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 #include <linux/module.h>
29 #include <linux/moduleparam.h>
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/string.h>
33 #include <linux/errno.h>
34 #include <linux/skbuff.h>
35 #include <linux/list.h>
36 #include <linux/compiler.h>
37 #include <linux/rbtree.h>
38 #include <net/netlink.h>
39 #include <net/pkt_sched.h>
43 ========================================================================
44 HTB is like TBF with multiple classes. It is also similar to CBQ because
45 it allows to assign priority to each class in hierarchy.
46 In fact it is another implementation of Floyd's formal sharing.
49 Each class is assigned level. Leaf has ALWAYS level 0 and root
50 classes have level TC_HTB_MAXDEPTH-1. Interior nodes has level
51 one less than their parent.
54 static int htb_hysteresis __read_mostly = 0; /* whether to use mode hysteresis for speedup */
55 #define HTB_VER 0x30011 /* major must be matched with number suplied by TC as version */
57 #if HTB_VER >> 16 != TC_HTB_PROTOVER
58 #error "Mismatched sch_htb.c and pkt_sch.h"
61 /* Module parameter and sysfs export */
62 module_param (htb_hysteresis, int, 0640);
63 MODULE_PARM_DESC(htb_hysteresis, "Hysteresis mode, less CPU load, less accurate");
65 /* used internaly to keep status of single class */
67 HTB_CANT_SEND, /* class can't send and can't borrow */
68 HTB_MAY_BORROW, /* class can't send but may borrow */
69 HTB_CAN_SEND /* class can send */
72 /* interior & leaf nodes; props specific to leaves are marked L: */
74 struct Qdisc_class_common common;
75 /* general class parameters */
76 struct gnet_stats_basic bstats;
77 struct gnet_stats_queue qstats;
78 struct gnet_stats_rate_est rate_est;
79 struct tc_htb_xstats xstats; /* our special stats */
80 int refcnt; /* usage count of this class */
83 int level; /* our level (see above) */
84 unsigned int children;
85 struct htb_class *parent; /* parent class */
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 Qdisc_class_hash clhash;
141 struct list_head drops[TC_HTB_NUMPRIO];/* active leaves (for drops) */
143 /* self list - roots of self generating tree */
144 struct rb_root row[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
145 int row_mask[TC_HTB_MAXDEPTH];
146 struct rb_node *ptr[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
147 u32 last_ptr_id[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
149 /* self wait list - roots of wait PQs per row */
150 struct rb_root wait_pq[TC_HTB_MAXDEPTH];
152 /* time of nearest event per level (row) */
153 psched_time_t near_ev_cache[TC_HTB_MAXDEPTH];
155 /* whether we hit non-work conserving class during this dequeue; we use */
156 int nwc_hit; /* this to disable mindelay complaint in dequeue */
158 int defcls; /* class where unclassified flows go to */
160 /* filters for qdisc itself */
161 struct tcf_proto *filter_list;
163 int rate2quantum; /* quant = rate / rate2quantum */
164 psched_time_t now; /* cached dequeue time */
165 struct qdisc_watchdog watchdog;
167 /* non shaped skbs; let them go directly thru */
168 struct sk_buff_head direct_queue;
169 int direct_qlen; /* max qlen of above */
174 /* find class in global hash table using given handle */
175 static inline struct htb_class *htb_find(u32 handle, struct Qdisc *sch)
177 struct htb_sched *q = qdisc_priv(sch);
178 struct Qdisc_class_common *clc;
180 clc = qdisc_class_find(&q->clhash, handle);
183 return container_of(clc, struct htb_class, common);
187 * htb_classify - classify a packet into class
189 * It returns NULL if the packet should be dropped or -1 if the packet
190 * should be passed directly thru. In all other cases leaf class is returned.
191 * We allow direct class selection by classid in priority. The we examine
192 * filters in qdisc and in inner nodes (if higher filter points to the inner
193 * node). If we end up with classid MAJOR:0 we enqueue the skb into special
194 * internal fifo (direct). These packets then go directly thru. If we still
195 * have no valid leaf we try to use MAJOR:default leaf. It still unsuccessfull
196 * then finish and return direct queue.
198 #define HTB_DIRECT (struct htb_class*)-1
200 static struct htb_class *htb_classify(struct sk_buff *skb, struct Qdisc *sch,
203 struct htb_sched *q = qdisc_priv(sch);
204 struct htb_class *cl;
205 struct tcf_result res;
206 struct tcf_proto *tcf;
209 /* allow to select class by setting skb->priority to valid classid;
210 note that nfmark can be used too by attaching filter fw with no
212 if (skb->priority == sch->handle)
213 return HTB_DIRECT; /* X:0 (direct flow) selected */
214 if ((cl = htb_find(skb->priority, sch)) != NULL && cl->level == 0)
217 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
218 tcf = q->filter_list;
219 while (tcf && (result = tc_classify(skb, tcf, &res)) >= 0) {
220 #ifdef CONFIG_NET_CLS_ACT
224 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
229 if ((cl = (void *)res.class) == NULL) {
230 if (res.classid == sch->handle)
231 return HTB_DIRECT; /* X:0 (direct flow) */
232 if ((cl = htb_find(res.classid, sch)) == NULL)
233 break; /* filter selected invalid classid */
236 return cl; /* we hit leaf; return it */
238 /* we have got inner class; apply inner filter chain */
239 tcf = cl->filter_list;
241 /* classification failed; try to use default class */
242 cl = htb_find(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
243 if (!cl || cl->level)
244 return HTB_DIRECT; /* bad default .. this is safe bet */
249 * htb_add_to_id_tree - adds class to the round robin list
251 * Routine adds class to the list (actually tree) sorted by classid.
