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 supplied 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 gnet_stats_basic_packed bstats_bias;
118 struct tc_htb_xstats xstats; /* our special stats */
120 /* token bucket parameters */
121 s64 tokens, ctokens;/* current number of tokens */
122 s64 t_c; /* checkpoint time */
125 struct htb_class_leaf {
126 int deficit[TC_HTB_MAXDEPTH];
129 struct htb_class_inner {
130 struct htb_prio clprio[TC_HTB_NUMPRIO];
135 int prio_activity; /* for which prios are we active */
136 enum htb_cmode cmode; /* current mode of the class */
137 struct rb_node pq_node; /* node for event queue */
138 struct rb_node node[TC_HTB_NUMPRIO]; /* node for self or feed tree */
140 unsigned int drops ____cacheline_aligned_in_smp;
141 unsigned int overlimits;
145 struct rb_root wait_pq;
146 struct htb_prio hprio[TC_HTB_NUMPRIO];
150 struct Qdisc_class_hash clhash;
151 int defcls; /* class where unclassified flows go to */
152 int rate2quantum; /* quant = rate / rate2quantum */
154 /* filters for qdisc itself */
155 struct tcf_proto __rcu *filter_list;
156 struct tcf_block *block;
158 #define HTB_WARN_TOOMANYEVENTS 0x1
159 unsigned int warned; /* only one warning */
161 struct work_struct work;
163 /* non shaped skbs; let them go directly thru */
164 struct qdisc_skb_head direct_queue;
168 struct qdisc_watchdog watchdog;
170 s64 now; /* cached dequeue time */
172 /* time of nearest event per level (row) */
173 s64 near_ev_cache[TC_HTB_MAXDEPTH];
175 int row_mask[TC_HTB_MAXDEPTH];
177 struct htb_level hlevel[TC_HTB_MAXDEPTH];
179 struct Qdisc **direct_qdiscs;
180 unsigned int num_direct_qdiscs;
185 /* find class in global hash table using given handle */
186 static inline struct htb_class *htb_find(u32 handle, struct Qdisc *sch)
188 struct htb_sched *q = qdisc_priv(sch);
189 struct Qdisc_class_common *clc;
191 clc = qdisc_class_find(&q->clhash, handle);
194 return container_of(clc, struct htb_class, common);
197 static unsigned long htb_search(struct Qdisc *sch, u32 handle)
199 return (unsigned long)htb_find(handle, sch);
202 * htb_classify - classify a packet into class
204 * It returns NULL if the packet should be dropped or -1 if the packet
205 * should be passed directly thru. In all other cases leaf class is returned.
206 * We allow direct class selection by classid in priority. The we examine
207 * filters in qdisc and in inner nodes (if higher filter points to the inner
208 * node). If we end up with classid MAJOR:0 we enqueue the skb into special
209 * internal fifo (direct). These packets then go directly thru. If we still
210 * have no valid leaf we try to use MAJOR:default leaf. It still unsuccessful
211 * then finish and return direct queue.
213 #define HTB_DIRECT ((struct htb_class *)-1L)
215 static struct htb_class *htb_classify(struct sk_buff *skb, struct Qdisc *sch,
218 struct htb_sched *q = qdisc_priv(sch);
219 struct htb_class *cl;
220 struct tcf_result res;
221 struct tcf_proto *tcf;
224 /* allow to select class by setting skb->priority to valid classid;
225 * note that nfmark can be used too by attaching filter fw with no
228 if (skb->priority == sch->handle)
229 return HTB_DIRECT; /* X:0 (direct flow) selected */
230 cl = htb_find(skb->priority, sch);
234 /* Start with inner filter chain if a non-leaf class is selected */
235 tcf = rcu_dereference_bh(cl->filter_list);
237 tcf = rcu_dereference_bh(q->filter_list);
240 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
241 while (tcf && (result = tcf_classify(skb, tcf, &res, false)) >= 0) {
242 #ifdef CONFIG_NET_CLS_ACT
247 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
253 cl = (void *)res.class;
255 if (res.classid == sch->handle)
256 return HTB_DIRECT; /* X:0 (direct flow) */
257 cl = htb_find(res.classid, sch);
259 break; /* filter selected invalid classid */
262 return cl; /* we hit leaf; return it */
264 /* we have got inner class; apply inner filter chain */
265 tcf = rcu_dereference_bh(cl->filter_list);
267 /* classification failed; try to use default class */
268 cl = htb_find(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
269 if (!cl || cl->level)
270 return HTB_DIRECT; /* bad default .. this is safe bet */
275 * htb_add_to_id_tree - adds class to the round robin list
276 * @root: the root of the tree
277 * @cl: the class to add
278 * @prio: the give prio in class
280 * Routine adds class to the list (actually tree) sorted by classid.
281 * Make sure that class is not already on such list for given prio.
283 static void htb_add_to_id_tree(struct rb_root *root,
284 struct htb_class *cl, int prio)
286 struct rb_node **p = &root->rb_node, *parent = NULL;
291 c = rb_entry(parent, struct htb_class, node[prio]);
293 if (cl->common.classid > c->common.classid)
294 p = &parent->rb_right;
296 p = &parent->rb_left;
298 rb_link_node(&cl->node[prio], parent, p);
299 rb_insert_color(&cl->node[prio], root);
303 * htb_add_to_wait_tree - adds class to the event queue with delay
304 * @q: the priority event queue
305 * @cl: the class to add
306 * @delay: delay in microseconds
308 * The class is added to priority event queue to indicate that class will
309 * change its mode in cl->pq_key microseconds. Make sure that class is not
310 * already in the queue.
312 static void htb_add_to_wait_tree(struct htb_sched *q,
313 struct htb_class *cl, s64 delay)
315 struct rb_node **p = &q->hlevel[cl->level].wait_pq.rb_node, *parent = NULL;
317 cl->pq_key = q->now + delay;
318 if (cl->pq_key == q->now)
321 /* update the nearest event cache */
322 if (q->near_ev_cache[cl->level] > cl->pq_key)
323 q->near_ev_cache[cl->level] = cl->pq_key;
328 c = rb_entry(parent, struct htb_class, pq_node);
329 if (cl->pq_key >= c->pq_key)
330 p = &parent->rb_right;
332 p = &parent->rb_left;
334 rb_link_node(&cl->pq_node, parent, p);
335 rb_insert_color(&cl->pq_node, &q->hlevel[cl->level].wait_pq);
339 * htb_next_rb_node - finds next node in binary tree
340 * @n: the current node in binary tree
342 * When we are past last key we return NULL.
343 * Average complexity is 2 steps per call.
345 static inline void htb_next_rb_node(struct rb_node **n)
351 * htb_add_class_to_row - add class to its row
352 * @q: the priority event queue
353 * @cl: the class to add
354 * @mask: the given priorities in class in bitmap
356 * The class is added to row at priorities marked in mask.
357 * It does nothing if mask == 0.
359 static inline void htb_add_class_to_row(struct htb_sched *q,
360 struct htb_class *cl, int mask)
362 q->row_mask[cl->level] |= mask;
364 int prio = ffz(~mask);
365 mask &= ~(1 << prio);
366 htb_add_to_id_tree(&q->hlevel[cl->level].hprio[prio].row, cl, prio);
370 /* If this triggers, it is a bug in this code, but it need not be fatal */
371 static void htb_safe_rb_erase(struct rb_node *rb, struct rb_root *root)
373 if (RB_EMPTY_NODE(rb)) {
383 * htb_remove_class_from_row - removes class from its row
384 * @q: the priority event queue
385 * @cl: the class to add
386 * @mask: the given priorities in class in bitmap
388 * The class is removed from row at priorities marked in mask.
