1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * net/sched/sch_htb.c Hierarchical token bucket, feed tree version
5 * Authors: Martin Devera, <devik@cdi.cz>
7 * Credits (in time order) for older HTB versions:
8 * Stef Coene <stef.coene@docum.org>
9 * HTB support at LARTC mailing list
10 * Ondrej Kraus, <krauso@barr.cz>
11 * found missing INIT_QDISC(htb)
12 * Vladimir Smelhaus, Aamer Akhter, Bert Hubert
13 * helped a lot to locate nasty class stall bug
14 * Andi Kleen, Jamal Hadi, Bert Hubert
15 * code review and helpful comments on shaping
16 * Tomasz Wrona, <tw@eter.tym.pl>
17 * created test case so that I was able to fix nasty bug
19 * spotted bug in dequeue code and helped with fix
21 * fixed requeue routine
22 * and many others. thanks.
24 #include <linux/module.h>
25 #include <linux/moduleparam.h>
26 #include <linux/types.h>
27 #include <linux/kernel.h>
28 #include <linux/string.h>
29 #include <linux/errno.h>
30 #include <linux/skbuff.h>
31 #include <linux/list.h>
32 #include <linux/compiler.h>
33 #include <linux/rbtree.h>
34 #include <linux/workqueue.h>
35 #include <linux/slab.h>
36 #include <net/netlink.h>
37 #include <net/sch_generic.h>
38 #include <net/pkt_sched.h>
39 #include <net/pkt_cls.h>
43 ========================================================================
44 HTB is like TBF with multiple classes. It is also similar to CBQ because
45 it allows to assign priority to each class in hierarchy.
46 In fact it is another implementation of Floyd's formal sharing.
49 Each class is assigned level. Leaf has ALWAYS level 0 and root
50 classes have level TC_HTB_MAXDEPTH-1. Interior nodes has level
51 one less than their parent.
54 static int htb_hysteresis __read_mostly = 0; /* whether to use mode hysteresis for speedup */
55 #define HTB_VER 0x30011 /* major must be matched with number suplied by TC as version */
57 #if HTB_VER >> 16 != TC_HTB_PROTOVER
58 #error "Mismatched sch_htb.c and pkt_sch.h"
61 /* Module parameter and sysfs export */
62 module_param (htb_hysteresis, int, 0640);
63 MODULE_PARM_DESC(htb_hysteresis, "Hysteresis mode, less CPU load, less accurate");
65 static int htb_rate_est = 0; /* htb classes have a default rate estimator */
66 module_param(htb_rate_est, int, 0640);
67 MODULE_PARM_DESC(htb_rate_est, "setup a default rate estimator (4sec 16sec) for htb classes");
69 /* used internaly to keep status of single class */
71 HTB_CANT_SEND, /* class can't send and can't borrow */
72 HTB_MAY_BORROW, /* class can't send but may borrow */
73 HTB_CAN_SEND /* class can send */
82 /* When class changes from state 1->2 and disconnects from
83 * parent's feed then we lost ptr value and start from the
84 * first child again. Here we store classid of the
85 * last valid ptr (used when ptr is NULL).
90 /* interior & leaf nodes; props specific to leaves are marked L:
91 * To reduce false sharing, place mostly read fields at beginning,
92 * and mostly written ones at the end.
95 struct Qdisc_class_common common;
96 struct psched_ratecfg rate;
97 struct psched_ratecfg ceil;
98 s64 buffer, cbuffer;/* token bucket depth/rate */
99 s64 mbuffer; /* max wait time */
100 u32 prio; /* these two are used only by leaves... */
101 int quantum; /* but stored for parent-to-leaf return */
103 struct tcf_proto __rcu *filter_list; /* class attached filters */
104 struct tcf_block *block;
107 int level; /* our level (see above) */
108 unsigned int children;
109 struct htb_class *parent; /* parent class */
111 struct net_rate_estimator __rcu *rate_est;
114 * Written often fields
116 struct gnet_stats_basic_packed bstats;
117 struct tc_htb_xstats xstats; /* our special stats */
119 /* token bucket parameters */
120 s64 tokens, ctokens;/* current number of tokens */
121 s64 t_c; /* checkpoint time */
124 struct htb_class_leaf {
125 int deficit[TC_HTB_MAXDEPTH];
128 struct htb_class_inner {
129 struct htb_prio clprio[TC_HTB_NUMPRIO];
134 int prio_activity; /* for which prios are we active */
135 enum htb_cmode cmode; /* current mode of the class */
136 struct rb_node pq_node; /* node for event queue */
137 struct rb_node node[TC_HTB_NUMPRIO]; /* node for self or feed tree */
139 unsigned int drops ____cacheline_aligned_in_smp;
140 unsigned int overlimits;
144 struct rb_root wait_pq;
145 struct htb_prio hprio[TC_HTB_NUMPRIO];
149 struct Qdisc_class_hash clhash;
150 int defcls; /* class where unclassified flows go to */
151 int rate2quantum; /* quant = rate / rate2quantum */
153 /* filters for qdisc itself */
154 struct tcf_proto __rcu *filter_list;
155 struct tcf_block *block;
157 #define HTB_WARN_TOOMANYEVENTS 0x1
158 unsigned int warned; /* only one warning */
160 struct work_struct work;
162 /* non shaped skbs; let them go directly thru */
163 struct qdisc_skb_head direct_queue;
167 struct qdisc_watchdog watchdog;
169 s64 now; /* cached dequeue time */
171 /* time of nearest event per level (row) */
172 s64 near_ev_cache[TC_HTB_MAXDEPTH];
174 int row_mask[TC_HTB_MAXDEPTH];
176 struct htb_level hlevel[TC_HTB_MAXDEPTH];
179 /* find class in global hash table using given handle */
180 static inline struct htb_class *htb_find(u32 handle, struct Qdisc *sch)
182 struct htb_sched *q = qdisc_priv(sch);
183 struct Qdisc_class_common *clc;
185 clc = qdisc_class_find(&q->clhash, handle);
188 return container_of(clc, struct htb_class, common);
191 static unsigned long htb_search(struct Qdisc *sch, u32 handle)
193 return (unsigned long)htb_find(handle, sch);
196 * htb_classify - classify a packet into class
198 * It returns NULL if the packet should be dropped or -1 if the packet
199 * should be passed directly thru. In all other cases leaf class is returned.
