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 */
117 #define QFQ_MAX_AGG_CLASSES 8 /* max num classes per aggregate allowed */
120 * Possible group states. These values are used as indexes for the bitmaps
121 * array of struct qfq_queue.
123 enum qfq_state { ER, IR, EB, IB, QFQ_MAX_STATE };
127 struct qfq_aggregate;
130 struct Qdisc_class_common common;
132 unsigned int filter_cnt;
134 struct gnet_stats_basic_packed bstats;
135 struct gnet_stats_queue qstats;
136 struct net_rate_estimator __rcu *rate_est;
138 struct list_head alist; /* Link for active-classes list. */
139 struct qfq_aggregate *agg; /* Parent aggregate. */
140 int deficit; /* DRR deficit counter. */
143 struct qfq_aggregate {
144 struct hlist_node next; /* Link for the slot list. */
145 u64 S, F; /* flow timestamps (exact) */
147 /* group we belong to. In principle we would need the index,
148 * which is log_2(lmax/weight), but we never reference it
149 * directly, only the group.
151 struct qfq_group *grp;
153 /* these are copied from the flowset. */
154 u32 class_weight; /* Weight of each class in this aggregate. */
155 /* Max pkt size for the classes in this aggregate, DRR quantum. */
158 u32 inv_w; /* ONE_FP/(sum of weights of classes in aggr.). */
159 u32 budgetmax; /* Max budget for this aggregate. */
160 u32 initial_budget, budget; /* Initial and current budget. */
162 int num_classes; /* Number of classes in this aggr. */
163 struct list_head active; /* DRR queue of active classes. */
165 struct hlist_node nonfull_next; /* See nonfull_aggs in qfq_sched. */
169 u64 S, F; /* group timestamps (approx). */
170 unsigned int slot_shift; /* Slot shift. */
171 unsigned int index; /* Group index. */
172 unsigned int front; /* Index of the front slot. */
173 unsigned long full_slots; /* non-empty slots */
175 /* Array of RR lists of active aggregates. */
176 struct hlist_head slots[QFQ_MAX_SLOTS];
180 struct tcf_proto __rcu *filter_list;
181 struct tcf_block *block;
182 struct Qdisc_class_hash clhash;
184 u64 oldV, V; /* Precise virtual times. */
185 struct qfq_aggregate *in_serv_agg; /* Aggregate being served. */
186 u32 wsum; /* weight sum */
187 u32 iwsum; /* inverse weight sum */
189 unsigned long bitmaps[QFQ_MAX_STATE]; /* Group bitmaps. */
190 struct qfq_group groups[QFQ_MAX_INDEX + 1]; /* The groups. */
191 u32 min_slot_shift; /* Index of the group-0 bit in the bitmaps. */
193 u32 max_agg_classes; /* Max number of classes per aggr. */
194 struct hlist_head nonfull_aggs; /* Aggs with room for more classes. */
198 * Possible reasons why the timestamps of an aggregate are updated
199 * enqueue: the aggregate switches from idle to active and must scheduled
201 * requeue: the aggregate finishes its budget, so it stops being served and
202 * must be rescheduled for service
204 enum update_reason {enqueue, requeue};
206 static struct qfq_class *qfq_find_class(struct Qdisc *sch, u32 classid)
208 struct qfq_sched *q = qdisc_priv(sch);
209 struct Qdisc_class_common *clc;
211 clc = qdisc_class_find(&q->clhash, classid);
214 return container_of(clc, struct qfq_class, common);
217 static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = {
218 [TCA_QFQ_WEIGHT] = { .type = NLA_U32 },
219 [TCA_QFQ_LMAX] = { .type = NLA_U32 },
223 * Calculate a flow index, given its weight and maximum packet length.
224 * index = log_2(maxlen/weight) but we need to apply the scaling.
225 * This is used only once at flow creation.
227 static int qfq_calc_index(u32 inv_w, unsigned int maxlen, u32 min_slot_shift)
229 u64 slot_size = (u64)maxlen * inv_w;
230 unsigned long size_map;
233 size_map = slot_size >> min_slot_shift;
237 index = __fls(size_map) + 1; /* basically a log_2 */
238 index -= !(slot_size - (1ULL << (index + min_slot_shift - 1)));
243 pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n",
244 (unsigned long) ONE_FP/inv_w, maxlen, index);
249 static void qfq_deactivate_agg(struct qfq_sched *, struct qfq_aggregate *);
250 static void qfq_activate_agg(struct qfq_sched *, struct qfq_aggregate *,
253 static void qfq_init_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
254 u32 lmax, u32 weight)
256 INIT_LIST_HEAD(&agg->active);
257 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
260 agg->class_weight = weight;
263 static struct qfq_aggregate *qfq_find_agg(struct qfq_sched *q,
264 u32 lmax, u32 weight)
266 struct qfq_aggregate *agg;
268 hlist_for_each_entry(agg, &q->nonfull_aggs, nonfull_next)
269 if (agg->lmax == lmax && agg->class_weight == weight)
276 /* Update aggregate as a function of the new number of classes. */
277 static void qfq_update_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
282 if (new_num_classes == q->max_agg_classes)
283 hlist_del_init(&agg->nonfull_next);
285 if (agg->num_classes > new_num_classes &&
286 new_num_classes == q->max_agg_classes - 1) /* agg no more full */
287 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
289 /* The next assignment may let
290 * agg->initial_budget > agg->budgetmax
291 * hold, we will take it into account in charge_actual_service().
