Loading...
1/*
2 * net/sched/sch_qfq.c Quick Fair Queueing Scheduler.
3 *
4 * Copyright (c) 2009 Fabio Checconi, Luigi Rizzo, and Paolo Valente.
5 *
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.
9 */
10
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>
20
21
22/* Quick Fair Queueing
23 ===================
24
25 Sources:
26
27 Fabio Checconi, Luigi Rizzo, and Paolo Valente: "QFQ: Efficient
28 Packet Scheduling with Tight Bandwidth Distribution Guarantees."
29
30 See also:
31 http://retis.sssup.it/~fabio/linux/qfq/
32 */
33
34/*
35
36 Virtual time computations.
37
38 S, F and V are all computed in fixed point arithmetic with
39 FRAC_BITS decimal bits.
40
41 QFQ_MAX_INDEX is the maximum index allowed for a group. We need
42 one bit per index.
43 QFQ_MAX_WSHIFT is the maximum power of two supported as a weight.
44
45 The layout of the bits is as below:
46
47 [ MTU_SHIFT ][ FRAC_BITS ]
48 [ MAX_INDEX ][ MIN_SLOT_SHIFT ]
49 ^.__grp->index = 0
50 *.__grp->slot_shift
51
52 where MIN_SLOT_SHIFT is derived by difference from the others.
53
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.
58
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.
66
67 The per-scheduler-instance data contain all the data structures
68 for the scheduler: bitmaps and bucket lists.
69
70 */
71
72/*
73 * Maximum number of consecutive slots occupied by backlogged classes
74 * inside a group.
75 */
76#define QFQ_MAX_SLOTS 32
77
78/*
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
82 * for the class.
83 *
84 * grp->index is the index of the group; and grp->slot_shift
85 * is the shift for the corresponding (scaled) sigma_i.
86 */
87#define QFQ_MAX_INDEX 19
88#define QFQ_MAX_WSHIFT 16
89
90#define QFQ_MAX_WEIGHT (1<<QFQ_MAX_WSHIFT)
91#define QFQ_MAX_WSUM (2*QFQ_MAX_WEIGHT)
92
93#define FRAC_BITS 30 /* fixed point arithmetic */
94#define ONE_FP (1UL << FRAC_BITS)
95#define IWSUM (ONE_FP/QFQ_MAX_WSUM)
96
97#define QFQ_MTU_SHIFT 11
98#define QFQ_MIN_SLOT_SHIFT (FRAC_BITS + QFQ_MTU_SHIFT - QFQ_MAX_INDEX)
99
100/*
101 * Possible group states. These values are used as indexes for the bitmaps
102 * array of struct qfq_queue.
103 */
104enum qfq_state { ER, IR, EB, IB, QFQ_MAX_STATE };
105
106struct qfq_group;
107
108struct qfq_class {
109 struct Qdisc_class_common common;
110
111 unsigned int refcnt;
112 unsigned int filter_cnt;
113
114 struct gnet_stats_basic_packed bstats;
115 struct gnet_stats_queue qstats;
116 struct gnet_stats_rate_est rate_est;
117 struct Qdisc *qdisc;
118
119 struct hlist_node next; /* Link for the slot list. */
120 u64 S, F; /* flow timestamps (exact) */
121
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.
125 */
126 struct qfq_group *grp;
127
128 /* these are copied from the flowset. */
129 u32 inv_w; /* ONE_FP/weight */
130 u32 lmax; /* Max packet size for this flow. */
131};
132
133struct qfq_group {
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 */
139
140 /* Array of RR lists of active classes. */
141 struct hlist_head slots[QFQ_MAX_SLOTS];
142};
143
144struct qfq_sched {
145 struct tcf_proto *filter_list;
146 struct Qdisc_class_hash clhash;
147
148 u64 V; /* Precise virtual time. */
149 u32 wsum; /* weight sum */
150
151 unsigned long bitmaps[QFQ_MAX_STATE]; /* Group bitmaps. */
152 struct qfq_group groups[QFQ_MAX_INDEX + 1]; /* The groups. */
153};
154
155static struct qfq_class *qfq_find_class(struct Qdisc *sch, u32 classid)
156{
157 struct qfq_sched *q = qdisc_priv(sch);
158 struct Qdisc_class_common *clc;
159
160 clc = qdisc_class_find(&q->clhash, classid);
161 if (clc == NULL)
162 return NULL;
163 return container_of(clc, struct qfq_class, common);
164}
165
166static void qfq_purge_queue(struct qfq_class *cl)
167{
168 unsigned int len = cl->qdisc->q.qlen;
169
170 qdisc_reset(cl->qdisc);
171 qdisc_tree_decrease_qlen(cl->qdisc, len);
172}
173
174static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = {
175 [TCA_QFQ_WEIGHT] = { .type = NLA_U32 },
176 [TCA_QFQ_LMAX] = { .type = NLA_U32 },
177};
178
179/*
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.
183 */
184static int qfq_calc_index(u32 inv_w, unsigned int maxlen)
185{
186 u64 slot_size = (u64)maxlen * inv_w;
187 unsigned long size_map;
188 int index = 0;
189
190 size_map = slot_size >> QFQ_MIN_SLOT_SHIFT;
191 if (!size_map)
192 goto out;
193
194 index = __fls(size_map) + 1; /* basically a log_2 */
195 index -= !(slot_size - (1ULL << (index + QFQ_MIN_SLOT_SHIFT - 1)));
196
197 if (index < 0)
198 index = 0;
199out:
200 pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n",
201 (unsigned long) ONE_FP/inv_w, maxlen, index);
202
203 return index;
204}
205
206static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
207 struct nlattr **tca, unsigned long *arg)
208{
209 struct qfq_sched *q = qdisc_priv(sch);
210 struct qfq_class *cl = (struct qfq_class *)*arg;
211 struct nlattr *tb[TCA_QFQ_MAX + 1];
212 u32 weight, lmax, inv_w;
213 int i, err;
214
215 if (tca[TCA_OPTIONS] == NULL) {
216 pr_notice("qfq: no options\n");
217 return -EINVAL;
218 }
219
220 err = nla_parse_nested(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS], qfq_policy);
221 if (err < 0)
222 return err;
223
224 if (tb[TCA_QFQ_WEIGHT]) {
225 weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]);
226 if (!weight || weight > (1UL << QFQ_MAX_WSHIFT)) {
227 pr_notice("qfq: invalid weight %u\n", weight);
228 return -EINVAL;
229 }
230 } else
231 weight = 1;
232
233 inv_w = ONE_FP / weight;
234 weight = ONE_FP / inv_w;
235 if (q->wsum + weight > QFQ_MAX_WSUM) {
236 pr_notice("qfq: total weight out of range (%u + %u)\n",
237 weight, q->wsum);
238 return -EINVAL;
239 }
240
241 if (tb[TCA_QFQ_LMAX]) {
242 lmax = nla_get_u32(tb[TCA_QFQ_LMAX]);
243 if (!lmax || lmax > (1UL << QFQ_MTU_SHIFT)) {
244 pr_notice("qfq: invalid max length %u\n", lmax);
245 return -EINVAL;
246 }
247 } else
248 lmax = 1UL << QFQ_MTU_SHIFT;
249
250 if (cl != NULL) {
251 if (tca[TCA_RATE]) {
252 err = gen_replace_estimator(&cl->bstats, &cl->rate_est,
253 qdisc_root_sleeping_lock(sch),
254 tca[TCA_RATE]);
255 if (err)
256 return err;
257 }
258
259 sch_tree_lock(sch);
260 if (tb[TCA_QFQ_WEIGHT]) {
261 q->wsum = weight - ONE_FP / cl->inv_w;
262 cl->inv_w = inv_w;
263 }
264 sch_tree_unlock(sch);
265
266 return 0;
267 }
268
269 cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
270 if (cl == NULL)
271 return -ENOBUFS;
272
273 cl->refcnt = 1;
274 cl->common.classid = classid;
275 cl->lmax = lmax;
276 cl->inv_w = inv_w;
277 i = qfq_calc_index(cl->inv_w, cl->lmax);
278
279 cl->grp = &q->groups[i];
280 q->wsum += weight;
281
282 cl->qdisc = qdisc_create_dflt(sch->dev_queue,
283 &pfifo_qdisc_ops, classid);
284 if (cl->qdisc == NULL)
285 cl->qdisc = &noop_qdisc;
286
287 if (tca[TCA_RATE]) {
288 err = gen_new_estimator(&cl->bstats, &cl->rate_est,
289 qdisc_root_sleeping_lock(sch),
290 tca[TCA_RATE]);
291 if (err) {
292 qdisc_destroy(cl->qdisc);
293 kfree(cl);
294 return err;
295 }
296 }
297
298 sch_tree_lock(sch);
299 qdisc_class_hash_insert(&q->clhash, &cl->common);
300 sch_tree_unlock(sch);
301
302 qdisc_class_hash_grow(sch, &q->clhash);
303
304 *arg = (unsigned long)cl;
305 return 0;
306}
307
308static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
309{
310 struct qfq_sched *q = qdisc_priv(sch);
311
312 if (cl->inv_w) {
313 q->wsum -= ONE_FP / cl->inv_w;
314 cl->inv_w = 0;
315 }
316
317 gen_kill_estimator(&cl->bstats, &cl->rate_est);
318 qdisc_destroy(cl->qdisc);
319 kfree(cl);
320}
321
322static int qfq_delete_class(struct Qdisc *sch, unsigned long arg)
323{
324 struct qfq_sched *q = qdisc_priv(sch);
325 struct qfq_class *cl = (struct qfq_class *)arg;
326
327 if (cl->filter_cnt > 0)
328 return -EBUSY;
329
330 sch_tree_lock(sch);
331
332 qfq_purge_queue(cl);
333 qdisc_class_hash_remove(&q->clhash, &cl->common);
334
335 BUG_ON(--cl->refcnt == 0);
336 /*
337 * This shouldn't happen: we "hold" one cops->get() when called
338 * from tc_ctl_tclass; the destroy method is done from cops->put().
