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>
46 ========================================================================
47 HTB is like TBF with multiple classes. It is also similar to CBQ because
48 it allows to assign priority to each class in hierarchy.
49 In fact it is another implementation of Floyd's formal sharing.
52 Each class is assigned level. Leaf has ALWAYS level 0 and root
53 classes have level TC_HTB_MAXDEPTH-1. Interior nodes has level
54 one less than their parent.
57 static int htb_hysteresis __read_mostly = 0; /* whether to use mode hysteresis for speedup */
58 #define HTB_VER 0x30011 /* major must be matched with number suplied by TC as version */
60 #if HTB_VER >> 16 != TC_HTB_PROTOVER
61 #error "Mismatched sch_htb.c and pkt_sch.h"
64 /* Module parameter and sysfs export */
65 module_param (htb_hysteresis, int, 0640);
66 MODULE_PARM_DESC(htb_hysteresis, "Hysteresis mode, less CPU load, less accurate");
68 static int htb_rate_est = 0; /* htb classes have a default rate estimator */
69 module_param(htb_rate_est, int, 0640);
70 MODULE_PARM_DESC(htb_rate_est, "setup a default rate estimator (4sec 16sec) for htb classes");
72 /* used internaly to keep status of single class */
74 HTB_CANT_SEND, /* class can't send and can't borrow */
75 HTB_MAY_BORROW, /* class can't send but may borrow */
76 HTB_CAN_SEND /* class can send */
85 /* When class changes from state 1->2 and disconnects from
86 * parent's feed then we lost ptr value and start from the
87 * first child again. Here we store classid of the
88 * last valid ptr (used when ptr is NULL).
93 /* interior & leaf nodes; props specific to leaves are marked L:
94 * To reduce false sharing, place mostly read fields at beginning,
95 * and mostly written ones at the end.
98 struct Qdisc_class_common common;
99 struct psched_ratecfg rate;
100 struct psched_ratecfg ceil;
101 s64 buffer, cbuffer;/* token bucket depth/rate */
102 s64 mbuffer; /* max wait time */
103 u32 prio; /* these two are used only by leaves... */
104 int quantum; /* but stored for parent-to-leaf return */
106 struct tcf_proto *filter_list; /* class attached filters */
108 int refcnt; /* usage count of this class */
110 int level; /* our level (see above) */
111 unsigned int children;
112 struct htb_class *parent; /* parent class */
114 struct gnet_stats_rate_est64 rate_est;
117 * Written often fields
119 struct gnet_stats_basic_packed bstats;
120 struct gnet_stats_queue qstats;
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 struct list_head drop_list;
130 int deficit[TC_HTB_MAXDEPTH];
133 struct htb_class_inner {
134 struct htb_prio clprio[TC_HTB_NUMPRIO];
139 int prio_activity; /* for which prios are we active */
140 enum htb_cmode cmode; /* current mode of the class */
141 struct rb_node pq_node; /* node for event queue */
142 struct rb_node node[TC_HTB_NUMPRIO]; /* node for self or feed tree */
146 struct rb_root wait_pq;
147 struct htb_prio hprio[TC_HTB_NUMPRIO];
151 struct Qdisc_class_hash clhash;
152 int defcls; /* class where unclassified flows go to */
153 int rate2quantum; /* quant = rate / rate2quantum */
155 /* filters for qdisc itself */
156 struct tcf_proto *filter_list;
158 #define HTB_WARN_TOOMANYEVENTS 0x1
159 unsigned int warned; /* only one warning */
161 struct work_struct work;
163 /* non shaped skbs; let them go directly thru */
164 struct sk_buff_head direct_queue;
167 struct qdisc_watchdog watchdog;
169 s64 now; /* cached dequeue time */
170 struct list_head drops[TC_HTB_NUMPRIO];/* active leaves (for drops) */
172 /* time of nearest event per level (row) */
173 s64 near_ev_cache[TC_HTB_MAXDEPTH];
175 int row_mask[TC_HTB_MAXDEPTH];
177 struct htb_level hlevel[TC_HTB_MAXDEPTH];
180 /* find class in global hash table using given handle */
181 static inline struct htb_class *htb_find(u32 handle, struct Qdisc *sch)
183 struct htb_sched *q = qdisc_priv(sch);
184 struct Qdisc_class_common *clc;
186 clc = qdisc_class_find(&q->clhash, handle);
189 return container_of(clc, struct htb_class, common);
193 * htb_classify - classify a packet into class
195 * It returns NULL if the packet should be dropped or -1 if the packet
196 * should be passed directly thru. In all other cases leaf class is returned.
197 * We allow direct class selection by classid in priority. The we examine
198 * filters in qdisc and in inner nodes (if higher filter points to the inner
199 * node). If we end up with classid MAJOR:0 we enqueue the skb into special
200 * internal fifo (direct). These packets then go directly thru. If we still
201 * have no valid leaf we try to use MAJOR:default leaf. It still unsuccessful
202 * then finish and return direct queue.
204 #define HTB_DIRECT ((struct htb_class *)-1L)
206 static struct htb_class *htb_classify(struct sk_buff *skb, struct Qdisc *sch,
209 struct htb_sched *q = qdisc_priv(sch);
210 struct htb_class *cl;
211 struct tcf_result res;
212 struct tcf_proto *tcf;
215 /* allow to select class by setting skb->priority to valid classid;
216 * note that nfmark can be used too by attaching filter fw with no
219 if (skb->priority == sch->handle)
220 return HTB_DIRECT; /* X:0 (direct flow) selected */
221 cl = htb_find(skb->priority, sch);
225 /* Start with inner filter chain if a non-leaf class is selected */
226 tcf = cl->filter_list;
228 tcf = q->filter_list;
231 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
232 while (tcf && (result = tc_classify(skb, tcf, &res)) >= 0) {
233 #ifdef CONFIG_NET_CLS_ACT
237 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
242 cl = (void *)res.class;
244 if (res.classid == sch->handle)
245 return HTB_DIRECT; /* X:0 (direct flow) */
246 cl = htb_find(res.classid, sch);
248 break; /* filter selected invalid classid */
251 return cl; /* we hit leaf; return it */
253 /* we have got inner class; apply inner filter chain */
254 tcf = cl->filter_list;
256 /* classification failed; try to use default class */
257 cl = htb_find(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
258 if (!cl || cl->level)
259 return HTB_DIRECT; /* bad default .. this is safe bet */
264 * htb_add_to_id_tree - adds class to the round robin list
266 * Routine adds class to the list (actually tree) sorted by classid.