293 agg->budgetmax = new_num_classes * agg->lmax;
294 new_agg_weight = agg->class_weight * new_num_classes;
295 agg->inv_w = ONE_FP/new_agg_weight;
297 if (agg->grp == NULL) {
298 int i = qfq_calc_index(agg->inv_w, agg->budgetmax,
300 agg->grp = &q->groups[i];
304 (int) agg->class_weight * (new_num_classes - agg->num_classes);
305 q->iwsum = ONE_FP / q->wsum;
307 agg->num_classes = new_num_classes;
310 /* Add class to aggregate. */
311 static void qfq_add_to_agg(struct qfq_sched *q,
312 struct qfq_aggregate *agg,
313 struct qfq_class *cl)
317 qfq_update_agg(q, agg, agg->num_classes+1);
318 if (cl->qdisc->q.qlen > 0) { /* adding an active class */
319 list_add_tail(&cl->alist, &agg->active);
320 if (list_first_entry(&agg->active, struct qfq_class, alist) ==
321 cl && q->in_serv_agg != agg) /* agg was inactive */
322 qfq_activate_agg(q, agg, enqueue); /* schedule agg */
326 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *);
328 static void qfq_destroy_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
330 hlist_del_init(&agg->nonfull_next);
331 q->wsum -= agg->class_weight;
333 q->iwsum = ONE_FP / q->wsum;
335 if (q->in_serv_agg == agg)
336 q->in_serv_agg = qfq_choose_next_agg(q);
340 /* Deschedule class from within its parent aggregate. */
341 static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
343 struct qfq_aggregate *agg = cl->agg;
346 list_del(&cl->alist); /* remove from RR queue of the aggregate */
347 if (list_empty(&agg->active)) /* agg is now inactive */
348 qfq_deactivate_agg(q, agg);
351 /* Remove class from its parent aggregate. */
352 static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
354 struct qfq_aggregate *agg = cl->agg;
357 if (agg->num_classes == 1) { /* agg being emptied, destroy it */
358 qfq_destroy_agg(q, agg);
361 qfq_update_agg(q, agg, agg->num_classes-1);
364 /* Deschedule class and remove it from its parent aggregate. */
365 static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
367 if (cl->qdisc->q.qlen > 0) /* class is active */
368 qfq_deactivate_class(q, cl);
370 qfq_rm_from_agg(q, cl);
373 /* Move class to a new aggregate, matching the new class weight and/or lmax */
374 static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight,
377 struct qfq_sched *q = qdisc_priv(sch);
378 struct qfq_aggregate *new_agg = qfq_find_agg(q, lmax, weight);
380 if (new_agg == NULL) { /* create new aggregate */
381 new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC);
384 qfq_init_agg(q, new_agg, lmax, weight);
386 qfq_deact_rm_from_agg(q, cl);
387 qfq_add_to_agg(q, new_agg, cl);
392 static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
393 struct nlattr **tca, unsigned long *arg,
394 struct netlink_ext_ack *extack)
396 struct qfq_sched *q = qdisc_priv(sch);
397 struct qfq_class *cl = (struct qfq_class *)*arg;
398 bool existing = false;
399 struct nlattr *tb[TCA_QFQ_MAX + 1];
400 struct qfq_aggregate *new_agg = NULL;
401 u32 weight, lmax, inv_w;
405 if (tca[TCA_OPTIONS] == NULL) {
406 pr_notice("qfq: no options\n");
410 err = nla_parse_nested_deprecated(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS],
415 if (tb[TCA_QFQ_WEIGHT]) {
416 weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]);
417 if (!weight || weight > (1UL << QFQ_MAX_WSHIFT)) {
418 pr_notice("qfq: invalid weight %u\n", weight);
424 if (tb[TCA_QFQ_LMAX]) {
425 lmax = nla_get_u32(tb[TCA_QFQ_LMAX]);
426 if (lmax < QFQ_MIN_LMAX || lmax > (1UL << QFQ_MTU_SHIFT)) {
427 pr_notice("qfq: invalid max length %u\n", lmax);
431 lmax = psched_mtu(qdisc_dev(sch));
433 inv_w = ONE_FP / weight;
434 weight = ONE_FP / inv_w;
437 lmax == cl->agg->lmax &&
438 weight == cl->agg->class_weight)
439 return 0; /* nothing to change */
441 delta_w = weight - (cl ? cl->agg->class_weight : 0);
443 if (q->wsum + delta_w > QFQ_MAX_WSUM) {
444 pr_notice("qfq: total weight out of range (%d + %u)\n",
449 if (cl != NULL) { /* modify existing class */
451 err = gen_replace_estimator(&cl->bstats, NULL,
454 qdisc_root_sleeping_running(sch),
463 /* create and init new class */
464 cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
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,
480 qdisc_root_sleeping_running(sch),
486 if (cl->qdisc != &noop_qdisc)
487 qdisc_hash_add(cl->qdisc, true);
489 qdisc_class_hash_insert(&q->clhash, &cl->common);
490 sch_tree_unlock(sch);
492 qdisc_class_hash_grow(sch, &q->clhash);
496 new_agg = qfq_find_agg(q, lmax, weight);
497 if (new_agg == NULL) { /* create new aggregate */
498 sch_tree_unlock(sch);
499 new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL);
