1 /*
2 * net/sched/sch_cbq.c Class-Based Queueing discipline.
3 *
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
7 * 2 of the License, or (at your option) any later version.
8 *
9 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
10 *
11 */
12
13 #include <linux/module.h>
14 #include <linux/slab.h>
15 #include <linux/types.h>
16 #include <linux/kernel.h>
17 #include <linux/string.h>
18 #include <linux/errno.h>
19 #include <linux/skbuff.h>
20 #include <net/netlink.h>
21 #include <net/pkt_sched.h>
22 #include <net/pkt_cls.h>
23
24
25 /* Class-Based Queueing (CBQ) algorithm.
26 =======================================
27
28 Sources: [1] Sally Floyd and Van Jacobson, "Link-sharing and Resource
29 Management Models for Packet Networks",
30 IEEE/ACM Transactions on Networking, Vol.3, No.4, 1995
31
32 [2] Sally Floyd, "Notes on CBQ and Guaranteed Service", 1995
33
34 [3] Sally Floyd, "Notes on Class-Based Queueing: Setting
35 Parameters", 1996
36
37 [4] Sally Floyd and Michael Speer, "Experimental Results
38 for Class-Based Queueing", 1998, not published.
39
40 -----------------------------------------------------------------------
41
42 Algorithm skeleton was taken from NS simulator cbq.cc.
43 If someone wants to check this code against the LBL version,
44 he should take into account that ONLY the skeleton was borrowed,
45 the implementation is different. Particularly:
46
47 --- The WRR algorithm is different. Our version looks more
48 reasonable (I hope) and works when quanta are allowed to be
49 less than MTU, which is always the case when real time classes
50 have small rates. Note, that the statement of [3] is
51 incomplete, delay may actually be estimated even if class
52 per-round allotment is less than MTU. Namely, if per-round
53 allotment is W*r_i, and r_1+...+r_k = r < 1
54
55 delay_i <= ([MTU/(W*r_i)]*W*r + W*r + k*MTU)/B
56
57 In the worst case we have IntServ estimate with D = W*r+k*MTU
58 and C = MTU*r. The proof (if correct at all) is trivial.
59
60
61 --- It seems that cbq-2.0 is not very accurate. At least, I cannot
62 interpret some places, which look like wrong translations
63 from NS. Anyone is advised to find these differences
64 and explain to me, why I am wrong 8).
65
66 --- Linux has no EOI event, so that we cannot estimate true class
67 idle time. Workaround is to consider the next dequeue event
68 as sign that previous packet is finished. This is wrong because of
69 internal device queueing, but on a permanently loaded link it is true.
70 Moreover, combined with clock integrator, this scheme looks
71 very close to an ideal solution. */
72
73 struct cbq_sched_data;
74
75
76 struct cbq_class {
77 struct Qdisc_class_common common;
78 struct cbq_class *next_alive; /* next class with backlog in this priority band */
79
80 /* Parameters */
81 unsigned char priority; /* class priority */
82 unsigned char priority2; /* priority to be used after overlimit */
83 unsigned char ewma_log; /* time constant for idle time calculation */
84
85 u32 defmap;
86
87 /* Link-sharing scheduler parameters */
88 long maxidle; /* Class parameters: see below. */
89 long offtime;
90 long minidle;
91 u32 avpkt;
92 struct qdisc_rate_table *R_tab;
93
94 /* General scheduler (WRR) parameters */
95 long allot;
96 long quantum; /* Allotment per WRR round */
97 long weight; /* Relative allotment: see below */
98
99 struct Qdisc *qdisc; /* Ptr to CBQ discipline */
100 struct cbq_class *split; /* Ptr to split node */
101 struct cbq_class *share; /* Ptr to LS parent in the class tree */
102 struct cbq_class *tparent; /* Ptr to tree parent in the class tree */
103 struct cbq_class *borrow; /* NULL if class is bandwidth limited;
104 parent otherwise */
105 struct cbq_class *sibling; /* Sibling chain */
106 struct cbq_class *children; /* Pointer to children chain */
107
108 struct Qdisc *q; /* Elementary queueing discipline */
109
110
111 /* Variables */
112 unsigned char cpriority; /* Effective priority */
113 unsigned char delayed;
114 unsigned char level; /* level of the class in hierarchy:
115 0 for leaf classes, and maximal
116 level of children + 1 for nodes.
117 */
118
119 psched_time_t last; /* Last end of service */
120 psched_time_t undertime;
121 long avgidle;
122 long deficit; /* Saved deficit for WRR */
123 psched_time_t penalized;
124 struct gnet_stats_basic_packed bstats;
125 struct gnet_stats_queue qstats;
126 struct net_rate_estimator __rcu *rate_est;
127 struct tc_cbq_xstats xstats;
128
129 struct tcf_proto __rcu *filter_list;
130 struct tcf_block *block;
131
132 int filters;
133
134 struct cbq_class *defaults[TC_PRIO_MAX + 1];
135 };
136
137 struct cbq_sched_data {
138 struct Qdisc_class_hash clhash; /* Hash table of all classes */
139 int nclasses[TC_CBQ_MAXPRIO + 1];
140 unsigned int quanta[TC_CBQ_MAXPRIO + 1];
141
142 struct cbq_class link;
143
144 unsigned int activemask;
145 struct cbq_class *active[TC_CBQ_MAXPRIO + 1]; /* List of all classes
146 with backlog */
147
148 #ifdef CONFIG_NET_CLS_ACT
149 struct cbq_class *rx_class;
150 #endif
151 struct cbq_class *tx_class;
152 struct cbq_class *tx_borrowed;
153 int tx_len;
154 psched_time_t now; /* Cached timestamp */
155 unsigned int pmask;
156
157 struct hrtimer delay_timer;
158 struct qdisc_watchdog watchdog; /* Watchdog timer,
159 started when CBQ has
160 backlog, but cannot
161 transmit just now */
162 psched_tdiff_t wd_expires;
163 int toplevel;
164 u32 hgenerator;
165 };
166
167
168 #define L2T(cl, len) qdisc_l2t((cl)->R_tab, len)
169
170 static inline struct cbq_class *
171 cbq_class_lookup(struct cbq_sched_data *q, u32 classid)
172 {
173 struct Qdisc_class_common *clc;
174
175 clc = qdisc_class_find(&q->clhash, classid);
176 if (clc == NULL)
177 return NULL;
178 return container_of(clc, struct cbq_class, common);
179 }
180
181 #ifdef CONFIG_NET_CLS_ACT
182
183 static struct cbq_class *
184 cbq_reclassify(struct sk_buff *skb, struct cbq_class *this)
185 {
186 struct cbq_class *cl;
187
188 for (cl = this->tparent; cl; cl = cl->tparent) {
189 struct cbq_class *new = cl->defaults[TC_PRIO_BESTEFFORT];
190
191 if (new != NULL && new != this)
192 return new;
193 }
194 return NULL;
195 }
196
197 #endif
198
199 /* Classify packet. The procedure is pretty complicated, but
200 * it allows us to combine link sharing and priority scheduling
201 * transparently.
202 *
203 * Namely, you can put link sharing rules (f.e. route based) at root of CBQ,
204 * so that it resolves to split nodes. Then packets are classified
205 * by logical priority, or a more specific classifier may be attached
206 * to the split node.
207 */
208
209 static struct cbq_class *
210 cbq_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
211 {
212 struct cbq_sched_data *q = qdisc_priv(sch);
213 struct cbq_class *head = &q->link;
214 struct cbq_class **defmap;
215 struct cbq_class *cl = NULL;
216 u32 prio = skb->priority;
217 struct tcf_proto *fl;
218 struct tcf_result res;
219
220 /*
221 * Step 1. If skb->priority points to one of our classes, use it.