267 * Make sure that class is not already on such list for given prio.
269 static void htb_add_to_id_tree(struct rb_root *root,
270 struct htb_class *cl, int prio)
272 struct rb_node **p = &root->rb_node, *parent = NULL;
277 c = rb_entry(parent, struct htb_class, node[prio]);
279 if (cl->common.classid > c->common.classid)
280 p = &parent->rb_right;
282 p = &parent->rb_left;
284 rb_link_node(&cl->node[prio], parent, p);
285 rb_insert_color(&cl->node[prio], root);
289 * htb_add_to_wait_tree - adds class to the event queue with delay
291 * The class is added to priority event queue to indicate that class will
292 * change its mode in cl->pq_key microseconds. Make sure that class is not
293 * already in the queue.
295 static void htb_add_to_wait_tree(struct htb_sched *q,
296 struct htb_class *cl, s64 delay)
298 struct rb_node **p = &q->hlevel[cl->level].wait_pq.rb_node, *parent = NULL;
300 cl->pq_key = q->now + delay;
301 if (cl->pq_key == q->now)
304 /* update the nearest event cache */
305 if (q->near_ev_cache[cl->level] > cl->pq_key)
306 q->near_ev_cache[cl->level] = cl->pq_key;
311 c = rb_entry(parent, struct htb_class, pq_node);
312 if (cl->pq_key >= c->pq_key)
313 p = &parent->rb_right;
315 p = &parent->rb_left;
317 rb_link_node(&cl->pq_node, parent, p);
318 rb_insert_color(&cl->pq_node, &q->hlevel[cl->level].wait_pq);
322 * htb_next_rb_node - finds next node in binary tree
324 * When we are past last key we return NULL.
325 * Average complexity is 2 steps per call.
327 static inline void htb_next_rb_node(struct rb_node **n)
333 * htb_add_class_to_row - add class to its row
335 * The class is added to row at priorities marked in mask.
336 * It does nothing if mask == 0.
338 static inline void htb_add_class_to_row(struct htb_sched *q,
339 struct htb_class *cl, int mask)
341 q->row_mask[cl->level] |= mask;
343 int prio = ffz(~mask);
344 mask &= ~(1 << prio);
345 htb_add_to_id_tree(&q->hlevel[cl->level].hprio[prio].row, cl, prio);
349 /* If this triggers, it is a bug in this code, but it need not be fatal */
350 static void htb_safe_rb_erase(struct rb_node *rb, struct rb_root *root)
352 if (RB_EMPTY_NODE(rb)) {
362 * htb_remove_class_from_row - removes class from its row
364 * The class is removed from row at priorities marked in mask.
365 * It does nothing if mask == 0.
367 static inline void htb_remove_class_from_row(struct htb_sched *q,
368 struct htb_class *cl, int mask)
371 struct htb_level *hlevel = &q->hlevel[cl->level];
374 int prio = ffz(~mask);
375 struct htb_prio *hprio = &hlevel->hprio[prio];
377 mask &= ~(1 << prio);
378 if (hprio->ptr == cl->node + prio)
379 htb_next_rb_node(&hprio->ptr);
381 htb_safe_rb_erase(cl->node + prio, &hprio->row);
382 if (!hprio->row.rb_node)
385 q->row_mask[cl->level] &= ~m;
389 * htb_activate_prios - creates active classe's feed chain
391 * The class is connected to ancestors and/or appropriate rows
392 * for priorities it is participating on. cl->cmode must be new
393 * (activated) mode. It does nothing if cl->prio_activity == 0.
395 static void htb_activate_prios(struct htb_sched *q, struct htb_class *cl)
397 struct htb_class *p = cl->parent;
398 long m, mask = cl->prio_activity;
400 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
406 if (p->un.inner.clprio[prio].feed.rb_node)
407 /* parent already has its feed in use so that
408 * reset bit in mask as parent is already ok
410 mask &= ~(1 << prio);
412 htb_add_to_id_tree(&p->un.inner.clprio[prio].feed, cl, prio);
414 p->prio_activity |= mask;
419 if (cl->cmode == HTB_CAN_SEND && mask)
420 htb_add_class_to_row(q, cl, mask);
424 * htb_deactivate_prios - remove class from feed chain
426 * cl->cmode must represent old mode (before deactivation). It does
427 * nothing if cl->prio_activity == 0. Class is removed from all feed
430 static void htb_deactivate_prios(struct htb_sched *q, struct htb_class *cl)
432 struct htb_class *p = cl->parent;
433 long m, mask = cl->prio_activity;
435 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
442 if (p->un.inner.clprio[prio].ptr == cl->node + prio) {
443 /* we are removing child which is pointed to from
444 * parent feed - forget the pointer but remember
447 p->un.inner.clprio[prio].last_ptr_id = cl->common.classid;
448 p->un.inner.clprio[prio].ptr = NULL;
451 htb_safe_rb_erase(cl->node + prio,
452 &p->un.inner.clprio[prio].feed);
454 if (!p->un.inner.clprio[prio].feed.rb_node)
458 p->prio_activity &= ~mask;
463 if (cl->cmode == HTB_CAN_SEND && mask)
464 htb_remove_class_from_row(q, cl, mask);
467 static inline s64 htb_lowater(const struct htb_class *cl)
470 return cl->cmode != HTB_CANT_SEND ? -cl->cbuffer : 0;
474 static inline s64 htb_hiwater(const struct htb_class *cl)
477 return cl->cmode == HTB_CAN_SEND ? -cl->buffer : 0;
484 * htb_class_mode - computes and returns current class mode
486 * It computes cl's mode at time cl->t_c+diff and returns it. If mode
487 * is not HTB_CAN_SEND then cl->pq_key is updated to time difference
488 * from now to time when cl will change its state.
