Commit | Line | Data |
---|---|---|
2874c5fd | 1 | // SPDX-License-Identifier: GPL-2.0-or-later |
87990467 | 2 | /* |
1da177e4 LT |
3 | * net/sched/sch_htb.c Hierarchical token bucket, feed tree version |
4 | * | |
1da177e4 LT |
5 | * Authors: Martin Devera, <devik@cdi.cz> |
6 | * | |
7 | * Credits (in time order) for older HTB versions: | |
8 | * Stef Coene <stef.coene@docum.org> | |
9 | * HTB support at LARTC mailing list | |
10297b99 | 10 | * Ondrej Kraus, <krauso@barr.cz> |
1da177e4 LT |
11 | * found missing INIT_QDISC(htb) |
12 | * Vladimir Smelhaus, Aamer Akhter, Bert Hubert | |
13 | * helped a lot to locate nasty class stall bug | |
14 | * Andi Kleen, Jamal Hadi, Bert Hubert | |
15 | * code review and helpful comments on shaping | |
16 | * Tomasz Wrona, <tw@eter.tym.pl> | |
17 | * created test case so that I was able to fix nasty bug | |
18 | * Wilfried Weissmann | |
19 | * spotted bug in dequeue code and helped with fix | |
20 | * Jiri Fojtasek | |
21 | * fixed requeue routine | |
22 | * and many others. thanks. | |
1da177e4 | 23 | */ |
1da177e4 | 24 | #include <linux/module.h> |
47083fc0 | 25 | #include <linux/moduleparam.h> |
1da177e4 LT |
26 | #include <linux/types.h> |
27 | #include <linux/kernel.h> | |
1da177e4 | 28 | #include <linux/string.h> |
1da177e4 | 29 | #include <linux/errno.h> |
1da177e4 LT |
30 | #include <linux/skbuff.h> |
31 | #include <linux/list.h> | |
32 | #include <linux/compiler.h> | |
0ba48053 | 33 | #include <linux/rbtree.h> |
1224736d | 34 | #include <linux/workqueue.h> |
5a0e3ad6 | 35 | #include <linux/slab.h> |
dc5fc579 | 36 | #include <net/netlink.h> |
292f1c7f | 37 | #include <net/sch_generic.h> |
1da177e4 | 38 | #include <net/pkt_sched.h> |
cf1facda | 39 | #include <net/pkt_cls.h> |
1da177e4 LT |
40 | |
41 | /* HTB algorithm. | |
42 | Author: devik@cdi.cz | |
43 | ======================================================================== | |
44 | HTB is like TBF with multiple classes. It is also similar to CBQ because | |
10297b99 | 45 | it allows to assign priority to each class in hierarchy. |
1da177e4 LT |
46 | In fact it is another implementation of Floyd's formal sharing. |
47 | ||
48 | Levels: | |
10297b99 | 49 | Each class is assigned level. Leaf has ALWAYS level 0 and root |
1da177e4 LT |
50 | classes have level TC_HTB_MAXDEPTH-1. Interior nodes has level |
51 | one less than their parent. | |
52 | */ | |
53 | ||
47083fc0 | 54 | static int htb_hysteresis __read_mostly = 0; /* whether to use mode hysteresis for speedup */ |
37f2ad2b | 55 | #define HTB_VER 0x30011 /* major must be matched with number supplied by TC as version */ |
1da177e4 LT |
56 | |
57 | #if HTB_VER >> 16 != TC_HTB_PROTOVER | |
58 | #error "Mismatched sch_htb.c and pkt_sch.h" | |
59 | #endif | |
60 | ||
47083fc0 JDB |
61 | /* Module parameter and sysfs export */ |
62 | module_param (htb_hysteresis, int, 0640); | |
63 | MODULE_PARM_DESC(htb_hysteresis, "Hysteresis mode, less CPU load, less accurate"); | |
64 | ||
64153ce0 ED |
65 | static int htb_rate_est = 0; /* htb classes have a default rate estimator */ |
66 | module_param(htb_rate_est, int, 0640); | |
67 | MODULE_PARM_DESC(htb_rate_est, "setup a default rate estimator (4sec 16sec) for htb classes"); | |
68 | ||
1da177e4 LT |
69 | /* used internaly to keep status of single class */ |
70 | enum htb_cmode { | |
87990467 SH |
71 | HTB_CANT_SEND, /* class can't send and can't borrow */ |
72 | HTB_MAY_BORROW, /* class can't send but may borrow */ | |
73 | HTB_CAN_SEND /* class can send */ | |
1da177e4 LT |
74 | }; |
75 | ||
c9364636 ED |
76 | struct htb_prio { |
77 | union { | |
78 | struct rb_root row; | |
79 | struct rb_root feed; | |
80 | }; | |
81 | struct rb_node *ptr; | |
82 | /* When class changes from state 1->2 and disconnects from | |
83 | * parent's feed then we lost ptr value and start from the | |
84 | * first child again. Here we store classid of the | |
85 | * last valid ptr (used when ptr is NULL). | |
86 | */ | |
87 | u32 last_ptr_id; | |
88 | }; | |
89 | ||
ca4ec90b ED |
90 | /* interior & leaf nodes; props specific to leaves are marked L: |
91 | * To reduce false sharing, place mostly read fields at beginning, | |
92 | * and mostly written ones at the end. | |
93 | */ | |
87990467 | 94 | struct htb_class { |
f4c1f3e0 | 95 | struct Qdisc_class_common common; |
ca4ec90b ED |
96 | struct psched_ratecfg rate; |
97 | struct psched_ratecfg ceil; | |
98 | s64 buffer, cbuffer;/* token bucket depth/rate */ | |
99 | s64 mbuffer; /* max wait time */ | |
cbd37556 | 100 | u32 prio; /* these two are used only by leaves... */ |
ca4ec90b ED |
101 | int quantum; /* but stored for parent-to-leaf return */ |
102 | ||
25d8c0d5 | 103 | struct tcf_proto __rcu *filter_list; /* class attached filters */ |
6529eaba | 104 | struct tcf_block *block; |
ca4ec90b | 105 | int filter_cnt; |
ca4ec90b ED |
106 | |
107 | int level; /* our level (see above) */ | |
108 | unsigned int children; | |
109 | struct htb_class *parent; /* parent class */ | |
110 | ||
1c0d32fd | 111 | struct net_rate_estimator __rcu *rate_est; |
1da177e4 | 112 | |
ca4ec90b ED |
113 | /* |
114 | * Written often fields | |
115 | */ | |
50dc9a85 AD |
116 | struct gnet_stats_basic_sync bstats; |
117 | struct gnet_stats_basic_sync bstats_bias; | |
ca4ec90b | 118 | struct tc_htb_xstats xstats; /* our special stats */ |
87990467 | 119 | |
ca4ec90b ED |
120 | /* token bucket parameters */ |
121 | s64 tokens, ctokens;/* current number of tokens */ | |
122 | s64 t_c; /* checkpoint time */ | |
c19f7a34 | 123 | |
87990467 SH |
124 | union { |
125 | struct htb_class_leaf { | |
c9364636 ED |
126 | int deficit[TC_HTB_MAXDEPTH]; |
127 | struct Qdisc *q; | |
ca49bfd9 | 128 | struct netdev_queue *offload_queue; |
87990467 SH |
129 | } leaf; |
130 | struct htb_class_inner { | |
c9364636 | 131 | struct htb_prio clprio[TC_HTB_NUMPRIO]; |
87990467 | 132 | } inner; |
11957be2 | 133 | }; |
ca4ec90b | 134 | s64 pq_key; |
87990467 | 135 | |
ca4ec90b ED |
136 | int prio_activity; /* for which prios are we active */ |
137 | enum htb_cmode cmode; /* current mode of the class */ | |
138 | struct rb_node pq_node; /* node for event queue */ | |
139 | struct rb_node node[TC_HTB_NUMPRIO]; /* node for self or feed tree */ | |
338ed9b4 ED |
140 | |
141 | unsigned int drops ____cacheline_aligned_in_smp; | |
3c75f6ee | 142 | unsigned int overlimits; |
1da177e4 LT |
143 | }; |
144 | ||
c9364636 ED |
145 | struct htb_level { |
146 | struct rb_root wait_pq; | |
147 | struct htb_prio hprio[TC_HTB_NUMPRIO]; | |
148 | }; | |
149 | ||
87990467 | 150 | struct htb_sched { |
f4c1f3e0 | 151 | struct Qdisc_class_hash clhash; |
c9364636 ED |
152 | int defcls; /* class where unclassified flows go to */ |
153 | int rate2quantum; /* quant = rate / rate2quantum */ | |
1da177e4 | 154 | |
c9364636 | 155 | /* filters for qdisc itself */ |
25d8c0d5 | 156 | struct tcf_proto __rcu *filter_list; |
6529eaba | 157 | struct tcf_block *block; |
1da177e4 | 158 | |
c9364636 ED |
159 | #define HTB_WARN_TOOMANYEVENTS 0x1 |
160 | unsigned int warned; /* only one warning */ | |
161 | int direct_qlen; | |
162 | struct work_struct work; | |
1da177e4 | 163 | |
c9364636 | 164 | /* non shaped skbs; let them go directly thru */ |
48da34b7 | 165 | struct qdisc_skb_head direct_queue; |
b362487a CW |
166 | u32 direct_pkts; |
167 | u32 overlimits; | |
1da177e4 | 168 | |
c9364636 | 169 | struct qdisc_watchdog watchdog; |
1da177e4 | 170 | |
c9364636 | 171 | s64 now; /* cached dequeue time */ |
1da177e4 | 172 | |
c9364636 ED |
173 | /* time of nearest event per level (row) */ |
174 | s64 near_ev_cache[TC_HTB_MAXDEPTH]; | |
87990467 | 175 | |
c9364636 | 176 | int row_mask[TC_HTB_MAXDEPTH]; |
e82181de | 177 | |
c9364636 | 178 | struct htb_level hlevel[TC_HTB_MAXDEPTH]; |
d03b195b MM |
179 | |
180 | struct Qdisc **direct_qdiscs; | |
181 | unsigned int num_direct_qdiscs; | |
182 | ||
183 | bool offload; | |
1da177e4 LT |
184 | }; |
185 | ||
1da177e4 | 186 | /* find class in global hash table using given handle */ |
87990467 | 187 | static inline struct htb_class *htb_find(u32 handle, struct Qdisc *sch) |
1da177e4 LT |
188 | { |
189 | struct htb_sched *q = qdisc_priv(sch); | |
f4c1f3e0 | 190 | struct Qdisc_class_common *clc; |
0cef296d | 191 | |
f4c1f3e0 PM |
192 | clc = qdisc_class_find(&q->clhash, handle); |
193 | if (clc == NULL) | |
1da177e4 | 194 | return NULL; |
f4c1f3e0 | 195 | return container_of(clc, struct htb_class, common); |
1da177e4 LT |
196 | } |
197 | ||
143976ce WC |
198 | static unsigned long htb_search(struct Qdisc *sch, u32 handle) |
199 | { | |
200 | return (unsigned long)htb_find(handle, sch); | |
201 | } | |
43d25378 RD |
202 | |
203 | #define HTB_DIRECT ((struct htb_class *)-1L) | |
204 | ||
1da177e4 LT |
205 | /** |
206 | * htb_classify - classify a packet into class | |
43d25378 RD |
207 | * @skb: the socket buffer |
208 | * @sch: the active queue discipline | |
209 | * @qerr: pointer for returned status code | |
1da177e4 LT |
210 | * |
211 | * It returns NULL if the packet should be dropped or -1 if the packet | |
212 | * should be passed directly thru. In all other cases leaf class is returned. | |
213 | * We allow direct class selection by classid in priority. The we examine | |
214 | * filters in qdisc and in inner nodes (if higher filter points to the inner | |
215 | * node). If we end up with classid MAJOR:0 we enqueue the skb into special | |
10297b99 | 216 | * internal fifo (direct). These packets then go directly thru. If we still |
25985edc | 217 | * have no valid leaf we try to use MAJOR:default leaf. It still unsuccessful |
1da177e4 LT |
218 | * then finish and return direct queue. |
219 | */ | |
87990467 SH |
220 | static struct htb_class *htb_classify(struct sk_buff *skb, struct Qdisc *sch, |
221 | int *qerr) | |
1da177e4 LT |
222 | { |
223 | struct htb_sched *q = qdisc_priv(sch); | |
224 | struct htb_class *cl; | |
225 | struct tcf_result res; | |
226 | struct tcf_proto *tcf; | |
227 | int result; | |
228 | ||
229 | /* allow to select class by setting skb->priority to valid classid; | |
cc7ec456 ED |
230 | * note that nfmark can be used too by attaching filter fw with no |
231 | * rules in it | |
232 | */ | |
1da177e4 | 233 | if (skb->priority == sch->handle) |
87990467 | 234 | return HTB_DIRECT; /* X:0 (direct flow) selected */ |
cc7ec456 | 235 | cl = htb_find(skb->priority, sch); |
29824310 HM |
236 | if (cl) { |
237 | if (cl->level == 0) | |
238 | return cl; | |
239 | /* Start with inner filter chain if a non-leaf class is selected */ | |
25d8c0d5 | 240 | tcf = rcu_dereference_bh(cl->filter_list); |
29824310 | 241 | } else { |
25d8c0d5 | 242 | tcf = rcu_dereference_bh(q->filter_list); |
29824310 | 243 | } |
1da177e4 | 244 | |
c27f339a | 245 | *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS; |
3aa26055 | 246 | while (tcf && (result = tcf_classify(skb, NULL, tcf, &res, false)) >= 0) { |
1da177e4 LT |
247 | #ifdef CONFIG_NET_CLS_ACT |
248 | switch (result) { | |
