Linux 3.8-rc7
[cris-mirror.git] / net / sched / sch_qfq.c
blob6ed37652a4c388df5f96dde4f4e865d86e1868f6
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)
116 #define IWSUM (ONE_FP/QFQ_MAX_WSUM)
118 #define QFQ_MTU_SHIFT 16 /* to support TSO/GSO */
119 #define QFQ_MIN_LMAX 512 /* see qfq_slot_insert */
121 #define QFQ_MAX_AGG_CLASSES 8 /* max num classes per aggregate allowed */
124 * Possible group states. These values are used as indexes for the bitmaps
125 * array of struct qfq_queue.
127 enum qfq_state { ER, IR, EB, IB, QFQ_MAX_STATE };
129 struct qfq_group;
131 struct qfq_aggregate;
133 struct qfq_class {
134 struct Qdisc_class_common common;
136 unsigned int refcnt;
137 unsigned int filter_cnt;
139 struct gnet_stats_basic_packed bstats;
140 struct gnet_stats_queue qstats;
141 struct gnet_stats_rate_est rate_est;
142 struct Qdisc *qdisc;
143 struct list_head alist; /* Link for active-classes list. */
144 struct qfq_aggregate *agg; /* Parent aggregate. */
145 int deficit; /* DRR deficit counter. */
148 struct qfq_aggregate {
149 struct hlist_node next; /* Link for the slot list. */
150 u64 S, F; /* flow timestamps (exact) */
152 /* group we belong to. In principle we would need the index,
153 * which is log_2(lmax/weight), but we never reference it
154 * directly, only the group.
156 struct qfq_group *grp;
158 /* these are copied from the flowset. */
159 u32 class_weight; /* Weight of each class in this aggregate. */
160 /* Max pkt size for the classes in this aggregate, DRR quantum. */
161 int lmax;
163 u32 inv_w; /* ONE_FP/(sum of weights of classes in aggr.). */
164 u32 budgetmax; /* Max budget for this aggregate. */
165 u32 initial_budget, budget; /* Initial and current budget. */
167 int num_classes; /* Number of classes in this aggr. */
168 struct list_head active; /* DRR queue of active classes. */
170 struct hlist_node nonfull_next; /* See nonfull_aggs in qfq_sched. */
173 struct qfq_group {
174 u64 S, F; /* group timestamps (approx). */
175 unsigned int slot_shift; /* Slot shift. */
176 unsigned int index; /* Group index. */
177 unsigned int front; /* Index of the front slot. */
178 unsigned long full_slots; /* non-empty slots */
180 /* Array of RR lists of active aggregates. */
181 struct hlist_head slots[QFQ_MAX_SLOTS];
184 struct qfq_sched {
185 struct tcf_proto *filter_list;
186 struct Qdisc_class_hash clhash;
188 u64 oldV, V; /* Precise virtual times. */
189 struct qfq_aggregate *in_serv_agg; /* Aggregate being served. */
190 u32 num_active_agg; /* Num. of active aggregates */
191 u32 wsum; /* weight sum */
193 unsigned long bitmaps[QFQ_MAX_STATE]; /* Group bitmaps. */
194 struct qfq_group groups[QFQ_MAX_INDEX + 1]; /* The groups. */
195 u32 min_slot_shift; /* Index of the group-0 bit in the bitmaps. */
197 u32 max_agg_classes; /* Max number of classes per aggr. */
198 struct hlist_head nonfull_aggs; /* Aggs with room for more classes. */
202 * Possible reasons why the timestamps of an aggregate are updated
203 * enqueue: the aggregate switches from idle to active and must scheduled
204 * for service
205 * requeue: the aggregate finishes its budget, so it stops being served and
206 * must be rescheduled for service
208 enum update_reason {enqueue, requeue};
210 static struct qfq_class *qfq_find_class(struct Qdisc *sch, u32 classid)
212 struct qfq_sched *q = qdisc_priv(sch);
213 struct Qdisc_class_common *clc;
215 clc = qdisc_class_find(&q->clhash, classid);
216 if (clc == NULL)
217 return NULL;
218 return container_of(clc, struct qfq_class, common);
221 static void qfq_purge_queue(struct qfq_class *cl)
223 unsigned int len = cl->qdisc->q.qlen;
225 qdisc_reset(cl->qdisc);
226 qdisc_tree_decrease_qlen(cl->qdisc, len);
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 },
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.
