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.
11 * Credits (in time order) for older HTB versions:
13 * HTB support at LARTC mailing list
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
21 * created test case so that I was able to fix nasty bug
23 * spotted bug in dequeue code and helped with fix
25 * fixed requeue routine
26 * and many others. thanks.
28 #include <linux/module.h>
29 #include <linux/moduleparam.h>
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/string.h>
33 #include <linux/errno.h>
34 #include <linux/skbuff.h>
35 #include <linux/list.h>
36 #include <linux/compiler.h>
37 #include <linux/rbtree.h>
38 #include <linux/workqueue.h>
39 #include <linux/slab.h>
40 #include <net/netlink.h>
41 #include <net/sch_generic.h>
42 #include <net/pkt_sched.h>
43 #include <net/pkt_cls.h>
47 ========================================================================
48 HTB is like TBF with multiple classes. It is also similar to CBQ because
49 it allows to assign priority to each class in hierarchy.
50 In fact it is another implementation of Floyd's formal sharing.
53 Each class is assigned level. Leaf has ALWAYS level 0 and root
54 classes have level TC_HTB_MAXDEPTH-1. Interior nodes has level
55 one less than their parent.
58 static int htb_hysteresis __read_mostly = 0; /* whether to use mode hysteresis for speedup */
59 #define HTB_VER 0x30011 /* major must be matched with number suplied by TC as version */
61 #if HTB_VER >> 16 != TC_HTB_PROTOVER
62 #error "Mismatched sch_htb.c and pkt_sch.h"
65 /* Module parameter and sysfs export */
66 module_param (htb_hysteresis, int, 0640);
67 MODULE_PARM_DESC(htb_hysteresis, "Hysteresis mode, less CPU load, less accurate");
69 static int htb_rate_est = 0; /* htb classes have a default rate estimator */
70 module_param(htb_rate_est, int, 0640);
71 MODULE_PARM_DESC(htb_rate_est, "setup a default rate estimator (4sec 16sec) for htb classes");
73 /* used internaly to keep status of single class */
75 HTB_CANT_SEND, /* class can't send and can't borrow */
76 HTB_MAY_BORROW, /* class can't send but may borrow */
77 HTB_CAN_SEND /* class can send */
86 /* When class changes from state 1->2 and disconnects from
87 * parent's feed then we lost ptr value and start from the
88 * first child again. Here we store classid of the
89 * last valid ptr (used when ptr is NULL).
94 /* interior & leaf nodes; props specific to leaves are marked L:
95 * To reduce false sharing, place mostly read fields at beginning,
96 * and mostly written ones at the end.
99 struct Qdisc_class_common common;
100 struct psched_ratecfg rate;
101 struct psched_ratecfg ceil;
102 s64 buffer, cbuffer;/* token bucket depth/rate */
103 s64 mbuffer; /* max wait time */
104 u32 prio; /* these two are used only by leaves... */
105 int quantum; /* but stored for parent-to-leaf return */
107 struct tcf_proto __rcu *filter_list; /* class attached filters */
108 struct tcf_block *block;
111 int level; /* our level (see above) */
112 unsigned int children;
113 struct htb_class *parent; /* parent class */
115 struct net_rate_estimator __rcu *rate_est;
118 * Written often fields
120 struct gnet_stats_basic_packed bstats;
121 struct tc_htb_xstats xstats; /* our special stats */
123 /* token bucket parameters */
124 s64 tokens, ctokens;/* current number of tokens */
125 s64 t_c; /* checkpoint time */
128 struct htb_class_leaf {
129 int deficit[TC_HTB_MAXDEPTH];
132 struct htb_class_inner {
133 struct htb_prio clprio[TC_HTB_NUMPRIO];
138 int prio_activity; /* for which prios are we active */
139 enum htb_cmode cmode; /* current mode of the class */
140 struct rb_node pq_node; /* node for event queue */
141 struct rb_node node[TC_HTB_NUMPRIO]; /* node for self or feed tree */
143 unsigned int drops ____cacheline_aligned_in_smp;
144 unsigned int overlimits;
148 struct rb_root wait_pq;
149 struct htb_prio hprio[TC_HTB_NUMPRIO];
153 struct Qdisc_class_hash clhash;
154 int defcls; /* class where unclassified flows go to */
155 int rate2quantum; /* quant = rate / rate2quantum */
157 /* filters for qdisc itself */
158 struct tcf_proto __rcu *filter_list;
159 struct tcf_block *block;
161 #define HTB_WARN_TOOMANYEVENTS 0x1
162 unsigned int warned; /* only one warning */
164 struct work_struct work;
166 /* non shaped skbs; let them go directly thru */
167 struct qdisc_skb_head direct_queue;
170 struct qdisc_watchdog watchdog;
172 s64 now; /* cached dequeue time */
174 /* time of nearest event per level (row) */
175 s64 near_ev_cache[TC_HTB_MAXDEPTH];
177 int row_mask[TC_HTB_MAXDEPTH];
179 struct htb_level hlevel[TC_HTB_MAXDEPTH];
182 /* find class in global hash table using given handle */
183 static inline struct htb_class *htb_find(u32 handle, struct Qdisc *sch)
185 struct htb_sched *q = qdisc_priv(sch);
186 struct Qdisc_class_common *clc;
188 clc = qdisc_class_find(&q->clhash, handle);
191 return container_of(clc, struct htb_class, common);
194 static unsigned long htb_search(struct Qdisc *sch, u32 handle)
196 return (unsigned long)htb_find(handle, sch);
199 * htb_classify - classify a packet into class
201 * It returns NULL if the packet should be dropped or -1 if the packet
202 * should be passed directly thru. In all other cases leaf class is returned.