339 */
340
341 sch_tree_unlock(sch);
342 return 0;
343}
344
345static unsigned long qfq_get_class(struct Qdisc *sch, u32 classid)
346{
347 struct qfq_class *cl = qfq_find_class(sch, classid);
348
349 if (cl != NULL)
350 cl->refcnt++;
351
352 return (unsigned long)cl;
353}
354
355static void qfq_put_class(struct Qdisc *sch, unsigned long arg)
356{
357 struct qfq_class *cl = (struct qfq_class *)arg;
358
359 if (--cl->refcnt == 0)
360 qfq_destroy_class(sch, cl);
361}
362
363static struct tcf_proto **qfq_tcf_chain(struct Qdisc *sch, unsigned long cl)
364{
365 struct qfq_sched *q = qdisc_priv(sch);
366
367 if (cl)
368 return NULL;
369
370 return &q->filter_list;
371}
372
373static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
374 u32 classid)
375{
376 struct qfq_class *cl = qfq_find_class(sch, classid);
377
378 if (cl != NULL)
379 cl->filter_cnt++;
380
381 return (unsigned long)cl;
382}
383
384static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
385{
386 struct qfq_class *cl = (struct qfq_class *)arg;
387
388 cl->filter_cnt--;
389}
390
391static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
392 struct Qdisc *new, struct Qdisc **old)
393{
394 struct qfq_class *cl = (struct qfq_class *)arg;
395
396 if (new == NULL) {
397 new = qdisc_create_dflt(sch->dev_queue,
398 &pfifo_qdisc_ops, cl->common.classid);
399 if (new == NULL)
400 new = &noop_qdisc;
401 }
402
403 sch_tree_lock(sch);
404 qfq_purge_queue(cl);
405 *old = cl->qdisc;
406 cl->qdisc = new;
407 sch_tree_unlock(sch);
408 return 0;
409}
410
411static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
412{
413 struct qfq_class *cl = (struct qfq_class *)arg;
414
415 return cl->qdisc;
416}
417
418static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
419 struct sk_buff *skb, struct tcmsg *tcm)
420{
421 struct qfq_class *cl = (struct qfq_class *)arg;
422 struct nlattr *nest;
423
424 tcm->tcm_parent = TC_H_ROOT;
425 tcm->tcm_handle = cl->common.classid;
426 tcm->tcm_info = cl->qdisc->handle;
427
428 nest = nla_nest_start(skb, TCA_OPTIONS);
429 if (nest == NULL)
430 goto nla_put_failure;
431 NLA_PUT_U32(skb, TCA_QFQ_WEIGHT, ONE_FP/cl->inv_w);
432 NLA_PUT_U32(skb, TCA_QFQ_LMAX, cl->lmax);
433 return nla_nest_end(skb, nest);
434
435nla_put_failure:
436 nla_nest_cancel(skb, nest);
437 return -EMSGSIZE;
438}
439
440static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
441 struct gnet_dump *d)
442{
443 struct qfq_class *cl = (struct qfq_class *)arg;
444 struct tc_qfq_stats xstats;
445
446 memset(&xstats, 0, sizeof(xstats));
447 cl->qdisc->qstats.qlen = cl->qdisc->q.qlen;
448
449 xstats.weight = ONE_FP/cl->inv_w;
450 xstats.lmax = cl->lmax;
451
452 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
453 gnet_stats_copy_rate_est(d, &cl->bstats, &cl->rate_est) < 0 ||
454 gnet_stats_copy_queue(d, &cl->qdisc->qstats) < 0)
455 return -1;
456
457 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
458}
459
460static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
461{
462 struct qfq_sched *q = qdisc_priv(sch);
463 struct qfq_class *cl;
464 struct hlist_node *n;
465 unsigned int i;
466
467 if (arg->stop)
468 return;
469
470 for (i = 0; i < q->clhash.hashsize; i++) {
471 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode) {
472 if (arg->count < arg->skip) {
473 arg->count++;
474 continue;
475 }
476 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
477 arg->stop = 1;
478 return;
479 }
480 arg->count++;
481 }
482 }
483}
484
485static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
486 int *qerr)
487{
488 struct qfq_sched *q = qdisc_priv(sch);
489 struct qfq_class *cl;
490 struct tcf_result res;
491 int result;
492
493 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
494 pr_debug("qfq_classify: found %d\n", skb->priority);
495 cl = qfq_find_class(sch, skb->priority);
496 if (cl != NULL)
497 return cl;
498 }
499
500 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
501 result = tc_classify(skb, q->filter_list, &res);
502 if (result >= 0) {
503#ifdef CONFIG_NET_CLS_ACT
504 switch (result) {
505 case TC_ACT_QUEUED:
506 case TC_ACT_STOLEN:
507 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
508 case TC_ACT_SHOT:
509 return NULL;
510 }
511#endif
512 cl = (struct qfq_class *)res.class;
513 if (cl == NULL)
514 cl = qfq_find_class(sch, res.classid);
515 return cl;
516 }
517
518 return NULL;
519}
520
521/* Generic comparison function, handling wraparound. */
522static inline int qfq_gt(u64 a, u64 b)
523{
524 return (s64)(a - b) > 0;
525}
526
527/* Round a precise timestamp to its slotted value. */
528static inline u64 qfq_round_down(u64 ts, unsigned int shift)
529{
530 return ts & ~((1ULL << shift) - 1);
531}
532
533/* return the pointer to the group with lowest index in the bitmap */
534static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
535 unsigned long bitmap)
536{
537 int index = __ffs(bitmap);
538 return &q->groups[index];
539}
540/* Calculate a mask to mimic what would be ffs_from(). */
541static inline unsigned long mask_from(unsigned long bitmap, int from)
542{
543 return bitmap & ~((1UL << from) - 1);
544}
545
546/*
547 * The state computation relies on ER=0, IR=1, EB=2, IB=3
548 * First compute eligibility comparing grp->S, q->V,
549 * then check if someone is blocking us and possibly add EB
550 */
551static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
552{
553 /* if S > V we are not eligible */
554 unsigned int state = qfq_gt(grp->S, q->V);
555 unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
556 struct qfq_group *next;
557
558 if (mask) {
559 next = qfq_ffs(q, mask);
560 if (qfq_gt(grp->F, next->F))
561 state |= EB;
562 }
563
564 return state;
565}
566
567
568/*
569 * In principle
570 * q->bitmaps[dst] |= q->bitmaps[src] & mask;
571 * q->bitmaps[src] &= ~mask;
572 * but we should make sure that src != dst
573 */
574static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
575 int src, int dst)
576{
577 q->bitmaps[dst] |= q->bitmaps[src] & mask;
578 q->bitmaps[src] &= ~mask;
579}
580
581static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
582{
583 unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
584 struct qfq_group *next;
585
586 if (mask) {
587 next = qfq_ffs(q, mask);
588 if (!qfq_gt(next->F, old_F))
589 return;
590 }
591
592 mask = (1UL << index) - 1;
593 qfq_move_groups(q, mask, EB, ER);
594 qfq_move_groups(q, mask, IB, IR);
595}
596
597/*
598 * perhaps
599 *
600 old_V ^= q->V;
601 old_V >>= QFQ_MIN_SLOT_SHIFT;
602 if (old_V) {
603 ...
604 }
605 *
606 */
607static void qfq_make_eligible(struct qfq_sched *q, u64 old_V)
608{
609 unsigned long vslot = q->V >> QFQ_MIN_SLOT_SHIFT;
610 unsigned long old_vslot = old_V >> QFQ_MIN_SLOT_SHIFT;
611
612 if (vslot != old_vslot) {
613 unsigned long mask = (1UL << fls(vslot ^ old_vslot)) - 1;
614 qfq_move_groups(q, mask, IR, ER);
615 qfq_move_groups(q, mask, IB, EB);
616 }
617}
618
619
620/*
621 * XXX we should make sure that slot becomes less than 32.
622 * This is guaranteed by the input values.
623 * roundedS is always cl->S rounded on grp->slot_shift bits.
624 */
625static void qfq_slot_insert(struct qfq_group *grp, struct qfq_class *cl,
626 u64 roundedS)
627{
628 u64 slot = (roundedS - grp->S) >> grp->slot_shift;
629 unsigned int i = (grp->front + slot) % QFQ_MAX_SLOTS;
630
631 hlist_add_head(&cl->next, &grp->slots[i]);
632 __set_bit(slot, &grp->full_slots);
633}
634
635/* Maybe introduce hlist_first_entry?? */
636static struct qfq_class *qfq_slot_head(struct qfq_group *grp)
637{
638 return hlist_entry(grp->slots[grp->front].first,
639 struct qfq_class, next);
640}
641
642/*
643 * remove the entry from the slot
644 */
645static void qfq_front_slot_remove(struct qfq_group *grp)
646{
647 struct qfq_class *cl = qfq_slot_head(grp);
648
649 BUG_ON(!cl);
650 hlist_del(&cl->next);
651 if (hlist_empty(&grp->slots[grp->front]))
652 __clear_bit(0, &grp->full_slots);
653}
654
655/*
656 * Returns the first full queue in a group. As a side effect,
657 * adjust the bucket list so the first non-empty bucket is at
658 * position 0 in full_slots.
659 */
660static struct qfq_class *qfq_slot_scan(struct qfq_group *grp)
661{
662 unsigned int i;
663
664 pr_debug("qfq slot_scan: grp %u full %#lx\n",
665 grp->index, grp->full_slots);
666
667 if (grp->full_slots == 0)
668 return NULL;
669
670 i = __ffs(grp->full_slots); /* zero based */
671 if (i > 0) {
672 grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
673 grp->full_slots >>= i;
674 }
675
676 return qfq_slot_head(grp);
677}
678
679/*
680 * adjust the bucket list. When the start time of a group decreases,
681 * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
682 * move the objects. The mask of occupied slots must be shifted
683 * because we use ffs() to find the first non-empty slot.
684 * This covers decreases in the group's start time, but what about
685 * increases of the start time ?
686 * Here too we should make sure that i is less than 32
687 */
688static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
689{
690 unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
691
692 grp->full_slots <<= i;
693 grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
694}
695
696static void qfq_update_eligible(struct qfq_sched *q, u64 old_V)
697{
698 struct qfq_group *grp;
699 unsigned long ineligible;
700
701 ineligible = q->bitmaps[IR] | q->bitmaps[IB];
702 if (ineligible) {
703 if (!q->bitmaps[ER]) {
704 grp = qfq_ffs(q, ineligible);
705 if (qfq_gt(grp->S, q->V))
706 q->V = grp->S;
707 }
708 qfq_make_eligible(q, old_V);
709 }
710}
711
712/* What is length of next packet in queue (0 if queue is empty) */
713static unsigned int qdisc_peek_len(struct Qdisc *sch)
714{
715 struct sk_buff *skb;
716
717 skb = sch->ops->peek(sch);
718 return skb ? qdisc_pkt_len(skb) : 0;
719}
720
721/*
722 * Updates the class, returns true if also the group needs to be updated.