239 static int qfq_calc_index(u32 inv_w, unsigned int maxlen, u32 min_slot_shift)
241 u64 slot_size = (u64)maxlen * inv_w;
242 unsigned long size_map;
243 int index = 0;
245 size_map = slot_size >> min_slot_shift;
246 if (!size_map)
247 goto out;
249 index = __fls(size_map) + 1; /* basically a log_2 */
250 index -= !(slot_size - (1ULL << (index + min_slot_shift - 1)));
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);
258 return index;
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);
265 static void qfq_init_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
266 u32 lmax, u32 weight)
268 INIT_LIST_HEAD(&agg->active);
269 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
271 agg->lmax = lmax;
272 agg->class_weight = weight;
275 static struct qfq_aggregate *qfq_find_agg(struct qfq_sched *q,
276 u32 lmax, u32 weight)
278 struct qfq_aggregate *agg;
279 struct hlist_node *n;
281 hlist_for_each_entry(agg, n, &q->nonfull_aggs, nonfull_next)
282 if (agg->lmax == lmax && agg->class_weight == weight)
283 return agg;
285 return NULL;
289 /* Update aggregate as a function of the new number of classes. */
290 static void qfq_update_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
291 int new_num_classes)
293 u32 new_agg_weight;
295 if (new_num_classes == q->max_agg_classes)
296 hlist_del_init(&agg->nonfull_next);
298 if (agg->num_classes > new_num_classes &&
299 new_num_classes == q->max_agg_classes - 1) /* agg no more full */
300 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
302 agg->budgetmax = new_num_classes * agg->lmax;
303 new_agg_weight = agg->class_weight * new_num_classes;
304 agg->inv_w = ONE_FP/new_agg_weight;
306 if (agg->grp == NULL) {
307 int i = qfq_calc_index(agg->inv_w, agg->budgetmax,
308 q->min_slot_shift);
309 agg->grp = &q->groups[i];
312 q->wsum +=
313 (int) agg->class_weight * (new_num_classes - agg->num_classes);
315 agg->num_classes = new_num_classes;
318 /* Add class to aggregate. */
319 static void qfq_add_to_agg(struct qfq_sched *q,
320 struct qfq_aggregate *agg,
321 struct qfq_class *cl)
323 cl->agg = agg;
325 qfq_update_agg(q, agg, agg->num_classes+1);
326 if (cl->qdisc->q.qlen > 0) { /* adding an active class */
327 list_add_tail(&cl->alist, &agg->active);
328 if (list_first_entry(&agg->active, struct qfq_class, alist) ==
329 cl && q->in_serv_agg != agg) /* agg was inactive */
330 qfq_activate_agg(q, agg, enqueue); /* schedule agg */
334 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *);
336 static void qfq_destroy_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
338 if (!hlist_unhashed(&agg->nonfull_next))
339 hlist_del_init(&agg->nonfull_next);
340 if (q->in_serv_agg == agg)
341 q->in_serv_agg = qfq_choose_next_agg(q);
342 kfree(agg);
345 /* Deschedule class from within its parent aggregate. */
346 static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
348 struct qfq_aggregate *agg = cl->agg;
351 list_del(&cl->alist); /* remove from RR queue of the aggregate */
352 if (list_empty(&agg->active)) /* agg is now inactive */
353 qfq_deactivate_agg(q, agg);
356 /* Remove class from its parent aggregate. */
357 static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
359 struct qfq_aggregate *agg = cl->agg;
361 cl->agg = NULL;
362 if (agg->num_classes == 1) { /* agg being emptied, destroy it */
363 qfq_destroy_agg(q, agg);
364 return;
366 qfq_update_agg(q, agg, agg->num_classes-1);
369 /* Deschedule class and remove it from its parent aggregate. */
370 static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
372 if (cl->qdisc->q.qlen > 0) /* class is active */
373 qfq_deactivate_class(q, cl);
375 qfq_rm_from_agg(q, cl);
378 /* Move class to a new aggregate, matching the new class weight and/or lmax */
379 static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight,
380 u32 lmax)
382 struct qfq_sched *q = qdisc_priv(sch);
383 struct qfq_aggregate *new_agg = qfq_find_agg(q, lmax, weight);
385 if (new_agg == NULL) { /* create new aggregate */
386 new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC);
387 if (new_agg == NULL)
388 return -ENOBUFS;
389 qfq_init_agg(q, new_agg, lmax, weight);
391 qfq_deact_rm_from_agg(q, cl);
392 qfq_add_to_agg(q, new_agg, cl);
394 return 0;
397 static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
398 struct nlattr **tca, unsigned long *arg)
400 struct qfq_sched *q = qdisc_priv(sch);
401 struct qfq_class *cl = (struct qfq_class *)*arg;
402 bool existing = false;
403 struct nlattr *tb[TCA_QFQ_MAX + 1];
404 struct qfq_aggregate *new_agg = NULL;
405 u32 weight, lmax, inv_w;
406 int err;
407 int delta_w;
409 if (tca[TCA_OPTIONS] == NULL) {
410 pr_notice("qfq: no options\n");
411 return -EINVAL;
414 err = nla_parse_nested(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS], qfq_policy);
415 if (err < 0)
416 return err;
418 if (tb[TCA_QFQ_WEIGHT]) {
419 weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]);
420 if (!weight || weight > (1UL << QFQ_MAX_WSHIFT)) {
421 pr_notice("qfq: invalid weight %u\n", weight);
422 return -EINVAL;
424 } else
425 weight = 1;
427 if (tb[TCA_QFQ_LMAX]) {
428 lmax = nla_get_u32(tb[TCA_QFQ_LMAX]);
429 if (lmax < QFQ_MIN_LMAX || lmax > (1UL << QFQ_MTU_SHIFT)) {
430 pr_notice("qfq: invalid max length %u\n", lmax);
431 return -EINVAL;
433 } else
434 lmax = psched_mtu(qdisc_dev(sch));