389 * It does nothing if mask == 0.
391 static inline void htb_remove_class_from_row(struct htb_sched *q,
392 struct htb_class *cl, int mask)
395 struct htb_level *hlevel = &q->hlevel[cl->level];
398 int prio = ffz(~mask);
399 struct htb_prio *hprio = &hlevel->hprio[prio];
401 mask &= ~(1 << prio);
402 if (hprio->ptr == cl->node + prio)
403 htb_next_rb_node(&hprio->ptr);
405 htb_safe_rb_erase(cl->node + prio, &hprio->row);
406 if (!hprio->row.rb_node)
409 q->row_mask[cl->level] &= ~m;
413 * htb_activate_prios - creates active classe's feed chain
414 * @q: the priority event queue
415 * @cl: the class to activate
417 * The class is connected to ancestors and/or appropriate rows
418 * for priorities it is participating on. cl->cmode must be new
419 * (activated) mode. It does nothing if cl->prio_activity == 0.
421 static void htb_activate_prios(struct htb_sched *q, struct htb_class *cl)
423 struct htb_class *p = cl->parent;
424 long m, mask = cl->prio_activity;
426 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
432 if (p->inner.clprio[prio].feed.rb_node)
433 /* parent already has its feed in use so that
434 * reset bit in mask as parent is already ok
436 mask &= ~(1 << prio);
438 htb_add_to_id_tree(&p->inner.clprio[prio].feed, cl, prio);
440 p->prio_activity |= mask;
445 if (cl->cmode == HTB_CAN_SEND && mask)
446 htb_add_class_to_row(q, cl, mask);
450 * htb_deactivate_prios - remove class from feed chain
451 * @q: the priority event queue
452 * @cl: the class to deactivate
454 * cl->cmode must represent old mode (before deactivation). It does
455 * nothing if cl->prio_activity == 0. Class is removed from all feed
458 static void htb_deactivate_prios(struct htb_sched *q, struct htb_class *cl)
460 struct htb_class *p = cl->parent;
461 long m, mask = cl->prio_activity;
463 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
470 if (p->inner.clprio[prio].ptr == cl->node + prio) {
471 /* we are removing child which is pointed to from
472 * parent feed - forget the pointer but remember
475 p->inner.clprio[prio].last_ptr_id = cl->common.classid;
476 p->inner.clprio[prio].ptr = NULL;
479 htb_safe_rb_erase(cl->node + prio,
480 &p->inner.clprio[prio].feed);
482 if (!p->inner.clprio[prio].feed.rb_node)
486 p->prio_activity &= ~mask;
491 if (cl->cmode == HTB_CAN_SEND && mask)
492 htb_remove_class_from_row(q, cl, mask);
495 static inline s64 htb_lowater(const struct htb_class *cl)
498 return cl->cmode != HTB_CANT_SEND ? -cl->cbuffer : 0;
502 static inline s64 htb_hiwater(const struct htb_class *cl)
505 return cl->cmode == HTB_CAN_SEND ? -cl->buffer : 0;
512 * htb_class_mode - computes and returns current class mode
513 * @cl: the target class
514 * @diff: diff time in microseconds
516 * It computes cl's mode at time cl->t_c+diff and returns it. If mode
517 * is not HTB_CAN_SEND then cl->pq_key is updated to time difference
518 * from now to time when cl will change its state.
519 * Also it is worth to note that class mode doesn't change simply
520 * at cl->{c,}tokens == 0 but there can rather be hysteresis of
521 * 0 .. -cl->{c,}buffer range. It is meant to limit number of
522 * mode transitions per time unit. The speed gain is about 1/6.
524 static inline enum htb_cmode
525 htb_class_mode(struct htb_class *cl, s64 *diff)
529 if ((toks = (cl->ctokens + *diff)) < htb_lowater(cl)) {
531 return HTB_CANT_SEND;
534 if ((toks = (cl->tokens + *diff)) >= htb_hiwater(cl))
538 return HTB_MAY_BORROW;
542 * htb_change_class_mode - changes classe's mode
543 * @q: the priority event queue
544 * @cl: the target class
545 * @diff: diff time in microseconds
547 * This should be the only way how to change classe's mode under normal
548 * circumstances. Routine will update feed lists linkage, change mode
549 * and add class to the wait event queue if appropriate. New mode should
550 * be different from old one and cl->pq_key has to be valid if changing
551 * to mode other than HTB_CAN_SEND (see htb_add_to_wait_tree).
554 htb_change_class_mode(struct htb_sched *q, struct htb_class *cl, s64 *diff)
556 enum htb_cmode new_mode = htb_class_mode(cl, diff);
558 if (new_mode == cl->cmode)
561 if (new_mode == HTB_CANT_SEND) {
566 if (cl->prio_activity) { /* not necessary: speed optimization */
567 if (cl->cmode != HTB_CANT_SEND)
568 htb_deactivate_prios(q, cl);
569 cl->cmode = new_mode;
570 if (new_mode != HTB_CANT_SEND)
571 htb_activate_prios(q, cl);
573 cl->cmode = new_mode;
577 * htb_activate - inserts leaf cl into appropriate active feeds
578 * @q: the priority event queue
579 * @cl: the target class
581 * Routine learns (new) priority of leaf and activates feed chain
582 * for the prio. It can be called on already active leaf safely.
583 * It also adds leaf into droplist.
585 static inline void htb_activate(struct htb_sched *q, struct htb_class *cl)
587 WARN_ON(cl->level || !cl->leaf.q || !cl->leaf.q->q.qlen);
589 if (!cl->prio_activity) {
590 cl->prio_activity = 1 << cl->prio;
591 htb_activate_prios(q, cl);
596 * htb_deactivate - remove leaf cl from active feeds
597 * @q: the priority event queue
598 * @cl: the target class
600 * Make sure that leaf is active. In the other words it can't be called
601 * with non-active leaf. It also removes class from the drop list.
