2 * net/sched/sch_htb.c Hierarchical token bucket, feed tree version
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License, or (at your option) any later version.
9 * Authors: Martin Devera, <devik@cdi.cz>
11 * Credits (in time order) for older HTB versions:
12 * Stef Coene <stef.coene@docum.org>
13 * HTB support at LARTC mailing list
14 * Ondrej Kraus, <krauso@barr.cz>
15 * found missing INIT_QDISC(htb)
16 * Vladimir Smelhaus, Aamer Akhter, Bert Hubert
17 * helped a lot to locate nasty class stall bug
18 * Andi Kleen, Jamal Hadi, Bert Hubert
19 * code review and helpful comments on shaping
20 * Tomasz Wrona, <tw@eter.tym.pl>
21 * created test case so that I was able to fix nasty bug
23 * spotted bug in dequeue code and helped with fix
25 * fixed requeue routine
26 * and many others. thanks.
28 * $Id: sch_htb.c,v 1.25 2003/12/07 11:08:25 devik Exp devik $
30 #include <linux/module.h>
31 #include <asm/uaccess.h>
32 #include <asm/system.h>
33 #include <linux/bitops.h>
34 #include <linux/types.h>
35 #include <linux/kernel.h>
36 #include <linux/sched.h>
37 #include <linux/string.h>
39 #include <linux/socket.h>
40 #include <linux/sockios.h>
42 #include <linux/errno.h>
43 #include <linux/interrupt.h>
44 #include <linux/if_ether.h>
45 #include <linux/inet.h>
46 #include <linux/netdevice.h>
47 #include <linux/etherdevice.h>
48 #include <linux/notifier.h>
50 #include <net/route.h>
51 #include <linux/skbuff.h>
52 #include <linux/list.h>
53 #include <linux/compiler.h>
55 #include <net/pkt_sched.h>
56 #include <linux/rbtree.h>
60 ========================================================================
61 HTB is like TBF with multiple classes. It is also similar to CBQ because
62 it allows to assign priority to each class in hierarchy.
63 In fact it is another implementation of Floyd's formal sharing.
66 Each class is assigned level. Leaf has ALWAYS level 0 and root
67 classes have level TC_HTB_MAXDEPTH-1. Interior nodes has level
68 one less than their parent.
71 #define HTB_HSIZE 16 /* classid hash size */
72 #define HTB_EWMAC 2 /* rate average over HTB_EWMAC*HTB_HSIZE sec */
73 #define HTB_RATECM 1 /* whether to use rate computer */
74 #define HTB_HYSTERESIS 1/* whether to use mode hysteresis for speedup */
75 #define HTB_VER 0x30011 /* major must be matched with number suplied by TC as version */
77 #if HTB_VER >> 16 != TC_HTB_PROTOVER
78 #error "Mismatched sch_htb.c and pkt_sch.h"
81 /* used internaly to keep status of single class */
83 HTB_CANT_SEND, /* class can't send and can't borrow */
84 HTB_MAY_BORROW, /* class can't send but may borrow */
85 HTB_CAN_SEND /* class can send */
88 /* interior & leaf nodes; props specific to leaves are marked L: */
91 /* general class parameters */
93 struct gnet_stats_basic bstats;
94 struct gnet_stats_queue qstats;
95 struct gnet_stats_rate_est rate_est;
96 struct tc_htb_xstats xstats;/* our special stats */
97 int refcnt; /* usage count of this class */
100 /* rate measurement counters */
101 unsigned long rate_bytes,sum_bytes;
102 unsigned long rate_packets,sum_packets;
106 int level; /* our level (see above) */
107 struct htb_class *parent; /* parent class */
108 struct list_head hlist; /* classid hash list item */
109 struct list_head sibling; /* sibling list item */
110 struct list_head children; /* children list */
113 struct htb_class_leaf {
118 int deficit[TC_HTB_MAXDEPTH];
119 struct list_head drop_list;
121 struct htb_class_inner {
122 struct rb_root feed[TC_HTB_NUMPRIO]; /* feed trees */
123 struct rb_node *ptr[TC_HTB_NUMPRIO]; /* current class ptr */
124 /* When class changes from state 1->2 and disconnects from
125 parent's feed then we lost ptr value and start from the
126 first child again. Here we store classid of the
127 last valid ptr (used when ptr is NULL). */
128 u32 last_ptr_id[TC_HTB_NUMPRIO];
131 struct rb_node node[TC_HTB_NUMPRIO]; /* node for self or feed tree */
132 struct rb_node pq_node; /* node for event queue */
133 unsigned long pq_key; /* the same type as jiffies global */
135 int prio_activity; /* for which prios are we active */
136 enum htb_cmode cmode; /* current mode of the class */
138 /* class attached filters */
139 struct tcf_proto *filter_list;
142 int warned; /* only one warning about non work conserving .. */
144 /* token bucket parameters */
145 struct qdisc_rate_table *rate; /* rate table of the class itself */
146 struct qdisc_rate_table *ceil; /* ceiling rate (limits borrows too) */
147 long buffer,cbuffer; /* token bucket depth/rate */
148 psched_tdiff_t mbuffer; /* max wait time */
149 long tokens,ctokens; /* current number of tokens */
150 psched_time_t t_c; /* checkpoint time */
153 /* TODO: maybe compute rate when size is too large .. or drop ? */
154 static __inline__ long L2T(struct htb_class *cl,struct qdisc_rate_table *rate,
157 int slot = size >> rate->rate.cell_log;
162 return rate->data[slot];
167 struct list_head root; /* root classes list */
168 struct list_head hash[HTB_HSIZE]; /* hashed by classid */
169 struct list_head drops[TC_HTB_NUMPRIO]; /* active leaves (for drops) */
171 /* self list - roots of self generating tree */
172 struct rb_root row[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
173 int row_mask[TC_HTB_MAXDEPTH];
174 struct rb_node *ptr[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
175 u32 last_ptr_id[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
177 /* self wait list - roots of wait PQs per row */
178 struct rb_root wait_pq[TC_HTB_MAXDEPTH];
180 /* time of nearest event per level (row) */
181 unsigned long near_ev_cache[TC_HTB_MAXDEPTH];
183 /* cached value of jiffies in dequeue */
184 unsigned long jiffies;
186 /* whether we hit non-work conserving class during this dequeue; we use */
187 int nwc_hit; /* this to disable mindelay complaint in dequeue */
189 int defcls; /* class where unclassified flows go to */
191 /* filters for qdisc itself */
192 struct tcf_proto *filter_list;
195 int rate2quantum; /* quant = rate / rate2quantum */
196 psched_time_t now; /* cached dequeue time */
197 struct timer_list timer; /* send delay timer */
199 struct timer_list rttim; /* rate computer timer */
200 int recmp_bucket; /* which hash bucket to recompute next */
203 /* non shaped skbs; let them go directly thru */
204 struct sk_buff_head direct_queue;
205 int direct_qlen; /* max qlen of above */
210 /* compute hash of size HTB_HSIZE for given handle */
211 static __inline__ int htb_hash(u32 h)
214 #error "Declare new hash for your HTB_HSIZE"
216 h ^= h>>8; /* stolen from cbq_hash */
221 /* find class in global hash table using given handle */
222 static __inline__ struct htb_class *htb_find(u32 handle, struct Qdisc *sch)
224 struct htb_sched *q = qdisc_priv(sch);
226 if (TC_H_MAJ(handle) != sch->handle)
229 list_for_each (p,q->hash+htb_hash(handle)) {
230 struct htb_class *cl = list_entry(p,struct htb_class,hlist);
231 if (cl->classid == handle)
238 * htb_classify - classify a packet into class
240 * It returns NULL if the packet should be dropped or -1 if the packet
241 * should be passed directly thru. In all other cases leaf class is returned.
