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 __rcu *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 __rcu *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 = rcu_dereference_bh(cl->filter_list);
228 tcf = rcu_dereference_bh(q->filter_list);
231 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
232 while (tcf && (result = tc_classify(skb, tcf, &res, false)) >= 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 = rcu_dereference_bh(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)
589 qdisc_qstats_drop(sch);
593 } else if ((ret = qdisc_enqueue(skb, cl->un.leaf.q)) != NET_XMIT_SUCCESS) {
594 if (net_xmit_drop_count(ret)) {
595 qdisc_qstats_drop(sch);
603 qdisc_qstats_backlog_inc(sch, skb);
605 return NET_XMIT_SUCCESS;
608 static inline void htb_accnt_tokens(struct htb_class *cl, int bytes, s64 diff)
610 s64 toks = diff + cl->tokens;
612 if (toks > cl->buffer)
614 toks -= (s64) psched_l2t_ns(&cl->rate, bytes);
615 if (toks <= -cl->mbuffer)
616 toks = 1 - cl->mbuffer;
621 static inline void htb_accnt_ctokens(struct htb_class *cl, int bytes, s64 diff)
623 s64 toks = diff + cl->ctokens;
625 if (toks > cl->cbuffer)
627 toks -= (s64) psched_l2t_ns(&cl->ceil, bytes);
628 if (toks <= -cl->mbuffer)
629 toks = 1 - cl->mbuffer;
635 * htb_charge_class - charges amount "bytes" to leaf and ancestors
637 * Routine assumes that packet "bytes" long was dequeued from leaf cl
638 * borrowing from "level". It accounts bytes to ceil leaky bucket for
639 * leaf and all ancestors and to rate bucket for ancestors at levels
640 * "level" and higher. It also handles possible change of mode resulting
641 * from the update. Note that mode can also increase here (MAY_BORROW to
642 * CAN_SEND) because we can use more precise clock that event queue here.
643 * In such case we remove class from event queue first.
645 static void htb_charge_class(struct htb_sched *q, struct htb_class *cl,
646 int level, struct sk_buff *skb)
648 int bytes = qdisc_pkt_len(skb);
649 enum htb_cmode old_mode;
653 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
654 if (cl->level >= level) {
655 if (cl->level == level)
657 htb_accnt_tokens(cl, bytes, diff);
659 cl->xstats.borrows++;
660 cl->tokens += diff; /* we moved t_c; update tokens */
662 htb_accnt_ctokens(cl, bytes, diff);
665 old_mode = cl->cmode;
667 htb_change_class_mode(q, cl, &diff);
668 if (old_mode != cl->cmode) {
669 if (old_mode != HTB_CAN_SEND)
670 htb_safe_rb_erase(&cl->pq_node, &q->hlevel[cl->level].wait_pq);
671 if (cl->cmode != HTB_CAN_SEND)
672 htb_add_to_wait_tree(q, cl, diff);
675 /* update basic stats except for leaves which are already updated */
677 bstats_update(&cl->bstats, skb);
684 * htb_do_events - make mode changes to classes at the level
686 * Scans event queue for pending events and applies them. Returns time of
687 * next pending event (0 for no event in pq, q->now for too many events).
688 * Note: Applied are events whose have cl->pq_key <= q->now.
690 static s64 htb_do_events(struct htb_sched *q, const int level,
693 /* don't run for longer than 2 jiffies; 2 is used instead of
694 * 1 to simplify things when jiffy is going to be incremented
697 unsigned long stop_at = start + 2;
698 struct rb_root *wait_pq = &q->hlevel[level].wait_pq;
700 while (time_before(jiffies, stop_at)) {
701 struct htb_class *cl;
703 struct rb_node *p = rb_first(wait_pq);
708 cl = rb_entry(p, struct htb_class, pq_node);
709 if (cl->pq_key > q->now)
712 htb_safe_rb_erase(p, wait_pq);
713 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
714 htb_change_class_mode(q, cl, &diff);
715 if (cl->cmode != HTB_CAN_SEND)
716 htb_add_to_wait_tree(q, cl, diff);
719 /* too much load - let's continue after a break for scheduling */
720 if (!(q->warned & HTB_WARN_TOOMANYEVENTS)) {
721 pr_warn("htb: too many events!\n");
722 q->warned |= HTB_WARN_TOOMANYEVENTS;
728 /* Returns class->node+prio from id-tree where classe's id is >= id. NULL
729 * is no such one exists.
