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();
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
++;
243 } else if (net_xmit_drop_count(ret
)) {
247 pr_debug("netem: enqueue ret %d\n", ret
);
251 /* Requeue packets but don't change time stamp */
252 static int netem_requeue(struct sk_buff
*skb
, struct Qdisc
*sch
)
254 struct netem_sched_data
*q
= qdisc_priv(sch
);
257 if ((ret
= q
->qdisc
->ops
->requeue(skb
, q
->qdisc
)) == 0) {
259 sch
->qstats
.requeues
++;
265 static unsigned int netem_drop(struct Qdisc
* sch
)
267 struct netem_sched_data
*q
= qdisc_priv(sch
);
268 unsigned int len
= 0;
270 if (q
->qdisc
->ops
->drop
&& (len
= q
->qdisc
->ops
->drop(q
->qdisc
)) != 0) {
277 static struct sk_buff
*netem_dequeue(struct Qdisc
*sch
)
279 struct netem_sched_data
*q
= qdisc_priv(sch
);
283 if (sch
->flags
& TCQ_F_THROTTLED
)
286 skb
= q
->qdisc
->dequeue(q
->qdisc
);
288 const struct netem_skb_cb
*cb
= netem_skb_cb(skb
);
289 psched_time_t now
= psched_get_time();
291 /* if more time remaining? */
292 if (cb
->time_to_send
<= now
) {
293 pr_debug("netem_dequeue: return skb=%p\n", skb
);
298 if (unlikely(q
->qdisc
->ops
->requeue(skb
, q
->qdisc
) != NET_XMIT_SUCCESS
)) {
299 qdisc_tree_decrease_qlen(q
->qdisc
, 1);
301 printk(KERN_ERR
"netem: %s could not requeue\n",
305 qdisc_watchdog_schedule(&q
->watchdog
, cb
->time_to_send
);
311 static void netem_reset(struct Qdisc
*sch
)
313 struct netem_sched_data
*q
= qdisc_priv(sch
);
315 qdisc_reset(q
->qdisc
);
317 qdisc_watchdog_cancel(&q
->watchdog
);
321 * Distribution data is a variable size payload containing
322 * signed 16 bit values.
324 static int get_dist_table(struct Qdisc
*sch
, const struct nlattr
*attr
)
326 struct netem_sched_data
*q
= qdisc_priv(sch
);
327 unsigned long n
= nla_len(attr
)/sizeof(__s16
);
328 const __s16
*data
= nla_data(attr
);
329 spinlock_t
*root_lock
;
336 d
= kmalloc(sizeof(*d
) + n
*sizeof(d
->table
[0]), GFP_KERNEL
);
341 for (i
= 0; i
< n
; i
++)
342 d
->table
[i
] = data
[i
];
344 root_lock
= qdisc_root_sleeping_lock(sch
);
346 spin_lock_bh(root_lock
);
347 d
= xchg(&q
->delay_dist
, d
);
348 spin_unlock_bh(root_lock
);
354 static int get_correlation(struct Qdisc
*sch
, const struct nlattr
*attr
)
356 struct netem_sched_data
*q
= qdisc_priv(sch
);
357 const struct tc_netem_corr
*c
= nla_data(attr
);
359 init_crandom(&q
->delay_cor
, c
->delay_corr
);
360 init_crandom(&q
->loss_cor
, c
->loss_corr
);
361 init_crandom(&q
->dup_cor
, c
->dup_corr
);
365 static int get_reorder(struct Qdisc
*sch
, const struct nlattr
*attr
)
367 struct netem_sched_data
*q
= qdisc_priv(sch
);
368 const struct tc_netem_reorder
*r
= nla_data(attr
);
370 q
->reorder
= r
->probability
;
371 init_crandom(&q
->reorder_cor
, r
->correlation
);
375 static int get_corrupt(struct Qdisc
*sch
, const struct nlattr
*attr
)
377 struct netem_sched_data
*q
= qdisc_priv(sch
);
378 const struct tc_netem_corrupt
*r
= nla_data(attr
);
380 q
->corrupt
= r
->probability
;
381 init_crandom(&q
->corrupt_cor
, r
->correlation
);
385 static const struct nla_policy netem_policy
[TCA_NETEM_MAX
+ 1] = {
386 [TCA_NETEM_CORR
] = { .len
= sizeof(struct tc_netem_corr
) },
387 [TCA_NETEM_REORDER
] = { .len
= sizeof(struct tc_netem_reorder
) },
388 [TCA_NETEM_CORRUPT
] = { .len
= sizeof(struct tc_netem_corrupt
) },
391 /* Parse netlink message to set options */
392 static int netem_change(struct Qdisc
*sch
, struct nlattr
*opt
)
394 struct netem_sched_data
*q
= qdisc_priv(sch
);
395 struct nlattr
*tb
[TCA_NETEM_MAX
+ 1];
396 struct tc_netem_qopt
*qopt
;
402 ret
= nla_parse_nested_compat(tb
, TCA_NETEM_MAX
, opt
, netem_policy
,
403 qopt
, sizeof(*qopt
));
407 ret
= fifo_set_limit(q
->qdisc
, qopt
->limit
);
409 pr_debug("netem: can't set fifo limit\n");
413 q
->latency
= qopt
->latency
;
414 q
->jitter
= qopt
->jitter
;
415 q
->limit
= qopt
->limit
;
418 q
->loss
= qopt
->loss
;
419 q
->duplicate
= qopt
->duplicate
;
421 /* for compatibility with earlier versions.
