xref: /linux/block/bfq-wf2q.c (revision 2d31c684)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Hierarchical Budget Worst-case Fair Weighted Fair Queueing
4  * (B-WF2Q+): hierarchical scheduling algorithm by which the BFQ I/O
5  * scheduler schedules generic entities. The latter can represent
6  * either single bfq queues (associated with processes) or groups of
7  * bfq queues (associated with cgroups).
8  */
9 #include "bfq-iosched.h"
10 
11 /**
12  * bfq_gt - compare two timestamps.
13  * @a: first ts.
14  * @b: second ts.
15  *
16  * Return @a > @b, dealing with wrapping correctly.
17  */
bfq_gt(u64 a,u64 b)18 static int bfq_gt(u64 a, u64 b)
19 {
20 	return (s64)(a - b) > 0;
21 }
22 
bfq_root_active_entity(struct rb_root * tree)23 static struct bfq_entity *bfq_root_active_entity(struct rb_root *tree)
24 {
25 	struct rb_node *node = tree->rb_node;
26 
27 	return rb_entry(node, struct bfq_entity, rb_node);
28 }
29 
bfq_class_idx(struct bfq_entity * entity)30 static unsigned int bfq_class_idx(struct bfq_entity *entity)
31 {
32 	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
33 
34 	return bfqq ? bfqq->ioprio_class - 1 :
35 		BFQ_DEFAULT_GRP_CLASS - 1;
36 }
37 
bfq_tot_busy_queues(struct bfq_data * bfqd)38 unsigned int bfq_tot_busy_queues(struct bfq_data *bfqd)
39 {
40 	return bfqd->busy_queues[0] + bfqd->busy_queues[1] +
41 		bfqd->busy_queues[2];
42 }
43 
44 static struct bfq_entity *bfq_lookup_next_entity(struct bfq_sched_data *sd,
45 						 bool expiration);
46 
47 static bool bfq_update_parent_budget(struct bfq_entity *next_in_service);
48 
49 /**
50  * bfq_update_next_in_service - update sd->next_in_service
51  * @sd: sched_data for which to perform the update.
52  * @new_entity: if not NULL, pointer to the entity whose activation,
53  *		requeueing or repositioning triggered the invocation of
54  *		this function.
55  * @expiration: id true, this function is being invoked after the
56  *             expiration of the in-service entity
57  *
58  * This function is called to update sd->next_in_service, which, in
59  * its turn, may change as a consequence of the insertion or
60  * extraction of an entity into/from one of the active trees of
61  * sd. These insertions/extractions occur as a consequence of
62  * activations/deactivations of entities, with some activations being
63  * 'true' activations, and other activations being requeueings (i.e.,
64  * implementing the second, requeueing phase of the mechanism used to
65  * reposition an entity in its active tree; see comments on
66  * __bfq_activate_entity and __bfq_requeue_entity for details). In
67  * both the last two activation sub-cases, new_entity points to the
68  * just activated or requeued entity.
69  *
70  * Returns true if sd->next_in_service changes in such a way that
71  * entity->parent may become the next_in_service for its parent
72  * entity.
73  */
bfq_update_next_in_service(struct bfq_sched_data * sd,struct bfq_entity * new_entity,bool expiration)74 static bool bfq_update_next_in_service(struct bfq_sched_data *sd,
75 				       struct bfq_entity *new_entity,
76 				       bool expiration)
77 {
78 	struct bfq_entity *next_in_service = sd->next_in_service;
79 	bool parent_sched_may_change = false;
80 	bool change_without_lookup = false;
81 
82 	/*
83 	 * If this update is triggered by the activation, requeueing
84 	 * or repositioning of an entity that does not coincide with
85 	 * sd->next_in_service, then a full lookup in the active tree
86 	 * can be avoided. In fact, it is enough to check whether the
87 	 * just-modified entity has the same priority as
88 	 * sd->next_in_service, is eligible and has a lower virtual
89 	 * finish time than sd->next_in_service. If this compound
90 	 * condition holds, then the new entity becomes the new
91 	 * next_in_service. Otherwise no change is needed.
92 	 */
93 	if (new_entity && new_entity != sd->next_in_service) {
94 		/*
95 		 * Flag used to decide whether to replace
96 		 * sd->next_in_service with new_entity. Tentatively
97 		 * set to true, and left as true if
98 		 * sd->next_in_service is NULL.
99 		 */
100 		change_without_lookup = true;
101 
102 		/*
103 		 * If there is already a next_in_service candidate
104 		 * entity, then compare timestamps to decide whether
105 		 * to replace sd->service_tree with new_entity.
106 		 */
107 		if (next_in_service) {
108 			unsigned int new_entity_class_idx =
109 				bfq_class_idx(new_entity);
110 			struct bfq_service_tree *st =
111 				sd->service_tree + new_entity_class_idx;
112 
113 			change_without_lookup =
114 				(new_entity_class_idx ==
115 				 bfq_class_idx(next_in_service)
116 				 &&
117 				 !bfq_gt(new_entity->start, st->vtime)
118 				 &&
119 				 bfq_gt(next_in_service->finish,
120 					new_entity->finish));
121 		}
122 
123 		if (change_without_lookup)
124 			next_in_service = new_entity;
125 	}
126 
127 	if (!change_without_lookup) /* lookup needed */
128 		next_in_service = bfq_lookup_next_entity(sd, expiration);
129 
130 	if (next_in_service) {
131 		bool new_budget_triggers_change =
132 			bfq_update_parent_budget(next_in_service);
133 
134 		parent_sched_may_change = !sd->next_in_service ||
135 			new_budget_triggers_change;
136 	}
137 
138 	sd->next_in_service = next_in_service;
139 
140 	return parent_sched_may_change;
141 }
142 
143 #ifdef CONFIG_BFQ_GROUP_IOSCHED
144 
145 /*
146  * Returns true if this budget changes may let next_in_service->parent
147  * become the next_in_service entity for its parent entity.
148  */
bfq_update_parent_budget(struct bfq_entity * next_in_service)149 static bool bfq_update_parent_budget(struct bfq_entity *next_in_service)
150 {
151 	struct bfq_entity *bfqg_entity;
152 	struct bfq_group *bfqg;
153 	struct bfq_sched_data *group_sd;
154 	bool ret = false;
155 
156 	group_sd = next_in_service->sched_data;
157 
158 	bfqg = container_of(group_sd, struct bfq_group, sched_data);
159 	/*
160 	 * bfq_group's my_entity field is not NULL only if the group
161 	 * is not the root group. We must not touch the root entity
162 	 * as it must never become an in-service entity.
163 	 */
164 	bfqg_entity = bfqg->my_entity;
165 	if (bfqg_entity) {
166 		if (bfqg_entity->budget > next_in_service->budget)
167 			ret = true;
168 		bfqg_entity->budget = next_in_service->budget;
169 	}
170 
171 	return ret;
172 }
173 
174 /*
175  * This function tells whether entity stops being a candidate for next
176  * service, according to the restrictive definition of the field
177  * next_in_service. In particular, this function is invoked for an
178  * entity that is about to be set in service.
179  *
180  * If entity is a queue, then the entity is no longer a candidate for
181  * next service according to the that definition, because entity is
182  * about to become the in-service queue. This function then returns
183  * true if entity is a queue.
184  *
185  * In contrast, entity could still be a candidate for next service if
186  * it is not a queue, and has more than one active child. In fact,
187  * even if one of its children is about to be set in service, other
188  * active children may still be the next to serve, for the parent
189  * entity, even according to the above definition. As a consequence, a
190  * non-queue entity is not a candidate for next-service only if it has
191  * only one active child. And only if this condition holds, then this
192  * function returns true for a non-queue entity.
193  */
bfq_no_longer_next_in_service(struct bfq_entity * entity)194 static bool bfq_no_longer_next_in_service(struct bfq_entity *entity)
195 {
196 	struct bfq_group *bfqg;
197 
198 	if (bfq_entity_to_bfqq(entity))
199 		return true;
200 
201 	bfqg = container_of(entity, struct bfq_group, entity);
202 
203 	/*
204 	 * The field active_entities does not always contain the
205 	 * actual number of active children entities: it happens to
206 	 * not account for the in-service entity in case the latter is
207 	 * removed from its active tree (which may get done after
208 	 * invoking the function bfq_no_longer_next_in_service in
209 	 * bfq_get_next_queue). Fortunately, here, i.e., while
210 	 * bfq_no_longer_next_in_service is not yet completed in
211 	 * bfq_get_next_queue, bfq_active_extract has not yet been
212 	 * invoked, and thus active_entities still coincides with the
213 	 * actual number of active entities.
