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 <linux/workqueue.h>
39 #include <linux/slab.h>
40 #include <net/netlink.h>
41 #include <net/pkt_sched.h>
45 ========================================================================
46 HTB is like TBF with multiple classes. It is also similar to CBQ because
47 it allows to assign priority to each class in hierarchy.
48 In fact it is another implementation of Floyd's formal sharing.
51 Each class is assigned level. Leaf has ALWAYS level 0 and root
52 classes have level TC_HTB_MAXDEPTH-1. Interior nodes has level
53 one less than their parent.
56 static int htb_hysteresis __read_mostly = 0; /* whether to use mode hysteresis for speedup */
57 #define HTB_VER 0x30011 /* major must be matched with number suplied by TC as version */
59 #if HTB_VER >> 16 != TC_HTB_PROTOVER
60 #error "Mismatched sch_htb.c and pkt_sch.h"
63 /* Module parameter and sysfs export */
64 module_param (htb_hysteresis, int, 0640);
65 MODULE_PARM_DESC(htb_hysteresis, "Hysteresis mode, less CPU load, less accurate");
67 /* used internaly to keep status of single class */
69 HTB_CANT_SEND, /* class can't send and can't borrow */
70 HTB_MAY_BORROW, /* class can't send but may borrow */
71 HTB_CAN_SEND /* class can send */
74 /* interior & leaf nodes; props specific to leaves are marked L: */
76 struct Qdisc_class_common common;
77 /* general class parameters */
78 struct gnet_stats_basic_packed bstats;
79 struct gnet_stats_queue qstats;
80 struct gnet_stats_rate_est rate_est;
81 struct tc_htb_xstats xstats; /* our special stats */
82 int refcnt; /* usage count of this class */
85 int level; /* our level (see above) */
86 unsigned int children;
87 struct htb_class *parent; /* parent class */
89 u32 prio; /* these two are used only by leaves... */
90 int quantum; /* but stored for parent-to-leaf return */
93 struct htb_class_leaf {
95 int deficit[TC_HTB_MAXDEPTH];
96 struct list_head drop_list;
98 struct htb_class_inner {
99 struct rb_root feed[TC_HTB_NUMPRIO]; /* feed trees */
100 struct rb_node *ptr[TC_HTB_NUMPRIO]; /* current class ptr */
101 /* When class changes from state 1->2 and disconnects from
102 * parent's feed then we lost ptr value and start from the
103 * first child again. Here we store classid of the
104 * last valid ptr (used when ptr is NULL).
106 u32 last_ptr_id[TC_HTB_NUMPRIO];
109 struct rb_node node[TC_HTB_NUMPRIO]; /* node for self or feed tree */
110 struct rb_node pq_node; /* node for event queue */
111 psched_time_t pq_key;
113 int prio_activity; /* for which prios are we active */
114 enum htb_cmode cmode; /* current mode of the class */
116 /* class attached filters */
117 struct tcf_proto *filter_list;
120 /* token bucket parameters */
121 struct qdisc_rate_table *rate; /* rate table of the class itself */
122 struct qdisc_rate_table *ceil; /* ceiling rate (limits borrows too) */
123 long buffer, cbuffer; /* token bucket depth/rate */
124 psched_tdiff_t mbuffer; /* max wait time */
125 long tokens, ctokens; /* current number of tokens */
126 psched_time_t t_c; /* checkpoint time */
130 struct Qdisc_class_hash clhash;
131 struct list_head drops[TC_HTB_NUMPRIO];/* active leaves (for drops) */
133 /* self list - roots of self generating tree */
134 struct rb_root row[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
135 int row_mask[TC_HTB_MAXDEPTH];
136 struct rb_node *ptr[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
137 u32 last_ptr_id[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
139 /* self wait list - roots of wait PQs per row */
140 struct rb_root wait_pq[TC_HTB_MAXDEPTH];
142 /* time of nearest event per level (row) */
143 psched_time_t near_ev_cache[TC_HTB_MAXDEPTH];
145 int defcls; /* class where unclassified flows go to */
147 /* filters for qdisc itself */
148 struct tcf_proto *filter_list;
150 int rate2quantum; /* quant = rate / rate2quantum */
151 psched_time_t now; /* cached dequeue time */
152 struct qdisc_watchdog watchdog;
154 /* non shaped skbs; let them go directly thru */
155 struct sk_buff_head direct_queue;
156 int direct_qlen; /* max qlen of above */
160 #define HTB_WARN_TOOMANYEVENTS 0x1
161 unsigned int warned; /* only one warning */
162 struct work_struct work;
165 /* find class in global hash table using given handle */
166 static inline struct htb_class *htb_find(u32 handle, struct Qdisc *sch)
168 struct htb_sched *q = qdisc_priv(sch);
169 struct Qdisc_class_common *clc;
171 clc = qdisc_class_find(&q->clhash, handle);
174 return container_of(clc, struct htb_class, common);
178 * htb_classify - classify a packet into class
180 * It returns NULL if the packet should be dropped or -1 if the packet
181 * should be passed directly thru. In all other cases leaf class is returned.
182 * We allow direct class selection by classid in priority. The we examine
183 * filters in qdisc and in inner nodes (if higher filter points to the inner
184 * node). If we end up with classid MAJOR:0 we enqueue the skb into special
185 * internal fifo (direct). These packets then go directly thru. If we still
186 * have no valid leaf we try to use MAJOR:default leaf. It still unsuccessful
187 * then finish and return direct queue.
