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/sch_generic.h>
42 #include <net/pkt_sched.h>
43 #include <net/pkt_cls.h>
47 ========================================================================
48 HTB is like TBF with multiple classes. It is also similar to CBQ because
49 it allows to assign priority to each class in hierarchy.
50 In fact it is another implementation of Floyd's formal sharing.
53 Each class is assigned level. Leaf has ALWAYS level 0 and root
54 classes have level TC_HTB_MAXDEPTH-1. Interior nodes has level
55 one less than their parent.
58 static int htb_hysteresis __read_mostly = 0; /* whether to use mode hysteresis for speedup */
59 #define HTB_VER 0x30011 /* major must be matched with number suplied by TC as version */
61 #if HTB_VER >> 16 != TC_HTB_PROTOVER
62 #error "Mismatched sch_htb.c and pkt_sch.h"
65 /* Module parameter and sysfs export */
66 module_param (htb_hysteresis, int, 0640);
67 MODULE_PARM_DESC(htb_hysteresis, "Hysteresis mode, less CPU load, less accurate");
69 static int htb_rate_est = 0; /* htb classes have a default rate estimator */
70 module_param(htb_rate_est, int, 0640);
71 MODULE_PARM_DESC(htb_rate_est, "setup a default rate estimator (4sec 16sec) for htb classes");
73 /* used internaly to keep status of single class */
75 HTB_CANT_SEND, /* class can't send and can't borrow */
76 HTB_MAY_BORROW, /* class can't send but may borrow */
77 HTB_CAN_SEND /* class can send */
86 /* When class changes from state 1->2 and disconnects from
87 * parent's feed then we lost ptr value and start from the
88 * first child again. Here we store classid of the
89 * last valid ptr (used when ptr is NULL).
94 /* interior & leaf nodes; props specific to leaves are marked L:
95 * To reduce false sharing, place mostly read fields at beginning,
96 * and mostly written ones at the end.
99 struct Qdisc_class_common common;
100 struct psched_ratecfg rate;
101 struct psched_ratecfg ceil;
102 s64 buffer, cbuffer;/* token bucket depth/rate */
103 s64 mbuffer; /* max wait time */
104 u32 prio; /* these two are used only by leaves... */
105 int quantum; /* but stored for parent-to-leaf return */
107 struct tcf_proto __rcu *filter_list; /* class attached filters */
108 struct tcf_block *block;
111 int level; /* our level (see above) */
112 unsigned int children;
113 struct htb_class *parent; /* parent class */
115 struct net_rate_estimator __rcu *rate_est;
118 * Written often fields
120 struct gnet_stats_basic_packed bstats;
121 struct tc_htb_xstats xstats; /* our special stats */
123 /* token bucket parameters */
124 s64 tokens, ctokens;/* current number of tokens */
125 s64 t_c; /* checkpoint time */
128 struct htb_class_leaf {
129 int deficit[TC_HTB_MAXDEPTH];
132 struct htb_class_inner {
133 struct htb_prio clprio[TC_HTB_NUMPRIO];
138 int prio_activity; /* for which prios are we active */
139 enum htb_cmode cmode; /* current mode of the class */
140 struct rb_node pq_node; /* node for event queue */
141 struct rb_node node[TC_HTB_NUMPRIO]; /* node for self or feed tree */
143 unsigned int drops ____cacheline_aligned_in_smp;
144 unsigned int overlimits;
148 struct rb_root wait_pq;
149 struct htb_prio hprio[TC_HTB_NUMPRIO];
153 struct Qdisc_class_hash clhash;
154 int defcls; /* class where unclassified flows go to */
155 int rate2quantum; /* quant = rate / rate2quantum */
157 /* filters for qdisc itself */
158 struct tcf_proto __rcu *filter_list;
159 struct tcf_block *block;
161 #define HTB_WARN_TOOMANYEVENTS 0x1
162 unsigned int warned; /* only one warning */
164 struct work_struct work;
166 /* non shaped skbs; let them go directly thru */
167 struct qdisc_skb_head direct_queue;
171 struct qdisc_watchdog watchdog;
173 s64 now; /* cached dequeue time */
175 /* time of nearest event per level (row) */
176 s64 near_ev_cache[TC_HTB_MAXDEPTH];
178 int row_mask[TC_HTB_MAXDEPTH];
180 struct htb_level hlevel[TC_HTB_MAXDEPTH];
183 /* find class in global hash table using given handle */
184 static inline struct htb_class *htb_find(u32 handle, struct Qdisc *sch)
186 struct htb_sched *q = qdisc_priv(sch);
187 struct Qdisc_class_common *clc;
189 clc = qdisc_class_find(&q->clhash, handle);
192 return container_of(clc, struct htb_class, common);
195 static unsigned long htb_search(struct Qdisc *sch, u32 handle)
197 return (unsigned long)htb_find(handle, sch);
200 * htb_classify - classify a packet into class
202 * It returns NULL if the packet should be dropped or -1 if the packet
203 * should be passed directly thru. In all other cases leaf class is returned.
204 * We allow direct class selection by classid in priority. The we examine
205 * filters in qdisc and in inner nodes (if higher filter points to the inner
206 * node). If we end up with classid MAJOR:0 we enqueue the skb into special
207 * internal fifo (direct). These packets then go directly thru. If we still
208 * have no valid leaf we try to use MAJOR:default leaf. It still unsuccessful
209 * then finish and return direct queue.
