xref: /openbmc/linux/net/sched/sch_qfq.c (revision 6de298ff)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * net/sched/sch_qfq.c         Quick Fair Queueing Plus Scheduler.
4  *
5  * Copyright (c) 2009 Fabio Checconi, Luigi Rizzo, and Paolo Valente.
6  * Copyright (c) 2012 Paolo Valente.
7  */
8 
9 #include <linux/module.h>
10 #include <linux/init.h>
11 #include <linux/bitops.h>
12 #include <linux/errno.h>
13 #include <linux/netdevice.h>
14 #include <linux/pkt_sched.h>
15 #include <net/sch_generic.h>
16 #include <net/pkt_sched.h>
17 #include <net/pkt_cls.h>
18 
19 
20 /*  Quick Fair Queueing Plus
21     ========================
22 
23     Sources:
24 
25     [1] Paolo Valente,
26     "Reducing the Execution Time of Fair-Queueing Schedulers."
27     http://algo.ing.unimo.it/people/paolo/agg-sched/agg-sched.pdf
28 
29     Sources for QFQ:
30 
31     [2] Fabio Checconi, Luigi Rizzo, and Paolo Valente: "QFQ: Efficient
32     Packet Scheduling with Tight Bandwidth Distribution Guarantees."
33 
34     See also:
35     http://retis.sssup.it/~fabio/linux/qfq/
36  */
37 
38 /*
39 
40   QFQ+ divides classes into aggregates of at most MAX_AGG_CLASSES
41   classes. Each aggregate is timestamped with a virtual start time S
42   and a virtual finish time F, and scheduled according to its
43   timestamps. S and F are computed as a function of a system virtual
44   time function V. The classes within each aggregate are instead
45   scheduled with DRR.
46 
47   To speed up operations, QFQ+ divides also aggregates into a limited
48   number of groups. Which group a class belongs to depends on the
49   ratio between the maximum packet length for the class and the weight
50   of the class. Groups have their own S and F. In the end, QFQ+
51   schedules groups, then aggregates within groups, then classes within
52   aggregates. See [1] and [2] for a full description.
53 
54   Virtual time computations.
55 
56   S, F and V are all computed in fixed point arithmetic with
57   FRAC_BITS decimal bits.
58 
59   QFQ_MAX_INDEX is the maximum index allowed for a group. We need
60 	one bit per index.
61   QFQ_MAX_WSHIFT is the maximum power of two supported as a weight.
62 
63   The layout of the bits is as below:
64 
65                    [ MTU_SHIFT ][      FRAC_BITS    ]
66                    [ MAX_INDEX    ][ MIN_SLOT_SHIFT ]
67 				 ^.__grp->index = 0
68 				 *.__grp->slot_shift
69 
70   where MIN_SLOT_SHIFT is derived by difference from the others.
71 
72   The max group index corresponds to Lmax/w_min, where
73   Lmax=1<<MTU_SHIFT, w_min = 1 .
74   From this, and knowing how many groups (MAX_INDEX) we want,
75   we can derive the shift corresponding to each group.
76 
77   Because we often need to compute
78 	F = S + len/w_i  and V = V + len/wsum
79   instead of storing w_i store the value
80 	inv_w = (1<<FRAC_BITS)/w_i
81   so we can do F = S + len * inv_w * wsum.
82   We use W_TOT in the formulas so we can easily move between
83   static and adaptive weight sum.
84 
85   The per-scheduler-instance data contain all the data structures
86   for the scheduler: bitmaps and bucket lists.
87 
88  */
89 
90 /*
91  * Maximum number of consecutive slots occupied by backlogged classes
92  * inside a group.
93  */
94 #define QFQ_MAX_SLOTS	32
95 
96 /*
97  * Shifts used for aggregate<->group mapping.  We allow class weights that are
98  * in the range [1, 2^MAX_WSHIFT], and we try to map each aggregate i to the
99  * group with the smallest index that can support the L_i / r_i configured
100  * for the classes in the aggregate.
101  *
102  * grp->index is the index of the group; and grp->slot_shift
103  * is the shift for the corresponding (scaled) sigma_i.
104  */
105 #define QFQ_MAX_INDEX		24
106 #define QFQ_MAX_WSHIFT		10
107 
108 #define	QFQ_MAX_WEIGHT		(1<<QFQ_MAX_WSHIFT) /* see qfq_slot_insert */
109 #define QFQ_MAX_WSUM		(64*QFQ_MAX_WEIGHT)
110 
111 #define FRAC_BITS		30	/* fixed point arithmetic */
112 #define ONE_FP			(1UL << FRAC_BITS)
113 
114 #define QFQ_MTU_SHIFT		16	/* to support TSO/GSO */
115 #define QFQ_MIN_LMAX		512	/* see qfq_slot_insert */
116 #define QFQ_MAX_LMAX		(1UL << QFQ_MTU_SHIFT)
117 
118 #define QFQ_MAX_AGG_CLASSES	8 /* max num classes per aggregate allowed */
119 
120 /*
121  * Possible group states.  These values are used as indexes for the bitmaps
122  * array of struct qfq_queue.
123  */
124 enum qfq_state { ER, IR, EB, IB, QFQ_MAX_STATE };
125 
126 struct qfq_group;
127 
128 struct qfq_aggregate;
129 
130 struct qfq_class {
131 	struct Qdisc_class_common common;
132 
133 	unsigned int filter_cnt;
134 
135 	struct gnet_stats_basic_sync bstats;
136 	struct gnet_stats_queue qstats;
137 	struct net_rate_estimator __rcu *rate_est;
138 	struct Qdisc *qdisc;
139 	struct list_head alist;		/* Link for active-classes list. */
140 	struct qfq_aggregate *agg;	/* Parent aggregate. */
141 	int deficit;			/* DRR deficit counter. */
142 };
143 
144 struct qfq_aggregate {
145 	struct hlist_node next;	/* Link for the slot list. */
146 	u64 S, F;		/* flow timestamps (exact) */
147 
148 	/* group we belong to. In principle we would need the index,
149 	 * which is log_2(lmax/weight), but we never reference it
150 	 * directly, only the group.
