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 /* used internaly to keep status of single class */
70 HTB_CANT_SEND, /* class can't send and can't borrow */
71 HTB_MAY_BORROW, /* class can't send but may borrow */
72 HTB_CAN_SEND /* class can send */
75 /* interior & leaf nodes; props specific to leaves are marked L: */
77 struct Qdisc_class_common common;
78 /* general class parameters */
79 struct gnet_stats_basic_packed bstats;
80 struct gnet_stats_queue qstats;
81 struct gnet_stats_rate_est rate_est;
82 struct tc_htb_xstats xstats; /* our special stats */
83 int refcnt; /* usage count of this class */
86 int level; /* our level (see above) */
87 unsigned int children;
88 struct htb_class *parent; /* parent class */
90 int prio; /* these two are used only by leaves... */
91 int quantum; /* but stored for parent-to-leaf return */
94 struct htb_class_leaf {
96 int deficit[TC_HTB_MAXDEPTH];
97 struct list_head drop_list;
99 struct htb_class_inner {
100 struct rb_root feed[TC_HTB_NUMPRIO]; /* feed trees */
101 struct rb_node *ptr[TC_HTB_NUMPRIO]; /* current class ptr */
102 /* When class changes from state 1->2 and disconnects from
103 * parent's feed then we lost ptr value and start from the
104 * first child again. Here we store classid of the
105 * last valid ptr (used when ptr is NULL).
107 u32 last_ptr_id[TC_HTB_NUMPRIO];
110 struct rb_node node[TC_HTB_NUMPRIO]; /* node for self or feed tree */
111 struct rb_node pq_node; /* node for event queue */
112 psched_time_t pq_key;
114 int prio_activity; /* for which prios are we active */
115 enum htb_cmode cmode; /* current mode of the class */
117 /* class attached filters */
118 struct tcf_proto *filter_list;
121 /* token bucket parameters */
122 struct psched_ratecfg rate;
123 struct psched_ratecfg ceil;
124 s64 buffer, cbuffer; /* token bucket depth/rate */
125 psched_tdiff_t mbuffer; /* max wait time */
126 s64 tokens, ctokens; /* current number of tokens */
127 psched_time_t t_c; /* checkpoint time */
131 struct Qdisc_class_hash clhash;
132 struct list_head drops[TC_HTB_NUMPRIO];/* active leaves (for drops) */
134 /* self list - roots of self generating tree */
135 struct rb_root row[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
136 int row_mask[TC_HTB_MAXDEPTH];
137 struct rb_node *ptr[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
138 u32 last_ptr_id[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
140 /* self wait list - roots of wait PQs per row */
141 struct rb_root wait_pq[TC_HTB_MAXDEPTH];
143 /* time of nearest event per level (row) */
144 psched_time_t near_ev_cache[TC_HTB_MAXDEPTH];
146 int defcls; /* class where unclassified flows go to */
148 /* filters for qdisc itself */
149 struct tcf_proto *filter_list;
151 int rate2quantum; /* quant = rate / rate2quantum */
152 psched_time_t now; /* cached dequeue time */
153 struct qdisc_watchdog watchdog;
155 /* non shaped skbs; let them go directly thru */
156 struct sk_buff_head direct_queue;
157 int direct_qlen; /* max qlen of above */
161 #define HTB_WARN_TOOMANYEVENTS 0x1
162 unsigned int warned; /* only one warning */
163 struct work_struct work;
166 /* find class in global hash table using given handle */
167 static inline struct htb_class *htb_find(u32 handle, struct Qdisc *sch)
169 struct htb_sched *q = qdisc_priv(sch);
170 struct Qdisc_class_common *clc;
172 clc = qdisc_class_find(&q->clhash, handle);
175 return container_of(clc, struct htb_class, common);
179 * htb_classify - classify a packet into class
181 * It returns NULL if the packet should be dropped or -1 if the packet
182 * should be passed directly thru. In all other cases leaf class is returned.
183 * We allow direct class selection by classid in priority. The we examine
184 * filters in qdisc and in inner nodes (if higher filter points to the inner
185 * node). If we end up with classid MAJOR:0 we enqueue the skb into special
186 * internal fifo (direct). These packets then go directly thru. If we still
187 * have no valid leaf we try to use MAJOR:default leaf. It still unsuccessful
188 * then finish and return direct queue.
190 #define HTB_DIRECT ((struct htb_class *)-1L)
192 static struct htb_class *htb_classify(struct sk_buff *skb, struct Qdisc *sch,
195 struct htb_sched *q = qdisc_priv(sch);
196 struct htb_class *cl;
197 struct tcf_result res;
198 struct tcf_proto *tcf;
201 /* allow to select class by setting skb->priority to valid classid;
202 * note that nfmark can be used too by attaching filter fw with no
205 if (skb->priority == sch->handle)
206 return HTB_DIRECT; /* X:0 (direct flow) selected */
207 cl = htb_find(skb->priority, sch);
208 if (cl && cl->level == 0)
211 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
212 tcf = q->filter_list;
213 while (tcf && (result = tc_classify(skb, tcf, &res)) >= 0) {
214 #ifdef CONFIG_NET_CLS_ACT
218 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
223 cl = (void *)res.class;
225 if (res.classid == sch->handle)
226 return HTB_DIRECT; /* X:0 (direct flow) */
227 cl = htb_find(res.classid, sch);
229 break; /* filter selected invalid classid */
232 return cl; /* we hit leaf; return it */
234 /* we have got inner class; apply inner filter chain */
235 tcf = cl->filter_list;
237 /* classification failed; try to use default class */
238 cl = htb_find(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
239 if (!cl || cl->level)
240 return HTB_DIRECT; /* bad default .. this is safe bet */
245 * htb_add_to_id_tree - adds class to the round robin list
247 * Routine adds class to the list (actually tree) sorted by classid.
