xref: /openbmc/linux/net/sched/sch_qfq.c (revision b755c25f)
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;
385 
386 	/* 'lmax' can range from [QFQ_MIN_LMAX, pktlen + stab overhead] */
387 	if (lmax > QFQ_MAX_LMAX)
388 		return -EINVAL;
389 
390 	new_agg = qfq_find_agg(q, lmax, weight);
391 	if (new_agg == NULL) { /* create new aggregate */
392 		new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC);
393 		if (new_agg == NULL)
394 			return -ENOBUFS;
395 		qfq_init_agg(q, new_agg, lmax, weight);
396 	}
397 	qfq_deact_rm_from_agg(q, cl);
398 	qfq_add_to_agg(q, new_agg, cl);
399 
400 	return 0;
401 }
402 
403 static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
404 			    struct nlattr **tca, unsigned long *arg,
405 			    struct netlink_ext_ack *extack)
406 {
407 	struct qfq_sched *q = qdisc_priv(sch);
408 	struct qfq_class *cl = (struct qfq_class *)*arg;
409 	bool existing = false;
410 	struct nlattr *tb[TCA_QFQ_MAX + 1];
411 	struct qfq_aggregate *new_agg = NULL;
412 	u32 weight, lmax, inv_w;
413 	int err;
414 	int delta_w;
415 
416 	if (NL_REQ_ATTR_CHECK(extack, NULL, tca, TCA_OPTIONS)) {
417 		NL_SET_ERR_MSG_MOD(extack, "missing options");
418 		return -EINVAL;
419 	}
420 
421 	err = nla_parse_nested_deprecated(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS],
422 					  qfq_policy, extack);
423 	if (err < 0)
424 		return err;
425 
426 	if (tb[TCA_QFQ_WEIGHT])
427 		weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]);
428 	else
429 		weight = 1;
430 
431 	if (tb[TCA_QFQ_LMAX]) {
432 		lmax = nla_get_u32(tb[TCA_QFQ_LMAX]);
433 	} else {
434 		/* MTU size is user controlled */
435 		lmax = psched_mtu(qdisc_dev(sch));
436 		if (lmax < QFQ_MIN_LMAX || lmax > QFQ_MAX_LMAX) {
437 			NL_SET_ERR_MSG_MOD(extack,
438 					   "MTU size out of bounds for qfq");
439 			return -EINVAL;
440 		}
441 	}
442 
443 	inv_w = ONE_FP / weight;
444 	weight = ONE_FP / inv_w;
445 
446 	if (cl != NULL &&
447 	    lmax == cl->agg->lmax &&
448 	    weight == cl->agg->class_weight)
449 		return 0; /* nothing to change */
450 
451 	delta_w = weight - (cl ? cl->agg->class_weight : 0);
452 
453 	if (q->wsum + delta_w > QFQ_MAX_WSUM) {
454 		NL_SET_ERR_MSG_FMT_MOD(extack,
455 				       "total weight out of range (%d + %u)\n",
456 				       delta_w, q->wsum);
457 		return -EINVAL;
458 	}
459 
460 	if (cl != NULL) { /* modify existing class */
461 		if (tca[TCA_RATE]) {
462 			err = gen_replace_estimator(&cl->bstats, NULL,
463 						    &cl->rate_est,
464 						    NULL,
465 						    true,
466 						    tca[TCA_RATE]);
467 			if (err)
468 				return err;
469 		}
470 		existing = true;
471 		goto set_change_agg;
472 	}
473 
474 	/* create and init new class */
475 	cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
476 	if (cl == NULL)
477 		return -ENOBUFS;
478 
479 	gnet_stats_basic_sync_init(&cl->bstats);
480 	cl->common.classid = classid;
481 	cl->deficit = lmax;
482 
483 	cl->qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
484 				      classid, NULL);
485 	if (cl->qdisc == NULL)
486 		cl->qdisc = &noop_qdisc;
487 
488 	if (tca[TCA_RATE]) {
489 		err = gen_new_estimator(&cl->bstats, NULL,
490 					&cl->rate_est,
491 					NULL,
492 					true,
493 					tca[TCA_RATE]);
494 		if (err)
495 			goto destroy_class;
496 	}
497 
498 	if (cl->qdisc != &noop_qdisc)
499 		qdisc_hash_add(cl->qdisc, true);
500 
501 set_change_agg:
502 	sch_tree_lock(sch);
503 	new_agg = qfq_find_agg(q, lmax, weight);
504 	if (new_agg == NULL) { /* create new aggregate */
505 		sch_tree_unlock(sch);
506 		new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL);
507 		if (new_agg == NULL) {
508 			err = -ENOBUFS;
509 			gen_kill_estimator(&cl->rate_est);
510 			goto destroy_class;
511 		}
512 		sch_tree_lock(sch);
513 		qfq_init_agg(q, new_agg, lmax, weight);
514 	}
515 	if (existing)
516 		qfq_deact_rm_from_agg(q, cl);
517 	else
518 		qdisc_class_hash_insert(&q->clhash, &cl->common);
519 	qfq_add_to_agg(q, new_agg, cl);
520 	sch_tree_unlock(sch);
521 	qdisc_class_hash_grow(sch, &q->clhash);
522 
523 	*arg = (unsigned long)cl;
524 	return 0;
525 
526 destroy_class:
527 	qdisc_put(cl->qdisc);
528 	kfree(cl);
529 	return err;
530 }
531 
532 static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
533 {
534 	struct qfq_sched *q = qdisc_priv(sch);
535 
536 	qfq_rm_from_agg(q, cl);
537 	gen_kill_estimator(&cl->rate_est);
538 	qdisc_put(cl->qdisc);
539 	kfree(cl);
540 }
541 
