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 = sch->dev->qdisc;
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);
313 /* Pass size change message down to embedded FIFO */
314 static int set_fifo_limit(struct Qdisc *q, int limit)
319 /* Hack to avoid sending change message to non-FIFO */
320 if (strncmp(q->ops->id + 1, "fifo", 4) != 0)
323 rta = kmalloc(RTA_LENGTH(sizeof(struct tc_fifo_qopt)), GFP_KERNEL);
325 rta->rta_type = RTM_NEWQDISC;
326 rta->rta_len = RTA_LENGTH(sizeof(struct tc_fifo_qopt));
327 ((struct tc_fifo_qopt *)RTA_DATA(rta))->limit = limit;
329 ret = q->ops->change(q, rta);
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 rtattr *attr)
341 struct netem_sched_data *q = qdisc_priv(sch);
342 unsigned long n = RTA_PAYLOAD(attr)/sizeof(__s16);
343 const __s16 *data = RTA_DATA(attr);
350 d = kmalloc(sizeof(*d) + n*sizeof(d->table[0]), GFP_KERNEL);
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);
366 static int get_correlation(struct Qdisc *sch, const struct rtattr *attr)
368 struct netem_sched_data *q = qdisc_priv(sch);
369 const struct tc_netem_corr *c = RTA_DATA(attr);
371 if (RTA_PAYLOAD(attr) != sizeof(*c))
374 init_crandom(&q->delay_cor, c->delay_corr);
375 init_crandom(&q->loss_cor, c->loss_corr);
376 init_crandom(&q->dup_cor, c->dup_corr);
380 static int get_reorder(struct Qdisc *sch, const struct rtattr *attr)
382 struct netem_sched_data *q = qdisc_priv(sch);
383 const struct tc_netem_reorder *r = RTA_DATA(attr);
385 if (RTA_PAYLOAD(attr) != sizeof(*r))
388 q->reorder = r->probability;
389 init_crandom(&q->reorder_cor, r->correlation);
393 static int get_corrupt(struct Qdisc *sch, const struct rtattr *attr)
395 struct netem_sched_data *q = qdisc_priv(sch);
396 const struct tc_netem_corrupt *r = RTA_DATA(attr);
398 if (RTA_PAYLOAD(attr) != sizeof(*r))
401 q->corrupt = r->probability;
402 init_crandom(&q->corrupt_cor, r->correlation);
406 /* Parse netlink message to set options */
407 static int netem_change(struct Qdisc *sch, struct rtattr *opt)
409 struct netem_sched_data *q = qdisc_priv(sch);
410 struct tc_netem_qopt *qopt;
413 if (opt == NULL || RTA_PAYLOAD(opt) < sizeof(*qopt))
416 qopt = RTA_DATA(opt);
417 ret = set_fifo_limit(q->qdisc, qopt->limit);
419 pr_debug("netem: can't set fifo limit\n");
423 q->latency = qopt->latency;
424 q->jitter = qopt->jitter;
425 q->limit = qopt->limit;
428 q->loss = qopt->loss;
429 q->duplicate = qopt->duplicate;
431 /* for compatibility with earlier versions.
432 * if gap is set, need to assume 100% probability
437 /* Handle nested options after initial queue options.
438 * Should have put all options in nested format but too late now.
440 if (RTA_PAYLOAD(opt) > sizeof(*qopt)) {
441 struct rtattr *tb[TCA_NETEM_MAX];
442 if (rtattr_parse(tb, TCA_NETEM_MAX,
443 RTA_DATA(opt) + sizeof(*qopt),
444 RTA_PAYLOAD(opt) - sizeof(*qopt)))
447 if (tb[TCA_NETEM_CORR-1]) {
448 ret = get_correlation(sch, tb[TCA_NETEM_CORR-1]);
453 if (tb[TCA_NETEM_DELAY_DIST-1]) {
454 ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST-1]);
459 if (tb[TCA_NETEM_REORDER-1]) {
460 ret = get_reorder(sch, tb[TCA_NETEM_REORDER-1]);
465 if (tb[TCA_NETEM_CORRUPT-1]) {
466 ret = get_corrupt(sch, tb[TCA_NETEM_CORRUPT-1]);
476 * Special case version of FIFO queue for use by netem.