723 */
724static bool qfq_update_class(struct qfq_group *grp, struct qfq_class *cl)
725{
726 unsigned int len = qdisc_peek_len(cl->qdisc);
727
728 cl->S = cl->F;
729 if (!len)
730 qfq_front_slot_remove(grp); /* queue is empty */
731 else {
732 u64 roundedS;
733
734 cl->F = cl->S + (u64)len * cl->inv_w;
735 roundedS = qfq_round_down(cl->S, grp->slot_shift);
736 if (roundedS == grp->S)
737 return false;
738
739 qfq_front_slot_remove(grp);
740 qfq_slot_insert(grp, cl, roundedS);
741 }
742
743 return true;
744}
745
746static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
747{
748 struct qfq_sched *q = qdisc_priv(sch);
749 struct qfq_group *grp;
750 struct qfq_class *cl;
751 struct sk_buff *skb;
752 unsigned int len;
753 u64 old_V;
754
755 if (!q->bitmaps[ER])
756 return NULL;
757
758 grp = qfq_ffs(q, q->bitmaps[ER]);
759
760 cl = qfq_slot_head(grp);
761 skb = qdisc_dequeue_peeked(cl->qdisc);
762 if (!skb) {
763 WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
764 return NULL;
765 }
766
767 sch->q.qlen--;
768 qdisc_bstats_update(sch, skb);
769
770 old_V = q->V;
771 len = qdisc_pkt_len(skb);
772 q->V += (u64)len * IWSUM;
773 pr_debug("qfq dequeue: len %u F %lld now %lld\n",
774 len, (unsigned long long) cl->F, (unsigned long long) q->V);
775
776 if (qfq_update_class(grp, cl)) {
777 u64 old_F = grp->F;
778
779 cl = qfq_slot_scan(grp);
780 if (!cl)
781 __clear_bit(grp->index, &q->bitmaps[ER]);
782 else {
783 u64 roundedS = qfq_round_down(cl->S, grp->slot_shift);
784 unsigned int s;
785
786 if (grp->S == roundedS)
787 goto skip_unblock;
788 grp->S = roundedS;
789 grp->F = roundedS + (2ULL << grp->slot_shift);
790 __clear_bit(grp->index, &q->bitmaps[ER]);
791 s = qfq_calc_state(q, grp);
792 __set_bit(grp->index, &q->bitmaps[s]);
793 }
794
795 qfq_unblock_groups(q, grp->index, old_F);
796 }
797
798skip_unblock:
799 qfq_update_eligible(q, old_V);
800
801 return skb;
802}
803
804/*
805 * Assign a reasonable start time for a new flow k in group i.
806 * Admissible values for \hat(F) are multiples of \sigma_i
807 * no greater than V+\sigma_i . Larger values mean that
808 * we had a wraparound so we consider the timestamp to be stale.
809 *
810 * If F is not stale and F >= V then we set S = F.
811 * Otherwise we should assign S = V, but this may violate
812 * the ordering in ER. So, if we have groups in ER, set S to
813 * the F_j of the first group j which would be blocking us.
814 * We are guaranteed not to move S backward because
815 * otherwise our group i would still be blocked.
816 */
817static void qfq_update_start(struct qfq_sched *q, struct qfq_class *cl)
818{
819 unsigned long mask;
820 uint32_t limit, roundedF;
821 int slot_shift = cl->grp->slot_shift;
822
823 roundedF = qfq_round_down(cl->F, slot_shift);
824 limit = qfq_round_down(q->V, slot_shift) + (1UL << slot_shift);
825
826 if (!qfq_gt(cl->F, q->V) || qfq_gt(roundedF, limit)) {
827 /* timestamp was stale */
828 mask = mask_from(q->bitmaps[ER], cl->grp->index);
829 if (mask) {
830 struct qfq_group *next = qfq_ffs(q, mask);
831 if (qfq_gt(roundedF, next->F)) {
832 cl->S = next->F;
833 return;
834 }
835 }
836 cl->S = q->V;
837 } else /* timestamp is not stale */
838 cl->S = cl->F;
839}
840
841static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
842{
843 struct qfq_sched *q = qdisc_priv(sch);
844 struct qfq_group *grp;
845 struct qfq_class *cl;
846 int err;
847 u64 roundedS;
848 int s;
849
850 cl = qfq_classify(skb, sch, &err);
851 if (cl == NULL) {
852 if (err & __NET_XMIT_BYPASS)
853 sch->qstats.drops++;
854 kfree_skb(skb);
855 return err;
856 }
857 pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
858
859 err = qdisc_enqueue(skb, cl->qdisc);
860 if (unlikely(err != NET_XMIT_SUCCESS)) {
861 pr_debug("qfq_enqueue: enqueue failed %d\n", err);
862 if (net_xmit_drop_count(err)) {
863 cl->qstats.drops++;
864 sch->qstats.drops++;
865 }
866 return err;
867 }
868
869 bstats_update(&cl->bstats, skb);
870 ++sch->q.qlen;
871
872 /* If the new skb is not the head of queue, then done here. */
873 if (cl->qdisc->q.qlen != 1)
874 return err;
875
876 /* If reach this point, queue q was idle */
877 grp = cl->grp;
878 qfq_update_start(q, cl);
879
880 /* compute new finish time and rounded start. */
881 cl->F = cl->S + (u64)qdisc_pkt_len(skb) * cl->inv_w;
882 roundedS = qfq_round_down(cl->S, grp->slot_shift);
883
884 /*
885 * insert cl in the correct bucket.
886 * If cl->S >= grp->S we don't need to adjust the
887 * bucket list and simply go to the insertion phase.
888 * Otherwise grp->S is decreasing, we must make room
889 * in the bucket list, and also recompute the group state.
890 * Finally, if there were no flows in this group and nobody
891 * was in ER make sure to adjust V.
892 */
893 if (grp->full_slots) {
894 if (!qfq_gt(grp->S, cl->S))
895 goto skip_update;
896
897 /* create a slot for this cl->S */
898 qfq_slot_rotate(grp, roundedS);
899 /* group was surely ineligible, remove */
900 __clear_bit(grp->index, &q->bitmaps[IR]);
901 __clear_bit(grp->index, &q->bitmaps[IB]);
902 } else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V))
903 q->V = roundedS;
904
905 grp->S = roundedS;
906 grp->F = roundedS + (2ULL << grp->slot_shift);
907 s = qfq_calc_state(q, grp);
908 __set_bit(grp->index, &q->bitmaps[s]);
909
910 pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
911 s, q->bitmaps[s],
912 (unsigned long long) cl->S,
913 (unsigned long long) cl->F,
914 (unsigned long long) q->V);
915
916skip_update:
917 qfq_slot_insert(grp, cl, roundedS);
918
919 return err;
920}
921
922
923static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
924 struct qfq_class *cl)
925{
926 unsigned int i, offset;
927 u64 roundedS;
928
929 roundedS = qfq_round_down(cl->S, grp->slot_shift);
930 offset = (roundedS - grp->S) >> grp->slot_shift;
931 i = (grp->front + offset) % QFQ_MAX_SLOTS;
932
933 hlist_del(&cl->next);
934 if (hlist_empty(&grp->slots[i]))
935 __clear_bit(offset, &grp->full_slots);
936}
937
938/*
939 * called to forcibly destroy a queue.
940 * If the queue is not in the front bucket, or if it has
941 * other queues in the front bucket, we can simply remove
942 * the queue with no other side effects.
943 * Otherwise we must propagate the event up.
944 */
945static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
946{
947 struct qfq_group *grp = cl->grp;
948 unsigned long mask;
949 u64 roundedS;
950 int s;
951
952 cl->F = cl->S;
953 qfq_slot_remove(q, grp, cl);
954
955 if (!grp->full_slots) {
956 __clear_bit(grp->index, &q->bitmaps[IR]);
957 __clear_bit(grp->index, &q->bitmaps[EB]);
958 __clear_bit(grp->index, &q->bitmaps[IB]);
959
960 if (test_bit(grp->index, &q->bitmaps[ER]) &&
961 !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
962 mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
963 if (mask)
964 mask = ~((1UL << __fls(mask)) - 1);
965 else
966 mask = ~0UL;
967 qfq_move_groups(q, mask, EB, ER);
968 qfq_move_groups(q, mask, IB, IR);
969 }
970 __clear_bit(grp->index, &q->bitmaps[ER]);
971 } else if (hlist_empty(&grp->slots[grp->front])) {
972 cl = qfq_slot_scan(grp);
973 roundedS = qfq_round_down(cl->S, grp->slot_shift);
974 if (grp->S != roundedS) {
975 __clear_bit(grp->index, &q->bitmaps[ER]);
976 __clear_bit(grp->index, &q->bitmaps[IR]);
977 __clear_bit(grp->index, &q->bitmaps[EB]);
978 __clear_bit(grp->index, &q->bitmaps[IB]);
979 grp->S = roundedS;
980 grp->F = roundedS + (2ULL << grp->slot_shift);
981 s = qfq_calc_state(q, grp);
982 __set_bit(grp->index, &q->bitmaps[s]);
983 }
984 }
985
986 qfq_update_eligible(q, q->V);
987}
988
989static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
990{
991 struct qfq_sched *q = qdisc_priv(sch);
992 struct qfq_class *cl = (struct qfq_class *)arg;
993
994 if (cl->qdisc->q.qlen == 0)
995 qfq_deactivate_class(q, cl);
996}
997
998static unsigned int qfq_drop(struct Qdisc *sch)
999{
1000 struct qfq_sched *q = qdisc_priv(sch);
1001 struct qfq_group *grp;
1002 unsigned int i, j, len;
1003
1004 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1005 grp = &q->groups[i];
1006 for (j = 0; j < QFQ_MAX_SLOTS; j++) {
1007 struct qfq_class *cl;
1008 struct hlist_node *n;
1009
1010 hlist_for_each_entry(cl, n, &grp->slots[j], next) {
1011
1012 if (!cl->qdisc->ops->drop)
1013 continue;
1014
1015 len = cl->qdisc->ops->drop(cl->qdisc);
1016 if (len > 0) {
1017 sch->q.qlen--;
1018 if (!cl->qdisc->q.qlen)
1019 qfq_deactivate_class(q, cl);
1020
1021 return len;
1022 }
1023 }
1024 }
1025 }
1026
1027 return 0;
1028}
1029
1030static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
1031{
1032 struct qfq_sched *q = qdisc_priv(sch);
1033 struct qfq_group *grp;
1034 int i, j, err;
1035
1036 err = qdisc_class_hash_init(&q->clhash);
1037 if (err < 0)
1038 return err;
1039
1040 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1041 grp = &q->groups[i];