500 if (new_agg == NULL) {
502 gen_kill_estimator(&cl->rate_est);
506 qfq_init_agg(q, new_agg, lmax, weight);
509 qfq_deact_rm_from_agg(q, cl);
510 qfq_add_to_agg(q, new_agg, cl);
511 sch_tree_unlock(sch);
513 *arg = (unsigned long)cl;
517 qdisc_put(cl->qdisc);
522 static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
524 struct qfq_sched *q = qdisc_priv(sch);
526 qfq_rm_from_agg(q, cl);
527 gen_kill_estimator(&cl->rate_est);
528 qdisc_put(cl->qdisc);
532 static int qfq_delete_class(struct Qdisc *sch, unsigned long arg,
533 struct netlink_ext_ack *extack)
535 struct qfq_sched *q = qdisc_priv(sch);
536 struct qfq_class *cl = (struct qfq_class *)arg;
538 if (cl->filter_cnt > 0)
543 qdisc_purge_queue(cl->qdisc);
544 qdisc_class_hash_remove(&q->clhash, &cl->common);
546 sch_tree_unlock(sch);
548 qfq_destroy_class(sch, cl);
552 static unsigned long qfq_search_class(struct Qdisc *sch, u32 classid)
554 return (unsigned long)qfq_find_class(sch, classid);
557 static struct tcf_block *qfq_tcf_block(struct Qdisc *sch, unsigned long cl,
558 struct netlink_ext_ack *extack)
560 struct qfq_sched *q = qdisc_priv(sch);
568 static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
571 struct qfq_class *cl = qfq_find_class(sch, classid);
576 return (unsigned long)cl;
579 static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
581 struct qfq_class *cl = (struct qfq_class *)arg;
586 static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
587 struct Qdisc *new, struct Qdisc **old,
588 struct netlink_ext_ack *extack)
590 struct qfq_class *cl = (struct qfq_class *)arg;
593 new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
594 cl->common.classid, NULL);
599 *old = qdisc_replace(sch, new, &cl->qdisc);
603 static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
605 struct qfq_class *cl = (struct qfq_class *)arg;
610 static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
611 struct sk_buff *skb, struct tcmsg *tcm)
613 struct qfq_class *cl = (struct qfq_class *)arg;
616 tcm->tcm_parent = TC_H_ROOT;
617 tcm->tcm_handle = cl->common.classid;
618 tcm->tcm_info = cl->qdisc->handle;
620 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
622 goto nla_put_failure;
623 if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) ||
624 nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax))
625 goto nla_put_failure;
626 return nla_nest_end(skb, nest);
629 nla_nest_cancel(skb, nest);
633 static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
636 struct qfq_class *cl = (struct qfq_class *)arg;
637 struct tc_qfq_stats xstats;
639 memset(&xstats, 0, sizeof(xstats));
641 xstats.weight = cl->agg->class_weight;
642 xstats.lmax = cl->agg->lmax;
644 if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
645 d, NULL, &cl->bstats) < 0 ||
646 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
647 qdisc_qstats_copy(d, cl->qdisc) < 0)
650 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
653 static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
655 struct qfq_sched *q = qdisc_priv(sch);
656 struct qfq_class *cl;
662 for (i = 0; i < q->clhash.hashsize; i++) {
663 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
664 if (arg->count < arg->skip) {
668 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
677 static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
680 struct qfq_sched *q = qdisc_priv(sch);
681 struct qfq_class *cl;
682 struct tcf_result res;
683 struct tcf_proto *fl;
686 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
687 pr_debug("qfq_classify: found %d\n", skb->priority);
688 cl = qfq_find_class(sch, skb->priority);
693 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
694 fl = rcu_dereference_bh(q->filter_list);
695 result = tcf_classify(skb, fl, &res, false);
697 #ifdef CONFIG_NET_CLS_ACT
702 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
708 cl = (struct qfq_class *)res.class;
710 cl = qfq_find_class(sch, res.classid);
717 /* Generic comparison function, handling wraparound. */
718 static inline int qfq_gt(u64 a, u64 b)
720 return (s64)(a - b) > 0;
723 /* Round a precise timestamp to its slotted value. */
724 static inline u64 qfq_round_down(u64 ts, unsigned int shift)
726 return ts & ~((1ULL << shift) - 1);
729 /* return the pointer to the group with lowest index in the bitmap */
730 static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
731 unsigned long bitmap)
733 int index = __ffs(bitmap);
734 return &q->groups[index];
736 /* Calculate a mask to mimic what would be ffs_from(). */
