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>
17 #include <linux/module.h>
18 #include <linux/slab.h>
19 #include <linux/types.h>
20 #include <linux/kernel.h>
21 #include <linux/errno.h>
22 #include <linux/skbuff.h>
23 #include <linux/vmalloc.h>
24 #include <linux/rtnetlink.h>
26 #include <net/netlink.h>
27 #include <net/pkt_sched.h>
31 /* Network Emulation Queuing algorithm.
32 ====================================
34 Sources: [1] Mark Carson, Darrin Santay, "NIST Net - A Linux-based
35 Network Emulation Tool
36 [2] Luigi Rizzo, DummyNet for FreeBSD
38 ----------------------------------------------------------------
40 This started out as a simple way to delay outgoing packets to
41 test TCP but has grown to include most of the functionality
42 of a full blown network emulator like NISTnet. It can delay
43 packets and add random jitter (and correlation). The random
44 distribution can be loaded from a table as well to provide
45 normal, Pareto, or experimental curves. Packet loss,
46 duplication, and reordering can also be emulated.
48 This qdisc does not do classification that can be handled in
49 layering other disciplines. It does not need to do bandwidth
50 control either since that can be handled by using token
51 bucket or other rate control.
53 Correlated Loss Generator models
55 Added generation of correlated loss according to the
56 "Gilbert-Elliot" model, a 4-state markov model.
59 [1] NetemCLG Home http://netgroup.uniroma2.it/NetemCLG
60 [2] S. Salsano, F. Ludovici, A. Ordine, "Definition of a general
61 and intuitive loss model for packet networks and its implementation
62 in the Netem module in the Linux kernel", available in [1]
64 Authors: Stefano Salsano <stefano.salsano at uniroma2.it
65 Fabio Ludovici <fabio.ludovici at yahoo.it>
68 struct netem_sched_data {
70 struct qdisc_watchdog watchdog;
72 psched_tdiff_t latency;
73 psched_tdiff_t jitter;
86 } delay_cor, loss_cor, dup_cor, reorder_cor, corrupt_cor;
99 /* Correlated Loss Generation models */
101 /* state of the Markov chain */
104 /* 4-states and Gilbert-Elliot models */
105 u32 a1; /* p13 for 4-states or p for GE */
106 u32 a2; /* p31 for 4-states or r for GE */
107 u32 a3; /* p32 for 4-states or h for GE */
108 u32 a4; /* p14 for 4-states or 1-k for GE */
109 u32 a5; /* p23 used only in 4-states */
114 /* Time stamp put into socket buffer control block */
115 struct netem_skb_cb {
116 psched_time_t time_to_send;
119 static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb)
121 qdisc_cb_private_validate(skb, sizeof(struct netem_skb_cb));
122 return (struct netem_skb_cb *)qdisc_skb_cb(skb)->data;
125 /* init_crandom - initialize correlated random number generator
126 * Use entropy source for initial seed.
128 static void init_crandom(struct crndstate *state, unsigned long rho)
131 state->last = net_random();
134 /* get_crandom - correlated random number generator
135 * Next number depends on last value.
136 * rho is scaled to avoid floating point.
138 static u32 get_crandom(struct crndstate *state)
141 unsigned long answer;
143 if (state->rho == 0) /* no correlation */
146 value = net_random();
147 rho = (u64)state->rho + 1;
148 answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32;
149 state->last = answer;
153 /* loss_4state - 4-state model loss generator
154 * Generates losses according to the 4-state Markov chain adopted in
155 * the GI (General and Intuitive) loss model.
157 static bool loss_4state(struct netem_sched_data *q)
159 struct clgstate *clg = &q->clg;
160 u32 rnd = net_random();
163 * Makes a comparison between rnd and the transition
164 * probabilities outgoing from the current state, then decides the
165 * next state and if the next packet has to be transmitted or lost.
166 * The four states correspond to:
167 * 1 => successfully transmitted packets within a gap period
168 * 4 => isolated losses within a gap period
169 * 3 => lost packets within a burst period
170 * 2 => successfully transmitted packets within a burst period
172 switch (clg->state) {
177 } else if (clg->a4 < rnd && rnd < clg->a1) {
180 } else if (clg->a1 < rnd)
195 else if (clg->a3 < rnd && rnd < clg->a2 + clg->a3) {
198 } else if (clg->a2 + clg->a3 < rnd) {
211 /* loss_gilb_ell - Gilbert-Elliot model loss generator
212 * Generates losses according to the Gilbert-Elliot loss model or
213 * its special cases (Gilbert or Simple Gilbert)
215 * Makes a comparison between random number and the transition
216 * probabilities outgoing from the current state, then decides the
217 * next state. A second random number is extracted and the comparison
218 * with the loss probability of the current state decides if the next
219 * packet will be transmitted or lost.
