Loading...
1/*
2 * net/sched/sch_htb.c Hierarchical token bucket, feed tree version
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License, or (at your option) any later version.
8 *
9 * Authors: Martin Devera, <devik@cdi.cz>
10 *
11 * Credits (in time order) for older HTB versions:
12 * Stef Coene <stef.coene@docum.org>
13 * HTB support at LARTC mailing list
14 * Ondrej Kraus, <krauso@barr.cz>
15 * found missing INIT_QDISC(htb)
16 * Vladimir Smelhaus, Aamer Akhter, Bert Hubert
17 * helped a lot to locate nasty class stall bug
18 * Andi Kleen, Jamal Hadi, Bert Hubert
19 * code review and helpful comments on shaping
20 * Tomasz Wrona, <tw@eter.tym.pl>
21 * created test case so that I was able to fix nasty bug
22 * Wilfried Weissmann
23 * spotted bug in dequeue code and helped with fix
24 * Jiri Fojtasek
25 * fixed requeue routine
26 * and many others. thanks.
27 */
28#include <linux/module.h>
29#include <linux/moduleparam.h>
30#include <linux/types.h>
31#include <linux/kernel.h>
32#include <linux/string.h>
33#include <linux/errno.h>
34#include <linux/skbuff.h>
35#include <linux/list.h>
36#include <linux/compiler.h>
37#include <linux/rbtree.h>
38#include <linux/workqueue.h>
39#include <linux/slab.h>
40#include <net/netlink.h>
41#include <net/pkt_sched.h>
42
43/* HTB algorithm.
44 Author: devik@cdi.cz
45 ========================================================================
46 HTB is like TBF with multiple classes. It is also similar to CBQ because
47 it allows to assign priority to each class in hierarchy.
48 In fact it is another implementation of Floyd's formal sharing.
49
50 Levels:
51 Each class is assigned level. Leaf has ALWAYS level 0 and root
52 classes have level TC_HTB_MAXDEPTH-1. Interior nodes has level
53 one less than their parent.
54*/
55
56static int htb_hysteresis __read_mostly = 0; /* whether to use mode hysteresis for speedup */
57#define HTB_VER 0x30011 /* major must be matched with number suplied by TC as version */
58
59#if HTB_VER >> 16 != TC_HTB_PROTOVER
60#error "Mismatched sch_htb.c and pkt_sch.h"
61#endif
62
63/* Module parameter and sysfs export */
64module_param (htb_hysteresis, int, 0640);
65MODULE_PARM_DESC(htb_hysteresis, "Hysteresis mode, less CPU load, less accurate");
66
67/* used internaly to keep status of single class */
68enum htb_cmode {
69 HTB_CANT_SEND, /* class can't send and can't borrow */
70 HTB_MAY_BORROW, /* class can't send but may borrow */
71 HTB_CAN_SEND /* class can send */
72};
73
74/* interior & leaf nodes; props specific to leaves are marked L: */
75struct htb_class {
76 struct Qdisc_class_common common;
77 /* general class parameters */
78 struct gnet_stats_basic_packed bstats;
79 struct gnet_stats_queue qstats;
80 struct gnet_stats_rate_est rate_est;
81 struct tc_htb_xstats xstats; /* our special stats */
82 int refcnt; /* usage count of this class */
83
84 /* topology */
85 int level; /* our level (see above) */
86 unsigned int children;
87 struct htb_class *parent; /* parent class */
88
89 int prio; /* these two are used only by leaves... */
90 int quantum; /* but stored for parent-to-leaf return */
91
92 union {
93 struct htb_class_leaf {
94 struct Qdisc *q;
95 int deficit[TC_HTB_MAXDEPTH];
96 struct list_head drop_list;
97 } leaf;
98 struct htb_class_inner {
99 struct rb_root feed[TC_HTB_NUMPRIO]; /* feed trees */
100 struct rb_node *ptr[TC_HTB_NUMPRIO]; /* current class ptr */
101 /* When class changes from state 1->2 and disconnects from
102 * parent's feed then we lost ptr value and start from the
103 * first child again. Here we store classid of the
104 * last valid ptr (used when ptr is NULL).
105 */
106 u32 last_ptr_id[TC_HTB_NUMPRIO];
107 } inner;
108 } un;
109 struct rb_node node[TC_HTB_NUMPRIO]; /* node for self or feed tree */
110 struct rb_node pq_node; /* node for event queue */
111 psched_time_t pq_key;
112
113 int prio_activity; /* for which prios are we active */
114 enum htb_cmode cmode; /* current mode of the class */
115
116 /* class attached filters */
117 struct tcf_proto *filter_list;
118 int filter_cnt;
119
120 /* token bucket parameters */
121 struct qdisc_rate_table *rate; /* rate table of the class itself */
122 struct qdisc_rate_table *ceil; /* ceiling rate (limits borrows too) */
123 long buffer, cbuffer; /* token bucket depth/rate */
124 psched_tdiff_t mbuffer; /* max wait time */
125 long tokens, ctokens; /* current number of tokens */
126 psched_time_t t_c; /* checkpoint time */
127};
128
129struct htb_sched {
130 struct Qdisc_class_hash clhash;
131 struct list_head drops[TC_HTB_NUMPRIO];/* active leaves (for drops) */
132
133 /* self list - roots of self generating tree */
134 struct rb_root row[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
135 int row_mask[TC_HTB_MAXDEPTH];
136 struct rb_node *ptr[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
137 u32 last_ptr_id[TC_HTB_MAXDEPTH][TC_HTB_NUMPRIO];
138
139 /* self wait list - roots of wait PQs per row */
140 struct rb_root wait_pq[TC_HTB_MAXDEPTH];
141
142 /* time of nearest event per level (row) */
143 psched_time_t near_ev_cache[TC_HTB_MAXDEPTH];
144
145 int defcls; /* class where unclassified flows go to */
146
147 /* filters for qdisc itself */
148 struct tcf_proto *filter_list;
149
150 int rate2quantum; /* quant = rate / rate2quantum */
151 psched_time_t now; /* cached dequeue time */
152 struct qdisc_watchdog watchdog;
153
154 /* non shaped skbs; let them go directly thru */
155 struct sk_buff_head direct_queue;
156 int direct_qlen; /* max qlen of above */
157
158 long direct_pkts;
159
160#define HTB_WARN_TOOMANYEVENTS 0x1
161 unsigned int warned; /* only one warning */
162 struct work_struct work;
163};
164
165/* find class in global hash table using given handle */
166static inline struct htb_class *htb_find(u32 handle, struct Qdisc *sch)
167{
168 struct htb_sched *q = qdisc_priv(sch);
169 struct Qdisc_class_common *clc;
170
171 clc = qdisc_class_find(&q->clhash, handle);
172 if (clc == NULL)
173 return NULL;
174 return container_of(clc, struct htb_class, common);
175}
176
177/**
178 * htb_classify - classify a packet into class
179 *
180 * It returns NULL if the packet should be dropped or -1 if the packet
181 * should be passed directly thru. In all other cases leaf class is returned.
182 * We allow direct class selection by classid in priority. The we examine
183 * filters in qdisc and in inner nodes (if higher filter points to the inner
184 * node). If we end up with classid MAJOR:0 we enqueue the skb into special
185 * internal fifo (direct). These packets then go directly thru. If we still
186 * have no valid leaf we try to use MAJOR:default leaf. It still unsuccessful
187 * then finish and return direct queue.
188 */
189#define HTB_DIRECT ((struct htb_class *)-1L)
190
191static struct htb_class *htb_classify(struct sk_buff *skb, struct Qdisc *sch,
192 int *qerr)
193{
194 struct htb_sched *q = qdisc_priv(sch);
195 struct htb_class *cl;
196 struct tcf_result res;
197 struct tcf_proto *tcf;
198 int result;
199
200 /* allow to select class by setting skb->priority to valid classid;
201 * note that nfmark can be used too by attaching filter fw with no
202 * rules in it
203 */
204 if (skb->priority == sch->handle)
205 return HTB_DIRECT; /* X:0 (direct flow) selected */
206 cl = htb_find(skb->priority, sch);
207 if (cl && cl->level == 0)
208 return cl;
209
210 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
211 tcf = q->filter_list;
212 while (tcf && (result = tc_classify(skb, tcf, &res)) >= 0) {
213#ifdef CONFIG_NET_CLS_ACT
214 switch (result) {
215 case TC_ACT_QUEUED:
216 case TC_ACT_STOLEN:
217 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
218 case TC_ACT_SHOT:
219 return NULL;
220 }
221#endif
222 cl = (void *)res.class;
223 if (!cl) {
224 if (res.classid == sch->handle)
225 return HTB_DIRECT; /* X:0 (direct flow) */
226 cl = htb_find(res.classid, sch);
227 if (!cl)
228 break; /* filter selected invalid classid */
229 }
230 if (!cl->level)
231 return cl; /* we hit leaf; return it */
232
233 /* we have got inner class; apply inner filter chain */
234 tcf = cl->filter_list;
235 }
236 /* classification failed; try to use default class */
237 cl = htb_find(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
238 if (!cl || cl->level)
239 return HTB_DIRECT; /* bad default .. this is safe bet */
240 return cl;
241}
242
243/**
244 * htb_add_to_id_tree - adds class to the round robin list
245 *
246 * Routine adds class to the list (actually tree) sorted by classid.
247 * Make sure that class is not already on such list for given prio.
248 */
249static void htb_add_to_id_tree(struct rb_root *root,
250 struct htb_class *cl, int prio)
251{
252 struct rb_node **p = &root->rb_node, *parent = NULL;
253
254 while (*p) {
255 struct htb_class *c;
256 parent = *p;
257 c = rb_entry(parent, struct htb_class, node[prio]);
258
259 if (cl->common.classid > c->common.classid)
260 p = &parent->rb_right;
261 else
262 p = &parent->rb_left;
263 }
264 rb_link_node(&cl->node[prio], parent, p);
265 rb_insert_color(&cl->node[prio], root);
266}
267
268/**
269 * htb_add_to_wait_tree - adds class to the event queue with delay
270 *
271 * The class is added to priority event queue to indicate that class will
272 * change its mode in cl->pq_key microseconds. Make sure that class is not
273 * already in the queue.
274 */
275static void htb_add_to_wait_tree(struct htb_sched *q,
276 struct htb_class *cl, long delay)
277{
278 struct rb_node **p = &q->wait_pq[cl->level].rb_node, *parent = NULL;
279
280 cl->pq_key = q->now + delay;
281 if (cl->pq_key == q->now)
282 cl->pq_key++;
283
284 /* update the nearest event cache */
285 if (q->near_ev_cache[cl->level] > cl->pq_key)
286 q->near_ev_cache[cl->level] = cl->pq_key;
287
288 while (*p) {
289 struct htb_class *c;
290 parent = *p;
291 c = rb_entry(parent, struct htb_class, pq_node);
292 if (cl->pq_key >= c->pq_key)
293 p = &parent->rb_right;
294 else
295 p = &parent->rb_left;
296 }
297 rb_link_node(&cl->pq_node, parent, p);
298 rb_insert_color(&cl->pq_node, &q->wait_pq[cl->level]);
299}
300
301/**
302 * htb_next_rb_node - finds next node in binary tree
303 *
304 * When we are past last key we return NULL.
305 * Average complexity is 2 steps per call.
306 */
307static inline void htb_next_rb_node(struct rb_node **n)
308{
309 *n = rb_next(*n);
310}
311
312/**
313 * htb_add_class_to_row - add class to its row
314 *
315 * The class is added to row at priorities marked in mask.
316 * It does nothing if mask == 0.
317 */
318static inline void htb_add_class_to_row(struct htb_sched *q,
319 struct htb_class *cl, int mask)
320{
321 q->row_mask[cl->level] |= mask;
322 while (mask) {
323 int prio = ffz(~mask);
324 mask &= ~(1 << prio);
325 htb_add_to_id_tree(q->row[cl->level] + prio, cl, prio);
326 }
327}
328
329/* If this triggers, it is a bug in this code, but it need not be fatal */
330static void htb_safe_rb_erase(struct rb_node *rb, struct rb_root *root)
331{
332 if (RB_EMPTY_NODE(rb)) {
333 WARN_ON(1);
334 } else {
335 rb_erase(rb, root);
336 RB_CLEAR_NODE(rb);
337 }
338}
339
340
341/**
342 * htb_remove_class_from_row - removes class from its row
343 *
344 * The class is removed from row at priorities marked in mask.
345 * It does nothing if mask == 0.
346 */
347static inline void htb_remove_class_from_row(struct htb_sched *q,
348 struct htb_class *cl, int mask)
349{
350 int m = 0;
351
352 while (mask) {
353 int prio = ffz(~mask);
354
355 mask &= ~(1 << prio);
356 if (q->ptr[cl->level][prio] == cl->node + prio)
357 htb_next_rb_node(q->ptr[cl->level] + prio);
358
359 htb_safe_rb_erase(cl->node + prio, q->row[cl->level] + prio);
360 if (!q->row[cl->level][prio].rb_node)
361 m |= 1 << prio;
362 }
363 q->row_mask[cl->level] &= ~m;
364}
365
366/**
367 * htb_activate_prios - creates active classe's feed chain
368 *
369 * The class is connected to ancestors and/or appropriate rows
370 * for priorities it is participating on. cl->cmode must be new
371 * (activated) mode. It does nothing if cl->prio_activity == 0.
372 */
373static void htb_activate_prios(struct htb_sched *q, struct htb_class *cl)
374{
375 struct htb_class *p = cl->parent;
376 long m, mask = cl->prio_activity;
377
378 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
379 m = mask;
380 while (m) {
381 int prio = ffz(~m);
382 m &= ~(1 << prio);
383
384 if (p->un.inner.feed[prio].rb_node)
385 /* parent already has its feed in use so that
386 * reset bit in mask as parent is already ok
387 */
388 mask &= ~(1 << prio);
389
390 htb_add_to_id_tree(p->un.inner.feed + prio, cl, prio);
391 }
392 p->prio_activity |= mask;
393 cl = p;
394 p = cl->parent;
395
396 }
397 if (cl->cmode == HTB_CAN_SEND && mask)
398 htb_add_class_to_row(q, cl, mask);
399}
400
401/**
402 * htb_deactivate_prios - remove class from feed chain
403 *
404 * cl->cmode must represent old mode (before deactivation). It does
405 * nothing if cl->prio_activity == 0. Class is removed from all feed
406 * chains and rows.
407 */
408static void htb_deactivate_prios(struct htb_sched *q, struct htb_class *cl)
409{
410 struct htb_class *p = cl->parent;
411 long m, mask = cl->prio_activity;
412
413 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
414 m = mask;
415 mask = 0;
416 while (m) {
417 int prio = ffz(~m);
418 m &= ~(1 << prio);
419
420 if (p->un.inner.ptr[prio] == cl->node + prio) {
421 /* we are removing child which is pointed to from
422 * parent feed - forget the pointer but remember
423 * classid
424 */
425 p->un.inner.last_ptr_id[prio] = cl->common.classid;
426 p->un.inner.ptr[prio] = NULL;
427 }
428
429 htb_safe_rb_erase(cl->node + prio, p->un.inner.feed + prio);
430
431 if (!p->un.inner.feed[prio].rb_node)
432 mask |= 1 << prio;
433 }
434
435 p->prio_activity &= ~mask;
436 cl = p;
437 p = cl->parent;
438
439 }
440 if (cl->cmode == HTB_CAN_SEND && mask)
441 htb_remove_class_from_row(q, cl, mask);
442}
443
444static inline long htb_lowater(const struct htb_class *cl)
445{
446 if (htb_hysteresis)
447 return cl->cmode != HTB_CANT_SEND ? -cl->cbuffer : 0;
448 else
449 return 0;
450}
451static inline long htb_hiwater(const struct htb_class *cl)
452{
453 if (htb_hysteresis)
454 return cl->cmode == HTB_CAN_SEND ? -cl->buffer : 0;
455 else
456 return 0;
457}
458
459
460/**
461 * htb_class_mode - computes and returns current class mode
462 *
463 * It computes cl's mode at time cl->t_c+diff and returns it. If mode
464 * is not HTB_CAN_SEND then cl->pq_key is updated to time difference
465 * from now to time when cl will change its state.
466 * Also it is worth to note that class mode doesn't change simply
467 * at cl->{c,}tokens == 0 but there can rather be hysteresis of
468 * 0 .. -cl->{c,}buffer range. It is meant to limit number of
469 * mode transitions per time unit. The speed gain is about 1/6.