477 * It queues in order based on timestamps in skb's
479 struct fifo_sched_data {
481 psched_time_t oldest;
484 static int tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
486 struct fifo_sched_data *q = qdisc_priv(sch);
487 struct sk_buff_head *list = &sch->q;
488 psched_time_t tnext = ((struct netem_skb_cb *)nskb->cb)->time_to_send;
491 if (likely(skb_queue_len(list) < q->limit)) {
492 /* Optimize for add at tail */
493 if (likely(skb_queue_empty(list) || tnext >= q->oldest)) {
495 return qdisc_enqueue_tail(nskb, sch);
498 skb_queue_reverse_walk(list, skb) {
499 const struct netem_skb_cb *cb
500 = (const struct netem_skb_cb *)skb->cb;
502 if (tnext >= cb->time_to_send)
506 __skb_queue_after(list, skb, nskb);
508 sch->qstats.backlog += nskb->len;
509 sch->bstats.bytes += nskb->len;
510 sch->bstats.packets++;
512 return NET_XMIT_SUCCESS;
515 return qdisc_reshape_fail(nskb, sch);
518 static int tfifo_init(struct Qdisc *sch, struct rtattr *opt)
520 struct fifo_sched_data *q = qdisc_priv(sch);
523 struct tc_fifo_qopt *ctl = RTA_DATA(opt);
524 if (RTA_PAYLOAD(opt) < sizeof(*ctl))
527 q->limit = ctl->limit;
529 q->limit = max_t(u32, sch->dev->tx_queue_len, 1);
531 q->oldest = PSCHED_PASTPERFECT;
535 static int tfifo_dump(struct Qdisc *sch, struct sk_buff *skb)
537 struct fifo_sched_data *q = qdisc_priv(sch);
538 struct tc_fifo_qopt opt = { .limit = q->limit };
540 RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
547 static struct Qdisc_ops tfifo_qdisc_ops __read_mostly = {
549 .priv_size = sizeof(struct fifo_sched_data),
550 .enqueue = tfifo_enqueue,
551 .dequeue = qdisc_dequeue_head,
552 .requeue = qdisc_requeue,
553 .drop = qdisc_queue_drop,
555 .reset = qdisc_reset_queue,
556 .change = tfifo_init,
560 static int netem_init(struct Qdisc *sch, struct rtattr *opt)
562 struct netem_sched_data *q = qdisc_priv(sch);
568 qdisc_watchdog_init(&q->watchdog, sch);
570 q->qdisc = qdisc_create_dflt(sch->dev, &tfifo_qdisc_ops,
571 TC_H_MAKE(sch->handle, 1));
573 pr_debug("netem: qdisc create failed\n");
577 ret = netem_change(sch, opt);
579 pr_debug("netem: change failed\n");
580 qdisc_destroy(q->qdisc);
585 static void netem_destroy(struct Qdisc *sch)
587 struct netem_sched_data *q = qdisc_priv(sch);
589 qdisc_watchdog_cancel(&q->watchdog);
590 qdisc_destroy(q->qdisc);
591 kfree(q->delay_dist);
594 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
596 const struct netem_sched_data *q = qdisc_priv(sch);
597 unsigned char *b = skb_tail_pointer(skb);
598 struct rtattr *rta = (struct rtattr *) b;
599 struct tc_netem_qopt qopt;
600 struct tc_netem_corr cor;
601 struct tc_netem_reorder reorder;
602 struct tc_netem_corrupt corrupt;
604 qopt.latency = q->latency;
605 qopt.jitter = q->jitter;
606 qopt.limit = q->limit;
609 qopt.duplicate = q->duplicate;
610 RTA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
612 cor.delay_corr = q->delay_cor.rho;
613 cor.loss_corr = q->loss_cor.rho;
614 cor.dup_corr = q->dup_cor.rho;
615 RTA_PUT(skb, TCA_NETEM_CORR, sizeof(cor), &cor);
617 reorder.probability = q->reorder;
618 reorder.correlation = q->reorder_cor.rho;
619 RTA_PUT(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder);
621 corrupt.probability = q->corrupt;
622 corrupt.correlation = q->corrupt_cor.rho;
623 RTA_PUT(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt);
625 rta->rta_len = skb_tail_pointer(skb) - b;
634 static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
635 struct sk_buff *skb, struct tcmsg *tcm)
637 struct netem_sched_data *q = qdisc_priv(sch);
639 if (cl != 1) /* only one class */
642 tcm->tcm_handle |= TC_H_MIN(1);
643 tcm->tcm_info = q->qdisc->handle;
648 static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
651 struct netem_sched_data *q = qdisc_priv(sch);
657 *old = xchg(&q->qdisc, new);
658 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
660 sch_tree_unlock(sch);
665 static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
667 struct netem_sched_data *q = qdisc_priv(sch);
671 static unsigned long netem_get(struct Qdisc *sch, u32 classid)
676 static void netem_put(struct Qdisc *sch, unsigned long arg)
680 static int netem_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
681 struct rtattr **tca, unsigned long *arg)
686 static int netem_delete(struct Qdisc *sch, unsigned long arg)
691 static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
694 if (walker->count >= walker->skip)
695 if (walker->fn(sch, 1, walker) < 0) {
703 static struct tcf_proto **netem_find_tcf(struct Qdisc *sch, unsigned long cl)
708 static const struct Qdisc_class_ops netem_class_ops = {
709 .graft = netem_graft,
713 .change = netem_change_class,
714 .delete = netem_delete,
716 .tcf_chain = netem_find_tcf,
717 .dump = netem_dump_class,
720 static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
722 .cl_ops = &netem_class_ops,
723 .priv_size = sizeof(struct netem_sched_data),
724 .enqueue = netem_enqueue,
725 .dequeue = netem_dequeue,
726 .requeue = netem_requeue,
729 .reset = netem_reset,
730 .destroy = netem_destroy,
731 .change = netem_change,
733 .owner = THIS_MODULE,
737 static int __init netem_module_init(void)
739 pr_info("netem: version " VERSION "\n");
740 return register_qdisc(&netem_qdisc_ops);
742 static void __exit netem_module_exit(void)
744 unregister_qdisc(&netem_qdisc_ops);
746 module_init(netem_module_init)
747 module_exit(netem_module_exit)
748 MODULE_LICENSE("GPL");