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
bfq_bfqq_to_bfqg(struct bfq_queue * bfqq)145 struct bfq_group *bfq_bfqq_to_bfqg(struct bfq_queue *bfqq)
146 {
147 struct bfq_entity *group_entity = bfqq->entity.parent;
148
149 if (!group_entity)
150 group_entity = &bfqq->bfqd->root_group->entity;
151
152 return container_of(group_entity, struct bfq_group, entity);
153 }
154
155 /*
156 * Returns true if this budget changes may let next_in_service->parent
157 * become the next_in_service entity for its parent entity.
158 */
bfq_update_parent_budget(struct bfq_entity * next_in_service)159 static bool bfq_update_parent_budget(struct bfq_entity *next_in_service)
160 {
161 struct bfq_entity *bfqg_entity;
162 struct bfq_group *bfqg;
163 struct bfq_sched_data *group_sd;
164 bool ret = false;
165
166 group_sd = next_in_service->sched_data;
167
168 bfqg = container_of(group_sd, struct bfq_group, sched_data);
169 /*
170 * bfq_group's my_entity field is not NULL only if the group
171 * is not the root group. We must not touch the root entity
172 * as it must never become an in-service entity.
173 */
174 bfqg_entity = bfqg->my_entity;
175 if (bfqg_entity) {
176 if (bfqg_entity->budget > next_in_service->budget)
177 ret = true;
178 bfqg_entity->budget = next_in_service->budget;
179 }
180
181 return ret;
182 }
183
184 /*
185 * This function tells whether entity stops being a candidate for next
186 * service, according to the restrictive definition of the field
187 * next_in_service. In particular, this function is invoked for an
188 * entity that is about to be set in service.
189 *
190 * If entity is a queue, then the entity is no longer a candidate for
191 * next service according to the that definition, because entity is
192 * about to become the in-service queue. This function then returns
193 * true if entity is a queue.
194 *
195 * In contrast, entity could still be a candidate for next service if
196 * it is not a queue, and has more than one active child. In fact,
197 * even if one of its children is about to be set in service, other
198 * active children may still be the next to serve, for the parent
199 * entity, even according to the above definition. As a consequence, a
200 * non-queue entity is not a candidate for next-service only if it has
201 * only one active child. And only if this condition holds, then this
202 * function returns true for a non-queue entity.
203 */
bfq_no_longer_next_in_service(struct bfq_entity * entity)204 static bool bfq_no_longer_next_in_service(struct bfq_entity *entity)
205 {
206 struct bfq_group *bfqg;
207
208 if (bfq_entity_to_bfqq(entity))
209 return true;
210
211 bfqg = container_of(entity, struct bfq_group, entity);
212
213 /*
214 * The field active_entities does not always contain the
215 * actual number of active children entities: it happens to
216 * not account for the in-service entity in case the latter is
217 * removed from its active tree (which may get done after
218 * invoking the function bfq_no_longer_next_in_service in
219 * bfq_get_next_queue). Fortunately, here, i.e., while
220 * bfq_no_longer_next_in_service is not yet completed in
221 * bfq_get_next_queue, bfq_active_extract has not yet been
222 * invoked, and thus active_entities still coincides with the
223 * actual number of active entities.
224 */
225 if (bfqg->active_entities == 1)
226 return true;
227
228 return false;
229 }
230
231 #else /* CONFIG_BFQ_GROUP_IOSCHED */
232
bfq_bfqq_to_bfqg(struct bfq_queue * bfqq)233 struct bfq_group *bfq_bfqq_to_bfqg(struct bfq_queue *bfqq)
234 {
235 return bfqq->bfqd->root_group;
236 }
237
bfq_update_parent_budget(struct bfq_entity * next_in_service)238 static bool bfq_update_parent_budget(struct bfq_entity *next_in_service)
239 {
240 return false;
241 }
242
bfq_no_longer_next_in_service(struct bfq_entity * entity)243 static bool bfq_no_longer_next_in_service(struct bfq_entity *entity)
244 {
245 return true;
246 }
247
248 #endif /* CONFIG_BFQ_GROUP_IOSCHED */
249
250 /*
251 * Shift for timestamp calculations. This actually limits the maximum
252 * service allowed in one timestamp delta (small shift values increase it),
253 * the maximum total weight that can be used for the queues in the system
254 * (big shift values increase it), and the period of virtual time
255 * wraparounds.
256 */
257 #define WFQ_SERVICE_SHIFT 22
258
bfq_entity_to_bfqq(struct bfq_entity * entity)259 struct bfq_queue *bfq_entity_to_bfqq(struct bfq_entity *entity)
260 {
261 struct bfq_queue *bfqq = NULL;
262
263 if (!entity->my_sched_data)
264 bfqq = container_of(entity, struct bfq_queue, entity);
265
266 return bfqq;
267 }
268
269
270 /**
271 * bfq_delta - map service into the virtual time domain.
272 * @service: amount of service.
273 * @weight: scale factor (weight of an entity or weight sum).
274 */
bfq_delta(unsigned long service,unsigned long weight)275 static u64 bfq_delta(unsigned long service, unsigned long weight)
276 {
277 return div64_ul((u64)service << WFQ_SERVICE_SHIFT, weight);
278 }
279
280 /**
281 * bfq_calc_finish - assign the finish time to an entity.
282 * @entity: the entity to act upon.
283 * @service: the service to be charged to the entity.
284 */
bfq_calc_finish(struct bfq_entity * entity,unsigned long service)285 static void bfq_calc_finish(struct bfq_entity *entity, unsigned long service)
286 {
287 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
288
289 entity->finish = entity->start +
290 bfq_delta(service, entity->weight);
291
292 if (bfqq) {
293 bfq_log_bfqq(bfqq->bfqd, bfqq,
294 "calc_finish: serv %lu, w %d",
295 service, entity->weight);
296 bfq_log_bfqq(bfqq->bfqd, bfqq,
297 "calc_finish: start %llu, finish %llu, delta %llu",
298 entity->start, entity->finish,
299 bfq_delta(service, entity->weight));
300 }
301 }
302
303 /**
304 * bfq_entity_of - get an entity from a node.
305 * @node: the node field of the entity.
306 *
307 * Convert a node pointer to the relative entity. This is used only
308 * to simplify the logic of some functions and not as the generic
309 * conversion mechanism because, e.g., in the tree walking functions,
310 * the check for a %NULL value would be redundant.
311 */
bfq_entity_of(struct rb_node * node)312 struct bfq_entity *bfq_entity_of(struct rb_node *node)
313 {
314 struct bfq_entity *entity = NULL;
315
316 if (node)
317 entity = rb_entry(node, struct bfq_entity, rb_node);
318
319 return entity;
320 }
321
322 /**
323 * bfq_extract - remove an entity from a tree.
324 * @root: the tree root.
325 * @entity: the entity to remove.
326 */
bfq_extract(struct rb_root * root,struct bfq_entity * entity)327 static void bfq_extract(struct rb_root *root, struct bfq_entity *entity)
328 {
329 entity->tree = NULL;
330 rb_erase(&entity->rb_node, root);
331 }
332
333 /**
334 * bfq_idle_extract - extract an entity from the idle tree.
335 * @st: the service tree of the owning @entity.
336 * @entity: the entity being removed.
337 */
bfq_idle_extract(struct bfq_service_tree * st,struct bfq_entity * entity)338 static void bfq_idle_extract(struct bfq_service_tree *st,
339 struct bfq_entity *entity)
340 {
341 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
342 struct rb_node *next;
343
344 if (entity == st->first_idle) {
345 next = rb_next(&entity->rb_node);
346 st->first_idle = bfq_entity_of(next);
347 }
348
349 if (entity == st->last_idle) {
350 next = rb_prev(&entity->rb_node);
351 st->last_idle = bfq_entity_of(next);
352 }
353
354 bfq_extract(&st->idle, entity);
355
356 if (bfqq)
357 list_del(&bfqq->bfqq_list);
358 }
359
360 /**
361 * bfq_insert - generic tree insertion.
