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];
128 struct netdev_queue *offload_queue;
130 struct htb_class_inner {
131 struct htb_prio clprio[TC_HTB_NUMPRIO];
136 int prio_activity; /* for which prios are we active */
137 enum htb_cmode cmode; /* current mode of the class */
138 struct rb_node pq_node; /* node for event queue */
139 struct rb_node node[TC_HTB_NUMPRIO]; /* node for self or feed tree */
141 unsigned int drops ____cacheline_aligned_in_smp;
142 unsigned int overlimits;
146 struct rb_root wait_pq;
147 struct htb_prio hprio[TC_HTB_NUMPRIO];
151 struct Qdisc_class_hash clhash;
152 int defcls; /* class where unclassified flows go to */
153 int rate2quantum; /* quant = rate / rate2quantum */
155 /* filters for qdisc itself */
156 struct tcf_proto __rcu *filter_list;
157 struct tcf_block *block;
159 #define HTB_WARN_TOOMANYEVENTS 0x1
160 unsigned int warned; /* only one warning */
162 struct work_struct work;
164 /* non shaped skbs; let them go directly thru */
165 struct qdisc_skb_head direct_queue;
169 struct qdisc_watchdog watchdog;
171 s64 now; /* cached dequeue time */
173 /* time of nearest event per level (row) */
174 s64 near_ev_cache[TC_HTB_MAXDEPTH];
176 int row_mask[TC_HTB_MAXDEPTH];
178 struct htb_level hlevel[TC_HTB_MAXDEPTH];
180 struct Qdisc **direct_qdiscs;
181 unsigned int num_direct_qdiscs;
186 /* find class in global hash table using given handle */
187 static inline struct htb_class *htb_find(u32 handle, struct Qdisc *sch)
189 struct htb_sched *q = qdisc_priv(sch);
190 struct Qdisc_class_common *clc;
192 clc = qdisc_class_find(&q->clhash, handle);
195 return container_of(clc, struct htb_class, common);
198 static unsigned long htb_search(struct Qdisc *sch, u32 handle)
200 return (unsigned long)htb_find(handle, sch);
203 * htb_classify - classify a packet into class
205 * It returns NULL if the packet should be dropped or -1 if the packet
206 * should be passed directly thru. In all other cases leaf class is returned.
207 * We allow direct class selection by classid in priority. The we examine
208 * filters in qdisc and in inner nodes (if higher filter points to the inner
209 * node). If we end up with classid MAJOR:0 we enqueue the skb into special
210 * internal fifo (direct). These packets then go directly thru. If we still
211 * have no valid leaf we try to use MAJOR:default leaf. It still unsuccessful
212 * then finish and return direct queue.
214 #define HTB_DIRECT ((struct htb_class *)-1L)
216 static struct htb_class *htb_classify(struct sk_buff *skb, struct Qdisc *sch,
219 struct htb_sched *q = qdisc_priv(sch);
220 struct htb_class *cl;
221 struct tcf_result res;
222 struct tcf_proto *tcf;
225 /* allow to select class by setting skb->priority to valid classid;
226 * note that nfmark can be used too by attaching filter fw with no
229 if (skb->priority == sch->handle)
230 return HTB_DIRECT; /* X:0 (direct flow) selected */
231 cl = htb_find(skb->priority, sch);
235 /* Start with inner filter chain if a non-leaf class is selected */
236 tcf = rcu_dereference_bh(cl->filter_list);
238 tcf = rcu_dereference_bh(q->filter_list);
241 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
242 while (tcf && (result = tcf_classify(skb, NULL, tcf, &res, false)) >= 0) {
243 #ifdef CONFIG_NET_CLS_ACT
248 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
254 cl = (void *)res.class;
256 if (res.classid == sch->handle)
257 return HTB_DIRECT; /* X:0 (direct flow) */
258 cl = htb_find(res.classid, sch);
260 break; /* filter selected invalid classid */
263 return cl; /* we hit leaf; return it */
265 /* we have got inner class; apply inner filter chain */
266 tcf = rcu_dereference_bh(cl->filter_list);
268 /* classification failed; try to use default class */
269 cl = htb_find(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
270 if (!cl || cl->level)
271 return HTB_DIRECT; /* bad default .. this is safe bet */
276 * htb_add_to_id_tree - adds class to the round robin list
277 * @root: the root of the tree
278 * @cl: the class to add
279 * @prio: the give prio in class
281 * Routine adds class to the list (actually tree) sorted by classid.
282 * Make sure that class is not already on such list for given prio.
284 static void htb_add_to_id_tree(struct rb_root *root,
285 struct htb_class *cl, int prio)
287 struct rb_node **p = &root->rb_node, *parent = NULL;
292 c = rb_entry(parent, struct htb_class, node[prio]);
294 if (cl->common.classid > c->common.classid)
295 p = &parent->rb_right;
297 p = &parent->rb_left;
299 rb_link_node(&cl->node[prio], parent, p);
300 rb_insert_color(&cl->node[prio], root);
304 * htb_add_to_wait_tree - adds class to the event queue with delay
305 * @q: the priority event queue
306 * @cl: the class to add
307 * @delay: delay in microseconds
309 * The class is added to priority event queue to indicate that class will
310 * change its mode in cl->pq_key microseconds. Make sure that class is not
311 * already in the queue.
313 static void htb_add_to_wait_tree(struct htb_sched *q,
314 struct htb_class *cl, s64 delay)
316 struct rb_node **p = &q->hlevel[cl->level].wait_pq.rb_node, *parent = NULL;
318 cl->pq_key = q->now + delay;
319 if (cl->pq_key == q->now)
322 /* update the nearest event cache */
323 if (q->near_ev_cache[cl->level] > cl->pq_key)
324 q->near_ev_cache[cl->level] = cl->pq_key;
329 c = rb_entry(parent, struct htb_class, pq_node);
330 if (cl->pq_key >= c->pq_key)
331 p = &parent->rb_right;
333 p = &parent->rb_left;
335 rb_link_node(&cl->pq_node, parent, p);
336 rb_insert_color(&cl->pq_node, &q->hlevel[cl->level].wait_pq);
340 * htb_next_rb_node - finds next node in binary tree
341 * @n: the current node in binary tree
343 * When we are past last key we return NULL.
344 * Average complexity is 2 steps per call.
346 static inline void htb_next_rb_node(struct rb_node **n)
352 * htb_add_class_to_row - add class to its row
353 * @q: the priority event queue
354 * @cl: the class to add
355 * @mask: the given priorities in class in bitmap
357 * The class is added to row at priorities marked in mask.
358 * It does nothing if mask == 0.
360 static inline void htb_add_class_to_row(struct htb_sched *q,
361 struct htb_class *cl, int mask)
363 q->row_mask[cl->level] |= mask;
365 int prio = ffz(~mask);
366 mask &= ~(1 << prio);
367 htb_add_to_id_tree(&q->hlevel[cl->level].hprio[prio].row, cl, prio);
371 /* If this triggers, it is a bug in this code, but it need not be fatal */
372 static void htb_safe_rb_erase(struct rb_node *rb, struct rb_root *root)
374 if (RB_EMPTY_NODE(rb)) {
384 * htb_remove_class_from_row - removes class from its row
385 * @q: the priority event queue
386 * @cl: the class to add
387 * @mask: the given priorities in class in bitmap
389 * The class is removed from row at priorities marked in mask.
390 * It does nothing if mask == 0.