151 	 */
152 	struct qfq_group *grp;
153 
154 	/* these are copied from the flowset. */
155 	u32	class_weight; /* Weight of each class in this aggregate. */
156 	/* Max pkt size for the classes in this aggregate, DRR quantum. */
157 	int	lmax;
158 
159 	u32	inv_w;	    /* ONE_FP/(sum of weights of classes in aggr.). */
160 	u32	budgetmax;  /* Max budget for this aggregate. */
161 	u32	initial_budget, budget;     /* Initial and current budget. */
162 
163 	int		  num_classes;	/* Number of classes in this aggr. */
164 	struct list_head  active;	/* DRR queue of active classes. */
165 
166 	struct hlist_node nonfull_next;	/* See nonfull_aggs in qfq_sched. */
167 };
168 
169 struct qfq_group {
170 	u64 S, F;			/* group timestamps (approx). */
171 	unsigned int slot_shift;	/* Slot shift. */
172 	unsigned int index;		/* Group index. */
173 	unsigned int front;		/* Index of the front slot. */
174 	unsigned long full_slots;	/* non-empty slots */
175 
176 	/* Array of RR lists of active aggregates. */
177 	struct hlist_head slots[QFQ_MAX_SLOTS];
178 };
179 
180 struct qfq_sched {
181 	struct tcf_proto __rcu *filter_list;
182 	struct tcf_block	*block;
183 	struct Qdisc_class_hash clhash;
184 
185 	u64			oldV, V;	/* Precise virtual times. */
186 	struct qfq_aggregate	*in_serv_agg;   /* Aggregate being served. */
187 	u32			wsum;		/* weight sum */
188 	u32			iwsum;		/* inverse weight sum */
189 
190 	unsigned long bitmaps[QFQ_MAX_STATE];	    /* Group bitmaps. */
191 	struct qfq_group groups[QFQ_MAX_INDEX + 1]; /* The groups. */
192 	u32 min_slot_shift;	/* Index of the group-0 bit in the bitmaps. */
193 
194 	u32 max_agg_classes;		/* Max number of classes per aggr. */
195 	struct hlist_head nonfull_aggs; /* Aggs with room for more classes. */
196 };
197 
198 /*
199  * Possible reasons why the timestamps of an aggregate are updated
200  * enqueue: the aggregate switches from idle to active and must scheduled
201  *	    for service
202  * requeue: the aggregate finishes its budget, so it stops being served and
203  *	    must be rescheduled for service
204  */
205 enum update_reason {enqueue, requeue};
206 
207 static struct qfq_class *qfq_find_class(struct Qdisc *sch, u32 classid)
208 {
209 	struct qfq_sched *q = qdisc_priv(sch);
210 	struct Qdisc_class_common *clc;
211 
212 	clc = qdisc_class_find(&q->clhash, classid);
213 	if (clc == NULL)
214 		return NULL;
215 	return container_of(clc, struct qfq_class, common);
216 }
217 
218 static struct netlink_range_validation lmax_range = {
219 	.min = QFQ_MIN_LMAX,
220 	.max = QFQ_MAX_LMAX,
221 };
222 
223 static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = {
224 	[TCA_QFQ_WEIGHT] = NLA_POLICY_RANGE(NLA_U32, 1, QFQ_MAX_WEIGHT),
225 	[TCA_QFQ_LMAX] = NLA_POLICY_FULL_RANGE(NLA_U32, &lmax_range),
226 };
227 
228 /*
229  * Calculate a flow index, given its weight and maximum packet length.
230  * index = log_2(maxlen/weight) but we need to apply the scaling.
231  * This is used only once at flow creation.
232  */
233 static int qfq_calc_index(u32 inv_w, unsigned int maxlen, u32 min_slot_shift)
234 {
235 	u64 slot_size = (u64)maxlen * inv_w;
236 	unsigned long size_map;
237 	int index = 0;
238 
239 	size_map = slot_size >> min_slot_shift;
240 	if (!size_map)
241 		goto out;
242 
243 	index = __fls(size_map) + 1;	/* basically a log_2 */
244 	index -= !(slot_size - (1ULL << (index + min_slot_shift - 1)));
245 
246 	if (index < 0)
247 		index = 0;
248 out:
249 	pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n",
250 		 (unsigned long) ONE_FP/inv_w, maxlen, index);
251 
252 	return index;
253 }
254 
255 static void qfq_deactivate_agg(struct qfq_sched *, struct qfq_aggregate *);
256 static void qfq_activate_agg(struct qfq_sched *, struct qfq_aggregate *,
257 			     enum update_reason);
258 
259 static void qfq_init_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
260 			 u32 lmax, u32 weight)
261 {
262 	INIT_LIST_HEAD(&agg->active);
263 	hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
264 
265 	agg->lmax = lmax;
266 	agg->class_weight = weight;
267 }
268 
269 static struct qfq_aggregate *qfq_find_agg(struct qfq_sched *q,
270 					  u32 lmax, u32 weight)
271 {
272 	struct qfq_aggregate *agg;
273 
274 	hlist_for_each_entry(agg, &q->nonfull_aggs, nonfull_next)
275 		if (agg->lmax == lmax && agg->class_weight == weight)
276 			return agg;
277 
278 	return NULL;
279 }
280 
281 
282 /* Update aggregate as a function of the new number of classes. */
283 static void qfq_update_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
284 			   int new_num_classes)
285 {
286 	u32 new_agg_weight;
287 
288 	if (new_num_classes == q->max_agg_classes)
289 		hlist_del_init(&agg->nonfull_next);
290 
291 	if (agg->num_classes > new_num_classes &&
292 	    new_num_classes == q->max_agg_classes - 1) /* agg no more full */
293 		hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
294 
295 	/* The next assignment may let
296 	 * agg->initial_budget > agg->budgetmax
297 	 * hold, we will take it into account in charge_actual_service().