200 * We allow direct class selection by classid in priority. The we examine
201 * filters in qdisc and in inner nodes (if higher filter points to the inner
202 * node). If we end up with classid MAJOR:0 we enqueue the skb into special
203 * internal fifo (direct). These packets then go directly thru. If we still
204 * have no valid leaf we try to use MAJOR:default leaf. It still unsuccessful
205 * then finish and return direct queue.
207 #define HTB_DIRECT ((struct htb_class *)-1L)
209 static struct htb_class *htb_classify(struct sk_buff *skb, struct Qdisc *sch,
212 struct htb_sched *q = qdisc_priv(sch);
213 struct htb_class *cl;
214 struct tcf_result res;
215 struct tcf_proto *tcf;
218 /* allow to select class by setting skb->priority to valid classid;
219 * note that nfmark can be used too by attaching filter fw with no
222 if (skb->priority == sch->handle)
223 return HTB_DIRECT; /* X:0 (direct flow) selected */
224 cl = htb_find(skb->priority, sch);
228 /* Start with inner filter chain if a non-leaf class is selected */
229 tcf = rcu_dereference_bh(cl->filter_list);
231 tcf = rcu_dereference_bh(q->filter_list);
234 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
235 while (tcf && (result = tcf_classify(skb, tcf, &res, false)) >= 0) {
236 #ifdef CONFIG_NET_CLS_ACT
241 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
247 cl = (void *)res.class;
249 if (res.classid == sch->handle)
250 return HTB_DIRECT; /* X:0 (direct flow) */
251 cl = htb_find(res.classid, sch);
253 break; /* filter selected invalid classid */
256 return cl; /* we hit leaf; return it */
258 /* we have got inner class; apply inner filter chain */
259 tcf = rcu_dereference_bh(cl->filter_list);
261 /* classification failed; try to use default class */
262 cl = htb_find(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
263 if (!cl || cl->level)
264 return HTB_DIRECT; /* bad default .. this is safe bet */
269 * htb_add_to_id_tree - adds class to the round robin list
271 * Routine adds class to the list (actually tree) sorted by classid.
272 * Make sure that class is not already on such list for given prio.
274 static void htb_add_to_id_tree(struct rb_root *root,
275 struct htb_class *cl, int prio)
277 struct rb_node **p = &root->rb_node, *parent = NULL;
282 c = rb_entry(parent, struct htb_class, node[prio]);
284 if (cl->common.classid > c->common.classid)
285 p = &parent->rb_right;
287 p = &parent->rb_left;
289 rb_link_node(&cl->node[prio], parent, p);
290 rb_insert_color(&cl->node[prio], root);
294 * htb_add_to_wait_tree - adds class to the event queue with delay
296 * The class is added to priority event queue to indicate that class will
297 * change its mode in cl->pq_key microseconds. Make sure that class is not
298 * already in the queue.
300 static void htb_add_to_wait_tree(struct htb_sched *q,
301 struct htb_class *cl, s64 delay)
303 struct rb_node **p = &q->hlevel[cl->level].wait_pq.rb_node, *parent = NULL;
305 cl->pq_key = q->now + delay;
306 if (cl->pq_key == q->now)
309 /* update the nearest event cache */
310 if (q->near_ev_cache[cl->level] > cl->pq_key)
311 q->near_ev_cache[cl->level] = cl->pq_key;
316 c = rb_entry(parent, struct htb_class, pq_node);
317 if (cl->pq_key >= c->pq_key)
318 p = &parent->rb_right;
320 p = &parent->rb_left;
322 rb_link_node(&cl->pq_node, parent, p);
323 rb_insert_color(&cl->pq_node, &q->hlevel[cl->level].wait_pq);
327 * htb_next_rb_node - finds next node in binary tree
329 * When we are past last key we return NULL.
330 * Average complexity is 2 steps per call.
332 static inline void htb_next_rb_node(struct rb_node **n)
338 * htb_add_class_to_row - add class to its row
340 * The class is added to row at priorities marked in mask.
341 * It does nothing if mask == 0.
343 static inline void htb_add_class_to_row(struct htb_sched *q,
344 struct htb_class *cl, int mask)
346 q->row_mask[cl->level] |= mask;
348 int prio = ffz(~mask);
349 mask &= ~(1 << prio);
350 htb_add_to_id_tree(&q->hlevel[cl->level].hprio[prio].row, cl, prio);
354 /* If this triggers, it is a bug in this code, but it need not be fatal */
355 static void htb_safe_rb_erase(struct rb_node *rb, struct rb_root *root)
357 if (RB_EMPTY_NODE(rb)) {
367 * htb_remove_class_from_row - removes class from its row
369 * The class is removed from row at priorities marked in mask.
