ARM: dts: kirkwood: gpio pin fixes for linkstation ls-wxl/wsxl
[linux/fpc-iii.git] / net / sched / sch_qfq.c
blob3dc3a6e560520a6f3b9b3bdf3bfa88277a08c266
1 /*
2 * net/sched/sch_qfq.c Quick Fair Queueing Plus Scheduler.
4 * Copyright (c) 2009 Fabio Checconi, Luigi Rizzo, and Paolo Valente.
5 * Copyright (c) 2012 Paolo Valente.
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.
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>
23 /* Quick Fair Queueing Plus
24 ========================
26 Sources:
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
32 Sources for QFQ:
34 [2] Fabio Checconi, Luigi Rizzo, and Paolo Valente: "QFQ: Efficient
35 Packet Scheduling with Tight Bandwidth Distribution Guarantees."
37 See also:
38 http://retis.sssup.it/~fabio/linux/qfq/
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.
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.
57 Virtual time computations.
59 S, F and V are all computed in fixed point arithmetic with
60 FRAC_BITS decimal bits.
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.
66 The layout of the bits is as below:
68 [ MTU_SHIFT ][ FRAC_BITS ]
69 [ MAX_INDEX ][ MIN_SLOT_SHIFT ]
70 ^.__grp->index = 0
71 *.__grp->slot_shift
73 where MIN_SLOT_SHIFT is derived by difference from the others.
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.
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.
88 The per-scheduler-instance data contain all the data structures
89 for the scheduler: bitmaps and bucket lists.
94 * Maximum number of consecutive slots occupied by backlogged classes
95 * inside a group.
97 #define QFQ_MAX_SLOTS 32
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.
105 * grp->index is the index of the group; and grp->slot_shift
106 * is the shift for the corresponding (scaled) sigma_i.
108 #define QFQ_MAX_INDEX 24
109 #define QFQ_MAX_WSHIFT 10
111 #define QFQ_MAX_WEIGHT (1<<QFQ_MAX_WSHIFT) /* see qfq_slot_insert */
112 #define QFQ_MAX_WSUM (64*QFQ_MAX_WEIGHT)
114 #define FRAC_BITS 30 /* fixed point arithmetic */
115 #define ONE_FP (1UL << FRAC_BITS)
117 #define QFQ_MTU_SHIFT 16 /* to support TSO/GSO */
118 #define QFQ_MIN_LMAX 512 /* see qfq_slot_insert */
120 #define QFQ_MAX_AGG_CLASSES 8 /* max num classes per aggregate allowed */
123 * Possible group states. These values are used as indexes for the bitmaps
124 * array of struct qfq_queue.
126 enum qfq_state { ER, IR, EB, IB, QFQ_MAX_STATE };
128 struct qfq_group;
130 struct qfq_aggregate;
132 struct qfq_class {
133 struct Qdisc_class_common common;
135 unsigned int refcnt;
136 unsigned int filter_cnt;
138 struct gnet_stats_basic_packed bstats;
139 struct gnet_stats_queue qstats;
140 struct gnet_stats_rate_est64 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. */
147 struct qfq_aggregate {
148 struct hlist_node next; /* Link for the slot list. */
149 u64 S, F; /* flow timestamps (exact) */
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.
155 struct qfq_group *grp;
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;
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. */
166 int num_classes; /* Number of classes in this aggr. */
167 struct list_head active; /* DRR queue of active classes. */
169 struct hlist_node nonfull_next; /* See nonfull_aggs in qfq_sched. */
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 */
179 /* Array of RR lists of active aggregates. */
180 struct hlist_head slots[QFQ_MAX_SLOTS];
183 struct qfq_sched {
184 struct tcf_proto __rcu *filter_list;
185 struct Qdisc_class_hash clhash;
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 */
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. */
196 u32 max_agg_classes; /* Max number of classes per aggr. */
197 struct hlist_head nonfull_aggs; /* Aggs with room for more classes. */
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
207 enum update_reason {enqueue, requeue};
209 static struct qfq_class *qfq_find_class(struct Qdisc *sch, u32 classid)
211 struct qfq_sched *q = qdisc_priv(sch);
212 struct Qdisc_class_common *clc;
214 clc = qdisc_class_find(&q->clhash, classid);
215 if (clc == NULL)
216 return NULL;
217 return container_of(clc, struct qfq_class, common);
220 static void qfq_purge_queue(struct qfq_class *cl)
222 unsigned int len = cl->qdisc->q.qlen;
224 qdisc_reset(cl->qdisc);
225 qdisc_tree_decrease_qlen(cl->qdisc, len);
228 static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = {
229 [TCA_QFQ_WEIGHT] = { .type = NLA_U32 },
230 [TCA_QFQ_LMAX] = { .type = NLA_U32 },
234 * Calculate a flow index, given its weight and maximum packet length.
235 * index = log_2(maxlen/weight) but we need to apply the scaling.
236 * This is used only once at flow creation.