603 static inline void htb_deactivate(struct htb_sched *q, struct htb_class *cl)
605 WARN_ON(!cl->prio_activity);
607 htb_deactivate_prios(q, cl);
608 cl->prio_activity = 0;
611 static int htb_enqueue(struct sk_buff *skb, struct Qdisc *sch,
612 struct sk_buff **to_free)
615 unsigned int len = qdisc_pkt_len(skb);
616 struct htb_sched *q = qdisc_priv(sch);
617 struct htb_class *cl = htb_classify(skb, sch, &ret);
619 if (cl == HTB_DIRECT) {
620 /* enqueue to helper queue */
621 if (q->direct_queue.qlen < q->direct_qlen) {
622 __qdisc_enqueue_tail(skb, &q->direct_queue);
625 return qdisc_drop(skb, sch, to_free);
627 #ifdef CONFIG_NET_CLS_ACT
629 if (ret & __NET_XMIT_BYPASS)
630 qdisc_qstats_drop(sch);
631 __qdisc_drop(skb, to_free);
634 } else if ((ret = qdisc_enqueue(skb, cl->leaf.q,
635 to_free)) != NET_XMIT_SUCCESS) {
636 if (net_xmit_drop_count(ret)) {
637 qdisc_qstats_drop(sch);
645 sch->qstats.backlog += len;
647 return NET_XMIT_SUCCESS;
650 static inline void htb_accnt_tokens(struct htb_class *cl, int bytes, s64 diff)
652 s64 toks = diff + cl->tokens;
654 if (toks > cl->buffer)
656 toks -= (s64) psched_l2t_ns(&cl->rate, bytes);
657 if (toks <= -cl->mbuffer)
658 toks = 1 - cl->mbuffer;
663 static inline void htb_accnt_ctokens(struct htb_class *cl, int bytes, s64 diff)
665 s64 toks = diff + cl->ctokens;
667 if (toks > cl->cbuffer)
669 toks -= (s64) psched_l2t_ns(&cl->ceil, bytes);
670 if (toks <= -cl->mbuffer)
671 toks = 1 - cl->mbuffer;
677 * htb_charge_class - charges amount "bytes" to leaf and ancestors
678 * @q: the priority event queue
679 * @cl: the class to start iterate
680 * @level: the minimum level to account
681 * @skb: the socket buffer
683 * Routine assumes that packet "bytes" long was dequeued from leaf cl
684 * borrowing from "level". It accounts bytes to ceil leaky bucket for
685 * leaf and all ancestors and to rate bucket for ancestors at levels
686 * "level" and higher. It also handles possible change of mode resulting
687 * from the update. Note that mode can also increase here (MAY_BORROW to
688 * CAN_SEND) because we can use more precise clock that event queue here.
689 * In such case we remove class from event queue first.
691 static void htb_charge_class(struct htb_sched *q, struct htb_class *cl,
692 int level, struct sk_buff *skb)
694 int bytes = qdisc_pkt_len(skb);
695 enum htb_cmode old_mode;
699 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
700 if (cl->level >= level) {
701 if (cl->level == level)
703 htb_accnt_tokens(cl, bytes, diff);
705 cl->xstats.borrows++;
706 cl->tokens += diff; /* we moved t_c; update tokens */
708 htb_accnt_ctokens(cl, bytes, diff);
711 old_mode = cl->cmode;
713 htb_change_class_mode(q, cl, &diff);
714 if (old_mode != cl->cmode) {
715 if (old_mode != HTB_CAN_SEND)
716 htb_safe_rb_erase(&cl->pq_node, &q->hlevel[cl->level].wait_pq);
717 if (cl->cmode != HTB_CAN_SEND)
718 htb_add_to_wait_tree(q, cl, diff);
721 /* update basic stats except for leaves which are already updated */
723 bstats_update(&cl->bstats, skb);
730 * htb_do_events - make mode changes to classes at the level
731 * @q: the priority event queue
732 * @level: which wait_pq in 'q->hlevel'
733 * @start: start jiffies
735 * Scans event queue for pending events and applies them. Returns time of
736 * next pending event (0 for no event in pq, q->now for too many events).
737 * Note: Applied are events whose have cl->pq_key <= q->now.
739 static s64 htb_do_events(struct htb_sched *q, const int level,
742 /* don't run for longer than 2 jiffies; 2 is used instead of
743 * 1 to simplify things when jiffy is going to be incremented
746 unsigned long stop_at = start + 2;
747 struct rb_root *wait_pq = &q->hlevel[level].wait_pq;
749 while (time_before(jiffies, stop_at)) {
750 struct htb_class *cl;
752 struct rb_node *p = rb_first(wait_pq);
757 cl = rb_entry(p, struct htb_class, pq_node);
758 if (cl->pq_key > q->now)
761 htb_safe_rb_erase(p, wait_pq);
762 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
763 htb_change_class_mode(q, cl, &diff);
764 if (cl->cmode != HTB_CAN_SEND)
765 htb_add_to_wait_tree(q, cl, diff);
768 /* too much load - let's continue after a break for scheduling */
769 if (!(q->warned & HTB_WARN_TOOMANYEVENTS)) {
770 pr_warn("htb: too many events!\n");
771 q->warned |= HTB_WARN_TOOMANYEVENTS;
777 /* Returns class->node+prio from id-tree where classe's id is >= id. NULL
778 * is no such one exists.
780 static struct rb_node *htb_id_find_next_upper(int prio, struct rb_node *n,
783 struct rb_node *r = NULL;
785 struct htb_class *cl =
786 rb_entry(n, struct htb_class, node[prio]);
788 if (id > cl->common.classid) {
790 } else if (id < cl->common.classid) {
801 * htb_lookup_leaf - returns next leaf class in DRR order
802 * @hprio: the current one
803 * @prio: which prio in class
805 * Find leaf where current feed pointers points to.
807 static struct htb_class *htb_lookup_leaf(struct htb_prio *hprio, const int prio)
811 struct rb_node *root;
812 struct rb_node **pptr;
814 } stk[TC_HTB_MAXDEPTH], *sp = stk;
816 BUG_ON(!hprio->row.rb_node);
817 sp->root = hprio->row.rb_node;
818 sp->pptr = &hprio->ptr;
819 sp->pid = &hprio->last_ptr_id;
821 for (i = 0; i < 65535; i++) {
822 if (!*sp->pptr && *sp->pid) {
823 /* ptr was invalidated but id is valid - try to recover
824 * the original or next ptr
827 htb_id_find_next_upper(prio, sp->root, *sp->pid);
829 *sp->pid = 0; /* ptr is valid now so that remove this hint as it
830 * can become out of date quickly
832 if (!*sp->pptr) { /* we are at right end; rewind & go up */
833 *sp->pptr = sp->root;
834 while ((*sp->pptr)->rb_left)
835 *sp->pptr = (*sp->pptr)->rb_left;
842 htb_next_rb_node(sp->pptr);
845 struct htb_class *cl;
846 struct htb_prio *clp;
848 cl = rb_entry(*sp->pptr, struct htb_class, node[prio]);
851 clp = &cl->inner.clprio[prio];
852 (++sp)->root = clp->feed.rb_node;
853 sp->pptr = &clp->ptr;
854 sp->pid = &clp->last_ptr_id;
861 /* dequeues packet at given priority and level; call only if
862 * you are sure that there is active class at prio/level
864 static struct sk_buff *htb_dequeue_tree(struct htb_sched *q, const int prio,
867 struct sk_buff *skb = NULL;
868 struct htb_class *cl, *start;
869 struct htb_level *hlevel = &q->hlevel[level];
870 struct htb_prio *hprio = &hlevel->hprio[prio];
872 /* look initial class up in the row */
873 start = cl = htb_lookup_leaf(hprio, prio);
