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 | ||
135 | unsigned int refcnt; | |
136 | unsigned int filter_cnt; | |
137 | ||
138 | struct gnet_stats_basic_packed bstats; | |
139 | struct gnet_stats_queue qstats; | |
45203a3b | 140 | struct gnet_stats_rate_est64 rate_est; |
0545a303 | 141 | struct Qdisc *qdisc; |
462dbc91 PV |
142 | struct list_head alist; /* Link for active-classes list. */ |
143 | struct qfq_aggregate *agg; /* Parent aggregate. */ | |
144 | int deficit; /* DRR deficit counter. */ | |
145 | }; | |
0545a303 | 146 | |
462dbc91 | 147 | struct qfq_aggregate { |
0545a303 | 148 | struct hlist_node next; /* Link for the slot list. */ |
149 | u64 S, F; /* flow timestamps (exact) */ | |
150 | ||
151 | /* group we belong to. In principle we would need the index, | |
152 | * which is log_2(lmax/weight), but we never reference it | |
153 | * directly, only the group. | |
154 | */ | |
155 | struct qfq_group *grp; | |
156 | ||
157 | /* these are copied from the flowset. */ | |
462dbc91 PV |
158 | u32 class_weight; /* Weight of each class in this aggregate. */ |
159 | /* Max pkt size for the classes in this aggregate, DRR quantum. */ | |
160 | int lmax; | |
161 | ||
162 | u32 inv_w; /* ONE_FP/(sum of weights of classes in aggr.). */ | |
163 | u32 budgetmax; /* Max budget for this aggregate. */ | |
164 | u32 initial_budget, budget; /* Initial and current budget. */ | |
165 | ||
166 | int num_classes; /* Number of classes in this aggr. */ | |
167 | struct list_head active; /* DRR queue of active classes. */ | |
168 | ||
169 | struct hlist_node nonfull_next; /* See nonfull_aggs in qfq_sched. */ | |
0545a303 | 170 | }; |
171 | ||
172 | struct qfq_group { | |
173 | u64 S, F; /* group timestamps (approx). */ | |
174 | unsigned int slot_shift; /* Slot shift. */ | |
175 | unsigned int index; /* Group index. */ | |
176 | unsigned int front; /* Index of the front slot. */ | |
177 | unsigned long full_slots; /* non-empty slots */ | |
178 | ||
462dbc91 | 179 | /* Array of RR lists of active aggregates. */ |
0545a303 | 180 | struct hlist_head slots[QFQ_MAX_SLOTS]; |
181 | }; | |
182 | ||
183 | struct qfq_sched { | |
25d8c0d5 | 184 | struct tcf_proto __rcu *filter_list; |
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; | |
223 | ||
224 | qdisc_reset(cl->qdisc); | |
225 | qdisc_tree_decrease_qlen(cl->qdisc, len); | |
226 | } | |
227 | ||
228 | static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = { | |
229 | [TCA_QFQ_WEIGHT] = { .type = NLA_U32 }, | |
230 | [TCA_QFQ_LMAX] = { .type = NLA_U32 }, | |
231 | }; | |
232 | ||
233 | /* | |
234 | * Calculate a flow index, given its weight and maximum packet length. | |
235 | * index = log_2(maxlen/weight) but we need to apply the scaling. | |
236 | * This is used only once at flow creation. | |
237 | */ | |
462dbc91 | 238 | static int qfq_calc_index(u32 inv_w, unsigned int maxlen, u32 min_slot_shift) |
0545a303 | 239 | { |
240 | u64 slot_size = (u64)maxlen * inv_w; | |
241 | unsigned long size_map; | |
242 | int index = 0; | |
243 | ||
462dbc91 | 244 | size_map = slot_size >> min_slot_shift; |
0545a303 | 245 | if (!size_map) |
246 | goto out; | |
247 | ||
248 | index = __fls(size_map) + 1; /* basically a log_2 */ | |
462dbc91 | 249 | index -= !(slot_size - (1ULL << (index + min_slot_shift - 1))); |
0545a303 | 250 | |
251 | if (index < 0) | |
252 | index = 0; | |
253 | out: | |
254 | pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n", | |
255 | (unsigned long) ONE_FP/inv_w, maxlen, index); | |
256 | ||
257 | return index; | |
258 | } | |
259 | ||
462dbc91 PV |
260 | static void qfq_deactivate_agg(struct qfq_sched *, struct qfq_aggregate *); |
261 | static void qfq_activate_agg(struct qfq_sched *, struct qfq_aggregate *, | |
262 | enum update_reason); | |
263 | ||
264 | static void qfq_init_agg(struct qfq_sched *q, struct qfq_aggregate *agg, | |
265 | u32 lmax, u32 weight) | |
be72f63b | 266 | { |
462dbc91 PV |
267 | INIT_LIST_HEAD(&agg->active); |
268 | hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs); | |
269 | ||
270 | agg->lmax = lmax; | |
271 | agg->class_weight = weight; | |
272 | } | |
273 | ||
274 | static struct qfq_aggregate *qfq_find_agg(struct qfq_sched *q, | |
275 | u32 lmax, u32 weight) | |
276 | { | |
277 | struct qfq_aggregate *agg; | |
462dbc91 | 278 | |
b67bfe0d | 279 | hlist_for_each_entry(agg, &q->nonfull_aggs, nonfull_next) |
462dbc91 PV |
280 | if (agg->lmax == lmax && agg->class_weight == weight) |
281 | return agg; | |
282 | ||
283 | return NULL; | |
284 | } | |
285 | ||
be72f63b | 286 | |
462dbc91 PV |
287 | /* Update aggregate as a function of the new number of classes. */ |
288 | static void qfq_update_agg(struct qfq_sched *q, struct qfq_aggregate *agg, | |
289 | int new_num_classes) | |
290 | { | |
291 | u32 new_agg_weight; | |
292 | ||
293 | if (new_num_classes == q->max_agg_classes) | |
294 | hlist_del_init(&agg->nonfull_next); | |
295 | ||
296 | if (agg->num_classes > new_num_classes && | |
297 | new_num_classes == q->max_agg_classes - 1) /* agg no more full */ | |
298 | hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs); | |
299 | ||
9b99b7e9 PV |
300 | /* The next assignment may let |
301 | * agg->initial_budget > agg->budgetmax | |
302 | * hold, we will take it into account in charge_actual_service(). | |
303 | */ | |
462dbc91 PV |
304 | agg->budgetmax = new_num_classes * agg->lmax; |
305 | new_agg_weight = agg->class_weight * new_num_classes; | |
306 | agg->inv_w = ONE_FP/new_agg_weight; | |
307 | ||
308 | if (agg->grp == NULL) { | |
309 | int i = qfq_calc_index(agg->inv_w, agg->budgetmax, | |
310 | q->min_slot_shift); | |
311 | agg->grp = &q->groups[i]; | |
312 | } | |
313 | ||
314 | q->wsum += | |
315 | (int) agg->class_weight * (new_num_classes - agg->num_classes); | |
87f40dd6 | 316 | q->iwsum = ONE_FP / q->wsum; |
462dbc91 PV |
317 | |
318 | agg->num_classes = new_num_classes; | |
319 | } | |
320 | ||
321 | /* Add class to aggregate. */ | |
322 | static void qfq_add_to_agg(struct qfq_sched *q, | |
323 | struct qfq_aggregate *agg, | |
324 | struct qfq_class *cl) | |
325 | { | |
326 | cl->agg = agg; | |
327 | ||
328 | qfq_update_agg(q, agg, agg->num_classes+1); | |
329 | if (cl->qdisc->q.qlen > 0) { /* adding an active class */ | |
330 | list_add_tail(&cl->alist, &agg->active); | |
331 | if (list_first_entry(&agg->active, struct qfq_class, alist) == | |
332 | cl && q->in_serv_agg != agg) /* agg was inactive */ | |
333 | qfq_activate_agg(q, agg, enqueue); /* schedule agg */ | |
334 | } | |
be72f63b PV |
335 | } |
336 | ||
462dbc91 | 337 | static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *); |
be72f63b | 338 | |
