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 if (!next_in_service)
141 return parent_sched_may_change;
142
143 return parent_sched_may_change;
144 }
145
146 #ifdef CONFIG_BFQ_GROUP_IOSCHED
147
bfq_bfqq_to_bfqg(struct bfq_queue * bfqq)148 struct bfq_group *bfq_bfqq_to_bfqg(struct bfq_queue *bfqq)
149 {
150 struct bfq_entity *group_entity = bfqq->entity.parent;
151
152 if (!group_entity)
153 group_entity = &bfqq->bfqd->root_group->entity;
154
155 return container_of(group_entity, struct bfq_group, entity);
156 }
157
158 /*
159 * Returns true if this budget changes may let next_in_service->parent
160 * become the next_in_service entity for its parent entity.
161 */
bfq_update_parent_budget(struct bfq_entity * next_in_service)162 static bool bfq_update_parent_budget(struct bfq_entity *next_in_service)
163 {
164 struct bfq_entity *bfqg_entity;
165 struct bfq_group *bfqg;
166 struct bfq_sched_data *group_sd;
167 bool ret = false;
168
169 group_sd = next_in_service->sched_data;
170
171 bfqg = container_of(group_sd, struct bfq_group, sched_data);
172 /*
173 * bfq_group's my_entity field is not NULL only if the group
174 * is not the root group. We must not touch the root entity
175 * as it must never become an in-service entity.
176 */
177 bfqg_entity = bfqg->my_entity;
178 if (bfqg_entity) {
179 if (bfqg_entity->budget > next_in_service->budget)
180 ret = true;
181 bfqg_entity->budget = next_in_service->budget;
182 }
183
184 return ret;
185 }
186
187 /*
188 * This function tells whether entity stops being a candidate for next
189 * service, according to the restrictive definition of the field
190 * next_in_service. In particular, this function is invoked for an
191 * entity that is about to be set in service.
192 *
193 * If entity is a queue, then the entity is no longer a candidate for
194 * next service according to the that definition, because entity is
195 * about to become the in-service queue. This function then returns
196 * true if entity is a queue.
197 *
198 * In contrast, entity could still be a candidate for next service if
199 * it is not a queue, and has more than one active child. In fact,
200 * even if one of its children is about to be set in service, other
201 * active children may still be the next to serve, for the parent
202 * entity, even according to the above definition. As a consequence, a
203 * non-queue entity is not a candidate for next-service only if it has
204 * only one active child. And only if this condition holds, then this
205 * function returns true for a non-queue entity.
206 */
bfq_no_longer_next_in_service(struct bfq_entity * entity)207 static bool bfq_no_longer_next_in_service(struct bfq_entity *entity)
208 {
209 struct bfq_group *bfqg;
210
211 if (bfq_entity_to_bfqq(entity))
212 return true;
213
214 bfqg = container_of(entity, struct bfq_group, entity);
215
216 /*
217 * The field active_entities does not always contain the
218 * actual number of active children entities: it happens to
219 * not account for the in-service entity in case the latter is
220 * removed from its active tree (which may get done after
221 * invoking the function bfq_no_longer_next_in_service in
222 * bfq_get_next_queue). Fortunately, here, i.e., while
223 * bfq_no_longer_next_in_service is not yet completed in
224 * bfq_get_next_queue, bfq_active_extract has not yet been
225 * invoked, and thus active_entities still coincides with the
226 * actual number of active entities.
227 */
228 if (bfqg->active_entities == 1)
229 return true;
230
231 return false;
232 }
233
234 #else /* CONFIG_BFQ_GROUP_IOSCHED */
235
bfq_bfqq_to_bfqg(struct bfq_queue * bfqq)236 struct bfq_group *bfq_bfqq_to_bfqg(struct bfq_queue *bfqq)
237 {
238 return bfqq->bfqd->root_group;
239 }
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
251 #endif /* CONFIG_BFQ_GROUP_IOSCHED */
252
253 /*
254 * Shift for timestamp calculations. This actually limits the maximum
255 * service allowed in one timestamp delta (small shift values increase it),
256 * the maximum total weight that can be used for the queues in the system
257 * (big shift values increase it), and the period of virtual time
258 * wraparounds.
259 */
260 #define WFQ_SERVICE_SHIFT 22
261
bfq_entity_to_bfqq(struct bfq_entity * entity)262 struct bfq_queue *bfq_entity_to_bfqq(struct bfq_entity *entity)
263 {
264 struct bfq_queue *bfqq = NULL;
265
266 if (!entity->my_sched_data)
267 bfqq = container_of(entity, struct bfq_queue, entity);
268
269 return bfqq;
270 }
271
272
273 /**
274 * bfq_delta - map service into the virtual time domain.
275 * @service: amount of service.
276 * @weight: scale factor (weight of an entity or weight sum).
277 */
bfq_delta(unsigned long service,unsigned long weight)278 static u64 bfq_delta(unsigned long service, unsigned long weight)
279 {
280 return div64_ul((u64)service << WFQ_SERVICE_SHIFT, weight);
281 }
282
283 /**
284 * bfq_calc_finish - assign the finish time to an entity.
285 * @entity: the entity to act upon.
286 * @service: the service to be charged to the entity.
287 */
bfq_calc_finish(struct bfq_entity * entity,unsigned long service)288 static void bfq_calc_finish(struct bfq_entity *entity, unsigned long service)
289 {
290 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
291
292 entity->finish = entity->start +
293 bfq_delta(service, entity->weight);
294
295 if (bfqq) {
296 bfq_log_bfqq(bfqq->bfqd, bfqq,
297 "calc_finish: serv %lu, w %d",
298 service, entity->weight);
299 bfq_log_bfqq(bfqq->bfqd, bfqq,
300 "calc_finish: start %llu, finish %llu, delta %llu",
301 entity->start, entity->finish,
302 bfq_delta(service, entity->weight));
303 }
304 }
305
306 /**
307 * bfq_entity_of - get an entity from a node.
308 * @node: the node field of the entity.
309 *
310 * Convert a node pointer to the relative entity. This is used only
311 * to simplify the logic of some functions and not as the generic
312 * conversion mechanism because, e.g., in the tree walking functions,
313 * the check for a %NULL value would be redundant.
314 */
bfq_entity_of(struct rb_node * node)315 struct bfq_entity *bfq_entity_of(struct rb_node *node)
316 {
317 struct bfq_entity *entity = NULL;
318
319 if (node)
320 entity = rb_entry(node, struct bfq_entity, rb_node);
321
322 return entity;
323 }
324
325 /**
326 * bfq_extract - remove an entity from a tree.
327 * @root: the tree root.
328 * @entity: the entity to remove.
329 */
bfq_extract(struct rb_root * root,struct bfq_entity * entity)330 static void bfq_extract(struct rb_root *root, struct bfq_entity *entity)
331 {
332 entity->tree = NULL;
333 rb_erase(&entity->rb_node, root);
334 }
335
336 /**
337 * bfq_idle_extract - extract an entity from the idle tree.
338 * @st: the service tree of the owning @entity.
339 * @entity: the entity being removed.
340 */
bfq_idle_extract(struct bfq_service_tree * st,struct bfq_entity * entity)341 static void bfq_idle_extract(struct bfq_service_tree *st,
342 struct bfq_entity *entity)
343 {
344 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
345 struct rb_node *next;
346
347 if (entity == st->first_idle) {
348 next = rb_next(&entity->rb_node);
349 st->first_idle = bfq_entity_of(next);
350 }
351
352 if (entity == st->last_idle) {
353 next = rb_prev(&entity->rb_node);
354 st->last_idle = bfq_entity_of(next);
355 }
356
357 bfq_extract(&st->idle, entity);
358
359 if (bfqq)
360 list_del(&bfqq->bfqq_list);
361 }
362
363 /**
364 * bfq_insert - generic tree insertion.