214 	 */
215 	if (bfqg->active_entities == 1)
216 		return true;
217 
218 	return false;
219 }
220 
bfq_inc_active_entities(struct bfq_entity * entity)221 static void bfq_inc_active_entities(struct bfq_entity *entity)
222 {
223 	struct bfq_sched_data *sd = entity->sched_data;
224 	struct bfq_group *bfqg = container_of(sd, struct bfq_group, sched_data);
225 
226 	if (bfqg != bfqg->bfqd->root_group)
227 		bfqg->active_entities++;
228 }
229 
bfq_dec_active_entities(struct bfq_entity * entity)230 static void bfq_dec_active_entities(struct bfq_entity *entity)
231 {
232 	struct bfq_sched_data *sd = entity->sched_data;
233 	struct bfq_group *bfqg = container_of(sd, struct bfq_group, sched_data);
234 
235 	if (bfqg != bfqg->bfqd->root_group)
236 		bfqg->active_entities--;
237 }
238 
239 #else /* CONFIG_BFQ_GROUP_IOSCHED */
240 
bfq_update_parent_budget(struct bfq_entity * next_in_service)241 static bool bfq_update_parent_budget(struct bfq_entity *next_in_service)
242 {
243 	return false;
244 }
245 
bfq_no_longer_next_in_service(struct bfq_entity * entity)246 static bool bfq_no_longer_next_in_service(struct bfq_entity *entity)
247 {
248 	return true;
249 }
250 
bfq_inc_active_entities(struct bfq_entity * entity)251 static void bfq_inc_active_entities(struct bfq_entity *entity)
252 {
253 }
254 
bfq_dec_active_entities(struct bfq_entity * entity)255 static void bfq_dec_active_entities(struct bfq_entity *entity)
256 {
257 }
258 
259 #endif /* CONFIG_BFQ_GROUP_IOSCHED */
260 
261 /*
262  * Shift for timestamp calculations.  This actually limits the maximum
263  * service allowed in one timestamp delta (small shift values increase it),
264  * the maximum total weight that can be used for the queues in the system
265  * (big shift values increase it), and the period of virtual time
266  * wraparounds.
267  */
268 #define WFQ_SERVICE_SHIFT	22
269 
bfq_entity_to_bfqq(struct bfq_entity * entity)270 struct bfq_queue *bfq_entity_to_bfqq(struct bfq_entity *entity)
271 {
272 	struct bfq_queue *bfqq = NULL;
273 
274 	if (!entity->my_sched_data)
275 		bfqq = container_of(entity, struct bfq_queue, entity);
276 
277 	return bfqq;
278 }
279 
280 
281 /**
282  * bfq_delta - map service into the virtual time domain.
283  * @service: amount of service.
284  * @weight: scale factor (weight of an entity or weight sum).
285  */
bfq_delta(unsigned long service,unsigned long weight)286 static u64 bfq_delta(unsigned long service, unsigned long weight)
287 {
288 	return div64_ul((u64)service << WFQ_SERVICE_SHIFT, weight);
289 }
290 
291 /**
292  * bfq_calc_finish - assign the finish time to an entity.
293  * @entity: the entity to act upon.
294  * @service: the service to be charged to the entity.
295  */
bfq_calc_finish(struct bfq_entity * entity,unsigned long service)296 static void bfq_calc_finish(struct bfq_entity *entity, unsigned long service)
297 {
298 	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
299 
300 	entity->finish = entity->start +
301 		bfq_delta(service, entity->weight);
302 
303 	if (bfqq) {
304 		bfq_log_bfqq(bfqq->bfqd, bfqq,
305 			"calc_finish: serv %lu, w %d",
306 			service, entity->weight);
307 		bfq_log_bfqq(bfqq->bfqd, bfqq,
308 			"calc_finish: start %llu, finish %llu, delta %llu",
309 			entity->start, entity->finish,
310 			bfq_delta(service, entity->weight));
311 	}
312 }
313 
314 /**
315  * bfq_entity_of - get an entity from a node.
316  * @node: the node field of the entity.
317  *
318  * Convert a node pointer to the relative entity.  This is used only
319  * to simplify the logic of some functions and not as the generic
320  * conversion mechanism because, e.g., in the tree walking functions,
321  * the check for a %NULL value would be redundant.
322  */
bfq_entity_of(struct rb_node * node)323 struct bfq_entity *bfq_entity_of(struct rb_node *node)
324 {
325 	struct bfq_entity *entity = NULL;
326 
327 	if (node)
328 		entity = rb_entry(node, struct bfq_entity, rb_node);
329 
330 	return entity;
331 }
332 
333 /**
334  * bfq_extract - remove an entity from a tree.
335  * @root: the tree root.
336  * @entity: the entity to remove.
337  */
bfq_extract(struct rb_root * root,struct bfq_entity * entity)338 static void bfq_extract(struct rb_root *root, struct bfq_entity *entity)
339 {
340 	entity->tree = NULL;
341 	rb_erase(&entity->rb_node, root);
342 }
343 
344 /**
345  * bfq_idle_extract - extract an entity from the idle tree.
346  * @st: the service tree of the owning @entity.
347  * @entity: the entity being removed.
348  */
bfq_idle_extract(struct bfq_service_tree * st,struct bfq_entity * entity)349 static void bfq_idle_extract(struct bfq_service_tree *st,
350 			     struct bfq_entity *entity)
351 {
352 	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
353 	struct rb_node *next;
354 
355 	if (entity == st->first_idle) {
356 		next = rb_next(&entity->rb_node);
357 		st->first_idle = bfq_entity_of(next);
358 	}
359 
360 	if (entity == st->last_idle) {
361 		next = rb_prev(&entity->rb_node);
362 		st->last_idle = bfq_entity_of(next);
363 	}
364 
365 	bfq_extract(&st->idle, entity);
366 
367 	if (bfqq)
368 		list_del(&bfqq->bfqq_list);
369 }
370 
371 /**
372  * bfq_insert - generic tree insertion.
373  * @root: tree root.
374  * @entity: entity to insert.
375  *
376  * This is used for the idle and the active tree, since they are both
377  * ordered by finish time.
378  */
bfq_insert(struct rb_root * root,struct bfq_entity * entity)379 static void bfq_insert(struct rb_root *root, struct bfq_entity *entity)
380 {
381 	struct bfq_entity *entry;
382 	struct rb_node **node = &root->rb_node;
383 	struct rb_node *parent = NULL;
384 
385 	while (*node) {
386 		parent = *node;
387 		entry = rb_entry(parent, struct bfq_entity, rb_node);
388 
389 		if (bfq_gt(entry->finish, entity->finish))
390 			node = &parent->rb_left;
391 		else
392 			node = &parent->rb_right;
393 	}
394 
395 	rb_link_node(&entity->rb_node, parent, node);
396 	rb_insert_color(&entity->rb_node, root);
397 
398 	entity->tree = root;
399 }
400 
401 /**
402  * bfq_update_min - update the min_start field of a entity.
403  * @entity: the entity to update.
404  * @node: one of its children.
405  *
406  * This function is called when @entity may store an invalid value for
407  * min_start due to updates to the active tree.  The function  assumes
408  * that the subtree rooted at @node (which may be its left or its right
409  * child) has a valid min_start value.
410  */
bfq_update_min(struct bfq_entity * entity,struct rb_node * node)411 static void bfq_update_min(struct bfq_entity *entity, struct rb_node *node)
412 {
413 	struct bfq_entity *child;
414 
415 	if (node) {
416 		child = rb_entry(node, struct bfq_entity, rb_node);
417 		if (bfq_gt(entity->min_start, child->min_start))
418 			entity->min_start = child->min_start;
419 	}
420 }
421 
422 /**
423  * bfq_update_active_node - recalculate min_start.
424  * @node: the node to update.
425  *
426  * @node may have changed position or one of its children may have moved,
427  * this function updates its min_start value.  The left and right subtrees
428  * are assumed to hold a correct min_start value.
429  */
bfq_update_active_node(struct rb_node * node)430 static void bfq_update_active_node(struct rb_node *node)
431 {
432 	struct bfq_entity *entity = rb_entry(node, struct bfq_entity, rb_node);
433 
434 	entity->min_start = entity->start;
435 	bfq_update_min(entity, node->rb_right);
436 	bfq_update_min(entity, node->rb_left);
437 }
438 
439 /**
440  * bfq_update_active_tree - update min_start for the whole active tree.
441  * @node: the starting node.
442  *
443  * @node must be the deepest modified node after an update.  This function
444  * updates its min_start using the values held by its children, assuming
445  * that they did not change, and then updates all the nodes that may have
446  * changed in the path to the root.  The only nodes that may have changed
447  * are the ones in the path or their siblings.
