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 return qdisc_drop(skb, sch);
562 }
563#ifdef CONFIG_NET_CLS_ACT
564 } else if (!cl) {
565 if (ret & __NET_XMIT_BYPASS)
566 sch->qstats.drops++;
567 kfree_skb(skb);
568 return ret;
569#endif
570 } else if ((ret = qdisc_enqueue(skb, cl->un.leaf.q)) != NET_XMIT_SUCCESS) {
571 if (net_xmit_drop_count(ret)) {
572 sch->qstats.drops++;
573 cl->qstats.drops++;
574 }
575 return ret;
576 } else {
577 htb_activate(q, cl);
578 }
579
580 sch->q.qlen++;
581 return NET_XMIT_SUCCESS;
582}
583
584static inline void htb_accnt_tokens(struct htb_class *cl, int bytes, long diff)
585{
586 long toks = diff + cl->tokens;
587
588 if (toks > cl->buffer)
589 toks = cl->buffer;
590 toks -= (long) qdisc_l2t(cl->rate, bytes);
591 if (toks <= -cl->mbuffer)
592 toks = 1 - cl->mbuffer;
593
594 cl->tokens = toks;
595}
596
597static inline void htb_accnt_ctokens(struct htb_class *cl, int bytes, long diff)
598{
599 long toks = diff + cl->ctokens;
600
601 if (toks > cl->cbuffer)
602 toks = cl->cbuffer;
603 toks -= (long) qdisc_l2t(cl->ceil, bytes);
604 if (toks <= -cl->mbuffer)
605 toks = 1 - cl->mbuffer;
606
607 cl->ctokens = toks;
608}
609
610/**
611 * htb_charge_class - charges amount "bytes" to leaf and ancestors
612 *
613 * Routine assumes that packet "bytes" long was dequeued from leaf cl
614 * borrowing from "level". It accounts bytes to ceil leaky bucket for
615 * leaf and all ancestors and to rate bucket for ancestors at levels
616 * "level" and higher. It also handles possible change of mode resulting
617 * from the update. Note that mode can also increase here (MAY_BORROW to
618 * CAN_SEND) because we can use more precise clock that event queue here.
619 * In such case we remove class from event queue first.
620 */
621static void htb_charge_class(struct htb_sched *q, struct htb_class *cl,
622 int level, struct sk_buff *skb)
623{
624 int bytes = qdisc_pkt_len(skb);
625 enum htb_cmode old_mode;
626 long diff;
627
628 while (cl) {
629 diff = psched_tdiff_bounded(q->now, cl->t_c, cl->mbuffer);
630 if (cl->level >= level) {
631 if (cl->level == level)
632 cl->xstats.lends++;
633 htb_accnt_tokens(cl, bytes, diff);
634 } else {
635 cl->xstats.borrows++;
636 cl->tokens += diff; /* we moved t_c; update tokens */
637 }
638 htb_accnt_ctokens(cl, bytes, diff);
639 cl->t_c = q->now;
640
641 old_mode = cl->cmode;
642 diff = 0;
643 htb_change_class_mode(q, cl, &diff);
644 if (old_mode != cl->cmode) {
645 if (old_mode != HTB_CAN_SEND)
646 htb_safe_rb_erase(&cl->pq_node, q->wait_pq + cl->level);
647 if (cl->cmode != HTB_CAN_SEND)
648 htb_add_to_wait_tree(q, cl, diff);
649 }
650
651 /* update basic stats except for leaves which are already updated */
652 if (cl->level)
653 bstats_update(&cl->bstats, skb);
654
655 cl = cl->parent;
656 }
657}
658
659/**
660 * htb_do_events - make mode changes to classes at the level
661 *
662 * Scans event queue for pending events and applies them. Returns time of
663 * next pending event (0 for no event in pq, q->now for too many events).
664 * Note: Applied are events whose have cl->pq_key <= q->now.
665 */
666static psched_time_t htb_do_events(struct htb_sched *q, int level,
667 unsigned long start)
668{
669 /* don't run for longer than 2 jiffies; 2 is used instead of
670 * 1 to simplify things when jiffy is going to be incremented
671 * too soon
672 */
673 unsigned long stop_at = start + 2;
674 while (time_before(jiffies, stop_at)) {
675 struct htb_class *cl;
676 long diff;
677 struct rb_node *p = rb_first(&q->wait_pq[level]);
678
679 if (!p)
680 return 0;
681
682 cl = rb_entry(p, struct htb_class, pq_node);
683 if (cl->pq_key > q->now)
684 return cl->pq_key;
685
686 htb_safe_rb_erase(p, q->wait_pq + level);
687 diff = psched_tdiff_bounded(q->now, cl->t_c, cl->mbuffer);
688 htb_change_class_mode(q, cl, &diff);
689 if (cl->cmode != HTB_CAN_SEND)
690 htb_add_to_wait_tree(q, cl, diff);
691 }
692
693 /* too much load - let's continue after a break for scheduling */
694 if (!(q->warned & HTB_WARN_TOOMANYEVENTS)) {
695 pr_warning("htb: too many events!\n");
696 q->warned |= HTB_WARN_TOOMANYEVENTS;
697 }
698
699 return q->now;
700}
701
702/* Returns class->node+prio from id-tree where classe's id is >= id. NULL
703 * is no such one exists.
704 */
705static struct rb_node *htb_id_find_next_upper(int prio, struct rb_node *n,
706 u32 id)
707{
708 struct rb_node *r = NULL;
709 while (n) {
710 struct htb_class *cl =
711 rb_entry(n, struct htb_class, node[prio]);
712
713 if (id > cl->common.classid) {
714 n = n->rb_right;
715 } else if (id < cl->common.classid) {
716 r = n;
717 n = n->rb_left;
718 } else {
719 return n;
720 }
721 }
722 return r;
723}
724
725/**
726 * htb_lookup_leaf - returns next leaf class in DRR order
727 *
728 * Find leaf where current feed pointers points to.
729 */
730static struct htb_class *htb_lookup_leaf(struct rb_root *tree, int prio,
731 struct rb_node **pptr, u32 * pid)
732{
733 int i;
734 struct {
735 struct rb_node *root;
736 struct rb_node **pptr;
737 u32 *pid;
738 } stk[TC_HTB_MAXDEPTH], *sp = stk;
739
740 BUG_ON(!tree->rb_node);
741 sp->root = tree->rb_node;
742 sp->pptr = pptr;
743 sp->pid = pid;
744
745 for (i = 0; i < 65535; i++) {
746 if (!*sp->pptr && *sp->pid) {
747 /* ptr was invalidated but id is valid - try to recover
748 * the original or next ptr
749 */
750 *sp->pptr =
751 htb_id_find_next_upper(prio, sp->root, *sp->pid);
752 }
753 *sp->pid = 0; /* ptr is valid now so that remove this hint as it
754 * can become out of date quickly
755 */
756 if (!*sp->pptr) { /* we are at right end; rewind & go up */
757 *sp->pptr = sp->root;
758 while ((*sp->pptr)->rb_left)
759 *sp->pptr = (*sp->pptr)->rb_left;
760 if (sp > stk) {
761 sp--;
762 if (!*sp->pptr) {
763 WARN_ON(1);
764 return NULL;
765 }
766 htb_next_rb_node(sp->pptr);
767 }
768 } else {
769 struct htb_class *cl;
770 cl = rb_entry(*sp->pptr, struct htb_class, node[prio]);
771 if (!cl->level)
772 return cl;
773 (++sp)->root = cl->un.inner.feed[prio].rb_node;
774 sp->pptr = cl->un.inner.ptr + prio;
775 sp->pid = cl->un.inner.last_ptr_id + prio;
776 }
777 }
778 WARN_ON(1);
779 return NULL;
780}
781
782/* dequeues packet at given priority and level; call only if
783 * you are sure that there is active class at prio/level
784 */
785static struct sk_buff *htb_dequeue_tree(struct htb_sched *q, int prio,
786 int level)
787{
788 struct sk_buff *skb = NULL;
789 struct htb_class *cl, *start;
790 /* look initial class up in the row */
791 start = cl = htb_lookup_leaf(q->row[level] + prio, prio,
792 q->ptr[level] + prio,
793 q->last_ptr_id[level] + prio);
794
795 do {
796next:
797 if (unlikely(!cl))
798 return NULL;
799
800 /* class can be empty - it is unlikely but can be true if leaf
801 * qdisc drops packets in enqueue routine or if someone used
802 * graft operation on the leaf since last dequeue;
803 * simply deactivate and skip such class
804 */
805 if (unlikely(cl->un.leaf.q->q.qlen == 0)) {
806 struct htb_class *next;
807 htb_deactivate(q, cl);
808
809 /* row/level might become empty */
810 if ((q->row_mask[level] & (1 << prio)) == 0)
811 return NULL;
812
813 next = htb_lookup_leaf(q->row[level] + prio,
814 prio, q->ptr[level] + prio,
815 q->last_ptr_id[level] + prio);
816
817 if (cl == start) /* fix start if we just deleted it */
818 start = next;
819 cl = next;
820 goto next;
821 }
822
823 skb = cl->un.leaf.q->dequeue(cl->un.leaf.q);
824 if (likely(skb != NULL))
825 break;
826
827 qdisc_warn_nonwc("htb", cl->un.leaf.q);
828 htb_next_rb_node((level ? cl->parent->un.inner.ptr : q->
829 ptr[0]) + prio);