737 static inline unsigned long mask_from(unsigned long bitmap, int from)
739 return bitmap & ~((1UL << from) - 1);
743 * The state computation relies on ER=0, IR=1, EB=2, IB=3
744 * First compute eligibility comparing grp->S, q->V,
745 * then check if someone is blocking us and possibly add EB
747 static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
749 /* if S > V we are not eligible */
750 unsigned int state = qfq_gt(grp->S, q->V);
751 unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
752 struct qfq_group *next;
755 next = qfq_ffs(q, mask);
756 if (qfq_gt(grp->F, next->F))
766 * q->bitmaps[dst] |= q->bitmaps[src] & mask;
767 * q->bitmaps[src] &= ~mask;
768 * but we should make sure that src != dst
770 static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
773 q->bitmaps[dst] |= q->bitmaps[src] & mask;
774 q->bitmaps[src] &= ~mask;
777 static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
779 unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
780 struct qfq_group *next;
783 next = qfq_ffs(q, mask);
784 if (!qfq_gt(next->F, old_F))
788 mask = (1UL << index) - 1;
789 qfq_move_groups(q, mask, EB, ER);
790 qfq_move_groups(q, mask, IB, IR);
797 old_V >>= q->min_slot_shift;
803 static void qfq_make_eligible(struct qfq_sched *q)
805 unsigned long vslot = q->V >> q->min_slot_shift;
806 unsigned long old_vslot = q->oldV >> q->min_slot_shift;
808 if (vslot != old_vslot) {
810 int last_flip_pos = fls(vslot ^ old_vslot);
812 if (last_flip_pos > 31) /* higher than the number of groups */
813 mask = ~0UL; /* make all groups eligible */
815 mask = (1UL << last_flip_pos) - 1;
817 qfq_move_groups(q, mask, IR, ER);
818 qfq_move_groups(q, mask, IB, EB);
823 * The index of the slot in which the input aggregate agg is to be
824 * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2'
825 * and not a '-1' because the start time of the group may be moved
826 * backward by one slot after the aggregate has been inserted, and
827 * this would cause non-empty slots to be right-shifted by one
830 * QFQ+ fully satisfies this bound to the slot index if the parameters
831 * of the classes are not changed dynamically, and if QFQ+ never
832 * happens to postpone the service of agg unjustly, i.e., it never
833 * happens that the aggregate becomes backlogged and eligible, or just
834 * eligible, while an aggregate with a higher approximated finish time
835 * is being served. In particular, in this case QFQ+ guarantees that
836 * the timestamps of agg are low enough that the slot index is never
837 * higher than 2. Unfortunately, QFQ+ cannot provide the same
838 * guarantee if it happens to unjustly postpone the service of agg, or
839 * if the parameters of some class are changed.
841 * As for the first event, i.e., an out-of-order service, the
842 * upper bound to the slot index guaranteed by QFQ+ grows to
844 * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
845 * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1.
847 * The following function deals with this problem by backward-shifting
848 * the timestamps of agg, if needed, so as to guarantee that the slot
849 * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may
850 * cause the service of other aggregates to be postponed, yet the
851 * worst-case guarantees of these aggregates are not violated. In
852 * fact, in case of no out-of-order service, the timestamps of agg
853 * would have been even lower than they are after the backward shift,
854 * because QFQ+ would have guaranteed a maximum value equal to 2 for
855 * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose
856 * service is postponed because of the backward-shift would have
857 * however waited for the service of agg before being served.
859 * The other event that may cause the slot index to be higher than 2
860 * for agg is a recent change of the parameters of some class. If the
861 * weight of a class is increased or the lmax (max_pkt_size) of the
862 * class is decreased, then a new aggregate with smaller slot size
863 * than the original parent aggregate of the class may happen to be
864 * activated. The activation of this aggregate should be properly
865 * delayed to when the service of the class has finished in the ideal
866 * system tracked by QFQ+. If the activation of the aggregate is not
867 * delayed to this reference time instant, then this aggregate may be
868 * unjustly served before other aggregates waiting for service. This
869 * may cause the above bound to the slot index to be violated for some
870 * of these unlucky aggregates.