392 static inline void htb_remove_class_from_row(struct htb_sched *q,
393 struct htb_class *cl, int mask)
396 struct htb_level *hlevel = &q->hlevel[cl->level];
399 int prio = ffz(~mask);
400 struct htb_prio *hprio = &hlevel->hprio[prio];
402 mask &= ~(1 << prio);
403 if (hprio->ptr == cl->node + prio)
404 htb_next_rb_node(&hprio->ptr);
406 htb_safe_rb_erase(cl->node + prio, &hprio->row);
407 if (!hprio->row.rb_node)
410 q->row_mask[cl->level] &= ~m;
414 * htb_activate_prios - creates active classe's feed chain
415 * @q: the priority event queue
416 * @cl: the class to activate
418 * The class is connected to ancestors and/or appropriate rows
419 * for priorities it is participating on. cl->cmode must be new
420 * (activated) mode. It does nothing if cl->prio_activity == 0.
422 static void htb_activate_prios(struct htb_sched *q, struct htb_class *cl)
424 struct htb_class *p = cl->parent;
425 long m, mask = cl->prio_activity;
427 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
433 if (p->inner.clprio[prio].feed.rb_node)
434 /* parent already has its feed in use so that
435 * reset bit in mask as parent is already ok
437 mask &= ~(1 << prio);
439 htb_add_to_id_tree(&p->inner.clprio[prio].feed, cl, prio);
441 p->prio_activity |= mask;
446 if (cl->cmode == HTB_CAN_SEND && mask)
447 htb_add_class_to_row(q, cl, mask);
451 * htb_deactivate_prios - remove class from feed chain
452 * @q: the priority event queue
453 * @cl: the class to deactivate
455 * cl->cmode must represent old mode (before deactivation). It does
456 * nothing if cl->prio_activity == 0. Class is removed from all feed
459 static void htb_deactivate_prios(struct htb_sched *q, struct htb_class *cl)
461 struct htb_class *p = cl->parent;
462 long m, mask = cl->prio_activity;
464 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
471 if (p->inner.clprio[prio].ptr == cl->node + prio) {
472 /* we are removing child which is pointed to from
473 * parent feed - forget the pointer but remember
476 p->inner.clprio[prio].last_ptr_id = cl->common.classid;
477 p->inner.clprio[prio].ptr = NULL;
480 htb_safe_rb_erase(cl->node + prio,
481 &p->inner.clprio[prio].feed);
483 if (!p->inner.clprio[prio].feed.rb_node)
487 p->prio_activity &= ~mask;
492 if (cl->cmode == HTB_CAN_SEND && mask)
493 htb_remove_class_from_row(q, cl, mask);
496 static inline s64 htb_lowater(const struct htb_class *cl)
499 return cl->cmode != HTB_CANT_SEND ? -cl->cbuffer : 0;
503 static inline s64 htb_hiwater(const struct htb_class *cl)
506 return cl->cmode == HTB_CAN_SEND ? -cl->buffer : 0;
513 * htb_class_mode - computes and returns current class mode
514 * @cl: the target class
515 * @diff: diff time in microseconds
517 * It computes cl's mode at time cl->t_c+diff and returns it. If mode
518 * is not HTB_CAN_SEND then cl->pq_key is updated to time difference
519 * from now to time when cl will change its state.
520 * Also it is worth to note that class mode doesn't change simply
521 * at cl->{c,}tokens == 0 but there can rather be hysteresis of
522 * 0 .. -cl->{c,}buffer range. It is meant to limit number of
523 * mode transitions per time unit. The speed gain is about 1/6.
525 static inline enum htb_cmode
526 htb_class_mode(struct htb_class *cl, s64 *diff)
530 if ((toks = (cl->ctokens + *diff)) < htb_lowater(cl)) {
532 return HTB_CANT_SEND;
535 if ((toks = (cl->tokens + *diff)) >= htb_hiwater(cl))
539 return HTB_MAY_BORROW;
543 * htb_change_class_mode - changes classe's mode
544 * @q: the priority event queue
545 * @cl: the target class
546 * @diff: diff time in microseconds
548 * This should be the only way how to change classe's mode under normal
549 * circumstances. Routine will update feed lists linkage, change mode
550 * and add class to the wait event queue if appropriate. New mode should
551 * be different from old one and cl->pq_key has to be valid if changing
552 * to mode other than HTB_CAN_SEND (see htb_add_to_wait_tree).
555 htb_change_class_mode(struct htb_sched *q, struct htb_class *cl, s64 *diff)
557 enum htb_cmode new_mode = htb_class_mode(cl, diff);
559 if (new_mode == cl->cmode)
562 if (new_mode == HTB_CANT_SEND) {
567 if (cl->prio_activity) { /* not necessary: speed optimization */
568 if (cl->cmode != HTB_CANT_SEND)
569 htb_deactivate_prios(q, cl);
570 cl->cmode = new_mode;
571 if (new_mode != HTB_CANT_SEND)
572 htb_activate_prios(q, cl);
574 cl->cmode = new_mode;
578 * htb_activate - inserts leaf cl into appropriate active feeds
579 * @q: the priority event queue
580 * @cl: the target class
582 * Routine learns (new) priority of leaf and activates feed chain
583 * for the prio. It can be called on already active leaf safely.
584 * It also adds leaf into droplist.
586 static inline void htb_activate(struct htb_sched *q, struct htb_class *cl)
588 WARN_ON(cl->level || !cl->leaf.q || !cl->leaf.q->q.qlen);
590 if (!cl->prio_activity) {
591 cl->prio_activity = 1 << cl->prio;
592 htb_activate_prios(q, cl);
597 * htb_deactivate - remove leaf cl from active feeds
598 * @q: the priority event queue
599 * @cl: the target class
601 * Make sure that leaf is active. In the other words it can't be called
602 * with non-active leaf. It also removes class from the drop list.
604 static inline void htb_deactivate(struct htb_sched *q, struct htb_class *cl)
606 WARN_ON(!cl->prio_activity);
608 htb_deactivate_prios(q, cl);
609 cl->prio_activity = 0;
612 static int htb_enqueue(struct sk_buff *skb, struct Qdisc *sch,
613 struct sk_buff **to_free)
616 unsigned int len = qdisc_pkt_len(skb);
617 struct htb_sched *q = qdisc_priv(sch);
618 struct htb_class *cl = htb_classify(skb, sch, &ret);
620 if (cl == HTB_DIRECT) {
621 /* enqueue to helper queue */
622 if (q->direct_queue.qlen < q->direct_qlen) {
623 __qdisc_enqueue_tail(skb, &q->direct_queue);
626 return qdisc_drop(skb, sch, to_free);
628 #ifdef CONFIG_NET_CLS_ACT
630 if (ret & __NET_XMIT_BYPASS)
631 qdisc_qstats_drop(sch);
632 __qdisc_drop(skb, to_free);
635 } else if ((ret = qdisc_enqueue(skb, cl->leaf.q,
636 to_free)) != NET_XMIT_SUCCESS) {
637 if (net_xmit_drop_count(ret)) {
638 qdisc_qstats_drop(sch);
646 sch->qstats.backlog += len;
648 return NET_XMIT_SUCCESS;
651 static inline void htb_accnt_tokens(struct htb_class *cl, int bytes, s64 diff)
653 s64 toks = diff + cl->tokens;
655 if (toks > cl->buffer)
657 toks -= (s64) psched_l2t_ns(&cl->rate, bytes);
658 if (toks <= -cl->mbuffer)
659 toks = 1 - cl->mbuffer;
664 static inline void htb_accnt_ctokens(struct htb_class *cl, int bytes, s64 diff)
666 s64 toks = diff + cl->ctokens;
668 if (toks > cl->cbuffer)
670 toks -= (s64) psched_l2t_ns(&cl->ceil, bytes);
671 if (toks <= -cl->mbuffer)
672 toks = 1 - cl->mbuffer;
678 * htb_charge_class - charges amount "bytes" to leaf and ancestors
679 * @q: the priority event queue
680 * @cl: the class to start iterate
681 * @level: the minimum level to account
682 * @skb: the socket buffer
684 * Routine assumes that packet "bytes" long was dequeued from leaf cl
685 * borrowing from "level". It accounts bytes to ceil leaky bucket for
686 * leaf and all ancestors and to rate bucket for ancestors at levels
687 * "level" and higher. It also handles possible change of mode resulting
688 * from the update. Note that mode can also increase here (MAY_BORROW to
689 * CAN_SEND) because we can use more precise clock that event queue here.