203 * We allow direct class selection by classid in priority. The we examine
204 * filters in qdisc and in inner nodes (if higher filter points to the inner
205 * node). If we end up with classid MAJOR:0 we enqueue the skb into special
206 * internal fifo (direct). These packets then go directly thru. If we still
207 * have no valid leaf we try to use MAJOR:default leaf. It still unsuccessful
208 * then finish and return direct queue.
210 #define HTB_DIRECT ((struct htb_class *)-1L)
212 static struct htb_class *htb_classify(struct sk_buff *skb, struct Qdisc *sch,
215 struct htb_sched *q = qdisc_priv(sch);
216 struct htb_class *cl;
217 struct tcf_result res;
218 struct tcf_proto *tcf;
221 /* allow to select class by setting skb->priority to valid classid;
222 * note that nfmark can be used too by attaching filter fw with no
225 if (skb->priority == sch->handle)
226 return HTB_DIRECT; /* X:0 (direct flow) selected */
227 cl = htb_find(skb->priority, sch);
231 /* Start with inner filter chain if a non-leaf class is selected */
232 tcf = rcu_dereference_bh(cl->filter_list);
234 tcf = rcu_dereference_bh(q->filter_list);
237 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
238 while (tcf && (result = tcf_classify(skb, tcf, &res, false)) >= 0) {
239 #ifdef CONFIG_NET_CLS_ACT
244 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
250 cl = (void *)res.class;
252 if (res.classid == sch->handle)
253 return HTB_DIRECT; /* X:0 (direct flow) */
254 cl = htb_find(res.classid, sch);
256 break; /* filter selected invalid classid */
259 return cl; /* we hit leaf; return it */
261 /* we have got inner class; apply inner filter chain */
262 tcf = rcu_dereference_bh(cl->filter_list);
264 /* classification failed; try to use default class */
265 cl = htb_find(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
266 if (!cl || cl->level)
267 return HTB_DIRECT; /* bad default .. this is safe bet */
272 * htb_add_to_id_tree - adds class to the round robin list
274 * Routine adds class to the list (actually tree) sorted by classid.
275 * Make sure that class is not already on such list for given prio.
277 static void htb_add_to_id_tree(struct rb_root *root,
278 struct htb_class *cl, int prio)
280 struct rb_node **p = &root->rb_node, *parent = NULL;
285 c = rb_entry(parent, struct htb_class, node[prio]);
287 if (cl->common.classid > c->common.classid)
288 p = &parent->rb_right;
290 p = &parent->rb_left;
292 rb_link_node(&cl->node[prio], parent, p);
293 rb_insert_color(&cl->node[prio], root);
297 * htb_add_to_wait_tree - adds class to the event queue with delay
299 * The class is added to priority event queue to indicate that class will
300 * change its mode in cl->pq_key microseconds. Make sure that class is not
301 * already in the queue.
303 static void htb_add_to_wait_tree(struct htb_sched *q,
304 struct htb_class *cl, s64 delay)
306 struct rb_node **p = &q->hlevel[cl->level].wait_pq.rb_node, *parent = NULL;
308 cl->pq_key = q->now + delay;
309 if (cl->pq_key == q->now)
312 /* update the nearest event cache */
313 if (q->near_ev_cache[cl->level] > cl->pq_key)
314 q->near_ev_cache[cl->level] = cl->pq_key;
319 c = rb_entry(parent, struct htb_class, pq_node);
320 if (cl->pq_key >= c->pq_key)
321 p = &parent->rb_right;
323 p = &parent->rb_left;
325 rb_link_node(&cl->pq_node, parent, p);
326 rb_insert_color(&cl->pq_node, &q->hlevel[cl->level].wait_pq);
330 * htb_next_rb_node - finds next node in binary tree
332 * When we are past last key we return NULL.
333 * Average complexity is 2 steps per call.
335 static inline void htb_next_rb_node(struct rb_node **n)
341 * htb_add_class_to_row - add class to its row
343 * The class is added to row at priorities marked in mask.
344 * It does nothing if mask == 0.
346 static inline void htb_add_class_to_row(struct htb_sched *q,
347 struct htb_class *cl, int mask)
349 q->row_mask[cl->level] |= mask;
351 int prio = ffz(~mask);
352 mask &= ~(1 << prio);
353 htb_add_to_id_tree(&q->hlevel[cl->level].hprio[prio].row, cl, prio);
357 /* If this triggers, it is a bug in this code, but it need not be fatal */
358 static void htb_safe_rb_erase(struct rb_node *rb, struct rb_root *root)
360 if (RB_EMPTY_NODE(rb)) {
370 * htb_remove_class_from_row - removes class from its row
372 * The class is removed from row at priorities marked in mask.
373 * It does nothing if mask == 0.
375 static inline void htb_remove_class_from_row(struct htb_sched *q,
376 struct htb_class *cl, int mask)
379 struct htb_level *hlevel = &q->hlevel[cl->level];
382 int prio = ffz(~mask);
383 struct htb_prio *hprio = &hlevel->hprio[prio];
385 mask &= ~(1 << prio);
386 if (hprio->ptr == cl->node + prio)
387 htb_next_rb_node(&hprio->ptr);
389 htb_safe_rb_erase(cl->node + prio, &hprio->row);
390 if (!hprio->row.rb_node)
393 q->row_mask[cl->level] &= ~m;
397 * htb_activate_prios - creates active classe's feed chain
399 * The class is connected to ancestors and/or appropriate rows
400 * for priorities it is participating on. cl->cmode must be new
401 * (activated) mode. It does nothing if cl->prio_activity == 0.