362 * @root: tree root.
363 * @entity: entity to insert.
364 *
365 * This is used for the idle and the active tree, since they are both
366 * ordered by finish time.
367 */
bfq_insert(struct rb_root * root,struct bfq_entity * entity)368 static void bfq_insert(struct rb_root *root, struct bfq_entity *entity)
369 {
370 struct bfq_entity *entry;
371 struct rb_node **node = &root->rb_node;
372 struct rb_node *parent = NULL;
373
374 while (*node) {
375 parent = *node;
376 entry = rb_entry(parent, struct bfq_entity, rb_node);
377
378 if (bfq_gt(entry->finish, entity->finish))
379 node = &parent->rb_left;
380 else
381 node = &parent->rb_right;
382 }
383
384 rb_link_node(&entity->rb_node, parent, node);
385 rb_insert_color(&entity->rb_node, root);
386
387 entity->tree = root;
388 }
389
390 /**
391 * bfq_update_min - update the min_start field of a entity.
392 * @entity: the entity to update.
393 * @node: one of its children.
394 *
395 * This function is called when @entity may store an invalid value for
396 * min_start due to updates to the active tree. The function assumes
397 * that the subtree rooted at @node (which may be its left or its right
398 * child) has a valid min_start value.
399 */
bfq_update_min(struct bfq_entity * entity,struct rb_node * node)400 static void bfq_update_min(struct bfq_entity *entity, struct rb_node *node)
401 {
402 struct bfq_entity *child;
403
404 if (node) {
405 child = rb_entry(node, struct bfq_entity, rb_node);
406 if (bfq_gt(entity->min_start, child->min_start))
407 entity->min_start = child->min_start;
408 }
409 }
410
411 /**
412 * bfq_update_active_node - recalculate min_start.
413 * @node: the node to update.
414 *
415 * @node may have changed position or one of its children may have moved,
416 * this function updates its min_start value. The left and right subtrees
417 * are assumed to hold a correct min_start value.
418 */
bfq_update_active_node(struct rb_node * node)419 static void bfq_update_active_node(struct rb_node *node)
420 {
421 struct bfq_entity *entity = rb_entry(node, struct bfq_entity, rb_node);
422
423 entity->min_start = entity->start;
424 bfq_update_min(entity, node->rb_right);
425 bfq_update_min(entity, node->rb_left);
426 }
427
428 /**
429 * bfq_update_active_tree - update min_start for the whole active tree.
430 * @node: the starting node.
431 *
432 * @node must be the deepest modified node after an update. This function
433 * updates its min_start using the values held by its children, assuming
434 * that they did not change, and then updates all the nodes that may have
435 * changed in the path to the root. The only nodes that may have changed
436 * are the ones in the path or their siblings.
437 */
bfq_update_active_tree(struct rb_node * node)438 static void bfq_update_active_tree(struct rb_node *node)
439 {
440 struct rb_node *parent;
441
442 up:
443 bfq_update_active_node(node);
444
445 parent = rb_parent(node);
446 if (!parent)
447 return;
448
449 if (node == parent->rb_left && parent->rb_right)
450 bfq_update_active_node(parent->rb_right);
451 else if (parent->rb_left)
452 bfq_update_active_node(parent->rb_left);
453
454 node = parent;
455 goto up;
456 }
457
458 /**
459 * bfq_active_insert - insert an entity in the active tree of its
460 * group/device.
461 * @st: the service tree of the entity.
462 * @entity: the entity being inserted.
463 *
464 * The active tree is ordered by finish time, but an extra key is kept
465 * per each node, containing the minimum value for the start times of
466 * its children (and the node itself), so it's possible to search for
467 * the eligible node with the lowest finish time in logarithmic time.
468 */
bfq_active_insert(struct bfq_service_tree * st,struct bfq_entity * entity)469 static void bfq_active_insert(struct bfq_service_tree *st,
470 struct bfq_entity *entity)
471 {
472 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
473 struct rb_node *node = &entity->rb_node;
474 #ifdef CONFIG_BFQ_GROUP_IOSCHED
475 struct bfq_sched_data *sd = NULL;
476 struct bfq_group *bfqg = NULL;
477 struct bfq_data *bfqd = NULL;
478 #endif
479
480 bfq_insert(&st->active, entity);
481
482 if (node->rb_left)
483 node = node->rb_left;
484 else if (node->rb_right)
485 node = node->rb_right;
486
487 bfq_update_active_tree(node);
488
489 #ifdef CONFIG_BFQ_GROUP_IOSCHED
490 sd = entity->sched_data;
491 bfqg = container_of(sd, struct bfq_group, sched_data);
492 bfqd = (struct bfq_data *)bfqg->bfqd;
493 #endif
494 if (bfqq)
495 list_add(&bfqq->bfqq_list, &bfqq->bfqd->active_list);
496 #ifdef CONFIG_BFQ_GROUP_IOSCHED
497 if (bfqg != bfqd->root_group)
498 bfqg->active_entities++;
499 #endif
500 }
501
502 /**
503 * bfq_ioprio_to_weight - calc a weight from an ioprio.
504 * @ioprio: the ioprio value to convert.
505 */
bfq_ioprio_to_weight(int ioprio)506 unsigned short bfq_ioprio_to_weight(int ioprio)
507 {
508 return (IOPRIO_NR_LEVELS - ioprio) * BFQ_WEIGHT_CONVERSION_COEFF;
509 }
510
511 /**
512 * bfq_weight_to_ioprio - calc an ioprio from a weight.
513 * @weight: the weight value to convert.
514 *
515 * To preserve as much as possible the old only-ioprio user interface,
516 * 0 is used as an escape ioprio value for weights (numerically) equal or
517 * larger than IOPRIO_NR_LEVELS * BFQ_WEIGHT_CONVERSION_COEFF.
518 */
bfq_weight_to_ioprio(int weight)519 static unsigned short bfq_weight_to_ioprio(int weight)
520 {
521 return max_t(int, 0,
522 IOPRIO_NR_LEVELS - weight / BFQ_WEIGHT_CONVERSION_COEFF);
523 }
524
bfq_get_entity(struct bfq_entity * entity)525 static void bfq_get_entity(struct bfq_entity *entity)
526 {
527 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
528
529 if (bfqq) {
530 bfqq->ref++;
531 bfq_log_bfqq(bfqq->bfqd, bfqq, "get_entity: %p %d",
532 bfqq, bfqq->ref);
533 }
534 }
535
536 /**
537 * bfq_find_deepest - find the deepest node that an extraction can modify.
538 * @node: the node being removed.
539 *
540 * Do the first step of an extraction in an rb tree, looking for the
541 * node that will replace @node, and returning the deepest node that
542 * the following modifications to the tree can touch. If @node is the
543 * last node in the tree return %NULL.
544 */
bfq_find_deepest(struct rb_node * node)545 static struct rb_node *bfq_find_deepest(struct rb_node *node)
546 {
547 struct rb_node *deepest;
548
549 if (!node->rb_right && !node->rb_left)
550 deepest = rb_parent(node);
551 else if (!node->rb_right)
552 deepest = node->rb_left;
553 else if (!node->rb_left)
554 deepest = node->rb_right;
555 else {
556 deepest = rb_next(node);
557 if (deepest->rb_right)
558 deepest = deepest->rb_right;
559 else if (rb_parent(deepest) != node)
560 deepest = rb_parent(deepest);
561 }
562
563 return deepest;
564 }
565
566 /**
567 * bfq_active_extract - remove an entity from the active tree.