298 	 */
299 	agg->budgetmax = new_num_classes * agg->lmax;
300 	new_agg_weight = agg->class_weight * new_num_classes;
301 	agg->inv_w = ONE_FP/new_agg_weight;
302 
303 	if (agg->grp == NULL) {
304 		int i = qfq_calc_index(agg->inv_w, agg->budgetmax,
305 				       q->min_slot_shift);
306 		agg->grp = &q->groups[i];
307 	}
308 
309 	q->wsum +=
310 		(int) agg->class_weight * (new_num_classes - agg->num_classes);
311 	q->iwsum = ONE_FP / q->wsum;
312 
313 	agg->num_classes = new_num_classes;
314 }
315 
316 /* Add class to aggregate. */
317 static void qfq_add_to_agg(struct qfq_sched *q,
318 			   struct qfq_aggregate *agg,
319 			   struct qfq_class *cl)
320 {
321 	cl->agg = agg;
322 
323 	qfq_update_agg(q, agg, agg->num_classes+1);
324 	if (cl->qdisc->q.qlen > 0) { /* adding an active class */
325 		list_add_tail(&cl->alist, &agg->active);
326 		if (list_first_entry(&agg->active, struct qfq_class, alist) ==
327 		    cl && q->in_serv_agg != agg) /* agg was inactive */
328 			qfq_activate_agg(q, agg, enqueue); /* schedule agg */
329 	}
330 }
331 
332 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *);
333 
334 static void qfq_destroy_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
335 {
336 	hlist_del_init(&agg->nonfull_next);
337 	q->wsum -= agg->class_weight;
338 	if (q->wsum != 0)
339 		q->iwsum = ONE_FP / q->wsum;
340 
341 	if (q->in_serv_agg == agg)
342 		q->in_serv_agg = qfq_choose_next_agg(q);
343 	kfree(agg);
344 }
345 
346 /* Deschedule class from within its parent aggregate. */
347 static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
348 {
349 	struct qfq_aggregate *agg = cl->agg;
350 
351 
352 	list_del(&cl->alist); /* remove from RR queue of the aggregate */
353 	if (list_empty(&agg->active)) /* agg is now inactive */
354 		qfq_deactivate_agg(q, agg);
355 }
356 
357 /* Remove class from its parent aggregate. */
358 static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
359 {
360 	struct qfq_aggregate *agg = cl->agg;
361 
362 	cl->agg = NULL;
363 	if (agg->num_classes == 1) { /* agg being emptied, destroy it */
364 		qfq_destroy_agg(q, agg);
365 		return;
366 	}
367 	qfq_update_agg(q, agg, agg->num_classes-1);
368 }
369 
370 /* Deschedule class and remove it from its parent aggregate. */
371 static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
372 {
373 	if (cl->qdisc->q.qlen > 0) /* class is active */
374 		qfq_deactivate_class(q, cl);
375 
376 	qfq_rm_from_agg(q, cl);
377 }
378 
379 /* Move class to a new aggregate, matching the new class weight and/or lmax */
380 static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight,
381 			   u32 lmax)
382 {
383 	struct qfq_sched *q = qdisc_priv(sch);
384 	struct qfq_aggregate *new_agg = qfq_find_agg(q, lmax, weight);
385 
386 	if (new_agg == NULL) { /* create new aggregate */
387 		new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC);
388 		if (new_agg == NULL)
389 			return -ENOBUFS;
390 		qfq_init_agg(q, new_agg, lmax, weight);
391 	}
392 	qfq_deact_rm_from_agg(q, cl);
393 	qfq_add_to_agg(q, new_agg, cl);
394 
395 	return 0;
396 }
397 
398 static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
399 			    struct nlattr **tca, unsigned long *arg,
400 			    struct netlink_ext_ack *extack)
401 {
402 	struct qfq_sched *q = qdisc_priv(sch);
403 	struct qfq_class *cl = (struct qfq_class *)*arg;
404 	bool existing = false;
405 	struct nlattr *tb[TCA_QFQ_MAX + 1];
406 	struct qfq_aggregate *new_agg = NULL;
407 	u32 weight, lmax, inv_w;
408 	int err;
409 	int delta_w;
410 
411 	if (NL_REQ_ATTR_CHECK(extack, NULL, tca, TCA_OPTIONS)) {
412 		NL_SET_ERR_MSG_MOD(extack, "missing options");
413 		return -EINVAL;
414 	}
415 
416 	err = nla_parse_nested_deprecated(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS],
417 					  qfq_policy, extack);
418 	if (err < 0)
419 		return err;
420 
421 	if (tb[TCA_QFQ_WEIGHT])
422 		weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]);
423 	else
424 		weight = 1;
425 
426 	if (tb[TCA_QFQ_LMAX])
427 		lmax = nla_get_u32(tb[TCA_QFQ_LMAX]);
428 	else
429 		lmax = psched_mtu(qdisc_dev(sch));
430 
431 	inv_w = ONE_FP / weight;
432 	weight = ONE_FP / inv_w;
433 
434 	if (cl != NULL &&
435 	    lmax == cl->agg->lmax &&
436 	    weight == cl->agg->class_weight)
437 		return 0; /* nothing to change */
438 
439 	delta_w = weight - (cl ? cl->agg->class_weight : 0);
440 
441 	if (q->wsum + delta_w > QFQ_MAX_WSUM) {
442 		NL_SET_ERR_MSG_FMT_MOD(extack,
443 				       "total weight out of range (%d + %u)\n",
444 				       delta_w, q->wsum);
445 		return -EINVAL;
446 	}
447 
448 	if (cl != NULL) { /* modify existing class */
449 		if (tca[TCA_RATE]) {
450 			err = gen_replace_estimator(&cl->bstats, NULL,
451 						    &cl->rate_est,
452 						    NULL,
453 						    true,
454 						    tca[TCA_RATE]);
455 			if (err)
456 				return err;
457 		}
458 		existing = true;
459 		goto set_change_agg;
460 	}
461 
462 	/* create and init new class */
463 	cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
464 	if (cl == NULL)
465 		return -ENOBUFS;
466 
467 	gnet_stats_basic_sync_init(&cl->bstats);
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 					true,
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 
489 set_change_agg:
490 	sch_tree_lock(sch);
491 	new_agg = qfq_find_agg(q, lmax, weight);
492 	if (new_agg == NULL) { /* create new aggregate */
493 		sch_tree_unlock(sch);
494 		new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL);
495 		if (new_agg == NULL) {
496 			err = -ENOBUFS;
497 			gen_kill_estimator(&cl->rate_est);
498 			goto destroy_class;
499 		}
500 		sch_tree_lock(sch);
501 		qfq_init_agg(q, new_agg, lmax, weight);
502 	}
503 	if (existing)
504 		qfq_deact_rm_from_agg(q, cl);
505 	else
506 		qdisc_class_hash_insert(&q->clhash, &cl->common);
507 	qfq_add_to_agg(q, new_agg, cl);
508 	sch_tree_unlock(sch);
509 	qdisc_class_hash_grow(sch, &q->clhash);
510 
511 	*arg = (unsigned long)cl;
512 	return 0;
513 
514 destroy_class:
515 	qdisc_put(cl->qdisc);
516 	kfree(cl);
517 	return err;
518 }
519 
520 static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
521 {
522 	struct qfq_sched *q = qdisc_priv(sch);
523 
524 	qfq_rm_from_agg(q, cl);
525 	gen_kill_estimator(&cl->rate_est);
526 	qdisc_put(cl->qdisc);
527 	kfree(cl);
528 }
529 
530 static int qfq_delete_class(struct Qdisc *sch, unsigned long arg,
531 			    struct netlink_ext_ack *extack)
532 {
533 	struct qfq_sched *q = qdisc_priv(sch);
534 	struct qfq_class *cl = (struct qfq_class *)arg;
535 
536 	if (cl->filter_cnt > 0)
537 		return -EBUSY;
538 
539 	sch_tree_lock(sch);
540 
541 	qdisc_purge_queue(cl->qdisc);
