• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_packed 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 
qfq_find_class(struct Qdisc * sch,u32 classid)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  */
qfq_calc_index(u32 inv_w,unsigned int maxlen,u32 min_slot_shift)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 
qfq_init_agg(struct qfq_sched * q,struct qfq_aggregate * agg,u32 lmax,u32 weight)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 
qfq_find_agg(struct qfq_sched * q,u32 lmax,u32 weight)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. */
qfq_update_agg(struct qfq_sched * q,struct qfq_aggregate * agg,int new_num_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. */
qfq_add_to_agg(struct qfq_sched * q,struct qfq_aggregate * agg,struct qfq_class * cl)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 
qfq_destroy_agg(struct qfq_sched * q,struct qfq_aggregate * agg)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. */
qfq_deactivate_class(struct qfq_sched * q,struct qfq_class * cl)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. */
qfq_rm_from_agg(struct qfq_sched * q,struct qfq_class * cl)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. */
qfq_deact_rm_from_agg(struct qfq_sched * q,struct qfq_class * cl)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 */
qfq_change_agg(struct Qdisc * sch,struct qfq_class * cl,u32 weight,u32 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 
qfq_change_class(struct Qdisc * sch,u32 classid,u32 parentid,struct nlattr ** tca,unsigned long * arg,struct netlink_ext_ack * extack)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 (tca[TCA_OPTIONS] == NULL) {
417 		pr_notice("qfq: no options\n");
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 		pr_notice("qfq: total weight out of range (%d + %u)\n",
455 			  delta_w, q->wsum);
456 		return -EINVAL;
457 	}
458 
459 	if (cl != NULL) { /* modify existing class */
460 		if (tca[TCA_RATE]) {
461 			err = gen_replace_estimator(&cl->bstats, NULL,
462 						    &cl->rate_est,
463 						    NULL,
464 						    qdisc_root_sleeping_running(sch),
465 						    tca[TCA_RATE]);
466 			if (err)
467 				return err;
468 		}
469 		existing = true;
470 		goto set_change_agg;
471 	}
472 
473 	/* create and init new class */
474 	cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
475 	if (cl == NULL)
476 		return -ENOBUFS;
477 
478 	cl->common.classid = classid;
479 	cl->deficit = lmax;
480 
481 	cl->qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
482 				      classid, NULL);
483 	if (cl->qdisc == NULL)
484 		cl->qdisc = &noop_qdisc;
485 
486 	if (tca[TCA_RATE]) {
487 		err = gen_new_estimator(&cl->bstats, NULL,
488 					&cl->rate_est,
489 					NULL,
490 					qdisc_root_sleeping_running(sch),
491 					tca[TCA_RATE]);
492 		if (err)
493 			goto destroy_class;
494 	}
495 
496 	if (cl->qdisc != &noop_qdisc)
497 		qdisc_hash_add(cl->qdisc, true);
498 
499 set_change_agg:
500 	sch_tree_lock(sch);
501 	new_agg = qfq_find_agg(q, lmax, weight);
502 	if (new_agg == NULL) { /* create new aggregate */
503 		sch_tree_unlock(sch);
504 		new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL);
505 		if (new_agg == NULL) {
506 			err = -ENOBUFS;
507 			gen_kill_estimator(&cl->rate_est);
508 			goto destroy_class;
509 		}
510 		sch_tree_lock(sch);
511 		qfq_init_agg(q, new_agg, lmax, weight);
512 	}
513 	if (existing)
514 		qfq_deact_rm_from_agg(q, cl);
515 	else
516 		qdisc_class_hash_insert(&q->clhash, &cl->common);
517 	qfq_add_to_agg(q, new_agg, cl);
518 	sch_tree_unlock(sch);
519 	qdisc_class_hash_grow(sch, &q->clhash);
520 
521 	*arg = (unsigned long)cl;
522 	return 0;
523 
524 destroy_class:
525 	qdisc_put(cl->qdisc);
526 	kfree(cl);
527 	return err;
528 }
529 
qfq_destroy_class(struct Qdisc * sch,struct qfq_class * cl)530 static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
531 {
532 	struct qfq_sched *q = qdisc_priv(sch);
533 
534 	qfq_rm_from_agg(q, cl);
535 	gen_kill_estimator(&cl->rate_est);
536 	qdisc_put(cl->qdisc);
537 	kfree(cl);
538 }
539 
qfq_delete_class(struct Qdisc * sch,unsigned long arg,struct netlink_ext_ack * extack)540 static int qfq_delete_class(struct Qdisc *sch, unsigned long arg,
541 			    struct netlink_ext_ack *extack)
542 {
543 	struct qfq_sched *q = qdisc_priv(sch);
544 	struct qfq_class *cl = (struct qfq_class *)arg;
545 
546 	if (cl->filter_cnt > 0)
547 		return -EBUSY;
548 
549 	sch_tree_lock(sch);
550 
551 	qdisc_purge_queue(cl->qdisc);
552 	qdisc_class_hash_remove(&q->clhash, &cl->common);
553 
554 	sch_tree_unlock(sch);
555 
556 	qfq_destroy_class(sch, cl);
557 	return 0;
558 }
559 
qfq_search_class(struct Qdisc * sch,u32 classid)560 static unsigned long qfq_search_class(struct Qdisc *sch, u32 classid)
561 {
562 	return (unsigned long)qfq_find_class(sch, classid);
563 }
564 
qfq_tcf_block(struct Qdisc * sch,unsigned long cl,struct netlink_ext_ack * extack)565 static struct tcf_block *qfq_tcf_block(struct Qdisc *sch, unsigned long cl,
566 				       struct netlink_ext_ack *extack)