731 static struct rb_node *htb_id_find_next_upper(int prio, struct rb_node *n,
734 struct rb_node *r = NULL;
736 struct htb_class *cl =
737 rb_entry(n, struct htb_class, node[prio]);
739 if (id > cl->common.classid) {
741 } else if (id < cl->common.classid) {
752 * htb_lookup_leaf - returns next leaf class in DRR order
754 * Find leaf where current feed pointers points to.
756 static struct htb_class *htb_lookup_leaf(struct htb_prio *hprio, const int prio)
760 struct rb_node *root;
761 struct rb_node **pptr;
763 } stk[TC_HTB_MAXDEPTH], *sp = stk;
765 BUG_ON(!hprio->row.rb_node);
766 sp->root = hprio->row.rb_node;
767 sp->pptr = &hprio->ptr;
768 sp->pid = &hprio->last_ptr_id;
770 for (i = 0; i < 65535; i++) {
771 if (!*sp->pptr && *sp->pid) {
772 /* ptr was invalidated but id is valid - try to recover
773 * the original or next ptr
776 htb_id_find_next_upper(prio, sp->root, *sp->pid);
778 *sp->pid = 0; /* ptr is valid now so that remove this hint as it
779 * can become out of date quickly
781 if (!*sp->pptr) { /* we are at right end; rewind & go up */
782 *sp->pptr = sp->root;
783 while ((*sp->pptr)->rb_left)
784 *sp->pptr = (*sp->pptr)->rb_left;
791 htb_next_rb_node(sp->pptr);
794 struct htb_class *cl;
795 struct htb_prio *clp;
797 cl = rb_entry(*sp->pptr, struct htb_class, node[prio]);
800 clp = &cl->un.inner.clprio[prio];
801 (++sp)->root = clp->feed.rb_node;
802 sp->pptr = &clp->ptr;
803 sp->pid = &clp->last_ptr_id;
810 /* dequeues packet at given priority and level; call only if
811 * you are sure that there is active class at prio/level
813 static struct sk_buff *htb_dequeue_tree(struct htb_sched *q, const int prio,
816 struct sk_buff *skb = NULL;
817 struct htb_class *cl, *start;
818 struct htb_level *hlevel = &q->hlevel[level];
819 struct htb_prio *hprio = &hlevel->hprio[prio];
821 /* look initial class up in the row */
822 start = cl = htb_lookup_leaf(hprio, prio);
829 /* class can be empty - it is unlikely but can be true if leaf
830 * qdisc drops packets in enqueue routine or if someone used
831 * graft operation on the leaf since last dequeue;
832 * simply deactivate and skip such class
834 if (unlikely(cl->un.leaf.q->q.qlen == 0)) {
835 struct htb_class *next;
836 htb_deactivate(q, cl);
838 /* row/level might become empty */
839 if ((q->row_mask[level] & (1 << prio)) == 0)
842 next = htb_lookup_leaf(hprio, prio);
844 if (cl == start) /* fix start if we just deleted it */
850 skb = cl->un.leaf.q->dequeue(cl->un.leaf.q);
851 if (likely(skb != NULL))
854 qdisc_warn_nonwc("htb", cl->un.leaf.q);
855 htb_next_rb_node(level ? &cl->parent->un.inner.clprio[prio].ptr:
856 &q->hlevel[0].hprio[prio].ptr);
857 cl = htb_lookup_leaf(hprio, prio);
859 } while (cl != start);
