2 * net/sched/sch_htb.c Hierarchical token bucket, feed tree version
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: Martin Devera, <devik@cdi.cz>
11 * Credits (in time order) for older HTB versions:
12 * Stef Coene <stef.coene@docum.org>
13 * HTB support at LARTC mailing list
14 * Ondrej Kraus, <krauso@barr.cz>
15 * found missing INIT_QDISC(htb)
16 * Vladimir Smelhaus, Aamer Akhter, Bert Hubert
17 * helped a lot to locate nasty class stall bug
18 * Andi Kleen, Jamal Hadi, Bert Hubert
19 * code review and helpful comments on shaping
20 * Tomasz Wrona, <tw@eter.tym.pl>
21 * created test case so that I was able to fix nasty bug
23 * spotted bug in dequeue code and helped with fix
25 * fixed requeue routine
26 * and many others. thanks.
28 * $Id: sch_htb.c,v 1.25 2003/12/07 11:08:25 devik Exp devik $
30 #include <linux/module.h>
31 #include <linux/types.h>
32 #include <linux/kernel.h>
33 #include <linux/string.h>
34 #include <linux/errno.h>
35 #include <linux/skbuff.h>
36 #include <linux/list.h>
37 #include <linux/compiler.h>
38 #include <linux/rbtree.h>
39 #include <net/netlink.h>
40 #include <net/pkt_sched.h>
44 ========================================================================
45 HTB is like TBF with multiple classes. It is also similar to CBQ because
46 it allows to assign priority to each class in hierarchy.
47 In fact it is another implementation of Floyd's formal sharing.
50 Each class is assigned level. Leaf has ALWAYS level 0 and root
51 classes have level TC_HTB_MAXDEPTH-1. Interior nodes has level
52 one less than their parent.
55 #define HTB_HSIZE 16 /* classid hash size */
56 #define HTB_HYSTERESIS 1 /* whether to use mode hysteresis for speedup */
57 #define HTB_VER 0x30011 /* major must be matched with number suplied by TC as version */
59 #if HTB_VER >> 16 != TC_HTB_PROTOVER
60 #error "Mismatched sch_htb.c and pkt_sch.h"
63 /* used internaly to keep status of single class */
65 HTB_CANT_SEND, /* class can't send and can't borrow */
66 HTB_MAY_BORROW, /* class can't send but may borrow */
67 HTB_CAN_SEND /* class can send */
70 /* interior & leaf nodes; props specific to leaves are marked L: */
72 /* general class parameters */
74 struct gnet_stats_basic bstats;
75 struct gnet_stats_queue qstats;
76 struct gnet_stats_rate_est rate_est;
77 struct tc_htb_xstats xstats; /* our special stats */
78 int refcnt; /* usage count of this class */
81 int level; /* our level (see above) */
82 struct htb_class *parent; /* parent class */
83 struct hlist_node hlist; /* classid hash list item */
84 struct list_head sibling; /* sibling list item */
85 struct list_head children; /* children list */
88 struct htb_class_leaf {
93 int deficit[TC_HTB_MAXDEPTH];
94 struct list_head drop_list;
96 struct htb_class_inner {
97 struct rb_root feed[TC_HTB_NUMPRIO]; /* feed trees */
98 struct rb_node *ptr[TC_HTB_NUMPRIO]; /* current class ptr */
99 /* When class changes from state 1->2 and disconnects from
100 parent's feed then we lost ptr value and start from the
101 first child again. Here we store classid of the
102 last valid ptr (used when ptr is NULL). */
103 u32 last_ptr_id[TC_HTB_NUMPRIO];
106 struct rb_node node[TC_HTB_NUMPRIO]; /* node for self or feed tree */
107 struct rb_node pq_node; /* node for event queue */
108 psched_time_t pq_key;
110 int prio_activity; /* for which prios are we active */
111 enum htb_cmode cmode; /* current mode of the class */
113 /* class attached filters */
114 struct tcf_proto *filter_list;
117 int warned; /* only one warning about non work conserving .. */
119 /* token bucket parameters */
120 struct qdisc_rate_table *rate; /* rate table of the class itself */
121 struct qdisc_rate_table *ceil; /* ceiling rate (limits borrows too) */
122 long buffer, cbuffer; /* token bucket depth/rate */
123 psched_tdiff_t mbuffer; /* max wait time */
124 long tokens, ctokens; /* current number of tokens */
125 psched_time_t t_c; /* checkpoint time */
127 int prio; /* For parent to leaf return possible here */
128 int quantum; /* we do backup. Finally full replacement */
129 /* of un.leaf originals should be done. */
132 static inline long L2T(struct htb_class *cl, struct qdisc_rate_table *rate,
135 long result = qdisc_l2t(rate, size);
140 struct list_head root; /* root classes list */
141 struct hlist_head hash[HTB_HSIZE]; /* hashed by classid */
142 struct list_head drops[TC_HTB_NUMPRIO];/* active leaves (for drops) */
144 /* self list - roots of self generating tree */
145 struct rb_root row[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
146 int row_mask[TC_HTB_MAXDEPTH];
147 struct rb_node *ptr[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
148 u32 last_ptr_id[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
150 /* self wait list - roots of wait PQs per row */
151 struct rb_root wait_pq[TC_HTB_MAXDEPTH];
153 /* time of nearest event per level (row) */
154 psched_time_t near_ev_cache[TC_HTB_MAXDEPTH];
156 /* whether we hit non-work conserving class during this dequeue; we use */
157 int nwc_hit; /* this to disable mindelay complaint in dequeue */
159 int defcls; /* class where unclassified flows go to */
161 /* filters for qdisc itself */
162 struct tcf_proto *filter_list;
165 int rate2quantum; /* quant = rate / rate2quantum */
166 psched_time_t now; /* cached dequeue time */
167 struct qdisc_watchdog watchdog;
169 /* non shaped skbs; let them go directly thru */
170 struct sk_buff_head direct_queue;
171 int direct_qlen; /* max qlen of above */
176 /* compute hash of size HTB_HSIZE for given handle */
177 static inline int htb_hash(u32 h)
180 #error "Declare new hash for your HTB_HSIZE"
182 h ^= h >> 8; /* stolen from cbq_hash */
187 /* find class in global hash table using given handle */
188 static inline struct htb_class *htb_find(u32 handle, struct Qdisc *sch)
190 struct htb_sched *q = qdisc_priv(sch);
191 struct hlist_node *p;
192 struct htb_class *cl;
194 if (TC_H_MAJ(handle) != sch->handle)
197 hlist_for_each_entry(cl, p, q->hash + htb_hash(handle), hlist) {
198 if (cl->classid == handle)
205 * htb_classify - classify a packet into class
207 * It returns NULL if the packet should be dropped or -1 if the packet
208 * should be passed directly thru. In all other cases leaf class is returned.
