xref: /openbmc/linux/net/sched/sch_qfq.c (revision 31e67366)
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  */
123 enum qfq_state { ER, IR, EB, IB, QFQ_MAX_STATE };
124 
125 struct qfq_group;
126 
127 struct qfq_aggregate;
128 
129 struct qfq_class {
130 	struct Qdisc_class_common common;
131 
132 	unsigned int filter_cnt;
133 
134 	struct gnet_stats_basic_packed 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 
143 struct 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 
168 struct 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 
179 struct 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  */
204 enum update_reason {enqueue, requeue};
205 
206 static 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 
217 static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = {
218 	[TCA_QFQ_WEIGHT] = { .type = NLA_U32 },
219 	[TCA_QFQ_LMAX] = { .type = NLA_U32 },
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  */
227 static 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;
242 out:
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 
249 static void qfq_deactivate_agg(struct qfq_sched *, struct qfq_aggregate *);
250 static void qfq_activate_agg(struct qfq_sched *, struct qfq_aggregate *,
251 			     enum update_reason);
252 
253 static 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 
263 static 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. */
277 static 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. */
311 static 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 
326 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *);
327 
328 static 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. */
341 static 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. */
352 static 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. */
365 static 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 */
374 static 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 
392 static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
393 			    struct nlattr **tca, unsigned long *arg,
394 			    struct netlink_ext_ack *extack)
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 						    qdisc_root_sleeping_running(sch),
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 	cl->common.classid = classid;
469 	cl->deficit = lmax;
470 
471 	cl->qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
472 				      classid, NULL);
473 	if (cl->qdisc == NULL)
474 		cl->qdisc = &noop_qdisc;
475 
476 	if (tca[TCA_RATE]) {
477 		err = gen_new_estimator(&cl->bstats, NULL,
478 					&cl->rate_est,
479 					NULL,
480 					qdisc_root_sleeping_running(sch),
481 					tca[TCA_RATE]);
482 		if (err)
483 			goto destroy_class;
484 	}
485 
486 	if (cl->qdisc != &noop_qdisc)
487 		qdisc_hash_add(cl->qdisc, true);
488 	sch_tree_lock(sch);
489 	qdisc_class_hash_insert(&q->clhash, &cl->common);
490 	sch_tree_unlock(sch);
491 
492 	qdisc_class_hash_grow(sch, &q->clhash);
493 
494 set_change_agg:
495 	sch_tree_lock(sch);
496 	new_agg = qfq_find_agg(q, lmax, weight);
497 	if (new_agg == NULL) { /* create new aggregate */
498 		sch_tree_unlock(sch);
499 		new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL);
500 		if (new_agg == NULL) {
501 			err = -ENOBUFS;
502 			gen_kill_estimator(&cl->rate_est);
503 			goto destroy_class;
504 		}
505 		sch_tree_lock(sch);
506 		qfq_init_agg(q, new_agg, lmax, weight);
507 	}
508 	if (existing)
509 		qfq_deact_rm_from_agg(q, cl);
510 	qfq_add_to_agg(q, new_agg, cl);
511 	sch_tree_unlock(sch);
512 
513 	*arg = (unsigned long)cl;
514 	return 0;
515 
516 destroy_class:
517 	qdisc_put(cl->qdisc);
518 	kfree(cl);
519 	return err;
520 }
521 
522 static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
523 {
524 	struct qfq_sched *q = qdisc_priv(sch);
525 
526 	qfq_rm_from_agg(q, cl);
527 	gen_kill_estimator(&cl->rate_est);
528 	qdisc_put(cl->qdisc);
529 	kfree(cl);
530 }
531 
532 static int qfq_delete_class(struct Qdisc *sch, unsigned long arg,
533 			    struct netlink_ext_ack *extack)
534 {
535 	struct qfq_sched *q = qdisc_priv(sch);
536 	struct qfq_class *cl = (struct qfq_class *)arg;
537 
538 	if (cl->filter_cnt > 0)
539 		return -EBUSY;
540 
541 	sch_tree_lock(sch);
542 
543 	qdisc_purge_queue(cl->qdisc);
544 	qdisc_class_hash_remove(&q->clhash, &cl->common);
545 
546 	sch_tree_unlock(sch);
547 
548 	qfq_destroy_class(sch, cl);
549 	return 0;
550 }
551 
552 static unsigned long qfq_search_class(struct Qdisc *sch, u32 classid)
553 {
554 	return (unsigned long)qfq_find_class(sch, classid);
555 }
556 
557 static struct tcf_block *qfq_tcf_block(struct Qdisc *sch, unsigned long cl,
558 				       struct netlink_ext_ack *extack)
559 {
560 	struct qfq_sched *q = qdisc_priv(sch);
561 
562 	if (cl)
563 		return NULL;
564 
565 	return q->block;
566 }
567 
568 static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
569 				  u32 classid)
570 {
571 	struct qfq_class *cl = qfq_find_class(sch, classid);
572 
