Commit | Line | Data |
---|---|---|
d2912cb1 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
0545a303 | 2 | /* |
462dbc91 | 3 | * net/sched/sch_qfq.c Quick Fair Queueing Plus Scheduler. |
0545a303 | 4 | * |
5 | * Copyright (c) 2009 Fabio Checconi, Luigi Rizzo, and Paolo Valente. | |
462dbc91 | 6 | * Copyright (c) 2012 Paolo Valente. |
0545a303 | 7 | */ |
8 | ||
9 | #include <linux/module.h> | |
10 | #include <linux/init.h> | |
11 | #include <linux/bitops.h> | |
12 | #include <linux/errno.h> | |
13 | #include <linux/netdevice.h> | |
14 | #include <linux/pkt_sched.h> | |
15 | #include <net/sch_generic.h> | |
16 | #include <net/pkt_sched.h> | |
17 | #include <net/pkt_cls.h> | |
18 | ||
19 | ||
462dbc91 PV |
20 | /* Quick Fair Queueing Plus |
21 | ======================== | |
0545a303 | 22 | |
23 | Sources: | |
24 | ||
462dbc91 PV |
25 | [1] Paolo Valente, |
26 | "Reducing the Execution Time of Fair-Queueing Schedulers." | |
27 | http://algo.ing.unimo.it/people/paolo/agg-sched/agg-sched.pdf | |
28 | ||
29 | Sources for QFQ: | |
30 | ||
31 | [2] Fabio Checconi, Luigi Rizzo, and Paolo Valente: "QFQ: Efficient | |
0545a303 | 32 | Packet Scheduling with Tight Bandwidth Distribution Guarantees." |
33 | ||
34 | See also: | |
35 | http://retis.sssup.it/~fabio/linux/qfq/ | |
36 | */ | |
37 | ||
38 | /* | |
39 | ||
462dbc91 PV |
40 | QFQ+ divides classes into aggregates of at most MAX_AGG_CLASSES |
41 | classes. Each aggregate is timestamped with a virtual start time S | |
42 | and a virtual finish time F, and scheduled according to its | |
43 | timestamps. S and F are computed as a function of a system virtual | |
44 | time function V. The classes within each aggregate are instead | |
45 | scheduled with DRR. | |
46 | ||
47 | To speed up operations, QFQ+ divides also aggregates into a limited | |
48 | number of groups. Which group a class belongs to depends on the | |
49 | ratio between the maximum packet length for the class and the weight | |
50 | of the class. Groups have their own S and F. In the end, QFQ+ | |
51 | schedules groups, then aggregates within groups, then classes within | |
52 | aggregates. See [1] and [2] for a full description. | |
53 | ||
0545a303 | 54 | Virtual time computations. |
55 | ||
56 | S, F and V are all computed in fixed point arithmetic with | |
57 | FRAC_BITS decimal bits. | |
58 | ||
59 | QFQ_MAX_INDEX is the maximum index allowed for a group. We need | |
60 | one bit per index. | |
61 | QFQ_MAX_WSHIFT is the maximum power of two supported as a weight. | |
62 | ||
63 | The layout of the bits is as below: | |
64 | ||
65 | [ MTU_SHIFT ][ FRAC_BITS ] | |
66 | [ MAX_INDEX ][ MIN_SLOT_SHIFT ] | |
67 | ^.__grp->index = 0 | |
68 | *.__grp->slot_shift | |
69 | ||
70 | where MIN_SLOT_SHIFT is derived by difference from the others. | |
71 | ||
72 | The max group index corresponds to Lmax/w_min, where | |
73 | Lmax=1<<MTU_SHIFT, w_min = 1 . | |
74 | From this, and knowing how many groups (MAX_INDEX) we want, | |
75 | we can derive the shift corresponding to each group. | |
76 | ||
77 | Because we often need to compute | |
78 | F = S + len/w_i and V = V + len/wsum | |
79 | instead of storing w_i store the value | |
80 | inv_w = (1<<FRAC_BITS)/w_i | |
81 | so we can do F = S + len * inv_w * wsum. | |
82 | We use W_TOT in the formulas so we can easily move between | |
83 | static and adaptive weight sum. | |
84 | ||
85 | The per-scheduler-instance data contain all the data structures | |
86 | for the scheduler: bitmaps and bucket lists. | |
87 | ||
88 | */ | |
89 | ||
90 | /* | |
91 | * Maximum number of consecutive slots occupied by backlogged classes | |
92 | * inside a group. | |
93 | */ | |
94 | #define QFQ_MAX_SLOTS 32 | |
95 | ||
96 | /* | |
462dbc91 PV |
97 | * Shifts used for aggregate<->group mapping. We allow class weights that are |
98 | * in the range [1, 2^MAX_WSHIFT], and we try to map each aggregate i to the | |
0545a303 | 99 | * group with the smallest index that can support the L_i / r_i configured |
462dbc91 | 100 | * for the classes in the aggregate. |
0545a303 | 101 | * |
102 | * grp->index is the index of the group; and grp->slot_shift | |
103 | * is the shift for the corresponding (scaled) sigma_i. | |
104 | */ | |
3015f3d2 | 105 | #define QFQ_MAX_INDEX 24 |
462dbc91 | 106 | #define QFQ_MAX_WSHIFT 10 |
0545a303 | 107 | |
462dbc91 PV |
108 | #define QFQ_MAX_WEIGHT (1<<QFQ_MAX_WSHIFT) /* see qfq_slot_insert */ |
109 | #define QFQ_MAX_WSUM (64*QFQ_MAX_WEIGHT) | |
0545a303 | 110 | |
111 | #define FRAC_BITS 30 /* fixed point arithmetic */ | |
112 | #define ONE_FP (1UL << FRAC_BITS) | |
0545a303 | 113 | |
3015f3d2 | 114 | #define QFQ_MTU_SHIFT 16 /* to support TSO/GSO */ |
462dbc91 PV |
115 | #define QFQ_MIN_LMAX 512 /* see qfq_slot_insert */ |
116 | ||
117 | #define QFQ_MAX_AGG_CLASSES 8 /* max num classes per aggregate allowed */ | |
0545a303 | 118 | |
119 | /* | |
120 | * Possible group states. These values are used as indexes for the bitmaps | |
121 | * array of struct qfq_queue. | |
122 | */ | |
123 | enum qfq_state { ER, IR, EB, IB, QFQ_MAX_STATE }; | |
124 | ||
125 | struct qfq_group; | |
126 | ||
462dbc91 PV |
127 | struct qfq_aggregate; |
128 | ||
0545a303 | 129 | struct qfq_class { |
130 | struct Qdisc_class_common common; | |
131 | ||
0545a303 | 132 | unsigned int filter_cnt; |
133 | ||
50dc9a85 | 134 | struct gnet_stats_basic_sync bstats; |
0545a303 | 135 | struct gnet_stats_queue qstats; |
1c0d32fd | 136 | struct net_rate_estimator __rcu *rate_est; |
0545a303 | 137 | struct Qdisc *qdisc; |
462dbc91 PV |
138 | struct list_head alist; /* Link for active-classes list. */ |
139 | struct qfq_aggregate *agg; /* Parent aggregate. */ | |
140 | int deficit; /* DRR deficit counter. */ | |
141 | }; | |
0545a303 | 142 | |
462dbc91 | 143 | struct qfq_aggregate { |
0545a303 | 144 | struct hlist_node next; /* Link for the slot list. */ |
145 | u64 S, F; /* flow timestamps (exact) */ | |
146 | ||
147 | /* group we belong to. In principle we would need the index, | |
148 | * which is log_2(lmax/weight), but we never reference it | |
149 | * directly, only the group. | |
150 | */ | |
151 | struct qfq_group *grp; | |
152 | ||
153 | /* these are copied from the flowset. */ | |
462dbc91 PV |
154 | u32 class_weight; /* Weight of each class in this aggregate. */ |
155 | /* Max pkt size for the classes in this aggregate, DRR quantum. */ | |
156 | int lmax; | |
157 | ||
158 | u32 inv_w; /* ONE_FP/(sum of weights of classes in aggr.). */ | |
159 | u32 budgetmax; /* Max budget for this aggregate. */ | |
160 | u32 initial_budget, budget; /* Initial and current budget. */ | |
161 | ||
162 | int num_classes; /* Number of classes in this aggr. */ | |
163 | struct list_head active; /* DRR queue of active classes. */ | |
164 | ||
165 | struct hlist_node nonfull_next; /* See nonfull_aggs in qfq_sched. */ | |
0545a303 | 166 | }; |
167 | ||
168 | struct qfq_group { | |
169 | u64 S, F; /* group timestamps (approx). */ | |
170 | unsigned int slot_shift; /* Slot shift. */ | |
171 | unsigned int index; /* Group index. */ | |
172 | unsigned int front; /* Index of the front slot. */ | |
173 | unsigned long full_slots; /* non-empty slots */ | |
174 | ||
462dbc91 | 175 | /* Array of RR lists of active aggregates. */ |
0545a303 | 176 | struct hlist_head slots[QFQ_MAX_SLOTS]; |
177 | }; | |
178 | ||
179 | struct qfq_sched { | |
25d8c0d5 | 180 | struct tcf_proto __rcu *filter_list; |
