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;
136 unsigned int filter_cnt;
138 struct gnet_stats_basic_packed bstats;
139 struct gnet_stats_queue qstats;
140 struct gnet_stats_rate_est64 rate_est;
142 struct list_head alist; /* Link for active-classes list. */
143 struct qfq_aggregate *agg; /* Parent aggregate. */
144 int deficit; /* DRR deficit counter. */
147 struct qfq_aggregate {
148 struct hlist_node next; /* Link for the slot list. */
149 u64 S, F; /* flow timestamps (exact) */
151 /* group we belong to. In principle we would need the index,
152 * which is log_2(lmax/weight), but we never reference it
153 * directly, only the group.
155 struct qfq_group *grp;
157 /* these are copied from the flowset. */
158 u32 class_weight; /* Weight of each class in this aggregate. */
159 /* Max pkt size for the classes in this aggregate, DRR quantum. */
162 u32 inv_w; /* ONE_FP/(sum of weights of classes in aggr.). */
163 u32 budgetmax; /* Max budget for this aggregate. */
164 u32 initial_budget, budget; /* Initial and current budget. */
166 int num_classes; /* Number of classes in this aggr. */
167 struct list_head active; /* DRR queue of active classes. */
169 struct hlist_node nonfull_next; /* See nonfull_aggs in qfq_sched. */
173 u64 S, F; /* group timestamps (approx). */
174 unsigned int slot_shift; /* Slot shift. */
175 unsigned int index; /* Group index. */
176 unsigned int front; /* Index of the front slot. */
177 unsigned long full_slots; /* non-empty slots */
179 /* Array of RR lists of active aggregates. */
180 struct hlist_head slots[QFQ_MAX_SLOTS];
184 struct tcf_proto __rcu *filter_list;
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 void qfq_purge_queue(struct qfq_class *cl)
222 unsigned int len = cl->qdisc->q.qlen;
223 unsigned int backlog = cl->qdisc->qstats.backlog;
225 qdisc_reset(cl->qdisc);
226 qdisc_tree_reduce_backlog(cl->qdisc, len, backlog);
229 static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = {
230 [TCA_QFQ_WEIGHT] = { .type = NLA_U32 },
231 [TCA_QFQ_LMAX] = { .type = NLA_U32 },
235 * Calculate a flow index, given its weight and maximum packet length.
236 * index = log_2(maxlen/weight) but we need to apply the scaling.
237 * This is used only once at flow creation.
239 static int qfq_calc_index(u32 inv_w, unsigned int maxlen, u32 min_slot_shift)
241 u64 slot_size = (u64)maxlen * inv_w;
242 unsigned long size_map;
245 size_map = slot_size >> min_slot_shift;
249 index = __fls(size_map) + 1; /* basically a log_2 */
250 index -= !(slot_size - (1ULL << (index + min_slot_shift - 1)));
255 pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n",
256 (unsigned long) ONE_FP/inv_w, maxlen, index);
261 static void qfq_deactivate_agg(struct qfq_sched *, struct qfq_aggregate *);
262 static void qfq_activate_agg(struct qfq_sched *, struct qfq_aggregate *,
265 static void qfq_init_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
266 u32 lmax, u32 weight)
268 INIT_LIST_HEAD(&agg->active);
269 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
272 agg->class_weight = weight;
275 static struct qfq_aggregate *qfq_find_agg(struct qfq_sched *q,
276 u32 lmax, u32 weight)
278 struct qfq_aggregate *agg;
280 hlist_for_each_entry(agg, &q->nonfull_aggs, nonfull_next)
281 if (agg->lmax == lmax && agg->class_weight == weight)
288 /* Update aggregate as a function of the new number of classes. */
289 static void qfq_update_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
294 if (new_num_classes == q->max_agg_classes)
295 hlist_del_init(&agg->nonfull_next);
297 if (agg->num_classes > new_num_classes &&
298 new_num_classes == q->max_agg_classes - 1) /* agg no more full */
299 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
301 /* The next assignment may let
302 * agg->initial_budget > agg->budgetmax
303 * hold, we will take it into account in charge_actual_service().