238 static int qfq_calc_index(u32 inv_w, unsigned int maxlen, u32 min_slot_shift)
240 u64 slot_size = (u64)maxlen * inv_w;
241 unsigned long size_map;
242 int index = 0;
244 size_map = slot_size >> min_slot_shift;
245 if (!size_map)
246 goto out;
248 index = __fls(size_map) + 1; /* basically a log_2 */
249 index -= !(slot_size - (1ULL << (index + min_slot_shift - 1)));
251 if (index < 0)
252 index = 0;
253 out:
254 pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n",
255 (unsigned long) ONE_FP/inv_w, maxlen, index);
257 return index;
260 static void qfq_deactivate_agg(struct qfq_sched *, struct qfq_aggregate *);
261 static void qfq_activate_agg(struct qfq_sched *, struct qfq_aggregate *,
262 enum update_reason);
264 static void qfq_init_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
265 u32 lmax, u32 weight)
267 INIT_LIST_HEAD(&agg->active);
268 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
270 agg->lmax = lmax;
271 agg->class_weight = weight;
274 static struct qfq_aggregate *qfq_find_agg(struct qfq_sched *q,
275 u32 lmax, u32 weight)
277 struct qfq_aggregate *agg;
279 hlist_for_each_entry(agg, &q->nonfull_aggs, nonfull_next)
280 if (agg->lmax == lmax && agg->class_weight == weight)
281 return agg;
283 return NULL;
287 /* Update aggregate as a function of the new number of classes. */
288 static void qfq_update_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
289 int new_num_classes)
291 u32 new_agg_weight;
293 if (new_num_classes == q->max_agg_classes)
294 hlist_del_init(&agg->nonfull_next);
296 if (agg->num_classes > new_num_classes &&
297 new_num_classes == q->max_agg_classes - 1) /* agg no more full */
298 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
300 /* The next assignment may let
301 * agg->initial_budget > agg->budgetmax
302 * hold, we will take it into account in charge_actual_service().
304 agg->budgetmax = new_num_classes * agg->lmax;
305 new_agg_weight = agg->class_weight * new_num_classes;
306 agg->inv_w = ONE_FP/new_agg_weight;
308 if (agg->grp == NULL) {
309 int i = qfq_calc_index(agg->inv_w, agg->budgetmax,
310 q->min_slot_shift);
311 agg->grp = &q->groups[i];
314 q->wsum +=
315 (int) agg->class_weight * (new_num_classes - agg->num_classes);
316 q->iwsum = ONE_FP / q->wsum;
318 agg->num_classes = new_num_classes;
321 /* Add class to aggregate. */
322 static void qfq_add_to_agg(struct qfq_sched *q,
323 struct qfq_aggregate *agg,
324 struct qfq_class *cl)
326 cl->agg = agg;
328 qfq_update_agg(q, agg, agg->num_classes+1);
329 if (cl->qdisc->q.qlen > 0) { /* adding an active class */
330 list_add_tail(&cl->alist, &agg->active);
331 if (list_first_entry(&agg->active, struct qfq_class, alist) ==
332 cl && q->in_serv_agg != agg) /* agg was inactive */
333 qfq_activate_agg(q, agg, enqueue); /* schedule agg */
337 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *);
339 static void qfq_destroy_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
341 hlist_del_init(&agg->nonfull_next);
342 q->wsum -= agg->class_weight;
343 if (q->wsum != 0)
344 q->iwsum = ONE_FP / q->wsum;
346 if (q->in_serv_agg == agg)
347 q->in_serv_agg = qfq_choose_next_agg(q);
348 kfree(agg);
351 /* Deschedule class from within its parent aggregate. */
352 static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
354 struct qfq_aggregate *agg = cl->agg;
357 list_del(&cl->alist); /* remove from RR queue of the aggregate */
358 if (list_empty(&agg->active)) /* agg is now inactive */
359 qfq_deactivate_agg(q, agg);
362 /* Remove class from its parent aggregate. */
363 static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
365 struct qfq_aggregate *agg = cl->agg;
367 cl->agg = NULL;
368 if (agg->num_classes == 1) { /* agg being emptied, destroy it */
369 qfq_destroy_agg(q, agg);
370 return;
372 qfq_update_agg(q, agg, agg->num_classes-1);
375 /* Deschedule class and remove it from its parent aggregate. */
376 static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
378 if (cl->qdisc->q.qlen > 0) /* class is active */
379 qfq_deactivate_class(q, cl);
381 qfq_rm_from_agg(q, cl);
384 /* Move class to a new aggregate, matching the new class weight and/or lmax */
385 static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight,
386 u32 lmax)
388 struct qfq_sched *q = qdisc_priv(sch);
389 struct qfq_aggregate *new_agg = qfq_find_agg(q, lmax, weight);
391 if (new_agg == NULL) { /* create new aggregate */
392 new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC);
393 if (new_agg == NULL)
394 return -ENOBUFS;
395 qfq_init_agg(q, new_agg, lmax, weight);
397 qfq_deact_rm_from_agg(q, cl);
398 qfq_add_to_agg(q, new_agg, cl);
400 return 0;
403 static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
404 struct nlattr **tca, unsigned long *arg)
406 struct qfq_sched *q = qdisc_priv(sch);
407 struct qfq_class *cl = (struct qfq_class *)*arg;
408 bool existing = false;
409 struct nlattr *tb[TCA_QFQ_MAX + 1];
410 struct qfq_aggregate *new_agg = NULL;
411 u32 weight, lmax, inv_w;
412 int err;
413 int delta_w;
415 if (tca[TCA_OPTIONS] == NULL) {
416 pr_notice("qfq: no options\n");
417 return -EINVAL;
420 err = nla_parse_nested(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS], qfq_policy);
421 if (err < 0)
422 return err;
424 if (tb[TCA_QFQ_WEIGHT]) {
425 weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]);
426 if (!weight || weight > (1UL << QFQ_MAX_WSHIFT)) {