542 	qdisc_class_hash_remove(&q->clhash, &cl->common);
543 
544 	sch_tree_unlock(sch);
545 
546 	qfq_destroy_class(sch, cl);
547 	return 0;
548 }
549 
550 static unsigned long qfq_search_class(struct Qdisc *sch, u32 classid)
551 {
552 	return (unsigned long)qfq_find_class(sch, classid);
553 }
554 
555 static struct tcf_block *qfq_tcf_block(struct Qdisc *sch, unsigned long cl,
556 				       struct netlink_ext_ack *extack)
557 {
558 	struct qfq_sched *q = qdisc_priv(sch);
559 
560 	if (cl)
561 		return NULL;
562 
563 	return q->block;
564 }
565 
566 static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
567 				  u32 classid)
568 {
569 	struct qfq_class *cl = qfq_find_class(sch, classid);
570 
571 	if (cl != NULL)
572 		cl->filter_cnt++;
573 
574 	return (unsigned long)cl;
575 }
576 
577 static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
578 {
579 	struct qfq_class *cl = (struct qfq_class *)arg;
580 
581 	cl->filter_cnt--;
582 }
583 
584 static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
585 			   struct Qdisc *new, struct Qdisc **old,
586 			   struct netlink_ext_ack *extack)
587 {
588 	struct qfq_class *cl = (struct qfq_class *)arg;
589 
590 	if (new == NULL) {
591 		new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
592 					cl->common.classid, NULL);
593 		if (new == NULL)
594 			new = &noop_qdisc;
595 	}
596 
597 	*old = qdisc_replace(sch, new, &cl->qdisc);
598 	return 0;
599 }
600 
601 static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
602 {
603 	struct qfq_class *cl = (struct qfq_class *)arg;
604 
605 	return cl->qdisc;
606 }
607 
608 static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
609 			  struct sk_buff *skb, struct tcmsg *tcm)
610 {
611 	struct qfq_class *cl = (struct qfq_class *)arg;
612 	struct nlattr *nest;
613 
614 	tcm->tcm_parent	= TC_H_ROOT;
615 	tcm->tcm_handle	= cl->common.classid;
616 	tcm->tcm_info	= cl->qdisc->handle;
617 
618 	nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
619 	if (nest == NULL)
620 		goto nla_put_failure;
621 	if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) ||
622 	    nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax))
623 		goto nla_put_failure;
624 	return nla_nest_end(skb, nest);
625 
626 nla_put_failure:
627 	nla_nest_cancel(skb, nest);
628 	return -EMSGSIZE;
629 }
630 
631 static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
632 				struct gnet_dump *d)
633 {
634 	struct qfq_class *cl = (struct qfq_class *)arg;
635 	struct tc_qfq_stats xstats;
636 
637 	memset(&xstats, 0, sizeof(xstats));
638 
639 	xstats.weight = cl->agg->class_weight;
640 	xstats.lmax = cl->agg->lmax;
641 
642 	if (gnet_stats_copy_basic(d, NULL, &cl->bstats, true) < 0 ||
643 	    gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
644 	    qdisc_qstats_copy(d, cl->qdisc) < 0)
645 		return -1;
646 
647 	return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
648 }
649 
650 static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
651 {
652 	struct qfq_sched *q = qdisc_priv(sch);
653 	struct qfq_class *cl;
654 	unsigned int i;
655 
656 	if (arg->stop)
657 		return;
658 
659 	for (i = 0; i < q->clhash.hashsize; i++) {
660 		hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
661 			if (!tc_qdisc_stats_dump(sch, (unsigned long)cl, arg))
662 				return;
663 		}
664 	}
665 }
666 
667 static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
668 				      int *qerr)
669 {
670 	struct qfq_sched *q = qdisc_priv(sch);
671 	struct qfq_class *cl;
672 	struct tcf_result res;
673 	struct tcf_proto *fl;
674 	int result;
675 
676 	if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
677 		pr_debug("qfq_classify: found %d\n", skb->priority);
678 		cl = qfq_find_class(sch, skb->priority);
679 		if (cl != NULL)
680 			return cl;
681 	}
682 
683 	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
684 	fl = rcu_dereference_bh(q->filter_list);
685 	result = tcf_classify(skb, NULL, fl, &res, false);
686 	if (result >= 0) {
687 #ifdef CONFIG_NET_CLS_ACT
688 		switch (result) {
689 		case TC_ACT_QUEUED:
690 		case TC_ACT_STOLEN:
691 		case TC_ACT_TRAP:
692 			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
693 			fallthrough;
694 		case TC_ACT_SHOT:
695 			return NULL;
696 		}
697 #endif
698 		cl = (struct qfq_class *)res.class;
699 		if (cl == NULL)
700 			cl = qfq_find_class(sch, res.classid);
701 		return cl;
702 	}
703 
704 	return NULL;
705 }
706 
707 /* Generic comparison function, handling wraparound. */
708 static inline int qfq_gt(u64 a, u64 b)
709 {
710 	return (s64)(a - b) > 0;
711 }
712 
713 /* Round a precise timestamp to its slotted value. */
714 static inline u64 qfq_round_down(u64 ts, unsigned int shift)
715 {
716 	return ts & ~((1ULL << shift) - 1);
717 }
718 
719 /* return the pointer to the group with lowest index in the bitmap */
720 static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
721 					unsigned long bitmap)
722 {
723 	int index = __ffs(bitmap);
724 	return &q->groups[index];
725 }
726 /* Calculate a mask to mimic what would be ffs_from(). */
727 static inline unsigned long mask_from(unsigned long bitmap, int from)
728 {
729 	return bitmap & ~((1UL << from) - 1);
730 }
731 
732 /*
733  * The state computation relies on ER=0, IR=1, EB=2, IB=3
734  * First compute eligibility comparing grp->S, q->V,
735  * then check if someone is blocking us and possibly add EB
736  */
737 static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
738 {
739 	/* if S > V we are not eligible */
740 	unsigned int state = qfq_gt(grp->S, q->V);
741 	unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
742 	struct qfq_group *next;
743 
744 	if (mask) {
745 		next = qfq_ffs(q, mask);
746 		if (qfq_gt(grp->F, next->F))
747 			state |= EB;
748 	}
749 
750 	return state;
751 }
752 
753 
754 /*
755  * In principle
756  *	q->bitmaps[dst] |= q->bitmaps[src] & mask;
757  *	q->bitmaps[src] &= ~mask;
758  * but we should make sure that src != dst
759  */
760 static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
761 				   int src, int dst)
762 {
763 	q->bitmaps[dst] |= q->bitmaps[src] & mask;
764 	q->bitmaps[src] &= ~mask;
765 }
766 
767 static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
768 {
769 	unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
770 	struct qfq_group *next;
771 
772 	if (mask) {
773 		next = qfq_ffs(q, mask);
774 		if (!qfq_gt(next->F, old_F))
775 			return;
776 	}
777 
778 	mask = (1UL << index) - 1;
779 	qfq_move_groups(q, mask, EB, ER);
780 	qfq_move_groups(q, mask, IB, IR);
781 }
782 
783 /*
784  * perhaps
785  *
786 	old_V ^= q->V;
787 	old_V >>= q->min_slot_shift;
788 	if (old_V) {
789 		...
