1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * net/sched/sch_netem.c Network emulator
4 *
5 * Many of the algorithms and ideas for this came from
6 * NIST Net which is not copyrighted.
7 *
8 * Authors: Stephen Hemminger <shemminger@osdl.org>
9 * Catalin(ux aka Dino) BOIE <catab at umbrella dot ro>
10 */
11
12 #include <linux/mm.h>
13 #include <linux/module.h>
14 #include <linux/slab.h>
15 #include <linux/types.h>
16 #include <linux/kernel.h>
17 #include <linux/errno.h>
18 #include <linux/skbuff.h>
19 #include <linux/vmalloc.h>
20 #include <linux/prandom.h>
21 #include <linux/rtnetlink.h>
22 #include <linux/reciprocal_div.h>
23 #include <linux/rbtree.h>
24
25 #include <net/gso.h>
26 #include <net/netlink.h>
27 #include <net/pkt_sched.h>
28 #include <net/inet_ecn.h>
29
30 #define VERSION "1.3"
31
32 /* Network Emulation Queuing algorithm.
33 ====================================
34
35 Sources: [1] Mark Carson, Darrin Santay, "NIST Net - A Linux-based
36 Network Emulation Tool
37 [2] Luigi Rizzo, DummyNet for FreeBSD
38
39 ----------------------------------------------------------------
40
41 This started out as a simple way to delay outgoing packets to
42 test TCP but has grown to include most of the functionality
43 of a full blown network emulator like NISTnet. It can delay
44 packets and add random jitter (and correlation). The random
45 distribution can be loaded from a table as well to provide
46 normal, Pareto, or experimental curves. Packet loss,
47 duplication, and reordering can also be emulated.
48
49 This qdisc does not do classification that can be handled in
50 layering other disciplines. It does not need to do bandwidth
51 control either since that can be handled by using token
52 bucket or other rate control.
53
54 Correlated Loss Generator models
55
56 Added generation of correlated loss according to the
57 "Gilbert-Elliot" model, a 4-state markov model.
58
59 References:
60 [1] NetemCLG Home http://netgroup.uniroma2.it/NetemCLG
61 [2] S. Salsano, F. Ludovici, A. Ordine, "Definition of a general
62 and intuitive loss model for packet networks and its implementation
63 in the Netem module in the Linux kernel", available in [1]
64
65 Authors: Stefano Salsano <stefano.salsano at uniroma2.it
66 Fabio Ludovici <fabio.ludovici at yahoo.it>
67 */
68
69 struct disttable {
70 u32 size;
71 s16 table[] __counted_by(size);
72 };
73
74 struct netem_sched_data {
75 /* internal t(ime)fifo qdisc uses t_root and sch->limit */
76 struct rb_root t_root;
77
78 /* a linear queue; reduces rbtree rebalancing when jitter is low */
79 struct sk_buff *t_head;
80 struct sk_buff *t_tail;
81
82 u32 t_len;
83
84 /* optional qdisc for classful handling (NULL at netem init) */
85 struct Qdisc *qdisc;
86
87 struct qdisc_watchdog watchdog;
88
89 s64 latency;
90 s64 jitter;
91
92 u32 loss;
93 u32 ecn;
94 u32 limit;
95 u32 counter;
96 u32 gap;
97 u32 duplicate;
98 u32 reorder;
99 u32 corrupt;
100 u64 rate;
101 s32 packet_overhead;
102 u32 cell_size;
103 struct reciprocal_value cell_size_reciprocal;
104 s32 cell_overhead;
105
106 struct crndstate {
107 u32 last;
108 u32 rho;
109 } delay_cor, loss_cor, dup_cor, reorder_cor, corrupt_cor;
110
111 struct prng {
112 u64 seed;
113 struct rnd_state prng_state;
114 } prng;
115
116 struct disttable *delay_dist;
117
118 enum {
119 CLG_RANDOM,
120 CLG_4_STATES,
121 CLG_GILB_ELL,
122 } loss_model;
123
124 enum {
125 TX_IN_GAP_PERIOD = 1,
126 TX_IN_BURST_PERIOD,
127 LOST_IN_GAP_PERIOD,
128 LOST_IN_BURST_PERIOD,
129 } _4_state_model;
130
131 enum {
132 GOOD_STATE = 1,
133 BAD_STATE,
134 } GE_state_model;
135
136 /* Correlated Loss Generation models */
137 struct clgstate {
138 /* state of the Markov chain */
139 u8 state;
140
141 /* 4-states and Gilbert-Elliot models */
142 u32 a1; /* p13 for 4-states or p for GE */
143 u32 a2; /* p31 for 4-states or r for GE */
144 u32 a3; /* p32 for 4-states or h for GE */
145 u32 a4; /* p14 for 4-states or 1-k for GE */
146 u32 a5; /* p23 used only in 4-states */
147 } clg;
148
149 struct tc_netem_slot slot_config;
150 struct slotstate {
151 u64 slot_next;
152 s32 packets_left;
153 s32 bytes_left;
154 } slot;
155
156 struct disttable *slot_dist;
157 };
158
159 /* Time stamp put into socket buffer control block
160 * Only valid when skbs are in our internal t(ime)fifo queue.
161 *
162 * As skb->rbnode uses same storage than skb->next, skb->prev and skb->tstamp,
163 * and skb->next & skb->prev are scratch space for a qdisc,
164 * we save skb->tstamp value in skb->cb[] before destroying it.
165 */
166 struct netem_skb_cb {
167 u64 time_to_send;
168 };
169
netem_skb_cb(struct sk_buff * skb)170 static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb)
171 {
172 /* we assume we can use skb next/prev/tstamp as storage for rb_node */
173 qdisc_cb_private_validate(skb, sizeof(struct netem_skb_cb));
174 return (struct netem_skb_cb *)qdisc_skb_cb(skb)->data;
175 }
176
177 /* init_crandom - initialize correlated random number generator
178 * Use entropy source for initial seed.
179 */
init_crandom(struct crndstate * state,unsigned long rho)180 static void init_crandom(struct crndstate *state, unsigned long rho)
181 {
182 state->rho = rho;
183 state->last = get_random_u32();
184 }
185
186 /* get_crandom - correlated random number generator
187 * Next number depends on last value.