189 #define HTB_DIRECT ((struct htb_class *)-1L)
191 static struct htb_class *htb_classify(struct sk_buff *skb, struct Qdisc *sch,
194 struct htb_sched *q = qdisc_priv(sch);
195 struct htb_class *cl;
196 struct tcf_result res;
197 struct tcf_proto *tcf;
200 /* allow to select class by setting skb->priority to valid classid;
201 * note that nfmark can be used too by attaching filter fw with no
204 if (skb->priority == sch->handle)
205 return HTB_DIRECT; /* X:0 (direct flow) selected */
206 cl = htb_find(skb->priority, sch);
207 if (cl && cl->level == 0)
210 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
211 tcf = q->filter_list;
212 while (tcf && (result = tc_classify(skb, tcf, &res)) >= 0) {
213 #ifdef CONFIG_NET_CLS_ACT
217 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
222 cl = (void *)res.class;
224 if (res.classid == sch->handle)
225 return HTB_DIRECT; /* X:0 (direct flow) */
226 cl = htb_find(res.classid, sch);
228 break; /* filter selected invalid classid */
231 return cl; /* we hit leaf; return it */
233 /* we have got inner class; apply inner filter chain */
234 tcf = cl->filter_list;
236 /* classification failed; try to use default class */
237 cl = htb_find(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
238 if (!cl || cl->level)
239 return HTB_DIRECT; /* bad default .. this is safe bet */
244 * htb_add_to_id_tree - adds class to the round robin list
246 * Routine adds class to the list (actually tree) sorted by classid.
247 * Make sure that class is not already on such list for given prio.
249 static void htb_add_to_id_tree(struct rb_root *root,
250 struct htb_class *cl, int prio)
252 struct rb_node **p = &root->rb_node, *parent = NULL;
257 c = rb_entry(parent, struct htb_class, node[prio]);
259 if (cl->common.classid > c->common.classid)
260 p = &parent->rb_right;
262 p = &parent->rb_left;
264 rb_link_node(&cl->node[prio], parent, p);
265 rb_insert_color(&cl->node[prio], root);
269 * htb_add_to_wait_tree - adds class to the event queue with delay
271 * The class is added to priority event queue to indicate that class will
272 * change its mode in cl->pq_key microseconds. Make sure that class is not
273 * already in the queue.
275 static void htb_add_to_wait_tree(struct htb_sched *q,
276 struct htb_class *cl, long delay)
278 struct rb_node **p = &q->wait_pq[cl->level].rb_node, *parent = NULL;
280 cl->pq_key = q->now + delay;
281 if (cl->pq_key == q->now)
284 /* update the nearest event cache */
285 if (q->near_ev_cache[cl->level] > cl->pq_key)
286 q->near_ev_cache[cl->level] = cl->pq_key;
291 c = rb_entry(parent, struct htb_class, pq_node);
292 if (cl->pq_key >= c->pq_key)
293 p = &parent->rb_right;
295 p = &parent->rb_left;
297 rb_link_node(&cl->pq_node, parent, p);
298 rb_insert_color(&cl->pq_node, &q->wait_pq[cl->level]);
302 * htb_next_rb_node - finds next node in binary tree
304 * When we are past last key we return NULL.
305 * Average complexity is 2 steps per call.
307 static inline void htb_next_rb_node(struct rb_node **n)
313 * htb_add_class_to_row - add class to its row
315 * The class is added to row at priorities marked in mask.
316 * It does nothing if mask == 0.
318 static inline void htb_add_class_to_row(struct htb_sched *q,
319 struct htb_class *cl, int mask)
321 q->row_mask[cl->level] |= mask;
323 int prio = ffz(~mask);
324 mask &= ~(1 << prio);
325 htb_add_to_id_tree(q->row[cl->level] + prio, cl, prio);
329 /* If this triggers, it is a bug in this code, but it need not be fatal */
330 static void htb_safe_rb_erase(struct rb_node *rb, struct rb_root *root)
332 if (RB_EMPTY_NODE(rb)) {
342 * htb_remove_class_from_row - removes class from its row
344 * The class is removed from row at priorities marked in mask.
345 * It does nothing if mask == 0.
347 static inline void htb_remove_class_from_row(struct htb_sched *q,
348 struct htb_class *cl, int mask)
353 int prio = ffz(~mask);
355 mask &= ~(1 << prio);
356 if (q->ptr[cl->level][prio] == cl->node + prio)
357 htb_next_rb_node(q->ptr[cl->level] + prio);
359 htb_safe_rb_erase(cl->node + prio, q->row[cl->level] + prio);
360 if (!q->row[cl->level][prio].rb_node)
363 q->row_mask[cl->level] &= ~m;
367 * htb_activate_prios - creates active classe's feed chain
369 * The class is connected to ancestors and/or appropriate rows
370 * for priorities it is participating on. cl->cmode must be new
371 * (activated) mode. It does nothing if cl->prio_activity == 0.