427 pr_notice("qfq: invalid weight %u\n", weight);
428 return -EINVAL;
430 } else
431 weight = 1;
433 if (tb[TCA_QFQ_LMAX]) {
434 lmax = nla_get_u32(tb[TCA_QFQ_LMAX]);
435 if (lmax < QFQ_MIN_LMAX || lmax > (1UL << QFQ_MTU_SHIFT)) {
436 pr_notice("qfq: invalid max length %u\n", lmax);
437 return -EINVAL;
439 } else
440 lmax = psched_mtu(qdisc_dev(sch));
442 inv_w = ONE_FP / weight;
443 weight = ONE_FP / inv_w;
445 if (cl != NULL &&
446 lmax == cl->agg->lmax &&
447 weight == cl->agg->class_weight)
448 return 0; /* nothing to change */
450 delta_w = weight - (cl ? cl->agg->class_weight : 0);
452 if (q->wsum + delta_w > QFQ_MAX_WSUM) {
453 pr_notice("qfq: total weight out of range (%d + %u)\n",
454 delta_w, q->wsum);
455 return -EINVAL;
458 if (cl != NULL) { /* modify existing class */
459 if (tca[TCA_RATE]) {
460 err = gen_replace_estimator(&cl->bstats, NULL,
461 &cl->rate_est,
462 qdisc_root_sleeping_lock(sch),
463 tca[TCA_RATE]);
464 if (err)
465 return err;
467 existing = true;
468 goto set_change_agg;
471 /* create and init new class */
472 cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
473 if (cl == NULL)
474 return -ENOBUFS;
476 cl->refcnt = 1;
477 cl->common.classid = classid;
478 cl->deficit = lmax;
480 cl->qdisc = qdisc_create_dflt(sch->dev_queue,
481 &pfifo_qdisc_ops, classid);
482 if (cl->qdisc == NULL)
483 cl->qdisc = &noop_qdisc;
485 if (tca[TCA_RATE]) {
486 err = gen_new_estimator(&cl->bstats, NULL,
487 &cl->rate_est,
488 qdisc_root_sleeping_lock(sch),
489 tca[TCA_RATE]);
490 if (err)
491 goto destroy_class;
494 sch_tree_lock(sch);
495 qdisc_class_hash_insert(&q->clhash, &cl->common);
496 sch_tree_unlock(sch);
498 qdisc_class_hash_grow(sch, &q->clhash);
500 set_change_agg:
501 sch_tree_lock(sch);
502 new_agg = qfq_find_agg(q, lmax, weight);
503 if (new_agg == NULL) { /* create new aggregate */
504 sch_tree_unlock(sch);
505 new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL);
506 if (new_agg == NULL) {
507 err = -ENOBUFS;
508 gen_kill_estimator(&cl->bstats, &cl->rate_est);
509 goto destroy_class;
511 sch_tree_lock(sch);
512 qfq_init_agg(q, new_agg, lmax, weight);
514 if (existing)
515 qfq_deact_rm_from_agg(q, cl);
516 qfq_add_to_agg(q, new_agg, cl);
517 sch_tree_unlock(sch);
519 *arg = (unsigned long)cl;
520 return 0;
522 destroy_class:
523 qdisc_destroy(cl->qdisc);
524 kfree(cl);
525 return err;
528 static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
530 struct qfq_sched *q = qdisc_priv(sch);
532 qfq_rm_from_agg(q, cl);
533 gen_kill_estimator(&cl->bstats, &cl->rate_est);
534 qdisc_destroy(cl->qdisc);
535 kfree(cl);
538 static int qfq_delete_class(struct Qdisc *sch, unsigned long arg)
540 struct qfq_sched *q = qdisc_priv(sch);
541 struct qfq_class *cl = (struct qfq_class *)arg;
543 if (cl->filter_cnt > 0)
544 return -EBUSY;
546 sch_tree_lock(sch);
548 qfq_purge_queue(cl);
549 qdisc_class_hash_remove(&q->clhash, &cl->common);
551 BUG_ON(--cl->refcnt == 0);
553 * This shouldn't happen: we "hold" one cops->get() when called
554 * from tc_ctl_tclass; the destroy method is done from cops->put().
557 sch_tree_unlock(sch);
558 return 0;
561 static unsigned long qfq_get_class(struct Qdisc *sch, u32 classid)
563 struct qfq_class *cl = qfq_find_class(sch, classid);
565 if (cl != NULL)
566 cl->refcnt++;
568 return (unsigned long)cl;
571 static void qfq_put_class(struct Qdisc *sch, unsigned long arg)
573 struct qfq_class *cl = (struct qfq_class *)arg;
575 if (--cl->refcnt == 0)
576 qfq_destroy_class(sch, cl);
579 static struct tcf_proto __rcu **qfq_tcf_chain(struct Qdisc *sch,
580 unsigned long cl)
582 struct qfq_sched *q = qdisc_priv(sch);
584 if (cl)
585 return NULL;
587 return &q->filter_list;
590 static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
591 u32 classid)
593 struct qfq_class *cl = qfq_find_class(sch, classid);
595 if (cl != NULL)
596 cl->filter_cnt++;
598 return (unsigned long)cl;
601 static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
603 struct qfq_class *cl = (struct qfq_class *)arg;
605 cl->filter_cnt--;
608 static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
609 struct Qdisc *new, struct Qdisc **old)
611 struct qfq_class *cl = (struct qfq_class *)arg;
613 if (new == NULL) {
614 new = qdisc_create_dflt(sch->dev_queue,
615 &pfifo_qdisc_ops, cl->common.classid);
616 if (new == NULL)
617 new = &noop_qdisc;
620 sch_tree_lock(sch);
621 qfq_purge_queue(cl);
622 *old = cl->qdisc;
623 cl->qdisc = new;
624 sch_tree_unlock(sch);
625 return 0;
628 static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
630 struct qfq_class *cl = (struct qfq_class *)arg;
632 return cl->qdisc;
635 static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
636 struct sk_buff *skb, struct tcmsg *tcm)
638 struct qfq_class *cl = (struct qfq_class *)arg;
639 struct nlattr *nest;
641 tcm->tcm_parent = TC_H_ROOT;
642 tcm->tcm_handle = cl->common.classid;
643 tcm->tcm_info = cl->qdisc->handle;
645 nest = nla_nest_start(skb, TCA_OPTIONS);
646 if (nest == NULL)
647 goto nla_put_failure;
648 if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) ||
649 nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax))
650 goto nla_put_failure;
651 return nla_nest_end(skb, nest);
653 nla_put_failure:
654 nla_nest_cancel(skb, nest);
655 return -EMSGSIZE;
658 static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
659 struct gnet_dump *d)