188 * rho is scaled to avoid floating point.
189 */
get_crandom(struct crndstate * state,struct prng * p)190 static u32 get_crandom(struct crndstate *state, struct prng *p)
191 {
192 u64 value, rho;
193 unsigned long answer;
194 struct rnd_state *s = &p->prng_state;
195
196 if (!state || state->rho == 0) /* no correlation */
197 return prandom_u32_state(s);
198
199 value = prandom_u32_state(s);
200 rho = (u64)state->rho + 1;
201 answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32;
202 state->last = answer;
203 return answer;
204 }
205
206 /* loss_4state - 4-state model loss generator
207 * Generates losses according to the 4-state Markov chain adopted in
208 * the GI (General and Intuitive) loss model.
209 */
loss_4state(struct netem_sched_data * q)210 static bool loss_4state(struct netem_sched_data *q)
211 {
212 struct clgstate *clg = &q->clg;
213 u32 rnd = prandom_u32_state(&q->prng.prng_state);
214
215 /*
216 * Makes a comparison between rnd and the transition
217 * probabilities outgoing from the current state, then decides the
218 * next state and if the next packet has to be transmitted or lost.
219 * The four states correspond to:
220 * TX_IN_GAP_PERIOD => successfully transmitted packets within a gap period
221 * LOST_IN_GAP_PERIOD => isolated losses within a gap period
222 * LOST_IN_BURST_PERIOD => lost packets within a burst period
223 * TX_IN_BURST_PERIOD => successfully transmitted packets within a burst period
224 */
225 switch (clg->state) {
226 case TX_IN_GAP_PERIOD:
227 if (rnd < clg->a4) {
228 clg->state = LOST_IN_GAP_PERIOD;
229 return true;
230 } else if (clg->a4 < rnd && rnd < clg->a1 + clg->a4) {
231 clg->state = LOST_IN_BURST_PERIOD;
232 return true;
233 } else if (clg->a1 + clg->a4 < rnd) {
234 clg->state = TX_IN_GAP_PERIOD;
235 }
236
237 break;
238 case TX_IN_BURST_PERIOD:
239 if (rnd < clg->a5) {
240 clg->state = LOST_IN_BURST_PERIOD;
241 return true;
242 } else {
243 clg->state = TX_IN_BURST_PERIOD;
244 }
245
246 break;
247 case LOST_IN_BURST_PERIOD:
248 if (rnd < clg->a3)
249 clg->state = TX_IN_BURST_PERIOD;
250 else if (clg->a3 < rnd && rnd < clg->a2 + clg->a3) {
251 clg->state = TX_IN_GAP_PERIOD;
252 } else if (clg->a2 + clg->a3 < rnd) {
253 clg->state = LOST_IN_BURST_PERIOD;
254 return true;
255 }
256 break;
257 case LOST_IN_GAP_PERIOD:
258 clg->state = TX_IN_GAP_PERIOD;
259 break;
260 }
261
262 return false;
263 }
264
265 /* loss_gilb_ell - Gilbert-Elliot model loss generator
266 * Generates losses according to the Gilbert-Elliot loss model or
267 * its special cases (Gilbert or Simple Gilbert)
268 *
269 * Makes a comparison between random number and the transition
270 * probabilities outgoing from the current state, then decides the
271 * next state. A second random number is extracted and the comparison
272 * with the loss probability of the current state decides if the next
273 * packet will be transmitted or lost.
274 */
loss_gilb_ell(struct netem_sched_data * q)275 static bool loss_gilb_ell(struct netem_sched_data *q)
276 {
277 struct clgstate *clg = &q->clg;
278 struct rnd_state *s = &q->prng.prng_state;
279
280 switch (clg->state) {
281 case GOOD_STATE:
282 if (prandom_u32_state(s) < clg->a1)
283 clg->state = BAD_STATE;
284 if (prandom_u32_state(s) < clg->a4)
285 return true;
286 break;
287 case BAD_STATE:
288 if (prandom_u32_state(s) < clg->a2)
289 clg->state = GOOD_STATE;
290 if (prandom_u32_state(s) > clg->a3)
291 return true;
292 }
293
294 return false;
295 }
296
loss_event(struct netem_sched_data * q)297 static bool loss_event(struct netem_sched_data *q)
298 {
299 switch (q->loss_model) {
300 case CLG_RANDOM:
301 /* Random packet drop 0 => none, ~0 => all */
302 return q->loss && q->loss >= get_crandom(&q->loss_cor, &q->prng);
303
304 case CLG_4_STATES:
305 /* 4state loss model algorithm (used also for GI model)
306 * Extracts a value from the markov 4 state loss generator,
307 * if it is 1 drops a packet and if needed writes the event in
308 * the kernel logs
309 */
310 return loss_4state(q);
311
312 case CLG_GILB_ELL:
313 /* Gilbert-Elliot loss model algorithm
314 * Extracts a value from the Gilbert-Elliot loss generator,
315 * if it is 1 drops a packet and if needed writes the event in
316 * the kernel logs
317 */
318 return loss_gilb_ell(q);
319 }
320
321 return false; /* not reached */
322 }
323
324
325 /* tabledist - return a pseudo-randomly distributed value with mean mu and
326 * std deviation sigma. Uses table lookup to approximate the desired
327 * distribution, and a uniformly-distributed pseudo-random source.