222 */
223 if (TC_H_MAJ(prio ^ sch->handle) == 0 &&
224 (cl = cbq_class_lookup(q, prio)) != NULL)
225 return cl;
226
227 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
228 for (;;) {
229 int result = 0;
230 defmap = head->defaults;
231
232 fl = rcu_dereference_bh(head->filter_list);
233 /*
234 * Step 2+n. Apply classifier.
235 */
236 result = tcf_classify(skb, fl, &res, true);
237 if (!fl || result < 0)
238 goto fallback;
239
240 cl = (void *)res.class;
241 if (!cl) {
242 if (TC_H_MAJ(res.classid))
243 cl = cbq_class_lookup(q, res.classid);
244 else if ((cl = defmap[res.classid & TC_PRIO_MAX]) == NULL)
245 cl = defmap[TC_PRIO_BESTEFFORT];
246
247 if (cl == NULL)
248 goto fallback;
249 }
250 if (cl->level >= head->level)
251 goto fallback;
252 #ifdef CONFIG_NET_CLS_ACT
253 switch (result) {
254 case TC_ACT_QUEUED:
255 case TC_ACT_STOLEN:
256 case TC_ACT_TRAP:
257 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
258 /* fall through */
259 case TC_ACT_SHOT:
260 return NULL;
261 case TC_ACT_RECLASSIFY:
262 return cbq_reclassify(skb, cl);
263 }
264 #endif
265 if (cl->level == 0)
266 return cl;
267
268 /*
269 * Step 3+n. If classifier selected a link sharing class,
270 * apply agency specific classifier.
271 * Repeat this procdure until we hit a leaf node.
272 */
273 head = cl;
274 }
275
276 fallback:
277 cl = head;
278
279 /*
280 * Step 4. No success...
281 */
282 if (TC_H_MAJ(prio) == 0 &&
283 !(cl = head->defaults[prio & TC_PRIO_MAX]) &&
284 !(cl = head->defaults[TC_PRIO_BESTEFFORT]))
285 return head;
286
287 return cl;
288 }
289
290 /*
291 * A packet has just been enqueued on the empty class.
292 * cbq_activate_class adds it to the tail of active class list
293 * of its priority band.
294 */
295
296 static inline void cbq_activate_class(struct cbq_class *cl)
297 {
298 struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
299 int prio = cl->cpriority;
300 struct cbq_class *cl_tail;
301
302 cl_tail = q->active[prio];
303 q->active[prio] = cl;
304
305 if (cl_tail != NULL) {
306 cl->next_alive = cl_tail->next_alive;
307 cl_tail->next_alive = cl;
308 } else {
309 cl->next_alive = cl;
310 q->activemask |= (1<<prio);
311 }
312 }
313
314 /*
315 * Unlink class from active chain.
316 * Note that this same procedure is done directly in cbq_dequeue*
317 * during round-robin procedure.
318 */
319
320 static void cbq_deactivate_class(struct cbq_class *this)
321 {
322 struct cbq_sched_data *q = qdisc_priv(this->qdisc);
323 int prio = this->cpriority;
324 struct cbq_class *cl;
325 struct cbq_class *cl_prev = q->active[prio];
326
327 do {
328 cl = cl_prev->next_alive;
329 if (cl == this) {
330 cl_prev->next_alive = cl->next_alive;
331 cl->next_alive = NULL;
332
333 if (cl == q->active[prio]) {
334 q->active[prio] = cl_prev;
335 if (cl == q->active[prio]) {
336 q->active[prio] = NULL;
337 q->activemask &= ~(1<<prio);
338 return;
339 }
340 }
341 return;
342 }
343 } while ((cl_prev = cl) != q->active[prio]);
344 }
345
346 static void
347 cbq_mark_toplevel(struct cbq_sched_data *q, struct cbq_class *cl)
348 {
349 int toplevel = q->toplevel;
350
351 if (toplevel > cl->level) {
352 psched_time_t now = psched_get_time();
353
354 do {
355 if (cl->undertime < now) {
356 q->toplevel = cl->level;
357 return;
358 }
359 } while ((cl = cl->borrow) != NULL && toplevel > cl->level);
360 }
361 }
362
363 static int
364 cbq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
365 struct sk_buff **to_free)
366 {
367 struct cbq_sched_data *q = qdisc_priv(sch);
368 int uninitialized_var(ret);
369 struct cbq_class *cl = cbq_classify(skb, sch, &ret);
370
371 #ifdef CONFIG_NET_CLS_ACT
372 q->rx_class = cl;
373 #endif
374 if (cl == NULL) {
375 if (ret & __NET_XMIT_BYPASS)
376 qdisc_qstats_drop(sch);
377 __qdisc_drop(skb, to_free);
378 return ret;
379 }
380
381 ret = qdisc_enqueue(skb, cl->q, to_free);
382 if (ret == NET_XMIT_SUCCESS) {
383 sch->q.qlen++;
384 cbq_mark_toplevel(q, cl);
385 if (!cl->next_alive)
386 cbq_activate_class(cl);
387 return ret;
388 }
389
390 if (net_xmit_drop_count(ret)) {
391 qdisc_qstats_drop(sch);
392 cbq_mark_toplevel(q, cl);
393 cl->qstats.drops++;
394 }
395 return ret;
396 }
397
398 /* Overlimit action: penalize leaf class by adding offtime */
399 static void cbq_overlimit(struct cbq_class *cl)
400 {
401 struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
402 psched_tdiff_t delay = cl->undertime - q->now;
403
404 if (!cl->delayed) {
405 delay += cl->offtime;
406
407 /*
408 * Class goes to sleep, so that it will have no
409 * chance to work avgidle. Let's forgive it 8)
410 *
411 * BTW cbq-2.0 has a crap in this
412 * place, apparently they forgot to shift it by cl->ewma_log.
413 */
414 if (cl->avgidle < 0)
415 delay -= (-cl->avgidle) - ((-cl->avgidle) >> cl->ewma_log);
416 if (cl->avgidle < cl->minidle)
417 cl->avgidle = cl->minidle;
418 if (delay <= 0)
419 delay = 1;
420 cl->undertime = q->now + delay;
421
422 cl->xstats.overactions++;
423 cl->delayed = 1;
424 }
425 if (q->wd_expires == 0 || q->wd_expires > delay)
426 q->wd_expires = delay;
427
428 /* Dirty work! We must schedule wakeups based on
429 * real available rate, rather than leaf rate,
430 * which may be tiny (even zero).
