1/*
2 * cgroups support for the BFQ I/O scheduler.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License as
6 * published by the Free Software Foundation; either version 2 of the
7 * License, or (at your option) any later version.
8 *
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * General Public License for more details.
13 */
14#include <linux/module.h>
15#include <linux/slab.h>
16#include <linux/blkdev.h>
17#include <linux/cgroup.h>
18#include <linux/elevator.h>
19#include <linux/ktime.h>
20#include <linux/rbtree.h>
21#include <linux/ioprio.h>
22#include <linux/sbitmap.h>
23#include <linux/delay.h>
24
25#include "bfq-iosched.h"
26
27#if defined(CONFIG_BFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
28
29/* bfqg stats flags */
30enum bfqg_stats_flags {
31 BFQG_stats_waiting = 0,
32 BFQG_stats_idling,
33 BFQG_stats_empty,
34};
35
36#define BFQG_FLAG_FNS(name) \
37static void bfqg_stats_mark_##name(struct bfqg_stats *stats) \
38{ \
39 stats->flags |= (1 << BFQG_stats_##name); \
40} \
41static void bfqg_stats_clear_##name(struct bfqg_stats *stats) \
42{ \
43 stats->flags &= ~(1 << BFQG_stats_##name); \
44} \
45static int bfqg_stats_##name(struct bfqg_stats *stats) \
46{ \
47 return (stats->flags & (1 << BFQG_stats_##name)) != 0; \
48} \
49
50BFQG_FLAG_FNS(waiting)
51BFQG_FLAG_FNS(idling)
52BFQG_FLAG_FNS(empty)
53#undef BFQG_FLAG_FNS
54
55/* This should be called with the scheduler lock held. */
56static void bfqg_stats_update_group_wait_time(struct bfqg_stats *stats)
57{
58 u64 now;
59
60 if (!bfqg_stats_waiting(stats))
61 return;
62
63 now = ktime_get_ns();
64 if (now > stats->start_group_wait_time)
65 blkg_stat_add(&stats->group_wait_time,
66 now - stats->start_group_wait_time);
67 bfqg_stats_clear_waiting(stats);
68}
69
70/* This should be called with the scheduler lock held. */
71static void bfqg_stats_set_start_group_wait_time(struct bfq_group *bfqg,
72 struct bfq_group *curr_bfqg)
73{
74 struct bfqg_stats *stats = &bfqg->stats;
75
76 if (bfqg_stats_waiting(stats))
77 return;
78 if (bfqg == curr_bfqg)
79 return;
80 stats->start_group_wait_time = ktime_get_ns();
81 bfqg_stats_mark_waiting(stats);
82}
83
84/* This should be called with the scheduler lock held. */
85static void bfqg_stats_end_empty_time(struct bfqg_stats *stats)
86{
87 u64 now;
88
89 if (!bfqg_stats_empty(stats))
90 return;
91
92 now = ktime_get_ns();
93 if (now > stats->start_empty_time)
94 blkg_stat_add(&stats->empty_time,
95 now - stats->start_empty_time);
96 bfqg_stats_clear_empty(stats);
97}
98
99void bfqg_stats_update_dequeue(struct bfq_group *bfqg)
100{
101 blkg_stat_add(&bfqg->stats.dequeue, 1);
102}
103
104void bfqg_stats_set_start_empty_time(struct bfq_group *bfqg)
105{
106 struct bfqg_stats *stats = &bfqg->stats;
107
108 if (blkg_rwstat_total(&stats->queued))
109 return;
110
111 /*
112 * group is already marked empty. This can happen if bfqq got new
113 * request in parent group and moved to this group while being added
114 * to service tree. Just ignore the event and move on.
115 */
116 if (bfqg_stats_empty(stats))
117 return;
118
119 stats->start_empty_time = ktime_get_ns();
120 bfqg_stats_mark_empty(stats);
121}
122
123void bfqg_stats_update_idle_time(struct bfq_group *bfqg)
124{
125 struct bfqg_stats *stats = &bfqg->stats;
126
127 if (bfqg_stats_idling(stats)) {
128 u64 now = ktime_get_ns();
129
130 if (now > stats->start_idle_time)
131 blkg_stat_add(&stats->idle_time,
132 now - stats->start_idle_time);
133 bfqg_stats_clear_idling(stats);
134 }
135}
136
137void bfqg_stats_set_start_idle_time(struct bfq_group *bfqg)
138{
139 struct bfqg_stats *stats = &bfqg->stats;
140
141 stats->start_idle_time = ktime_get_ns();
142 bfqg_stats_mark_idling(stats);
143}
144
145void bfqg_stats_update_avg_queue_size(struct bfq_group *bfqg)
146{
147 struct bfqg_stats *stats = &bfqg->stats;
148
149 blkg_stat_add(&stats->avg_queue_size_sum,
150 blkg_rwstat_total(&stats->queued));
151 blkg_stat_add(&stats->avg_queue_size_samples, 1);
152 bfqg_stats_update_group_wait_time(stats);
153}
154
155void bfqg_stats_update_io_add(struct bfq_group *bfqg, struct bfq_queue *bfqq,
156 unsigned int op)
157{
158 blkg_rwstat_add(&bfqg->stats.queued, op, 1);
159 bfqg_stats_end_empty_time(&bfqg->stats);
160 if (!(bfqq == ((struct bfq_data *)bfqg->bfqd)->in_service_queue))
161 bfqg_stats_set_start_group_wait_time(bfqg, bfqq_group(bfqq));
162}
163
164void bfqg_stats_update_io_remove(struct bfq_group *bfqg, unsigned int op)
165{
166 blkg_rwstat_add(&bfqg->stats.queued, op, -1);
167}
168
169void bfqg_stats_update_io_merged(struct bfq_group *bfqg, unsigned int op)
170{
171 blkg_rwstat_add(&bfqg->stats.merged, op, 1);
172}
173
174void bfqg_stats_update_completion(struct bfq_group *bfqg, u64 start_time_ns,
175 u64 io_start_time_ns, unsigned int op)
