]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - net/sched/sch_qfq.c
Merge tag 'libnvdimm-for-4.12' of git://git.kernel.org/pub/scm/linux/kernel/git/nvdim...
[mirror_ubuntu-artful-kernel.git] / net / sched / sch_qfq.c
1 /*
2 * net/sched/sch_qfq.c Quick Fair Queueing Plus Scheduler.
3 *
4 * Copyright (c) 2009 Fabio Checconi, Luigi Rizzo, and Paolo Valente.
5 * Copyright (c) 2012 Paolo Valente.
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
23 /* Quick Fair Queueing Plus
24 ========================
25
26 Sources:
27
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
35 Packet Scheduling with Tight Bandwidth Distribution Guarantees."
36
37 See also:
38 http://retis.sssup.it/~fabio/linux/qfq/
39 */
40
41 /*
42
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
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 /*
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
102 * group with the smallest index that can support the L_i / r_i configured
103 * for the classes in the aggregate.
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 */
108 #define QFQ_MAX_INDEX 24
109 #define QFQ_MAX_WSHIFT 10
110
111 #define QFQ_MAX_WEIGHT (1<<QFQ_MAX_WSHIFT) /* see qfq_slot_insert */
112 #define QFQ_MAX_WSUM (64*QFQ_MAX_WEIGHT)
113
114 #define FRAC_BITS 30 /* fixed point arithmetic */
115 #define ONE_FP (1UL << FRAC_BITS)
116
117 #define QFQ_MTU_SHIFT 16 /* to support TSO/GSO */
118 #define QFQ_MIN_LMAX 512 /* see qfq_slot_insert */
119
120 #define QFQ_MAX_AGG_CLASSES 8 /* max num classes per aggregate allowed */
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
130 struct qfq_aggregate;
131
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;
140 struct net_rate_estimator __rcu *rate_est;
141 struct Qdisc *qdisc;
142 struct list_head alist; /* Link for active-classes list. */
143 struct qfq_aggregate *agg; /* Parent aggregate. */
144 int deficit; /* DRR deficit counter. */
145 };
146
147 struct qfq_aggregate {
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. */
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. */
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
179 /* Array of RR lists of active aggregates. */
180 struct hlist_head slots[QFQ_MAX_SLOTS];
181 };
182
183 struct qfq_sched {
184 struct tcf_proto __rcu *filter_list;
185 struct Qdisc_class_hash clhash;
186
187 u64 oldV, V; /* Precise virtual times. */
188 struct qfq_aggregate *in_serv_agg; /* Aggregate being served. */
189 u32 wsum; /* weight sum */
190 u32 iwsum; /* inverse weight sum */
191
192 unsigned long bitmaps[QFQ_MAX_STATE]; /* Group bitmaps. */
193 struct qfq_group groups[QFQ_MAX_INDEX + 1]; /* The groups. */
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. */
198 };
199
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
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 unsigned int backlog = cl->qdisc->qstats.backlog;
224
225 qdisc_reset(cl->qdisc);
226 qdisc_tree_reduce_backlog(cl->qdisc, len, backlog);
227 }
228
229 static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = {
230 [TCA_QFQ_WEIGHT] = { .type = NLA_U32 },
231 [TCA_QFQ_LMAX] = { .type = NLA_U32 },
232 };
233
234 /*
235 * Calculate a flow index, given its weight and maximum packet length.
236 * index = log_2(maxlen/weight) but we need to apply the scaling.
237 * This is used only once at flow creation.
238 */
239 static int qfq_calc_index(u32 inv_w, unsigned int maxlen, u32 min_slot_shift)
240 {
241 u64 slot_size = (u64)maxlen * inv_w;
242 unsigned long size_map;
243 int index = 0;
244
245 size_map = slot_size >> min_slot_shift;
246 if (!size_map)
247 goto out;
248
249 index = __fls(size_map) + 1; /* basically a log_2 */
250 index -= !(slot_size - (1ULL << (index + min_slot_shift - 1)));
251
252 if (index < 0)
253 index = 0;
254 out:
255 pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n",
256 (unsigned long) ONE_FP/inv_w, maxlen, index);
257
258 return index;
259 }
260
261 static void qfq_deactivate_agg(struct qfq_sched *, struct qfq_aggregate *);
262 static void qfq_activate_agg(struct qfq_sched *, struct qfq_aggregate *,
263 enum update_reason);
264
265 static void qfq_init_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
266 u32 lmax, u32 weight)
267 {
268 INIT_LIST_HEAD(&agg->active);
269 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
270
271 agg->lmax = lmax;
272 agg->class_weight = weight;
273 }
274
275 static struct qfq_aggregate *qfq_find_agg(struct qfq_sched *q,
276 u32 lmax, u32 weight)
277 {
278 struct qfq_aggregate *agg;
279
280 hlist_for_each_entry(agg, &q->nonfull_aggs, nonfull_next)
281 if (agg->lmax == lmax && agg->class_weight == weight)
282 return agg;
283
284 return NULL;
285 }
286
287
288 /* Update aggregate as a function of the new number of classes. */
289 static void qfq_update_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
290 int new_num_classes)
291 {
292 u32 new_agg_weight;
293
294 if (new_num_classes == q->max_agg_classes)
295 hlist_del_init(&agg->nonfull_next);
296
297 if (agg->num_classes > new_num_classes &&
298 new_num_classes == q->max_agg_classes - 1) /* agg no more full */
299 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
300
301 /* The next assignment may let
302 * agg->initial_budget > agg->budgetmax
303 * hold, we will take it into account in charge_actual_service().