209 * We allow direct class selection by classid in priority. The we examine
210 * filters in qdisc and in inner nodes (if higher filter points to the inner
211 * node). If we end up with classid MAJOR:0 we enqueue the skb into special
212 * internal fifo (direct). These packets then go directly thru. If we still
213 * have no valid leaf we try to use MAJOR:default leaf. It still unsuccessfull
214 * then finish and return direct queue.
216 #define HTB_DIRECT (struct htb_class*)-1
218 static struct htb_class *htb_classify(struct sk_buff *skb, struct Qdisc *sch,
221 struct htb_sched *q = qdisc_priv(sch);
222 struct htb_class *cl;
223 struct tcf_result res;
224 struct tcf_proto *tcf;
227 /* allow to select class by setting skb->priority to valid classid;
228 note that nfmark can be used too by attaching filter fw with no
230 if (skb->priority == sch->handle)
231 return HTB_DIRECT; /* X:0 (direct flow) selected */
232 if ((cl = htb_find(skb->priority, sch)) != NULL && cl->level == 0)
235 *qerr = NET_XMIT_BYPASS;
236 tcf = q->filter_list;
237 while (tcf && (result = tc_classify(skb, tcf, &res)) >= 0) {
238 #ifdef CONFIG_NET_CLS_ACT
242 *qerr = NET_XMIT_SUCCESS;
247 if ((cl = (void *)res.class) == NULL) {
248 if (res.classid == sch->handle)
249 return HTB_DIRECT; /* X:0 (direct flow) */
250 if ((cl = htb_find(res.classid, sch)) == NULL)
251 break; /* filter selected invalid classid */
254 return cl; /* we hit leaf; return it */
256 /* we have got inner class; apply inner filter chain */
257 tcf = cl->filter_list;
259 /* classification failed; try to use default class */
260 cl = htb_find(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
261 if (!cl || cl->level)
262 return HTB_DIRECT; /* bad default .. this is safe bet */
267 * htb_add_to_id_tree - adds class to the round robin list
269 * Routine adds class to the list (actually tree) sorted by classid.
270 * Make sure that class is not already on such list for given prio.
272 static void htb_add_to_id_tree(struct rb_root *root,
273 struct htb_class *cl, int prio)
275 struct rb_node **p = &root->rb_node, *parent = NULL;
280 c = rb_entry(parent, struct htb_class, node[prio]);
282 if (cl->classid > c->classid)
283 p = &parent->rb_right;
285 p = &parent->rb_left;
287 rb_link_node(&cl->node[prio], parent, p);
288 rb_insert_color(&cl->node[prio], root);
292 * htb_add_to_wait_tree - adds class to the event queue with delay
294 * The class is added to priority event queue to indicate that class will
295 * change its mode in cl->pq_key microseconds. Make sure that class is not
296 * already in the queue.
298 static void htb_add_to_wait_tree(struct htb_sched *q,
299 struct htb_class *cl, long delay)
301 struct rb_node **p = &q->wait_pq[cl->level].rb_node, *parent = NULL;
303 cl->pq_key = q->now + delay;
304 if (cl->pq_key == q->now)
307 /* update the nearest event cache */
308 if (q->near_ev_cache[cl->level] > cl->pq_key)
309 q->near_ev_cache[cl->level] = cl->pq_key;
314 c = rb_entry(parent, struct htb_class, pq_node);
315 if (cl->pq_key >= c->pq_key)
316 p = &parent->rb_right;
318 p = &parent->rb_left;
320 rb_link_node(&cl->pq_node, parent, p);
321 rb_insert_color(&cl->pq_node, &q->wait_pq[cl->level]);
325 * htb_next_rb_node - finds next node in binary tree
327 * When we are past last key we return NULL.
328 * Average complexity is 2 steps per call.
330 static inline void htb_next_rb_node(struct rb_node **n)
336 * htb_add_class_to_row - add class to its row
338 * The class is added to row at priorities marked in mask.
339 * It does nothing if mask == 0.
341 static inline void htb_add_class_to_row(struct htb_sched *q,
342 struct htb_class *cl, int mask)
344 q->row_mask[cl->level] |= mask;
346 int prio = ffz(~mask);
347 mask &= ~(1 << prio);
348 htb_add_to_id_tree(q->row[cl->level] + prio, cl, prio);
352 /* If this triggers, it is a bug in this code, but it need not be fatal */
353 static void htb_safe_rb_erase(struct rb_node *rb, struct rb_root *root)
355 if (RB_EMPTY_NODE(rb)) {
365 * htb_remove_class_from_row - removes class from its row
367 * The class is removed from row at priorities marked in mask.
368 * It does nothing if mask == 0.
370 static inline void htb_remove_class_from_row(struct htb_sched *q,
371 struct htb_class *cl, int mask)
376 int prio = ffz(~mask);
378 mask &= ~(1 << prio);
379 if (q->ptr[cl->level][prio] == cl->node + prio)
380 htb_next_rb_node(q->ptr[cl->level] + prio);
382 htb_safe_rb_erase(cl->node + prio, q->row[cl->level] + prio);
383 if (!q->row[cl->level][prio].rb_node)
386 q->row_mask[cl->level] &= ~m;
390 * htb_activate_prios - creates active classe's feed chain
392 * The class is connected to ancestors and/or appropriate rows
393 * for priorities it is participating on. cl->cmode must be new
394 * (activated) mode. It does nothing if cl->prio_activity == 0.
