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_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;
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 BUG_TRAP(!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 BUG_TRAP(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 (qdisc_enqueue(skb, cl->un.leaf.q) != NET_XMIT_SUCCESS) {
578 return NET_XMIT_DROP;
580 cl->bstats.packets +=
581 skb_is_gso(skb)?skb_shinfo(skb)->gso_segs:1;
582 cl->bstats.bytes += skb->len;
587 sch->bstats.packets += skb_is_gso(skb)?skb_shinfo(skb)->gso_segs:1;
588 sch->bstats.bytes += skb->len;
589 return NET_XMIT_SUCCESS;
592 /* TODO: requeuing packet charges it to policers again !! */
593 static int htb_requeue(struct sk_buff *skb, struct Qdisc *sch)
596 struct htb_sched *q = qdisc_priv(sch);
597 struct htb_class *cl = htb_classify(skb, sch, &ret);
598 struct sk_buff *tskb;
600 if (cl == HTB_DIRECT) {
601 /* enqueue to helper queue */
602 if (q->direct_queue.qlen < q->direct_qlen) {
603 __skb_queue_head(&q->direct_queue, skb);
605 __skb_queue_head(&q->direct_queue, skb);
606 tskb = __skb_dequeue_tail(&q->direct_queue);
611 #ifdef CONFIG_NET_CLS_ACT
613 if (ret == NET_XMIT_BYPASS)
618 } else if (cl->un.leaf.q->ops->requeue(skb, cl->un.leaf.q) !=
622 return NET_XMIT_DROP;
627 sch->qstats.requeues++;
628 return NET_XMIT_SUCCESS;
632 * htb_charge_class - charges amount "bytes" to leaf and ancestors
634 * Routine assumes that packet "bytes" long was dequeued from leaf cl
635 * borrowing from "level". It accounts bytes to ceil leaky bucket for
636 * leaf and all ancestors and to rate bucket for ancestors at levels
637 * "level" and higher. It also handles possible change of mode resulting
638 * from the update. Note that mode can also increase here (MAY_BORROW to
639 * CAN_SEND) because we can use more precise clock that event queue here.
640 * In such case we remove class from event queue first.
642 static void htb_charge_class(struct htb_sched *q, struct htb_class *cl,
643 int level, struct sk_buff *skb)
645 int bytes = skb->len;
647 enum htb_cmode old_mode;
649 #define HTB_ACCNT(T,B,R) toks = diff + cl->T; \
650 if (toks > cl->B) toks = cl->B; \
651 toks -= L2T(cl, cl->R, bytes); \
652 if (toks <= -cl->mbuffer) toks = 1-cl->mbuffer; \
656 diff = psched_tdiff_bounded(q->now, cl->t_c, cl->mbuffer);
657 if (cl->level >= level) {
658 if (cl->level == level)
660 HTB_ACCNT(tokens, buffer, rate);
662 cl->xstats.borrows++;
663 cl->tokens += diff; /* we moved t_c; update tokens */
665 HTB_ACCNT(ctokens, cbuffer, ceil);
668 old_mode = cl->cmode;
670 htb_change_class_mode(q, cl, &diff);
671 if (old_mode != cl->cmode) {
672 if (old_mode != HTB_CAN_SEND)
673 htb_safe_rb_erase(&cl->pq_node, q->wait_pq + cl->level);
674 if (cl->cmode != HTB_CAN_SEND)
675 htb_add_to_wait_tree(q, cl, diff);
678 /* update byte stats except for leaves which are already updated */
680 cl->bstats.bytes += bytes;
681 cl->bstats.packets += skb_is_gso(skb)?
682 skb_shinfo(skb)->gso_segs:1;
689 * htb_do_events - make mode changes to classes at the level
691 * Scans event queue for pending events and applies them. Returns time of
692 * next pending event (0 for no event in pq).
693 * Note: Applied are events whose have cl->pq_key <= q->now.