252 * Make sure that class is not already on such list for given prio.
254 static void htb_add_to_id_tree(struct rb_root *root,
255 struct htb_class *cl, int prio)
257 struct rb_node **p = &root->rb_node, *parent = NULL;
262 c = rb_entry(parent, struct htb_class, node[prio]);
264 if (cl->common.classid > c->common.classid)
265 p = &parent->rb_right;
267 p = &parent->rb_left;
269 rb_link_node(&cl->node[prio], parent, p);
270 rb_insert_color(&cl->node[prio], root);
274 * htb_add_to_wait_tree - adds class to the event queue with delay
276 * The class is added to priority event queue to indicate that class will
277 * change its mode in cl->pq_key microseconds. Make sure that class is not
278 * already in the queue.
280 static void htb_add_to_wait_tree(struct htb_sched *q,
281 struct htb_class *cl, long delay)
283 struct rb_node **p = &q->wait_pq[cl->level].rb_node, *parent = NULL;
285 cl->pq_key = q->now + delay;
286 if (cl->pq_key == q->now)
289 /* update the nearest event cache */
290 if (q->near_ev_cache[cl->level] > cl->pq_key)
291 q->near_ev_cache[cl->level] = cl->pq_key;
296 c = rb_entry(parent, struct htb_class, pq_node);
297 if (cl->pq_key >= c->pq_key)
298 p = &parent->rb_right;
300 p = &parent->rb_left;
302 rb_link_node(&cl->pq_node, parent, p);
303 rb_insert_color(&cl->pq_node, &q->wait_pq[cl->level]);
307 * htb_next_rb_node - finds next node in binary tree
309 * When we are past last key we return NULL.
310 * Average complexity is 2 steps per call.
312 static inline void htb_next_rb_node(struct rb_node **n)
318 * htb_add_class_to_row - add class to its row
320 * The class is added to row at priorities marked in mask.
321 * It does nothing if mask == 0.
323 static inline void htb_add_class_to_row(struct htb_sched *q,
324 struct htb_class *cl, int mask)
326 q->row_mask[cl->level] |= mask;
328 int prio = ffz(~mask);
329 mask &= ~(1 << prio);
330 htb_add_to_id_tree(q->row[cl->level] + prio, cl, prio);
334 /* If this triggers, it is a bug in this code, but it need not be fatal */
335 static void htb_safe_rb_erase(struct rb_node *rb, struct rb_root *root)
337 if (RB_EMPTY_NODE(rb)) {
347 * htb_remove_class_from_row - removes class from its row
349 * The class is removed from row at priorities marked in mask.
350 * It does nothing if mask == 0.
352 static inline void htb_remove_class_from_row(struct htb_sched *q,
353 struct htb_class *cl, int mask)
358 int prio = ffz(~mask);
360 mask &= ~(1 << prio);
361 if (q->ptr[cl->level][prio] == cl->node + prio)
362 htb_next_rb_node(q->ptr[cl->level] + prio);
364 htb_safe_rb_erase(cl->node + prio, q->row[cl->level] + prio);
365 if (!q->row[cl->level][prio].rb_node)
368 q->row_mask[cl->level] &= ~m;
372 * htb_activate_prios - creates active classe's feed chain
374 * The class is connected to ancestors and/or appropriate rows
375 * for priorities it is participating on. cl->cmode must be new
376 * (activated) mode. It does nothing if cl->prio_activity == 0.
378 static void htb_activate_prios(struct htb_sched *q, struct htb_class *cl)
380 struct htb_class *p = cl->parent;
381 long m, mask = cl->prio_activity;
383 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
389 if (p->un.inner.feed[prio].rb_node)
390 /* parent already has its feed in use so that
391 reset bit in mask as parent is already ok */
392 mask &= ~(1 << prio);
394 htb_add_to_id_tree(p->un.inner.feed + prio, cl, prio);
396 p->prio_activity |= mask;
401 if (cl->cmode == HTB_CAN_SEND && mask)
402 htb_add_class_to_row(q, cl, mask);
406 * htb_deactivate_prios - remove class from feed chain
408 * cl->cmode must represent old mode (before deactivation). It does
409 * nothing if cl->prio_activity == 0. Class is removed from all feed
412 static void htb_deactivate_prios(struct htb_sched *q, struct htb_class *cl)
414 struct htb_class *p = cl->parent;
415 long m, mask = cl->prio_activity;
417 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
424 if (p->un.inner.ptr[prio] == cl->node + prio) {
425 /* we are removing child which is pointed to from
426 parent feed - forget the pointer but remember
428 p->un.inner.last_ptr_id[prio] = cl->common.classid;
429 p->un.inner.ptr[prio] = NULL;
432 htb_safe_rb_erase(cl->node + prio, p->un.inner.feed + prio);
434 if (!p->un.inner.feed[prio].rb_node)
438 p->prio_activity &= ~mask;
443 if (cl->cmode == HTB_CAN_SEND && mask)
444 htb_remove_class_from_row(q, cl, mask);
447 static inline long htb_lowater(const struct htb_class *cl)
450 return cl->cmode != HTB_CANT_SEND ? -cl->cbuffer : 0;
454 static inline long htb_hiwater(const struct htb_class *cl)
457 return cl->cmode == HTB_CAN_SEND ? -cl->buffer : 0;
464 * htb_class_mode - computes and returns current class mode
466 * It computes cl's mode at time cl->t_c+diff and returns it. If mode
467 * is not HTB_CAN_SEND then cl->pq_key is updated to time difference
468 * from now to time when cl will change its state.