861 if (likely(skb != NULL)) {
862 bstats_update(&cl->bstats, skb);
863 cl->un.leaf.deficit[level] -= qdisc_pkt_len(skb);
864 if (cl->un.leaf.deficit[level] < 0) {
865 cl->un.leaf.deficit[level] += cl->quantum;
866 htb_next_rb_node(level ? &cl->parent->un.inner.clprio[prio].ptr :
867 &q->hlevel[0].hprio[prio].ptr);
869 /* this used to be after charge_class but this constelation
870 * gives us slightly better performance
872 if (!cl->un.leaf.q->q.qlen)
873 htb_deactivate(q, cl);
874 htb_charge_class(q, cl, level, skb);
879 static struct sk_buff *htb_dequeue(struct Qdisc *sch)
882 struct htb_sched *q = qdisc_priv(sch);
885 unsigned long start_at;
887 /* try to dequeue direct packets as high prio (!) to minimize cpu work */
888 skb = __skb_dequeue(&q->direct_queue);
891 qdisc_bstats_update(sch, skb);
892 qdisc_unthrottled(sch);
893 qdisc_qstats_backlog_dec(sch, skb);
900 q->now = ktime_get_ns();
903 next_event = q->now + 5LLU * NSEC_PER_SEC;
905 for (level = 0; level < TC_HTB_MAXDEPTH; level++) {
906 /* common case optimization - skip event handler quickly */
908 s64 event = q->near_ev_cache[level];
910 if (q->now >= event) {
911 event = htb_do_events(q, level, start_at);
913 event = q->now + NSEC_PER_SEC;
914 q->near_ev_cache[level] = event;
917 if (next_event > event)
920 m = ~q->row_mask[level];
921 while (m != (int)(-1)) {
925 skb = htb_dequeue_tree(q, prio, level);
926 if (likely(skb != NULL))
930 qdisc_qstats_overlimit(sch);
931 if (likely(next_event > q->now))
932 qdisc_watchdog_schedule_ns(&q->watchdog, next_event, true);
934 schedule_work(&q->work);
939 /* try to drop from each class (by prio) until one succeed */
940 static unsigned int htb_drop(struct Qdisc *sch)
942 struct htb_sched *q = qdisc_priv(sch);
945 for (prio = TC_HTB_NUMPRIO - 1; prio >= 0; prio--) {
947 list_for_each(p, q->drops + prio) {
948 struct htb_class *cl = list_entry(p, struct htb_class,
951 if (cl->un.leaf.q->ops->drop &&
952 (len = cl->un.leaf.q->ops->drop(cl->un.leaf.q))) {
953 sch->qstats.backlog -= len;
955 if (!cl->un.leaf.q->q.qlen)
956 htb_deactivate(q, cl);
964 /* reset all classes */
965 /* always caled under BH & queue lock */
966 static void htb_reset(struct Qdisc *sch)
968 struct htb_sched *q = qdisc_priv(sch);
969 struct htb_class *cl;
972 for (i = 0; i < q->clhash.hashsize; i++) {
973 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
975 memset(&cl->un.inner, 0, sizeof(cl->un.inner));
978 qdisc_reset(cl->un.leaf.q);
979 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
981 cl->prio_activity = 0;
982 cl->cmode = HTB_CAN_SEND;
985 qdisc_watchdog_cancel(&q->watchdog);
986 __skb_queue_purge(&q->direct_queue);
988 sch->qstats.backlog = 0;
989 memset(q->hlevel, 0, sizeof(q->hlevel));