695 static psched_time_t htb_do_events(struct htb_sched *q, int level)
697 /* don't run for longer than 2 jiffies; 2 is used instead of
698 1 to simplify things when jiffy is going to be incremented
700 unsigned long stop_at = jiffies + 2;
701 while (time_before(jiffies, stop_at)) {
702 struct htb_class *cl;
704 struct rb_node *p = rb_first(&q->wait_pq[level]);
709 cl = rb_entry(p, struct htb_class, pq_node);
710 if (cl->pq_key > q->now)
713 htb_safe_rb_erase(p, q->wait_pq + level);
714 diff = psched_tdiff_bounded(q->now, cl->t_c, cl->mbuffer);
715 htb_change_class_mode(q, cl, &diff);
716 if (cl->cmode != HTB_CAN_SEND)
717 htb_add_to_wait_tree(q, cl, diff);
719 /* too much load - let's continue on next jiffie */
720 return q->now + PSCHED_TICKS_PER_SEC / HZ;
723 /* Returns class->node+prio from id-tree where classe's id is >= id. NULL
724 is no such one exists. */
725 static struct rb_node *htb_id_find_next_upper(int prio, struct rb_node *n,
728 struct rb_node *r = NULL;
730 struct htb_class *cl =
731 rb_entry(n, struct htb_class, node[prio]);
732 if (id == cl->common.classid)
735 if (id > cl->common.classid) {
746 * htb_lookup_leaf - returns next leaf class in DRR order
748 * Find leaf where current feed pointers points to.
750 static struct htb_class *htb_lookup_leaf(struct rb_root *tree, int prio,
751 struct rb_node **pptr, u32 * pid)
755 struct rb_node *root;
756 struct rb_node **pptr;
758 } stk[TC_HTB_MAXDEPTH], *sp = stk;
760 BUG_TRAP(tree->rb_node);
761 sp->root = tree->rb_node;
765 for (i = 0; i < 65535; i++) {
766 if (!*sp->pptr && *sp->pid) {
767 /* ptr was invalidated but id is valid - try to recover
768 the original or next ptr */
770 htb_id_find_next_upper(prio, sp->root, *sp->pid);
772 *sp->pid = 0; /* ptr is valid now so that remove this hint as it
773 can become out of date quickly */
774 if (!*sp->pptr) { /* we are at right end; rewind & go up */
775 *sp->pptr = sp->root;
776 while ((*sp->pptr)->rb_left)
777 *sp->pptr = (*sp->pptr)->rb_left;
783 htb_next_rb_node(sp->pptr);
786 struct htb_class *cl;
787 cl = rb_entry(*sp->pptr, struct htb_class, node[prio]);
790 (++sp)->root = cl->un.inner.feed[prio].rb_node;
791 sp->pptr = cl->un.inner.ptr + prio;
792 sp->pid = cl->un.inner.last_ptr_id + prio;
799 /* dequeues packet at given priority and level; call only if
800 you are sure that there is active class at prio/level */
801 static struct sk_buff *htb_dequeue_tree(struct htb_sched *q, int prio,
804 struct sk_buff *skb = NULL;
805 struct htb_class *cl, *start;
806 /* look initial class up in the row */
807 start = cl = htb_lookup_leaf(q->row[level] + prio, prio,
808 q->ptr[level] + prio,
809 q->last_ptr_id[level] + prio);
817 /* class can be empty - it is unlikely but can be true if leaf
818 qdisc drops packets in enqueue routine or if someone used
819 graft operation on the leaf since last dequeue;
820 simply deactivate and skip such class */
821 if (unlikely(cl->un.leaf.q->q.qlen == 0)) {
822 struct htb_class *next;
823 htb_deactivate(q, cl);
825 /* row/level might become empty */
826 if ((q->row_mask[level] & (1 << prio)) == 0)
829 next = htb_lookup_leaf(q->row[level] + prio,
830 prio, q->ptr[level] + prio,
831 q->last_ptr_id[level] + prio);
833 if (cl == start) /* fix start if we just deleted it */
839 skb = cl->un.leaf.q->dequeue(cl->un.leaf.q);
