2 * net/sched/sch_cbq.c Class-Based Queueing discipline.
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.
9 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
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>
25 /* Class-Based Queueing (CBQ) algorithm.
26 =======================================
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
32 [2] Sally Floyd, "Notes on CBQ and Guaranteed Service", 1995
34 [3] Sally Floyd, "Notes on Class-Based Queueing: Setting
37 [4] Sally Floyd and Michael Speer, "Experimental Results
38 for Class-Based Queueing", 1998, not published.
40 -----------------------------------------------------------------------
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:
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
55 delay_i <= ([MTU/(W*r_i)]*W*r + W*r + k*MTU)/B
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.
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).
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. */
73 struct cbq_sched_data;
77 struct Qdisc_class_common common;
78 struct cbq_class *next_alive; /* next class with backlog in this priority band */
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 */
87 /* Link-sharing scheduler parameters */
88 long maxidle; /* Class parameters: see below. */
92 struct qdisc_rate_table *R_tab;
94 /* General scheduler (WRR) parameters */
96 long quantum; /* Allotment per WRR round */
97 long weight; /* Relative allotment: see below */
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;
105 struct cbq_class *sibling; /* Sibling chain */
106 struct cbq_class *children; /* Pointer to children chain */
108 struct Qdisc *q; /* Elementary queueing discipline */
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.
119 psched_time_t last; /* Last end of service */
120 psched_time_t undertime;
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;
129 struct tcf_proto __rcu *filter_list;
130 struct tcf_block *block;
135 struct cbq_class *defaults[TC_PRIO_MAX + 1];
138 struct cbq_sched_data {
139 struct Qdisc_class_hash clhash; /* Hash table of all classes */
140 int nclasses[TC_CBQ_MAXPRIO + 1];
141 unsigned int quanta[TC_CBQ_MAXPRIO + 1];
143 struct cbq_class link;
145 unsigned int activemask;
146 struct cbq_class *active[TC_CBQ_MAXPRIO + 1]; /* List of all classes
149 #ifdef CONFIG_NET_CLS_ACT
150 struct cbq_class *rx_class;
152 struct cbq_class *tx_class;
153 struct cbq_class *tx_borrowed;
155 psched_time_t now; /* Cached timestamp */
158 struct hrtimer delay_timer;
159 struct qdisc_watchdog watchdog; /* Watchdog timer,
163 psched_tdiff_t wd_expires;
169 #define L2T(cl, len) qdisc_l2t((cl)->R_tab, len)
171 static inline struct cbq_class *
172 cbq_class_lookup(struct cbq_sched_data *q, u32 classid)
174 struct Qdisc_class_common *clc;
176 clc = qdisc_class_find(&q->clhash, classid);
179 return container_of(clc, struct cbq_class, common);
182 #ifdef CONFIG_NET_CLS_ACT
184 static struct cbq_class *
185 cbq_reclassify(struct sk_buff *skb, struct cbq_class *this)
187 struct cbq_class *cl;
189 for (cl = this->tparent; cl; cl = cl->tparent) {
190 struct cbq_class *new = cl->defaults[TC_PRIO_BESTEFFORT];
192 if (new != NULL && new != this)
200 /* Classify packet. The procedure is pretty complicated, but
201 * it allows us to combine link sharing and priority scheduling
204 * Namely, you can put link sharing rules (f.e. route based) at root of CBQ,
205 * so that it resolves to split nodes. Then packets are classified
206 * by logical priority, or a more specific classifier may be attached
210 static struct cbq_class *
211 cbq_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
213 struct cbq_sched_data *q = qdisc_priv(sch);
214 struct cbq_class *head = &q->link;
215 struct cbq_class **defmap;
216 struct cbq_class *cl = NULL;
217 u32 prio = skb->priority;
218 struct tcf_proto *fl;
219 struct tcf_result res;
222 * Step 1. If skb->priority points to one of our classes, use it.
224 if (TC_H_MAJ(prio ^ sch->handle) == 0 &&
225 (cl = cbq_class_lookup(q, prio)) != NULL)
228 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
231 defmap = head->defaults;
233 fl = rcu_dereference_bh(head->filter_list);
235 * Step 2+n. Apply classifier.
237 result = tcf_classify(skb, fl, &res, true);
238 if (!fl || result < 0)
241 cl = (void *)res.class;
243 if (TC_H_MAJ(res.classid))
244 cl = cbq_class_lookup(q, res.classid);
245 else if ((cl = defmap[res.classid & TC_PRIO_MAX]) == NULL)
246 cl = defmap[TC_PRIO_BESTEFFORT];
251 if (cl->level >= head->level)
253 #ifdef CONFIG_NET_CLS_ACT
258 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
261 case TC_ACT_RECLASSIFY:
262 return cbq_reclassify(skb, cl);
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.
280 * Step 4. No success...