6529eaba | 181 | struct tcf_block *block; |
0545a303 | 182 | struct Qdisc_class_hash clhash; |
183 | ||
462dbc91 PV |
184 | u64 oldV, V; /* Precise virtual times. */ |
185 | struct qfq_aggregate *in_serv_agg; /* Aggregate being served. */ | |
462dbc91 | 186 | u32 wsum; /* weight sum */ |
87f40dd6 | 187 | u32 iwsum; /* inverse weight sum */ |
0545a303 | 188 | |
189 | unsigned long bitmaps[QFQ_MAX_STATE]; /* Group bitmaps. */ | |
190 | struct qfq_group groups[QFQ_MAX_INDEX + 1]; /* The groups. */ | |
462dbc91 PV |
191 | u32 min_slot_shift; /* Index of the group-0 bit in the bitmaps. */ |
192 | ||
193 | u32 max_agg_classes; /* Max number of classes per aggr. */ | |
194 | struct hlist_head nonfull_aggs; /* Aggs with room for more classes. */ | |
0545a303 | 195 | }; |
196 | ||
462dbc91 PV |
197 | /* |
198 | * Possible reasons why the timestamps of an aggregate are updated | |
199 | * enqueue: the aggregate switches from idle to active and must scheduled | |
200 | * for service | |
201 | * requeue: the aggregate finishes its budget, so it stops being served and | |
202 | * must be rescheduled for service | |
203 | */ | |
204 | enum update_reason {enqueue, requeue}; | |
205 | ||
0545a303 | 206 | static struct qfq_class *qfq_find_class(struct Qdisc *sch, u32 classid) |
207 | { | |
208 | struct qfq_sched *q = qdisc_priv(sch); | |
209 | struct Qdisc_class_common *clc; | |
210 | ||
211 | clc = qdisc_class_find(&q->clhash, classid); | |
212 | if (clc == NULL) | |
213 | return NULL; | |
214 | return container_of(clc, struct qfq_class, common); | |
215 | } | |
216 | ||
0545a303 | 217 | static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = { |
218 | [TCA_QFQ_WEIGHT] = { .type = NLA_U32 }, | |
219 | [TCA_QFQ_LMAX] = { .type = NLA_U32 }, | |
220 | }; | |
221 | ||
222 | /* | |
223 | * Calculate a flow index, given its weight and maximum packet length. | |
224 | * index = log_2(maxlen/weight) but we need to apply the scaling. | |
225 | * This is used only once at flow creation. | |
226 | */ | |
462dbc91 | 227 | static int qfq_calc_index(u32 inv_w, unsigned int maxlen, u32 min_slot_shift) |
0545a303 | 228 | { |
229 | u64 slot_size = (u64)maxlen * inv_w; | |
230 | unsigned long size_map; | |
231 | int index = 0; | |
232 | ||
462dbc91 | 233 | size_map = slot_size >> min_slot_shift; |
0545a303 | 234 | if (!size_map) |
235 | goto out; | |
236 | ||
237 | index = __fls(size_map) + 1; /* basically a log_2 */ | |
462dbc91 | 238 | index -= !(slot_size - (1ULL << (index + min_slot_shift - 1))); |
0545a303 | 239 | |
240 | if (index < 0) | |
241 | index = 0; | |
242 | out: | |
243 | pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n", | |
244 | (unsigned long) ONE_FP/inv_w, maxlen, index); | |
245 | ||
246 | return index; | |
247 | } | |
248 | ||
462dbc91 PV |
249 | static void qfq_deactivate_agg(struct qfq_sched *, struct qfq_aggregate *); |
250 | static void qfq_activate_agg(struct qfq_sched *, struct qfq_aggregate *, | |
251 | enum update_reason); | |
252 | ||
253 | static void qfq_init_agg(struct qfq_sched *q, struct qfq_aggregate *agg, | |
254 | u32 lmax, u32 weight) | |
be72f63b | 255 | { |
462dbc91 PV |
256 | INIT_LIST_HEAD(&agg->active); |
257 | hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs); | |
258 | ||
259 | agg->lmax = lmax; | |
260 | agg->class_weight = weight; | |
261 | } | |
262 | ||
263 | static struct qfq_aggregate *qfq_find_agg(struct qfq_sched *q, | |
264 | u32 lmax, u32 weight) | |
265 | { | |
266 | struct qfq_aggregate *agg; | |
462dbc91 | 267 | |
b67bfe0d | 268 | hlist_for_each_entry(agg, &q->nonfull_aggs, nonfull_next) |
462dbc91 PV |
269 | if (agg->lmax == lmax && agg->class_weight == weight) |
270 | return agg; | |
271 | ||
272 | return NULL; | |
273 | } | |
274 | ||
be72f63b | 275 | |
462dbc91 PV |
276 | /* Update aggregate as a function of the new number of classes. */ |
277 | static void qfq_update_agg(struct qfq_sched *q, struct qfq_aggregate *agg, | |
278 | int new_num_classes) | |
279 | { | |
280 | u32 new_agg_weight; | |
281 | ||
282 | if (new_num_classes == q->max_agg_classes) | |
283 | hlist_del_init(&agg->nonfull_next); | |
284 | ||
285 | if (agg->num_classes > new_num_classes && | |
286 | new_num_classes == q->max_agg_classes - 1) /* agg no more full */ | |
287 | hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs); | |
288 | ||
9b99b7e9 PV |
289 | /* The next assignment may let |
290 | * agg->initial_budget > agg->budgetmax | |
291 | * hold, we will take it into account in charge_actual_service(). | |
292 | */ | |
462dbc91 PV |
293 | agg->budgetmax = new_num_classes * agg->lmax; |
294 | new_agg_weight = agg->class_weight * new_num_classes; | |
295 | agg->inv_w = ONE_FP/new_agg_weight; | |
296 | ||
297 | if (agg->grp == NULL) { | |
298 | int i = qfq_calc_index(agg->inv_w, agg->budgetmax, | |
299 | q->min_slot_shift); | |
300 | agg->grp = &q->groups[i]; | |
301 | } | |
302 | ||
303 | q->wsum += | |
304 | (int) agg->class_weight * (new_num_classes - agg->num_classes); | |
87f40dd6 | 305 | q->iwsum = ONE_FP / q->wsum; |
462dbc91 PV |
306 | |
307 | agg->num_classes = new_num_classes; | |
308 | } | |
309 | ||
310 | /* Add class to aggregate. */ | |
311 | static void qfq_add_to_agg(struct qfq_sched *q, | |
312 | struct qfq_aggregate *agg, | |
313 | struct qfq_class *cl) | |
314 | { | |
315 | cl->agg = agg; | |
316 | ||
317 | qfq_update_agg(q, agg, agg->num_classes+1); | |
318 | if (cl->qdisc->q.qlen > 0) { /* adding an active class */ | |
319 | list_add_tail(&cl->alist, &agg->active); | |
320 | if (list_first_entry(&agg->active, struct qfq_class, alist) == | |
321 | cl && q->in_serv_agg != agg) /* agg was inactive */ | |
322 | qfq_activate_agg(q, agg, enqueue); /* schedule agg */ | |
323 | } | |
be72f63b PV |
324 | } |
325 | ||
462dbc91 | 326 | static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *); |
be72f63b | 327 | |
462dbc91 | 328 | static void qfq_destroy_agg(struct qfq_sched *q, struct qfq_aggregate *agg) |
be72f63b | 329 | { |
a55e1c5c | 330 | hlist_del_init(&agg->nonfull_next); |
87f40dd6 PV |
331 | q->wsum -= agg->class_weight; |
332 | if (q->wsum != 0) | |
333 | q->iwsum = ONE_FP / q->wsum; | |
334 | ||
462dbc91 PV |
335 | if (q->in_serv_agg == agg) |
336 | q->in_serv_agg = qfq_choose_next_agg(q); | |
337 | kfree(agg); | |
338 | } | |
be72f63b | 339 | |
462dbc91 PV |
340 | /* Deschedule class from within its parent aggregate. */ |
341 | static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl) | |
342 | { | |
343 | struct qfq_aggregate *agg = cl->agg; | |
be72f63b | 344 | |
be72f63b | 345 | |
462dbc91 PV |
346 | list_del(&cl->alist); /* remove from RR queue of the aggregate */ |
347 | if (list_empty(&agg->active)) /* agg is now inactive */ | |
348 | qfq_deactivate_agg(q, agg); | |
be72f63b PV |
349 | } |
350 | ||
462dbc91 PV |
351 | /* Remove class from its parent aggregate. */ |
352 | static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl) | |
3015f3d2 | 353 | { |
462dbc91 | 354 | struct qfq_aggregate *agg = cl->agg; |
3015f3d2 | 355 | |
462dbc91 PV |
356 | cl->agg = NULL; |
357 | if (agg->num_classes == 1) { /* agg being emptied, destroy it */ | |
358 | qfq_destroy_agg(q, agg); | |
359 | return; | |
3015f3d2 | 360 | } |
462dbc91 PV |
361 | qfq_update_agg(q, agg, agg->num_classes-1); |
362 | } | |
3015f3d2 | 363 | |
462dbc91 PV |
364 | /* Deschedule class and remove it from its parent aggregate. */ |
365 | static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl) | |
