xref: /linux/net/sched/sch_qfq.c (revision d257f9bf06129613de539ea71ecea60848b662cd)
1 /*
2  * net/sched/sch_qfq.c         Quick Fair Queueing Plus Scheduler.
3  *
4  * Copyright (c) 2009 Fabio Checconi, Luigi Rizzo, and Paolo Valente.
5  * Copyright (c) 2012 Paolo Valente.
6  *
7  * This program is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU General Public License
9  * version 2 as published by the Free Software Foundation.
10  */
11 
12 #include <linux/module.h>
13 #include <linux/init.h>
14 #include <linux/bitops.h>
15 #include <linux/errno.h>
16 #include <linux/netdevice.h>
17 #include <linux/pkt_sched.h>
18 #include <net/sch_generic.h>
19 #include <net/pkt_sched.h>
20 #include <net/pkt_cls.h>
21 
22 
23 /*  Quick Fair Queueing Plus
24     ========================
25 
26     Sources:
27 
28     [1] Paolo Valente,
29     "Reducing the Execution Time of Fair-Queueing Schedulers."
30     http://algo.ing.unimo.it/people/paolo/agg-sched/agg-sched.pdf
31 
32     Sources for QFQ:
33 
34     [2] Fabio Checconi, Luigi Rizzo, and Paolo Valente: "QFQ: Efficient
35     Packet Scheduling with Tight Bandwidth Distribution Guarantees."
36 
37     See also:
38     http://retis.sssup.it/~fabio/linux/qfq/
39  */
40 
41 /*
42 
43   QFQ+ divides classes into aggregates of at most MAX_AGG_CLASSES
44   classes. Each aggregate is timestamped with a virtual start time S
45   and a virtual finish time F, and scheduled according to its
46   timestamps. S and F are computed as a function of a system virtual
47   time function V. The classes within each aggregate are instead
48   scheduled with DRR.
49 
50   To speed up operations, QFQ+ divides also aggregates into a limited
51   number of groups. Which group a class belongs to depends on the
52   ratio between the maximum packet length for the class and the weight
53   of the class. Groups have their own S and F. In the end, QFQ+
54   schedules groups, then aggregates within groups, then classes within
55   aggregates. See [1] and [2] for a full description.
56 
57   Virtual time computations.
58 
59   S, F and V are all computed in fixed point arithmetic with
60   FRAC_BITS decimal bits.
61 
62   QFQ_MAX_INDEX is the maximum index allowed for a group. We need
63 	one bit per index.
64   QFQ_MAX_WSHIFT is the maximum power of two supported as a weight.
65 
66   The layout of the bits is as below:
67 
68                    [ MTU_SHIFT ][      FRAC_BITS    ]
69                    [ MAX_INDEX    ][ MIN_SLOT_SHIFT ]
70 				 ^.__grp->index = 0
71 				 *.__grp->slot_shift
72 
73   where MIN_SLOT_SHIFT is derived by difference from the others.
74 
75   The max group index corresponds to Lmax/w_min, where
76   Lmax=1<<MTU_SHIFT, w_min = 1 .
77   From this, and knowing how many groups (MAX_INDEX) we want,
78   we can derive the shift corresponding to each group.
79 
80   Because we often need to compute
81 	F = S + len/w_i  and V = V + len/wsum
82   instead of storing w_i store the value
83 	inv_w = (1<<FRAC_BITS)/w_i
84   so we can do F = S + len * inv_w * wsum.
85   We use W_TOT in the formulas so we can easily move between
86   static and adaptive weight sum.
87 
88   The per-scheduler-instance data contain all the data structures
89   for the scheduler: bitmaps and bucket lists.
90 
91  */
92 
93 /*
94  * Maximum number of consecutive slots occupied by backlogged classes
95  * inside a group.
96  */
97 #define QFQ_MAX_SLOTS	32
98 
99 /*
100  * Shifts used for aggregate<->group mapping.  We allow class weights that are
101  * in the range [1, 2^MAX_WSHIFT], and we try to map each aggregate i to the
102  * group with the smallest index that can support the L_i / r_i configured
103  * for the classes in the aggregate.
104  *
105  * grp->index is the index of the group; and grp->slot_shift
106  * is the shift for the corresponding (scaled) sigma_i.
107  */
108 #define QFQ_MAX_INDEX		24
109 #define QFQ_MAX_WSHIFT		10
110 
111 #define	QFQ_MAX_WEIGHT		(1<<QFQ_MAX_WSHIFT) /* see qfq_slot_insert */
112 #define QFQ_MAX_WSUM		(64*QFQ_MAX_WEIGHT)
113 
114 #define FRAC_BITS		30	/* fixed point arithmetic */
115 #define ONE_FP			(1UL << FRAC_BITS)
116 
117 #define QFQ_MTU_SHIFT		16	/* to support TSO/GSO */
118 #define QFQ_MIN_LMAX		512	/* see qfq_slot_insert */
119 
120 #define QFQ_MAX_AGG_CLASSES	8 /* max num classes per aggregate allowed */
121 
122 /*
123  * Possible group states.  These values are used as indexes for the bitmaps
124  * array of struct qfq_queue.
125  */
126 enum qfq_state { ER, IR, EB, IB, QFQ_MAX_STATE };
127 
128 struct qfq_group;
129 
130 struct qfq_aggregate;
131 
132 struct qfq_class {
133 	struct Qdisc_class_common common;
134 
135 	unsigned int refcnt;
136 	unsigned int filter_cnt;
137 
138 	struct gnet_stats_basic_packed bstats;
139 	struct gnet_stats_queue qstats;
140 	struct net_rate_estimator __rcu *rate_est;
141 	struct Qdisc *qdisc;
142 	struct list_head alist;		/* Link for active-classes list. */
143 	struct qfq_aggregate *agg;	/* Parent aggregate. */
144 	int deficit;			/* DRR deficit counter. */
145 };
146 
147 struct qfq_aggregate {
148 	struct hlist_node next;	/* Link for the slot list. */
149 	u64 S, F;		/* flow timestamps (exact) */
150 
151 	/* group we belong to. In principle we would need the index,
152 	 * which is log_2(lmax/weight), but we never reference it
153 	 * directly, only the group.