305 agg->budgetmax = new_num_classes * agg->lmax;
306 new_agg_weight = agg->class_weight * new_num_classes;
307 agg->inv_w = ONE_FP/new_agg_weight;
309 if (agg->grp == NULL) {
310 int i = qfq_calc_index(agg->inv_w, agg->budgetmax,
312 agg->grp = &q->groups[i];
316 (int) agg->class_weight * (new_num_classes - agg->num_classes);
317 q->iwsum = ONE_FP / q->wsum;
319 agg->num_classes = new_num_classes;
322 /* Add class to aggregate. */
323 static void qfq_add_to_agg(struct qfq_sched *q,
324 struct qfq_aggregate *agg,
325 struct qfq_class *cl)
329 qfq_update_agg(q, agg, agg->num_classes+1);
330 if (cl->qdisc->q.qlen > 0) { /* adding an active class */
331 list_add_tail(&cl->alist, &agg->active);
332 if (list_first_entry(&agg->active, struct qfq_class, alist) ==
333 cl && q->in_serv_agg != agg) /* agg was inactive */
334 qfq_activate_agg(q, agg, enqueue); /* schedule agg */
338 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *);
340 static void qfq_destroy_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
342 hlist_del_init(&agg->nonfull_next);
343 q->wsum -= agg->class_weight;
345 q->iwsum = ONE_FP / q->wsum;
347 if (q->in_serv_agg == agg)
348 q->in_serv_agg = qfq_choose_next_agg(q);
352 /* Deschedule class from within its parent aggregate. */
353 static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
355 struct qfq_aggregate *agg = cl->agg;
358 list_del(&cl->alist); /* remove from RR queue of the aggregate */
359 if (list_empty(&agg->active)) /* agg is now inactive */
360 qfq_deactivate_agg(q, agg);
363 /* Remove class from its parent aggregate. */
364 static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
366 struct qfq_aggregate *agg = cl->agg;
369 if (agg->num_classes == 1) { /* agg being emptied, destroy it */
370 qfq_destroy_agg(q, agg);
373 qfq_update_agg(q, agg, agg->num_classes-1);
376 /* Deschedule class and remove it from its parent aggregate. */
377 static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
379 if (cl->qdisc->q.qlen > 0) /* class is active */
380 qfq_deactivate_class(q, cl);
382 qfq_rm_from_agg(q, cl);
385 /* Move class to a new aggregate, matching the new class weight and/or lmax */
386 static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight,
389 struct qfq_sched *q = qdisc_priv(sch);
390 struct qfq_aggregate *new_agg = qfq_find_agg(q, lmax, weight);
392 if (new_agg == NULL) { /* create new aggregate */
393 new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC);
396 qfq_init_agg(q, new_agg, lmax, weight);
398 qfq_deact_rm_from_agg(q, cl);
399 qfq_add_to_agg(q, new_agg, cl);
404 static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
405 struct nlattr **tca, unsigned long *arg)
407 struct qfq_sched *q = qdisc_priv(sch);
408 struct qfq_class *cl = (struct qfq_class *)*arg;
409 bool existing = false;
410 struct nlattr *tb[TCA_QFQ_MAX + 1];
411 struct qfq_aggregate *new_agg = NULL;
412 u32 weight, lmax, inv_w;
416 if (tca[TCA_OPTIONS] == NULL) {
417 pr_notice("qfq: no options\n");
421 err = nla_parse_nested(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS], qfq_policy);
425 if (tb[TCA_QFQ_WEIGHT]) {
426 weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]);
427 if (!weight || weight > (1UL << QFQ_MAX_WSHIFT)) {
428 pr_notice("qfq: invalid weight %u\n", weight);
434 if (tb[TCA_QFQ_LMAX]) {
435 lmax = nla_get_u32(tb[TCA_QFQ_LMAX]);
436 if (lmax < QFQ_MIN_LMAX || lmax > (1UL << QFQ_MTU_SHIFT)) {
437 pr_notice("qfq: invalid max length %u\n", lmax);
441 lmax = psched_mtu(qdisc_dev(sch));
443 inv_w = ONE_FP / weight;
444 weight = ONE_FP / inv_w;
447 lmax == cl->agg->lmax &&
448 weight == cl->agg->class_weight)
449 return 0; /* nothing to change */
451 delta_w = weight - (cl ? cl->agg->class_weight : 0);
453 if (q->wsum + delta_w > QFQ_MAX_WSUM) {
454 pr_notice("qfq: total weight out of range (%d + %u)\n",
459 if (cl != NULL) { /* modify existing class */
461 err = gen_replace_estimator(&cl->bstats, NULL,
464 qdisc_root_sleeping_running(sch),
473 /* create and init new class */
474 cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
479 cl->common.classid = classid;
482 cl->qdisc = qdisc_create_dflt(sch->dev_queue,
483 &pfifo_qdisc_ops, classid);
484 if (cl->qdisc == NULL)
485 cl->qdisc = &noop_qdisc;
488 err = gen_new_estimator(&cl->bstats, NULL,
491 qdisc_root_sleeping_running(sch),
498 qdisc_class_hash_insert(&q->clhash, &cl->common);
499 sch_tree_unlock(sch);
501 qdisc_class_hash_grow(sch, &q->clhash);
505 new_agg = qfq_find_agg(q, lmax, weight);
506 if (new_agg == NULL) { /* create new aggregate */
507 sch_tree_unlock(sch);
508 new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL);
509 if (new_agg == NULL) {
511 gen_kill_estimator(&cl->bstats, &cl->rate_est);
515 qfq_init_agg(q, new_agg, lmax, weight);
518 qfq_deact_rm_from_agg(q, cl);
519 qfq_add_to_agg(q, new_agg, cl);
520 sch_tree_unlock(sch);
522 *arg = (unsigned long)cl;
526 qdisc_destroy(cl->qdisc);
531 static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
533 struct qfq_sched *q = qdisc_priv(sch);
535 qfq_rm_from_agg(q, cl);
536 gen_kill_estimator(&cl->bstats, &cl->rate_est);
537 qdisc_destroy(cl->qdisc);
541 static int qfq_delete_class(struct Qdisc *sch, unsigned long arg)
543 struct qfq_sched *q = qdisc_priv(sch);
544 struct qfq_class *cl = (struct qfq_class *)arg;
546 if (cl->filter_cnt > 0)
552 qdisc_class_hash_remove(&q->clhash, &cl->common);
554 BUG_ON(--cl->refcnt == 0);
556 * This shouldn't happen: we "hold" one cops->get() when called
557 * from tc_ctl_tclass; the destroy method is done from cops->put().
