Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net
[linux-2.6/libata-dev.git] / net / sched / sch_qfq.c
blobd51852bba01c981c9f9834dad82cfbcfec904508
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;
280 hlist_for_each_entry(agg, &q->nonfull_aggs, nonfull_next)
281 if (agg->lmax == lmax && agg->class_weight == weight)
282 return agg;
284 return NULL;
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)
292 u32 new_agg_weight;
294 if (new_num_classes == q->max_agg_classes)
295 hlist_del_init(&agg->nonfull_next);
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);
301 /* The next assignment may let
302 * agg->initial_budget > agg->budgetmax
303 * hold, we will take it into account in charge_actual_service().
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;
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];
315 q->wsum +=
316 (int) agg->class_weight * (new_num_classes - agg->num_classes);
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 if (!hlist_unhashed(&agg->nonfull_next))
342 hlist_del_init(&agg->nonfull_next);
343 if (q->in_serv_agg == agg)
344 q->in_serv_agg = qfq_choose_next_agg(q);
345 kfree(agg);
348 /* Deschedule class from within its parent aggregate. */
349 static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
351 struct qfq_aggregate *agg = cl->agg;
354 list_del(&cl->alist); /* remove from RR queue of the aggregate */
355 if (list_empty(&agg->active)) /* agg is now inactive */
356 qfq_deactivate_agg(q, agg);
359 /* Remove class from its parent aggregate. */
360 static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
362 struct qfq_aggregate *agg = cl->agg;
364 cl->agg = NULL;
365 if (agg->num_classes == 1) { /* agg being emptied, destroy it */
366 qfq_destroy_agg(q, agg);
367 return;
369 qfq_update_agg(q, agg, agg->num_classes-1);
372 /* Deschedule class and remove it from its parent aggregate. */
373 static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
375 if (cl->qdisc->q.qlen > 0) /* class is active */
376 qfq_deactivate_class(q, cl);
378 qfq_rm_from_agg(q, cl);
381 /* Move class to a new aggregate, matching the new class weight and/or lmax */
382 static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight,
383 u32 lmax)
385 struct qfq_sched *q = qdisc_priv(sch);
386 struct qfq_aggregate *new_agg = qfq_find_agg(q, lmax, weight);
388 if (new_agg == NULL) { /* create new aggregate */
389 new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC);
390 if (new_agg == NULL)
391 return -ENOBUFS;
392 qfq_init_agg(q, new_agg, lmax, weight);
394 qfq_deact_rm_from_agg(q, cl);
395 qfq_add_to_agg(q, new_agg, cl);
397 return 0;
400 static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
401 struct nlattr **tca, unsigned long *arg)
403 struct qfq_sched *q = qdisc_priv(sch);
404 struct qfq_class *cl = (struct qfq_class *)*arg;
405 bool existing = false;
406 struct nlattr *tb[TCA_QFQ_MAX + 1];
407 struct qfq_aggregate *new_agg = NULL;
408 u32 weight, lmax, inv_w;
409 int err;
410 int delta_w;
412 if (tca[TCA_OPTIONS] == NULL) {
413 pr_notice("qfq: no options\n");
414 return -EINVAL;
417 err = nla_parse_nested(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS], qfq_policy);
418 if (err < 0)
419 return err;
421 if (tb[TCA_QFQ_WEIGHT]) {
422 weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]);
423 if (!weight || weight > (1UL << QFQ_MAX_WSHIFT)) {
424 pr_notice("qfq: invalid weight %u\n", weight);
425 return -EINVAL;
427 } else
428 weight = 1;
430 if (tb[TCA_QFQ_LMAX]) {
431 lmax = nla_get_u32(tb[TCA_QFQ_LMAX]);
432 if (lmax < QFQ_MIN_LMAX || lmax > (1UL << QFQ_MTU_SHIFT)) {
433 pr_notice("qfq: invalid max length %u\n", lmax);
434 return -EINVAL;
436 } else
437 lmax = psched_mtu(qdisc_dev(sch));
439 inv_w = ONE_FP / weight;
440 weight = ONE_FP / inv_w;
442 if (cl != NULL &&
443 lmax == cl->agg->lmax &&
444 weight == cl->agg->class_weight)
445 return 0; /* nothing to change */
447 delta_w = weight - (cl ? cl->agg->class_weight : 0);
449 if (q->wsum + delta_w > QFQ_MAX_WSUM) {
450 pr_notice("qfq: total weight out of range (%d + %u)\n",
451 delta_w, q->wsum);
452 return -EINVAL;
455 if (cl != NULL) { /* modify existing class */
456 if (tca[TCA_RATE]) {
457 err = gen_replace_estimator(&cl->bstats, &cl->rate_est,
458 qdisc_root_sleeping_lock(sch),
459 tca[TCA_RATE]);
460 if (err)
461 return err;
463 existing = true;
464 goto set_change_agg;