568 * @st: the service_tree containing the tree.
569 * @entity: the entity being removed.
570 */
bfq_active_extract(struct bfq_service_tree * st,struct bfq_entity * entity)571 static void bfq_active_extract(struct bfq_service_tree *st,
572 struct bfq_entity *entity)
573 {
574 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
575 struct rb_node *node;
576 #ifdef CONFIG_BFQ_GROUP_IOSCHED
577 struct bfq_sched_data *sd = NULL;
578 struct bfq_group *bfqg = NULL;
579 struct bfq_data *bfqd = NULL;
580 #endif
581
582 node = bfq_find_deepest(&entity->rb_node);
583 bfq_extract(&st->active, entity);
584
585 if (node)
586 bfq_update_active_tree(node);
587
588 #ifdef CONFIG_BFQ_GROUP_IOSCHED
589 sd = entity->sched_data;
590 bfqg = container_of(sd, struct bfq_group, sched_data);
591 bfqd = (struct bfq_data *)bfqg->bfqd;
592 #endif
593 if (bfqq)
594 list_del(&bfqq->bfqq_list);
595 #ifdef CONFIG_BFQ_GROUP_IOSCHED
596 if (bfqg != bfqd->root_group)
597 bfqg->active_entities--;
598 #endif
599 }
600
601 /**
602 * bfq_idle_insert - insert an entity into the idle tree.
603 * @st: the service tree containing the tree.
604 * @entity: the entity to insert.
605 */
bfq_idle_insert(struct bfq_service_tree * st,struct bfq_entity * entity)606 static void bfq_idle_insert(struct bfq_service_tree *st,
607 struct bfq_entity *entity)
608 {
609 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
610 struct bfq_entity *first_idle = st->first_idle;
611 struct bfq_entity *last_idle = st->last_idle;
612
613 if (!first_idle || bfq_gt(first_idle->finish, entity->finish))
614 st->first_idle = entity;
615 if (!last_idle || bfq_gt(entity->finish, last_idle->finish))
616 st->last_idle = entity;
617
618 bfq_insert(&st->idle, entity);
619
620 if (bfqq)
621 list_add(&bfqq->bfqq_list, &bfqq->bfqd->idle_list);
622 }
623
624 /**
625 * bfq_forget_entity - do not consider entity any longer for scheduling
626 * @st: the service tree.
627 * @entity: the entity being removed.
628 * @is_in_service: true if entity is currently the in-service entity.
629 *
630 * Forget everything about @entity. In addition, if entity represents
631 * a queue, and the latter is not in service, then release the service
632 * reference to the queue (the one taken through bfq_get_entity). In
633 * fact, in this case, there is really no more service reference to
634 * the queue, as the latter is also outside any service tree. If,
635 * instead, the queue is in service, then __bfq_bfqd_reset_in_service
636 * will take care of putting the reference when the queue finally
637 * stops being served.
638 */
bfq_forget_entity(struct bfq_service_tree * st,struct bfq_entity * entity,bool is_in_service)639 static void bfq_forget_entity(struct bfq_service_tree *st,
640 struct bfq_entity *entity,
641 bool is_in_service)
642 {
643 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
644
645 entity->on_st_or_in_serv = false;
646 st->wsum -= entity->weight;
647 if (bfqq && !is_in_service)
648 bfq_put_queue(bfqq);
649 }
650
651 /**
652 * bfq_put_idle_entity - release the idle tree ref of an entity.
653 * @st: service tree for the entity.
654 * @entity: the entity being released.
655 */
bfq_put_idle_entity(struct bfq_service_tree * st,struct bfq_entity * entity)656 void bfq_put_idle_entity(struct bfq_service_tree *st, struct bfq_entity *entity)
657 {
658 bfq_idle_extract(st, entity);
659 bfq_forget_entity(st, entity,
660 entity == entity->sched_data->in_service_entity);
661 }
662
663 /**
664 * bfq_forget_idle - update the idle tree if necessary.
665 * @st: the service tree to act upon.
666 *
667 * To preserve the global O(log N) complexity we only remove one entry here;
668 * as the idle tree will not grow indefinitely this can be done safely.
669 */
bfq_forget_idle(struct bfq_service_tree * st)670 static void bfq_forget_idle(struct bfq_service_tree *st)
671 {
672 struct bfq_entity *first_idle = st->first_idle;
673 struct bfq_entity *last_idle = st->last_idle;
674
675 if (RB_EMPTY_ROOT(&st->active) && last_idle &&
676 !bfq_gt(last_idle->finish, st->vtime)) {
677 /*
678 * Forget the whole idle tree, increasing the vtime past
679 * the last finish time of idle entities.
680 */
681 st->vtime = last_idle->finish;
682 }
683
684 if (first_idle && !bfq_gt(first_idle->finish, st->vtime))
685 bfq_put_idle_entity(st, first_idle);
686 }
687
bfq_entity_service_tree(struct bfq_entity * entity)688 struct bfq_service_tree *bfq_entity_service_tree(struct bfq_entity *entity)
689 {
690 struct bfq_sched_data *sched_data = entity->sched_data;
691 unsigned int idx = bfq_class_idx(entity);
692
693 return sched_data->service_tree + idx;
694 }
695
696 /*
697 * Update weight and priority of entity. If update_class_too is true,
698 * then update the ioprio_class of entity too.
699 *
700 * The reason why the update of ioprio_class is controlled through the
701 * last parameter is as follows. Changing the ioprio class of an
702 * entity implies changing the destination service trees for that
703 * entity. If such a change occurred when the entity is already on one
704 * of the service trees for its previous class, then the state of the
705 * entity would become more complex: none of the new possible service
706 * trees for the entity, according to bfq_entity_service_tree(), would
707 * match any of the possible service trees on which the entity
708 * is. Complex operations involving these trees, such as entity
709 * activations and deactivations, should take into account this
710 * additional complexity. To avoid this issue, this function is
711 * invoked with update_class_too unset in the points in the code where
712 * entity may happen to be on some tree.
713 */
714 struct bfq_service_tree *
__bfq_entity_update_weight_prio(struct bfq_service_tree * old_st,struct bfq_entity * entity,bool update_class_too)715 __bfq_entity_update_weight_prio(struct bfq_service_tree *old_st,
716 struct bfq_entity *entity,
717 bool update_class_too)
718 {
719 struct bfq_service_tree *new_st = old_st;
720
721 if (entity->prio_changed) {
722 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
723 unsigned int prev_weight, new_weight;
724 struct bfq_data *bfqd = NULL;
725 struct rb_root_cached *root;
726 #ifdef CONFIG_BFQ_GROUP_IOSCHED
727 struct bfq_sched_data *sd;
728 struct bfq_group *bfqg;
729 #endif
730
731 if (bfqq)
732 bfqd = bfqq->bfqd;
733 #ifdef CONFIG_BFQ_GROUP_IOSCHED
734 else {
735 sd = entity->my_sched_data;
736 bfqg = container_of(sd, struct bfq_group, sched_data);
737 bfqd = (struct bfq_data *)bfqg->bfqd;
738 }
739 #endif
740
741 /* Matches the smp_wmb() in bfq_group_set_weight. */
742 smp_rmb();
743 old_st->wsum -= entity->weight;
744
745 if (entity->new_weight != entity->orig_weight) {
746 if (entity->new_weight < BFQ_MIN_WEIGHT ||
747 entity->new_weight > BFQ_MAX_WEIGHT) {
748 pr_crit("update_weight_prio: new_weight %d\n",
749 entity->new_weight);
750 if (entity->new_weight < BFQ_MIN_WEIGHT)
751 entity->new_weight = BFQ_MIN_WEIGHT;
752 else
753 entity->new_weight = BFQ_MAX_WEIGHT;
754 }
755 entity->orig_weight = entity->new_weight;
756 if (bfqq)
757 bfqq->ioprio =
758 bfq_weight_to_ioprio(entity->orig_weight);
759 }
760
761 if (bfqq && update_class_too)
762 bfqq->ioprio_class = bfqq->new_ioprio_class;
763
764 /*
765 * Reset prio_changed only if the ioprio_class change
766 * is not pending any longer.