221 static bool loss_gilb_ell(struct netem_sched_data *q)
223 struct clgstate *clg = &q->clg;
225 switch (clg->state) {
227 if (net_random() < clg->a1)
229 if (net_random() < clg->a4)
232 if (net_random() < clg->a2)
234 if (clg->a3 > net_random())
241 static bool loss_event(struct netem_sched_data *q)
243 switch (q->loss_model) {
245 /* Random packet drop 0 => none, ~0 => all */
246 return q->loss && q->loss >= get_crandom(&q->loss_cor);
249 /* 4state loss model algorithm (used also for GI model)
250 * Extracts a value from the markov 4 state loss generator,
251 * if it is 1 drops a packet and if needed writes the event in
254 return loss_4state(q);
257 /* Gilbert-Elliot loss model algorithm
258 * Extracts a value from the Gilbert-Elliot loss generator,
259 * if it is 1 drops a packet and if needed writes the event in
262 return loss_gilb_ell(q);
265 return false; /* not reached */
269 /* tabledist - return a pseudo-randomly distributed value with mean mu and
270 * std deviation sigma. Uses table lookup to approximate the desired
271 * distribution, and a uniformly-distributed pseudo-random source.
273 static psched_tdiff_t tabledist(psched_tdiff_t mu, psched_tdiff_t sigma,
274 struct crndstate *state,
275 const struct disttable *dist)
284 rnd = get_crandom(state);
286 /* default uniform distribution */
288 return (rnd % (2*sigma)) - sigma + mu;
290 t = dist->table[rnd % dist->size];
291 x = (sigma % NETEM_DIST_SCALE) * t;
293 x += NETEM_DIST_SCALE/2;
295 x -= NETEM_DIST_SCALE/2;
297 return x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu;
301 * Insert one skb into qdisc.
302 * Note: parent depends on return value to account for queue length.
303 * NET_XMIT_DROP: queue length didn't change.
304 * NET_XMIT_SUCCESS: one skb was queued.
306 static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch)
308 struct netem_sched_data *q = qdisc_priv(sch);
309 /* We don't fill cb now as skb_unshare() may invalidate it */
310 struct netem_skb_cb *cb;
311 struct sk_buff *skb2;
315 /* Random duplication */
316 if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor))
326 return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
332 * If we need to duplicate packet, then re-insert at top of the
333 * qdisc tree, since parent queuer expects that only one
334 * skb will be queued.
336 if (count > 1 && (skb2 = skb_clone(skb, GFP_ATOMIC)) != NULL) {
337 struct Qdisc *rootq = qdisc_root(sch);
338 u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
341 qdisc_enqueue_root(skb2, rootq);
342 q->duplicate = dupsave;
346 * Randomized packet corruption.
347 * Make copy if needed since we are modifying
348 * If packet is going to be hardware checksummed, then
349 * do it now in software before we mangle it.
351 if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor)) {
352 if (!(skb = skb_unshare(skb, GFP_ATOMIC)) ||
353 (skb->ip_summed == CHECKSUM_PARTIAL &&
354 skb_checksum_help(skb))) {
356 return NET_XMIT_DROP;
359 skb->data[net_random() % skb_headlen(skb)] ^= 1<<(net_random() % 8);
362 cb = netem_skb_cb(skb);
363 if (q->gap == 0 || /* not doing reordering */
364 q->counter < q->gap || /* inside last reordering gap */
365 q->reorder < get_crandom(&q->reorder_cor)) {
367 psched_tdiff_t delay;
369 delay = tabledist(q->latency, q->jitter,
370 &q->delay_cor, q->delay_dist);
372 now = psched_get_time();
373 cb->time_to_send = now + delay;
375 ret = qdisc_enqueue(skb, q->qdisc);
378 * Do re-ordering by putting one out of N packets at the front
381 cb->time_to_send = psched_get_time();
384 __skb_queue_head(&q->qdisc->q, skb);
385 sch->qstats.backlog += qdisc_pkt_len(skb);
386 sch->qstats.requeues++;
387 ret = NET_XMIT_SUCCESS;
390 if (ret != NET_XMIT_SUCCESS) {
391 if (net_xmit_drop_count(ret)) {
398 return NET_XMIT_SUCCESS;
401 static unsigned int netem_drop(struct Qdisc *sch)
403 struct netem_sched_data *q = qdisc_priv(sch);
404 unsigned int len = 0;
406 if (q->qdisc->ops->drop && (len = q->qdisc->ops->drop(q->qdisc)) != 0) {
413 static struct sk_buff *netem_dequeue(struct Qdisc *sch)
415 struct netem_sched_data *q = qdisc_priv(sch);
418 if (qdisc_is_throttled(sch))
421 skb = q->qdisc->ops->peek(q->qdisc);
423 const struct netem_skb_cb *cb = netem_skb_cb(skb);
424 psched_time_t now = psched_get_time();
426 /* if more time remaining? */
427 if (cb->time_to_send <= now) {
428 skb = qdisc_dequeue_peeked(q->qdisc);
432 #ifdef CONFIG_NET_CLS_ACT
434 * If it's at ingress let's pretend the delay is
435 * from the network (tstamp will be updated).