403 static void htb_activate_prios(struct htb_sched *q, struct htb_class *cl)
405 struct htb_class *p = cl->parent;
406 long m, mask = cl->prio_activity;
408 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
414 if (p->inner.clprio[prio].feed.rb_node)
415 /* parent already has its feed in use so that
416 * reset bit in mask as parent is already ok
418 mask &= ~(1 << prio);
420 htb_add_to_id_tree(&p->inner.clprio[prio].feed, cl, prio);
422 p->prio_activity |= mask;
427 if (cl->cmode == HTB_CAN_SEND && mask)
428 htb_add_class_to_row(q, cl, mask);
432 * htb_deactivate_prios - remove class from feed chain
434 * cl->cmode must represent old mode (before deactivation). It does
435 * nothing if cl->prio_activity == 0. Class is removed from all feed
438 static void htb_deactivate_prios(struct htb_sched *q, struct htb_class *cl)
440 struct htb_class *p = cl->parent;
441 long m, mask = cl->prio_activity;
443 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
450 if (p->inner.clprio[prio].ptr == cl->node + prio) {
451 /* we are removing child which is pointed to from
452 * parent feed - forget the pointer but remember
455 p->inner.clprio[prio].last_ptr_id = cl->common.classid;
456 p->inner.clprio[prio].ptr = NULL;
459 htb_safe_rb_erase(cl->node + prio,
460 &p->inner.clprio[prio].feed);
462 if (!p->inner.clprio[prio].feed.rb_node)
466 p->prio_activity &= ~mask;
471 if (cl->cmode == HTB_CAN_SEND && mask)
472 htb_remove_class_from_row(q, cl, mask);
475 static inline s64 htb_lowater(const struct htb_class *cl)
478 return cl->cmode != HTB_CANT_SEND ? -cl->cbuffer : 0;
482 static inline s64 htb_hiwater(const struct htb_class *cl)
485 return cl->cmode == HTB_CAN_SEND ? -cl->buffer : 0;
492 * htb_class_mode - computes and returns current class mode
494 * It computes cl's mode at time cl->t_c+diff and returns it. If mode
495 * is not HTB_CAN_SEND then cl->pq_key is updated to time difference
496 * from now to time when cl will change its state.
497 * Also it is worth to note that class mode doesn't change simply
498 * at cl->{c,}tokens == 0 but there can rather be hysteresis of
499 * 0 .. -cl->{c,}buffer range. It is meant to limit number of
500 * mode transitions per time unit. The speed gain is about 1/6.
502 static inline enum htb_cmode
503 htb_class_mode(struct htb_class *cl, s64 *diff)
507 if ((toks = (cl->ctokens + *diff)) < htb_lowater(cl)) {
509 return HTB_CANT_SEND;
512 if ((toks = (cl->tokens + *diff)) >= htb_hiwater(cl))
516 return HTB_MAY_BORROW;
520 * htb_change_class_mode - changes classe's mode
522 * This should be the only way how to change classe's mode under normal
523 * cirsumstances. Routine will update feed lists linkage, change mode
524 * and add class to the wait event queue if appropriate. New mode should
525 * be different from old one and cl->pq_key has to be valid if changing
526 * to mode other than HTB_CAN_SEND (see htb_add_to_wait_tree).
529 htb_change_class_mode(struct htb_sched *q, struct htb_class *cl, s64 *diff)
531 enum htb_cmode new_mode = htb_class_mode(cl, diff);
533 if (new_mode == cl->cmode)
536 if (new_mode == HTB_CANT_SEND)
539 if (cl->prio_activity) { /* not necessary: speed optimization */
540 if (cl->cmode != HTB_CANT_SEND)
541 htb_deactivate_prios(q, cl);
542 cl->cmode = new_mode;
543 if (new_mode != HTB_CANT_SEND)
544 htb_activate_prios(q, cl);
546 cl->cmode = new_mode;
550 * htb_activate - inserts leaf cl into appropriate active feeds
552 * Routine learns (new) priority of leaf and activates feed chain
553 * for the prio. It can be called on already active leaf safely.
554 * It also adds leaf into droplist.
556 static inline void htb_activate(struct htb_sched *q, struct htb_class *cl)
558 WARN_ON(cl->level || !cl->leaf.q || !cl->leaf.q->q.qlen);
560 if (!cl->prio_activity) {
561 cl->prio_activity = 1 << cl->prio;
562 htb_activate_prios(q, cl);
567 * htb_deactivate - remove leaf cl from active feeds
569 * Make sure that leaf is active. In the other words it can't be called
570 * with non-active leaf. It also removes class from the drop list.
572 static inline void htb_deactivate(struct htb_sched *q, struct htb_class *cl)
574 WARN_ON(!cl->prio_activity);
576 htb_deactivate_prios(q, cl);
577 cl->prio_activity = 0;
580 static int htb_enqueue(struct sk_buff *skb, struct Qdisc *sch,
581 struct sk_buff **to_free)
583 int uninitialized_var(ret);
584 unsigned int len = qdisc_pkt_len(skb);
585 struct htb_sched *q = qdisc_priv(sch);
586 struct htb_class *cl = htb_classify(skb, sch, &ret);
588 if (cl == HTB_DIRECT) {
589 /* enqueue to helper queue */
590 if (q->direct_queue.qlen < q->direct_qlen) {
591 __qdisc_enqueue_tail(skb, &q->direct_queue);
594 return qdisc_drop(skb, sch, to_free);
596 #ifdef CONFIG_NET_CLS_ACT
598 if (ret & __NET_XMIT_BYPASS)
599 qdisc_qstats_drop(sch);
600 __qdisc_drop(skb, to_free);
603 } else if ((ret = qdisc_enqueue(skb, cl->leaf.q,
604 to_free)) != NET_XMIT_SUCCESS) {
605 if (net_xmit_drop_count(ret)) {
606 qdisc_qstats_drop(sch);
614 sch->qstats.backlog += len;
616 return NET_XMIT_SUCCESS;
619 static inline void htb_accnt_tokens(struct htb_class *cl, int bytes, s64 diff)
621 s64 toks = diff + cl->tokens;
623 if (toks > cl->buffer)
625 toks -= (s64) psched_l2t_ns(&cl->rate, bytes);
626 if (toks <= -cl->mbuffer)
627 toks = 1 - cl->mbuffer;
632 static inline void htb_accnt_ctokens(struct htb_class *cl, int bytes, s64 diff)
634 s64 toks = diff + cl->ctokens;
636 if (toks > cl->cbuffer)
638 toks -= (s64) psched_l2t_ns(&cl->ceil, bytes);
639 if (toks <= -cl->mbuffer)
640 toks = 1 - cl->mbuffer;
646 * htb_charge_class - charges amount "bytes" to leaf and ancestors
648 * Routine assumes that packet "bytes" long was dequeued from leaf cl
649 * borrowing from "level". It accounts bytes to ceil leaky bucket for
650 * leaf and all ancestors and to rate bucket for ancestors at levels
651 * "level" and higher. It also handles possible change of mode resulting
652 * from the update. Note that mode can also increase here (MAY_BORROW to
653 * CAN_SEND) because we can use more precise clock that event queue here.
