]>
Commit | Line | Data |
---|---|---|
0545a303 | 1 | /* |
462dbc91 | 2 | * net/sched/sch_qfq.c Quick Fair Queueing Plus Scheduler. |
0545a303 | 3 | * |
4 | * Copyright (c) 2009 Fabio Checconi, Luigi Rizzo, and Paolo Valente. | |
462dbc91 | 5 | * Copyright (c) 2012 Paolo Valente. |
0545a303 | 6 | * |
7 | * This program is free software; you can redistribute it and/or | |
8 | * modify it under the terms of the GNU General Public License | |
9 | * version 2 as published by the Free Software Foundation. | |
10 | */ | |
11 | ||
12 | #include <linux/module.h> | |
13 | #include <linux/init.h> | |
14 | #include <linux/bitops.h> | |
15 | #include <linux/errno.h> | |
16 | #include <linux/netdevice.h> | |
17 | #include <linux/pkt_sched.h> | |
18 | #include <net/sch_generic.h> | |
19 | #include <net/pkt_sched.h> | |
20 | #include <net/pkt_cls.h> | |
21 | ||
22 | ||
462dbc91 PV |
23 | /* Quick Fair Queueing Plus |
24 | ======================== | |
0545a303 | 25 | |
26 | Sources: | |
27 | ||
462dbc91 PV |
28 | [1] Paolo Valente, |
29 | "Reducing the Execution Time of Fair-Queueing Schedulers." | |
30 | http://algo.ing.unimo.it/people/paolo/agg-sched/agg-sched.pdf | |
31 | ||
32 | Sources for QFQ: | |
33 | ||
34 | [2] Fabio Checconi, Luigi Rizzo, and Paolo Valente: "QFQ: Efficient | |
0545a303 | 35 | Packet Scheduling with Tight Bandwidth Distribution Guarantees." |
36 | ||
37 | See also: | |
38 | http://retis.sssup.it/~fabio/linux/qfq/ | |
39 | */ | |
40 | ||
41 | /* | |
42 | ||
462dbc91 PV |
43 | QFQ+ divides classes into aggregates of at most MAX_AGG_CLASSES |
44 | classes. Each aggregate is timestamped with a virtual start time S | |
45 | and a virtual finish time F, and scheduled according to its | |
46 | timestamps. S and F are computed as a function of a system virtual | |
47 | time function V. The classes within each aggregate are instead | |
48 | scheduled with DRR. | |
49 | ||
50 | To speed up operations, QFQ+ divides also aggregates into a limited | |
51 | number of groups. Which group a class belongs to depends on the | |
52 | ratio between the maximum packet length for the class and the weight | |
53 | of the class. Groups have their own S and F. In the end, QFQ+ | |
54 | schedules groups, then aggregates within groups, then classes within | |
55 | aggregates. See [1] and [2] for a full description. | |
56 | ||
0545a303 | 57 | Virtual time computations. |
58 | ||
59 | S, F and V are all computed in fixed point arithmetic with | |
60 | FRAC_BITS decimal bits. | |
61 | ||
62 | QFQ_MAX_INDEX is the maximum index allowed for a group. We need | |
63 | one bit per index. | |
64 | QFQ_MAX_WSHIFT is the maximum power of two supported as a weight. | |
65 | ||
66 | The layout of the bits is as below: | |
67 | ||
68 | [ MTU_SHIFT ][ FRAC_BITS ] | |
69 | [ MAX_INDEX ][ MIN_SLOT_SHIFT ] | |
70 | ^.__grp->index = 0 | |
71 | *.__grp->slot_shift | |
72 | ||
73 | where MIN_SLOT_SHIFT is derived by difference from the others. | |
74 | ||
75 | The max group index corresponds to Lmax/w_min, where | |
76 | Lmax=1<<MTU_SHIFT, w_min = 1 . | |
77 | From this, and knowing how many groups (MAX_INDEX) we want, | |
78 | we can derive the shift corresponding to each group. | |
79 | ||
80 | Because we often need to compute | |
81 | F = S + len/w_i and V = V + len/wsum | |
82 | instead of storing w_i store the value | |
83 | inv_w = (1<<FRAC_BITS)/w_i | |
84 | so we can do F = S + len * inv_w * wsum. | |
85 | We use W_TOT in the formulas so we can easily move between | |
86 | static and adaptive weight sum. | |
87 | ||
88 | The per-scheduler-instance data contain all the data structures | |
89 | for the scheduler: bitmaps and bucket lists. | |
90 | ||
91 | */ | |
92 | ||
93 | /* | |
94 | * Maximum number of consecutive slots occupied by backlogged classes | |
95 | * inside a group. | |
96 | */ | |
97 | #define QFQ_MAX_SLOTS 32 | |
98 | ||
99 | /* | |
462dbc91 PV |
100 | * Shifts used for aggregate<->group mapping. We allow class weights that are |
101 | * in the range [1, 2^MAX_WSHIFT], and we try to map each aggregate i to the | |
0545a303 | 102 | * group with the smallest index that can support the L_i / r_i configured |
462dbc91 | 103 | * for the classes in the aggregate. |
0545a303 | 104 | * |
105 | * grp->index is the index of the group; and grp->slot_shift | |
106 | * is the shift for the corresponding (scaled) sigma_i. | |
107 | */ | |
3015f3d2 | 108 | #define QFQ_MAX_INDEX 24 |
462dbc91 | 109 | #define QFQ_MAX_WSHIFT 10 |
0545a303 | 110 | |
462dbc91 PV |
111 | #define QFQ_MAX_WEIGHT (1<<QFQ_MAX_WSHIFT) /* see qfq_slot_insert */ |
112 | #define QFQ_MAX_WSUM (64*QFQ_MAX_WEIGHT) | |
0545a303 | 113 | |
114 | #define FRAC_BITS 30 /* fixed point arithmetic */ | |
115 | #define ONE_FP (1UL << FRAC_BITS) | |
0545a303 | 116 | |
3015f3d2 | 117 | #define QFQ_MTU_SHIFT 16 /* to support TSO/GSO */ |
462dbc91 PV |
118 | #define QFQ_MIN_LMAX 512 /* see qfq_slot_insert */ |
119 | ||
120 | #define QFQ_MAX_AGG_CLASSES 8 /* max num classes per aggregate allowed */ | |
0545a303 | 121 | |
122 | /* | |
123 | * Possible group states. These values are used as indexes for the bitmaps | |
124 | * array of struct qfq_queue. | |
125 | */ | |
126 | enum qfq_state { ER, IR, EB, IB, QFQ_MAX_STATE }; | |
127 | ||
128 | struct qfq_group; | |
129 | ||
462dbc91 PV |
130 | struct qfq_aggregate; |
131 | ||
0545a303 | 132 | struct qfq_class { |
133 | struct Qdisc_class_common common; | |
134 | ||
0545a303 | 135 | unsigned int filter_cnt; |
136 | ||
137 | struct gnet_stats_basic_packed bstats; | |
138 | struct gnet_stats_queue qstats; | |
1c0d32fd | 139 | struct net_rate_estimator __rcu *rate_est; |
0545a303 | 140 | struct Qdisc *qdisc; |
462dbc91 PV |
141 | struct list_head alist; /* Link for active-classes list. */ |
142 | struct qfq_aggregate *agg; /* Parent aggregate. */ | |
143 | int deficit; /* DRR deficit counter. */ | |
144 | }; | |
0545a303 | 145 | |
462dbc91 | 146 | struct qfq_aggregate { |
0545a303 | 147 | struct hlist_node next; /* Link for the slot list. */ |
148 | u64 S, F; /* flow timestamps (exact) */ | |
149 | ||
150 | /* group we belong to. In principle we would need the index, | |
151 | * which is log_2(lmax/weight), but we never reference it | |
152 | * directly, only the group. | |
153 | */ | |
154 | struct qfq_group *grp; | |
155 | ||
156 | /* these are copied from the flowset. */ | |
462dbc91 PV |
157 | u32 class_weight; /* Weight of each class in this aggregate. */ |
158 | /* Max pkt size for the classes in this aggregate, DRR quantum. */ | |
159 | int lmax; | |
160 | ||
161 | u32 inv_w; /* ONE_FP/(sum of weights of classes in aggr.). */ | |
162 | u32 budgetmax; /* Max budget for this aggregate. */ | |
163 | u32 initial_budget, budget; /* Initial and current budget. */ | |
164 | ||
165 | int num_classes; /* Number of classes in this aggr. */ | |
166 | struct list_head active; /* DRR queue of active classes. */ | |
167 | ||
168 | struct hlist_node nonfull_next; /* See nonfull_aggs in qfq_sched. */ | |
0545a303 | 169 | }; |
170 | ||
171 | struct qfq_group { | |
172 | u64 S, F; /* group timestamps (approx). */ | |
173 | unsigned int slot_shift; /* Slot shift. */ | |
174 | unsigned int index; /* Group index. */ | |
175 | unsigned int front; /* Index of the front slot. */ | |
176 | unsigned long full_slots; /* non-empty slots */ | |
177 | ||
462dbc91 | 178 | /* Array of RR lists of active aggregates. */ |
0545a303 | 179 | struct hlist_head slots[QFQ_MAX_SLOTS]; |
180 | }; | |