470 */
471static inline enum htb_cmode
472htb_class_mode(struct htb_class *cl, long *diff)
473{
474 long toks;
475
476 if ((toks = (cl->ctokens + *diff)) < htb_lowater(cl)) {
477 *diff = -toks;
478 return HTB_CANT_SEND;
479 }
480
481 if ((toks = (cl->tokens + *diff)) >= htb_hiwater(cl))
482 return HTB_CAN_SEND;
483
484 *diff = -toks;
485 return HTB_MAY_BORROW;
486}
487
488/**
489 * htb_change_class_mode - changes classe's mode
490 *
491 * This should be the only way how to change classe's mode under normal
492 * cirsumstances. Routine will update feed lists linkage, change mode
493 * and add class to the wait event queue if appropriate. New mode should
494 * be different from old one and cl->pq_key has to be valid if changing
495 * to mode other than HTB_CAN_SEND (see htb_add_to_wait_tree).
496 */
497static void
498htb_change_class_mode(struct htb_sched *q, struct htb_class *cl, long *diff)
499{
500 enum htb_cmode new_mode = htb_class_mode(cl, diff);
501
502 if (new_mode == cl->cmode)
503 return;
504
505 if (cl->prio_activity) { /* not necessary: speed optimization */
506 if (cl->cmode != HTB_CANT_SEND)
507 htb_deactivate_prios(q, cl);
508 cl->cmode = new_mode;
509 if (new_mode != HTB_CANT_SEND)
510 htb_activate_prios(q, cl);
511 } else
512 cl->cmode = new_mode;
513}
514
515/**
516 * htb_activate - inserts leaf cl into appropriate active feeds
517 *
518 * Routine learns (new) priority of leaf and activates feed chain
519 * for the prio. It can be called on already active leaf safely.
520 * It also adds leaf into droplist.
521 */
522static inline void htb_activate(struct htb_sched *q, struct htb_class *cl)
523{
524 WARN_ON(cl->level || !cl->un.leaf.q || !cl->un.leaf.q->q.qlen);
525
526 if (!cl->prio_activity) {
527 cl->prio_activity = 1 << cl->prio;
528 htb_activate_prios(q, cl);
529 list_add_tail(&cl->un.leaf.drop_list,
530 q->drops + cl->prio);
531 }
532}
533
534/**
535 * htb_deactivate - remove leaf cl from active feeds
536 *
537 * Make sure that leaf is active. In the other words it can't be called
538 * with non-active leaf. It also removes class from the drop list.
539 */
540static inline void htb_deactivate(struct htb_sched *q, struct htb_class *cl)
541{
542 WARN_ON(!cl->prio_activity);
543
544 htb_deactivate_prios(q, cl);
545 cl->prio_activity = 0;
546 list_del_init(&cl->un.leaf.drop_list);
547}
548
549static int htb_enqueue(struct sk_buff *skb, struct Qdisc *sch)
550{
551 int uninitialized_var(ret);
552 struct htb_sched *q = qdisc_priv(sch);
553 struct htb_class *cl = htb_classify(skb, sch, &ret);
554
555 if (cl == HTB_DIRECT) {
556 /* enqueue to helper queue */
557 if (q->direct_queue.qlen < q->direct_qlen) {
558 __skb_queue_tail(&q->direct_queue, skb);
559 q->direct_pkts++;
560 } else {
561 kfree_skb(skb);
562 sch->qstats.drops++;
563 return NET_XMIT_DROP;
564 }
565#ifdef CONFIG_NET_CLS_ACT
566 } else if (!cl) {
567 if (ret & __NET_XMIT_BYPASS)
568 sch->qstats.drops++;
569 kfree_skb(skb);
570 return ret;
571#endif
572 } else if ((ret = qdisc_enqueue(skb, cl->un.leaf.q)) != NET_XMIT_SUCCESS) {
573 if (net_xmit_drop_count(ret)) {
574 sch->qstats.drops++;
575 cl->qstats.drops++;
576 }
577 return ret;
578 } else {
579 bstats_update(&cl->bstats, skb);
580 htb_activate(q, cl);
581 }
582
583 sch->q.qlen++;
584 return NET_XMIT_SUCCESS;
585}
586
587static inline void htb_accnt_tokens(struct htb_class *cl, int bytes, long diff)
588{
589 long toks = diff + cl->tokens;
590
591 if (toks > cl->buffer)
592 toks = cl->buffer;
593 toks -= (long) qdisc_l2t(cl->rate, bytes);
594 if (toks <= -cl->mbuffer)
595 toks = 1 - cl->mbuffer;
596
597 cl->tokens = toks;
598}
599
600static inline void htb_accnt_ctokens(struct htb_class *cl, int bytes, long diff)
601{
602 long toks = diff + cl->ctokens;
603
604 if (toks > cl->cbuffer)
605 toks = cl->cbuffer;
606 toks -= (long) qdisc_l2t(cl->ceil, bytes);
607 if (toks <= -cl->mbuffer)
608 toks = 1 - cl->mbuffer;
609
610 cl->ctokens = toks;
611}
612
613/**
614 * htb_charge_class - charges amount "bytes" to leaf and ancestors
615 *
616 * Routine assumes that packet "bytes" long was dequeued from leaf cl
617 * borrowing from "level". It accounts bytes to ceil leaky bucket for
618 * leaf and all ancestors and to rate bucket for ancestors at levels
619 * "level" and higher. It also handles possible change of mode resulting
620 * from the update. Note that mode can also increase here (MAY_BORROW to
621 * CAN_SEND) because we can use more precise clock that event queue here.
622 * In such case we remove class from event queue first.
623 */
624static void htb_charge_class(struct htb_sched *q, struct htb_class *cl,
625 int level, struct sk_buff *skb)
626{
627 int bytes = qdisc_pkt_len(skb);
628 enum htb_cmode old_mode;
629 long diff;
630
631 while (cl) {
632 diff = psched_tdiff_bounded(q->now, cl->t_c, cl->mbuffer);
633 if (cl->level >= level) {
634 if (cl->level == level)
635 cl->xstats.lends++;
636 htb_accnt_tokens(cl, bytes, diff);
637 } else {
638 cl->xstats.borrows++;
639 cl->tokens += diff; /* we moved t_c; update tokens */
640 }
641 htb_accnt_ctokens(cl, bytes, diff);
642 cl->t_c = q->now;
643
644 old_mode = cl->cmode;
645 diff = 0;
646 htb_change_class_mode(q, cl, &diff);
647 if (old_mode != cl->cmode) {
648 if (old_mode != HTB_CAN_SEND)
649 htb_safe_rb_erase(&cl->pq_node, q->wait_pq + cl->level);
650 if (cl->cmode != HTB_CAN_SEND)
651 htb_add_to_wait_tree(q, cl, diff);
652 }
653
654 /* update basic stats except for leaves which are already updated */
655 if (cl->level)
656 bstats_update(&cl->bstats, skb);
657
658 cl = cl->parent;
659 }
660}
661
662/**
663 * htb_do_events - make mode changes to classes at the level
664 *
665 * Scans event queue for pending events and applies them. Returns time of
666 * next pending event (0 for no event in pq, q->now for too many events).
667 * Note: Applied are events whose have cl->pq_key <= q->now.
668 */
669static psched_time_t htb_do_events(struct htb_sched *q, int level,
670 unsigned long start)
671{
672 /* don't run for longer than 2 jiffies; 2 is used instead of
673 * 1 to simplify things when jiffy is going to be incremented
674 * too soon
675 */
676 unsigned long stop_at = start + 2;
677 while (time_before(jiffies, stop_at)) {
678 struct htb_class *cl;
679 long diff;
680 struct rb_node *p = rb_first(&q->wait_pq[level]);
681
682 if (!p)
683 return 0;
684
685 cl = rb_entry(p, struct htb_class, pq_node);
686 if (cl->pq_key > q->now)
687 return cl->pq_key;
688
689 htb_safe_rb_erase(p, q->wait_pq + level);
690 diff = psched_tdiff_bounded(q->now, cl->t_c, cl->mbuffer);
691 htb_change_class_mode(q, cl, &diff);
692 if (cl->cmode != HTB_CAN_SEND)
693 htb_add_to_wait_tree(q, cl, diff);
694 }
695
696 /* too much load - let's continue after a break for scheduling */
697 if (!(q->warned & HTB_WARN_TOOMANYEVENTS)) {
698 pr_warning("htb: too many events!\n");
699 q->warned |= HTB_WARN_TOOMANYEVENTS;
700 }
701
702 return q->now;
703}
704
705/* Returns class->node+prio from id-tree where classe's id is >= id. NULL
706 * is no such one exists.
707 */
708static struct rb_node *htb_id_find_next_upper(int prio, struct rb_node *n,
709 u32 id)
710{
711 struct rb_node *r = NULL;
712 while (n) {
713 struct htb_class *cl =
714 rb_entry(n, struct htb_class, node[prio]);
715
716 if (id > cl->common.classid) {
717 n = n->rb_right;
718 } else if (id < cl->common.classid) {
719 r = n;
720 n = n->rb_left;
721 } else {
722 return n;
723 }
724 }
725 return r;
726}
727
728/**
729 * htb_lookup_leaf - returns next leaf class in DRR order
730 *
731 * Find leaf where current feed pointers points to.
732 */
733static struct htb_class *htb_lookup_leaf(struct rb_root *tree, int prio,
734 struct rb_node **pptr, u32 * pid)
735{
736 int i;
737 struct {
738 struct rb_node *root;
739 struct rb_node **pptr;
740 u32 *pid;
741 } stk[TC_HTB_MAXDEPTH], *sp = stk;
742
743 BUG_ON(!tree->rb_node);
744 sp->root = tree->rb_node;
745 sp->pptr = pptr;
746 sp->pid = pid;
747
748 for (i = 0; i < 65535; i++) {
749 if (!*sp->pptr && *sp->pid) {
750 /* ptr was invalidated but id is valid - try to recover
751 * the original or next ptr
752 */
753 *sp->pptr =
754 htb_id_find_next_upper(prio, sp->root, *sp->pid);
755 }
756 *sp->pid = 0; /* ptr is valid now so that remove this hint as it
757 * can become out of date quickly
758 */
759 if (!*sp->pptr) { /* we are at right end; rewind & go up */
760 *sp->pptr = sp->root;
761 while ((*sp->pptr)->rb_left)
762 *sp->pptr = (*sp->pptr)->rb_left;
763 if (sp > stk) {
764 sp--;
765 if (!*sp->pptr) {
766 WARN_ON(1);
767 return NULL;
768 }
769 htb_next_rb_node(sp->pptr);
770 }
771 } else {
772 struct htb_class *cl;
773 cl = rb_entry(*sp->pptr, struct htb_class, node[prio]);
774 if (!cl->level)
775 return cl;
776 (++sp)->root = cl->un.inner.feed[prio].rb_node;
777 sp->pptr = cl->un.inner.ptr + prio;
778 sp->pid = cl->un.inner.last_ptr_id + prio;
779 }
780 }
781 WARN_ON(1);
782 return NULL;
783}
784
785/* dequeues packet at given priority and level; call only if
786 * you are sure that there is active class at prio/level
787 */
788static struct sk_buff *htb_dequeue_tree(struct htb_sched *q, int prio,
789 int level)
790{
791 struct sk_buff *skb = NULL;
792 struct htb_class *cl, *start;
793 /* look initial class up in the row */
794 start = cl = htb_lookup_leaf(q->row[level] + prio, prio,
795 q->ptr[level] + prio,
796 q->last_ptr_id[level] + prio);
797
798 do {
799next:
800 if (unlikely(!cl))
801 return NULL;
802
803 /* class can be empty - it is unlikely but can be true if leaf
804 * qdisc drops packets in enqueue routine or if someone used
805 * graft operation on the leaf since last dequeue;
806 * simply deactivate and skip such class
807 */
808 if (unlikely(cl->un.leaf.q->q.qlen == 0)) {
809 struct htb_class *next;
810 htb_deactivate(q, cl);
811
812 /* row/level might become empty */
813 if ((q->row_mask[level] & (1 << prio)) == 0)
814 return NULL;
815
816 next = htb_lookup_leaf(q->row[level] + prio,
817 prio, q->ptr[level] + prio,
818 q->last_ptr_id[level] + prio);
819
820 if (cl == start) /* fix start if we just deleted it */
821 start = next;
822 cl = next;
823 goto next;
824 }
825
826 skb = cl->un.leaf.q->dequeue(cl->un.leaf.q);
827 if (likely(skb != NULL))
828 break;
829
830 qdisc_warn_nonwc("htb", cl->un.leaf.q);
831 htb_next_rb_node((level ? cl->parent->un.inner.ptr : q->
832 ptr[0]) + prio);
833 cl = htb_lookup_leaf(q->row[level] + prio, prio,
834 q->ptr[level] + prio,
835 q->last_ptr_id[level] + prio);
836
837 } while (cl != start);
838
839 if (likely(skb != NULL)) {
840 cl->un.leaf.deficit[level] -= qdisc_pkt_len(skb);
841 if (cl->un.leaf.deficit[level] < 0) {
842 cl->un.leaf.deficit[level] += cl->quantum;
843 htb_next_rb_node((level ? cl->parent->un.inner.ptr : q->
844 ptr[0]) + prio);
845 }
846 /* this used to be after charge_class but this constelation
847 * gives us slightly better performance
848 */
849 if (!cl->un.leaf.q->q.qlen)
850 htb_deactivate(q, cl);
851 htb_charge_class(q, cl, level, skb);
852 }
853 return skb;
854}
855
856static struct sk_buff *htb_dequeue(struct Qdisc *sch)
857{
858 struct sk_buff *skb;
859 struct htb_sched *q = qdisc_priv(sch);
860 int level;
861 psched_time_t next_event;
862 unsigned long start_at;
863
864 /* try to dequeue direct packets as high prio (!) to minimize cpu work */
865 skb = __skb_dequeue(&q->direct_queue);
866 if (skb != NULL) {
867ok:
868 qdisc_bstats_update(sch, skb);
869 qdisc_unthrottled(sch);
870 sch->q.qlen--;
871 return skb;
872 }
873
874 if (!sch->q.qlen)
875 goto fin;
876 q->now = psched_get_time();
877 start_at = jiffies;
878
879 next_event = q->now + 5 * PSCHED_TICKS_PER_SEC;
880
881 for (level = 0; level < TC_HTB_MAXDEPTH; level++) {
882 /* common case optimization - skip event handler quickly */
883 int m;
884 psched_time_t event;
885
886 if (q->now >= q->near_ev_cache[level]) {
887 event = htb_do_events(q, level, start_at);
888 if (!event)
889 event = q->now + PSCHED_TICKS_PER_SEC;
890 q->near_ev_cache[level] = event;
891 } else
892 event = q->near_ev_cache[level];
893
894 if (next_event > event)
895 next_event = event;
896
897 m = ~q->row_mask[level];
898 while (m != (int)(-1)) {
899 int prio = ffz(m);
900
901 m |= 1 << prio;
902 skb = htb_dequeue_tree(q, prio, level);
903 if (likely(skb != NULL))
904 goto ok;
905 }
906 }
907 sch->qstats.overlimits++;
908 if (likely(next_event > q->now))
909 qdisc_watchdog_schedule(&q->watchdog, next_event);
910 else
911 schedule_work(&q->work);
912fin:
913 return skb;
914}
915
916/* try to drop from each class (by prio) until one succeed */
917static unsigned int htb_drop(struct Qdisc *sch)
918{
919 struct htb_sched *q = qdisc_priv(sch);
920 int prio;
921
922 for (prio = TC_HTB_NUMPRIO - 1; prio >= 0; prio--) {
923 struct list_head *p;
924 list_for_each(p, q->drops + prio) {
925 struct htb_class *cl = list_entry(p, struct htb_class,
926 un.leaf.drop_list);
927 unsigned int len;
928 if (cl->un.leaf.q->ops->drop &&
929 (len = cl->un.leaf.q->ops->drop(cl->un.leaf.q))) {
930 sch->q.qlen--;
931 if (!cl->un.leaf.q->q.qlen)
932 htb_deactivate(q, cl);
933 return len;
934 }
935 }
936 }
937 return 0;
938}
939
940/* reset all classes */
941/* always caled under BH & queue lock */
942static void htb_reset(struct Qdisc *sch)
943{
944 struct htb_sched *q = qdisc_priv(sch);
945 struct htb_class *cl;
946 struct hlist_node *n;
947 unsigned int i;
948
949 for (i = 0; i < q->clhash.hashsize; i++) {
950 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode) {
951 if (cl->level)
952 memset(&cl->un.inner, 0, sizeof(cl->un.inner));
953 else {
954 if (cl->un.leaf.q)
955 qdisc_reset(cl->un.leaf.q);
956 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
957 }
958 cl->prio_activity = 0;
959 cl->cmode = HTB_CAN_SEND;
960
961 }
962 }
963 qdisc_watchdog_cancel(&q->watchdog);
964 __skb_queue_purge(&q->direct_queue);
965 sch->q.qlen = 0;
966 memset(q->row, 0, sizeof(q->row));
