xref: /openbmc/linux/net/sched/sch_qfq.c (revision 875e5771536f8f631f38f0c6090a108cd611fcf3)
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 	struct gnet_stats_basic_sync bstats;
134 	struct gnet_stats_queue qstats;
135 	struct net_rate_estimator __rcu *rate_est;
136 	struct Qdisc *qdisc;
137 	struct list_head alist;		/* Link for active-classes list. */
138 	struct qfq_aggregate *agg;	/* Parent aggregate. */
139 	int deficit;			/* DRR deficit counter. */
140 };
141 
142 struct qfq_aggregate {
143 	struct hlist_node next;	/* Link for the slot list. */
144 	u64 S, F;		/* flow timestamps (exact) */
145 
146 	/* group we belong to. In principle we would need the index,
147 	 * which is log_2(lmax/weight), but we never reference it
148 	 * directly, only the group.
149 	 */
150 	struct qfq_group *grp;
151 
152 	/* these are copied from the flowset. */
153 	u32	class_weight; /* Weight of each class in this aggregate. */
154 	/* Max pkt size for the classes in this aggregate, DRR quantum. */
155 	int	lmax;
156 
157 	u32	inv_w;	    /* ONE_FP/(sum of weights of classes in aggr.). */
158 	u32	budgetmax;  /* Max budget for this aggregate. */
159 	u32	initial_budget, budget;     /* Initial and current budget. */
160 
161 	int		  num_classes;	/* Number of classes in this aggr. */
162 	struct list_head  active;	/* DRR queue of active classes. */
163 
164 	struct hlist_node nonfull_next;	/* See nonfull_aggs in qfq_sched. */
165 };
166 
167 struct qfq_group {
168 	u64 S, F;			/* group timestamps (approx). */
169 	unsigned int slot_shift;	/* Slot shift. */
170 	unsigned int index;		/* Group index. */
171 	unsigned int front;		/* Index of the front slot. */
172 	unsigned long full_slots;	/* non-empty slots */
173 
174 	/* Array of RR lists of active aggregates. */
175 	struct hlist_head slots[QFQ_MAX_SLOTS];
176 };
177 
178 struct qfq_sched {
179 	struct tcf_proto __rcu *filter_list;
180 	struct tcf_block	*block;
181 	struct Qdisc_class_hash clhash;
182 
183 	u64			oldV, V;	/* Precise virtual times. */
184 	struct qfq_aggregate	*in_serv_agg;   /* Aggregate being served. */
185 	u32			wsum;		/* weight sum */
186 	u32			iwsum;		/* inverse weight sum */
187 
188 	unsigned long bitmaps[QFQ_MAX_STATE];	    /* Group bitmaps. */
189 	struct qfq_group groups[QFQ_MAX_INDEX + 1]; /* The groups. */
190 	u32 min_slot_shift;	/* Index of the group-0 bit in the bitmaps. */
191 
192 	u32 max_agg_classes;		/* Max number of classes per aggr. */
193 	struct hlist_head nonfull_aggs; /* Aggs with room for more classes. */
194 };
195 
196 /*
197  * Possible reasons why the timestamps of an aggregate are updated
198  * enqueue: the aggregate switches from idle to active and must scheduled
199  *	    for service
200  * requeue: the aggregate finishes its budget, so it stops being served and
201  *	    must be rescheduled for service
202  */
203 enum update_reason {enqueue, requeue};
204 
cl_is_active(struct qfq_class * cl)205 static bool cl_is_active(struct qfq_class *cl)
206 {
207 	return !list_empty(&cl->alist);
208 }
209 
qfq_find_class(struct Qdisc * sch,u32 classid)210 static struct qfq_class *qfq_find_class(struct Qdisc *sch, u32 classid)
211 {
212 	struct qfq_sched *q = qdisc_priv(sch);
213 	struct Qdisc_class_common *clc;
214 
215 	clc = qdisc_class_find(&q->clhash, classid);
216 	if (clc == NULL)
217 		return NULL;
218 	return container_of(clc, struct qfq_class, common);
219 }
220 
221 static struct netlink_range_validation lmax_range = {
222 	.min = QFQ_MIN_LMAX,
223 	.max = QFQ_MAX_LMAX,
224 };
225 
226 static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = {
227 	[TCA_QFQ_WEIGHT] = NLA_POLICY_RANGE(NLA_U32, 1, QFQ_MAX_WEIGHT),
228 	[TCA_QFQ_LMAX] = NLA_POLICY_FULL_RANGE(NLA_U32, &lmax_range),
229 };
230 
231 /*
232  * Calculate a flow index, given its weight and maximum packet length.
233  * index = log_2(maxlen/weight) but we need to apply the scaling.
234  * This is used only once at flow creation.
235  */
qfq_calc_index(u32 inv_w,unsigned int maxlen,u32 min_slot_shift)236 static int qfq_calc_index(u32 inv_w, unsigned int maxlen, u32 min_slot_shift)
237 {
238 	u64 slot_size = (u64)maxlen * inv_w;
239 	unsigned long size_map;
240 	int index = 0;
241 
242 	size_map = slot_size >> min_slot_shift;
243 	if (!size_map)
244 		goto out;
245 
246 	index = __fls(size_map) + 1;	/* basically a log_2 */
247 	index -= !(slot_size - (1ULL << (index + min_slot_shift - 1)));
248 
249 	if (index < 0)
250 		index = 0;
251 out:
252 	pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n",
253 		 (unsigned long) ONE_FP/inv_w, maxlen, index);
254 
255 	return index;
256 }
257 
258 static void qfq_deactivate_agg(struct qfq_sched *, struct qfq_aggregate *);
259 static void qfq_activate_agg(struct qfq_sched *, struct qfq_aggregate *,
260 			     enum update_reason);
261 
qfq_init_agg(struct qfq_sched * q,struct qfq_aggregate * agg,u32 lmax,u32 weight)262 static void qfq_init_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
263 			 u32 lmax, u32 weight)
264 {
265 	INIT_LIST_HEAD(&agg->active);
266 	hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
267 
268 	agg->lmax = lmax;
269 	agg->class_weight = weight;
270 }
271 
qfq_find_agg(struct qfq_sched * q,u32 lmax,u32 weight)272 static struct qfq_aggregate *qfq_find_agg(struct qfq_sched *q,
273 					  u32 lmax, u32 weight)
274 {
275 	struct qfq_aggregate *agg;
276 
277 	hlist_for_each_entry(agg, &q->nonfull_aggs, nonfull_next)
278 		if (agg->lmax == lmax && agg->class_weight == weight)
279 			return agg;
280 
281 	return NULL;
282 }
283 
284 
285 /* 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)286 static void qfq_update_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
287 			   int new_num_classes)
288 {
289 	u32 new_agg_weight;
290 
291 	if (new_num_classes == q->max_agg_classes)
292 		hlist_del_init(&agg->nonfull_next);
293 
294 	if (agg->num_classes > new_num_classes &&
295 	    new_num_classes == q->max_agg_classes - 1) /* agg no more full */
296 		hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
297 
298 	/* The next assignment may let
299 	 * agg->initial_budget > agg->budgetmax
300 	 * hold, we will take it into account in charge_actual_service().
