xref: /openbmc/linux/net/sched/sch_qfq.c (revision 8fc134fe)
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 
qfq_find_class(struct Qdisc * sch,u32 classid)205 static struct qfq_class *qfq_find_class(struct Qdisc *sch, u32 classid)
206 {
207 	struct qfq_sched *q = qdisc_priv(sch);
208 	struct Qdisc_class_common *clc;
209 
210 	clc = qdisc_class_find(&q->clhash, classid);
211 	if (clc == NULL)
212 		return NULL;
213 	return container_of(clc, struct qfq_class, common);
214 }
215 
216 static struct netlink_range_validation lmax_range = {
217 	.min = QFQ_MIN_LMAX,
218 	.max = QFQ_MAX_LMAX,
219 };
220 
221 static const struct nla_policy qfq_policy[TCA_QFQ_MAX + 1] = {
222 	[TCA_QFQ_WEIGHT] = NLA_POLICY_RANGE(NLA_U32, 1, QFQ_MAX_WEIGHT),
223 	[TCA_QFQ_LMAX] = NLA_POLICY_FULL_RANGE(NLA_U32, &lmax_range),
224 };
225 
226 /*
227  * Calculate a flow index, given its weight and maximum packet length.
228  * index = log_2(maxlen/weight) but we need to apply the scaling.
229  * This is used only once at flow creation.
230  */
qfq_calc_index(u32 inv_w,unsigned int maxlen,u32 min_slot_shift)231 static int qfq_calc_index(u32 inv_w, unsigned int maxlen, u32 min_slot_shift)
232 {
233 	u64 slot_size = (u64)maxlen * inv_w;
234 	unsigned long size_map;
235 	int index = 0;
236 
237 	size_map = slot_size >> min_slot_shift;
238 	if (!size_map)
239 		goto out;
240 
241 	index = __fls(size_map) + 1;	/* basically a log_2 */
242 	index -= !(slot_size - (1ULL << (index + min_slot_shift - 1)));
243 
244 	if (index < 0)
245 		index = 0;
246 out:
247 	pr_debug("qfq calc_index: W = %lu, L = %u, I = %d\n",
248 		 (unsigned long) ONE_FP/inv_w, maxlen, index);
249 
250 	return index;
251 }
252 
253 static void qfq_deactivate_agg(struct qfq_sched *, struct qfq_aggregate *);
254 static void qfq_activate_agg(struct qfq_sched *, struct qfq_aggregate *,
255 			     enum update_reason);
256 
qfq_init_agg(struct qfq_sched * q,struct qfq_aggregate * agg,u32 lmax,u32 weight)257 static void qfq_init_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
258 			 u32 lmax, u32 weight)
259 {
260 	INIT_LIST_HEAD(&agg->active);
261 	hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
262 
263 	agg->lmax = lmax;
264 	agg->class_weight = weight;
265 }
266 
qfq_find_agg(struct qfq_sched * q,u32 lmax,u32 weight)267 static struct qfq_aggregate *qfq_find_agg(struct qfq_sched *q,
268 					  u32 lmax, u32 weight)
269 {
270 	struct qfq_aggregate *agg;
271 
272 	hlist_for_each_entry(agg, &q->nonfull_aggs, nonfull_next)
273 		if (agg->lmax == lmax && agg->class_weight == weight)
274 			return agg;
275 
276 	return NULL;
277 }
278 
279 
280 /* 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)281 static void qfq_update_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
282 			   int new_num_classes)
283 {
284 	u32 new_agg_weight;
285 
286 	if (new_num_classes == q->max_agg_classes)
287 		hlist_del_init(&agg->nonfull_next);
288 
289 	if (agg->num_classes > new_num_classes &&
290 	    new_num_classes == q->max_agg_classes - 1) /* agg no more full */
291 		hlist_add_head(&agg->nonfull_next, &q->nonfull_aggs);
292 
293 	/* The next assignment may let
294 	 * agg->initial_budget > agg->budgetmax
295 	 * hold, we will take it into account in charge_actual_service().
296 	 */
297 	agg->budgetmax = new_num_classes * agg->lmax;
298 	new_agg_weight = agg->class_weight * new_num_classes;
299 	agg->inv_w = ONE_FP/new_agg_weight;
300 
301 	if (agg->grp == NULL) {
302 		int i = qfq_calc_index(agg->inv_w, agg->budgetmax,
303 				       q->min_slot_shift);
304 		agg->grp = &q->groups[i];
305 	}
306 
307 	q->wsum +=
308 		(int) agg->class_weight * (new_num_classes - agg->num_classes);
309 	q->iwsum = ONE_FP / q->wsum;
310 
311 	agg->num_classes = new_num_classes;
312 }
313 
314 /* Add class to aggregate. */
qfq_add_to_agg(struct qfq_sched * q,struct qfq_aggregate * agg,struct qfq_class * cl)315 static void qfq_add_to_agg(struct qfq_sched *q,
316 			   struct qfq_aggregate *agg,
317 			   struct qfq_class *cl)
318 {
319 	cl->agg = agg;
320 
321 	qfq_update_agg(q, agg, agg->num_classes+1);
322 	if (cl->qdisc->q.qlen > 0) { /* adding an active class */
323 		list_add_tail(&cl->alist, &agg->active);
324 		if (list_first_entry(&agg->active, struct qfq_class, alist) ==
325 		    cl && q->in_serv_agg != agg) /* agg was inactive */
326 			qfq_activate_agg(q, agg, enqueue); /* schedule agg */
327 	}
328 }
329 
330 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *);
331 
qfq_destroy_agg(struct qfq_sched * q,struct qfq_aggregate * agg)332 static void qfq_destroy_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
333 {
334 	hlist_del_init(&agg->nonfull_next);
335 	q->wsum -= agg->class_weight;
336 	if (q->wsum != 0)
337 		q->iwsum = ONE_FP / q->wsum;
338 
339 	if (q->in_serv_agg == agg)
340 		q->in_serv_agg = qfq_choose_next_agg(q);
341 	kfree(agg);
342 }
343 
344 /* Deschedule class from within its parent aggregate. */
qfq_deactivate_class(struct qfq_sched * q,struct qfq_class * cl)345 static void qfq_deactivate_class(struct qfq_sched *q, struct qfq_class *cl)
346 {
347 	struct qfq_aggregate *agg = cl->agg;
348 
349 
350 	list_del(&cl->alist); /* remove from RR queue of the aggregate */
351 	if (list_empty(&agg->active)) /* agg is now inactive */
352 		qfq_deactivate_agg(q, agg);
353 }
354 
355 /* Remove class from its parent aggregate. */
qfq_rm_from_agg(struct qfq_sched * q,struct qfq_class * cl)356 static void qfq_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
357 {
358 	struct qfq_aggregate *agg = cl->agg;
359 
360 	cl->agg = NULL;
361 	if (agg->num_classes == 1) { /* agg being emptied, destroy it */
362 		qfq_destroy_agg(q, agg);
363 		return;
364 	}
365 	qfq_update_agg(q, agg, agg->num_classes-1);
366 }
367 
