1 /*
2  * Hierarchical Budget Worst-case Fair Weighted Fair Queueing
3  * (B-WF2Q+): hierarchical scheduling algorithm by which the BFQ I/O
4  * scheduler schedules generic entities. The latter can represent
5  * either single bfq queues (associated with processes) or groups of
6  * bfq queues (associated with cgroups).
7  *
8  *  This program is free software; you can redistribute it and/or
9  *  modify it under the terms of the GNU General Public License as
10  *  published by the Free Software Foundation; either version 2 of the
11  *  License, or (at your option) any later version.
12  *
13  *  This program is distributed in the hope that it will be useful,
14  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
15  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  *  General Public License for more details.
17  */
18 #include "bfq-iosched.h"
19 
20 /**
21  * bfq_gt - compare two timestamps.
22  * @a: first ts.
23  * @b: second ts.
24  *
25  * Return @a > @b, dealing with wrapping correctly.
26  */
bfq_gt(u64 a,u64 b)27 static int bfq_gt(u64 a, u64 b)
28 {
29 	return (s64)(a - b) > 0;
30 }
31 
bfq_root_active_entity(struct rb_root * tree)32 static struct bfq_entity *bfq_root_active_entity(struct rb_root *tree)
33 {
34 	struct rb_node *node = tree->rb_node;
35 
36 	return rb_entry(node, struct bfq_entity, rb_node);
37 }
38 
bfq_class_idx(struct bfq_entity * entity)39 static unsigned int bfq_class_idx(struct bfq_entity *entity)
40 {
41 	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
42 
43 	return bfqq ? bfqq->ioprio_class - 1 :
44 		BFQ_DEFAULT_GRP_CLASS - 1;
45 }
46 
47 static struct bfq_entity *bfq_lookup_next_entity(struct bfq_sched_data *sd,
48 						 bool expiration);
49 
50 static bool bfq_update_parent_budget(struct bfq_entity *next_in_service);
51 
52 /**
53  * bfq_update_next_in_service - update sd->next_in_service
54  * @sd: sched_data for which to perform the update.
55  * @new_entity: if not NULL, pointer to the entity whose activation,
56  *		requeueing or repositionig triggered the invocation of
57  *		this function.
58  * @expiration: id true, this function is being invoked after the
59  *             expiration of the in-service entity
60  *
61  * This function is called to update sd->next_in_service, which, in
62  * its turn, may change as a consequence of the insertion or
63  * extraction of an entity into/from one of the active trees of
64  * sd. These insertions/extractions occur as a consequence of
65  * activations/deactivations of entities, with some activations being
66  * 'true' activations, and other activations being requeueings (i.e.,
67  * implementing the second, requeueing phase of the mechanism used to
68  * reposition an entity in its active tree; see comments on
69  * __bfq_activate_entity and __bfq_requeue_entity for details). In
70  * both the last two activation sub-cases, new_entity points to the
71  * just activated or requeued entity.
72  *
73  * Returns true if sd->next_in_service changes in such a way that
74  * entity->parent may become the next_in_service for its parent
75  * entity.
76  */
bfq_update_next_in_service(struct bfq_sched_data * sd,struct bfq_entity * new_entity,bool expiration)77 static bool bfq_update_next_in_service(struct bfq_sched_data *sd,
78 				       struct bfq_entity *new_entity,
79 				       bool expiration)
80 {
81 	struct bfq_entity *next_in_service = sd->next_in_service;
82 	bool parent_sched_may_change = false;
83 	bool change_without_lookup = false;
84 
85 	/*
86 	 * If this update is triggered by the activation, requeueing
87 	 * or repositiong of an entity that does not coincide with
88 	 * sd->next_in_service, then a full lookup in the active tree
89 	 * can be avoided. In fact, it is enough to check whether the
90 	 * just-modified entity has the same priority as
91 	 * sd->next_in_service, is eligible and has a lower virtual
92 	 * finish time than sd->next_in_service. If this compound
93 	 * condition holds, then the new entity becomes the new
94 	 * next_in_service. Otherwise no change is needed.
95 	 */
96 	if (new_entity && new_entity != sd->next_in_service) {
97 		/*
98 		 * Flag used to decide whether to replace
99 		 * sd->next_in_service with new_entity. Tentatively
100 		 * set to true, and left as true if
101 		 * sd->next_in_service is NULL.
102 		 */
103 		change_without_lookup = true;
104 
105 		/*
106 		 * If there is already a next_in_service candidate
107 		 * entity, then compare timestamps to decide whether
108 		 * to replace sd->service_tree with new_entity.
109 		 */
110 		if (next_in_service) {
111 			unsigned int new_entity_class_idx =
112 				bfq_class_idx(new_entity);
113 			struct bfq_service_tree *st =
114 				sd->service_tree + new_entity_class_idx;
115 
116 			change_without_lookup =
117 				(new_entity_class_idx ==
118 				 bfq_class_idx(next_in_service)
119 				 &&
120 				 !bfq_gt(new_entity->start, st->vtime)
121 				 &&
122 				 bfq_gt(next_in_service->finish,
123 					new_entity->finish));
124 		}
125 
126 		if (change_without_lookup)
127 			next_in_service = new_entity;
128 	}
129 
130 	if (!change_without_lookup) /* lookup needed */
131 		next_in_service = bfq_lookup_next_entity(sd, expiration);
132 
133 	if (next_in_service) {
134 		bool new_budget_triggers_change =
135 			bfq_update_parent_budget(next_in_service);
136 
137 		parent_sched_may_change = !sd->next_in_service ||
138 			new_budget_triggers_change;
139 	}
140 
141 	sd->next_in_service = next_in_service;
142 
143 	if (!next_in_service)
144 		return parent_sched_may_change;
145 
146 	return parent_sched_may_change;
147 }
148 
149 #ifdef CONFIG_BFQ_GROUP_IOSCHED
150 
bfq_bfqq_to_bfqg(struct bfq_queue * bfqq)151 struct bfq_group *bfq_bfqq_to_bfqg(struct bfq_queue *bfqq)
152 {
153 	struct bfq_entity *group_entity = bfqq->entity.parent;
154 
155 	if (!group_entity)
156 		group_entity = &bfqq->bfqd->root_group->entity;
157 
158 	return container_of(group_entity, struct bfq_group, entity);
159 }
160 
161 /*
162  * Returns true if this budget changes may let next_in_service->parent
163  * become the next_in_service entity for its parent entity.
164  */
bfq_update_parent_budget(struct bfq_entity * next_in_service)165 static bool bfq_update_parent_budget(struct bfq_entity *next_in_service)
166 {
167 	struct bfq_entity *bfqg_entity;
168 	struct bfq_group *bfqg;
169 	struct bfq_sched_data *group_sd;
170 	bool ret = false;
171 
172 	group_sd = next_in_service->sched_data;
173 
174 	bfqg = container_of(group_sd, struct bfq_group, sched_data);
175 	/*
176 	 * bfq_group's my_entity field is not NULL only if the group
177 	 * is not the root group. We must not touch the root entity
178 	 * as it must never become an in-service entity.
179 	 */
180 	bfqg_entity = bfqg->my_entity;
181 	if (bfqg_entity) {
182 		if (bfqg_entity->budget > next_in_service->budget)
183 			ret = true;
184 		bfqg_entity->budget = next_in_service->budget;
185 	}
186 
187 	return ret;
188 }
189 
190 /*
191  * This function tells whether entity stops being a candidate for next
192  * service, according to the restrictive definition of the field
193  * next_in_service. In particular, this function is invoked for an
194  * entity that is about to be set in service.
195  *
196  * If entity is a queue, then the entity is no longer a candidate for
197  * next service according to the that definition, because entity is
198  * about to become the in-service queue. This function then returns
199  * true if entity is a queue.
200  *
201  * In contrast, entity could still be a candidate for next service if
202  * it is not a queue, and has more than one active child. In fact,
203  * even if one of its children is about to be set in service, other
204  * active children may still be the next to serve, for the parent
205  * entity, even according to the above definition. As a consequence, a
206  * non-queue entity is not a candidate for next-service only if it has
207  * only one active child. And only if this condition holds, then this
208  * function returns true for a non-queue entity.
209  */
bfq_no_longer_next_in_service(struct bfq_entity * entity)210 static bool bfq_no_longer_next_in_service(struct bfq_entity *entity)
211 {
212 	struct bfq_group *bfqg;
213 
214 	if (bfq_entity_to_bfqq(entity))
215 		return true;
216 
217 	bfqg = container_of(entity, struct bfq_group, entity);
218 
219 	/*
220 	 * The field active_entities does not always contain the
221 	 * actual number of active children entities: it happens to
222 	 * not account for the in-service entity in case the latter is
223 	 * removed from its active tree (which may get done after
224 	 * invoking the function bfq_no_longer_next_in_service in
225 	 * bfq_get_next_queue). Fortunately, here, i.e., while
226 	 * bfq_no_longer_next_in_service is not yet completed in
227 	 * bfq_get_next_queue, bfq_active_extract has not yet been
228 	 * invoked, and thus active_entities still coincides with the
229 	 * actual number of active entities.