567 {
568 	struct qfq_sched *q = qdisc_priv(sch);
569 
570 	if (cl)
571 		return NULL;
572 
573 	return q->block;
574 }
575 
qfq_bind_tcf(struct Qdisc * sch,unsigned long parent,u32 classid)576 static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
577 				  u32 classid)
578 {
579 	struct qfq_class *cl = qfq_find_class(sch, classid);
580 
581 	if (cl != NULL)
582 		cl->filter_cnt++;
583 
584 	return (unsigned long)cl;
585 }
586 
qfq_unbind_tcf(struct Qdisc * sch,unsigned long arg)587 static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
588 {
589 	struct qfq_class *cl = (struct qfq_class *)arg;
590 
591 	cl->filter_cnt--;
592 }
593 
qfq_graft_class(struct Qdisc * sch,unsigned long arg,struct Qdisc * new,struct Qdisc ** old,struct netlink_ext_ack * extack)594 static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
595 			   struct Qdisc *new, struct Qdisc **old,
596 			   struct netlink_ext_ack *extack)
597 {
598 	struct qfq_class *cl = (struct qfq_class *)arg;
599 
600 	if (new == NULL) {
601 		new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
602 					cl->common.classid, NULL);
603 		if (new == NULL)
604 			new = &noop_qdisc;
605 	}
606 
607 	*old = qdisc_replace(sch, new, &cl->qdisc);
608 	return 0;
609 }
610 
qfq_class_leaf(struct Qdisc * sch,unsigned long arg)611 static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
612 {
613 	struct qfq_class *cl = (struct qfq_class *)arg;
614 
615 	return cl->qdisc;
616 }
617 
qfq_dump_class(struct Qdisc * sch,unsigned long arg,struct sk_buff * skb,struct tcmsg * tcm)618 static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
619 			  struct sk_buff *skb, struct tcmsg *tcm)
620 {
621 	struct qfq_class *cl = (struct qfq_class *)arg;
622 	struct nlattr *nest;
623 
624 	tcm->tcm_parent	= TC_H_ROOT;
625 	tcm->tcm_handle	= cl->common.classid;
626 	tcm->tcm_info	= cl->qdisc->handle;
627 
628 	nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
629 	if (nest == NULL)
630 		goto nla_put_failure;
631 	if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) ||
632 	    nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax))
633 		goto nla_put_failure;
634 	return nla_nest_end(skb, nest);
635 
636 nla_put_failure:
637 	nla_nest_cancel(skb, nest);
638 	return -EMSGSIZE;
639 }
640 
qfq_dump_class_stats(struct Qdisc * sch,unsigned long arg,struct gnet_dump * d)641 static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
642 				struct gnet_dump *d)
643 {
644 	struct qfq_class *cl = (struct qfq_class *)arg;
645 	struct tc_qfq_stats xstats;
646 
647 	memset(&xstats, 0, sizeof(xstats));
648 
649 	xstats.weight = cl->agg->class_weight;
650 	xstats.lmax = cl->agg->lmax;
651 
652 	if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
653 				  d, NULL, &cl->bstats) < 0 ||
654 	    gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
655 	    qdisc_qstats_copy(d, cl->qdisc) < 0)
656 		return -1;
657 
658 	return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
659 }
660 
qfq_walk(struct Qdisc * sch,struct qdisc_walker * arg)661 static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
662 {
663 	struct qfq_sched *q = qdisc_priv(sch);
664 	struct qfq_class *cl;
665 	unsigned int i;
666 
667 	if (arg->stop)
668 		return;
669 
670 	for (i = 0; i < q->clhash.hashsize; i++) {
671 		hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
672 			if (arg->count < arg->skip) {
673 				arg->count++;
674 				continue;
675 			}
676 			if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
677 				arg->stop = 1;
678 				return;
679 			}
680 			arg->count++;
681 		}
682 	}
683 }
684 
qfq_classify(struct sk_buff * skb,struct Qdisc * sch,int * qerr)685 static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
686 				      int *qerr)
687 {
688 	struct qfq_sched *q = qdisc_priv(sch);
689 	struct qfq_class *cl;
690 	struct tcf_result res;
691 	struct tcf_proto *fl;
692 	int result;
693 
694 	if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
695 		pr_debug("qfq_classify: found %d\n", skb->priority);
696 		cl = qfq_find_class(sch, skb->priority);
697 		if (cl != NULL)
698 			return cl;
699 	}
700 
701 	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
702 	fl = rcu_dereference_bh(q->filter_list);
703 	result = tcf_classify(skb, NULL, fl, &res, false);
704 	if (result >= 0) {
705 #ifdef CONFIG_NET_CLS_ACT
706 		switch (result) {
707 		case TC_ACT_QUEUED:
708 		case TC_ACT_STOLEN:
709 		case TC_ACT_TRAP:
710 			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
711 			fallthrough;
712 		case TC_ACT_SHOT:
713 			return NULL;
714 		}
715 #endif
716 		cl = (struct qfq_class *)res.class;
717 		if (cl == NULL)
718 			cl = qfq_find_class(sch, res.classid);
719 		return cl;
720 	}
721 
722 	return NULL;
723 }
724 
725 /* Generic comparison function, handling wraparound. */
qfq_gt(u64 a,u64 b)726 static inline int qfq_gt(u64 a, u64 b)
727 {
728 	return (s64)(a - b) > 0;
729 }
730 
731 /* Round a precise timestamp to its slotted value. */
qfq_round_down(u64 ts,unsigned int shift)732 static inline u64 qfq_round_down(u64 ts, unsigned int shift)