366 | { | |
367 | if (cl->qdisc->q.qlen > 0) /* class is active */ | |
368 | qfq_deactivate_class(q, cl); | |
3015f3d2 | 369 | |
462dbc91 | 370 | qfq_rm_from_agg(q, cl); |
3015f3d2 PV |
371 | } |
372 | ||
462dbc91 PV |
373 | /* Move class to a new aggregate, matching the new class weight and/or lmax */ |
374 | static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight, | |
375 | u32 lmax) | |
376 | { | |
377 | struct qfq_sched *q = qdisc_priv(sch); | |
378 | struct qfq_aggregate *new_agg = qfq_find_agg(q, lmax, weight); | |
379 | ||
380 | if (new_agg == NULL) { /* create new aggregate */ | |
381 | new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC); | |
382 | if (new_agg == NULL) | |
383 | return -ENOBUFS; | |
384 | qfq_init_agg(q, new_agg, lmax, weight); | |
385 | } | |
386 | qfq_deact_rm_from_agg(q, cl); | |
387 | qfq_add_to_agg(q, new_agg, cl); | |
388 | ||
389 | return 0; | |
390 | } | |
3015f3d2 | 391 | |
0545a303 | 392 | static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid, |
793d81d6 AA |
393 | struct nlattr **tca, unsigned long *arg, |
394 | struct netlink_ext_ack *extack) | |
0545a303 | 395 | { |
396 | struct qfq_sched *q = qdisc_priv(sch); | |
397 | struct qfq_class *cl = (struct qfq_class *)*arg; | |
462dbc91 | 398 | bool existing = false; |
0545a303 | 399 | struct nlattr *tb[TCA_QFQ_MAX + 1]; |
462dbc91 | 400 | struct qfq_aggregate *new_agg = NULL; |
0545a303 | 401 | u32 weight, lmax, inv_w; |
3015f3d2 | 402 | int err; |
d32ae76f | 403 | int delta_w; |
0545a303 | 404 | |
405 | if (tca[TCA_OPTIONS] == NULL) { | |
406 | pr_notice("qfq: no options\n"); | |
407 | return -EINVAL; | |
408 | } | |
409 | ||
8cb08174 JB |
410 | err = nla_parse_nested_deprecated(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS], |
411 | qfq_policy, NULL); | |
0545a303 | 412 | if (err < 0) |
413 | return err; | |
414 | ||
415 | if (tb[TCA_QFQ_WEIGHT]) { | |
416 | weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]); | |
417 | if (!weight || weight > (1UL << QFQ_MAX_WSHIFT)) { | |
418 | pr_notice("qfq: invalid weight %u\n", weight); | |
419 | return -EINVAL; | |
420 | } | |
421 | } else | |
422 | weight = 1; | |
423 | ||
30379334 | 424 | if (tb[TCA_QFQ_LMAX]) |
0545a303 | 425 | lmax = nla_get_u32(tb[TCA_QFQ_LMAX]); |
30379334 | 426 | else |
3015f3d2 | 427 | lmax = psched_mtu(qdisc_dev(sch)); |
0545a303 | 428 | |
30379334 GJ |
429 | if (lmax < QFQ_MIN_LMAX || lmax > (1UL << QFQ_MTU_SHIFT)) { |
430 | pr_notice("qfq: invalid max length %u\n", lmax); | |
431 | return -EINVAL; | |
432 | } | |
433 | ||
462dbc91 PV |
434 | inv_w = ONE_FP / weight; |
435 | weight = ONE_FP / inv_w; | |
436 | ||
437 | if (cl != NULL && | |
438 | lmax == cl->agg->lmax && | |
439 | weight == cl->agg->class_weight) | |
440 | return 0; /* nothing to change */ | |
441 | ||
442 | delta_w = weight - (cl ? cl->agg->class_weight : 0); | |
443 | ||
444 | if (q->wsum + delta_w > QFQ_MAX_WSUM) { | |
445 | pr_notice("qfq: total weight out of range (%d + %u)\n", | |
446 | delta_w, q->wsum); | |
447 | return -EINVAL; | |
448 | } | |
449 | ||
450 | if (cl != NULL) { /* modify existing class */ | |
0545a303 | 451 | if (tca[TCA_RATE]) { |
22e0f8b9 JF |
452 | err = gen_replace_estimator(&cl->bstats, NULL, |
453 | &cl->rate_est, | |
edb09eb1 | 454 | NULL, |
29cbcd85 | 455 | true, |
0545a303 | 456 | tca[TCA_RATE]); |
457 | if (err) | |
458 | return err; | |
459 | } | |
462dbc91 PV |
460 | existing = true; |
461 | goto set_change_agg; | |
0545a303 | 462 | } |
463 | ||
462dbc91 | 464 | /* create and init new class */ |
0545a303 | 465 | cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL); |
466 | if (cl == NULL) | |
467 | return -ENOBUFS; | |
468 | ||
50dc9a85 | 469 | gnet_stats_basic_sync_init(&cl->bstats); |
0545a303 | 470 | cl->common.classid = classid; |
462dbc91 | 471 | cl->deficit = lmax; |
0545a303 | 472 | |
a38a9882 AA |
473 | cl->qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops, |
474 | classid, NULL); | |
0545a303 | 475 | if (cl->qdisc == NULL) |
476 | cl->qdisc = &noop_qdisc; | |
477 | ||
478 | if (tca[TCA_RATE]) { | |
22e0f8b9 JF |
479 | err = gen_new_estimator(&cl->bstats, NULL, |
480 | &cl->rate_est, | |
edb09eb1 | 481 | NULL, |
29cbcd85 | 482 | true, |
0545a303 | 483 | tca[TCA_RATE]); |
462dbc91 PV |
484 | if (err) |
485 | goto destroy_class; | |
0545a303 | 486 | } |
487 | ||
49b49971 JK |
488 | if (cl->qdisc != &noop_qdisc) |
489 | qdisc_hash_add(cl->qdisc, true); | |
0545a303 | 490 | |
462dbc91 PV |
491 | set_change_agg: |
492 | sch_tree_lock(sch); | |
493 | new_agg = qfq_find_agg(q, lmax, weight); | |
494 | if (new_agg == NULL) { /* create new aggregate */ | |
495 | sch_tree_unlock(sch); | |
496 | new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL); | |
497 | if (new_agg == NULL) { | |
498 | err = -ENOBUFS; | |
1c0d32fd | 499 | gen_kill_estimator(&cl->rate_est); |
462dbc91 PV |
500 | goto destroy_class; |
501 | } | |
502 | sch_tree_lock(sch); | |
503 | qfq_init_agg(q, new_agg, lmax, weight); | |
504 | } | |
505 | if (existing) | |
506 | qfq_deact_rm_from_agg(q, cl); | |
0cd58e5c ED |
507 | else |
508 | qdisc_class_hash_insert(&q->clhash, &cl->common); | |
462dbc91 PV |
509 | qfq_add_to_agg(q, new_agg, cl); |
510 | sch_tree_unlock(sch); | |
0cd58e5c | 511 | qdisc_class_hash_grow(sch, &q->clhash); |
462dbc91 | 512 | |
0545a303 | 513 | *arg = (unsigned long)cl; |
514 | return 0; | |
462dbc91 PV |
515 | |
516 | destroy_class: | |
86bd446b | 517 | qdisc_put(cl->qdisc); |
462dbc91 PV |
518 | kfree(cl); |
519 | return err; | |
0545a303 | 520 | } |
521 | ||
522 | static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl) | |
523 | { | |
524 | struct qfq_sched *q = qdisc_priv(sch); | |
525 | ||
462dbc91 | 526 | qfq_rm_from_agg(q, cl); |
1c0d32fd | 527 | gen_kill_estimator(&cl->rate_est); |
86bd446b | 528 | qdisc_put(cl->qdisc); |
0545a303 | 529 | kfree(cl); |
530 | } | |
531 | ||
4dd78a73 MM |
532 | static int qfq_delete_class(struct Qdisc *sch, unsigned long arg, |
533 | struct netlink_ext_ack *extack) | |
0545a303 | 534 | { |
535 | struct qfq_sched *q = qdisc_priv(sch); | |
536 | struct qfq_class *cl = (struct qfq_class *)arg; | |
537 | ||
538 | if (cl->filter_cnt > 0) | |
539 | return -EBUSY; | |
540 | ||
541 | sch_tree_lock(sch); | |
542 | ||
e5f0e8f8 | 543 | qdisc_purge_queue(cl->qdisc); |
0545a303 | 544 | qdisc_class_hash_remove(&q->clhash, &cl->common); |
545 | ||
0545a303 | 546 | sch_tree_unlock(sch); |
0545a303 | 547 | |
143976ce WC |
548 | qfq_destroy_class(sch, cl); |
549 | return 0; | |
0545a303 | 550 | } |
551 | ||
143976ce | 552 | static unsigned long qfq_search_class(struct Qdisc *sch, u32 classid) |
0545a303 | 553 | { |
143976ce | 554 | return (unsigned long)qfq_find_class(sch, classid); |
0545a303 | 555 | } |
556 | ||
cbaacc4e AA |
557 | static struct tcf_block *qfq_tcf_block(struct Qdisc *sch, unsigned long cl, |
558 | struct netlink_ext_ack *extack) | |
0545a303 | 559 | { |
560 | struct qfq_sched *q = qdisc_priv(sch); | |
561 | ||
562 | if (cl) | |
563 | return NULL; | |
564 | ||
6529eaba | 565 | return q->block; |
0545a303 | 566 | } |
567 | ||
568 | static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent, | |
569 | u32 classid) | |
570 | { | |
571 | struct qfq_class *cl = qfq_find_class(sch, classid); | |
572 | ||
573 | if (cl != NULL) | |
574 | cl->filter_cnt++; | |
575 | ||
576 | return (unsigned long)cl; | |
577 | } | |