967 memset(q->row_mask, 0, sizeof(q->row_mask));
968 memset(q->wait_pq, 0, sizeof(q->wait_pq));
969 memset(q->ptr, 0, sizeof(q->ptr));
970 for (i = 0; i < TC_HTB_NUMPRIO; i++)
971 INIT_LIST_HEAD(q->drops + i);
972}
973
974static const struct nla_policy htb_policy[TCA_HTB_MAX + 1] = {
975 [TCA_HTB_PARMS] = { .len = sizeof(struct tc_htb_opt) },
976 [TCA_HTB_INIT] = { .len = sizeof(struct tc_htb_glob) },
977 [TCA_HTB_CTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
978 [TCA_HTB_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
979};
980
981static void htb_work_func(struct work_struct *work)
982{
983 struct htb_sched *q = container_of(work, struct htb_sched, work);
984 struct Qdisc *sch = q->watchdog.qdisc;
985
986 __netif_schedule(qdisc_root(sch));
987}
988
989static int htb_init(struct Qdisc *sch, struct nlattr *opt)
990{
991 struct htb_sched *q = qdisc_priv(sch);
992 struct nlattr *tb[TCA_HTB_INIT + 1];
993 struct tc_htb_glob *gopt;
994 int err;
995 int i;
996
997 if (!opt)
998 return -EINVAL;
999
1000 err = nla_parse_nested(tb, TCA_HTB_INIT, opt, htb_policy);
1001 if (err < 0)
1002 return err;
1003
1004 if (tb[TCA_HTB_INIT] == NULL) {
1005 pr_err("HTB: hey probably you have bad tc tool ?\n");
1006 return -EINVAL;
1007 }
1008 gopt = nla_data(tb[TCA_HTB_INIT]);
1009 if (gopt->version != HTB_VER >> 16) {
1010 pr_err("HTB: need tc/htb version %d (minor is %d), you have %d\n",
1011 HTB_VER >> 16, HTB_VER & 0xffff, gopt->version);
1012 return -EINVAL;
1013 }
1014
1015 err = qdisc_class_hash_init(&q->clhash);
1016 if (err < 0)
1017 return err;
1018 for (i = 0; i < TC_HTB_NUMPRIO; i++)
1019 INIT_LIST_HEAD(q->drops + i);
1020
1021 qdisc_watchdog_init(&q->watchdog, sch);
1022 INIT_WORK(&q->work, htb_work_func);
1023 skb_queue_head_init(&q->direct_queue);
1024
1025 q->direct_qlen = qdisc_dev(sch)->tx_queue_len;
1026 if (q->direct_qlen < 2) /* some devices have zero tx_queue_len */
1027 q->direct_qlen = 2;
1028
1029 if ((q->rate2quantum = gopt->rate2quantum) < 1)
1030 q->rate2quantum = 1;
1031 q->defcls = gopt->defcls;
1032
1033 return 0;
1034}
1035
1036static int htb_dump(struct Qdisc *sch, struct sk_buff *skb)
1037{
1038 spinlock_t *root_lock = qdisc_root_sleeping_lock(sch);
1039 struct htb_sched *q = qdisc_priv(sch);
1040 struct nlattr *nest;
1041 struct tc_htb_glob gopt;
1042
1043 spin_lock_bh(root_lock);
1044
1045 gopt.direct_pkts = q->direct_pkts;
1046 gopt.version = HTB_VER;
1047 gopt.rate2quantum = q->rate2quantum;
1048 gopt.defcls = q->defcls;
1049 gopt.debug = 0;
1050
1051 nest = nla_nest_start(skb, TCA_OPTIONS);
1052 if (nest == NULL)
1053 goto nla_put_failure;
1054 NLA_PUT(skb, TCA_HTB_INIT, sizeof(gopt), &gopt);
1055 nla_nest_end(skb, nest);
1056
1057 spin_unlock_bh(root_lock);
1058 return skb->len;
1059
1060nla_put_failure:
1061 spin_unlock_bh(root_lock);
1062 nla_nest_cancel(skb, nest);
1063 return -1;
1064}
1065
1066static int htb_dump_class(struct Qdisc *sch, unsigned long arg,
1067 struct sk_buff *skb, struct tcmsg *tcm)
1068{
1069 struct htb_class *cl = (struct htb_class *)arg;
1070 spinlock_t *root_lock = qdisc_root_sleeping_lock(sch);
1071 struct nlattr *nest;
1072 struct tc_htb_opt opt;
1073
1074 spin_lock_bh(root_lock);
1075 tcm->tcm_parent = cl->parent ? cl->parent->common.classid : TC_H_ROOT;
1076 tcm->tcm_handle = cl->common.classid;
1077 if (!cl->level && cl->un.leaf.q)
1078 tcm->tcm_info = cl->un.leaf.q->handle;
1079
1080 nest = nla_nest_start(skb, TCA_OPTIONS);
1081 if (nest == NULL)
1082 goto nla_put_failure;
1083
1084 memset(&opt, 0, sizeof(opt));
1085
1086 opt.rate = cl->rate->rate;
1087 opt.buffer = cl->buffer;
1088 opt.ceil = cl->ceil->rate;
1089 opt.cbuffer = cl->cbuffer;
1090 opt.quantum = cl->quantum;
1091 opt.prio = cl->prio;
1092 opt.level = cl->level;
1093 NLA_PUT(skb, TCA_HTB_PARMS, sizeof(opt), &opt);
1094
1095 nla_nest_end(skb, nest);
1096 spin_unlock_bh(root_lock);
1097 return skb->len;
1098
1099nla_put_failure:
1100 spin_unlock_bh(root_lock);
1101 nla_nest_cancel(skb, nest);
1102 return -1;
1103}
1104
1105static int
1106htb_dump_class_stats(struct Qdisc *sch, unsigned long arg, struct gnet_dump *d)
1107{
1108 struct htb_class *cl = (struct htb_class *)arg;
1109
1110 if (!cl->level && cl->un.leaf.q)
1111 cl->qstats.qlen = cl->un.leaf.q->q.qlen;
1112 cl->xstats.tokens = cl->tokens;
1113 cl->xstats.ctokens = cl->ctokens;
1114
1115 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
1116 gnet_stats_copy_rate_est(d, NULL, &cl->rate_est) < 0 ||
1117 gnet_stats_copy_queue(d, &cl->qstats) < 0)
1118 return -1;
1119
1120 return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1121}
1122
1123static int htb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1124 struct Qdisc **old)
1125{
1126 struct htb_class *cl = (struct htb_class *)arg;
1127
1128 if (cl->level)
1129 return -EINVAL;
1130 if (new == NULL &&
1131 (new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1132 cl->common.classid)) == NULL)
1133 return -ENOBUFS;
1134
1135 sch_tree_lock(sch);
1136 *old = cl->un.leaf.q;
1137 cl->un.leaf.q = new;
1138 if (*old != NULL) {
1139 qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
1140 qdisc_reset(*old);
1141 }
1142 sch_tree_unlock(sch);
1143 return 0;
1144}
1145
1146static struct Qdisc *htb_leaf(struct Qdisc *sch, unsigned long arg)
1147{
1148 struct htb_class *cl = (struct htb_class *)arg;
1149 return !cl->level ? cl->un.leaf.q : NULL;
1150}
1151
1152static void htb_qlen_notify(struct Qdisc *sch, unsigned long arg)
1153{
1154 struct htb_class *cl = (struct htb_class *)arg;
1155
1156 if (cl->un.leaf.q->q.qlen == 0)
1157 htb_deactivate(qdisc_priv(sch), cl);
1158}
1159
1160static unsigned long htb_get(struct Qdisc *sch, u32 classid)
1161{
1162 struct htb_class *cl = htb_find(classid, sch);
1163 if (cl)
1164 cl->refcnt++;
1165 return (unsigned long)cl;
1166}
1167
1168static inline int htb_parent_last_child(struct htb_class *cl)
1169{
1170 if (!cl->parent)
1171 /* the root class */
1172 return 0;
1173 if (cl->parent->children > 1)
1174 /* not the last child */
1175 return 0;
1176 return 1;
1177}
1178
1179static void htb_parent_to_leaf(struct htb_sched *q, struct htb_class *cl,
1180 struct Qdisc *new_q)
1181{
1182 struct htb_class *parent = cl->parent;
1183
1184 WARN_ON(cl->level || !cl->un.leaf.q || cl->prio_activity);
1185
1186 if (parent->cmode != HTB_CAN_SEND)
1187 htb_safe_rb_erase(&parent->pq_node, q->wait_pq + parent->level);
1188
1189 parent->level = 0;
1190 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1191 INIT_LIST_HEAD(&parent->un.leaf.drop_list);
1192 parent->un.leaf.q = new_q ? new_q : &noop_qdisc;
1193 parent->tokens = parent->buffer;
1194 parent->ctokens = parent->cbuffer;
1195 parent->t_c = psched_get_time();
1196 parent->cmode = HTB_CAN_SEND;
1197}
1198
1199static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
1200{
1201 if (!cl->level) {
1202 WARN_ON(!cl->un.leaf.q);
1203 qdisc_destroy(cl->un.leaf.q);
1204 }
1205 gen_kill_estimator(&cl->bstats, &cl->rate_est);
1206 qdisc_put_rtab(cl->rate);
1207 qdisc_put_rtab(cl->ceil);
1208
1209 tcf_destroy_chain(&cl->filter_list);
1210 kfree(cl);
1211}
1212
1213static void htb_destroy(struct Qdisc *sch)
1214{
1215 struct htb_sched *q = qdisc_priv(sch);
1216 struct hlist_node *n, *next;
1217 struct htb_class *cl;
1218 unsigned int i;
1219
1220 cancel_work_sync(&q->work);
1221 qdisc_watchdog_cancel(&q->watchdog);
1222 /* This line used to be after htb_destroy_class call below
1223 * and surprisingly it worked in 2.4. But it must precede it
1224 * because filter need its target class alive to be able to call
1225 * unbind_filter on it (without Oops).
1226 */
1227 tcf_destroy_chain(&q->filter_list);
1228
1229 for (i = 0; i < q->clhash.hashsize; i++) {
1230 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode)
1231 tcf_destroy_chain(&cl->filter_list);
1232 }
1233 for (i = 0; i < q->clhash.hashsize; i++) {
1234 hlist_for_each_entry_safe(cl, n, next, &q->clhash.hash[i],
1235 common.hnode)
1236 htb_destroy_class(sch, cl);
1237 }
1238 qdisc_class_hash_destroy(&q->clhash);
1239 __skb_queue_purge(&q->direct_queue);
1240}
1241
1242static int htb_delete(struct Qdisc *sch, unsigned long arg)
1243{
1244 struct htb_sched *q = qdisc_priv(sch);
1245 struct htb_class *cl = (struct htb_class *)arg;
1246 unsigned int qlen;
1247 struct Qdisc *new_q = NULL;
1248 int last_child = 0;
1249
1250 // TODO: why don't allow to delete subtree ? references ? does
1251 // tc subsys quarantee us that in htb_destroy it holds no class
1252 // refs so that we can remove children safely there ?
1253 if (cl->children || cl->filter_cnt)
1254 return -EBUSY;
1255
1256 if (!cl->level && htb_parent_last_child(cl)) {
1257 new_q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1258 cl->parent->common.classid);
1259 last_child = 1;
1260 }
1261
1262 sch_tree_lock(sch);
1263
1264 if (!cl->level) {
1265 qlen = cl->un.leaf.q->q.qlen;
1266 qdisc_reset(cl->un.leaf.q);
1267 qdisc_tree_decrease_qlen(cl->un.leaf.q, qlen);
1268 }
1269
1270 /* delete from hash and active; remainder in destroy_class */
1271 qdisc_class_hash_remove(&q->clhash, &cl->common);
1272 if (cl->parent)
1273 cl->parent->children--;
1274
1275 if (cl->prio_activity)
1276 htb_deactivate(q, cl);
1277
1278 if (cl->cmode != HTB_CAN_SEND)
1279 htb_safe_rb_erase(&cl->pq_node, q->wait_pq + cl->level);
1280
1281 if (last_child)
1282 htb_parent_to_leaf(q, cl, new_q);
1283
1284 BUG_ON(--cl->refcnt == 0);
1285 /*
1286 * This shouldn't happen: we "hold" one cops->get() when called
1287 * from tc_ctl_tclass; the destroy method is done from cops->put().
1288 */
1289
1290 sch_tree_unlock(sch);
1291 return 0;
1292}
1293
1294static void htb_put(struct Qdisc *sch, unsigned long arg)
1295{
1296 struct htb_class *cl = (struct htb_class *)arg;
1297
1298 if (--cl->refcnt == 0)
1299 htb_destroy_class(sch, cl);
1300}
1301
1302static int htb_change_class(struct Qdisc *sch, u32 classid,
1303 u32 parentid, struct nlattr **tca,
1304 unsigned long *arg)
1305{
1306 int err = -EINVAL;
1307 struct htb_sched *q = qdisc_priv(sch);
1308 struct htb_class *cl = (struct htb_class *)*arg, *parent;
1309 struct nlattr *opt = tca[TCA_OPTIONS];
1310 struct qdisc_rate_table *rtab = NULL, *ctab = NULL;
1311 struct nlattr *tb[__TCA_HTB_MAX];
1312 struct tc_htb_opt *hopt;
1313
1314 /* extract all subattrs from opt attr */
1315 if (!opt)
1316 goto failure;
1317
1318 err = nla_parse_nested(tb, TCA_HTB_MAX, opt, htb_policy);
1319 if (err < 0)
1320 goto failure;
1321
1322 err = -EINVAL;
1323 if (tb[TCA_HTB_PARMS] == NULL)
1324 goto failure;
1325
1326 parent = parentid == TC_H_ROOT ? NULL : htb_find(parentid, sch);
1327
1328 hopt = nla_data(tb[TCA_HTB_PARMS]);
1329
1330 rtab = qdisc_get_rtab(&hopt->rate, tb[TCA_HTB_RTAB]);
1331 ctab = qdisc_get_rtab(&hopt->ceil, tb[TCA_HTB_CTAB]);
1332 if (!rtab || !ctab)
1333 goto failure;
1334
1335 if (!cl) { /* new class */
1336 struct Qdisc *new_q;
1337 int prio;
1338 struct {
1339 struct nlattr nla;
1340 struct gnet_estimator opt;
1341 } est = {
1342 .nla = {
1343 .nla_len = nla_attr_size(sizeof(est.opt)),
1344 .nla_type = TCA_RATE,
1345 },
1346 .opt = {
1347 /* 4s interval, 16s averaging constant */
1348 .interval = 2,
1349 .ewma_log = 2,
1350 },
1351 };
1352
1353 /* check for valid classid */
1354 if (!classid || TC_H_MAJ(classid ^ sch->handle) ||
1355 htb_find(classid, sch))
1356 goto failure;
1357
1358 /* check maximal depth */
1359 if (parent && parent->parent && parent->parent->level < 2) {
1360 pr_err("htb: tree is too deep\n");
1361 goto failure;
1362 }
1363 err = -ENOBUFS;
1364 cl = kzalloc(sizeof(*cl), GFP_KERNEL);
1365 if (!cl)
1366 goto failure;
1367
1368 err = gen_new_estimator(&cl->bstats, &cl->rate_est,
1369 qdisc_root_sleeping_lock(sch),
1370 tca[TCA_RATE] ? : &est.nla);
1371 if (err) {
1372 kfree(cl);
1373 goto failure;
1374 }
1375
1376 cl->refcnt = 1;
1377 cl->children = 0;
1378 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
1379 RB_CLEAR_NODE(&cl->pq_node);
1380
1381 for (prio = 0; prio < TC_HTB_NUMPRIO; prio++)
1382 RB_CLEAR_NODE(&cl->node[prio]);
1383
1384 /* create leaf qdisc early because it uses kmalloc(GFP_KERNEL)
1385 * so that can't be used inside of sch_tree_lock
1386 * -- thanks to Karlis Peisenieks
1387 */
1388 new_q = qdisc_create_dflt(sch->dev_queue,
1389 &pfifo_qdisc_ops, classid);
1390 sch_tree_lock(sch);
1391 if (parent && !parent->level) {
1392 unsigned int qlen = parent->un.leaf.q->q.qlen;
1393
1394 /* turn parent into inner node */
1395 qdisc_reset(parent->un.leaf.q);
1396 qdisc_tree_decrease_qlen(parent->un.leaf.q, qlen);
1397 qdisc_destroy(parent->un.leaf.q);
1398 if (parent->prio_activity)
1399 htb_deactivate(q, parent);
1400
1401 /* remove from evt list because of level change */
1402 if (parent->cmode != HTB_CAN_SEND) {
1403 htb_safe_rb_erase(&parent->pq_node, q->wait_pq);
1404 parent->cmode = HTB_CAN_SEND;
1405 }
1406 parent->level = (parent->parent ? parent->parent->level
1407 : TC_HTB_MAXDEPTH) - 1;
1408 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1409 }
1410 /* leaf (we) needs elementary qdisc */
1411 cl->un.leaf.q = new_q ? new_q : &noop_qdisc;
1412
1413 cl->common.classid = classid;
1414 cl->parent = parent;
1415
1416 /* set class to be in HTB_CAN_SEND state */
1417 cl->tokens = hopt->buffer;
1418 cl->ctokens = hopt->cbuffer;
1419 cl->mbuffer = 60 * PSCHED_TICKS_PER_SEC; /* 1min */
1420 cl->t_c = psched_get_time();
1421 cl->cmode = HTB_CAN_SEND;
1422
1423 /* attach to the hash list and parent's family */
1424 qdisc_class_hash_insert(&q->clhash, &cl->common);
1425 if (parent)
1426 parent->children++;
1427 } else {
1428 if (tca[TCA_RATE]) {
1429 err = gen_replace_estimator(&cl->bstats, &cl->rate_est,
1430 qdisc_root_sleeping_lock(sch),
1431 tca[TCA_RATE]);
1432 if (err)
1433 return err;
1434 }
1435 sch_tree_lock(sch);
1436 }
1437
1438 /* it used to be a nasty bug here, we have to check that node
1439 * is really leaf before changing cl->un.leaf !
