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 ========================
29 "Reducing the Execution Time of Fair-Queueing Schedulers."
30 http://algo.ing.unimo.it/people/paolo/agg-sched/agg-sched.pdf
34 [2] Fabio Checconi, Luigi Rizzo, and Paolo Valente: "QFQ: Efficient
35 Packet Scheduling with Tight Bandwidth Distribution Guarantees."
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
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
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 ]
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
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 };
130 struct qfq_aggregate;
133 struct Qdisc_class_common common;
135 unsigned int filter_cnt;
137 struct gnet_stats_basic_packed bstats;
138 struct gnet_stats_queue qstats;
139 struct net_rate_estimator __rcu *rate_est;
141 struct list_head alist; /* Link for active-classes list. */
142 struct qfq_aggregate *agg; /* Parent aggregate. */
143 int deficit; /* DRR deficit counter. */
146 struct qfq_aggregate {
147 struct hlist_node next; /* Link for the slot list. */
148 u64 S, F; /* flow timestamps (exact) */
150 /* group we belong to. In principle we would need the index,
151 * which is log_2(lmax/weight), but we never reference it
152 * directly, only the group.
154 struct qfq_group *grp;
156 /* these are copied from the flowset. */
157 u32 class_weight; /* Weight of each class in this aggregate. */
158 /* Max pkt size for the classes in this aggregate, DRR quantum. */
161 u32 inv_w; /* ONE_FP/(sum of weights of classes in aggr.). */
162 u32 budgetmax; /* Max budget for this aggregate. */
163 u32 initial_budget, budget; /* Initial and current budget. */
165 int num_classes; /* Number of classes in this aggr. */
166 struct list_head active; /* DRR queue of active classes. */
168 struct hlist_node nonfull_next; /* See nonfull_aggs in qfq_sched. */
172 u64 S, F; /* group timestamps (approx). */
173 unsigned int slot_shift; /* Slot shift. */
174 unsigned int index; /* Group index. */
175 unsigned int front; /* Index of the front slot. */
176 unsigned long full_slots; /* non-empty slots */
178 /* Array of RR lists of active aggregates. */
179 struct hlist_head slots[QFQ_MAX_SLOTS];
183 struct tcf_proto __rcu *filter_list;
184 struct tcf_block *block;
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
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);
217 return container_of(clc, struct qfq_class, common);
220 static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = {
221 [TCA_QFQ_WEIGHT] = { .type = NLA_U32 },
222 [TCA_QFQ_LMAX] = { .type = NLA_U32 },
226 * Calculate a flow index, given its weight and maximum packet length.
227 * index = log_2(maxlen/weight) but we need to apply the scaling.
228 * This is used only once at flow creation.
230 static int qfq_calc_index(u32 inv_w, unsigned int maxlen, u32 min_slot_shift)
232 u64 slot_size = (u64)maxlen * inv_w;
233 unsigned long size_map;
236 size_map = slot_size >> min_slot_shift;
240 index = __fls(size_map) + 1; /* basically a log_2 */
241 index -= !(slot_size - (1ULL << (index + min_slot_shift - 1)));
246 pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n",
247 (unsigned long) ONE_FP/inv_w, maxlen, index);
252 static void qfq_deactivate_agg(struct qfq_sched *, struct qfq_aggregate *);
253 static void qfq_activate_agg(struct qfq_sched *, struct qfq_aggregate *,
256 static void qfq_init_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
257 u32 lmax, u32 weight)
259 INIT_LIST_HEAD(&agg->active);
260 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
263 agg->class_weight = weight;
266 static struct qfq_aggregate *qfq_find_agg(struct qfq_sched *q,
267 u32 lmax, u32 weight)
269 struct qfq_aggregate *agg;
271 hlist_for_each_entry(agg, &q->nonfull_aggs, nonfull_next)
272 if (agg->lmax == lmax && agg->class_weight == weight)
279 /* Update aggregate as a function of the new number of classes. */
280 static void qfq_update_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
285 if (new_num_classes == q->max_agg_classes)
286 hlist_del_init(&agg->nonfull_next);
288 if (agg->num_classes > new_num_classes &&
289 new_num_classes == q->max_agg_classes - 1) /* agg no more full */
290 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
292 /* The next assignment may let
293 * agg->initial_budget > agg->budgetmax
294 * hold, we will take it into account in charge_actual_service().