248 * Make sure that class is not already on such list for given prio.
250 static void htb_add_to_id_tree(struct rb_root *root,
251 struct htb_class *cl, int prio)
253 struct rb_node **p = &root->rb_node, *parent = NULL;
258 c = rb_entry(parent, struct htb_class, node[prio]);
260 if (cl->common.classid > c->common.classid)
261 p = &parent->rb_right;
263 p = &parent->rb_left;
265 rb_link_node(&cl->node[prio], parent, p);
266 rb_insert_color(&cl->node[prio], root);
270 * htb_add_to_wait_tree - adds class to the event queue with delay
272 * The class is added to priority event queue to indicate that class will
273 * change its mode in cl->pq_key microseconds. Make sure that class is not
274 * already in the queue.
276 static void htb_add_to_wait_tree(struct htb_sched *q,
277 struct htb_class *cl, s64 delay)
279 struct rb_node **p = &q->wait_pq[cl->level].rb_node, *parent = NULL;
281 cl->pq_key = q->now + delay;
282 if (cl->pq_key == q->now)
285 /* update the nearest event cache */
286 if (q->near_ev_cache[cl->level] > cl->pq_key)
287 q->near_ev_cache[cl->level] = cl->pq_key;
292 c = rb_entry(parent, struct htb_class, pq_node);
293 if (cl->pq_key >= c->pq_key)
294 p = &parent->rb_right;
296 p = &parent->rb_left;
298 rb_link_node(&cl->pq_node, parent, p);
299 rb_insert_color(&cl->pq_node, &q->wait_pq[cl->level]);
303 * htb_next_rb_node - finds next node in binary tree
305 * When we are past last key we return NULL.
306 * Average complexity is 2 steps per call.
308 static inline void htb_next_rb_node(struct rb_node **n)
314 * htb_add_class_to_row - add class to its row
316 * The class is added to row at priorities marked in mask.
317 * It does nothing if mask == 0.
319 static inline void htb_add_class_to_row(struct htb_sched *q,
320 struct htb_class *cl, int mask)
322 q->row_mask[cl->level] |= mask;
324 int prio = ffz(~mask);
325 mask &= ~(1 << prio);
326 htb_add_to_id_tree(q->row[cl->level] + prio, cl, prio);
330 /* If this triggers, it is a bug in this code, but it need not be fatal */
331 static void htb_safe_rb_erase(struct rb_node *rb, struct rb_root *root)
333 if (RB_EMPTY_NODE(rb)) {
343 * htb_remove_class_from_row - removes class from its row
345 * The class is removed from row at priorities marked in mask.
346 * It does nothing if mask == 0.
348 static inline void htb_remove_class_from_row(struct htb_sched *q,
349 struct htb_class *cl, int mask)
354 int prio = ffz(~mask);
356 mask &= ~(1 << prio);
357 if (q->ptr[cl->level][prio] == cl->node + prio)
358 htb_next_rb_node(q->ptr[cl->level] + prio);
360 htb_safe_rb_erase(cl->node + prio, q->row[cl->level] + prio);
361 if (!q->row[cl->level][prio].rb_node)
364 q->row_mask[cl->level] &= ~m;
368 * htb_activate_prios - creates active classe's feed chain
370 * The class is connected to ancestors and/or appropriate rows
371 * for priorities it is participating on. cl->cmode must be new
372 * (activated) mode. It does nothing if cl->prio_activity == 0.
374 static void htb_activate_prios(struct htb_sched *q, struct htb_class *cl)
376 struct htb_class *p = cl->parent;
377 long m, mask = cl->prio_activity;
379 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
385 if (p->un.inner.feed[prio].rb_node)
386 /* parent already has its feed in use so that
387 * reset bit in mask as parent is already ok
389 mask &= ~(1 << prio);
391 htb_add_to_id_tree(p->un.inner.feed + prio, cl, prio);
393 p->prio_activity |= mask;
398 if (cl->cmode == HTB_CAN_SEND && mask)
399 htb_add_class_to_row(q, cl, mask);
403 * htb_deactivate_prios - remove class from feed chain
405 * cl->cmode must represent old mode (before deactivation). It does
406 * nothing if cl->prio_activity == 0. Class is removed from all feed
409 static void htb_deactivate_prios(struct htb_sched *q, struct htb_class *cl)
411 struct htb_class *p = cl->parent;
412 long m, mask = cl->prio_activity;
414 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
421 if (p->un.inner.ptr[prio] == cl->node + prio) {
422 /* we are removing child which is pointed to from
423 * parent feed - forget the pointer but remember
426 p->un.inner.last_ptr_id[prio] = cl->common.classid;
427 p->un.inner.ptr[prio] = NULL;
430 htb_safe_rb_erase(cl->node + prio, p->un.inner.feed + prio);
432 if (!p->un.inner.feed[prio].rb_node)
436 p->prio_activity &= ~mask;
441 if (cl->cmode == HTB_CAN_SEND && mask)
442 htb_remove_class_from_row(q, cl, mask);
445 static inline s64 htb_lowater(const struct htb_class *cl)
448 return cl->cmode != HTB_CANT_SEND ? -cl->cbuffer : 0;
452 static inline s64 htb_hiwater(const struct htb_class *cl)
455 return cl->cmode == HTB_CAN_SEND ? -cl->buffer : 0;
462 * htb_class_mode - computes and returns current class mode
464 * It computes cl's mode at time cl->t_c+diff and returns it. If mode
465 * is not HTB_CAN_SEND then cl->pq_key is updated to time difference
466 * from now to time when cl will change its state.