211 #define HTB_DIRECT ((struct htb_class *)-1L)
213 static struct htb_class *htb_classify(struct sk_buff *skb, struct Qdisc *sch,
216 struct htb_sched *q = qdisc_priv(sch);
217 struct htb_class *cl;
218 struct tcf_result res;
219 struct tcf_proto *tcf;
222 /* allow to select class by setting skb->priority to valid classid;
223 * note that nfmark can be used too by attaching filter fw with no
226 if (skb->priority == sch->handle)
227 return HTB_DIRECT; /* X:0 (direct flow) selected */
228 cl = htb_find(skb->priority, sch);
232 /* Start with inner filter chain if a non-leaf class is selected */
233 tcf = rcu_dereference_bh(cl->filter_list);
235 tcf = rcu_dereference_bh(q->filter_list);
238 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
239 while (tcf && (result = tcf_classify(skb, tcf, &res, false)) >= 0) {
240 #ifdef CONFIG_NET_CLS_ACT
245 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
251 cl = (void *)res.class;
253 if (res.classid == sch->handle)
254 return HTB_DIRECT; /* X:0 (direct flow) */
255 cl = htb_find(res.classid, sch);
257 break; /* filter selected invalid classid */
260 return cl; /* we hit leaf; return it */
262 /* we have got inner class; apply inner filter chain */
263 tcf = rcu_dereference_bh(cl->filter_list);
265 /* classification failed; try to use default class */
266 cl = htb_find(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
267 if (!cl || cl->level)
268 return HTB_DIRECT; /* bad default .. this is safe bet */
273 * htb_add_to_id_tree - adds class to the round robin list
275 * Routine adds class to the list (actually tree) sorted by classid.
276 * Make sure that class is not already on such list for given prio.
278 static void htb_add_to_id_tree(struct rb_root *root,
279 struct htb_class *cl, int prio)
281 struct rb_node **p = &root->rb_node, *parent = NULL;
286 c = rb_entry(parent, struct htb_class, node[prio]);
288 if (cl->common.classid > c->common.classid)
289 p = &parent->rb_right;
291 p = &parent->rb_left;
293 rb_link_node(&cl->node[prio], parent, p);
294 rb_insert_color(&cl->node[prio], root);
298 * htb_add_to_wait_tree - adds class to the event queue with delay
300 * The class is added to priority event queue to indicate that class will
301 * change its mode in cl->pq_key microseconds. Make sure that class is not
302 * already in the queue.
304 static void htb_add_to_wait_tree(struct htb_sched *q,
305 struct htb_class *cl, s64 delay)
307 struct rb_node **p = &q->hlevel[cl->level].wait_pq.rb_node, *parent = NULL;
309 cl->pq_key = q->now + delay;
310 if (cl->pq_key == q->now)
313 /* update the nearest event cache */
314 if (q->near_ev_cache[cl->level] > cl->pq_key)
315 q->near_ev_cache[cl->level] = cl->pq_key;
320 c = rb_entry(parent, struct htb_class, pq_node);
321 if (cl->pq_key >= c->pq_key)
322 p = &parent->rb_right;
324 p = &parent->rb_left;
326 rb_link_node(&cl->pq_node, parent, p);
327 rb_insert_color(&cl->pq_node, &q->hlevel[cl->level].wait_pq);
331 * htb_next_rb_node - finds next node in binary tree
333 * When we are past last key we return NULL.
334 * Average complexity is 2 steps per call.
336 static inline void htb_next_rb_node(struct rb_node **n)
342 * htb_add_class_to_row - add class to its row
344 * The class is added to row at priorities marked in mask.
345 * It does nothing if mask == 0.
347 static inline void htb_add_class_to_row(struct htb_sched *q,
348 struct htb_class *cl, int mask)
350 q->row_mask[cl->level] |= mask;
352 int prio = ffz(~mask);
353 mask &= ~(1 << prio);
354 htb_add_to_id_tree(&q->hlevel[cl->level].hprio[prio].row, cl, prio);
358 /* If this triggers, it is a bug in this code, but it need not be fatal */
359 static void htb_safe_rb_erase(struct rb_node *rb, struct rb_root *root)
361 if (RB_EMPTY_NODE(rb)) {
371 * htb_remove_class_from_row - removes class from its row
373 * The class is removed from row at priorities marked in mask.
374 * It does nothing if mask == 0.
376 static inline void htb_remove_class_from_row(struct htb_sched *q,
377 struct htb_class *cl, int mask)
380 struct htb_level *hlevel = &q->hlevel[cl->level];
383 int prio = ffz(~mask);
384 struct htb_prio *hprio = &hlevel->hprio[prio];
386 mask &= ~(1 << prio);
387 if (hprio->ptr == cl->node + prio)
388 htb_next_rb_node(&hprio->ptr);
390 htb_safe_rb_erase(cl->node + prio, &hprio->row);
391 if (!hprio->row.rb_node)
394 q->row_mask[cl->level] &= ~m;
398 * htb_activate_prios - creates active classe's feed chain
400 * The class is connected to ancestors and/or appropriate rows
401 * for priorities it is participating on. cl->cmode must be new
402 * (activated) mode. It does nothing if cl->prio_activity == 0.