790 	}
791  *
792  */
793 static void qfq_make_eligible(struct qfq_sched *q)
794 {
795 	unsigned long vslot = q->V >> q->min_slot_shift;
796 	unsigned long old_vslot = q->oldV >> q->min_slot_shift;
797 
798 	if (vslot != old_vslot) {
799 		unsigned long mask;
800 		int last_flip_pos = fls(vslot ^ old_vslot);
801 
802 		if (last_flip_pos > 31) /* higher than the number of groups */
803 			mask = ~0UL;    /* make all groups eligible */
804 		else
805 			mask = (1UL << last_flip_pos) - 1;
806 
807 		qfq_move_groups(q, mask, IR, ER);
808 		qfq_move_groups(q, mask, IB, EB);
809 	}
810 }
811 
812 /*
813  * The index of the slot in which the input aggregate agg is to be
814  * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2'
815  * and not a '-1' because the start time of the group may be moved
816  * backward by one slot after the aggregate has been inserted, and
817  * this would cause non-empty slots to be right-shifted by one
818  * position.
819  *
820  * QFQ+ fully satisfies this bound to the slot index if the parameters
821  * of the classes are not changed dynamically, and if QFQ+ never
822  * happens to postpone the service of agg unjustly, i.e., it never
823  * happens that the aggregate becomes backlogged and eligible, or just
824  * eligible, while an aggregate with a higher approximated finish time
825  * is being served. In particular, in this case QFQ+ guarantees that
826  * the timestamps of agg are low enough that the slot index is never
827  * higher than 2. Unfortunately, QFQ+ cannot provide the same
828  * guarantee if it happens to unjustly postpone the service of agg, or
829  * if the parameters of some class are changed.
830  *
831  * As for the first event, i.e., an out-of-order service, the
832  * upper bound to the slot index guaranteed by QFQ+ grows to
833  * 2 +
834  * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
835  * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1.
836  *
837  * The following function deals with this problem by backward-shifting
838  * the timestamps of agg, if needed, so as to guarantee that the slot
839  * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may
840  * cause the service of other aggregates to be postponed, yet the
841  * worst-case guarantees of these aggregates are not violated.  In
842  * fact, in case of no out-of-order service, the timestamps of agg
843  * would have been even lower than they are after the backward shift,
844  * because QFQ+ would have guaranteed a maximum value equal to 2 for
845  * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose
846  * service is postponed because of the backward-shift would have
847  * however waited for the service of agg before being served.
848  *
849  * The other event that may cause the slot index to be higher than 2
850  * for agg is a recent change of the parameters of some class. If the
851  * weight of a class is increased or the lmax (max_pkt_size) of the
852  * class is decreased, then a new aggregate with smaller slot size
853  * than the original parent aggregate of the class may happen to be
854  * activated. The activation of this aggregate should be properly
855  * delayed to when the service of the class has finished in the ideal
856  * system tracked by QFQ+. If the activation of the aggregate is not
857  * delayed to this reference time instant, then this aggregate may be
858  * unjustly served before other aggregates waiting for service. This
859  * may cause the above bound to the slot index to be violated for some
860  * of these unlucky aggregates.
861  *
862  * Instead of delaying the activation of the new aggregate, which is
863  * quite complex, the above-discussed capping of the slot index is
864  * used to handle also the consequences of a change of the parameters
865  * of a class.
866  */
867 static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg,
868 			    u64 roundedS)
869 {
870 	u64 slot = (roundedS - grp->S) >> grp->slot_shift;
871 	unsigned int i; /* slot index in the bucket list */
872 
873 	if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
874 		u64 deltaS = roundedS - grp->S -
875 			((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
876 		agg->S -= deltaS;
877 		agg->F -= deltaS;
878 		slot = QFQ_MAX_SLOTS - 2;
879 	}
880 
881 	i = (grp->front + slot) % QFQ_MAX_SLOTS;
882 
883 	hlist_add_head(&agg->next, &grp->slots[i]);
884 	__set_bit(slot, &grp->full_slots);
885 }
886 
887 /* Maybe introduce hlist_first_entry?? */
888 static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp)
889 {
890 	return hlist_entry(grp->slots[grp->front].first,
891 			   struct qfq_aggregate, next);
892 }
893 
894 /*
895  * remove the entry from the slot
896  */
897 static void qfq_front_slot_remove(struct qfq_group *grp)
898 {
899 	struct qfq_aggregate *agg = qfq_slot_head(grp);
900 
901 	BUG_ON(!agg);
902 	hlist_del(&agg->next);
903 	if (hlist_empty(&grp->slots[grp->front]))
904 		__clear_bit(0, &grp->full_slots);
905 }
906 
907 /*
908  * Returns the first aggregate in the first non-empty bucket of the
909  * group. As a side effect, adjusts the bucket list so the first
910  * non-empty bucket is at position 0 in full_slots.
911  */
912 static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp)
913 {
914 	unsigned int i;
915 
916 	pr_debug("qfq slot_scan: grp %u full %#lx\n",
917 		 grp->index, grp->full_slots);
918 
919 	if (grp->full_slots == 0)
920 		return NULL;
921 
922 	i = __ffs(grp->full_slots);  /* zero based */
923 	if (i > 0) {
924 		grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
925 		grp->full_slots >>= i;
926 	}
927 
928 	return qfq_slot_head(grp);
929 }
930 
931 /*
932  * adjust the bucket list. When the start time of a group decreases,
933  * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
934  * move the objects. The mask of occupied slots must be shifted
935  * because we use ffs() to find the first non-empty slot.
936  * This covers decreases in the group's start time, but what about
937  * increases of the start time ?
938  * Here too we should make sure that i is less than 32
939  */
940 static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
941 {
942 	unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
943 
944 	grp->full_slots <<= i;
945 	grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
946 }
947 
948 static void qfq_update_eligible(struct qfq_sched *q)
949 {
950 	struct qfq_group *grp;
951 	unsigned long ineligible;
952 
953 	ineligible = q->bitmaps[IR] | q->bitmaps[IB];
954 	if (ineligible) {
955 		if (!q->bitmaps[ER]) {
956 			grp = qfq_ffs(q, ineligible);
957 			if (qfq_gt(grp->S, q->V))
958 				q->V = grp->S;
959 		}
960 		qfq_make_eligible(q);
961 	}
962 }
963 
964 /* Dequeue head packet of the head class in the DRR queue of the aggregate. */
965 static void agg_dequeue(struct qfq_aggregate *agg,
966 			struct qfq_class *cl, unsigned int len)
967 {
968 	qdisc_dequeue_peeked(cl->qdisc);
969 
970 	cl->deficit -= (int) len;
971 
972 	if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */
973 		list_del(&cl->alist);
974 	else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) {
975 		cl->deficit += agg->lmax;
976 		list_move_tail(&cl->alist, &agg->active);
977 	}
978 }
979 
980 static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg,
981 					   struct qfq_class **cl,
982 					   unsigned int *len)
983 {
984 	struct sk_buff *skb;
985 
986 	*cl = list_first_entry(&agg->active, struct qfq_class, alist);
987 	skb = (*cl)->qdisc->ops->peek((*cl)->qdisc);
988 	if (skb == NULL)
989 		WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
990 	else
991 		*len = qdisc_pkt_len(skb);
992 
993 	return skb;
994 }
995 
996 /* Update F according to the actual service received by the aggregate. */
997 static inline void charge_actual_service(struct qfq_aggregate *agg)
998 {
999 	/* Compute the service received by the aggregate, taking into
1000 	 * account that, after decreasing the number of classes in
1001 	 * agg, it may happen that
1002 	 * agg->initial_budget - agg->budget > agg->bugdetmax
1003 	 */
1004 	u32 service_received = min(agg->budgetmax,
1005 				   agg->initial_budget - agg->budget);
1006 
1007 	agg->F = agg->S + (u64)service_received * agg->inv_w;
1008 }
1009 
1010 /* Assign a reasonable start time for a new aggregate in group i.
