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
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>
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 {
56 struct qdisc_watchdog watchdog;
58 psched_tdiff_t latency;
59 psched_tdiff_t jitter;
72 } delay_cor, loss_cor, dup_cor, reorder_cor, corrupt_cor;
80 /* Time stamp put into socket buffer control block */
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)
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)
101 unsigned long answer;
103 if (state->rho == 0) /* no correlation */
106 value = net_random();
107 rho = (u64)state->rho + 1;
108 answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32;
109 state->last = 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)
128 rnd = get_crandom(state);
130 /* default uniform distribution */
132 return (rnd % (2*sigma)) - sigma + mu;
134 t = dist->table[rnd % dist->size];
135 x = (sigma % NETEM_DIST_SCALE) * t;
137 x += NETEM_DIST_SCALE/2;
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;
159 pr_debug("netem_enqueue skb=%p\n", skb);
161 /* Random duplication */
162 if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor))
165 /* Random packet drop 0 => none, ~0 => all */
166 if (q->loss && q->loss >= get_crandom(&q->loss_cor))
172 return NET_XMIT_BYPASS;
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 = qdisc_root(sch);
184 u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
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))) {
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)) {
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;
221 ret = q->qdisc->enqueue(skb, q->qdisc);
224 * Do re-ordering by putting one out of N packets at the front
227 cb->time_to_send = psched_get_time();
229 ret = q->qdisc->ops->requeue(skb, q->qdisc);
232 if (likely(ret == NET_XMIT_SUCCESS)) {
234 sch->bstats.bytes += skb->len;
235 sch->bstats.packets++;
239 pr_debug("netem: enqueue ret %d\n", 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);
249 if ((ret = q->qdisc->ops->requeue(skb, q->qdisc)) == 0) {
251 sch->qstats.requeues++;
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) {
269 static struct sk_buff *netem_dequeue(struct Qdisc *sch)
271 struct netem_sched_data *q = qdisc_priv(sch);
275 if (sch->flags & TCQ_F_THROTTLED)
278 skb = q->qdisc->dequeue(q->qdisc);
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);
291 if (unlikely(q->qdisc->ops->requeue(skb, q->qdisc) != NET_XMIT_SUCCESS)) {
292 qdisc_tree_decrease_qlen(q->qdisc, 1);
294 printk(KERN_ERR "netem: %s could not requeue\n",
298 qdisc_watchdog_schedule(&q->watchdog, cb->time_to_send);
304 static void netem_reset(struct Qdisc *sch)
306 struct netem_sched_data *q = qdisc_priv(sch);
308 qdisc_reset(q->qdisc);
310 qdisc_watchdog_cancel(&q->watchdog);
314 * Distribution data is a variable size payload containing
315 * signed 16 bit values.
317 static int get_dist_table(struct Qdisc *sch, const struct nlattr *attr)
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;
329 d = kmalloc(sizeof(*d) + n*sizeof(d->table[0]), GFP_KERNEL);
334 for (i = 0; i < n; i++)
335 d->table[i] = data[i];
337 root_lock = qdisc_root_lock(sch);
339 spin_lock_bh(root_lock);
340 d = xchg(&q->delay_dist, d);
341 spin_unlock_bh(root_lock);
347 static int get_correlation(struct Qdisc *sch, const struct nlattr *attr)
349 struct netem_sched_data *q = qdisc_priv(sch);
350 const struct tc_netem_corr *c = nla_data(attr);
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);
358 static int get_reorder(struct Qdisc *sch, const struct nlattr *attr)
360 struct netem_sched_data *q = qdisc_priv(sch);
361 const struct tc_netem_reorder *r = nla_data(attr);
363 q->reorder = r->probability;
364 init_crandom(&q->reorder_cor, r->correlation);
368 static int get_corrupt(struct Qdisc *sch, const struct nlattr *attr)
370 struct netem_sched_data *q = qdisc_priv(sch);
371 const struct tc_netem_corrupt *r = nla_data(attr);
373 q->corrupt = r->probability;
374 init_crandom(&q->corrupt_cor, r->correlation);
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) },
384 /* Parse netlink message to set options */
385 static int netem_change(struct Qdisc *sch, struct nlattr *opt)
387 struct netem_sched_data *q = qdisc_priv(sch);
388 struct nlattr *tb[TCA_NETEM_MAX + 1];
389 struct tc_netem_qopt *qopt;
395 ret = nla_parse_nested_compat(tb, TCA_NETEM_MAX, opt, netem_policy,
396 qopt, sizeof(*qopt));
400 ret = fifo_set_limit(q->qdisc, qopt->limit);
402 pr_debug("netem: can't set fifo limit\n");
406 q->latency = qopt->latency;
407 q->jitter = qopt->jitter;
408 q->limit = qopt->limit;
411 q->loss = qopt->loss;
412 q->duplicate = qopt->duplicate;
414 /* for compatibility with earlier versions.