296 agg->budgetmax = new_num_classes * agg->lmax;
297 new_agg_weight = agg->class_weight * new_num_classes;
298 agg->inv_w = ONE_FP/new_agg_weight;
300 if (agg->grp == NULL) {
301 int i = qfq_calc_index(agg->inv_w, agg->budgetmax,
303 agg->grp = &q->groups[i];
307 (int) agg->class_weight * (new_num_classes - agg->num_classes);
308 q->iwsum = ONE_FP / q->wsum;
310 agg->num_classes = new_num_classes;
313 /* Add class to aggregate. */
314 static void qfq_add_to_agg(struct qfq_sched *q,
315 struct qfq_aggregate *agg,
316 struct qfq_class *cl)
320 qfq_update_agg(q, agg, agg->num_classes+1);
321 if (cl->qdisc->q.qlen > 0) { /* adding an active class */
322 list_add_tail(&cl->alist, &agg->active);
323 if (list_first_entry(&agg->active, struct qfq_class, alist) ==
324 cl && q->in_serv_agg != agg) /* agg was inactive */
325 qfq_activate_agg(q, agg, enqueue); /* schedule agg */
329 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *);
331 static void qfq_destroy_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
333 hlist_del_init(&agg->nonfull_next);
334 q->wsum -= agg->class_weight;
336 q->iwsum = ONE_FP / q->wsum;
338 if (q->in_serv_agg == agg)
339 q->in_serv_agg = qfq_choose_next_agg(q);
343 /* Deschedule class from within its parent aggregate. */
344 static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
346 struct qfq_aggregate *agg = cl->agg;
349 list_del(&cl->alist); /* remove from RR queue of the aggregate */
350 if (list_empty(&agg->active)) /* agg is now inactive */
351 qfq_deactivate_agg(q, agg);
354 /* Remove class from its parent aggregate. */
355 static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
357 struct qfq_aggregate *agg = cl->agg;
360 if (agg->num_classes == 1) { /* agg being emptied, destroy it */
361 qfq_destroy_agg(q, agg);
364 qfq_update_agg(q, agg, agg->num_classes-1);
367 /* Deschedule class and remove it from its parent aggregate. */
368 static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
370 if (cl->qdisc->q.qlen > 0) /* class is active */
371 qfq_deactivate_class(q, cl);
373 qfq_rm_from_agg(q, cl);
376 /* Move class to a new aggregate, matching the new class weight and/or lmax */
377 static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight,
380 struct qfq_sched *q = qdisc_priv(sch);
381 struct qfq_aggregate *new_agg = qfq_find_agg(q, lmax, weight);
383 if (new_agg == NULL) { /* create new aggregate */
384 new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC);
387 qfq_init_agg(q, new_agg, lmax, weight);
389 qfq_deact_rm_from_agg(q, cl);
390 qfq_add_to_agg(q, new_agg, cl);
395 static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
396 struct nlattr **tca, unsigned long *arg,
397 struct netlink_ext_ack *extack)
399 struct qfq_sched *q = qdisc_priv(sch);
400 struct qfq_class *cl = (struct qfq_class *)*arg;
401 bool existing = false;
402 struct nlattr *tb[TCA_QFQ_MAX + 1];
403 struct qfq_aggregate *new_agg = NULL;
404 u32 weight, lmax, inv_w;
408 if (tca[TCA_OPTIONS] == NULL) {
409 pr_notice("qfq: no options\n");
413 err = nla_parse_nested_deprecated(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS],
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);
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);
434 lmax = psched_mtu(qdisc_dev(sch));
436 inv_w = ONE_FP / weight;
437 weight = ONE_FP / inv_w;
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",
452 if (cl != NULL) { /* modify existing class */
454 err = gen_replace_estimator(&cl->bstats, NULL,
457 qdisc_root_sleeping_running(sch),
466 /* create and init new class */
467 cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
471 cl->common.classid = classid;
474 cl->qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
476 if (cl->qdisc == NULL)
477 cl->qdisc = &noop_qdisc;
480 err = gen_new_estimator(&cl->bstats, NULL,
483 qdisc_root_sleeping_running(sch),
489 if (cl->qdisc != &noop_qdisc)
490 qdisc_hash_add(cl->qdisc, true);
492 qdisc_class_hash_insert(&q->clhash, &cl->common);
493 sch_tree_unlock(sch);
495 qdisc_class_hash_grow(sch, &q->clhash);
499 new_agg = qfq_find_agg(q, lmax, weight);
500 if (new_agg == NULL) { /* create new aggregate */
501 sch_tree_unlock(sch);
502 new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL);
503 if (new_agg == NULL) {
505 gen_kill_estimator(&cl->rate_est);
509 qfq_init_agg(q, new_agg, lmax, weight);
512 qfq_deact_rm_from_agg(q, cl);
513 qfq_add_to_agg(q, new_agg, cl);