373 static void htb_activate_prios(struct htb_sched *q, struct htb_class *cl)
375 struct htb_class *p = cl->parent;
376 long m, mask = cl->prio_activity;
378 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
384 if (p->un.inner.feed[prio].rb_node)
385 /* parent already has its feed in use so that
386 * 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
425 p->un.inner.last_ptr_id[prio] = cl->common.classid;
426 p->un.inner.ptr[prio] = NULL;
429 htb_safe_rb_erase(cl->node + prio, p->un.inner.feed + prio);
431 if (!p->un.inner.feed[prio].rb_node)
435 p->prio_activity &= ~mask;
440 if (cl->cmode == HTB_CAN_SEND && mask)
441 htb_remove_class_from_row(q, cl, mask);
444 static inline long htb_lowater(const struct htb_class *cl)
447 return cl->cmode != HTB_CANT_SEND ? -cl->cbuffer : 0;
451 static inline long htb_hiwater(const struct htb_class *cl)
454 return cl->cmode == HTB_CAN_SEND ? -cl->buffer : 0;
461 * htb_class_mode - computes and returns current class mode
463 * It computes cl's mode at time cl->t_c+diff and returns it. If mode
464 * is not HTB_CAN_SEND then cl->pq_key is updated to time difference
465 * from now to time when cl will change its state.
466 * Also it is worth to note that class mode doesn't change simply
467 * at cl->{c,}tokens == 0 but there can rather be hysteresis of
468 * 0 .. -cl->{c,}buffer range. It is meant to limit number of
469 * mode transitions per time unit. The speed gain is about 1/6.
471 static inline enum htb_cmode
472 htb_class_mode(struct htb_class *cl, long *diff)
476 if ((toks = (cl->ctokens + *diff)) < htb_lowater(cl)) {
478 return HTB_CANT_SEND;
481 if ((toks = (cl->tokens + *diff)) >= htb_hiwater(cl))
485 return HTB_MAY_BORROW;
489 * htb_change_class_mode - changes classe's mode
491 * This should be the only way how to change classe's mode under normal
492 * cirsumstances. Routine will update feed lists linkage, change mode
493 * and add class to the wait event queue if appropriate. New mode should
494 * be different from old one and cl->pq_key has to be valid if changing
495 * to mode other than HTB_CAN_SEND (see htb_add_to_wait_tree).
498 htb_change_class_mode(struct htb_sched *q, struct htb_class *cl, long *diff)
500 enum htb_cmode new_mode = htb_class_mode(cl, diff);
502 if (new_mode == cl->cmode)
505 if (cl->prio_activity) { /* not necessary: speed optimization */
506 if (cl->cmode != HTB_CANT_SEND)
507 htb_deactivate_prios(q, cl);
508 cl->cmode = new_mode;
509 if (new_mode != HTB_CANT_SEND)
510 htb_activate_prios(q, cl);
512 cl->cmode = new_mode;
516 * htb_activate - inserts leaf cl into appropriate active feeds
518 * Routine learns (new) priority of leaf and activates feed chain
519 * for the prio. It can be called on already active leaf safely.
520 * It also adds leaf into droplist.
522 static inline void htb_activate(struct htb_sched *q, struct htb_class *cl)
524 WARN_ON(cl->level || !cl->un.leaf.q || !cl->un.leaf.q->q.qlen);
526 if (!cl->prio_activity) {
527 cl->prio_activity = 1 << cl->prio;
528 htb_activate_prios(q, cl);
529 list_add_tail(&cl->un.leaf.drop_list,
530 q->drops + cl->prio);
535 * htb_deactivate - remove leaf cl from active feeds
537 * Make sure that leaf is active. In the other words it can't be called
538 * with non-active leaf. It also removes class from the drop list.
540 static inline void htb_deactivate(struct htb_sched *q, struct htb_class *cl)
542 WARN_ON(!cl->prio_activity);
544 htb_deactivate_prios(q, cl);
545 cl->prio_activity = 0;
546 list_del_init(&cl->un.leaf.drop_list);
549 static int htb_enqueue(struct sk_buff *skb, struct Qdisc *sch)
551 int uninitialized_var(ret);
552 struct htb_sched *q = qdisc_priv(sch);
553 struct htb_class *cl = htb_classify(skb, sch, &ret);
555 if (cl == HTB_DIRECT) {
556 /* enqueue to helper queue */
557 if (q->direct_queue.qlen < q->direct_qlen) {
558 __skb_queue_tail(&q->direct_queue, skb);
563 return NET_XMIT_DROP;
565 #ifdef CONFIG_NET_CLS_ACT
567 if (ret & __NET_XMIT_BYPASS)
572 } else if ((ret = qdisc_enqueue(skb, cl->un.leaf.q)) != NET_XMIT_SUCCESS) {
573 if (net_xmit_drop_count(ret)) {
579 bstats_update(&cl->bstats, skb);
583 sch->qstats.backlog += qdisc_pkt_len(skb);
585 return NET_XMIT_SUCCESS;
588 static inline void htb_accnt_tokens(struct htb_class *cl, int bytes, long diff)
590 long toks = diff + cl->tokens;
592 if (toks > cl->buffer)
594 toks -= (long) qdisc_l2t(cl->rate, bytes);
595 if (toks <= -cl->mbuffer)
596 toks = 1 - cl->mbuffer;
601 static inline void htb_accnt_ctokens(struct htb_class *cl, int bytes, long diff)
603 long toks = diff + cl->ctokens;
605 if (toks > cl->cbuffer)
607 toks -= (long) qdisc_l2t(cl->ceil, bytes);
608 if (toks <= -cl->mbuffer)
609 toks = 1 - cl->mbuffer;
615 * htb_charge_class - charges amount "bytes" to leaf and ancestors
617 * Routine assumes that packet "bytes" long was dequeued from leaf cl
618 * borrowing from "level". It accounts bytes to ceil leaky bucket for
619 * leaf and all ancestors and to rate bucket for ancestors at levels
620 * "level" and higher. It also handles possible change of mode resulting
621 * from the update. Note that mode can also increase here (MAY_BORROW to
622 * CAN_SEND) because we can use more precise clock that event queue here.
