?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
drm/modes: Fix drm_mode_vrefres() docs
[drm/drm-misc.git] / net / sched / sch_netem.c
blobfe6fed291a7b8697aa204782561cc45568269e8c
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 /* optional qdisc for classful handling (NULL at netem init) */
83 struct Qdisc *qdisc;
84
85 struct qdisc_watchdog watchdog;
86
87 s64 latency;
88 s64 jitter;
89
90 u32 loss;
91 u32 ecn;
92 u32 limit;
93 u32 counter;
94 u32 gap;
95 u32 duplicate;
96 u32 reorder;
97 u32 corrupt;
98 u64 rate;
99 s32 packet_overhead;
100 u32 cell_size;
101 struct reciprocal_value cell_size_reciprocal;
102 s32 cell_overhead;
103
104 struct crndstate {
105 u32 last;
106 u32 rho;
107 } delay_cor, loss_cor, dup_cor, reorder_cor, corrupt_cor;
108
109 struct prng {
110 u64 seed;
111 struct rnd_state prng_state;
112 } prng;
113
114 struct disttable *delay_dist;
115
116 enum {
117 CLG_RANDOM,
118 CLG_4_STATES,
119 CLG_GILB_ELL,
120 } loss_model;
121
122 enum {
123 TX_IN_GAP_PERIOD = 1,
124 TX_IN_BURST_PERIOD,
125 LOST_IN_GAP_PERIOD,
126 LOST_IN_BURST_PERIOD,
127 } _4_state_model;
128
129 enum {
130 GOOD_STATE = 1,
131 BAD_STATE,
132 } GE_state_model;
133
134 /* Correlated Loss Generation models */
135 struct clgstate {
136 /* state of the Markov chain */
137 u8 state;
138
139 /* 4-states and Gilbert-Elliot models */
140 u32 a1; /* p13 for 4-states or p for GE */
141 u32 a2; /* p31 for 4-states or r for GE */
142 u32 a3; /* p32 for 4-states or h for GE */
143 u32 a4; /* p14 for 4-states or 1-k for GE */
144 u32 a5; /* p23 used only in 4-states */
145 } clg;
146
147 struct tc_netem_slot slot_config;
148 struct slotstate {
149 u64 slot_next;
150 s32 packets_left;
151 s32 bytes_left;
152 } slot;
153
154 struct disttable *slot_dist;
155 };
156
157 /* Time stamp put into socket buffer control block
158 * Only valid when skbs are in our internal t(ime)fifo queue.
159 *
160 * As skb->rbnode uses same storage than skb->next, skb->prev and skb->tstamp,
161 * and skb->next & skb->prev are scratch space for a qdisc,
162 * we save skb->tstamp value in skb->cb[] before destroying it.
163 */
164 struct netem_skb_cb {
165 u64 time_to_send;
166 };
167
168 static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb)
169 {
170 /* we assume we can use skb next/prev/tstamp as storage for rb_node */
171 qdisc_cb_private_validate(skb, sizeof(struct netem_skb_cb));
172 return (struct netem_skb_cb *)qdisc_skb_cb(skb)->data;
173 }
174
175 /* init_crandom - initialize correlated random number generator
176 * Use entropy source for initial seed.
177 */
178 static void init_crandom(struct crndstate *state, unsigned long rho)
179 {
180 state->rho = rho;
181 state->last = get_random_u32();
182 }
183
184 /* get_crandom - correlated random number generator
185 * Next number depends on last value.
186 * rho is scaled to avoid floating point.
187 */
188 static u32 get_crandom(struct crndstate *state, struct prng *p)
189 {
190 u64 value, rho;
191 unsigned long answer;
192 struct rnd_state *s = &p->prng_state;
193
194 if (!state || state->rho == 0) /* no correlation */
195 return prandom_u32_state(s);
196
197 value = prandom_u32_state(s);
198 rho = (u64)state->rho + 1;
199 answer = (value * ((1ull<<32) - rho) + state->last * rho) >> 32;
200 state->last = answer;
201 return answer;
202 }
203
204 /* loss_4state - 4-state model loss generator
205 * Generates losses according to the 4-state Markov chain adopted in
206 * the GI (General and Intuitive) loss model.
207 */
208 static bool loss_4state(struct netem_sched_data *q)
209 {
210 struct clgstate *clg = &q->clg;
211 u32 rnd = prandom_u32_state(&q->prng.prng_state);
212
213 /*
214 * Makes a comparison between rnd and the transition
215 * probabilities outgoing from the current state, then decides the
216 * next state and if the next packet has to be transmitted or lost.