154 	 */
155 	struct qfq_group *grp;
156 
157 	/* these are copied from the flowset. */
158 	u32	class_weight; /* Weight of each class in this aggregate. */
159 	/* Max pkt size for the classes in this aggregate, DRR quantum. */
160 	int	lmax;
161 
162 	u32	inv_w;	    /* ONE_FP/(sum of weights of classes in aggr.). */
163 	u32	budgetmax;  /* Max budget for this aggregate. */
164 	u32	initial_budget, budget;     /* Initial and current budget. */
165 
166 	int		  num_classes;	/* Number of classes in this aggr. */
167 	struct list_head  active;	/* DRR queue of active classes. */
168 
169 	struct hlist_node nonfull_next;	/* See nonfull_aggs in qfq_sched. */
170 };
171 
172 struct qfq_group {
173 	u64 S, F;			/* group timestamps (approx). */
174 	unsigned int slot_shift;	/* Slot shift. */
175 	unsigned int index;		/* Group index. */
176 	unsigned int front;		/* Index of the front slot. */
177 	unsigned long full_slots;	/* non-empty slots */
178 
179 	/* Array of RR lists of active aggregates. */
180 	struct hlist_head slots[QFQ_MAX_SLOTS];
181 };
182 
183 struct qfq_sched {
184 	struct tcf_proto __rcu *filter_list;
185 	struct tcf_block	*block;
186 	struct Qdisc_class_hash clhash;
187 
188 	u64			oldV, V;	/* Precise virtual times. */
189 	struct qfq_aggregate	*in_serv_agg;   /* Aggregate being served. */
190 	u32			wsum;		/* weight sum */
191 	u32			iwsum;		/* inverse weight sum */
192 
193 	unsigned long bitmaps[QFQ_MAX_STATE];	    /* Group bitmaps. */
194 	struct qfq_group groups[QFQ_MAX_INDEX + 1]; /* The groups. */
195 	u32 min_slot_shift;	/* Index of the group-0 bit in the bitmaps. */
196 
197 	u32 max_agg_classes;		/* Max number of classes per aggr. */
198 	struct hlist_head nonfull_aggs; /* Aggs with room for more classes. */
199 };
200 
201 /*
202  * Possible reasons why the timestamps of an aggregate are updated
203  * enqueue: the aggregate switches from idle to active and must scheduled
204  *	    for service
205  * requeue: the aggregate finishes its budget, so it stops being served and
206  *	    must be rescheduled for service
207  */
208 enum update_reason {enqueue, requeue};
209 
210 static struct qfq_class *qfq_find_class(struct Qdisc *sch, u32 classid)
211 {
212 	struct qfq_sched *q = qdisc_priv(sch);
213 	struct Qdisc_class_common *clc;
214 
215 	clc = qdisc_class_find(&q->clhash, classid);
216 	if (clc == NULL)
217 		return NULL;
218 	return container_of(clc, struct qfq_class, common);
219 }
220 
221 static void qfq_purge_queue(struct qfq_class *cl)
222 {
223 	unsigned int len = cl->qdisc->q.qlen;
224 	unsigned int backlog = cl->qdisc->qstats.backlog;
225 
226 	qdisc_reset(cl->qdisc);
227 	qdisc_tree_reduce_backlog(cl->qdisc, len, backlog);
228 }
229 
230 static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = {
231 	[TCA_QFQ_WEIGHT] = { .type = NLA_U32 },
232 	[TCA_QFQ_LMAX] = { .type = NLA_U32 },
233 };
234 
235 /*
236  * Calculate a flow index, given its weight and maximum packet length.
237  * index = log_2(maxlen/weight) but we need to apply the scaling.
238  * This is used only once at flow creation.
239  */
240 static int qfq_calc_index(u32 inv_w, unsigned int maxlen, u32 min_slot_shift)
241 {
242 	u64 slot_size = (u64)maxlen * inv_w;
243 	unsigned long size_map;
244 	int index = 0;
245 
246 	size_map = slot_size >> min_slot_shift;
247 	if (!size_map)
248 		goto out;
249 
250 	index = __fls(size_map) + 1;	/* basically a log_2 */
251 	index -= !(slot_size - (1ULL << (index + min_slot_shift - 1)));
252 
253 	if (index < 0)
254 		index = 0;
255 out:
256 	pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n",
257 		 (unsigned long) ONE_FP/inv_w, maxlen, index);
258 
259 	return index;
260 }
261 
262 static void qfq_deactivate_agg(struct qfq_sched *, struct qfq_aggregate *);
263 static void qfq_activate_agg(struct qfq_sched *, struct qfq_aggregate *,
264 			     enum update_reason);
265 
266 static void qfq_init_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
267 			 u32 lmax, u32 weight)
268 {
269 	INIT_LIST_HEAD(&agg->active);
270 	hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
271 
272 	agg->lmax = lmax;
273 	agg->class_weight = weight;
274 }
275 
276 static struct qfq_aggregate *qfq_find_agg(struct qfq_sched *q,
277 					  u32 lmax, u32 weight)
278 {
279 	struct qfq_aggregate *agg;
280 
281 	hlist_for_each_entry(agg, &q->nonfull_aggs, nonfull_next)
282 		if (agg->lmax == lmax && agg->class_weight == weight)
283 			return agg;
284 
285 	return NULL;
286 }
287 
288 
289 /* Update aggregate as a function of the new number of classes. */
290 static void qfq_update_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
291 			   int new_num_classes)
292 {
293 	u32 new_agg_weight;
294 
295 	if (new_num_classes == q->max_agg_classes)
296 		hlist_del_init(&agg->nonfull_next);
297 
298 	if (agg->num_classes > new_num_classes &&
299 	    new_num_classes == q->max_agg_classes - 1) /* agg no more full */
300 		hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
301 
302 	/* The next assignment may let
303 	 * agg->initial_budget > agg->budgetmax
304 	 * hold, we will take it into account in charge_actual_service().
305 	 */
306 	agg->budgetmax = new_num_classes * agg->lmax;
307 	new_agg_weight = agg->class_weight * new_num_classes;
308 	agg->inv_w = ONE_FP/new_agg_weight;
309 
310 	if (agg->grp == NULL) {
311 		int i = qfq_calc_index(agg->inv_w, agg->budgetmax,
312 				       q->min_slot_shift);
313 		agg->grp = &q->groups[i];
314 	}
315 
316 	q->wsum +=
317 		(int) agg->class_weight * (new_num_classes - agg->num_classes);
318 	q->iwsum = ONE_FP / q->wsum;
319 
320 	agg->num_classes = new_num_classes;
321 }
322 
323 /* Add class to aggregate. */
324 static void qfq_add_to_agg(struct qfq_sched *q,
325 			   struct qfq_aggregate *agg,
326 			   struct qfq_class *cl)
327 {
328 	cl->agg = agg;
329 
330 	qfq_update_agg(q, agg, agg->num_classes+1);
331 	if (cl->qdisc->q.qlen > 0) { /* adding an active class */
332 		list_add_tail(&cl->alist, &agg->active);
333 		if (list_first_entry(&agg->active, struct qfq_class, alist) ==
334 		    cl && q->in_serv_agg != agg) /* agg was inactive */
335 			qfq_activate_agg(q, agg, enqueue); /* schedule agg */
336 	}
337 }
338 
339 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *);
340 