1440 */
1441 if (!cl->level) {
1442 cl->quantum = rtab->rate.rate / q->rate2quantum;
1443 if (!hopt->quantum && cl->quantum < 1000) {
1444 pr_warning(
1445 "HTB: quantum of class %X is small. Consider r2q change.\n",
1446 cl->common.classid);
1447 cl->quantum = 1000;
1448 }
1449 if (!hopt->quantum && cl->quantum > 200000) {
1450 pr_warning(
1451 "HTB: quantum of class %X is big. Consider r2q change.\n",
1452 cl->common.classid);
1453 cl->quantum = 200000;
1454 }
1455 if (hopt->quantum)
1456 cl->quantum = hopt->quantum;
1457 if ((cl->prio = hopt->prio) >= TC_HTB_NUMPRIO)
1458 cl->prio = TC_HTB_NUMPRIO - 1;
1459 }
1460
1461 cl->buffer = hopt->buffer;
1462 cl->cbuffer = hopt->cbuffer;
1463 if (cl->rate)
1464 qdisc_put_rtab(cl->rate);
1465 cl->rate = rtab;
1466 if (cl->ceil)
1467 qdisc_put_rtab(cl->ceil);
1468 cl->ceil = ctab;
1469 sch_tree_unlock(sch);
1470
1471 qdisc_class_hash_grow(sch, &q->clhash);
1472
1473 *arg = (unsigned long)cl;
1474 return 0;
1475
1476failure:
1477 if (rtab)
1478 qdisc_put_rtab(rtab);
1479 if (ctab)
1480 qdisc_put_rtab(ctab);
1481 return err;
1482}
1483
1484static struct tcf_proto **htb_find_tcf(struct Qdisc *sch, unsigned long arg)
1485{
1486 struct htb_sched *q = qdisc_priv(sch);
1487 struct htb_class *cl = (struct htb_class *)arg;
1488 struct tcf_proto **fl = cl ? &cl->filter_list : &q->filter_list;
1489
1490 return fl;
1491}
1492
1493static unsigned long htb_bind_filter(struct Qdisc *sch, unsigned long parent,
1494 u32 classid)
1495{
1496 struct htb_class *cl = htb_find(classid, sch);
1497
1498 /*if (cl && !cl->level) return 0;
1499 * The line above used to be there to prevent attaching filters to
1500 * leaves. But at least tc_index filter uses this just to get class
1501 * for other reasons so that we have to allow for it.
1502 * ----
1503 * 19.6.2002 As Werner explained it is ok - bind filter is just
1504 * another way to "lock" the class - unlike "get" this lock can
1505 * be broken by class during destroy IIUC.
1506 */
1507 if (cl)
1508 cl->filter_cnt++;
1509 return (unsigned long)cl;
1510}
1511
1512static void htb_unbind_filter(struct Qdisc *sch, unsigned long arg)
1513{
1514 struct htb_class *cl = (struct htb_class *)arg;
1515
1516 if (cl)
1517 cl->filter_cnt--;
1518}
1519
1520static void htb_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1521{
1522 struct htb_sched *q = qdisc_priv(sch);
1523 struct htb_class *cl;
1524 struct hlist_node *n;
1525 unsigned int i;
1526
1527 if (arg->stop)
1528 return;
1529
1530 for (i = 0; i < q->clhash.hashsize; i++) {
1531 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode) {
1532 if (arg->count < arg->skip) {
1533 arg->count++;
1534 continue;
1535 }
1536 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1537 arg->stop = 1;
1538 return;
1539 }
1540 arg->count++;
1541 }
1542 }
1543}
1544
1545static const struct Qdisc_class_ops htb_class_ops = {
1546 .graft = htb_graft,
1547 .leaf = htb_leaf,
1548 .qlen_notify = htb_qlen_notify,
1549 .get = htb_get,
1550 .put = htb_put,
1551 .change = htb_change_class,
1552 .delete = htb_delete,
1553 .walk = htb_walk,
1554 .tcf_chain = htb_find_tcf,
1555 .bind_tcf = htb_bind_filter,
1556 .unbind_tcf = htb_unbind_filter,
1557 .dump = htb_dump_class,
1558 .dump_stats = htb_dump_class_stats,
1559};
1560
1561static struct Qdisc_ops htb_qdisc_ops __read_mostly = {
1562 .cl_ops = &htb_class_ops,
1563 .id = "htb",
1564 .priv_size = sizeof(struct htb_sched),
1565 .enqueue = htb_enqueue,
1566 .dequeue = htb_dequeue,
1567 .peek = qdisc_peek_dequeued,
1568 .drop = htb_drop,
1569 .init = htb_init,
1570 .reset = htb_reset,
1571 .destroy = htb_destroy,
1572 .dump = htb_dump,
1573 .owner = THIS_MODULE,
1574};
1575
1576static int __init htb_module_init(void)
1577{
1578 return register_qdisc(&htb_qdisc_ops);
1579}
1580static void __exit htb_module_exit(void)
1581{
1582 unregister_qdisc(&htb_qdisc_ops);
1583}
1584
1585module_init(htb_module_init)
1586module_exit(htb_module_exit)
1587MODULE_LICENSE("GPL");
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * net/sched/sch_htb.c Hierarchical token bucket, feed tree version
4 *
5 * Authors: Martin Devera, <devik@cdi.cz>
6 *
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
18 * Wilfried Weissmann
19 * spotted bug in dequeue code and helped with fix
20 * Jiri Fojtasek
21 * fixed requeue routine
22 * and many others. thanks.
23 */
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>
40
41/* HTB algorithm.
42 Author: devik@cdi.cz
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.
47
48 Levels:
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.
52*/
53
54static 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 */
56
57#if HTB_VER >> 16 != TC_HTB_PROTOVER
58#error "Mismatched sch_htb.c and pkt_sch.h"
59#endif
60
61/* Module parameter and sysfs export */
62module_param (htb_hysteresis, int, 0640);
63MODULE_PARM_DESC(htb_hysteresis, "Hysteresis mode, less CPU load, less accurate");
64
65static int htb_rate_est = 0; /* htb classes have a default rate estimator */
66module_param(htb_rate_est, int, 0640);
67MODULE_PARM_DESC(htb_rate_est, "setup a default rate estimator (4sec 16sec) for htb classes");
68
69/* used internaly to keep status of single class */
70enum htb_cmode {
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 */
74};
75
76struct htb_prio {
77 union {
78 struct rb_root row;
79 struct rb_root feed;
80 };
81 struct rb_node *ptr;
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).
86 */
87 u32 last_ptr_id;
88};
89
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.
93 */
94struct htb_class {
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 */
102
103 struct tcf_proto __rcu *filter_list; /* class attached filters */
104 struct tcf_block *block;
105
106 int level; /* our level (see above) */
107 unsigned int children;
108 struct htb_class *parent; /* parent class */
109
110 struct net_rate_estimator __rcu *rate_est;
111
112 /*
113 * Written often fields
114 */
115 struct gnet_stats_basic_sync bstats;
116 struct gnet_stats_basic_sync bstats_bias;
117 struct tc_htb_xstats xstats; /* our special stats */
118
119 /* token bucket parameters */
120 s64 tokens, ctokens;/* current number of tokens */
121 s64 t_c; /* checkpoint time */
122
123 union {
124 struct htb_class_leaf {
125 int deficit[TC_HTB_MAXDEPTH];
126 struct Qdisc *q;
127 struct netdev_queue *offload_queue;
128 } leaf;
129 struct htb_class_inner {
130 struct htb_prio clprio[TC_HTB_NUMPRIO];
131 } inner;
132 };
133 s64 pq_key;
134
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 */
139
140 unsigned int drops ____cacheline_aligned_in_smp;
141 unsigned int overlimits;
142};
143
144struct htb_level {
145 struct rb_root wait_pq;
146 struct htb_prio hprio[TC_HTB_NUMPRIO];
147};
148
149struct htb_sched {
150 struct Qdisc_class_hash clhash;
151 int defcls; /* class where unclassified flows go to */
152 int rate2quantum; /* quant = rate / rate2quantum */
153
154 /* filters for qdisc itself */
155 struct tcf_proto __rcu *filter_list;
156 struct tcf_block *block;
157
158#define HTB_WARN_TOOMANYEVENTS 0x1
159 unsigned int warned; /* only one warning */
160 int direct_qlen;
161 struct work_struct work;
162
163 /* non shaped skbs; let them go directly thru */
164 struct qdisc_skb_head direct_queue;
165 u32 direct_pkts;
166 u32 overlimits;
167
168 struct qdisc_watchdog watchdog;
169
170 s64 now; /* cached dequeue time */
171
172 /* time of nearest event per level (row) */
173 s64 near_ev_cache[TC_HTB_MAXDEPTH];
174
175 int row_mask[TC_HTB_MAXDEPTH];
176
177 struct htb_level hlevel[TC_HTB_MAXDEPTH];
178
179 struct Qdisc **direct_qdiscs;
180 unsigned int num_direct_qdiscs;
181
182 bool offload;
183};
184
185/* find class in global hash table using given handle */
186static inline struct htb_class *htb_find(u32 handle, struct Qdisc *sch)
187{
188 struct htb_sched *q = qdisc_priv(sch);
189 struct Qdisc_class_common *clc;
190
191 clc = qdisc_class_find(&q->clhash, handle);
192 if (clc == NULL)
193 return NULL;
194 return container_of(clc, struct htb_class, common);
195}
196
197static unsigned long htb_search(struct Qdisc *sch, u32 handle)
198{
199 return (unsigned long)htb_find(handle, sch);
200}
201
202#define HTB_DIRECT ((struct htb_class *)-1L)
203
204/**
205 * htb_classify - classify a packet into class
206 * @skb: the socket buffer
207 * @sch: the active queue discipline
208 * @qerr: pointer for returned status code
209 *
210 * It returns NULL if the packet should be dropped or -1 if the packet
211 * should be passed directly thru. In all other cases leaf class is returned.
212 * We allow direct class selection by classid in priority. The we examine
213 * filters in qdisc and in inner nodes (if higher filter points to the inner
214 * node). If we end up with classid MAJOR:0 we enqueue the skb into special
215 * internal fifo (direct). These packets then go directly thru. If we still
216 * have no valid leaf we try to use MAJOR:default leaf. It still unsuccessful
217 * then finish and return direct queue.
218 */
219static struct htb_class *htb_classify(struct sk_buff *skb, struct Qdisc *sch,
220 int *qerr)
221{
222 struct htb_sched *q = qdisc_priv(sch);
223 struct htb_class *cl;
224 struct tcf_result res;
225 struct tcf_proto *tcf;
226 int result;
227
228 /* allow to select class by setting skb->priority to valid classid;
229 * note that nfmark can be used too by attaching filter fw with no
230 * rules in it
231 */
232 if (skb->priority == sch->handle)
233 return HTB_DIRECT; /* X:0 (direct flow) selected */
234 cl = htb_find(skb->priority, sch);
235 if (cl) {
236 if (cl->level == 0)
237 return cl;
238 /* Start with inner filter chain if a non-leaf class is selected */
239 tcf = rcu_dereference_bh(cl->filter_list);
240 } else {
241 tcf = rcu_dereference_bh(q->filter_list);
242 }
243
244 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
245 while (tcf && (result = tcf_classify(skb, NULL, tcf, &res, false)) >= 0) {
246#ifdef CONFIG_NET_CLS_ACT
247 switch (result) {
248 case TC_ACT_QUEUED:
249 case TC_ACT_STOLEN:
250 case TC_ACT_TRAP:
251 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
252 fallthrough;
253 case TC_ACT_SHOT:
254 return NULL;
255 }
256#endif
257 cl = (void *)res.class;
258 if (!cl) {
259 if (res.classid == sch->handle)
260 return HTB_DIRECT; /* X:0 (direct flow) */
261 cl = htb_find(res.classid, sch);
262 if (!cl)
263 break; /* filter selected invalid classid */
264 }
265 if (!cl->level)
266 return cl; /* we hit leaf; return it */
267
268 /* we have got inner class; apply inner filter chain */
269 tcf = rcu_dereference_bh(cl->filter_list);
270 }
271 /* classification failed; try to use default class */
272 cl = htb_find(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
273 if (!cl || cl->level)
274 return HTB_DIRECT; /* bad default .. this is safe bet */
275 return cl;
276}
277
278/**
279 * htb_add_to_id_tree - adds class to the round robin list
280 * @root: the root of the tree
281 * @cl: the class to add
282 * @prio: the give prio in class
283 *
284 * Routine adds class to the list (actually tree) sorted by classid.
285 * Make sure that class is not already on such list for given prio.
286 */
287static void htb_add_to_id_tree(struct rb_root *root,
288 struct htb_class *cl, int prio)
289{
290 struct rb_node **p = &root->rb_node, *parent = NULL;
291
292 while (*p) {
293 struct htb_class *c;
294 parent = *p;
295 c = rb_entry(parent, struct htb_class, node[prio]);
296
297 if (cl->common.classid > c->common.classid)
298 p = &parent->rb_right;
299 else
300 p = &parent->rb_left;
301 }
302 rb_link_node(&cl->node[prio], parent, p);
303 rb_insert_color(&cl->node[prio], root);
304}
305
306/**
307 * htb_add_to_wait_tree - adds class to the event queue with delay
308 * @q: the priority event queue
309 * @cl: the class to add
310 * @delay: delay in microseconds
311 *
312 * The class is added to priority event queue to indicate that class will
313 * change its mode in cl->pq_key microseconds. Make sure that class is not
314 * already in the queue.
315 */
316static void htb_add_to_wait_tree(struct htb_sched *q,
317 struct htb_class *cl, s64 delay)
318{
319 struct rb_node **p = &q->hlevel[cl->level].wait_pq.rb_node, *parent = NULL;
320
321 cl->pq_key = q->now + delay;
322 if (cl->pq_key == q->now)
323 cl->pq_key++;
324
325 /* update the nearest event cache */
326 if (q->near_ev_cache[cl->level] > cl->pq_key)
327 q->near_ev_cache[cl->level] = cl->pq_key;
328
329 while (*p) {
330 struct htb_class *c;
331 parent = *p;
332 c = rb_entry(parent, struct htb_class, pq_node);
333 if (cl->pq_key >= c->pq_key)
334 p = &parent->rb_right;
335 else
336 p = &parent->rb_left;
337 }
338 rb_link_node(&cl->pq_node, parent, p);
339 rb_insert_color(&cl->pq_node, &q->hlevel[cl->level].wait_pq);
340}
341
342/**
343 * htb_next_rb_node - finds next node in binary tree
344 * @n: the current node in binary tree
345 *
346 * When we are past last key we return NULL.
347 * Average complexity is 2 steps per call.
348 */
349static inline void htb_next_rb_node(struct rb_node **n)
350{
351 *n = rb_next(*n);
352}
353
354/**
355 * htb_add_class_to_row - add class to its row
356 * @q: the priority event queue
357 * @cl: the class to add
358 * @mask: the given priorities in class in bitmap
359 *
360 * The class is added to row at priorities marked in mask.
361 * It does nothing if mask == 0.