181 | ||
182 | struct qfq_sched { | |
25d8c0d5 | 183 | struct tcf_proto __rcu *filter_list; |
6529eaba | 184 | struct tcf_block *block; |
0545a303 | 185 | struct Qdisc_class_hash clhash; |
186 | ||
462dbc91 PV |
187 | u64 oldV, V; /* Precise virtual times. */ |
188 | struct qfq_aggregate *in_serv_agg; /* Aggregate being served. */ | |
462dbc91 | 189 | u32 wsum; /* weight sum */ |
87f40dd6 | 190 | u32 iwsum; /* inverse weight sum */ |
0545a303 | 191 | |
192 | unsigned long bitmaps[QFQ_MAX_STATE]; /* Group bitmaps. */ | |
193 | struct qfq_group groups[QFQ_MAX_INDEX + 1]; /* The groups. */ | |
462dbc91 PV |
194 | u32 min_slot_shift; /* Index of the group-0 bit in the bitmaps. */ |
195 | ||
196 | u32 max_agg_classes; /* Max number of classes per aggr. */ | |
197 | struct hlist_head nonfull_aggs; /* Aggs with room for more classes. */ | |
0545a303 | 198 | }; |
199 | ||
462dbc91 PV |
200 | /* |
201 | * Possible reasons why the timestamps of an aggregate are updated | |
202 | * enqueue: the aggregate switches from idle to active and must scheduled | |
203 | * for service | |
204 | * requeue: the aggregate finishes its budget, so it stops being served and | |
205 | * must be rescheduled for service | |
206 | */ | |
207 | enum update_reason {enqueue, requeue}; | |
208 | ||
0545a303 | 209 | static struct qfq_class *qfq_find_class(struct Qdisc *sch, u32 classid) |
210 | { | |
211 | struct qfq_sched *q = qdisc_priv(sch); | |
212 | struct Qdisc_class_common *clc; | |
213 | ||
214 | clc = qdisc_class_find(&q->clhash, classid); | |
215 | if (clc == NULL) | |
216 | return NULL; | |
217 | return container_of(clc, struct qfq_class, common); | |
218 | } | |
219 | ||
220 | static void qfq_purge_queue(struct qfq_class *cl) | |
221 | { | |
222 | unsigned int len = cl->qdisc->q.qlen; | |
2ccccf5f | 223 | unsigned int backlog = cl->qdisc->qstats.backlog; |
0545a303 | 224 | |
225 | qdisc_reset(cl->qdisc); | |
2ccccf5f | 226 | qdisc_tree_reduce_backlog(cl->qdisc, len, backlog); |
0545a303 | 227 | } |
228 | ||
229 | static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = { | |
230 | [TCA_QFQ_WEIGHT] = { .type = NLA_U32 }, | |
231 | [TCA_QFQ_LMAX] = { .type = NLA_U32 }, | |
232 | }; | |
233 | ||
234 | /* | |
235 | * Calculate a flow index, given its weight and maximum packet length. | |
236 | * index = log_2(maxlen/weight) but we need to apply the scaling. | |
237 | * This is used only once at flow creation. | |
238 | */ | |
462dbc91 | 239 | static int qfq_calc_index(u32 inv_w, unsigned int maxlen, u32 min_slot_shift) |
0545a303 | 240 | { |
241 | u64 slot_size = (u64)maxlen * inv_w; | |
242 | unsigned long size_map; | |
243 | int index = 0; | |
244 | ||
462dbc91 | 245 | size_map = slot_size >> min_slot_shift; |
0545a303 | 246 | if (!size_map) |
247 | goto out; | |
248 | ||
249 | index = __fls(size_map) + 1; /* basically a log_2 */ | |
462dbc91 | 250 | index -= !(slot_size - (1ULL << (index + min_slot_shift - 1))); |
0545a303 | 251 | |
252 | if (index < 0) | |
253 | index = 0; | |
254 | out: | |
255 | pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n", | |
256 | (unsigned long) ONE_FP/inv_w, maxlen, index); | |
257 | ||
258 | return index; | |
259 | } | |
260 | ||
462dbc91 PV |
261 | static void qfq_deactivate_agg(struct qfq_sched *, struct qfq_aggregate *); |
262 | static void qfq_activate_agg(struct qfq_sched *, struct qfq_aggregate *, | |
263 | enum update_reason); | |
264 | ||
265 | static void qfq_init_agg(struct qfq_sched *q, struct qfq_aggregate *agg, | |
266 | u32 lmax, u32 weight) | |
be72f63b | 267 | { |
462dbc91 PV |
268 | INIT_LIST_HEAD(&agg->active); |
269 | hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs); | |
270 | ||
271 | agg->lmax = lmax; | |
272 | agg->class_weight = weight; | |
273 | } | |
274 | ||
275 | static struct qfq_aggregate *qfq_find_agg(struct qfq_sched *q, | |
276 | u32 lmax, u32 weight) | |
277 | { | |
278 | struct qfq_aggregate *agg; | |
462dbc91 | 279 | |
b67bfe0d | 280 | hlist_for_each_entry(agg, &q->nonfull_aggs, nonfull_next) |
462dbc91 PV |
281 | if (agg->lmax == lmax && agg->class_weight == weight) |
282 | return agg; | |
283 | ||
284 | return NULL; | |
285 | } | |
286 | ||
be72f63b | 287 | |
462dbc91 PV |
288 | /* Update aggregate as a function of the new number of classes. */ |
289 | static void qfq_update_agg(struct qfq_sched *q, struct qfq_aggregate *agg, | |
290 | int new_num_classes) | |
291 | { | |
292 | u32 new_agg_weight; | |
293 | ||
294 | if (new_num_classes == q->max_agg_classes) | |
295 | hlist_del_init(&agg->nonfull_next); | |
296 | ||
297 | if (agg->num_classes > new_num_classes && | |
298 | new_num_classes == q->max_agg_classes - 1) /* agg no more full */ | |
299 | hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs); | |
300 | ||
9b99b7e9 PV |
301 | /* The next assignment may let |
302 | * agg->initial_budget > agg->budgetmax | |
303 | * hold, we will take it into account in charge_actual_service(). | |
304 | */ | |
462dbc91 PV |
305 | agg->budgetmax = new_num_classes * agg->lmax; |
306 | new_agg_weight = agg->class_weight * new_num_classes; | |
307 | agg->inv_w = ONE_FP/new_agg_weight; | |
308 | ||
309 | if (agg->grp == NULL) { | |
310 | int i = qfq_calc_index(agg->inv_w, agg->budgetmax, | |
311 | q->min_slot_shift); | |
312 | agg->grp = &q->groups[i]; | |
313 | } | |
314 | ||
315 | q->wsum += | |
316 | (int) agg->class_weight * (new_num_classes - agg->num_classes); | |
87f40dd6 | 317 | q->iwsum = ONE_FP / q->wsum; |
462dbc91 PV |
318 | |
319 | agg->num_classes = new_num_classes; | |
320 | } | |
321 | ||
322 | /* Add class to aggregate. */ | |
323 | static void qfq_add_to_agg(struct qfq_sched *q, | |
324 | struct qfq_aggregate *agg, | |
325 | struct qfq_class *cl) | |
326 | { | |
327 | cl->agg = agg; | |
328 | ||
329 | qfq_update_agg(q, agg, agg->num_classes+1); | |
330 | if (cl->qdisc->q.qlen > 0) { /* adding an active class */ | |
331 | list_add_tail(&cl->alist, &agg->active); | |
332 | if (list_first_entry(&agg->active, struct qfq_class, alist) == | |
333 | cl && q->in_serv_agg != agg) /* agg was inactive */ | |
334 | qfq_activate_agg(q, agg, enqueue); /* schedule agg */ | |
335 | } | |
be72f63b PV |
336 | } |
337 | ||
462dbc91 | 338 | static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *); |
be72f63b | 339 | |
462dbc91 | 340 | static void qfq_destroy_agg(struct qfq_sched *q, struct qfq_aggregate *agg) |
be72f63b | 341 | { |
a55e1c5c | 342 | hlist_del_init(&agg->nonfull_next); |
87f40dd6 PV |
343 | q->wsum -= agg->class_weight; |
344 | if (q->wsum != 0) | |
345 | q->iwsum = ONE_FP / q->wsum; | |
346 | ||
462dbc91 PV |
347 | if (q->in_serv_agg == agg) |
348 | q->in_serv_agg = qfq_choose_next_agg(q); | |
349 | kfree(agg); | |
350 | } | |
be72f63b | 351 | |
462dbc91 PV |
352 | /* Deschedule class from within its parent aggregate. */ |
353 | static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl) | |
354 | { | |
355 | struct qfq_aggregate *agg = cl->agg; | |
be72f63b | 356 | |
be72f63b | 357 | |
462dbc91 PV |
358 | list_del(&cl->alist); /* remove from RR queue of the aggregate */ |
359 | if (list_empty(&agg->active)) /* agg is now inactive */ | |
360 | qfq_deactivate_agg(q, agg); | |
be72f63b PV |
361 | } |
362 | ||
462dbc91 PV |
363 | /* Remove class from its parent aggregate. */ |
364 | static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl) | |
3015f3d2 | 365 | { |
462dbc91 | 366 | struct qfq_aggregate *agg = cl->agg; |
3015f3d2 | 367 | |
462dbc91 PV |