990 memset(q->row_mask, 0, sizeof(q->row_mask));
991 for (i = 0; i < TC_HTB_NUMPRIO; i++)
992 INIT_LIST_HEAD(q->drops + i);
995 static const struct nla_policy htb_policy[TCA_HTB_MAX + 1] = {
996 [TCA_HTB_PARMS] = { .len = sizeof(struct tc_htb_opt) },
997 [TCA_HTB_INIT] = { .len = sizeof(struct tc_htb_glob) },
998 [TCA_HTB_CTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
999 [TCA_HTB_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
1000 [TCA_HTB_DIRECT_QLEN] = { .type = NLA_U32 },
1001 [TCA_HTB_RATE64] = { .type = NLA_U64 },
1002 [TCA_HTB_CEIL64] = { .type = NLA_U64 },
1005 static void htb_work_func(struct work_struct *work)
1007 struct htb_sched *q = container_of(work, struct htb_sched, work);
1008 struct Qdisc *sch = q->watchdog.qdisc;
1011 __netif_schedule(qdisc_root(sch));
1015 static int htb_init(struct Qdisc *sch, struct nlattr *opt)
1017 struct htb_sched *q = qdisc_priv(sch);
1018 struct nlattr *tb[TCA_HTB_MAX + 1];
1019 struct tc_htb_glob *gopt;
1026 err = nla_parse_nested(tb, TCA_HTB_MAX, opt, htb_policy);
1030 if (!tb[TCA_HTB_INIT])
1033 gopt = nla_data(tb[TCA_HTB_INIT]);
1034 if (gopt->version != HTB_VER >> 16)
1037 err = qdisc_class_hash_init(&q->clhash);
1040 for (i = 0; i < TC_HTB_NUMPRIO; i++)
1041 INIT_LIST_HEAD(q->drops + i);
1043 qdisc_watchdog_init(&q->watchdog, sch);
1044 INIT_WORK(&q->work, htb_work_func);
1045 __skb_queue_head_init(&q->direct_queue);
1047 if (tb[TCA_HTB_DIRECT_QLEN])
1048 q->direct_qlen = nla_get_u32(tb[TCA_HTB_DIRECT_QLEN]);
1050 q->direct_qlen = qdisc_dev(sch)->tx_queue_len;
1052 if ((q->rate2quantum = gopt->rate2quantum) < 1)
1053 q->rate2quantum = 1;
1054 q->defcls = gopt->defcls;
1059 static int htb_dump(struct Qdisc *sch, struct sk_buff *skb)
1061 struct htb_sched *q = qdisc_priv(sch);
1062 struct nlattr *nest;
1063 struct tc_htb_glob gopt;
1065 /* Its safe to not acquire qdisc lock. As we hold RTNL,
1066 * no change can happen on the qdisc parameters.
1069 gopt.direct_pkts = q->direct_pkts;
1070 gopt.version = HTB_VER;
1071 gopt.rate2quantum = q->rate2quantum;
1072 gopt.defcls = q->defcls;
1075 nest = nla_nest_start(skb, TCA_OPTIONS);
1077 goto nla_put_failure;
1078 if (nla_put(skb, TCA_HTB_INIT, sizeof(gopt), &gopt) ||
1079 nla_put_u32(skb, TCA_HTB_DIRECT_QLEN, q->direct_qlen))
1080 goto nla_put_failure;
1082 return nla_nest_end(skb, nest);
1085 nla_nest_cancel(skb, nest);
1089 static int htb_dump_class(struct Qdisc *sch, unsigned long arg,
1090 struct sk_buff *skb, struct tcmsg *tcm)
1092 struct htb_class *cl = (struct htb_class *)arg;
1093 struct nlattr *nest;
1094 struct tc_htb_opt opt;
1096 /* Its safe to not acquire qdisc lock. As we hold RTNL,
1097 * no change can happen on the class parameters.