365 * @root: tree root.
366 * @entity: entity to insert.
367 *
368 * This is used for the idle and the active tree, since they are both
369 * ordered by finish time.
370 */
bfq_insert(struct rb_root * root,struct bfq_entity * entity)371 static void bfq_insert(struct rb_root *root, struct bfq_entity *entity)
372 {
373 struct bfq_entity *entry;
374 struct rb_node **node = &root->rb_node;
375 struct rb_node *parent = NULL;
376
377 while (*node) {
378 parent = *node;
379 entry = rb_entry(parent, struct bfq_entity, rb_node);
380
381 if (bfq_gt(entry->finish, entity->finish))
382 node = &parent->rb_left;
383 else
384 node = &parent->rb_right;
385 }
386
387 rb_link_node(&entity->rb_node, parent, node);
388 rb_insert_color(&entity->rb_node, root);
389
390 entity->tree = root;
391 }
392
393 /**
394 * bfq_update_min - update the min_start field of a entity.
395 * @entity: the entity to update.
396 * @node: one of its children.
397 *
398 * This function is called when @entity may store an invalid value for
399 * min_start due to updates to the active tree. The function assumes
400 * that the subtree rooted at @node (which may be its left or its right
401 * child) has a valid min_start value.
402 */
bfq_update_min(struct bfq_entity * entity,struct rb_node * node)403 static void bfq_update_min(struct bfq_entity *entity, struct rb_node *node)
404 {
405 struct bfq_entity *child;
406
407 if (node) {
408 child = rb_entry(node, struct bfq_entity, rb_node);
409 if (bfq_gt(entity->min_start, child->min_start))
410 entity->min_start = child->min_start;
411 }
412 }
413
414 /**
415 * bfq_update_active_node - recalculate min_start.
416 * @node: the node to update.
417 *
418 * @node may have changed position or one of its children may have moved,
419 * this function updates its min_start value. The left and right subtrees
420 * are assumed to hold a correct min_start value.
421 */
bfq_update_active_node(struct rb_node * node)422 static void bfq_update_active_node(struct rb_node *node)
423 {
424 struct bfq_entity *entity = rb_entry(node, struct bfq_entity, rb_node);
425
426 entity->min_start = entity->start;
427 bfq_update_min(entity, node->rb_right);
428 bfq_update_min(entity, node->rb_left);
429 }
430
431 /**
432 * bfq_update_active_tree - update min_start for the whole active tree.
433 * @node: the starting node.
434 *
435 * @node must be the deepest modified node after an update. This function
436 * updates its min_start using the values held by its children, assuming
437 * that they did not change, and then updates all the nodes that may have
438 * changed in the path to the root. The only nodes that may have changed
439 * are the ones in the path or their siblings.
440 */
bfq_update_active_tree(struct rb_node * node)441 static void bfq_update_active_tree(struct rb_node *node)
442 {
443 struct rb_node *parent;
444
445 up:
446 bfq_update_active_node(node);
447
448 parent = rb_parent(node);
449 if (!parent)
450 return;
451
452 if (node == parent->rb_left && parent->rb_right)
453 bfq_update_active_node(parent->rb_right);
454 else if (parent->rb_left)
455 bfq_update_active_node(parent->rb_left);
456
457 node = parent;
458 goto up;
459 }
460
461 /**
462 * bfq_active_insert - insert an entity in the active tree of its
463 * group/device.
464 * @st: the service tree of the entity.
465 * @entity: the entity being inserted.
466 *
467 * The active tree is ordered by finish time, but an extra key is kept
468 * per each node, containing the minimum value for the start times of
469 * its children (and the node itself), so it's possible to search for
470 * the eligible node with the lowest finish time in logarithmic time.
471 */
bfq_active_insert(struct bfq_service_tree * st,struct bfq_entity * entity)472 static void bfq_active_insert(struct bfq_service_tree *st,
473 struct bfq_entity *entity)
474 {
475 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
476 struct rb_node *node = &entity->rb_node;
477 #ifdef CONFIG_BFQ_GROUP_IOSCHED
478 struct bfq_sched_data *sd = NULL;
479 struct bfq_group *bfqg = NULL;
480 struct bfq_data *bfqd = NULL;
481 #endif
482
483 bfq_insert(&st->active, entity);
484
485 if (node->rb_left)
486 node = node->rb_left;
487 else if (node->rb_right)
488 node = node->rb_right;
489
490 bfq_update_active_tree(node);
491
492 #ifdef CONFIG_BFQ_GROUP_IOSCHED
493 sd = entity->sched_data;
494 bfqg = container_of(sd, struct bfq_group, sched_data);
495 bfqd = (struct bfq_data *)bfqg->bfqd;
496 #endif
497 if (bfqq)
498 list_add(&bfqq->bfqq_list, &bfqq->bfqd->active_list);
499 #ifdef CONFIG_BFQ_GROUP_IOSCHED
500 if (bfqg != bfqd->root_group)
501 bfqg->active_entities++;
502 #endif
503 }
504
505 /**
506 * bfq_ioprio_to_weight - calc a weight from an ioprio.
507 * @ioprio: the ioprio value to convert.
508 */
bfq_ioprio_to_weight(int ioprio)509 unsigned short bfq_ioprio_to_weight(int ioprio)
510 {
511 return (IOPRIO_BE_NR - ioprio) * BFQ_WEIGHT_CONVERSION_COEFF;
512 }
513
514 /**
515 * bfq_weight_to_ioprio - calc an ioprio from a weight.
516 * @weight: the weight value to convert.
517 *
518 * To preserve as much as possible the old only-ioprio user interface,
519 * 0 is used as an escape ioprio value for weights (numerically) equal or
520 * larger than IOPRIO_BE_NR * BFQ_WEIGHT_CONVERSION_COEFF.
521 */
bfq_weight_to_ioprio(int weight)522 static unsigned short bfq_weight_to_ioprio(int weight)
523 {
524 return max_t(int, 0,
525 IOPRIO_BE_NR * BFQ_WEIGHT_CONVERSION_COEFF - weight);
526 }
527
bfq_get_entity(struct bfq_entity * entity)528 static void bfq_get_entity(struct bfq_entity *entity)
529 {
530 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
531
532 if (bfqq) {
533 bfqq->ref++;
534 bfq_log_bfqq(bfqq->bfqd, bfqq, "get_entity: %p %d",
535 bfqq, bfqq->ref);
536 }
537 }
538
539 /**
540 * bfq_find_deepest - find the deepest node that an extraction can modify.
541 * @node: the node being removed.
542 *
543 * Do the first step of an extraction in an rb tree, looking for the
544 * node that will replace @node, and returning the deepest node that
545 * the following modifications to the tree can touch. If @node is the
546 * last node in the tree return %NULL.
547 */
bfq_find_deepest(struct rb_node * node)548 static struct rb_node *bfq_find_deepest(struct rb_node *node)
549 {
550 struct rb_node *deepest;
551
552 if (!node->rb_right && !node->rb_left)
553 deepest = rb_parent(node);
554 else if (!node->rb_right)
555 deepest = node->rb_left;
556 else if (!node->rb_left)
557 deepest = node->rb_right;
558 else {
559 deepest = rb_next(node);
560 if (deepest->rb_right)
561 deepest = deepest->rb_right;
562 else if (rb_parent(deepest) != node)
563 deepest = rb_parent(deepest);
564 }
565
566 return deepest;
567 }
568
569 /**
570 * bfq_active_extract - remove an entity from the active tree.