230 	 */
231 	if (bfqg->active_entities == 1)
232 		return true;
233 
234 	return false;
235 }
236 
237 #else /* CONFIG_BFQ_GROUP_IOSCHED */
238 
bfq_bfqq_to_bfqg(struct bfq_queue * bfqq)239 struct bfq_group *bfq_bfqq_to_bfqg(struct bfq_queue *bfqq)
240 {
241 	return bfqq->bfqd->root_group;
242 }
243 
bfq_update_parent_budget(struct bfq_entity * next_in_service)244 static bool bfq_update_parent_budget(struct bfq_entity *next_in_service)
245 {
246 	return false;
247 }
248 
bfq_no_longer_next_in_service(struct bfq_entity * entity)249 static bool bfq_no_longer_next_in_service(struct bfq_entity *entity)
250 {
251 	return true;
252 }
253 
254 #endif /* CONFIG_BFQ_GROUP_IOSCHED */
255 
256 /*
257  * Shift for timestamp calculations.  This actually limits the maximum
258  * service allowed in one timestamp delta (small shift values increase it),
259  * the maximum total weight that can be used for the queues in the system
260  * (big shift values increase it), and the period of virtual time
261  * wraparounds.
262  */
263 #define WFQ_SERVICE_SHIFT	22
264 
bfq_entity_to_bfqq(struct bfq_entity * entity)265 struct bfq_queue *bfq_entity_to_bfqq(struct bfq_entity *entity)
266 {
267 	struct bfq_queue *bfqq = NULL;
268 
269 	if (!entity->my_sched_data)
270 		bfqq = container_of(entity, struct bfq_queue, entity);
271 
272 	return bfqq;
273 }
274 
275 
276 /**
277  * bfq_delta - map service into the virtual time domain.
278  * @service: amount of service.
279  * @weight: scale factor (weight of an entity or weight sum).
280  */
bfq_delta(unsigned long service,unsigned long weight)281 static u64 bfq_delta(unsigned long service, unsigned long weight)
282 {
283 	u64 d = (u64)service << WFQ_SERVICE_SHIFT;
284 
285 	do_div(d, weight);
286 	return d;
287 }
288 
289 /**
290  * bfq_calc_finish - assign the finish time to an entity.
291  * @entity: the entity to act upon.
292  * @service: the service to be charged to the entity.
293  */
bfq_calc_finish(struct bfq_entity * entity,unsigned long service)294 static void bfq_calc_finish(struct bfq_entity *entity, unsigned long service)
295 {
296 	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
297 
298 	entity->finish = entity->start +
299 		bfq_delta(service, entity->weight);
300 
301 	if (bfqq) {
302 		bfq_log_bfqq(bfqq->bfqd, bfqq,
303 			"calc_finish: serv %lu, w %d",
304 			service, entity->weight);
305 		bfq_log_bfqq(bfqq->bfqd, bfqq,
306 			"calc_finish: start %llu, finish %llu, delta %llu",
307 			entity->start, entity->finish,
308 			bfq_delta(service, entity->weight));
309 	}
310 }
311 
312 /**
313  * bfq_entity_of - get an entity from a node.
314  * @node: the node field of the entity.
315  *
316  * Convert a node pointer to the relative entity.  This is used only
317  * to simplify the logic of some functions and not as the generic
318  * conversion mechanism because, e.g., in the tree walking functions,
319  * the check for a %NULL value would be redundant.
320  */
bfq_entity_of(struct rb_node * node)321 struct bfq_entity *bfq_entity_of(struct rb_node *node)
322 {
323 	struct bfq_entity *entity = NULL;
324 
325 	if (node)
326 		entity = rb_entry(node, struct bfq_entity, rb_node);
327 
328 	return entity;
329 }
330 
331 /**
332  * bfq_extract - remove an entity from a tree.
333  * @root: the tree root.
334  * @entity: the entity to remove.
335  */
bfq_extract(struct rb_root * root,struct bfq_entity * entity)336 static void bfq_extract(struct rb_root *root, struct bfq_entity *entity)
337 {
338 	entity->tree = NULL;
339 	rb_erase(&entity->rb_node, root);
340 }
341 
342 /**
343  * bfq_idle_extract - extract an entity from the idle tree.
344  * @st: the service tree of the owning @entity.
345  * @entity: the entity being removed.
346  */
bfq_idle_extract(struct bfq_service_tree * st,struct bfq_entity * entity)347 static void bfq_idle_extract(struct bfq_service_tree *st,
348 			     struct bfq_entity *entity)
349 {
350 	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
351 	struct rb_node *next;
352 
353 	if (entity == st->first_idle) {
354 		next = rb_next(&entity->rb_node);
355 		st->first_idle = bfq_entity_of(next);
356 	}
357 
358 	if (entity == st->last_idle) {
359 		next = rb_prev(&entity->rb_node);
360 		st->last_idle = bfq_entity_of(next);
361 	}
362 
363 	bfq_extract(&st->idle, entity);
364 
365 	if (bfqq)
366 		list_del(&bfqq->bfqq_list);
367 }
368 
369 /**
370  * bfq_insert - generic tree insertion.
371  * @root: tree root.
372  * @entity: entity to insert.
373  *
374  * This is used for the idle and the active tree, since they are both
375  * ordered by finish time.
376  */
bfq_insert(struct rb_root * root,struct bfq_entity * entity)377 static void bfq_insert(struct rb_root *root, struct bfq_entity *entity)
378 {
379 	struct bfq_entity *entry;
380 	struct rb_node **node = &root->rb_node;
381 	struct rb_node *parent = NULL;
382 
383 	while (*node) {
384 		parent = *node;
385 		entry = rb_entry(parent, struct bfq_entity, rb_node);
386 
387 		if (bfq_gt(entry->finish, entity->finish))
388 			node = &parent->rb_left;
389 		else
390 			node = &parent->rb_right;
391 	}
392 
393 	rb_link_node(&entity->rb_node, parent, node);
394 	rb_insert_color(&entity->rb_node, root);
395 
396 	entity->tree = root;
397 }
398 
399 /**
400  * bfq_update_min - update the min_start field of a entity.
401  * @entity: the entity to update.
402  * @node: one of its children.
403  *
404  * This function is called when @entity may store an invalid value for
405  * min_start due to updates to the active tree.  The function  assumes
406  * that the subtree rooted at @node (which may be its left or its right
407  * child) has a valid min_start value.
408  */
bfq_update_min(struct bfq_entity * entity,struct rb_node * node)409 static void bfq_update_min(struct bfq_entity *entity, struct rb_node *node)
410 {
411 	struct bfq_entity *child;
412 
413 	if (node) {
414 		child = rb_entry(node, struct bfq_entity, rb_node);
415 		if (bfq_gt(entity->min_start, child->min_start))
416 			entity->min_start = child->min_start;
417 	}
418 }
419 
420 /**
421  * bfq_update_active_node - recalculate min_start.
422  * @node: the node to update.
423  *
424  * @node may have changed position or one of its children may have moved,
425  * this function updates its min_start value.  The left and right subtrees
426  * are assumed to hold a correct min_start value.
427  */
bfq_update_active_node(struct rb_node * node)428 static void bfq_update_active_node(struct rb_node *node)
429 {
430 	struct bfq_entity *entity = rb_entry(node, struct bfq_entity, rb_node);
431 
432 	entity->min_start = entity->start;
433 	bfq_update_min(entity, node->rb_right);
434 	bfq_update_min(entity, node->rb_left);
435 }
436 
437 /**
438  * bfq_update_active_tree - update min_start for the whole active tree.
439  * @node: the starting node.
440  *
441  * @node must be the deepest modified node after an update.  This function
442  * updates its min_start using the values held by its children, assuming
443  * that they did not change, and then updates all the nodes that may have
444  * changed in the path to the root.  The only nodes that may have changed
445  * are the ones in the path or their siblings.
446  */
bfq_update_active_tree(struct rb_node * node)447 static void bfq_update_active_tree(struct rb_node *node)
448 {
449 	struct rb_node *parent;
450 
451 up:
452 	bfq_update_active_node(node);
453 
454 	parent = rb_parent(node);
455 	if (!parent)
456 		return;
457 
458 	if (node == parent->rb_left && parent->rb_right)
459 		bfq_update_active_node(parent->rb_right);
460 	else if (parent->rb_left)
461 		bfq_update_active_node(parent->rb_left);
462 
463 	node = parent;
464 	goto up;
465 }
466 
467 /**
468  * bfq_active_insert - insert an entity in the active tree of its
469  *                     group/device.