341 static void qfq_destroy_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
342 {
343 	hlist_del_init(&agg->nonfull_next);
344 	q->wsum -= agg->class_weight;
345 	if (q->wsum != 0)
346 		q->iwsum = ONE_FP / q->wsum;
347 
348 	if (q->in_serv_agg == agg)
349 		q->in_serv_agg = qfq_choose_next_agg(q);
350 	kfree(agg);
351 }
352 
353 /* Deschedule class from within its parent aggregate. */
354 static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
355 {
356 	struct qfq_aggregate *agg = cl->agg;
357 
358 
359 	list_del(&cl->alist); /* remove from RR queue of the aggregate */
360 	if (list_empty(&agg->active)) /* agg is now inactive */
361 		qfq_deactivate_agg(q, agg);
362 }
363 
364 /* Remove class from its parent aggregate. */
365 static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
366 {
367 	struct qfq_aggregate *agg = cl->agg;
368 
369 	cl->agg = NULL;
370 	if (agg->num_classes == 1) { /* agg being emptied, destroy it */
371 		qfq_destroy_agg(q, agg);
372 		return;
373 	}
374 	qfq_update_agg(q, agg, agg->num_classes-1);
375 }
376 
377 /* Deschedule class and remove it from its parent aggregate. */
378 static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
379 {
380 	if (cl->qdisc->q.qlen > 0) /* class is active */
381 		qfq_deactivate_class(q, cl);
382 
383 	qfq_rm_from_agg(q, cl);
384 }
385 
386 /* Move class to a new aggregate, matching the new class weight and/or lmax */
387 static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight,
388 			   u32 lmax)
389 {
390 	struct qfq_sched *q = qdisc_priv(sch);
391 	struct qfq_aggregate *new_agg = qfq_find_agg(q, lmax, weight);
392 
393 	if (new_agg == NULL) { /* create new aggregate */
394 		new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC);
395 		if (new_agg == NULL)
396 			return -ENOBUFS;
397 		qfq_init_agg(q, new_agg, lmax, weight);
398 	}
399 	qfq_deact_rm_from_agg(q, cl);
400 	qfq_add_to_agg(q, new_agg, cl);
401 
402 	return 0;
403 }
404 
405 static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
406 			    struct nlattr **tca, unsigned long *arg)
407 {
408 	struct qfq_sched *q = qdisc_priv(sch);
409 	struct qfq_class *cl = (struct qfq_class *)*arg;
410 	bool existing = false;
411 	struct nlattr *tb[TCA_QFQ_MAX + 1];
412 	struct qfq_aggregate *new_agg = NULL;
413 	u32 weight, lmax, inv_w;
414 	int err;
415 	int delta_w;
416 
417 	if (tca[TCA_OPTIONS] == NULL) {
418 		pr_notice("qfq: no options\n");
419 		return -EINVAL;
420 	}
421 
422 	err = nla_parse_nested(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS], qfq_policy,
423 			       NULL);
424 	if (err < 0)
425 		return err;
426 
427 	if (tb[TCA_QFQ_WEIGHT]) {
428 		weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]);
429 		if (!weight || weight > (1UL << QFQ_MAX_WSHIFT)) {
430 			pr_notice("qfq: invalid weight %u\n", weight);
431 			return -EINVAL;
432 		}
433 	} else
434 		weight = 1;
435 
436 	if (tb[TCA_QFQ_LMAX]) {
437 		lmax = nla_get_u32(tb[TCA_QFQ_LMAX]);
438 		if (lmax < QFQ_MIN_LMAX || lmax > (1UL << QFQ_MTU_SHIFT)) {
439 			pr_notice("qfq: invalid max length %u\n", lmax);
440 			return -EINVAL;
441 		}
442 	} else
443 		lmax = psched_mtu(qdisc_dev(sch));
444 
445 	inv_w = ONE_FP / weight;
446 	weight = ONE_FP / inv_w;
447 
448 	if (cl != NULL &&
449 	    lmax == cl->agg->lmax &&
450 	    weight == cl->agg->class_weight)
451 		return 0; /* nothing to change */
452 
453 	delta_w = weight - (cl ? cl->agg->class_weight : 0);
454 
455 	if (q->wsum + delta_w > QFQ_MAX_WSUM) {
456 		pr_notice("qfq: total weight out of range (%d + %u)\n",
457 			  delta_w, q->wsum);
458 		return -EINVAL;
459 	}
460 
461 	if (cl != NULL) { /* modify existing class */
462 		if (tca[TCA_RATE]) {
463 			err = gen_replace_estimator(&cl->bstats, NULL,
464 						    &cl->rate_est,
465 						    NULL,
466 						    qdisc_root_sleeping_running(sch),
467 						    tca[TCA_RATE]);
468 			if (err)
469 				return err;
470 		}
471 		existing = true;
472 		goto set_change_agg;
473 	}
474 
475 	/* create and init new class */
476 	cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
477 	if (cl == NULL)
478 		return -ENOBUFS;
479 
480 	cl->refcnt = 1;
481 	cl->common.classid = classid;
482 	cl->deficit = lmax;
483 
484 	cl->qdisc = qdisc_create_dflt(sch->dev_queue,
485 				      &pfifo_qdisc_ops, classid);
486 	if (cl->qdisc == NULL)
487 		cl->qdisc = &noop_qdisc;
488 
489 	if (tca[TCA_RATE]) {
490 		err = gen_new_estimator(&cl->bstats, NULL,
491 					&cl->rate_est,
492 					NULL,
493 					qdisc_root_sleeping_running(sch),
494 					tca[TCA_RATE]);
495 		if (err)
496 			goto destroy_class;
497 	}
498 
499 	if (cl->qdisc != &noop_qdisc)
500 		qdisc_hash_add(cl->qdisc, true);
501 	sch_tree_lock(sch);
502 	qdisc_class_hash_insert(&q->clhash, &cl->common);
503 	sch_tree_unlock(sch);
504 
505 	qdisc_class_hash_grow(sch, &q->clhash);
506 
507 set_change_agg:
508 	sch_tree_lock(sch);
509 	new_agg = qfq_find_agg(q, lmax, weight);
510 	if (new_agg == NULL) { /* create new aggregate */
511 		sch_tree_unlock(sch);
512 		new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL);
513 		if (new_agg == NULL) {
514 			err = -ENOBUFS;
515 			gen_kill_estimator(&cl->rate_est);
516 			goto destroy_class;
517 		}
518 		sch_tree_lock(sch);
519 		qfq_init_agg(q, new_agg, lmax, weight);
520 	}
521 	if (existing)
522 		qfq_deact_rm_from_agg(q, cl);
523 	qfq_add_to_agg(q, new_agg, cl);
524 	sch_tree_unlock(sch);
525 
526 	*arg = (unsigned long)cl;
527 	return 0;
528 
529 destroy_class:
530 	qdisc_destroy(cl->qdisc);
531 	kfree(cl);
532 	return err;
533 }
534 
535 static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
536 {
537 	struct qfq_sched *q = qdisc_priv(sch);
538 
539 	qfq_rm_from_agg(q, cl);
540 	gen_kill_estimator(&cl->rate_est);
541 	qdisc_destroy(cl->qdisc);
542 	kfree(cl);
543 }
544 
545 static int qfq_delete_class(struct Qdisc *sch, unsigned long arg)
546 {
547 	struct qfq_sched *q = qdisc_priv(sch);
548 	struct qfq_class *cl = (struct qfq_class *)arg;
549 
550 	if (cl->filter_cnt > 0)
551 		return -EBUSY;
552 
553 	sch_tree_lock(sch);
554 
555 	qfq_purge_queue(cl);
556 	qdisc_class_hash_remove(&q->clhash, &cl->common);
557 
558 	BUG_ON(--cl->refcnt == 0);
559 	/*
560 	 * This shouldn't happen: we "hold" one cops->get() when called
561 	 * from tc_ctl_tclass; the destroy method is done from cops->put().