176{
177 struct bfqg_stats *stats = &bfqg->stats;
178 u64 now = ktime_get_ns();
179
180 if (now > io_start_time_ns)
181 blkg_rwstat_add(&stats->service_time, op,
182 now - io_start_time_ns);
183 if (io_start_time_ns > start_time_ns)
184 blkg_rwstat_add(&stats->wait_time, op,
185 io_start_time_ns - start_time_ns);
186}
187
188#else /* CONFIG_BFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
189
190void bfqg_stats_update_io_add(struct bfq_group *bfqg, struct bfq_queue *bfqq,
191 unsigned int op) { }
192void bfqg_stats_update_io_remove(struct bfq_group *bfqg, unsigned int op) { }
193void bfqg_stats_update_io_merged(struct bfq_group *bfqg, unsigned int op) { }
194void bfqg_stats_update_completion(struct bfq_group *bfqg, u64 start_time_ns,
195 u64 io_start_time_ns, unsigned int op) { }
196void bfqg_stats_update_dequeue(struct bfq_group *bfqg) { }
197void bfqg_stats_set_start_empty_time(struct bfq_group *bfqg) { }
198void bfqg_stats_update_idle_time(struct bfq_group *bfqg) { }
199void bfqg_stats_set_start_idle_time(struct bfq_group *bfqg) { }
200void bfqg_stats_update_avg_queue_size(struct bfq_group *bfqg) { }
201
202#endif /* CONFIG_BFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
203
204#ifdef CONFIG_BFQ_GROUP_IOSCHED
205
206/*
207 * blk-cgroup policy-related handlers
208 * The following functions help in converting between blk-cgroup
209 * internal structures and BFQ-specific structures.
210 */
211
212static struct bfq_group *pd_to_bfqg(struct blkg_policy_data *pd)
213{
214 return pd ? container_of(pd, struct bfq_group, pd) : NULL;
215}
216
217struct blkcg_gq *bfqg_to_blkg(struct bfq_group *bfqg)
218{
219 return pd_to_blkg(&bfqg->pd);
220}
221
222static struct bfq_group *blkg_to_bfqg(struct blkcg_gq *blkg)
223{
224 return pd_to_bfqg(blkg_to_pd(blkg, &blkcg_policy_bfq));
225}
226
227/*
228 * bfq_group handlers
229 * The following functions help in navigating the bfq_group hierarchy
230 * by allowing to find the parent of a bfq_group or the bfq_group
231 * associated to a bfq_queue.
232 */
233
234static struct bfq_group *bfqg_parent(struct bfq_group *bfqg)
235{
236 struct blkcg_gq *pblkg = bfqg_to_blkg(bfqg)->parent;
237
238 return pblkg ? blkg_to_bfqg(pblkg) : NULL;
239}
240
241struct bfq_group *bfqq_group(struct bfq_queue *bfqq)
242{
243 struct bfq_entity *group_entity = bfqq->entity.parent;
244
245 return group_entity ? container_of(group_entity, struct bfq_group,
246 entity) :
247 bfqq->bfqd->root_group;
248}
249
250/*
251 * The following two functions handle get and put of a bfq_group by
252 * wrapping the related blk-cgroup hooks.
253 */
254
255static void bfqg_get(struct bfq_group *bfqg)
256{
257 bfqg->ref++;
258}
259
260static void bfqg_put(struct bfq_group *bfqg)
261{
262 bfqg->ref--;
263
264 if (bfqg->ref == 0)
265 kfree(bfqg);
266}
267
268static void bfqg_and_blkg_get(struct bfq_group *bfqg)
269{
270 /* see comments in bfq_bic_update_cgroup for why refcounting bfqg */
271 bfqg_get(bfqg);
272
273 blkg_get(bfqg_to_blkg(bfqg));
274}
275
276void bfqg_and_blkg_put(struct bfq_group *bfqg)
277{
278 blkg_put(bfqg_to_blkg(bfqg));
279
280 bfqg_put(bfqg);
281}
282
283/* @stats = 0 */
284static void bfqg_stats_reset(struct bfqg_stats *stats)
285{
286#ifdef CONFIG_DEBUG_BLK_CGROUP
287 /* queued stats shouldn't be cleared */
288 blkg_rwstat_reset(&stats->merged);
289 blkg_rwstat_reset(&stats->service_time);
290 blkg_rwstat_reset(&stats->wait_time);
291 blkg_stat_reset(&stats->time);
292 blkg_stat_reset(&stats->avg_queue_size_sum);
293 blkg_stat_reset(&stats->avg_queue_size_samples);
294 blkg_stat_reset(&stats->dequeue);
295 blkg_stat_reset(&stats->group_wait_time);
296 blkg_stat_reset(&stats->idle_time);
297 blkg_stat_reset(&stats->empty_time);
298#endif
299}
300
301/* @to += @from */
302static void bfqg_stats_add_aux(struct bfqg_stats *to, struct bfqg_stats *from)
303{
304 if (!to || !from)
305 return;
306
307#ifdef CONFIG_DEBUG_BLK_CGROUP
308 /* queued stats shouldn't be cleared */
309 blkg_rwstat_add_aux(&to->merged, &from->merged);
310 blkg_rwstat_add_aux(&to->service_time, &from->service_time);
311 blkg_rwstat_add_aux(&to->wait_time, &from->wait_time);
312 blkg_stat_add_aux(&from->time, &from->time);
313 blkg_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
314 blkg_stat_add_aux(&to->avg_queue_size_samples,
315 &from->avg_queue_size_samples);
316 blkg_stat_add_aux(&to->dequeue, &from->dequeue);
317 blkg_stat_add_aux(&to->group_wait_time, &from->group_wait_time);
318 blkg_stat_add_aux(&to->idle_time, &from->idle_time);
319 blkg_stat_add_aux(&to->empty_time, &from->empty_time);
320#endif
321}
322
323/*
324 * Transfer @bfqg's stats to its parent's aux counts so that the ancestors'
325 * recursive stats can still account for the amount used by this bfqg after
326 * it's gone.
