Loading...
1/*
2 * net/sched/sch_qfq.c Quick Fair Queueing Plus Scheduler.
3 *
4 * Copyright (c) 2009 Fabio Checconi, Luigi Rizzo, and Paolo Valente.
5 * Copyright (c) 2012 Paolo Valente.
6 *
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * version 2 as published by the Free Software Foundation.
10 */
11
12#include <linux/module.h>
13#include <linux/init.h>
14#include <linux/bitops.h>
15#include <linux/errno.h>
16#include <linux/netdevice.h>
17#include <linux/pkt_sched.h>
18#include <net/sch_generic.h>
19#include <net/pkt_sched.h>
20#include <net/pkt_cls.h>
21
22
23/* Quick Fair Queueing Plus
24 ========================
25
26 Sources:
27
28 [1] Paolo Valente,
29 "Reducing the Execution Time of Fair-Queueing Schedulers."
30 http://algo.ing.unimo.it/people/paolo/agg-sched/agg-sched.pdf
31
32 Sources for QFQ:
33
34 [2] Fabio Checconi, Luigi Rizzo, and Paolo Valente: "QFQ: Efficient
35 Packet Scheduling with Tight Bandwidth Distribution Guarantees."
36
37 See also:
38 http://retis.sssup.it/~fabio/linux/qfq/
39 */
40
41/*
42
43 QFQ+ divides classes into aggregates of at most MAX_AGG_CLASSES
44 classes. Each aggregate is timestamped with a virtual start time S
45 and a virtual finish time F, and scheduled according to its
46 timestamps. S and F are computed as a function of a system virtual
47 time function V. The classes within each aggregate are instead
48 scheduled with DRR.
49
50 To speed up operations, QFQ+ divides also aggregates into a limited
51 number of groups. Which group a class belongs to depends on the
52 ratio between the maximum packet length for the class and the weight
53 of the class. Groups have their own S and F. In the end, QFQ+
54 schedules groups, then aggregates within groups, then classes within
55 aggregates. See [1] and [2] for a full description.
56
57 Virtual time computations.
58
59 S, F and V are all computed in fixed point arithmetic with
60 FRAC_BITS decimal bits.
61
62 QFQ_MAX_INDEX is the maximum index allowed for a group. We need
63 one bit per index.
64 QFQ_MAX_WSHIFT is the maximum power of two supported as a weight.
65
66 The layout of the bits is as below:
67
68 [ MTU_SHIFT ][ FRAC_BITS ]
69 [ MAX_INDEX ][ MIN_SLOT_SHIFT ]
70 ^.__grp->index = 0
71 *.__grp->slot_shift
72
73 where MIN_SLOT_SHIFT is derived by difference from the others.
74
75 The max group index corresponds to Lmax/w_min, where
76 Lmax=1<<MTU_SHIFT, w_min = 1 .
77 From this, and knowing how many groups (MAX_INDEX) we want,
78 we can derive the shift corresponding to each group.
79
80 Because we often need to compute
81 F = S + len/w_i and V = V + len/wsum
82 instead of storing w_i store the value
83 inv_w = (1<<FRAC_BITS)/w_i
84 so we can do F = S + len * inv_w * wsum.
85 We use W_TOT in the formulas so we can easily move between
86 static and adaptive weight sum.
87
88 The per-scheduler-instance data contain all the data structures
89 for the scheduler: bitmaps and bucket lists.
90
91 */
92
93/*
94 * Maximum number of consecutive slots occupied by backlogged classes
95 * inside a group.
96 */
97#define QFQ_MAX_SLOTS 32
98
99/*
100 * Shifts used for aggregate<->group mapping. We allow class weights that are
101 * in the range [1, 2^MAX_WSHIFT], and we try to map each aggregate i to the
102 * group with the smallest index that can support the L_i / r_i configured
103 * for the classes in the aggregate.
104 *
105 * grp->index is the index of the group; and grp->slot_shift
106 * is the shift for the corresponding (scaled) sigma_i.
107 */
108#define QFQ_MAX_INDEX 24
109#define QFQ_MAX_WSHIFT 10
110
111#define QFQ_MAX_WEIGHT (1<<QFQ_MAX_WSHIFT) /* see qfq_slot_insert */
112#define QFQ_MAX_WSUM (64*QFQ_MAX_WEIGHT)
113
114#define FRAC_BITS 30 /* fixed point arithmetic */
115#define ONE_FP (1UL << FRAC_BITS)
116
117#define QFQ_MTU_SHIFT 16 /* to support TSO/GSO */
118#define QFQ_MIN_LMAX 512 /* see qfq_slot_insert */
119
120#define QFQ_MAX_AGG_CLASSES 8 /* max num classes per aggregate allowed */
121
122/*
123 * Possible group states. These values are used as indexes for the bitmaps
124 * array of struct qfq_queue.
125 */
126enum qfq_state { ER, IR, EB, IB, QFQ_MAX_STATE };
127
128struct qfq_group;
129
130struct qfq_aggregate;
131
132struct qfq_class {
133 struct Qdisc_class_common common;
134
135 unsigned int refcnt;
136 unsigned int filter_cnt;
137
138 struct gnet_stats_basic_packed bstats;
139 struct gnet_stats_queue qstats;
140 struct gnet_stats_rate_est64 rate_est;
141 struct Qdisc *qdisc;
142 struct list_head alist; /* Link for active-classes list. */
143 struct qfq_aggregate *agg; /* Parent aggregate. */
144 int deficit; /* DRR deficit counter. */
145};
146
147struct qfq_aggregate {
148 struct hlist_node next; /* Link for the slot list. */
149 u64 S, F; /* flow timestamps (exact) */
150
151 /* group we belong to. In principle we would need the index,
152 * which is log_2(lmax/weight), but we never reference it
153 * directly, only the group.
154 */
155 struct qfq_group *grp;
156
157 /* these are copied from the flowset. */
158 u32 class_weight; /* Weight of each class in this aggregate. */
159 /* Max pkt size for the classes in this aggregate, DRR quantum. */
160 int lmax;
161
162 u32 inv_w; /* ONE_FP/(sum of weights of classes in aggr.). */
163 u32 budgetmax; /* Max budget for this aggregate. */
164 u32 initial_budget, budget; /* Initial and current budget. */
165
166 int num_classes; /* Number of classes in this aggr. */
167 struct list_head active; /* DRR queue of active classes. */
168
169 struct hlist_node nonfull_next; /* See nonfull_aggs in qfq_sched. */
170};
171
172struct qfq_group {
173 u64 S, F; /* group timestamps (approx). */
174 unsigned int slot_shift; /* Slot shift. */
175 unsigned int index; /* Group index. */
176 unsigned int front; /* Index of the front slot. */
177 unsigned long full_slots; /* non-empty slots */
178
179 /* Array of RR lists of active aggregates. */
180 struct hlist_head slots[QFQ_MAX_SLOTS];
181};
182
183struct qfq_sched {
184 struct tcf_proto *filter_list;
185 struct Qdisc_class_hash clhash;
186
187 u64 oldV, V; /* Precise virtual times. */
188 struct qfq_aggregate *in_serv_agg; /* Aggregate being served. */
189 u32 num_active_agg; /* Num. of active aggregates */
190 u32 wsum; /* weight sum */
191 u32 iwsum; /* inverse weight sum */
192
193 unsigned long bitmaps[QFQ_MAX_STATE]; /* Group bitmaps. */
194 struct qfq_group groups[QFQ_MAX_INDEX + 1]; /* The groups. */
195 u32 min_slot_shift; /* Index of the group-0 bit in the bitmaps. */
196
197 u32 max_agg_classes; /* Max number of classes per aggr. */
198 struct hlist_head nonfull_aggs; /* Aggs with room for more classes. */
199};
200
201/*
202 * Possible reasons why the timestamps of an aggregate are updated
203 * enqueue: the aggregate switches from idle to active and must scheduled
204 * for service
205 * requeue: the aggregate finishes its budget, so it stops being served and
206 * must be rescheduled for service
207 */
208enum update_reason {enqueue, requeue};
209
210static struct qfq_class *qfq_find_class(struct Qdisc *sch, u32 classid)
211{
212 struct qfq_sched *q = qdisc_priv(sch);
213 struct Qdisc_class_common *clc;
214
215 clc = qdisc_class_find(&q->clhash, classid);
216 if (clc == NULL)
217 return NULL;
218 return container_of(clc, struct qfq_class, common);
219}
220
221static void qfq_purge_queue(struct qfq_class *cl)
222{
223 unsigned int len = cl->qdisc->q.qlen;
224
225 qdisc_reset(cl->qdisc);
226 qdisc_tree_decrease_qlen(cl->qdisc, len);
227}
228
229static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = {
230 [TCA_QFQ_WEIGHT] = { .type = NLA_U32 },
231 [TCA_QFQ_LMAX] = { .type = NLA_U32 },
232};
233
234/*
235 * Calculate a flow index, given its weight and maximum packet length.
236 * index = log_2(maxlen/weight) but we need to apply the scaling.
237 * This is used only once at flow creation.
238 */
239static int qfq_calc_index(u32 inv_w, unsigned int maxlen, u32 min_slot_shift)
240{
241 u64 slot_size = (u64)maxlen * inv_w;
242 unsigned long size_map;
243 int index = 0;
244
245 size_map = slot_size >> min_slot_shift;
246 if (!size_map)
247 goto out;
248
249 index = __fls(size_map) + 1; /* basically a log_2 */
250 index -= !(slot_size - (1ULL << (index + min_slot_shift - 1)));
251
252 if (index < 0)
253 index = 0;
254out:
255 pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n",
256 (unsigned long) ONE_FP/inv_w, maxlen, index);
257
258 return index;
259}
260
261static void qfq_deactivate_agg(struct qfq_sched *, struct qfq_aggregate *);
262static void qfq_activate_agg(struct qfq_sched *, struct qfq_aggregate *,
263 enum update_reason);
264
265static void qfq_init_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
266 u32 lmax, u32 weight)
267{
268 INIT_LIST_HEAD(&agg->active);
269 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
270
271 agg->lmax = lmax;
272 agg->class_weight = weight;
273}
274
275static struct qfq_aggregate *qfq_find_agg(struct qfq_sched *q,
276 u32 lmax, u32 weight)
277{
278 struct qfq_aggregate *agg;
279
280 hlist_for_each_entry(agg, &q->nonfull_aggs, nonfull_next)
281 if (agg->lmax == lmax && agg->class_weight == weight)
282 return agg;
283
284 return NULL;
285}
286
287
288/* Update aggregate as a function of the new number of classes. */
289static void qfq_update_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
290 int new_num_classes)
291{
292 u32 new_agg_weight;
293
294 if (new_num_classes == q->max_agg_classes)
295 hlist_del_init(&agg->nonfull_next);
296
297 if (agg->num_classes > new_num_classes &&
298 new_num_classes == q->max_agg_classes - 1) /* agg no more full */
299 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
300
301 /* The next assignment may let
302 * agg->initial_budget > agg->budgetmax
303 * hold, we will take it into account in charge_actual_service().