560 sch_tree_unlock(sch);
564 static unsigned long qfq_get_class(struct Qdisc *sch, u32 classid)
566 struct qfq_class *cl = qfq_find_class(sch, classid);
571 return (unsigned long)cl;
574 static void qfq_put_class(struct Qdisc *sch, unsigned long arg)
576 struct qfq_class *cl = (struct qfq_class *)arg;
578 if (--cl->refcnt == 0)
579 qfq_destroy_class(sch, cl);
582 static struct tcf_proto __rcu **qfq_tcf_chain(struct Qdisc *sch,
585 struct qfq_sched *q = qdisc_priv(sch);
590 return &q->filter_list;
593 static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
596 struct qfq_class *cl = qfq_find_class(sch, classid);
601 return (unsigned long)cl;
604 static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
606 struct qfq_class *cl = (struct qfq_class *)arg;
611 static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
612 struct Qdisc *new, struct Qdisc **old)
614 struct qfq_class *cl = (struct qfq_class *)arg;
617 new = qdisc_create_dflt(sch->dev_queue,
618 &pfifo_qdisc_ops, cl->common.classid);
623 *old = qdisc_replace(sch, new, &cl->qdisc);
627 static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
629 struct qfq_class *cl = (struct qfq_class *)arg;
634 static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
635 struct sk_buff *skb, struct tcmsg *tcm)
637 struct qfq_class *cl = (struct qfq_class *)arg;
640 tcm->tcm_parent = TC_H_ROOT;
641 tcm->tcm_handle = cl->common.classid;
642 tcm->tcm_info = cl->qdisc->handle;
644 nest = nla_nest_start(skb, TCA_OPTIONS);
646 goto nla_put_failure;
647 if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) ||
648 nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax))
649 goto nla_put_failure;
650 return nla_nest_end(skb, nest);
653 nla_nest_cancel(skb, nest);
657 static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
660 struct qfq_class *cl = (struct qfq_class *)arg;
661 struct tc_qfq_stats xstats;
663 memset(&xstats, 0, sizeof(xstats));
665 xstats.weight = cl->agg->class_weight;
666 xstats.lmax = cl->agg->lmax;
668 if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
669 d, NULL, &cl->bstats) < 0 ||
670 gnet_stats_copy_rate_est(d, &cl->bstats, &cl->rate_est) < 0 ||
671 gnet_stats_copy_queue(d, NULL,
672 &cl->qdisc->qstats, cl->qdisc->q.qlen) < 0)
675 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
678 static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
680 struct qfq_sched *q = qdisc_priv(sch);
681 struct qfq_class *cl;
687 for (i = 0; i < q->clhash.hashsize; i++) {
688 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
689 if (arg->count < arg->skip) {
693 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
702 static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
705 struct qfq_sched *q = qdisc_priv(sch);
706 struct qfq_class *cl;
707 struct tcf_result res;
708 struct tcf_proto *fl;
711 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
712 pr_debug("qfq_classify: found %d\n", skb->priority);
713 cl = qfq_find_class(sch, skb->priority);
718 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
719 fl = rcu_dereference_bh(q->filter_list);
720 result = tc_classify(skb, fl, &res, false);
722 #ifdef CONFIG_NET_CLS_ACT
726 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
731 cl = (struct qfq_class *)res.class;
733 cl = qfq_find_class(sch, res.classid);
740 /* Generic comparison function, handling wraparound. */
741 static inline int qfq_gt(u64 a, u64 b)
743 return (s64)(a - b) > 0;
746 /* Round a precise timestamp to its slotted value. */
747 static inline u64 qfq_round_down(u64 ts, unsigned int shift)
749 return ts & ~((1ULL << shift) - 1);
752 /* return the pointer to the group with lowest index in the bitmap */
753 static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
754 unsigned long bitmap)
756 int index = __ffs(bitmap);
757 return &q->groups[index];
759 /* Calculate a mask to mimic what would be ffs_from(). */
760 static inline unsigned long mask_from(unsigned long bitmap, int from)
762 return bitmap & ~((1UL << from) - 1);
766 * The state computation relies on ER=0, IR=1, EB=2, IB=3
767 * First compute eligibility comparing grp->S, q->V,
768 * then check if someone is blocking us and possibly add EB
770 static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