573 	if (cl != NULL)
574 		cl->filter_cnt++;
575 
576 	return (unsigned long)cl;
577 }
578 
579 static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
580 {
581 	struct qfq_class *cl = (struct qfq_class *)arg;
582 
583 	cl->filter_cnt--;
584 }
585 
586 static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
587 			   struct Qdisc *new, struct Qdisc **old,
588 			   struct netlink_ext_ack *extack)
589 {
590 	struct qfq_class *cl = (struct qfq_class *)arg;
591 
592 	if (new == NULL) {
593 		new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
594 					cl->common.classid, NULL);
595 		if (new == NULL)
596 			new = &noop_qdisc;
597 	}
598 
599 	*old = qdisc_replace(sch, new, &cl->qdisc);
600 	return 0;
601 }
602 
603 static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
604 {
605 	struct qfq_class *cl = (struct qfq_class *)arg;
606 
607 	return cl->qdisc;
608 }
609 
610 static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
611 			  struct sk_buff *skb, struct tcmsg *tcm)
612 {
613 	struct qfq_class *cl = (struct qfq_class *)arg;
614 	struct nlattr *nest;
615 
616 	tcm->tcm_parent	= TC_H_ROOT;
617 	tcm->tcm_handle	= cl->common.classid;
618 	tcm->tcm_info	= cl->qdisc->handle;
619 
620 	nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
621 	if (nest == NULL)
622 		goto nla_put_failure;
623 	if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) ||
624 	    nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax))
625 		goto nla_put_failure;
626 	return nla_nest_end(skb, nest);
627 
628 nla_put_failure:
629 	nla_nest_cancel(skb, nest);
630 	return -EMSGSIZE;
631 }
632 
633 static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
634 				struct gnet_dump *d)
635 {
636 	struct qfq_class *cl = (struct qfq_class *)arg;
637 	struct tc_qfq_stats xstats;
638 
639 	memset(&xstats, 0, sizeof(xstats));
640 
641 	xstats.weight = cl->agg->class_weight;
642 	xstats.lmax = cl->agg->lmax;
643 
644 	if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
645 				  d, NULL, &cl->bstats) < 0 ||
646 	    gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
647 	    qdisc_qstats_copy(d, cl->qdisc) < 0)
648 		return -1;
649 
650 	return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
651 }
652 
653 static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
654 {
655 	struct qfq_sched *q = qdisc_priv(sch);
656 	struct qfq_class *cl;
657 	unsigned int i;
658 
659 	if (arg->stop)
660 		return;
661 
662 	for (i = 0; i < q->clhash.hashsize; i++) {
663 		hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
664 			if (arg->count < arg->skip) {
665 				arg->count++;
666 				continue;
667 			}
668 			if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
669 				arg->stop = 1;
670 				return;
671 			}
672 			arg->count++;
673 		}
674 	}
675 }
676 
677 static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
678 				      int *qerr)
679 {
680 	struct qfq_sched *q = qdisc_priv(sch);
681 	struct qfq_class *cl;
682 	struct tcf_result res;
683 	struct tcf_proto *fl;
684 	int result;
685 
686 	if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
687 		pr_debug("qfq_classify: found %d\n", skb->priority);
688 		cl = qfq_find_class(sch, skb->priority);
689 		if (cl != NULL)
690 			return cl;
691 	}
692 
693 	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
694 	fl = rcu_dereference_bh(q->filter_list);
695 	result = tcf_classify(skb, fl, &res, false);
696 	if (result >= 0) {
697 #ifdef CONFIG_NET_CLS_ACT
698 		switch (result) {
699 		case TC_ACT_QUEUED:
700 		case TC_ACT_STOLEN:
701 		case TC_ACT_TRAP:
702 			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
703 			fallthrough;
704 		case TC_ACT_SHOT:
705 			return NULL;
706 		}
707 #endif
708 		cl = (struct qfq_class *)res.class;
709 		if (cl == NULL)
710 			cl = qfq_find_class(sch, res.classid);
711 		return cl;
712 	}
713 
714 	return NULL;
715 }
716 
717 /* Generic comparison function, handling wraparound. */
718 static inline int qfq_gt(u64 a, u64 b)
719 {
720 	return (s64)(a - b) > 0;
721 }
722 
723 /* Round a precise timestamp to its slotted value. */
724 static inline u64 qfq_round_down(u64 ts, unsigned int shift)
725 {
726 	return ts & ~((1ULL << shift) - 1);
727 }
728 
729 /* return the pointer to the group with lowest index in the bitmap */
730 static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
731 					unsigned long bitmap)
732 {
733 	int index = __ffs(bitmap);
734 	return &q->groups[index];
735 }
736 /* Calculate a mask to mimic what would be ffs_from(). */
737 static inline unsigned long mask_from(unsigned long bitmap, int from)
738 {
739 	return bitmap & ~((1UL << from) - 1);
740 }
741 
742 /*
743  * The state computation relies on ER=0, IR=1, EB=2, IB=3
744  * First compute eligibility comparing grp->S, q->V,
745  * then check if someone is blocking us and possibly add EB
746  */
747 static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
748 {
749 	/* if S > V we are not eligible */