328 */
tabledist(s64 mu,s32 sigma,struct crndstate * state,struct prng * prng,const struct disttable * dist)329 static s64 tabledist(s64 mu, s32 sigma,
330 struct crndstate *state,
331 struct prng *prng,
332 const struct disttable *dist)
333 {
334 s64 x;
335 long t;
336 u32 rnd;
337
338 if (sigma == 0)
339 return mu;
340
341 rnd = get_crandom(state, prng);
342
343 /* default uniform distribution */
344 if (dist == NULL)
345 return ((rnd % (2 * (u32)sigma)) + mu) - sigma;
346
347 t = dist->table[rnd % dist->size];
348 x = (sigma % NETEM_DIST_SCALE) * t;
349 if (x >= 0)
350 x += NETEM_DIST_SCALE/2;
351 else
352 x -= NETEM_DIST_SCALE/2;
353
354 return x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu;
355 }
356
packet_time_ns(u64 len,const struct netem_sched_data * q)357 static u64 packet_time_ns(u64 len, const struct netem_sched_data *q)
358 {
359 len += q->packet_overhead;
360
361 if (q->cell_size) {
362 u32 cells = reciprocal_divide(len, q->cell_size_reciprocal);
363
364 if (len > cells * q->cell_size) /* extra cell needed for remainder */
365 cells++;
366 len = cells * (q->cell_size + q->cell_overhead);
367 }
368
369 return div64_u64(len * NSEC_PER_SEC, q->rate);
370 }
371
tfifo_reset(struct Qdisc * sch)372 static void tfifo_reset(struct Qdisc *sch)
373 {
374 struct netem_sched_data *q = qdisc_priv(sch);
375 struct rb_node *p = rb_first(&q->t_root);
376
377 while (p) {
378 struct sk_buff *skb = rb_to_skb(p);
379
380 p = rb_next(p);
381 rb_erase(&skb->rbnode, &q->t_root);
382 rtnl_kfree_skbs(skb, skb);
383 }
384
385 rtnl_kfree_skbs(q->t_head, q->t_tail);
386 q->t_head = NULL;
387 q->t_tail = NULL;
388 q->t_len = 0;
389 }
390
tfifo_enqueue(struct sk_buff * nskb,struct Qdisc * sch)391 static void tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
392 {
393 struct netem_sched_data *q = qdisc_priv(sch);
394 u64 tnext = netem_skb_cb(nskb)->time_to_send;
395
396 if (!q->t_tail || tnext >= netem_skb_cb(q->t_tail)->time_to_send) {
397 if (q->t_tail)
398 q->t_tail->next = nskb;
399 else
400 q->t_head = nskb;
401 q->t_tail = nskb;
402 } else {
403 struct rb_node **p = &q->t_root.rb_node, *parent = NULL;
404
405 while (*p) {
406 struct sk_buff *skb;
407
408 parent = *p;
409 skb = rb_to_skb(parent);
410 if (tnext >= netem_skb_cb(skb)->time_to_send)
411 p = &parent->rb_right;
412 else
413 p = &parent->rb_left;
414 }
415 rb_link_node(&nskb->rbnode, parent, p);
416 rb_insert_color(&nskb->rbnode, &q->t_root);
417 }
418 q->t_len++;
419 sch->q.qlen++;
420 }
421
422 /* netem can't properly corrupt a megapacket (like we get from GSO), so instead
423 * when we statistically choose to corrupt one, we instead segment it, returning
424 * the first packet to be corrupted, and re-enqueue the remaining frames
425 */
netem_segment(struct sk_buff * skb,struct Qdisc * sch,struct sk_buff ** to_free)426 static struct sk_buff *netem_segment(struct sk_buff *skb, struct Qdisc *sch,
427 struct sk_buff **to_free)
428 {
429 struct sk_buff *segs;
430 netdev_features_t features = netif_skb_features(skb);
431
432 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
433
434 if (IS_ERR_OR_NULL(segs)) {
435 qdisc_drop(skb, sch, to_free);
436 return NULL;
437 }
438 consume_skb(skb);
439 return segs;
440 }
441
442 /*
443 * Insert one skb into qdisc.
444 * Note: parent depends on return value to account for queue length.
445 * NET_XMIT_DROP: queue length didn't change.
446 * NET_XMIT_SUCCESS: one skb was queued.
447 */
netem_enqueue(struct sk_buff * skb,struct Qdisc * sch,struct sk_buff ** to_free)448 static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch,
449 struct sk_buff **to_free)
450 {
451 struct netem_sched_data *q = qdisc_priv(sch);
452 /* We don't fill cb now as skb_unshare() may invalidate it */
453 struct netem_skb_cb *cb;
454 struct sk_buff *skb2 = NULL;
455 struct sk_buff *segs = NULL;
456 unsigned int prev_len = qdisc_pkt_len(skb);
457 int count = 1;
458
459 /* Do not fool qdisc_drop_all() */
460 skb->prev = NULL;
461
462 /* Random duplication */
463 if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor, &q->prng))
464 ++count;
465
466 /* Drop packet? */
467 if (loss_event(q)) {
468 if (q->ecn && INET_ECN_set_ce(skb))
469 qdisc_qstats_drop(sch); /* mark packet */
470 else
471 --count;
472 }
473 if (count == 0) {
474 qdisc_qstats_drop(sch);
475 __qdisc_drop(skb, to_free);
476 return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
477 }
478
479 /* If a delay is expected, orphan the skb. (orphaning usually takes
480 * place at TX completion time, so _before_ the link transit delay)
481 */
482 if (q->latency || q->jitter || q->rate)
483 skb_orphan_partial(skb);
484
485 /*
486 * If we need to duplicate packet, then clone it before
487 * original is modified.
488 */
489 if (count > 1)
490 skb2 = skb_clone(skb, GFP_ATOMIC);
491
492 /*
493 * Randomized packet corruption.
494 * Make copy if needed since we are modifying
495 * If packet is going to be hardware checksummed, then
496 * do it now in software before we mangle it.
497 */
498 if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor, &q->prng)) {
499 if (skb_is_gso(skb)) {
500 skb = netem_segment(skb, sch, to_free);
501 if (!skb)
502 goto finish_segs;
503
504 segs = skb->next;
505 skb_mark_not_on_list(skb);
506 qdisc_skb_cb(skb)->pkt_len = skb->len;
507 }
508
509 skb = skb_unshare(skb, GFP_ATOMIC);
510 if (unlikely(!skb)) {
511 qdisc_qstats_drop(sch);
512 goto finish_segs;
513 }
514 if (skb->ip_summed == CHECKSUM_PARTIAL &&
515 skb_checksum_help(skb)) {
516 qdisc_drop(skb, sch, to_free);
517 skb = NULL;
518 goto finish_segs;
519 }
520
521 skb->data[get_random_u32_below(skb_headlen(skb))] ^=
522 1<<get_random_u32_below(8);
523 }
524
525 if (unlikely(q->t_len >= sch->limit)) {
526 /* re-link segs, so that qdisc_drop_all() frees them all */
527 skb->next = segs;
528 qdisc_drop_all(skb, sch, to_free);
529 if (skb2)
530 __qdisc_drop(skb2, to_free);
531 return NET_XMIT_DROP;
532 }
533
534 /*
535 * If doing duplication then re-insert at top of the
536 * qdisc tree, since parent queuer expects that only one
537 * skb will be queued.