467 * Also it is worth to note that class mode doesn't change simply
468 * at cl->{c,}tokens == 0 but there can rather be hysteresis of
469 * 0 .. -cl->{c,}buffer range. It is meant to limit number of
470 * mode transitions per time unit. The speed gain is about 1/6.
472 static inline enum htb_cmode
473 htb_class_mode(struct htb_class *cl, s64 *diff)
477 if ((toks = (cl->ctokens + *diff)) < htb_lowater(cl)) {
479 return HTB_CANT_SEND;
482 if ((toks = (cl->tokens + *diff)) >= htb_hiwater(cl))
486 return HTB_MAY_BORROW;
490 * htb_change_class_mode - changes classe's mode
492 * This should be the only way how to change classe's mode under normal
493 * cirsumstances. Routine will update feed lists linkage, change mode
494 * and add class to the wait event queue if appropriate. New mode should
495 * be different from old one and cl->pq_key has to be valid if changing
496 * to mode other than HTB_CAN_SEND (see htb_add_to_wait_tree).
499 htb_change_class_mode(struct htb_sched *q, struct htb_class *cl, s64 *diff)
501 enum htb_cmode new_mode = htb_class_mode(cl, diff);
503 if (new_mode == cl->cmode)
506 if (cl->prio_activity) { /* not necessary: speed optimization */
507 if (cl->cmode != HTB_CANT_SEND)
508 htb_deactivate_prios(q, cl);
509 cl->cmode = new_mode;
510 if (new_mode != HTB_CANT_SEND)
511 htb_activate_prios(q, cl);
513 cl->cmode = new_mode;
517 * htb_activate - inserts leaf cl into appropriate active feeds
519 * Routine learns (new) priority of leaf and activates feed chain
520 * for the prio. It can be called on already active leaf safely.
521 * It also adds leaf into droplist.
523 static inline void htb_activate(struct htb_sched *q, struct htb_class *cl)
525 WARN_ON(cl->level || !cl->un.leaf.q || !cl->un.leaf.q->q.qlen);
527 if (!cl->prio_activity) {
528 cl->prio_activity = 1 << cl->prio;
529 htb_activate_prios(q, cl);
530 list_add_tail(&cl->un.leaf.drop_list,
531 q->drops + cl->prio);
536 * htb_deactivate - remove leaf cl from active feeds
538 * Make sure that leaf is active. In the other words it can't be called
539 * with non-active leaf. It also removes class from the drop list.
541 static inline void htb_deactivate(struct htb_sched *q, struct htb_class *cl)
543 WARN_ON(!cl->prio_activity);
545 htb_deactivate_prios(q, cl);
546 cl->prio_activity = 0;
547 list_del_init(&cl->un.leaf.drop_list);
550 static int htb_enqueue(struct sk_buff *skb, struct Qdisc *sch)
552 int uninitialized_var(ret);
553 struct htb_sched *q = qdisc_priv(sch);
554 struct htb_class *cl = htb_classify(skb, sch, &ret);
556 if (cl == HTB_DIRECT) {
557 /* enqueue to helper queue */
558 if (q->direct_queue.qlen < q->direct_qlen) {
559 __skb_queue_tail(&q->direct_queue, skb);
562 return qdisc_drop(skb, sch);
564 #ifdef CONFIG_NET_CLS_ACT
566 if (ret & __NET_XMIT_BYPASS)
571 } else if ((ret = qdisc_enqueue(skb, cl->un.leaf.q)) != NET_XMIT_SUCCESS) {
572 if (net_xmit_drop_count(ret)) {
582 return NET_XMIT_SUCCESS;
585 static inline void htb_accnt_tokens(struct htb_class *cl, int bytes, s64 diff)
587 s64 toks = diff + cl->tokens;
589 if (toks > cl->buffer)
591 toks -= (s64) psched_l2t_ns(&cl->rate, bytes);
592 if (toks <= -cl->mbuffer)
593 toks = 1 - cl->mbuffer;
598 static inline void htb_accnt_ctokens(struct htb_class *cl, int bytes, s64 diff)
600 s64 toks = diff + cl->ctokens;
602 if (toks > cl->cbuffer)
604 toks -= (s64) psched_l2t_ns(&cl->ceil, bytes);
605 if (toks <= -cl->mbuffer)
606 toks = 1 - cl->mbuffer;
612 * htb_charge_class - charges amount "bytes" to leaf and ancestors
614 * Routine assumes that packet "bytes" long was dequeued from leaf cl
615 * borrowing from "level". It accounts bytes to ceil leaky bucket for
616 * leaf and all ancestors and to rate bucket for ancestors at levels
617 * "level" and higher. It also handles possible change of mode resulting
618 * from the update. Note that mode can also increase here (MAY_BORROW to
619 * CAN_SEND) because we can use more precise clock that event queue here.
620 * In such case we remove class from event queue first.