448  */
bfq_update_active_tree(struct rb_node * node)449 static void bfq_update_active_tree(struct rb_node *node)
450 {
451 	struct rb_node *parent;
452 
453 up:
454 	bfq_update_active_node(node);
455 
456 	parent = rb_parent(node);
457 	if (!parent)
458 		return;
459 
460 	if (node == parent->rb_left && parent->rb_right)
461 		bfq_update_active_node(parent->rb_right);
462 	else if (parent->rb_left)
463 		bfq_update_active_node(parent->rb_left);
464 
465 	node = parent;
466 	goto up;
467 }
468 
469 /**
470  * bfq_active_insert - insert an entity in the active tree of its
471  *                     group/device.
472  * @st: the service tree of the entity.
473  * @entity: the entity being inserted.
474  *
475  * The active tree is ordered by finish time, but an extra key is kept
476  * per each node, containing the minimum value for the start times of
477  * its children (and the node itself), so it's possible to search for
478  * the eligible node with the lowest finish time in logarithmic time.
479  */
bfq_active_insert(struct bfq_service_tree * st,struct bfq_entity * entity)480 static void bfq_active_insert(struct bfq_service_tree *st,
481 			      struct bfq_entity *entity)
482 {
483 	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
484 	struct rb_node *node = &entity->rb_node;
485 
486 	bfq_insert(&st->active, entity);
487 
488 	if (node->rb_left)
489 		node = node->rb_left;
490 	else if (node->rb_right)
491 		node = node->rb_right;
492 
493 	bfq_update_active_tree(node);
494 
495 	if (bfqq)
496 		list_add(&bfqq->bfqq_list, &bfqq->bfqd->active_list[bfqq->actuator_idx]);
497 
498 	bfq_inc_active_entities(entity);
499 }
500 
501 /**
502  * bfq_ioprio_to_weight - calc a weight from an ioprio.
503  * @ioprio: the ioprio value to convert.
504  */
bfq_ioprio_to_weight(int ioprio)505 unsigned short bfq_ioprio_to_weight(int ioprio)
506 {
507 	return (IOPRIO_NR_LEVELS - ioprio) * BFQ_WEIGHT_CONVERSION_COEFF;
508 }
509 
510 /**
511  * bfq_weight_to_ioprio - calc an ioprio from a weight.
512  * @weight: the weight value to convert.
513  *
514  * To preserve as much as possible the old only-ioprio user interface,
515  * 0 is used as an escape ioprio value for weights (numerically) equal or
516  * larger than IOPRIO_NR_LEVELS * BFQ_WEIGHT_CONVERSION_COEFF.
517  */
bfq_weight_to_ioprio(int weight)518 static unsigned short bfq_weight_to_ioprio(int weight)
519 {
520 	return max_t(int, 0,
521 		     IOPRIO_NR_LEVELS - weight / BFQ_WEIGHT_CONVERSION_COEFF);
522 }
523 
bfq_get_entity(struct bfq_entity * entity)524 static void bfq_get_entity(struct bfq_entity *entity)
525 {
526 	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
527 
528 	if (bfqq) {
529 		bfqq->ref++;
530 		bfq_log_bfqq(bfqq->bfqd, bfqq, "get_entity: %p %d",
531 			     bfqq, bfqq->ref);
532 	}
533 }
534 
535 /**
536  * bfq_find_deepest - find the deepest node that an extraction can modify.
537  * @node: the node being removed.
538  *
539  * Do the first step of an extraction in an rb tree, looking for the
540  * node that will replace @node, and returning the deepest node that
541  * the following modifications to the tree can touch.  If @node is the
542  * last node in the tree return %NULL.
543  */
bfq_find_deepest(struct rb_node * node)544 static struct rb_node *bfq_find_deepest(struct rb_node *node)
545 {
546 	struct rb_node *deepest;
547 
548 	if (!node->rb_right && !node->rb_left)
549 		deepest = rb_parent(node);
550 	else if (!node->rb_right)
551 		deepest = node->rb_left;
552 	else if (!node->rb_left)
553 		deepest = node->rb_right;
554 	else {
555 		deepest = rb_next(node);
556 		if (deepest->rb_right)
557 			deepest = deepest->rb_right;
558 		else if (rb_parent(deepest) != node)
559 			deepest = rb_parent(deepest);
560 	}
561 
562 	return deepest;
563 }
564 
565 /**
566  * bfq_active_extract - remove an entity from the active tree.
567  * @st: the service_tree containing the tree.
568  * @entity: the entity being removed.
569  */
bfq_active_extract(struct bfq_service_tree * st,struct bfq_entity * entity)570 static void bfq_active_extract(struct bfq_service_tree *st,
571 			       struct bfq_entity *entity)
572 {
573 	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
574 	struct rb_node *node;
575 
576 	node = bfq_find_deepest(&entity->rb_node);
577 	bfq_extract(&st->active, entity);
578 
579 	if (node)
580 		bfq_update_active_tree(node);
581 	if (bfqq)
582 		list_del(&bfqq->bfqq_list);
583 
584 	bfq_dec_active_entities(entity);
585 }
586 
587 /**
588  * bfq_idle_insert - insert an entity into the idle tree.
589  * @st: the service tree containing the tree.
590  * @entity: the entity to insert.
591  */
bfq_idle_insert(struct bfq_service_tree * st,struct bfq_entity * entity)592 static void bfq_idle_insert(struct bfq_service_tree *st,
593 			    struct bfq_entity *entity)
594 {
595 	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
596 	struct bfq_entity *first_idle = st->first_idle;
597 	struct bfq_entity *last_idle = st->last_idle;
598 
599 	if (!first_idle || bfq_gt(first_idle->finish, entity->finish))
600 		st->first_idle = entity;
601 	if (!last_idle || bfq_gt(entity->finish, last_idle->finish))
602 		st->last_idle = entity;
603 
604 	bfq_insert(&st->idle, entity);
605 
606 	if (bfqq)
607 		list_add(&bfqq->bfqq_list, &bfqq->bfqd->idle_list);
608 }
609 
610 /**
611  * bfq_forget_entity - do not consider entity any longer for scheduling
612  * @st: the service tree.
613  * @entity: the entity being removed.
614  * @is_in_service: true if entity is currently the in-service entity.
615  *
616  * Forget everything about @entity. In addition, if entity represents
617  * a queue, and the latter is not in service, then release the service
618  * reference to the queue (the one taken through bfq_get_entity). In
619  * fact, in this case, there is really no more service reference to
620  * the queue, as the latter is also outside any service tree. If,
621  * instead, the queue is in service, then __bfq_bfqd_reset_in_service
622  * will take care of putting the reference when the queue finally
623  * stops being served.
624  */
bfq_forget_entity(struct bfq_service_tree * st,struct bfq_entity * entity,bool is_in_service)625 static void bfq_forget_entity(struct bfq_service_tree *st,
626 			      struct bfq_entity *entity,
627 			      bool is_in_service)
628 {
629 	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
630 
631 	entity->on_st_or_in_serv = false;
632 	st->wsum -= entity->weight;
633 	if (bfqq && !is_in_service)
634 		bfq_put_queue(bfqq);
635 }
636 
637 /**
638  * bfq_put_idle_entity - release the idle tree ref of an entity.
639  * @st: service tree for the entity.
640  * @entity: the entity being released.
641  */
bfq_put_idle_entity(struct bfq_service_tree * st,struct bfq_entity * entity)642 void bfq_put_idle_entity(struct bfq_service_tree *st, struct bfq_entity *entity)
643 {
644 	bfq_idle_extract(st, entity);
645 	bfq_forget_entity(st, entity,
646 			  entity == entity->sched_data->in_service_entity);
647 }
648 
649 /**
650  * bfq_forget_idle - update the idle tree if necessary.
651  * @st: the service tree to act upon.
652  *
653  * To preserve the global O(log N) complexity we only remove one entry here;
654  * as the idle tree will not grow indefinitely this can be done safely.
655  */
bfq_forget_idle(struct bfq_service_tree * st)656 static void bfq_forget_idle(struct bfq_service_tree *st)
657 {
658 	struct bfq_entity *first_idle = st->first_idle;
659 	struct bfq_entity *last_idle = st->last_idle;
660 
661 	if (RB_EMPTY_ROOT(&st->active) && last_idle &&
662 	    !bfq_gt(last_idle->finish, st->vtime)) {
663 		/*
664 		 * Forget the whole idle tree, increasing the vtime past
665 		 * the last finish time of idle entities.
666 		 */
667 		st->vtime = last_idle->finish;
668 	}
669 
670 	if (first_idle && !bfq_gt(first_idle->finish, st->vtime))
671 		bfq_put_idle_entity(st, first_idle);
672 }
673 
bfq_entity_service_tree(struct bfq_entity * entity)674 struct bfq_service_tree *bfq_entity_service_tree(struct bfq_entity *entity)
675 {
676 	struct bfq_sched_data *sched_data = entity->sched_data;
677 	unsigned int idx = bfq_class_idx(entity);
678 
679 	return sched_data->service_tree + idx;
680 }
681 
682 /*
683  * Update weight and priority of entity. If update_class_too is true,
684  * then update the ioprio_class of entity too.