327 */
328static void bfqg_stats_xfer_dead(struct bfq_group *bfqg)
329{
330 struct bfq_group *parent;
331
332 if (!bfqg) /* root_group */
333 return;
334
335 parent = bfqg_parent(bfqg);
336
337 lockdep_assert_held(&bfqg_to_blkg(bfqg)->q->queue_lock);
338
339 if (unlikely(!parent))
340 return;
341
342 bfqg_stats_add_aux(&parent->stats, &bfqg->stats);
343 bfqg_stats_reset(&bfqg->stats);
344}
345
346void bfq_init_entity(struct bfq_entity *entity, struct bfq_group *bfqg)
347{
348 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
349
350 entity->weight = entity->new_weight;
351 entity->orig_weight = entity->new_weight;
352 if (bfqq) {
353 bfqq->ioprio = bfqq->new_ioprio;
354 bfqq->ioprio_class = bfqq->new_ioprio_class;
355 /*
356 * Make sure that bfqg and its associated blkg do not
357 * disappear before entity.
358 */
359 bfqg_and_blkg_get(bfqg);
360 }
361 entity->parent = bfqg->my_entity; /* NULL for root group */
362 entity->sched_data = &bfqg->sched_data;
363}
364
365static void bfqg_stats_exit(struct bfqg_stats *stats)
366{
367#ifdef CONFIG_DEBUG_BLK_CGROUP
368 blkg_rwstat_exit(&stats->merged);
369 blkg_rwstat_exit(&stats->service_time);
370 blkg_rwstat_exit(&stats->wait_time);
371 blkg_rwstat_exit(&stats->queued);
372 blkg_stat_exit(&stats->time);
373 blkg_stat_exit(&stats->avg_queue_size_sum);
374 blkg_stat_exit(&stats->avg_queue_size_samples);
375 blkg_stat_exit(&stats->dequeue);
376 blkg_stat_exit(&stats->group_wait_time);
377 blkg_stat_exit(&stats->idle_time);
378 blkg_stat_exit(&stats->empty_time);
379#endif
380}
381
382static int bfqg_stats_init(struct bfqg_stats *stats, gfp_t gfp)
383{
384#ifdef CONFIG_DEBUG_BLK_CGROUP
385 if (blkg_rwstat_init(&stats->merged, gfp) ||
386 blkg_rwstat_init(&stats->service_time, gfp) ||
387 blkg_rwstat_init(&stats->wait_time, gfp) ||
388 blkg_rwstat_init(&stats->queued, gfp) ||
389 blkg_stat_init(&stats->time, gfp) ||
390 blkg_stat_init(&stats->avg_queue_size_sum, gfp) ||
391 blkg_stat_init(&stats->avg_queue_size_samples, gfp) ||
392 blkg_stat_init(&stats->dequeue, gfp) ||
393 blkg_stat_init(&stats->group_wait_time, gfp) ||
394 blkg_stat_init(&stats->idle_time, gfp) ||
395 blkg_stat_init(&stats->empty_time, gfp)) {
396 bfqg_stats_exit(stats);
397 return -ENOMEM;
398 }
399#endif
400
401 return 0;
402}
403
404static struct bfq_group_data *cpd_to_bfqgd(struct blkcg_policy_data *cpd)
405{
406 return cpd ? container_of(cpd, struct bfq_group_data, pd) : NULL;
407}
408
409static struct bfq_group_data *blkcg_to_bfqgd(struct blkcg *blkcg)
410{
411 return cpd_to_bfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_bfq));
412}
413
414static struct blkcg_policy_data *bfq_cpd_alloc(gfp_t gfp)
415{
416 struct bfq_group_data *bgd;
417
418 bgd = kzalloc(sizeof(*bgd), gfp);
419 if (!bgd)
420 return NULL;
421 return &bgd->pd;
422}
423
424static void bfq_cpd_init(struct blkcg_policy_data *cpd)
425{
426 struct bfq_group_data *d = cpd_to_bfqgd(cpd);
427
428 d->weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ?
