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.
51 struct netem_sched_data {
53 struct qdisc_watchdog watchdog;
55 psched_tdiff_t latency;
56 psched_tdiff_t jitter;
69 } delay_cor, loss_cor, dup_cor, reorder_cor, corrupt_cor;
77 /* Time stamp put into socket buffer control block */
79 psched_time_t time_to_send;
82 static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb)
84 BUILD_BUG_ON(sizeof(skb->cb) <
85 sizeof(struct qdisc_skb_cb) + sizeof(struct netem_skb_cb));
86 return (struct netem_skb_cb *)qdisc_skb_cb(skb)->data;
89 /* init_crandom - initialize correlated random number generator
90 * Use entropy source for initial seed.
92 static void init_crandom(struct crndstate *state, unsigned long rho)
95 state->last = net_random();
98 /* get_crandom - correlated random number generator
99 * Next number depends on last value.
100 * rho is scaled to avoid floating point.
102 static u32 get_crandom(struct crndstate *state)
105 unsigned long answer;
107 if (state->rho == 0) /* no correlation */
110 value = net_random();
111 rho = (u64)state->rho + 1;
112 answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32;
113 state->last = answer;
117 /* tabledist - return a pseudo-randomly distributed value with mean mu and
118 * std deviation sigma. Uses table lookup to approximate the desired
119 * distribution, and a uniformly-distributed pseudo-random source.
121 static psched_tdiff_t tabledist(psched_tdiff_t mu, psched_tdiff_t sigma,
122 struct crndstate *state,
123 const struct disttable *dist)
132 rnd = get_crandom(state);
134 /* default uniform distribution */
136 return (rnd % (2*sigma)) - sigma + mu;
138 t = dist->table[rnd % dist->size];
139 x = (sigma % NETEM_DIST_SCALE) * t;
141 x += NETEM_DIST_SCALE/2;
143 x -= NETEM_DIST_SCALE/2;
145 return x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu;
149 * Insert one skb into qdisc.
150 * Note: parent depends on return value to account for queue length.
151 * NET_XMIT_DROP: queue length didn't change.
152 * NET_XMIT_SUCCESS: one skb was queued.
154 static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch)
156 struct netem_sched_data *q = qdisc_priv(sch);
157 /* We don't fill cb now as skb_unshare() may invalidate it */
158 struct netem_skb_cb *cb;
159 struct sk_buff *skb2;
163 pr_debug("netem_enqueue skb=%p\n", skb);
165 /* Random duplication */
166 if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor))
169 /* Random packet drop 0 => none, ~0 => all */
170 if (q->loss && q->loss >= get_crandom(&q->loss_cor))
176 return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
182 * If we need to duplicate packet, then re-insert at top of the
183 * qdisc tree, since parent queuer expects that only one
184 * skb will be queued.
186 if (count > 1 && (skb2 = skb_clone(skb, GFP_ATOMIC)) != NULL) {
187 struct Qdisc *rootq = qdisc_root(sch);
188 u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
191 qdisc_enqueue_root(skb2, rootq);
192 q->duplicate = dupsave;
196 * Randomized packet corruption.
197 * Make copy if needed since we are modifying
198 * If packet is going to be hardware checksummed, then
199 * do it now in software before we mangle it.