542 static int qfq_delete_class(struct Qdisc *sch, unsigned long arg,
543 			    struct netlink_ext_ack *extack)
544 {
545 	struct qfq_sched *q = qdisc_priv(sch);
546 	struct qfq_class *cl = (struct qfq_class *)arg;
547 
548 	if (cl->filter_cnt > 0)
549 		return -EBUSY;
550 
551 	sch_tree_lock(sch);
552 
553 	qdisc_purge_queue(cl->qdisc);
554 	qdisc_class_hash_remove(&q->clhash, &cl->common);
555 
556 	sch_tree_unlock(sch);
557 
558 	qfq_destroy_class(sch, cl);
559 	return 0;
560 }
561 
562 static unsigned long qfq_search_class(struct Qdisc *sch, u32 classid)
563 {
564 	return (unsigned long)qfq_find_class(sch, classid);
565 }
566 
567 static struct tcf_block *qfq_tcf_block(struct Qdisc *sch, unsigned long cl,
568 				       struct netlink_ext_ack *extack)
569 {
570 	struct qfq_sched *q = qdisc_priv(sch);
571 
572 	if (cl)
573 		return NULL;
574 
575 	return q->block;
576 }
577 
578 static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
579 				  u32 classid)
580 {
581 	struct qfq_class *cl = qfq_find_class(sch, classid);
582 
583 	if (cl != NULL)
584 		cl->filter_cnt++;
585 
586 	return (unsigned long)cl;
587 }
588 
589 static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
590 {
591 	struct qfq_class *cl = (struct qfq_class *)arg;
592 
593 	cl->filter_cnt--;
594 }
595 
596 static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
597 			   struct Qdisc *new, struct Qdisc **old,
598 			   struct netlink_ext_ack *extack)
599 {
600 	struct qfq_class *cl = (struct qfq_class *)arg;
601 
602 	if (new == NULL) {
603 		new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
604 					cl->common.classid, NULL);
605 		if (new == NULL)
606 			new = &noop_qdisc;
607 	}
608 
609 	*old = qdisc_replace(sch, new, &cl->qdisc);
610 	return 0;
611 }
612 
613 static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
614 {
615 	struct qfq_class *cl = (struct qfq_class *)arg;
616 
617 	return cl->qdisc;
618 }
619 
620 static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
621 			  struct sk_buff *skb, struct tcmsg *tcm)
622 {
623 	struct qfq_class *cl = (struct qfq_class *)arg;
624 	struct nlattr *nest;
625 
626 	tcm->tcm_parent	= TC_H_ROOT;
627 	tcm->tcm_handle	= cl->common.classid;
628 	tcm->tcm_info	= cl->qdisc->handle;
629 
630 	nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
631 	if (nest == NULL)
632 		goto nla_put_failure;
633 	if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) ||
634 	    nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax))
635 		goto nla_put_failure;
636 	return nla_nest_end(skb, nest);
637 
638 nla_put_failure:
639 	nla_nest_cancel(skb, nest);
640 	return -EMSGSIZE;
641 }
642 
643 static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
644 				struct gnet_dump *d)
645 {
646 	struct qfq_class *cl = (struct qfq_class *)arg;
647 	struct tc_qfq_stats xstats;
648 
649 	memset(&xstats, 0, sizeof(xstats));
650 
651 	xstats.weight = cl->agg->class_weight;
652 	xstats.lmax = cl->agg->lmax;
653 
654 	if (gnet_stats_copy_basic(d, NULL, &cl->bstats, true) < 0 ||
655 	    gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
656 	    qdisc_qstats_copy(d, cl->qdisc) < 0)
657 		return -1;
658 
659 	return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
660 }
661 
662 static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
663 {
664 	struct qfq_sched *q = qdisc_priv(sch);
665 	struct qfq_class *cl;
666 	unsigned int i;
667 
668 	if (arg->stop)
669 		return;
670 
671 	for (i = 0; i < q->clhash.hashsize; i++) {
672 		hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
673 			if (!tc_qdisc_stats_dump(sch, (unsigned long)cl, arg))
674 				return;
675 		}
676 	}
677 }
678 
679 static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
680 				      int *qerr)
681 {
682 	struct qfq_sched *q = qdisc_priv(sch);
683 	struct qfq_class *cl;
684 	struct tcf_result res;
685 	struct tcf_proto *fl;
686 	int result;
687 
688 	if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
689 		pr_debug("qfq_classify: found %d\n", skb->priority);
690 		cl = qfq_find_class(sch, skb->priority);
691 		if (cl != NULL)
692 			return cl;
693 	}
694 
695 	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
696 	fl = rcu_dereference_bh(q->filter_list);
697 	result = tcf_classify(skb, NULL, fl, &res, false);
698 	if (result >= 0) {
699 #ifdef CONFIG_NET_CLS_ACT
700 		switch (result) {
701 		case TC_ACT_QUEUED:
702 		case TC_ACT_STOLEN:
703 		case TC_ACT_TRAP:
704 			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
705 			fallthrough;
706 		case TC_ACT_SHOT:
707 			return NULL;
708 		}
709 #endif
710 		cl = (struct qfq_class *)res.class;
711 		if (cl == NULL)
712 			cl = qfq_find_class(sch, res.classid);
713 		return cl;
714 	}
715 
716 	return NULL;
717 }
718 
719 /* Generic comparison function, handling wraparound. */
720 static inline int qfq_gt(u64 a, u64 b)
721 {
722 	return (s64)(a - b) > 0;
723 }
724 
725 /* Round a precise timestamp to its slotted value. */
726 static inline u64 qfq_round_down(u64 ts, unsigned int shift)
727 {
728 	return ts & ~((1ULL << shift) - 1);
729 }
730 
731 /* return the pointer to the group with lowest index in the bitmap */
732 static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
733 					unsigned long bitmap)
734 {
735 	int index = __ffs(bitmap);
736 	return &q->groups[index];
737 }
738 /* Calculate a mask to mimic what would be ffs_from(). */
739 static inline unsigned long mask_from(unsigned long bitmap, int from)
740 {
741 	return bitmap & ~((1UL << from) - 1);
742 }
743 
744 /*
745  * The state computation relies on ER=0, IR=1, EB=2, IB=3
746  * First compute eligibility comparing grp->S, q->V,
747  * then check if someone is blocking us and possibly add EB
748  */
749 static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
750 {
751 	/* if S > V we are not eligible */
752 	unsigned int state = qfq_gt(grp->S, q->V);
753 	unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
754 	struct qfq_group *next;
755 
756 	if (mask) {
757 		next = qfq_ffs(q, mask);
758 		if (qfq_gt(grp->F, next->F))
759 			state |= EB;
760 	}
761 
762 	return state;
763 }
764 
765 
766 /*
767  * In principle
768  *	q->bitmaps[dst] |= q->bitmaps[src] & mask;
769  *	q->bitmaps[src] &= ~mask;
770  * but we should make sure that src != dst
771  */
772 static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
773 				   int src, int dst)
774 {
775 	q->bitmaps[dst] |= q->bitmaps[src] & mask;
776 	q->bitmaps[src] &= ~mask;
777 }
778 
779 static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
780 {
781 	unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
782 	struct qfq_group *next;
783 
784 	if (mask) {
785 		next = qfq_ffs(q, mask);
786 		if (!qfq_gt(next->F, old_F))
787 			return;
788 	}
789 
790 	mask = (1UL << index) - 1;
791 	qfq_move_groups(q, mask, EB, ER);
792 	qfq_move_groups(q, mask, IB, IR);
793 }
794 
795 /*
796  * perhaps
797  *
798 	old_V ^= q->V;
799 	old_V >>= q->min_slot_shift;
800 	if (old_V) {
801 		...
802 	}
803  *
804  */
805 static void qfq_make_eligible(struct qfq_sched *q)
806 {
807 	unsigned long vslot = q->V >> q->min_slot_shift;
808 	unsigned long old_vslot = q->oldV >> q->min_slot_shift;
809 
810 	if (vslot != old_vslot) {
811 		unsigned long mask;
812 		int last_flip_pos = fls(vslot ^ old_vslot);
813 
814 		if (last_flip_pos > 31) /* higher than the number of groups */
815 			mask = ~0UL;    /* make all groups eligible */
816 		else
817 			mask = (1UL << last_flip_pos) - 1;
818 
819 		qfq_move_groups(q, mask, IR, ER);
820 		qfq_move_groups(q, mask, IB, EB);
821 	}
822 }
823 
824 /*
825  * The index of the slot in which the input aggregate agg is to be
826  * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2'
827  * and not a '-1' because the start time of the group may be moved
828  * backward by one slot after the aggregate has been inserted, and
829  * this would cause non-empty slots to be right-shifted by one
830  * position.
831  *
832  * QFQ+ fully satisfies this bound to the slot index if the parameters
833  * of the classes are not changed dynamically, and if QFQ+ never
834  * happens to postpone the service of agg unjustly, i.e., it never
835  * happens that the aggregate becomes backlogged and eligible, or just
836  * eligible, while an aggregate with a higher approximated finish time
837  * is being served. In particular, in this case QFQ+ guarantees that
838  * the timestamps of agg are low enough that the slot index is never
839  * higher than 2. Unfortunately, QFQ+ cannot provide the same
840  * guarantee if it happens to unjustly postpone the service of agg, or
841  * if the parameters of some class are changed.
842  *
843  * As for the first event, i.e., an out-of-order service, the
844  * upper bound to the slot index guaranteed by QFQ+ grows to
845  * 2 +
846  * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
847  * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1.