733 {
734 	return ts & ~((1ULL << shift) - 1);
735 }
736 
737 /* return the pointer to the group with lowest index in the bitmap */
qfq_ffs(struct qfq_sched * q,unsigned long bitmap)738 static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
739 					unsigned long bitmap)
740 {
741 	int index = __ffs(bitmap);
742 	return &q->groups[index];
743 }
744 /* Calculate a mask to mimic what would be ffs_from(). */
mask_from(unsigned long bitmap,int from)745 static inline unsigned long mask_from(unsigned long bitmap, int from)
746 {
747 	return bitmap & ~((1UL << from) - 1);
748 }
749 
750 /*
751  * The state computation relies on ER=0, IR=1, EB=2, IB=3
752  * First compute eligibility comparing grp->S, q->V,
753  * then check if someone is blocking us and possibly add EB
754  */
qfq_calc_state(struct qfq_sched * q,const struct qfq_group * grp)755 static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
756 {
757 	/* if S > V we are not eligible */
758 	unsigned int state = qfq_gt(grp->S, q->V);
759 	unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
760 	struct qfq_group *next;
761 
762 	if (mask) {
763 		next = qfq_ffs(q, mask);
764 		if (qfq_gt(grp->F, next->F))
765 			state |= EB;
766 	}
767 
768 	return state;
769 }
770 
771 
772 /*
773  * In principle
774  *	q->bitmaps[dst] |= q->bitmaps[src] & mask;
775  *	q->bitmaps[src] &= ~mask;
776  * but we should make sure that src != dst
777  */
qfq_move_groups(struct qfq_sched * q,unsigned long mask,int src,int dst)778 static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
779 				   int src, int dst)
780 {
781 	q->bitmaps[dst] |= q->bitmaps[src] & mask;
782 	q->bitmaps[src] &= ~mask;
783 }
784 
qfq_unblock_groups(struct qfq_sched * q,int index,u64 old_F)785 static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
786 {
787 	unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
788 	struct qfq_group *next;
789 
790 	if (mask) {
791 		next = qfq_ffs(q, mask);
792 		if (!qfq_gt(next->F, old_F))
793 			return;
794 	}
795 
796 	mask = (1UL << index) - 1;
797 	qfq_move_groups(q, mask, EB, ER);
798 	qfq_move_groups(q, mask, IB, IR);
799 }
800 
801 /*
802  * perhaps
803  *
804 	old_V ^= q->V;
805 	old_V >>= q->min_slot_shift;
806 	if (old_V) {
807 		...
808 	}
809  *
810  */
qfq_make_eligible(struct qfq_sched * q)811 static void qfq_make_eligible(struct qfq_sched *q)
812 {
813 	unsigned long vslot = q->V >> q->min_slot_shift;
814 	unsigned long old_vslot = q->oldV >> q->min_slot_shift;
815 
816 	if (vslot != old_vslot) {
817 		unsigned long mask;
818 		int last_flip_pos = fls(vslot ^ old_vslot);
819 
820 		if (last_flip_pos > 31) /* higher than the number of groups */
821 			mask = ~0UL;    /* make all groups eligible */
822 		else
823 			mask = (1UL << last_flip_pos) - 1;
824 
825 		qfq_move_groups(q, mask, IR, ER);
826 		qfq_move_groups(q, mask, IB, EB);
827 	}
828 }
829 
830 /*
831  * The index of the slot in which the input aggregate agg is to be
832  * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2'
833  * and not a '-1' because the start time of the group may be moved
834  * backward by one slot after the aggregate has been inserted, and
835  * this would cause non-empty slots to be right-shifted by one
836  * position.
837  *
838  * QFQ+ fully satisfies this bound to the slot index if the parameters
839  * of the classes are not changed dynamically, and if QFQ+ never
840  * happens to postpone the service of agg unjustly, i.e., it never
841  * happens that the aggregate becomes backlogged and eligible, or just
842  * eligible, while an aggregate with a higher approximated finish time
843  * is being served. In particular, in this case QFQ+ guarantees that
844  * the timestamps of agg are low enough that the slot index is never
845  * higher than 2. Unfortunately, QFQ+ cannot provide the same
846  * guarantee if it happens to unjustly postpone the service of agg, or
847  * if the parameters of some class are changed.
848  *
849  * As for the first event, i.e., an out-of-order service, the
850  * upper bound to the slot index guaranteed by QFQ+ grows to
851  * 2 +
852  * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
853  * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1.
854  *
855  * The following function deals with this problem by backward-shifting
856  * the timestamps of agg, if needed, so as to guarantee that the slot
857  * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may
858  * cause the service of other aggregates to be postponed, yet the
859  * worst-case guarantees of these aggregates are not violated.  In
860  * fact, in case of no out-of-order service, the timestamps of agg
861  * would have been even lower than they are after the backward shift,
862  * because QFQ+ would have guaranteed a maximum value equal to 2 for
863  * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose
864  * service is postponed because of the backward-shift would have
865  * however waited for the service of agg before being served.
