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 {
92 int deficit[TC_HTB_MAXDEPTH];
93 struct list_head drop_list;
95 struct htb_class_inner {
96 struct rb_root feed[TC_HTB_NUMPRIO]; /* feed trees */
97 struct rb_node *ptr[TC_HTB_NUMPRIO]; /* current class ptr */
98 /* When class changes from state 1->2 and disconnects from
99 parent's feed then we lost ptr value and start from the
100 first child again. Here we store classid of the
101 last valid ptr (used when ptr is NULL). */
102 u32 last_ptr_id[TC_HTB_NUMPRIO];
105 struct rb_node node[TC_HTB_NUMPRIO]; /* node for self or feed tree */
106 struct rb_node pq_node; /* node for event queue */
107 psched_time_t pq_key;
109 int prio_activity; /* for which prios are we active */
110 enum htb_cmode cmode; /* current mode of the class */
112 /* class attached filters */
113 struct tcf_proto *filter_list;
116 int warned; /* only one warning about non work conserving .. */
118 /* token bucket parameters */
119 struct qdisc_rate_table *rate; /* rate table of the class itself */
120 struct qdisc_rate_table *ceil; /* ceiling rate (limits borrows too) */
121 long buffer, cbuffer; /* token bucket depth/rate */
122 psched_tdiff_t mbuffer; /* max wait time */
123 long tokens, ctokens; /* current number of tokens */
124 psched_time_t t_c; /* checkpoint time */
126 int prio; /* For parent to leaf return possible here */
127 int quantum; /* we do backup. Finally full replacement */
128 /* of un.leaf originals should be done. */
131 static inline long L2T(struct htb_class *cl, struct qdisc_rate_table *rate,
134 long result = qdisc_l2t(rate, size);
139 struct Qdisc_class_hash clhash;
140 struct list_head drops[TC_HTB_NUMPRIO];/* active leaves (for drops) */
142 /* self list - roots of self generating tree */
143 struct rb_root row[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
144 int row_mask[TC_HTB_MAXDEPTH];
145 struct rb_node *ptr[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
146 u32 last_ptr_id[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
148 /* self wait list - roots of wait PQs per row */
149 struct rb_root wait_pq[TC_HTB_MAXDEPTH];
151 /* time of nearest event per level (row) */
152 psched_time_t near_ev_cache[TC_HTB_MAXDEPTH];
154 int defcls; /* class where unclassified flows go to */
156 /* filters for qdisc itself */
157 struct tcf_proto *filter_list;
159 int rate2quantum; /* quant = rate / rate2quantum */
160 psched_time_t now; /* cached dequeue time */
161 struct qdisc_watchdog watchdog;
163 /* non shaped skbs; let them go directly thru */
164 struct sk_buff_head direct_queue;
165 int direct_qlen; /* max qlen of above */
170 /* find class in global hash table using given handle */
171 static inline struct htb_class *htb_find(u32 handle, struct Qdisc *sch)
173 struct htb_sched *q = qdisc_priv(sch);
174 struct Qdisc_class_common *clc;
176 clc = qdisc_class_find(&q->clhash, handle);
179 return container_of(clc, struct htb_class, common);
183 * htb_classify - classify a packet into class
185 * It returns NULL if the packet should be dropped or -1 if the packet
186 * should be passed directly thru. In all other cases leaf class is returned.
187 * We allow direct class selection by classid in priority. The we examine
188 * filters in qdisc and in inner nodes (if higher filter points to the inner
189 * node). If we end up with classid MAJOR:0 we enqueue the skb into special
190 * internal fifo (direct). These packets then go directly thru. If we still
191 * have no valid leaf we try to use MAJOR:default leaf. It still unsuccessfull
192 * then finish and return direct queue.
194 #define HTB_DIRECT (struct htb_class*)-1
196 static struct htb_class *htb_classify(struct sk_buff *skb, struct Qdisc *sch,
199 struct htb_sched *q = qdisc_priv(sch);
200 struct htb_class *cl;
201 struct tcf_result res;
202 struct tcf_proto *tcf;
205 /* allow to select class by setting skb->priority to valid classid;
206 note that nfmark can be used too by attaching filter fw with no
208 if (skb->priority == sch->handle)
209 return HTB_DIRECT; /* X:0 (direct flow) selected */
210 if ((cl = htb_find(skb->priority, sch)) != NULL && cl->level == 0)
213 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
214 tcf = q->filter_list;
215 while (tcf && (result = tc_classify(skb, tcf, &res)) >= 0) {
216 #ifdef CONFIG_NET_CLS_ACT
220 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
225 if ((cl = (void *)res.class) == NULL) {
226 if (res.classid == sch->handle)
227 return HTB_DIRECT; /* X:0 (direct flow) */
228 if ((cl = htb_find(res.classid, sch)) == NULL)
229 break; /* filter selected invalid classid */
232 return cl; /* we hit leaf; return it */
234 /* we have got inner class; apply inner filter chain */
235 tcf = cl->filter_list;
237 /* classification failed; try to use default class */
238 cl = htb_find(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
239 if (!cl || cl->level)
240 return HTB_DIRECT; /* bad default .. this is safe bet */
245 * htb_add_to_id_tree - adds class to the round robin list
247 * Routine adds class to the list (actually tree) sorted by classid.
248 * Make sure that class is not already on such list for given prio.
