1 // SPDX-License-Identifier: GPL-2.0-only
3 * net/sched/sch_qfq.c Quick Fair Queueing Plus Scheduler.
5 * Copyright (c) 2009 Fabio Checconi, Luigi Rizzo, and Paolo Valente.
6 * Copyright (c) 2012 Paolo Valente.
9 #include <linux/module.h>
10 #include <linux/init.h>
11 #include <linux/bitops.h>
12 #include <linux/errno.h>
13 #include <linux/netdevice.h>
14 #include <linux/pkt_sched.h>
15 #include <net/sch_generic.h>
16 #include <net/pkt_sched.h>
17 #include <net/pkt_cls.h>
20 /* Quick Fair Queueing Plus
21 ========================
26 "Reducing the Execution Time of Fair-Queueing Schedulers."
27 http://algo.ing.unimo.it/people/paolo/agg-sched/agg-sched.pdf
31 [2] Fabio Checconi, Luigi Rizzo, and Paolo Valente: "QFQ: Efficient
32 Packet Scheduling with Tight Bandwidth Distribution Guarantees."
35 http://retis.sssup.it/~fabio/linux/qfq/
40 QFQ+ divides classes into aggregates of at most MAX_AGG_CLASSES
41 classes. Each aggregate is timestamped with a virtual start time S
42 and a virtual finish time F, and scheduled according to its
43 timestamps. S and F are computed as a function of a system virtual
44 time function V. The classes within each aggregate are instead
47 To speed up operations, QFQ+ divides also aggregates into a limited
48 number of groups. Which group a class belongs to depends on the
49 ratio between the maximum packet length for the class and the weight
50 of the class. Groups have their own S and F. In the end, QFQ+
51 schedules groups, then aggregates within groups, then classes within
52 aggregates. See [1] and [2] for a full description.
54 Virtual time computations.
56 S, F and V are all computed in fixed point arithmetic with
57 FRAC_BITS decimal bits.
59 QFQ_MAX_INDEX is the maximum index allowed for a group. We need
61 QFQ_MAX_WSHIFT is the maximum power of two supported as a weight.
63 The layout of the bits is as below:
65 [ MTU_SHIFT ][ FRAC_BITS ]
66 [ MAX_INDEX ][ MIN_SLOT_SHIFT ]
70 where MIN_SLOT_SHIFT is derived by difference from the others.
72 The max group index corresponds to Lmax/w_min, where
73 Lmax=1<<MTU_SHIFT, w_min = 1 .
74 From this, and knowing how many groups (MAX_INDEX) we want,
75 we can derive the shift corresponding to each group.
77 Because we often need to compute
78 F = S + len/w_i and V = V + len/wsum
79 instead of storing w_i store the value
80 inv_w = (1<<FRAC_BITS)/w_i
81 so we can do F = S + len * inv_w * wsum.
82 We use W_TOT in the formulas so we can easily move between
83 static and adaptive weight sum.
85 The per-scheduler-instance data contain all the data structures
86 for the scheduler: bitmaps and bucket lists.
91 * Maximum number of consecutive slots occupied by backlogged classes
94 #define QFQ_MAX_SLOTS 32
97 * Shifts used for aggregate<->group mapping. We allow class weights that are
98 * in the range [1, 2^MAX_WSHIFT], and we try to map each aggregate i to the
99 * group with the smallest index that can support the L_i / r_i configured
100 * for the classes in the aggregate.
102 * grp->index is the index of the group; and grp->slot_shift
103 * is the shift for the corresponding (scaled) sigma_i.
105 #define QFQ_MAX_INDEX 24
106 #define QFQ_MAX_WSHIFT 10
108 #define QFQ_MAX_WEIGHT (1<<QFQ_MAX_WSHIFT) /* see qfq_slot_insert */
109 #define QFQ_MAX_WSUM (64*QFQ_MAX_WEIGHT)
111 #define FRAC_BITS 30 /* fixed point arithmetic */
112 #define ONE_FP (1UL << FRAC_BITS)
114 #define QFQ_MTU_SHIFT 16 /* to support TSO/GSO */
115 #define QFQ_MIN_LMAX 512 /* see qfq_slot_insert */
116 #define QFQ_MAX_LMAX (1UL << QFQ_MTU_SHIFT)
118 #define QFQ_MAX_AGG_CLASSES 8 /* max num classes per aggregate allowed */
121 * Possible group states. These values are used as indexes for the bitmaps
122 * array of struct qfq_queue.
124 enum qfq_state { ER, IR, EB, IB, QFQ_MAX_STATE };
128 struct qfq_aggregate;
131 struct Qdisc_class_common common;
133 unsigned int filter_cnt;
135 struct gnet_stats_basic_sync bstats;
136 struct gnet_stats_queue qstats;
137 struct net_rate_estimator __rcu *rate_est;
139 struct list_head alist; /* Link for active-classes list. */
140 struct qfq_aggregate *agg; /* Parent aggregate. */
141 int deficit; /* DRR deficit counter. */
144 struct qfq_aggregate {
145 struct hlist_node next; /* Link for the slot list. */
146 u64 S, F; /* flow timestamps (exact) */
148 /* group we belong to. In principle we would need the index,
149 * which is log_2(lmax/weight), but we never reference it
150 * directly, only the group.