848  *
849  * The following function deals with this problem by backward-shifting
850  * the timestamps of agg, if needed, so as to guarantee that the slot
851  * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may
852  * cause the service of other aggregates to be postponed, yet the
853  * worst-case guarantees of these aggregates are not violated.  In
854  * fact, in case of no out-of-order service, the timestamps of agg
855  * would have been even lower than they are after the backward shift,
856  * because QFQ+ would have guaranteed a maximum value equal to 2 for
857  * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose
858  * service is postponed because of the backward-shift would have
859  * however waited for the service of agg before being served.
860  *
861  * The other event that may cause the slot index to be higher than 2
862  * for agg is a recent change of the parameters of some class. If the
863  * weight of a class is increased or the lmax (max_pkt_size) of the
864  * class is decreased, then a new aggregate with smaller slot size
865  * than the original parent aggregate of the class may happen to be
866  * activated. The activation of this aggregate should be properly
867  * delayed to when the service of the class has finished in the ideal
868  * system tracked by QFQ+. If the activation of the aggregate is not
869  * delayed to this reference time instant, then this aggregate may be
870  * unjustly served before other aggregates waiting for service. This
871  * may cause the above bound to the slot index to be violated for some
872  * of these unlucky aggregates.
873  *
874  * Instead of delaying the activation of the new aggregate, which is
875  * quite complex, the above-discussed capping of the slot index is
876  * used to handle also the consequences of a change of the parameters
877  * of a class.
878  */
879 static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg,
880 			    u64 roundedS)
881 {
882 	u64 slot = (roundedS - grp->S) >> grp->slot_shift;
883 	unsigned int i; /* slot index in the bucket list */
884 
885 	if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
886 		u64 deltaS = roundedS - grp->S -
887 			((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
888 		agg->S -= deltaS;
889 		agg->F -= deltaS;
890 		slot = QFQ_MAX_SLOTS - 2;
891 	}
892 
893 	i = (grp->front + slot) % QFQ_MAX_SLOTS;
894 
895 	hlist_add_head(&agg->next, &grp->slots[i]);
896 	__set_bit(slot, &grp->full_slots);
897 }
898 
899 /* Maybe introduce hlist_first_entry?? */
900 static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp)
901 {
902 	return hlist_entry(grp->slots[grp->front].first,
903 			   struct qfq_aggregate, next);
904 }
905 
906 /*
907  * remove the entry from the slot
908  */
909 static void qfq_front_slot_remove(struct qfq_group *grp)
910 {
911 	struct qfq_aggregate *agg = qfq_slot_head(grp);
912 
913 	BUG_ON(!agg);
914 	hlist_del(&agg->next);
915 	if (hlist_empty(&grp->slots[grp->front]))
916 		__clear_bit(0, &grp->full_slots);
917 }
918 
919 /*
920  * Returns the first aggregate in the first non-empty bucket of the
921  * group. As a side effect, adjusts the bucket list so the first
922  * non-empty bucket is at position 0 in full_slots.
923  */
924 static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp)
925 {
926 	unsigned int i;
927 
928 	pr_debug("qfq slot_scan: grp %u full %#lx\n",
929 		 grp->index, grp->full_slots);
930 
931 	if (grp->full_slots == 0)
932 		return NULL;
933 
934 	i = __ffs(grp->full_slots);  /* zero based */
935 	if (i > 0) {
936 		grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
937 		grp->full_slots >>= i;
938 	}
939 
940 	return qfq_slot_head(grp);
941 }
942 
943 /*
944  * adjust the bucket list. When the start time of a group decreases,
945  * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
946  * move the objects. The mask of occupied slots must be shifted
947  * because we use ffs() to find the first non-empty slot.
948  * This covers decreases in the group's start time, but what about
949  * increases of the start time ?
950  * Here too we should make sure that i is less than 32
951  */
952 static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
953 {
954 	unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
955 
956 	grp->full_slots <<= i;
957 	grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
958 }
959 
960 static void qfq_update_eligible(struct qfq_sched *q)
961 {
962 	struct qfq_group *grp;
963 	unsigned long ineligible;
964 
965 	ineligible = q->bitmaps[IR] | q->bitmaps[IB];
966 	if (ineligible) {
967 		if (!q->bitmaps[ER]) {
968 			grp = qfq_ffs(q, ineligible);
969 			if (qfq_gt(grp->S, q->V))
970 				q->V = grp->S;
971 		}
972 		qfq_make_eligible(q);
973 	}
974 }
975 
976 /* Dequeue head packet of the head class in the DRR queue of the aggregate. */
977 static void agg_dequeue(struct qfq_aggregate *agg,
978 			struct qfq_class *cl, unsigned int len)
979 {
980 	qdisc_dequeue_peeked(cl->qdisc);
981 
982 	cl->deficit -= (int) len;
983 
984 	if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */
985 		list_del(&cl->alist);
986 	else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) {
987 		cl->deficit += agg->lmax;
988 		list_move_tail(&cl->alist, &agg->active);
989 	}
990 }
991 
992 static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg,
993 					   struct qfq_class **cl,
994 					   unsigned int *len)
995 {
996 	struct sk_buff *skb;
997 
998 	*cl = list_first_entry(&agg->active, struct qfq_class, alist);
999 	skb = (*cl)->qdisc->ops->peek((*cl)->qdisc);
1000 	if (skb == NULL)
1001 		WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
1002 	else
1003 		*len = qdisc_pkt_len(skb);
1004 
1005 	return skb;
1006 }
1007 
1008 /* Update F according to the actual service received by the aggregate. */
1009 static inline void charge_actual_service(struct qfq_aggregate *agg)
1010 {
1011 	/* Compute the service received by the aggregate, taking into
1012 	 * account that, after decreasing the number of classes in
1013 	 * agg, it may happen that
1014 	 * agg->initial_budget - agg->budget > agg->bugdetmax
1015 	 */
1016 	u32 service_received = min(agg->budgetmax,
1017 				   agg->initial_budget - agg->budget);
1018 
1019 	agg->F = agg->S + (u64)service_received * agg->inv_w;
1020 }
1021 
1022 /* Assign a reasonable start time for a new aggregate in group i.