304 */
305 agg->budgetmax = new_num_classes * agg->lmax;
306 new_agg_weight = agg->class_weight * new_num_classes;
307 agg->inv_w = ONE_FP/new_agg_weight;
308
309 if (agg->grp == NULL) {
310 int i = qfq_calc_index(agg->inv_w, agg->budgetmax,
311 q->min_slot_shift);
312 agg->grp = &q->groups[i];
313 }
314
315 q->wsum +=
316 (int) agg->class_weight * (new_num_classes - agg->num_classes);
317 q->iwsum = ONE_FP / q->wsum;
318
319 agg->num_classes = new_num_classes;
320 }
321
322 /* Add class to aggregate. */
323 static void qfq_add_to_agg(struct qfq_sched *q,
324 struct qfq_aggregate *agg,
325 struct qfq_class *cl)
326 {
327 cl->agg = agg;
328
329 qfq_update_agg(q, agg, agg->num_classes+1);
330 if (cl->qdisc->q.qlen > 0) { /* adding an active class */
331 list_add_tail(&cl->alist, &agg->active);
332 if (list_first_entry(&agg->active, struct qfq_class, alist) ==
333 cl && q->in_serv_agg != agg) /* agg was inactive */
334 qfq_activate_agg(q, agg, enqueue); /* schedule agg */
335 }
336 }
337
338 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *);
339
340 static void qfq_destroy_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
341 {
342 hlist_del_init(&agg->nonfull_next);
343 q->wsum -= agg->class_weight;
344 if (q->wsum != 0)
345 q->iwsum = ONE_FP / q->wsum;
346
347 if (q->in_serv_agg == agg)
348 q->in_serv_agg = qfq_choose_next_agg(q);
349 kfree(agg);
350 }
351
352 /* Deschedule class from within its parent aggregate. */
353 static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
354 {
355 struct qfq_aggregate *agg = cl->agg;
356
357
358 list_del(&cl->alist); /* remove from RR queue of the aggregate */
359 if (list_empty(&agg->active)) /* agg is now inactive */
360 qfq_deactivate_agg(q, agg);
361 }
362
363 /* Remove class from its parent aggregate. */
364 static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
365 {
366 struct qfq_aggregate *agg = cl->agg;
367
368 cl->agg = NULL;
369 if (agg->num_classes == 1) { /* agg being emptied, destroy it */
370 qfq_destroy_agg(q, agg);
371 return;
372 }
373 qfq_update_agg(q, agg, agg->num_classes-1);
374 }
375
376 /* Deschedule class and remove it from its parent aggregate. */
377 static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
378 {
379 if (cl->qdisc->q.qlen > 0) /* class is active */
380 qfq_deactivate_class(q, cl);
381
382 qfq_rm_from_agg(q, cl);
383 }
384
385 /* Move class to a new aggregate, matching the new class weight and/or lmax */
386 static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight,
387 u32 lmax)
388 {
389 struct qfq_sched *q = qdisc_priv(sch);
390 struct qfq_aggregate *new_agg = qfq_find_agg(q, lmax, weight);
391
392 if (new_agg == NULL) { /* create new aggregate */
393 new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC);
394 if (new_agg == NULL)
395 return -ENOBUFS;
396 qfq_init_agg(q, new_agg, lmax, weight);
397 }
398 qfq_deact_rm_from_agg(q, cl);
399 qfq_add_to_agg(q, new_agg, cl);
400
401 return 0;
402 }
403
404 static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
405 struct nlattr **tca, unsigned long *arg)
406 {
407 struct qfq_sched *q = qdisc_priv(sch);
408 struct qfq_class *cl = (struct qfq_class *)*arg;
409 bool existing = false;
410 struct nlattr *tb[TCA_QFQ_MAX + 1];
411 struct qfq_aggregate *new_agg = NULL;
412 u32 weight, lmax, inv_w;
413 int err;
414 int delta_w;
415
416 if (tca[TCA_OPTIONS] == NULL) {
417 pr_notice("qfq: no options\n");
418 return -EINVAL;
419 }
420
421 err = nla_parse_nested(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS], qfq_policy,
422 NULL);
423 if (err < 0)
424 return err;
425
426 if (tb[TCA_QFQ_WEIGHT]) {
427 weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]);
428 if (!weight || weight > (1UL << QFQ_MAX_WSHIFT)) {
429 pr_notice("qfq: invalid weight %u\n", weight);
430 return -EINVAL;
431 }
432 } else
433 weight = 1;
434
435 if (tb[TCA_QFQ_LMAX]) {
436 lmax = nla_get_u32(tb[TCA_QFQ_LMAX]);
437 if (lmax < QFQ_MIN_LMAX || lmax > (1UL << QFQ_MTU_SHIFT)) {
438 pr_notice("qfq: invalid max length %u\n", lmax);
439 return -EINVAL;
440 }
441 } else
442 lmax = psched_mtu(qdisc_dev(sch));
443
444 inv_w = ONE_FP / weight;
445 weight = ONE_FP / inv_w;
446
447 if (cl != NULL &&
448 lmax == cl->agg->lmax &&
449 weight == cl->agg->class_weight)
450 return 0; /* nothing to change */
451
452 delta_w = weight - (cl ? cl->agg->class_weight : 0);
453
454 if (q->wsum + delta_w > QFQ_MAX_WSUM) {
455 pr_notice("qfq: total weight out of range (%d + %u)\n",
456 delta_w, q->wsum);
457 return -EINVAL;
458 }
459
460 if (cl != NULL) { /* modify existing class */
461 if (tca[TCA_RATE]) {
462 err = gen_replace_estimator(&cl->bstats, NULL,
463 &cl->rate_est,
464 NULL,
465 qdisc_root_sleeping_running(sch),
466 tca[TCA_RATE]);
467 if (err)
468 return err;
469 }
470 existing = true;
471 goto set_change_agg;
472 }
473
474 /* create and init new class */
475 cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
476 if (cl == NULL)
477 return -ENOBUFS;
478
479 cl->refcnt = 1;
480 cl->common.classid = classid;
481 cl->deficit = lmax;
482
483 cl->qdisc = qdisc_create_dflt(sch->dev_queue,
484 &pfifo_qdisc_ops, classid);
485 if (cl->qdisc == NULL)
486 cl->qdisc = &noop_qdisc;
487
488 if (tca[TCA_RATE]) {
489 err = gen_new_estimator(&cl->bstats, NULL,
490 &cl->rate_est,
491 NULL,
492 qdisc_root_sleeping_running(sch),
493 tca[TCA_RATE]);
494 if (err)
495 goto destroy_class;
496 }
497
498 if (cl->qdisc != &noop_qdisc)