201 if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor)) {
202 if (!(skb = skb_unshare(skb, GFP_ATOMIC))
203 || (skb->ip_summed == CHECKSUM_PARTIAL
204 && skb_checksum_help(skb))) {
206 return NET_XMIT_DROP;
209 skb->data[net_random() % skb_headlen(skb)] ^= 1<<(net_random() % 8);
212 cb = netem_skb_cb(skb);
213 if (q->gap == 0 /* not doing reordering */
214 || q->counter < q->gap /* inside last reordering gap */
215 || q->reorder < get_crandom(&q->reorder_cor)) {
217 psched_tdiff_t delay;
219 delay = tabledist(q->latency, q->jitter,
220 &q->delay_cor, q->delay_dist);
222 now = psched_get_time();
223 cb->time_to_send = now + delay;
225 ret = qdisc_enqueue(skb, q->qdisc);
228 * Do re-ordering by putting one out of N packets at the front
231 cb->time_to_send = psched_get_time();
233 ret = q->qdisc->ops->requeue(skb, q->qdisc);
236 if (likely(ret == NET_XMIT_SUCCESS)) {
238 sch->bstats.bytes += qdisc_pkt_len(skb);
239 sch->bstats.packets++;
240 } else if (net_xmit_drop_count(ret)) {
244 pr_debug("netem: enqueue ret %d\n", ret);
248 /* Requeue packets but don't change time stamp */
249 static int netem_requeue(struct sk_buff *skb, struct Qdisc *sch)
251 struct netem_sched_data *q = qdisc_priv(sch);
254 if ((ret = q->qdisc->ops->requeue(skb, q->qdisc)) == 0) {
256 sch->qstats.requeues++;
262 static unsigned int netem_drop(struct Qdisc* sch)
264 struct netem_sched_data *q = qdisc_priv(sch);
265 unsigned int len = 0;
267 if (q->qdisc->ops->drop && (len = q->qdisc->ops->drop(q->qdisc)) != 0) {
274 static struct sk_buff *netem_dequeue(struct Qdisc *sch)
276 struct netem_sched_data *q = qdisc_priv(sch);
280 if (sch->flags & TCQ_F_THROTTLED)
283 skb = q->qdisc->dequeue(q->qdisc);
285 const struct netem_skb_cb *cb = netem_skb_cb(skb);
286 psched_time_t now = psched_get_time();
288 /* if more time remaining? */
289 if (cb->time_to_send <= now) {
290 pr_debug("netem_dequeue: return skb=%p\n", skb);
295 if (unlikely(q->qdisc->ops->requeue(skb, q->qdisc) != NET_XMIT_SUCCESS)) {
296 qdisc_tree_decrease_qlen(q->qdisc, 1);
298 printk(KERN_ERR "netem: %s could not requeue\n",
302 qdisc_watchdog_schedule(&q->watchdog, cb->time_to_send);
308 static void netem_reset(struct Qdisc *sch)
310 struct netem_sched_data *q = qdisc_priv(sch);
312 qdisc_reset(q->qdisc);
314 qdisc_watchdog_cancel(&q->watchdog);
318 * Distribution data is a variable size payload containing
319 * signed 16 bit values.
321 static int get_dist_table(struct Qdisc *sch, const struct nlattr *attr)
323 struct netem_sched_data *q = qdisc_priv(sch);
324 unsigned long n = nla_len(attr)/sizeof(__s16);
325 const __s16 *data = nla_data(attr);
326 spinlock_t *root_lock;
333 d = kmalloc(sizeof(*d) + n*sizeof(d->table[0]), GFP_KERNEL);
338 for (i = 0; i < n; i++)
339 d->table[i] = data[i];
341 root_lock = qdisc_root_sleeping_lock(sch);
343 spin_lock_bh(root_lock);
344 d = xchg(&q->delay_dist, d);
345 spin_unlock_bh(root_lock);
351 static int get_correlation(struct Qdisc *sch, const struct nlattr *attr)
353 struct netem_sched_data *q = qdisc_priv(sch);
354 const struct tc_netem_corr *c = nla_data(attr);
356 init_crandom(&q->delay_cor, c->delay_corr);
357 init_crandom(&q->loss_cor, c->loss_corr);
358 init_crandom(&q->dup_cor, c->dup_corr);
362 static int get_reorder(struct Qdisc *sch, const struct nlattr *attr)
364 struct netem_sched_data *q = qdisc_priv(sch);
365 const struct tc_netem_reorder *r = nla_data(attr);
367 q->reorder = r->probability;
368 init_crandom(&q->reorder_cor, r->correlation);
372 static int get_corrupt(struct Qdisc *sch, const struct nlattr *attr)
374 struct netem_sched_data *q = qdisc_priv(sch);