767 */
768 if (!bfqq || bfqq->ioprio_class == bfqq->new_ioprio_class)
769 entity->prio_changed = 0;
770
771 /*
772 * NOTE: here we may be changing the weight too early,
773 * this will cause unfairness. The correct approach
774 * would have required additional complexity to defer
775 * weight changes to the proper time instants (i.e.,
776 * when entity->finish <= old_st->vtime).
777 */
778 new_st = bfq_entity_service_tree(entity);
779
780 prev_weight = entity->weight;
781 new_weight = entity->orig_weight *
782 (bfqq ? bfqq->wr_coeff : 1);
783 /*
784 * If the weight of the entity changes, and the entity is a
785 * queue, remove the entity from its old weight counter (if
786 * there is a counter associated with the entity).
787 */
788 if (prev_weight != new_weight && bfqq) {
789 root = &bfqd->queue_weights_tree;
790 __bfq_weights_tree_remove(bfqd, bfqq, root);
791 }
792 entity->weight = new_weight;
793 /*
794 * Add the entity, if it is not a weight-raised queue,
795 * to the counter associated with its new weight.
796 */
797 if (prev_weight != new_weight && bfqq && bfqq->wr_coeff == 1) {
798 /* If we get here, root has been initialized. */
799 bfq_weights_tree_add(bfqd, bfqq, root);
800 }
801
802 new_st->wsum += entity->weight;
803
804 if (new_st != old_st)
805 entity->start = new_st->vtime;
806 }
807
808 return new_st;
809 }
810
811 /**
812 * bfq_bfqq_served - update the scheduler status after selection for
813 * service.
814 * @bfqq: the queue being served.
815 * @served: bytes to transfer.
816 *
817 * NOTE: this can be optimized, as the timestamps of upper level entities
818 * are synchronized every time a new bfqq is selected for service. By now,
819 * we keep it to better check consistency.
820 */
bfq_bfqq_served(struct bfq_queue * bfqq,int served)821 void bfq_bfqq_served(struct bfq_queue *bfqq, int served)
822 {
823 struct bfq_entity *entity = &bfqq->entity;
824 struct bfq_service_tree *st;
825
826 if (!bfqq->service_from_backlogged)
827 bfqq->first_IO_time = jiffies;
828
829 if (bfqq->wr_coeff > 1)
830 bfqq->service_from_wr += served;
831
832 bfqq->service_from_backlogged += served;
833 for_each_entity(entity) {
834 st = bfq_entity_service_tree(entity);
835
836 entity->service += served;
837
838 st->vtime += bfq_delta(served, st->wsum);
839 bfq_forget_idle(st);
840 }
841 bfq_log_bfqq(bfqq->bfqd, bfqq, "bfqq_served %d secs", served);
842 }
843
844 /**
845 * bfq_bfqq_charge_time - charge an amount of service equivalent to the length
846 * of the time interval during which bfqq has been in
847 * service.
848 * @bfqd: the device
849 * @bfqq: the queue that needs a service update.
850 * @time_ms: the amount of time during which the queue has received service
851 *
852 * If a queue does not consume its budget fast enough, then providing
853 * the queue with service fairness may impair throughput, more or less
854 * severely. For this reason, queues that consume their budget slowly
855 * are provided with time fairness instead of service fairness. This
856 * goal is achieved through the BFQ scheduling engine, even if such an
857 * engine works in the service, and not in the time domain. The trick
858 * is charging these queues with an inflated amount of service, equal
859 * to the amount of service that they would have received during their
860 * service slot if they had been fast, i.e., if their requests had
861 * been dispatched at a rate equal to the estimated peak rate.
862 *
863 * It is worth noting that time fairness can cause important
864 * distortions in terms of bandwidth distribution, on devices with
865 * internal queueing. The reason is that I/O requests dispatched
866 * during the service slot of a queue may be served after that service
867 * slot is finished, and may have a total processing time loosely
868 * correlated with the duration of the service slot. This is
869 * especially true for short service slots.
870 */
bfq_bfqq_charge_time(struct bfq_data * bfqd,struct bfq_queue * bfqq,unsigned long time_ms)871 void bfq_bfqq_charge_time(struct bfq_data *bfqd, struct bfq_queue *bfqq,
872 unsigned long time_ms)
873 {
874 struct bfq_entity *entity = &bfqq->entity;
875 unsigned long timeout_ms = jiffies_to_msecs(bfq_timeout);
876 unsigned long bounded_time_ms = min(time_ms, timeout_ms);
877 int serv_to_charge_for_time =
878 (bfqd->bfq_max_budget * bounded_time_ms) / timeout_ms;
879 int tot_serv_to_charge = max(serv_to_charge_for_time, entity->service);
880
881 /* Increase budget to avoid inconsistencies */
882 if (tot_serv_to_charge > entity->budget)
883 entity->budget = tot_serv_to_charge;
884
885 bfq_bfqq_served(bfqq,
886 max_t(int, 0, tot_serv_to_charge - entity->service));
887 }
888
bfq_update_fin_time_enqueue(struct bfq_entity * entity,struct bfq_service_tree * st,bool backshifted)889 static void bfq_update_fin_time_enqueue(struct bfq_entity *entity,
890 struct bfq_service_tree *st,
891 bool backshifted)
892 {
893 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
894
895 /*
896 * When this function is invoked, entity is not in any service
897 * tree, then it is safe to invoke next function with the last
898 * parameter set (see the comments on the function).
899 */
900 st = __bfq_entity_update_weight_prio(st, entity, true);
901 bfq_calc_finish(entity, entity->budget);
902
903 /*
904 * If some queues enjoy backshifting for a while, then their
905 * (virtual) finish timestamps may happen to become lower and
906 * lower than the system virtual time. In particular, if
907 * these queues often happen to be idle for short time
908 * periods, and during such time periods other queues with
909 * higher timestamps happen to be busy, then the backshifted
910 * timestamps of the former queues can become much lower than
911 * the system virtual time. In fact, to serve the queues with
912 * higher timestamps while the ones with lower timestamps are
913 * idle, the system virtual time may be pushed-up to much
914 * higher values than the finish timestamps of the idle
915 * queues. As a consequence, the finish timestamps of all new
916 * or newly activated queues may end up being much larger than
917 * those of lucky queues with backshifted timestamps. The
918 * latter queues may then monopolize the device for a lot of
919 * time. This would simply break service guarantees.
920 *
921 * To reduce this problem, push up a little bit the
922 * backshifted timestamps of the queue associated with this
923 * entity (only a queue can happen to have the backshifted
924 * flag set): just enough to let the finish timestamp of the
925 * queue be equal to the current value of the system virtual
926 * time. This may introduce a little unfairness among queues
927 * with backshifted timestamps, but it does not break
928 * worst-case fairness guarantees.
929 *
930 * As a special case, if bfqq is weight-raised, push up
931 * timestamps much less, to keep very low the probability that
932 * this push up causes the backshifted finish timestamps of
933 * weight-raised queues to become higher than the backshifted
934 * finish timestamps of non weight-raised queues.