499 qdisc_hash_add(cl->qdisc, true);
500 sch_tree_lock(sch);
501 qdisc_class_hash_insert(&q->clhash, &cl->common);
502 sch_tree_unlock(sch);
503
504 qdisc_class_hash_grow(sch, &q->clhash);
505
506 set_change_agg:
507 sch_tree_lock(sch);
508 new_agg = qfq_find_agg(q, lmax, weight);
509 if (new_agg == NULL) { /* create new aggregate */
510 sch_tree_unlock(sch);
511 new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL);
512 if (new_agg == NULL) {
513 err = -ENOBUFS;
514 gen_kill_estimator(&cl->rate_est);
515 goto destroy_class;
516 }
517 sch_tree_lock(sch);
518 qfq_init_agg(q, new_agg, lmax, weight);
519 }
520 if (existing)
521 qfq_deact_rm_from_agg(q, cl);
522 qfq_add_to_agg(q, new_agg, cl);
523 sch_tree_unlock(sch);
524
525 *arg = (unsigned long)cl;
526 return 0;
527
528 destroy_class:
529 qdisc_destroy(cl->qdisc);
530 kfree(cl);
531 return err;
532 }
533
534 static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
535 {
536 struct qfq_sched *q = qdisc_priv(sch);
537
538 qfq_rm_from_agg(q, cl);
539 gen_kill_estimator(&cl->rate_est);
540 qdisc_destroy(cl->qdisc);
541 kfree(cl);
542 }
543
544 static int qfq_delete_class(struct Qdisc *sch, unsigned long arg)
545 {
546 struct qfq_sched *q = qdisc_priv(sch);
547 struct qfq_class *cl = (struct qfq_class *)arg;
548
549 if (cl->filter_cnt > 0)
550 return -EBUSY;
551
552 sch_tree_lock(sch);
553
554 qfq_purge_queue(cl);
555 qdisc_class_hash_remove(&q->clhash, &cl->common);
556
557 BUG_ON(--cl->refcnt == 0);
558 /*
559 * This shouldn't happen: we "hold" one cops->get() when called
560 * from tc_ctl_tclass; the destroy method is done from cops->put().
561 */
562
563 sch_tree_unlock(sch);
564 return 0;
565 }
566
567 static unsigned long qfq_get_class(struct Qdisc *sch, u32 classid)
568 {
569 struct qfq_class *cl = qfq_find_class(sch, classid);
570
571 if (cl != NULL)
572 cl->refcnt++;
573
574 return (unsigned long)cl;
575 }
576
577 static void qfq_put_class(struct Qdisc *sch, unsigned long arg)
578 {
579 struct qfq_class *cl = (struct qfq_class *)arg;
580
581 if (--cl->refcnt == 0)
582 qfq_destroy_class(sch, cl);
583 }
584
585 static struct tcf_proto __rcu **qfq_tcf_chain(struct Qdisc *sch,
586 unsigned long cl)
587 {
588 struct qfq_sched *q = qdisc_priv(sch);
589
590 if (cl)
591 return NULL;
592
593 return &q->filter_list;
594 }
595
596 static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
597 u32 classid)
598 {
599 struct qfq_class *cl = qfq_find_class(sch, classid);
600
601 if (cl != NULL)
602 cl->filter_cnt++;
603
604 return (unsigned long)cl;
605 }
606
607 static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
608 {
609 struct qfq_class *cl = (struct qfq_class *)arg;
610
611 cl->filter_cnt--;
612 }
613
614 static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
615 struct Qdisc *new, struct Qdisc **old)
616 {
617 struct qfq_class *cl = (struct qfq_class *)arg;
618
619 if (new == NULL) {
620 new = qdisc_create_dflt(sch->dev_queue,
621 &pfifo_qdisc_ops, cl->common.classid);
622 if (new == NULL)
623 new = &noop_qdisc;
624 }
625
626 *old = qdisc_replace(sch, new, &cl->qdisc);
627 return 0;
628 }
629
630 static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
631 {
632 struct qfq_class *cl = (struct qfq_class *)arg;
633
634 return cl->qdisc;
635 }
636
637 static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
638 struct sk_buff *skb, struct tcmsg *tcm)
639 {
640 struct qfq_class *cl = (struct qfq_class *)arg;
641 struct nlattr *nest;
642
643 tcm->tcm_parent = TC_H_ROOT;
644 tcm->tcm_handle = cl->common.classid;
645 tcm->tcm_info = cl->qdisc->handle;
646
647 nest = nla_nest_start(skb, TCA_OPTIONS);
648 if (nest == NULL)
649 goto nla_put_failure;
650 if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) ||
651 nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax))
652 goto nla_put_failure;
653 return nla_nest_end(skb, nest);
654
655 nla_put_failure:
656 nla_nest_cancel(skb, nest);
657 return -EMSGSIZE;
658 }
659
660 static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
661 struct gnet_dump *d)
662 {
663 struct qfq_class *cl = (struct qfq_class *)arg;
664 struct tc_qfq_stats xstats;
665
666 memset(&xstats, 0, sizeof(xstats));
667
668 xstats.weight = cl->agg->class_weight;
669 xstats.lmax = cl->agg->lmax;
670
671 if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
672 d, NULL, &cl->bstats) < 0 ||
673 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
674 gnet_stats_copy_queue(d, NULL,
675 &cl->qdisc->qstats, cl->qdisc->q.qlen) < 0)
676 return -1;
677
678 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
679 }
680
681 static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
682 {
683 struct qfq_sched *q = qdisc_priv(sch);
684 struct qfq_class *cl;
685 unsigned int i;
686
687 if (arg->stop)
688 return;
689
690 for (i = 0; i < q->clhash.hashsize; i++) {
691 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
692 if (arg->count < arg->skip) {
693 arg->count++;
694 continue;
695 }
696 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
697 arg->stop = 1;
698 return;
699 }
700 arg->count++;
701 }
702 }
703 }
704
705 static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
706 int *qerr)
707 {
708 struct qfq_sched *q = qdisc_priv(sch);
709 struct qfq_class *cl;
710 struct tcf_result res;
711 struct tcf_proto *fl;
712 int result;
713
714 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
715 pr_debug("qfq_classify: found %d\n", skb->priority);