872 * Instead of delaying the activation of the new aggregate, which is
873 * quite complex, the above-discussed capping of the slot index is
874 * used to handle also the consequences of a change of the parameters
877 static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg,
880 u64 slot = (roundedS - grp->S) >> grp->slot_shift;
881 unsigned int i; /* slot index in the bucket list */
883 if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
884 u64 deltaS = roundedS - grp->S -
885 ((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
888 slot = QFQ_MAX_SLOTS - 2;
891 i = (grp->front + slot) % QFQ_MAX_SLOTS;
893 hlist_add_head(&agg->next, &grp->slots[i]);
894 __set_bit(slot, &grp->full_slots);
897 /* Maybe introduce hlist_first_entry?? */
898 static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp)
900 return hlist_entry(grp->slots[grp->front].first,
901 struct qfq_aggregate, next);
905 * remove the entry from the slot
907 static void qfq_front_slot_remove(struct qfq_group *grp)
909 struct qfq_aggregate *agg = qfq_slot_head(grp);
912 hlist_del(&agg->next);
913 if (hlist_empty(&grp->slots[grp->front]))
914 __clear_bit(0, &grp->full_slots);
918 * Returns the first aggregate in the first non-empty bucket of the
919 * group. As a side effect, adjusts the bucket list so the first
920 * non-empty bucket is at position 0 in full_slots.
922 static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp)
926 pr_debug("qfq slot_scan: grp %u full %#lx\n",
927 grp->index, grp->full_slots);
929 if (grp->full_slots == 0)
932 i = __ffs(grp->full_slots); /* zero based */
934 grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
935 grp->full_slots >>= i;
938 return qfq_slot_head(grp);
942 * adjust the bucket list. When the start time of a group decreases,
943 * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
944 * move the objects. The mask of occupied slots must be shifted
945 * because we use ffs() to find the first non-empty slot.
946 * This covers decreases in the group's start time, but what about
947 * increases of the start time ?
948 * Here too we should make sure that i is less than 32
950 static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
952 unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
954 grp->full_slots <<= i;
955 grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
958 static void qfq_update_eligible(struct qfq_sched *q)
960 struct qfq_group *grp;
961 unsigned long ineligible;
963 ineligible = q->bitmaps[IR] | q->bitmaps[IB];
965 if (!q->bitmaps[ER]) {
966 grp = qfq_ffs(q, ineligible);
967 if (qfq_gt(grp->S, q->V))
970 qfq_make_eligible(q);
974 /* Dequeue head packet of the head class in the DRR queue of the aggregate. */
975 static void agg_dequeue(struct qfq_aggregate *agg,
976 struct qfq_class *cl, unsigned int len)
978 qdisc_dequeue_peeked(cl->qdisc);
980 cl->deficit -= (int) len;
982 if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */
983 list_del(&cl->alist);
984 else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) {
985 cl->deficit += agg->lmax;
986 list_move_tail(&cl->alist, &agg->active);
990 static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg,
991 struct qfq_class **cl,
996 *cl = list_first_entry(&agg->active, struct qfq_class, alist);
997 skb = (*cl)->qdisc->ops->peek((*cl)->qdisc);
999 WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
1001 *len = qdisc_pkt_len(skb);
1006 /* Update F according to the actual service received by the aggregate. */
1007 static inline void charge_actual_service(struct qfq_aggregate *agg)
1009 /* Compute the service received by the aggregate, taking into
1010 * account that, after decreasing the number of classes in
1011 * agg, it may happen that
1012 * agg->initial_budget - agg->budget > agg->bugdetmax
1014 u32 service_received = min(agg->budgetmax,
1015 agg->initial_budget - agg->budget);
1017 agg->F = agg->S + (u64)service_received * agg->inv_w;
1020 /* Assign a reasonable start time for a new aggregate in group i.
1021 * Admissible values for \hat(F) are multiples of \sigma_i
1022 * no greater than V+\sigma_i . Larger values mean that
1023 * we had a wraparound so we consider the timestamp to be stale.
1025 * If F is not stale and F >= V then we set S = F.
1026 * Otherwise we should assign S = V, but this may violate
1027 * the ordering in EB (see [2]). So, if we have groups in ER,
1028 * set S to the F_j of the first group j which would be blocking us.
1029 * We are guaranteed not to move S backward because
1030 * otherwise our group i would still be blocked.