750 	unsigned int state = qfq_gt(grp->S, q->V);
751 	unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
752 	struct qfq_group *next;
753 
754 	if (mask) {
755 		next = qfq_ffs(q, mask);
756 		if (qfq_gt(grp->F, next->F))
757 			state |= EB;
758 	}
759 
760 	return state;
761 }
762 
763 
764 /*
765  * In principle
766  *	q->bitmaps[dst] |= q->bitmaps[src] & mask;
767  *	q->bitmaps[src] &= ~mask;
768  * but we should make sure that src != dst
769  */
770 static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
771 				   int src, int dst)
772 {
773 	q->bitmaps[dst] |= q->bitmaps[src] & mask;
774 	q->bitmaps[src] &= ~mask;
775 }
776 
777 static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
778 {
779 	unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
780 	struct qfq_group *next;
781 
782 	if (mask) {
783 		next = qfq_ffs(q, mask);
784 		if (!qfq_gt(next->F, old_F))
785 			return;
786 	}
787 
788 	mask = (1UL << index) - 1;
789 	qfq_move_groups(q, mask, EB, ER);
790 	qfq_move_groups(q, mask, IB, IR);
791 }
792 
793 /*
794  * perhaps
795  *
796 	old_V ^= q->V;
797 	old_V >>= q->min_slot_shift;
798 	if (old_V) {
799 		...
800 	}
801  *
802  */
803 static void qfq_make_eligible(struct qfq_sched *q)
804 {
805 	unsigned long vslot = q->V >> q->min_slot_shift;
806 	unsigned long old_vslot = q->oldV >> q->min_slot_shift;
807 
808 	if (vslot != old_vslot) {
809 		unsigned long mask;
810 		int last_flip_pos = fls(vslot ^ old_vslot);
811 
812 		if (last_flip_pos > 31) /* higher than the number of groups */
813 			mask = ~0UL;    /* make all groups eligible */
814 		else
815 			mask = (1UL << last_flip_pos) - 1;
816 
817 		qfq_move_groups(q, mask, IR, ER);
818 		qfq_move_groups(q, mask, IB, EB);
819 	}
820 }
821 
822 /*
823  * The index of the slot in which the input aggregate agg is to be
824  * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2'
825  * and not a '-1' because the start time of the group may be moved
826  * backward by one slot after the aggregate has been inserted, and
827  * this would cause non-empty slots to be right-shifted by one
828  * position.
829  *
830  * QFQ+ fully satisfies this bound to the slot index if the parameters
831  * of the classes are not changed dynamically, and if QFQ+ never
832  * happens to postpone the service of agg unjustly, i.e., it never
833  * happens that the aggregate becomes backlogged and eligible, or just
834  * eligible, while an aggregate with a higher approximated finish time
835  * is being served. In particular, in this case QFQ+ guarantees that
836  * the timestamps of agg are low enough that the slot index is never
837  * higher than 2. Unfortunately, QFQ+ cannot provide the same
838  * guarantee if it happens to unjustly postpone the service of agg, or
839  * if the parameters of some class are changed.
840  *
841  * As for the first event, i.e., an out-of-order service, the
842  * upper bound to the slot index guaranteed by QFQ+ grows to
843  * 2 +
844  * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
845  * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1.
846  *
847  * The following function deals with this problem by backward-shifting
848  * the timestamps of agg, if needed, so as to guarantee that the slot
849  * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may
850  * cause the service of other aggregates to be postponed, yet the
851  * worst-case guarantees of these aggregates are not violated.  In
852  * fact, in case of no out-of-order service, the timestamps of agg
853  * would have been even lower than they are after the backward shift,
854  * because QFQ+ would have guaranteed a maximum value equal to 2 for
855  * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose
856  * service is postponed because of the backward-shift would have
857  * however waited for the service of agg before being served.
858  *
859  * The other event that may cause the slot index to be higher than 2
860  * for agg is a recent change of the parameters of some class. If the
861  * weight of a class is increased or the lmax (max_pkt_size) of the
862  * class is decreased, then a new aggregate with smaller slot size
863  * than the original parent aggregate of the class may happen to be
864  * activated. The activation of this aggregate should be properly
865  * delayed to when the service of the class has finished in the ideal
866  * system tracked by QFQ+. If the activation of the aggregate is not
867  * delayed to this reference time instant, then this aggregate may be
868  * unjustly served before other aggregates waiting for service. This
869  * may cause the above bound to the slot index to be violated for some
870  * of these unlucky aggregates.
871  *
872  * Instead of delaying the activation of the new aggregate, which is
873  * quite complex, the above-discussed capping of the slot index is
874  * used to handle also the consequences of a change of the parameters
875  * of a class.