489 * Also it is worth to note that class mode doesn't change simply
490 * at cl->{c,}tokens == 0 but there can rather be hysteresis of
491 * 0 .. -cl->{c,}buffer range. It is meant to limit number of
492 * mode transitions per time unit. The speed gain is about 1/6.
494 static inline enum htb_cmode
495 htb_class_mode(struct htb_class *cl, s64 *diff)
499 if ((toks = (cl->ctokens + *diff)) < htb_lowater(cl)) {
501 return HTB_CANT_SEND;
504 if ((toks = (cl->tokens + *diff)) >= htb_hiwater(cl))
508 return HTB_MAY_BORROW;
512 * htb_change_class_mode - changes classe's mode
514 * This should be the only way how to change classe's mode under normal
515 * cirsumstances. Routine will update feed lists linkage, change mode
516 * and add class to the wait event queue if appropriate. New mode should
517 * be different from old one and cl->pq_key has to be valid if changing
518 * to mode other than HTB_CAN_SEND (see htb_add_to_wait_tree).
521 htb_change_class_mode(struct htb_sched *q, struct htb_class *cl, s64 *diff)
523 enum htb_cmode new_mode = htb_class_mode(cl, diff);
525 if (new_mode == cl->cmode)
528 if (cl->prio_activity) { /* not necessary: speed optimization */
529 if (cl->cmode != HTB_CANT_SEND)
530 htb_deactivate_prios(q, cl);
531 cl->cmode = new_mode;
532 if (new_mode != HTB_CANT_SEND)
533 htb_activate_prios(q, cl);
535 cl->cmode = new_mode;
539 * htb_activate - inserts leaf cl into appropriate active feeds
541 * Routine learns (new) priority of leaf and activates feed chain
542 * for the prio. It can be called on already active leaf safely.
543 * It also adds leaf into droplist.
545 static inline void htb_activate(struct htb_sched *q, struct htb_class *cl)
547 WARN_ON(cl->level || !cl->un.leaf.q || !cl->un.leaf.q->q.qlen);
549 if (!cl->prio_activity) {
550 cl->prio_activity = 1 << cl->prio;
551 htb_activate_prios(q, cl);
552 list_add_tail(&cl->un.leaf.drop_list,
553 q->drops + cl->prio);
558 * htb_deactivate - remove leaf cl from active feeds
560 * Make sure that leaf is active. In the other words it can't be called
561 * with non-active leaf. It also removes class from the drop list.
563 static inline void htb_deactivate(struct htb_sched *q, struct htb_class *cl)
565 WARN_ON(!cl->prio_activity);
567 htb_deactivate_prios(q, cl);
568 cl->prio_activity = 0;
569 list_del_init(&cl->un.leaf.drop_list);
572 static int htb_enqueue(struct sk_buff *skb, struct Qdisc *sch)
574 int uninitialized_var(ret);
575 struct htb_sched *q = qdisc_priv(sch);
576 struct htb_class *cl = htb_classify(skb, sch, &ret);
578 if (cl == HTB_DIRECT) {
579 /* enqueue to helper queue */
580 if (q->direct_queue.qlen < q->direct_qlen) {
581 __skb_queue_tail(&q->direct_queue, skb);
584 return qdisc_drop(skb, sch);
586 #ifdef CONFIG_NET_CLS_ACT
588 if (ret & __NET_XMIT_BYPASS)
593 } else if ((ret = qdisc_enqueue(skb, cl->un.leaf.q)) != NET_XMIT_SUCCESS) {
594 if (net_xmit_drop_count(ret)) {
604 return NET_XMIT_SUCCESS;
607 static inline void htb_accnt_tokens(struct htb_class *cl, int bytes, s64 diff)
609 s64 toks = diff + cl->tokens;
611 if (toks > cl->buffer)
613 toks -= (s64) psched_l2t_ns(&cl->rate, bytes);
614 if (toks <= -cl->mbuffer)
615 toks = 1 - cl->mbuffer;
620 static inline void htb_accnt_ctokens(struct htb_class *cl, int bytes, s64 diff)
622 s64 toks = diff + cl->ctokens;
624 if (toks > cl->cbuffer)
626 toks -= (s64) psched_l2t_ns(&cl->ceil, bytes);
627 if (toks <= -cl->mbuffer)
628 toks = 1 - cl->mbuffer;
634 * htb_charge_class - charges amount "bytes" to leaf and ancestors
636 * Routine assumes that packet "bytes" long was dequeued from leaf cl
637 * borrowing from "level". It accounts bytes to ceil leaky bucket for
638 * leaf and all ancestors and to rate bucket for ancestors at levels
639 * "level" and higher. It also handles possible change of mode resulting
640 * from the update. Note that mode can also increase here (MAY_BORROW to
641 * CAN_SEND) because we can use more precise clock that event queue here.
642 * In such case we remove class from event queue first.