250 static void htb_add_to_id_tree(struct rb_root *root,
251 struct htb_class *cl, int prio)
253 struct rb_node **p = &root->rb_node, *parent = NULL;
258 c = rb_entry(parent, struct htb_class, node[prio]);
260 if (cl->common.classid > c->common.classid)
261 p = &parent->rb_right;
263 p = &parent->rb_left;
265 rb_link_node(&cl->node[prio], parent, p);
266 rb_insert_color(&cl->node[prio], root);
270 * htb_add_to_wait_tree - adds class to the event queue with delay
272 * The class is added to priority event queue to indicate that class will
273 * change its mode in cl->pq_key microseconds. Make sure that class is not
274 * already in the queue.
276 static void htb_add_to_wait_tree(struct htb_sched *q,
277 struct htb_class *cl, long delay)
279 struct rb_node **p = &q->wait_pq[cl->level].rb_node, *parent = NULL;
281 cl->pq_key = q->now + delay;
282 if (cl->pq_key == q->now)
285 /* update the nearest event cache */
286 if (q->near_ev_cache[cl->level] > cl->pq_key)
287 q->near_ev_cache[cl->level] = cl->pq_key;
292 c = rb_entry(parent, struct htb_class, pq_node);
293 if (cl->pq_key >= c->pq_key)
294 p = &parent->rb_right;
296 p = &parent->rb_left;
298 rb_link_node(&cl->pq_node, parent, p);
299 rb_insert_color(&cl->pq_node, &q->wait_pq[cl->level]);
303 * htb_next_rb_node - finds next node in binary tree
305 * When we are past last key we return NULL.
306 * Average complexity is 2 steps per call.
308 static inline void htb_next_rb_node(struct rb_node **n)
314 * htb_add_class_to_row - add class to its row
316 * The class is added to row at priorities marked in mask.
317 * It does nothing if mask == 0.
319 static inline void htb_add_class_to_row(struct htb_sched *q,
320 struct htb_class *cl, int mask)
322 q->row_mask[cl->level] |= mask;
324 int prio = ffz(~mask);
325 mask &= ~(1 << prio);
326 htb_add_to_id_tree(q->row[cl->level] + prio, cl, prio);
330 /* If this triggers, it is a bug in this code, but it need not be fatal */
331 static void htb_safe_rb_erase(struct rb_node *rb, struct rb_root *root)
333 if (RB_EMPTY_NODE(rb)) {
343 * htb_remove_class_from_row - removes class from its row
345 * The class is removed from row at priorities marked in mask.
346 * It does nothing if mask == 0.
348 static inline void htb_remove_class_from_row(struct htb_sched *q,
349 struct htb_class *cl, int mask)
354 int prio = ffz(~mask);
356 mask &= ~(1 << prio);
357 if (q->ptr[cl->level][prio] == cl->node + prio)
358 htb_next_rb_node(q->ptr[cl->level] + prio);
360 htb_safe_rb_erase(cl->node + prio, q->row[cl->level] + prio);
361 if (!q->row[cl->level][prio].rb_node)
364 q->row_mask[cl->level] &= ~m;
368 * htb_activate_prios - creates active classe's feed chain
370 * The class is connected to ancestors and/or appropriate rows
371 * for priorities it is participating on. cl->cmode must be new
372 * (activated) mode. It does nothing if cl->prio_activity == 0.
374 static void htb_activate_prios(struct htb_sched *q, struct htb_class *cl)
376 struct htb_class *p = cl->parent;
377 long m, mask = cl->prio_activity;
379 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
385 if (p->un.inner.feed[prio].rb_node)
386 /* parent already has its feed in use so that
387 reset bit in mask as parent is already ok */
388 mask &= ~(1 << prio);
390 htb_add_to_id_tree(p->un.inner.feed + prio, cl, prio);
392 p->prio_activity |= mask;
397 if (cl->cmode == HTB_CAN_SEND && mask)
398 htb_add_class_to_row(q, cl, mask);
402 * htb_deactivate_prios - remove class from feed chain
404 * cl->cmode must represent old mode (before deactivation). It does
405 * nothing if cl->prio_activity == 0. Class is removed from all feed
408 static void htb_deactivate_prios(struct htb_sched *q, struct htb_class *cl)
410 struct htb_class *p = cl->parent;
411 long m, mask = cl->prio_activity;
413 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
420 if (p->un.inner.ptr[prio] == cl->node + prio) {
421 /* we are removing child which is pointed to from
422 parent feed - forget the pointer but remember
424 p->un.inner.last_ptr_id[prio] = cl->common.classid;
425 p->un.inner.ptr[prio] = NULL;
428 htb_safe_rb_erase(cl->node + prio, p->un.inner.feed + prio);
430 if (!p->un.inner.feed[prio].rb_node)
434 p->prio_activity &= ~mask;
439 if (cl->cmode == HTB_CAN_SEND && mask)
440 htb_remove_class_from_row(q, cl, mask);
443 static inline long htb_lowater(const struct htb_class *cl)
446 return cl->cmode != HTB_CANT_SEND ? -cl->cbuffer : 0;
450 static inline long htb_hiwater(const struct htb_class *cl)
453 return cl->cmode == HTB_CAN_SEND ? -cl->buffer : 0;
460 * htb_class_mode - computes and returns current class mode
462 * It computes cl's mode at time cl->t_c+diff and returns it. If mode
463 * is not HTB_CAN_SEND then cl->pq_key is updated to time difference
464 * from now to time when cl will change its state.