1099 tcm->tcm_parent = cl->parent ? cl->parent->common.classid : TC_H_ROOT;
1100 tcm->tcm_handle = cl->common.classid;
1101 if (!cl->level && cl->un.leaf.q)
1102 tcm->tcm_info = cl->un.leaf.q->handle;
1104 nest = nla_nest_start(skb, TCA_OPTIONS);
1106 goto nla_put_failure;
1108 memset(&opt, 0, sizeof(opt));
1110 psched_ratecfg_getrate(&opt.rate, &cl->rate);
1111 opt.buffer = PSCHED_NS2TICKS(cl->buffer);
1112 psched_ratecfg_getrate(&opt.ceil, &cl->ceil);
1113 opt.cbuffer = PSCHED_NS2TICKS(cl->cbuffer);
1114 opt.quantum = cl->quantum;
1115 opt.prio = cl->prio;
1116 opt.level = cl->level;
1117 if (nla_put(skb, TCA_HTB_PARMS, sizeof(opt), &opt))
1118 goto nla_put_failure;
1119 if ((cl->rate.rate_bytes_ps >= (1ULL << 32)) &&
1120 nla_put_u64_64bit(skb, TCA_HTB_RATE64, cl->rate.rate_bytes_ps,
1122 goto nla_put_failure;
1123 if ((cl->ceil.rate_bytes_ps >= (1ULL << 32)) &&
1124 nla_put_u64_64bit(skb, TCA_HTB_CEIL64, cl->ceil.rate_bytes_ps,
1126 goto nla_put_failure;
1128 return nla_nest_end(skb, nest);
1131 nla_nest_cancel(skb, nest);
1136 htb_dump_class_stats(struct Qdisc *sch, unsigned long arg, struct gnet_dump *d)
1138 struct htb_class *cl = (struct htb_class *)arg;
1141 if (!cl->level && cl->un.leaf.q)
1142 qlen = cl->un.leaf.q->q.qlen;
1143 cl->xstats.tokens = clamp_t(s64, PSCHED_NS2TICKS(cl->tokens),
1145 cl->xstats.ctokens = clamp_t(s64, PSCHED_NS2TICKS(cl->ctokens),
1148 if (gnet_stats_copy_basic(d, NULL, &cl->bstats) < 0 ||
1149 gnet_stats_copy_rate_est(d, NULL, &cl->rate_est) < 0 ||
1150 gnet_stats_copy_queue(d, NULL, &cl->qstats, qlen) < 0)
1153 return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1156 static int htb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1159 struct htb_class *cl = (struct htb_class *)arg;
1164 (new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1165 cl->common.classid)) == NULL)
1168 *old = qdisc_replace(sch, new, &cl->un.leaf.q);
1172 static struct Qdisc *htb_leaf(struct Qdisc *sch, unsigned long arg)
1174 struct htb_class *cl = (struct htb_class *)arg;
1175 return !cl->level ? cl->un.leaf.q : NULL;
1178 static void htb_qlen_notify(struct Qdisc *sch, unsigned long arg)
1180 struct htb_class *cl = (struct htb_class *)arg;
1182 if (cl->un.leaf.q->q.qlen == 0)
1183 htb_deactivate(qdisc_priv(sch), cl);
1186 static unsigned long htb_get(struct Qdisc *sch, u32 classid)
1188 struct htb_class *cl = htb_find(classid, sch);
1191 return (unsigned long)cl;
1194 static inline int htb_parent_last_child(struct htb_class *cl)
1197 /* the root class */
1199 if (cl->parent->children > 1)
1200 /* not the last child */
1205 static void htb_parent_to_leaf(struct htb_sched *q, struct htb_class *cl,
1206 struct Qdisc *new_q)
1208 struct htb_class *parent = cl->parent;
1210 WARN_ON(cl->level || !cl->un.leaf.q || cl->prio_activity);
1212 if (parent->cmode != HTB_CAN_SEND)
1213 htb_safe_rb_erase(&parent->pq_node,
1214 &q->hlevel[parent->level].wait_pq);
1217 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1218 INIT_LIST_HEAD(&parent->un.leaf.drop_list);
1219 parent->un.leaf.q = new_q ? new_q : &noop_qdisc;
1220 parent->tokens = parent->buffer;
1221 parent->ctokens = parent->cbuffer;
1222 parent->t_c = ktime_get_ns();
1223 parent->cmode = HTB_CAN_SEND;
1226 static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
1229 WARN_ON(!cl->un.leaf.q);
1230 qdisc_destroy(cl->un.leaf.q);
1232 gen_kill_estimator(&cl->bstats, &cl->rate_est);
1233 tcf_destroy_chain(&cl->filter_list);
1237 static void htb_destroy(struct Qdisc *sch)
1239 struct htb_sched *q = qdisc_priv(sch);
1240 struct hlist_node *next;
1241 struct htb_class *cl;
1244 cancel_work_sync(&q->work);
1245 qdisc_watchdog_cancel(&q->watchdog);
1246 /* This line used to be after htb_destroy_class call below
1247 * and surprisingly it worked in 2.4. But it must precede it
1248 * because filter need its target class alive to be able to call
1249 * unbind_filter on it (without Oops).