242 * We allow direct class selection by classid in priority. The we examine
243 * filters in qdisc and in inner nodes (if higher filter points to the inner
244 * node). If we end up with classid MAJOR:0 we enqueue the skb into special
245 * internal fifo (direct). These packets then go directly thru. If we still
246 * have no valid leaf we try to use MAJOR:default leaf. It still unsuccessfull
247 * then finish and return direct queue.
249 #define HTB_DIRECT (struct htb_class*)-1
250 static inline u32 htb_classid(struct htb_class *cl)
252 return (cl && cl != HTB_DIRECT) ? cl->classid : TC_H_UNSPEC;
255 static struct htb_class *htb_classify(struct sk_buff *skb, struct Qdisc *sch, int *qerr)
257 struct htb_sched *q = qdisc_priv(sch);
258 struct htb_class *cl;
259 struct tcf_result res;
260 struct tcf_proto *tcf;
263 /* allow to select class by setting skb->priority to valid classid;
264 note that nfmark can be used too by attaching filter fw with no
266 if (skb->priority == sch->handle)
267 return HTB_DIRECT; /* X:0 (direct flow) selected */
268 if ((cl = htb_find(skb->priority,sch)) != NULL && cl->level == 0)
271 *qerr = NET_XMIT_BYPASS;
272 tcf = q->filter_list;
273 while (tcf && (result = tc_classify(skb, tcf, &res)) >= 0) {
274 #ifdef CONFIG_NET_CLS_ACT
278 *qerr = NET_XMIT_SUCCESS;
282 #elif defined(CONFIG_NET_CLS_POLICE)
283 if (result == TC_POLICE_SHOT)
286 if ((cl = (void*)res.class) == NULL) {
287 if (res.classid == sch->handle)
288 return HTB_DIRECT; /* X:0 (direct flow) */
289 if ((cl = htb_find(res.classid,sch)) == NULL)
290 break; /* filter selected invalid classid */
293 return cl; /* we hit leaf; return it */
295 /* we have got inner class; apply inner filter chain */
296 tcf = cl->filter_list;
298 /* classification failed; try to use default class */
299 cl = htb_find(TC_H_MAKE(TC_H_MAJ(sch->handle),q->defcls),sch);
300 if (!cl || cl->level)
301 return HTB_DIRECT; /* bad default .. this is safe bet */
306 * htb_add_to_id_tree - adds class to the round robin list
308 * Routine adds class to the list (actually tree) sorted by classid.
309 * Make sure that class is not already on such list for given prio.
311 static void htb_add_to_id_tree (struct rb_root *root,
312 struct htb_class *cl,int prio)
314 struct rb_node **p = &root->rb_node, *parent = NULL;
317 struct htb_class *c; parent = *p;
318 c = rb_entry(parent, struct htb_class, node[prio]);
320 if (cl->classid > c->classid)
321 p = &parent->rb_right;
323 p = &parent->rb_left;
325 rb_link_node(&cl->node[prio], parent, p);
326 rb_insert_color(&cl->node[prio], root);
330 * htb_add_to_wait_tree - adds class to the event queue with delay
332 * The class is added to priority event queue to indicate that class will
333 * change its mode in cl->pq_key microseconds. Make sure that class is not
334 * already in the queue.
336 static void htb_add_to_wait_tree (struct htb_sched *q,
337 struct htb_class *cl,long delay)
339 struct rb_node **p = &q->wait_pq[cl->level].rb_node, *parent = NULL;
341 cl->pq_key = q->jiffies + PSCHED_US2JIFFIE(delay);
342 if (cl->pq_key == q->jiffies)
345 /* update the nearest event cache */
346 if (time_after(q->near_ev_cache[cl->level], cl->pq_key))
347 q->near_ev_cache[cl->level] = cl->pq_key;
350 struct htb_class *c; parent = *p;
351 c = rb_entry(parent, struct htb_class, pq_node);
352 if (time_after_eq(cl->pq_key, c->pq_key))
353 p = &parent->rb_right;
355 p = &parent->rb_left;
357 rb_link_node(&cl->pq_node, parent, p);
358 rb_insert_color(&cl->pq_node, &q->wait_pq[cl->level]);
362 * htb_next_rb_node - finds next node in binary tree
364 * When we are past last key we return NULL.
365 * Average complexity is 2 steps per call.
367 static void htb_next_rb_node(struct rb_node **n)
373 * htb_add_class_to_row - add class to its row
375 * The class is added to row at priorities marked in mask.
376 * It does nothing if mask == 0.