562 	 */
563 
564 	sch_tree_unlock(sch);
565 	return 0;
566 }
567 
568 static unsigned long qfq_get_class(struct Qdisc *sch, u32 classid)
569 {
570 	struct qfq_class *cl = qfq_find_class(sch, classid);
571 
572 	if (cl != NULL)
573 		cl->refcnt++;
574 
575 	return (unsigned long)cl;
576 }
577 
578 static void qfq_put_class(struct Qdisc *sch, unsigned long arg)
579 {
580 	struct qfq_class *cl = (struct qfq_class *)arg;
581 
582 	if (--cl->refcnt == 0)
583 		qfq_destroy_class(sch, cl);
584 }
585 
586 static struct tcf_block *qfq_tcf_block(struct Qdisc *sch, unsigned long cl)
587 {
588 	struct qfq_sched *q = qdisc_priv(sch);
589 
590 	if (cl)
591 		return NULL;
592 
593 	return q->block;
594 }
595 
596 static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
597 				  u32 classid)
598 {
599 	struct qfq_class *cl = qfq_find_class(sch, classid);
600 
601 	if (cl != NULL)
602 		cl->filter_cnt++;
603 
604 	return (unsigned long)cl;
605 }
606 
607 static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
608 {
609 	struct qfq_class *cl = (struct qfq_class *)arg;
610 
611 	cl->filter_cnt--;
612 }
613 
614 static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
615 			   struct Qdisc *new, struct Qdisc **old)
616 {
617 	struct qfq_class *cl = (struct qfq_class *)arg;
618 
619 	if (new == NULL) {
620 		new = qdisc_create_dflt(sch->dev_queue,
621 					&pfifo_qdisc_ops, cl->common.classid);
622 		if (new == NULL)
623 			new = &noop_qdisc;
624 	}
625 
626 	*old = qdisc_replace(sch, new, &cl->qdisc);
627 	return 0;
628 }
629 
630 static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
631 {
632 	struct qfq_class *cl = (struct qfq_class *)arg;
633 
634 	return cl->qdisc;
635 }
636 
637 static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
638 			  struct sk_buff *skb, struct tcmsg *tcm)
639 {
640 	struct qfq_class *cl = (struct qfq_class *)arg;
641 	struct nlattr *nest;
642 
643 	tcm->tcm_parent	= TC_H_ROOT;
644 	tcm->tcm_handle	= cl->common.classid;
645 	tcm->tcm_info	= cl->qdisc->handle;
646 
647 	nest = nla_nest_start(skb, TCA_OPTIONS);
648 	if (nest == NULL)
649 		goto nla_put_failure;
650 	if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) ||
651 	    nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax))
652 		goto nla_put_failure;
653 	return nla_nest_end(skb, nest);
654 
655 nla_put_failure:
656 	nla_nest_cancel(skb, nest);
657 	return -EMSGSIZE;
658 }
659 
660 static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
661 				struct gnet_dump *d)
662 {
663 	struct qfq_class *cl = (struct qfq_class *)arg;
664 	struct tc_qfq_stats xstats;
665 
666 	memset(&xstats, 0, sizeof(xstats));
667 
668 	xstats.weight = cl->agg->class_weight;
669 	xstats.lmax = cl->agg->lmax;
670 
671 	if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
672 				  d, NULL, &cl->bstats) < 0 ||
673 	    gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
674 	    gnet_stats_copy_queue(d, NULL,
675 				  &cl->qdisc->qstats, cl->qdisc->q.qlen) < 0)
676 		return -1;
677 
678 	return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
679 }
680 
681 static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
682 {
683 	struct qfq_sched *q = qdisc_priv(sch);
684 	struct qfq_class *cl;
685 	unsigned int i;
686 
687 	if (arg->stop)
688 		return;
689 
690 	for (i = 0; i < q->clhash.hashsize; i++) {
691 		hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
692 			if (arg->count < arg->skip) {
693 				arg->count++;
694 				continue;
695 			}
696 			if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
697 				arg->stop = 1;
698 				return;
699 			}
700 			arg->count++;
701 		}
702 	}
703 }
704 
705 static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
706 				      int *qerr)
707 {
708 	struct qfq_sched *q = qdisc_priv(sch);
709 	struct qfq_class *cl;
710 	struct tcf_result res;
711 	struct tcf_proto *fl;
712 	int result;
713 
714 	if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
715 		pr_debug("qfq_classify: found %d\n", skb->priority);
716 		cl = qfq_find_class(sch, skb->priority);
717 		if (cl != NULL)
718 			return cl;
719 	}
720 
721 	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
722 	fl = rcu_dereference_bh(q->filter_list);
723 	result = tcf_classify(skb, fl, &res, false);
724 	if (result >= 0) {
725 #ifdef CONFIG_NET_CLS_ACT
726 		switch (result) {
727 		case TC_ACT_QUEUED:
728 		case TC_ACT_STOLEN:
729 		case TC_ACT_TRAP:
730 			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
731 		case TC_ACT_SHOT:
732 			return NULL;
733 		}
734 #endif
735 		cl = (struct qfq_class *)res.class;
736 		if (cl == NULL)
737 			cl = qfq_find_class(sch, res.classid);
738 		return cl;
739 	}
740 
741 	return NULL;
742 }
743 
744 /* Generic comparison function, handling wraparound. */
745 static inline int qfq_gt(u64 a, u64 b)
746 {
747 	return (s64)(a - b) > 0;
748 }
749 
750 /* Round a precise timestamp to its slotted value. */
751 static inline u64 qfq_round_down(u64 ts, unsigned int shift)
752 {
753 	return ts & ~((1ULL << shift) - 1);
754 }
755 
756 /* return the pointer to the group with lowest index in the bitmap */
757 static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
758 					unsigned long bitmap)
759 {
760 	int index = __ffs(bitmap);
761 	return &q->groups[index];
762 }
763 /* Calculate a mask to mimic what would be ffs_from(). */
764 static inline unsigned long mask_from(unsigned long bitmap, int from)
765 {
766 	return bitmap & ~((1UL << from) - 1);
767 }
768 
769 /*
770  * The state computation relies on ER=0, IR=1, EB=2, IB=3
771  * First compute eligibility comparing grp->S, q->V,
772  * then check if someone is blocking us and possibly add EB
773  */
774 static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
775 {
776 	/* if S > V we are not eligible */
777 	unsigned int state = qfq_gt(grp->S, q->V);
778 	unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
779 	struct qfq_group *next;
780 
781 	if (mask) {
782 		next = qfq_ffs(q, mask);
783 		if (qfq_gt(grp->F, next->F))
784 			state |= EB;
785 	}
786 
787 	return state;
788 }
789 
790 
791 /*
792  * In principle
793  *	q->bitmaps[dst] |= q->bitmaps[src] & mask;
794  *	q->bitmaps[src] &= ~mask;
795  * but we should make sure that src != dst
796  */
797 static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
798 				   int src, int dst)
799 {
800 	q->bitmaps[dst] |= q->bitmaps[src] & mask;
801 	q->bitmaps[src] &= ~mask;
802 }
803 
804 static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
805 {
806 	unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
807 	struct qfq_group *next;
808 
809 	if (mask) {
810 		next = qfq_ffs(q, mask);
811 		if (!qfq_gt(next->F, old_F))
812 			return;
813 	}
814 
815 	mask = (1UL << index) - 1;
816 	qfq_move_groups(q, mask, EB, ER);
817 	qfq_move_groups(q, mask, IB, IR);
818 }
819 
820 /*
821  * perhaps
822  *
823 	old_V ^= q->V;
824 	old_V >>= q->min_slot_shift;
825 	if (old_V) {
826 		...