431 */
432 if (q->toplevel == TC_CBQ_MAXLEVEL) {
433 struct cbq_class *b;
434 psched_tdiff_t base_delay = q->wd_expires;
435
436 for (b = cl->borrow; b; b = b->borrow) {
437 delay = b->undertime - q->now;
438 if (delay < base_delay) {
439 if (delay <= 0)
440 delay = 1;
441 base_delay = delay;
442 }
443 }
444
445 q->wd_expires = base_delay;
446 }
447 }
448
449 static psched_tdiff_t cbq_undelay_prio(struct cbq_sched_data *q, int prio,
450 psched_time_t now)
451 {
452 struct cbq_class *cl;
453 struct cbq_class *cl_prev = q->active[prio];
454 psched_time_t sched = now;
455
456 if (cl_prev == NULL)
457 return 0;
458
459 do {
460 cl = cl_prev->next_alive;
461 if (now - cl->penalized > 0) {
462 cl_prev->next_alive = cl->next_alive;
463 cl->next_alive = NULL;
464 cl->cpriority = cl->priority;
465 cl->delayed = 0;
466 cbq_activate_class(cl);
467
468 if (cl == q->active[prio]) {
469 q->active[prio] = cl_prev;
470 if (cl == q->active[prio]) {
471 q->active[prio] = NULL;
472 return 0;
473 }
474 }
475
476 cl = cl_prev->next_alive;
477 } else if (sched - cl->penalized > 0)
478 sched = cl->penalized;
479 } while ((cl_prev = cl) != q->active[prio]);
480
481 return sched - now;
482 }
483
484 static enum hrtimer_restart cbq_undelay(struct hrtimer *timer)
485 {
486 struct cbq_sched_data *q = container_of(timer, struct cbq_sched_data,
487 delay_timer);
488 struct Qdisc *sch = q->watchdog.qdisc;
489 psched_time_t now;
490 psched_tdiff_t delay = 0;
491 unsigned int pmask;
492
493 now = psched_get_time();
494
495 pmask = q->pmask;
496 q->pmask = 0;
497
498 while (pmask) {
499 int prio = ffz(~pmask);
500 psched_tdiff_t tmp;
501
502 pmask &= ~(1<<prio);
503
504 tmp = cbq_undelay_prio(q, prio, now);
505 if (tmp > 0) {
506 q->pmask |= 1<<prio;
507 if (tmp < delay || delay == 0)
508 delay = tmp;
509 }
510 }
511
512 if (delay) {
513 ktime_t time;
514
515 time = 0;
516 time = ktime_add_ns(time, PSCHED_TICKS2NS(now + delay));
517 hrtimer_start(&q->delay_timer, time, HRTIMER_MODE_ABS_PINNED);
518 }
519
520 __netif_schedule(qdisc_root(sch));
521 return HRTIMER_NORESTART;
522 }
523
524 /*
525 * It is mission critical procedure.
526 *
527 * We "regenerate" toplevel cutoff, if transmitting class
528 * has backlog and it is not regulated. It is not part of
529 * original CBQ description, but looks more reasonable.
530 * Probably, it is wrong. This question needs further investigation.
531 */
532
533 static inline void
534 cbq_update_toplevel(struct cbq_sched_data *q, struct cbq_class *cl,
535 struct cbq_class *borrowed)
536 {
537 if (cl && q->toplevel >= borrowed->level) {
538 if (cl->q->q.qlen > 1) {
539 do {
540 if (borrowed->undertime == PSCHED_PASTPERFECT) {
541 q->toplevel = borrowed->level;
542 return;
543 }
544 } while ((borrowed = borrowed->borrow) != NULL);
545 }
546 #if 0
547 /* It is not necessary now. Uncommenting it
548 will save CPU cycles, but decrease fairness.
549 */
550 q->toplevel = TC_CBQ_MAXLEVEL;
551 #endif
552 }
553 }
554
555 static void
556 cbq_update(struct cbq_sched_data *q)
557 {
558 struct cbq_class *this = q->tx_class;
559 struct cbq_class *cl = this;
560 int len = q->tx_len;
561 psched_time_t now;
562
563 q->tx_class = NULL;
564 /* Time integrator. We calculate EOS time
565 * by adding expected packet transmission time.
566 */
567 now = q->now + L2T(&q->link, len);
568
569 for ( ; cl; cl = cl->share) {
570 long avgidle = cl->avgidle;
571 long idle;
572
573 cl->bstats.packets++;
574 cl->bstats.bytes += len;
575
576 /*
577 * (now - last) is total time between packet right edges.
578 * (last_pktlen/rate) is "virtual" busy time, so that
579 *
580 * idle = (now - last) - last_pktlen/rate
581 */
582
583 idle = now - cl->last;
584 if ((unsigned long)idle > 128*1024*1024) {
585 avgidle = cl->maxidle;
586 } else {
587 idle -= L2T(cl, len);
588
589 /* true_avgidle := (1-W)*true_avgidle + W*idle,
590 * where W=2^{-ewma_log}. But cl->avgidle is scaled:
591 * cl->avgidle == true_avgidle/W,
592 * hence:
593 */
594 avgidle += idle - (avgidle>>cl->ewma_log);
595 }
596
597 if (avgidle <= 0) {
598 /* Overlimit or at-limit */
599
600 if (avgidle < cl->minidle)
601 avgidle = cl->minidle;
602
603 cl->avgidle = avgidle;
604
605 /* Calculate expected time, when this class
606 * will be allowed to send.
607 * It will occur, when:
608 * (1-W)*true_avgidle + W*delay = 0, i.e.
609 * idle = (1/W - 1)*(-true_avgidle)
610 * or
611 * idle = (1 - W)*(-cl->avgidle);
612 */
613 idle = (-avgidle) - ((-avgidle) >> cl->ewma_log);
614
615 /*
616 * That is not all.
617 * To maintain the rate allocated to the class,
618 * we add to undertime virtual clock,
619 * necessary to complete transmitted packet.
620 * (len/phys_bandwidth has been already passed
621 * to the moment of cbq_update)
622 */
623
624 idle -= L2T(&q->link, len);
625 idle += L2T(cl, len);
626
627 cl->undertime = now + idle;
628 } else {
629 /* Underlimit */
630
631 cl->undertime = PSCHED_PASTPERFECT;
632 if (avgidle > cl->maxidle)
633 cl->avgidle = cl->maxidle;
634 else
635 cl->avgidle = avgidle;
636 }
637 if ((s64)(now - cl->last) > 0)
638 cl->last = now;
639 }
640
641 cbq_update_toplevel(q, this, q->tx_borrowed);
642 }
643
644 static inline struct cbq_class *
645 cbq_under_limit(struct cbq_class *cl)
646 {
647 struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
648 struct cbq_class *this_cl = cl;
649
650 if (cl->tparent == NULL)
651 return cl;
652
653 if (cl->undertime == PSCHED_PASTPERFECT || q->now >= cl->undertime) {
654 cl->delayed = 0;
655 return cl;
656 }
657
658 do {
659 /* It is very suspicious place. Now overlimit
660 * action is generated for not bounded classes
661 * only if link is completely congested.
662 * Though it is in agree with ancestor-only paradigm,
663 * it looks very stupid. Particularly,
664 * it means that this chunk of code will either
665 * never be called or result in strong amplification
666 * of burstiness. Dangerous, silly, and, however,
667 * no another solution exists.
668 */
669 cl = cl->borrow;
670 if (!cl) {
671 this_cl->qstats.overlimits++;
672 cbq_overlimit(this_cl);
673 return NULL;
674 }
675 if (cl->level > q->toplevel)
676 return NULL;
677 } while (cl->undertime != PSCHED_PASTPERFECT && q->now < cl->undertime);
678
679 cl->delayed = 0;
680 return cl;
681 }
682
683 static inline struct sk_buff *
684 cbq_dequeue_prio(struct Qdisc *sch, int prio)
685 {
686 struct cbq_sched_data *q = qdisc_priv(sch);
687 struct cbq_class *cl_tail, *cl_prev, *cl;
688 struct sk_buff *skb;
689 int deficit;
690
691 cl_tail = cl_prev = q->active[prio];
692 cl = cl_prev->next_alive;
693
694 do {
695 deficit = 0;
696
697 /* Start round */
698 do {
699 struct cbq_class *borrow = cl;
700
701 if (cl->q->q.qlen &&
702 (borrow = cbq_under_limit(cl)) == NULL)
703 goto skip_class;
704
705 if (cl->deficit <= 0) {
706 /* Class exhausted its allotment per
707 * this round. Switch to the next one.