840 if (likely(skb != NULL))
844 "htb: class %X isn't work conserving ?!\n",
849 htb_next_rb_node((level ? cl->parent->un.inner.ptr : q->
851 cl = htb_lookup_leaf(q->row[level] + prio, prio,
852 q->ptr[level] + prio,
853 q->last_ptr_id[level] + prio);
855 } while (cl != start);
857 if (likely(skb != NULL)) {
858 if ((cl->un.leaf.deficit[level] -= skb->len) < 0) {
859 cl->un.leaf.deficit[level] += cl->un.leaf.quantum;
860 htb_next_rb_node((level ? cl->parent->un.inner.ptr : q->
863 /* this used to be after charge_class but this constelation
864 gives us slightly better performance */
865 if (!cl->un.leaf.q->q.qlen)
866 htb_deactivate(q, cl);
867 htb_charge_class(q, cl, level, skb);
872 static struct sk_buff *htb_dequeue(struct Qdisc *sch)
874 struct sk_buff *skb = NULL;
875 struct htb_sched *q = qdisc_priv(sch);
877 psched_time_t next_event;
879 /* try to dequeue direct packets as high prio (!) to minimize cpu work */
880 skb = __skb_dequeue(&q->direct_queue);
882 sch->flags &= ~TCQ_F_THROTTLED;
889 q->now = psched_get_time();
891 next_event = q->now + 5 * PSCHED_TICKS_PER_SEC;
893 for (level = 0; level < TC_HTB_MAXDEPTH; level++) {
894 /* common case optimization - skip event handler quickly */
898 if (q->now >= q->near_ev_cache[level]) {
899 event = htb_do_events(q, level);
901 event = q->now + PSCHED_TICKS_PER_SEC;
902 q->near_ev_cache[level] = event;
904 event = q->near_ev_cache[level];
906 if (event && next_event > event)
909 m = ~q->row_mask[level];
910 while (m != (int)(-1)) {
913 skb = htb_dequeue_tree(q, prio, level);
914 if (likely(skb != NULL)) {
916 sch->flags &= ~TCQ_F_THROTTLED;
921 sch->qstats.overlimits++;
922 qdisc_watchdog_schedule(&q->watchdog, next_event);
927 /* try to drop from each class (by prio) until one succeed */
928 static unsigned int htb_drop(struct Qdisc *sch)
930 struct htb_sched *q = qdisc_priv(sch);
933 for (prio = TC_HTB_NUMPRIO - 1; prio >= 0; prio--) {
935 list_for_each(p, q->drops + prio) {
936 struct htb_class *cl = list_entry(p, struct htb_class,
939 if (cl->un.leaf.q->ops->drop &&
940 (len = cl->un.leaf.q->ops->drop(cl->un.leaf.q))) {
942 if (!cl->un.leaf.q->q.qlen)
943 htb_deactivate(q, cl);
951 /* reset all classes */
952 /* always caled under BH & queue lock */
953 static void htb_reset(struct Qdisc *sch)
955 struct htb_sched *q = qdisc_priv(sch);
956 struct htb_class *cl;
957 struct hlist_node *n;
960 for (i = 0; i < q->clhash.hashsize; i++) {
961 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode) {
963 memset(&cl->un.inner, 0, sizeof(cl->un.inner));
966 qdisc_reset(cl->un.leaf.q);
967 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
969 cl->prio_activity = 0;
970 cl->cmode = HTB_CAN_SEND;
974 qdisc_watchdog_cancel(&q->watchdog);
975 __skb_queue_purge(&q->direct_queue);
977 memset(q->row, 0, sizeof(q->row));
978 memset(q->row_mask, 0, sizeof(q->row_mask));
979 memset(q->wait_pq, 0, sizeof(q->wait_pq));
980 memset(q->ptr, 0, sizeof(q->ptr));
981 for (i = 0; i < TC_HTB_NUMPRIO; i++)
982 INIT_LIST_HEAD(q->drops + i);
985 static const struct nla_policy htb_policy[TCA_HTB_MAX + 1] = {
986 [TCA_HTB_PARMS] = { .len = sizeof(struct tc_htb_opt) },
987 [TCA_HTB_INIT] = { .len = sizeof(struct tc_htb_glob) },
988 [TCA_HTB_CTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