429 CGROUP_WEIGHT_DFL : BFQ_WEIGHT_LEGACY_DFL;
430}
431
432static void bfq_cpd_free(struct blkcg_policy_data *cpd)
433{
434 kfree(cpd_to_bfqgd(cpd));
435}
436
437static struct blkg_policy_data *bfq_pd_alloc(gfp_t gfp, int node)
438{
439 struct bfq_group *bfqg;
440
441 bfqg = kzalloc_node(sizeof(*bfqg), gfp, node);
442 if (!bfqg)
443 return NULL;
444
445 if (bfqg_stats_init(&bfqg->stats, gfp)) {
446 kfree(bfqg);
447 return NULL;
448 }
449
450 /* see comments in bfq_bic_update_cgroup for why refcounting */
451 bfqg_get(bfqg);
452 return &bfqg->pd;
453}
454
455static void bfq_pd_init(struct blkg_policy_data *pd)
456{
457 struct blkcg_gq *blkg = pd_to_blkg(pd);
458 struct bfq_group *bfqg = blkg_to_bfqg(blkg);
459 struct bfq_data *bfqd = blkg->q->elevator->elevator_data;
460 struct bfq_entity *entity = &bfqg->entity;
461 struct bfq_group_data *d = blkcg_to_bfqgd(blkg->blkcg);
462
463 entity->orig_weight = entity->weight = entity->new_weight = d->weight;
464 entity->my_sched_data = &bfqg->sched_data;
465 bfqg->my_entity = entity; /*
466 * the root_group's will be set to NULL
467 * in bfq_init_queue()
468 */
469 bfqg->bfqd = bfqd;
470 bfqg->active_entities = 0;
471 bfqg->rq_pos_tree = RB_ROOT;
472}
473
474static void bfq_pd_free(struct blkg_policy_data *pd)
475{
476 struct bfq_group *bfqg = pd_to_bfqg(pd);
477
478 bfqg_stats_exit(&bfqg->stats);
479 bfqg_put(bfqg);
480}
481
482static void bfq_pd_reset_stats(struct blkg_policy_data *pd)
483{
484 struct bfq_group *bfqg = pd_to_bfqg(pd);
485
486 bfqg_stats_reset(&bfqg->stats);
487}
488
489static void bfq_group_set_parent(struct bfq_group *bfqg,
490 struct bfq_group *parent)
491{
492 struct bfq_entity *entity;
493
494 entity = &bfqg->entity;
495 entity->parent = parent->my_entity;
496 entity->sched_data = &parent->sched_data;
497}
498
499static struct bfq_group *bfq_lookup_bfqg(struct bfq_data *bfqd,
500 struct blkcg *blkcg)
501{
502 struct blkcg_gq *blkg;
503
504 blkg = blkg_lookup(blkcg, bfqd->queue);
505 if (likely(blkg))
506 return blkg_to_bfqg(blkg);
507 return NULL;
508}
509
510struct bfq_group *bfq_find_set_group(struct bfq_data *bfqd,
511 struct blkcg *blkcg)
512{
513 struct bfq_group *bfqg, *parent;
514 struct bfq_entity *entity;
515
516 bfqg = bfq_lookup_bfqg(bfqd, blkcg);
517
518 if (unlikely(!bfqg))
519 return NULL;
520
521 /*
522 * Update chain of bfq_groups as we might be handling a leaf group
523 * which, along with some of its relatives, has not been hooked yet
524 * to the private hierarchy of BFQ.
525 */
526 entity = &bfqg->entity;
527 for_each_entity(entity) {
528 bfqg = container_of(entity, struct bfq_group, entity);
529 if (bfqg != bfqd->root_group) {
530 parent = bfqg_parent(bfqg);
531 if (!parent)
532 parent = bfqd->root_group;
533 bfq_group_set_parent(bfqg, parent);
534 }
535 }
536
537 return bfqg;
538}
539
540/**
541 * bfq_bfqq_move - migrate @bfqq to @bfqg.
542 * @bfqd: queue descriptor.
543 * @bfqq: the queue to move.
544 * @bfqg: the group to move to.
545 *
546 * Move @bfqq to @bfqg, deactivating it from its old group and reactivating
547 * it on the new one. Avoid putting the entity on the old group idle tree.
548 *
549 * Must be called under the scheduler lock, to make sure that the blkg
550 * owning @bfqg does not disappear (see comments in
551 * bfq_bic_update_cgroup on guaranteeing the consistency of blkg
552 * objects).
553 */
554void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq,
555 struct bfq_group *bfqg)
556{
557 struct bfq_entity *entity = &bfqq->entity;
558
559 /* If bfqq is empty, then bfq_bfqq_expire also invokes
560 * bfq_del_bfqq_busy, thereby removing bfqq and its entity
561 * from data structures related to current group. Otherwise we
562 * need to remove bfqq explicitly with bfq_deactivate_bfqq, as
563 * we do below.
564 */
565 if (bfqq == bfqd->in_service_queue)
566 bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
567 false, BFQQE_PREEMPTED);
568
569 if (bfq_bfqq_busy(bfqq))
570 bfq_deactivate_bfqq(bfqd, bfqq, false, false);
571 else if (entity->on_st)
572 bfq_put_idle_entity(bfq_entity_service_tree(entity), entity);
573 bfqg_and_blkg_put(bfqq_group(bfqq));
574
575 entity->parent = bfqg->my_entity;
576 entity->sched_data = &bfqg->sched_data;
577 /* pin down bfqg and its associated blkg */
578 bfqg_and_blkg_get(bfqg);
579
580 if (bfq_bfqq_busy(bfqq)) {
581 bfq_pos_tree_add_move(bfqd, bfqq);
582 bfq_activate_bfqq(bfqd, bfqq);
583 }
584
585 if (!bfqd->in_service_queue && !bfqd->rq_in_driver)
586 bfq_schedule_dispatch(bfqd);
587}
588
589/**
590 * __bfq_bic_change_cgroup - move @bic to @cgroup.
591 * @bfqd: the queue descriptor.
592 * @bic: the bic to move.
593 * @blkcg: the blk-cgroup to move to.
594 *
595 * Move bic to blkcg, assuming that bfqd->lock is held; which makes
596 * sure that the reference to cgroup is valid across the call (see
597 * comments in bfq_bic_update_cgroup on this issue)
598 *
599 * NOTE: an alternative approach might have been to store the current
600 * cgroup in bfqq and getting a reference to it, reducing the lookup
601 * time here, at the price of slightly more complex code.