378 static inline void htb_add_class_to_row(struct htb_sched *q,
379 struct htb_class *cl,int mask)
381 q->row_mask[cl->level] |= mask;
383 int prio = ffz(~mask);
384 mask &= ~(1 << prio);
385 htb_add_to_id_tree(q->row[cl->level]+prio,cl,prio);
390 * htb_remove_class_from_row - removes class from its row
392 * The class is removed from row at priorities marked in mask.
393 * It does nothing if mask == 0.
395 static __inline__ void htb_remove_class_from_row(struct htb_sched *q,
396 struct htb_class *cl,int mask)
401 int prio = ffz(~mask);
402 mask &= ~(1 << prio);
403 if (q->ptr[cl->level][prio] == cl->node+prio)
404 htb_next_rb_node(q->ptr[cl->level]+prio);
405 rb_erase(cl->node + prio,q->row[cl->level]+prio);
406 if (!q->row[cl->level][prio].rb_node)
409 q->row_mask[cl->level] &= ~m;
413 * htb_activate_prios - creates active classe's feed chain
415 * The class is connected to ancestors and/or appropriate rows
416 * for priorities it is participating on. cl->cmode must be new
417 * (activated) mode. It does nothing if cl->prio_activity == 0.
419 static void htb_activate_prios(struct htb_sched *q,struct htb_class *cl)
421 struct htb_class *p = cl->parent;
422 long m,mask = cl->prio_activity;
424 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
426 m = mask; while (m) {
430 if (p->un.inner.feed[prio].rb_node)
431 /* parent already has its feed in use so that
432 reset bit in mask as parent is already ok */
433 mask &= ~(1 << prio);
435 htb_add_to_id_tree(p->un.inner.feed+prio,cl,prio);
437 p->prio_activity |= mask;
438 cl = p; p = cl->parent;
441 if (cl->cmode == HTB_CAN_SEND && mask)
442 htb_add_class_to_row(q,cl,mask);
446 * htb_deactivate_prios - remove class from feed chain
448 * cl->cmode must represent old mode (before deactivation). It does
449 * nothing if cl->prio_activity == 0. Class is removed from all feed
452 static void htb_deactivate_prios(struct htb_sched *q, struct htb_class *cl)
454 struct htb_class *p = cl->parent;
455 long m,mask = cl->prio_activity;
458 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
464 if (p->un.inner.ptr[prio] == cl->node+prio) {
465 /* we are removing child which is pointed to from
466 parent feed - forget the pointer but remember
468 p->un.inner.last_ptr_id[prio] = cl->classid;
469 p->un.inner.ptr[prio] = NULL;
472 rb_erase(cl->node + prio,p->un.inner.feed + prio);
474 if (!p->un.inner.feed[prio].rb_node)
478 p->prio_activity &= ~mask;
479 cl = p; p = cl->parent;
482 if (cl->cmode == HTB_CAN_SEND && mask)
483 htb_remove_class_from_row(q,cl,mask);
487 * htb_class_mode - computes and returns current class mode
489 * It computes cl's mode at time cl->t_c+diff and returns it. If mode
490 * is not HTB_CAN_SEND then cl->pq_key is updated to time difference
491 * from now to time when cl will change its state.
492 * Also it is worth to note that class mode doesn't change simply
493 * at cl->{c,}tokens == 0 but there can rather be hysteresis of
494 * 0 .. -cl->{c,}buffer range. It is meant to limit number of
495 * mode transitions per time unit. The speed gain is about 1/6.
497 static __inline__ enum htb_cmode
498 htb_class_mode(struct htb_class *cl,long *diff)
502 if ((toks = (cl->ctokens + *diff)) < (
504 cl->cmode != HTB_CANT_SEND ? -cl->cbuffer :
508 return HTB_CANT_SEND;
510 if ((toks = (cl->tokens + *diff)) >= (
512 cl->cmode == HTB_CAN_SEND ? -cl->buffer :
518 return HTB_MAY_BORROW;
522 * htb_change_class_mode - changes classe's mode
524 * This should be the only way how to change classe's mode under normal
525 * cirsumstances. Routine will update feed lists linkage, change mode
526 * and add class to the wait event queue if appropriate. New mode should
527 * be different from old one and cl->pq_key has to be valid if changing
528 * to mode other than HTB_CAN_SEND (see htb_add_to_wait_tree).
531 htb_change_class_mode(struct htb_sched *q, struct htb_class *cl, long *diff)
533 enum htb_cmode new_mode = htb_class_mode(cl,diff);
536 if (new_mode == cl->cmode)
539 if (cl->prio_activity) { /* not necessary: speed optimization */
540 if (cl->cmode != HTB_CANT_SEND)
541 htb_deactivate_prios(q,cl);
542 cl->cmode = new_mode;
543 if (new_mode != HTB_CANT_SEND)
544 htb_activate_prios(q,cl);
546 cl->cmode = new_mode;
550 * htb_activate - inserts leaf cl into appropriate active feeds
552 * Routine learns (new) priority of leaf and activates feed chain
553 * for the prio. It can be called on already active leaf safely.
554 * It also adds leaf into droplist.
556 static __inline__ void htb_activate(struct htb_sched *q,struct htb_class *cl)
558 BUG_TRAP(!cl->level && cl->un.leaf.q && cl->un.leaf.q->q.qlen);
560 if (!cl->prio_activity) {
561 cl->prio_activity = 1 << (cl->un.leaf.aprio = cl->un.leaf.prio);
562 htb_activate_prios(q,cl);
563 list_add_tail(&cl->un.leaf.drop_list,q->drops+cl->un.leaf.aprio);
568 * htb_deactivate - remove leaf cl from active feeds
570 * Make sure that leaf is active. In the other words it can't be called
571 * with non-active leaf. It also removes class from the drop list.