866  *
867  * The other event that may cause the slot index to be higher than 2
868  * for agg is a recent change of the parameters of some class. If the
869  * weight of a class is increased or the lmax (max_pkt_size) of the
870  * class is decreased, then a new aggregate with smaller slot size
871  * than the original parent aggregate of the class may happen to be
872  * activated. The activation of this aggregate should be properly
873  * delayed to when the service of the class has finished in the ideal
874  * system tracked by QFQ+. If the activation of the aggregate is not
875  * delayed to this reference time instant, then this aggregate may be
876  * unjustly served before other aggregates waiting for service. This
877  * may cause the above bound to the slot index to be violated for some
878  * of these unlucky aggregates.
879  *
880  * Instead of delaying the activation of the new aggregate, which is
881  * quite complex, the above-discussed capping of the slot index is
882  * used to handle also the consequences of a change of the parameters
883  * of a class.
884  */
qfq_slot_insert(struct qfq_group * grp,struct qfq_aggregate * agg,u64 roundedS)885 static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg,
886 			    u64 roundedS)
887 {
888 	u64 slot = (roundedS - grp->S) >> grp->slot_shift;
889 	unsigned int i; /* slot index in the bucket list */
890 
891 	if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
892 		u64 deltaS = roundedS - grp->S -
893 			((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
894 		agg->S -= deltaS;
895 		agg->F -= deltaS;
896 		slot = QFQ_MAX_SLOTS - 2;
897 	}
898 
899 	i = (grp->front + slot) % QFQ_MAX_SLOTS;
900 
901 	hlist_add_head(&agg->next, &grp->slots[i]);
902 	__set_bit(slot, &grp->full_slots);
903 }
904 
905 /* Maybe introduce hlist_first_entry?? */
qfq_slot_head(struct qfq_group * grp)906 static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp)
907 {
908 	return hlist_entry(grp->slots[grp->front].first,
909 			   struct qfq_aggregate, next);
910 }
911 
912 /*
913  * remove the entry from the slot
914  */
qfq_front_slot_remove(struct qfq_group * grp)915 static void qfq_front_slot_remove(struct qfq_group *grp)
916 {
917 	struct qfq_aggregate *agg = qfq_slot_head(grp);
918 
919 	BUG_ON(!agg);
920 	hlist_del(&agg->next);
921 	if (hlist_empty(&grp->slots[grp->front]))
922 		__clear_bit(0, &grp->full_slots);
923 }
924 
925 /*
926  * Returns the first aggregate in the first non-empty bucket of the
927  * group. As a side effect, adjusts the bucket list so the first
928  * non-empty bucket is at position 0 in full_slots.
929  */
qfq_slot_scan(struct qfq_group * grp)930 static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp)
931 {
932 	unsigned int i;
933 
934 	pr_debug("qfq slot_scan: grp %u full %#lx\n",
935 		 grp->index, grp->full_slots);
936 
937 	if (grp->full_slots == 0)
938 		return NULL;
939 
940 	i = __ffs(grp->full_slots);  /* zero based */
941 	if (i > 0) {
942 		grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
943 		grp->full_slots >>= i;
944 	}
945 
946 	return qfq_slot_head(grp);
947 }
948 
949 /*
950  * adjust the bucket list. When the start time of a group decreases,
951  * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
952  * move the objects. The mask of occupied slots must be shifted
953  * because we use ffs() to find the first non-empty slot.
954  * This covers decreases in the group's start time, but what about
955  * increases of the start time ?
956  * Here too we should make sure that i is less than 32
957  */
qfq_slot_rotate(struct qfq_group * grp,u64 roundedS)958 static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
959 {
960 	unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
961 
962 	grp->full_slots <<= i;
963 	grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
964 }
965 
qfq_update_eligible(struct qfq_sched * q)966 static void qfq_update_eligible(struct qfq_sched *q)
967 {
968 	struct qfq_group *grp;
969 	unsigned long ineligible;
970 
971 	ineligible = q->bitmaps[IR] | q->bitmaps[IB];
972 	if (ineligible) {
973 		if (!q->bitmaps[ER]) {
974 			grp = qfq_ffs(q, ineligible);
975 			if (qfq_gt(grp->S, q->V))
976 				q->V = grp->S;
977 		}
978 		qfq_make_eligible(q);
979 	}
980 }
981 
982 /* Dequeue head packet of the head class in the DRR queue of the aggregate. */
agg_dequeue(struct qfq_aggregate * agg,struct qfq_class * cl,unsigned int len)983 static struct sk_buff *agg_dequeue(struct qfq_aggregate *agg,
984 				   struct qfq_class *cl, unsigned int len)
985 {
986 	struct sk_buff *skb = qdisc_dequeue_peeked(cl->qdisc);
987 
988 	if (!skb)
989 		return NULL;
990 
991 	cl->deficit -= (int) len;
992 
993 	if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */
994 		list_del(&cl->alist);
995 	else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) {
996 		cl->deficit += agg->lmax;
997 		list_move_tail(&cl->alist, &agg->active);
998 	}
999 
1000 	return skb;
1001 }
1002 
qfq_peek_skb(struct qfq_aggregate * agg,struct qfq_class ** cl,unsigned int * len)1003 static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg,
1004 					   struct qfq_class **cl,
1005 					   unsigned int *len)
1006 {
1007 	struct sk_buff *skb;
1008 
1009 	*cl = list_first_entry(&agg->active, struct qfq_class, alist);
1010 	skb = (*cl)->qdisc->ops->peek((*cl)->qdisc);
1011 	if (skb == NULL)
1012 		WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
1013 	else
1014 		*len = qdisc_pkt_len(skb);
1015 
1016 	return skb;
1017 }
1018 
1019 /* Update F according to the actual service received by the aggregate. */
charge_actual_service(struct qfq_aggregate * agg)1020 static inline void charge_actual_service(struct qfq_aggregate *agg)
1021 {
1022 	/* Compute the service received by the aggregate, taking into
1023 	 * account that, after decreasing the number of classes in
1024 	 * agg, it may happen that
1025 	 * agg->initial_budget - agg->budget > agg->bugdetmax
1026 	 */
1027 	u32 service_received = min(agg->budgetmax,
1028 				   agg->initial_budget - agg->budget);
1029 
1030 	agg->F = agg->S + (u64)service_received * agg->inv_w;
1031 }
1032 
1033 /* Assign a reasonable start time for a new aggregate in group i.