935 */
936 if (backshifted && bfq_gt(st->vtime, entity->finish)) {
937 unsigned long delta = st->vtime - entity->finish;
938
939 if (bfqq)
940 delta /= bfqq->wr_coeff;
941
942 entity->start += delta;
943 entity->finish += delta;
944 }
945
946 bfq_active_insert(st, entity);
947 }
948
949 /**
950 * __bfq_activate_entity - handle activation of entity.
951 * @entity: the entity being activated.
952 * @non_blocking_wait_rq: true if entity was waiting for a request
953 *
954 * Called for a 'true' activation, i.e., if entity is not active and
955 * one of its children receives a new request.
956 *
957 * Basically, this function updates the timestamps of entity and
958 * inserts entity into its active tree, after possibly extracting it
959 * from its idle tree.
960 */
__bfq_activate_entity(struct bfq_entity * entity,bool non_blocking_wait_rq)961 static void __bfq_activate_entity(struct bfq_entity *entity,
962 bool non_blocking_wait_rq)
963 {
964 struct bfq_service_tree *st = bfq_entity_service_tree(entity);
965 bool backshifted = false;
966 unsigned long long min_vstart;
967
968 /* See comments on bfq_fqq_update_budg_for_activation */
969 if (non_blocking_wait_rq && bfq_gt(st->vtime, entity->finish)) {
970 backshifted = true;
971 min_vstart = entity->finish;
972 } else
973 min_vstart = st->vtime;
974
975 if (entity->tree == &st->idle) {
976 /*
977 * Must be on the idle tree, bfq_idle_extract() will
978 * check for that.
979 */
980 bfq_idle_extract(st, entity);
981 entity->start = bfq_gt(min_vstart, entity->finish) ?
982 min_vstart : entity->finish;
983 } else {
984 /*
985 * The finish time of the entity may be invalid, and
986 * it is in the past for sure, otherwise the queue
987 * would have been on the idle tree.
988 */
989 entity->start = min_vstart;
990 st->wsum += entity->weight;
991 /*
992 * entity is about to be inserted into a service tree,
993 * and then set in service: get a reference to make
994 * sure entity does not disappear until it is no
995 * longer in service or scheduled for service.
996 */
997 bfq_get_entity(entity);
998
999 entity->on_st_or_in_serv = true;
1000 }
1001
1002 #ifdef CONFIG_BFQ_GROUP_IOSCHED
1003 if (!bfq_entity_to_bfqq(entity)) { /* bfq_group */
1004 struct bfq_group *bfqg =
1005 container_of(entity, struct bfq_group, entity);
1006 struct bfq_data *bfqd = bfqg->bfqd;
1007
1008 if (!entity->in_groups_with_pending_reqs) {
1009 entity->in_groups_with_pending_reqs = true;
1010 bfqd->num_groups_with_pending_reqs++;
1011 }
1012 }
1013 #endif
1014
1015 bfq_update_fin_time_enqueue(entity, st, backshifted);
1016 }
1017
1018 /**
1019 * __bfq_requeue_entity - handle requeueing or repositioning of an entity.
1020 * @entity: the entity being requeued or repositioned.
1021 *
1022 * Requeueing is needed if this entity stops being served, which
1023 * happens if a leaf descendant entity has expired. On the other hand,
1024 * repositioning is needed if the next_inservice_entity for the child
1025 * entity has changed. See the comments inside the function for
1026 * details.
1027 *
1028 * Basically, this function: 1) removes entity from its active tree if
1029 * present there, 2) updates the timestamps of entity and 3) inserts
1030 * entity back into its active tree (in the new, right position for
1031 * the new values of the timestamps).
1032 */
__bfq_requeue_entity(struct bfq_entity * entity)1033 static void __bfq_requeue_entity(struct bfq_entity *entity)
1034 {
1035 struct bfq_sched_data *sd = entity->sched_data;
1036 struct bfq_service_tree *st = bfq_entity_service_tree(entity);
1037
1038 if (entity == sd->in_service_entity) {
1039 /*
1040 * We are requeueing the current in-service entity,
1041 * which may have to be done for one of the following
1042 * reasons:
1043 * - entity represents the in-service queue, and the
1044 * in-service queue is being requeued after an
1045 * expiration;
1046 * - entity represents a group, and its budget has
1047 * changed because one of its child entities has
1048 * just been either activated or requeued for some
1049 * reason; the timestamps of the entity need then to
1050 * be updated, and the entity needs to be enqueued
1051 * or repositioned accordingly.
1052 *
1053 * In particular, before requeueing, the start time of
1054 * the entity must be moved forward to account for the
1055 * service that the entity has received while in
1056 * service. This is done by the next instructions. The
1057 * finish time will then be updated according to this
1058 * new value of the start time, and to the budget of
1059 * the entity.
1060 */
1061 bfq_calc_finish(entity, entity->service);
1062 entity->start = entity->finish;
1063 /*
1064 * In addition, if the entity had more than one child
1065 * when set in service, then it was not extracted from
1066 * the active tree. This implies that the position of
1067 * the entity in the active tree may need to be
1068 * changed now, because we have just updated the start
1069 * time of the entity, and we will update its finish
1070 * time in a moment (the requeueing is then, more
1071 * precisely, a repositioning in this case). To
1072 * implement this repositioning, we: 1) dequeue the
1073 * entity here, 2) update the finish time and requeue
1074 * the entity according to the new timestamps below.
1075 */
1076 if (entity->tree)
1077 bfq_active_extract(st, entity);
1078 } else { /* The entity is already active, and not in service */
1079 /*
1080 * In this case, this function gets called only if the
1081 * next_in_service entity below this entity has
1082 * changed, and this change has caused the budget of
1083 * this entity to change, which, finally implies that
1084 * the finish time of this entity must be
1085 * updated. Such an update may cause the scheduling,
1086 * i.e., the position in the active tree, of this
1087 * entity to change. We handle this change by: 1)
1088 * dequeueing the entity here, 2) updating the finish
1089 * time and requeueing the entity according to the new
1090 * timestamps below. This is the same approach as the
1091 * non-extracted-entity sub-case above.
1092 */
1093 bfq_active_extract(st, entity);
1094 }
1095
1096 bfq_update_fin_time_enqueue(entity, st, false);
1097 }
1098
__bfq_activate_requeue_entity(struct bfq_entity * entity,struct bfq_sched_data * sd,bool non_blocking_wait_rq)1099 static void __bfq_activate_requeue_entity(struct bfq_entity *entity,
1100 struct bfq_sched_data *sd,
1101 bool non_blocking_wait_rq)
1102 {
1103 struct bfq_service_tree *st = bfq_entity_service_tree(entity);
1104
1105 if (sd->in_service_entity == entity || entity->tree == &st->active)
1106 /*
1107 * in service or already queued on the active tree,
1108 * requeue or reposition
1109 */
1110 __bfq_requeue_entity(entity);
1111 else
1112 /*
1113 * Not in service and not queued on its active tree:
1114 * the activity is idle and this is a true activation.
1115 */
1116 __bfq_activate_entity(entity, non_blocking_wait_rq);
1117 }
1118
1119
1120 /**
1121 * bfq_activate_requeue_entity - activate or requeue an entity representing a
1122 * bfq_queue, and activate, requeue or reposition
1123 * all ancestors for which such an update becomes
1124 * necessary.
1125 * @entity: the entity to activate.