622 static void htb_charge_class(struct htb_sched *q, struct htb_class *cl,
623 int level, struct sk_buff *skb)
625 int bytes = qdisc_pkt_len(skb);
626 enum htb_cmode old_mode;
630 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
631 if (cl->level >= level) {
632 if (cl->level == level)
634 htb_accnt_tokens(cl, bytes, diff);
636 cl->xstats.borrows++;
637 cl->tokens += diff; /* we moved t_c; update tokens */
639 htb_accnt_ctokens(cl, bytes, diff);
642 old_mode = cl->cmode;
644 htb_change_class_mode(q, cl, &diff);
645 if (old_mode != cl->cmode) {
646 if (old_mode != HTB_CAN_SEND)
647 htb_safe_rb_erase(&cl->pq_node, q->wait_pq + cl->level);
648 if (cl->cmode != HTB_CAN_SEND)
649 htb_add_to_wait_tree(q, cl, diff);
652 /* update basic stats except for leaves which are already updated */
654 bstats_update(&cl->bstats, skb);
661 * htb_do_events - make mode changes to classes at the level
663 * Scans event queue for pending events and applies them. Returns time of
664 * next pending event (0 for no event in pq, q->now for too many events).
665 * Note: Applied are events whose have cl->pq_key <= q->now.
667 static psched_time_t htb_do_events(struct htb_sched *q, int level,
670 /* don't run for longer than 2 jiffies; 2 is used instead of
671 * 1 to simplify things when jiffy is going to be incremented
674 unsigned long stop_at = start + 2;
675 while (time_before(jiffies, stop_at)) {
676 struct htb_class *cl;
678 struct rb_node *p = rb_first(&q->wait_pq[level]);
683 cl = rb_entry(p, struct htb_class, pq_node);
684 if (cl->pq_key > q->now)
687 htb_safe_rb_erase(p, q->wait_pq + level);
688 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
689 htb_change_class_mode(q, cl, &diff);
690 if (cl->cmode != HTB_CAN_SEND)
691 htb_add_to_wait_tree(q, cl, diff);
694 /* too much load - let's continue after a break for scheduling */
695 if (!(q->warned & HTB_WARN_TOOMANYEVENTS)) {
696 pr_warning("htb: too many events!\n");
697 q->warned |= HTB_WARN_TOOMANYEVENTS;
703 /* Returns class->node+prio from id-tree where classe's id is >= id. NULL
704 * is no such one exists.
706 static struct rb_node *htb_id_find_next_upper(int prio, struct rb_node *n,
709 struct rb_node *r = NULL;
711 struct htb_class *cl =
712 rb_entry(n, struct htb_class, node[prio]);
714 if (id > cl->common.classid) {
716 } else if (id < cl->common.classid) {
727 * htb_lookup_leaf - returns next leaf class in DRR order
729 * Find leaf where current feed pointers points to.
731 static struct htb_class *htb_lookup_leaf(struct rb_root *tree, int prio,
732 struct rb_node **pptr, u32 * pid)
736 struct rb_node *root;
737 struct rb_node **pptr;
739 } stk[TC_HTB_MAXDEPTH], *sp = stk;
741 BUG_ON(!tree->rb_node);
742 sp->root = tree->rb_node;
746 for (i = 0; i < 65535; i++) {
747 if (!*sp->pptr && *sp->pid) {
748 /* ptr was invalidated but id is valid - try to recover
749 * the original or next ptr
752 htb_id_find_next_upper(prio, sp->root, *sp->pid);
754 *sp->pid = 0; /* ptr is valid now so that remove this hint as it
755 * can become out of date quickly
757 if (!*sp->pptr) { /* we are at right end; rewind & go up */
758 *sp->pptr = sp->root;
759 while ((*sp->pptr)->rb_left)
760 *sp->pptr = (*sp->pptr)->rb_left;
767 htb_next_rb_node(sp->pptr);
770 struct htb_class *cl;
771 cl = rb_entry(*sp->pptr, struct htb_class, node[prio]);
774 (++sp)->root = cl->un.inner.feed[prio].rb_node;
775 sp->pptr = cl->un.inner.ptr + prio;
776 sp->pid = cl->un.inner.last_ptr_id + prio;
783 /* dequeues packet at given priority and level; call only if
784 * you are sure that there is active class at prio/level
786 static struct sk_buff *htb_dequeue_tree(struct htb_sched *q, int prio,
789 struct sk_buff *skb = NULL;
790 struct htb_class *cl, *start;
791 /* look initial class up in the row */
792 start = cl = htb_lookup_leaf(q->row[level] + prio, prio,
793 q->ptr[level] + prio,
794 q->last_ptr_id[level] + prio);
801 /* class can be empty - it is unlikely but can be true if leaf
802 * qdisc drops packets in enqueue routine or if someone used
803 * graft operation on the leaf since last dequeue;
804 * simply deactivate and skip such class
806 if (unlikely(cl->un.leaf.q->q.qlen == 0)) {
807 struct htb_class *next;
808 htb_deactivate(q, cl);
810 /* row/level might become empty */
811 if ((q->row_mask[level] & (1 << prio)) == 0)
814 next = htb_lookup_leaf(q->row[level] + prio,
815 prio, q->ptr[level] + prio,
816 q->last_ptr_id[level] + prio);
818 if (cl == start) /* fix start if we just deleted it */
824 skb = cl->un.leaf.q->dequeue(cl->un.leaf.q);
825 if (likely(skb != NULL))
828 qdisc_warn_nonwc("htb", cl->un.leaf.q);