602 */
603static struct bfq_group *__bfq_bic_change_cgroup(struct bfq_data *bfqd,
604 struct bfq_io_cq *bic,
605 struct blkcg *blkcg)
606{
607 struct bfq_queue *async_bfqq = bic_to_bfqq(bic, 0);
608 struct bfq_queue *sync_bfqq = bic_to_bfqq(bic, 1);
609 struct bfq_group *bfqg;
610 struct bfq_entity *entity;
611
612 bfqg = bfq_find_set_group(bfqd, blkcg);
613
614 if (unlikely(!bfqg))
615 bfqg = bfqd->root_group;
616
617 if (async_bfqq) {
618 entity = &async_bfqq->entity;
619
620 if (entity->sched_data != &bfqg->sched_data) {
621 bic_set_bfqq(bic, NULL, 0);
622 bfq_log_bfqq(bfqd, async_bfqq,
623 "bic_change_group: %p %d",
624 async_bfqq, async_bfqq->ref);
625 bfq_put_queue(async_bfqq);
626 }
627 }
628
629 if (sync_bfqq) {
630 entity = &sync_bfqq->entity;
631 if (entity->sched_data != &bfqg->sched_data)
632 bfq_bfqq_move(bfqd, sync_bfqq, bfqg);
633 }
634
635 return bfqg;
636}
637
638void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio)
639{
640 struct bfq_data *bfqd = bic_to_bfqd(bic);
641 struct bfq_group *bfqg = NULL;
642 uint64_t serial_nr;
643
644 rcu_read_lock();
645 serial_nr = __bio_blkcg(bio)->css.serial_nr;
646
647 /*
648 * Check whether blkcg has changed. The condition may trigger
649 * spuriously on a newly created cic but there's no harm.
650 */
651 if (unlikely(!bfqd) || likely(bic->blkcg_serial_nr == serial_nr))
652 goto out;
653
654 bfqg = __bfq_bic_change_cgroup(bfqd, bic, __bio_blkcg(bio));
655 /*
656 * Update blkg_path for bfq_log_* functions. We cache this
657 * path, and update it here, for the following
658 * reasons. Operations on blkg objects in blk-cgroup are
659 * protected with the request_queue lock, and not with the
660 * lock that protects the instances of this scheduler
661 * (bfqd->lock). This exposes BFQ to the following sort of
662 * race.
663 *
664 * The blkg_lookup performed in bfq_get_queue, protected
665 * through rcu, may happen to return the address of a copy of
666 * the original blkg. If this is the case, then the
667 * bfqg_and_blkg_get performed in bfq_get_queue, to pin down
668 * the blkg, is useless: it does not prevent blk-cgroup code
669 * from destroying both the original blkg and all objects
670 * directly or indirectly referred by the copy of the
671 * blkg.
672 *
673 * On the bright side, destroy operations on a blkg invoke, as
674 * a first step, hooks of the scheduler associated with the
675 * blkg. And these hooks are executed with bfqd->lock held for
676 * BFQ. As a consequence, for any blkg associated with the
677 * request queue this instance of the scheduler is attached
678 * to, we are guaranteed that such a blkg is not destroyed, and
679 * that all the pointers it contains are consistent, while we
680 * are holding bfqd->lock. A blkg_lookup performed with
681 * bfqd->lock held then returns a fully consistent blkg, which
682 * remains consistent until this lock is held.
683 *
684 * Thanks to the last fact, and to the fact that: (1) bfqg has
685 * been obtained through a blkg_lookup in the above
686 * assignment, and (2) bfqd->lock is being held, here we can
687 * safely use the policy data for the involved blkg (i.e., the
688 * field bfqg->pd) to get to the blkg associated with bfqg,
689 * and then we can safely use any field of blkg. After we
690 * release bfqd->lock, even just getting blkg through this
691 * bfqg may cause dangling references to be traversed, as
692 * bfqg->pd may not exist any more.
693 *
694 * In view of the above facts, here we cache, in the bfqg, any
695 * blkg data we may need for this bic, and for its associated
696 * bfq_queue. As of now, we need to cache only the path of the
697 * blkg, which is used in the bfq_log_* functions.
698 *
699 * Finally, note that bfqg itself needs to be protected from
700 * destruction on the blkg_free of the original blkg (which
701 * invokes bfq_pd_free). We use an additional private
702 * refcounter for bfqg, to let it disappear only after no
703 * bfq_queue refers to it any longer.
704 */
705 blkg_path(bfqg_to_blkg(bfqg), bfqg->blkg_path, sizeof(bfqg->blkg_path));
706 bic->blkcg_serial_nr = serial_nr;
707out:
708 rcu_read_unlock();
709}
710
711/**
712 * bfq_flush_idle_tree - deactivate any entity on the idle tree of @st.
713 * @st: the service tree being flushed.
714 */
715static void bfq_flush_idle_tree(struct bfq_service_tree *st)
716{
717 struct bfq_entity *entity = st->first_idle;
718
719 for (; entity ; entity = st->first_idle)
720 __bfq_deactivate_entity(entity, false);
721}
722
723/**
724 * bfq_reparent_leaf_entity - move leaf entity to the root_group.
725 * @bfqd: the device data structure with the root group.
726 * @entity: the entity to move.
727 */
728static void bfq_reparent_leaf_entity(struct bfq_data *bfqd,
729 struct bfq_entity *entity)
730{
731 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
732
733 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
734}
735
736/**
737 * bfq_reparent_active_entities - move to the root group all active
738 * entities.
739 * @bfqd: the device data structure with the root group.
740 * @bfqg: the group to move from.
741 * @st: the service tree with the entities.