716 cl = qfq_find_class(sch, skb->priority);
717 if (cl != NULL)
718 return cl;
719 }
720
721 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
722 fl = rcu_dereference_bh(q->filter_list);
723 result = tc_classify(skb, fl, &res, false);
724 if (result >= 0) {
725 #ifdef CONFIG_NET_CLS_ACT
726 switch (result) {
727 case TC_ACT_QUEUED:
728 case TC_ACT_STOLEN:
729 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
730 case TC_ACT_SHOT:
731 return NULL;
732 }
733 #endif
734 cl = (struct qfq_class *)res.class;
735 if (cl == NULL)
736 cl = qfq_find_class(sch, res.classid);
737 return cl;
738 }
739
740 return NULL;
741 }
742
743 /* Generic comparison function, handling wraparound. */
744 static inline int qfq_gt(u64 a, u64 b)
745 {
746 return (s64)(a - b) > 0;
747 }
748
749 /* Round a precise timestamp to its slotted value. */
750 static inline u64 qfq_round_down(u64 ts, unsigned int shift)
751 {
752 return ts & ~((1ULL << shift) - 1);
753 }
754
755 /* return the pointer to the group with lowest index in the bitmap */
756 static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
757 unsigned long bitmap)
758 {
759 int index = __ffs(bitmap);
760 return &q->groups[index];
761 }
762 /* Calculate a mask to mimic what would be ffs_from(). */
763 static inline unsigned long mask_from(unsigned long bitmap, int from)
764 {
765 return bitmap & ~((1UL << from) - 1);
766 }
767
768 /*
769 * The state computation relies on ER=0, IR=1, EB=2, IB=3
770 * First compute eligibility comparing grp->S, q->V,
771 * then check if someone is blocking us and possibly add EB
772 */
773 static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
774 {
775 /* if S > V we are not eligible */
776 unsigned int state = qfq_gt(grp->S, q->V);
777 unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
778 struct qfq_group *next;
779
780 if (mask) {
781 next = qfq_ffs(q, mask);
782 if (qfq_gt(grp->F, next->F))
783 state |= EB;
784 }
785
786 return state;
787 }
788
789
790 /*
791 * In principle
792 * q->bitmaps[dst] |= q->bitmaps[src] & mask;
793 * q->bitmaps[src] &= ~mask;
794 * but we should make sure that src != dst
795 */
796 static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
797 int src, int dst)
798 {
799 q->bitmaps[dst] |= q->bitmaps[src] & mask;
800 q->bitmaps[src] &= ~mask;
801 }
802
803 static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
804 {
805 unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
806 struct qfq_group *next;
807
808 if (mask) {
809 next = qfq_ffs(q, mask);
810 if (!qfq_gt(next->F, old_F))
811 return;
812 }
813
814 mask = (1UL << index) - 1;
815 qfq_move_groups(q, mask, EB, ER);
816 qfq_move_groups(q, mask, IB, IR);
817 }
818
819 /*
820 * perhaps
821 *
822 old_V ^= q->V;
823 old_V >>= q->min_slot_shift;
824 if (old_V) {
825 ...
826 }
827 *
828 */
829 static void qfq_make_eligible(struct qfq_sched *q)
830 {
831 unsigned long vslot = q->V >> q->min_slot_shift;
832 unsigned long old_vslot = q->oldV >> q->min_slot_shift;
833
834 if (vslot != old_vslot) {
835 unsigned long mask;
836 int last_flip_pos = fls(vslot ^ old_vslot);
837
838 if (last_flip_pos > 31) /* higher than the number of groups */
839 mask = ~0UL; /* make all groups eligible */
840 else
841 mask = (1UL << last_flip_pos) - 1;
842
843 qfq_move_groups(q, mask, IR, ER);
844 qfq_move_groups(q, mask, IB, EB);
845 }
846 }
847
848 /*
849 * The index of the slot in which the input aggregate agg is to be
850 * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2'
851 * and not a '-1' because the start time of the group may be moved
852 * backward by one slot after the aggregate has been inserted, and
853 * this would cause non-empty slots to be right-shifted by one
854 * position.
855 *
856 * QFQ+ fully satisfies this bound to the slot index if the parameters
857 * of the classes are not changed dynamically, and if QFQ+ never
858 * happens to postpone the service of agg unjustly, i.e., it never
859 * happens that the aggregate becomes backlogged and eligible, or just
860 * eligible, while an aggregate with a higher approximated finish time
861 * is being served. In particular, in this case QFQ+ guarantees that
862 * the timestamps of agg are low enough that the slot index is never
863 * higher than 2. Unfortunately, QFQ+ cannot provide the same
864 * guarantee if it happens to unjustly postpone the service of agg, or
865 * if the parameters of some class are changed.
866 *
867 * As for the first event, i.e., an out-of-order service, the
868 * upper bound to the slot index guaranteed by QFQ+ grows to
869 * 2 +
870 * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
871 * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1.
872 *
873 * The following function deals with this problem by backward-shifting
874 * the timestamps of agg, if needed, so as to guarantee that the slot
875 * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may
876 * cause the service of other aggregates to be postponed, yet the
877 * worst-case guarantees of these aggregates are not violated. In
878 * fact, in case of no out-of-order service, the timestamps of agg
879 * would have been even lower than they are after the backward shift,
880 * because QFQ+ would have guaranteed a maximum value equal to 2 for
881 * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose
882 * service is postponed because of the backward-shift would have
883 * however waited for the service of agg before being served.
