2 * net/sched/sch_qfq.c Quick Fair Queueing Scheduler.
4 * Copyright (c) 2009 Fabio Checconi, Luigi Rizzo, and Paolo Valente.
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * version 2 as published by the Free Software Foundation.
11 #include <linux/module.h>
12 #include <linux/init.h>
13 #include <linux/bitops.h>
14 #include <linux/errno.h>
15 #include <linux/netdevice.h>
16 #include <linux/pkt_sched.h>
17 #include <net/sch_generic.h>
18 #include <net/pkt_sched.h>
19 #include <net/pkt_cls.h>
22 /* Quick Fair Queueing
27 Fabio Checconi, Luigi Rizzo, and Paolo Valente: "QFQ: Efficient
28 Packet Scheduling with Tight Bandwidth Distribution Guarantees."
31 http://retis.sssup.it/~fabio/linux/qfq/
36 Virtual time computations.
38 S, F and V are all computed in fixed point arithmetic with
39 FRAC_BITS decimal bits.
41 QFQ_MAX_INDEX is the maximum index allowed for a group. We need
43 QFQ_MAX_WSHIFT is the maximum power of two supported as a weight.
45 The layout of the bits is as below:
47 [ MTU_SHIFT ][ FRAC_BITS ]
48 [ MAX_INDEX ][ MIN_SLOT_SHIFT ]
52 where MIN_SLOT_SHIFT is derived by difference from the others.
54 The max group index corresponds to Lmax/w_min, where
55 Lmax=1<<MTU_SHIFT, w_min = 1 .
56 From this, and knowing how many groups (MAX_INDEX) we want,
57 we can derive the shift corresponding to each group.
59 Because we often need to compute
60 F = S + len/w_i and V = V + len/wsum
61 instead of storing w_i store the value
62 inv_w = (1<<FRAC_BITS)/w_i
63 so we can do F = S + len * inv_w * wsum.
64 We use W_TOT in the formulas so we can easily move between
65 static and adaptive weight sum.
67 The per-scheduler-instance data contain all the data structures
68 for the scheduler: bitmaps and bucket lists.
73 * Maximum number of consecutive slots occupied by backlogged classes
76 #define QFQ_MAX_SLOTS 32
79 * Shifts used for class<->group mapping. We allow class weights that are
80 * in the range [1, 2^MAX_WSHIFT], and we try to map each class i to the
81 * group with the smallest index that can support the L_i / r_i configured
84 * grp->index is the index of the group; and grp->slot_shift
85 * is the shift for the corresponding (scaled) sigma_i.
87 #define QFQ_MAX_INDEX 19
88 #define QFQ_MAX_WSHIFT 16
90 #define QFQ_MAX_WEIGHT (1<<QFQ_MAX_WSHIFT)
91 #define QFQ_MAX_WSUM (2*QFQ_MAX_WEIGHT)
93 #define FRAC_BITS 30 /* fixed point arithmetic */
94 #define ONE_FP (1UL << FRAC_BITS)
95 #define IWSUM (ONE_FP/QFQ_MAX_WSUM)
97 #define QFQ_MTU_SHIFT 11
98 #define QFQ_MIN_SLOT_SHIFT (FRAC_BITS + QFQ_MTU_SHIFT - QFQ_MAX_INDEX)
101 * Possible group states. These values are used as indexes for the bitmaps
102 * array of struct qfq_queue.
104 enum qfq_state { ER, IR, EB, IB, QFQ_MAX_STATE };
109 struct Qdisc_class_common common;
112 unsigned int filter_cnt;
114 struct gnet_stats_basic_packed bstats;
115 struct gnet_stats_queue qstats;
116 struct gnet_stats_rate_est rate_est;
119 struct hlist_node next; /* Link for the slot list. */
120 u64 S, F; /* flow timestamps (exact) */
122 /* group we belong to. In principle we would need the index,
123 * which is log_2(lmax/weight), but we never reference it
124 * directly, only the group.
