priv(sch);
260 unsigned int len = 0;
262 if (q->qdisc->ops->drop && (len = q->qdisc->ops->drop(q->qdisc)) != 0) {
263 sch->q.qlen--;
264 sch->qstats.drops++;
266 return len;
269 static struct sk_buff *netem_dequeue(struct Qdisc *sch)
271 struct netem_sched_data *q = qdisc_priv(sch);
272 struct sk_buff *skb;
274 smp_mb();
275 if (sch->flags & TCQ_F_THROTTLED)
276 return NULL;
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();
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;
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);
298 qdisc_watchdog_schedule(&q->watchdog, cb->time_to_send);
301 return NULL;
304 static void netem_reset(struct Qdisc *sch)
306 struct netem_sched_data *q = qdisc_priv(sch);
308 qdisc_reset(q->qdisc);
309 sch->q.qlen = 0;
310 qdisc_watchdog_cancel(&q->watchdog);
313 /* Pass size change message down to embedded FIFO */
314 static int set_fifo_limit(struct Qdisc *q, int limit)
316 struct nlattr *nla;
317 int ret = -ENOMEM;
319 /* Hack to avoid sending change message to non-FIFO */
320 if (strncmp(q->ops->id + 1, "fifo", 4) != 0)
321 return 0;
323 nla = kmalloc(nla_attr_size(sizeof(struct tc_fifo_qopt)), GFP_KERNEL);
324 if (nla) {
325 nla->nla_type = RTM_NEWQDISC;
326 nla->nla_len = nla_attr_size(sizeof(struct tc_fifo_qopt));
327 ((struct tc_fifo_qopt *)nla_data(nla))->limit = limit;
329 ret = q->ops->change(q, nla);
330 kfree(nla);
332 return ret;
336 * Distribution data is a variable size payload containing
337 * signed 16 bit values.
339 static int get_dist_table(struct Qdisc *sch, const struct nlattr *attr)
341 struct netem_sched_data *q = qdisc_priv(sch);
342 unsigned long n = nla_len(attr)/sizeof(__s16);
343 const __s16 *data = nla_data(attr);
344 struct disttable *d;
345 int i;
347 if (n > 65536)
348 return -EINVAL;
350 d = kmalloc(sizeof(*d) + n*sizeof(d->table[0]), GFP_KERNEL);
351 if (!d)
352 return -ENOMEM;
354 d->size = n;
355 for (i = 0; i < n; i++)
356 d->table[i] = data[i];
358 spin_lock_bh(&sch->dev->queue_lock);
359 d = xchg(&q->delay_dist, d);
360 spin_unlock_bh(&sch->dev->queue_lock);
362 kfree(d);
363 return 0;
366 static int get_correlation(struct Qdisc *sch, const struct nlattr *attr)
368 struct netem_sched_data *q = qdisc_priv(sch);
369 const struct tc_netem_corr *c = nla_data(attr);
371 init_crandom(&q->delay_cor, c->delay_corr);
372 init_crandom(&q->loss_cor, c->loss_corr);
373 init_crandom(&q->dup_cor, c->dup_corr);
374 return 0;
377 static int get_reorder(struct Qdisc *sch, const struct nlattr *attr)
379 struct netem_sched_data *q = qdisc_priv(sch);
380 const struct tc_netem_reorder *r = nla_data(attr);
382 q->reorder = r->probability;
383 init_crandom(&q->reorder_cor, r->correlation);
384 return 0;
387 static int get_corrupt(struct Qdisc *sch, const struct nlattr *attr)
389 struct netem_sched_data *q = qdisc_priv(sch);
390 const struct tc_netem_corrupt *r = nla_data(attr);
392 q->corrupt = r->probability;
393 init_crandom(&q->corrupt_cor, r->correlation);
394 return 0;
397 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
398 [TCA_NETEM_CORR] = { .len = sizeof(struct tc_netem_corr) },
399 [TCA_NETEM_REORDER] = { .len = sizeof(struct tc_netem_reorder) },
400 [TCA_NETEM_CORRUPT] = { .len = sizeof(struct tc_netem_corrupt) },
403 /* Parse netlink message to set options */
404 static int netem_change(struct Qdisc *sch, struct nlattr *opt)
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;
411 if (opt == NULL)
412 return -EINVAL;
414 ret = nla_parse_nested_compat(tb, TCA_NETEM_MAX, opt, netem_policy,
415 qopt, sizeof(*qopt));
416 if (ret < 0)
417 return ret;
419 ret = set_fifo_limit(q->qdisc, qopt->limit);
420 if (ret) {
421 pr_debug("netem: can't set fifo limit\n");
422 return ret;
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;
433 /* for compatibility with earlier versions.