1126 * @non_blocking_wait_rq: true if this entity was waiting for a request
1127 * @requeue: true if this is a requeue, which implies that bfqq is
1128 * being expired; thus ALL its ancestors stop being served and must
1129 * therefore be requeued
1130 * @expiration: true if this function is being invoked in the expiration path
1131 * of the in-service queue
1132 */
bfq_activate_requeue_entity(struct bfq_entity * entity,bool non_blocking_wait_rq,bool requeue,bool expiration)1133 static void bfq_activate_requeue_entity(struct bfq_entity *entity,
1134 bool non_blocking_wait_rq,
1135 bool requeue, bool expiration)
1136 {
1137 struct bfq_sched_data *sd;
1138
1139 for_each_entity(entity) {
1140 sd = entity->sched_data;
1141 __bfq_activate_requeue_entity(entity, sd, non_blocking_wait_rq);
1142
1143 if (!bfq_update_next_in_service(sd, entity, expiration) &&
1144 !requeue)
1145 break;
1146 }
1147 }
1148
1149 /**
1150 * __bfq_deactivate_entity - update sched_data and service trees for
1151 * entity, so as to represent entity as inactive
1152 * @entity: the entity being deactivated.
1153 * @ins_into_idle_tree: if false, the entity will not be put into the
1154 * idle tree.
1155 *
1156 * If necessary and allowed, puts entity into the idle tree. NOTE:
1157 * entity may be on no tree if in service.
1158 */
__bfq_deactivate_entity(struct bfq_entity * entity,bool ins_into_idle_tree)1159 bool __bfq_deactivate_entity(struct bfq_entity *entity, bool ins_into_idle_tree)
1160 {
1161 struct bfq_sched_data *sd = entity->sched_data;
1162 struct bfq_service_tree *st;
1163 bool is_in_service;
1164
1165 if (!entity->on_st_or_in_serv) /*
1166 * entity never activated, or
1167 * already inactive
1168 */
1169 return false;
1170
1171 /*
1172 * If we get here, then entity is active, which implies that
1173 * bfq_group_set_parent has already been invoked for the group
1174 * represented by entity. Therefore, the field
1175 * entity->sched_data has been set, and we can safely use it.
1176 */
1177 st = bfq_entity_service_tree(entity);
1178 is_in_service = entity == sd->in_service_entity;
1179
1180 bfq_calc_finish(entity, entity->service);
1181
1182 if (is_in_service)
1183 sd->in_service_entity = NULL;
1184 else
1185 /*
1186 * Non in-service entity: nobody will take care of
1187 * resetting its service counter on expiration. Do it
1188 * now.
1189 */
1190 entity->service = 0;
1191
1192 if (entity->tree == &st->active)
1193 bfq_active_extract(st, entity);
1194 else if (!is_in_service && entity->tree == &st->idle)
1195 bfq_idle_extract(st, entity);
1196
1197 if (!ins_into_idle_tree || !bfq_gt(entity->finish, st->vtime))
1198 bfq_forget_entity(st, entity, is_in_service);
1199 else
1200 bfq_idle_insert(st, entity);
1201
1202 return true;
1203 }
1204
1205 /**
1206 * bfq_deactivate_entity - deactivate an entity representing a bfq_queue.
1207 * @entity: the entity to deactivate.
1208 * @ins_into_idle_tree: true if the entity can be put into the idle tree
1209 * @expiration: true if this function is being invoked in the expiration path
1210 * of the in-service queue
1211 */
bfq_deactivate_entity(struct bfq_entity * entity,bool ins_into_idle_tree,bool expiration)1212 static void bfq_deactivate_entity(struct bfq_entity *entity,
1213 bool ins_into_idle_tree,
1214 bool expiration)
1215 {
1216 struct bfq_sched_data *sd;
1217 struct bfq_entity *parent = NULL;
1218
1219 for_each_entity_safe(entity, parent) {
1220 sd = entity->sched_data;
1221
1222 if (!__bfq_deactivate_entity(entity, ins_into_idle_tree)) {
1223 /*
1224 * entity is not in any tree any more, so
1225 * this deactivation is a no-op, and there is
1226 * nothing to change for upper-level entities
1227 * (in case of expiration, this can never
1228 * happen).
1229 */
1230 return;
1231 }
1232
1233 if (sd->next_in_service == entity)
1234 /*
1235 * entity was the next_in_service entity,
1236 * then, since entity has just been
1237 * deactivated, a new one must be found.
1238 */
1239 bfq_update_next_in_service(sd, NULL, expiration);
1240
1241 if (sd->next_in_service || sd->in_service_entity) {
1242 /*
1243 * The parent entity is still active, because
1244 * either next_in_service or in_service_entity
1245 * is not NULL. So, no further upwards
1246 * deactivation must be performed. Yet,
1247 * next_in_service has changed. Then the
1248 * schedule does need to be updated upwards.
1249 *
1250 * NOTE If in_service_entity is not NULL, then
1251 * next_in_service may happen to be NULL,
1252 * although the parent entity is evidently
1253 * active. This happens if 1) the entity
1254 * pointed by in_service_entity is the only
1255 * active entity in the parent entity, and 2)
1256 * according to the definition of
1257 * next_in_service, the in_service_entity
1258 * cannot be considered as
1259 * next_in_service. See the comments on the
1260 * definition of next_in_service for details.
1261 */
1262 break;
1263 }
1264
1265 /*
1266 * If we get here, then the parent is no more
1267 * backlogged and we need to propagate the
1268 * deactivation upwards. Thus let the loop go on.
1269 */
1270
1271 /*
1272 * Also let parent be queued into the idle tree on
1273 * deactivation, to preserve service guarantees, and
1274 * assuming that who invoked this function does not
1275 * need parent entities too to be removed completely.
1276 */
1277 ins_into_idle_tree = true;
1278 }
1279
1280 /*
1281 * If the deactivation loop is fully executed, then there are
1282 * no more entities to touch and next loop is not executed at
1283 * all. Otherwise, requeue remaining entities if they are
1284 * about to stop receiving service, or reposition them if this
1285 * is not the case.
1286 */
1287 entity = parent;
1288 for_each_entity(entity) {
1289 /*
1290 * Invoke __bfq_requeue_entity on entity, even if
1291 * already active, to requeue/reposition it in the
1292 * active tree (because sd->next_in_service has
1293 * changed)
1294 */
1295 __bfq_requeue_entity(entity);
1296
1297 sd = entity->sched_data;
1298 if (!bfq_update_next_in_service(sd, entity, expiration) &&
1299 !expiration)
1300 /*
1301 * next_in_service unchanged or not causing
1302 * any change in entity->parent->sd, and no
1303 * requeueing needed for expiration: stop
1304 * here.
1305 */
1306 break;
1307 }
1308 }
1309
1310 /**
1311 * bfq_calc_vtime_jump - compute the value to which the vtime should jump,
1312 * if needed, to have at least one entity eligible.
1313 * @st: the service tree to act upon.
1314 *
1315 * Assumes that st is not empty.
1316 */
bfq_calc_vtime_jump(struct bfq_service_tree * st)1317 static u64 bfq_calc_vtime_jump(struct bfq_service_tree *st)
1318 {
1319 struct bfq_entity *root_entity = bfq_root_active_entity(&st->active);
1320
1321 if (bfq_gt(root_entity->min_start, st->vtime))
1322 return root_entity->min_start;
1323
1324 return st->vtime;
1325 }
1326
bfq_update_vtime(struct bfq_service_tree * st,u64 new_value)1327 static void bfq_update_vtime(struct bfq_service_tree *st, u64 new_value)
1328 {
1329 if (new_value > st->vtime) {
1330 st->vtime = new_value;
1331 bfq_forget_idle(st);
1332 }
1333 }
1334
1335 /**
1336 * bfq_first_active_entity - find the eligible entity with
1337 * the smallest finish time
1338 * @st: the service tree to select from.