217 * The four states correspond to:
218 * TX_IN_GAP_PERIOD => successfully transmitted packets within a gap period
219 * LOST_IN_GAP_PERIOD => isolated losses within a gap period
220 * LOST_IN_BURST_PERIOD => lost packets within a burst period
221 * TX_IN_BURST_PERIOD => successfully transmitted packets within a burst period
222 */
223 switch (clg->state) {
224 case TX_IN_GAP_PERIOD:
225 if (rnd < clg->a4) {
226 clg->state = LOST_IN_GAP_PERIOD;
227 return true;
228 } else if (clg->a4 < rnd && rnd < clg->a1 + clg->a4) {
229 clg->state = LOST_IN_BURST_PERIOD;
230 return true;
231 } else if (clg->a1 + clg->a4 < rnd) {
232 clg->state = TX_IN_GAP_PERIOD;
233 }
234
235 break;
236 case TX_IN_BURST_PERIOD:
237 if (rnd < clg->a5) {
238 clg->state = LOST_IN_BURST_PERIOD;
239 return true;
240 } else {
241 clg->state = TX_IN_BURST_PERIOD;
242 }
243
244 break;
245 case LOST_IN_BURST_PERIOD:
246 if (rnd < clg->a3)
247 clg->state = TX_IN_BURST_PERIOD;
248 else if (clg->a3 < rnd && rnd < clg->a2 + clg->a3) {
249 clg->state = TX_IN_GAP_PERIOD;
250 } else if (clg->a2 + clg->a3 < rnd) {
251 clg->state = LOST_IN_BURST_PERIOD;
252 return true;
253 }
254 break;
255 case LOST_IN_GAP_PERIOD:
256 clg->state = TX_IN_GAP_PERIOD;
257 break;
258 }
259
260 return false;
261 }
262
263 /* loss_gilb_ell - Gilbert-Elliot model loss generator
264 * Generates losses according to the Gilbert-Elliot loss model or
265 * its special cases (Gilbert or Simple Gilbert)
266 *
267 * Makes a comparison between random number and the transition
268 * probabilities outgoing from the current state, then decides the
269 * next state. A second random number is extracted and the comparison
270 * with the loss probability of the current state decides if the next
271 * packet will be transmitted or lost.
272 */
273 static bool loss_gilb_ell(struct netem_sched_data *q)
274 {
275 struct clgstate *clg = &q->clg;
276 struct rnd_state *s = &q->prng.prng_state;
277
278 switch (clg->state) {
279 case GOOD_STATE:
280 if (prandom_u32_state(s) < clg->a1)
281 clg->state = BAD_STATE;
282 if (prandom_u32_state(s) < clg->a4)
283 return true;
284 break;
285 case BAD_STATE:
286 if (prandom_u32_state(s) < clg->a2)
287 clg->state = GOOD_STATE;
288 if (prandom_u32_state(s) > clg->a3)
289 return true;
290 }
291
292 return false;
293 }
294
295 static bool loss_event(struct netem_sched_data *q)
296 {
297 switch (q->loss_model) {
298 case CLG_RANDOM:
299 /* Random packet drop 0 => none, ~0 => all */
300 return q->loss && q->loss >= get_crandom(&q->loss_cor, &q->prng);
301
302 case CLG_4_STATES:
303 /* 4state loss model algorithm (used also for GI model)
304 * Extracts a value from the markov 4 state loss generator,
305 * if it is 1 drops a packet and if needed writes the event in
306 * the kernel logs
307 */
308 return loss_4state(q);
309
310 case CLG_GILB_ELL:
311 /* Gilbert-Elliot loss model algorithm
312 * Extracts a value from the Gilbert-Elliot loss generator,
313 * if it is 1 drops a packet and if needed writes the event in
314 * the kernel logs
315 */
316 return loss_gilb_ell(q);
317 }
318
319 return false; /* not reached */
320 }
321
322
323 /* tabledist - return a pseudo-randomly distributed value with mean mu and
324 * std deviation sigma. Uses table lookup to approximate the desired
325 * distribution, and a uniformly-distributed pseudo-random source.
326 */
327 static s64 tabledist(s64 mu, s32 sigma,
328 struct crndstate *state,
329 struct prng *prng,
330 const struct disttable *dist)
331 {
332 s64 x;
333 long t;
334 u32 rnd;
335
336 if (sigma == 0)
337 return mu;
338
339 rnd = get_crandom(state, prng);
340
341 /* default uniform distribution */
342 if (dist == NULL)
343 return ((rnd % (2 * (u32)sigma)) + mu) - sigma;
344
345 t = dist->table[rnd % dist->size];
346 x = (sigma % NETEM_DIST_SCALE) * t;
347 if (x >= 0)
348 x += NETEM_DIST_SCALE/2;
349 else
350 x -= NETEM_DIST_SCALE/2;
351
352 return x / NETEM_DIST_SCALE + (sigma / NETEM_DIST_SCALE) * t + mu;
353 }
354
355 static u64 packet_time_ns(u64 len, const struct netem_sched_data *q)
356 {
357 len += q->packet_overhead;
358
359 if (q->cell_size) {
360 u32 cells = reciprocal_divide(len, q->cell_size_reciprocal);
361
362 if (len > cells * q->cell_size) /* extra cell needed for remainder */
363 cells++;
364 len = cells * (q->cell_size + q->cell_overhead);
365 }
366
367 return div64_u64(len * NSEC_PER_SEC, q->rate);
368 }
369
370 static void tfifo_reset(struct Qdisc *sch)
371 {
372 struct netem_sched_data *q = qdisc_priv(sch);
373 struct rb_node *p = rb_first(&q->t_root);
374
375 while (p) {
376 struct sk_buff *skb = rb_to_skb(p);
377
378 p = rb_next(p);
379 rb_erase(&skb->rbnode, &q->t_root);
380 rtnl_kfree_skbs(skb, skb);
381 }
382
383 rtnl_kfree_skbs(q->t_head, q->t_tail);
384 q->t_head = NULL;
385 q->t_tail = NULL;
386 }
387
388 static void tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
389 {
390 struct netem_sched_data *q = qdisc_priv(sch);
391 u64 tnext = netem_skb_cb(nskb)->time_to_send;
392
393 if (!q->t_tail || tnext >= netem_skb_cb(q->t_tail)->time_to_send) {
394 if (q->t_tail)
395 q->t_tail->next = nskb;
396 else
397 q->t_head = nskb;
398 q->t_tail = nskb;
399 } else {
400 struct rb_node **p = &q->t_root.rb_node, *parent = NULL;
401
402 while (*p) {
403 struct sk_buff *skb;
404
405 parent = *p;
406 skb = rb_to_skb(parent);
407 if (tnext >= netem_skb_cb(skb)->time_to_send)
408 p = &parent->rb_right;
409 else
410 p = &parent->rb_left;
411 }
412 rb_link_node(&nskb->rbnode, parent, p);
413 rb_insert_color(&nskb->rbnode, &q->t_root);
414 }
415 sch->q.qlen++;
416 }
417
418 /* netem can't properly corrupt a megapacket (like we get from GSO), so instead
419 * when we statistically choose to corrupt one, we instead segment it, returning
420 * the first packet to be corrupted, and re-enqueue the remaining frames
421 */
422 static struct sk_buff *netem_segment(struct sk_buff *skb, struct Qdisc *sch,
423 struct sk_buff **to_free)
424 {
425 struct sk_buff *segs;
426 netdev_features_t features = netif_skb_features(skb);
427
428 segs = skb_gso_segment(skb, features & ~NETIF_F_GSO_MASK);
429
430 if (IS_ERR_OR_NULL(segs)) {
431 qdisc_drop(skb, sch, to_free);
432 return NULL;
433 }
434 consume_skb(skb);
435 return segs;
436 }
437
438 /*
439 * Insert one skb into qdisc.