368 | cl->agg = NULL; |
369 | if (agg->num_classes == 1) { /* agg being emptied, destroy it */ | |
370 | qfq_destroy_agg(q, agg); | |
371 | return; | |
3015f3d2 | 372 | } |
462dbc91 PV |
373 | qfq_update_agg(q, agg, agg->num_classes-1); |
374 | } | |
3015f3d2 | 375 | |
462dbc91 PV |
376 | /* Deschedule class and remove it from its parent aggregate. */ |
377 | static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl) | |
378 | { | |
379 | if (cl->qdisc->q.qlen > 0) /* class is active */ | |
380 | qfq_deactivate_class(q, cl); | |
3015f3d2 | 381 | |
462dbc91 | 382 | qfq_rm_from_agg(q, cl); |
3015f3d2 PV |
383 | } |
384 | ||
462dbc91 PV |
385 | /* Move class to a new aggregate, matching the new class weight and/or lmax */ |
386 | static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight, | |
387 | u32 lmax) | |
388 | { | |
389 | struct qfq_sched *q = qdisc_priv(sch); | |
390 | struct qfq_aggregate *new_agg = qfq_find_agg(q, lmax, weight); | |
391 | ||
392 | if (new_agg == NULL) { /* create new aggregate */ | |
393 | new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC); | |
394 | if (new_agg == NULL) | |
395 | return -ENOBUFS; | |
396 | qfq_init_agg(q, new_agg, lmax, weight); | |
397 | } | |
398 | qfq_deact_rm_from_agg(q, cl); | |
399 | qfq_add_to_agg(q, new_agg, cl); | |
400 | ||
401 | return 0; | |
402 | } | |
3015f3d2 | 403 | |
0545a303 | 404 | static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid, |
793d81d6 AA |
405 | struct nlattr **tca, unsigned long *arg, |
406 | struct netlink_ext_ack *extack) | |
0545a303 | 407 | { |
408 | struct qfq_sched *q = qdisc_priv(sch); | |
409 | struct qfq_class *cl = (struct qfq_class *)*arg; | |
462dbc91 | 410 | bool existing = false; |
0545a303 | 411 | struct nlattr *tb[TCA_QFQ_MAX + 1]; |
462dbc91 | 412 | struct qfq_aggregate *new_agg = NULL; |
0545a303 | 413 | u32 weight, lmax, inv_w; |
3015f3d2 | 414 | int err; |
d32ae76f | 415 | int delta_w; |
0545a303 | 416 | |
417 | if (tca[TCA_OPTIONS] == NULL) { | |
418 | pr_notice("qfq: no options\n"); | |
419 | return -EINVAL; | |
420 | } | |
421 | ||
fceb6435 JB |
422 | err = nla_parse_nested(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS], qfq_policy, |
423 | NULL); | |
0545a303 | 424 | if (err < 0) |
425 | return err; | |
426 | ||
427 | if (tb[TCA_QFQ_WEIGHT]) { | |
428 | weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]); | |
429 | if (!weight || weight > (1UL << QFQ_MAX_WSHIFT)) { | |
430 | pr_notice("qfq: invalid weight %u\n", weight); | |
431 | return -EINVAL; | |
432 | } | |
433 | } else | |
434 | weight = 1; | |
435 | ||
0545a303 | 436 | if (tb[TCA_QFQ_LMAX]) { |
437 | lmax = nla_get_u32(tb[TCA_QFQ_LMAX]); | |
3015f3d2 | 438 | if (lmax < QFQ_MIN_LMAX || lmax > (1UL << QFQ_MTU_SHIFT)) { |
0545a303 | 439 | pr_notice("qfq: invalid max length %u\n", lmax); |
440 | return -EINVAL; | |
441 | } | |
442 | } else | |
3015f3d2 | 443 | lmax = psched_mtu(qdisc_dev(sch)); |
0545a303 | 444 | |
462dbc91 PV |
445 | inv_w = ONE_FP / weight; |
446 | weight = ONE_FP / inv_w; | |
447 | ||
448 | if (cl != NULL && | |
449 | lmax == cl->agg->lmax && | |
450 | weight == cl->agg->class_weight) | |
451 | return 0; /* nothing to change */ | |
452 | ||
453 | delta_w = weight - (cl ? cl->agg->class_weight : 0); | |
454 | ||
455 | if (q->wsum + delta_w > QFQ_MAX_WSUM) { | |
456 | pr_notice("qfq: total weight out of range (%d + %u)\n", | |
457 | delta_w, q->wsum); | |
458 | return -EINVAL; | |
459 | } | |
460 | ||
461 | if (cl != NULL) { /* modify existing class */ | |
0545a303 | 462 | if (tca[TCA_RATE]) { |
22e0f8b9 JF |
463 | err = gen_replace_estimator(&cl->bstats, NULL, |
464 | &cl->rate_est, | |
edb09eb1 ED |
465 | NULL, |
466 | qdisc_root_sleeping_running(sch), | |
0545a303 | 467 | tca[TCA_RATE]); |
468 | if (err) | |
469 | return err; | |
470 | } | |
462dbc91 PV |
471 | existing = true; |
472 | goto set_change_agg; | |
0545a303 | 473 | } |
474 | ||
462dbc91 | 475 | /* create and init new class */ |
0545a303 | 476 | cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL); |
477 | if (cl == NULL) | |
478 | return -ENOBUFS; | |
479 | ||
0545a303 | 480 | cl->common.classid = classid; |
462dbc91 | 481 | cl->deficit = lmax; |
0545a303 | 482 | |
483 | cl->qdisc = qdisc_create_dflt(sch->dev_queue, | |
484 | &pfifo_qdisc_ops, classid); | |
485 | if (cl->qdisc == NULL) | |
486 | cl->qdisc = &noop_qdisc; | |
487 | ||
488 | if (tca[TCA_RATE]) { | |
22e0f8b9 JF |
489 | err = gen_new_estimator(&cl->bstats, NULL, |
490 | &cl->rate_est, | |
edb09eb1 ED |
491 | NULL, |
492 | qdisc_root_sleeping_running(sch), | |
0545a303 | 493 | tca[TCA_RATE]); |
462dbc91 PV |
494 | if (err) |
495 | goto destroy_class; | |
0545a303 | 496 | } |
497 | ||
49b49971 JK |
498 | if (cl->qdisc != &noop_qdisc) |
499 | qdisc_hash_add(cl->qdisc, true); | |
0545a303 | 500 | sch_tree_lock(sch); |
501 | qdisc_class_hash_insert(&q->clhash, &cl->common); | |
502 | sch_tree_unlock(sch); | |
503 | ||
504 | qdisc_class_hash_grow(sch, &q->clhash); | |
505 | ||
462dbc91 PV |
506 | set_change_agg: |
507 | sch_tree_lock(sch); | |
508 | new_agg = qfq_find_agg(q, lmax, weight); | |
509 | if (new_agg == NULL) { /* create new aggregate */ | |
510 | sch_tree_unlock(sch); | |
511 | new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL); | |
512 | if (new_agg == NULL) { | |
513 | err = -ENOBUFS; | |
1c0d32fd | 514 | gen_kill_estimator(&cl->rate_est); |
462dbc91 PV |
515 | goto destroy_class; |
516 | } | |
517 | sch_tree_lock(sch); | |
518 | qfq_init_agg(q, new_agg, lmax, weight); | |
519 | } | |
520 | if (existing) | |
521 | qfq_deact_rm_from_agg(q, cl); | |
522 | qfq_add_to_agg(q, new_agg, cl); | |
523 | sch_tree_unlock(sch); | |
524 | ||
0545a303 | 525 | *arg = (unsigned long)cl; |
526 | return 0; | |
462dbc91 PV |
527 | |
528 | destroy_class: | |
529 | qdisc_destroy(cl->qdisc); | |
530 | kfree(cl); | |
531 | return err; | |
0545a303 | 532 | } |
533 | ||
534 | static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl) | |
535 | { | |
536 | struct qfq_sched *q = qdisc_priv(sch); | |
537 | ||
462dbc91 | 538 | qfq_rm_from_agg(q, cl); |
1c0d32fd | 539 | gen_kill_estimator(&cl->rate_est); |
0545a303 | 540 | qdisc_destroy(cl->qdisc); |
541 | kfree(cl); | |
542 | } | |
543 | ||
544 | static int qfq_delete_class(struct Qdisc *sch, unsigned long arg) | |
545 | { | |
546 | struct qfq_sched *q = qdisc_priv(sch); | |
547 | struct qfq_class *cl = (struct qfq_class *)arg; | |
548 | ||
549 | if (cl->filter_cnt > 0) | |
550 | return -EBUSY; | |
551 | ||
552 | sch_tree_lock(sch); | |
553 | ||
554 | qfq_purge_queue(cl); | |
555 | qdisc_class_hash_remove(&q->clhash, &cl->common); | |
556 | ||
0545a303 | 557 | sch_tree_unlock(sch); |
0545a303 | 558 | |
143976ce WC |
559 | qfq_destroy_class(sch, cl); |
560 | return 0; | |
0545a303 | 561 | } |
562 | ||
143976ce | 563 | static unsigned long qfq_search_class(struct Qdisc *sch, u32 classid) |
0545a303 | 564 | { |
143976ce | 565 | return (unsigned long)qfq_find_class(sch, classid); |
0545a303 | 566 | } |
567 | ||
cbaacc4e AA |
568 | static struct tcf_block *qfq_tcf_block(struct Qdisc *sch, unsigned long cl, |
569 | struct netlink_ext_ack *extack) | |
0545a303 | 570 | { |
571 | struct qfq_sched *q = qdisc_priv(sch); | |
572 | ||
573 | if (cl) | |
574 | return NULL; | |
575 | ||
6529eaba | 576 | return q->block; |
0545a303 | 577 | } |
578 | ||
579 | static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent, | |
580 | u32 classid) | |
581 | { | |