436 inv_w = ONE_FP / weight;
437 weight = ONE_FP / inv_w;
439 if (cl != NULL &&
440 lmax == cl->agg->lmax &&
441 weight == cl->agg->class_weight)
442 return 0; /* nothing to change */
444 delta_w = weight - (cl ? cl->agg->class_weight : 0);
446 if (q->wsum + delta_w > QFQ_MAX_WSUM) {
447 pr_notice("qfq: total weight out of range (%d + %u)\n",
448 delta_w, q->wsum);
449 return -EINVAL;
452 if (cl != NULL) { /* modify existing class */
453 if (tca[TCA_RATE]) {
454 err = gen_replace_estimator(&cl->bstats, &cl->rate_est,
455 qdisc_root_sleeping_lock(sch),
456 tca[TCA_RATE]);
457 if (err)
458 return err;
460 existing = true;
461 goto set_change_agg;
464 /* create and init new class */
465 cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
466 if (cl == NULL)
467 return -ENOBUFS;
469 cl->refcnt = 1;
470 cl->common.classid = classid;
471 cl->deficit = lmax;
473 cl->qdisc = qdisc_create_dflt(sch->dev_queue,
474 &pfifo_qdisc_ops, classid);
475 if (cl->qdisc == NULL)
476 cl->qdisc = &noop_qdisc;
478 if (tca[TCA_RATE]) {
479 err = gen_new_estimator(&cl->bstats, &cl->rate_est,
480 qdisc_root_sleeping_lock(sch),
481 tca[TCA_RATE]);
482 if (err)
483 goto destroy_class;
486 sch_tree_lock(sch);
487 qdisc_class_hash_insert(&q->clhash, &cl->common);
488 sch_tree_unlock(sch);
490 qdisc_class_hash_grow(sch, &q->clhash);
492 set_change_agg:
493 sch_tree_lock(sch);
494 new_agg = qfq_find_agg(q, lmax, weight);
495 if (new_agg == NULL) { /* create new aggregate */
496 sch_tree_unlock(sch);
497 new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL);
498 if (new_agg == NULL) {
499 err = -ENOBUFS;
500 gen_kill_estimator(&cl->bstats, &cl->rate_est);
501 goto destroy_class;
503 sch_tree_lock(sch);
504 qfq_init_agg(q, new_agg, lmax, weight);
506 if (existing)
507 qfq_deact_rm_from_agg(q, cl);
508 qfq_add_to_agg(q, new_agg, cl);
509 sch_tree_unlock(sch);
511 *arg = (unsigned long)cl;
512 return 0;
514 destroy_class:
515 qdisc_destroy(cl->qdisc);
516 kfree(cl);
517 return err;
520 static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
522 struct qfq_sched *q = qdisc_priv(sch);
524 qfq_rm_from_agg(q, cl);
525 gen_kill_estimator(&cl->bstats, &cl->rate_est);
526 qdisc_destroy(cl->qdisc);
527 kfree(cl);
530 static int qfq_delete_class(struct Qdisc *sch, unsigned long arg)
532 struct qfq_sched *q = qdisc_priv(sch);
533 struct qfq_class *cl = (struct qfq_class *)arg;
535 if (cl->filter_cnt > 0)
536 return -EBUSY;
538 sch_tree_lock(sch);
540 qfq_purge_queue(cl);
541 qdisc_class_hash_remove(&q->clhash, &cl->common);
543 BUG_ON(--cl->refcnt == 0);
545 * This shouldn't happen: we "hold" one cops->get() when called
546 * from tc_ctl_tclass; the destroy method is done from cops->put().
549 sch_tree_unlock(sch);
550 return 0;
553 static unsigned long qfq_get_class(struct Qdisc *sch, u32 classid)
555 struct qfq_class *cl = qfq_find_class(sch, classid);
557 if (cl != NULL)
558 cl->refcnt++;
560 return (unsigned long)cl;
563 static void qfq_put_class(struct Qdisc *sch, unsigned long arg)
565 struct qfq_class *cl = (struct qfq_class *)arg;
567 if (--cl->refcnt == 0)
568 qfq_destroy_class(sch, cl);
571 static struct tcf_proto **qfq_tcf_chain(struct Qdisc *sch, unsigned long cl)
573 struct qfq_sched *q = qdisc_priv(sch);
575 if (cl)
576 return NULL;
578 return &q->filter_list;
581 static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
582 u32 classid)
584 struct qfq_class *cl = qfq_find_class(sch, classid);
586 if (cl != NULL)
587 cl->filter_cnt++;
589 return (unsigned long)cl;
592 static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
594 struct qfq_class *cl = (struct qfq_class *)arg;
596 cl->filter_cnt--;
599 static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
600 struct Qdisc *new, struct Qdisc **old)
602 struct qfq_class *cl = (struct qfq_class *)arg;
604 if (new == NULL) {
605 new = qdisc_create_dflt(sch->dev_queue,
606 &pfifo_qdisc_ops, cl->common.classid);
607 if (new == NULL)
608 new = &noop_qdisc;
611 sch_tree_lock(sch);
612 qfq_purge_queue(cl);
613 *old = cl->qdisc;
614 cl->qdisc = new;
615 sch_tree_unlock(sch);
616 return 0;
619 static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
621 struct qfq_class *cl = (struct qfq_class *)arg;
623 return cl->qdisc;
626 static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
627 struct sk_buff *skb, struct tcmsg *tcm)
629 struct qfq_class *cl = (struct qfq_class *)arg;
630 struct nlattr *nest;
632 tcm->tcm_parent = TC_H_ROOT;
633 tcm->tcm_handle = cl->common.classid;
634 tcm->tcm_info = cl->qdisc->handle;
636 nest = nla_nest_start(skb, TCA_OPTIONS);
637 if (nest == NULL)
638 goto nla_put_failure;
639 if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) ||
640 nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax))
641 goto nla_put_failure;
642 return nla_nest_end(skb, nest);
644 nla_put_failure:
645 nla_nest_cancel(skb, nest);
646 return -EMSGSIZE;
649 static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
650 struct gnet_dump *d)
652 struct qfq_class *cl = (struct qfq_class *)arg;
653 struct tc_qfq_stats xstats;
655 memset(&xstats, 0, sizeof(xstats));
656 cl->qdisc->qstats.qlen = cl->qdisc->q.qlen;
658 xstats.weight = cl->agg->class_weight;
659 xstats.lmax = cl->agg->lmax;
661 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