1042 grp->index = i;
1043 grp->slot_shift = QFQ_MTU_SHIFT + FRAC_BITS
1044 - (QFQ_MAX_INDEX - i);
1045 for (j = 0; j < QFQ_MAX_SLOTS; j++)
1046 INIT_HLIST_HEAD(&grp->slots[j]);
1047 }
1048
1049 return 0;
1050}
1051
1052static void qfq_reset_qdisc(struct Qdisc *sch)
1053{
1054 struct qfq_sched *q = qdisc_priv(sch);
1055 struct qfq_group *grp;
1056 struct qfq_class *cl;
1057 struct hlist_node *n, *tmp;
1058 unsigned int i, j;
1059
1060 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1061 grp = &q->groups[i];
1062 for (j = 0; j < QFQ_MAX_SLOTS; j++) {
1063 hlist_for_each_entry_safe(cl, n, tmp,
1064 &grp->slots[j], next) {
1065 qfq_deactivate_class(q, cl);
1066 }
1067 }
1068 }
1069
1070 for (i = 0; i < q->clhash.hashsize; i++) {
1071 hlist_for_each_entry(cl, n, &q->clhash.hash[i], common.hnode)
1072 qdisc_reset(cl->qdisc);
1073 }
1074 sch->q.qlen = 0;
1075}
1076
1077static void qfq_destroy_qdisc(struct Qdisc *sch)
1078{
1079 struct qfq_sched *q = qdisc_priv(sch);
1080 struct qfq_class *cl;
1081 struct hlist_node *n, *next;
1082 unsigned int i;
1083
1084 tcf_destroy_chain(&q->filter_list);
1085
1086 for (i = 0; i < q->clhash.hashsize; i++) {
1087 hlist_for_each_entry_safe(cl, n, next, &q->clhash.hash[i],
1088 common.hnode) {
1089 qfq_destroy_class(sch, cl);
1090 }
1091 }
1092 qdisc_class_hash_destroy(&q->clhash);
1093}
1094
1095static const struct Qdisc_class_ops qfq_class_ops = {
1096 .change = qfq_change_class,
1097 .delete = qfq_delete_class,
1098 .get = qfq_get_class,
1099 .put = qfq_put_class,
1100 .tcf_chain = qfq_tcf_chain,
1101 .bind_tcf = qfq_bind_tcf,
1102 .unbind_tcf = qfq_unbind_tcf,
1103 .graft = qfq_graft_class,
1104 .leaf = qfq_class_leaf,
1105 .qlen_notify = qfq_qlen_notify,
1106 .dump = qfq_dump_class,
1107 .dump_stats = qfq_dump_class_stats,
1108 .walk = qfq_walk,
1109};
1110
1111static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1112 .cl_ops = &qfq_class_ops,
1113 .id = "qfq",
1114 .priv_size = sizeof(struct qfq_sched),
1115 .enqueue = qfq_enqueue,
1116 .dequeue = qfq_dequeue,
1117 .peek = qdisc_peek_dequeued,
1118 .drop = qfq_drop,
1119 .init = qfq_init_qdisc,
1120 .reset = qfq_reset_qdisc,
1121 .destroy = qfq_destroy_qdisc,
1122 .owner = THIS_MODULE,
1123};
1124
1125static int __init qfq_init(void)
1126{
1127 return register_qdisc(&qfq_qdisc_ops);
1128}
1129
1130static void __exit qfq_exit(void)
1131{
1132 unregister_qdisc(&qfq_qdisc_ops);
1133}
1134
1135module_init(qfq_init);
1136module_exit(qfq_exit);
1137MODULE_LICENSE("GPL");
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * net/sched/sch_qfq.c Quick Fair Queueing Plus Scheduler.
4 *
5 * Copyright (c) 2009 Fabio Checconi, Luigi Rizzo, and Paolo Valente.
6 * Copyright (c) 2012 Paolo Valente.
7 */
8
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>
18
19
20/* Quick Fair Queueing Plus
21 ========================
22
23 Sources:
24
25 [1] Paolo Valente,
26 "Reducing the Execution Time of Fair-Queueing Schedulers."
27 http://algo.ing.unimo.it/people/paolo/agg-sched/agg-sched.pdf
28
29 Sources for QFQ:
30
31 [2] Fabio Checconi, Luigi Rizzo, and Paolo Valente: "QFQ: Efficient
32 Packet Scheduling with Tight Bandwidth Distribution Guarantees."
33
34 See also:
35 http://retis.sssup.it/~fabio/linux/qfq/
36 */
37
38/*
39
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
45 scheduled with DRR.
46
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.
53
54 Virtual time computations.
55
56 S, F and V are all computed in fixed point arithmetic with
57 FRAC_BITS decimal bits.
58
59 QFQ_MAX_INDEX is the maximum index allowed for a group. We need
60 one bit per index.
61 QFQ_MAX_WSHIFT is the maximum power of two supported as a weight.
62
63 The layout of the bits is as below:
64
65 [ MTU_SHIFT ][ FRAC_BITS ]
66 [ MAX_INDEX ][ MIN_SLOT_SHIFT ]
67 ^.__grp->index = 0
68 *.__grp->slot_shift
69
70 where MIN_SLOT_SHIFT is derived by difference from the others.
71
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.
76
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.
84
85 The per-scheduler-instance data contain all the data structures
86 for the scheduler: bitmaps and bucket lists.
87
88 */
89
90/*
91 * Maximum number of consecutive slots occupied by backlogged classes
92 * inside a group.
93 */
94#define QFQ_MAX_SLOTS 32
95
96/*
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.
101 *
102 * grp->index is the index of the group; and grp->slot_shift
103 * is the shift for the corresponding (scaled) sigma_i.
104 */
105#define QFQ_MAX_INDEX 24
106#define QFQ_MAX_WSHIFT 10
107
108#define QFQ_MAX_WEIGHT (1<<QFQ_MAX_WSHIFT) /* see qfq_slot_insert */
109#define QFQ_MAX_WSUM (64*QFQ_MAX_WEIGHT)
110
111#define FRAC_BITS 30 /* fixed point arithmetic */
112#define ONE_FP (1UL << FRAC_BITS)
113
114#define QFQ_MTU_SHIFT 16 /* to support TSO/GSO */
115#define QFQ_MIN_LMAX 512 /* see qfq_slot_insert */
116#define QFQ_MAX_LMAX (1UL << QFQ_MTU_SHIFT)
117
118#define QFQ_MAX_AGG_CLASSES 8 /* max num classes per aggregate allowed */
119
120/*
121 * Possible group states. These values are used as indexes for the bitmaps
122 * array of struct qfq_queue.
123 */
124enum qfq_state { ER, IR, EB, IB, QFQ_MAX_STATE };
125
126struct qfq_group;
127
128struct qfq_aggregate;
129
130struct qfq_class {
131 struct Qdisc_class_common common;
132
133 struct gnet_stats_basic_sync bstats;
134 struct gnet_stats_queue qstats;
135 struct net_rate_estimator __rcu *rate_est;
136 struct Qdisc *qdisc;
137 struct list_head alist; /* Link for active-classes list. */
138 struct qfq_aggregate *agg; /* Parent aggregate. */
139 int deficit; /* DRR deficit counter. */
140};
141
142struct qfq_aggregate {
143 struct hlist_node next; /* Link for the slot list. */
144 u64 S, F; /* flow timestamps (exact) */
145
146 /* group we belong to. In principle we would need the index,
147 * which is log_2(lmax/weight), but we never reference it
148 * directly, only the group.
149 */
150 struct qfq_group *grp;
151
152 /* these are copied from the flowset. */
153 u32 class_weight; /* Weight of each class in this aggregate. */
154 /* Max pkt size for the classes in this aggregate, DRR quantum. */
155 int lmax;
156
157 u32 inv_w; /* ONE_FP/(sum of weights of classes in aggr.). */
158 u32 budgetmax; /* Max budget for this aggregate. */
159 u32 initial_budget, budget; /* Initial and current budget. */
160
161 int num_classes; /* Number of classes in this aggr. */
162 struct list_head active; /* DRR queue of active classes. */
163
164 struct hlist_node nonfull_next; /* See nonfull_aggs in qfq_sched. */
165};
166
167struct qfq_group {
168 u64 S, F; /* group timestamps (approx). */
169 unsigned int slot_shift; /* Slot shift. */
170 unsigned int index; /* Group index. */
171 unsigned int front; /* Index of the front slot. */
172 unsigned long full_slots; /* non-empty slots */
173
174 /* Array of RR lists of active aggregates. */
175 struct hlist_head slots[QFQ_MAX_SLOTS];
176};
177
178struct qfq_sched {
179 struct tcf_proto __rcu *filter_list;
180 struct tcf_block *block;
181 struct Qdisc_class_hash clhash;
182
183 u64 oldV, V; /* Precise virtual times. */
184 struct qfq_aggregate *in_serv_agg; /* Aggregate being served. */
185 u32 wsum; /* weight sum */
186 u32 iwsum; /* inverse weight sum */
187
188 unsigned long bitmaps[QFQ_MAX_STATE]; /* Group bitmaps. */
189 struct qfq_group groups[QFQ_MAX_INDEX + 1]; /* The groups. */
190 u32 min_slot_shift; /* Index of the group-0 bit in the bitmaps. */
191
192 u32 max_agg_classes; /* Max number of classes per aggr. */
193 struct hlist_head nonfull_aggs; /* Aggs with room for more classes. */
194};
195
196/*
197 * Possible reasons why the timestamps of an aggregate are updated
198 * enqueue: the aggregate switches from idle to active and must scheduled
199 * for service
200 * requeue: the aggregate finishes its budget, so it stops being served and
201 * must be rescheduled for service
202 */
203enum update_reason {enqueue, requeue};
204
205static struct qfq_class *qfq_find_class(struct Qdisc *sch, u32 classid)
206{
207 struct qfq_sched *q = qdisc_priv(sch);
208 struct Qdisc_class_common *clc;
209
210 clc = qdisc_class_find(&q->clhash, classid);
211 if (clc == NULL)
212 return NULL;
213 return container_of(clc, struct qfq_class, common);
214}
215
216static const struct netlink_range_validation lmax_range = {
217 .min = QFQ_MIN_LMAX,
218 .max = QFQ_MAX_LMAX,
219};
220
221static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = {
222 [TCA_QFQ_WEIGHT] = NLA_POLICY_RANGE(NLA_U32, 1, QFQ_MAX_WEIGHT),
223 [TCA_QFQ_LMAX] = NLA_POLICY_FULL_RANGE(NLA_U32, &lmax_range),
224};
225
226/*
227 * Calculate a flow index, given its weight and maximum packet length.
228 * index = log_2(maxlen/weight) but we need to apply the scaling.
229 * This is used only once at flow creation.