623 * In such case we remove class from event queue first.
625 static void htb_charge_class(struct htb_sched *q, struct htb_class *cl,
626 int level, struct sk_buff *skb)
628 int bytes = qdisc_pkt_len(skb);
629 enum htb_cmode old_mode;
633 diff = psched_tdiff_bounded(q->now, cl->t_c, cl->mbuffer);
634 if (cl->level >= level) {
635 if (cl->level == level)
637 htb_accnt_tokens(cl, bytes, diff);
639 cl->xstats.borrows++;
640 cl->tokens += diff; /* we moved t_c; update tokens */
642 htb_accnt_ctokens(cl, bytes, diff);
645 old_mode = cl->cmode;
647 htb_change_class_mode(q, cl, &diff);
648 if (old_mode != cl->cmode) {
649 if (old_mode != HTB_CAN_SEND)
650 htb_safe_rb_erase(&cl->pq_node, q->wait_pq + cl->level);
651 if (cl->cmode != HTB_CAN_SEND)
652 htb_add_to_wait_tree(q, cl, diff);
655 /* update basic stats except for leaves which are already updated */
657 bstats_update(&cl->bstats, skb);
664 * htb_do_events - make mode changes to classes at the level
666 * Scans event queue for pending events and applies them. Returns time of
667 * next pending event (0 for no event in pq, q->now for too many events).
668 * Note: Applied are events whose have cl->pq_key <= q->now.
670 static psched_time_t htb_do_events(struct htb_sched *q, int level,
673 /* don't run for longer than 2 jiffies; 2 is used instead of
674 * 1 to simplify things when jiffy is going to be incremented
677 unsigned long stop_at = start + 2;
678 while (time_before(jiffies, stop_at)) {
679 struct htb_class *cl;
681 struct rb_node *p = rb_first(&q->wait_pq[level]);
686 cl = rb_entry(p, struct htb_class, pq_node);
687 if (cl->pq_key > q->now)
690 htb_safe_rb_erase(p, q->wait_pq + level);
691 diff = psched_tdiff_bounded(q->now, cl->t_c, cl->mbuffer);
692 htb_change_class_mode(q, cl, &diff);
693 if (cl->cmode != HTB_CAN_SEND)
694 htb_add_to_wait_tree(q, cl, diff);
697 /* too much load - let's continue after a break for scheduling */
698 if (!(q->warned & HTB_WARN_TOOMANYEVENTS)) {
699 pr_warning("htb: too many events!\n");
700 q->warned |= HTB_WARN_TOOMANYEVENTS;
706 /* Returns class->node+prio from id-tree where classe's id is >= id. NULL
707 * is no such one exists.
709 static struct rb_node *htb_id_find_next_upper(int prio, struct rb_node *n,
712 struct rb_node *r = NULL;
714 struct htb_class *cl =
715 rb_entry(n, struct htb_class, node[prio]);
717 if (id > cl->common.classid) {
719 } else if (id < cl->common.classid) {
730 * htb_lookup_leaf - returns next leaf class in DRR order
732 * Find leaf where current feed pointers points to.
734 static struct htb_class *htb_lookup_leaf(struct rb_root *tree, int prio,
735 struct rb_node **pptr, u32 * pid)
739 struct rb_node *root;
740 struct rb_node **pptr;
742 } stk[TC_HTB_MAXDEPTH], *sp = stk;
744 BUG_ON(!tree->rb_node);
745 sp->root = tree->rb_node;
749 for (i = 0; i < 65535; i++) {
750 if (!*sp->pptr && *sp->pid) {
751 /* ptr was invalidated but id is valid - try to recover
752 * the original or next ptr
755 htb_id_find_next_upper(prio, sp->root, *sp->pid);
757 *sp->pid = 0; /* ptr is valid now so that remove this hint as it
758 * can become out of date quickly
760 if (!*sp->pptr) { /* we are at right end; rewind & go up */
761 *sp->pptr = sp->root;
762 while ((*sp->pptr)->rb_left)
763 *sp->pptr = (*sp->pptr)->rb_left;
770 htb_next_rb_node(sp->pptr);
773 struct htb_class *cl;
774 cl = rb_entry(*sp->pptr, struct htb_class, node[prio]);
777 (++sp)->root = cl->un.inner.feed[prio].rb_node;
778 sp->pptr = cl->un.inner.ptr + prio;
779 sp->pid = cl->un.inner.last_ptr_id + prio;
786 /* dequeues packet at given priority and level; call only if
787 * you are sure that there is active class at prio/level
789 static struct sk_buff *htb_dequeue_tree(struct htb_sched *q, int prio,
792 struct sk_buff *skb = NULL;
793 struct htb_class *cl, *start;
794 /* look initial class up in the row */
795 start = cl = htb_lookup_leaf(q->row[level] + prio, prio,
796 q->ptr[level] + prio,
797 q->last_ptr_id[level] + prio);
804 /* class can be empty - it is unlikely but can be true if leaf
805 * qdisc drops packets in enqueue routine or if someone used
806 * graft operation on the leaf since last dequeue;