1251 tcf_destroy_chain(&q->filter_list);
1253 for (i = 0; i < q->clhash.hashsize; i++) {
1254 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode)
1255 tcf_destroy_chain(&cl->filter_list);
1257 for (i = 0; i < q->clhash.hashsize; i++) {
1258 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1260 htb_destroy_class(sch, cl);
1262 qdisc_class_hash_destroy(&q->clhash);
1263 __skb_queue_purge(&q->direct_queue);
1266 static int htb_delete(struct Qdisc *sch, unsigned long arg)
1268 struct htb_sched *q = qdisc_priv(sch);
1269 struct htb_class *cl = (struct htb_class *)arg;
1270 struct Qdisc *new_q = NULL;
1273 /* TODO: why don't allow to delete subtree ? references ? does
1274 * tc subsys guarantee us that in htb_destroy it holds no class
1275 * refs so that we can remove children safely there ?
1277 if (cl->children || cl->filter_cnt)
1280 if (!cl->level && htb_parent_last_child(cl)) {
1281 new_q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1282 cl->parent->common.classid);
1289 unsigned int qlen = cl->un.leaf.q->q.qlen;
1290 unsigned int backlog = cl->un.leaf.q->qstats.backlog;
1292 qdisc_reset(cl->un.leaf.q);
1293 qdisc_tree_reduce_backlog(cl->un.leaf.q, qlen, backlog);
1296 /* delete from hash and active; remainder in destroy_class */
1297 qdisc_class_hash_remove(&q->clhash, &cl->common);
1299 cl->parent->children--;
1301 if (cl->prio_activity)
1302 htb_deactivate(q, cl);
1304 if (cl->cmode != HTB_CAN_SEND)
1305 htb_safe_rb_erase(&cl->pq_node,
1306 &q->hlevel[cl->level].wait_pq);
1309 htb_parent_to_leaf(q, cl, new_q);
1311 BUG_ON(--cl->refcnt == 0);
1313 * This shouldn't happen: we "hold" one cops->get() when called
1314 * from tc_ctl_tclass; the destroy method is done from cops->put().
1317 sch_tree_unlock(sch);
1321 static void htb_put(struct Qdisc *sch, unsigned long arg)
1323 struct htb_class *cl = (struct htb_class *)arg;
1325 if (--cl->refcnt == 0)
1326 htb_destroy_class(sch, cl);
1329 static int htb_change_class(struct Qdisc *sch, u32 classid,
1330 u32 parentid, struct nlattr **tca,
1334 struct htb_sched *q = qdisc_priv(sch);
1335 struct htb_class *cl = (struct htb_class *)*arg, *parent;
1336 struct nlattr *opt = tca[TCA_OPTIONS];
1337 struct nlattr *tb[TCA_HTB_MAX + 1];
1338 struct tc_htb_opt *hopt;
1341 /* extract all subattrs from opt attr */
1345 err = nla_parse_nested(tb, TCA_HTB_MAX, opt, htb_policy);
1350 if (tb[TCA_HTB_PARMS] == NULL)
1353 parent = parentid == TC_H_ROOT ? NULL : htb_find(parentid, sch);
1355 hopt = nla_data(tb[TCA_HTB_PARMS]);
1356 if (!hopt->rate.rate || !hopt->ceil.rate)
1359 /* Keeping backward compatible with rate_table based iproute2 tc */
1360 if (hopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
1361 qdisc_put_rtab(qdisc_get_rtab(&hopt->rate, tb[TCA_HTB_RTAB]));
1363 if (hopt->ceil.linklayer == TC_LINKLAYER_UNAWARE)
1364 qdisc_put_rtab(qdisc_get_rtab(&hopt->ceil, tb[TCA_HTB_CTAB]));