685  *
686  * The reason why the update of ioprio_class is controlled through the
687  * last parameter is as follows. Changing the ioprio class of an
688  * entity implies changing the destination service trees for that
689  * entity. If such a change occurred when the entity is already on one
690  * of the service trees for its previous class, then the state of the
691  * entity would become more complex: none of the new possible service
692  * trees for the entity, according to bfq_entity_service_tree(), would
693  * match any of the possible service trees on which the entity
694  * is. Complex operations involving these trees, such as entity
695  * activations and deactivations, should take into account this
696  * additional complexity.  To avoid this issue, this function is
697  * invoked with update_class_too unset in the points in the code where
698  * entity may happen to be on some tree.
699  */
700 struct bfq_service_tree *
__bfq_entity_update_weight_prio(struct bfq_service_tree * old_st,struct bfq_entity * entity,bool update_class_too)701 __bfq_entity_update_weight_prio(struct bfq_service_tree *old_st,
702 				struct bfq_entity *entity,
703 				bool update_class_too)
704 {
705 	struct bfq_service_tree *new_st = old_st;
706 
707 	if (entity->prio_changed) {
708 		struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
709 		unsigned int prev_weight, new_weight;
710 
711 		/* Matches the smp_wmb() in bfq_group_set_weight. */
712 		smp_rmb();
713 		old_st->wsum -= entity->weight;
714 
715 		if (entity->new_weight != entity->orig_weight) {
716 			if (entity->new_weight < BFQ_MIN_WEIGHT ||
717 			    entity->new_weight > BFQ_MAX_WEIGHT) {
718 				pr_crit("update_weight_prio: new_weight %d\n",
719 					entity->new_weight);
720 				if (entity->new_weight < BFQ_MIN_WEIGHT)
721 					entity->new_weight = BFQ_MIN_WEIGHT;
722 				else
723 					entity->new_weight = BFQ_MAX_WEIGHT;
724 			}
725 			entity->orig_weight = entity->new_weight;
726 			if (bfqq)
727 				bfqq->ioprio =
728 				  bfq_weight_to_ioprio(entity->orig_weight);
729 		}
730 
731 		if (bfqq && update_class_too)
732 			bfqq->ioprio_class = bfqq->new_ioprio_class;
733 
734 		/*
735 		 * Reset prio_changed only if the ioprio_class change
736 		 * is not pending any longer.
737 		 */
738 		if (!bfqq || bfqq->ioprio_class == bfqq->new_ioprio_class)
739 			entity->prio_changed = 0;
740 
741 		/*
742 		 * NOTE: here we may be changing the weight too early,
743 		 * this will cause unfairness.  The correct approach
744 		 * would have required additional complexity to defer
745 		 * weight changes to the proper time instants (i.e.,
746 		 * when entity->finish <= old_st->vtime).
747 		 */
748 		new_st = bfq_entity_service_tree(entity);
749 
750 		prev_weight = entity->weight;
751 		new_weight = entity->orig_weight *
752 			     (bfqq ? bfqq->wr_coeff : 1);
753 		/*
754 		 * If the weight of the entity changes, and the entity is a
755 		 * queue, remove the entity from its old weight counter (if
756 		 * there is a counter associated with the entity).
757 		 */
758 		if (prev_weight != new_weight && bfqq)
759 			bfq_weights_tree_remove(bfqq);
760 		entity->weight = new_weight;
761 		/*
762 		 * Add the entity, if it is not a weight-raised queue,
763 		 * to the counter associated with its new weight.
764 		 */
765 		if (prev_weight != new_weight && bfqq && bfqq->wr_coeff == 1)
766 			bfq_weights_tree_add(bfqq);
767 
768 		new_st->wsum += entity->weight;
769 
770 		if (new_st != old_st)
771 			entity->start = new_st->vtime;
772 	}
773 
774 	return new_st;
775 }
776 
777 /**
778  * bfq_bfqq_served - update the scheduler status after selection for
779  *                   service.
780  * @bfqq: the queue being served.
781  * @served: bytes to transfer.
782  *
783  * NOTE: this can be optimized, as the timestamps of upper level entities
784  * are synchronized every time a new bfqq is selected for service.  By now,
785  * we keep it to better check consistency.
786  */
bfq_bfqq_served(struct bfq_queue * bfqq,int served)787 void bfq_bfqq_served(struct bfq_queue *bfqq, int served)
788 {
789 	struct bfq_entity *entity = &bfqq->entity;
790 	struct bfq_service_tree *st;
791 
792 	if (!bfqq->service_from_backlogged)
793 		bfqq->first_IO_time = jiffies;
794 
795 	if (bfqq->wr_coeff > 1)
796 		bfqq->service_from_wr += served;
797 
798 	bfqq->service_from_backlogged += served;
799 	for_each_entity(entity) {
800 		st = bfq_entity_service_tree(entity);
801 
802 		entity->service += served;
803 
804 		st->vtime += bfq_delta(served, st->wsum);
805 		bfq_forget_idle(st);
806 	}
807 	bfq_log_bfqq(bfqq->bfqd, bfqq, "bfqq_served %d secs", served);
808 }
809 
810 /**
811  * bfq_bfqq_charge_time - charge an amount of service equivalent to the length
812  *			  of the time interval during which bfqq has been in
813  *			  service.
814  * @bfqd: the device
815  * @bfqq: the queue that needs a service update.
816  * @time_ms: the amount of time during which the queue has received service
817  *
818  * If a queue does not consume its budget fast enough, then providing
819  * the queue with service fairness may impair throughput, more or less
820  * severely. For this reason, queues that consume their budget slowly
821  * are provided with time fairness instead of service fairness. This
822  * goal is achieved through the BFQ scheduling engine, even if such an
823  * engine works in the service, and not in the time domain. The trick
824  * is charging these queues with an inflated amount of service, equal
825  * to the amount of service that they would have received during their
826  * service slot if they had been fast, i.e., if their requests had
827  * been dispatched at a rate equal to the estimated peak rate.
828  *
829  * It is worth noting that time fairness can cause important
830  * distortions in terms of bandwidth distribution, on devices with
831  * internal queueing. The reason is that I/O requests dispatched
832  * during the service slot of a queue may be served after that service
833  * slot is finished, and may have a total processing time loosely
834  * correlated with the duration of the service slot. This is
835  * especially true for short service slots.
836  */
bfq_bfqq_charge_time(struct bfq_data * bfqd,struct bfq_queue * bfqq,unsigned long time_ms)837 void bfq_bfqq_charge_time(struct bfq_data *bfqd, struct bfq_queue *bfqq,
838 			  unsigned long time_ms)
839 {
840 	struct bfq_entity *entity = &bfqq->entity;
841 	unsigned long timeout_ms = jiffies_to_msecs(bfq_timeout);
842 	unsigned long bounded_time_ms = min(time_ms, timeout_ms);
843 	int serv_to_charge_for_time =
844 		(bfqd->bfq_max_budget * bounded_time_ms) / timeout_ms;
845 	int tot_serv_to_charge = max(serv_to_charge_for_time, entity->service);
846 
847 	/* Increase budget to avoid inconsistencies */
848 	if (tot_serv_to_charge > entity->budget)
849 		entity->budget = tot_serv_to_charge;
850 
851 	bfq_bfqq_served(bfqq,
852 			max_t(int, 0, tot_serv_to_charge - entity->service));
853 }
854 
bfq_update_fin_time_enqueue(struct bfq_entity * entity,struct bfq_service_tree * st,bool backshifted)855 static void bfq_update_fin_time_enqueue(struct bfq_entity *entity,
856 					struct bfq_service_tree *st,
857 					bool backshifted)
858 {
859 	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
860 
861 	/*
862 	 * When this function is invoked, entity is not in any service
863 	 * tree, then it is safe to invoke next function with the last
864 	 * parameter set (see the comments on the function).
865 	 */
866 	st = __bfq_entity_update_weight_prio(st, entity, true);
867 	bfq_calc_finish(entity, entity->budget);
868 
869 	/*
870 	 * If some queues enjoy backshifting for a while, then their
871 	 * (virtual) finish timestamps may happen to become lower and
872 	 * lower than the system virtual time.	In particular, if
873 	 * these queues often happen to be idle for short time
874 	 * periods, and during such time periods other queues with
875 	 * higher timestamps happen to be busy, then the backshifted
876 	 * timestamps of the former queues can become much lower than
877 	 * the system virtual time. In fact, to serve the queues with
878 	 * higher timestamps while the ones with lower timestamps are
879 	 * idle, the system virtual time may be pushed-up to much
880 	 * higher values than the finish timestamps of the idle
881 	 * queues. As a consequence, the finish timestamps of all new
882 	 * or newly activated queues may end up being much larger than
883 	 * those of lucky queues with backshifted timestamps. The
884 	 * latter queues may then monopolize the device for a lot of
885 	 * time. This would simply break service guarantees.