375 const struct tc_netem_corrupt *r = nla_data(attr);
377 q->corrupt = r->probability;
378 init_crandom(&q->corrupt_cor, r->correlation);
382 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
383 [TCA_NETEM_CORR] = { .len = sizeof(struct tc_netem_corr) },
384 [TCA_NETEM_REORDER] = { .len = sizeof(struct tc_netem_reorder) },
385 [TCA_NETEM_CORRUPT] = { .len = sizeof(struct tc_netem_corrupt) },
388 static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
389 const struct nla_policy *policy, int len)
391 int nested_len = nla_len(nla) - NLA_ALIGN(len);
395 if (nested_len >= nla_attr_size(0))
396 return nla_parse(tb, maxtype, nla_data(nla) + NLA_ALIGN(len),
398 memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
402 /* Parse netlink message to set options */
403 static int netem_change(struct Qdisc *sch, struct nlattr *opt)
405 struct netem_sched_data *q = qdisc_priv(sch);
406 struct nlattr *tb[TCA_NETEM_MAX + 1];
407 struct tc_netem_qopt *qopt;
413 qopt = nla_data(opt);
414 ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
418 ret = fifo_set_limit(q->qdisc, qopt->limit);
420 pr_debug("netem: can't set fifo limit\n");
424 q->latency = qopt->latency;
425 q->jitter = qopt->jitter;
426 q->limit = qopt->limit;
429 q->loss = qopt->loss;
430 q->duplicate = qopt->duplicate;
432 /* for compatibility with earlier versions.
433 * if gap is set, need to assume 100% probability
438 if (tb[TCA_NETEM_CORR]) {
439 ret = get_correlation(sch, tb[TCA_NETEM_CORR]);
444 if (tb[TCA_NETEM_DELAY_DIST]) {
445 ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST]);
450 if (tb[TCA_NETEM_REORDER]) {
451 ret = get_reorder(sch, tb[TCA_NETEM_REORDER]);
456 if (tb[TCA_NETEM_CORRUPT]) {
457 ret = get_corrupt(sch, tb[TCA_NETEM_CORRUPT]);
466 * Special case version of FIFO queue for use by netem.
467 * It queues in order based on timestamps in skb's
469 struct fifo_sched_data {
471 psched_time_t oldest;
474 static int tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
476 struct fifo_sched_data *q = qdisc_priv(sch);
477 struct sk_buff_head *list = &sch->q;
478 psched_time_t tnext = netem_skb_cb(nskb)->time_to_send;
481 if (likely(skb_queue_len(list) < q->limit)) {
482 /* Optimize for add at tail */
483 if (likely(skb_queue_empty(list) || tnext >= q->oldest)) {
485 return qdisc_enqueue_tail(nskb, sch);
488 skb_queue_reverse_walk(list, skb) {
489 const struct netem_skb_cb *cb = netem_skb_cb(skb);
491 if (tnext >= cb->time_to_send)
495 __skb_queue_after(list, skb, nskb);
497 sch->qstats.backlog += qdisc_pkt_len(nskb);
498 sch->bstats.bytes += qdisc_pkt_len(nskb);
499 sch->bstats.packets++;
501 return NET_XMIT_SUCCESS;
504 return qdisc_reshape_fail(nskb, sch);
507 static int tfifo_init(struct Qdisc *sch, struct nlattr *opt)
509 struct fifo_sched_data *q = qdisc_priv(sch);
512 struct tc_fifo_qopt *ctl = nla_data(opt);
513 if (nla_len(opt) < sizeof(*ctl))
516 q->limit = ctl->limit;
518 q->limit = max_t(u32, qdisc_dev(sch)->tx_queue_len, 1);
520 q->oldest = PSCHED_PASTPERFECT;
524 static int tfifo_dump(struct Qdisc *sch, struct sk_buff *skb)
526 struct fifo_sched_data *q = qdisc_priv(sch);
527 struct tc_fifo_qopt opt = { .limit = q->limit };
529 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
536 static struct Qdisc_ops tfifo_qdisc_ops __read_mostly = {
538 .priv_size = sizeof(struct fifo_sched_data),
539 .enqueue = tfifo_enqueue,
540 .dequeue = qdisc_dequeue_head,
541 .requeue = qdisc_requeue,
542 .drop = qdisc_queue_drop,
544 .reset = qdisc_reset_queue,
545 .change = tfifo_init,
549 static int netem_init(struct Qdisc *sch, struct nlattr *opt)