362 */
363static inline void htb_add_class_to_row(struct htb_sched *q,
364 struct htb_class *cl, int mask)
365{
366 q->row_mask[cl->level] |= mask;
367 while (mask) {
368 int prio = ffz(~mask);
369 mask &= ~(1 << prio);
370 htb_add_to_id_tree(&q->hlevel[cl->level].hprio[prio].row, cl, prio);
371 }
372}
373
374/* If this triggers, it is a bug in this code, but it need not be fatal */
375static void htb_safe_rb_erase(struct rb_node *rb, struct rb_root *root)
376{
377 if (RB_EMPTY_NODE(rb)) {
378 WARN_ON(1);
379 } else {
380 rb_erase(rb, root);
381 RB_CLEAR_NODE(rb);
382 }
383}
384
385
386/**
387 * htb_remove_class_from_row - removes class from its row
388 * @q: the priority event queue
389 * @cl: the class to add
390 * @mask: the given priorities in class in bitmap
391 *
392 * The class is removed from row at priorities marked in mask.
393 * It does nothing if mask == 0.
394 */
395static inline void htb_remove_class_from_row(struct htb_sched *q,
396 struct htb_class *cl, int mask)
397{
398 int m = 0;
399 struct htb_level *hlevel = &q->hlevel[cl->level];
400
401 while (mask) {
402 int prio = ffz(~mask);
403 struct htb_prio *hprio = &hlevel->hprio[prio];
404
405 mask &= ~(1 << prio);
406 if (hprio->ptr == cl->node + prio)
407 htb_next_rb_node(&hprio->ptr);
408
409 htb_safe_rb_erase(cl->node + prio, &hprio->row);
410 if (!hprio->row.rb_node)
411 m |= 1 << prio;
412 }
413 q->row_mask[cl->level] &= ~m;
414}
415
416/**
417 * htb_activate_prios - creates active classe's feed chain
418 * @q: the priority event queue
419 * @cl: the class to activate
420 *
421 * The class is connected to ancestors and/or appropriate rows
422 * for priorities it is participating on. cl->cmode must be new
423 * (activated) mode. It does nothing if cl->prio_activity == 0.
424 */
425static void htb_activate_prios(struct htb_sched *q, struct htb_class *cl)
426{
427 struct htb_class *p = cl->parent;
428 long m, mask = cl->prio_activity;
429
430 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
431 m = mask;
432 while (m) {
433 unsigned int prio = ffz(~m);
434
435 if (WARN_ON_ONCE(prio >= ARRAY_SIZE(p->inner.clprio)))
436 break;
437 m &= ~(1 << prio);
438
439 if (p->inner.clprio[prio].feed.rb_node)
440 /* parent already has its feed in use so that
441 * reset bit in mask as parent is already ok
442 */
443 mask &= ~(1 << prio);
444
445 htb_add_to_id_tree(&p->inner.clprio[prio].feed, cl, prio);
446 }
447 p->prio_activity |= mask;
448 cl = p;
449 p = cl->parent;
450
451 }
452 if (cl->cmode == HTB_CAN_SEND && mask)
453 htb_add_class_to_row(q, cl, mask);
454}
455
456/**
457 * htb_deactivate_prios - remove class from feed chain
458 * @q: the priority event queue
459 * @cl: the class to deactivate
460 *
461 * cl->cmode must represent old mode (before deactivation). It does
462 * nothing if cl->prio_activity == 0. Class is removed from all feed
463 * chains and rows.
464 */
465static void htb_deactivate_prios(struct htb_sched *q, struct htb_class *cl)
466{
467 struct htb_class *p = cl->parent;
468 long m, mask = cl->prio_activity;
469
470 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
471 m = mask;
472 mask = 0;
473 while (m) {
474 int prio = ffz(~m);
475 m &= ~(1 << prio);
476
477 if (p->inner.clprio[prio].ptr == cl->node + prio) {
478 /* we are removing child which is pointed to from
479 * parent feed - forget the pointer but remember
480 * classid
481 */
482 p->inner.clprio[prio].last_ptr_id = cl->common.classid;
483 p->inner.clprio[prio].ptr = NULL;
484 }
485
486 htb_safe_rb_erase(cl->node + prio,
487 &p->inner.clprio[prio].feed);
488
489 if (!p->inner.clprio[prio].feed.rb_node)
490 mask |= 1 << prio;
491 }
492
493 p->prio_activity &= ~mask;
494 cl = p;
495 p = cl->parent;
496
497 }
498 if (cl->cmode == HTB_CAN_SEND && mask)
499 htb_remove_class_from_row(q, cl, mask);
500}
501
502static inline s64 htb_lowater(const struct htb_class *cl)
503{
504 if (htb_hysteresis)
505 return cl->cmode != HTB_CANT_SEND ? -cl->cbuffer : 0;
506 else
507 return 0;
508}
509static inline s64 htb_hiwater(const struct htb_class *cl)
510{
511 if (htb_hysteresis)
512 return cl->cmode == HTB_CAN_SEND ? -cl->buffer : 0;
513 else
514 return 0;
515}
516
517
518/**
519 * htb_class_mode - computes and returns current class mode
520 * @cl: the target class
521 * @diff: diff time in microseconds
522 *
523 * It computes cl's mode at time cl->t_c+diff and returns it. If mode
524 * is not HTB_CAN_SEND then cl->pq_key is updated to time difference
525 * from now to time when cl will change its state.
526 * Also it is worth to note that class mode doesn't change simply
527 * at cl->{c,}tokens == 0 but there can rather be hysteresis of
528 * 0 .. -cl->{c,}buffer range. It is meant to limit number of
529 * mode transitions per time unit. The speed gain is about 1/6.
530 */
531static inline enum htb_cmode
532htb_class_mode(struct htb_class *cl, s64 *diff)
533{
534 s64 toks;
535
536 if ((toks = (cl->ctokens + *diff)) < htb_lowater(cl)) {
537 *diff = -toks;
538 return HTB_CANT_SEND;
539 }
540
541 if ((toks = (cl->tokens + *diff)) >= htb_hiwater(cl))
542 return HTB_CAN_SEND;
543
544 *diff = -toks;
545 return HTB_MAY_BORROW;
546}
547
548/**
549 * htb_change_class_mode - changes classe's mode
550 * @q: the priority event queue
551 * @cl: the target class
552 * @diff: diff time in microseconds
553 *
554 * This should be the only way how to change classe's mode under normal
555 * circumstances. Routine will update feed lists linkage, change mode
556 * and add class to the wait event queue if appropriate. New mode should
557 * be different from old one and cl->pq_key has to be valid if changing
558 * to mode other than HTB_CAN_SEND (see htb_add_to_wait_tree).
559 */
560static void
561htb_change_class_mode(struct htb_sched *q, struct htb_class *cl, s64 *diff)
562{
563 enum htb_cmode new_mode = htb_class_mode(cl, diff);
564
565 if (new_mode == cl->cmode)
566 return;
567
568 if (new_mode == HTB_CANT_SEND) {
569 cl->overlimits++;
570 q->overlimits++;
571 }
572
573 if (cl->prio_activity) { /* not necessary: speed optimization */
574 if (cl->cmode != HTB_CANT_SEND)
575 htb_deactivate_prios(q, cl);
576 cl->cmode = new_mode;
577 if (new_mode != HTB_CANT_SEND)
578 htb_activate_prios(q, cl);
579 } else
580 cl->cmode = new_mode;
581}
582
583/**
584 * htb_activate - inserts leaf cl into appropriate active feeds
585 * @q: the priority event queue
586 * @cl: the target class
587 *
588 * Routine learns (new) priority of leaf and activates feed chain
589 * for the prio. It can be called on already active leaf safely.
590 * It also adds leaf into droplist.
591 */
592static inline void htb_activate(struct htb_sched *q, struct htb_class *cl)
593{
594 WARN_ON(cl->level || !cl->leaf.q || !cl->leaf.q->q.qlen);
595
596 if (!cl->prio_activity) {
597 cl->prio_activity = 1 << cl->prio;
598 htb_activate_prios(q, cl);
599 }
600}
601
602/**
603 * htb_deactivate - remove leaf cl from active feeds
604 * @q: the priority event queue
605 * @cl: the target class
606 *
607 * Make sure that leaf is active. In the other words it can't be called
608 * with non-active leaf. It also removes class from the drop list.
609 */
610static inline void htb_deactivate(struct htb_sched *q, struct htb_class *cl)
611{
612 WARN_ON(!cl->prio_activity);
613
614 htb_deactivate_prios(q, cl);
615 cl->prio_activity = 0;
616}
617
618static int htb_enqueue(struct sk_buff *skb, struct Qdisc *sch,
619 struct sk_buff **to_free)
620{
621 int ret;
622 unsigned int len = qdisc_pkt_len(skb);
623 struct htb_sched *q = qdisc_priv(sch);
624 struct htb_class *cl = htb_classify(skb, sch, &ret);
625
626 if (cl == HTB_DIRECT) {
627 /* enqueue to helper queue */
628 if (q->direct_queue.qlen < q->direct_qlen) {
629 __qdisc_enqueue_tail(skb, &q->direct_queue);
630 q->direct_pkts++;
631 } else {
632 return qdisc_drop(skb, sch, to_free);
633 }
634#ifdef CONFIG_NET_CLS_ACT
635 } else if (!cl) {
636 if (ret & __NET_XMIT_BYPASS)
637 qdisc_qstats_drop(sch);
638 __qdisc_drop(skb, to_free);
639 return ret;
640#endif
641 } else if ((ret = qdisc_enqueue(skb, cl->leaf.q,
642 to_free)) != NET_XMIT_SUCCESS) {
643 if (net_xmit_drop_count(ret)) {
644 qdisc_qstats_drop(sch);
645 cl->drops++;
646 }
647 return ret;
648 } else {
649 htb_activate(q, cl);
650 }
651
652 sch->qstats.backlog += len;
653 sch->q.qlen++;
654 return NET_XMIT_SUCCESS;
655}
656
657static inline void htb_accnt_tokens(struct htb_class *cl, int bytes, s64 diff)
658{
659 s64 toks = diff + cl->tokens;
660
661 if (toks > cl->buffer)
662 toks = cl->buffer;
663 toks -= (s64) psched_l2t_ns(&cl->rate, bytes);
664 if (toks <= -cl->mbuffer)
665 toks = 1 - cl->mbuffer;
666
667 cl->tokens = toks;
668}
669
670static inline void htb_accnt_ctokens(struct htb_class *cl, int bytes, s64 diff)
671{
672 s64 toks = diff + cl->ctokens;
673
674 if (toks > cl->cbuffer)
675 toks = cl->cbuffer;
676 toks -= (s64) psched_l2t_ns(&cl->ceil, bytes);
677 if (toks <= -cl->mbuffer)
678 toks = 1 - cl->mbuffer;
679
680 cl->ctokens = toks;
681}
682
683/**
684 * htb_charge_class - charges amount "bytes" to leaf and ancestors
685 * @q: the priority event queue
686 * @cl: the class to start iterate
687 * @level: the minimum level to account
688 * @skb: the socket buffer
689 *
690 * Routine assumes that packet "bytes" long was dequeued from leaf cl
691 * borrowing from "level". It accounts bytes to ceil leaky bucket for
692 * leaf and all ancestors and to rate bucket for ancestors at levels
693 * "level" and higher. It also handles possible change of mode resulting
694 * from the update. Note that mode can also increase here (MAY_BORROW to
695 * CAN_SEND) because we can use more precise clock that event queue here.
696 * In such case we remove class from event queue first.
697 */
698static void htb_charge_class(struct htb_sched *q, struct htb_class *cl,
699 int level, struct sk_buff *skb)
700{
701 int bytes = qdisc_pkt_len(skb);
702 enum htb_cmode old_mode;
703 s64 diff;
704
705 while (cl) {
706 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
707 if (cl->level >= level) {
708 if (cl->level == level)
709 cl->xstats.lends++;
710 htb_accnt_tokens(cl, bytes, diff);
711 } else {
712 cl->xstats.borrows++;
713 cl->tokens += diff; /* we moved t_c; update tokens */
714 }
715 htb_accnt_ctokens(cl, bytes, diff);
716 cl->t_c = q->now;
717
718 old_mode = cl->cmode;
719 diff = 0;
720 htb_change_class_mode(q, cl, &diff);
721 if (old_mode != cl->cmode) {
722 if (old_mode != HTB_CAN_SEND)
723 htb_safe_rb_erase(&cl->pq_node, &q->hlevel[cl->level].wait_pq);
724 if (cl->cmode != HTB_CAN_SEND)
725 htb_add_to_wait_tree(q, cl, diff);
726 }
727
728 /* update basic stats except for leaves which are already updated */
729 if (cl->level)
730 bstats_update(&cl->bstats, skb);
731
732 cl = cl->parent;
733 }
734}
735
736/**
737 * htb_do_events - make mode changes to classes at the level
738 * @q: the priority event queue
739 * @level: which wait_pq in 'q->hlevel'
740 * @start: start jiffies
741 *
742 * Scans event queue for pending events and applies them. Returns time of
743 * next pending event (0 for no event in pq, q->now for too many events).
744 * Note: Applied are events whose have cl->pq_key <= q->now.
745 */
746static s64 htb_do_events(struct htb_sched *q, const int level,
747 unsigned long start)
748{
749 /* don't run for longer than 2 jiffies; 2 is used instead of
750 * 1 to simplify things when jiffy is going to be incremented
751 * too soon
752 */
753 unsigned long stop_at = start + 2;
754 struct rb_root *wait_pq = &q->hlevel[level].wait_pq;
755
756 while (time_before(jiffies, stop_at)) {
757 struct htb_class *cl;
758 s64 diff;
759 struct rb_node *p = rb_first(wait_pq);
760
761 if (!p)
762 return 0;
763
764 cl = rb_entry(p, struct htb_class, pq_node);
765 if (cl->pq_key > q->now)
766 return cl->pq_key;
767
768 htb_safe_rb_erase(p, wait_pq);
769 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
770 htb_change_class_mode(q, cl, &diff);
771 if (cl->cmode != HTB_CAN_SEND)
772 htb_add_to_wait_tree(q, cl, diff);
773 }
774
775 /* too much load - let's continue after a break for scheduling */
776 if (!(q->warned & HTB_WARN_TOOMANYEVENTS)) {
777 pr_warn("htb: too many events!\n");
778 q->warned |= HTB_WARN_TOOMANYEVENTS;
779 }
780
781 return q->now;
782}
783
784/* Returns class->node+prio from id-tree where classe's id is >= id. NULL
785 * is no such one exists.
786 */
787static struct rb_node *htb_id_find_next_upper(int prio, struct rb_node *n,
788 u32 id)
789{
790 struct rb_node *r = NULL;
791 while (n) {
792 struct htb_class *cl =
793 rb_entry(n, struct htb_class, node[prio]);
794
795 if (id > cl->common.classid) {
796 n = n->rb_right;
797 } else if (id < cl->common.classid) {
798 r = n;
799 n = n->rb_left;
800 } else {
801 return n;
802 }
803 }
804 return r;
805}
806
807/**
808 * htb_lookup_leaf - returns next leaf class in DRR order
809 * @hprio: the current one
810 * @prio: which prio in class
811 *
812 * Find leaf where current feed pointers points to.