404 static void htb_activate_prios(struct htb_sched *q, struct htb_class *cl)
406 struct htb_class *p = cl->parent;
407 long m, mask = cl->prio_activity;
409 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
415 if (p->inner.clprio[prio].feed.rb_node)
416 /* parent already has its feed in use so that
417 * reset bit in mask as parent is already ok
419 mask &= ~(1 << prio);
421 htb_add_to_id_tree(&p->inner.clprio[prio].feed, cl, prio);
423 p->prio_activity |= mask;
428 if (cl->cmode == HTB_CAN_SEND && mask)
429 htb_add_class_to_row(q, cl, mask);
433 * htb_deactivate_prios - remove class from feed chain
435 * cl->cmode must represent old mode (before deactivation). It does
436 * nothing if cl->prio_activity == 0. Class is removed from all feed
439 static void htb_deactivate_prios(struct htb_sched *q, struct htb_class *cl)
441 struct htb_class *p = cl->parent;
442 long m, mask = cl->prio_activity;
444 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
451 if (p->inner.clprio[prio].ptr == cl->node + prio) {
452 /* we are removing child which is pointed to from
453 * parent feed - forget the pointer but remember
456 p->inner.clprio[prio].last_ptr_id = cl->common.classid;
457 p->inner.clprio[prio].ptr = NULL;
460 htb_safe_rb_erase(cl->node + prio,
461 &p->inner.clprio[prio].feed);
463 if (!p->inner.clprio[prio].feed.rb_node)
467 p->prio_activity &= ~mask;
472 if (cl->cmode == HTB_CAN_SEND && mask)
473 htb_remove_class_from_row(q, cl, mask);
476 static inline s64 htb_lowater(const struct htb_class *cl)
479 return cl->cmode != HTB_CANT_SEND ? -cl->cbuffer : 0;
483 static inline s64 htb_hiwater(const struct htb_class *cl)
486 return cl->cmode == HTB_CAN_SEND ? -cl->buffer : 0;
493 * htb_class_mode - computes and returns current class mode
495 * It computes cl's mode at time cl->t_c+diff and returns it. If mode
496 * is not HTB_CAN_SEND then cl->pq_key is updated to time difference
497 * from now to time when cl will change its state.
498 * Also it is worth to note that class mode doesn't change simply
499 * at cl->{c,}tokens == 0 but there can rather be hysteresis of
500 * 0 .. -cl->{c,}buffer range. It is meant to limit number of
501 * mode transitions per time unit. The speed gain is about 1/6.
503 static inline enum htb_cmode
504 htb_class_mode(struct htb_class *cl, s64 *diff)
508 if ((toks = (cl->ctokens + *diff)) < htb_lowater(cl)) {
510 return HTB_CANT_SEND;
513 if ((toks = (cl->tokens + *diff)) >= htb_hiwater(cl))
517 return HTB_MAY_BORROW;
521 * htb_change_class_mode - changes classe's mode
523 * This should be the only way how to change classe's mode under normal
524 * cirsumstances. Routine will update feed lists linkage, change mode
525 * and add class to the wait event queue if appropriate. New mode should
526 * be different from old one and cl->pq_key has to be valid if changing
527 * to mode other than HTB_CAN_SEND (see htb_add_to_wait_tree).
530 htb_change_class_mode(struct htb_sched *q, struct htb_class *cl, s64 *diff)
532 enum htb_cmode new_mode = htb_class_mode(cl, diff);
534 if (new_mode == cl->cmode)
537 if (new_mode == HTB_CANT_SEND) {
542 if (cl->prio_activity) { /* not necessary: speed optimization */
543 if (cl->cmode != HTB_CANT_SEND)
544 htb_deactivate_prios(q, cl);
545 cl->cmode = new_mode;
546 if (new_mode != HTB_CANT_SEND)
547 htb_activate_prios(q, cl);
549 cl->cmode = new_mode;
553 * htb_activate - inserts leaf cl into appropriate active feeds
555 * Routine learns (new) priority of leaf and activates feed chain
556 * for the prio. It can be called on already active leaf safely.
557 * It also adds leaf into droplist.
559 static inline void htb_activate(struct htb_sched *q, struct htb_class *cl)
561 WARN_ON(cl->level || !cl->leaf.q || !cl->leaf.q->q.qlen);
563 if (!cl->prio_activity) {
564 cl->prio_activity = 1 << cl->prio;
565 htb_activate_prios(q, cl);
570 * htb_deactivate - remove leaf cl from active feeds
572 * Make sure that leaf is active. In the other words it can't be called
573 * with non-active leaf. It also removes class from the drop list.
575 static inline void htb_deactivate(struct htb_sched *q, struct htb_class *cl)
577 WARN_ON(!cl->prio_activity);
579 htb_deactivate_prios(q, cl);
580 cl->prio_activity = 0;
583 static int htb_enqueue(struct sk_buff *skb, struct Qdisc *sch,
584 struct sk_buff **to_free)
586 int uninitialized_var(ret);
587 unsigned int len = qdisc_pkt_len(skb);
588 struct htb_sched *q = qdisc_priv(sch);
589 struct htb_class *cl = htb_classify(skb, sch, &ret);
591 if (cl == HTB_DIRECT) {
592 /* enqueue to helper queue */
593 if (q->direct_queue.qlen < q->direct_qlen) {
594 __qdisc_enqueue_tail(skb, &q->direct_queue);
597 return qdisc_drop(skb, sch, to_free);
599 #ifdef CONFIG_NET_CLS_ACT
601 if (ret & __NET_XMIT_BYPASS)
602 qdisc_qstats_drop(sch);
603 __qdisc_drop(skb, to_free);
606 } else if ((ret = qdisc_enqueue(skb, cl->leaf.q,
607 to_free)) != NET_XMIT_SUCCESS) {
608 if (net_xmit_drop_count(ret)) {
609 qdisc_qstats_drop(sch);
617 sch->qstats.backlog += len;
619 return NET_XMIT_SUCCESS;
622 static inline void htb_accnt_tokens(struct htb_class *cl, int bytes, s64 diff)
624 s64 toks = diff + cl->tokens;
626 if (toks > cl->buffer)
628 toks -= (s64) psched_l2t_ns(&cl->rate, bytes);
629 if (toks <= -cl->mbuffer)
630 toks = 1 - cl->mbuffer;
635 static inline void htb_accnt_ctokens(struct htb_class *cl, int bytes, s64 diff)
637 s64 toks = diff + cl->ctokens;
639 if (toks > cl->cbuffer)
641 toks -= (s64) psched_l2t_ns(&cl->ceil, bytes);
642 if (toks <= -cl->mbuffer)
643 toks = 1 - cl->mbuffer;
649 * htb_charge_class - charges amount "bytes" to leaf and ancestors
651 * Routine assumes that packet "bytes" long was dequeued from leaf cl
652 * borrowing from "level". It accounts bytes to ceil leaky bucket for
653 * leaf and all ancestors and to rate bucket for ancestors at levels
654 * "level" and higher. It also handles possible change of mode resulting
655 * from the update. Note that mode can also increase here (MAY_BORROW to
656 * CAN_SEND) because we can use more precise clock that event queue here.