152 struct qfq_group *grp;
154 /* these are copied from the flowset. */
155 u32 class_weight; /* Weight of each class in this aggregate. */
156 /* Max pkt size for the classes in this aggregate, DRR quantum. */
159 u32 inv_w; /* ONE_FP/(sum of weights of classes in aggr.). */
160 u32 budgetmax; /* Max budget for this aggregate. */
161 u32 initial_budget, budget; /* Initial and current budget. */
163 int num_classes; /* Number of classes in this aggr. */
164 struct list_head active; /* DRR queue of active classes. */
166 struct hlist_node nonfull_next; /* See nonfull_aggs in qfq_sched. */
170 u64 S, F; /* group timestamps (approx). */
171 unsigned int slot_shift; /* Slot shift. */
172 unsigned int index; /* Group index. */
173 unsigned int front; /* Index of the front slot. */
174 unsigned long full_slots; /* non-empty slots */
176 /* Array of RR lists of active aggregates. */
177 struct hlist_head slots[QFQ_MAX_SLOTS];
181 struct tcf_proto __rcu *filter_list;
182 struct tcf_block *block;
183 struct Qdisc_class_hash clhash;
185 u64 oldV, V; /* Precise virtual times. */
186 struct qfq_aggregate *in_serv_agg; /* Aggregate being served. */
187 u32 wsum; /* weight sum */
188 u32 iwsum; /* inverse weight sum */
190 unsigned long bitmaps[QFQ_MAX_STATE]; /* Group bitmaps. */
191 struct qfq_group groups[QFQ_MAX_INDEX + 1]; /* The groups. */
192 u32 min_slot_shift; /* Index of the group-0 bit in the bitmaps. */
194 u32 max_agg_classes; /* Max number of classes per aggr. */
195 struct hlist_head nonfull_aggs; /* Aggs with room for more classes. */
199 * Possible reasons why the timestamps of an aggregate are updated
200 * enqueue: the aggregate switches from idle to active and must scheduled
202 * requeue: the aggregate finishes its budget, so it stops being served and
203 * must be rescheduled for service
205 enum update_reason {enqueue, requeue};
207 static struct qfq_class *qfq_find_class(struct Qdisc *sch, u32 classid)
209 struct qfq_sched *q = qdisc_priv(sch);
210 struct Qdisc_class_common *clc;
212 clc = qdisc_class_find(&q->clhash, classid);
215 return container_of(clc, struct qfq_class, common);
218 static struct netlink_range_validation lmax_range = {
223 static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = {
224 [TCA_QFQ_WEIGHT] = NLA_POLICY_RANGE(NLA_U32, 1, QFQ_MAX_WEIGHT),
225 [TCA_QFQ_LMAX] = NLA_POLICY_FULL_RANGE(NLA_U32, &lmax_range),
229 * Calculate a flow index, given its weight and maximum packet length.
230 * index = log_2(maxlen/weight) but we need to apply the scaling.
231 * This is used only once at flow creation.
233 static int qfq_calc_index(u32 inv_w, unsigned int maxlen, u32 min_slot_shift)
235 u64 slot_size = (u64)maxlen * inv_w;
236 unsigned long size_map;
239 size_map = slot_size >> min_slot_shift;
243 index = __fls(size_map) + 1; /* basically a log_2 */
244 index -= !(slot_size - (1ULL << (index + min_slot_shift - 1)));
249 pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n",
250 (unsigned long) ONE_FP/inv_w, maxlen, index);
255 static void qfq_deactivate_agg(struct qfq_sched *, struct qfq_aggregate *);
256 static void qfq_activate_agg(struct qfq_sched *, struct qfq_aggregate *,
259 static void qfq_init_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
260 u32 lmax, u32 weight)
262 INIT_LIST_HEAD(&agg->active);
263 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
266 agg->class_weight = weight;
269 static struct qfq_aggregate *qfq_find_agg(struct qfq_sched *q,
270 u32 lmax, u32 weight)
272 struct qfq_aggregate *agg;
274 hlist_for_each_entry(agg, &q->nonfull_aggs, nonfull_next)
275 if (agg->lmax == lmax && agg->class_weight == weight)
282 /* Update aggregate as a function of the new number of classes. */
283 static void qfq_update_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
288 if (new_num_classes == q->max_agg_classes)
289 hlist_del_init(&agg->nonfull_next);
291 if (agg->num_classes > new_num_classes &&
292 new_num_classes == q->max_agg_classes - 1) /* agg no more full */
293 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
295 /* The next assignment may let
296 * agg->initial_budget > agg->budgetmax
297 * hold, we will take it into account in charge_actual_service().
