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 */
87 int prio; /* these two are used only by leaves... */
88 int quantum; /* but stored for parent-to-leaf return */
91 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 */
129 struct Qdisc_class_hash clhash;
130 struct list_head drops[TC_HTB_NUMPRIO];/* active leaves (for drops) */
132 /* self list - roots of self generating tree */
133 struct rb_root row[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
134 int row_mask[TC_HTB_MAXDEPTH];
135 struct rb_node *ptr[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
136 u32 last_ptr_id[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
138 /* self wait list - roots of wait PQs per row */
139 struct rb_root wait_pq[TC_HTB_MAXDEPTH];
141 /* time of nearest event per level (row) */
142 psched_time_t near_ev_cache[TC_HTB_MAXDEPTH];
144 int defcls; /* class where unclassified flows go to */
146 /* filters for qdisc itself */
147 struct tcf_proto *filter_list;
149 int rate2quantum; /* quant = rate / rate2quantum */
150 psched_time_t now; /* cached dequeue time */
151 struct qdisc_watchdog watchdog;
153 /* non shaped skbs; let them go directly thru */
154 struct sk_buff_head direct_queue;
155 int direct_qlen; /* max qlen of above */
160 /* find class in global hash table using given handle */
161 static inline struct htb_class *htb_find(u32 handle, struct Qdisc *sch)
163 struct htb_sched *q = qdisc_priv(sch);
164 struct Qdisc_class_common *clc;
166 clc = qdisc_class_find(&q->clhash, handle);
169 return container_of(clc, struct htb_class, common);
173 * htb_classify - classify a packet into class
175 * It returns NULL if the packet should be dropped or -1 if the packet
176 * should be passed directly thru. In all other cases leaf class is returned.
177 * We allow direct class selection by classid in priority. The we examine
178 * filters in qdisc and in inner nodes (if higher filter points to the inner
179 * node). If we end up with classid MAJOR:0 we enqueue the skb into special
180 * internal fifo (direct). These packets then go directly thru. If we still
181 * have no valid leaf we try to use MAJOR:default leaf. It still unsuccessfull
182 * then finish and return direct queue.
184 #define HTB_DIRECT (struct htb_class*)-1
186 static struct htb_class *htb_classify(struct sk_buff *skb, struct Qdisc *sch,
189 struct htb_sched *q = qdisc_priv(sch);
190 struct htb_class *cl;
191 struct tcf_result res;
192 struct tcf_proto *tcf;
195 /* allow to select class by setting skb->priority to valid classid;
196 note that nfmark can be used too by attaching filter fw with no
198 if (skb->priority == sch->handle)
199 return HTB_DIRECT; /* X:0 (direct flow) selected */
200 if ((cl = htb_find(skb->priority, sch)) != NULL && cl->level == 0)
203 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
204 tcf = q->filter_list;
205 while (tcf && (result = tc_classify(skb, tcf, &res)) >= 0) {
206 #ifdef CONFIG_NET_CLS_ACT
210 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
215 if ((cl = (void *)res.class) == NULL) {
216 if (res.classid == sch->handle)
217 return HTB_DIRECT; /* X:0 (direct flow) */
218 if ((cl = htb_find(res.classid, sch)) == NULL)
219 break; /* filter selected invalid classid */
222 return cl; /* we hit leaf; return it */
224 /* we have got inner class; apply inner filter chain */
225 tcf = cl->filter_list;
227 /* classification failed; try to use default class */
228 cl = htb_find(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
229 if (!cl || cl->level)
230 return HTB_DIRECT; /* bad default .. this is safe bet */
235 * htb_add_to_id_tree - adds class to the round robin list
237 * Routine adds class to the list (actually tree) sorted by classid.
238 * Make sure that class is not already on such list for given prio.
240 static void htb_add_to_id_tree(struct rb_root *root,
241 struct htb_class *cl, int prio)
243 struct rb_node **p = &root->rb_node, *parent = NULL;
248 c = rb_entry(parent, struct htb_class, node[prio]);
250 if (cl->common.classid > c->common.classid)
251 p = &parent->rb_right;
253 p = &parent->rb_left;
255 rb_link_node(&cl->node[prio], parent, p);
256 rb_insert_color(&cl->node[prio], root);
260 * htb_add_to_wait_tree - adds class to the event queue with delay
262 * The class is added to priority event queue to indicate that class will
263 * change its mode in cl->pq_key microseconds. Make sure that class is not
264 * already in the queue.
266 static void htb_add_to_wait_tree(struct htb_sched *q,
267 struct htb_class *cl, long delay)
269 struct rb_node **p = &q->wait_pq[cl->level].rb_node, *parent = NULL;
271 cl->pq_key = q->now + delay;
272 if (cl->pq_key == q->now)
275 /* update the nearest event cache */
276 if (q->near_ev_cache[cl->level] > cl->pq_key)
277 q->near_ev_cache[cl->level] = cl->pq_key;
282 c = rb_entry(parent, struct htb_class, pq_node);
283 if (cl->pq_key >= c->pq_key)
284 p = &parent->rb_right;
286 p = &parent->rb_left;
288 rb_link_node(&cl->pq_node, parent, p);
289 rb_insert_color(&cl->pq_node, &q->wait_pq[cl->level]);
293 * htb_next_rb_node - finds next node in binary tree
295 * When we are past last key we return NULL.