1339 * @vtime: the system virtual to use as a reference for eligibility
1340 *
1341 * This function searches the first schedulable entity, starting from the
1342 * root of the tree and going on the left every time on this side there is
1343 * a subtree with at least one eligible (start <= vtime) entity. The path on
1344 * the right is followed only if a) the left subtree contains no eligible
1345 * entities and b) no eligible entity has been found yet.
1346 */
bfq_first_active_entity(struct bfq_service_tree * st,u64 vtime)1347 static struct bfq_entity *bfq_first_active_entity(struct bfq_service_tree *st,
1348 u64 vtime)
1349 {
1350 struct bfq_entity *entry, *first = NULL;
1351 struct rb_node *node = st->active.rb_node;
1352
1353 while (node) {
1354 entry = rb_entry(node, struct bfq_entity, rb_node);
1355 left:
1356 if (!bfq_gt(entry->start, vtime))
1357 first = entry;
1358
1359 if (node->rb_left) {
1360 entry = rb_entry(node->rb_left,
1361 struct bfq_entity, rb_node);
1362 if (!bfq_gt(entry->min_start, vtime)) {
1363 node = node->rb_left;
1364 goto left;
1365 }
1366 }
1367 if (first)
1368 break;
1369 node = node->rb_right;
1370 }
1371
1372 return first;
1373 }
1374
1375 /**
1376 * __bfq_lookup_next_entity - return the first eligible entity in @st.
1377 * @st: the service tree.
1378 *
1379 * If there is no in-service entity for the sched_data st belongs to,
1380 * then return the entity that will be set in service if:
1381 * 1) the parent entity this st belongs to is set in service;
1382 * 2) no entity belonging to such parent entity undergoes a state change
1383 * that would influence the timestamps of the entity (e.g., becomes idle,
1384 * becomes backlogged, changes its budget, ...).
1385 *
1386 * In this first case, update the virtual time in @st too (see the
1387 * comments on this update inside the function).
1388 *
1389 * In contrast, if there is an in-service entity, then return the
1390 * entity that would be set in service if not only the above
1391 * conditions, but also the next one held true: the currently
1392 * in-service entity, on expiration,
1393 * 1) gets a finish time equal to the current one, or
1394 * 2) is not eligible any more, or
1395 * 3) is idle.
1396 */
1397 static struct bfq_entity *
__bfq_lookup_next_entity(struct bfq_service_tree * st,bool in_service)1398 __bfq_lookup_next_entity(struct bfq_service_tree *st, bool in_service)
1399 {
1400 struct bfq_entity *entity;
1401 u64 new_vtime;
1402
1403 if (RB_EMPTY_ROOT(&st->active))
1404 return NULL;
1405
1406 /*
1407 * Get the value of the system virtual time for which at
1408 * least one entity is eligible.
1409 */
1410 new_vtime = bfq_calc_vtime_jump(st);
1411
1412 /*
1413 * If there is no in-service entity for the sched_data this
1414 * active tree belongs to, then push the system virtual time
1415 * up to the value that guarantees that at least one entity is
1416 * eligible. If, instead, there is an in-service entity, then
1417 * do not make any such update, because there is already an
1418 * eligible entity, namely the in-service one (even if the
1419 * entity is not on st, because it was extracted when set in
1420 * service).
1421 */
1422 if (!in_service)
1423 bfq_update_vtime(st, new_vtime);
1424
1425 entity = bfq_first_active_entity(st, new_vtime);
1426
1427 return entity;
1428 }
1429
1430 /**
1431 * bfq_lookup_next_entity - return the first eligible entity in @sd.
1432 * @sd: the sched_data.
1433 * @expiration: true if we are on the expiration path of the in-service queue
1434 *
1435 * This function is invoked when there has been a change in the trees
1436 * for sd, and we need to know what is the new next entity to serve
1437 * after this change.
1438 */
bfq_lookup_next_entity(struct bfq_sched_data * sd,bool expiration)1439 static struct bfq_entity *bfq_lookup_next_entity(struct bfq_sched_data *sd,
1440 bool expiration)
1441 {
1442 struct bfq_service_tree *st = sd->service_tree;
1443 struct bfq_service_tree *idle_class_st = st + (BFQ_IOPRIO_CLASSES - 1);
1444 struct bfq_entity *entity = NULL;
1445 int class_idx = 0;
1446
1447 /*
1448 * Choose from idle class, if needed to guarantee a minimum
1449 * bandwidth to this class (and if there is some active entity
1450 * in idle class). This should also mitigate
1451 * priority-inversion problems in case a low priority task is
1452 * holding file system resources.
1453 */
1454 if (time_is_before_jiffies(sd->bfq_class_idle_last_service +
1455 BFQ_CL_IDLE_TIMEOUT)) {
1456 if (!RB_EMPTY_ROOT(&idle_class_st->active))
1457 class_idx = BFQ_IOPRIO_CLASSES - 1;
1458 /* About to be served if backlogged, or not yet backlogged */
1459 sd->bfq_class_idle_last_service = jiffies;
1460 }
1461
1462 /*
1463 * Find the next entity to serve for the highest-priority
1464 * class, unless the idle class needs to be served.
1465 */
1466 for (; class_idx < BFQ_IOPRIO_CLASSES; class_idx++) {
1467 /*
1468 * If expiration is true, then bfq_lookup_next_entity
1469 * is being invoked as a part of the expiration path
1470 * of the in-service queue. In this case, even if
1471 * sd->in_service_entity is not NULL,
1472 * sd->in_service_entity at this point is actually not
1473 * in service any more, and, if needed, has already
1474 * been properly queued or requeued into the right
1475 * tree. The reason why sd->in_service_entity is still
1476 * not NULL here, even if expiration is true, is that
1477 * sd->in_service_entity is reset as a last step in the
1478 * expiration path. So, if expiration is true, tell
1479 * __bfq_lookup_next_entity that there is no
1480 * sd->in_service_entity.
1481 */
1482 entity = __bfq_lookup_next_entity(st + class_idx,
1483 sd->in_service_entity &&
1484 !expiration);
1485
1486 if (entity)
1487 break;
1488 }
1489
1490 if (!entity)
1491 return NULL;
1492
1493 return entity;
1494 }
1495
next_queue_may_preempt(struct bfq_data * bfqd)1496 bool next_queue_may_preempt(struct bfq_data *bfqd)
1497 {
1498 struct bfq_sched_data *sd = &bfqd->root_group->sched_data;
1499
1500 return sd->next_in_service != sd->in_service_entity;
1501 }
1502
1503 /*
1504 * Get next queue for service.
1505 */
bfq_get_next_queue(struct bfq_data * bfqd)1506 struct bfq_queue *bfq_get_next_queue(struct bfq_data *bfqd)
1507 {
1508 struct bfq_entity *entity = NULL;
1509 struct bfq_sched_data *sd;
1510 struct bfq_queue *bfqq;
1511
1512 if (bfq_tot_busy_queues(bfqd) == 0)
1513 return NULL;
1514
1515 /*
1516 * Traverse the path from the root to the leaf entity to
1517 * serve. Set in service all the entities visited along the
1518 * way.
1519 */
1520 sd = &bfqd->root_group->sched_data;
1521 for (; sd ; sd = entity->my_sched_data) {
1522 /*
1523 * WARNING. We are about to set the in-service entity
1524 * to sd->next_in_service, i.e., to the (cached) value
1525 * returned by bfq_lookup_next_entity(sd) the last
1526 * time it was invoked, i.e., the last time when the
1527 * service order in sd changed as a consequence of the
1528 * activation or deactivation of an entity. In this
1529 * respect, if we execute bfq_lookup_next_entity(sd)
1530 * in this very moment, it may, although with low
1531 * probability, yield a different entity than that
1532 * pointed to by sd->next_in_service. This rare event
1533 * happens in case there was no CLASS_IDLE entity to
1534 * serve for sd when bfq_lookup_next_entity(sd) was
1535 * invoked for the last time, while there is now one
1536 * such entity.