830 cl = htb_lookup_leaf(q->row[level] + prio, prio,
831 q->ptr[level] + prio,
832 q->last_ptr_id[level] + prio);
833
834 } while (cl != start);
835
836 if (likely(skb != NULL)) {
837 bstats_update(&cl->bstats, skb);
838 cl->un.leaf.deficit[level] -= qdisc_pkt_len(skb);
839 if (cl->un.leaf.deficit[level] < 0) {
840 cl->un.leaf.deficit[level] += cl->quantum;
841 htb_next_rb_node((level ? cl->parent->un.inner.ptr : q->
842 ptr[0]) + prio);
843 }
844 /* this used to be after charge_class but this constelation
845 * gives us slightly better performance
846 */
847 if (!cl->un.leaf.q->q.qlen)
848 htb_deactivate(q, cl);
849 htb_charge_class(q, cl, level, skb);
850 }
851 return skb;
852}
853
854static struct sk_buff *htb_dequeue(struct Qdisc *sch)
855{
856 struct sk_buff *skb;
857 struct htb_sched *q = qdisc_priv(sch);
858 int level;
859 psched_time_t next_event;
860 unsigned long start_at;
861
862 /* try to dequeue direct packets as high prio (!) to minimize cpu work */
863 skb = __skb_dequeue(&q->direct_queue);
864 if (skb != NULL) {
865ok:
866 qdisc_bstats_update(sch, skb);
867 qdisc_unthrottled(sch);
868 sch->q.qlen--;
869 return skb;
870 }
871
872 if (!sch->q.qlen)
873 goto fin;
874 q->now = psched_get_time();
875 start_at = jiffies;
876
877 next_event = q->now + 5 * PSCHED_TICKS_PER_SEC;
878
879 for (level = 0; level < TC_HTB_MAXDEPTH; level++) {
880 /* common case optimization - skip event handler quickly */
881 int m;
882 psched_time_t event;
883
884 if (q->now >= q->near_ev_cache[level]) {
885 event = htb_do_events(q, level, start_at);
886 if (!event)
887 event = q->now + PSCHED_TICKS_PER_SEC;
888 q->near_ev_cache[level] = event;
889 } else
890 event = q->near_ev_cache[level];
891
892 if (next_event > event)
893 next_event = event;
894
895 m = ~q->row_mask[level];
896 while (m != (int)(-1)) {
897 int prio = ffz(m);
898
899 m |= 1 << prio;
900 skb = htb_dequeue_tree(q, prio, level);
901 if (likely(skb != NULL))
902 goto ok;
903 }
904 }
905 sch->qstats.overlimits++;
906 if (likely(next_event > q->now))
907 qdisc_watchdog_schedule(&q->watchdog, next_event);
908 else
909 schedule_work(&q->work);
910fin:
911 return skb;
912}
913
914/* try to drop from each class (by prio) until one succeed */
915static unsigned int htb_drop(struct Qdisc *sch)
916{
917 struct htb_sched *q = qdisc_priv(sch);
918 int prio;
919
920 for (prio = TC_HTB_NUMPRIO - 1; prio >= 0; prio--) {
921 struct list_head *p;
922 list_for_each(p, q->drops + prio) {
923 struct htb_class *cl = list_entry(p, struct htb_class,
924 un.leaf.drop_list);
925 unsigned int len;
926 if (cl->un.leaf.q->ops->drop &&
927 (len = cl->un.leaf.q->ops->drop(cl->un.leaf.q))) {
928 sch->q.qlen--;
929 if (!cl->un.leaf.q->q.qlen)
930 htb_deactivate(q, cl);
931 return len;
932 }
933 }
934 }
935 return 0;
936}
937
938/* reset all classes */
939/* always caled under BH & queue lock */
940static void htb_reset(struct Qdisc *sch)
941{
942 struct htb_sched *q = qdisc_priv(sch);
943 struct htb_class *cl;
944 struct hlist_node *n;
945 unsigned int i;
946
947 for (i = 0; i < q->clhash.hashsize; i++) {
948 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode) {
949 if (cl->level)
950 memset(&cl->un.inner, 0, sizeof(cl->un.inner));
951 else {
952 if (cl->un.leaf.q)
953 qdisc_reset(cl->un.leaf.q);
954 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
955 }
956 cl->prio_activity = 0;
957 cl->cmode = HTB_CAN_SEND;
958
959 }
960 }
961 qdisc_watchdog_cancel(&q->watchdog);
962 __skb_queue_purge(&q->direct_queue);
963 sch->q.qlen = 0;
964 memset(q->row, 0, sizeof(q->row));
965 memset(q->row_mask, 0, sizeof(q->row_mask));
966 memset(q->wait_pq, 0, sizeof(q->wait_pq));
967 memset(q->ptr, 0, sizeof(q->ptr));
968 for (i = 0; i < TC_HTB_NUMPRIO; i++)
969 INIT_LIST_HEAD(q->drops + i);
970}
971
972static const struct nla_policy htb_policy[TCA_HTB_MAX + 1] = {
973 [TCA_HTB_PARMS] = { .len = sizeof(struct tc_htb_opt) },
974 [TCA_HTB_INIT] = { .len = sizeof(struct tc_htb_glob) },
975 [TCA_HTB_CTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
976 [TCA_HTB_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
977};
978
979static void htb_work_func(struct work_struct *work)
980{
981 struct htb_sched *q = container_of(work, struct htb_sched, work);
982 struct Qdisc *sch = q->watchdog.qdisc;
983
984 __netif_schedule(qdisc_root(sch));
985}
986
987static int htb_init(struct Qdisc *sch, struct nlattr *opt)
988{
989 struct htb_sched *q = qdisc_priv(sch);
990 struct nlattr *tb[TCA_HTB_INIT + 1];
991 struct tc_htb_glob *gopt;
992 int err;
993 int i;
994
995 if (!opt)
996 return -EINVAL;
997
998 err = nla_parse_nested(tb, TCA_HTB_INIT, opt, htb_policy);
999 if (err < 0)
1000 return err;
1001
1002 if (tb[TCA_HTB_INIT] == NULL) {
1003 pr_err("HTB: hey probably you have bad tc tool ?\n");
1004 return -EINVAL;
1005 }
1006 gopt = nla_data(tb[TCA_HTB_INIT]);
1007 if (gopt->version != HTB_VER >> 16) {
1008 pr_err("HTB: need tc/htb version %d (minor is %d), you have %d\n",
1009 HTB_VER >> 16, HTB_VER & 0xffff, gopt->version);
1010 return -EINVAL;
1011 }
1012
1013 err = qdisc_class_hash_init(&q->clhash);
1014 if (err < 0)
1015 return err;
1016 for (i = 0; i < TC_HTB_NUMPRIO; i++)
1017 INIT_LIST_HEAD(q->drops + i);
1018
1019 qdisc_watchdog_init(&q->watchdog, sch);
1020 INIT_WORK(&q->work, htb_work_func);
1021 skb_queue_head_init(&q->direct_queue);
1022
1023 q->direct_qlen = qdisc_dev(sch)->tx_queue_len;
1024 if (q->direct_qlen < 2) /* some devices have zero tx_queue_len */
1025 q->direct_qlen = 2;
1026
1027 if ((q->rate2quantum = gopt->rate2quantum) < 1)
1028 q->rate2quantum = 1;
1029 q->defcls = gopt->defcls;
1030
1031 return 0;
1032}
1033
1034static int htb_dump(struct Qdisc *sch, struct sk_buff *skb)
1035{
1036 spinlock_t *root_lock = qdisc_root_sleeping_lock(sch);
1037 struct htb_sched *q = qdisc_priv(sch);
1038 struct nlattr *nest;
1039 struct tc_htb_glob gopt;
1040
1041 spin_lock_bh(root_lock);
1042
1043 gopt.direct_pkts = q->direct_pkts;
1044 gopt.version = HTB_VER;
1045 gopt.rate2quantum = q->rate2quantum;
1046 gopt.defcls = q->defcls;
1047 gopt.debug = 0;
1048
1049 nest = nla_nest_start(skb, TCA_OPTIONS);
1050 if (nest == NULL)
1051 goto nla_put_failure;
1052 if (nla_put(skb, TCA_HTB_INIT, sizeof(gopt), &gopt))
1053 goto nla_put_failure;
1054 nla_nest_end(skb, nest);
1055
1056 spin_unlock_bh(root_lock);
1057 return skb->len;
1058
1059nla_put_failure:
1060 spin_unlock_bh(root_lock);
1061 nla_nest_cancel(skb, nest);
1062 return -1;
1063}
1064
1065static int htb_dump_class(struct Qdisc *sch, unsigned long arg,
1066 struct sk_buff *skb, struct tcmsg *tcm)
1067{
1068 struct htb_class *cl = (struct htb_class *)arg;
1069 spinlock_t *root_lock = qdisc_root_sleeping_lock(sch);
1070 struct nlattr *nest;
1071 struct tc_htb_opt opt;
1072
1073 spin_lock_bh(root_lock);
1074 tcm->tcm_parent = cl->parent ? cl->parent->common.classid : TC_H_ROOT;
1075 tcm->tcm_handle = cl->common.classid;
1076 if (!cl->level && cl->un.leaf.q)
1077 tcm->tcm_info = cl->un.leaf.q->handle;
1078
1079 nest = nla_nest_start(skb, TCA_OPTIONS);
1080 if (nest == NULL)
1081 goto nla_put_failure;
1082
1083 memset(&opt, 0, sizeof(opt));
1084
1085 opt.rate = cl->rate->rate;
1086 opt.buffer = cl->buffer;
1087 opt.ceil = cl->ceil->rate;
1088 opt.cbuffer = cl->cbuffer;
1089 opt.quantum = cl->quantum;
1090 opt.prio = cl->prio;
1091 opt.level = cl->level;
1092 if (nla_put(skb, TCA_HTB_PARMS, sizeof(opt), &opt))
1093 goto nla_put_failure;
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/*
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/sch_generic.h>
42#include <net/pkt_sched.h>
43
44/* HTB algorithm.