370 * It does nothing if mask == 0.
372 static inline void htb_remove_class_from_row(struct htb_sched *q,
373 struct htb_class *cl, int mask)
376 struct htb_level *hlevel = &q->hlevel[cl->level];
379 int prio = ffz(~mask);
380 struct htb_prio *hprio = &hlevel->hprio[prio];
382 mask &= ~(1 << prio);
383 if (hprio->ptr == cl->node + prio)
384 htb_next_rb_node(&hprio->ptr);
386 htb_safe_rb_erase(cl->node + prio, &hprio->row);
387 if (!hprio->row.rb_node)
390 q->row_mask[cl->level] &= ~m;
394 * htb_activate_prios - creates active classe's feed chain
396 * The class is connected to ancestors and/or appropriate rows
397 * for priorities it is participating on. cl->cmode must be new
398 * (activated) mode. It does nothing if cl->prio_activity == 0.
400 static void htb_activate_prios(struct htb_sched *q, struct htb_class *cl)
402 struct htb_class *p = cl->parent;
403 long m, mask = cl->prio_activity;
405 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
408 unsigned int prio = ffz(~m);
410 if (WARN_ON_ONCE(prio >= ARRAY_SIZE(p->inner.clprio)))
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) {
541 if (cl->prio_activity) { /* not necessary: speed optimization */
542 if (cl->cmode != HTB_CANT_SEND)
543 htb_deactivate_prios(q, cl);
544 cl->cmode = new_mode;
545 if (new_mode != HTB_CANT_SEND)
546 htb_activate_prios(q, cl);
548 cl->cmode = new_mode;
552 * htb_activate - inserts leaf cl into appropriate active feeds
554 * Routine learns (new) priority of leaf and activates feed chain
555 * for the prio. It can be called on already active leaf safely.
556 * It also adds leaf into droplist.
558 static inline void htb_activate(struct htb_sched *q, struct htb_class *cl)
560 WARN_ON(cl->level || !cl->leaf.q || !cl->leaf.q->q.qlen);
562 if (!cl->prio_activity) {
563 cl->prio_activity = 1 << cl->prio;
564 htb_activate_prios(q, cl);
569 * htb_deactivate - remove leaf cl from active feeds
571 * Make sure that leaf is active. In the other words it can't be called
572 * with non-active leaf. It also removes class from the drop list.
574 static inline void htb_deactivate(struct htb_sched *q, struct htb_class *cl)
576 WARN_ON(!cl->prio_activity);
578 htb_deactivate_prios(q, cl);
579 cl->prio_activity = 0;
582 static int htb_enqueue(struct sk_buff *skb, struct Qdisc *sch,
583 struct sk_buff **to_free)
586 unsigned int len = qdisc_pkt_len(skb);
587 struct htb_sched *q = qdisc_priv(sch);
588 struct htb_class *cl = htb_classify(skb, sch, &ret);
590 if (cl == HTB_DIRECT) {
591 /* enqueue to helper queue */
592 if (q->direct_queue.qlen < q->direct_qlen) {
593 __qdisc_enqueue_tail(skb, &q->direct_queue);
596 return qdisc_drop(skb, sch, to_free);
598 #ifdef CONFIG_NET_CLS_ACT
600 if (ret & __NET_XMIT_BYPASS)
601 qdisc_qstats_drop(sch);
602 __qdisc_drop(skb, to_free);
605 } else if ((ret = qdisc_enqueue(skb, cl->leaf.q,
606 to_free)) != NET_XMIT_SUCCESS) {
607 if (net_xmit_drop_count(ret)) {
608 qdisc_qstats_drop(sch);
616 sch->qstats.backlog += len;
618 return NET_XMIT_SUCCESS;
621 static inline void htb_accnt_tokens(struct htb_class *cl, int bytes, s64 diff)
623 s64 toks = diff + cl->tokens;
625 if (toks > cl->buffer)
627 toks -= (s64) psched_l2t_ns(&cl->rate, bytes);
628 if (toks <= -cl->mbuffer)
629 toks = 1 - cl->mbuffer;
634 static inline void htb_accnt_ctokens(struct htb_class *cl, int bytes, s64 diff)
636 s64 toks = diff + cl->ctokens;
638 if (toks > cl->cbuffer)
640 toks -= (s64) psched_l2t_ns(&cl->ceil, bytes);
641 if (toks <= -cl->mbuffer)
642 toks = 1 - cl->mbuffer;
648 * htb_charge_class - charges amount "bytes" to leaf and ancestors
650 * Routine assumes that packet "bytes" long was dequeued from leaf cl
651 * borrowing from "level". It accounts bytes to ceil leaky bucket for
652 * leaf and all ancestors and to rate bucket for ancestors at levels
653 * "level" and higher. It also handles possible change of mode resulting
654 * from the update. Note that mode can also increase here (MAY_BORROW to
655 * CAN_SEND) because we can use more precise clock that event queue here.
656 * In such case we remove class from event queue first.