742 */
743static void bfq_reparent_active_entities(struct bfq_data *bfqd,
744 struct bfq_group *bfqg,
745 struct bfq_service_tree *st)
746{
747 struct rb_root *active = &st->active;
748 struct bfq_entity *entity = NULL;
749
750 if (!RB_EMPTY_ROOT(&st->active))
751 entity = bfq_entity_of(rb_first(active));
752
753 for (; entity ; entity = bfq_entity_of(rb_first(active)))
754 bfq_reparent_leaf_entity(bfqd, entity);
755
756 if (bfqg->sched_data.in_service_entity)
757 bfq_reparent_leaf_entity(bfqd,
758 bfqg->sched_data.in_service_entity);
759}
760
761/**
762 * bfq_pd_offline - deactivate the entity associated with @pd,
763 * and reparent its children entities.
764 * @pd: descriptor of the policy going offline.
765 *
766 * blkio already grabs the queue_lock for us, so no need to use
767 * RCU-based magic
768 */
769static void bfq_pd_offline(struct blkg_policy_data *pd)
770{
771 struct bfq_service_tree *st;
772 struct bfq_group *bfqg = pd_to_bfqg(pd);
773 struct bfq_data *bfqd = bfqg->bfqd;
774 struct bfq_entity *entity = bfqg->my_entity;
775 unsigned long flags;
776 int i;
777
778 spin_lock_irqsave(&bfqd->lock, flags);
779
780 if (!entity) /* root group */
781 goto put_async_queues;
782
783 /*
784 * Empty all service_trees belonging to this group before
785 * deactivating the group itself.
786 */
787 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) {
788 st = bfqg->sched_data.service_tree + i;
789
790 /*
791 * The idle tree may still contain bfq_queues belonging
792 * to exited task because they never migrated to a different
793 * cgroup from the one being destroyed now.
794 */
795 bfq_flush_idle_tree(st);
796
797 /*
798 * It may happen that some queues are still active
799 * (busy) upon group destruction (if the corresponding
800 * processes have been forced to terminate). We move
801 * all the leaf entities corresponding to these queues
802 * to the root_group.
803 * Also, it may happen that the group has an entity
804 * in service, which is disconnected from the active
805 * tree: it must be moved, too.
806 * There is no need to put the sync queues, as the
807 * scheduler has taken no reference.
808 */
809 bfq_reparent_active_entities(bfqd, bfqg, st);
810 }
811
812 __bfq_deactivate_entity(entity, false);
813
814put_async_queues:
815 bfq_put_async_queues(bfqd, bfqg);
816
817 spin_unlock_irqrestore(&bfqd->lock, flags);
818 /*
819 * @blkg is going offline and will be ignored by
820 * blkg_[rw]stat_recursive_sum(). Transfer stats to the parent so
821 * that they don't get lost. If IOs complete after this point, the
822 * stats for them will be lost. Oh well...
823 */
824 bfqg_stats_xfer_dead(bfqg);
825}
826
827void bfq_end_wr_async(struct bfq_data *bfqd)
828{
829 struct blkcg_gq *blkg;
830
831 list_for_each_entry(blkg, &bfqd->queue->blkg_list, q_node) {
832 struct bfq_group *bfqg = blkg_to_bfqg(blkg);
833
834 bfq_end_wr_async_queues(bfqd, bfqg);
835 }
836 bfq_end_wr_async_queues(bfqd, bfqd->root_group);
837}
838
839static int bfq_io_show_weight(struct seq_file *sf, void *v)
840{
841 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
842 struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
843 unsigned int val = 0;
844
845 if (bfqgd)
846 val = bfqgd->weight;
847
848 seq_printf(sf, "%u\n", val);
849
850 return 0;
851}
852
853static int bfq_io_set_weight_legacy(struct cgroup_subsys_state *css,
854 struct cftype *cftype,
855 u64 val)
856{
857 struct blkcg *blkcg = css_to_blkcg(css);
858 struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
859 struct blkcg_gq *blkg;
860 int ret = -ERANGE;
861
862 if (val < BFQ_MIN_WEIGHT || val > BFQ_MAX_WEIGHT)
863 return ret;
864
865 ret = 0;
866 spin_lock_irq(&blkcg->lock);
867 bfqgd->weight = (unsigned short)val;
868 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
869 struct bfq_group *bfqg = blkg_to_bfqg(blkg);
870
871 if (!bfqg)
872 continue;
873 /*
874 * Setting the prio_changed flag of the entity
875 * to 1 with new_weight == weight would re-set
876 * the value of the weight to its ioprio mapping.
877 * Set the flag only if necessary.
878 */
879 if ((unsigned short)val != bfqg->entity.new_weight) {
880 bfqg->entity.new_weight = (unsigned short)val;
881 /*
882 * Make sure that the above new value has been
883 * stored in bfqg->entity.new_weight before
884 * setting the prio_changed flag. In fact,
885 * this flag may be read asynchronously (in
886 * critical sections protected by a different
887 * lock than that held here), and finding this
888 * flag set may cause the execution of the code
889 * for updating parameters whose value may
890 * depend also on bfqg->entity.new_weight (in
891 * __bfq_entity_update_weight_prio).
892 * This barrier makes sure that the new value
893 * of bfqg->entity.new_weight is correctly
894 * seen in that code.
