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/slab.h>
18 #include <linux/types.h>
19 #include <linux/kernel.h>
20 #include <linux/errno.h>
21 #include <linux/skbuff.h>
22 #include <linux/vmalloc.h>
23 #include <linux/rtnetlink.h>
25 #include <net/netlink.h>
26 #include <net/pkt_sched.h>
30 /* Network Emulation Queuing algorithm.
31 ====================================
33 Sources: [1] Mark Carson, Darrin Santay, "NIST Net - A Linux-based
34 Network Emulation Tool
35 [2] Luigi Rizzo, DummyNet for FreeBSD
37 ----------------------------------------------------------------
39 This started out as a simple way to delay outgoing packets to
40 test TCP but has grown to include most of the functionality
41 of a full blown network emulator like NISTnet. It can delay
42 packets and add random jitter (and correlation). The random
43 distribution can be loaded from a table as well to provide
44 normal, Pareto, or experimental curves. Packet loss,
45 duplication, and reordering can also be emulated.
47 This qdisc does not do classification that can be handled in
48 layering other disciplines. It does not need to do bandwidth
49 control either since that can be handled by using token
50 bucket or other rate control.
52 Correlated Loss Generator models
54 Added generation of correlated loss according to the
55 "Gilbert-Elliot" model, a 4-state markov model.
58 [1] NetemCLG Home http://netgroup.uniroma2.it/NetemCLG
59 [2] S. Salsano, F. Ludovici, A. Ordine, "Definition of a general
60 and intuitive loss model for packet networks and its implementation
61 in the Netem module in the Linux kernel", available in [1]
63 Authors: Stefano Salsano <stefano.salsano at uniroma2.it
64 Fabio Ludovici <fabio.ludovici at yahoo.it>
67 struct netem_sched_data {
69 struct qdisc_watchdog watchdog;
71 psched_tdiff_t latency;
72 psched_tdiff_t jitter;
85 } delay_cor, loss_cor, dup_cor, reorder_cor, corrupt_cor;
98 /* Correlated Loss Generation models */
100 /* state of the Markov chain */
103 /* 4-states and Gilbert-Elliot models */
104 u32 a1; /* p13 for 4-states or p for GE */
105 u32 a2; /* p31 for 4-states or r for GE */
106 u32 a3; /* p32 for 4-states or h for GE */
107 u32 a4; /* p14 for 4-states or 1-k for GE */
108 u32 a5; /* p23 used only in 4-states */
113 /* Time stamp put into socket buffer control block */
114 struct netem_skb_cb {
115 psched_time_t time_to_send;
118 static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb)
120 qdisc_cb_private_validate(skb, sizeof(struct netem_skb_cb));
121 return (struct netem_skb_cb *)qdisc_skb_cb(skb)->data;
124 /* init_crandom - initialize correlated random number generator
125 * Use entropy source for initial seed.
127 static void init_crandom(struct crndstate *state, unsigned long rho)
130 state->last = net_random();
133 /* get_crandom - correlated random number generator
134 * Next number depends on last value.
135 * rho is scaled to avoid floating point.
137 static u32 get_crandom(struct crndstate *state)
140 unsigned long answer;
142 if (state->rho == 0) /* no correlation */
145 value = net_random();
146 rho = (u64)state->rho + 1;
147 answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32;
148 state->last = answer;
152 /* loss_4state - 4-state model loss generator
153 * Generates losses according to the 4-state Markov chain adopted in
154 * the GI (General and Intuitive) loss model.
156 static bool loss_4state(struct netem_sched_data *q)
158 struct clgstate *clg = &q->clg;
159 u32 rnd = net_random();
162 * Makes a comparison between rnd and the transition
163 * probabilities outgoing from the current state, then decides the
164 * next state and if the next packet has to be transmitted or lost.
165 * The four states correspond to:
166 * 1 => successfully transmitted packets within a gap period
167 * 4 => isolated losses within a gap period
168 * 3 => lost packets within a burst period
169 * 2 => successfully transmitted packets within a burst period
171 switch (clg->state) {
176 } else if (clg->a4 < rnd && rnd < clg->a1) {
179 } else if (clg->a1 < rnd)
194 else if (clg->a3 < rnd && rnd < clg->a2 + clg->a3) {
197 } else if (clg->a2 + clg->a3 < rnd) {
210 /* loss_gilb_ell - Gilbert-Elliot model loss generator
211 * Generates losses according to the Gilbert-Elliot loss model or
212 * its special cases (Gilbert or Simple Gilbert)
214 * Makes a comparison between random number and the transition
215 * probabilities outgoing from the current state, then decides the
216 * next state. A second random number is extracted and the comparison
217 * with the loss probability of the current state decides if the next
218 * packet will be transmitted or lost.