301 	 */
302 	agg->budgetmax = new_num_classes * agg->lmax;
303 	new_agg_weight = agg->class_weight * new_num_classes;
304 	agg->inv_w = ONE_FP/new_agg_weight;
305 
306 	if (agg->grp == NULL) {
307 		int i = qfq_calc_index(agg->inv_w, agg->budgetmax,
308 				       q->min_slot_shift);
309 		agg->grp = &q->groups[i];
310 	}
311 
312 	q->wsum +=
313 		(int) agg->class_weight * (new_num_classes - agg->num_classes);
314 	q->iwsum = ONE_FP / q->wsum;
315 
316 	agg->num_classes = new_num_classes;
317 }
318 
319 /* Add class to aggregate. */
qfq_add_to_agg(struct qfq_sched * q,struct qfq_aggregate * agg,struct qfq_class * cl)320 static void qfq_add_to_agg(struct qfq_sched *q,
321 			   struct qfq_aggregate *agg,
322 			   struct qfq_class *cl)
323 {
324 	cl->agg = agg;
325 
326 	qfq_update_agg(q, agg, agg->num_classes+1);
327 	if (cl->qdisc->q.qlen > 0) { /* adding an active class */
328 		list_add_tail(&cl->alist, &agg->active);
329 		if (list_first_entry(&agg->active, struct qfq_class, alist) ==
330 		    cl && q->in_serv_agg != agg) /* agg was inactive */
331 			qfq_activate_agg(q, agg, enqueue); /* schedule agg */
332 	}
333 }
334 
335 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *);
336 
qfq_destroy_agg(struct qfq_sched * q,struct qfq_aggregate * agg)337 static void qfq_destroy_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
338 {
339 	hlist_del_init(&agg->nonfull_next);
340 	q->wsum -= agg->class_weight;
341 	if (q->wsum != 0)
342 		q->iwsum = ONE_FP / q->wsum;
343 
344 	if (q->in_serv_agg == agg)
345 		q->in_serv_agg = qfq_choose_next_agg(q);
346 	kfree(agg);
347 }
348 
349 /* Deschedule class from within its parent aggregate. */
qfq_deactivate_class(struct qfq_sched * q,struct qfq_class * cl)350 static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
351 {
352 	struct qfq_aggregate *agg = cl->agg;
353 
354 
355 	list_del_init(&cl->alist); /* remove from RR queue of the aggregate */
356 	if (list_empty(&agg->active)) /* agg is now inactive */
357 		qfq_deactivate_agg(q, agg);
358 }
359 
360 /* Remove class from its parent aggregate. */
qfq_rm_from_agg(struct qfq_sched * q,struct qfq_class * cl)361 static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
362 {
363 	struct qfq_aggregate *agg = cl->agg;
364 
365 	cl->agg = NULL;
366 	if (agg->num_classes == 1) { /* agg being emptied, destroy it */
367 		qfq_destroy_agg(q, agg);
368 		return;
369 	}
370 	qfq_update_agg(q, agg, agg->num_classes-1);
371 }
372 
373 /* Deschedule class and remove it from its parent aggregate. */
qfq_deact_rm_from_agg(struct qfq_sched * q,struct qfq_class * cl)374 static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
375 {
376 	if (cl->qdisc->q.qlen > 0) /* class is active */
377 		qfq_deactivate_class(q, cl);
378 
379 	qfq_rm_from_agg(q, cl);
380 }
381 
382 /* 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)383 static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight,
384 			   u32 lmax)
385 {
386 	struct qfq_sched *q = qdisc_priv(sch);
387 	struct qfq_aggregate *new_agg;
388 
389 	/* 'lmax' can range from [QFQ_MIN_LMAX, pktlen + stab overhead] */
390 	if (lmax > QFQ_MAX_LMAX)
391 		return -EINVAL;
392 
393 	new_agg = qfq_find_agg(q, lmax, weight);
394 	if (new_agg == NULL) { /* create new aggregate */
395 		new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC);
396 		if (new_agg == NULL)
397 			return -ENOBUFS;
398 		qfq_init_agg(q, new_agg, lmax, weight);
399 	}
400 	qfq_deact_rm_from_agg(q, cl);
401 	qfq_add_to_agg(q, new_agg, cl);
402 
403 	return 0;
404 }
405 
qfq_change_class(struct Qdisc * sch,u32 classid,u32 parentid,struct nlattr ** tca,unsigned long * arg,struct netlink_ext_ack * extack)406 static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
407 			    struct nlattr **tca, unsigned long *arg,
408 			    struct netlink_ext_ack *extack)
409 {
410 	struct qfq_sched *q = qdisc_priv(sch);
411 	struct qfq_class *cl = (struct qfq_class *)*arg;
412 	bool existing = false;
413 	struct nlattr *tb[TCA_QFQ_MAX + 1];
414 	struct qfq_aggregate *new_agg = NULL;
415 	u32 weight, lmax, inv_w, old_weight, old_lmax;
416 	int err;
417 	int delta_w;
418 
419 	if (NL_REQ_ATTR_CHECK(extack, NULL, tca, TCA_OPTIONS)) {
420 		NL_SET_ERR_MSG_MOD(extack, "missing options");
421 		return -EINVAL;
422 	}
423 
424 	err = nla_parse_nested_deprecated(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS],
425 					  qfq_policy, extack);
426 	if (err < 0)
427 		return err;
428 
429 	if (tb[TCA_QFQ_WEIGHT])
430 		weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]);
431 	else
432 		weight = 1;
433 
434 	if (tb[TCA_QFQ_LMAX]) {
435 		lmax = nla_get_u32(tb[TCA_QFQ_LMAX]);
436 	} else {
437 		/* MTU size is user controlled */
438 		lmax = psched_mtu(qdisc_dev(sch));
439 		if (lmax < QFQ_MIN_LMAX || lmax > QFQ_MAX_LMAX) {
440 			NL_SET_ERR_MSG_MOD(extack,
441 					   "MTU size out of bounds for qfq");
442 			return -EINVAL;
443 		}
444 	}
445 
446 	inv_w = ONE_FP / weight;
447 	weight = ONE_FP / inv_w;
448 
449 	if (cl != NULL) {
450 		sch_tree_lock(sch);
451 		old_weight = cl->agg->class_weight;
452 		old_lmax   = cl->agg->lmax;
453 		sch_tree_unlock(sch);
454 		if (lmax == old_lmax && weight == old_weight)
455 			return 0; /* nothing to change */
456 	}
457 
458 	delta_w = weight - (cl ? old_weight : 0);
459 
460 	if (q->wsum + delta_w > QFQ_MAX_WSUM) {
461 		NL_SET_ERR_MSG_FMT_MOD(extack,
462 				       "total weight out of range (%d + %u)\n",
463 				       delta_w, q->wsum);
464 		return -EINVAL;
465 	}
466 
467 	if (cl != NULL) { /* modify existing class */
468 		if (tca[TCA_RATE]) {
469 			err = gen_replace_estimator(&cl->bstats, NULL,
470 						    &cl->rate_est,
471 						    NULL,
472 						    true,