876  */
877 static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg,
878 			    u64 roundedS)
879 {
880 	u64 slot = (roundedS - grp->S) >> grp->slot_shift;
881 	unsigned int i; /* slot index in the bucket list */
882 
883 	if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
884 		u64 deltaS = roundedS - grp->S -
885 			((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
886 		agg->S -= deltaS;
887 		agg->F -= deltaS;
888 		slot = QFQ_MAX_SLOTS - 2;
889 	}
890 
891 	i = (grp->front + slot) % QFQ_MAX_SLOTS;
892 
893 	hlist_add_head(&agg->next, &grp->slots[i]);
894 	__set_bit(slot, &grp->full_slots);
895 }
896 
897 /* Maybe introduce hlist_first_entry?? */
898 static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp)
899 {
900 	return hlist_entry(grp->slots[grp->front].first,
901 			   struct qfq_aggregate, next);
902 }
903 
904 /*
905  * remove the entry from the slot
906  */
907 static void qfq_front_slot_remove(struct qfq_group *grp)
908 {
909 	struct qfq_aggregate *agg = qfq_slot_head(grp);
910 
911 	BUG_ON(!agg);
912 	hlist_del(&agg->next);
913 	if (hlist_empty(&grp->slots[grp->front]))
914 		__clear_bit(0, &grp->full_slots);
915 }
916 
917 /*
918  * Returns the first aggregate in the first non-empty bucket of the
919  * group. As a side effect, adjusts the bucket list so the first
920  * non-empty bucket is at position 0 in full_slots.
921  */
922 static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp)
923 {
924 	unsigned int i;
925 
926 	pr_debug("qfq slot_scan: grp %u full %#lx\n",
927 		 grp->index, grp->full_slots);
928 
929 	if (grp->full_slots == 0)
930 		return NULL;
931 
932 	i = __ffs(grp->full_slots);  /* zero based */
933 	if (i > 0) {
934 		grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
935 		grp->full_slots >>= i;
936 	}
937 
938 	return qfq_slot_head(grp);
939 }
940 
941 /*
942  * adjust the bucket list. When the start time of a group decreases,
943  * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
944  * move the objects. The mask of occupied slots must be shifted
945  * because we use ffs() to find the first non-empty slot.
946  * This covers decreases in the group's start time, but what about
947  * increases of the start time ?
948  * Here too we should make sure that i is less than 32
949  */
950 static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
951 {
952 	unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
953 
954 	grp->full_slots <<= i;
955 	grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
956 }
957 
958 static void qfq_update_eligible(struct qfq_sched *q)
959 {
960 	struct qfq_group *grp;
961 	unsigned long ineligible;
962 
963 	ineligible = q->bitmaps[IR] | q->bitmaps[IB];
964 	if (ineligible) {
965 		if (!q->bitmaps[ER]) {
966 			grp = qfq_ffs(q, ineligible);
967 			if (qfq_gt(grp->S, q->V))
968 				q->V = grp->S;
969 		}
970 		qfq_make_eligible(q);
971 	}
972 }
973 
974 /* Dequeue head packet of the head class in the DRR queue of the aggregate. */
975 static void agg_dequeue(struct qfq_aggregate *agg,
976 			struct qfq_class *cl, unsigned int len)
977 {
978 	qdisc_dequeue_peeked(cl->qdisc);
979 
980 	cl->deficit -= (int) len;
981 
982 	if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */
983 		list_del(&cl->alist);
984 	else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) {
985 		cl->deficit += agg->lmax;
986 		list_move_tail(&cl->alist, &agg->active);
987 	}
988 }
989 
990 static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg,
991 					   struct qfq_class **cl,
992 					   unsigned int *len)
993 {
994 	struct sk_buff *skb;
995 
996 	*cl = list_first_entry(&agg->active, struct qfq_class, alist);
997 	skb = (*cl)->qdisc->ops->peek((*cl)->qdisc);
998 	if (skb == NULL)
999 		WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
1000 	else
1001 		*len = qdisc_pkt_len(skb);
1002 
1003 	return skb;
1004 }
1005 
1006 /* Update F according to the actual service received by the aggregate. */
1007 static inline void charge_actual_service(struct qfq_aggregate *agg)
1008 {
1009 	/* Compute the service received by the aggregate, taking into
1010 	 * account that, after decreasing the number of classes in
1011 	 * agg, it may happen that
1012 	 * agg->initial_budget - agg->budget > agg->bugdetmax
1013 	 */
1014 	u32 service_received = min(agg->budgetmax,
1015 				   agg->initial_budget - agg->budget);
1016 
1017 	agg->F = agg->S + (u64)service_received * agg->inv_w;
1018 }
1019 
1020 /* Assign a reasonable start time for a new aggregate in group i.
1021  * Admissible values for \hat(F) are multiples of \sigma_i
1022  * no greater than V+\sigma_i . Larger values mean that
1023  * we had a wraparound so we consider the timestamp to be stale.