467 /* create and init new class */
468 cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
469 if (cl == NULL)
470 return -ENOBUFS;
472 cl->refcnt = 1;
473 cl->common.classid = classid;
474 cl->deficit = lmax;
476 cl->qdisc = qdisc_create_dflt(sch->dev_queue,
477 &pfifo_qdisc_ops, classid);
478 if (cl->qdisc == NULL)
479 cl->qdisc = &noop_qdisc;
481 if (tca[TCA_RATE]) {
482 err = gen_new_estimator(&cl->bstats, &cl->rate_est,
483 qdisc_root_sleeping_lock(sch),
484 tca[TCA_RATE]);
485 if (err)
486 goto destroy_class;
489 sch_tree_lock(sch);
490 qdisc_class_hash_insert(&q->clhash, &cl->common);
491 sch_tree_unlock(sch);
493 qdisc_class_hash_grow(sch, &q->clhash);
495 set_change_agg:
496 sch_tree_lock(sch);
497 new_agg = qfq_find_agg(q, lmax, weight);
498 if (new_agg == NULL) { /* create new aggregate */
499 sch_tree_unlock(sch);
500 new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL);
501 if (new_agg == NULL) {
502 err = -ENOBUFS;
503 gen_kill_estimator(&cl->bstats, &cl->rate_est);
504 goto destroy_class;
506 sch_tree_lock(sch);
507 qfq_init_agg(q, new_agg, lmax, weight);
509 if (existing)
510 qfq_deact_rm_from_agg(q, cl);
511 qfq_add_to_agg(q, new_agg, cl);
512 sch_tree_unlock(sch);
514 *arg = (unsigned long)cl;
515 return 0;
517 destroy_class:
518 qdisc_destroy(cl->qdisc);
519 kfree(cl);
520 return err;
523 static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
525 struct qfq_sched *q = qdisc_priv(sch);
527 qfq_rm_from_agg(q, cl);
528 gen_kill_estimator(&cl->bstats, &cl->rate_est);
529 qdisc_destroy(cl->qdisc);
530 kfree(cl);
533 static int qfq_delete_class(struct Qdisc *sch, unsigned long arg)
535 struct qfq_sched *q = qdisc_priv(sch);
536 struct qfq_class *cl = (struct qfq_class *)arg;
538 if (cl->filter_cnt > 0)
539 return -EBUSY;
541 sch_tree_lock(sch);
543 qfq_purge_queue(cl);
544 qdisc_class_hash_remove(&q->clhash, &cl->common);
546 BUG_ON(--cl->refcnt == 0);
548 * This shouldn't happen: we "hold" one cops->get() when called
549 * from tc_ctl_tclass; the destroy method is done from cops->put().
552 sch_tree_unlock(sch);
553 return 0;
556 static unsigned long qfq_get_class(struct Qdisc *sch, u32 classid)
558 struct qfq_class *cl = qfq_find_class(sch, classid);
560 if (cl != NULL)
561 cl->refcnt++;
563 return (unsigned long)cl;
566 static void qfq_put_class(struct Qdisc *sch, unsigned long arg)
568 struct qfq_class *cl = (struct qfq_class *)arg;
570 if (--cl->refcnt == 0)
571 qfq_destroy_class(sch, cl);
574 static struct tcf_proto **qfq_tcf_chain(struct Qdisc *sch, unsigned long cl)
576 struct qfq_sched *q = qdisc_priv(sch);
578 if (cl)
579 return NULL;
581 return &q->filter_list;
584 static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
585 u32 classid)
587 struct qfq_class *cl = qfq_find_class(sch, classid);
589 if (cl != NULL)
590 cl->filter_cnt++;
592 return (unsigned long)cl;
595 static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
597 struct qfq_class *cl = (struct qfq_class *)arg;
599 cl->filter_cnt--;
602 static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
603 struct Qdisc *new, struct Qdisc **old)
605 struct qfq_class *cl = (struct qfq_class *)arg;
607 if (new == NULL) {
608 new = qdisc_create_dflt(sch->dev_queue,
609 &pfifo_qdisc_ops, cl->common.classid);
610 if (new == NULL)
611 new = &noop_qdisc;
614 sch_tree_lock(sch);
615 qfq_purge_queue(cl);
616 *old = cl->qdisc;
617 cl->qdisc = new;
618 sch_tree_unlock(sch);
619 return 0;
622 static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
624 struct qfq_class *cl = (struct qfq_class *)arg;
626 return cl->qdisc;
629 static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
630 struct sk_buff *skb, struct tcmsg *tcm)
632 struct qfq_class *cl = (struct qfq_class *)arg;
633 struct nlattr *nest;
635 tcm->tcm_parent = TC_H_ROOT;
636 tcm->tcm_handle = cl->common.classid;
637 tcm->tcm_info = cl->qdisc->handle;
639 nest = nla_nest_start(skb, TCA_OPTIONS);
640 if (nest == NULL)
641 goto nla_put_failure;
642 if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) ||
643 nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax))
644 goto nla_put_failure;
645 return nla_nest_end(skb, nest);
647 nla_put_failure:
648 nla_nest_cancel(skb, nest);
649 return -EMSGSIZE;
652 static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
653 struct gnet_dump *d)
655 struct qfq_class *cl = (struct qfq_class *)arg;
656 struct tc_qfq_stats xstats;