249 | case TC_ACT_QUEUED: | |
87990467 | 250 | case TC_ACT_STOLEN: |
e25ea21f | 251 | case TC_ACT_TRAP: |
378a2f09 | 252 | *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN; |
964201de | 253 | fallthrough; |
1da177e4 LT |
254 | case TC_ACT_SHOT: |
255 | return NULL; | |
256 | } | |
1da177e4 | 257 | #endif |
cc7ec456 ED |
258 | cl = (void *)res.class; |
259 | if (!cl) { | |
1da177e4 | 260 | if (res.classid == sch->handle) |
87990467 | 261 | return HTB_DIRECT; /* X:0 (direct flow) */ |
cc7ec456 ED |
262 | cl = htb_find(res.classid, sch); |
263 | if (!cl) | |
87990467 | 264 | break; /* filter selected invalid classid */ |
1da177e4 LT |
265 | } |
266 | if (!cl->level) | |
87990467 | 267 | return cl; /* we hit leaf; return it */ |
1da177e4 LT |
268 | |
269 | /* we have got inner class; apply inner filter chain */ | |
25d8c0d5 | 270 | tcf = rcu_dereference_bh(cl->filter_list); |
1da177e4 LT |
271 | } |
272 | /* classification failed; try to use default class */ | |
87990467 | 273 | cl = htb_find(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch); |
1da177e4 | 274 | if (!cl || cl->level) |
87990467 | 275 | return HTB_DIRECT; /* bad default .. this is safe bet */ |
1da177e4 LT |
276 | return cl; |
277 | } | |
278 | ||
1da177e4 LT |
279 | /** |
280 | * htb_add_to_id_tree - adds class to the round robin list | |
a10541f5 YK |
281 | * @root: the root of the tree |
282 | * @cl: the class to add | |
283 | * @prio: the give prio in class | |
1da177e4 LT |
284 | * |
285 | * Routine adds class to the list (actually tree) sorted by classid. | |
286 | * Make sure that class is not already on such list for given prio. | |
287 | */ | |
87990467 SH |
288 | static void htb_add_to_id_tree(struct rb_root *root, |
289 | struct htb_class *cl, int prio) | |
1da177e4 LT |
290 | { |
291 | struct rb_node **p = &root->rb_node, *parent = NULL; | |
3bf72957 | 292 | |
1da177e4 | 293 | while (*p) { |
87990467 SH |
294 | struct htb_class *c; |
295 | parent = *p; | |
1da177e4 | 296 | c = rb_entry(parent, struct htb_class, node[prio]); |
3bf72957 | 297 | |
f4c1f3e0 | 298 | if (cl->common.classid > c->common.classid) |
1da177e4 | 299 | p = &parent->rb_right; |
87990467 | 300 | else |
1da177e4 LT |
301 | p = &parent->rb_left; |
302 | } | |
303 | rb_link_node(&cl->node[prio], parent, p); | |
304 | rb_insert_color(&cl->node[prio], root); | |
305 | } | |
306 | ||
307 | /** | |
308 | * htb_add_to_wait_tree - adds class to the event queue with delay | |
4d7efa73 YK |
309 | * @q: the priority event queue |
310 | * @cl: the class to add | |
311 | * @delay: delay in microseconds | |
1da177e4 LT |
312 | * |
313 | * The class is added to priority event queue to indicate that class will | |
314 | * change its mode in cl->pq_key microseconds. Make sure that class is not | |
315 | * already in the queue. | |
316 | */ | |
87990467 | 317 | static void htb_add_to_wait_tree(struct htb_sched *q, |
56b765b7 | 318 | struct htb_class *cl, s64 delay) |
1da177e4 | 319 | { |
c9364636 | 320 | struct rb_node **p = &q->hlevel[cl->level].wait_pq.rb_node, *parent = NULL; |
3bf72957 | 321 | |
fb983d45 PM |
322 | cl->pq_key = q->now + delay; |
323 | if (cl->pq_key == q->now) | |
1da177e4 LT |
324 | cl->pq_key++; |
325 | ||
326 | /* update the nearest event cache */ | |
fb983d45 | 327 | if (q->near_ev_cache[cl->level] > cl->pq_key) |
1da177e4 | 328 | q->near_ev_cache[cl->level] = cl->pq_key; |
87990467 | 329 | |
1da177e4 | 330 | while (*p) { |
87990467 SH |
331 | struct htb_class *c; |
332 | parent = *p; | |
1da177e4 | 333 | c = rb_entry(parent, struct htb_class, pq_node); |
fb983d45 | 334 | if (cl->pq_key >= c->pq_key) |
1da177e4 | 335 | p = &parent->rb_right; |
87990467 | 336 | else |
1da177e4 LT |
337 | p = &parent->rb_left; |
338 | } | |
339 | rb_link_node(&cl->pq_node, parent, p); | |
c9364636 | 340 | rb_insert_color(&cl->pq_node, &q->hlevel[cl->level].wait_pq); |
1da177e4 LT |
341 | } |
342 | ||
343 | /** | |
344 | * htb_next_rb_node - finds next node in binary tree | |
274e5d0e | 345 | * @n: the current node in binary tree |
1da177e4 LT |
346 | * |
347 | * When we are past last key we return NULL. | |
348 | * Average complexity is 2 steps per call. | |
349 | */ | |
3696f625 | 350 | static inline void htb_next_rb_node(struct rb_node **n) |
1da177e4 LT |
351 | { |
352 | *n = rb_next(*n); | |
353 | } | |
354 | ||
355 | /** | |
356 | * htb_add_class_to_row - add class to its row | |
996bccc3 YK |
357 | * @q: the priority event queue |
358 | * @cl: the class to add | |
359 | * @mask: the given priorities in class in bitmap | |
1da177e4 LT |
360 | * |
361 | * The class is added to row at priorities marked in mask. | |
362 | * It does nothing if mask == 0. | |
363 | */ | |
87990467 SH |
364 | static inline void htb_add_class_to_row(struct htb_sched *q, |
365 | struct htb_class *cl, int mask) | |
1da177e4 | 366 | { |
1da177e4 LT |
367 | q->row_mask[cl->level] |= mask; |
368 | while (mask) { | |
369 | int prio = ffz(~mask); | |
370 | mask &= ~(1 << prio); | |
c9364636 | 371 | htb_add_to_id_tree(&q->hlevel[cl->level].hprio[prio].row, cl, prio); |
1da177e4 LT |
372 | } |
373 | } | |
374 | ||
3696f625 SH |
375 | /* If this triggers, it is a bug in this code, but it need not be fatal */ |
376 | static void htb_safe_rb_erase(struct rb_node *rb, struct rb_root *root) | |
377 | { | |
81771b3b | 378 | if (RB_EMPTY_NODE(rb)) { |
3696f625 SH |
379 | WARN_ON(1); |
380 | } else { | |
381 | rb_erase(rb, root); | |
382 | RB_CLEAR_NODE(rb); | |
383 | } | |
384 | } | |
385 | ||
386 | ||
1da177e4 LT |
387 | /** |
388 | * htb_remove_class_from_row - removes class from its row | |
5f8c6d05 YK |
389 | * @q: the priority event queue |
390 | * @cl: the class to add | |
391 | * @mask: the given priorities in class in bitmap | |
1da177e4 LT |
392 | * |
393 | * The class is removed from row at priorities marked in mask. | |
394 | * It does nothing if mask == 0. | |
395 | */ | |
87990467 SH |
396 | static inline void htb_remove_class_from_row(struct htb_sched *q, |
397 | struct htb_class *cl, int mask) | |
1da177e4 LT |
398 | { |
399 | int m = 0; | |
c9364636 | 400 | struct htb_level *hlevel = &q->hlevel[cl->level]; |
3bf72957 | 401 | |
1da177e4 LT |
402 | while (mask) { |
403 | int prio = ffz(~mask); | |
c9364636 | 404 | struct htb_prio *hprio = &hlevel->hprio[prio]; |
3696f625 | 405 | |
1da177e4 | 406 | mask &= ~(1 << prio); |
c9364636 ED |
407 | if (hprio->ptr == cl->node + prio) |
408 | htb_next_rb_node(&hprio->ptr); | |
3696f625 | 409 | |
c9364636 ED |
410 | htb_safe_rb_erase(cl->node + prio, &hprio->row); |
411 | if (!hprio->row.rb_node) | |
1da177e4 LT |
412 | m |= 1 << prio; |
413 | } | |
1da177e4 LT |
414 | q->row_mask[cl->level] &= ~m; |
415 | } | |
416 | ||
417 | /** | |
418 | * htb_activate_prios - creates active classe's feed chain | |
876b5fc0 YK |
419 | * @q: the priority event queue |
420 | * @cl: the class to activate | |
1da177e4 LT |
421 | * |
422 | * The class is connected to ancestors and/or appropriate rows | |
10297b99 | 423 | * for priorities it is participating on. cl->cmode must be new |
1da177e4 LT |
424 | * (activated) mode. It does nothing if cl->prio_activity == 0. |
425 | */ | |
87990467 | 426 | static void htb_activate_prios(struct htb_sched *q, struct htb_class *cl) |
1da177e4 LT |
427 | { |
428 | struct htb_class *p = cl->parent; | |
87990467 | 429 | long m, mask = cl->prio_activity; |
1da177e4 LT |
430 | |
431 | while (cl->cmode == HTB_MAY_BORROW && p && mask) { | |
87990467 SH |
432 | m = mask; |
433 | while (m) { | |
de5ca4c3 KC |
434 | unsigned int prio = ffz(~m); |
435 | ||
9cec2aaf | 436 | if (WARN_ON_ONCE(prio >= ARRAY_SIZE(p->inner.clprio))) |
de5ca4c3 | 437 | break; |
1da177e4 | 438 | m &= ~(1 << prio); |
87990467 | 439 | |
11957be2 | 440 | if (p->inner.clprio[prio].feed.rb_node) |
1da177e4 | 441 | /* parent already has its feed in use so that |
cc7ec456 ED |
442 | * reset bit in mask as parent is already ok |
443 | */ | |
1da177e4 | 444 | mask &= ~(1 << prio); |
87990467 | 445 | |
11957be2 | 446 | htb_add_to_id_tree(&p->inner.clprio[prio].feed, cl, prio); |
1da177e4 | 447 | } |
1da177e4 | 448 | p->prio_activity |= mask; |
87990467 SH |
449 | cl = p; |
450 | p = cl->parent; | |
3bf72957 | 451 | |
1da177e4 LT |
452 | } |
453 | if (cl->cmode == HTB_CAN_SEND && mask) | |
87990467 | 454 | htb_add_class_to_row(q, cl, mask); |
1da177e4 LT |
455 | } |
456 | ||
457 | /** | |
458 | * htb_deactivate_prios - remove class from feed chain | |
4113be20 YK |
459 | * @q: the priority event queue |
460 | * @cl: the class to deactivate | |
1da177e4 | 461 | * |
10297b99 | 462 | * cl->cmode must represent old mode (before deactivation). It does |
1da177e4 LT |
463 | * nothing if cl->prio_activity == 0. Class is removed from all feed |
464 | * chains and rows. | |
465 | */ | |
466 | static void htb_deactivate_prios(struct htb_sched *q, struct htb_class *cl) | |
467 | { | |
468 | struct htb_class *p = cl->parent; | |
87990467 | 469 | long m, mask = cl->prio_activity; |
1da177e4 LT |
470 | |
471 | while (cl->cmode == HTB_MAY_BORROW && p && mask) { | |
87990467 SH |
472 | m = mask; |
473 | mask = 0; | |
1da177e4 LT |
474 | while (m) { |
475 | int prio = ffz(~m); | |
476 | m &= ~(1 << prio); | |
87990467 | 477 | |
11957be2 | 478 | if (p->inner.clprio[prio].ptr == cl->node + prio) { |
1da177e4 | 479 | /* we are removing child which is pointed to from |
cc7ec456 ED |
480 | * parent feed - forget the pointer but remember |
481 | * classid | |
482 | */ | |
11957be2 CW |
483 | p->inner.clprio[prio].last_ptr_id = cl->common.classid; |
484 | p->inner.clprio[prio].ptr = NULL; | |
1da177e4 | 485 | } |
87990467 | 486 | |
c9364636 | 487 | htb_safe_rb_erase(cl->node + prio, |
11957be2 | 488 | &p->inner.clprio[prio].feed); |
87990467 | 489 | |
11957be2 | 490 | if (!p->inner.clprio[prio].feed.rb_node) |
1da177e4 LT |
491 | mask |= 1 << prio; |
492 | } | |
3bf72957 | 493 | |
1da177e4 | 494 | p->prio_activity &= ~mask; |
87990467 SH |
495 | cl = p; |
496 | p = cl->parent; | |
3bf72957 | 497 | |
1da177e4 | 498 | } |
87990467 SH |
499 | if (cl->cmode == HTB_CAN_SEND && mask) |
500 | htb_remove_class_from_row(q, cl, mask); | |
1da177e4 LT |
501 | } |
502 | ||
56b765b7 | 503 | static inline s64 htb_lowater(const struct htb_class *cl) |
18a63e86 | 504 | { |
47083fc0 JDB |
505 | if (htb_hysteresis) |
506 | return cl->cmode != HTB_CANT_SEND ? -cl->cbuffer : 0; | |
507 | else | |
508 | return 0; | |
18a63e86 | 509 | } |
56b765b7 | 510 | static inline s64 htb_hiwater(const struct htb_class *cl) |
18a63e86 | 511 | { |
47083fc0 JDB |
512 | if (htb_hysteresis) |
513 | return cl->cmode == HTB_CAN_SEND ? -cl->buffer : 0; | |
514 | else | |
515 | return 0; | |
18a63e86 | 516 | } |
47083fc0 | 517 | |
18a63e86 | 518 | |
1da177e4 LT |
519 | /** |
520 | * htb_class_mode - computes and returns current class mode | |
1e955952 YK |
521 | * @cl: the target class |
522 | * @diff: diff time in microseconds | |
1da177e4 LT |
523 | * |