45 Author: devik@cdi.cz
46 ========================================================================
47 HTB is like TBF with multiple classes. It is also similar to CBQ because
48 it allows to assign priority to each class in hierarchy.
49 In fact it is another implementation of Floyd's formal sharing.
50
51 Levels:
52 Each class is assigned level. Leaf has ALWAYS level 0 and root
53 classes have level TC_HTB_MAXDEPTH-1. Interior nodes has level
54 one less than their parent.
55*/
56
57static int htb_hysteresis __read_mostly = 0; /* whether to use mode hysteresis for speedup */
58#define HTB_VER 0x30011 /* major must be matched with number suplied by TC as version */
59
60#if HTB_VER >> 16 != TC_HTB_PROTOVER
61#error "Mismatched sch_htb.c and pkt_sch.h"
62#endif
63
64/* Module parameter and sysfs export */
65module_param (htb_hysteresis, int, 0640);
66MODULE_PARM_DESC(htb_hysteresis, "Hysteresis mode, less CPU load, less accurate");
67
68static int htb_rate_est = 0; /* htb classes have a default rate estimator */
69module_param(htb_rate_est, int, 0640);
70MODULE_PARM_DESC(htb_rate_est, "setup a default rate estimator (4sec 16sec) for htb classes");
71
72/* used internaly to keep status of single class */
73enum htb_cmode {
74 HTB_CANT_SEND, /* class can't send and can't borrow */
75 HTB_MAY_BORROW, /* class can't send but may borrow */
76 HTB_CAN_SEND /* class can send */
77};
78
79struct htb_prio {
80 union {
81 struct rb_root row;
82 struct rb_root feed;
83 };
84 struct rb_node *ptr;
85 /* When class changes from state 1->2 and disconnects from
86 * parent's feed then we lost ptr value and start from the
87 * first child again. Here we store classid of the
88 * last valid ptr (used when ptr is NULL).
89 */
90 u32 last_ptr_id;
91};
92
93/* interior & leaf nodes; props specific to leaves are marked L:
94 * To reduce false sharing, place mostly read fields at beginning,
95 * and mostly written ones at the end.
96 */
97struct htb_class {
98 struct Qdisc_class_common common;
99 struct psched_ratecfg rate;
100 struct psched_ratecfg ceil;
101 s64 buffer, cbuffer;/* token bucket depth/rate */
102 s64 mbuffer; /* max wait time */
103 u32 prio; /* these two are used only by leaves... */
104 int quantum; /* but stored for parent-to-leaf return */
105
106 struct tcf_proto __rcu *filter_list; /* class attached filters */
107 int filter_cnt;
108 int refcnt; /* usage count of this class */
109
110 int level; /* our level (see above) */
111 unsigned int children;
112 struct htb_class *parent; /* parent class */
113
114 struct gnet_stats_rate_est64 rate_est;
115
116 /*
117 * Written often fields
118 */
119 struct gnet_stats_basic_packed bstats;
120 struct gnet_stats_queue qstats;
121 struct tc_htb_xstats xstats; /* our special stats */
122
123 /* token bucket parameters */
124 s64 tokens, ctokens;/* current number of tokens */
125 s64 t_c; /* checkpoint time */
126
127 union {
128 struct htb_class_leaf {
129 struct list_head drop_list;
130 int deficit[TC_HTB_MAXDEPTH];
131 struct Qdisc *q;
132 } leaf;
133 struct htb_class_inner {
134 struct htb_prio clprio[TC_HTB_NUMPRIO];
135 } inner;
136 } un;
137 s64 pq_key;
138
139 int prio_activity; /* for which prios are we active */
140 enum htb_cmode cmode; /* current mode of the class */
141 struct rb_node pq_node; /* node for event queue */
142 struct rb_node node[TC_HTB_NUMPRIO]; /* node for self or feed tree */
143};
144
145struct htb_level {
146 struct rb_root wait_pq;
147 struct htb_prio hprio[TC_HTB_NUMPRIO];
148};
149
150struct htb_sched {
151 struct Qdisc_class_hash clhash;
152 int defcls; /* class where unclassified flows go to */
153 int rate2quantum; /* quant = rate / rate2quantum */
154
155 /* filters for qdisc itself */
156 struct tcf_proto __rcu *filter_list;
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 sk_buff_head direct_queue;
165 long direct_pkts;
166
167 struct qdisc_watchdog watchdog;
168
169 s64 now; /* cached dequeue time */
170 struct list_head drops[TC_HTB_NUMPRIO];/* active leaves (for drops) */
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
180/* find class in global hash table using given handle */
181static inline struct htb_class *htb_find(u32 handle, struct Qdisc *sch)
182{
183 struct htb_sched *q = qdisc_priv(sch);
184 struct Qdisc_class_common *clc;
185
186 clc = qdisc_class_find(&q->clhash, handle);
187 if (clc == NULL)
188 return NULL;
189 return container_of(clc, struct htb_class, common);
190}
191
192/**
193 * htb_classify - classify a packet into class
194 *
195 * It returns NULL if the packet should be dropped or -1 if the packet
196 * should be passed directly thru. In all other cases leaf class is returned.
197 * We allow direct class selection by classid in priority. The we examine
198 * filters in qdisc and in inner nodes (if higher filter points to the inner
199 * node). If we end up with classid MAJOR:0 we enqueue the skb into special
200 * internal fifo (direct). These packets then go directly thru. If we still
201 * have no valid leaf we try to use MAJOR:default leaf. It still unsuccessful
202 * then finish and return direct queue.
203 */
204#define HTB_DIRECT ((struct htb_class *)-1L)
205
206static struct htb_class *htb_classify(struct sk_buff *skb, struct Qdisc *sch,
207 int *qerr)
208{
209 struct htb_sched *q = qdisc_priv(sch);
210 struct htb_class *cl;
211 struct tcf_result res;
212 struct tcf_proto *tcf;
213 int result;
214
215 /* allow to select class by setting skb->priority to valid classid;
216 * note that nfmark can be used too by attaching filter fw with no
217 * rules in it
218 */
219 if (skb->priority == sch->handle)
220 return HTB_DIRECT; /* X:0 (direct flow) selected */
221 cl = htb_find(skb->priority, sch);
222 if (cl) {
223 if (cl->level == 0)
224 return cl;
225 /* Start with inner filter chain if a non-leaf class is selected */
226 tcf = rcu_dereference_bh(cl->filter_list);
227 } else {
228 tcf = rcu_dereference_bh(q->filter_list);
229 }
230
231 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
232 while (tcf && (result = tc_classify(skb, tcf, &res, false)) >= 0) {
233#ifdef CONFIG_NET_CLS_ACT
234 switch (result) {
235 case TC_ACT_QUEUED:
236 case TC_ACT_STOLEN:
237 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
238 case TC_ACT_SHOT:
239 return NULL;
240 }
241#endif
242 cl = (void *)res.class;
243 if (!cl) {
244 if (res.classid == sch->handle)
245 return HTB_DIRECT; /* X:0 (direct flow) */
246 cl = htb_find(res.classid, sch);
247 if (!cl)
248 break; /* filter selected invalid classid */
249 }
250 if (!cl->level)
251 return cl; /* we hit leaf; return it */
252
253 /* we have got inner class; apply inner filter chain */
254 tcf = rcu_dereference_bh(cl->filter_list);
255 }
256 /* classification failed; try to use default class */
257 cl = htb_find(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
258 if (!cl || cl->level)
259 return HTB_DIRECT; /* bad default .. this is safe bet */
260 return cl;
261}
262
263/**
264 * htb_add_to_id_tree - adds class to the round robin list
265 *
266 * Routine adds class to the list (actually tree) sorted by classid.
267 * Make sure that class is not already on such list for given prio.
268 */
269static void htb_add_to_id_tree(struct rb_root *root,
270 struct htb_class *cl, int prio)
271{
272 struct rb_node **p = &root->rb_node, *parent = NULL;
273
274 while (*p) {
275 struct htb_class *c;
276 parent = *p;
277 c = rb_entry(parent, struct htb_class, node[prio]);
278
279 if (cl->common.classid > c->common.classid)
280 p = &parent->rb_right;
281 else
282 p = &parent->rb_left;
283 }
284 rb_link_node(&cl->node[prio], parent, p);
285 rb_insert_color(&cl->node[prio], root);
286}
287
288/**
289 * htb_add_to_wait_tree - adds class to the event queue with delay
290 *
291 * The class is added to priority event queue to indicate that class will
292 * change its mode in cl->pq_key microseconds. Make sure that class is not
293 * already in the queue.
294 */
295static void htb_add_to_wait_tree(struct htb_sched *q,
296 struct htb_class *cl, s64 delay)
297{
298 struct rb_node **p = &q->hlevel[cl->level].wait_pq.rb_node, *parent = NULL;
299
300 cl->pq_key = q->now + delay;
301 if (cl->pq_key == q->now)
302 cl->pq_key++;
303
304 /* update the nearest event cache */
305 if (q->near_ev_cache[cl->level] > cl->pq_key)
306 q->near_ev_cache[cl->level] = cl->pq_key;
307
308 while (*p) {
309 struct htb_class *c;
310 parent = *p;
311 c = rb_entry(parent, struct htb_class, pq_node);
312 if (cl->pq_key >= c->pq_key)
313 p = &parent->rb_right;
314 else
315 p = &parent->rb_left;
316 }
317 rb_link_node(&cl->pq_node, parent, p);
318 rb_insert_color(&cl->pq_node, &q->hlevel[cl->level].wait_pq);
319}
320
321/**
322 * htb_next_rb_node - finds next node in binary tree
323 *
324 * When we are past last key we return NULL.