708 */
709 deficit = 1;
710 cl->deficit += cl->quantum;
711 goto next_class;
712 }
713
714 skb = cl->q->dequeue(cl->q);
715
716 /* Class did not give us any skb :-(
717 * It could occur even if cl->q->q.qlen != 0
718 * f.e. if cl->q == "tbf"
719 */
720 if (skb == NULL)
721 goto skip_class;
722
723 cl->deficit -= qdisc_pkt_len(skb);
724 q->tx_class = cl;
725 q->tx_borrowed = borrow;
726 if (borrow != cl) {
727 #ifndef CBQ_XSTATS_BORROWS_BYTES
728 borrow->xstats.borrows++;
729 cl->xstats.borrows++;
730 #else
731 borrow->xstats.borrows += qdisc_pkt_len(skb);
732 cl->xstats.borrows += qdisc_pkt_len(skb);
733 #endif
734 }
735 q->tx_len = qdisc_pkt_len(skb);
736
737 if (cl->deficit <= 0) {
738 q->active[prio] = cl;
739 cl = cl->next_alive;
740 cl->deficit += cl->quantum;
741 }
742 return skb;
743
744 skip_class:
745 if (cl->q->q.qlen == 0 || prio != cl->cpriority) {
746 /* Class is empty or penalized.
747 * Unlink it from active chain.
748 */
749 cl_prev->next_alive = cl->next_alive;
750 cl->next_alive = NULL;
751
752 /* Did cl_tail point to it? */
753 if (cl == cl_tail) {
754 /* Repair it! */
755 cl_tail = cl_prev;
756
757 /* Was it the last class in this band? */
758 if (cl == cl_tail) {
759 /* Kill the band! */
760 q->active[prio] = NULL;
761 q->activemask &= ~(1<<prio);
762 if (cl->q->q.qlen)
763 cbq_activate_class(cl);
764 return NULL;
765 }
766
767 q->active[prio] = cl_tail;
768 }
769 if (cl->q->q.qlen)
770 cbq_activate_class(cl);
771
772 cl = cl_prev;
773 }
774
775 next_class:
776 cl_prev = cl;
777 cl = cl->next_alive;
778 } while (cl_prev != cl_tail);
779 } while (deficit);
780
781 q->active[prio] = cl_prev;
782
783 return NULL;
784 }
785
786 static inline struct sk_buff *
787 cbq_dequeue_1(struct Qdisc *sch)
788 {
789 struct cbq_sched_data *q = qdisc_priv(sch);
790 struct sk_buff *skb;
791 unsigned int activemask;
792
793 activemask = q->activemask & 0xFF;
794 while (activemask) {
795 int prio = ffz(~activemask);
796 activemask &= ~(1<<prio);
797 skb = cbq_dequeue_prio(sch, prio);
798 if (skb)
799 return skb;
800 }
801 return NULL;
802 }
803
804 static struct sk_buff *
805 cbq_dequeue(struct Qdisc *sch)
806 {
807 struct sk_buff *skb;
808 struct cbq_sched_data *q = qdisc_priv(sch);
809 psched_time_t now;
810
811 now = psched_get_time();
812
813 if (q->tx_class)
814 cbq_update(q);
815
816 q->now = now;
817
818 for (;;) {
819 q->wd_expires = 0;
820
821 skb = cbq_dequeue_1(sch);
822 if (skb) {
823 qdisc_bstats_update(sch, skb);
824 sch->q.qlen--;
825 return skb;
826 }
827
828 /* All the classes are overlimit.
829 *
830 * It is possible, if:
831 *
832 * 1. Scheduler is empty.
833 * 2. Toplevel cutoff inhibited borrowing.
834 * 3. Root class is overlimit.
835 *
836 * Reset 2d and 3d conditions and retry.
837 *
838 * Note, that NS and cbq-2.0 are buggy, peeking
839 * an arbitrary class is appropriate for ancestor-only
840 * sharing, but not for toplevel algorithm.
841 *
842 * Our version is better, but slower, because it requires
843 * two passes, but it is unavoidable with top-level sharing.
844 */
845
846 if (q->toplevel == TC_CBQ_MAXLEVEL &&
847 q->link.undertime == PSCHED_PASTPERFECT)
848 break;
849
850 q->toplevel = TC_CBQ_MAXLEVEL;
851 q->link.undertime = PSCHED_PASTPERFECT;
852 }
853
854 /* No packets in scheduler or nobody wants to give them to us :-(
855 * Sigh... start watchdog timer in the last case.
856 */
857
858 if (sch->q.qlen) {
859 qdisc_qstats_overlimit(sch);
860 if (q->wd_expires)
861 qdisc_watchdog_schedule(&q->watchdog,
862 now + q->wd_expires);
863 }
864 return NULL;
865 }
866
867 /* CBQ class maintanance routines */
868
869 static void cbq_adjust_levels(struct cbq_class *this)
870 {
871 if (this == NULL)
872 return;
873
874 do {
875 int level = 0;
876 struct cbq_class *cl;
877
878 cl = this->children;
879 if (cl) {
880 do {
881 if (cl->level > level)
882 level = cl->level;
883 } while ((cl = cl->sibling) != this->children);
884 }
885 this->level = level + 1;
886 } while ((this = this->tparent) != NULL);
887 }
888
889 static void cbq_normalize_quanta(struct cbq_sched_data *q, int prio)
890 {
891 struct cbq_class *cl;
892 unsigned int h;
893
894 if (q->quanta[prio] == 0)
895 return;
896
897 for (h = 0; h < q->clhash.hashsize; h++) {
898 hlist_for_each_entry(cl, &q->clhash.hash[h], common.hnode) {
899 /* BUGGGG... Beware! This expression suffer of
900 * arithmetic overflows!