827 	}
828  *
829  */
830 static void qfq_make_eligible(struct qfq_sched *q)
831 {
832 	unsigned long vslot = q->V >> q->min_slot_shift;
833 	unsigned long old_vslot = q->oldV >> q->min_slot_shift;
834 
835 	if (vslot != old_vslot) {
836 		unsigned long mask;
837 		int last_flip_pos = fls(vslot ^ old_vslot);
838 
839 		if (last_flip_pos > 31) /* higher than the number of groups */
840 			mask = ~0UL;    /* make all groups eligible */
841 		else
842 			mask = (1UL << last_flip_pos) - 1;
843 
844 		qfq_move_groups(q, mask, IR, ER);
845 		qfq_move_groups(q, mask, IB, EB);
846 	}
847 }
848 
849 /*
850  * The index of the slot in which the input aggregate agg is to be
851  * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2'
852  * and not a '-1' because the start time of the group may be moved
853  * backward by one slot after the aggregate has been inserted, and
854  * this would cause non-empty slots to be right-shifted by one
855  * position.
856  *
857  * QFQ+ fully satisfies this bound to the slot index if the parameters
858  * of the classes are not changed dynamically, and if QFQ+ never
859  * happens to postpone the service of agg unjustly, i.e., it never
860  * happens that the aggregate becomes backlogged and eligible, or just
861  * eligible, while an aggregate with a higher approximated finish time
862  * is being served. In particular, in this case QFQ+ guarantees that
863  * the timestamps of agg are low enough that the slot index is never
864  * higher than 2. Unfortunately, QFQ+ cannot provide the same
865  * guarantee if it happens to unjustly postpone the service of agg, or
866  * if the parameters of some class are changed.
867  *
868  * As for the first event, i.e., an out-of-order service, the
869  * upper bound to the slot index guaranteed by QFQ+ grows to
870  * 2 +
871  * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
872  * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1.
873  *
874  * The following function deals with this problem by backward-shifting
875  * the timestamps of agg, if needed, so as to guarantee that the slot
876  * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may
877  * cause the service of other aggregates to be postponed, yet the
878  * worst-case guarantees of these aggregates are not violated.  In
879  * fact, in case of no out-of-order service, the timestamps of agg
880  * would have been even lower than they are after the backward shift,
881  * because QFQ+ would have guaranteed a maximum value equal to 2 for
882  * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose
883  * service is postponed because of the backward-shift would have
884  * however waited for the service of agg before being served.
885  *
886  * The other event that may cause the slot index to be higher than 2
887  * for agg is a recent change of the parameters of some class. If the
888  * weight of a class is increased or the lmax (max_pkt_size) of the
889  * class is decreased, then a new aggregate with smaller slot size
890  * than the original parent aggregate of the class may happen to be
891  * activated. The activation of this aggregate should be properly
892  * delayed to when the service of the class has finished in the ideal
893  * system tracked by QFQ+. If the activation of the aggregate is not
894  * delayed to this reference time instant, then this aggregate may be
895  * unjustly served before other aggregates waiting for service. This
896  * may cause the above bound to the slot index to be violated for some
897  * of these unlucky aggregates.
898  *
899  * Instead of delaying the activation of the new aggregate, which is
900  * quite complex, the above-discussed capping of the slot index is
901  * used to handle also the consequences of a change of the parameters
902  * of a class.
903  */
904 static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg,
905 			    u64 roundedS)
906 {
907 	u64 slot = (roundedS - grp->S) >> grp->slot_shift;
908 	unsigned int i; /* slot index in the bucket list */
909 
910 	if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
911 		u64 deltaS = roundedS - grp->S -
912 			((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
913 		agg->S -= deltaS;
914 		agg->F -= deltaS;
915 		slot = QFQ_MAX_SLOTS - 2;
916 	}
917 
918 	i = (grp->front + slot) % QFQ_MAX_SLOTS;
919 
920 	hlist_add_head(&agg->next, &grp->slots[i]);
921 	__set_bit(slot, &grp->full_slots);
922 }
923 
924 /* Maybe introduce hlist_first_entry?? */
925 static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp)
926 {
927 	return hlist_entry(grp->slots[grp->front].first,
928 			   struct qfq_aggregate, next);
929 }
930 
931 /*
932  * remove the entry from the slot
933  */
934 static void qfq_front_slot_remove(struct qfq_group *grp)
935 {
936 	struct qfq_aggregate *agg = qfq_slot_head(grp);
937 
938 	BUG_ON(!agg);
939 	hlist_del(&agg->next);
940 	if (hlist_empty(&grp->slots[grp->front]))
941 		__clear_bit(0, &grp->full_slots);
942 }
943 
944 /*
945  * Returns the first aggregate in the first non-empty bucket of the
946  * group. As a side effect, adjusts the bucket list so the first
947  * non-empty bucket is at position 0 in full_slots.
948  */
949 static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp)
950 {
951 	unsigned int i;
952 
953 	pr_debug("qfq slot_scan: grp %u full %#lx\n",
954 		 grp->index, grp->full_slots);
955 
956 	if (grp->full_slots == 0)
957 		return NULL;
958 
959 	i = __ffs(grp->full_slots);  /* zero based */
960 	if (i > 0) {
961 		grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
962 		grp->full_slots >>= i;
963 	}
964 
965 	return qfq_slot_head(grp);
966 }
967 
968 /*
969  * adjust the bucket list. When the start time of a group decreases,
970  * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
971  * move the objects. The mask of occupied slots must be shifted
972  * because we use ffs() to find the first non-empty slot.
973  * This covers decreases in the group's start time, but what about
974  * increases of the start time ?