1366 if (!cl) { /* new class */
1367 struct Qdisc *new_q;
1371 struct gnet_estimator opt;
1374 .nla_len = nla_attr_size(sizeof(est.opt)),
1375 .nla_type = TCA_RATE,
1378 /* 4s interval, 16s averaging constant */
1384 /* check for valid classid */
1385 if (!classid || TC_H_MAJ(classid ^ sch->handle) ||
1386 htb_find(classid, sch))
1389 /* check maximal depth */
1390 if (parent && parent->parent && parent->parent->level < 2) {
1391 pr_err("htb: tree is too deep\n");
1395 cl = kzalloc(sizeof(*cl), GFP_KERNEL);
1399 if (htb_rate_est || tca[TCA_RATE]) {
1400 err = gen_new_estimator(&cl->bstats, NULL,
1402 qdisc_root_sleeping_lock(sch),
1403 tca[TCA_RATE] ? : &est.nla);
1412 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
1413 RB_CLEAR_NODE(&cl->pq_node);
1415 for (prio = 0; prio < TC_HTB_NUMPRIO; prio++)
1416 RB_CLEAR_NODE(&cl->node[prio]);
1418 /* create leaf qdisc early because it uses kmalloc(GFP_KERNEL)
1419 * so that can't be used inside of sch_tree_lock
1420 * -- thanks to Karlis Peisenieks
1422 new_q = qdisc_create_dflt(sch->dev_queue,
1423 &pfifo_qdisc_ops, classid);
1425 if (parent && !parent->level) {
1426 unsigned int qlen = parent->un.leaf.q->q.qlen;
1427 unsigned int backlog = parent->un.leaf.q->qstats.backlog;
1429 /* turn parent into inner node */
1430 qdisc_reset(parent->un.leaf.q);
1431 qdisc_tree_reduce_backlog(parent->un.leaf.q, qlen, backlog);
1432 qdisc_destroy(parent->un.leaf.q);
1433 if (parent->prio_activity)
1434 htb_deactivate(q, parent);
1436 /* remove from evt list because of level change */
1437 if (parent->cmode != HTB_CAN_SEND) {
1438 htb_safe_rb_erase(&parent->pq_node, &q->hlevel[0].wait_pq);
1439 parent->cmode = HTB_CAN_SEND;
1441 parent->level = (parent->parent ? parent->parent->level
1442 : TC_HTB_MAXDEPTH) - 1;
1443 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1445 /* leaf (we) needs elementary qdisc */
1446 cl->un.leaf.q = new_q ? new_q : &noop_qdisc;
1448 cl->common.classid = classid;
1449 cl->parent = parent;
1451 /* set class to be in HTB_CAN_SEND state */
1452 cl->tokens = PSCHED_TICKS2NS(hopt->buffer);
1453 cl->ctokens = PSCHED_TICKS2NS(hopt->cbuffer);
1454 cl->mbuffer = 60ULL * NSEC_PER_SEC; /* 1min */
1455 cl->t_c = ktime_get_ns();
1456 cl->cmode = HTB_CAN_SEND;
1458 /* attach to the hash list and parent's family */
1459 qdisc_class_hash_insert(&q->clhash, &cl->common);
1463 if (tca[TCA_RATE]) {
1464 spinlock_t *lock = qdisc_root_sleeping_lock(sch);
1466 err = gen_replace_estimator(&cl->bstats, NULL,
1476 rate64 = tb[TCA_HTB_RATE64] ? nla_get_u64(tb[TCA_HTB_RATE64]) : 0;
1478 ceil64 = tb[TCA_HTB_CEIL64] ? nla_get_u64(tb[TCA_HTB_CEIL64]) : 0;
1480 psched_ratecfg_precompute(&cl->rate, &hopt->rate, rate64);
1481 psched_ratecfg_precompute(&cl->ceil, &hopt->ceil, ceil64);
1483 /* it used to be a nasty bug here, we have to check that node
1484 * is really leaf before changing cl->un.leaf !