470  * @st: the service tree of the entity.
471  * @entity: the entity being inserted.
472  *
473  * The active tree is ordered by finish time, but an extra key is kept
474  * per each node, containing the minimum value for the start times of
475  * its children (and the node itself), so it's possible to search for
476  * the eligible node with the lowest finish time in logarithmic time.
477  */
bfq_active_insert(struct bfq_service_tree * st,struct bfq_entity * entity)478 static void bfq_active_insert(struct bfq_service_tree *st,
479 			      struct bfq_entity *entity)
480 {
481 	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
482 	struct rb_node *node = &entity->rb_node;
483 #ifdef CONFIG_BFQ_GROUP_IOSCHED
484 	struct bfq_sched_data *sd = NULL;
485 	struct bfq_group *bfqg = NULL;
486 	struct bfq_data *bfqd = NULL;
487 #endif
488 
489 	bfq_insert(&st->active, entity);
490 
491 	if (node->rb_left)
492 		node = node->rb_left;
493 	else if (node->rb_right)
494 		node = node->rb_right;
495 
496 	bfq_update_active_tree(node);
497 
498 #ifdef CONFIG_BFQ_GROUP_IOSCHED
499 	sd = entity->sched_data;
500 	bfqg = container_of(sd, struct bfq_group, sched_data);
501 	bfqd = (struct bfq_data *)bfqg->bfqd;
502 #endif
503 	if (bfqq)
504 		list_add(&bfqq->bfqq_list, &bfqq->bfqd->active_list);
505 #ifdef CONFIG_BFQ_GROUP_IOSCHED
506 	if (bfqg != bfqd->root_group)
507 		bfqg->active_entities++;
508 #endif
509 }
510 
511 /**
512  * bfq_ioprio_to_weight - calc a weight from an ioprio.
513  * @ioprio: the ioprio value to convert.
514  */
bfq_ioprio_to_weight(int ioprio)515 unsigned short bfq_ioprio_to_weight(int ioprio)
516 {
517 	return (IOPRIO_BE_NR - ioprio) * BFQ_WEIGHT_CONVERSION_COEFF;
518 }
519 
520 /**
521  * bfq_weight_to_ioprio - calc an ioprio from a weight.
522  * @weight: the weight value to convert.
523  *
524  * To preserve as much as possible the old only-ioprio user interface,
525  * 0 is used as an escape ioprio value for weights (numerically) equal or
526  * larger than IOPRIO_BE_NR * BFQ_WEIGHT_CONVERSION_COEFF.
527  */
bfq_weight_to_ioprio(int weight)528 static unsigned short bfq_weight_to_ioprio(int weight)
529 {
530 	return max_t(int, 0,
531 		     IOPRIO_BE_NR * BFQ_WEIGHT_CONVERSION_COEFF - weight);
532 }
533 
bfq_get_entity(struct bfq_entity * entity)534 static void bfq_get_entity(struct bfq_entity *entity)
535 {
536 	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
537 
538 	if (bfqq) {
539 		bfqq->ref++;
540 		bfq_log_bfqq(bfqq->bfqd, bfqq, "get_entity: %p %d",
541 			     bfqq, bfqq->ref);
542 	}
543 }
544 
545 /**
546  * bfq_find_deepest - find the deepest node that an extraction can modify.
547  * @node: the node being removed.
548  *
549  * Do the first step of an extraction in an rb tree, looking for the
550  * node that will replace @node, and returning the deepest node that
551  * the following modifications to the tree can touch.  If @node is the
552  * last node in the tree return %NULL.
553  */
bfq_find_deepest(struct rb_node * node)554 static struct rb_node *bfq_find_deepest(struct rb_node *node)
555 {
556 	struct rb_node *deepest;
557 
558 	if (!node->rb_right && !node->rb_left)
559 		deepest = rb_parent(node);
560 	else if (!node->rb_right)
561 		deepest = node->rb_left;
562 	else if (!node->rb_left)
563 		deepest = node->rb_right;
564 	else {
565 		deepest = rb_next(node);
566 		if (deepest->rb_right)
567 			deepest = deepest->rb_right;
568 		else if (rb_parent(deepest) != node)
569 			deepest = rb_parent(deepest);
570 	}
571 
572 	return deepest;
573 }
574 
575 /**
576  * bfq_active_extract - remove an entity from the active tree.
577  * @st: the service_tree containing the tree.
578  * @entity: the entity being removed.
579  */
bfq_active_extract(struct bfq_service_tree * st,struct bfq_entity * entity)580 static void bfq_active_extract(struct bfq_service_tree *st,
581 			       struct bfq_entity *entity)
582 {
583 	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
584 	struct rb_node *node;
585 #ifdef CONFIG_BFQ_GROUP_IOSCHED
586 	struct bfq_sched_data *sd = NULL;
587 	struct bfq_group *bfqg = NULL;
588 	struct bfq_data *bfqd = NULL;
589 #endif
590 
591 	node = bfq_find_deepest(&entity->rb_node);
592 	bfq_extract(&st->active, entity);
593 
594 	if (node)
595 		bfq_update_active_tree(node);
596 
597 #ifdef CONFIG_BFQ_GROUP_IOSCHED
598 	sd = entity->sched_data;
599 	bfqg = container_of(sd, struct bfq_group, sched_data);
600 	bfqd = (struct bfq_data *)bfqg->bfqd;
601 #endif
602 	if (bfqq)
603 		list_del(&bfqq->bfqq_list);
604 #ifdef CONFIG_BFQ_GROUP_IOSCHED
605 	if (bfqg != bfqd->root_group)
606 		bfqg->active_entities--;
607 #endif
608 }
609 
610 /**
611  * bfq_idle_insert - insert an entity into the idle tree.
612  * @st: the service tree containing the tree.
613  * @entity: the entity to insert.
614  */
bfq_idle_insert(struct bfq_service_tree * st,struct bfq_entity * entity)615 static void bfq_idle_insert(struct bfq_service_tree *st,
616 			    struct bfq_entity *entity)
617 {
618 	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
619 	struct bfq_entity *first_idle = st->first_idle;
620 	struct bfq_entity *last_idle = st->last_idle;
621 
622 	if (!first_idle || bfq_gt(first_idle->finish, entity->finish))
623 		st->first_idle = entity;
624 	if (!last_idle || bfq_gt(entity->finish, last_idle->finish))
625 		st->last_idle = entity;
626 
627 	bfq_insert(&st->idle, entity);
628 
629 	if (bfqq)
630 		list_add(&bfqq->bfqq_list, &bfqq->bfqd->idle_list);
631 }
632 
633 /**
634  * bfq_forget_entity - do not consider entity any longer for scheduling
635  * @st: the service tree.
636  * @entity: the entity being removed.
637  * @is_in_service: true if entity is currently the in-service entity.
638  *
639  * Forget everything about @entity. In addition, if entity represents
640  * a queue, and the latter is not in service, then release the service
641  * reference to the queue (the one taken through bfq_get_entity). In
642  * fact, in this case, there is really no more service reference to
643  * the queue, as the latter is also outside any service tree. If,
644  * instead, the queue is in service, then __bfq_bfqd_reset_in_service
645  * will take care of putting the reference when the queue finally
646  * stops being served.
647  */
bfq_forget_entity(struct bfq_service_tree * st,struct bfq_entity * entity,bool is_in_service)648 static void bfq_forget_entity(struct bfq_service_tree *st,
649 			      struct bfq_entity *entity,
650 			      bool is_in_service)
651 {
652 	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
653 
654 	entity->on_st = false;
655 	st->wsum -= entity->weight;
656 	if (bfqq && !is_in_service)
657 		bfq_put_queue(bfqq);
658 }
659 
660 /**
661  * bfq_put_idle_entity - release the idle tree ref of an entity.
662  * @st: service tree for the entity.
663  * @entity: the entity being released.
664  */
bfq_put_idle_entity(struct bfq_service_tree * st,struct bfq_entity * entity)665 void bfq_put_idle_entity(struct bfq_service_tree *st, struct bfq_entity *entity)
666 {
667 	bfq_idle_extract(st, entity);
668 	bfq_forget_entity(st, entity,
669 			  entity == entity->sched_data->in_service_entity);
670 }
671 
672 /**
673  * bfq_forget_idle - update the idle tree if necessary.
674  * @st: the service tree to act upon.
675  *
676  * To preserve the global O(log N) complexity we only remove one entry here;
677  * as the idle tree will not grow indefinitely this can be done safely.