895 */
896 smp_wmb();
897 bfqg->entity.prio_changed = 1;
898 }
899 }
900 spin_unlock_irq(&blkcg->lock);
901
902 return ret;
903}
904
905static ssize_t bfq_io_set_weight(struct kernfs_open_file *of,
906 char *buf, size_t nbytes,
907 loff_t off)
908{
909 u64 weight;
910 /* First unsigned long found in the file is used */
911 int ret = kstrtoull(strim(buf), 0, &weight);
912
913 if (ret)
914 return ret;
915
916 ret = bfq_io_set_weight_legacy(of_css(of), NULL, weight);
917 return ret ?: nbytes;
918}
919
920#ifdef CONFIG_DEBUG_BLK_CGROUP
921static int bfqg_print_stat(struct seq_file *sf, void *v)
922{
923 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
924 &blkcg_policy_bfq, seq_cft(sf)->private, false);
925 return 0;
926}
927
928static int bfqg_print_rwstat(struct seq_file *sf, void *v)
929{
930 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
931 &blkcg_policy_bfq, seq_cft(sf)->private, true);
932 return 0;
933}
934
935static u64 bfqg_prfill_stat_recursive(struct seq_file *sf,
936 struct blkg_policy_data *pd, int off)
937{
938 u64 sum = blkg_stat_recursive_sum(pd_to_blkg(pd),
939 &blkcg_policy_bfq, off);
940 return __blkg_prfill_u64(sf, pd, sum);
941}
942
943static u64 bfqg_prfill_rwstat_recursive(struct seq_file *sf,
944 struct blkg_policy_data *pd, int off)
945{
946 struct blkg_rwstat sum = blkg_rwstat_recursive_sum(pd_to_blkg(pd),
947 &blkcg_policy_bfq,
948 off);
949 return __blkg_prfill_rwstat(sf, pd, &sum);
950}
951
952static int bfqg_print_stat_recursive(struct seq_file *sf, void *v)
953{
954 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
955 bfqg_prfill_stat_recursive, &blkcg_policy_bfq,
956 seq_cft(sf)->private, false);
957 return 0;
958}
959
960static int bfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
961{
962 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
963 bfqg_prfill_rwstat_recursive, &blkcg_policy_bfq,
964 seq_cft(sf)->private, true);
965 return 0;
966}
967
968static u64 bfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
969 int off)
970{
971 u64 sum = blkg_rwstat_total(&pd->blkg->stat_bytes);
972
973 return __blkg_prfill_u64(sf, pd, sum >> 9);
974}
975
976static int bfqg_print_stat_sectors(struct seq_file *sf, void *v)
977{
978 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
979 bfqg_prfill_sectors, &blkcg_policy_bfq, 0, false);
980 return 0;
981}
982
983static u64 bfqg_prfill_sectors_recursive(struct seq_file *sf,
984 struct blkg_policy_data *pd, int off)
985{
986 struct blkg_rwstat tmp = blkg_rwstat_recursive_sum(pd->blkg, NULL,
987 offsetof(struct blkcg_gq, stat_bytes));
988 u64 sum = atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_READ]) +
989 atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_WRITE]);
990
991 return __blkg_prfill_u64(sf, pd, sum >> 9);
992}
993
994static int bfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
995{
996 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
997 bfqg_prfill_sectors_recursive, &blkcg_policy_bfq, 0,
998 false);
999 return 0;
1000}
1001
1002static u64 bfqg_prfill_avg_queue_size(struct seq_file *sf,
1003 struct blkg_policy_data *pd, int off)
1004{
1005 struct bfq_group *bfqg = pd_to_bfqg(pd);
1006 u64 samples = blkg_stat_read(&bfqg->stats.avg_queue_size_samples);
1007 u64 v = 0;
1008
1009 if (samples) {
1010 v = blkg_stat_read(&bfqg->stats.avg_queue_size_sum);
1011 v = div64_u64(v, samples);
1012 }
1013 __blkg_prfill_u64(sf, pd, v);
1014 return 0;
1015}
1016
1017/* print avg_queue_size */
1018static int bfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1019{
1020 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1021 bfqg_prfill_avg_queue_size, &blkcg_policy_bfq,
1022 0, false);
1023 return 0;
1024}
1025#endif /* CONFIG_DEBUG_BLK_CGROUP */
1026
1027struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node)
1028{
1029 int ret;
1030
1031 ret = blkcg_activate_policy(bfqd->queue, &blkcg_policy_bfq);
1032 if (ret)
1033 return NULL;
1034
1035 return blkg_to_bfqg(bfqd->queue->root_blkg);
1036}
1037
1038struct blkcg_policy blkcg_policy_bfq = {
1039 .dfl_cftypes = bfq_blkg_files,
1040 .legacy_cftypes = bfq_blkcg_legacy_files,
1041
1042 .cpd_alloc_fn = bfq_cpd_alloc,
1043 .cpd_init_fn = bfq_cpd_init,
1044 .cpd_bind_fn = bfq_cpd_init,
1045 .cpd_free_fn = bfq_cpd_free,
1046
1047 .pd_alloc_fn = bfq_pd_alloc,
1048 .pd_init_fn = bfq_pd_init,
1049 .pd_offline_fn = bfq_pd_offline,
1050 .pd_free_fn = bfq_pd_free,
1051 .pd_reset_stats_fn = bfq_pd_reset_stats,
1052};
1053
1054struct cftype bfq_blkcg_legacy_files[] = {
1055 {
1056 .name = "bfq.weight",
1057 .flags = CFTYPE_NOT_ON_ROOT,
1058 .seq_show = bfq_io_show_weight,
1059 .write_u64 = bfq_io_set_weight_legacy,
1060 },
1061
1062 /* statistics, covers only the tasks in the bfqg */
1063 {
1064 .name = "bfq.io_service_bytes",