690 * In such case we remove class from event queue first.
692 static void htb_charge_class(struct htb_sched *q, struct htb_class *cl,
693 int level, struct sk_buff *skb)
695 int bytes = qdisc_pkt_len(skb);
696 enum htb_cmode old_mode;
700 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
701 if (cl->level >= level) {
702 if (cl->level == level)
704 htb_accnt_tokens(cl, bytes, diff);
706 cl->xstats.borrows++;
707 cl->tokens += diff; /* we moved t_c; update tokens */
709 htb_accnt_ctokens(cl, bytes, diff);
712 old_mode = cl->cmode;
714 htb_change_class_mode(q, cl, &diff);
715 if (old_mode != cl->cmode) {
716 if (old_mode != HTB_CAN_SEND)
717 htb_safe_rb_erase(&cl->pq_node, &q->hlevel[cl->level].wait_pq);
718 if (cl->cmode != HTB_CAN_SEND)
719 htb_add_to_wait_tree(q, cl, diff);
722 /* update basic stats except for leaves which are already updated */
724 bstats_update(&cl->bstats, skb);
731 * htb_do_events - make mode changes to classes at the level
732 * @q: the priority event queue
733 * @level: which wait_pq in 'q->hlevel'
734 * @start: start jiffies
736 * Scans event queue for pending events and applies them. Returns time of
737 * next pending event (0 for no event in pq, q->now for too many events).
738 * Note: Applied are events whose have cl->pq_key <= q->now.
740 static s64 htb_do_events(struct htb_sched *q, const int level,
743 /* don't run for longer than 2 jiffies; 2 is used instead of
744 * 1 to simplify things when jiffy is going to be incremented
747 unsigned long stop_at = start + 2;
748 struct rb_root *wait_pq = &q->hlevel[level].wait_pq;
750 while (time_before(jiffies, stop_at)) {
751 struct htb_class *cl;
753 struct rb_node *p = rb_first(wait_pq);
758 cl = rb_entry(p, struct htb_class, pq_node);
759 if (cl->pq_key > q->now)
762 htb_safe_rb_erase(p, wait_pq);
763 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
764 htb_change_class_mode(q, cl, &diff);
765 if (cl->cmode != HTB_CAN_SEND)
766 htb_add_to_wait_tree(q, cl, diff);
769 /* too much load - let's continue after a break for scheduling */
770 if (!(q->warned & HTB_WARN_TOOMANYEVENTS)) {
771 pr_warn("htb: too many events!\n");
772 q->warned |= HTB_WARN_TOOMANYEVENTS;
778 /* Returns class->node+prio from id-tree where classe's id is >= id. NULL
779 * is no such one exists.
781 static struct rb_node *htb_id_find_next_upper(int prio, struct rb_node *n,
784 struct rb_node *r = NULL;
786 struct htb_class *cl =
787 rb_entry(n, struct htb_class, node[prio]);
789 if (id > cl->common.classid) {
791 } else if (id < cl->common.classid) {
802 * htb_lookup_leaf - returns next leaf class in DRR order
803 * @hprio: the current one
804 * @prio: which prio in class
806 * Find leaf where current feed pointers points to.
808 static struct htb_class *htb_lookup_leaf(struct htb_prio *hprio, const int prio)
812 struct rb_node *root;
813 struct rb_node **pptr;
815 } stk[TC_HTB_MAXDEPTH], *sp = stk;
817 BUG_ON(!hprio->row.rb_node);
818 sp->root = hprio->row.rb_node;
819 sp->pptr = &hprio->ptr;
820 sp->pid = &hprio->last_ptr_id;
822 for (i = 0; i < 65535; i++) {
823 if (!*sp->pptr && *sp->pid) {
824 /* ptr was invalidated but id is valid - try to recover
825 * the original or next ptr
828 htb_id_find_next_upper(prio, sp->root, *sp->pid);
830 *sp->pid = 0; /* ptr is valid now so that remove this hint as it
831 * can become out of date quickly
833 if (!*sp->pptr) { /* we are at right end; rewind & go up */
834 *sp->pptr = sp->root;
835 while ((*sp->pptr)->rb_left)
836 *sp->pptr = (*sp->pptr)->rb_left;
843 htb_next_rb_node(sp->pptr);
846 struct htb_class *cl;
847 struct htb_prio *clp;
849 cl = rb_entry(*sp->pptr, struct htb_class, node[prio]);
852 clp = &cl->inner.clprio[prio];
853 (++sp)->root = clp->feed.rb_node;
854 sp->pptr = &clp->ptr;
855 sp->pid = &clp->last_ptr_id;
862 /* dequeues packet at given priority and level; call only if
863 * you are sure that there is active class at prio/level
865 static struct sk_buff *htb_dequeue_tree(struct htb_sched *q, const int prio,
868 struct sk_buff *skb = NULL;
869 struct htb_class *cl, *start;
870 struct htb_level *hlevel = &q->hlevel[level];
871 struct htb_prio *hprio = &hlevel->hprio[prio];
873 /* look initial class up in the row */
874 start = cl = htb_lookup_leaf(hprio, prio);
881 /* class can be empty - it is unlikely but can be true if leaf
882 * qdisc drops packets in enqueue routine or if someone used
883 * graft operation on the leaf since last dequeue;
884 * simply deactivate and skip such class
886 if (unlikely(cl->leaf.q->q.qlen == 0)) {
887 struct htb_class *next;
888 htb_deactivate(q, cl);
890 /* row/level might become empty */
891 if ((q->row_mask[level] & (1 << prio)) == 0)
894 next = htb_lookup_leaf(hprio, prio);
896 if (cl == start) /* fix start if we just deleted it */
902 skb = cl->leaf.q->dequeue(cl->leaf.q);
903 if (likely(skb != NULL))
906 qdisc_warn_nonwc("htb", cl->leaf.q);
907 htb_next_rb_node(level ? &cl->parent->inner.clprio[prio].ptr:
908 &q->hlevel[0].hprio[prio].ptr);
909 cl = htb_lookup_leaf(hprio, prio);
911 } while (cl != start);
913 if (likely(skb != NULL)) {
914 bstats_update(&cl->bstats, skb);
915 cl->leaf.deficit[level] -= qdisc_pkt_len(skb);
916 if (cl->leaf.deficit[level] < 0) {
917 cl->leaf.deficit[level] += cl->quantum;
918 htb_next_rb_node(level ? &cl->parent->inner.clprio[prio].ptr :