299 agg->budgetmax = new_num_classes * agg->lmax;
300 new_agg_weight = agg->class_weight * new_num_classes;
301 agg->inv_w = ONE_FP/new_agg_weight;
303 if (agg->grp == NULL) {
304 int i = qfq_calc_index(agg->inv_w, agg->budgetmax,
306 agg->grp = &q->groups[i];
310 (int) agg->class_weight * (new_num_classes - agg->num_classes);
311 q->iwsum = ONE_FP / q->wsum;
313 agg->num_classes = new_num_classes;
316 /* Add class to aggregate. */
317 static void qfq_add_to_agg(struct qfq_sched *q,
318 struct qfq_aggregate *agg,
319 struct qfq_class *cl)
323 qfq_update_agg(q, agg, agg->num_classes+1);
324 if (cl->qdisc->q.qlen > 0) { /* adding an active class */
325 list_add_tail(&cl->alist, &agg->active);
326 if (list_first_entry(&agg->active, struct qfq_class, alist) ==
327 cl && q->in_serv_agg != agg) /* agg was inactive */
328 qfq_activate_agg(q, agg, enqueue); /* schedule agg */
332 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *);
334 static void qfq_destroy_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
336 hlist_del_init(&agg->nonfull_next);
337 q->wsum -= agg->class_weight;
339 q->iwsum = ONE_FP / q->wsum;
341 if (q->in_serv_agg == agg)
342 q->in_serv_agg = qfq_choose_next_agg(q);
346 /* Deschedule class from within its parent aggregate. */
347 static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
349 struct qfq_aggregate *agg = cl->agg;
352 list_del(&cl->alist); /* remove from RR queue of the aggregate */
353 if (list_empty(&agg->active)) /* agg is now inactive */
354 qfq_deactivate_agg(q, agg);
357 /* Remove class from its parent aggregate. */
358 static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
360 struct qfq_aggregate *agg = cl->agg;
363 if (agg->num_classes == 1) { /* agg being emptied, destroy it */
364 qfq_destroy_agg(q, agg);
367 qfq_update_agg(q, agg, agg->num_classes-1);
370 /* Deschedule class and remove it from its parent aggregate. */
371 static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
373 if (cl->qdisc->q.qlen > 0) /* class is active */
374 qfq_deactivate_class(q, cl);
376 qfq_rm_from_agg(q, cl);
379 /* Move class to a new aggregate, matching the new class weight and/or lmax */
380 static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight,
383 struct qfq_sched *q = qdisc_priv(sch);
384 struct qfq_aggregate *new_agg = qfq_find_agg(q, lmax, weight);
386 if (new_agg == NULL) { /* create new aggregate */
387 new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC);
390 qfq_init_agg(q, new_agg, lmax, weight);
392 qfq_deact_rm_from_agg(q, cl);
393 qfq_add_to_agg(q, new_agg, cl);
398 static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
399 struct nlattr **tca, unsigned long *arg,
400 struct netlink_ext_ack *extack)
402 struct qfq_sched *q = qdisc_priv(sch);
403 struct qfq_class *cl = (struct qfq_class *)*arg;
404 bool existing = false;
405 struct nlattr *tb[TCA_QFQ_MAX + 1];
406 struct qfq_aggregate *new_agg = NULL;
407 u32 weight, lmax, inv_w;
411 if (NL_REQ_ATTR_CHECK(extack, NULL, tca, TCA_OPTIONS)) {
412 NL_SET_ERR_MSG_MOD(extack, "missing options");
416 err = nla_parse_nested_deprecated(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS],
421 if (tb[TCA_QFQ_WEIGHT])
422 weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]);
426 if (tb[TCA_QFQ_LMAX])
427 lmax = nla_get_u32(tb[TCA_QFQ_LMAX]);
429 lmax = psched_mtu(qdisc_dev(sch));
431 inv_w = ONE_FP / weight;
432 weight = ONE_FP / inv_w;
435 lmax == cl->agg->lmax &&
436 weight == cl->agg->class_weight)
437 return 0; /* nothing to change */
439 delta_w = weight - (cl ? cl->agg->class_weight : 0);
441 if (q->wsum + delta_w > QFQ_MAX_WSUM) {
442 NL_SET_ERR_MSG_FMT_MOD(extack,
443 "total weight out of range (%d + %u)\n",
448 if (cl != NULL) { /* modify existing class */
450 err = gen_replace_estimator(&cl->bstats, NULL,
462 /* create and init new class */
463 cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
467 gnet_stats_basic_sync_init(&cl->bstats);
468 cl->common.classid = classid;
471 cl->qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
473 if (cl->qdisc == NULL)
474 cl->qdisc = &noop_qdisc;
477 err = gen_new_estimator(&cl->bstats, NULL,
486 if (cl->qdisc != &noop_qdisc)
487 qdisc_hash_add(cl->qdisc, true);
491 new_agg = qfq_find_agg(q, lmax, weight);
492 if (new_agg == NULL) { /* create new aggregate */
493 sch_tree_unlock(sch);
494 new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL);
495 if (new_agg == NULL) {
497 gen_kill_estimator(&cl->rate_est);
501 qfq_init_agg(q, new_agg, lmax, weight);
504 qfq_deact_rm_from_agg(q, cl);
506 qdisc_class_hash_insert(&q->clhash, &cl->common);
507 qfq_add_to_agg(q, new_agg, cl);
508 sch_tree_unlock(sch);
509 qdisc_class_hash_grow(sch, &q->clhash);
511 *arg = (unsigned long)cl;
515 qdisc_put(cl->qdisc);
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->rate_est);
526 qdisc_put(cl->qdisc);
530 static int qfq_delete_class(struct Qdisc *sch, unsigned long arg,
531 struct netlink_ext_ack *extack)
533 struct qfq_sched *q = qdisc_priv(sch);
534 struct qfq_class *cl = (struct qfq_class *)arg;
536 if (cl->filter_cnt > 0)
541 qdisc_purge_queue(cl->qdisc);
542 qdisc_class_hash_remove(&q->clhash, &cl->common);
544 sch_tree_unlock(sch);
546 qfq_destroy_class(sch, cl);
550 static unsigned long qfq_search_class(struct Qdisc *sch, u32 classid)
552 return (unsigned long)qfq_find_class(sch, classid);
555 static struct tcf_block *qfq_tcf_block(struct Qdisc *sch, unsigned long cl,
556 struct netlink_ext_ack *extack)
558 struct qfq_sched *q = qdisc_priv(sch);
566 static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
569 struct qfq_class *cl = qfq_find_class(sch, classid);
574 return (unsigned long)cl;
577 static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
579 struct qfq_class *cl = (struct qfq_class *)arg;