462dbc91 | 339 | static void qfq_destroy_agg(struct qfq_sched *q, struct qfq_aggregate *agg) |
be72f63b | 340 | { |
a55e1c5c | 341 | hlist_del_init(&agg->nonfull_next); |
87f40dd6 PV |
342 | q->wsum -= agg->class_weight; |
343 | if (q->wsum != 0) | |
344 | q->iwsum = ONE_FP / q->wsum; | |
345 | ||
462dbc91 PV |
346 | if (q->in_serv_agg == agg) |
347 | q->in_serv_agg = qfq_choose_next_agg(q); | |
348 | kfree(agg); | |
349 | } | |
be72f63b | 350 | |
462dbc91 PV |
351 | /* Deschedule class from within its parent aggregate. */ |
352 | static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl) | |
353 | { | |
354 | struct qfq_aggregate *agg = cl->agg; | |
be72f63b | 355 | |
be72f63b | 356 | |
462dbc91 PV |
357 | list_del(&cl->alist); /* remove from RR queue of the aggregate */ |
358 | if (list_empty(&agg->active)) /* agg is now inactive */ | |
359 | qfq_deactivate_agg(q, agg); | |
be72f63b PV |
360 | } |
361 | ||
462dbc91 PV |
362 | /* Remove class from its parent aggregate. */ |
363 | static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl) | |
3015f3d2 | 364 | { |
462dbc91 | 365 | struct qfq_aggregate *agg = cl->agg; |
3015f3d2 | 366 | |
462dbc91 PV |
367 | cl->agg = NULL; |
368 | if (agg->num_classes == 1) { /* agg being emptied, destroy it */ | |
369 | qfq_destroy_agg(q, agg); | |
370 | return; | |
3015f3d2 | 371 | } |
462dbc91 PV |
372 | qfq_update_agg(q, agg, agg->num_classes-1); |
373 | } | |
3015f3d2 | 374 | |
462dbc91 PV |
375 | /* Deschedule class and remove it from its parent aggregate. */ |
376 | static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl) | |
377 | { | |
378 | if (cl->qdisc->q.qlen > 0) /* class is active */ | |
379 | qfq_deactivate_class(q, cl); | |
3015f3d2 | 380 | |
462dbc91 | 381 | qfq_rm_from_agg(q, cl); |
3015f3d2 PV |
382 | } |
383 | ||
462dbc91 PV |
384 | /* Move class to a new aggregate, matching the new class weight and/or lmax */ |
385 | static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight, | |
386 | u32 lmax) | |
387 | { | |
388 | struct qfq_sched *q = qdisc_priv(sch); | |
389 | struct qfq_aggregate *new_agg = qfq_find_agg(q, lmax, weight); | |
390 | ||
391 | if (new_agg == NULL) { /* create new aggregate */ | |
392 | new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC); | |
393 | if (new_agg == NULL) | |
394 | return -ENOBUFS; | |
395 | qfq_init_agg(q, new_agg, lmax, weight); | |
396 | } | |
397 | qfq_deact_rm_from_agg(q, cl); | |
398 | qfq_add_to_agg(q, new_agg, cl); | |
399 | ||
400 | return 0; | |
401 | } | |
3015f3d2 | 402 | |
0545a303 | 403 | static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid, |
404 | struct nlattr **tca, unsigned long *arg) | |
405 | { | |
406 | struct qfq_sched *q = qdisc_priv(sch); | |
407 | struct qfq_class *cl = (struct qfq_class *)*arg; | |
462dbc91 | 408 | bool existing = false; |
0545a303 | 409 | struct nlattr *tb[TCA_QFQ_MAX + 1]; |
462dbc91 | 410 | struct qfq_aggregate *new_agg = NULL; |
0545a303 | 411 | u32 weight, lmax, inv_w; |
3015f3d2 | 412 | int err; |
d32ae76f | 413 | int delta_w; |
0545a303 | 414 | |
415 | if (tca[TCA_OPTIONS] == NULL) { | |
416 | pr_notice("qfq: no options\n"); | |
417 | return -EINVAL; | |
418 | } | |
419 | ||
420 | err = nla_parse_nested(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS], qfq_policy); | |
421 | if (err < 0) | |
422 | return err; | |
423 | ||
424 | if (tb[TCA_QFQ_WEIGHT]) { | |
425 | weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]); | |
426 | if (!weight || weight > (1UL << QFQ_MAX_WSHIFT)) { | |
427 | pr_notice("qfq: invalid weight %u\n", weight); | |
428 | return -EINVAL; | |
429 | } | |
430 | } else | |
431 | weight = 1; | |
432 | ||
0545a303 | 433 | if (tb[TCA_QFQ_LMAX]) { |
434 | lmax = nla_get_u32(tb[TCA_QFQ_LMAX]); | |
3015f3d2 | 435 | if (lmax < QFQ_MIN_LMAX || lmax > (1UL << QFQ_MTU_SHIFT)) { |
0545a303 | 436 | pr_notice("qfq: invalid max length %u\n", lmax); |
437 | return -EINVAL; | |
438 | } | |
439 | } else | |
3015f3d2 | 440 | lmax = psched_mtu(qdisc_dev(sch)); |
0545a303 | 441 | |
462dbc91 PV |
442 | inv_w = ONE_FP / weight; |
443 | weight = ONE_FP / inv_w; | |
444 | ||
445 | if (cl != NULL && | |
446 | lmax == cl->agg->lmax && | |
447 | weight == cl->agg->class_weight) | |
448 | return 0; /* nothing to change */ | |
449 | ||
450 | delta_w = weight - (cl ? cl->agg->class_weight : 0); | |
451 | ||
452 | if (q->wsum + delta_w > QFQ_MAX_WSUM) { | |
453 | pr_notice("qfq: total weight out of range (%d + %u)\n", | |
454 | delta_w, q->wsum); | |
455 | return -EINVAL; | |
456 | } | |
457 | ||
458 | if (cl != NULL) { /* modify existing class */ | |
0545a303 | 459 | if (tca[TCA_RATE]) { |
22e0f8b9 JF |
460 | err = gen_replace_estimator(&cl->bstats, NULL, |
461 | &cl->rate_est, | |
0545a303 | 462 | qdisc_root_sleeping_lock(sch), |
463 | tca[TCA_RATE]); | |
464 | if (err) | |
465 | return err; | |
466 | } | |
462dbc91 PV |
467 | existing = true; |
468 | goto set_change_agg; | |
0545a303 | 469 | } |
470 | ||
462dbc91 | 471 | /* create and init new class */ |
0545a303 | 472 | cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL); |
473 | if (cl == NULL) | |
474 | return -ENOBUFS; | |
475 | ||
476 | cl->refcnt = 1; | |
477 | cl->common.classid = classid; | |
462dbc91 | 478 | cl->deficit = lmax; |
0545a303 | 479 | |
480 | cl->qdisc = qdisc_create_dflt(sch->dev_queue, | |
481 | &pfifo_qdisc_ops, classid); | |
482 | if (cl->qdisc == NULL) | |
483 | cl->qdisc = &noop_qdisc; | |
484 | ||
485 | if (tca[TCA_RATE]) { | |
22e0f8b9 JF |
486 | err = gen_new_estimator(&cl->bstats, NULL, |
487 | &cl->rate_est, | |
0545a303 | 488 | qdisc_root_sleeping_lock(sch), |
489 | tca[TCA_RATE]); | |
462dbc91 PV |
490 | if (err) |
491 | goto destroy_class; | |
0545a303 | 492 | } |
493 | ||
494 | sch_tree_lock(sch); | |
495 | qdisc_class_hash_insert(&q->clhash, &cl->common); | |
496 | sch_tree_unlock(sch); | |
497 | ||
498 | qdisc_class_hash_grow(sch, &q->clhash); | |
499 | ||
462dbc91 PV |
500 | set_change_agg: |
501 | sch_tree_lock(sch); | |
502 | new_agg = qfq_find_agg(q, lmax, weight); | |
503 | if (new_agg == NULL) { /* create new aggregate */ | |
504 | sch_tree_unlock(sch); | |
505 | new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL); | |
506 | if (new_agg == NULL) { | |
507 | err = -ENOBUFS; | |
508 | gen_kill_estimator(&cl->bstats, &cl->rate_est); | |
509 | goto destroy_class; | |
510 | } | |
511 | sch_tree_lock(sch); | |
512 | qfq_init_agg(q, new_agg, lmax, weight); | |
513 | } | |
514 | if (existing) | |
515 | qfq_deact_rm_from_agg(q, cl); | |
516 | qfq_add_to_agg(q, new_agg, cl); | |
517 | sch_tree_unlock(sch); | |
518 | ||
0545a303 | 519 | *arg = (unsigned long)cl; |
520 | return 0; | |
462dbc91 PV |
521 | |
522 | destroy_class: | |