772 /* if S > V we are not eligible */
773 unsigned int state = qfq_gt(grp->S, q->V);
774 unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
775 struct qfq_group *next;
778 next = qfq_ffs(q, mask);
779 if (qfq_gt(grp->F, next->F))
789 * q->bitmaps[dst] |= q->bitmaps[src] & mask;
790 * q->bitmaps[src] &= ~mask;
791 * but we should make sure that src != dst
793 static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
796 q->bitmaps[dst] |= q->bitmaps[src] & mask;
797 q->bitmaps[src] &= ~mask;
800 static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
802 unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
803 struct qfq_group *next;
806 next = qfq_ffs(q, mask);
807 if (!qfq_gt(next->F, old_F))
811 mask = (1UL << index) - 1;
812 qfq_move_groups(q, mask, EB, ER);
813 qfq_move_groups(q, mask, IB, IR);
820 old_V >>= q->min_slot_shift;
826 static void qfq_make_eligible(struct qfq_sched *q)
828 unsigned long vslot = q->V >> q->min_slot_shift;
829 unsigned long old_vslot = q->oldV >> q->min_slot_shift;
831 if (vslot != old_vslot) {
833 int last_flip_pos = fls(vslot ^ old_vslot);
835 if (last_flip_pos > 31) /* higher than the number of groups */
836 mask = ~0UL; /* make all groups eligible */
838 mask = (1UL << last_flip_pos) - 1;
840 qfq_move_groups(q, mask, IR, ER);
841 qfq_move_groups(q, mask, IB, EB);
846 * The index of the slot in which the input aggregate agg is to be
847 * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2'
848 * and not a '-1' because the start time of the group may be moved
849 * backward by one slot after the aggregate has been inserted, and
850 * this would cause non-empty slots to be right-shifted by one
853 * QFQ+ fully satisfies this bound to the slot index if the parameters
854 * of the classes are not changed dynamically, and if QFQ+ never
855 * happens to postpone the service of agg unjustly, i.e., it never
856 * happens that the aggregate becomes backlogged and eligible, or just
857 * eligible, while an aggregate with a higher approximated finish time
858 * is being served. In particular, in this case QFQ+ guarantees that
859 * the timestamps of agg are low enough that the slot index is never
860 * higher than 2. Unfortunately, QFQ+ cannot provide the same
861 * guarantee if it happens to unjustly postpone the service of agg, or
862 * if the parameters of some class are changed.
864 * As for the first event, i.e., an out-of-order service, the
865 * upper bound to the slot index guaranteed by QFQ+ grows to
867 * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
868 * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1.
870 * The following function deals with this problem by backward-shifting
871 * the timestamps of agg, if needed, so as to guarantee that the slot
872 * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may
873 * cause the service of other aggregates to be postponed, yet the
874 * worst-case guarantees of these aggregates are not violated. In
875 * fact, in case of no out-of-order service, the timestamps of agg
876 * would have been even lower than they are after the backward shift,
877 * because QFQ+ would have guaranteed a maximum value equal to 2 for
878 * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose
879 * service is postponed because of the backward-shift would have
880 * however waited for the service of agg before being served.
882 * The other event that may cause the slot index to be higher than 2
883 * for agg is a recent change of the parameters of some class. If the
884 * weight of a class is increased or the lmax (max_pkt_size) of the
885 * class is decreased, then a new aggregate with smaller slot size
886 * than the original parent aggregate of the class may happen to be
887 * activated. The activation of this aggregate should be properly
888 * delayed to when the service of the class has finished in the ideal
889 * system tracked by QFQ+. If the activation of the aggregate is not
890 * delayed to this reference time instant, then this aggregate may be
891 * unjustly served before other aggregates waiting for service. This
892 * may cause the above bound to the slot index to be violated for some
893 * of these unlucky aggregates.