469 * Also it is worth to note that class mode doesn't change simply
470 * at cl->{c,}tokens == 0 but there can rather be hysteresis of
471 * 0 .. -cl->{c,}buffer range. It is meant to limit number of
472 * mode transitions per time unit. The speed gain is about 1/6.
474 static inline enum htb_cmode
475 htb_class_mode(struct htb_class *cl, long *diff)
479 if ((toks = (cl->ctokens + *diff)) < htb_lowater(cl)) {
481 return HTB_CANT_SEND;
484 if ((toks = (cl->tokens + *diff)) >= htb_hiwater(cl))
488 return HTB_MAY_BORROW;
492 * htb_change_class_mode - changes classe's mode
494 * This should be the only way how to change classe's mode under normal
495 * cirsumstances. Routine will update feed lists linkage, change mode
496 * and add class to the wait event queue if appropriate. New mode should
497 * be different from old one and cl->pq_key has to be valid if changing
498 * to mode other than HTB_CAN_SEND (see htb_add_to_wait_tree).
501 htb_change_class_mode(struct htb_sched *q, struct htb_class *cl, long *diff)
503 enum htb_cmode new_mode = htb_class_mode(cl, diff);
505 if (new_mode == cl->cmode)
508 if (cl->prio_activity) { /* not necessary: speed optimization */
509 if (cl->cmode != HTB_CANT_SEND)
510 htb_deactivate_prios(q, cl);
511 cl->cmode = new_mode;
512 if (new_mode != HTB_CANT_SEND)
513 htb_activate_prios(q, cl);
515 cl->cmode = new_mode;
519 * htb_activate - inserts leaf cl into appropriate active feeds
521 * Routine learns (new) priority of leaf and activates feed chain
522 * for the prio. It can be called on already active leaf safely.
523 * It also adds leaf into droplist.
525 static inline void htb_activate(struct htb_sched *q, struct htb_class *cl)
527 WARN_ON(cl->level || !cl->un.leaf.q || !cl->un.leaf.q->q.qlen);
529 if (!cl->prio_activity) {
530 cl->prio_activity = 1 << (cl->un.leaf.aprio = cl->un.leaf.prio);
531 htb_activate_prios(q, cl);
532 list_add_tail(&cl->un.leaf.drop_list,
533 q->drops + cl->un.leaf.aprio);
538 * htb_deactivate - remove leaf cl from active feeds
540 * Make sure that leaf is active. In the other words it can't be called
541 * with non-active leaf. It also removes class from the drop list.
543 static inline void htb_deactivate(struct htb_sched *q, struct htb_class *cl)
545 WARN_ON(!cl->prio_activity);
547 htb_deactivate_prios(q, cl);
548 cl->prio_activity = 0;
549 list_del_init(&cl->un.leaf.drop_list);
552 static int htb_enqueue(struct sk_buff *skb, struct Qdisc *sch)
555 struct htb_sched *q = qdisc_priv(sch);
556 struct htb_class *cl = htb_classify(skb, sch, &ret);
558 if (cl == HTB_DIRECT) {
559 /* enqueue to helper queue */
560 if (q->direct_queue.qlen < q->direct_qlen) {
561 __skb_queue_tail(&q->direct_queue, skb);
566 return NET_XMIT_DROP;
568 #ifdef CONFIG_NET_CLS_ACT
570 if (ret & __NET_XMIT_BYPASS)
575 } else if ((ret = qdisc_enqueue(skb, cl->un.leaf.q)) != NET_XMIT_SUCCESS) {
576 if (net_xmit_drop_count(ret)) {
582 cl->bstats.packets +=
583 skb_is_gso(skb)?skb_shinfo(skb)->gso_segs:1;
584 cl->bstats.bytes += qdisc_pkt_len(skb);
589 sch->bstats.packets += skb_is_gso(skb)?skb_shinfo(skb)->gso_segs:1;
590 sch->bstats.bytes += qdisc_pkt_len(skb);
591 return NET_XMIT_SUCCESS;
594 /* TODO: requeuing packet charges it to policers again !! */
595 static int htb_requeue(struct sk_buff *skb, struct Qdisc *sch)
598 struct htb_sched *q = qdisc_priv(sch);
599 struct htb_class *cl = htb_classify(skb, sch, &ret);
600 struct sk_buff *tskb;
602 if (cl == HTB_DIRECT) {
603 /* enqueue to helper queue */
604 if (q->direct_queue.qlen < q->direct_qlen) {
605 __skb_queue_head(&q->direct_queue, skb);
607 __skb_queue_head(&q->direct_queue, skb);
608 tskb = __skb_dequeue_tail(&q->direct_queue);
613 #ifdef CONFIG_NET_CLS_ACT
615 if (ret & __NET_XMIT_BYPASS)
620 } else if ((ret = cl->un.leaf.q->ops->requeue(skb, cl->un.leaf.q)) !=
622 if (net_xmit_drop_count(ret)) {
631 sch->qstats.requeues++;
632 return NET_XMIT_SUCCESS;
636 * htb_charge_class - charges amount "bytes" to leaf and ancestors
638 * Routine assumes that packet "bytes" long was dequeued from leaf cl
639 * borrowing from "level". It accounts bytes to ceil leaky bucket for
640 * leaf and all ancestors and to rate bucket for ancestors at levels
641 * "level" and higher. It also handles possible change of mode resulting
642 * from the update. Note that mode can also increase here (MAY_BORROW to
643 * CAN_SEND) because we can use more precise clock that event queue here.