473 						    tca[TCA_RATE]);
474 			if (err)
475 				return err;
476 		}
477 		existing = true;
478 		goto set_change_agg;
479 	}
480 
481 	/* create and init new class */
482 	cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
483 	if (cl == NULL)
484 		return -ENOBUFS;
485 
486 	gnet_stats_basic_sync_init(&cl->bstats);
487 	cl->common.classid = classid;
488 	cl->deficit = lmax;
489 	INIT_LIST_HEAD(&cl->alist);
490 
491 	cl->qdisc = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
492 				      classid, NULL);
493 	if (cl->qdisc == NULL)
494 		cl->qdisc = &noop_qdisc;
495 
496 	if (tca[TCA_RATE]) {
497 		err = gen_new_estimator(&cl->bstats, NULL,
498 					&cl->rate_est,
499 					NULL,
500 					true,
501 					tca[TCA_RATE]);
502 		if (err)
503 			goto destroy_class;
504 	}
505 
506 	if (cl->qdisc != &noop_qdisc)
507 		qdisc_hash_add(cl->qdisc, true);
508 
509 set_change_agg:
510 	sch_tree_lock(sch);
511 	new_agg = qfq_find_agg(q, lmax, weight);
512 	if (new_agg == NULL) { /* create new aggregate */
513 		sch_tree_unlock(sch);
514 		new_agg = kzalloc(sizeof(*new_agg), GFP_KERNEL);
515 		if (new_agg == NULL) {
516 			err = -ENOBUFS;
517 			gen_kill_estimator(&cl->rate_est);
518 			goto destroy_class;
519 		}
520 		sch_tree_lock(sch);
521 		qfq_init_agg(q, new_agg, lmax, weight);
522 	}
523 	if (existing)
524 		qfq_deact_rm_from_agg(q, cl);
525 	else
526 		qdisc_class_hash_insert(&q->clhash, &cl->common);
527 	qfq_add_to_agg(q, new_agg, cl);
528 	sch_tree_unlock(sch);
529 	qdisc_class_hash_grow(sch, &q->clhash);
530 
531 	*arg = (unsigned long)cl;
532 	return 0;
533 
534 destroy_class:
535 	qdisc_put(cl->qdisc);
536 	kfree(cl);
537 	return err;
538 }
539 
qfq_destroy_class(struct Qdisc * sch,struct qfq_class * cl)540 static void qfq_destroy_class(struct Qdisc *sch, struct qfq_class *cl)
541 {
542 	struct qfq_sched *q = qdisc_priv(sch);
543 
544 	qfq_rm_from_agg(q, cl);
545 	gen_kill_estimator(&cl->rate_est);
546 	qdisc_put(cl->qdisc);
547 	kfree(cl);
548 }
549 
qfq_delete_class(struct Qdisc * sch,unsigned long arg,struct netlink_ext_ack * extack)550 static int qfq_delete_class(struct Qdisc *sch, unsigned long arg,
551 			    struct netlink_ext_ack *extack)
552 {
553 	struct qfq_sched *q = qdisc_priv(sch);
554 	struct qfq_class *cl = (struct qfq_class *)arg;
555 
556 	if (qdisc_class_in_use(&cl->common)) {
557 		NL_SET_ERR_MSG_MOD(extack, "QFQ class in use");
558 		return -EBUSY;
559 	}
560 
561 	sch_tree_lock(sch);
562 
563 	qdisc_purge_queue(cl->qdisc);
564 	qdisc_class_hash_remove(&q->clhash, &cl->common);
565 	qfq_destroy_class(sch, cl);
566 
567 	sch_tree_unlock(sch);
568 
569 	return 0;
570 }
571 
qfq_search_class(struct Qdisc * sch,u32 classid)572 static unsigned long qfq_search_class(struct Qdisc *sch, u32 classid)
573 {
574 	return (unsigned long)qfq_find_class(sch, classid);
575 }
576 
qfq_tcf_block(struct Qdisc * sch,unsigned long cl,struct netlink_ext_ack * extack)577 static struct tcf_block *qfq_tcf_block(struct Qdisc *sch, unsigned long cl,
578 				       struct netlink_ext_ack *extack)
579 {
580 	struct qfq_sched *q = qdisc_priv(sch);
581 
582 	if (cl)
583 		return NULL;
584 
585 	return q->block;
586 }
587 
qfq_bind_tcf(struct Qdisc * sch,unsigned long parent,u32 classid)588 static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
589 				  u32 classid)
590 {
591 	struct qfq_class *cl = qfq_find_class(sch, classid);
592 
593 	if (cl)
594 		qdisc_class_get(&cl->common);
595 
596 	return (unsigned long)cl;
597 }
598 
qfq_unbind_tcf(struct Qdisc * sch,unsigned long arg)599 static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
600 {
601 	struct qfq_class *cl = (struct qfq_class *)arg;
602 
603 	qdisc_class_put(&cl->common);
604 }
605 
qfq_graft_class(struct Qdisc * sch,unsigned long arg,struct Qdisc * new,struct Qdisc ** old,struct netlink_ext_ack * extack)606 static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
607 			   struct Qdisc *new, struct Qdisc **old,
608 			   struct netlink_ext_ack *extack)
609 {
610 	struct qfq_class *cl = (struct qfq_class *)arg;
611 
612 	if (new == NULL) {
613 		new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
614 					cl->common.classid, NULL);
615 		if (new == NULL)
616 			new = &noop_qdisc;
617 	}
618 
619 	*old = qdisc_replace(sch, new, &cl->qdisc);
620 	return 0;
621 }
622 
qfq_class_leaf(struct Qdisc * sch,unsigned long arg)623 static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
624 {
625 	struct qfq_class *cl = (struct qfq_class *)arg;
626 
627 	return cl->qdisc;
628 }
629 
qfq_dump_class(struct Qdisc * sch,unsigned long arg,struct sk_buff * skb,struct tcmsg * tcm)630 static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
631 			  struct sk_buff *skb, struct tcmsg *tcm)
632 {
633 	struct qfq_class *cl = (struct qfq_class *)arg;
634 	struct nlattr *nest;
635 	u32 class_weight, lmax;
636 
637 	tcm->tcm_parent	= TC_H_ROOT;
638 	tcm->tcm_handle	= cl->common.classid;
639 	tcm->tcm_info	= cl->qdisc->handle;
640 
641 	nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
642 	if (nest == NULL)
643 		goto nla_put_failure;
644 
645 	sch_tree_lock(sch);
646 	class_weight	= cl->agg->class_weight;
647 	lmax		= cl->agg->lmax;
648 	sch_tree_unlock(sch);
649 	if (nla_put_u32(skb, TCA_QFQ_WEIGHT, class_weight) ||
650 	    nla_put_u32(skb, TCA_QFQ_LMAX, lmax))
651 		goto nla_put_failure;
652 	return nla_nest_end(skb, nest);
653 
654 nla_put_failure:
655 	nla_nest_cancel(skb, nest);
656 	return -EMSGSIZE;
657 }
658 
qfq_dump_class_stats(struct Qdisc * sch,unsigned long arg,struct gnet_dump * d)659 static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
660 				struct gnet_dump *d)
661 {
662 	struct qfq_class *cl = (struct qfq_class *)arg;