1034  * Admissible values for \hat(F) are multiples of \sigma_i
1035  * no greater than V+\sigma_i . Larger values mean that
1036  * we had a wraparound so we consider the timestamp to be stale.
1037  *
1038  * If F is not stale and F >= V then we set S = F.
1039  * Otherwise we should assign S = V, but this may violate
1040  * the ordering in EB (see [2]). So, if we have groups in ER,
1041  * set S to the F_j of the first group j which would be blocking us.
1042  * We are guaranteed not to move S backward because
1043  * otherwise our group i would still be blocked.
1044  */
qfq_update_start(struct qfq_sched * q,struct qfq_aggregate * agg)1045 static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg)
1046 {
1047 	unsigned long mask;
1048 	u64 limit, roundedF;
1049 	int slot_shift = agg->grp->slot_shift;
1050 
1051 	roundedF = qfq_round_down(agg->F, slot_shift);
1052 	limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
1053 
1054 	if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) {
1055 		/* timestamp was stale */
1056 		mask = mask_from(q->bitmaps[ER], agg->grp->index);
1057 		if (mask) {
1058 			struct qfq_group *next = qfq_ffs(q, mask);
1059 			if (qfq_gt(roundedF, next->F)) {
1060 				if (qfq_gt(limit, next->F))
1061 					agg->S = next->F;
1062 				else /* preserve timestamp correctness */
1063 					agg->S = limit;
1064 				return;
1065 			}
1066 		}
1067 		agg->S = q->V;
1068 	} else  /* timestamp is not stale */
1069 		agg->S = agg->F;
1070 }
1071 
1072 /* Update the timestamps of agg before scheduling/rescheduling it for
1073  * service.  In particular, assign to agg->F its maximum possible
1074  * value, i.e., the virtual finish time with which the aggregate
1075  * should be labeled if it used all its budget once in service.
1076  */
1077 static inline void
qfq_update_agg_ts(struct qfq_sched * q,struct qfq_aggregate * agg,enum update_reason reason)1078 qfq_update_agg_ts(struct qfq_sched *q,
1079 		    struct qfq_aggregate *agg, enum update_reason reason)
1080 {
1081 	if (reason != requeue)
1082 		qfq_update_start(q, agg);
1083 	else /* just charge agg for the service received */
1084 		agg->S = agg->F;
1085 
1086 	agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w;
1087 }
1088 
1089 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg);
1090 
qfq_dequeue(struct Qdisc * sch)1091 static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
1092 {
1093 	struct qfq_sched *q = qdisc_priv(sch);
1094 	struct qfq_aggregate *in_serv_agg = q->in_serv_agg;
1095 	struct qfq_class *cl;
1096 	struct sk_buff *skb = NULL;
1097 	/* next-packet len, 0 means no more active classes in in-service agg */
1098 	unsigned int len = 0;
1099 
1100 	if (in_serv_agg == NULL)
1101 		return NULL;
1102 
1103 	if (!list_empty(&in_serv_agg->active))
1104 		skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1105 
1106 	/*
1107 	 * If there are no active classes in the in-service aggregate,
1108 	 * or if the aggregate has not enough budget to serve its next
1109 	 * class, then choose the next aggregate to serve.
1110 	 */
1111 	if (len == 0 || in_serv_agg->budget < len) {
1112 		charge_actual_service(in_serv_agg);
1113 
1114 		/* recharge the budget of the aggregate */
1115 		in_serv_agg->initial_budget = in_serv_agg->budget =
1116 			in_serv_agg->budgetmax;
1117 
1118 		if (!list_empty(&in_serv_agg->active)) {
1119 			/*
1120 			 * Still active: reschedule for
1121 			 * service. Possible optimization: if no other
1122 			 * aggregate is active, then there is no point
1123 			 * in rescheduling this aggregate, and we can
1124 			 * just keep it as the in-service one. This
1125 			 * should be however a corner case, and to
1126 			 * handle it, we would need to maintain an
1127 			 * extra num_active_aggs field.
1128 			*/
1129 			qfq_update_agg_ts(q, in_serv_agg, requeue);
1130 			qfq_schedule_agg(q, in_serv_agg);
1131 		} else if (sch->q.qlen == 0) { /* no aggregate to serve */
1132 			q->in_serv_agg = NULL;
1133 			return NULL;
1134 		}
1135 
1136 		/*
1137 		 * If we get here, there are other aggregates queued:
1138 		 * choose the new aggregate to serve.