880 /* class can be empty - it is unlikely but can be true if leaf
881 * qdisc drops packets in enqueue routine or if someone used
882 * graft operation on the leaf since last dequeue;
883 * simply deactivate and skip such class
885 if (unlikely(cl->leaf.q->q.qlen == 0)) {
886 struct htb_class *next;
887 htb_deactivate(q, cl);
889 /* row/level might become empty */
890 if ((q->row_mask[level] & (1 << prio)) == 0)
893 next = htb_lookup_leaf(hprio, prio);
895 if (cl == start) /* fix start if we just deleted it */
901 skb = cl->leaf.q->dequeue(cl->leaf.q);
902 if (likely(skb != NULL))
905 qdisc_warn_nonwc("htb", cl->leaf.q);
906 htb_next_rb_node(level ? &cl->parent->inner.clprio[prio].ptr:
907 &q->hlevel[0].hprio[prio].ptr);
908 cl = htb_lookup_leaf(hprio, prio);
910 } while (cl != start);
912 if (likely(skb != NULL)) {
913 bstats_update(&cl->bstats, skb);
914 cl->leaf.deficit[level] -= qdisc_pkt_len(skb);
915 if (cl->leaf.deficit[level] < 0) {
916 cl->leaf.deficit[level] += cl->quantum;
917 htb_next_rb_node(level ? &cl->parent->inner.clprio[prio].ptr :
918 &q->hlevel[0].hprio[prio].ptr);
920 /* this used to be after charge_class but this constelation
921 * gives us slightly better performance
923 if (!cl->leaf.q->q.qlen)
924 htb_deactivate(q, cl);
925 htb_charge_class(q, cl, level, skb);
930 static struct sk_buff *htb_dequeue(struct Qdisc *sch)
933 struct htb_sched *q = qdisc_priv(sch);
936 unsigned long start_at;
938 /* try to dequeue direct packets as high prio (!) to minimize cpu work */
939 skb = __qdisc_dequeue_head(&q->direct_queue);
942 qdisc_bstats_update(sch, skb);
943 qdisc_qstats_backlog_dec(sch, skb);
950 q->now = ktime_get_ns();
953 next_event = q->now + 5LLU * NSEC_PER_SEC;
955 for (level = 0; level < TC_HTB_MAXDEPTH; level++) {
956 /* common case optimization - skip event handler quickly */
958 s64 event = q->near_ev_cache[level];
960 if (q->now >= event) {
961 event = htb_do_events(q, level, start_at);
963 event = q->now + NSEC_PER_SEC;
964 q->near_ev_cache[level] = event;
967 if (next_event > event)
970 m = ~q->row_mask[level];
971 while (m != (int)(-1)) {
975 skb = htb_dequeue_tree(q, prio, level);
976 if (likely(skb != NULL))
980 if (likely(next_event > q->now))
981 qdisc_watchdog_schedule_ns(&q->watchdog, next_event);
983 schedule_work(&q->work);
988 /* reset all classes */
989 /* always caled under BH & queue lock */
990 static void htb_reset(struct Qdisc *sch)
992 struct htb_sched *q = qdisc_priv(sch);
993 struct htb_class *cl;
996 for (i = 0; i < q->clhash.hashsize; i++) {
997 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
999 memset(&cl->inner, 0, sizeof(cl->inner));
1001 if (cl->leaf.q && !q->offload)
1002 qdisc_reset(cl->leaf.q);
1004 cl->prio_activity = 0;
1005 cl->cmode = HTB_CAN_SEND;
1008 qdisc_watchdog_cancel(&q->watchdog);
1009 __qdisc_reset_queue(&q->direct_queue);
1011 sch->qstats.backlog = 0;
1012 memset(q->hlevel, 0, sizeof(q->hlevel));
1013 memset(q->row_mask, 0, sizeof(q->row_mask));
1016 static const struct nla_policy htb_policy[TCA_HTB_MAX + 1] = {
1017 [TCA_HTB_PARMS] = { .len = sizeof(struct tc_htb_opt) },
1018 [TCA_HTB_INIT] = { .len = sizeof(struct tc_htb_glob) },
1019 [TCA_HTB_CTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
1020 [TCA_HTB_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
1021 [TCA_HTB_DIRECT_QLEN] = { .type = NLA_U32 },
1022 [TCA_HTB_RATE64] = { .type = NLA_U64 },
1023 [TCA_HTB_CEIL64] = { .type = NLA_U64 },
1024 [TCA_HTB_OFFLOAD] = { .type = NLA_FLAG },
1027 static void htb_work_func(struct work_struct *work)
1029 struct htb_sched *q = container_of(work, struct htb_sched, work);
1030 struct Qdisc *sch = q->watchdog.qdisc;
1033 __netif_schedule(qdisc_root(sch));
1037 static void htb_set_lockdep_class_child(struct Qdisc *q)
1039 static struct lock_class_key child_key;
1041 lockdep_set_class(qdisc_lock(q), &child_key);
1044 static int htb_offload(struct net_device *dev, struct tc_htb_qopt_offload *opt)
1046 return dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_HTB, opt);
1049 static int htb_init(struct Qdisc *sch, struct nlattr *opt,
1050 struct netlink_ext_ack *extack)
1052 struct net_device *dev = qdisc_dev(sch);
1053 struct tc_htb_qopt_offload offload_opt;
1054 struct htb_sched *q = qdisc_priv(sch);
1055 struct nlattr *tb[TCA_HTB_MAX + 1];
1056 struct tc_htb_glob *gopt;
1061 qdisc_watchdog_init(&q->watchdog, sch);
1062 INIT_WORK(&q->work, htb_work_func);
1067 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
1071 err = nla_parse_nested_deprecated(tb, TCA_HTB_MAX, opt, htb_policy,
1076 if (!tb[TCA_HTB_INIT])
1079 gopt = nla_data(tb[TCA_HTB_INIT]);
1080 if (gopt->version != HTB_VER >> 16)
1083 offload = nla_get_flag(tb[TCA_HTB_OFFLOAD]);
1086 if (sch->parent != TC_H_ROOT)
1089 if (!tc_can_offload(dev) || !dev->netdev_ops->ndo_setup_tc)
1092 q->num_direct_qdiscs = dev->real_num_tx_queues;
1093 q->direct_qdiscs = kcalloc(q->num_direct_qdiscs,
1094 sizeof(*q->direct_qdiscs),
1096 if (!q->direct_qdiscs)
1100 err = qdisc_class_hash_init(&q->clhash);
1102 goto err_free_direct_qdiscs;
1104 qdisc_skb_head_init(&q->direct_queue);
1106 if (tb[TCA_HTB_DIRECT_QLEN])
1107 q->direct_qlen = nla_get_u32(tb[TCA_HTB_DIRECT_QLEN]);
1109 q->direct_qlen = qdisc_dev(sch)->tx_queue_len;
1111 if ((q->rate2quantum = gopt->rate2quantum) < 1)
1112 q->rate2quantum = 1;
1113 q->defcls = gopt->defcls;
1118 for (ntx = 0; ntx < q->num_direct_qdiscs; ntx++) {
1119 struct netdev_queue *dev_queue = netdev_get_tx_queue(dev, ntx);
1120 struct Qdisc *qdisc;
1122 qdisc = qdisc_create_dflt(dev_queue, &pfifo_qdisc_ops,
1123 TC_H_MAKE(sch->handle, 0), extack);
1126 goto err_free_qdiscs;
1129 htb_set_lockdep_class_child(qdisc);
1130 q->direct_qdiscs[ntx] = qdisc;
1131 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1134 sch->flags |= TCQ_F_MQROOT;
1136 offload_opt = (struct tc_htb_qopt_offload) {
1137 .command = TC_HTB_CREATE,
1138 .parent_classid = TC_H_MAJ(sch->handle) >> 16,
1139 .classid = TC_H_MIN(q->defcls),
1142 err = htb_offload(dev, &offload_opt);
1144 goto err_free_qdiscs;
1146 /* Defer this assignment, so that htb_destroy skips offload-related
1147 * parts (especially calling ndo_setup_tc) on errors.
1154 for (ntx = 0; ntx < q->num_direct_qdiscs && q->direct_qdiscs[ntx];
1156 qdisc_put(q->direct_qdiscs[ntx]);
1158 qdisc_class_hash_destroy(&q->clhash);
1159 /* Prevent use-after-free and double-free when htb_destroy gets called.