523 | qdisc_destroy(cl->qdisc); | |
524 | kfree(cl); | |
525 | return err; | |
0545a303 | 526 | } |
527 | ||
528 | static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl) | |
529 | { | |
530 | struct qfq_sched *q = qdisc_priv(sch); | |
531 | ||
462dbc91 | 532 | qfq_rm_from_agg(q, cl); |
0545a303 | 533 | gen_kill_estimator(&cl->bstats, &cl->rate_est); |
534 | qdisc_destroy(cl->qdisc); | |
535 | kfree(cl); | |
536 | } | |
537 | ||
538 | static int qfq_delete_class(struct Qdisc *sch, unsigned long arg) | |
539 | { | |
540 | struct qfq_sched *q = qdisc_priv(sch); | |
541 | struct qfq_class *cl = (struct qfq_class *)arg; | |
542 | ||
543 | if (cl->filter_cnt > 0) | |
544 | return -EBUSY; | |
545 | ||
546 | sch_tree_lock(sch); | |
547 | ||
548 | qfq_purge_queue(cl); | |
549 | qdisc_class_hash_remove(&q->clhash, &cl->common); | |
550 | ||
551 | BUG_ON(--cl->refcnt == 0); | |
552 | /* | |
553 | * This shouldn't happen: we "hold" one cops->get() when called | |
554 | * from tc_ctl_tclass; the destroy method is done from cops->put(). | |
555 | */ | |
556 | ||
557 | sch_tree_unlock(sch); | |
558 | return 0; | |
559 | } | |
560 | ||
561 | static unsigned long qfq_get_class(struct Qdisc *sch, u32 classid) | |
562 | { | |
563 | struct qfq_class *cl = qfq_find_class(sch, classid); | |
564 | ||
565 | if (cl != NULL) | |
566 | cl->refcnt++; | |
567 | ||
568 | return (unsigned long)cl; | |
569 | } | |
570 | ||
571 | static void qfq_put_class(struct Qdisc *sch, unsigned long arg) | |
572 | { | |
573 | struct qfq_class *cl = (struct qfq_class *)arg; | |
574 | ||
575 | if (--cl->refcnt == 0) | |
576 | qfq_destroy_class(sch, cl); | |
577 | } | |
578 | ||
25d8c0d5 JF |
579 | static struct tcf_proto __rcu **qfq_tcf_chain(struct Qdisc *sch, |
580 | unsigned long cl) | |
0545a303 | 581 | { |
582 | struct qfq_sched *q = qdisc_priv(sch); | |
583 | ||
584 | if (cl) | |
585 | return NULL; | |
586 | ||
587 | return &q->filter_list; | |
588 | } | |
589 | ||
590 | static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent, | |
591 | u32 classid) | |
592 | { | |
593 | struct qfq_class *cl = qfq_find_class(sch, classid); | |
594 | ||
595 | if (cl != NULL) | |
596 | cl->filter_cnt++; | |
597 | ||
598 | return (unsigned long)cl; | |
599 | } | |
600 | ||
601 | static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg) | |
602 | { | |
603 | struct qfq_class *cl = (struct qfq_class *)arg; | |
604 | ||
605 | cl->filter_cnt--; | |
606 | } | |
607 | ||
608 | static int qfq_graft_class(struct Qdisc *sch, unsigned long arg, | |
609 | struct Qdisc *new, struct Qdisc **old) | |
610 | { | |
611 | struct qfq_class *cl = (struct qfq_class *)arg; | |
612 | ||
613 | if (new == NULL) { | |
614 | new = qdisc_create_dflt(sch->dev_queue, | |
615 | &pfifo_qdisc_ops, cl->common.classid); | |
616 | if (new == NULL) | |
617 | new = &noop_qdisc; | |
618 | } | |
619 | ||
86a7996c | 620 | *old = qdisc_replace(sch, new, &cl->qdisc); |
0545a303 | 621 | return 0; |
622 | } | |
623 | ||
624 | static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg) | |
625 | { | |
626 | struct qfq_class *cl = (struct qfq_class *)arg; | |
627 | ||
628 | return cl->qdisc; | |
629 | } | |
630 | ||
631 | static int qfq_dump_class(struct Qdisc *sch, unsigned long arg, | |
632 | struct sk_buff *skb, struct tcmsg *tcm) | |
633 | { | |
634 | struct qfq_class *cl = (struct qfq_class *)arg; | |
635 | struct nlattr *nest; | |
636 | ||
637 | tcm->tcm_parent = TC_H_ROOT; | |
638 | tcm->tcm_handle = cl->common.classid; | |
639 | tcm->tcm_info = cl->qdisc->handle; | |
640 | ||
641 | nest = nla_nest_start(skb, TCA_OPTIONS); | |
642 | if (nest == NULL) | |
643 | goto nla_put_failure; | |
462dbc91 PV |
644 | if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) || |
645 | nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax)) | |
1b34ec43 | 646 | goto nla_put_failure; |
0545a303 | 647 | return nla_nest_end(skb, nest); |
648 | ||
649 | nla_put_failure: | |
650 | nla_nest_cancel(skb, nest); | |
651 | return -EMSGSIZE; | |
652 | } | |
653 | ||
654 | static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg, | |
655 | struct gnet_dump *d) | |
656 | { | |
657 | struct qfq_class *cl = (struct qfq_class *)arg; | |
658 | struct tc_qfq_stats xstats; | |
659 | ||
660 | memset(&xstats, 0, sizeof(xstats)); | |
0545a303 | 661 | |
462dbc91 PV |
662 | xstats.weight = cl->agg->class_weight; |
663 | xstats.lmax = cl->agg->lmax; | |
0545a303 | 664 | |
22e0f8b9 | 665 | if (gnet_stats_copy_basic(d, NULL, &cl->bstats) < 0 || |
0545a303 | 666 | gnet_stats_copy_rate_est(d, &cl->bstats, &cl->rate_est) < 0 || |
b0ab6f92 JF |
667 | gnet_stats_copy_queue(d, NULL, |
668 | &cl->qdisc->qstats, cl->qdisc->q.qlen) < 0) | |
0545a303 | 669 | return -1; |
670 | ||
671 | return gnet_stats_copy_app(d, &xstats, sizeof(xstats)); | |
672 | } | |
673 | ||
674 | static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg) | |
675 | { | |
676 | struct qfq_sched *q = qdisc_priv(sch); | |
677 | struct qfq_class *cl; | |
0545a303 | 678 | unsigned int i; |
679 | ||
680 | if (arg->stop) | |
681 | return; | |
682 | ||
683 | for (i = 0; i < q->clhash.hashsize; i++) { | |
b67bfe0d | 684 | hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) { |
0545a303 | 685 | if (arg->count < arg->skip) { |
686 | arg->count++; | |
687 | continue; | |
688 | } | |
689 | if (arg->fn(sch, (unsigned long)cl, arg) < 0) { | |
690 | arg->stop = 1; | |
691 | return; | |
692 | } | |
693 | arg->count++; | |
694 | } | |
695 | } | |
696 | } | |
697 | ||
698 | static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch, | |
699 | int *qerr) | |
700 | { | |
701 | struct qfq_sched *q = qdisc_priv(sch); | |
702 | struct qfq_class *cl; | |
703 | struct tcf_result res; | |
25d8c0d5 | 704 | struct tcf_proto *fl; |
0545a303 | 705 | int result; |
706 | ||
707 | if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) { | |
708 | pr_debug("qfq_classify: found %d\n", skb->priority); | |
709 | cl = qfq_find_class(sch, skb->priority); | |
710 | if (cl != NULL) | |
711 | return cl; | |
712 | } | |
713 | ||
714 | *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS; | |
25d8c0d5 | 715 | fl = rcu_dereference_bh(q->filter_list); |
3b3ae880 | 716 | result = tc_classify(skb, fl, &res, false); |
0545a303 | 717 | if (result >= 0) { |
718 | #ifdef CONFIG_NET_CLS_ACT | |
719 | switch (result) { | |
720 | case TC_ACT_QUEUED: | |
721 | case TC_ACT_STOLEN: | |
722 | *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN; | |
723 | case TC_ACT_SHOT: | |
724 | return NULL; | |
725 | } | |
726 | #endif | |
727 | cl = (struct qfq_class *)res.class; | |
728 | if (cl == NULL) | |
729 | cl = qfq_find_class(sch, res.classid); | |