126 struct qfq_group *grp;
128 /* these are copied from the flowset. */
129 u32 inv_w; /* ONE_FP/weight */
130 u32 lmax; /* Max packet size for this flow. */
134 u64 S, F; /* group timestamps (approx). */
135 unsigned int slot_shift; /* Slot shift. */
136 unsigned int index; /* Group index. */
137 unsigned int front; /* Index of the front slot. */
138 unsigned long full_slots; /* non-empty slots */
140 /* Array of RR lists of active classes. */
141 struct hlist_head slots[QFQ_MAX_SLOTS];
145 struct tcf_proto *filter_list;
146 struct Qdisc_class_hash clhash;
148 u64 V; /* Precise virtual time. */
149 u32 wsum; /* weight sum */
151 unsigned long bitmaps[QFQ_MAX_STATE]; /* Group bitmaps. */
152 struct qfq_group groups[QFQ_MAX_INDEX + 1]; /* The groups. */
155 static struct qfq_class *qfq_find_class(struct Qdisc *sch, u32 classid)
157 struct qfq_sched *q = qdisc_priv(sch);
158 struct Qdisc_class_common *clc;
160 clc = qdisc_class_find(&q->clhash, classid);
163 return container_of(clc, struct qfq_class, common);
166 static void qfq_purge_queue(struct qfq_class *cl)
168 unsigned int len = cl->qdisc->q.qlen;
170 qdisc_reset(cl->qdisc);
171 qdisc_tree_decrease_qlen(cl->qdisc, len);
174 static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = {
175 [TCA_QFQ_WEIGHT] = { .type = NLA_U32 },
176 [TCA_QFQ_LMAX] = { .type = NLA_U32 },
180 * Calculate a flow index, given its weight and maximum packet length.
181 * index = log_2(maxlen/weight) but we need to apply the scaling.
182 * This is used only once at flow creation.
184 static int qfq_calc_index(u32 inv_w, unsigned int maxlen)
186 u64 slot_size = (u64)maxlen * inv_w;
187 unsigned long size_map;
190 size_map = slot_size >> QFQ_MIN_SLOT_SHIFT;
194 index = __fls(size_map) + 1; /* basically a log_2 */
195 index -= !(slot_size - (1ULL << (index + QFQ_MIN_SLOT_SHIFT - 1)));
200 pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n",
201 (unsigned long) ONE_FP/inv_w, maxlen, index);
206 /* Length of the next packet (0 if the queue is empty). */
207 static unsigned int qdisc_peek_len(struct Qdisc *sch)
211 skb = sch->ops->peek(sch);
212 return skb ? qdisc_pkt_len(skb) : 0;
215 static void qfq_deactivate_class(struct qfq_sched *, struct qfq_class *);
216 static void qfq_activate_class(struct qfq_sched *q, struct qfq_class *cl,
219 static void qfq_update_class_params(struct qfq_sched *q, struct qfq_class *cl,
220 u32 lmax, u32 inv_w, int delta_w)
224 /* update qfq-specific data */
227 i = qfq_calc_index(cl->inv_w, cl->lmax);
229 cl->grp = &q->groups[i];
234 static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
235 struct nlattr **tca, unsigned long *arg)
237 struct qfq_sched *q = qdisc_priv(sch);
238 struct qfq_class *cl = (struct qfq_class *)*arg;