1161 q->clhash.hash = NULL;
1162 q->clhash.hashsize = 0;
1164 err_free_direct_qdiscs:
1165 kfree(q->direct_qdiscs);
1166 q->direct_qdiscs = NULL;
1170 static void htb_attach_offload(struct Qdisc *sch)
1172 struct net_device *dev = qdisc_dev(sch);
1173 struct htb_sched *q = qdisc_priv(sch);
1176 for (ntx = 0; ntx < q->num_direct_qdiscs; ntx++) {
1177 struct Qdisc *old, *qdisc = q->direct_qdiscs[ntx];
1179 old = dev_graft_qdisc(qdisc->dev_queue, qdisc);
1181 qdisc_hash_add(qdisc, false);
1183 for (ntx = q->num_direct_qdiscs; ntx < dev->num_tx_queues; ntx++) {
1184 struct netdev_queue *dev_queue = netdev_get_tx_queue(dev, ntx);
1185 struct Qdisc *old = dev_graft_qdisc(dev_queue, NULL);
1190 kfree(q->direct_qdiscs);
1191 q->direct_qdiscs = NULL;
1194 static void htb_attach_software(struct Qdisc *sch)
1196 struct net_device *dev = qdisc_dev(sch);
1199 /* Resemble qdisc_graft behavior. */
1200 for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
1201 struct netdev_queue *dev_queue = netdev_get_tx_queue(dev, ntx);
1202 struct Qdisc *old = dev_graft_qdisc(dev_queue, sch);
1204 qdisc_refcount_inc(sch);
1210 static void htb_attach(struct Qdisc *sch)
1212 struct htb_sched *q = qdisc_priv(sch);
1215 htb_attach_offload(sch);
1217 htb_attach_software(sch);
1220 static int htb_dump(struct Qdisc *sch, struct sk_buff *skb)
1222 struct htb_sched *q = qdisc_priv(sch);
1223 struct nlattr *nest;
1224 struct tc_htb_glob gopt;
1227 sch->flags |= TCQ_F_OFFLOADED;
1229 sch->flags &= ~TCQ_F_OFFLOADED;
1231 sch->qstats.overlimits = q->overlimits;
1232 /* Its safe to not acquire qdisc lock. As we hold RTNL,
1233 * no change can happen on the qdisc parameters.
1236 gopt.direct_pkts = q->direct_pkts;
1237 gopt.version = HTB_VER;
1238 gopt.rate2quantum = q->rate2quantum;
1239 gopt.defcls = q->defcls;
1242 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1244 goto nla_put_failure;
1245 if (nla_put(skb, TCA_HTB_INIT, sizeof(gopt), &gopt) ||
1246 nla_put_u32(skb, TCA_HTB_DIRECT_QLEN, q->direct_qlen))
1247 goto nla_put_failure;
1248 if (q->offload && nla_put_flag(skb, TCA_HTB_OFFLOAD))
1249 goto nla_put_failure;
1251 return nla_nest_end(skb, nest);
1254 nla_nest_cancel(skb, nest);
1258 static int htb_dump_class(struct Qdisc *sch, unsigned long arg,
1259 struct sk_buff *skb, struct tcmsg *tcm)
1261 struct htb_class *cl = (struct htb_class *)arg;
1262 struct htb_sched *q = qdisc_priv(sch);
1263 struct nlattr *nest;
1264 struct tc_htb_opt opt;
1266 /* Its safe to not acquire qdisc lock. As we hold RTNL,
1267 * no change can happen on the class parameters.
1269 tcm->tcm_parent = cl->parent ? cl->parent->common.classid : TC_H_ROOT;
1270 tcm->tcm_handle = cl->common.classid;
1271 if (!cl->level && cl->leaf.q)
1272 tcm->tcm_info = cl->leaf.q->handle;
1274 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1276 goto nla_put_failure;
1278 memset(&opt, 0, sizeof(opt));
1280 psched_ratecfg_getrate(&opt.rate, &cl->rate);
1281 opt.buffer = PSCHED_NS2TICKS(cl->buffer);
1282 psched_ratecfg_getrate(&opt.ceil, &cl->ceil);
1283 opt.cbuffer = PSCHED_NS2TICKS(cl->cbuffer);
1284 opt.quantum = cl->quantum;
1285 opt.prio = cl->prio;
1286 opt.level = cl->level;
1287 if (nla_put(skb, TCA_HTB_PARMS, sizeof(opt), &opt))
1288 goto nla_put_failure;
1289 if (q->offload && nla_put_flag(skb, TCA_HTB_OFFLOAD))
1290 goto nla_put_failure;
1291 if ((cl->rate.rate_bytes_ps >= (1ULL << 32)) &&
1292 nla_put_u64_64bit(skb, TCA_HTB_RATE64, cl->rate.rate_bytes_ps,
1294 goto nla_put_failure;
1295 if ((cl->ceil.rate_bytes_ps >= (1ULL << 32)) &&
1296 nla_put_u64_64bit(skb, TCA_HTB_CEIL64, cl->ceil.rate_bytes_ps,
1298 goto nla_put_failure;
1300 return nla_nest_end(skb, nest);
1303 nla_nest_cancel(skb, nest);
1307 static void htb_offload_aggregate_stats(struct htb_sched *q,
1308 struct htb_class *cl)
1310 struct htb_class *c;
1313 memset(&cl->bstats, 0, sizeof(cl->bstats));
1315 for (i = 0; i < q->clhash.hashsize; i++) {
1316 hlist_for_each_entry(c, &q->clhash.hash[i], common.hnode) {
1317 struct htb_class *p = c;
1319 while (p && p->level < cl->level)
1325 cl->bstats.bytes += c->bstats_bias.bytes;
1326 cl->bstats.packets += c->bstats_bias.packets;
1327 if (c->level == 0) {
1328 cl->bstats.bytes += c->leaf.q->bstats.bytes;
1329 cl->bstats.packets += c->leaf.q->bstats.packets;
1336 htb_dump_class_stats(struct Qdisc *sch, unsigned long arg, struct gnet_dump *d)
1338 struct htb_class *cl = (struct htb_class *)arg;
1339 struct htb_sched *q = qdisc_priv(sch);
1340 struct gnet_stats_queue qs = {
1342 .overlimits = cl->overlimits,
1346 if (!cl->level && cl->leaf.q)
1347 qdisc_qstats_qlen_backlog(cl->leaf.q, &qlen, &qs.backlog);
1349 cl->xstats.tokens = clamp_t(s64, PSCHED_NS2TICKS(cl->tokens),
1351 cl->xstats.ctokens = clamp_t(s64, PSCHED_NS2TICKS(cl->ctokens),
1357 cl->bstats = cl->leaf.q->bstats;
1359 memset(&cl->bstats, 0, sizeof(cl->bstats));
1360 cl->bstats.bytes += cl->bstats_bias.bytes;
1361 cl->bstats.packets += cl->bstats_bias.packets;
1363 htb_offload_aggregate_stats(q, cl);
1367 if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
1368 d, NULL, &cl->bstats) < 0 ||
1369 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1370 gnet_stats_copy_queue(d, NULL, &qs, qlen) < 0)
1373 return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1376 static struct netdev_queue *
1377 htb_select_queue(struct Qdisc *sch, struct tcmsg *tcm)
1379 struct net_device *dev = qdisc_dev(sch);
1380 struct tc_htb_qopt_offload offload_opt;