654 * In such case we remove class from event queue first.
656 static void htb_charge_class(struct htb_sched *q, struct htb_class *cl,
657 int level, struct sk_buff *skb)
659 int bytes = qdisc_pkt_len(skb);
660 enum htb_cmode old_mode;
664 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
665 if (cl->level >= level) {
666 if (cl->level == level)
668 htb_accnt_tokens(cl, bytes, diff);
670 cl->xstats.borrows++;
671 cl->tokens += diff; /* we moved t_c; update tokens */
673 htb_accnt_ctokens(cl, bytes, diff);
676 old_mode = cl->cmode;
678 htb_change_class_mode(q, cl, &diff);
679 if (old_mode != cl->cmode) {
680 if (old_mode != HTB_CAN_SEND)
681 htb_safe_rb_erase(&cl->pq_node, &q->hlevel[cl->level].wait_pq);
682 if (cl->cmode != HTB_CAN_SEND)
683 htb_add_to_wait_tree(q, cl, diff);
686 /* update basic stats except for leaves which are already updated */
688 bstats_update(&cl->bstats, skb);
695 * htb_do_events - make mode changes to classes at the level
697 * Scans event queue for pending events and applies them. Returns time of
698 * next pending event (0 for no event in pq, q->now for too many events).
699 * Note: Applied are events whose have cl->pq_key <= q->now.
701 static s64 htb_do_events(struct htb_sched *q, const int level,
704 /* don't run for longer than 2 jiffies; 2 is used instead of
705 * 1 to simplify things when jiffy is going to be incremented
708 unsigned long stop_at = start + 2;
709 struct rb_root *wait_pq = &q->hlevel[level].wait_pq;
711 while (time_before(jiffies, stop_at)) {
712 struct htb_class *cl;
714 struct rb_node *p = rb_first(wait_pq);
719 cl = rb_entry(p, struct htb_class, pq_node);
720 if (cl->pq_key > q->now)
723 htb_safe_rb_erase(p, wait_pq);
724 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
725 htb_change_class_mode(q, cl, &diff);
726 if (cl->cmode != HTB_CAN_SEND)
727 htb_add_to_wait_tree(q, cl, diff);
730 /* too much load - let's continue after a break for scheduling */
731 if (!(q->warned & HTB_WARN_TOOMANYEVENTS)) {
732 pr_warn("htb: too many events!\n");
733 q->warned |= HTB_WARN_TOOMANYEVENTS;
739 /* Returns class->node+prio from id-tree where classe's id is >= id. NULL
740 * is no such one exists.
742 static struct rb_node *htb_id_find_next_upper(int prio, struct rb_node *n,
745 struct rb_node *r = NULL;
747 struct htb_class *cl =
748 rb_entry(n, struct htb_class, node[prio]);
750 if (id > cl->common.classid) {
752 } else if (id < cl->common.classid) {
763 * htb_lookup_leaf - returns next leaf class in DRR order
765 * Find leaf where current feed pointers points to.