895 * Instead of delaying the activation of the new aggregate, which is
896 * quite complex, the above-discussed capping of the slot index is
897 * used to handle also the consequences of a change of the parameters
900 static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg,
903 u64 slot = (roundedS - grp->S) >> grp->slot_shift;
904 unsigned int i; /* slot index in the bucket list */
906 if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
907 u64 deltaS = roundedS - grp->S -
908 ((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
911 slot = QFQ_MAX_SLOTS - 2;
914 i = (grp->front + slot) % QFQ_MAX_SLOTS;
916 hlist_add_head(&agg->next, &grp->slots[i]);
917 __set_bit(slot, &grp->full_slots);
920 /* Maybe introduce hlist_first_entry?? */
921 static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp)
923 return hlist_entry(grp->slots[grp->front].first,
924 struct qfq_aggregate, next);
928 * remove the entry from the slot
930 static void qfq_front_slot_remove(struct qfq_group *grp)
932 struct qfq_aggregate *agg = qfq_slot_head(grp);
935 hlist_del(&agg->next);
936 if (hlist_empty(&grp->slots[grp->front]))
937 __clear_bit(0, &grp->full_slots);
941 * Returns the first aggregate in the first non-empty bucket of the
942 * group. As a side effect, adjusts the bucket list so the first
943 * non-empty bucket is at position 0 in full_slots.
945 static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp)
949 pr_debug("qfq slot_scan: grp %u full %#lx\n",
950 grp->index, grp->full_slots);
952 if (grp->full_slots == 0)
955 i = __ffs(grp->full_slots); /* zero based */
957 grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
958 grp->full_slots >>= i;
961 return qfq_slot_head(grp);
965 * adjust the bucket list. When the start time of a group decreases,
966 * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
967 * move the objects. The mask of occupied slots must be shifted
968 * because we use ffs() to find the first non-empty slot.
969 * This covers decreases in the group's start time, but what about
970 * increases of the start time ?
971 * Here too we should make sure that i is less than 32
973 static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
975 unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
977 grp->full_slots <<= i;
978 grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
981 static void qfq_update_eligible(struct qfq_sched *q)
983 struct qfq_group *grp;
984 unsigned long ineligible;
986 ineligible = q->bitmaps[IR] | q->bitmaps[IB];
988 if (!q->bitmaps[ER]) {
989 grp = qfq_ffs(q, ineligible);
990 if (qfq_gt(grp->S, q->V))
993 qfq_make_eligible(q);
997 /* Dequeue head packet of the head class in the DRR queue of the aggregate. */
998 static void agg_dequeue(struct qfq_aggregate *agg,
999 struct qfq_class *cl, unsigned int len)
1001 qdisc_dequeue_peeked(cl->qdisc);
1003 cl->deficit -= (int) len;
1005 if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */
1006 list_del(&cl->alist);
1007 else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) {
1008 cl->deficit += agg->lmax;
1009 list_move_tail(&cl->alist, &agg->active);
1013 static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg,
1014 struct qfq_class **cl,
1017 struct sk_buff *skb;
1019 *cl = list_first_entry(&agg->active, struct qfq_class, alist);
1020 skb = (*cl)->qdisc->ops->peek((*cl)->qdisc);
1022 WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
1024 *len = qdisc_pkt_len(skb);
1029 /* Update F according to the actual service received by the aggregate. */
1030 static inline void charge_actual_service(struct qfq_aggregate *agg)
1032 /* Compute the service received by the aggregate, taking into
1033 * account that, after decreasing the number of classes in
1034 * agg, it may happen that
1035 * agg->initial_budget - agg->budget > agg->bugdetmax
1037 u32 service_received = min(agg->budgetmax,
1038 agg->initial_budget - agg->budget);
1040 agg->F = agg->S + (u64)service_received * agg->inv_w;
1043 /* Assign a reasonable start time for a new aggregate in group i.
1044 * Admissible values for \hat(F) are multiples of \sigma_i
1045 * no greater than V+\sigma_i . Larger values mean that
1046 * we had a wraparound so we consider the timestamp to be stale.
1048 * If F is not stale and F >= V then we set S = F.
1049 * Otherwise we should assign S = V, but this may violate
1050 * the ordering in EB (see [2]). So, if we have groups in ER,
1051 * set S to the F_j of the first group j which would be blocking us.
1052 * We are guaranteed not to move S backward because
1053 * otherwise our group i would still be blocked.