884 *
885 * The other event that may cause the slot index to be higher than 2
886 * for agg is a recent change of the parameters of some class. If the
887 * weight of a class is increased or the lmax (max_pkt_size) of the
888 * class is decreased, then a new aggregate with smaller slot size
889 * than the original parent aggregate of the class may happen to be
890 * activated. The activation of this aggregate should be properly
891 * delayed to when the service of the class has finished in the ideal
892 * system tracked by QFQ+. If the activation of the aggregate is not
893 * delayed to this reference time instant, then this aggregate may be
894 * unjustly served before other aggregates waiting for service. This
895 * may cause the above bound to the slot index to be violated for some
896 * of these unlucky aggregates.
897 *
898 * Instead of delaying the activation of the new aggregate, which is
899 * quite complex, the above-discussed capping of the slot index is
900 * used to handle also the consequences of a change of the parameters
901 * of a class.
902 */
903 static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg,
904 u64 roundedS)
905 {
906 u64 slot = (roundedS - grp->S) >> grp->slot_shift;
907 unsigned int i; /* slot index in the bucket list */
908
909 if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
910 u64 deltaS = roundedS - grp->S -
911 ((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
912 agg->S -= deltaS;
913 agg->F -= deltaS;
914 slot = QFQ_MAX_SLOTS - 2;
915 }
916
917 i = (grp->front + slot) % QFQ_MAX_SLOTS;
918
919 hlist_add_head(&agg->next, &grp->slots[i]);
920 __set_bit(slot, &grp->full_slots);
921 }
922
923 /* Maybe introduce hlist_first_entry?? */
924 static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp)
925 {
926 return hlist_entry(grp->slots[grp->front].first,
927 struct qfq_aggregate, next);
928 }
929
930 /*
931 * remove the entry from the slot
932 */
933 static void qfq_front_slot_remove(struct qfq_group *grp)
934 {
935 struct qfq_aggregate *agg = qfq_slot_head(grp);
936
937 BUG_ON(!agg);
938 hlist_del(&agg->next);
939 if (hlist_empty(&grp->slots[grp->front]))
940 __clear_bit(0, &grp->full_slots);
941 }
942
943 /*
944 * Returns the first aggregate in the first non-empty bucket of the
945 * group. As a side effect, adjusts the bucket list so the first
946 * non-empty bucket is at position 0 in full_slots.
947 */
948 static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp)
949 {
950 unsigned int i;
951
952 pr_debug("qfq slot_scan: grp %u full %#lx\n",
953 grp->index, grp->full_slots);
954
955 if (grp->full_slots == 0)
956 return NULL;
957
958 i = __ffs(grp->full_slots); /* zero based */
959 if (i > 0) {
960 grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
961 grp->full_slots >>= i;
962 }
963
964 return qfq_slot_head(grp);
965 }
966
967 /*
968 * adjust the bucket list. When the start time of a group decreases,
969 * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
970 * move the objects. The mask of occupied slots must be shifted
971 * because we use ffs() to find the first non-empty slot.
972 * This covers decreases in the group's start time, but what about
973 * increases of the start time ?
974 * Here too we should make sure that i is less than 32
975 */
976 static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
977 {
978 unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
979
980 grp->full_slots <<= i;
981 grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
982 }
983
984 static void qfq_update_eligible(struct qfq_sched *q)
985 {
986 struct qfq_group *grp;
987 unsigned long ineligible;
988
989 ineligible = q->bitmaps[IR] | q->bitmaps[IB];
990 if (ineligible) {
991 if (!q->bitmaps[ER]) {
992 grp = qfq_ffs(q, ineligible);
993 if (qfq_gt(grp->S, q->V))
994 q->V = grp->S;
995 }
996 qfq_make_eligible(q);
997 }
998 }
999
1000 /* Dequeue head packet of the head class in the DRR queue of the aggregate. */
1001 static void agg_dequeue(struct qfq_aggregate *agg,
1002 struct qfq_class *cl, unsigned int len)
1003 {
1004 qdisc_dequeue_peeked(cl->qdisc);
1005
1006 cl->deficit -= (int) len;
1007
1008 if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */
1009 list_del(&cl->alist);
1010 else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) {
1011 cl->deficit += agg->lmax;
1012 list_move_tail(&cl->alist, &agg->active);
1013 }
1014 }
1015
1016 static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg,
1017 struct qfq_class **cl,
1018 unsigned int *len)
1019 {
1020 struct sk_buff *skb;
1021
1022 *cl = list_first_entry(&agg->active, struct qfq_class, alist);
1023 skb = (*cl)->qdisc->ops->peek((*cl)->qdisc);
1024 if (skb == NULL)
1025 WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
1026 else
1027 *len = qdisc_pkt_len(skb);
1028
1029 return skb;
1030 }
1031
1032 /* Update F according to the actual service received by the aggregate. */
1033 static inline void charge_actual_service(struct qfq_aggregate *agg)
1034 {
1035 /* Compute the service received by the aggregate, taking into
1036 * account that, after decreasing the number of classes in
1037 * agg, it may happen that
1038 * agg->initial_budget - agg->budget > agg->bugdetmax
1039 */
1040 u32 service_received = min(agg->budgetmax,
1041 agg->initial_budget - agg->budget);
1042
1043 agg->F = agg->S + (u64)service_received * agg->inv_w;
1044 }
1045
1046 /* Assign a reasonable start time for a new aggregate in group i.
1047 * Admissible values for \hat(F) are multiples of \sigma_i
1048 * no greater than V+\sigma_i . Larger values mean that
1049 * we had a wraparound so we consider the timestamp to be stale.
1050 *
1051 * If F is not stale and F >= V then we set S = F.
1052 * Otherwise we should assign S = V, but this may violate
1053 * the ordering in EB (see [2]). So, if we have groups in ER,
1054 * set S to the F_j of the first group j which would be blocking us.