571 * @st: the service_tree containing the tree.
572 * @entity: the entity being removed.
573 */
bfq_active_extract(struct bfq_service_tree * st,struct bfq_entity * entity)574 static void bfq_active_extract(struct bfq_service_tree *st,
575 struct bfq_entity *entity)
576 {
577 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
578 struct rb_node *node;
579 #ifdef CONFIG_BFQ_GROUP_IOSCHED
580 struct bfq_sched_data *sd = NULL;
581 struct bfq_group *bfqg = NULL;
582 struct bfq_data *bfqd = NULL;
583 #endif
584
585 node = bfq_find_deepest(&entity->rb_node);
586 bfq_extract(&st->active, entity);
587
588 if (node)
589 bfq_update_active_tree(node);
590
591 #ifdef CONFIG_BFQ_GROUP_IOSCHED
592 sd = entity->sched_data;
593 bfqg = container_of(sd, struct bfq_group, sched_data);
594 bfqd = (struct bfq_data *)bfqg->bfqd;
595 #endif
596 if (bfqq)
597 list_del(&bfqq->bfqq_list);
598 #ifdef CONFIG_BFQ_GROUP_IOSCHED
599 if (bfqg != bfqd->root_group)
600 bfqg->active_entities--;
601 #endif
602 }
603
604 /**
605 * bfq_idle_insert - insert an entity into the idle tree.
606 * @st: the service tree containing the tree.
607 * @entity: the entity to insert.
608 */
bfq_idle_insert(struct bfq_service_tree * st,struct bfq_entity * entity)609 static void bfq_idle_insert(struct bfq_service_tree *st,
610 struct bfq_entity *entity)
611 {
612 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
613 struct bfq_entity *first_idle = st->first_idle;
614 struct bfq_entity *last_idle = st->last_idle;
615
616 if (!first_idle || bfq_gt(first_idle->finish, entity->finish))
617 st->first_idle = entity;
618 if (!last_idle || bfq_gt(entity->finish, last_idle->finish))
619 st->last_idle = entity;
620
621 bfq_insert(&st->idle, entity);
622
623 if (bfqq)
624 list_add(&bfqq->bfqq_list, &bfqq->bfqd->idle_list);
625 }
626
627 /**
628 * bfq_forget_entity - do not consider entity any longer for scheduling
629 * @st: the service tree.
630 * @entity: the entity being removed.
631 * @is_in_service: true if entity is currently the in-service entity.
632 *
633 * Forget everything about @entity. In addition, if entity represents
634 * a queue, and the latter is not in service, then release the service
635 * reference to the queue (the one taken through bfq_get_entity). In
636 * fact, in this case, there is really no more service reference to
637 * the queue, as the latter is also outside any service tree. If,
638 * instead, the queue is in service, then __bfq_bfqd_reset_in_service
639 * will take care of putting the reference when the queue finally
640 * stops being served.
641 */
bfq_forget_entity(struct bfq_service_tree * st,struct bfq_entity * entity,bool is_in_service)642 static void bfq_forget_entity(struct bfq_service_tree *st,
643 struct bfq_entity *entity,
644 bool is_in_service)
645 {
646 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
647
648 entity->on_st_or_in_serv = false;
649 st->wsum -= entity->weight;
650 if (bfqq && !is_in_service)
651 bfq_put_queue(bfqq);
652 }
653
654 /**
655 * bfq_put_idle_entity - release the idle tree ref of an entity.
656 * @st: service tree for the entity.
657 * @entity: the entity being released.
658 */
bfq_put_idle_entity(struct bfq_service_tree * st,struct bfq_entity * entity)659 void bfq_put_idle_entity(struct bfq_service_tree *st, struct bfq_entity *entity)
660 {
661 bfq_idle_extract(st, entity);
662 bfq_forget_entity(st, entity,
663 entity == entity->sched_data->in_service_entity);
664 }
665
666 /**
667 * bfq_forget_idle - update the idle tree if necessary.
668 * @st: the service tree to act upon.
669 *
670 * To preserve the global O(log N) complexity we only remove one entry here;
671 * as the idle tree will not grow indefinitely this can be done safely.
672 */
bfq_forget_idle(struct bfq_service_tree * st)673 static void bfq_forget_idle(struct bfq_service_tree *st)
674 {
675 struct bfq_entity *first_idle = st->first_idle;
676 struct bfq_entity *last_idle = st->last_idle;
677
678 if (RB_EMPTY_ROOT(&st->active) && last_idle &&
679 !bfq_gt(last_idle->finish, st->vtime)) {
680 /*
681 * Forget the whole idle tree, increasing the vtime past
682 * the last finish time of idle entities.
683 */
684 st->vtime = last_idle->finish;
685 }
686
687 if (first_idle && !bfq_gt(first_idle->finish, st->vtime))
688 bfq_put_idle_entity(st, first_idle);
689 }
690
bfq_entity_service_tree(struct bfq_entity * entity)691 struct bfq_service_tree *bfq_entity_service_tree(struct bfq_entity *entity)
692 {
693 struct bfq_sched_data *sched_data = entity->sched_data;
694 unsigned int idx = bfq_class_idx(entity);
695
696 return sched_data->service_tree + idx;
697 }
698
699 /*
700 * Update weight and priority of entity. If update_class_too is true,
701 * then update the ioprio_class of entity too.
702 *
703 * The reason why the update of ioprio_class is controlled through the
704 * last parameter is as follows. Changing the ioprio class of an
705 * entity implies changing the destination service trees for that
706 * entity. If such a change occurred when the entity is already on one
707 * of the service trees for its previous class, then the state of the
708 * entity would become more complex: none of the new possible service
709 * trees for the entity, according to bfq_entity_service_tree(), would
710 * match any of the possible service trees on which the entity
711 * is. Complex operations involving these trees, such as entity
712 * activations and deactivations, should take into account this
713 * additional complexity. To avoid this issue, this function is
714 * invoked with update_class_too unset in the points in the code where
715 * entity may happen to be on some tree.
716 */
717 struct bfq_service_tree *
__bfq_entity_update_weight_prio(struct bfq_service_tree * old_st,struct bfq_entity * entity,bool update_class_too)718 __bfq_entity_update_weight_prio(struct bfq_service_tree *old_st,
719 struct bfq_entity *entity,
720 bool update_class_too)
721 {
722 struct bfq_service_tree *new_st = old_st;
723
724 if (entity->prio_changed) {
725 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
726 unsigned int prev_weight, new_weight;
727 struct bfq_data *bfqd = NULL;
728 struct rb_root_cached *root;
729 #ifdef CONFIG_BFQ_GROUP_IOSCHED
730 struct bfq_sched_data *sd;
731 struct bfq_group *bfqg;
732 #endif
733
734 if (bfqq)
735 bfqd = bfqq->bfqd;
736 #ifdef CONFIG_BFQ_GROUP_IOSCHED
737 else {
738 sd = entity->my_sched_data;
739 bfqg = container_of(sd, struct bfq_group, sched_data);
740 bfqd = (struct bfq_data *)bfqg->bfqd;
741 }
742 #endif
743
744 /* Matches the smp_wmb() in bfq_group_set_weight. */
745 smp_rmb();
746 old_st->wsum -= entity->weight;
747
748 if (entity->new_weight != entity->orig_weight) {
749 if (entity->new_weight < BFQ_MIN_WEIGHT ||
750 entity->new_weight > BFQ_MAX_WEIGHT) {
751 pr_crit("update_weight_prio: new_weight %d\n",
752 entity->new_weight);
753 if (entity->new_weight < BFQ_MIN_WEIGHT)
754 entity->new_weight = BFQ_MIN_WEIGHT;
755 else
756 entity->new_weight = BFQ_MAX_WEIGHT;
757 }
758 entity->orig_weight = entity->new_weight;
759 if (bfqq)
760 bfqq->ioprio =
761 bfq_weight_to_ioprio(entity->orig_weight);
762 }
763
764 if (bfqq && update_class_too)
765 bfqq->ioprio_class = bfqq->new_ioprio_class;
766
767 /*
768 * Reset prio_changed only if the ioprio_class change
769 * is not pending any longer.