368 /* Deschedule class and remove it from its parent aggregate. */
qfq_deact_rm_from_agg(struct qfq_sched * q,struct qfq_class * cl)369 static void qfq_deact_rm_from_agg(struct qfq_sched *q, struct qfq_class *cl)
370 {
371 	if (cl->qdisc->q.qlen > 0) /* class is active */
372 		qfq_deactivate_class(q, cl);
373 
374 	qfq_rm_from_agg(q, cl);
375 }
376 
377 /* 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)378 static int qfq_change_agg(struct Qdisc *sch, struct qfq_class *cl, u32 weight,
379 			   u32 lmax)
380 {
381 	struct qfq_sched *q = qdisc_priv(sch);
382 	struct qfq_aggregate *new_agg;
383 
384 	/* 'lmax' can range from [QFQ_MIN_LMAX, pktlen + stab overhead] */
385 	if (lmax > QFQ_MAX_LMAX)
386 		return -EINVAL;
387 
388 	new_agg = qfq_find_agg(q, lmax, weight);
389 	if (new_agg == NULL) { /* create new aggregate */
390 		new_agg = kzalloc(sizeof(*new_agg), GFP_ATOMIC);
391 		if (new_agg == NULL)
392 			return -ENOBUFS;
393 		qfq_init_agg(q, new_agg, lmax, weight);
394 	}
395 	qfq_deact_rm_from_agg(q, cl);
396 	qfq_add_to_agg(q, new_agg, cl);
397 
398 	return 0;
399 }
400 
qfq_change_class(struct Qdisc * sch,u32 classid,u32 parentid,struct nlattr ** tca,unsigned long * arg,struct netlink_ext_ack * extack)401 static int qfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
402 			    struct nlattr **tca, unsigned long *arg,
403 			    struct netlink_ext_ack *extack)
404 {
405 	struct qfq_sched *q = qdisc_priv(sch);
406 	struct qfq_class *cl = (struct qfq_class *)*arg;
407 	bool existing = false;
408 	struct nlattr *tb[TCA_QFQ_MAX + 1];
409 	struct qfq_aggregate *new_agg = NULL;
410 	u32 weight, lmax, inv_w;
411 	int err;
412 	int delta_w;
413 
414 	if (NL_REQ_ATTR_CHECK(extack, NULL, tca, TCA_OPTIONS)) {
415 		NL_SET_ERR_MSG_MOD(extack, "missing options");
416 		return -EINVAL;
417 	}
418 
419 	err = nla_parse_nested_deprecated(tb, TCA_QFQ_MAX, tca[TCA_OPTIONS],
420 					  qfq_policy, extack);
421 	if (err < 0)
422 		return err;
423 
424 	if (tb[TCA_QFQ_WEIGHT])
425 		weight = nla_get_u32(tb[TCA_QFQ_WEIGHT]);
426 	else
427 		weight = 1;
428 
429 	if (tb[TCA_QFQ_LMAX]) {
430 		lmax = nla_get_u32(tb[TCA_QFQ_LMAX]);
431 	} else {
432 		/* MTU size is user controlled */
433 		lmax = psched_mtu(qdisc_dev(sch));
434 		if (lmax < QFQ_MIN_LMAX || lmax > QFQ_MAX_LMAX) {
435 			NL_SET_ERR_MSG_MOD(extack,
436 					   "MTU size out of bounds for qfq");
437 			return -EINVAL;
438 		}
439 	}
440 
441 	inv_w = ONE_FP / weight;
442 	weight = ONE_FP / inv_w;
443 
444 	if (cl != NULL &&
445 	    lmax == cl->agg->lmax &&
446 	    weight == cl->agg->class_weight)
447 		return 0; /* nothing to change */
448 
449 	delta_w = weight - (cl ? cl->agg->class_weight : 0);
450 
451 	if (q->wsum + delta_w > QFQ_MAX_WSUM) {
452 		NL_SET_ERR_MSG_FMT_MOD(extack,
453 				       "total weight out of range (%d + %u)\n",
454 				       delta_w, q->wsum);
455 		return -EINVAL;
456 	}
457 
458 	if (cl != NULL) { /* modify existing class */
459 		if (tca[TCA_RATE]) {
460 			err = gen_replace_estimator(&cl->bstats, NULL,
461 						    &cl->rate_est,
462 						    NULL,
463 						    true,
464 						    tca[TCA_RATE]);
465 			if (err)
466 				return err;
467 		}
468 		existing = true;
469 		goto set_change_agg;
470 	}
471 
472 	/* create and init new class */
473 	cl = kzalloc(sizeof(struct qfq_class), GFP_KERNEL);
474 	if (cl == NULL)
475 		return -ENOBUFS;
476 
477 	gnet_stats_basic_sync_init(&cl->bstats);
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 					true,
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 (qdisc_class_in_use(&cl->common)) {
547 		NL_SET_ERR_MSG_MOD(extack, "QFQ class in use");
548 		return -EBUSY;
549 	}
550 
551 	sch_tree_lock(sch);
552 
553 	qdisc_purge_queue(cl->qdisc);
554 	qdisc_class_hash_remove(&q->clhash, &cl->common);
555 
556 	sch_tree_unlock(sch);
557 
558 	qfq_destroy_class(sch, cl);
559 	return 0;
560 }
561 
qfq_search_class(struct Qdisc * sch,u32 classid)562 static unsigned long qfq_search_class(struct Qdisc *sch, u32 classid)
563 {
564 	return (unsigned long)qfq_find_class(sch, classid);
565 }
566 
qfq_tcf_block(struct Qdisc * sch,unsigned long cl,struct netlink_ext_ack * extack)567 static struct tcf_block *qfq_tcf_block(struct Qdisc *sch, unsigned long cl,
568 				       struct netlink_ext_ack *extack)
569 {
570 	struct qfq_sched *q = qdisc_priv(sch);
571 
572 	if (cl)
573 		return NULL;
574 
575 	return q->block;
576 }
577 
qfq_bind_tcf(struct Qdisc * sch,unsigned long parent,u32 classid)578 static unsigned long qfq_bind_tcf(struct Qdisc *sch, unsigned long parent,
579 				  u32 classid)
580 {
581 	struct qfq_class *cl = qfq_find_class(sch, classid);
582 
583 	if (cl)
584 		qdisc_class_get(&cl->common);
585 
586 	return (unsigned long)cl;
587 }
588 
qfq_unbind_tcf(struct Qdisc * sch,unsigned long arg)589 static void qfq_unbind_tcf(struct Qdisc *sch, unsigned long arg)
590 {
591 	struct qfq_class *cl = (struct qfq_class *)arg;
592 
593 	qdisc_class_put(&cl->common);
594 }
595 
qfq_graft_class(struct Qdisc * sch,unsigned long arg,struct Qdisc * new,struct Qdisc ** old,struct netlink_ext_ack * extack)596 static int qfq_graft_class(struct Qdisc *sch, unsigned long arg,
597 			   struct Qdisc *new, struct Qdisc **old,
598 			   struct netlink_ext_ack *extack)
599 {
600 	struct qfq_class *cl = (struct qfq_class *)arg;
601 
602 	if (new == NULL) {
603 		new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
604 					cl->common.classid, NULL);
605 		if (new == NULL)
606 			new = &noop_qdisc;
607 	}
608 
609 	*old = qdisc_replace(sch, new, &cl->qdisc);
610 	return 0;
611 }
612 
qfq_class_leaf(struct Qdisc * sch,unsigned long arg)613 static struct Qdisc *qfq_class_leaf(struct Qdisc *sch, unsigned long arg)
614 {
615 	struct qfq_class *cl = (struct qfq_class *)arg;
616 
617 	return cl->qdisc;