886 	 *
887 	 * To reduce this problem, push up a little bit the
888 	 * backshifted timestamps of the queue associated with this
889 	 * entity (only a queue can happen to have the backshifted
890 	 * flag set): just enough to let the finish timestamp of the
891 	 * queue be equal to the current value of the system virtual
892 	 * time. This may introduce a little unfairness among queues
893 	 * with backshifted timestamps, but it does not break
894 	 * worst-case fairness guarantees.
895 	 *
896 	 * As a special case, if bfqq is weight-raised, push up
897 	 * timestamps much less, to keep very low the probability that
898 	 * this push up causes the backshifted finish timestamps of
899 	 * weight-raised queues to become higher than the backshifted
900 	 * finish timestamps of non weight-raised queues.
901 	 */
902 	if (backshifted && bfq_gt(st->vtime, entity->finish)) {
903 		unsigned long delta = st->vtime - entity->finish;
904 
905 		if (bfqq)
906 			delta /= bfqq->wr_coeff;
907 
908 		entity->start += delta;
909 		entity->finish += delta;
910 	}
911 
912 	bfq_active_insert(st, entity);
913 }
914 
915 /**
916  * __bfq_activate_entity - handle activation of entity.
917  * @entity: the entity being activated.
918  * @non_blocking_wait_rq: true if entity was waiting for a request
919  *
920  * Called for a 'true' activation, i.e., if entity is not active and
921  * one of its children receives a new request.
922  *
923  * Basically, this function updates the timestamps of entity and
924  * inserts entity into its active tree, after possibly extracting it
925  * from its idle tree.
926  */
__bfq_activate_entity(struct bfq_entity * entity,bool non_blocking_wait_rq)927 static void __bfq_activate_entity(struct bfq_entity *entity,
928 				  bool non_blocking_wait_rq)
929 {
930 	struct bfq_service_tree *st = bfq_entity_service_tree(entity);
931 	bool backshifted = false;
932 	unsigned long long min_vstart;
933 
934 	/* See comments on bfq_fqq_update_budg_for_activation */
935 	if (non_blocking_wait_rq && bfq_gt(st->vtime, entity->finish)) {
936 		backshifted = true;
937 		min_vstart = entity->finish;
938 	} else
939 		min_vstart = st->vtime;
940 
941 	if (entity->tree == &st->idle) {
942 		/*
943 		 * Must be on the idle tree, bfq_idle_extract() will
944 		 * check for that.
945 		 */
946 		bfq_idle_extract(st, entity);
947 		entity->start = bfq_gt(min_vstart, entity->finish) ?
948 			min_vstart : entity->finish;
949 	} else {
950 		/*
951 		 * The finish time of the entity may be invalid, and
952 		 * it is in the past for sure, otherwise the queue
953 		 * would have been on the idle tree.
954 		 */
955 		entity->start = min_vstart;
956 		st->wsum += entity->weight;
957 		/*
958 		 * entity is about to be inserted into a service tree,
959 		 * and then set in service: get a reference to make
960 		 * sure entity does not disappear until it is no
961 		 * longer in service or scheduled for service.
962 		 */
963 		bfq_get_entity(entity);
964 
965 		entity->on_st_or_in_serv = true;
966 	}
967 
968 	bfq_update_fin_time_enqueue(entity, st, backshifted);
969 }
970 
971 /**
972  * __bfq_requeue_entity - handle requeueing or repositioning of an entity.
973  * @entity: the entity being requeued or repositioned.
974  *
975  * Requeueing is needed if this entity stops being served, which
976  * happens if a leaf descendant entity has expired. On the other hand,
977  * repositioning is needed if the next_inservice_entity for the child
978  * entity has changed. See the comments inside the function for
979  * details.
980  *
981  * Basically, this function: 1) removes entity from its active tree if
982  * present there, 2) updates the timestamps of entity and 3) inserts
983  * entity back into its active tree (in the new, right position for
984  * the new values of the timestamps).
985  */
__bfq_requeue_entity(struct bfq_entity * entity)986 static void __bfq_requeue_entity(struct bfq_entity *entity)
987 {
988 	struct bfq_sched_data *sd = entity->sched_data;
989 	struct bfq_service_tree *st = bfq_entity_service_tree(entity);
990 
991 	if (entity == sd->in_service_entity) {
992 		/*
993 		 * We are requeueing the current in-service entity,
994 		 * which may have to be done for one of the following
995 		 * reasons:
996 		 * - entity represents the in-service queue, and the
997 		 *   in-service queue is being requeued after an
998 		 *   expiration;
999 		 * - entity represents a group, and its budget has
1000 		 *   changed because one of its child entities has
1001 		 *   just been either activated or requeued for some
1002 		 *   reason; the timestamps of the entity need then to
1003 		 *   be updated, and the entity needs to be enqueued
1004 		 *   or repositioned accordingly.
1005 		 *
1006 		 * In particular, before requeueing, the start time of
1007 		 * the entity must be moved forward to account for the
1008 		 * service that the entity has received while in
1009 		 * service. This is done by the next instructions. The
1010 		 * finish time will then be updated according to this
1011 		 * new value of the start time, and to the budget of
1012 		 * the entity.
1013 		 */
1014 		bfq_calc_finish(entity, entity->service);
1015 		entity->start = entity->finish;
1016 		/*
1017 		 * In addition, if the entity had more than one child
1018 		 * when set in service, then it was not extracted from
1019 		 * the active tree. This implies that the position of
1020 		 * the entity in the active tree may need to be
1021 		 * changed now, because we have just updated the start
1022 		 * time of the entity, and we will update its finish
1023 		 * time in a moment (the requeueing is then, more
1024 		 * precisely, a repositioning in this case). To
1025 		 * implement this repositioning, we: 1) dequeue the
1026 		 * entity here, 2) update the finish time and requeue
1027 		 * the entity according to the new timestamps below.
1028 		 */
1029 		if (entity->tree)
1030 			bfq_active_extract(st, entity);
1031 	} else { /* The entity is already active, and not in service */
1032 		/*
1033 		 * In this case, this function gets called only if the
1034 		 * next_in_service entity below this entity has
1035 		 * changed, and this change has caused the budget of
1036 		 * this entity to change, which, finally implies that
1037 		 * the finish time of this entity must be
1038 		 * updated. Such an update may cause the scheduling,
1039 		 * i.e., the position in the active tree, of this
1040 		 * entity to change. We handle this change by: 1)
1041 		 * dequeueing the entity here, 2) updating the finish
1042 		 * time and requeueing the entity according to the new
1043 		 * timestamps below. This is the same approach as the
1044 		 * non-extracted-entity sub-case above.
1045 		 */
1046 		bfq_active_extract(st, entity);
1047 	}
1048 
1049 	bfq_update_fin_time_enqueue(entity, st, false);
1050 }
1051 
__bfq_activate_requeue_entity(struct bfq_entity * entity,bool non_blocking_wait_rq)1052 static void __bfq_activate_requeue_entity(struct bfq_entity *entity,
1053 					  bool non_blocking_wait_rq)
1054 {
1055 	struct bfq_service_tree *st = bfq_entity_service_tree(entity);
1056 
1057 	if (entity->sched_data->in_service_entity == entity ||
1058 	    entity->tree == &st->active)
1059 		 /*
1060 		  * in service or already queued on the active tree,
1061 		  * requeue or reposition
1062 		  */
1063 		__bfq_requeue_entity(entity);
1064 	else
1065 		/*
1066 		 * Not in service and not queued on its active tree:
1067 		 * the activity is idle and this is a true activation.
1068 		 */
1069 		__bfq_activate_entity(entity, non_blocking_wait_rq);
1070 }
1071 
1072 
1073 /**
1074  * bfq_activate_requeue_entity - activate or requeue an entity representing a
1075  *				 bfq_queue, and activate, requeue or reposition
1076  *				 all ancestors for which such an update becomes
1077  *				 necessary.
1078  * @entity: the entity to activate.