1024  *
1025  * If F is not stale and F >= V then we set S = F.
1026  * Otherwise we should assign S = V, but this may violate
1027  * the ordering in EB (see [2]). So, if we have groups in ER,
1028  * set S to the F_j of the first group j which would be blocking us.
1029  * We are guaranteed not to move S backward because
1030  * otherwise our group i would still be blocked.
1031  */
1032 static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg)
1033 {
1034 	unsigned long mask;
1035 	u64 limit, roundedF;
1036 	int slot_shift = agg->grp->slot_shift;
1037 
1038 	roundedF = qfq_round_down(agg->F, slot_shift);
1039 	limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
1040 
1041 	if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) {
1042 		/* timestamp was stale */
1043 		mask = mask_from(q->bitmaps[ER], agg->grp->index);
1044 		if (mask) {
1045 			struct qfq_group *next = qfq_ffs(q, mask);
1046 			if (qfq_gt(roundedF, next->F)) {
1047 				if (qfq_gt(limit, next->F))
1048 					agg->S = next->F;
1049 				else /* preserve timestamp correctness */
1050 					agg->S = limit;
1051 				return;
1052 			}
1053 		}
1054 		agg->S = q->V;
1055 	} else  /* timestamp is not stale */
1056 		agg->S = agg->F;
1057 }
1058 
1059 /* Update the timestamps of agg before scheduling/rescheduling it for
1060  * service.  In particular, assign to agg->F its maximum possible
1061  * value, i.e., the virtual finish time with which the aggregate
1062  * should be labeled if it used all its budget once in service.
1063  */
1064 static inline void
1065 qfq_update_agg_ts(struct qfq_sched *q,
1066 		    struct qfq_aggregate *agg, enum update_reason reason)
1067 {
1068 	if (reason != requeue)
1069 		qfq_update_start(q, agg);
1070 	else /* just charge agg for the service received */
1071 		agg->S = agg->F;
1072 
1073 	agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w;
1074 }
1075 
1076 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg);
1077 
1078 static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
1079 {
1080 	struct qfq_sched *q = qdisc_priv(sch);
1081 	struct qfq_aggregate *in_serv_agg = q->in_serv_agg;
1082 	struct qfq_class *cl;
1083 	struct sk_buff *skb = NULL;
1084 	/* next-packet len, 0 means no more active classes in in-service agg */
1085 	unsigned int len = 0;
1086 
1087 	if (in_serv_agg == NULL)
1088 		return NULL;
1089 
1090 	if (!list_empty(&in_serv_agg->active))
1091 		skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1092 
1093 	/*
1094 	 * If there are no active classes in the in-service aggregate,
1095 	 * or if the aggregate has not enough budget to serve its next
1096 	 * class, then choose the next aggregate to serve.
1097 	 */
1098 	if (len == 0 || in_serv_agg->budget < len) {
1099 		charge_actual_service(in_serv_agg);
1100 
1101 		/* recharge the budget of the aggregate */
1102 		in_serv_agg->initial_budget = in_serv_agg->budget =
1103 			in_serv_agg->budgetmax;
1104 
1105 		if (!list_empty(&in_serv_agg->active)) {
1106 			/*
1107 			 * Still active: reschedule for
1108 			 * service. Possible optimization: if no other
1109 			 * aggregate is active, then there is no point
1110 			 * in rescheduling this aggregate, and we can
1111 			 * just keep it as the in-service one. This
1112 			 * should be however a corner case, and to
1113 			 * handle it, we would need to maintain an
1114 			 * extra num_active_aggs field.
1115 			*/
1116 			qfq_update_agg_ts(q, in_serv_agg, requeue);
1117 			qfq_schedule_agg(q, in_serv_agg);
1118 		} else if (sch->q.qlen == 0) { /* no aggregate to serve */
1119 			q->in_serv_agg = NULL;
1120 			return NULL;
1121 		}
1122 
1123 		/*
1124 		 * If we get here, there are other aggregates queued:
1125 		 * choose the new aggregate to serve.
1126 		 */
1127 		in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
1128 		skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1129 	}
1130 	if (!skb)
1131 		return NULL;
1132 
1133 	qdisc_qstats_backlog_dec(sch, skb);
1134 	sch->q.qlen--;
1135 	qdisc_bstats_update(sch, skb);
1136 
1137 	agg_dequeue(in_serv_agg, cl, len);
1138 	/* If lmax is lowered, through qfq_change_class, for a class
1139 	 * owning pending packets with larger size than the new value
1140 	 * of lmax, then the following condition may hold.
