Loading...
1/*
2 * Common Block IO controller cgroup interface
3 *
4 * Based on ideas and code from CFQ, CFS and BFQ:
5 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
6 *
7 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
8 * Paolo Valente <paolo.valente@unimore.it>
9 *
10 * Copyright (C) 2009 Vivek Goyal <vgoyal@redhat.com>
11 * Nauman Rafique <nauman@google.com>
12 *
13 * For policy-specific per-blkcg data:
14 * Copyright (C) 2015 Paolo Valente <paolo.valente@unimore.it>
15 * Arianna Avanzini <avanzini.arianna@gmail.com>
16 */
17#include <linux/ioprio.h>
18#include <linux/kdev_t.h>
19#include <linux/module.h>
20#include <linux/err.h>
21#include <linux/blkdev.h>
22#include <linux/backing-dev.h>
23#include <linux/slab.h>
24#include <linux/genhd.h>
25#include <linux/delay.h>
26#include <linux/atomic.h>
27#include <linux/ctype.h>
28#include <linux/blk-cgroup.h>
29#include "blk.h"
30
31#define MAX_KEY_LEN 100
32
33/*
34 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
35 * blkcg_pol_register_mutex nests outside of it and synchronizes entire
36 * policy [un]register operations including cgroup file additions /
37 * removals. Putting cgroup file registration outside blkcg_pol_mutex
38 * allows grabbing it from cgroup callbacks.
39 */
40static DEFINE_MUTEX(blkcg_pol_register_mutex);
41static DEFINE_MUTEX(blkcg_pol_mutex);
42
43struct blkcg blkcg_root;
44EXPORT_SYMBOL_GPL(blkcg_root);
45
46struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
47
48static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
49
50static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */
51
52static bool blkcg_policy_enabled(struct request_queue *q,
53 const struct blkcg_policy *pol)
54{
55 return pol && test_bit(pol->plid, q->blkcg_pols);
56}
57
58/**
59 * blkg_free - free a blkg
60 * @blkg: blkg to free
61 *
62 * Free @blkg which may be partially allocated.
63 */
64static void blkg_free(struct blkcg_gq *blkg)
65{
66 int i;
67
68 if (!blkg)
69 return;
70
71 for (i = 0; i < BLKCG_MAX_POLS; i++)
72 if (blkg->pd[i])
73 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
74
75 if (blkg->blkcg != &blkcg_root)
76 blk_exit_rl(&blkg->rl);
77
78 blkg_rwstat_exit(&blkg->stat_ios);
79 blkg_rwstat_exit(&blkg->stat_bytes);
80 kfree(blkg);
81}
82
83/**
84 * blkg_alloc - allocate a blkg
85 * @blkcg: block cgroup the new blkg is associated with
86 * @q: request_queue the new blkg is associated with
87 * @gfp_mask: allocation mask to use
88 *
89 * Allocate a new blkg assocating @blkcg and @q.
90 */
91static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
92 gfp_t gfp_mask)
93{
94 struct blkcg_gq *blkg;
95 int i;
96
97 /* alloc and init base part */
98 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
99 if (!blkg)
100 return NULL;
101
102 if (blkg_rwstat_init(&blkg->stat_bytes, gfp_mask) ||
103 blkg_rwstat_init(&blkg->stat_ios, gfp_mask))
104 goto err_free;
105
106 blkg->q = q;
107 INIT_LIST_HEAD(&blkg->q_node);
108 blkg->blkcg = blkcg;
109 atomic_set(&blkg->refcnt, 1);
110
111 /* root blkg uses @q->root_rl, init rl only for !root blkgs */
112 if (blkcg != &blkcg_root) {
113 if (blk_init_rl(&blkg->rl, q, gfp_mask))
114 goto err_free;
115 blkg->rl.blkg = blkg;
116 }
117
118 for (i = 0; i < BLKCG_MAX_POLS; i++) {
119 struct blkcg_policy *pol = blkcg_policy[i];
120 struct blkg_policy_data *pd;
121
122 if (!blkcg_policy_enabled(q, pol))
123 continue;
124
125 /* alloc per-policy data and attach it to blkg */
126 pd = pol->pd_alloc_fn(gfp_mask, q->node);
127 if (!pd)
128 goto err_free;
129
130 blkg->pd[i] = pd;
131 pd->blkg = blkg;
132 pd->plid = i;
133 }
134
135 return blkg;
136
137err_free:
138 blkg_free(blkg);
139 return NULL;
140}
141
142struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg,
143 struct request_queue *q, bool update_hint)
144{
145 struct blkcg_gq *blkg;
146
147 /*
148 * Hint didn't match. Look up from the radix tree. Note that the
149 * hint can only be updated under queue_lock as otherwise @blkg
150 * could have already been removed from blkg_tree. The caller is
151 * responsible for grabbing queue_lock if @update_hint.
152 */
153 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
154 if (blkg && blkg->q == q) {
155 if (update_hint) {
156 lockdep_assert_held(q->queue_lock);
157 rcu_assign_pointer(blkcg->blkg_hint, blkg);
158 }
159 return blkg;
160 }
161
162 return NULL;
163}
164EXPORT_SYMBOL_GPL(blkg_lookup_slowpath);
165
166/*
167 * If @new_blkg is %NULL, this function tries to allocate a new one as
168 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
169 */
170static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
171 struct request_queue *q,
172 struct blkcg_gq *new_blkg)
173{
174 struct blkcg_gq *blkg;
175 struct bdi_writeback_congested *wb_congested;
176 int i, ret;
177
178 WARN_ON_ONCE(!rcu_read_lock_held());
179 lockdep_assert_held(q->queue_lock);
180
181 /* blkg holds a reference to blkcg */
182 if (!css_tryget_online(&blkcg->css)) {
183 ret = -ENODEV;
184 goto err_free_blkg;
185 }
186
187 wb_congested = wb_congested_get_create(&q->backing_dev_info,
188 blkcg->css.id, GFP_NOWAIT);
189 if (!wb_congested) {
190 ret = -ENOMEM;
191 goto err_put_css;
192 }
193
194 /* allocate */
195 if (!new_blkg) {
196 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT);
197 if (unlikely(!new_blkg)) {
198 ret = -ENOMEM;
199 goto err_put_congested;
200 }
201 }
202 blkg = new_blkg;
203 blkg->wb_congested = wb_congested;
204
205 /* link parent */
206 if (blkcg_parent(blkcg)) {
207 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
208 if (WARN_ON_ONCE(!blkg->parent)) {
209 ret = -ENODEV;
210 goto err_put_congested;
211 }
212 blkg_get(blkg->parent);
213 }
214
215 /* invoke per-policy init */
216 for (i = 0; i < BLKCG_MAX_POLS; i++) {
217 struct blkcg_policy *pol = blkcg_policy[i];
218
219 if (blkg->pd[i] && pol->pd_init_fn)
220 pol->pd_init_fn(blkg->pd[i]);
221 }
222
223 /* insert */
224 spin_lock(&blkcg->lock);
225 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
226 if (likely(!ret)) {
227 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
228 list_add(&blkg->q_node, &q->blkg_list);
229
230 for (i = 0; i < BLKCG_MAX_POLS; i++) {
231 struct blkcg_policy *pol = blkcg_policy[i];
232
233 if (blkg->pd[i] && pol->pd_online_fn)
234 pol->pd_online_fn(blkg->pd[i]);
235 }
236 }
237 blkg->online = true;
238 spin_unlock(&blkcg->lock);
239
240 if (!ret)
241 return blkg;
242
243 /* @blkg failed fully initialized, use the usual release path */
244 blkg_put(blkg);
245 return ERR_PTR(ret);
246
247err_put_congested:
248 wb_congested_put(wb_congested);
249err_put_css:
250 css_put(&blkcg->css);
251err_free_blkg:
252 blkg_free(new_blkg);
253 return ERR_PTR(ret);
254}
255
256/**
257 * blkg_lookup_create - lookup blkg, try to create one if not there
258 * @blkcg: blkcg of interest
259 * @q: request_queue of interest
260 *
261 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
262 * create one. blkg creation is performed recursively from blkcg_root such
263 * that all non-root blkg's have access to the parent blkg. This function
264 * should be called under RCU read lock and @q->queue_lock.