1055 static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg)
1058 u64 limit, roundedF;
1059 int slot_shift = agg->grp->slot_shift;
1061 roundedF = qfq_round_down(agg->F, slot_shift);
1062 limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
1064 if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) {
1065 /* timestamp was stale */
1066 mask = mask_from(q->bitmaps[ER], agg->grp->index);
1068 struct qfq_group *next = qfq_ffs(q, mask);
1069 if (qfq_gt(roundedF, next->F)) {
1070 if (qfq_gt(limit, next->F))
1072 else /* preserve timestamp correctness */
1078 } else /* timestamp is not stale */
1082 /* Update the timestamps of agg before scheduling/rescheduling it for
1083 * service. In particular, assign to agg->F its maximum possible
1084 * value, i.e., the virtual finish time with which the aggregate
1085 * should be labeled if it used all its budget once in service.
1088 qfq_update_agg_ts(struct qfq_sched *q,
1089 struct qfq_aggregate *agg, enum update_reason reason)
1091 if (reason != requeue)
1092 qfq_update_start(q, agg);
1093 else /* just charge agg for the service received */
1096 agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w;
1099 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg);
1101 static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
1103 struct qfq_sched *q = qdisc_priv(sch);
1104 struct qfq_aggregate *in_serv_agg = q->in_serv_agg;
1105 struct qfq_class *cl;
1106 struct sk_buff *skb = NULL;
1107 /* next-packet len, 0 means no more active classes in in-service agg */
1108 unsigned int len = 0;
1110 if (in_serv_agg == NULL)
1113 if (!list_empty(&in_serv_agg->active))
1114 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1117 * If there are no active classes in the in-service aggregate,
1118 * or if the aggregate has not enough budget to serve its next
1119 * class, then choose the next aggregate to serve.
1121 if (len == 0 || in_serv_agg->budget < len) {
1122 charge_actual_service(in_serv_agg);
1124 /* recharge the budget of the aggregate */
1125 in_serv_agg->initial_budget = in_serv_agg->budget =
1126 in_serv_agg->budgetmax;
1128 if (!list_empty(&in_serv_agg->active)) {
1130 * Still active: reschedule for
1131 * service. Possible optimization: if no other
1132 * aggregate is active, then there is no point
1133 * in rescheduling this aggregate, and we can
1134 * just keep it as the in-service one. This
1135 * should be however a corner case, and to
1136 * handle it, we would need to maintain an
1137 * extra num_active_aggs field.
1139 qfq_update_agg_ts(q, in_serv_agg, requeue);
1140 qfq_schedule_agg(q, in_serv_agg);
1141 } else if (sch->q.qlen == 0) { /* no aggregate to serve */
1142 q->in_serv_agg = NULL;
1147 * If we get here, there are other aggregates queued:
1148 * choose the new aggregate to serve.
1150 in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
1151 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1157 qdisc_bstats_update(sch, skb);
1159 agg_dequeue(in_serv_agg, cl, len);
1160 /* If lmax is lowered, through qfq_change_class, for a class
1161 * owning pending packets with larger size than the new value
1162 * of lmax, then the following condition may hold.
1164 if (unlikely(in_serv_agg->budget < len))
1165 in_serv_agg->budget = 0;
1167 in_serv_agg->budget -= len;
1169 q->V += (u64)len * q->iwsum;
1170 pr_debug("qfq dequeue: len %u F %lld now %lld\n",
1171 len, (unsigned long long) in_serv_agg->F,
1172 (unsigned long long) q->V);
1177 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)
1179 struct qfq_group *grp;
1180 struct qfq_aggregate *agg, *new_front_agg;
1183 qfq_update_eligible(q);
1186 if (!q->bitmaps[ER])
1189 grp = qfq_ffs(q, q->bitmaps[ER]);
1192 agg = qfq_slot_head(grp);
1194 /* agg starts to be served, remove it from schedule */
1195 qfq_front_slot_remove(grp);
1197 new_front_agg = qfq_slot_scan(grp);
1199 if (new_front_agg == NULL) /* group is now inactive, remove from ER */
1200 __clear_bit(grp->index, &q->bitmaps[ER]);
1202 u64 roundedS = qfq_round_down(new_front_agg->S,
1206 if (grp->S == roundedS)
1209 grp->F = roundedS + (2ULL << grp->slot_shift);
1210 __clear_bit(grp->index, &q->bitmaps[ER]);
1211 s = qfq_calc_state(q, grp);