618 }
619 
qfq_dump_class(struct Qdisc * sch,unsigned long arg,struct sk_buff * skb,struct tcmsg * tcm)620 static int qfq_dump_class(struct Qdisc *sch, unsigned long arg,
621 			  struct sk_buff *skb, struct tcmsg *tcm)
622 {
623 	struct qfq_class *cl = (struct qfq_class *)arg;
624 	struct nlattr *nest;
625 
626 	tcm->tcm_parent	= TC_H_ROOT;
627 	tcm->tcm_handle	= cl->common.classid;
628 	tcm->tcm_info	= cl->qdisc->handle;
629 
630 	nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
631 	if (nest == NULL)
632 		goto nla_put_failure;
633 	if (nla_put_u32(skb, TCA_QFQ_WEIGHT, cl->agg->class_weight) ||
634 	    nla_put_u32(skb, TCA_QFQ_LMAX, cl->agg->lmax))
635 		goto nla_put_failure;
636 	return nla_nest_end(skb, nest);
637 
638 nla_put_failure:
639 	nla_nest_cancel(skb, nest);
640 	return -EMSGSIZE;
641 }
642 
qfq_dump_class_stats(struct Qdisc * sch,unsigned long arg,struct gnet_dump * d)643 static int qfq_dump_class_stats(struct Qdisc *sch, unsigned long arg,
644 				struct gnet_dump *d)
645 {
646 	struct qfq_class *cl = (struct qfq_class *)arg;
647 	struct tc_qfq_stats xstats;
648 
649 	memset(&xstats, 0, sizeof(xstats));
650 
651 	xstats.weight = cl->agg->class_weight;
652 	xstats.lmax = cl->agg->lmax;
653 
654 	if (gnet_stats_copy_basic(d, NULL, &cl->bstats, true) < 0 ||
655 	    gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
656 	    qdisc_qstats_copy(d, cl->qdisc) < 0)
657 		return -1;
658 
659 	return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
660 }
661 
qfq_walk(struct Qdisc * sch,struct qdisc_walker * arg)662 static void qfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
663 {
664 	struct qfq_sched *q = qdisc_priv(sch);
665 	struct qfq_class *cl;
666 	unsigned int i;
667 
668 	if (arg->stop)
669 		return;
670 
671 	for (i = 0; i < q->clhash.hashsize; i++) {
672 		hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
673 			if (!tc_qdisc_stats_dump(sch, (unsigned long)cl, arg))
674 				return;
675 		}
676 	}
677 }
678 
qfq_classify(struct sk_buff * skb,struct Qdisc * sch,int * qerr)679 static struct qfq_class *qfq_classify(struct sk_buff *skb, struct Qdisc *sch,
680 				      int *qerr)
681 {
682 	struct qfq_sched *q = qdisc_priv(sch);
683 	struct qfq_class *cl;
684 	struct tcf_result res;
685 	struct tcf_proto *fl;
686 	int result;
687 
688 	if (TC_H_MAJ(skb->priority ^ sch->handle) == 0) {
689 		pr_debug("qfq_classify: found %d\n", skb->priority);
690 		cl = qfq_find_class(sch, skb->priority);
691 		if (cl != NULL)
692 			return cl;
693 	}
694 
695 	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
696 	fl = rcu_dereference_bh(q->filter_list);
697 	result = tcf_classify(skb, NULL, fl, &res, false);
698 	if (result >= 0) {
699 #ifdef CONFIG_NET_CLS_ACT
700 		switch (result) {
701 		case TC_ACT_QUEUED:
702 		case TC_ACT_STOLEN:
703 		case TC_ACT_TRAP:
704 			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
705 			fallthrough;
706 		case TC_ACT_SHOT:
707 			return NULL;
708 		}
709 #endif
710 		cl = (struct qfq_class *)res.class;
711 		if (cl == NULL)
712 			cl = qfq_find_class(sch, res.classid);
713 		return cl;
714 	}
715 
716 	return NULL;
717 }
718 
719 /* Generic comparison function, handling wraparound. */
qfq_gt(u64 a,u64 b)720 static inline int qfq_gt(u64 a, u64 b)
721 {
722 	return (s64)(a - b) > 0;
723 }
724 
725 /* Round a precise timestamp to its slotted value. */
qfq_round_down(u64 ts,unsigned int shift)726 static inline u64 qfq_round_down(u64 ts, unsigned int shift)
727 {
728 	return ts & ~((1ULL << shift) - 1);
729 }
730 
731 /* return the pointer to the group with lowest index in the bitmap */
qfq_ffs(struct qfq_sched * q,unsigned long bitmap)732 static inline struct qfq_group *qfq_ffs(struct qfq_sched *q,
733 					unsigned long bitmap)
734 {
735 	int index = __ffs(bitmap);
736 	return &q->groups[index];
737 }
738 /* Calculate a mask to mimic what would be ffs_from(). */
mask_from(unsigned long bitmap,int from)739 static inline unsigned long mask_from(unsigned long bitmap, int from)
740 {
741 	return bitmap & ~((1UL << from) - 1);
742 }
743 
744 /*
745  * The state computation relies on ER=0, IR=1, EB=2, IB=3
746  * First compute eligibility comparing grp->S, q->V,
747  * then check if someone is blocking us and possibly add EB
748  */
qfq_calc_state(struct qfq_sched * q,const struct qfq_group * grp)749 static int qfq_calc_state(struct qfq_sched *q, const struct qfq_group *grp)
750 {
751 	/* if S > V we are not eligible */
752 	unsigned int state = qfq_gt(grp->S, q->V);
753 	unsigned long mask = mask_from(q->bitmaps[ER], grp->index);
754 	struct qfq_group *next;
755 
756 	if (mask) {
757 		next = qfq_ffs(q, mask);
758 		if (qfq_gt(grp->F, next->F))
759 			state |= EB;
760 	}
761 
762 	return state;
763 }
764 
765 
766 /*
767  * In principle
768  *	q->bitmaps[dst] |= q->bitmaps[src] & mask;
769  *	q->bitmaps[src] &= ~mask;
770  * but we should make sure that src != dst
771  */
qfq_move_groups(struct qfq_sched * q,unsigned long mask,int src,int dst)772 static inline void qfq_move_groups(struct qfq_sched *q, unsigned long mask,
773 				   int src, int dst)
774 {
775 	q->bitmaps[dst] |= q->bitmaps[src] & mask;
776 	q->bitmaps[src] &= ~mask;
777 }
778 
qfq_unblock_groups(struct qfq_sched * q,int index,u64 old_F)779 static void qfq_unblock_groups(struct qfq_sched *q, int index, u64 old_F)
780 {
781 	unsigned long mask = mask_from(q->bitmaps[ER], index + 1);
782 	struct qfq_group *next;
783 
784 	if (mask) {
785 		next = qfq_ffs(q, mask);
786 		if (!qfq_gt(next->F, old_F))
787 			return;
788 	}
789 
790 	mask = (1UL << index) - 1;
791 	qfq_move_groups(q, mask, EB, ER);
792 	qfq_move_groups(q, mask, IB, IR);
793 }
794 
795 /*
796  * perhaps
797  *
798 	old_V ^= q->V;
799 	old_V >>= q->min_slot_shift;
800 	if (old_V) {
801 		...