573 static __inline__ void
574 htb_deactivate(struct htb_sched *q,struct htb_class *cl)
576 BUG_TRAP(cl->prio_activity);
578 htb_deactivate_prios(q,cl);
579 cl->prio_activity = 0;
580 list_del_init(&cl->un.leaf.drop_list);
583 static int htb_enqueue(struct sk_buff *skb, struct Qdisc *sch)
586 struct htb_sched *q = qdisc_priv(sch);
587 struct htb_class *cl = htb_classify(skb,sch,&ret);
589 if (cl == HTB_DIRECT) {
590 /* enqueue to helper queue */
591 if (q->direct_queue.qlen < q->direct_qlen) {
592 __skb_queue_tail(&q->direct_queue, skb);
597 return NET_XMIT_DROP;
599 #ifdef CONFIG_NET_CLS_ACT
601 if (ret == NET_XMIT_BYPASS)
606 } else if (cl->un.leaf.q->enqueue(skb, cl->un.leaf.q) != NET_XMIT_SUCCESS) {
609 return NET_XMIT_DROP;
611 cl->bstats.packets++; cl->bstats.bytes += skb->len;
616 sch->bstats.packets++; sch->bstats.bytes += skb->len;
617 return NET_XMIT_SUCCESS;
620 /* TODO: requeuing packet charges it to policers again !! */
621 static int htb_requeue(struct sk_buff *skb, struct Qdisc *sch)
623 struct htb_sched *q = qdisc_priv(sch);
624 int ret = NET_XMIT_SUCCESS;
625 struct htb_class *cl = htb_classify(skb,sch, &ret);
626 struct sk_buff *tskb;
628 if (cl == HTB_DIRECT || !cl) {
629 /* enqueue to helper queue */
630 if (q->direct_queue.qlen < q->direct_qlen && cl) {
631 __skb_queue_head(&q->direct_queue, skb);
633 __skb_queue_head(&q->direct_queue, skb);
634 tskb = __skb_dequeue_tail(&q->direct_queue);
639 } else if (cl->un.leaf.q->ops->requeue(skb, cl->un.leaf.q) != NET_XMIT_SUCCESS) {
642 return NET_XMIT_DROP;
647 sch->qstats.requeues++;
648 return NET_XMIT_SUCCESS;
651 static void htb_timer(unsigned long arg)
653 struct Qdisc *sch = (struct Qdisc*)arg;
654 sch->flags &= ~TCQ_F_THROTTLED;
656 netif_schedule(sch->dev);
660 #define RT_GEN(D,R) R+=D-(R/HTB_EWMAC);D=0
661 static void htb_rate_timer(unsigned long arg)
663 struct Qdisc *sch = (struct Qdisc*)arg;
664 struct htb_sched *q = qdisc_priv(sch);
667 /* lock queue so that we can muck with it */
668 spin_lock_bh(&sch->dev->queue_lock);
670 q->rttim.expires = jiffies + HZ;
671 add_timer(&q->rttim);
673 /* scan and recompute one bucket at time */
674 if (++q->recmp_bucket >= HTB_HSIZE)
676 list_for_each (p,q->hash+q->recmp_bucket) {
677 struct htb_class *cl = list_entry(p,struct htb_class,hlist);
679 RT_GEN (cl->sum_bytes,cl->rate_bytes);
680 RT_GEN (cl->sum_packets,cl->rate_packets);
682 spin_unlock_bh(&sch->dev->queue_lock);
687 * htb_charge_class - charges amount "bytes" to leaf and ancestors
689 * Routine assumes that packet "bytes" long was dequeued from leaf cl
690 * borrowing from "level". It accounts bytes to ceil leaky bucket for
691 * leaf and all ancestors and to rate bucket for ancestors at levels
692 * "level" and higher. It also handles possible change of mode resulting
693 * from the update. Note that mode can also increase here (MAY_BORROW to
694 * CAN_SEND) because we can use more precise clock that event queue here.
695 * In such case we remove class from event queue first.
697 static void htb_charge_class(struct htb_sched *q,struct htb_class *cl,
701 enum htb_cmode old_mode;
703 #define HTB_ACCNT(T,B,R) toks = diff + cl->T; \
704 if (toks > cl->B) toks = cl->B; \
705 toks -= L2T(cl, cl->R, bytes); \
706 if (toks <= -cl->mbuffer) toks = 1-cl->mbuffer; \
710 diff = PSCHED_TDIFF_SAFE(q->now, cl->t_c, (u32)cl->mbuffer);
711 if (cl->level >= level) {
712 if (cl->level == level) cl->xstats.lends++;
713 HTB_ACCNT (tokens,buffer,rate);
715 cl->xstats.borrows++;
716 cl->tokens += diff; /* we moved t_c; update tokens */
718 HTB_ACCNT (ctokens,cbuffer,ceil);
721 old_mode = cl->cmode; diff = 0;
722 htb_change_class_mode(q,cl,&diff);
723 if (old_mode != cl->cmode) {
724 if (old_mode != HTB_CAN_SEND)
725 rb_erase(&cl->pq_node,q->wait_pq+cl->level);
726 if (cl->cmode != HTB_CAN_SEND)
727 htb_add_to_wait_tree (q,cl,diff);
731 /* update rate counters */
732 cl->sum_bytes += bytes; cl->sum_packets++;
735 /* update byte stats except for leaves which are already updated */
737 cl->bstats.bytes += bytes;
738 cl->bstats.packets++;
745 * htb_do_events - make mode changes to classes at the level
747 * Scans event queue for pending events and applies them. Returns jiffies to
748 * next pending event (0 for no event in pq).
749 * Note: Aplied are events whose have cl->pq_key <= jiffies.
751 static long htb_do_events(struct htb_sched *q,int level)
755 for (i = 0; i < 500; i++) {
756 struct htb_class *cl;
758 struct rb_node *p = q->wait_pq[level].rb_node;
760 while (p->rb_left) p = p->rb_left;
762 cl = rb_entry(p, struct htb_class, pq_node);
763 if (time_after(cl->pq_key, q->jiffies)) {
764 return cl->pq_key - q->jiffies;
766 rb_erase(p,q->wait_pq+level);
767 diff = PSCHED_TDIFF_SAFE(q->now, cl->t_c, (u32)cl->mbuffer);
768 htb_change_class_mode(q,cl,&diff);
769 if (cl->cmode != HTB_CAN_SEND)
770 htb_add_to_wait_tree (q,cl,diff);
773 printk(KERN_WARNING "htb: too many events !\n");
777 /* Returns class->node+prio from id-tree where classe's id is >= id. NULL
778 is no such one exists. */
779 static struct rb_node *
780 htb_id_find_next_upper(int prio,struct rb_node *n,u32 id)
782 struct rb_node *r = NULL;
784 struct htb_class *cl = rb_entry(n,struct htb_class,node[prio]);
785 if (id == cl->classid) return n;
787 if (id > cl->classid) {
798 * htb_lookup_leaf - returns next leaf class in DRR order
800 * Find leaf where current feed pointers points to.