989 [TCA_HTB_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
992 static int htb_init(struct Qdisc *sch, struct nlattr *opt)
994 struct htb_sched *q = qdisc_priv(sch);
995 struct nlattr *tb[TCA_HTB_INIT + 1];
996 struct tc_htb_glob *gopt;
1003 err = nla_parse_nested(tb, TCA_HTB_INIT, opt, htb_policy);
1007 if (tb[TCA_HTB_INIT] == NULL) {
1008 printk(KERN_ERR "HTB: hey probably you have bad tc tool ?\n");
1011 gopt = nla_data(tb[TCA_HTB_INIT]);
1012 if (gopt->version != HTB_VER >> 16) {
1014 "HTB: need tc/htb version %d (minor is %d), you have %d\n",
1015 HTB_VER >> 16, HTB_VER & 0xffff, gopt->version);
1019 err = qdisc_class_hash_init(&q->clhash);
1022 for (i = 0; i < TC_HTB_NUMPRIO; i++)
1023 INIT_LIST_HEAD(q->drops + i);
1025 qdisc_watchdog_init(&q->watchdog, sch);
1026 skb_queue_head_init(&q->direct_queue);
1028 q->direct_qlen = qdisc_dev(sch)->tx_queue_len;
1029 if (q->direct_qlen < 2) /* some devices have zero tx_queue_len */
1032 if ((q->rate2quantum = gopt->rate2quantum) < 1)
1033 q->rate2quantum = 1;
1034 q->defcls = gopt->defcls;
1039 static int htb_dump(struct Qdisc *sch, struct sk_buff *skb)
1041 spinlock_t *root_lock = qdisc_root_lock(sch);
1042 struct htb_sched *q = qdisc_priv(sch);
1043 struct nlattr *nest;
1044 struct tc_htb_glob gopt;
1046 spin_lock_bh(root_lock);
1048 gopt.direct_pkts = q->direct_pkts;
1049 gopt.version = HTB_VER;
1050 gopt.rate2quantum = q->rate2quantum;
1051 gopt.defcls = q->defcls;
1054 nest = nla_nest_start(skb, TCA_OPTIONS);
1056 goto nla_put_failure;
1057 NLA_PUT(skb, TCA_HTB_INIT, sizeof(gopt), &gopt);
1058 nla_nest_end(skb, nest);
1060 spin_unlock_bh(root_lock);
1064 spin_unlock_bh(root_lock);
1065 nla_nest_cancel(skb, nest);
1069 static int htb_dump_class(struct Qdisc *sch, unsigned long arg,
1070 struct sk_buff *skb, struct tcmsg *tcm)
1072 struct htb_class *cl = (struct htb_class *)arg;
1073 spinlock_t *root_lock = qdisc_root_lock(sch);
1074 struct nlattr *nest;
1075 struct tc_htb_opt opt;
1077 spin_lock_bh(root_lock);
1078 tcm->tcm_parent = cl->parent ? cl->parent->common.classid : TC_H_ROOT;
1079 tcm->tcm_handle = cl->common.classid;
1080 if (!cl->level && cl->un.leaf.q)
1081 tcm->tcm_info = cl->un.leaf.q->handle;
1083 nest = nla_nest_start(skb, TCA_OPTIONS);
1085 goto nla_put_failure;
1087 memset(&opt, 0, sizeof(opt));
1089 opt.rate = cl->rate->rate;
1090 opt.buffer = cl->buffer;
1091 opt.ceil = cl->ceil->rate;
1092 opt.cbuffer = cl->cbuffer;
1093 opt.quantum = cl->un.leaf.quantum;
1094 opt.prio = cl->un.leaf.prio;
1095 opt.level = cl->level;
1096 NLA_PUT(skb, TCA_HTB_PARMS, sizeof(opt), &opt);
1098 nla_nest_end(skb, nest);
1099 spin_unlock_bh(root_lock);
1103 spin_unlock_bh(root_lock);
1104 nla_nest_cancel(skb, nest);
1109 htb_dump_class_stats(struct Qdisc *sch, unsigned long arg, struct gnet_dump *d)
1111 struct htb_class *cl = (struct htb_class *)arg;
1113 if (!cl->level && cl->un.leaf.q)
1114 cl->qstats.qlen = cl->un.leaf.q->q.qlen;
1115 cl->xstats.tokens = cl->tokens;
1116 cl->xstats.ctokens = cl->ctokens;
1118 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
1119 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1120 gnet_stats_copy_queue(d, &cl->qstats) < 0)
1123 return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1126 static int htb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1129 struct htb_class *cl = (struct htb_class *)arg;