440 * Note: parent depends on return value to account for queue length.
441 * NET_XMIT_DROP: queue length didn't change.
442 * NET_XMIT_SUCCESS: one skb was queued.
443 */
444 static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch,
445 struct sk_buff **to_free)
446 {
447 struct netem_sched_data *q = qdisc_priv(sch);
448 /* We don't fill cb now as skb_unshare() may invalidate it */
449 struct netem_skb_cb *cb;
450 struct sk_buff *skb2 = NULL;
451 struct sk_buff *segs = NULL;
452 unsigned int prev_len = qdisc_pkt_len(skb);
453 int count = 1;
454
455 /* Do not fool qdisc_drop_all() */
456 skb->prev = NULL;
457
458 /* Random duplication */
459 if (q->duplicate && q->duplicate >= get_crandom(&q->dup_cor, &q->prng))
460 ++count;
461
462 /* Drop packet? */
463 if (loss_event(q)) {
464 if (q->ecn && INET_ECN_set_ce(skb))
465 qdisc_qstats_drop(sch); /* mark packet */
466 else
467 --count;
468 }
469 if (count == 0) {
470 qdisc_qstats_drop(sch);
471 __qdisc_drop(skb, to_free);
472 return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
473 }
474
475 /* If a delay is expected, orphan the skb. (orphaning usually takes
476 * place at TX completion time, so _before_ the link transit delay)
477 */
478 if (q->latency || q->jitter || q->rate)
479 skb_orphan_partial(skb);
480
481 /*
482 * If we need to duplicate packet, then clone it before
483 * original is modified.
484 */
485 if (count > 1)
486 skb2 = skb_clone(skb, GFP_ATOMIC);
487
488 /*
489 * Randomized packet corruption.
490 * Make copy if needed since we are modifying
491 * If packet is going to be hardware checksummed, then
492 * do it now in software before we mangle it.
493 */
494 if (q->corrupt && q->corrupt >= get_crandom(&q->corrupt_cor, &q->prng)) {
495 if (skb_is_gso(skb)) {
496 skb = netem_segment(skb, sch, to_free);
497 if (!skb)
498 goto finish_segs;
499
500 segs = skb->next;
501 skb_mark_not_on_list(skb);
502 qdisc_skb_cb(skb)->pkt_len = skb->len;
503 }
504
505 skb = skb_unshare(skb, GFP_ATOMIC);
506 if (unlikely(!skb)) {
507 qdisc_qstats_drop(sch);
508 goto finish_segs;
509 }
510 if (skb->ip_summed == CHECKSUM_PARTIAL &&
511 skb_checksum_help(skb)) {
512 qdisc_drop(skb, sch, to_free);
513 skb = NULL;
514 goto finish_segs;
515 }
516
517 skb->data[get_random_u32_below(skb_headlen(skb))] ^=
518 1<<get_random_u32_below(8);
519 }
520
521 if (unlikely(sch->q.qlen >= sch->limit)) {
522 /* re-link segs, so that qdisc_drop_all() frees them all */
523 skb->next = segs;
524 qdisc_drop_all(skb, sch, to_free);
525 if (skb2)
526 __qdisc_drop(skb2, to_free);
527 return NET_XMIT_DROP;
528 }
529
530 /*
531 * If doing duplication then re-insert at top of the
532 * qdisc tree, since parent queuer expects that only one
533 * skb will be queued.
534 */
535 if (skb2) {
536 struct Qdisc *rootq = qdisc_root_bh(sch);
537 u32 dupsave = q->duplicate; /* prevent duplicating a dup... */
538
539 q->duplicate = 0;
540 rootq->enqueue(skb2, rootq, to_free);
541 q->duplicate = dupsave;
542 skb2 = NULL;
543 }
544
545 qdisc_qstats_backlog_inc(sch, skb);
546
547 cb = netem_skb_cb(skb);
548 if (q->gap == 0 || /* not doing reordering */
549 q->counter < q->gap - 1 || /* inside last reordering gap */
550 q->reorder < get_crandom(&q->reorder_cor, &q->prng)) {
551 u64 now;
552 s64 delay;
553
554 delay = tabledist(q->latency, q->jitter,
555 &q->delay_cor, &q->prng, q->delay_dist);
556
557 now = ktime_get_ns();
558
559 if (q->rate) {
560 struct netem_skb_cb *last = NULL;
561
562 if (sch->q.tail)
563 last = netem_skb_cb(sch->q.tail);
564 if (q->t_root.rb_node) {
565 struct sk_buff *t_skb;
566 struct netem_skb_cb *t_last;
567
568 t_skb = skb_rb_last(&q->t_root);
569 t_last = netem_skb_cb(t_skb);
570 if (!last ||
571 t_last->time_to_send > last->time_to_send)
572 last = t_last;
573 }
574 if (q->t_tail) {
575 struct netem_skb_cb *t_last =
576 netem_skb_cb(q->t_tail);
577
578 if (!last ||
579 t_last->time_to_send > last->time_to_send)
580 last = t_last;
581 }
582
583 if (last) {
584 /*
585 * Last packet in queue is reference point (now),
586 * calculate this time bonus and subtract
587 * from delay.