661 struct qfq_class *cl = (struct qfq_class *)arg;
662 struct tc_qfq_stats xstats;
664 memset(&xstats, 0, sizeof(xstats));
666 xstats.weight = cl->agg->class_weight;
667 xstats.lmax = cl->agg->lmax;
669 if (gnet_stats_copy_basic(d, NULL, &cl->bstats) < 0 ||
670 gnet_stats_copy_rate_est(d, &cl->bstats, &cl->rate_est) < 0 ||
671 gnet_stats_copy_queue(d, NULL,
672 &cl->qdisc->qstats, cl->qdisc->q.qlen) < 0)
673 return -1;
675 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
678 static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
680 struct qfq_sched *q = qdisc_priv(sch);
681 struct qfq_class *cl;
682 unsigned int i;
684 if (arg->stop)
685 return;
687 for (i = 0; i < q->clhash.hashsize; i++) {
688 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
689 if (arg->count < arg->skip) {
690 arg->count++;
691 continue;
693 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
694 arg->stop = 1;
695 return;
697 arg->count++;
702 static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
703 int *qerr)
705 struct qfq_sched *q = qdisc_priv(sch);
706 struct qfq_class *cl;
707 struct tcf_result res;
708 struct tcf_proto *fl;
709 int result;
711 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
712 pr_debug("qfq_classify: found %d\n", skb->priority);
713 cl = qfq_find_class(sch, skb->priority);
714 if (cl != NULL)
715 return cl;
718 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
719 fl = rcu_dereference_bh(q->filter_list);
720 result = tc_classify(skb, fl, &res, false);
721 if (result >= 0) {
722 #ifdef CONFIG_NET_CLS_ACT
723 switch (result) {
724 case TC_ACT_QUEUED:
725 case TC_ACT_STOLEN:
726 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
727 case TC_ACT_SHOT:
728 return NULL;
730 #endif
731 cl = (struct qfq_class *)res.class;
732 if (cl == NULL)
733 cl = qfq_find_class(sch, res.classid);
734 return cl;
737 return NULL;
740 /* Generic comparison function, handling wraparound. */
741 static inline int qfq_gt(u64 a, u64 b)
743 return (s64)(a - b) > 0;
746 /* Round a precise timestamp to its slotted value. */
747 static inline u64 qfq_round_down(u64 ts, unsigned int shift)
749 return ts & ~((1ULL << shift) - 1);
752 /* return the pointer to the group with lowest index in the bitmap */
753 static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
754 unsigned long bitmap)
756 int index = __ffs(bitmap);
757 return &q->groups[index];
759 /* Calculate a mask to mimic what would be ffs_from(). */
760 static inline unsigned long mask_from(unsigned long bitmap, int from)
762 return bitmap & ~((1UL << from) - 1);
766 * The state computation relies on ER=0, IR=1, EB=2, IB=3
767 * First compute eligibility comparing grp->S, q->V,
768 * then check if someone is blocking us and possibly add EB
770 static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
772 /* if S > V we are not eligible */
773 unsigned int state = qfq_gt(grp->S, q->V);
774 unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
775 struct qfq_group *next;
777 if (mask) {
778 next = qfq_ffs(q, mask);
779 if (qfq_gt(grp->F, next->F))
780 state |= EB;
783 return state;
788 * In principle
789 * q->bitmaps[dst] |= q->bitmaps[src] & mask;
790 * q->bitmaps[src] &= ~mask;
791 * but we should make sure that src != dst
793 static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
794 int src, int dst)
796 q->bitmaps[dst] |= q->bitmaps[src] & mask;
797 q->bitmaps[src] &= ~mask;
800 static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
802 unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
803 struct qfq_group *next;
805 if (mask) {
806 next = qfq_ffs(q, mask);
807 if (!qfq_gt(next->F, old_F))
808 return;
811 mask = (1UL << index) - 1;
812 qfq_move_groups(q, mask, EB, ER);
813 qfq_move_groups(q, mask, IB, IR);
817 * perhaps
819 old_V ^= q->V;
820 old_V >>= q->min_slot_shift;
821 if (old_V) {
826 static void qfq_make_eligible(struct qfq_sched *q)
828 unsigned long vslot = q->V >> q->min_slot_shift;
829 unsigned long old_vslot = q->oldV >> q->min_slot_shift;
831 if (vslot != old_vslot) {
832 unsigned long mask;
833 int last_flip_pos = fls(vslot ^ old_vslot);
835 if (last_flip_pos > 31) /* higher than the number of groups */
836 mask = ~0UL; /* make all groups eligible */
837 else
838 mask = (1UL << last_flip_pos) - 1;
840 qfq_move_groups(q, mask, IR, ER);
841 qfq_move_groups(q, mask, IB, EB);
846 * The index of the slot in which the input aggregate agg is to be
847 * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2'
848 * and not a '-1' because the start time of the group may be moved
849 * backward by one slot after the aggregate has been inserted, and
850 * this would cause non-empty slots to be right-shifted by one
851 * position.
853 * QFQ+ fully satisfies this bound to the slot index if the parameters
854 * of the classes are not changed dynamically, and if QFQ+ never
855 * happens to postpone the service of agg unjustly, i.e., it never
856 * happens that the aggregate becomes backlogged and eligible, or just
857 * eligible, while an aggregate with a higher approximated finish time
858 * is being served. In particular, in this case QFQ+ guarantees that
859 * the timestamps of agg are low enough that the slot index is never
860 * higher than 2. Unfortunately, QFQ+ cannot provide the same
861 * guarantee if it happens to unjustly postpone the service of agg, or
862 * if the parameters of some class are changed.