465 * Also it is worth to note that class mode doesn't change simply
466 * at cl->{c,}tokens == 0 but there can rather be hysteresis of
467 * 0 .. -cl->{c,}buffer range. It is meant to limit number of
468 * mode transitions per time unit. The speed gain is about 1/6.
470 static inline enum htb_cmode
471 htb_class_mode(struct htb_class *cl, long *diff)
475 if ((toks = (cl->ctokens + *diff)) < htb_lowater(cl)) {
477 return HTB_CANT_SEND;
480 if ((toks = (cl->tokens + *diff)) >= htb_hiwater(cl))
484 return HTB_MAY_BORROW;
488 * htb_change_class_mode - changes classe's mode
490 * This should be the only way how to change classe's mode under normal
491 * cirsumstances. Routine will update feed lists linkage, change mode
492 * and add class to the wait event queue if appropriate. New mode should
493 * be different from old one and cl->pq_key has to be valid if changing
494 * to mode other than HTB_CAN_SEND (see htb_add_to_wait_tree).
497 htb_change_class_mode(struct htb_sched *q, struct htb_class *cl, long *diff)
499 enum htb_cmode new_mode = htb_class_mode(cl, diff);
501 if (new_mode == cl->cmode)
504 if (cl->prio_activity) { /* not necessary: speed optimization */
505 if (cl->cmode != HTB_CANT_SEND)
506 htb_deactivate_prios(q, cl);
507 cl->cmode = new_mode;
508 if (new_mode != HTB_CANT_SEND)
509 htb_activate_prios(q, cl);
511 cl->cmode = new_mode;
515 * htb_activate - inserts leaf cl into appropriate active feeds
517 * Routine learns (new) priority of leaf and activates feed chain
518 * for the prio. It can be called on already active leaf safely.
519 * It also adds leaf into droplist.
521 static inline void htb_activate(struct htb_sched *q, struct htb_class *cl)
523 WARN_ON(cl->level || !cl->un.leaf.q || !cl->un.leaf.q->q.qlen);
525 if (!cl->prio_activity) {
526 cl->prio_activity = 1 << cl->un.leaf.prio;
527 htb_activate_prios(q, cl);
528 list_add_tail(&cl->un.leaf.drop_list,
529 q->drops + cl->un.leaf.prio);
534 * htb_deactivate - remove leaf cl from active feeds
536 * Make sure that leaf is active. In the other words it can't be called
537 * with non-active leaf. It also removes class from the drop list.
539 static inline void htb_deactivate(struct htb_sched *q, struct htb_class *cl)
541 WARN_ON(!cl->prio_activity);
543 htb_deactivate_prios(q, cl);
544 cl->prio_activity = 0;
545 list_del_init(&cl->un.leaf.drop_list);
548 static int htb_enqueue(struct sk_buff *skb, struct Qdisc *sch)
550 int uninitialized_var(ret);
551 struct htb_sched *q = qdisc_priv(sch);
552 struct htb_class *cl = htb_classify(skb, sch, &ret);
554 if (cl == HTB_DIRECT) {
555 /* enqueue to helper queue */
556 if (q->direct_queue.qlen < q->direct_qlen) {
557 __skb_queue_tail(&q->direct_queue, skb);
562 return NET_XMIT_DROP;
564 #ifdef CONFIG_NET_CLS_ACT
566 if (ret & __NET_XMIT_BYPASS)
571 } else if ((ret = qdisc_enqueue(skb, cl->un.leaf.q)) != NET_XMIT_SUCCESS) {
572 if (net_xmit_drop_count(ret)) {
578 cl->bstats.packets +=
579 skb_is_gso(skb)?skb_shinfo(skb)->gso_segs:1;
580 cl->bstats.bytes += qdisc_pkt_len(skb);
585 sch->bstats.packets += skb_is_gso(skb)?skb_shinfo(skb)->gso_segs:1;
586 sch->bstats.bytes += qdisc_pkt_len(skb);
587 return NET_XMIT_SUCCESS;
591 * htb_charge_class - charges amount "bytes" to leaf and ancestors
593 * Routine assumes that packet "bytes" long was dequeued from leaf cl
594 * borrowing from "level". It accounts bytes to ceil leaky bucket for
595 * leaf and all ancestors and to rate bucket for ancestors at levels
596 * "level" and higher. It also handles possible change of mode resulting
597 * from the update. Note that mode can also increase here (MAY_BORROW to
598 * CAN_SEND) because we can use more precise clock that event queue here.
599 * In such case we remove class from event queue first.
601 static void htb_charge_class(struct htb_sched *q, struct htb_class *cl,
602 int level, struct sk_buff *skb)
604 int bytes = qdisc_pkt_len(skb);
606 enum htb_cmode old_mode;
608 #define HTB_ACCNT(T,B,R) toks = diff + cl->T; \
609 if (toks > cl->B) toks = cl->B; \
610 toks -= L2T(cl, cl->R, bytes); \
611 if (toks <= -cl->mbuffer) toks = 1-cl->mbuffer; \
615 diff = psched_tdiff_bounded(q->now, cl->t_c, cl->mbuffer);
616 if (cl->level >= level) {
617 if (cl->level == level)
619 HTB_ACCNT(tokens, buffer, rate);
621 cl->xstats.borrows++;
622 cl->tokens += diff; /* we moved t_c; update tokens */
624 HTB_ACCNT(ctokens, cbuffer, ceil);
627 old_mode = cl->cmode;
629 htb_change_class_mode(q, cl, &diff);
630 if (old_mode != cl->cmode) {
631 if (old_mode != HTB_CAN_SEND)
632 htb_safe_rb_erase(&cl->pq_node, q->wait_pq + cl->level);
633 if (cl->cmode != HTB_CAN_SEND)
634 htb_add_to_wait_tree(q, cl, diff);
637 /* update byte stats except for leaves which are already updated */
639 cl->bstats.bytes += bytes;
640 cl->bstats.packets += skb_is_gso(skb)?