1023  * Admissible values for \hat(F) are multiples of \sigma_i
1024  * no greater than V+\sigma_i . Larger values mean that
1025  * we had a wraparound so we consider the timestamp to be stale.
1026  *
1027  * If F is not stale and F >= V then we set S = F.
1028  * Otherwise we should assign S = V, but this may violate
1029  * the ordering in EB (see [2]). So, if we have groups in ER,
1030  * set S to the F_j of the first group j which would be blocking us.
1031  * We are guaranteed not to move S backward because
1032  * otherwise our group i would still be blocked.
1033  */
1034 static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg)
1035 {
1036 	unsigned long mask;
1037 	u64 limit, roundedF;
1038 	int slot_shift = agg->grp->slot_shift;
1039 
1040 	roundedF = qfq_round_down(agg->F, slot_shift);
1041 	limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
1042 
1043 	if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) {
1044 		/* timestamp was stale */
1045 		mask = mask_from(q->bitmaps[ER], agg->grp->index);
1046 		if (mask) {
1047 			struct qfq_group *next = qfq_ffs(q, mask);
1048 			if (qfq_gt(roundedF, next->F)) {
1049 				if (qfq_gt(limit, next->F))
1050 					agg->S = next->F;
1051 				else /* preserve timestamp correctness */
1052 					agg->S = limit;
1053 				return;
1054 			}
1055 		}
1056 		agg->S = q->V;
1057 	} else  /* timestamp is not stale */
1058 		agg->S = agg->F;
1059 }
1060 
1061 /* Update the timestamps of agg before scheduling/rescheduling it for
1062  * service.  In particular, assign to agg->F its maximum possible
1063  * value, i.e., the virtual finish time with which the aggregate
1064  * should be labeled if it used all its budget once in service.
1065  */
1066 static inline void
1067 qfq_update_agg_ts(struct qfq_sched *q,
1068 		    struct qfq_aggregate *agg, enum update_reason reason)
1069 {
1070 	if (reason != requeue)
1071 		qfq_update_start(q, agg);
1072 	else /* just charge agg for the service received */
1073 		agg->S = agg->F;
1074 
1075 	agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w;
1076 }
1077 
1078 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg);
1079 
1080 static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
1081 {
1082 	struct qfq_sched *q = qdisc_priv(sch);
1083 	struct qfq_aggregate *in_serv_agg = q->in_serv_agg;
1084 	struct qfq_class *cl;
1085 	struct sk_buff *skb = NULL;
1086 	/* next-packet len, 0 means no more active classes in in-service agg */
1087 	unsigned int len = 0;
1088 
1089 	if (in_serv_agg == NULL)
1090 		return NULL;
1091 
1092 	if (!list_empty(&in_serv_agg->active))
1093 		skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1094 
1095 	/*
1096 	 * If there are no active classes in the in-service aggregate,
1097 	 * or if the aggregate has not enough budget to serve its next
1098 	 * class, then choose the next aggregate to serve.
1099 	 */
1100 	if (len == 0 || in_serv_agg->budget < len) {
1101 		charge_actual_service(in_serv_agg);
1102 
1103 		/* recharge the budget of the aggregate */
1104 		in_serv_agg->initial_budget = in_serv_agg->budget =
1105 			in_serv_agg->budgetmax;
1106 
1107 		if (!list_empty(&in_serv_agg->active)) {
1108 			/*
1109 			 * Still active: reschedule for
1110 			 * service. Possible optimization: if no other
1111 			 * aggregate is active, then there is no point
1112 			 * in rescheduling this aggregate, and we can
1113 			 * just keep it as the in-service one. This
1114 			 * should be however a corner case, and to
1115 			 * handle it, we would need to maintain an
1116 			 * extra num_active_aggs field.
1117 			*/
1118 			qfq_update_agg_ts(q, in_serv_agg, requeue);
1119 			qfq_schedule_agg(q, in_serv_agg);
1120 		} else if (sch->q.qlen == 0) { /* no aggregate to serve */
1121 			q->in_serv_agg = NULL;
1122 			return NULL;
1123 		}
1124 
1125 		/*
1126 		 * If we get here, there are other aggregates queued:
1127 		 * choose the new aggregate to serve.