802 	}
803  *
804  */
qfq_make_eligible(struct qfq_sched * q)805 static void qfq_make_eligible(struct qfq_sched *q)
806 {
807 	unsigned long vslot = q->V >> q->min_slot_shift;
808 	unsigned long old_vslot = q->oldV >> q->min_slot_shift;
809 
810 	if (vslot != old_vslot) {
811 		unsigned long mask;
812 		int last_flip_pos = fls(vslot ^ old_vslot);
813 
814 		if (last_flip_pos > 31) /* higher than the number of groups */
815 			mask = ~0UL;    /* make all groups eligible */
816 		else
817 			mask = (1UL << last_flip_pos) - 1;
818 
819 		qfq_move_groups(q, mask, IR, ER);
820 		qfq_move_groups(q, mask, IB, EB);
821 	}
822 }
823 
824 /*
825  * The index of the slot in which the input aggregate agg is to be
826  * inserted must not be higher than QFQ_MAX_SLOTS-2. There is a '-2'
827  * and not a '-1' because the start time of the group may be moved
828  * backward by one slot after the aggregate has been inserted, and
829  * this would cause non-empty slots to be right-shifted by one
830  * position.
831  *
832  * QFQ+ fully satisfies this bound to the slot index if the parameters
833  * of the classes are not changed dynamically, and if QFQ+ never
834  * happens to postpone the service of agg unjustly, i.e., it never
835  * happens that the aggregate becomes backlogged and eligible, or just
836  * eligible, while an aggregate with a higher approximated finish time
837  * is being served. In particular, in this case QFQ+ guarantees that
838  * the timestamps of agg are low enough that the slot index is never
839  * higher than 2. Unfortunately, QFQ+ cannot provide the same
840  * guarantee if it happens to unjustly postpone the service of agg, or
841  * if the parameters of some class are changed.
842  *
843  * As for the first event, i.e., an out-of-order service, the
844  * upper bound to the slot index guaranteed by QFQ+ grows to
845  * 2 +
846  * QFQ_MAX_AGG_CLASSES * ((1<<QFQ_MTU_SHIFT)/QFQ_MIN_LMAX) *
847  * (current_max_weight/current_wsum) <= 2 + 8 * 128 * 1.
848  *
849  * The following function deals with this problem by backward-shifting
850  * the timestamps of agg, if needed, so as to guarantee that the slot
851  * index is never higher than QFQ_MAX_SLOTS-2. This backward-shift may
852  * cause the service of other aggregates to be postponed, yet the
853  * worst-case guarantees of these aggregates are not violated.  In
854  * fact, in case of no out-of-order service, the timestamps of agg
855  * would have been even lower than they are after the backward shift,
856  * because QFQ+ would have guaranteed a maximum value equal to 2 for
857  * the slot index, and 2 < QFQ_MAX_SLOTS-2. Hence the aggregates whose
858  * service is postponed because of the backward-shift would have
859  * however waited for the service of agg before being served.
860  *
861  * The other event that may cause the slot index to be higher than 2
862  * for agg is a recent change of the parameters of some class. If the
863  * weight of a class is increased or the lmax (max_pkt_size) of the
864  * class is decreased, then a new aggregate with smaller slot size
865  * than the original parent aggregate of the class may happen to be
866  * activated. The activation of this aggregate should be properly
867  * delayed to when the service of the class has finished in the ideal
868  * system tracked by QFQ+. If the activation of the aggregate is not
869  * delayed to this reference time instant, then this aggregate may be
870  * unjustly served before other aggregates waiting for service. This
871  * may cause the above bound to the slot index to be violated for some
872  * of these unlucky aggregates.
873  *
874  * Instead of delaying the activation of the new aggregate, which is
875  * quite complex, the above-discussed capping of the slot index is
876  * used to handle also the consequences of a change of the parameters
877  * of a class.
878  */
qfq_slot_insert(struct qfq_group * grp,struct qfq_aggregate * agg,u64 roundedS)879 static void qfq_slot_insert(struct qfq_group *grp, struct qfq_aggregate *agg,
880 			    u64 roundedS)
881 {
882 	u64 slot = (roundedS - grp->S) >> grp->slot_shift;
883 	unsigned int i; /* slot index in the bucket list */
884 
885 	if (unlikely(slot > QFQ_MAX_SLOTS - 2)) {
886 		u64 deltaS = roundedS - grp->S -
887 			((u64)(QFQ_MAX_SLOTS - 2)<<grp->slot_shift);
888 		agg->S -= deltaS;
889 		agg->F -= deltaS;
890 		slot = QFQ_MAX_SLOTS - 2;
891 	}
892 
893 	i = (grp->front + slot) % QFQ_MAX_SLOTS;
894 
895 	hlist_add_head(&agg->next, &grp->slots[i]);
896 	__set_bit(slot, &grp->full_slots);
897 }
898 
899 /* Maybe introduce hlist_first_entry?? */
qfq_slot_head(struct qfq_group * grp)900 static struct qfq_aggregate *qfq_slot_head(struct qfq_group *grp)
901 {
902 	return hlist_entry(grp->slots[grp->front].first,
903 			   struct qfq_aggregate, next);
904 }
905 
906 /*
907  * remove the entry from the slot
908  */
qfq_front_slot_remove(struct qfq_group * grp)909 static void qfq_front_slot_remove(struct qfq_group *grp)
910 {
911 	struct qfq_aggregate *agg = qfq_slot_head(grp);
912 
913 	BUG_ON(!agg);
914 	hlist_del(&agg->next);
915 	if (hlist_empty(&grp->slots[grp->front]))
916 		__clear_bit(0, &grp->full_slots);
917 }
918 
919 /*
920  * Returns the first aggregate in the first non-empty bucket of the
921  * group. As a side effect, adjusts the bucket list so the first
922  * non-empty bucket is at position 0 in full_slots.