396 static void htb_activate_prios(struct htb_sched *q, struct htb_class *cl)
398 struct htb_class *p = cl->parent;
399 long m, mask = cl->prio_activity;
401 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
407 if (p->un.inner.feed[prio].rb_node)
408 /* parent already has its feed in use so that
409 reset bit in mask as parent is already ok */
410 mask &= ~(1 << prio);
412 htb_add_to_id_tree(p->un.inner.feed + prio, cl, prio);
414 p->prio_activity |= mask;
419 if (cl->cmode == HTB_CAN_SEND && mask)
420 htb_add_class_to_row(q, cl, mask);
424 * htb_deactivate_prios - remove class from feed chain
426 * cl->cmode must represent old mode (before deactivation). It does
427 * nothing if cl->prio_activity == 0. Class is removed from all feed
430 static void htb_deactivate_prios(struct htb_sched *q, struct htb_class *cl)
432 struct htb_class *p = cl->parent;
433 long m, mask = cl->prio_activity;
435 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
442 if (p->un.inner.ptr[prio] == cl->node + prio) {
443 /* we are removing child which is pointed to from
444 parent feed - forget the pointer but remember
446 p->un.inner.last_ptr_id[prio] = cl->classid;
447 p->un.inner.ptr[prio] = NULL;
450 htb_safe_rb_erase(cl->node + prio, p->un.inner.feed + prio);
452 if (!p->un.inner.feed[prio].rb_node)
456 p->prio_activity &= ~mask;
461 if (cl->cmode == HTB_CAN_SEND && mask)
462 htb_remove_class_from_row(q, cl, mask);
466 static inline long htb_lowater(const struct htb_class *cl)
468 return cl->cmode != HTB_CANT_SEND ? -cl->cbuffer : 0;
470 static inline long htb_hiwater(const struct htb_class *cl)
472 return cl->cmode == HTB_CAN_SEND ? -cl->buffer : 0;
475 #define htb_lowater(cl) (0)
476 #define htb_hiwater(cl) (0)
480 * htb_class_mode - computes and returns current class mode
482 * It computes cl's mode at time cl->t_c+diff and returns it. If mode
483 * is not HTB_CAN_SEND then cl->pq_key is updated to time difference
484 * from now to time when cl will change its state.
485 * Also it is worth to note that class mode doesn't change simply
486 * at cl->{c,}tokens == 0 but there can rather be hysteresis of
487 * 0 .. -cl->{c,}buffer range. It is meant to limit number of
488 * mode transitions per time unit. The speed gain is about 1/6.
490 static inline enum htb_cmode
491 htb_class_mode(struct htb_class *cl, long *diff)
495 if ((toks = (cl->ctokens + *diff)) < htb_lowater(cl)) {
497 return HTB_CANT_SEND;
500 if ((toks = (cl->tokens + *diff)) >= htb_hiwater(cl))
504 return HTB_MAY_BORROW;
508 * htb_change_class_mode - changes classe's mode
510 * This should be the only way how to change classe's mode under normal
511 * cirsumstances. Routine will update feed lists linkage, change mode
512 * and add class to the wait event queue if appropriate. New mode should
513 * be different from old one and cl->pq_key has to be valid if changing
514 * to mode other than HTB_CAN_SEND (see htb_add_to_wait_tree).
517 htb_change_class_mode(struct htb_sched *q, struct htb_class *cl, long *diff)
519 enum htb_cmode new_mode = htb_class_mode(cl, diff);
521 if (new_mode == cl->cmode)
524 if (cl->prio_activity) { /* not necessary: speed optimization */
525 if (cl->cmode != HTB_CANT_SEND)
526 htb_deactivate_prios(q, cl);
527 cl->cmode = new_mode;
528 if (new_mode != HTB_CANT_SEND)
529 htb_activate_prios(q, cl);
531 cl->cmode = new_mode;
535 * htb_activate - inserts leaf cl into appropriate active feeds
537 * Routine learns (new) priority of leaf and activates feed chain
538 * for the prio. It can be called on already active leaf safely.
539 * It also adds leaf into droplist.
541 static inline void htb_activate(struct htb_sched *q, struct htb_class *cl)
543 BUG_TRAP(!cl->level && cl->un.leaf.q && cl->un.leaf.q->q.qlen);
545 if (!cl->prio_activity) {
546 cl->prio_activity = 1 << (cl->un.leaf.aprio = cl->un.leaf.prio);
547 htb_activate_prios(q, cl);
548 list_add_tail(&cl->un.leaf.drop_list,
549 q->drops + cl->un.leaf.aprio);
554 * htb_deactivate - remove leaf cl from active feeds
556 * Make sure that leaf is active. In the other words it can't be called
557 * with non-active leaf. It also removes class from the drop list.