658 memset(&xstats, 0, sizeof(xstats));
659 cl->qdisc->qstats.qlen = cl->qdisc->q.qlen;
661 xstats.weight = cl->agg->class_weight;
662 xstats.lmax = cl->agg->lmax;
664 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
665 gnet_stats_copy_rate_est(d, &cl->bstats, &cl->rate_est) < 0 ||
666 gnet_stats_copy_queue(d, &cl->qdisc->qstats) < 0)
667 return -1;
669 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
672 static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
674 struct qfq_sched *q = qdisc_priv(sch);
675 struct qfq_class *cl;
676 unsigned int i;
678 if (arg->stop)
679 return;
681 for (i = 0; i < q->clhash.hashsize; i++) {
682 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
683 if (arg->count < arg->skip) {
684 arg->count++;
685 continue;
687 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
688 arg->stop = 1;
689 return;
691 arg->count++;
696 static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
697 int *qerr)
699 struct qfq_sched *q = qdisc_priv(sch);
700 struct qfq_class *cl;
701 struct tcf_result res;
702 int result;
704 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
705 pr_debug("qfq_classify: found %d\n", skb->priority);
706 cl = qfq_find_class(sch, skb->priority);
707 if (cl != NULL)
708 return cl;
711 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
712 result = tc_classify(skb, q->filter_list, &res);
713 if (result >= 0) {
714 #ifdef CONFIG_NET_CLS_ACT
715 switch (result) {
716 case TC_ACT_QUEUED:
717 case TC_ACT_STOLEN:
718 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
719 case TC_ACT_SHOT:
720 return NULL;
722 #endif
723 cl = (struct qfq_class *)res.class;
724 if (cl == NULL)
725 cl = qfq_find_class(sch, res.classid);
726 return cl;
729 return NULL;
732 /* Generic comparison function, handling wraparound. */
733 static inline int qfq_gt(u64 a, u64 b)
735 return (s64)(a - b) > 0;
738 /* Round a precise timestamp to its slotted value. */
739 static inline u64 qfq_round_down(u64 ts, unsigned int shift)
741 return ts & ~((1ULL << shift) - 1);
744 /* return the pointer to the group with lowest index in the bitmap */
745 static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
746 unsigned long bitmap)
748 int index = __ffs(bitmap);
749 return &q->groups[index];
751 /* Calculate a mask to mimic what would be ffs_from(). */
752 static inline unsigned long mask_from(unsigned long bitmap, int from)
754 return bitmap & ~((1UL << from) - 1);
758 * The state computation relies on ER=0, IR=1, EB=2, IB=3
759 * First compute eligibility comparing grp->S, q->V,
760 * then check if someone is blocking us and possibly add EB
762 static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
764 /* if S > V we are not eligible */
765 unsigned int state = qfq_gt(grp->S, q->V);
766 unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
767 struct qfq_group *next;
769 if (mask) {
770 next = qfq_ffs(q, mask);
771 if (qfq_gt(grp->F, next->F))
772 state |= EB;
775 return state;
780 * In principle
781 * q->bitmaps[dst] |= q->bitmaps[src] & mask;
782 * q->bitmaps[src] &= ~mask;
783 * but we should make sure that src != dst
785 static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
786 int src, int dst)
788 q->bitmaps[dst] |= q->bitmaps[src] & mask;
789 q->bitmaps[src] &= ~mask;
792 static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
794 unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
795 struct qfq_group *next;
797 if (mask) {
798 next = qfq_ffs(q, mask);
799 if (!qfq_gt(next->F, old_F))
800 return;
803 mask = (1UL << index) - 1;
804 qfq_move_groups(q, mask, EB, ER);
805 qfq_move_groups(q, mask, IB, IR);
809 * perhaps
811 old_V ^= q->V;
812 old_V >>= q->min_slot_shift;
813 if (old_V) {
818 static void qfq_make_eligible(struct qfq_sched *q)
820 unsigned long vslot = q->V >> q->min_slot_shift;
821 unsigned long old_vslot = q->oldV >> q->min_slot_shift;
823 if (vslot != old_vslot) {
824 unsigned long mask = (1ULL << fls(vslot ^ old_vslot)) - 1;
825 qfq_move_groups(q, mask, IR, ER);
826 qfq_move_groups(q, mask, IB, EB);
832 * The index of the slot in which the aggregate is to be inserted must
833 * not be higher than QFQ_MAX_SLOTS-2. There is a '-2' and not a '-1'
834 * because the start time of the group may be moved backward by one
835 * slot after the aggregate has been inserted, and this would cause
836 * non-empty slots to be right-shifted by one position.