770 */
771 if (!bfqq || bfqq->ioprio_class == bfqq->new_ioprio_class)
772 entity->prio_changed = 0;
773
774 /*
775 * NOTE: here we may be changing the weight too early,
776 * this will cause unfairness. The correct approach
777 * would have required additional complexity to defer
778 * weight changes to the proper time instants (i.e.,
779 * when entity->finish <= old_st->vtime).
780 */
781 new_st = bfq_entity_service_tree(entity);
782
783 prev_weight = entity->weight;
784 new_weight = entity->orig_weight *
785 (bfqq ? bfqq->wr_coeff : 1);
786 /*
787 * If the weight of the entity changes, and the entity is a
788 * queue, remove the entity from its old weight counter (if
789 * there is a counter associated with the entity).
790 */
791 if (prev_weight != new_weight && bfqq) {
792 root = &bfqd->queue_weights_tree;
793 __bfq_weights_tree_remove(bfqd, bfqq, root);
794 }
795 entity->weight = new_weight;
796 /*
797 * Add the entity, if it is not a weight-raised queue,
798 * to the counter associated with its new weight.
799 */
800 if (prev_weight != new_weight && bfqq && bfqq->wr_coeff == 1) {
801 /* If we get here, root has been initialized. */
802 bfq_weights_tree_add(bfqd, bfqq, root);
803 }
804
805 new_st->wsum += entity->weight;
806
807 if (new_st != old_st)
808 entity->start = new_st->vtime;
809 }
810
811 return new_st;
812 }
813
814 /**
815 * bfq_bfqq_served - update the scheduler status after selection for
816 * service.
817 * @bfqq: the queue being served.
818 * @served: bytes to transfer.
819 *
820 * NOTE: this can be optimized, as the timestamps of upper level entities
821 * are synchronized every time a new bfqq is selected for service. By now,
822 * we keep it to better check consistency.
823 */
bfq_bfqq_served(struct bfq_queue * bfqq,int served)824 void bfq_bfqq_served(struct bfq_queue *bfqq, int served)
825 {
826 struct bfq_entity *entity = &bfqq->entity;
827 struct bfq_service_tree *st;
828
829 if (!bfqq->service_from_backlogged)
830 bfqq->first_IO_time = jiffies;
831
832 if (bfqq->wr_coeff > 1)
833 bfqq->service_from_wr += served;
834
835 bfqq->service_from_backlogged += served;
836 for_each_entity(entity) {
837 st = bfq_entity_service_tree(entity);
838
839 entity->service += served;
840
841 st->vtime += bfq_delta(served, st->wsum);
842 bfq_forget_idle(st);
843 }
844 bfq_log_bfqq(bfqq->bfqd, bfqq, "bfqq_served %d secs", served);
845 }
846
847 /**
848 * bfq_bfqq_charge_time - charge an amount of service equivalent to the length
849 * of the time interval during which bfqq has been in
850 * service.
851 * @bfqd: the device
852 * @bfqq: the queue that needs a service update.
853 * @time_ms: the amount of time during which the queue has received service
854 *
855 * If a queue does not consume its budget fast enough, then providing
856 * the queue with service fairness may impair throughput, more or less
857 * severely. For this reason, queues that consume their budget slowly
858 * are provided with time fairness instead of service fairness. This
859 * goal is achieved through the BFQ scheduling engine, even if such an
860 * engine works in the service, and not in the time domain. The trick
861 * is charging these queues with an inflated amount of service, equal
862 * to the amount of service that they would have received during their
863 * service slot if they had been fast, i.e., if their requests had
864 * been dispatched at a rate equal to the estimated peak rate.
865 *
866 * It is worth noting that time fairness can cause important
867 * distortions in terms of bandwidth distribution, on devices with
868 * internal queueing. The reason is that I/O requests dispatched
869 * during the service slot of a queue may be served after that service
870 * slot is finished, and may have a total processing time loosely
871 * correlated with the duration of the service slot. This is
872 * especially true for short service slots.
873 */
bfq_bfqq_charge_time(struct bfq_data * bfqd,struct bfq_queue * bfqq,unsigned long time_ms)874 void bfq_bfqq_charge_time(struct bfq_data *bfqd, struct bfq_queue *bfqq,
875 unsigned long time_ms)
876 {
877 struct bfq_entity *entity = &bfqq->entity;
878 unsigned long timeout_ms = jiffies_to_msecs(bfq_timeout);
879 unsigned long bounded_time_ms = min(time_ms, timeout_ms);
880 int serv_to_charge_for_time =
881 (bfqd->bfq_max_budget * bounded_time_ms) / timeout_ms;
882 int tot_serv_to_charge = max(serv_to_charge_for_time, entity->service);
883
884 /* Increase budget to avoid inconsistencies */
885 if (tot_serv_to_charge > entity->budget)
886 entity->budget = tot_serv_to_charge;
887
888 bfq_bfqq_served(bfqq,
889 max_t(int, 0, tot_serv_to_charge - entity->service));
890 }
891
bfq_update_fin_time_enqueue(struct bfq_entity * entity,struct bfq_service_tree * st,bool backshifted)892 static void bfq_update_fin_time_enqueue(struct bfq_entity *entity,
893 struct bfq_service_tree *st,
894 bool backshifted)
895 {
896 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
897
898 /*
899 * When this function is invoked, entity is not in any service
900 * tree, then it is safe to invoke next function with the last
901 * parameter set (see the comments on the function).
902 */
903 st = __bfq_entity_update_weight_prio(st, entity, true);
904 bfq_calc_finish(entity, entity->budget);
905
906 /*
907 * If some queues enjoy backshifting for a while, then their
908 * (virtual) finish timestamps may happen to become lower and
909 * lower than the system virtual time. In particular, if
910 * these queues often happen to be idle for short time
911 * periods, and during such time periods other queues with
912 * higher timestamps happen to be busy, then the backshifted
913 * timestamps of the former queues can become much lower than
914 * the system virtual time. In fact, to serve the queues with
915 * higher timestamps while the ones with lower timestamps are
916 * idle, the system virtual time may be pushed-up to much
917 * higher values than the finish timestamps of the idle
918 * queues. As a consequence, the finish timestamps of all new
919 * or newly activated queues may end up being much larger than
920 * those of lucky queues with backshifted timestamps. The
921 * latter queues may then monopolize the device for a lot of
922 * time. This would simply break service guarantees.
923 *
924 * To reduce this problem, push up a little bit the
925 * backshifted timestamps of the queue associated with this
926 * entity (only a queue can happen to have the backshifted
927 * flag set): just enough to let the finish timestamp of the
928 * queue be equal to the current value of the system virtual
929 * time. This may introduce a little unfairness among queues
930 * with backshifted timestamps, but it does not break
931 * worst-case fairness guarantees.