1032 static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg)
1035 u64 limit, roundedF;
1036 int slot_shift = agg->grp->slot_shift;
1038 roundedF = qfq_round_down(agg->F, slot_shift);
1039 limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
1041 if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) {
1042 /* timestamp was stale */
1043 mask = mask_from(q->bitmaps[ER], agg->grp->index);
1045 struct qfq_group *next = qfq_ffs(q, mask);
1046 if (qfq_gt(roundedF, next->F)) {
1047 if (qfq_gt(limit, next->F))
1049 else /* preserve timestamp correctness */
1055 } else /* timestamp is not stale */
1059 /* Update the timestamps of agg before scheduling/rescheduling it for
1060 * service. In particular, assign to agg->F its maximum possible
1061 * value, i.e., the virtual finish time with which the aggregate
1062 * should be labeled if it used all its budget once in service.
1065 qfq_update_agg_ts(struct qfq_sched *q,
1066 struct qfq_aggregate *agg, enum update_reason reason)
1068 if (reason != requeue)
1069 qfq_update_start(q, agg);
1070 else /* just charge agg for the service received */
1073 agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w;
1076 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg);
1078 static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
1080 struct qfq_sched *q = qdisc_priv(sch);
1081 struct qfq_aggregate *in_serv_agg = q->in_serv_agg;
1082 struct qfq_class *cl;
1083 struct sk_buff *skb = NULL;
1084 /* next-packet len, 0 means no more active classes in in-service agg */
1085 unsigned int len = 0;
1087 if (in_serv_agg == NULL)
1090 if (!list_empty(&in_serv_agg->active))
1091 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1094 * If there are no active classes in the in-service aggregate,
1095 * or if the aggregate has not enough budget to serve its next
1096 * class, then choose the next aggregate to serve.
1098 if (len == 0 || in_serv_agg->budget < len) {
1099 charge_actual_service(in_serv_agg);
1101 /* recharge the budget of the aggregate */
1102 in_serv_agg->initial_budget = in_serv_agg->budget =
1103 in_serv_agg->budgetmax;
1105 if (!list_empty(&in_serv_agg->active)) {
1107 * Still active: reschedule for
1108 * service. Possible optimization: if no other
1109 * aggregate is active, then there is no point
1110 * in rescheduling this aggregate, and we can
1111 * just keep it as the in-service one. This
1112 * should be however a corner case, and to
1113 * handle it, we would need to maintain an
1114 * extra num_active_aggs field.
1116 qfq_update_agg_ts(q, in_serv_agg, requeue);
1117 qfq_schedule_agg(q, in_serv_agg);
1118 } else if (sch->q.qlen == 0) { /* no aggregate to serve */
1119 q->in_serv_agg = NULL;
1124 * If we get here, there are other aggregates queued:
1125 * choose the new aggregate to serve.
1127 in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
1128 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1133 qdisc_qstats_backlog_dec(sch, skb);
1135 qdisc_bstats_update(sch, skb);
1137 agg_dequeue(in_serv_agg, cl, len);
1138 /* If lmax is lowered, through qfq_change_class, for a class
1139 * owning pending packets with larger size than the new value
1140 * of lmax, then the following condition may hold.
1142 if (unlikely(in_serv_agg->budget < len))
1143 in_serv_agg->budget = 0;
1145 in_serv_agg->budget -= len;
1147 q->V += (u64)len * q->iwsum;
1148 pr_debug("qfq dequeue: len %u F %lld now %lld\n",
1149 len, (unsigned long long) in_serv_agg->F,
1150 (unsigned long long) q->V);
1155 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)
1157 struct qfq_group *grp;
1158 struct qfq_aggregate *agg, *new_front_agg;
1161 qfq_update_eligible(q);
1164 if (!q->bitmaps[ER])
1167 grp = qfq_ffs(q, q->bitmaps[ER]);
1170 agg = qfq_slot_head(grp);
1172 /* agg starts to be served, remove it from schedule */
1173 qfq_front_slot_remove(grp);
1175 new_front_agg = qfq_slot_scan(grp);
1177 if (new_front_agg == NULL) /* group is now inactive, remove from ER */
1178 __clear_bit(grp->index, &q->bitmaps[ER]);
1180 u64 roundedS = qfq_round_down(new_front_agg->S,
1184 if (grp->S == roundedS)
1187 grp->F = roundedS + (2ULL << grp->slot_shift);
1188 __clear_bit(grp->index, &q->bitmaps[ER]);
1189 s = qfq_calc_state(q, grp);
1190 __set_bit(grp->index, &q->bitmaps[s]);