584 static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
585 struct Qdisc *new, struct Qdisc **old,
586 struct netlink_ext_ack *extack)
588 struct qfq_class *cl = (struct qfq_class *)arg;
591 new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
592 cl->common.classid, NULL);
597 *old = qdisc_replace(sch, new, &cl->qdisc);
601 static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
603 struct qfq_class *cl = (struct qfq_class *)arg;
608 static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
609 struct sk_buff *skb, struct tcmsg *tcm)
611 struct qfq_class *cl = (struct qfq_class *)arg;
614 tcm->tcm_parent = TC_H_ROOT;
615 tcm->tcm_handle = cl->common.classid;
616 tcm->tcm_info = cl->qdisc->handle;
618 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
620 goto nla_put_failure;
621 if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) ||
622 nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax))
623 goto nla_put_failure;
624 return nla_nest_end(skb, nest);
627 nla_nest_cancel(skb, nest);
631 static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
634 struct qfq_class *cl = (struct qfq_class *)arg;
635 struct tc_qfq_stats xstats;
637 memset(&xstats, 0, sizeof(xstats));
639 xstats.weight = cl->agg->class_weight;
640 xstats.lmax = cl->agg->lmax;
642 if (gnet_stats_copy_basic(d, NULL, &cl->bstats, true) < 0 ||
643 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
644 qdisc_qstats_copy(d, cl->qdisc) < 0)
647 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
650 static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
652 struct qfq_sched *q = qdisc_priv(sch);
653 struct qfq_class *cl;
659 for (i = 0; i < q->clhash.hashsize; i++) {
660 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
661 if (!tc_qdisc_stats_dump(sch, (unsigned long)cl, arg))
667 static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
670 struct qfq_sched *q = qdisc_priv(sch);
671 struct qfq_class *cl;
672 struct tcf_result res;
673 struct tcf_proto *fl;
676 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
677 pr_debug("qfq_classify: found %d\n", skb->priority);
678 cl = qfq_find_class(sch, skb->priority);
683 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
684 fl = rcu_dereference_bh(q->filter_list);
685 result = tcf_classify(skb, NULL, fl, &res, false);
687 #ifdef CONFIG_NET_CLS_ACT
692 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
698 cl = (struct qfq_class *)res.class;
700 cl = qfq_find_class(sch, res.classid);
707 /* Generic comparison function, handling wraparound. */
708 static inline int qfq_gt(u64 a, u64 b)
710 return (s64)(a - b) > 0;
713 /* Round a precise timestamp to its slotted value. */
714 static inline u64 qfq_round_down(u64 ts, unsigned int shift)
716 return ts & ~((1ULL << shift) - 1);
719 /* return the pointer to the group with lowest index in the bitmap */
720 static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
721 unsigned long bitmap)
723 int index = __ffs(bitmap);
724 return &q->groups[index];
726 /* Calculate a mask to mimic what would be ffs_from(). */
727 static inline unsigned long mask_from(unsigned long bitmap, int from)
729 return bitmap & ~((1UL << from) - 1);
733 * The state computation relies on ER=0, IR=1, EB=2, IB=3
734 * First compute eligibility comparing grp->S, q->V,
735 * then check if someone is blocking us and possibly add EB
737 static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
739 /* if S > V we are not eligible */
740 unsigned int state = qfq_gt(grp->S, q->V);
741 unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
742 struct qfq_group *next;
745 next = qfq_ffs(q, mask);
746 if (qfq_gt(grp->F, next->F))
756 * q->bitmaps[dst] |= q->bitmaps[src] & mask;
757 * q->bitmaps[src] &= ~mask;
758 * but we should make sure that src != dst
760 static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
763 q->bitmaps[dst] |= q->bitmaps[src] & mask;
764 q->bitmaps[src] &= ~mask;
767 static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
769 unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
770 struct qfq_group *next;
773 next = qfq_ffs(q, mask);
774 if (!qfq_gt(next->F, old_F))
778 mask = (1UL << index) - 1;
779 qfq_move_groups(q, mask, EB, ER);
780 qfq_move_groups(q, mask, IB, IR);
787 old_V >>= q->min_slot_shift;
793 static void qfq_make_eligible(struct qfq_sched *q)
795 unsigned long vslot = q->V >> q->min_slot_shift;
796 unsigned long old_vslot = q->oldV >> q->min_slot_shift;
798 if (vslot != old_vslot) {
800 int last_flip_pos = fls(vslot ^ old_vslot);
802 if (last_flip_pos > 31) /* higher than the number of groups */
803 mask = ~0UL; /* make all groups eligible */
805 mask = (1UL << last_flip_pos) - 1;
807 qfq_move_groups(q, mask, IR, ER);
808 qfq_move_groups(q, mask, IB, EB);
813 * The index of the slot in which the input aggregate agg is to be
814 * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2'
815 * and not a '-1' because the start time of the group may be moved
816 * backward by one slot after the aggregate has been inserted, and
817 * this would cause non-empty slots to be right-shifted by one
820 * QFQ+ fully satisfies this bound to the slot index if the parameters
821 * of the classes are not changed dynamically, and if QFQ+ never
822 * happens to postpone the service of agg unjustly, i.e., it never
823 * happens that the aggregate becomes backlogged and eligible, or just
824 * eligible, while an aggregate with a higher approximated finish time
825 * is being served. In particular, in this case QFQ+ guarantees that
826 * the timestamps of agg are low enough that the slot index is never
827 * higher than 2. Unfortunately, QFQ+ cannot provide the same
828 * guarantee if it happens to unjustly postpone the service of agg, or
829 * if the parameters of some class are changed.