559 static inline void htb_deactivate(struct htb_sched *q, struct htb_class *cl)
561 BUG_TRAP(cl->prio_activity);
563 htb_deactivate_prios(q, cl);
564 cl->prio_activity = 0;
565 list_del_init(&cl->un.leaf.drop_list);
568 static int htb_enqueue(struct sk_buff *skb, struct Qdisc *sch)
571 struct htb_sched *q = qdisc_priv(sch);
572 struct htb_class *cl = htb_classify(skb, sch, &ret);
574 if (cl == HTB_DIRECT) {
575 /* enqueue to helper queue */
576 if (q->direct_queue.qlen < q->direct_qlen) {
577 __skb_queue_tail(&q->direct_queue, skb);
582 return NET_XMIT_DROP;
584 #ifdef CONFIG_NET_CLS_ACT
586 if (ret == NET_XMIT_BYPASS)
591 } else if (cl->un.leaf.q->enqueue(skb, cl->un.leaf.q) !=
595 return NET_XMIT_DROP;
597 cl->bstats.packets +=
598 skb_is_gso(skb)?skb_shinfo(skb)->gso_segs:1;
599 cl->bstats.bytes += skb->len;
604 sch->bstats.packets += skb_is_gso(skb)?skb_shinfo(skb)->gso_segs:1;
605 sch->bstats.bytes += skb->len;
606 return NET_XMIT_SUCCESS;
609 /* TODO: requeuing packet charges it to policers again !! */
610 static int htb_requeue(struct sk_buff *skb, struct Qdisc *sch)
612 struct htb_sched *q = qdisc_priv(sch);
613 int ret = NET_XMIT_SUCCESS;
614 struct htb_class *cl = htb_classify(skb, sch, &ret);
615 struct sk_buff *tskb;
617 if (cl == HTB_DIRECT || !cl) {
618 /* enqueue to helper queue */
619 if (q->direct_queue.qlen < q->direct_qlen && cl) {
620 __skb_queue_head(&q->direct_queue, skb);
622 __skb_queue_head(&q->direct_queue, skb);
623 tskb = __skb_dequeue_tail(&q->direct_queue);
628 } else if (cl->un.leaf.q->ops->requeue(skb, cl->un.leaf.q) !=
632 return NET_XMIT_DROP;
637 sch->qstats.requeues++;
638 return NET_XMIT_SUCCESS;
642 * htb_charge_class - charges amount "bytes" to leaf and ancestors
644 * Routine assumes that packet "bytes" long was dequeued from leaf cl
645 * borrowing from "level". It accounts bytes to ceil leaky bucket for
646 * leaf and all ancestors and to rate bucket for ancestors at levels
647 * "level" and higher. It also handles possible change of mode resulting
648 * from the update. Note that mode can also increase here (MAY_BORROW to
649 * CAN_SEND) because we can use more precise clock that event queue here.
650 * In such case we remove class from event queue first.
652 static void htb_charge_class(struct htb_sched *q, struct htb_class *cl,
653 int level, struct sk_buff *skb)
655 int bytes = skb->len;
657 enum htb_cmode old_mode;
659 #define HTB_ACCNT(T,B,R) toks = diff + cl->T; \
660 if (toks > cl->B) toks = cl->B; \
661 toks -= L2T(cl, cl->R, bytes); \
662 if (toks <= -cl->mbuffer) toks = 1-cl->mbuffer; \
666 diff = psched_tdiff_bounded(q->now, cl->t_c, cl->mbuffer);
667 if (cl->level >= level) {
668 if (cl->level == level)
670 HTB_ACCNT(tokens, buffer, rate);
672 cl->xstats.borrows++;
673 cl->tokens += diff; /* we moved t_c; update tokens */
675 HTB_ACCNT(ctokens, cbuffer, ceil);
678 old_mode = cl->cmode;
680 htb_change_class_mode(q, cl, &diff);
681 if (old_mode != cl->cmode) {
682 if (old_mode != HTB_CAN_SEND)
683 htb_safe_rb_erase(&cl->pq_node, q->wait_pq + cl->level);
684 if (cl->cmode != HTB_CAN_SEND)
685 htb_add_to_wait_tree(q, cl, diff);
688 /* update byte stats except for leaves which are already updated */
690 cl->bstats.bytes += bytes;
691 cl->bstats.packets += skb_is_gso(skb)?
692 skb_shinfo(skb)->gso_segs:1;
699 * htb_do_events - make mode changes to classes at the level
701 * Scans event queue for pending events and applies them. Returns time of
702 * next pending event (0 for no event in pq).
703 * Note: Applied are events whose have cl->pq_key <= q->now.
705 static psched_time_t htb_do_events(struct htb_sched *q, int level)
709 for (i = 0; i < 500; i++) {
710 struct htb_class *cl;
712 struct rb_node *p = rb_first(&q->wait_pq[level]);
717 cl = rb_entry(p, struct htb_class, pq_node);
718 if (cl->pq_key > q->now)
721 htb_safe_rb_erase(p, q->wait_pq + level);
722 diff = psched_tdiff_bounded(q->now, cl->t_c, cl->mbuffer);
723 htb_change_class_mode(q, cl, &diff);
724 if (cl->cmode != HTB_CAN_SEND)
725 htb_add_to_wait_tree(q, cl, diff);
728 printk(KERN_WARNING "htb: too many events !\n");
729 return q->now + PSCHED_TICKS_PER_SEC / 10;
732 /* Returns class->node+prio from id-tree where classe's id is >= id. NULL
733 is no such one exists. */
734 static struct rb_node *htb_id_find_next_upper(int prio, struct rb_node *n,
737 struct rb_node *r = NULL;
739 struct htb_class *cl =
740 rb_entry(n, struct htb_class, node[prio]);
741 if (id == cl->classid)
744 if (id > cl->classid) {
755 * htb_lookup_leaf - returns next leaf class in DRR order
757 * Find leaf where current feed pointers points to.