829 htb_next_rb_node((level ? cl->parent->un.inner.ptr : q->
831 cl = htb_lookup_leaf(q->row[level] + prio, prio,
832 q->ptr[level] + prio,
833 q->last_ptr_id[level] + prio);
835 } while (cl != start);
837 if (likely(skb != NULL)) {
838 bstats_update(&cl->bstats, skb);
839 cl->un.leaf.deficit[level] -= qdisc_pkt_len(skb);
840 if (cl->un.leaf.deficit[level] < 0) {
841 cl->un.leaf.deficit[level] += cl->quantum;
842 htb_next_rb_node((level ? cl->parent->un.inner.ptr : q->
845 /* this used to be after charge_class but this constelation
846 * gives us slightly better performance
848 if (!cl->un.leaf.q->q.qlen)
849 htb_deactivate(q, cl);
850 htb_charge_class(q, cl, level, skb);
855 static struct sk_buff *htb_dequeue(struct Qdisc *sch)
858 struct htb_sched *q = qdisc_priv(sch);
860 psched_time_t next_event;
861 unsigned long start_at;
863 /* try to dequeue direct packets as high prio (!) to minimize cpu work */
864 skb = __skb_dequeue(&q->direct_queue);
867 qdisc_bstats_update(sch, skb);
868 qdisc_unthrottled(sch);
875 q->now = ktime_to_ns(ktime_get());
878 next_event = q->now + 5LLU * NSEC_PER_SEC;
880 for (level = 0; level < TC_HTB_MAXDEPTH; level++) {
881 /* common case optimization - skip event handler quickly */
885 if (q->now >= q->near_ev_cache[level]) {
886 event = htb_do_events(q, level, start_at);
888 event = q->now + NSEC_PER_SEC;
889 q->near_ev_cache[level] = event;
891 event = q->near_ev_cache[level];
893 if (next_event > event)
896 m = ~q->row_mask[level];
897 while (m != (int)(-1)) {
901 skb = htb_dequeue_tree(q, prio, level);
902 if (likely(skb != NULL))
906 sch->qstats.overlimits++;
907 if (likely(next_event > q->now)) {
908 if (!test_bit(__QDISC_STATE_DEACTIVATED,
909 &qdisc_root_sleeping(q->watchdog.qdisc)->state)) {
910 ktime_t time = ns_to_ktime(next_event);
911 qdisc_throttled(q->watchdog.qdisc);
912 hrtimer_start(&q->watchdog.timer, time,
916 schedule_work(&q->work);
922 /* try to drop from each class (by prio) until one succeed */
923 static unsigned int htb_drop(struct Qdisc *sch)
925 struct htb_sched *q = qdisc_priv(sch);
928 for (prio = TC_HTB_NUMPRIO - 1; prio >= 0; prio--) {
930 list_for_each(p, q->drops + prio) {
931 struct htb_class *cl = list_entry(p, struct htb_class,
934 if (cl->un.leaf.q->ops->drop &&
935 (len = cl->un.leaf.q->ops->drop(cl->un.leaf.q))) {
937 if (!cl->un.leaf.q->q.qlen)
938 htb_deactivate(q, cl);
946 /* reset all classes */
947 /* always caled under BH & queue lock */
948 static void htb_reset(struct Qdisc *sch)
950 struct htb_sched *q = qdisc_priv(sch);
951 struct htb_class *cl;
954 for (i = 0; i < q->clhash.hashsize; i++) {
955 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
957 memset(&cl->un.inner, 0, sizeof(cl->un.inner));
960 qdisc_reset(cl->un.leaf.q);
961 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
963 cl->prio_activity = 0;
964 cl->cmode = HTB_CAN_SEND;
968 qdisc_watchdog_cancel(&q->watchdog);
969 __skb_queue_purge(&q->direct_queue);
971 memset(q->row, 0, sizeof(q->row));
972 memset(q->row_mask, 0, sizeof(q->row_mask));
973 memset(q->wait_pq, 0, sizeof(q->wait_pq));
974 memset(q->ptr, 0, sizeof(q->ptr));
975 for (i = 0; i < TC_HTB_NUMPRIO; i++)
976 INIT_LIST_HEAD(q->drops + i);
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 },
986 static void htb_work_func(struct work_struct *work)
988 struct htb_sched *q = container_of(work, struct htb_sched, work);
989 struct Qdisc *sch = q->watchdog.qdisc;
991 __netif_schedule(qdisc_root(sch));
994 static int htb_init(struct Qdisc *sch, struct nlattr *opt)
996 struct htb_sched *q = qdisc_priv(sch);
997 struct nlattr *tb[TCA_HTB_INIT + 1];
998 struct tc_htb_glob *gopt;
1005 err = nla_parse_nested(tb, TCA_HTB_INIT, opt, htb_policy);
1009 if (tb[TCA_HTB_INIT] == NULL) {
1010 pr_err("HTB: hey probably you have bad tc tool ?\n");
1013 gopt = nla_data(tb[TCA_HTB_INIT]);
1014 if (gopt->version != HTB_VER >> 16) {
1015 pr_err("HTB: need tc/htb version %d (minor is %d), you have %d\n",
1016 HTB_VER >> 16, HTB_VER & 0xffff, gopt->version);
1020 err = qdisc_class_hash_init(&q->clhash);
1023 for (i = 0; i < TC_HTB_NUMPRIO; i++)
1024 INIT_LIST_HEAD(q->drops + i);
1026 qdisc_watchdog_init(&q->watchdog, sch);
1027 INIT_WORK(&q->work, htb_work_func);
1028 skb_queue_head_init(&q->direct_queue);
1030 q->direct_qlen = qdisc_dev(sch)->tx_queue_len;
1031 if (q->direct_qlen < 2) /* some devices have zero tx_queue_len */
1034 if ((q->rate2quantum = gopt->rate2quantum) < 1)
1035 q->rate2quantum = 1;
1036 q->defcls = gopt->defcls;
1041 static int htb_dump(struct Qdisc *sch, struct sk_buff *skb)
1043 spinlock_t *root_lock = qdisc_root_sleeping_lock(sch);