864 * As for the first event, i.e., an out-of-order service, the
865 * upper bound to the slot index guaranteed by QFQ+ grows to
866 * 2 +
867 * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
868 * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1.
870 * The following function deals with this problem by backward-shifting
871 * the timestamps of agg, if needed, so as to guarantee that the slot
872 * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may
873 * cause the service of other aggregates to be postponed, yet the
874 * worst-case guarantees of these aggregates are not violated. In
875 * fact, in case of no out-of-order service, the timestamps of agg
876 * would have been even lower than they are after the backward shift,
877 * because QFQ+ would have guaranteed a maximum value equal to 2 for
878 * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose
879 * service is postponed because of the backward-shift would have
880 * however waited for the service of agg before being served.
882 * The other event that may cause the slot index to be higher than 2
883 * for agg is a recent change of the parameters of some class. If the
884 * weight of a class is increased or the lmax (max_pkt_size) of the
885 * class is decreased, then a new aggregate with smaller slot size
886 * than the original parent aggregate of the class may happen to be
887 * activated. The activation of this aggregate should be properly
888 * delayed to when the service of the class has finished in the ideal
889 * system tracked by QFQ+. If the activation of the aggregate is not
890 * delayed to this reference time instant, then this aggregate may be
891 * unjustly served before other aggregates waiting for service. This
892 * may cause the above bound to the slot index to be violated for some
893 * of these unlucky aggregates.
895 * Instead of delaying the activation of the new aggregate, which is
896 * quite complex, the above-discussed capping of the slot index is
897 * used to handle also the consequences of a change of the parameters
898 * of a class.
900 static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg,
901 u64 roundedS)
903 u64 slot = (roundedS - grp->S) >> grp->slot_shift;
904 unsigned int i; /* slot index in the bucket list */
906 if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
907 u64 deltaS = roundedS - grp->S -
908 ((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
909 agg->S -= deltaS;
910 agg->F -= deltaS;
911 slot = QFQ_MAX_SLOTS - 2;
914 i = (grp->front + slot) % QFQ_MAX_SLOTS;
916 hlist_add_head(&agg->next, &grp->slots[i]);
917 __set_bit(slot, &grp->full_slots);
920 /* Maybe introduce hlist_first_entry?? */
921 static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp)
923 return hlist_entry(grp->slots[grp->front].first,
924 struct qfq_aggregate, next);
928 * remove the entry from the slot
930 static void qfq_front_slot_remove(struct qfq_group *grp)
932 struct qfq_aggregate *agg = qfq_slot_head(grp);
934 BUG_ON(!agg);
935 hlist_del(&agg->next);
936 if (hlist_empty(&grp->slots[grp->front]))
937 __clear_bit(0, &grp->full_slots);
941 * Returns the first aggregate in the first non-empty bucket of the
942 * group. As a side effect, adjusts the bucket list so the first
943 * non-empty bucket is at position 0 in full_slots.
945 static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp)
947 unsigned int i;
949 pr_debug("qfq slot_scan: grp %u full %#lx\n",
950 grp->index, grp->full_slots);
952 if (grp->full_slots == 0)
953 return NULL;
955 i = __ffs(grp->full_slots); /* zero based */
956 if (i > 0) {
957 grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
958 grp->full_slots >>= i;
961 return qfq_slot_head(grp);
965 * adjust the bucket list. When the start time of a group decreases,
966 * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
967 * move the objects. The mask of occupied slots must be shifted
968 * because we use ffs() to find the first non-empty slot.
969 * This covers decreases in the group's start time, but what about
970 * increases of the start time ?
971 * Here too we should make sure that i is less than 32
973 static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
975 unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
977 grp->full_slots <<= i;
978 grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
981 static void qfq_update_eligible(struct qfq_sched *q)
983 struct qfq_group *grp;
984 unsigned long ineligible;
986 ineligible = q->bitmaps[IR] | q->bitmaps[IB];
987 if (ineligible) {
988 if (!q->bitmaps[ER]) {
989 grp = qfq_ffs(q, ineligible);
990 if (qfq_gt(grp->S, q->V))
991 q->V = grp->S;
993 qfq_make_eligible(q);
997 /* Dequeue head packet of the head class in the DRR queue of the aggregate. */
998 static void agg_dequeue(struct qfq_aggregate *agg,
999 struct qfq_class *cl, unsigned int len)
1001 qdisc_dequeue_peeked(cl->qdisc);
1003 cl->deficit -= (int) len;
1005 if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */
1006 list_del(&cl->alist);
1007 else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) {
1008 cl->deficit += agg->lmax;
1009 list_move_tail(&cl->alist, &agg->active);
1013 static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg,
1014 struct qfq_class **cl,
1015 unsigned int *len)
1017 struct sk_buff *skb;
1019 *cl = list_first_entry(&agg->active, struct qfq_class, alist);
1020 skb = (*cl)->qdisc->ops->peek((*cl)->qdisc);
1021 if (skb == NULL)
1022 WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
1023 else
1024 *len = qdisc_pkt_len(skb);
1026 return skb;
1029 /* Update F according to the actual service received by the aggregate. */
1030 static inline void charge_actual_service(struct qfq_aggregate *agg)
1032 /* Compute the service received by the aggregate, taking into
1033 * account that, after decreasing the number of classes in
1034 * agg, it may happen that
1035 * agg->initial_budget - agg->budget > agg->bugdetmax
1037 u32 service_received = min(agg->budgetmax,
1038 agg->initial_budget - agg->budget);
1040 agg->F = agg->S + (u64)service_received * agg->inv_w;
1043 /* Assign a reasonable start time for a new aggregate in group i.
1044 * Admissible values for \hat(F) are multiples of \sigma_i
1045 * no greater than V+\sigma_i . Larger values mean that
1046 * we had a wraparound so we consider the timestamp to be stale.
1048 * If F is not stale and F >= V then we set S = F.
