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