1139 		 */
1140 		in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
1141 		skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1142 	}
1143 	if (!skb)
1144 		return NULL;
1145 
1146 	sch->q.qlen--;
1147 
1148 	skb = agg_dequeue(in_serv_agg, cl, len);
1149 
1150 	if (!skb) {
1151 		sch->q.qlen++;
1152 		return NULL;
1153 	}
1154 
1155 	qdisc_qstats_backlog_dec(sch, skb);
1156 	qdisc_bstats_update(sch, skb);
1157 
1158 	/* If lmax is lowered, through qfq_change_class, for a class
1159 	 * owning pending packets with larger size than the new value
1160 	 * of lmax, then the following condition may hold.
1161 	 */
1162 	if (unlikely(in_serv_agg->budget < len))
1163 		in_serv_agg->budget = 0;
1164 	else
1165 		in_serv_agg->budget -= len;
1166 
1167 	q->V += (u64)len * q->iwsum;
1168 	pr_debug("qfq dequeue: len %u F %lld now %lld\n",
1169 		 len, (unsigned long long) in_serv_agg->F,
1170 		 (unsigned long long) q->V);
1171 
1172 	return skb;
1173 }
1174 
qfq_choose_next_agg(struct qfq_sched * q)1175 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)
1176 {
1177 	struct qfq_group *grp;
1178 	struct qfq_aggregate *agg, *new_front_agg;
1179 	u64 old_F;
1180 
1181 	qfq_update_eligible(q);
1182 	q->oldV = q->V;
1183 
1184 	if (!q->bitmaps[ER])
1185 		return NULL;
1186 
1187 	grp = qfq_ffs(q, q->bitmaps[ER]);
1188 	old_F = grp->F;
1189 
1190 	agg = qfq_slot_head(grp);
1191 
1192 	/* agg starts to be served, remove it from schedule */
1193 	qfq_front_slot_remove(grp);
1194 
1195 	new_front_agg = qfq_slot_scan(grp);
1196 
1197 	if (new_front_agg == NULL) /* group is now inactive, remove from ER */
1198 		__clear_bit(grp->index, &q->bitmaps[ER]);
1199 	else {
1200 		u64 roundedS = qfq_round_down(new_front_agg->S,
1201 					      grp->slot_shift);
1202 		unsigned int s;
1203 
1204 		if (grp->S == roundedS)
1205 			return agg;
1206 		grp->S = roundedS;
1207 		grp->F = roundedS + (2ULL << grp->slot_shift);
1208 		__clear_bit(grp->index, &q->bitmaps[ER]);
1209 		s = qfq_calc_state(q, grp);
1210 		__set_bit(grp->index, &q->bitmaps[s]);
1211 	}
1212 
1213 	qfq_unblock_groups(q, grp->index, old_F);
1214 
1215 	return agg;
1216 }
1217 
qfq_enqueue(struct sk_buff * skb,struct Qdisc * sch,struct sk_buff ** to_free)1218 static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
1219 		       struct sk_buff **to_free)
1220 {
1221 	unsigned int len = qdisc_pkt_len(skb), gso_segs;
1222 	struct qfq_sched *q = qdisc_priv(sch);
1223 	struct qfq_class *cl;
1224 	struct qfq_aggregate *agg;
1225 	int err = 0;
1226 	bool first;
1227 
1228 	cl = qfq_classify(skb, sch, &err);
1229 	if (cl == NULL) {
1230 		if (err & __NET_XMIT_BYPASS)
1231 			qdisc_qstats_drop(sch);
1232 		__qdisc_drop(skb, to_free);
1233 		return err;
1234 	}
1235 	pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
1236 
1237 	if (unlikely(cl->agg->lmax < len)) {
1238 		pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
1239 			 cl->agg->lmax, len, cl->common.classid);
1240 		err = qfq_change_agg(sch, cl, cl->agg->class_weight, len);
1241 		if (err) {
1242 			cl->qstats.drops++;
1243 			return qdisc_drop(skb, sch, to_free);
1244 		}
1245 	}
1246 
1247 	gso_segs = skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 1;
1248 	first = !cl->qdisc->q.qlen;
1249 	err = qdisc_enqueue(skb, cl->qdisc, to_free);
1250 	if (unlikely(err != NET_XMIT_SUCCESS)) {
1251 		pr_debug("qfq_enqueue: enqueue failed %d\n", err);
1252 		if (net_xmit_drop_count(err)) {
1253 			cl->qstats.drops++;
1254 			qdisc_qstats_drop(sch);
1255 		}
1256 		return err;
1257 	}
1258 
1259 	cl->bstats.bytes += len;
1260 	cl->bstats.packets += gso_segs;
1261 	sch->qstats.backlog += len;
1262 	++sch->q.qlen;
1263 
1264 	agg = cl->agg;
1265 	/* if the queue was not empty, then done here */
1266 	if (!first) {
1267 		if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) &&
1268 		    list_first_entry(&agg->active, struct qfq_class, alist)
1269 		    == cl && cl->deficit < len)
1270 			list_move_tail(&cl->alist, &agg->active);
1271 
1272 		return err;
1273 	}
1274 
1275 	/* schedule class for service within the aggregate */
1276 	cl->deficit = agg->lmax;
1277 	list_add_tail(&cl->alist, &agg->active);
1278 
1279 	if (list_first_entry(&agg->active, struct qfq_class, alist) != cl ||
1280 	    q->in_serv_agg == agg)
1281 		return err; /* non-empty or in service, nothing else to do */
1282 
1283 	qfq_activate_agg(q, agg, enqueue);
1284 
1285 	return err;
1286 }
1287 
1288 /*
1289  * Schedule aggregate according to its timestamps.