813 */
814static struct htb_class *htb_lookup_leaf(struct htb_prio *hprio, const int prio)
815{
816 int i;
817 struct {
818 struct rb_node *root;
819 struct rb_node **pptr;
820 u32 *pid;
821 } stk[TC_HTB_MAXDEPTH], *sp = stk;
822
823 BUG_ON(!hprio->row.rb_node);
824 sp->root = hprio->row.rb_node;
825 sp->pptr = &hprio->ptr;
826 sp->pid = &hprio->last_ptr_id;
827
828 for (i = 0; i < 65535; i++) {
829 if (!*sp->pptr && *sp->pid) {
830 /* ptr was invalidated but id is valid - try to recover
831 * the original or next ptr
832 */
833 *sp->pptr =
834 htb_id_find_next_upper(prio, sp->root, *sp->pid);
835 }
836 *sp->pid = 0; /* ptr is valid now so that remove this hint as it
837 * can become out of date quickly
838 */
839 if (!*sp->pptr) { /* we are at right end; rewind & go up */
840 *sp->pptr = sp->root;
841 while ((*sp->pptr)->rb_left)
842 *sp->pptr = (*sp->pptr)->rb_left;
843 if (sp > stk) {
844 sp--;
845 if (!*sp->pptr) {
846 WARN_ON(1);
847 return NULL;
848 }
849 htb_next_rb_node(sp->pptr);
850 }
851 } else {
852 struct htb_class *cl;
853 struct htb_prio *clp;
854
855 cl = rb_entry(*sp->pptr, struct htb_class, node[prio]);
856 if (!cl->level)
857 return cl;
858 clp = &cl->inner.clprio[prio];
859 (++sp)->root = clp->feed.rb_node;
860 sp->pptr = &clp->ptr;
861 sp->pid = &clp->last_ptr_id;
862 }
863 }
864 WARN_ON(1);
865 return NULL;
866}
867
868/* dequeues packet at given priority and level; call only if
869 * you are sure that there is active class at prio/level
870 */
871static struct sk_buff *htb_dequeue_tree(struct htb_sched *q, const int prio,
872 const int level)
873{
874 struct sk_buff *skb = NULL;
875 struct htb_class *cl, *start;
876 struct htb_level *hlevel = &q->hlevel[level];
877 struct htb_prio *hprio = &hlevel->hprio[prio];
878
879 /* look initial class up in the row */
880 start = cl = htb_lookup_leaf(hprio, prio);
881
882 do {
883next:
884 if (unlikely(!cl))
885 return NULL;
886
887 /* class can be empty - it is unlikely but can be true if leaf
888 * qdisc drops packets in enqueue routine or if someone used
889 * graft operation on the leaf since last dequeue;
890 * simply deactivate and skip such class
891 */
892 if (unlikely(cl->leaf.q->q.qlen == 0)) {
893 struct htb_class *next;
894 htb_deactivate(q, cl);
895
896 /* row/level might become empty */
897 if ((q->row_mask[level] & (1 << prio)) == 0)
898 return NULL;
899
900 next = htb_lookup_leaf(hprio, prio);
901
902 if (cl == start) /* fix start if we just deleted it */
903 start = next;
904 cl = next;
905 goto next;
906 }
907
908 skb = cl->leaf.q->dequeue(cl->leaf.q);
909 if (likely(skb != NULL))
910 break;
911
912 qdisc_warn_nonwc("htb", cl->leaf.q);
913 htb_next_rb_node(level ? &cl->parent->inner.clprio[prio].ptr:
914 &q->hlevel[0].hprio[prio].ptr);
915 cl = htb_lookup_leaf(hprio, prio);
916
917 } while (cl != start);
918
919 if (likely(skb != NULL)) {
920 bstats_update(&cl->bstats, skb);
921 cl->leaf.deficit[level] -= qdisc_pkt_len(skb);
922 if (cl->leaf.deficit[level] < 0) {
923 cl->leaf.deficit[level] += cl->quantum;
924 htb_next_rb_node(level ? &cl->parent->inner.clprio[prio].ptr :
925 &q->hlevel[0].hprio[prio].ptr);
926 }
927 /* this used to be after charge_class but this constelation
928 * gives us slightly better performance
929 */
930 if (!cl->leaf.q->q.qlen)
931 htb_deactivate(q, cl);
932 htb_charge_class(q, cl, level, skb);
933 }
934 return skb;
935}
936
937static struct sk_buff *htb_dequeue(struct Qdisc *sch)
938{
939 struct sk_buff *skb;
940 struct htb_sched *q = qdisc_priv(sch);
941 int level;
942 s64 next_event;
943 unsigned long start_at;
944
945 /* try to dequeue direct packets as high prio (!) to minimize cpu work */
946 skb = __qdisc_dequeue_head(&q->direct_queue);
947 if (skb != NULL) {
948ok:
949 qdisc_bstats_update(sch, skb);
950 qdisc_qstats_backlog_dec(sch, skb);
951 sch->q.qlen--;
952 return skb;
953 }
954
955 if (!sch->q.qlen)
956 goto fin;
957 q->now = ktime_get_ns();
958 start_at = jiffies;
959
960 next_event = q->now + 5LLU * NSEC_PER_SEC;
961
962 for (level = 0; level < TC_HTB_MAXDEPTH; level++) {
963 /* common case optimization - skip event handler quickly */
964 int m;
965 s64 event = q->near_ev_cache[level];
966
967 if (q->now >= event) {
968 event = htb_do_events(q, level, start_at);
969 if (!event)
970 event = q->now + NSEC_PER_SEC;
971 q->near_ev_cache[level] = event;
972 }
973
974 if (next_event > event)
975 next_event = event;
976
977 m = ~q->row_mask[level];
978 while (m != (int)(-1)) {
979 int prio = ffz(m);
980
981 m |= 1 << prio;
982 skb = htb_dequeue_tree(q, prio, level);
983 if (likely(skb != NULL))
984 goto ok;
985 }
986 }
987 if (likely(next_event > q->now))
988 qdisc_watchdog_schedule_ns(&q->watchdog, next_event);
989 else
990 schedule_work(&q->work);
991fin:
992 return skb;
993}
994
995/* reset all classes */
996/* always caled under BH & queue lock */
997static void htb_reset(struct Qdisc *sch)
998{
999 struct htb_sched *q = qdisc_priv(sch);
1000 struct htb_class *cl;
1001 unsigned int i;
1002
1003 for (i = 0; i < q->clhash.hashsize; i++) {
1004 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1005 if (cl->level)
1006 memset(&cl->inner, 0, sizeof(cl->inner));
1007 else {
1008 if (cl->leaf.q && !q->offload)
1009 qdisc_reset(cl->leaf.q);
1010 }
1011 cl->prio_activity = 0;
1012 cl->cmode = HTB_CAN_SEND;
1013 }
1014 }
1015 qdisc_watchdog_cancel(&q->watchdog);
1016 __qdisc_reset_queue(&q->direct_queue);
1017 memset(q->hlevel, 0, sizeof(q->hlevel));
1018 memset(q->row_mask, 0, sizeof(q->row_mask));
1019}
1020
1021static const struct nla_policy htb_policy[TCA_HTB_MAX + 1] = {
1022 [TCA_HTB_PARMS] = { .len = sizeof(struct tc_htb_opt) },
1023 [TCA_HTB_INIT] = { .len = sizeof(struct tc_htb_glob) },
1024 [TCA_HTB_CTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
1025 [TCA_HTB_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
1026 [TCA_HTB_DIRECT_QLEN] = { .type = NLA_U32 },
1027 [TCA_HTB_RATE64] = { .type = NLA_U64 },
1028 [TCA_HTB_CEIL64] = { .type = NLA_U64 },
1029 [TCA_HTB_OFFLOAD] = { .type = NLA_FLAG },
1030};
1031
1032static void htb_work_func(struct work_struct *work)
1033{
1034 struct htb_sched *q = container_of(work, struct htb_sched, work);
1035 struct Qdisc *sch = q->watchdog.qdisc;
1036
1037 rcu_read_lock();
1038 __netif_schedule(qdisc_root(sch));
1039 rcu_read_unlock();
1040}
1041
1042static void htb_set_lockdep_class_child(struct Qdisc *q)
1043{
1044 static struct lock_class_key child_key;
1045
1046 lockdep_set_class(qdisc_lock(q), &child_key);
1047}
1048
1049static int htb_offload(struct net_device *dev, struct tc_htb_qopt_offload *opt)
1050{
1051 return dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_HTB, opt);
1052}
1053
1054static int htb_init(struct Qdisc *sch, struct nlattr *opt,
1055 struct netlink_ext_ack *extack)
1056{
1057 struct net_device *dev = qdisc_dev(sch);
1058 struct tc_htb_qopt_offload offload_opt;
1059 struct htb_sched *q = qdisc_priv(sch);
1060 struct nlattr *tb[TCA_HTB_MAX + 1];
1061 struct tc_htb_glob *gopt;
1062 unsigned int ntx;
1063 bool offload;
1064 int err;
1065
1066 qdisc_watchdog_init(&q->watchdog, sch);
1067 INIT_WORK(&q->work, htb_work_func);
1068
1069 if (!opt)
1070 return -EINVAL;
1071
1072 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
1073 if (err)
1074 return err;
1075
1076 err = nla_parse_nested_deprecated(tb, TCA_HTB_MAX, opt, htb_policy,
1077 NULL);
1078 if (err < 0)
1079 return err;
1080
1081 if (!tb[TCA_HTB_INIT])
1082 return -EINVAL;
1083
1084 gopt = nla_data(tb[TCA_HTB_INIT]);
1085 if (gopt->version != HTB_VER >> 16)
1086 return -EINVAL;
1087
1088 offload = nla_get_flag(tb[TCA_HTB_OFFLOAD]);
1089
1090 if (offload) {
1091 if (sch->parent != TC_H_ROOT) {
1092 NL_SET_ERR_MSG(extack, "HTB must be the root qdisc to use offload");
1093 return -EOPNOTSUPP;
1094 }
1095
1096 if (!tc_can_offload(dev) || !dev->netdev_ops->ndo_setup_tc) {
1097 NL_SET_ERR_MSG(extack, "hw-tc-offload ethtool feature flag must be on");
1098 return -EOPNOTSUPP;
1099 }
1100
1101 q->num_direct_qdiscs = dev->real_num_tx_queues;
1102 q->direct_qdiscs = kcalloc(q->num_direct_qdiscs,
1103 sizeof(*q->direct_qdiscs),
1104 GFP_KERNEL);
1105 if (!q->direct_qdiscs)
1106 return -ENOMEM;
1107 }
1108
1109 err = qdisc_class_hash_init(&q->clhash);
1110 if (err < 0)
1111 return err;
1112
1113 if (tb[TCA_HTB_DIRECT_QLEN])
1114 q->direct_qlen = nla_get_u32(tb[TCA_HTB_DIRECT_QLEN]);
1115 else
1116 q->direct_qlen = qdisc_dev(sch)->tx_queue_len;
1117
1118 if ((q->rate2quantum = gopt->rate2quantum) < 1)
1119 q->rate2quantum = 1;
1120 q->defcls = gopt->defcls;
1121
1122 if (!offload)
1123 return 0;
1124
1125 for (ntx = 0; ntx < q->num_direct_qdiscs; ntx++) {
1126 struct netdev_queue *dev_queue = netdev_get_tx_queue(dev, ntx);
1127 struct Qdisc *qdisc;
1128
1129 qdisc = qdisc_create_dflt(dev_queue, &pfifo_qdisc_ops,
1130 TC_H_MAKE(sch->handle, 0), extack);
1131 if (!qdisc) {
1132 return -ENOMEM;
1133 }
1134
1135 htb_set_lockdep_class_child(qdisc);
1136 q->direct_qdiscs[ntx] = qdisc;
1137 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1138 }
1139
1140 sch->flags |= TCQ_F_MQROOT;
1141
1142 offload_opt = (struct tc_htb_qopt_offload) {
1143 .command = TC_HTB_CREATE,
1144 .parent_classid = TC_H_MAJ(sch->handle) >> 16,
1145 .classid = TC_H_MIN(q->defcls),
1146 .extack = extack,
1147 };
1148 err = htb_offload(dev, &offload_opt);
1149 if (err)
1150 return err;
1151
1152 /* Defer this assignment, so that htb_destroy skips offload-related
1153 * parts (especially calling ndo_setup_tc) on errors.
1154 */
1155 q->offload = true;
1156
1157 return 0;
1158}
1159
1160static void htb_attach_offload(struct Qdisc *sch)
1161{
1162 struct net_device *dev = qdisc_dev(sch);
1163 struct htb_sched *q = qdisc_priv(sch);
1164 unsigned int ntx;
1165
1166 for (ntx = 0; ntx < q->num_direct_qdiscs; ntx++) {
1167 struct Qdisc *old, *qdisc = q->direct_qdiscs[ntx];
1168
1169 old = dev_graft_qdisc(qdisc->dev_queue, qdisc);
1170 qdisc_put(old);
1171 qdisc_hash_add(qdisc, false);
1172 }
1173 for (ntx = q->num_direct_qdiscs; ntx < dev->num_tx_queues; ntx++) {
1174 struct netdev_queue *dev_queue = netdev_get_tx_queue(dev, ntx);
1175 struct Qdisc *old = dev_graft_qdisc(dev_queue, NULL);
1176
1177 qdisc_put(old);
1178 }
1179
1180 kfree(q->direct_qdiscs);
1181 q->direct_qdiscs = NULL;
1182}
1183
1184static void htb_attach_software(struct Qdisc *sch)
1185{
1186 struct net_device *dev = qdisc_dev(sch);
1187 unsigned int ntx;
1188
1189 /* Resemble qdisc_graft behavior. */
1190 for (ntx = 0; ntx < dev->num_tx_queues; ntx++) {
1191 struct netdev_queue *dev_queue = netdev_get_tx_queue(dev, ntx);
1192 struct Qdisc *old = dev_graft_qdisc(dev_queue, sch);
1193
1194 qdisc_refcount_inc(sch);
1195
1196 qdisc_put(old);
1197 }
1198}
1199
1200static void htb_attach(struct Qdisc *sch)
1201{
1202 struct htb_sched *q = qdisc_priv(sch);
1203
1204 if (q->offload)
1205 htb_attach_offload(sch);
1206 else
1207 htb_attach_software(sch);
1208}
1209
1210static int htb_dump(struct Qdisc *sch, struct sk_buff *skb)
1211{
1212 struct htb_sched *q = qdisc_priv(sch);
1213 struct nlattr *nest;
1214 struct tc_htb_glob gopt;
1215
1216 if (q->offload)
1217 sch->flags |= TCQ_F_OFFLOADED;
1218 else
1219 sch->flags &= ~TCQ_F_OFFLOADED;
1220
1221 sch->qstats.overlimits = q->overlimits;
1222 /* Its safe to not acquire qdisc lock. As we hold RTNL,
1223 * no change can happen on the qdisc parameters.
1224 */
1225
1226 gopt.direct_pkts = q->direct_pkts;
1227 gopt.version = HTB_VER;
1228 gopt.rate2quantum = q->rate2quantum;
1229 gopt.defcls = q->defcls;
1230 gopt.debug = 0;
1231
1232 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1233 if (nest == NULL)
1234 goto nla_put_failure;
1235 if (nla_put(skb, TCA_HTB_INIT, sizeof(gopt), &gopt) ||
1236 nla_put_u32(skb, TCA_HTB_DIRECT_QLEN, q->direct_qlen))
1237 goto nla_put_failure;
1238 if (q->offload && nla_put_flag(skb, TCA_HTB_OFFLOAD))
1239 goto nla_put_failure;
1240
1241 return nla_nest_end(skb, nest);
1242
1243nla_put_failure:
1244 nla_nest_cancel(skb, nest);
1245 return -1;
1246}
1247
1248static int htb_dump_class(struct Qdisc *sch, unsigned long arg,
1249 struct sk_buff *skb, struct tcmsg *tcm)
1250{
1251 struct htb_class *cl = (struct htb_class *)arg;
1252 struct htb_sched *q = qdisc_priv(sch);
1253 struct nlattr *nest;
1254 struct tc_htb_opt opt;
1255
1256 /* Its safe to not acquire qdisc lock. As we hold RTNL,
1257 * no change can happen on the class parameters.