1487 u64 quantum = cl->rate.rate_bytes_ps;
1489 do_div(quantum, q->rate2quantum);
1490 cl->quantum = min_t(u64, quantum, INT_MAX);
1492 if (!hopt->quantum && cl->quantum < 1000) {
1493 pr_warn("HTB: quantum of class %X is small. Consider r2q change.\n",
1494 cl->common.classid);
1497 if (!hopt->quantum && cl->quantum > 200000) {
1498 pr_warn("HTB: quantum of class %X is big. Consider r2q change.\n",
1499 cl->common.classid);
1500 cl->quantum = 200000;
1503 cl->quantum = hopt->quantum;
1504 if ((cl->prio = hopt->prio) >= TC_HTB_NUMPRIO)
1505 cl->prio = TC_HTB_NUMPRIO - 1;
1508 cl->buffer = PSCHED_TICKS2NS(hopt->buffer);
1509 cl->cbuffer = PSCHED_TICKS2NS(hopt->cbuffer);
1511 sch_tree_unlock(sch);
1513 qdisc_class_hash_grow(sch, &q->clhash);
1515 *arg = (unsigned long)cl;
1522 static struct tcf_proto __rcu **htb_find_tcf(struct Qdisc *sch,
1525 struct htb_sched *q = qdisc_priv(sch);
1526 struct htb_class *cl = (struct htb_class *)arg;
1527 struct tcf_proto __rcu **fl = cl ? &cl->filter_list : &q->filter_list;
1532 static unsigned long htb_bind_filter(struct Qdisc *sch, unsigned long parent,
1535 struct htb_class *cl = htb_find(classid, sch);
1537 /*if (cl && !cl->level) return 0;
1538 * The line above used to be there to prevent attaching filters to
1539 * leaves. But at least tc_index filter uses this just to get class
1540 * for other reasons so that we have to allow for it.
1542 * 19.6.2002 As Werner explained it is ok - bind filter is just
1543 * another way to "lock" the class - unlike "get" this lock can
1544 * be broken by class during destroy IIUC.
1548 return (unsigned long)cl;
1551 static void htb_unbind_filter(struct Qdisc *sch, unsigned long arg)
1553 struct htb_class *cl = (struct htb_class *)arg;
1559 static void htb_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1561 struct htb_sched *q = qdisc_priv(sch);
1562 struct htb_class *cl;
1568 for (i = 0; i < q->clhash.hashsize; i++) {
1569 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1570 if (arg->count < arg->skip) {
1574 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1583 static const struct Qdisc_class_ops htb_class_ops = {
1586 .qlen_notify = htb_qlen_notify,
1589 .change = htb_change_class,
1590 .delete = htb_delete,
1592 .tcf_chain = htb_find_tcf,
1593 .bind_tcf = htb_bind_filter,
1594 .unbind_tcf = htb_unbind_filter,
1595 .dump = htb_dump_class,
1596 .dump_stats = htb_dump_class_stats,
1599 static struct Qdisc_ops htb_qdisc_ops __read_mostly = {
1600 .cl_ops = &htb_class_ops,
1602 .priv_size = sizeof(struct htb_sched),
1603 .enqueue = htb_enqueue,
1604 .dequeue = htb_dequeue,
1605 .peek = qdisc_peek_dequeued,
1609 .destroy = htb_destroy,
1611 .owner = THIS_MODULE,
1614 static int __init htb_module_init(void)
1616 return register_qdisc(&htb_qdisc_ops);
1618 static void __exit htb_module_exit(void)
1620 unregister_qdisc(&htb_qdisc_ops);
1623 module_init(htb_module_init)
1624 module_exit(htb_module_exit)
1625 MODULE_LICENSE("GPL");