657 * In such case we remove class from event queue first.
659 static void htb_charge_class(struct htb_sched *q, struct htb_class *cl,
660 int level, struct sk_buff *skb)
662 int bytes = qdisc_pkt_len(skb);
663 enum htb_cmode old_mode;
667 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
668 if (cl->level >= level) {
669 if (cl->level == level)
671 htb_accnt_tokens(cl, bytes, diff);
673 cl->xstats.borrows++;
674 cl->tokens += diff; /* we moved t_c; update tokens */
676 htb_accnt_ctokens(cl, bytes, diff);
679 old_mode = cl->cmode;
681 htb_change_class_mode(q, cl, &diff);
682 if (old_mode != cl->cmode) {
683 if (old_mode != HTB_CAN_SEND)
684 htb_safe_rb_erase(&cl->pq_node, &q->hlevel[cl->level].wait_pq);
685 if (cl->cmode != HTB_CAN_SEND)
686 htb_add_to_wait_tree(q, cl, diff);
689 /* update basic stats except for leaves which are already updated */
691 bstats_update(&cl->bstats, skb);
698 * htb_do_events - make mode changes to classes at the level
700 * Scans event queue for pending events and applies them. Returns time of
701 * next pending event (0 for no event in pq, q->now for too many events).
702 * Note: Applied are events whose have cl->pq_key <= q->now.
704 static s64 htb_do_events(struct htb_sched *q, const int level,
707 /* don't run for longer than 2 jiffies; 2 is used instead of
708 * 1 to simplify things when jiffy is going to be incremented
711 unsigned long stop_at = start + 2;
712 struct rb_root *wait_pq = &q->hlevel[level].wait_pq;
714 while (time_before(jiffies, stop_at)) {
715 struct htb_class *cl;
717 struct rb_node *p = rb_first(wait_pq);
722 cl = rb_entry(p, struct htb_class, pq_node);
723 if (cl->pq_key > q->now)
726 htb_safe_rb_erase(p, wait_pq);
727 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
728 htb_change_class_mode(q, cl, &diff);
729 if (cl->cmode != HTB_CAN_SEND)
730 htb_add_to_wait_tree(q, cl, diff);
733 /* too much load - let's continue after a break for scheduling */
734 if (!(q->warned & HTB_WARN_TOOMANYEVENTS)) {
735 pr_warn("htb: too many events!\n");
736 q->warned |= HTB_WARN_TOOMANYEVENTS;
742 /* Returns class->node+prio from id-tree where classe's id is >= id. NULL
743 * is no such one exists.
745 static struct rb_node *htb_id_find_next_upper(int prio, struct rb_node *n,
748 struct rb_node *r = NULL;
750 struct htb_class *cl =
751 rb_entry(n, struct htb_class, node[prio]);
753 if (id > cl->common.classid) {
755 } else if (id < cl->common.classid) {
766 * htb_lookup_leaf - returns next leaf class in DRR order
768 * Find leaf where current feed pointers points to.
770 static struct htb_class *htb_lookup_leaf(struct htb_prio *hprio, const int prio)
774 struct rb_node *root;
775 struct rb_node **pptr;
777 } stk[TC_HTB_MAXDEPTH], *sp = stk;
779 BUG_ON(!hprio->row.rb_node);
780 sp->root = hprio->row.rb_node;
781 sp->pptr = &hprio->ptr;
782 sp->pid = &hprio->last_ptr_id;
784 for (i = 0; i < 65535; i++) {
785 if (!*sp->pptr && *sp->pid) {
786 /* ptr was invalidated but id is valid - try to recover
787 * the original or next ptr
790 htb_id_find_next_upper(prio, sp->root, *sp->pid);
792 *sp->pid = 0; /* ptr is valid now so that remove this hint as it
793 * can become out of date quickly
795 if (!*sp->pptr) { /* we are at right end; rewind & go up */
796 *sp->pptr = sp->root;
797 while ((*sp->pptr)->rb_left)
798 *sp->pptr = (*sp->pptr)->rb_left;
805 htb_next_rb_node(sp->pptr);
808 struct htb_class *cl;
809 struct htb_prio *clp;
811 cl = rb_entry(*sp->pptr, struct htb_class, node[prio]);
814 clp = &cl->inner.clprio[prio];
815 (++sp)->root = clp->feed.rb_node;
816 sp->pptr = &clp->ptr;
817 sp->pid = &clp->last_ptr_id;
824 /* dequeues packet at given priority and level; call only if
825 * you are sure that there is active class at prio/level
827 static struct sk_buff *htb_dequeue_tree(struct htb_sched *q, const int prio,
830 struct sk_buff *skb = NULL;
831 struct htb_class *cl, *start;
832 struct htb_level *hlevel = &q->hlevel[level];
833 struct htb_prio *hprio = &hlevel->hprio[prio];
835 /* look initial class up in the row */
836 start = cl = htb_lookup_leaf(hprio, prio);
843 /* class can be empty - it is unlikely but can be true if leaf
844 * qdisc drops packets in enqueue routine or if someone used
845 * graft operation on the leaf since last dequeue;
846 * simply deactivate and skip such class
848 if (unlikely(cl->leaf.q->q.qlen == 0)) {
849 struct htb_class *next;