296 * Average complexity is 2 steps per call.
298 static inline void htb_next_rb_node(struct rb_node **n)
304 * htb_add_class_to_row - add class to its row
306 * The class is added to row at priorities marked in mask.
307 * It does nothing if mask == 0.
309 static inline void htb_add_class_to_row(struct htb_sched *q,
310 struct htb_class *cl, int mask)
312 q->row_mask[cl->level] |= mask;
314 int prio = ffz(~mask);
315 mask &= ~(1 << prio);
316 htb_add_to_id_tree(q->row[cl->level] + prio, cl, prio);
320 /* If this triggers, it is a bug in this code, but it need not be fatal */
321 static void htb_safe_rb_erase(struct rb_node *rb, struct rb_root *root)
323 if (RB_EMPTY_NODE(rb)) {
333 * htb_remove_class_from_row - removes class from its row
335 * The class is removed from row at priorities marked in mask.
336 * It does nothing if mask == 0.
338 static inline void htb_remove_class_from_row(struct htb_sched *q,
339 struct htb_class *cl, int mask)
344 int prio = ffz(~mask);
346 mask &= ~(1 << prio);
347 if (q->ptr[cl->level][prio] == cl->node + prio)
348 htb_next_rb_node(q->ptr[cl->level] + prio);
350 htb_safe_rb_erase(cl->node + prio, q->row[cl->level] + prio);
351 if (!q->row[cl->level][prio].rb_node)
354 q->row_mask[cl->level] &= ~m;
358 * htb_activate_prios - creates active classe's feed chain
360 * The class is connected to ancestors and/or appropriate rows
361 * for priorities it is participating on. cl->cmode must be new
362 * (activated) mode. It does nothing if cl->prio_activity == 0.
364 static void htb_activate_prios(struct htb_sched *q, struct htb_class *cl)
366 struct htb_class *p = cl->parent;
367 long m, mask = cl->prio_activity;
369 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
375 if (p->un.inner.feed[prio].rb_node)
376 /* parent already has its feed in use so that
377 reset bit in mask as parent is already ok */
378 mask &= ~(1 << prio);
380 htb_add_to_id_tree(p->un.inner.feed + prio, cl, prio);
382 p->prio_activity |= mask;
387 if (cl->cmode == HTB_CAN_SEND && mask)
388 htb_add_class_to_row(q, cl, mask);
392 * htb_deactivate_prios - remove class from feed chain
394 * cl->cmode must represent old mode (before deactivation). It does
395 * nothing if cl->prio_activity == 0. Class is removed from all feed
398 static void htb_deactivate_prios(struct htb_sched *q, struct htb_class *cl)
400 struct htb_class *p = cl->parent;
401 long m, mask = cl->prio_activity;
403 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
410 if (p->un.inner.ptr[prio] == cl->node + prio) {
411 /* we are removing child which is pointed to from
412 parent feed - forget the pointer but remember
414 p->un.inner.last_ptr_id[prio] = cl->common.classid;
415 p->un.inner.ptr[prio] = NULL;
418 htb_safe_rb_erase(cl->node + prio, p->un.inner.feed + prio);
420 if (!p->un.inner.feed[prio].rb_node)
424 p->prio_activity &= ~mask;
429 if (cl->cmode == HTB_CAN_SEND && mask)
430 htb_remove_class_from_row(q, cl, mask);
433 static inline long htb_lowater(const struct htb_class *cl)
436 return cl->cmode != HTB_CANT_SEND ? -cl->cbuffer : 0;
440 static inline long htb_hiwater(const struct htb_class *cl)
443 return cl->cmode == HTB_CAN_SEND ? -cl->buffer : 0;
450 * htb_class_mode - computes and returns current class mode
452 * It computes cl's mode at time cl->t_c+diff and returns it. If mode
453 * is not HTB_CAN_SEND then cl->pq_key is updated to time difference
454 * from now to time when cl will change its state.
455 * Also it is worth to note that class mode doesn't change simply
456 * at cl->{c,}tokens == 0 but there can rather be hysteresis of
457 * 0 .. -cl->{c,}buffer range. It is meant to limit number of
458 * mode transitions per time unit. The speed gain is about 1/6.
460 static inline enum htb_cmode
461 htb_class_mode(struct htb_class *cl, long *diff)
465 if ((toks = (cl->ctokens + *diff)) < htb_lowater(cl)) {
467 return HTB_CANT_SEND;
470 if ((toks = (cl->tokens + *diff)) >= htb_hiwater(cl))
474 return HTB_MAY_BORROW;
478 * htb_change_class_mode - changes classe's mode
480 * This should be the only way how to change classe's mode under normal
481 * cirsumstances. Routine will update feed lists linkage, change mode
482 * and add class to the wait event queue if appropriate. New mode should
483 * be different from old one and cl->pq_key has to be valid if changing
484 * to mode other than HTB_CAN_SEND (see htb_add_to_wait_tree).