578 | ||
579 | static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg) | |
580 | { | |
581 | struct qfq_class *cl = (struct qfq_class *)arg; | |
582 | ||
583 | cl->filter_cnt--; | |
584 | } | |
585 | ||
586 | static int qfq_graft_class(struct Qdisc *sch, unsigned long arg, | |
653d6fd6 AA |
587 | struct Qdisc *new, struct Qdisc **old, |
588 | struct netlink_ext_ack *extack) | |
0545a303 | 589 | { |
590 | struct qfq_class *cl = (struct qfq_class *)arg; | |
591 | ||
592 | if (new == NULL) { | |
a38a9882 AA |
593 | new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops, |
594 | cl->common.classid, NULL); | |
0545a303 | 595 | if (new == NULL) |
596 | new = &noop_qdisc; | |
597 | } | |
598 | ||
86a7996c | 599 | *old = qdisc_replace(sch, new, &cl->qdisc); |
0545a303 | 600 | return 0; |
601 | } | |
602 | ||
603 | static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg) | |
604 | { | |
605 | struct qfq_class *cl = (struct qfq_class *)arg; | |
606 | ||
607 | return cl->qdisc; | |
608 | } | |
609 | ||
610 | static int qfq_dump_class(struct Qdisc *sch, unsigned long arg, | |
611 | struct sk_buff *skb, struct tcmsg *tcm) | |
612 | { | |
613 | struct qfq_class *cl = (struct qfq_class *)arg; | |
614 | struct nlattr *nest; | |
615 | ||
616 | tcm->tcm_parent = TC_H_ROOT; | |
617 | tcm->tcm_handle = cl->common.classid; | |
618 | tcm->tcm_info = cl->qdisc->handle; | |
619 | ||
ae0be8de | 620 | nest = nla_nest_start_noflag(skb, TCA_OPTIONS); |
0545a303 | 621 | if (nest == NULL) |
622 | goto nla_put_failure; | |
462dbc91 PV |
623 | if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) || |
624 | nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax)) | |
1b34ec43 | 625 | goto nla_put_failure; |
0545a303 | 626 | return nla_nest_end(skb, nest); |
627 | ||
628 | nla_put_failure: | |
629 | nla_nest_cancel(skb, nest); | |
630 | return -EMSGSIZE; | |
631 | } | |
632 | ||
633 | static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg, | |
634 | struct gnet_dump *d) | |
635 | { | |
636 | struct qfq_class *cl = (struct qfq_class *)arg; | |
637 | struct tc_qfq_stats xstats; | |
638 | ||
639 | memset(&xstats, 0, sizeof(xstats)); | |
0545a303 | 640 | |
462dbc91 PV |
641 | xstats.weight = cl->agg->class_weight; |
642 | xstats.lmax = cl->agg->lmax; | |
0545a303 | 643 | |
29cbcd85 | 644 | if (gnet_stats_copy_basic(d, NULL, &cl->bstats, true) < 0 || |
1c0d32fd | 645 | gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 || |
5dd431b6 | 646 | qdisc_qstats_copy(d, cl->qdisc) < 0) |
0545a303 | 647 | return -1; |
648 | ||
649 | return gnet_stats_copy_app(d, &xstats, sizeof(xstats)); | |
650 | } | |
651 | ||
652 | static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg) | |
653 | { | |
654 | struct qfq_sched *q = qdisc_priv(sch); | |
655 | struct qfq_class *cl; | |
0545a303 | 656 | unsigned int i; |
657 | ||
658 | if (arg->stop) | |
659 | return; | |
660 | ||
661 | for (i = 0; i < q->clhash.hashsize; i++) { | |
b67bfe0d | 662 | hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) { |
e046fa89 | 663 | if (!tc_qdisc_stats_dump(sch, (unsigned long)cl, arg)) |
0545a303 | 664 | return; |
0545a303 | 665 | } |
666 | } | |
667 | } | |
668 | ||
669 | static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch, | |
670 | int *qerr) | |
671 | { | |
672 | struct qfq_sched *q = qdisc_priv(sch); | |
673 | struct qfq_class *cl; | |
674 | struct tcf_result res; | |
25d8c0d5 | 675 | struct tcf_proto *fl; |
0545a303 | 676 | int result; |
677 | ||
678 | if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) { | |
679 | pr_debug("qfq_classify: found %d\n", skb->priority); | |
680 | cl = qfq_find_class(sch, skb->priority); | |
681 | if (cl != NULL) | |
682 | return cl; | |
683 | } | |
684 | ||
685 | *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS; | |
25d8c0d5 | 686 | fl = rcu_dereference_bh(q->filter_list); |
3aa26055 | 687 | result = tcf_classify(skb, NULL, fl, &res, false); |
0545a303 | 688 | if (result >= 0) { |
689 | #ifdef CONFIG_NET_CLS_ACT | |
690 | switch (result) { | |
691 | case TC_ACT_QUEUED: | |
692 | case TC_ACT_STOLEN: | |
e25ea21f | 693 | case TC_ACT_TRAP: |
0545a303 | 694 | *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN; |
964201de | 695 | fallthrough; |
0545a303 | 696 | case TC_ACT_SHOT: |
697 | return NULL; | |
698 | } | |
699 | #endif | |
700 | cl = (struct qfq_class *)res.class; | |
701 | if (cl == NULL) | |
702 | cl = qfq_find_class(sch, res.classid); | |
703 | return cl; | |
704 | } | |
705 | ||
706 | return NULL; | |
707 | } | |
708 | ||
709 | /* Generic comparison function, handling wraparound. */ | |
710 | static inline int qfq_gt(u64 a, u64 b) | |
711 | { | |
712 | return (s64)(a - b) > 0; | |
713 | } | |
714 | ||
715 | /* Round a precise timestamp to its slotted value. */ | |
716 | static inline u64 qfq_round_down(u64 ts, unsigned int shift) | |
717 | { | |
718 | return ts & ~((1ULL << shift) - 1); | |
719 | } | |
720 | ||
721 | /* return the pointer to the group with lowest index in the bitmap */ | |
722 | static inline struct qfq_group *qfq_ffs(struct qfq_sched *q, | |
723 | unsigned long bitmap) | |
724 | { | |
725 | int index = __ffs(bitmap); | |
726 | return &q->groups[index]; | |
727 | } | |
728 | /* Calculate a mask to mimic what would be ffs_from(). */ | |
729 | static inline unsigned long mask_from(unsigned long bitmap, int from) | |
730 | { | |
731 | return bitmap & ~((1UL << from) - 1); | |
732 | } | |
733 | ||
734 | /* | |
735 | * The state computation relies on ER=0, IR=1, EB=2, IB=3 | |
736 | * First compute eligibility comparing grp->S, q->V, | |
737 | * then check if someone is blocking us and possibly add EB | |
738 | */ | |
739 | static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp) | |
740 | { | |
741 | /* if S > V we are not eligible */ | |
742 | unsigned int state = qfq_gt(grp->S, q->V); | |
743 | unsigned long mask = mask_from(q->bitmaps[ER], grp->index); | |
744 | struct qfq_group *next; | |
745 | ||
746 | if (mask) { | |
747 | next = qfq_ffs(q, mask); | |
748 | if (qfq_gt(grp->F, next->F)) | |
749 | state |= EB; | |
750 | } | |
751 | ||
752 | return state; | |
753 | } | |
754 | ||
755 | ||
756 | /* | |
757 | * In principle | |
758 | * q->bitmaps[dst] |= q->bitmaps[src] & mask; | |
759 | * q->bitmaps[src] &= ~mask; | |
760 | * but we should make sure that src != dst | |
761 | */ | |
762 | static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask, | |
763 | int src, int dst) | |
764 | { | |
765 | q->bitmaps[dst] |= q->bitmaps[src] & mask; | |
766 | q->bitmaps[src] &= ~mask; | |
767 | } | |
768 | ||
769 | static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F) | |
770 | { | |
771 | unsigned long mask = mask_from(q->bitmaps[ER], index + 1); | |
772 | struct qfq_group *next; | |
773 | ||
774 | if (mask) { | |
775 | next = qfq_ffs(q, mask); | |
776 | if (!qfq_gt(next->F, old_F)) | |
777 | return; | |
778 | } | |
779 | ||
780 | mask = (1UL << index) - 1; | |
781 | qfq_move_groups(q, mask, EB, ER); | |
782 | qfq_move_groups(q, mask, IB, IR); | |
783 | } | |
784 | ||
785 | /* | |
786 | * perhaps | |
787 | * | |
788 | old_V ^= q->V; | |
462dbc91 | 789 | old_V >>= q->min_slot_shift; |
0545a303 | 790 | if (old_V) { |
791 | ... | |
792 | } | |
793 | * | |
794 | */ | |
462dbc91 | 795 | static void qfq_make_eligible(struct qfq_sched *q) |
0545a303 | 796 | { |
462dbc91 PV |
797 | unsigned long vslot = q->V >> q->min_slot_shift; |