850 htb_deactivate(q, cl);
852 /* row/level might become empty */
853 if ((q->row_mask[level] & (1 << prio)) == 0)
856 next = htb_lookup_leaf(hprio, prio);
858 if (cl == start) /* fix start if we just deleted it */
864 skb = cl->leaf.q->dequeue(cl->leaf.q);
865 if (likely(skb != NULL))
868 qdisc_warn_nonwc("htb", cl->leaf.q);
869 htb_next_rb_node(level ? &cl->parent->inner.clprio[prio].ptr:
870 &q->hlevel[0].hprio[prio].ptr);
871 cl = htb_lookup_leaf(hprio, prio);
873 } while (cl != start);
875 if (likely(skb != NULL)) {
876 bstats_update(&cl->bstats, skb);
877 cl->leaf.deficit[level] -= qdisc_pkt_len(skb);
878 if (cl->leaf.deficit[level] < 0) {
879 cl->leaf.deficit[level] += cl->quantum;
880 htb_next_rb_node(level ? &cl->parent->inner.clprio[prio].ptr :
881 &q->hlevel[0].hprio[prio].ptr);
883 /* this used to be after charge_class but this constelation
884 * gives us slightly better performance
886 if (!cl->leaf.q->q.qlen)
887 htb_deactivate(q, cl);
888 htb_charge_class(q, cl, level, skb);
893 static struct sk_buff *htb_dequeue(struct Qdisc *sch)
896 struct htb_sched *q = qdisc_priv(sch);
899 unsigned long start_at;
901 /* try to dequeue direct packets as high prio (!) to minimize cpu work */
902 skb = __qdisc_dequeue_head(&q->direct_queue);
905 qdisc_bstats_update(sch, skb);
906 qdisc_qstats_backlog_dec(sch, skb);
913 q->now = ktime_get_ns();
916 next_event = q->now + 5LLU * NSEC_PER_SEC;
918 for (level = 0; level < TC_HTB_MAXDEPTH; level++) {
919 /* common case optimization - skip event handler quickly */
921 s64 event = q->near_ev_cache[level];
923 if (q->now >= event) {
924 event = htb_do_events(q, level, start_at);
926 event = q->now + NSEC_PER_SEC;
927 q->near_ev_cache[level] = event;
930 if (next_event > event)
933 m = ~q->row_mask[level];
934 while (m != (int)(-1)) {
938 skb = htb_dequeue_tree(q, prio, level);
939 if (likely(skb != NULL))
943 if (likely(next_event > q->now))
944 qdisc_watchdog_schedule_ns(&q->watchdog, next_event);
946 schedule_work(&q->work);
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;
959 for (i = 0; i < q->clhash.hashsize; i++) {
960 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
962 memset(&cl->inner, 0, sizeof(cl->inner));
965 qdisc_reset(cl->leaf.q);
967 cl->prio_activity = 0;
968 cl->cmode = HTB_CAN_SEND;
971 qdisc_watchdog_cancel(&q->watchdog);
972 __qdisc_reset_queue(&q->direct_queue);
974 sch->qstats.backlog = 0;
975 memset(q->hlevel, 0, sizeof(q->hlevel));
976 memset(q->row_mask, 0, sizeof(q->row_mask));
979 static const struct nla_policy htb_policy[TCA_HTB_MAX + 1] = {
980 [TCA_HTB_PARMS] = { .len = sizeof(struct tc_htb_opt) },
981 [TCA_HTB_INIT] = { .len = sizeof(struct tc_htb_glob) },
982 [TCA_HTB_CTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
983 [TCA_HTB_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
984 [TCA_HTB_DIRECT_QLEN] = { .type = NLA_U32 },
985 [TCA_HTB_RATE64] = { .type = NLA_U64 },
986 [TCA_HTB_CEIL64] = { .type = NLA_U64 },
989 static void htb_work_func(struct work_struct *work)
991 struct htb_sched *q = container_of(work, struct htb_sched, work);
992 struct Qdisc *sch = q->watchdog.qdisc;
995 __netif_schedule(qdisc_root(sch));
999 static int htb_init(struct Qdisc *sch, struct nlattr *opt,
1000 struct netlink_ext_ack *extack)
1002 struct htb_sched *q = qdisc_priv(sch);
1003 struct nlattr *tb[TCA_HTB_MAX + 1];
1004 struct tc_htb_glob *gopt;
1007 qdisc_watchdog_init(&q->watchdog, sch);
1008 INIT_WORK(&q->work, htb_work_func);
1013 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
1017 err = nla_parse_nested_deprecated(tb, TCA_HTB_MAX, opt, htb_policy,
1022 if (!tb[TCA_HTB_INIT])
1025 gopt = nla_data(tb[TCA_HTB_INIT]);
1026 if (gopt->version != HTB_VER >> 16)
1029 err = qdisc_class_hash_init(&q->clhash);
1033 qdisc_skb_head_init(&q->direct_queue);
1035 if (tb[TCA_HTB_DIRECT_QLEN])
1036 q->direct_qlen = nla_get_u32(tb[TCA_HTB_DIRECT_QLEN]);
1038 q->direct_qlen = qdisc_dev(sch)->tx_queue_len;
1040 if ((q->rate2quantum = gopt->rate2quantum) < 1)
1041 q->rate2quantum = 1;
1042 q->defcls = gopt->defcls;
1047 static int htb_dump(struct Qdisc *sch, struct sk_buff *skb)
1049 struct htb_sched *q = qdisc_priv(sch);
1050 struct nlattr *nest;
1051 struct tc_htb_glob gopt;
1053 sch->qstats.overlimits = q->overlimits;
1054 /* Its safe to not acquire qdisc lock. As we hold RTNL,
1055 * no change can happen on the qdisc parameters.