304 */
305 agg->budgetmax = new_num_classes * agg->lmax;
306 new_agg_weight = agg->class_weight * new_num_classes;
307 agg->inv_w = ONE_FP/new_agg_weight;
308
309 if (agg->grp == NULL) {
310 int i = qfq_calc_index(agg->inv_w, agg->budgetmax,
311 q->min_slot_shift);
312 agg->grp = &q->groups[i];
313 }
314
315 q->wsum +=
316 (int) agg->class_weight * (new_num_classes - agg->num_classes);
317 q->iwsum = ONE_FP / q->wsum;
318
319 agg->num_classes = new_num_classes;
320}
321
322/* Add class to aggregate. */
323static void qfq_add_to_agg(struct qfq_sched *q,
324 struct qfq_aggregate *agg,
325 struct qfq_class *cl)
326{
327 cl->agg = agg;
328
329 qfq_update_agg(q, agg, agg->num_classes+1);
330 if (cl->qdisc->q.qlen > 0) { /* adding an active class */
331 list_add_tail(&cl->alist, &agg->active);
332 if (list_first_entry(&agg->active, struct qfq_class, alist) ==
333 cl && q->in_serv_agg != agg) /* agg was inactive */
334 qfq_activate_agg(q, agg, enqueue); /* schedule agg */
335 }
336}
337
338static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *);
339
340static void qfq_destroy_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
341{
342 if (!hlist_unhashed(&agg->nonfull_next))
343 hlist_del_init(&agg->nonfull_next);
344 q->wsum -= agg->class_weight;
345 if (q->wsum != 0)
346 q->iwsum = ONE_FP / q->wsum;
347
348 if (q->in_serv_agg == agg)
349 q->in_serv_agg = qfq_choose_next_agg(q);
350 kfree(agg);
351}
352
353/* Deschedule class from within its parent aggregate. */
354static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
355{
356 struct qfq_aggregate *agg = cl->agg;
357
358
359 list_del(&cl->alist); /* remove from RR queue of the aggregate */
360 if (list_empty(&agg->active)) /* agg is now inactive */
361 qfq_deactivate_agg(q, agg);
362}
363
364/* Remove class from its parent aggregate. */
365static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
366{
367 struct qfq_aggregate *agg = cl->agg;
368
369 cl->agg = NULL;
370 if (agg->num_classes == 1) { /* agg being emptied, destroy it */
371 qfq_destroy_agg(q, agg);
372 return;
373 }
374 qfq_update_agg(q, agg, agg->num_classes-1);
375}
376
377/* Deschedule class and remove it from its parent aggregate. */
378static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
379{
380 if (cl->qdisc->q.qlen > 0) /* class is active */
381 qfq_deactivate_class(q, cl);
382
383 qfq_rm_from_agg(q, cl);
384}
385
386/* Move class to a new aggregate, matching the new class weight and/or lmax */
387static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight,
388 u32 lmax)
389{
390 struct qfq_sched *q = qdisc_priv(sch);
391 struct qfq_aggregate *new_agg = qfq_find_agg(q, lmax, weight);
392
393 if (new_agg == NULL) { /* create new aggregate */
394 new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC);
395 if (new_agg == NULL)
396 return -ENOBUFS;
397 qfq_init_agg(q, new_agg, lmax, weight);
398 }
399 qfq_deact_rm_from_agg(q, cl);
400 qfq_add_to_agg(q, new_agg, cl);
401
402 return 0;
403}
404
405static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
406 struct nlattr **tca, unsigned long *arg)
407{
408 struct qfq_sched *q = qdisc_priv(sch);
409 struct qfq_class *cl = (struct qfq_class *)*arg;
410 bool existing = false;
411 struct nlattr *tb[TCA_QFQ_MAX + 1];
412 struct qfq_aggregate *new_agg = NULL;
413 u32 weight, lmax, inv_w;
414 int err;
415 int delta_w;
416
417 if (tca[TCA_OPTIONS] == NULL) {
418 pr_notice("qfq: no options\n");
419 return -EINVAL;
420 }
421
422 err = nla_parse_nested(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS], qfq_policy);
423 if (err < 0)
424 return err;
425
426 if (tb[TCA_QFQ_WEIGHT]) {
427 weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]);
428 if (!weight || weight > (1UL << QFQ_MAX_WSHIFT)) {
429 pr_notice("qfq: invalid weight %u\n", weight);
430 return -EINVAL;
431 }
432 } else
433 weight = 1;
434
435 if (tb[TCA_QFQ_LMAX]) {
436 lmax = nla_get_u32(tb[TCA_QFQ_LMAX]);
437 if (lmax < QFQ_MIN_LMAX || lmax > (1UL << QFQ_MTU_SHIFT)) {
438 pr_notice("qfq: invalid max length %u\n", lmax);
439 return -EINVAL;
440 }
441 } else
442 lmax = psched_mtu(qdisc_dev(sch));
443
444 inv_w = ONE_FP / weight;
445 weight = ONE_FP / inv_w;
446
447 if (cl != NULL &&
448 lmax == cl->agg->lmax &&
449 weight == cl->agg->class_weight)
450 return 0; /* nothing to change */
451
452 delta_w = weight - (cl ? cl->agg->class_weight : 0);
453
454 if (q->wsum + delta_w > QFQ_MAX_WSUM) {
455 pr_notice("qfq: total weight out of range (%d + %u)\n",
456 delta_w, q->wsum);
457 return -EINVAL;
458 }
459
460 if (cl != NULL) { /* modify existing class */
461 if (tca[TCA_RATE]) {
462 err = gen_replace_estimator(&cl->bstats, &cl->rate_est,
463 qdisc_root_sleeping_lock(sch),
464 tca[TCA_RATE]);
465 if (err)
466 return err;
467 }
468 existing = true;
469 goto set_change_agg;
470 }
471
472 /* create and init new class */
473 cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
474 if (cl == NULL)
475 return -ENOBUFS;
476
477 cl->refcnt = 1;
478 cl->common.classid = classid;
479 cl->deficit = lmax;
480
481 cl->qdisc = qdisc_create_dflt(sch->dev_queue,
482 &pfifo_qdisc_ops, classid);
483 if (cl->qdisc == NULL)
484 cl->qdisc = &noop_qdisc;
485
486 if (tca[TCA_RATE]) {
487 err = gen_new_estimator(&cl->bstats, &cl->rate_est,
488 qdisc_root_sleeping_lock(sch),
489 tca[TCA_RATE]);
490 if (err)
491 goto destroy_class;
492 }
493
494 sch_tree_lock(sch);
495 qdisc_class_hash_insert(&q->clhash, &cl->common);
496 sch_tree_unlock(sch);
497
498 qdisc_class_hash_grow(sch, &q->clhash);
499
500set_change_agg:
501 sch_tree_lock(sch);
502 new_agg = qfq_find_agg(q, lmax, weight);
503 if (new_agg == NULL) { /* create new aggregate */
504 sch_tree_unlock(sch);
505 new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL);
506 if (new_agg == NULL) {
507 err = -ENOBUFS;
508 gen_kill_estimator(&cl->bstats, &cl->rate_est);
509 goto destroy_class;
510 }
511 sch_tree_lock(sch);
512 qfq_init_agg(q, new_agg, lmax, weight);
513 }
514 if (existing)
515 qfq_deact_rm_from_agg(q, cl);
516 qfq_add_to_agg(q, new_agg, cl);
517 sch_tree_unlock(sch);
518
519 *arg = (unsigned long)cl;
520 return 0;
521
522destroy_class:
523 qdisc_destroy(cl->qdisc);
524 kfree(cl);
525 return err;
526}
527
528static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
529{
530 struct qfq_sched *q = qdisc_priv(sch);
531
532 qfq_rm_from_agg(q, cl);
533 gen_kill_estimator(&cl->bstats, &cl->rate_est);
534 qdisc_destroy(cl->qdisc);
535 kfree(cl);
536}
537
538static int qfq_delete_class(struct Qdisc *sch, unsigned long arg)
539{
540 struct qfq_sched *q = qdisc_priv(sch);
541 struct qfq_class *cl = (struct qfq_class *)arg;
542
543 if (cl->filter_cnt > 0)
544 return -EBUSY;
545
546 sch_tree_lock(sch);
547
548 qfq_purge_queue(cl);
549 qdisc_class_hash_remove(&q->clhash, &cl->common);
550
551 BUG_ON(--cl->refcnt == 0);
552 /*
553 * This shouldn't happen: we "hold" one cops->get() when called
554 * from tc_ctl_tclass; the destroy method is done from cops->put().
555 */
556
557 sch_tree_unlock(sch);
558 return 0;
559}
560
561static unsigned long qfq_get_class(struct Qdisc *sch, u32 classid)
562{
563 struct qfq_class *cl = qfq_find_class(sch, classid);
564
565 if (cl != NULL)
566 cl->refcnt++;
567
568 return (unsigned long)cl;
569}
570
571static void qfq_put_class(struct Qdisc *sch, unsigned long arg)
572{
573 struct qfq_class *cl = (struct qfq_class *)arg;
574
575 if (--cl->refcnt == 0)
576 qfq_destroy_class(sch, cl);
577}
578
579static struct tcf_proto **qfq_tcf_chain(struct Qdisc *sch, unsigned long cl)
580{
581 struct qfq_sched *q = qdisc_priv(sch);
582
583 if (cl)
584 return NULL;
585
586 return &q->filter_list;
587}
588
589static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
590 u32 classid)
591{
592 struct qfq_class *cl = qfq_find_class(sch, classid);
593
594 if (cl != NULL)
595 cl->filter_cnt++;
596
597 return (unsigned long)cl;
598}
599
600static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
601{
602 struct qfq_class *cl = (struct qfq_class *)arg;
603
604 cl->filter_cnt--;
605}
606
607static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
608 struct Qdisc *new, struct Qdisc **old)
609{
610 struct qfq_class *cl = (struct qfq_class *)arg;
611
612 if (new == NULL) {
613 new = qdisc_create_dflt(sch->dev_queue,
614 &pfifo_qdisc_ops, cl->common.classid);
615 if (new == NULL)
616 new = &noop_qdisc;
617 }
618
619 sch_tree_lock(sch);
620 qfq_purge_queue(cl);
621 *old = cl->qdisc;
622 cl->qdisc = new;
623 sch_tree_unlock(sch);
624 return 0;
625}
626
627static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
628{
629 struct qfq_class *cl = (struct qfq_class *)arg;
630
631 return cl->qdisc;
632}
633
634static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
635 struct sk_buff *skb, struct tcmsg *tcm)
636{
637 struct qfq_class *cl = (struct qfq_class *)arg;
638 struct nlattr *nest;
639
640 tcm->tcm_parent = TC_H_ROOT;
641 tcm->tcm_handle = cl->common.classid;
642 tcm->tcm_info = cl->qdisc->handle;
643
644 nest = nla_nest_start(skb, TCA_OPTIONS);
645 if (nest == NULL)
646 goto nla_put_failure;
647 if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) ||
648 nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax))
649 goto nla_put_failure;
650 return nla_nest_end(skb, nest);
651
652nla_put_failure:
653 nla_nest_cancel(skb, nest);
654 return -EMSGSIZE;
655}
656
657static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
658 struct gnet_dump *d)
659{
660 struct qfq_class *cl = (struct qfq_class *)arg;
661 struct tc_qfq_stats xstats;
662
663 memset(&xstats, 0, sizeof(xstats));
664 cl->qdisc->qstats.qlen = cl->qdisc->q.qlen;
665
666 xstats.weight = cl->agg->class_weight;
667 xstats.lmax = cl->agg->lmax;
668
669 if (gnet_stats_copy_basic(d, &cl->bstats) < 0 ||
670 gnet_stats_copy_rate_est(d, &cl->bstats, &cl->rate_est) < 0 ||
671 gnet_stats_copy_queue(d, &cl->qdisc->qstats) < 0)
672 return -1;
673
674 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
675}
676
677static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
678{
679 struct qfq_sched *q = qdisc_priv(sch);
680 struct qfq_class *cl;
681 unsigned int i;
682
683 if (arg->stop)
684 return;
685
686 for (i = 0; i < q->clhash.hashsize; i++) {
687 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
688 if (arg->count < arg->skip) {
689 arg->count++;
690 continue;
691 }
692 if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
693 arg->stop = 1;
694 return;
695 }
696 arg->count++;
697 }
698 }
699}
700
701static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
702 int *qerr)
703{
704 struct qfq_sched *q = qdisc_priv(sch);
705 struct qfq_class *cl;
706 struct tcf_result res;
707 int result;
708
709 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
710 pr_debug("qfq_classify: found %d\n", skb->priority);
711 cl = qfq_find_class(sch, skb->priority);
712 if (cl != NULL)
713 return cl;
714 }
715
716 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
717 result = tc_classify(skb, q->filter_list, &res);
718 if (result >= 0) {
719#ifdef CONFIG_NET_CLS_ACT
720 switch (result) {
721 case TC_ACT_QUEUED:
722 case TC_ACT_STOLEN:
723 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
724 case TC_ACT_SHOT:
725 return NULL;
726 }
727#endif
728 cl = (struct qfq_class *)res.class;
729 if (cl == NULL)
730 cl = qfq_find_class(sch, res.classid);
731 return cl;
732 }
733
734 return NULL;
735}
736
737/* Generic comparison function, handling wraparound. */
738static inline int qfq_gt(u64 a, u64 b)
739{
740 return (s64)(a - b) > 0;
741}
742
743/* Round a precise timestamp to its slotted value. */
744static inline u64 qfq_round_down(u64 ts, unsigned int shift)
745{
746 return ts & ~((1ULL << shift) - 1);
747}
748
749/* return the pointer to the group with lowest index in the bitmap */
750static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
751 unsigned long bitmap)
752{
753 int index = __ffs(bitmap);
754 return &q->groups[index];
755}
756/* Calculate a mask to mimic what would be ffs_from(). */
757static inline unsigned long mask_from(unsigned long bitmap, int from)
758{
759 return bitmap & ~((1UL << from) - 1);
760}
761
762/*
763 * The state computation relies on ER=0, IR=1, EB=2, IB=3
764 * First compute eligibility comparing grp->S, q->V,
765 * then check if someone is blocking us and possibly add EB
766 */
767static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
768{
769 /* if S > V we are not eligible */
770 unsigned int state = qfq_gt(grp->S, q->V);
771 unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
772 struct qfq_group *next;
773
774 if (mask) {
775 next = qfq_ffs(q, mask);
776 if (qfq_gt(grp->F, next->F))
777 state |= EB;
778 }
779
780 return state;
781}
782
783
784/*
785 * In principle
786 * q->bitmaps[dst] |= q->bitmaps[src] & mask;
787 * q->bitmaps[src] &= ~mask;
788 * but we should make sure that src != dst
789 */
790static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
791 int src, int dst)
792{
793 q->bitmaps[dst] |= q->bitmaps[src] & mask;
794 q->bitmaps[src] &= ~mask;
795}
796
797static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
798{
799 unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
800 struct qfq_group *next;
801
802 if (mask) {
803 next = qfq_ffs(q, mask);
804 if (!qfq_gt(next->F, old_F))
805 return;
806 }
807
808 mask = (1UL << index) - 1;
809 qfq_move_groups(q, mask, EB, ER);
810 qfq_move_groups(q, mask, IB, IR);
811}
812
813/*
814 * perhaps
815 *
816 old_V ^= q->V;
817 old_V >>= q->min_slot_shift;
818 if (old_V) {
819 ...