798 | unsigned long old_vslot = q->oldV >> q->min_slot_shift; | |
0545a303 | 799 | |
800 | if (vslot != old_vslot) { | |
87f1369d PV |
801 | unsigned long mask; |
802 | int last_flip_pos = fls(vslot ^ old_vslot); | |
803 | ||
804 | if (last_flip_pos > 31) /* higher than the number of groups */ | |
805 | mask = ~0UL; /* make all groups eligible */ | |
806 | else | |
807 | mask = (1UL << last_flip_pos) - 1; | |
808 | ||
0545a303 | 809 | qfq_move_groups(q, mask, IR, ER); |
810 | qfq_move_groups(q, mask, IB, EB); | |
811 | } | |
812 | } | |
813 | ||
0545a303 | 814 | /* |
87f40dd6 PV |
815 | * The index of the slot in which the input aggregate agg is to be |
816 | * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2' | |
817 | * and not a '-1' because the start time of the group may be moved | |
818 | * backward by one slot after the aggregate has been inserted, and | |
819 | * this would cause non-empty slots to be right-shifted by one | |
820 | * position. | |
821 | * | |
822 | * QFQ+ fully satisfies this bound to the slot index if the parameters | |
823 | * of the classes are not changed dynamically, and if QFQ+ never | |
824 | * happens to postpone the service of agg unjustly, i.e., it never | |
825 | * happens that the aggregate becomes backlogged and eligible, or just | |
826 | * eligible, while an aggregate with a higher approximated finish time | |
827 | * is being served. In particular, in this case QFQ+ guarantees that | |
828 | * the timestamps of agg are low enough that the slot index is never | |
829 | * higher than 2. Unfortunately, QFQ+ cannot provide the same | |
830 | * guarantee if it happens to unjustly postpone the service of agg, or | |
831 | * if the parameters of some class are changed. | |
832 | * | |
833 | * As for the first event, i.e., an out-of-order service, the | |
834 | * upper bound to the slot index guaranteed by QFQ+ grows to | |
835 | * 2 + | |
836 | * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) * | |
837 | * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1. | |
3015f3d2 | 838 | * |
87f40dd6 PV |
839 | * The following function deals with this problem by backward-shifting |
840 | * the timestamps of agg, if needed, so as to guarantee that the slot | |
841 | * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may | |
842 | * cause the service of other aggregates to be postponed, yet the | |
843 | * worst-case guarantees of these aggregates are not violated. In | |
844 | * fact, in case of no out-of-order service, the timestamps of agg | |
845 | * would have been even lower than they are after the backward shift, | |
846 | * because QFQ+ would have guaranteed a maximum value equal to 2 for | |
847 | * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose | |
848 | * service is postponed because of the backward-shift would have | |
849 | * however waited for the service of agg before being served. | |
3015f3d2 | 850 | * |
87f40dd6 PV |
851 | * The other event that may cause the slot index to be higher than 2 |
852 | * for agg is a recent change of the parameters of some class. If the | |
853 | * weight of a class is increased or the lmax (max_pkt_size) of the | |
854 | * class is decreased, then a new aggregate with smaller slot size | |
855 | * than the original parent aggregate of the class may happen to be | |
856 | * activated. The activation of this aggregate should be properly | |
857 | * delayed to when the service of the class has finished in the ideal | |
858 | * system tracked by QFQ+. If the activation of the aggregate is not | |
859 | * delayed to this reference time instant, then this aggregate may be | |
860 | * unjustly served before other aggregates waiting for service. This | |
861 | * may cause the above bound to the slot index to be violated for some | |
862 | * of these unlucky aggregates. | |
3015f3d2 | 863 | * |
462dbc91 | 864 | * Instead of delaying the activation of the new aggregate, which is |
87f40dd6 PV |
865 | * quite complex, the above-discussed capping of the slot index is |
866 | * used to handle also the consequences of a change of the parameters | |
867 | * of a class. | |
0545a303 | 868 | */ |
462dbc91 | 869 | static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg, |
0545a303 | 870 | u64 roundedS) |
871 | { | |
872 | u64 slot = (roundedS - grp->S) >> grp->slot_shift; | |
3015f3d2 PV |
873 | unsigned int i; /* slot index in the bucket list */ |
874 | ||
875 | if (unlikely(slot > QFQ_MAX_SLOTS - 2)) { | |
876 | u64 deltaS = roundedS - grp->S - | |
877 | ((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift); | |
462dbc91 PV |
878 | agg->S -= deltaS; |
879 | agg->F -= deltaS; | |
3015f3d2 PV |
880 | slot = QFQ_MAX_SLOTS - 2; |
881 | } | |
882 | ||
883 | i = (grp->front + slot) % QFQ_MAX_SLOTS; | |
0545a303 | 884 | |
462dbc91 | 885 | hlist_add_head(&agg->next, &grp->slots[i]); |
0545a303 | 886 | __set_bit(slot, &grp->full_slots); |
887 | } | |
888 | ||
889 | /* Maybe introduce hlist_first_entry?? */ | |
462dbc91 | 890 | static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp) |
0545a303 | 891 | { |
892 | return hlist_entry(grp->slots[grp->front].first, | |
462dbc91 | 893 | struct qfq_aggregate, next); |
0545a303 | 894 | } |
895 | ||
896 | /* | |
897 | * remove the entry from the slot | |
898 | */ | |
899 | static void qfq_front_slot_remove(struct qfq_group *grp) | |
900 | { | |
462dbc91 | 901 | struct qfq_aggregate *agg = qfq_slot_head(grp); |
0545a303 | 902 | |
462dbc91 PV |
903 | BUG_ON(!agg); |
904 | hlist_del(&agg->next); | |
0545a303 | 905 | if (hlist_empty(&grp->slots[grp->front])) |
906 | __clear_bit(0, &grp->full_slots); | |
907 | } | |
908 | ||
909 | /* | |
462dbc91 PV |
910 | * Returns the first aggregate in the first non-empty bucket of the |
911 | * group. As a side effect, adjusts the bucket list so the first | |
912 | * non-empty bucket is at position 0 in full_slots. | |
0545a303 | 913 | */ |
462dbc91 | 914 | static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp) |
0545a303 | 915 | { |
916 | unsigned int i; | |
917 | ||
918 | pr_debug("qfq slot_scan: grp %u full %#lx\n", | |
919 | grp->index, grp->full_slots); | |
920 | ||
921 | if (grp->full_slots == 0) | |
922 | return NULL; | |
923 | ||
924 | i = __ffs(grp->full_slots); /* zero based */ | |
925 | if (i > 0) { | |
926 | grp->front = (grp->front + i) % QFQ_MAX_SLOTS; | |
927 | grp->full_slots >>= i; | |
928 | } | |
929 | ||
930 | return qfq_slot_head(grp); | |
931 | } | |
932 | ||
933 | /* | |
934 | * adjust the bucket list. When the start time of a group decreases, | |
935 | * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to | |
936 | * move the objects. The mask of occupied slots must be shifted | |
937 | * because we use ffs() to find the first non-empty slot. | |
938 | * This covers decreases in the group's start time, but what about | |
939 | * increases of the start time ? | |
940 | * Here too we should make sure that i is less than 32 | |
941 | */ | |
942 | static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS) | |
943 | { | |
944 | unsigned int i = (grp->S - roundedS) >> grp->slot_shift; | |
945 | ||
946 | grp->full_slots <<= i; | |
947 | grp->front = (grp->front - i) % QFQ_MAX_SLOTS; | |
948 | } | |
949 | ||
462dbc91 | 950 | static void qfq_update_eligible(struct qfq_sched *q) |
0545a303 | 951 | { |
952 | struct qfq_group *grp; | |
953 | unsigned long ineligible; | |
954 | ||
955 | ineligible = q->bitmaps[IR] | q->bitmaps[IB]; | |