663 	struct tc_qfq_stats xstats;
664 
665 	memset(&xstats, 0, sizeof(xstats));
666 
667 	sch_tree_lock(sch);
668 	xstats.weight = cl->agg->class_weight;
669 	xstats.lmax = cl->agg->lmax;
670 	sch_tree_unlock(sch);
671 
672 	if (gnet_stats_copy_basic(d, NULL, &cl->bstats, true) < 0 ||
673 	    gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
674 	    qdisc_qstats_copy(d, cl->qdisc) < 0)
675 		return -1;
676 
677 	return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
678 }
679 
qfq_walk(struct Qdisc * sch,struct qdisc_walker * arg)680 static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
681 {
682 	struct qfq_sched *q = qdisc_priv(sch);
683 	struct qfq_class *cl;
684 	unsigned int i;
685 
686 	if (arg->stop)
687 		return;
688 
689 	for (i = 0; i < q->clhash.hashsize; i++) {
690 		hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
691 			if (!tc_qdisc_stats_dump(sch, (unsigned long)cl, arg))
692 				return;
693 		}
694 	}
695 }
696 
qfq_classify(struct sk_buff * skb,struct Qdisc * sch,int * qerr)697 static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
698 				      int *qerr)
699 {
700 	struct qfq_sched *q = qdisc_priv(sch);
701 	struct qfq_class *cl;
702 	struct tcf_result res;
703 	struct tcf_proto *fl;
704 	int result;
705 
706 	if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
707 		pr_debug("qfq_classify: found %d\n", skb->priority);
708 		cl = qfq_find_class(sch, skb->priority);
709 		if (cl != NULL)
710 			return cl;
711 	}
712 
713 	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
714 	fl = rcu_dereference_bh(q->filter_list);
715 	result = tcf_classify(skb, NULL, fl, &res, false);
716 	if (result >= 0) {
717 #ifdef CONFIG_NET_CLS_ACT
718 		switch (result) {
719 		case TC_ACT_QUEUED:
720 		case TC_ACT_STOLEN:
721 		case TC_ACT_TRAP:
722 			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
723 			fallthrough;
724 		case TC_ACT_SHOT:
725 			return NULL;
726 		}
727 #endif
728 		cl = (struct qfq_class *)res.class;
729 		if (cl == NULL)
730 			cl = qfq_find_class(sch, res.classid);
731 		return cl;
732 	}
733 
734 	return NULL;
735 }
736 
737 /* Generic comparison function, handling wraparound. */
qfq_gt(u64 a,u64 b)738 static inline int qfq_gt(u64 a, u64 b)
739 {
740 	return (s64)(a - b) > 0;
741 }
742 
743 /* Round a precise timestamp to its slotted value. */
qfq_round_down(u64 ts,unsigned int shift)744 static inline u64 qfq_round_down(u64 ts, unsigned int shift)
745 {
746 	return ts & ~((1ULL << shift) - 1);
747 }
748 
749 /* return the pointer to the group with lowest index in the bitmap */
qfq_ffs(struct qfq_sched * q,unsigned long bitmap)750 static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
751 					unsigned long bitmap)
752 {
753 	int index = __ffs(bitmap);
754 	return &q->groups[index];
755 }
756 /* Calculate a mask to mimic what would be ffs_from(). */
mask_from(unsigned long bitmap,int from)757 static inline unsigned long mask_from(unsigned long bitmap, int from)
758 {
759 	return bitmap & ~((1UL << from) - 1);
760 }
761 
762 /*
763  * The state computation relies on ER=0, IR=1, EB=2, IB=3
764  * First compute eligibility comparing grp->S, q->V,
765  * then check if someone is blocking us and possibly add EB
766  */
qfq_calc_state(struct qfq_sched * q,const struct qfq_group * grp)767 static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
768 {
769 	/* if S > V we are not eligible */
770 	unsigned int state = qfq_gt(grp->S, q->V);
771 	unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
772 	struct qfq_group *next;
773 
774 	if (mask) {
775 		next = qfq_ffs(q, mask);
776 		if (qfq_gt(grp->F, next->F))
777 			state |= EB;
778 	}
779 
780 	return state;
781 }
782 
783 
784 /*
785  * In principle
786  *	q->bitmaps[dst] |= q->bitmaps[src] & mask;
787  *	q->bitmaps[src] &= ~mask;
788  * but we should make sure that src != dst
789  */
qfq_move_groups(struct qfq_sched * q,unsigned long mask,int src,int dst)790 static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
791 				   int src, int dst)
792 {
793 	q->bitmaps[dst] |= q->bitmaps[src] & mask;
794 	q->bitmaps[src] &= ~mask;
795 }
796 
qfq_unblock_groups(struct qfq_sched * q,int index,u64 old_F)797 static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
798 {
799 	unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
800 	struct qfq_group *next;
801 
802 	if (mask) {
803 		next = qfq_ffs(q, mask);
804 		if (!qfq_gt(next->F, old_F))
805 			return;
806 	}
807 
808 	mask = (1UL << index) - 1;
809 	qfq_move_groups(q, mask, EB, ER);
810 	qfq_move_groups(q, mask, IB, IR);
811 }
812 
813 /*
814  * perhaps
815  *
816 	old_V ^= q->V;
817 	old_V >>= q->min_slot_shift;
818 	if (old_V) {
819 		...
820 	}
821  *
822  */
qfq_make_eligible(struct qfq_sched * q)823 static void qfq_make_eligible(struct qfq_sched *q)
824 {
825 	unsigned long vslot = q->V >> q->min_slot_shift;
826 	unsigned long old_vslot = q->oldV >> q->min_slot_shift;
827 
828 	if (vslot != old_vslot) {
829 		unsigned long mask;
830 		int last_flip_pos = fls(vslot ^ old_vslot);
831 
832 		if (last_flip_pos > 31) /* higher than the number of groups */
833 			mask = ~0UL;    /* make all groups eligible */
834 		else
835 			mask = (1UL << last_flip_pos) - 1;
836 
837 		qfq_move_groups(q, mask, IR, ER);
838 		qfq_move_groups(q, mask, IB, EB);
839 	}
840 }
841 
842 /*
843  * The index of the slot in which the input aggregate agg is to be
844  * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2'
845  * and not a '-1' because the start time of the group may be moved
846  * backward by one slot after the aggregate has been inserted, and
847  * this would cause non-empty slots to be right-shifted by one
848  * position.