1055 * We are guaranteed not to move S backward because
1056 * otherwise our group i would still be blocked.
1057 */
1058 static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg)
1059 {
1060 unsigned long mask;
1061 u64 limit, roundedF;
1062 int slot_shift = agg->grp->slot_shift;
1063
1064 roundedF = qfq_round_down(agg->F, slot_shift);
1065 limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
1066
1067 if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) {
1068 /* timestamp was stale */
1069 mask = mask_from(q->bitmaps[ER], agg->grp->index);
1070 if (mask) {
1071 struct qfq_group *next = qfq_ffs(q, mask);
1072 if (qfq_gt(roundedF, next->F)) {
1073 if (qfq_gt(limit, next->F))
1074 agg->S = next->F;
1075 else /* preserve timestamp correctness */
1076 agg->S = limit;
1077 return;
1078 }
1079 }
1080 agg->S = q->V;
1081 } else /* timestamp is not stale */
1082 agg->S = agg->F;
1083 }
1084
1085 /* Update the timestamps of agg before scheduling/rescheduling it for
1086 * service. In particular, assign to agg->F its maximum possible
1087 * value, i.e., the virtual finish time with which the aggregate
1088 * should be labeled if it used all its budget once in service.
1089 */
1090 static inline void
1091 qfq_update_agg_ts(struct qfq_sched *q,
1092 struct qfq_aggregate *agg, enum update_reason reason)
1093 {
1094 if (reason != requeue)
1095 qfq_update_start(q, agg);
1096 else /* just charge agg for the service received */
1097 agg->S = agg->F;
1098
1099 agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w;
1100 }
1101
1102 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg);
1103
1104 static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
1105 {
1106 struct qfq_sched *q = qdisc_priv(sch);
1107 struct qfq_aggregate *in_serv_agg = q->in_serv_agg;
1108 struct qfq_class *cl;
1109 struct sk_buff *skb = NULL;
1110 /* next-packet len, 0 means no more active classes in in-service agg */
1111 unsigned int len = 0;
1112
1113 if (in_serv_agg == NULL)
1114 return NULL;
1115
1116 if (!list_empty(&in_serv_agg->active))
1117 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1118
1119 /*
1120 * If there are no active classes in the in-service aggregate,
1121 * or if the aggregate has not enough budget to serve its next
1122 * class, then choose the next aggregate to serve.
1123 */
1124 if (len == 0 || in_serv_agg->budget < len) {
1125 charge_actual_service(in_serv_agg);
1126
1127 /* recharge the budget of the aggregate */
1128 in_serv_agg->initial_budget = in_serv_agg->budget =
1129 in_serv_agg->budgetmax;
1130
1131 if (!list_empty(&in_serv_agg->active)) {
1132 /*
1133 * Still active: reschedule for
1134 * service. Possible optimization: if no other
1135 * aggregate is active, then there is no point
1136 * in rescheduling this aggregate, and we can
1137 * just keep it as the in-service one. This
1138 * should be however a corner case, and to
1139 * handle it, we would need to maintain an
1140 * extra num_active_aggs field.
1141 */
1142 qfq_update_agg_ts(q, in_serv_agg, requeue);
1143 qfq_schedule_agg(q, in_serv_agg);
1144 } else if (sch->q.qlen == 0) { /* no aggregate to serve */
1145 q->in_serv_agg = NULL;
1146 return NULL;
1147 }
1148
1149 /*
1150 * If we get here, there are other aggregates queued:
1151 * choose the new aggregate to serve.
1152 */
1153 in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
1154 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1155 }
1156 if (!skb)
1157 return NULL;
1158
1159 qdisc_qstats_backlog_dec(sch, skb);
1160 sch->q.qlen--;
1161 qdisc_bstats_update(sch, skb);
1162
1163 agg_dequeue(in_serv_agg, cl, len);
1164 /* If lmax is lowered, through qfq_change_class, for a class
1165 * owning pending packets with larger size than the new value
1166 * of lmax, then the following condition may hold.
1167 */
1168 if (unlikely(in_serv_agg->budget < len))
1169 in_serv_agg->budget = 0;
1170 else
1171 in_serv_agg->budget -= len;
1172
1173 q->V += (u64)len * q->iwsum;
1174 pr_debug("qfq dequeue: len %u F %lld now %lld\n",
1175 len, (unsigned long long) in_serv_agg->F,
1176 (unsigned long long) q->V);
1177
1178 return skb;
1179 }
1180
1181 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)
1182 {
1183 struct qfq_group *grp;
1184 struct qfq_aggregate *agg, *new_front_agg;
1185 u64 old_F;
1186
1187 qfq_update_eligible(q);
1188 q->oldV = q->V;
1189
1190 if (!q->bitmaps[ER])
1191 return NULL;
1192
1193 grp = qfq_ffs(q, q->bitmaps[ER]);
1194 old_F = grp->F;
1195
1196 agg = qfq_slot_head(grp);
1197
1198 /* agg starts to be served, remove it from schedule */
1199 qfq_front_slot_remove(grp);
1200
1201 new_front_agg = qfq_slot_scan(grp);
1202
1203 if (new_front_agg == NULL) /* group is now inactive, remove from ER */
1204 __clear_bit(grp->index, &q->bitmaps[ER]);
1205 else {
1206 u64 roundedS = qfq_round_down(new_front_agg->S,
1207 grp->slot_shift);
1208 unsigned int s;
1209
1210 if (grp->S == roundedS)
1211 return agg;
1212 grp->S = roundedS;
1213 grp->F = roundedS + (2ULL << grp->slot_shift);