901 */
902 if (cl->priority == prio) {
903 cl->quantum = (cl->weight*cl->allot*q->nclasses[prio])/
904 q->quanta[prio];
905 }
906 if (cl->quantum <= 0 ||
907 cl->quantum > 32*qdisc_dev(cl->qdisc)->mtu) {
908 pr_warn("CBQ: class %08x has bad quantum==%ld, repaired.\n",
909 cl->common.classid, cl->quantum);
910 cl->quantum = qdisc_dev(cl->qdisc)->mtu/2 + 1;
911 }
912 }
913 }
914 }
915
916 static void cbq_sync_defmap(struct cbq_class *cl)
917 {
918 struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
919 struct cbq_class *split = cl->split;
920 unsigned int h;
921 int i;
922
923 if (split == NULL)
924 return;
925
926 for (i = 0; i <= TC_PRIO_MAX; i++) {
927 if (split->defaults[i] == cl && !(cl->defmap & (1<<i)))
928 split->defaults[i] = NULL;
929 }
930
931 for (i = 0; i <= TC_PRIO_MAX; i++) {
932 int level = split->level;
933
934 if (split->defaults[i])
935 continue;
936
937 for (h = 0; h < q->clhash.hashsize; h++) {
938 struct cbq_class *c;
939
940 hlist_for_each_entry(c, &q->clhash.hash[h],
941 common.hnode) {
942 if (c->split == split && c->level < level &&
943 c->defmap & (1<<i)) {
944 split->defaults[i] = c;
945 level = c->level;
946 }
947 }
948 }
949 }
950 }
951
952 static void cbq_change_defmap(struct cbq_class *cl, u32 splitid, u32 def, u32 mask)
953 {
954 struct cbq_class *split = NULL;
955
956 if (splitid == 0) {
957 split = cl->split;
958 if (!split)
959 return;
960 splitid = split->common.classid;
961 }
962
963 if (split == NULL || split->common.classid != splitid) {
964 for (split = cl->tparent; split; split = split->tparent)
965 if (split->common.classid == splitid)
966 break;
967 }
968
969 if (split == NULL)
970 return;
971
972 if (cl->split != split) {
973 cl->defmap = 0;
974 cbq_sync_defmap(cl);
975 cl->split = split;
976 cl->defmap = def & mask;
977 } else
978 cl->defmap = (cl->defmap & ~mask) | (def & mask);
979
980 cbq_sync_defmap(cl);
981 }
982
983 static void cbq_unlink_class(struct cbq_class *this)
984 {
985 struct cbq_class *cl, **clp;
986 struct cbq_sched_data *q = qdisc_priv(this->qdisc);
987
988 qdisc_class_hash_remove(&q->clhash, &this->common);
989
990 if (this->tparent) {
991 clp = &this->sibling;
992 cl = *clp;
993 do {
994 if (cl == this) {
995 *clp = cl->sibling;
996 break;
997 }
998 clp = &cl->sibling;
999 } while ((cl = *clp) != this->sibling);
1000
1001 if (this->tparent->children == this) {
1002 this->tparent->children = this->sibling;
1003 if (this->sibling == this)
1004 this->tparent->children = NULL;
1005 }
1006 } else {
1007 WARN_ON(this->sibling != this);
1008 }
1009 }
1010
1011 static void cbq_link_class(struct cbq_class *this)
1012 {
1013 struct cbq_sched_data *q = qdisc_priv(this->qdisc);
1014 struct cbq_class *parent = this->tparent;
1015
1016 this->sibling = this;
1017 qdisc_class_hash_insert(&q->clhash, &this->common);
1018
1019 if (parent == NULL)
1020 return;
1021
1022 if (parent->children == NULL) {
1023 parent->children = this;
1024 } else {
1025 this->sibling = parent->children->sibling;
1026 parent->children->sibling = this;
1027 }
1028 }
1029
1030 static void
1031 cbq_reset(struct Qdisc *sch)
1032 {
1033 struct cbq_sched_data *q = qdisc_priv(sch);
1034 struct cbq_class *cl;
1035 int prio;
1036 unsigned int h;
1037
1038 q->activemask = 0;
1039 q->pmask = 0;
1040 q->tx_class = NULL;
1041 q->tx_borrowed = NULL;
1042 qdisc_watchdog_cancel(&q->watchdog);
1043 hrtimer_cancel(&q->delay_timer);
1044 q->toplevel = TC_CBQ_MAXLEVEL;
1045 q->now = psched_get_time();
1046
1047 for (prio = 0; prio <= TC_CBQ_MAXPRIO; prio++)
1048 q->active[prio] = NULL;
1049
1050 for (h = 0; h < q->clhash.hashsize; h++) {
1051 hlist_for_each_entry(cl, &q->clhash.hash[h], common.hnode) {
1052 qdisc_reset(cl->q);
1053
1054 cl->next_alive = NULL;
1055 cl->undertime = PSCHED_PASTPERFECT;
1056 cl->avgidle = cl->maxidle;
1057 cl->deficit = cl->quantum;
1058 cl->cpriority = cl->priority;
1059 }
1060 }
1061 sch->q.qlen = 0;
1062 }
1063
1064
1065 static int cbq_set_lss(struct cbq_class *cl, struct tc_cbq_lssopt *lss)
1066 {
1067 if (lss->change & TCF_CBQ_LSS_FLAGS) {
1068 cl->share = (lss->flags & TCF_CBQ_LSS_ISOLATED) ? NULL : cl->tparent;
1069 cl->borrow = (lss->flags & TCF_CBQ_LSS_BOUNDED) ? NULL : cl->tparent;
1070 }
1071 if (lss->change & TCF_CBQ_LSS_EWMA)
1072 cl->ewma_log = lss->ewma_log;
1073 if (lss->change & TCF_CBQ_LSS_AVPKT)
1074 cl->avpkt = lss->avpkt;
1075 if (lss->change & TCF_CBQ_LSS_MINIDLE)
1076 cl->minidle = -(long)lss->minidle;
1077 if (lss->change & TCF_CBQ_LSS_MAXIDLE) {
1078 cl->maxidle = lss->maxidle;
1079 cl->avgidle = lss->maxidle;
1080 }
1081 if (lss->change & TCF_CBQ_LSS_OFFTIME)
1082 cl->offtime = lss->offtime;
1083 return 0;
1084 }
1085
1086 static void cbq_rmprio(struct cbq_sched_data *q, struct cbq_class *cl)
1087 {
1088 q->nclasses[cl->priority]--;
1089 q->quanta[cl->priority] -= cl->weight;
1090 cbq_normalize_quanta(q, cl->priority);
1091 }
1092
1093 static void cbq_addprio(struct cbq_sched_data *q, struct cbq_class *cl)
1094 {
1095 q->nclasses[cl->priority]++;
1096 q->quanta[cl->priority] += cl->weight;
1097 cbq_normalize_quanta(q, cl->priority);
1098 }
1099
1100 static int cbq_set_wrr(struct cbq_class *cl, struct tc_cbq_wrropt *wrr)
1101 {
1102 struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
1103
1104 if (wrr->allot)
1105 cl->allot = wrr->allot;
1106 if (wrr->weight)
1107 cl->weight = wrr->weight;
1108 if (wrr->priority) {
1109 cl->priority = wrr->priority - 1;
1110 cl->cpriority = cl->priority;
1111 if (cl->priority >= cl->priority2)
1112 cl->priority2 = TC_CBQ_MAXPRIO - 1;
1113 }
1114
1115 cbq_addprio(q, cl);
1116 return 0;
1117 }
1118
1119 static int cbq_set_fopt(struct cbq_class *cl, struct tc_cbq_fopt *fopt)
1120 {
1121 cbq_change_defmap(cl, fopt->split, fopt->defmap, fopt->defchange);
1122 return 0;
1123 }
1124
1125 static const struct nla_policy cbq_policy[TCA_CBQ_MAX + 1] = {