956 | if (ineligible) { | |
957 | if (!q->bitmaps[ER]) { | |
958 | grp = qfq_ffs(q, ineligible); | |
959 | if (qfq_gt(grp->S, q->V)) | |
960 | q->V = grp->S; | |
961 | } | |
462dbc91 | 962 | qfq_make_eligible(q); |
0545a303 | 963 | } |
964 | } | |
965 | ||
462dbc91 PV |
966 | /* Dequeue head packet of the head class in the DRR queue of the aggregate. */ |
967 | static void agg_dequeue(struct qfq_aggregate *agg, | |
968 | struct qfq_class *cl, unsigned int len) | |
0545a303 | 969 | { |
462dbc91 | 970 | qdisc_dequeue_peeked(cl->qdisc); |
0545a303 | 971 | |
462dbc91 | 972 | cl->deficit -= (int) len; |
0545a303 | 973 | |
462dbc91 PV |
974 | if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */ |
975 | list_del(&cl->alist); | |
976 | else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) { | |
977 | cl->deficit += agg->lmax; | |
978 | list_move_tail(&cl->alist, &agg->active); | |
0545a303 | 979 | } |
462dbc91 PV |
980 | } |
981 | ||
982 | static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg, | |
983 | struct qfq_class **cl, | |
984 | unsigned int *len) | |
985 | { | |
986 | struct sk_buff *skb; | |
0545a303 | 987 | |
462dbc91 PV |
988 | *cl = list_first_entry(&agg->active, struct qfq_class, alist); |
989 | skb = (*cl)->qdisc->ops->peek((*cl)->qdisc); | |
990 | if (skb == NULL) | |
991 | WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n"); | |
992 | else | |
993 | *len = qdisc_pkt_len(skb); | |
994 | ||
995 | return skb; | |
996 | } | |
997 | ||
998 | /* Update F according to the actual service received by the aggregate. */ | |
999 | static inline void charge_actual_service(struct qfq_aggregate *agg) | |
1000 | { | |
9b99b7e9 PV |
1001 | /* Compute the service received by the aggregate, taking into |
1002 | * account that, after decreasing the number of classes in | |
1003 | * agg, it may happen that | |
1004 | * agg->initial_budget - agg->budget > agg->bugdetmax | |
1005 | */ | |
1006 | u32 service_received = min(agg->budgetmax, | |
1007 | agg->initial_budget - agg->budget); | |
462dbc91 PV |
1008 | |
1009 | agg->F = agg->S + (u64)service_received * agg->inv_w; | |
0545a303 | 1010 | } |
1011 | ||
88d4f419 PV |
1012 | /* Assign a reasonable start time for a new aggregate in group i. |
1013 | * Admissible values for \hat(F) are multiples of \sigma_i | |
1014 | * no greater than V+\sigma_i . Larger values mean that | |
1015 | * we had a wraparound so we consider the timestamp to be stale. | |
1016 | * | |
1017 | * If F is not stale and F >= V then we set S = F. | |
1018 | * Otherwise we should assign S = V, but this may violate | |
1019 | * the ordering in EB (see [2]). So, if we have groups in ER, | |
1020 | * set S to the F_j of the first group j which would be blocking us. | |
1021 | * We are guaranteed not to move S backward because | |
1022 | * otherwise our group i would still be blocked. | |
1023 | */ | |
1024 | static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg) | |
1025 | { | |
1026 | unsigned long mask; | |
1027 | u64 limit, roundedF; | |
1028 | int slot_shift = agg->grp->slot_shift; | |
1029 | ||
1030 | roundedF = qfq_round_down(agg->F, slot_shift); | |
1031 | limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift); | |
1032 | ||
1033 | if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) { | |
1034 | /* timestamp was stale */ | |
1035 | mask = mask_from(q->bitmaps[ER], agg->grp->index); | |
1036 | if (mask) { | |
1037 | struct qfq_group *next = qfq_ffs(q, mask); | |
1038 | if (qfq_gt(roundedF, next->F)) { | |
1039 | if (qfq_gt(limit, next->F)) | |
1040 | agg->S = next->F; | |
1041 | else /* preserve timestamp correctness */ | |
1042 | agg->S = limit; | |
1043 | return; | |
1044 | } | |
1045 | } | |
1046 | agg->S = q->V; | |
1047 | } else /* timestamp is not stale */ | |
1048 | agg->S = agg->F; | |
1049 | } | |
1050 | ||
1051 | /* Update the timestamps of agg before scheduling/rescheduling it for | |
1052 | * service. In particular, assign to agg->F its maximum possible | |
1053 | * value, i.e., the virtual finish time with which the aggregate | |
1054 | * should be labeled if it used all its budget once in service. | |
1055 | */ | |
1056 | static inline void | |
1057 | qfq_update_agg_ts(struct qfq_sched *q, | |
1058 | struct qfq_aggregate *agg, enum update_reason reason) | |
1059 | { | |
1060 | if (reason != requeue) | |
1061 | qfq_update_start(q, agg); | |
1062 | else /* just charge agg for the service received */ | |
1063 | agg->S = agg->F; | |
1064 | ||
1065 | agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w; | |
1066 | } | |
2f3b89a1 PV |
1067 | |
1068 | static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg); | |
1069 | ||
0545a303 | 1070 | static struct sk_buff *qfq_dequeue(struct Qdisc *sch) |
1071 | { | |
1072 | struct qfq_sched *q = qdisc_priv(sch); | |
462dbc91 | 1073 | struct qfq_aggregate *in_serv_agg = q->in_serv_agg; |
0545a303 | 1074 | struct qfq_class *cl; |
462dbc91 PV |
1075 | struct sk_buff *skb = NULL; |
1076 | /* next-packet len, 0 means no more active classes in in-service agg */ | |
1077 | unsigned int len = 0; | |
0545a303 | 1078 | |
462dbc91 | 1079 | if (in_serv_agg == NULL) |
0545a303 | 1080 | return NULL; |
1081 | ||
462dbc91 PV |
1082 | if (!list_empty(&in_serv_agg->active)) |
1083 | skb = qfq_peek_skb(in_serv_agg, &cl, &len); | |
0545a303 | 1084 | |
462dbc91 PV |
1085 | /* |
1086 | * If there are no active classes in the in-service aggregate, | |
1087 | * or if the aggregate has not enough budget to serve its next | |
1088 | * class, then choose the next aggregate to serve. | |
1089 | */ | |
1090 | if (len == 0 || in_serv_agg->budget < len) { | |
1091 | charge_actual_service(in_serv_agg); | |
1092 | ||
1093 | /* recharge the budget of the aggregate */ | |
1094 | in_serv_agg->initial_budget = in_serv_agg->budget = | |
1095 | in_serv_agg->budgetmax; | |
1096 | ||
2f3b89a1 | 1097 | if (!list_empty(&in_serv_agg->active)) { |
462dbc91 PV |
1098 | /* |
1099 | * Still active: reschedule for | |
1100 | * service. Possible optimization: if no other | |
1101 | * aggregate is active, then there is no point | |
1102 | * in rescheduling this aggregate, and we can | |
1103 | * just keep it as the in-service one. This | |
1104 | * should be however a corner case, and to | |
1105 | * handle it, we would need to maintain an | |
1106 | * extra num_active_aggs field. | |
1107 | */ | |
2f3b89a1 PV |
1108 | qfq_update_agg_ts(q, in_serv_agg, requeue); |
1109 | qfq_schedule_agg(q, in_serv_agg); | |
1110 | } else if (sch->q.qlen == 0) { /* no aggregate to serve */ | |
462dbc91 PV |
1111 | q->in_serv_agg = NULL; |
1112 | return NULL; | |
1113 | } | |
1114 | ||
1115 | /* | |
1116 | * If we get here, there are other aggregates queued: | |
1117 | * choose the new aggregate to serve. | |
1118 | */ | |
1119 | in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q); | |
1120 | skb = qfq_peek_skb(in_serv_agg, &cl, &len); | |
0545a303 | 1121 | } |
462dbc91 PV |
1122 | if (!skb) |
1123 | return NULL; | |
0545a303 | 1124 | |
2ed5c3f0 | 1125 | qdisc_qstats_backlog_dec(sch, skb); |
0545a303 | 1126 | sch->q.qlen--; |
1127 | qdisc_bstats_update(sch, skb); | |
1128 | ||
462dbc91 | 1129 | agg_dequeue(in_serv_agg, cl, len); |
a0143efa PV |
1130 | /* If lmax is lowered, through qfq_change_class, for a class |
1131 | * owning pending packets with larger size than the new value | |
1132 | * of lmax, then the following condition may hold. | |