1049 * Otherwise we should assign S = V, but this may violate
1050 * the ordering in EB (see [2]). So, if we have groups in ER,
1051 * set S to the F_j of the first group j which would be blocking us.
1052 * We are guaranteed not to move S backward because
1053 * otherwise our group i would still be blocked.
1055 static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg)
1057 unsigned long mask;
1058 u64 limit, roundedF;
1059 int slot_shift = agg->grp->slot_shift;
1061 roundedF = qfq_round_down(agg->F, slot_shift);
1062 limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
1064 if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) {
1065 /* timestamp was stale */
1066 mask = mask_from(q->bitmaps[ER], agg->grp->index);
1067 if (mask) {
1068 struct qfq_group *next = qfq_ffs(q, mask);
1069 if (qfq_gt(roundedF, next->F)) {
1070 if (qfq_gt(limit, next->F))
1071 agg->S = next->F;
1072 else /* preserve timestamp correctness */
1073 agg->S = limit;
1074 return;
1077 agg->S = q->V;
1078 } else /* timestamp is not stale */
1079 agg->S = agg->F;
1082 /* Update the timestamps of agg before scheduling/rescheduling it for
1083 * service. In particular, assign to agg->F its maximum possible
1084 * value, i.e., the virtual finish time with which the aggregate
1085 * should be labeled if it used all its budget once in service.
1087 static inline void
1088 qfq_update_agg_ts(struct qfq_sched *q,
1089 struct qfq_aggregate *agg, enum update_reason reason)
1091 if (reason != requeue)
1092 qfq_update_start(q, agg);
1093 else /* just charge agg for the service received */
1094 agg->S = agg->F;
1096 agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w;
1099 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg);
1101 static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
1103 struct qfq_sched *q = qdisc_priv(sch);
1104 struct qfq_aggregate *in_serv_agg = q->in_serv_agg;
1105 struct qfq_class *cl;
1106 struct sk_buff *skb = NULL;
1107 /* next-packet len, 0 means no more active classes in in-service agg */
1108 unsigned int len = 0;
1110 if (in_serv_agg == NULL)
1111 return NULL;
1113 if (!list_empty(&in_serv_agg->active))
1114 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1117 * If there are no active classes in the in-service aggregate,
1118 * or if the aggregate has not enough budget to serve its next
1119 * class, then choose the next aggregate to serve.
1121 if (len == 0 || in_serv_agg->budget < len) {
1122 charge_actual_service(in_serv_agg);
1124 /* recharge the budget of the aggregate */
1125 in_serv_agg->initial_budget = in_serv_agg->budget =
1126 in_serv_agg->budgetmax;
1128 if (!list_empty(&in_serv_agg->active)) {
1130 * Still active: reschedule for
1131 * service. Possible optimization: if no other
1132 * aggregate is active, then there is no point
1133 * in rescheduling this aggregate, and we can
1134 * just keep it as the in-service one. This
1135 * should be however a corner case, and to
1136 * handle it, we would need to maintain an
1137 * extra num_active_aggs field.
1139 qfq_update_agg_ts(q, in_serv_agg, requeue);
1140 qfq_schedule_agg(q, in_serv_agg);
1141 } else if (sch->q.qlen == 0) { /* no aggregate to serve */
1142 q->in_serv_agg = NULL;
1143 return NULL;
1147 * If we get here, there are other aggregates queued:
1148 * choose the new aggregate to serve.
1150 in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
1151 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1153 if (!skb)
1154 return NULL;
1156 sch->q.qlen--;
1157 qdisc_bstats_update(sch, skb);
1159 agg_dequeue(in_serv_agg, cl, len);
1160 /* If lmax is lowered, through qfq_change_class, for a class
1161 * owning pending packets with larger size than the new value
1162 * of lmax, then the following condition may hold.
1164 if (unlikely(in_serv_agg->budget < len))
1165 in_serv_agg->budget = 0;
1166 else
1167 in_serv_agg->budget -= len;
1169 q->V += (u64)len * q->iwsum;
1170 pr_debug("qfq dequeue: len %u F %lld now %lld\n",
1171 len, (unsigned long long) in_serv_agg->F,
1172 (unsigned long long) q->V);
1174 return skb;
1177 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)
1179 struct qfq_group *grp;
1180 struct qfq_aggregate *agg, *new_front_agg;
1181 u64 old_F;
1183 qfq_update_eligible(q);
1184 q->oldV = q->V;
1186 if (!q->bitmaps[ER])
1187 return NULL;
1189 grp = qfq_ffs(q, q->bitmaps[ER]);
1190 old_F = grp->F;
1192 agg = qfq_slot_head(grp);
1194 /* agg starts to be served, remove it from schedule */
1195 qfq_front_slot_remove(grp);
1197 new_front_agg = qfq_slot_scan(grp);
1199 if (new_front_agg == NULL) /* group is now inactive, remove from ER */
1200 __clear_bit(grp->index, &q->bitmaps[ER]);
1201 else {
1202 u64 roundedS = qfq_round_down(new_front_agg->S,