588 */
589 delay -= last->time_to_send - now;
590 delay = max_t(s64, 0, delay);
591 now = last->time_to_send;
592 }
593
594 delay += packet_time_ns(qdisc_pkt_len(skb), q);
595 }
596
597 cb->time_to_send = now + delay;
598 ++q->counter;
599 tfifo_enqueue(skb, sch);
600 } else {
601 /*
602 * Do re-ordering by putting one out of N packets at the front
603 * of the queue.
604 */
605 cb->time_to_send = ktime_get_ns();
606 q->counter = 0;
607
608 __qdisc_enqueue_head(skb, &sch->q);
609 sch->qstats.requeues++;
610 }
611
612 finish_segs:
613 if (skb2)
614 __qdisc_drop(skb2, to_free);
615
616 if (segs) {
617 unsigned int len, last_len;
618 int rc, nb;
619
620 len = skb ? skb->len : 0;
621 nb = skb ? 1 : 0;
622
623 while (segs) {
624 skb2 = segs->next;
625 skb_mark_not_on_list(segs);
626 qdisc_skb_cb(segs)->pkt_len = segs->len;
627 last_len = segs->len;
628 rc = qdisc_enqueue(segs, sch, to_free);
629 if (rc != NET_XMIT_SUCCESS) {
630 if (net_xmit_drop_count(rc))
631 qdisc_qstats_drop(sch);
632 } else {
633 nb++;
634 len += last_len;
635 }
636 segs = skb2;
637 }
638 /* Parent qdiscs accounted for 1 skb of size @prev_len */
639 qdisc_tree_reduce_backlog(sch, -(nb - 1), -(len - prev_len));
640 } else if (!skb) {
641 return NET_XMIT_DROP;
642 }
643 return NET_XMIT_SUCCESS;
644 }
645
646 /* Delay the next round with a new future slot with a
647 * correct number of bytes and packets.
648 */
649
650 static void get_slot_next(struct netem_sched_data *q, u64 now)
651 {
652 s64 next_delay;
653
654 if (!q->slot_dist)
655 next_delay = q->slot_config.min_delay +
656 (get_random_u32() *
657 (q->slot_config.max_delay -
658 q->slot_config.min_delay) >> 32);
659 else
660 next_delay = tabledist(q->slot_config.dist_delay,
661 (s32)(q->slot_config.dist_jitter),
662 NULL, &q->prng, q->slot_dist);
663
664 q->slot.slot_next = now + next_delay;
665 q->slot.packets_left = q->slot_config.max_packets;
666 q->slot.bytes_left = q->slot_config.max_bytes;
667 }
668
669 static struct sk_buff *netem_peek(struct netem_sched_data *q)
670 {
671 struct sk_buff *skb = skb_rb_first(&q->t_root);
672 u64 t1, t2;
673
674 if (!skb)
675 return q->t_head;
676 if (!q->t_head)
677 return skb;
678
679 t1 = netem_skb_cb(skb)->time_to_send;
680 t2 = netem_skb_cb(q->t_head)->time_to_send;
681 if (t1 < t2)
682 return skb;
683 return q->t_head;
684 }
685
686 static void netem_erase_head(struct netem_sched_data *q, struct sk_buff *skb)
687 {
688 if (skb == q->t_head) {
689 q->t_head = skb->next;
690 if (!q->t_head)
691 q->t_tail = NULL;
692 } else {
693 rb_erase(&skb->rbnode, &q->t_root);
694 }
695 }
696
697 static struct sk_buff *netem_dequeue(struct Qdisc *sch)
698 {
699 struct netem_sched_data *q = qdisc_priv(sch);
700 struct sk_buff *skb;
701
702 tfifo_dequeue:
703 skb = __qdisc_dequeue_head(&sch->q);
704 if (skb) {
705 qdisc_qstats_backlog_dec(sch, skb);
706 deliver:
707 qdisc_bstats_update(sch, skb);
708 return skb;
709 }
710 skb = netem_peek(q);
711 if (skb) {
712 u64 time_to_send;
713 u64 now = ktime_get_ns();
714
715 /* if more time remaining? */
716 time_to_send = netem_skb_cb(skb)->time_to_send;
717 if (q->slot.slot_next && q->slot.slot_next < time_to_send)
718 get_slot_next(q, now);
719
720 if (time_to_send <= now && q->slot.slot_next <= now) {
721 netem_erase_head(q, skb);
722 sch->q.qlen--;
723 qdisc_qstats_backlog_dec(sch, skb);
724 skb->next = NULL;
725 skb->prev = NULL;
726 /* skb->dev shares skb->rbnode area,
727 * we need to restore its value.