265 *
266 * Returns pointer to the looked up or created blkg on success, ERR_PTR()
267 * value on error. If @q is dead, returns ERR_PTR(-EINVAL). If @q is not
268 * dead and bypassing, returns ERR_PTR(-EBUSY).
269 */
270struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
271 struct request_queue *q)
272{
273 struct blkcg_gq *blkg;
274
275 WARN_ON_ONCE(!rcu_read_lock_held());
276 lockdep_assert_held(q->queue_lock);
277
278 /*
279 * This could be the first entry point of blkcg implementation and
280 * we shouldn't allow anything to go through for a bypassing queue.
281 */
282 if (unlikely(blk_queue_bypass(q)))
283 return ERR_PTR(blk_queue_dying(q) ? -ENODEV : -EBUSY);
284
285 blkg = __blkg_lookup(blkcg, q, true);
286 if (blkg)
287 return blkg;
288
289 /*
290 * Create blkgs walking down from blkcg_root to @blkcg, so that all
291 * non-root blkgs have access to their parents.
292 */
293 while (true) {
294 struct blkcg *pos = blkcg;
295 struct blkcg *parent = blkcg_parent(blkcg);
296
297 while (parent && !__blkg_lookup(parent, q, false)) {
298 pos = parent;
299 parent = blkcg_parent(parent);
300 }
301
302 blkg = blkg_create(pos, q, NULL);
303 if (pos == blkcg || IS_ERR(blkg))
304 return blkg;
305 }
306}
307
308static void blkg_destroy(struct blkcg_gq *blkg)
309{
310 struct blkcg *blkcg = blkg->blkcg;
311 struct blkcg_gq *parent = blkg->parent;
312 int i;
313
314 lockdep_assert_held(blkg->q->queue_lock);
315 lockdep_assert_held(&blkcg->lock);
316
317 /* Something wrong if we are trying to remove same group twice */
318 WARN_ON_ONCE(list_empty(&blkg->q_node));
319 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
320
321 for (i = 0; i < BLKCG_MAX_POLS; i++) {
322 struct blkcg_policy *pol = blkcg_policy[i];
323
324 if (blkg->pd[i] && pol->pd_offline_fn)
325 pol->pd_offline_fn(blkg->pd[i]);
326 }
327
328 if (parent) {
329 blkg_rwstat_add_aux(&parent->stat_bytes, &blkg->stat_bytes);
330 blkg_rwstat_add_aux(&parent->stat_ios, &blkg->stat_ios);
331 }
332
333 blkg->online = false;
334
335 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
336 list_del_init(&blkg->q_node);
337 hlist_del_init_rcu(&blkg->blkcg_node);
338
339 /*
340 * Both setting lookup hint to and clearing it from @blkg are done
341 * under queue_lock. If it's not pointing to @blkg now, it never
342 * will. Hint assignment itself can race safely.
343 */
344 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
345 rcu_assign_pointer(blkcg->blkg_hint, NULL);
346
347 /*
348 * Put the reference taken at the time of creation so that when all
349 * queues are gone, group can be destroyed.
350 */
351 blkg_put(blkg);
352}
353
354/**
355 * blkg_destroy_all - destroy all blkgs associated with a request_queue
356 * @q: request_queue of interest
357 *
358 * Destroy all blkgs associated with @q.
359 */
360static void blkg_destroy_all(struct request_queue *q)
361{
362 struct blkcg_gq *blkg, *n;
363
364 lockdep_assert_held(q->queue_lock);
365
366 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
367 struct blkcg *blkcg = blkg->blkcg;
368
369 spin_lock(&blkcg->lock);
370 blkg_destroy(blkg);
371 spin_unlock(&blkcg->lock);
372 }
373
374 q->root_blkg = NULL;
375 q->root_rl.blkg = NULL;
376}
377
378/*
379 * A group is RCU protected, but having an rcu lock does not mean that one
380 * can access all the fields of blkg and assume these are valid. For
381 * example, don't try to follow throtl_data and request queue links.
382 *
383 * Having a reference to blkg under an rcu allows accesses to only values
384 * local to groups like group stats and group rate limits.
385 */
386void __blkg_release_rcu(struct rcu_head *rcu_head)
387{
388 struct blkcg_gq *blkg = container_of(rcu_head, struct blkcg_gq, rcu_head);
389
390 /* release the blkcg and parent blkg refs this blkg has been holding */
391 css_put(&blkg->blkcg->css);
392 if (blkg->parent)
393 blkg_put(blkg->parent);
394
395 wb_congested_put(blkg->wb_congested);
396
397 blkg_free(blkg);
398}
399EXPORT_SYMBOL_GPL(__blkg_release_rcu);
400
401/*
402 * The next function used by blk_queue_for_each_rl(). It's a bit tricky
403 * because the root blkg uses @q->root_rl instead of its own rl.
404 */
405struct request_list *__blk_queue_next_rl(struct request_list *rl,
406 struct request_queue *q)
407{
408 struct list_head *ent;
409 struct blkcg_gq *blkg;
410
411 /*
412 * Determine the current blkg list_head. The first entry is
413 * root_rl which is off @q->blkg_list and mapped to the head.
414 */
415 if (rl == &q->root_rl) {
416 ent = &q->blkg_list;
417 /* There are no more block groups, hence no request lists */
418 if (list_empty(ent))
419 return NULL;
420 } else {
421 blkg = container_of(rl, struct blkcg_gq, rl);
422 ent = &blkg->q_node;
423 }
424
425 /* walk to the next list_head, skip root blkcg */
426 ent = ent->next;
427 if (ent == &q->root_blkg->q_node)
428 ent = ent->next;
429 if (ent == &q->blkg_list)
430 return NULL;
431
432 blkg = container_of(ent, struct blkcg_gq, q_node);
433 return &blkg->rl;
434}
435
436static int blkcg_reset_stats(struct cgroup_subsys_state *css,
437 struct cftype *cftype, u64 val)
438{
439 struct blkcg *blkcg = css_to_blkcg(css);
440 struct blkcg_gq *blkg;
441 int i;
442
443 mutex_lock(&blkcg_pol_mutex);
444 spin_lock_irq(&blkcg->lock);
445
446 /*
447 * Note that stat reset is racy - it doesn't synchronize against
448 * stat updates. This is a debug feature which shouldn't exist
449 * anyway. If you get hit by a race, retry.