807 * simply deactivate and skip such class
809 if (unlikely(cl->un.leaf.q->q.qlen == 0)) {
810 struct htb_class *next;
811 htb_deactivate(q, cl);
813 /* row/level might become empty */
814 if ((q->row_mask[level] & (1 << prio)) == 0)
817 next = htb_lookup_leaf(q->row[level] + prio,
818 prio, q->ptr[level] + prio,
819 q->last_ptr_id[level] + prio);
821 if (cl == start) /* fix start if we just deleted it */
827 skb = cl->un.leaf.q->dequeue(cl->un.leaf.q);
828 if (likely(skb != NULL))
831 qdisc_warn_nonwc("htb", cl->un.leaf.q);
832 htb_next_rb_node((level ? cl->parent->un.inner.ptr : q->
834 cl = htb_lookup_leaf(q->row[level] + prio, prio,
835 q->ptr[level] + prio,
836 q->last_ptr_id[level] + prio);
838 } while (cl != start);
840 if (likely(skb != NULL)) {
841 cl->un.leaf.deficit[level] -= qdisc_pkt_len(skb);
842 if (cl->un.leaf.deficit[level] < 0) {
843 cl->un.leaf.deficit[level] += cl->quantum;
844 htb_next_rb_node((level ? cl->parent->un.inner.ptr : q->
847 /* this used to be after charge_class but this constelation
848 * gives us slightly better performance
850 if (!cl->un.leaf.q->q.qlen)
851 htb_deactivate(q, cl);
852 htb_charge_class(q, cl, level, skb);
857 static struct sk_buff *htb_dequeue(struct Qdisc *sch)
860 struct htb_sched *q = qdisc_priv(sch);
862 psched_time_t next_event;
863 unsigned long start_at;
865 /* try to dequeue direct packets as high prio (!) to minimize cpu work */
866 skb = __skb_dequeue(&q->direct_queue);
869 qdisc_bstats_update(sch, skb);
870 qdisc_unthrottled(sch);
871 sch->qstats.backlog -= qdisc_pkt_len(skb);
878 q->now = psched_get_time();
881 next_event = q->now + 5LLU * PSCHED_TICKS_PER_SEC;
883 for (level = 0; level < TC_HTB_MAXDEPTH; level++) {
884 /* common case optimization - skip event handler quickly */
888 if (q->now >= q->near_ev_cache[level]) {
889 event = htb_do_events(q, level, start_at);
891 event = q->now + PSCHED_TICKS_PER_SEC;
892 q->near_ev_cache[level] = event;
894 event = q->near_ev_cache[level];
896 if (next_event > event)
899 m = ~q->row_mask[level];
900 while (m != (int)(-1)) {
904 skb = htb_dequeue_tree(q, prio, level);
905 if (likely(skb != NULL))
909 sch->qstats.overlimits++;
910 if (likely(next_event > q->now))
911 qdisc_watchdog_schedule(&q->watchdog, next_event);
913 schedule_work(&q->work);
918 /* try to drop from each class (by prio) until one succeed */
919 static unsigned int htb_drop(struct Qdisc *sch)
921 struct htb_sched *q = qdisc_priv(sch);
924 for (prio = TC_HTB_NUMPRIO - 1; prio >= 0; prio--) {
926 list_for_each(p, q->drops + prio) {
927 struct htb_class *cl = list_entry(p, struct htb_class,
930 if (cl->un.leaf.q->ops->drop &&
931 (len = cl->un.leaf.q->ops->drop(cl->un.leaf.q))) {
932 sch->qstats.backlog -= len;
934 if (!cl->un.leaf.q->q.qlen)
935 htb_deactivate(q, cl);
943 /* reset all classes */
944 /* always caled under BH & queue lock */
945 static void htb_reset(struct Qdisc *sch)
947 struct htb_sched *q = qdisc_priv(sch);
948 struct htb_class *cl;
949 struct hlist_node *n;
952 for (i = 0; i < q->clhash.hashsize; i++) {
953 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode) {
955 memset(&cl->un.inner, 0, sizeof(cl->un.inner));
958 qdisc_reset(cl->un.leaf.q);
959 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
961 cl->prio_activity = 0;
962 cl->cmode = HTB_CAN_SEND;
965 qdisc_watchdog_cancel(&q->watchdog);
966 __skb_queue_purge(&q->direct_queue);
968 sch->qstats.backlog = 0;
969 memset(q->row, 0, sizeof(q->row));
970 memset(q->row_mask, 0, sizeof(q->row_mask));
971 memset(q->wait_pq, 0, sizeof(q->wait_pq));
972 memset(q->ptr, 0, sizeof(q->ptr));
973 for (i = 0; i < TC_HTB_NUMPRIO; i++)
974 INIT_LIST_HEAD(q->drops + i);
977 static const struct nla_policy htb_policy[TCA_HTB_MAX + 1] = {
978 [TCA_HTB_PARMS] = { .len = sizeof(struct tc_htb_opt) },
979 [TCA_HTB_INIT] = { .len = sizeof(struct tc_htb_glob) },
980 [TCA_HTB_CTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
981 [TCA_HTB_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
984 static void htb_work_func(struct work_struct *work)
986 struct htb_sched *q = container_of(work, struct htb_sched, work);
987 struct Qdisc *sch = q->watchdog.qdisc;