1011  * Admissible values for \hat(F) are multiples of \sigma_i
1012  * no greater than V+\sigma_i . Larger values mean that
1013  * we had a wraparound so we consider the timestamp to be stale.
1014  *
1015  * If F is not stale and F >= V then we set S = F.
1016  * Otherwise we should assign S = V, but this may violate
1017  * the ordering in EB (see [2]). So, if we have groups in ER,
1018  * set S to the F_j of the first group j which would be blocking us.
1019  * We are guaranteed not to move S backward because
1020  * otherwise our group i would still be blocked.
1021  */
1022 static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg)
1023 {
1024 	unsigned long mask;
1025 	u64 limit, roundedF;
1026 	int slot_shift = agg->grp->slot_shift;
1027 
1028 	roundedF = qfq_round_down(agg->F, slot_shift);
1029 	limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
1030 
1031 	if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) {
1032 		/* timestamp was stale */
1033 		mask = mask_from(q->bitmaps[ER], agg->grp->index);
1034 		if (mask) {
1035 			struct qfq_group *next = qfq_ffs(q, mask);
1036 			if (qfq_gt(roundedF, next->F)) {
1037 				if (qfq_gt(limit, next->F))
1038 					agg->S = next->F;
1039 				else /* preserve timestamp correctness */
1040 					agg->S = limit;
1041 				return;
1042 			}
1043 		}
1044 		agg->S = q->V;
1045 	} else  /* timestamp is not stale */
1046 		agg->S = agg->F;
1047 }
1048 
1049 /* Update the timestamps of agg before scheduling/rescheduling it for
1050  * service.  In particular, assign to agg->F its maximum possible
1051  * value, i.e., the virtual finish time with which the aggregate
1052  * should be labeled if it used all its budget once in service.
1053  */
1054 static inline void
1055 qfq_update_agg_ts(struct qfq_sched *q,
1056 		    struct qfq_aggregate *agg, enum update_reason reason)
1057 {
1058 	if (reason != requeue)
1059 		qfq_update_start(q, agg);
1060 	else /* just charge agg for the service received */
1061 		agg->S = agg->F;
1062 
1063 	agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w;
1064 }
1065 
1066 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg);
1067 
1068 static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
1069 {
1070 	struct qfq_sched *q = qdisc_priv(sch);
1071 	struct qfq_aggregate *in_serv_agg = q->in_serv_agg;
1072 	struct qfq_class *cl;
1073 	struct sk_buff *skb = NULL;
1074 	/* next-packet len, 0 means no more active classes in in-service agg */
1075 	unsigned int len = 0;
1076 
1077 	if (in_serv_agg == NULL)
1078 		return NULL;
1079 
1080 	if (!list_empty(&in_serv_agg->active))
1081 		skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1082 
1083 	/*
1084 	 * If there are no active classes in the in-service aggregate,
1085 	 * or if the aggregate has not enough budget to serve its next
1086 	 * class, then choose the next aggregate to serve.
1087 	 */
1088 	if (len == 0 || in_serv_agg->budget < len) {
1089 		charge_actual_service(in_serv_agg);
1090 
1091 		/* recharge the budget of the aggregate */
1092 		in_serv_agg->initial_budget = in_serv_agg->budget =
1093 			in_serv_agg->budgetmax;
1094 
1095 		if (!list_empty(&in_serv_agg->active)) {
1096 			/*
1097 			 * Still active: reschedule for
1098 			 * service. Possible optimization: if no other
1099 			 * aggregate is active, then there is no point
1100 			 * in rescheduling this aggregate, and we can
1101 			 * just keep it as the in-service one. This
1102 			 * should be however a corner case, and to
1103 			 * handle it, we would need to maintain an
1104 			 * extra num_active_aggs field.
1105 			*/
1106 			qfq_update_agg_ts(q, in_serv_agg, requeue);
1107 			qfq_schedule_agg(q, in_serv_agg);
1108 		} else if (sch->q.qlen == 0) { /* no aggregate to serve */
1109 			q->in_serv_agg = NULL;
1110 			return NULL;
1111 		}
1112 
1113 		/*
1114 		 * If we get here, there are other aggregates queued:
1115 		 * choose the new aggregate to serve.
1116 		 */
1117 		in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
1118 		skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1119 	}
1120 	if (!skb)
1121 		return NULL;
1122 
1123 	qdisc_qstats_backlog_dec(sch, skb);
1124 	sch->q.qlen--;
1125 	qdisc_bstats_update(sch, skb);
1126 
1127 	agg_dequeue(in_serv_agg, cl, len);
1128 	/* If lmax is lowered, through qfq_change_class, for a class
1129 	 * owning pending packets with larger size than the new value
1130 	 * of lmax, then the following condition may hold.