538 */
539 if (skb2) {
540 struct Qdisc *rootq = qdisc_root_bh(sch);
541 u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
542
543 q->duplicate = 0;
544 rootq->enqueue(skb2, rootq, to_free);
545 q->duplicate = dupsave;
546 skb2 = NULL;
547 }
548
549 qdisc_qstats_backlog_inc(sch, skb);
550
551 cb = netem_skb_cb(skb);
552 if (q->gap == 0 || /* not doing reordering */
553 q->counter < q->gap - 1 || /* inside last reordering gap */
554 q->reorder < get_crandom(&q->reorder_cor, &q->prng)) {
555 u64 now;
556 s64 delay;
557
558 delay = tabledist(q->latency, q->jitter,
559 &q->delay_cor, &q->prng, q->delay_dist);
560
561 now = ktime_get_ns();
562
563 if (q->rate) {
564 struct netem_skb_cb *last = NULL;
565
566 if (sch->q.tail)
567 last = netem_skb_cb(sch->q.tail);
568 if (q->t_root.rb_node) {
569 struct sk_buff *t_skb;
570 struct netem_skb_cb *t_last;
571
572 t_skb = skb_rb_last(&q->t_root);
573 t_last = netem_skb_cb(t_skb);
574 if (!last ||
575 t_last->time_to_send > last->time_to_send)
576 last = t_last;
577 }
578 if (q->t_tail) {
579 struct netem_skb_cb *t_last =
580 netem_skb_cb(q->t_tail);
581
582 if (!last ||
583 t_last->time_to_send > last->time_to_send)
584 last = t_last;
585 }
586
587 if (last) {
588 /*
589 * Last packet in queue is reference point (now),
590 * calculate this time bonus and subtract
591 * from delay.
592 */
593 delay -= last->time_to_send - now;
594 delay = max_t(s64, 0, delay);
595 now = last->time_to_send;
596 }
597
598 delay += packet_time_ns(qdisc_pkt_len(skb), q);
599 }
600
601 cb->time_to_send = now + delay;
602 ++q->counter;
603 tfifo_enqueue(skb, sch);
604 } else {
605 /*
606 * Do re-ordering by putting one out of N packets at the front
607 * of the queue.
608 */
609 cb->time_to_send = ktime_get_ns();
610 q->counter = 0;
611
612 __qdisc_enqueue_head(skb, &sch->q);
613 sch->qstats.requeues++;
614 }
615
616 finish_segs:
617 if (skb2)
618 __qdisc_drop(skb2, to_free);
619
620 if (segs) {
621 unsigned int len, last_len;
622 int rc, nb;
623
624 len = skb ? skb->len : 0;
625 nb = skb ? 1 : 0;
626
627 while (segs) {
628 skb2 = segs->next;
629 skb_mark_not_on_list(segs);
630 qdisc_skb_cb(segs)->pkt_len = segs->len;
631 last_len = segs->len;
632 rc = qdisc_enqueue(segs, sch, to_free);
633 if (rc != NET_XMIT_SUCCESS) {
634 if (net_xmit_drop_count(rc))
635 qdisc_qstats_drop(sch);
636 } else {
637 nb++;
638 len += last_len;
639 }
640 segs = skb2;
641 }
642 /* Parent qdiscs accounted for 1 skb of size @prev_len */
643 qdisc_tree_reduce_backlog(sch, -(nb - 1), -(len - prev_len));
644 } else if (!skb) {
645 return NET_XMIT_DROP;
646 }
647 return NET_XMIT_SUCCESS;
648 }
649
650 /* Delay the next round with a new future slot with a
651 * correct number of bytes and packets.
652 */
653
get_slot_next(struct netem_sched_data * q,u64 now)654 static void get_slot_next(struct netem_sched_data *q, u64 now)
655 {
656 s64 next_delay;
657
658 if (!q->slot_dist)
659 next_delay = q->slot_config.min_delay +
660 (get_random_u32() *
661 (q->slot_config.max_delay -
662 q->slot_config.min_delay) >> 32);
663 else
664 next_delay = tabledist(q->slot_config.dist_delay,
665 (s32)(q->slot_config.dist_jitter),
666 NULL, &q->prng, q->slot_dist);
667
668 q->slot.slot_next = now + next_delay;
669 q->slot.packets_left = q->slot_config.max_packets;
670 q->slot.bytes_left = q->slot_config.max_bytes;
671 }
672
netem_peek(struct netem_sched_data * q)673 static struct sk_buff *netem_peek(struct netem_sched_data *q)
674 {
675 struct sk_buff *skb = skb_rb_first(&q->t_root);
676 u64 t1, t2;
677
678 if (!skb)
679 return q->t_head;
680 if (!q->t_head)
681 return skb;
682
683 t1 = netem_skb_cb(skb)->time_to_send;
684 t2 = netem_skb_cb(q->t_head)->time_to_send;
685 if (t1 < t2)
686 return skb;
687 return q->t_head;
688 }
689
netem_erase_head(struct netem_sched_data * q,struct sk_buff * skb)690 static void netem_erase_head(struct netem_sched_data *q, struct sk_buff *skb)
691 {
692 if (skb == q->t_head) {
693 q->t_head = skb->next;
694 if (!q->t_head)
695 q->t_tail = NULL;
696 } else {
697 rb_erase(&skb->rbnode, &q->t_root);
698 }
699 }
700
netem_dequeue(struct Qdisc * sch)701 static struct sk_buff *netem_dequeue(struct Qdisc *sch)
702 {
703 struct netem_sched_data *q = qdisc_priv(sch);
704 struct sk_buff *skb;
705
706 tfifo_dequeue:
707 skb = __qdisc_dequeue_head(&sch->q);
708 if (skb) {
709 deliver:
710 qdisc_qstats_backlog_dec(sch, skb);
711 qdisc_bstats_update(sch, skb);
712 return skb;
713 }
714 skb = netem_peek(q);
715 if (skb) {
716 u64 time_to_send;
717 u64 now = ktime_get_ns();
718
719 /* if more time remaining? */
720 time_to_send = netem_skb_cb(skb)->time_to_send;
721 if (q->slot.slot_next && q->slot.slot_next < time_to_send)
722 get_slot_next(q, now);
723
724 if (time_to_send <= now && q->slot.slot_next <= now) {
725 netem_erase_head(q, skb);
726 q->t_len--;
727 skb->next = NULL;
728 skb->prev = NULL;
729 /* skb->dev shares skb->rbnode area,
730 * we need to restore its value.