514 sch_tree_unlock(sch);
516 *arg = (unsigned long)cl;
520 qdisc_put(cl->qdisc);
525 static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
527 struct qfq_sched *q = qdisc_priv(sch);
529 qfq_rm_from_agg(q, cl);
530 gen_kill_estimator(&cl->rate_est);
531 qdisc_put(cl->qdisc);
535 static int qfq_delete_class(struct Qdisc *sch, unsigned long arg)
537 struct qfq_sched *q = qdisc_priv(sch);
538 struct qfq_class *cl = (struct qfq_class *)arg;
540 if (cl->filter_cnt > 0)
545 qdisc_purge_queue(cl->qdisc);
546 qdisc_class_hash_remove(&q->clhash, &cl->common);
548 sch_tree_unlock(sch);
550 qfq_destroy_class(sch, cl);
554 static unsigned long qfq_search_class(struct Qdisc *sch, u32 classid)
556 return (unsigned long)qfq_find_class(sch, classid);
559 static struct tcf_block *qfq_tcf_block(struct Qdisc *sch, unsigned long cl,
560 struct netlink_ext_ack *extack)
562 struct qfq_sched *q = qdisc_priv(sch);
570 static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
573 struct qfq_class *cl = qfq_find_class(sch, classid);
578 return (unsigned long)cl;
581 static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
583 struct qfq_class *cl = (struct qfq_class *)arg;
588 static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
589 struct Qdisc *new, struct Qdisc **old,
590 struct netlink_ext_ack *extack)
592 struct qfq_class *cl = (struct qfq_class *)arg;
595 new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
596 cl->common.classid, NULL);
601 *old = qdisc_replace(sch, new, &cl->qdisc);
605 static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
607 struct qfq_class *cl = (struct qfq_class *)arg;
612 static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
613 struct sk_buff *skb, struct tcmsg *tcm)
615 struct qfq_class *cl = (struct qfq_class *)arg;
618 tcm->tcm_parent = TC_H_ROOT;
619 tcm->tcm_handle = cl->common.classid;
620 tcm->tcm_info = cl->qdisc->handle;
622 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
624 goto nla_put_failure;
625 if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) ||
626 nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax))
627 goto nla_put_failure;
628 return nla_nest_end(skb, nest);
631 nla_nest_cancel(skb, nest);
635 static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
638 struct qfq_class *cl = (struct qfq_class *)arg;
639 struct tc_qfq_stats xstats;
641 memset(&xstats, 0, sizeof(xstats));
643 xstats.weight = cl->agg->class_weight;
644 xstats.lmax = cl->agg->lmax;
646 if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
647 d, NULL, &cl->bstats) < 0 ||
648 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
649 qdisc_qstats_copy(d, cl->qdisc) < 0)
652 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
655 static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
657 struct qfq_sched *q = qdisc_priv(sch);
658 struct qfq_class *cl;
664 for (i = 0; i < q->clhash.hashsize; i++) {
665 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
666 if (arg->count < arg->skip) {
670 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
679 static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
682 struct qfq_sched *q = qdisc_priv(sch);
683 struct qfq_class *cl;
684 struct tcf_result res;
685 struct tcf_proto *fl;
688 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
689 pr_debug("qfq_classify: found %d\n", skb->priority);
690 cl = qfq_find_class(sch, skb->priority);
695 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
696 fl = rcu_dereference_bh(q->filter_list);
697 result = tcf_classify(skb, fl, &res, false);
699 #ifdef CONFIG_NET_CLS_ACT
704 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
710 cl = (struct qfq_class *)res.class;
712 cl = qfq_find_class(sch, res.classid);
719 /* Generic comparison function, handling wraparound. */
720 static inline int qfq_gt(u64 a, u64 b)
722 return (s64)(a - b) > 0;
725 /* Round a precise timestamp to its slotted value. */
726 static inline u64 qfq_round_down(u64 ts, unsigned int shift)
728 return ts & ~((1ULL << shift) - 1);
731 /* return the pointer to the group with lowest index in the bitmap */
732 static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
733 unsigned long bitmap)
735 int index = __ffs(bitmap);
736 return &q->groups[index];
738 /* Calculate a mask to mimic what would be ffs_from(). */
739 static inline unsigned long mask_from(unsigned long bitmap, int from)
741 return bitmap & ~((1UL << from) - 1);