582 | struct qfq_class *cl = qfq_find_class(sch, classid); | |
583 | ||
584 | if (cl != NULL) | |
585 | cl->filter_cnt++; | |
586 | ||
587 | return (unsigned long)cl; | |
588 | } | |
589 | ||
590 | static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg) | |
591 | { | |
592 | struct qfq_class *cl = (struct qfq_class *)arg; | |
593 | ||
594 | cl->filter_cnt--; | |
595 | } | |
596 | ||
597 | static int qfq_graft_class(struct Qdisc *sch, unsigned long arg, | |
598 | struct Qdisc *new, struct Qdisc **old) | |
599 | { | |
600 | struct qfq_class *cl = (struct qfq_class *)arg; | |
601 | ||
602 | if (new == NULL) { | |
603 | new = qdisc_create_dflt(sch->dev_queue, | |
604 | &pfifo_qdisc_ops, cl->common.classid); | |
605 | if (new == NULL) | |
606 | new = &noop_qdisc; | |
607 | } | |
608 | ||
86a7996c | 609 | *old = qdisc_replace(sch, new, &cl->qdisc); |
0545a303 | 610 | return 0; |
611 | } | |
612 | ||
613 | static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg) | |
614 | { | |
615 | struct qfq_class *cl = (struct qfq_class *)arg; | |
616 | ||
617 | return cl->qdisc; | |
618 | } | |
619 | ||
620 | static int qfq_dump_class(struct Qdisc *sch, unsigned long arg, | |
621 | struct sk_buff *skb, struct tcmsg *tcm) | |
622 | { | |
623 | struct qfq_class *cl = (struct qfq_class *)arg; | |
624 | struct nlattr *nest; | |
625 | ||
626 | tcm->tcm_parent = TC_H_ROOT; | |
627 | tcm->tcm_handle = cl->common.classid; | |
628 | tcm->tcm_info = cl->qdisc->handle; | |
629 | ||
630 | nest = nla_nest_start(skb, TCA_OPTIONS); | |
631 | if (nest == NULL) | |
632 | goto nla_put_failure; | |
462dbc91 PV |
633 | if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) || |
634 | nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax)) | |
1b34ec43 | 635 | goto nla_put_failure; |
0545a303 | 636 | return nla_nest_end(skb, nest); |
637 | ||
638 | nla_put_failure: | |
639 | nla_nest_cancel(skb, nest); | |
640 | return -EMSGSIZE; | |
641 | } | |
642 | ||
643 | static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg, | |
644 | struct gnet_dump *d) | |
645 | { | |
646 | struct qfq_class *cl = (struct qfq_class *)arg; | |
647 | struct tc_qfq_stats xstats; | |
648 | ||
649 | memset(&xstats, 0, sizeof(xstats)); | |
0545a303 | 650 | |
462dbc91 PV |
651 | xstats.weight = cl->agg->class_weight; |
652 | xstats.lmax = cl->agg->lmax; | |
0545a303 | 653 | |
edb09eb1 ED |
654 | if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch), |
655 | d, NULL, &cl->bstats) < 0 || | |
1c0d32fd | 656 | gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 || |
b0ab6f92 JF |
657 | gnet_stats_copy_queue(d, NULL, |
658 | &cl->qdisc->qstats, cl->qdisc->q.qlen) < 0) | |
0545a303 | 659 | return -1; |
660 | ||
661 | return gnet_stats_copy_app(d, &xstats, sizeof(xstats)); | |
662 | } | |
663 | ||
664 | static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg) | |
665 | { | |
666 | struct qfq_sched *q = qdisc_priv(sch); | |
667 | struct qfq_class *cl; | |
0545a303 | 668 | unsigned int i; |
669 | ||
670 | if (arg->stop) | |
671 | return; | |
672 | ||
673 | for (i = 0; i < q->clhash.hashsize; i++) { | |
b67bfe0d | 674 | hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) { |
0545a303 | 675 | if (arg->count < arg->skip) { |
676 | arg->count++; | |
677 | continue; | |
678 | } | |
679 | if (arg->fn(sch, (unsigned long)cl, arg) < 0) { | |
680 | arg->stop = 1; | |
681 | return; | |
682 | } | |
683 | arg->count++; | |
684 | } | |
685 | } | |
686 | } | |
687 | ||
688 | static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch, | |
689 | int *qerr) | |
690 | { | |
691 | struct qfq_sched *q = qdisc_priv(sch); | |
692 | struct qfq_class *cl; | |
693 | struct tcf_result res; | |
25d8c0d5 | 694 | struct tcf_proto *fl; |
0545a303 | 695 | int result; |
696 | ||
697 | if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) { | |
698 | pr_debug("qfq_classify: found %d\n", skb->priority); | |
699 | cl = qfq_find_class(sch, skb->priority); | |
700 | if (cl != NULL) | |
701 | return cl; | |
702 | } | |
703 | ||
704 | *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS; | |
25d8c0d5 | 705 | fl = rcu_dereference_bh(q->filter_list); |
87d83093 | 706 | result = tcf_classify(skb, fl, &res, false); |
0545a303 | 707 | if (result >= 0) { |
708 | #ifdef CONFIG_NET_CLS_ACT | |
709 | switch (result) { | |
710 | case TC_ACT_QUEUED: | |
711 | case TC_ACT_STOLEN: | |
e25ea21f | 712 | case TC_ACT_TRAP: |
0545a303 | 713 | *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN; |
f3ae608e | 714 | /* fall through */ |
0545a303 | 715 | case TC_ACT_SHOT: |
716 | return NULL; | |
717 | } | |
718 | #endif | |
719 | cl = (struct qfq_class *)res.class; | |
720 | if (cl == NULL) | |
721 | cl = qfq_find_class(sch, res.classid); | |
722 | return cl; | |
723 | } | |
724 | ||
725 | return NULL; | |
726 | } | |
727 | ||
728 | /* Generic comparison function, handling wraparound. */ | |
729 | static inline int qfq_gt(u64 a, u64 b) | |
730 | { | |
731 | return (s64)(a - b) > 0; | |
732 | } | |
733 | ||
734 | /* Round a precise timestamp to its slotted value. */ | |
735 | static inline u64 qfq_round_down(u64 ts, unsigned int shift) | |
736 | { | |
737 | return ts & ~((1ULL << shift) - 1); | |
738 | } | |
739 | ||
740 | /* return the pointer to the group with lowest index in the bitmap */ | |
741 | static inline struct qfq_group *qfq_ffs(struct qfq_sched *q, | |
742 | unsigned long bitmap) | |
743 | { | |
744 | int index = __ffs(bitmap); | |
745 | return &q->groups[index]; | |
746 | } | |
747 | /* Calculate a mask to mimic what would be ffs_from(). */ | |
748 | static inline unsigned long mask_from(unsigned long bitmap, int from) | |
749 | { | |
750 | return bitmap & ~((1UL << from) - 1); | |
751 | } | |
752 | ||
753 | /* | |
754 | * The state computation relies on ER=0, IR=1, EB=2, IB=3 | |
755 | * First compute eligibility comparing grp->S, q->V, | |
756 | * then check if someone is blocking us and possibly add EB | |
757 | */ | |
758 | static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp) | |
759 | { | |
760 | /* if S > V we are not eligible */ | |
761 | unsigned int state = qfq_gt(grp->S, q->V); | |
762 | unsigned long mask = mask_from(q->bitmaps[ER], grp->index); | |
763 | struct qfq_group *next; | |
764 | ||
765 | if (mask) { | |
766 | next = qfq_ffs(q, mask); | |
767 | if (qfq_gt(grp->F, next->F)) | |
768 | state |= EB; | |
769 | } | |
770 | ||
771 | return state; | |
772 | } | |
773 | ||
774 | ||
775 | /* | |
776 | * In principle | |
777 | * q->bitmaps[dst] |= q->bitmaps[src] & mask; | |
778 | * q->bitmaps[src] &= ~mask; | |
779 | * but we should make sure that src != dst | |
780 | */ | |
781 | static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask, | |
782 | int src, int dst) | |
783 | { | |
784 | q->bitmaps[dst] |= q->bitmaps[src] & mask; | |
785 | q->bitmaps[src] &= ~mask; | |
786 | } | |
787 | ||
788 | static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F) | |
789 | { | |
790 | unsigned long mask = mask_from(q->bitmaps[ER], index + 1); | |
791 | struct qfq_group *next; | |
792 | ||
793 | if (mask) { | |
794 | next = qfq_ffs(q, mask); | |
795 | if (!qfq_gt(next->F, old_F)) | |
796 | return; | |
797 | } | |
798 | ||
799 | mask = (1UL << index) - 1; | |
800 | qfq_move_groups(q, mask, EB, ER); | |
801 | qfq_move_groups(q, mask, IB, IR); | |
802 | } | |
803 | ||
804 | /* | |
805 | * perhaps | |
806 | * | |
807 | old_V ^= q->V; | |
462dbc91 | 808 | old_V >>= q->min_slot_shift; |