1126 [TCA_CBQ_LSSOPT] = { .len = sizeof(struct tc_cbq_lssopt) },
1127 [TCA_CBQ_WRROPT] = { .len = sizeof(struct tc_cbq_wrropt) },
1128 [TCA_CBQ_FOPT] = { .len = sizeof(struct tc_cbq_fopt) },
1129 [TCA_CBQ_OVL_STRATEGY] = { .len = sizeof(struct tc_cbq_ovl) },
1130 [TCA_CBQ_RATE] = { .len = sizeof(struct tc_ratespec) },
1131 [TCA_CBQ_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
1132 [TCA_CBQ_POLICE] = { .len = sizeof(struct tc_cbq_police) },
1133 };
1134
1135 static int cbq_init(struct Qdisc *sch, struct nlattr *opt)
1136 {
1137 struct cbq_sched_data *q = qdisc_priv(sch);
1138 struct nlattr *tb[TCA_CBQ_MAX + 1];
1139 struct tc_ratespec *r;
1140 int err;
1141
1142 qdisc_watchdog_init(&q->watchdog, sch);
1143 hrtimer_init(&q->delay_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
1144 q->delay_timer.function = cbq_undelay;
1145
1146 if (!opt)
1147 return -EINVAL;
1148
1149 err = nla_parse_nested(tb, TCA_CBQ_MAX, opt, cbq_policy, NULL);
1150 if (err < 0)
1151 return err;
1152
1153 if (tb[TCA_CBQ_RTAB] == NULL || tb[TCA_CBQ_RATE] == NULL)
1154 return -EINVAL;
1155
1156 r = nla_data(tb[TCA_CBQ_RATE]);
1157
1158 if ((q->link.R_tab = qdisc_get_rtab(r, tb[TCA_CBQ_RTAB])) == NULL)
1159 return -EINVAL;
1160
1161 err = tcf_block_get(&q->link.block, &q->link.filter_list, sch);
1162 if (err)
1163 goto put_rtab;
1164
1165 err = qdisc_class_hash_init(&q->clhash);
1166 if (err < 0)
1167 goto put_block;
1168
1169 q->link.sibling = &q->link;
1170 q->link.common.classid = sch->handle;
1171 q->link.qdisc = sch;
1172 q->link.q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1173 sch->handle);
1174 if (!q->link.q)
1175 q->link.q = &noop_qdisc;
1176 else
1177 qdisc_hash_add(q->link.q, true);
1178
1179 q->link.priority = TC_CBQ_MAXPRIO - 1;
1180 q->link.priority2 = TC_CBQ_MAXPRIO - 1;
1181 q->link.cpriority = TC_CBQ_MAXPRIO - 1;
1182 q->link.allot = psched_mtu(qdisc_dev(sch));
1183 q->link.quantum = q->link.allot;
1184 q->link.weight = q->link.R_tab->rate.rate;
1185
1186 q->link.ewma_log = TC_CBQ_DEF_EWMA;
1187 q->link.avpkt = q->link.allot/2;
1188 q->link.minidle = -0x7FFFFFFF;
1189
1190 q->toplevel = TC_CBQ_MAXLEVEL;
1191 q->now = psched_get_time();
1192
1193 cbq_link_class(&q->link);
1194
1195 if (tb[TCA_CBQ_LSSOPT])
1196 cbq_set_lss(&q->link, nla_data(tb[TCA_CBQ_LSSOPT]));
1197
1198 cbq_addprio(q, &q->link);
1199 return 0;
1200
1201 put_block:
1202 tcf_block_put(q->link.block);
1203
1204 put_rtab:
1205 qdisc_put_rtab(q->link.R_tab);
1206 return err;
1207 }
1208
1209 static int cbq_dump_rate(struct sk_buff *skb, struct cbq_class *cl)
1210 {
1211 unsigned char *b = skb_tail_pointer(skb);
1212
1213 if (nla_put(skb, TCA_CBQ_RATE, sizeof(cl->R_tab->rate), &cl->R_tab->rate))
1214 goto nla_put_failure;
1215 return skb->len;
1216
1217 nla_put_failure:
1218 nlmsg_trim(skb, b);
1219 return -1;
1220 }
1221
1222 static int cbq_dump_lss(struct sk_buff *skb, struct cbq_class *cl)
1223 {
1224 unsigned char *b = skb_tail_pointer(skb);
1225 struct tc_cbq_lssopt opt;
1226
1227 opt.flags = 0;
1228 if (cl->borrow == NULL)
1229 opt.flags |= TCF_CBQ_LSS_BOUNDED;
1230 if (cl->share == NULL)
1231 opt.flags |= TCF_CBQ_LSS_ISOLATED;
1232 opt.ewma_log = cl->ewma_log;
1233 opt.level = cl->level;
1234 opt.avpkt = cl->avpkt;
1235 opt.maxidle = cl->maxidle;
1236 opt.minidle = (u32)(-cl->minidle);
1237 opt.offtime = cl->offtime;
1238 opt.change = ~0;
1239 if (nla_put(skb, TCA_CBQ_LSSOPT, sizeof(opt), &opt))
1240 goto nla_put_failure;
1241 return skb->len;
1242
1243 nla_put_failure:
1244 nlmsg_trim(skb, b);
1245 return -1;
1246 }
1247
1248 static int cbq_dump_wrr(struct sk_buff *skb, struct cbq_class *cl)
1249 {
1250 unsigned char *b = skb_tail_pointer(skb);
1251 struct tc_cbq_wrropt opt;
1252
1253 memset(&opt, 0, sizeof(opt));
1254 opt.flags = 0;
1255 opt.allot = cl->allot;
1256 opt.priority = cl->priority + 1;
1257 opt.cpriority = cl->cpriority + 1;
1258 opt.weight = cl->weight;
1259 if (nla_put(skb, TCA_CBQ_WRROPT, sizeof(opt), &opt))
1260 goto nla_put_failure;
1261 return skb->len;
1262
1263 nla_put_failure:
1264 nlmsg_trim(skb, b);
1265 return -1;
1266 }
1267
1268 static int cbq_dump_fopt(struct sk_buff *skb, struct cbq_class *cl)
1269 {
1270 unsigned char *b = skb_tail_pointer(skb);
1271 struct tc_cbq_fopt opt;
1272
1273 if (cl->split || cl->defmap) {
1274 opt.split = cl->split ? cl->split->common.classid : 0;
1275 opt.defmap = cl->defmap;
1276 opt.defchange = ~0;
1277 if (nla_put(skb, TCA_CBQ_FOPT, sizeof(opt), &opt))
1278 goto nla_put_failure;
1279 }
1280 return skb->len;
1281
1282 nla_put_failure:
1283 nlmsg_trim(skb, b);
1284 return -1;
1285 }
1286
1287 static int cbq_dump_attr(struct sk_buff *skb, struct cbq_class *cl)
1288 {
1289 if (cbq_dump_lss(skb, cl) < 0 ||
1290 cbq_dump_rate(skb, cl) < 0 ||
1291 cbq_dump_wrr(skb, cl) < 0 ||
1292 cbq_dump_fopt(skb, cl) < 0)
1293 return -1;
1294 return 0;
1295 }
1296
1297 static int cbq_dump(struct Qdisc *sch, struct sk_buff *skb)
1298 {
1299 struct cbq_sched_data *q = qdisc_priv(sch);
1300 struct nlattr *nest;
1301
1302 nest = nla_nest_start(skb, TCA_OPTIONS);
1303 if (nest == NULL)
1304 goto nla_put_failure;
1305 if (cbq_dump_attr(skb, &q->link) < 0)
1306 goto nla_put_failure;
1307 return nla_nest_end(skb, nest);
1308
1309 nla_put_failure:
1310 nla_nest_cancel(skb, nest);
1311 return -1;
1312 }
1313
1314 static int
1315 cbq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
1316 {
1317 struct cbq_sched_data *q = qdisc_priv(sch);
1318
1319 q->link.xstats.avgidle = q->link.avgidle;
1320 return gnet_stats_copy_app(d, &q->link.xstats, sizeof(q->link.xstats));
1321 }
1322
1323 static int
1324 cbq_dump_class(struct Qdisc *sch, unsigned long arg,
1325 struct sk_buff *skb, struct tcmsg *tcm)
1326 {
1327 struct cbq_class *cl = (struct cbq_class *)arg;
1328 struct nlattr *nest;
1329
1330 if (cl->tparent)