230 */
231static int qfq_calc_index(u32 inv_w, unsigned int maxlen, u32 min_slot_shift)
232{
233 u64 slot_size = (u64)maxlen * inv_w;
234 unsigned long size_map;
235 int index = 0;
236
237 size_map = slot_size >> min_slot_shift;
238 if (!size_map)
239 goto out;
240
241 index = __fls(size_map) + 1; /* basically a log_2 */
242 index -= !(slot_size - (1ULL << (index + min_slot_shift - 1)));
243
244 if (index < 0)
245 index = 0;
246out:
247 pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n",
248 (unsigned long) ONE_FP/inv_w, maxlen, index);
249
250 return index;
251}
252
253static void qfq_deactivate_agg(struct qfq_sched *, struct qfq_aggregate *);
254static void qfq_activate_agg(struct qfq_sched *, struct qfq_aggregate *,
255 enum update_reason);
256
257static void qfq_init_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
258 u32 lmax, u32 weight)
259{
260 INIT_LIST_HEAD(&agg->active);
261 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
262
263 agg->lmax = lmax;
264 agg->class_weight = weight;
265}
266
267static struct qfq_aggregate *qfq_find_agg(struct qfq_sched *q,
268 u32 lmax, u32 weight)
269{
270 struct qfq_aggregate *agg;
271
272 hlist_for_each_entry(agg, &q->nonfull_aggs, nonfull_next)
273 if (agg->lmax == lmax && agg->class_weight == weight)
274 return agg;
275
276 return NULL;
277}
278
279
280/* Update aggregate as a function of the new number of classes. */
281static void qfq_update_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
282 int new_num_classes)
283{
284 u32 new_agg_weight;
285
286 if (new_num_classes == q->max_agg_classes)
287 hlist_del_init(&agg->nonfull_next);
288
289 if (agg->num_classes > new_num_classes &&
290 new_num_classes == q->max_agg_classes - 1) /* agg no more full */
291 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
292
293 /* The next assignment may let
294 * agg->initial_budget > agg->budgetmax
295 * hold, we will take it into account in charge_actual_service().
296 */
297 agg->budgetmax = new_num_classes * agg->lmax;
298 new_agg_weight = agg->class_weight * new_num_classes;
299 agg->inv_w = ONE_FP/new_agg_weight;
300
301 if (agg->grp == NULL) {
302 int i = qfq_calc_index(agg->inv_w, agg->budgetmax,
303 q->min_slot_shift);
304 agg->grp = &q->groups[i];
305 }
306
307 q->wsum +=
308 (int) agg->class_weight * (new_num_classes - agg->num_classes);
309 q->iwsum = ONE_FP / q->wsum;
310
311 agg->num_classes = new_num_classes;
312}
313
314/* Add class to aggregate. */
315static void qfq_add_to_agg(struct qfq_sched *q,
316 struct qfq_aggregate *agg,
317 struct qfq_class *cl)
318{
319 cl->agg = agg;
320
321 qfq_update_agg(q, agg, agg->num_classes+1);
322 if (cl->qdisc->q.qlen > 0) { /* adding an active class */
323 list_add_tail(&cl->alist, &agg->active);
324 if (list_first_entry(&agg->active, struct qfq_class, alist) ==
325 cl && q->in_serv_agg != agg) /* agg was inactive */
326 qfq_activate_agg(q, agg, enqueue); /* schedule agg */
327 }
328}
329
330static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *);
331
332static void qfq_destroy_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
333{
334 hlist_del_init(&agg->nonfull_next);
335 q->wsum -= agg->class_weight;
336 if (q->wsum != 0)
337 q->iwsum = ONE_FP / q->wsum;
338
339 if (q->in_serv_agg == agg)
340 q->in_serv_agg = qfq_choose_next_agg(q);
341 kfree(agg);
342}
343
344/* Deschedule class from within its parent aggregate. */
345static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
346{
347 struct qfq_aggregate *agg = cl->agg;
348
349
350 list_del(&cl->alist); /* remove from RR queue of the aggregate */
351 if (list_empty(&agg->active)) /* agg is now inactive */
352 qfq_deactivate_agg(q, agg);
353}
354
355/* Remove class from its parent aggregate. */
356static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
357{
358 struct qfq_aggregate *agg = cl->agg;
359
360 cl->agg = NULL;
361 if (agg->num_classes == 1) { /* agg being emptied, destroy it */
362 qfq_destroy_agg(q, agg);
363 return;
364 }
365 qfq_update_agg(q, agg, agg->num_classes-1);
366}
367
368/* Deschedule class and remove it from its parent aggregate. */
369static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
370{
371 if (cl->qdisc->q.qlen > 0) /* class is active */
372 qfq_deactivate_class(q, cl);
373
374 qfq_rm_from_agg(q, cl);
375}
376
377/* Move class to a new aggregate, matching the new class weight and/or lmax */
378static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight,
379 u32 lmax)
380{
381 struct qfq_sched *q = qdisc_priv(sch);
382 struct qfq_aggregate *new_agg;
383
384 /* 'lmax' can range from [QFQ_MIN_LMAX, pktlen + stab overhead] */
385 if (lmax > QFQ_MAX_LMAX)
386 return -EINVAL;
387
388 new_agg = qfq_find_agg(q, lmax, weight);
389 if (new_agg == NULL) { /* create new aggregate */
390 new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC);
391 if (new_agg == NULL)
392 return -ENOBUFS;
393 qfq_init_agg(q, new_agg, lmax, weight);
394 }
395 qfq_deact_rm_from_agg(q, cl);
396 qfq_add_to_agg(q, new_agg, cl);
397
398 return 0;
399}
400
401static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
402 struct nlattr **tca, unsigned long *arg,
403 struct netlink_ext_ack *extack)
404{
405 struct qfq_sched *q = qdisc_priv(sch);
406 struct qfq_class *cl = (struct qfq_class *)*arg;
407 bool existing = false;
408 struct nlattr *tb[TCA_QFQ_MAX + 1];
409 struct qfq_aggregate *new_agg = NULL;
410 u32 weight, lmax, inv_w;
411 int err;
412 int delta_w;
413
414 if (NL_REQ_ATTR_CHECK(extack, NULL, tca, TCA_OPTIONS)) {
415 NL_SET_ERR_MSG_MOD(extack, "missing options");
416 return -EINVAL;
417 }
418
419 err = nla_parse_nested_deprecated(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS],
420 qfq_policy, extack);
421 if (err < 0)
422 return err;
423
424 weight = nla_get_u32_default(tb[TCA_QFQ_WEIGHT], 1);
425
426 if (tb[TCA_QFQ_LMAX]) {
427 lmax = nla_get_u32(tb[TCA_QFQ_LMAX]);
428 } else {
429 /* MTU size is user controlled */
430 lmax = psched_mtu(qdisc_dev(sch));
431 if (lmax < QFQ_MIN_LMAX || lmax > QFQ_MAX_LMAX) {
432 NL_SET_ERR_MSG_MOD(extack,
433 "MTU size out of bounds for qfq");
434 return -EINVAL;
435 }
436 }
437
438 inv_w = ONE_FP / weight;
439 weight = ONE_FP / inv_w;
440
441 if (cl != NULL &&
442 lmax == cl->agg->lmax &&
443 weight == cl->agg->class_weight)
444 return 0; /* nothing to change */
445
446 delta_w = weight - (cl ? cl->agg->class_weight : 0);
447
448 if (q->wsum + delta_w > QFQ_MAX_WSUM) {
449 NL_SET_ERR_MSG_FMT_MOD(extack,
450 "total weight out of range (%d + %u)\n",
451 delta_w, q->wsum);
452 return -EINVAL;
453 }
454
455 if (cl != NULL) { /* modify existing class */
456 if (tca[TCA_RATE]) {
457 err = gen_replace_estimator(&cl->bstats, NULL,
458 &cl->rate_est,
459 NULL,
460 true,
461 tca[TCA_RATE]);
462 if (err)
463 return err;
464 }
465 existing = true;
466 goto set_change_agg;
467 }
468
469 /* create and init new class */
470 cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
471 if (cl == NULL)
472 return -ENOBUFS;
473
474 gnet_stats_basic_sync_init(&cl->bstats);
475 cl->common.classid = classid;
476 cl->deficit = lmax;
477
478 cl->qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
479 classid, NULL);
480 if (cl->qdisc == NULL)
481 cl->qdisc = &noop_qdisc;
482
483 if (tca[TCA_RATE]) {
484 err = gen_new_estimator(&cl->bstats, NULL,
485 &cl->rate_est,
486 NULL,
487 true,
488 tca[TCA_RATE]);
489 if (err)
490 goto destroy_class;
491 }
492
493 if (cl->qdisc != &noop_qdisc)
494 qdisc_hash_add(cl->qdisc, true);
495
496set_change_agg:
497 sch_tree_lock(sch);
498 new_agg = qfq_find_agg(q, lmax, weight);
499 if (new_agg == NULL) { /* create new aggregate */
500 sch_tree_unlock(sch);
501 new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL);
502 if (new_agg == NULL) {
503 err = -ENOBUFS;
504 gen_kill_estimator(&cl->rate_est);
505 goto destroy_class;
506 }
507 sch_tree_lock(sch);
508 qfq_init_agg(q, new_agg, lmax, weight);
509 }
510 if (existing)
511 qfq_deact_rm_from_agg(q, cl);
512 else
513 qdisc_class_hash_insert(&q->clhash, &cl->common);
514 qfq_add_to_agg(q, new_agg, cl);
515 sch_tree_unlock(sch);
516 qdisc_class_hash_grow(sch, &q->clhash);
517
518 *arg = (unsigned long)cl;
519 return 0;
520
521destroy_class:
522 qdisc_put(cl->qdisc);
523 kfree(cl);
524 return err;
525}
526
527static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
528{
529 struct qfq_sched *q = qdisc_priv(sch);
530
531 qfq_rm_from_agg(q, cl);
532 gen_kill_estimator(&cl->rate_est);
533 qdisc_put(cl->qdisc);
534 kfree(cl);
535}
536
537static int qfq_delete_class(struct Qdisc *sch, unsigned long arg,
538 struct netlink_ext_ack *extack)
539{
540 struct qfq_sched *q = qdisc_priv(sch);
541 struct qfq_class *cl = (struct qfq_class *)arg;
542
543 if (qdisc_class_in_use(&cl->common)) {
544 NL_SET_ERR_MSG_MOD(extack, "QFQ class in use");
545 return -EBUSY;
546 }
547
548 sch_tree_lock(sch);
549
550 qdisc_purge_queue(cl->qdisc);
551 qdisc_class_hash_remove(&q->clhash, &cl->common);
552
553 sch_tree_unlock(sch);
554
555 qfq_destroy_class(sch, cl);
556 return 0;
557}
558
559static unsigned long qfq_search_class(struct Qdisc *sch, u32 classid)
560{
561 return (unsigned long)qfq_find_class(sch, classid);
562}
563
564static struct tcf_block *qfq_tcf_block(struct Qdisc *sch, unsigned long cl,
565 struct netlink_ext_ack *extack)
566{
567 struct qfq_sched *q = qdisc_priv(sch);
568
569 if (cl)
570 return NULL;
571
572 return q->block;
573}
574
575static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
576 u32 classid)
577{
578 struct qfq_class *cl = qfq_find_class(sch, classid);
579
580 if (cl)
581 qdisc_class_get(&cl->common);
582
583 return (unsigned long)cl;
584}
585
586static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
587{
588 struct qfq_class *cl = (struct qfq_class *)arg;
589
590 qdisc_class_put(&cl->common);
591}
592
593static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
594 struct Qdisc *new, struct Qdisc **old,
595 struct netlink_ext_ack *extack)
596{
597 struct qfq_class *cl = (struct qfq_class *)arg;
598
599 if (new == NULL) {
600 new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
601 cl->common.classid, NULL);