932 *
933 * As a special case, if bfqq is weight-raised, push up
934 * timestamps much less, to keep very low the probability that
935 * this push up causes the backshifted finish timestamps of
936 * weight-raised queues to become higher than the backshifted
937 * finish timestamps of non weight-raised queues.
938 */
939 if (backshifted && bfq_gt(st->vtime, entity->finish)) {
940 unsigned long delta = st->vtime - entity->finish;
941
942 if (bfqq)
943 delta /= bfqq->wr_coeff;
944
945 entity->start += delta;
946 entity->finish += delta;
947 }
948
949 bfq_active_insert(st, entity);
950 }
951
952 /**
953 * __bfq_activate_entity - handle activation of entity.
954 * @entity: the entity being activated.
955 * @non_blocking_wait_rq: true if entity was waiting for a request
956 *
957 * Called for a 'true' activation, i.e., if entity is not active and
958 * one of its children receives a new request.
959 *
960 * Basically, this function updates the timestamps of entity and
961 * inserts entity into its active tree, after possibly extracting it
962 * from its idle tree.
963 */
__bfq_activate_entity(struct bfq_entity * entity,bool non_blocking_wait_rq)964 static void __bfq_activate_entity(struct bfq_entity *entity,
965 bool non_blocking_wait_rq)
966 {
967 struct bfq_service_tree *st = bfq_entity_service_tree(entity);
968 bool backshifted = false;
969 unsigned long long min_vstart;
970
971 /* See comments on bfq_fqq_update_budg_for_activation */
972 if (non_blocking_wait_rq && bfq_gt(st->vtime, entity->finish)) {
973 backshifted = true;
974 min_vstart = entity->finish;
975 } else
976 min_vstart = st->vtime;
977
978 if (entity->tree == &st->idle) {
979 /*
980 * Must be on the idle tree, bfq_idle_extract() will
981 * check for that.
982 */
983 bfq_idle_extract(st, entity);
984 entity->start = bfq_gt(min_vstart, entity->finish) ?
985 min_vstart : entity->finish;
986 } else {
987 /*
988 * The finish time of the entity may be invalid, and
989 * it is in the past for sure, otherwise the queue
990 * would have been on the idle tree.
991 */
992 entity->start = min_vstart;
993 st->wsum += entity->weight;
994 /*
995 * entity is about to be inserted into a service tree,
996 * and then set in service: get a reference to make
997 * sure entity does not disappear until it is no
998 * longer in service or scheduled for service.
999 */
1000 bfq_get_entity(entity);
1001
1002 entity->on_st_or_in_serv = true;
1003 }
1004
1005 #ifdef CONFIG_BFQ_GROUP_IOSCHED
1006 if (!bfq_entity_to_bfqq(entity)) { /* bfq_group */
1007 struct bfq_group *bfqg =
1008 container_of(entity, struct bfq_group, entity);
1009 struct bfq_data *bfqd = bfqg->bfqd;
1010
1011 if (!entity->in_groups_with_pending_reqs) {
1012 entity->in_groups_with_pending_reqs = true;
1013 bfqd->num_groups_with_pending_reqs++;
1014 }
1015 }
1016 #endif
1017
1018 bfq_update_fin_time_enqueue(entity, st, backshifted);
1019 }
1020
1021 /**
1022 * __bfq_requeue_entity - handle requeueing or repositioning of an entity.
1023 * @entity: the entity being requeued or repositioned.
1024 *
1025 * Requeueing is needed if this entity stops being served, which
1026 * happens if a leaf descendant entity has expired. On the other hand,
1027 * repositioning is needed if the next_inservice_entity for the child
1028 * entity has changed. See the comments inside the function for
1029 * details.
1030 *
1031 * Basically, this function: 1) removes entity from its active tree if
1032 * present there, 2) updates the timestamps of entity and 3) inserts
1033 * entity back into its active tree (in the new, right position for
1034 * the new values of the timestamps).
1035 */
__bfq_requeue_entity(struct bfq_entity * entity)1036 static void __bfq_requeue_entity(struct bfq_entity *entity)
1037 {
1038 struct bfq_sched_data *sd = entity->sched_data;
1039 struct bfq_service_tree *st = bfq_entity_service_tree(entity);
1040
1041 if (entity == sd->in_service_entity) {
1042 /*
1043 * We are requeueing the current in-service entity,
1044 * which may have to be done for one of the following
1045 * reasons:
1046 * - entity represents the in-service queue, and the
1047 * in-service queue is being requeued after an
1048 * expiration;
1049 * - entity represents a group, and its budget has
1050 * changed because one of its child entities has
1051 * just been either activated or requeued for some
1052 * reason; the timestamps of the entity need then to
1053 * be updated, and the entity needs to be enqueued
1054 * or repositioned accordingly.
1055 *
1056 * In particular, before requeueing, the start time of
1057 * the entity must be moved forward to account for the
1058 * service that the entity has received while in
1059 * service. This is done by the next instructions. The
1060 * finish time will then be updated according to this
1061 * new value of the start time, and to the budget of
1062 * the entity.
1063 */
1064 bfq_calc_finish(entity, entity->service);
1065 entity->start = entity->finish;
1066 /*
1067 * In addition, if the entity had more than one child
1068 * when set in service, then it was not extracted from
1069 * the active tree. This implies that the position of
1070 * the entity in the active tree may need to be
1071 * changed now, because we have just updated the start
1072 * time of the entity, and we will update its finish
1073 * time in a moment (the requeueing is then, more
1074 * precisely, a repositioning in this case). To
1075 * implement this repositioning, we: 1) dequeue the
1076 * entity here, 2) update the finish time and requeue
1077 * the entity according to the new timestamps below.
1078 */
1079 if (entity->tree)
1080 bfq_active_extract(st, entity);
1081 } else { /* The entity is already active, and not in service */
1082 /*
1083 * In this case, this function gets called only if the
1084 * next_in_service entity below this entity has
1085 * changed, and this change has caused the budget of
1086 * this entity to change, which, finally implies that
1087 * the finish time of this entity must be
1088 * updated. Such an update may cause the scheduling,
1089 * i.e., the position in the active tree, of this
1090 * entity to change. We handle this change by: 1)
1091 * dequeueing the entity here, 2) updating the finish
1092 * time and requeueing the entity according to the new
1093 * timestamps below. This is the same approach as the
1094 * non-extracted-entity sub-case above.
1095 */
1096 bfq_active_extract(st, entity);
1097 }
1098
1099 bfq_update_fin_time_enqueue(entity, st, false);
1100 }
1101
__bfq_activate_requeue_entity(struct bfq_entity * entity,struct bfq_sched_data * sd,bool non_blocking_wait_rq)1102 static void __bfq_activate_requeue_entity(struct bfq_entity *entity,
1103 struct bfq_sched_data *sd,
1104 bool non_blocking_wait_rq)
1105 {
1106 struct bfq_service_tree *st = bfq_entity_service_tree(entity);
1107
1108 if (sd->in_service_entity == entity || entity->tree == &st->active)
1109 /*
1110 * in service or already queued on the active tree,
1111 * requeue or reposition
1112 */
1113 __bfq_requeue_entity(entity);
1114 else
1115 /*
1116 * Not in service and not queued on its active tree:
1117 * the activity is idle and this is a true activation.
1118 */
1119 __bfq_activate_entity(entity, non_blocking_wait_rq);
1120 }
1121
1122
1123 /**
1124 * bfq_activate_requeue_entity - activate or requeue an entity representing a
1125 * bfq_queue, and activate, requeue or reposition
1126 * all ancestors for which such an update becomes
1127 * necessary.
1128 * @entity: the entity to activate.