658 static void htb_charge_class(struct htb_sched *q, struct htb_class *cl,
659 int level, struct sk_buff *skb)
661 int bytes = qdisc_pkt_len(skb);
662 enum htb_cmode old_mode;
666 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
667 if (cl->level >= level) {
668 if (cl->level == level)
670 htb_accnt_tokens(cl, bytes, diff);
672 cl->xstats.borrows++;
673 cl->tokens += diff; /* we moved t_c; update tokens */
675 htb_accnt_ctokens(cl, bytes, diff);
678 old_mode = cl->cmode;
680 htb_change_class_mode(q, cl, &diff);
681 if (old_mode != cl->cmode) {
682 if (old_mode != HTB_CAN_SEND)
683 htb_safe_rb_erase(&cl->pq_node, &q->hlevel[cl->level].wait_pq);
684 if (cl->cmode != HTB_CAN_SEND)
685 htb_add_to_wait_tree(q, cl, diff);
688 /* update basic stats except for leaves which are already updated */
690 bstats_update(&cl->bstats, skb);
697 * htb_do_events - make mode changes to classes at the level
699 * Scans event queue for pending events and applies them. Returns time of
700 * next pending event (0 for no event in pq, q->now for too many events).
701 * Note: Applied are events whose have cl->pq_key <= q->now.
703 static s64 htb_do_events(struct htb_sched *q, const int level,
706 /* don't run for longer than 2 jiffies; 2 is used instead of
707 * 1 to simplify things when jiffy is going to be incremented
710 unsigned long stop_at = start + 2;
711 struct rb_root *wait_pq = &q->hlevel[level].wait_pq;
713 while (time_before(jiffies, stop_at)) {
714 struct htb_class *cl;
716 struct rb_node *p = rb_first(wait_pq);
721 cl = rb_entry(p, struct htb_class, pq_node);
722 if (cl->pq_key > q->now)
725 htb_safe_rb_erase(p, wait_pq);
726 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
727 htb_change_class_mode(q, cl, &diff);
728 if (cl->cmode != HTB_CAN_SEND)
729 htb_add_to_wait_tree(q, cl, diff);
732 /* too much load - let's continue after a break for scheduling */
733 if (!(q->warned & HTB_WARN_TOOMANYEVENTS)) {
734 pr_warn("htb: too many events!\n");
735 q->warned |= HTB_WARN_TOOMANYEVENTS;
741 /* Returns class->node+prio from id-tree where classe's id is >= id. NULL
742 * is no such one exists.
744 static struct rb_node *htb_id_find_next_upper(int prio, struct rb_node *n,
747 struct rb_node *r = NULL;
749 struct htb_class *cl =
750 rb_entry(n, struct htb_class, node[prio]);
752 if (id > cl->common.classid) {
754 } else if (id < cl->common.classid) {
765 * htb_lookup_leaf - returns next leaf class in DRR order
767 * Find leaf where current feed pointers points to.
769 static struct htb_class *htb_lookup_leaf(struct htb_prio *hprio, const int prio)
773 struct rb_node *root;
774 struct rb_node **pptr;
776 } stk[TC_HTB_MAXDEPTH], *sp = stk;
778 BUG_ON(!hprio->row.rb_node);
779 sp->root = hprio->row.rb_node;
780 sp->pptr = &hprio->ptr;
781 sp->pid = &hprio->last_ptr_id;
783 for (i = 0; i < 65535; i++) {
784 if (!*sp->pptr && *sp->pid) {
785 /* ptr was invalidated but id is valid - try to recover
786 * the original or next ptr
789 htb_id_find_next_upper(prio, sp->root, *sp->pid);
791 *sp->pid = 0; /* ptr is valid now so that remove this hint as it
792 * can become out of date quickly
794 if (!*sp->pptr) { /* we are at right end; rewind & go up */
795 *sp->pptr = sp->root;
796 while ((*sp->pptr)->rb_left)
797 *sp->pptr = (*sp->pptr)->rb_left;
804 htb_next_rb_node(sp->pptr);
807 struct htb_class *cl;
808 struct htb_prio *clp;
810 cl = rb_entry(*sp->pptr, struct htb_class, node[prio]);
813 clp = &cl->inner.clprio[prio];
814 (++sp)->root = clp->feed.rb_node;
815 sp->pptr = &clp->ptr;
816 sp->pid = &clp->last_ptr_id;
823 /* dequeues packet at given priority and level; call only if
824 * you are sure that there is active class at prio/level
826 static struct sk_buff *htb_dequeue_tree(struct htb_sched *q, const int prio,
829 struct sk_buff *skb = NULL;
830 struct htb_class *cl, *start;
831 struct htb_level *hlevel = &q->hlevel[level];
832 struct htb_prio *hprio = &hlevel->hprio[prio];
834 /* look initial class up in the row */
835 start = cl = htb_lookup_leaf(hprio, prio);
842 /* class can be empty - it is unlikely but can be true if leaf
843 * qdisc drops packets in enqueue routine or if someone used
844 * graft operation on the leaf since last dequeue;
845 * simply deactivate and skip such class
847 if (unlikely(cl->leaf.q->q.qlen == 0)) {
848 struct htb_class *next;
849 htb_deactivate(q, cl);
851 /* row/level might become empty */
852 if ((q->row_mask[level] & (1 << prio)) == 0)