1381 struct htb_sched *q = qdisc_priv(sch);
1385 return sch->dev_queue;
1387 offload_opt = (struct tc_htb_qopt_offload) {
1388 .command = TC_HTB_LEAF_QUERY_QUEUE,
1389 .classid = TC_H_MIN(tcm->tcm_parent),
1391 err = htb_offload(dev, &offload_opt);
1392 if (err || offload_opt.qid >= dev->num_tx_queues)
1394 return netdev_get_tx_queue(dev, offload_opt.qid);
1397 static struct Qdisc *
1398 htb_graft_helper(struct netdev_queue *dev_queue, struct Qdisc *new_q)
1400 struct net_device *dev = dev_queue->dev;
1401 struct Qdisc *old_q;
1403 if (dev->flags & IFF_UP)
1404 dev_deactivate(dev);
1405 old_q = dev_graft_qdisc(dev_queue, new_q);
1407 new_q->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1408 if (dev->flags & IFF_UP)
1414 static void htb_offload_move_qdisc(struct Qdisc *sch, u16 qid_old, u16 qid_new)
1416 struct netdev_queue *queue_old, *queue_new;
1417 struct net_device *dev = qdisc_dev(sch);
1418 struct Qdisc *qdisc;
1420 queue_old = netdev_get_tx_queue(dev, qid_old);
1421 queue_new = netdev_get_tx_queue(dev, qid_new);
1423 if (dev->flags & IFF_UP)
1424 dev_deactivate(dev);
1425 qdisc = dev_graft_qdisc(queue_old, NULL);
1426 qdisc->dev_queue = queue_new;
1427 qdisc = dev_graft_qdisc(queue_new, qdisc);
1428 if (dev->flags & IFF_UP)
1431 WARN_ON(!(qdisc->flags & TCQ_F_BUILTIN));
1434 static int htb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1435 struct Qdisc **old, struct netlink_ext_ack *extack)
1437 struct netdev_queue *dev_queue = sch->dev_queue;
1438 struct htb_class *cl = (struct htb_class *)arg;
1439 struct htb_sched *q = qdisc_priv(sch);
1440 struct Qdisc *old_q;
1446 dev_queue = new->dev_queue;
1447 WARN_ON(dev_queue != cl->leaf.q->dev_queue);
1451 new = qdisc_create_dflt(dev_queue, &pfifo_qdisc_ops,
1452 cl->common.classid, extack);
1458 htb_set_lockdep_class_child(new);
1459 /* One ref for cl->leaf.q, the other for dev_queue->qdisc. */
1460 qdisc_refcount_inc(new);
1461 old_q = htb_graft_helper(dev_queue, new);
1464 *old = qdisc_replace(sch, new, &cl->leaf.q);
1467 WARN_ON(old_q != *old);
1474 static struct Qdisc *htb_leaf(struct Qdisc *sch, unsigned long arg)
1476 struct htb_class *cl = (struct htb_class *)arg;
1477 return !cl->level ? cl->leaf.q : NULL;
1480 static void htb_qlen_notify(struct Qdisc *sch, unsigned long arg)
1482 struct htb_class *cl = (struct htb_class *)arg;
1484 htb_deactivate(qdisc_priv(sch), cl);
1487 static inline int htb_parent_last_child(struct htb_class *cl)
1490 /* the root class */
1492 if (cl->parent->children > 1)
1493 /* not the last child */
1498 static void htb_parent_to_leaf(struct Qdisc *sch, struct htb_class *cl,
1499 struct Qdisc *new_q)
1501 struct htb_sched *q = qdisc_priv(sch);
1502 struct htb_class *parent = cl->parent;
1504 WARN_ON(cl->level || !cl->leaf.q || cl->prio_activity);
1506 if (parent->cmode != HTB_CAN_SEND)
1507 htb_safe_rb_erase(&parent->pq_node,
1508 &q->hlevel[parent->level].wait_pq);
1511 memset(&parent->inner, 0, sizeof(parent->inner));
1512 parent->leaf.q = new_q ? new_q : &noop_qdisc;
1513 parent->tokens = parent->buffer;
1514 parent->ctokens = parent->cbuffer;
1515 parent->t_c = ktime_get_ns();
1516 parent->cmode = HTB_CAN_SEND;
1519 static void htb_parent_to_leaf_offload(struct Qdisc *sch,
1520 struct netdev_queue *dev_queue,
1521 struct Qdisc *new_q)
1523 struct Qdisc *old_q;
1525 /* One ref for cl->leaf.q, the other for dev_queue->qdisc. */
1527 qdisc_refcount_inc(new_q);
1528 old_q = htb_graft_helper(dev_queue, new_q);
1529 WARN_ON(!(old_q->flags & TCQ_F_BUILTIN));
1532 static int htb_destroy_class_offload(struct Qdisc *sch, struct htb_class *cl,
1533 bool last_child, bool destroying,
1534 struct netlink_ext_ack *extack)
1536 struct tc_htb_qopt_offload offload_opt;
1537 struct Qdisc *q = cl->leaf.q;
1538 struct Qdisc *old = NULL;
1546 /* On destroy of HTB, two cases are possible:
1547 * 1. q is a normal qdisc, but q->dev_queue has noop qdisc.
1548 * 2. q is a noop qdisc (for nodes that were inner),
1549 * q->dev_queue is noop_netdev_queue.
1551 old = htb_graft_helper(q->dev_queue, NULL);
1557 cl->parent->bstats_bias.bytes += q->bstats.bytes;
1558 cl->parent->bstats_bias.packets += q->bstats.packets;
1561 offload_opt = (struct tc_htb_qopt_offload) {
1562 .command = !last_child ? TC_HTB_LEAF_DEL :
1563 destroying ? TC_HTB_LEAF_DEL_LAST_FORCE :
1564 TC_HTB_LEAF_DEL_LAST,
1565 .classid = cl->common.classid,
1568 err = htb_offload(qdisc_dev(sch), &offload_opt);
1570 if (!err || destroying)
1573 htb_graft_helper(q->dev_queue, old);
1578 if (!err && offload_opt.moved_qid != 0) {
1580 q->dev_queue = netdev_get_tx_queue(qdisc_dev(sch),
1583 htb_offload_move_qdisc(sch, offload_opt.moved_qid,
1590 static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
1593 WARN_ON(!cl->leaf.q);
1594 qdisc_put(cl->leaf.q);
1596 gen_kill_estimator(&cl->rate_est);
1597 tcf_block_put(cl->block);
1601 static void htb_destroy(struct Qdisc *sch)
1603 struct net_device *dev = qdisc_dev(sch);
1604 struct tc_htb_qopt_offload offload_opt;
1605 struct htb_sched *q = qdisc_priv(sch);
1606 struct hlist_node *next;
1607 bool nonempty, changed;
1608 struct htb_class *cl;
1611 cancel_work_sync(&q->work);
1612 qdisc_watchdog_cancel(&q->watchdog);
1613 /* This line used to be after htb_destroy_class call below
1614 * and surprisingly it worked in 2.4. But it must precede it
1615 * because filter need its target class alive to be able to call
1616 * unbind_filter on it (without Oops).