487 htb_change_class_mode(struct htb_sched *q, struct htb_class *cl, long *diff)
489 enum htb_cmode new_mode = htb_class_mode(cl, diff);
491 if (new_mode == cl->cmode)
494 if (cl->prio_activity) { /* not necessary: speed optimization */
495 if (cl->cmode != HTB_CANT_SEND)
496 htb_deactivate_prios(q, cl);
497 cl->cmode = new_mode;
498 if (new_mode != HTB_CANT_SEND)
499 htb_activate_prios(q, cl);
501 cl->cmode = new_mode;
505 * htb_activate - inserts leaf cl into appropriate active feeds
507 * Routine learns (new) priority of leaf and activates feed chain
508 * for the prio. It can be called on already active leaf safely.
509 * It also adds leaf into droplist.
511 static inline void htb_activate(struct htb_sched *q, struct htb_class *cl)
513 WARN_ON(cl->level || !cl->un.leaf.q || !cl->un.leaf.q->q.qlen);
515 if (!cl->prio_activity) {
516 cl->prio_activity = 1 << cl->prio;
517 htb_activate_prios(q, cl);
518 list_add_tail(&cl->un.leaf.drop_list,
519 q->drops + cl->prio);
524 * htb_deactivate - remove leaf cl from active feeds
526 * Make sure that leaf is active. In the other words it can't be called
527 * with non-active leaf. It also removes class from the drop list.
529 static inline void htb_deactivate(struct htb_sched *q, struct htb_class *cl)
531 WARN_ON(!cl->prio_activity);
533 htb_deactivate_prios(q, cl);
534 cl->prio_activity = 0;
535 list_del_init(&cl->un.leaf.drop_list);
538 static int htb_enqueue(struct sk_buff *skb, struct Qdisc *sch)
540 int uninitialized_var(ret);
541 struct htb_sched *q = qdisc_priv(sch);
542 struct htb_class *cl = htb_classify(skb, sch, &ret);
544 if (cl == HTB_DIRECT) {
545 /* enqueue to helper queue */
546 if (q->direct_queue.qlen < q->direct_qlen) {
547 __skb_queue_tail(&q->direct_queue, skb);
552 return NET_XMIT_DROP;
554 #ifdef CONFIG_NET_CLS_ACT
556 if (ret & __NET_XMIT_BYPASS)
561 } else if ((ret = qdisc_enqueue(skb, cl->un.leaf.q)) != NET_XMIT_SUCCESS) {
562 if (net_xmit_drop_count(ret)) {
568 cl->bstats.packets +=
569 skb_is_gso(skb)?skb_shinfo(skb)->gso_segs:1;
570 cl->bstats.bytes += qdisc_pkt_len(skb);
575 sch->bstats.packets += skb_is_gso(skb)?skb_shinfo(skb)->gso_segs:1;
576 sch->bstats.bytes += qdisc_pkt_len(skb);
577 return NET_XMIT_SUCCESS;
580 static inline void htb_accnt_tokens(struct htb_class *cl, int bytes, long diff)
582 long toks = diff + cl->tokens;
584 if (toks > cl->buffer)
586 toks -= (long) qdisc_l2t(cl->rate, bytes);
587 if (toks <= -cl->mbuffer)
588 toks = 1 - cl->mbuffer;
593 static inline void htb_accnt_ctokens(struct htb_class *cl, int bytes, long diff)
595 long toks = diff + cl->ctokens;
597 if (toks > cl->cbuffer)
599 toks -= (long) qdisc_l2t(cl->ceil, bytes);
600 if (toks <= -cl->mbuffer)
601 toks = 1 - cl->mbuffer;
607 * htb_charge_class - charges amount "bytes" to leaf and ancestors
609 * Routine assumes that packet "bytes" long was dequeued from leaf cl
610 * borrowing from "level". It accounts bytes to ceil leaky bucket for
611 * leaf and all ancestors and to rate bucket for ancestors at levels
612 * "level" and higher. It also handles possible change of mode resulting
613 * from the update. Note that mode can also increase here (MAY_BORROW to
614 * CAN_SEND) because we can use more precise clock that event queue here.
615 * In such case we remove class from event queue first.
617 static void htb_charge_class(struct htb_sched *q, struct htb_class *cl,
618 int level, struct sk_buff *skb)
620 int bytes = qdisc_pkt_len(skb);
621 enum htb_cmode old_mode;
625 diff = psched_tdiff_bounded(q->now, cl->t_c, cl->mbuffer);
626 if (cl->level >= level) {
627 if (cl->level == level)
629 htb_accnt_tokens(cl, bytes, diff);
631 cl->xstats.borrows++;
632 cl->tokens += diff; /* we moved t_c; update tokens */
634 htb_accnt_ctokens(cl, bytes, diff);
637 old_mode = cl->cmode;
639 htb_change_class_mode(q, cl, &diff);
640 if (old_mode != cl->cmode) {
641 if (old_mode != HTB_CAN_SEND)
642 htb_safe_rb_erase(&cl->pq_node, q->wait_pq + cl->level);
643 if (cl->cmode != HTB_CAN_SEND)
644 htb_add_to_wait_tree(q, cl, diff);
647 /* update byte stats except for leaves which are already updated */
649 cl->bstats.bytes += bytes;
650 cl->bstats.packets += skb_is_gso(skb)?
651 skb_shinfo(skb)->gso_segs:1;
658 * htb_do_events - make mode changes to classes at the level
660 * Scans event queue for pending events and applies them. Returns time of
661 * next pending event (0 for no event in pq).