728 */
729 skb->dev = qdisc_dev(sch);
730
731 if (q->slot.slot_next) {
732 q->slot.packets_left--;
733 q->slot.bytes_left -= qdisc_pkt_len(skb);
734 if (q->slot.packets_left <= 0 ||
735 q->slot.bytes_left <= 0)
736 get_slot_next(q, now);
737 }
738
739 if (q->qdisc) {
740 unsigned int pkt_len = qdisc_pkt_len(skb);
741 struct sk_buff *to_free = NULL;
742 int err;
743
744 err = qdisc_enqueue(skb, q->qdisc, &to_free);
745 kfree_skb_list(to_free);
746 if (err != NET_XMIT_SUCCESS) {
747 if (net_xmit_drop_count(err))
748 qdisc_qstats_drop(sch);
749 qdisc_tree_reduce_backlog(sch, 1, pkt_len);
750 }
751 goto tfifo_dequeue;
752 }
753 goto deliver;
754 }
755
756 if (q->qdisc) {
757 skb = q->qdisc->ops->dequeue(q->qdisc);
758 if (skb)
759 goto deliver;
760 }
761
762 qdisc_watchdog_schedule_ns(&q->watchdog,
763 max(time_to_send,
764 q->slot.slot_next));
765 }
766
767 if (q->qdisc) {
768 skb = q->qdisc->ops->dequeue(q->qdisc);
769 if (skb)
770 goto deliver;
771 }
772 return NULL;
773 }
774
775 static void netem_reset(struct Qdisc *sch)
776 {
777 struct netem_sched_data *q = qdisc_priv(sch);
778
779 qdisc_reset_queue(sch);
780 tfifo_reset(sch);
781 if (q->qdisc)
782 qdisc_reset(q->qdisc);
783 qdisc_watchdog_cancel(&q->watchdog);
784 }
785
786 static void dist_free(struct disttable *d)
787 {
788 kvfree(d);
789 }
790
791 /*
792 * Distribution data is a variable size payload containing
793 * signed 16 bit values.
794 */
795
796 static int get_dist_table(struct disttable **tbl, const struct nlattr *attr)
797 {
798 size_t n = nla_len(attr)/sizeof(__s16);
799 const __s16 *data = nla_data(attr);
800 struct disttable *d;
801 int i;
802
803 if (!n || n > NETEM_DIST_MAX)
804 return -EINVAL;
805
806 d = kvmalloc(struct_size(d, table, n), GFP_KERNEL);
807 if (!d)
808 return -ENOMEM;
809
810 d->size = n;
811 for (i = 0; i < n; i++)
812 d->table[i] = data[i];
813
814 *tbl = d;
815 return 0;
816 }
817
818 static void get_slot(struct netem_sched_data *q, const struct nlattr *attr)
819 {
820 const struct tc_netem_slot *c = nla_data(attr);
821
822 q->slot_config = *c;
823 if (q->slot_config.max_packets == 0)
824 q->slot_config.max_packets = INT_MAX;
825 if (q->slot_config.max_bytes == 0)
826 q->slot_config.max_bytes = INT_MAX;
827
828 /* capping dist_jitter to the range acceptable by tabledist() */
829 q->slot_config.dist_jitter = min_t(__s64, INT_MAX, abs(q->slot_config.dist_jitter));
830
831 q->slot.packets_left = q->slot_config.max_packets;
832 q->slot.bytes_left = q->slot_config.max_bytes;
833 if (q->slot_config.min_delay | q->slot_config.max_delay |
834 q->slot_config.dist_jitter)
835 q->slot.slot_next = ktime_get_ns();
836 else
837 q->slot.slot_next = 0;
838 }
839
840 static void get_correlation(struct netem_sched_data *q, const struct nlattr *attr)
841 {
842 const struct tc_netem_corr *c = nla_data(attr);
843
844 init_crandom(&q->delay_cor, c->delay_corr);
845 init_crandom(&q->loss_cor, c->loss_corr);
846 init_crandom(&q->dup_cor, c->dup_corr);
847 }
848
849 static void get_reorder(struct netem_sched_data *q, const struct nlattr *attr)
850 {
851 const struct tc_netem_reorder *r = nla_data(attr);
852
853 q->reorder = r->probability;
854 init_crandom(&q->reorder_cor, r->correlation);
855 }
856
857 static void get_corrupt(struct netem_sched_data *q, const struct nlattr *attr)
858 {
859 const struct tc_netem_corrupt *r = nla_data(attr);
860
861 q->corrupt = r->probability;
862 init_crandom(&q->corrupt_cor, r->correlation);
863 }
864
865 static void get_rate(struct netem_sched_data *q, const struct nlattr *attr)
866 {
867 const struct tc_netem_rate *r = nla_data(attr);
868
869 q->rate = r->rate;
870 q->packet_overhead = r->packet_overhead;
871 q->cell_size = r->cell_size;
872 q->cell_overhead = r->cell_overhead;
873 if (q->cell_size)
874 q->cell_size_reciprocal = reciprocal_value(q->cell_size);
875 else
876 q->cell_size_reciprocal = (struct reciprocal_value) { 0 };
877 }
878
879 static int get_loss_clg(struct netem_sched_data *q, const struct nlattr *attr)
880 {