220 static bool loss_gilb_ell(struct netem_sched_data *q)
222 struct clgstate *clg = &q->clg;
224 switch (clg->state) {
226 if (net_random() < clg->a1)
228 if (net_random() < clg->a4)
231 if (net_random() < clg->a2)
233 if (clg->a3 > net_random())
240 static bool loss_event(struct netem_sched_data *q)
242 switch (q->loss_model) {
244 /* Random packet drop 0 => none, ~0 => all */
245 return q->loss && q->loss >= get_crandom(&q->loss_cor);
248 /* 4state loss model algorithm (used also for GI model)
249 * Extracts a value from the markov 4 state loss generator,
250 * if it is 1 drops a packet and if needed writes the event in
253 return loss_4state(q);
256 /* Gilbert-Elliot loss model algorithm
257 * Extracts a value from the Gilbert-Elliot loss generator,
258 * if it is 1 drops a packet and if needed writes the event in
261 return loss_gilb_ell(q);
264 return false; /* not reached */
268 /* tabledist - return a pseudo-randomly distributed value with mean mu and
269 * std deviation sigma. Uses table lookup to approximate the desired
270 * distribution, and a uniformly-distributed pseudo-random source.
272 static psched_tdiff_t tabledist(psched_tdiff_t mu, psched_tdiff_t sigma,
273 struct crndstate *state,
274 const struct disttable *dist)
283 rnd = get_crandom(state);
285 /* default uniform distribution */
287 return (rnd % (2*sigma)) - sigma + mu;
289 t = dist->table[rnd % dist->size];
290 x = (sigma % NETEM_DIST_SCALE) * t;
292 x += NETEM_DIST_SCALE/2;
294 x -= NETEM_DIST_SCALE/2;
296 return x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu;
300 * Insert one skb into qdisc.
301 * Note: parent depends on return value to account for queue length.
302 * NET_XMIT_DROP: queue length didn't change.
303 * NET_XMIT_SUCCESS: one skb was queued.
305 static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch)
307 struct netem_sched_data *q = qdisc_priv(sch);
308 /* We don't fill cb now as skb_unshare() may invalidate it */
309 struct netem_skb_cb *cb;
310 struct sk_buff *skb2;
314 /* Random duplication */
315 if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor))
325 return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
331 * If we need to duplicate packet, then re-insert at top of the
332 * qdisc tree, since parent queuer expects that only one
333 * skb will be queued.
335 if (count > 1 && (skb2 = skb_clone(skb, GFP_ATOMIC)) != NULL) {
336 struct Qdisc *rootq = qdisc_root(sch);
337 u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
340 qdisc_enqueue_root(skb2, rootq);
341 q->duplicate = dupsave;
345 * Randomized packet corruption.
346 * Make copy if needed since we are modifying
347 * If packet is going to be hardware checksummed, then
348 * do it now in software before we mangle it.
350 if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor)) {
351 if (!(skb = skb_unshare(skb, GFP_ATOMIC)) ||
352 (skb->ip_summed == CHECKSUM_PARTIAL &&
353 skb_checksum_help(skb))) {
355 return NET_XMIT_DROP;
358 skb->data[net_random() % skb_headlen(skb)] ^= 1<<(net_random() % 8);
361 cb = netem_skb_cb(skb);
362 if (q->gap == 0 || /* not doing reordering */
363 q->counter < q->gap || /* inside last reordering gap */
364 q->reorder < get_crandom(&q->reorder_cor)) {
366 psched_tdiff_t delay;
368 delay = tabledist(q->latency, q->jitter,
369 &q->delay_cor, q->delay_dist);
371 now = psched_get_time();
372 cb->time_to_send = now + delay;
374 ret = qdisc_enqueue(skb, q->qdisc);
377 * Do re-ordering by putting one out of N packets at the front
380 cb->time_to_send = psched_get_time();
383 __skb_queue_head(&q->qdisc->q, skb);
384 sch->qstats.backlog += qdisc_pkt_len(skb);
385 sch->qstats.requeues++;
386 ret = NET_XMIT_SUCCESS;
389 if (ret != NET_XMIT_SUCCESS) {
390 if (net_xmit_drop_count(ret)) {
397 return NET_XMIT_SUCCESS;
400 static unsigned int netem_drop(struct Qdisc *sch)
402 struct netem_sched_data *q = qdisc_priv(sch);
403 unsigned int len = 0;
405 if (q->qdisc->ops->drop && (len = q->qdisc->ops->drop(q->qdisc)) != 0) {
412 static struct sk_buff *netem_dequeue(struct Qdisc *sch)
414 struct netem_sched_data *q = qdisc_priv(sch);
417 if (qdisc_is_throttled(sch))
420 skb = q->qdisc->ops->peek(q->qdisc);
422 const struct netem_skb_cb *cb = netem_skb_cb(skb);
423 psched_time_t now = psched_get_time();
425 /* if more time remaining? */
426 if (cb->time_to_send <= now) {
427 skb = qdisc_dequeue_peeked(q->qdisc);
431 #ifdef CONFIG_NET_CLS_ACT
433 * If it's at ingress let's pretend the delay is
434 * from the network (tstamp will be updated).