1128 		 */
1129 		in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
1130 		skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1131 	}
1132 	if (!skb)
1133 		return NULL;
1134 
1135 	qdisc_qstats_backlog_dec(sch, skb);
1136 	sch->q.qlen--;
1137 	qdisc_bstats_update(sch, skb);
1138 
1139 	agg_dequeue(in_serv_agg, cl, len);
1140 	/* If lmax is lowered, through qfq_change_class, for a class
1141 	 * owning pending packets with larger size than the new value
1142 	 * of lmax, then the following condition may hold.
1143 	 */
1144 	if (unlikely(in_serv_agg->budget < len))
1145 		in_serv_agg->budget = 0;
1146 	else
1147 		in_serv_agg->budget -= len;
1148 
1149 	q->V += (u64)len * q->iwsum;
1150 	pr_debug("qfq dequeue: len %u F %lld now %lld\n",
1151 		 len, (unsigned long long) in_serv_agg->F,
1152 		 (unsigned long long) q->V);
1153 
1154 	return skb;
1155 }
1156 
1157 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)
1158 {
1159 	struct qfq_group *grp;
1160 	struct qfq_aggregate *agg, *new_front_agg;
1161 	u64 old_F;
1162 
1163 	qfq_update_eligible(q);
1164 	q->oldV = q->V;
1165 
1166 	if (!q->bitmaps[ER])
1167 		return NULL;
1168 
1169 	grp = qfq_ffs(q, q->bitmaps[ER]);
1170 	old_F = grp->F;
1171 
1172 	agg = qfq_slot_head(grp);
1173 
1174 	/* agg starts to be served, remove it from schedule */
1175 	qfq_front_slot_remove(grp);
1176 
1177 	new_front_agg = qfq_slot_scan(grp);
1178 
1179 	if (new_front_agg == NULL) /* group is now inactive, remove from ER */
1180 		__clear_bit(grp->index, &q->bitmaps[ER]);
1181 	else {
1182 		u64 roundedS = qfq_round_down(new_front_agg->S,
1183 					      grp->slot_shift);
1184 		unsigned int s;
1185 
1186 		if (grp->S == roundedS)
1187 			return agg;
1188 		grp->S = roundedS;
1189 		grp->F = roundedS + (2ULL << grp->slot_shift);
1190 		__clear_bit(grp->index, &q->bitmaps[ER]);
1191 		s = qfq_calc_state(q, grp);
1192 		__set_bit(grp->index, &q->bitmaps[s]);
1193 	}
1194 
1195 	qfq_unblock_groups(q, grp->index, old_F);
1196 
1197 	return agg;
1198 }
1199 
1200 static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
1201 		       struct sk_buff **to_free)
1202 {
1203 	unsigned int len = qdisc_pkt_len(skb), gso_segs;
1204 	struct qfq_sched *q = qdisc_priv(sch);
1205 	struct qfq_class *cl;
1206 	struct qfq_aggregate *agg;
1207 	int err = 0;
1208 	bool first;
1209 
1210 	cl = qfq_classify(skb, sch, &err);
1211 	if (cl == NULL) {
1212 		if (err & __NET_XMIT_BYPASS)
1213 			qdisc_qstats_drop(sch);
1214 		__qdisc_drop(skb, to_free);
1215 		return err;
1216 	}
1217 	pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
1218 
1219 	if (unlikely(cl->agg->lmax < len)) {
1220 		pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
1221 			 cl->agg->lmax, len, cl->common.classid);
1222 		err = qfq_change_agg(sch, cl, cl->agg->class_weight, len);
1223 		if (err) {
1224 			cl->qstats.drops++;
1225 			return qdisc_drop(skb, sch, to_free);
1226 		}
1227 	}
1228 
1229 	gso_segs = skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 1;
1230 	first = !cl->qdisc->q.qlen;
1231 	err = qdisc_enqueue(skb, cl->qdisc, to_free);
1232 	if (unlikely(err != NET_XMIT_SUCCESS)) {
1233 		pr_debug("qfq_enqueue: enqueue failed %d\n", err);
1234 		if (net_xmit_drop_count(err)) {
1235 			cl->qstats.drops++;
1236 			qdisc_qstats_drop(sch);
1237 		}
1238 		return err;
1239 	}
1240 
1241 	_bstats_update(&cl->bstats, len, gso_segs);
1242 	sch->qstats.backlog += len;
1243 	++sch->q.qlen;
1244 
1245 	agg = cl->agg;
1246 	/* if the queue was not empty, then done here */
1247 	if (!first) {
1248 		if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) &&
1249 		    list_first_entry(&agg->active, struct qfq_class, alist)
1250 		    == cl && cl->deficit < len)
1251 			list_move_tail(&cl->alist, &agg->active);
1252 
1253 		return err;
1254 	}
1255 
1256 	/* schedule class for service within the aggregate */
1257 	cl->deficit = agg->lmax;
1258 	list_add_tail(&cl->alist, &agg->active);
1259 
1260 	if (list_first_entry(&agg->active, struct qfq_class, alist) != cl ||
1261 	    q->in_serv_agg == agg)
1262 		return err; /* non-empty or in service, nothing else to do */
1263 
1264 	qfq_activate_agg(q, agg, enqueue);
1265 
1266 	return err;
1267 }
1268 
1269 /*
1270  * Schedule aggregate according to its timestamps.