641 skb_shinfo(skb)->gso_segs:1;
648 * htb_do_events - make mode changes to classes at the level
650 * Scans event queue for pending events and applies them. Returns time of
651 * next pending event (0 for no event in pq).
652 * Note: Applied are events whose have cl->pq_key <= q->now.
654 static psched_time_t htb_do_events(struct htb_sched *q, int level)
656 /* don't run for longer than 2 jiffies; 2 is used instead of
657 1 to simplify things when jiffy is going to be incremented
659 unsigned long stop_at = jiffies + 2;
660 while (time_before(jiffies, stop_at)) {
661 struct htb_class *cl;
663 struct rb_node *p = rb_first(&q->wait_pq[level]);
668 cl = rb_entry(p, struct htb_class, pq_node);
669 if (cl->pq_key > q->now)
672 htb_safe_rb_erase(p, q->wait_pq + level);
673 diff = psched_tdiff_bounded(q->now, cl->t_c, cl->mbuffer);
674 htb_change_class_mode(q, cl, &diff);
675 if (cl->cmode != HTB_CAN_SEND)
676 htb_add_to_wait_tree(q, cl, diff);
678 /* too much load - let's continue on next jiffie */
679 return q->now + PSCHED_TICKS_PER_SEC / HZ;
682 /* Returns class->node+prio from id-tree where classe's id is >= id. NULL
683 is no such one exists. */
684 static struct rb_node *htb_id_find_next_upper(int prio, struct rb_node *n,
687 struct rb_node *r = NULL;
689 struct htb_class *cl =
690 rb_entry(n, struct htb_class, node[prio]);
691 if (id == cl->common.classid)
694 if (id > cl->common.classid) {
705 * htb_lookup_leaf - returns next leaf class in DRR order
707 * Find leaf where current feed pointers points to.
709 static struct htb_class *htb_lookup_leaf(struct rb_root *tree, int prio,
710 struct rb_node **pptr, u32 * pid)
714 struct rb_node *root;
715 struct rb_node **pptr;
717 } stk[TC_HTB_MAXDEPTH], *sp = stk;
719 WARN_ON(!tree->rb_node);
720 sp->root = tree->rb_node;
724 for (i = 0; i < 65535; i++) {
725 if (!*sp->pptr && *sp->pid) {
726 /* ptr was invalidated but id is valid - try to recover
727 the original or next ptr */
729 htb_id_find_next_upper(prio, sp->root, *sp->pid);
731 *sp->pid = 0; /* ptr is valid now so that remove this hint as it
732 can become out of date quickly */
733 if (!*sp->pptr) { /* we are at right end; rewind & go up */
734 *sp->pptr = sp->root;
735 while ((*sp->pptr)->rb_left)
736 *sp->pptr = (*sp->pptr)->rb_left;
742 htb_next_rb_node(sp->pptr);
745 struct htb_class *cl;
746 cl = rb_entry(*sp->pptr, struct htb_class, node[prio]);
749 (++sp)->root = cl->un.inner.feed[prio].rb_node;
750 sp->pptr = cl->un.inner.ptr + prio;
751 sp->pid = cl->un.inner.last_ptr_id + prio;
758 /* dequeues packet at given priority and level; call only if
759 you are sure that there is active class at prio/level */
760 static struct sk_buff *htb_dequeue_tree(struct htb_sched *q, int prio,
763 struct sk_buff *skb = NULL;
764 struct htb_class *cl, *start;
765 /* look initial class up in the row */
766 start = cl = htb_lookup_leaf(q->row[level] + prio, prio,
767 q->ptr[level] + prio,
768 q->last_ptr_id[level] + prio);
776 /* class can be empty - it is unlikely but can be true if leaf
777 qdisc drops packets in enqueue routine or if someone used
778 graft operation on the leaf since last dequeue;
779 simply deactivate and skip such class */
780 if (unlikely(cl->un.leaf.q->q.qlen == 0)) {
781 struct htb_class *next;
782 htb_deactivate(q, cl);
784 /* row/level might become empty */
785 if ((q->row_mask[level] & (1 << prio)) == 0)
788 next = htb_lookup_leaf(q->row[level] + prio,
789 prio, q->ptr[level] + prio,
790 q->last_ptr_id[level] + prio);
792 if (cl == start) /* fix start if we just deleted it */
798 skb = cl->un.leaf.q->dequeue(cl->un.leaf.q);
799 if (likely(skb != NULL))
803 "htb: class %X isn't work conserving ?!\n",
808 htb_next_rb_node((level ? cl->parent->un.inner.ptr : q->
810 cl = htb_lookup_leaf(q->row[level] + prio, prio,
811 q->ptr[level] + prio,
812 q->last_ptr_id[level] + prio);
814 } while (cl != start);
816 if (likely(skb != NULL)) {
817 cl->un.leaf.deficit[level] -= qdisc_pkt_len(skb);
818 if (cl->un.leaf.deficit[level] < 0) {
819 cl->un.leaf.deficit[level] += cl->un.leaf.quantum;
820 htb_next_rb_node((level ? cl->parent->un.inner.ptr : q->
823 /* this used to be after charge_class but this constelation
824 gives us slightly better performance */
825 if (!cl->un.leaf.q->q.qlen)
826 htb_deactivate(q, cl);
827 htb_charge_class(q, cl, level, skb);
832 static struct sk_buff *htb_dequeue(struct Qdisc *sch)
834 struct sk_buff *skb = NULL;
835 struct htb_sched *q = qdisc_priv(sch);
837 psched_time_t next_event;