644 * In such case we remove class from event queue first.
646 static void htb_charge_class(struct htb_sched *q, struct htb_class *cl,
647 int level, struct sk_buff *skb)
649 int bytes = qdisc_pkt_len(skb);
651 enum htb_cmode old_mode;
653 #define HTB_ACCNT(T,B,R) toks = diff + cl->T; \
654 if (toks > cl->B) toks = cl->B; \
655 toks -= L2T(cl, cl->R, bytes); \
656 if (toks <= -cl->mbuffer) toks = 1-cl->mbuffer; \
660 diff = psched_tdiff_bounded(q->now, cl->t_c, cl->mbuffer);
661 if (cl->level >= level) {
662 if (cl->level == level)
664 HTB_ACCNT(tokens, buffer, rate);
666 cl->xstats.borrows++;
667 cl->tokens += diff; /* we moved t_c; update tokens */
669 HTB_ACCNT(ctokens, cbuffer, ceil);
672 old_mode = cl->cmode;
674 htb_change_class_mode(q, cl, &diff);
675 if (old_mode != cl->cmode) {
676 if (old_mode != HTB_CAN_SEND)
677 htb_safe_rb_erase(&cl->pq_node, q->wait_pq + cl->level);
678 if (cl->cmode != HTB_CAN_SEND)
679 htb_add_to_wait_tree(q, cl, diff);
682 /* update byte stats except for leaves which are already updated */
684 cl->bstats.bytes += bytes;
685 cl->bstats.packets += skb_is_gso(skb)?
686 skb_shinfo(skb)->gso_segs:1;
693 * htb_do_events - make mode changes to classes at the level
695 * Scans event queue for pending events and applies them. Returns time of
696 * next pending event (0 for no event in pq).
697 * Note: Applied are events whose have cl->pq_key <= q->now.
699 static psched_time_t htb_do_events(struct htb_sched *q, int level)
701 /* don't run for longer than 2 jiffies; 2 is used instead of
702 1 to simplify things when jiffy is going to be incremented
704 unsigned long stop_at = jiffies + 2;
705 while (time_before(jiffies, stop_at)) {
706 struct htb_class *cl;
708 struct rb_node *p = rb_first(&q->wait_pq[level]);
713 cl = rb_entry(p, struct htb_class, pq_node);
714 if (cl->pq_key > q->now)
717 htb_safe_rb_erase(p, q->wait_pq + level);
718 diff = psched_tdiff_bounded(q->now, cl->t_c, cl->mbuffer);
719 htb_change_class_mode(q, cl, &diff);
720 if (cl->cmode != HTB_CAN_SEND)
721 htb_add_to_wait_tree(q, cl, diff);
723 /* too much load - let's continue on next jiffie */
724 return q->now + PSCHED_TICKS_PER_SEC / HZ;
727 /* Returns class->node+prio from id-tree where classe's id is >= id. NULL
728 is no such one exists. */
729 static struct rb_node *htb_id_find_next_upper(int prio, struct rb_node *n,
732 struct rb_node *r = NULL;
734 struct htb_class *cl =
735 rb_entry(n, struct htb_class, node[prio]);
736 if (id == cl->common.classid)
739 if (id > cl->common.classid) {
750 * htb_lookup_leaf - returns next leaf class in DRR order
752 * Find leaf where current feed pointers points to.
754 static struct htb_class *htb_lookup_leaf(struct rb_root *tree, int prio,
755 struct rb_node **pptr, u32 * pid)
759 struct rb_node *root;
760 struct rb_node **pptr;
762 } stk[TC_HTB_MAXDEPTH], *sp = stk;
764 WARN_ON(!tree->rb_node);
765 sp->root = tree->rb_node;
769 for (i = 0; i < 65535; i++) {
770 if (!*sp->pptr && *sp->pid) {
771 /* ptr was invalidated but id is valid - try to recover
772 the original or next ptr */
774 htb_id_find_next_upper(prio, sp->root, *sp->pid);
776 *sp->pid = 0; /* ptr is valid now so that remove this hint as it
777 can become out of date quickly */
778 if (!*sp->pptr) { /* we are at right end; rewind & go up */
779 *sp->pptr = sp->root;
780 while ((*sp->pptr)->rb_left)
781 *sp->pptr = (*sp->pptr)->rb_left;
787 htb_next_rb_node(sp->pptr);
790 struct htb_class *cl;
791 cl = rb_entry(*sp->pptr, struct htb_class, node[prio]);
794 (++sp)->root = cl->un.inner.feed[prio].rb_node;
795 sp->pptr = cl->un.inner.ptr + prio;
796 sp->pid = cl->un.inner.last_ptr_id + prio;
803 /* dequeues packet at given priority and level; call only if
804 you are sure that there is active class at prio/level */
805 static struct sk_buff *htb_dequeue_tree(struct htb_sched *q, int prio,
808 struct sk_buff *skb = NULL;
809 struct htb_class *cl, *start;
810 /* look initial class up in the row */
811 start = cl = htb_lookup_leaf(q->row[level] + prio, prio,
812 q->ptr[level] + prio,
813 q->last_ptr_id[level] + prio);
821 /* class can be empty - it is unlikely but can be true if leaf
822 qdisc drops packets in enqueue routine or if someone used
823 graft operation on the leaf since last dequeue;
824 simply deactivate and skip such class */