1079  * @non_blocking_wait_rq: true if this entity was waiting for a request
1080  * @requeue: true if this is a requeue, which implies that bfqq is
1081  *	     being expired; thus ALL its ancestors stop being served and must
1082  *	     therefore be requeued
1083  * @expiration: true if this function is being invoked in the expiration path
1084  *             of the in-service queue
1085  */
bfq_activate_requeue_entity(struct bfq_entity * entity,bool non_blocking_wait_rq,bool requeue,bool expiration)1086 static void bfq_activate_requeue_entity(struct bfq_entity *entity,
1087 					bool non_blocking_wait_rq,
1088 					bool requeue, bool expiration)
1089 {
1090 	for_each_entity(entity) {
1091 		__bfq_activate_requeue_entity(entity, non_blocking_wait_rq);
1092 		if (!bfq_update_next_in_service(entity->sched_data, entity,
1093 						expiration) && !requeue)
1094 			break;
1095 	}
1096 }
1097 
1098 /**
1099  * __bfq_deactivate_entity - update sched_data and service trees for
1100  * entity, so as to represent entity as inactive
1101  * @entity: the entity being deactivated.
1102  * @ins_into_idle_tree: if false, the entity will not be put into the
1103  *			idle tree.
1104  *
1105  * If necessary and allowed, puts entity into the idle tree. NOTE:
1106  * entity may be on no tree if in service.
1107  */
__bfq_deactivate_entity(struct bfq_entity * entity,bool ins_into_idle_tree)1108 bool __bfq_deactivate_entity(struct bfq_entity *entity, bool ins_into_idle_tree)
1109 {
1110 	struct bfq_sched_data *sd = entity->sched_data;
1111 	struct bfq_service_tree *st;
1112 	bool is_in_service;
1113 
1114 	if (!entity->on_st_or_in_serv) /*
1115 					* entity never activated, or
1116 					* already inactive
1117 					*/
1118 		return false;
1119 
1120 	/*
1121 	 * If we get here, then entity is active, which implies that
1122 	 * bfq_group_set_parent has already been invoked for the group
1123 	 * represented by entity. Therefore, the field
1124 	 * entity->sched_data has been set, and we can safely use it.
1125 	 */
1126 	st = bfq_entity_service_tree(entity);
1127 	is_in_service = entity == sd->in_service_entity;
1128 
1129 	bfq_calc_finish(entity, entity->service);
1130 
1131 	if (is_in_service)
1132 		sd->in_service_entity = NULL;
1133 	else
1134 		/*
1135 		 * Non in-service entity: nobody will take care of
1136 		 * resetting its service counter on expiration. Do it
1137 		 * now.
1138 		 */
1139 		entity->service = 0;
1140 
1141 	if (entity->tree == &st->active)
1142 		bfq_active_extract(st, entity);
1143 	else if (!is_in_service && entity->tree == &st->idle)
1144 		bfq_idle_extract(st, entity);
1145 
1146 	if (!ins_into_idle_tree || !bfq_gt(entity->finish, st->vtime))
1147 		bfq_forget_entity(st, entity, is_in_service);
1148 	else
1149 		bfq_idle_insert(st, entity);
1150 
1151 	return true;
1152 }
1153 
1154 /**
1155  * bfq_deactivate_entity - deactivate an entity representing a bfq_queue.
1156  * @entity: the entity to deactivate.
1157  * @ins_into_idle_tree: true if the entity can be put into the idle tree
1158  * @expiration: true if this function is being invoked in the expiration path
1159  *             of the in-service queue
1160  */
bfq_deactivate_entity(struct bfq_entity * entity,bool ins_into_idle_tree,bool expiration)1161 static void bfq_deactivate_entity(struct bfq_entity *entity,
1162 				  bool ins_into_idle_tree,
1163 				  bool expiration)
1164 {
1165 	struct bfq_sched_data *sd;
1166 	struct bfq_entity *parent = NULL;
1167 
1168 	for_each_entity_safe(entity, parent) {
1169 		sd = entity->sched_data;
1170 
1171 		if (!__bfq_deactivate_entity(entity, ins_into_idle_tree)) {
1172 			/*
1173 			 * entity is not in any tree any more, so
1174 			 * this deactivation is a no-op, and there is
1175 			 * nothing to change for upper-level entities
1176 			 * (in case of expiration, this can never
1177 			 * happen).
1178 			 */
1179 			return;
1180 		}
1181 
1182 		if (sd->next_in_service == entity)
1183 			/*
1184 			 * entity was the next_in_service entity,
1185 			 * then, since entity has just been
1186 			 * deactivated, a new one must be found.
1187 			 */
1188 			bfq_update_next_in_service(sd, NULL, expiration);
1189 
1190 		if (sd->next_in_service || sd->in_service_entity) {
1191 			/*
1192 			 * The parent entity is still active, because
1193 			 * either next_in_service or in_service_entity
1194 			 * is not NULL. So, no further upwards
1195 			 * deactivation must be performed.  Yet,
1196 			 * next_in_service has changed.	Then the
1197 			 * schedule does need to be updated upwards.
1198 			 *
1199 			 * NOTE If in_service_entity is not NULL, then
1200 			 * next_in_service may happen to be NULL,
1201 			 * although the parent entity is evidently
1202 			 * active. This happens if 1) the entity
1203 			 * pointed by in_service_entity is the only
1204 			 * active entity in the parent entity, and 2)
1205 			 * according to the definition of
1206 			 * next_in_service, the in_service_entity
1207 			 * cannot be considered as
1208 			 * next_in_service. See the comments on the
1209 			 * definition of next_in_service for details.
1210 			 */
1211 			break;
1212 		}
1213 
1214 		/*
1215 		 * If we get here, then the parent is no more
1216 		 * backlogged and we need to propagate the
1217 		 * deactivation upwards. Thus let the loop go on.
1218 		 */
1219 
1220 		/*
1221 		 * Also let parent be queued into the idle tree on
1222 		 * deactivation, to preserve service guarantees, and
1223 		 * assuming that who invoked this function does not
1224 		 * need parent entities too to be removed completely.
1225 		 */
1226 		ins_into_idle_tree = true;
1227 	}
1228 
1229 	/*
1230 	 * If the deactivation loop is fully executed, then there are
1231 	 * no more entities to touch and next loop is not executed at
1232 	 * all. Otherwise, requeue remaining entities if they are
1233 	 * about to stop receiving service, or reposition them if this
1234 	 * is not the case.
1235 	 */
1236 	entity = parent;
1237 	for_each_entity(entity) {
1238 		/*
1239 		 * Invoke __bfq_requeue_entity on entity, even if
1240 		 * already active, to requeue/reposition it in the
1241 		 * active tree (because sd->next_in_service has
1242 		 * changed)
1243 		 */
1244 		__bfq_requeue_entity(entity);
1245 
1246 		sd = entity->sched_data;
1247 		if (!bfq_update_next_in_service(sd, entity, expiration) &&
1248 		    !expiration)
1249 			/*
1250 			 * next_in_service unchanged or not causing
1251 			 * any change in entity->parent->sd, and no
1252 			 * requeueing needed for expiration: stop
1253 			 * here.
1254 			 */
1255 			break;
1256 	}
1257 }
1258 
1259 /**
1260  * bfq_calc_vtime_jump - compute the value to which the vtime should jump,
1261  *                       if needed, to have at least one entity eligible.
1262  * @st: the service tree to act upon.
1263  *
1264  * Assumes that st is not empty.
1265  */
bfq_calc_vtime_jump(struct bfq_service_tree * st)1266 static u64 bfq_calc_vtime_jump(struct bfq_service_tree *st)
1267 {
1268 	struct bfq_entity *root_entity = bfq_root_active_entity(&st->active);
1269 
1270 	if (bfq_gt(root_entity->min_start, st->vtime))
1271 		return root_entity->min_start;
1272 
1273 	return st->vtime;
1274 }
1275 
bfq_update_vtime(struct bfq_service_tree * st,u64 new_value)1276 static void bfq_update_vtime(struct bfq_service_tree *st, u64 new_value)
1277 {
1278 	if (new_value > st->vtime) {
1279 		st->vtime = new_value;
1280 		bfq_forget_idle(st);
1281 	}
1282 }
1283 
1284 /**
1285  * bfq_first_active_entity - find the eligible entity with
1286  *                           the smallest finish time
1287  * @st: the service tree to select from.
1288  * @vtime: the system virtual to use as a reference for eligibility
1289  *
1290  * This function searches the first schedulable entity, starting from the
1291  * root of the tree and going on the left every time on this side there is
1292  * a subtree with at least one eligible (start <= vtime) entity. The path on
1293  * the right is followed only if a) the left subtree contains no eligible
1294  * entities and b) no eligible entity has been found yet.