855 next = htb_lookup_leaf(hprio, prio);
857 if (cl == start) /* fix start if we just deleted it */
863 skb = cl->leaf.q->dequeue(cl->leaf.q);
864 if (likely(skb != NULL))
867 qdisc_warn_nonwc("htb", cl->leaf.q);
868 htb_next_rb_node(level ? &cl->parent->inner.clprio[prio].ptr:
869 &q->hlevel[0].hprio[prio].ptr);
870 cl = htb_lookup_leaf(hprio, prio);
872 } while (cl != start);
874 if (likely(skb != NULL)) {
875 bstats_update(&cl->bstats, skb);
876 cl->leaf.deficit[level] -= qdisc_pkt_len(skb);
877 if (cl->leaf.deficit[level] < 0) {
878 cl->leaf.deficit[level] += cl->quantum;
879 htb_next_rb_node(level ? &cl->parent->inner.clprio[prio].ptr :
880 &q->hlevel[0].hprio[prio].ptr);
882 /* this used to be after charge_class but this constelation
883 * gives us slightly better performance
885 if (!cl->leaf.q->q.qlen)
886 htb_deactivate(q, cl);
887 htb_charge_class(q, cl, level, skb);
892 static struct sk_buff *htb_dequeue(struct Qdisc *sch)
895 struct htb_sched *q = qdisc_priv(sch);
898 unsigned long start_at;
900 /* try to dequeue direct packets as high prio (!) to minimize cpu work */
901 skb = __qdisc_dequeue_head(&q->direct_queue);
904 qdisc_bstats_update(sch, skb);
905 qdisc_qstats_backlog_dec(sch, skb);
912 q->now = ktime_get_ns();
915 next_event = q->now + 5LLU * NSEC_PER_SEC;
917 for (level = 0; level < TC_HTB_MAXDEPTH; level++) {
918 /* common case optimization - skip event handler quickly */
920 s64 event = q->near_ev_cache[level];
922 if (q->now >= event) {
923 event = htb_do_events(q, level, start_at);
925 event = q->now + NSEC_PER_SEC;
926 q->near_ev_cache[level] = event;
929 if (next_event > event)
932 m = ~q->row_mask[level];
933 while (m != (int)(-1)) {
937 skb = htb_dequeue_tree(q, prio, level);
938 if (likely(skb != NULL))
942 if (likely(next_event > q->now))
943 qdisc_watchdog_schedule_ns(&q->watchdog, next_event);
945 schedule_work(&q->work);
950 /* reset all classes */
951 /* always caled under BH & queue lock */
952 static void htb_reset(struct Qdisc *sch)
954 struct htb_sched *q = qdisc_priv(sch);
955 struct htb_class *cl;
958 for (i = 0; i < q->clhash.hashsize; i++) {
959 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
961 memset(&cl->inner, 0, sizeof(cl->inner));
964 qdisc_reset(cl->leaf.q);
966 cl->prio_activity = 0;
967 cl->cmode = HTB_CAN_SEND;
970 qdisc_watchdog_cancel(&q->watchdog);
971 __qdisc_reset_queue(&q->direct_queue);
973 sch->qstats.backlog = 0;
974 memset(q->hlevel, 0, sizeof(q->hlevel));
975 memset(q->row_mask, 0, sizeof(q->row_mask));
978 static const struct nla_policy htb_policy[TCA_HTB_MAX + 1] = {
979 [TCA_HTB_PARMS] = { .len = sizeof(struct tc_htb_opt) },
980 [TCA_HTB_INIT] = { .len = sizeof(struct tc_htb_glob) },
981 [TCA_HTB_CTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
982 [TCA_HTB_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
983 [TCA_HTB_DIRECT_QLEN] = { .type = NLA_U32 },
984 [TCA_HTB_RATE64] = { .type = NLA_U64 },
985 [TCA_HTB_CEIL64] = { .type = NLA_U64 },
988 static void htb_work_func(struct work_struct *work)
990 struct htb_sched *q = container_of(work, struct htb_sched, work);
991 struct Qdisc *sch = q->watchdog.qdisc;
994 __netif_schedule(qdisc_root(sch));
998 static int htb_init(struct Qdisc *sch, struct nlattr *opt,
999 struct netlink_ext_ack *extack)
1001 struct htb_sched *q = qdisc_priv(sch);
1002 struct nlattr *tb[TCA_HTB_MAX + 1];
1003 struct tc_htb_glob *gopt;
1006 qdisc_watchdog_init(&q->watchdog, sch);
1007 INIT_WORK(&q->work, htb_work_func);
1012 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
1016 err = nla_parse_nested_deprecated(tb, TCA_HTB_MAX, opt, htb_policy,
1021 if (!tb[TCA_HTB_INIT])
1024 gopt = nla_data(tb[TCA_HTB_INIT]);
1025 if (gopt->version != HTB_VER >> 16)
1028 err = qdisc_class_hash_init(&q->clhash);
1032 qdisc_skb_head_init(&q->direct_queue);
1034 if (tb[TCA_HTB_DIRECT_QLEN])
1035 q->direct_qlen = nla_get_u32(tb[TCA_HTB_DIRECT_QLEN]);
1037 q->direct_qlen = qdisc_dev(sch)->tx_queue_len;
1039 if ((q->rate2quantum = gopt->rate2quantum) < 1)
1040 q->rate2quantum = 1;
1041 q->defcls = gopt->defcls;
1046 static int htb_dump(struct Qdisc *sch, struct sk_buff *skb)
1048 struct htb_sched *q = qdisc_priv(sch);
1049 struct nlattr *nest;
1050 struct tc_htb_glob gopt;
1052 sch->qstats.overlimits = q->overlimits;
1053 /* Its safe to not acquire qdisc lock. As we hold RTNL,
1054 * no change can happen on the qdisc parameters.