1131 if (cl && !cl->level) {
1133 (new = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
1135 cl->common.classid))
1139 if ((*old = xchg(&cl->un.leaf.q, new)) != NULL) {
1140 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
1143 sch_tree_unlock(sch);
1149 static struct Qdisc *htb_leaf(struct Qdisc *sch, unsigned long arg)
1151 struct htb_class *cl = (struct htb_class *)arg;
1152 return (cl && !cl->level) ? cl->un.leaf.q : NULL;
1155 static void htb_qlen_notify(struct Qdisc *sch, unsigned long arg)
1157 struct htb_class *cl = (struct htb_class *)arg;
1159 if (cl->un.leaf.q->q.qlen == 0)
1160 htb_deactivate(qdisc_priv(sch), cl);
1163 static unsigned long htb_get(struct Qdisc *sch, u32 classid)
1165 struct htb_class *cl = htb_find(classid, sch);
1168 return (unsigned long)cl;
1171 static inline int htb_parent_last_child(struct htb_class *cl)
1174 /* the root class */
1176 if (cl->parent->children > 1)
1177 /* not the last child */
1182 static void htb_parent_to_leaf(struct htb_sched *q, struct htb_class *cl,
1183 struct Qdisc *new_q)
1185 struct htb_class *parent = cl->parent;
1187 BUG_TRAP(!cl->level && cl->un.leaf.q && !cl->prio_activity);
1189 if (parent->cmode != HTB_CAN_SEND)
1190 htb_safe_rb_erase(&parent->pq_node, q->wait_pq + parent->level);
1193 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1194 INIT_LIST_HEAD(&parent->un.leaf.drop_list);
1195 parent->un.leaf.q = new_q ? new_q : &noop_qdisc;
1196 parent->un.leaf.quantum = parent->quantum;
1197 parent->un.leaf.prio = parent->prio;
1198 parent->tokens = parent->buffer;
1199 parent->ctokens = parent->cbuffer;
1200 parent->t_c = psched_get_time();
1201 parent->cmode = HTB_CAN_SEND;
1204 static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
1207 BUG_TRAP(cl->un.leaf.q);
1208 qdisc_destroy(cl->un.leaf.q);
1210 gen_kill_estimator(&cl->bstats, &cl->rate_est);
1211 qdisc_put_rtab(cl->rate);
1212 qdisc_put_rtab(cl->ceil);
1214 tcf_destroy_chain(&cl->filter_list);
1218 /* always caled under BH & queue lock */
1219 static void htb_destroy(struct Qdisc *sch)
1221 struct htb_sched *q = qdisc_priv(sch);
1222 struct hlist_node *n, *next;
1223 struct htb_class *cl;
1226 qdisc_watchdog_cancel(&q->watchdog);
1227 /* This line used to be after htb_destroy_class call below
1228 and surprisingly it worked in 2.4. But it must precede it
1229 because filter need its target class alive to be able to call
1230 unbind_filter on it (without Oops). */
1231 tcf_destroy_chain(&q->filter_list);
1233 for (i = 0; i < q->clhash.hashsize; i++) {
1234 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode)
1235 tcf_destroy_chain(&cl->filter_list);
1237 for (i = 0; i < q->clhash.hashsize; i++) {
1238 hlist_for_each_entry_safe(cl, n, next, &q->clhash.hash[i],
1240 htb_destroy_class(sch, cl);
1242 qdisc_class_hash_destroy(&q->clhash);
1243 __skb_queue_purge(&q->direct_queue);
1246 static int htb_delete(struct Qdisc *sch, unsigned long arg)
1248 struct htb_sched *q = qdisc_priv(sch);
1249 struct htb_class *cl = (struct htb_class *)arg;
1251 struct Qdisc *new_q = NULL;
1254 // TODO: why don't allow to delete subtree ? references ? does
1255 // tc subsys quarantee us that in htb_destroy it holds no class
1256 // refs so that we can remove children safely there ?