820 }
821 *
822 */
823static void qfq_make_eligible(struct qfq_sched *q)
824{
825 unsigned long vslot = q->V >> q->min_slot_shift;
826 unsigned long old_vslot = q->oldV >> q->min_slot_shift;
827
828 if (vslot != old_vslot) {
829 unsigned long mask;
830 int last_flip_pos = fls(vslot ^ old_vslot);
831
832 if (last_flip_pos > 31) /* higher than the number of groups */
833 mask = ~0UL; /* make all groups eligible */
834 else
835 mask = (1UL << last_flip_pos) - 1;
836
837 qfq_move_groups(q, mask, IR, ER);
838 qfq_move_groups(q, mask, IB, EB);
839 }
840}
841
842/*
843 * The index of the slot in which the input aggregate agg is to be
844 * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2'
845 * and not a '-1' because the start time of the group may be moved
846 * backward by one slot after the aggregate has been inserted, and
847 * this would cause non-empty slots to be right-shifted by one
848 * position.
849 *
850 * QFQ+ fully satisfies this bound to the slot index if the parameters
851 * of the classes are not changed dynamically, and if QFQ+ never
852 * happens to postpone the service of agg unjustly, i.e., it never
853 * happens that the aggregate becomes backlogged and eligible, or just
854 * eligible, while an aggregate with a higher approximated finish time
855 * is being served. In particular, in this case QFQ+ guarantees that
856 * the timestamps of agg are low enough that the slot index is never
857 * higher than 2. Unfortunately, QFQ+ cannot provide the same
858 * guarantee if it happens to unjustly postpone the service of agg, or
859 * if the parameters of some class are changed.
860 *
861 * As for the first event, i.e., an out-of-order service, the
862 * upper bound to the slot index guaranteed by QFQ+ grows to
863 * 2 +
864 * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
865 * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1.
866 *
867 * The following function deals with this problem by backward-shifting
868 * the timestamps of agg, if needed, so as to guarantee that the slot
869 * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may
870 * cause the service of other aggregates to be postponed, yet the
871 * worst-case guarantees of these aggregates are not violated. In
872 * fact, in case of no out-of-order service, the timestamps of agg
873 * would have been even lower than they are after the backward shift,
874 * because QFQ+ would have guaranteed a maximum value equal to 2 for
875 * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose
876 * service is postponed because of the backward-shift would have
877 * however waited for the service of agg before being served.
878 *
879 * The other event that may cause the slot index to be higher than 2
880 * for agg is a recent change of the parameters of some class. If the
881 * weight of a class is increased or the lmax (max_pkt_size) of the
882 * class is decreased, then a new aggregate with smaller slot size
883 * than the original parent aggregate of the class may happen to be
884 * activated. The activation of this aggregate should be properly
885 * delayed to when the service of the class has finished in the ideal
886 * system tracked by QFQ+. If the activation of the aggregate is not
887 * delayed to this reference time instant, then this aggregate may be
888 * unjustly served before other aggregates waiting for service. This
889 * may cause the above bound to the slot index to be violated for some
890 * of these unlucky aggregates.
891 *
892 * Instead of delaying the activation of the new aggregate, which is
893 * quite complex, the above-discussed capping of the slot index is
894 * used to handle also the consequences of a change of the parameters
895 * of a class.
896 */
897static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg,
898 u64 roundedS)
899{
900 u64 slot = (roundedS - grp->S) >> grp->slot_shift;
901 unsigned int i; /* slot index in the bucket list */
902
903 if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
904 u64 deltaS = roundedS - grp->S -
905 ((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
906 agg->S -= deltaS;
907 agg->F -= deltaS;
908 slot = QFQ_MAX_SLOTS - 2;
909 }
910
911 i = (grp->front + slot) % QFQ_MAX_SLOTS;
912
913 hlist_add_head(&agg->next, &grp->slots[i]);
914 __set_bit(slot, &grp->full_slots);
915}
916
917/* Maybe introduce hlist_first_entry?? */
918static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp)
919{
920 return hlist_entry(grp->slots[grp->front].first,
921 struct qfq_aggregate, next);
922}
923
924/*
925 * remove the entry from the slot
926 */
927static void qfq_front_slot_remove(struct qfq_group *grp)
928{
929 struct qfq_aggregate *agg = qfq_slot_head(grp);
930
931 BUG_ON(!agg);
932 hlist_del(&agg->next);
933 if (hlist_empty(&grp->slots[grp->front]))
934 __clear_bit(0, &grp->full_slots);
935}
936
937/*
938 * Returns the first aggregate in the first non-empty bucket of the
939 * group. As a side effect, adjusts the bucket list so the first
940 * non-empty bucket is at position 0 in full_slots.
941 */
942static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp)
943{
944 unsigned int i;
945
946 pr_debug("qfq slot_scan: grp %u full %#lx\n",
947 grp->index, grp->full_slots);
948
949 if (grp->full_slots == 0)
950 return NULL;
951
952 i = __ffs(grp->full_slots); /* zero based */
953 if (i > 0) {
954 grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
955 grp->full_slots >>= i;
956 }
957
958 return qfq_slot_head(grp);
959}
960
961/*
962 * adjust the bucket list. When the start time of a group decreases,
963 * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
964 * move the objects. The mask of occupied slots must be shifted
965 * because we use ffs() to find the first non-empty slot.
966 * This covers decreases in the group's start time, but what about
967 * increases of the start time ?
968 * Here too we should make sure that i is less than 32
969 */
970static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
971{
972 unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
973
974 grp->full_slots <<= i;
975 grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
976}
977
978static void qfq_update_eligible(struct qfq_sched *q)
979{
980 struct qfq_group *grp;
981 unsigned long ineligible;
982
983 ineligible = q->bitmaps[IR] | q->bitmaps[IB];
984 if (ineligible) {
985 if (!q->bitmaps[ER]) {
986 grp = qfq_ffs(q, ineligible);
987 if (qfq_gt(grp->S, q->V))
988 q->V = grp->S;
989 }
990 qfq_make_eligible(q);
991 }
992}
993
994/* Dequeue head packet of the head class in the DRR queue of the aggregate. */
995static void agg_dequeue(struct qfq_aggregate *agg,
996 struct qfq_class *cl, unsigned int len)
997{
998 qdisc_dequeue_peeked(cl->qdisc);
999
1000 cl->deficit -= (int) len;
1001
1002 if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */
1003 list_del(&cl->alist);
1004 else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) {
1005 cl->deficit += agg->lmax;
1006 list_move_tail(&cl->alist, &agg->active);
1007 }
1008}
1009
1010static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg,
1011 struct qfq_class **cl,
1012 unsigned int *len)
1013{
1014 struct sk_buff *skb;
1015
1016 *cl = list_first_entry(&agg->active, struct qfq_class, alist);
1017 skb = (*cl)->qdisc->ops->peek((*cl)->qdisc);
1018 if (skb == NULL)
1019 WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
1020 else
1021 *len = qdisc_pkt_len(skb);
1022
1023 return skb;
1024}
1025
1026/* Update F according to the actual service received by the aggregate. */
1027static inline void charge_actual_service(struct qfq_aggregate *agg)
1028{
1029 /* Compute the service received by the aggregate, taking into
1030 * account that, after decreasing the number of classes in
1031 * agg, it may happen that
1032 * agg->initial_budget - agg->budget > agg->bugdetmax
1033 */
1034 u32 service_received = min(agg->budgetmax,
1035 agg->initial_budget - agg->budget);
1036
1037 agg->F = agg->S + (u64)service_received * agg->inv_w;
1038}
1039
1040/* Assign a reasonable start time for a new aggregate in group i.
1041 * Admissible values for \hat(F) are multiples of \sigma_i
1042 * no greater than V+\sigma_i . Larger values mean that
1043 * we had a wraparound so we consider the timestamp to be stale.
1044 *
1045 * If F is not stale and F >= V then we set S = F.
1046 * Otherwise we should assign S = V, but this may violate
1047 * the ordering in EB (see [2]). So, if we have groups in ER,
1048 * set S to the F_j of the first group j which would be blocking us.
1049 * We are guaranteed not to move S backward because
1050 * otherwise our group i would still be blocked.
1051 */
1052static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg)
1053{
1054 unsigned long mask;
1055 u64 limit, roundedF;
1056 int slot_shift = agg->grp->slot_shift;
1057
1058 roundedF = qfq_round_down(agg->F, slot_shift);
1059 limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
1060
1061 if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) {
1062 /* timestamp was stale */
1063 mask = mask_from(q->bitmaps[ER], agg->grp->index);
1064 if (mask) {
1065 struct qfq_group *next = qfq_ffs(q, mask);
1066 if (qfq_gt(roundedF, next->F)) {
1067 if (qfq_gt(limit, next->F))
1068 agg->S = next->F;
1069 else /* preserve timestamp correctness */
1070 agg->S = limit;
1071 return;
1072 }
1073 }
1074 agg->S = q->V;
1075 } else /* timestamp is not stale */
1076 agg->S = agg->F;
1077}
1078
1079/* Update the timestamps of agg before scheduling/rescheduling it for
1080 * service. In particular, assign to agg->F its maximum possible
1081 * value, i.e., the virtual finish time with which the aggregate
1082 * should be labeled if it used all its budget once in service.
1083 */
1084static inline void
1085qfq_update_agg_ts(struct qfq_sched *q,
1086 struct qfq_aggregate *agg, enum update_reason reason)
1087{
1088 if (reason != requeue)
1089 qfq_update_start(q, agg);
1090 else /* just charge agg for the service received */
1091 agg->S = agg->F;
1092
1093 agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w;
1094}
1095
1096static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg);
1097
1098static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
1099{
1100 struct qfq_sched *q = qdisc_priv(sch);
1101 struct qfq_aggregate *in_serv_agg = q->in_serv_agg;
1102 struct qfq_class *cl;
1103 struct sk_buff *skb = NULL;
1104 /* next-packet len, 0 means no more active classes in in-service agg */
1105 unsigned int len = 0;
1106
1107 if (in_serv_agg == NULL)
1108 return NULL;
1109
1110 if (!list_empty(&in_serv_agg->active))
1111 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1112
1113 /*
1114 * If there are no active classes in the in-service aggregate,
1115 * or if the aggregate has not enough budget to serve its next
1116 * class, then choose the next aggregate to serve.