759 static struct htb_class *htb_lookup_leaf(struct rb_root *tree, int prio,
760 struct rb_node **pptr, u32 * pid)
764 struct rb_node *root;
765 struct rb_node **pptr;
767 } stk[TC_HTB_MAXDEPTH], *sp = stk;
769 BUG_TRAP(tree->rb_node);
770 sp->root = tree->rb_node;
774 for (i = 0; i < 65535; i++) {
775 if (!*sp->pptr && *sp->pid) {
776 /* ptr was invalidated but id is valid - try to recover
777 the original or next ptr */
779 htb_id_find_next_upper(prio, sp->root, *sp->pid);
781 *sp->pid = 0; /* ptr is valid now so that remove this hint as it
782 can become out of date quickly */
783 if (!*sp->pptr) { /* we are at right end; rewind & go up */
784 *sp->pptr = sp->root;
785 while ((*sp->pptr)->rb_left)
786 *sp->pptr = (*sp->pptr)->rb_left;
792 htb_next_rb_node(sp->pptr);
795 struct htb_class *cl;
796 cl = rb_entry(*sp->pptr, struct htb_class, node[prio]);
799 (++sp)->root = cl->un.inner.feed[prio].rb_node;
800 sp->pptr = cl->un.inner.ptr + prio;
801 sp->pid = cl->un.inner.last_ptr_id + prio;
808 /* dequeues packet at given priority and level; call only if
809 you are sure that there is active class at prio/level */
810 static struct sk_buff *htb_dequeue_tree(struct htb_sched *q, int prio,
813 struct sk_buff *skb = NULL;
814 struct htb_class *cl, *start;
815 /* look initial class up in the row */
816 start = cl = htb_lookup_leaf(q->row[level] + prio, prio,
817 q->ptr[level] + prio,
818 q->last_ptr_id[level] + prio);
826 /* class can be empty - it is unlikely but can be true if leaf
827 qdisc drops packets in enqueue routine or if someone used
828 graft operation on the leaf since last dequeue;
829 simply deactivate and skip such class */
830 if (unlikely(cl->un.leaf.q->q.qlen == 0)) {
831 struct htb_class *next;
832 htb_deactivate(q, cl);
834 /* row/level might become empty */
835 if ((q->row_mask[level] & (1 << prio)) == 0)
838 next = htb_lookup_leaf(q->row[level] + prio,
839 prio, q->ptr[level] + prio,
840 q->last_ptr_id[level] + prio);
842 if (cl == start) /* fix start if we just deleted it */
848 skb = cl->un.leaf.q->dequeue(cl->un.leaf.q);
849 if (likely(skb != NULL))
853 "htb: class %X isn't work conserving ?!\n",
858 htb_next_rb_node((level ? cl->parent->un.inner.ptr : q->
860 cl = htb_lookup_leaf(q->row[level] + prio, prio,
861 q->ptr[level] + prio,
862 q->last_ptr_id[level] + prio);
864 } while (cl != start);
866 if (likely(skb != NULL)) {
867 if ((cl->un.leaf.deficit[level] -= skb->len) < 0) {
868 cl->un.leaf.deficit[level] += cl->un.leaf.quantum;
869 htb_next_rb_node((level ? cl->parent->un.inner.ptr : q->
872 /* this used to be after charge_class but this constelation
873 gives us slightly better performance */
874 if (!cl->un.leaf.q->q.qlen)
875 htb_deactivate(q, cl);
876 htb_charge_class(q, cl, level, skb);
881 static struct sk_buff *htb_dequeue(struct Qdisc *sch)
883 struct sk_buff *skb = NULL;
884 struct htb_sched *q = qdisc_priv(sch);
886 psched_time_t next_event;
888 /* try to dequeue direct packets as high prio (!) to minimize cpu work */
889 skb = __skb_dequeue(&q->direct_queue);
891 sch->flags &= ~TCQ_F_THROTTLED;
898 q->now = psched_get_time();
900 next_event = q->now + 5 * PSCHED_TICKS_PER_SEC;
902 for (level = 0; level < TC_HTB_MAXDEPTH; level++) {
903 /* common case optimization - skip event handler quickly */
907 if (q->now >= q->near_ev_cache[level]) {
908 event = htb_do_events(q, level);
910 event = q->now + PSCHED_TICKS_PER_SEC;
911 q->near_ev_cache[level] = event;
913 event = q->near_ev_cache[level];
915 if (event && next_event > event)
918 m = ~q->row_mask[level];
919 while (m != (int)(-1)) {
922 skb = htb_dequeue_tree(q, prio, level);
923 if (likely(skb != NULL)) {
925 sch->flags &= ~TCQ_F_THROTTLED;
930 sch->qstats.overlimits++;
931 qdisc_watchdog_schedule(&q->watchdog, next_event);
936 /* try to drop from each class (by prio) until one succeed */
937 static unsigned int htb_drop(struct Qdisc *sch)
939 struct htb_sched *q = qdisc_priv(sch);
942 for (prio = TC_HTB_NUMPRIO - 1; prio >= 0; prio--) {
944 list_for_each(p, q->drops + prio) {
945 struct htb_class *cl = list_entry(p, struct htb_class,
948 if (cl->un.leaf.q->ops->drop &&
949 (len = cl->un.leaf.q->ops->drop(cl->un.leaf.q))) {
951 if (!cl->un.leaf.q->q.qlen)
952 htb_deactivate(q, cl);
960 /* reset all classes */
961 /* always caled under BH & queue lock */
962 static void htb_reset(struct Qdisc *sch)
964 struct htb_sched *q = qdisc_priv(sch);
967 for (i = 0; i < HTB_HSIZE; i++) {
968 struct hlist_node *p;
969 struct htb_class *cl;
971 hlist_for_each_entry(cl, p, q->hash + i, hlist) {
973 memset(&cl->un.inner, 0, sizeof(cl->un.inner));
976 qdisc_reset(cl->un.leaf.q);
977 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
979 cl->prio_activity = 0;
980 cl->cmode = HTB_CAN_SEND;
984 qdisc_watchdog_cancel(&q->watchdog);