1618 tcf_block_put(q->block);
1620 for (i = 0; i < q->clhash.hashsize; i++) {
1621 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1622 tcf_block_put(cl->block);
1630 for (i = 0; i < q->clhash.hashsize; i++) {
1631 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1636 htb_destroy_class(sch, cl);
1647 last_child = htb_parent_last_child(cl);
1648 htb_destroy_class_offload(sch, cl, last_child,
1650 qdisc_class_hash_remove(&q->clhash,
1653 cl->parent->children--;
1655 htb_parent_to_leaf(sch, cl, NULL);
1656 htb_destroy_class(sch, cl);
1662 qdisc_class_hash_destroy(&q->clhash);
1663 __qdisc_reset_queue(&q->direct_queue);
1668 offload_opt = (struct tc_htb_qopt_offload) {
1669 .command = TC_HTB_DESTROY,
1671 htb_offload(dev, &offload_opt);
1673 if (!q->direct_qdiscs)
1675 for (i = 0; i < q->num_direct_qdiscs && q->direct_qdiscs[i]; i++)
1676 qdisc_put(q->direct_qdiscs[i]);
1677 kfree(q->direct_qdiscs);
1680 static int htb_delete(struct Qdisc *sch, unsigned long arg,
1681 struct netlink_ext_ack *extack)
1683 struct htb_sched *q = qdisc_priv(sch);
1684 struct htb_class *cl = (struct htb_class *)arg;
1685 struct Qdisc *new_q = NULL;
1689 /* TODO: why don't allow to delete subtree ? references ? does
1690 * tc subsys guarantee us that in htb_destroy it holds no class
1691 * refs so that we can remove children safely there ?
1693 if (cl->children || cl->filter_cnt)
1696 if (!cl->level && htb_parent_last_child(cl))
1700 err = htb_destroy_class_offload(sch, cl, last_child, false,
1707 struct netdev_queue *dev_queue;
1709 dev_queue = q->offload ? cl->leaf.q->dev_queue : sch->dev_queue;
1710 new_q = qdisc_create_dflt(dev_queue, &pfifo_qdisc_ops,
1711 cl->parent->common.classid,
1715 htb_set_lockdep_class_child(new_q);
1716 htb_parent_to_leaf_offload(sch, dev_queue, new_q);
1723 qdisc_purge_queue(cl->leaf.q);
1725 /* delete from hash and active; remainder in destroy_class */
1726 qdisc_class_hash_remove(&q->clhash, &cl->common);
1728 cl->parent->children--;
1730 if (cl->prio_activity)
1731 htb_deactivate(q, cl);
1733 if (cl->cmode != HTB_CAN_SEND)
1734 htb_safe_rb_erase(&cl->pq_node,
1735 &q->hlevel[cl->level].wait_pq);
1738 htb_parent_to_leaf(sch, cl, new_q);
1740 sch_tree_unlock(sch);
1742 htb_destroy_class(sch, cl);
1746 static int htb_change_class(struct Qdisc *sch, u32 classid,
1747 u32 parentid, struct nlattr **tca,
1748 unsigned long *arg, struct netlink_ext_ack *extack)
1751 struct htb_sched *q = qdisc_priv(sch);
1752 struct htb_class *cl = (struct htb_class *)*arg, *parent;
1753 struct tc_htb_qopt_offload offload_opt;
1754 struct nlattr *opt = tca[TCA_OPTIONS];
1755 struct nlattr *tb[TCA_HTB_MAX + 1];
1756 struct Qdisc *parent_qdisc = NULL;
1757 struct netdev_queue *dev_queue;
1758 struct tc_htb_opt *hopt;
1762 /* extract all subattrs from opt attr */
1766 err = nla_parse_nested_deprecated(tb, TCA_HTB_MAX, opt, htb_policy,
1772 if (tb[TCA_HTB_PARMS] == NULL)
1775 parent = parentid == TC_H_ROOT ? NULL : htb_find(parentid, sch);
1777 hopt = nla_data(tb[TCA_HTB_PARMS]);
1778 if (!hopt->rate.rate || !hopt->ceil.rate)
1781 /* Keeping backward compatible with rate_table based iproute2 tc */
1782 if (hopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
1783 qdisc_put_rtab(qdisc_get_rtab(&hopt->rate, tb[TCA_HTB_RTAB],
1786 if (hopt->ceil.linklayer == TC_LINKLAYER_UNAWARE)
1787 qdisc_put_rtab(qdisc_get_rtab(&hopt->ceil, tb[TCA_HTB_CTAB],
1790 rate64 = tb[TCA_HTB_RATE64] ? nla_get_u64(tb[TCA_HTB_RATE64]) : 0;
1791 ceil64 = tb[TCA_HTB_CEIL64] ? nla_get_u64(tb[TCA_HTB_CEIL64]) : 0;
1793 if (!cl) { /* new class */
1794 struct net_device *dev = qdisc_dev(sch);
1795 struct Qdisc *new_q, *old_q;
1799 struct gnet_estimator opt;
1802 .nla_len = nla_attr_size(sizeof(est.opt)),
1803 .nla_type = TCA_RATE,
1806 /* 4s interval, 16s averaging constant */
1812 /* check for valid classid */
1813 if (!classid || TC_H_MAJ(classid ^ sch->handle) ||
1814 htb_find(classid, sch))
1817 /* check maximal depth */
1818 if (parent && parent->parent && parent->parent->level < 2) {
1819 pr_err("htb: tree is too deep\n");
1823 cl = kzalloc(sizeof(*cl), GFP_KERNEL);
1827 err = tcf_block_get(&cl->block, &cl->filter_list, sch, extack);
1832 if (htb_rate_est || tca[TCA_RATE]) {
1833 err = gen_new_estimator(&cl->bstats, NULL,
1836 qdisc_root_sleeping_running(sch),
1837 tca[TCA_RATE] ? : &est.nla);
1843 RB_CLEAR_NODE(&cl->pq_node);
1845 for (prio = 0; prio < TC_HTB_NUMPRIO; prio++)
1846 RB_CLEAR_NODE(&cl->node[prio]);
1848 cl->common.classid = classid;
1850 /* Make sure nothing interrupts us in between of two
1851 * ndo_setup_tc calls.
1855 /* create leaf qdisc early because it uses kmalloc(GFP_KERNEL)
1856 * so that can't be used inside of sch_tree_lock
1857 * -- thanks to Karlis Peisenieks
1860 dev_queue = sch->dev_queue;
1861 } else if (!(parent && !parent->level)) {
1862 /* Assign a dev_queue to this classid. */
1863 offload_opt = (struct tc_htb_qopt_offload) {
1864 .command = TC_HTB_LEAF_ALLOC_QUEUE,
1865 .classid = cl->common.classid,
1866 .parent_classid = parent ?