325 * Average complexity is 2 steps per call.
326 */
327static inline void htb_next_rb_node(struct rb_node **n)
328{
329 *n = rb_next(*n);
330}
331
332/**
333 * htb_add_class_to_row - add class to its row
334 *
335 * The class is added to row at priorities marked in mask.
336 * It does nothing if mask == 0.
337 */
338static inline void htb_add_class_to_row(struct htb_sched *q,
339 struct htb_class *cl, int mask)
340{
341 q->row_mask[cl->level] |= mask;
342 while (mask) {
343 int prio = ffz(~mask);
344 mask &= ~(1 << prio);
345 htb_add_to_id_tree(&q->hlevel[cl->level].hprio[prio].row, cl, prio);
346 }
347}
348
349/* If this triggers, it is a bug in this code, but it need not be fatal */
350static void htb_safe_rb_erase(struct rb_node *rb, struct rb_root *root)
351{
352 if (RB_EMPTY_NODE(rb)) {
353 WARN_ON(1);
354 } else {
355 rb_erase(rb, root);
356 RB_CLEAR_NODE(rb);
357 }
358}
359
360
361/**
362 * htb_remove_class_from_row - removes class from its row
363 *
364 * The class is removed from row at priorities marked in mask.
365 * It does nothing if mask == 0.
366 */
367static inline void htb_remove_class_from_row(struct htb_sched *q,
368 struct htb_class *cl, int mask)
369{
370 int m = 0;
371 struct htb_level *hlevel = &q->hlevel[cl->level];
372
373 while (mask) {
374 int prio = ffz(~mask);
375 struct htb_prio *hprio = &hlevel->hprio[prio];
376
377 mask &= ~(1 << prio);
378 if (hprio->ptr == cl->node + prio)
379 htb_next_rb_node(&hprio->ptr);
380
381 htb_safe_rb_erase(cl->node + prio, &hprio->row);
382 if (!hprio->row.rb_node)
383 m |= 1 << prio;
384 }
385 q->row_mask[cl->level] &= ~m;
386}
387
388/**
389 * htb_activate_prios - creates active classe's feed chain
390 *
391 * The class is connected to ancestors and/or appropriate rows
392 * for priorities it is participating on. cl->cmode must be new
393 * (activated) mode. It does nothing if cl->prio_activity == 0.
394 */
395static void htb_activate_prios(struct htb_sched *q, struct htb_class *cl)
396{
397 struct htb_class *p = cl->parent;
398 long m, mask = cl->prio_activity;
399
400 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
401 m = mask;
402 while (m) {
403 int prio = ffz(~m);
404 m &= ~(1 << prio);
405
406 if (p->un.inner.clprio[prio].feed.rb_node)
407 /* parent already has its feed in use so that
408 * reset bit in mask as parent is already ok
409 */
410 mask &= ~(1 << prio);
411
412 htb_add_to_id_tree(&p->un.inner.clprio[prio].feed, cl, prio);
413 }
414 p->prio_activity |= mask;
415 cl = p;
416 p = cl->parent;
417
418 }
419 if (cl->cmode == HTB_CAN_SEND && mask)
420 htb_add_class_to_row(q, cl, mask);
421}
422
423/**
424 * htb_deactivate_prios - remove class from feed chain
425 *
426 * cl->cmode must represent old mode (before deactivation). It does
427 * nothing if cl->prio_activity == 0. Class is removed from all feed
428 * chains and rows.
429 */
430static void htb_deactivate_prios(struct htb_sched *q, struct htb_class *cl)
431{
432 struct htb_class *p = cl->parent;
433 long m, mask = cl->prio_activity;
434
435 while (cl->cmode == HTB_MAY_BORROW && p && mask) {
436 m = mask;
437 mask = 0;
438 while (m) {
439 int prio = ffz(~m);
440 m &= ~(1 << prio);
441
442 if (p->un.inner.clprio[prio].ptr == cl->node + prio) {
443 /* we are removing child which is pointed to from
444 * parent feed - forget the pointer but remember
445 * classid
446 */
447 p->un.inner.clprio[prio].last_ptr_id = cl->common.classid;
448 p->un.inner.clprio[prio].ptr = NULL;
449 }
450
451 htb_safe_rb_erase(cl->node + prio,
452 &p->un.inner.clprio[prio].feed);
453
454 if (!p->un.inner.clprio[prio].feed.rb_node)
455 mask |= 1 << prio;
456 }
457
458 p->prio_activity &= ~mask;
459 cl = p;
460 p = cl->parent;
461
462 }
463 if (cl->cmode == HTB_CAN_SEND && mask)
464 htb_remove_class_from_row(q, cl, mask);
465}
466
467static inline s64 htb_lowater(const struct htb_class *cl)
468{
469 if (htb_hysteresis)
470 return cl->cmode != HTB_CANT_SEND ? -cl->cbuffer : 0;
471 else
472 return 0;
473}
474static inline s64 htb_hiwater(const struct htb_class *cl)
475{
476 if (htb_hysteresis)
477 return cl->cmode == HTB_CAN_SEND ? -cl->buffer : 0;
478 else
479 return 0;
480}
481
482
483/**
484 * htb_class_mode - computes and returns current class mode
485 *
486 * It computes cl's mode at time cl->t_c+diff and returns it. If mode
487 * is not HTB_CAN_SEND then cl->pq_key is updated to time difference
488 * from now to time when cl will change its state.
489 * Also it is worth to note that class mode doesn't change simply
490 * at cl->{c,}tokens == 0 but there can rather be hysteresis of
491 * 0 .. -cl->{c,}buffer range. It is meant to limit number of
492 * mode transitions per time unit. The speed gain is about 1/6.
493 */
494static inline enum htb_cmode
495htb_class_mode(struct htb_class *cl, s64 *diff)
496{
497 s64 toks;
498
499 if ((toks = (cl->ctokens + *diff)) < htb_lowater(cl)) {
500 *diff = -toks;
501 return HTB_CANT_SEND;
502 }
503
504 if ((toks = (cl->tokens + *diff)) >= htb_hiwater(cl))
505 return HTB_CAN_SEND;
506
507 *diff = -toks;
508 return HTB_MAY_BORROW;
509}
510
511/**
512 * htb_change_class_mode - changes classe's mode
513 *
514 * This should be the only way how to change classe's mode under normal
515 * cirsumstances. Routine will update feed lists linkage, change mode
516 * and add class to the wait event queue if appropriate. New mode should
517 * be different from old one and cl->pq_key has to be valid if changing
518 * to mode other than HTB_CAN_SEND (see htb_add_to_wait_tree).
519 */
520static void
521htb_change_class_mode(struct htb_sched *q, struct htb_class *cl, s64 *diff)
522{
523 enum htb_cmode new_mode = htb_class_mode(cl, diff);
524
525 if (new_mode == cl->cmode)
526 return;
527
528 if (cl->prio_activity) { /* not necessary: speed optimization */
529 if (cl->cmode != HTB_CANT_SEND)
530 htb_deactivate_prios(q, cl);
531 cl->cmode = new_mode;
532 if (new_mode != HTB_CANT_SEND)
533 htb_activate_prios(q, cl);
534 } else
535 cl->cmode = new_mode;
536}
537
538/**
539 * htb_activate - inserts leaf cl into appropriate active feeds
540 *
541 * Routine learns (new) priority of leaf and activates feed chain
542 * for the prio. It can be called on already active leaf safely.
543 * It also adds leaf into droplist.
544 */
545static inline void htb_activate(struct htb_sched *q, struct htb_class *cl)
546{
547 WARN_ON(cl->level || !cl->un.leaf.q || !cl->un.leaf.q->q.qlen);
548
549 if (!cl->prio_activity) {
550 cl->prio_activity = 1 << cl->prio;
551 htb_activate_prios(q, cl);
552 list_add_tail(&cl->un.leaf.drop_list,
553 q->drops + cl->prio);
554 }
555}
556
557/**
558 * htb_deactivate - remove leaf cl from active feeds
559 *
560 * Make sure that leaf is active. In the other words it can't be called
561 * with non-active leaf. It also removes class from the drop list.