602 if (new == NULL)
603 new = &noop_qdisc;
604 }
605
606 *old = qdisc_replace(sch, new, &cl->qdisc);
607 return 0;
608}
609
610static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
611{
612 struct qfq_class *cl = (struct qfq_class *)arg;
613
614 return cl->qdisc;
615}
616
617static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
618 struct sk_buff *skb, struct tcmsg *tcm)
619{
620 struct qfq_class *cl = (struct qfq_class *)arg;
621 struct nlattr *nest;
622
623 tcm->tcm_parent = TC_H_ROOT;
624 tcm->tcm_handle = cl->common.classid;
625 tcm->tcm_info = cl->qdisc->handle;
626
627 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
628 if (nest == NULL)
629 goto nla_put_failure;
630 if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) ||
631 nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax))
632 goto nla_put_failure;
633 return nla_nest_end(skb, nest);
634
635nla_put_failure:
636 nla_nest_cancel(skb, nest);
637 return -EMSGSIZE;
638}
639
640static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
641 struct gnet_dump *d)
642{
643 struct qfq_class *cl = (struct qfq_class *)arg;
644 struct tc_qfq_stats xstats;
645
646 memset(&xstats, 0, sizeof(xstats));
647
648 xstats.weight = cl->agg->class_weight;
649 xstats.lmax = cl->agg->lmax;
650
651 if (gnet_stats_copy_basic(d, NULL, &cl->bstats, true) < 0 ||
652 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
653 qdisc_qstats_copy(d, cl->qdisc) < 0)
654 return -1;
655
656 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
657}
658
659static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
660{
661 struct qfq_sched *q = qdisc_priv(sch);
662 struct qfq_class *cl;
663 unsigned int i;
664
665 if (arg->stop)
666 return;
667
668 for (i = 0; i < q->clhash.hashsize; i++) {
669 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
670 if (!tc_qdisc_stats_dump(sch, (unsigned long)cl, arg))
671 return;
672 }
673 }
674}
675
676static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
677 int *qerr)
678{
679 struct qfq_sched *q = qdisc_priv(sch);
680 struct qfq_class *cl;
681 struct tcf_result res;
682 struct tcf_proto *fl;
683 int result;
684
685 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
686 pr_debug("qfq_classify: found %d\n", skb->priority);
687 cl = qfq_find_class(sch, skb->priority);
688 if (cl != NULL)
689 return cl;
690 }
691
692 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
693 fl = rcu_dereference_bh(q->filter_list);
694 result = tcf_classify(skb, NULL, fl, &res, false);
695 if (result >= 0) {
696#ifdef CONFIG_NET_CLS_ACT
697 switch (result) {
698 case TC_ACT_QUEUED:
699 case TC_ACT_STOLEN:
700 case TC_ACT_TRAP:
701 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
702 fallthrough;
703 case TC_ACT_SHOT:
704 return NULL;
705 }
706#endif
707 cl = (struct qfq_class *)res.class;
708 if (cl == NULL)
709 cl = qfq_find_class(sch, res.classid);
710 return cl;
711 }
712
713 return NULL;
714}
715
716/* Generic comparison function, handling wraparound. */
717static inline int qfq_gt(u64 a, u64 b)
718{
719 return (s64)(a - b) > 0;
720}
721
722/* Round a precise timestamp to its slotted value. */
723static inline u64 qfq_round_down(u64 ts, unsigned int shift)
724{
725 return ts & ~((1ULL << shift) - 1);
726}
727
728/* return the pointer to the group with lowest index in the bitmap */
729static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
730 unsigned long bitmap)
731{
732 int index = __ffs(bitmap);
733 return &q->groups[index];
734}
735/* Calculate a mask to mimic what would be ffs_from(). */
736static inline unsigned long mask_from(unsigned long bitmap, int from)
737{
738 return bitmap & ~((1UL << from) - 1);
739}
740
741/*
742 * The state computation relies on ER=0, IR=1, EB=2, IB=3
743 * First compute eligibility comparing grp->S, q->V,
744 * then check if someone is blocking us and possibly add EB
745 */
746static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
747{
748 /* if S > V we are not eligible */
749 unsigned int state = qfq_gt(grp->S, q->V);
750 unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
751 struct qfq_group *next;
752
753 if (mask) {
754 next = qfq_ffs(q, mask);
755 if (qfq_gt(grp->F, next->F))
756 state |= EB;
757 }
758
759 return state;
760}
761
762
763/*
764 * In principle
765 * q->bitmaps[dst] |= q->bitmaps[src] & mask;
766 * q->bitmaps[src] &= ~mask;
767 * but we should make sure that src != dst
768 */
769static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
770 int src, int dst)
771{
772 q->bitmaps[dst] |= q->bitmaps[src] & mask;
773 q->bitmaps[src] &= ~mask;
774}
775
776static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
777{
778 unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
779 struct qfq_group *next;
780
781 if (mask) {
782 next = qfq_ffs(q, mask);
783 if (!qfq_gt(next->F, old_F))
784 return;
785 }
786
787 mask = (1UL << index) - 1;
788 qfq_move_groups(q, mask, EB, ER);
789 qfq_move_groups(q, mask, IB, IR);
790}
791
792/*
793 * perhaps
794 *
795 old_V ^= q->V;
796 old_V >>= q->min_slot_shift;
797 if (old_V) {
798 ...
799 }
800 *
801 */
802static void qfq_make_eligible(struct qfq_sched *q)
803{
804 unsigned long vslot = q->V >> q->min_slot_shift;
805 unsigned long old_vslot = q->oldV >> q->min_slot_shift;
806
807 if (vslot != old_vslot) {
808 unsigned long mask;
809 int last_flip_pos = fls(vslot ^ old_vslot);
810
811 if (last_flip_pos > 31) /* higher than the number of groups */
812 mask = ~0UL; /* make all groups eligible */
813 else
814 mask = (1UL << last_flip_pos) - 1;
815
816 qfq_move_groups(q, mask, IR, ER);
817 qfq_move_groups(q, mask, IB, EB);
818 }
819}
820
821/*
822 * The index of the slot in which the input aggregate agg is to be
823 * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2'
824 * and not a '-1' because the start time of the group may be moved
825 * backward by one slot after the aggregate has been inserted, and
826 * this would cause non-empty slots to be right-shifted by one
827 * position.
828 *
829 * QFQ+ fully satisfies this bound to the slot index if the parameters
830 * of the classes are not changed dynamically, and if QFQ+ never
831 * happens to postpone the service of agg unjustly, i.e., it never
832 * happens that the aggregate becomes backlogged and eligible, or just
833 * eligible, while an aggregate with a higher approximated finish time
834 * is being served. In particular, in this case QFQ+ guarantees that
835 * the timestamps of agg are low enough that the slot index is never
836 * higher than 2. Unfortunately, QFQ+ cannot provide the same
837 * guarantee if it happens to unjustly postpone the service of agg, or
838 * if the parameters of some class are changed.
839 *
840 * As for the first event, i.e., an out-of-order service, the
841 * upper bound to the slot index guaranteed by QFQ+ grows to
842 * 2 +
843 * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
844 * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1.
845 *
846 * The following function deals with this problem by backward-shifting
847 * the timestamps of agg, if needed, so as to guarantee that the slot
848 * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may
849 * cause the service of other aggregates to be postponed, yet the
850 * worst-case guarantees of these aggregates are not violated. In
851 * fact, in case of no out-of-order service, the timestamps of agg
852 * would have been even lower than they are after the backward shift,
853 * because QFQ+ would have guaranteed a maximum value equal to 2 for
854 * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose
855 * service is postponed because of the backward-shift would have
856 * however waited for the service of agg before being served.
857 *
858 * The other event that may cause the slot index to be higher than 2
859 * for agg is a recent change of the parameters of some class. If the
860 * weight of a class is increased or the lmax (max_pkt_size) of the
861 * class is decreased, then a new aggregate with smaller slot size
862 * than the original parent aggregate of the class may happen to be
863 * activated. The activation of this aggregate should be properly
864 * delayed to when the service of the class has finished in the ideal
865 * system tracked by QFQ+. If the activation of the aggregate is not
866 * delayed to this reference time instant, then this aggregate may be
867 * unjustly served before other aggregates waiting for service. This
868 * may cause the above bound to the slot index to be violated for some
869 * of these unlucky aggregates.
870 *
871 * Instead of delaying the activation of the new aggregate, which is
872 * quite complex, the above-discussed capping of the slot index is
873 * used to handle also the consequences of a change of the parameters
874 * of a class.
875 */
876static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg,
877 u64 roundedS)
878{
879 u64 slot = (roundedS - grp->S) >> grp->slot_shift;
880 unsigned int i; /* slot index in the bucket list */
881
882 if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
883 u64 deltaS = roundedS - grp->S -
884 ((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
885 agg->S -= deltaS;
886 agg->F -= deltaS;
887 slot = QFQ_MAX_SLOTS - 2;
888 }
889
890 i = (grp->front + slot) % QFQ_MAX_SLOTS;
891
892 hlist_add_head(&agg->next, &grp->slots[i]);
893 __set_bit(slot, &grp->full_slots);
894}
895
896/* Maybe introduce hlist_first_entry?? */
897static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp)
898{
899 return hlist_entry(grp->slots[grp->front].first,
900 struct qfq_aggregate, next);
901}
902
903/*
904 * remove the entry from the slot
905 */
906static void qfq_front_slot_remove(struct qfq_group *grp)
907{
908 struct qfq_aggregate *agg = qfq_slot_head(grp);
909
910 BUG_ON(!agg);
911 hlist_del(&agg->next);
912 if (hlist_empty(&grp->slots[grp->front]))
913 __clear_bit(0, &grp->full_slots);
914}
915
916/*
917 * Returns the first aggregate in the first non-empty bucket of the
918 * group. As a side effect, adjusts the bucket list so the first
919 * non-empty bucket is at position 0 in full_slots.