923  */
qfq_slot_scan(struct qfq_group * grp)924 static struct qfq_aggregate *qfq_slot_scan(struct qfq_group *grp)
925 {
926 	unsigned int i;
927 
928 	pr_debug("qfq slot_scan: grp %u full %#lx\n",
929 		 grp->index, grp->full_slots);
930 
931 	if (grp->full_slots == 0)
932 		return NULL;
933 
934 	i = __ffs(grp->full_slots);  /* zero based */
935 	if (i > 0) {
936 		grp->front = (grp->front + i) % QFQ_MAX_SLOTS;
937 		grp->full_slots >>= i;
938 	}
939 
940 	return qfq_slot_head(grp);
941 }
942 
943 /*
944  * adjust the bucket list. When the start time of a group decreases,
945  * we move the index down (modulo QFQ_MAX_SLOTS) so we don't need to
946  * move the objects. The mask of occupied slots must be shifted
947  * because we use ffs() to find the first non-empty slot.
948  * This covers decreases in the group's start time, but what about
949  * increases of the start time ?
950  * Here too we should make sure that i is less than 32
951  */
qfq_slot_rotate(struct qfq_group * grp,u64 roundedS)952 static void qfq_slot_rotate(struct qfq_group *grp, u64 roundedS)
953 {
954 	unsigned int i = (grp->S - roundedS) >> grp->slot_shift;
955 
956 	grp->full_slots <<= i;
957 	grp->front = (grp->front - i) % QFQ_MAX_SLOTS;
958 }
959 
qfq_update_eligible(struct qfq_sched * q)960 static void qfq_update_eligible(struct qfq_sched *q)
961 {
962 	struct qfq_group *grp;
963 	unsigned long ineligible;
964 
965 	ineligible = q->bitmaps[IR] | q->bitmaps[IB];
966 	if (ineligible) {
967 		if (!q->bitmaps[ER]) {
968 			grp = qfq_ffs(q, ineligible);
969 			if (qfq_gt(grp->S, q->V))
970 				q->V = grp->S;
971 		}
972 		qfq_make_eligible(q);
973 	}
974 }
975 
976 /* Dequeue head packet of the head class in the DRR queue of the aggregate. */
agg_dequeue(struct qfq_aggregate * agg,struct qfq_class * cl,unsigned int len)977 static struct sk_buff *agg_dequeue(struct qfq_aggregate *agg,
978 				   struct qfq_class *cl, unsigned int len)
979 {
980 	struct sk_buff *skb = qdisc_dequeue_peeked(cl->qdisc);
981 
982 	if (!skb)
983 		return NULL;
984 
985 	cl->deficit -= (int) len;
986 
987 	if (cl->qdisc->q.qlen == 0) /* no more packets, remove from list */
988 		list_del(&cl->alist);
989 	else if (cl->deficit < qdisc_pkt_len(cl->qdisc->ops->peek(cl->qdisc))) {
990 		cl->deficit += agg->lmax;
991 		list_move_tail(&cl->alist, &agg->active);
992 	}
993 
994 	return skb;
995 }
996 
qfq_peek_skb(struct qfq_aggregate * agg,struct qfq_class ** cl,unsigned int * len)997 static inline struct sk_buff *qfq_peek_skb(struct qfq_aggregate *agg,
998 					   struct qfq_class **cl,
999 					   unsigned int *len)
1000 {
1001 	struct sk_buff *skb;
1002 
1003 	*cl = list_first_entry(&agg->active, struct qfq_class, alist);
1004 	skb = (*cl)->qdisc->ops->peek((*cl)->qdisc);
1005 	if (skb == NULL)
1006 		WARN_ONCE(1, "qfq_dequeue: non-workconserving leaf\n");
1007 	else
1008 		*len = qdisc_pkt_len(skb);
1009 
1010 	return skb;
1011 }
1012 
1013 /* Update F according to the actual service received by the aggregate. */
charge_actual_service(struct qfq_aggregate * agg)1014 static inline void charge_actual_service(struct qfq_aggregate *agg)
1015 {
1016 	/* Compute the service received by the aggregate, taking into
1017 	 * account that, after decreasing the number of classes in
1018 	 * agg, it may happen that
1019 	 * agg->initial_budget - agg->budget > agg->bugdetmax
1020 	 */
1021 	u32 service_received = min(agg->budgetmax,
1022 				   agg->initial_budget - agg->budget);
1023 
1024 	agg->F = agg->S + (u64)service_received * agg->inv_w;
1025 }
1026 
1027 /* Assign a reasonable start time for a new aggregate in group i.
1028  * Admissible values for \hat(F) are multiples of \sigma_i
1029  * no greater than V+\sigma_i . Larger values mean that
1030  * we had a wraparound so we consider the timestamp to be stale.
1031  *
1032  * If F is not stale and F >= V then we set S = F.
1033  * Otherwise we should assign S = V, but this may violate
1034  * the ordering in EB (see [2]). So, if we have groups in ER,
1035  * set S to the F_j of the first group j which would be blocking us.
1036  * We are guaranteed not to move S backward because
1037  * otherwise our group i would still be blocked.
1038  */
qfq_update_start(struct qfq_sched * q,struct qfq_aggregate * agg)1039 static void qfq_update_start(struct qfq_sched *q, struct qfq_aggregate *agg)
1040 {
1041 	unsigned long mask;
1042 	u64 limit, roundedF;
1043 	int slot_shift = agg->grp->slot_shift;
1044 
1045 	roundedF = qfq_round_down(agg->F, slot_shift);
1046 	limit = qfq_round_down(q->V, slot_shift) + (1ULL << slot_shift);
1047 
1048 	if (!qfq_gt(agg->F, q->V) || qfq_gt(roundedF, limit)) {
1049 		/* timestamp was stale */
1050 		mask = mask_from(q->bitmaps[ER], agg->grp->index);
1051 		if (mask) {
1052 			struct qfq_group *next = qfq_ffs(q, mask);
1053 			if (qfq_gt(roundedF, next->F)) {
1054 				if (qfq_gt(limit, next->F))
1055 					agg->S = next->F;
1056 				else /* preserve timestamp correctness */
1057 					agg->S = limit;
1058 				return;
1059 			}
1060 		}
1061 		agg->S = q->V;
1062 	} else  /* timestamp is not stale */
1063 		agg->S = agg->F;
1064 }
1065 
1066 /* Update the timestamps of agg before scheduling/rescheduling it for
1067  * service.  In particular, assign to agg->F its maximum possible
1068  * value, i.e., the virtual finish time with which the aggregate
1069  * should be labeled if it used all its budget once in service.