282 if (TC_H_MAJ(prio) == 0 &&
283 !(cl = head->defaults[prio & TC_PRIO_MAX]) &&
284 !(cl = head->defaults[TC_PRIO_BESTEFFORT]))
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.
296 static inline void cbq_activate_class(struct cbq_class *cl)
298 struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
299 int prio = cl->cpriority;
300 struct cbq_class *cl_tail;
302 cl_tail = q->active[prio];
303 q->active[prio] = cl;
305 if (cl_tail != NULL) {
306 cl->next_alive = cl_tail->next_alive;
307 cl_tail->next_alive = cl;
310 q->activemask |= (1<<prio);
315 * Unlink class from active chain.
316 * Note that this same procedure is done directly in cbq_dequeue*
317 * during round-robin procedure.
320 static void cbq_deactivate_class(struct cbq_class *this)
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];
328 cl = cl_prev->next_alive;
330 cl_prev->next_alive = cl->next_alive;
331 cl->next_alive = NULL;
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);
343 } while ((cl_prev = cl) != q->active[prio]);
347 cbq_mark_toplevel(struct cbq_sched_data *q, struct cbq_class *cl)
349 int toplevel = q->toplevel;
351 if (toplevel > cl->level) {
352 psched_time_t now = psched_get_time();
355 if (cl->undertime < now) {
356 q->toplevel = cl->level;
359 } while ((cl = cl->borrow) != NULL && toplevel > cl->level);
364 cbq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
365 struct sk_buff **to_free)
367 struct cbq_sched_data *q = qdisc_priv(sch);
368 int uninitialized_var(ret);
369 struct cbq_class *cl = cbq_classify(skb, sch, &ret);
371 #ifdef CONFIG_NET_CLS_ACT
375 if (ret & __NET_XMIT_BYPASS)
376 qdisc_qstats_drop(sch);
377 __qdisc_drop(skb, to_free);
381 ret = qdisc_enqueue(skb, cl->q, to_free);
382 if (ret == NET_XMIT_SUCCESS) {
384 cbq_mark_toplevel(q, cl);
386 cbq_activate_class(cl);
390 if (net_xmit_drop_count(ret)) {
391 qdisc_qstats_drop(sch);
392 cbq_mark_toplevel(q, cl);
398 /* Overlimit action: penalize leaf class by adding offtime */
399 static void cbq_overlimit(struct cbq_class *cl)
401 struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
402 psched_tdiff_t delay = cl->undertime - q->now;
405 delay += cl->offtime;
408 * Class goes to sleep, so that it will have no
409 * chance to work avgidle. Let's forgive it 8)
411 * BTW cbq-2.0 has a crap in this
412 * place, apparently they forgot to shift it by cl->ewma_log.
415 delay -= (-cl->avgidle) - ((-cl->avgidle) >> cl->ewma_log);
416 if (cl->avgidle < cl->minidle)
417 cl->avgidle = cl->minidle;
420 cl->undertime = q->now + delay;
422 cl->xstats.overactions++;
425 if (q->wd_expires == 0 || q->wd_expires > delay)
426 q->wd_expires = delay;
428 /* Dirty work! We must schedule wakeups based on
429 * real available rate, rather than leaf rate,
430 * which may be tiny (even zero).
432 if (q->toplevel == TC_CBQ_MAXLEVEL) {
434 psched_tdiff_t base_delay = q->wd_expires;
436 for (b = cl->borrow; b; b = b->borrow) {
437 delay = b->undertime - q->now;
438 if (delay < base_delay) {
445 q->wd_expires = base_delay;
449 static psched_tdiff_t cbq_undelay_prio(struct cbq_sched_data *q, int prio,
452 struct cbq_class *cl;
453 struct cbq_class *cl_prev = q->active[prio];
454 psched_time_t sched = now;
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;
466 cbq_activate_class(cl);
468 if (cl == q->active[prio]) {
469 q->active[prio] = cl_prev;
470 if (cl == q->active[prio]) {
471 q->active[prio] = NULL;
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]);
484 static enum hrtimer_restart cbq_undelay(struct hrtimer *timer)
486 struct cbq_sched_data *q = container_of(timer, struct cbq_sched_data,
488 struct Qdisc *sch = q->watchdog.qdisc;
490 psched_tdiff_t delay = 0;
493 now = psched_get_time();
499 int prio = ffz(~pmask);
504 tmp = cbq_undelay_prio(q, prio, now);
507 if (tmp < delay || delay == 0)
516 time = ktime_add_ns(time, PSCHED_TICKS2NS(now + delay));
517 hrtimer_start(&q->delay_timer, time, HRTIMER_MODE_ABS_PINNED);
520 __netif_schedule(qdisc_root(sch));
521 return HRTIMER_NORESTART;
525 * It is mission critical procedure.
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.