1214 __clear_bit(grp->index, &q->bitmaps[ER]);
1215 s = qfq_calc_state(q, grp);
1216 __set_bit(grp->index, &q->bitmaps[s]);
1217 }
1218
1219 qfq_unblock_groups(q, grp->index, old_F);
1220
1221 return agg;
1222 }
1223
1224 static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
1225 struct sk_buff **to_free)
1226 {
1227 struct qfq_sched *q = qdisc_priv(sch);
1228 struct qfq_class *cl;
1229 struct qfq_aggregate *agg;
1230 int err = 0;
1231
1232 cl = qfq_classify(skb, sch, &err);
1233 if (cl == NULL) {
1234 if (err & __NET_XMIT_BYPASS)
1235 qdisc_qstats_drop(sch);
1236 kfree_skb(skb);
1237 return err;
1238 }
1239 pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
1240
1241 if (unlikely(cl->agg->lmax < qdisc_pkt_len(skb))) {
1242 pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
1243 cl->agg->lmax, qdisc_pkt_len(skb), cl->common.classid);
1244 err = qfq_change_agg(sch, cl, cl->agg->class_weight,
1245 qdisc_pkt_len(skb));
1246 if (err) {
1247 cl->qstats.drops++;
1248 return qdisc_drop(skb, sch, to_free);
1249 }
1250 }
1251
1252 err = qdisc_enqueue(skb, cl->qdisc, to_free);
1253 if (unlikely(err != NET_XMIT_SUCCESS)) {
1254 pr_debug("qfq_enqueue: enqueue failed %d\n", err);
1255 if (net_xmit_drop_count(err)) {
1256 cl->qstats.drops++;
1257 qdisc_qstats_drop(sch);
1258 }
1259 return err;
1260 }
1261
1262 bstats_update(&cl->bstats, skb);
1263 qdisc_qstats_backlog_inc(sch, skb);
1264 ++sch->q.qlen;
1265
1266 agg = cl->agg;
1267 /* if the queue was not empty, then done here */
1268 if (cl->qdisc->q.qlen != 1) {
1269 if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) &&
1270 list_first_entry(&agg->active, struct qfq_class, alist)
1271 == cl && cl->deficit < qdisc_pkt_len(skb))
1272 list_move_tail(&cl->alist, &agg->active);
1273
1274 return err;
1275 }
1276
1277 /* schedule class for service within the aggregate */
1278 cl->deficit = agg->lmax;
1279 list_add_tail(&cl->alist, &agg->active);
1280
1281 if (list_first_entry(&agg->active, struct qfq_class, alist) != cl ||
1282 q->in_serv_agg == agg)
1283 return err; /* non-empty or in service, nothing else to do */
1284
1285 qfq_activate_agg(q, agg, enqueue);
1286
1287 return err;
1288 }
1289
1290 /*
1291 * Schedule aggregate according to its timestamps.
1292 */
1293 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1294 {
1295 struct qfq_group *grp = agg->grp;
1296 u64 roundedS;
1297 int s;
1298
1299 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1300
1301 /*
1302 * Insert agg in the correct bucket.
1303 * If agg->S >= grp->S we don't need to adjust the
1304 * bucket list and simply go to the insertion phase.
1305 * Otherwise grp->S is decreasing, we must make room
1306 * in the bucket list, and also recompute the group state.
1307 * Finally, if there were no flows in this group and nobody
1308 * was in ER make sure to adjust V.
1309 */
1310 if (grp->full_slots) {
1311 if (!qfq_gt(grp->S, agg->S))
1312 goto skip_update;
1313
1314 /* create a slot for this agg->S */
1315 qfq_slot_rotate(grp, roundedS);
1316 /* group was surely ineligible, remove */
1317 __clear_bit(grp->index, &q->bitmaps[IR]);
1318 __clear_bit(grp->index, &q->bitmaps[IB]);
1319 } else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) &&
1320 q->in_serv_agg == NULL)
1321 q->V = roundedS;
1322
1323 grp->S = roundedS;
1324 grp->F = roundedS + (2ULL << grp->slot_shift);
1325 s = qfq_calc_state(q, grp);
1326 __set_bit(grp->index, &q->bitmaps[s]);
1327
1328 pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
1329 s, q->bitmaps[s],
1330 (unsigned long long) agg->S,
1331 (unsigned long long) agg->F,
1332 (unsigned long long) q->V);
1333
1334 skip_update:
1335 qfq_slot_insert(grp, agg, roundedS);
1336 }
1337
1338
1339 /* Update agg ts and schedule agg for service */
1340 static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
1341 enum update_reason reason)
1342 {
1343 agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */
1344
1345 qfq_update_agg_ts(q, agg, reason);
1346 if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */
1347 q->in_serv_agg = agg; /* start serving this aggregate */
1348 /* update V: to be in service, agg must be eligible */
1349 q->oldV = q->V = agg->S;
1350 } else if (agg != q->in_serv_agg)
1351 qfq_schedule_agg(q, agg);
1352 }
1353
1354 static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
1355 struct qfq_aggregate *agg)
1356 {
1357 unsigned int i, offset;
1358 u64 roundedS;
1359
1360 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1361 offset = (roundedS - grp->S) >> grp->slot_shift;
1362
1363 i = (grp->front + offset) % QFQ_MAX_SLOTS;
1364
1365 hlist_del(&agg->next);
1366 if (hlist_empty(&grp->slots[i]))
1367 __clear_bit(offset, &grp->full_slots);
1368 }
1369
1370 /*
1371 * Called to forcibly deschedule an aggregate. If the aggregate is
1372 * not in the front bucket, or if the latter has other aggregates in
1373 * the front bucket, we can simply remove the aggregate with no other
1374 * side effects.
1375 * Otherwise we must propagate the event up.