1070  */
1071 static inline void
qfq_update_agg_ts(struct qfq_sched * q,struct qfq_aggregate * agg,enum update_reason reason)1072 qfq_update_agg_ts(struct qfq_sched *q,
1073 		    struct qfq_aggregate *agg, enum update_reason reason)
1074 {
1075 	if (reason != requeue)
1076 		qfq_update_start(q, agg);
1077 	else /* just charge agg for the service received */
1078 		agg->S = agg->F;
1079 
1080 	agg->F = agg->S + (u64)agg->budgetmax * agg->inv_w;
1081 }
1082 
1083 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg);
1084 
qfq_dequeue(struct Qdisc * sch)1085 static struct sk_buff *qfq_dequeue(struct Qdisc *sch)
1086 {
1087 	struct qfq_sched *q = qdisc_priv(sch);
1088 	struct qfq_aggregate *in_serv_agg = q->in_serv_agg;
1089 	struct qfq_class *cl;
1090 	struct sk_buff *skb = NULL;
1091 	/* next-packet len, 0 means no more active classes in in-service agg */
1092 	unsigned int len = 0;
1093 
1094 	if (in_serv_agg == NULL)
1095 		return NULL;
1096 
1097 	if (!list_empty(&in_serv_agg->active))
1098 		skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1099 
1100 	/*
1101 	 * If there are no active classes in the in-service aggregate,
1102 	 * or if the aggregate has not enough budget to serve its next
1103 	 * class, then choose the next aggregate to serve.
1104 	 */
1105 	if (len == 0 || in_serv_agg->budget < len) {
1106 		charge_actual_service(in_serv_agg);
1107 
1108 		/* recharge the budget of the aggregate */
1109 		in_serv_agg->initial_budget = in_serv_agg->budget =
1110 			in_serv_agg->budgetmax;
1111 
1112 		if (!list_empty(&in_serv_agg->active)) {
1113 			/*
1114 			 * Still active: reschedule for
1115 			 * service. Possible optimization: if no other
1116 			 * aggregate is active, then there is no point
1117 			 * in rescheduling this aggregate, and we can
1118 			 * just keep it as the in-service one. This
1119 			 * should be however a corner case, and to
1120 			 * handle it, we would need to maintain an
1121 			 * extra num_active_aggs field.
1122 			*/
1123 			qfq_update_agg_ts(q, in_serv_agg, requeue);
1124 			qfq_schedule_agg(q, in_serv_agg);
1125 		} else if (sch->q.qlen == 0) { /* no aggregate to serve */
1126 			q->in_serv_agg = NULL;
1127 			return NULL;
1128 		}
1129 
1130 		/*
1131 		 * If we get here, there are other aggregates queued:
1132 		 * choose the new aggregate to serve.
1133 		 */
1134 		in_serv_agg = q->in_serv_agg = qfq_choose_next_agg(q);
1135 		skb = qfq_peek_skb(in_serv_agg, &cl, &len);
1136 	}
1137 	if (!skb)
1138 		return NULL;
1139 
1140 	sch->q.qlen--;
1141 
1142 	skb = agg_dequeue(in_serv_agg, cl, len);
1143 
1144 	if (!skb) {
1145 		sch->q.qlen++;
1146 		return NULL;
1147 	}
1148 
1149 	qdisc_qstats_backlog_dec(sch, skb);
1150 	qdisc_bstats_update(sch, skb);
1151 
1152 	/* If lmax is lowered, through qfq_change_class, for a class
1153 	 * owning pending packets with larger size than the new value
1154 	 * of lmax, then the following condition may hold.
1155 	 */
1156 	if (unlikely(in_serv_agg->budget < len))
1157 		in_serv_agg->budget = 0;
1158 	else
1159 		in_serv_agg->budget -= len;
1160 
1161 	q->V += (u64)len * q->iwsum;
1162 	pr_debug("qfq dequeue: len %u F %lld now %lld\n",
1163 		 len, (unsigned long long) in_serv_agg->F,
1164 		 (unsigned long long) q->V);
1165 
1166 	return skb;
1167 }
1168 
qfq_choose_next_agg(struct qfq_sched * q)1169 static struct qfq_aggregate *qfq_choose_next_agg(struct qfq_sched *q)
1170 {
1171 	struct qfq_group *grp;
1172 	struct qfq_aggregate *agg, *new_front_agg;
1173 	u64 old_F;
1174 
1175 	qfq_update_eligible(q);
1176 	q->oldV = q->V;
1177 
1178 	if (!q->bitmaps[ER])
1179 		return NULL;
1180 
1181 	grp = qfq_ffs(q, q->bitmaps[ER]);
1182 	old_F = grp->F;
1183 
1184 	agg = qfq_slot_head(grp);
1185 
1186 	/* agg starts to be served, remove it from schedule */
1187 	qfq_front_slot_remove(grp);
1188 
1189 	new_front_agg = qfq_slot_scan(grp);
1190 
1191 	if (new_front_agg == NULL) /* group is now inactive, remove from ER */
1192 		__clear_bit(grp->index, &q->bitmaps[ER]);
1193 	else {
1194 		u64 roundedS = qfq_round_down(new_front_agg->S,
1195 					      grp->slot_shift);
1196 		unsigned int s;
1197 
1198 		if (grp->S == roundedS)
1199 			return agg;
1200 		grp->S = roundedS;
1201 		grp->F = roundedS + (2ULL << grp->slot_shift);
1202 		__clear_bit(grp->index, &q->bitmaps[ER]);
1203 		s = qfq_calc_state(q, grp);
1204 		__set_bit(grp->index, &q->bitmaps[s]);
1205 	}
1206 
1207 	qfq_unblock_groups(q, grp->index, old_F);
1208 
1209 	return agg;
1210 }
1211 
qfq_enqueue(struct sk_buff * skb,struct Qdisc * sch,struct sk_buff ** to_free)1212 static int qfq_enqueue(struct sk_buff *skb, struct Qdisc *sch,
1213 		       struct sk_buff **to_free)
1214 {
1215 	unsigned int len = qdisc_pkt_len(skb), gso_segs;
1216 	struct qfq_sched *q = qdisc_priv(sch);
1217 	struct qfq_class *cl;
1218 	struct qfq_aggregate *agg;
1219 	int err = 0;
1220 	bool first;
1221 
1222 	cl = qfq_classify(skb, sch, &err);
1223 	if (cl == NULL) {
1224 		if (err & __NET_XMIT_BYPASS)
1225 			qdisc_qstats_drop(sch);
1226 		__qdisc_drop(skb, to_free);
1227 		return err;
1228 	}
1229 	pr_debug("qfq_enqueue: cl = %x\n", cl->common.classid);
1230 
1231 	if (unlikely(cl->agg->lmax < len)) {
1232 		pr_debug("qfq: increasing maxpkt from %u to %u for class %u",
1233 			 cl->agg->lmax, len, cl->common.classid);
1234 		err = qfq_change_agg(sch, cl, cl->agg->class_weight, len);
1235 		if (err) {
1236 			cl->qstats.drops++;
1237 			return qdisc_drop(skb, sch, to_free);
1238 		}
1239 	}
1240 
1241 	gso_segs = skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 1;
1242 	first = !cl->qdisc->q.qlen;
1243 	err = qdisc_enqueue(skb, cl->qdisc, to_free);
1244 	if (unlikely(err != NET_XMIT_SUCCESS)) {
1245 		pr_debug("qfq_enqueue: enqueue failed %d\n", err);
1246 		if (net_xmit_drop_count(err)) {
1247 			cl->qstats.drops++;
1248 			qdisc_qstats_drop(sch);
1249 		}
1250 		return err;
1251 	}
1252 
1253 	_bstats_update(&cl->bstats, len, gso_segs);
1254 	sch->qstats.backlog += len;
1255 	++sch->q.qlen;
1256 
1257 	agg = cl->agg;
1258 	/* if the queue was not empty, then done here */
1259 	if (!first) {
1260 		if (unlikely(skb == cl->qdisc->ops->peek(cl->qdisc)) &&
1261 		    list_first_entry(&agg->active, struct qfq_class, alist)
1262 		    == cl && cl->deficit < len)
1263 			list_move_tail(&cl->alist, &agg->active);
1264 
1265 		return err;
1266 	}
1267 
1268 	/* schedule class for service within the aggregate */
1269 	cl->deficit = agg->lmax;
1270 	list_add_tail(&cl->alist, &agg->active);
1271 
1272 	if (list_first_entry(&agg->active, struct qfq_class, alist) != cl ||
1273 	    q->in_serv_agg == agg)
1274 		return err; /* non-empty or in service, nothing else to do */
1275 
1276 	qfq_activate_agg(q, agg, enqueue);
1277 
1278 	return err;
1279 }
1280 
1281 /*
1282  * Schedule aggregate according to its timestamps.