1044 struct htb_sched *q = qdisc_priv(sch);
1045 struct nlattr *nest;
1046 struct tc_htb_glob gopt;
1048 spin_lock_bh(root_lock);
1050 gopt.direct_pkts = q->direct_pkts;
1051 gopt.version = HTB_VER;
1052 gopt.rate2quantum = q->rate2quantum;
1053 gopt.defcls = q->defcls;
1056 nest = nla_nest_start(skb, TCA_OPTIONS);
1058 goto nla_put_failure;
1059 if (nla_put(skb, TCA_HTB_INIT, sizeof(gopt), &gopt))
1060 goto nla_put_failure;
1061 nla_nest_end(skb, nest);
1063 spin_unlock_bh(root_lock);
1067 spin_unlock_bh(root_lock);
1068 nla_nest_cancel(skb, nest);
1072 static int htb_dump_class(struct Qdisc *sch, unsigned long arg,
1073 struct sk_buff *skb, struct tcmsg *tcm)
1075 struct htb_class *cl = (struct htb_class *)arg;
1076 spinlock_t *root_lock = qdisc_root_sleeping_lock(sch);
1077 struct nlattr *nest;
1078 struct tc_htb_opt opt;
1080 spin_lock_bh(root_lock);
1081 tcm->tcm_parent = cl->parent ? cl->parent->common.classid : TC_H_ROOT;
1082 tcm->tcm_handle = cl->common.classid;
1083 if (!cl->level && cl->un.leaf.q)
1084 tcm->tcm_info = cl->un.leaf.q->handle;
1086 nest = nla_nest_start(skb, TCA_OPTIONS);
1088 goto nla_put_failure;
1090 memset(&opt, 0, sizeof(opt));
1092 opt.rate.rate = psched_ratecfg_getrate(&cl->rate);
1093 opt.buffer = PSCHED_NS2TICKS(cl->buffer);
1094 opt.ceil.rate = psched_ratecfg_getrate(&cl->ceil);
1095 opt.cbuffer = PSCHED_NS2TICKS(cl->cbuffer);
1096 opt.quantum = cl->quantum;
1097 opt.prio = cl->prio;
1098 opt.level = cl->level;
1099 if (nla_put(skb, TCA_HTB_PARMS, sizeof(opt), &opt))
1100 goto nla_put_failure;
1102 nla_nest_end(skb, nest);
1103 spin_unlock_bh(root_lock);
1107 spin_unlock_bh(root_lock);
1108 nla_nest_cancel(skb, nest);
1113 htb_dump_class_stats(struct Qdisc *sch, unsigned long arg, struct gnet_dump *d)
1115 struct htb_class *cl = (struct htb_class *)arg;
1117 if (!cl->level && cl->un.leaf.q)
1118 cl->qstats.qlen = cl->un.leaf.q->q.qlen;
1119 cl->xstats.tokens = cl->tokens;
1120 cl->xstats.ctokens = cl->ctokens;
1122 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
1123 gnet_stats_copy_rate_est(d, NULL, &cl->rate_est) < 0 ||
1124 gnet_stats_copy_queue(d, &cl->qstats) < 0)
1127 return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1130 static int htb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1133 struct htb_class *cl = (struct htb_class *)arg;
1138 (new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1139 cl->common.classid)) == NULL)
1143 *old = cl->un.leaf.q;
1144 cl->un.leaf.q = new;
1146 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
1149 sch_tree_unlock(sch);
1153 static struct Qdisc *htb_leaf(struct Qdisc *sch, unsigned long arg)
1155 struct htb_class *cl = (struct htb_class *)arg;
1156 return !cl->level ? cl->un.leaf.q : NULL;
1159 static void htb_qlen_notify(struct Qdisc *sch, unsigned long arg)
1161 struct htb_class *cl = (struct htb_class *)arg;
1163 if (cl->un.leaf.q->q.qlen == 0)
1164 htb_deactivate(qdisc_priv(sch), cl);
1167 static unsigned long htb_get(struct Qdisc *sch, u32 classid)
1169 struct htb_class *cl = htb_find(classid, sch);
1172 return (unsigned long)cl;
1175 static inline int htb_parent_last_child(struct htb_class *cl)
1178 /* the root class */
1180 if (cl->parent->children > 1)
1181 /* not the last child */
1186 static void htb_parent_to_leaf(struct htb_sched *q, struct htb_class *cl,
1187 struct Qdisc *new_q)
1189 struct htb_class *parent = cl->parent;
1191 WARN_ON(cl->level || !cl->un.leaf.q || cl->prio_activity);
1193 if (parent->cmode != HTB_CAN_SEND)
1194 htb_safe_rb_erase(&parent->pq_node, q->wait_pq + parent->level);
1197 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1198 INIT_LIST_HEAD(&parent->un.leaf.drop_list);
1199 parent->un.leaf.q = new_q ? new_q : &noop_qdisc;
1200 parent->tokens = parent->buffer;
1201 parent->ctokens = parent->cbuffer;
1202 parent->t_c = psched_get_time();
1203 parent->cmode = HTB_CAN_SEND;
1206 static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
1209 WARN_ON(!cl->un.leaf.q);
1210 qdisc_destroy(cl->un.leaf.q);
1212 gen_kill_estimator(&cl->bstats, &cl->rate_est);
1213 tcf_destroy_chain(&cl->filter_list);
1217 static void htb_destroy(struct Qdisc *sch)
1219 struct htb_sched *q = qdisc_priv(sch);
1220 struct hlist_node *next;
1221 struct htb_class *cl;
1224 cancel_work_sync(&q->work);
1225 qdisc_watchdog_cancel(&q->watchdog);
1226 /* This line used to be after htb_destroy_class call below
1227 * and surprisingly it worked in 2.4. But it must precede it
1228 * because filter need its target class alive to be able to call
1229 * unbind_filter on it (without Oops).