0545a303 | 809 | if (old_V) { |
810 | ... | |
811 | } | |
812 | * | |
813 | */ | |
462dbc91 | 814 | static void qfq_make_eligible(struct qfq_sched *q) |
0545a303 | 815 | { |
462dbc91 PV |
816 | unsigned long vslot = q->V >> q->min_slot_shift; |
817 | unsigned long old_vslot = q->oldV >> q->min_slot_shift; | |
0545a303 | 818 | |
819 | if (vslot != old_vslot) { | |
87f1369d PV |
820 | unsigned long mask; |
821 | int last_flip_pos = fls(vslot ^ old_vslot); | |
822 | ||
823 | if (last_flip_pos > 31) /* higher than the number of groups */ | |
824 | mask = ~0UL; /* make all groups eligible */ | |
825 | else | |
826 | mask = (1UL << last_flip_pos) - 1; | |
827 | ||
0545a303 | 828 | qfq_move_groups(q, mask, IR, ER); |
829 | qfq_move_groups(q, mask, IB, EB); | |
830 | } | |
831 | } | |
832 | ||
0545a303 | 833 | /* |
87f40dd6 PV |
834 | * The index of the slot in which the input aggregate agg is to be |
835 | * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2' | |
836 | * and not a '-1' because the start time of the group may be moved | |
837 | * backward by one slot after the aggregate has been inserted, and | |
838 | * this would cause non-empty slots to be right-shifted by one | |
839 | * position. | |
840 | * | |
841 | * QFQ+ fully satisfies this bound to the slot index if the parameters | |
842 | * of the classes are not changed dynamically, and if QFQ+ never | |
843 | * happens to postpone the service of agg unjustly, i.e., it never | |
844 | * happens that the aggregate becomes backlogged and eligible, or just | |
845 | * eligible, while an aggregate with a higher approximated finish time | |
846 | * is being served. In particular, in this case QFQ+ guarantees that | |
847 | * the timestamps of agg are low enough that the slot index is never | |
848 | * higher than 2. Unfortunately, QFQ+ cannot provide the same | |
849 | * guarantee if it happens to unjustly postpone the service of agg, or | |
850 | * if the parameters of some class are changed. | |
851 | * | |
852 | * As for the first event, i.e., an out-of-order service, the | |
853 | * upper bound to the slot index guaranteed by QFQ+ grows to | |
854 | * 2 + | |
855 | * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) * | |
856 | * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1. | |
3015f3d2 | 857 | * |
87f40dd6 PV |
858 | * The following function deals with this problem by backward-shifting |
859 | * the timestamps of agg, if needed, so as to guarantee that the slot | |
860 | * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may | |
861 | * cause the service of other aggregates to be postponed, yet the | |
862 | * worst-case guarantees of these aggregates are not violated. In | |
863 | * fact, in case of no out-of-order service, the timestamps of agg | |
864 | * would have been even lower than they are after the backward shift, | |
865 | * because QFQ+ would have guaranteed a maximum value equal to 2 for | |
866 | * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose | |
867 | * service is postponed because of the backward-shift would have | |
868 | * however waited for the service of agg before being served. | |
3015f3d2 | 869 | * |
87f40dd6 PV |
870 | * The other event that may cause the slot index to be higher than 2 |
871 | * for agg is a recent change of the parameters of some class. If the | |
872 | * weight of a class is increased or the lmax (max_pkt_size) of the | |
873 | * class is decreased, then a new aggregate with smaller slot size | |
874 | * than the original parent aggregate of the class may happen to be | |
875 | * activated. The activation of this aggregate should be properly | |
876 | * delayed to when the service of the class has finished in the ideal | |
877 | * system tracked by QFQ+. If the activation of the aggregate is not | |
878 | * delayed to this reference time instant, then this aggregate may be | |
879 | * unjustly served before other aggregates waiting for service. This | |
880 | * may cause the above bound to the slot index to be violated for some | |
881 | * of these unlucky aggregates. | |
3015f3d2 | 882 | * |
462dbc91 | 883 | * Instead of delaying the activation of the new aggregate, which is |
87f40dd6 PV |
884 | * quite complex, the above-discussed capping of the slot index is |
885 | * used to handle also the consequences of a change of the parameters | |
886 | * of a class. | |
0545a303 | 887 | */ |
462dbc91 | 888 | static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg, |
0545a303 | 889 | u64 roundedS) |
890 | { | |
891 | u64 slot = (roundedS - grp->S) >> grp->slot_shift; | |
3015f3d2 PV |
892 | unsigned int i; /* slot index in the bucket list */ |
893 | ||
894 | if (unlikely(slot > QFQ_MAX_SLOTS - 2)) { | |
895 | u64 deltaS = roundedS - grp->S - | |
896 | ((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift); | |
462dbc91 PV |
897 | agg->S -= deltaS; |
898 | agg->F -= deltaS; | |
3015f3d2 PV |
899 | slot = QFQ_MAX_SLOTS - 2; |
900 | } | |
901 | ||
902 | i = (grp->front + slot) % QFQ_MAX_SLOTS; | |
0545a303 | 903 | |
462dbc91 | 904 | hlist_add_head(&agg->next, &grp->slots[i]); |
0545a303 | 905 | __set_bit(slot, &grp->full_slots); |
906 | } | |
907 | ||
908 | /* Maybe introduce hlist_first_entry?? */ | |
462dbc91 | 909 | static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp) |
0545a303 | 910 | { |
911 | return hlist_entry(grp->slots[grp->front].first, | |
462dbc91 | 912 | struct qfq_aggregate, next); |
0545a303 | 913 | } |
914 | ||
915 | /* | |
916 | * remove the entry from the slot | |
917 | */ | |
918 | static void qfq_front_slot_remove(struct qfq_group *grp) | |
919 | { | |
462dbc91 | 920 | struct qfq_aggregate *agg = qfq_slot_head(grp); |
0545a303 | 921 | |
462dbc91 PV |
922 | BUG_ON(!agg); |
923 | hlist_del(&agg->next); | |
0545a303 | 924 | if (hlist_empty(&grp->slots[grp->front])) |
925 | __clear_bit(0, &grp->full_slots); | |
926 | } | |
927 | ||
928 | /* | |
462dbc91 PV |
929 | * Returns the first aggregate in the first non-empty bucket of the |
930 | * group. As a side effect, adjusts the bucket list so the first | |
931 | * non-empty bucket is at position 0 in full_slots. | |
0545a303 | 932 | */ |
462dbc91 | 933 | static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp) |
0545a303 | 934 | { |
935 | unsigned int i; | |
936 | ||
937 | pr_debug("qfq slot_scan: grp %u full %#lx\n", | |
938 | grp->index, grp->full_slots); | |
939 | ||
940 | if (grp->full_slots == 0) | |
941 | return NULL; | |
942 | ||
943 | i = __ffs(grp->full_slots); /* zero based */ | |
944 | if (i > 0) { | |
945 | grp->front = (grp->front + i) % QFQ_MAX_SLOTS; | |
946 | grp->full_slots >>= i; | |
947 | } | |
948 | ||
949 | return qfq_slot_head(grp); | |
950 | } | |
951 | ||
952 | /* | |
953 | * adjust the bucket list. When the start time of a group decreases, | |
954 | * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to | |
955 | * move the objects. The mask of occupied slots must be shifted | |
956 | * because we use ffs() to find the first non-empty slot. | |
957 | * This covers decreases in the group's start time, but what about | |
958 | * increases of the start time ? | |
959 | * Here too we should make sure that i is less than 32 | |
960 | */ | |
961 | static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS) | |
962 | { | |
963 | unsigned int i = (grp->S - roundedS) >> grp->slot_shift; | |
964 | ||
965 | grp->full_slots <<= i; | |
966 | grp->front = (grp->front - i) % QFQ_MAX_SLOTS; | |
967 | } | |
968 | ||