802 static struct htb_class *
803 htb_lookup_leaf(struct rb_root *tree,int prio,struct rb_node **pptr,u32 *pid)
807 struct rb_node *root;
808 struct rb_node **pptr;
810 } stk[TC_HTB_MAXDEPTH],*sp = stk;
812 BUG_TRAP(tree->rb_node);
813 sp->root = tree->rb_node;
817 for (i = 0; i < 65535; i++) {
818 if (!*sp->pptr && *sp->pid) {
819 /* ptr was invalidated but id is valid - try to recover
820 the original or next ptr */
821 *sp->pptr = htb_id_find_next_upper(prio,sp->root,*sp->pid);
823 *sp->pid = 0; /* ptr is valid now so that remove this hint as it
824 can become out of date quickly */
825 if (!*sp->pptr) { /* we are at right end; rewind & go up */
826 *sp->pptr = sp->root;
827 while ((*sp->pptr)->rb_left)
828 *sp->pptr = (*sp->pptr)->rb_left;
831 BUG_TRAP(*sp->pptr); if(!*sp->pptr) return NULL;
832 htb_next_rb_node (sp->pptr);
835 struct htb_class *cl;
836 cl = rb_entry(*sp->pptr,struct htb_class,node[prio]);
839 (++sp)->root = cl->un.inner.feed[prio].rb_node;
840 sp->pptr = cl->un.inner.ptr+prio;
841 sp->pid = cl->un.inner.last_ptr_id+prio;
848 /* dequeues packet at given priority and level; call only if
849 you are sure that there is active class at prio/level */
850 static struct sk_buff *
851 htb_dequeue_tree(struct htb_sched *q,int prio,int level)
853 struct sk_buff *skb = NULL;
854 struct htb_class *cl,*start;
855 /* look initial class up in the row */
856 start = cl = htb_lookup_leaf (q->row[level]+prio,prio,
857 q->ptr[level]+prio,q->last_ptr_id[level]+prio);
862 if (!cl) return NULL;
864 /* class can be empty - it is unlikely but can be true if leaf
865 qdisc drops packets in enqueue routine or if someone used
866 graft operation on the leaf since last dequeue;
867 simply deactivate and skip such class */
868 if (unlikely(cl->un.leaf.q->q.qlen == 0)) {
869 struct htb_class *next;
870 htb_deactivate(q,cl);
872 /* row/level might become empty */
873 if ((q->row_mask[level] & (1 << prio)) == 0)
876 next = htb_lookup_leaf (q->row[level]+prio,
877 prio,q->ptr[level]+prio,q->last_ptr_id[level]+prio);
879 if (cl == start) /* fix start if we just deleted it */
885 if (likely((skb = cl->un.leaf.q->dequeue(cl->un.leaf.q)) != NULL))
888 printk(KERN_WARNING "htb: class %X isn't work conserving ?!\n",cl->classid);
892 htb_next_rb_node((level?cl->parent->un.inner.ptr:q->ptr[0])+prio);
893 cl = htb_lookup_leaf (q->row[level]+prio,prio,q->ptr[level]+prio,
894 q->last_ptr_id[level]+prio);
896 } while (cl != start);
898 if (likely(skb != NULL)) {
899 if ((cl->un.leaf.deficit[level] -= skb->len) < 0) {
900 cl->un.leaf.deficit[level] += cl->un.leaf.quantum;
901 htb_next_rb_node((level?cl->parent->un.inner.ptr:q->ptr[0])+prio);
903 /* this used to be after charge_class but this constelation
904 gives us slightly better performance */
905 if (!cl->un.leaf.q->q.qlen)
906 htb_deactivate (q,cl);
907 htb_charge_class (q,cl,level,skb->len);
912 static void htb_delay_by(struct Qdisc *sch,long delay)
914 struct htb_sched *q = qdisc_priv(sch);
915 if (delay <= 0) delay = 1;
916 if (unlikely(delay > 5*HZ)) {
918 printk(KERN_INFO "HTB delay %ld > 5sec\n", delay);
921 /* why don't use jiffies here ? because expires can be in past */
922 mod_timer(&q->timer, q->jiffies + delay);
923 sch->flags |= TCQ_F_THROTTLED;
924 sch->qstats.overlimits++;
927 static struct sk_buff *htb_dequeue(struct Qdisc *sch)
929 struct sk_buff *skb = NULL;
930 struct htb_sched *q = qdisc_priv(sch);
934 q->jiffies = jiffies;
936 /* try to dequeue direct packets as high prio (!) to minimize cpu work */
937 if ((skb = __skb_dequeue(&q->direct_queue)) != NULL) {
938 sch->flags &= ~TCQ_F_THROTTLED;
943 if (!sch->q.qlen) goto fin;
944 PSCHED_GET_TIME(q->now);
946 min_delay = LONG_MAX;
948 for (level = 0; level < TC_HTB_MAXDEPTH; level++) {
949 /* common case optimization - skip event handler quickly */
952 if (time_after_eq(q->jiffies, q->near_ev_cache[level])) {
953 delay = htb_do_events(q,level);
954 q->near_ev_cache[level] = q->jiffies + (delay ? delay : HZ);
956 delay = q->near_ev_cache[level] - q->jiffies;
958 if (delay && min_delay > delay)
960 m = ~q->row_mask[level];
961 while (m != (int)(-1)) {
964 skb = htb_dequeue_tree(q,prio,level);
965 if (likely(skb != NULL)) {
967 sch->flags &= ~TCQ_F_THROTTLED;
972 htb_delay_by (sch,min_delay > 5*HZ ? 5*HZ : min_delay);
977 /* try to drop from each class (by prio) until one succeed */
978 static unsigned int htb_drop(struct Qdisc* sch)
980 struct htb_sched *q = qdisc_priv(sch);
983 for (prio = TC_HTB_NUMPRIO - 1; prio >= 0; prio--) {
985 list_for_each (p,q->drops+prio) {
986 struct htb_class *cl = list_entry(p, struct htb_class,
989 if (cl->un.leaf.q->ops->drop &&