551 struct netem_sched_data *q = qdisc_priv(sch);
557 qdisc_watchdog_init(&q->watchdog, sch);
559 q->qdisc = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
561 TC_H_MAKE(sch->handle, 1));
563 pr_debug("netem: qdisc create failed\n");
567 ret = netem_change(sch, opt);
569 pr_debug("netem: change failed\n");
570 qdisc_destroy(q->qdisc);
575 static void netem_destroy(struct Qdisc *sch)
577 struct netem_sched_data *q = qdisc_priv(sch);
579 qdisc_watchdog_cancel(&q->watchdog);
580 qdisc_destroy(q->qdisc);
581 kfree(q->delay_dist);
584 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
586 const struct netem_sched_data *q = qdisc_priv(sch);
587 unsigned char *b = skb_tail_pointer(skb);
588 struct nlattr *nla = (struct nlattr *) b;
589 struct tc_netem_qopt qopt;
590 struct tc_netem_corr cor;
591 struct tc_netem_reorder reorder;
592 struct tc_netem_corrupt corrupt;
594 qopt.latency = q->latency;
595 qopt.jitter = q->jitter;
596 qopt.limit = q->limit;
599 qopt.duplicate = q->duplicate;
600 NLA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
602 cor.delay_corr = q->delay_cor.rho;
603 cor.loss_corr = q->loss_cor.rho;
604 cor.dup_corr = q->dup_cor.rho;
605 NLA_PUT(skb, TCA_NETEM_CORR, sizeof(cor), &cor);
607 reorder.probability = q->reorder;
608 reorder.correlation = q->reorder_cor.rho;
609 NLA_PUT(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder);
611 corrupt.probability = q->corrupt;
612 corrupt.correlation = q->corrupt_cor.rho;
613 NLA_PUT(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt);
615 nla->nla_len = skb_tail_pointer(skb) - b;
624 static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
625 struct sk_buff *skb, struct tcmsg *tcm)
627 struct netem_sched_data *q = qdisc_priv(sch);
629 if (cl != 1) /* only one class */
632 tcm->tcm_handle |= TC_H_MIN(1);
633 tcm->tcm_info = q->qdisc->handle;
638 static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
641 struct netem_sched_data *q = qdisc_priv(sch);
647 *old = xchg(&q->qdisc, new);
648 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
650 sch_tree_unlock(sch);
655 static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
657 struct netem_sched_data *q = qdisc_priv(sch);
661 static unsigned long netem_get(struct Qdisc *sch, u32 classid)
666 static void netem_put(struct Qdisc *sch, unsigned long arg)
670 static int netem_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
671 struct nlattr **tca, unsigned long *arg)
676 static int netem_delete(struct Qdisc *sch, unsigned long arg)
681 static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
684 if (walker->count >= walker->skip)
685 if (walker->fn(sch, 1, walker) < 0) {
693 static struct tcf_proto **netem_find_tcf(struct Qdisc *sch, unsigned long cl)
698 static const struct Qdisc_class_ops netem_class_ops = {
699 .graft = netem_graft,
703 .change = netem_change_class,
704 .delete = netem_delete,
706 .tcf_chain = netem_find_tcf,
707 .dump = netem_dump_class,
710 static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
712 .cl_ops = &netem_class_ops,
713 .priv_size = sizeof(struct netem_sched_data),
714 .enqueue = netem_enqueue,
715 .dequeue = netem_dequeue,
716 .requeue = netem_requeue,
719 .reset = netem_reset,
720 .destroy = netem_destroy,
721 .change = netem_change,
723 .owner = THIS_MODULE,
727 static int __init netem_module_init(void)
729 pr_info("netem: version " VERSION "\n");
730 return register_qdisc(&netem_qdisc_ops);
732 static void __exit netem_module_exit(void)
734 unregister_qdisc(&netem_qdisc_ops);
736 module_init(netem_module_init)
737 module_exit(netem_module_exit)
738 MODULE_LICENSE("GPL");