849  *
850  * QFQ+ fully satisfies this bound to the slot index if the parameters
851  * of the classes are not changed dynamically, and if QFQ+ never
852  * happens to postpone the service of agg unjustly, i.e., it never
853  * happens that the aggregate becomes backlogged and eligible, or just
854  * eligible, while an aggregate with a higher approximated finish time
855  * is being served. In particular, in this case QFQ+ guarantees that
856  * the timestamps of agg are low enough that the slot index is never
857  * higher than 2. Unfortunately, QFQ+ cannot provide the same
858  * guarantee if it happens to unjustly postpone the service of agg, or
859  * if the parameters of some class are changed.
860  *
861  * As for the first event, i.e., an out-of-order service, the
862  * upper bound to the slot index guaranteed by QFQ+ grows to
863  * 2 +
864  * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
865  * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1.
866  *
867  * The following function deals with this problem by backward-shifting
868  * the timestamps of agg, if needed, so as to guarantee that the slot
869  * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may
870  * cause the service of other aggregates to be postponed, yet the
871  * worst-case guarantees of these aggregates are not violated.  In
872  * fact, in case of no out-of-order service, the timestamps of agg
873  * would have been even lower than they are after the backward shift,
874  * because QFQ+ would have guaranteed a maximum value equal to 2 for
875  * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose
876  * service is postponed because of the backward-shift would have
877  * however waited for the service of agg before being served.
878  *
879  * The other event that may cause the slot index to be higher than 2
880  * for agg is a recent change of the parameters of some class. If the
881  * weight of a class is increased or the lmax (max_pkt_size) of the
882  * class is decreased, then a new aggregate with smaller slot size
883  * than the original parent aggregate of the class may happen to be
884  * activated. The activation of this aggregate should be properly
885  * delayed to when the service of the class has finished in the ideal
886  * system tracked by QFQ+. If the activation of the aggregate is not
887  * delayed to this reference time instant, then this aggregate may be
888  * unjustly served before other aggregates waiting for service. This
889  * may cause the above bound to the slot index to be violated for some
890  * of these unlucky aggregates.
891  *
892  * Instead of delaying the activation of the new aggregate, which is
893  * quite complex, the above-discussed capping of the slot index is
894  * used to handle also the consequences of a change of the parameters
895  * of a class.
896  */
qfq_slot_insert(struct qfq_group * grp,struct qfq_aggregate * agg,u64 roundedS)897 static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg,
898 			    u64 roundedS)
899 {
900 	u64 slot = (roundedS - grp->S) >> grp->slot_shift;
901 	unsigned int i; /* slot index in the bucket list */
902 
903 	if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
904 		u64 deltaS = roundedS - grp->S -
905 			((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
906 		agg->S -= deltaS;
907 		agg->F -= deltaS;
908 		slot = QFQ_MAX_SLOTS - 2;
909 	}
910 
911 	i = (grp->front + slot) % QFQ_MAX_SLOTS;
912 
913 	hlist_add_head(&agg->next, &grp->slots[i]);
914 	__set_bit(slot, &grp->full_slots);
915 }
916 
917 /* Maybe introduce hlist_first_entry?? */
qfq_slot_head(struct qfq_group * grp)918 static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp)
919 {
920 	return hlist_entry(grp->slots[grp->front].first,
921 			   struct qfq_aggregate, next);
922 }
923 
924 /*
925  * remove the entry from the slot
926  */
qfq_front_slot_remove(struct qfq_group * grp)927 static void qfq_front_slot_remove(struct qfq_group *grp)
928 {
929 	struct qfq_aggregate *agg = qfq_slot_head(grp);
930 
931 	BUG_ON(!agg);
932 	hlist_del(&agg->next);
933 	if (hlist_empty(&grp->slots[grp->front]))
934 		__clear_bit(0, &grp->full_slots);
935 }
936 
937 /*
938  * Returns the first aggregate in the first non-empty bucket of the
939  * group. As a side effect, adjusts the bucket list so the first
940  * non-empty bucket is at position 0 in full_slots.
941  */
qfq_slot_scan(struct qfq_group * grp)942 static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp)
943 {
944 	unsigned int i;
945 
946 	pr_debug("qfq slot_scan: grp %u full %#lx\n",
947 		 grp->index, grp->full_slots);
948 
949 	if (grp->full_slots == 0)
950 		return NULL;
951 
952 	i = __ffs(grp->full_slots);  /* zero based */
953 	if (i > 0) {
954 		grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
955 		grp->full_slots >>= i;
956 	}
957 
958 	return qfq_slot_head(grp);
959 }
960 
961 /*
962  * adjust the bucket list. When the start time of a group decreases,
963  * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
964  * move the objects. The mask of occupied slots must be shifted
965  * because we use ffs() to find the first non-empty slot.
966  * This covers decreases in the group's start time, but what about
967  * increases of the start time ?
968  * Here too we should make sure that i is less than 32
969  */
qfq_slot_rotate(struct qfq_group * grp,u64 roundedS)970 static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
971 {
972 	unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
973 
974 	grp->full_slots <<= i;
975 	grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
976 }
977 
qfq_update_eligible(struct qfq_sched * q)978 static void qfq_update_eligible(struct qfq_sched *q)
979 {
980 	struct qfq_group *grp;
981 	unsigned long ineligible;
982 
983 	ineligible = q->bitmaps[IR] | q->bitmaps[IB];
984 	if (ineligible) {
985 		if (!q->bitmaps[ER]) {
986 			grp = qfq_ffs(q, ineligible);
987 			if (qfq_gt(grp->S, q->V))
988 				q->V = grp->S;
989 		}
990 		qfq_make_eligible(q);
991 	}
992 }
993 
994 /* 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)995 static struct sk_buff *agg_dequeue(struct qfq_aggregate *agg,
996 				   struct qfq_class *cl, unsigned int len)
997 {
998 	struct sk_buff *skb = qdisc_dequeue_peeked(cl->qdisc);
999 
1000 	if (!skb)
1001 		return NULL;
1002 
1003 	cl->deficit -= (int) len;
1004 
1005 	if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */
1006 		list_del_init(&cl->alist);
1007 	else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) {
1008 		cl->deficit += agg->lmax;
1009 		list_move_tail(&cl->alist, &agg->active);
1010 	}
1011 
1012 	return skb;
1013 }
1014 
qfq_peek_skb(struct qfq_aggregate * agg,struct qfq_class ** cl,unsigned int * len)1015 static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg,
1016 					   struct qfq_class **cl,
1017 					   unsigned int *len)
1018 {
1019 	struct sk_buff *skb;
1020 
1021 	*cl = list_first_entry(&agg->active, struct qfq_class, alist);
1022 	skb = (*cl)->qdisc->ops->peek((*cl)->qdisc);
1023 	if (skb == NULL)
1024 		WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
1025 	else
1026 		*len = qdisc_pkt_len(skb);
1027 
1028 	return skb;
1029 }
1030 
1031 /* Update F according to the actual service received by the aggregate. */
charge_actual_service(struct qfq_aggregate * agg)1032 static inline void charge_actual_service(struct qfq_aggregate *agg)
1033 {
1034 	/* Compute the service received by the aggregate, taking into
1035 	 * account that, after decreasing the number of classes in
1036 	 * agg, it may happen that
1037 	 * agg->initial_budget - agg->budget > agg->bugdetmax
1038 	 */
1039 	u32 service_received = min(agg->budgetmax,
1040 				   agg->initial_budget - agg->budget);
1041 
1042 	agg->F = agg->S + (u64)service_received * agg->inv_w;
1043 }
1044 
1045 /* Assign a reasonable start time for a new aggregate in group i.