731 */
732 skb->dev = qdisc_dev(sch);
733
734 if (q->slot.slot_next) {
735 q->slot.packets_left--;
736 q->slot.bytes_left -= qdisc_pkt_len(skb);
737 if (q->slot.packets_left <= 0 ||
738 q->slot.bytes_left <= 0)
739 get_slot_next(q, now);
740 }
741
742 if (q->qdisc) {
743 unsigned int pkt_len = qdisc_pkt_len(skb);
744 struct sk_buff *to_free = NULL;
745 int err;
746
747 err = qdisc_enqueue(skb, q->qdisc, &to_free);
748 kfree_skb_list(to_free);
749 if (err != NET_XMIT_SUCCESS) {
750 if (net_xmit_drop_count(err))
751 qdisc_qstats_drop(sch);
752 qdisc_tree_reduce_backlog(sch, 1, pkt_len);
753 sch->qstats.backlog -= pkt_len;
754 sch->q.qlen--;
755 }
756 goto tfifo_dequeue;
757 }
758 sch->q.qlen--;
759 goto deliver;
760 }
761
762 if (q->qdisc) {
763 skb = q->qdisc->ops->dequeue(q->qdisc);
764 if (skb) {
765 sch->q.qlen--;
766 goto deliver;
767 }
768 }
769
770 qdisc_watchdog_schedule_ns(&q->watchdog,
771 max(time_to_send,
772 q->slot.slot_next));
773 }
774
775 if (q->qdisc) {
776 skb = q->qdisc->ops->dequeue(q->qdisc);
777 if (skb) {
778 sch->q.qlen--;
779 goto deliver;
780 }
781 }
782 return NULL;
783 }
784
netem_reset(struct Qdisc * sch)785 static void netem_reset(struct Qdisc *sch)
786 {
787 struct netem_sched_data *q = qdisc_priv(sch);
788
789 qdisc_reset_queue(sch);
790 tfifo_reset(sch);
791 if (q->qdisc)
792 qdisc_reset(q->qdisc);
793 qdisc_watchdog_cancel(&q->watchdog);
794 }
795
dist_free(struct disttable * d)796 static void dist_free(struct disttable *d)
797 {
798 kvfree(d);
799 }
800
801 /*
802 * Distribution data is a variable size payload containing
803 * signed 16 bit values.
804 */
805
get_dist_table(struct disttable ** tbl,const struct nlattr * attr)806 static int get_dist_table(struct disttable **tbl, const struct nlattr *attr)
807 {
808 size_t n = nla_len(attr)/sizeof(__s16);
809 const __s16 *data = nla_data(attr);
810 struct disttable *d;
811 int i;
812
813 if (!n || n > NETEM_DIST_MAX)
814 return -EINVAL;
815
816 d = kvmalloc(struct_size(d, table, n), GFP_KERNEL);
817 if (!d)
818 return -ENOMEM;
819
820 d->size = n;
821 for (i = 0; i < n; i++)
822 d->table[i] = data[i];
823
824 *tbl = d;
825 return 0;
826 }
827
get_slot(struct netem_sched_data * q,const struct nlattr * attr)828 static void get_slot(struct netem_sched_data *q, const struct nlattr *attr)
829 {
830 const struct tc_netem_slot *c = nla_data(attr);
831
832 q->slot_config = *c;
833 if (q->slot_config.max_packets == 0)
834 q->slot_config.max_packets = INT_MAX;
835 if (q->slot_config.max_bytes == 0)
836 q->slot_config.max_bytes = INT_MAX;
837
838 /* capping dist_jitter to the range acceptable by tabledist() */
839 q->slot_config.dist_jitter = min_t(__s64, INT_MAX, abs(q->slot_config.dist_jitter));
840
841 q->slot.packets_left = q->slot_config.max_packets;
842 q->slot.bytes_left = q->slot_config.max_bytes;
843 if (q->slot_config.min_delay | q->slot_config.max_delay |
844 q->slot_config.dist_jitter)
845 q->slot.slot_next = ktime_get_ns();
846 else
847 q->slot.slot_next = 0;
848 }
849
get_correlation(struct netem_sched_data * q,const struct nlattr * attr)850 static void get_correlation(struct netem_sched_data *q, const struct nlattr *attr)
851 {
852 const struct tc_netem_corr *c = nla_data(attr);
853
854 init_crandom(&q->delay_cor, c->delay_corr);
855 init_crandom(&q->loss_cor, c->loss_corr);
856 init_crandom(&q->dup_cor, c->dup_corr);
857 }
858
get_reorder(struct netem_sched_data * q,const struct nlattr * attr)859 static void get_reorder(struct netem_sched_data *q, const struct nlattr *attr)
860 {
861 const struct tc_netem_reorder *r = nla_data(attr);
862
863 q->reorder = r->probability;
864 init_crandom(&q->reorder_cor, r->correlation);
865 }
866
get_corrupt(struct netem_sched_data * q,const struct nlattr * attr)867 static void get_corrupt(struct netem_sched_data *q, const struct nlattr *attr)
868 {
869 const struct tc_netem_corrupt *r = nla_data(attr);
870
871 q->corrupt = r->probability;
872 init_crandom(&q->corrupt_cor, r->correlation);
873 }
874
get_rate(struct netem_sched_data * q,const struct nlattr * attr)875 static void get_rate(struct netem_sched_data *q, const struct nlattr *attr)
876 {
877 const struct tc_netem_rate *r = nla_data(attr);
878
879 q->rate = r->rate;
880 q->packet_overhead = r->packet_overhead;
881 q->cell_size = r->cell_size;
882 q->cell_overhead = r->cell_overhead;
883 if (q->cell_size)
884 q->cell_size_reciprocal = reciprocal_value(q->cell_size);