450 */
451 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
452 blkg_rwstat_reset(&blkg->stat_bytes);
453 blkg_rwstat_reset(&blkg->stat_ios);
454
455 for (i = 0; i < BLKCG_MAX_POLS; i++) {
456 struct blkcg_policy *pol = blkcg_policy[i];
457
458 if (blkg->pd[i] && pol->pd_reset_stats_fn)
459 pol->pd_reset_stats_fn(blkg->pd[i]);
460 }
461 }
462
463 spin_unlock_irq(&blkcg->lock);
464 mutex_unlock(&blkcg_pol_mutex);
465 return 0;
466}
467
468const char *blkg_dev_name(struct blkcg_gq *blkg)
469{
470 /* some drivers (floppy) instantiate a queue w/o disk registered */
471 if (blkg->q->backing_dev_info.dev)
472 return dev_name(blkg->q->backing_dev_info.dev);
473 return NULL;
474}
475EXPORT_SYMBOL_GPL(blkg_dev_name);
476
477/**
478 * blkcg_print_blkgs - helper for printing per-blkg data
479 * @sf: seq_file to print to
480 * @blkcg: blkcg of interest
481 * @prfill: fill function to print out a blkg
482 * @pol: policy in question
483 * @data: data to be passed to @prfill
484 * @show_total: to print out sum of prfill return values or not
485 *
486 * This function invokes @prfill on each blkg of @blkcg if pd for the
487 * policy specified by @pol exists. @prfill is invoked with @sf, the
488 * policy data and @data and the matching queue lock held. If @show_total
489 * is %true, the sum of the return values from @prfill is printed with
490 * "Total" label at the end.
491 *
492 * This is to be used to construct print functions for
493 * cftype->read_seq_string method.
494 */
495void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
496 u64 (*prfill)(struct seq_file *,
497 struct blkg_policy_data *, int),
498 const struct blkcg_policy *pol, int data,
499 bool show_total)
500{
501 struct blkcg_gq *blkg;
502 u64 total = 0;
503
504 rcu_read_lock();
505 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
506 spin_lock_irq(blkg->q->queue_lock);
507 if (blkcg_policy_enabled(blkg->q, pol))
508 total += prfill(sf, blkg->pd[pol->plid], data);
509 spin_unlock_irq(blkg->q->queue_lock);
510 }
511 rcu_read_unlock();
512
513 if (show_total)
514 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
515}
516EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
517
518/**
519 * __blkg_prfill_u64 - prfill helper for a single u64 value
520 * @sf: seq_file to print to
521 * @pd: policy private data of interest
522 * @v: value to print
523 *
524 * Print @v to @sf for the device assocaited with @pd.
525 */
526u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
527{
528 const char *dname = blkg_dev_name(pd->blkg);
529
530 if (!dname)
531 return 0;
532
533 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
534 return v;
535}
536EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
537
538/**
539 * __blkg_prfill_rwstat - prfill helper for a blkg_rwstat
540 * @sf: seq_file to print to
541 * @pd: policy private data of interest
542 * @rwstat: rwstat to print
543 *
544 * Print @rwstat to @sf for the device assocaited with @pd.
545 */
546u64 __blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
547 const struct blkg_rwstat *rwstat)
548{
549 static const char *rwstr[] = {
550 [BLKG_RWSTAT_READ] = "Read",
551 [BLKG_RWSTAT_WRITE] = "Write",
552 [BLKG_RWSTAT_SYNC] = "Sync",
553 [BLKG_RWSTAT_ASYNC] = "Async",
554 };
555 const char *dname = blkg_dev_name(pd->blkg);
556 u64 v;
557 int i;
558
559 if (!dname)
560 return 0;
561
562 for (i = 0; i < BLKG_RWSTAT_NR; i++)
563 seq_printf(sf, "%s %s %llu\n", dname, rwstr[i],
564 (unsigned long long)atomic64_read(&rwstat->aux_cnt[i]));
565
566 v = atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_READ]) +
567 atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_WRITE]);
568 seq_printf(sf, "%s Total %llu\n", dname, (unsigned long long)v);
569 return v;
570}
571EXPORT_SYMBOL_GPL(__blkg_prfill_rwstat);
572
573/**
574 * blkg_prfill_stat - prfill callback for blkg_stat
575 * @sf: seq_file to print to
576 * @pd: policy private data of interest
577 * @off: offset to the blkg_stat in @pd
578 *
579 * prfill callback for printing a blkg_stat.
580 */
581u64 blkg_prfill_stat(struct seq_file *sf, struct blkg_policy_data *pd, int off)
582{
583 return __blkg_prfill_u64(sf, pd, blkg_stat_read((void *)pd + off));
584}
585EXPORT_SYMBOL_GPL(blkg_prfill_stat);
586
587/**
588 * blkg_prfill_rwstat - prfill callback for blkg_rwstat
589 * @sf: seq_file to print to
590 * @pd: policy private data of interest
591 * @off: offset to the blkg_rwstat in @pd
592 *
593 * prfill callback for printing a blkg_rwstat.
594 */
595u64 blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
596 int off)
597{
598 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd + off);
599
600 return __blkg_prfill_rwstat(sf, pd, &rwstat);
601}
602EXPORT_SYMBOL_GPL(blkg_prfill_rwstat);
603
604static u64 blkg_prfill_rwstat_field(struct seq_file *sf,
605 struct blkg_policy_data *pd, int off)
606{
607 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd->blkg + off);
608
609 return __blkg_prfill_rwstat(sf, pd, &rwstat);
610}
611
612/**
613 * blkg_print_stat_bytes - seq_show callback for blkg->stat_bytes
614 * @sf: seq_file to print to
615 * @v: unused
616 *
617 * To be used as cftype->seq_show to print blkg->stat_bytes.
618 * cftype->private must be set to the blkcg_policy.
619 */
620int blkg_print_stat_bytes(struct seq_file *sf, void *v)
621{
622 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
623 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
624 offsetof(struct blkcg_gq, stat_bytes), true);
625 return 0;
626}
627EXPORT_SYMBOL_GPL(blkg_print_stat_bytes);
628
629/**
630 * blkg_print_stat_bytes - seq_show callback for blkg->stat_ios
631 * @sf: seq_file to print to
632 * @v: unused
633 *
634 * To be used as cftype->seq_show to print blkg->stat_ios. cftype->private
635 * must be set to the blkcg_policy.