662 * Note: Applied are events whose have cl->pq_key <= q->now.
664 static psched_time_t htb_do_events(struct htb_sched *q, int level,
667 /* don't run for longer than 2 jiffies; 2 is used instead of
668 1 to simplify things when jiffy is going to be incremented
670 unsigned long stop_at = start + 2;
671 while (time_before(jiffies, stop_at)) {
672 struct htb_class *cl;
674 struct rb_node *p = rb_first(&q->wait_pq[level]);
679 cl = rb_entry(p, struct htb_class, pq_node);
680 if (cl->pq_key > q->now)
683 htb_safe_rb_erase(p, q->wait_pq + level);
684 diff = psched_tdiff_bounded(q->now, cl->t_c, cl->mbuffer);
685 htb_change_class_mode(q, cl, &diff);
686 if (cl->cmode != HTB_CAN_SEND)
687 htb_add_to_wait_tree(q, cl, diff);
689 /* too much load - let's continue on next jiffie (including above) */
690 return q->now + 2 * PSCHED_TICKS_PER_SEC / HZ;
693 /* Returns class->node+prio from id-tree where classe's id is >= id. NULL
694 is no such one exists. */
695 static struct rb_node *htb_id_find_next_upper(int prio, struct rb_node *n,
698 struct rb_node *r = NULL;
700 struct htb_class *cl =
701 rb_entry(n, struct htb_class, node[prio]);
703 if (id > cl->common.classid) {
705 } else if (id < cl->common.classid) {
716 * htb_lookup_leaf - returns next leaf class in DRR order
718 * Find leaf where current feed pointers points to.
720 static struct htb_class *htb_lookup_leaf(struct rb_root *tree, int prio,
721 struct rb_node **pptr, u32 * pid)
725 struct rb_node *root;
726 struct rb_node **pptr;
728 } stk[TC_HTB_MAXDEPTH], *sp = stk;
730 BUG_ON(!tree->rb_node);
731 sp->root = tree->rb_node;
735 for (i = 0; i < 65535; i++) {
736 if (!*sp->pptr && *sp->pid) {
737 /* ptr was invalidated but id is valid - try to recover
738 the original or next ptr */
740 htb_id_find_next_upper(prio, sp->root, *sp->pid);
742 *sp->pid = 0; /* ptr is valid now so that remove this hint as it
743 can become out of date quickly */
744 if (!*sp->pptr) { /* we are at right end; rewind & go up */
745 *sp->pptr = sp->root;
746 while ((*sp->pptr)->rb_left)
747 *sp->pptr = (*sp->pptr)->rb_left;
754 htb_next_rb_node(sp->pptr);
757 struct htb_class *cl;
758 cl = rb_entry(*sp->pptr, struct htb_class, node[prio]);
761 (++sp)->root = cl->un.inner.feed[prio].rb_node;
762 sp->pptr = cl->un.inner.ptr + prio;
763 sp->pid = cl->un.inner.last_ptr_id + prio;
770 /* dequeues packet at given priority and level; call only if
771 you are sure that there is active class at prio/level */
772 static struct sk_buff *htb_dequeue_tree(struct htb_sched *q, int prio,
775 struct sk_buff *skb = NULL;
776 struct htb_class *cl, *start;
777 /* look initial class up in the row */
778 start = cl = htb_lookup_leaf(q->row[level] + prio, prio,
779 q->ptr[level] + prio,
780 q->last_ptr_id[level] + prio);
787 /* class can be empty - it is unlikely but can be true if leaf
788 qdisc drops packets in enqueue routine or if someone used
789 graft operation on the leaf since last dequeue;
790 simply deactivate and skip such class */
791 if (unlikely(cl->un.leaf.q->q.qlen == 0)) {
792 struct htb_class *next;
793 htb_deactivate(q, cl);
795 /* row/level might become empty */
796 if ((q->row_mask[level] & (1 << prio)) == 0)
799 next = htb_lookup_leaf(q->row[level] + prio,
800 prio, q->ptr[level] + prio,
801 q->last_ptr_id[level] + prio);
803 if (cl == start) /* fix start if we just deleted it */
809 skb = cl->un.leaf.q->dequeue(cl->un.leaf.q);
810 if (likely(skb != NULL))
814 "htb: class %X isn't work conserving ?!\n",
819 htb_next_rb_node((level ? cl->parent->un.inner.ptr : q->
821 cl = htb_lookup_leaf(q->row[level] + prio, prio,
822 q->ptr[level] + prio,
823 q->last_ptr_id[level] + prio);
825 } while (cl != start);
827 if (likely(skb != NULL)) {
828 cl->un.leaf.deficit[level] -= qdisc_pkt_len(skb);
829 if (cl->un.leaf.deficit[level] < 0) {
830 cl->un.leaf.deficit[level] += cl->quantum;
831 htb_next_rb_node((level ? cl->parent->un.inner.ptr : q->
834 /* this used to be after charge_class but this constelation
835 gives us slightly better performance */
836 if (!cl->un.leaf.q->q.qlen)
837 htb_deactivate(q, cl);
838 htb_charge_class(q, cl, level, skb);
843 static struct sk_buff *htb_dequeue(struct Qdisc *sch)
845 struct sk_buff *skb = NULL;
846 struct htb_sched *q = qdisc_priv(sch);
848 psched_time_t next_event;
849 unsigned long start_at;
851 /* try to dequeue direct packets as high prio (!) to minimize cpu work */
852 skb = __skb_dequeue(&q->direct_queue);