825 if (unlikely(cl->un.leaf.q->q.qlen == 0)) {
826 struct htb_class *next;
827 htb_deactivate(q, cl);
829 /* row/level might become empty */
830 if ((q->row_mask[level] & (1 << prio)) == 0)
833 next = htb_lookup_leaf(q->row[level] + prio,
834 prio, q->ptr[level] + prio,
835 q->last_ptr_id[level] + prio);
837 if (cl == start) /* fix start if we just deleted it */
843 skb = cl->un.leaf.q->dequeue(cl->un.leaf.q);
844 if (likely(skb != NULL))
848 "htb: class %X isn't work conserving ?!\n",
853 htb_next_rb_node((level ? cl->parent->un.inner.ptr : q->
855 cl = htb_lookup_leaf(q->row[level] + prio, prio,
856 q->ptr[level] + prio,
857 q->last_ptr_id[level] + prio);
859 } while (cl != start);
861 if (likely(skb != NULL)) {
862 cl->un.leaf.deficit[level] -= qdisc_pkt_len(skb);
863 if (cl->un.leaf.deficit[level] < 0) {
864 cl->un.leaf.deficit[level] += cl->un.leaf.quantum;
865 htb_next_rb_node((level ? cl->parent->un.inner.ptr : q->
868 /* this used to be after charge_class but this constelation
869 gives us slightly better performance */
870 if (!cl->un.leaf.q->q.qlen)
871 htb_deactivate(q, cl);
872 htb_charge_class(q, cl, level, skb);
877 static struct sk_buff *htb_dequeue(struct Qdisc *sch)
879 struct sk_buff *skb = NULL;
880 struct htb_sched *q = qdisc_priv(sch);
882 psched_time_t next_event;
884 /* try to dequeue direct packets as high prio (!) to minimize cpu work */
885 skb = __skb_dequeue(&q->direct_queue);
887 sch->flags &= ~TCQ_F_THROTTLED;
894 q->now = psched_get_time();
896 next_event = q->now + 5 * PSCHED_TICKS_PER_SEC;
898 for (level = 0; level < TC_HTB_MAXDEPTH; level++) {
899 /* common case optimization - skip event handler quickly */
903 if (q->now >= q->near_ev_cache[level]) {
904 event = htb_do_events(q, level);
906 event = q->now + PSCHED_TICKS_PER_SEC;
907 q->near_ev_cache[level] = event;
909 event = q->near_ev_cache[level];
911 if (event && next_event > event)
914 m = ~q->row_mask[level];
915 while (m != (int)(-1)) {
918 skb = htb_dequeue_tree(q, prio, level);
919 if (likely(skb != NULL)) {
921 sch->flags &= ~TCQ_F_THROTTLED;
926 sch->qstats.overlimits++;
927 qdisc_watchdog_schedule(&q->watchdog, next_event);
932 /* try to drop from each class (by prio) until one succeed */
933 static unsigned int htb_drop(struct Qdisc *sch)
935 struct htb_sched *q = qdisc_priv(sch);
938 for (prio = TC_HTB_NUMPRIO - 1; prio >= 0; prio--) {
940 list_for_each(p, q->drops + prio) {
941 struct htb_class *cl = list_entry(p, struct htb_class,
944 if (cl->un.leaf.q->ops->drop &&
945 (len = cl->un.leaf.q->ops->drop(cl->un.leaf.q))) {
947 if (!cl->un.leaf.q->q.qlen)
948 htb_deactivate(q, cl);
956 /* reset all classes */
957 /* always caled under BH & queue lock */
958 static void htb_reset(struct Qdisc *sch)
960 struct htb_sched *q = qdisc_priv(sch);
961 struct htb_class *cl;
962 struct hlist_node *n;
965 for (i = 0; i < q->clhash.hashsize; i++) {
966 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode) {
968 memset(&cl->un.inner, 0, sizeof(cl->un.inner));
971 qdisc_reset(cl->un.leaf.q);
972 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
974 cl->prio_activity = 0;
975 cl->cmode = HTB_CAN_SEND;
979 qdisc_watchdog_cancel(&q->watchdog);
980 __skb_queue_purge(&q->direct_queue);
982 memset(q->row, 0, sizeof(q->row));
983 memset(q->row_mask, 0, sizeof(q->row_mask));
984 memset(q->wait_pq, 0, sizeof(q->wait_pq));
985 memset(q->ptr, 0, sizeof(q->ptr));
986 for (i = 0; i < TC_HTB_NUMPRIO; i++)
987 INIT_LIST_HEAD(q->drops + i);
990 static const struct nla_policy htb_policy[TCA_HTB_MAX + 1] = {
991 [TCA_HTB_PARMS] = { .len = sizeof(struct tc_htb_opt) },
992 [TCA_HTB_INIT] = { .len = sizeof(struct tc_htb_glob) },
993 [TCA_HTB_CTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
994 [TCA_HTB_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
997 static int htb_init(struct Qdisc *sch, struct nlattr *opt)
999 struct htb_sched *q = qdisc_priv(sch);
1000 struct nlattr *tb[TCA_HTB_INIT + 1];
1001 struct tc_htb_glob *gopt;
1008 err = nla_parse_nested(tb, TCA_HTB_INIT, opt, htb_policy);
1012 if (tb[TCA_HTB_INIT] == NULL) {
1013 printk(KERN_ERR "HTB: hey probably you have bad tc tool ?\n");
1016 gopt = nla_data(tb[TCA_HTB_INIT]);
1017 if (gopt->version != HTB_VER >> 16) {
1019 "HTB: need tc/htb version %d (minor is %d), you have %d\n",
1020 HTB_VER >> 16, HTB_VER & 0xffff, gopt->version);