1290  */
qfq_schedule_agg(struct qfq_sched * q,struct qfq_aggregate * agg)1291 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1292 {
1293 	struct qfq_group *grp = agg->grp;
1294 	u64 roundedS;
1295 	int s;
1296 
1297 	roundedS = qfq_round_down(agg->S, grp->slot_shift);
1298 
1299 	/*
1300 	 * Insert agg in the correct bucket.
1301 	 * If agg->S >= grp->S we don't need to adjust the
1302 	 * bucket list and simply go to the insertion phase.
1303 	 * Otherwise grp->S is decreasing, we must make room
1304 	 * in the bucket list, and also recompute the group state.
1305 	 * Finally, if there were no flows in this group and nobody
1306 	 * was in ER make sure to adjust V.
1307 	 */
1308 	if (grp->full_slots) {
1309 		if (!qfq_gt(grp->S, agg->S))
1310 			goto skip_update;
1311 
1312 		/* create a slot for this agg->S */
1313 		qfq_slot_rotate(grp, roundedS);
1314 		/* group was surely ineligible, remove */
1315 		__clear_bit(grp->index, &q->bitmaps[IR]);
1316 		__clear_bit(grp->index, &q->bitmaps[IB]);
1317 	} else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) &&
1318 		   q->in_serv_agg == NULL)
1319 		q->V = roundedS;
1320 
1321 	grp->S = roundedS;
1322 	grp->F = roundedS + (2ULL << grp->slot_shift);
1323 	s = qfq_calc_state(q, grp);
1324 	__set_bit(grp->index, &q->bitmaps[s]);
1325 
1326 	pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
1327 		 s, q->bitmaps[s],
1328 		 (unsigned long long) agg->S,
1329 		 (unsigned long long) agg->F,
1330 		 (unsigned long long) q->V);
1331 
1332 skip_update:
1333 	qfq_slot_insert(grp, agg, roundedS);
1334 }
1335 
1336 
1337 /* Update agg ts and schedule agg for service */
qfq_activate_agg(struct qfq_sched * q,struct qfq_aggregate * agg,enum update_reason reason)1338 static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
1339 			     enum update_reason reason)
1340 {
1341 	agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */
1342 
1343 	qfq_update_agg_ts(q, agg, reason);
1344 	if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */
1345 		q->in_serv_agg = agg; /* start serving this aggregate */
1346 		 /* update V: to be in service, agg must be eligible */
1347 		q->oldV = q->V = agg->S;
1348 	} else if (agg != q->in_serv_agg)
1349 		qfq_schedule_agg(q, agg);
1350 }
1351 
qfq_slot_remove(struct qfq_sched * q,struct qfq_group * grp,struct qfq_aggregate * agg)1352 static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
1353 			    struct qfq_aggregate *agg)
1354 {
1355 	unsigned int i, offset;
1356 	u64 roundedS;
1357 
1358 	roundedS = qfq_round_down(agg->S, grp->slot_shift);
1359 	offset = (roundedS - grp->S) >> grp->slot_shift;
1360 
1361 	i = (grp->front + offset) % QFQ_MAX_SLOTS;
1362 
1363 	hlist_del(&agg->next);
1364 	if (hlist_empty(&grp->slots[i]))
1365 		__clear_bit(offset, &grp->full_slots);
1366 }
1367 
1368 /*
1369  * Called to forcibly deschedule an aggregate.  If the aggregate is
1370  * not in the front bucket, or if the latter has other aggregates in
1371  * the front bucket, we can simply remove the aggregate with no other
1372  * side effects.
1373  * Otherwise we must propagate the event up.
1374  */
qfq_deactivate_agg(struct qfq_sched * q,struct qfq_aggregate * agg)1375 static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1376 {
1377 	struct qfq_group *grp = agg->grp;
1378 	unsigned long mask;
1379 	u64 roundedS;
1380 	int s;
1381 
1382 	if (agg == q->in_serv_agg) {
1383 		charge_actual_service(agg);
1384 		q->in_serv_agg = qfq_choose_next_agg(q);
1385 		return;
1386 	}
1387 
1388 	agg->F = agg->S;
1389 	qfq_slot_remove(q, grp, agg);
1390 
1391 	if (!grp->full_slots) {
1392 		__clear_bit(grp->index, &q->bitmaps[IR]);
1393 		__clear_bit(grp->index, &q->bitmaps[EB]);
1394 		__clear_bit(grp->index, &q->bitmaps[IB]);
1395 
1396 		if (test_bit(grp->index, &q->bitmaps[ER]) &&
1397 		    !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
1398 			mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
1399 			if (mask)
1400 				mask = ~((1UL << __fls(mask)) - 1);
1401 			else
1402 				mask = ~0UL;
1403 			qfq_move_groups(q, mask, EB, ER);
1404 			qfq_move_groups(q, mask, IB, IR);
1405 		}
1406 		__clear_bit(grp->index, &q->bitmaps[ER]);
1407 	} else if (hlist_empty(&grp->slots[grp->front])) {