636 */
637int blkg_print_stat_ios(struct seq_file *sf, void *v)
638{
639 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
640 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
641 offsetof(struct blkcg_gq, stat_ios), true);
642 return 0;
643}
644EXPORT_SYMBOL_GPL(blkg_print_stat_ios);
645
646static u64 blkg_prfill_rwstat_field_recursive(struct seq_file *sf,
647 struct blkg_policy_data *pd,
648 int off)
649{
650 struct blkg_rwstat rwstat = blkg_rwstat_recursive_sum(pd->blkg,
651 NULL, off);
652 return __blkg_prfill_rwstat(sf, pd, &rwstat);
653}
654
655/**
656 * blkg_print_stat_bytes_recursive - recursive version of blkg_print_stat_bytes
657 * @sf: seq_file to print to
658 * @v: unused
659 */
660int blkg_print_stat_bytes_recursive(struct seq_file *sf, void *v)
661{
662 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
663 blkg_prfill_rwstat_field_recursive,
664 (void *)seq_cft(sf)->private,
665 offsetof(struct blkcg_gq, stat_bytes), true);
666 return 0;
667}
668EXPORT_SYMBOL_GPL(blkg_print_stat_bytes_recursive);
669
670/**
671 * blkg_print_stat_ios_recursive - recursive version of blkg_print_stat_ios
672 * @sf: seq_file to print to
673 * @v: unused
674 */
675int blkg_print_stat_ios_recursive(struct seq_file *sf, void *v)
676{
677 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
678 blkg_prfill_rwstat_field_recursive,
679 (void *)seq_cft(sf)->private,
680 offsetof(struct blkcg_gq, stat_ios), true);
681 return 0;
682}
683EXPORT_SYMBOL_GPL(blkg_print_stat_ios_recursive);
684
685/**
686 * blkg_stat_recursive_sum - collect hierarchical blkg_stat
687 * @blkg: blkg of interest
688 * @pol: blkcg_policy which contains the blkg_stat
689 * @off: offset to the blkg_stat in blkg_policy_data or @blkg
690 *
691 * Collect the blkg_stat specified by @blkg, @pol and @off and all its
692 * online descendants and their aux counts. The caller must be holding the
693 * queue lock for online tests.
694 *
695 * If @pol is NULL, blkg_stat is at @off bytes into @blkg; otherwise, it is
696 * at @off bytes into @blkg's blkg_policy_data of the policy.
697 */
698u64 blkg_stat_recursive_sum(struct blkcg_gq *blkg,
699 struct blkcg_policy *pol, int off)
700{
701 struct blkcg_gq *pos_blkg;
702 struct cgroup_subsys_state *pos_css;
703 u64 sum = 0;
704
705 lockdep_assert_held(blkg->q->queue_lock);
706
707 rcu_read_lock();
708 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
709 struct blkg_stat *stat;
710
711 if (!pos_blkg->online)
712 continue;
713
714 if (pol)
715 stat = (void *)blkg_to_pd(pos_blkg, pol) + off;
716 else
717 stat = (void *)blkg + off;
718
719 sum += blkg_stat_read(stat) + atomic64_read(&stat->aux_cnt);
720 }
721 rcu_read_unlock();
722
723 return sum;
724}
725EXPORT_SYMBOL_GPL(blkg_stat_recursive_sum);
726
727/**
728 * blkg_rwstat_recursive_sum - collect hierarchical blkg_rwstat
729 * @blkg: blkg of interest
730 * @pol: blkcg_policy which contains the blkg_rwstat
731 * @off: offset to the blkg_rwstat in blkg_policy_data or @blkg
732 *
733 * Collect the blkg_rwstat specified by @blkg, @pol and @off and all its
734 * online descendants and their aux counts. The caller must be holding the
735 * queue lock for online tests.
736 *
737 * If @pol is NULL, blkg_rwstat is at @off bytes into @blkg; otherwise, it
738 * is at @off bytes into @blkg's blkg_policy_data of the policy.
739 */
740struct blkg_rwstat blkg_rwstat_recursive_sum(struct blkcg_gq *blkg,
741 struct blkcg_policy *pol, int off)
742{
743 struct blkcg_gq *pos_blkg;
744 struct cgroup_subsys_state *pos_css;
745 struct blkg_rwstat sum = { };
746 int i;
747
748 lockdep_assert_held(blkg->q->queue_lock);
749
750 rcu_read_lock();
751 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
752 struct blkg_rwstat *rwstat;
753
754 if (!pos_blkg->online)
755 continue;
756
757 if (pol)
758 rwstat = (void *)blkg_to_pd(pos_blkg, pol) + off;
759 else
760 rwstat = (void *)pos_blkg + off;
761
762 for (i = 0; i < BLKG_RWSTAT_NR; i++)
763 atomic64_add(atomic64_read(&rwstat->aux_cnt[i]) +
764 percpu_counter_sum_positive(&rwstat->cpu_cnt[i]),
765 &sum.aux_cnt[i]);
766 }
767 rcu_read_unlock();
768
769 return sum;
770}
771EXPORT_SYMBOL_GPL(blkg_rwstat_recursive_sum);
772
773/**
774 * blkg_conf_prep - parse and prepare for per-blkg config update
775 * @blkcg: target block cgroup
776 * @pol: target policy
777 * @input: input string
778 * @ctx: blkg_conf_ctx to be filled
779 *
780 * Parse per-blkg config update from @input and initialize @ctx with the
781 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
782 * part of @input following MAJ:MIN. This function returns with RCU read
783 * lock and queue lock held and must be paired with blkg_conf_finish().
784 */
785int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
786 char *input, struct blkg_conf_ctx *ctx)
787 __acquires(rcu) __acquires(disk->queue->queue_lock)
788{
789 struct gendisk *disk;
790 struct blkcg_gq *blkg;
791 struct module *owner;
792 unsigned int major, minor;
793 int key_len, part, ret;
794 char *body;
795
796 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
797 return -EINVAL;
798
799 body = input + key_len;
800 if (!isspace(*body))
801 return -EINVAL;
802 body = skip_spaces(body);
803
804 disk = get_gendisk(MKDEV(major, minor), &part);
805 if (!disk)
806 return -ENODEV;
807 if (part) {
808 owner = disk->fops->owner;
809 put_disk(disk);
810 module_put(owner);
811 return -ENODEV;
812 }
813
814 rcu_read_lock();
815 spin_lock_irq(disk->queue->queue_lock);
816
817 if (blkcg_policy_enabled(disk->queue, pol))
818 blkg = blkg_lookup_create(blkcg, disk->queue);
819 else
820 blkg = ERR_PTR(-EOPNOTSUPP);
821
822 if (IS_ERR(blkg)) {
823 ret = PTR_ERR(blkg);
824 rcu_read_unlock();
825 spin_unlock_irq(disk->queue->queue_lock);
826 owner = disk->fops->owner;
827 put_disk(disk);
828 module_put(owner);
829 /*
830 * If queue was bypassing, we should retry. Do so after a
831 * short msleep(). It isn't strictly necessary but queue
832 * can be bypassing for some time and it's always nice to
833 * avoid busy looping.
834 */
835 if (ret == -EBUSY) {
836 msleep(10);
837 ret = restart_syscall();
838 }
839 return ret;
840 }
841
842 ctx->disk = disk;
843 ctx->blkg = blkg;
844 ctx->body = body;
845 return 0;
846}
847EXPORT_SYMBOL_GPL(blkg_conf_prep);
848
849/**
850 * blkg_conf_finish - finish up per-blkg config update
851 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
852 *
853 * Finish up after per-blkg config update. This function must be paired
854 * with blkg_conf_prep().