1046  * Admissible values for \hat(F) are multiples of \sigma_i
1047  * no greater than V+\sigma_i . Larger values mean that
1048  * we had a wraparound so we consider the timestamp to be stale.
1049  *
1050  * If F is not stale and F >= V then we set S = F.
1051  * Otherwise we should assign S = V, but this may violate
1052  * the ordering in EB (see [2]). So, if we have groups in ER,
1053  * set S to the F_j of the first group j which would be blocking us.
1054  * We are guaranteed not to move S backward because
1055  * otherwise our group i would still be blocked.
1056  */
qfq_update_start(struct qfq_sched * q,struct qfq_aggregate * agg)1057 static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg)
1058 {
1059 	unsigned long mask;
1060 	u64 limit, roundedF;
1061 	int slot_shift = agg->grp->slot_shift;
1062 
1063 	roundedF = qfq_round_down(agg->F, slot_shift);
1064 	limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
1065 
1066 	if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) {
1067 		/* timestamp was stale */
1068 		mask = mask_from(q->bitmaps[ER], agg->grp->index);
1069 		if (mask) {
1070 			struct qfq_group *next = qfq_ffs(q, mask);
1071 			if (qfq_gt(roundedF, next->F)) {
1072 				if (qfq_gt(limit, next->F))
1073 					agg->S = next->F;
1074 				else /* preserve timestamp correctness */
1075 					agg->S = limit;
1076 				return;
1077 			}
1078 		}
1079 		agg->S = q->V;
1080 	} else  /* timestamp is not stale */
1081 		agg->S = agg->F;
1082 }
1083 
1084 /* Update the timestamps of agg before scheduling/rescheduling it for
1085  * service.  In particular, assign to agg->F its maximum possible
1086  * value, i.e., the virtual finish time with which the aggregate
1087  * should be labeled if it used all its budget once in service.
1088  */
1089 static inline void
qfq_update_agg_ts(struct qfq_sched * q,struct qfq_aggregate * agg,enum update_reason reason)1090 qfq_update_agg_ts(struct qfq_sched *q,
1091 		    struct qfq_aggregate *agg, enum update_reason reason)
1092 {
1093 	if (reason != requeue)
1094 		qfq_update_start(q, agg);
1095 	else /* just charge agg for the service received */
1096 		agg->S = agg->F;
1097 
1098 	agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w;
1099 }
1100 
1101 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg);
1102 
qfq_dequeue(struct Qdisc * sch)1103 static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
1104 {
1105 	struct qfq_sched *q = qdisc_priv(sch);
1106 	struct qfq_aggregate *in_serv_agg = q->in_serv_agg;
1107 	struct qfq_class *cl;
1108 	struct sk_buff *skb = NULL;
1109 	/* next-packet len, 0 means no more active classes in in-service agg */
1110 	unsigned int len = 0;
1111 
1112 	if (in_serv_agg == NULL)
1113 		return NULL;
1114 
1115 	if (!list_empty(&in_serv_agg->active))
1116 		skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1117 
1118 	/*
1119 	 * If there are no active classes in the in-service aggregate,
1120 	 * or if the aggregate has not enough budget to serve its next
1121 	 * class, then choose the next aggregate to serve.
1122 	 */
1123 	if (len == 0 || in_serv_agg->budget < len) {
1124 		charge_actual_service(in_serv_agg);
1125 
1126 		/* recharge the budget of the aggregate */
1127 		in_serv_agg->initial_budget = in_serv_agg->budget =
1128 			in_serv_agg->budgetmax;
1129 
1130 		if (!list_empty(&in_serv_agg->active)) {
1131 			/*
1132 			 * Still active: reschedule for
1133 			 * service. Possible optimization: if no other
1134 			 * aggregate is active, then there is no point
1135 			 * in rescheduling this aggregate, and we can
1136 			 * just keep it as the in-service one. This
1137 			 * should be however a corner case, and to
1138 			 * handle it, we would need to maintain an
1139 			 * extra num_active_aggs field.
1140 			*/
1141 			qfq_update_agg_ts(q, in_serv_agg, requeue);
1142 			qfq_schedule_agg(q, in_serv_agg);
1143 		} else if (sch->q.qlen == 0) { /* no aggregate to serve */
1144 			q->in_serv_agg = NULL;
1145 			return NULL;
1146 		}
1147 
1148 		/*
1149 		 * If we get here, there are other aggregates queued:
1150 		 * choose the new aggregate to serve.
1151 		 */
1152 		in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
1153 		skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1154 	}
1155 	if (!skb)
1156 		return NULL;
1157 
1158 	sch->q.qlen--;
1159 
1160 	skb = agg_dequeue(in_serv_agg, cl, len);
1161 
1162 	if (!skb) {
1163 		sch->q.qlen++;
1164 		return NULL;
1165 	}
1166 
1167 	qdisc_qstats_backlog_dec(sch, skb);
1168 	qdisc_bstats_update(sch, skb);
1169 
1170 	/* If lmax is lowered, through qfq_change_class, for a class
1171 	 * owning pending packets with larger size than the new value
1172 	 * of lmax, then the following condition may hold.
