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 static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb)
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;
92 /* init_crandom - initialize correlated random number generator
93 * Use entropy source for initial seed.
95 static void init_crandom(struct crndstate *state, unsigned long rho)
98 state->last = net_random();
101 /* get_crandom - correlated random number generator
102 * Next number depends on last value.
103 * rho is scaled to avoid floating point.
105 static u32 get_crandom(struct crndstate *state)
108 unsigned long answer;
110 if (state->rho == 0) /* no correlation */
113 value = net_random();
114 rho = (u64)state->rho + 1;
115 answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32;
116 state->last = answer;
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.
124 static psched_tdiff_t tabledist(psched_tdiff_t mu, psched_tdiff_t sigma,
125 struct crndstate *state,
126 const struct disttable *dist)
135 rnd = get_crandom(state);
137 /* default uniform distribution */
139 return (rnd % (2*sigma)) - sigma + mu;
141 t = dist->table[rnd % dist->size];
142 x = (sigma % NETEM_DIST_SCALE) * t;
144 x += NETEM_DIST_SCALE/2;
146 x -= NETEM_DIST_SCALE/2;
148 return x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu;
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.
157 static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch)
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;
166 pr_debug("netem_enqueue skb=%p\n", skb);
168 /* Random duplication */
169 if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor))
172 /* Random packet drop 0 => none, ~0 => all */
173 if (q->loss && q->loss >= get_crandom(&q->loss_cor))
179 return NET_XMIT_BYPASS;
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.
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... */
194 qdisc_enqueue_root(skb2, rootq);
195 q->duplicate = dupsave;
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.
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))) {
209 return NET_XMIT_DROP;
212 skb->data[net_random() % skb_headlen(skb)] ^= 1<<(net_random() % 8);
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)) {
220 psched_tdiff_t delay;
222 delay = tabledist(q->latency, q->jitter,
223 &q->delay_cor, q->delay_dist);
225 now = psched_get_time();
226 cb->time_to_send = now + delay;
228 ret = qdisc_enqueue(skb, q->qdisc);
231 * Do re-ordering by putting one out of N packets at the front
234 cb->time_to_send = psched_get_time();
236 ret = q->qdisc->ops->requeue(skb, q->qdisc);
239 if (likely(ret == NET_XMIT_SUCCESS)) {
241 sch->bstats.bytes += qdisc_pkt_len(skb);
242 sch->bstats.packets++;
246 pr_debug("netem: enqueue ret %d\n", ret);
250 /* Requeue packets but don't change time stamp */
251 static int netem_requeue(struct sk_buff *skb, struct Qdisc *sch)
253 struct netem_sched_data *q = qdisc_priv(sch);
256 if ((ret = q->qdisc->ops->requeue(skb, q->qdisc)) == 0) {
258 sch->qstats.requeues++;
264 static unsigned int netem_drop(struct Qdisc* sch)
266 struct netem_sched_data *q = qdisc_priv(sch);
267 unsigned int len = 0;
269 if (q->qdisc->ops->drop && (len = q->qdisc->ops->drop(q->qdisc)) != 0) {
276 static struct sk_buff *netem_dequeue(struct Qdisc *sch)
278 struct netem_sched_data *q = qdisc_priv(sch);
282 if (sch->flags & TCQ_F_THROTTLED)
285 skb = q->qdisc->dequeue(q->qdisc);
287 const struct netem_skb_cb *cb = netem_skb_cb(skb);
288 psched_time_t now = psched_get_time();
290 /* if more time remaining? */
291 if (cb->time_to_send <= now) {
292 pr_debug("netem_dequeue: return skb=%p\n", skb);
297 if (unlikely(q->qdisc->ops->requeue(skb, q->qdisc) != NET_XMIT_SUCCESS)) {
298 qdisc_tree_decrease_qlen(q->qdisc, 1);
300 printk(KERN_ERR "netem: %s could not requeue\n",
304 qdisc_watchdog_schedule(&q->watchdog, cb->time_to_send);
310 static void netem_reset(struct Qdisc *sch)
312 struct netem_sched_data *q = qdisc_priv(sch);
314 qdisc_reset(q->qdisc);
316 qdisc_watchdog_cancel(&q->watchdog);
320 * Distribution data is a variable size payload containing
321 * signed 16 bit values.