1257 if (cl->children || cl->filter_cnt)
1260 if (!cl->level && htb_parent_last_child(cl)) {
1261 new_q = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
1263 cl->parent->common.classid);
1270 qlen = cl->un.leaf.q->q.qlen;
1271 qdisc_reset(cl->un.leaf.q);
1272 qdisc_tree_decrease_qlen(cl->un.leaf.q, qlen);
1275 /* delete from hash and active; remainder in destroy_class */
1276 qdisc_class_hash_remove(&q->clhash, &cl->common);
1277 cl->parent->children--;
1279 if (cl->prio_activity)
1280 htb_deactivate(q, cl);
1282 if (cl->cmode != HTB_CAN_SEND)
1283 htb_safe_rb_erase(&cl->pq_node, q->wait_pq + cl->level);
1286 htb_parent_to_leaf(q, cl, new_q);
1288 if (--cl->refcnt == 0)
1289 htb_destroy_class(sch, cl);
1291 sch_tree_unlock(sch);
1295 static void htb_put(struct Qdisc *sch, unsigned long arg)
1297 struct htb_class *cl = (struct htb_class *)arg;
1299 if (--cl->refcnt == 0)
1300 htb_destroy_class(sch, cl);
1303 static int htb_change_class(struct Qdisc *sch, u32 classid,
1304 u32 parentid, struct nlattr **tca,
1308 struct htb_sched *q = qdisc_priv(sch);
1309 struct htb_class *cl = (struct htb_class *)*arg, *parent;
1310 struct nlattr *opt = tca[TCA_OPTIONS];
1311 struct qdisc_rate_table *rtab = NULL, *ctab = NULL;
1312 struct nlattr *tb[TCA_HTB_RTAB + 1];
1313 struct tc_htb_opt *hopt;
1315 /* extract all subattrs from opt attr */
1319 err = nla_parse_nested(tb, TCA_HTB_RTAB, opt, htb_policy);
1324 if (tb[TCA_HTB_PARMS] == NULL)
1327 parent = parentid == TC_H_ROOT ? NULL : htb_find(parentid, sch);
1329 hopt = nla_data(tb[TCA_HTB_PARMS]);
1331 rtab = qdisc_get_rtab(&hopt->rate, tb[TCA_HTB_RTAB]);
1332 ctab = qdisc_get_rtab(&hopt->ceil, tb[TCA_HTB_CTAB]);
1336 if (!cl) { /* new class */
1337 struct Qdisc *new_q;
1341 struct gnet_estimator opt;
1344 .nla_len = nla_attr_size(sizeof(est.opt)),
1345 .nla_type = TCA_RATE,
1348 /* 4s interval, 16s averaging constant */
1354 /* check for valid classid */
1355 if (!classid || TC_H_MAJ(classid ^ sch->handle)
1356 || htb_find(classid, sch))
1359 /* check maximal depth */
1360 if (parent && parent->parent && parent->parent->level < 2) {
1361 printk(KERN_ERR "htb: tree is too deep\n");
1365 if ((cl = kzalloc(sizeof(*cl), GFP_KERNEL)) == NULL)
1368 gen_new_estimator(&cl->bstats, &cl->rate_est,
1369 qdisc_root_lock(sch),
1370 tca[TCA_RATE] ? : &est.nla);
1373 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
1374 RB_CLEAR_NODE(&cl->pq_node);
1376 for (prio = 0; prio < TC_HTB_NUMPRIO; prio++)
1377 RB_CLEAR_NODE(&cl->node[prio]);
1379 /* create leaf qdisc early because it uses kmalloc(GFP_KERNEL)
1380 so that can't be used inside of sch_tree_lock
1381 -- thanks to Karlis Peisenieks */
1382 new_q = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
1383 &pfifo_qdisc_ops, classid);
1385 if (parent && !parent->level) {
1386 unsigned int qlen = parent->un.leaf.q->q.qlen;
1388 /* turn parent into inner node */
1389 qdisc_reset(parent->un.leaf.q);
1390 qdisc_tree_decrease_qlen(parent->un.leaf.q, qlen);
1391 qdisc_destroy(parent->un.leaf.q);
1392 if (parent->prio_activity)
1393 htb_deactivate(q, parent);
1395 /* remove from evt list because of level change */
1396 if (parent->cmode != HTB_CAN_SEND) {
1397 htb_safe_rb_erase(&parent->pq_node, q->wait_pq);
1398 parent->cmode = HTB_CAN_SEND;
1400 parent->level = (parent->parent ? parent->parent->level
1401 : TC_HTB_MAXDEPTH) - 1;
1402 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1404 /* leaf (we) needs elementary qdisc */
1405 cl->un.leaf.q = new_q ? new_q : &noop_qdisc;
1407 cl->common.classid = classid;
1408 cl->parent = parent;
1410 /* set class to be in HTB_CAN_SEND state */
1411 cl->tokens = hopt->buffer;