524 | * It computes cl's mode at time cl->t_c+diff and returns it. If mode | |
525 | * is not HTB_CAN_SEND then cl->pq_key is updated to time difference | |
10297b99 | 526 | * from now to time when cl will change its state. |
1da177e4 | 527 | * Also it is worth to note that class mode doesn't change simply |
10297b99 | 528 | * at cl->{c,}tokens == 0 but there can rather be hysteresis of |
1da177e4 LT |
529 | * 0 .. -cl->{c,}buffer range. It is meant to limit number of |
530 | * mode transitions per time unit. The speed gain is about 1/6. | |
531 | */ | |
87990467 | 532 | static inline enum htb_cmode |
56b765b7 | 533 | htb_class_mode(struct htb_class *cl, s64 *diff) |
1da177e4 | 534 | { |
56b765b7 | 535 | s64 toks; |
1da177e4 | 536 | |
87990467 SH |
537 | if ((toks = (cl->ctokens + *diff)) < htb_lowater(cl)) { |
538 | *diff = -toks; | |
539 | return HTB_CANT_SEND; | |
540 | } | |
18a63e86 | 541 | |
87990467 SH |
542 | if ((toks = (cl->tokens + *diff)) >= htb_hiwater(cl)) |
543 | return HTB_CAN_SEND; | |
1da177e4 | 544 | |
87990467 SH |
545 | *diff = -toks; |
546 | return HTB_MAY_BORROW; | |
1da177e4 LT |
547 | } |
548 | ||
549 | /** | |
550 | * htb_change_class_mode - changes classe's mode | |
4b479e98 YK |
551 | * @q: the priority event queue |
552 | * @cl: the target class | |
553 | * @diff: diff time in microseconds | |
1da177e4 LT |
554 | * |
555 | * This should be the only way how to change classe's mode under normal | |
37f2ad2b | 556 | * circumstances. Routine will update feed lists linkage, change mode |
1da177e4 LT |
557 | * and add class to the wait event queue if appropriate. New mode should |
558 | * be different from old one and cl->pq_key has to be valid if changing | |
559 | * to mode other than HTB_CAN_SEND (see htb_add_to_wait_tree). | |
560 | */ | |
87990467 | 561 | static void |
56b765b7 | 562 | htb_change_class_mode(struct htb_sched *q, struct htb_class *cl, s64 *diff) |
87990467 SH |
563 | { |
564 | enum htb_cmode new_mode = htb_class_mode(cl, diff); | |
1da177e4 LT |
565 | |
566 | if (new_mode == cl->cmode) | |
87990467 SH |
567 | return; |
568 | ||
b362487a | 569 | if (new_mode == HTB_CANT_SEND) { |
3c75f6ee | 570 | cl->overlimits++; |
b362487a CW |
571 | q->overlimits++; |
572 | } | |
3c75f6ee | 573 | |
87990467 SH |
574 | if (cl->prio_activity) { /* not necessary: speed optimization */ |
575 | if (cl->cmode != HTB_CANT_SEND) | |
576 | htb_deactivate_prios(q, cl); | |
1da177e4 | 577 | cl->cmode = new_mode; |
87990467 SH |
578 | if (new_mode != HTB_CANT_SEND) |
579 | htb_activate_prios(q, cl); | |
580 | } else | |
1da177e4 LT |
581 | cl->cmode = new_mode; |
582 | } | |
583 | ||
584 | /** | |
10297b99 | 585 | * htb_activate - inserts leaf cl into appropriate active feeds |
8df7e8ff YK |
586 | * @q: the priority event queue |
587 | * @cl: the target class | |
1da177e4 LT |
588 | * |
589 | * Routine learns (new) priority of leaf and activates feed chain | |
590 | * for the prio. It can be called on already active leaf safely. | |
591 | * It also adds leaf into droplist. | |
592 | */ | |
87990467 | 593 | static inline void htb_activate(struct htb_sched *q, struct htb_class *cl) |
1da177e4 | 594 | { |
11957be2 | 595 | WARN_ON(cl->level || !cl->leaf.q || !cl->leaf.q->q.qlen); |
3bf72957 | 596 | |
1da177e4 | 597 | if (!cl->prio_activity) { |
c19f7a34 | 598 | cl->prio_activity = 1 << cl->prio; |
87990467 | 599 | htb_activate_prios(q, cl); |
1da177e4 LT |
600 | } |
601 | } | |
602 | ||
603 | /** | |
10297b99 | 604 | * htb_deactivate - remove leaf cl from active feeds |
9a034f25 YK |
605 | * @q: the priority event queue |
606 | * @cl: the target class | |
1da177e4 LT |
607 | * |
608 | * Make sure that leaf is active. In the other words it can't be called | |
609 | * with non-active leaf. It also removes class from the drop list. | |
610 | */ | |
87990467 | 611 | static inline void htb_deactivate(struct htb_sched *q, struct htb_class *cl) |
1da177e4 | 612 | { |
547b792c | 613 | WARN_ON(!cl->prio_activity); |
3bf72957 | 614 | |
87990467 | 615 | htb_deactivate_prios(q, cl); |
1da177e4 | 616 | cl->prio_activity = 0; |
1da177e4 LT |
617 | } |
618 | ||
520ac30f ED |
619 | static int htb_enqueue(struct sk_buff *skb, struct Qdisc *sch, |
620 | struct sk_buff **to_free) | |
1da177e4 | 621 | { |
3f649ab7 | 622 | int ret; |
f6bab199 | 623 | unsigned int len = qdisc_pkt_len(skb); |
87990467 SH |
624 | struct htb_sched *q = qdisc_priv(sch); |
625 | struct htb_class *cl = htb_classify(skb, sch, &ret); | |
626 | ||
627 | if (cl == HTB_DIRECT) { | |
628 | /* enqueue to helper queue */ | |
629 | if (q->direct_queue.qlen < q->direct_qlen) { | |
aea890b8 | 630 | __qdisc_enqueue_tail(skb, &q->direct_queue); |
87990467 SH |
631 | q->direct_pkts++; |
632 | } else { | |
520ac30f | 633 | return qdisc_drop(skb, sch, to_free); |
87990467 | 634 | } |
1da177e4 | 635 | #ifdef CONFIG_NET_CLS_ACT |
87990467 | 636 | } else if (!cl) { |
c27f339a | 637 | if (ret & __NET_XMIT_BYPASS) |
25331d6c | 638 | qdisc_qstats_drop(sch); |
520ac30f | 639 | __qdisc_drop(skb, to_free); |
87990467 | 640 | return ret; |
1da177e4 | 641 | #endif |
11957be2 | 642 | } else if ((ret = qdisc_enqueue(skb, cl->leaf.q, |
520ac30f | 643 | to_free)) != NET_XMIT_SUCCESS) { |
378a2f09 | 644 | if (net_xmit_drop_count(ret)) { |
25331d6c | 645 | qdisc_qstats_drop(sch); |
338ed9b4 | 646 | cl->drops++; |
378a2f09 | 647 | } |
69747650 | 648 | return ret; |
87990467 | 649 | } else { |
87990467 SH |
650 | htb_activate(q, cl); |
651 | } | |
652 | ||
f6bab199 | 653 | sch->qstats.backlog += len; |
87990467 | 654 | sch->q.qlen++; |
87990467 | 655 | return NET_XMIT_SUCCESS; |
1da177e4 LT |
656 | } |
657 | ||
56b765b7 | 658 | static inline void htb_accnt_tokens(struct htb_class *cl, int bytes, s64 diff) |
59e4220a | 659 | { |
56b765b7 | 660 | s64 toks = diff + cl->tokens; |
59e4220a JP |
661 | |
662 | if (toks > cl->buffer) | |
663 | toks = cl->buffer; | |
292f1c7f | 664 | toks -= (s64) psched_l2t_ns(&cl->rate, bytes); |
59e4220a JP |
665 | if (toks <= -cl->mbuffer) |
666 | toks = 1 - cl->mbuffer; | |
667 | ||
668 | cl->tokens = toks; | |
669 | } | |
670 | ||
56b765b7 | 671 | static inline void htb_accnt_ctokens(struct htb_class *cl, int bytes, s64 diff) |
59e4220a | 672 | { |
56b765b7 | 673 | s64 toks = diff + cl->ctokens; |
59e4220a JP |
674 | |
675 | if (toks > cl->cbuffer) | |
676 | toks = cl->cbuffer; | |
292f1c7f | 677 | toks -= (s64) psched_l2t_ns(&cl->ceil, bytes); |
59e4220a JP |
678 | if (toks <= -cl->mbuffer) |
679 | toks = 1 - cl->mbuffer; | |
680 | ||
681 | cl->ctokens = toks; | |
682 | } | |
683 | ||
1da177e4 LT |
684 | /** |
685 | * htb_charge_class - charges amount "bytes" to leaf and ancestors | |
0e5c9084 YK |
686 | * @q: the priority event queue |
687 | * @cl: the class to start iterate | |
688 | * @level: the minimum level to account | |
689 | * @skb: the socket buffer | |
1da177e4 LT |
690 | * |
691 | * Routine assumes that packet "bytes" long was dequeued from leaf cl | |
692 | * borrowing from "level". It accounts bytes to ceil leaky bucket for | |
693 | * leaf and all ancestors and to rate bucket for ancestors at levels | |
694 | * "level" and higher. It also handles possible change of mode resulting | |
695 | * from the update. Note that mode can also increase here (MAY_BORROW to | |
696 | * CAN_SEND) because we can use more precise clock that event queue here. | |
697 | * In such case we remove class from event queue first. | |
698 | */ | |
87990467 | 699 | static void htb_charge_class(struct htb_sched *q, struct htb_class *cl, |
c9726d68 | 700 | int level, struct sk_buff *skb) |
87990467 | 701 | { |
0abf77e5 | 702 | int bytes = qdisc_pkt_len(skb); |
1da177e4 | 703 | enum htb_cmode old_mode; |
56b765b7 | 704 | s64 diff; |
1da177e4 LT |
705 | |
706 | while (cl) { | |
56b765b7 | 707 | diff = min_t(s64, q->now - cl->t_c, cl->mbuffer); |
1da177e4 | 708 | if (cl->level >= level) { |
87990467 SH |
709 | if (cl->level == level) |
710 | cl->xstats.lends++; | |
59e4220a | 711 | htb_accnt_tokens(cl, bytes, diff); |
1da177e4 LT |
712 | } else { |
713 | cl->xstats.borrows++; | |
87990467 | 714 | cl->tokens += diff; /* we moved t_c; update tokens */ |
1da177e4 | 715 | } |
59e4220a | 716 | htb_accnt_ctokens(cl, bytes, diff); |
1da177e4 | 717 | cl->t_c = q->now; |
1da177e4 | 718 | |
87990467 SH |
719 | old_mode = cl->cmode; |
720 | diff = 0; | |
721 | htb_change_class_mode(q, cl, &diff); | |
1da177e4 LT |
722 | if (old_mode != cl->cmode) { |
723 | if (old_mode != HTB_CAN_SEND) | |
c9364636 | 724 | htb_safe_rb_erase(&cl->pq_node, &q->hlevel[cl->level].wait_pq); |
1da177e4 | 725 | if (cl->cmode != HTB_CAN_SEND) |
87990467 | 726 | htb_add_to_wait_tree(q, cl, diff); |
1da177e4 | 727 | } |
1da177e4 | 728 | |
bfe0d029 ED |
729 | /* update basic stats except for leaves which are already updated */ |
730 | if (cl->level) | |
731 | bstats_update(&cl->bstats, skb); | |
732 | ||
1da177e4 LT |
733 | cl = cl->parent; |
734 | } | |
735 | } | |
736 | ||
737 | /** | |
738 | * htb_do_events - make mode changes to classes at the level | |
2c3ee53e YK |
739 | * @q: the priority event queue |
740 | * @level: which wait_pq in 'q->hlevel' | |
741 | * @start: start jiffies | |
1da177e4 | 742 | * |
fb983d45 | 743 | * Scans event queue for pending events and applies them. Returns time of |
1224736d | 744 | * next pending event (0 for no event in pq, q->now for too many events). |
fb983d45 | 745 | * Note: Applied are events whose have cl->pq_key <= q->now. |
1da177e4 | 746 | */ |
c9364636 | 747 | static s64 htb_do_events(struct htb_sched *q, const int level, |
5343a7f8 | 748 | unsigned long start) |
1da177e4 | 749 | { |
8f3ea33a | 750 | /* don't run for longer than 2 jiffies; 2 is used instead of |
cc7ec456 ED |
751 | * 1 to simplify things when jiffy is going to be incremented |
752 | * too soon | |
753 | */ | |
a73be040 | 754 | unsigned long stop_at = start + 2; |
c9364636 ED |
755 | struct rb_root *wait_pq = &q->hlevel[level].wait_pq; |
756 | ||
8f3ea33a | 757 | while (time_before(jiffies, stop_at)) { |
1da177e4 | 758 | struct htb_class *cl; |
56b765b7 | 759 | s64 diff; |
c9364636 | 760 | struct rb_node *p = rb_first(wait_pq); |
30bdbe39 | 761 | |
87990467 SH |
762 | if (!p) |
763 | return 0; | |
1da177e4 LT |
764 | |
765 | cl = rb_entry(p, struct htb_class, pq_node); | |
fb983d45 PM |
766 | if (cl->pq_key > q->now) |
767 | return cl->pq_key; | |
768 | ||
c9364636 | 769 | htb_safe_rb_erase(p, wait_pq); |
56b765b7 | 770 | diff = min_t(s64, q->now - cl->t_c, cl->mbuffer); |
87990467 | 771 | htb_change_class_mode(q, cl, &diff); |
1da177e4 | 772 | if (cl->cmode != HTB_CAN_SEND) |
87990467 | 773 | htb_add_to_wait_tree(q, cl, diff); |
1da177e4 | 774 | } |
1224736d JP |
775 | |
776 | /* too much load - let's continue after a break for scheduling */ | |
e82181de | 777 | if (!(q->warned & HTB_WARN_TOOMANYEVENTS)) { |
c17988a9 | 778 | pr_warn("htb: too many events!\n"); |
e82181de JP |
779 | q->warned |= HTB_WARN_TOOMANYEVENTS; |
780 | } | |