562 */
563static inline void htb_deactivate(struct htb_sched *q, struct htb_class *cl)
564{
565 WARN_ON(!cl->prio_activity);
566
567 htb_deactivate_prios(q, cl);
568 cl->prio_activity = 0;
569 list_del_init(&cl->un.leaf.drop_list);
570}
571
572static int htb_enqueue(struct sk_buff *skb, struct Qdisc *sch)
573{
574 int uninitialized_var(ret);
575 struct htb_sched *q = qdisc_priv(sch);
576 struct htb_class *cl = htb_classify(skb, sch, &ret);
577
578 if (cl == HTB_DIRECT) {
579 /* enqueue to helper queue */
580 if (q->direct_queue.qlen < q->direct_qlen) {
581 __skb_queue_tail(&q->direct_queue, skb);
582 q->direct_pkts++;
583 } else {
584 return qdisc_drop(skb, sch);
585 }
586#ifdef CONFIG_NET_CLS_ACT
587 } else if (!cl) {
588 if (ret & __NET_XMIT_BYPASS)
589 qdisc_qstats_drop(sch);
590 kfree_skb(skb);
591 return ret;
592#endif
593 } else if ((ret = qdisc_enqueue(skb, cl->un.leaf.q)) != NET_XMIT_SUCCESS) {
594 if (net_xmit_drop_count(ret)) {
595 qdisc_qstats_drop(sch);
596 cl->qstats.drops++;
597 }
598 return ret;
599 } else {
600 htb_activate(q, cl);
601 }
602
603 qdisc_qstats_backlog_inc(sch, skb);
604 sch->q.qlen++;
605 return NET_XMIT_SUCCESS;
606}
607
608static inline void htb_accnt_tokens(struct htb_class *cl, int bytes, s64 diff)
609{
610 s64 toks = diff + cl->tokens;
611
612 if (toks > cl->buffer)
613 toks = cl->buffer;
614 toks -= (s64) psched_l2t_ns(&cl->rate, bytes);
615 if (toks <= -cl->mbuffer)
616 toks = 1 - cl->mbuffer;
617
618 cl->tokens = toks;
619}
620
621static inline void htb_accnt_ctokens(struct htb_class *cl, int bytes, s64 diff)
622{
623 s64 toks = diff + cl->ctokens;
624
625 if (toks > cl->cbuffer)
626 toks = cl->cbuffer;
627 toks -= (s64) psched_l2t_ns(&cl->ceil, bytes);
628 if (toks <= -cl->mbuffer)
629 toks = 1 - cl->mbuffer;
630
631 cl->ctokens = toks;
632}
633
634/**
635 * htb_charge_class - charges amount "bytes" to leaf and ancestors
636 *
637 * Routine assumes that packet "bytes" long was dequeued from leaf cl
638 * borrowing from "level". It accounts bytes to ceil leaky bucket for
639 * leaf and all ancestors and to rate bucket for ancestors at levels
640 * "level" and higher. It also handles possible change of mode resulting
641 * from the update. Note that mode can also increase here (MAY_BORROW to
642 * CAN_SEND) because we can use more precise clock that event queue here.
643 * In such case we remove class from event queue first.
644 */
645static void htb_charge_class(struct htb_sched *q, struct htb_class *cl,
646 int level, struct sk_buff *skb)
647{
648 int bytes = qdisc_pkt_len(skb);
649 enum htb_cmode old_mode;
650 s64 diff;
651
652 while (cl) {
653 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
654 if (cl->level >= level) {
655 if (cl->level == level)
656 cl->xstats.lends++;
657 htb_accnt_tokens(cl, bytes, diff);
658 } else {
659 cl->xstats.borrows++;
660 cl->tokens += diff; /* we moved t_c; update tokens */
661 }
662 htb_accnt_ctokens(cl, bytes, diff);
663 cl->t_c = q->now;
664
665 old_mode = cl->cmode;
666 diff = 0;
667 htb_change_class_mode(q, cl, &diff);
668 if (old_mode != cl->cmode) {
669 if (old_mode != HTB_CAN_SEND)
670 htb_safe_rb_erase(&cl->pq_node, &q->hlevel[cl->level].wait_pq);
671 if (cl->cmode != HTB_CAN_SEND)
672 htb_add_to_wait_tree(q, cl, diff);
673 }
674
675 /* update basic stats except for leaves which are already updated */
676 if (cl->level)
677 bstats_update(&cl->bstats, skb);
678
679 cl = cl->parent;
680 }
681}
682
683/**
684 * htb_do_events - make mode changes to classes at the level
685 *
686 * Scans event queue for pending events and applies them. Returns time of
687 * next pending event (0 for no event in pq, q->now for too many events).
688 * Note: Applied are events whose have cl->pq_key <= q->now.
689 */
690static s64 htb_do_events(struct htb_sched *q, const int level,
691 unsigned long start)
692{
693 /* don't run for longer than 2 jiffies; 2 is used instead of
694 * 1 to simplify things when jiffy is going to be incremented
695 * too soon
696 */
697 unsigned long stop_at = start + 2;
698 struct rb_root *wait_pq = &q->hlevel[level].wait_pq;
699
700 while (time_before(jiffies, stop_at)) {
701 struct htb_class *cl;
702 s64 diff;
703 struct rb_node *p = rb_first(wait_pq);
704
705 if (!p)
706 return 0;
707
708 cl = rb_entry(p, struct htb_class, pq_node);
709 if (cl->pq_key > q->now)
710 return cl->pq_key;
711
712 htb_safe_rb_erase(p, wait_pq);
713 diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
714 htb_change_class_mode(q, cl, &diff);
715 if (cl->cmode != HTB_CAN_SEND)
716 htb_add_to_wait_tree(q, cl, diff);
717 }
718
719 /* too much load - let's continue after a break for scheduling */
720 if (!(q->warned & HTB_WARN_TOOMANYEVENTS)) {
721 pr_warn("htb: too many events!\n");
722 q->warned |= HTB_WARN_TOOMANYEVENTS;
723 }
724
725 return q->now;
726}
727
728/* Returns class->node+prio from id-tree where classe's id is >= id. NULL
729 * is no such one exists.
730 */
731static struct rb_node *htb_id_find_next_upper(int prio, struct rb_node *n,
732 u32 id)
733{
734 struct rb_node *r = NULL;
735 while (n) {
736 struct htb_class *cl =
737 rb_entry(n, struct htb_class, node[prio]);
738
739 if (id > cl->common.classid) {
740 n = n->rb_right;
741 } else if (id < cl->common.classid) {
742 r = n;
743 n = n->rb_left;
744 } else {
745 return n;
746 }
747 }
748 return r;
749}
750
751/**
752 * htb_lookup_leaf - returns next leaf class in DRR order
753 *
754 * Find leaf where current feed pointers points to.
755 */
756static struct htb_class *htb_lookup_leaf(struct htb_prio *hprio, const int prio)
757{
758 int i;
759 struct {
760 struct rb_node *root;
761 struct rb_node **pptr;
762 u32 *pid;
763 } stk[TC_HTB_MAXDEPTH], *sp = stk;
764
765 BUG_ON(!hprio->row.rb_node);
766 sp->root = hprio->row.rb_node;
767 sp->pptr = &hprio->ptr;
768 sp->pid = &hprio->last_ptr_id;
769
770 for (i = 0; i < 65535; i++) {
771 if (!*sp->pptr && *sp->pid) {
772 /* ptr was invalidated but id is valid - try to recover
773 * the original or next ptr
774 */
775 *sp->pptr =
776 htb_id_find_next_upper(prio, sp->root, *sp->pid);
777 }
778 *sp->pid = 0; /* ptr is valid now so that remove this hint as it
779 * can become out of date quickly
780 */
781 if (!*sp->pptr) { /* we are at right end; rewind & go up */
782 *sp->pptr = sp->root;
783 while ((*sp->pptr)->rb_left)
784 *sp->pptr = (*sp->pptr)->rb_left;
785 if (sp > stk) {
786 sp--;
787 if (!*sp->pptr) {
788 WARN_ON(1);
789 return NULL;
790 }
791 htb_next_rb_node(sp->pptr);
792 }
793 } else {
794 struct htb_class *cl;
795 struct htb_prio *clp;
796
797 cl = rb_entry(*sp->pptr, struct htb_class, node[prio]);
798 if (!cl->level)
799 return cl;
800 clp = &cl->un.inner.clprio[prio];
801 (++sp)->root = clp->feed.rb_node;
802 sp->pptr = &clp->ptr;
803 sp->pid = &clp->last_ptr_id;
804 }
805 }
806 WARN_ON(1);
807 return NULL;
808}
809
810/* dequeues packet at given priority and level; call only if
811 * you are sure that there is active class at prio/level
812 */
813static struct sk_buff *htb_dequeue_tree(struct htb_sched *q, const int prio,
814 const int level)
815{
816 struct sk_buff *skb = NULL;
817 struct htb_class *cl, *start;
818 struct htb_level *hlevel = &q->hlevel[level];
819 struct htb_prio *hprio = &hlevel->hprio[prio];
820
821 /* look initial class up in the row */
822 start = cl = htb_lookup_leaf(hprio, prio);
823
824 do {
825next:
826 if (unlikely(!cl))
827 return NULL;
828
829 /* class can be empty - it is unlikely but can be true if leaf
830 * qdisc drops packets in enqueue routine or if someone used
831 * graft operation on the leaf since last dequeue;