1117 */
1118 if (len == 0 || in_serv_agg->budget < len) {
1119 charge_actual_service(in_serv_agg);
1120
1121 /* recharge the budget of the aggregate */
1122 in_serv_agg->initial_budget = in_serv_agg->budget =
1123 in_serv_agg->budgetmax;
1124
1125 if (!list_empty(&in_serv_agg->active)) {
1126 /*
1127 * Still active: reschedule for
1128 * service. Possible optimization: if no other
1129 * aggregate is active, then there is no point
1130 * in rescheduling this aggregate, and we can
1131 * just keep it as the in-service one. This
1132 * should be however a corner case, and to
1133 * handle it, we would need to maintain an
1134 * extra num_active_aggs field.
1135 */
1136 qfq_update_agg_ts(q, in_serv_agg, requeue);
1137 qfq_schedule_agg(q, in_serv_agg);
1138 } else if (sch->q.qlen == 0) { /* no aggregate to serve */
1139 q->in_serv_agg = NULL;
1140 return NULL;
1141 }
1142
1143 /*
1144 * If we get here, there are other aggregates queued:
1145 * choose the new aggregate to serve.
1146 */
1147 in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
1148 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1149 }
1150 if (!skb)
1151 return NULL;
1152
1153 sch->q.qlen--;
1154 qdisc_bstats_update(sch, skb);
1155
1156 agg_dequeue(in_serv_agg, cl, len);
1157 /* If lmax is lowered, through qfq_change_class, for a class
1158 * owning pending packets with larger size than the new value
1159 * of lmax, then the following condition may hold.
1160 */
1161 if (unlikely(in_serv_agg->budget < len))
1162 in_serv_agg->budget = 0;
1163 else
1164 in_serv_agg->budget -= len;
1165
1166 q->V += (u64)len * q->iwsum;
1167 pr_debug("qfq dequeue: len %u F %lld now %lld\n",
1168 len, (unsigned long long) in_serv_agg->F,
1169 (unsigned long long) q->V);
1170
1171 return skb;
1172}
1173
1174static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)
1175{
1176 struct qfq_group *grp;
1177 struct qfq_aggregate *agg, *new_front_agg;
1178 u64 old_F;
1179
1180 qfq_update_eligible(q);
1181 q->oldV = q->V;
1182
1183 if (!q->bitmaps[ER])
1184 return NULL;
1185
1186 grp = qfq_ffs(q, q->bitmaps[ER]);
1187 old_F = grp->F;
1188
1189 agg = qfq_slot_head(grp);
1190
1191 /* agg starts to be served, remove it from schedule */
1192 qfq_front_slot_remove(grp);
1193
1194 new_front_agg = qfq_slot_scan(grp);
1195
1196 if (new_front_agg == NULL) /* group is now inactive, remove from ER */
1197 __clear_bit(grp->index, &q->bitmaps[ER]);
1198 else {
1199 u64 roundedS = qfq_round_down(new_front_agg->S,
1200 grp->slot_shift);
1201 unsigned int s;
1202
1203 if (grp->S == roundedS)
1204 return agg;
1205 grp->S = roundedS;
1206 grp->F = roundedS + (2ULL << grp->slot_shift);
1207 __clear_bit(grp->index, &q->bitmaps[ER]);
1208 s = qfq_calc_state(q, grp);
1209 __set_bit(grp->index, &q->bitmaps[s]);
1210 }
1211
1212 qfq_unblock_groups(q, grp->index, old_F);
1213
1214 return agg;
1215}
1216
1217static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
1218{
1219 struct qfq_sched *q = qdisc_priv(sch);
1220 struct qfq_class *cl;
1221 struct qfq_aggregate *agg;
1222 int err = 0;
1223
1224 cl = qfq_classify(skb, sch, &err);
1225 if (cl == NULL) {
1226 if (err & __NET_XMIT_BYPASS)
1227 sch->qstats.drops++;
1228 kfree_skb(skb);
1229 return err;
1230 }
1231 pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
1232
1233 if (unlikely(cl->agg->lmax < qdisc_pkt_len(skb))) {
1234 pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
1235 cl->agg->lmax, qdisc_pkt_len(skb), cl->common.classid);
1236 err = qfq_change_agg(sch, cl, cl->agg->class_weight,
1237 qdisc_pkt_len(skb));
1238 if (err)
1239 return err;
1240 }
1241
1242 err = qdisc_enqueue(skb, cl->qdisc);
1243 if (unlikely(err != NET_XMIT_SUCCESS)) {
1244 pr_debug("qfq_enqueue: enqueue failed %d\n", err);
1245 if (net_xmit_drop_count(err)) {
1246 cl->qstats.drops++;
1247 sch->qstats.drops++;
1248 }
1249 return err;
1250 }
1251
1252 bstats_update(&cl->bstats, skb);
1253 ++sch->q.qlen;
1254
1255 agg = cl->agg;
1256 /* if the queue was not empty, then done here */
1257 if (cl->qdisc->q.qlen != 1) {
1258 if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) &&
1259 list_first_entry(&agg->active, struct qfq_class, alist)
1260 == cl && cl->deficit < qdisc_pkt_len(skb))
1261 list_move_tail(&cl->alist, &agg->active);
1262
1263 return err;
1264 }
1265
1266 /* schedule class for service within the aggregate */
1267 cl->deficit = agg->lmax;
1268 list_add_tail(&cl->alist, &agg->active);
1269
1270 if (list_first_entry(&agg->active, struct qfq_class, alist) != cl ||
1271 q->in_serv_agg == agg)
1272 return err; /* non-empty or in service, nothing else to do */
1273
1274 qfq_activate_agg(q, agg, enqueue);
1275
1276 return err;
1277}
1278
1279/*
1280 * Schedule aggregate according to its timestamps.
1281 */
1282static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1283{
1284 struct qfq_group *grp = agg->grp;
1285 u64 roundedS;
1286 int s;
1287
1288 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1289
1290 /*
1291 * Insert agg in the correct bucket.
1292 * If agg->S >= grp->S we don't need to adjust the
1293 * bucket list and simply go to the insertion phase.
1294 * Otherwise grp->S is decreasing, we must make room
1295 * in the bucket list, and also recompute the group state.
1296 * Finally, if there were no flows in this group and nobody
1297 * was in ER make sure to adjust V.
1298 */
1299 if (grp->full_slots) {
1300 if (!qfq_gt(grp->S, agg->S))
1301 goto skip_update;
1302
1303 /* create a slot for this agg->S */
1304 qfq_slot_rotate(grp, roundedS);
1305 /* group was surely ineligible, remove */
1306 __clear_bit(grp->index, &q->bitmaps[IR]);
1307 __clear_bit(grp->index, &q->bitmaps[IB]);
1308 } else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) &&
1309 q->in_serv_agg == NULL)
1310 q->V = roundedS;
1311
1312 grp->S = roundedS;
1313 grp->F = roundedS + (2ULL << grp->slot_shift);
1314 s = qfq_calc_state(q, grp);
1315 __set_bit(grp->index, &q->bitmaps[s]);
1316
1317 pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
1318 s, q->bitmaps[s],
1319 (unsigned long long) agg->S,
1320 (unsigned long long) agg->F,
1321 (unsigned long long) q->V);
1322
1323skip_update:
1324 qfq_slot_insert(grp, agg, roundedS);
1325}
1326
1327
1328/* Update agg ts and schedule agg for service */
1329static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
1330 enum update_reason reason)
1331{
1332 agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */
1333
1334 qfq_update_agg_ts(q, agg, reason);
1335 if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */
1336 q->in_serv_agg = agg; /* start serving this aggregate */
1337 /* update V: to be in service, agg must be eligible */
1338 q->oldV = q->V = agg->S;
1339 } else if (agg != q->in_serv_agg)
1340 qfq_schedule_agg(q, agg);
1341}
1342
1343static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
1344 struct qfq_aggregate *agg)
1345{
1346 unsigned int i, offset;
1347 u64 roundedS;
1348
1349 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1350 offset = (roundedS - grp->S) >> grp->slot_shift;
1351
1352 i = (grp->front + offset) % QFQ_MAX_SLOTS;
1353
1354 hlist_del(&agg->next);
1355 if (hlist_empty(&grp->slots[i]))
1356 __clear_bit(offset, &grp->full_slots);
1357}
1358
1359/*
1360 * Called to forcibly deschedule an aggregate. If the aggregate is
1361 * not in the front bucket, or if the latter has other aggregates in
1362 * the front bucket, we can simply remove the aggregate with no other
1363 * side effects.
1364 * Otherwise we must propagate the event up.
1365 */
1366static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1367{
1368 struct qfq_group *grp = agg->grp;
1369 unsigned long mask;
1370 u64 roundedS;
1371 int s;
1372
1373 if (agg == q->in_serv_agg) {
1374 charge_actual_service(agg);
1375 q->in_serv_agg = qfq_choose_next_agg(q);
1376 return;
1377 }
1378
1379 agg->F = agg->S;
1380 qfq_slot_remove(q, grp, agg);
1381
1382 if (!grp->full_slots) {
1383 __clear_bit(grp->index, &q->bitmaps[IR]);
1384 __clear_bit(grp->index, &q->bitmaps[EB]);
1385 __clear_bit(grp->index, &q->bitmaps[IB]);
1386
1387 if (test_bit(grp->index, &q->bitmaps[ER]) &&
1388 !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
1389 mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
1390 if (mask)
1391 mask = ~((1UL << __fls(mask)) - 1);
1392 else
1393 mask = ~0UL;
1394 qfq_move_groups(q, mask, EB, ER);
1395 qfq_move_groups(q, mask, IB, IR);
1396 }
1397 __clear_bit(grp->index, &q->bitmaps[ER]);
1398 } else if (hlist_empty(&grp->slots[grp->front])) {
1399 agg = qfq_slot_scan(grp);
1400 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1401 if (grp->S != roundedS) {
1402 __clear_bit(grp->index, &q->bitmaps[ER]);
1403 __clear_bit(grp->index, &q->bitmaps[IR]);
1404 __clear_bit(grp->index, &q->bitmaps[EB]);
1405 __clear_bit(grp->index, &q->bitmaps[IB]);
1406 grp->S = roundedS;
1407 grp->F = roundedS + (2ULL << grp->slot_shift);
1408 s = qfq_calc_state(q, grp);
1409 __set_bit(grp->index, &q->bitmaps[s]);
1410 }
1411 }
1412}
1413
1414static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1415{
1416 struct qfq_sched *q = qdisc_priv(sch);
1417 struct qfq_class *cl = (struct qfq_class *)arg;
1418
1419 if (cl->qdisc->q.qlen == 0)
1420 qfq_deactivate_class(q, cl);
1421}
1422
1423static unsigned int qfq_drop_from_slot(struct qfq_sched *q,
1424 struct hlist_head *slot)
1425{
1426 struct qfq_aggregate *agg;
1427 struct qfq_class *cl;
1428 unsigned int len;
1429
1430 hlist_for_each_entry(agg, slot, next) {
1431 list_for_each_entry(cl, &agg->active, alist) {
1432
1433 if (!cl->qdisc->ops->drop)