1224736d JP |
781 | |
782 | return q->now; | |
1da177e4 LT |
783 | } |
784 | ||
785 | /* Returns class->node+prio from id-tree where classe's id is >= id. NULL | |
cc7ec456 ED |
786 | * is no such one exists. |
787 | */ | |
87990467 SH |
788 | static struct rb_node *htb_id_find_next_upper(int prio, struct rb_node *n, |
789 | u32 id) | |
1da177e4 LT |
790 | { |
791 | struct rb_node *r = NULL; | |
792 | while (n) { | |
87990467 SH |
793 | struct htb_class *cl = |
794 | rb_entry(n, struct htb_class, node[prio]); | |
87990467 | 795 | |
f4c1f3e0 | 796 | if (id > cl->common.classid) { |
1da177e4 | 797 | n = n->rb_right; |
1b5c0077 | 798 | } else if (id < cl->common.classid) { |
1da177e4 LT |
799 | r = n; |
800 | n = n->rb_left; | |
1b5c0077 JP |
801 | } else { |
802 | return n; | |
1da177e4 LT |
803 | } |
804 | } | |
805 | return r; | |
806 | } | |
807 | ||
808 | /** | |
809 | * htb_lookup_leaf - returns next leaf class in DRR order | |
9977d6f5 YK |
810 | * @hprio: the current one |
811 | * @prio: which prio in class | |
1da177e4 LT |
812 | * |
813 | * Find leaf where current feed pointers points to. | |
814 | */ | |
c9364636 | 815 | static struct htb_class *htb_lookup_leaf(struct htb_prio *hprio, const int prio) |
1da177e4 LT |
816 | { |
817 | int i; | |
818 | struct { | |
819 | struct rb_node *root; | |
820 | struct rb_node **pptr; | |
821 | u32 *pid; | |
87990467 SH |
822 | } stk[TC_HTB_MAXDEPTH], *sp = stk; |
823 | ||
c9364636 ED |
824 | BUG_ON(!hprio->row.rb_node); |
825 | sp->root = hprio->row.rb_node; | |
826 | sp->pptr = &hprio->ptr; | |
827 | sp->pid = &hprio->last_ptr_id; | |
1da177e4 LT |
828 | |
829 | for (i = 0; i < 65535; i++) { | |
87990467 | 830 | if (!*sp->pptr && *sp->pid) { |
10297b99 | 831 | /* ptr was invalidated but id is valid - try to recover |
cc7ec456 ED |
832 | * the original or next ptr |
833 | */ | |
87990467 SH |
834 | *sp->pptr = |
835 | htb_id_find_next_upper(prio, sp->root, *sp->pid); | |
1da177e4 | 836 | } |
87990467 | 837 | *sp->pid = 0; /* ptr is valid now so that remove this hint as it |
cc7ec456 ED |
838 | * can become out of date quickly |
839 | */ | |
87990467 | 840 | if (!*sp->pptr) { /* we are at right end; rewind & go up */ |
1da177e4 | 841 | *sp->pptr = sp->root; |
87990467 | 842 | while ((*sp->pptr)->rb_left) |
1da177e4 LT |
843 | *sp->pptr = (*sp->pptr)->rb_left; |
844 | if (sp > stk) { | |
845 | sp--; | |
512bb43e JP |
846 | if (!*sp->pptr) { |
847 | WARN_ON(1); | |
87990467 | 848 | return NULL; |
512bb43e | 849 | } |
87990467 | 850 | htb_next_rb_node(sp->pptr); |
1da177e4 LT |
851 | } |
852 | } else { | |
853 | struct htb_class *cl; | |
c9364636 ED |
854 | struct htb_prio *clp; |
855 | ||
87990467 SH |
856 | cl = rb_entry(*sp->pptr, struct htb_class, node[prio]); |
857 | if (!cl->level) | |
1da177e4 | 858 | return cl; |
11957be2 | 859 | clp = &cl->inner.clprio[prio]; |
c9364636 ED |
860 | (++sp)->root = clp->feed.rb_node; |
861 | sp->pptr = &clp->ptr; | |
862 | sp->pid = &clp->last_ptr_id; | |
1da177e4 LT |
863 | } |
864 | } | |
547b792c | 865 | WARN_ON(1); |
1da177e4 LT |
866 | return NULL; |
867 | } | |
868 | ||
869 | /* dequeues packet at given priority and level; call only if | |
cc7ec456 ED |
870 | * you are sure that there is active class at prio/level |
871 | */ | |
c9364636 ED |
872 | static struct sk_buff *htb_dequeue_tree(struct htb_sched *q, const int prio, |
873 | const int level) | |
1da177e4 LT |
874 | { |
875 | struct sk_buff *skb = NULL; | |
87990467 | 876 | struct htb_class *cl, *start; |
c9364636 ED |
877 | struct htb_level *hlevel = &q->hlevel[level]; |
878 | struct htb_prio *hprio = &hlevel->hprio[prio]; | |
879 | ||
1da177e4 | 880 | /* look initial class up in the row */ |
c9364636 | 881 | start = cl = htb_lookup_leaf(hprio, prio); |
87990467 | 882 | |
1da177e4 LT |
883 | do { |
884 | next: | |
512bb43e | 885 | if (unlikely(!cl)) |
87990467 | 886 | return NULL; |
1da177e4 LT |
887 | |
888 | /* class can be empty - it is unlikely but can be true if leaf | |
cc7ec456 ED |
889 | * qdisc drops packets in enqueue routine or if someone used |
890 | * graft operation on the leaf since last dequeue; | |
891 | * simply deactivate and skip such class | |
892 | */ | |
11957be2 | 893 | if (unlikely(cl->leaf.q->q.qlen == 0)) { |
1da177e4 | 894 | struct htb_class *next; |
87990467 | 895 | htb_deactivate(q, cl); |
1da177e4 LT |
896 | |
897 | /* row/level might become empty */ | |
898 | if ((q->row_mask[level] & (1 << prio)) == 0) | |
87990467 | 899 | return NULL; |
1da177e4 | 900 | |
c9364636 | 901 | next = htb_lookup_leaf(hprio, prio); |
87990467 SH |
902 | |
903 | if (cl == start) /* fix start if we just deleted it */ | |
1da177e4 LT |
904 | start = next; |
905 | cl = next; | |
906 | goto next; | |
907 | } | |
87990467 | 908 | |
11957be2 | 909 | skb = cl->leaf.q->dequeue(cl->leaf.q); |
87990467 | 910 | if (likely(skb != NULL)) |
1da177e4 | 911 | break; |
633fe66e | 912 | |
11957be2 CW |
913 | qdisc_warn_nonwc("htb", cl->leaf.q); |
914 | htb_next_rb_node(level ? &cl->parent->inner.clprio[prio].ptr: | |
c9364636 ED |
915 | &q->hlevel[0].hprio[prio].ptr); |
916 | cl = htb_lookup_leaf(hprio, prio); | |
1da177e4 LT |
917 | |
918 | } while (cl != start); | |
919 | ||
920 | if (likely(skb != NULL)) { | |
196d97f6 | 921 | bstats_update(&cl->bstats, skb); |
11957be2 CW |
922 | cl->leaf.deficit[level] -= qdisc_pkt_len(skb); |
923 | if (cl->leaf.deficit[level] < 0) { | |
924 | cl->leaf.deficit[level] += cl->quantum; | |
925 | htb_next_rb_node(level ? &cl->parent->inner.clprio[prio].ptr : | |
c9364636 | 926 | &q->hlevel[0].hprio[prio].ptr); |
1da177e4 LT |
927 | } |
928 | /* this used to be after charge_class but this constelation | |
cc7ec456 ED |
929 | * gives us slightly better performance |
930 | */ | |
11957be2 | 931 | if (!cl->leaf.q->q.qlen) |
87990467 | 932 | htb_deactivate(q, cl); |
c9726d68 | 933 | htb_charge_class(q, cl, level, skb); |
1da177e4 LT |
934 | } |
935 | return skb; | |
936 | } | |
937 | ||
1da177e4 LT |
938 | static struct sk_buff *htb_dequeue(struct Qdisc *sch) |
939 | { | |
9190b3b3 | 940 | struct sk_buff *skb; |
1da177e4 LT |
941 | struct htb_sched *q = qdisc_priv(sch); |
942 | int level; | |
5343a7f8 | 943 | s64 next_event; |
a73be040 | 944 | unsigned long start_at; |
1da177e4 LT |
945 | |
946 | /* try to dequeue direct packets as high prio (!) to minimize cpu work */ | |
48da34b7 | 947 | skb = __qdisc_dequeue_head(&q->direct_queue); |
87990467 | 948 | if (skb != NULL) { |
9190b3b3 ED |
949 | ok: |
950 | qdisc_bstats_update(sch, skb); | |
431e3a8e | 951 | qdisc_qstats_backlog_dec(sch, skb); |
1da177e4 LT |
952 | sch->q.qlen--; |
953 | return skb; | |
954 | } | |
955 | ||
87990467 SH |
956 | if (!sch->q.qlen) |
957 | goto fin; | |
d2de875c | 958 | q->now = ktime_get_ns(); |
a73be040 | 959 | start_at = jiffies; |
1da177e4 | 960 | |
d2fe85da | 961 | next_event = q->now + 5LLU * NSEC_PER_SEC; |
633fe66e | 962 | |
1da177e4 LT |
963 | for (level = 0; level < TC_HTB_MAXDEPTH; level++) { |
964 | /* common case optimization - skip event handler quickly */ | |
965 | int m; | |
c9364636 | 966 | s64 event = q->near_ev_cache[level]; |
fb983d45 | 967 | |
c9364636 | 968 | if (q->now >= event) { |
a73be040 | 969 | event = htb_do_events(q, level, start_at); |
2e4b3b0e | 970 | if (!event) |
56b765b7 | 971 | event = q->now + NSEC_PER_SEC; |
2e4b3b0e | 972 | q->near_ev_cache[level] = event; |
c9364636 | 973 | } |
fb983d45 | 974 | |
c0851347 | 975 | if (next_event > event) |
fb983d45 | 976 | next_event = event; |
87990467 | 977 | |
1da177e4 LT |
978 | m = ~q->row_mask[level]; |
979 | while (m != (int)(-1)) { | |
87990467 | 980 | int prio = ffz(m); |
cc7ec456 | 981 | |
1da177e4 | 982 | m |= 1 << prio; |
87990467 | 983 | skb = htb_dequeue_tree(q, prio, level); |
9190b3b3 ED |
984 | if (likely(skb != NULL)) |
985 | goto ok; | |
1da177e4 LT |
986 | } |
987 | } | |
a9efad8b | 988 | if (likely(next_event > q->now)) |
45f50bed | 989 | qdisc_watchdog_schedule_ns(&q->watchdog, next_event); |
a9efad8b | 990 | else |
1224736d | 991 | schedule_work(&q->work); |
1da177e4 | 992 | fin: |
1da177e4 LT |
993 | return skb; |
994 | } | |
995 | ||
1da177e4 LT |
996 | /* reset all classes */ |
997 | /* always caled under BH & queue lock */ | |
87990467 | 998 | static void htb_reset(struct Qdisc *sch) |
1da177e4 LT |
999 | { |
1000 | struct htb_sched *q = qdisc_priv(sch); | |
f4c1f3e0 | 1001 | struct htb_class *cl; |
f4c1f3e0 | 1002 | unsigned int i; |
0cef296d | 1003 | |
f4c1f3e0 | 1004 | for (i = 0; i < q->clhash.hashsize; i++) { |
b67bfe0d | 1005 | hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) { |
1da177e4 | 1006 | if (cl->level) |
11957be2 | 1007 | memset(&cl->inner, 0, sizeof(cl->inner)); |
1da177e4 | 1008 | else { |
d03b195b | 1009 | if (cl->leaf.q && !q->offload) |
11957be2 | 1010 | qdisc_reset(cl->leaf.q); |
1da177e4 LT |
1011 | } |
1012 | cl->prio_activity = 0; | |
1013 | cl->cmode = HTB_CAN_SEND; | |
1da177e4 LT |
1014 | } |
1015 | } | |
fb983d45 | 1016 | qdisc_watchdog_cancel(&q->watchdog); |
a5a9f534 | 1017 | __qdisc_reset_queue(&q->direct_queue); |
c9364636 | 1018 | memset(q->hlevel, 0, sizeof(q->hlevel)); |
87990467 | 1019 | memset(q->row_mask, 0, sizeof(q->row_mask)); |
1da177e4 LT |
1020 | } |
1021 | ||
27a3421e PM |
1022 | static const struct nla_policy htb_policy[TCA_HTB_MAX + 1] = { |
1023 | [TCA_HTB_PARMS] = { .len = sizeof(struct tc_htb_opt) }, | |
1024 | [TCA_HTB_INIT] = { .len = sizeof(struct tc_htb_glob) }, | |
1025 | [TCA_HTB_CTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE }, | |
1026 | [TCA_HTB_RTAB] = { .type = NLA_BINARY, .len = TC_RTAB_SIZE }, | |
6906f4ed | 1027 | [TCA_HTB_DIRECT_QLEN] = { .type = NLA_U32 }, |
df62cdf3 ED |
1028 | [TCA_HTB_RATE64] = { .type = NLA_U64 }, |
1029 | [TCA_HTB_CEIL64] = { .type = NLA_U64 }, | |
d03b195b | 1030 | [TCA_HTB_OFFLOAD] = { .type = NLA_FLAG }, |
27a3421e PM |
1031 | }; |
1032 | ||
1224736d JP |
1033 | static void htb_work_func(struct work_struct *work) |
1034 | { | |
1035 | struct htb_sched *q = container_of(work, struct htb_sched, work); | |
1036 | struct Qdisc *sch = q->watchdog.qdisc; | |
1037 | ||
0ee13627 | 1038 | rcu_read_lock(); |
1224736d | 1039 | __netif_schedule(qdisc_root(sch)); |
0ee13627 | 1040 | rcu_read_unlock(); |
1224736d JP |
1041 | } |
1042 | ||
d03b195b MM |
1043 | static void htb_set_lockdep_class_child(struct Qdisc *q) |
1044 | { | |
1045 | static struct lock_class_key child_key; | |
1046 | ||
1047 | lockdep_set_class(qdisc_lock(q), &child_key); | |
1048 | } | |
1049 | ||
1050 | static int htb_offload(struct net_device *dev, struct tc_htb_qopt_offload *opt) | |
1051 | { | |
1052 | return dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_HTB, opt); | |
1053 | } | |
1054 | ||
e63d7dfd AA |
1055 | static int htb_init(struct Qdisc *sch, struct nlattr *opt, |
1056 | struct netlink_ext_ack *extack) | |
1da177e4 | 1057 | { |
d03b195b MM |
1058 | struct net_device *dev = qdisc_dev(sch); |