422 * if gap is set, need to assume 100% probability
427 if (tb
[TCA_NETEM_CORR
]) {
428 ret
= get_correlation(sch
, tb
[TCA_NETEM_CORR
]);
433 if (tb
[TCA_NETEM_DELAY_DIST
]) {
434 ret
= get_dist_table(sch
, tb
[TCA_NETEM_DELAY_DIST
]);
439 if (tb
[TCA_NETEM_REORDER
]) {
440 ret
= get_reorder(sch
, tb
[TCA_NETEM_REORDER
]);
445 if (tb
[TCA_NETEM_CORRUPT
]) {
446 ret
= get_corrupt(sch
, tb
[TCA_NETEM_CORRUPT
]);
455 * Special case version of FIFO queue for use by netem.
456 * It queues in order based on timestamps in skb's
458 struct fifo_sched_data
{
460 psched_time_t oldest
;
463 static int tfifo_enqueue(struct sk_buff
*nskb
, struct Qdisc
*sch
)
465 struct fifo_sched_data
*q
= qdisc_priv(sch
);
466 struct sk_buff_head
*list
= &sch
->q
;
467 psched_time_t tnext
= netem_skb_cb(nskb
)->time_to_send
;
470 if (likely(skb_queue_len(list
) < q
->limit
)) {
471 /* Optimize for add at tail */
472 if (likely(skb_queue_empty(list
) || tnext
>= q
->oldest
)) {
474 return qdisc_enqueue_tail(nskb
, sch
);
477 skb_queue_reverse_walk(list
, skb
) {
478 const struct netem_skb_cb
*cb
= netem_skb_cb(skb
);
480 if (tnext
>= cb
->time_to_send
)
484 __skb_queue_after(list
, skb
, nskb
);
486 sch
->qstats
.backlog
+= qdisc_pkt_len(nskb
);
487 sch
->bstats
.bytes
+= qdisc_pkt_len(nskb
);
488 sch
->bstats
.packets
++;
490 return NET_XMIT_SUCCESS
;
493 return qdisc_reshape_fail(nskb
, sch
);
496 static int tfifo_init(struct Qdisc
*sch
, struct nlattr
*opt
)
498 struct fifo_sched_data
*q
= qdisc_priv(sch
);
501 struct tc_fifo_qopt
*ctl
= nla_data(opt
);
502 if (nla_len(opt
) < sizeof(*ctl
))
505 q
->limit
= ctl
->limit
;
507 q
->limit
= max_t(u32
, qdisc_dev(sch
)->tx_queue_len
, 1);
509 q
->oldest
= PSCHED_PASTPERFECT
;
513 static int tfifo_dump(struct Qdisc
*sch
, struct sk_buff
*skb
)
515 struct fifo_sched_data
*q
= qdisc_priv(sch
);
516 struct tc_fifo_qopt opt
= { .limit
= q
->limit
};
518 NLA_PUT(skb
, TCA_OPTIONS
, sizeof(opt
), &opt
);
525 static struct Qdisc_ops tfifo_qdisc_ops __read_mostly
= {
527 .priv_size
= sizeof(struct fifo_sched_data
),
528 .enqueue
= tfifo_enqueue
,
529 .dequeue
= qdisc_dequeue_head
,
530 .requeue
= qdisc_requeue
,
531 .drop
= qdisc_queue_drop
,
533 .reset
= qdisc_reset_queue
,
534 .change
= tfifo_init
,
538 static int netem_init(struct Qdisc
*sch
, struct nlattr
*opt
)
540 struct netem_sched_data
*q
= qdisc_priv(sch
);
546 qdisc_watchdog_init(&q
->watchdog
, sch
);
548 q
->qdisc
= qdisc_create_dflt(qdisc_dev(sch
), sch
->dev_queue
,
550 TC_H_MAKE(sch
->handle
, 1));
552 pr_debug("netem: qdisc create failed\n");
556 ret
= netem_change(sch
, opt
);
558 pr_debug("netem: change failed\n");
559 qdisc_destroy(q
->qdisc
);
564 static void netem_destroy(struct Qdisc
*sch
)
566 struct netem_sched_data
*q
= qdisc_priv(sch
);
568 qdisc_watchdog_cancel(&q
->watchdog
);
569 qdisc_destroy(q
->qdisc
);
570 kfree(q
->delay_dist
);
573 static int netem_dump(struct Qdisc
*sch
, struct sk_buff
*skb
)
575 const struct netem_sched_data
*q
= qdisc_priv(sch
);
576 unsigned char *b
= skb_tail_pointer(skb
);