919 &q->hlevel[0].hprio[prio].ptr);
921 /* this used to be after charge_class but this constelation
922 * gives us slightly better performance
924 if (!cl->leaf.q->q.qlen)
925 htb_deactivate(q, cl);
926 htb_charge_class(q, cl, level, skb);
931 static struct sk_buff *htb_dequeue(struct Qdisc *sch)
934 struct htb_sched *q = qdisc_priv(sch);
937 unsigned long start_at;
939 /* try to dequeue direct packets as high prio (!) to minimize cpu work */
940 skb = __qdisc_dequeue_head(&q->direct_queue);
943 qdisc_bstats_update(sch, skb);
944 qdisc_qstats_backlog_dec(sch, skb);
951 q->now = ktime_get_ns();
954 next_event = q->now + 5LLU * NSEC_PER_SEC;
956 for (level = 0; level < TC_HTB_MAXDEPTH; level++) {
957 /* common case optimization - skip event handler quickly */
959 s64 event = q->near_ev_cache[level];
961 if (q->now >= event) {
962 event = htb_do_events(q, level, start_at);
964 event = q->now + NSEC_PER_SEC;
965 q->near_ev_cache[level] = event;
968 if (next_event > event)
971 m = ~q->row_mask[level];
972 while (m != (int)(-1)) {
976 skb = htb_dequeue_tree(q, prio, level);
977 if (likely(skb != NULL))
981 if (likely(next_event > q->now))
982 qdisc_watchdog_schedule_ns(&q->watchdog, next_event);
984 schedule_work(&q->work);
989 /* reset all classes */
990 /* always caled under BH & queue lock */
991 static void htb_reset(struct Qdisc *sch)
993 struct htb_sched *q = qdisc_priv(sch);
994 struct htb_class *cl;
997 for (i = 0; i < q->clhash.hashsize; i++) {
998 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1000 memset(&cl->inner, 0, sizeof(cl->inner));
1002 if (cl->leaf.q && !q->offload)
1003 qdisc_reset(cl->leaf.q);
1005 cl->prio_activity = 0;
1006 cl->cmode = HTB_CAN_SEND;
1009 qdisc_watchdog_cancel(&q->watchdog);
1010 __qdisc_reset_queue(&q->direct_queue);
1011 memset(q->hlevel, 0, sizeof(q->hlevel));
1012 memset(q->row_mask, 0, sizeof(q->row_mask));
1015 static const struct nla_policy htb_policy[TCA_HTB_MAX + 1] = {
1016 [TCA_HTB_PARMS] = { .len = sizeof(struct tc_htb_opt) },
1017 [TCA_HTB_INIT] = { .len = sizeof(struct tc_htb_glob) },
1018 [TCA_HTB_CTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
1019 [TCA_HTB_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
1020 [TCA_HTB_DIRECT_QLEN] = { .type = NLA_U32 },
1021 [TCA_HTB_RATE64] = { .type = NLA_U64 },
1022 [TCA_HTB_CEIL64] = { .type = NLA_U64 },
1023 [TCA_HTB_OFFLOAD] = { .type = NLA_FLAG },
1026 static void htb_work_func(struct work_struct *work)
1028 struct htb_sched *q = container_of(work, struct htb_sched, work);
1029 struct Qdisc *sch = q->watchdog.qdisc;
1032 __netif_schedule(qdisc_root(sch));
1036 static void htb_set_lockdep_class_child(struct Qdisc *q)
1038 static struct lock_class_key child_key;
1040 lockdep_set_class(qdisc_lock(q), &child_key);
1043 static int htb_offload(struct net_device *dev, struct tc_htb_qopt_offload *opt)
1045 return dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_HTB, opt);
1048 static int htb_init(struct Qdisc *sch, struct nlattr *opt,
1049 struct netlink_ext_ack *extack)
1051 struct net_device *dev = qdisc_dev(sch);
1052 struct tc_htb_qopt_offload offload_opt;
1053 struct htb_sched *q = qdisc_priv(sch);
1054 struct nlattr *tb[TCA_HTB_MAX + 1];
1055 struct tc_htb_glob *gopt;
1060 qdisc_watchdog_init(&q->watchdog, sch);
1061 INIT_WORK(&q->work, htb_work_func);
1066 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
1070 err = nla_parse_nested_deprecated(tb, TCA_HTB_MAX, opt, htb_policy,
1075 if (!tb[TCA_HTB_INIT])
1078 gopt = nla_data(tb[TCA_HTB_INIT]);
1079 if (gopt->version != HTB_VER >> 16)
1082 offload = nla_get_flag(tb[TCA_HTB_OFFLOAD]);
1085 if (sch->parent != TC_H_ROOT)
1088 if (!tc_can_offload(dev) || !dev->netdev_ops->ndo_setup_tc)
1091 q->num_direct_qdiscs = dev->real_num_tx_queues;
1092 q->direct_qdiscs = kcalloc(q->num_direct_qdiscs,
1093 sizeof(*q->direct_qdiscs),
1095 if (!q->direct_qdiscs)
1099 err = qdisc_class_hash_init(&q->clhash);
1101 goto err_free_direct_qdiscs;
1103 qdisc_skb_head_init(&q->direct_queue);
1105 if (tb[TCA_HTB_DIRECT_QLEN])
1106 q->direct_qlen = nla_get_u32(tb[TCA_HTB_DIRECT_QLEN]);
1108 q->direct_qlen = qdisc_dev(sch)->tx_queue_len;
1110 if ((q->rate2quantum = gopt->rate2quantum) < 1)
1111 q->rate2quantum = 1;
1112 q->defcls = gopt->defcls;
1117 for (ntx = 0; ntx < q->num_direct_qdiscs; ntx++) {
1118 struct netdev_queue *dev_queue = netdev_get_tx_queue(dev, ntx);
1119 struct Qdisc *qdisc;
1121 qdisc = qdisc_create_dflt(dev_queue, &pfifo_qdisc_ops,
1122 TC_H_MAKE(sch->handle, 0), extack);
1125 goto err_free_qdiscs;
1128 htb_set_lockdep_class_child(qdisc);
1129 q->direct_qdiscs[ntx] = qdisc;
1130 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1133 sch->flags |= TCQ_F_MQROOT;
1135 offload_opt = (struct tc_htb_qopt_offload) {
1136 .command = TC_HTB_CREATE,
1137 .parent_classid = TC_H_MAJ(sch->handle) >> 16,
1138 .classid = TC_H_MIN(q->defcls),
1141 err = htb_offload(dev, &offload_opt);
1143 goto err_free_qdiscs;
1145 /* Defer this assignment, so that htb_destroy skips offload-related
1146 * parts (especially calling ndo_setup_tc) on errors.
1153 for (ntx = 0; ntx < q->num_direct_qdiscs && q->direct_qdiscs[ntx];
1155 qdisc_put(q->direct_qdiscs[ntx]);
1157 qdisc_class_hash_destroy(&q->clhash);
1158 /* Prevent use-after-free and double-free when htb_destroy gets called.