437 if (G_TC_FROM(skb->tc_verd) & AT_INGRESS)
438 skb->tstamp.tv64 = 0;
442 qdisc_unthrottled(sch);
443 qdisc_bstats_update(sch, skb);
447 qdisc_watchdog_schedule(&q->watchdog, cb->time_to_send);
453 static void netem_reset(struct Qdisc *sch)
455 struct netem_sched_data *q = qdisc_priv(sch);
457 qdisc_reset(q->qdisc);
459 qdisc_watchdog_cancel(&q->watchdog);
462 static void dist_free(struct disttable *d)
465 if (is_vmalloc_addr(d))
473 * Distribution data is a variable size payload containing
474 * signed 16 bit values.
476 static int get_dist_table(struct Qdisc *sch, const struct nlattr *attr)
478 struct netem_sched_data *q = qdisc_priv(sch);
479 size_t n = nla_len(attr)/sizeof(__s16);
480 const __s16 *data = nla_data(attr);
481 spinlock_t *root_lock;
486 if (n > NETEM_DIST_MAX)
489 s = sizeof(struct disttable) + n * sizeof(s16);
490 d = kmalloc(s, GFP_KERNEL | __GFP_NOWARN);
497 for (i = 0; i < n; i++)
498 d->table[i] = data[i];
500 root_lock = qdisc_root_sleeping_lock(sch);
502 spin_lock_bh(root_lock);
503 swap(q->delay_dist, d);
504 spin_unlock_bh(root_lock);
510 static void get_correlation(struct Qdisc *sch, const struct nlattr *attr)
512 struct netem_sched_data *q = qdisc_priv(sch);
513 const struct tc_netem_corr *c = nla_data(attr);
515 init_crandom(&q->delay_cor, c->delay_corr);
516 init_crandom(&q->loss_cor, c->loss_corr);
517 init_crandom(&q->dup_cor, c->dup_corr);
520 static void get_reorder(struct Qdisc *sch, const struct nlattr *attr)
522 struct netem_sched_data *q = qdisc_priv(sch);
523 const struct tc_netem_reorder *r = nla_data(attr);
525 q->reorder = r->probability;
526 init_crandom(&q->reorder_cor, r->correlation);
529 static void get_corrupt(struct Qdisc *sch, const struct nlattr *attr)
531 struct netem_sched_data *q = qdisc_priv(sch);
532 const struct tc_netem_corrupt *r = nla_data(attr);
534 q->corrupt = r->probability;
535 init_crandom(&q->corrupt_cor, r->correlation);
538 static int get_loss_clg(struct Qdisc *sch, const struct nlattr *attr)
540 struct netem_sched_data *q = qdisc_priv(sch);
541 const struct nlattr *la;
544 nla_for_each_nested(la, attr, rem) {
545 u16 type = nla_type(la);
548 case NETEM_LOSS_GI: {
549 const struct tc_netem_gimodel *gi = nla_data(la);
551 if (nla_len(la) != sizeof(struct tc_netem_gimodel)) {
552 pr_info("netem: incorrect gi model size\n");
556 q->loss_model = CLG_4_STATES;
567 case NETEM_LOSS_GE: {
568 const struct tc_netem_gemodel *ge = nla_data(la);
570 if (nla_len(la) != sizeof(struct tc_netem_gemodel)) {
571 pr_info("netem: incorrect gi model size\n");
575 q->loss_model = CLG_GILB_ELL;
585 pr_info("netem: unknown loss type %u\n", type);
593 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
594 [TCA_NETEM_CORR] = { .len = sizeof(struct tc_netem_corr) },
595 [TCA_NETEM_REORDER] = { .len = sizeof(struct tc_netem_reorder) },
596 [TCA_NETEM_CORRUPT] = { .len = sizeof(struct tc_netem_corrupt) },
597 [TCA_NETEM_LOSS] = { .type = NLA_NESTED },
600 static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
601 const struct nla_policy *policy, int len)
603 int nested_len = nla_len(nla) - NLA_ALIGN(len);
605 if (nested_len < 0) {
606 pr_info("netem: invalid attributes len %d\n", nested_len);
610 if (nested_len >= nla_attr_size(0))
611 return nla_parse(tb, maxtype, nla_data(nla) + NLA_ALIGN(len),
614 memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