323 static int get_dist_table(struct Qdisc *sch, const struct nlattr *attr)
325 struct netem_sched_data *q = qdisc_priv(sch);
326 unsigned long n = nla_len(attr)/sizeof(__s16);
327 const __s16 *data = nla_data(attr);
328 spinlock_t *root_lock;
335 d = kmalloc(sizeof(*d) + n*sizeof(d->table[0]), GFP_KERNEL);
340 for (i = 0; i < n; i++)
341 d->table[i] = data[i];
343 root_lock = qdisc_root_lock(sch);
345 spin_lock_bh(root_lock);
346 d = xchg(&q->delay_dist, d);
347 spin_unlock_bh(root_lock);
353 static int get_correlation(struct Qdisc *sch, const struct nlattr *attr)
355 struct netem_sched_data *q = qdisc_priv(sch);
356 const struct tc_netem_corr *c = nla_data(attr);
358 init_crandom(&q->delay_cor, c->delay_corr);
359 init_crandom(&q->loss_cor, c->loss_corr);
360 init_crandom(&q->dup_cor, c->dup_corr);
364 static int get_reorder(struct Qdisc *sch, const struct nlattr *attr)
366 struct netem_sched_data *q = qdisc_priv(sch);
367 const struct tc_netem_reorder *r = nla_data(attr);
369 q->reorder = r->probability;
370 init_crandom(&q->reorder_cor, r->correlation);
374 static int get_corrupt(struct Qdisc *sch, const struct nlattr *attr)
376 struct netem_sched_data *q = qdisc_priv(sch);
377 const struct tc_netem_corrupt *r = nla_data(attr);
379 q->corrupt = r->probability;
380 init_crandom(&q->corrupt_cor, r->correlation);
384 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
385 [TCA_NETEM_CORR] = { .len = sizeof(struct tc_netem_corr) },
386 [TCA_NETEM_REORDER] = { .len = sizeof(struct tc_netem_reorder) },
387 [TCA_NETEM_CORRUPT] = { .len = sizeof(struct tc_netem_corrupt) },
390 /* Parse netlink message to set options */
391 static int netem_change(struct Qdisc *sch, struct nlattr *opt)
393 struct netem_sched_data *q = qdisc_priv(sch);
394 struct nlattr *tb[TCA_NETEM_MAX + 1];
395 struct tc_netem_qopt *qopt;
401 ret = nla_parse_nested_compat(tb, TCA_NETEM_MAX, opt, netem_policy,
402 qopt, sizeof(*qopt));
406 ret = fifo_set_limit(q->qdisc, qopt->limit);
408 pr_debug("netem: can't set fifo limit\n");
412 q->latency = qopt->latency;
413 q->jitter = qopt->jitter;
414 q->limit = qopt->limit;
417 q->loss = qopt->loss;
418 q->duplicate = qopt->duplicate;
420 /* for compatibility with earlier versions.
421 * if gap is set, need to assume 100% probability
426 if (tb[TCA_NETEM_CORR]) {
427 ret = get_correlation(sch, tb[TCA_NETEM_CORR]);
432 if (tb[TCA_NETEM_DELAY_DIST]) {
433 ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST]);
438 if (tb[TCA_NETEM_REORDER]) {
439 ret = get_reorder(sch, tb[TCA_NETEM_REORDER]);
444 if (tb[TCA_NETEM_CORRUPT]) {
445 ret = get_corrupt(sch, tb[TCA_NETEM_CORRUPT]);
454 * Special case version of FIFO queue for use by netem.
455 * It queues in order based on timestamps in skb's
457 struct fifo_sched_data {
459 psched_time_t oldest;
462 static int tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
464 struct fifo_sched_data *q = qdisc_priv(sch);
465 struct sk_buff_head *list = &sch->q;
466 psched_time_t tnext = netem_skb_cb(nskb)->time_to_send;
469 if (likely(skb_queue_len(list) < q->limit)) {
470 /* Optimize for add at tail */
471 if (likely(skb_queue_empty(list) || tnext >= q->oldest)) {
473 return qdisc_enqueue_tail(nskb, sch);
476 skb_queue_reverse_walk(list, skb) {
477 const struct netem_skb_cb *cb = netem_skb_cb(skb);
479 if (tnext >= cb->time_to_send)
483 __skb_queue_after(list, skb, nskb);
485 sch->qstats.backlog += qdisc_pkt_len(nskb);
486 sch->bstats.bytes += qdisc_pkt_len(nskb);
487 sch->bstats.packets++;
489 return NET_XMIT_SUCCESS;
492 return qdisc_reshape_fail(nskb, sch);
495 static int tfifo_init(struct Qdisc *sch, struct nlattr *opt)
497 struct fifo_sched_data *q = qdisc_priv(sch);
500 struct tc_fifo_qopt *ctl = nla_data(opt);
501 if (nla_len(opt) < sizeof(*ctl))
504 q->limit = ctl->limit;
506 q->limit = max_t(u32, qdisc_dev(sch)->tx_queue_len, 1);
508 q->oldest = PSCHED_PASTPERFECT;
512 static int tfifo_dump(struct Qdisc *sch, struct sk_buff *skb)
514 struct fifo_sched_data *q = qdisc_priv(sch);
515 struct tc_fifo_qopt opt = { .limit = q->limit };