885 else
886 q->cell_size_reciprocal = (struct reciprocal_value) { 0 };
887 }
888
get_loss_clg(struct netem_sched_data * q,const struct nlattr * attr)889 static int get_loss_clg(struct netem_sched_data *q, const struct nlattr *attr)
890 {
891 const struct nlattr *la;
892 int rem;
893
894 nla_for_each_nested(la, attr, rem) {
895 u16 type = nla_type(la);
896
897 switch (type) {
898 case NETEM_LOSS_GI: {
899 const struct tc_netem_gimodel *gi = nla_data(la);
900
901 if (nla_len(la) < sizeof(struct tc_netem_gimodel)) {
902 pr_info("netem: incorrect gi model size\n");
903 return -EINVAL;
904 }
905
906 q->loss_model = CLG_4_STATES;
907
908 q->clg.state = TX_IN_GAP_PERIOD;
909 q->clg.a1 = gi->p13;
910 q->clg.a2 = gi->p31;
911 q->clg.a3 = gi->p32;
912 q->clg.a4 = gi->p14;
913 q->clg.a5 = gi->p23;
914 break;
915 }
916
917 case NETEM_LOSS_GE: {
918 const struct tc_netem_gemodel *ge = nla_data(la);
919
920 if (nla_len(la) < sizeof(struct tc_netem_gemodel)) {
921 pr_info("netem: incorrect ge model size\n");
922 return -EINVAL;
923 }
924
925 q->loss_model = CLG_GILB_ELL;
926 q->clg.state = GOOD_STATE;
927 q->clg.a1 = ge->p;
928 q->clg.a2 = ge->r;
929 q->clg.a3 = ge->h;
930 q->clg.a4 = ge->k1;
931 break;
932 }
933
934 default:
935 pr_info("netem: unknown loss type %u\n", type);
936 return -EINVAL;
937 }
938 }
939
940 return 0;
941 }
942
943 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
944 [TCA_NETEM_CORR] = { .len = sizeof(struct tc_netem_corr) },
945 [TCA_NETEM_REORDER] = { .len = sizeof(struct tc_netem_reorder) },
946 [TCA_NETEM_CORRUPT] = { .len = sizeof(struct tc_netem_corrupt) },
947 [TCA_NETEM_RATE] = { .len = sizeof(struct tc_netem_rate) },
948 [TCA_NETEM_LOSS] = { .type = NLA_NESTED },
949 [TCA_NETEM_ECN] = { .type = NLA_U32 },
950 [TCA_NETEM_RATE64] = { .type = NLA_U64 },
951 [TCA_NETEM_LATENCY64] = { .type = NLA_S64 },
952 [TCA_NETEM_JITTER64] = { .type = NLA_S64 },
953 [TCA_NETEM_SLOT] = { .len = sizeof(struct tc_netem_slot) },
954 [TCA_NETEM_PRNG_SEED] = { .type = NLA_U64 },
955 };
956
parse_attr(struct nlattr * tb[],int maxtype,struct nlattr * nla,const struct nla_policy * policy,int len)957 static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
958 const struct nla_policy *policy, int len)
959 {
960 int nested_len = nla_len(nla) - NLA_ALIGN(len);
961
962 if (nested_len < 0) {
963 pr_info("netem: invalid attributes len %d\n", nested_len);
964 return -EINVAL;
965 }
966
967 if (nested_len >= nla_attr_size(0))
968 return nla_parse_deprecated(tb, maxtype,
969 nla_data(nla) + NLA_ALIGN(len),
970 nested_len, policy, NULL);
971
972 memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
973 return 0;
974 }
975
976 /* Parse netlink message to set options */
netem_change(struct Qdisc * sch,struct nlattr * opt,struct netlink_ext_ack * extack)977 static int netem_change(struct Qdisc *sch, struct nlattr *opt,
978 struct netlink_ext_ack *extack)
979 {
980 struct netem_sched_data *q = qdisc_priv(sch);
981 struct nlattr *tb[TCA_NETEM_MAX + 1];
982 struct disttable *delay_dist = NULL;
983 struct disttable *slot_dist = NULL;
984 struct tc_netem_qopt *qopt;
985 struct clgstate old_clg;
986 int old_loss_model = CLG_RANDOM;
987 int ret;
988
989 qopt = nla_data(opt);
990 ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
991 if (ret < 0)
992 return ret;
993
994 if (tb[TCA_NETEM_DELAY_DIST]) {
995 ret = get_dist_table(&delay_dist, tb[TCA_NETEM_DELAY_DIST]);
996 if (ret)
997 goto table_free;
998 }
999
1000 if (tb[TCA_NETEM_SLOT_DIST]) {
1001 ret = get_dist_table(&slot_dist, tb[TCA_NETEM_SLOT_DIST]);
1002 if (ret)
1003 goto table_free;
1004 }
1005
1006 sch_tree_lock(sch);
1007 /* backup q->clg and q->loss_model */
1008 old_clg = q->clg;
1009 old_loss_model = q->loss_model;
1010
1011 if (tb[TCA_NETEM_LOSS]) {
1012 ret = get_loss_clg(q, tb[TCA_NETEM_LOSS]);
1013 if (ret) {
1014 q->loss_model = old_loss_model;
1015 q->clg = old_clg;
1016 goto unlock;
1017 }
1018 } else {
1019 q->loss_model = CLG_RANDOM;
1020 }
1021
1022 if (delay_dist)
1023 swap(q->delay_dist, delay_dist);
1024 if (slot_dist)
1025 swap(q->slot_dist, slot_dist);
1026 sch->limit = qopt->limit;
1027
1028 q->latency = PSCHED_TICKS2NS(qopt->latency);
1029 q->jitter = PSCHED_TICKS2NS(qopt->jitter);
1030 q->limit = qopt->limit;
1031 q->gap = qopt->gap;
1032 q->counter = 0;
1033 q->loss = qopt->loss;
1034 q->duplicate = qopt->duplicate;
1035
1036 /* for compatibility with earlier versions.