832 * simply deactivate and skip such class
833 */
834 if (unlikely(cl->un.leaf.q->q.qlen == 0)) {
835 struct htb_class *next;
836 htb_deactivate(q, cl);
837
838 /* row/level might become empty */
839 if ((q->row_mask[level] & (1 << prio)) == 0)
840 return NULL;
841
842 next = htb_lookup_leaf(hprio, prio);
843
844 if (cl == start) /* fix start if we just deleted it */
845 start = next;
846 cl = next;
847 goto next;
848 }
849
850 skb = cl->un.leaf.q->dequeue(cl->un.leaf.q);
851 if (likely(skb != NULL))
852 break;
853
854 qdisc_warn_nonwc("htb", cl->un.leaf.q);
855 htb_next_rb_node(level ? &cl->parent->un.inner.clprio[prio].ptr:
856 &q->hlevel[0].hprio[prio].ptr);
857 cl = htb_lookup_leaf(hprio, prio);
858
859 } while (cl != start);
860
861 if (likely(skb != NULL)) {
862 bstats_update(&cl->bstats, skb);
863 cl->un.leaf.deficit[level] -= qdisc_pkt_len(skb);
864 if (cl->un.leaf.deficit[level] < 0) {
865 cl->un.leaf.deficit[level] += cl->quantum;
866 htb_next_rb_node(level ? &cl->parent->un.inner.clprio[prio].ptr :
867 &q->hlevel[0].hprio[prio].ptr);
868 }
869 /* this used to be after charge_class but this constelation
870 * gives us slightly better performance
871 */
872 if (!cl->un.leaf.q->q.qlen)
873 htb_deactivate(q, cl);
874 htb_charge_class(q, cl, level, skb);
875 }
876 return skb;
877}
878
879static struct sk_buff *htb_dequeue(struct Qdisc *sch)
880{
881 struct sk_buff *skb;
882 struct htb_sched *q = qdisc_priv(sch);
883 int level;
884 s64 next_event;
885 unsigned long start_at;
886
887 /* try to dequeue direct packets as high prio (!) to minimize cpu work */
888 skb = __skb_dequeue(&q->direct_queue);
889 if (skb != NULL) {
890ok:
891 qdisc_bstats_update(sch, skb);
892 qdisc_unthrottled(sch);
893 qdisc_qstats_backlog_dec(sch, skb);
894 sch->q.qlen--;
895 return skb;
896 }
897
898 if (!sch->q.qlen)
899 goto fin;
900 q->now = ktime_get_ns();
901 start_at = jiffies;
902
903 next_event = q->now + 5LLU * NSEC_PER_SEC;
904
905 for (level = 0; level < TC_HTB_MAXDEPTH; level++) {
906 /* common case optimization - skip event handler quickly */
907 int m;
908 s64 event = q->near_ev_cache[level];
909
910 if (q->now >= event) {
911 event = htb_do_events(q, level, start_at);
912 if (!event)
913 event = q->now + NSEC_PER_SEC;
914 q->near_ev_cache[level] = event;
915 }
916
917 if (next_event > event)
918 next_event = event;
919
920 m = ~q->row_mask[level];
921 while (m != (int)(-1)) {
922 int prio = ffz(m);
923
924 m |= 1 << prio;
925 skb = htb_dequeue_tree(q, prio, level);
926 if (likely(skb != NULL))
927 goto ok;
928 }
929 }
930 qdisc_qstats_overlimit(sch);
931 if (likely(next_event > q->now)) {
932 if (!test_bit(__QDISC_STATE_DEACTIVATED,
933 &qdisc_root_sleeping(q->watchdog.qdisc)->state)) {
934 ktime_t time = ns_to_ktime(next_event);
935 qdisc_throttled(q->watchdog.qdisc);
936 hrtimer_start(&q->watchdog.timer, time,
937 HRTIMER_MODE_ABS_PINNED);
938 }
939 } else {
940 schedule_work(&q->work);
941 }
942fin:
943 return skb;
944}
945
946/* try to drop from each class (by prio) until one succeed */
947static unsigned int htb_drop(struct Qdisc *sch)
948{
949 struct htb_sched *q = qdisc_priv(sch);
950 int prio;
951
952 for (prio = TC_HTB_NUMPRIO - 1; prio >= 0; prio--) {
953 struct list_head *p;
954 list_for_each(p, q->drops + prio) {
955 struct htb_class *cl = list_entry(p, struct htb_class,
956 un.leaf.drop_list);
957 unsigned int len;
958 if (cl->un.leaf.q->ops->drop &&
959 (len = cl->un.leaf.q->ops->drop(cl->un.leaf.q))) {
960 sch->qstats.backlog -= len;
961 sch->q.qlen--;
962 if (!cl->un.leaf.q->q.qlen)
963 htb_deactivate(q, cl);
964 return len;
965 }
966 }
967 }
968 return 0;
969}
970
971/* reset all classes */
972/* always caled under BH & queue lock */
973static void htb_reset(struct Qdisc *sch)
974{
975 struct htb_sched *q = qdisc_priv(sch);
976 struct htb_class *cl;
977 unsigned int i;
978
979 for (i = 0; i < q->clhash.hashsize; i++) {
980 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
981 if (cl->level)
982 memset(&cl->un.inner, 0, sizeof(cl->un.inner));
983 else {
984 if (cl->un.leaf.q)
985 qdisc_reset(cl->un.leaf.q);
986 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
987 }
988 cl->prio_activity = 0;
989 cl->cmode = HTB_CAN_SEND;
990 }
991 }
992 qdisc_watchdog_cancel(&q->watchdog);
993 __skb_queue_purge(&q->direct_queue);
994 sch->q.qlen = 0;
995 sch->qstats.backlog = 0;
996 memset(q->hlevel, 0, sizeof(q->hlevel));
997 memset(q->row_mask, 0, sizeof(q->row_mask));
998 for (i = 0; i < TC_HTB_NUMPRIO; i++)
999 INIT_LIST_HEAD(q->drops + i);
1000}
1001
1002static const struct nla_policy htb_policy[TCA_HTB_MAX + 1] = {
1003 [TCA_HTB_PARMS] = { .len = sizeof(struct tc_htb_opt) },
1004 [TCA_HTB_INIT] = { .len = sizeof(struct tc_htb_glob) },
1005 [TCA_HTB_CTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
1006 [TCA_HTB_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
1007 [TCA_HTB_DIRECT_QLEN] = { .type = NLA_U32 },
1008 [TCA_HTB_RATE64] = { .type = NLA_U64 },
1009 [TCA_HTB_CEIL64] = { .type = NLA_U64 },
1010};
1011
1012static void htb_work_func(struct work_struct *work)
1013{
1014 struct htb_sched *q = container_of(work, struct htb_sched, work);
1015 struct Qdisc *sch = q->watchdog.qdisc;
1016
1017 __netif_schedule(qdisc_root(sch));
1018}
1019
1020static int htb_init(struct Qdisc *sch, struct nlattr *opt)
1021{
1022 struct htb_sched *q = qdisc_priv(sch);
1023 struct nlattr *tb[TCA_HTB_MAX + 1];
1024 struct tc_htb_glob *gopt;
1025 int err;
1026 int i;
1027
1028 if (!opt)
1029 return -EINVAL;
1030
1031 err = nla_parse_nested(tb, TCA_HTB_MAX, opt, htb_policy);
1032 if (err < 0)
1033 return err;
1034
1035 if (!tb[TCA_HTB_INIT])
1036 return -EINVAL;
1037
1038 gopt = nla_data(tb[TCA_HTB_INIT]);
1039 if (gopt->version != HTB_VER >> 16)
1040 return -EINVAL;
1041
1042 err = qdisc_class_hash_init(&q->clhash);
1043 if (err < 0)
1044 return err;
1045 for (i = 0; i < TC_HTB_NUMPRIO; i++)
1046 INIT_LIST_HEAD(q->drops + i);
1047
1048 qdisc_watchdog_init(&q->watchdog, sch);
1049 INIT_WORK(&q->work, htb_work_func);
1050 __skb_queue_head_init(&q->direct_queue);
1051
1052 if (tb[TCA_HTB_DIRECT_QLEN])
1053 q->direct_qlen = nla_get_u32(tb[TCA_HTB_DIRECT_QLEN]);
1054 else
1055 q->direct_qlen = qdisc_dev(sch)->tx_queue_len;
1056
1057 if ((q->rate2quantum = gopt->rate2quantum) < 1)
1058 q->rate2quantum = 1;
1059 q->defcls = gopt->defcls;
1060
1061 return 0;
1062}
1063
1064static int htb_dump(struct Qdisc *sch, struct sk_buff *skb)
1065{
1066 struct htb_sched *q = qdisc_priv(sch);
1067 struct nlattr *nest;
1068 struct tc_htb_glob gopt;
1069
1070 /* Its safe to not acquire qdisc lock. As we hold RTNL,
1071 * no change can happen on the qdisc parameters.
1072 */
1073
1074 gopt.direct_pkts = q->direct_pkts;
1075 gopt.version = HTB_VER;
1076 gopt.rate2quantum = q->rate2quantum;
1077 gopt.defcls = q->defcls;
1078 gopt.debug = 0;
1079
1080 nest = nla_nest_start(skb, TCA_OPTIONS);
1081 if (nest == NULL)
1082 goto nla_put_failure;
1083 if (nla_put(skb, TCA_HTB_INIT, sizeof(gopt), &gopt) ||
1084 nla_put_u32(skb, TCA_HTB_DIRECT_QLEN, q->direct_qlen))
1085 goto nla_put_failure;
1086
1087 return nla_nest_end(skb, nest);
1088
1089nla_put_failure:
1090 nla_nest_cancel(skb, nest);
1091 return -1;
1092}
1093
1094static int htb_dump_class(struct Qdisc *sch, unsigned long arg,
1095 struct sk_buff *skb, struct tcmsg *tcm)
1096{
1097 struct htb_class *cl = (struct htb_class *)arg;
1098 struct nlattr *nest;
1099 struct tc_htb_opt opt;
1100
1101 /* Its safe to not acquire qdisc lock. As we hold RTNL,
1102 * no change can happen on the class parameters.