730 | return cl; | |
731 | } | |
732 | ||
733 | return NULL; | |
734 | } | |
735 | ||
736 | /* Generic comparison function, handling wraparound. */ | |
737 | static inline int qfq_gt(u64 a, u64 b) | |
738 | { | |
739 | return (s64)(a - b) > 0; | |
740 | } | |
741 | ||
742 | /* Round a precise timestamp to its slotted value. */ | |
743 | static inline u64 qfq_round_down(u64 ts, unsigned int shift) | |
744 | { | |
745 | return ts & ~((1ULL << shift) - 1); | |
746 | } | |
747 | ||
748 | /* return the pointer to the group with lowest index in the bitmap */ | |
749 | static inline struct qfq_group *qfq_ffs(struct qfq_sched *q, | |
750 | unsigned long bitmap) | |
751 | { | |
752 | int index = __ffs(bitmap); | |
753 | return &q->groups[index]; | |
754 | } | |
755 | /* Calculate a mask to mimic what would be ffs_from(). */ | |
756 | static inline unsigned long mask_from(unsigned long bitmap, int from) | |
757 | { | |
758 | return bitmap & ~((1UL << from) - 1); | |
759 | } | |
760 | ||
761 | /* | |
762 | * The state computation relies on ER=0, IR=1, EB=2, IB=3 | |
763 | * First compute eligibility comparing grp->S, q->V, | |
764 | * then check if someone is blocking us and possibly add EB | |
765 | */ | |
766 | static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp) | |
767 | { | |
768 | /* if S > V we are not eligible */ | |
769 | unsigned int state = qfq_gt(grp->S, q->V); | |
770 | unsigned long mask = mask_from(q->bitmaps[ER], grp->index); | |
771 | struct qfq_group *next; | |
772 | ||
773 | if (mask) { | |
774 | next = qfq_ffs(q, mask); | |
775 | if (qfq_gt(grp->F, next->F)) | |
776 | state |= EB; | |
777 | } | |
778 | ||
779 | return state; | |
780 | } | |
781 | ||
782 | ||
783 | /* | |
784 | * In principle | |
785 | * q->bitmaps[dst] |= q->bitmaps[src] & mask; | |
786 | * q->bitmaps[src] &= ~mask; | |
787 | * but we should make sure that src != dst | |
788 | */ | |
789 | static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask, | |
790 | int src, int dst) | |
791 | { | |
792 | q->bitmaps[dst] |= q->bitmaps[src] & mask; | |
793 | q->bitmaps[src] &= ~mask; | |
794 | } | |
795 | ||
796 | static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F) | |
797 | { | |
798 | unsigned long mask = mask_from(q->bitmaps[ER], index + 1); | |
799 | struct qfq_group *next; | |
800 | ||
801 | if (mask) { | |
802 | next = qfq_ffs(q, mask); | |
803 | if (!qfq_gt(next->F, old_F)) | |
804 | return; | |
805 | } | |
806 | ||
807 | mask = (1UL << index) - 1; | |
808 | qfq_move_groups(q, mask, EB, ER); | |
809 | qfq_move_groups(q, mask, IB, IR); | |
810 | } | |
811 | ||
812 | /* | |
813 | * perhaps | |
814 | * | |
815 | old_V ^= q->V; | |
462dbc91 | 816 | old_V >>= q->min_slot_shift; |
0545a303 | 817 | if (old_V) { |
818 | ... | |
819 | } | |
820 | * | |
821 | */ | |
462dbc91 | 822 | static void qfq_make_eligible(struct qfq_sched *q) |
0545a303 | 823 | { |
462dbc91 PV |
824 | unsigned long vslot = q->V >> q->min_slot_shift; |
825 | unsigned long old_vslot = q->oldV >> q->min_slot_shift; | |
0545a303 | 826 | |
827 | if (vslot != old_vslot) { | |
87f1369d PV |
828 | unsigned long mask; |
829 | int last_flip_pos = fls(vslot ^ old_vslot); | |
830 | ||
831 | if (last_flip_pos > 31) /* higher than the number of groups */ | |
832 | mask = ~0UL; /* make all groups eligible */ | |
833 | else | |
834 | mask = (1UL << last_flip_pos) - 1; | |
835 | ||
0545a303 | 836 | qfq_move_groups(q, mask, IR, ER); |
837 | qfq_move_groups(q, mask, IB, EB); | |
838 | } | |
839 | } | |
840 | ||
0545a303 | 841 | /* |
87f40dd6 PV |
842 | * The index of the slot in which the input aggregate agg is to be |
843 | * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2' | |
844 | * and not a '-1' because the start time of the group may be moved | |
845 | * backward by one slot after the aggregate has been inserted, and | |
846 | * this would cause non-empty slots to be right-shifted by one | |
847 | * position. | |
848 | * | |
849 | * QFQ+ fully satisfies this bound to the slot index if the parameters | |
850 | * of the classes are not changed dynamically, and if QFQ+ never | |
851 | * happens to postpone the service of agg unjustly, i.e., it never | |
852 | * happens that the aggregate becomes backlogged and eligible, or just | |
853 | * eligible, while an aggregate with a higher approximated finish time | |
854 | * is being served. In particular, in this case QFQ+ guarantees that | |
855 | * the timestamps of agg are low enough that the slot index is never | |
856 | * higher than 2. Unfortunately, QFQ+ cannot provide the same | |
857 | * guarantee if it happens to unjustly postpone the service of agg, or | |
858 | * if the parameters of some class are changed. | |
859 | * | |
860 | * As for the first event, i.e., an out-of-order service, the | |
861 | * upper bound to the slot index guaranteed by QFQ+ grows to | |
862 | * 2 + | |
863 | * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) * | |
864 | * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1. | |
3015f3d2 | 865 | * |
87f40dd6 PV |
866 | * The following function deals with this problem by backward-shifting |
867 | * the timestamps of agg, if needed, so as to guarantee that the slot | |
868 | * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may | |
869 | * cause the service of other aggregates to be postponed, yet the | |
870 | * worst-case guarantees of these aggregates are not violated. In | |
871 | * fact, in case of no out-of-order service, the timestamps of agg | |
872 | * would have been even lower than they are after the backward shift, | |
873 | * because QFQ+ would have guaranteed a maximum value equal to 2 for | |
874 | * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose | |
875 | * service is postponed because of the backward-shift would have | |
876 | * however waited for the service of agg before being served. | |
3015f3d2 | 877 | * |
87f40dd6 PV |
878 | * The other event that may cause the slot index to be higher than 2 |
879 | * for agg is a recent change of the parameters of some class. If the | |
880 | * weight of a class is increased or the lmax (max_pkt_size) of the | |
881 | * class is decreased, then a new aggregate with smaller slot size | |
882 | * than the original parent aggregate of the class may happen to be | |
883 | * activated. The activation of this aggregate should be properly | |
884 | * delayed to when the service of the class has finished in the ideal | |
885 | * system tracked by QFQ+. If the activation of the aggregate is not | |
886 | * delayed to this reference time instant, then this aggregate may be | |
887 | * unjustly served before other aggregates waiting for service. This | |