1160 q->clhash.hash = NULL;
1161 q->clhash.hashsize = 0;
1163 err_free_direct_qdiscs:
1164 kfree(q->direct_qdiscs);
1165 q->direct_qdiscs = NULL;
1169 static void htb_attach_offload(struct Qdisc *sch)
1171 struct net_device *dev = qdisc_dev(sch);
1172 struct htb_sched *q = qdisc_priv(sch);
1175 for (ntx = 0; ntx < q->num_direct_qdiscs; ntx++) {
1176 struct Qdisc *old, *qdisc = q->direct_qdiscs[ntx];
1178 old = dev_graft_qdisc(qdisc->dev_queue, qdisc);
1180 qdisc_hash_add(qdisc, false);
1182 for (ntx = q->num_direct_qdiscs; ntx < dev->num_tx_queues; ntx++) {
1183 struct netdev_queue *dev_queue = netdev_get_tx_queue(dev, ntx);
1184 struct Qdisc *old = dev_graft_qdisc(dev_queue, NULL);
1189 kfree(q->direct_qdiscs);
1190 q->direct_qdiscs = NULL;
1193 static void htb_attach_software(struct Qdisc *sch)
1195 struct net_device *dev = qdisc_dev(sch);
1198 /* Resemble qdisc_graft behavior. */
1199 for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
1200 struct netdev_queue *dev_queue = netdev_get_tx_queue(dev, ntx);
1201 struct Qdisc *old = dev_graft_qdisc(dev_queue, sch);
1203 qdisc_refcount_inc(sch);
1209 static void htb_attach(struct Qdisc *sch)
1211 struct htb_sched *q = qdisc_priv(sch);
1214 htb_attach_offload(sch);
1216 htb_attach_software(sch);
1219 static int htb_dump(struct Qdisc *sch, struct sk_buff *skb)
1221 struct htb_sched *q = qdisc_priv(sch);
1222 struct nlattr *nest;
1223 struct tc_htb_glob gopt;
1226 sch->flags |= TCQ_F_OFFLOADED;
1228 sch->flags &= ~TCQ_F_OFFLOADED;
1230 sch->qstats.overlimits = q->overlimits;
1231 /* Its safe to not acquire qdisc lock. As we hold RTNL,
1232 * no change can happen on the qdisc parameters.
1235 gopt.direct_pkts = q->direct_pkts;
1236 gopt.version = HTB_VER;
1237 gopt.rate2quantum = q->rate2quantum;
1238 gopt.defcls = q->defcls;
1241 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1243 goto nla_put_failure;
1244 if (nla_put(skb, TCA_HTB_INIT, sizeof(gopt), &gopt) ||
1245 nla_put_u32(skb, TCA_HTB_DIRECT_QLEN, q->direct_qlen))
1246 goto nla_put_failure;
1247 if (q->offload && nla_put_flag(skb, TCA_HTB_OFFLOAD))
1248 goto nla_put_failure;
1250 return nla_nest_end(skb, nest);
1253 nla_nest_cancel(skb, nest);
1257 static int htb_dump_class(struct Qdisc *sch, unsigned long arg,
1258 struct sk_buff *skb, struct tcmsg *tcm)
1260 struct htb_class *cl = (struct htb_class *)arg;
1261 struct htb_sched *q = qdisc_priv(sch);
1262 struct nlattr *nest;
1263 struct tc_htb_opt opt;
1265 /* Its safe to not acquire qdisc lock. As we hold RTNL,
1266 * no change can happen on the class parameters.
1268 tcm->tcm_parent = cl->parent ? cl->parent->common.classid : TC_H_ROOT;
1269 tcm->tcm_handle = cl->common.classid;
1270 if (!cl->level && cl->leaf.q)
1271 tcm->tcm_info = cl->leaf.q->handle;
1273 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1275 goto nla_put_failure;
1277 memset(&opt, 0, sizeof(opt));
1279 psched_ratecfg_getrate(&opt.rate, &cl->rate);
1280 opt.buffer = PSCHED_NS2TICKS(cl->buffer);
1281 psched_ratecfg_getrate(&opt.ceil, &cl->ceil);
1282 opt.cbuffer = PSCHED_NS2TICKS(cl->cbuffer);
1283 opt.quantum = cl->quantum;
1284 opt.prio = cl->prio;
1285 opt.level = cl->level;
1286 if (nla_put(skb, TCA_HTB_PARMS, sizeof(opt), &opt))
1287 goto nla_put_failure;
1288 if (q->offload && nla_put_flag(skb, TCA_HTB_OFFLOAD))
1289 goto nla_put_failure;
1290 if ((cl->rate.rate_bytes_ps >= (1ULL << 32)) &&
1291 nla_put_u64_64bit(skb, TCA_HTB_RATE64, cl->rate.rate_bytes_ps,
1293 goto nla_put_failure;
1294 if ((cl->ceil.rate_bytes_ps >= (1ULL << 32)) &&
1295 nla_put_u64_64bit(skb, TCA_HTB_CEIL64, cl->ceil.rate_bytes_ps,
1297 goto nla_put_failure;
1299 return nla_nest_end(skb, nest);
1302 nla_nest_cancel(skb, nest);
1306 static void htb_offload_aggregate_stats(struct htb_sched *q,
1307 struct htb_class *cl)
1309 struct htb_class *c;
1312 memset(&cl->bstats, 0, sizeof(cl->bstats));
1314 for (i = 0; i < q->clhash.hashsize; i++) {
1315 hlist_for_each_entry(c, &q->clhash.hash[i], common.hnode) {
1316 struct htb_class *p = c;
1318 while (p && p->level < cl->level)
1324 cl->bstats.bytes += c->bstats_bias.bytes;
1325 cl->bstats.packets += c->bstats_bias.packets;
1326 if (c->level == 0) {
1327 cl->bstats.bytes += c->leaf.q->bstats.bytes;
1328 cl->bstats.packets += c->leaf.q->bstats.packets;
1335 htb_dump_class_stats(struct Qdisc *sch, unsigned long arg, struct gnet_dump *d)
1337 struct htb_class *cl = (struct htb_class *)arg;
1338 struct htb_sched *q = qdisc_priv(sch);
1339 struct gnet_stats_queue qs = {
1341 .overlimits = cl->overlimits,
1345 if (!cl->level && cl->leaf.q)
1346 qdisc_qstats_qlen_backlog(cl->leaf.q, &qlen, &qs.backlog);
1348 cl->xstats.tokens = clamp_t(s64, PSCHED_NS2TICKS(cl->tokens),
1350 cl->xstats.ctokens = clamp_t(s64, PSCHED_NS2TICKS(cl->ctokens),
1356 cl->bstats = cl->leaf.q->bstats;
1358 memset(&cl->bstats, 0, sizeof(cl->bstats));
1359 cl->bstats.bytes += cl->bstats_bias.bytes;
1360 cl->bstats.packets += cl->bstats_bias.packets;
1362 htb_offload_aggregate_stats(q, cl);
1366 if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
1367 d, NULL, &cl->bstats) < 0 ||
1368 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1369 gnet_stats_copy_queue(d, NULL, &qs, qlen) < 0)
1372 return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1375 static struct netdev_queue *
1376 htb_select_queue(struct Qdisc *sch, struct tcmsg *tcm)
1378 struct net_device *dev = qdisc_dev(sch);
1379 struct tc_htb_qopt_offload offload_opt;
1380 struct htb_sched *q = qdisc_priv(sch);
1384 return sch->dev_queue;
1386 offload_opt = (struct tc_htb_qopt_offload) {
1387 .command = TC_HTB_LEAF_QUERY_QUEUE,
1388 .classid = TC_H_MIN(tcm->tcm_parent),
1390 err = htb_offload(dev, &offload_opt);
1391 if (err || offload_opt.qid >= dev->num_tx_queues)
1393 return netdev_get_tx_queue(dev, offload_opt.qid);