1203 grp->slot_shift);
1204 unsigned int s;
1206 if (grp->S == roundedS)
1207 return agg;
1208 grp->S = roundedS;
1209 grp->F = roundedS + (2ULL << grp->slot_shift);
1210 __clear_bit(grp->index, &q->bitmaps[ER]);
1211 s = qfq_calc_state(q, grp);
1212 __set_bit(grp->index, &q->bitmaps[s]);
1215 qfq_unblock_groups(q, grp->index, old_F);
1217 return agg;
1220 static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
1222 struct qfq_sched *q = qdisc_priv(sch);
1223 struct qfq_class *cl;
1224 struct qfq_aggregate *agg;
1225 int err = 0;
1227 cl = qfq_classify(skb, sch, &err);
1228 if (cl == NULL) {
1229 if (err & __NET_XMIT_BYPASS)
1230 qdisc_qstats_drop(sch);
1231 kfree_skb(skb);
1232 return err;
1234 pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
1236 if (unlikely(cl->agg->lmax < qdisc_pkt_len(skb))) {
1237 pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
1238 cl->agg->lmax, qdisc_pkt_len(skb), cl->common.classid);
1239 err = qfq_change_agg(sch, cl, cl->agg->class_weight,
1240 qdisc_pkt_len(skb));
1241 if (err)
1242 return err;
1245 err = qdisc_enqueue(skb, cl->qdisc);
1246 if (unlikely(err != NET_XMIT_SUCCESS)) {
1247 pr_debug("qfq_enqueue: enqueue failed %d\n", err);
1248 if (net_xmit_drop_count(err)) {
1249 cl->qstats.drops++;
1250 qdisc_qstats_drop(sch);
1252 return err;
1255 bstats_update(&cl->bstats, skb);
1256 ++sch->q.qlen;
1258 agg = cl->agg;
1259 /* if the queue was not empty, then done here */
1260 if (cl->qdisc->q.qlen != 1) {
1261 if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) &&
1262 list_first_entry(&agg->active, struct qfq_class, alist)
1263 == cl && cl->deficit < qdisc_pkt_len(skb))
1264 list_move_tail(&cl->alist, &agg->active);
1266 return err;
1269 /* schedule class for service within the aggregate */
1270 cl->deficit = agg->lmax;
1271 list_add_tail(&cl->alist, &agg->active);
1273 if (list_first_entry(&agg->active, struct qfq_class, alist) != cl ||
1274 q->in_serv_agg == agg)
1275 return err; /* non-empty or in service, nothing else to do */
1277 qfq_activate_agg(q, agg, enqueue);
1279 return err;
1283 * Schedule aggregate according to its timestamps.
1285 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1287 struct qfq_group *grp = agg->grp;
1288 u64 roundedS;
1289 int s;
1291 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1294 * Insert agg in the correct bucket.
1295 * If agg->S >= grp->S we don't need to adjust the
1296 * bucket list and simply go to the insertion phase.
1297 * Otherwise grp->S is decreasing, we must make room
1298 * in the bucket list, and also recompute the group state.
1299 * Finally, if there were no flows in this group and nobody
1300 * was in ER make sure to adjust V.
1302 if (grp->full_slots) {
1303 if (!qfq_gt(grp->S, agg->S))
1304 goto skip_update;
1306 /* create a slot for this agg->S */
1307 qfq_slot_rotate(grp, roundedS);
1308 /* group was surely ineligible, remove */
1309 __clear_bit(grp->index, &q->bitmaps[IR]);
1310 __clear_bit(grp->index, &q->bitmaps[IB]);
1311 } else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) &&
1312 q->in_serv_agg == NULL)
1313 q->V = roundedS;
1315 grp->S = roundedS;
1316 grp->F = roundedS + (2ULL << grp->slot_shift);
1317 s = qfq_calc_state(q, grp);
1318 __set_bit(grp->index, &q->bitmaps[s]);
1320 pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
1321 s, q->bitmaps[s],
1322 (unsigned long long) agg->S,
1323 (unsigned long long) agg->F,
1324 (unsigned long long) q->V);
1326 skip_update:
1327 qfq_slot_insert(grp, agg, roundedS);
1331 /* Update agg ts and schedule agg for service */
1332 static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
1333 enum update_reason reason)
1335 agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */
1337 qfq_update_agg_ts(q, agg, reason);
1338 if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */
1339 q->in_serv_agg = agg; /* start serving this aggregate */
1340 /* update V: to be in service, agg must be eligible */
1341 q->oldV = q->V = agg->S;
1342 } else if (agg != q->in_serv_agg)
1343 qfq_schedule_agg(q, agg);
1346 static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
1347 struct qfq_aggregate *agg)
1349 unsigned int i, offset;
1350 u64 roundedS;
1352 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1353 offset = (roundedS - grp->S) >> grp->slot_shift;
1355 i = (grp->front + offset) % QFQ_MAX_SLOTS;
1357 hlist_del(&agg->next);
1358 if (hlist_empty(&grp->slots[i]))
1359 __clear_bit(offset, &grp->full_slots);
1363 * Called to forcibly deschedule an aggregate. If the aggregate is
1364 * not in the front bucket, or if the latter has other aggregates in
1365 * the front bucket, we can simply remove the aggregate with no other
1366 * side effects.
1367 * Otherwise we must propagate the event up.