838 * If the weight and lmax (max_pkt_size) of the classes do not change,
839 * then QFQ+ does meet the above contraint according to the current
840 * values of its parameters. In fact, if the weight and lmax of the
841 * classes do not change, then, from the theory, QFQ+ guarantees that
842 * the slot index is never higher than
843 * 2 + QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
844 * (QFQ_MAX_WEIGHT/QFQ_MAX_WSUM) = 2 + 8 * 128 * (1 / 64) = 18
846 * When the weight of a class is increased or the lmax of the class is
847 * decreased, a new aggregate with smaller slot size than the original
848 * parent aggregate of the class may happen to be activated. The
849 * activation of this aggregate should be properly delayed to when the
850 * service of the class has finished in the ideal system tracked by
851 * QFQ+. If the activation of the aggregate is not delayed to this
852 * reference time instant, then this aggregate may be unjustly served
853 * before other aggregates waiting for service. This may cause the
854 * above bound to the slot index to be violated for some of these
855 * unlucky aggregates.
857 * Instead of delaying the activation of the new aggregate, which is
858 * quite complex, the following inaccurate but simple solution is used:
859 * if the slot index is higher than QFQ_MAX_SLOTS-2, then the
860 * timestamps of the aggregate are shifted backward so as to let the
861 * slot index become equal to QFQ_MAX_SLOTS-2.
863 static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg,
864 u64 roundedS)
866 u64 slot = (roundedS - grp->S) >> grp->slot_shift;
867 unsigned int i; /* slot index in the bucket list */
869 if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
870 u64 deltaS = roundedS - grp->S -
871 ((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
872 agg->S -= deltaS;
873 agg->F -= deltaS;
874 slot = QFQ_MAX_SLOTS - 2;
877 i = (grp->front + slot) % QFQ_MAX_SLOTS;
879 hlist_add_head(&agg->next, &grp->slots[i]);
880 __set_bit(slot, &grp->full_slots);
883 /* Maybe introduce hlist_first_entry?? */
884 static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp)
886 return hlist_entry(grp->slots[grp->front].first,
887 struct qfq_aggregate, next);
891 * remove the entry from the slot
893 static void qfq_front_slot_remove(struct qfq_group *grp)
895 struct qfq_aggregate *agg = qfq_slot_head(grp);
897 BUG_ON(!agg);
898 hlist_del(&agg->next);
899 if (hlist_empty(&grp->slots[grp->front]))
900 __clear_bit(0, &grp->full_slots);
904 * Returns the first aggregate in the first non-empty bucket of the
905 * group. As a side effect, adjusts the bucket list so the first
906 * non-empty bucket is at position 0 in full_slots.
908 static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp)
910 unsigned int i;
912 pr_debug("qfq slot_scan: grp %u full %#lx\n",
913 grp->index, grp->full_slots);
915 if (grp->full_slots == 0)
916 return NULL;
918 i = __ffs(grp->full_slots); /* zero based */
919 if (i > 0) {
920 grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
921 grp->full_slots >>= i;
924 return qfq_slot_head(grp);
928 * adjust the bucket list. When the start time of a group decreases,
929 * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
930 * move the objects. The mask of occupied slots must be shifted
931 * because we use ffs() to find the first non-empty slot.
932 * This covers decreases in the group's start time, but what about
933 * increases of the start time ?