985 __skb_queue_purge(&q->direct_queue);
987 memset(q->row, 0, sizeof(q->row));
988 memset(q->row_mask, 0, sizeof(q->row_mask));
989 memset(q->wait_pq, 0, sizeof(q->wait_pq));
990 memset(q->ptr, 0, sizeof(q->ptr));
991 for (i = 0; i < TC_HTB_NUMPRIO; i++)
992 INIT_LIST_HEAD(q->drops + i);
995 static int htb_init(struct Qdisc *sch, struct nlattr *opt)
997 struct htb_sched *q = qdisc_priv(sch);
998 struct nlattr *tb[TCA_HTB_INIT + 1];
999 struct tc_htb_glob *gopt;
1006 err = nla_parse_nested(tb, TCA_HTB_INIT, opt, NULL);
1010 if (tb[TCA_HTB_INIT] == NULL ||
1011 nla_len(tb[TCA_HTB_INIT]) < sizeof(*gopt)) {
1012 printk(KERN_ERR "HTB: hey probably you have bad tc tool ?\n");
1015 gopt = nla_data(tb[TCA_HTB_INIT]);
1016 if (gopt->version != HTB_VER >> 16) {
1018 "HTB: need tc/htb version %d (minor is %d), you have %d\n",
1019 HTB_VER >> 16, HTB_VER & 0xffff, gopt->version);
1023 INIT_LIST_HEAD(&q->root);
1024 for (i = 0; i < HTB_HSIZE; i++)
1025 INIT_HLIST_HEAD(q->hash + i);
1026 for (i = 0; i < TC_HTB_NUMPRIO; i++)
1027 INIT_LIST_HEAD(q->drops + i);
1029 qdisc_watchdog_init(&q->watchdog, sch);
1030 skb_queue_head_init(&q->direct_queue);
1032 q->direct_qlen = sch->dev->tx_queue_len;
1033 if (q->direct_qlen < 2) /* some devices have zero tx_queue_len */
1036 if ((q->rate2quantum = gopt->rate2quantum) < 1)
1037 q->rate2quantum = 1;
1038 q->defcls = gopt->defcls;
1043 static int htb_dump(struct Qdisc *sch, struct sk_buff *skb)
1045 struct htb_sched *q = qdisc_priv(sch);
1046 struct nlattr *nest;
1047 struct tc_htb_glob gopt;
1049 spin_lock_bh(&sch->dev->queue_lock);
1051 gopt.direct_pkts = q->direct_pkts;
1052 gopt.version = HTB_VER;
1053 gopt.rate2quantum = q->rate2quantum;
1054 gopt.defcls = q->defcls;
1057 nest = nla_nest_start(skb, TCA_OPTIONS);
1059 goto nla_put_failure;
1060 NLA_PUT(skb, TCA_HTB_INIT, sizeof(gopt), &gopt);
1061 nla_nest_end(skb, nest);
1063 spin_unlock_bh(&sch->dev->queue_lock);
1067 spin_unlock_bh(&sch->dev->queue_lock);
1068 nla_nest_cancel(skb, nest);
1072 static int htb_dump_class(struct Qdisc *sch, unsigned long arg,
1073 struct sk_buff *skb, struct tcmsg *tcm)
1075 struct htb_class *cl = (struct htb_class *)arg;
1076 struct nlattr *nest;
1077 struct tc_htb_opt opt;
1079 spin_lock_bh(&sch->dev->queue_lock);
1080 tcm->tcm_parent = cl->parent ? cl->parent->classid : TC_H_ROOT;
1081 tcm->tcm_handle = cl->classid;
1082 if (!cl->level && cl->un.leaf.q)
1083 tcm->tcm_info = cl->un.leaf.q->handle;
1085 nest = nla_nest_start(skb, TCA_OPTIONS);
1087 goto nla_put_failure;
1089 memset(&opt, 0, sizeof(opt));
1091 opt.rate = cl->rate->rate;
1092 opt.buffer = cl->buffer;
1093 opt.ceil = cl->ceil->rate;
1094 opt.cbuffer = cl->cbuffer;
1095 opt.quantum = cl->un.leaf.quantum;
1096 opt.prio = cl->un.leaf.prio;
1097 opt.level = cl->level;
1098 NLA_PUT(skb, TCA_HTB_PARMS, sizeof(opt), &opt);
1100 nla_nest_end(skb, nest);
1101 spin_unlock_bh(&sch->dev->queue_lock);
1105 spin_unlock_bh(&sch->dev->queue_lock);
1106 nla_nest_cancel(skb, nest);
1111 htb_dump_class_stats(struct Qdisc *sch, unsigned long arg, struct gnet_dump *d)
1113 struct htb_class *cl = (struct htb_class *)arg;
1115 if (!cl->level && cl->un.leaf.q)
1116 cl->qstats.qlen = cl->un.leaf.q->q.qlen;
1117 cl->xstats.tokens = cl->tokens;
1118 cl->xstats.ctokens = cl->ctokens;
1120 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
1121 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1122 gnet_stats_copy_queue(d, &cl->qstats) < 0)
1125 return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1128 static int htb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1131 struct htb_class *cl = (struct htb_class *)arg;
1133 if (cl && !cl->level) {
1135 (new = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops,
1140 if ((*old = xchg(&cl->un.leaf.q, new)) != NULL) {
1141 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
1144 sch_tree_unlock(sch);
1150 static struct Qdisc *htb_leaf(struct Qdisc *sch, unsigned long arg)
1152 struct htb_class *cl = (struct htb_class *)arg;
1153 return (cl && !cl->level) ? cl->un.leaf.q : NULL;
1156 static void htb_qlen_notify(struct Qdisc *sch, unsigned long arg)
1158 struct htb_class *cl = (struct htb_class *)arg;
1160 if (cl->un.leaf.q->q.qlen == 0)
1161 htb_deactivate(qdisc_priv(sch), cl);
1164 static unsigned long htb_get(struct Qdisc *sch, u32 classid)
1166 struct htb_class *cl = htb_find(classid, sch);
1169 return (unsigned long)cl;
1172 static inline int htb_parent_last_child(struct htb_class *cl)
1175 /* the root class */
1178 if (!(cl->parent->children.next == &cl->sibling &&
1179 cl->parent->children.prev == &cl->sibling))
1180 /* not the last child */
1186 static void htb_parent_to_leaf(struct htb_class *cl, struct Qdisc *new_q)
1188 struct htb_class *parent = cl->parent;