831 * As for the first event, i.e., an out-of-order service, the
832 * upper bound to the slot index guaranteed by QFQ+ grows to
834 * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
835 * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1.
837 * The following function deals with this problem by backward-shifting
838 * the timestamps of agg, if needed, so as to guarantee that the slot
839 * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may
840 * cause the service of other aggregates to be postponed, yet the
841 * worst-case guarantees of these aggregates are not violated. In
842 * fact, in case of no out-of-order service, the timestamps of agg
843 * would have been even lower than they are after the backward shift,
844 * because QFQ+ would have guaranteed a maximum value equal to 2 for
845 * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose
846 * service is postponed because of the backward-shift would have
847 * however waited for the service of agg before being served.
849 * The other event that may cause the slot index to be higher than 2
850 * for agg is a recent change of the parameters of some class. If the
851 * weight of a class is increased or the lmax (max_pkt_size) of the
852 * class is decreased, then a new aggregate with smaller slot size
853 * than the original parent aggregate of the class may happen to be
854 * activated. The activation of this aggregate should be properly
855 * delayed to when the service of the class has finished in the ideal
856 * system tracked by QFQ+. If the activation of the aggregate is not
857 * delayed to this reference time instant, then this aggregate may be
858 * unjustly served before other aggregates waiting for service. This
859 * may cause the above bound to the slot index to be violated for some
860 * of these unlucky aggregates.
862 * Instead of delaying the activation of the new aggregate, which is
863 * quite complex, the above-discussed capping of the slot index is
864 * used to handle also the consequences of a change of the parameters
867 static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg,
870 u64 slot = (roundedS - grp->S) >> grp->slot_shift;
871 unsigned int i; /* slot index in the bucket list */
873 if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
874 u64 deltaS = roundedS - grp->S -
875 ((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
878 slot = QFQ_MAX_SLOTS - 2;
881 i = (grp->front + slot) % QFQ_MAX_SLOTS;
883 hlist_add_head(&agg->next, &grp->slots[i]);
884 __set_bit(slot, &grp->full_slots);
887 /* Maybe introduce hlist_first_entry?? */
888 static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp)
890 return hlist_entry(grp->slots[grp->front].first,
891 struct qfq_aggregate, next);
895 * remove the entry from the slot
897 static void qfq_front_slot_remove(struct qfq_group *grp)
899 struct qfq_aggregate *agg = qfq_slot_head(grp);
902 hlist_del(&agg->next);
903 if (hlist_empty(&grp->slots[grp->front]))
904 __clear_bit(0, &grp->full_slots);
908 * Returns the first aggregate in the first non-empty bucket of the
909 * group. As a side effect, adjusts the bucket list so the first
910 * non-empty bucket is at position 0 in full_slots.
912 static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp)
916 pr_debug("qfq slot_scan: grp %u full %#lx\n",
917 grp->index, grp->full_slots);
919 if (grp->full_slots == 0)
922 i = __ffs(grp->full_slots); /* zero based */
924 grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
925 grp->full_slots >>= i;
928 return qfq_slot_head(grp);
932 * adjust the bucket list. When the start time of a group decreases,
933 * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
934 * move the objects. The mask of occupied slots must be shifted
935 * because we use ffs() to find the first non-empty slot.
936 * This covers decreases in the group's start time, but what about
937 * increases of the start time ?