1258 */
1259 tcm->tcm_parent = cl->parent ? cl->parent->common.classid : TC_H_ROOT;
1260 tcm->tcm_handle = cl->common.classid;
1261 if (!cl->level && cl->leaf.q)
1262 tcm->tcm_info = cl->leaf.q->handle;
1263
1264 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
1265 if (nest == NULL)
1266 goto nla_put_failure;
1267
1268 memset(&opt, 0, sizeof(opt));
1269
1270 psched_ratecfg_getrate(&opt.rate, &cl->rate);
1271 opt.buffer = PSCHED_NS2TICKS(cl->buffer);
1272 psched_ratecfg_getrate(&opt.ceil, &cl->ceil);
1273 opt.cbuffer = PSCHED_NS2TICKS(cl->cbuffer);
1274 opt.quantum = cl->quantum;
1275 opt.prio = cl->prio;
1276 opt.level = cl->level;
1277 if (nla_put(skb, TCA_HTB_PARMS, sizeof(opt), &opt))
1278 goto nla_put_failure;
1279 if (q->offload && nla_put_flag(skb, TCA_HTB_OFFLOAD))
1280 goto nla_put_failure;
1281 if ((cl->rate.rate_bytes_ps >= (1ULL << 32)) &&
1282 nla_put_u64_64bit(skb, TCA_HTB_RATE64, cl->rate.rate_bytes_ps,
1283 TCA_HTB_PAD))
1284 goto nla_put_failure;
1285 if ((cl->ceil.rate_bytes_ps >= (1ULL << 32)) &&
1286 nla_put_u64_64bit(skb, TCA_HTB_CEIL64, cl->ceil.rate_bytes_ps,
1287 TCA_HTB_PAD))
1288 goto nla_put_failure;
1289
1290 return nla_nest_end(skb, nest);
1291
1292nla_put_failure:
1293 nla_nest_cancel(skb, nest);
1294 return -1;
1295}
1296
1297static void htb_offload_aggregate_stats(struct htb_sched *q,
1298 struct htb_class *cl)
1299{
1300 u64 bytes = 0, packets = 0;
1301 struct htb_class *c;
1302 unsigned int i;
1303
1304 gnet_stats_basic_sync_init(&cl->bstats);
1305
1306 for (i = 0; i < q->clhash.hashsize; i++) {
1307 hlist_for_each_entry(c, &q->clhash.hash[i], common.hnode) {
1308 struct htb_class *p = c;
1309
1310 while (p && p->level < cl->level)
1311 p = p->parent;
1312
1313 if (p != cl)
1314 continue;
1315
1316 bytes += u64_stats_read(&c->bstats_bias.bytes);
1317 packets += u64_stats_read(&c->bstats_bias.packets);
1318 if (c->level == 0) {
1319 bytes += u64_stats_read(&c->leaf.q->bstats.bytes);
1320 packets += u64_stats_read(&c->leaf.q->bstats.packets);
1321 }
1322 }
1323 }
1324 _bstats_update(&cl->bstats, bytes, packets);
1325}
1326
1327static int
1328htb_dump_class_stats(struct Qdisc *sch, unsigned long arg, struct gnet_dump *d)
1329{
1330 struct htb_class *cl = (struct htb_class *)arg;
1331 struct htb_sched *q = qdisc_priv(sch);
1332 struct gnet_stats_queue qs = {
1333 .drops = cl->drops,
1334 .overlimits = cl->overlimits,
1335 };
1336 __u32 qlen = 0;
1337
1338 if (!cl->level && cl->leaf.q)
1339 qdisc_qstats_qlen_backlog(cl->leaf.q, &qlen, &qs.backlog);
1340
1341 cl->xstats.tokens = clamp_t(s64, PSCHED_NS2TICKS(cl->tokens),
1342 INT_MIN, INT_MAX);
1343 cl->xstats.ctokens = clamp_t(s64, PSCHED_NS2TICKS(cl->ctokens),
1344 INT_MIN, INT_MAX);
1345
1346 if (q->offload) {
1347 if (!cl->level) {
1348 if (cl->leaf.q)
1349 cl->bstats = cl->leaf.q->bstats;
1350 else
1351 gnet_stats_basic_sync_init(&cl->bstats);
1352 _bstats_update(&cl->bstats,
1353 u64_stats_read(&cl->bstats_bias.bytes),
1354 u64_stats_read(&cl->bstats_bias.packets));
1355 } else {
1356 htb_offload_aggregate_stats(q, cl);
1357 }
1358 }
1359
1360 if (gnet_stats_copy_basic(d, NULL, &cl->bstats, true) < 0 ||
1361 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1362 gnet_stats_copy_queue(d, NULL, &qs, qlen) < 0)
1363 return -1;
1364
1365 return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1366}
1367
1368static struct netdev_queue *
1369htb_select_queue(struct Qdisc *sch, struct tcmsg *tcm)
1370{
1371 struct net_device *dev = qdisc_dev(sch);
1372 struct tc_htb_qopt_offload offload_opt;
1373 struct htb_sched *q = qdisc_priv(sch);
1374 int err;
1375
1376 if (!q->offload)
1377 return sch->dev_queue;
1378
1379 offload_opt = (struct tc_htb_qopt_offload) {
1380 .command = TC_HTB_LEAF_QUERY_QUEUE,
1381 .classid = TC_H_MIN(tcm->tcm_parent),
1382 };
1383 err = htb_offload(dev, &offload_opt);
1384 if (err || offload_opt.qid >= dev->num_tx_queues)
1385 return NULL;
1386 return netdev_get_tx_queue(dev, offload_opt.qid);
1387}
1388
1389static struct Qdisc *
1390htb_graft_helper(struct netdev_queue *dev_queue, struct Qdisc *new_q)
1391{
1392 struct net_device *dev = dev_queue->dev;
1393 struct Qdisc *old_q;
1394
1395 if (dev->flags & IFF_UP)
1396 dev_deactivate(dev);
1397 old_q = dev_graft_qdisc(dev_queue, new_q);
1398 if (new_q)
1399 new_q->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1400 if (dev->flags & IFF_UP)
1401 dev_activate(dev);
1402
1403 return old_q;
1404}
1405
1406static struct netdev_queue *htb_offload_get_queue(struct htb_class *cl)
1407{
1408 struct netdev_queue *queue;
1409
1410 queue = cl->leaf.offload_queue;
1411 if (!(cl->leaf.q->flags & TCQ_F_BUILTIN))
1412 WARN_ON(cl->leaf.q->dev_queue != queue);
1413
1414 return queue;
1415}
1416
1417static void htb_offload_move_qdisc(struct Qdisc *sch, struct htb_class *cl_old,
1418 struct htb_class *cl_new, bool destroying)
1419{
1420 struct netdev_queue *queue_old, *queue_new;
1421 struct net_device *dev = qdisc_dev(sch);
1422
1423 queue_old = htb_offload_get_queue(cl_old);
1424 queue_new = htb_offload_get_queue(cl_new);
1425
1426 if (!destroying) {
1427 struct Qdisc *qdisc;
1428
1429 if (dev->flags & IFF_UP)
1430 dev_deactivate(dev);
1431 qdisc = dev_graft_qdisc(queue_old, NULL);
1432 WARN_ON(qdisc != cl_old->leaf.q);
1433 }
1434
1435 if (!(cl_old->leaf.q->flags & TCQ_F_BUILTIN))
1436 cl_old->leaf.q->dev_queue = queue_new;
1437 cl_old->leaf.offload_queue = queue_new;
1438
1439 if (!destroying) {
1440 struct Qdisc *qdisc;
1441
1442 qdisc = dev_graft_qdisc(queue_new, cl_old->leaf.q);
1443 if (dev->flags & IFF_UP)
1444 dev_activate(dev);
1445 WARN_ON(!(qdisc->flags & TCQ_F_BUILTIN));
1446 }
1447}
1448
1449static int htb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1450 struct Qdisc **old, struct netlink_ext_ack *extack)
1451{
1452 struct netdev_queue *dev_queue = sch->dev_queue;
1453 struct htb_class *cl = (struct htb_class *)arg;
1454 struct htb_sched *q = qdisc_priv(sch);
1455 struct Qdisc *old_q;
1456
1457 if (cl->level)
1458 return -EINVAL;
1459
1460 if (q->offload)
1461 dev_queue = htb_offload_get_queue(cl);
1462
1463 if (!new) {
1464 new = qdisc_create_dflt(dev_queue, &pfifo_qdisc_ops,
1465 cl->common.classid, extack);
1466 if (!new)
1467 return -ENOBUFS;
1468 }
1469
1470 if (q->offload) {
1471 htb_set_lockdep_class_child(new);
1472 /* One ref for cl->leaf.q, the other for dev_queue->qdisc. */
1473 qdisc_refcount_inc(new);
1474 old_q = htb_graft_helper(dev_queue, new);
1475 }
1476
1477 *old = qdisc_replace(sch, new, &cl->leaf.q);
1478
1479 if (q->offload) {
1480 WARN_ON(old_q != *old);
1481 qdisc_put(old_q);
1482 }
1483
1484 return 0;
1485}
1486
1487static struct Qdisc *htb_leaf(struct Qdisc *sch, unsigned long arg)
1488{
1489 struct htb_class *cl = (struct htb_class *)arg;
1490 return !cl->level ? cl->leaf.q : NULL;
1491}
1492
1493static void htb_qlen_notify(struct Qdisc *sch, unsigned long arg)
1494{
1495 struct htb_class *cl = (struct htb_class *)arg;
1496
1497 htb_deactivate(qdisc_priv(sch), cl);
1498}
1499
1500static inline int htb_parent_last_child(struct htb_class *cl)
1501{
1502 if (!cl->parent)
1503 /* the root class */
1504 return 0;
1505 if (cl->parent->children > 1)
1506 /* not the last child */
1507 return 0;
1508 return 1;
1509}
1510
1511static void htb_parent_to_leaf(struct Qdisc *sch, struct htb_class *cl,
1512 struct Qdisc *new_q)
1513{
1514 struct htb_sched *q = qdisc_priv(sch);
1515 struct htb_class *parent = cl->parent;
1516
1517 WARN_ON(cl->level || !cl->leaf.q || cl->prio_activity);
1518
1519 if (parent->cmode != HTB_CAN_SEND)
1520 htb_safe_rb_erase(&parent->pq_node,
1521 &q->hlevel[parent->level].wait_pq);
1522
1523 parent->level = 0;
1524 memset(&parent->inner, 0, sizeof(parent->inner));
1525 parent->leaf.q = new_q ? new_q : &noop_qdisc;
1526 parent->tokens = parent->buffer;
1527 parent->ctokens = parent->cbuffer;
1528 parent->t_c = ktime_get_ns();
1529 parent->cmode = HTB_CAN_SEND;
1530 if (q->offload)
1531 parent->leaf.offload_queue = cl->leaf.offload_queue;
1532}
1533
1534static void htb_parent_to_leaf_offload(struct Qdisc *sch,
1535 struct netdev_queue *dev_queue,
1536 struct Qdisc *new_q)
1537{
1538 struct Qdisc *old_q;
1539
1540 /* One ref for cl->leaf.q, the other for dev_queue->qdisc. */
1541 if (new_q)
1542 qdisc_refcount_inc(new_q);
1543 old_q = htb_graft_helper(dev_queue, new_q);
1544 WARN_ON(!(old_q->flags & TCQ_F_BUILTIN));
1545}
1546
1547static int htb_destroy_class_offload(struct Qdisc *sch, struct htb_class *cl,
1548 bool last_child, bool destroying,
1549 struct netlink_ext_ack *extack)
1550{
1551 struct tc_htb_qopt_offload offload_opt;
1552 struct netdev_queue *dev_queue;
1553 struct Qdisc *q = cl->leaf.q;
1554 struct Qdisc *old;
1555 int err;
1556
1557 if (cl->level)
1558 return -EINVAL;
1559
1560 WARN_ON(!q);
1561 dev_queue = htb_offload_get_queue(cl);
1562 /* When destroying, caller qdisc_graft grafts the new qdisc and invokes
1563 * qdisc_put for the qdisc being destroyed. htb_destroy_class_offload
1564 * does not need to graft or qdisc_put the qdisc being destroyed.
1565 */
1566 if (!destroying) {
1567 old = htb_graft_helper(dev_queue, NULL);
1568 /* Last qdisc grafted should be the same as cl->leaf.q when
1569 * calling htb_delete.
1570 */
1571 WARN_ON(old != q);
1572 }
1573
1574 if (cl->parent) {
1575 _bstats_update(&cl->parent->bstats_bias,
1576 u64_stats_read(&q->bstats.bytes),
1577 u64_stats_read(&q->bstats.packets));
1578 }
1579
1580 offload_opt = (struct tc_htb_qopt_offload) {
1581 .command = !last_child ? TC_HTB_LEAF_DEL :
1582 destroying ? TC_HTB_LEAF_DEL_LAST_FORCE :
1583 TC_HTB_LEAF_DEL_LAST,
1584 .classid = cl->common.classid,
1585 .extack = extack,
1586 };
1587 err = htb_offload(qdisc_dev(sch), &offload_opt);
1588
1589 if (!destroying) {
1590 if (!err)
1591 qdisc_put(old);
1592 else
1593 htb_graft_helper(dev_queue, old);
1594 }
1595
1596 if (last_child)
1597 return err;
1598
1599 if (!err && offload_opt.classid != TC_H_MIN(cl->common.classid)) {
1600 u32 classid = TC_H_MAJ(sch->handle) |
1601 TC_H_MIN(offload_opt.classid);
1602 struct htb_class *moved_cl = htb_find(classid, sch);
1603
1604 htb_offload_move_qdisc(sch, moved_cl, cl, destroying);
1605 }
1606
1607 return err;
1608}
1609
1610static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
1611{
1612 if (!cl->level) {
1613 WARN_ON(!cl->leaf.q);
1614 qdisc_put(cl->leaf.q);
1615 }
1616 gen_kill_estimator(&cl->rate_est);
1617 tcf_block_put(cl->block);
1618 kfree(cl);
1619}
1620
1621static void htb_destroy(struct Qdisc *sch)
1622{
1623 struct net_device *dev = qdisc_dev(sch);
1624 struct tc_htb_qopt_offload offload_opt;
1625 struct htb_sched *q = qdisc_priv(sch);
1626 struct hlist_node *next;
1627 bool nonempty, changed;
1628 struct htb_class *cl;
1629 unsigned int i;
1630
1631 cancel_work_sync(&q->work);
1632 qdisc_watchdog_cancel(&q->watchdog);
1633 /* This line used to be after htb_destroy_class call below
1634 * and surprisingly it worked in 2.4. But it must precede it
1635 * because filter need its target class alive to be able to call
1636 * unbind_filter on it (without Oops).
1637 */
1638 tcf_block_put(q->block);
1639
1640 for (i = 0; i < q->clhash.hashsize; i++) {
1641 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1642 tcf_block_put(cl->block);
1643 cl->block = NULL;
1644 }
1645 }
1646
1647 do {
1648 nonempty = false;
1649 changed = false;
1650 for (i = 0; i < q->clhash.hashsize; i++) {
1651 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1652 common.hnode) {
1653 bool last_child;
1654
1655 if (!q->offload) {
1656 htb_destroy_class(sch, cl);
1657 continue;
1658 }
1659
1660 nonempty = true;
1661
1662 if (cl->level)
1663 continue;
1664
1665 changed = true;
1666
1667 last_child = htb_parent_last_child(cl);
1668 htb_destroy_class_offload(sch, cl, last_child,
1669 true, NULL);
1670 qdisc_class_hash_remove(&q->clhash,
1671 &cl->common);
1672 if (cl->parent)
1673 cl->parent->children--;
1674 if (last_child)
1675 htb_parent_to_leaf(sch, cl, NULL);
1676 htb_destroy_class(sch, cl);
1677 }
1678 }
1679 } while (changed);
1680 WARN_ON(nonempty);
1681
1682 qdisc_class_hash_destroy(&q->clhash);
1683 __qdisc_reset_queue(&q->direct_queue);
1684
1685 if (q->offload) {
1686 offload_opt = (struct tc_htb_qopt_offload) {
1687 .command = TC_HTB_DESTROY,
1688 };
1689 htb_offload(dev, &offload_opt);
1690 }
1691
1692 if (!q->direct_qdiscs)
1693 return;
1694 for (i = 0; i < q->num_direct_qdiscs && q->direct_qdiscs[i]; i++)
1695 qdisc_put(q->direct_qdiscs[i]);
1696 kfree(q->direct_qdiscs);
1697}
1698
1699static int htb_delete(struct Qdisc *sch, unsigned long arg,
1700 struct netlink_ext_ack *extack)
1701{
1702 struct htb_sched *q = qdisc_priv(sch);
1703 struct htb_class *cl = (struct htb_class *)arg;
1704 struct Qdisc *new_q = NULL;
1705 int last_child = 0;
1706 int err;
1707
1708 /* TODO: why don't allow to delete subtree ? references ? does
1709 * tc subsys guarantee us that in htb_destroy it holds no class
1710 * refs so that we can remove children safely there ?