618 /* Parse netlink message to set options */
619 static int netem_change(struct Qdisc *sch, struct nlattr *opt)
621 struct netem_sched_data *q = qdisc_priv(sch);
622 struct nlattr *tb[TCA_NETEM_MAX + 1];
623 struct tc_netem_qopt *qopt;
629 qopt = nla_data(opt);
630 ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
634 ret = fifo_set_limit(q->qdisc, qopt->limit);
636 pr_info("netem: can't set fifo limit\n");
640 q->latency = qopt->latency;
641 q->jitter = qopt->jitter;
642 q->limit = qopt->limit;
645 q->loss = qopt->loss;
646 q->duplicate = qopt->duplicate;
648 /* for compatibility with earlier versions.
649 * if gap is set, need to assume 100% probability
654 if (tb[TCA_NETEM_CORR])
655 get_correlation(sch, tb[TCA_NETEM_CORR]);
657 if (tb[TCA_NETEM_DELAY_DIST]) {
658 ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST]);
663 if (tb[TCA_NETEM_REORDER])
664 get_reorder(sch, tb[TCA_NETEM_REORDER]);
666 if (tb[TCA_NETEM_CORRUPT])
667 get_corrupt(sch, tb[TCA_NETEM_CORRUPT]);
669 q->loss_model = CLG_RANDOM;
670 if (tb[TCA_NETEM_LOSS])
671 ret = get_loss_clg(sch, tb[TCA_NETEM_LOSS]);
677 * Special case version of FIFO queue for use by netem.
678 * It queues in order based on timestamps in skb's
680 struct fifo_sched_data {
682 psched_time_t oldest;
685 static int tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
687 struct fifo_sched_data *q = qdisc_priv(sch);
688 struct sk_buff_head *list = &sch->q;
689 psched_time_t tnext = netem_skb_cb(nskb)->time_to_send;
692 if (likely(skb_queue_len(list) < q->limit)) {
693 /* Optimize for add at tail */
694 if (likely(skb_queue_empty(list) || tnext >= q->oldest)) {
696 return qdisc_enqueue_tail(nskb, sch);
699 skb_queue_reverse_walk(list, skb) {
700 const struct netem_skb_cb *cb = netem_skb_cb(skb);
702 if (tnext >= cb->time_to_send)
706 __skb_queue_after(list, skb, nskb);
708 sch->qstats.backlog += qdisc_pkt_len(nskb);
710 return NET_XMIT_SUCCESS;
713 return qdisc_reshape_fail(nskb, sch);
716 static int tfifo_init(struct Qdisc *sch, struct nlattr *opt)
718 struct fifo_sched_data *q = qdisc_priv(sch);
721 struct tc_fifo_qopt *ctl = nla_data(opt);
722 if (nla_len(opt) < sizeof(*ctl))
725 q->limit = ctl->limit;
727 q->limit = max_t(u32, qdisc_dev(sch)->tx_queue_len, 1);
729 q->oldest = PSCHED_PASTPERFECT;
733 static int tfifo_dump(struct Qdisc *sch, struct sk_buff *skb)
735 struct fifo_sched_data *q = qdisc_priv(sch);
736 struct tc_fifo_qopt opt = { .limit = q->limit };
738 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
745 static struct Qdisc_ops tfifo_qdisc_ops __read_mostly = {
747 .priv_size = sizeof(struct fifo_sched_data),
748 .enqueue = tfifo_enqueue,
749 .dequeue = qdisc_dequeue_head,
750 .peek = qdisc_peek_head,
751 .drop = qdisc_queue_drop,
753 .reset = qdisc_reset_queue,
754 .change = tfifo_init,
758 static int netem_init(struct Qdisc *sch, struct nlattr *opt)
760 struct netem_sched_data *q = qdisc_priv(sch);
766 qdisc_watchdog_init(&q->watchdog, sch);
768 q->loss_model = CLG_RANDOM;
769 q->qdisc = qdisc_create_dflt(sch->dev_queue, &tfifo_qdisc_ops,
770 TC_H_MAKE(sch->handle, 1));
772 pr_notice("netem: qdisc create tfifo qdisc failed\n");
776 ret = netem_change(sch, opt);
778 pr_info("netem: change failed\n");
779 qdisc_destroy(q->qdisc);
784 static void netem_destroy(struct Qdisc *sch)