855 */
856void blkg_conf_finish(struct blkg_conf_ctx *ctx)
857 __releases(ctx->disk->queue->queue_lock) __releases(rcu)
858{
859 struct module *owner;
860
861 spin_unlock_irq(ctx->disk->queue->queue_lock);
862 rcu_read_unlock();
863 owner = ctx->disk->fops->owner;
864 put_disk(ctx->disk);
865 module_put(owner);
866}
867EXPORT_SYMBOL_GPL(blkg_conf_finish);
868
869static int blkcg_print_stat(struct seq_file *sf, void *v)
870{
871 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
872 struct blkcg_gq *blkg;
873
874 rcu_read_lock();
875
876 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
877 const char *dname;
878 struct blkg_rwstat rwstat;
879 u64 rbytes, wbytes, rios, wios;
880
881 dname = blkg_dev_name(blkg);
882 if (!dname)
883 continue;
884
885 spin_lock_irq(blkg->q->queue_lock);
886
887 rwstat = blkg_rwstat_recursive_sum(blkg, NULL,
888 offsetof(struct blkcg_gq, stat_bytes));
889 rbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_READ]);
890 wbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_WRITE]);
891
892 rwstat = blkg_rwstat_recursive_sum(blkg, NULL,
893 offsetof(struct blkcg_gq, stat_ios));
894 rios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_READ]);
895 wios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_WRITE]);
896
897 spin_unlock_irq(blkg->q->queue_lock);
898
899 if (rbytes || wbytes || rios || wios)
900 seq_printf(sf, "%s rbytes=%llu wbytes=%llu rios=%llu wios=%llu\n",
901 dname, rbytes, wbytes, rios, wios);
902 }
903
904 rcu_read_unlock();
905 return 0;
906}
907
908struct cftype blkcg_files[] = {
909 {
910 .name = "stat",
911 .flags = CFTYPE_NOT_ON_ROOT,
912 .seq_show = blkcg_print_stat,
913 },
914 { } /* terminate */
915};
916
917struct cftype blkcg_legacy_files[] = {
918 {
919 .name = "reset_stats",
920 .write_u64 = blkcg_reset_stats,
921 },
922 { } /* terminate */
923};
924
925/**
926 * blkcg_css_offline - cgroup css_offline callback
927 * @css: css of interest
928 *
929 * This function is called when @css is about to go away and responsible
930 * for shooting down all blkgs associated with @css. blkgs should be
931 * removed while holding both q and blkcg locks. As blkcg lock is nested
932 * inside q lock, this function performs reverse double lock dancing.
933 *
934 * This is the blkcg counterpart of ioc_release_fn().
935 */
936static void blkcg_css_offline(struct cgroup_subsys_state *css)
937{
938 struct blkcg *blkcg = css_to_blkcg(css);
939
940 spin_lock_irq(&blkcg->lock);
941
942 while (!hlist_empty(&blkcg->blkg_list)) {
943 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
944 struct blkcg_gq, blkcg_node);
945 struct request_queue *q = blkg->q;
946
947 if (spin_trylock(q->queue_lock)) {
948 blkg_destroy(blkg);
949 spin_unlock(q->queue_lock);
950 } else {
951 spin_unlock_irq(&blkcg->lock);
952 cpu_relax();
953 spin_lock_irq(&blkcg->lock);
954 }
955 }
956
957 spin_unlock_irq(&blkcg->lock);
958
959 wb_blkcg_offline(blkcg);
960}
961
962static void blkcg_css_free(struct cgroup_subsys_state *css)
963{
964 struct blkcg *blkcg = css_to_blkcg(css);
965 int i;
966
967 mutex_lock(&blkcg_pol_mutex);
968
969 list_del(&blkcg->all_blkcgs_node);
970
971 for (i = 0; i < BLKCG_MAX_POLS; i++)
972 if (blkcg->cpd[i])
973 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
974
975 mutex_unlock(&blkcg_pol_mutex);
976
977 kfree(blkcg);
978}
979
980static struct cgroup_subsys_state *
981blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
982{
983 struct blkcg *blkcg;
984 struct cgroup_subsys_state *ret;
985 int i;
986
987 mutex_lock(&blkcg_pol_mutex);
988
989 if (!parent_css) {
990 blkcg = &blkcg_root;
991 } else {
992 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
993 if (!blkcg) {
994 ret = ERR_PTR(-ENOMEM);
995 goto free_blkcg;
996 }
997 }
998
999 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1000 struct blkcg_policy *pol = blkcg_policy[i];
1001 struct blkcg_policy_data *cpd;
1002
1003 /*
1004 * If the policy hasn't been attached yet, wait for it
1005 * to be attached before doing anything else. Otherwise,
1006 * check if the policy requires any specific per-cgroup
1007 * data: if it does, allocate and initialize it.
1008 */
1009 if (!pol || !pol->cpd_alloc_fn)
1010 continue;
1011
1012 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1013 if (!cpd) {
1014 ret = ERR_PTR(-ENOMEM);
1015 goto free_pd_blkcg;
1016 }
1017 blkcg->cpd[i] = cpd;
1018 cpd->blkcg = blkcg;
1019 cpd->plid = i;
1020 if (pol->cpd_init_fn)
1021 pol->cpd_init_fn(cpd);
1022 }
1023
1024 spin_lock_init(&blkcg->lock);
1025 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT);
1026 INIT_HLIST_HEAD(&blkcg->blkg_list);
1027#ifdef CONFIG_CGROUP_WRITEBACK
1028 INIT_LIST_HEAD(&blkcg->cgwb_list);
1029#endif
1030 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1031
1032 mutex_unlock(&blkcg_pol_mutex);
1033 return &blkcg->css;
1034
1035free_pd_blkcg:
1036 for (i--; i >= 0; i--)
1037 if (blkcg->cpd[i])
1038 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1039free_blkcg:
1040 kfree(blkcg);
1041 mutex_unlock(&blkcg_pol_mutex);
1042 return ret;
1043}
1044
1045/**
1046 * blkcg_init_queue - initialize blkcg part of request queue
1047 * @q: request_queue to initialize
1048 *
1049 * Called from blk_alloc_queue_node(). Responsible for initializing blkcg
1050 * part of new request_queue @q.
1051 *
1052 * RETURNS:
1053 * 0 on success, -errno on failure.
1054 */
1055int blkcg_init_queue(struct request_queue *q)
1056{
1057 struct blkcg_gq *new_blkg, *blkg;
1058 bool preloaded;
1059 int ret;
1060
1061 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1062 if (!new_blkg)
1063 return -ENOMEM;
1064
1065 preloaded = !radix_tree_preload(GFP_KERNEL);
1066
1067 /*
1068 * Make sure the root blkg exists and count the existing blkgs. As
1069 * @q is bypassing at this point, blkg_lookup_create() can't be
1070 * used. Open code insertion.
1071 */
1072 rcu_read_lock();
1073 spin_lock_irq(q->queue_lock);
1074 blkg = blkg_create(&blkcg_root, q, new_blkg);
1075 spin_unlock_irq(q->queue_lock);
1076 rcu_read_unlock();
1077
1078 if (preloaded)
1079 radix_tree_preload_end();
1080
1081 if (IS_ERR(blkg)) {
1082 blkg_free(new_blkg);
1083 return PTR_ERR(blkg);
1084 }
1085
1086 q->root_blkg = blkg;
1087 q->root_rl.blkg = blkg;
1088
1089 ret = blk_throtl_init(q);
1090 if (ret) {
1091 spin_lock_irq(q->queue_lock);
1092 blkg_destroy_all(q);
1093 spin_unlock_irq(q->queue_lock);
1094 }
1095 return ret;
1096}
1097
1098/**
1099 * blkcg_drain_queue - drain blkcg part of request_queue
1100 * @q: request_queue to drain
1101 *
1102 * Called from blk_drain_queue(). Responsible for draining blkcg part.