239 struct nlattr *tb[TCA_QFQ_MAX + 1];
240 u32 weight, lmax, inv_w;
244 if (tca[TCA_OPTIONS] == NULL) {
245 pr_notice("qfq: no options\n");
249 err = nla_parse_nested(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS], qfq_policy);
253 if (tb[TCA_QFQ_WEIGHT]) {
254 weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]);
255 if (!weight || weight > (1UL << QFQ_MAX_WSHIFT)) {
256 pr_notice("qfq: invalid weight %u\n", weight);
262 inv_w = ONE_FP / weight;
263 weight = ONE_FP / inv_w;
264 delta_w = weight - (cl ? ONE_FP / cl->inv_w : 0);
265 if (q->wsum + delta_w > QFQ_MAX_WSUM) {
266 pr_notice("qfq: total weight out of range (%u + %u)\n",
271 if (tb[TCA_QFQ_LMAX]) {
272 lmax = nla_get_u32(tb[TCA_QFQ_LMAX]);
273 if (!lmax || lmax > (1UL << QFQ_MTU_SHIFT)) {
274 pr_notice("qfq: invalid max length %u\n", lmax);
278 lmax = 1UL << QFQ_MTU_SHIFT;
281 bool need_reactivation = false;
284 err = gen_replace_estimator(&cl->bstats, &cl->rate_est,
285 qdisc_root_sleeping_lock(sch),
291 if (lmax == cl->lmax && inv_w == cl->inv_w)
292 return 0; /* nothing to update */
294 i = qfq_calc_index(inv_w, lmax);
296 if (&q->groups[i] != cl->grp && cl->qdisc->q.qlen > 0) {
298 * shift cl->F back, to not charge the
299 * class for the not-yet-served head
303 /* remove class from its slot in the old group */
304 qfq_deactivate_class(q, cl);
305 need_reactivation = true;
308 qfq_update_class_params(q, cl, lmax, inv_w, delta_w);
310 if (need_reactivation) /* activate in new group */
311 qfq_activate_class(q, cl, qdisc_peek_len(cl->qdisc));
312 sch_tree_unlock(sch);
317 cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
322 cl->common.classid = classid;
324 qfq_update_class_params(q, cl, lmax, inv_w, delta_w);
326 cl->qdisc = qdisc_create_dflt(sch->dev_queue,
327 &pfifo_qdisc_ops, classid);
328 if (cl->qdisc == NULL)
329 cl->qdisc = &noop_qdisc;
332 err = gen_new_estimator(&cl->bstats, &cl->rate_est,
333 qdisc_root_sleeping_lock(sch),
336 qdisc_destroy(cl->qdisc);
343 qdisc_class_hash_insert(&q->clhash, &cl->common);
344 sch_tree_unlock(sch);
346 qdisc_class_hash_grow(sch, &q->clhash);
348 *arg = (unsigned long)cl;
352 static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
354 struct qfq_sched *q = qdisc_priv(sch);
357 q->wsum -= ONE_FP / cl->inv_w;
361 gen_kill_estimator(&cl->bstats, &cl->rate_est);
362 qdisc_destroy(cl->qdisc);
366 static int qfq_delete_class(struct Qdisc *sch, unsigned long arg)
368 struct qfq_sched *q = qdisc_priv(sch);
369 struct qfq_class *cl = (struct qfq_class *)arg;
371 if (cl->filter_cnt > 0)
377 qdisc_class_hash_remove(&q->clhash, &cl->common);
379 BUG_ON(--cl->refcnt == 0);
381 * This shouldn't happen: we "hold" one cops->get() when called
382 * from tc_ctl_tclass; the destroy method is done from cops->put().