934 * Here too we should make sure that i is less than 32
936 static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
938 unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
940 grp->full_slots <<= i;
941 grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
944 static void qfq_update_eligible(struct qfq_sched *q)
946 struct qfq_group *grp;
947 unsigned long ineligible;
949 ineligible = q->bitmaps[IR] | q->bitmaps[IB];
950 if (ineligible) {
951 if (!q->bitmaps[ER]) {
952 grp = qfq_ffs(q, ineligible);
953 if (qfq_gt(grp->S, q->V))
954 q->V = grp->S;
956 qfq_make_eligible(q);
960 /* Dequeue head packet of the head class in the DRR queue of the aggregate. */
961 static void agg_dequeue(struct qfq_aggregate *agg,
962 struct qfq_class *cl, unsigned int len)
964 qdisc_dequeue_peeked(cl->qdisc);
966 cl->deficit -= (int) len;
968 if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */
969 list_del(&cl->alist);
970 else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) {
971 cl->deficit += agg->lmax;
972 list_move_tail(&cl->alist, &agg->active);
976 static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg,
977 struct qfq_class **cl,
978 unsigned int *len)
980 struct sk_buff *skb;
982 *cl = list_first_entry(&agg->active, struct qfq_class, alist);
983 skb = (*cl)->qdisc->ops->peek((*cl)->qdisc);
984 if (skb == NULL)
985 WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
986 else
987 *len = qdisc_pkt_len(skb);
989 return skb;
992 /* Update F according to the actual service received by the aggregate. */
993 static inline void charge_actual_service(struct qfq_aggregate *agg)
995 /* Compute the service received by the aggregate, taking into
996 * account that, after decreasing the number of classes in
997 * agg, it may happen that
998 * agg->initial_budget - agg->budget > agg->bugdetmax
1000 u32 service_received = min(agg->budgetmax,
1001 agg->initial_budget - agg->budget);
1003 agg->F = agg->S + (u64)service_received * agg->inv_w;
1006 static inline void qfq_update_agg_ts(struct qfq_sched *q,
1007 struct qfq_aggregate *agg,
1008 enum update_reason reason);
1010 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg);
1012 static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
1014 struct qfq_sched *q = qdisc_priv(sch);
1015 struct qfq_aggregate *in_serv_agg = q->in_serv_agg;
1016 struct qfq_class *cl;
1017 struct sk_buff *skb = NULL;
1018 /* next-packet len, 0 means no more active classes in in-service agg */
1019 unsigned int len = 0;
1021 if (in_serv_agg == NULL)
1022 return NULL;
1024 if (!list_empty(&in_serv_agg->active))
1025 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1028 * If there are no active classes in the in-service aggregate,
1029 * or if the aggregate has not enough budget to serve its next
1030 * class, then choose the next aggregate to serve.
1032 if (len == 0 || in_serv_agg->budget < len) {
1033 charge_actual_service(in_serv_agg);
1035 /* recharge the budget of the aggregate */
1036 in_serv_agg->initial_budget = in_serv_agg->budget =
1037 in_serv_agg->budgetmax;
1039 if (!list_empty(&in_serv_agg->active)) {
1041 * Still active: reschedule for
1042 * service. Possible optimization: if no other
1043 * aggregate is active, then there is no point
1044 * in rescheduling this aggregate, and we can
1045 * just keep it as the in-service one. This
1046 * should be however a corner case, and to
1047 * handle it, we would need to maintain an
1048 * extra num_active_aggs field.
1050 qfq_update_agg_ts(q, in_serv_agg, requeue);
1051 qfq_schedule_agg(q, in_serv_agg);
1052 } else if (sch->q.qlen == 0) { /* no aggregate to serve */
1053 q->in_serv_agg = NULL;
1054 return NULL;
1058 * If we get here, there are other aggregates queued:
1059 * choose the new aggregate to serve.
1061 in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
1062 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1064 if (!skb)
1065 return NULL;
1067 sch->q.qlen--;
1068 qdisc_bstats_update(sch, skb);
1070 agg_dequeue(in_serv_agg, cl, len);
1071 /* If lmax is lowered, through qfq_change_class, for a class
1072 * owning pending packets with larger size than the new value
1073 * of lmax, then the following condition may hold.
1075 if (unlikely(in_serv_agg->budget < len))
1076 in_serv_agg->budget = 0;
1077 else
1078 in_serv_agg->budget -= len;
1080 q->V += (u64)len * IWSUM;
1081 pr_debug("qfq dequeue: len %u F %lld now %lld\n",
1082 len, (unsigned long long) in_serv_agg->F,
1083 (unsigned long long) q->V);
1085 return skb;
1088 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)
1090 struct qfq_group *grp;
1091 struct qfq_aggregate *agg, *new_front_agg;
1092 u64 old_F;
1094 qfq_update_eligible(q);
1095 q->oldV = q->V;
1097 if (!q->bitmaps[ER])
1098 return NULL;
1100 grp = qfq_ffs(q, q->bitmaps[ER]);
1101 old_F = grp->F;
1103 agg = qfq_slot_head(grp);
1105 /* agg starts to be served, remove it from schedule */
1106 qfq_front_slot_remove(grp);
1108 new_front_agg = qfq_slot_scan(grp);
1110 if (new_front_agg == NULL) /* group is now inactive, remove from ER */
1111 __clear_bit(grp->index, &q->bitmaps[ER]);
1112 else {
1113 u64 roundedS = qfq_round_down(new_front_agg->S,
1114 grp->slot_shift);
1115 unsigned int s;
1117 if (grp->S == roundedS)
1118 return agg;
1119 grp->S = roundedS;
1120 grp->F = roundedS + (2ULL << grp->slot_shift);
1121 __clear_bit(grp->index, &q->bitmaps[ER]);
1122 s = qfq_calc_state(q, grp);
1123 __set_bit(grp->index, &q->bitmaps[s]);
1126 qfq_unblock_groups(q, grp->index, old_F);
1128 return agg;
1132 * Assign a reasonable start time for a new aggregate in group i.