1103 */
1104 tcm->tcm_parent = cl->parent ? cl->parent->common.classid : TC_H_ROOT;
1105 tcm->tcm_handle = cl->common.classid;
1106 if (!cl->level && cl->un.leaf.q)
1107 tcm->tcm_info = cl->un.leaf.q->handle;
1108
1109 nest = nla_nest_start(skb, TCA_OPTIONS);
1110 if (nest == NULL)
1111 goto nla_put_failure;
1112
1113 memset(&opt, 0, sizeof(opt));
1114
1115 psched_ratecfg_getrate(&opt.rate, &cl->rate);
1116 opt.buffer = PSCHED_NS2TICKS(cl->buffer);
1117 psched_ratecfg_getrate(&opt.ceil, &cl->ceil);
1118 opt.cbuffer = PSCHED_NS2TICKS(cl->cbuffer);
1119 opt.quantum = cl->quantum;
1120 opt.prio = cl->prio;
1121 opt.level = cl->level;
1122 if (nla_put(skb, TCA_HTB_PARMS, sizeof(opt), &opt))
1123 goto nla_put_failure;
1124 if ((cl->rate.rate_bytes_ps >= (1ULL << 32)) &&
1125 nla_put_u64(skb, TCA_HTB_RATE64, cl->rate.rate_bytes_ps))
1126 goto nla_put_failure;
1127 if ((cl->ceil.rate_bytes_ps >= (1ULL << 32)) &&
1128 nla_put_u64(skb, TCA_HTB_CEIL64, cl->ceil.rate_bytes_ps))
1129 goto nla_put_failure;
1130
1131 return nla_nest_end(skb, nest);
1132
1133nla_put_failure:
1134 nla_nest_cancel(skb, nest);
1135 return -1;
1136}
1137
1138static int
1139htb_dump_class_stats(struct Qdisc *sch, unsigned long arg, struct gnet_dump *d)
1140{
1141 struct htb_class *cl = (struct htb_class *)arg;
1142 __u32 qlen = 0;
1143
1144 if (!cl->level && cl->un.leaf.q)
1145 qlen = cl->un.leaf.q->q.qlen;
1146 cl->xstats.tokens = PSCHED_NS2TICKS(cl->tokens);
1147 cl->xstats.ctokens = PSCHED_NS2TICKS(cl->ctokens);
1148
1149 if (gnet_stats_copy_basic(d, NULL, &cl->bstats) < 0 ||
1150 gnet_stats_copy_rate_est(d, NULL, &cl->rate_est) < 0 ||
1151 gnet_stats_copy_queue(d, NULL, &cl->qstats, qlen) < 0)
1152 return -1;
1153
1154 return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1155}
1156
1157static int htb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1158 struct Qdisc **old)
1159{
1160 struct htb_class *cl = (struct htb_class *)arg;
1161
1162 if (cl->level)
1163 return -EINVAL;
1164 if (new == NULL &&
1165 (new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1166 cl->common.classid)) == NULL)
1167 return -ENOBUFS;
1168
1169 *old = qdisc_replace(sch, new, &cl->un.leaf.q);
1170 return 0;
1171}
1172
1173static struct Qdisc *htb_leaf(struct Qdisc *sch, unsigned long arg)
1174{
1175 struct htb_class *cl = (struct htb_class *)arg;
1176 return !cl->level ? cl->un.leaf.q : NULL;
1177}
1178
1179static void htb_qlen_notify(struct Qdisc *sch, unsigned long arg)
1180{
1181 struct htb_class *cl = (struct htb_class *)arg;
1182
1183 if (cl->un.leaf.q->q.qlen == 0)
1184 htb_deactivate(qdisc_priv(sch), cl);
1185}
1186
1187static unsigned long htb_get(struct Qdisc *sch, u32 classid)
1188{
1189 struct htb_class *cl = htb_find(classid, sch);
1190 if (cl)
1191 cl->refcnt++;
1192 return (unsigned long)cl;
1193}
1194
1195static inline int htb_parent_last_child(struct htb_class *cl)
1196{
1197 if (!cl->parent)
1198 /* the root class */
1199 return 0;
1200 if (cl->parent->children > 1)
1201 /* not the last child */
1202 return 0;
1203 return 1;
1204}
1205
1206static void htb_parent_to_leaf(struct htb_sched *q, struct htb_class *cl,
1207 struct Qdisc *new_q)
1208{
1209 struct htb_class *parent = cl->parent;
1210
1211 WARN_ON(cl->level || !cl->un.leaf.q || cl->prio_activity);
1212
1213 if (parent->cmode != HTB_CAN_SEND)
1214 htb_safe_rb_erase(&parent->pq_node,
1215 &q->hlevel[parent->level].wait_pq);
1216
1217 parent->level = 0;
1218 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1219 INIT_LIST_HEAD(&parent->un.leaf.drop_list);
1220 parent->un.leaf.q = new_q ? new_q : &noop_qdisc;
1221 parent->tokens = parent->buffer;
1222 parent->ctokens = parent->cbuffer;
1223 parent->t_c = ktime_get_ns();
1224 parent->cmode = HTB_CAN_SEND;
1225}
1226
1227static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
1228{
1229 if (!cl->level) {
1230 WARN_ON(!cl->un.leaf.q);
1231 qdisc_destroy(cl->un.leaf.q);
1232 }
1233 gen_kill_estimator(&cl->bstats, &cl->rate_est);
1234 tcf_destroy_chain(&cl->filter_list);
1235 kfree(cl);
1236}
1237
1238static void htb_destroy(struct Qdisc *sch)
1239{
1240 struct htb_sched *q = qdisc_priv(sch);
1241 struct hlist_node *next;
1242 struct htb_class *cl;
1243 unsigned int i;
1244
1245 cancel_work_sync(&q->work);
1246 qdisc_watchdog_cancel(&q->watchdog);
1247 /* This line used to be after htb_destroy_class call below
1248 * and surprisingly it worked in 2.4. But it must precede it
1249 * because filter need its target class alive to be able to call
1250 * unbind_filter on it (without Oops).
1251 */
1252 tcf_destroy_chain(&q->filter_list);
1253
1254 for (i = 0; i < q->clhash.hashsize; i++) {
1255 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode)
1256 tcf_destroy_chain(&cl->filter_list);
1257 }
1258 for (i = 0; i < q->clhash.hashsize; i++) {
1259 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1260 common.hnode)
1261 htb_destroy_class(sch, cl);
1262 }
1263 qdisc_class_hash_destroy(&q->clhash);
1264 __skb_queue_purge(&q->direct_queue);
1265}
1266
1267static int htb_delete(struct Qdisc *sch, unsigned long arg)
1268{
1269 struct htb_sched *q = qdisc_priv(sch);
1270 struct htb_class *cl = (struct htb_class *)arg;
1271 struct Qdisc *new_q = NULL;
1272 int last_child = 0;
1273
1274 /* TODO: why don't allow to delete subtree ? references ? does
1275 * tc subsys guarantee us that in htb_destroy it holds no class
1276 * refs so that we can remove children safely there ?
1277 */
1278 if (cl->children || cl->filter_cnt)
1279 return -EBUSY;
1280
1281 if (!cl->level && htb_parent_last_child(cl)) {
1282 new_q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1283 cl->parent->common.classid);
1284 last_child = 1;
1285 }
1286
1287 sch_tree_lock(sch);
1288
1289 if (!cl->level) {
1290 unsigned int qlen = cl->un.leaf.q->q.qlen;
1291 unsigned int backlog = cl->un.leaf.q->qstats.backlog;
1292
1293 qdisc_reset(cl->un.leaf.q);
1294 qdisc_tree_reduce_backlog(cl->un.leaf.q, qlen, backlog);
1295 }
1296
1297 /* delete from hash and active; remainder in destroy_class */
1298 qdisc_class_hash_remove(&q->clhash, &cl->common);
1299 if (cl->parent)
1300 cl->parent->children--;
1301
1302 if (cl->prio_activity)
1303 htb_deactivate(q, cl);
1304
1305 if (cl->cmode != HTB_CAN_SEND)
1306 htb_safe_rb_erase(&cl->pq_node,
1307 &q->hlevel[cl->level].wait_pq);
1308
1309 if (last_child)
1310 htb_parent_to_leaf(q, cl, new_q);
1311
1312 BUG_ON(--cl->refcnt == 0);
1313 /*
1314 * This shouldn't happen: we "hold" one cops->get() when called
1315 * from tc_ctl_tclass; the destroy method is done from cops->put().