385 sch_tree_unlock(sch);
389 static unsigned long qfq_get_class(struct Qdisc *sch, u32 classid)
391 struct qfq_class *cl = qfq_find_class(sch, classid);
396 return (unsigned long)cl;
399 static void qfq_put_class(struct Qdisc *sch, unsigned long arg)
401 struct qfq_class *cl = (struct qfq_class *)arg;
403 if (--cl->refcnt == 0)
404 qfq_destroy_class(sch, cl);
407 static struct tcf_proto **qfq_tcf_chain(struct Qdisc *sch, unsigned long cl)
409 struct qfq_sched *q = qdisc_priv(sch);
414 return &q->filter_list;
417 static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
420 struct qfq_class *cl = qfq_find_class(sch, classid);
425 return (unsigned long)cl;
428 static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
430 struct qfq_class *cl = (struct qfq_class *)arg;
435 static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
436 struct Qdisc *new, struct Qdisc **old)
438 struct qfq_class *cl = (struct qfq_class *)arg;
441 new = qdisc_create_dflt(sch->dev_queue,
442 &pfifo_qdisc_ops, cl->common.classid);
451 sch_tree_unlock(sch);
455 static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
457 struct qfq_class *cl = (struct qfq_class *)arg;
462 static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
463 struct sk_buff *skb, struct tcmsg *tcm)
465 struct qfq_class *cl = (struct qfq_class *)arg;
468 tcm->tcm_parent = TC_H_ROOT;
469 tcm->tcm_handle = cl->common.classid;
470 tcm->tcm_info = cl->qdisc->handle;
472 nest = nla_nest_start(skb, TCA_OPTIONS);
474 goto nla_put_failure;
475 if (nla_put_u32(skb, TCA_QFQ_WEIGHT, ONE_FP/cl->inv_w) ||
476 nla_put_u32(skb, TCA_QFQ_LMAX, cl->lmax))
477 goto nla_put_failure;
478 return nla_nest_end(skb, nest);
481 nla_nest_cancel(skb, nest);
485 static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
488 struct qfq_class *cl = (struct qfq_class *)arg;
489 struct tc_qfq_stats xstats;
491 memset(&xstats, 0, sizeof(xstats));
492 cl->qdisc->qstats.qlen = cl->qdisc->q.qlen;
494 xstats.weight = ONE_FP/cl->inv_w;
495 xstats.lmax = cl->lmax;
497 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
498 gnet_stats_copy_rate_est(d, &cl->bstats, &cl->rate_est) < 0 ||
499 gnet_stats_copy_queue(d, &cl->qdisc->qstats) < 0)
502 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
505 static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
507 struct qfq_sched *q = qdisc_priv(sch);
508 struct qfq_class *cl;
509 struct hlist_node *n;
515 for (i = 0; i < q->clhash.hashsize; i++) {
516 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode) {
517 if (arg->count < arg->skip) {
521 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
530 static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
533 struct qfq_sched *q = qdisc_priv(sch);
534 struct qfq_class *cl;
535 struct tcf_result res;
538 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
539 pr_debug("qfq_classify: found %d\n", skb->priority);
540 cl = qfq_find_class(sch, skb->priority);
545 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
546 result = tc_classify(skb, q->filter_list, &res);
548 #ifdef CONFIG_NET_CLS_ACT
552 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
557 cl = (struct qfq_class *)res.class;
559 cl = qfq_find_class(sch, res.classid);
566 /* Generic comparison function, handling wraparound. */
567 static inline int qfq_gt(u64 a, u64 b)
569 return (s64)(a - b) > 0;
572 /* Round a precise timestamp to its slotted value. */
573 static inline u64 qfq_round_down(u64 ts, unsigned int shift)
575 return ts & ~((1ULL << shift) - 1);
578 /* return the pointer to the group with lowest index in the bitmap */
579 static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
580 unsigned long bitmap)
582 int index = __ffs(bitmap);
583 return &q->groups[index];
585 /* Calculate a mask to mimic what would be ffs_from(). */
586 static inline unsigned long mask_from(unsigned long bitmap, int from)