1133 * Admissible values for \hat(F) are multiples of \sigma_i
1134 * no greater than V+\sigma_i . Larger values mean that
1135 * we had a wraparound so we consider the timestamp to be stale.
1137 * If F is not stale and F >= V then we set S = F.
1138 * Otherwise we should assign S = V, but this may violate
1139 * the ordering in EB (see [2]). So, if we have groups in ER,
1140 * set S to the F_j of the first group j which would be blocking us.
1141 * We are guaranteed not to move S backward because
1142 * otherwise our group i would still be blocked.
1144 static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg)
1146 unsigned long mask;
1147 u64 limit, roundedF;
1148 int slot_shift = agg->grp->slot_shift;
1150 roundedF = qfq_round_down(agg->F, slot_shift);
1151 limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
1153 if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) {
1154 /* timestamp was stale */
1155 mask = mask_from(q->bitmaps[ER], agg->grp->index);
1156 if (mask) {
1157 struct qfq_group *next = qfq_ffs(q, mask);
1158 if (qfq_gt(roundedF, next->F)) {
1159 if (qfq_gt(limit, next->F))
1160 agg->S = next->F;
1161 else /* preserve timestamp correctness */
1162 agg->S = limit;
1163 return;
1166 agg->S = q->V;
1167 } else /* timestamp is not stale */
1168 agg->S = agg->F;
1172 * Update the timestamps of agg before scheduling/rescheduling it for
1173 * service. In particular, assign to agg->F its maximum possible
1174 * value, i.e., the virtual finish time with which the aggregate
1175 * should be labeled if it used all its budget once in service.
1177 static inline void
1178 qfq_update_agg_ts(struct qfq_sched *q,
1179 struct qfq_aggregate *agg, enum update_reason reason)
1181 if (reason != requeue)
1182 qfq_update_start(q, agg);
1183 else /* just charge agg for the service received */
1184 agg->S = agg->F;
1186 agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w;
1189 static void qfq_schedule_agg(struct qfq_sched *, struct qfq_aggregate *);
1191 static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
1193 struct qfq_sched *q = qdisc_priv(sch);
1194 struct qfq_class *cl;
1195 struct qfq_aggregate *agg;
1196 int err = 0;
1198 cl = qfq_classify(skb, sch, &err);
1199 if (cl == NULL) {
1200 if (err & __NET_XMIT_BYPASS)
1201 sch->qstats.drops++;
1202 kfree_skb(skb);
1203 return err;
1205 pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
1207 if (unlikely(cl->agg->lmax < qdisc_pkt_len(skb))) {
1208 pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
1209 cl->agg->lmax, qdisc_pkt_len(skb), cl->common.classid);
1210 err = qfq_change_agg(sch, cl, cl->agg->class_weight,
1211 qdisc_pkt_len(skb));
1212 if (err)
1213 return err;
1216 err = qdisc_enqueue(skb, cl->qdisc);
1217 if (unlikely(err != NET_XMIT_SUCCESS)) {
1218 pr_debug("qfq_enqueue: enqueue failed %d\n", err);
1219 if (net_xmit_drop_count(err)) {
1220 cl->qstats.drops++;
1221 sch->qstats.drops++;
1223 return err;
1226 bstats_update(&cl->bstats, skb);
1227 ++sch->q.qlen;
1229 agg = cl->agg;
1230 /* if the queue was not empty, then done here */
1231 if (cl->qdisc->q.qlen != 1) {
1232 if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) &&
1233 list_first_entry(&agg->active, struct qfq_class, alist)
1234 == cl && cl->deficit < qdisc_pkt_len(skb))
1235 list_move_tail(&cl->alist, &agg->active);
1237 return err;
1240 /* schedule class for service within the aggregate */
1241 cl->deficit = agg->lmax;
1242 list_add_tail(&cl->alist, &agg->active);
1244 if (list_first_entry(&agg->active, struct qfq_class, alist) != cl ||
1245 q->in_serv_agg == agg)
1246 return err; /* non-empty or in service, nothing else to do */
1248 qfq_activate_agg(q, agg, enqueue);
1250 return err;
1254 * Schedule aggregate according to its timestamps.
1256 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1258 struct qfq_group *grp = agg->grp;
1259 u64 roundedS;
1260 int s;
1262 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1265 * Insert agg in the correct bucket.
1266 * If agg->S >= grp->S we don't need to adjust the
1267 * bucket list and simply go to the insertion phase.