881 const struct nlattr *la;
882 int rem;
883
884 nla_for_each_nested(la, attr, rem) {
885 u16 type = nla_type(la);
886
887 switch (type) {
888 case NETEM_LOSS_GI: {
889 const struct tc_netem_gimodel *gi = nla_data(la);
890
891 if (nla_len(la) < sizeof(struct tc_netem_gimodel)) {
892 pr_info("netem: incorrect gi model size\n");
893 return -EINVAL;
894 }
895
896 q->loss_model = CLG_4_STATES;
897
898 q->clg.state = TX_IN_GAP_PERIOD;
899 q->clg.a1 = gi->p13;
900 q->clg.a2 = gi->p31;
901 q->clg.a3 = gi->p32;
902 q->clg.a4 = gi->p14;
903 q->clg.a5 = gi->p23;
904 break;
905 }
906
907 case NETEM_LOSS_GE: {
908 const struct tc_netem_gemodel *ge = nla_data(la);
909
910 if (nla_len(la) < sizeof(struct tc_netem_gemodel)) {
911 pr_info("netem: incorrect ge model size\n");
912 return -EINVAL;
913 }
914
915 q->loss_model = CLG_GILB_ELL;
916 q->clg.state = GOOD_STATE;
917 q->clg.a1 = ge->p;
918 q->clg.a2 = ge->r;
919 q->clg.a3 = ge->h;
920 q->clg.a4 = ge->k1;
921 break;
922 }
923
924 default:
925 pr_info("netem: unknown loss type %u\n", type);
926 return -EINVAL;
927 }
928 }
929
930 return 0;
931 }
932
933 static const struct nla_policy netem_policy[TCA_NETEM_MAX + 1] = {
934 [TCA_NETEM_CORR] = { .len = sizeof(struct tc_netem_corr) },
935 [TCA_NETEM_REORDER] = { .len = sizeof(struct tc_netem_reorder) },
936 [TCA_NETEM_CORRUPT] = { .len = sizeof(struct tc_netem_corrupt) },
937 [TCA_NETEM_RATE] = { .len = sizeof(struct tc_netem_rate) },
938 [TCA_NETEM_LOSS] = { .type = NLA_NESTED },
939 [TCA_NETEM_ECN] = { .type = NLA_U32 },
940 [TCA_NETEM_RATE64] = { .type = NLA_U64 },
941 [TCA_NETEM_LATENCY64] = { .type = NLA_S64 },
942 [TCA_NETEM_JITTER64] = { .type = NLA_S64 },
943 [TCA_NETEM_SLOT] = { .len = sizeof(struct tc_netem_slot) },
944 [TCA_NETEM_PRNG_SEED] = { .type = NLA_U64 },
945 };
946
947 static int parse_attr(struct nlattr *tb[], int maxtype, struct nlattr *nla,
948 const struct nla_policy *policy, int len)
949 {
950 int nested_len = nla_len(nla) - NLA_ALIGN(len);
951
952 if (nested_len < 0) {
953 pr_info("netem: invalid attributes len %d\n", nested_len);
954 return -EINVAL;
955 }
956
957 if (nested_len >= nla_attr_size(0))
958 return nla_parse_deprecated(tb, maxtype,
959 nla_data(nla) + NLA_ALIGN(len),
960 nested_len, policy, NULL);
961
962 memset(tb, 0, sizeof(struct nlattr *) * (maxtype + 1));
963 return 0;
964 }
965
966 /* Parse netlink message to set options */
967 static int netem_change(struct Qdisc *sch, struct nlattr *opt,
968 struct netlink_ext_ack *extack)
969 {
970 struct netem_sched_data *q = qdisc_priv(sch);
971 struct nlattr *tb[TCA_NETEM_MAX + 1];
972 struct disttable *delay_dist = NULL;
973 struct disttable *slot_dist = NULL;
974 struct tc_netem_qopt *qopt;
975 struct clgstate old_clg;
976 int old_loss_model = CLG_RANDOM;
977 int ret;
978
979 qopt = nla_data(opt);
980 ret = parse_attr(tb, TCA_NETEM_MAX, opt, netem_policy, sizeof(*qopt));
981 if (ret < 0)
982 return ret;
983
984 if (tb[TCA_NETEM_DELAY_DIST]) {
985 ret = get_dist_table(&delay_dist, tb[TCA_NETEM_DELAY_DIST]);
986 if (ret)
987 goto table_free;
988 }
989
990 if (tb[TCA_NETEM_SLOT_DIST]) {
991 ret = get_dist_table(&slot_dist, tb[TCA_NETEM_SLOT_DIST]);
992 if (ret)
993 goto table_free;
994 }
995
996 sch_tree_lock(sch);
997 /* backup q->clg and q->loss_model */
998 old_clg = q->clg;
999 old_loss_model = q->loss_model;
1000
1001 if (tb[TCA_NETEM_LOSS]) {
1002 ret = get_loss_clg(q, tb[TCA_NETEM_LOSS]);
1003 if (ret) {
1004 q->loss_model = old_loss_model;
1005 q->clg = old_clg;
1006 goto unlock;
1007 }
1008 } else {
1009 q->loss_model = CLG_RANDOM;
1010 }
1011
1012 if (delay_dist)
1013 swap(q->delay_dist, delay_dist);
1014 if (slot_dist)
1015 swap(q->slot_dist, slot_dist);
1016 sch->limit = qopt->limit;
1017
1018 q->latency = PSCHED_TICKS2NS(qopt->latency);
1019 q->jitter = PSCHED_TICKS2NS(qopt->jitter);
1020 q->limit = qopt->limit;
1021 q->gap = qopt->gap;
1022 q->counter = 0;
1023 q->loss = qopt->loss;
1024 q->duplicate = qopt->duplicate;
1025
1026 /* for compatibility with earlier versions.