1376 */
1377 static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1378 {
1379 struct qfq_group *grp = agg->grp;
1380 unsigned long mask;
1381 u64 roundedS;
1382 int s;
1383
1384 if (agg == q->in_serv_agg) {
1385 charge_actual_service(agg);
1386 q->in_serv_agg = qfq_choose_next_agg(q);
1387 return;
1388 }
1389
1390 agg->F = agg->S;
1391 qfq_slot_remove(q, grp, agg);
1392
1393 if (!grp->full_slots) {
1394 __clear_bit(grp->index, &q->bitmaps[IR]);
1395 __clear_bit(grp->index, &q->bitmaps[EB]);
1396 __clear_bit(grp->index, &q->bitmaps[IB]);
1397
1398 if (test_bit(grp->index, &q->bitmaps[ER]) &&
1399 !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
1400 mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
1401 if (mask)
1402 mask = ~((1UL << __fls(mask)) - 1);
1403 else
1404 mask = ~0UL;
1405 qfq_move_groups(q, mask, EB, ER);
1406 qfq_move_groups(q, mask, IB, IR);
1407 }
1408 __clear_bit(grp->index, &q->bitmaps[ER]);
1409 } else if (hlist_empty(&grp->slots[grp->front])) {
1410 agg = qfq_slot_scan(grp);
1411 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1412 if (grp->S != roundedS) {
1413 __clear_bit(grp->index, &q->bitmaps[ER]);
1414 __clear_bit(grp->index, &q->bitmaps[IR]);
1415 __clear_bit(grp->index, &q->bitmaps[EB]);
1416 __clear_bit(grp->index, &q->bitmaps[IB]);
1417 grp->S = roundedS;
1418 grp->F = roundedS + (2ULL << grp->slot_shift);
1419 s = qfq_calc_state(q, grp);
1420 __set_bit(grp->index, &q->bitmaps[s]);
1421 }
1422 }
1423 }
1424
1425 static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1426 {
1427 struct qfq_sched *q = qdisc_priv(sch);
1428 struct qfq_class *cl = (struct qfq_class *)arg;
1429
1430 if (cl->qdisc->q.qlen == 0)
1431 qfq_deactivate_class(q, cl);
1432 }
1433
1434 static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
1435 {
1436 struct qfq_sched *q = qdisc_priv(sch);
1437 struct qfq_group *grp;
1438 int i, j, err;
1439 u32 max_cl_shift, maxbudg_shift, max_classes;
1440
1441 err = qdisc_class_hash_init(&q->clhash);
1442 if (err < 0)
1443 return err;
1444
1445 if (qdisc_dev(sch)->tx_queue_len + 1 > QFQ_MAX_AGG_CLASSES)
1446 max_classes = QFQ_MAX_AGG_CLASSES;
1447 else
1448 max_classes = qdisc_dev(sch)->tx_queue_len + 1;
1449 /* max_cl_shift = floor(log_2(max_classes)) */
1450 max_cl_shift = __fls(max_classes);
1451 q->max_agg_classes = 1<<max_cl_shift;
1452
1453 /* maxbudg_shift = log2(max_len * max_classes_per_agg) */
1454 maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift;
1455 q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX;
1456
1457 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1458 grp = &q->groups[i];
1459 grp->index = i;
1460 grp->slot_shift = q->min_slot_shift + i;
1461 for (j = 0; j < QFQ_MAX_SLOTS; j++)
1462 INIT_HLIST_HEAD(&grp->slots[j]);
1463 }
1464
1465 INIT_HLIST_HEAD(&q->nonfull_aggs);
1466
1467 return 0;
1468 }
1469
1470 static void qfq_reset_qdisc(struct Qdisc *sch)
1471 {
1472 struct qfq_sched *q = qdisc_priv(sch);
1473 struct qfq_class *cl;
1474 unsigned int i;
1475
1476 for (i = 0; i < q->clhash.hashsize; i++) {
1477 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1478 if (cl->qdisc->q.qlen > 0)
1479 qfq_deactivate_class(q, cl);
1480
1481 qdisc_reset(cl->qdisc);
1482 }
1483 }
1484 sch->qstats.backlog = 0;
1485 sch->q.qlen = 0;
1486 }
1487
1488 static void qfq_destroy_qdisc(struct Qdisc *sch)
1489 {
1490 struct qfq_sched *q = qdisc_priv(sch);
1491 struct qfq_class *cl;
1492 struct hlist_node *next;
1493 unsigned int i;
1494
1495 tcf_destroy_chain(&q->filter_list);
1496
1497 for (i = 0; i < q->clhash.hashsize; i++) {
1498 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1499 common.hnode) {
1500 qfq_destroy_class(sch, cl);
1501 }
1502 }
1503 qdisc_class_hash_destroy(&q->clhash);
1504 }
1505
1506 static const struct Qdisc_class_ops qfq_class_ops = {
1507 .change = qfq_change_class,
1508 .delete = qfq_delete_class,
1509 .get = qfq_get_class,
1510 .put = qfq_put_class,
1511 .tcf_chain = qfq_tcf_chain,
1512 .bind_tcf = qfq_bind_tcf,
1513 .unbind_tcf = qfq_unbind_tcf,
1514 .graft = qfq_graft_class,
1515 .leaf = qfq_class_leaf,
1516 .qlen_notify = qfq_qlen_notify,
1517 .dump = qfq_dump_class,
1518 .dump_stats = qfq_dump_class_stats,
1519 .walk = qfq_walk,
1520 };
1521
1522 static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1523 .cl_ops = &qfq_class_ops,
1524 .id = "qfq",
1525 .priv_size = sizeof(struct qfq_sched),
1526 .enqueue = qfq_enqueue,
1527 .dequeue = qfq_dequeue,
1528 .peek = qdisc_peek_dequeued,
1529 .init = qfq_init_qdisc,
1530 .reset = qfq_reset_qdisc,
1531 .destroy = qfq_destroy_qdisc,
1532 .owner = THIS_MODULE,
1533 };
1534
1535 static int __init qfq_init(void)
1536 {
1537 return register_qdisc(&qfq_qdisc_ops);
1538 }
1539
1540 static void __exit qfq_exit(void)
1541 {
1542 unregister_qdisc(&qfq_qdisc_ops);
1543 }
1544
1545 module_init(qfq_init);
1546 module_exit(qfq_exit);
1547 MODULE_LICENSE("GPL");