1369 static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1371 struct qfq_group *grp = agg->grp;
1372 unsigned long mask;
1373 u64 roundedS;
1374 int s;
1376 if (agg == q->in_serv_agg) {
1377 charge_actual_service(agg);
1378 q->in_serv_agg = qfq_choose_next_agg(q);
1379 return;
1382 agg->F = agg->S;
1383 qfq_slot_remove(q, grp, agg);
1385 if (!grp->full_slots) {
1386 __clear_bit(grp->index, &q->bitmaps[IR]);
1387 __clear_bit(grp->index, &q->bitmaps[EB]);
1388 __clear_bit(grp->index, &q->bitmaps[IB]);
1390 if (test_bit(grp->index, &q->bitmaps[ER]) &&
1391 !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
1392 mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
1393 if (mask)
1394 mask = ~((1UL << __fls(mask)) - 1);
1395 else
1396 mask = ~0UL;
1397 qfq_move_groups(q, mask, EB, ER);
1398 qfq_move_groups(q, mask, IB, IR);
1400 __clear_bit(grp->index, &q->bitmaps[ER]);
1401 } else if (hlist_empty(&grp->slots[grp->front])) {
1402 agg = qfq_slot_scan(grp);
1403 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1404 if (grp->S != roundedS) {
1405 __clear_bit(grp->index, &q->bitmaps[ER]);
1406 __clear_bit(grp->index, &q->bitmaps[IR]);
1407 __clear_bit(grp->index, &q->bitmaps[EB]);
1408 __clear_bit(grp->index, &q->bitmaps[IB]);
1409 grp->S = roundedS;
1410 grp->F = roundedS + (2ULL << grp->slot_shift);
1411 s = qfq_calc_state(q, grp);
1412 __set_bit(grp->index, &q->bitmaps[s]);
1417 static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1419 struct qfq_sched *q = qdisc_priv(sch);
1420 struct qfq_class *cl = (struct qfq_class *)arg;
1422 if (cl->qdisc->q.qlen == 0)
1423 qfq_deactivate_class(q, cl);
1426 static unsigned int qfq_drop_from_slot(struct qfq_sched *q,
1427 struct hlist_head *slot)
1429 struct qfq_aggregate *agg;
1430 struct qfq_class *cl;
1431 unsigned int len;
1433 hlist_for_each_entry(agg, slot, next) {
1434 list_for_each_entry(cl, &agg->active, alist) {
1436 if (!cl->qdisc->ops->drop)
1437 continue;
1439 len = cl->qdisc->ops->drop(cl->qdisc);
1440 if (len > 0) {
1441 if (cl->qdisc->q.qlen == 0)
1442 qfq_deactivate_class(q, cl);
1444 return len;
1448 return 0;
1451 static unsigned int qfq_drop(struct Qdisc *sch)
1453 struct qfq_sched *q = qdisc_priv(sch);
1454 struct qfq_group *grp;
1455 unsigned int i, j, len;
1457 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1458 grp = &q->groups[i];
1459 for (j = 0; j < QFQ_MAX_SLOTS; j++) {
1460 len = qfq_drop_from_slot(q, &grp->slots[j]);
1461 if (len > 0) {
1462 sch->q.qlen--;
1463 return len;
1469 return 0;
1472 static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
1474 struct qfq_sched *q = qdisc_priv(sch);
1475 struct qfq_group *grp;
1476 int i, j, err;
1477 u32 max_cl_shift, maxbudg_shift, max_classes;
1479 err = qdisc_class_hash_init(&q->clhash);
1480 if (err < 0)
1481 return err;
1483 if (qdisc_dev(sch)->tx_queue_len + 1 > QFQ_MAX_AGG_CLASSES)
1484 max_classes = QFQ_MAX_AGG_CLASSES;
1485 else
1486 max_classes = qdisc_dev(sch)->tx_queue_len + 1;
1487 /* max_cl_shift = floor(log_2(max_classes)) */
1488 max_cl_shift = __fls(max_classes);
1489 q->max_agg_classes = 1<<max_cl_shift;
1491 /* maxbudg_shift = log2(max_len * max_classes_per_agg) */
1492 maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift;
1493 q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX;
1495 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1496 grp = &q->groups[i];
1497 grp->index = i;
1498 grp->slot_shift = q->min_slot_shift + i;
1499 for (j = 0; j < QFQ_MAX_SLOTS; j++)
1500 INIT_HLIST_HEAD(&grp->slots[j]);
1503 INIT_HLIST_HEAD(&q->nonfull_aggs);
1505 return 0;
1508 static void qfq_reset_qdisc(struct Qdisc *sch)
1510 struct qfq_sched *q = qdisc_priv(sch);
1511 struct qfq_class *cl;
1512 unsigned int i;
1514 for (i = 0; i < q->clhash.hashsize; i++) {
1515 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1516 if (cl->qdisc->q.qlen > 0)
1517 qfq_deactivate_class(q, cl);
1519 qdisc_reset(cl->qdisc);
1522 sch->q.qlen = 0;
1525 static void qfq_destroy_qdisc(struct Qdisc *sch)
1527 struct qfq_sched *q = qdisc_priv(sch);
1528 struct qfq_class *cl;
1529 struct hlist_node *next;
1530 unsigned int i;
1532 tcf_destroy_chain(&q->filter_list);
1534 for (i = 0; i < q->clhash.hashsize; i++) {
1535 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1536 common.hnode) {
1537 qfq_destroy_class(sch, cl);
1540 qdisc_class_hash_destroy(&q->clhash);
1543 static const struct Qdisc_class_ops qfq_class_ops = {
1544 .change = qfq_change_class,
1545 .delete = qfq_delete_class,
1546 .get = qfq_get_class,
1547 .put = qfq_put_class,
1548 .tcf_chain = qfq_tcf_chain,
1549 .bind_tcf = qfq_bind_tcf,
1550 .unbind_tcf = qfq_unbind_tcf,
1551 .graft = qfq_graft_class,
1552 .leaf = qfq_class_leaf,
1553 .qlen_notify = qfq_qlen_notify,
1554 .dump = qfq_dump_class,
1555 .dump_stats = qfq_dump_class_stats,
1556 .walk = qfq_walk,
1559 static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1560 .cl_ops = &qfq_class_ops,
1561 .id = "qfq",
1562 .priv_size = sizeof(struct qfq_sched),
1563 .enqueue = qfq_enqueue,
1564 .dequeue = qfq_dequeue,
1565 .peek = qdisc_peek_dequeued,
1566 .drop = qfq_drop,
1567 .init = qfq_init_qdisc,
1568 .reset = qfq_reset_qdisc,
1569 .destroy = qfq_destroy_qdisc,
1570 .owner = THIS_MODULE,
1573 static int __init qfq_init(void)
1575 return register_qdisc(&qfq_qdisc_ops);
1578 static void __exit qfq_exit(void)
1580 unregister_qdisc(&qfq_qdisc_ops);
1583 module_init(qfq_init);
1584 module_exit(qfq_exit);
1585 MODULE_LICENSE("GPL");