1434 continue;
1435
1436 len = cl->qdisc->ops->drop(cl->qdisc);
1437 if (len > 0) {
1438 if (cl->qdisc->q.qlen == 0)
1439 qfq_deactivate_class(q, cl);
1440
1441 return len;
1442 }
1443 }
1444 }
1445 return 0;
1446}
1447
1448static unsigned int qfq_drop(struct Qdisc *sch)
1449{
1450 struct qfq_sched *q = qdisc_priv(sch);
1451 struct qfq_group *grp;
1452 unsigned int i, j, len;
1453
1454 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1455 grp = &q->groups[i];
1456 for (j = 0; j < QFQ_MAX_SLOTS; j++) {
1457 len = qfq_drop_from_slot(q, &grp->slots[j]);
1458 if (len > 0) {
1459 sch->q.qlen--;
1460 return len;
1461 }
1462 }
1463
1464 }
1465
1466 return 0;
1467}
1468
1469static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
1470{
1471 struct qfq_sched *q = qdisc_priv(sch);
1472 struct qfq_group *grp;
1473 int i, j, err;
1474 u32 max_cl_shift, maxbudg_shift, max_classes;
1475
1476 err = qdisc_class_hash_init(&q->clhash);
1477 if (err < 0)
1478 return err;
1479
1480 if (qdisc_dev(sch)->tx_queue_len + 1 > QFQ_MAX_AGG_CLASSES)
1481 max_classes = QFQ_MAX_AGG_CLASSES;
1482 else
1483 max_classes = qdisc_dev(sch)->tx_queue_len + 1;
1484 /* max_cl_shift = floor(log_2(max_classes)) */
1485 max_cl_shift = __fls(max_classes);
1486 q->max_agg_classes = 1<<max_cl_shift;
1487
1488 /* maxbudg_shift = log2(max_len * max_classes_per_agg) */
1489 maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift;
1490 q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX;
1491
1492 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1493 grp = &q->groups[i];
1494 grp->index = i;
1495 grp->slot_shift = q->min_slot_shift + i;
1496 for (j = 0; j < QFQ_MAX_SLOTS; j++)
1497 INIT_HLIST_HEAD(&grp->slots[j]);
1498 }
1499
1500 INIT_HLIST_HEAD(&q->nonfull_aggs);
1501
1502 return 0;
1503}
1504
1505static void qfq_reset_qdisc(struct Qdisc *sch)
1506{
1507 struct qfq_sched *q = qdisc_priv(sch);
1508 struct qfq_class *cl;
1509 unsigned int i;
1510
1511 for (i = 0; i < q->clhash.hashsize; i++) {
1512 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1513 if (cl->qdisc->q.qlen > 0)
1514 qfq_deactivate_class(q, cl);
1515
1516 qdisc_reset(cl->qdisc);
1517 }
1518 }
1519 sch->q.qlen = 0;
1520}
1521
1522static void qfq_destroy_qdisc(struct Qdisc *sch)
1523{
1524 struct qfq_sched *q = qdisc_priv(sch);
1525 struct qfq_class *cl;
1526 struct hlist_node *next;
1527 unsigned int i;
1528
1529 tcf_destroy_chain(&q->filter_list);
1530
1531 for (i = 0; i < q->clhash.hashsize; i++) {
1532 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1533 common.hnode) {
1534 qfq_destroy_class(sch, cl);
1535 }
1536 }
1537 qdisc_class_hash_destroy(&q->clhash);
1538}
1539
1540static const struct Qdisc_class_ops qfq_class_ops = {
1541 .change = qfq_change_class,
1542 .delete = qfq_delete_class,
1543 .get = qfq_get_class,
1544 .put = qfq_put_class,
1545 .tcf_chain = qfq_tcf_chain,
1546 .bind_tcf = qfq_bind_tcf,
1547 .unbind_tcf = qfq_unbind_tcf,
1548 .graft = qfq_graft_class,
1549 .leaf = qfq_class_leaf,
1550 .qlen_notify = qfq_qlen_notify,
1551 .dump = qfq_dump_class,
1552 .dump_stats = qfq_dump_class_stats,
1553 .walk = qfq_walk,
1554};
1555
1556static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1557 .cl_ops = &qfq_class_ops,
1558 .id = "qfq",
1559 .priv_size = sizeof(struct qfq_sched),
1560 .enqueue = qfq_enqueue,
1561 .dequeue = qfq_dequeue,
1562 .peek = qdisc_peek_dequeued,
1563 .drop = qfq_drop,
1564 .init = qfq_init_qdisc,
1565 .reset = qfq_reset_qdisc,
1566 .destroy = qfq_destroy_qdisc,
1567 .owner = THIS_MODULE,
1568};
1569
1570static int __init qfq_init(void)
1571{
1572 return register_qdisc(&qfq_qdisc_ops);
1573}
1574
1575static void __exit qfq_exit(void)
1576{
1577 unregister_qdisc(&qfq_qdisc_ops);
1578}
1579
1580module_init(qfq_init);
1581module_exit(qfq_exit);
1582MODULE_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
117#define QFQ_MAX_AGG_CLASSES 8 /* max num classes per aggregate allowed */
118
119/*
120 * Possible group states. These values are used as indexes for the bitmaps
121 * array of struct qfq_queue.
122 */
123enum qfq_state { ER, IR, EB, IB, QFQ_MAX_STATE };
124
125struct qfq_group;
126
127struct qfq_aggregate;
128
129struct qfq_class {
130 struct Qdisc_class_common common;
131
132 unsigned int filter_cnt;
133
134 struct gnet_stats_basic_sync bstats;
135 struct gnet_stats_queue qstats;
136 struct net_rate_estimator __rcu *rate_est;
137 struct Qdisc *qdisc;
138 struct list_head alist; /* Link for active-classes list. */
139 struct qfq_aggregate *agg; /* Parent aggregate. */
140 int deficit; /* DRR deficit counter. */
141};
142
143struct qfq_aggregate {
144 struct hlist_node next; /* Link for the slot list. */
145 u64 S, F; /* flow timestamps (exact) */
146
147 /* group we belong to. In principle we would need the index,
148 * which is log_2(lmax/weight), but we never reference it
149 * directly, only the group.
150 */
151 struct qfq_group *grp;
152
153 /* these are copied from the flowset. */
154 u32 class_weight; /* Weight of each class in this aggregate. */
155 /* Max pkt size for the classes in this aggregate, DRR quantum. */
156 int lmax;
157
158 u32 inv_w; /* ONE_FP/(sum of weights of classes in aggr.). */
159 u32 budgetmax; /* Max budget for this aggregate. */
160 u32 initial_budget, budget; /* Initial and current budget. */
161
162 int num_classes; /* Number of classes in this aggr. */
163 struct list_head active; /* DRR queue of active classes. */
164
165 struct hlist_node nonfull_next; /* See nonfull_aggs in qfq_sched. */
166};
167
168struct qfq_group {
169 u64 S, F; /* group timestamps (approx). */
170 unsigned int slot_shift; /* Slot shift. */
171 unsigned int index; /* Group index. */
172 unsigned int front; /* Index of the front slot. */
173 unsigned long full_slots; /* non-empty slots */
174
175 /* Array of RR lists of active aggregates. */
176 struct hlist_head slots[QFQ_MAX_SLOTS];
177};
178
179struct qfq_sched {
180 struct tcf_proto __rcu *filter_list;
181 struct tcf_block *block;
182 struct Qdisc_class_hash clhash;
183
184 u64 oldV, V; /* Precise virtual times. */
185 struct qfq_aggregate *in_serv_agg; /* Aggregate being served. */
186 u32 wsum; /* weight sum */
187 u32 iwsum; /* inverse weight sum */
188
189 unsigned long bitmaps[QFQ_MAX_STATE]; /* Group bitmaps. */
190 struct qfq_group groups[QFQ_MAX_INDEX + 1]; /* The groups. */
191 u32 min_slot_shift; /* Index of the group-0 bit in the bitmaps. */
192
193 u32 max_agg_classes; /* Max number of classes per aggr. */
194 struct hlist_head nonfull_aggs; /* Aggs with room for more classes. */
195};
196
197/*
198 * Possible reasons why the timestamps of an aggregate are updated
199 * enqueue: the aggregate switches from idle to active and must scheduled
200 * for service
201 * requeue: the aggregate finishes its budget, so it stops being served and
202 * must be rescheduled for service
203 */
204enum update_reason {enqueue, requeue};
205
206static struct qfq_class *qfq_find_class(struct Qdisc *sch, u32 classid)
207{
208 struct qfq_sched *q = qdisc_priv(sch);
209 struct Qdisc_class_common *clc;
210
211 clc = qdisc_class_find(&q->clhash, classid);
212 if (clc == NULL)
213 return NULL;
214 return container_of(clc, struct qfq_class, common);
215}
216
217static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = {
218 [TCA_QFQ_WEIGHT] = { .type = NLA_U32 },
219 [TCA_QFQ_LMAX] = { .type = NLA_U32 },
220};
221
222/*
223 * Calculate a flow index, given its weight and maximum packet length.
224 * index = log_2(maxlen/weight) but we need to apply the scaling.
225 * This is used only once at flow creation.
226 */
227static int qfq_calc_index(u32 inv_w, unsigned int maxlen, u32 min_slot_shift)
228{
229 u64 slot_size = (u64)maxlen * inv_w;
230 unsigned long size_map;
231 int index = 0;
232
233 size_map = slot_size >> min_slot_shift;
234 if (!size_map)
235 goto out;
236
237 index = __fls(size_map) + 1; /* basically a log_2 */
238 index -= !(slot_size - (1ULL << (index + min_slot_shift - 1)));
239
240 if (index < 0)
241 index = 0;
242out:
243 pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n",
244 (unsigned long) ONE_FP/inv_w, maxlen, index);
245
246 return index;
247}
248
249static void qfq_deactivate_agg(struct qfq_sched *, struct qfq_aggregate *);
250static void qfq_activate_agg(struct qfq_sched *, struct qfq_aggregate *,
251 enum update_reason);
252
253static void qfq_init_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
254 u32 lmax, u32 weight)
255{
256 INIT_LIST_HEAD(&agg->active);
257 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
258
259 agg->lmax = lmax;
260 agg->class_weight = weight;
261}
262
263static struct qfq_aggregate *qfq_find_agg(struct qfq_sched *q,
264 u32 lmax, u32 weight)
265{
266 struct qfq_aggregate *agg;
267
268 hlist_for_each_entry(agg, &q->nonfull_aggs, nonfull_next)
269 if (agg->lmax == lmax && agg->class_weight == weight)
270 return agg;
271
272 return NULL;
273}
274
275
276/* Update aggregate as a function of the new number of classes. */
277static void qfq_update_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
278 int new_num_classes)
279{
280 u32 new_agg_weight;
281
282 if (new_num_classes == q->max_agg_classes)
283 hlist_del_init(&agg->nonfull_next);
284
285 if (agg->num_classes > new_num_classes &&
286 new_num_classes == q->max_agg_classes - 1) /* agg no more full */
287 hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
288
289 /* The next assignment may let
290 * agg->initial_budget > agg->budgetmax
291 * hold, we will take it into account in charge_actual_service().