1212 __set_bit(grp->index, &q->bitmaps[s]);
1215 qfq_unblock_groups(q, grp->index, old_F);
1220 static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
1221 struct sk_buff **to_free)
1223 struct qfq_sched *q = qdisc_priv(sch);
1224 struct qfq_class *cl;
1225 struct qfq_aggregate *agg;
1228 cl = qfq_classify(skb, sch, &err);
1230 if (err & __NET_XMIT_BYPASS)
1231 qdisc_qstats_drop(sch);
1235 pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
1237 if (unlikely(cl->agg->lmax < qdisc_pkt_len(skb))) {
1238 pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
1239 cl->agg->lmax, qdisc_pkt_len(skb), cl->common.classid);
1240 err = qfq_change_agg(sch, cl, cl->agg->class_weight,
1241 qdisc_pkt_len(skb));
1244 return qdisc_drop(skb, sch, to_free);
1248 err = qdisc_enqueue(skb, cl->qdisc, to_free);
1249 if (unlikely(err != NET_XMIT_SUCCESS)) {
1250 pr_debug("qfq_enqueue: enqueue failed %d\n", err);
1251 if (net_xmit_drop_count(err)) {
1253 qdisc_qstats_drop(sch);
1258 bstats_update(&cl->bstats, skb);
1262 /* if the queue was not empty, then done here */
1263 if (cl->qdisc->q.qlen != 1) {
1264 if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) &&
1265 list_first_entry(&agg->active, struct qfq_class, alist)
1266 == cl && cl->deficit < qdisc_pkt_len(skb))
1267 list_move_tail(&cl->alist, &agg->active);
1272 /* schedule class for service within the aggregate */
1273 cl->deficit = agg->lmax;
1274 list_add_tail(&cl->alist, &agg->active);
1276 if (list_first_entry(&agg->active, struct qfq_class, alist) != cl ||
1277 q->in_serv_agg == agg)
1278 return err; /* non-empty or in service, nothing else to do */
1280 qfq_activate_agg(q, agg, enqueue);
1286 * Schedule aggregate according to its timestamps.
1288 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1290 struct qfq_group *grp = agg->grp;
1294 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1297 * Insert agg in the correct bucket.
1298 * If agg->S >= grp->S we don't need to adjust the
1299 * bucket list and simply go to the insertion phase.
1300 * Otherwise grp->S is decreasing, we must make room
1301 * in the bucket list, and also recompute the group state.
1302 * Finally, if there were no flows in this group and nobody
1303 * was in ER make sure to adjust V.
1305 if (grp->full_slots) {
1306 if (!qfq_gt(grp->S, agg->S))
1309 /* create a slot for this agg->S */
1310 qfq_slot_rotate(grp, roundedS);
1311 /* group was surely ineligible, remove */
1312 __clear_bit(grp->index, &q->bitmaps[IR]);
1313 __clear_bit(grp->index, &q->bitmaps[IB]);
1314 } else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) &&
1315 q->in_serv_agg == NULL)
1319 grp->F = roundedS + (2ULL << grp->slot_shift);
1320 s = qfq_calc_state(q, grp);
1321 __set_bit(grp->index, &q->bitmaps[s]);
1323 pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
1325 (unsigned long long) agg->S,
1326 (unsigned long long) agg->F,
1327 (unsigned long long) q->V);
1330 qfq_slot_insert(grp, agg, roundedS);
1334 /* Update agg ts and schedule agg for service */
1335 static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
1336 enum update_reason reason)
1338 agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */
1340 qfq_update_agg_ts(q, agg, reason);
1341 if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */
1342 q->in_serv_agg = agg; /* start serving this aggregate */
1343 /* update V: to be in service, agg must be eligible */
1344 q->oldV = q->V = agg->S;
1345 } else if (agg != q->in_serv_agg)
1346 qfq_schedule_agg(q, agg);
1349 static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
1350 struct qfq_aggregate *agg)
1352 unsigned int i, offset;
1355 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1356 offset = (roundedS - grp->S) >> grp->slot_shift;
1358 i = (grp->front + offset) % QFQ_MAX_SLOTS;
1360 hlist_del(&agg->next);
1361 if (hlist_empty(&grp->slots[i]))
1362 __clear_bit(offset, &grp->full_slots);
1366 * Called to forcibly deschedule an aggregate. If the aggregate is
1367 * not in the front bucket, or if the latter has other aggregates in
1368 * the front bucket, we can simply remove the aggregate with no other
1370 * Otherwise we must propagate the event up.