434 * if gap is set, need to assume 100% probability
436 if (q->gap)
437 q->reorder = ~0;
439 if (tb[TCA_NETEM_CORR]) {
440 ret = get_correlation(sch, tb[TCA_NETEM_CORR]);
441 if (ret)
442 return ret;
445 if (tb[TCA_NETEM_DELAY_DIST]) {
446 ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST]);
447 if (ret)
448 return ret;
451 if (tb[TCA_NETEM_REORDER]) {
452 ret = get_reorder(sch, tb[TCA_NETEM_REORDER]);
453 if (ret)
454 return ret;
457 if (tb[TCA_NETEM_CORRUPT]) {
458 ret = get_corrupt(sch, tb[TCA_NETEM_CORRUPT]);
459 if (ret)
460 return ret;
463 return 0;
467 * Special case version of FIFO queue for use by netem.
468 * It queues in order based on timestamps in skb's
470 struct fifo_sched_data {
471 u32 limit;
472 psched_time_t oldest;
475 static int tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
477 struct fifo_sched_data *q = qdisc_priv(sch);
478 struct sk_buff_head *list = &sch->q;
479 psched_time_t tnext = ((struct netem_skb_cb *)nskb->cb)->time_to_send;
480 struct sk_buff *skb;
482 if (likely(skb_queue_len(list) < q->limit)) {
483 /* Optimize for add at tail */
484 if (likely(skb_queue_empty(list) || tnext >= q->oldest)) {
485 q->oldest = tnext;
486 return qdisc_enqueue_tail(nskb, sch);
489 skb_queue_reverse_walk(list, skb) {
490 const struct netem_skb_cb *cb
491 = (const struct netem_skb_cb *)skb->cb;
493 if (tnext >= cb->time_to_send)
494 break;
497 __skb_queue_after(list, skb, nskb);
499 sch->qstats.backlog += nskb->len;
500 sch->bstats.bytes += nskb->len;
501 sch->bstats.packets++;
503 return NET_XMIT_SUCCESS;
506 return qdisc_reshape_fail(nskb, sch);
509 static int tfifo_init(struct Qdisc *sch, struct nlattr *opt)
511 struct fifo_sched_data *q = qdisc_priv(sch);
513 if (opt) {
514 struct tc_fifo_qopt *ctl = nla_data(opt);
515 if (nla_len(opt) < sizeof(*ctl))
516 return -EINVAL;
518 q->limit = ctl->limit;
519 } else
520 q->limit = max_t(u32, sch->dev->tx_queue_len, 1);
522 q->oldest = PSCHED_PASTPERFECT;
523 return 0;
526 static int tfifo_dump(struct Qdisc *sch, struct sk_buff *skb)
528 struct fifo_sched_data *q = qdisc_priv(sch);
529 struct tc_fifo_qopt opt = { .limit = q->limit };
531 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
532 return skb->len;
534 nla_put_failure:
535 return -1;
538 static struct Qdisc_ops tfifo_qdisc_ops __read_mostly = {
539 .id = "tfifo",
540 .priv_size = sizeof(struct fifo_sched_data),
541 .enqueue = tfifo_enqueue,
542 .dequeue = qdisc_dequeue_head,
543 .requeue = qdisc_requeue,
544 .drop = qdisc_queue_drop,
545 .init = tfifo_init,
546 .reset = qdisc_reset_queue,
547 .change = tfifo_init,
548 .dump = tfifo_dump,
551 static int netem_init(struct Qdisc *sch, struct nlattr *opt)
553 struct netem_sched_data *q = qdisc_priv(sch);
554 int ret;
556 if (!opt)
557 return -EINVAL;
559 qdisc_watchdog_init(&q->watchdog, sch);
561 q->qdisc = qdisc_create_dflt(sch->dev, &tfifo_qdisc_ops,
562 TC_H_MAKE(sch->handle, 1));
563 if (!q->qdisc) {
564 pr_debug("netem: qdisc create failed\n");
565 return -ENOMEM;
568 ret = netem_change(sch, opt);
569 if (ret) {
570 pr_debug("netem: change failed\n");
571 qdisc_destroy(q->qdisc);
573 return ret;
576 static void netem_destroy(struct Qdisc *sch)
578 struct netem_sched_data *q = qdisc_priv(sch);
580 qdisc_watchdog_cancel(&q->watchdog);
581 qdisc_destroy(q->qdisc);
582 kfree(q->delay_dist);
585 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
587 const struct netem_sched_data *q = qdisc_priv(sch);
588 unsigned char *b = skb_tail_pointer(skb);
589 struct nlattr *nla = (struct nlattr *) b;