1141 	 */
1142 	if (unlikely(in_serv_agg->budget < len))
1143 		in_serv_agg->budget = 0;
1144 	else
1145 		in_serv_agg->budget -= len;
1146 
1147 	q->V += (u64)len * q->iwsum;
1148 	pr_debug("qfq dequeue: len %u F %lld now %lld\n",
1149 		 len, (unsigned long long) in_serv_agg->F,
1150 		 (unsigned long long) q->V);
1151 
1152 	return skb;
1153 }
1154 
1155 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)
1156 {
1157 	struct qfq_group *grp;
1158 	struct qfq_aggregate *agg, *new_front_agg;
1159 	u64 old_F;
1160 
1161 	qfq_update_eligible(q);
1162 	q->oldV = q->V;
1163 
1164 	if (!q->bitmaps[ER])
1165 		return NULL;
1166 
1167 	grp = qfq_ffs(q, q->bitmaps[ER]);
1168 	old_F = grp->F;
1169 
1170 	agg = qfq_slot_head(grp);
1171 
1172 	/* agg starts to be served, remove it from schedule */
1173 	qfq_front_slot_remove(grp);
1174 
1175 	new_front_agg = qfq_slot_scan(grp);
1176 
1177 	if (new_front_agg == NULL) /* group is now inactive, remove from ER */
1178 		__clear_bit(grp->index, &q->bitmaps[ER]);
1179 	else {
1180 		u64 roundedS = qfq_round_down(new_front_agg->S,
1181 					      grp->slot_shift);
1182 		unsigned int s;
1183 
1184 		if (grp->S == roundedS)
1185 			return agg;
1186 		grp->S = roundedS;
1187 		grp->F = roundedS + (2ULL << grp->slot_shift);
1188 		__clear_bit(grp->index, &q->bitmaps[ER]);
1189 		s = qfq_calc_state(q, grp);
1190 		__set_bit(grp->index, &q->bitmaps[s]);
1191 	}
1192 
1193 	qfq_unblock_groups(q, grp->index, old_F);
1194 
1195 	return agg;
1196 }
1197 
1198 static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
1199 		       struct sk_buff **to_free)
1200 {
1201 	unsigned int len = qdisc_pkt_len(skb), gso_segs;
1202 	struct qfq_sched *q = qdisc_priv(sch);
1203 	struct qfq_class *cl;
1204 	struct qfq_aggregate *agg;
1205 	int err = 0;
1206 	bool first;
1207 
1208 	cl = qfq_classify(skb, sch, &err);
1209 	if (cl == NULL) {
1210 		if (err & __NET_XMIT_BYPASS)
1211 			qdisc_qstats_drop(sch);
1212 		__qdisc_drop(skb, to_free);
1213 		return err;
1214 	}
1215 	pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
1216 
1217 	if (unlikely(cl->agg->lmax < len)) {
1218 		pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
1219 			 cl->agg->lmax, len, cl->common.classid);
1220 		err = qfq_change_agg(sch, cl, cl->agg->class_weight, len);
1221 		if (err) {
1222 			cl->qstats.drops++;
1223 			return qdisc_drop(skb, sch, to_free);
1224 		}
1225 	}
1226 
1227 	gso_segs = skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 1;
1228 	first = !cl->qdisc->q.qlen;
1229 	err = qdisc_enqueue(skb, cl->qdisc, to_free);
1230 	if (unlikely(err != NET_XMIT_SUCCESS)) {
1231 		pr_debug("qfq_enqueue: enqueue failed %d\n", err);
1232 		if (net_xmit_drop_count(err)) {
1233 			cl->qstats.drops++;
1234 			qdisc_qstats_drop(sch);
1235 		}
1236 		return err;
1237 	}
1238 
1239 	cl->bstats.bytes += len;
1240 	cl->bstats.packets += gso_segs;
1241 	sch->qstats.backlog += len;
1242 	++sch->q.qlen;
1243 
1244 	agg = cl->agg;
1245 	/* if the queue was not empty, then done here */
1246 	if (!first) {
1247 		if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) &&
1248 		    list_first_entry(&agg->active, struct qfq_class, alist)
1249 		    == cl && cl->deficit < len)
1250 			list_move_tail(&cl->alist, &agg->active);
1251 
1252 		return err;
1253 	}
1254 
1255 	/* schedule class for service within the aggregate */
1256 	cl->deficit = agg->lmax;
1257 	list_add_tail(&cl->alist, &agg->active);
1258 
1259 	if (list_first_entry(&agg->active, struct qfq_class, alist) != cl ||
1260 	    q->in_serv_agg == agg)
1261 		return err; /* non-empty or in service, nothing else to do */
1262 
1263 	qfq_activate_agg(q, agg, enqueue);
1264 
1265 	return err;
1266 }
1267 
1268 /*
1269  * Schedule aggregate according to its timestamps.
1270  */
1271 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1272 {
1273 	struct qfq_group *grp = agg->grp;
1274 	u64 roundedS;
1275 	int s;
1276 
1277 	roundedS = qfq_round_down(agg->S, grp->slot_shift);
1278 
1279 	/*
1280 	 * Insert agg in the correct bucket.
1281 	 * If agg->S >= grp->S we don't need to adjust the
1282 	 * bucket list and simply go to the insertion phase.