1058 gopt.direct_pkts = q->direct_pkts;
1059 gopt.version = HTB_VER;
1060 gopt.rate2quantum = q->rate2quantum;
1061 gopt.defcls = q->defcls;
1064 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1066 goto nla_put_failure;
1067 if (nla_put(skb, TCA_HTB_INIT, sizeof(gopt), &gopt) ||
1068 nla_put_u32(skb, TCA_HTB_DIRECT_QLEN, q->direct_qlen))
1069 goto nla_put_failure;
1071 return nla_nest_end(skb, nest);
1074 nla_nest_cancel(skb, nest);
1078 static int htb_dump_class(struct Qdisc *sch, unsigned long arg,
1079 struct sk_buff *skb, struct tcmsg *tcm)
1081 struct htb_class *cl = (struct htb_class *)arg;
1082 struct nlattr *nest;
1083 struct tc_htb_opt opt;
1085 /* Its safe to not acquire qdisc lock. As we hold RTNL,
1086 * no change can happen on the class parameters.
1088 tcm->tcm_parent = cl->parent ? cl->parent->common.classid : TC_H_ROOT;
1089 tcm->tcm_handle = cl->common.classid;
1090 if (!cl->level && cl->leaf.q)
1091 tcm->tcm_info = cl->leaf.q->handle;
1093 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1095 goto nla_put_failure;
1097 memset(&opt, 0, sizeof(opt));
1099 psched_ratecfg_getrate(&opt.rate, &cl->rate);
1100 opt.buffer = PSCHED_NS2TICKS(cl->buffer);
1101 psched_ratecfg_getrate(&opt.ceil, &cl->ceil);
1102 opt.cbuffer = PSCHED_NS2TICKS(cl->cbuffer);
1103 opt.quantum = cl->quantum;
1104 opt.prio = cl->prio;
1105 opt.level = cl->level;
1106 if (nla_put(skb, TCA_HTB_PARMS, sizeof(opt), &opt))
1107 goto nla_put_failure;
1108 if ((cl->rate.rate_bytes_ps >= (1ULL << 32)) &&
1109 nla_put_u64_64bit(skb, TCA_HTB_RATE64, cl->rate.rate_bytes_ps,
1111 goto nla_put_failure;
1112 if ((cl->ceil.rate_bytes_ps >= (1ULL << 32)) &&
1113 nla_put_u64_64bit(skb, TCA_HTB_CEIL64, cl->ceil.rate_bytes_ps,
1115 goto nla_put_failure;
1117 return nla_nest_end(skb, nest);
1120 nla_nest_cancel(skb, nest);
1125 htb_dump_class_stats(struct Qdisc *sch, unsigned long arg, struct gnet_dump *d)
1127 struct htb_class *cl = (struct htb_class *)arg;
1128 struct gnet_stats_queue qs = {
1130 .overlimits = cl->overlimits,
1134 if (!cl->level && cl->leaf.q)
1135 qdisc_qstats_qlen_backlog(cl->leaf.q, &qlen, &qs.backlog);
1137 cl->xstats.tokens = clamp_t(s64, PSCHED_NS2TICKS(cl->tokens),
1139 cl->xstats.ctokens = clamp_t(s64, PSCHED_NS2TICKS(cl->ctokens),
1142 if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
1143 d, NULL, &cl->bstats) < 0 ||
1144 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1145 gnet_stats_copy_queue(d, NULL, &qs, qlen) < 0)
1148 return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1151 static int htb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1152 struct Qdisc **old, struct netlink_ext_ack *extack)
1154 struct htb_class *cl = (struct htb_class *)arg;
1159 (new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1160 cl->common.classid, extack)) == NULL)
1163 *old = qdisc_replace(sch, new, &cl->leaf.q);
1167 static struct Qdisc *htb_leaf(struct Qdisc *sch, unsigned long arg)
1169 struct htb_class *cl = (struct htb_class *)arg;
1170 return !cl->level ? cl->leaf.q : NULL;
1173 static void htb_qlen_notify(struct Qdisc *sch, unsigned long arg)
1175 struct htb_class *cl = (struct htb_class *)arg;
1177 htb_deactivate(qdisc_priv(sch), cl);
1180 static inline int htb_parent_last_child(struct htb_class *cl)
1183 /* the root class */
1185 if (cl->parent->children > 1)
1186 /* not the last child */
1191 static void htb_parent_to_leaf(struct htb_sched *q, struct htb_class *cl,
1192 struct Qdisc *new_q)
1194 struct htb_class *parent = cl->parent;
1196 WARN_ON(cl->level || !cl->leaf.q || cl->prio_activity);
1198 if (parent->cmode != HTB_CAN_SEND)
1199 htb_safe_rb_erase(&parent->pq_node,
1200 &q->hlevel[parent->level].wait_pq);
1203 memset(&parent->inner, 0, sizeof(parent->inner));
1204 parent->leaf.q = new_q ? new_q : &noop_qdisc;
1205 parent->tokens = parent->buffer;
1206 parent->ctokens = parent->cbuffer;
1207 parent->t_c = ktime_get_ns();
1208 parent->cmode = HTB_CAN_SEND;
1211 static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
1214 WARN_ON(!cl->leaf.q);
1215 qdisc_put(cl->leaf.q);
1217 gen_kill_estimator(&cl->rate_est);
1218 tcf_block_put(cl->block);
1222 static void htb_destroy(struct Qdisc *sch)
1224 struct htb_sched *q = qdisc_priv(sch);
1225 struct hlist_node *next;
1226 struct htb_class *cl;
1229 cancel_work_sync(&q->work);
1230 qdisc_watchdog_cancel(&q->watchdog);
1231 /* This line used to be after htb_destroy_class call below
1232 * and surprisingly it worked in 2.4. But it must precede it
1233 * because filter need its target class alive to be able to call
1234 * unbind_filter on it (without Oops).