1190 BUG_TRAP(!cl->level && cl->un.leaf.q && !cl->prio_activity);
1193 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1194 INIT_LIST_HEAD(&parent->un.leaf.drop_list);
1195 parent->un.leaf.q = new_q ? new_q : &noop_qdisc;
1196 parent->un.leaf.quantum = parent->quantum;
1197 parent->un.leaf.prio = parent->prio;
1198 parent->tokens = parent->buffer;
1199 parent->ctokens = parent->cbuffer;
1200 parent->t_c = psched_get_time();
1201 parent->cmode = HTB_CAN_SEND;
1204 static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
1206 struct htb_sched *q = qdisc_priv(sch);
1209 BUG_TRAP(cl->un.leaf.q);
1210 qdisc_destroy(cl->un.leaf.q);
1212 gen_kill_estimator(&cl->bstats, &cl->rate_est);
1213 qdisc_put_rtab(cl->rate);
1214 qdisc_put_rtab(cl->ceil);
1216 tcf_destroy_chain(cl->filter_list);
1218 while (!list_empty(&cl->children))
1219 htb_destroy_class(sch, list_entry(cl->children.next,
1220 struct htb_class, sibling));
1222 /* note: this delete may happen twice (see htb_delete) */
1223 hlist_del_init(&cl->hlist);
1224 list_del(&cl->sibling);
1226 if (cl->prio_activity)
1227 htb_deactivate(q, cl);
1229 if (cl->cmode != HTB_CAN_SEND)
1230 htb_safe_rb_erase(&cl->pq_node, q->wait_pq + cl->level);
1235 /* always caled under BH & queue lock */
1236 static void htb_destroy(struct Qdisc *sch)
1238 struct htb_sched *q = qdisc_priv(sch);
1240 qdisc_watchdog_cancel(&q->watchdog);
1241 /* This line used to be after htb_destroy_class call below
1242 and surprisingly it worked in 2.4. But it must precede it
1243 because filter need its target class alive to be able to call
1244 unbind_filter on it (without Oops). */
1245 tcf_destroy_chain(q->filter_list);
1247 while (!list_empty(&q->root))
1248 htb_destroy_class(sch, list_entry(q->root.next,
1249 struct htb_class, sibling));
1251 __skb_queue_purge(&q->direct_queue);
1254 static int htb_delete(struct Qdisc *sch, unsigned long arg)
1256 struct htb_sched *q = qdisc_priv(sch);
1257 struct htb_class *cl = (struct htb_class *)arg;
1259 struct Qdisc *new_q = NULL;
1262 // TODO: why don't allow to delete subtree ? references ? does
1263 // tc subsys quarantee us that in htb_destroy it holds no class
1264 // refs so that we can remove children safely there ?
1265 if (!list_empty(&cl->children) || cl->filter_cnt)
1268 if (!cl->level && htb_parent_last_child(cl)) {
1269 new_q = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops,
1270 cl->parent->classid);
1277 qlen = cl->un.leaf.q->q.qlen;
1278 qdisc_reset(cl->un.leaf.q);
1279 qdisc_tree_decrease_qlen(cl->un.leaf.q, qlen);
1282 /* delete from hash and active; remainder in destroy_class */
1283 hlist_del_init(&cl->hlist);
1285 if (cl->prio_activity)
1286 htb_deactivate(q, cl);
1289 htb_parent_to_leaf(cl, new_q);
1291 if (--cl->refcnt == 0)
1292 htb_destroy_class(sch, cl);
1294 sch_tree_unlock(sch);
1298 static void htb_put(struct Qdisc *sch, unsigned long arg)
1300 struct htb_class *cl = (struct htb_class *)arg;
1302 if (--cl->refcnt == 0)
1303 htb_destroy_class(sch, cl);
1306 static int htb_change_class(struct Qdisc *sch, u32 classid,
1307 u32 parentid, struct nlattr **tca,
1311 struct htb_sched *q = qdisc_priv(sch);
1312 struct htb_class *cl = (struct htb_class *)*arg, *parent;
1313 struct nlattr *opt = tca[TCA_OPTIONS];
1314 struct qdisc_rate_table *rtab = NULL, *ctab = NULL;
1315 struct nlattr *tb[TCA_HTB_RTAB + 1];
1316 struct tc_htb_opt *hopt;
1318 /* extract all subattrs from opt attr */
1322 err = nla_parse_nested(tb, TCA_HTB_RTAB, opt, NULL);
1327 if (tb[TCA_HTB_PARMS] == NULL ||
1328 nla_len(tb[TCA_HTB_PARMS]) < sizeof(*hopt))
1331 parent = parentid == TC_H_ROOT ? NULL : htb_find(parentid, sch);
1333 hopt = nla_data(tb[TCA_HTB_PARMS]);
1335 rtab = qdisc_get_rtab(&hopt->rate, tb[TCA_HTB_RTAB]);
1336 ctab = qdisc_get_rtab(&hopt->ceil, tb[TCA_HTB_CTAB]);
1340 if (!cl) { /* new class */
1341 struct Qdisc *new_q;
1345 struct gnet_estimator opt;
1348 .nla_len = nla_attr_size(sizeof(est.opt)),
1349 .nla_type = TCA_RATE,
1352 /* 4s interval, 16s averaging constant */
1358 /* check for valid classid */
1359 if (!classid || TC_H_MAJ(classid ^ sch->handle)
1360 || htb_find(classid, sch))
1363 /* check maximal depth */
1364 if (parent && parent->parent && parent->parent->level < 2) {
1365 printk(KERN_ERR "htb: tree is too deep\n");
1369 if ((cl = kzalloc(sizeof(*cl), GFP_KERNEL)) == NULL)
1372 gen_new_estimator(&cl->bstats, &cl->rate_est,
1373 &sch->dev->queue_lock,
1374 tca[TCA_RATE] ? : &est.nla);
1376 INIT_LIST_HEAD(&cl->sibling);
1377 INIT_HLIST_NODE(&cl->hlist);
1378 INIT_LIST_HEAD(&cl->children);
1379 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
1380 RB_CLEAR_NODE(&cl->pq_node);
1382 for (prio = 0; prio < TC_HTB_NUMPRIO; prio++)
1383 RB_CLEAR_NODE(&cl->node[prio]);
1385 /* create leaf qdisc early because it uses kmalloc(GFP_KERNEL)