938 * Here too we should make sure that i is less than 32
940 static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
942 unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
944 grp->full_slots <<= i;
945 grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
948 static void qfq_update_eligible(struct qfq_sched *q)
950 struct qfq_group *grp;
951 unsigned long ineligible;
953 ineligible = q->bitmaps[IR] | q->bitmaps[IB];
955 if (!q->bitmaps[ER]) {
956 grp = qfq_ffs(q, ineligible);
957 if (qfq_gt(grp->S, q->V))
960 qfq_make_eligible(q);
964 /* Dequeue head packet of the head class in the DRR queue of the aggregate. */
965 static void agg_dequeue(struct qfq_aggregate *agg,
966 struct qfq_class *cl, unsigned int len)
968 qdisc_dequeue_peeked(cl->qdisc);
970 cl->deficit -= (int) len;
972 if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */
973 list_del(&cl->alist);
974 else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) {
975 cl->deficit += agg->lmax;
976 list_move_tail(&cl->alist, &agg->active);
980 static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg,
981 struct qfq_class **cl,
986 *cl = list_first_entry(&agg->active, struct qfq_class, alist);
987 skb = (*cl)->qdisc->ops->peek((*cl)->qdisc);
989 WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
991 *len = qdisc_pkt_len(skb);
996 /* Update F according to the actual service received by the aggregate. */
997 static inline void charge_actual_service(struct qfq_aggregate *agg)
999 /* Compute the service received by the aggregate, taking into
1000 * account that, after decreasing the number of classes in
1001 * agg, it may happen that
1002 * agg->initial_budget - agg->budget > agg->bugdetmax
1004 u32 service_received = min(agg->budgetmax,
1005 agg->initial_budget - agg->budget);
1007 agg->F = agg->S + (u64)service_received * agg->inv_w;
1010 /* Assign a reasonable start time for a new aggregate in group i.
1011 * Admissible values for \hat(F) are multiples of \sigma_i
1012 * no greater than V+\sigma_i . Larger values mean that
1013 * we had a wraparound so we consider the timestamp to be stale.
1015 * If F is not stale and F >= V then we set S = F.
1016 * Otherwise we should assign S = V, but this may violate
1017 * the ordering in EB (see [2]). So, if we have groups in ER,
1018 * set S to the F_j of the first group j which would be blocking us.
1019 * We are guaranteed not to move S backward because
1020 * otherwise our group i would still be blocked.
1022 static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg)
1025 u64 limit, roundedF;
1026 int slot_shift = agg->grp->slot_shift;
1028 roundedF = qfq_round_down(agg->F, slot_shift);
1029 limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
1031 if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) {
1032 /* timestamp was stale */
1033 mask = mask_from(q->bitmaps[ER], agg->grp->index);
1035 struct qfq_group *next = qfq_ffs(q, mask);
1036 if (qfq_gt(roundedF, next->F)) {
1037 if (qfq_gt(limit, next->F))
1039 else /* preserve timestamp correctness */
1045 } else /* timestamp is not stale */
1049 /* Update the timestamps of agg before scheduling/rescheduling it for
1050 * service. In particular, assign to agg->F its maximum possible
1051 * value, i.e., the virtual finish time with which the aggregate
1052 * should be labeled if it used all its budget once in service.
1055 qfq_update_agg_ts(struct qfq_sched *q,
1056 struct qfq_aggregate *agg, enum update_reason reason)
1058 if (reason != requeue)
1059 qfq_update_start(q, agg);
1060 else /* just charge agg for the service received */
1063 agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w;
1066 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg);
1068 static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
1070 struct qfq_sched *q = qdisc_priv(sch);
1071 struct qfq_aggregate *in_serv_agg = q->in_serv_agg;
1072 struct qfq_class *cl;
1073 struct sk_buff *skb = NULL;
1074 /* next-packet len, 0 means no more active classes in in-service agg */
1075 unsigned int len = 0;
1077 if (in_serv_agg == NULL)
1080 if (!list_empty(&in_serv_agg->active))
1081 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1084 * If there are no active classes in the in-service aggregate,
1085 * or if the aggregate has not enough budget to serve its next
1086 * class, then choose the next aggregate to serve.
1088 if (len == 0 || in_serv_agg->budget < len) {
1089 charge_actual_service(in_serv_agg);
1091 /* recharge the budget of the aggregate */
1092 in_serv_agg->initial_budget = in_serv_agg->budget =
1093 in_serv_agg->budgetmax;
1095 if (!list_empty(&in_serv_agg->active)) {
1097 * Still active: reschedule for
1098 * service. Possible optimization: if no other
1099 * aggregate is active, then there is no point
1100 * in rescheduling this aggregate, and we can
1101 * just keep it as the in-service one. This
1102 * should be however a corner case, and to
1103 * handle it, we would need to maintain an
1104 * extra num_active_aggs field.
1106 qfq_update_agg_ts(q, in_serv_agg, requeue);
1107 qfq_schedule_agg(q, in_serv_agg);
1108 } else if (sch->q.qlen == 0) { /* no aggregate to serve */
1109 q->in_serv_agg = NULL;
1114 * If we get here, there are other aggregates queued:
1115 * choose the new aggregate to serve.
1117 in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
1118 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1123 qdisc_qstats_backlog_dec(sch, skb);
1125 qdisc_bstats_update(sch, skb);
1127 agg_dequeue(in_serv_agg, cl, len);
1128 /* If lmax is lowered, through qfq_change_class, for a class
1129 * owning pending packets with larger size than the new value
1130 * of lmax, then the following condition may hold.