577 struct nlattr
*nla
= (struct nlattr
*) b
;
578 struct tc_netem_qopt qopt
;
579 struct tc_netem_corr cor
;
580 struct tc_netem_reorder reorder
;
581 struct tc_netem_corrupt corrupt
;
583 qopt
.latency
= q
->latency
;
584 qopt
.jitter
= q
->jitter
;
585 qopt
.limit
= q
->limit
;
588 qopt
.duplicate
= q
->duplicate
;
589 NLA_PUT(skb
, TCA_OPTIONS
, sizeof(qopt
), &qopt
);
591 cor
.delay_corr
= q
->delay_cor
.rho
;
592 cor
.loss_corr
= q
->loss_cor
.rho
;
593 cor
.dup_corr
= q
->dup_cor
.rho
;
594 NLA_PUT(skb
, TCA_NETEM_CORR
, sizeof(cor
), &cor
);
596 reorder
.probability
= q
->reorder
;
597 reorder
.correlation
= q
->reorder_cor
.rho
;
598 NLA_PUT(skb
, TCA_NETEM_REORDER
, sizeof(reorder
), &reorder
);
600 corrupt
.probability
= q
->corrupt
;
601 corrupt
.correlation
= q
->corrupt_cor
.rho
;
602 NLA_PUT(skb
, TCA_NETEM_CORRUPT
, sizeof(corrupt
), &corrupt
);
604 nla
->nla_len
= skb_tail_pointer(skb
) - b
;
613 static int netem_dump_class(struct Qdisc
*sch
, unsigned long cl
,
614 struct sk_buff
*skb
, struct tcmsg
*tcm
)
616 struct netem_sched_data
*q
= qdisc_priv(sch
);
618 if (cl
!= 1) /* only one class */
621 tcm
->tcm_handle
|= TC_H_MIN(1);
622 tcm
->tcm_info
= q
->qdisc
->handle
;
627 static int netem_graft(struct Qdisc
*sch
, unsigned long arg
, struct Qdisc
*new,
630 struct netem_sched_data
*q
= qdisc_priv(sch
);
636 *old
= xchg(&q
->qdisc
, new);
637 qdisc_tree_decrease_qlen(*old
, (*old
)->q
.qlen
);
639 sch_tree_unlock(sch
);
644 static struct Qdisc
*netem_leaf(struct Qdisc
*sch
, unsigned long arg
)
646 struct netem_sched_data
*q
= qdisc_priv(sch
);
650 static unsigned long netem_get(struct Qdisc
*sch
, u32 classid
)
655 static void netem_put(struct Qdisc
*sch
, unsigned long arg
)
659 static int netem_change_class(struct Qdisc
*sch
, u32 classid
, u32 parentid
,
660 struct nlattr
**tca
, unsigned long *arg
)
665 static int netem_delete(struct Qdisc
*sch
, unsigned long arg
)
670 static void netem_walk(struct Qdisc
*sch
, struct qdisc_walker
*walker
)
673 if (walker
->count
>= walker
->skip
)
674 if (walker
->fn(sch
, 1, walker
) < 0) {
682 static struct tcf_proto
**netem_find_tcf(struct Qdisc
*sch
, unsigned long cl
)
687 static const struct Qdisc_class_ops netem_class_ops
= {
688 .graft
= netem_graft
,
692 .change
= netem_change_class
,
693 .delete = netem_delete
,
695 .tcf_chain
= netem_find_tcf
,
696 .dump
= netem_dump_class
,
699 static struct Qdisc_ops netem_qdisc_ops __read_mostly
= {
701 .cl_ops
= &netem_class_ops
,
702 .priv_size
= sizeof(struct netem_sched_data
),
703 .enqueue
= netem_enqueue
,
704 .dequeue
= netem_dequeue
,
705 .requeue
= netem_requeue
,
708 .reset
= netem_reset
,
709 .destroy
= netem_destroy
,
710 .change
= netem_change
,
712 .owner
= THIS_MODULE
,
716 static int __init
netem_module_init(void)
718 pr_info("netem: version " VERSION
"\n");
719 return register_qdisc(&netem_qdisc_ops
);
721 static void __exit
netem_module_exit(void)
723 unregister_qdisc(&netem_qdisc_ops
);
725 module_init(netem_module_init
)
726 module_exit(netem_module_exit
)
727 MODULE_LICENSE("GPL");