1236 tcf_block_put(q->block);
1238 for (i = 0; i < q->clhash.hashsize; i++) {
1239 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1240 tcf_block_put(cl->block);
1244 for (i = 0; i < q->clhash.hashsize; i++) {
1245 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1247 htb_destroy_class(sch, cl);
1249 qdisc_class_hash_destroy(&q->clhash);
1250 __qdisc_reset_queue(&q->direct_queue);
1253 static int htb_delete(struct Qdisc *sch, unsigned long arg)
1255 struct htb_sched *q = qdisc_priv(sch);
1256 struct htb_class *cl = (struct htb_class *)arg;
1257 struct Qdisc *new_q = NULL;
1260 /* TODO: why don't allow to delete subtree ? references ? does
1261 * tc subsys guarantee us that in htb_destroy it holds no class
1262 * refs so that we can remove children safely there ?
1264 if (cl->children || cl->filter_cnt)
1267 if (!cl->level && htb_parent_last_child(cl)) {
1268 new_q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1269 cl->parent->common.classid,
1277 qdisc_purge_queue(cl->leaf.q);
1279 /* delete from hash and active; remainder in destroy_class */
1280 qdisc_class_hash_remove(&q->clhash, &cl->common);
1282 cl->parent->children--;
1284 if (cl->prio_activity)
1285 htb_deactivate(q, cl);
1287 if (cl->cmode != HTB_CAN_SEND)
1288 htb_safe_rb_erase(&cl->pq_node,
1289 &q->hlevel[cl->level].wait_pq);
1292 htb_parent_to_leaf(q, cl, new_q);
1294 sch_tree_unlock(sch);
1296 htb_destroy_class(sch, cl);
1300 static int htb_change_class(struct Qdisc *sch, u32 classid,
1301 u32 parentid, struct nlattr **tca,
1302 unsigned long *arg, struct netlink_ext_ack *extack)
1305 struct htb_sched *q = qdisc_priv(sch);
1306 struct htb_class *cl = (struct htb_class *)*arg, *parent;
1307 struct nlattr *opt = tca[TCA_OPTIONS];
1308 struct nlattr *tb[TCA_HTB_MAX + 1];
1309 struct tc_htb_opt *hopt;
1313 /* extract all subattrs from opt attr */
1317 err = nla_parse_nested_deprecated(tb, TCA_HTB_MAX, opt, htb_policy,
1323 if (tb[TCA_HTB_PARMS] == NULL)
1326 parent = parentid == TC_H_ROOT ? NULL : htb_find(parentid, sch);
1328 hopt = nla_data(tb[TCA_HTB_PARMS]);
1329 if (!hopt->rate.rate || !hopt->ceil.rate)
1332 /* Keeping backward compatible with rate_table based iproute2 tc */
1333 if (hopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
1334 qdisc_put_rtab(qdisc_get_rtab(&hopt->rate, tb[TCA_HTB_RTAB],
1337 if (hopt->ceil.linklayer == TC_LINKLAYER_UNAWARE)
1338 qdisc_put_rtab(qdisc_get_rtab(&hopt->ceil, tb[TCA_HTB_CTAB],
1341 if (!cl) { /* new class */
1342 struct Qdisc *new_q;
1346 struct gnet_estimator opt;
1349 .nla_len = nla_attr_size(sizeof(est.opt)),
1350 .nla_type = TCA_RATE,
1353 /* 4s interval, 16s averaging constant */
1359 /* check for valid classid */
1360 if (!classid || TC_H_MAJ(classid ^ sch->handle) ||
1361 htb_find(classid, sch))
1364 /* check maximal depth */
1365 if (parent && parent->parent && parent->parent->level < 2) {
1366 pr_err("htb: tree is too deep\n");
1370 cl = kzalloc(sizeof(*cl), GFP_KERNEL);
1374 err = tcf_block_get(&cl->block, &cl->filter_list, sch, extack);
1379 if (htb_rate_est || tca[TCA_RATE]) {
1380 err = gen_new_estimator(&cl->bstats, NULL,
1383 qdisc_root_sleeping_running(sch),
1384 tca[TCA_RATE] ? : &est.nla);
1386 tcf_block_put(cl->block);
1393 RB_CLEAR_NODE(&cl->pq_node);
1395 for (prio = 0; prio < TC_HTB_NUMPRIO; prio++)
1396 RB_CLEAR_NODE(&cl->node[prio]);
1398 /* create leaf qdisc early because it uses kmalloc(GFP_KERNEL)
1399 * so that can't be used inside of sch_tree_lock
1400 * -- thanks to Karlis Peisenieks
1402 new_q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1405 if (parent && !parent->level) {
1406 /* turn parent into inner node */
1407 qdisc_purge_queue(parent->leaf.q);
1408 qdisc_put(parent->leaf.q);
1409 if (parent->prio_activity)
1410 htb_deactivate(q, parent);
1412 /* remove from evt list because of level change */
1413 if (parent->cmode != HTB_CAN_SEND) {
1414 htb_safe_rb_erase(&parent->pq_node, &q->hlevel[0].wait_pq);
1415 parent->cmode = HTB_CAN_SEND;
1417 parent->level = (parent->parent ? parent->parent->level
1418 : TC_HTB_MAXDEPTH) - 1;