534 cbq_update_toplevel(struct cbq_sched_data *q, struct cbq_class *cl,
535 struct cbq_class *borrowed)
537 if (cl && q->toplevel >= borrowed->level) {
538 if (cl->q->q.qlen > 1) {
540 if (borrowed->undertime == PSCHED_PASTPERFECT) {
541 q->toplevel = borrowed->level;
544 } while ((borrowed = borrowed->borrow) != NULL);
547 /* It is not necessary now. Uncommenting it
548 will save CPU cycles, but decrease fairness.
550 q->toplevel = TC_CBQ_MAXLEVEL;
556 cbq_update(struct cbq_sched_data *q)
558 struct cbq_class *this = q->tx_class;
559 struct cbq_class *cl = this;
564 /* Time integrator. We calculate EOS time
565 * by adding expected packet transmission time.
567 now = q->now + L2T(&q->link, len);
569 for ( ; cl; cl = cl->share) {
570 long avgidle = cl->avgidle;
573 cl->bstats.packets++;
574 cl->bstats.bytes += len;
577 * (now - last) is total time between packet right edges.
578 * (last_pktlen/rate) is "virtual" busy time, so that
580 * idle = (now - last) - last_pktlen/rate
583 idle = now - cl->last;
584 if ((unsigned long)idle > 128*1024*1024) {
585 avgidle = cl->maxidle;
587 idle -= L2T(cl, len);
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,
594 avgidle += idle - (avgidle>>cl->ewma_log);
598 /* Overlimit or at-limit */
600 if (avgidle < cl->minidle)
601 avgidle = cl->minidle;
603 cl->avgidle = avgidle;
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)
611 * idle = (1 - W)*(-cl->avgidle);
613 idle = (-avgidle) - ((-avgidle) >> cl->ewma_log);
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)
624 idle -= L2T(&q->link, len);
625 idle += L2T(cl, len);
627 cl->undertime = now + idle;
631 cl->undertime = PSCHED_PASTPERFECT;
632 if (avgidle > cl->maxidle)
633 cl->avgidle = cl->maxidle;
635 cl->avgidle = avgidle;
637 if ((s64)(now - cl->last) > 0)
641 cbq_update_toplevel(q, this, q->tx_borrowed);
644 static inline struct cbq_class *
645 cbq_under_limit(struct cbq_class *cl)
647 struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
648 struct cbq_class *this_cl = cl;
650 if (cl->tparent == NULL)
653 if (cl->undertime == PSCHED_PASTPERFECT || q->now >= cl->undertime) {
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.
671 this_cl->qstats.overlimits++;
672 cbq_overlimit(this_cl);
675 if (cl->level > q->toplevel)
677 } while (cl->undertime != PSCHED_PASTPERFECT && q->now < cl->undertime);
683 static inline struct sk_buff *
684 cbq_dequeue_prio(struct Qdisc *sch, int prio)
686 struct cbq_sched_data *q = qdisc_priv(sch);
687 struct cbq_class *cl_tail, *cl_prev, *cl;
691 cl_tail = cl_prev = q->active[prio];
692 cl = cl_prev->next_alive;
699 struct cbq_class *borrow = cl;
702 (borrow = cbq_under_limit(cl)) == NULL)
705 if (cl->deficit <= 0) {
706 /* Class exhausted its allotment per
707 * this round. Switch to the next one.
710 cl->deficit += cl->quantum;
714 skb = cl->q->dequeue(cl->q);
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"
723 cl->deficit -= qdisc_pkt_len(skb);
725 q->tx_borrowed = borrow;
727 #ifndef CBQ_XSTATS_BORROWS_BYTES
728 borrow->xstats.borrows++;
729 cl->xstats.borrows++;
731 borrow->xstats.borrows += qdisc_pkt_len(skb);
732 cl->xstats.borrows += qdisc_pkt_len(skb);
735 q->tx_len = qdisc_pkt_len(skb);
737 if (cl->deficit <= 0) {
738 q->active[prio] = cl;
740 cl->deficit += cl->quantum;
745 if (cl->q->q.qlen == 0 || prio != cl->cpriority) {
746 /* Class is empty or penalized.
747 * Unlink it from active chain.
749 cl_prev->next_alive = cl->next_alive;
750 cl->next_alive = NULL;
752 /* Did cl_tail point to it? */
757 /* Was it the last class in this band? */
760 q->active[prio] = NULL;
761 q->activemask &= ~(1<<prio);
763 cbq_activate_class(cl);
767 q->active[prio] = cl_tail;
770 cbq_activate_class(cl);
778 } while (cl_prev != cl_tail);
781 q->active[prio] = cl_prev;
786 static inline struct sk_buff *
787 cbq_dequeue_1(struct Qdisc *sch)
789 struct cbq_sched_data *q = qdisc_priv(sch);
791 unsigned int activemask;
793 activemask = q->activemask & 0xFF;
795 int prio = ffz(~activemask);
796 activemask &= ~(1<<prio);
797 skb = cbq_dequeue_prio(sch, prio);
804 static struct sk_buff *
805 cbq_dequeue(struct Qdisc *sch)
808 struct cbq_sched_data *q = qdisc_priv(sch);
811 now = psched_get_time();
821 skb = cbq_dequeue_1(sch);
823 qdisc_bstats_update(sch, skb);
828 /* All the classes are overlimit.