1372 static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1374 struct qfq_group *grp = agg->grp;
1379 if (agg == q->in_serv_agg) {
1380 charge_actual_service(agg);
1381 q->in_serv_agg = qfq_choose_next_agg(q);
1386 qfq_slot_remove(q, grp, agg);
1388 if (!grp->full_slots) {
1389 __clear_bit(grp->index, &q->bitmaps[IR]);
1390 __clear_bit(grp->index, &q->bitmaps[EB]);
1391 __clear_bit(grp->index, &q->bitmaps[IB]);
1393 if (test_bit(grp->index, &q->bitmaps[ER]) &&
1394 !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
1395 mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
1397 mask = ~((1UL << __fls(mask)) - 1);
1400 qfq_move_groups(q, mask, EB, ER);
1401 qfq_move_groups(q, mask, IB, IR);
1403 __clear_bit(grp->index, &q->bitmaps[ER]);
1404 } else if (hlist_empty(&grp->slots[grp->front])) {
1405 agg = qfq_slot_scan(grp);
1406 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1407 if (grp->S != roundedS) {
1408 __clear_bit(grp->index, &q->bitmaps[ER]);
1409 __clear_bit(grp->index, &q->bitmaps[IR]);
1410 __clear_bit(grp->index, &q->bitmaps[EB]);
1411 __clear_bit(grp->index, &q->bitmaps[IB]);
1413 grp->F = roundedS + (2ULL << grp->slot_shift);
1414 s = qfq_calc_state(q, grp);
1415 __set_bit(grp->index, &q->bitmaps[s]);
1420 static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1422 struct qfq_sched *q = qdisc_priv(sch);
1423 struct qfq_class *cl = (struct qfq_class *)arg;
1425 if (cl->qdisc->q.qlen == 0)
1426 qfq_deactivate_class(q, cl);
1429 static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
1431 struct qfq_sched *q = qdisc_priv(sch);
1432 struct qfq_group *grp;
1434 u32 max_cl_shift, maxbudg_shift, max_classes;
1436 err = qdisc_class_hash_init(&q->clhash);
1440 if (qdisc_dev(sch)->tx_queue_len + 1 > QFQ_MAX_AGG_CLASSES)
1441 max_classes = QFQ_MAX_AGG_CLASSES;
1443 max_classes = qdisc_dev(sch)->tx_queue_len + 1;
1444 /* max_cl_shift = floor(log_2(max_classes)) */
1445 max_cl_shift = __fls(max_classes);
1446 q->max_agg_classes = 1<<max_cl_shift;
1448 /* maxbudg_shift = log2(max_len * max_classes_per_agg) */
1449 maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift;
1450 q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX;
1452 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1453 grp = &q->groups[i];
1455 grp->slot_shift = q->min_slot_shift + i;
1456 for (j = 0; j < QFQ_MAX_SLOTS; j++)
1457 INIT_HLIST_HEAD(&grp->slots[j]);
1460 INIT_HLIST_HEAD(&q->nonfull_aggs);
1465 static void qfq_reset_qdisc(struct Qdisc *sch)
1467 struct qfq_sched *q = qdisc_priv(sch);
1468 struct qfq_class *cl;
1471 for (i = 0; i < q->clhash.hashsize; i++) {
1472 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1473 if (cl->qdisc->q.qlen > 0)
1474 qfq_deactivate_class(q, cl);
1476 qdisc_reset(cl->qdisc);
1482 static void qfq_destroy_qdisc(struct Qdisc *sch)
1484 struct qfq_sched *q = qdisc_priv(sch);
1485 struct qfq_class *cl;
1486 struct hlist_node *next;
1489 tcf_destroy_chain(&q->filter_list);
1491 for (i = 0; i < q->clhash.hashsize; i++) {
1492 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1494 qfq_destroy_class(sch, cl);
1497 qdisc_class_hash_destroy(&q->clhash);
1500 static const struct Qdisc_class_ops qfq_class_ops = {
1501 .change = qfq_change_class,
1502 .delete = qfq_delete_class,
1503 .get = qfq_get_class,
1504 .put = qfq_put_class,
1505 .tcf_chain = qfq_tcf_chain,
1506 .bind_tcf = qfq_bind_tcf,
1507 .unbind_tcf = qfq_unbind_tcf,
1508 .graft = qfq_graft_class,
1509 .leaf = qfq_class_leaf,
1510 .qlen_notify = qfq_qlen_notify,
1511 .dump = qfq_dump_class,
1512 .dump_stats = qfq_dump_class_stats,
1516 static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1517 .cl_ops = &qfq_class_ops,
1519 .priv_size = sizeof(struct qfq_sched),
1520 .enqueue = qfq_enqueue,
1521 .dequeue = qfq_dequeue,
1522 .peek = qdisc_peek_dequeued,
1523 .init = qfq_init_qdisc,
1524 .reset = qfq_reset_qdisc,
1525 .destroy = qfq_destroy_qdisc,
1526 .owner = THIS_MODULE,
1529 static int __init qfq_init(void)
1531 return register_qdisc(&qfq_qdisc_ops);
1534 static void __exit qfq_exit(void)
1536 unregister_qdisc(&qfq_qdisc_ops);
1539 module_init(qfq_init);
1540 module_exit(qfq_exit);
1541 MODULE_LICENSE("GPL");