745 * The state computation relies on ER=0, IR=1, EB=2, IB=3
746 * First compute eligibility comparing grp->S, q->V,
747 * then check if someone is blocking us and possibly add EB
749 static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
751 /* if S > V we are not eligible */
752 unsigned int state = qfq_gt(grp->S, q->V);
753 unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
754 struct qfq_group *next;
757 next = qfq_ffs(q, mask);
758 if (qfq_gt(grp->F, next->F))
768 * q->bitmaps[dst] |= q->bitmaps[src] & mask;
769 * q->bitmaps[src] &= ~mask;
770 * but we should make sure that src != dst
772 static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
775 q->bitmaps[dst] |= q->bitmaps[src] & mask;
776 q->bitmaps[src] &= ~mask;
779 static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
781 unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
782 struct qfq_group *next;
785 next = qfq_ffs(q, mask);
786 if (!qfq_gt(next->F, old_F))
790 mask = (1UL << index) - 1;
791 qfq_move_groups(q, mask, EB, ER);
792 qfq_move_groups(q, mask, IB, IR);
799 old_V >>= q->min_slot_shift;
805 static void qfq_make_eligible(struct qfq_sched *q)
807 unsigned long vslot = q->V >> q->min_slot_shift;
808 unsigned long old_vslot = q->oldV >> q->min_slot_shift;
810 if (vslot != old_vslot) {
812 int last_flip_pos = fls(vslot ^ old_vslot);
814 if (last_flip_pos > 31) /* higher than the number of groups */
815 mask = ~0UL; /* make all groups eligible */
817 mask = (1UL << last_flip_pos) - 1;
819 qfq_move_groups(q, mask, IR, ER);
820 qfq_move_groups(q, mask, IB, EB);
825 * The index of the slot in which the input aggregate agg is to be
826 * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2'
827 * and not a '-1' because the start time of the group may be moved
828 * backward by one slot after the aggregate has been inserted, and
829 * this would cause non-empty slots to be right-shifted by one
832 * QFQ+ fully satisfies this bound to the slot index if the parameters
833 * of the classes are not changed dynamically, and if QFQ+ never
834 * happens to postpone the service of agg unjustly, i.e., it never
835 * happens that the aggregate becomes backlogged and eligible, or just
836 * eligible, while an aggregate with a higher approximated finish time
837 * is being served. In particular, in this case QFQ+ guarantees that
838 * the timestamps of agg are low enough that the slot index is never
839 * higher than 2. Unfortunately, QFQ+ cannot provide the same
840 * guarantee if it happens to unjustly postpone the service of agg, or
841 * if the parameters of some class are changed.
843 * As for the first event, i.e., an out-of-order service, the
844 * upper bound to the slot index guaranteed by QFQ+ grows to
846 * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
847 * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1.
849 * The following function deals with this problem by backward-shifting
850 * the timestamps of agg, if needed, so as to guarantee that the slot
851 * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may
852 * cause the service of other aggregates to be postponed, yet the
853 * worst-case guarantees of these aggregates are not violated. In
854 * fact, in case of no out-of-order service, the timestamps of agg
855 * would have been even lower than they are after the backward shift,
856 * because QFQ+ would have guaranteed a maximum value equal to 2 for
857 * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose
858 * service is postponed because of the backward-shift would have
859 * however waited for the service of agg before being served.
861 * The other event that may cause the slot index to be higher than 2
862 * for agg is a recent change of the parameters of some class. If the
863 * weight of a class is increased or the lmax (max_pkt_size) of the
864 * class is decreased, then a new aggregate with smaller slot size
865 * than the original parent aggregate of the class may happen to be
866 * activated. The activation of this aggregate should be properly
867 * delayed to when the service of the class has finished in the ideal
868 * system tracked by QFQ+. If the activation of the aggregate is not
869 * delayed to this reference time instant, then this aggregate may be
870 * unjustly served before other aggregates waiting for service. This
871 * may cause the above bound to the slot index to be violated for some
872 * of these unlucky aggregates.