1024 err = qdisc_class_hash_init(&q->clhash);
1027 for (i = 0; i < TC_HTB_NUMPRIO; i++)
1028 INIT_LIST_HEAD(q->drops + i);
1030 qdisc_watchdog_init(&q->watchdog, sch);
1031 skb_queue_head_init(&q->direct_queue);
1033 q->direct_qlen = qdisc_dev(sch)->tx_queue_len;
1034 if (q->direct_qlen < 2) /* some devices have zero tx_queue_len */
1037 if ((q->rate2quantum = gopt->rate2quantum) < 1)
1038 q->rate2quantum = 1;
1039 q->defcls = gopt->defcls;
1044 static int htb_dump(struct Qdisc *sch, struct sk_buff *skb)
1046 spinlock_t *root_lock = qdisc_root_sleeping_lock(sch);
1047 struct htb_sched *q = qdisc_priv(sch);
1048 struct nlattr *nest;
1049 struct tc_htb_glob gopt;
1051 spin_lock_bh(root_lock);
1053 gopt.direct_pkts = q->direct_pkts;
1054 gopt.version = HTB_VER;
1055 gopt.rate2quantum = q->rate2quantum;
1056 gopt.defcls = q->defcls;
1059 nest = nla_nest_start(skb, TCA_OPTIONS);
1061 goto nla_put_failure;
1062 NLA_PUT(skb, TCA_HTB_INIT, sizeof(gopt), &gopt);
1063 nla_nest_end(skb, nest);
1065 spin_unlock_bh(root_lock);
1069 spin_unlock_bh(root_lock);
1070 nla_nest_cancel(skb, nest);
1074 static int htb_dump_class(struct Qdisc *sch, unsigned long arg,
1075 struct sk_buff *skb, struct tcmsg *tcm)
1077 struct htb_class *cl = (struct htb_class *)arg;
1078 spinlock_t *root_lock = qdisc_root_sleeping_lock(sch);
1079 struct nlattr *nest;
1080 struct tc_htb_opt opt;
1082 spin_lock_bh(root_lock);
1083 tcm->tcm_parent = cl->parent ? cl->parent->common.classid : TC_H_ROOT;
1084 tcm->tcm_handle = cl->common.classid;
1085 if (!cl->level && cl->un.leaf.q)
1086 tcm->tcm_info = cl->un.leaf.q->handle;
1088 nest = nla_nest_start(skb, TCA_OPTIONS);
1090 goto nla_put_failure;
1092 memset(&opt, 0, sizeof(opt));
1094 opt.rate = cl->rate->rate;
1095 opt.buffer = cl->buffer;
1096 opt.ceil = cl->ceil->rate;
1097 opt.cbuffer = cl->cbuffer;
1098 opt.quantum = cl->un.leaf.quantum;
1099 opt.prio = cl->un.leaf.prio;
1100 opt.level = cl->level;
1101 NLA_PUT(skb, TCA_HTB_PARMS, sizeof(opt), &opt);
1103 nla_nest_end(skb, nest);
1104 spin_unlock_bh(root_lock);
1108 spin_unlock_bh(root_lock);
1109 nla_nest_cancel(skb, nest);
1114 htb_dump_class_stats(struct Qdisc *sch, unsigned long arg, struct gnet_dump *d)
1116 struct htb_class *cl = (struct htb_class *)arg;
1118 if (!cl->level && cl->un.leaf.q)
1119 cl->qstats.qlen = cl->un.leaf.q->q.qlen;
1120 cl->xstats.tokens = cl->tokens;
1121 cl->xstats.ctokens = cl->ctokens;
1123 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
1124 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1125 gnet_stats_copy_queue(d, &cl->qstats) < 0)
1128 return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1131 static int htb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1134 struct htb_class *cl = (struct htb_class *)arg;
1136 if (cl && !cl->level) {
1138 (new = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
1140 cl->common.classid))
1144 if ((*old = xchg(&cl->un.leaf.q, new)) != NULL) {
1145 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
1148 sch_tree_unlock(sch);
1154 static struct Qdisc *htb_leaf(struct Qdisc *sch, unsigned long arg)
1156 struct htb_class *cl = (struct htb_class *)arg;
1157 return (cl && !cl->level) ? cl->un.leaf.q : NULL;
1160 static void htb_qlen_notify(struct Qdisc *sch, unsigned long arg)
1162 struct htb_class *cl = (struct htb_class *)arg;
1164 if (cl->un.leaf.q->q.qlen == 0)
1165 htb_deactivate(qdisc_priv(sch), cl);
1168 static unsigned long htb_get(struct Qdisc *sch, u32 classid)
1170 struct htb_class *cl = htb_find(classid, sch);
1173 return (unsigned long)cl;
1176 static inline int htb_parent_last_child(struct htb_class *cl)
1179 /* the root class */
1181 if (cl->parent->children > 1)
1182 /* not the last child */
1187 static void htb_parent_to_leaf(struct htb_sched *q, struct htb_class *cl,
1188 struct Qdisc *new_q)
1190 struct htb_class *parent = cl->parent;
1192 WARN_ON(cl->level || !cl->un.leaf.q || cl->prio_activity);
1194 if (parent->cmode != HTB_CAN_SEND)
1195 htb_safe_rb_erase(&parent->pq_node, q->wait_pq + parent->level);
1198 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1199 INIT_LIST_HEAD(&parent->un.leaf.drop_list);
1200 parent->un.leaf.q = new_q ? new_q : &noop_qdisc;
1201 parent->un.leaf.quantum = parent->quantum;
1202 parent->un.leaf.prio = parent->prio;
1203 parent->tokens = parent->buffer;