662 gnet_stats_copy_rate_est(d, &cl->bstats, &cl->rate_est) < 0 ||
663 gnet_stats_copy_queue(d, &cl->qdisc->qstats) < 0)
664 return -1;
666 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
669 static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
671 struct qfq_sched *q = qdisc_priv(sch);
672 struct qfq_class *cl;
673 struct hlist_node *n;
674 unsigned int i;
676 if (arg->stop)
677 return;
679 for (i = 0; i < q->clhash.hashsize; i++) {
680 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode) {
681 if (arg->count < arg->skip) {
682 arg->count++;
683 continue;
685 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
686 arg->stop = 1;
687 return;
689 arg->count++;
694 static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
695 int *qerr)
697 struct qfq_sched *q = qdisc_priv(sch);
698 struct qfq_class *cl;
699 struct tcf_result res;
700 int result;
702 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
703 pr_debug("qfq_classify: found %d\n", skb->priority);
704 cl = qfq_find_class(sch, skb->priority);
705 if (cl != NULL)
706 return cl;
709 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
710 result = tc_classify(skb, q->filter_list, &res);
711 if (result >= 0) {
712 #ifdef CONFIG_NET_CLS_ACT
713 switch (result) {
714 case TC_ACT_QUEUED:
715 case TC_ACT_STOLEN:
716 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
717 case TC_ACT_SHOT:
718 return NULL;
720 #endif
721 cl = (struct qfq_class *)res.class;
722 if (cl == NULL)
723 cl = qfq_find_class(sch, res.classid);
724 return cl;
727 return NULL;
730 /* Generic comparison function, handling wraparound. */
731 static inline int qfq_gt(u64 a, u64 b)
733 return (s64)(a - b) > 0;
736 /* Round a precise timestamp to its slotted value. */
737 static inline u64 qfq_round_down(u64 ts, unsigned int shift)
739 return ts & ~((1ULL << shift) - 1);
742 /* return the pointer to the group with lowest index in the bitmap */
743 static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
744 unsigned long bitmap)
746 int index = __ffs(bitmap);
747 return &q->groups[index];
749 /* Calculate a mask to mimic what would be ffs_from(). */
750 static inline unsigned long mask_from(unsigned long bitmap, int from)
752 return bitmap & ~((1UL << from) - 1);
756 * The state computation relies on ER=0, IR=1, EB=2, IB=3
757 * First compute eligibility comparing grp->S, q->V,
758 * then check if someone is blocking us and possibly add EB
760 static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
762 /* if S > V we are not eligible */
763 unsigned int state = qfq_gt(grp->S, q->V);
764 unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
765 struct qfq_group *next;
767 if (mask) {
768 next = qfq_ffs(q, mask);
769 if (qfq_gt(grp->F, next->F))
770 state |= EB;
773 return state;
778 * In principle
779 * q->bitmaps[dst] |= q->bitmaps[src] & mask;
780 * q->bitmaps[src] &= ~mask;
781 * but we should make sure that src != dst
783 static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
784 int src, int dst)
786 q->bitmaps[dst] |= q->bitmaps[src] & mask;
787 q->bitmaps[src] &= ~mask;
790 static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
792 unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
793 struct qfq_group *next;
795 if (mask) {
796 next = qfq_ffs(q, mask);
797 if (!qfq_gt(next->F, old_F))
798 return;
801 mask = (1UL << index) - 1;
802 qfq_move_groups(q, mask, EB, ER);
803 qfq_move_groups(q, mask, IB, IR);
807 * perhaps
809 old_V ^= q->V;
810 old_V >>= q->min_slot_shift;
811 if (old_V) {
816 static void qfq_make_eligible(struct qfq_sched *q)
818 unsigned long vslot = q->V >> q->min_slot_shift;
819 unsigned long old_vslot = q->oldV >> q->min_slot_shift;
821 if (vslot != old_vslot) {
822 unsigned long mask = (1UL << fls(vslot ^ old_vslot)) - 1;
823 qfq_move_groups(q, mask, IR, ER);
824 qfq_move_groups(q, mask, IB, EB);
830 * The index of the slot in which the aggregate is to be inserted must
831 * not be higher than QFQ_MAX_SLOTS-2. There is a '-2' and not a '-1'
832 * because the start time of the group may be moved backward by one
833 * slot after the aggregate has been inserted, and this would cause
834 * non-empty slots to be right-shifted by one position.
836 * If the weight and lmax (max_pkt_size) of the classes do not change,
837 * then QFQ+ does meet the above contraint according to the current
838 * values of its parameters. In fact, if the weight and lmax of the
839 * classes do not change, then, from the theory, QFQ+ guarantees that
840 * the slot index is never higher than
841 * 2 + QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
842 * (QFQ_MAX_WEIGHT/QFQ_MAX_WSUM) = 2 + 8 * 128 * (1 / 64) = 18
844 * When the weight of a class is increased or the lmax of the class is
845 * decreased, a new aggregate with smaller slot size than the original
846 * parent aggregate of the class may happen to be activated. The
847 * activation of this aggregate should be properly delayed to when the
848 * service of the class has finished in the ideal system tracked by
849 * QFQ+. If the activation of the aggregate is not delayed to this
850 * reference time instant, then this aggregate may be unjustly served
851 * before other aggregates waiting for service. This may cause the
852 * above bound to the slot index to be violated for some of these
853 * unlucky aggregates.