1295  */
bfq_first_active_entity(struct bfq_service_tree * st,u64 vtime)1296 static struct bfq_entity *bfq_first_active_entity(struct bfq_service_tree *st,
1297 						  u64 vtime)
1298 {
1299 	struct bfq_entity *entry, *first = NULL;
1300 	struct rb_node *node = st->active.rb_node;
1301 
1302 	while (node) {
1303 		entry = rb_entry(node, struct bfq_entity, rb_node);
1304 left:
1305 		if (!bfq_gt(entry->start, vtime))
1306 			first = entry;
1307 
1308 		if (node->rb_left) {
1309 			entry = rb_entry(node->rb_left,
1310 					 struct bfq_entity, rb_node);
1311 			if (!bfq_gt(entry->min_start, vtime)) {
1312 				node = node->rb_left;
1313 				goto left;
1314 			}
1315 		}
1316 		if (first)
1317 			break;
1318 		node = node->rb_right;
1319 	}
1320 
1321 	return first;
1322 }
1323 
1324 /**
1325  * __bfq_lookup_next_entity - return the first eligible entity in @st.
1326  * @st: the service tree.
1327  * @in_service: whether or not there is an in-service entity for the sched_data
1328  *	this active tree belongs to.
1329  *
1330  * If there is no in-service entity for the sched_data st belongs to,
1331  * then return the entity that will be set in service if:
1332  * 1) the parent entity this st belongs to is set in service;
1333  * 2) no entity belonging to such parent entity undergoes a state change
1334  * that would influence the timestamps of the entity (e.g., becomes idle,
1335  * becomes backlogged, changes its budget, ...).
1336  *
1337  * In this first case, update the virtual time in @st too (see the
1338  * comments on this update inside the function).
1339  *
1340  * In contrast, if there is an in-service entity, then return the
1341  * entity that would be set in service if not only the above
1342  * conditions, but also the next one held true: the currently
1343  * in-service entity, on expiration,
1344  * 1) gets a finish time equal to the current one, or
1345  * 2) is not eligible any more, or
1346  * 3) is idle.
1347  */
1348 static struct bfq_entity *
__bfq_lookup_next_entity(struct bfq_service_tree * st,bool in_service)1349 __bfq_lookup_next_entity(struct bfq_service_tree *st, bool in_service)
1350 {
1351 	struct bfq_entity *entity;
1352 	u64 new_vtime;
1353 
1354 	if (RB_EMPTY_ROOT(&st->active))
1355 		return NULL;
1356 
1357 	/*
1358 	 * Get the value of the system virtual time for which at
1359 	 * least one entity is eligible.
1360 	 */
1361 	new_vtime = bfq_calc_vtime_jump(st);
1362 
1363 	/*
1364 	 * If there is no in-service entity for the sched_data this
1365 	 * active tree belongs to, then push the system virtual time
1366 	 * up to the value that guarantees that at least one entity is
1367 	 * eligible. If, instead, there is an in-service entity, then
1368 	 * do not make any such update, because there is already an
1369 	 * eligible entity, namely the in-service one (even if the
1370 	 * entity is not on st, because it was extracted when set in
1371 	 * service).
1372 	 */
1373 	if (!in_service)
1374 		bfq_update_vtime(st, new_vtime);
1375 
1376 	entity = bfq_first_active_entity(st, new_vtime);
1377 
1378 	return entity;
1379 }
1380 
1381 /**
1382  * bfq_lookup_next_entity - return the first eligible entity in @sd.
1383  * @sd: the sched_data.
1384  * @expiration: true if we are on the expiration path of the in-service queue
1385  *
1386  * This function is invoked when there has been a change in the trees
1387  * for sd, and we need to know what is the new next entity to serve
1388  * after this change.
1389  */
bfq_lookup_next_entity(struct bfq_sched_data * sd,bool expiration)1390 static struct bfq_entity *bfq_lookup_next_entity(struct bfq_sched_data *sd,
1391 						 bool expiration)
1392 {
1393 	struct bfq_service_tree *st = sd->service_tree;
1394 	struct bfq_service_tree *idle_class_st = st + (BFQ_IOPRIO_CLASSES - 1);
1395 	struct bfq_entity *entity = NULL;
1396 	int class_idx = 0;
1397 
1398 	/*
1399 	 * Choose from idle class, if needed to guarantee a minimum
1400 	 * bandwidth to this class (and if there is some active entity
1401 	 * in idle class). This should also mitigate
1402 	 * priority-inversion problems in case a low priority task is
1403 	 * holding file system resources.
1404 	 */
1405 	if (time_is_before_jiffies(sd->bfq_class_idle_last_service +
1406 				   BFQ_CL_IDLE_TIMEOUT)) {
1407 		if (!RB_EMPTY_ROOT(&idle_class_st->active))
1408 			class_idx = BFQ_IOPRIO_CLASSES - 1;
1409 		/* About to be served if backlogged, or not yet backlogged */
1410 		sd->bfq_class_idle_last_service = jiffies;
1411 	}
1412 
1413 	/*
1414 	 * Find the next entity to serve for the highest-priority
1415 	 * class, unless the idle class needs to be served.
1416 	 */
1417 	for (; class_idx < BFQ_IOPRIO_CLASSES; class_idx++) {
1418 		/*
1419 		 * If expiration is true, then bfq_lookup_next_entity
1420 		 * is being invoked as a part of the expiration path
1421 		 * of the in-service queue. In this case, even if
1422 		 * sd->in_service_entity is not NULL,
1423 		 * sd->in_service_entity at this point is actually not
1424 		 * in service any more, and, if needed, has already
1425 		 * been properly queued or requeued into the right
1426 		 * tree. The reason why sd->in_service_entity is still
1427 		 * not NULL here, even if expiration is true, is that
1428 		 * sd->in_service_entity is reset as a last step in the
1429 		 * expiration path. So, if expiration is true, tell
1430 		 * __bfq_lookup_next_entity that there is no
1431 		 * sd->in_service_entity.
1432 		 */
1433 		entity = __bfq_lookup_next_entity(st + class_idx,
1434 						  sd->in_service_entity &&
1435 						  !expiration);
1436 
1437 		if (entity)
1438 			break;
1439 	}
1440 
1441 	return entity;
1442 }
1443 
next_queue_may_preempt(struct bfq_data * bfqd)1444 bool next_queue_may_preempt(struct bfq_data *bfqd)
1445 {
1446 	struct bfq_sched_data *sd = &bfqd->root_group->sched_data;
1447 
1448 	return sd->next_in_service != sd->in_service_entity;
1449 }
1450 
1451 /*
1452  * Get next queue for service.
1453  */
bfq_get_next_queue(struct bfq_data * bfqd)1454 struct bfq_queue *bfq_get_next_queue(struct bfq_data *bfqd)
1455 {
1456 	struct bfq_entity *entity = NULL;
1457 	struct bfq_sched_data *sd;
1458 	struct bfq_queue *bfqq;
1459 
1460 	if (bfq_tot_busy_queues(bfqd) == 0)
1461 		return NULL;
1462 
1463 	/*
1464 	 * Traverse the path from the root to the leaf entity to
1465 	 * serve. Set in service all the entities visited along the
1466 	 * way.
1467 	 */
1468 	sd = &bfqd->root_group->sched_data;
1469 	for (; sd ; sd = entity->my_sched_data) {
1470 		/*
1471 		 * WARNING. We are about to set the in-service entity
1472 		 * to sd->next_in_service, i.e., to the (cached) value
1473 		 * returned by bfq_lookup_next_entity(sd) the last
1474 		 * time it was invoked, i.e., the last time when the
1475 		 * service order in sd changed as a consequence of the
1476 		 * activation or deactivation of an entity. In this
1477 		 * respect, if we execute bfq_lookup_next_entity(sd)
1478 		 * in this very moment, it may, although with low
1479 		 * probability, yield a different entity than that
1480 		 * pointed to by sd->next_in_service. This rare event
1481 		 * happens in case there was no CLASS_IDLE entity to
1482 		 * serve for sd when bfq_lookup_next_entity(sd) was
1483 		 * invoked for the last time, while there is now one
1484 		 * such entity.
1485 		 *
1486 		 * If the above event happens, then the scheduling of
1487 		 * such entity in CLASS_IDLE is postponed until the
1488 		 * service of the sd->next_in_service entity
1489 		 * finishes. In fact, when the latter is expired,
1490 		 * bfq_lookup_next_entity(sd) gets called again,
1491 		 * exactly to update sd->next_in_service.
1492 		 */
1493 
1494 		/* Make next_in_service entity become in_service_entity */
1495 		entity = sd->next_in_service;
1496 		sd->in_service_entity = entity;
1497 
1498 		/*
1499 		 * If entity is no longer a candidate for next
1500 		 * service, then it must be extracted from its active
1501 		 * tree, so as to make sure that it won't be
1502 		 * considered when computing next_in_service. See the
1503 		 * comments on the function
1504 		 * bfq_no_longer_next_in_service() for details.