989 __netif_schedule(qdisc_root(sch));
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;
1000 qdisc_watchdog_init(&q->watchdog, sch);
1001 INIT_WORK(&q->work, htb_work_func);
1006 err = nla_parse_nested(tb, TCA_HTB_INIT, opt, htb_policy);
1010 if (tb[TCA_HTB_INIT] == NULL) {
1011 pr_err("HTB: hey probably you have bad tc tool ?\n");
1014 gopt = nla_data(tb[TCA_HTB_INIT]);
1015 if (gopt->version != HTB_VER >> 16) {
1016 pr_err("HTB: need tc/htb version %d (minor is %d), you have %d\n",
1017 HTB_VER >> 16, HTB_VER & 0xffff, gopt->version);
1021 err = qdisc_class_hash_init(&q->clhash);
1024 for (i = 0; i < TC_HTB_NUMPRIO; i++)
1025 INIT_LIST_HEAD(q->drops + i);
1027 skb_queue_head_init(&q->direct_queue);
1029 q->direct_qlen = qdisc_dev(sch)->tx_queue_len;
1030 if (q->direct_qlen < 2) /* some devices have zero tx_queue_len */
1033 if ((q->rate2quantum = gopt->rate2quantum) < 1)
1034 q->rate2quantum = 1;
1035 q->defcls = gopt->defcls;
1040 static int htb_dump(struct Qdisc *sch, struct sk_buff *skb)
1042 spinlock_t *root_lock = qdisc_root_sleeping_lock(sch);
1043 struct htb_sched *q = qdisc_priv(sch);
1044 struct nlattr *nest;
1045 struct tc_htb_glob gopt;
1047 spin_lock_bh(root_lock);
1049 gopt.direct_pkts = q->direct_pkts;
1050 gopt.version = HTB_VER;
1051 gopt.rate2quantum = q->rate2quantum;
1052 gopt.defcls = q->defcls;
1055 nest = nla_nest_start(skb, TCA_OPTIONS);
1057 goto nla_put_failure;
1058 NLA_PUT(skb, TCA_HTB_INIT, sizeof(gopt), &gopt);
1059 nla_nest_end(skb, nest);
1061 spin_unlock_bh(root_lock);
1065 spin_unlock_bh(root_lock);
1066 nla_nest_cancel(skb, nest);
1070 static int htb_dump_class(struct Qdisc *sch, unsigned long arg,
1071 struct sk_buff *skb, struct tcmsg *tcm)
1073 struct htb_class *cl = (struct htb_class *)arg;
1074 spinlock_t *root_lock = qdisc_root_sleeping_lock(sch);
1075 struct nlattr *nest;
1076 struct tc_htb_opt opt;
1078 spin_lock_bh(root_lock);
1079 tcm->tcm_parent = cl->parent ? cl->parent->common.classid : TC_H_ROOT;
1080 tcm->tcm_handle = cl->common.classid;
1081 if (!cl->level && cl->un.leaf.q)
1082 tcm->tcm_info = cl->un.leaf.q->handle;
1084 nest = nla_nest_start(skb, TCA_OPTIONS);
1086 goto nla_put_failure;
1088 memset(&opt, 0, sizeof(opt));
1090 opt.rate = cl->rate->rate;
1091 opt.buffer = cl->buffer;
1092 opt.ceil = cl->ceil->rate;
1093 opt.cbuffer = cl->cbuffer;
1094 opt.quantum = cl->quantum;
1095 opt.prio = cl->prio;
1096 opt.level = cl->level;
1097 NLA_PUT(skb, TCA_HTB_PARMS, sizeof(opt), &opt);
1099 nla_nest_end(skb, nest);
1100 spin_unlock_bh(root_lock);
1104 spin_unlock_bh(root_lock);
1105 nla_nest_cancel(skb, nest);
1110 htb_dump_class_stats(struct Qdisc *sch, unsigned long arg, struct gnet_dump *d)
1112 struct htb_class *cl = (struct htb_class *)arg;
1114 if (!cl->level && cl->un.leaf.q)
1115 cl->qstats.qlen = cl->un.leaf.q->q.qlen;
1116 cl->xstats.tokens = cl->tokens;
1117 cl->xstats.ctokens = cl->ctokens;
1119 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
1120 gnet_stats_copy_rate_est(d, NULL, &cl->rate_est) < 0 ||
1121 gnet_stats_copy_queue(d, &cl->qstats) < 0)
1124 return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1127 static int htb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1130 struct htb_class *cl = (struct htb_class *)arg;
1135 (new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1136 cl->common.classid)) == NULL)
1140 *old = cl->un.leaf.q;
1141 cl->un.leaf.q = new;
1143 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
1146 sch_tree_unlock(sch);
1150 static struct Qdisc *htb_leaf(struct Qdisc *sch, unsigned long arg)
1152 struct htb_class *cl = (struct htb_class *)arg;
1153 return !cl->level ? cl->un.leaf.q : NULL;
1156 static void htb_qlen_notify(struct Qdisc *sch, unsigned long arg)
1158 struct htb_class *cl = (struct htb_class *)arg;
1160 if (cl->un.leaf.q->q.qlen == 0)
1161 htb_deactivate(qdisc_priv(sch), cl);
1164 static unsigned long htb_get(struct Qdisc *sch, u32 classid)
1166 struct htb_class *cl = htb_find(classid, sch);
1169 return (unsigned long)cl;
1172 static inline int htb_parent_last_child(struct htb_class *cl)
1175 /* the root class */
1177 if (cl->parent->children > 1)