1059 | struct tc_htb_qopt_offload offload_opt; | |
1da177e4 | 1060 | struct htb_sched *q = qdisc_priv(sch); |
6906f4ed | 1061 | struct nlattr *tb[TCA_HTB_MAX + 1]; |
1da177e4 | 1062 | struct tc_htb_glob *gopt; |
d03b195b | 1063 | unsigned int ntx; |
fb3a3e37 | 1064 | bool offload; |
cee63723 | 1065 | int err; |
cee63723 | 1066 | |
88c2ace6 NA |
1067 | qdisc_watchdog_init(&q->watchdog, sch); |
1068 | INIT_WORK(&q->work, htb_work_func); | |
1069 | ||
cee63723 PM |
1070 | if (!opt) |
1071 | return -EINVAL; | |
1072 | ||
8d1a77f9 | 1073 | err = tcf_block_get(&q->block, &q->filter_list, sch, extack); |
6529eaba JP |
1074 | if (err) |
1075 | return err; | |
1076 | ||
8cb08174 JB |
1077 | err = nla_parse_nested_deprecated(tb, TCA_HTB_MAX, opt, htb_policy, |
1078 | NULL); | |
cee63723 PM |
1079 | if (err < 0) |
1080 | return err; | |
1081 | ||
6906f4ed | 1082 | if (!tb[TCA_HTB_INIT]) |
1da177e4 | 1083 | return -EINVAL; |
6906f4ed | 1084 | |
1e90474c | 1085 | gopt = nla_data(tb[TCA_HTB_INIT]); |
6906f4ed | 1086 | if (gopt->version != HTB_VER >> 16) |
1da177e4 | 1087 | return -EINVAL; |
1da177e4 | 1088 | |
fb3a3e37 | 1089 | offload = nla_get_flag(tb[TCA_HTB_OFFLOAD]); |
d03b195b | 1090 | |
fb3a3e37 | 1091 | if (offload) { |
648a991c MM |
1092 | if (sch->parent != TC_H_ROOT) { |
1093 | NL_SET_ERR_MSG(extack, "HTB must be the root qdisc to use offload"); | |
d03b195b | 1094 | return -EOPNOTSUPP; |
648a991c | 1095 | } |
d03b195b | 1096 | |
648a991c MM |
1097 | if (!tc_can_offload(dev) || !dev->netdev_ops->ndo_setup_tc) { |
1098 | NL_SET_ERR_MSG(extack, "hw-tc-offload ethtool feature flag must be on"); | |
d03b195b | 1099 | return -EOPNOTSUPP; |
648a991c | 1100 | } |
d03b195b MM |
1101 | |
1102 | q->num_direct_qdiscs = dev->real_num_tx_queues; | |
1103 | q->direct_qdiscs = kcalloc(q->num_direct_qdiscs, | |
1104 | sizeof(*q->direct_qdiscs), | |
1105 | GFP_KERNEL); | |
1106 | if (!q->direct_qdiscs) | |
1107 | return -ENOMEM; | |
1108 | } | |
1109 | ||
f4c1f3e0 PM |
1110 | err = qdisc_class_hash_init(&q->clhash); |
1111 | if (err < 0) | |
d59f4e1d | 1112 | return err; |
1da177e4 | 1113 | |
6906f4ed ED |
1114 | if (tb[TCA_HTB_DIRECT_QLEN]) |
1115 | q->direct_qlen = nla_get_u32(tb[TCA_HTB_DIRECT_QLEN]); | |
348e3435 | 1116 | else |
6906f4ed | 1117 | q->direct_qlen = qdisc_dev(sch)->tx_queue_len; |
348e3435 | 1118 | |
1da177e4 LT |
1119 | if ((q->rate2quantum = gopt->rate2quantum) < 1) |
1120 | q->rate2quantum = 1; | |
1121 | q->defcls = gopt->defcls; | |
1122 | ||
fb3a3e37 | 1123 | if (!offload) |
d03b195b MM |
1124 | return 0; |
1125 | ||
1126 | for (ntx = 0; ntx < q->num_direct_qdiscs; ntx++) { | |
1127 | struct netdev_queue *dev_queue = netdev_get_tx_queue(dev, ntx); | |
1128 | struct Qdisc *qdisc; | |
1129 | ||
1130 | qdisc = qdisc_create_dflt(dev_queue, &pfifo_qdisc_ops, | |
1131 | TC_H_MAKE(sch->handle, 0), extack); | |
1132 | if (!qdisc) { | |
d59f4e1d | 1133 | return -ENOMEM; |
d03b195b MM |
1134 | } |
1135 | ||
1136 | htb_set_lockdep_class_child(qdisc); | |
1137 | q->direct_qdiscs[ntx] = qdisc; | |
1138 | qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT; | |
1139 | } | |
1140 | ||
1141 | sch->flags |= TCQ_F_MQROOT; | |
1142 | ||
1143 | offload_opt = (struct tc_htb_qopt_offload) { | |
1144 | .command = TC_HTB_CREATE, | |
1145 | .parent_classid = TC_H_MAJ(sch->handle) >> 16, | |
1146 | .classid = TC_H_MIN(q->defcls), | |
1147 | .extack = extack, | |
1148 | }; | |
1149 | err = htb_offload(dev, &offload_opt); | |
1150 | if (err) | |
d59f4e1d | 1151 | return err; |
d03b195b | 1152 | |
fb3a3e37 MM |
1153 | /* Defer this assignment, so that htb_destroy skips offload-related |
1154 | * parts (especially calling ndo_setup_tc) on errors. | |
1155 | */ | |
1156 | q->offload = true; | |
1157 | ||
1da177e4 | 1158 | return 0; |
d03b195b MM |
1159 | } |
1160 | ||
1161 | static void htb_attach_offload(struct Qdisc *sch) | |
1162 | { | |
1163 | struct net_device *dev = qdisc_dev(sch); | |
1164 | struct htb_sched *q = qdisc_priv(sch); | |
1165 | unsigned int ntx; | |
1166 | ||
1167 | for (ntx = 0; ntx < q->num_direct_qdiscs; ntx++) { | |
1168 | struct Qdisc *old, *qdisc = q->direct_qdiscs[ntx]; | |
1169 | ||
1170 | old = dev_graft_qdisc(qdisc->dev_queue, qdisc); | |
1171 | qdisc_put(old); | |
1172 | qdisc_hash_add(qdisc, false); | |
1173 | } | |
1174 | for (ntx = q->num_direct_qdiscs; ntx < dev->num_tx_queues; ntx++) { | |
1175 | struct netdev_queue *dev_queue = netdev_get_tx_queue(dev, ntx); | |
1176 | struct Qdisc *old = dev_graft_qdisc(dev_queue, NULL); | |
1177 | ||
1178 | qdisc_put(old); | |
1179 | } | |
1180 | ||
1181 | kfree(q->direct_qdiscs); | |
1182 | q->direct_qdiscs = NULL; | |
1183 | } | |
1184 | ||
1185 | static void htb_attach_software(struct Qdisc *sch) | |
1186 | { | |
1187 | struct net_device *dev = qdisc_dev(sch); | |
1188 | unsigned int ntx; | |
1189 | ||
1190 | /* Resemble qdisc_graft behavior. */ | |
1191 | for (ntx = 0; ntx < dev->num_tx_queues; ntx++) { | |
1192 | struct netdev_queue *dev_queue = netdev_get_tx_queue(dev, ntx); | |
1193 | struct Qdisc *old = dev_graft_qdisc(dev_queue, sch); | |
1194 | ||
1195 | qdisc_refcount_inc(sch); | |
1196 | ||
1197 | qdisc_put(old); | |
1198 | } | |
1199 | } | |
1200 | ||
1201 | static void htb_attach(struct Qdisc *sch) | |
1202 | { | |
1203 | struct htb_sched *q = qdisc_priv(sch); | |
1204 | ||
1205 | if (q->offload) | |
1206 | htb_attach_offload(sch); | |
1207 | else | |
1208 | htb_attach_software(sch); | |
1da177e4 LT |
1209 | } |
1210 | ||
1211 | static int htb_dump(struct Qdisc *sch, struct sk_buff *skb) | |
1212 | { | |
1213 | struct htb_sched *q = qdisc_priv(sch); | |
4b3550ef | 1214 | struct nlattr *nest; |
1da177e4 | 1215 | struct tc_htb_glob gopt; |
4b3550ef | 1216 | |
d03b195b MM |
1217 | if (q->offload) |
1218 | sch->flags |= TCQ_F_OFFLOADED; | |
1219 | else | |
1220 | sch->flags &= ~TCQ_F_OFFLOADED; | |
1221 | ||
b362487a | 1222 | sch->qstats.overlimits = q->overlimits; |
6f542efc ED |
1223 | /* Its safe to not acquire qdisc lock. As we hold RTNL, |
1224 | * no change can happen on the qdisc parameters. | |
1225 | */ | |
1da177e4 | 1226 | |
4b3550ef | 1227 | gopt.direct_pkts = q->direct_pkts; |
1da177e4 LT |
1228 | gopt.version = HTB_VER; |
1229 | gopt.rate2quantum = q->rate2quantum; | |
1230 | gopt.defcls = q->defcls; | |
3bf72957 | 1231 | gopt.debug = 0; |
4b3550ef | 1232 | |
ae0be8de | 1233 | nest = nla_nest_start_noflag(skb, TCA_OPTIONS); |
4b3550ef PM |
1234 | if (nest == NULL) |
1235 | goto nla_put_failure; | |
6906f4ed ED |
1236 | if (nla_put(skb, TCA_HTB_INIT, sizeof(gopt), &gopt) || |
1237 | nla_put_u32(skb, TCA_HTB_DIRECT_QLEN, q->direct_qlen)) | |
1b34ec43 | 1238 | goto nla_put_failure; |
d03b195b MM |
1239 | if (q->offload && nla_put_flag(skb, TCA_HTB_OFFLOAD)) |
1240 | goto nla_put_failure; | |
4b3550ef | 1241 | |
6f542efc | 1242 | return nla_nest_end(skb, nest); |
4b3550ef | 1243 | |
1e90474c | 1244 | nla_put_failure: |
4b3550ef | 1245 | nla_nest_cancel(skb, nest); |
1da177e4 LT |
1246 | return -1; |
1247 | } | |
1248 | ||
1249 | static int htb_dump_class(struct Qdisc *sch, unsigned long arg, | |
87990467 | 1250 | struct sk_buff *skb, struct tcmsg *tcm) |
1da177e4 | 1251 | { |
87990467 | 1252 | struct htb_class *cl = (struct htb_class *)arg; |
83271586 | 1253 | struct htb_sched *q = qdisc_priv(sch); |
4b3550ef | 1254 | struct nlattr *nest; |
1da177e4 LT |
1255 | struct tc_htb_opt opt; |
1256 | ||
6f542efc ED |
1257 | /* Its safe to not acquire qdisc lock. As we hold RTNL, |
1258 | * no change can happen on the class parameters. | |
1259 | */ | |
f4c1f3e0 PM |
1260 | tcm->tcm_parent = cl->parent ? cl->parent->common.classid : TC_H_ROOT; |
1261 | tcm->tcm_handle = cl->common.classid; | |
11957be2 CW |
1262 | if (!cl->level && cl->leaf.q) |
1263 | tcm->tcm_info = cl->leaf.q->handle; | |
1da177e4 | 1264 | |
ae0be8de | 1265 | nest = nla_nest_start_noflag(skb, TCA_OPTIONS); |
4b3550ef PM |
1266 | if (nest == NULL) |
1267 | goto nla_put_failure; | |
1da177e4 | 1268 | |
87990467 | 1269 | memset(&opt, 0, sizeof(opt)); |
1da177e4 | 1270 | |
01cb71d2 | 1271 | psched_ratecfg_getrate(&opt.rate, &cl->rate); |
9c10f411 | 1272 | opt.buffer = PSCHED_NS2TICKS(cl->buffer); |
01cb71d2 | 1273 | psched_ratecfg_getrate(&opt.ceil, &cl->ceil); |
9c10f411 | 1274 | opt.cbuffer = PSCHED_NS2TICKS(cl->cbuffer); |
c19f7a34 JP |
1275 | opt.quantum = cl->quantum; |
1276 | opt.prio = cl->prio; | |
87990467 | 1277 | opt.level = cl->level; |
1b34ec43 DM |
1278 | if (nla_put(skb, TCA_HTB_PARMS, sizeof(opt), &opt)) |
1279 | goto nla_put_failure; | |
83271586 MM |
1280 | if (q->offload && nla_put_flag(skb, TCA_HTB_OFFLOAD)) |
1281 | goto nla_put_failure; | |
df62cdf3 | 1282 | if ((cl->rate.rate_bytes_ps >= (1ULL << 32)) && |
2a51c1e8 ND |
1283 | nla_put_u64_64bit(skb, TCA_HTB_RATE64, cl->rate.rate_bytes_ps, |
1284 | TCA_HTB_PAD)) | |
df62cdf3 ED |
1285 | goto nla_put_failure; |
1286 | if ((cl->ceil.rate_bytes_ps >= (1ULL << 32)) && | |
2a51c1e8 ND |
1287 | nla_put_u64_64bit(skb, TCA_HTB_CEIL64, cl->ceil.rate_bytes_ps, |
1288 | TCA_HTB_PAD)) | |
df62cdf3 | 1289 | goto nla_put_failure; |
4b3550ef | 1290 | |
6f542efc | 1291 | return nla_nest_end(skb, nest); |
4b3550ef | 1292 | |
1e90474c | 1293 | nla_put_failure: |
4b3550ef | 1294 | nla_nest_cancel(skb, nest); |
1da177e4 LT |
1295 | return -1; |
1296 | } | |
1297 | ||
83271586 MM |
1298 | static void htb_offload_aggregate_stats(struct htb_sched *q, |
1299 | struct htb_class *cl) | |
1300 | { | |
f56940da | 1301 | u64 bytes = 0, packets = 0; |
83271586 MM |
1302 | struct htb_class *c; |
1303 | unsigned int i; | |
1304 | ||
50dc9a85 | 1305 | gnet_stats_basic_sync_init(&cl->bstats); |
83271586 MM |
1306 | |
1307 | for (i = 0; i < q->clhash.hashsize; i++) { | |
1308 | hlist_for_each_entry(c, &q->clhash.hash[i], common.hnode) { | |
1309 | struct htb_class *p = c; | |
1310 | ||
1311 | while (p && p->level < cl->level) | |
1312 | p = p->parent; | |
1313 | ||
1314 | if (p != cl) | |
1315 | continue; | |
1316 | ||
50dc9a85 AD |
1317 | bytes += u64_stats_read(&c->bstats_bias.bytes); |
1318 | packets += u64_stats_read(&c->bstats_bias.packets); | |
83271586 | 1319 | if (c->level == 0) { |
50dc9a85 AD |
1320 | bytes += u64_stats_read(&c->leaf.q->bstats.bytes); |
1321 | packets += u64_stats_read(&c->leaf.q->bstats.packets); | |
83271586 MM |
1322 | } |
1323 | } | |
1324 | } | |
f56940da | 1325 | _bstats_update(&cl->bstats, bytes, packets); |
83271586 MM |
1326 | } |
1327 | ||
1da177e4 | 1328 | static int |
87990467 | 1329 | htb_dump_class_stats(struct Qdisc *sch, unsigned long arg, struct gnet_dump *d) |
1da177e4 | 1330 | { |
87990467 | 1331 | struct htb_class *cl = (struct htb_class *)arg; |
83271586 | 1332 | struct htb_sched *q = qdisc_priv(sch); |
338ed9b4 ED |
1333 | struct gnet_stats_queue qs = { |
1334 | .drops = cl->drops, | |
3c75f6ee | 1335 | .overlimits = cl->overlimits, |
338ed9b4 | 1336 | }; |
64015853 | 1337 | __u32 qlen = 0; |
1da177e4 | 1338 | |
5dd431b6 PA |