678  */
bfq_forget_idle(struct bfq_service_tree * st)679 static void bfq_forget_idle(struct bfq_service_tree *st)
680 {
681 	struct bfq_entity *first_idle = st->first_idle;
682 	struct bfq_entity *last_idle = st->last_idle;
683 
684 	if (RB_EMPTY_ROOT(&st->active) && last_idle &&
685 	    !bfq_gt(last_idle->finish, st->vtime)) {
686 		/*
687 		 * Forget the whole idle tree, increasing the vtime past
688 		 * the last finish time of idle entities.
689 		 */
690 		st->vtime = last_idle->finish;
691 	}
692 
693 	if (first_idle && !bfq_gt(first_idle->finish, st->vtime))
694 		bfq_put_idle_entity(st, first_idle);
695 }
696 
bfq_entity_service_tree(struct bfq_entity * entity)697 struct bfq_service_tree *bfq_entity_service_tree(struct bfq_entity *entity)
698 {
699 	struct bfq_sched_data *sched_data = entity->sched_data;
700 	unsigned int idx = bfq_class_idx(entity);
701 
702 	return sched_data->service_tree + idx;
703 }
704 
705 /*
706  * Update weight and priority of entity. If update_class_too is true,
707  * then update the ioprio_class of entity too.
708  *
709  * The reason why the update of ioprio_class is controlled through the
710  * last parameter is as follows. Changing the ioprio class of an
711  * entity implies changing the destination service trees for that
712  * entity. If such a change occurred when the entity is already on one
713  * of the service trees for its previous class, then the state of the
714  * entity would become more complex: none of the new possible service
715  * trees for the entity, according to bfq_entity_service_tree(), would
716  * match any of the possible service trees on which the entity
717  * is. Complex operations involving these trees, such as entity
718  * activations and deactivations, should take into account this
719  * additional complexity.  To avoid this issue, this function is
720  * invoked with update_class_too unset in the points in the code where
721  * entity may happen to be on some tree.
722  */
723 struct bfq_service_tree *
__bfq_entity_update_weight_prio(struct bfq_service_tree * old_st,struct bfq_entity * entity,bool update_class_too)724 __bfq_entity_update_weight_prio(struct bfq_service_tree *old_st,
725 				struct bfq_entity *entity,
726 				bool update_class_too)
727 {
728 	struct bfq_service_tree *new_st = old_st;
729 
730 	if (entity->prio_changed) {
731 		struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
732 		unsigned int prev_weight, new_weight;
733 		struct bfq_data *bfqd = NULL;
734 		struct rb_root *root;
735 #ifdef CONFIG_BFQ_GROUP_IOSCHED
736 		struct bfq_sched_data *sd;
737 		struct bfq_group *bfqg;
738 #endif
739 
740 		if (bfqq)
741 			bfqd = bfqq->bfqd;
742 #ifdef CONFIG_BFQ_GROUP_IOSCHED
743 		else {
744 			sd = entity->my_sched_data;
745 			bfqg = container_of(sd, struct bfq_group, sched_data);
746 			bfqd = (struct bfq_data *)bfqg->bfqd;
747 		}
748 #endif
749 
750 		old_st->wsum -= entity->weight;
751 
752 		if (entity->new_weight != entity->orig_weight) {
753 			if (entity->new_weight < BFQ_MIN_WEIGHT ||
754 			    entity->new_weight > BFQ_MAX_WEIGHT) {
755 				pr_crit("update_weight_prio: new_weight %d\n",
756 					entity->new_weight);
757 				if (entity->new_weight < BFQ_MIN_WEIGHT)
758 					entity->new_weight = BFQ_MIN_WEIGHT;
759 				else
760 					entity->new_weight = BFQ_MAX_WEIGHT;
761 			}
762 			entity->orig_weight = entity->new_weight;
763 			if (bfqq)
764 				bfqq->ioprio =
765 				  bfq_weight_to_ioprio(entity->orig_weight);
766 		}
767 
768 		if (bfqq && update_class_too)
769 			bfqq->ioprio_class = bfqq->new_ioprio_class;
770 
771 		/*
772 		 * Reset prio_changed only if the ioprio_class change
773 		 * is not pending any longer.
774 		 */
775 		if (!bfqq || bfqq->ioprio_class == bfqq->new_ioprio_class)
776 			entity->prio_changed = 0;
777 
778 		/*
779 		 * NOTE: here we may be changing the weight too early,
780 		 * this will cause unfairness.  The correct approach
781 		 * would have required additional complexity to defer
782 		 * weight changes to the proper time instants (i.e.,
783 		 * when entity->finish <= old_st->vtime).
784 		 */
785 		new_st = bfq_entity_service_tree(entity);
786 
787 		prev_weight = entity->weight;
788 		new_weight = entity->orig_weight *
789 			     (bfqq ? bfqq->wr_coeff : 1);
790 		/*
791 		 * If the weight of the entity changes, and the entity is a
792 		 * queue, remove the entity from its old weight counter (if
793 		 * there is a counter associated with the entity).
794 		 */
795 		if (prev_weight != new_weight && bfqq) {
796 			root = &bfqd->queue_weights_tree;
797 			__bfq_weights_tree_remove(bfqd, bfqq, root);
798 		}
799 		entity->weight = new_weight;
800 		/*
801 		 * Add the entity, if it is not a weight-raised queue,
802 		 * to the counter associated with its new weight.
803 		 */
804 		if (prev_weight != new_weight && bfqq && bfqq->wr_coeff == 1) {
805 			/* If we get here, root has been initialized. */
806 			bfq_weights_tree_add(bfqd, bfqq, root);
807 		}
808 
809 		new_st->wsum += entity->weight;
810 
811 		if (new_st != old_st)
812 			entity->start = new_st->vtime;
813 	}
814 
815 	return new_st;
816 }
817 
818 /**
819  * bfq_bfqq_served - update the scheduler status after selection for
820  *                   service.
821  * @bfqq: the queue being served.
822  * @served: bytes to transfer.
823  *
824  * NOTE: this can be optimized, as the timestamps of upper level entities
825  * are synchronized every time a new bfqq is selected for service.  By now,
826  * we keep it to better check consistency.
827  */
bfq_bfqq_served(struct bfq_queue * bfqq,int served)828 void bfq_bfqq_served(struct bfq_queue *bfqq, int served)
829 {
830 	struct bfq_entity *entity = &bfqq->entity;
831 	struct bfq_service_tree *st;
832 
833 	if (!bfqq->service_from_backlogged)
834 		bfqq->first_IO_time = jiffies;
835 
836 	if (bfqq->wr_coeff > 1)
837 		bfqq->service_from_wr += served;
838 
839 	bfqq->service_from_backlogged += served;
840 	for_each_entity(entity) {
841 		st = bfq_entity_service_tree(entity);
842 
843 		entity->service += served;
844 
845 		st->vtime += bfq_delta(served, st->wsum);
846 		bfq_forget_idle(st);
847 	}
848 	bfq_log_bfqq(bfqq->bfqd, bfqq, "bfqq_served %d secs", served);
849 }
850 
851 /**
852  * bfq_bfqq_charge_time - charge an amount of service equivalent to the length
853  *			  of the time interval during which bfqq has been in
854  *			  service.
855  * @bfqd: the device
856  * @bfqq: the queue that needs a service update.
857  * @time_ms: the amount of time during which the queue has received service
858  *
859  * If a queue does not consume its budget fast enough, then providing
860  * the queue with service fairness may impair throughput, more or less
861  * severely. For this reason, queues that consume their budget slowly
862  * are provided with time fairness instead of service fairness. This
863  * goal is achieved through the BFQ scheduling engine, even if such an
864  * engine works in the service, and not in the time domain. The trick
865  * is charging these queues with an inflated amount of service, equal
866  * to the amount of service that they would have received during their
867  * service slot if they had been fast, i.e., if their requests had
868  * been dispatched at a rate equal to the estimated peak rate.
869  *
870  * It is worth noting that time fairness can cause important
871  * distortions in terms of bandwidth distribution, on devices with
872  * internal queueing. The reason is that I/O requests dispatched
873  * during the service slot of a queue may be served after that service
874  * slot is finished, and may have a total processing time loosely
875  * correlated with the duration of the service slot. This is
876  * especially true for short service slots.
