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