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_SUCCESS | __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();
237 __skb_queue_head(&q->qdisc->q, skb);
238 q->qdisc->qstats.backlog += qdisc_pkt_len(skb);
239 q->qdisc->qstats.requeues++;
240 ret = NET_XMIT_SUCCESS;
243 if (likely(ret == NET_XMIT_SUCCESS)) {
245 sch->bstats.bytes += qdisc_pkt_len(skb);
246 sch->bstats.packets++;
247 } else if (net_xmit_drop_count(ret)) {
251 pr_debug("netem: enqueue ret %d\n", ret);
255 /* Requeue packets but don't change time stamp */
256 static int netem_requeue(struct sk_buff *skb, struct Qdisc *sch)
258 struct netem_sched_data *q = qdisc_priv(sch);
261 if ((ret = q->qdisc->ops->requeue(skb, q->qdisc)) == 0) {
263 sch->qstats.requeues++;
269 static unsigned int netem_drop(struct Qdisc* sch)
271 struct netem_sched_data *q = qdisc_priv(sch);
272 unsigned int len = 0;
274 if (q->qdisc->ops->drop && (len = q->qdisc->ops->drop(q->qdisc)) != 0) {
281 static struct sk_buff *netem_dequeue(struct Qdisc *sch)
283 struct netem_sched_data *q = qdisc_priv(sch);
287 if (sch->flags & TCQ_F_THROTTLED)
290 skb = q->qdisc->ops->peek(q->qdisc);
292 const struct netem_skb_cb *cb = netem_skb_cb(skb);
293 psched_time_t now = psched_get_time();
295 /* if more time remaining? */
296 if (cb->time_to_send <= now) {
297 skb = qdisc_dequeue_peeked(q->qdisc);
301 pr_debug("netem_dequeue: return skb=%p\n", skb);
306 qdisc_watchdog_schedule(&q->watchdog, cb->time_to_send);
312 static void netem_reset(struct Qdisc *sch)
314 struct netem_sched_data *q = qdisc_priv(sch);
316 qdisc_reset(q->qdisc);
318 qdisc_watchdog_cancel(&q->watchdog);
322 * Distribution data is a variable size payload containing
323 * signed 16 bit values.
325 static int get_dist_table(struct Qdisc *sch, const struct nlattr *attr)
327 struct netem_sched_data *q = qdisc_priv(sch);
328 unsigned long n = nla_len(attr)/sizeof(__s16);
329 const __s16 *data = nla_data(attr);
330 spinlock_t *root_lock;
337 d = kmalloc(sizeof(*d) + n*sizeof(d->table[0]), GFP_KERNEL);
342 for (i = 0; i < n; i++)
343 d->table[i] = data[i];
345 root_lock = qdisc_root_sleeping_lock(sch);
347 spin_lock_bh(root_lock);
348 d = xchg(&q->delay_dist, d);
349 spin_unlock_bh(root_lock);
355 static int get_correlation(struct Qdisc *sch, const struct nlattr *attr)
357 struct netem_sched_data *q = qdisc_priv(sch);
358 const struct tc_netem_corr *c = nla_data(attr);
360 init_crandom(&q->delay_cor, c->delay_corr);
361 init_crandom(&q->loss_cor, c->loss_corr);
362 init_crandom(&q->dup_cor, c->dup_corr);
366 static int get_reorder(struct Qdisc *sch, const struct nlattr *attr)
368 struct netem_sched_data *q = qdisc_priv(sch);
369 const struct tc_netem_reorder *r = nla_data(attr);
371 q->reorder = r->probability;
372 init_crandom(&q->reorder_cor, r->correlation);
376 static int get_corrupt(struct Qdisc *sch, const struct nlattr *attr)
378 struct netem_sched_data *q = qdisc_priv(sch);
379 const struct tc_netem_corrupt *r = nla_data(attr);
381 q->corrupt = r->probability;
382 init_crandom(&q->corrupt_cor, r->correlation);
386 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
387 [TCA_NETEM_CORR] = { .len = sizeof(struct tc_netem_corr) },
388 [TCA_NETEM_REORDER] = { .len = sizeof(struct tc_netem_reorder) },
389 [TCA_NETEM_CORRUPT] = { .len = sizeof(struct tc_netem_corrupt) },
392 static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
393 const struct nla_policy *policy, int len)
395 int nested_len = nla_len(nla) - NLA_ALIGN(len);
399 if (nested_len >= nla_attr_size(0))
400 return nla_parse(tb, maxtype, nla_data(nla) + NLA_ALIGN(len),
402 memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
406 /* Parse netlink message to set options */
407 static int netem_change(struct Qdisc *sch, struct nlattr *opt)
409 struct netem_sched_data *q = qdisc_priv(sch);
410 struct nlattr *tb[TCA_NETEM_MAX + 1];
411 struct tc_netem_qopt *qopt;
417 qopt = nla_data(opt);
418 ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
422 ret = fifo_set_limit(q->qdisc, qopt->limit);
424 pr_debug("netem: can't set fifo limit\n");
428 q->latency = qopt->latency;
429 q->jitter = qopt->jitter;
430 q->limit = qopt->limit;
433 q->loss = qopt->loss;
434 q->duplicate = qopt->duplicate;
436 /* for compatibility with earlier versions.