1037 * if gap is set, need to assume 100% probability
1038 */
1039 if (q->gap)
1040 q->reorder = ~0;
1041
1042 if (tb[TCA_NETEM_CORR])
1043 get_correlation(q, tb[TCA_NETEM_CORR]);
1044
1045 if (tb[TCA_NETEM_REORDER])
1046 get_reorder(q, tb[TCA_NETEM_REORDER]);
1047
1048 if (tb[TCA_NETEM_CORRUPT])
1049 get_corrupt(q, tb[TCA_NETEM_CORRUPT]);
1050
1051 if (tb[TCA_NETEM_RATE])
1052 get_rate(q, tb[TCA_NETEM_RATE]);
1053
1054 if (tb[TCA_NETEM_RATE64])
1055 q->rate = max_t(u64, q->rate,
1056 nla_get_u64(tb[TCA_NETEM_RATE64]));
1057
1058 if (tb[TCA_NETEM_LATENCY64])
1059 q->latency = nla_get_s64(tb[TCA_NETEM_LATENCY64]);
1060
1061 if (tb[TCA_NETEM_JITTER64])
1062 q->jitter = nla_get_s64(tb[TCA_NETEM_JITTER64]);
1063
1064 if (tb[TCA_NETEM_ECN])
1065 q->ecn = nla_get_u32(tb[TCA_NETEM_ECN]);
1066
1067 if (tb[TCA_NETEM_SLOT])
1068 get_slot(q, tb[TCA_NETEM_SLOT]);
1069
1070 /* capping jitter to the range acceptable by tabledist() */
1071 q->jitter = min_t(s64, abs(q->jitter), INT_MAX);
1072
1073 if (tb[TCA_NETEM_PRNG_SEED])
1074 q->prng.seed = nla_get_u64(tb[TCA_NETEM_PRNG_SEED]);
1075 else
1076 q->prng.seed = get_random_u64();
1077 prandom_seed_state(&q->prng.prng_state, q->prng.seed);
1078
1079 unlock:
1080 sch_tree_unlock(sch);
1081
1082 table_free:
1083 dist_free(delay_dist);
1084 dist_free(slot_dist);
1085 return ret;
1086 }
1087
netem_init(struct Qdisc * sch,struct nlattr * opt,struct netlink_ext_ack * extack)1088 static int netem_init(struct Qdisc *sch, struct nlattr *opt,
1089 struct netlink_ext_ack *extack)
1090 {
1091 struct netem_sched_data *q = qdisc_priv(sch);
1092 int ret;
1093
1094 qdisc_watchdog_init(&q->watchdog, sch);
1095
1096 if (!opt)
1097 return -EINVAL;
1098
1099 q->loss_model = CLG_RANDOM;
1100 ret = netem_change(sch, opt, extack);
1101 if (ret)
1102 pr_info("netem: change failed\n");
1103 return ret;
1104 }
1105
netem_destroy(struct Qdisc * sch)1106 static void netem_destroy(struct Qdisc *sch)
1107 {
1108 struct netem_sched_data *q = qdisc_priv(sch);
1109
1110 qdisc_watchdog_cancel(&q->watchdog);
1111 if (q->qdisc)
1112 qdisc_put(q->qdisc);
1113 dist_free(q->delay_dist);
1114 dist_free(q->slot_dist);
1115 }
1116
dump_loss_model(const struct netem_sched_data * q,struct sk_buff * skb)1117 static int dump_loss_model(const struct netem_sched_data *q,
1118 struct sk_buff *skb)
1119 {
1120 struct nlattr *nest;
1121
1122 nest = nla_nest_start_noflag(skb, TCA_NETEM_LOSS);
1123 if (nest == NULL)
1124 goto nla_put_failure;
1125
1126 switch (q->loss_model) {
1127 case CLG_RANDOM:
1128 /* legacy loss model */
1129 nla_nest_cancel(skb, nest);
1130 return 0; /* no data */
1131
1132 case CLG_4_STATES: {
1133 struct tc_netem_gimodel gi = {
1134 .p13 = q->clg.a1,
1135 .p31 = q->clg.a2,
1136 .p32 = q->clg.a3,
1137 .p14 = q->clg.a4,
1138 .p23 = q->clg.a5,
1139 };
1140
1141 if (nla_put(skb, NETEM_LOSS_GI, sizeof(gi), &gi))
1142 goto nla_put_failure;
1143 break;
1144 }
1145 case CLG_GILB_ELL: {
1146 struct tc_netem_gemodel ge = {
1147 .p = q->clg.a1,
1148 .r = q->clg.a2,
1149 .h = q->clg.a3,
1150 .k1 = q->clg.a4,
1151 };
1152
1153 if (nla_put(skb, NETEM_LOSS_GE, sizeof(ge), &ge))
1154 goto nla_put_failure;
1155 break;
1156 }
1157 }
1158
1159 nla_nest_end(skb, nest);
1160 return 0;
1161
1162 nla_put_failure:
1163 nla_nest_cancel(skb, nest);
1164 return -1;
1165 }
1166
netem_dump(struct Qdisc * sch,struct sk_buff * skb)1167 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
1168 {
1169 const struct netem_sched_data *q = qdisc_priv(sch);
1170 struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb);
1171 struct tc_netem_qopt qopt;
1172 struct tc_netem_corr cor;
1173 struct tc_netem_reorder reorder;
1174 struct tc_netem_corrupt corrupt;
1175 struct tc_netem_rate rate;
1176 struct tc_netem_slot slot;
1177
1178 qopt.latency = min_t(psched_time_t, PSCHED_NS2TICKS(q->latency),
1179 UINT_MAX);
1180 qopt.jitter = min_t(psched_time_t, PSCHED_NS2TICKS(q->jitter),
1181 UINT_MAX);
1182 qopt.limit = q->limit;
1183 qopt.loss = q->loss;
1184 qopt.gap = q->gap;
1185 qopt.duplicate = q->duplicate;
1186 if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
1187 goto nla_put_failure;
1188
1189 if (nla_put(skb, TCA_NETEM_LATENCY64, sizeof(q->latency), &q->latency))
1190 goto nla_put_failure;
1191
1192 if (nla_put(skb, TCA_NETEM_JITTER64, sizeof(q->jitter), &q->jitter))
1193 goto nla_put_failure;