1129 * @non_blocking_wait_rq: true if this entity was waiting for a request
1130 * @requeue: true if this is a requeue, which implies that bfqq is
1131 * being expired; thus ALL its ancestors stop being served and must
1132 * therefore be requeued
1133 * @expiration: true if this function is being invoked in the expiration path
1134 * of the in-service queue
1135 */
bfq_activate_requeue_entity(struct bfq_entity * entity,bool non_blocking_wait_rq,bool requeue,bool expiration)1136 static void bfq_activate_requeue_entity(struct bfq_entity *entity,
1137 bool non_blocking_wait_rq,
1138 bool requeue, bool expiration)
1139 {
1140 struct bfq_sched_data *sd;
1141
1142 for_each_entity(entity) {
1143 sd = entity->sched_data;
1144 __bfq_activate_requeue_entity(entity, sd, non_blocking_wait_rq);
1145
1146 if (!bfq_update_next_in_service(sd, entity, expiration) &&
1147 !requeue)
1148 break;
1149 }
1150 }
1151
1152 /**
1153 * __bfq_deactivate_entity - update sched_data and service trees for
1154 * entity, so as to represent entity as inactive
1155 * @entity: the entity being deactivated.
1156 * @ins_into_idle_tree: if false, the entity will not be put into the
1157 * idle tree.
1158 *
1159 * If necessary and allowed, puts entity into the idle tree. NOTE:
1160 * entity may be on no tree if in service.
1161 */
__bfq_deactivate_entity(struct bfq_entity * entity,bool ins_into_idle_tree)1162 bool __bfq_deactivate_entity(struct bfq_entity *entity, bool ins_into_idle_tree)
1163 {
1164 struct bfq_sched_data *sd = entity->sched_data;
1165 struct bfq_service_tree *st;
1166 bool is_in_service;
1167
1168 if (!entity->on_st_or_in_serv) /*
1169 * entity never activated, or
1170 * already inactive
1171 */
1172 return false;
1173
1174 /*
1175 * If we get here, then entity is active, which implies that
1176 * bfq_group_set_parent has already been invoked for the group
1177 * represented by entity. Therefore, the field
1178 * entity->sched_data has been set, and we can safely use it.
1179 */
1180 st = bfq_entity_service_tree(entity);
1181 is_in_service = entity == sd->in_service_entity;
1182
1183 bfq_calc_finish(entity, entity->service);
1184
1185 if (is_in_service)
1186 sd->in_service_entity = NULL;
1187 else
1188 /*
1189 * Non in-service entity: nobody will take care of
1190 * resetting its service counter on expiration. Do it
1191 * now.
1192 */
1193 entity->service = 0;
1194
1195 if (entity->tree == &st->active)
1196 bfq_active_extract(st, entity);
1197 else if (!is_in_service && entity->tree == &st->idle)
1198 bfq_idle_extract(st, entity);
1199
1200 if (!ins_into_idle_tree || !bfq_gt(entity->finish, st->vtime))
1201 bfq_forget_entity(st, entity, is_in_service);
1202 else
1203 bfq_idle_insert(st, entity);
1204
1205 return true;
1206 }
1207
1208 /**
1209 * bfq_deactivate_entity - deactivate an entity representing a bfq_queue.
1210 * @entity: the entity to deactivate.
1211 * @ins_into_idle_tree: true if the entity can be put into the idle tree
1212 * @expiration: true if this function is being invoked in the expiration path
1213 * of the in-service queue
1214 */
bfq_deactivate_entity(struct bfq_entity * entity,bool ins_into_idle_tree,bool expiration)1215 static void bfq_deactivate_entity(struct bfq_entity *entity,
1216 bool ins_into_idle_tree,
1217 bool expiration)
1218 {
1219 struct bfq_sched_data *sd;
1220 struct bfq_entity *parent = NULL;
1221
1222 for_each_entity_safe(entity, parent) {
1223 sd = entity->sched_data;
1224
1225 if (!__bfq_deactivate_entity(entity, ins_into_idle_tree)) {
1226 /*
1227 * entity is not in any tree any more, so
1228 * this deactivation is a no-op, and there is
1229 * nothing to change for upper-level entities
1230 * (in case of expiration, this can never
1231 * happen).
1232 */
1233 return;
1234 }
1235
1236 if (sd->next_in_service == entity)
1237 /*
1238 * entity was the next_in_service entity,
1239 * then, since entity has just been
1240 * deactivated, a new one must be found.
1241 */
1242 bfq_update_next_in_service(sd, NULL, expiration);
1243
1244 if (sd->next_in_service || sd->in_service_entity) {
1245 /*
1246 * The parent entity is still active, because
1247 * either next_in_service or in_service_entity
1248 * is not NULL. So, no further upwards
1249 * deactivation must be performed. Yet,
1250 * next_in_service has changed. Then the
1251 * schedule does need to be updated upwards.
1252 *
1253 * NOTE If in_service_entity is not NULL, then
1254 * next_in_service may happen to be NULL,
1255 * although the parent entity is evidently
1256 * active. This happens if 1) the entity
1257 * pointed by in_service_entity is the only
1258 * active entity in the parent entity, and 2)
1259 * according to the definition of
1260 * next_in_service, the in_service_entity
1261 * cannot be considered as
1262 * next_in_service. See the comments on the
1263 * definition of next_in_service for details.
1264 */
1265 break;
1266 }
1267
1268 /*
1269 * If we get here, then the parent is no more
1270 * backlogged and we need to propagate the
1271 * deactivation upwards. Thus let the loop go on.
1272 */
1273
1274 /*
1275 * Also let parent be queued into the idle tree on
1276 * deactivation, to preserve service guarantees, and
1277 * assuming that who invoked this function does not
1278 * need parent entities too to be removed completely.
1279 */
1280 ins_into_idle_tree = true;
1281 }
1282
1283 /*
1284 * If the deactivation loop is fully executed, then there are
1285 * no more entities to touch and next loop is not executed at
1286 * all. Otherwise, requeue remaining entities if they are
1287 * about to stop receiving service, or reposition them if this
1288 * is not the case.
1289 */
1290 entity = parent;
1291 for_each_entity(entity) {
1292 /*
1293 * Invoke __bfq_requeue_entity on entity, even if
1294 * already active, to requeue/reposition it in the
1295 * active tree (because sd->next_in_service has
1296 * changed)
1297 */
1298 __bfq_requeue_entity(entity);
1299
1300 sd = entity->sched_data;
1301 if (!bfq_update_next_in_service(sd, entity, expiration) &&
1302 !expiration)
1303 /*
1304 * next_in_service unchanged or not causing
1305 * any change in entity->parent->sd, and no
1306 * requeueing needed for expiration: stop
1307 * here.
1308 */
1309 break;
1310 }
1311 }
1312
1313 /**
1314 * bfq_calc_vtime_jump - compute the value to which the vtime should jump,
1315 * if needed, to have at least one entity eligible.
1316 * @st: the service tree to act upon.
1317 *
1318 * Assumes that st is not empty.
1319 */
bfq_calc_vtime_jump(struct bfq_service_tree * st)1320 static u64 bfq_calc_vtime_jump(struct bfq_service_tree *st)
1321 {
1322 struct bfq_entity *root_entity = bfq_root_active_entity(&st->active);
1323
1324 if (bfq_gt(root_entity->min_start, st->vtime))
1325 return root_entity->min_start;
1326
1327 return st->vtime;
1328 }
1329
bfq_update_vtime(struct bfq_service_tree * st,u64 new_value)1330 static void bfq_update_vtime(struct bfq_service_tree *st, u64 new_value)
1331 {
1332 if (new_value > st->vtime) {
1333 st->vtime = new_value;
1334 bfq_forget_idle(st);
1335 }
1336 }
1337
1338 /**
1339 * bfq_first_active_entity - find the eligible entity with
1340 * the smallest finish time
1341 * @st: the service tree to select from.