292 */
293 agg->budgetmax = new_num_classes * agg->lmax;
294 new_agg_weight = agg->class_weight * new_num_classes;
295 agg->inv_w = ONE_FP/new_agg_weight;
296
297 if (agg->grp == NULL) {
298 int i = qfq_calc_index(agg->inv_w, agg->budgetmax,
299 q->min_slot_shift);
300 agg->grp = &q->groups[i];
301 }
302
303 q->wsum +=
304 (int) agg->class_weight * (new_num_classes - agg->num_classes);
305 q->iwsum = ONE_FP / q->wsum;
306
307 agg->num_classes = new_num_classes;
308}
309
310/* Add class to aggregate. */
311static void qfq_add_to_agg(struct qfq_sched *q,
312 struct qfq_aggregate *agg,
313 struct qfq_class *cl)
314{
315 cl->agg = agg;
316
317 qfq_update_agg(q, agg, agg->num_classes+1);
318 if (cl->qdisc->q.qlen > 0) { /* adding an active class */
319 list_add_tail(&cl->alist, &agg->active);
320 if (list_first_entry(&agg->active, struct qfq_class, alist) ==
321 cl && q->in_serv_agg != agg) /* agg was inactive */
322 qfq_activate_agg(q, agg, enqueue); /* schedule agg */
323 }
324}
325
326static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *);
327
328static void qfq_destroy_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
329{
330 hlist_del_init(&agg->nonfull_next);
331 q->wsum -= agg->class_weight;
332 if (q->wsum != 0)
333 q->iwsum = ONE_FP / q->wsum;
334
335 if (q->in_serv_agg == agg)
336 q->in_serv_agg = qfq_choose_next_agg(q);
337 kfree(agg);
338}
339
340/* Deschedule class from within its parent aggregate. */
341static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
342{
343 struct qfq_aggregate *agg = cl->agg;
344
345
346 list_del(&cl->alist); /* remove from RR queue of the aggregate */
347 if (list_empty(&agg->active)) /* agg is now inactive */
348 qfq_deactivate_agg(q, agg);
349}
350
351/* Remove class from its parent aggregate. */
352static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
353{
354 struct qfq_aggregate *agg = cl->agg;
355
356 cl->agg = NULL;
357 if (agg->num_classes == 1) { /* agg being emptied, destroy it */
358 qfq_destroy_agg(q, agg);
359 return;
360 }
361 qfq_update_agg(q, agg, agg->num_classes-1);
362}
363
364/* Deschedule class and remove it from its parent aggregate. */
365static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
366{
367 if (cl->qdisc->q.qlen > 0) /* class is active */
368 qfq_deactivate_class(q, cl);
369
370 qfq_rm_from_agg(q, cl);
371}
372
373/* Move class to a new aggregate, matching the new class weight and/or lmax */
374static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight,
375 u32 lmax)
376{
377 struct qfq_sched *q = qdisc_priv(sch);
378 struct qfq_aggregate *new_agg = qfq_find_agg(q, lmax, weight);
379
380 if (new_agg == NULL) { /* create new aggregate */
381 new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC);
382 if (new_agg == NULL)
383 return -ENOBUFS;
384 qfq_init_agg(q, new_agg, lmax, weight);
385 }
386 qfq_deact_rm_from_agg(q, cl);
387 qfq_add_to_agg(q, new_agg, cl);
388
389 return 0;
390}
391
392static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
393 struct nlattr **tca, unsigned long *arg,
394 struct netlink_ext_ack *extack)
395{
396 struct qfq_sched *q = qdisc_priv(sch);
397 struct qfq_class *cl = (struct qfq_class *)*arg;
398 bool existing = false;
399 struct nlattr *tb[TCA_QFQ_MAX + 1];
400 struct qfq_aggregate *new_agg = NULL;
401 u32 weight, lmax, inv_w;
402 int err;
403 int delta_w;
404
405 if (tca[TCA_OPTIONS] == NULL) {
406 pr_notice("qfq: no options\n");
407 return -EINVAL;
408 }
409
410 err = nla_parse_nested_deprecated(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS],
411 qfq_policy, NULL);
412 if (err < 0)
413 return err;
414
415 if (tb[TCA_QFQ_WEIGHT]) {
416 weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]);
417 if (!weight || weight > (1UL << QFQ_MAX_WSHIFT)) {
418 pr_notice("qfq: invalid weight %u\n", weight);
419 return -EINVAL;
420 }
421 } else
422 weight = 1;
423
424 if (tb[TCA_QFQ_LMAX]) {
425 lmax = nla_get_u32(tb[TCA_QFQ_LMAX]);
426 if (lmax < QFQ_MIN_LMAX || lmax > (1UL << QFQ_MTU_SHIFT)) {
427 pr_notice("qfq: invalid max length %u\n", lmax);
428 return -EINVAL;
429 }
430 } else
431 lmax = psched_mtu(qdisc_dev(sch));
432
433 inv_w = ONE_FP / weight;
434 weight = ONE_FP / inv_w;
435
436 if (cl != NULL &&
437 lmax == cl->agg->lmax &&
438 weight == cl->agg->class_weight)
439 return 0; /* nothing to change */
440
441 delta_w = weight - (cl ? cl->agg->class_weight : 0);
442
443 if (q->wsum + delta_w > QFQ_MAX_WSUM) {
444 pr_notice("qfq: total weight out of range (%d + %u)\n",
445 delta_w, q->wsum);
446 return -EINVAL;
447 }
448
449 if (cl != NULL) { /* modify existing class */
450 if (tca[TCA_RATE]) {
451 err = gen_replace_estimator(&cl->bstats, NULL,
452 &cl->rate_est,
453 NULL,
454 true,
455 tca[TCA_RATE]);
456 if (err)
457 return err;
458 }
459 existing = true;
460 goto set_change_agg;
461 }
462
463 /* create and init new class */
464 cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
465 if (cl == NULL)
466 return -ENOBUFS;
467
468 gnet_stats_basic_sync_init(&cl->bstats);
469 cl->common.classid = classid;
470 cl->deficit = lmax;
471
472 cl->qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
473 classid, NULL);
474 if (cl->qdisc == NULL)
475 cl->qdisc = &noop_qdisc;
476
477 if (tca[TCA_RATE]) {
478 err = gen_new_estimator(&cl->bstats, NULL,
479 &cl->rate_est,
480 NULL,
481 true,
482 tca[TCA_RATE]);
483 if (err)
484 goto destroy_class;
485 }
486
487 if (cl->qdisc != &noop_qdisc)
488 qdisc_hash_add(cl->qdisc, true);
489
490set_change_agg:
491 sch_tree_lock(sch);
492 new_agg = qfq_find_agg(q, lmax, weight);
493 if (new_agg == NULL) { /* create new aggregate */
494 sch_tree_unlock(sch);
495 new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL);
496 if (new_agg == NULL) {
497 err = -ENOBUFS;
498 gen_kill_estimator(&cl->rate_est);
499 goto destroy_class;
500 }
501 sch_tree_lock(sch);
502 qfq_init_agg(q, new_agg, lmax, weight);
503 }
504 if (existing)
505 qfq_deact_rm_from_agg(q, cl);
506 else
507 qdisc_class_hash_insert(&q->clhash, &cl->common);
508 qfq_add_to_agg(q, new_agg, cl);
509 sch_tree_unlock(sch);
510 qdisc_class_hash_grow(sch, &q->clhash);
511
512 *arg = (unsigned long)cl;
513 return 0;
514
515destroy_class:
516 qdisc_put(cl->qdisc);
517 kfree(cl);
518 return err;
519}
520
521static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
522{
523 struct qfq_sched *q = qdisc_priv(sch);
524
525 qfq_rm_from_agg(q, cl);
526 gen_kill_estimator(&cl->rate_est);
527 qdisc_put(cl->qdisc);
528 kfree(cl);
529}
530
531static int qfq_delete_class(struct Qdisc *sch, unsigned long arg,
532 struct netlink_ext_ack *extack)
533{
534 struct qfq_sched *q = qdisc_priv(sch);
535 struct qfq_class *cl = (struct qfq_class *)arg;
536
537 if (cl->filter_cnt > 0)
538 return -EBUSY;
539
540 sch_tree_lock(sch);
541
542 qdisc_purge_queue(cl->qdisc);
543 qdisc_class_hash_remove(&q->clhash, &cl->common);
544
545 sch_tree_unlock(sch);
546
547 qfq_destroy_class(sch, cl);
548 return 0;
549}
550
551static unsigned long qfq_search_class(struct Qdisc *sch, u32 classid)
552{
553 return (unsigned long)qfq_find_class(sch, classid);
554}
555
556static struct tcf_block *qfq_tcf_block(struct Qdisc *sch, unsigned long cl,
557 struct netlink_ext_ack *extack)
558{
559 struct qfq_sched *q = qdisc_priv(sch);
560
561 if (cl)
562 return NULL;
563
564 return q->block;
565}
566
567static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
568 u32 classid)
569{
570 struct qfq_class *cl = qfq_find_class(sch, classid);
571
572 if (cl != NULL)
573 cl->filter_cnt++;
574
575 return (unsigned long)cl;
576}
577
578static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
579{
580 struct qfq_class *cl = (struct qfq_class *)arg;
581
582 cl->filter_cnt--;
583}
584
585static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
586 struct Qdisc *new, struct Qdisc **old,
587 struct netlink_ext_ack *extack)
588{
589 struct qfq_class *cl = (struct qfq_class *)arg;
590
591 if (new == NULL) {
592 new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
593 cl->common.classid, NULL);
594 if (new == NULL)
595 new = &noop_qdisc;
596 }
597
598 *old = qdisc_replace(sch, new, &cl->qdisc);
599 return 0;
600}
601
602static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
603{
604 struct qfq_class *cl = (struct qfq_class *)arg;
605
606 return cl->qdisc;
607}
608
609static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
610 struct sk_buff *skb, struct tcmsg *tcm)
611{
612 struct qfq_class *cl = (struct qfq_class *)arg;
613 struct nlattr *nest;
614
615 tcm->tcm_parent = TC_H_ROOT;
616 tcm->tcm_handle = cl->common.classid;
617 tcm->tcm_info = cl->qdisc->handle;
618
619 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
620 if (nest == NULL)
621 goto nla_put_failure;
622 if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) ||
623 nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax))
624 goto nla_put_failure;
625 return nla_nest_end(skb, nest);
626
627nla_put_failure:
628 nla_nest_cancel(skb, nest);
629 return -EMSGSIZE;
630}
631
632static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
633 struct gnet_dump *d)
634{
635 struct qfq_class *cl = (struct qfq_class *)arg;
636 struct tc_qfq_stats xstats;
637
638 memset(&xstats, 0, sizeof(xstats));
639
640 xstats.weight = cl->agg->class_weight;
641 xstats.lmax = cl->agg->lmax;
642
643 if (gnet_stats_copy_basic(d, NULL, &cl->bstats, true) < 0 ||
644 gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
645 qdisc_qstats_copy(d, cl->qdisc) < 0)
646 return -1;
647
648 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
649}
650
651static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
652{
653 struct qfq_sched *q = qdisc_priv(sch);
654 struct qfq_class *cl;
655 unsigned int i;
656
657 if (arg->stop)
658 return;
659
660 for (i = 0; i < q->clhash.hashsize; i++) {
661 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
662 if (!tc_qdisc_stats_dump(sch, (unsigned long)cl, arg))
663 return;
664 }
665 }
666}
667
668static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
669 int *qerr)
670{
671 struct qfq_sched *q = qdisc_priv(sch);
672 struct qfq_class *cl;
673 struct tcf_result res;
674 struct tcf_proto *fl;
675 int result;
676
677 if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
678 pr_debug("qfq_classify: found %d\n", skb->priority);
679 cl = qfq_find_class(sch, skb->priority);
680 if (cl != NULL)
681 return cl;
682 }
683
684 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
685 fl = rcu_dereference_bh(q->filter_list);
686 result = tcf_classify(skb, NULL, fl, &res, false);
687 if (result >= 0) {
688#ifdef CONFIG_NET_CLS_ACT
689 switch (result) {
690 case TC_ACT_QUEUED:
691 case TC_ACT_STOLEN:
692 case TC_ACT_TRAP:
693 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
694 fallthrough;
695 case TC_ACT_SHOT:
696 return NULL;
697 }
698#endif
699 cl = (struct qfq_class *)res.class;
700 if (cl == NULL)
701 cl = qfq_find_class(sch, res.classid);
702 return cl;
703 }
704
705 return NULL;
706}
707
708/* Generic comparison function, handling wraparound. */
709static inline int qfq_gt(u64 a, u64 b)
710{
711 return (s64)(a - b) > 0;
712}
713
714/* Round a precise timestamp to its slotted value. */
715static inline u64 qfq_round_down(u64 ts, unsigned int shift)
716{
717 return ts & ~((1ULL << shift) - 1);
718}
719
720/* return the pointer to the group with lowest index in the bitmap */
721static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
722 unsigned long bitmap)
723{
724 int index = __ffs(bitmap);
725 return &q->groups[index];
726}
727/* Calculate a mask to mimic what would be ffs_from(). */
728static inline unsigned long mask_from(unsigned long bitmap, int from)
729{
730 return bitmap & ~((1UL << from) - 1);
731}
732
733/*
734 * The state computation relies on ER=0, IR=1, EB=2, IB=3
735 * First compute eligibility comparing grp->S, q->V,
736 * then check if someone is blocking us and possibly add EB
737 */
738static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
739{
740 /* if S > V we are not eligible */
741 unsigned int state = qfq_gt(grp->S, q->V);
742 unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
743 struct qfq_group *next;
744
745 if (mask) {
746 next = qfq_ffs(q, mask);
747 if (qfq_gt(grp->F, next->F))
748 state |= EB;
749 }
750
751 return state;
752}
753
754
755/*
756 * In principle
757 * q->bitmaps[dst] |= q->bitmaps[src] & mask;
758 * q->bitmaps[src] &= ~mask;
759 * but we should make sure that src != dst
760 */
761static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
762 int src, int dst)
763{
764 q->bitmaps[dst] |= q->bitmaps[src] & mask;
765 q->bitmaps[src] &= ~mask;
766}
767
768static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
769{
770 unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
771 struct qfq_group *next;
772
773 if (mask) {
774 next = qfq_ffs(q, mask);
775 if (!qfq_gt(next->F, old_F))
776 return;
777 }
778
779 mask = (1UL << index) - 1;
780 qfq_move_groups(q, mask, EB, ER);
781 qfq_move_groups(q, mask, IB, IR);
782}
783
784/*
785 * perhaps
786 *
787 old_V ^= q->V;
788 old_V >>= q->min_slot_shift;
789 if (old_V) {
790 ...