854 sch->flags &= ~TCQ_F_THROTTLED;
861 q->now = psched_get_time();
864 next_event = q->now + 5 * PSCHED_TICKS_PER_SEC;
866 for (level = 0; level < TC_HTB_MAXDEPTH; level++) {
867 /* common case optimization - skip event handler quickly */
871 if (q->now >= q->near_ev_cache[level]) {
872 event = htb_do_events(q, level, start_at);
874 event = q->now + PSCHED_TICKS_PER_SEC;
875 q->near_ev_cache[level] = event;
877 event = q->near_ev_cache[level];
879 if (next_event > event)
882 m = ~q->row_mask[level];
883 while (m != (int)(-1)) {
886 skb = htb_dequeue_tree(q, prio, level);
887 if (likely(skb != NULL)) {
889 sch->flags &= ~TCQ_F_THROTTLED;
894 sch->qstats.overlimits++;
895 qdisc_watchdog_schedule(&q->watchdog, next_event);
900 /* try to drop from each class (by prio) until one succeed */
901 static unsigned int htb_drop(struct Qdisc *sch)
903 struct htb_sched *q = qdisc_priv(sch);
906 for (prio = TC_HTB_NUMPRIO - 1; prio >= 0; prio--) {
908 list_for_each(p, q->drops + prio) {
909 struct htb_class *cl = list_entry(p, struct htb_class,
912 if (cl->un.leaf.q->ops->drop &&
913 (len = cl->un.leaf.q->ops->drop(cl->un.leaf.q))) {
915 if (!cl->un.leaf.q->q.qlen)
916 htb_deactivate(q, cl);
924 /* reset all classes */
925 /* always caled under BH & queue lock */
926 static void htb_reset(struct Qdisc *sch)
928 struct htb_sched *q = qdisc_priv(sch);
929 struct htb_class *cl;
930 struct hlist_node *n;
933 for (i = 0; i < q->clhash.hashsize; i++) {
934 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode) {
936 memset(&cl->un.inner, 0, sizeof(cl->un.inner));
939 qdisc_reset(cl->un.leaf.q);
940 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
942 cl->prio_activity = 0;
943 cl->cmode = HTB_CAN_SEND;
947 qdisc_watchdog_cancel(&q->watchdog);
948 __skb_queue_purge(&q->direct_queue);
950 memset(q->row, 0, sizeof(q->row));
951 memset(q->row_mask, 0, sizeof(q->row_mask));
952 memset(q->wait_pq, 0, sizeof(q->wait_pq));
953 memset(q->ptr, 0, sizeof(q->ptr));
954 for (i = 0; i < TC_HTB_NUMPRIO; i++)
955 INIT_LIST_HEAD(q->drops + i);
958 static const struct nla_policy htb_policy[TCA_HTB_MAX + 1] = {
959 [TCA_HTB_PARMS] = { .len = sizeof(struct tc_htb_opt) },
960 [TCA_HTB_INIT] = { .len = sizeof(struct tc_htb_glob) },
961 [TCA_HTB_CTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
962 [TCA_HTB_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
965 static int htb_init(struct Qdisc *sch, struct nlattr *opt)
967 struct htb_sched *q = qdisc_priv(sch);
968 struct nlattr *tb[TCA_HTB_INIT + 1];
969 struct tc_htb_glob *gopt;
976 err = nla_parse_nested(tb, TCA_HTB_INIT, opt, htb_policy);
980 if (tb[TCA_HTB_INIT] == NULL) {
981 printk(KERN_ERR "HTB: hey probably you have bad tc tool ?\n");
984 gopt = nla_data(tb[TCA_HTB_INIT]);
985 if (gopt->version != HTB_VER >> 16) {
987 "HTB: need tc/htb version %d (minor is %d), you have %d\n",
988 HTB_VER >> 16, HTB_VER & 0xffff, gopt->version);
992 err = qdisc_class_hash_init(&q->clhash);
995 for (i = 0; i < TC_HTB_NUMPRIO; i++)
996 INIT_LIST_HEAD(q->drops + i);
998 qdisc_watchdog_init(&q->watchdog, sch);
999 skb_queue_head_init(&q->direct_queue);
1001 q->direct_qlen = qdisc_dev(sch)->tx_queue_len;
1002 if (q->direct_qlen < 2) /* some devices have zero tx_queue_len */
1005 if ((q->rate2quantum = gopt->rate2quantum) < 1)
1006 q->rate2quantum = 1;
1007 q->defcls = gopt->defcls;
1012 static int htb_dump(struct Qdisc *sch, struct sk_buff *skb)
1014 spinlock_t *root_lock = qdisc_root_sleeping_lock(sch);
1015 struct htb_sched *q = qdisc_priv(sch);
1016 struct nlattr *nest;
1017 struct tc_htb_glob gopt;
1019 spin_lock_bh(root_lock);
1021 gopt.direct_pkts = q->direct_pkts;
1022 gopt.version = HTB_VER;
1023 gopt.rate2quantum = q->rate2quantum;
1024 gopt.defcls = q->defcls;
1027 nest = nla_nest_start(skb, TCA_OPTIONS);
1029 goto nla_put_failure;
1030 NLA_PUT(skb, TCA_HTB_INIT, sizeof(gopt), &gopt);
1031 nla_nest_end(skb, nest);
1033 spin_unlock_bh(root_lock);
1037 spin_unlock_bh(root_lock);
1038 nla_nest_cancel(skb, nest);
1042 static int htb_dump_class(struct Qdisc *sch, unsigned long arg,
1043 struct sk_buff *skb, struct tcmsg *tcm)
1045 struct htb_class *cl = (struct htb_class *)arg;
1046 spinlock_t *root_lock = qdisc_root_sleeping_lock(sch);
1047 struct nlattr *nest;
1048 struct tc_htb_opt opt;
1050 spin_lock_bh(root_lock);
1051 tcm->tcm_parent = cl->parent ? cl->parent->common.classid : TC_H_ROOT;