1173 	 */
1174 	if (unlikely(in_serv_agg->budget < len))
1175 		in_serv_agg->budget = 0;
1176 	else
1177 		in_serv_agg->budget -= len;
1178 
1179 	q->V += (u64)len * q->iwsum;
1180 	pr_debug("qfq dequeue: len %u F %lld now %lld\n",
1181 		 len, (unsigned long long) in_serv_agg->F,
1182 		 (unsigned long long) q->V);
1183 
1184 	return skb;
1185 }
1186 
qfq_choose_next_agg(struct qfq_sched * q)1187 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)
1188 {
1189 	struct qfq_group *grp;
1190 	struct qfq_aggregate *agg, *new_front_agg;
1191 	u64 old_F;
1192 
1193 	qfq_update_eligible(q);
1194 	q->oldV = q->V;
1195 
1196 	if (!q->bitmaps[ER])
1197 		return NULL;
1198 
1199 	grp = qfq_ffs(q, q->bitmaps[ER]);
1200 	old_F = grp->F;
1201 
1202 	agg = qfq_slot_head(grp);
1203 
1204 	/* agg starts to be served, remove it from schedule */
1205 	qfq_front_slot_remove(grp);
1206 
1207 	new_front_agg = qfq_slot_scan(grp);
1208 
1209 	if (new_front_agg == NULL) /* group is now inactive, remove from ER */
1210 		__clear_bit(grp->index, &q->bitmaps[ER]);
1211 	else {
1212 		u64 roundedS = qfq_round_down(new_front_agg->S,
1213 					      grp->slot_shift);
1214 		unsigned int s;
1215 
1216 		if (grp->S == roundedS)
1217 			return agg;
1218 		grp->S = roundedS;
1219 		grp->F = roundedS + (2ULL << grp->slot_shift);
1220 		__clear_bit(grp->index, &q->bitmaps[ER]);
1221 		s = qfq_calc_state(q, grp);
1222 		__set_bit(grp->index, &q->bitmaps[s]);
1223 	}
1224 
1225 	qfq_unblock_groups(q, grp->index, old_F);
1226 
1227 	return agg;
1228 }
1229 
qfq_enqueue(struct sk_buff * skb,struct Qdisc * sch,struct sk_buff ** to_free)1230 static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
1231 		       struct sk_buff **to_free)
1232 {
1233 	unsigned int len = qdisc_pkt_len(skb), gso_segs;
1234 	struct qfq_sched *q = qdisc_priv(sch);
1235 	struct qfq_class *cl;
1236 	struct qfq_aggregate *agg;
1237 	int err = 0;
1238 
1239 	cl = qfq_classify(skb, sch, &err);
1240 	if (cl == NULL) {
1241 		if (err & __NET_XMIT_BYPASS)
1242 			qdisc_qstats_drop(sch);
1243 		__qdisc_drop(skb, to_free);
1244 		return err;
1245 	}
1246 	pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
1247 
1248 	if (unlikely(cl->agg->lmax < len)) {
1249 		pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
1250 			 cl->agg->lmax, len, cl->common.classid);
1251 		err = qfq_change_agg(sch, cl, cl->agg->class_weight, len);
1252 		if (err) {
1253 			cl->qstats.drops++;
1254 			return qdisc_drop(skb, sch, to_free);
1255 		}
1256 	}
1257 
1258 	gso_segs = skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 1;
1259 	err = qdisc_enqueue(skb, cl->qdisc, to_free);
1260 	if (unlikely(err != NET_XMIT_SUCCESS)) {
1261 		pr_debug("qfq_enqueue: enqueue failed %d\n", err);
1262 		if (net_xmit_drop_count(err)) {
1263 			cl->qstats.drops++;
1264 			qdisc_qstats_drop(sch);
1265 		}
1266 		return err;
1267 	}
1268 
1269 	_bstats_update(&cl->bstats, len, gso_segs);
1270 	sch->qstats.backlog += len;
1271 	++sch->q.qlen;
1272 
1273 	agg = cl->agg;
1274 	/* if the class is active, then done here */
1275 	if (cl_is_active(cl)) {
1276 		if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) &&
1277 		    list_first_entry(&agg->active, struct qfq_class, alist)
1278 		    == cl && cl->deficit < len)
1279 			list_move_tail(&cl->alist, &agg->active);
1280 
1281 		return err;
1282 	}
1283 
1284 	/* schedule class for service within the aggregate */
1285 	cl->deficit = agg->lmax;
1286 	list_add_tail(&cl->alist, &agg->active);
1287 
1288 	if (list_first_entry(&agg->active, struct qfq_class, alist) != cl ||
1289 	    q->in_serv_agg == agg)
1290 		return err; /* non-empty or in service, nothing else to do */
1291 
1292 	qfq_activate_agg(q, agg, enqueue);
1293 
1294 	return err;
1295 }
1296 
1297 /*
1298  * Schedule aggregate according to its timestamps.
1299  */
qfq_schedule_agg(struct qfq_sched * q,struct qfq_aggregate * agg)1300 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1301 {
1302 	struct qfq_group *grp = agg->grp;
1303 	u64 roundedS;
1304 	int s;
1305 
1306 	roundedS = qfq_round_down(agg->S, grp->slot_shift);
1307 
1308 	/*
1309 	 * Insert agg in the correct bucket.
1310 	 * If agg->S >= grp->S we don't need to adjust the
1311 	 * bucket list and simply go to the insertion phase.
1312 	 * Otherwise grp->S is decreasing, we must make room
1313 	 * in the bucket list, and also recompute the group state.
1314 	 * Finally, if there were no flows in this group and nobody
1315 	 * was in ER make sure to adjust V.
1316 	 */
1317 	if (grp->full_slots) {
1318 		if (!qfq_gt(grp->S, agg->S))
1319 			goto skip_update;
1320 
1321 		/* create a slot for this agg->S */
1322 		qfq_slot_rotate(grp, roundedS);
1323 		/* group was surely ineligible, remove */
1324 		__clear_bit(grp->index, &q->bitmaps[IR]);
1325 		__clear_bit(grp->index, &q->bitmaps[IB]);
1326 	} else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) &&
1327 		   q->in_serv_agg == NULL)
1328 		q->V = roundedS;
1329 
1330 	grp->S = roundedS;
1331 	grp->F = roundedS + (2ULL << grp->slot_shift);
1332 	s = qfq_calc_state(q, grp);
1333 	__set_bit(grp->index, &q->bitmaps[s]);
1334 
1335 	pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
1336 		 s, q->bitmaps[s],
1337 		 (unsigned long long) agg->S,
1338 		 (unsigned long long) agg->F,
1339 		 (unsigned long long) q->V);
1340 
1341 skip_update:
1342 	qfq_slot_insert(grp, agg, roundedS);
1343 }
1344 
1345 
1346 /* Update agg ts and schedule agg for service */
qfq_activate_agg(struct qfq_sched * q,struct qfq_aggregate * agg,enum update_reason reason)1347 static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
1348 			     enum update_reason reason)
1349 {
1350 	agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */
1351 
1352 	qfq_update_agg_ts(q, agg, reason);
1353 	if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */
1354 		q->in_serv_agg = agg; /* start serving this aggregate */
1355 		 /* update V: to be in service, agg must be eligible */
1356 		q->oldV = q->V = agg->S;
1357 	} else if (agg != q->in_serv_agg)
1358 		qfq_schedule_agg(q, agg);
1359 }
1360 
qfq_slot_remove(struct qfq_sched * q,struct qfq_group * grp,struct qfq_aggregate * agg)1361 static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
1362 			    struct qfq_aggregate *agg)
1363 {
1364 	unsigned int i, offset;
1365 	u64 roundedS;
1366 
1367 	roundedS = qfq_round_down(agg->S, grp->slot_shift);
1368 	offset = (roundedS - grp->S) >> grp->slot_shift;
1369 
1370 	i = (grp->front + offset) % QFQ_MAX_SLOTS;
1371 
1372 	hlist_del(&agg->next);
1373 	if (hlist_empty(&grp->slots[i]))
1374 		__clear_bit(offset, &grp->full_slots);
1375 }
1376 
1377 /*
1378  * Called to forcibly deschedule an aggregate.  If the aggregate is
1379  * not in the front bucket, or if the latter has other aggregates in
1380  * the front bucket, we can simply remove the aggregate with no other
1381  * side effects.
1382  * Otherwise we must propagate the event up.