990 (len = cl->un.leaf.q->ops->drop(cl->un.leaf.q))) {
992 if (!cl->un.leaf.q->q.qlen)
993 htb_deactivate (q,cl);
1001 /* reset all classes */
1002 /* always caled under BH & queue lock */
1003 static void htb_reset(struct Qdisc* sch)
1005 struct htb_sched *q = qdisc_priv(sch);
1008 for (i = 0; i < HTB_HSIZE; i++) {
1009 struct list_head *p;
1010 list_for_each (p,q->hash+i) {
1011 struct htb_class *cl = list_entry(p,struct htb_class,hlist);
1013 memset(&cl->un.inner,0,sizeof(cl->un.inner));
1016 qdisc_reset(cl->un.leaf.q);
1017 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
1019 cl->prio_activity = 0;
1020 cl->cmode = HTB_CAN_SEND;
1024 sch->flags &= ~TCQ_F_THROTTLED;
1025 del_timer(&q->timer);
1026 __skb_queue_purge(&q->direct_queue);
1028 memset(q->row,0,sizeof(q->row));
1029 memset(q->row_mask,0,sizeof(q->row_mask));
1030 memset(q->wait_pq,0,sizeof(q->wait_pq));
1031 memset(q->ptr,0,sizeof(q->ptr));
1032 for (i = 0; i < TC_HTB_NUMPRIO; i++)
1033 INIT_LIST_HEAD(q->drops+i);
1036 static int htb_init(struct Qdisc *sch, struct rtattr *opt)
1038 struct htb_sched *q = qdisc_priv(sch);
1039 struct rtattr *tb[TCA_HTB_INIT];
1040 struct tc_htb_glob *gopt;
1042 if (!opt || rtattr_parse_nested(tb, TCA_HTB_INIT, opt) ||
1043 tb[TCA_HTB_INIT-1] == NULL ||
1044 RTA_PAYLOAD(tb[TCA_HTB_INIT-1]) < sizeof(*gopt)) {
1045 printk(KERN_ERR "HTB: hey probably you have bad tc tool ?\n");
1048 gopt = RTA_DATA(tb[TCA_HTB_INIT-1]);
1049 if (gopt->version != HTB_VER >> 16) {
1050 printk(KERN_ERR "HTB: need tc/htb version %d (minor is %d), you have %d\n",
1051 HTB_VER >> 16,HTB_VER & 0xffff,gopt->version);
1055 INIT_LIST_HEAD(&q->root);
1056 for (i = 0; i < HTB_HSIZE; i++)
1057 INIT_LIST_HEAD(q->hash+i);
1058 for (i = 0; i < TC_HTB_NUMPRIO; i++)
1059 INIT_LIST_HEAD(q->drops+i);
1061 init_timer(&q->timer);
1062 skb_queue_head_init(&q->direct_queue);
1064 q->direct_qlen = sch->dev->tx_queue_len;
1065 if (q->direct_qlen < 2) /* some devices have zero tx_queue_len */
1067 q->timer.function = htb_timer;
1068 q->timer.data = (unsigned long)sch;
1071 init_timer(&q->rttim);
1072 q->rttim.function = htb_rate_timer;
1073 q->rttim.data = (unsigned long)sch;
1074 q->rttim.expires = jiffies + HZ;
1075 add_timer(&q->rttim);
1077 if ((q->rate2quantum = gopt->rate2quantum) < 1)
1078 q->rate2quantum = 1;
1079 q->defcls = gopt->defcls;
1084 static int htb_dump(struct Qdisc *sch, struct sk_buff *skb)
1086 struct htb_sched *q = qdisc_priv(sch);
1087 unsigned char *b = skb->tail;
1089 struct tc_htb_glob gopt;
1090 spin_lock_bh(&sch->dev->queue_lock);
1091 gopt.direct_pkts = q->direct_pkts;
1093 gopt.version = HTB_VER;
1094 gopt.rate2quantum = q->rate2quantum;
1095 gopt.defcls = q->defcls;
1097 rta = (struct rtattr*)b;
1098 RTA_PUT(skb, TCA_OPTIONS, 0, NULL);
1099 RTA_PUT(skb, TCA_HTB_INIT, sizeof(gopt), &gopt);
1100 rta->rta_len = skb->tail - b;
1101 spin_unlock_bh(&sch->dev->queue_lock);
1104 spin_unlock_bh(&sch->dev->queue_lock);
1105 skb_trim(skb, skb->tail - skb->data);
1109 static int htb_dump_class(struct Qdisc *sch, unsigned long arg,
1110 struct sk_buff *skb, struct tcmsg *tcm)
1112 struct htb_class *cl = (struct htb_class*)arg;
1113 unsigned char *b = skb->tail;
1115 struct tc_htb_opt opt;
1117 spin_lock_bh(&sch->dev->queue_lock);
1118 tcm->tcm_parent = cl->parent ? cl->parent->classid : TC_H_ROOT;
1119 tcm->tcm_handle = cl->classid;
1120 if (!cl->level && cl->un.leaf.q)
1121 tcm->tcm_info = cl->un.leaf.q->handle;
1123 rta = (struct rtattr*)b;
1124 RTA_PUT(skb, TCA_OPTIONS, 0, NULL);
1126 memset (&opt,0,sizeof(opt));
1128 opt.rate = cl->rate->rate; opt.buffer = cl->buffer;
1129 opt.ceil = cl->ceil->rate; opt.cbuffer = cl->cbuffer;
1130 opt.quantum = cl->un.leaf.quantum; opt.prio = cl->un.leaf.prio;
1131 opt.level = cl->level;
1132 RTA_PUT(skb, TCA_HTB_PARMS, sizeof(opt), &opt);
1133 rta->rta_len = skb->tail - b;
1134 spin_unlock_bh(&sch->dev->queue_lock);
1137 spin_unlock_bh(&sch->dev->queue_lock);
1138 skb_trim(skb, b - skb->data);
1143 htb_dump_class_stats(struct Qdisc *sch, unsigned long arg,
1144 struct gnet_dump *d)
1146 struct htb_class *cl = (struct htb_class*)arg;
1149 cl->rate_est.bps = cl->rate_bytes/(HTB_EWMAC*HTB_HSIZE);
1150 cl->rate_est.pps = cl->rate_packets/(HTB_EWMAC*HTB_HSIZE);
1153 if (!cl->level && cl->un.leaf.q)
1154 cl->qstats.qlen = cl->un.leaf.q->q.qlen;
1155 cl->xstats.tokens = cl->tokens;
1156 cl->xstats.ctokens = cl->ctokens;
1158 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
1159 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1160 gnet_stats_copy_queue(d, &cl->qstats) < 0)