1342 * @vtime: the system virtual to use as a reference for eligibility
1343 *
1344 * This function searches the first schedulable entity, starting from the
1345 * root of the tree and going on the left every time on this side there is
1346 * a subtree with at least one eligible (start <= vtime) entity. The path on
1347 * the right is followed only if a) the left subtree contains no eligible
1348 * entities and b) no eligible entity has been found yet.
1349 */
bfq_first_active_entity(struct bfq_service_tree * st,u64 vtime)1350 static struct bfq_entity *bfq_first_active_entity(struct bfq_service_tree *st,
1351 u64 vtime)
1352 {
1353 struct bfq_entity *entry, *first = NULL;
1354 struct rb_node *node = st->active.rb_node;
1355
1356 while (node) {
1357 entry = rb_entry(node, struct bfq_entity, rb_node);
1358 left:
1359 if (!bfq_gt(entry->start, vtime))
1360 first = entry;
1361
1362 if (node->rb_left) {
1363 entry = rb_entry(node->rb_left,
1364 struct bfq_entity, rb_node);
1365 if (!bfq_gt(entry->min_start, vtime)) {
1366 node = node->rb_left;
1367 goto left;
1368 }
1369 }
1370 if (first)
1371 break;
1372 node = node->rb_right;
1373 }
1374
1375 return first;
1376 }
1377
1378 /**
1379 * __bfq_lookup_next_entity - return the first eligible entity in @st.
1380 * @st: the service tree.
1381 *
1382 * If there is no in-service entity for the sched_data st belongs to,
1383 * then return the entity that will be set in service if:
1384 * 1) the parent entity this st belongs to is set in service;
1385 * 2) no entity belonging to such parent entity undergoes a state change
1386 * that would influence the timestamps of the entity (e.g., becomes idle,
1387 * becomes backlogged, changes its budget, ...).
1388 *
1389 * In this first case, update the virtual time in @st too (see the
1390 * comments on this update inside the function).
1391 *
1392 * In contrast, if there is an in-service entity, then return the
1393 * entity that would be set in service if not only the above
1394 * conditions, but also the next one held true: the currently
1395 * in-service entity, on expiration,
1396 * 1) gets a finish time equal to the current one, or
1397 * 2) is not eligible any more, or
1398 * 3) is idle.
1399 */
1400 static struct bfq_entity *
__bfq_lookup_next_entity(struct bfq_service_tree * st,bool in_service)1401 __bfq_lookup_next_entity(struct bfq_service_tree *st, bool in_service)
1402 {
1403 struct bfq_entity *entity;
1404 u64 new_vtime;
1405
1406 if (RB_EMPTY_ROOT(&st->active))
1407 return NULL;
1408
1409 /*
1410 * Get the value of the system virtual time for which at
1411 * least one entity is eligible.
1412 */
1413 new_vtime = bfq_calc_vtime_jump(st);
1414
1415 /*
1416 * If there is no in-service entity for the sched_data this
1417 * active tree belongs to, then push the system virtual time
1418 * up to the value that guarantees that at least one entity is
1419 * eligible. If, instead, there is an in-service entity, then
1420 * do not make any such update, because there is already an
1421 * eligible entity, namely the in-service one (even if the
1422 * entity is not on st, because it was extracted when set in
1423 * service).
1424 */
1425 if (!in_service)
1426 bfq_update_vtime(st, new_vtime);
1427
1428 entity = bfq_first_active_entity(st, new_vtime);
1429
1430 return entity;
1431 }
1432
1433 /**
1434 * bfq_lookup_next_entity - return the first eligible entity in @sd.
1435 * @sd: the sched_data.
1436 * @expiration: true if we are on the expiration path of the in-service queue
1437 *
1438 * This function is invoked when there has been a change in the trees
1439 * for sd, and we need to know what is the new next entity to serve
1440 * after this change.
1441 */
bfq_lookup_next_entity(struct bfq_sched_data * sd,bool expiration)1442 static struct bfq_entity *bfq_lookup_next_entity(struct bfq_sched_data *sd,
1443 bool expiration)
1444 {
1445 struct bfq_service_tree *st = sd->service_tree;
1446 struct bfq_service_tree *idle_class_st = st + (BFQ_IOPRIO_CLASSES - 1);
1447 struct bfq_entity *entity = NULL;
1448 int class_idx = 0;
1449
1450 /*
1451 * Choose from idle class, if needed to guarantee a minimum
1452 * bandwidth to this class (and if there is some active entity
1453 * in idle class). This should also mitigate
1454 * priority-inversion problems in case a low priority task is
1455 * holding file system resources.
1456 */
1457 if (time_is_before_jiffies(sd->bfq_class_idle_last_service +
1458 BFQ_CL_IDLE_TIMEOUT)) {
1459 if (!RB_EMPTY_ROOT(&idle_class_st->active))
1460 class_idx = BFQ_IOPRIO_CLASSES - 1;
1461 /* About to be served if backlogged, or not yet backlogged */
1462 sd->bfq_class_idle_last_service = jiffies;
1463 }
1464
1465 /*
1466 * Find the next entity to serve for the highest-priority
1467 * class, unless the idle class needs to be served.
1468 */
1469 for (; class_idx < BFQ_IOPRIO_CLASSES; class_idx++) {
1470 /*
1471 * If expiration is true, then bfq_lookup_next_entity
1472 * is being invoked as a part of the expiration path
1473 * of the in-service queue. In this case, even if
1474 * sd->in_service_entity is not NULL,
1475 * sd->in_service_entity at this point is actually not
1476 * in service any more, and, if needed, has already
1477 * been properly queued or requeued into the right
1478 * tree. The reason why sd->in_service_entity is still
1479 * not NULL here, even if expiration is true, is that
1480 * sd->in_service_entity is reset as a last step in the
1481 * expiration path. So, if expiration is true, tell
1482 * __bfq_lookup_next_entity that there is no
1483 * sd->in_service_entity.
1484 */
1485 entity = __bfq_lookup_next_entity(st + class_idx,
1486 sd->in_service_entity &&
1487 !expiration);
1488
1489 if (entity)
1490 break;
1491 }
1492
1493 if (!entity)
1494 return NULL;
1495
1496 return entity;
1497 }
1498
next_queue_may_preempt(struct bfq_data * bfqd)1499 bool next_queue_may_preempt(struct bfq_data *bfqd)
1500 {
1501 struct bfq_sched_data *sd = &bfqd->root_group->sched_data;
1502
1503 return sd->next_in_service != sd->in_service_entity;
1504 }
1505
1506 /*
1507 * Get next queue for service.
1508 */
bfq_get_next_queue(struct bfq_data * bfqd)1509 struct bfq_queue *bfq_get_next_queue(struct bfq_data *bfqd)
1510 {
1511 struct bfq_entity *entity = NULL;
1512 struct bfq_sched_data *sd;
1513 struct bfq_queue *bfqq;
1514
1515 if (bfq_tot_busy_queues(bfqd) == 0)
1516 return NULL;
1517
1518 /*
1519 * Traverse the path from the root to the leaf entity to
1520 * serve. Set in service all the entities visited along the
1521 * way.