877  */
bfq_bfqq_charge_time(struct bfq_data * bfqd,struct bfq_queue * bfqq,unsigned long time_ms)878 void bfq_bfqq_charge_time(struct bfq_data *bfqd, struct bfq_queue *bfqq,
879 			  unsigned long time_ms)
880 {
881 	struct bfq_entity *entity = &bfqq->entity;
882 	unsigned long timeout_ms = jiffies_to_msecs(bfq_timeout);
883 	unsigned long bounded_time_ms = min(time_ms, timeout_ms);
884 	int serv_to_charge_for_time =
885 		(bfqd->bfq_max_budget * bounded_time_ms) / timeout_ms;
886 	int tot_serv_to_charge = max(serv_to_charge_for_time, entity->service);
887 
888 	/* Increase budget to avoid inconsistencies */
889 	if (tot_serv_to_charge > entity->budget)
890 		entity->budget = tot_serv_to_charge;
891 
892 	bfq_bfqq_served(bfqq,
893 			max_t(int, 0, tot_serv_to_charge - entity->service));
894 }
895 
bfq_update_fin_time_enqueue(struct bfq_entity * entity,struct bfq_service_tree * st,bool backshifted)896 static void bfq_update_fin_time_enqueue(struct bfq_entity *entity,
897 					struct bfq_service_tree *st,
898 					bool backshifted)
899 {
900 	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
901 
902 	/*
903 	 * When this function is invoked, entity is not in any service
904 	 * tree, then it is safe to invoke next function with the last
905 	 * parameter set (see the comments on the function).
906 	 */
907 	st = __bfq_entity_update_weight_prio(st, entity, true);
908 	bfq_calc_finish(entity, entity->budget);
909 
910 	/*
911 	 * If some queues enjoy backshifting for a while, then their
912 	 * (virtual) finish timestamps may happen to become lower and
913 	 * lower than the system virtual time.	In particular, if
914 	 * these queues often happen to be idle for short time
915 	 * periods, and during such time periods other queues with
916 	 * higher timestamps happen to be busy, then the backshifted
917 	 * timestamps of the former queues can become much lower than
918 	 * the system virtual time. In fact, to serve the queues with
919 	 * higher timestamps while the ones with lower timestamps are
920 	 * idle, the system virtual time may be pushed-up to much
921 	 * higher values than the finish timestamps of the idle
922 	 * queues. As a consequence, the finish timestamps of all new
923 	 * or newly activated queues may end up being much larger than
924 	 * those of lucky queues with backshifted timestamps. The
925 	 * latter queues may then monopolize the device for a lot of
926 	 * time. This would simply break service guarantees.
927 	 *
928 	 * To reduce this problem, push up a little bit the
929 	 * backshifted timestamps of the queue associated with this
930 	 * entity (only a queue can happen to have the backshifted
931 	 * flag set): just enough to let the finish timestamp of the
932 	 * queue be equal to the current value of the system virtual
933 	 * time. This may introduce a little unfairness among queues
934 	 * with backshifted timestamps, but it does not break
935 	 * worst-case fairness guarantees.
936 	 *
937 	 * As a special case, if bfqq is weight-raised, push up
938 	 * timestamps much less, to keep very low the probability that
939 	 * this push up causes the backshifted finish timestamps of
940 	 * weight-raised queues to become higher than the backshifted
941 	 * finish timestamps of non weight-raised queues.
942 	 */
943 	if (backshifted && bfq_gt(st->vtime, entity->finish)) {
944 		unsigned long delta = st->vtime - entity->finish;
945 
946 		if (bfqq)
947 			delta /= bfqq->wr_coeff;
948 
949 		entity->start += delta;
950 		entity->finish += delta;
951 	}
952 
953 	bfq_active_insert(st, entity);
954 }
955 
956 /**
957  * __bfq_activate_entity - handle activation of entity.
958  * @entity: the entity being activated.
959  * @non_blocking_wait_rq: true if entity was waiting for a request
960  *
961  * Called for a 'true' activation, i.e., if entity is not active and
962  * one of its children receives a new request.
963  *
964  * Basically, this function updates the timestamps of entity and
965  * inserts entity into its active tree, after possibly extracting it
966  * from its idle tree.
967  */
__bfq_activate_entity(struct bfq_entity * entity,bool non_blocking_wait_rq)968 static void __bfq_activate_entity(struct bfq_entity *entity,
969 				  bool non_blocking_wait_rq)
970 {
971 	struct bfq_service_tree *st = bfq_entity_service_tree(entity);
972 	bool backshifted = false;
973 	unsigned long long min_vstart;
974 
975 	/* See comments on bfq_fqq_update_budg_for_activation */
976 	if (non_blocking_wait_rq && bfq_gt(st->vtime, entity->finish)) {
977 		backshifted = true;
978 		min_vstart = entity->finish;
979 	} else
980 		min_vstart = st->vtime;
981 
982 	if (entity->tree == &st->idle) {
983 		/*
984 		 * Must be on the idle tree, bfq_idle_extract() will
985 		 * check for that.
986 		 */
987 		bfq_idle_extract(st, entity);
988 		entity->start = bfq_gt(min_vstart, entity->finish) ?
989 			min_vstart : entity->finish;
990 	} else {
991 		/*
992 		 * The finish time of the entity may be invalid, and
993 		 * it is in the past for sure, otherwise the queue
994 		 * would have been on the idle tree.
995 		 */
996 		entity->start = min_vstart;
997 		st->wsum += entity->weight;
998 		/*
999 		 * entity is about to be inserted into a service tree,
1000 		 * and then set in service: get a reference to make
1001 		 * sure entity does not disappear until it is no
1002 		 * longer in service or scheduled for service.
1003 		 */
1004 		bfq_get_entity(entity);
1005 
1006 		entity->on_st = true;
1007 	}
1008 
1009 #ifdef BFQ_GROUP_IOSCHED_ENABLED
1010 	if (!bfq_entity_to_bfqq(entity)) { /* bfq_group */
1011 		struct bfq_group *bfqg =
1012 			container_of(entity, struct bfq_group, entity);
1013 		struct bfq_data *bfqd = bfqg->bfqd;
1014 
1015 		if (!entity->in_groups_with_pending_reqs) {
1016 			entity->in_groups_with_pending_reqs = true;
1017 			bfqd->num_groups_with_pending_reqs++;
1018 		}
1019 	}
1020 #endif
1021 
1022 	bfq_update_fin_time_enqueue(entity, st, backshifted);
1023 }
1024 
1025 /**
1026  * __bfq_requeue_entity - handle requeueing or repositioning of an entity.
1027  * @entity: the entity being requeued or repositioned.
1028  *
1029  * Requeueing is needed if this entity stops being served, which
1030  * happens if a leaf descendant entity has expired. On the other hand,
1031  * repositioning is needed if the next_inservice_entity for the child
1032  * entity has changed. See the comments inside the function for
1033  * details.
1034  *
1035  * Basically, this function: 1) removes entity from its active tree if
1036  * present there, 2) updates the timestamps of entity and 3) inserts
1037  * entity back into its active tree (in the new, right position for
1038  * the new values of the timestamps).
1039  */
__bfq_requeue_entity(struct bfq_entity * entity)1040 static void __bfq_requeue_entity(struct bfq_entity *entity)
1041 {
1042 	struct bfq_sched_data *sd = entity->sched_data;
1043 	struct bfq_service_tree *st = bfq_entity_service_tree(entity);
1044 
1045 	if (entity == sd->in_service_entity) {
1046 		/*
1047 		 * We are requeueing the current in-service entity,
1048 		 * which may have to be done for one of the following
1049 		 * reasons:
1050 		 * - entity represents the in-service queue, and the
1051 		 *   in-service queue is being requeued after an
1052 		 *   expiration;
1053 		 * - entity represents a group, and its budget has
1054 		 *   changed because one of its child entities has
1055 		 *   just been either activated or requeued for some
1056 		 *   reason; the timestamps of the entity need then to
1057 		 *   be updated, and the entity needs to be enqueued
1058 		 *   or repositioned accordingly.
1059 		 *
1060 		 * In particular, before requeueing, the start time of
1061 		 * the entity must be moved forward to account for the
1062 		 * service that the entity has received while in
1063 		 * service. This is done by the next instructions. The
1064 		 * finish time will then be updated according to this
1065 		 * new value of the start time, and to the budget of
1066 		 * the entity.
1067 		 */
1068 		bfq_calc_finish(entity, entity->service);
1069 		entity->start = entity->finish;
1070 		/*
1071 		 * In addition, if the entity had more than one child
1072 		 * when set in service, then it was not extracted from
1073 		 * the active tree. This implies that the position of
1074 		 * the entity in the active tree may need to be
1075 		 * changed now, because we have just updated the start
1076 		 * time of the entity, and we will update its finish
1077 		 * time in a moment (the requeueing is then, more
1078 		 * precisely, a repositioning in this case). To
1079 		 * implement this repositioning, we: 1) dequeue the
1080 		 * entity here, 2) update the finish time and requeue
1081 		 * the entity according to the new timestamps below.