1283 	 * Otherwise grp->S is decreasing, we must make room
1284 	 * in the bucket list, and also recompute the group state.
1285 	 * Finally, if there were no flows in this group and nobody
1286 	 * was in ER make sure to adjust V.
1287 	 */
1288 	if (grp->full_slots) {
1289 		if (!qfq_gt(grp->S, agg->S))
1290 			goto skip_update;
1291 
1292 		/* create a slot for this agg->S */
1293 		qfq_slot_rotate(grp, roundedS);
1294 		/* group was surely ineligible, remove */
1295 		__clear_bit(grp->index, &q->bitmaps[IR]);
1296 		__clear_bit(grp->index, &q->bitmaps[IB]);
1297 	} else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) &&
1298 		   q->in_serv_agg == NULL)
1299 		q->V = roundedS;
1300 
1301 	grp->S = roundedS;
1302 	grp->F = roundedS + (2ULL << grp->slot_shift);
1303 	s = qfq_calc_state(q, grp);
1304 	__set_bit(grp->index, &q->bitmaps[s]);
1305 
1306 	pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
1307 		 s, q->bitmaps[s],
1308 		 (unsigned long long) agg->S,
1309 		 (unsigned long long) agg->F,
1310 		 (unsigned long long) q->V);
1311 
1312 skip_update:
1313 	qfq_slot_insert(grp, agg, roundedS);
1314 }
1315 
1316 
1317 /* Update agg ts and schedule agg for service */
1318 static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
1319 			     enum update_reason reason)
1320 {
1321 	agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */
1322 
1323 	qfq_update_agg_ts(q, agg, reason);
1324 	if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */
1325 		q->in_serv_agg = agg; /* start serving this aggregate */
1326 		 /* update V: to be in service, agg must be eligible */
1327 		q->oldV = q->V = agg->S;
1328 	} else if (agg != q->in_serv_agg)
1329 		qfq_schedule_agg(q, agg);
1330 }
1331 
1332 static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
1333 			    struct qfq_aggregate *agg)
1334 {
1335 	unsigned int i, offset;
1336 	u64 roundedS;
1337 
1338 	roundedS = qfq_round_down(agg->S, grp->slot_shift);
1339 	offset = (roundedS - grp->S) >> grp->slot_shift;
1340 
1341 	i = (grp->front + offset) % QFQ_MAX_SLOTS;
1342 
1343 	hlist_del(&agg->next);
1344 	if (hlist_empty(&grp->slots[i]))
1345 		__clear_bit(offset, &grp->full_slots);
1346 }
1347 
1348 /*
1349  * Called to forcibly deschedule an aggregate.  If the aggregate is
1350  * not in the front bucket, or if the latter has other aggregates in
1351  * the front bucket, we can simply remove the aggregate with no other
1352  * side effects.
1353  * Otherwise we must propagate the event up.
1354  */
1355 static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1356 {
1357 	struct qfq_group *grp = agg->grp;
1358 	unsigned long mask;
1359 	u64 roundedS;
1360 	int s;
1361 
1362 	if (agg == q->in_serv_agg) {
1363 		charge_actual_service(agg);
1364 		q->in_serv_agg = qfq_choose_next_agg(q);
1365 		return;
1366 	}
1367 
1368 	agg->F = agg->S;
1369 	qfq_slot_remove(q, grp, agg);
1370 
1371 	if (!grp->full_slots) {
1372 		__clear_bit(grp->index, &q->bitmaps[IR]);
1373 		__clear_bit(grp->index, &q->bitmaps[EB]);
1374 		__clear_bit(grp->index, &q->bitmaps[IB]);
1375 
1376 		if (test_bit(grp->index, &q->bitmaps[ER]) &&
1377 		    !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
1378 			mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
1379 			if (mask)
1380 				mask = ~((1UL << __fls(mask)) - 1);
1381 			else
1382 				mask = ~0UL;
1383 			qfq_move_groups(q, mask, EB, ER);
1384 			qfq_move_groups(q, mask, IB, IR);
1385 		}
1386 		__clear_bit(grp->index, &q->bitmaps[ER]);
1387 	} else if (hlist_empty(&grp->slots[grp->front])) {
1388 		agg = qfq_slot_scan(grp);
1389 		roundedS = qfq_round_down(agg->S, grp->slot_shift);
1390 		if (grp->S != roundedS) {
1391 			__clear_bit(grp->index, &q->bitmaps[ER]);
1392 			__clear_bit(grp->index, &q->bitmaps[IR]);
1393 			__clear_bit(grp->index, &q->bitmaps[EB]);
1394 			__clear_bit(grp->index, &q->bitmaps[IB]);