1057 gopt.direct_pkts = q->direct_pkts;
1058 gopt.version = HTB_VER;
1059 gopt.rate2quantum = q->rate2quantum;
1060 gopt.defcls = q->defcls;
1063 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1065 goto nla_put_failure;
1066 if (nla_put(skb, TCA_HTB_INIT, sizeof(gopt), &gopt) ||
1067 nla_put_u32(skb, TCA_HTB_DIRECT_QLEN, q->direct_qlen))
1068 goto nla_put_failure;
1070 return nla_nest_end(skb, nest);
1073 nla_nest_cancel(skb, nest);
1077 static int htb_dump_class(struct Qdisc *sch, unsigned long arg,
1078 struct sk_buff *skb, struct tcmsg *tcm)
1080 struct htb_class *cl = (struct htb_class *)arg;
1081 struct nlattr *nest;
1082 struct tc_htb_opt opt;
1084 /* Its safe to not acquire qdisc lock. As we hold RTNL,
1085 * no change can happen on the class parameters.
1087 tcm->tcm_parent = cl->parent ? cl->parent->common.classid : TC_H_ROOT;
1088 tcm->tcm_handle = cl->common.classid;
1089 if (!cl->level && cl->leaf.q)
1090 tcm->tcm_info = cl->leaf.q->handle;
1092 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1094 goto nla_put_failure;
1096 memset(&opt, 0, sizeof(opt));
1098 psched_ratecfg_getrate(&opt.rate, &cl->rate);
1099 opt.buffer = PSCHED_NS2TICKS(cl->buffer);
1100 psched_ratecfg_getrate(&opt.ceil, &cl->ceil);
1101 opt.cbuffer = PSCHED_NS2TICKS(cl->cbuffer);
1102 opt.quantum = cl->quantum;
1103 opt.prio = cl->prio;
1104 opt.level = cl->level;
1105 if (nla_put(skb, TCA_HTB_PARMS, sizeof(opt), &opt))
1106 goto nla_put_failure;
1107 if ((cl->rate.rate_bytes_ps >= (1ULL << 32)) &&
1108 nla_put_u64_64bit(skb, TCA_HTB_RATE64, cl->rate.rate_bytes_ps,
1110 goto nla_put_failure;
1111 if ((cl->ceil.rate_bytes_ps >= (1ULL << 32)) &&
1112 nla_put_u64_64bit(skb, TCA_HTB_CEIL64, cl->ceil.rate_bytes_ps,
1114 goto nla_put_failure;
1116 return nla_nest_end(skb, nest);
1119 nla_nest_cancel(skb, nest);
1124 htb_dump_class_stats(struct Qdisc *sch, unsigned long arg, struct gnet_dump *d)
1126 struct htb_class *cl = (struct htb_class *)arg;
1127 struct gnet_stats_queue qs = {
1129 .overlimits = cl->overlimits,
1133 if (!cl->level && cl->leaf.q)
1134 qdisc_qstats_qlen_backlog(cl->leaf.q, &qlen, &qs.backlog);
1136 cl->xstats.tokens = clamp_t(s64, PSCHED_NS2TICKS(cl->tokens),
1138 cl->xstats.ctokens = clamp_t(s64, PSCHED_NS2TICKS(cl->ctokens),
1141 if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
1142 d, NULL, &cl->bstats) < 0 ||
1143 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1144 gnet_stats_copy_queue(d, NULL, &qs, qlen) < 0)
1147 return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1150 static int htb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1151 struct Qdisc **old, struct netlink_ext_ack *extack)
1153 struct htb_class *cl = (struct htb_class *)arg;
1158 (new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1159 cl->common.classid, extack)) == NULL)
1162 *old = qdisc_replace(sch, new, &cl->leaf.q);
1166 static struct Qdisc *htb_leaf(struct Qdisc *sch, unsigned long arg)
1168 struct htb_class *cl = (struct htb_class *)arg;
1169 return !cl->level ? cl->leaf.q : NULL;
1172 static void htb_qlen_notify(struct Qdisc *sch, unsigned long arg)
1174 struct htb_class *cl = (struct htb_class *)arg;
1176 htb_deactivate(qdisc_priv(sch), cl);
1179 static inline int htb_parent_last_child(struct htb_class *cl)
1182 /* the root class */
1184 if (cl->parent->children > 1)
1185 /* not the last child */
1190 static void htb_parent_to_leaf(struct htb_sched *q, struct htb_class *cl,
1191 struct Qdisc *new_q)
1193 struct htb_class *parent = cl->parent;
1195 WARN_ON(cl->level || !cl->leaf.q || cl->prio_activity);
1197 if (parent->cmode != HTB_CAN_SEND)
1198 htb_safe_rb_erase(&parent->pq_node,
1199 &q->hlevel[parent->level].wait_pq);
1202 memset(&parent->inner, 0, sizeof(parent->inner));
1203 parent->leaf.q = new_q ? new_q : &noop_qdisc;
1204 parent->tokens = parent->buffer;
1205 parent->ctokens = parent->cbuffer;
1206 parent->t_c = ktime_get_ns();
1207 parent->cmode = HTB_CAN_SEND;
1210 static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
1213 WARN_ON(!cl->leaf.q);
1214 qdisc_put(cl->leaf.q);
1216 gen_kill_estimator(&cl->rate_est);
1217 tcf_block_put(cl->block);
1221 static void htb_destroy(struct Qdisc *sch)
1223 struct htb_sched *q = qdisc_priv(sch);
1224 struct hlist_node *next;
1225 struct htb_class *cl;
1228 cancel_work_sync(&q->work);
1229 qdisc_watchdog_cancel(&q->watchdog);
1230 /* This line used to be after htb_destroy_class call below
1231 * and surprisingly it worked in 2.4. But it must precede it
1232 * because filter need its target class alive to be able to call
1233 * unbind_filter on it (without Oops).