588 return bitmap & ~((1UL << from) - 1);
592 * The state computation relies on ER=0, IR=1, EB=2, IB=3
593 * First compute eligibility comparing grp->S, q->V,
594 * then check if someone is blocking us and possibly add EB
596 static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
598 /* if S > V we are not eligible */
599 unsigned int state = qfq_gt(grp->S, q->V);
600 unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
601 struct qfq_group *next;
604 next = qfq_ffs(q, mask);
605 if (qfq_gt(grp->F, next->F))
615 * q->bitmaps[dst] |= q->bitmaps[src] & mask;
616 * q->bitmaps[src] &= ~mask;
617 * but we should make sure that src != dst
619 static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
622 q->bitmaps[dst] |= q->bitmaps[src] & mask;
623 q->bitmaps[src] &= ~mask;
626 static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
628 unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
629 struct qfq_group *next;
632 next = qfq_ffs(q, mask);
633 if (!qfq_gt(next->F, old_F))
637 mask = (1UL << index) - 1;
638 qfq_move_groups(q, mask, EB, ER);
639 qfq_move_groups(q, mask, IB, IR);
646 old_V >>= QFQ_MIN_SLOT_SHIFT;
652 static void qfq_make_eligible(struct qfq_sched *q, u64 old_V)
654 unsigned long vslot = q->V >> QFQ_MIN_SLOT_SHIFT;
655 unsigned long old_vslot = old_V >> QFQ_MIN_SLOT_SHIFT;
657 if (vslot != old_vslot) {
658 unsigned long mask = (1UL << fls(vslot ^ old_vslot)) - 1;
659 qfq_move_groups(q, mask, IR, ER);
660 qfq_move_groups(q, mask, IB, EB);
666 * XXX we should make sure that slot becomes less than 32.
667 * This is guaranteed by the input values.
668 * roundedS is always cl->S rounded on grp->slot_shift bits.
670 static void qfq_slot_insert(struct qfq_group *grp, struct qfq_class *cl,
673 u64 slot = (roundedS - grp->S) >> grp->slot_shift;
674 unsigned int i = (grp->front + slot) % QFQ_MAX_SLOTS;
676 hlist_add_head(&cl->next, &grp->slots[i]);
677 __set_bit(slot, &grp->full_slots);
680 /* Maybe introduce hlist_first_entry?? */
681 static struct qfq_class *qfq_slot_head(struct qfq_group *grp)
683 return hlist_entry(grp->slots[grp->front].first,
684 struct qfq_class, next);
688 * remove the entry from the slot
690 static void qfq_front_slot_remove(struct qfq_group *grp)
692 struct qfq_class *cl = qfq_slot_head(grp);
695 hlist_del(&cl->next);
696 if (hlist_empty(&grp->slots[grp->front]))
697 __clear_bit(0, &grp->full_slots);
701 * Returns the first full queue in a group. As a side effect,
702 * adjust the bucket list so the first non-empty bucket is at
703 * position 0 in full_slots.
705 static struct qfq_class *qfq_slot_scan(struct qfq_group *grp)
709 pr_debug("qfq slot_scan: grp %u full %#lx\n",
710 grp->index, grp->full_slots);
712 if (grp->full_slots == 0)
715 i = __ffs(grp->full_slots); /* zero based */
717 grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
718 grp->full_slots >>= i;
721 return qfq_slot_head(grp);
725 * adjust the bucket list. When the start time of a group decreases,
726 * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
727 * move the objects. The mask of occupied slots must be shifted
728 * because we use ffs() to find the first non-empty slot.
729 * This covers decreases in the group's start time, but what about
730 * increases of the start time ?
731 * Here too we should make sure that i is less than 32
733 static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
735 unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
737 grp->full_slots <<= i;
738 grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
741 static void qfq_update_eligible(struct qfq_sched *q, u64 old_V)
743 struct qfq_group *grp;
744 unsigned long ineligible;
746 ineligible = q->bitmaps[IR] | q->bitmaps[IB];
748 if (!q->bitmaps[ER]) {
749 grp = qfq_ffs(q, ineligible);
750 if (qfq_gt(grp->S, q->V))
753 qfq_make_eligible(q, old_V);
758 * Updates the class, returns true if also the group needs to be updated.