830 * It is possible, if:
832 * 1. Scheduler is empty.
833 * 2. Toplevel cutoff inhibited borrowing.
834 * 3. Root class is overlimit.
836 * Reset 2d and 3d conditions and retry.
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.
842 * Our version is better, but slower, because it requires
843 * two passes, but it is unavoidable with top-level sharing.
846 if (q->toplevel == TC_CBQ_MAXLEVEL &&
847 q->link.undertime == PSCHED_PASTPERFECT)
850 q->toplevel = TC_CBQ_MAXLEVEL;
851 q->link.undertime = PSCHED_PASTPERFECT;
854 /* No packets in scheduler or nobody wants to give them to us :-(
855 * Sigh... start watchdog timer in the last case.
859 qdisc_qstats_overlimit(sch);
861 qdisc_watchdog_schedule(&q->watchdog,
862 now + q->wd_expires);
867 /* CBQ class maintanance routines */
869 static void cbq_adjust_levels(struct cbq_class *this)
876 struct cbq_class *cl;
881 if (cl->level > level)
883 } while ((cl = cl->sibling) != this->children);
885 this->level = level + 1;
886 } while ((this = this->tparent) != NULL);
889 static void cbq_normalize_quanta(struct cbq_sched_data *q, int prio)
891 struct cbq_class *cl;
894 if (q->quanta[prio] == 0)
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!
902 if (cl->priority == prio) {
903 cl->quantum = (cl->weight*cl->allot*q->nclasses[prio])/
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;
916 static void cbq_sync_defmap(struct cbq_class *cl)
918 struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
919 struct cbq_class *split = cl->split;
926 for (i = 0; i <= TC_PRIO_MAX; i++) {
927 if (split->defaults[i] == cl && !(cl->defmap & (1<<i)))
928 split->defaults[i] = NULL;
931 for (i = 0; i <= TC_PRIO_MAX; i++) {
932 int level = split->level;
934 if (split->defaults[i])
937 for (h = 0; h < q->clhash.hashsize; h++) {
940 hlist_for_each_entry(c, &q->clhash.hash[h],
942 if (c->split == split && c->level < level &&
943 c->defmap & (1<<i)) {
944 split->defaults[i] = c;
952 static void cbq_change_defmap(struct cbq_class *cl, u32 splitid, u32 def, u32 mask)
954 struct cbq_class *split = NULL;
960 splitid = split->common.classid;
963 if (split == NULL || split->common.classid != splitid) {
964 for (split = cl->tparent; split; split = split->tparent)
965 if (split->common.classid == splitid)
972 if (cl->split != split) {
976 cl->defmap = def & mask;
978 cl->defmap = (cl->defmap & ~mask) | (def & mask);
983 static void cbq_unlink_class(struct cbq_class *this)
985 struct cbq_class *cl, **clp;
986 struct cbq_sched_data *q = qdisc_priv(this->qdisc);
988 qdisc_class_hash_remove(&q->clhash, &this->common);
991 clp = &this->sibling;
999 } while ((cl = *clp) != this->sibling);
1001 if (this->tparent->children == this) {
1002 this->tparent->children = this->sibling;
1003 if (this->sibling == this)
1004 this->tparent->children = NULL;
1007 WARN_ON(this->sibling != this);
1011 static void cbq_link_class(struct cbq_class *this)
1013 struct cbq_sched_data *q = qdisc_priv(this->qdisc);
1014 struct cbq_class *parent = this->tparent;
1016 this->sibling = this;
1017 qdisc_class_hash_insert(&q->clhash, &this->common);
1022 if (parent->children == NULL) {
1023 parent->children = this;
1025 this->sibling = parent->children->sibling;
1026 parent->children->sibling = this;
1031 cbq_reset(struct Qdisc *sch)
1033 struct cbq_sched_data *q = qdisc_priv(sch);
1034 struct cbq_class *cl;
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();
1047 for (prio = 0; prio <= TC_CBQ_MAXPRIO; prio++)
1048 q->active[prio] = NULL;
1050 for (h = 0; h < q->clhash.hashsize; h++) {