1711 */
1712 if (cl->children || qdisc_class_in_use(&cl->common)) {
1713 NL_SET_ERR_MSG(extack, "HTB class in use");
1714 return -EBUSY;
1715 }
1716
1717 if (!cl->level && htb_parent_last_child(cl))
1718 last_child = 1;
1719
1720 if (q->offload) {
1721 err = htb_destroy_class_offload(sch, cl, last_child, false,
1722 extack);
1723 if (err)
1724 return err;
1725 }
1726
1727 if (last_child) {
1728 struct netdev_queue *dev_queue = sch->dev_queue;
1729
1730 if (q->offload)
1731 dev_queue = htb_offload_get_queue(cl);
1732
1733 new_q = qdisc_create_dflt(dev_queue, &pfifo_qdisc_ops,
1734 cl->parent->common.classid,
1735 NULL);
1736 if (q->offload) {
1737 if (new_q)
1738 htb_set_lockdep_class_child(new_q);
1739 htb_parent_to_leaf_offload(sch, dev_queue, new_q);
1740 }
1741 }
1742
1743 sch_tree_lock(sch);
1744
1745 if (!cl->level)
1746 qdisc_purge_queue(cl->leaf.q);
1747
1748 /* delete from hash and active; remainder in destroy_class */
1749 qdisc_class_hash_remove(&q->clhash, &cl->common);
1750 if (cl->parent)
1751 cl->parent->children--;
1752
1753 if (cl->prio_activity)
1754 htb_deactivate(q, cl);
1755
1756 if (cl->cmode != HTB_CAN_SEND)
1757 htb_safe_rb_erase(&cl->pq_node,
1758 &q->hlevel[cl->level].wait_pq);
1759
1760 if (last_child)
1761 htb_parent_to_leaf(sch, cl, new_q);
1762
1763 sch_tree_unlock(sch);
1764
1765 htb_destroy_class(sch, cl);
1766 return 0;
1767}
1768
1769static int htb_change_class(struct Qdisc *sch, u32 classid,
1770 u32 parentid, struct nlattr **tca,
1771 unsigned long *arg, struct netlink_ext_ack *extack)
1772{
1773 int err = -EINVAL;
1774 struct htb_sched *q = qdisc_priv(sch);
1775 struct htb_class *cl = (struct htb_class *)*arg, *parent;
1776 struct tc_htb_qopt_offload offload_opt;
1777 struct nlattr *opt = tca[TCA_OPTIONS];
1778 struct nlattr *tb[TCA_HTB_MAX + 1];
1779 struct Qdisc *parent_qdisc = NULL;
1780 struct netdev_queue *dev_queue;
1781 struct tc_htb_opt *hopt;
1782 u64 rate64, ceil64;
1783 int warn = 0;
1784
1785 /* extract all subattrs from opt attr */
1786 if (!opt)
1787 goto failure;
1788
1789 err = nla_parse_nested_deprecated(tb, TCA_HTB_MAX, opt, htb_policy,
1790 extack);
1791 if (err < 0)
1792 goto failure;
1793
1794 err = -EINVAL;
1795 if (tb[TCA_HTB_PARMS] == NULL)
1796 goto failure;
1797
1798 parent = parentid == TC_H_ROOT ? NULL : htb_find(parentid, sch);
1799
1800 hopt = nla_data(tb[TCA_HTB_PARMS]);
1801 if (!hopt->rate.rate || !hopt->ceil.rate)
1802 goto failure;
1803
1804 if (q->offload) {
1805 /* Options not supported by the offload. */
1806 if (hopt->rate.overhead || hopt->ceil.overhead) {
1807 NL_SET_ERR_MSG(extack, "HTB offload doesn't support the overhead parameter");
1808 goto failure;
1809 }
1810 if (hopt->rate.mpu || hopt->ceil.mpu) {
1811 NL_SET_ERR_MSG(extack, "HTB offload doesn't support the mpu parameter");
1812 goto failure;
1813 }
1814 }
1815
1816 /* Keeping backward compatible with rate_table based iproute2 tc */
1817 if (hopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
1818 qdisc_put_rtab(qdisc_get_rtab(&hopt->rate, tb[TCA_HTB_RTAB],
1819 NULL));
1820
1821 if (hopt->ceil.linklayer == TC_LINKLAYER_UNAWARE)
1822 qdisc_put_rtab(qdisc_get_rtab(&hopt->ceil, tb[TCA_HTB_CTAB],
1823 NULL));
1824
1825 rate64 = tb[TCA_HTB_RATE64] ? nla_get_u64(tb[TCA_HTB_RATE64]) : 0;
1826 ceil64 = tb[TCA_HTB_CEIL64] ? nla_get_u64(tb[TCA_HTB_CEIL64]) : 0;
1827
1828 if (!cl) { /* new class */
1829 struct net_device *dev = qdisc_dev(sch);
1830 struct Qdisc *new_q, *old_q;
1831 int prio;
1832 struct {
1833 struct nlattr nla;
1834 struct gnet_estimator opt;
1835 } est = {
1836 .nla = {
1837 .nla_len = nla_attr_size(sizeof(est.opt)),
1838 .nla_type = TCA_RATE,
1839 },
1840 .opt = {
1841 /* 4s interval, 16s averaging constant */
1842 .interval = 2,
1843 .ewma_log = 2,
1844 },
1845 };
1846
1847 /* check for valid classid */
1848 if (!classid || TC_H_MAJ(classid ^ sch->handle) ||
1849 htb_find(classid, sch))
1850 goto failure;
1851
1852 /* check maximal depth */
1853 if (parent && parent->parent && parent->parent->level < 2) {
1854 NL_SET_ERR_MSG_MOD(extack, "tree is too deep");
1855 goto failure;
1856 }
1857 err = -ENOBUFS;
1858 cl = kzalloc(sizeof(*cl), GFP_KERNEL);
1859 if (!cl)
1860 goto failure;
1861
1862 gnet_stats_basic_sync_init(&cl->bstats);
1863 gnet_stats_basic_sync_init(&cl->bstats_bias);
1864
1865 err = tcf_block_get(&cl->block, &cl->filter_list, sch, extack);
1866 if (err) {
1867 kfree(cl);
1868 goto failure;
1869 }
1870 if (htb_rate_est || tca[TCA_RATE]) {
1871 err = gen_new_estimator(&cl->bstats, NULL,
1872 &cl->rate_est,
1873 NULL,
1874 true,
1875 tca[TCA_RATE] ? : &est.nla);
1876 if (err)
1877 goto err_block_put;
1878 }
1879
1880 cl->children = 0;
1881 RB_CLEAR_NODE(&cl->pq_node);
1882
1883 for (prio = 0; prio < TC_HTB_NUMPRIO; prio++)
1884 RB_CLEAR_NODE(&cl->node[prio]);
1885
1886 cl->common.classid = classid;
1887
1888 /* Make sure nothing interrupts us in between of two
1889 * ndo_setup_tc calls.
1890 */
1891 ASSERT_RTNL();
1892
1893 /* create leaf qdisc early because it uses kmalloc(GFP_KERNEL)
1894 * so that can't be used inside of sch_tree_lock
1895 * -- thanks to Karlis Peisenieks
1896 */
1897 if (!q->offload) {
1898 dev_queue = sch->dev_queue;
1899 } else if (!(parent && !parent->level)) {
1900 /* Assign a dev_queue to this classid. */
1901 offload_opt = (struct tc_htb_qopt_offload) {
1902 .command = TC_HTB_LEAF_ALLOC_QUEUE,
1903 .classid = cl->common.classid,
1904 .parent_classid = parent ?
1905 TC_H_MIN(parent->common.classid) :
1906 TC_HTB_CLASSID_ROOT,
1907 .rate = max_t(u64, hopt->rate.rate, rate64),
1908 .ceil = max_t(u64, hopt->ceil.rate, ceil64),
1909 .prio = hopt->prio,
1910 .quantum = hopt->quantum,
1911 .extack = extack,
1912 };
1913 err = htb_offload(dev, &offload_opt);
1914 if (err) {
1915 NL_SET_ERR_MSG_WEAK(extack,
1916 "Failed to offload TC_HTB_LEAF_ALLOC_QUEUE");
1917 goto err_kill_estimator;
1918 }
1919 dev_queue = netdev_get_tx_queue(dev, offload_opt.qid);
1920 } else { /* First child. */
1921 dev_queue = htb_offload_get_queue(parent);
1922 old_q = htb_graft_helper(dev_queue, NULL);
1923 WARN_ON(old_q != parent->leaf.q);
1924 offload_opt = (struct tc_htb_qopt_offload) {
1925 .command = TC_HTB_LEAF_TO_INNER,
1926 .classid = cl->common.classid,
1927 .parent_classid =
1928 TC_H_MIN(parent->common.classid),
1929 .rate = max_t(u64, hopt->rate.rate, rate64),
1930 .ceil = max_t(u64, hopt->ceil.rate, ceil64),
1931 .prio = hopt->prio,
1932 .quantum = hopt->quantum,
1933 .extack = extack,
1934 };
1935 err = htb_offload(dev, &offload_opt);
1936 if (err) {
1937 NL_SET_ERR_MSG_WEAK(extack,
1938 "Failed to offload TC_HTB_LEAF_TO_INNER");
1939 htb_graft_helper(dev_queue, old_q);
1940 goto err_kill_estimator;
1941 }
1942 _bstats_update(&parent->bstats_bias,
1943 u64_stats_read(&old_q->bstats.bytes),
1944 u64_stats_read(&old_q->bstats.packets));
1945 qdisc_put(old_q);
1946 }
1947 new_q = qdisc_create_dflt(dev_queue, &pfifo_qdisc_ops,
1948 classid, NULL);
1949 if (q->offload) {
1950 if (new_q) {
1951 htb_set_lockdep_class_child(new_q);
1952 /* One ref for cl->leaf.q, the other for
1953 * dev_queue->qdisc.
1954 */
1955 qdisc_refcount_inc(new_q);
1956 }
1957 old_q = htb_graft_helper(dev_queue, new_q);
1958 /* No qdisc_put needed. */
1959 WARN_ON(!(old_q->flags & TCQ_F_BUILTIN));
1960 }
1961 sch_tree_lock(sch);
1962 if (parent && !parent->level) {
1963 /* turn parent into inner node */
1964 qdisc_purge_queue(parent->leaf.q);
1965 parent_qdisc = parent->leaf.q;
1966 if (parent->prio_activity)
1967 htb_deactivate(q, parent);
1968
1969 /* remove from evt list because of level change */
1970 if (parent->cmode != HTB_CAN_SEND) {
1971 htb_safe_rb_erase(&parent->pq_node, &q->hlevel[0].wait_pq);
1972 parent->cmode = HTB_CAN_SEND;
1973 }
1974 parent->level = (parent->parent ? parent->parent->level
1975 : TC_HTB_MAXDEPTH) - 1;
1976 memset(&parent->inner, 0, sizeof(parent->inner));
1977 }
1978
1979 /* leaf (we) needs elementary qdisc */
1980 cl->leaf.q = new_q ? new_q : &noop_qdisc;
1981 if (q->offload)
1982 cl->leaf.offload_queue = dev_queue;
1983
1984 cl->parent = parent;
1985
1986 /* set class to be in HTB_CAN_SEND state */
1987 cl->tokens = PSCHED_TICKS2NS(hopt->buffer);
1988 cl->ctokens = PSCHED_TICKS2NS(hopt->cbuffer);
1989 cl->mbuffer = 60ULL * NSEC_PER_SEC; /* 1min */
1990 cl->t_c = ktime_get_ns();
1991 cl->cmode = HTB_CAN_SEND;
1992
1993 /* attach to the hash list and parent's family */
1994 qdisc_class_hash_insert(&q->clhash, &cl->common);
1995 if (parent)
1996 parent->children++;
1997 if (cl->leaf.q != &noop_qdisc)
1998 qdisc_hash_add(cl->leaf.q, true);
1999 } else {
2000 if (tca[TCA_RATE]) {
2001 err = gen_replace_estimator(&cl->bstats, NULL,
2002 &cl->rate_est,
2003 NULL,
2004 true,
2005 tca[TCA_RATE]);
2006 if (err)
2007 return err;
2008 }
2009
2010 if (q->offload) {
2011 struct net_device *dev = qdisc_dev(sch);
2012
2013 offload_opt = (struct tc_htb_qopt_offload) {
2014 .command = TC_HTB_NODE_MODIFY,
2015 .classid = cl->common.classid,
2016 .rate = max_t(u64, hopt->rate.rate, rate64),
2017 .ceil = max_t(u64, hopt->ceil.rate, ceil64),
2018 .prio = hopt->prio,
2019 .quantum = hopt->quantum,
2020 .extack = extack,
2021 };
2022 err = htb_offload(dev, &offload_opt);
2023 if (err)
2024 /* Estimator was replaced, and rollback may fail
2025 * as well, so we don't try to recover it, and
2026 * the estimator won't work property with the
2027 * offload anyway, because bstats are updated
2028 * only when the stats are queried.
2029 */
2030 return err;
2031 }
2032
2033 sch_tree_lock(sch);
2034 }
2035
2036 psched_ratecfg_precompute(&cl->rate, &hopt->rate, rate64);
2037 psched_ratecfg_precompute(&cl->ceil, &hopt->ceil, ceil64);
2038
2039 /* it used to be a nasty bug here, we have to check that node
2040 * is really leaf before changing cl->leaf !
2041 */
2042 if (!cl->level) {
2043 u64 quantum = cl->rate.rate_bytes_ps;
2044
2045 do_div(quantum, q->rate2quantum);
2046 cl->quantum = min_t(u64, quantum, INT_MAX);
2047
2048 if (!hopt->quantum && cl->quantum < 1000) {
2049 warn = -1;
2050 cl->quantum = 1000;
2051 }
2052 if (!hopt->quantum && cl->quantum > 200000) {
2053 warn = 1;
2054 cl->quantum = 200000;
2055 }
2056 if (hopt->quantum)
2057 cl->quantum = hopt->quantum;
2058 if ((cl->prio = hopt->prio) >= TC_HTB_NUMPRIO)
2059 cl->prio = TC_HTB_NUMPRIO - 1;
2060 }
2061
2062 cl->buffer = PSCHED_TICKS2NS(hopt->buffer);
2063 cl->cbuffer = PSCHED_TICKS2NS(hopt->cbuffer);
2064
2065 sch_tree_unlock(sch);
2066 qdisc_put(parent_qdisc);
2067
2068 if (warn)
2069 NL_SET_ERR_MSG_FMT_MOD(extack,
2070 "quantum of class %X is %s. Consider r2q change.",
2071 cl->common.classid, (warn == -1 ? "small" : "big"));
2072
2073 qdisc_class_hash_grow(sch, &q->clhash);
2074
2075 *arg = (unsigned long)cl;
2076 return 0;
2077
2078err_kill_estimator:
2079 gen_kill_estimator(&cl->rate_est);
2080err_block_put:
2081 tcf_block_put(cl->block);
2082 kfree(cl);
2083failure:
2084 return err;
2085}
2086
2087static struct tcf_block *htb_tcf_block(struct Qdisc *sch, unsigned long arg,
2088 struct netlink_ext_ack *extack)
2089{
2090 struct htb_sched *q = qdisc_priv(sch);
2091 struct htb_class *cl = (struct htb_class *)arg;
2092
2093 return cl ? cl->block : q->block;
2094}
2095
2096static unsigned long htb_bind_filter(struct Qdisc *sch, unsigned long parent,
2097 u32 classid)
2098{
2099 struct htb_class *cl = htb_find(classid, sch);
2100
2101 /*if (cl && !cl->level) return 0;
2102 * The line above used to be there to prevent attaching filters to
2103 * leaves. But at least tc_index filter uses this just to get class
2104 * for other reasons so that we have to allow for it.
2105 * ----
2106 * 19.6.2002 As Werner explained it is ok - bind filter is just
2107 * another way to "lock" the class - unlike "get" this lock can
2108 * be broken by class during destroy IIUC.
2109 */
2110 if (cl)
2111 qdisc_class_get(&cl->common);
2112 return (unsigned long)cl;
2113}
2114
2115static void htb_unbind_filter(struct Qdisc *sch, unsigned long arg)
2116{
2117 struct htb_class *cl = (struct htb_class *)arg;
2118
2119 qdisc_class_put(&cl->common);
2120}
2121
2122static void htb_walk(struct Qdisc *sch, struct qdisc_walker *arg)
2123{
2124 struct htb_sched *q = qdisc_priv(sch);
2125 struct htb_class *cl;
2126 unsigned int i;
2127
2128 if (arg->stop)
2129 return;
2130
2131 for (i = 0; i < q->clhash.hashsize; i++) {
2132 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
2133 if (!tc_qdisc_stats_dump(sch, (unsigned long)cl, arg))
2134 return;
2135 }
2136 }
2137}
2138
2139static const struct Qdisc_class_ops htb_class_ops = {
2140 .select_queue = htb_select_queue,
2141 .graft = htb_graft,
2142 .leaf = htb_leaf,
2143 .qlen_notify = htb_qlen_notify,
2144 .find = htb_search,
2145 .change = htb_change_class,
2146 .delete = htb_delete,
2147 .walk = htb_walk,
2148 .tcf_block = htb_tcf_block,
2149 .bind_tcf = htb_bind_filter,
2150 .unbind_tcf = htb_unbind_filter,
2151 .dump = htb_dump_class,
2152 .dump_stats = htb_dump_class_stats,
2153};
2154
2155static struct Qdisc_ops htb_qdisc_ops __read_mostly = {
2156 .cl_ops = &htb_class_ops,
2157 .id = "htb",
2158 .priv_size = sizeof(struct htb_sched),
2159 .enqueue = htb_enqueue,
2160 .dequeue = htb_dequeue,
2161 .peek = qdisc_peek_dequeued,
2162 .init = htb_init,
2163 .attach = htb_attach,
2164 .reset = htb_reset,
2165 .destroy = htb_destroy,
2166 .dump = htb_dump,
2167 .owner = THIS_MODULE,
2168};
2169MODULE_ALIAS_NET_SCH("htb");
2170
2171static int __init htb_module_init(void)
2172{
2173 return register_qdisc(&htb_qdisc_ops);
2174}
2175static void __exit htb_module_exit(void)
2176{
2177 unregister_qdisc(&htb_qdisc_ops);
2178}
2179
2180module_init(htb_module_init)
2181module_exit(htb_module_exit)
2182MODULE_LICENSE("GPL");
2183MODULE_DESCRIPTION("Hierarchical Token Bucket scheduler");