975  * Here too we should make sure that i is less than 32
976  */
977 static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
978 {
979 	unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
980 
981 	grp->full_slots <<= i;
982 	grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
983 }
984 
985 static void qfq_update_eligible(struct qfq_sched *q)
986 {
987 	struct qfq_group *grp;
988 	unsigned long ineligible;
989 
990 	ineligible = q->bitmaps[IR] | q->bitmaps[IB];
991 	if (ineligible) {
992 		if (!q->bitmaps[ER]) {
993 			grp = qfq_ffs(q, ineligible);
994 			if (qfq_gt(grp->S, q->V))
995 				q->V = grp->S;
996 		}
997 		qfq_make_eligible(q);
998 	}
999 }
1000 
1001 /* Dequeue head packet of the head class in the DRR queue of the aggregate. */
1002 static void agg_dequeue(struct qfq_aggregate *agg,
1003 			struct qfq_class *cl, unsigned int len)
1004 {
1005 	qdisc_dequeue_peeked(cl->qdisc);
1006 
1007 	cl->deficit -= (int) len;
1008 
1009 	if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */
1010 		list_del(&cl->alist);
1011 	else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) {
1012 		cl->deficit += agg->lmax;
1013 		list_move_tail(&cl->alist, &agg->active);
1014 	}
1015 }
1016 
1017 static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg,
1018 					   struct qfq_class **cl,
1019 					   unsigned int *len)
1020 {
1021 	struct sk_buff *skb;
1022 
1023 	*cl = list_first_entry(&agg->active, struct qfq_class, alist);
1024 	skb = (*cl)->qdisc->ops->peek((*cl)->qdisc);
1025 	if (skb == NULL)
1026 		WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
1027 	else
1028 		*len = qdisc_pkt_len(skb);
1029 
1030 	return skb;
1031 }
1032 
1033 /* Update F according to the actual service received by the aggregate. */
1034 static inline void charge_actual_service(struct qfq_aggregate *agg)
1035 {
1036 	/* Compute the service received by the aggregate, taking into
1037 	 * account that, after decreasing the number of classes in
1038 	 * agg, it may happen that
1039 	 * agg->initial_budget - agg->budget > agg->bugdetmax
1040 	 */
1041 	u32 service_received = min(agg->budgetmax,
1042 				   agg->initial_budget - agg->budget);
1043 
1044 	agg->F = agg->S + (u64)service_received * agg->inv_w;
1045 }
1046 
1047 /* Assign a reasonable start time for a new aggregate in group i.
1048  * Admissible values for \hat(F) are multiples of \sigma_i
1049  * no greater than V+\sigma_i . Larger values mean that
1050  * we had a wraparound so we consider the timestamp to be stale.
1051  *
1052  * If F is not stale and F >= V then we set S = F.
1053  * Otherwise we should assign S = V, but this may violate
1054  * the ordering in EB (see [2]). So, if we have groups in ER,
1055  * set S to the F_j of the first group j which would be blocking us.
1056  * We are guaranteed not to move S backward because
1057  * otherwise our group i would still be blocked.
1058  */
1059 static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg)
1060 {
1061 	unsigned long mask;
1062 	u64 limit, roundedF;
1063 	int slot_shift = agg->grp->slot_shift;
1064 
1065 	roundedF = qfq_round_down(agg->F, slot_shift);
1066 	limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
1067 
1068 	if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) {
1069 		/* timestamp was stale */
1070 		mask = mask_from(q->bitmaps[ER], agg->grp->index);
1071 		if (mask) {
1072 			struct qfq_group *next = qfq_ffs(q, mask);
1073 			if (qfq_gt(roundedF, next->F)) {
1074 				if (qfq_gt(limit, next->F))
1075 					agg->S = next->F;
1076 				else /* preserve timestamp correctness */
1077 					agg->S = limit;
1078 				return;
1079 			}
1080 		}
1081 		agg->S = q->V;
1082 	} else  /* timestamp is not stale */
1083 		agg->S = agg->F;
1084 }
1085 
1086 /* Update the timestamps of agg before scheduling/rescheduling it for
1087  * service.  In particular, assign to agg->F its maximum possible
1088  * value, i.e., the virtual finish time with which the aggregate
1089  * should be labeled if it used all its budget once in service.
1090  */
1091 static inline void
1092 qfq_update_agg_ts(struct qfq_sched *q,
1093 		    struct qfq_aggregate *agg, enum update_reason reason)
1094 {
1095 	if (reason != requeue)
1096 		qfq_update_start(q, agg);
1097 	else /* just charge agg for the service received */
1098 		agg->S = agg->F;
1099 
1100 	agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w;
1101 }
1102 
1103 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg);
1104 
1105 static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
1106 {
1107 	struct qfq_sched *q = qdisc_priv(sch);
1108 	struct qfq_aggregate *in_serv_agg = q->in_serv_agg;
1109 	struct qfq_class *cl;
1110 	struct sk_buff *skb = NULL;
1111 	/* next-packet len, 0 means no more active classes in in-service agg */
1112 	unsigned int len = 0;
1113 
1114 	if (in_serv_agg == NULL)
1115 		return NULL;
1116 
1117 	if (!list_empty(&in_serv_agg->active))
1118 		skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1119 
1120 	/*
1121 	 * If there are no active classes in the in-service aggregate,
1122 	 * or if the aggregate has not enough budget to serve its next
1123 	 * class, then choose the next aggregate to serve.
1124 	 */
1125 	if (len == 0 || in_serv_agg->budget < len) {
1126 		charge_actual_service(in_serv_agg);
1127 
1128 		/* recharge the budget of the aggregate */
1129 		in_serv_agg->initial_budget = in_serv_agg->budget =
1130 			in_serv_agg->budgetmax;
1131 
1132 		if (!list_empty(&in_serv_agg->active)) {
1133 			/*
1134 			 * Still active: reschedule for
1135 			 * service. Possible optimization: if no other
1136 			 * aggregate is active, then there is no point
1137 			 * in rescheduling this aggregate, and we can
1138 			 * just keep it as the in-service one. This
1139 			 * should be however a corner case, and to
1140 			 * handle it, we would need to maintain an
1141 			 * extra num_active_aggs field.
1142 			*/
1143 			qfq_update_agg_ts(q, in_serv_agg, requeue);
1144 			qfq_schedule_agg(q, in_serv_agg);
1145 		} else if (sch->q.qlen == 0) { /* no aggregate to serve */
1146 			q->in_serv_agg = NULL;
1147 			return NULL;
1148 		}
1149 
1150 		/*
1151 		 * If we get here, there are other aggregates queued:
1152 		 * choose the new aggregate to serve.
1153 		 */
1154 		in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
1155 		skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1156 	}
1157 	if (!skb)
1158 		return NULL;
1159 
1160 	qdisc_qstats_backlog_dec(sch, skb);
1161 	sch->q.qlen--;
1162 	qdisc_bstats_update(sch, skb);
1163 
1164 	agg_dequeue(in_serv_agg, cl, len);
1165 	/* If lmax is lowered, through qfq_change_class, for a class
1166 	 * owning pending packets with larger size than the new value
1167 	 * of lmax, then the following condition may hold.