791 }
792 *
793 */
794static void qfq_make_eligible(struct qfq_sched *q)
795{
796 unsigned long vslot = q->V >> q->min_slot_shift;
797 unsigned long old_vslot = q->oldV >> q->min_slot_shift;
798
799 if (vslot != old_vslot) {
800 unsigned long mask;
801 int last_flip_pos = fls(vslot ^ old_vslot);
802
803 if (last_flip_pos > 31) /* higher than the number of groups */
804 mask = ~0UL; /* make all groups eligible */
805 else
806 mask = (1UL << last_flip_pos) - 1;
807
808 qfq_move_groups(q, mask, IR, ER);
809 qfq_move_groups(q, mask, IB, EB);
810 }
811}
812
813/*
814 * The index of the slot in which the input aggregate agg is to be
815 * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2'
816 * and not a '-1' because the start time of the group may be moved
817 * backward by one slot after the aggregate has been inserted, and
818 * this would cause non-empty slots to be right-shifted by one
819 * position.
820 *
821 * QFQ+ fully satisfies this bound to the slot index if the parameters
822 * of the classes are not changed dynamically, and if QFQ+ never
823 * happens to postpone the service of agg unjustly, i.e., it never
824 * happens that the aggregate becomes backlogged and eligible, or just
825 * eligible, while an aggregate with a higher approximated finish time
826 * is being served. In particular, in this case QFQ+ guarantees that
827 * the timestamps of agg are low enough that the slot index is never
828 * higher than 2. Unfortunately, QFQ+ cannot provide the same
829 * guarantee if it happens to unjustly postpone the service of agg, or
830 * if the parameters of some class are changed.
831 *
832 * As for the first event, i.e., an out-of-order service, the
833 * upper bound to the slot index guaranteed by QFQ+ grows to
834 * 2 +
835 * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
836 * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1.
837 *
838 * The following function deals with this problem by backward-shifting
839 * the timestamps of agg, if needed, so as to guarantee that the slot
840 * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may
841 * cause the service of other aggregates to be postponed, yet the
842 * worst-case guarantees of these aggregates are not violated. In
843 * fact, in case of no out-of-order service, the timestamps of agg
844 * would have been even lower than they are after the backward shift,
845 * because QFQ+ would have guaranteed a maximum value equal to 2 for
846 * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose
847 * service is postponed because of the backward-shift would have
848 * however waited for the service of agg before being served.
849 *
850 * The other event that may cause the slot index to be higher than 2
851 * for agg is a recent change of the parameters of some class. If the
852 * weight of a class is increased or the lmax (max_pkt_size) of the
853 * class is decreased, then a new aggregate with smaller slot size
854 * than the original parent aggregate of the class may happen to be
855 * activated. The activation of this aggregate should be properly
856 * delayed to when the service of the class has finished in the ideal
857 * system tracked by QFQ+. If the activation of the aggregate is not
858 * delayed to this reference time instant, then this aggregate may be
859 * unjustly served before other aggregates waiting for service. This
860 * may cause the above bound to the slot index to be violated for some
861 * of these unlucky aggregates.
862 *
863 * Instead of delaying the activation of the new aggregate, which is
864 * quite complex, the above-discussed capping of the slot index is
865 * used to handle also the consequences of a change of the parameters
866 * of a class.
867 */
868static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg,
869 u64 roundedS)
870{
871 u64 slot = (roundedS - grp->S) >> grp->slot_shift;
872 unsigned int i; /* slot index in the bucket list */
873
874 if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
875 u64 deltaS = roundedS - grp->S -
876 ((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
877 agg->S -= deltaS;
878 agg->F -= deltaS;
879 slot = QFQ_MAX_SLOTS - 2;
880 }
881
882 i = (grp->front + slot) % QFQ_MAX_SLOTS;
883
884 hlist_add_head(&agg->next, &grp->slots[i]);
885 __set_bit(slot, &grp->full_slots);
886}
887
888/* Maybe introduce hlist_first_entry?? */
889static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp)
890{
891 return hlist_entry(grp->slots[grp->front].first,
892 struct qfq_aggregate, next);
893}
894
895/*
896 * remove the entry from the slot
897 */
898static void qfq_front_slot_remove(struct qfq_group *grp)
899{
900 struct qfq_aggregate *agg = qfq_slot_head(grp);
901
902 BUG_ON(!agg);
903 hlist_del(&agg->next);
904 if (hlist_empty(&grp->slots[grp->front]))
905 __clear_bit(0, &grp->full_slots);
906}
907
908/*
909 * Returns the first aggregate in the first non-empty bucket of the
910 * group. As a side effect, adjusts the bucket list so the first
911 * non-empty bucket is at position 0 in full_slots.
912 */
913static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp)
914{
915 unsigned int i;
916
917 pr_debug("qfq slot_scan: grp %u full %#lx\n",
918 grp->index, grp->full_slots);
919
920 if (grp->full_slots == 0)
921 return NULL;
922
923 i = __ffs(grp->full_slots); /* zero based */
924 if (i > 0) {
925 grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
926 grp->full_slots >>= i;
927 }
928
929 return qfq_slot_head(grp);
930}
931
932/*
933 * adjust the bucket list. When the start time of a group decreases,
934 * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
935 * move the objects. The mask of occupied slots must be shifted
936 * because we use ffs() to find the first non-empty slot.
937 * This covers decreases in the group's start time, but what about
938 * increases of the start time ?
939 * Here too we should make sure that i is less than 32
940 */
941static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
942{
943 unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
944
945 grp->full_slots <<= i;
946 grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
947}
948
949static void qfq_update_eligible(struct qfq_sched *q)
950{
951 struct qfq_group *grp;
952 unsigned long ineligible;
953
954 ineligible = q->bitmaps[IR] | q->bitmaps[IB];
955 if (ineligible) {
956 if (!q->bitmaps[ER]) {
957 grp = qfq_ffs(q, ineligible);
958 if (qfq_gt(grp->S, q->V))
959 q->V = grp->S;
960 }
961 qfq_make_eligible(q);
962 }
963}
964
965/* Dequeue head packet of the head class in the DRR queue of the aggregate. */
966static void agg_dequeue(struct qfq_aggregate *agg,
967 struct qfq_class *cl, unsigned int len)
968{
969 qdisc_dequeue_peeked(cl->qdisc);
970
971 cl->deficit -= (int) len;
972
973 if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */
974 list_del(&cl->alist);
975 else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) {
976 cl->deficit += agg->lmax;
977 list_move_tail(&cl->alist, &agg->active);
978 }
979}
980
981static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg,
982 struct qfq_class **cl,
983 unsigned int *len)
984{
985 struct sk_buff *skb;
986
987 *cl = list_first_entry(&agg->active, struct qfq_class, alist);
988 skb = (*cl)->qdisc->ops->peek((*cl)->qdisc);
989 if (skb == NULL)
990 WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
991 else
992 *len = qdisc_pkt_len(skb);
993
994 return skb;
995}
996
997/* Update F according to the actual service received by the aggregate. */
998static inline void charge_actual_service(struct qfq_aggregate *agg)
999{
1000 /* Compute the service received by the aggregate, taking into
1001 * account that, after decreasing the number of classes in
1002 * agg, it may happen that
1003 * agg->initial_budget - agg->budget > agg->bugdetmax
1004 */
1005 u32 service_received = min(agg->budgetmax,
1006 agg->initial_budget - agg->budget);
1007
1008 agg->F = agg->S + (u64)service_received * agg->inv_w;
1009}
1010
1011/* Assign a reasonable start time for a new aggregate in group i.
1012 * Admissible values for \hat(F) are multiples of \sigma_i
1013 * no greater than V+\sigma_i . Larger values mean that
1014 * we had a wraparound so we consider the timestamp to be stale.
1015 *
1016 * If F is not stale and F >= V then we set S = F.
1017 * Otherwise we should assign S = V, but this may violate
1018 * the ordering in EB (see [2]). So, if we have groups in ER,
1019 * set S to the F_j of the first group j which would be blocking us.
1020 * We are guaranteed not to move S backward because
1021 * otherwise our group i would still be blocked.
1022 */
1023static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg)
1024{
1025 unsigned long mask;
1026 u64 limit, roundedF;
1027 int slot_shift = agg->grp->slot_shift;
1028
1029 roundedF = qfq_round_down(agg->F, slot_shift);
1030 limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
1031
1032 if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) {
1033 /* timestamp was stale */
1034 mask = mask_from(q->bitmaps[ER], agg->grp->index);
1035 if (mask) {
1036 struct qfq_group *next = qfq_ffs(q, mask);
1037 if (qfq_gt(roundedF, next->F)) {
1038 if (qfq_gt(limit, next->F))
1039 agg->S = next->F;
1040 else /* preserve timestamp correctness */
1041 agg->S = limit;
1042 return;
1043 }
1044 }
1045 agg->S = q->V;
1046 } else /* timestamp is not stale */
1047 agg->S = agg->F;
1048}
1049
1050/* Update the timestamps of agg before scheduling/rescheduling it for
1051 * service. In particular, assign to agg->F its maximum possible
1052 * value, i.e., the virtual finish time with which the aggregate
1053 * should be labeled if it used all its budget once in service.
1054 */
1055static inline void
1056qfq_update_agg_ts(struct qfq_sched *q,
1057 struct qfq_aggregate *agg, enum update_reason reason)
1058{
1059 if (reason != requeue)
1060 qfq_update_start(q, agg);
1061 else /* just charge agg for the service received */
1062 agg->S = agg->F;
1063
1064 agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w;
1065}
1066
1067static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg);
1068
1069static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
1070{
1071 struct qfq_sched *q = qdisc_priv(sch);
1072 struct qfq_aggregate *in_serv_agg = q->in_serv_agg;
1073 struct qfq_class *cl;
1074 struct sk_buff *skb = NULL;
1075 /* next-packet len, 0 means no more active classes in in-service agg */
1076 unsigned int len = 0;
1077
1078 if (in_serv_agg == NULL)
1079 return NULL;
1080
1081 if (!list_empty(&in_serv_agg->active))
1082 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1083
1084 /*
1085 * If there are no active classes in the in-service aggregate,
1086 * or if the aggregate has not enough budget to serve its next
1087 * class, then choose the next aggregate to serve.
1088 */
1089 if (len == 0 || in_serv_agg->budget < len) {
1090 charge_actual_service(in_serv_agg);
1091
1092 /* recharge the budget of the aggregate */
1093 in_serv_agg->initial_budget = in_serv_agg->budget =
1094 in_serv_agg->budgetmax;
1095
1096 if (!list_empty(&in_serv_agg->active)) {
1097 /*
1098 * Still active: reschedule for
1099 * service. Possible optimization: if no other
1100 * aggregate is active, then there is no point
1101 * in rescheduling this aggregate, and we can
1102 * just keep it as the in-service one. This
1103 * should be however a corner case, and to
1104 * handle it, we would need to maintain an
1105 * extra num_active_aggs field.
1106 */
1107 qfq_update_agg_ts(q, in_serv_agg, requeue);
1108 qfq_schedule_agg(q, in_serv_agg);
1109 } else if (sch->q.qlen == 0) { /* no aggregate to serve */
1110 q->in_serv_agg = NULL;
1111 return NULL;
1112 }
1113
1114 /*
1115 * If we get here, there are other aggregates queued:
1116 * choose the new aggregate to serve.
1117 */
1118 in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
1119 skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1120 }
1121 if (!skb)
1122 return NULL;
1123
1124 qdisc_qstats_backlog_dec(sch, skb);
1125 sch->q.qlen--;
1126 qdisc_bstats_update(sch, skb);
1127
1128 agg_dequeue(in_serv_agg, cl, len);
1129 /* If lmax is lowered, through qfq_change_class, for a class
1130 * owning pending packets with larger size than the new value
1131 * of lmax, then the following condition may hold.