1027 * if gap is set, need to assume 100% probability
1028 */
1029 if (q->gap)
1030 q->reorder = ~0;
1031
1032 if (tb[TCA_NETEM_CORR])
1033 get_correlation(q, tb[TCA_NETEM_CORR]);
1034
1035 if (tb[TCA_NETEM_REORDER])
1036 get_reorder(q, tb[TCA_NETEM_REORDER]);
1037
1038 if (tb[TCA_NETEM_CORRUPT])
1039 get_corrupt(q, tb[TCA_NETEM_CORRUPT]);
1040
1041 if (tb[TCA_NETEM_RATE])
1042 get_rate(q, tb[TCA_NETEM_RATE]);
1043
1044 if (tb[TCA_NETEM_RATE64])
1045 q->rate = max_t(u64, q->rate,
1046 nla_get_u64(tb[TCA_NETEM_RATE64]));
1047
1048 if (tb[TCA_NETEM_LATENCY64])
1049 q->latency = nla_get_s64(tb[TCA_NETEM_LATENCY64]);
1050
1051 if (tb[TCA_NETEM_JITTER64])
1052 q->jitter = nla_get_s64(tb[TCA_NETEM_JITTER64]);
1053
1054 if (tb[TCA_NETEM_ECN])
1055 q->ecn = nla_get_u32(tb[TCA_NETEM_ECN]);
1056
1057 if (tb[TCA_NETEM_SLOT])
1058 get_slot(q, tb[TCA_NETEM_SLOT]);
1059
1060 /* capping jitter to the range acceptable by tabledist() */
1061 q->jitter = min_t(s64, abs(q->jitter), INT_MAX);
1062
1063 if (tb[TCA_NETEM_PRNG_SEED])
1064 q->prng.seed = nla_get_u64(tb[TCA_NETEM_PRNG_SEED]);
1065 else
1066 q->prng.seed = get_random_u64();
1067 prandom_seed_state(&q->prng.prng_state, q->prng.seed);
1068
1069 unlock:
1070 sch_tree_unlock(sch);
1071
1072 table_free:
1073 dist_free(delay_dist);
1074 dist_free(slot_dist);
1075 return ret;
1076 }
1077
1078 static int netem_init(struct Qdisc *sch, struct nlattr *opt,
1079 struct netlink_ext_ack *extack)
1080 {
1081 struct netem_sched_data *q = qdisc_priv(sch);
1082 int ret;
1083
1084 qdisc_watchdog_init(&q->watchdog, sch);
1085
1086 if (!opt)
1087 return -EINVAL;
1088
1089 q->loss_model = CLG_RANDOM;
1090 ret = netem_change(sch, opt, extack);
1091 if (ret)
1092 pr_info("netem: change failed\n");
1093 return ret;
1094 }
1095
1096 static void netem_destroy(struct Qdisc *sch)
1097 {
1098 struct netem_sched_data *q = qdisc_priv(sch);
1099
1100 qdisc_watchdog_cancel(&q->watchdog);
1101 if (q->qdisc)
1102 qdisc_put(q->qdisc);
1103 dist_free(q->delay_dist);
1104 dist_free(q->slot_dist);
1105 }
1106
1107 static int dump_loss_model(const struct netem_sched_data *q,
1108 struct sk_buff *skb)
1109 {
1110 struct nlattr *nest;
1111
1112 nest = nla_nest_start_noflag(skb, TCA_NETEM_LOSS);
1113 if (nest == NULL)
1114 goto nla_put_failure;
1115
1116 switch (q->loss_model) {
1117 case CLG_RANDOM:
1118 /* legacy loss model */
1119 nla_nest_cancel(skb, nest);
1120 return 0; /* no data */
1121
1122 case CLG_4_STATES: {
1123 struct tc_netem_gimodel gi = {
1124 .p13 = q->clg.a1,
1125 .p31 = q->clg.a2,
1126 .p32 = q->clg.a3,
1127 .p14 = q->clg.a4,
1128 .p23 = q->clg.a5,
1129 };
1130
1131 if (nla_put(skb, NETEM_LOSS_GI, sizeof(gi), &gi))
1132 goto nla_put_failure;
1133 break;
1134 }
1135 case CLG_GILB_ELL: {
1136 struct tc_netem_gemodel ge = {
1137 .p = q->clg.a1,
1138 .r = q->clg.a2,
1139 .h = q->clg.a3,
1140 .k1 = q->clg.a4,
1141 };
1142
1143 if (nla_put(skb, NETEM_LOSS_GE, sizeof(ge), &ge))
1144 goto nla_put_failure;
1145 break;
1146 }
1147 }
1148
1149 nla_nest_end(skb, nest);
1150 return 0;
1151
1152 nla_put_failure:
1153 nla_nest_cancel(skb, nest);
1154 return -1;
1155 }
1156
1157 static int netem_dump(struct Qdisc *sch, struct sk_buff *skb)
1158 {
1159 const struct netem_sched_data *q = qdisc_priv(sch);
1160 struct nlattr *nla = (struct nlattr *) skb_tail_pointer(skb);
1161 struct tc_netem_qopt qopt;
1162 struct tc_netem_corr cor;
1163 struct tc_netem_reorder reorder;
1164 struct tc_netem_corrupt corrupt;
1165 struct tc_netem_rate rate;
1166 struct tc_netem_slot slot;
1167
1168 qopt.latency = min_t(psched_time_t, PSCHED_NS2TICKS(q->latency),
1169 UINT_MAX);
1170 qopt.jitter = min_t(psched_time_t, PSCHED_NS2TICKS(q->jitter),
1171 UINT_MAX);
1172 qopt.limit = q->limit;
1173 qopt.loss = q->loss;
1174 qopt.gap = q->gap;
1175 qopt.duplicate = q->duplicate;
1176 if (nla_put(skb, TCA_OPTIONS, sizeof(qopt), &qopt))
1177 goto nla_put_failure;
1178
1179 if (nla_put(skb, TCA_NETEM_LATENCY64, sizeof(q->latency), &q->latency))
1180 goto nla_put_failure;
1181
1182 if (nla_put(skb, TCA_NETEM_JITTER64, sizeof(q->jitter), &q->jitter))