1331 tcm->tcm_parent = cl->tparent->common.classid;
1332 else
1333 tcm->tcm_parent = TC_H_ROOT;
1334 tcm->tcm_handle = cl->common.classid;
1335 tcm->tcm_info = cl->q->handle;
1336
1337 nest = nla_nest_start(skb, TCA_OPTIONS);
1338 if (nest == NULL)
1339 goto nla_put_failure;
1340 if (cbq_dump_attr(skb, cl) < 0)
1341 goto nla_put_failure;
1342 return nla_nest_end(skb, nest);
1343
1344 nla_put_failure:
1345 nla_nest_cancel(skb, nest);
1346 return -1;
1347 }
1348
1349 static int
1350 cbq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1351 struct gnet_dump *d)
1352 {
1353 struct cbq_sched_data *q = qdisc_priv(sch);
1354 struct cbq_class *cl = (struct cbq_class *)arg;
1355
1356 cl->xstats.avgidle = cl->avgidle;
1357 cl->xstats.undertime = 0;
1358
1359 if (cl->undertime != PSCHED_PASTPERFECT)
1360 cl->xstats.undertime = cl->undertime - q->now;
1361
1362 if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
1363 d, NULL, &cl->bstats) < 0 ||
1364 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1365 gnet_stats_copy_queue(d, NULL, &cl->qstats, cl->q->q.qlen) < 0)
1366 return -1;
1367
1368 return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1369 }
1370
1371 static int cbq_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1372 struct Qdisc **old)
1373 {
1374 struct cbq_class *cl = (struct cbq_class *)arg;
1375
1376 if (new == NULL) {
1377 new = qdisc_create_dflt(sch->dev_queue,
1378 &pfifo_qdisc_ops, cl->common.classid);
1379 if (new == NULL)
1380 return -ENOBUFS;
1381 }
1382
1383 *old = qdisc_replace(sch, new, &cl->q);
1384 return 0;
1385 }
1386
1387 static struct Qdisc *cbq_leaf(struct Qdisc *sch, unsigned long arg)
1388 {
1389 struct cbq_class *cl = (struct cbq_class *)arg;
1390
1391 return cl->q;
1392 }
1393
1394 static void cbq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1395 {
1396 struct cbq_class *cl = (struct cbq_class *)arg;
1397
1398 cbq_deactivate_class(cl);
1399 }
1400
1401 static unsigned long cbq_find(struct Qdisc *sch, u32 classid)
1402 {
1403 struct cbq_sched_data *q = qdisc_priv(sch);
1404
1405 return (unsigned long)cbq_class_lookup(q, classid);
1406 }
1407
1408 static void cbq_destroy_class(struct Qdisc *sch, struct cbq_class *cl)
1409 {
1410 struct cbq_sched_data *q = qdisc_priv(sch);
1411
1412 WARN_ON(cl->filters);
1413
1414 tcf_block_put(cl->block);
1415 qdisc_destroy(cl->q);
1416 qdisc_put_rtab(cl->R_tab);
1417 gen_kill_estimator(&cl->rate_est);
1418 if (cl != &q->link)
1419 kfree(cl);
1420 }
1421
1422 static void cbq_destroy(struct Qdisc *sch)
1423 {
1424 struct cbq_sched_data *q = qdisc_priv(sch);
1425 struct hlist_node *next;
1426 struct cbq_class *cl;
1427 unsigned int h;
1428
1429 #ifdef CONFIG_NET_CLS_ACT
1430 q->rx_class = NULL;
1431 #endif
1432 /*
1433 * Filters must be destroyed first because we don't destroy the
1434 * classes from root to leafs which means that filters can still
1435 * be bound to classes which have been destroyed already. --TGR '04
1436 */
1437 for (h = 0; h < q->clhash.hashsize; h++) {
1438 hlist_for_each_entry(cl, &q->clhash.hash[h], common.hnode) {
1439 tcf_block_put(cl->block);
1440 cl->block = NULL;
1441 }
1442 }
1443 for (h = 0; h < q->clhash.hashsize; h++) {
1444 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[h],
1445 common.hnode)
1446 cbq_destroy_class(sch, cl);
1447 }
1448 qdisc_class_hash_destroy(&q->clhash);
1449 }
1450
1451 static int
1452 cbq_change_class(struct Qdisc *sch, u32 classid, u32 parentid, struct nlattr **tca,
1453 unsigned long *arg)
1454 {
1455 int err;
1456 struct cbq_sched_data *q = qdisc_priv(sch);
1457 struct cbq_class *cl = (struct cbq_class *)*arg;
1458 struct nlattr *opt = tca[TCA_OPTIONS];
1459 struct nlattr *tb[TCA_CBQ_MAX + 1];
1460 struct cbq_class *parent;
1461 struct qdisc_rate_table *rtab = NULL;
1462
1463 if (opt == NULL)
1464 return -EINVAL;
1465
1466 err = nla_parse_nested(tb, TCA_CBQ_MAX, opt, cbq_policy, NULL);
1467 if (err < 0)
1468 return err;
1469
1470 if (tb[TCA_CBQ_OVL_STRATEGY] || tb[TCA_CBQ_POLICE])
1471 return -EOPNOTSUPP;
1472
1473 if (cl) {
1474 /* Check parent */
1475 if (parentid) {
1476 if (cl->tparent &&
1477 cl->tparent->common.classid != parentid)
1478 return -EINVAL;
1479 if (!cl->tparent && parentid != TC_H_ROOT)
1480 return -EINVAL;
1481 }
1482
1483 if (tb[TCA_CBQ_RATE]) {
1484 rtab = qdisc_get_rtab(nla_data(tb[TCA_CBQ_RATE]),
1485 tb[TCA_CBQ_RTAB]);
1486 if (rtab == NULL)
1487 return -EINVAL;
1488 }
1489
1490 if (tca[TCA_RATE]) {
1491 err = gen_replace_estimator(&cl->bstats, NULL,
1492 &cl->rate_est,
1493 NULL,
1494 qdisc_root_sleeping_running(sch),
1495 tca[TCA_RATE]);
1496 if (err) {
1497 qdisc_put_rtab(rtab);
1498 return err;
1499 }
1500 }
1501
1502 /* Change class parameters */
1503 sch_tree_lock(sch);
1504
1505 if (cl->next_alive != NULL)
1506 cbq_deactivate_class(cl);
1507
1508 if (rtab) {
1509 qdisc_put_rtab(cl->R_tab);
1510 cl->R_tab = rtab;
1511 }
1512
1513 if (tb[TCA_CBQ_LSSOPT])
1514 cbq_set_lss(cl, nla_data(tb[TCA_CBQ_LSSOPT]));
1515
1516 if (tb[TCA_CBQ_WRROPT]) {
1517 cbq_rmprio(q, cl);
1518 cbq_set_wrr(cl, nla_data(tb[TCA_CBQ_WRROPT]));
1519 }
1520
1521 if (tb[TCA_CBQ_FOPT])
1522 cbq_set_fopt(cl, nla_data(tb[TCA_CBQ_FOPT]));
1523
1524 if (cl->q->q.qlen)
1525 cbq_activate_class(cl);
1526
1527 sch_tree_unlock(sch);
1528
1529 return 0;
1530 }
1531
1532 if (parentid == TC_H_ROOT)
1533 return -EINVAL;
1534
1535 if (tb[TCA_CBQ_WRROPT] == NULL || tb[TCA_CBQ_RATE] == NULL ||
1536 tb[TCA_CBQ_LSSOPT] == NULL)
1537 return -EINVAL;
1538
1539 rtab = qdisc_get_rtab(nla_data(tb[TCA_CBQ_RATE]), tb[TCA_CBQ_RTAB]);
1540 if (rtab == NULL)
1541 return -EINVAL;
1542
1543 if (classid) {
1544 err = -EINVAL;
1545 if (TC_H_MAJ(classid ^ sch->handle) ||
1546 cbq_class_lookup(q, classid))
1547 goto failure;
1548 } else {
1549 int i;
1550 classid = TC_H_MAKE(sch->handle, 0x8000);
1551