436 if (G_TC_FROM(skb->tc_verd) & AT_INGRESS)
437 skb->tstamp.tv64 = 0;
441 qdisc_unthrottled(sch);
442 qdisc_bstats_update(sch, skb);
446 qdisc_watchdog_schedule(&q->watchdog, cb->time_to_send);
452 static void netem_reset(struct Qdisc *sch)
454 struct netem_sched_data *q = qdisc_priv(sch);
456 qdisc_reset(q->qdisc);
458 qdisc_watchdog_cancel(&q->watchdog);
461 static void dist_free(struct disttable *d)
464 if (is_vmalloc_addr(d))
472 * Distribution data is a variable size payload containing
473 * signed 16 bit values.
475 static int get_dist_table(struct Qdisc *sch, const struct nlattr *attr)
477 struct netem_sched_data *q = qdisc_priv(sch);
478 size_t n = nla_len(attr)/sizeof(__s16);
479 const __s16 *data = nla_data(attr);
480 spinlock_t *root_lock;
485 if (n > NETEM_DIST_MAX)
488 s = sizeof(struct disttable) + n * sizeof(s16);
489 d = kmalloc(s, GFP_KERNEL);
496 for (i = 0; i < n; i++)
497 d->table[i] = data[i];
499 root_lock = qdisc_root_sleeping_lock(sch);
501 spin_lock_bh(root_lock);
502 dist_free(q->delay_dist);
504 spin_unlock_bh(root_lock);
508 static void get_correlation(struct Qdisc *sch, const struct nlattr *attr)
510 struct netem_sched_data *q = qdisc_priv(sch);
511 const struct tc_netem_corr *c = nla_data(attr);
513 init_crandom(&q->delay_cor, c->delay_corr);
514 init_crandom(&q->loss_cor, c->loss_corr);
515 init_crandom(&q->dup_cor, c->dup_corr);
518 static void get_reorder(struct Qdisc *sch, const struct nlattr *attr)
520 struct netem_sched_data *q = qdisc_priv(sch);
521 const struct tc_netem_reorder *r = nla_data(attr);
523 q->reorder = r->probability;
524 init_crandom(&q->reorder_cor, r->correlation);
527 static void get_corrupt(struct Qdisc *sch, const struct nlattr *attr)
529 struct netem_sched_data *q = qdisc_priv(sch);
530 const struct tc_netem_corrupt *r = nla_data(attr);
532 q->corrupt = r->probability;
533 init_crandom(&q->corrupt_cor, r->correlation);
536 static int get_loss_clg(struct Qdisc *sch, const struct nlattr *attr)
538 struct netem_sched_data *q = qdisc_priv(sch);
539 const struct nlattr *la;
542 nla_for_each_nested(la, attr, rem) {
543 u16 type = nla_type(la);
546 case NETEM_LOSS_GI: {
547 const struct tc_netem_gimodel *gi = nla_data(la);
549 if (nla_len(la) != sizeof(struct tc_netem_gimodel)) {
550 pr_info("netem: incorrect gi model size\n");
554 q->loss_model = CLG_4_STATES;
565 case NETEM_LOSS_GE: {
566 const struct tc_netem_gemodel *ge = nla_data(la);
568 if (nla_len(la) != sizeof(struct tc_netem_gemodel)) {
569 pr_info("netem: incorrect gi model size\n");
573 q->loss_model = CLG_GILB_ELL;
583 pr_info("netem: unknown loss type %u\n", type);
591 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
592 [TCA_NETEM_CORR] = { .len = sizeof(struct tc_netem_corr) },
593 [TCA_NETEM_REORDER] = { .len = sizeof(struct tc_netem_reorder) },
594 [TCA_NETEM_CORRUPT] = { .len = sizeof(struct tc_netem_corrupt) },
595 [TCA_NETEM_LOSS] = { .type = NLA_NESTED },
598 static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
599 const struct nla_policy *policy, int len)
601 int nested_len = nla_len(nla) - NLA_ALIGN(len);
603 if (nested_len < 0) {
604 pr_info("netem: invalid attributes len %d\n", nested_len);
608 if (nested_len >= nla_attr_size(0))
609 return nla_parse(tb, maxtype, nla_data(nla) + NLA_ALIGN(len),
612 memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