1082 		 */
1083 		if (entity->tree)
1084 			bfq_active_extract(st, entity);
1085 	} else { /* The entity is already active, and not in service */
1086 		/*
1087 		 * In this case, this function gets called only if the
1088 		 * next_in_service entity below this entity has
1089 		 * changed, and this change has caused the budget of
1090 		 * this entity to change, which, finally implies that
1091 		 * the finish time of this entity must be
1092 		 * updated. Such an update may cause the scheduling,
1093 		 * i.e., the position in the active tree, of this
1094 		 * entity to change. We handle this change by: 1)
1095 		 * dequeueing the entity here, 2) updating the finish
1096 		 * time and requeueing the entity according to the new
1097 		 * timestamps below. This is the same approach as the
1098 		 * non-extracted-entity sub-case above.
1099 		 */
1100 		bfq_active_extract(st, entity);
1101 	}
1102 
1103 	bfq_update_fin_time_enqueue(entity, st, false);
1104 }
1105 
__bfq_activate_requeue_entity(struct bfq_entity * entity,struct bfq_sched_data * sd,bool non_blocking_wait_rq)1106 static void __bfq_activate_requeue_entity(struct bfq_entity *entity,
1107 					  struct bfq_sched_data *sd,
1108 					  bool non_blocking_wait_rq)
1109 {
1110 	struct bfq_service_tree *st = bfq_entity_service_tree(entity);
1111 
1112 	if (sd->in_service_entity == entity || entity->tree == &st->active)
1113 		 /*
1114 		  * in service or already queued on the active tree,
1115 		  * requeue or reposition
1116 		  */
1117 		__bfq_requeue_entity(entity);
1118 	else
1119 		/*
1120 		 * Not in service and not queued on its active tree:
1121 		 * the activity is idle and this is a true activation.
1122 		 */
1123 		__bfq_activate_entity(entity, non_blocking_wait_rq);
1124 }
1125 
1126 
1127 /**
1128  * bfq_activate_requeue_entity - activate or requeue an entity representing a
1129  *				 bfq_queue, and activate, requeue or reposition
1130  *				 all ancestors for which such an update becomes
1131  *				 necessary.
1132  * @entity: the entity to activate.
1133  * @non_blocking_wait_rq: true if this entity was waiting for a request
1134  * @requeue: true if this is a requeue, which implies that bfqq is
1135  *	     being expired; thus ALL its ancestors stop being served and must
1136  *	     therefore be requeued
1137  * @expiration: true if this function is being invoked in the expiration path
1138  *             of the in-service queue
1139  */
bfq_activate_requeue_entity(struct bfq_entity * entity,bool non_blocking_wait_rq,bool requeue,bool expiration)1140 static void bfq_activate_requeue_entity(struct bfq_entity *entity,
1141 					bool non_blocking_wait_rq,
1142 					bool requeue, bool expiration)
1143 {
1144 	struct bfq_sched_data *sd;
1145 
1146 	for_each_entity(entity) {
1147 		sd = entity->sched_data;
1148 		__bfq_activate_requeue_entity(entity, sd, non_blocking_wait_rq);
1149 
1150 		if (!bfq_update_next_in_service(sd, entity, expiration) &&
1151 		    !requeue)
1152 			break;
1153 	}
1154 }
1155 
1156 /**
1157  * __bfq_deactivate_entity - deactivate an entity from its service tree.
1158  * @entity: the entity to deactivate.
1159  * @ins_into_idle_tree: if false, the entity will not be put into the
1160  *			idle tree.
1161  *
1162  * Deactivates an entity, independently of its previous state.  Must
1163  * be invoked only if entity is on a service tree. Extracts the entity
1164  * from that tree, and if necessary and allowed, puts it into the idle
1165  * tree.
1166  */
__bfq_deactivate_entity(struct bfq_entity * entity,bool ins_into_idle_tree)1167 bool __bfq_deactivate_entity(struct bfq_entity *entity, bool ins_into_idle_tree)
1168 {
1169 	struct bfq_sched_data *sd = entity->sched_data;
1170 	struct bfq_service_tree *st;
1171 	bool is_in_service;
1172 
1173 	if (!entity->on_st) /* entity never activated, or already inactive */
1174 		return false;
1175 
1176 	/*
1177 	 * If we get here, then entity is active, which implies that
1178 	 * bfq_group_set_parent has already been invoked for the group
1179 	 * represented by entity. Therefore, the field
1180 	 * entity->sched_data has been set, and we can safely use it.
1181 	 */
1182 	st = bfq_entity_service_tree(entity);
1183 	is_in_service = entity == sd->in_service_entity;
1184 
1185 	bfq_calc_finish(entity, entity->service);
1186 
1187 	if (is_in_service)
1188 		sd->in_service_entity = NULL;
1189 	else
1190 		/*
1191 		 * Non in-service entity: nobody will take care of
1192 		 * resetting its service counter on expiration. Do it
1193 		 * now.
1194 		 */
1195 		entity->service = 0;
1196 
1197 	if (entity->tree == &st->active)
1198 		bfq_active_extract(st, entity);
1199 	else if (!is_in_service && entity->tree == &st->idle)
1200 		bfq_idle_extract(st, entity);
1201 
1202 	if (!ins_into_idle_tree || !bfq_gt(entity->finish, st->vtime))
1203 		bfq_forget_entity(st, entity, is_in_service);
1204 	else
1205 		bfq_idle_insert(st, entity);
1206 
1207 	return true;
1208 }
1209 
1210 /**
1211  * bfq_deactivate_entity - deactivate an entity representing a bfq_queue.
1212  * @entity: the entity to deactivate.
1213  * @ins_into_idle_tree: true if the entity can be put into the idle tree
1214  * @expiration: true if this function is being invoked in the expiration path
1215  *             of the in-service queue
1216  */
bfq_deactivate_entity(struct bfq_entity * entity,bool ins_into_idle_tree,bool expiration)1217 static void bfq_deactivate_entity(struct bfq_entity *entity,
1218 				  bool ins_into_idle_tree,
1219 				  bool expiration)
1220 {
1221 	struct bfq_sched_data *sd;
1222 	struct bfq_entity *parent = NULL;
1223 
1224 	for_each_entity_safe(entity, parent) {
1225 		sd = entity->sched_data;
1226 
1227 		if (!__bfq_deactivate_entity(entity, ins_into_idle_tree)) {
1228 			/*
1229 			 * entity is not in any tree any more, so
1230 			 * this deactivation is a no-op, and there is
1231 			 * nothing to change for upper-level entities
1232 			 * (in case of expiration, this can never
1233 			 * happen).
1234 			 */
1235 			return;
1236 		}
1237 
1238 		if (sd->next_in_service == entity)
1239 			/*
1240 			 * entity was the next_in_service entity,
1241 			 * then, since entity has just been
1242 			 * deactivated, a new one must be found.
1243 			 */
1244 			bfq_update_next_in_service(sd, NULL, expiration);
1245 
1246 		if (sd->next_in_service || sd->in_service_entity) {
1247 			/*
1248 			 * The parent entity is still active, because
1249 			 * either next_in_service or in_service_entity
1250 			 * is not NULL. So, no further upwards
1251 			 * deactivation must be performed.  Yet,
1252 			 * next_in_service has changed.	Then the
1253 			 * schedule does need to be updated upwards.
1254 			 *
1255 			 * NOTE If in_service_entity is not NULL, then
1256 			 * next_in_service may happen to be NULL,
1257 			 * although the parent entity is evidently
1258 			 * active. This happens if 1) the entity
1259 			 * pointed by in_service_entity is the only
1260 			 * active entity in the parent entity, and 2)
1261 			 * according to the definition of
1262 			 * next_in_service, the in_service_entity
1263 			 * cannot be considered as
1264 			 * next_in_service. See the comments on the
1265 			 * definition of next_in_service for details.
1266 			 */
1267 			break;
1268 		}
1269 
1270 		/*
1271 		 * If we get here, then the parent is no more
1272 		 * backlogged and we need to propagate the
1273 		 * deactivation upwards. Thus let the loop go on.
1274 		 */
1275 
1276 		/*
1277 		 * Also let parent be queued into the idle tree on
1278 		 * deactivation, to preserve service guarantees, and
1279 		 * assuming that who invoked this function does not
1280 		 * need parent entities too to be removed completely.
1281 		 */
1282 		ins_into_idle_tree = true;
1283 	}
1284 
1285 	/*
1286 	 * If the deactivation loop is fully executed, then there are
1287 	 * no more entities to touch and next loop is not executed at
1288 	 * all. Otherwise, requeue remaining entities if they are
1289 	 * about to stop receiving service, or reposition them if this
1290 	 * is not the case.