1419 memset(&parent->inner, 0, sizeof(parent->inner));
1421 /* leaf (we) needs elementary qdisc */
1422 cl->leaf.q = new_q ? new_q : &noop_qdisc;
1424 cl->common.classid = classid;
1425 cl->parent = parent;
1427 /* set class to be in HTB_CAN_SEND state */
1428 cl->tokens = PSCHED_TICKS2NS(hopt->buffer);
1429 cl->ctokens = PSCHED_TICKS2NS(hopt->cbuffer);
1430 cl->mbuffer = 60ULL * NSEC_PER_SEC; /* 1min */
1431 cl->t_c = ktime_get_ns();
1432 cl->cmode = HTB_CAN_SEND;
1434 /* attach to the hash list and parent's family */
1435 qdisc_class_hash_insert(&q->clhash, &cl->common);
1438 if (cl->leaf.q != &noop_qdisc)
1439 qdisc_hash_add(cl->leaf.q, true);
1441 if (tca[TCA_RATE]) {
1442 err = gen_replace_estimator(&cl->bstats, NULL,
1445 qdisc_root_sleeping_running(sch),
1453 rate64 = tb[TCA_HTB_RATE64] ? nla_get_u64(tb[TCA_HTB_RATE64]) : 0;
1455 ceil64 = tb[TCA_HTB_CEIL64] ? nla_get_u64(tb[TCA_HTB_CEIL64]) : 0;
1457 psched_ratecfg_precompute(&cl->rate, &hopt->rate, rate64);
1458 psched_ratecfg_precompute(&cl->ceil, &hopt->ceil, ceil64);
1460 /* it used to be a nasty bug here, we have to check that node
1461 * is really leaf before changing cl->leaf !
1464 u64 quantum = cl->rate.rate_bytes_ps;
1466 do_div(quantum, q->rate2quantum);
1467 cl->quantum = min_t(u64, quantum, INT_MAX);
1469 if (!hopt->quantum && cl->quantum < 1000) {
1473 if (!hopt->quantum && cl->quantum > 200000) {
1475 cl->quantum = 200000;
1478 cl->quantum = hopt->quantum;
1479 if ((cl->prio = hopt->prio) >= TC_HTB_NUMPRIO)
1480 cl->prio = TC_HTB_NUMPRIO - 1;
1483 cl->buffer = PSCHED_TICKS2NS(hopt->buffer);
1484 cl->cbuffer = PSCHED_TICKS2NS(hopt->cbuffer);
1486 sch_tree_unlock(sch);
1489 pr_warn("HTB: quantum of class %X is %s. Consider r2q change.\n",
1490 cl->common.classid, (warn == -1 ? "small" : "big"));
1492 qdisc_class_hash_grow(sch, &q->clhash);
1494 *arg = (unsigned long)cl;
1501 static struct tcf_block *htb_tcf_block(struct Qdisc *sch, unsigned long arg,
1502 struct netlink_ext_ack *extack)
1504 struct htb_sched *q = qdisc_priv(sch);
1505 struct htb_class *cl = (struct htb_class *)arg;
1507 return cl ? cl->block : q->block;
1510 static unsigned long htb_bind_filter(struct Qdisc *sch, unsigned long parent,
1513 struct htb_class *cl = htb_find(classid, sch);
1515 /*if (cl && !cl->level) return 0;
1516 * The line above used to be there to prevent attaching filters to
1517 * leaves. But at least tc_index filter uses this just to get class
1518 * for other reasons so that we have to allow for it.
1520 * 19.6.2002 As Werner explained it is ok - bind filter is just
1521 * another way to "lock" the class - unlike "get" this lock can
1522 * be broken by class during destroy IIUC.
1526 return (unsigned long)cl;
1529 static void htb_unbind_filter(struct Qdisc *sch, unsigned long arg)
1531 struct htb_class *cl = (struct htb_class *)arg;
1537 static void htb_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1539 struct htb_sched *q = qdisc_priv(sch);
1540 struct htb_class *cl;
1546 for (i = 0; i < q->clhash.hashsize; i++) {
1547 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1548 if (arg->count < arg->skip) {
1552 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1561 static const struct Qdisc_class_ops htb_class_ops = {
1564 .qlen_notify = htb_qlen_notify,
1566 .change = htb_change_class,
1567 .delete = htb_delete,
1569 .tcf_block = htb_tcf_block,
1570 .bind_tcf = htb_bind_filter,
1571 .unbind_tcf = htb_unbind_filter,
1572 .dump = htb_dump_class,
1573 .dump_stats = htb_dump_class_stats,
1576 static struct Qdisc_ops htb_qdisc_ops __read_mostly = {
1577 .cl_ops = &htb_class_ops,
1579 .priv_size = sizeof(struct htb_sched),
1580 .enqueue = htb_enqueue,
1581 .dequeue = htb_dequeue,
1582 .peek = qdisc_peek_dequeued,
1585 .destroy = htb_destroy,
1587 .owner = THIS_MODULE,
1590 static int __init htb_module_init(void)
1592 return register_qdisc(&htb_qdisc_ops);
1594 static void __exit htb_module_exit(void)
1596 unregister_qdisc(&htb_qdisc_ops);
1599 module_init(htb_module_init)
1600 module_exit(htb_module_exit)
1601 MODULE_LICENSE("GPL");