767 static struct htb_class *htb_lookup_leaf(struct htb_prio *hprio, const int prio)
771 struct rb_node *root;
772 struct rb_node **pptr;
774 } stk[TC_HTB_MAXDEPTH], *sp = stk;
776 BUG_ON(!hprio->row.rb_node);
777 sp->root = hprio->row.rb_node;
778 sp->pptr = &hprio->ptr;
779 sp->pid = &hprio->last_ptr_id;
781 for (i = 0; i < 65535; i++) {
782 if (!*sp->pptr && *sp->pid) {
783 /* ptr was invalidated but id is valid - try to recover
784 * the original or next ptr
787 htb_id_find_next_upper(prio, sp->root, *sp->pid);
789 *sp->pid = 0; /* ptr is valid now so that remove this hint as it
790 * can become out of date quickly
792 if (!*sp->pptr) { /* we are at right end; rewind & go up */
793 *sp->pptr = sp->root;
794 while ((*sp->pptr)->rb_left)
795 *sp->pptr = (*sp->pptr)->rb_left;
802 htb_next_rb_node(sp->pptr);
805 struct htb_class *cl;
806 struct htb_prio *clp;
808 cl = rb_entry(*sp->pptr, struct htb_class, node[prio]);
811 clp = &cl->inner.clprio[prio];
812 (++sp)->root = clp->feed.rb_node;
813 sp->pptr = &clp->ptr;
814 sp->pid = &clp->last_ptr_id;
821 /* dequeues packet at given priority and level; call only if
822 * you are sure that there is active class at prio/level
824 static struct sk_buff *htb_dequeue_tree(struct htb_sched *q, const int prio,
827 struct sk_buff *skb = NULL;
828 struct htb_class *cl, *start;
829 struct htb_level *hlevel = &q->hlevel[level];
830 struct htb_prio *hprio = &hlevel->hprio[prio];
832 /* look initial class up in the row */
833 start = cl = htb_lookup_leaf(hprio, prio);
840 /* class can be empty - it is unlikely but can be true if leaf
841 * qdisc drops packets in enqueue routine or if someone used
842 * graft operation on the leaf since last dequeue;
843 * simply deactivate and skip such class
845 if (unlikely(cl->leaf.q->q.qlen == 0)) {
846 struct htb_class *next;
847 htb_deactivate(q, cl);
849 /* row/level might become empty */
850 if ((q->row_mask[level] & (1 << prio)) == 0)
853 next = htb_lookup_leaf(hprio, prio);
855 if (cl == start) /* fix start if we just deleted it */
861 skb = cl->leaf.q->dequeue(cl->leaf.q);
862 if (likely(skb != NULL))
865 qdisc_warn_nonwc("htb", cl->leaf.q);
866 htb_next_rb_node(level ? &cl->parent->inner.clprio[prio].ptr:
867 &q->hlevel[0].hprio[prio].ptr);
868 cl = htb_lookup_leaf(hprio, prio);
870 } while (cl != start);
872 if (likely(skb != NULL)) {
873 bstats_update(&cl->bstats, skb);
874 cl->leaf.deficit[level] -= qdisc_pkt_len(skb);
875 if (cl->leaf.deficit[level] < 0) {
876 cl->leaf.deficit[level] += cl->quantum;
877 htb_next_rb_node(level ? &cl->parent->inner.clprio[prio].ptr :
878 &q->hlevel[0].hprio[prio].ptr);
880 /* this used to be after charge_class but this constelation
881 * gives us slightly better performance
883 if (!cl->leaf.q->q.qlen)
884 htb_deactivate(q, cl);
885 htb_charge_class(q, cl, level, skb);
890 static struct sk_buff *htb_dequeue(struct Qdisc *sch)
893 struct htb_sched *q = qdisc_priv(sch);
896 unsigned long start_at;
898 /* try to dequeue direct packets as high prio (!) to minimize cpu work */
899 skb = __qdisc_dequeue_head(&q->direct_queue);
902 qdisc_bstats_update(sch, skb);
903 qdisc_qstats_backlog_dec(sch, skb);
910 q->now = ktime_get_ns();
913 next_event = q->now + 5LLU * NSEC_PER_SEC;
915 for (level = 0; level < TC_HTB_MAXDEPTH; level++) {
916 /* common case optimization - skip event handler quickly */
918 s64 event = q->near_ev_cache[level];
920 if (q->now >= event) {
921 event = htb_do_events(q, level, start_at);
923 event = q->now + NSEC_PER_SEC;
924 q->near_ev_cache[level] = event;
927 if (next_event > event)
930 m = ~q->row_mask[level];
931 while (m != (int)(-1)) {
935 skb = htb_dequeue_tree(q, prio, level);
936 if (likely(skb != NULL))
940 qdisc_qstats_overlimit(sch);
941 if (likely(next_event > q->now))
942 qdisc_watchdog_schedule_ns(&q->watchdog, next_event);
944 schedule_work(&q->work);
949 /* reset all classes */
950 /* always caled under BH & queue lock */
951 static void htb_reset(struct Qdisc *sch)
953 struct htb_sched *q = qdisc_priv(sch);
954 struct htb_class *cl;
957 for (i = 0; i < q->clhash.hashsize; i++) {
958 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
960 memset(&cl->inner, 0, sizeof(cl->inner));
963 qdisc_reset(cl->leaf.q);
965 cl->prio_activity = 0;
966 cl->cmode = HTB_CAN_SEND;
969 qdisc_watchdog_cancel(&q->watchdog);
970 __qdisc_reset_queue(&q->direct_queue);
972 sch->qstats.backlog = 0;
973 memset(q->hlevel, 0, sizeof(q->hlevel));
974 memset(q->row_mask, 0, sizeof(q->row_mask));
977 static const struct nla_policy htb_policy[TCA_HTB_MAX + 1] = {
978 [TCA_HTB_PARMS] = { .len = sizeof(struct tc_htb_opt) },
979 [TCA_HTB_INIT] = { .len = sizeof(struct tc_htb_glob) },
980 [TCA_HTB_CTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
981 [TCA_HTB_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
982 [TCA_HTB_DIRECT_QLEN] = { .type = NLA_U32 },
983 [TCA_HTB_RATE64] = { .type = NLA_U64 },
984 [TCA_HTB_CEIL64] = { .type = NLA_U64 },
987 static void htb_work_func(struct work_struct *work)
989 struct htb_sched *q = container_of(work, struct htb_sched, work);
990 struct Qdisc *sch = q->watchdog.qdisc;
993 __netif_schedule(qdisc_root(sch));
997 static int htb_init(struct Qdisc *sch, struct nlattr *opt,
998 struct netlink_ext_ack *extack)
1000 struct htb_sched *q = qdisc_priv(sch);
1001 struct nlattr *tb[TCA_HTB_MAX + 1];
1002 struct tc_htb_glob *gopt;
1005 qdisc_watchdog_init(&q->watchdog, sch);
1006 INIT_WORK(&q->work, htb_work_func);
1011 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
1015 err = nla_parse_nested(tb, TCA_HTB_MAX, opt, htb_policy, NULL);
1019 if (!tb[TCA_HTB_INIT])
1022 gopt = nla_data(tb[TCA_HTB_INIT]);
1023 if (gopt->version != HTB_VER >> 16)
1026 err = qdisc_class_hash_init(&q->clhash);
1030 qdisc_skb_head_init(&q->direct_queue);
1032 if (tb[TCA_HTB_DIRECT_QLEN])
1033 q->direct_qlen = nla_get_u32(tb[TCA_HTB_DIRECT_QLEN]);
1035 q->direct_qlen = qdisc_dev(sch)->tx_queue_len;
1037 if ((q->rate2quantum = gopt->rate2quantum) < 1)
1038 q->rate2quantum = 1;
1039 q->defcls = gopt->defcls;
1044 static int htb_dump(struct Qdisc *sch, struct sk_buff *skb)
1046 struct htb_sched *q = qdisc_priv(sch);
1047 struct nlattr *nest;
1048 struct tc_htb_glob gopt;
1050 /* Its safe to not acquire qdisc lock. As we hold RTNL,
1051 * no change can happen on the qdisc parameters.