644 static void htb_charge_class(struct htb_sched *q, struct htb_class *cl,
645 int level, struct sk_buff *skb)
647 int bytes = qdisc_pkt_len(skb);
648 enum htb_cmode old_mode;
652 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
653 if (cl->level >= level) {
654 if (cl->level == level)
656 htb_accnt_tokens(cl, bytes, diff);
658 cl->xstats.borrows++;
659 cl->tokens += diff; /* we moved t_c; update tokens */
661 htb_accnt_ctokens(cl, bytes, diff);
664 old_mode = cl->cmode;
666 htb_change_class_mode(q, cl, &diff);
667 if (old_mode != cl->cmode) {
668 if (old_mode != HTB_CAN_SEND)
669 htb_safe_rb_erase(&cl->pq_node, &q->hlevel[cl->level].wait_pq);
670 if (cl->cmode != HTB_CAN_SEND)
671 htb_add_to_wait_tree(q, cl, diff);
674 /* update basic stats except for leaves which are already updated */
676 bstats_update(&cl->bstats, skb);
683 * htb_do_events - make mode changes to classes at the level
685 * Scans event queue for pending events and applies them. Returns time of
686 * next pending event (0 for no event in pq, q->now for too many events).
687 * Note: Applied are events whose have cl->pq_key <= q->now.
689 static s64 htb_do_events(struct htb_sched *q, const int level,
692 /* don't run for longer than 2 jiffies; 2 is used instead of
693 * 1 to simplify things when jiffy is going to be incremented
696 unsigned long stop_at = start + 2;
697 struct rb_root *wait_pq = &q->hlevel[level].wait_pq;
699 while (time_before(jiffies, stop_at)) {
700 struct htb_class *cl;
702 struct rb_node *p = rb_first(wait_pq);
707 cl = rb_entry(p, struct htb_class, pq_node);
708 if (cl->pq_key > q->now)
711 htb_safe_rb_erase(p, wait_pq);
712 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
713 htb_change_class_mode(q, cl, &diff);
714 if (cl->cmode != HTB_CAN_SEND)
715 htb_add_to_wait_tree(q, cl, diff);
718 /* too much load - let's continue after a break for scheduling */
719 if (!(q->warned & HTB_WARN_TOOMANYEVENTS)) {
720 pr_warn("htb: too many events!\n");
721 q->warned |= HTB_WARN_TOOMANYEVENTS;
727 /* Returns class->node+prio from id-tree where classe's id is >= id. NULL
728 * is no such one exists.
730 static struct rb_node *htb_id_find_next_upper(int prio, struct rb_node *n,
733 struct rb_node *r = NULL;
735 struct htb_class *cl =
736 rb_entry(n, struct htb_class, node[prio]);
738 if (id > cl->common.classid) {
740 } else if (id < cl->common.classid) {
751 * htb_lookup_leaf - returns next leaf class in DRR order
753 * Find leaf where current feed pointers points to.
755 static struct htb_class *htb_lookup_leaf(struct htb_prio *hprio, const int prio)
759 struct rb_node *root;
760 struct rb_node **pptr;
762 } stk[TC_HTB_MAXDEPTH], *sp = stk;
764 BUG_ON(!hprio->row.rb_node);
765 sp->root = hprio->row.rb_node;
766 sp->pptr = &hprio->ptr;
767 sp->pid = &hprio->last_ptr_id;
769 for (i = 0; i < 65535; i++) {
770 if (!*sp->pptr && *sp->pid) {
771 /* ptr was invalidated but id is valid - try to recover
772 * the original or next ptr
775 htb_id_find_next_upper(prio, sp->root, *sp->pid);
777 *sp->pid = 0; /* ptr is valid now so that remove this hint as it
778 * can become out of date quickly
780 if (!*sp->pptr) { /* we are at right end; rewind & go up */
781 *sp->pptr = sp->root;
782 while ((*sp->pptr)->rb_left)
783 *sp->pptr = (*sp->pptr)->rb_left;
790 htb_next_rb_node(sp->pptr);
793 struct htb_class *cl;
794 struct htb_prio *clp;
796 cl = rb_entry(*sp->pptr, struct htb_class, node[prio]);
799 clp = &cl->un.inner.clprio[prio];
800 (++sp)->root = clp->feed.rb_node;
801 sp->pptr = &clp->ptr;
802 sp->pid = &clp->last_ptr_id;
809 /* dequeues packet at given priority and level; call only if
810 * you are sure that there is active class at prio/level
812 static struct sk_buff *htb_dequeue_tree(struct htb_sched *q, const int prio,
815 struct sk_buff *skb = NULL;
816 struct htb_class *cl, *start;
817 struct htb_level *hlevel = &q->hlevel[level];
818 struct htb_prio *hprio = &hlevel->hprio[prio];
820 /* look initial class up in the row */
821 start = cl = htb_lookup_leaf(hprio, prio);
828 /* class can be empty - it is unlikely but can be true if leaf
829 * qdisc drops packets in enqueue routine or if someone used
830 * graft operation on the leaf since last dequeue;
831 * simply deactivate and skip such class
833 if (unlikely(cl->un.leaf.q->q.qlen == 0)) {
834 struct htb_class *next;
835 htb_deactivate(q, cl);
837 /* row/level might become empty */
838 if ((q->row_mask[level] & (1 << prio)) == 0)
841 next = htb_lookup_leaf(hprio, prio);
843 if (cl == start) /* fix start if we just deleted it */
849 skb = cl->un.leaf.q->dequeue(cl->un.leaf.q);
850 if (likely(skb != NULL))
853 qdisc_warn_nonwc("htb", cl->un.leaf.q);
854 htb_next_rb_node(level ? &cl->parent->un.inner.clprio[prio].ptr:
855 &q->hlevel[0].hprio[prio].ptr);