1235 tcf_block_put(q->block);
1237 for (i = 0; i < q->clhash.hashsize; i++) {
1238 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1239 tcf_block_put(cl->block);
1243 for (i = 0; i < q->clhash.hashsize; i++) {
1244 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1246 htb_destroy_class(sch, cl);
1248 qdisc_class_hash_destroy(&q->clhash);
1249 __qdisc_reset_queue(&q->direct_queue);
1252 static int htb_delete(struct Qdisc *sch, unsigned long arg)
1254 struct htb_sched *q = qdisc_priv(sch);
1255 struct htb_class *cl = (struct htb_class *)arg;
1256 struct Qdisc *new_q = NULL;
1259 /* TODO: why don't allow to delete subtree ? references ? does
1260 * tc subsys guarantee us that in htb_destroy it holds no class
1261 * refs so that we can remove children safely there ?
1263 if (cl->children || cl->filter_cnt)
1266 if (!cl->level && htb_parent_last_child(cl)) {
1267 new_q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1268 cl->parent->common.classid,
1276 qdisc_purge_queue(cl->leaf.q);
1278 /* delete from hash and active; remainder in destroy_class */
1279 qdisc_class_hash_remove(&q->clhash, &cl->common);
1281 cl->parent->children--;
1283 if (cl->prio_activity)
1284 htb_deactivate(q, cl);
1286 if (cl->cmode != HTB_CAN_SEND)
1287 htb_safe_rb_erase(&cl->pq_node,
1288 &q->hlevel[cl->level].wait_pq);
1291 htb_parent_to_leaf(q, cl, new_q);
1293 sch_tree_unlock(sch);
1295 htb_destroy_class(sch, cl);
1299 static int htb_change_class(struct Qdisc *sch, u32 classid,
1300 u32 parentid, struct nlattr **tca,
1301 unsigned long *arg, struct netlink_ext_ack *extack)
1304 struct htb_sched *q = qdisc_priv(sch);
1305 struct htb_class *cl = (struct htb_class *)*arg, *parent;
1306 struct nlattr *opt = tca[TCA_OPTIONS];
1307 struct nlattr *tb[TCA_HTB_MAX + 1];
1308 struct Qdisc *parent_qdisc = NULL;
1309 struct tc_htb_opt *hopt;
1313 /* extract all subattrs from opt attr */
1317 err = nla_parse_nested_deprecated(tb, TCA_HTB_MAX, opt, htb_policy,
1323 if (tb[TCA_HTB_PARMS] == NULL)
1326 parent = parentid == TC_H_ROOT ? NULL : htb_find(parentid, sch);
1328 hopt = nla_data(tb[TCA_HTB_PARMS]);
1329 if (!hopt->rate.rate || !hopt->ceil.rate)
1332 /* Keeping backward compatible with rate_table based iproute2 tc */
1333 if (hopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
1334 qdisc_put_rtab(qdisc_get_rtab(&hopt->rate, tb[TCA_HTB_RTAB],
1337 if (hopt->ceil.linklayer == TC_LINKLAYER_UNAWARE)
1338 qdisc_put_rtab(qdisc_get_rtab(&hopt->ceil, tb[TCA_HTB_CTAB],
1341 if (!cl) { /* new class */
1342 struct Qdisc *new_q;
1346 struct gnet_estimator opt;
1349 .nla_len = nla_attr_size(sizeof(est.opt)),
1350 .nla_type = TCA_RATE,
1353 /* 4s interval, 16s averaging constant */
1359 /* check for valid classid */
1360 if (!classid || TC_H_MAJ(classid ^ sch->handle) ||
1361 htb_find(classid, sch))
1364 /* check maximal depth */
1365 if (parent && parent->parent && parent->parent->level < 2) {
1366 pr_err("htb: tree is too deep\n");
1370 cl = kzalloc(sizeof(*cl), GFP_KERNEL);
1374 err = tcf_block_get(&cl->block, &cl->filter_list, sch, extack);
1379 if (htb_rate_est || tca[TCA_RATE]) {
1380 err = gen_new_estimator(&cl->bstats, NULL,
1383 qdisc_root_sleeping_running(sch),
1384 tca[TCA_RATE] ? : &est.nla);
1386 tcf_block_put(cl->block);
1393 RB_CLEAR_NODE(&cl->pq_node);
1395 for (prio = 0; prio < TC_HTB_NUMPRIO; prio++)
1396 RB_CLEAR_NODE(&cl->node[prio]);
1398 /* create leaf qdisc early because it uses kmalloc(GFP_KERNEL)
1399 * so that can't be used inside of sch_tree_lock
1400 * -- thanks to Karlis Peisenieks
1402 new_q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1405 if (parent && !parent->level) {
1406 /* turn parent into inner node */
1407 qdisc_purge_queue(parent->leaf.q);
1408 parent_qdisc = parent->leaf.q;
1409 if (parent->prio_activity)
1410 htb_deactivate(q, parent);
1412 /* remove from evt list because of level change */
1413 if (parent->cmode != HTB_CAN_SEND) {
1414 htb_safe_rb_erase(&parent->pq_node, &q->hlevel[0].wait_pq);
1415 parent->cmode = HTB_CAN_SEND;
1417 parent->level = (parent->parent ? parent->parent->level
1418 : TC_HTB_MAXDEPTH) - 1;