1867 TC_H_MIN(parent->common.classid) :
1868 TC_HTB_CLASSID_ROOT,
1869 .rate = max_t(u64, hopt->rate.rate, rate64),
1870 .ceil = max_t(u64, hopt->ceil.rate, ceil64),
1873 err = htb_offload(dev, &offload_opt);
1875 pr_err("htb: TC_HTB_LEAF_ALLOC_QUEUE failed with err = %d\n",
1877 goto err_kill_estimator;
1879 dev_queue = netdev_get_tx_queue(dev, offload_opt.qid);
1880 } else { /* First child. */
1881 dev_queue = parent->leaf.q->dev_queue;
1882 old_q = htb_graft_helper(dev_queue, NULL);
1883 WARN_ON(old_q != parent->leaf.q);
1884 offload_opt = (struct tc_htb_qopt_offload) {
1885 .command = TC_HTB_LEAF_TO_INNER,
1886 .classid = cl->common.classid,
1888 TC_H_MIN(parent->common.classid),
1889 .rate = max_t(u64, hopt->rate.rate, rate64),
1890 .ceil = max_t(u64, hopt->ceil.rate, ceil64),
1893 err = htb_offload(dev, &offload_opt);
1895 pr_err("htb: TC_HTB_LEAF_TO_INNER failed with err = %d\n",
1897 htb_graft_helper(dev_queue, old_q);
1898 goto err_kill_estimator;
1900 parent->bstats_bias.bytes += old_q->bstats.bytes;
1901 parent->bstats_bias.packets += old_q->bstats.packets;
1904 new_q = qdisc_create_dflt(dev_queue, &pfifo_qdisc_ops,
1908 htb_set_lockdep_class_child(new_q);
1909 /* One ref for cl->leaf.q, the other for
1912 qdisc_refcount_inc(new_q);
1914 old_q = htb_graft_helper(dev_queue, new_q);
1915 /* No qdisc_put needed. */
1916 WARN_ON(!(old_q->flags & TCQ_F_BUILTIN));
1919 if (parent && !parent->level) {
1920 /* turn parent into inner node */
1921 qdisc_purge_queue(parent->leaf.q);
1922 parent_qdisc = parent->leaf.q;
1923 if (parent->prio_activity)
1924 htb_deactivate(q, parent);
1926 /* remove from evt list because of level change */
1927 if (parent->cmode != HTB_CAN_SEND) {
1928 htb_safe_rb_erase(&parent->pq_node, &q->hlevel[0].wait_pq);
1929 parent->cmode = HTB_CAN_SEND;
1931 parent->level = (parent->parent ? parent->parent->level
1932 : TC_HTB_MAXDEPTH) - 1;
1933 memset(&parent->inner, 0, sizeof(parent->inner));
1936 /* leaf (we) needs elementary qdisc */
1937 cl->leaf.q = new_q ? new_q : &noop_qdisc;
1939 cl->parent = parent;
1941 /* set class to be in HTB_CAN_SEND state */
1942 cl->tokens = PSCHED_TICKS2NS(hopt->buffer);
1943 cl->ctokens = PSCHED_TICKS2NS(hopt->cbuffer);
1944 cl->mbuffer = 60ULL * NSEC_PER_SEC; /* 1min */
1945 cl->t_c = ktime_get_ns();
1946 cl->cmode = HTB_CAN_SEND;
1948 /* attach to the hash list and parent's family */
1949 qdisc_class_hash_insert(&q->clhash, &cl->common);
1952 if (cl->leaf.q != &noop_qdisc)
1953 qdisc_hash_add(cl->leaf.q, true);
1955 if (tca[TCA_RATE]) {
1956 err = gen_replace_estimator(&cl->bstats, NULL,
1959 qdisc_root_sleeping_running(sch),
1966 struct net_device *dev = qdisc_dev(sch);
1968 offload_opt = (struct tc_htb_qopt_offload) {
1969 .command = TC_HTB_NODE_MODIFY,
1970 .classid = cl->common.classid,
1971 .rate = max_t(u64, hopt->rate.rate, rate64),
1972 .ceil = max_t(u64, hopt->ceil.rate, ceil64),
1975 err = htb_offload(dev, &offload_opt);
1977 /* Estimator was replaced, and rollback may fail
1978 * as well, so we don't try to recover it, and
1979 * the estimator won't work property with the
1980 * offload anyway, because bstats are updated
1981 * only when the stats are queried.
1989 psched_ratecfg_precompute(&cl->rate, &hopt->rate, rate64);
1990 psched_ratecfg_precompute(&cl->ceil, &hopt->ceil, ceil64);
1992 /* it used to be a nasty bug here, we have to check that node
1993 * is really leaf before changing cl->leaf !
1996 u64 quantum = cl->rate.rate_bytes_ps;
1998 do_div(quantum, q->rate2quantum);
1999 cl->quantum = min_t(u64, quantum, INT_MAX);
2001 if (!hopt->quantum && cl->quantum < 1000) {
2005 if (!hopt->quantum && cl->quantum > 200000) {
2007 cl->quantum = 200000;
2010 cl->quantum = hopt->quantum;
2011 if ((cl->prio = hopt->prio) >= TC_HTB_NUMPRIO)
2012 cl->prio = TC_HTB_NUMPRIO - 1;
2015 cl->buffer = PSCHED_TICKS2NS(hopt->buffer);
2016 cl->cbuffer = PSCHED_TICKS2NS(hopt->cbuffer);
2018 sch_tree_unlock(sch);
2019 qdisc_put(parent_qdisc);
2022 pr_warn("HTB: quantum of class %X is %s. Consider r2q change.\n",
2023 cl->common.classid, (warn == -1 ? "small" : "big"));
2025 qdisc_class_hash_grow(sch, &q->clhash);
2027 *arg = (unsigned long)cl;
2031 gen_kill_estimator(&cl->rate_est);
2033 tcf_block_put(cl->block);
2039 static struct tcf_block *htb_tcf_block(struct Qdisc *sch, unsigned long arg,
2040 struct netlink_ext_ack *extack)
2042 struct htb_sched *q = qdisc_priv(sch);
2043 struct htb_class *cl = (struct htb_class *)arg;
2045 return cl ? cl->block : q->block;
2048 static unsigned long htb_bind_filter(struct Qdisc *sch, unsigned long parent,
2051 struct htb_class *cl = htb_find(classid, sch);
2053 /*if (cl && !cl->level) return 0;
2054 * The line above used to be there to prevent attaching filters to
2055 * leaves. But at least tc_index filter uses this just to get class
2056 * for other reasons so that we have to allow for it.
2058 * 19.6.2002 As Werner explained it is ok - bind filter is just
2059 * another way to "lock" the class - unlike "get" this lock can
2060 * be broken by class during destroy IIUC.
2064 return (unsigned long)cl;
2067 static void htb_unbind_filter(struct Qdisc *sch, unsigned long arg)
2069 struct htb_class *cl = (struct htb_class *)arg;
2075 static void htb_walk(struct Qdisc *sch, struct qdisc_walker *arg)
2077 struct htb_sched *q = qdisc_priv(sch);
2078 struct htb_class *cl;
2084 for (i = 0; i < q->clhash.hashsize; i++) {
2085 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
2086 if (arg->count < arg->skip) {
2090 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
2099 static const struct Qdisc_class_ops htb_class_ops = {
2100 .select_queue = htb_select_queue,
2103 .qlen_notify = htb_qlen_notify,
2105 .change = htb_change_class,
2106 .delete = htb_delete,
2108 .tcf_block = htb_tcf_block,
2109 .bind_tcf = htb_bind_filter,
2110 .unbind_tcf = htb_unbind_filter,
2111 .dump = htb_dump_class,
2112 .dump_stats = htb_dump_class_stats,
2115 static struct Qdisc_ops htb_qdisc_ops __read_mostly = {
2116 .cl_ops = &htb_class_ops,
2118 .priv_size = sizeof(struct htb_sched),
2119 .enqueue = htb_enqueue,
2120 .dequeue = htb_dequeue,
2121 .peek = qdisc_peek_dequeued,
2123 .attach = htb_attach,
2125 .destroy = htb_destroy,
2127 .owner = THIS_MODULE,
2130 static int __init htb_module_init(void)
2132 return register_qdisc(&htb_qdisc_ops);
2134 static void __exit htb_module_exit(void)
2136 unregister_qdisc(&htb_qdisc_ops);
2139 module_init(htb_module_init)
2140 module_exit(htb_module_exit)
2141 MODULE_LICENSE("GPL");