1552 for (i = 0; i < 0x8000; i++) {
1553 if (++q->hgenerator >= 0x8000)
1554 q->hgenerator = 1;
1555 if (cbq_class_lookup(q, classid|q->hgenerator) == NULL)
1556 break;
1557 }
1558 err = -ENOSR;
1559 if (i >= 0x8000)
1560 goto failure;
1561 classid = classid|q->hgenerator;
1562 }
1563
1564 parent = &q->link;
1565 if (parentid) {
1566 parent = cbq_class_lookup(q, parentid);
1567 err = -EINVAL;
1568 if (parent == NULL)
1569 goto failure;
1570 }
1571
1572 err = -ENOBUFS;
1573 cl = kzalloc(sizeof(*cl), GFP_KERNEL);
1574 if (cl == NULL)
1575 goto failure;
1576
1577 err = tcf_block_get(&cl->block, &cl->filter_list, sch);
1578 if (err) {
1579 kfree(cl);
1580 return err;
1581 }
1582
1583 if (tca[TCA_RATE]) {
1584 err = gen_new_estimator(&cl->bstats, NULL, &cl->rate_est,
1585 NULL,
1586 qdisc_root_sleeping_running(sch),
1587 tca[TCA_RATE]);
1588 if (err) {
1589 tcf_block_put(cl->block);
1590 kfree(cl);
1591 goto failure;
1592 }
1593 }
1594
1595 cl->R_tab = rtab;
1596 rtab = NULL;
1597 cl->q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops, classid);
1598 if (!cl->q)
1599 cl->q = &noop_qdisc;
1600 else
1601 qdisc_hash_add(cl->q, true);
1602
1603 cl->common.classid = classid;
1604 cl->tparent = parent;
1605 cl->qdisc = sch;
1606 cl->allot = parent->allot;
1607 cl->quantum = cl->allot;
1608 cl->weight = cl->R_tab->rate.rate;
1609
1610 sch_tree_lock(sch);
1611 cbq_link_class(cl);
1612 cl->borrow = cl->tparent;
1613 if (cl->tparent != &q->link)
1614 cl->share = cl->tparent;
1615 cbq_adjust_levels(parent);
1616 cl->minidle = -0x7FFFFFFF;
1617 cbq_set_lss(cl, nla_data(tb[TCA_CBQ_LSSOPT]));
1618 cbq_set_wrr(cl, nla_data(tb[TCA_CBQ_WRROPT]));
1619 if (cl->ewma_log == 0)
1620 cl->ewma_log = q->link.ewma_log;
1621 if (cl->maxidle == 0)
1622 cl->maxidle = q->link.maxidle;
1623 if (cl->avpkt == 0)
1624 cl->avpkt = q->link.avpkt;
1625 if (tb[TCA_CBQ_FOPT])
1626 cbq_set_fopt(cl, nla_data(tb[TCA_CBQ_FOPT]));
1627 sch_tree_unlock(sch);
1628
1629 qdisc_class_hash_grow(sch, &q->clhash);
1630
1631 *arg = (unsigned long)cl;
1632 return 0;
1633
1634 failure:
1635 qdisc_put_rtab(rtab);
1636 return err;
1637 }
1638
1639 static int cbq_delete(struct Qdisc *sch, unsigned long arg)
1640 {
1641 struct cbq_sched_data *q = qdisc_priv(sch);
1642 struct cbq_class *cl = (struct cbq_class *)arg;
1643 unsigned int qlen, backlog;
1644
1645 if (cl->filters || cl->children || cl == &q->link)
1646 return -EBUSY;
1647
1648 sch_tree_lock(sch);
1649
1650 qlen = cl->q->q.qlen;
1651 backlog = cl->q->qstats.backlog;
1652 qdisc_reset(cl->q);
1653 qdisc_tree_reduce_backlog(cl->q, qlen, backlog);
1654
1655 if (cl->next_alive)
1656 cbq_deactivate_class(cl);
1657
1658 if (q->tx_borrowed == cl)
1659 q->tx_borrowed = q->tx_class;
1660 if (q->tx_class == cl) {
1661 q->tx_class = NULL;
1662 q->tx_borrowed = NULL;
1663 }
1664 #ifdef CONFIG_NET_CLS_ACT
1665 if (q->rx_class == cl)
1666 q->rx_class = NULL;
1667 #endif
1668
1669 cbq_unlink_class(cl);
1670 cbq_adjust_levels(cl->tparent);
1671 cl->defmap = 0;
1672 cbq_sync_defmap(cl);
1673
1674 cbq_rmprio(q, cl);
1675 sch_tree_unlock(sch);
1676
1677 cbq_destroy_class(sch, cl);
1678 return 0;
1679 }
1680
1681 static struct tcf_block *cbq_tcf_block(struct Qdisc *sch, unsigned long arg)
1682 {
1683 struct cbq_sched_data *q = qdisc_priv(sch);
1684 struct cbq_class *cl = (struct cbq_class *)arg;
1685
1686 if (cl == NULL)
1687 cl = &q->link;
1688
1689 return cl->block;
1690 }
1691
1692 static unsigned long cbq_bind_filter(struct Qdisc *sch, unsigned long parent,
1693 u32 classid)
1694 {
1695 struct cbq_sched_data *q = qdisc_priv(sch);
1696 struct cbq_class *p = (struct cbq_class *)parent;
1697 struct cbq_class *cl = cbq_class_lookup(q, classid);
1698
1699 if (cl) {
1700 if (p && p->level <= cl->level)
1701 return 0;
1702 cl->filters++;
1703 return (unsigned long)cl;
1704 }
1705 return 0;
1706 }
1707
1708 static void cbq_unbind_filter(struct Qdisc *sch, unsigned long arg)
1709 {
1710 struct cbq_class *cl = (struct cbq_class *)arg;
1711
1712 cl->filters--;
1713 }
1714
1715 static void cbq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1716 {
1717 struct cbq_sched_data *q = qdisc_priv(sch);
1718 struct cbq_class *cl;
1719 unsigned int h;
1720
1721 if (arg->stop)
1722 return;
1723
1724 for (h = 0; h < q->clhash.hashsize; h++) {
1725 hlist_for_each_entry(cl, &q->clhash.hash[h], common.hnode) {
1726 if (arg->count < arg->skip) {
1727 arg->count++;
1728 continue;
1729 }
1730 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1731 arg->stop = 1;
1732 return;
1733 }
1734 arg->count++;
1735 }
1736 }
1737 }
1738
1739 static const struct Qdisc_class_ops cbq_class_ops = {
1740 .graft = cbq_graft,
1741 .leaf = cbq_leaf,
1742 .qlen_notify = cbq_qlen_notify,
1743 .find = cbq_find,
1744 .change = cbq_change_class,
1745 .delete = cbq_delete,
1746 .walk = cbq_walk,
1747 .tcf_block = cbq_tcf_block,
1748 .bind_tcf = cbq_bind_filter,
1749 .unbind_tcf = cbq_unbind_filter,
1750 .dump = cbq_dump_class,
1751 .dump_stats = cbq_dump_class_stats,
1752 };
1753
1754 static struct Qdisc_ops cbq_qdisc_ops __read_mostly = {
1755 .next = NULL,
1756 .cl_ops = &cbq_class_ops,
1757 .id = "cbq",
1758 .priv_size = sizeof(struct cbq_sched_data),
1759 .enqueue = cbq_enqueue,
1760 .dequeue = cbq_dequeue,
1761 .peek = qdisc_peek_dequeued,
1762 .init = cbq_init,
1763 .reset = cbq_reset,
1764 .destroy = cbq_destroy,
1765 .change = NULL,
1766 .dump = cbq_dump,
1767 .dump_stats = cbq_dump_stats,
1768 .owner = THIS_MODULE,
1769 };
1770
1771 static int __init cbq_module_init(void)
1772 {
1773 return register_qdisc(&cbq_qdisc_ops);
1774 }
1775 static void __exit cbq_module_exit(void)
1776 {
1777 unregister_qdisc(&cbq_qdisc_ops);
1778 }
1779 module_init(cbq_module_init)
1780 module_exit(cbq_module_exit)
1781 MODULE_LICENSE("GPL");
1782
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.