874 * Instead of delaying the activation of the new aggregate, which is
875 * quite complex, the above-discussed capping of the slot index is
876 * used to handle also the consequences of a change of the parameters
879 static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg,
882 u64 slot = (roundedS - grp->S) >> grp->slot_shift;
883 unsigned int i; /* slot index in the bucket list */
885 if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
886 u64 deltaS = roundedS - grp->S -
887 ((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
890 slot = QFQ_MAX_SLOTS - 2;
893 i = (grp->front + slot) % QFQ_MAX_SLOTS;
895 hlist_add_head(&agg->next, &grp->slots[i]);
896 __set_bit(slot, &grp->full_slots);
899 /* Maybe introduce hlist_first_entry?? */
900 static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp)
902 return hlist_entry(grp->slots[grp->front].first,
903 struct qfq_aggregate, next);
907 * remove the entry from the slot
909 static void qfq_front_slot_remove(struct qfq_group *grp)
911 struct qfq_aggregate *agg = qfq_slot_head(grp);
914 hlist_del(&agg->next);
915 if (hlist_empty(&grp->slots[grp->front]))
916 __clear_bit(0, &grp->full_slots);
920 * Returns the first aggregate in the first non-empty bucket of the
921 * group. As a side effect, adjusts the bucket list so the first
922 * non-empty bucket is at position 0 in full_slots.
924 static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp)
928 pr_debug("qfq slot_scan: grp %u full %#lx\n",
929 grp->index, grp->full_slots);
931 if (grp->full_slots == 0)
934 i = __ffs(grp->full_slots); /* zero based */
936 grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
937 grp->full_slots >>= i;
940 return qfq_slot_head(grp);
944 * adjust the bucket list. When the start time of a group decreases,
945 * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
946 * move the objects. The mask of occupied slots must be shifted
947 * because we use ffs() to find the first non-empty slot.
948 * This covers decreases in the group's start time, but what about
949 * increases of the start time ?
950 * Here too we should make sure that i is less than 32
952 static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
954 unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
956 grp->full_slots <<= i;
957 grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
960 static void qfq_update_eligible(struct qfq_sched *q)
962 struct qfq_group *grp;
963 unsigned long ineligible;
965 ineligible = q->bitmaps[IR] | q->bitmaps[IB];
967 if (!q->bitmaps[ER]) {
968 grp = qfq_ffs(q, ineligible);
969 if (qfq_gt(grp->S, q->V))
972 qfq_make_eligible(q);
976 /* Dequeue head packet of the head class in the DRR queue of the aggregate. */
977 static void agg_dequeue(struct qfq_aggregate *agg,
978 struct qfq_class *cl, unsigned int len)
980 qdisc_dequeue_peeked(cl->qdisc);
982 cl->deficit -= (int) len;
984 if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */
985 list_del(&cl->alist);
986 else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) {
987 cl->deficit += agg->lmax;
988 list_move_tail(&cl->alist, &agg->active);
992 static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg,
993 struct qfq_class **cl,
998 *cl = list_first_entry(&agg->active, struct qfq_class, alist);
999 skb = (*cl)->qdisc->ops->peek((*cl)->qdisc);
1001 WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
1003 *len = qdisc_pkt_len(skb);
1008 /* Update F according to the actual service received by the aggregate. */
1009 static inline void charge_actual_service(struct qfq_aggregate *agg)
1011 /* Compute the service received by the aggregate, taking into
1012 * account that, after decreasing the number of classes in
1013 * agg, it may happen that
1014 * agg->initial_budget - agg->budget > agg->bugdetmax
1016 u32 service_received = min(agg->budgetmax,
1017 agg->initial_budget - agg->budget);
1019 agg->F = agg->S + (u64)service_received * agg->inv_w;
1022 /* Assign a reasonable start time for a new aggregate in group i.
1023 * Admissible values for \hat(F) are multiples of \sigma_i
1024 * no greater than V+\sigma_i . Larger values mean that
1025 * we had a wraparound so we consider the timestamp to be stale.
1027 * If F is not stale and F >= V then we set S = F.
1028 * Otherwise we should assign S = V, but this may violate
1029 * the ordering in EB (see [2]). So, if we have groups in ER,
1030 * set S to the F_j of the first group j which would be blocking us.
1031 * We are guaranteed not to move S backward because
1032 * otherwise our group i would still be blocked.
1034 static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg)
1037 u64 limit, roundedF;
1038 int slot_shift = agg->grp->slot_shift;
1040 roundedF = qfq_round_down(agg->F, slot_shift);
1041 limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
1043 if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) {
1044 /* timestamp was stale */
1045 mask = mask_from(q->bitmaps[ER], agg->grp->index);
1047 struct qfq_group *next = qfq_ffs(q, mask);
1048 if (qfq_gt(roundedF, next->F)) {
1049 if (qfq_gt(limit, next->F))
1051 else /* preserve timestamp correctness */
1057 } else /* timestamp is not stale */
1061 /* Update the timestamps of agg before scheduling/rescheduling it for
1062 * service. In particular, assign to agg->F its maximum possible
1063 * value, i.e., the virtual finish time with which the aggregate
1064 * should be labeled if it used all its budget once in service.