1168 	 */
1169 	if (unlikely(in_serv_agg->budget < len))
1170 		in_serv_agg->budget = 0;
1171 	else
1172 		in_serv_agg->budget -= len;
1173 
1174 	q->V += (u64)len * q->iwsum;
1175 	pr_debug("qfq dequeue: len %u F %lld now %lld\n",
1176 		 len, (unsigned long long) in_serv_agg->F,
1177 		 (unsigned long long) q->V);
1178 
1179 	return skb;
1180 }
1181 
1182 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)
1183 {
1184 	struct qfq_group *grp;
1185 	struct qfq_aggregate *agg, *new_front_agg;
1186 	u64 old_F;
1187 
1188 	qfq_update_eligible(q);
1189 	q->oldV = q->V;
1190 
1191 	if (!q->bitmaps[ER])
1192 		return NULL;
1193 
1194 	grp = qfq_ffs(q, q->bitmaps[ER]);
1195 	old_F = grp->F;
1196 
1197 	agg = qfq_slot_head(grp);
1198 
1199 	/* agg starts to be served, remove it from schedule */
1200 	qfq_front_slot_remove(grp);
1201 
1202 	new_front_agg = qfq_slot_scan(grp);
1203 
1204 	if (new_front_agg == NULL) /* group is now inactive, remove from ER */
1205 		__clear_bit(grp->index, &q->bitmaps[ER]);
1206 	else {
1207 		u64 roundedS = qfq_round_down(new_front_agg->S,
1208 					      grp->slot_shift);
1209 		unsigned int s;
1210 
1211 		if (grp->S == roundedS)
1212 			return agg;
1213 		grp->S = roundedS;
1214 		grp->F = roundedS + (2ULL << grp->slot_shift);
1215 		__clear_bit(grp->index, &q->bitmaps[ER]);
1216 		s = qfq_calc_state(q, grp);
1217 		__set_bit(grp->index, &q->bitmaps[s]);
1218 	}
1219 
1220 	qfq_unblock_groups(q, grp->index, old_F);
1221 
1222 	return agg;
1223 }
1224 
1225 static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
1226 		       struct sk_buff **to_free)
1227 {
1228 	struct qfq_sched *q = qdisc_priv(sch);
1229 	struct qfq_class *cl;
1230 	struct qfq_aggregate *agg;
1231 	int err = 0;
1232 
1233 	cl = qfq_classify(skb, sch, &err);
1234 	if (cl == NULL) {
1235 		if (err & __NET_XMIT_BYPASS)
1236 			qdisc_qstats_drop(sch);
1237 		kfree_skb(skb);
1238 		return err;
1239 	}
1240 	pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
1241 
1242 	if (unlikely(cl->agg->lmax < qdisc_pkt_len(skb))) {
1243 		pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
1244 			 cl->agg->lmax, qdisc_pkt_len(skb), cl->common.classid);
1245 		err = qfq_change_agg(sch, cl, cl->agg->class_weight,
1246 				     qdisc_pkt_len(skb));
1247 		if (err) {
1248 			cl->qstats.drops++;
1249 			return qdisc_drop(skb, sch, to_free);
1250 		}
1251 	}
1252 
1253 	err = qdisc_enqueue(skb, cl->qdisc, to_free);
1254 	if (unlikely(err != NET_XMIT_SUCCESS)) {
1255 		pr_debug("qfq_enqueue: enqueue failed %d\n", err);
1256 		if (net_xmit_drop_count(err)) {
1257 			cl->qstats.drops++;
1258 			qdisc_qstats_drop(sch);
1259 		}
1260 		return err;
1261 	}
1262 
1263 	bstats_update(&cl->bstats, skb);
1264 	qdisc_qstats_backlog_inc(sch, skb);
1265 	++sch->q.qlen;
1266 
1267 	agg = cl->agg;
1268 	/* if the queue was not empty, then done here */
1269 	if (cl->qdisc->q.qlen != 1) {
1270 		if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) &&
1271 		    list_first_entry(&agg->active, struct qfq_class, alist)
1272 		    == cl && cl->deficit < qdisc_pkt_len(skb))
1273 			list_move_tail(&cl->alist, &agg->active);
1274 
1275 		return err;
1276 	}
1277 
1278 	/* schedule class for service within the aggregate */
1279 	cl->deficit = agg->lmax;
1280 	list_add_tail(&cl->alist, &agg->active);
1281 
1282 	if (list_first_entry(&agg->active, struct qfq_class, alist) != cl ||
1283 	    q->in_serv_agg == agg)
1284 		return err; /* non-empty or in service, nothing else to do */
1285 
1286 	qfq_activate_agg(q, agg, enqueue);
1287 
1288 	return err;
1289 }
1290 
1291 /*
1292  * Schedule aggregate according to its timestamps.
1293  */
1294 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1295 {
1296 	struct qfq_group *grp = agg->grp;
1297 	u64 roundedS;
1298 	int s;
1299 
1300 	roundedS = qfq_round_down(agg->S, grp->slot_shift);
1301 
1302 	/*
1303 	 * Insert agg in the correct bucket.
1304 	 * If agg->S >= grp->S we don't need to adjust the
1305 	 * bucket list and simply go to the insertion phase.
1306 	 * Otherwise grp->S is decreasing, we must make room
1307 	 * in the bucket list, and also recompute the group state.
1308 	 * Finally, if there were no flows in this group and nobody
1309 	 * was in ER make sure to adjust V.
1310 	 */
1311 	if (grp->full_slots) {
1312 		if (!qfq_gt(grp->S, agg->S))
1313 			goto skip_update;
1314 
1315 		/* create a slot for this agg->S */
1316 		qfq_slot_rotate(grp, roundedS);
1317 		/* group was surely ineligible, remove */
1318 		__clear_bit(grp->index, &q->bitmaps[IR]);
1319 		__clear_bit(grp->index, &q->bitmaps[IB]);
1320 	} else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) &&
1321 		   q->in_serv_agg == NULL)
1322 		q->V = roundedS;
1323 
1324 	grp->S = roundedS;
1325 	grp->F = roundedS + (2ULL << grp->slot_shift);
1326 	s = qfq_calc_state(q, grp);
1327 	__set_bit(grp->index, &q->bitmaps[s]);
1328 
1329 	pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
1330 		 s, q->bitmaps[s],
1331 		 (unsigned long long) agg->S,
1332 		 (unsigned long long) agg->F,
1333 		 (unsigned long long) q->V);
1334 
1335 skip_update:
1336 	qfq_slot_insert(grp, agg, roundedS);
1337 }
1338 
1339 
1340 /* Update agg ts and schedule agg for service */
1341 static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
1342 			     enum update_reason reason)
1343 {
1344 	agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */
1345 
1346 	qfq_update_agg_ts(q, agg, reason);
1347 	if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */
1348 		q->in_serv_agg = agg; /* start serving this aggregate */
1349 		 /* update V: to be in service, agg must be eligible */
1350 		q->oldV = q->V = agg->S;
1351 	} else if (agg != q->in_serv_agg)
1352 		qfq_schedule_agg(q, agg);
1353 }
1354 
1355 static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
1356 			    struct qfq_aggregate *agg)
1357 {
1358 	unsigned int i, offset;
1359 	u64 roundedS;
1360 
1361 	roundedS = qfq_round_down(agg->S, grp->slot_shift);
1362 	offset = (roundedS - grp->S) >> grp->slot_shift;
1363 
1364 	i = (grp->front + offset) % QFQ_MAX_SLOTS;
1365 
1366 	hlist_del(&agg->next);
1367 	if (hlist_empty(&grp->slots[i]))
1368 		__clear_bit(offset, &grp->full_slots);
1369 }
1370 
1371 /*
1372  * Called to forcibly deschedule an aggregate.  If the aggregate is
1373  * not in the front bucket, or if the latter has other aggregates in
1374  * the front bucket, we can simply remove the aggregate with no other
1375  * side effects.