1163 return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1166 static int htb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1169 struct htb_class *cl = (struct htb_class*)arg;
1171 if (cl && !cl->level) {
1172 if (new == NULL && (new = qdisc_create_dflt(sch->dev,
1173 &pfifo_qdisc_ops)) == NULL)
1176 if ((*old = xchg(&cl->un.leaf.q, new)) != NULL) {
1177 if (cl->prio_activity)
1178 htb_deactivate (qdisc_priv(sch),cl);
1180 /* TODO: is it correct ? Why CBQ doesn't do it ? */
1181 sch->q.qlen -= (*old)->q.qlen;
1184 sch_tree_unlock(sch);
1190 static struct Qdisc * htb_leaf(struct Qdisc *sch, unsigned long arg)
1192 struct htb_class *cl = (struct htb_class*)arg;
1193 return (cl && !cl->level) ? cl->un.leaf.q : NULL;
1196 static unsigned long htb_get(struct Qdisc *sch, u32 classid)
1198 struct htb_class *cl = htb_find(classid,sch);
1201 return (unsigned long)cl;
1204 static void htb_destroy_filters(struct tcf_proto **fl)
1206 struct tcf_proto *tp;
1208 while ((tp = *fl) != NULL) {
1214 static void htb_destroy_class(struct Qdisc* sch,struct htb_class *cl)
1216 struct htb_sched *q = qdisc_priv(sch);
1218 BUG_TRAP(cl->un.leaf.q);
1219 sch->q.qlen -= cl->un.leaf.q->q.qlen;
1220 qdisc_destroy(cl->un.leaf.q);
1222 qdisc_put_rtab(cl->rate);
1223 qdisc_put_rtab(cl->ceil);
1225 htb_destroy_filters (&cl->filter_list);
1227 while (!list_empty(&cl->children))
1228 htb_destroy_class (sch,list_entry(cl->children.next,
1229 struct htb_class,sibling));
1231 /* note: this delete may happen twice (see htb_delete) */
1232 list_del(&cl->hlist);
1233 list_del(&cl->sibling);
1235 if (cl->prio_activity)
1236 htb_deactivate (q,cl);
1238 if (cl->cmode != HTB_CAN_SEND)
1239 rb_erase(&cl->pq_node,q->wait_pq+cl->level);
1244 /* always caled under BH & queue lock */
1245 static void htb_destroy(struct Qdisc* sch)
1247 struct htb_sched *q = qdisc_priv(sch);
1249 del_timer_sync (&q->timer);
1251 del_timer_sync (&q->rttim);
1253 /* This line used to be after htb_destroy_class call below
1254 and surprisingly it worked in 2.4. But it must precede it
1255 because filter need its target class alive to be able to call
1256 unbind_filter on it (without Oops). */
1257 htb_destroy_filters(&q->filter_list);
1259 while (!list_empty(&q->root))
1260 htb_destroy_class (sch,list_entry(q->root.next,
1261 struct htb_class,sibling));
1263 __skb_queue_purge(&q->direct_queue);
1266 static int htb_delete(struct Qdisc *sch, unsigned long arg)
1268 struct htb_sched *q = qdisc_priv(sch);
1269 struct htb_class *cl = (struct htb_class*)arg;
1271 // TODO: why don't allow to delete subtree ? references ? does
1272 // tc subsys quarantee us that in htb_destroy it holds no class
1273 // refs so that we can remove children safely there ?
1274 if (!list_empty(&cl->children) || cl->filter_cnt)
1279 /* delete from hash and active; remainder in destroy_class */
1280 list_del_init(&cl->hlist);
1281 if (cl->prio_activity)
1282 htb_deactivate (q,cl);
1284 if (--cl->refcnt == 0)
1285 htb_destroy_class(sch,cl);
1287 sch_tree_unlock(sch);
1291 static void htb_put(struct Qdisc *sch, unsigned long arg)
1293 struct htb_class *cl = (struct htb_class*)arg;
1295 if (--cl->refcnt == 0)
1296 htb_destroy_class(sch,cl);
1299 static int htb_change_class(struct Qdisc *sch, u32 classid,
1300 u32 parentid, struct rtattr **tca, unsigned long *arg)
1303 struct htb_sched *q = qdisc_priv(sch);
1304 struct htb_class *cl = (struct htb_class*)*arg,*parent;
1305 struct rtattr *opt = tca[TCA_OPTIONS-1];
1306 struct qdisc_rate_table *rtab = NULL, *ctab = NULL;
1307 struct rtattr *tb[TCA_HTB_RTAB];
1308 struct tc_htb_opt *hopt;
1310 /* extract all subattrs from opt attr */
1311 if (!opt || rtattr_parse_nested(tb, TCA_HTB_RTAB, opt) ||
1312 tb[TCA_HTB_PARMS-1] == NULL ||
1313 RTA_PAYLOAD(tb[TCA_HTB_PARMS-1]) < sizeof(*hopt))
1316 parent = parentid == TC_H_ROOT ? NULL : htb_find (parentid,sch);
1318 hopt = RTA_DATA(tb[TCA_HTB_PARMS-1]);
1320 rtab = qdisc_get_rtab(&hopt->rate, tb[TCA_HTB_RTAB-1]);
1321 ctab = qdisc_get_rtab(&hopt->ceil, tb[TCA_HTB_CTAB-1]);
1322 if (!rtab || !ctab) goto failure;
1324 if (!cl) { /* new class */
1325 struct Qdisc *new_q;
1326 /* check for valid classid */
1327 if (!classid || TC_H_MAJ(classid^sch->handle) || htb_find(classid,sch))
1330 /* check maximal depth */
1331 if (parent && parent->parent && parent->parent->level < 2) {
1332 printk(KERN_ERR "htb: tree is too deep\n");
1336 if ((cl = kzalloc(sizeof(*cl), GFP_KERNEL)) == NULL)
1340 INIT_LIST_HEAD(&cl->sibling);
1341 INIT_LIST_HEAD(&cl->hlist);
1342 INIT_LIST_HEAD(&cl->children);
1343 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
1345 /* create leaf qdisc early because it uses kmalloc(GFP_KERNEL)