920 */
921static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp)
922{
923 unsigned int i;
924
925 pr_debug("qfq slot_scan: grp %u full %#lx\n",
926 grp->index, grp->full_slots);
927
928 if (grp->full_slots == 0)
929 return NULL;
930
931 i = __ffs(grp->full_slots); /* zero based */
932 if (i > 0) {
933 grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
934 grp->full_slots >>= i;
935 }
936
937 return qfq_slot_head(grp);
938}
939
940/*
941 * adjust the bucket list. When the start time of a group decreases,
942 * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
943 * move the objects. The mask of occupied slots must be shifted
944 * because we use ffs() to find the first non-empty slot.
945 * This covers decreases in the group's start time, but what about
946 * increases of the start time ?
947 * Here too we should make sure that i is less than 32
948 */
949static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
950{
951 unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
952
953 grp->full_slots <<= i;
954 grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
955}
956
957static void qfq_update_eligible(struct qfq_sched *q)
958{
959 struct qfq_group *grp;
960 unsigned long ineligible;
961
962 ineligible = q->bitmaps[IR] | q->bitmaps[IB];
963 if (ineligible) {
964 if (!q->bitmaps[ER]) {
965 grp = qfq_ffs(q, ineligible);
966 if (qfq_gt(grp->S, q->V))
967 q->V = grp->S;
968 }
969 qfq_make_eligible(q);
970 }
971}
972
973/* Dequeue head packet of the head class in the DRR queue of the aggregate. */
974static struct sk_buff *agg_dequeue(struct qfq_aggregate *agg,
975 struct qfq_class *cl, unsigned int len)
976{
977 struct sk_buff *skb = qdisc_dequeue_peeked(cl->qdisc);
978
979 if (!skb)
980 return NULL;
981
982 cl->deficit -= (int) len;
983
984 if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */
985 list_del(&cl->alist);
986 else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) {
987 cl->deficit += agg->lmax;
988 list_move_tail(&cl->alist, &agg->active);
989 }
990
991 return skb;
992}
993
994static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg,
995 struct qfq_class **cl,
996 unsigned int *len)
997{
998 struct sk_buff *skb;
999
1000 *cl = list_first_entry(&agg->active, struct qfq_class, alist);
1001 skb = (*cl)->qdisc->ops->peek((*cl)->qdisc);
1002 if (skb == NULL)
1003 qdisc_warn_nonwc("qfq_dequeue", (*cl)->qdisc);
1004 else
1005 *len = qdisc_pkt_len(skb);
1006
1007 return skb;
1008}
1009
1010/* Update F according to the actual service received by the aggregate. */
1011static inline void charge_actual_service(struct qfq_aggregate *agg)
1012{
1013 /* Compute the service received by the aggregate, taking into
1014 * account that, after decreasing the number of classes in
1015 * agg, it may happen that
1016 * agg->initial_budget - agg->budget > agg->bugdetmax
1017 */
1018 u32 service_received = min(agg->budgetmax,
1019 agg->initial_budget - agg->budget);
1020
1021 agg->F = agg->S + (u64)service_received * agg->inv_w;
1022}
1023
1024/* Assign a reasonable start time for a new aggregate in group i.
1025 * Admissible values for \hat(F) are multiples of \sigma_i
1026 * no greater than V+\sigma_i . Larger values mean that
1027 * we had a wraparound so we consider the timestamp to be stale.
1028 *
1029 * If F is not stale and F >= V then we set S = F.
1030 * Otherwise we should assign S = V, but this may violate
1031 * the ordering in EB (see [2]). So, if we have groups in ER,
1032 * set S to the F_j of the first group j which would be blocking us.
1033 * We are guaranteed not to move S backward because
1034 * otherwise our group i would still be blocked.
1035 */
1036static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg)
1037{
1038 unsigned long mask;
1039 u64 limit, roundedF;
1040 int slot_shift = agg->grp->slot_shift;
1041
1042 roundedF = qfq_round_down(agg->F, slot_shift);
1043 limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
1044
1045 if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) {
1046 /* timestamp was stale */
1047 mask = mask_from(q->bitmaps[ER], agg->grp->index);
1048 if (mask) {
1049 struct qfq_group *next = qfq_ffs(q, mask);
1050 if (qfq_gt(roundedF, next->F)) {
1051 if (qfq_gt(limit, next->F))
1052 agg->S = next->F;
1053 else /* preserve timestamp correctness */
1054 agg->S = limit;
1055 return;
1056 }
1057 }
1058 agg->S = q->V;
1059 } else /* timestamp is not stale */
1060 agg->S = agg->F;
1061}
1062
1063/* Update the timestamps of agg before scheduling/rescheduling it for
1064 * service. In particular, assign to agg->F its maximum possible
1065 * value, i.e., the virtual finish time with which the aggregate
1066 * should be labeled if it used all its budget once in service.
1067 */
1068static inline void
1069qfq_update_agg_ts(struct qfq_sched *q,
1070 struct qfq_aggregate *agg, enum update_reason reason)
1071{
1072 if (reason != requeue)
1073 qfq_update_start(q, agg);
1074 else /* just charge agg for the service received */
1075 agg->S = agg->F;
1076
1077 agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w;
1078}
1079
1080static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg);
1081
1082static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
1083{
1084 struct qfq_sched *q = qdisc_priv(sch);
1085 struct qfq_aggregate *in_serv_agg = q->in_serv_agg;
1086 struct qfq_class *cl;
1087 struct sk_buff *skb = NULL;
1088 /* next-packet len, 0 means no more active classes in in-service agg */
1089 unsigned int len = 0;
1090
1091 if (in_serv_agg == NULL)
1092 return NULL;
1093
1094 if (!list_empty(&in_serv_agg->active))
1095 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1096
1097 /*
1098 * If there are no active classes in the in-service aggregate,
1099 * or if the aggregate has not enough budget to serve its next
1100 * class, then choose the next aggregate to serve.
1101 */
1102 if (len == 0 || in_serv_agg->budget < len) {
1103 charge_actual_service(in_serv_agg);
1104
1105 /* recharge the budget of the aggregate */
1106 in_serv_agg->initial_budget = in_serv_agg->budget =
1107 in_serv_agg->budgetmax;
1108
1109 if (!list_empty(&in_serv_agg->active)) {
1110 /*
1111 * Still active: reschedule for
1112 * service. Possible optimization: if no other
1113 * aggregate is active, then there is no point
1114 * in rescheduling this aggregate, and we can
1115 * just keep it as the in-service one. This
1116 * should be however a corner case, and to
1117 * handle it, we would need to maintain an
1118 * extra num_active_aggs field.
1119 */
1120 qfq_update_agg_ts(q, in_serv_agg, requeue);
1121 qfq_schedule_agg(q, in_serv_agg);
1122 } else if (sch->q.qlen == 0) { /* no aggregate to serve */
1123 q->in_serv_agg = NULL;
1124 return NULL;
1125 }
1126
1127 /*
1128 * If we get here, there are other aggregates queued:
1129 * choose the new aggregate to serve.
1130 */
1131 in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
1132 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1133 }
1134 if (!skb)
1135 return NULL;
1136
1137 sch->q.qlen--;
1138
1139 skb = agg_dequeue(in_serv_agg, cl, len);
1140
1141 if (!skb) {
1142 sch->q.qlen++;
1143 return NULL;
1144 }
1145
1146 qdisc_qstats_backlog_dec(sch, skb);
1147 qdisc_bstats_update(sch, skb);
1148
1149 /* If lmax is lowered, through qfq_change_class, for a class
1150 * owning pending packets with larger size than the new value
1151 * of lmax, then the following condition may hold.
1152 */
1153 if (unlikely(in_serv_agg->budget < len))
1154 in_serv_agg->budget = 0;
1155 else
1156 in_serv_agg->budget -= len;
1157
1158 q->V += (u64)len * q->iwsum;
1159 pr_debug("qfq dequeue: len %u F %lld now %lld\n",
1160 len, (unsigned long long) in_serv_agg->F,
1161 (unsigned long long) q->V);
1162
1163 return skb;
1164}
1165
1166static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)
1167{
1168 struct qfq_group *grp;
1169 struct qfq_aggregate *agg, *new_front_agg;
1170 u64 old_F;
1171
1172 qfq_update_eligible(q);
1173 q->oldV = q->V;
1174
1175 if (!q->bitmaps[ER])
1176 return NULL;
1177
1178 grp = qfq_ffs(q, q->bitmaps[ER]);
1179 old_F = grp->F;
1180
1181 agg = qfq_slot_head(grp);
1182
1183 /* agg starts to be served, remove it from schedule */
1184 qfq_front_slot_remove(grp);
1185
1186 new_front_agg = qfq_slot_scan(grp);
1187
1188 if (new_front_agg == NULL) /* group is now inactive, remove from ER */
1189 __clear_bit(grp->index, &q->bitmaps[ER]);
1190 else {
1191 u64 roundedS = qfq_round_down(new_front_agg->S,
1192 grp->slot_shift);
1193 unsigned int s;
1194
1195 if (grp->S == roundedS)
1196 return agg;
1197 grp->S = roundedS;
1198 grp->F = roundedS + (2ULL << grp->slot_shift);
1199 __clear_bit(grp->index, &q->bitmaps[ER]);
1200 s = qfq_calc_state(q, grp);
1201 __set_bit(grp->index, &q->bitmaps[s]);
1202 }
1203
1204 qfq_unblock_groups(q, grp->index, old_F);
1205
1206 return agg;
1207}
1208
1209static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
1210 struct sk_buff **to_free)
1211{
1212 unsigned int len = qdisc_pkt_len(skb), gso_segs;
1213 struct qfq_sched *q = qdisc_priv(sch);
1214 struct qfq_class *cl;
1215 struct qfq_aggregate *agg;
1216 int err = 0;
1217 bool first;
1218
1219 cl = qfq_classify(skb, sch, &err);
1220 if (cl == NULL) {
1221 if (err & __NET_XMIT_BYPASS)
1222 qdisc_qstats_drop(sch);
1223 __qdisc_drop(skb, to_free);
1224 return err;
1225 }
1226 pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
1227
1228 if (unlikely(cl->agg->lmax < len)) {
1229 pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
1230 cl->agg->lmax, len, cl->common.classid);
1231 err = qfq_change_agg(sch, cl, cl->agg->class_weight, len);
1232 if (err) {
1233 cl->qstats.drops++;
1234 return qdisc_drop(skb, sch, to_free);
1235 }
1236 }
1237
1238 gso_segs = skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 1;
1239 first = !cl->qdisc->q.qlen;
1240 err = qdisc_enqueue(skb, cl->qdisc, to_free);
1241 if (unlikely(err != NET_XMIT_SUCCESS)) {
1242 pr_debug("qfq_enqueue: enqueue failed %d\n", err);
1243 if (net_xmit_drop_count(err)) {
1244 cl->qstats.drops++;
1245 qdisc_qstats_drop(sch);
1246 }
1247 return err;
1248 }
1249
1250 _bstats_update(&cl->bstats, len, gso_segs);
1251 sch->qstats.backlog += len;
1252 ++sch->q.qlen;
1253
1254 agg = cl->agg;
1255 /* if the queue was not empty, then done here */
1256 if (!first) {
1257 if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) &&
1258 list_first_entry(&agg->active, struct qfq_class, alist)
1259 == cl && cl->deficit < len)
1260 list_move_tail(&cl->alist, &agg->active);
1261
1262 return err;
1263 }
1264
1265 /* schedule class for service within the aggregate */
1266 cl->deficit = agg->lmax;
1267 list_add_tail(&cl->alist, &agg->active);
1268
1269 if (list_first_entry(&agg->active, struct qfq_class, alist) != cl ||
1270 q->in_serv_agg == agg)
1271 return err; /* non-empty or in service, nothing else to do */
1272
1273 qfq_activate_agg(q, agg, enqueue);
1274
1275 return err;
1276}
1277
1278/*
1279 * Schedule aggregate according to its timestamps.