1054 gopt.direct_pkts = q->direct_pkts;
1055 gopt.version = HTB_VER;
1056 gopt.rate2quantum = q->rate2quantum;
1057 gopt.defcls = q->defcls;
1060 nest = nla_nest_start(skb, TCA_OPTIONS);
1062 goto nla_put_failure;
1063 if (nla_put(skb, TCA_HTB_INIT, sizeof(gopt), &gopt) ||
1064 nla_put_u32(skb, TCA_HTB_DIRECT_QLEN, q->direct_qlen))
1065 goto nla_put_failure;
1067 return nla_nest_end(skb, nest);
1070 nla_nest_cancel(skb, nest);
1074 static int htb_dump_class(struct Qdisc *sch, unsigned long arg,
1075 struct sk_buff *skb, struct tcmsg *tcm)
1077 struct htb_class *cl = (struct htb_class *)arg;
1078 struct nlattr *nest;
1079 struct tc_htb_opt opt;
1081 /* Its safe to not acquire qdisc lock. As we hold RTNL,
1082 * no change can happen on the class parameters.
1084 tcm->tcm_parent = cl->parent ? cl->parent->common.classid : TC_H_ROOT;
1085 tcm->tcm_handle = cl->common.classid;
1086 if (!cl->level && cl->leaf.q)
1087 tcm->tcm_info = cl->leaf.q->handle;
1089 nest = nla_nest_start(skb, TCA_OPTIONS);
1091 goto nla_put_failure;
1093 memset(&opt, 0, sizeof(opt));
1095 psched_ratecfg_getrate(&opt.rate, &cl->rate);
1096 opt.buffer = PSCHED_NS2TICKS(cl->buffer);
1097 psched_ratecfg_getrate(&opt.ceil, &cl->ceil);
1098 opt.cbuffer = PSCHED_NS2TICKS(cl->cbuffer);
1099 opt.quantum = cl->quantum;
1100 opt.prio = cl->prio;
1101 opt.level = cl->level;
1102 if (nla_put(skb, TCA_HTB_PARMS, sizeof(opt), &opt))
1103 goto nla_put_failure;
1104 if ((cl->rate.rate_bytes_ps >= (1ULL << 32)) &&
1105 nla_put_u64_64bit(skb, TCA_HTB_RATE64, cl->rate.rate_bytes_ps,
1107 goto nla_put_failure;
1108 if ((cl->ceil.rate_bytes_ps >= (1ULL << 32)) &&
1109 nla_put_u64_64bit(skb, TCA_HTB_CEIL64, cl->ceil.rate_bytes_ps,
1111 goto nla_put_failure;
1113 return nla_nest_end(skb, nest);
1116 nla_nest_cancel(skb, nest);
1121 htb_dump_class_stats(struct Qdisc *sch, unsigned long arg, struct gnet_dump *d)
1123 struct htb_class *cl = (struct htb_class *)arg;
1124 struct gnet_stats_queue qs = {
1126 .overlimits = cl->overlimits,
1130 if (!cl->level && cl->leaf.q)
1131 qdisc_qstats_qlen_backlog(cl->leaf.q, &qlen, &qs.backlog);
1133 cl->xstats.tokens = clamp_t(s64, PSCHED_NS2TICKS(cl->tokens),
1135 cl->xstats.ctokens = clamp_t(s64, PSCHED_NS2TICKS(cl->ctokens),
1138 if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
1139 d, NULL, &cl->bstats) < 0 ||
1140 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1141 gnet_stats_copy_queue(d, NULL, &qs, qlen) < 0)
1144 return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1147 static int htb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1148 struct Qdisc **old, struct netlink_ext_ack *extack)
1150 struct htb_class *cl = (struct htb_class *)arg;
1155 (new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1156 cl->common.classid, extack)) == NULL)
1159 *old = qdisc_replace(sch, new, &cl->leaf.q);
1163 static struct Qdisc *htb_leaf(struct Qdisc *sch, unsigned long arg)
1165 struct htb_class *cl = (struct htb_class *)arg;
1166 return !cl->level ? cl->leaf.q : NULL;
1169 static void htb_qlen_notify(struct Qdisc *sch, unsigned long arg)
1171 struct htb_class *cl = (struct htb_class *)arg;
1173 htb_deactivate(qdisc_priv(sch), cl);
1176 static inline int htb_parent_last_child(struct htb_class *cl)
1179 /* the root class */
1181 if (cl->parent->children > 1)
1182 /* not the last child */
1187 static void htb_parent_to_leaf(struct htb_sched *q, struct htb_class *cl,
1188 struct Qdisc *new_q)
1190 struct htb_class *parent = cl->parent;
1192 WARN_ON(cl->level || !cl->leaf.q || cl->prio_activity);
1194 if (parent->cmode != HTB_CAN_SEND)
1195 htb_safe_rb_erase(&parent->pq_node,
1196 &q->hlevel[parent->level].wait_pq);
1199 memset(&parent->inner, 0, sizeof(parent->inner));
1200 parent->leaf.q = new_q ? new_q : &noop_qdisc;
1201 parent->tokens = parent->buffer;
1202 parent->ctokens = parent->cbuffer;
1203 parent->t_c = ktime_get_ns();
1204 parent->cmode = HTB_CAN_SEND;
1207 static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
1210 WARN_ON(!cl->leaf.q);
1211 qdisc_put(cl->leaf.q);
1213 gen_kill_estimator(&cl->rate_est);
1214 tcf_block_put(cl->block);
1218 static void htb_destroy(struct Qdisc *sch)
1220 struct htb_sched *q = qdisc_priv(sch);
1221 struct hlist_node *next;
1222 struct htb_class *cl;
1225 cancel_work_sync(&q->work);
1226 qdisc_watchdog_cancel(&q->watchdog);
1227 /* This line used to be after htb_destroy_class call below
1228 * and surprisingly it worked in 2.4. But it must precede it
1229 * because filter need its target class alive to be able to call
1230 * unbind_filter on it (without Oops).