1346 so that can't be used inside of sch_tree_lock
1347 -- thanks to Karlis Peisenieks */
1348 new_q = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops);
1350 if (parent && !parent->level) {
1351 /* turn parent into inner node */
1352 sch->q.qlen -= parent->un.leaf.q->q.qlen;
1353 qdisc_destroy (parent->un.leaf.q);
1354 if (parent->prio_activity)
1355 htb_deactivate (q,parent);
1357 /* remove from evt list because of level change */
1358 if (parent->cmode != HTB_CAN_SEND) {
1359 rb_erase(&parent->pq_node,q->wait_pq);
1360 parent->cmode = HTB_CAN_SEND;
1362 parent->level = (parent->parent ? parent->parent->level
1363 : TC_HTB_MAXDEPTH) - 1;
1364 memset (&parent->un.inner,0,sizeof(parent->un.inner));
1366 /* leaf (we) needs elementary qdisc */
1367 cl->un.leaf.q = new_q ? new_q : &noop_qdisc;
1369 cl->classid = classid; cl->parent = parent;
1371 /* set class to be in HTB_CAN_SEND state */
1372 cl->tokens = hopt->buffer;
1373 cl->ctokens = hopt->cbuffer;
1374 cl->mbuffer = PSCHED_JIFFIE2US(HZ*60); /* 1min */
1375 PSCHED_GET_TIME(cl->t_c);
1376 cl->cmode = HTB_CAN_SEND;
1378 /* attach to the hash list and parent's family */
1379 list_add_tail(&cl->hlist, q->hash+htb_hash(classid));
1380 list_add_tail(&cl->sibling, parent ? &parent->children : &q->root);
1381 } else sch_tree_lock(sch);
1383 /* it used to be a nasty bug here, we have to check that node
1384 is really leaf before changing cl->un.leaf ! */
1386 cl->un.leaf.quantum = rtab->rate.rate / q->rate2quantum;
1387 if (!hopt->quantum && cl->un.leaf.quantum < 1000) {
1388 printk(KERN_WARNING "HTB: quantum of class %X is small. Consider r2q change.\n", cl->classid);
1389 cl->un.leaf.quantum = 1000;
1391 if (!hopt->quantum && cl->un.leaf.quantum > 200000) {
1392 printk(KERN_WARNING "HTB: quantum of class %X is big. Consider r2q change.\n", cl->classid);
1393 cl->un.leaf.quantum = 200000;
1396 cl->un.leaf.quantum = hopt->quantum;
1397 if ((cl->un.leaf.prio = hopt->prio) >= TC_HTB_NUMPRIO)
1398 cl->un.leaf.prio = TC_HTB_NUMPRIO - 1;
1401 cl->buffer = hopt->buffer;
1402 cl->cbuffer = hopt->cbuffer;
1403 if (cl->rate) qdisc_put_rtab(cl->rate); cl->rate = rtab;
1404 if (cl->ceil) qdisc_put_rtab(cl->ceil); cl->ceil = ctab;
1405 sch_tree_unlock(sch);
1407 *arg = (unsigned long)cl;
1411 if (rtab) qdisc_put_rtab(rtab);
1412 if (ctab) qdisc_put_rtab(ctab);
1416 static struct tcf_proto **htb_find_tcf(struct Qdisc *sch, unsigned long arg)
1418 struct htb_sched *q = qdisc_priv(sch);
1419 struct htb_class *cl = (struct htb_class *)arg;
1420 struct tcf_proto **fl = cl ? &cl->filter_list : &q->filter_list;
1425 static unsigned long htb_bind_filter(struct Qdisc *sch, unsigned long parent,
1428 struct htb_sched *q = qdisc_priv(sch);
1429 struct htb_class *cl = htb_find (classid,sch);
1431 /*if (cl && !cl->level) return 0;
1432 The line above used to be there to prevent attaching filters to
1433 leaves. But at least tc_index filter uses this just to get class
1434 for other reasons so that we have to allow for it.
1436 19.6.2002 As Werner explained it is ok - bind filter is just
1437 another way to "lock" the class - unlike "get" this lock can
1438 be broken by class during destroy IIUC.
1444 return (unsigned long)cl;
1447 static void htb_unbind_filter(struct Qdisc *sch, unsigned long arg)
1449 struct htb_sched *q = qdisc_priv(sch);
1450 struct htb_class *cl = (struct htb_class *)arg;
1458 static void htb_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1460 struct htb_sched *q = qdisc_priv(sch);
1466 for (i = 0; i < HTB_HSIZE; i++) {
1467 struct list_head *p;
1468 list_for_each (p,q->hash+i) {
1469 struct htb_class *cl = list_entry(p,struct htb_class,hlist);
1470 if (arg->count < arg->skip) {
1474 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1483 static struct Qdisc_class_ops htb_class_ops = {
1488 .change = htb_change_class,
1489 .delete = htb_delete,
1491 .tcf_chain = htb_find_tcf,
1492 .bind_tcf = htb_bind_filter,
1493 .unbind_tcf = htb_unbind_filter,
1494 .dump = htb_dump_class,
1495 .dump_stats = htb_dump_class_stats,
1498 static struct Qdisc_ops htb_qdisc_ops = {
1500 .cl_ops = &htb_class_ops,
1502 .priv_size = sizeof(struct htb_sched),
1503 .enqueue = htb_enqueue,
1504 .dequeue = htb_dequeue,
1505 .requeue = htb_requeue,
1509 .destroy = htb_destroy,
1510 .change = NULL /* htb_change */,
1512 .owner = THIS_MODULE,
1515 static int __init htb_module_init(void)
1517 return register_qdisc(&htb_qdisc_ops);
1519 static void __exit htb_module_exit(void)
1521 unregister_qdisc(&htb_qdisc_ops);
1523 module_init(htb_module_init)
1524 module_exit(htb_module_exit)
1525 MODULE_LICENSE("GPL");