1183 goto nla_put_failure;
1184
1185 cor.delay_corr = q->delay_cor.rho;
1186 cor.loss_corr = q->loss_cor.rho;
1187 cor.dup_corr = q->dup_cor.rho;
1188 if (nla_put(skb, TCA_NETEM_CORR, sizeof(cor), &cor))
1189 goto nla_put_failure;
1190
1191 reorder.probability = q->reorder;
1192 reorder.correlation = q->reorder_cor.rho;
1193 if (nla_put(skb, TCA_NETEM_REORDER, sizeof(reorder), &reorder))
1194 goto nla_put_failure;
1195
1196 corrupt.probability = q->corrupt;
1197 corrupt.correlation = q->corrupt_cor.rho;
1198 if (nla_put(skb, TCA_NETEM_CORRUPT, sizeof(corrupt), &corrupt))
1199 goto nla_put_failure;
1200
1201 if (q->rate >= (1ULL << 32)) {
1202 if (nla_put_u64_64bit(skb, TCA_NETEM_RATE64, q->rate,
1203 TCA_NETEM_PAD))
1204 goto nla_put_failure;
1205 rate.rate = ~0U;
1206 } else {
1207 rate.rate = q->rate;
1208 }
1209 rate.packet_overhead = q->packet_overhead;
1210 rate.cell_size = q->cell_size;
1211 rate.cell_overhead = q->cell_overhead;
1212 if (nla_put(skb, TCA_NETEM_RATE, sizeof(rate), &rate))
1213 goto nla_put_failure;
1214
1215 if (q->ecn && nla_put_u32(skb, TCA_NETEM_ECN, q->ecn))
1216 goto nla_put_failure;
1217
1218 if (dump_loss_model(q, skb) != 0)
1219 goto nla_put_failure;
1220
1221 if (q->slot_config.min_delay | q->slot_config.max_delay |
1222 q->slot_config.dist_jitter) {
1223 slot = q->slot_config;
1224 if (slot.max_packets == INT_MAX)
1225 slot.max_packets = 0;
1226 if (slot.max_bytes == INT_MAX)
1227 slot.max_bytes = 0;
1228 if (nla_put(skb, TCA_NETEM_SLOT, sizeof(slot), &slot))
1229 goto nla_put_failure;
1230 }
1231
1232 if (nla_put_u64_64bit(skb, TCA_NETEM_PRNG_SEED, q->prng.seed,
1233 TCA_NETEM_PAD))
1234 goto nla_put_failure;
1235
1236 return nla_nest_end(skb, nla);
1237
1238 nla_put_failure:
1239 nlmsg_trim(skb, nla);
1240 return -1;
1241 }
1242
1243 static int netem_dump_class(struct Qdisc *sch, unsigned long cl,
1244 struct sk_buff *skb, struct tcmsg *tcm)
1245 {
1246 struct netem_sched_data *q = qdisc_priv(sch);
1247
1248 if (cl != 1 || !q->qdisc) /* only one class */
1249 return -ENOENT;
1250
1251 tcm->tcm_handle |= TC_H_MIN(1);
1252 tcm->tcm_info = q->qdisc->handle;
1253
1254 return 0;
1255 }
1256
1257 static int netem_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1258 struct Qdisc **old, struct netlink_ext_ack *extack)
1259 {
1260 struct netem_sched_data *q = qdisc_priv(sch);
1261
1262 *old = qdisc_replace(sch, new, &q->qdisc);
1263 return 0;
1264 }
1265
1266 static struct Qdisc *netem_leaf(struct Qdisc *sch, unsigned long arg)
1267 {
1268 struct netem_sched_data *q = qdisc_priv(sch);
1269 return q->qdisc;
1270 }
1271
1272 static unsigned long netem_find(struct Qdisc *sch, u32 classid)
1273 {
1274 return 1;
1275 }
1276
1277 static void netem_walk(struct Qdisc *sch, struct qdisc_walker *walker)
1278 {
1279 if (!walker->stop) {
1280 if (!tc_qdisc_stats_dump(sch, 1, walker))
1281 return;
1282 }
1283 }
1284
1285 static const struct Qdisc_class_ops netem_class_ops = {
1286 .graft = netem_graft,
1287 .leaf = netem_leaf,
1288 .find = netem_find,
1289 .walk = netem_walk,
1290 .dump = netem_dump_class,
1291 };
1292
1293 static struct Qdisc_ops netem_qdisc_ops __read_mostly = {
1294 .id = "netem",
1295 .cl_ops = &netem_class_ops,
1296 .priv_size = sizeof(struct netem_sched_data),
1297 .enqueue = netem_enqueue,
1298 .dequeue = netem_dequeue,
1299 .peek = qdisc_peek_dequeued,
1300 .init = netem_init,
1301 .reset = netem_reset,
1302 .destroy = netem_destroy,
1303 .change = netem_change,
1304 .dump = netem_dump,
1305 .owner = THIS_MODULE,
1306 };
1307 MODULE_ALIAS_NET_SCH("netem");
1308
1309
1310 static int __init netem_module_init(void)
1311 {
1312 pr_info("netem: version " VERSION "\n");
1313 return register_qdisc(&netem_qdisc_ops);
1314 }
1315 static void __exit netem_module_exit(void)
1316 {
1317 unregister_qdisc(&netem_qdisc_ops);
1318 }
1319 module_init(netem_module_init)
1320 module_exit(netem_module_exit)
1321 MODULE_LICENSE("GPL");
1322 MODULE_DESCRIPTION("Network characteristics emulator qdisc");
Kernel DRM miscellaneous fixes and cross-tree changes