1133 | */ | |
1134 | if (unlikely(in_serv_agg->budget < len)) | |
1135 | in_serv_agg->budget = 0; | |
1136 | else | |
1137 | in_serv_agg->budget -= len; | |
1138 | ||
87f40dd6 | 1139 | q->V += (u64)len * q->iwsum; |
0545a303 | 1140 | pr_debug("qfq dequeue: len %u F %lld now %lld\n", |
462dbc91 PV |
1141 | len, (unsigned long long) in_serv_agg->F, |
1142 | (unsigned long long) q->V); | |
0545a303 | 1143 | |
462dbc91 PV |
1144 | return skb; |
1145 | } | |
0545a303 | 1146 | |
462dbc91 PV |
1147 | static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q) |
1148 | { | |
1149 | struct qfq_group *grp; | |
1150 | struct qfq_aggregate *agg, *new_front_agg; | |
1151 | u64 old_F; | |
0545a303 | 1152 | |
462dbc91 PV |
1153 | qfq_update_eligible(q); |
1154 | q->oldV = q->V; | |
1155 | ||
1156 | if (!q->bitmaps[ER]) | |
1157 | return NULL; | |
1158 | ||
1159 | grp = qfq_ffs(q, q->bitmaps[ER]); | |
1160 | old_F = grp->F; | |
1161 | ||
1162 | agg = qfq_slot_head(grp); | |
0545a303 | 1163 | |
462dbc91 PV |
1164 | /* agg starts to be served, remove it from schedule */ |
1165 | qfq_front_slot_remove(grp); | |
1166 | ||
1167 | new_front_agg = qfq_slot_scan(grp); | |
1168 | ||
1169 | if (new_front_agg == NULL) /* group is now inactive, remove from ER */ | |
1170 | __clear_bit(grp->index, &q->bitmaps[ER]); | |
1171 | else { | |
1172 | u64 roundedS = qfq_round_down(new_front_agg->S, | |
1173 | grp->slot_shift); | |
1174 | unsigned int s; | |
1175 | ||
1176 | if (grp->S == roundedS) | |
1177 | return agg; | |
1178 | grp->S = roundedS; | |
1179 | grp->F = roundedS + (2ULL << grp->slot_shift); | |
1180 | __clear_bit(grp->index, &q->bitmaps[ER]); | |
1181 | s = qfq_calc_state(q, grp); | |
1182 | __set_bit(grp->index, &q->bitmaps[s]); | |
0545a303 | 1183 | } |
1184 | ||
462dbc91 | 1185 | qfq_unblock_groups(q, grp->index, old_F); |
0545a303 | 1186 | |
462dbc91 | 1187 | return agg; |
0545a303 | 1188 | } |
1189 | ||
ac5c66f2 | 1190 | static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch, |
520ac30f | 1191 | struct sk_buff **to_free) |
0545a303 | 1192 | { |
f6bab199 | 1193 | unsigned int len = qdisc_pkt_len(skb), gso_segs; |
0545a303 | 1194 | struct qfq_sched *q = qdisc_priv(sch); |
0545a303 | 1195 | struct qfq_class *cl; |
462dbc91 | 1196 | struct qfq_aggregate *agg; |
f54ba779 | 1197 | int err = 0; |
37d9cf1a | 1198 | bool first; |
0545a303 | 1199 | |
1200 | cl = qfq_classify(skb, sch, &err); | |
1201 | if (cl == NULL) { | |
1202 | if (err & __NET_XMIT_BYPASS) | |
25331d6c | 1203 | qdisc_qstats_drop(sch); |
39ad1297 | 1204 | __qdisc_drop(skb, to_free); |
0545a303 | 1205 | return err; |
1206 | } | |
1207 | pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid); | |
1208 | ||
f6bab199 | 1209 | if (unlikely(cl->agg->lmax < len)) { |
3015f3d2 | 1210 | pr_debug("qfq: increasing maxpkt from %u to %u for class %u", |
f6bab199 THJ |
1211 | cl->agg->lmax, len, cl->common.classid); |
1212 | err = qfq_change_agg(sch, cl, cl->agg->class_weight, len); | |
9b15350f FW |
1213 | if (err) { |
1214 | cl->qstats.drops++; | |
520ac30f | 1215 | return qdisc_drop(skb, sch, to_free); |
9b15350f | 1216 | } |
3015f3d2 PV |
1217 | } |
1218 | ||
f6bab199 | 1219 | gso_segs = skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 1; |
37d9cf1a | 1220 | first = !cl->qdisc->q.qlen; |
ac5c66f2 | 1221 | err = qdisc_enqueue(skb, cl->qdisc, to_free); |
0545a303 | 1222 | if (unlikely(err != NET_XMIT_SUCCESS)) { |
1223 | pr_debug("qfq_enqueue: enqueue failed %d\n", err); | |
1224 | if (net_xmit_drop_count(err)) { | |
1225 | cl->qstats.drops++; | |
25331d6c | 1226 | qdisc_qstats_drop(sch); |
0545a303 | 1227 | } |
1228 | return err; | |
1229 | } | |
1230 | ||
f56940da | 1231 | _bstats_update(&cl->bstats, len, gso_segs); |
f6bab199 | 1232 | sch->qstats.backlog += len; |
0545a303 | 1233 | ++sch->q.qlen; |
1234 | ||
462dbc91 PV |
1235 | agg = cl->agg; |
1236 | /* if the queue was not empty, then done here */ | |
37d9cf1a | 1237 | if (!first) { |
462dbc91 PV |
1238 | if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) && |
1239 | list_first_entry(&agg->active, struct qfq_class, alist) | |
f6bab199 | 1240 | == cl && cl->deficit < len) |
462dbc91 PV |
1241 | list_move_tail(&cl->alist, &agg->active); |
1242 | ||
0545a303 | 1243 | return err; |
462dbc91 PV |
1244 | } |
1245 | ||
1246 | /* schedule class for service within the aggregate */ | |
1247 | cl->deficit = agg->lmax; | |
1248 | list_add_tail(&cl->alist, &agg->active); | |
0545a303 | 1249 | |
2f3b89a1 PV |
1250 | if (list_first_entry(&agg->active, struct qfq_class, alist) != cl || |
1251 | q->in_serv_agg == agg) | |
1252 | return err; /* non-empty or in service, nothing else to do */ | |
462dbc91 | 1253 | |
2f3b89a1 | 1254 | qfq_activate_agg(q, agg, enqueue); |
be72f63b PV |
1255 | |
1256 | return err; | |
1257 | } | |
1258 | ||
1259 | /* | |
462dbc91 | 1260 | * Schedule aggregate according to its timestamps. |
be72f63b | 1261 | */ |
462dbc91 | 1262 | static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg) |
be72f63b | 1263 | { |
462dbc91 | 1264 | struct qfq_group *grp = agg->grp; |
be72f63b PV |
1265 | u64 roundedS; |
1266 | int s; | |
1267 | ||
462dbc91 | 1268 | roundedS = qfq_round_down(agg->S, grp->slot_shift); |
0545a303 | 1269 | |
1270 | /* | |
462dbc91 PV |
1271 | * Insert agg in the correct bucket. |
1272 | * If agg->S >= grp->S we don't need to adjust the | |
0545a303 | 1273 | * bucket list and simply go to the insertion phase. |
1274 | * Otherwise grp->S is decreasing, we must make room | |
1275 | * in the bucket list, and also recompute the group state. | |
1276 | * Finally, if there were no flows in this group and nobody | |
1277 | * was in ER make sure to adjust V. | |
1278 | */ | |
1279 | if (grp->full_slots) { | |
462dbc91 | 1280 | if (!qfq_gt(grp->S, agg->S)) |
0545a303 | 1281 | goto skip_update; |
1282 | ||
462dbc91 | 1283 | /* create a slot for this agg->S */ |
0545a303 | 1284 | qfq_slot_rotate(grp, roundedS); |
1285 | /* group was surely ineligible, remove */ | |
1286 | __clear_bit(grp->index, &q->bitmaps[IR]); | |
1287 | __clear_bit(grp->index, &q->bitmaps[IB]); | |
40dd2d54 PV |
1288 | } else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) && |
1289 | q->in_serv_agg == NULL) | |
0545a303 | 1290 | q->V = roundedS; |
1291 | ||
1292 | grp->S = roundedS; | |
1293 | grp->F = roundedS + (2ULL << grp->slot_shift); | |
1294 | s = qfq_calc_state(q, grp); | |
1295 | __set_bit(grp->index, &q->bitmaps[s]); | |
1296 | ||
1297 | pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n", | |
1298 | s, q->bitmaps[s], | |
462dbc91 PV |
1299 | (unsigned long long) agg->S, |
1300 | (unsigned long long) agg->F, | |
0545a303 | 1301 | (unsigned long long) q->V); |
1302 | ||
1303 | skip_update: | |
462dbc91 | 1304 | qfq_slot_insert(grp, agg, roundedS); |
0545a303 | 1305 | } |
1306 | ||
1307 | ||
462dbc91 PV |
1308 | /* Update agg ts and schedule agg for service */ |
1309 | static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg, | |
1310 | enum update_reason reason) | |
1311 | { | |
2f3b89a1 PV |
1312 | agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */ |
1313 | ||
462dbc91 | 1314 | qfq_update_agg_ts(q, agg, reason); |
2f3b89a1 PV |
1315 | if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */ |
1316 | q->in_serv_agg = agg; /* start serving this aggregate */ | |
1317 | /* update V: to be in service, agg must be eligible */ | |