1131 	 */
1132 	if (unlikely(in_serv_agg->budget < len))
1133 		in_serv_agg->budget = 0;
1134 	else
1135 		in_serv_agg->budget -= len;
1136 
1137 	q->V += (u64)len * q->iwsum;
1138 	pr_debug("qfq dequeue: len %u F %lld now %lld\n",
1139 		 len, (unsigned long long) in_serv_agg->F,
1140 		 (unsigned long long) q->V);
1141 
1142 	return skb;
1143 }
1144 
1145 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)
1146 {
1147 	struct qfq_group *grp;
1148 	struct qfq_aggregate *agg, *new_front_agg;
1149 	u64 old_F;
1150 
1151 	qfq_update_eligible(q);
1152 	q->oldV = q->V;
1153 
1154 	if (!q->bitmaps[ER])
1155 		return NULL;
1156 
1157 	grp = qfq_ffs(q, q->bitmaps[ER]);
1158 	old_F = grp->F;
1159 
1160 	agg = qfq_slot_head(grp);
1161 
1162 	/* agg starts to be served, remove it from schedule */
1163 	qfq_front_slot_remove(grp);
1164 
1165 	new_front_agg = qfq_slot_scan(grp);
1166 
1167 	if (new_front_agg == NULL) /* group is now inactive, remove from ER */
1168 		__clear_bit(grp->index, &q->bitmaps[ER]);
1169 	else {
1170 		u64 roundedS = qfq_round_down(new_front_agg->S,
1171 					      grp->slot_shift);
1172 		unsigned int s;
1173 
1174 		if (grp->S == roundedS)
1175 			return agg;
1176 		grp->S = roundedS;
1177 		grp->F = roundedS + (2ULL << grp->slot_shift);
1178 		__clear_bit(grp->index, &q->bitmaps[ER]);
1179 		s = qfq_calc_state(q, grp);
1180 		__set_bit(grp->index, &q->bitmaps[s]);
1181 	}
1182 
1183 	qfq_unblock_groups(q, grp->index, old_F);
1184 
1185 	return agg;
1186 }
1187 
1188 static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
1189 		       struct sk_buff **to_free)
1190 {
1191 	unsigned int len = qdisc_pkt_len(skb), gso_segs;
1192 	struct qfq_sched *q = qdisc_priv(sch);
1193 	struct qfq_class *cl;
1194 	struct qfq_aggregate *agg;
1195 	int err = 0;
1196 	bool first;
1197 
1198 	cl = qfq_classify(skb, sch, &err);
1199 	if (cl == NULL) {
1200 		if (err & __NET_XMIT_BYPASS)
1201 			qdisc_qstats_drop(sch);
1202 		__qdisc_drop(skb, to_free);
1203 		return err;
1204 	}
1205 	pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
1206 
1207 	if (unlikely(cl->agg->lmax < len)) {
1208 		pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
1209 			 cl->agg->lmax, len, cl->common.classid);
1210 		err = qfq_change_agg(sch, cl, cl->agg->class_weight, len);
1211 		if (err) {
1212 			cl->qstats.drops++;
1213 			return qdisc_drop(skb, sch, to_free);
1214 		}
1215 	}
1216 
1217 	gso_segs = skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 1;
1218 	first = !cl->qdisc->q.qlen;
1219 	err = qdisc_enqueue(skb, cl->qdisc, to_free);
1220 	if (unlikely(err != NET_XMIT_SUCCESS)) {
1221 		pr_debug("qfq_enqueue: enqueue failed %d\n", err);
1222 		if (net_xmit_drop_count(err)) {
1223 			cl->qstats.drops++;
1224 			qdisc_qstats_drop(sch);
1225 		}
1226 		return err;
1227 	}
1228 
1229 	_bstats_update(&cl->bstats, len, gso_segs);
1230 	sch->qstats.backlog += len;
1231 	++sch->q.qlen;
1232 
1233 	agg = cl->agg;
1234 	/* if the queue was not empty, then done here */
1235 	if (!first) {
1236 		if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) &&
1237 		    list_first_entry(&agg->active, struct qfq_class, alist)
1238 		    == cl && cl->deficit < len)
1239 			list_move_tail(&cl->alist, &agg->active);
1240 
1241 		return err;
1242 	}
1243 
1244 	/* schedule class for service within the aggregate */
1245 	cl->deficit = agg->lmax;
1246 	list_add_tail(&cl->alist, &agg->active);
1247 
1248 	if (list_first_entry(&agg->active, struct qfq_class, alist) != cl ||
1249 	    q->in_serv_agg == agg)
1250 		return err; /* non-empty or in service, nothing else to do */
1251 
1252 	qfq_activate_agg(q, agg, enqueue);
1253 
1254 	return err;
1255 }
1256 
1257 /*
1258  * Schedule aggregate according to its timestamps.
1259  */
1260 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1261 {
1262 	struct qfq_group *grp = agg->grp;
1263 	u64 roundedS;
1264 	int s;
1265 
1266 	roundedS = qfq_round_down(agg->S, grp->slot_shift);
1267 
1268 	/*
1269 	 * Insert agg in the correct bucket.
1270 	 * If agg->S >= grp->S we don't need to adjust the
1271 	 * bucket list and simply go to the insertion phase.
1272 	 * Otherwise grp->S is decreasing, we must make room
1273 	 * in the bucket list, and also recompute the group state.
1274 	 * Finally, if there were no flows in this group and nobody
1275 	 * was in ER make sure to adjust V.
1276 	 */
1277 	if (grp->full_slots) {
1278 		if (!qfq_gt(grp->S, agg->S))
1279 			goto skip_update;
1280 
1281 		/* create a slot for this agg->S */
1282 		qfq_slot_rotate(grp, roundedS);
1283 		/* group was surely ineligible, remove */
1284 		__clear_bit(grp->index, &q->bitmaps[IR]);
1285 		__clear_bit(grp->index, &q->bitmaps[IB]);
1286 	} else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) &&
1287 		   q->in_serv_agg == NULL)
1288 		q->V = roundedS;
1289 
1290 	grp->S = roundedS;
1291 	grp->F = roundedS + (2ULL << grp->slot_shift);
1292 	s = qfq_calc_state(q, grp);
1293 	__set_bit(grp->index, &q->bitmaps[s]);
1294 
1295 	pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
1296 		 s, q->bitmaps[s],
1297 		 (unsigned long long) agg->S,
1298 		 (unsigned long long) agg->F,
1299 		 (unsigned long long) q->V);
1300 
1301 skip_update:
1302 	qfq_slot_insert(grp, agg, roundedS);
1303 }
1304 
1305 
1306 /* Update agg ts and schedule agg for service */
1307 static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
1308 			     enum update_reason reason)
1309 {
1310 	agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */
1311 
1312 	qfq_update_agg_ts(q, agg, reason);
1313 	if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */
1314 		q->in_serv_agg = agg; /* start serving this aggregate */
1315 		 /* update V: to be in service, agg must be eligible */
1316 		q->oldV = q->V = agg->S;
1317 	} else if (agg != q->in_serv_agg)
1318 		qfq_schedule_agg(q, agg);
1319 }
1320 
1321 static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
1322 			    struct qfq_aggregate *agg)
1323 {
1324 	unsigned int i, offset;
1325 	u64 roundedS;
1326 
1327 	roundedS = qfq_round_down(agg->S, grp->slot_shift);
1328 	offset = (roundedS - grp->S) >> grp->slot_shift;
1329 
1330 	i = (grp->front + offset) % QFQ_MAX_SLOTS;
1331 
1332 	hlist_del(&agg->next);
1333 	if (hlist_empty(&grp->slots[i]))
1334 		__clear_bit(offset, &grp->full_slots);
1335 }
1336 
1337 /*
1338  * Called to forcibly deschedule an aggregate.  If the aggregate is
1339  * not in the front bucket, or if the latter has other aggregates in
1340  * the front bucket, we can simply remove the aggregate with no other
1341  * side effects.