888 | * may cause the above bound to the slot index to be violated for some | |
889 | * of these unlucky aggregates. | |
3015f3d2 | 890 | * |
462dbc91 | 891 | * Instead of delaying the activation of the new aggregate, which is |
87f40dd6 PV |
892 | * quite complex, the above-discussed capping of the slot index is |
893 | * used to handle also the consequences of a change of the parameters | |
894 | * of a class. | |
0545a303 | 895 | */ |
462dbc91 | 896 | static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg, |
0545a303 | 897 | u64 roundedS) |
898 | { | |
899 | u64 slot = (roundedS - grp->S) >> grp->slot_shift; | |
3015f3d2 PV |
900 | unsigned int i; /* slot index in the bucket list */ |
901 | ||
902 | if (unlikely(slot > QFQ_MAX_SLOTS - 2)) { | |
903 | u64 deltaS = roundedS - grp->S - | |
904 | ((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift); | |
462dbc91 PV |
905 | agg->S -= deltaS; |
906 | agg->F -= deltaS; | |
3015f3d2 PV |
907 | slot = QFQ_MAX_SLOTS - 2; |
908 | } | |
909 | ||
910 | i = (grp->front + slot) % QFQ_MAX_SLOTS; | |
0545a303 | 911 | |
462dbc91 | 912 | hlist_add_head(&agg->next, &grp->slots[i]); |
0545a303 | 913 | __set_bit(slot, &grp->full_slots); |
914 | } | |
915 | ||
916 | /* Maybe introduce hlist_first_entry?? */ | |
462dbc91 | 917 | static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp) |
0545a303 | 918 | { |
919 | return hlist_entry(grp->slots[grp->front].first, | |
462dbc91 | 920 | struct qfq_aggregate, next); |
0545a303 | 921 | } |
922 | ||
923 | /* | |
924 | * remove the entry from the slot | |
925 | */ | |
926 | static void qfq_front_slot_remove(struct qfq_group *grp) | |
927 | { | |
462dbc91 | 928 | struct qfq_aggregate *agg = qfq_slot_head(grp); |
0545a303 | 929 | |
462dbc91 PV |
930 | BUG_ON(!agg); |
931 | hlist_del(&agg->next); | |
0545a303 | 932 | if (hlist_empty(&grp->slots[grp->front])) |
933 | __clear_bit(0, &grp->full_slots); | |
934 | } | |
935 | ||
936 | /* | |
462dbc91 PV |
937 | * Returns the first aggregate in the first non-empty bucket of the |
938 | * group. As a side effect, adjusts the bucket list so the first | |
939 | * non-empty bucket is at position 0 in full_slots. | |
0545a303 | 940 | */ |
462dbc91 | 941 | static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp) |
0545a303 | 942 | { |
943 | unsigned int i; | |
944 | ||
945 | pr_debug("qfq slot_scan: grp %u full %#lx\n", | |
946 | grp->index, grp->full_slots); | |
947 | ||
948 | if (grp->full_slots == 0) | |
949 | return NULL; | |
950 | ||
951 | i = __ffs(grp->full_slots); /* zero based */ | |
952 | if (i > 0) { | |
953 | grp->front = (grp->front + i) % QFQ_MAX_SLOTS; | |
954 | grp->full_slots >>= i; | |
955 | } | |
956 | ||
957 | return qfq_slot_head(grp); | |
958 | } | |
959 | ||
960 | /* | |
961 | * adjust the bucket list. When the start time of a group decreases, | |
962 | * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to | |
963 | * move the objects. The mask of occupied slots must be shifted | |
964 | * because we use ffs() to find the first non-empty slot. | |
965 | * This covers decreases in the group's start time, but what about | |
966 | * increases of the start time ? | |
967 | * Here too we should make sure that i is less than 32 | |
968 | */ | |
969 | static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS) | |
970 | { | |
971 | unsigned int i = (grp->S - roundedS) >> grp->slot_shift; | |
972 | ||
973 | grp->full_slots <<= i; | |
974 | grp->front = (grp->front - i) % QFQ_MAX_SLOTS; | |
975 | } | |
976 | ||
462dbc91 | 977 | static void qfq_update_eligible(struct qfq_sched *q) |
0545a303 | 978 | { |
979 | struct qfq_group *grp; | |
980 | unsigned long ineligible; | |
981 | ||
982 | ineligible = q->bitmaps[IR] | q->bitmaps[IB]; | |
983 | if (ineligible) { | |
984 | if (!q->bitmaps[ER]) { | |
985 | grp = qfq_ffs(q, ineligible); | |
986 | if (qfq_gt(grp->S, q->V)) | |
987 | q->V = grp->S; | |
988 | } | |
462dbc91 | 989 | qfq_make_eligible(q); |
0545a303 | 990 | } |
991 | } | |
992 | ||
462dbc91 PV |
993 | /* Dequeue head packet of the head class in the DRR queue of the aggregate. */ |
994 | static void agg_dequeue(struct qfq_aggregate *agg, | |
995 | struct qfq_class *cl, unsigned int len) | |
0545a303 | 996 | { |
462dbc91 | 997 | qdisc_dequeue_peeked(cl->qdisc); |
0545a303 | 998 | |
462dbc91 | 999 | cl->deficit -= (int) len; |
0545a303 | 1000 | |
462dbc91 PV |
1001 | if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */ |
1002 | list_del(&cl->alist); | |
1003 | else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) { | |
1004 | cl->deficit += agg->lmax; | |
1005 | list_move_tail(&cl->alist, &agg->active); | |
0545a303 | 1006 | } |
462dbc91 PV |
1007 | } |
1008 | ||
1009 | static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg, | |
1010 | struct qfq_class **cl, | |
1011 | unsigned int *len) | |
1012 | { | |
1013 | struct sk_buff *skb; | |
0545a303 | 1014 | |
462dbc91 PV |
1015 | *cl = list_first_entry(&agg->active, struct qfq_class, alist); |
1016 | skb = (*cl)->qdisc->ops->peek((*cl)->qdisc); | |
1017 | if (skb == NULL) | |
1018 | WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n"); | |
1019 | else | |
1020 | *len = qdisc_pkt_len(skb); | |
1021 | ||
1022 | return skb; | |
1023 | } | |
1024 | ||
1025 | /* Update F according to the actual service received by the aggregate. */ | |
1026 | static inline void charge_actual_service(struct qfq_aggregate *agg) | |
1027 | { | |
9b99b7e9 PV |
1028 | /* Compute the service received by the aggregate, taking into |
1029 | * account that, after decreasing the number of classes in | |
1030 | * agg, it may happen that | |
1031 | * agg->initial_budget - agg->budget > agg->bugdetmax | |
1032 | */ | |
1033 | u32 service_received = min(agg->budgetmax, | |
1034 | agg->initial_budget - agg->budget); | |
462dbc91 PV |
1035 | |
1036 | agg->F = agg->S + (u64)service_received * agg->inv_w; | |
0545a303 | 1037 | } |
1038 | ||
88d4f419 PV |
1039 | /* Assign a reasonable start time for a new aggregate in group i. |
1040 | * Admissible values for \hat(F) are multiples of \sigma_i | |
1041 | * no greater than V+\sigma_i . Larger values mean that | |
1042 | * we had a wraparound so we consider the timestamp to be stale. | |
1043 | * | |
1044 | * If F is not stale and F >= V then we set S = F. | |
1045 | * Otherwise we should assign S = V, but this may violate | |
1046 | * the ordering in EB (see [2]). So, if we have groups in ER, | |
1047 | * set S to the F_j of the first group j which would be blocking us. | |
1048 | * We are guaranteed not to move S backward because | |
1049 | * otherwise our group i would still be blocked. | |
1050 | */ | |