1051 hlist_for_each_entry(cl, &q->clhash.hash[h], common.hnode) {
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;
1065 static int cbq_set_lss(struct cbq_class *cl, struct tc_cbq_lssopt *lss)
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;
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;
1081 if (lss->change & TCF_CBQ_LSS_OFFTIME)
1082 cl->offtime = lss->offtime;
1086 static void cbq_rmprio(struct cbq_sched_data *q, struct cbq_class *cl)
1088 q->nclasses[cl->priority]--;
1089 q->quanta[cl->priority] -= cl->weight;
1090 cbq_normalize_quanta(q, cl->priority);
1093 static void cbq_addprio(struct cbq_sched_data *q, struct cbq_class *cl)
1095 q->nclasses[cl->priority]++;
1096 q->quanta[cl->priority] += cl->weight;
1097 cbq_normalize_quanta(q, cl->priority);
1100 static int cbq_set_wrr(struct cbq_class *cl, struct tc_cbq_wrropt *wrr)
1102 struct cbq_sched_data *q = qdisc_priv(cl->qdisc);
1105 cl->allot = wrr->allot;
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;
1119 static int cbq_set_fopt(struct cbq_class *cl, struct tc_cbq_fopt *fopt)
1121 cbq_change_defmap(cl, fopt->split, fopt->defmap, fopt->defchange);
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) },
1135 static int cbq_init(struct Qdisc *sch, struct nlattr *opt)
1137 struct cbq_sched_data *q = qdisc_priv(sch);
1138 struct nlattr *tb[TCA_CBQ_MAX + 1];
1139 struct tc_ratespec *r;
1142 err = nla_parse_nested(tb, TCA_CBQ_MAX, opt, cbq_policy, NULL);
1146 if (tb[TCA_CBQ_RTAB] == NULL || tb[TCA_CBQ_RATE] == NULL)
1149 r = nla_data(tb[TCA_CBQ_RATE]);
1151 if ((q->link.R_tab = qdisc_get_rtab(r, tb[TCA_CBQ_RTAB])) == NULL)
1154 err = qdisc_class_hash_init(&q->clhash);
1159 q->link.sibling = &q->link;
1160 q->link.common.classid = sch->handle;
1161 q->link.qdisc = sch;
1162 q->link.q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1165 q->link.q = &noop_qdisc;
1167 qdisc_hash_add(q->link.q, true);
1169 q->link.priority = TC_CBQ_MAXPRIO - 1;
1170 q->link.priority2 = TC_CBQ_MAXPRIO - 1;
1171 q->link.cpriority = TC_CBQ_MAXPRIO - 1;
1172 q->link.allot = psched_mtu(qdisc_dev(sch));
1173 q->link.quantum = q->link.allot;
1174 q->link.weight = q->link.R_tab->rate.rate;
1176 q->link.ewma_log = TC_CBQ_DEF_EWMA;
1177 q->link.avpkt = q->link.allot/2;
1178 q->link.minidle = -0x7FFFFFFF;
1180 qdisc_watchdog_init(&q->watchdog, sch);
1181 hrtimer_init(&q->delay_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
1182 q->delay_timer.function = cbq_undelay;
1183 q->toplevel = TC_CBQ_MAXLEVEL;
1184 q->now = psched_get_time();
1186 cbq_link_class(&q->link);
1188 if (tb[TCA_CBQ_LSSOPT])
1189 cbq_set_lss(&q->link, nla_data(tb[TCA_CBQ_LSSOPT]));
1191 cbq_addprio(q, &q->link);
1195 qdisc_put_rtab(q->link.R_tab);
1199 static int cbq_dump_rate(struct sk_buff *skb, struct cbq_class *cl)
1201 unsigned char *b = skb_tail_pointer(skb);
1203 if (nla_put(skb, TCA_CBQ_RATE, sizeof(cl->R_tab->rate), &cl->R_tab->rate))
1204 goto nla_put_failure;
1212 static int cbq_dump_lss(struct sk_buff *skb, struct cbq_class *cl)
1214 unsigned char *b = skb_tail_pointer(skb);
1215 struct tc_cbq_lssopt opt;
1218 if (cl->borrow == NULL)
1219 opt.flags |= TCF_CBQ_LSS_BOUNDED;
1220 if (cl->share == NULL)
1221 opt.flags |= TCF_CBQ_LSS_ISOLATED;
1222 opt.ewma_log = cl->ewma_log;
1223 opt.level = cl->level;
1224 opt.avpkt = cl->avpkt;
1225 opt.maxidle = cl->maxidle;
1226 opt.minidle = (u32)(-cl->minidle);
1227 opt.offtime = cl->offtime;
1229 if (nla_put(skb, TCA_CBQ_LSSOPT, sizeof(opt), &opt))
1230 goto nla_put_failure;
1238 static int cbq_dump_wrr(struct sk_buff *skb, struct cbq_class *cl)