839 /* try to dequeue direct packets as high prio (!) to minimize cpu work */
840 skb = __skb_dequeue(&q->direct_queue);
842 sch->flags &= ~TCQ_F_THROTTLED;
849 q->now = psched_get_time();
851 next_event = q->now + 5 * PSCHED_TICKS_PER_SEC;
853 for (level = 0; level < TC_HTB_MAXDEPTH; level++) {
854 /* common case optimization - skip event handler quickly */
858 if (q->now >= q->near_ev_cache[level]) {
859 event = htb_do_events(q, level);
861 event = q->now + PSCHED_TICKS_PER_SEC;
862 q->near_ev_cache[level] = event;
864 event = q->near_ev_cache[level];
866 if (event && next_event > event)
869 m = ~q->row_mask[level];
870 while (m != (int)(-1)) {
873 skb = htb_dequeue_tree(q, prio, level);
874 if (likely(skb != NULL)) {
876 sch->flags &= ~TCQ_F_THROTTLED;
881 sch->qstats.overlimits++;
882 qdisc_watchdog_schedule(&q->watchdog, next_event);
887 /* try to drop from each class (by prio) until one succeed */
888 static unsigned int htb_drop(struct Qdisc *sch)
890 struct htb_sched *q = qdisc_priv(sch);
893 for (prio = TC_HTB_NUMPRIO - 1; prio >= 0; prio--) {
895 list_for_each(p, q->drops + prio) {
896 struct htb_class *cl = list_entry(p, struct htb_class,
899 if (cl->un.leaf.q->ops->drop &&
900 (len = cl->un.leaf.q->ops->drop(cl->un.leaf.q))) {
902 if (!cl->un.leaf.q->q.qlen)
903 htb_deactivate(q, cl);
911 /* reset all classes */
912 /* always caled under BH & queue lock */
913 static void htb_reset(struct Qdisc *sch)
915 struct htb_sched *q = qdisc_priv(sch);
916 struct htb_class *cl;
917 struct hlist_node *n;
920 for (i = 0; i < q->clhash.hashsize; i++) {
921 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode) {
923 memset(&cl->un.inner, 0, sizeof(cl->un.inner));
926 qdisc_reset(cl->un.leaf.q);
927 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
929 cl->prio_activity = 0;
930 cl->cmode = HTB_CAN_SEND;
934 qdisc_watchdog_cancel(&q->watchdog);
935 __skb_queue_purge(&q->direct_queue);
937 memset(q->row, 0, sizeof(q->row));
938 memset(q->row_mask, 0, sizeof(q->row_mask));
939 memset(q->wait_pq, 0, sizeof(q->wait_pq));
940 memset(q->ptr, 0, sizeof(q->ptr));
941 for (i = 0; i < TC_HTB_NUMPRIO; i++)
942 INIT_LIST_HEAD(q->drops + i);
945 static const struct nla_policy htb_policy[TCA_HTB_MAX + 1] = {
946 [TCA_HTB_PARMS] = { .len = sizeof(struct tc_htb_opt) },
947 [TCA_HTB_INIT] = { .len = sizeof(struct tc_htb_glob) },
948 [TCA_HTB_CTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
949 [TCA_HTB_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
952 static int htb_init(struct Qdisc *sch, struct nlattr *opt)
954 struct htb_sched *q = qdisc_priv(sch);
955 struct nlattr *tb[TCA_HTB_INIT + 1];
956 struct tc_htb_glob *gopt;
963 err = nla_parse_nested(tb, TCA_HTB_INIT, opt, htb_policy);
967 if (tb[TCA_HTB_INIT] == NULL) {
968 printk(KERN_ERR "HTB: hey probably you have bad tc tool ?\n");
971 gopt = nla_data(tb[TCA_HTB_INIT]);
972 if (gopt->version != HTB_VER >> 16) {
974 "HTB: need tc/htb version %d (minor is %d), you have %d\n",
975 HTB_VER >> 16, HTB_VER & 0xffff, gopt->version);
979 err = qdisc_class_hash_init(&q->clhash);
982 for (i = 0; i < TC_HTB_NUMPRIO; i++)
983 INIT_LIST_HEAD(q->drops + i);
985 qdisc_watchdog_init(&q->watchdog, sch);
986 skb_queue_head_init(&q->direct_queue);
988 q->direct_qlen = qdisc_dev(sch)->tx_queue_len;
989 if (q->direct_qlen < 2) /* some devices have zero tx_queue_len */
992 if ((q->rate2quantum = gopt->rate2quantum) < 1)
994 q->defcls = gopt->defcls;
999 static int htb_dump(struct Qdisc *sch, struct sk_buff *skb)
1001 spinlock_t *root_lock = qdisc_root_sleeping_lock(sch);
1002 struct htb_sched *q = qdisc_priv(sch);
1003 struct nlattr *nest;
1004 struct tc_htb_glob gopt;
1006 spin_lock_bh(root_lock);
1008 gopt.direct_pkts = q->direct_pkts;
1009 gopt.version = HTB_VER;
1010 gopt.rate2quantum = q->rate2quantum;
1011 gopt.defcls = q->defcls;
1014 nest = nla_nest_start(skb, TCA_OPTIONS);
1016 goto nla_put_failure;
1017 NLA_PUT(skb, TCA_HTB_INIT, sizeof(gopt), &gopt);
1018 nla_nest_end(skb, nest);
1020 spin_unlock_bh(root_lock);
1024 spin_unlock_bh(root_lock);
1025 nla_nest_cancel(skb, nest);
1029 static int htb_dump_class(struct Qdisc *sch, unsigned long arg,
1030 struct sk_buff *skb, struct tcmsg *tcm)
1032 struct htb_class *cl = (struct htb_class *)arg;
1033 spinlock_t *root_lock = qdisc_root_sleeping_lock(sch);
1034 struct nlattr *nest;
1035 struct tc_htb_opt opt;
1037 spin_lock_bh(root_lock);
1038 tcm->tcm_parent = cl->parent ? cl->parent->common.classid : TC_H_ROOT;
1039 tcm->tcm_handle = cl->common.classid;