1103 */
1104void blkcg_drain_queue(struct request_queue *q)
1105{
1106 lockdep_assert_held(q->queue_lock);
1107
1108 /*
1109 * @q could be exiting and already have destroyed all blkgs as
1110 * indicated by NULL root_blkg. If so, don't confuse policies.
1111 */
1112 if (!q->root_blkg)
1113 return;
1114
1115 blk_throtl_drain(q);
1116}
1117
1118/**
1119 * blkcg_exit_queue - exit and release blkcg part of request_queue
1120 * @q: request_queue being released
1121 *
1122 * Called from blk_release_queue(). Responsible for exiting blkcg part.
1123 */
1124void blkcg_exit_queue(struct request_queue *q)
1125{
1126 spin_lock_irq(q->queue_lock);
1127 blkg_destroy_all(q);
1128 spin_unlock_irq(q->queue_lock);
1129
1130 blk_throtl_exit(q);
1131}
1132
1133/*
1134 * We cannot support shared io contexts, as we have no mean to support
1135 * two tasks with the same ioc in two different groups without major rework
1136 * of the main cic data structures. For now we allow a task to change
1137 * its cgroup only if it's the only owner of its ioc.
1138 */
1139static int blkcg_can_attach(struct cgroup_taskset *tset)
1140{
1141 struct task_struct *task;
1142 struct cgroup_subsys_state *dst_css;
1143 struct io_context *ioc;
1144 int ret = 0;
1145
1146 /* task_lock() is needed to avoid races with exit_io_context() */
1147 cgroup_taskset_for_each(task, dst_css, tset) {
1148 task_lock(task);
1149 ioc = task->io_context;
1150 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
1151 ret = -EINVAL;
1152 task_unlock(task);
1153 if (ret)
1154 break;
1155 }
1156 return ret;
1157}
1158
1159static void blkcg_bind(struct cgroup_subsys_state *root_css)
1160{
1161 int i;
1162
1163 mutex_lock(&blkcg_pol_mutex);
1164
1165 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1166 struct blkcg_policy *pol = blkcg_policy[i];
1167 struct blkcg *blkcg;
1168
1169 if (!pol || !pol->cpd_bind_fn)
1170 continue;
1171
1172 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1173 if (blkcg->cpd[pol->plid])
1174 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1175 }
1176 mutex_unlock(&blkcg_pol_mutex);
1177}
1178
1179struct cgroup_subsys io_cgrp_subsys = {
1180 .css_alloc = blkcg_css_alloc,
1181 .css_offline = blkcg_css_offline,
1182 .css_free = blkcg_css_free,
1183 .can_attach = blkcg_can_attach,
1184 .bind = blkcg_bind,
1185 .dfl_cftypes = blkcg_files,
1186 .legacy_cftypes = blkcg_legacy_files,
1187 .legacy_name = "blkio",
1188#ifdef CONFIG_MEMCG
1189 /*
1190 * This ensures that, if available, memcg is automatically enabled
1191 * together on the default hierarchy so that the owner cgroup can
1192 * be retrieved from writeback pages.
1193 */
1194 .depends_on = 1 << memory_cgrp_id,
1195#endif
1196};
1197EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1198
1199/**
1200 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1201 * @q: request_queue of interest
1202 * @pol: blkcg policy to activate
1203 *
1204 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1205 * bypass mode to populate its blkgs with policy_data for @pol.
1206 *
1207 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1208 * from IO path. Update of each blkg is protected by both queue and blkcg
1209 * locks so that holding either lock and testing blkcg_policy_enabled() is
1210 * always enough for dereferencing policy data.
1211 *
1212 * The caller is responsible for synchronizing [de]activations and policy
1213 * [un]registerations. Returns 0 on success, -errno on failure.
1214 */
1215int blkcg_activate_policy(struct request_queue *q,
1216 const struct blkcg_policy *pol)
1217{
1218 struct blkg_policy_data *pd_prealloc = NULL;
1219 struct blkcg_gq *blkg;
1220 int ret;
1221
1222 if (blkcg_policy_enabled(q, pol))
1223 return 0;
1224
1225 blk_queue_bypass_start(q);
1226pd_prealloc:
1227 if (!pd_prealloc) {
1228 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q->node);
1229 if (!pd_prealloc) {
1230 ret = -ENOMEM;
1231 goto out_bypass_end;
1232 }
1233 }
1234
1235 spin_lock_irq(q->queue_lock);
1236
1237 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1238 struct blkg_policy_data *pd;
1239
1240 if (blkg->pd[pol->plid])
1241 continue;
1242
1243 pd = pol->pd_alloc_fn(GFP_NOWAIT, q->node);
1244 if (!pd)
1245 swap(pd, pd_prealloc);
1246 if (!pd) {
1247 spin_unlock_irq(q->queue_lock);
1248 goto pd_prealloc;
1249 }
1250
1251 blkg->pd[pol->plid] = pd;
1252 pd->blkg = blkg;
1253 pd->plid = pol->plid;
1254 if (pol->pd_init_fn)
1255 pol->pd_init_fn(pd);
1256 }
1257
1258 __set_bit(pol->plid, q->blkcg_pols);
1259 ret = 0;
1260
1261 spin_unlock_irq(q->queue_lock);
1262out_bypass_end:
1263 blk_queue_bypass_end(q);
1264 if (pd_prealloc)
1265 pol->pd_free_fn(pd_prealloc);
1266 return ret;
1267}
1268EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1269
1270/**
1271 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1272 * @q: request_queue of interest
1273 * @pol: blkcg policy to deactivate
1274 *
1275 * Deactivate @pol on @q. Follows the same synchronization rules as
1276 * blkcg_activate_policy().
1277 */
1278void blkcg_deactivate_policy(struct request_queue *q,
1279 const struct blkcg_policy *pol)
1280{
1281 struct blkcg_gq *blkg;
1282
1283 if (!blkcg_policy_enabled(q, pol))
1284 return;
1285
1286 blk_queue_bypass_start(q);
1287 spin_lock_irq(q->queue_lock);
1288
1289 __clear_bit(pol->plid, q->blkcg_pols);
1290
1291 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1292 /* grab blkcg lock too while removing @pd from @blkg */
1293 spin_lock(&blkg->blkcg->lock);
1294
1295 if (blkg->pd[pol->plid]) {
1296 if (pol->pd_offline_fn)
1297 pol->pd_offline_fn(blkg->pd[pol->plid]);
1298 pol->pd_free_fn(blkg->pd[pol->plid]);
1299 blkg->pd[pol->plid] = NULL;
1300 }
1301
1302 spin_unlock(&blkg->blkcg->lock);
1303 }
1304
1305 spin_unlock_irq(q->queue_lock);
1306 blk_queue_bypass_end(q);
1307}
1308EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1309
1310/**
1311 * blkcg_policy_register - register a blkcg policy
1312 * @pol: blkcg policy to register
1313 *
1314 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1315 * successful registration. Returns 0 on success and -errno on failure.