1231 tcf_destroy_chain(&q->filter_list);
1233 for (i = 0; i < q->clhash.hashsize; i++) {
1234 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode)
1235 tcf_destroy_chain(&cl->filter_list);
1237 for (i = 0; i < q->clhash.hashsize; i++) {
1238 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1240 htb_destroy_class(sch, cl);
1242 qdisc_class_hash_destroy(&q->clhash);
1243 __skb_queue_purge(&q->direct_queue);
1246 static int htb_delete(struct Qdisc *sch, unsigned long arg)
1248 struct htb_sched *q = qdisc_priv(sch);
1249 struct htb_class *cl = (struct htb_class *)arg;
1251 struct Qdisc *new_q = NULL;
1254 // TODO: why don't allow to delete subtree ? references ? does
1255 // tc subsys quarantee us that in htb_destroy it holds no class
1256 // refs so that we can remove children safely there ?
1257 if (cl->children || cl->filter_cnt)
1260 if (!cl->level && htb_parent_last_child(cl)) {
1261 new_q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1262 cl->parent->common.classid);
1269 qlen = cl->un.leaf.q->q.qlen;
1270 qdisc_reset(cl->un.leaf.q);
1271 qdisc_tree_decrease_qlen(cl->un.leaf.q, qlen);
1274 /* delete from hash and active; remainder in destroy_class */
1275 qdisc_class_hash_remove(&q->clhash, &cl->common);
1277 cl->parent->children--;
1279 if (cl->prio_activity)
1280 htb_deactivate(q, cl);
1282 if (cl->cmode != HTB_CAN_SEND)
1283 htb_safe_rb_erase(&cl->pq_node, q->wait_pq + cl->level);
1286 htb_parent_to_leaf(q, cl, new_q);
1288 BUG_ON(--cl->refcnt == 0);
1290 * This shouldn't happen: we "hold" one cops->get() when called
1291 * from tc_ctl_tclass; the destroy method is done from cops->put().
1294 sch_tree_unlock(sch);
1298 static void htb_put(struct Qdisc *sch, unsigned long arg)
1300 struct htb_class *cl = (struct htb_class *)arg;
1302 if (--cl->refcnt == 0)
1303 htb_destroy_class(sch, cl);
1306 static int htb_change_class(struct Qdisc *sch, u32 classid,
1307 u32 parentid, struct nlattr **tca,
1311 struct htb_sched *q = qdisc_priv(sch);
1312 struct htb_class *cl = (struct htb_class *)*arg, *parent;
1313 struct nlattr *opt = tca[TCA_OPTIONS];
1314 struct nlattr *tb[__TCA_HTB_MAX];
1315 struct tc_htb_opt *hopt;
1317 /* extract all subattrs from opt attr */
1321 err = nla_parse_nested(tb, TCA_HTB_MAX, opt, htb_policy);
1326 if (tb[TCA_HTB_PARMS] == NULL)
1329 parent = parentid == TC_H_ROOT ? NULL : htb_find(parentid, sch);
1331 hopt = nla_data(tb[TCA_HTB_PARMS]);
1332 if (!hopt->rate.rate || !hopt->ceil.rate)
1335 if (!cl) { /* new class */
1336 struct Qdisc *new_q;
1340 struct gnet_estimator opt;
1343 .nla_len = nla_attr_size(sizeof(est.opt)),
1344 .nla_type = TCA_RATE,
1347 /* 4s interval, 16s averaging constant */
1353 /* check for valid classid */
1354 if (!classid || TC_H_MAJ(classid ^ sch->handle) ||
1355 htb_find(classid, sch))
1358 /* check maximal depth */
1359 if (parent && parent->parent && parent->parent->level < 2) {
1360 pr_err("htb: tree is too deep\n");
1364 cl = kzalloc(sizeof(*cl), GFP_KERNEL);
1368 err = gen_new_estimator(&cl->bstats, &cl->rate_est,
1369 qdisc_root_sleeping_lock(sch),
1370 tca[TCA_RATE] ? : &est.nla);
1378 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
1379 RB_CLEAR_NODE(&cl->pq_node);
1381 for (prio = 0; prio < TC_HTB_NUMPRIO; prio++)
1382 RB_CLEAR_NODE(&cl->node[prio]);
1384 /* create leaf qdisc early because it uses kmalloc(GFP_KERNEL)
1385 * so that can't be used inside of sch_tree_lock
1386 * -- thanks to Karlis Peisenieks
1388 new_q = qdisc_create_dflt(sch->dev_queue,
1389 &pfifo_qdisc_ops, classid);
1391 if (parent && !parent->level) {
1392 unsigned int qlen = parent->un.leaf.q->q.qlen;
1394 /* turn parent into inner node */
1395 qdisc_reset(parent->un.leaf.q);
1396 qdisc_tree_decrease_qlen(parent->un.leaf.q, qlen);
1397 qdisc_destroy(parent->un.leaf.q);
1398 if (parent->prio_activity)
1399 htb_deactivate(q, parent);
1401 /* remove from evt list because of level change */
1402 if (parent->cmode != HTB_CAN_SEND) {
1403 htb_safe_rb_erase(&parent->pq_node, q->wait_pq);