1376  * Otherwise we must propagate the event up.
1377  */
1378 static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1379 {
1380 	struct qfq_group *grp = agg->grp;
1381 	unsigned long mask;
1382 	u64 roundedS;
1383 	int s;
1384 
1385 	if (agg == q->in_serv_agg) {
1386 		charge_actual_service(agg);
1387 		q->in_serv_agg = qfq_choose_next_agg(q);
1388 		return;
1389 	}
1390 
1391 	agg->F = agg->S;
1392 	qfq_slot_remove(q, grp, agg);
1393 
1394 	if (!grp->full_slots) {
1395 		__clear_bit(grp->index, &q->bitmaps[IR]);
1396 		__clear_bit(grp->index, &q->bitmaps[EB]);
1397 		__clear_bit(grp->index, &q->bitmaps[IB]);
1398 
1399 		if (test_bit(grp->index, &q->bitmaps[ER]) &&
1400 		    !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
1401 			mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
1402 			if (mask)
1403 				mask = ~((1UL << __fls(mask)) - 1);
1404 			else
1405 				mask = ~0UL;
1406 			qfq_move_groups(q, mask, EB, ER);
1407 			qfq_move_groups(q, mask, IB, IR);
1408 		}
1409 		__clear_bit(grp->index, &q->bitmaps[ER]);
1410 	} else if (hlist_empty(&grp->slots[grp->front])) {
1411 		agg = qfq_slot_scan(grp);
1412 		roundedS = qfq_round_down(agg->S, grp->slot_shift);
1413 		if (grp->S != roundedS) {
1414 			__clear_bit(grp->index, &q->bitmaps[ER]);
1415 			__clear_bit(grp->index, &q->bitmaps[IR]);
1416 			__clear_bit(grp->index, &q->bitmaps[EB]);
1417 			__clear_bit(grp->index, &q->bitmaps[IB]);
1418 			grp->S = roundedS;
1419 			grp->F = roundedS + (2ULL << grp->slot_shift);
1420 			s = qfq_calc_state(q, grp);
1421 			__set_bit(grp->index, &q->bitmaps[s]);
1422 		}
1423 	}
1424 }
1425 
1426 static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1427 {
1428 	struct qfq_sched *q = qdisc_priv(sch);
1429 	struct qfq_class *cl = (struct qfq_class *)arg;
1430 
1431 	if (cl->qdisc->q.qlen == 0)
1432 		qfq_deactivate_class(q, cl);
1433 }
1434 
1435 static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt)
1436 {
1437 	struct qfq_sched *q = qdisc_priv(sch);
1438 	struct qfq_group *grp;
1439 	int i, j, err;
1440 	u32 max_cl_shift, maxbudg_shift, max_classes;
1441 
1442 	err = tcf_block_get(&q->block, &q->filter_list);
1443 	if (err)
1444 		return err;
1445 
1446 	err = qdisc_class_hash_init(&q->clhash);
1447 	if (err < 0)
1448 		return err;
1449 
1450 	if (qdisc_dev(sch)->tx_queue_len + 1 > QFQ_MAX_AGG_CLASSES)
1451 		max_classes = QFQ_MAX_AGG_CLASSES;
1452 	else
1453 		max_classes = qdisc_dev(sch)->tx_queue_len + 1;
1454 	/* max_cl_shift = floor(log_2(max_classes)) */
1455 	max_cl_shift = __fls(max_classes);
1456 	q->max_agg_classes = 1<<max_cl_shift;
1457 
1458 	/* maxbudg_shift = log2(max_len * max_classes_per_agg) */
1459 	maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift;
1460 	q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX;
1461 
1462 	for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1463 		grp = &q->groups[i];
1464 		grp->index = i;
1465 		grp->slot_shift = q->min_slot_shift + i;
1466 		for (j = 0; j < QFQ_MAX_SLOTS; j++)
1467 			INIT_HLIST_HEAD(&grp->slots[j]);
1468 	}
1469 
1470 	INIT_HLIST_HEAD(&q->nonfull_aggs);
1471 
1472 	return 0;
1473 }
1474 
1475 static void qfq_reset_qdisc(struct Qdisc *sch)
1476 {
1477 	struct qfq_sched *q = qdisc_priv(sch);
1478 	struct qfq_class *cl;
1479 	unsigned int i;
1480 
1481 	for (i = 0; i < q->clhash.hashsize; i++) {
1482 		hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1483 			if (cl->qdisc->q.qlen > 0)
1484 				qfq_deactivate_class(q, cl);
1485 
1486 			qdisc_reset(cl->qdisc);
1487 		}
1488 	}
1489 	sch->qstats.backlog = 0;
1490 	sch->q.qlen = 0;
1491 }
1492 
1493 static void qfq_destroy_qdisc(struct Qdisc *sch)
1494 {
1495 	struct qfq_sched *q = qdisc_priv(sch);
1496 	struct qfq_class *cl;
1497 	struct hlist_node *next;
1498 	unsigned int i;
1499 
1500 	tcf_block_put(q->block);
1501 
1502 	for (i = 0; i < q->clhash.hashsize; i++) {
1503 		hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1504 					  common.hnode) {
1505 			qfq_destroy_class(sch, cl);
1506 		}
1507 	}
1508 	qdisc_class_hash_destroy(&q->clhash);
1509 }
1510 
1511 static const struct Qdisc_class_ops qfq_class_ops = {
1512 	.change		= qfq_change_class,
1513 	.delete		= qfq_delete_class,
1514 	.get		= qfq_get_class,
1515 	.put		= qfq_put_class,
1516 	.tcf_block	= qfq_tcf_block,
1517 	.bind_tcf	= qfq_bind_tcf,
1518 	.unbind_tcf	= qfq_unbind_tcf,
1519 	.graft		= qfq_graft_class,
1520 	.leaf		= qfq_class_leaf,
1521 	.qlen_notify	= qfq_qlen_notify,
1522 	.dump		= qfq_dump_class,
1523 	.dump_stats	= qfq_dump_class_stats,
1524 	.walk		= qfq_walk,
1525 };
1526 
1527 static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1528 	.cl_ops		= &qfq_class_ops,
1529 	.id		= "qfq",
1530 	.priv_size	= sizeof(struct qfq_sched),
1531 	.enqueue	= qfq_enqueue,
1532 	.dequeue	= qfq_dequeue,
1533 	.peek		= qdisc_peek_dequeued,
1534 	.init		= qfq_init_qdisc,
1535 	.reset		= qfq_reset_qdisc,
1536 	.destroy	= qfq_destroy_qdisc,
1537 	.owner		= THIS_MODULE,
1538 };
1539 
1540 static int __init qfq_init(void)
1541 {
1542 	return register_qdisc(&qfq_qdisc_ops);
1543 }
1544 
1545 static void __exit qfq_exit(void)
1546 {
1547 	unregister_qdisc(&qfq_qdisc_ops);
1548 }
1549 
1550 module_init(qfq_init);
1551 module_exit(qfq_exit);
1552 MODULE_LICENSE("GPL");
1553