1232 tcf_block_put(q->block);
1234 for (i = 0; i < q->clhash.hashsize; i++) {
1235 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1236 tcf_block_put(cl->block);
1240 for (i = 0; i < q->clhash.hashsize; i++) {
1241 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1243 htb_destroy_class(sch, cl);
1245 qdisc_class_hash_destroy(&q->clhash);
1246 __qdisc_reset_queue(&q->direct_queue);
1249 static int htb_delete(struct Qdisc *sch, unsigned long arg)
1251 struct htb_sched *q = qdisc_priv(sch);
1252 struct htb_class *cl = (struct htb_class *)arg;
1253 struct Qdisc *new_q = NULL;
1256 /* TODO: why don't allow to delete subtree ? references ? does
1257 * tc subsys guarantee us that in htb_destroy it holds no class
1258 * refs so that we can remove children safely there ?
1260 if (cl->children || cl->filter_cnt)
1263 if (!cl->level && htb_parent_last_child(cl)) {
1264 new_q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1265 cl->parent->common.classid,
1273 qdisc_purge_queue(cl->leaf.q);
1275 /* delete from hash and active; remainder in destroy_class */
1276 qdisc_class_hash_remove(&q->clhash, &cl->common);
1278 cl->parent->children--;
1280 if (cl->prio_activity)
1281 htb_deactivate(q, cl);
1283 if (cl->cmode != HTB_CAN_SEND)
1284 htb_safe_rb_erase(&cl->pq_node,
1285 &q->hlevel[cl->level].wait_pq);
1288 htb_parent_to_leaf(q, cl, new_q);
1290 sch_tree_unlock(sch);
1292 htb_destroy_class(sch, cl);
1296 static int htb_change_class(struct Qdisc *sch, u32 classid,
1297 u32 parentid, struct nlattr **tca,
1298 unsigned long *arg, struct netlink_ext_ack *extack)
1301 struct htb_sched *q = qdisc_priv(sch);
1302 struct htb_class *cl = (struct htb_class *)*arg, *parent;
1303 struct nlattr *opt = tca[TCA_OPTIONS];
1304 struct nlattr *tb[TCA_HTB_MAX + 1];
1305 struct tc_htb_opt *hopt;
1309 /* extract all subattrs from opt attr */
1313 err = nla_parse_nested(tb, TCA_HTB_MAX, opt, htb_policy, NULL);
1318 if (tb[TCA_HTB_PARMS] == NULL)
1321 parent = parentid == TC_H_ROOT ? NULL : htb_find(parentid, sch);
1323 hopt = nla_data(tb[TCA_HTB_PARMS]);
1324 if (!hopt->rate.rate || !hopt->ceil.rate)
1327 /* Keeping backward compatible with rate_table based iproute2 tc */
1328 if (hopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
1329 qdisc_put_rtab(qdisc_get_rtab(&hopt->rate, tb[TCA_HTB_RTAB],
1332 if (hopt->ceil.linklayer == TC_LINKLAYER_UNAWARE)
1333 qdisc_put_rtab(qdisc_get_rtab(&hopt->ceil, tb[TCA_HTB_CTAB],
1336 if (!cl) { /* new class */
1337 struct Qdisc *new_q;
1341 struct gnet_estimator opt;
1344 .nla_len = nla_attr_size(sizeof(est.opt)),
1345 .nla_type = TCA_RATE,
1348 /* 4s interval, 16s averaging constant */
1354 /* check for valid classid */
1355 if (!classid || TC_H_MAJ(classid ^ sch->handle) ||
1356 htb_find(classid, sch))
1359 /* check maximal depth */
1360 if (parent && parent->parent && parent->parent->level < 2) {
1361 pr_err("htb: tree is too deep\n");
1365 cl = kzalloc(sizeof(*cl), GFP_KERNEL);
1369 err = tcf_block_get(&cl->block, &cl->filter_list, sch, extack);
1374 if (htb_rate_est || tca[TCA_RATE]) {
1375 err = gen_new_estimator(&cl->bstats, NULL,
1378 qdisc_root_sleeping_running(sch),
1379 tca[TCA_RATE] ? : &est.nla);
1381 tcf_block_put(cl->block);
1388 RB_CLEAR_NODE(&cl->pq_node);
1390 for (prio = 0; prio < TC_HTB_NUMPRIO; prio++)
1391 RB_CLEAR_NODE(&cl->node[prio]);
1393 /* create leaf qdisc early because it uses kmalloc(GFP_KERNEL)
1394 * so that can't be used inside of sch_tree_lock
1395 * -- thanks to Karlis Peisenieks
1397 new_q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1400 if (parent && !parent->level) {
1401 /* turn parent into inner node */
1402 qdisc_purge_queue(parent->leaf.q);
1403 qdisc_put(parent->leaf.q);
1404 if (parent->prio_activity)
1405 htb_deactivate(q, parent);
1407 /* remove from evt list because of level change */
1408 if (parent->cmode != HTB_CAN_SEND) {
1409 htb_safe_rb_erase(&parent->pq_node, &q->hlevel[0].wait_pq);
1410 parent->cmode = HTB_CAN_SEND;
1412 parent->level = (parent->parent ? parent->parent->level
1413 : TC_HTB_MAXDEPTH) - 1;