1283  */
qfq_schedule_agg(struct qfq_sched * q,struct qfq_aggregate * agg)1284 static void qfq_schedule_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1285 {
1286 	struct qfq_group *grp = agg->grp;
1287 	u64 roundedS;
1288 	int s;
1289 
1290 	roundedS = qfq_round_down(agg->S, grp->slot_shift);
1291 
1292 	/*
1293 	 * Insert agg in the correct bucket.
1294 	 * If agg->S >= grp->S we don't need to adjust the
1295 	 * bucket list and simply go to the insertion phase.
1296 	 * Otherwise grp->S is decreasing, we must make room
1297 	 * in the bucket list, and also recompute the group state.
1298 	 * Finally, if there were no flows in this group and nobody
1299 	 * was in ER make sure to adjust V.
1300 	 */
1301 	if (grp->full_slots) {
1302 		if (!qfq_gt(grp->S, agg->S))
1303 			goto skip_update;
1304 
1305 		/* create a slot for this agg->S */
1306 		qfq_slot_rotate(grp, roundedS);
1307 		/* group was surely ineligible, remove */
1308 		__clear_bit(grp->index, &q->bitmaps[IR]);
1309 		__clear_bit(grp->index, &q->bitmaps[IB]);
1310 	} else if (!q->bitmaps[ER] && qfq_gt(roundedS, q->V) &&
1311 		   q->in_serv_agg == NULL)
1312 		q->V = roundedS;
1313 
1314 	grp->S = roundedS;
1315 	grp->F = roundedS + (2ULL << grp->slot_shift);
1316 	s = qfq_calc_state(q, grp);
1317 	__set_bit(grp->index, &q->bitmaps[s]);
1318 
1319 	pr_debug("qfq enqueue: new state %d %#lx S %lld F %lld V %lld\n",
1320 		 s, q->bitmaps[s],
1321 		 (unsigned long long) agg->S,
1322 		 (unsigned long long) agg->F,
1323 		 (unsigned long long) q->V);
1324 
1325 skip_update:
1326 	qfq_slot_insert(grp, agg, roundedS);
1327 }
1328 
1329 
1330 /* Update agg ts and schedule agg for service */
qfq_activate_agg(struct qfq_sched * q,struct qfq_aggregate * agg,enum update_reason reason)1331 static void qfq_activate_agg(struct qfq_sched *q, struct qfq_aggregate *agg,
1332 			     enum update_reason reason)
1333 {
1334 	agg->initial_budget = agg->budget = agg->budgetmax; /* recharge budg. */
1335 
1336 	qfq_update_agg_ts(q, agg, reason);
1337 	if (q->in_serv_agg == NULL) { /* no aggr. in service or scheduled */
1338 		q->in_serv_agg = agg; /* start serving this aggregate */
1339 		 /* update V: to be in service, agg must be eligible */
1340 		q->oldV = q->V = agg->S;
1341 	} else if (agg != q->in_serv_agg)
1342 		qfq_schedule_agg(q, agg);
1343 }
1344 
qfq_slot_remove(struct qfq_sched * q,struct qfq_group * grp,struct qfq_aggregate * agg)1345 static void qfq_slot_remove(struct qfq_sched *q, struct qfq_group *grp,
1346 			    struct qfq_aggregate *agg)
1347 {
1348 	unsigned int i, offset;
1349 	u64 roundedS;
1350 
1351 	roundedS = qfq_round_down(agg->S, grp->slot_shift);
1352 	offset = (roundedS - grp->S) >> grp->slot_shift;
1353 
1354 	i = (grp->front + offset) % QFQ_MAX_SLOTS;
1355 
1356 	hlist_del(&agg->next);
1357 	if (hlist_empty(&grp->slots[i]))
1358 		__clear_bit(offset, &grp->full_slots);
1359 }
1360 
1361 /*
1362  * Called to forcibly deschedule an aggregate.  If the aggregate is
1363  * not in the front bucket, or if the latter has other aggregates in
1364  * the front bucket, we can simply remove the aggregate with no other
1365  * side effects.
1366  * Otherwise we must propagate the event up.