856 cl = htb_lookup_leaf(hprio, prio);
858 } while (cl != start);
860 if (likely(skb != NULL)) {
861 bstats_update(&cl->bstats, skb);
862 cl->un.leaf.deficit[level] -= qdisc_pkt_len(skb);
863 if (cl->un.leaf.deficit[level] < 0) {
864 cl->un.leaf.deficit[level] += cl->quantum;
865 htb_next_rb_node(level ? &cl->parent->un.inner.clprio[prio].ptr :
866 &q->hlevel[0].hprio[prio].ptr);
868 /* this used to be after charge_class but this constelation
869 * gives us slightly better performance
871 if (!cl->un.leaf.q->q.qlen)
872 htb_deactivate(q, cl);
873 htb_charge_class(q, cl, level, skb);
878 static struct sk_buff *htb_dequeue(struct Qdisc *sch)
881 struct htb_sched *q = qdisc_priv(sch);
884 unsigned long start_at;
886 /* try to dequeue direct packets as high prio (!) to minimize cpu work */
887 skb = __skb_dequeue(&q->direct_queue);
890 qdisc_bstats_update(sch, skb);
891 qdisc_unthrottled(sch);
898 q->now = ktime_to_ns(ktime_get());
901 next_event = q->now + 5LLU * NSEC_PER_SEC;
903 for (level = 0; level < TC_HTB_MAXDEPTH; level++) {
904 /* common case optimization - skip event handler quickly */
906 s64 event = q->near_ev_cache[level];
908 if (q->now >= event) {
909 event = htb_do_events(q, level, start_at);
911 event = q->now + NSEC_PER_SEC;
912 q->near_ev_cache[level] = event;
915 if (next_event > event)
918 m = ~q->row_mask[level];
919 while (m != (int)(-1)) {
923 skb = htb_dequeue_tree(q, prio, level);
924 if (likely(skb != NULL))
928 sch->qstats.overlimits++;
929 if (likely(next_event > q->now)) {
930 if (!test_bit(__QDISC_STATE_DEACTIVATED,
931 &qdisc_root_sleeping(q->watchdog.qdisc)->state)) {
932 ktime_t time = ns_to_ktime(next_event);
933 qdisc_throttled(q->watchdog.qdisc);
934 hrtimer_start(&q->watchdog.timer, time,
938 schedule_work(&q->work);
944 /* try to drop from each class (by prio) until one succeed */
945 static unsigned int htb_drop(struct Qdisc *sch)
947 struct htb_sched *q = qdisc_priv(sch);
950 for (prio = TC_HTB_NUMPRIO - 1; prio >= 0; prio--) {
952 list_for_each(p, q->drops + prio) {
953 struct htb_class *cl = list_entry(p, struct htb_class,
956 if (cl->un.leaf.q->ops->drop &&
957 (len = cl->un.leaf.q->ops->drop(cl->un.leaf.q))) {
959 if (!cl->un.leaf.q->q.qlen)
960 htb_deactivate(q, cl);
968 /* reset all classes */
969 /* always caled under BH & queue lock */
970 static void htb_reset(struct Qdisc *sch)
972 struct htb_sched *q = qdisc_priv(sch);
973 struct htb_class *cl;
976 for (i = 0; i < q->clhash.hashsize; i++) {
977 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
979 memset(&cl->un.inner, 0, sizeof(cl->un.inner));
982 qdisc_reset(cl->un.leaf.q);
983 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
985 cl->prio_activity = 0;
986 cl->cmode = HTB_CAN_SEND;
990 qdisc_watchdog_cancel(&q->watchdog);
991 __skb_queue_purge(&q->direct_queue);
993 memset(q->hlevel, 0, sizeof(q->hlevel));
994 memset(q->row_mask, 0, sizeof(q->row_mask));
995 for (i = 0; i < TC_HTB_NUMPRIO; i++)
996 INIT_LIST_HEAD(q->drops + i);
999 static const struct nla_policy htb_policy[TCA_HTB_MAX + 1] = {
1000 [TCA_HTB_PARMS] = { .len = sizeof(struct tc_htb_opt) },
1001 [TCA_HTB_INIT] = { .len = sizeof(struct tc_htb_glob) },
1002 [TCA_HTB_CTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
1003 [TCA_HTB_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
1004 [TCA_HTB_DIRECT_QLEN] = { .type = NLA_U32 },
1005 [TCA_HTB_RATE64] = { .type = NLA_U64 },
1006 [TCA_HTB_CEIL64] = { .type = NLA_U64 },
1009 static void htb_work_func(struct work_struct *work)
1011 struct htb_sched *q = container_of(work, struct htb_sched, work);
1012 struct Qdisc *sch = q->watchdog.qdisc;
1014 __netif_schedule(qdisc_root(sch));
1017 static int htb_init(struct Qdisc *sch, struct nlattr *opt)
1019 struct htb_sched *q = qdisc_priv(sch);
1020 struct nlattr *tb[TCA_HTB_MAX + 1];
1021 struct tc_htb_glob *gopt;
1028 err = nla_parse_nested(tb, TCA_HTB_MAX, opt, htb_policy);
1032 if (!tb[TCA_HTB_INIT])
1035 gopt = nla_data(tb[TCA_HTB_INIT]);
1036 if (gopt->version != HTB_VER >> 16)
1039 err = qdisc_class_hash_init(&q->clhash);
1042 for (i = 0; i < TC_HTB_NUMPRIO; i++)
1043 INIT_LIST_HEAD(q->drops + i);
1045 qdisc_watchdog_init(&q->watchdog, sch);
1046 INIT_WORK(&q->work, htb_work_func);
1047 skb_queue_head_init(&q->direct_queue);
1049 if (tb[TCA_HTB_DIRECT_QLEN])
1050 q->direct_qlen = nla_get_u32(tb[TCA_HTB_DIRECT_QLEN]);
1052 q->direct_qlen = qdisc_dev(sch)->tx_queue_len;
1053 if (q->direct_qlen < 2) /* some devices have zero tx_queue_len */
1056 if ((q->rate2quantum = gopt->rate2quantum) < 1)
1057 q->rate2quantum = 1;
1058 q->defcls = gopt->defcls;
1063 static int htb_dump(struct Qdisc *sch, struct sk_buff *skb)
1065 spinlock_t *root_lock = qdisc_root_sleeping_lock(sch);
1066 struct htb_sched *q = qdisc_priv(sch);
1067 struct nlattr *nest;