462dbc91 | 969 | static void qfq_update_eligible(struct qfq_sched *q) |
0545a303 | 970 | { |
971 | struct qfq_group *grp; | |
972 | unsigned long ineligible; | |
973 | ||
974 | ineligible = q->bitmaps[IR] | q->bitmaps[IB]; | |
975 | if (ineligible) { | |
976 | if (!q->bitmaps[ER]) { | |
977 | grp = qfq_ffs(q, ineligible); | |
978 | if (qfq_gt(grp->S, q->V)) | |
979 | q->V = grp->S; | |
980 | } | |
462dbc91 | 981 | qfq_make_eligible(q); |
0545a303 | 982 | } |
983 | } | |
984 | ||
462dbc91 PV |
985 | /* Dequeue head packet of the head class in the DRR queue of the aggregate. */ |
986 | static void agg_dequeue(struct qfq_aggregate *agg, | |
987 | struct qfq_class *cl, unsigned int len) | |
0545a303 | 988 | { |
462dbc91 | 989 | qdisc_dequeue_peeked(cl->qdisc); |
0545a303 | 990 | |
462dbc91 | 991 | cl->deficit -= (int) len; |
0545a303 | 992 | |
462dbc91 PV |
993 | if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */ |
994 | list_del(&cl->alist); | |
995 | else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) { | |
996 | cl->deficit += agg->lmax; | |
997 | list_move_tail(&cl->alist, &agg->active); | |
0545a303 | 998 | } |
462dbc91 PV |
999 | } |
1000 | ||
1001 | static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg, | |
1002 | struct qfq_class **cl, | |
1003 | unsigned int *len) | |
1004 | { | |
1005 | struct sk_buff *skb; | |
0545a303 | 1006 | |
462dbc91 PV |
1007 | *cl = list_first_entry(&agg->active, struct qfq_class, alist); |
1008 | skb = (*cl)->qdisc->ops->peek((*cl)->qdisc); | |
1009 | if (skb == NULL) | |
1010 | WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n"); | |
1011 | else | |
1012 | *len = qdisc_pkt_len(skb); | |
1013 | ||
1014 | return skb; | |
1015 | } | |
1016 | ||
1017 | /* Update F according to the actual service received by the aggregate. */ | |
1018 | static inline void charge_actual_service(struct qfq_aggregate *agg) | |
1019 | { | |
9b99b7e9 PV |
1020 | /* Compute the service received by the aggregate, taking into |
1021 | * account that, after decreasing the number of classes in | |
1022 | * agg, it may happen that | |
1023 | * agg->initial_budget - agg->budget > agg->bugdetmax | |
1024 | */ | |
1025 | u32 service_received = min(agg->budgetmax, | |
1026 | agg->initial_budget - agg->budget); | |
462dbc91 PV |
1027 | |
1028 | agg->F = agg->S + (u64)service_received * agg->inv_w; | |
0545a303 | 1029 | } |
1030 | ||
88d4f419 PV |
1031 | /* Assign a reasonable start time for a new aggregate in group i. |
1032 | * Admissible values for \hat(F) are multiples of \sigma_i | |
1033 | * no greater than V+\sigma_i . Larger values mean that | |
1034 | * we had a wraparound so we consider the timestamp to be stale. | |
1035 | * | |
1036 | * If F is not stale and F >= V then we set S = F. | |
1037 | * Otherwise we should assign S = V, but this may violate | |
1038 | * the ordering in EB (see [2]). So, if we have groups in ER, | |
1039 | * set S to the F_j of the first group j which would be blocking us. | |
1040 | * We are guaranteed not to move S backward because | |
1041 | * otherwise our group i would still be blocked. | |
1042 | */ | |
1043 | static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg) | |
1044 | { | |
1045 | unsigned long mask; | |
1046 | u64 limit, roundedF; | |
1047 | int slot_shift = agg->grp->slot_shift; | |
1048 | ||
1049 | roundedF = qfq_round_down(agg->F, slot_shift); | |
1050 | limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift); | |
1051 | ||
1052 | if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) { | |
1053 | /* timestamp was stale */ | |
1054 | mask = mask_from(q->bitmaps[ER], agg->grp->index); | |
1055 | if (mask) { | |
1056 | struct qfq_group *next = qfq_ffs(q, mask); | |
1057 | if (qfq_gt(roundedF, next->F)) { | |
1058 | if (qfq_gt(limit, next->F)) | |
1059 | agg->S = next->F; | |
1060 | else /* preserve timestamp correctness */ | |
1061 | agg->S = limit; | |
1062 | return; | |
1063 | } | |
1064 | } | |
1065 | agg->S = q->V; | |
1066 | } else /* timestamp is not stale */ | |
1067 | agg->S = agg->F; | |
1068 | } | |
1069 | ||
1070 | /* Update the timestamps of agg before scheduling/rescheduling it for | |
1071 | * service. In particular, assign to agg->F its maximum possible | |
1072 | * value, i.e., the virtual finish time with which the aggregate | |
1073 | * should be labeled if it used all its budget once in service. | |
1074 | */ | |
1075 | static inline void | |
1076 | qfq_update_agg_ts(struct qfq_sched *q, | |
1077 | struct qfq_aggregate *agg, enum update_reason reason) | |
1078 | { | |
1079 | if (reason != requeue) | |
1080 | qfq_update_start(q, agg); | |
1081 | else /* just charge agg for the service received */ | |
1082 | agg->S = agg->F; | |
1083 | ||
1084 | agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w; | |
1085 | } | |
2f3b89a1 PV |
1086 | |
1087 | static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg); | |
1088 | ||
0545a303 | 1089 | static struct sk_buff *qfq_dequeue(struct Qdisc *sch) |
1090 | { | |
1091 | struct qfq_sched *q = qdisc_priv(sch); | |
462dbc91 | 1092 | struct qfq_aggregate *in_serv_agg = q->in_serv_agg; |
0545a303 | 1093 | struct qfq_class *cl; |
462dbc91 PV |
1094 | struct sk_buff *skb = NULL; |
1095 | /* next-packet len, 0 means no more active classes in in-service agg */ | |
1096 | unsigned int len = 0; | |
0545a303 | 1097 | |
462dbc91 | 1098 | if (in_serv_agg == NULL) |
0545a303 | 1099 | return NULL; |
1100 | ||
462dbc91 PV |
1101 | if (!list_empty(&in_serv_agg->active)) |
1102 | skb = qfq_peek_skb(in_serv_agg, &cl, &len); | |
0545a303 | 1103 | |
462dbc91 PV |
1104 | /* |
1105 | * If there are no active classes in the in-service aggregate, | |
1106 | * or if the aggregate has not enough budget to serve its next | |
1107 | * class, then choose the next aggregate to serve. | |
1108 | */ | |
1109 | if (len == 0 || in_serv_agg->budget < len) { | |
1110 | charge_actual_service(in_serv_agg); | |
1111 | ||
1112 | /* recharge the budget of the aggregate */ | |
1113 | in_serv_agg->initial_budget = in_serv_agg->budget = | |
1114 | in_serv_agg->budgetmax; | |
1115 | ||
2f3b89a1 | 1116 | if (!list_empty(&in_serv_agg->active)) { |
462dbc91 PV |
1117 | /* |
1118 | * Still active: reschedule for | |
1119 | * service. Possible optimization: if no other | |
1120 | * aggregate is active, then there is no point | |
1121 | * in rescheduling this aggregate, and we can | |
1122 | * just keep it as the in-service one. This | |
1123 | * should be however a corner case, and to | |
1124 | * handle it, we would need to maintain an | |
1125 | * extra num_active_aggs field. | |
1126 | */ | |
2f3b89a1 PV |
1127 | qfq_update_agg_ts(q, in_serv_agg, requeue); |
1128 | qfq_schedule_agg(q, in_serv_agg); | |
1129 | } else if (sch->q.qlen == 0) { /* no aggregate to serve */ | |
462dbc91 PV |
1130 | q->in_serv_agg = NULL; |
1131 | return NULL; | |
1132 | } | |
1133 | ||
1134 | /* | |
1135 | * If we get here, there are other aggregates queued: | |
1136 | * choose the new aggregate to serve. | |
1137 | */ | |
1138 | in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q); | |
1139 | skb = qfq_peek_skb(in_serv_agg, &cl, &len); | |
0545a303 | 1140 | } |
462dbc91 PV |
1141 | if (!skb) |
1142 | return NULL; | |
0545a303 | 1143 | |
2ed5c3f0 | 1144 | qdisc_qstats_backlog_dec(sch, skb); |
0545a303 | 1145 | sch->q.qlen--; |
1146 | qdisc_bstats_update(sch, skb); | |
1147 | ||
462dbc91 | 1148 | agg_dequeue(in_serv_agg, cl, len); |
a0143efa PV |