1204 parent->ctokens = parent->cbuffer;
1205 parent->t_c = psched_get_time();
1206 parent->cmode = HTB_CAN_SEND;
1209 static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
1212 WARN_ON(!cl->un.leaf.q);
1213 qdisc_destroy(cl->un.leaf.q);
1215 gen_kill_estimator(&cl->bstats, &cl->rate_est);
1216 qdisc_put_rtab(cl->rate);
1217 qdisc_put_rtab(cl->ceil);
1219 tcf_destroy_chain(&cl->filter_list);
1223 /* always caled under BH & queue lock */
1224 static void htb_destroy(struct Qdisc *sch)
1226 struct htb_sched *q = qdisc_priv(sch);
1227 struct hlist_node *n, *next;
1228 struct htb_class *cl;
1231 qdisc_watchdog_cancel(&q->watchdog);
1232 /* This line used to be after htb_destroy_class call below
1233 and surprisingly it worked in 2.4. But it must precede it
1234 because filter need its target class alive to be able to call
1235 unbind_filter on it (without Oops). */
1236 tcf_destroy_chain(&q->filter_list);
1238 for (i = 0; i < q->clhash.hashsize; i++) {
1239 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode)
1240 tcf_destroy_chain(&cl->filter_list);
1242 for (i = 0; i < q->clhash.hashsize; i++) {
1243 hlist_for_each_entry_safe(cl, n, next, &q->clhash.hash[i],
1245 htb_destroy_class(sch, cl);
1247 qdisc_class_hash_destroy(&q->clhash);
1248 __skb_queue_purge(&q->direct_queue);
1251 static int htb_delete(struct Qdisc *sch, unsigned long arg)
1253 struct htb_sched *q = qdisc_priv(sch);
1254 struct htb_class *cl = (struct htb_class *)arg;
1256 struct Qdisc *new_q = NULL;
1259 // TODO: why don't allow to delete subtree ? references ? does
1260 // tc subsys quarantee us that in htb_destroy it holds no class
1261 // refs so that we can remove children safely there ?
1262 if (cl->children || cl->filter_cnt)
1265 if (!cl->level && htb_parent_last_child(cl)) {
1266 new_q = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
1268 cl->parent->common.classid);
1275 qlen = cl->un.leaf.q->q.qlen;
1276 qdisc_reset(cl->un.leaf.q);
1277 qdisc_tree_decrease_qlen(cl->un.leaf.q, qlen);
1280 /* delete from hash and active; remainder in destroy_class */
1281 qdisc_class_hash_remove(&q->clhash, &cl->common);
1283 cl->parent->children--;
1285 if (cl->prio_activity)
1286 htb_deactivate(q, cl);
1288 if (cl->cmode != HTB_CAN_SEND)
1289 htb_safe_rb_erase(&cl->pq_node, q->wait_pq + cl->level);
1292 htb_parent_to_leaf(q, cl, new_q);
1294 if (--cl->refcnt == 0)
1295 htb_destroy_class(sch, cl);
1297 sch_tree_unlock(sch);
1301 static void htb_put(struct Qdisc *sch, unsigned long arg)
1303 struct htb_class *cl = (struct htb_class *)arg;
1305 if (--cl->refcnt == 0)
1306 htb_destroy_class(sch, cl);
1309 static int htb_change_class(struct Qdisc *sch, u32 classid,
1310 u32 parentid, struct nlattr **tca,
1314 struct htb_sched *q = qdisc_priv(sch);
1315 struct htb_class *cl = (struct htb_class *)*arg, *parent;
1316 struct nlattr *opt = tca[TCA_OPTIONS];
1317 struct qdisc_rate_table *rtab = NULL, *ctab = NULL;
1318 struct nlattr *tb[TCA_HTB_RTAB + 1];
1319 struct tc_htb_opt *hopt;
1321 /* extract all subattrs from opt attr */
1325 err = nla_parse_nested(tb, TCA_HTB_RTAB, opt, htb_policy);
1330 if (tb[TCA_HTB_PARMS] == NULL)
1333 parent = parentid == TC_H_ROOT ? NULL : htb_find(parentid, sch);
1335 hopt = nla_data(tb[TCA_HTB_PARMS]);
1337 rtab = qdisc_get_rtab(&hopt->rate, tb[TCA_HTB_RTAB]);
1338 ctab = qdisc_get_rtab(&hopt->ceil, tb[TCA_HTB_CTAB]);
1342 if (!cl) { /* new class */
1343 struct Qdisc *new_q;
1347 struct gnet_estimator opt;
1350 .nla_len = nla_attr_size(sizeof(est.opt)),
1351 .nla_type = TCA_RATE,
1354 /* 4s interval, 16s averaging constant */
1360 /* check for valid classid */
1361 if (!classid || TC_H_MAJ(classid ^ sch->handle)
1362 || htb_find(classid, sch))
1365 /* check maximal depth */
1366 if (parent && parent->parent && parent->parent->level < 2) {
1367 printk(KERN_ERR "htb: tree is too deep\n");
1371 if ((cl = kzalloc(sizeof(*cl), GFP_KERNEL)) == NULL)
1374 gen_new_estimator(&cl->bstats, &cl->rate_est,
1375 qdisc_root_sleeping_lock(sch),
1376 tca[TCA_RATE] ? : &est.nla);
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(qdisc_dev(sch), sch->dev_queue,
1389 &pfifo_qdisc_ops, classid);
1391 if (parent && !parent->level) {
1392 unsigned int qlen = parent->un.leaf.q->q.qlen;
1394 /* turn parent into inner node */
1395 qdisc_reset(parent->un.leaf.q);