1316 */
1317
1318 sch_tree_unlock(sch);
1319 return 0;
1320}
1321
1322static void htb_put(struct Qdisc *sch, unsigned long arg)
1323{
1324 struct htb_class *cl = (struct htb_class *)arg;
1325
1326 if (--cl->refcnt == 0)
1327 htb_destroy_class(sch, cl);
1328}
1329
1330static int htb_change_class(struct Qdisc *sch, u32 classid,
1331 u32 parentid, struct nlattr **tca,
1332 unsigned long *arg)
1333{
1334 int err = -EINVAL;
1335 struct htb_sched *q = qdisc_priv(sch);
1336 struct htb_class *cl = (struct htb_class *)*arg, *parent;
1337 struct nlattr *opt = tca[TCA_OPTIONS];
1338 struct nlattr *tb[TCA_HTB_MAX + 1];
1339 struct tc_htb_opt *hopt;
1340 u64 rate64, ceil64;
1341
1342 /* extract all subattrs from opt attr */
1343 if (!opt)
1344 goto failure;
1345
1346 err = nla_parse_nested(tb, TCA_HTB_MAX, opt, htb_policy);
1347 if (err < 0)
1348 goto failure;
1349
1350 err = -EINVAL;
1351 if (tb[TCA_HTB_PARMS] == NULL)
1352 goto failure;
1353
1354 parent = parentid == TC_H_ROOT ? NULL : htb_find(parentid, sch);
1355
1356 hopt = nla_data(tb[TCA_HTB_PARMS]);
1357 if (!hopt->rate.rate || !hopt->ceil.rate)
1358 goto failure;
1359
1360 /* Keeping backward compatible with rate_table based iproute2 tc */
1361 if (hopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
1362 qdisc_put_rtab(qdisc_get_rtab(&hopt->rate, tb[TCA_HTB_RTAB]));
1363
1364 if (hopt->ceil.linklayer == TC_LINKLAYER_UNAWARE)
1365 qdisc_put_rtab(qdisc_get_rtab(&hopt->ceil, tb[TCA_HTB_CTAB]));
1366
1367 if (!cl) { /* new class */
1368 struct Qdisc *new_q;
1369 int prio;
1370 struct {
1371 struct nlattr nla;
1372 struct gnet_estimator opt;
1373 } est = {
1374 .nla = {
1375 .nla_len = nla_attr_size(sizeof(est.opt)),
1376 .nla_type = TCA_RATE,
1377 },
1378 .opt = {
1379 /* 4s interval, 16s averaging constant */
1380 .interval = 2,
1381 .ewma_log = 2,
1382 },
1383 };
1384
1385 /* check for valid classid */
1386 if (!classid || TC_H_MAJ(classid ^ sch->handle) ||
1387 htb_find(classid, sch))
1388 goto failure;
1389
1390 /* check maximal depth */
1391 if (parent && parent->parent && parent->parent->level < 2) {
1392 pr_err("htb: tree is too deep\n");
1393 goto failure;
1394 }
1395 err = -ENOBUFS;
1396 cl = kzalloc(sizeof(*cl), GFP_KERNEL);
1397 if (!cl)
1398 goto failure;
1399
1400 if (htb_rate_est || tca[TCA_RATE]) {
1401 err = gen_new_estimator(&cl->bstats, NULL,
1402 &cl->rate_est,
1403 qdisc_root_sleeping_lock(sch),
1404 tca[TCA_RATE] ? : &est.nla);
1405 if (err) {
1406 kfree(cl);
1407 goto failure;
1408 }
1409 }
1410
1411 cl->refcnt = 1;
1412 cl->children = 0;
1413 INIT_LIST_HEAD(&cl->un.leaf.drop_list);
1414 RB_CLEAR_NODE(&cl->pq_node);
1415
1416 for (prio = 0; prio < TC_HTB_NUMPRIO; prio++)
1417 RB_CLEAR_NODE(&cl->node[prio]);
1418
1419 /* create leaf qdisc early because it uses kmalloc(GFP_KERNEL)
1420 * so that can't be used inside of sch_tree_lock
1421 * -- thanks to Karlis Peisenieks
1422 */
1423 new_q = qdisc_create_dflt(sch->dev_queue,
1424 &pfifo_qdisc_ops, classid);
1425 sch_tree_lock(sch);
1426 if (parent && !parent->level) {
1427 unsigned int qlen = parent->un.leaf.q->q.qlen;
1428 unsigned int backlog = parent->un.leaf.q->qstats.backlog;
1429
1430 /* turn parent into inner node */
1431 qdisc_reset(parent->un.leaf.q);
1432 qdisc_tree_reduce_backlog(parent->un.leaf.q, qlen, backlog);
1433 qdisc_destroy(parent->un.leaf.q);
1434 if (parent->prio_activity)
1435 htb_deactivate(q, parent);
1436
1437 /* remove from evt list because of level change */
1438 if (parent->cmode != HTB_CAN_SEND) {
1439 htb_safe_rb_erase(&parent->pq_node, &q->hlevel[0].wait_pq);
1440 parent->cmode = HTB_CAN_SEND;
1441 }
1442 parent->level = (parent->parent ? parent->parent->level
1443 : TC_HTB_MAXDEPTH) - 1;
1444 memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1445 }
1446 /* leaf (we) needs elementary qdisc */
1447 cl->un.leaf.q = new_q ? new_q : &noop_qdisc;
1448
1449 cl->common.classid = classid;
1450 cl->parent = parent;
1451
1452 /* set class to be in HTB_CAN_SEND state */
1453 cl->tokens = PSCHED_TICKS2NS(hopt->buffer);
1454 cl->ctokens = PSCHED_TICKS2NS(hopt->cbuffer);
1455 cl->mbuffer = 60ULL * NSEC_PER_SEC; /* 1min */
1456 cl->t_c = ktime_get_ns();
1457 cl->cmode = HTB_CAN_SEND;
1458
1459 /* attach to the hash list and parent's family */
1460 qdisc_class_hash_insert(&q->clhash, &cl->common);
1461 if (parent)
1462 parent->children++;
1463 } else {
1464 if (tca[TCA_RATE]) {
1465 spinlock_t *lock = qdisc_root_sleeping_lock(sch);
1466
1467 err = gen_replace_estimator(&cl->bstats, NULL,
1468 &cl->rate_est,
1469 lock,
1470 tca[TCA_RATE]);
1471 if (err)
1472 return err;
1473 }
1474 sch_tree_lock(sch);
1475 }
1476
1477 rate64 = tb[TCA_HTB_RATE64] ? nla_get_u64(tb[TCA_HTB_RATE64]) : 0;
1478
1479 ceil64 = tb[TCA_HTB_CEIL64] ? nla_get_u64(tb[TCA_HTB_CEIL64]) : 0;
1480
1481 psched_ratecfg_precompute(&cl->rate, &hopt->rate, rate64);
1482 psched_ratecfg_precompute(&cl->ceil, &hopt->ceil, ceil64);
1483
1484 /* it used to be a nasty bug here, we have to check that node
1485 * is really leaf before changing cl->un.leaf !
1486 */
1487 if (!cl->level) {
1488 u64 quantum = cl->rate.rate_bytes_ps;
1489
1490 do_div(quantum, q->rate2quantum);
1491 cl->quantum = min_t(u64, quantum, INT_MAX);
1492
1493 if (!hopt->quantum && cl->quantum < 1000) {
1494 pr_warn("HTB: quantum of class %X is small. Consider r2q change.\n",
1495 cl->common.classid);
1496 cl->quantum = 1000;
1497 }
1498 if (!hopt->quantum && cl->quantum > 200000) {
1499 pr_warn("HTB: quantum of class %X is big. Consider r2q change.\n",
1500 cl->common.classid);
1501 cl->quantum = 200000;
1502 }
1503 if (hopt->quantum)
1504 cl->quantum = hopt->quantum;
1505 if ((cl->prio = hopt->prio) >= TC_HTB_NUMPRIO)
1506 cl->prio = TC_HTB_NUMPRIO - 1;
1507 }
1508
1509 cl->buffer = PSCHED_TICKS2NS(hopt->buffer);
1510 cl->cbuffer = PSCHED_TICKS2NS(hopt->cbuffer);
1511
1512 sch_tree_unlock(sch);
1513
1514 qdisc_class_hash_grow(sch, &q->clhash);
1515
1516 *arg = (unsigned long)cl;
1517 return 0;
1518
1519failure:
1520 return err;
1521}
1522
1523static struct tcf_proto __rcu **htb_find_tcf(struct Qdisc *sch,
1524 unsigned long arg)
1525{
1526 struct htb_sched *q = qdisc_priv(sch);
1527 struct htb_class *cl = (struct htb_class *)arg;
1528 struct tcf_proto __rcu **fl = cl ? &cl->filter_list : &q->filter_list;
1529
1530 return fl;
1531}
1532
1533static unsigned long htb_bind_filter(struct Qdisc *sch, unsigned long parent,
1534 u32 classid)
1535{
1536 struct htb_class *cl = htb_find(classid, sch);
1537
1538 /*if (cl && !cl->level) return 0;
1539 * The line above used to be there to prevent attaching filters to
1540 * leaves. But at least tc_index filter uses this just to get class
1541 * for other reasons so that we have to allow for it.
1542 * ----
1543 * 19.6.2002 As Werner explained it is ok - bind filter is just
1544 * another way to "lock" the class - unlike "get" this lock can
1545 * be broken by class during destroy IIUC.
1546 */
1547 if (cl)
1548 cl->filter_cnt++;
1549 return (unsigned long)cl;
1550}
1551
1552static void htb_unbind_filter(struct Qdisc *sch, unsigned long arg)
1553{
1554 struct htb_class *cl = (struct htb_class *)arg;
1555
1556 if (cl)
1557 cl->filter_cnt--;
1558}
1559
1560static void htb_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1561{
1562 struct htb_sched *q = qdisc_priv(sch);
1563 struct htb_class *cl;
1564 unsigned int i;
1565
1566 if (arg->stop)
1567 return;
1568
1569 for (i = 0; i < q->clhash.hashsize; i++) {
1570 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1571 if (arg->count < arg->skip) {
1572 arg->count++;
1573 continue;
1574 }
1575 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1576 arg->stop = 1;
1577 return;
1578 }
1579 arg->count++;
1580 }
1581 }
1582}
1583
1584static const struct Qdisc_class_ops htb_class_ops = {
1585 .graft = htb_graft,
1586 .leaf = htb_leaf,
1587 .qlen_notify = htb_qlen_notify,
1588 .get = htb_get,
1589 .put = htb_put,
1590 .change = htb_change_class,
1591 .delete = htb_delete,
1592 .walk = htb_walk,
1593 .tcf_chain = htb_find_tcf,
1594 .bind_tcf = htb_bind_filter,
1595 .unbind_tcf = htb_unbind_filter,
1596 .dump = htb_dump_class,
1597 .dump_stats = htb_dump_class_stats,
1598};
1599
1600static struct Qdisc_ops htb_qdisc_ops __read_mostly = {
1601 .cl_ops = &htb_class_ops,
1602 .id = "htb",
1603 .priv_size = sizeof(struct htb_sched),
1604 .enqueue = htb_enqueue,
1605 .dequeue = htb_dequeue,
1606 .peek = qdisc_peek_dequeued,
1607 .drop = htb_drop,
1608 .init = htb_init,
1609 .reset = htb_reset,
1610 .destroy = htb_destroy,
1611 .dump = htb_dump,
1612 .owner = THIS_MODULE,
1613};
1614
1615static int __init htb_module_init(void)
1616{
1617 return register_qdisc(&htb_qdisc_ops);
1618}
1619static void __exit htb_module_exit(void)
1620{
1621 unregister_qdisc(&htb_qdisc_ops);
1622}
1623
1624module_init(htb_module_init)
1625module_exit(htb_module_exit)
1626MODULE_LICENSE("GPL");