1316 */
1317int blkcg_policy_register(struct blkcg_policy *pol)
1318{
1319 struct blkcg *blkcg;
1320 int i, ret;
1321
1322 mutex_lock(&blkcg_pol_register_mutex);
1323 mutex_lock(&blkcg_pol_mutex);
1324
1325 /* find an empty slot */
1326 ret = -ENOSPC;
1327 for (i = 0; i < BLKCG_MAX_POLS; i++)
1328 if (!blkcg_policy[i])
1329 break;
1330 if (i >= BLKCG_MAX_POLS)
1331 goto err_unlock;
1332
1333 /* register @pol */
1334 pol->plid = i;
1335 blkcg_policy[pol->plid] = pol;
1336
1337 /* allocate and install cpd's */
1338 if (pol->cpd_alloc_fn) {
1339 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1340 struct blkcg_policy_data *cpd;
1341
1342 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1343 if (!cpd) {
1344 mutex_unlock(&blkcg_pol_mutex);
1345 goto err_free_cpds;
1346 }
1347
1348 blkcg->cpd[pol->plid] = cpd;
1349 cpd->blkcg = blkcg;
1350 cpd->plid = pol->plid;
1351 pol->cpd_init_fn(cpd);
1352 }
1353 }
1354
1355 mutex_unlock(&blkcg_pol_mutex);
1356
1357 /* everything is in place, add intf files for the new policy */
1358 if (pol->dfl_cftypes)
1359 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1360 pol->dfl_cftypes));
1361 if (pol->legacy_cftypes)
1362 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1363 pol->legacy_cftypes));
1364 mutex_unlock(&blkcg_pol_register_mutex);
1365 return 0;
1366
1367err_free_cpds:
1368 if (pol->cpd_alloc_fn) {
1369 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1370 if (blkcg->cpd[pol->plid]) {
1371 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1372 blkcg->cpd[pol->plid] = NULL;
1373 }
1374 }
1375 }
1376 blkcg_policy[pol->plid] = NULL;
1377err_unlock:
1378 mutex_unlock(&blkcg_pol_mutex);
1379 mutex_unlock(&blkcg_pol_register_mutex);
1380 return ret;
1381}
1382EXPORT_SYMBOL_GPL(blkcg_policy_register);
1383
1384/**
1385 * blkcg_policy_unregister - unregister a blkcg policy
1386 * @pol: blkcg policy to unregister
1387 *
1388 * Undo blkcg_policy_register(@pol). Might sleep.
1389 */
1390void blkcg_policy_unregister(struct blkcg_policy *pol)
1391{
1392 struct blkcg *blkcg;
1393
1394 mutex_lock(&blkcg_pol_register_mutex);
1395
1396 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1397 goto out_unlock;
1398
1399 /* kill the intf files first */
1400 if (pol->dfl_cftypes)
1401 cgroup_rm_cftypes(pol->dfl_cftypes);
1402 if (pol->legacy_cftypes)
1403 cgroup_rm_cftypes(pol->legacy_cftypes);
1404
1405 /* remove cpds and unregister */
1406 mutex_lock(&blkcg_pol_mutex);
1407
1408 if (pol->cpd_alloc_fn) {
1409 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1410 if (blkcg->cpd[pol->plid]) {
1411 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1412 blkcg->cpd[pol->plid] = NULL;
1413 }
1414 }
1415 }
1416 blkcg_policy[pol->plid] = NULL;
1417
1418 mutex_unlock(&blkcg_pol_mutex);
1419out_unlock:
1420 mutex_unlock(&blkcg_pol_register_mutex);
1421}
1422EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Common Block IO controller cgroup interface
4 *
5 * Based on ideas and code from CFQ, CFS and BFQ:
6 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
7 *
8 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
9 * Paolo Valente <paolo.valente@unimore.it>
10 *
11 * Copyright (C) 2009 Vivek Goyal <vgoyal@redhat.com>
12 * Nauman Rafique <nauman@google.com>
13 *
14 * For policy-specific per-blkcg data:
15 * Copyright (C) 2015 Paolo Valente <paolo.valente@unimore.it>
16 * Arianna Avanzini <avanzini.arianna@gmail.com>
17 */
18#include <linux/ioprio.h>
19#include <linux/kdev_t.h>
20#include <linux/module.h>
21#include <linux/sched/signal.h>
22#include <linux/err.h>
23#include <linux/blkdev.h>
24#include <linux/backing-dev.h>
25#include <linux/slab.h>
26#include <linux/delay.h>
27#include <linux/atomic.h>
28#include <linux/ctype.h>
29#include <linux/resume_user_mode.h>
30#include <linux/psi.h>
31#include <linux/part_stat.h>
32#include "blk.h"
33#include "blk-cgroup.h"
34#include "blk-ioprio.h"
35#include "blk-throttle.h"
36
37static void __blkcg_rstat_flush(struct blkcg *blkcg, int cpu);
38
39/*
40 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
41 * blkcg_pol_register_mutex nests outside of it and synchronizes entire
42 * policy [un]register operations including cgroup file additions /
43 * removals. Putting cgroup file registration outside blkcg_pol_mutex
44 * allows grabbing it from cgroup callbacks.
45 */
46static DEFINE_MUTEX(blkcg_pol_register_mutex);
47static DEFINE_MUTEX(blkcg_pol_mutex);
48
49struct blkcg blkcg_root;
50EXPORT_SYMBOL_GPL(blkcg_root);
51
52struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
53EXPORT_SYMBOL_GPL(blkcg_root_css);
54
55static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
56
57static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */
58
59bool blkcg_debug_stats = false;
60
61static DEFINE_RAW_SPINLOCK(blkg_stat_lock);
62
63#define BLKG_DESTROY_BATCH_SIZE 64
64
65/*
66 * Lockless lists for tracking IO stats update
67 *
68 * New IO stats are stored in the percpu iostat_cpu within blkcg_gq (blkg).
69 * There are multiple blkg's (one for each block device) attached to each
70 * blkcg. The rstat code keeps track of which cpu has IO stats updated,
71 * but it doesn't know which blkg has the updated stats. If there are many
72 * block devices in a system, the cost of iterating all the blkg's to flush
73 * out the IO stats can be high. To reduce such overhead, a set of percpu
74 * lockless lists (lhead) per blkcg are used to track the set of recently
75 * updated iostat_cpu's since the last flush. An iostat_cpu will be put
76 * onto the lockless list on the update side [blk_cgroup_bio_start()] if
77 * not there yet and then removed when being flushed [blkcg_rstat_flush()].
78 * References to blkg are gotten and then put back in the process to
79 * protect against blkg removal.
80 *
81 * Return: 0 if successful or -ENOMEM if allocation fails.
82 */
83static int init_blkcg_llists(struct blkcg *blkcg)
84{
85 int cpu;
86
87 blkcg->lhead = alloc_percpu_gfp(struct llist_head, GFP_KERNEL);
88 if (!blkcg->lhead)
89 return -ENOMEM;
90
91 for_each_possible_cpu(cpu)
92 init_llist_head(per_cpu_ptr(blkcg->lhead, cpu));
93 return 0;
94}
95
96/**
97 * blkcg_css - find the current css
98 *
99 * Find the css associated with either the kthread or the current task.
100 * This may return a dying css, so it is up to the caller to use tryget logic
101 * to confirm it is alive and well.
102 */
103static struct cgroup_subsys_state *blkcg_css(void)
104{
105 struct cgroup_subsys_state *css;
106
107 css = kthread_blkcg();
108 if (css)
109 return css;
110 return task_css(current, io_cgrp_id);
111}
112
113static bool blkcg_policy_enabled(struct request_queue *q,
114 const struct blkcg_policy *pol)
115{
116 return pol && test_bit(pol->plid, q->blkcg_pols);
117}
118
119static void blkg_free_workfn(struct work_struct *work)
120{
121 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
122 free_work);
123 struct request_queue *q = blkg->q;
124 int i;
125
126 /*
127 * pd_free_fn() can also be called from blkcg_deactivate_policy(),
128 * in order to make sure pd_free_fn() is called in order, the deletion
129 * of the list blkg->q_node is delayed to here from blkg_destroy(), and
130 * blkcg_mutex is used to synchronize blkg_free_workfn() and
131 * blkcg_deactivate_policy().