1339 | if (!cl->level && cl->leaf.q) |
1340 | qdisc_qstats_qlen_backlog(cl->leaf.q, &qlen, &qs.backlog); | |
1341 | ||
0564bf0a KK |
1342 | cl->xstats.tokens = clamp_t(s64, PSCHED_NS2TICKS(cl->tokens), |
1343 | INT_MIN, INT_MAX); | |
1344 | cl->xstats.ctokens = clamp_t(s64, PSCHED_NS2TICKS(cl->ctokens), | |
1345 | INT_MIN, INT_MAX); | |
1da177e4 | 1346 | |
83271586 MM |
1347 | if (q->offload) { |
1348 | if (!cl->level) { | |
1349 | if (cl->leaf.q) | |
1350 | cl->bstats = cl->leaf.q->bstats; | |
1351 | else | |
50dc9a85 | 1352 | gnet_stats_basic_sync_init(&cl->bstats); |
f56940da | 1353 | _bstats_update(&cl->bstats, |
50dc9a85 AD |
1354 | u64_stats_read(&cl->bstats_bias.bytes), |
1355 | u64_stats_read(&cl->bstats_bias.packets)); | |
83271586 MM |
1356 | } else { |
1357 | htb_offload_aggregate_stats(q, cl); | |
1358 | } | |
1359 | } | |
1360 | ||
29cbcd85 | 1361 | if (gnet_stats_copy_basic(d, NULL, &cl->bstats, true) < 0 || |
1c0d32fd | 1362 | gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 || |
338ed9b4 | 1363 | gnet_stats_copy_queue(d, NULL, &qs, qlen) < 0) |
1da177e4 LT |
1364 | return -1; |
1365 | ||
1366 | return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats)); | |
1367 | } | |
1368 | ||
d03b195b MM |
1369 | static struct netdev_queue * |
1370 | htb_select_queue(struct Qdisc *sch, struct tcmsg *tcm) | |
1371 | { | |
1372 | struct net_device *dev = qdisc_dev(sch); | |
1373 | struct tc_htb_qopt_offload offload_opt; | |
93bde210 | 1374 | struct htb_sched *q = qdisc_priv(sch); |
d03b195b MM |
1375 | int err; |
1376 | ||
93bde210 MM |
1377 | if (!q->offload) |
1378 | return sch->dev_queue; | |
1379 | ||
d03b195b MM |
1380 | offload_opt = (struct tc_htb_qopt_offload) { |
1381 | .command = TC_HTB_LEAF_QUERY_QUEUE, | |
1382 | .classid = TC_H_MIN(tcm->tcm_parent), | |
1383 | }; | |
1384 | err = htb_offload(dev, &offload_opt); | |
1385 | if (err || offload_opt.qid >= dev->num_tx_queues) | |
1386 | return NULL; | |
1387 | return netdev_get_tx_queue(dev, offload_opt.qid); | |
1388 | } | |
1389 | ||
1390 | static struct Qdisc * | |
1391 | htb_graft_helper(struct netdev_queue *dev_queue, struct Qdisc *new_q) | |
1392 | { | |
1393 | struct net_device *dev = dev_queue->dev; | |
1394 | struct Qdisc *old_q; | |
1395 | ||
1396 | if (dev->flags & IFF_UP) | |
1397 | dev_deactivate(dev); | |
1398 | old_q = dev_graft_qdisc(dev_queue, new_q); | |
1399 | if (new_q) | |
1400 | new_q->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT; | |
1401 | if (dev->flags & IFF_UP) | |
1402 | dev_activate(dev); | |
1403 | ||
1404 | return old_q; | |
1405 | } | |
1406 | ||
ca49bfd9 MM |
1407 | static struct netdev_queue *htb_offload_get_queue(struct htb_class *cl) |
1408 | { | |
1409 | struct netdev_queue *queue; | |
1410 | ||
1411 | queue = cl->leaf.offload_queue; | |
1412 | if (!(cl->leaf.q->flags & TCQ_F_BUILTIN)) | |
1413 | WARN_ON(cl->leaf.q->dev_queue != queue); | |
1414 | ||
1415 | return queue; | |
1416 | } | |
1417 | ||
1418 | static void htb_offload_move_qdisc(struct Qdisc *sch, struct htb_class *cl_old, | |
1419 | struct htb_class *cl_new, bool destroying) | |
d03b195b MM |
1420 | { |
1421 | struct netdev_queue *queue_old, *queue_new; | |
1422 | struct net_device *dev = qdisc_dev(sch); | |
d03b195b | 1423 | |
ca49bfd9 MM |
1424 | queue_old = htb_offload_get_queue(cl_old); |
1425 | queue_new = htb_offload_get_queue(cl_new); | |
d03b195b | 1426 | |
ca49bfd9 MM |
1427 | if (!destroying) { |
1428 | struct Qdisc *qdisc; | |
d03b195b | 1429 | |
ca49bfd9 MM |
1430 | if (dev->flags & IFF_UP) |
1431 | dev_deactivate(dev); | |
1432 | qdisc = dev_graft_qdisc(queue_old, NULL); | |
1433 | WARN_ON(qdisc != cl_old->leaf.q); | |
1434 | } | |
1435 | ||
1436 | if (!(cl_old->leaf.q->flags & TCQ_F_BUILTIN)) | |
1437 | cl_old->leaf.q->dev_queue = queue_new; | |
1438 | cl_old->leaf.offload_queue = queue_new; | |
1439 | ||
1440 | if (!destroying) { | |
1441 | struct Qdisc *qdisc; | |
1442 | ||
1443 | qdisc = dev_graft_qdisc(queue_new, cl_old->leaf.q); | |
1444 | if (dev->flags & IFF_UP) | |
1445 | dev_activate(dev); | |
1446 | WARN_ON(!(qdisc->flags & TCQ_F_BUILTIN)); | |
1447 | } | |
d03b195b MM |
1448 | } |
1449 | ||
1da177e4 | 1450 | static int htb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new, |
653d6fd6 | 1451 | struct Qdisc **old, struct netlink_ext_ack *extack) |
1da177e4 | 1452 | { |
d03b195b | 1453 | struct netdev_queue *dev_queue = sch->dev_queue; |
87990467 | 1454 | struct htb_class *cl = (struct htb_class *)arg; |
d03b195b MM |
1455 | struct htb_sched *q = qdisc_priv(sch); |
1456 | struct Qdisc *old_q; | |
1da177e4 | 1457 | |
5b9a9ccf PM |
1458 | if (cl->level) |
1459 | return -EINVAL; | |
d03b195b | 1460 | |
ca49bfd9 MM |
1461 | if (q->offload) |
1462 | dev_queue = htb_offload_get_queue(cl); | |
d03b195b MM |
1463 | |
1464 | if (!new) { | |
1465 | new = qdisc_create_dflt(dev_queue, &pfifo_qdisc_ops, | |
1466 | cl->common.classid, extack); | |
1467 | if (!new) | |
1468 | return -ENOBUFS; | |
1469 | } | |
1470 | ||
1471 | if (q->offload) { | |
1472 | htb_set_lockdep_class_child(new); | |
1473 | /* One ref for cl->leaf.q, the other for dev_queue->qdisc. */ | |
1474 | qdisc_refcount_inc(new); | |
1475 | old_q = htb_graft_helper(dev_queue, new); | |
1476 | } | |
5b9a9ccf | 1477 | |
11957be2 | 1478 | *old = qdisc_replace(sch, new, &cl->leaf.q); |
d03b195b MM |
1479 | |
1480 | if (q->offload) { | |
1481 | WARN_ON(old_q != *old); | |
1482 | qdisc_put(old_q); | |
1483 | } | |
1484 | ||
5b9a9ccf | 1485 | return 0; |
1da177e4 LT |
1486 | } |
1487 | ||
87990467 | 1488 | static struct Qdisc *htb_leaf(struct Qdisc *sch, unsigned long arg) |
1da177e4 | 1489 | { |
87990467 | 1490 | struct htb_class *cl = (struct htb_class *)arg; |
11957be2 | 1491 | return !cl->level ? cl->leaf.q : NULL; |
1da177e4 LT |
1492 | } |
1493 | ||
256d61b8 PM |
1494 | static void htb_qlen_notify(struct Qdisc *sch, unsigned long arg) |
1495 | { | |
1496 | struct htb_class *cl = (struct htb_class *)arg; | |
1497 | ||
95946658 | 1498 | htb_deactivate(qdisc_priv(sch), cl); |
256d61b8 PM |
1499 | } |
1500 | ||
160d5e10 JP |
1501 | static inline int htb_parent_last_child(struct htb_class *cl) |
1502 | { | |
1503 | if (!cl->parent) | |
1504 | /* the root class */ | |
1505 | return 0; | |
42077599 | 1506 | if (cl->parent->children > 1) |
160d5e10 JP |
1507 | /* not the last child */ |
1508 | return 0; | |
160d5e10 JP |
1509 | return 1; |
1510 | } | |
1511 | ||
d03b195b | 1512 | static void htb_parent_to_leaf(struct Qdisc *sch, struct htb_class *cl, |
3ba08b00 | 1513 | struct Qdisc *new_q) |
160d5e10 | 1514 | { |
d03b195b | 1515 | struct htb_sched *q = qdisc_priv(sch); |
160d5e10 JP |
1516 | struct htb_class *parent = cl->parent; |
1517 | ||
11957be2 | 1518 | WARN_ON(cl->level || !cl->leaf.q || cl->prio_activity); |
160d5e10 | 1519 | |
3ba08b00 | 1520 | if (parent->cmode != HTB_CAN_SEND) |
c9364636 ED |
1521 | htb_safe_rb_erase(&parent->pq_node, |
1522 | &q->hlevel[parent->level].wait_pq); | |
3ba08b00 | 1523 | |
160d5e10 | 1524 | parent->level = 0; |
11957be2 CW |
1525 | memset(&parent->inner, 0, sizeof(parent->inner)); |
1526 | parent->leaf.q = new_q ? new_q : &noop_qdisc; | |
160d5e10 JP |
1527 | parent->tokens = parent->buffer; |
1528 | parent->ctokens = parent->cbuffer; | |
d2de875c | 1529 | parent->t_c = ktime_get_ns(); |
160d5e10 | 1530 | parent->cmode = HTB_CAN_SEND; |
ca49bfd9 MM |
1531 | if (q->offload) |
1532 | parent->leaf.offload_queue = cl->leaf.offload_queue; | |
160d5e10 JP |
1533 | } |
1534 | ||
d03b195b MM |
1535 | static void htb_parent_to_leaf_offload(struct Qdisc *sch, |
1536 | struct netdev_queue *dev_queue, | |
1537 | struct Qdisc *new_q) | |
1538 | { | |
1539 | struct Qdisc *old_q; | |
1540 | ||
1541 | /* One ref for cl->leaf.q, the other for dev_queue->qdisc. */ | |
944d671d YW |
1542 | if (new_q) |
1543 | qdisc_refcount_inc(new_q); | |
d03b195b MM |
1544 | old_q = htb_graft_helper(dev_queue, new_q); |
1545 | WARN_ON(!(old_q->flags & TCQ_F_BUILTIN)); | |
1546 | } | |
1547 | ||
1548 | static int htb_destroy_class_offload(struct Qdisc *sch, struct htb_class *cl, | |
1549 | bool last_child, bool destroying, | |
1550 | struct netlink_ext_ack *extack) | |
1551 | { | |
1552 | struct tc_htb_qopt_offload offload_opt; | |
ca49bfd9 | 1553 | struct netdev_queue *dev_queue; |
d03b195b | 1554 | struct Qdisc *q = cl->leaf.q; |
a22b7388 | 1555 | struct Qdisc *old; |
d03b195b MM |
1556 | int err; |
1557 | ||
1558 | if (cl->level) | |
1559 | return -EINVAL; | |
1560 | ||
1561 | WARN_ON(!q); | |
ca49bfd9 | 1562 | dev_queue = htb_offload_get_queue(cl); |
a22b7388 RR |
1563 | /* When destroying, caller qdisc_graft grafts the new qdisc and invokes |
1564 | * qdisc_put for the qdisc being destroyed. htb_destroy_class_offload | |
1565 | * does not need to graft or qdisc_put the qdisc being destroyed. | |
1566 | */ | |
1567 | if (!destroying) { | |
1568 | old = htb_graft_helper(dev_queue, NULL); | |
1569 | /* Last qdisc grafted should be the same as cl->leaf.q when | |
1570 | * calling htb_delete. | |
d03b195b | 1571 | */ |
d03b195b | 1572 | WARN_ON(old != q); |
a22b7388 | 1573 | } |
d03b195b | 1574 | |
83271586 | 1575 | if (cl->parent) { |
f56940da | 1576 | _bstats_update(&cl->parent->bstats_bias, |
50dc9a85 AD |
1577 | u64_stats_read(&q->bstats.bytes), |
1578 | u64_stats_read(&q->bstats.packets)); | |
83271586 MM |
1579 | } |
1580 | ||
d03b195b MM |
1581 | offload_opt = (struct tc_htb_qopt_offload) { |
1582 | .command = !last_child ? TC_HTB_LEAF_DEL : | |
1583 | destroying ? TC_HTB_LEAF_DEL_LAST_FORCE : | |
1584 | TC_HTB_LEAF_DEL_LAST, | |
1585 | .classid = cl->common.classid, | |
1586 | .extack = extack, | |
1587 | }; | |
1588 | err = htb_offload(qdisc_dev(sch), &offload_opt); | |
1589 | ||
a22b7388 RR |
1590 | if (!destroying) { |
1591 | if (!err) | |
1592 | qdisc_put(old); | |
1593 | else | |
1594 | htb_graft_helper(dev_queue, old); | |
1595 | } | |
d03b195b MM |
1596 | |
1597 | if (last_child) | |
1598 | return err; | |
1599 | ||
ca49bfd9 MM |
1600 | if (!err && offload_opt.classid != TC_H_MIN(cl->common.classid)) { |
1601 | u32 classid = TC_H_MAJ(sch->handle) | | |
1602 | TC_H_MIN(offload_opt.classid); | |
1603 | struct htb_class *moved_cl = htb_find(classid, sch); | |
1604 | ||
1605 | htb_offload_move_qdisc(sch, moved_cl, cl, destroying); | |
d03b195b MM |
1606 | } |
1607 | ||
1608 | return err; | |
1609 | } | |
1610 | ||
87990467 | 1611 | static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl) |
1da177e4 | 1612 | { |
1da177e4 | 1613 | if (!cl->level) { |
11957be2 | 1614 | WARN_ON(!cl->leaf.q); |
86bd446b | 1615 | qdisc_put(cl->leaf.q); |
1da177e4 | 1616 | } |
1c0d32fd | 1617 | gen_kill_estimator(&cl->rate_est); |
6529eaba | 1618 | tcf_block_put(cl->block); |
1da177e4 LT |
1619 | kfree(cl); |
1620 | } | |
1621 | ||
87990467 | 1622 | static void htb_destroy(struct Qdisc *sch) |
1da177e4 | 1623 | { |
d03b195b MM |
1624 | struct net_device *dev = qdisc_dev(sch); |
1625 | struct tc_htb_qopt_offload offload_opt; | |