1367  */
qfq_deactivate_agg(struct qfq_sched * q,struct qfq_aggregate * agg)1368 static void qfq_deactivate_agg(struct qfq_sched *q, struct qfq_aggregate *agg)
1369 {
1370 	struct qfq_group *grp = agg->grp;
1371 	unsigned long mask;
1372 	u64 roundedS;
1373 	int s;
1374 
1375 	if (agg == q->in_serv_agg) {
1376 		charge_actual_service(agg);
1377 		q->in_serv_agg = qfq_choose_next_agg(q);
1378 		return;
1379 	}
1380 
1381 	agg->F = agg->S;
1382 	qfq_slot_remove(q, grp, agg);
1383 
1384 	if (!grp->full_slots) {
1385 		__clear_bit(grp->index, &q->bitmaps[IR]);
1386 		__clear_bit(grp->index, &q->bitmaps[EB]);
1387 		__clear_bit(grp->index, &q->bitmaps[IB]);
1388 
1389 		if (test_bit(grp->index, &q->bitmaps[ER]) &&
1390 		    !(q->bitmaps[ER] & ~((1UL << grp->index) - 1))) {
1391 			mask = q->bitmaps[ER] & ((1UL << grp->index) - 1);
1392 			if (mask)
1393 				mask = ~((1UL << __fls(mask)) - 1);
1394 			else
1395 				mask = ~0UL;
1396 			qfq_move_groups(q, mask, EB, ER);
1397 			qfq_move_groups(q, mask, IB, IR);
1398 		}
1399 		__clear_bit(grp->index, &q->bitmaps[ER]);
1400 	} else if (hlist_empty(&grp->slots[grp->front])) {
1401 		agg = qfq_slot_scan(grp);
1402 		roundedS = qfq_round_down(agg->S, grp->slot_shift);
1403 		if (grp->S != roundedS) {
1404 			__clear_bit(grp->index, &q->bitmaps[ER]);
1405 			__clear_bit(grp->index, &q->bitmaps[IR]);
1406 			__clear_bit(grp->index, &q->bitmaps[EB]);
1407 			__clear_bit(grp->index, &q->bitmaps[IB]);
1408 			grp->S = roundedS;
1409 			grp->F = roundedS + (2ULL << grp->slot_shift);
1410 			s = qfq_calc_state(q, grp);
1411 			__set_bit(grp->index, &q->bitmaps[s]);
1412 		}
1413 	}
1414 }
1415 
qfq_qlen_notify(struct Qdisc * sch,unsigned long arg)1416 static void qfq_qlen_notify(struct Qdisc *sch, unsigned long arg)
1417 {
1418 	struct qfq_sched *q = qdisc_priv(sch);
1419 	struct qfq_class *cl = (struct qfq_class *)arg;
1420 
1421 	qfq_deactivate_class(q, cl);
1422 }
1423 
qfq_init_qdisc(struct Qdisc * sch,struct nlattr * opt,struct netlink_ext_ack * extack)1424 static int qfq_init_qdisc(struct Qdisc *sch, struct nlattr *opt,
1425 			  struct netlink_ext_ack *extack)
1426 {
1427 	struct qfq_sched *q = qdisc_priv(sch);
1428 	struct qfq_group *grp;
1429 	int i, j, err;
1430 	u32 max_cl_shift, maxbudg_shift, max_classes;
1431 
1432 	err = tcf_block_get(&q->block, &q->filter_list, sch, extack);
1433 	if (err)
1434 		return err;
1435 
1436 	err = qdisc_class_hash_init(&q->clhash);
1437 	if (err < 0)
1438 		return err;
1439 
1440 	max_classes = min_t(u64, (u64)qdisc_dev(sch)->tx_queue_len + 1,
1441 			    QFQ_MAX_AGG_CLASSES);
1442 	/* max_cl_shift = floor(log_2(max_classes)) */
1443 	max_cl_shift = __fls(max_classes);
1444 	q->max_agg_classes = 1<<max_cl_shift;
1445 
1446 	/* maxbudg_shift = log2(max_len * max_classes_per_agg) */
1447 	maxbudg_shift = QFQ_MTU_SHIFT + max_cl_shift;
1448 	q->min_slot_shift = FRAC_BITS + maxbudg_shift - QFQ_MAX_INDEX;
1449 
1450 	for (i = 0; i <= QFQ_MAX_INDEX; i++) {
1451 		grp = &q->groups[i];
1452 		grp->index = i;
1453 		grp->slot_shift = q->min_slot_shift + i;
1454 		for (j = 0; j < QFQ_MAX_SLOTS; j++)
1455 			INIT_HLIST_HEAD(&grp->slots[j]);
1456 	}
1457 
1458 	INIT_HLIST_HEAD(&q->nonfull_aggs);
1459 
1460 	return 0;
1461 }
1462 
qfq_reset_qdisc(struct Qdisc * sch)1463 static void qfq_reset_qdisc(struct Qdisc *sch)
1464 {
1465 	struct qfq_sched *q = qdisc_priv(sch);
1466 	struct qfq_class *cl;
1467 	unsigned int i;
1468 
1469 	for (i = 0; i < q->clhash.hashsize; i++) {
1470 		hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1471 			if (cl->qdisc->q.qlen > 0)
1472 				qfq_deactivate_class(q, cl);
1473 
1474 			qdisc_reset(cl->qdisc);
1475 		}
1476 	}
1477 }
1478 
qfq_destroy_qdisc(struct Qdisc * sch)1479 static void qfq_destroy_qdisc(struct Qdisc *sch)
1480 {
1481 	struct qfq_sched *q = qdisc_priv(sch);
1482 	struct qfq_class *cl;
1483 	struct hlist_node *next;
1484 	unsigned int i;
1485 
1486 	tcf_block_put(q->block);
1487 
1488 	for (i = 0; i < q->clhash.hashsize; i++) {
1489 		hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1490 					  common.hnode) {
1491 			qfq_destroy_class(sch, cl);
1492 		}
1493 	}
1494 	qdisc_class_hash_destroy(&q->clhash);
1495 }
1496 
1497 static const struct Qdisc_class_ops qfq_class_ops = {
1498 	.change		= qfq_change_class,
1499 	.delete		= qfq_delete_class,
1500 	.find		= qfq_search_class,
1501 	.tcf_block	= qfq_tcf_block,
1502 	.bind_tcf	= qfq_bind_tcf,
1503 	.unbind_tcf	= qfq_unbind_tcf,
1504 	.graft		= qfq_graft_class,
1505 	.leaf		= qfq_class_leaf,
1506 	.qlen_notify	= qfq_qlen_notify,
1507 	.dump		= qfq_dump_class,
1508 	.dump_stats	= qfq_dump_class_stats,
1509 	.walk		= qfq_walk,
1510 };
1511 
1512 static struct Qdisc_ops qfq_qdisc_ops __read_mostly = {
1513 	.cl_ops		= &qfq_class_ops,
1514 	.id		= "qfq",
1515 	.priv_size	= sizeof(struct qfq_sched),
1516 	.enqueue	= qfq_enqueue,
1517 	.dequeue	= qfq_dequeue,
1518 	.peek		= qdisc_peek_dequeued,
1519 	.init		= qfq_init_qdisc,
1520 	.reset		= qfq_reset_qdisc,
1521 	.destroy	= qfq_destroy_qdisc,
1522 	.owner		= THIS_MODULE,
1523 };
1524 
qfq_init(void)1525 static int __init qfq_init(void)
1526 {
1527 	return register_qdisc(&qfq_qdisc_ops);
1528 }
1529 
qfq_exit(void)1530 static void __exit qfq_exit(void)
1531 {
1532 	unregister_qdisc(&qfq_qdisc_ops);
1533 }
1534 
1535 module_init(qfq_init);
1536 module_exit(qfq_exit);
1537 MODULE_LICENSE("GPL");
1538