760 static bool qfq_update_class(struct qfq_group *grp, struct qfq_class *cl)
762 unsigned int len = qdisc_peek_len(cl->qdisc);
766 qfq_front_slot_remove(grp); /* queue is empty */
770 cl->F = cl->S + (u64)len * cl->inv_w;
771 roundedS = qfq_round_down(cl->S, grp->slot_shift);
772 if (roundedS == grp->S)
775 qfq_front_slot_remove(grp);
776 qfq_slot_insert(grp, cl, roundedS);
782 static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
784 struct qfq_sched *q = qdisc_priv(sch);
785 struct qfq_group *grp;
786 struct qfq_class *cl;
794 grp = qfq_ffs(q, q->bitmaps[ER]);
796 cl = qfq_slot_head(grp);
797 skb = qdisc_dequeue_peeked(cl->qdisc);
799 WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
804 qdisc_bstats_update(sch, skb);
807 len = qdisc_pkt_len(skb);
808 q->V += (u64)len * IWSUM;
809 pr_debug("qfq dequeue: len %u F %lld now %lld\n",
810 len, (unsigned long long) cl->F, (unsigned long long) q->V);
812 if (qfq_update_class(grp, cl)) {
815 cl = qfq_slot_scan(grp);
817 __clear_bit(grp->index, &q->bitmaps[ER]);
819 u64 roundedS = qfq_round_down(cl->S, grp->slot_shift);
822 if (grp->S == roundedS)
825 grp->F = roundedS + (2ULL << grp->slot_shift);
826 __clear_bit(grp->index, &q->bitmaps[ER]);
827 s = qfq_calc_state(q, grp);
828 __set_bit(grp->index, &q->bitmaps[s]);
831 qfq_unblock_groups(q, grp->index, old_F);
835 qfq_update_eligible(q, old_V);
841 * Assign a reasonable start time for a new flow k in group i.
842 * Admissible values for \hat(F) are multiples of \sigma_i
843 * no greater than V+\sigma_i . Larger values mean that
844 * we had a wraparound so we consider the timestamp to be stale.
846 * If F is not stale and F >= V then we set S = F.
847 * Otherwise we should assign S = V, but this may violate
848 * the ordering in ER. So, if we have groups in ER, set S to
849 * the F_j of the first group j which would be blocking us.
850 * We are guaranteed not to move S backward because
851 * otherwise our group i would still be blocked.
853 static void qfq_update_start(struct qfq_sched *q, struct qfq_class *cl)
857 int slot_shift = cl->grp->slot_shift;
859 roundedF = qfq_round_down(cl->F, slot_shift);
860 limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
862 if (!qfq_gt(cl->F, q->V) || qfq_gt(roundedF, limit)) {
863 /* timestamp was stale */
864 mask = mask_from(q->bitmaps[ER], cl->grp->index);
866 struct qfq_group *next = qfq_ffs(q, mask);
867 if (qfq_gt(roundedF, next->F)) {
873 } else /* timestamp is not stale */
877 static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
879 struct qfq_sched *q = qdisc_priv(sch);
880 struct qfq_class *cl;
883 cl = qfq_classify(skb, sch, &err);
885 if (err & __NET_XMIT_BYPASS)
890 pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
892 err = qdisc_enqueue(skb, cl->qdisc);
893 if (unlikely(err != NET_XMIT_SUCCESS)) {
894 pr_debug("qfq_enqueue: enqueue failed %d\n", err);
895 if (net_xmit_drop_count(err)) {
902 bstats_update(&cl->bstats, skb);
905 /* If the new skb is not the head of queue, then done here. */
906 if (cl->qdisc->q.qlen != 1)
909 /* If reach this point, queue q was idle */
910 qfq_activate_class(q, cl, qdisc_pkt_len(skb));
916 * Handle class switch from idle to backlogged.