415 * if gap is set, need to assume 100% probability
420 if (tb[TCA_NETEM_CORR]) {
421 ret = get_correlation(sch, tb[TCA_NETEM_CORR]);
426 if (tb[TCA_NETEM_DELAY_DIST]) {
427 ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST]);
432 if (tb[TCA_NETEM_REORDER]) {
433 ret = get_reorder(sch, tb[TCA_NETEM_REORDER]);
438 if (tb[TCA_NETEM_CORRUPT]) {
439 ret = get_corrupt(sch, tb[TCA_NETEM_CORRUPT]);
448 * Special case version of FIFO queue for use by netem.
449 * It queues in order based on timestamps in skb's
451 struct fifo_sched_data {
453 psched_time_t oldest;
456 static int tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
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;
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)) {
467 return qdisc_enqueue_tail(nskb, sch);
470 skb_queue_reverse_walk(list, skb) {
471 const struct netem_skb_cb *cb
472 = (const struct netem_skb_cb *)skb->cb;
474 if (tnext >= cb->time_to_send)
478 __skb_queue_after(list, skb, nskb);
480 sch->qstats.backlog += nskb->len;
481 sch->bstats.bytes += nskb->len;
482 sch->bstats.packets++;
484 return NET_XMIT_SUCCESS;
487 return qdisc_reshape_fail(nskb, sch);
490 static int tfifo_init(struct Qdisc *sch, struct nlattr *opt)
492 struct fifo_sched_data *q = qdisc_priv(sch);
495 struct tc_fifo_qopt *ctl = nla_data(opt);
496 if (nla_len(opt) < sizeof(*ctl))
499 q->limit = ctl->limit;
501 q->limit = max_t(u32, qdisc_dev(sch)->tx_queue_len, 1);
503 q->oldest = PSCHED_PASTPERFECT;
507 static int tfifo_dump(struct Qdisc *sch, struct sk_buff *skb)
509 struct fifo_sched_data *q = qdisc_priv(sch);
510 struct tc_fifo_qopt opt = { .limit = q->limit };
512 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
519 static struct Qdisc_ops tfifo_qdisc_ops __read_mostly = {
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,
527 .reset = qdisc_reset_queue,
528 .change = tfifo_init,
532 static int netem_init(struct Qdisc *sch, struct nlattr *opt)
534 struct netem_sched_data *q = qdisc_priv(sch);
540 qdisc_watchdog_init(&q->watchdog, sch);
542 q->qdisc = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
544 TC_H_MAKE(sch->handle, 1));
546 pr_debug("netem: qdisc create failed\n");
550 ret = netem_change(sch, opt);
552 pr_debug("netem: change failed\n");
553 qdisc_destroy(q->qdisc);
558 static void netem_destroy(struct Qdisc *sch)
560 struct netem_sched_data *q = qdisc_priv(sch);
562 qdisc_watchdog_cancel(&q->watchdog);
563 qdisc_destroy(q->qdisc);
564 kfree(q->delay_dist);
567 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
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;
577 qopt.latency = q->latency;
578 qopt.jitter = q->jitter;
579 qopt.limit = q->limit;
582 qopt.duplicate = q->duplicate;
583 NLA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
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);
590 reorder.probability = q->reorder;
591 reorder.correlation = q->reorder_cor.rho;
592 NLA_PUT(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder);
594 corrupt.probability = q->corrupt;
595 corrupt.correlation = q->corrupt_cor.rho;
596 NLA_PUT(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt);
598 nla->nla_len = skb_tail_pointer(skb) - b;
607 static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
608 struct sk_buff *skb, struct tcmsg *tcm)
610 struct netem_sched_data *q = qdisc_priv(sch);
612 if (cl != 1) /* only one class */
615 tcm->tcm_handle |= TC_H_MIN(1);
616 tcm->tcm_info = q->qdisc->handle;
621 static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
624 struct netem_sched_data *q = qdisc_priv(sch);
630 *old = xchg(&q->qdisc, new);
631 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
633 sch_tree_unlock(sch);
638 static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
640 struct netem_sched_data *q = qdisc_priv(sch);
644 static unsigned long netem_get(struct Qdisc *sch, u32 classid)
649 static void netem_put(struct Qdisc *sch, unsigned long arg)
653 static int netem_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
654 struct nlattr **tca, unsigned long *arg)
659 static int netem_delete(struct Qdisc *sch, unsigned long arg)
664 static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
667 if (walker->count >= walker->skip)
668 if (walker->fn(sch, 1, walker) < 0) {
676 static struct tcf_proto **netem_find_tcf(struct Qdisc *sch, unsigned long cl)
681 static const struct Qdisc_class_ops netem_class_ops = {
682 .graft = netem_graft,
686 .change = netem_change_class,
687 .delete = netem_delete,
689 .tcf_chain = netem_find_tcf,
690 .dump = netem_dump_class,
693 static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
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,
702 .reset = netem_reset,
703 .destroy = netem_destroy,
704 .change = netem_change,
706 .owner = THIS_MODULE,
710 static int __init netem_module_init(void)
712 pr_info("netem: version " VERSION "\n");
713 return register_qdisc(&netem_qdisc_ops);
715 static void __exit netem_module_exit(void)
717 unregister_qdisc(&netem_qdisc_ops);
719 module_init(netem_module_init)
720 module_exit(netem_module_exit)
721 MODULE_LICENSE("GPL");