1132 if (unlikely(in_serv_agg->budget < len))
1133 in_serv_agg->budget = 0;
1135 in_serv_agg->budget -= len;
1137 q->V += (u64)len * q->iwsum;
1138 pr_debug("qfq dequeue: len %u F %lld now %lld\n",
1139 len, (unsigned long long) in_serv_agg->F,
1140 (unsigned long long) q->V);
1145 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)
1147 struct qfq_group *grp;
1148 struct qfq_aggregate *agg, *new_front_agg;
1151 qfq_update_eligible(q);
1154 if (!q->bitmaps[ER])
1157 grp = qfq_ffs(q, q->bitmaps[ER]);
1160 agg = qfq_slot_head(grp);
1162 /* agg starts to be served, remove it from schedule */
1163 qfq_front_slot_remove(grp);
1165 new_front_agg = qfq_slot_scan(grp);
1167 if (new_front_agg == NULL) /* group is now inactive, remove from ER */
1168 __clear_bit(grp->index, &q->bitmaps[ER]);
1170 u64 roundedS = qfq_round_down(new_front_agg->S,
1174 if (grp->S == roundedS)
1177 grp->F = roundedS + (2ULL << grp->slot_shift);
1178 __clear_bit(grp->index, &q->bitmaps[ER]);
1179 s = qfq_calc_state(q, grp);
1180 __set_bit(grp->index, &q->bitmaps[s]);
1183 qfq_unblock_groups(q, grp->index, old_F);
1188 static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
1189 struct sk_buff **to_free)
1191 unsigned int len = qdisc_pkt_len(skb), gso_segs;
1192 struct qfq_sched *q = qdisc_priv(sch);
1193 struct qfq_class *cl;
1194 struct qfq_aggregate *agg;
1198 cl = qfq_classify(skb, sch, &err);
1200 if (err & __NET_XMIT_BYPASS)
1201 qdisc_qstats_drop(sch);
1202 __qdisc_drop(skb, to_free);
1205 pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
1207 if (unlikely(cl->agg->lmax < len)) {
1208 pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
1209 cl->agg->lmax, len, cl->common.classid);
1210 err = qfq_change_agg(sch, cl, cl->agg->class_weight, len);
1213 return qdisc_drop(skb, sch, to_free);
1217 gso_segs = skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 1;
1218 first = !cl->qdisc->q.qlen;
1219 err = qdisc_enqueue(skb, cl->qdisc, to_free);
1220 if (unlikely(err != NET_XMIT_SUCCESS)) {
1221 pr_debug("qfq_enqueue: enqueue failed %d\n", err);
1222 if (net_xmit_drop_count(err)) {
1224 qdisc_qstats_drop(sch);
1229 _bstats_update(&cl->bstats, len, gso_segs);
1230 sch->qstats.backlog += len;
1234 /* if the queue was not empty, then done here */
1236 if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) &&
1237 list_first_entry(&agg->active, struct qfq_class, alist)
1238 == cl && cl->deficit < len)
1239 list_move_tail(&cl->alist, &agg->active);
1244 /* schedule class for service within the aggregate */
1245 cl->deficit = agg->lmax;
1246 list_add_tail(&cl->alist, &agg->active);
1248 if (list_first_entry(&agg->active, struct qfq_class, alist) != cl ||
1249 q->in_serv_agg == agg)
1250 return err; /* non-empty or in service, nothing else to do */
1252 qfq_activate_agg(q, agg, enqueue);
1258 * Schedule aggregate according to its timestamps.
1260 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1262 struct qfq_group *grp = agg->grp;
1266 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1269 * Insert agg in the correct bucket.
1270 * If agg->S >= grp->S we don't need to adjust the
1271 * bucket list and simply go to the insertion phase.
1272 * Otherwise grp->S is decreasing, we must make room
1273 * in the bucket list, and also recompute the group state.
1274 * Finally, if there were no flows in this group and nobody
1275 * was in ER make sure to adjust V.
1277 if (grp->full_slots) {
1278 if (!qfq_gt(grp->S, agg->S))
1281 /* create a slot for this agg->S */
1282 qfq_slot_rotate(grp, roundedS);
1283 /* group was surely ineligible, remove */
1284 __clear_bit(grp->index, &q->bitmaps[IR]);
1285 __clear_bit(grp->index, &q->bitmaps[IB]);
1286 } else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) &&
1287 q->in_serv_agg == NULL)
1291 grp->F = roundedS + (2ULL << grp->slot_shift);
1292 s = qfq_calc_state(q, grp);
1293 __set_bit(grp->index, &q->bitmaps[s]);
1295 pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
1297 (unsigned long long) agg->S,
1298 (unsigned long long) agg->F,
1299 (unsigned long long) q->V);
1302 qfq_slot_insert(grp, agg, roundedS);
1306 /* Update agg ts and schedule agg for service */
1307 static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
1308 enum update_reason reason)
1310 agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */
1312 qfq_update_agg_ts(q, agg, reason);
1313 if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */
1314 q->in_serv_agg = agg; /* start serving this aggregate */
1315 /* update V: to be in service, agg must be eligible */
1316 q->oldV = q->V = agg->S;
1317 } else if (agg != q->in_serv_agg)
1318 qfq_schedule_agg(q, agg);
1321 static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
1322 struct qfq_aggregate *agg)
1324 unsigned int i, offset;
1327 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1328 offset = (roundedS - grp->S) >> grp->slot_shift;
1330 i = (grp->front + offset) % QFQ_MAX_SLOTS;
1332 hlist_del(&agg->next);
1333 if (hlist_empty(&grp->slots[i]))
1334 __clear_bit(offset, &grp->full_slots);
1338 * Called to forcibly deschedule an aggregate. If the aggregate is
1339 * not in the front bucket, or if the latter has other aggregates in
1340 * the front bucket, we can simply remove the aggregate with no other
1342 * Otherwise we must propagate the event up.