918 static void qfq_activate_class(struct qfq_sched *q, struct qfq_class *cl,
919 unsigned int pkt_len)
921 struct qfq_group *grp = cl->grp;
925 qfq_update_start(q, cl);
927 /* compute new finish time and rounded start. */
928 cl->F = cl->S + (u64)pkt_len * cl->inv_w;
929 roundedS = qfq_round_down(cl->S, grp->slot_shift);
932 * insert cl in the correct bucket.
933 * If cl->S >= grp->S we don't need to adjust the
934 * bucket list and simply go to the insertion phase.
935 * Otherwise grp->S is decreasing, we must make room
936 * in the bucket list, and also recompute the group state.
937 * Finally, if there were no flows in this group and nobody
938 * was in ER make sure to adjust V.
940 if (grp->full_slots) {
941 if (!qfq_gt(grp->S, cl->S))
944 /* create a slot for this cl->S */
945 qfq_slot_rotate(grp, roundedS);
946 /* group was surely ineligible, remove */
947 __clear_bit(grp->index, &q->bitmaps[IR]);
948 __clear_bit(grp->index, &q->bitmaps[IB]);
949 } else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V))
953 grp->F = roundedS + (2ULL << grp->slot_shift);
954 s = qfq_calc_state(q, grp);
955 __set_bit(grp->index, &q->bitmaps[s]);
957 pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
959 (unsigned long long) cl->S,
960 (unsigned long long) cl->F,
961 (unsigned long long) q->V);
964 qfq_slot_insert(grp, cl, roundedS);
968 static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
969 struct qfq_class *cl)
971 unsigned int i, offset;
974 roundedS = qfq_round_down(cl->S, grp->slot_shift);
975 offset = (roundedS - grp->S) >> grp->slot_shift;
976 i = (grp->front + offset) % QFQ_MAX_SLOTS;
978 hlist_del(&cl->next);
979 if (hlist_empty(&grp->slots[i]))
980 __clear_bit(offset, &grp->full_slots);
984 * called to forcibly destroy a queue.
985 * If the queue is not in the front bucket, or if it has
986 * other queues in the front bucket, we can simply remove
987 * the queue with no other side effects.
988 * Otherwise we must propagate the event up.
990 static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
992 struct qfq_group *grp = cl->grp;
998 qfq_slot_remove(q, grp, cl);
1000 if (!grp->full_slots) {
1001 __clear_bit(grp->index, &q->bitmaps[IR]);
1002 __clear_bit(grp->index, &q->bitmaps[EB]);
1003 __clear_bit(grp->index, &q->bitmaps[IB]);
1005 if (test_bit(grp->index, &q->bitmaps[ER]) &&
1006 !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
1007 mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
1009 mask = ~((1UL << __fls(mask)) - 1);
1012 qfq_move_groups(q, mask, EB, ER);
1013 qfq_move_groups(q, mask, IB, IR);
1015 __clear_bit(grp->index, &q->bitmaps[ER]);
1016 } else if (hlist_empty(&grp->slots[grp->front])) {
1017 cl = qfq_slot_scan(grp);
1018 roundedS = qfq_round_down(cl->S, grp->slot_shift);
1019 if (grp->S != roundedS) {
1020 __clear_bit(grp->index, &q->bitmaps[ER]);
1021 __clear_bit(grp->index, &q->bitmaps[IR]);
1022 __clear_bit(grp->index, &q->bitmaps[EB]);
1023 __clear_bit(grp->index, &q->bitmaps[IB]);
1025 grp->F = roundedS + (2ULL << grp->slot_shift);
1026 s = qfq_calc_state(q, grp);
1027 __set_bit(grp->index, &q->bitmaps[s]);
1031 qfq_update_eligible(q, q->V);