1040 if (!cl->level && cl->un.leaf.q)
1041 tcm->tcm_info = cl->un.leaf.q->handle;
1043 nest = nla_nest_start(skb, TCA_OPTIONS);
1045 goto nla_put_failure;
1047 memset(&opt, 0, sizeof(opt));
1049 opt.rate = cl->rate->rate;
1050 opt.buffer = cl->buffer;
1051 opt.ceil = cl->ceil->rate;
1052 opt.cbuffer = cl->cbuffer;
1053 opt.quantum = cl->un.leaf.quantum;
1054 opt.prio = cl->un.leaf.prio;
1055 opt.level = cl->level;
1056 NLA_PUT(skb, TCA_HTB_PARMS, sizeof(opt), &opt);
1058 nla_nest_end(skb, nest);
1059 spin_unlock_bh(root_lock);
1063 spin_unlock_bh(root_lock);
1064 nla_nest_cancel(skb, nest);
1069 htb_dump_class_stats(struct Qdisc *sch, unsigned long arg, struct gnet_dump *d)
1071 struct htb_class *cl = (struct htb_class *)arg;
1073 if (!cl->level && cl->un.leaf.q)
1074 cl->qstats.qlen = cl->un.leaf.q->q.qlen;
1075 cl->xstats.tokens = cl->tokens;
1076 cl->xstats.ctokens = cl->ctokens;
1078 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
1079 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1080 gnet_stats_copy_queue(d, &cl->qstats) < 0)
1083 return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1086 static int htb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1089 struct htb_class *cl = (struct htb_class *)arg;
1091 if (cl && !cl->level) {
1093 (new = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
1095 cl->common.classid))
1099 *old = cl->un.leaf.q;
1100 cl->un.leaf.q = new;
1102 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
1105 sch_tree_unlock(sch);
1111 static struct Qdisc *htb_leaf(struct Qdisc *sch, unsigned long arg)
1113 struct htb_class *cl = (struct htb_class *)arg;
1114 return (cl && !cl->level) ? cl->un.leaf.q : NULL;
1117 static void htb_qlen_notify(struct Qdisc *sch, unsigned long arg)
1119 struct htb_class *cl = (struct htb_class *)arg;
1121 if (cl->un.leaf.q->q.qlen == 0)
1122 htb_deactivate(qdisc_priv(sch), cl);
1125 static unsigned long htb_get(struct Qdisc *sch, u32 classid)
1127 struct htb_class *cl = htb_find(classid, sch);
1130 return (unsigned long)cl;
1133 static inline int htb_parent_last_child(struct htb_class *cl)
1136 /* the root class */
1138 if (cl->parent->children > 1)
1139 /* not the last child */
1144 static void htb_parent_to_leaf(struct htb_sched *q, struct htb_class *cl,
1145 struct Qdisc *new_q)
1147 struct htb_class *parent = cl->parent;
1149 WARN_ON(cl->level || !cl->un.leaf.q || cl->prio_activity);
1151 if (parent->cmode != HTB_CAN_SEND)
1152 htb_safe_rb_erase(&parent->pq_node, q->wait_pq + parent->level);
1155 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1156 INIT_LIST_HEAD(&parent->un.leaf.drop_list);
1157 parent->un.leaf.q = new_q ? new_q : &noop_qdisc;
1158 parent->un.leaf.quantum = parent->quantum;
1159 parent->un.leaf.prio = parent->prio;
1160 parent->tokens = parent->buffer;
1161 parent->ctokens = parent->cbuffer;
1162 parent->t_c = psched_get_time();
1163 parent->cmode = HTB_CAN_SEND;
1166 static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
1169 WARN_ON(!cl->un.leaf.q);
1170 qdisc_destroy(cl->un.leaf.q);
1172 gen_kill_estimator(&cl->bstats, &cl->rate_est);
1173 qdisc_put_rtab(cl->rate);
1174 qdisc_put_rtab(cl->ceil);
1176 tcf_destroy_chain(&cl->filter_list);
1180 /* always caled under BH & queue lock */
1181 static void htb_destroy(struct Qdisc *sch)
1183 struct htb_sched *q = qdisc_priv(sch);
1184 struct hlist_node *n, *next;
1185 struct htb_class *cl;
1188 qdisc_watchdog_cancel(&q->watchdog);
1189 /* This line used to be after htb_destroy_class call below
1190 and surprisingly it worked in 2.4. But it must precede it
1191 because filter need its target class alive to be able to call
1192 unbind_filter on it (without Oops). */
1193 tcf_destroy_chain(&q->filter_list);
1195 for (i = 0; i < q->clhash.hashsize; i++) {
1196 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode)
1197 tcf_destroy_chain(&cl->filter_list);
1199 for (i = 0; i < q->clhash.hashsize; i++) {
1200 hlist_for_each_entry_safe(cl, n, next, &q->clhash.hash[i],
1202 htb_destroy_class(sch, cl);
1204 qdisc_class_hash_destroy(&q->clhash);
1205 __skb_queue_purge(&q->direct_queue);
1208 static int htb_delete(struct Qdisc *sch, unsigned long arg)
1210 struct htb_sched *q = qdisc_priv(sch);
1211 struct htb_class *cl = (struct htb_class *)arg;
1213 struct Qdisc *new_q = NULL;
1216 // TODO: why don't allow to delete subtree ? references ? does
1217 // tc subsys quarantee us that in htb_destroy it holds no class
1218 // refs so that we can remove children safely there ?