1051 | static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg) | |
1052 | { | |
1053 | unsigned long mask; | |
1054 | u64 limit, roundedF; | |
1055 | int slot_shift = agg->grp->slot_shift; | |
1056 | ||
1057 | roundedF = qfq_round_down(agg->F, slot_shift); | |
1058 | limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift); | |
1059 | ||
1060 | if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) { | |
1061 | /* timestamp was stale */ | |
1062 | mask = mask_from(q->bitmaps[ER], agg->grp->index); | |
1063 | if (mask) { | |
1064 | struct qfq_group *next = qfq_ffs(q, mask); | |
1065 | if (qfq_gt(roundedF, next->F)) { | |
1066 | if (qfq_gt(limit, next->F)) | |
1067 | agg->S = next->F; | |
1068 | else /* preserve timestamp correctness */ | |
1069 | agg->S = limit; | |
1070 | return; | |
1071 | } | |
1072 | } | |
1073 | agg->S = q->V; | |
1074 | } else /* timestamp is not stale */ | |
1075 | agg->S = agg->F; | |
1076 | } | |
1077 | ||
1078 | /* Update the timestamps of agg before scheduling/rescheduling it for | |
1079 | * service. In particular, assign to agg->F its maximum possible | |
1080 | * value, i.e., the virtual finish time with which the aggregate | |
1081 | * should be labeled if it used all its budget once in service. | |
1082 | */ | |
1083 | static inline void | |
1084 | qfq_update_agg_ts(struct qfq_sched *q, | |
1085 | struct qfq_aggregate *agg, enum update_reason reason) | |
1086 | { | |
1087 | if (reason != requeue) | |
1088 | qfq_update_start(q, agg); | |
1089 | else /* just charge agg for the service received */ | |
1090 | agg->S = agg->F; | |
1091 | ||
1092 | agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w; | |
1093 | } | |
2f3b89a1 PV |
1094 | |
1095 | static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg); | |
1096 | ||
0545a303 | 1097 | static struct sk_buff *qfq_dequeue(struct Qdisc *sch) |
1098 | { | |
1099 | struct qfq_sched *q = qdisc_priv(sch); | |
462dbc91 | 1100 | struct qfq_aggregate *in_serv_agg = q->in_serv_agg; |
0545a303 | 1101 | struct qfq_class *cl; |
462dbc91 PV |
1102 | struct sk_buff *skb = NULL; |
1103 | /* next-packet len, 0 means no more active classes in in-service agg */ | |
1104 | unsigned int len = 0; | |
0545a303 | 1105 | |
462dbc91 | 1106 | if (in_serv_agg == NULL) |
0545a303 | 1107 | return NULL; |
1108 | ||
462dbc91 PV |
1109 | if (!list_empty(&in_serv_agg->active)) |
1110 | skb = qfq_peek_skb(in_serv_agg, &cl, &len); | |
0545a303 | 1111 | |
462dbc91 PV |
1112 | /* |
1113 | * If there are no active classes in the in-service aggregate, | |
1114 | * or if the aggregate has not enough budget to serve its next | |
1115 | * class, then choose the next aggregate to serve. | |
1116 | */ | |
1117 | if (len == 0 || in_serv_agg->budget < len) { | |
1118 | charge_actual_service(in_serv_agg); | |
1119 | ||
1120 | /* recharge the budget of the aggregate */ | |
1121 | in_serv_agg->initial_budget = in_serv_agg->budget = | |
1122 | in_serv_agg->budgetmax; | |
1123 | ||
2f3b89a1 | 1124 | if (!list_empty(&in_serv_agg->active)) { |
462dbc91 PV |
1125 | /* |
1126 | * Still active: reschedule for | |
1127 | * service. Possible optimization: if no other | |
1128 | * aggregate is active, then there is no point | |
1129 | * in rescheduling this aggregate, and we can | |
1130 | * just keep it as the in-service one. This | |
1131 | * should be however a corner case, and to | |
1132 | * handle it, we would need to maintain an | |
1133 | * extra num_active_aggs field. | |
1134 | */ | |
2f3b89a1 PV |
1135 | qfq_update_agg_ts(q, in_serv_agg, requeue); |
1136 | qfq_schedule_agg(q, in_serv_agg); | |
1137 | } else if (sch->q.qlen == 0) { /* no aggregate to serve */ | |
462dbc91 PV |
1138 | q->in_serv_agg = NULL; |
1139 | return NULL; | |
1140 | } | |
1141 | ||
1142 | /* | |
1143 | * If we get here, there are other aggregates queued: | |
1144 | * choose the new aggregate to serve. | |
1145 | */ | |
1146 | in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q); | |
1147 | skb = qfq_peek_skb(in_serv_agg, &cl, &len); | |
0545a303 | 1148 | } |
462dbc91 PV |
1149 | if (!skb) |
1150 | return NULL; | |
0545a303 | 1151 | |
1152 | sch->q.qlen--; | |
1153 | qdisc_bstats_update(sch, skb); | |
1154 | ||
462dbc91 | 1155 | agg_dequeue(in_serv_agg, cl, len); |
a0143efa PV |
1156 | /* If lmax is lowered, through qfq_change_class, for a class |
1157 | * owning pending packets with larger size than the new value | |
1158 | * of lmax, then the following condition may hold. | |
1159 | */ | |
1160 | if (unlikely(in_serv_agg->budget < len)) | |
1161 | in_serv_agg->budget = 0; | |
1162 | else | |
1163 | in_serv_agg->budget -= len; | |
1164 | ||
87f40dd6 | 1165 | q->V += (u64)len * q->iwsum; |
0545a303 | 1166 | pr_debug("qfq dequeue: len %u F %lld now %lld\n", |
462dbc91 PV |
1167 | len, (unsigned long long) in_serv_agg->F, |
1168 | (unsigned long long) q->V); | |
0545a303 | 1169 | |
462dbc91 PV |
1170 | return skb; |
1171 | } | |
0545a303 | 1172 | |
462dbc91 PV |
1173 | static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q) |
1174 | { | |
1175 | struct qfq_group *grp; | |
1176 | struct qfq_aggregate *agg, *new_front_agg; | |
1177 | u64 old_F; | |
0545a303 | 1178 | |
462dbc91 PV |
1179 | qfq_update_eligible(q); |
1180 | q->oldV = q->V; | |
1181 | ||
1182 | if (!q->bitmaps[ER]) | |
1183 | return NULL; | |
1184 | ||
1185 | grp = qfq_ffs(q, q->bitmaps[ER]); | |
1186 | old_F = grp->F; | |
1187 | ||
1188 | agg = qfq_slot_head(grp); | |
0545a303 | 1189 | |
462dbc91 PV |
1190 | /* agg starts to be served, remove it from schedule */ |
1191 | qfq_front_slot_remove(grp); | |
1192 | ||
1193 | new_front_agg = qfq_slot_scan(grp); | |
1194 | ||
1195 | if (new_front_agg == NULL) /* group is now inactive, remove from ER */ | |
1196 | __clear_bit(grp->index, &q->bitmaps[ER]); | |
1197 | else { | |
1198 | u64 roundedS = qfq_round_down(new_front_agg->S, | |
1199 | grp->slot_shift); | |
1200 | unsigned int s; | |
1201 | ||
1202 | if (grp->S == roundedS) | |
1203 | return agg; | |
1204 | grp->S = roundedS; | |
1205 | grp->F = roundedS + (2ULL << grp->slot_shift); | |
1206 | __clear_bit(grp->index, &q->bitmaps[ER]); | |
1207 | s = qfq_calc_state(q, grp); | |
1208 | __set_bit(grp->index, &q->bitmaps[s]); | |
0545a303 | 1209 | } |
1210 | ||
462dbc91 | 1211 | qfq_unblock_groups(q, grp->index, old_F); |
0545a303 | 1212 | |
462dbc91 | 1213 | return agg; |
0545a303 | 1214 | } |
1215 | ||
0545a303 | 1216 | static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch) |
1217 | { | |
1218 | struct qfq_sched *q = qdisc_priv(sch); | |
0545a303 | 1219 | struct qfq_class *cl; |
462dbc91 | 1220 | struct qfq_aggregate *agg; |
f54ba779 | 1221 | int err = 0; |
0545a303 | 1222 | |
1223 | cl = qfq_classify(skb, sch, &err); | |