1052 tcm->tcm_handle = cl->common.classid;
1053 if (!cl->level && cl->un.leaf.q)
1054 tcm->tcm_info = cl->un.leaf.q->handle;
1056 nest = nla_nest_start(skb, TCA_OPTIONS);
1058 goto nla_put_failure;
1060 memset(&opt, 0, sizeof(opt));
1062 opt.rate = cl->rate->rate;
1063 opt.buffer = cl->buffer;
1064 opt.ceil = cl->ceil->rate;
1065 opt.cbuffer = cl->cbuffer;
1066 opt.quantum = cl->quantum;
1067 opt.prio = cl->prio;
1068 opt.level = cl->level;
1069 NLA_PUT(skb, TCA_HTB_PARMS, sizeof(opt), &opt);
1071 nla_nest_end(skb, nest);
1072 spin_unlock_bh(root_lock);
1076 spin_unlock_bh(root_lock);
1077 nla_nest_cancel(skb, nest);
1082 htb_dump_class_stats(struct Qdisc *sch, unsigned long arg, struct gnet_dump *d)
1084 struct htb_class *cl = (struct htb_class *)arg;
1086 if (!cl->level && cl->un.leaf.q)
1087 cl->qstats.qlen = cl->un.leaf.q->q.qlen;
1088 cl->xstats.tokens = cl->tokens;
1089 cl->xstats.ctokens = cl->ctokens;
1091 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
1092 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1093 gnet_stats_copy_queue(d, &cl->qstats) < 0)
1096 return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1099 static int htb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1102 struct htb_class *cl = (struct htb_class *)arg;
1104 if (cl && !cl->level) {
1106 (new = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
1108 cl->common.classid))
1112 *old = cl->un.leaf.q;
1113 cl->un.leaf.q = new;
1115 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
1118 sch_tree_unlock(sch);
1124 static struct Qdisc *htb_leaf(struct Qdisc *sch, unsigned long arg)
1126 struct htb_class *cl = (struct htb_class *)arg;
1127 return (cl && !cl->level) ? cl->un.leaf.q : NULL;
1130 static void htb_qlen_notify(struct Qdisc *sch, unsigned long arg)
1132 struct htb_class *cl = (struct htb_class *)arg;
1134 if (cl->un.leaf.q->q.qlen == 0)
1135 htb_deactivate(qdisc_priv(sch), cl);
1138 static unsigned long htb_get(struct Qdisc *sch, u32 classid)
1140 struct htb_class *cl = htb_find(classid, sch);
1143 return (unsigned long)cl;
1146 static inline int htb_parent_last_child(struct htb_class *cl)
1149 /* the root class */
1151 if (cl->parent->children > 1)
1152 /* not the last child */
1157 static void htb_parent_to_leaf(struct htb_sched *q, struct htb_class *cl,
1158 struct Qdisc *new_q)
1160 struct htb_class *parent = cl->parent;
1162 WARN_ON(cl->level || !cl->un.leaf.q || cl->prio_activity);
1164 if (parent->cmode != HTB_CAN_SEND)
1165 htb_safe_rb_erase(&parent->pq_node, q->wait_pq + parent->level);
1168 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1169 INIT_LIST_HEAD(&parent->un.leaf.drop_list);
1170 parent->un.leaf.q = new_q ? new_q : &noop_qdisc;
1171 parent->tokens = parent->buffer;
1172 parent->ctokens = parent->cbuffer;
1173 parent->t_c = psched_get_time();
1174 parent->cmode = HTB_CAN_SEND;
1177 static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
1180 WARN_ON(!cl->un.leaf.q);
1181 qdisc_destroy(cl->un.leaf.q);
1183 gen_kill_estimator(&cl->bstats, &cl->rate_est);
1184 qdisc_put_rtab(cl->rate);
1185 qdisc_put_rtab(cl->ceil);
1187 tcf_destroy_chain(&cl->filter_list);
1191 /* always caled under BH & queue lock */
1192 static void htb_destroy(struct Qdisc *sch)
1194 struct htb_sched *q = qdisc_priv(sch);
1195 struct hlist_node *n, *next;
1196 struct htb_class *cl;
1199 qdisc_watchdog_cancel(&q->watchdog);
1200 /* This line used to be after htb_destroy_class call below
1201 and surprisingly it worked in 2.4. But it must precede it
1202 because filter need its target class alive to be able to call
1203 unbind_filter on it (without Oops). */
1204 tcf_destroy_chain(&q->filter_list);
1206 for (i = 0; i < q->clhash.hashsize; i++) {
1207 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode)
1208 tcf_destroy_chain(&cl->filter_list);
1210 for (i = 0; i < q->clhash.hashsize; i++) {
1211 hlist_for_each_entry_safe(cl, n, next, &q->clhash.hash[i],
1213 htb_destroy_class(sch, cl);
1215 qdisc_class_hash_destroy(&q->clhash);
1216 __skb_queue_purge(&q->direct_queue);
1219 static int htb_delete(struct Qdisc *sch, unsigned long arg)
1221 struct htb_sched *q = qdisc_priv(sch);
1222 struct htb_class *cl = (struct htb_class *)arg;
1224 struct Qdisc *new_q = NULL;
1227 // TODO: why don't allow to delete subtree ? references ? does
1228 // tc subsys quarantee us that in htb_destroy it holds no class
1229 // refs so that we can remove children safely there ?