1395 			grp->S = roundedS;
1396 			grp->F = roundedS + (2ULL << grp->slot_shift);
1397 			s = qfq_calc_state(q, grp);
1398 			__set_bit(grp->index, &q->bitmaps[s]);
1399 		}
1400 	}
1401 }
1402 
1403 static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1404 {
1405 	struct qfq_sched *q = qdisc_priv(sch);
1406 	struct qfq_class *cl = (struct qfq_class *)arg;
1407 
1408 	qfq_deactivate_class(q, cl);
1409 }
1410 
1411 static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt,
1412 			  struct netlink_ext_ack *extack)
1413 {
1414 	struct qfq_sched *q = qdisc_priv(sch);
1415 	struct qfq_group *grp;
1416 	int i, j, err;
1417 	u32 max_cl_shift, maxbudg_shift, max_classes;
1418 
1419 	err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
1420 	if (err)
1421 		return err;
1422 
1423 	err = qdisc_class_hash_init(&q->clhash);
1424 	if (err < 0)
1425 		return err;
1426 
1427 	if (qdisc_dev(sch)->tx_queue_len + 1 > QFQ_MAX_AGG_CLASSES)
1428 		max_classes = QFQ_MAX_AGG_CLASSES;
1429 	else
1430 		max_classes = qdisc_dev(sch)->tx_queue_len + 1;
1431 	/* max_cl_shift = floor(log_2(max_classes)) */
1432 	max_cl_shift = __fls(max_classes);
1433 	q->max_agg_classes = 1<<max_cl_shift;
1434 
1435 	/* maxbudg_shift = log2(max_len * max_classes_per_agg) */
1436 	maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift;
1437 	q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX;
1438 
1439 	for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1440 		grp = &q->groups[i];
1441 		grp->index = i;
1442 		grp->slot_shift = q->min_slot_shift + i;
1443 		for (j = 0; j < QFQ_MAX_SLOTS; j++)
1444 			INIT_HLIST_HEAD(&grp->slots[j]);
1445 	}
1446 
1447 	INIT_HLIST_HEAD(&q->nonfull_aggs);
1448 
1449 	return 0;
1450 }
1451 
1452 static void qfq_reset_qdisc(struct Qdisc *sch)
1453 {
1454 	struct qfq_sched *q = qdisc_priv(sch);
1455 	struct qfq_class *cl;
1456 	unsigned int i;
1457 
1458 	for (i = 0; i < q->clhash.hashsize; i++) {
1459 		hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1460 			if (cl->qdisc->q.qlen > 0)
1461 				qfq_deactivate_class(q, cl);
1462 
1463 			qdisc_reset(cl->qdisc);
1464 		}
1465 	}
1466 	sch->qstats.backlog = 0;
1467 	sch->q.qlen = 0;
1468 }
1469 
1470 static void qfq_destroy_qdisc(struct Qdisc *sch)
1471 {
1472 	struct qfq_sched *q = qdisc_priv(sch);
1473 	struct qfq_class *cl;
1474 	struct hlist_node *next;
1475 	unsigned int i;
1476 
1477 	tcf_block_put(q->block);
1478 
1479 	for (i = 0; i < q->clhash.hashsize; i++) {
1480 		hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1481 					  common.hnode) {
1482 			qfq_destroy_class(sch, cl);
1483 		}
1484 	}
1485 	qdisc_class_hash_destroy(&q->clhash);
1486 }
1487 
1488 static const struct Qdisc_class_ops qfq_class_ops = {
1489 	.change		= qfq_change_class,
1490 	.delete		= qfq_delete_class,
1491 	.find		= qfq_search_class,
1492 	.tcf_block	= qfq_tcf_block,
1493 	.bind_tcf	= qfq_bind_tcf,
1494 	.unbind_tcf	= qfq_unbind_tcf,
1495 	.graft		= qfq_graft_class,
1496 	.leaf		= qfq_class_leaf,
1497 	.qlen_notify	= qfq_qlen_notify,
1498 	.dump		= qfq_dump_class,
1499 	.dump_stats	= qfq_dump_class_stats,
1500 	.walk		= qfq_walk,
1501 };
1502 
1503 static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1504 	.cl_ops		= &qfq_class_ops,
1505 	.id		= "qfq",
1506 	.priv_size	= sizeof(struct qfq_sched),
1507 	.enqueue	= qfq_enqueue,
1508 	.dequeue	= qfq_dequeue,
1509 	.peek		= qdisc_peek_dequeued,
1510 	.init		= qfq_init_qdisc,
1511 	.reset		= qfq_reset_qdisc,
1512 	.destroy	= qfq_destroy_qdisc,
1513 	.owner		= THIS_MODULE,
1514 };
1515 
1516 static int __init qfq_init(void)
1517 {
1518 	return register_qdisc(&qfq_qdisc_ops);
1519 }
1520 
1521 static void __exit qfq_exit(void)
1522 {
1523 	unregister_qdisc(&qfq_qdisc_ops);
1524 }
1525 
1526 module_init(qfq_init);
1527 module_exit(qfq_exit);
1528 MODULE_LICENSE("GPL");
1529