1271  */
1272 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1273 {
1274 	struct qfq_group *grp = agg->grp;
1275 	u64 roundedS;
1276 	int s;
1277 
1278 	roundedS = qfq_round_down(agg->S, grp->slot_shift);
1279 
1280 	/*
1281 	 * Insert agg in the correct bucket.
1282 	 * If agg->S >= grp->S we don't need to adjust the
1283 	 * bucket list and simply go to the insertion phase.
1284 	 * Otherwise grp->S is decreasing, we must make room
1285 	 * in the bucket list, and also recompute the group state.
1286 	 * Finally, if there were no flows in this group and nobody
1287 	 * was in ER make sure to adjust V.
1288 	 */
1289 	if (grp->full_slots) {
1290 		if (!qfq_gt(grp->S, agg->S))
1291 			goto skip_update;
1292 
1293 		/* create a slot for this agg->S */
1294 		qfq_slot_rotate(grp, roundedS);
1295 		/* group was surely ineligible, remove */
1296 		__clear_bit(grp->index, &q->bitmaps[IR]);
1297 		__clear_bit(grp->index, &q->bitmaps[IB]);
1298 	} else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) &&
1299 		   q->in_serv_agg == NULL)
1300 		q->V = roundedS;
1301 
1302 	grp->S = roundedS;
1303 	grp->F = roundedS + (2ULL << grp->slot_shift);
1304 	s = qfq_calc_state(q, grp);
1305 	__set_bit(grp->index, &q->bitmaps[s]);
1306 
1307 	pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
1308 		 s, q->bitmaps[s],
1309 		 (unsigned long long) agg->S,
1310 		 (unsigned long long) agg->F,
1311 		 (unsigned long long) q->V);
1312 
1313 skip_update:
1314 	qfq_slot_insert(grp, agg, roundedS);
1315 }
1316 
1317 
1318 /* Update agg ts and schedule agg for service */
1319 static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
1320 			     enum update_reason reason)
1321 {
1322 	agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */
1323 
1324 	qfq_update_agg_ts(q, agg, reason);
1325 	if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */
1326 		q->in_serv_agg = agg; /* start serving this aggregate */
1327 		 /* update V: to be in service, agg must be eligible */
1328 		q->oldV = q->V = agg->S;
1329 	} else if (agg != q->in_serv_agg)
1330 		qfq_schedule_agg(q, agg);
1331 }
1332 
1333 static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
1334 			    struct qfq_aggregate *agg)
1335 {
1336 	unsigned int i, offset;
1337 	u64 roundedS;
1338 
1339 	roundedS = qfq_round_down(agg->S, grp->slot_shift);
1340 	offset = (roundedS - grp->S) >> grp->slot_shift;
1341 
1342 	i = (grp->front + offset) % QFQ_MAX_SLOTS;
1343 
1344 	hlist_del(&agg->next);
1345 	if (hlist_empty(&grp->slots[i]))
1346 		__clear_bit(offset, &grp->full_slots);
1347 }
1348 
1349 /*
1350  * Called to forcibly deschedule an aggregate.  If the aggregate is
1351  * not in the front bucket, or if the latter has other aggregates in
1352  * the front bucket, we can simply remove the aggregate with no other
1353  * side effects.
1354  * Otherwise we must propagate the event up.
1355  */
1356 static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1357 {
1358 	struct qfq_group *grp = agg->grp;
1359 	unsigned long mask;
1360 	u64 roundedS;
1361 	int s;
1362 
1363 	if (agg == q->in_serv_agg) {
1364 		charge_actual_service(agg);
1365 		q->in_serv_agg = qfq_choose_next_agg(q);
1366 		return;
1367 	}
1368 
1369 	agg->F = agg->S;
1370 	qfq_slot_remove(q, grp, agg);
1371 
1372 	if (!grp->full_slots) {
1373 		__clear_bit(grp->index, &q->bitmaps[IR]);
1374 		__clear_bit(grp->index, &q->bitmaps[EB]);
1375 		__clear_bit(grp->index, &q->bitmaps[IB]);
1376 
1377 		if (test_bit(grp->index, &q->bitmaps[ER]) &&
1378 		    !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
1379 			mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
1380 			if (mask)
1381 				mask = ~((1UL << __fls(mask)) - 1);
1382 			else
1383 				mask = ~0UL;
1384 			qfq_move_groups(q, mask, EB, ER);
1385 			qfq_move_groups(q, mask, IB, IR);
1386 		}
1387 		__clear_bit(grp->index, &q->bitmaps[ER]);
1388 	} else if (hlist_empty(&grp->slots[grp->front])) {
1389 		agg = qfq_slot_scan(grp);