1344 static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1346 struct qfq_group *grp = agg->grp;
1351 if (agg == q->in_serv_agg) {
1352 charge_actual_service(agg);
1353 q->in_serv_agg = qfq_choose_next_agg(q);
1358 qfq_slot_remove(q, grp, agg);
1360 if (!grp->full_slots) {
1361 __clear_bit(grp->index, &q->bitmaps[IR]);
1362 __clear_bit(grp->index, &q->bitmaps[EB]);
1363 __clear_bit(grp->index, &q->bitmaps[IB]);
1365 if (test_bit(grp->index, &q->bitmaps[ER]) &&
1366 !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
1367 mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
1369 mask = ~((1UL << __fls(mask)) - 1);
1372 qfq_move_groups(q, mask, EB, ER);
1373 qfq_move_groups(q, mask, IB, IR);
1375 __clear_bit(grp->index, &q->bitmaps[ER]);
1376 } else if (hlist_empty(&grp->slots[grp->front])) {
1377 agg = qfq_slot_scan(grp);
1378 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1379 if (grp->S != roundedS) {
1380 __clear_bit(grp->index, &q->bitmaps[ER]);
1381 __clear_bit(grp->index, &q->bitmaps[IR]);
1382 __clear_bit(grp->index, &q->bitmaps[EB]);
1383 __clear_bit(grp->index, &q->bitmaps[IB]);
1385 grp->F = roundedS + (2ULL << grp->slot_shift);
1386 s = qfq_calc_state(q, grp);
1387 __set_bit(grp->index, &q->bitmaps[s]);
1392 static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1394 struct qfq_sched *q = qdisc_priv(sch);
1395 struct qfq_class *cl = (struct qfq_class *)arg;
1397 qfq_deactivate_class(q, cl);
1400 static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt,
1401 struct netlink_ext_ack *extack)
1403 struct qfq_sched *q = qdisc_priv(sch);
1404 struct qfq_group *grp;
1406 u32 max_cl_shift, maxbudg_shift, max_classes;
1408 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
1412 err = qdisc_class_hash_init(&q->clhash);
1416 max_classes = min_t(u64, (u64)qdisc_dev(sch)->tx_queue_len + 1,
1417 QFQ_MAX_AGG_CLASSES);
1418 /* max_cl_shift = floor(log_2(max_classes)) */
1419 max_cl_shift = __fls(max_classes);
1420 q->max_agg_classes = 1<<max_cl_shift;
1422 /* maxbudg_shift = log2(max_len * max_classes_per_agg) */
1423 maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift;
1424 q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX;
1426 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1427 grp = &q->groups[i];
1429 grp->slot_shift = q->min_slot_shift + i;
1430 for (j = 0; j < QFQ_MAX_SLOTS; j++)
1431 INIT_HLIST_HEAD(&grp->slots[j]);
1434 INIT_HLIST_HEAD(&q->nonfull_aggs);
1439 static void qfq_reset_qdisc(struct Qdisc *sch)
1441 struct qfq_sched *q = qdisc_priv(sch);
1442 struct qfq_class *cl;
1445 for (i = 0; i < q->clhash.hashsize; i++) {
1446 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1447 if (cl->qdisc->q.qlen > 0)
1448 qfq_deactivate_class(q, cl);
1450 qdisc_reset(cl->qdisc);
1455 static void qfq_destroy_qdisc(struct Qdisc *sch)
1457 struct qfq_sched *q = qdisc_priv(sch);
1458 struct qfq_class *cl;
1459 struct hlist_node *next;
1462 tcf_block_put(q->block);
1464 for (i = 0; i < q->clhash.hashsize; i++) {
1465 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1467 qfq_destroy_class(sch, cl);
1470 qdisc_class_hash_destroy(&q->clhash);
1473 static const struct Qdisc_class_ops qfq_class_ops = {
1474 .change = qfq_change_class,
1475 .delete = qfq_delete_class,
1476 .find = qfq_search_class,
1477 .tcf_block = qfq_tcf_block,
1478 .bind_tcf = qfq_bind_tcf,
1479 .unbind_tcf = qfq_unbind_tcf,
1480 .graft = qfq_graft_class,
1481 .leaf = qfq_class_leaf,
1482 .qlen_notify = qfq_qlen_notify,
1483 .dump = qfq_dump_class,
1484 .dump_stats = qfq_dump_class_stats,
1488 static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1489 .cl_ops = &qfq_class_ops,
1491 .priv_size = sizeof(struct qfq_sched),
1492 .enqueue = qfq_enqueue,
1493 .dequeue = qfq_dequeue,
1494 .peek = qdisc_peek_dequeued,
1495 .init = qfq_init_qdisc,
1496 .reset = qfq_reset_qdisc,
1497 .destroy = qfq_destroy_qdisc,
1498 .owner = THIS_MODULE,
1501 static int __init qfq_init(void)
1503 return register_qdisc(&qfq_qdisc_ops);
1506 static void __exit qfq_exit(void)
1508 unregister_qdisc(&qfq_qdisc_ops);
1511 module_init(qfq_init);
1512 module_exit(qfq_exit);
1513 MODULE_LICENSE("GPL");