1268 * Otherwise grp->S is decreasing, we must make room
1269 * in the bucket list, and also recompute the group state.
1270 * Finally, if there were no flows in this group and nobody
1271 * was in ER make sure to adjust V.
1273 if (grp->full_slots) {
1274 if (!qfq_gt(grp->S, agg->S))
1275 goto skip_update;
1277 /* create a slot for this agg->S */
1278 qfq_slot_rotate(grp, roundedS);
1279 /* group was surely ineligible, remove */
1280 __clear_bit(grp->index, &q->bitmaps[IR]);
1281 __clear_bit(grp->index, &q->bitmaps[IB]);
1282 } else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) &&
1283 q->in_serv_agg == NULL)
1284 q->V = roundedS;
1286 grp->S = roundedS;
1287 grp->F = roundedS + (2ULL << grp->slot_shift);
1288 s = qfq_calc_state(q, grp);
1289 __set_bit(grp->index, &q->bitmaps[s]);
1291 pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
1292 s, q->bitmaps[s],
1293 (unsigned long long) agg->S,
1294 (unsigned long long) agg->F,
1295 (unsigned long long) q->V);
1297 skip_update:
1298 qfq_slot_insert(grp, agg, roundedS);
1302 /* Update agg ts and schedule agg for service */
1303 static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
1304 enum update_reason reason)
1306 agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */
1308 qfq_update_agg_ts(q, agg, reason);
1309 if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */
1310 q->in_serv_agg = agg; /* start serving this aggregate */
1311 /* update V: to be in service, agg must be eligible */
1312 q->oldV = q->V = agg->S;
1313 } else if (agg != q->in_serv_agg)
1314 qfq_schedule_agg(q, agg);
1317 static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
1318 struct qfq_aggregate *agg)
1320 unsigned int i, offset;
1321 u64 roundedS;
1323 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1324 offset = (roundedS - grp->S) >> grp->slot_shift;
1326 i = (grp->front + offset) % QFQ_MAX_SLOTS;
1328 hlist_del(&agg->next);
1329 if (hlist_empty(&grp->slots[i]))
1330 __clear_bit(offset, &grp->full_slots);
1334 * Called to forcibly deschedule an aggregate. If the aggregate is
1335 * not in the front bucket, or if the latter has other aggregates in
1336 * the front bucket, we can simply remove the aggregate with no other
1337 * side effects.
1338 * Otherwise we must propagate the event up.
1340 static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1342 struct qfq_group *grp = agg->grp;
1343 unsigned long mask;
1344 u64 roundedS;
1345 int s;
1347 if (agg == q->in_serv_agg) {
1348 charge_actual_service(agg);
1349 q->in_serv_agg = qfq_choose_next_agg(q);
1350 return;
1353 agg->F = agg->S;
1354 qfq_slot_remove(q, grp, agg);
1356 if (!grp->full_slots) {
1357 __clear_bit(grp->index, &q->bitmaps[IR]);
1358 __clear_bit(grp->index, &q->bitmaps[EB]);
1359 __clear_bit(grp->index, &q->bitmaps[IB]);
1361 if (test_bit(grp->index, &q->bitmaps[ER]) &&
1362 !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
1363 mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
1364 if (mask)
1365 mask = ~((1UL << __fls(mask)) - 1);
1366 else
1367 mask = ~0UL;
1368 qfq_move_groups(q, mask, EB, ER);
1369 qfq_move_groups(q, mask, IB, IR);
1371 __clear_bit(grp->index, &q->bitmaps[ER]);
1372 } else if (hlist_empty(&grp->slots[grp->front])) {
1373 agg = qfq_slot_scan(grp);
1374 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1375 if (grp->S != roundedS) {
1376 __clear_bit(grp->index, &q->bitmaps[ER]);
1377 __clear_bit(grp->index, &q->bitmaps[IR]);
1378 __clear_bit(grp->index, &q->bitmaps[EB]);
1379 __clear_bit(grp->index, &q->bitmaps[IB]);
1380 grp->S = roundedS;
1381 grp->F = roundedS + (2ULL << grp->slot_shift);
1382 s = qfq_calc_state(q, grp);
1383 __set_bit(grp->index, &q->bitmaps[s]);
1388 static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1390 struct qfq_sched *q = qdisc_priv(sch);
1391 struct qfq_class *cl = (struct qfq_class *)arg;
1393 if (cl->qdisc->q.qlen == 0)