1068 struct tc_htb_glob gopt;
1070 spin_lock_bh(root_lock);
1072 gopt.direct_pkts = q->direct_pkts;
1073 gopt.version = HTB_VER;
1074 gopt.rate2quantum = q->rate2quantum;
1075 gopt.defcls = q->defcls;
1078 nest = nla_nest_start(skb, TCA_OPTIONS);
1080 goto nla_put_failure;
1081 if (nla_put(skb, TCA_HTB_INIT, sizeof(gopt), &gopt) ||
1082 nla_put_u32(skb, TCA_HTB_DIRECT_QLEN, q->direct_qlen))
1083 goto nla_put_failure;
1084 nla_nest_end(skb, nest);
1086 spin_unlock_bh(root_lock);
1090 spin_unlock_bh(root_lock);
1091 nla_nest_cancel(skb, nest);
1095 static int htb_dump_class(struct Qdisc *sch, unsigned long arg,
1096 struct sk_buff *skb, struct tcmsg *tcm)
1098 struct htb_class *cl = (struct htb_class *)arg;
1099 spinlock_t *root_lock = qdisc_root_sleeping_lock(sch);
1100 struct nlattr *nest;
1101 struct tc_htb_opt opt;
1103 spin_lock_bh(root_lock);
1104 tcm->tcm_parent = cl->parent ? cl->parent->common.classid : TC_H_ROOT;
1105 tcm->tcm_handle = cl->common.classid;
1106 if (!cl->level && cl->un.leaf.q)
1107 tcm->tcm_info = cl->un.leaf.q->handle;
1109 nest = nla_nest_start(skb, TCA_OPTIONS);
1111 goto nla_put_failure;
1113 memset(&opt, 0, sizeof(opt));
1115 psched_ratecfg_getrate(&opt.rate, &cl->rate);
1116 opt.buffer = PSCHED_NS2TICKS(cl->buffer);
1117 psched_ratecfg_getrate(&opt.ceil, &cl->ceil);
1118 opt.cbuffer = PSCHED_NS2TICKS(cl->cbuffer);
1119 opt.quantum = cl->quantum;
1120 opt.prio = cl->prio;
1121 opt.level = cl->level;
1122 if (nla_put(skb, TCA_HTB_PARMS, sizeof(opt), &opt))
1123 goto nla_put_failure;
1124 if ((cl->rate.rate_bytes_ps >= (1ULL << 32)) &&
1125 nla_put_u64(skb, TCA_HTB_RATE64, cl->rate.rate_bytes_ps))
1126 goto nla_put_failure;
1127 if ((cl->ceil.rate_bytes_ps >= (1ULL << 32)) &&
1128 nla_put_u64(skb, TCA_HTB_CEIL64, cl->ceil.rate_bytes_ps))
1129 goto nla_put_failure;
1131 nla_nest_end(skb, nest);
1132 spin_unlock_bh(root_lock);
1136 spin_unlock_bh(root_lock);
1137 nla_nest_cancel(skb, nest);
1142 htb_dump_class_stats(struct Qdisc *sch, unsigned long arg, struct gnet_dump *d)
1144 struct htb_class *cl = (struct htb_class *)arg;
1146 if (!cl->level && cl->un.leaf.q)
1147 cl->qstats.qlen = cl->un.leaf.q->q.qlen;
1148 cl->xstats.tokens = PSCHED_NS2TICKS(cl->tokens);
1149 cl->xstats.ctokens = PSCHED_NS2TICKS(cl->ctokens);
1151 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
1152 gnet_stats_copy_rate_est(d, NULL, &cl->rate_est) < 0 ||
1153 gnet_stats_copy_queue(d, &cl->qstats) < 0)
1156 return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1159 static int htb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1162 struct htb_class *cl = (struct htb_class *)arg;
1167 (new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1168 cl->common.classid)) == NULL)
1172 *old = cl->un.leaf.q;
1173 cl->un.leaf.q = new;
1175 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
1178 sch_tree_unlock(sch);
1182 static struct Qdisc *htb_leaf(struct Qdisc *sch, unsigned long arg)
1184 struct htb_class *cl = (struct htb_class *)arg;
1185 return !cl->level ? cl->un.leaf.q : NULL;
1188 static void htb_qlen_notify(struct Qdisc *sch, unsigned long arg)
1190 struct htb_class *cl = (struct htb_class *)arg;
1192 if (cl->un.leaf.q->q.qlen == 0)
1193 htb_deactivate(qdisc_priv(sch), cl);
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 inline int htb_parent_last_child(struct htb_class *cl)
1207 /* the root class */
1209 if (cl->parent->children > 1)
1210 /* not the last child */
1215 static void htb_parent_to_leaf(struct htb_sched *q, struct htb_class *cl,
1216 struct Qdisc *new_q)
1218 struct htb_class *parent = cl->parent;
1220 WARN_ON(cl->level || !cl->un.leaf.q || cl->prio_activity);
1222 if (parent->cmode != HTB_CAN_SEND)
1223 htb_safe_rb_erase(&parent->pq_node,
1224 &q->hlevel[parent->level].wait_pq);
1227 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1228 INIT_LIST_HEAD(&parent->un.leaf.drop_list);
1229 parent->un.leaf.q = new_q ? new_q : &noop_qdisc;
1230 parent->tokens = parent->buffer;
1231 parent->ctokens = parent->cbuffer;
1232 parent->t_c = ktime_to_ns(ktime_get());
1233 parent->cmode = HTB_CAN_SEND;
1236 static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
1239 WARN_ON(!cl->un.leaf.q);
1240 qdisc_destroy(cl->un.leaf.q);
1242 gen_kill_estimator(&cl->bstats, &cl->rate_est);
1243 tcf_destroy_chain(&cl->filter_list);
1247 static void htb_destroy(struct Qdisc *sch)
1249 struct htb_sched *q = qdisc_priv(sch);
1250 struct hlist_node *next;
1251 struct htb_class *cl;
1254 cancel_work_sync(&q->work);
1255 qdisc_watchdog_cancel(&q->watchdog);
1256 /* This line used to be after htb_destroy_class call below
1257 * and surprisingly it worked in 2.4. But it must precede it
1258 * because filter need its target class alive to be able to call
1259 * unbind_filter on it (without Oops).