1280 */
1281static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1282{
1283 struct qfq_group *grp = agg->grp;
1284 u64 roundedS;
1285 int s;
1286
1287 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1288
1289 /*
1290 * Insert agg in the correct bucket.
1291 * If agg->S >= grp->S we don't need to adjust the
1292 * bucket list and simply go to the insertion phase.
1293 * Otherwise grp->S is decreasing, we must make room
1294 * in the bucket list, and also recompute the group state.
1295 * Finally, if there were no flows in this group and nobody
1296 * was in ER make sure to adjust V.
1297 */
1298 if (grp->full_slots) {
1299 if (!qfq_gt(grp->S, agg->S))
1300 goto skip_update;
1301
1302 /* create a slot for this agg->S */
1303 qfq_slot_rotate(grp, roundedS);
1304 /* group was surely ineligible, remove */
1305 __clear_bit(grp->index, &q->bitmaps[IR]);
1306 __clear_bit(grp->index, &q->bitmaps[IB]);
1307 } else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) &&
1308 q->in_serv_agg == NULL)
1309 q->V = roundedS;
1310
1311 grp->S = roundedS;
1312 grp->F = roundedS + (2ULL << grp->slot_shift);
1313 s = qfq_calc_state(q, grp);
1314 __set_bit(grp->index, &q->bitmaps[s]);
1315
1316 pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
1317 s, q->bitmaps[s],
1318 (unsigned long long) agg->S,
1319 (unsigned long long) agg->F,
1320 (unsigned long long) q->V);
1321
1322skip_update:
1323 qfq_slot_insert(grp, agg, roundedS);
1324}
1325
1326
1327/* Update agg ts and schedule agg for service */
1328static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
1329 enum update_reason reason)
1330{
1331 agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */
1332
1333 qfq_update_agg_ts(q, agg, reason);
1334 if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */
1335 q->in_serv_agg = agg; /* start serving this aggregate */
1336 /* update V: to be in service, agg must be eligible */
1337 q->oldV = q->V = agg->S;
1338 } else if (agg != q->in_serv_agg)
1339 qfq_schedule_agg(q, agg);
1340}
1341
1342static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
1343 struct qfq_aggregate *agg)
1344{
1345 unsigned int i, offset;
1346 u64 roundedS;
1347
1348 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1349 offset = (roundedS - grp->S) >> grp->slot_shift;
1350
1351 i = (grp->front + offset) % QFQ_MAX_SLOTS;
1352
1353 hlist_del(&agg->next);
1354 if (hlist_empty(&grp->slots[i]))
1355 __clear_bit(offset, &grp->full_slots);
1356}
1357
1358/*
1359 * Called to forcibly deschedule an aggregate. If the aggregate is
1360 * not in the front bucket, or if the latter has other aggregates in
1361 * the front bucket, we can simply remove the aggregate with no other
1362 * side effects.
1363 * Otherwise we must propagate the event up.
1364 */
1365static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1366{
1367 struct qfq_group *grp = agg->grp;
1368 unsigned long mask;
1369 u64 roundedS;
1370 int s;
1371
1372 if (agg == q->in_serv_agg) {
1373 charge_actual_service(agg);
1374 q->in_serv_agg = qfq_choose_next_agg(q);
1375 return;
1376 }
1377
1378 agg->F = agg->S;
1379 qfq_slot_remove(q, grp, agg);
1380
1381 if (!grp->full_slots) {
1382 __clear_bit(grp->index, &q->bitmaps[IR]);
1383 __clear_bit(grp->index, &q->bitmaps[EB]);
1384 __clear_bit(grp->index, &q->bitmaps[IB]);
1385
1386 if (test_bit(grp->index, &q->bitmaps[ER]) &&
1387 !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
1388 mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
1389 if (mask)
1390 mask = ~((1UL << __fls(mask)) - 1);
1391 else
1392 mask = ~0UL;
1393 qfq_move_groups(q, mask, EB, ER);
1394 qfq_move_groups(q, mask, IB, IR);
1395 }
1396 __clear_bit(grp->index, &q->bitmaps[ER]);
1397 } else if (hlist_empty(&grp->slots[grp->front])) {
1398 agg = qfq_slot_scan(grp);
1399 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1400 if (grp->S != roundedS) {
1401 __clear_bit(grp->index, &q->bitmaps[ER]);
1402 __clear_bit(grp->index, &q->bitmaps[IR]);
1403 __clear_bit(grp->index, &q->bitmaps[EB]);
1404 __clear_bit(grp->index, &q->bitmaps[IB]);
1405 grp->S = roundedS;
1406 grp->F = roundedS + (2ULL << grp->slot_shift);
1407 s = qfq_calc_state(q, grp);
1408 __set_bit(grp->index, &q->bitmaps[s]);
1409 }
1410 }
1411}
1412
1413static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1414{
1415 struct qfq_sched *q = qdisc_priv(sch);
1416 struct qfq_class *cl = (struct qfq_class *)arg;
1417
1418 qfq_deactivate_class(q, cl);
1419}
1420
1421static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt,
1422 struct netlink_ext_ack *extack)
1423{
1424 struct qfq_sched *q = qdisc_priv(sch);
1425 struct qfq_group *grp;
1426 int i, j, err;
1427 u32 max_cl_shift, maxbudg_shift, max_classes;
1428
1429 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
1430 if (err)
1431 return err;
1432
1433 err = qdisc_class_hash_init(&q->clhash);
1434 if (err < 0)
1435 return err;
1436
1437 max_classes = min_t(u64, (u64)qdisc_dev(sch)->tx_queue_len + 1,
1438 QFQ_MAX_AGG_CLASSES);
1439 /* max_cl_shift = floor(log_2(max_classes)) */
1440 max_cl_shift = __fls(max_classes);
1441 q->max_agg_classes = 1<<max_cl_shift;
1442
1443 /* maxbudg_shift = log2(max_len * max_classes_per_agg) */
1444 maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift;
1445 q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX;
1446
1447 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1448 grp = &q->groups[i];
1449 grp->index = i;
1450 grp->slot_shift = q->min_slot_shift + i;
1451 for (j = 0; j < QFQ_MAX_SLOTS; j++)
1452 INIT_HLIST_HEAD(&grp->slots[j]);
1453 }
1454
1455 INIT_HLIST_HEAD(&q->nonfull_aggs);
1456
1457 return 0;
1458}
1459
1460static void qfq_reset_qdisc(struct Qdisc *sch)
1461{
1462 struct qfq_sched *q = qdisc_priv(sch);
1463 struct qfq_class *cl;
1464 unsigned int i;
1465
1466 for (i = 0; i < q->clhash.hashsize; i++) {
1467 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1468 if (cl->qdisc->q.qlen > 0)
1469 qfq_deactivate_class(q, cl);
1470
1471 qdisc_reset(cl->qdisc);
1472 }
1473 }
1474}
1475
1476static void qfq_destroy_qdisc(struct Qdisc *sch)
1477{
1478 struct qfq_sched *q = qdisc_priv(sch);
1479 struct qfq_class *cl;
1480 struct hlist_node *next;
1481 unsigned int i;
1482
1483 tcf_block_put(q->block);
1484
1485 for (i = 0; i < q->clhash.hashsize; i++) {
1486 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1487 common.hnode) {
1488 qfq_destroy_class(sch, cl);
1489 }
1490 }
1491 qdisc_class_hash_destroy(&q->clhash);
1492}
1493
1494static const struct Qdisc_class_ops qfq_class_ops = {
1495 .change = qfq_change_class,
1496 .delete = qfq_delete_class,
1497 .find = qfq_search_class,
1498 .tcf_block = qfq_tcf_block,
1499 .bind_tcf = qfq_bind_tcf,
1500 .unbind_tcf = qfq_unbind_tcf,
1501 .graft = qfq_graft_class,
1502 .leaf = qfq_class_leaf,
1503 .qlen_notify = qfq_qlen_notify,
1504 .dump = qfq_dump_class,
1505 .dump_stats = qfq_dump_class_stats,
1506 .walk = qfq_walk,
1507};
1508
1509static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1510 .cl_ops = &qfq_class_ops,
1511 .id = "qfq",
1512 .priv_size = sizeof(struct qfq_sched),
1513 .enqueue = qfq_enqueue,
1514 .dequeue = qfq_dequeue,
1515 .peek = qdisc_peek_dequeued,
1516 .init = qfq_init_qdisc,
1517 .reset = qfq_reset_qdisc,
1518 .destroy = qfq_destroy_qdisc,
1519 .owner = THIS_MODULE,
1520};
1521MODULE_ALIAS_NET_SCH("qfq");
1522
1523static int __init qfq_init(void)
1524{
1525 return register_qdisc(&qfq_qdisc_ops);
1526}
1527
1528static void __exit qfq_exit(void)
1529{
1530 unregister_qdisc(&qfq_qdisc_ops);
1531}
1532
1533module_init(qfq_init);
1534module_exit(qfq_exit);
1535MODULE_LICENSE("GPL");
1536MODULE_DESCRIPTION("Quick Fair Queueing Plus qdisc");