1414 memset(&parent->inner, 0, sizeof(parent->inner));
1416 /* leaf (we) needs elementary qdisc */
1417 cl->leaf.q = new_q ? new_q : &noop_qdisc;
1419 cl->common.classid = classid;
1420 cl->parent = parent;
1422 /* set class to be in HTB_CAN_SEND state */
1423 cl->tokens = PSCHED_TICKS2NS(hopt->buffer);
1424 cl->ctokens = PSCHED_TICKS2NS(hopt->cbuffer);
1425 cl->mbuffer = 60ULL * NSEC_PER_SEC; /* 1min */
1426 cl->t_c = ktime_get_ns();
1427 cl->cmode = HTB_CAN_SEND;
1429 /* attach to the hash list and parent's family */
1430 qdisc_class_hash_insert(&q->clhash, &cl->common);
1433 if (cl->leaf.q != &noop_qdisc)
1434 qdisc_hash_add(cl->leaf.q, true);
1436 if (tca[TCA_RATE]) {
1437 err = gen_replace_estimator(&cl->bstats, NULL,
1440 qdisc_root_sleeping_running(sch),
1448 rate64 = tb[TCA_HTB_RATE64] ? nla_get_u64(tb[TCA_HTB_RATE64]) : 0;
1450 ceil64 = tb[TCA_HTB_CEIL64] ? nla_get_u64(tb[TCA_HTB_CEIL64]) : 0;
1452 psched_ratecfg_precompute(&cl->rate, &hopt->rate, rate64);
1453 psched_ratecfg_precompute(&cl->ceil, &hopt->ceil, ceil64);
1455 /* it used to be a nasty bug here, we have to check that node
1456 * is really leaf before changing cl->leaf !
1459 u64 quantum = cl->rate.rate_bytes_ps;
1461 do_div(quantum, q->rate2quantum);
1462 cl->quantum = min_t(u64, quantum, INT_MAX);
1464 if (!hopt->quantum && cl->quantum < 1000) {
1468 if (!hopt->quantum && cl->quantum > 200000) {
1470 cl->quantum = 200000;
1473 cl->quantum = hopt->quantum;
1474 if ((cl->prio = hopt->prio) >= TC_HTB_NUMPRIO)
1475 cl->prio = TC_HTB_NUMPRIO - 1;
1478 cl->buffer = PSCHED_TICKS2NS(hopt->buffer);
1479 cl->cbuffer = PSCHED_TICKS2NS(hopt->cbuffer);
1481 sch_tree_unlock(sch);
1484 pr_warn("HTB: quantum of class %X is %s. Consider r2q change.\n",
1485 cl->common.classid, (warn == -1 ? "small" : "big"));
1487 qdisc_class_hash_grow(sch, &q->clhash);
1489 *arg = (unsigned long)cl;
1496 static struct tcf_block *htb_tcf_block(struct Qdisc *sch, unsigned long arg,
1497 struct netlink_ext_ack *extack)
1499 struct htb_sched *q = qdisc_priv(sch);
1500 struct htb_class *cl = (struct htb_class *)arg;
1502 return cl ? cl->block : q->block;
1505 static unsigned long htb_bind_filter(struct Qdisc *sch, unsigned long parent,
1508 struct htb_class *cl = htb_find(classid, sch);
1510 /*if (cl && !cl->level) return 0;
1511 * The line above used to be there to prevent attaching filters to
1512 * leaves. But at least tc_index filter uses this just to get class
1513 * for other reasons so that we have to allow for it.
1515 * 19.6.2002 As Werner explained it is ok - bind filter is just
1516 * another way to "lock" the class - unlike "get" this lock can
1517 * be broken by class during destroy IIUC.
1521 return (unsigned long)cl;
1524 static void htb_unbind_filter(struct Qdisc *sch, unsigned long arg)
1526 struct htb_class *cl = (struct htb_class *)arg;
1532 static void htb_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1534 struct htb_sched *q = qdisc_priv(sch);
1535 struct htb_class *cl;
1541 for (i = 0; i < q->clhash.hashsize; i++) {
1542 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1543 if (arg->count < arg->skip) {
1547 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1556 static const struct Qdisc_class_ops htb_class_ops = {
1559 .qlen_notify = htb_qlen_notify,
1561 .change = htb_change_class,
1562 .delete = htb_delete,
1564 .tcf_block = htb_tcf_block,
1565 .bind_tcf = htb_bind_filter,
1566 .unbind_tcf = htb_unbind_filter,
1567 .dump = htb_dump_class,
1568 .dump_stats = htb_dump_class_stats,
1571 static struct Qdisc_ops htb_qdisc_ops __read_mostly = {
1572 .cl_ops = &htb_class_ops,
1574 .priv_size = sizeof(struct htb_sched),
1575 .enqueue = htb_enqueue,
1576 .dequeue = htb_dequeue,
1577 .peek = qdisc_peek_dequeued,
1580 .destroy = htb_destroy,
1582 .owner = THIS_MODULE,
1585 static int __init htb_module_init(void)
1587 return register_qdisc(&htb_qdisc_ops);
1589 static void __exit htb_module_exit(void)
1591 unregister_qdisc(&htb_qdisc_ops);
1594 module_init(htb_module_init)
1595 module_exit(htb_module_exit)
1596 MODULE_LICENSE("GPL");