786 struct netem_sched_data *q = qdisc_priv(sch);
788 qdisc_watchdog_cancel(&q->watchdog);
789 qdisc_destroy(q->qdisc);
790 dist_free(q->delay_dist);
793 static int dump_loss_model(const struct netem_sched_data *q,
798 nest = nla_nest_start(skb, TCA_NETEM_LOSS);
800 goto nla_put_failure;
802 switch (q->loss_model) {
804 /* legacy loss model */
805 nla_nest_cancel(skb, nest);
806 return 0; /* no data */
809 struct tc_netem_gimodel gi = {
817 NLA_PUT(skb, NETEM_LOSS_GI, sizeof(gi), &gi);
821 struct tc_netem_gemodel ge = {
828 NLA_PUT(skb, NETEM_LOSS_GE, sizeof(ge), &ge);
833 nla_nest_end(skb, nest);
837 nla_nest_cancel(skb, nest);
841 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
843 const struct netem_sched_data *q = qdisc_priv(sch);
844 struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb);
845 struct tc_netem_qopt qopt;
846 struct tc_netem_corr cor;
847 struct tc_netem_reorder reorder;
848 struct tc_netem_corrupt corrupt;
850 qopt.latency = q->latency;
851 qopt.jitter = q->jitter;
852 qopt.limit = q->limit;
855 qopt.duplicate = q->duplicate;
856 NLA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
858 cor.delay_corr = q->delay_cor.rho;
859 cor.loss_corr = q->loss_cor.rho;
860 cor.dup_corr = q->dup_cor.rho;
861 NLA_PUT(skb, TCA_NETEM_CORR, sizeof(cor), &cor);
863 reorder.probability = q->reorder;
864 reorder.correlation = q->reorder_cor.rho;
865 NLA_PUT(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder);
867 corrupt.probability = q->corrupt;
868 corrupt.correlation = q->corrupt_cor.rho;
869 NLA_PUT(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt);
871 if (dump_loss_model(q, skb) != 0)
872 goto nla_put_failure;
874 return nla_nest_end(skb, nla);
877 nlmsg_trim(skb, nla);
881 static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
882 struct sk_buff *skb, struct tcmsg *tcm)
884 struct netem_sched_data *q = qdisc_priv(sch);
886 if (cl != 1) /* only one class */
889 tcm->tcm_handle |= TC_H_MIN(1);
890 tcm->tcm_info = q->qdisc->handle;
895 static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
898 struct netem_sched_data *q = qdisc_priv(sch);
906 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
908 sch_tree_unlock(sch);
913 static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
915 struct netem_sched_data *q = qdisc_priv(sch);
919 static unsigned long netem_get(struct Qdisc *sch, u32 classid)
924 static void netem_put(struct Qdisc *sch, unsigned long arg)
928 static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
931 if (walker->count >= walker->skip)
932 if (walker->fn(sch, 1, walker) < 0) {
940 static const struct Qdisc_class_ops netem_class_ops = {
941 .graft = netem_graft,
946 .dump = netem_dump_class,
949 static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
951 .cl_ops = &netem_class_ops,
952 .priv_size = sizeof(struct netem_sched_data),
953 .enqueue = netem_enqueue,
954 .dequeue = netem_dequeue,
955 .peek = qdisc_peek_dequeued,
958 .reset = netem_reset,
959 .destroy = netem_destroy,
960 .change = netem_change,
962 .owner = THIS_MODULE,
966 static int __init netem_module_init(void)
968 pr_info("netem: version " VERSION "\n");
969 return register_qdisc(&netem_qdisc_ops);
971 static void __exit netem_module_exit(void)
973 unregister_qdisc(&netem_qdisc_ops);
975 module_init(netem_module_init)
976 module_exit(netem_module_exit)
977 MODULE_LICENSE("GPL");