1537 *
1538 * If the above event happens, then the scheduling of
1539 * such entity in CLASS_IDLE is postponed until the
1540 * service of the sd->next_in_service entity
1541 * finishes. In fact, when the latter is expired,
1542 * bfq_lookup_next_entity(sd) gets called again,
1543 * exactly to update sd->next_in_service.
1544 */
1545
1546 /* Make next_in_service entity become in_service_entity */
1547 entity = sd->next_in_service;
1548 sd->in_service_entity = entity;
1549
1550 /*
1551 * If entity is no longer a candidate for next
1552 * service, then it must be extracted from its active
1553 * tree, so as to make sure that it won't be
1554 * considered when computing next_in_service. See the
1555 * comments on the function
1556 * bfq_no_longer_next_in_service() for details.
1557 */
1558 if (bfq_no_longer_next_in_service(entity))
1559 bfq_active_extract(bfq_entity_service_tree(entity),
1560 entity);
1561
1562 /*
1563 * Even if entity is not to be extracted according to
1564 * the above check, a descendant entity may get
1565 * extracted in one of the next iterations of this
1566 * loop. Such an event could cause a change in
1567 * next_in_service for the level of the descendant
1568 * entity, and thus possibly back to this level.
1569 *
1570 * However, we cannot perform the resulting needed
1571 * update of next_in_service for this level before the
1572 * end of the whole loop, because, to know which is
1573 * the correct next-to-serve candidate entity for each
1574 * level, we need first to find the leaf entity to set
1575 * in service. In fact, only after we know which is
1576 * the next-to-serve leaf entity, we can discover
1577 * whether the parent entity of the leaf entity
1578 * becomes the next-to-serve, and so on.
1579 */
1580 }
1581
1582 bfqq = bfq_entity_to_bfqq(entity);
1583
1584 /*
1585 * We can finally update all next-to-serve entities along the
1586 * path from the leaf entity just set in service to the root.
1587 */
1588 for_each_entity(entity) {
1589 struct bfq_sched_data *sd = entity->sched_data;
1590
1591 if (!bfq_update_next_in_service(sd, NULL, false))
1592 break;
1593 }
1594
1595 return bfqq;
1596 }
1597
1598 /* returns true if the in-service queue gets freed */
__bfq_bfqd_reset_in_service(struct bfq_data * bfqd)1599 bool __bfq_bfqd_reset_in_service(struct bfq_data *bfqd)
1600 {
1601 struct bfq_queue *in_serv_bfqq = bfqd->in_service_queue;
1602 struct bfq_entity *in_serv_entity = &in_serv_bfqq->entity;
1603 struct bfq_entity *entity = in_serv_entity;
1604
1605 bfq_clear_bfqq_wait_request(in_serv_bfqq);
1606 hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
1607 bfqd->in_service_queue = NULL;
1608
1609 /*
1610 * When this function is called, all in-service entities have
1611 * been properly deactivated or requeued, so we can safely
1612 * execute the final step: reset in_service_entity along the
1613 * path from entity to the root.
1614 */
1615 for_each_entity(entity)
1616 entity->sched_data->in_service_entity = NULL;
1617
1618 /*
1619 * in_serv_entity is no longer in service, so, if it is in no
1620 * service tree either, then release the service reference to
1621 * the queue it represents (taken with bfq_get_entity).
1622 */
1623 if (!in_serv_entity->on_st_or_in_serv) {
1624 /*
1625 * If no process is referencing in_serv_bfqq any
1626 * longer, then the service reference may be the only
1627 * reference to the queue. If this is the case, then
1628 * bfqq gets freed here.
1629 */
1630 int ref = in_serv_bfqq->ref;
1631 bfq_put_queue(in_serv_bfqq);
1632 if (ref == 1)
1633 return true;
1634 }
1635
1636 return false;
1637 }
1638
bfq_deactivate_bfqq(struct bfq_data * bfqd,struct bfq_queue * bfqq,bool ins_into_idle_tree,bool expiration)1639 void bfq_deactivate_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1640 bool ins_into_idle_tree, bool expiration)
1641 {
1642 struct bfq_entity *entity = &bfqq->entity;
1643
1644 bfq_deactivate_entity(entity, ins_into_idle_tree, expiration);
1645 }
1646
bfq_activate_bfqq(struct bfq_data * bfqd,struct bfq_queue * bfqq)1647 void bfq_activate_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1648 {
1649 struct bfq_entity *entity = &bfqq->entity;
1650
1651 bfq_activate_requeue_entity(entity, bfq_bfqq_non_blocking_wait_rq(bfqq),
1652 false, false);
1653 bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1654 }
1655
bfq_requeue_bfqq(struct bfq_data * bfqd,struct bfq_queue * bfqq,bool expiration)1656 void bfq_requeue_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1657 bool expiration)
1658 {
1659 struct bfq_entity *entity = &bfqq->entity;
1660
1661 bfq_activate_requeue_entity(entity, false,
1662 bfqq == bfqd->in_service_queue, expiration);
1663 }
1664
1665 /*
1666 * Called when the bfqq no longer has requests pending, remove it from
1667 * the service tree. As a special case, it can be invoked during an
1668 * expiration.
1669 */
bfq_del_bfqq_busy(struct bfq_data * bfqd,struct bfq_queue * bfqq,bool expiration)1670 void bfq_del_bfqq_busy(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1671 bool expiration)
1672 {
1673 bfq_log_bfqq(bfqd, bfqq, "del from busy");
1674
1675 bfq_clear_bfqq_busy(bfqq);
1676
1677 bfqd->busy_queues[bfqq->ioprio_class - 1]--;
1678
1679 if (bfqq->wr_coeff > 1)
1680 bfqd->wr_busy_queues--;
1681
1682 bfqg_stats_update_dequeue(bfqq_group(bfqq));
1683
1684 bfq_deactivate_bfqq(bfqd, bfqq, true, expiration);
1685
1686 if (!bfqq->dispatched)
1687 bfq_weights_tree_remove(bfqd, bfqq);
1688 }
1689
1690 /*
1691 * Called when an inactive queue receives a new request.
1692 */
bfq_add_bfqq_busy(struct bfq_data * bfqd,struct bfq_queue * bfqq)1693 void bfq_add_bfqq_busy(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1694 {
1695 bfq_log_bfqq(bfqd, bfqq, "add to busy");
1696
1697 bfq_activate_bfqq(bfqd, bfqq);
1698
1699 bfq_mark_bfqq_busy(bfqq);
1700 bfqd->busy_queues[bfqq->ioprio_class - 1]++;
1701
1702 if (!bfqq->dispatched)
1703 if (bfqq->wr_coeff == 1)
1704 bfq_weights_tree_add(bfqd, bfqq,
1705 &bfqd->queue_weights_tree);
1706
1707 if (bfqq->wr_coeff > 1)
1708 bfqd->wr_busy_queues++;
1709
1710 /* Move bfqq to the head of the woken list of its waker */
1711 if (!hlist_unhashed(&bfqq->woken_list_node) &&
1712 &bfqq->woken_list_node != bfqq->waker_bfqq->woken_list.first) {
1713 hlist_del_init(&bfqq->woken_list_node);
1714 hlist_add_head(&bfqq->woken_list_node,
1715 &bfqq->waker_bfqq->woken_list);
1716 }
1717 }
1718