1193 qfq_unblock_groups(q, grp->index, old_F);
1198 static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
1199 struct sk_buff **to_free)
1201 unsigned int len = qdisc_pkt_len(skb), gso_segs;
1202 struct qfq_sched *q = qdisc_priv(sch);
1203 struct qfq_class *cl;
1204 struct qfq_aggregate *agg;
1208 cl = qfq_classify(skb, sch, &err);
1210 if (err & __NET_XMIT_BYPASS)
1211 qdisc_qstats_drop(sch);
1212 __qdisc_drop(skb, to_free);
1215 pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
1217 if (unlikely(cl->agg->lmax < len)) {
1218 pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
1219 cl->agg->lmax, len, cl->common.classid);
1220 err = qfq_change_agg(sch, cl, cl->agg->class_weight, len);
1223 return qdisc_drop(skb, sch, to_free);
1227 gso_segs = skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 1;
1228 first = !cl->qdisc->q.qlen;
1229 err = qdisc_enqueue(skb, cl->qdisc, to_free);
1230 if (unlikely(err != NET_XMIT_SUCCESS)) {
1231 pr_debug("qfq_enqueue: enqueue failed %d\n", err);
1232 if (net_xmit_drop_count(err)) {
1234 qdisc_qstats_drop(sch);
1239 cl->bstats.bytes += len;
1240 cl->bstats.packets += gso_segs;
1241 sch->qstats.backlog += len;
1245 /* if the queue was not empty, then done here */
1247 if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) &&
1248 list_first_entry(&agg->active, struct qfq_class, alist)
1249 == cl && cl->deficit < len)
1250 list_move_tail(&cl->alist, &agg->active);
1255 /* schedule class for service within the aggregate */
1256 cl->deficit = agg->lmax;
1257 list_add_tail(&cl->alist, &agg->active);
1259 if (list_first_entry(&agg->active, struct qfq_class, alist) != cl ||
1260 q->in_serv_agg == agg)
1261 return err; /* non-empty or in service, nothing else to do */
1263 qfq_activate_agg(q, agg, enqueue);
1269 * Schedule aggregate according to its timestamps.
1271 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1273 struct qfq_group *grp = agg->grp;
1277 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1280 * Insert agg in the correct bucket.
1281 * If agg->S >= grp->S we don't need to adjust the
1282 * bucket list and simply go to the insertion phase.
1283 * Otherwise grp->S is decreasing, we must make room
1284 * in the bucket list, and also recompute the group state.
1285 * Finally, if there were no flows in this group and nobody
1286 * was in ER make sure to adjust V.
1288 if (grp->full_slots) {
1289 if (!qfq_gt(grp->S, agg->S))
1292 /* create a slot for this agg->S */
1293 qfq_slot_rotate(grp, roundedS);
1294 /* group was surely ineligible, remove */
1295 __clear_bit(grp->index, &q->bitmaps[IR]);
1296 __clear_bit(grp->index, &q->bitmaps[IB]);
1297 } else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) &&
1298 q->in_serv_agg == NULL)
1302 grp->F = roundedS + (2ULL << grp->slot_shift);
1303 s = qfq_calc_state(q, grp);
1304 __set_bit(grp->index, &q->bitmaps[s]);
1306 pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
1308 (unsigned long long) agg->S,
1309 (unsigned long long) agg->F,
1310 (unsigned long long) q->V);
1313 qfq_slot_insert(grp, agg, roundedS);
1317 /* Update agg ts and schedule agg for service */
1318 static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
1319 enum update_reason reason)
1321 agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */
1323 qfq_update_agg_ts(q, agg, reason);
1324 if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */
1325 q->in_serv_agg = agg; /* start serving this aggregate */
1326 /* update V: to be in service, agg must be eligible */
1327 q->oldV = q->V = agg->S;
1328 } else if (agg != q->in_serv_agg)
1329 qfq_schedule_agg(q, agg);
1332 static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
1333 struct qfq_aggregate *agg)
1335 unsigned int i, offset;
1338 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1339 offset = (roundedS - grp->S) >> grp->slot_shift;
1341 i = (grp->front + offset) % QFQ_MAX_SLOTS;
1343 hlist_del(&agg->next);
1344 if (hlist_empty(&grp->slots[i]))
1345 __clear_bit(offset, &grp->full_slots);
1349 * Called to forcibly deschedule an aggregate. If the aggregate is
1350 * not in the front bucket, or if the latter has other aggregates in
1351 * the front bucket, we can simply remove the aggregate with no other
1353 * Otherwise we must propagate the event up.