1132 */
1133 if (unlikely(in_serv_agg->budget < len))
1134 in_serv_agg->budget = 0;
1135 else
1136 in_serv_agg->budget -= len;
1137
1138 q->V += (u64)len * q->iwsum;
1139 pr_debug("qfq dequeue: len %u F %lld now %lld\n",
1140 len, (unsigned long long) in_serv_agg->F,
1141 (unsigned long long) q->V);
1142
1143 return skb;
1144}
1145
1146static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)
1147{
1148 struct qfq_group *grp;
1149 struct qfq_aggregate *agg, *new_front_agg;
1150 u64 old_F;
1151
1152 qfq_update_eligible(q);
1153 q->oldV = q->V;
1154
1155 if (!q->bitmaps[ER])
1156 return NULL;
1157
1158 grp = qfq_ffs(q, q->bitmaps[ER]);
1159 old_F = grp->F;
1160
1161 agg = qfq_slot_head(grp);
1162
1163 /* agg starts to be served, remove it from schedule */
1164 qfq_front_slot_remove(grp);
1165
1166 new_front_agg = qfq_slot_scan(grp);
1167
1168 if (new_front_agg == NULL) /* group is now inactive, remove from ER */
1169 __clear_bit(grp->index, &q->bitmaps[ER]);
1170 else {
1171 u64 roundedS = qfq_round_down(new_front_agg->S,
1172 grp->slot_shift);
1173 unsigned int s;
1174
1175 if (grp->S == roundedS)
1176 return agg;
1177 grp->S = roundedS;
1178 grp->F = roundedS + (2ULL << grp->slot_shift);
1179 __clear_bit(grp->index, &q->bitmaps[ER]);
1180 s = qfq_calc_state(q, grp);
1181 __set_bit(grp->index, &q->bitmaps[s]);
1182 }
1183
1184 qfq_unblock_groups(q, grp->index, old_F);
1185
1186 return agg;
1187}
1188
1189static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
1190 struct sk_buff **to_free)
1191{
1192 unsigned int len = qdisc_pkt_len(skb), gso_segs;
1193 struct qfq_sched *q = qdisc_priv(sch);
1194 struct qfq_class *cl;
1195 struct qfq_aggregate *agg;
1196 int err = 0;
1197 bool first;
1198
1199 cl = qfq_classify(skb, sch, &err);
1200 if (cl == NULL) {
1201 if (err & __NET_XMIT_BYPASS)
1202 qdisc_qstats_drop(sch);
1203 __qdisc_drop(skb, to_free);
1204 return err;
1205 }
1206 pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
1207
1208 if (unlikely(cl->agg->lmax < len)) {
1209 pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
1210 cl->agg->lmax, len, cl->common.classid);
1211 err = qfq_change_agg(sch, cl, cl->agg->class_weight, len);
1212 if (err) {
1213 cl->qstats.drops++;
1214 return qdisc_drop(skb, sch, to_free);
1215 }
1216 }
1217
1218 gso_segs = skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 1;
1219 first = !cl->qdisc->q.qlen;
1220 err = qdisc_enqueue(skb, cl->qdisc, to_free);
1221 if (unlikely(err != NET_XMIT_SUCCESS)) {
1222 pr_debug("qfq_enqueue: enqueue failed %d\n", err);
1223 if (net_xmit_drop_count(err)) {
1224 cl->qstats.drops++;
1225 qdisc_qstats_drop(sch);
1226 }
1227 return err;
1228 }
1229
1230 _bstats_update(&cl->bstats, len, gso_segs);
1231 sch->qstats.backlog += len;
1232 ++sch->q.qlen;
1233
1234 agg = cl->agg;
1235 /* if the queue was not empty, then done here */
1236 if (!first) {
1237 if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) &&
1238 list_first_entry(&agg->active, struct qfq_class, alist)
1239 == cl && cl->deficit < len)
1240 list_move_tail(&cl->alist, &agg->active);
1241
1242 return err;
1243 }
1244
1245 /* schedule class for service within the aggregate */
1246 cl->deficit = agg->lmax;
1247 list_add_tail(&cl->alist, &agg->active);
1248
1249 if (list_first_entry(&agg->active, struct qfq_class, alist) != cl ||
1250 q->in_serv_agg == agg)
1251 return err; /* non-empty or in service, nothing else to do */
1252
1253 qfq_activate_agg(q, agg, enqueue);
1254
1255 return err;
1256}
1257
1258/*
1259 * Schedule aggregate according to its timestamps.
1260 */
1261static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1262{
1263 struct qfq_group *grp = agg->grp;
1264 u64 roundedS;
1265 int s;
1266
1267 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1268
1269 /*
1270 * Insert agg in the correct bucket.
1271 * If agg->S >= grp->S we don't need to adjust the
1272 * bucket list and simply go to the insertion phase.
1273 * Otherwise grp->S is decreasing, we must make room
1274 * in the bucket list, and also recompute the group state.
1275 * Finally, if there were no flows in this group and nobody
1276 * was in ER make sure to adjust V.
1277 */
1278 if (grp->full_slots) {
1279 if (!qfq_gt(grp->S, agg->S))
1280 goto skip_update;
1281
1282 /* create a slot for this agg->S */
1283 qfq_slot_rotate(grp, roundedS);
1284 /* group was surely ineligible, remove */
1285 __clear_bit(grp->index, &q->bitmaps[IR]);
1286 __clear_bit(grp->index, &q->bitmaps[IB]);
1287 } else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) &&
1288 q->in_serv_agg == NULL)
1289 q->V = roundedS;
1290
1291 grp->S = roundedS;
1292 grp->F = roundedS + (2ULL << grp->slot_shift);
1293 s = qfq_calc_state(q, grp);
1294 __set_bit(grp->index, &q->bitmaps[s]);
1295
1296 pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
1297 s, q->bitmaps[s],
1298 (unsigned long long) agg->S,
1299 (unsigned long long) agg->F,
1300 (unsigned long long) q->V);
1301
1302skip_update:
1303 qfq_slot_insert(grp, agg, roundedS);
1304}
1305
1306
1307/* Update agg ts and schedule agg for service */
1308static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
1309 enum update_reason reason)
1310{
1311 agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */
1312
1313 qfq_update_agg_ts(q, agg, reason);
1314 if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */
1315 q->in_serv_agg = agg; /* start serving this aggregate */
1316 /* update V: to be in service, agg must be eligible */
1317 q->oldV = q->V = agg->S;
1318 } else if (agg != q->in_serv_agg)
1319 qfq_schedule_agg(q, agg);
1320}
1321
1322static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
1323 struct qfq_aggregate *agg)
1324{
1325 unsigned int i, offset;
1326 u64 roundedS;
1327
1328 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1329 offset = (roundedS - grp->S) >> grp->slot_shift;
1330
1331 i = (grp->front + offset) % QFQ_MAX_SLOTS;
1332
1333 hlist_del(&agg->next);
1334 if (hlist_empty(&grp->slots[i]))
1335 __clear_bit(offset, &grp->full_slots);
1336}
1337
1338/*
1339 * Called to forcibly deschedule an aggregate. If the aggregate is
1340 * not in the front bucket, or if the latter has other aggregates in
1341 * the front bucket, we can simply remove the aggregate with no other
1342 * side effects.
1343 * Otherwise we must propagate the event up.
1344 */
1345static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1346{
1347 struct qfq_group *grp = agg->grp;
1348 unsigned long mask;
1349 u64 roundedS;
1350 int s;
1351
1352 if (agg == q->in_serv_agg) {
1353 charge_actual_service(agg);
1354 q->in_serv_agg = qfq_choose_next_agg(q);
1355 return;
1356 }
1357
1358 agg->F = agg->S;
1359 qfq_slot_remove(q, grp, agg);
1360
1361 if (!grp->full_slots) {
1362 __clear_bit(grp->index, &q->bitmaps[IR]);
1363 __clear_bit(grp->index, &q->bitmaps[EB]);
1364 __clear_bit(grp->index, &q->bitmaps[IB]);
1365
1366 if (test_bit(grp->index, &q->bitmaps[ER]) &&
1367 !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
1368 mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
1369 if (mask)
1370 mask = ~((1UL << __fls(mask)) - 1);
1371 else
1372 mask = ~0UL;
1373 qfq_move_groups(q, mask, EB, ER);
1374 qfq_move_groups(q, mask, IB, IR);
1375 }
1376 __clear_bit(grp->index, &q->bitmaps[ER]);
1377 } else if (hlist_empty(&grp->slots[grp->front])) {
1378 agg = qfq_slot_scan(grp);
1379 roundedS = qfq_round_down(agg->S, grp->slot_shift);
1380 if (grp->S != roundedS) {
1381 __clear_bit(grp->index, &q->bitmaps[ER]);
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 grp->S = roundedS;
1386 grp->F = roundedS + (2ULL << grp->slot_shift);
1387 s = qfq_calc_state(q, grp);
1388 __set_bit(grp->index, &q->bitmaps[s]);
1389 }
1390 }
1391}
1392
1393static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1394{
1395 struct qfq_sched *q = qdisc_priv(sch);
1396 struct qfq_class *cl = (struct qfq_class *)arg;
1397
1398 qfq_deactivate_class(q, cl);
1399}
1400
1401static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt,
1402 struct netlink_ext_ack *extack)
1403{
1404 struct qfq_sched *q = qdisc_priv(sch);
1405 struct qfq_group *grp;
1406 int i, j, err;
1407 u32 max_cl_shift, maxbudg_shift, max_classes;
1408
1409 err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
1410 if (err)
1411 return err;
1412
1413 err = qdisc_class_hash_init(&q->clhash);
1414 if (err < 0)
1415 return err;
1416
1417 max_classes = min_t(u64, (u64)qdisc_dev(sch)->tx_queue_len + 1,
1418 QFQ_MAX_AGG_CLASSES);
1419 /* max_cl_shift = floor(log_2(max_classes)) */
1420 max_cl_shift = __fls(max_classes);
1421 q->max_agg_classes = 1<<max_cl_shift;
1422
1423 /* maxbudg_shift = log2(max_len * max_classes_per_agg) */
1424 maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift;
1425 q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX;
1426
1427 for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1428 grp = &q->groups[i];
1429 grp->index = i;
1430 grp->slot_shift = q->min_slot_shift + i;
1431 for (j = 0; j < QFQ_MAX_SLOTS; j++)
1432 INIT_HLIST_HEAD(&grp->slots[j]);
1433 }
1434
1435 INIT_HLIST_HEAD(&q->nonfull_aggs);
1436
1437 return 0;
1438}
1439
1440static void qfq_reset_qdisc(struct Qdisc *sch)
1441{
1442 struct qfq_sched *q = qdisc_priv(sch);
1443 struct qfq_class *cl;
1444 unsigned int i;
1445
1446 for (i = 0; i < q->clhash.hashsize; i++) {
1447 hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1448 if (cl->qdisc->q.qlen > 0)
1449 qfq_deactivate_class(q, cl);
1450
1451 qdisc_reset(cl->qdisc);
1452 }
1453 }
1454}
1455
1456static void qfq_destroy_qdisc(struct Qdisc *sch)
1457{
1458 struct qfq_sched *q = qdisc_priv(sch);
1459 struct qfq_class *cl;
1460 struct hlist_node *next;
1461 unsigned int i;
1462
1463 tcf_block_put(q->block);
1464
1465 for (i = 0; i < q->clhash.hashsize; i++) {
1466 hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1467 common.hnode) {
1468 qfq_destroy_class(sch, cl);
1469 }
1470 }
1471 qdisc_class_hash_destroy(&q->clhash);
1472}
1473
1474static const struct Qdisc_class_ops qfq_class_ops = {
1475 .change = qfq_change_class,
1476 .delete = qfq_delete_class,
1477 .find = qfq_search_class,
1478 .tcf_block = qfq_tcf_block,
1479 .bind_tcf = qfq_bind_tcf,
1480 .unbind_tcf = qfq_unbind_tcf,
1481 .graft = qfq_graft_class,
1482 .leaf = qfq_class_leaf,
1483 .qlen_notify = qfq_qlen_notify,
1484 .dump = qfq_dump_class,
1485 .dump_stats = qfq_dump_class_stats,
1486 .walk = qfq_walk,
1487};
1488
1489static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1490 .cl_ops = &qfq_class_ops,
1491 .id = "qfq",
1492 .priv_size = sizeof(struct qfq_sched),
1493 .enqueue = qfq_enqueue,
1494 .dequeue = qfq_dequeue,
1495 .peek = qdisc_peek_dequeued,
1496 .init = qfq_init_qdisc,
1497 .reset = qfq_reset_qdisc,
1498 .destroy = qfq_destroy_qdisc,
1499 .owner = THIS_MODULE,
1500};
1501
1502static int __init qfq_init(void)
1503{
1504 return register_qdisc(&qfq_qdisc_ops);
1505}
1506
1507static void __exit qfq_exit(void)
1508{
1509 unregister_qdisc(&qfq_qdisc_ops);
1510}
1511
1512module_init(qfq_init);
1513module_exit(qfq_exit);
1514MODULE_LICENSE("GPL");