590 struct tc_netem_qopt qopt;
591 struct tc_netem_corr cor;
592 struct tc_netem_reorder reorder;
593 struct tc_netem_corrupt corrupt;
595 qopt.latency = q->latency;
596 qopt.jitter = q->jitter;
597 qopt.limit = q->limit;
598 qopt.loss = q->loss;
599 qopt.gap = q->gap;
600 qopt.duplicate = q->duplicate;
601 NLA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
603 cor.delay_corr = q->delay_cor.rho;
604 cor.loss_corr = q->loss_cor.rho;
605 cor.dup_corr = q->dup_cor.rho;
606 NLA_PUT(skb, TCA_NETEM_CORR, sizeof(cor), &cor);
608 reorder.probability = q->reorder;
609 reorder.correlation = q->reorder_cor.rho;
610 NLA_PUT(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder);
612 corrupt.probability = q->corrupt;
613 corrupt.correlation = q->corrupt_cor.rho;
614 NLA_PUT(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt);
616 nla->nla_len = skb_tail_pointer(skb) - b;
618 return skb->len;
620 nla_put_failure:
621 nlmsg_trim(skb, b);
622 return -1;
625 static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
626 struct sk_buff *skb, struct tcmsg *tcm)
628 struct netem_sched_data *q = qdisc_priv(sch);
630 if (cl != 1) /* only one class */
631 return -ENOENT;
633 tcm->tcm_handle |= TC_H_MIN(1);
634 tcm->tcm_info = q->qdisc->handle;
636 return 0;
639 static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
640 struct Qdisc **old)
642 struct netem_sched_data *q = qdisc_priv(sch);
644 if (new == NULL)
645 new = &noop_qdisc;
647 sch_tree_lock(sch);
648 *old = xchg(&q->qdisc, new);
649 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
650 qdisc_reset(*old);
651 sch_tree_unlock(sch);
653 return 0;
656 static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
658 struct netem_sched_data *q = qdisc_priv(sch);
659 return q->qdisc;
662 static unsigned long netem_get(struct Qdisc *sch, u32 classid)
664 return 1;
667 static void netem_put(struct Qdisc *sch, unsigned long arg)
671 static int netem_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
672 struct nlattr **tca, unsigned long *arg)
674 return -ENOSYS;
677 static int netem_delete(struct Qdisc *sch, unsigned long arg)
679 return -ENOSYS;
682 static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
684 if (!walker->stop) {
685 if (walker->count >= walker->skip)
686 if (walker->fn(sch, 1, walker) < 0) {
687 walker->stop = 1;
688 return;
690 walker->count++;
694 static struct tcf_proto **netem_find_tcf(struct Qdisc *sch, unsigned long cl)
696 return NULL;
699 static const struct Qdisc_class_ops netem_class_ops = {
700 .graft = netem_graft,
701 .leaf = netem_leaf,
702 .get = netem_get,
703 .put = netem_put,
704 .change = netem_change_class,
705 .delete = netem_delete,
706 .walk = netem_walk,
707 .tcf_chain = netem_find_tcf,
708 .dump = netem_dump_class,
711 static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
712 .id = "netem",
713 .cl_ops = &netem_class_ops,
714 .priv_size = sizeof(struct netem_sched_data),
715 .enqueue = netem_enqueue,
716 .dequeue = netem_dequeue,
717 .requeue = netem_requeue,
718 .drop = netem_drop,
719 .init = netem_init,
720 .reset = netem_reset,
721 .destroy = netem_destroy,
722 .change = netem_change,
723 .dump = netem_dump,
724 .owner = THIS_MODULE,
728 static int __init netem_module_init(void)
730 pr_info("netem: version " VERSION "\n");
731 return register_qdisc(&netem_qdisc_ops);
733 static void __exit netem_module_exit(void)
735 unregister_qdisc(&netem_qdisc_ops);
737 module_init(netem_module_init)
738 module_exit(netem_module_exit)
739 MODULE_LICENSE("GPL");