1291 	 */
1292 	entity = parent;
1293 	for_each_entity(entity) {
1294 		/*
1295 		 * Invoke __bfq_requeue_entity on entity, even if
1296 		 * already active, to requeue/reposition it in the
1297 		 * active tree (because sd->next_in_service has
1298 		 * changed)
1299 		 */
1300 		__bfq_requeue_entity(entity);
1301 
1302 		sd = entity->sched_data;
1303 		if (!bfq_update_next_in_service(sd, entity, expiration) &&
1304 		    !expiration)
1305 			/*
1306 			 * next_in_service unchanged or not causing
1307 			 * any change in entity->parent->sd, and no
1308 			 * requeueing needed for expiration: stop
1309 			 * here.
1310 			 */
1311 			break;
1312 	}
1313 }
1314 
1315 /**
1316  * bfq_calc_vtime_jump - compute the value to which the vtime should jump,
1317  *                       if needed, to have at least one entity eligible.
1318  * @st: the service tree to act upon.
1319  *
1320  * Assumes that st is not empty.
1321  */
bfq_calc_vtime_jump(struct bfq_service_tree * st)1322 static u64 bfq_calc_vtime_jump(struct bfq_service_tree *st)
1323 {
1324 	struct bfq_entity *root_entity = bfq_root_active_entity(&st->active);
1325 
1326 	if (bfq_gt(root_entity->min_start, st->vtime))
1327 		return root_entity->min_start;
1328 
1329 	return st->vtime;
1330 }
1331 
bfq_update_vtime(struct bfq_service_tree * st,u64 new_value)1332 static void bfq_update_vtime(struct bfq_service_tree *st, u64 new_value)
1333 {
1334 	if (new_value > st->vtime) {
1335 		st->vtime = new_value;
1336 		bfq_forget_idle(st);
1337 	}
1338 }
1339 
1340 /**
1341  * bfq_first_active_entity - find the eligible entity with
1342  *                           the smallest finish time
1343  * @st: the service tree to select from.
1344  * @vtime: the system virtual to use as a reference for eligibility
1345  *
1346  * This function searches the first schedulable entity, starting from the
1347  * root of the tree and going on the left every time on this side there is
1348  * a subtree with at least one eligible (start <= vtime) entity. The path on
1349  * the right is followed only if a) the left subtree contains no eligible
1350  * entities and b) no eligible entity has been found yet.
1351  */
bfq_first_active_entity(struct bfq_service_tree * st,u64 vtime)1352 static struct bfq_entity *bfq_first_active_entity(struct bfq_service_tree *st,
1353 						  u64 vtime)
1354 {
1355 	struct bfq_entity *entry, *first = NULL;
1356 	struct rb_node *node = st->active.rb_node;
1357 
1358 	while (node) {
1359 		entry = rb_entry(node, struct bfq_entity, rb_node);
1360 left:
1361 		if (!bfq_gt(entry->start, vtime))
1362 			first = entry;
1363 
1364 		if (node->rb_left) {
1365 			entry = rb_entry(node->rb_left,
1366 					 struct bfq_entity, rb_node);
1367 			if (!bfq_gt(entry->min_start, vtime)) {
1368 				node = node->rb_left;
1369 				goto left;
1370 			}
1371 		}
1372 		if (first)
1373 			break;
1374 		node = node->rb_right;
1375 	}
1376 
1377 	return first;
1378 }
1379 
1380 /**
1381  * __bfq_lookup_next_entity - return the first eligible entity in @st.
1382  * @st: the service tree.
1383  *
1384  * If there is no in-service entity for the sched_data st belongs to,
1385  * then return the entity that will be set in service if:
1386  * 1) the parent entity this st belongs to is set in service;
1387  * 2) no entity belonging to such parent entity undergoes a state change
1388  * that would influence the timestamps of the entity (e.g., becomes idle,
1389  * becomes backlogged, changes its budget, ...).
1390  *
1391  * In this first case, update the virtual time in @st too (see the
1392  * comments on this update inside the function).
1393  *
1394  * In constrast, if there is an in-service entity, then return the
1395  * entity that would be set in service if not only the above
1396  * conditions, but also the next one held true: the currently
1397  * in-service entity, on expiration,
1398  * 1) gets a finish time equal to the current one, or
1399  * 2) is not eligible any more, or
1400  * 3) is idle.
1401  */
1402 static struct bfq_entity *
__bfq_lookup_next_entity(struct bfq_service_tree * st,bool in_service)1403 __bfq_lookup_next_entity(struct bfq_service_tree *st, bool in_service)
1404 {
1405 	struct bfq_entity *entity;
1406 	u64 new_vtime;
1407 
1408 	if (RB_EMPTY_ROOT(&st->active))
1409 		return NULL;
1410 
1411 	/*
1412 	 * Get the value of the system virtual time for which at
1413 	 * least one entity is eligible.
1414 	 */
1415 	new_vtime = bfq_calc_vtime_jump(st);
1416 
1417 	/*
1418 	 * If there is no in-service entity for the sched_data this
1419 	 * active tree belongs to, then push the system virtual time
1420 	 * up to the value that guarantees that at least one entity is
1421 	 * eligible. If, instead, there is an in-service entity, then
1422 	 * do not make any such update, because there is already an
1423 	 * eligible entity, namely the in-service one (even if the
1424 	 * entity is not on st, because it was extracted when set in
1425 	 * service).
1426 	 */
1427 	if (!in_service)
1428 		bfq_update_vtime(st, new_vtime);
1429 
1430 	entity = bfq_first_active_entity(st, new_vtime);
1431 
1432 	return entity;
1433 }
1434 
1435 /**
1436  * bfq_lookup_next_entity - return the first eligible entity in @sd.
1437  * @sd: the sched_data.
1438  * @expiration: true if we are on the expiration path of the in-service queue
1439  *
1440  * This function is invoked when there has been a change in the trees
1441  * for sd, and we need to know what is the new next entity to serve
1442  * after this change.
1443  */
bfq_lookup_next_entity(struct bfq_sched_data * sd,bool expiration)1444 static struct bfq_entity *bfq_lookup_next_entity(struct bfq_sched_data *sd,
1445 						 bool expiration)
1446 {
1447 	struct bfq_service_tree *st = sd->service_tree;
1448 	struct bfq_service_tree *idle_class_st = st + (BFQ_IOPRIO_CLASSES - 1);
1449 	struct bfq_entity *entity = NULL;
1450 	int class_idx = 0;
1451 
1452 	/*
1453 	 * Choose from idle class, if needed to guarantee a minimum
1454 	 * bandwidth to this class (and if there is some active entity
1455 	 * in idle class). This should also mitigate
1456 	 * priority-inversion problems in case a low priority task is
1457 	 * holding file system resources.
1458 	 */
1459 	if (time_is_before_jiffies(sd->bfq_class_idle_last_service +
1460 				   BFQ_CL_IDLE_TIMEOUT)) {
1461 		if (!RB_EMPTY_ROOT(&idle_class_st->active))
1462 			class_idx = BFQ_IOPRIO_CLASSES - 1;
1463 		/* About to be served if backlogged, or not yet backlogged */
1464 		sd->bfq_class_idle_last_service = jiffies;
1465 	}
1466 
1467 	/*
1468 	 * Find the next entity to serve for the highest-priority
1469 	 * class, unless the idle class needs to be served.
1470 	 */
1471 	for (; class_idx < BFQ_IOPRIO_CLASSES; class_idx++) {
1472 		/*
1473 		 * If expiration is true, then bfq_lookup_next_entity
1474 		 * is being invoked as a part of the expiration path
1475 		 * of the in-service queue. In this case, even if
1476 		 * sd->in_service_entity is not NULL,
1477 		 * sd->in_service_entiy at this point is actually not
1478 		 * in service any more, and, if needed, has already
1479 		 * been properly queued or requeued into the right
1480 		 * tree. The reason why sd->in_service_entity is still
1481 		 * not NULL here, even if expiration is true, is that
1482 		 * sd->in_service_entiy is reset as a last step in the
1483 		 * expiration path. So, if expiration is true, tell
1484 		 * __bfq_lookup_next_entity that there is no
1485 		 * sd->in_service_entity.
1486 		 */
1487 		entity = __bfq_lookup_next_entity(st + class_idx,
1488 						  sd->in_service_entity &&
1489 						  !expiration);
1490 
1491 		if (entity)
1492 			break;
1493 	}
1494 
1495 	if (!entity)
1496 		return NULL;
1497 
1498 	return entity;
1499 }
1500 
next_queue_may_preempt(struct bfq_data * bfqd)1501 bool next_queue_may_preempt(struct bfq_data *bfqd)
1502 {
1503 	struct bfq_sched_data *sd = &bfqd->root_group->sched_data;
1504 
1505 	return sd->next_in_service != sd->in_service_entity;
1506 }
1507 
1508 /*
1509  * Get next queue for service.