1396 static struct Qdisc *
1397 htb_graft_helper(struct netdev_queue *dev_queue, struct Qdisc *new_q)
1399 struct net_device *dev = dev_queue->dev;
1400 struct Qdisc *old_q;
1402 if (dev->flags & IFF_UP)
1403 dev_deactivate(dev);
1404 old_q = dev_graft_qdisc(dev_queue, new_q);
1406 new_q->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1407 if (dev->flags & IFF_UP)
1413 static struct netdev_queue *htb_offload_get_queue(struct htb_class *cl)
1415 struct netdev_queue *queue;
1417 queue = cl->leaf.offload_queue;
1418 if (!(cl->leaf.q->flags & TCQ_F_BUILTIN))
1419 WARN_ON(cl->leaf.q->dev_queue != queue);
1424 static void htb_offload_move_qdisc(struct Qdisc *sch, struct htb_class *cl_old,
1425 struct htb_class *cl_new, bool destroying)
1427 struct netdev_queue *queue_old, *queue_new;
1428 struct net_device *dev = qdisc_dev(sch);
1430 queue_old = htb_offload_get_queue(cl_old);
1431 queue_new = htb_offload_get_queue(cl_new);
1434 struct Qdisc *qdisc;
1436 if (dev->flags & IFF_UP)
1437 dev_deactivate(dev);
1438 qdisc = dev_graft_qdisc(queue_old, NULL);
1439 WARN_ON(qdisc != cl_old->leaf.q);
1442 if (!(cl_old->leaf.q->flags & TCQ_F_BUILTIN))
1443 cl_old->leaf.q->dev_queue = queue_new;
1444 cl_old->leaf.offload_queue = queue_new;
1447 struct Qdisc *qdisc;
1449 qdisc = dev_graft_qdisc(queue_new, cl_old->leaf.q);
1450 if (dev->flags & IFF_UP)
1452 WARN_ON(!(qdisc->flags & TCQ_F_BUILTIN));
1456 static int htb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1457 struct Qdisc **old, struct netlink_ext_ack *extack)
1459 struct netdev_queue *dev_queue = sch->dev_queue;
1460 struct htb_class *cl = (struct htb_class *)arg;
1461 struct htb_sched *q = qdisc_priv(sch);
1462 struct Qdisc *old_q;
1468 dev_queue = htb_offload_get_queue(cl);
1471 new = qdisc_create_dflt(dev_queue, &pfifo_qdisc_ops,
1472 cl->common.classid, extack);
1478 htb_set_lockdep_class_child(new);
1479 /* One ref for cl->leaf.q, the other for dev_queue->qdisc. */
1480 qdisc_refcount_inc(new);
1481 old_q = htb_graft_helper(dev_queue, new);
1484 *old = qdisc_replace(sch, new, &cl->leaf.q);
1487 WARN_ON(old_q != *old);
1494 static struct Qdisc *htb_leaf(struct Qdisc *sch, unsigned long arg)
1496 struct htb_class *cl = (struct htb_class *)arg;
1497 return !cl->level ? cl->leaf.q : NULL;
1500 static void htb_qlen_notify(struct Qdisc *sch, unsigned long arg)
1502 struct htb_class *cl = (struct htb_class *)arg;
1504 htb_deactivate(qdisc_priv(sch), cl);
1507 static inline int htb_parent_last_child(struct htb_class *cl)
1510 /* the root class */
1512 if (cl->parent->children > 1)
1513 /* not the last child */
1518 static void htb_parent_to_leaf(struct Qdisc *sch, struct htb_class *cl,
1519 struct Qdisc *new_q)
1521 struct htb_sched *q = qdisc_priv(sch);
1522 struct htb_class *parent = cl->parent;
1524 WARN_ON(cl->level || !cl->leaf.q || cl->prio_activity);
1526 if (parent->cmode != HTB_CAN_SEND)
1527 htb_safe_rb_erase(&parent->pq_node,
1528 &q->hlevel[parent->level].wait_pq);
1531 memset(&parent->inner, 0, sizeof(parent->inner));
1532 parent->leaf.q = new_q ? new_q : &noop_qdisc;
1533 parent->tokens = parent->buffer;
1534 parent->ctokens = parent->cbuffer;
1535 parent->t_c = ktime_get_ns();
1536 parent->cmode = HTB_CAN_SEND;
1538 parent->leaf.offload_queue = cl->leaf.offload_queue;
1541 static void htb_parent_to_leaf_offload(struct Qdisc *sch,
1542 struct netdev_queue *dev_queue,
1543 struct Qdisc *new_q)
1545 struct Qdisc *old_q;
1547 /* One ref for cl->leaf.q, the other for dev_queue->qdisc. */
1549 qdisc_refcount_inc(new_q);
1550 old_q = htb_graft_helper(dev_queue, new_q);
1551 WARN_ON(!(old_q->flags & TCQ_F_BUILTIN));
1554 static int htb_destroy_class_offload(struct Qdisc *sch, struct htb_class *cl,
1555 bool last_child, bool destroying,
1556 struct netlink_ext_ack *extack)
1558 struct tc_htb_qopt_offload offload_opt;
1559 struct netdev_queue *dev_queue;
1560 struct Qdisc *q = cl->leaf.q;
1568 dev_queue = htb_offload_get_queue(cl);
1569 /* When destroying, caller qdisc_graft grafts the new qdisc and invokes
1570 * qdisc_put for the qdisc being destroyed. htb_destroy_class_offload
1571 * does not need to graft or qdisc_put the qdisc being destroyed.
1574 old = htb_graft_helper(dev_queue, NULL);
1575 /* Last qdisc grafted should be the same as cl->leaf.q when
1576 * calling htb_delete.
1582 cl->parent->bstats_bias.bytes += q->bstats.bytes;
1583 cl->parent->bstats_bias.packets += q->bstats.packets;
1586 offload_opt = (struct tc_htb_qopt_offload) {
1587 .command = !last_child ? TC_HTB_LEAF_DEL :
1588 destroying ? TC_HTB_LEAF_DEL_LAST_FORCE :
1589 TC_HTB_LEAF_DEL_LAST,
1590 .classid = cl->common.classid,
1593 err = htb_offload(qdisc_dev(sch), &offload_opt);
1599 htb_graft_helper(dev_queue, old);
1605 if (!err && offload_opt.classid != TC_H_MIN(cl->common.classid)) {
1606 u32 classid = TC_H_MAJ(sch->handle) |
1607 TC_H_MIN(offload_opt.classid);
1608 struct htb_class *moved_cl = htb_find(classid, sch);
1610 htb_offload_move_qdisc(sch, moved_cl, cl, destroying);
1616 static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
1619 WARN_ON(!cl->leaf.q);
1620 qdisc_put(cl->leaf.q);
1622 gen_kill_estimator(&cl->rate_est);
1623 tcf_block_put(cl->block);
1627 static void htb_destroy(struct Qdisc *sch)
1629 struct net_device *dev = qdisc_dev(sch);
1630 struct tc_htb_qopt_offload offload_opt;
1631 struct htb_sched *q = qdisc_priv(sch);
1632 struct hlist_node *next;
1633 bool nonempty, changed;
1634 struct htb_class *cl;
1637 cancel_work_sync(&q->work);
1638 qdisc_watchdog_cancel(&q->watchdog);
1639 /* This line used to be after htb_destroy_class call below
1640 * and surprisingly it worked in 2.4. But it must precede it
1641 * because filter need its target class alive to be able to call
1642 * unbind_filter on it (without Oops).