1240 unsigned char *b = skb_tail_pointer(skb);
1241 struct tc_cbq_wrropt opt;
1243 memset(&opt, 0, sizeof(opt));
1245 opt.allot = cl->allot;
1246 opt.priority = cl->priority + 1;
1247 opt.cpriority = cl->cpriority + 1;
1248 opt.weight = cl->weight;
1249 if (nla_put(skb, TCA_CBQ_WRROPT, sizeof(opt), &opt))
1250 goto nla_put_failure;
1258 static int cbq_dump_fopt(struct sk_buff *skb, struct cbq_class *cl)
1260 unsigned char *b = skb_tail_pointer(skb);
1261 struct tc_cbq_fopt opt;
1263 if (cl->split || cl->defmap) {
1264 opt.split = cl->split ? cl->split->common.classid : 0;
1265 opt.defmap = cl->defmap;
1267 if (nla_put(skb, TCA_CBQ_FOPT, sizeof(opt), &opt))
1268 goto nla_put_failure;
1277 static int cbq_dump_attr(struct sk_buff *skb, struct cbq_class *cl)
1279 if (cbq_dump_lss(skb, cl) < 0 ||
1280 cbq_dump_rate(skb, cl) < 0 ||
1281 cbq_dump_wrr(skb, cl) < 0 ||
1282 cbq_dump_fopt(skb, cl) < 0)
1287 static int cbq_dump(struct Qdisc *sch, struct sk_buff *skb)
1289 struct cbq_sched_data *q = qdisc_priv(sch);
1290 struct nlattr *nest;
1292 nest = nla_nest_start(skb, TCA_OPTIONS);
1294 goto nla_put_failure;
1295 if (cbq_dump_attr(skb, &q->link) < 0)
1296 goto nla_put_failure;
1297 return nla_nest_end(skb, nest);
1300 nla_nest_cancel(skb, nest);
1305 cbq_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
1307 struct cbq_sched_data *q = qdisc_priv(sch);
1309 q->link.xstats.avgidle = q->link.avgidle;
1310 return gnet_stats_copy_app(d, &q->link.xstats, sizeof(q->link.xstats));
1314 cbq_dump_class(struct Qdisc *sch, unsigned long arg,
1315 struct sk_buff *skb, struct tcmsg *tcm)
1317 struct cbq_class *cl = (struct cbq_class *)arg;
1318 struct nlattr *nest;
1321 tcm->tcm_parent = cl->tparent->common.classid;
1323 tcm->tcm_parent = TC_H_ROOT;
1324 tcm->tcm_handle = cl->common.classid;
1325 tcm->tcm_info = cl->q->handle;
1327 nest = nla_nest_start(skb, TCA_OPTIONS);
1329 goto nla_put_failure;
1330 if (cbq_dump_attr(skb, cl) < 0)
1331 goto nla_put_failure;
1332 return nla_nest_end(skb, nest);
1335 nla_nest_cancel(skb, nest);
1340 cbq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1341 struct gnet_dump *d)
1343 struct cbq_sched_data *q = qdisc_priv(sch);
1344 struct cbq_class *cl = (struct cbq_class *)arg;
1346 cl->xstats.avgidle = cl->avgidle;
1347 cl->xstats.undertime = 0;
1349 if (cl->undertime != PSCHED_PASTPERFECT)
1350 cl->xstats.undertime = cl->undertime - q->now;
1352 if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
1353 d, NULL, &cl->bstats) < 0 ||
1354 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1355 gnet_stats_copy_queue(d, NULL, &cl->qstats, cl->q->q.qlen) < 0)
1358 return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1361 static int cbq_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1364 struct cbq_class *cl = (struct cbq_class *)arg;
1367 new = qdisc_create_dflt(sch->dev_queue,
1368 &pfifo_qdisc_ops, cl->common.classid);
1373 *old = qdisc_replace(sch, new, &cl->q);
1377 static struct Qdisc *cbq_leaf(struct Qdisc *sch, unsigned long arg)
1379 struct cbq_class *cl = (struct cbq_class *)arg;
1384 static void cbq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1386 struct cbq_class *cl = (struct cbq_class *)arg;
1388 if (cl->q->q.qlen == 0)
1389 cbq_deactivate_class(cl);
1392 static unsigned long cbq_get(struct Qdisc *sch, u32 classid)
1394 struct cbq_sched_data *q = qdisc_priv(sch);
1395 struct cbq_class *cl = cbq_class_lookup(q, classid);
1399 return (unsigned long)cl;
1404 static void cbq_destroy_class(struct Qdisc *sch, struct cbq_class *cl)