1149 | /* If lmax is lowered, through qfq_change_class, for a class |
1150 | * owning pending packets with larger size than the new value | |
1151 | * of lmax, then the following condition may hold. | |
1152 | */ | |
1153 | if (unlikely(in_serv_agg->budget < len)) | |
1154 | in_serv_agg->budget = 0; | |
1155 | else | |
1156 | in_serv_agg->budget -= len; | |
1157 | ||
87f40dd6 | 1158 | q->V += (u64)len * q->iwsum; |
0545a303 | 1159 | pr_debug("qfq dequeue: len %u F %lld now %lld\n", |
462dbc91 PV |
1160 | len, (unsigned long long) in_serv_agg->F, |
1161 | (unsigned long long) q->V); | |
0545a303 | 1162 | |
462dbc91 PV |
1163 | return skb; |
1164 | } | |
0545a303 | 1165 | |
462dbc91 PV |
1166 | static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q) |
1167 | { | |
1168 | struct qfq_group *grp; | |
1169 | struct qfq_aggregate *agg, *new_front_agg; | |
1170 | u64 old_F; | |
0545a303 | 1171 | |
462dbc91 PV |
1172 | qfq_update_eligible(q); |
1173 | q->oldV = q->V; | |
1174 | ||
1175 | if (!q->bitmaps[ER]) | |
1176 | return NULL; | |
1177 | ||
1178 | grp = qfq_ffs(q, q->bitmaps[ER]); | |
1179 | old_F = grp->F; | |
1180 | ||
1181 | agg = qfq_slot_head(grp); | |
0545a303 | 1182 | |
462dbc91 PV |
1183 | /* agg starts to be served, remove it from schedule */ |
1184 | qfq_front_slot_remove(grp); | |
1185 | ||
1186 | new_front_agg = qfq_slot_scan(grp); | |
1187 | ||
1188 | if (new_front_agg == NULL) /* group is now inactive, remove from ER */ | |
1189 | __clear_bit(grp->index, &q->bitmaps[ER]); | |
1190 | else { | |
1191 | u64 roundedS = qfq_round_down(new_front_agg->S, | |
1192 | grp->slot_shift); | |
1193 | unsigned int s; | |
1194 | ||
1195 | if (grp->S == roundedS) | |
1196 | return agg; | |
1197 | grp->S = roundedS; | |
1198 | grp->F = roundedS + (2ULL << grp->slot_shift); | |
1199 | __clear_bit(grp->index, &q->bitmaps[ER]); | |
1200 | s = qfq_calc_state(q, grp); | |
1201 | __set_bit(grp->index, &q->bitmaps[s]); | |
0545a303 | 1202 | } |
1203 | ||
462dbc91 | 1204 | qfq_unblock_groups(q, grp->index, old_F); |
0545a303 | 1205 | |
462dbc91 | 1206 | return agg; |
0545a303 | 1207 | } |
1208 | ||
520ac30f ED |
1209 | static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch, |
1210 | struct sk_buff **to_free) | |
0545a303 | 1211 | { |
1212 | struct qfq_sched *q = qdisc_priv(sch); | |
0545a303 | 1213 | struct qfq_class *cl; |
462dbc91 | 1214 | struct qfq_aggregate *agg; |
f54ba779 | 1215 | int err = 0; |
0545a303 | 1216 | |
1217 | cl = qfq_classify(skb, sch, &err); | |
1218 | if (cl == NULL) { | |
1219 | if (err & __NET_XMIT_BYPASS) | |
25331d6c | 1220 | qdisc_qstats_drop(sch); |
39ad1297 | 1221 | __qdisc_drop(skb, to_free); |
0545a303 | 1222 | return err; |
1223 | } | |
1224 | pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid); | |
1225 | ||
462dbc91 | 1226 | if (unlikely(cl->agg->lmax < qdisc_pkt_len(skb))) { |
3015f3d2 | 1227 | pr_debug("qfq: increasing maxpkt from %u to %u for class %u", |
462dbc91 PV |
1228 | cl->agg->lmax, qdisc_pkt_len(skb), cl->common.classid); |
1229 | err = qfq_change_agg(sch, cl, cl->agg->class_weight, | |
1230 | qdisc_pkt_len(skb)); | |
9b15350f FW |
1231 | if (err) { |
1232 | cl->qstats.drops++; | |
520ac30f | 1233 | return qdisc_drop(skb, sch, to_free); |
9b15350f | 1234 | } |
3015f3d2 PV |
1235 | } |
1236 | ||
520ac30f | 1237 | err = qdisc_enqueue(skb, cl->qdisc, to_free); |
0545a303 | 1238 | if (unlikely(err != NET_XMIT_SUCCESS)) { |
1239 | pr_debug("qfq_enqueue: enqueue failed %d\n", err); | |
1240 | if (net_xmit_drop_count(err)) { | |
1241 | cl->qstats.drops++; | |
25331d6c | 1242 | qdisc_qstats_drop(sch); |
0545a303 | 1243 | } |
1244 | return err; | |
1245 | } | |
1246 | ||
1247 | bstats_update(&cl->bstats, skb); | |
2ed5c3f0 | 1248 | qdisc_qstats_backlog_inc(sch, skb); |
0545a303 | 1249 | ++sch->q.qlen; |
1250 | ||
462dbc91 PV |
1251 | agg = cl->agg; |
1252 | /* if the queue was not empty, then done here */ | |
1253 | if (cl->qdisc->q.qlen != 1) { | |
1254 | if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) && | |
1255 | list_first_entry(&agg->active, struct qfq_class, alist) | |
1256 | == cl && cl->deficit < qdisc_pkt_len(skb)) | |
1257 | list_move_tail(&cl->alist, &agg->active); | |
1258 | ||
0545a303 | 1259 | return err; |
462dbc91 PV |
1260 | } |
1261 | ||
1262 | /* schedule class for service within the aggregate */ | |
1263 | cl->deficit = agg->lmax; | |
1264 | list_add_tail(&cl->alist, &agg->active); | |
0545a303 | 1265 | |
2f3b89a1 PV |
1266 | if (list_first_entry(&agg->active, struct qfq_class, alist) != cl || |
1267 | q->in_serv_agg == agg) | |
1268 | return err; /* non-empty or in service, nothing else to do */ | |
462dbc91 | 1269 | |
2f3b89a1 | 1270 | qfq_activate_agg(q, agg, enqueue); |
be72f63b PV |
1271 | |
1272 | return err; | |
1273 | } | |
1274 | ||
1275 | /* | |
462dbc91 | 1276 | * Schedule aggregate according to its timestamps. |
be72f63b | 1277 | */ |
462dbc91 | 1278 | static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg) |
be72f63b | 1279 | { |
462dbc91 | 1280 | struct qfq_group *grp = agg->grp; |
be72f63b PV |
1281 | u64 roundedS; |
1282 | int s; | |
1283 | ||
462dbc91 | 1284 | roundedS = qfq_round_down(agg->S, grp->slot_shift); |
0545a303 | 1285 | |
1286 | /* | |
462dbc91 PV |
1287 | * Insert agg in the correct bucket. |
1288 | * If agg->S >= grp->S we don't need to adjust the | |
0545a303 | 1289 | * bucket list and simply go to the insertion phase. |
1290 | * Otherwise grp->S is decreasing, we must make room | |
1291 | * in the bucket list, and also recompute the group state. | |
1292 | * Finally, if there were no flows in this group and nobody | |
1293 | * was in ER make sure to adjust V. | |
1294 | */ | |
1295 | if (grp->full_slots) { | |
462dbc91 | 1296 | if (!qfq_gt(grp->S, agg->S)) |
0545a303 | 1297 | goto skip_update; |
1298 | ||
462dbc91 | 1299 | /* create a slot for this agg->S */ |
0545a303 | 1300 | qfq_slot_rotate(grp, roundedS); |
1301 | /* group was surely ineligible, remove */ | |
1302 | __clear_bit(grp->index, &q->bitmaps[IR]); | |
1303 | __clear_bit(grp->index, &q->bitmaps[IB]); | |
40dd2d54 PV |
1304 | } else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) && |
1305 | q->in_serv_agg == NULL) | |
0545a303 | 1306 | q->V = roundedS; |
1307 | ||
1308 | grp->S = roundedS; | |
1309 | grp->F = roundedS + (2ULL << grp->slot_shift); | |
1310 | s = qfq_calc_state(q, grp); | |
1311 | __set_bit(grp->index, &q->bitmaps[s]); | |
1312 | ||
1313 | pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n", | |
1314 | s, q->bitmaps[s], | |
462dbc91 PV |
1315 | (unsigned long long) agg->S, |
1316 | (unsigned long long) agg->F, | |
0545a303 | 1317 | (unsigned long long) q->V); |
1318 | ||
1319 | skip_update: | |
462dbc91 | 1320 | qfq_slot_insert(grp, agg, roundedS); |
0545a303 | 1321 | } |
1322 | ||
1323 | ||
462dbc91 PV |
1324 | /* Update agg ts and schedule agg for service */ |
1325 | static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg, | |
1326 | enum update_reason reason) | |
1327 | { | |
2f3b89a1 PV |
1328 | agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */ |
1329 | ||
462dbc91 | 1330 | qfq_update_agg_ts(q, agg, reason); |
2f3b89a1 PV |
1331 | if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */ |
1332 | q->in_serv_agg = agg; /* start serving this aggregate */ | |
1333 | /* update V: to be in service, agg must be eligible */ | |
1334 | q->oldV = q->V = agg->S; | |
1335 | } else if (agg != q->in_serv_agg) | |
1336 | qfq_schedule_agg(q, agg); | |