1065 .private = (unsigned long)&blkcg_policy_bfq,
1066 .seq_show = blkg_print_stat_bytes,
1067 },
1068 {
1069 .name = "bfq.io_serviced",
1070 .private = (unsigned long)&blkcg_policy_bfq,
1071 .seq_show = blkg_print_stat_ios,
1072 },
1073#ifdef CONFIG_DEBUG_BLK_CGROUP
1074 {
1075 .name = "bfq.time",
1076 .private = offsetof(struct bfq_group, stats.time),
1077 .seq_show = bfqg_print_stat,
1078 },
1079 {
1080 .name = "bfq.sectors",
1081 .seq_show = bfqg_print_stat_sectors,
1082 },
1083 {
1084 .name = "bfq.io_service_time",
1085 .private = offsetof(struct bfq_group, stats.service_time),
1086 .seq_show = bfqg_print_rwstat,
1087 },
1088 {
1089 .name = "bfq.io_wait_time",
1090 .private = offsetof(struct bfq_group, stats.wait_time),
1091 .seq_show = bfqg_print_rwstat,
1092 },
1093 {
1094 .name = "bfq.io_merged",
1095 .private = offsetof(struct bfq_group, stats.merged),
1096 .seq_show = bfqg_print_rwstat,
1097 },
1098 {
1099 .name = "bfq.io_queued",
1100 .private = offsetof(struct bfq_group, stats.queued),
1101 .seq_show = bfqg_print_rwstat,
1102 },
1103#endif /* CONFIG_DEBUG_BLK_CGROUP */
1104
1105 /* the same statictics which cover the bfqg and its descendants */
1106 {
1107 .name = "bfq.io_service_bytes_recursive",
1108 .private = (unsigned long)&blkcg_policy_bfq,
1109 .seq_show = blkg_print_stat_bytes_recursive,
1110 },
1111 {
1112 .name = "bfq.io_serviced_recursive",
1113 .private = (unsigned long)&blkcg_policy_bfq,
1114 .seq_show = blkg_print_stat_ios_recursive,
1115 },
1116#ifdef CONFIG_DEBUG_BLK_CGROUP
1117 {
1118 .name = "bfq.time_recursive",
1119 .private = offsetof(struct bfq_group, stats.time),
1120 .seq_show = bfqg_print_stat_recursive,
1121 },
1122 {
1123 .name = "bfq.sectors_recursive",
1124 .seq_show = bfqg_print_stat_sectors_recursive,
1125 },
1126 {
1127 .name = "bfq.io_service_time_recursive",
1128 .private = offsetof(struct bfq_group, stats.service_time),
1129 .seq_show = bfqg_print_rwstat_recursive,
1130 },
1131 {
1132 .name = "bfq.io_wait_time_recursive",
1133 .private = offsetof(struct bfq_group, stats.wait_time),
1134 .seq_show = bfqg_print_rwstat_recursive,
1135 },
1136 {
1137 .name = "bfq.io_merged_recursive",
1138 .private = offsetof(struct bfq_group, stats.merged),
1139 .seq_show = bfqg_print_rwstat_recursive,
1140 },
1141 {
1142 .name = "bfq.io_queued_recursive",
1143 .private = offsetof(struct bfq_group, stats.queued),
1144 .seq_show = bfqg_print_rwstat_recursive,
1145 },
1146 {
1147 .name = "bfq.avg_queue_size",
1148 .seq_show = bfqg_print_avg_queue_size,
1149 },
1150 {
1151 .name = "bfq.group_wait_time",
1152 .private = offsetof(struct bfq_group, stats.group_wait_time),
1153 .seq_show = bfqg_print_stat,
1154 },
1155 {
1156 .name = "bfq.idle_time",
1157 .private = offsetof(struct bfq_group, stats.idle_time),
1158 .seq_show = bfqg_print_stat,
1159 },
1160 {
1161 .name = "bfq.empty_time",
1162 .private = offsetof(struct bfq_group, stats.empty_time),
1163 .seq_show = bfqg_print_stat,
1164 },
1165 {
1166 .name = "bfq.dequeue",
1167 .private = offsetof(struct bfq_group, stats.dequeue),
1168 .seq_show = bfqg_print_stat,
1169 },
1170#endif /* CONFIG_DEBUG_BLK_CGROUP */
1171 { } /* terminate */
1172};
1173
1174struct cftype bfq_blkg_files[] = {
1175 {
1176 .name = "bfq.weight",
1177 .flags = CFTYPE_NOT_ON_ROOT,
1178 .seq_show = bfq_io_show_weight,
1179 .write = bfq_io_set_weight,
1180 },
1181 {} /* terminate */
1182};
1183
1184#else /* CONFIG_BFQ_GROUP_IOSCHED */
1185
1186void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1187 struct bfq_group *bfqg) {}
1188
1189void bfq_init_entity(struct bfq_entity *entity, struct bfq_group *bfqg)
1190{
1191 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
1192
1193 entity->weight = entity->new_weight;
1194 entity->orig_weight = entity->new_weight;
1195 if (bfqq) {
1196 bfqq->ioprio = bfqq->new_ioprio;
1197 bfqq->ioprio_class = bfqq->new_ioprio_class;
1198 }
1199 entity->sched_data = &bfqg->sched_data;
1200}
1201
1202void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio) {}
1203
1204void bfq_end_wr_async(struct bfq_data *bfqd)
1205{
1206 bfq_end_wr_async_queues(bfqd, bfqd->root_group);
1207}
1208
1209struct bfq_group *bfq_find_set_group(struct bfq_data *bfqd, struct blkcg *blkcg)
1210{
1211 return bfqd->root_group;
1212}
1213
1214struct bfq_group *bfqq_group(struct bfq_queue *bfqq)
1215{
1216 return bfqq->bfqd->root_group;
1217}
1218
1219struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node)
1220{
1221 struct bfq_group *bfqg;
1222 int i;
1223
1224 bfqg = kmalloc_node(sizeof(*bfqg), GFP_KERNEL | __GFP_ZERO, node);
1225 if (!bfqg)
1226 return NULL;
1227
1228 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
1229 bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
1230
1231 return bfqg;
1232}
1233#endif /* CONFIG_BFQ_GROUP_IOSCHED */
1234