1386 so that can't be used inside of sch_tree_lock
1387 -- thanks to Karlis Peisenieks */
1388 new_q = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops, classid);
1390 if (parent && !parent->level) {
1391 unsigned int qlen = parent->un.leaf.q->q.qlen;
1393 /* turn parent into inner node */
1394 qdisc_reset(parent->un.leaf.q);
1395 qdisc_tree_decrease_qlen(parent->un.leaf.q, qlen);
1396 qdisc_destroy(parent->un.leaf.q);
1397 if (parent->prio_activity)
1398 htb_deactivate(q, parent);
1400 /* remove from evt list because of level change */
1401 if (parent->cmode != HTB_CAN_SEND) {
1402 htb_safe_rb_erase(&parent->pq_node, q->wait_pq);
1403 parent->cmode = HTB_CAN_SEND;
1405 parent->level = (parent->parent ? parent->parent->level
1406 : TC_HTB_MAXDEPTH) - 1;
1407 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1409 /* leaf (we) needs elementary qdisc */
1410 cl->un.leaf.q = new_q ? new_q : &noop_qdisc;
1412 cl->classid = classid;
1413 cl->parent = parent;
1415 /* set class to be in HTB_CAN_SEND state */
1416 cl->tokens = hopt->buffer;
1417 cl->ctokens = hopt->cbuffer;
1418 cl->mbuffer = 60 * PSCHED_TICKS_PER_SEC; /* 1min */
1419 cl->t_c = psched_get_time();
1420 cl->cmode = HTB_CAN_SEND;
1422 /* attach to the hash list and parent's family */
1423 hlist_add_head(&cl->hlist, q->hash + htb_hash(classid));
1424 list_add_tail(&cl->sibling,
1425 parent ? &parent->children : &q->root);
1428 gen_replace_estimator(&cl->bstats, &cl->rate_est,
1429 &sch->dev->queue_lock,
1434 /* it used to be a nasty bug here, we have to check that node
1435 is really leaf before changing cl->un.leaf ! */
1437 cl->un.leaf.quantum = rtab->rate.rate / q->rate2quantum;
1438 if (!hopt->quantum && cl->un.leaf.quantum < 1000) {
1440 "HTB: quantum of class %X is small. Consider r2q change.\n",
1442 cl->un.leaf.quantum = 1000;
1444 if (!hopt->quantum && cl->un.leaf.quantum > 200000) {
1446 "HTB: quantum of class %X is big. Consider r2q change.\n",
1448 cl->un.leaf.quantum = 200000;
1451 cl->un.leaf.quantum = hopt->quantum;
1452 if ((cl->un.leaf.prio = hopt->prio) >= TC_HTB_NUMPRIO)
1453 cl->un.leaf.prio = TC_HTB_NUMPRIO - 1;
1455 /* backup for htb_parent_to_leaf */
1456 cl->quantum = cl->un.leaf.quantum;
1457 cl->prio = cl->un.leaf.prio;
1460 cl->buffer = hopt->buffer;
1461 cl->cbuffer = hopt->cbuffer;
1463 qdisc_put_rtab(cl->rate);
1466 qdisc_put_rtab(cl->ceil);
1468 sch_tree_unlock(sch);
1470 *arg = (unsigned long)cl;
1475 qdisc_put_rtab(rtab);
1477 qdisc_put_rtab(ctab);
1481 static struct tcf_proto **htb_find_tcf(struct Qdisc *sch, unsigned long arg)
1483 struct htb_sched *q = qdisc_priv(sch);
1484 struct htb_class *cl = (struct htb_class *)arg;
1485 struct tcf_proto **fl = cl ? &cl->filter_list : &q->filter_list;
1490 static unsigned long htb_bind_filter(struct Qdisc *sch, unsigned long parent,
1493 struct htb_sched *q = qdisc_priv(sch);
1494 struct htb_class *cl = htb_find(classid, sch);
1496 /*if (cl && !cl->level) return 0;
1497 The line above used to be there to prevent attaching filters to
1498 leaves. But at least tc_index filter uses this just to get class
1499 for other reasons so that we have to allow for it.
1501 19.6.2002 As Werner explained it is ok - bind filter is just
1502 another way to "lock" the class - unlike "get" this lock can
1503 be broken by class during destroy IIUC.
1509 return (unsigned long)cl;
1512 static void htb_unbind_filter(struct Qdisc *sch, unsigned long arg)
1514 struct htb_sched *q = qdisc_priv(sch);
1515 struct htb_class *cl = (struct htb_class *)arg;
1523 static void htb_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1525 struct htb_sched *q = qdisc_priv(sch);
1531 for (i = 0; i < HTB_HSIZE; i++) {
1532 struct hlist_node *p;
1533 struct htb_class *cl;
1535 hlist_for_each_entry(cl, p, q->hash + i, hlist) {
1536 if (arg->count < arg->skip) {
1540 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1549 static const struct Qdisc_class_ops htb_class_ops = {
1552 .qlen_notify = htb_qlen_notify,
1555 .change = htb_change_class,
1556 .delete = htb_delete,
1558 .tcf_chain = htb_find_tcf,
1559 .bind_tcf = htb_bind_filter,
1560 .unbind_tcf = htb_unbind_filter,
1561 .dump = htb_dump_class,
1562 .dump_stats = htb_dump_class_stats,
1565 static struct Qdisc_ops htb_qdisc_ops __read_mostly = {
1567 .cl_ops = &htb_class_ops,
1569 .priv_size = sizeof(struct htb_sched),
1570 .enqueue = htb_enqueue,
1571 .dequeue = htb_dequeue,
1572 .requeue = htb_requeue,
1576 .destroy = htb_destroy,
1577 .change = NULL /* htb_change */,
1579 .owner = THIS_MODULE,
1582 static int __init htb_module_init(void)
1584 return register_qdisc(&htb_qdisc_ops);
1586 static void __exit htb_module_exit(void)
1588 unregister_qdisc(&htb_qdisc_ops);
1591 module_init(htb_module_init)
1592 module_exit(htb_module_exit)
1593 MODULE_LICENSE("GPL");