1408 		agg = qfq_slot_scan(grp);
1409 		roundedS = qfq_round_down(agg->S, grp->slot_shift);
1410 		if (grp->S != roundedS) {
1411 			__clear_bit(grp->index, &q->bitmaps[ER]);
1412 			__clear_bit(grp->index, &q->bitmaps[IR]);
1413 			__clear_bit(grp->index, &q->bitmaps[EB]);
1414 			__clear_bit(grp->index, &q->bitmaps[IB]);
1415 			grp->S = roundedS;
1416 			grp->F = roundedS + (2ULL << grp->slot_shift);
1417 			s = qfq_calc_state(q, grp);
1418 			__set_bit(grp->index, &q->bitmaps[s]);
1419 		}
1420 	}
1421 }
1422 
qfq_qlen_notify(struct Qdisc * sch,unsigned long arg)1423 static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1424 {
1425 	struct qfq_sched *q = qdisc_priv(sch);
1426 	struct qfq_class *cl = (struct qfq_class *)arg;
1427 
1428 	qfq_deactivate_class(q, cl);
1429 }
1430 
qfq_init_qdisc(struct Qdisc * sch,struct nlattr * opt,struct netlink_ext_ack * extack)1431 static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt,
1432 			  struct netlink_ext_ack *extack)
1433 {
1434 	struct qfq_sched *q = qdisc_priv(sch);
1435 	struct qfq_group *grp;
1436 	int i, j, err;
1437 	u32 max_cl_shift, maxbudg_shift, max_classes;
1438 
1439 	err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
1440 	if (err)
1441 		return err;
1442 
1443 	err = qdisc_class_hash_init(&q->clhash);
1444 	if (err < 0)
1445 		return err;
1446 
1447 	max_classes = min_t(u64, (u64)qdisc_dev(sch)->tx_queue_len + 1,
1448 			    QFQ_MAX_AGG_CLASSES);
1449 	/* max_cl_shift = floor(log_2(max_classes)) */
1450 	max_cl_shift = __fls(max_classes);
1451 	q->max_agg_classes = 1<<max_cl_shift;
1452 
1453 	/* maxbudg_shift = log2(max_len * max_classes_per_agg) */
1454 	maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift;
1455 	q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX;
1456 
1457 	for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1458 		grp = &q->groups[i];
1459 		grp->index = i;
1460 		grp->slot_shift = q->min_slot_shift + i;
1461 		for (j = 0; j < QFQ_MAX_SLOTS; j++)
1462 			INIT_HLIST_HEAD(&grp->slots[j]);
1463 	}
1464 
1465 	INIT_HLIST_HEAD(&q->nonfull_aggs);
1466 
1467 	return 0;
1468 }
1469 
qfq_reset_qdisc(struct Qdisc * sch)1470 static void qfq_reset_qdisc(struct Qdisc *sch)
1471 {
1472 	struct qfq_sched *q = qdisc_priv(sch);
1473 	struct qfq_class *cl;
1474 	unsigned int i;
1475 
1476 	for (i = 0; i < q->clhash.hashsize; i++) {
1477 		hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1478 			if (cl->qdisc->q.qlen > 0)
1479 				qfq_deactivate_class(q, cl);
1480 
1481 			qdisc_reset(cl->qdisc);
1482 		}
1483 	}
1484 }
1485 
qfq_destroy_qdisc(struct Qdisc * sch)1486 static void qfq_destroy_qdisc(struct Qdisc *sch)
1487 {
1488 	struct qfq_sched *q = qdisc_priv(sch);
1489 	struct qfq_class *cl;
1490 	struct hlist_node *next;
1491 	unsigned int i;
1492 
1493 	tcf_block_put(q->block);
1494 
1495 	for (i = 0; i < q->clhash.hashsize; i++) {
1496 		hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1497 					  common.hnode) {
1498 			qfq_destroy_class(sch, cl);
1499 		}
1500 	}
1501 	qdisc_class_hash_destroy(&q->clhash);
1502 }
1503 
1504 static const struct Qdisc_class_ops qfq_class_ops = {
1505 	.change		= qfq_change_class,
1506 	.delete		= qfq_delete_class,
1507 	.find		= qfq_search_class,
1508 	.tcf_block	= qfq_tcf_block,
1509 	.bind_tcf	= qfq_bind_tcf,
1510 	.unbind_tcf	= qfq_unbind_tcf,
1511 	.graft		= qfq_graft_class,
1512 	.leaf		= qfq_class_leaf,
1513 	.qlen_notify	= qfq_qlen_notify,
1514 	.dump		= qfq_dump_class,
1515 	.dump_stats	= qfq_dump_class_stats,
1516 	.walk		= qfq_walk,
1517 };
1518 
1519 static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1520 	.cl_ops		= &qfq_class_ops,
1521 	.id		= "qfq",
1522 	.priv_size	= sizeof(struct qfq_sched),
1523 	.enqueue	= qfq_enqueue,
1524 	.dequeue	= qfq_dequeue,
1525 	.peek		= qdisc_peek_dequeued,
1526 	.init		= qfq_init_qdisc,
1527 	.reset		= qfq_reset_qdisc,
1528 	.destroy	= qfq_destroy_qdisc,
1529 	.owner		= THIS_MODULE,
1530 };
1531 
qfq_init(void)1532 static int __init qfq_init(void)
1533 {
1534 	return register_qdisc(&qfq_qdisc_ops);
1535 }
1536 
qfq_exit(void)1537 static void __exit qfq_exit(void)
1538 {
1539 	unregister_qdisc(&qfq_qdisc_ops);
1540 }
1541 
1542 module_init(qfq_init);
1543 module_exit(qfq_exit);
1544 MODULE_LICENSE("GPL");
1545