1230 if (cl->children || cl->filter_cnt)
1233 if (!cl->level && htb_parent_last_child(cl)) {
1234 new_q = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
1236 cl->parent->common.classid);
1243 qlen = cl->un.leaf.q->q.qlen;
1244 qdisc_reset(cl->un.leaf.q);
1245 qdisc_tree_decrease_qlen(cl->un.leaf.q, qlen);
1248 /* delete from hash and active; remainder in destroy_class */
1249 qdisc_class_hash_remove(&q->clhash, &cl->common);
1251 cl->parent->children--;
1253 if (cl->prio_activity)
1254 htb_deactivate(q, cl);
1256 if (cl->cmode != HTB_CAN_SEND)
1257 htb_safe_rb_erase(&cl->pq_node, q->wait_pq + cl->level);
1260 htb_parent_to_leaf(q, cl, new_q);
1262 if (--cl->refcnt == 0)
1263 htb_destroy_class(sch, cl);
1265 sch_tree_unlock(sch);
1269 static void htb_put(struct Qdisc *sch, unsigned long arg)
1271 struct htb_class *cl = (struct htb_class *)arg;
1273 if (--cl->refcnt == 0)
1274 htb_destroy_class(sch, cl);
1277 static int htb_change_class(struct Qdisc *sch, u32 classid,
1278 u32 parentid, struct nlattr **tca,
1282 struct htb_sched *q = qdisc_priv(sch);
1283 struct htb_class *cl = (struct htb_class *)*arg, *parent;
1284 struct nlattr *opt = tca[TCA_OPTIONS];
1285 struct qdisc_rate_table *rtab = NULL, *ctab = NULL;
1286 struct nlattr *tb[TCA_HTB_RTAB + 1];
1287 struct tc_htb_opt *hopt;
1289 /* extract all subattrs from opt attr */
1293 err = nla_parse_nested(tb, TCA_HTB_RTAB, opt, htb_policy);
1298 if (tb[TCA_HTB_PARMS] == NULL)
1301 parent = parentid == TC_H_ROOT ? NULL : htb_find(parentid, sch);
1303 hopt = nla_data(tb[TCA_HTB_PARMS]);
1305 rtab = qdisc_get_rtab(&hopt->rate, tb[TCA_HTB_RTAB]);
1306 ctab = qdisc_get_rtab(&hopt->ceil, tb[TCA_HTB_CTAB]);
1310 if (!cl) { /* new class */
1311 struct Qdisc *new_q;
1315 struct gnet_estimator opt;
1318 .nla_len = nla_attr_size(sizeof(est.opt)),
1319 .nla_type = TCA_RATE,
1322 /* 4s interval, 16s averaging constant */
1328 /* check for valid classid */
1329 if (!classid || TC_H_MAJ(classid ^ sch->handle)
1330 || htb_find(classid, sch))
1333 /* check maximal depth */
1334 if (parent && parent->parent && parent->parent->level < 2) {
1335 printk(KERN_ERR "htb: tree is too deep\n");
1339 if ((cl = kzalloc(sizeof(*cl), GFP_KERNEL)) == NULL)
1342 err = gen_new_estimator(&cl->bstats, &cl->rate_est,
1343 qdisc_root_sleeping_lock(sch),
1344 tca[TCA_RATE] ? : &est.nla);
1352 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
1353 RB_CLEAR_NODE(&cl->pq_node);
1355 for (prio = 0; prio < TC_HTB_NUMPRIO; prio++)
1356 RB_CLEAR_NODE(&cl->node[prio]);
1358 /* create leaf qdisc early because it uses kmalloc(GFP_KERNEL)
1359 so that can't be used inside of sch_tree_lock
1360 -- thanks to Karlis Peisenieks */
1361 new_q = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
1362 &pfifo_qdisc_ops, classid);
1364 if (parent && !parent->level) {
1365 unsigned int qlen = parent->un.leaf.q->q.qlen;
1367 /* turn parent into inner node */
1368 qdisc_reset(parent->un.leaf.q);
1369 qdisc_tree_decrease_qlen(parent->un.leaf.q, qlen);
1370 qdisc_destroy(parent->un.leaf.q);
1371 if (parent->prio_activity)
1372 htb_deactivate(q, parent);
1374 /* remove from evt list because of level change */
1375 if (parent->cmode != HTB_CAN_SEND) {
1376 htb_safe_rb_erase(&parent->pq_node, q->wait_pq);
1377 parent->cmode = HTB_CAN_SEND;
1379 parent->level = (parent->parent ? parent->parent->level
1380 : TC_HTB_MAXDEPTH) - 1;
1381 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1383 /* leaf (we) needs elementary qdisc */
1384 cl->un.leaf.q = new_q ? new_q : &noop_qdisc;
1386 cl->common.classid = classid;
1387 cl->parent = parent;