1067 qfq_update_agg_ts(struct qfq_sched *q,
1068 struct qfq_aggregate *agg, enum update_reason reason)
1070 if (reason != requeue)
1071 qfq_update_start(q, agg);
1072 else /* just charge agg for the service received */
1075 agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w;
1078 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg);
1080 static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
1082 struct qfq_sched *q = qdisc_priv(sch);
1083 struct qfq_aggregate *in_serv_agg = q->in_serv_agg;
1084 struct qfq_class *cl;
1085 struct sk_buff *skb = NULL;
1086 /* next-packet len, 0 means no more active classes in in-service agg */
1087 unsigned int len = 0;
1089 if (in_serv_agg == NULL)
1092 if (!list_empty(&in_serv_agg->active))
1093 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1096 * If there are no active classes in the in-service aggregate,
1097 * or if the aggregate has not enough budget to serve its next
1098 * class, then choose the next aggregate to serve.
1100 if (len == 0 || in_serv_agg->budget < len) {
1101 charge_actual_service(in_serv_agg);
1103 /* recharge the budget of the aggregate */
1104 in_serv_agg->initial_budget = in_serv_agg->budget =
1105 in_serv_agg->budgetmax;
1107 if (!list_empty(&in_serv_agg->active)) {
1109 * Still active: reschedule for
1110 * service. Possible optimization: if no other
1111 * aggregate is active, then there is no point
1112 * in rescheduling this aggregate, and we can
1113 * just keep it as the in-service one. This
1114 * should be however a corner case, and to
1115 * handle it, we would need to maintain an
1116 * extra num_active_aggs field.
1118 qfq_update_agg_ts(q, in_serv_agg, requeue);
1119 qfq_schedule_agg(q, in_serv_agg);
1120 } else if (sch->q.qlen == 0) { /* no aggregate to serve */
1121 q->in_serv_agg = NULL;
1126 * If we get here, there are other aggregates queued:
1127 * choose the new aggregate to serve.
1129 in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
1130 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1135 qdisc_qstats_backlog_dec(sch, skb);
1137 qdisc_bstats_update(sch, skb);
1139 agg_dequeue(in_serv_agg, cl, len);
1140 /* If lmax is lowered, through qfq_change_class, for a class
1141 * owning pending packets with larger size than the new value
1142 * of lmax, then the following condition may hold.
1144 if (unlikely(in_serv_agg->budget < len))
1145 in_serv_agg->budget = 0;
1147 in_serv_agg->budget -= len;
1149 q->V += (u64)len * q->iwsum;
1150 pr_debug("qfq dequeue: len %u F %lld now %lld\n",
1151 len, (unsigned long long) in_serv_agg->F,
1152 (unsigned long long) q->V);
1157 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)
1159 struct qfq_group *grp;
1160 struct qfq_aggregate *agg, *new_front_agg;
1163 qfq_update_eligible(q);
1166 if (!q->bitmaps[ER])
1169 grp = qfq_ffs(q, q->bitmaps[ER]);
1172 agg = qfq_slot_head(grp);
1174 /* agg starts to be served, remove it from schedule */
1175 qfq_front_slot_remove(grp);
1177 new_front_agg = qfq_slot_scan(grp);
1179 if (new_front_agg == NULL) /* group is now inactive, remove from ER */
1180 __clear_bit(grp->index, &q->bitmaps[ER]);
1182 u64 roundedS = qfq_round_down(new_front_agg->S,
1186 if (grp->S == roundedS)
1189 grp->F = roundedS + (2ULL << grp->slot_shift);
1190 __clear_bit(grp->index, &q->bitmaps[ER]);
1191 s = qfq_calc_state(q, grp);
1192 __set_bit(grp->index, &q->bitmaps[s]);
1195 qfq_unblock_groups(q, grp->index, old_F);
1200 static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
1201 struct sk_buff **to_free)
1203 unsigned int len = qdisc_pkt_len(skb), gso_segs;
1204 struct qfq_sched *q = qdisc_priv(sch);
1205 struct qfq_class *cl;
1206 struct qfq_aggregate *agg;
1210 cl = qfq_classify(skb, sch, &err);
1212 if (err & __NET_XMIT_BYPASS)
1213 qdisc_qstats_drop(sch);
1214 __qdisc_drop(skb, to_free);
1217 pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
1219 if (unlikely(cl->agg->lmax < len)) {
1220 pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
1221 cl->agg->lmax, len, cl->common.classid);
1222 err = qfq_change_agg(sch, cl, cl->agg->class_weight, len);
1225 return qdisc_drop(skb, sch, to_free);
1229 gso_segs = skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 1;
1230 first = !cl->qdisc->q.qlen;
1231 err = qdisc_enqueue(skb, cl->qdisc, to_free);
1232 if (unlikely(err != NET_XMIT_SUCCESS)) {
1233 pr_debug("qfq_enqueue: enqueue failed %d\n", err);
1234 if (net_xmit_drop_count(err)) {
1236 qdisc_qstats_drop(sch);
1241 cl->bstats.bytes += len;
1242 cl->bstats.packets += gso_segs;
1243 sch->qstats.backlog += len;
1247 /* if the queue was not empty, then done here */
1249 if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) &&
1250 list_first_entry(&agg->active, struct qfq_class, alist)
1251 == cl && cl->deficit < len)
1252 list_move_tail(&cl->alist, &agg->active);
1257 /* schedule class for service within the aggregate */
1258 cl->deficit = agg->lmax;
1259 list_add_tail(&cl->alist, &agg->active);
1261 if (list_first_entry(&agg->active, struct qfq_class, alist) != cl ||
1262 q->in_serv_agg == agg)
1263 return err; /* non-empty or in service, nothing else to do */
1265 qfq_activate_agg(q, agg, enqueue);
1271 * Schedule aggregate according to its timestamps.