855 * Instead of delaying the activation of the new aggregate, which is
856 * quite complex, the following inaccurate but simple solution is used:
857 * if the slot index is higher than QFQ_MAX_SLOTS-2, then the
858 * timestamps of the aggregate are shifted backward so as to let the
859 * slot index become equal to QFQ_MAX_SLOTS-2.
861 static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg,
862 u64 roundedS)
864 u64 slot = (roundedS - grp->S) >> grp->slot_shift;
865 unsigned int i; /* slot index in the bucket list */
867 if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
868 u64 deltaS = roundedS - grp->S -
869 ((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
870 agg->S -= deltaS;
871 agg->F -= deltaS;
872 slot = QFQ_MAX_SLOTS - 2;
875 i = (grp->front + slot) % QFQ_MAX_SLOTS;
877 hlist_add_head(&agg->next, &grp->slots[i]);
878 __set_bit(slot, &grp->full_slots);
881 /* Maybe introduce hlist_first_entry?? */
882 static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp)
884 return hlist_entry(grp->slots[grp->front].first,
885 struct qfq_aggregate, next);
889 * remove the entry from the slot
891 static void qfq_front_slot_remove(struct qfq_group *grp)
893 struct qfq_aggregate *agg = qfq_slot_head(grp);
895 BUG_ON(!agg);
896 hlist_del(&agg->next);
897 if (hlist_empty(&grp->slots[grp->front]))
898 __clear_bit(0, &grp->full_slots);
902 * Returns the first aggregate in the first non-empty bucket of the
903 * group. As a side effect, adjusts the bucket list so the first
904 * non-empty bucket is at position 0 in full_slots.
906 static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp)
908 unsigned int i;
910 pr_debug("qfq slot_scan: grp %u full %#lx\n",
911 grp->index, grp->full_slots);
913 if (grp->full_slots == 0)
914 return NULL;
916 i = __ffs(grp->full_slots); /* zero based */
917 if (i > 0) {
918 grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
919 grp->full_slots >>= i;
922 return qfq_slot_head(grp);
926 * adjust the bucket list. When the start time of a group decreases,
927 * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
928 * move the objects. The mask of occupied slots must be shifted
929 * because we use ffs() to find the first non-empty slot.
930 * This covers decreases in the group's start time, but what about
931 * increases of the start time ?
932 * Here too we should make sure that i is less than 32
934 static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
936 unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
938 grp->full_slots <<= i;
939 grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
942 static void qfq_update_eligible(struct qfq_sched *q)
944 struct qfq_group *grp;
945 unsigned long ineligible;
947 ineligible = q->bitmaps[IR] | q->bitmaps[IB];
948 if (ineligible) {
949 if (!q->bitmaps[ER]) {
950 grp = qfq_ffs(q, ineligible);
951 if (qfq_gt(grp->S, q->V))
952 q->V = grp->S;
954 qfq_make_eligible(q);
958 /* Dequeue head packet of the head class in the DRR queue of the aggregate. */
959 static void agg_dequeue(struct qfq_aggregate *agg,
960 struct qfq_class *cl, unsigned int len)
962 qdisc_dequeue_peeked(cl->qdisc);
964 cl->deficit -= (int) len;
966 if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */
967 list_del(&cl->alist);
968 else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) {
969 cl->deficit += agg->lmax;
970 list_move_tail(&cl->alist, &agg->active);
974 static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg,
975 struct qfq_class **cl,
976 unsigned int *len)
978 struct sk_buff *skb;
980 *cl = list_first_entry(&agg->active, struct qfq_class, alist);
981 skb = (*cl)->qdisc->ops->peek((*cl)->qdisc);
982 if (skb == NULL)
983 WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
984 else
985 *len = qdisc_pkt_len(skb);
987 return skb;
990 /* Update F according to the actual service received by the aggregate. */
991 static inline void charge_actual_service(struct qfq_aggregate *agg)
993 /* compute the service received by the aggregate */
994 u32 service_received = agg->initial_budget - agg->budget;
996 agg->F = agg->S + (u64)service_received * agg->inv_w;
999 static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
1001 struct qfq_sched *q = qdisc_priv(sch);
1002 struct qfq_aggregate *in_serv_agg = q->in_serv_agg;
1003 struct qfq_class *cl;
1004 struct sk_buff *skb = NULL;
1005 /* next-packet len, 0 means no more active classes in in-service agg */
1006 unsigned int len = 0;
1008 if (in_serv_agg == NULL)
1009 return NULL;
1011 if (!list_empty(&in_serv_agg->active))
1012 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1015 * If there are no active classes in the in-service aggregate,
1016 * or if the aggregate has not enough budget to serve its next
1017 * class, then choose the next aggregate to serve.
1019 if (len == 0 || in_serv_agg->budget < len) {
1020 charge_actual_service(in_serv_agg);
1022 /* recharge the budget of the aggregate */
1023 in_serv_agg->initial_budget = in_serv_agg->budget =
1024 in_serv_agg->budgetmax;
1026 if (!list_empty(&in_serv_agg->active))
1028 * Still active: reschedule for
1029 * service. Possible optimization: if no other
1030 * aggregate is active, then there is no point
1031 * in rescheduling this aggregate, and we can
1032 * just keep it as the in-service one. This
1033 * should be however a corner case, and to
1034 * handle it, we would need to maintain an
1035 * extra num_active_aggs field.