1389 /* set class to be in HTB_CAN_SEND state */
1390 cl->tokens = hopt->buffer;
1391 cl->ctokens = hopt->cbuffer;
1392 cl->mbuffer = 60 * PSCHED_TICKS_PER_SEC; /* 1min */
1393 cl->t_c = psched_get_time();
1394 cl->cmode = HTB_CAN_SEND;
1396 /* attach to the hash list and parent's family */
1397 qdisc_class_hash_insert(&q->clhash, &cl->common);
1401 if (tca[TCA_RATE]) {
1402 err = gen_replace_estimator(&cl->bstats, &cl->rate_est,
1403 qdisc_root_sleeping_lock(sch),
1411 /* it used to be a nasty bug here, we have to check that node
1412 is really leaf before changing cl->un.leaf ! */
1414 cl->quantum = rtab->rate.rate / q->rate2quantum;
1415 if (!hopt->quantum && cl->quantum < 1000) {
1417 "HTB: quantum of class %X is small. Consider r2q change.\n",
1418 cl->common.classid);
1421 if (!hopt->quantum && cl->quantum > 200000) {
1423 "HTB: quantum of class %X is big. Consider r2q change.\n",
1424 cl->common.classid);
1425 cl->quantum = 200000;
1428 cl->quantum = hopt->quantum;
1429 if ((cl->prio = hopt->prio) >= TC_HTB_NUMPRIO)
1430 cl->prio = TC_HTB_NUMPRIO - 1;
1433 cl->buffer = hopt->buffer;
1434 cl->cbuffer = hopt->cbuffer;
1436 qdisc_put_rtab(cl->rate);
1439 qdisc_put_rtab(cl->ceil);
1441 sch_tree_unlock(sch);
1443 qdisc_class_hash_grow(sch, &q->clhash);
1445 *arg = (unsigned long)cl;
1450 qdisc_put_rtab(rtab);
1452 qdisc_put_rtab(ctab);
1456 static struct tcf_proto **htb_find_tcf(struct Qdisc *sch, unsigned long arg)
1458 struct htb_sched *q = qdisc_priv(sch);
1459 struct htb_class *cl = (struct htb_class *)arg;
1460 struct tcf_proto **fl = cl ? &cl->filter_list : &q->filter_list;
1465 static unsigned long htb_bind_filter(struct Qdisc *sch, unsigned long parent,
1468 struct htb_class *cl = htb_find(classid, sch);
1470 /*if (cl && !cl->level) return 0;
1471 The line above used to be there to prevent attaching filters to
1472 leaves. But at least tc_index filter uses this just to get class
1473 for other reasons so that we have to allow for it.
1475 19.6.2002 As Werner explained it is ok - bind filter is just
1476 another way to "lock" the class - unlike "get" this lock can
1477 be broken by class during destroy IIUC.
1481 return (unsigned long)cl;
1484 static void htb_unbind_filter(struct Qdisc *sch, unsigned long arg)
1486 struct htb_class *cl = (struct htb_class *)arg;
1492 static void htb_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1494 struct htb_sched *q = qdisc_priv(sch);
1495 struct htb_class *cl;
1496 struct hlist_node *n;
1502 for (i = 0; i < q->clhash.hashsize; i++) {
1503 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode) {
1504 if (arg->count < arg->skip) {
1508 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1517 static const struct Qdisc_class_ops htb_class_ops = {
1520 .qlen_notify = htb_qlen_notify,
1523 .change = htb_change_class,
1524 .delete = htb_delete,
1526 .tcf_chain = htb_find_tcf,
1527 .bind_tcf = htb_bind_filter,
1528 .unbind_tcf = htb_unbind_filter,
1529 .dump = htb_dump_class,
1530 .dump_stats = htb_dump_class_stats,
1533 static struct Qdisc_ops htb_qdisc_ops __read_mostly = {
1535 .cl_ops = &htb_class_ops,
1537 .priv_size = sizeof(struct htb_sched),
1538 .enqueue = htb_enqueue,
1539 .dequeue = htb_dequeue,
1540 .peek = qdisc_peek_dequeued,
1544 .destroy = htb_destroy,
1545 .change = NULL /* htb_change */,
1547 .owner = THIS_MODULE,
1550 static int __init htb_module_init(void)
1552 return register_qdisc(&htb_qdisc_ops);
1554 static void __exit htb_module_exit(void)
1556 unregister_qdisc(&htb_qdisc_ops);
1559 module_init(htb_module_init)
1560 module_exit(htb_module_exit)
1561 MODULE_LICENSE("GPL");