1da177e4 | 1626 | struct htb_sched *q = qdisc_priv(sch); |
b67bfe0d | 1627 | struct hlist_node *next; |
d03b195b | 1628 | bool nonempty, changed; |
fbd8f137 PM |
1629 | struct htb_class *cl; |
1630 | unsigned int i; | |
1da177e4 | 1631 | |
1224736d | 1632 | cancel_work_sync(&q->work); |
fb983d45 | 1633 | qdisc_watchdog_cancel(&q->watchdog); |
1da177e4 | 1634 | /* This line used to be after htb_destroy_class call below |
cc7ec456 ED |
1635 | * and surprisingly it worked in 2.4. But it must precede it |
1636 | * because filter need its target class alive to be able to call | |
1637 | * unbind_filter on it (without Oops). | |
1638 | */ | |
6529eaba | 1639 | tcf_block_put(q->block); |
87990467 | 1640 | |
f4c1f3e0 | 1641 | for (i = 0; i < q->clhash.hashsize; i++) { |
89890422 | 1642 | hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) { |
6529eaba | 1643 | tcf_block_put(cl->block); |
89890422 KK |
1644 | cl->block = NULL; |
1645 | } | |
fbd8f137 | 1646 | } |
d03b195b MM |
1647 | |
1648 | do { | |
1649 | nonempty = false; | |
1650 | changed = false; | |
1651 | for (i = 0; i < q->clhash.hashsize; i++) { | |
1652 | hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i], | |
1653 | common.hnode) { | |
1654 | bool last_child; | |
1655 | ||
1656 | if (!q->offload) { | |
1657 | htb_destroy_class(sch, cl); | |
1658 | continue; | |
1659 | } | |
1660 | ||
1661 | nonempty = true; | |
1662 | ||
1663 | if (cl->level) | |
1664 | continue; | |
1665 | ||
1666 | changed = true; | |
1667 | ||
1668 | last_child = htb_parent_last_child(cl); | |
1669 | htb_destroy_class_offload(sch, cl, last_child, | |
1670 | true, NULL); | |
1671 | qdisc_class_hash_remove(&q->clhash, | |
1672 | &cl->common); | |
1673 | if (cl->parent) | |
1674 | cl->parent->children--; | |
1675 | if (last_child) | |
1676 | htb_parent_to_leaf(sch, cl, NULL); | |
1677 | htb_destroy_class(sch, cl); | |
1678 | } | |
1679 | } | |
1680 | } while (changed); | |
1681 | WARN_ON(nonempty); | |
1682 | ||
f4c1f3e0 | 1683 | qdisc_class_hash_destroy(&q->clhash); |
a5a9f534 | 1684 | __qdisc_reset_queue(&q->direct_queue); |
d03b195b | 1685 | |
d59f4e1d ZS |
1686 | if (q->offload) { |
1687 | offload_opt = (struct tc_htb_qopt_offload) { | |
1688 | .command = TC_HTB_DESTROY, | |
1689 | }; | |
1690 | htb_offload(dev, &offload_opt); | |
1691 | } | |
d03b195b MM |
1692 | |
1693 | if (!q->direct_qdiscs) | |
1694 | return; | |
1695 | for (i = 0; i < q->num_direct_qdiscs && q->direct_qdiscs[i]; i++) | |
1696 | qdisc_put(q->direct_qdiscs[i]); | |
1697 | kfree(q->direct_qdiscs); | |
1da177e4 LT |
1698 | } |
1699 | ||
4dd78a73 MM |
1700 | static int htb_delete(struct Qdisc *sch, unsigned long arg, |
1701 | struct netlink_ext_ack *extack) | |
1da177e4 LT |
1702 | { |
1703 | struct htb_sched *q = qdisc_priv(sch); | |
87990467 | 1704 | struct htb_class *cl = (struct htb_class *)arg; |
160d5e10 JP |
1705 | struct Qdisc *new_q = NULL; |
1706 | int last_child = 0; | |
d03b195b | 1707 | int err; |
1da177e4 | 1708 | |
a071d272 YY |
1709 | /* TODO: why don't allow to delete subtree ? references ? does |
1710 | * tc subsys guarantee us that in htb_destroy it holds no class | |
1711 | * refs so that we can remove children safely there ? | |
1712 | */ | |
42077599 | 1713 | if (cl->children || cl->filter_cnt) |
1da177e4 | 1714 | return -EBUSY; |
87990467 | 1715 | |
d03b195b MM |
1716 | if (!cl->level && htb_parent_last_child(cl)) |
1717 | last_child = 1; | |
1718 | ||
1719 | if (q->offload) { | |
1720 | err = htb_destroy_class_offload(sch, cl, last_child, false, | |
1721 | extack); | |
1722 | if (err) | |
1723 | return err; | |
1724 | } | |
1725 | ||
1726 | if (last_child) { | |
ca49bfd9 MM |
1727 | struct netdev_queue *dev_queue = sch->dev_queue; |
1728 | ||
1729 | if (q->offload) | |
1730 | dev_queue = htb_offload_get_queue(cl); | |
d03b195b | 1731 | |
d03b195b | 1732 | new_q = qdisc_create_dflt(dev_queue, &pfifo_qdisc_ops, |
a38a9882 AA |
1733 | cl->parent->common.classid, |
1734 | NULL); | |
d03b195b | 1735 | if (q->offload) { |
944d671d | 1736 | if (new_q) |
d03b195b | 1737 | htb_set_lockdep_class_child(new_q); |
944d671d | 1738 | htb_parent_to_leaf_offload(sch, dev_queue, new_q); |
d03b195b | 1739 | } |
160d5e10 JP |
1740 | } |
1741 | ||
1da177e4 | 1742 | sch_tree_lock(sch); |
87990467 | 1743 | |
e5f0e8f8 PA |
1744 | if (!cl->level) |
1745 | qdisc_purge_queue(cl->leaf.q); | |
814a175e | 1746 | |
f4c1f3e0 PM |
1747 | /* delete from hash and active; remainder in destroy_class */ |
1748 | qdisc_class_hash_remove(&q->clhash, &cl->common); | |
26b284de JP |
1749 | if (cl->parent) |
1750 | cl->parent->children--; | |
c38c83cb | 1751 | |
1da177e4 | 1752 | if (cl->prio_activity) |
87990467 | 1753 | htb_deactivate(q, cl); |
1da177e4 | 1754 | |
fbd8f137 | 1755 | if (cl->cmode != HTB_CAN_SEND) |
c9364636 ED |
1756 | htb_safe_rb_erase(&cl->pq_node, |
1757 | &q->hlevel[cl->level].wait_pq); | |
fbd8f137 | 1758 | |
160d5e10 | 1759 | if (last_child) |
d03b195b | 1760 | htb_parent_to_leaf(sch, cl, new_q); |
160d5e10 | 1761 | |
1da177e4 | 1762 | sch_tree_unlock(sch); |
1da177e4 | 1763 | |
143976ce WC |
1764 | htb_destroy_class(sch, cl); |
1765 | return 0; | |
1da177e4 LT |
1766 | } |
1767 | ||
87990467 | 1768 | static int htb_change_class(struct Qdisc *sch, u32 classid, |
1e90474c | 1769 | u32 parentid, struct nlattr **tca, |
793d81d6 | 1770 | unsigned long *arg, struct netlink_ext_ack *extack) |
1da177e4 LT |
1771 | { |
1772 | int err = -EINVAL; | |
1773 | struct htb_sched *q = qdisc_priv(sch); | |
87990467 | 1774 | struct htb_class *cl = (struct htb_class *)*arg, *parent; |
d03b195b | 1775 | struct tc_htb_qopt_offload offload_opt; |
1e90474c | 1776 | struct nlattr *opt = tca[TCA_OPTIONS]; |
6906f4ed | 1777 | struct nlattr *tb[TCA_HTB_MAX + 1]; |
4ce70b4a | 1778 | struct Qdisc *parent_qdisc = NULL; |
d03b195b | 1779 | struct netdev_queue *dev_queue; |
1da177e4 | 1780 | struct tc_htb_opt *hopt; |
df62cdf3 | 1781 | u64 rate64, ceil64; |
da01ec4e | 1782 | int warn = 0; |
1da177e4 LT |
1783 | |
1784 | /* extract all subattrs from opt attr */ | |
cee63723 PM |
1785 | if (!opt) |
1786 | goto failure; | |
1787 | ||
8cb08174 JB |
1788 | err = nla_parse_nested_deprecated(tb, TCA_HTB_MAX, opt, htb_policy, |
1789 | NULL); | |
cee63723 PM |
1790 | if (err < 0) |
1791 | goto failure; | |
1792 | ||
1793 | err = -EINVAL; | |
27a3421e | 1794 | if (tb[TCA_HTB_PARMS] == NULL) |
1da177e4 | 1795 | goto failure; |
1da177e4 | 1796 | |
87990467 SH |
1797 | parent = parentid == TC_H_ROOT ? NULL : htb_find(parentid, sch); |
1798 | ||
1e90474c | 1799 | hopt = nla_data(tb[TCA_HTB_PARMS]); |
196d97f6 | 1800 | if (!hopt->rate.rate || !hopt->ceil.rate) |
87990467 | 1801 | goto failure; |
1da177e4 | 1802 | |
429c3be8 MM |
1803 | if (q->offload) { |
1804 | /* Options not supported by the offload. */ | |
1805 | if (hopt->rate.overhead || hopt->ceil.overhead) { | |
1806 | NL_SET_ERR_MSG(extack, "HTB offload doesn't support the overhead parameter"); | |
1807 | goto failure; | |
1808 | } | |
1809 | if (hopt->rate.mpu || hopt->ceil.mpu) { | |
1810 | NL_SET_ERR_MSG(extack, "HTB offload doesn't support the mpu parameter"); | |
1811 | goto failure; | |
1812 | } | |
1813 | if (hopt->quantum) { | |
1814 | NL_SET_ERR_MSG(extack, "HTB offload doesn't support the quantum parameter"); | |
1815 | goto failure; | |
1816 | } | |
1817 | if (hopt->prio) { | |
1818 | NL_SET_ERR_MSG(extack, "HTB offload doesn't support the prio parameter"); | |
1819 | goto failure; | |
1820 | } | |
1821 | } | |
1822 | ||
8a8e3d84 | 1823 | /* Keeping backward compatible with rate_table based iproute2 tc */ |
6b1dd856 | 1824 | if (hopt->rate.linklayer == TC_LINKLAYER_UNAWARE) |
e9bc3fa2 AA |
1825 | qdisc_put_rtab(qdisc_get_rtab(&hopt->rate, tb[TCA_HTB_RTAB], |
1826 | NULL)); | |
6b1dd856 YY |
1827 | |
1828 | if (hopt->ceil.linklayer == TC_LINKLAYER_UNAWARE) | |
e9bc3fa2 AA |
1829 | qdisc_put_rtab(qdisc_get_rtab(&hopt->ceil, tb[TCA_HTB_CTAB], |
1830 | NULL)); | |
8a8e3d84 | 1831 | |
d03b195b MM |
1832 | rate64 = tb[TCA_HTB_RATE64] ? nla_get_u64(tb[TCA_HTB_RATE64]) : 0; |
1833 | ceil64 = tb[TCA_HTB_CEIL64] ? nla_get_u64(tb[TCA_HTB_CEIL64]) : 0; | |
1834 | ||
87990467 | 1835 | if (!cl) { /* new class */ |
d03b195b MM |
1836 | struct net_device *dev = qdisc_dev(sch); |
1837 | struct Qdisc *new_q, *old_q; | |
3696f625 | 1838 | int prio; |
ee39e10c | 1839 | struct { |
1e90474c | 1840 | struct nlattr nla; |
ee39e10c PM |
1841 | struct gnet_estimator opt; |
1842 | } est = { | |
1e90474c PM |
1843 | .nla = { |
1844 | .nla_len = nla_attr_size(sizeof(est.opt)), | |
1845 | .nla_type = TCA_RATE, | |
ee39e10c PM |
1846 | }, |
1847 | .opt = { | |
1848 | /* 4s interval, 16s averaging constant */ | |
1849 | .interval = 2, | |
1850 | .ewma_log = 2, | |
1851 | }, | |
1852 | }; | |
3696f625 | 1853 | |
1da177e4 | 1854 | /* check for valid classid */ |
f64f9e71 JP |
1855 | if (!classid || TC_H_MAJ(classid ^ sch->handle) || |
1856 | htb_find(classid, sch)) | |
1da177e4 LT |
1857 | goto failure; |
1858 | ||
1859 | /* check maximal depth */ | |
1860 | if (parent && parent->parent && parent->parent->level < 2) { | |
cc7ec456 | 1861 | pr_err("htb: tree is too deep\n"); |
1da177e4 LT |
1862 | goto failure; |
1863 | } | |
1864 | err = -ENOBUFS; | |
cc7ec456 ED |
1865 | cl = kzalloc(sizeof(*cl), GFP_KERNEL); |
1866 | if (!cl) | |
1da177e4 | 1867 | goto failure; |
87990467 | 1868 | |
50dc9a85 AD |
1869 | gnet_stats_basic_sync_init(&cl->bstats); |
1870 | gnet_stats_basic_sync_init(&cl->bstats_bias); | |
67c9e627 | 1871 | |
8d1a77f9 | 1872 | err = tcf_block_get(&cl->block, &cl->filter_list, sch, extack); |
6529eaba JP |
1873 | if (err) { |
1874 | kfree(cl); | |
1875 | goto failure; | |
1876 | } | |
64153ce0 | 1877 | if (htb_rate_est || tca[TCA_RATE]) { |
22e0f8b9 JF |
1878 | err = gen_new_estimator(&cl->bstats, NULL, |
1879 | &cl->rate_est, | |
edb09eb1 | 1880 | NULL, |
29cbcd85 | 1881 | true, |
64153ce0 | 1882 | tca[TCA_RATE] ? : &est.nla); |
d03b195b MM |
1883 | if (err) |
1884 | goto err_block_put; | |
71bcb09a SH |
1885 | } |
1886 | ||
42077599 | 1887 | cl->children = 0; |
3696f625 SH |
1888 | RB_CLEAR_NODE(&cl->pq_node); |
1889 | ||
1890 | for (prio = 0; prio < TC_HTB_NUMPRIO; prio++) | |
1891 | RB_CLEAR_NODE(&cl->node[prio]); | |
1da177e4 | 1892 | |
d03b195b MM |
1893 | cl->common.classid = classid; |
1894 | ||
1895 | /* Make sure nothing interrupts us in between of two | |
1896 | * ndo_setup_tc calls. | |
1897 | */ | |
1898 | ASSERT_RTNL(); | |
1899 | ||
1da177e4 | 1900 | /* create leaf qdisc early because it uses kmalloc(GFP_KERNEL) |
cc7ec456 ED |
1901 | * so that can't be used inside of sch_tree_lock |
1902 | * -- thanks to Karlis Peisenieks | |
1903 | */ | |
d03b195b MM |
1904 | if (!q->offload) { |
1905 | dev_queue = sch->dev_queue; | |
1906 | } else if (!(parent && !parent->level)) { | |
1907 | /* Assign a dev_queue to this classid. */ | |
1908 | offload_opt = (struct tc_htb_qopt_offload) { | |