1194
1195 cor.delay_corr = q->delay_cor.rho;
1196 cor.loss_corr = q->loss_cor.rho;
1197 cor.dup_corr = q->dup_cor.rho;
1198 if (nla_put(skb, TCA_NETEM_CORR, sizeof(cor), &cor))
1199 goto nla_put_failure;
1200
1201 reorder.probability = q->reorder;
1202 reorder.correlation = q->reorder_cor.rho;
1203 if (nla_put(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder))
1204 goto nla_put_failure;
1205
1206 corrupt.probability = q->corrupt;
1207 corrupt.correlation = q->corrupt_cor.rho;
1208 if (nla_put(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt))
1209 goto nla_put_failure;
1210
1211 if (q->rate >= (1ULL << 32)) {
1212 if (nla_put_u64_64bit(skb, TCA_NETEM_RATE64, q->rate,
1213 TCA_NETEM_PAD))
1214 goto nla_put_failure;
1215 rate.rate = ~0U;
1216 } else {
1217 rate.rate = q->rate;
1218 }
1219 rate.packet_overhead = q->packet_overhead;
1220 rate.cell_size = q->cell_size;
1221 rate.cell_overhead = q->cell_overhead;
1222 if (nla_put(skb, TCA_NETEM_RATE, sizeof(rate), &rate))
1223 goto nla_put_failure;
1224
1225 if (q->ecn && nla_put_u32(skb, TCA_NETEM_ECN, q->ecn))
1226 goto nla_put_failure;
1227
1228 if (dump_loss_model(q, skb) != 0)
1229 goto nla_put_failure;
1230
1231 if (q->slot_config.min_delay | q->slot_config.max_delay |
1232 q->slot_config.dist_jitter) {
1233 slot = q->slot_config;
1234 if (slot.max_packets == INT_MAX)
1235 slot.max_packets = 0;
1236 if (slot.max_bytes == INT_MAX)
1237 slot.max_bytes = 0;
1238 if (nla_put(skb, TCA_NETEM_SLOT, sizeof(slot), &slot))
1239 goto nla_put_failure;
1240 }
1241
1242 if (nla_put_u64_64bit(skb, TCA_NETEM_PRNG_SEED, q->prng.seed,
1243 TCA_NETEM_PAD))
1244 goto nla_put_failure;
1245
1246 return nla_nest_end(skb, nla);
1247
1248 nla_put_failure:
1249 nlmsg_trim(skb, nla);
1250 return -1;
1251 }
1252
netem_dump_class(struct Qdisc * sch,unsigned long cl,struct sk_buff * skb,struct tcmsg * tcm)1253 static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
1254 struct sk_buff *skb, struct tcmsg *tcm)
1255 {
1256 struct netem_sched_data *q = qdisc_priv(sch);
1257
1258 if (cl != 1 || !q->qdisc) /* only one class */
1259 return -ENOENT;
1260
1261 tcm->tcm_handle |= TC_H_MIN(1);
1262 tcm->tcm_info = q->qdisc->handle;
1263
1264 return 0;
1265 }
1266
netem_graft(struct Qdisc * sch,unsigned long arg,struct Qdisc * new,struct Qdisc ** old,struct netlink_ext_ack * extack)1267 static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1268 struct Qdisc **old, struct netlink_ext_ack *extack)
1269 {
1270 struct netem_sched_data *q = qdisc_priv(sch);
1271
1272 *old = qdisc_replace(sch, new, &q->qdisc);
1273 return 0;
1274 }
1275
netem_leaf(struct Qdisc * sch,unsigned long arg)1276 static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
1277 {
1278 struct netem_sched_data *q = qdisc_priv(sch);
1279 return q->qdisc;
1280 }
1281
netem_find(struct Qdisc * sch,u32 classid)1282 static unsigned long netem_find(struct Qdisc *sch, u32 classid)
1283 {
1284 return 1;
1285 }
1286
netem_walk(struct Qdisc * sch,struct qdisc_walker * walker)1287 static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
1288 {
1289 if (!walker->stop) {
1290 if (!tc_qdisc_stats_dump(sch, 1, walker))
1291 return;
1292 }
1293 }
1294
1295 static const struct Qdisc_class_ops netem_class_ops = {
1296 .graft = netem_graft,
1297 .leaf = netem_leaf,
1298 .find = netem_find,
1299 .walk = netem_walk,
1300 .dump = netem_dump_class,
1301 };
1302
1303 static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
1304 .id = "netem",
1305 .cl_ops = &netem_class_ops,
1306 .priv_size = sizeof(struct netem_sched_data),
1307 .enqueue = netem_enqueue,
1308 .dequeue = netem_dequeue,
1309 .peek = qdisc_peek_dequeued,
1310 .init = netem_init,
1311 .reset = netem_reset,
1312 .destroy = netem_destroy,
1313 .change = netem_change,
1314 .dump = netem_dump,
1315 .owner = THIS_MODULE,
1316 };
1317 MODULE_ALIAS_NET_SCH("netem");
1318
1319
netem_module_init(void)1320 static int __init netem_module_init(void)
1321 {
1322 pr_info("netem: version " VERSION "\n");
1323 return register_qdisc(&netem_qdisc_ops);
1324 }
netem_module_exit(void)1325 static void __exit netem_module_exit(void)
1326 {
1327 unregister_qdisc(&netem_qdisc_ops);
1328 }
1329 module_init(netem_module_init)
1330 module_exit(netem_module_exit)
1331 MODULE_LICENSE("GPL");
1332 MODULE_DESCRIPTION("Network characteristics emulator qdisc");
1333