1505 		 */
1506 		if (bfq_no_longer_next_in_service(entity))
1507 			bfq_active_extract(bfq_entity_service_tree(entity),
1508 					   entity);
1509 
1510 		/*
1511 		 * Even if entity is not to be extracted according to
1512 		 * the above check, a descendant entity may get
1513 		 * extracted in one of the next iterations of this
1514 		 * loop. Such an event could cause a change in
1515 		 * next_in_service for the level of the descendant
1516 		 * entity, and thus possibly back to this level.
1517 		 *
1518 		 * However, we cannot perform the resulting needed
1519 		 * update of next_in_service for this level before the
1520 		 * end of the whole loop, because, to know which is
1521 		 * the correct next-to-serve candidate entity for each
1522 		 * level, we need first to find the leaf entity to set
1523 		 * in service. In fact, only after we know which is
1524 		 * the next-to-serve leaf entity, we can discover
1525 		 * whether the parent entity of the leaf entity
1526 		 * becomes the next-to-serve, and so on.
1527 		 */
1528 	}
1529 
1530 	bfqq = bfq_entity_to_bfqq(entity);
1531 
1532 	/*
1533 	 * We can finally update all next-to-serve entities along the
1534 	 * path from the leaf entity just set in service to the root.
1535 	 */
1536 	for_each_entity(entity) {
1537 		struct bfq_sched_data *sd = entity->sched_data;
1538 
1539 		if (!bfq_update_next_in_service(sd, NULL, false))
1540 			break;
1541 	}
1542 
1543 	return bfqq;
1544 }
1545 
1546 /* returns true if the in-service queue gets freed */
__bfq_bfqd_reset_in_service(struct bfq_data * bfqd)1547 bool __bfq_bfqd_reset_in_service(struct bfq_data *bfqd)
1548 {
1549 	struct bfq_queue *in_serv_bfqq = bfqd->in_service_queue;
1550 	struct bfq_entity *in_serv_entity = &in_serv_bfqq->entity;
1551 	struct bfq_entity *entity = in_serv_entity;
1552 
1553 	bfq_clear_bfqq_wait_request(in_serv_bfqq);
1554 	hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
1555 	bfqd->in_service_queue = NULL;
1556 
1557 	/*
1558 	 * When this function is called, all in-service entities have
1559 	 * been properly deactivated or requeued, so we can safely
1560 	 * execute the final step: reset in_service_entity along the
1561 	 * path from entity to the root.
1562 	 */
1563 	for_each_entity(entity)
1564 		entity->sched_data->in_service_entity = NULL;
1565 
1566 	/*
1567 	 * in_serv_entity is no longer in service, so, if it is in no
1568 	 * service tree either, then release the service reference to
1569 	 * the queue it represents (taken with bfq_get_entity).
1570 	 */
1571 	if (!in_serv_entity->on_st_or_in_serv) {
1572 		/*
1573 		 * If no process is referencing in_serv_bfqq any
1574 		 * longer, then the service reference may be the only
1575 		 * reference to the queue. If this is the case, then
1576 		 * bfqq gets freed here.
1577 		 */
1578 		int ref = in_serv_bfqq->ref;
1579 		bfq_put_queue(in_serv_bfqq);
1580 		if (ref == 1)
1581 			return true;
1582 	}
1583 
1584 	return false;
1585 }
1586 
bfq_deactivate_bfqq(struct bfq_data * bfqd,struct bfq_queue * bfqq,bool ins_into_idle_tree,bool expiration)1587 void bfq_deactivate_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1588 			 bool ins_into_idle_tree, bool expiration)
1589 {
1590 	struct bfq_entity *entity = &bfqq->entity;
1591 
1592 	bfq_deactivate_entity(entity, ins_into_idle_tree, expiration);
1593 }
1594 
bfq_activate_bfqq(struct bfq_data * bfqd,struct bfq_queue * bfqq)1595 void bfq_activate_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1596 {
1597 	struct bfq_entity *entity = &bfqq->entity;
1598 
1599 	bfq_activate_requeue_entity(entity, bfq_bfqq_non_blocking_wait_rq(bfqq),
1600 				    false, false);
1601 	bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1602 }
1603 
bfq_requeue_bfqq(struct bfq_data * bfqd,struct bfq_queue * bfqq,bool expiration)1604 void bfq_requeue_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1605 		      bool expiration)
1606 {
1607 	struct bfq_entity *entity = &bfqq->entity;
1608 
1609 	bfq_activate_requeue_entity(entity, false,
1610 				    bfqq == bfqd->in_service_queue, expiration);
1611 }
1612 
bfq_add_bfqq_in_groups_with_pending_reqs(struct bfq_queue * bfqq)1613 void bfq_add_bfqq_in_groups_with_pending_reqs(struct bfq_queue *bfqq)
1614 {
1615 #ifdef CONFIG_BFQ_GROUP_IOSCHED
1616 	struct bfq_entity *entity = &bfqq->entity;
1617 
1618 	if (!entity->in_groups_with_pending_reqs) {
1619 		entity->in_groups_with_pending_reqs = true;
1620 		if (!(bfqq_group(bfqq)->num_queues_with_pending_reqs++))
1621 			bfqq->bfqd->num_groups_with_pending_reqs++;
1622 	}
1623 #endif
1624 }
1625 
bfq_del_bfqq_in_groups_with_pending_reqs(struct bfq_queue * bfqq)1626 void bfq_del_bfqq_in_groups_with_pending_reqs(struct bfq_queue *bfqq)
1627 {
1628 #ifdef CONFIG_BFQ_GROUP_IOSCHED
1629 	struct bfq_entity *entity = &bfqq->entity;
1630 
1631 	if (entity->in_groups_with_pending_reqs) {
1632 		entity->in_groups_with_pending_reqs = false;
1633 		if (!(--bfqq_group(bfqq)->num_queues_with_pending_reqs))
1634 			bfqq->bfqd->num_groups_with_pending_reqs--;
1635 	}
1636 #endif
1637 }
1638 
1639 /*
1640  * Called when the bfqq no longer has requests pending, remove it from
1641  * the service tree. As a special case, it can be invoked during an
1642  * expiration.
1643  */
bfq_del_bfqq_busy(struct bfq_queue * bfqq,bool expiration)1644 void bfq_del_bfqq_busy(struct bfq_queue *bfqq, bool expiration)
1645 {
1646 	struct bfq_data *bfqd = bfqq->bfqd;
1647 
1648 	bfq_log_bfqq(bfqd, bfqq, "del from busy");
1649 
1650 	bfq_clear_bfqq_busy(bfqq);
1651 
1652 	bfqd->busy_queues[bfqq->ioprio_class - 1]--;
1653 
1654 	if (bfqq->wr_coeff > 1)
1655 		bfqd->wr_busy_queues--;
1656 
1657 	bfqg_stats_update_dequeue(bfqq_group(bfqq));
1658 
1659 	bfq_deactivate_bfqq(bfqd, bfqq, true, expiration);
1660 
1661 	if (!bfqq->dispatched) {
1662 		bfq_del_bfqq_in_groups_with_pending_reqs(bfqq);
1663 		/*
1664 		 * Next function is invoked last, because it causes bfqq to be
1665 		 * freed. DO NOT use bfqq after the next function invocation.
1666 		 */
1667 		bfq_weights_tree_remove(bfqq);
1668 	}
1669 }
1670 
1671 /*
1672  * Called when an inactive queue receives a new request.
1673  */
bfq_add_bfqq_busy(struct bfq_queue * bfqq)1674 void bfq_add_bfqq_busy(struct bfq_queue *bfqq)
1675 {
1676 	struct bfq_data *bfqd = bfqq->bfqd;
1677 
1678 	bfq_log_bfqq(bfqd, bfqq, "add to busy");
1679 
1680 	bfq_activate_bfqq(bfqd, bfqq);
1681 
1682 	bfq_mark_bfqq_busy(bfqq);
1683 	bfqd->busy_queues[bfqq->ioprio_class - 1]++;
1684 
1685 	if (!bfqq->dispatched) {
1686 		bfq_add_bfqq_in_groups_with_pending_reqs(bfqq);
1687 		if (bfqq->wr_coeff == 1)
1688 			bfq_weights_tree_add(bfqq);
1689 	}
1690 
1691 	if (bfqq->wr_coeff > 1)
1692 		bfqd->wr_busy_queues++;
1693 
1694 	/* Move bfqq to the head of the woken list of its waker */
1695 	if (!hlist_unhashed(&bfqq->woken_list_node) &&
1696 	    &bfqq->woken_list_node != bfqq->waker_bfqq->woken_list.first) {
1697 		hlist_del_init(&bfqq->woken_list_node);
1698 		hlist_add_head(&bfqq->woken_list_node,
1699 			       &bfqq->waker_bfqq->woken_list);
1700 	}
1701 }
1702