1404 parent->cmode = HTB_CAN_SEND;
1406 parent->level = (parent->parent ? parent->parent->level
1407 : TC_HTB_MAXDEPTH) - 1;
1408 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1410 /* leaf (we) needs elementary qdisc */
1411 cl->un.leaf.q = new_q ? new_q : &noop_qdisc;
1413 cl->common.classid = classid;
1414 cl->parent = parent;
1416 /* set class to be in HTB_CAN_SEND state */
1417 cl->tokens = PSCHED_TICKS2NS(hopt->buffer);
1418 cl->ctokens = PSCHED_TICKS2NS(hopt->cbuffer);
1419 cl->mbuffer = 60 * PSCHED_TICKS_PER_SEC; /* 1min */
1420 cl->t_c = psched_get_time();
1421 cl->cmode = HTB_CAN_SEND;
1423 /* attach to the hash list and parent's family */
1424 qdisc_class_hash_insert(&q->clhash, &cl->common);
1428 if (tca[TCA_RATE]) {
1429 err = gen_replace_estimator(&cl->bstats, &cl->rate_est,
1430 qdisc_root_sleeping_lock(sch),
1438 /* it used to be a nasty bug here, we have to check that node
1439 * is really leaf before changing cl->un.leaf !
1442 cl->quantum = hopt->rate.rate / q->rate2quantum;
1443 if (!hopt->quantum && cl->quantum < 1000) {
1445 "HTB: quantum of class %X is small. Consider r2q change.\n",
1446 cl->common.classid);
1449 if (!hopt->quantum && cl->quantum > 200000) {
1451 "HTB: quantum of class %X is big. Consider r2q change.\n",
1452 cl->common.classid);
1453 cl->quantum = 200000;
1456 cl->quantum = hopt->quantum;
1457 if ((cl->prio = hopt->prio) >= TC_HTB_NUMPRIO)
1458 cl->prio = TC_HTB_NUMPRIO - 1;
1461 psched_ratecfg_precompute(&cl->rate, hopt->rate.rate);
1462 psched_ratecfg_precompute(&cl->ceil, hopt->ceil.rate);
1464 cl->buffer = PSCHED_TICKS2NS(hopt->buffer);
1465 cl->cbuffer = PSCHED_TICKS2NS(hopt->buffer);
1467 sch_tree_unlock(sch);
1469 qdisc_class_hash_grow(sch, &q->clhash);
1471 *arg = (unsigned long)cl;
1478 static struct tcf_proto **htb_find_tcf(struct Qdisc *sch, unsigned long arg)
1480 struct htb_sched *q = qdisc_priv(sch);
1481 struct htb_class *cl = (struct htb_class *)arg;
1482 struct tcf_proto **fl = cl ? &cl->filter_list : &q->filter_list;
1487 static unsigned long htb_bind_filter(struct Qdisc *sch, unsigned long parent,
1490 struct htb_class *cl = htb_find(classid, sch);
1492 /*if (cl && !cl->level) return 0;
1493 * The line above used to be there to prevent attaching filters to
1494 * leaves. But at least tc_index filter uses this just to get class
1495 * for other reasons so that we have to allow for it.
1497 * 19.6.2002 As Werner explained it is ok - bind filter is just
1498 * another way to "lock" the class - unlike "get" this lock can
1499 * be broken by class during destroy IIUC.
1503 return (unsigned long)cl;
1506 static void htb_unbind_filter(struct Qdisc *sch, unsigned long arg)
1508 struct htb_class *cl = (struct htb_class *)arg;
1514 static void htb_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1516 struct htb_sched *q = qdisc_priv(sch);
1517 struct htb_class *cl;
1523 for (i = 0; i < q->clhash.hashsize; i++) {
1524 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1525 if (arg->count < arg->skip) {
1529 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1538 static const struct Qdisc_class_ops htb_class_ops = {
1541 .qlen_notify = htb_qlen_notify,
1544 .change = htb_change_class,
1545 .delete = htb_delete,
1547 .tcf_chain = htb_find_tcf,
1548 .bind_tcf = htb_bind_filter,
1549 .unbind_tcf = htb_unbind_filter,
1550 .dump = htb_dump_class,
1551 .dump_stats = htb_dump_class_stats,
1554 static struct Qdisc_ops htb_qdisc_ops __read_mostly = {
1555 .cl_ops = &htb_class_ops,
1557 .priv_size = sizeof(struct htb_sched),
1558 .enqueue = htb_enqueue,
1559 .dequeue = htb_dequeue,
1560 .peek = qdisc_peek_dequeued,
1564 .destroy = htb_destroy,
1566 .owner = THIS_MODULE,
1569 static int __init htb_module_init(void)
1571 return register_qdisc(&htb_qdisc_ops);
1573 static void __exit htb_module_exit(void)
1575 unregister_qdisc(&htb_qdisc_ops);
1578 module_init(htb_module_init)
1579 module_exit(htb_module_exit)
1580 MODULE_LICENSE("GPL");