1034 static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1036 struct qfq_sched *q = qdisc_priv(sch);
1037 struct qfq_class *cl = (struct qfq_class *)arg;
1039 if (cl->qdisc->q.qlen == 0)
1040 qfq_deactivate_class(q, cl);
1043 static unsigned int qfq_drop(struct Qdisc *sch)
1045 struct qfq_sched *q = qdisc_priv(sch);
1046 struct qfq_group *grp;
1047 unsigned int i, j, len;
1049 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1050 grp = &q->groups[i];
1051 for (j = 0; j < QFQ_MAX_SLOTS; j++) {
1052 struct qfq_class *cl;
1053 struct hlist_node *n;
1055 hlist_for_each_entry(cl, n, &grp->slots[j], next) {
1057 if (!cl->qdisc->ops->drop)
1060 len = cl->qdisc->ops->drop(cl->qdisc);
1063 if (!cl->qdisc->q.qlen)
1064 qfq_deactivate_class(q, cl);
1075 static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
1077 struct qfq_sched *q = qdisc_priv(sch);
1078 struct qfq_group *grp;
1081 err = qdisc_class_hash_init(&q->clhash);
1085 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1086 grp = &q->groups[i];
1088 grp->slot_shift = QFQ_MTU_SHIFT + FRAC_BITS
1089 - (QFQ_MAX_INDEX - i);
1090 for (j = 0; j < QFQ_MAX_SLOTS; j++)
1091 INIT_HLIST_HEAD(&grp->slots[j]);
1097 static void qfq_reset_qdisc(struct Qdisc *sch)
1099 struct qfq_sched *q = qdisc_priv(sch);
1100 struct qfq_group *grp;
1101 struct qfq_class *cl;
1102 struct hlist_node *n, *tmp;
1105 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1106 grp = &q->groups[i];
1107 for (j = 0; j < QFQ_MAX_SLOTS; j++) {
1108 hlist_for_each_entry_safe(cl, n, tmp,
1109 &grp->slots[j], next) {
1110 qfq_deactivate_class(q, cl);
1115 for (i = 0; i < q->clhash.hashsize; i++) {
1116 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode)
1117 qdisc_reset(cl->qdisc);
1122 static void qfq_destroy_qdisc(struct Qdisc *sch)
1124 struct qfq_sched *q = qdisc_priv(sch);
1125 struct qfq_class *cl;
1126 struct hlist_node *n, *next;
1129 tcf_destroy_chain(&q->filter_list);
1131 for (i = 0; i < q->clhash.hashsize; i++) {
1132 hlist_for_each_entry_safe(cl, n, next, &q->clhash.hash[i],
1134 qfq_destroy_class(sch, cl);
1137 qdisc_class_hash_destroy(&q->clhash);
1140 static const struct Qdisc_class_ops qfq_class_ops = {
1141 .change = qfq_change_class,
1142 .delete = qfq_delete_class,
1143 .get = qfq_get_class,
1144 .put = qfq_put_class,
1145 .tcf_chain = qfq_tcf_chain,
1146 .bind_tcf = qfq_bind_tcf,
1147 .unbind_tcf = qfq_unbind_tcf,
1148 .graft = qfq_graft_class,
1149 .leaf = qfq_class_leaf,
1150 .qlen_notify = qfq_qlen_notify,
1151 .dump = qfq_dump_class,
1152 .dump_stats = qfq_dump_class_stats,
1156 static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1157 .cl_ops = &qfq_class_ops,
1159 .priv_size = sizeof(struct qfq_sched),
1160 .enqueue = qfq_enqueue,
1161 .dequeue = qfq_dequeue,
1162 .peek = qdisc_peek_dequeued,
1164 .init = qfq_init_qdisc,
1165 .reset = qfq_reset_qdisc,
1166 .destroy = qfq_destroy_qdisc,
1167 .owner = THIS_MODULE,
1170 static int __init qfq_init(void)
1172 return register_qdisc(&qfq_qdisc_ops);
1175 static void __exit qfq_exit(void)
1177 unregister_qdisc(&qfq_qdisc_ops);
1180 module_init(qfq_init);
1181 module_exit(qfq_exit);
1182 MODULE_LICENSE("GPL");