1644 tcf_block_put(q->block);
1646 for (i = 0; i < q->clhash.hashsize; i++) {
1647 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1648 tcf_block_put(cl->block);
1656 for (i = 0; i < q->clhash.hashsize; i++) {
1657 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1662 htb_destroy_class(sch, cl);
1673 last_child = htb_parent_last_child(cl);
1674 htb_destroy_class_offload(sch, cl, last_child,
1676 qdisc_class_hash_remove(&q->clhash,
1679 cl->parent->children--;
1681 htb_parent_to_leaf(sch, cl, NULL);
1682 htb_destroy_class(sch, cl);
1688 qdisc_class_hash_destroy(&q->clhash);
1689 __qdisc_reset_queue(&q->direct_queue);
1694 offload_opt = (struct tc_htb_qopt_offload) {
1695 .command = TC_HTB_DESTROY,
1697 htb_offload(dev, &offload_opt);
1699 if (!q->direct_qdiscs)
1701 for (i = 0; i < q->num_direct_qdiscs && q->direct_qdiscs[i]; i++)
1702 qdisc_put(q->direct_qdiscs[i]);
1703 kfree(q->direct_qdiscs);
1706 static int htb_delete(struct Qdisc *sch, unsigned long arg,
1707 struct netlink_ext_ack *extack)
1709 struct htb_sched *q = qdisc_priv(sch);
1710 struct htb_class *cl = (struct htb_class *)arg;
1711 struct Qdisc *new_q = NULL;
1715 /* TODO: why don't allow to delete subtree ? references ? does
1716 * tc subsys guarantee us that in htb_destroy it holds no class
1717 * refs so that we can remove children safely there ?
1719 if (cl->children || cl->filter_cnt)
1722 if (!cl->level && htb_parent_last_child(cl))
1726 err = htb_destroy_class_offload(sch, cl, last_child, false,
1733 struct netdev_queue *dev_queue = sch->dev_queue;
1736 dev_queue = htb_offload_get_queue(cl);
1738 new_q = qdisc_create_dflt(dev_queue, &pfifo_qdisc_ops,
1739 cl->parent->common.classid,
1743 htb_set_lockdep_class_child(new_q);
1744 htb_parent_to_leaf_offload(sch, dev_queue, new_q);
1751 qdisc_purge_queue(cl->leaf.q);
1753 /* delete from hash and active; remainder in destroy_class */
1754 qdisc_class_hash_remove(&q->clhash, &cl->common);
1756 cl->parent->children--;
1758 if (cl->prio_activity)
1759 htb_deactivate(q, cl);
1761 if (cl->cmode != HTB_CAN_SEND)
1762 htb_safe_rb_erase(&cl->pq_node,
1763 &q->hlevel[cl->level].wait_pq);
1766 htb_parent_to_leaf(sch, cl, new_q);
1768 sch_tree_unlock(sch);
1770 htb_destroy_class(sch, cl);
1774 static int htb_change_class(struct Qdisc *sch, u32 classid,
1775 u32 parentid, struct nlattr **tca,
1776 unsigned long *arg, struct netlink_ext_ack *extack)
1779 struct htb_sched *q = qdisc_priv(sch);
1780 struct htb_class *cl = (struct htb_class *)*arg, *parent;
1781 struct tc_htb_qopt_offload offload_opt;
1782 struct nlattr *opt = tca[TCA_OPTIONS];
1783 struct nlattr *tb[TCA_HTB_MAX + 1];
1784 struct Qdisc *parent_qdisc = NULL;
1785 struct netdev_queue *dev_queue;
1786 struct tc_htb_opt *hopt;
1790 /* extract all subattrs from opt attr */
1794 err = nla_parse_nested_deprecated(tb, TCA_HTB_MAX, opt, htb_policy,
1800 if (tb[TCA_HTB_PARMS] == NULL)
1803 parent = parentid == TC_H_ROOT ? NULL : htb_find(parentid, sch);
1805 hopt = nla_data(tb[TCA_HTB_PARMS]);
1806 if (!hopt->rate.rate || !hopt->ceil.rate)
1810 /* Options not supported by the offload. */
1811 if (hopt->rate.overhead || hopt->ceil.overhead) {
1812 NL_SET_ERR_MSG(extack, "HTB offload doesn't support the overhead parameter");
1815 if (hopt->rate.mpu || hopt->ceil.mpu) {
1816 NL_SET_ERR_MSG(extack, "HTB offload doesn't support the mpu parameter");
1819 if (hopt->quantum) {
1820 NL_SET_ERR_MSG(extack, "HTB offload doesn't support the quantum parameter");
1824 NL_SET_ERR_MSG(extack, "HTB offload doesn't support the prio parameter");
1829 /* Keeping backward compatible with rate_table based iproute2 tc */
1830 if (hopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
1831 qdisc_put_rtab(qdisc_get_rtab(&hopt->rate, tb[TCA_HTB_RTAB],
1834 if (hopt->ceil.linklayer == TC_LINKLAYER_UNAWARE)
1835 qdisc_put_rtab(qdisc_get_rtab(&hopt->ceil, tb[TCA_HTB_CTAB],
1838 rate64 = tb[TCA_HTB_RATE64] ? nla_get_u64(tb[TCA_HTB_RATE64]) : 0;
1839 ceil64 = tb[TCA_HTB_CEIL64] ? nla_get_u64(tb[TCA_HTB_CEIL64]) : 0;
1841 if (!cl) { /* new class */
1842 struct net_device *dev = qdisc_dev(sch);
1843 struct Qdisc *new_q, *old_q;
1847 struct gnet_estimator opt;
1850 .nla_len = nla_attr_size(sizeof(est.opt)),
1851 .nla_type = TCA_RATE,
1854 /* 4s interval, 16s averaging constant */
1860 /* check for valid classid */
1861 if (!classid || TC_H_MAJ(classid ^ sch->handle) ||
1862 htb_find(classid, sch))
1865 /* check maximal depth */
1866 if (parent && parent->parent && parent->parent->level < 2) {
1867 pr_err("htb: tree is too deep\n");
1871 cl = kzalloc(sizeof(*cl), GFP_KERNEL);
1875 err = tcf_block_get(&cl->block, &cl->filter_list, sch, extack);
1880 if (htb_rate_est || tca[TCA_RATE]) {
1881 err = gen_new_estimator(&cl->bstats, NULL,
1884 qdisc_root_sleeping_running(sch),
1885 tca[TCA_RATE] ? : &est.nla);
1891 RB_CLEAR_NODE(&cl->pq_node);
1893 for (prio = 0; prio < TC_HTB_NUMPRIO; prio++)
1894 RB_CLEAR_NODE(&cl->node[prio]);
1896 cl->common.classid = classid;
1898 /* Make sure nothing interrupts us in between of two
1899 * ndo_setup_tc calls.
1903 /* create leaf qdisc early because it uses kmalloc(GFP_KERNEL)
1904 * so that can't be used inside of sch_tree_lock
1905 * -- thanks to Karlis Peisenieks
1908 dev_queue = sch->dev_queue;
1909 } else if (!(parent && !parent->level)) {
1910 /* Assign a dev_queue to this classid. */
1911 offload_opt = (struct tc_htb_qopt_offload) {
1912 .command = TC_HTB_LEAF_ALLOC_QUEUE,
1913 .classid = cl->common.classid,
1914 .parent_classid = parent ?