1273 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1275 struct qfq_group *grp = agg->grp;
1279 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1282 * Insert agg in the correct bucket.
1283 * If agg->S >= grp->S we don't need to adjust the
1284 * bucket list and simply go to the insertion phase.
1285 * Otherwise grp->S is decreasing, we must make room
1286 * in the bucket list, and also recompute the group state.
1287 * Finally, if there were no flows in this group and nobody
1288 * was in ER make sure to adjust V.
1290 if (grp->full_slots) {
1291 if (!qfq_gt(grp->S, agg->S))
1294 /* create a slot for this agg->S */
1295 qfq_slot_rotate(grp, roundedS);
1296 /* group was surely ineligible, remove */
1297 __clear_bit(grp->index, &q->bitmaps[IR]);
1298 __clear_bit(grp->index, &q->bitmaps[IB]);
1299 } else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) &&
1300 q->in_serv_agg == NULL)
1304 grp->F = roundedS + (2ULL << grp->slot_shift);
1305 s = qfq_calc_state(q, grp);
1306 __set_bit(grp->index, &q->bitmaps[s]);
1308 pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
1310 (unsigned long long) agg->S,
1311 (unsigned long long) agg->F,
1312 (unsigned long long) q->V);
1315 qfq_slot_insert(grp, agg, roundedS);
1319 /* Update agg ts and schedule agg for service */
1320 static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
1321 enum update_reason reason)
1323 agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */
1325 qfq_update_agg_ts(q, agg, reason);
1326 if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */
1327 q->in_serv_agg = agg; /* start serving this aggregate */
1328 /* update V: to be in service, agg must be eligible */
1329 q->oldV = q->V = agg->S;
1330 } else if (agg != q->in_serv_agg)
1331 qfq_schedule_agg(q, agg);
1334 static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
1335 struct qfq_aggregate *agg)
1337 unsigned int i, offset;
1340 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1341 offset = (roundedS - grp->S) >> grp->slot_shift;
1343 i = (grp->front + offset) % QFQ_MAX_SLOTS;
1345 hlist_del(&agg->next);
1346 if (hlist_empty(&grp->slots[i]))
1347 __clear_bit(offset, &grp->full_slots);
1351 * Called to forcibly deschedule an aggregate. If the aggregate is
1352 * not in the front bucket, or if the latter has other aggregates in
1353 * the front bucket, we can simply remove the aggregate with no other
1355 * Otherwise we must propagate the event up.