1510  */
bfq_get_next_queue(struct bfq_data * bfqd)1511 struct bfq_queue *bfq_get_next_queue(struct bfq_data *bfqd)
1512 {
1513 	struct bfq_entity *entity = NULL;
1514 	struct bfq_sched_data *sd;
1515 	struct bfq_queue *bfqq;
1516 
1517 	if (bfqd->busy_queues == 0)
1518 		return NULL;
1519 
1520 	/*
1521 	 * Traverse the path from the root to the leaf entity to
1522 	 * serve. Set in service all the entities visited along the
1523 	 * way.
1524 	 */
1525 	sd = &bfqd->root_group->sched_data;
1526 	for (; sd ; sd = entity->my_sched_data) {
1527 		/*
1528 		 * WARNING. We are about to set the in-service entity
1529 		 * to sd->next_in_service, i.e., to the (cached) value
1530 		 * returned by bfq_lookup_next_entity(sd) the last
1531 		 * time it was invoked, i.e., the last time when the
1532 		 * service order in sd changed as a consequence of the
1533 		 * activation or deactivation of an entity. In this
1534 		 * respect, if we execute bfq_lookup_next_entity(sd)
1535 		 * in this very moment, it may, although with low
1536 		 * probability, yield a different entity than that
1537 		 * pointed to by sd->next_in_service. This rare event
1538 		 * happens in case there was no CLASS_IDLE entity to
1539 		 * serve for sd when bfq_lookup_next_entity(sd) was
1540 		 * invoked for the last time, while there is now one
1541 		 * such entity.
1542 		 *
1543 		 * If the above event happens, then the scheduling of
1544 		 * such entity in CLASS_IDLE is postponed until the
1545 		 * service of the sd->next_in_service entity
1546 		 * finishes. In fact, when the latter is expired,
1547 		 * bfq_lookup_next_entity(sd) gets called again,
1548 		 * exactly to update sd->next_in_service.
1549 		 */
1550 
1551 		/* Make next_in_service entity become in_service_entity */
1552 		entity = sd->next_in_service;
1553 		sd->in_service_entity = entity;
1554 
1555 		/*
1556 		 * If entity is no longer a candidate for next
1557 		 * service, then it must be extracted from its active
1558 		 * tree, so as to make sure that it won't be
1559 		 * considered when computing next_in_service. See the
1560 		 * comments on the function
1561 		 * bfq_no_longer_next_in_service() for details.
1562 		 */
1563 		if (bfq_no_longer_next_in_service(entity))
1564 			bfq_active_extract(bfq_entity_service_tree(entity),
1565 					   entity);
1566 
1567 		/*
1568 		 * Even if entity is not to be extracted according to
1569 		 * the above check, a descendant entity may get
1570 		 * extracted in one of the next iterations of this
1571 		 * loop. Such an event could cause a change in
1572 		 * next_in_service for the level of the descendant
1573 		 * entity, and thus possibly back to this level.
1574 		 *
1575 		 * However, we cannot perform the resulting needed
1576 		 * update of next_in_service for this level before the
1577 		 * end of the whole loop, because, to know which is
1578 		 * the correct next-to-serve candidate entity for each
1579 		 * level, we need first to find the leaf entity to set
1580 		 * in service. In fact, only after we know which is
1581 		 * the next-to-serve leaf entity, we can discover
1582 		 * whether the parent entity of the leaf entity
1583 		 * becomes the next-to-serve, and so on.
1584 		 */
1585 	}
1586 
1587 	bfqq = bfq_entity_to_bfqq(entity);
1588 
1589 	/*
1590 	 * We can finally update all next-to-serve entities along the
1591 	 * path from the leaf entity just set in service to the root.
1592 	 */
1593 	for_each_entity(entity) {
1594 		struct bfq_sched_data *sd = entity->sched_data;
1595 
1596 		if (!bfq_update_next_in_service(sd, NULL, false))
1597 			break;
1598 	}
1599 
1600 	return bfqq;
1601 }
1602 
1603 /* returns true if the in-service queue gets freed */
__bfq_bfqd_reset_in_service(struct bfq_data * bfqd)1604 bool __bfq_bfqd_reset_in_service(struct bfq_data *bfqd)
1605 {
1606 	struct bfq_queue *in_serv_bfqq = bfqd->in_service_queue;
1607 	struct bfq_entity *in_serv_entity = &in_serv_bfqq->entity;
1608 	struct bfq_entity *entity = in_serv_entity;
1609 
1610 	bfq_clear_bfqq_wait_request(in_serv_bfqq);
1611 	hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
1612 	bfqd->in_service_queue = NULL;
1613 
1614 	/*
1615 	 * When this function is called, all in-service entities have
1616 	 * been properly deactivated or requeued, so we can safely
1617 	 * execute the final step: reset in_service_entity along the
1618 	 * path from entity to the root.
1619 	 */
1620 	for_each_entity(entity)
1621 		entity->sched_data->in_service_entity = NULL;
1622 
1623 	/*
1624 	 * in_serv_entity is no longer in service, so, if it is in no
1625 	 * service tree either, then release the service reference to
1626 	 * the queue it represents (taken with bfq_get_entity).
1627 	 */
1628 	if (!in_serv_entity->on_st) {
1629 		/*
1630 		 * If no process is referencing in_serv_bfqq any
1631 		 * longer, then the service reference may be the only
1632 		 * reference to the queue. If this is the case, then
1633 		 * bfqq gets freed here.
1634 		 */
1635 		int ref = in_serv_bfqq->ref;
1636 		bfq_put_queue(in_serv_bfqq);
1637 		if (ref == 1)
1638 			return true;
1639 	}
1640 
1641 	return false;
1642 }
1643 
bfq_deactivate_bfqq(struct bfq_data * bfqd,struct bfq_queue * bfqq,bool ins_into_idle_tree,bool expiration)1644 void bfq_deactivate_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1645 			 bool ins_into_idle_tree, bool expiration)
1646 {
1647 	struct bfq_entity *entity = &bfqq->entity;
1648 
1649 	bfq_deactivate_entity(entity, ins_into_idle_tree, expiration);
1650 }
1651 
bfq_activate_bfqq(struct bfq_data * bfqd,struct bfq_queue * bfqq)1652 void bfq_activate_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1653 {
1654 	struct bfq_entity *entity = &bfqq->entity;
1655 
1656 	bfq_activate_requeue_entity(entity, bfq_bfqq_non_blocking_wait_rq(bfqq),
1657 				    false, false);
1658 	bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1659 }
1660 
bfq_requeue_bfqq(struct bfq_data * bfqd,struct bfq_queue * bfqq,bool expiration)1661 void bfq_requeue_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1662 		      bool expiration)
1663 {
1664 	struct bfq_entity *entity = &bfqq->entity;
1665 
1666 	bfq_activate_requeue_entity(entity, false,
1667 				    bfqq == bfqd->in_service_queue, expiration);
1668 }
1669 
1670 /*
1671  * Called when the bfqq no longer has requests pending, remove it from
1672  * the service tree. As a special case, it can be invoked during an
1673  * expiration.
1674  */
bfq_del_bfqq_busy(struct bfq_data * bfqd,struct bfq_queue * bfqq,bool expiration)1675 void bfq_del_bfqq_busy(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1676 		       bool expiration)
1677 {
1678 	bfq_log_bfqq(bfqd, bfqq, "del from busy");
1679 
1680 	bfq_clear_bfqq_busy(bfqq);
1681 
1682 	bfqd->busy_queues--;
1683 
1684 	if (bfqq->wr_coeff > 1)
1685 		bfqd->wr_busy_queues--;
1686 
1687 	bfqg_stats_update_dequeue(bfqq_group(bfqq));
1688 
1689 	bfq_deactivate_bfqq(bfqd, bfqq, true, expiration);
1690 
1691 	if (!bfqq->dispatched)
1692 		bfq_weights_tree_remove(bfqd, bfqq);
1693 }
1694 
1695 /*
1696  * Called when an inactive queue receives a new request.
1697  */
bfq_add_bfqq_busy(struct bfq_data * bfqd,struct bfq_queue * bfqq)1698 void bfq_add_bfqq_busy(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1699 {
1700 	bfq_log_bfqq(bfqd, bfqq, "add to busy");
1701 
1702 	bfq_activate_bfqq(bfqd, bfqq);
1703 
1704 	bfq_mark_bfqq_busy(bfqq);
1705 	bfqd->busy_queues++;
1706 
1707 	if (!bfqq->dispatched)
1708 		if (bfqq->wr_coeff == 1)
1709 			bfq_weights_tree_add(bfqd, bfqq,
1710 					     &bfqd->queue_weights_tree);
1711 
1712 	if (bfqq->wr_coeff > 1)
1713 		bfqd->wr_busy_queues++;
1714 }
1715