1915 TC_H_MIN(parent->common.classid) :
1916 TC_HTB_CLASSID_ROOT,
1917 .rate = max_t(u64, hopt->rate.rate, rate64),
1918 .ceil = max_t(u64, hopt->ceil.rate, ceil64),
1921 err = htb_offload(dev, &offload_opt);
1923 pr_err("htb: TC_HTB_LEAF_ALLOC_QUEUE failed with err = %d\n",
1925 goto err_kill_estimator;
1927 dev_queue = netdev_get_tx_queue(dev, offload_opt.qid);
1928 } else { /* First child. */
1929 dev_queue = htb_offload_get_queue(parent);
1930 old_q = htb_graft_helper(dev_queue, NULL);
1931 WARN_ON(old_q != parent->leaf.q);
1932 offload_opt = (struct tc_htb_qopt_offload) {
1933 .command = TC_HTB_LEAF_TO_INNER,
1934 .classid = cl->common.classid,
1936 TC_H_MIN(parent->common.classid),
1937 .rate = max_t(u64, hopt->rate.rate, rate64),
1938 .ceil = max_t(u64, hopt->ceil.rate, ceil64),
1941 err = htb_offload(dev, &offload_opt);
1943 pr_err("htb: TC_HTB_LEAF_TO_INNER failed with err = %d\n",
1945 htb_graft_helper(dev_queue, old_q);
1946 goto err_kill_estimator;
1948 parent->bstats_bias.bytes += old_q->bstats.bytes;
1949 parent->bstats_bias.packets += old_q->bstats.packets;
1952 new_q = qdisc_create_dflt(dev_queue, &pfifo_qdisc_ops,
1956 htb_set_lockdep_class_child(new_q);
1957 /* One ref for cl->leaf.q, the other for
1960 qdisc_refcount_inc(new_q);
1962 old_q = htb_graft_helper(dev_queue, new_q);
1963 /* No qdisc_put needed. */
1964 WARN_ON(!(old_q->flags & TCQ_F_BUILTIN));
1967 if (parent && !parent->level) {
1968 /* turn parent into inner node */
1969 qdisc_purge_queue(parent->leaf.q);
1970 parent_qdisc = parent->leaf.q;
1971 if (parent->prio_activity)
1972 htb_deactivate(q, parent);
1974 /* remove from evt list because of level change */
1975 if (parent->cmode != HTB_CAN_SEND) {
1976 htb_safe_rb_erase(&parent->pq_node, &q->hlevel[0].wait_pq);
1977 parent->cmode = HTB_CAN_SEND;
1979 parent->level = (parent->parent ? parent->parent->level
1980 : TC_HTB_MAXDEPTH) - 1;
1981 memset(&parent->inner, 0, sizeof(parent->inner));
1984 /* leaf (we) needs elementary qdisc */
1985 cl->leaf.q = new_q ? new_q : &noop_qdisc;
1987 cl->leaf.offload_queue = dev_queue;
1989 cl->parent = parent;
1991 /* set class to be in HTB_CAN_SEND state */
1992 cl->tokens = PSCHED_TICKS2NS(hopt->buffer);
1993 cl->ctokens = PSCHED_TICKS2NS(hopt->cbuffer);
1994 cl->mbuffer = 60ULL * NSEC_PER_SEC; /* 1min */
1995 cl->t_c = ktime_get_ns();
1996 cl->cmode = HTB_CAN_SEND;
1998 /* attach to the hash list and parent's family */
1999 qdisc_class_hash_insert(&q->clhash, &cl->common);
2002 if (cl->leaf.q != &noop_qdisc)
2003 qdisc_hash_add(cl->leaf.q, true);
2005 if (tca[TCA_RATE]) {
2006 err = gen_replace_estimator(&cl->bstats, NULL,
2009 qdisc_root_sleeping_running(sch),
2016 struct net_device *dev = qdisc_dev(sch);
2018 offload_opt = (struct tc_htb_qopt_offload) {
2019 .command = TC_HTB_NODE_MODIFY,
2020 .classid = cl->common.classid,
2021 .rate = max_t(u64, hopt->rate.rate, rate64),
2022 .ceil = max_t(u64, hopt->ceil.rate, ceil64),
2025 err = htb_offload(dev, &offload_opt);
2027 /* Estimator was replaced, and rollback may fail
2028 * as well, so we don't try to recover it, and
2029 * the estimator won't work property with the
2030 * offload anyway, because bstats are updated
2031 * only when the stats are queried.
2039 psched_ratecfg_precompute(&cl->rate, &hopt->rate, rate64);
2040 psched_ratecfg_precompute(&cl->ceil, &hopt->ceil, ceil64);
2042 /* it used to be a nasty bug here, we have to check that node
2043 * is really leaf before changing cl->leaf !
2046 u64 quantum = cl->rate.rate_bytes_ps;
2048 do_div(quantum, q->rate2quantum);
2049 cl->quantum = min_t(u64, quantum, INT_MAX);
2051 if (!hopt->quantum && cl->quantum < 1000) {
2055 if (!hopt->quantum && cl->quantum > 200000) {
2057 cl->quantum = 200000;
2060 cl->quantum = hopt->quantum;
2061 if ((cl->prio = hopt->prio) >= TC_HTB_NUMPRIO)
2062 cl->prio = TC_HTB_NUMPRIO - 1;
2065 cl->buffer = PSCHED_TICKS2NS(hopt->buffer);
2066 cl->cbuffer = PSCHED_TICKS2NS(hopt->cbuffer);
2068 sch_tree_unlock(sch);
2069 qdisc_put(parent_qdisc);
2072 pr_warn("HTB: quantum of class %X is %s. Consider r2q change.\n",
2073 cl->common.classid, (warn == -1 ? "small" : "big"));
2075 qdisc_class_hash_grow(sch, &q->clhash);
2077 *arg = (unsigned long)cl;
2081 gen_kill_estimator(&cl->rate_est);
2083 tcf_block_put(cl->block);
2089 static struct tcf_block *htb_tcf_block(struct Qdisc *sch, unsigned long arg,
2090 struct netlink_ext_ack *extack)
2092 struct htb_sched *q = qdisc_priv(sch);
2093 struct htb_class *cl = (struct htb_class *)arg;
2095 return cl ? cl->block : q->block;
2098 static unsigned long htb_bind_filter(struct Qdisc *sch, unsigned long parent,
2101 struct htb_class *cl = htb_find(classid, sch);
2103 /*if (cl && !cl->level) return 0;
2104 * The line above used to be there to prevent attaching filters to
2105 * leaves. But at least tc_index filter uses this just to get class
2106 * for other reasons so that we have to allow for it.
2108 * 19.6.2002 As Werner explained it is ok - bind filter is just
2109 * another way to "lock" the class - unlike "get" this lock can
2110 * be broken by class during destroy IIUC.
2114 return (unsigned long)cl;
2117 static void htb_unbind_filter(struct Qdisc *sch, unsigned long arg)
2119 struct htb_class *cl = (struct htb_class *)arg;
2125 static void htb_walk(struct Qdisc *sch, struct qdisc_walker *arg)
2127 struct htb_sched *q = qdisc_priv(sch);
2128 struct htb_class *cl;
2134 for (i = 0; i < q->clhash.hashsize; i++) {
2135 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
2136 if (arg->count < arg->skip) {
2140 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
2149 static const struct Qdisc_class_ops htb_class_ops = {
2150 .select_queue = htb_select_queue,
2153 .qlen_notify = htb_qlen_notify,
2155 .change = htb_change_class,
2156 .delete = htb_delete,
2158 .tcf_block = htb_tcf_block,
2159 .bind_tcf = htb_bind_filter,
2160 .unbind_tcf = htb_unbind_filter,
2161 .dump = htb_dump_class,
2162 .dump_stats = htb_dump_class_stats,
2165 static struct Qdisc_ops htb_qdisc_ops __read_mostly = {
2166 .cl_ops = &htb_class_ops,
2168 .priv_size = sizeof(struct htb_sched),
2169 .enqueue = htb_enqueue,
2170 .dequeue = htb_dequeue,
2171 .peek = qdisc_peek_dequeued,
2173 .attach = htb_attach,
2175 .destroy = htb_destroy,
2177 .owner = THIS_MODULE,
2180 static int __init htb_module_init(void)
2182 return register_qdisc(&htb_qdisc_ops);
2184 static void __exit htb_module_exit(void)
2186 unregister_qdisc(&htb_qdisc_ops);
2189 module_init(htb_module_init)
2190 module_exit(htb_module_exit)
2191 MODULE_LICENSE("GPL");