1357 static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1359 struct qfq_group *grp = agg->grp;
1364 if (agg == q->in_serv_agg) {
1365 charge_actual_service(agg);
1366 q->in_serv_agg = qfq_choose_next_agg(q);
1371 qfq_slot_remove(q, grp, agg);
1373 if (!grp->full_slots) {
1374 __clear_bit(grp->index, &q->bitmaps[IR]);
1375 __clear_bit(grp->index, &q->bitmaps[EB]);
1376 __clear_bit(grp->index, &q->bitmaps[IB]);
1378 if (test_bit(grp->index, &q->bitmaps[ER]) &&
1379 !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
1380 mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
1382 mask = ~((1UL << __fls(mask)) - 1);
1385 qfq_move_groups(q, mask, EB, ER);
1386 qfq_move_groups(q, mask, IB, IR);
1388 __clear_bit(grp->index, &q->bitmaps[ER]);
1389 } else if (hlist_empty(&grp->slots[grp->front])) {
1390 agg = qfq_slot_scan(grp);
1391 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1392 if (grp->S != roundedS) {
1393 __clear_bit(grp->index, &q->bitmaps[ER]);
1394 __clear_bit(grp->index, &q->bitmaps[IR]);
1395 __clear_bit(grp->index, &q->bitmaps[EB]);
1396 __clear_bit(grp->index, &q->bitmaps[IB]);
1398 grp->F = roundedS + (2ULL << grp->slot_shift);
1399 s = qfq_calc_state(q, grp);
1400 __set_bit(grp->index, &q->bitmaps[s]);
1405 static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1407 struct qfq_sched *q = qdisc_priv(sch);
1408 struct qfq_class *cl = (struct qfq_class *)arg;
1410 qfq_deactivate_class(q, cl);
1413 static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt,
1414 struct netlink_ext_ack *extack)
1416 struct qfq_sched *q = qdisc_priv(sch);
1417 struct qfq_group *grp;
1419 u32 max_cl_shift, maxbudg_shift, max_classes;
1421 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
1425 err = qdisc_class_hash_init(&q->clhash);
1429 if (qdisc_dev(sch)->tx_queue_len + 1 > QFQ_MAX_AGG_CLASSES)
1430 max_classes = QFQ_MAX_AGG_CLASSES;
1432 max_classes = qdisc_dev(sch)->tx_queue_len + 1;
1433 /* max_cl_shift = floor(log_2(max_classes)) */
1434 max_cl_shift = __fls(max_classes);
1435 q->max_agg_classes = 1<<max_cl_shift;
1437 /* maxbudg_shift = log2(max_len * max_classes_per_agg) */
1438 maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift;
1439 q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX;
1441 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1442 grp = &q->groups[i];
1444 grp->slot_shift = q->min_slot_shift + i;
1445 for (j = 0; j < QFQ_MAX_SLOTS; j++)
1446 INIT_HLIST_HEAD(&grp->slots[j]);
1449 INIT_HLIST_HEAD(&q->nonfull_aggs);
1454 static void qfq_reset_qdisc(struct Qdisc *sch)
1456 struct qfq_sched *q = qdisc_priv(sch);
1457 struct qfq_class *cl;
1460 for (i = 0; i < q->clhash.hashsize; i++) {
1461 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1462 if (cl->qdisc->q.qlen > 0)
1463 qfq_deactivate_class(q, cl);
1465 qdisc_reset(cl->qdisc);
1468 sch->qstats.backlog = 0;
1472 static void qfq_destroy_qdisc(struct Qdisc *sch)
1474 struct qfq_sched *q = qdisc_priv(sch);
1475 struct qfq_class *cl;
1476 struct hlist_node *next;
1479 tcf_block_put(q->block);
1481 for (i = 0; i < q->clhash.hashsize; i++) {
1482 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1484 qfq_destroy_class(sch, cl);
1487 qdisc_class_hash_destroy(&q->clhash);
1490 static const struct Qdisc_class_ops qfq_class_ops = {
1491 .change = qfq_change_class,
1492 .delete = qfq_delete_class,
1493 .find = qfq_search_class,
1494 .tcf_block = qfq_tcf_block,
1495 .bind_tcf = qfq_bind_tcf,
1496 .unbind_tcf = qfq_unbind_tcf,
1497 .graft = qfq_graft_class,
1498 .leaf = qfq_class_leaf,
1499 .qlen_notify = qfq_qlen_notify,
1500 .dump = qfq_dump_class,
1501 .dump_stats = qfq_dump_class_stats,
1505 static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1506 .cl_ops = &qfq_class_ops,
1508 .priv_size = sizeof(struct qfq_sched),
1509 .enqueue = qfq_enqueue,
1510 .dequeue = qfq_dequeue,
1511 .peek = qdisc_peek_dequeued,
1512 .init = qfq_init_qdisc,
1513 .reset = qfq_reset_qdisc,
1514 .destroy = qfq_destroy_qdisc,
1515 .owner = THIS_MODULE,
1518 static int __init qfq_init(void)
1520 return register_qdisc(&qfq_qdisc_ops);
1523 static void __exit qfq_exit(void)
1525 unregister_qdisc(&qfq_qdisc_ops);
1528 module_init(qfq_init);
1529 module_exit(qfq_exit);
1530 MODULE_LICENSE("GPL");