1261 tcf_destroy_chain(&q->filter_list);
1263 for (i = 0; i < q->clhash.hashsize; i++) {
1264 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode)
1265 tcf_destroy_chain(&cl->filter_list);
1267 for (i = 0; i < q->clhash.hashsize; i++) {
1268 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1270 htb_destroy_class(sch, cl);
1272 qdisc_class_hash_destroy(&q->clhash);
1273 __skb_queue_purge(&q->direct_queue);
1276 static int htb_delete(struct Qdisc *sch, unsigned long arg)
1278 struct htb_sched *q = qdisc_priv(sch);
1279 struct htb_class *cl = (struct htb_class *)arg;
1281 struct Qdisc *new_q = NULL;
1284 /* TODO: why don't allow to delete subtree ? references ? does
1285 * tc subsys guarantee us that in htb_destroy it holds no class
1286 * refs so that we can remove children safely there ?
1288 if (cl->children || cl->filter_cnt)
1291 if (!cl->level && htb_parent_last_child(cl)) {
1292 new_q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1293 cl->parent->common.classid);
1300 qlen = cl->un.leaf.q->q.qlen;
1301 qdisc_reset(cl->un.leaf.q);
1302 qdisc_tree_decrease_qlen(cl->un.leaf.q, qlen);
1305 /* delete from hash and active; remainder in destroy_class */
1306 qdisc_class_hash_remove(&q->clhash, &cl->common);
1308 cl->parent->children--;
1310 if (cl->prio_activity)
1311 htb_deactivate(q, cl);
1313 if (cl->cmode != HTB_CAN_SEND)
1314 htb_safe_rb_erase(&cl->pq_node,
1315 &q->hlevel[cl->level].wait_pq);
1318 htb_parent_to_leaf(q, cl, new_q);
1320 BUG_ON(--cl->refcnt == 0);
1322 * This shouldn't happen: we "hold" one cops->get() when called
1323 * from tc_ctl_tclass; the destroy method is done from cops->put().
1326 sch_tree_unlock(sch);
1330 static void htb_put(struct Qdisc *sch, unsigned long arg)
1332 struct htb_class *cl = (struct htb_class *)arg;
1334 if (--cl->refcnt == 0)
1335 htb_destroy_class(sch, cl);
1338 static int htb_change_class(struct Qdisc *sch, u32 classid,
1339 u32 parentid, struct nlattr **tca,
1343 struct htb_sched *q = qdisc_priv(sch);
1344 struct htb_class *cl = (struct htb_class *)*arg, *parent;
1345 struct nlattr *opt = tca[TCA_OPTIONS];
1346 struct nlattr *tb[TCA_HTB_MAX + 1];
1347 struct tc_htb_opt *hopt;
1350 /* extract all subattrs from opt attr */
1354 err = nla_parse_nested(tb, TCA_HTB_MAX, opt, htb_policy);
1359 if (tb[TCA_HTB_PARMS] == NULL)
1362 parent = parentid == TC_H_ROOT ? NULL : htb_find(parentid, sch);
1364 hopt = nla_data(tb[TCA_HTB_PARMS]);
1365 if (!hopt->rate.rate || !hopt->ceil.rate)
1368 /* Keeping backward compatible with rate_table based iproute2 tc */
1369 if (hopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
1370 qdisc_put_rtab(qdisc_get_rtab(&hopt->rate, tb[TCA_HTB_RTAB]));
1372 if (hopt->ceil.linklayer == TC_LINKLAYER_UNAWARE)
1373 qdisc_put_rtab(qdisc_get_rtab(&hopt->ceil, tb[TCA_HTB_CTAB]));
1375 if (!cl) { /* new class */
1376 struct Qdisc *new_q;
1380 struct gnet_estimator opt;
1383 .nla_len = nla_attr_size(sizeof(est.opt)),
1384 .nla_type = TCA_RATE,
1387 /* 4s interval, 16s averaging constant */
1393 /* check for valid classid */
1394 if (!classid || TC_H_MAJ(classid ^ sch->handle) ||
1395 htb_find(classid, sch))
1398 /* check maximal depth */
1399 if (parent && parent->parent && parent->parent->level < 2) {
1400 pr_err("htb: tree is too deep\n");
1404 cl = kzalloc(sizeof(*cl), GFP_KERNEL);
1408 if (htb_rate_est || tca[TCA_RATE]) {
1409 err = gen_new_estimator(&cl->bstats, &cl->rate_est,
1410 qdisc_root_sleeping_lock(sch),
1411 tca[TCA_RATE] ? : &est.nla);
1420 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
1421 RB_CLEAR_NODE(&cl->pq_node);
1423 for (prio = 0; prio < TC_HTB_NUMPRIO; prio++)
1424 RB_CLEAR_NODE(&cl->node[prio]);
1426 /* create leaf qdisc early because it uses kmalloc(GFP_KERNEL)
1427 * so that can't be used inside of sch_tree_lock
1428 * -- thanks to Karlis Peisenieks
1430 new_q = qdisc_create_dflt(sch->dev_queue,
1431 &pfifo_qdisc_ops, classid);
1433 if (parent && !parent->level) {
1434 unsigned int qlen = parent->un.leaf.q->q.qlen;
1436 /* turn parent into inner node */
1437 qdisc_reset(parent->un.leaf.q);
1438 qdisc_tree_decrease_qlen(parent->un.leaf.q, qlen);
1439 qdisc_destroy(parent->un.leaf.q);
1440 if (parent->prio_activity)
1441 htb_deactivate(q, parent);
1443 /* remove from evt list because of level change */