462dbc91 PV |
1337 | } |
1338 | ||
0545a303 | 1339 | static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp, |
462dbc91 | 1340 | struct qfq_aggregate *agg) |
0545a303 | 1341 | { |
1342 | unsigned int i, offset; | |
1343 | u64 roundedS; | |
1344 | ||
462dbc91 | 1345 | roundedS = qfq_round_down(agg->S, grp->slot_shift); |
0545a303 | 1346 | offset = (roundedS - grp->S) >> grp->slot_shift; |
462dbc91 | 1347 | |
0545a303 | 1348 | i = (grp->front + offset) % QFQ_MAX_SLOTS; |
1349 | ||
462dbc91 | 1350 | hlist_del(&agg->next); |
0545a303 | 1351 | if (hlist_empty(&grp->slots[i])) |
1352 | __clear_bit(offset, &grp->full_slots); | |
1353 | } | |
1354 | ||
1355 | /* | |
462dbc91 PV |
1356 | * Called to forcibly deschedule an aggregate. If the aggregate is |
1357 | * not in the front bucket, or if the latter has other aggregates in | |
1358 | * the front bucket, we can simply remove the aggregate with no other | |
1359 | * side effects. | |
0545a303 | 1360 | * Otherwise we must propagate the event up. |
1361 | */ | |
462dbc91 | 1362 | static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg) |
0545a303 | 1363 | { |
462dbc91 | 1364 | struct qfq_group *grp = agg->grp; |
0545a303 | 1365 | unsigned long mask; |
1366 | u64 roundedS; | |
1367 | int s; | |
1368 | ||
462dbc91 PV |
1369 | if (agg == q->in_serv_agg) { |
1370 | charge_actual_service(agg); | |
1371 | q->in_serv_agg = qfq_choose_next_agg(q); | |
1372 | return; | |
1373 | } | |
1374 | ||
1375 | agg->F = agg->S; | |
1376 | qfq_slot_remove(q, grp, agg); | |
0545a303 | 1377 | |
1378 | if (!grp->full_slots) { | |
1379 | __clear_bit(grp->index, &q->bitmaps[IR]); | |
1380 | __clear_bit(grp->index, &q->bitmaps[EB]); | |
1381 | __clear_bit(grp->index, &q->bitmaps[IB]); | |
1382 | ||
1383 | if (test_bit(grp->index, &q->bitmaps[ER]) && | |
1384 | !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) { | |
1385 | mask = q->bitmaps[ER] & ((1UL << grp->index) - 1); | |
1386 | if (mask) | |
1387 | mask = ~((1UL << __fls(mask)) - 1); | |
1388 | else | |
1389 | mask = ~0UL; | |
1390 | qfq_move_groups(q, mask, EB, ER); | |
1391 | qfq_move_groups(q, mask, IB, IR); | |
1392 | } | |
1393 | __clear_bit(grp->index, &q->bitmaps[ER]); | |
1394 | } else if (hlist_empty(&grp->slots[grp->front])) { | |
462dbc91 PV |
1395 | agg = qfq_slot_scan(grp); |
1396 | roundedS = qfq_round_down(agg->S, grp->slot_shift); | |
0545a303 | 1397 | if (grp->S != roundedS) { |
1398 | __clear_bit(grp->index, &q->bitmaps[ER]); | |
1399 | __clear_bit(grp->index, &q->bitmaps[IR]); | |
1400 | __clear_bit(grp->index, &q->bitmaps[EB]); | |
1401 | __clear_bit(grp->index, &q->bitmaps[IB]); | |
1402 | grp->S = roundedS; | |
1403 | grp->F = roundedS + (2ULL << grp->slot_shift); | |
1404 | s = qfq_calc_state(q, grp); | |
1405 | __set_bit(grp->index, &q->bitmaps[s]); | |
1406 | } | |
1407 | } | |
0545a303 | 1408 | } |
1409 | ||
1410 | static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg) | |
1411 | { | |
1412 | struct qfq_sched *q = qdisc_priv(sch); | |
1413 | struct qfq_class *cl = (struct qfq_class *)arg; | |
1414 | ||
95946658 | 1415 | qfq_deactivate_class(q, cl); |
0545a303 | 1416 | } |
1417 | ||
e63d7dfd AA |
1418 | static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt, |
1419 | struct netlink_ext_ack *extack) | |
0545a303 | 1420 | { |
1421 | struct qfq_sched *q = qdisc_priv(sch); | |
1422 | struct qfq_group *grp; | |
1423 | int i, j, err; | |
462dbc91 | 1424 | u32 max_cl_shift, maxbudg_shift, max_classes; |
0545a303 | 1425 | |
69d78ef2 | 1426 | err = tcf_block_get(&q->block, &q->filter_list, sch); |
6529eaba JP |
1427 | if (err) |
1428 | return err; | |
1429 | ||
0545a303 | 1430 | err = qdisc_class_hash_init(&q->clhash); |
1431 | if (err < 0) | |
1432 | return err; | |
1433 | ||
462dbc91 PV |
1434 | if (qdisc_dev(sch)->tx_queue_len + 1 > QFQ_MAX_AGG_CLASSES) |
1435 | max_classes = QFQ_MAX_AGG_CLASSES; | |
1436 | else | |
1437 | max_classes = qdisc_dev(sch)->tx_queue_len + 1; | |
1438 | /* max_cl_shift = floor(log_2(max_classes)) */ | |
1439 | max_cl_shift = __fls(max_classes); | |
1440 | q->max_agg_classes = 1<<max_cl_shift; | |
1441 | ||
1442 | /* maxbudg_shift = log2(max_len * max_classes_per_agg) */ | |
1443 | maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift; | |
1444 | q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX; | |
1445 | ||
0545a303 | 1446 | for (i = 0; i <= QFQ_MAX_INDEX; i++) { |
1447 | grp = &q->groups[i]; | |
1448 | grp->index = i; | |
462dbc91 | 1449 | grp->slot_shift = q->min_slot_shift + i; |
0545a303 | 1450 | for (j = 0; j < QFQ_MAX_SLOTS; j++) |
1451 | INIT_HLIST_HEAD(&grp->slots[j]); | |
1452 | } | |
1453 | ||
462dbc91 PV |
1454 | INIT_HLIST_HEAD(&q->nonfull_aggs); |
1455 | ||
0545a303 | 1456 | return 0; |
1457 | } | |
1458 | ||
1459 | static void qfq_reset_qdisc(struct Qdisc *sch) | |
1460 | { | |
1461 | struct qfq_sched *q = qdisc_priv(sch); | |
0545a303 | 1462 | struct qfq_class *cl; |
462dbc91 | 1463 | unsigned int i; |
0545a303 | 1464 | |
462dbc91 | 1465 | for (i = 0; i < q->clhash.hashsize; i++) { |
b67bfe0d | 1466 | hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) { |
462dbc91 | 1467 | if (cl->qdisc->q.qlen > 0) |
0545a303 | 1468 | qfq_deactivate_class(q, cl); |
0545a303 | 1469 | |
0545a303 | 1470 | qdisc_reset(cl->qdisc); |
462dbc91 | 1471 | } |
0545a303 | 1472 | } |
2ed5c3f0 | 1473 | sch->qstats.backlog = 0; |
0545a303 | 1474 | sch->q.qlen = 0; |
1475 | } | |
1476 | ||
1477 | static void qfq_destroy_qdisc(struct Qdisc *sch) | |
1478 | { | |
1479 | struct qfq_sched *q = qdisc_priv(sch); | |
1480 | struct qfq_class *cl; | |
b67bfe0d | 1481 | struct hlist_node *next; |
0545a303 | 1482 | unsigned int i; |
1483 | ||
6529eaba | 1484 | tcf_block_put(q->block); |
0545a303 | 1485 | |
1486 | for (i = 0; i < q->clhash.hashsize; i++) { | |
b67bfe0d | 1487 | hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i], |
0545a303 | 1488 | common.hnode) { |
1489 | qfq_destroy_class(sch, cl); | |
1490 | } | |
1491 | } | |
1492 | qdisc_class_hash_destroy(&q->clhash); | |
1493 | } | |
1494 | ||
1495 | static const struct Qdisc_class_ops qfq_class_ops = { | |
1496 | .change = qfq_change_class, | |
1497 | .delete = qfq_delete_class, | |
143976ce | 1498 | .find = qfq_search_class, |
6529eaba | 1499 | .tcf_block = qfq_tcf_block, |
0545a303 | 1500 | .bind_tcf = qfq_bind_tcf, |
1501 | .unbind_tcf = qfq_unbind_tcf, | |
1502 | .graft = qfq_graft_class, | |
1503 | .leaf = qfq_class_leaf, | |
1504 | .qlen_notify = qfq_qlen_notify, | |
1505 | .dump = qfq_dump_class, | |
1506 | .dump_stats = qfq_dump_class_stats, | |
1507 | .walk = qfq_walk, | |
1508 | }; | |
1509 | ||
1510 | static struct Qdisc_ops qfq_qdisc_ops __read_mostly = { | |
1511 | .cl_ops = &qfq_class_ops, | |
1512 | .id = "qfq", | |
1513 | .priv_size = sizeof(struct qfq_sched), | |
1514 | .enqueue = qfq_enqueue, | |
1515 | .dequeue = qfq_dequeue, | |
1516 | .peek = qdisc_peek_dequeued, | |
0545a303 | 1517 | .init = qfq_init_qdisc, |
1518 | .reset = qfq_reset_qdisc, | |
1519 | .destroy = qfq_destroy_qdisc, | |
1520 | .owner = THIS_MODULE, | |
1521 | }; | |
1522 | ||
1523 | static int __init qfq_init(void) | |
1524 | { | |
1525 | return register_qdisc(&qfq_qdisc_ops); | |
1526 | } | |
1527 | ||
1528 | static void __exit qfq_exit(void) | |
1529 | { | |
1530 | unregister_qdisc(&qfq_qdisc_ops); | |
1531 | } | |
1532 | ||
1533 | module_init(qfq_init); | |
1534 | module_exit(qfq_exit); | |
1535 | MODULE_LICENSE("GPL"); |