616 /* Parse netlink message to set options */
617 static int netem_change(struct Qdisc *sch, struct nlattr *opt)
619 struct netem_sched_data *q = qdisc_priv(sch);
620 struct nlattr *tb[TCA_NETEM_MAX + 1];
621 struct tc_netem_qopt *qopt;
627 qopt = nla_data(opt);
628 ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
632 ret = fifo_set_limit(q->qdisc, qopt->limit);
634 pr_info("netem: can't set fifo limit\n");
638 q->latency = qopt->latency;
639 q->jitter = qopt->jitter;
640 q->limit = qopt->limit;
643 q->loss = qopt->loss;
644 q->duplicate = qopt->duplicate;
646 /* for compatibility with earlier versions.
647 * if gap is set, need to assume 100% probability
652 if (tb[TCA_NETEM_CORR])
653 get_correlation(sch, tb[TCA_NETEM_CORR]);
655 if (tb[TCA_NETEM_DELAY_DIST]) {
656 ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST]);
661 if (tb[TCA_NETEM_REORDER])
662 get_reorder(sch, tb[TCA_NETEM_REORDER]);
664 if (tb[TCA_NETEM_CORRUPT])
665 get_corrupt(sch, tb[TCA_NETEM_CORRUPT]);
667 q->loss_model = CLG_RANDOM;
668 if (tb[TCA_NETEM_LOSS])
669 ret = get_loss_clg(sch, tb[TCA_NETEM_LOSS]);
675 * Special case version of FIFO queue for use by netem.
676 * It queues in order based on timestamps in skb's
678 struct fifo_sched_data {
680 psched_time_t oldest;
683 static int tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
685 struct fifo_sched_data *q = qdisc_priv(sch);
686 struct sk_buff_head *list = &sch->q;
687 psched_time_t tnext = netem_skb_cb(nskb)->time_to_send;
690 if (likely(skb_queue_len(list) < q->limit)) {
691 /* Optimize for add at tail */
692 if (likely(skb_queue_empty(list) || tnext >= q->oldest)) {
694 return qdisc_enqueue_tail(nskb, sch);
697 skb_queue_reverse_walk(list, skb) {
698 const struct netem_skb_cb *cb = netem_skb_cb(skb);
700 if (tnext >= cb->time_to_send)
704 __skb_queue_after(list, skb, nskb);
706 sch->qstats.backlog += qdisc_pkt_len(nskb);
708 return NET_XMIT_SUCCESS;
711 return qdisc_reshape_fail(nskb, sch);
714 static int tfifo_init(struct Qdisc *sch, struct nlattr *opt)
716 struct fifo_sched_data *q = qdisc_priv(sch);
719 struct tc_fifo_qopt *ctl = nla_data(opt);
720 if (nla_len(opt) < sizeof(*ctl))
723 q->limit = ctl->limit;
725 q->limit = max_t(u32, qdisc_dev(sch)->tx_queue_len, 1);
727 q->oldest = PSCHED_PASTPERFECT;
731 static int tfifo_dump(struct Qdisc *sch, struct sk_buff *skb)
733 struct fifo_sched_data *q = qdisc_priv(sch);
734 struct tc_fifo_qopt opt = { .limit = q->limit };
736 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
743 static struct Qdisc_ops tfifo_qdisc_ops __read_mostly = {
745 .priv_size = sizeof(struct fifo_sched_data),
746 .enqueue = tfifo_enqueue,
747 .dequeue = qdisc_dequeue_head,
748 .peek = qdisc_peek_head,
749 .drop = qdisc_queue_drop,
751 .reset = qdisc_reset_queue,
752 .change = tfifo_init,
756 static int netem_init(struct Qdisc *sch, struct nlattr *opt)
758 struct netem_sched_data *q = qdisc_priv(sch);
764 qdisc_watchdog_init(&q->watchdog, sch);
766 q->loss_model = CLG_RANDOM;
767 q->qdisc = qdisc_create_dflt(sch->dev_queue, &tfifo_qdisc_ops,
768 TC_H_MAKE(sch->handle, 1));
770 pr_notice("netem: qdisc create tfifo qdisc failed\n");
774 ret = netem_change(sch, opt);
776 pr_info("netem: change failed\n");
777 qdisc_destroy(q->qdisc);
782 static void netem_destroy(struct Qdisc *sch)
784 struct netem_sched_data *q = qdisc_priv(sch);