1522 */
1523 sd = &bfqd->root_group->sched_data;
1524 for (; sd ; sd = entity->my_sched_data) {
1525 /*
1526 * WARNING. We are about to set the in-service entity
1527 * to sd->next_in_service, i.e., to the (cached) value
1528 * returned by bfq_lookup_next_entity(sd) the last
1529 * time it was invoked, i.e., the last time when the
1530 * service order in sd changed as a consequence of the
1531 * activation or deactivation of an entity. In this
1532 * respect, if we execute bfq_lookup_next_entity(sd)
1533 * in this very moment, it may, although with low
1534 * probability, yield a different entity than that
1535 * pointed to by sd->next_in_service. This rare event
1536 * happens in case there was no CLASS_IDLE entity to
1537 * serve for sd when bfq_lookup_next_entity(sd) was
1538 * invoked for the last time, while there is now one
1539 * such entity.
1540 *
1541 * If the above event happens, then the scheduling of
1542 * such entity in CLASS_IDLE is postponed until the
1543 * service of the sd->next_in_service entity
1544 * finishes. In fact, when the latter is expired,
1545 * bfq_lookup_next_entity(sd) gets called again,
1546 * exactly to update sd->next_in_service.
1547 */
1548
1549 /* Make next_in_service entity become in_service_entity */
1550 entity = sd->next_in_service;
1551 sd->in_service_entity = entity;
1552
1553 /*
1554 * If entity is no longer a candidate for next
1555 * service, then it must be extracted from its active
1556 * tree, so as to make sure that it won't be
1557 * considered when computing next_in_service. See the
1558 * comments on the function
1559 * bfq_no_longer_next_in_service() for details.
1560 */
1561 if (bfq_no_longer_next_in_service(entity))
1562 bfq_active_extract(bfq_entity_service_tree(entity),
1563 entity);
1564
1565 /*
1566 * Even if entity is not to be extracted according to
1567 * the above check, a descendant entity may get
1568 * extracted in one of the next iterations of this
1569 * loop. Such an event could cause a change in
1570 * next_in_service for the level of the descendant
1571 * entity, and thus possibly back to this level.
1572 *
1573 * However, we cannot perform the resulting needed
1574 * update of next_in_service for this level before the
1575 * end of the whole loop, because, to know which is
1576 * the correct next-to-serve candidate entity for each
1577 * level, we need first to find the leaf entity to set
1578 * in service. In fact, only after we know which is
1579 * the next-to-serve leaf entity, we can discover
1580 * whether the parent entity of the leaf entity
1581 * becomes the next-to-serve, and so on.
1582 */
1583 }
1584
1585 bfqq = bfq_entity_to_bfqq(entity);
1586
1587 /*
1588 * We can finally update all next-to-serve entities along the
1589 * path from the leaf entity just set in service to the root.
1590 */
1591 for_each_entity(entity) {
1592 struct bfq_sched_data *sd = entity->sched_data;
1593
1594 if (!bfq_update_next_in_service(sd, NULL, false))
1595 break;
1596 }
1597
1598 return bfqq;
1599 }
1600
1601 /* returns true if the in-service queue gets freed */
__bfq_bfqd_reset_in_service(struct bfq_data * bfqd)1602 bool __bfq_bfqd_reset_in_service(struct bfq_data *bfqd)
1603 {
1604 struct bfq_queue *in_serv_bfqq = bfqd->in_service_queue;
1605 struct bfq_entity *in_serv_entity = &in_serv_bfqq->entity;
1606 struct bfq_entity *entity = in_serv_entity;
1607
1608 bfq_clear_bfqq_wait_request(in_serv_bfqq);
1609 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
1610 bfqd->in_service_queue = NULL;
1611
1612 /*
1613 * When this function is called, all in-service entities have
1614 * been properly deactivated or requeued, so we can safely
1615 * execute the final step: reset in_service_entity along the
1616 * path from entity to the root.
1617 */
1618 for_each_entity(entity)
1619 entity->sched_data->in_service_entity = NULL;
1620
1621 /*
1622 * in_serv_entity is no longer in service, so, if it is in no
1623 * service tree either, then release the service reference to
1624 * the queue it represents (taken with bfq_get_entity).
1625 */
1626 if (!in_serv_entity->on_st_or_in_serv) {
1627 /*
1628 * If no process is referencing in_serv_bfqq any
1629 * longer, then the service reference may be the only
1630 * reference to the queue. If this is the case, then
1631 * bfqq gets freed here.
1632 */
1633 int ref = in_serv_bfqq->ref;
1634 bfq_put_queue(in_serv_bfqq);
1635 if (ref == 1)
1636 return true;
1637 }
1638
1639 return false;
1640 }
1641
bfq_deactivate_bfqq(struct bfq_data * bfqd,struct bfq_queue * bfqq,bool ins_into_idle_tree,bool expiration)1642 void bfq_deactivate_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1643 bool ins_into_idle_tree, bool expiration)
1644 {
1645 struct bfq_entity *entity = &bfqq->entity;
1646
1647 bfq_deactivate_entity(entity, ins_into_idle_tree, expiration);
1648 }
1649
bfq_activate_bfqq(struct bfq_data * bfqd,struct bfq_queue * bfqq)1650 void bfq_activate_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1651 {
1652 struct bfq_entity *entity = &bfqq->entity;
1653
1654 bfq_activate_requeue_entity(entity, bfq_bfqq_non_blocking_wait_rq(bfqq),
1655 false, false);
1656 bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1657 }
1658
bfq_requeue_bfqq(struct bfq_data * bfqd,struct bfq_queue * bfqq,bool expiration)1659 void bfq_requeue_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1660 bool expiration)
1661 {
1662 struct bfq_entity *entity = &bfqq->entity;
1663
1664 bfq_activate_requeue_entity(entity, false,
1665 bfqq == bfqd->in_service_queue, expiration);
1666 }
1667
1668 /*
1669 * Called when the bfqq no longer has requests pending, remove it from
1670 * the service tree. As a special case, it can be invoked during an
1671 * expiration.
1672 */
bfq_del_bfqq_busy(struct bfq_data * bfqd,struct bfq_queue * bfqq,bool expiration)1673 void bfq_del_bfqq_busy(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1674 bool expiration)
1675 {
1676 bfq_log_bfqq(bfqd, bfqq, "del from busy");
1677
1678 bfq_clear_bfqq_busy(bfqq);
1679
1680 bfqd->busy_queues[bfqq->ioprio_class - 1]--;
1681
1682 if (bfqq->wr_coeff > 1)
1683 bfqd->wr_busy_queues--;
1684
1685 bfqg_stats_update_dequeue(bfqq_group(bfqq));
1686
1687 bfq_deactivate_bfqq(bfqd, bfqq, true, expiration);
1688
1689 if (!bfqq->dispatched)
1690 bfq_weights_tree_remove(bfqd, bfqq);
1691 }
1692
1693 /*
1694 * Called when an inactive queue receives a new request.
1695 */
bfq_add_bfqq_busy(struct bfq_data * bfqd,struct bfq_queue * bfqq)1696 void bfq_add_bfqq_busy(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1697 {
1698 bfq_log_bfqq(bfqd, bfqq, "add to busy");
1699
1700 bfq_activate_bfqq(bfqd, bfqq);
1701
1702 bfq_mark_bfqq_busy(bfqq);
1703 bfqd->busy_queues[bfqq->ioprio_class - 1]++;
1704
1705 if (!bfqq->dispatched)
1706 if (bfqq->wr_coeff == 1)
1707 bfq_weights_tree_add(bfqd, bfqq,
1708 &bfqd->queue_weights_tree);
1709
1710 if (bfqq->wr_coeff > 1)
1711 bfqd->wr_busy_queues++;
1712 }
1713