1444 if (parent->cmode != HTB_CAN_SEND) {
1445 htb_safe_rb_erase(&parent->pq_node, &q->hlevel[0].wait_pq);
1446 parent->cmode = HTB_CAN_SEND;
1448 parent->level = (parent->parent ? parent->parent->level
1449 : TC_HTB_MAXDEPTH) - 1;
1450 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1452 /* leaf (we) needs elementary qdisc */
1453 cl->un.leaf.q = new_q ? new_q : &noop_qdisc;
1455 cl->common.classid = classid;
1456 cl->parent = parent;
1458 /* set class to be in HTB_CAN_SEND state */
1459 cl->tokens = PSCHED_TICKS2NS(hopt->buffer);
1460 cl->ctokens = PSCHED_TICKS2NS(hopt->cbuffer);
1461 cl->mbuffer = 60ULL * NSEC_PER_SEC; /* 1min */
1462 cl->t_c = ktime_to_ns(ktime_get());
1463 cl->cmode = HTB_CAN_SEND;
1465 /* attach to the hash list and parent's family */
1466 qdisc_class_hash_insert(&q->clhash, &cl->common);
1470 if (tca[TCA_RATE]) {
1471 err = gen_replace_estimator(&cl->bstats, &cl->rate_est,
1472 qdisc_root_sleeping_lock(sch),
1480 rate64 = tb[TCA_HTB_RATE64] ? nla_get_u64(tb[TCA_HTB_RATE64]) : 0;
1482 ceil64 = tb[TCA_HTB_CEIL64] ? nla_get_u64(tb[TCA_HTB_CEIL64]) : 0;
1484 psched_ratecfg_precompute(&cl->rate, &hopt->rate, rate64);
1485 psched_ratecfg_precompute(&cl->ceil, &hopt->ceil, ceil64);
1487 /* it used to be a nasty bug here, we have to check that node
1488 * is really leaf before changing cl->un.leaf !
1491 u64 quantum = cl->rate.rate_bytes_ps;
1493 do_div(quantum, q->rate2quantum);
1494 cl->quantum = min_t(u64, quantum, INT_MAX);
1496 if (!hopt->quantum && cl->quantum < 1000) {
1497 pr_warn("HTB: quantum of class %X is small. Consider r2q change.\n",
1498 cl->common.classid);
1501 if (!hopt->quantum && cl->quantum > 200000) {
1502 pr_warn("HTB: quantum of class %X is big. Consider r2q change.\n",
1503 cl->common.classid);
1504 cl->quantum = 200000;
1507 cl->quantum = hopt->quantum;
1508 if ((cl->prio = hopt->prio) >= TC_HTB_NUMPRIO)
1509 cl->prio = TC_HTB_NUMPRIO - 1;
1512 cl->buffer = PSCHED_TICKS2NS(hopt->buffer);
1513 cl->cbuffer = PSCHED_TICKS2NS(hopt->cbuffer);
1515 sch_tree_unlock(sch);
1517 qdisc_class_hash_grow(sch, &q->clhash);
1519 *arg = (unsigned long)cl;
1526 static struct tcf_proto **htb_find_tcf(struct Qdisc *sch, unsigned long arg)
1528 struct htb_sched *q = qdisc_priv(sch);
1529 struct htb_class *cl = (struct htb_class *)arg;
1530 struct tcf_proto **fl = cl ? &cl->filter_list : &q->filter_list;
1535 static unsigned long htb_bind_filter(struct Qdisc *sch, unsigned long parent,
1538 struct htb_class *cl = htb_find(classid, sch);
1540 /*if (cl && !cl->level) return 0;
1541 * The line above used to be there to prevent attaching filters to
1542 * leaves. But at least tc_index filter uses this just to get class
1543 * for other reasons so that we have to allow for it.
1545 * 19.6.2002 As Werner explained it is ok - bind filter is just
1546 * another way to "lock" the class - unlike "get" this lock can
1547 * be broken by class during destroy IIUC.
1551 return (unsigned long)cl;
1554 static void htb_unbind_filter(struct Qdisc *sch, unsigned long arg)
1556 struct htb_class *cl = (struct htb_class *)arg;
1562 static void htb_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1564 struct htb_sched *q = qdisc_priv(sch);
1565 struct htb_class *cl;
1571 for (i = 0; i < q->clhash.hashsize; i++) {
1572 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1573 if (arg->count < arg->skip) {
1577 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1586 static const struct Qdisc_class_ops htb_class_ops = {
1589 .qlen_notify = htb_qlen_notify,
1592 .change = htb_change_class,
1593 .delete = htb_delete,
1595 .tcf_chain = htb_find_tcf,
1596 .bind_tcf = htb_bind_filter,
1597 .unbind_tcf = htb_unbind_filter,
1598 .dump = htb_dump_class,
1599 .dump_stats = htb_dump_class_stats,
1602 static struct Qdisc_ops htb_qdisc_ops __read_mostly = {
1603 .cl_ops = &htb_class_ops,
1605 .priv_size = sizeof(struct htb_sched),
1606 .enqueue = htb_enqueue,
1607 .dequeue = htb_dequeue,
1608 .peek = qdisc_peek_dequeued,
1612 .destroy = htb_destroy,
1614 .owner = THIS_MODULE,
1617 static int __init htb_module_init(void)
1619 return register_qdisc(&htb_qdisc_ops);
1621 static void __exit htb_module_exit(void)
1623 unregister_qdisc(&htb_qdisc_ops);
1626 module_init(htb_module_init)
1627 module_exit(htb_module_exit)
1628 MODULE_LICENSE("GPL");