1318 | q->oldV = q->V = agg->S; | |
1319 | } else if (agg != q->in_serv_agg) | |
1320 | qfq_schedule_agg(q, agg); | |
462dbc91 PV |
1321 | } |
1322 | ||
0545a303 | 1323 | static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp, |
462dbc91 | 1324 | struct qfq_aggregate *agg) |
0545a303 | 1325 | { |
1326 | unsigned int i, offset; | |
1327 | u64 roundedS; | |
1328 | ||
462dbc91 | 1329 | roundedS = qfq_round_down(agg->S, grp->slot_shift); |
0545a303 | 1330 | offset = (roundedS - grp->S) >> grp->slot_shift; |
462dbc91 | 1331 | |
0545a303 | 1332 | i = (grp->front + offset) % QFQ_MAX_SLOTS; |
1333 | ||
462dbc91 | 1334 | hlist_del(&agg->next); |
0545a303 | 1335 | if (hlist_empty(&grp->slots[i])) |
1336 | __clear_bit(offset, &grp->full_slots); | |
1337 | } | |
1338 | ||
1339 | /* | |
462dbc91 PV |
1340 | * Called to forcibly deschedule an aggregate. If the aggregate is |
1341 | * not in the front bucket, or if the latter has other aggregates in | |
1342 | * the front bucket, we can simply remove the aggregate with no other | |
1343 | * side effects. | |
0545a303 | 1344 | * Otherwise we must propagate the event up. |
1345 | */ | |
462dbc91 | 1346 | static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg) |
0545a303 | 1347 | { |
462dbc91 | 1348 | struct qfq_group *grp = agg->grp; |
0545a303 | 1349 | unsigned long mask; |
1350 | u64 roundedS; | |
1351 | int s; | |
1352 | ||
462dbc91 PV |
1353 | if (agg == q->in_serv_agg) { |
1354 | charge_actual_service(agg); | |
1355 | q->in_serv_agg = qfq_choose_next_agg(q); | |
1356 | return; | |
1357 | } | |
1358 | ||
1359 | agg->F = agg->S; | |
1360 | qfq_slot_remove(q, grp, agg); | |
0545a303 | 1361 | |
1362 | if (!grp->full_slots) { | |
1363 | __clear_bit(grp->index, &q->bitmaps[IR]); | |
1364 | __clear_bit(grp->index, &q->bitmaps[EB]); | |
1365 | __clear_bit(grp->index, &q->bitmaps[IB]); | |
1366 | ||
1367 | if (test_bit(grp->index, &q->bitmaps[ER]) && | |
1368 | !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) { | |
1369 | mask = q->bitmaps[ER] & ((1UL << grp->index) - 1); | |
1370 | if (mask) | |
1371 | mask = ~((1UL << __fls(mask)) - 1); | |
1372 | else | |
1373 | mask = ~0UL; | |
1374 | qfq_move_groups(q, mask, EB, ER); | |
1375 | qfq_move_groups(q, mask, IB, IR); | |
1376 | } | |
1377 | __clear_bit(grp->index, &q->bitmaps[ER]); | |
1378 | } else if (hlist_empty(&grp->slots[grp->front])) { | |
462dbc91 PV |
1379 | agg = qfq_slot_scan(grp); |
1380 | roundedS = qfq_round_down(agg->S, grp->slot_shift); | |
0545a303 | 1381 | if (grp->S != roundedS) { |
1382 | __clear_bit(grp->index, &q->bitmaps[ER]); | |
1383 | __clear_bit(grp->index, &q->bitmaps[IR]); | |
1384 | __clear_bit(grp->index, &q->bitmaps[EB]); | |
1385 | __clear_bit(grp->index, &q->bitmaps[IB]); | |
1386 | grp->S = roundedS; | |
1387 | grp->F = roundedS + (2ULL << grp->slot_shift); | |
1388 | s = qfq_calc_state(q, grp); | |
1389 | __set_bit(grp->index, &q->bitmaps[s]); | |
1390 | } | |
1391 | } | |
0545a303 | 1392 | } |
1393 | ||
1394 | static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg) | |
1395 | { | |
1396 | struct qfq_sched *q = qdisc_priv(sch); | |
1397 | struct qfq_class *cl = (struct qfq_class *)arg; | |
1398 | ||
95946658 | 1399 | qfq_deactivate_class(q, cl); |
0545a303 | 1400 | } |
1401 | ||
e63d7dfd AA |
1402 | static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt, |
1403 | struct netlink_ext_ack *extack) | |
0545a303 | 1404 | { |
1405 | struct qfq_sched *q = qdisc_priv(sch); | |
1406 | struct qfq_group *grp; | |
1407 | int i, j, err; | |
462dbc91 | 1408 | u32 max_cl_shift, maxbudg_shift, max_classes; |
0545a303 | 1409 | |
8d1a77f9 | 1410 | err = tcf_block_get(&q->block, &q->filter_list, sch, extack); |
6529eaba JP |
1411 | if (err) |
1412 | return err; | |
1413 | ||
0545a303 | 1414 | err = qdisc_class_hash_init(&q->clhash); |
1415 | if (err < 0) | |
1416 | return err; | |
1417 | ||
7d18a078 ED |
1418 | max_classes = min_t(u64, (u64)qdisc_dev(sch)->tx_queue_len + 1, |
1419 | QFQ_MAX_AGG_CLASSES); | |
462dbc91 PV |
1420 | /* max_cl_shift = floor(log_2(max_classes)) */ |
1421 | max_cl_shift = __fls(max_classes); | |
1422 | q->max_agg_classes = 1<<max_cl_shift; | |
1423 | ||
1424 | /* maxbudg_shift = log2(max_len * max_classes_per_agg) */ | |
1425 | maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift; | |
1426 | q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX; | |
1427 | ||
0545a303 | 1428 | for (i = 0; i <= QFQ_MAX_INDEX; i++) { |
1429 | grp = &q->groups[i]; | |
1430 | grp->index = i; | |
462dbc91 | 1431 | grp->slot_shift = q->min_slot_shift + i; |
0545a303 | 1432 | for (j = 0; j < QFQ_MAX_SLOTS; j++) |
1433 | INIT_HLIST_HEAD(&grp->slots[j]); | |
1434 | } | |
1435 | ||
462dbc91 PV |
1436 | INIT_HLIST_HEAD(&q->nonfull_aggs); |
1437 | ||
0545a303 | 1438 | return 0; |
1439 | } | |
1440 | ||
1441 | static void qfq_reset_qdisc(struct Qdisc *sch) | |
1442 | { | |
1443 | struct qfq_sched *q = qdisc_priv(sch); | |
0545a303 | 1444 | struct qfq_class *cl; |
462dbc91 | 1445 | unsigned int i; |
0545a303 | 1446 | |
462dbc91 | 1447 | for (i = 0; i < q->clhash.hashsize; i++) { |
b67bfe0d | 1448 | hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) { |
462dbc91 | 1449 | if (cl->qdisc->q.qlen > 0) |
0545a303 | 1450 | qfq_deactivate_class(q, cl); |
0545a303 | 1451 | |
0545a303 | 1452 | qdisc_reset(cl->qdisc); |
462dbc91 | 1453 | } |
0545a303 | 1454 | } |
0545a303 | 1455 | } |
1456 | ||
1457 | static void qfq_destroy_qdisc(struct Qdisc *sch) | |
1458 | { | |
1459 | struct qfq_sched *q = qdisc_priv(sch); | |
1460 | struct qfq_class *cl; | |
b67bfe0d | 1461 | struct hlist_node *next; |
0545a303 | 1462 | unsigned int i; |
1463 | ||
6529eaba | 1464 | tcf_block_put(q->block); |
0545a303 | 1465 | |
1466 | for (i = 0; i < q->clhash.hashsize; i++) { | |
b67bfe0d | 1467 | hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i], |
0545a303 | 1468 | common.hnode) { |
1469 | qfq_destroy_class(sch, cl); | |
1470 | } | |
1471 | } | |
1472 | qdisc_class_hash_destroy(&q->clhash); | |
1473 | } | |
1474 | ||
1475 | static const struct Qdisc_class_ops qfq_class_ops = { | |
1476 | .change = qfq_change_class, | |
1477 | .delete = qfq_delete_class, | |
143976ce | 1478 | .find = qfq_search_class, |
6529eaba | 1479 | .tcf_block = qfq_tcf_block, |
0545a303 | 1480 | .bind_tcf = qfq_bind_tcf, |
1481 | .unbind_tcf = qfq_unbind_tcf, | |
1482 | .graft = qfq_graft_class, | |
1483 | .leaf = qfq_class_leaf, | |
1484 | .qlen_notify = qfq_qlen_notify, | |
1485 | .dump = qfq_dump_class, | |
1486 | .dump_stats = qfq_dump_class_stats, | |
1487 | .walk = qfq_walk, | |
1488 | }; | |
1489 | ||
1490 | static struct Qdisc_ops qfq_qdisc_ops __read_mostly = { | |
1491 | .cl_ops = &qfq_class_ops, | |
1492 | .id = "qfq", | |
1493 | .priv_size = sizeof(struct qfq_sched), | |
1494 | .enqueue = qfq_enqueue, | |
1495 | .dequeue = qfq_dequeue, | |
1496 | .peek = qdisc_peek_dequeued, | |
0545a303 | 1497 | .init = qfq_init_qdisc, |
1498 | .reset = qfq_reset_qdisc, | |
1499 | .destroy = qfq_destroy_qdisc, | |
1500 | .owner = THIS_MODULE, | |
1501 | }; | |
1502 | ||
1503 | static int __init qfq_init(void) | |
1504 | { | |
1505 | return register_qdisc(&qfq_qdisc_ops); | |
1506 | } | |
1507 | ||
1508 | static void __exit qfq_exit(void) | |
1509 | { | |
1510 | unregister_qdisc(&qfq_qdisc_ops); | |
1511 | } | |
1512 | ||
1513 | module_init(qfq_init); | |
1514 | module_exit(qfq_exit); | |
1515 | MODULE_LICENSE("GPL"); |