1342  * Otherwise we must propagate the event up.
1343  */
1344 static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1345 {
1346 	struct qfq_group *grp = agg->grp;
1347 	unsigned long mask;
1348 	u64 roundedS;
1349 	int s;
1350 
1351 	if (agg == q->in_serv_agg) {
1352 		charge_actual_service(agg);
1353 		q->in_serv_agg = qfq_choose_next_agg(q);
1354 		return;
1355 	}
1356 
1357 	agg->F = agg->S;
1358 	qfq_slot_remove(q, grp, agg);
1359 
1360 	if (!grp->full_slots) {
1361 		__clear_bit(grp->index, &q->bitmaps[IR]);
1362 		__clear_bit(grp->index, &q->bitmaps[EB]);
1363 		__clear_bit(grp->index, &q->bitmaps[IB]);
1364 
1365 		if (test_bit(grp->index, &q->bitmaps[ER]) &&
1366 		    !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
1367 			mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
1368 			if (mask)
1369 				mask = ~((1UL << __fls(mask)) - 1);
1370 			else
1371 				mask = ~0UL;
1372 			qfq_move_groups(q, mask, EB, ER);
1373 			qfq_move_groups(q, mask, IB, IR);
1374 		}
1375 		__clear_bit(grp->index, &q->bitmaps[ER]);
1376 	} else if (hlist_empty(&grp->slots[grp->front])) {
1377 		agg = qfq_slot_scan(grp);
1378 		roundedS = qfq_round_down(agg->S, grp->slot_shift);
1379 		if (grp->S != roundedS) {
1380 			__clear_bit(grp->index, &q->bitmaps[ER]);
1381 			__clear_bit(grp->index, &q->bitmaps[IR]);
1382 			__clear_bit(grp->index, &q->bitmaps[EB]);
1383 			__clear_bit(grp->index, &q->bitmaps[IB]);
1384 			grp->S = roundedS;
1385 			grp->F = roundedS + (2ULL << grp->slot_shift);
1386 			s = qfq_calc_state(q, grp);
1387 			__set_bit(grp->index, &q->bitmaps[s]);
1388 		}
1389 	}
1390 }
1391 
1392 static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1393 {
1394 	struct qfq_sched *q = qdisc_priv(sch);
1395 	struct qfq_class *cl = (struct qfq_class *)arg;
1396 
1397 	qfq_deactivate_class(q, cl);
1398 }
1399 
1400 static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt,
1401 			  struct netlink_ext_ack *extack)
1402 {
1403 	struct qfq_sched *q = qdisc_priv(sch);
1404 	struct qfq_group *grp;
1405 	int i, j, err;
1406 	u32 max_cl_shift, maxbudg_shift, max_classes;
1407 
1408 	err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
1409 	if (err)
1410 		return err;
1411 
1412 	err = qdisc_class_hash_init(&q->clhash);
1413 	if (err < 0)
1414 		return err;
1415 
1416 	max_classes = min_t(u64, (u64)qdisc_dev(sch)->tx_queue_len + 1,
1417 			    QFQ_MAX_AGG_CLASSES);
1418 	/* max_cl_shift = floor(log_2(max_classes)) */
1419 	max_cl_shift = __fls(max_classes);
1420 	q->max_agg_classes = 1<<max_cl_shift;
1421 
1422 	/* maxbudg_shift = log2(max_len * max_classes_per_agg) */
1423 	maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift;
1424 	q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX;
1425 
1426 	for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1427 		grp = &q->groups[i];
1428 		grp->index = i;
1429 		grp->slot_shift = q->min_slot_shift + i;
1430 		for (j = 0; j < QFQ_MAX_SLOTS; j++)
1431 			INIT_HLIST_HEAD(&grp->slots[j]);
1432 	}
1433 
1434 	INIT_HLIST_HEAD(&q->nonfull_aggs);
1435 
1436 	return 0;
1437 }
1438 
1439 static void qfq_reset_qdisc(struct Qdisc *sch)
1440 {
1441 	struct qfq_sched *q = qdisc_priv(sch);
1442 	struct qfq_class *cl;
1443 	unsigned int i;
1444 
1445 	for (i = 0; i < q->clhash.hashsize; i++) {
1446 		hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1447 			if (cl->qdisc->q.qlen > 0)
1448 				qfq_deactivate_class(q, cl);
1449 
1450 			qdisc_reset(cl->qdisc);
1451 		}
1452 	}
1453 }
1454 
1455 static void qfq_destroy_qdisc(struct Qdisc *sch)
1456 {
1457 	struct qfq_sched *q = qdisc_priv(sch);
1458 	struct qfq_class *cl;
1459 	struct hlist_node *next;
1460 	unsigned int i;
1461 
1462 	tcf_block_put(q->block);
1463 
1464 	for (i = 0; i < q->clhash.hashsize; i++) {
1465 		hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1466 					  common.hnode) {
1467 			qfq_destroy_class(sch, cl);
1468 		}
1469 	}
1470 	qdisc_class_hash_destroy(&q->clhash);
1471 }
1472 
1473 static const struct Qdisc_class_ops qfq_class_ops = {
1474 	.change		= qfq_change_class,
1475 	.delete		= qfq_delete_class,
1476 	.find		= qfq_search_class,
1477 	.tcf_block	= qfq_tcf_block,
1478 	.bind_tcf	= qfq_bind_tcf,
1479 	.unbind_tcf	= qfq_unbind_tcf,
1480 	.graft		= qfq_graft_class,
1481 	.leaf		= qfq_class_leaf,
1482 	.qlen_notify	= qfq_qlen_notify,
1483 	.dump		= qfq_dump_class,
1484 	.dump_stats	= qfq_dump_class_stats,
1485 	.walk		= qfq_walk,
1486 };
1487 
1488 static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1489 	.cl_ops		= &qfq_class_ops,
1490 	.id		= "qfq",
1491 	.priv_size	= sizeof(struct qfq_sched),
1492 	.enqueue	= qfq_enqueue,
1493 	.dequeue	= qfq_dequeue,
1494 	.peek		= qdisc_peek_dequeued,
1495 	.init		= qfq_init_qdisc,
1496 	.reset		= qfq_reset_qdisc,
1497 	.destroy	= qfq_destroy_qdisc,
1498 	.owner		= THIS_MODULE,
1499 };
1500 
1501 static int __init qfq_init(void)
1502 {
1503 	return register_qdisc(&qfq_qdisc_ops);
1504 }
1505 
1506 static void __exit qfq_exit(void)
1507 {
1508 	unregister_qdisc(&qfq_qdisc_ops);
1509 }
1510 
1511 module_init(qfq_init);
1512 module_exit(qfq_exit);
1513 MODULE_LICENSE("GPL");
1514