1406 struct cbq_sched_data *q = qdisc_priv(sch);
1408 WARN_ON(cl->filters);
1410 tcf_block_put(cl->block);
1411 qdisc_destroy(cl->q);
1412 qdisc_put_rtab(cl->R_tab);
1413 gen_kill_estimator(&cl->rate_est);
1418 static void cbq_destroy(struct Qdisc *sch)
1420 struct cbq_sched_data *q = qdisc_priv(sch);
1421 struct hlist_node *next;
1422 struct cbq_class *cl;
1425 #ifdef CONFIG_NET_CLS_ACT
1429 * Filters must be destroyed first because we don't destroy the
1430 * classes from root to leafs which means that filters can still
1431 * be bound to classes which have been destroyed already. --TGR '04
1433 for (h = 0; h < q->clhash.hashsize; h++) {
1434 hlist_for_each_entry(cl, &q->clhash.hash[h], common.hnode)
1435 tcf_block_put(cl->block);
1437 for (h = 0; h < q->clhash.hashsize; h++) {
1438 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[h],
1440 cbq_destroy_class(sch, cl);
1442 qdisc_class_hash_destroy(&q->clhash);
1445 static void cbq_put(struct Qdisc *sch, unsigned long arg)
1447 struct cbq_class *cl = (struct cbq_class *)arg;
1449 if (--cl->refcnt == 0) {
1450 #ifdef CONFIG_NET_CLS_ACT
1451 spinlock_t *root_lock = qdisc_root_sleeping_lock(sch);
1452 struct cbq_sched_data *q = qdisc_priv(sch);
1454 spin_lock_bh(root_lock);
1455 if (q->rx_class == cl)
1457 spin_unlock_bh(root_lock);
1460 cbq_destroy_class(sch, cl);
1465 cbq_change_class(struct Qdisc *sch, u32 classid, u32 parentid, struct nlattr **tca,
1469 struct cbq_sched_data *q = qdisc_priv(sch);
1470 struct cbq_class *cl = (struct cbq_class *)*arg;
1471 struct nlattr *opt = tca[TCA_OPTIONS];
1472 struct nlattr *tb[TCA_CBQ_MAX + 1];
1473 struct cbq_class *parent;
1474 struct qdisc_rate_table *rtab = NULL;
1479 err = nla_parse_nested(tb, TCA_CBQ_MAX, opt, cbq_policy, NULL);
1483 if (tb[TCA_CBQ_OVL_STRATEGY] || tb[TCA_CBQ_POLICE])
1490 cl->tparent->common.classid != parentid)
1492 if (!cl->tparent && parentid != TC_H_ROOT)
1496 if (tb[TCA_CBQ_RATE]) {
1497 rtab = qdisc_get_rtab(nla_data(tb[TCA_CBQ_RATE]),
1503 if (tca[TCA_RATE]) {
1504 err = gen_replace_estimator(&cl->bstats, NULL,
1507 qdisc_root_sleeping_running(sch),
1510 qdisc_put_rtab(rtab);
1515 /* Change class parameters */
1518 if (cl->next_alive != NULL)
1519 cbq_deactivate_class(cl);
1522 qdisc_put_rtab(cl->R_tab);
1526 if (tb[TCA_CBQ_LSSOPT])
1527 cbq_set_lss(cl, nla_data(tb[TCA_CBQ_LSSOPT]));
1529 if (tb[TCA_CBQ_WRROPT]) {
1531 cbq_set_wrr(cl, nla_data(tb[TCA_CBQ_WRROPT]));
1534 if (tb[TCA_CBQ_FOPT])
1535 cbq_set_fopt(cl, nla_data(tb[TCA_CBQ_FOPT]));
1538 cbq_activate_class(cl);
1540 sch_tree_unlock(sch);
1545 if (parentid == TC_H_ROOT)
1548 if (tb[TCA_CBQ_WRROPT] == NULL || tb[TCA_CBQ_RATE] == NULL ||
1549 tb[TCA_CBQ_LSSOPT] == NULL)
1552 rtab = qdisc_get_rtab(nla_data(tb[TCA_CBQ_RATE]), tb[TCA_CBQ_RTAB]);
1558 if (TC_H_MAJ(classid ^ sch->handle) ||
1559 cbq_class_lookup(q, classid))
1563 classid = TC_H_MAKE(sch->handle, 0x8000);
1565 for (i = 0; i < 0x8000; i++) {
1566 if (++q->hgenerator >= 0x8000)
1568 if (cbq_class_lookup(q, classid|q->hgenerator) == NULL)
1574 classid = classid|q->hgenerator;
1579 parent = cbq_class_lookup(q, parentid);
1586 cl = kzalloc(sizeof(*cl), GFP_KERNEL);
1590 err = tcf_block_get(&cl->block, &cl->filter_list);
1596 if (tca[TCA_RATE]) {
1597 err = gen_new_estimator(&cl->bstats, NULL, &cl->rate_est,
1599 qdisc_root_sleeping_running(sch),
1602 tcf_block_put(cl->block);
1611 cl->q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops, classid);
1613 cl->q = &noop_qdisc;
1615 qdisc_hash_add(cl->q, true);
1617 cl->common.classid = classid;