1037 qfq_activate_agg(q, in_serv_agg, requeue);
1038 else if (sch->q.qlen == 0) { /* no aggregate to serve */
1039 q->in_serv_agg = NULL;
1040 return NULL;
1044 * If we get here, there are other aggregates queued:
1045 * choose the new aggregate to serve.
1047 in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
1048 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1050 if (!skb)
1051 return NULL;
1053 sch->q.qlen--;
1054 qdisc_bstats_update(sch, skb);
1056 agg_dequeue(in_serv_agg, cl, len);
1057 in_serv_agg->budget -= len;
1058 q->V += (u64)len * IWSUM;
1059 pr_debug("qfq dequeue: len %u F %lld now %lld\n",
1060 len, (unsigned long long) in_serv_agg->F,
1061 (unsigned long long) q->V);
1063 return skb;
1066 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)
1068 struct qfq_group *grp;
1069 struct qfq_aggregate *agg, *new_front_agg;
1070 u64 old_F;
1072 qfq_update_eligible(q);
1073 q->oldV = q->V;
1075 if (!q->bitmaps[ER])
1076 return NULL;
1078 grp = qfq_ffs(q, q->bitmaps[ER]);
1079 old_F = grp->F;
1081 agg = qfq_slot_head(grp);
1083 /* agg starts to be served, remove it from schedule */
1084 qfq_front_slot_remove(grp);
1086 new_front_agg = qfq_slot_scan(grp);
1088 if (new_front_agg == NULL) /* group is now inactive, remove from ER */
1089 __clear_bit(grp->index, &q->bitmaps[ER]);
1090 else {
1091 u64 roundedS = qfq_round_down(new_front_agg->S,
1092 grp->slot_shift);
1093 unsigned int s;
1095 if (grp->S == roundedS)
1096 return agg;
1097 grp->S = roundedS;
1098 grp->F = roundedS + (2ULL << grp->slot_shift);
1099 __clear_bit(grp->index, &q->bitmaps[ER]);
1100 s = qfq_calc_state(q, grp);
1101 __set_bit(grp->index, &q->bitmaps[s]);
1104 qfq_unblock_groups(q, grp->index, old_F);
1106 return agg;
1110 * Assign a reasonable start time for a new aggregate in group i.
1111 * Admissible values for \hat(F) are multiples of \sigma_i
1112 * no greater than V+\sigma_i . Larger values mean that
1113 * we had a wraparound so we consider the timestamp to be stale.
1115 * If F is not stale and F >= V then we set S = F.
1116 * Otherwise we should assign S = V, but this may violate
1117 * the ordering in EB (see [2]). So, if we have groups in ER,
1118 * set S to the F_j of the first group j which would be blocking us.
1119 * We are guaranteed not to move S backward because
1120 * otherwise our group i would still be blocked.
1122 static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg)
1124 unsigned long mask;
1125 u64 limit, roundedF;
1126 int slot_shift = agg->grp->slot_shift;
1128 roundedF = qfq_round_down(agg->F, slot_shift);
1129 limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
1131 if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) {
1132 /* timestamp was stale */
1133 mask = mask_from(q->bitmaps[ER], agg->grp->index);
1134 if (mask) {
1135 struct qfq_group *next = qfq_ffs(q, mask);
1136 if (qfq_gt(roundedF, next->F)) {
1137 if (qfq_gt(limit, next->F))
1138 agg->S = next->F;
1139 else /* preserve timestamp correctness */
1140 agg->S = limit;
1141 return;
1144 agg->S = q->V;
1145 } else /* timestamp is not stale */
1146 agg->S = agg->F;
1150 * Update the timestamps of agg before scheduling/rescheduling it for
1151 * service. In particular, assign to agg->F its maximum possible
1152 * value, i.e., the virtual finish time with which the aggregate
1153 * should be labeled if it used all its budget once in service.
1155 static inline void
1156 qfq_update_agg_ts(struct qfq_sched *q,
1157 struct qfq_aggregate *agg, enum update_reason reason)
1159 if (reason != requeue)
1160 qfq_update_start(q, agg);
1161 else /* just charge agg for the service received */
1162 agg->S = agg->F;
1164 agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w;
1167 static void qfq_schedule_agg(struct qfq_sched *, struct qfq_aggregate *);
1169 static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
1171 struct qfq_sched *q = qdisc_priv(sch);
1172 struct qfq_class *cl;
1173 struct qfq_aggregate *agg;
1174 int err = 0;
1176 cl = qfq_classify(skb, sch, &err);
1177 if (cl == NULL) {
1178 if (err & __NET_XMIT_BYPASS)
1179 sch->qstats.drops++;
1180 kfree_skb(skb);
1181 return err;
1183 pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
1185 if (unlikely(cl->agg->lmax < qdisc_pkt_len(skb))) {
1186 pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
1187 cl->agg->lmax, qdisc_pkt_len(skb), cl->common.classid);
1188 err = qfq_change_agg(sch, cl, cl->agg->class_weight,
1189 qdisc_pkt_len(skb));
1190 if (err)
1191 return err;
1194 err = qdisc_enqueue(skb, cl->qdisc);
1195 if (unlikely(err != NET_XMIT_SUCCESS)) {
1196 pr_debug("qfq_enqueue: enqueue failed %d\n", err);
1197 if (net_xmit_drop_count(err)) {
1198 cl->qstats.drops++;
1199 sch->qstats.drops++;
1201 return err;
1204 bstats_update(&cl->bstats, skb);
1205 ++sch->q.qlen;
1207 agg = cl->agg;
1208 /* if the queue was not empty, then done here */
1209 if (cl->qdisc->q.qlen != 1) {
1210 if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) &&
1211 list_first_entry(&agg->active, struct qfq_class, alist)
1212 == cl && cl->deficit < qdisc_pkt_len(skb))
1213 list_move_tail(&cl->alist, &agg->active);
1215 return err;
1218 /* schedule class for service within the aggregate */
1219 cl->deficit = agg->lmax;
1220 list_add_tail(&cl->alist, &agg->active);
1222 if (list_first_entry(&agg->active, struct qfq_class, alist) != cl)
1223 return err; /* aggregate was not empty, nothing else to do */
1225 /* recharge budget */
1226 agg->initial_budget = agg->budget = agg->budgetmax;
1228 qfq_update_agg_ts(q, agg, enqueue);
1229 if (q->in_serv_agg == NULL)
1230 q->in_serv_agg = agg;
1231 else if (agg != q->in_serv_agg)
1232 qfq_schedule_agg(q, agg);
1234 return err;
1238 * Schedule aggregate according to its timestamps.