1383  */
qfq_deactivate_agg(struct qfq_sched * q,struct qfq_aggregate * agg)1384 static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1385 {
1386 	struct qfq_group *grp = agg->grp;
1387 	unsigned long mask;
1388 	u64 roundedS;
1389 	int s;
1390 
1391 	if (agg == q->in_serv_agg) {
1392 		charge_actual_service(agg);
1393 		q->in_serv_agg = qfq_choose_next_agg(q);
1394 		return;
1395 	}
1396 
1397 	agg->F = agg->S;
1398 	qfq_slot_remove(q, grp, agg);
1399 
1400 	if (!grp->full_slots) {
1401 		__clear_bit(grp->index, &q->bitmaps[IR]);
1402 		__clear_bit(grp->index, &q->bitmaps[EB]);
1403 		__clear_bit(grp->index, &q->bitmaps[IB]);
1404 
1405 		if (test_bit(grp->index, &q->bitmaps[ER]) &&
1406 		    !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
1407 			mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
1408 			if (mask)
1409 				mask = ~((1UL << __fls(mask)) - 1);
1410 			else
1411 				mask = ~0UL;
1412 			qfq_move_groups(q, mask, EB, ER);
1413 			qfq_move_groups(q, mask, IB, IR);
1414 		}
1415 		__clear_bit(grp->index, &q->bitmaps[ER]);
1416 	} else if (hlist_empty(&grp->slots[grp->front])) {
1417 		agg = qfq_slot_scan(grp);
1418 		roundedS = qfq_round_down(agg->S, grp->slot_shift);
1419 		if (grp->S != roundedS) {
1420 			__clear_bit(grp->index, &q->bitmaps[ER]);
1421 			__clear_bit(grp->index, &q->bitmaps[IR]);
1422 			__clear_bit(grp->index, &q->bitmaps[EB]);
1423 			__clear_bit(grp->index, &q->bitmaps[IB]);
1424 			grp->S = roundedS;
1425 			grp->F = roundedS + (2ULL << grp->slot_shift);
1426 			s = qfq_calc_state(q, grp);
1427 			__set_bit(grp->index, &q->bitmaps[s]);
1428 		}
1429 	}
1430 }
1431 
qfq_qlen_notify(struct Qdisc * sch,unsigned long arg)1432 static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1433 {
1434 	struct qfq_sched *q = qdisc_priv(sch);
1435 	struct qfq_class *cl = (struct qfq_class *)arg;
1436 
1437 	if (list_empty(&cl->alist))
1438 		return;
1439 	qfq_deactivate_class(q, cl);
1440 }
1441 
qfq_init_qdisc(struct Qdisc * sch,struct nlattr * opt,struct netlink_ext_ack * extack)1442 static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt,
1443 			  struct netlink_ext_ack *extack)
1444 {
1445 	struct qfq_sched *q = qdisc_priv(sch);
1446 	struct qfq_group *grp;
1447 	int i, j, err;
1448 	u32 max_cl_shift, maxbudg_shift, max_classes;
1449 
1450 	err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
1451 	if (err)
1452 		return err;
1453 
1454 	err = qdisc_class_hash_init(&q->clhash);
1455 	if (err < 0)
1456 		return err;
1457 
1458 	max_classes = min_t(u64, (u64)qdisc_dev(sch)->tx_queue_len + 1,
1459 			    QFQ_MAX_AGG_CLASSES);
1460 	/* max_cl_shift = floor(log_2(max_classes)) */
1461 	max_cl_shift = __fls(max_classes);
1462 	q->max_agg_classes = 1<<max_cl_shift;
1463 
1464 	/* maxbudg_shift = log2(max_len * max_classes_per_agg) */
1465 	maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift;
1466 	q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX;
1467 
1468 	for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1469 		grp = &q->groups[i];
1470 		grp->index = i;
1471 		grp->slot_shift = q->min_slot_shift + i;
1472 		for (j = 0; j < QFQ_MAX_SLOTS; j++)
1473 			INIT_HLIST_HEAD(&grp->slots[j]);
1474 	}
1475 
1476 	INIT_HLIST_HEAD(&q->nonfull_aggs);
1477 
1478 	return 0;
1479 }
1480 
qfq_reset_qdisc(struct Qdisc * sch)1481 static void qfq_reset_qdisc(struct Qdisc *sch)
1482 {
1483 	struct qfq_sched *q = qdisc_priv(sch);
1484 	struct qfq_class *cl;
1485 	unsigned int i;
1486 
1487 	for (i = 0; i < q->clhash.hashsize; i++) {
1488 		hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1489 			if (cl->qdisc->q.qlen > 0)
1490 				qfq_deactivate_class(q, cl);
1491 
1492 			qdisc_reset(cl->qdisc);
1493 		}
1494 	}
1495 }
1496 
qfq_destroy_qdisc(struct Qdisc * sch)1497 static void qfq_destroy_qdisc(struct Qdisc *sch)
1498 {
1499 	struct qfq_sched *q = qdisc_priv(sch);
1500 	struct qfq_class *cl;
1501 	struct hlist_node *next;
1502 	unsigned int i;
1503 
1504 	tcf_block_put(q->block);
1505 
1506 	for (i = 0; i < q->clhash.hashsize; i++) {
1507 		hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1508 					  common.hnode) {
1509 			qfq_destroy_class(sch, cl);
1510 		}
1511 	}
1512 	qdisc_class_hash_destroy(&q->clhash);
1513 }
1514 
1515 static const struct Qdisc_class_ops qfq_class_ops = {
1516 	.change		= qfq_change_class,
1517 	.delete		= qfq_delete_class,
1518 	.find		= qfq_search_class,
1519 	.tcf_block	= qfq_tcf_block,
1520 	.bind_tcf	= qfq_bind_tcf,
1521 	.unbind_tcf	= qfq_unbind_tcf,
1522 	.graft		= qfq_graft_class,
1523 	.leaf		= qfq_class_leaf,
1524 	.qlen_notify	= qfq_qlen_notify,
1525 	.dump		= qfq_dump_class,
1526 	.dump_stats	= qfq_dump_class_stats,
1527 	.walk		= qfq_walk,
1528 };
1529 
1530 static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1531 	.cl_ops		= &qfq_class_ops,
1532 	.id		= "qfq",
1533 	.priv_size	= sizeof(struct qfq_sched),
1534 	.enqueue	= qfq_enqueue,
1535 	.dequeue	= qfq_dequeue,
1536 	.peek		= qdisc_peek_dequeued,
1537 	.init		= qfq_init_qdisc,
1538 	.reset		= qfq_reset_qdisc,
1539 	.destroy	= qfq_destroy_qdisc,
1540 	.owner		= THIS_MODULE,
1541 };
1542 
qfq_init(void)1543 static int __init qfq_init(void)
1544 {
1545 	return register_qdisc(&qfq_qdisc_ops);
1546 }
1547 
qfq_exit(void)1548 static void __exit qfq_exit(void)
1549 {
1550 	unregister_qdisc(&qfq_qdisc_ops);
1551 }
1552 
1553 module_init(qfq_init);
1554 module_exit(qfq_exit);
1555 MODULE_LICENSE("GPL");
1556