786 qdisc_watchdog_cancel(&q->watchdog);
787 qdisc_destroy(q->qdisc);
788 dist_free(q->delay_dist);
791 static int dump_loss_model(const struct netem_sched_data *q,
796 nest = nla_nest_start(skb, TCA_NETEM_LOSS);
798 goto nla_put_failure;
800 switch (q->loss_model) {
802 /* legacy loss model */
803 nla_nest_cancel(skb, nest);
804 return 0; /* no data */
807 struct tc_netem_gimodel gi = {
815 NLA_PUT(skb, NETEM_LOSS_GI, sizeof(gi), &gi);
819 struct tc_netem_gemodel ge = {
826 NLA_PUT(skb, NETEM_LOSS_GE, sizeof(ge), &ge);
831 nla_nest_end(skb, nest);
835 nla_nest_cancel(skb, nest);
839 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
841 const struct netem_sched_data *q = qdisc_priv(sch);
842 struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb);
843 struct tc_netem_qopt qopt;
844 struct tc_netem_corr cor;
845 struct tc_netem_reorder reorder;
846 struct tc_netem_corrupt corrupt;
848 qopt.latency = q->latency;
849 qopt.jitter = q->jitter;
850 qopt.limit = q->limit;
853 qopt.duplicate = q->duplicate;
854 NLA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
856 cor.delay_corr = q->delay_cor.rho;
857 cor.loss_corr = q->loss_cor.rho;
858 cor.dup_corr = q->dup_cor.rho;
859 NLA_PUT(skb, TCA_NETEM_CORR, sizeof(cor), &cor);
861 reorder.probability = q->reorder;
862 reorder.correlation = q->reorder_cor.rho;
863 NLA_PUT(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder);
865 corrupt.probability = q->corrupt;
866 corrupt.correlation = q->corrupt_cor.rho;
867 NLA_PUT(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt);
869 if (dump_loss_model(q, skb) != 0)
870 goto nla_put_failure;
872 return nla_nest_end(skb, nla);
875 nlmsg_trim(skb, nla);
879 static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
880 struct sk_buff *skb, struct tcmsg *tcm)
882 struct netem_sched_data *q = qdisc_priv(sch);
884 if (cl != 1) /* only one class */
887 tcm->tcm_handle |= TC_H_MIN(1);
888 tcm->tcm_info = q->qdisc->handle;
893 static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
896 struct netem_sched_data *q = qdisc_priv(sch);
904 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
906 sch_tree_unlock(sch);
911 static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
913 struct netem_sched_data *q = qdisc_priv(sch);
917 static unsigned long netem_get(struct Qdisc *sch, u32 classid)
922 static void netem_put(struct Qdisc *sch, unsigned long arg)
926 static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
929 if (walker->count >= walker->skip)
930 if (walker->fn(sch, 1, walker) < 0) {
938 static const struct Qdisc_class_ops netem_class_ops = {
939 .graft = netem_graft,
944 .dump = netem_dump_class,
947 static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
949 .cl_ops = &netem_class_ops,
950 .priv_size = sizeof(struct netem_sched_data),
951 .enqueue = netem_enqueue,
952 .dequeue = netem_dequeue,
953 .peek = qdisc_peek_dequeued,
956 .reset = netem_reset,
957 .destroy = netem_destroy,
958 .change = netem_change,
960 .owner = THIS_MODULE,
964 static int __init netem_module_init(void)
966 pr_info("netem: version " VERSION "\n");
967 return register_qdisc(&netem_qdisc_ops);
969 static void __exit netem_module_exit(void)
971 unregister_qdisc(&netem_qdisc_ops);
973 module_init(netem_module_init)
974 module_exit(netem_module_exit)
975 MODULE_LICENSE("GPL");