1219 if (cl->children || cl->filter_cnt)
1222 if (!cl->level && htb_parent_last_child(cl)) {
1223 new_q = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
1225 cl->parent->common.classid);
1232 qlen = cl->un.leaf.q->q.qlen;
1233 qdisc_reset(cl->un.leaf.q);
1234 qdisc_tree_decrease_qlen(cl->un.leaf.q, qlen);
1237 /* delete from hash and active; remainder in destroy_class */
1238 qdisc_class_hash_remove(&q->clhash, &cl->common);
1240 cl->parent->children--;
1242 if (cl->prio_activity)
1243 htb_deactivate(q, cl);
1245 if (cl->cmode != HTB_CAN_SEND)
1246 htb_safe_rb_erase(&cl->pq_node, q->wait_pq + cl->level);
1249 htb_parent_to_leaf(q, cl, new_q);
1251 if (--cl->refcnt == 0)
1252 htb_destroy_class(sch, cl);
1254 sch_tree_unlock(sch);
1258 static void htb_put(struct Qdisc *sch, unsigned long arg)
1260 struct htb_class *cl = (struct htb_class *)arg;
1262 if (--cl->refcnt == 0)
1263 htb_destroy_class(sch, cl);
1266 static int htb_change_class(struct Qdisc *sch, u32 classid,
1267 u32 parentid, struct nlattr **tca,
1271 struct htb_sched *q = qdisc_priv(sch);
1272 struct htb_class *cl = (struct htb_class *)*arg, *parent;
1273 struct nlattr *opt = tca[TCA_OPTIONS];
1274 struct qdisc_rate_table *rtab = NULL, *ctab = NULL;
1275 struct nlattr *tb[TCA_HTB_RTAB + 1];
1276 struct tc_htb_opt *hopt;
1278 /* extract all subattrs from opt attr */
1282 err = nla_parse_nested(tb, TCA_HTB_RTAB, opt, htb_policy);
1287 if (tb[TCA_HTB_PARMS] == NULL)
1290 parent = parentid == TC_H_ROOT ? NULL : htb_find(parentid, sch);
1292 hopt = nla_data(tb[TCA_HTB_PARMS]);
1294 rtab = qdisc_get_rtab(&hopt->rate, tb[TCA_HTB_RTAB]);
1295 ctab = qdisc_get_rtab(&hopt->ceil, tb[TCA_HTB_CTAB]);
1299 if (!cl) { /* new class */
1300 struct Qdisc *new_q;
1304 struct gnet_estimator opt;
1307 .nla_len = nla_attr_size(sizeof(est.opt)),
1308 .nla_type = TCA_RATE,
1311 /* 4s interval, 16s averaging constant */
1317 /* check for valid classid */
1318 if (!classid || TC_H_MAJ(classid ^ sch->handle)
1319 || htb_find(classid, sch))
1322 /* check maximal depth */
1323 if (parent && parent->parent && parent->parent->level < 2) {
1324 printk(KERN_ERR "htb: tree is too deep\n");
1328 if ((cl = kzalloc(sizeof(*cl), GFP_KERNEL)) == NULL)
1331 err = gen_new_estimator(&cl->bstats, &cl->rate_est,
1332 qdisc_root_sleeping_lock(sch),
1333 tca[TCA_RATE] ? : &est.nla);
1341 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
1342 RB_CLEAR_NODE(&cl->pq_node);
1344 for (prio = 0; prio < TC_HTB_NUMPRIO; prio++)
1345 RB_CLEAR_NODE(&cl->node[prio]);
1347 /* create leaf qdisc early because it uses kmalloc(GFP_KERNEL)
1348 so that can't be used inside of sch_tree_lock
1349 -- thanks to Karlis Peisenieks */
1350 new_q = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
1351 &pfifo_qdisc_ops, classid);
1353 if (parent && !parent->level) {
1354 unsigned int qlen = parent->un.leaf.q->q.qlen;
1356 /* turn parent into inner node */
1357 qdisc_reset(parent->un.leaf.q);
1358 qdisc_tree_decrease_qlen(parent->un.leaf.q, qlen);
1359 qdisc_destroy(parent->un.leaf.q);
1360 if (parent->prio_activity)
1361 htb_deactivate(q, parent);
1363 /* remove from evt list because of level change */
1364 if (parent->cmode != HTB_CAN_SEND) {
1365 htb_safe_rb_erase(&parent->pq_node, q->wait_pq);
1366 parent->cmode = HTB_CAN_SEND;
1368 parent->level = (parent->parent ? parent->parent->level
1369 : TC_HTB_MAXDEPTH) - 1;
1370 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1372 /* leaf (we) needs elementary qdisc */
1373 cl->un.leaf.q = new_q ? new_q : &noop_qdisc;
1375 cl->common.classid = classid;
1376 cl->parent = parent;
1378 /* set class to be in HTB_CAN_SEND state */
1379 cl->tokens = hopt->buffer;
1380 cl->ctokens = hopt->cbuffer;