1419 memset(&parent->inner, 0, sizeof(parent->inner));
1421 /* leaf (we) needs elementary qdisc */
1422 cl->leaf.q = new_q ? new_q : &noop_qdisc;
1424 cl->common.classid = classid;
1425 cl->parent = parent;
1427 /* set class to be in HTB_CAN_SEND state */
1428 cl->tokens = PSCHED_TICKS2NS(hopt->buffer);
1429 cl->ctokens = PSCHED_TICKS2NS(hopt->cbuffer);
1430 cl->mbuffer = 60ULL * NSEC_PER_SEC; /* 1min */
1431 cl->t_c = ktime_get_ns();
1432 cl->cmode = HTB_CAN_SEND;
1434 /* attach to the hash list and parent's family */
1435 qdisc_class_hash_insert(&q->clhash, &cl->common);
1438 if (cl->leaf.q != &noop_qdisc)
1439 qdisc_hash_add(cl->leaf.q, true);
1441 if (tca[TCA_RATE]) {
1442 err = gen_replace_estimator(&cl->bstats, NULL,
1445 qdisc_root_sleeping_running(sch),
1453 rate64 = tb[TCA_HTB_RATE64] ? nla_get_u64(tb[TCA_HTB_RATE64]) : 0;
1455 ceil64 = tb[TCA_HTB_CEIL64] ? nla_get_u64(tb[TCA_HTB_CEIL64]) : 0;
1457 psched_ratecfg_precompute(&cl->rate, &hopt->rate, rate64);
1458 psched_ratecfg_precompute(&cl->ceil, &hopt->ceil, ceil64);
1460 /* it used to be a nasty bug here, we have to check that node
1461 * is really leaf before changing cl->leaf !
1464 u64 quantum = cl->rate.rate_bytes_ps;
1466 do_div(quantum, q->rate2quantum);
1467 cl->quantum = min_t(u64, quantum, INT_MAX);
1469 if (!hopt->quantum && cl->quantum < 1000) {
1473 if (!hopt->quantum && cl->quantum > 200000) {
1475 cl->quantum = 200000;
1478 cl->quantum = hopt->quantum;
1479 if ((cl->prio = hopt->prio) >= TC_HTB_NUMPRIO)
1480 cl->prio = TC_HTB_NUMPRIO - 1;
1483 cl->buffer = PSCHED_TICKS2NS(hopt->buffer);
1484 cl->cbuffer = PSCHED_TICKS2NS(hopt->cbuffer);
1486 sch_tree_unlock(sch);
1487 qdisc_put(parent_qdisc);
1490 pr_warn("HTB: quantum of class %X is %s. Consider r2q change.\n",
1491 cl->common.classid, (warn == -1 ? "small" : "big"));
1493 qdisc_class_hash_grow(sch, &q->clhash);
1495 *arg = (unsigned long)cl;
1502 static struct tcf_block *htb_tcf_block(struct Qdisc *sch, unsigned long arg,
1503 struct netlink_ext_ack *extack)
1505 struct htb_sched *q = qdisc_priv(sch);
1506 struct htb_class *cl = (struct htb_class *)arg;
1508 return cl ? cl->block : q->block;
1511 static unsigned long htb_bind_filter(struct Qdisc *sch, unsigned long parent,
1514 struct htb_class *cl = htb_find(classid, sch);
1516 /*if (cl && !cl->level) return 0;
1517 * The line above used to be there to prevent attaching filters to
1518 * leaves. But at least tc_index filter uses this just to get class
1519 * for other reasons so that we have to allow for it.
1521 * 19.6.2002 As Werner explained it is ok - bind filter is just
1522 * another way to "lock" the class - unlike "get" this lock can
1523 * be broken by class during destroy IIUC.
1527 return (unsigned long)cl;
1530 static void htb_unbind_filter(struct Qdisc *sch, unsigned long arg)
1532 struct htb_class *cl = (struct htb_class *)arg;
1538 static void htb_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1540 struct htb_sched *q = qdisc_priv(sch);
1541 struct htb_class *cl;
1547 for (i = 0; i < q->clhash.hashsize; i++) {
1548 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1549 if (arg->count < arg->skip) {
1553 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1562 static const struct Qdisc_class_ops htb_class_ops = {
1565 .qlen_notify = htb_qlen_notify,
1567 .change = htb_change_class,
1568 .delete = htb_delete,
1570 .tcf_block = htb_tcf_block,
1571 .bind_tcf = htb_bind_filter,
1572 .unbind_tcf = htb_unbind_filter,
1573 .dump = htb_dump_class,
1574 .dump_stats = htb_dump_class_stats,
1577 static struct Qdisc_ops htb_qdisc_ops __read_mostly = {
1578 .cl_ops = &htb_class_ops,
1580 .priv_size = sizeof(struct htb_sched),
1581 .enqueue = htb_enqueue,
1582 .dequeue = htb_dequeue,
1583 .peek = qdisc_peek_dequeued,
1586 .destroy = htb_destroy,
1588 .owner = THIS_MODULE,
1591 static int __init htb_module_init(void)
1593 return register_qdisc(&htb_qdisc_ops);
1595 static void __exit htb_module_exit(void)
1597 unregister_qdisc(&htb_qdisc_ops);
1600 module_init(htb_module_init)
1601 module_exit(htb_module_exit)
1602 MODULE_LICENSE("GPL");