1240 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1242 struct qfq_group *grp = agg->grp;
1243 u64 roundedS;
1244 int s;
1246 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1249 * Insert agg in the correct bucket.
1250 * If agg->S >= grp->S we don't need to adjust the
1251 * bucket list and simply go to the insertion phase.
1252 * Otherwise grp->S is decreasing, we must make room
1253 * in the bucket list, and also recompute the group state.
1254 * Finally, if there were no flows in this group and nobody
1255 * was in ER make sure to adjust V.
1257 if (grp->full_slots) {
1258 if (!qfq_gt(grp->S, agg->S))
1259 goto skip_update;
1261 /* create a slot for this agg->S */
1262 qfq_slot_rotate(grp, roundedS);
1263 /* group was surely ineligible, remove */
1264 __clear_bit(grp->index, &q->bitmaps[IR]);
1265 __clear_bit(grp->index, &q->bitmaps[IB]);
1266 } else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V))
1267 q->V = roundedS;
1269 grp->S = roundedS;
1270 grp->F = roundedS + (2ULL << grp->slot_shift);
1271 s = qfq_calc_state(q, grp);
1272 __set_bit(grp->index, &q->bitmaps[s]);
1274 pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
1275 s, q->bitmaps[s],
1276 (unsigned long long) agg->S,
1277 (unsigned long long) agg->F,
1278 (unsigned long long) q->V);
1280 skip_update:
1281 qfq_slot_insert(grp, agg, roundedS);
1285 /* Update agg ts and schedule agg for service */
1286 static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
1287 enum update_reason reason)
1289 qfq_update_agg_ts(q, agg, reason);
1290 qfq_schedule_agg(q, agg);
1293 static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
1294 struct qfq_aggregate *agg)
1296 unsigned int i, offset;
1297 u64 roundedS;
1299 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1300 offset = (roundedS - grp->S) >> grp->slot_shift;
1302 i = (grp->front + offset) % QFQ_MAX_SLOTS;
1304 hlist_del(&agg->next);
1305 if (hlist_empty(&grp->slots[i]))
1306 __clear_bit(offset, &grp->full_slots);
1310 * Called to forcibly deschedule an aggregate. If the aggregate is
1311 * not in the front bucket, or if the latter has other aggregates in
1312 * the front bucket, we can simply remove the aggregate with no other
1313 * side effects.
1314 * Otherwise we must propagate the event up.
1316 static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1318 struct qfq_group *grp = agg->grp;
1319 unsigned long mask;
1320 u64 roundedS;
1321 int s;
1323 if (agg == q->in_serv_agg) {
1324 charge_actual_service(agg);
1325 q->in_serv_agg = qfq_choose_next_agg(q);
1326 return;
1329 agg->F = agg->S;
1330 qfq_slot_remove(q, grp, agg);
1332 if (!grp->full_slots) {
1333 __clear_bit(grp->index, &q->bitmaps[IR]);
1334 __clear_bit(grp->index, &q->bitmaps[EB]);
1335 __clear_bit(grp->index, &q->bitmaps[IB]);
1337 if (test_bit(grp->index, &q->bitmaps[ER]) &&
1338 !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
1339 mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
1340 if (mask)
1341 mask = ~((1UL << __fls(mask)) - 1);
1342 else
1343 mask = ~0UL;
1344 qfq_move_groups(q, mask, EB, ER);
1345 qfq_move_groups(q, mask, IB, IR);
1347 __clear_bit(grp->index, &q->bitmaps[ER]);
1348 } else if (hlist_empty(&grp->slots[grp->front])) {
1349 agg = qfq_slot_scan(grp);
1350 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1351 if (grp->S != roundedS) {
1352 __clear_bit(grp->index, &q->bitmaps[ER]);
1353 __clear_bit(grp->index, &q->bitmaps[IR]);
1354 __clear_bit(grp->index, &q->bitmaps[EB]);
1355 __clear_bit(grp->index, &q->bitmaps[IB]);
1356 grp->S = roundedS;
1357 grp->F = roundedS + (2ULL << grp->slot_shift);
1358 s = qfq_calc_state(q, grp);
1359 __set_bit(grp->index, &q->bitmaps[s]);
1363 qfq_update_eligible(q);
1366 static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1368 struct qfq_sched *q = qdisc_priv(sch);
1369 struct qfq_class *cl = (struct qfq_class *)arg;
1371 if (cl->qdisc->q.qlen == 0)
1372 qfq_deactivate_class(q, cl);