1909 | .command = TC_HTB_LEAF_ALLOC_QUEUE, | |
1910 | .classid = cl->common.classid, | |
1911 | .parent_classid = parent ? | |
1912 | TC_H_MIN(parent->common.classid) : | |
1913 | TC_HTB_CLASSID_ROOT, | |
1914 | .rate = max_t(u64, hopt->rate.rate, rate64), | |
1915 | .ceil = max_t(u64, hopt->ceil.rate, ceil64), | |
1916 | .extack = extack, | |
1917 | }; | |
1918 | err = htb_offload(dev, &offload_opt); | |
1919 | if (err) { | |
1920 | pr_err("htb: TC_HTB_LEAF_ALLOC_QUEUE failed with err = %d\n", | |
1921 | err); | |
1922 | goto err_kill_estimator; | |
1923 | } | |
1924 | dev_queue = netdev_get_tx_queue(dev, offload_opt.qid); | |
1925 | } else { /* First child. */ | |
ca49bfd9 | 1926 | dev_queue = htb_offload_get_queue(parent); |
d03b195b MM |
1927 | old_q = htb_graft_helper(dev_queue, NULL); |
1928 | WARN_ON(old_q != parent->leaf.q); | |
1929 | offload_opt = (struct tc_htb_qopt_offload) { | |
1930 | .command = TC_HTB_LEAF_TO_INNER, | |
1931 | .classid = cl->common.classid, | |
1932 | .parent_classid = | |
1933 | TC_H_MIN(parent->common.classid), | |
1934 | .rate = max_t(u64, hopt->rate.rate, rate64), | |
1935 | .ceil = max_t(u64, hopt->ceil.rate, ceil64), | |
1936 | .extack = extack, | |
1937 | }; | |
1938 | err = htb_offload(dev, &offload_opt); | |
1939 | if (err) { | |
1940 | pr_err("htb: TC_HTB_LEAF_TO_INNER failed with err = %d\n", | |
1941 | err); | |
1942 | htb_graft_helper(dev_queue, old_q); | |
1943 | goto err_kill_estimator; | |
1944 | } | |
f56940da | 1945 | _bstats_update(&parent->bstats_bias, |
50dc9a85 AD |
1946 | u64_stats_read(&old_q->bstats.bytes), |
1947 | u64_stats_read(&old_q->bstats.packets)); | |
d03b195b MM |
1948 | qdisc_put(old_q); |
1949 | } | |
1950 | new_q = qdisc_create_dflt(dev_queue, &pfifo_qdisc_ops, | |
a38a9882 | 1951 | classid, NULL); |
d03b195b MM |
1952 | if (q->offload) { |
1953 | if (new_q) { | |
1954 | htb_set_lockdep_class_child(new_q); | |
1955 | /* One ref for cl->leaf.q, the other for | |
1956 | * dev_queue->qdisc. | |
1957 | */ | |
1958 | qdisc_refcount_inc(new_q); | |
1959 | } | |
1960 | old_q = htb_graft_helper(dev_queue, new_q); | |
1961 | /* No qdisc_put needed. */ | |
1962 | WARN_ON(!(old_q->flags & TCQ_F_BUILTIN)); | |
1963 | } | |
1da177e4 LT |
1964 | sch_tree_lock(sch); |
1965 | if (parent && !parent->level) { | |
1966 | /* turn parent into inner node */ | |
e5f0e8f8 | 1967 | qdisc_purge_queue(parent->leaf.q); |
4ce70b4a | 1968 | parent_qdisc = parent->leaf.q; |
87990467 SH |
1969 | if (parent->prio_activity) |
1970 | htb_deactivate(q, parent); | |
1da177e4 LT |
1971 | |
1972 | /* remove from evt list because of level change */ | |
1973 | if (parent->cmode != HTB_CAN_SEND) { | |
c9364636 | 1974 | htb_safe_rb_erase(&parent->pq_node, &q->hlevel[0].wait_pq); |
1da177e4 LT |
1975 | parent->cmode = HTB_CAN_SEND; |
1976 | } | |
1977 | parent->level = (parent->parent ? parent->parent->level | |
87990467 | 1978 | : TC_HTB_MAXDEPTH) - 1; |
11957be2 | 1979 | memset(&parent->inner, 0, sizeof(parent->inner)); |
1da177e4 | 1980 | } |
d03b195b | 1981 | |
1da177e4 | 1982 | /* leaf (we) needs elementary qdisc */ |
11957be2 | 1983 | cl->leaf.q = new_q ? new_q : &noop_qdisc; |
ca49bfd9 MM |
1984 | if (q->offload) |
1985 | cl->leaf.offload_queue = dev_queue; | |
1da177e4 | 1986 | |
87990467 | 1987 | cl->parent = parent; |
1da177e4 LT |
1988 | |
1989 | /* set class to be in HTB_CAN_SEND state */ | |
b9a7afde JP |
1990 | cl->tokens = PSCHED_TICKS2NS(hopt->buffer); |
1991 | cl->ctokens = PSCHED_TICKS2NS(hopt->cbuffer); | |
5343a7f8 | 1992 | cl->mbuffer = 60ULL * NSEC_PER_SEC; /* 1min */ |
d2de875c | 1993 | cl->t_c = ktime_get_ns(); |
1da177e4 LT |
1994 | cl->cmode = HTB_CAN_SEND; |
1995 | ||
1996 | /* attach to the hash list and parent's family */ | |
f4c1f3e0 | 1997 | qdisc_class_hash_insert(&q->clhash, &cl->common); |
42077599 PM |
1998 | if (parent) |
1999 | parent->children++; | |
11957be2 CW |
2000 | if (cl->leaf.q != &noop_qdisc) |
2001 | qdisc_hash_add(cl->leaf.q, true); | |
ee39e10c | 2002 | } else { |
71bcb09a | 2003 | if (tca[TCA_RATE]) { |
22e0f8b9 JF |
2004 | err = gen_replace_estimator(&cl->bstats, NULL, |
2005 | &cl->rate_est, | |
edb09eb1 | 2006 | NULL, |
29cbcd85 | 2007 | true, |
71bcb09a SH |
2008 | tca[TCA_RATE]); |
2009 | if (err) | |
2010 | return err; | |
2011 | } | |
1da177e4 | 2012 | |
d03b195b MM |
2013 | if (q->offload) { |
2014 | struct net_device *dev = qdisc_dev(sch); | |
2015 | ||
2016 | offload_opt = (struct tc_htb_qopt_offload) { | |
2017 | .command = TC_HTB_NODE_MODIFY, | |
2018 | .classid = cl->common.classid, | |
2019 | .rate = max_t(u64, hopt->rate.rate, rate64), | |
2020 | .ceil = max_t(u64, hopt->ceil.rate, ceil64), | |
2021 | .extack = extack, | |
2022 | }; | |
2023 | err = htb_offload(dev, &offload_opt); | |
2024 | if (err) | |
2025 | /* Estimator was replaced, and rollback may fail | |
2026 | * as well, so we don't try to recover it, and | |
2027 | * the estimator won't work property with the | |
2028 | * offload anyway, because bstats are updated | |
2029 | * only when the stats are queried. | |
2030 | */ | |
2031 | return err; | |
2032 | } | |
1598f7cb | 2033 | |
d03b195b MM |
2034 | sch_tree_lock(sch); |
2035 | } | |
1598f7cb YY |
2036 | |
2037 | psched_ratecfg_precompute(&cl->rate, &hopt->rate, rate64); | |
2038 | psched_ratecfg_precompute(&cl->ceil, &hopt->ceil, ceil64); | |
2039 | ||
1da177e4 | 2040 | /* it used to be a nasty bug here, we have to check that node |
11957be2 | 2041 | * is really leaf before changing cl->leaf ! |
cc7ec456 | 2042 | */ |
1da177e4 | 2043 | if (!cl->level) { |
1598f7cb YY |
2044 | u64 quantum = cl->rate.rate_bytes_ps; |
2045 | ||
2046 | do_div(quantum, q->rate2quantum); | |
2047 | cl->quantum = min_t(u64, quantum, INT_MAX); | |
2048 | ||
c19f7a34 | 2049 | if (!hopt->quantum && cl->quantum < 1000) { |
da01ec4e | 2050 | warn = -1; |
c19f7a34 | 2051 | cl->quantum = 1000; |
1da177e4 | 2052 | } |
c19f7a34 | 2053 | if (!hopt->quantum && cl->quantum > 200000) { |
da01ec4e | 2054 | warn = 1; |
c19f7a34 | 2055 | cl->quantum = 200000; |
1da177e4 LT |
2056 | } |
2057 | if (hopt->quantum) | |
c19f7a34 JP |
2058 | cl->quantum = hopt->quantum; |
2059 | if ((cl->prio = hopt->prio) >= TC_HTB_NUMPRIO) | |
2060 | cl->prio = TC_HTB_NUMPRIO - 1; | |
1da177e4 LT |
2061 | } |
2062 | ||
324f5aa5 | 2063 | cl->buffer = PSCHED_TICKS2NS(hopt->buffer); |
f3ad857e | 2064 | cl->cbuffer = PSCHED_TICKS2NS(hopt->cbuffer); |
56b765b7 | 2065 | |
1da177e4 | 2066 | sch_tree_unlock(sch); |
4ce70b4a | 2067 | qdisc_put(parent_qdisc); |
1da177e4 | 2068 | |
da01ec4e LR |
2069 | if (warn) |
2070 | pr_warn("HTB: quantum of class %X is %s. Consider r2q change.\n", | |
2071 | cl->common.classid, (warn == -1 ? "small" : "big")); | |
2072 | ||
f4c1f3e0 PM |
2073 | qdisc_class_hash_grow(sch, &q->clhash); |
2074 | ||
1da177e4 LT |
2075 | *arg = (unsigned long)cl; |
2076 | return 0; | |
2077 | ||
d03b195b MM |
2078 | err_kill_estimator: |
2079 | gen_kill_estimator(&cl->rate_est); | |
2080 | err_block_put: | |
2081 | tcf_block_put(cl->block); | |
2082 | kfree(cl); | |
1da177e4 | 2083 | failure: |
1da177e4 LT |
2084 | return err; |
2085 | } | |
2086 | ||
cbaacc4e AA |
2087 | static struct tcf_block *htb_tcf_block(struct Qdisc *sch, unsigned long arg, |
2088 | struct netlink_ext_ack *extack) | |
1da177e4 LT |
2089 | { |
2090 | struct htb_sched *q = qdisc_priv(sch); | |
2091 | struct htb_class *cl = (struct htb_class *)arg; | |
3bf72957 | 2092 | |
6529eaba | 2093 | return cl ? cl->block : q->block; |
1da177e4 LT |
2094 | } |
2095 | ||
2096 | static unsigned long htb_bind_filter(struct Qdisc *sch, unsigned long parent, | |
87990467 | 2097 | u32 classid) |
1da177e4 | 2098 | { |
87990467 | 2099 | struct htb_class *cl = htb_find(classid, sch); |
3bf72957 | 2100 | |
1da177e4 | 2101 | /*if (cl && !cl->level) return 0; |
cc7ec456 ED |
2102 | * The line above used to be there to prevent attaching filters to |
2103 | * leaves. But at least tc_index filter uses this just to get class | |
2104 | * for other reasons so that we have to allow for it. | |
2105 | * ---- | |
2106 | * 19.6.2002 As Werner explained it is ok - bind filter is just | |
2107 | * another way to "lock" the class - unlike "get" this lock can | |
2108 | * be broken by class during destroy IIUC. | |
1da177e4 | 2109 | */ |
87990467 SH |
2110 | if (cl) |
2111 | cl->filter_cnt++; | |
1da177e4 LT |
2112 | return (unsigned long)cl; |
2113 | } | |
2114 | ||
2115 | static void htb_unbind_filter(struct Qdisc *sch, unsigned long arg) | |
2116 | { | |
1da177e4 | 2117 | struct htb_class *cl = (struct htb_class *)arg; |
3bf72957 | 2118 | |
87990467 SH |
2119 | if (cl) |
2120 | cl->filter_cnt--; | |
1da177e4 LT |
2121 | } |
2122 | ||
2123 | static void htb_walk(struct Qdisc *sch, struct qdisc_walker *arg) | |
2124 | { | |
2125 | struct htb_sched *q = qdisc_priv(sch); | |
f4c1f3e0 | 2126 | struct htb_class *cl; |
f4c1f3e0 | 2127 | unsigned int i; |
1da177e4 LT |
2128 | |
2129 | if (arg->stop) | |
2130 | return; | |
2131 | ||
f4c1f3e0 | 2132 | for (i = 0; i < q->clhash.hashsize; i++) { |
b67bfe0d | 2133 | hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) { |
e046fa89 | 2134 | if (!tc_qdisc_stats_dump(sch, (unsigned long)cl, arg)) |
1da177e4 | 2135 | return; |
1da177e4 LT |
2136 | } |
2137 | } | |
2138 | } | |
2139 | ||
20fea08b | 2140 | static const struct Qdisc_class_ops htb_class_ops = { |
d03b195b | 2141 | .select_queue = htb_select_queue, |
1da177e4 LT |
2142 | .graft = htb_graft, |
2143 | .leaf = htb_leaf, | |
256d61b8 | 2144 | .qlen_notify = htb_qlen_notify, |
143976ce | 2145 | .find = htb_search, |
1da177e4 LT |
2146 | .change = htb_change_class, |
2147 | .delete = htb_delete, | |
2148 | .walk = htb_walk, | |
6529eaba | 2149 | .tcf_block = htb_tcf_block, |
1da177e4 LT |
2150 | .bind_tcf = htb_bind_filter, |
2151 | .unbind_tcf = htb_unbind_filter, | |
2152 | .dump = htb_dump_class, | |
2153 | .dump_stats = htb_dump_class_stats, | |
2154 | }; | |
2155 | ||
20fea08b | 2156 | static struct Qdisc_ops htb_qdisc_ops __read_mostly = { |
1da177e4 LT |
2157 | .cl_ops = &htb_class_ops, |
2158 | .id = "htb", | |
2159 | .priv_size = sizeof(struct htb_sched), | |
2160 | .enqueue = htb_enqueue, | |
2161 | .dequeue = htb_dequeue, | |
77be155c | 2162 | .peek = qdisc_peek_dequeued, |
1da177e4 | 2163 | .init = htb_init, |
d03b195b | 2164 | .attach = htb_attach, |
1da177e4 LT |
2165 | .reset = htb_reset, |
2166 | .destroy = htb_destroy, | |
1da177e4 LT |
2167 | .dump = htb_dump, |
2168 | .owner = THIS_MODULE, | |
2169 | }; | |
2170 | ||
2171 | static int __init htb_module_init(void) | |
2172 | { | |
87990467 | 2173 | return register_qdisc(&htb_qdisc_ops); |
1da177e4 | 2174 | } |
87990467 | 2175 | static void __exit htb_module_exit(void) |
1da177e4 | 2176 | { |
87990467 | 2177 | unregister_qdisc(&htb_qdisc_ops); |
1da177e4 | 2178 | } |
87990467 | 2179 | |
1da177e4 LT |
2180 | module_init(htb_module_init) |
2181 | module_exit(htb_module_exit) | |
2182 | MODULE_LICENSE("GPL"); |