1390 		roundedS = qfq_round_down(agg->S, grp->slot_shift);
1391 		if (grp->S != roundedS) {
1392 			__clear_bit(grp->index, &q->bitmaps[ER]);
1393 			__clear_bit(grp->index, &q->bitmaps[IR]);
1394 			__clear_bit(grp->index, &q->bitmaps[EB]);
1395 			__clear_bit(grp->index, &q->bitmaps[IB]);
1396 			grp->S = roundedS;
1397 			grp->F = roundedS + (2ULL << grp->slot_shift);
1398 			s = qfq_calc_state(q, grp);
1399 			__set_bit(grp->index, &q->bitmaps[s]);
1400 		}
1401 	}
1402 }
1403 
1404 static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1405 {
1406 	struct qfq_sched *q = qdisc_priv(sch);
1407 	struct qfq_class *cl = (struct qfq_class *)arg;
1408 
1409 	qfq_deactivate_class(q, cl);
1410 }
1411 
1412 static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt,
1413 			  struct netlink_ext_ack *extack)
1414 {
1415 	struct qfq_sched *q = qdisc_priv(sch);
1416 	struct qfq_group *grp;
1417 	int i, j, err;
1418 	u32 max_cl_shift, maxbudg_shift, max_classes;
1419 
1420 	err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
1421 	if (err)
1422 		return err;
1423 
1424 	err = qdisc_class_hash_init(&q->clhash);
1425 	if (err < 0)
1426 		return err;
1427 
1428 	max_classes = min_t(u64, (u64)qdisc_dev(sch)->tx_queue_len + 1,
1429 			    QFQ_MAX_AGG_CLASSES);
1430 	/* max_cl_shift = floor(log_2(max_classes)) */
1431 	max_cl_shift = __fls(max_classes);
1432 	q->max_agg_classes = 1<<max_cl_shift;
1433 
1434 	/* maxbudg_shift = log2(max_len * max_classes_per_agg) */
1435 	maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift;
1436 	q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX;
1437 
1438 	for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1439 		grp = &q->groups[i];
1440 		grp->index = i;
1441 		grp->slot_shift = q->min_slot_shift + i;
1442 		for (j = 0; j < QFQ_MAX_SLOTS; j++)
1443 			INIT_HLIST_HEAD(&grp->slots[j]);
1444 	}
1445 
1446 	INIT_HLIST_HEAD(&q->nonfull_aggs);
1447 
1448 	return 0;
1449 }
1450 
1451 static void qfq_reset_qdisc(struct Qdisc *sch)
1452 {
1453 	struct qfq_sched *q = qdisc_priv(sch);
1454 	struct qfq_class *cl;
1455 	unsigned int i;
1456 
1457 	for (i = 0; i < q->clhash.hashsize; i++) {
1458 		hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1459 			if (cl->qdisc->q.qlen > 0)
1460 				qfq_deactivate_class(q, cl);
1461 
1462 			qdisc_reset(cl->qdisc);
1463 		}
1464 	}
1465 }
1466 
1467 static void qfq_destroy_qdisc(struct Qdisc *sch)
1468 {
1469 	struct qfq_sched *q = qdisc_priv(sch);
1470 	struct qfq_class *cl;
1471 	struct hlist_node *next;
1472 	unsigned int i;
1473 
1474 	tcf_block_put(q->block);
1475 
1476 	for (i = 0; i < q->clhash.hashsize; i++) {
1477 		hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1478 					  common.hnode) {
1479 			qfq_destroy_class(sch, cl);
1480 		}
1481 	}
1482 	qdisc_class_hash_destroy(&q->clhash);
1483 }
1484 
1485 static const struct Qdisc_class_ops qfq_class_ops = {
1486 	.change		= qfq_change_class,
1487 	.delete		= qfq_delete_class,
1488 	.find		= qfq_search_class,
1489 	.tcf_block	= qfq_tcf_block,
1490 	.bind_tcf	= qfq_bind_tcf,
1491 	.unbind_tcf	= qfq_unbind_tcf,
1492 	.graft		= qfq_graft_class,
1493 	.leaf		= qfq_class_leaf,
1494 	.qlen_notify	= qfq_qlen_notify,
1495 	.dump		= qfq_dump_class,
1496 	.dump_stats	= qfq_dump_class_stats,
1497 	.walk		= qfq_walk,
1498 };
1499 
1500 static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1501 	.cl_ops		= &qfq_class_ops,
1502 	.id		= "qfq",
1503 	.priv_size	= sizeof(struct qfq_sched),
1504 	.enqueue	= qfq_enqueue,
1505 	.dequeue	= qfq_dequeue,
1506 	.peek		= qdisc_peek_dequeued,
1507 	.init		= qfq_init_qdisc,
1508 	.reset		= qfq_reset_qdisc,
1509 	.destroy	= qfq_destroy_qdisc,
1510 	.owner		= THIS_MODULE,
1511 };
1512 
1513 static int __init qfq_init(void)
1514 {
1515 	return register_qdisc(&qfq_qdisc_ops);
1516 }
1517 
1518 static void __exit qfq_exit(void)
1519 {
1520 	unregister_qdisc(&qfq_qdisc_ops);
1521 }
1522 
1523 module_init(qfq_init);
1524 module_exit(qfq_exit);
1525 MODULE_LICENSE("GPL");
1526