437 * if gap is set, need to assume 100% probability
442 if (tb[TCA_NETEM_CORR]) {
443 ret = get_correlation(sch, tb[TCA_NETEM_CORR]);
448 if (tb[TCA_NETEM_DELAY_DIST]) {
449 ret = get_dist_table(sch, tb[TCA_NETEM_DELAY_DIST]);
454 if (tb[TCA_NETEM_REORDER]) {
455 ret = get_reorder(sch, tb[TCA_NETEM_REORDER]);
460 if (tb[TCA_NETEM_CORRUPT]) {
461 ret = get_corrupt(sch, tb[TCA_NETEM_CORRUPT]);
470 * Special case version of FIFO queue for use by netem.
471 * It queues in order based on timestamps in skb's
473 struct fifo_sched_data {
475 psched_time_t oldest;
478 static int tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
480 struct fifo_sched_data *q = qdisc_priv(sch);
481 struct sk_buff_head *list = &sch->q;
482 psched_time_t tnext = netem_skb_cb(nskb)->time_to_send;
485 if (likely(skb_queue_len(list) < q->limit)) {
486 /* Optimize for add at tail */
487 if (likely(skb_queue_empty(list) || tnext >= q->oldest)) {
489 return qdisc_enqueue_tail(nskb, sch);
492 skb_queue_reverse_walk(list, skb) {
493 const struct netem_skb_cb *cb = netem_skb_cb(skb);
495 if (tnext >= cb->time_to_send)
499 __skb_queue_after(list, skb, nskb);
501 sch->qstats.backlog += qdisc_pkt_len(nskb);
502 sch->bstats.bytes += qdisc_pkt_len(nskb);
503 sch->bstats.packets++;
505 return NET_XMIT_SUCCESS;
508 return qdisc_reshape_fail(nskb, sch);
511 static int tfifo_init(struct Qdisc *sch, struct nlattr *opt)
513 struct fifo_sched_data *q = qdisc_priv(sch);
516 struct tc_fifo_qopt *ctl = nla_data(opt);
517 if (nla_len(opt) < sizeof(*ctl))
520 q->limit = ctl->limit;
522 q->limit = max_t(u32, qdisc_dev(sch)->tx_queue_len, 1);
524 q->oldest = PSCHED_PASTPERFECT;
528 static int tfifo_dump(struct Qdisc *sch, struct sk_buff *skb)
530 struct fifo_sched_data *q = qdisc_priv(sch);
531 struct tc_fifo_qopt opt = { .limit = q->limit };
533 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
540 static struct Qdisc_ops tfifo_qdisc_ops __read_mostly = {
542 .priv_size = sizeof(struct fifo_sched_data),
543 .enqueue = tfifo_enqueue,
544 .dequeue = qdisc_dequeue_head,
545 .peek = qdisc_peek_head,
546 .requeue = qdisc_requeue,
547 .drop = qdisc_queue_drop,
549 .reset = qdisc_reset_queue,
550 .change = tfifo_init,
554 static int netem_init(struct Qdisc *sch, struct nlattr *opt)
556 struct netem_sched_data *q = qdisc_priv(sch);
562 qdisc_watchdog_init(&q->watchdog, sch);
564 q->qdisc = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
566 TC_H_MAKE(sch->handle, 1));
568 pr_debug("netem: qdisc create failed\n");
572 ret = netem_change(sch, opt);
574 pr_debug("netem: change failed\n");
575 qdisc_destroy(q->qdisc);
580 static void netem_destroy(struct Qdisc *sch)
582 struct netem_sched_data *q = qdisc_priv(sch);
584 qdisc_watchdog_cancel(&q->watchdog);
585 qdisc_destroy(q->qdisc);
586 kfree(q->delay_dist);
589 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
591 const struct netem_sched_data *q = qdisc_priv(sch);
592 unsigned char *b = skb_tail_pointer(skb);
593 struct nlattr *nla = (struct nlattr *) b;
594 struct tc_netem_qopt qopt;
595 struct tc_netem_corr cor;
596 struct tc_netem_reorder reorder;
597 struct tc_netem_corrupt corrupt;
599 qopt.latency = q->latency;
600 qopt.jitter = q->jitter;
601 qopt.limit = q->limit;
604 qopt.duplicate = q->duplicate;
605 NLA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt);
607 cor.delay_corr = q->delay_cor.rho;
608 cor.loss_corr = q->loss_cor.rho;
609 cor.dup_corr = q->dup_cor.rho;
610 NLA_PUT(skb, TCA_NETEM_CORR, sizeof(cor), &cor);
612 reorder.probability = q->reorder;
613 reorder.correlation = q->reorder_cor.rho;
614 NLA_PUT(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder);
616 corrupt.probability = q->corrupt;
617 corrupt.correlation = q->corrupt_cor.rho;
618 NLA_PUT(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt);
620 nla->nla_len = skb_tail_pointer(skb) - b;
629 static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
631 .priv_size = sizeof(struct netem_sched_data),
632 .enqueue = netem_enqueue,
633 .dequeue = netem_dequeue,
634 .peek = qdisc_peek_dequeued,
635 .requeue = netem_requeue,
638 .reset = netem_reset,
639 .destroy = netem_destroy,
640 .change = netem_change,
642 .owner = THIS_MODULE,
646 static int __init netem_module_init(void)
648 pr_info("netem: version " VERSION "\n");
649 return register_qdisc(&netem_qdisc_ops);
651 static void __exit netem_module_exit(void)
653 unregister_qdisc(&netem_qdisc_ops);
655 module_init(netem_module_init)
656 module_exit(netem_module_exit)
657 MODULE_LICENSE("GPL");