1412 cl->ctokens = hopt->cbuffer;
1413 cl->mbuffer = 60 * PSCHED_TICKS_PER_SEC; /* 1min */
1414 cl->t_c = psched_get_time();
1415 cl->cmode = HTB_CAN_SEND;
1417 /* attach to the hash list and parent's family */
1418 qdisc_class_hash_insert(&q->clhash, &cl->common);
1423 gen_replace_estimator(&cl->bstats, &cl->rate_est,
1424 qdisc_root_lock(sch),
1429 /* it used to be a nasty bug here, we have to check that node
1430 is really leaf before changing cl->un.leaf ! */
1432 cl->un.leaf.quantum = rtab->rate.rate / q->rate2quantum;
1433 if (!hopt->quantum && cl->un.leaf.quantum < 1000) {
1435 "HTB: quantum of class %X is small. Consider r2q change.\n",
1436 cl->common.classid);
1437 cl->un.leaf.quantum = 1000;
1439 if (!hopt->quantum && cl->un.leaf.quantum > 200000) {
1441 "HTB: quantum of class %X is big. Consider r2q change.\n",
1442 cl->common.classid);
1443 cl->un.leaf.quantum = 200000;
1446 cl->un.leaf.quantum = hopt->quantum;
1447 if ((cl->un.leaf.prio = hopt->prio) >= TC_HTB_NUMPRIO)
1448 cl->un.leaf.prio = TC_HTB_NUMPRIO - 1;
1450 /* backup for htb_parent_to_leaf */
1451 cl->quantum = cl->un.leaf.quantum;
1452 cl->prio = cl->un.leaf.prio;
1455 cl->buffer = hopt->buffer;
1456 cl->cbuffer = hopt->cbuffer;
1458 qdisc_put_rtab(cl->rate);
1461 qdisc_put_rtab(cl->ceil);
1463 sch_tree_unlock(sch);
1465 qdisc_class_hash_grow(sch, &q->clhash);
1467 *arg = (unsigned long)cl;
1472 qdisc_put_rtab(rtab);
1474 qdisc_put_rtab(ctab);
1478 static struct tcf_proto **htb_find_tcf(struct Qdisc *sch, unsigned long arg)
1480 struct htb_sched *q = qdisc_priv(sch);
1481 struct htb_class *cl = (struct htb_class *)arg;
1482 struct tcf_proto **fl = cl ? &cl->filter_list : &q->filter_list;
1487 static unsigned long htb_bind_filter(struct Qdisc *sch, unsigned long parent,
1490 struct htb_class *cl = htb_find(classid, sch);
1492 /*if (cl && !cl->level) return 0;
1493 The line above used to be there to prevent attaching filters to
1494 leaves. But at least tc_index filter uses this just to get class
1495 for other reasons so that we have to allow for it.
1497 19.6.2002 As Werner explained it is ok - bind filter is just
1498 another way to "lock" the class - unlike "get" this lock can
1499 be broken by class during destroy IIUC.
1503 return (unsigned long)cl;
1506 static void htb_unbind_filter(struct Qdisc *sch, unsigned long arg)
1508 struct htb_class *cl = (struct htb_class *)arg;
1514 static void htb_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1516 struct htb_sched *q = qdisc_priv(sch);
1517 struct htb_class *cl;
1518 struct hlist_node *n;
1524 for (i = 0; i < q->clhash.hashsize; i++) {
1525 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode) {
1526 if (arg->count < arg->skip) {
1530 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1539 static const struct Qdisc_class_ops htb_class_ops = {
1542 .qlen_notify = htb_qlen_notify,
1545 .change = htb_change_class,
1546 .delete = htb_delete,
1548 .tcf_chain = htb_find_tcf,
1549 .bind_tcf = htb_bind_filter,
1550 .unbind_tcf = htb_unbind_filter,
1551 .dump = htb_dump_class,
1552 .dump_stats = htb_dump_class_stats,
1555 static struct Qdisc_ops htb_qdisc_ops __read_mostly = {
1557 .cl_ops = &htb_class_ops,
1559 .priv_size = sizeof(struct htb_sched),
1560 .enqueue = htb_enqueue,
1561 .dequeue = htb_dequeue,
1562 .requeue = htb_requeue,
1566 .destroy = htb_destroy,
1567 .change = NULL /* htb_change */,
1569 .owner = THIS_MODULE,
1572 static int __init htb_module_init(void)
1574 return register_qdisc(&htb_qdisc_ops);
1576 static void __exit htb_module_exit(void)
1578 unregister_qdisc(&htb_qdisc_ops);
1581 module_init(htb_module_init)
1582 module_exit(htb_module_exit)
1583 MODULE_LICENSE("GPL");