1178 /* not the last child */
1183 static void htb_parent_to_leaf(struct htb_sched *q, struct htb_class *cl,
1184 struct Qdisc *new_q)
1186 struct htb_class *parent = cl->parent;
1188 WARN_ON(cl->level || !cl->un.leaf.q || cl->prio_activity);
1190 if (parent->cmode != HTB_CAN_SEND)
1191 htb_safe_rb_erase(&parent->pq_node, q->wait_pq + parent->level);
1194 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1195 INIT_LIST_HEAD(&parent->un.leaf.drop_list);
1196 parent->un.leaf.q = new_q ? new_q : &noop_qdisc;
1197 parent->tokens = parent->buffer;
1198 parent->ctokens = parent->cbuffer;
1199 parent->t_c = psched_get_time();
1200 parent->cmode = HTB_CAN_SEND;
1203 static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
1206 WARN_ON(!cl->un.leaf.q);
1207 qdisc_destroy(cl->un.leaf.q);
1209 gen_kill_estimator(&cl->bstats, &cl->rate_est);
1210 qdisc_put_rtab(cl->rate);
1211 qdisc_put_rtab(cl->ceil);
1213 tcf_destroy_chain(&cl->filter_list);
1217 static void htb_destroy(struct Qdisc *sch)
1219 struct htb_sched *q = qdisc_priv(sch);
1220 struct hlist_node *n, *next;
1221 struct htb_class *cl;
1224 cancel_work_sync(&q->work);
1225 qdisc_watchdog_cancel(&q->watchdog);
1226 /* This line used to be after htb_destroy_class call below
1227 * and surprisingly it worked in 2.4. But it must precede it
1228 * because filter need its target class alive to be able to call
1229 * 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(sch->dev_queue, &pfifo_qdisc_ops,
1262 cl->parent->common.classid);
1269 qlen = cl->un.leaf.q->q.qlen;
1270 qdisc_reset(cl->un.leaf.q);
1271 qdisc_tree_decrease_qlen(cl->un.leaf.q, qlen);
1274 /* delete from hash and active; remainder in destroy_class */
1275 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 BUG_ON(--cl->refcnt == 0);
1290 * This shouldn't happen: we "hold" one cops->get() when called
1291 * from tc_ctl_tclass; the destroy method is done from cops->put().
1294 sch_tree_unlock(sch);
1298 static void htb_put(struct Qdisc *sch, unsigned long arg)
1300 struct htb_class *cl = (struct htb_class *)arg;
1302 if (--cl->refcnt == 0)
1303 htb_destroy_class(sch, cl);
1306 static int htb_change_class(struct Qdisc *sch, u32 classid,
1307 u32 parentid, struct nlattr **tca,
1311 struct htb_sched *q = qdisc_priv(sch);
1312 struct htb_class *cl = (struct htb_class *)*arg, *parent;
1313 struct nlattr *opt = tca[TCA_OPTIONS];
1314 struct qdisc_rate_table *rtab = NULL, *ctab = NULL;
1315 struct nlattr *tb[__TCA_HTB_MAX];
1316 struct tc_htb_opt *hopt;
1318 /* extract all subattrs from opt attr */
1322 err = nla_parse_nested(tb, TCA_HTB_MAX, opt, htb_policy);
1327 if (tb[TCA_HTB_PARMS] == NULL)
1330 parent = parentid == TC_H_ROOT ? NULL : htb_find(parentid, sch);
1332 hopt = nla_data(tb[TCA_HTB_PARMS]);
1334 rtab = qdisc_get_rtab(&hopt->rate, tb[TCA_HTB_RTAB]);
1335 ctab = qdisc_get_rtab(&hopt->ceil, tb[TCA_HTB_CTAB]);
1339 if (!cl) { /* new class */
1340 struct Qdisc *new_q;
1344 struct gnet_estimator opt;
1347 .nla_len = nla_attr_size(sizeof(est.opt)),
1348 .nla_type = TCA_RATE,
1351 /* 4s interval, 16s averaging constant */
1357 /* check for valid classid */
1358 if (!classid || TC_H_MAJ(classid ^ sch->handle) ||
1359 htb_find(classid, sch))
1362 /* check maximal depth */
1363 if (parent && parent->parent && parent->parent->level < 2) {
1364 pr_err("htb: tree is too deep\n");
1368 cl = kzalloc(sizeof(*cl), GFP_KERNEL);
1372 err = gen_new_estimator(&cl->bstats, &cl->rate_est,
1373 qdisc_root_sleeping_lock(sch),
1374 tca[TCA_RATE] ? : &est.nla);
1382 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
1383 RB_CLEAR_NODE(&cl->pq_node);
1385 for (prio = 0; prio < TC_HTB_NUMPRIO; prio++)
1386 RB_CLEAR_NODE(&cl->node[prio]);
1388 /* create leaf qdisc early because it uses kmalloc(GFP_KERNEL)
1389 * so that can't be used inside of sch_tree_lock
1390 * -- thanks to Karlis Peisenieks
1392 new_q = qdisc_create_dflt(sch->dev_queue,
1393 &pfifo_qdisc_ops, classid);
1395 if (parent && !parent->level) {
1396 unsigned int qlen = parent->un.leaf.q->q.qlen;
1398 /* turn parent into inner node */
1399 qdisc_reset(parent->un.leaf.q);
1400 qdisc_tree_decrease_qlen(parent->un.leaf.q, qlen);