1355 static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1357 struct qfq_group *grp = agg->grp;
1362 if (agg == q->in_serv_agg) {
1363 charge_actual_service(agg);
1364 q->in_serv_agg = qfq_choose_next_agg(q);
1369 qfq_slot_remove(q, grp, agg);
1371 if (!grp->full_slots) {
1372 __clear_bit(grp->index, &q->bitmaps[IR]);
1373 __clear_bit(grp->index, &q->bitmaps[EB]);
1374 __clear_bit(grp->index, &q->bitmaps[IB]);
1376 if (test_bit(grp->index, &q->bitmaps[ER]) &&
1377 !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
1378 mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
1380 mask = ~((1UL << __fls(mask)) - 1);
1383 qfq_move_groups(q, mask, EB, ER);
1384 qfq_move_groups(q, mask, IB, IR);
1386 __clear_bit(grp->index, &q->bitmaps[ER]);
1387 } else if (hlist_empty(&grp->slots[grp->front])) {
1388 agg = qfq_slot_scan(grp);
1389 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1390 if (grp->S != roundedS) {
1391 __clear_bit(grp->index, &q->bitmaps[ER]);
1392 __clear_bit(grp->index, &q->bitmaps[IR]);
1393 __clear_bit(grp->index, &q->bitmaps[EB]);
1394 __clear_bit(grp->index, &q->bitmaps[IB]);
1396 grp->F = roundedS + (2ULL << grp->slot_shift);
1397 s = qfq_calc_state(q, grp);
1398 __set_bit(grp->index, &q->bitmaps[s]);
1403 static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1405 struct qfq_sched *q = qdisc_priv(sch);
1406 struct qfq_class *cl = (struct qfq_class *)arg;
1408 qfq_deactivate_class(q, cl);
1411 static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt,
1412 struct netlink_ext_ack *extack)
1414 struct qfq_sched *q = qdisc_priv(sch);
1415 struct qfq_group *grp;
1417 u32 max_cl_shift, maxbudg_shift, max_classes;
1419 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
1423 err = qdisc_class_hash_init(&q->clhash);
1427 if (qdisc_dev(sch)->tx_queue_len + 1 > QFQ_MAX_AGG_CLASSES)
1428 max_classes = QFQ_MAX_AGG_CLASSES;
1430 max_classes = qdisc_dev(sch)->tx_queue_len + 1;
1431 /* max_cl_shift = floor(log_2(max_classes)) */
1432 max_cl_shift = __fls(max_classes);
1433 q->max_agg_classes = 1<<max_cl_shift;
1435 /* maxbudg_shift = log2(max_len * max_classes_per_agg) */
1436 maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift;
1437 q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX;
1439 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1440 grp = &q->groups[i];
1442 grp->slot_shift = q->min_slot_shift + i;
1443 for (j = 0; j < QFQ_MAX_SLOTS; j++)
1444 INIT_HLIST_HEAD(&grp->slots[j]);
1447 INIT_HLIST_HEAD(&q->nonfull_aggs);
1452 static void qfq_reset_qdisc(struct Qdisc *sch)
1454 struct qfq_sched *q = qdisc_priv(sch);
1455 struct qfq_class *cl;
1458 for (i = 0; i < q->clhash.hashsize; i++) {
1459 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1460 if (cl->qdisc->q.qlen > 0)
1461 qfq_deactivate_class(q, cl);
1463 qdisc_reset(cl->qdisc);
1466 sch->qstats.backlog = 0;
1470 static void qfq_destroy_qdisc(struct Qdisc *sch)
1472 struct qfq_sched *q = qdisc_priv(sch);
1473 struct qfq_class *cl;
1474 struct hlist_node *next;
1477 tcf_block_put(q->block);
1479 for (i = 0; i < q->clhash.hashsize; i++) {
1480 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1482 qfq_destroy_class(sch, cl);
1485 qdisc_class_hash_destroy(&q->clhash);
1488 static const struct Qdisc_class_ops qfq_class_ops = {
1489 .change = qfq_change_class,
1490 .delete = qfq_delete_class,
1491 .find = qfq_search_class,
1492 .tcf_block = qfq_tcf_block,
1493 .bind_tcf = qfq_bind_tcf,
1494 .unbind_tcf = qfq_unbind_tcf,
1495 .graft = qfq_graft_class,
1496 .leaf = qfq_class_leaf,
1497 .qlen_notify = qfq_qlen_notify,
1498 .dump = qfq_dump_class,
1499 .dump_stats = qfq_dump_class_stats,
1503 static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1504 .cl_ops = &qfq_class_ops,
1506 .priv_size = sizeof(struct qfq_sched),
1507 .enqueue = qfq_enqueue,
1508 .dequeue = qfq_dequeue,
1509 .peek = qdisc_peek_dequeued,
1510 .init = qfq_init_qdisc,
1511 .reset = qfq_reset_qdisc,
1512 .destroy = qfq_destroy_qdisc,
1513 .owner = THIS_MODULE,
1516 static int __init qfq_init(void)
1518 return register_qdisc(&qfq_qdisc_ops);
1521 static void __exit qfq_exit(void)
1523 unregister_qdisc(&qfq_qdisc_ops);
1526 module_init(qfq_init);
1527 module_exit(qfq_exit);
1528 MODULE_LICENSE("GPL");