1396 qdisc_tree_decrease_qlen(parent->un.leaf.q, qlen);
1397 qdisc_destroy(parent->un.leaf.q);
1398 if (parent->prio_activity)
1399 htb_deactivate(q, parent);
1401 /* remove from evt list because of level change */
1402 if (parent->cmode != HTB_CAN_SEND) {
1403 htb_safe_rb_erase(&parent->pq_node, q->wait_pq);
1404 parent->cmode = HTB_CAN_SEND;
1406 parent->level = (parent->parent ? parent->parent->level
1407 : TC_HTB_MAXDEPTH) - 1;
1408 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1410 /* leaf (we) needs elementary qdisc */
1411 cl->un.leaf.q = new_q ? new_q : &noop_qdisc;
1413 cl->common.classid = classid;
1414 cl->parent = parent;
1416 /* set class to be in HTB_CAN_SEND state */
1417 cl->tokens = hopt->buffer;
1418 cl->ctokens = hopt->cbuffer;
1419 cl->mbuffer = 60 * PSCHED_TICKS_PER_SEC; /* 1min */
1420 cl->t_c = psched_get_time();
1421 cl->cmode = HTB_CAN_SEND;
1423 /* attach to the hash list and parent's family */
1424 qdisc_class_hash_insert(&q->clhash, &cl->common);
1429 gen_replace_estimator(&cl->bstats, &cl->rate_est,
1430 qdisc_root_sleeping_lock(sch),
1435 /* it used to be a nasty bug here, we have to check that node
1436 is really leaf before changing cl->un.leaf ! */
1438 cl->un.leaf.quantum = rtab->rate.rate / q->rate2quantum;
1439 if (!hopt->quantum && cl->un.leaf.quantum < 1000) {
1441 "HTB: quantum of class %X is small. Consider r2q change.\n",
1442 cl->common.classid);
1443 cl->un.leaf.quantum = 1000;
1445 if (!hopt->quantum && cl->un.leaf.quantum > 200000) {
1447 "HTB: quantum of class %X is big. Consider r2q change.\n",
1448 cl->common.classid);
1449 cl->un.leaf.quantum = 200000;
1452 cl->un.leaf.quantum = hopt->quantum;
1453 if ((cl->un.leaf.prio = hopt->prio) >= TC_HTB_NUMPRIO)
1454 cl->un.leaf.prio = TC_HTB_NUMPRIO - 1;
1456 /* backup for htb_parent_to_leaf */
1457 cl->quantum = cl->un.leaf.quantum;
1458 cl->prio = cl->un.leaf.prio;
1461 cl->buffer = hopt->buffer;
1462 cl->cbuffer = hopt->cbuffer;
1464 qdisc_put_rtab(cl->rate);
1467 qdisc_put_rtab(cl->ceil);
1469 sch_tree_unlock(sch);
1471 qdisc_class_hash_grow(sch, &q->clhash);
1473 *arg = (unsigned long)cl;
1478 qdisc_put_rtab(rtab);
1480 qdisc_put_rtab(ctab);
1484 static struct tcf_proto **htb_find_tcf(struct Qdisc *sch, unsigned long arg)
1486 struct htb_sched *q = qdisc_priv(sch);
1487 struct htb_class *cl = (struct htb_class *)arg;
1488 struct tcf_proto **fl = cl ? &cl->filter_list : &q->filter_list;
1493 static unsigned long htb_bind_filter(struct Qdisc *sch, unsigned long parent,
1496 struct htb_class *cl = htb_find(classid, sch);
1498 /*if (cl && !cl->level) return 0;
1499 The line above used to be there to prevent attaching filters to
1500 leaves. But at least tc_index filter uses this just to get class
1501 for other reasons so that we have to allow for it.
1503 19.6.2002 As Werner explained it is ok - bind filter is just
1504 another way to "lock" the class - unlike "get" this lock can
1505 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_class *cl = (struct htb_class *)arg;
1520 static void htb_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1522 struct htb_sched *q = qdisc_priv(sch);
1523 struct htb_class *cl;
1524 struct hlist_node *n;
1530 for (i = 0; i < q->clhash.hashsize; i++) {
1531 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode) {
1532 if (arg->count < arg->skip) {
1536 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1545 static const struct Qdisc_class_ops htb_class_ops = {
1548 .qlen_notify = htb_qlen_notify,
1551 .change = htb_change_class,
1552 .delete = htb_delete,
1554 .tcf_chain = htb_find_tcf,
1555 .bind_tcf = htb_bind_filter,
1556 .unbind_tcf = htb_unbind_filter,
1557 .dump = htb_dump_class,
1558 .dump_stats = htb_dump_class_stats,
1561 static struct Qdisc_ops htb_qdisc_ops __read_mostly = {
1563 .cl_ops = &htb_class_ops,
1565 .priv_size = sizeof(struct htb_sched),
1566 .enqueue = htb_enqueue,
1567 .dequeue = htb_dequeue,
1568 .requeue = htb_requeue,
1572 .destroy = htb_destroy,
1573 .change = NULL /* htb_change */,
1575 .owner = THIS_MODULE,
1578 static int __init htb_module_init(void)
1580 return register_qdisc(&htb_qdisc_ops);
1582 static void __exit htb_module_exit(void)
1584 unregister_qdisc(&htb_qdisc_ops);
1587 module_init(htb_module_init)
1588 module_exit(htb_module_exit)
1589 MODULE_LICENSE("GPL");