1401 qdisc_destroy(parent->un.leaf.q);
1402 if (parent->prio_activity)
1403 htb_deactivate(q, parent);
1405 /* remove from evt list because of level change */
1406 if (parent->cmode != HTB_CAN_SEND) {
1407 htb_safe_rb_erase(&parent->pq_node, q->wait_pq);
1408 parent->cmode = HTB_CAN_SEND;
1410 parent->level = (parent->parent ? parent->parent->level
1411 : TC_HTB_MAXDEPTH) - 1;
1412 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1414 /* leaf (we) needs elementary qdisc */
1415 cl->un.leaf.q = new_q ? new_q : &noop_qdisc;
1417 cl->common.classid = classid;
1418 cl->parent = parent;
1420 /* set class to be in HTB_CAN_SEND state */
1421 cl->tokens = hopt->buffer;
1422 cl->ctokens = hopt->cbuffer;
1423 cl->mbuffer = 60 * PSCHED_TICKS_PER_SEC; /* 1min */
1424 cl->t_c = psched_get_time();
1425 cl->cmode = HTB_CAN_SEND;
1427 /* attach to the hash list and parent's family */
1428 qdisc_class_hash_insert(&q->clhash, &cl->common);
1432 if (tca[TCA_RATE]) {
1433 err = gen_replace_estimator(&cl->bstats, &cl->rate_est,
1434 qdisc_root_sleeping_lock(sch),
1442 /* it used to be a nasty bug here, we have to check that node
1443 * is really leaf before changing cl->un.leaf !
1446 cl->quantum = rtab->rate.rate / q->rate2quantum;
1447 if (!hopt->quantum && cl->quantum < 1000) {
1449 "HTB: quantum of class %X is small. Consider r2q change.\n",
1450 cl->common.classid);
1453 if (!hopt->quantum && cl->quantum > 200000) {
1455 "HTB: quantum of class %X is big. Consider r2q change.\n",
1456 cl->common.classid);
1457 cl->quantum = 200000;
1460 cl->quantum = hopt->quantum;
1461 if ((cl->prio = hopt->prio) >= TC_HTB_NUMPRIO)
1462 cl->prio = TC_HTB_NUMPRIO - 1;
1465 cl->buffer = hopt->buffer;
1466 cl->cbuffer = hopt->cbuffer;
1468 qdisc_put_rtab(cl->rate);
1471 qdisc_put_rtab(cl->ceil);
1473 sch_tree_unlock(sch);
1475 qdisc_class_hash_grow(sch, &q->clhash);
1477 *arg = (unsigned long)cl;
1482 qdisc_put_rtab(rtab);
1484 qdisc_put_rtab(ctab);
1488 static struct tcf_proto **htb_find_tcf(struct Qdisc *sch, unsigned long arg)
1490 struct htb_sched *q = qdisc_priv(sch);
1491 struct htb_class *cl = (struct htb_class *)arg;
1492 struct tcf_proto **fl = cl ? &cl->filter_list : &q->filter_list;
1497 static unsigned long htb_bind_filter(struct Qdisc *sch, unsigned long parent,
1500 struct htb_class *cl = htb_find(classid, sch);
1502 /*if (cl && !cl->level) return 0;
1503 * The line above used to be there to prevent attaching filters to
1504 * leaves. But at least tc_index filter uses this just to get class
1505 * for other reasons so that we have to allow for it.
1507 * 19.6.2002 As Werner explained it is ok - bind filter is just
1508 * another way to "lock" the class - unlike "get" this lock can
1509 * be broken by class during destroy IIUC.
1513 return (unsigned long)cl;
1516 static void htb_unbind_filter(struct Qdisc *sch, unsigned long arg)
1518 struct htb_class *cl = (struct htb_class *)arg;
1524 static void htb_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1526 struct htb_sched *q = qdisc_priv(sch);
1527 struct htb_class *cl;
1528 struct hlist_node *n;
1534 for (i = 0; i < q->clhash.hashsize; i++) {
1535 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode) {
1536 if (arg->count < arg->skip) {
1540 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1549 static const struct Qdisc_class_ops htb_class_ops = {
1552 .qlen_notify = htb_qlen_notify,
1555 .change = htb_change_class,
1556 .delete = htb_delete,
1558 .tcf_chain = htb_find_tcf,
1559 .bind_tcf = htb_bind_filter,
1560 .unbind_tcf = htb_unbind_filter,
1561 .dump = htb_dump_class,
1562 .dump_stats = htb_dump_class_stats,
1565 static struct Qdisc_ops htb_qdisc_ops __read_mostly = {
1566 .cl_ops = &htb_class_ops,
1568 .priv_size = sizeof(struct htb_sched),
1569 .enqueue = htb_enqueue,
1570 .dequeue = htb_dequeue,
1571 .peek = qdisc_peek_dequeued,
1575 .destroy = htb_destroy,
1577 .owner = THIS_MODULE,
1580 static int __init htb_module_init(void)
1582 return register_qdisc(&htb_qdisc_ops);
1584 static void __exit htb_module_exit(void)
1586 unregister_qdisc(&htb_qdisc_ops);
1589 module_init(htb_module_init)
1590 module_exit(htb_module_exit)
1591 MODULE_LICENSE("GPL");