1375 static unsigned int qfq_drop_from_slot(struct qfq_sched *q,
1376 struct hlist_head *slot)
1378 struct qfq_aggregate *agg;
1379 struct hlist_node *n;
1380 struct qfq_class *cl;
1381 unsigned int len;
1383 hlist_for_each_entry(agg, n, slot, next) {
1384 list_for_each_entry(cl, &agg->active, alist) {
1386 if (!cl->qdisc->ops->drop)
1387 continue;
1389 len = cl->qdisc->ops->drop(cl->qdisc);
1390 if (len > 0) {
1391 if (cl->qdisc->q.qlen == 0)
1392 qfq_deactivate_class(q, cl);
1394 return len;
1398 return 0;
1401 static unsigned int qfq_drop(struct Qdisc *sch)
1403 struct qfq_sched *q = qdisc_priv(sch);
1404 struct qfq_group *grp;
1405 unsigned int i, j, len;
1407 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1408 grp = &q->groups[i];
1409 for (j = 0; j < QFQ_MAX_SLOTS; j++) {
1410 len = qfq_drop_from_slot(q, &grp->slots[j]);
1411 if (len > 0) {
1412 sch->q.qlen--;
1413 return len;
1419 return 0;
1422 static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
1424 struct qfq_sched *q = qdisc_priv(sch);
1425 struct qfq_group *grp;
1426 int i, j, err;
1427 u32 max_cl_shift, maxbudg_shift, max_classes;
1429 err = qdisc_class_hash_init(&q->clhash);
1430 if (err < 0)
1431 return err;
1433 if (qdisc_dev(sch)->tx_queue_len + 1 > QFQ_MAX_AGG_CLASSES)
1434 max_classes = QFQ_MAX_AGG_CLASSES;
1435 else
1436 max_classes = qdisc_dev(sch)->tx_queue_len + 1;
1437 /* max_cl_shift = floor(log_2(max_classes)) */
1438 max_cl_shift = __fls(max_classes);
1439 q->max_agg_classes = 1<<max_cl_shift;
1441 /* maxbudg_shift = log2(max_len * max_classes_per_agg) */
1442 maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift;
1443 q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX;
1445 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1446 grp = &q->groups[i];
1447 grp->index = i;
1448 grp->slot_shift = q->min_slot_shift + i;
1449 for (j = 0; j < QFQ_MAX_SLOTS; j++)
1450 INIT_HLIST_HEAD(&grp->slots[j]);
1453 INIT_HLIST_HEAD(&q->nonfull_aggs);
1455 return 0;
1458 static void qfq_reset_qdisc(struct Qdisc *sch)
1460 struct qfq_sched *q = qdisc_priv(sch);
1461 struct qfq_class *cl;
1462 struct hlist_node *n;
1463 unsigned int i;
1465 for (i = 0; i < q->clhash.hashsize; i++) {
1466 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode) {
1467 if (cl->qdisc->q.qlen > 0)
1468 qfq_deactivate_class(q, cl);
1470 qdisc_reset(cl->qdisc);
1473 sch->q.qlen = 0;
1476 static void qfq_destroy_qdisc(struct Qdisc *sch)
1478 struct qfq_sched *q = qdisc_priv(sch);
1479 struct qfq_class *cl;
1480 struct hlist_node *n, *next;
1481 unsigned int i;
1483 tcf_destroy_chain(&q->filter_list);
1485 for (i = 0; i < q->clhash.hashsize; i++) {
1486 hlist_for_each_entry_safe(cl, n, next, &q->clhash.hash[i],
1487 common.hnode) {
1488 qfq_destroy_class(sch, cl);
1491 qdisc_class_hash_destroy(&q->clhash);
1494 static const struct Qdisc_class_ops qfq_class_ops = {
1495 .change = qfq_change_class,
1496 .delete = qfq_delete_class,
1497 .get = qfq_get_class,
1498 .put = qfq_put_class,
1499 .tcf_chain = qfq_tcf_chain,
1500 .bind_tcf = qfq_bind_tcf,
1501 .unbind_tcf = qfq_unbind_tcf,
1502 .graft = qfq_graft_class,
1503 .leaf = qfq_class_leaf,
1504 .qlen_notify = qfq_qlen_notify,
1505 .dump = qfq_dump_class,
1506 .dump_stats = qfq_dump_class_stats,
1507 .walk = qfq_walk,
1510 static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1511 .cl_ops = &qfq_class_ops,
1512 .id = "qfq",
1513 .priv_size = sizeof(struct qfq_sched),
1514 .enqueue = qfq_enqueue,
1515 .dequeue = qfq_dequeue,
1516 .peek = qdisc_peek_dequeued,
1517 .drop = qfq_drop,
1518 .init = qfq_init_qdisc,
1519 .reset = qfq_reset_qdisc,
1520 .destroy = qfq_destroy_qdisc,
1521 .owner = THIS_MODULE,
1524 static int __init qfq_init(void)
1526 return register_qdisc(&qfq_qdisc_ops);
1529 static void __exit qfq_exit(void)
1531 unregister_qdisc(&qfq_qdisc_ops);
1534 module_init(qfq_init);
1535 module_exit(qfq_exit);
1536 MODULE_LICENSE("GPL");