1224 | if (cl == NULL) { | |
1225 | if (err & __NET_XMIT_BYPASS) | |
25331d6c | 1226 | qdisc_qstats_drop(sch); |
0545a303 | 1227 | kfree_skb(skb); |
1228 | return err; | |
1229 | } | |
1230 | pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid); | |
1231 | ||
462dbc91 | 1232 | if (unlikely(cl->agg->lmax < qdisc_pkt_len(skb))) { |
3015f3d2 | 1233 | pr_debug("qfq: increasing maxpkt from %u to %u for class %u", |
462dbc91 PV |
1234 | cl->agg->lmax, qdisc_pkt_len(skb), cl->common.classid); |
1235 | err = qfq_change_agg(sch, cl, cl->agg->class_weight, | |
1236 | qdisc_pkt_len(skb)); | |
1237 | if (err) | |
1238 | return err; | |
3015f3d2 PV |
1239 | } |
1240 | ||
0545a303 | 1241 | err = qdisc_enqueue(skb, cl->qdisc); |
1242 | if (unlikely(err != NET_XMIT_SUCCESS)) { | |
1243 | pr_debug("qfq_enqueue: enqueue failed %d\n", err); | |
1244 | if (net_xmit_drop_count(err)) { | |
1245 | cl->qstats.drops++; | |
25331d6c | 1246 | qdisc_qstats_drop(sch); |
0545a303 | 1247 | } |
1248 | return err; | |
1249 | } | |
1250 | ||
1251 | bstats_update(&cl->bstats, skb); | |
1252 | ++sch->q.qlen; | |
1253 | ||
462dbc91 PV |
1254 | agg = cl->agg; |
1255 | /* if the queue was not empty, then done here */ | |
1256 | if (cl->qdisc->q.qlen != 1) { | |
1257 | if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) && | |
1258 | list_first_entry(&agg->active, struct qfq_class, alist) | |
1259 | == cl && cl->deficit < qdisc_pkt_len(skb)) | |
1260 | list_move_tail(&cl->alist, &agg->active); | |
1261 | ||
0545a303 | 1262 | return err; |
462dbc91 PV |
1263 | } |
1264 | ||
1265 | /* schedule class for service within the aggregate */ | |
1266 | cl->deficit = agg->lmax; | |
1267 | list_add_tail(&cl->alist, &agg->active); | |
0545a303 | 1268 | |
2f3b89a1 PV |
1269 | if (list_first_entry(&agg->active, struct qfq_class, alist) != cl || |
1270 | q->in_serv_agg == agg) | |
1271 | return err; /* non-empty or in service, nothing else to do */ | |
462dbc91 | 1272 | |
2f3b89a1 | 1273 | qfq_activate_agg(q, agg, enqueue); |
be72f63b PV |
1274 | |
1275 | return err; | |
1276 | } | |
1277 | ||
1278 | /* | |
462dbc91 | 1279 | * Schedule aggregate according to its timestamps. |
be72f63b | 1280 | */ |
462dbc91 | 1281 | static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg) |
be72f63b | 1282 | { |
462dbc91 | 1283 | struct qfq_group *grp = agg->grp; |
be72f63b PV |
1284 | u64 roundedS; |
1285 | int s; | |
1286 | ||
462dbc91 | 1287 | roundedS = qfq_round_down(agg->S, grp->slot_shift); |
0545a303 | 1288 | |
1289 | /* | |
462dbc91 PV |
1290 | * Insert agg in the correct bucket. |
1291 | * If agg->S >= grp->S we don't need to adjust the | |
0545a303 | 1292 | * bucket list and simply go to the insertion phase. |
1293 | * Otherwise grp->S is decreasing, we must make room | |
1294 | * in the bucket list, and also recompute the group state. | |
1295 | * Finally, if there were no flows in this group and nobody | |
1296 | * was in ER make sure to adjust V. | |
1297 | */ | |
1298 | if (grp->full_slots) { | |
462dbc91 | 1299 | if (!qfq_gt(grp->S, agg->S)) |
0545a303 | 1300 | goto skip_update; |
1301 | ||
462dbc91 | 1302 | /* create a slot for this agg->S */ |
0545a303 | 1303 | qfq_slot_rotate(grp, roundedS); |
1304 | /* group was surely ineligible, remove */ | |
1305 | __clear_bit(grp->index, &q->bitmaps[IR]); | |
1306 | __clear_bit(grp->index, &q->bitmaps[IB]); | |
40dd2d54 PV |
1307 | } else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) && |
1308 | q->in_serv_agg == NULL) | |
0545a303 | 1309 | q->V = roundedS; |
1310 | ||
1311 | grp->S = roundedS; | |
1312 | grp->F = roundedS + (2ULL << grp->slot_shift); | |
1313 | s = qfq_calc_state(q, grp); | |
1314 | __set_bit(grp->index, &q->bitmaps[s]); | |
1315 | ||
1316 | pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n", | |
1317 | s, q->bitmaps[s], | |
462dbc91 PV |
1318 | (unsigned long long) agg->S, |
1319 | (unsigned long long) agg->F, | |
0545a303 | 1320 | (unsigned long long) q->V); |
1321 | ||
1322 | skip_update: | |
462dbc91 | 1323 | qfq_slot_insert(grp, agg, roundedS); |
0545a303 | 1324 | } |
1325 | ||
1326 | ||
462dbc91 PV |
1327 | /* Update agg ts and schedule agg for service */ |
1328 | static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg, | |
1329 | enum update_reason reason) | |
1330 | { | |
2f3b89a1 PV |
1331 | agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */ |
1332 | ||
462dbc91 | 1333 | qfq_update_agg_ts(q, agg, reason); |
2f3b89a1 PV |
1334 | if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */ |
1335 | q->in_serv_agg = agg; /* start serving this aggregate */ | |
1336 | /* update V: to be in service, agg must be eligible */ | |
1337 | q->oldV = q->V = agg->S; | |
1338 | } else if (agg != q->in_serv_agg) | |
1339 | qfq_schedule_agg(q, agg); | |
462dbc91 PV |
1340 | } |
1341 | ||
0545a303 | 1342 | static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp, |
462dbc91 | 1343 | struct qfq_aggregate *agg) |
0545a303 | 1344 | { |
1345 | unsigned int i, offset; | |
1346 | u64 roundedS; | |
1347 | ||
462dbc91 | 1348 | roundedS = qfq_round_down(agg->S, grp->slot_shift); |
0545a303 | 1349 | offset = (roundedS - grp->S) >> grp->slot_shift; |
462dbc91 | 1350 | |
0545a303 | 1351 | i = (grp->front + offset) % QFQ_MAX_SLOTS; |
1352 | ||
462dbc91 | 1353 | hlist_del(&agg->next); |
0545a303 | 1354 | if (hlist_empty(&grp->slots[i])) |
1355 | __clear_bit(offset, &grp->full_slots); | |
1356 | } | |
1357 | ||
1358 | /* | |
462dbc91 PV |
1359 | * Called to forcibly deschedule an aggregate. If the aggregate is |
1360 | * not in the front bucket, or if the latter has other aggregates in | |
1361 | * the front bucket, we can simply remove the aggregate with no other | |
1362 | * side effects. | |
0545a303 | 1363 | * Otherwise we must propagate the event up. |
1364 | */ | |
462dbc91 | 1365 | static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg) |
0545a303 | 1366 | { |
462dbc91 | 1367 | struct qfq_group *grp = agg->grp; |
0545a303 | 1368 | unsigned long mask; |
1369 | u64 roundedS; | |
1370 | int s; | |
1371 | ||
462dbc91 PV |
1372 | if (agg == q->in_serv_agg) { |
1373 | charge_actual_service(agg); | |
1374 | q->in_serv_agg = qfq_choose_next_agg(q); | |
1375 | return; | |
1376 | } | |
1377 | ||
1378 | agg->F = agg->S; | |
1379 | qfq_slot_remove(q, grp, agg); | |
0545a303 | 1380 | |
1381 | if (!grp->full_slots) { | |
1382 | __clear_bit(grp->index, &q->bitmaps[IR]); | |
1383 | __clear_bit(grp->index, &q->bitmaps[EB]); | |
1384 | __clear_bit(grp->index, &q->bitmaps[IB]); | |
1385 | ||
1386 | if (test_bit(grp->index, &q->bitmaps[ER]) && | |
1387 | !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) { | |