517 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
524 static struct Qdisc_ops tfifo_qdisc_ops __read_mostly = {
526 .priv_size = sizeof(struct fifo_sched_data),
527 .enqueue = tfifo_enqueue,
528 .dequeue = qdisc_dequeue_head,
529 .requeue = qdisc_requeue,
530 .drop = qdisc_queue_drop,
532 .reset = qdisc_reset_queue,
533 .change = tfifo_init,
537 static int netem_init(struct Qdisc *sch, struct nlattr *opt)
539 struct netem_sched_data *q = qdisc_priv(sch);
545 qdisc_watchdog_init(&q->watchdog, sch);
547 q->qdisc = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
549 TC_H_MAKE(sch->handle, 1));
551 pr_debug("netem: qdisc create failed\n");
555 ret = netem_change(sch, opt);
557 pr_debug("netem: change failed\n");
558 qdisc_destroy(q->qdisc);
563 static void netem_destroy(struct Qdisc *sch)
565 struct netem_sched_data *q = qdisc_priv(sch);
567 qdisc_watchdog_cancel(&q->watchdog);
568 qdisc_destroy(q->qdisc);
569 kfree(q->delay_dist);
572 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
574 const struct netem_sched_data *q = qdisc_priv(sch);
575 unsigned char *b = skb_tail_pointer(skb);
576 struct nlattr *nla = (struct nlattr *) b;
577 struct tc_netem_qopt qopt;
578 struct tc_netem_corr cor;
579 struct tc_netem_reorder reorder;
580 struct tc_netem_corrupt corrupt;
582 qopt.latency = q->latency;
583 qopt.jitter = q->jitter;
584 qopt.limit = q->limit;
587 qopt.duplicate = q->duplicate;
588 NLA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
590 cor.delay_corr = q->delay_cor.rho;
591 cor.loss_corr = q->loss_cor.rho;
592 cor.dup_corr = q->dup_cor.rho;
593 NLA_PUT(skb, TCA_NETEM_CORR, sizeof(cor), &cor);
595 reorder.probability = q->reorder;
596 reorder.correlation = q->reorder_cor.rho;
597 NLA_PUT(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder);
599 corrupt.probability = q->corrupt;
600 corrupt.correlation = q->corrupt_cor.rho;
601 NLA_PUT(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt);
603 nla->nla_len = skb_tail_pointer(skb) - b;
612 static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
613 struct sk_buff *skb, struct tcmsg *tcm)
615 struct netem_sched_data *q = qdisc_priv(sch);
617 if (cl != 1) /* only one class */
620 tcm->tcm_handle |= TC_H_MIN(1);
621 tcm->tcm_info = q->qdisc->handle;
626 static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
629 struct netem_sched_data *q = qdisc_priv(sch);
635 *old = xchg(&q->qdisc, new);
636 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
638 sch_tree_unlock(sch);
643 static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
645 struct netem_sched_data *q = qdisc_priv(sch);
649 static unsigned long netem_get(struct Qdisc *sch, u32 classid)
654 static void netem_put(struct Qdisc *sch, unsigned long arg)
658 static int netem_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
659 struct nlattr **tca, unsigned long *arg)
664 static int netem_delete(struct Qdisc *sch, unsigned long arg)
669 static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
672 if (walker->count >= walker->skip)
673 if (walker->fn(sch, 1, walker) < 0) {
681 static struct tcf_proto **netem_find_tcf(struct Qdisc *sch, unsigned long cl)
686 static const struct Qdisc_class_ops netem_class_ops = {
687 .graft = netem_graft,
691 .change = netem_change_class,
692 .delete = netem_delete,
694 .tcf_chain = netem_find_tcf,
695 .dump = netem_dump_class,
698 static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
700 .cl_ops = &netem_class_ops,
701 .priv_size = sizeof(struct netem_sched_data),
702 .enqueue = netem_enqueue,
703 .dequeue = netem_dequeue,
704 .requeue = netem_requeue,
707 .reset = netem_reset,
708 .destroy = netem_destroy,
709 .change = netem_change,
711 .owner = THIS_MODULE,
715 static int __init netem_module_init(void)
717 pr_info("netem: version " VERSION "\n");
718 return register_qdisc(&netem_qdisc_ops);
720 static void __exit netem_module_exit(void)
722 unregister_qdisc(&netem_qdisc_ops);
724 module_init(netem_module_init)
725 module_exit(netem_module_exit)
726 MODULE_LICENSE("GPL");