1381 cl->mbuffer = 60 * PSCHED_TICKS_PER_SEC; /* 1min */
1382 cl->t_c = psched_get_time();
1383 cl->cmode = HTB_CAN_SEND;
1385 /* attach to the hash list and parent's family */
1386 qdisc_class_hash_insert(&q->clhash, &cl->common);
1390 if (tca[TCA_RATE]) {
1391 err = gen_replace_estimator(&cl->bstats, &cl->rate_est,
1392 qdisc_root_sleeping_lock(sch),
1400 /* it used to be a nasty bug here, we have to check that node
1401 is really leaf before changing cl->un.leaf ! */
1403 cl->un.leaf.quantum = rtab->rate.rate / q->rate2quantum;
1404 if (!hopt->quantum && cl->un.leaf.quantum < 1000) {
1406 "HTB: quantum of class %X is small. Consider r2q change.\n",
1407 cl->common.classid);
1408 cl->un.leaf.quantum = 1000;
1410 if (!hopt->quantum && cl->un.leaf.quantum > 200000) {
1412 "HTB: quantum of class %X is big. Consider r2q change.\n",
1413 cl->common.classid);
1414 cl->un.leaf.quantum = 200000;
1417 cl->un.leaf.quantum = hopt->quantum;
1418 if ((cl->un.leaf.prio = hopt->prio) >= TC_HTB_NUMPRIO)
1419 cl->un.leaf.prio = TC_HTB_NUMPRIO - 1;
1421 /* backup for htb_parent_to_leaf */
1422 cl->quantum = cl->un.leaf.quantum;
1423 cl->prio = cl->un.leaf.prio;
1426 cl->buffer = hopt->buffer;
1427 cl->cbuffer = hopt->cbuffer;
1429 qdisc_put_rtab(cl->rate);
1432 qdisc_put_rtab(cl->ceil);
1434 sch_tree_unlock(sch);
1436 qdisc_class_hash_grow(sch, &q->clhash);
1438 *arg = (unsigned long)cl;
1443 qdisc_put_rtab(rtab);
1445 qdisc_put_rtab(ctab);
1449 static struct tcf_proto **htb_find_tcf(struct Qdisc *sch, unsigned long arg)
1451 struct htb_sched *q = qdisc_priv(sch);
1452 struct htb_class *cl = (struct htb_class *)arg;
1453 struct tcf_proto **fl = cl ? &cl->filter_list : &q->filter_list;
1458 static unsigned long htb_bind_filter(struct Qdisc *sch, unsigned long parent,
1461 struct htb_class *cl = htb_find(classid, sch);
1463 /*if (cl && !cl->level) return 0;
1464 The line above used to be there to prevent attaching filters to
1465 leaves. But at least tc_index filter uses this just to get class
1466 for other reasons so that we have to allow for it.
1468 19.6.2002 As Werner explained it is ok - bind filter is just
1469 another way to "lock" the class - unlike "get" this lock can
1470 be broken by class during destroy IIUC.
1474 return (unsigned long)cl;
1477 static void htb_unbind_filter(struct Qdisc *sch, unsigned long arg)
1479 struct htb_class *cl = (struct htb_class *)arg;
1485 static void htb_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1487 struct htb_sched *q = qdisc_priv(sch);
1488 struct htb_class *cl;
1489 struct hlist_node *n;
1495 for (i = 0; i < q->clhash.hashsize; i++) {
1496 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode) {
1497 if (arg->count < arg->skip) {
1501 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1510 static const struct Qdisc_class_ops htb_class_ops = {
1513 .qlen_notify = htb_qlen_notify,
1516 .change = htb_change_class,
1517 .delete = htb_delete,
1519 .tcf_chain = htb_find_tcf,
1520 .bind_tcf = htb_bind_filter,
1521 .unbind_tcf = htb_unbind_filter,
1522 .dump = htb_dump_class,
1523 .dump_stats = htb_dump_class_stats,
1526 static struct Qdisc_ops htb_qdisc_ops __read_mostly = {
1528 .cl_ops = &htb_class_ops,
1530 .priv_size = sizeof(struct htb_sched),
1531 .enqueue = htb_enqueue,
1532 .dequeue = htb_dequeue,
1533 .peek = qdisc_peek_dequeued,
1537 .destroy = htb_destroy,
1538 .change = NULL /* htb_change */,
1540 .owner = THIS_MODULE,
1543 static int __init htb_module_init(void)
1545 return register_qdisc(&htb_qdisc_ops);
1547 static void __exit htb_module_exit(void)
1549 unregister_qdisc(&htb_qdisc_ops);
1552 module_init(htb_module_init)
1553 module_exit(htb_module_exit)
1554 MODULE_LICENSE("GPL");