1388 | mask = q->bitmaps[ER] & ((1UL << grp->index) - 1); | |
1389 | if (mask) | |
1390 | mask = ~((1UL << __fls(mask)) - 1); | |
1391 | else | |
1392 | mask = ~0UL; | |
1393 | qfq_move_groups(q, mask, EB, ER); | |
1394 | qfq_move_groups(q, mask, IB, IR); | |
1395 | } | |
1396 | __clear_bit(grp->index, &q->bitmaps[ER]); | |
1397 | } else if (hlist_empty(&grp->slots[grp->front])) { | |
462dbc91 PV |
1398 | agg = qfq_slot_scan(grp); |
1399 | roundedS = qfq_round_down(agg->S, grp->slot_shift); | |
0545a303 | 1400 | if (grp->S != roundedS) { |
1401 | __clear_bit(grp->index, &q->bitmaps[ER]); | |
1402 | __clear_bit(grp->index, &q->bitmaps[IR]); | |
1403 | __clear_bit(grp->index, &q->bitmaps[EB]); | |
1404 | __clear_bit(grp->index, &q->bitmaps[IB]); | |
1405 | grp->S = roundedS; | |
1406 | grp->F = roundedS + (2ULL << grp->slot_shift); | |
1407 | s = qfq_calc_state(q, grp); | |
1408 | __set_bit(grp->index, &q->bitmaps[s]); | |
1409 | } | |
1410 | } | |
0545a303 | 1411 | } |
1412 | ||
1413 | static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg) | |
1414 | { | |
1415 | struct qfq_sched *q = qdisc_priv(sch); | |
1416 | struct qfq_class *cl = (struct qfq_class *)arg; | |
1417 | ||
1418 | if (cl->qdisc->q.qlen == 0) | |
1419 | qfq_deactivate_class(q, cl); | |
1420 | } | |
1421 | ||
462dbc91 PV |
1422 | static unsigned int qfq_drop_from_slot(struct qfq_sched *q, |
1423 | struct hlist_head *slot) | |
1424 | { | |
1425 | struct qfq_aggregate *agg; | |
462dbc91 PV |
1426 | struct qfq_class *cl; |
1427 | unsigned int len; | |
1428 | ||
b67bfe0d | 1429 | hlist_for_each_entry(agg, slot, next) { |
462dbc91 PV |
1430 | list_for_each_entry(cl, &agg->active, alist) { |
1431 | ||
1432 | if (!cl->qdisc->ops->drop) | |
1433 | continue; | |
1434 | ||
1435 | len = cl->qdisc->ops->drop(cl->qdisc); | |
1436 | if (len > 0) { | |
1437 | if (cl->qdisc->q.qlen == 0) | |
1438 | qfq_deactivate_class(q, cl); | |
1439 | ||
1440 | return len; | |
1441 | } | |
1442 | } | |
1443 | } | |
1444 | return 0; | |
1445 | } | |
1446 | ||
0545a303 | 1447 | static unsigned int qfq_drop(struct Qdisc *sch) |
1448 | { | |
1449 | struct qfq_sched *q = qdisc_priv(sch); | |
1450 | struct qfq_group *grp; | |
1451 | unsigned int i, j, len; | |
1452 | ||
1453 | for (i = 0; i <= QFQ_MAX_INDEX; i++) { | |
1454 | grp = &q->groups[i]; | |
1455 | for (j = 0; j < QFQ_MAX_SLOTS; j++) { | |
462dbc91 PV |
1456 | len = qfq_drop_from_slot(q, &grp->slots[j]); |
1457 | if (len > 0) { | |
1458 | sch->q.qlen--; | |
1459 | return len; | |
0545a303 | 1460 | } |
1461 | } | |
462dbc91 | 1462 | |
0545a303 | 1463 | } |
1464 | ||
1465 | return 0; | |
1466 | } | |
1467 | ||
1468 | static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt) | |
1469 | { | |
1470 | struct qfq_sched *q = qdisc_priv(sch); | |
1471 | struct qfq_group *grp; | |
1472 | int i, j, err; | |
462dbc91 | 1473 | u32 max_cl_shift, maxbudg_shift, max_classes; |
0545a303 | 1474 | |
1475 | err = qdisc_class_hash_init(&q->clhash); | |
1476 | if (err < 0) | |
1477 | return err; | |
1478 | ||
462dbc91 PV |
1479 | if (qdisc_dev(sch)->tx_queue_len + 1 > QFQ_MAX_AGG_CLASSES) |
1480 | max_classes = QFQ_MAX_AGG_CLASSES; | |
1481 | else | |
1482 | max_classes = qdisc_dev(sch)->tx_queue_len + 1; | |
1483 | /* max_cl_shift = floor(log_2(max_classes)) */ | |
1484 | max_cl_shift = __fls(max_classes); | |
1485 | q->max_agg_classes = 1<<max_cl_shift; | |
1486 | ||
1487 | /* maxbudg_shift = log2(max_len * max_classes_per_agg) */ | |
1488 | maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift; | |
1489 | q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX; | |
1490 | ||
0545a303 | 1491 | for (i = 0; i <= QFQ_MAX_INDEX; i++) { |
1492 | grp = &q->groups[i]; | |
1493 | grp->index = i; | |
462dbc91 | 1494 | grp->slot_shift = q->min_slot_shift + i; |
0545a303 | 1495 | for (j = 0; j < QFQ_MAX_SLOTS; j++) |
1496 | INIT_HLIST_HEAD(&grp->slots[j]); | |
1497 | } | |
1498 | ||
462dbc91 PV |
1499 | INIT_HLIST_HEAD(&q->nonfull_aggs); |
1500 | ||
0545a303 | 1501 | return 0; |
1502 | } | |
1503 | ||
1504 | static void qfq_reset_qdisc(struct Qdisc *sch) | |
1505 | { | |
1506 | struct qfq_sched *q = qdisc_priv(sch); | |
0545a303 | 1507 | struct qfq_class *cl; |
462dbc91 | 1508 | unsigned int i; |
0545a303 | 1509 | |
462dbc91 | 1510 | for (i = 0; i < q->clhash.hashsize; i++) { |
b67bfe0d | 1511 | hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) { |
462dbc91 | 1512 | if (cl->qdisc->q.qlen > 0) |
0545a303 | 1513 | qfq_deactivate_class(q, cl); |
0545a303 | 1514 | |
0545a303 | 1515 | qdisc_reset(cl->qdisc); |
462dbc91 | 1516 | } |
0545a303 | 1517 | } |
1518 | sch->q.qlen = 0; | |
1519 | } | |
1520 | ||
1521 | static void qfq_destroy_qdisc(struct Qdisc *sch) | |
1522 | { | |
1523 | struct qfq_sched *q = qdisc_priv(sch); | |
1524 | struct qfq_class *cl; | |
b67bfe0d | 1525 | struct hlist_node *next; |
0545a303 | 1526 | unsigned int i; |
1527 | ||
1528 | tcf_destroy_chain(&q->filter_list); | |
1529 | ||
1530 | for (i = 0; i < q->clhash.hashsize; i++) { | |
b67bfe0d | 1531 | hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i], |
0545a303 | 1532 | common.hnode) { |
1533 | qfq_destroy_class(sch, cl); | |
1534 | } | |
1535 | } | |
1536 | qdisc_class_hash_destroy(&q->clhash); | |
1537 | } | |
1538 | ||
1539 | static const struct Qdisc_class_ops qfq_class_ops = { | |
1540 | .change = qfq_change_class, | |
1541 | .delete = qfq_delete_class, | |
1542 | .get = qfq_get_class, | |
1543 | .put = qfq_put_class, | |
1544 | .tcf_chain = qfq_tcf_chain, | |
1545 | .bind_tcf = qfq_bind_tcf, | |
1546 | .unbind_tcf = qfq_unbind_tcf, | |
1547 | .graft = qfq_graft_class, | |
1548 | .leaf = qfq_class_leaf, | |
1549 | .qlen_notify = qfq_qlen_notify, | |
1550 | .dump = qfq_dump_class, | |
1551 | .dump_stats = qfq_dump_class_stats, | |
1552 | .walk = qfq_walk, | |
1553 | }; | |
1554 | ||
1555 | static struct Qdisc_ops qfq_qdisc_ops __read_mostly = { | |
1556 | .cl_ops = &qfq_class_ops, | |
1557 | .id = "qfq", | |
1558 | .priv_size = sizeof(struct qfq_sched), | |
1559 | .enqueue = qfq_enqueue, | |
1560 | .dequeue = qfq_dequeue, | |
1561 | .peek = qdisc_peek_dequeued, | |
1562 | .drop = qfq_drop, | |
1563 | .init = qfq_init_qdisc, | |
1564 | .reset = qfq_reset_qdisc, | |
1565 | .destroy = qfq_destroy_qdisc, | |
1566 | .owner = THIS_MODULE, | |
1567 | }; | |
1568 | ||
1569 | static int __init qfq_init(void) | |
1570 | { | |
1571 | return register_qdisc(&qfq_qdisc_ops); | |
1572 | } | |
1573 | ||
1574 | static void __exit qfq_exit(void) | |
1575 | { | |
1576 | unregister_qdisc(&qfq_qdisc_ops); | |
1577 | } | |
1578 | ||
1579 | module_init(qfq_init); | |
1580 | module_exit(qfq_exit); | |
1581 | MODULE_LICENSE("GPL"); |