1618 cl->tparent = parent;
1620 cl->allot = parent->allot;
1621 cl->quantum = cl->allot;
1622 cl->weight = cl->R_tab->rate.rate;
1626 cl->borrow = cl->tparent;
1627 if (cl->tparent != &q->link)
1628 cl->share = cl->tparent;
1629 cbq_adjust_levels(parent);
1630 cl->minidle = -0x7FFFFFFF;
1631 cbq_set_lss(cl, nla_data(tb[TCA_CBQ_LSSOPT]));
1632 cbq_set_wrr(cl, nla_data(tb[TCA_CBQ_WRROPT]));
1633 if (cl->ewma_log == 0)
1634 cl->ewma_log = q->link.ewma_log;
1635 if (cl->maxidle == 0)
1636 cl->maxidle = q->link.maxidle;
1638 cl->avpkt = q->link.avpkt;
1639 if (tb[TCA_CBQ_FOPT])
1640 cbq_set_fopt(cl, nla_data(tb[TCA_CBQ_FOPT]));
1641 sch_tree_unlock(sch);
1643 qdisc_class_hash_grow(sch, &q->clhash);
1645 *arg = (unsigned long)cl;
1649 qdisc_put_rtab(rtab);
1653 static int cbq_delete(struct Qdisc *sch, unsigned long arg)
1655 struct cbq_sched_data *q = qdisc_priv(sch);
1656 struct cbq_class *cl = (struct cbq_class *)arg;
1657 unsigned int qlen, backlog;
1659 if (cl->filters || cl->children || cl == &q->link)
1664 qlen = cl->q->q.qlen;
1665 backlog = cl->q->qstats.backlog;
1667 qdisc_tree_reduce_backlog(cl->q, qlen, backlog);
1670 cbq_deactivate_class(cl);
1672 if (q->tx_borrowed == cl)
1673 q->tx_borrowed = q->tx_class;
1674 if (q->tx_class == cl) {
1676 q->tx_borrowed = NULL;
1678 #ifdef CONFIG_NET_CLS_ACT
1679 if (q->rx_class == cl)
1683 cbq_unlink_class(cl);
1684 cbq_adjust_levels(cl->tparent);
1686 cbq_sync_defmap(cl);
1689 sch_tree_unlock(sch);
1691 BUG_ON(--cl->refcnt == 0);
1693 * This shouldn't happen: we "hold" one cops->get() when called
1694 * from tc_ctl_tclass; the destroy method is done from cops->put().
1700 static struct tcf_block *cbq_tcf_block(struct Qdisc *sch, unsigned long arg)
1702 struct cbq_sched_data *q = qdisc_priv(sch);
1703 struct cbq_class *cl = (struct cbq_class *)arg;
1711 static unsigned long cbq_bind_filter(struct Qdisc *sch, unsigned long parent,
1714 struct cbq_sched_data *q = qdisc_priv(sch);
1715 struct cbq_class *p = (struct cbq_class *)parent;
1716 struct cbq_class *cl = cbq_class_lookup(q, classid);
1719 if (p && p->level <= cl->level)
1722 return (unsigned long)cl;
1727 static void cbq_unbind_filter(struct Qdisc *sch, unsigned long arg)
1729 struct cbq_class *cl = (struct cbq_class *)arg;
1734 static void cbq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1736 struct cbq_sched_data *q = qdisc_priv(sch);
1737 struct cbq_class *cl;
1743 for (h = 0; h < q->clhash.hashsize; h++) {
1744 hlist_for_each_entry(cl, &q->clhash.hash[h], common.hnode) {
1745 if (arg->count < arg->skip) {
1749 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1758 static const struct Qdisc_class_ops cbq_class_ops = {
1761 .qlen_notify = cbq_qlen_notify,
1764 .change = cbq_change_class,
1765 .delete = cbq_delete,
1767 .tcf_block = cbq_tcf_block,
1768 .bind_tcf = cbq_bind_filter,
1769 .unbind_tcf = cbq_unbind_filter,
1770 .dump = cbq_dump_class,
1771 .dump_stats = cbq_dump_class_stats,
1774 static struct Qdisc_ops cbq_qdisc_ops __read_mostly = {
1776 .cl_ops = &cbq_class_ops,
1778 .priv_size = sizeof(struct cbq_sched_data),
1779 .enqueue = cbq_enqueue,
1780 .dequeue = cbq_dequeue,
1781 .peek = qdisc_peek_dequeued,
1784 .destroy = cbq_destroy,
1787 .dump_stats = cbq_dump_stats,
1788 .owner = THIS_MODULE,
1791 static int __init cbq_module_init(void)
1793 return register_qdisc(&cbq_qdisc_ops);
1795 static void __exit cbq_module_exit(void)
1797 unregister_qdisc(&cbq_qdisc_ops);
1799 module_init(cbq_module_init)
1800 module_exit(cbq_module_exit)
1801 MODULE_LICENSE("GPL");