1394 qfq_deactivate_class(q, cl);
1397 static unsigned int qfq_drop_from_slot(struct qfq_sched *q,
1398 struct hlist_head *slot)
1400 struct qfq_aggregate *agg;
1401 struct qfq_class *cl;
1402 unsigned int len;
1404 hlist_for_each_entry(agg, slot, next) {
1405 list_for_each_entry(cl, &agg->active, alist) {
1407 if (!cl->qdisc->ops->drop)
1408 continue;
1410 len = cl->qdisc->ops->drop(cl->qdisc);
1411 if (len > 0) {
1412 if (cl->qdisc->q.qlen == 0)
1413 qfq_deactivate_class(q, cl);
1415 return len;
1419 return 0;
1422 static unsigned int qfq_drop(struct Qdisc *sch)
1424 struct qfq_sched *q = qdisc_priv(sch);
1425 struct qfq_group *grp;
1426 unsigned int i, j, len;
1428 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1429 grp = &q->groups[i];
1430 for (j = 0; j < QFQ_MAX_SLOTS; j++) {
1431 len = qfq_drop_from_slot(q, &grp->slots[j]);
1432 if (len > 0) {
1433 sch->q.qlen--;
1434 return len;
1440 return 0;
1443 static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
1445 struct qfq_sched *q = qdisc_priv(sch);
1446 struct qfq_group *grp;
1447 int i, j, err;
1448 u32 max_cl_shift, maxbudg_shift, max_classes;
1450 err = qdisc_class_hash_init(&q->clhash);
1451 if (err < 0)
1452 return err;
1454 if (qdisc_dev(sch)->tx_queue_len + 1 > QFQ_MAX_AGG_CLASSES)
1455 max_classes = QFQ_MAX_AGG_CLASSES;
1456 else
1457 max_classes = qdisc_dev(sch)->tx_queue_len + 1;
1458 /* max_cl_shift = floor(log_2(max_classes)) */
1459 max_cl_shift = __fls(max_classes);
1460 q->max_agg_classes = 1<<max_cl_shift;
1462 /* maxbudg_shift = log2(max_len * max_classes_per_agg) */
1463 maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift;
1464 q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX;
1466 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1467 grp = &q->groups[i];
1468 grp->index = i;
1469 grp->slot_shift = q->min_slot_shift + i;
1470 for (j = 0; j < QFQ_MAX_SLOTS; j++)
1471 INIT_HLIST_HEAD(&grp->slots[j]);
1474 INIT_HLIST_HEAD(&q->nonfull_aggs);
1476 return 0;
1479 static void qfq_reset_qdisc(struct Qdisc *sch)
1481 struct qfq_sched *q = qdisc_priv(sch);
1482 struct qfq_class *cl;
1483 unsigned int i;
1485 for (i = 0; i < q->clhash.hashsize; i++) {
1486 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1487 if (cl->qdisc->q.qlen > 0)
1488 qfq_deactivate_class(q, cl);
1490 qdisc_reset(cl->qdisc);
1493 sch->q.qlen = 0;
1496 static void qfq_destroy_qdisc(struct Qdisc *sch)
1498 struct qfq_sched *q = qdisc_priv(sch);
1499 struct qfq_class *cl;
1500 struct hlist_node *next;
1501 unsigned int i;
1503 tcf_destroy_chain(&q->filter_list);
1505 for (i = 0; i < q->clhash.hashsize; i++) {
1506 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1507 common.hnode) {
1508 qfq_destroy_class(sch, cl);
1511 qdisc_class_hash_destroy(&q->clhash);
1514 static const struct Qdisc_class_ops qfq_class_ops = {
1515 .change = qfq_change_class,
1516 .delete = qfq_delete_class,
1517 .get = qfq_get_class,
1518 .put = qfq_put_class,
1519 .tcf_chain = qfq_tcf_chain,
1520 .bind_tcf = qfq_bind_tcf,
1521 .unbind_tcf = qfq_unbind_tcf,
1522 .graft = qfq_graft_class,
1523 .leaf = qfq_class_leaf,
1524 .qlen_notify = qfq_qlen_notify,
1525 .dump = qfq_dump_class,
1526 .dump_stats = qfq_dump_class_stats,
1527 .walk = qfq_walk,
1530 static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1531 .cl_ops = &qfq_class_ops,
1532 .id = "qfq",
1533 .priv_size = sizeof(struct qfq_sched),
1534 .enqueue = qfq_enqueue,
1535 .dequeue = qfq_dequeue,
1536 .peek = qdisc_peek_dequeued,
1537 .drop = qfq_drop,
1538 .init = qfq_init_qdisc,
1539 .reset = qfq_reset_qdisc,
1540 .destroy = qfq_destroy_qdisc,
1541 .owner = THIS_MODULE,
1544 static int __init qfq_init(void)
1546 return register_qdisc(&qfq_qdisc_ops);
1549 static void __exit qfq_exit(void)
1551 unregister_qdisc(&qfq_qdisc_ops);
1554 module_init(qfq_init);
1555 module_exit(qfq_exit);
1556 MODULE_LICENSE("GPL");