132 */
133 mutex_lock(&q->blkcg_mutex);
134 for (i = 0; i < BLKCG_MAX_POLS; i++)
135 if (blkg->pd[i])
136 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
137 if (blkg->parent)
138 blkg_put(blkg->parent);
139 spin_lock_irq(&q->queue_lock);
140 list_del_init(&blkg->q_node);
141 spin_unlock_irq(&q->queue_lock);
142 mutex_unlock(&q->blkcg_mutex);
143
144 blk_put_queue(q);
145 free_percpu(blkg->iostat_cpu);
146 percpu_ref_exit(&blkg->refcnt);
147 kfree(blkg);
148}
149
150/**
151 * blkg_free - free a blkg
152 * @blkg: blkg to free
153 *
154 * Free @blkg which may be partially allocated.
155 */
156static void blkg_free(struct blkcg_gq *blkg)
157{
158 if (!blkg)
159 return;
160
161 /*
162 * Both ->pd_free_fn() and request queue's release handler may
163 * sleep, so free us by scheduling one work func
164 */
165 INIT_WORK(&blkg->free_work, blkg_free_workfn);
166 schedule_work(&blkg->free_work);
167}
168
169static void __blkg_release(struct rcu_head *rcu)
170{
171 struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head);
172 struct blkcg *blkcg = blkg->blkcg;
173 int cpu;
174
175#ifdef CONFIG_BLK_CGROUP_PUNT_BIO
176 WARN_ON(!bio_list_empty(&blkg->async_bios));
177#endif
178 /*
179 * Flush all the non-empty percpu lockless lists before releasing
180 * us, given these stat belongs to us.
181 *
182 * blkg_stat_lock is for serializing blkg stat update
183 */
184 for_each_possible_cpu(cpu)
185 __blkcg_rstat_flush(blkcg, cpu);
186
187 /* release the blkcg and parent blkg refs this blkg has been holding */
188 css_put(&blkg->blkcg->css);
189 blkg_free(blkg);
190}
191
192/*
193 * A group is RCU protected, but having an rcu lock does not mean that one
194 * can access all the fields of blkg and assume these are valid. For
195 * example, don't try to follow throtl_data and request queue links.
196 *
197 * Having a reference to blkg under an rcu allows accesses to only values
198 * local to groups like group stats and group rate limits.
199 */
200static void blkg_release(struct percpu_ref *ref)
201{
202 struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt);
203
204 call_rcu(&blkg->rcu_head, __blkg_release);
205}
206
207#ifdef CONFIG_BLK_CGROUP_PUNT_BIO
208static struct workqueue_struct *blkcg_punt_bio_wq;
209
210static void blkg_async_bio_workfn(struct work_struct *work)
211{
212 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
213 async_bio_work);
214 struct bio_list bios = BIO_EMPTY_LIST;
215 struct bio *bio;
216 struct blk_plug plug;
217 bool need_plug = false;
218
219 /* as long as there are pending bios, @blkg can't go away */
220 spin_lock(&blkg->async_bio_lock);
221 bio_list_merge(&bios, &blkg->async_bios);
222 bio_list_init(&blkg->async_bios);
223 spin_unlock(&blkg->async_bio_lock);
224
225 /* start plug only when bio_list contains at least 2 bios */
226 if (bios.head && bios.head->bi_next) {
227 need_plug = true;
228 blk_start_plug(&plug);
229 }
230 while ((bio = bio_list_pop(&bios)))
231 submit_bio(bio);
232 if (need_plug)
233 blk_finish_plug(&plug);
234}
235
236/*
237 * When a shared kthread issues a bio for a cgroup, doing so synchronously can
238 * lead to priority inversions as the kthread can be trapped waiting for that
239 * cgroup. Use this helper instead of submit_bio to punt the actual issuing to
240 * a dedicated per-blkcg work item to avoid such priority inversions.
241 */
242void blkcg_punt_bio_submit(struct bio *bio)
243{
244 struct blkcg_gq *blkg = bio->bi_blkg;
245
246 if (blkg->parent) {
247 spin_lock(&blkg->async_bio_lock);
248 bio_list_add(&blkg->async_bios, bio);
249 spin_unlock(&blkg->async_bio_lock);
250 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
251 } else {
252 /* never bounce for the root cgroup */
253 submit_bio(bio);
254 }
255}
256EXPORT_SYMBOL_GPL(blkcg_punt_bio_submit);
257
258static int __init blkcg_punt_bio_init(void)
259{
260 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
261 WQ_MEM_RECLAIM | WQ_FREEZABLE |
262 WQ_UNBOUND | WQ_SYSFS, 0);
263 if (!blkcg_punt_bio_wq)
264 return -ENOMEM;
265 return 0;
266}
267subsys_initcall(blkcg_punt_bio_init);
268#endif /* CONFIG_BLK_CGROUP_PUNT_BIO */
269
270/**
271 * bio_blkcg_css - return the blkcg CSS associated with a bio
272 * @bio: target bio
273 *
274 * This returns the CSS for the blkcg associated with a bio, or %NULL if not
275 * associated. Callers are expected to either handle %NULL or know association
276 * has been done prior to calling this.
277 */
278struct cgroup_subsys_state *bio_blkcg_css(struct bio *bio)
279{
280 if (!bio || !bio->bi_blkg)
281 return NULL;
282 return &bio->bi_blkg->blkcg->css;
283}
284EXPORT_SYMBOL_GPL(bio_blkcg_css);
285
286/**
287 * blkcg_parent - get the parent of a blkcg
288 * @blkcg: blkcg of interest
289 *
290 * Return the parent blkcg of @blkcg. Can be called anytime.
291 */
292static inline struct blkcg *blkcg_parent(struct blkcg *blkcg)
293{
294 return css_to_blkcg(blkcg->css.parent);
295}
296
297/**
298 * blkg_alloc - allocate a blkg
299 * @blkcg: block cgroup the new blkg is associated with
300 * @disk: gendisk the new blkg is associated with
301 * @gfp_mask: allocation mask to use
302 *
303 * Allocate a new blkg associating @blkcg and @disk.
304 */
305static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct gendisk *disk,
306 gfp_t gfp_mask)
307{
308 struct blkcg_gq *blkg;
309 int i, cpu;
310
311 /* alloc and init base part */
312 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, disk->queue->node);
313 if (!blkg)
314 return NULL;
315 if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask))
316 goto out_free_blkg;
317 blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask);
318 if (!blkg->iostat_cpu)
319 goto out_exit_refcnt;
320 if (!blk_get_queue(disk->queue))
321 goto out_free_iostat;
322
323 blkg->q = disk->queue;
324 INIT_LIST_HEAD(&blkg->q_node);
325 blkg->blkcg = blkcg;
326 blkg->iostat.blkg = blkg;
327#ifdef CONFIG_BLK_CGROUP_PUNT_BIO
328 spin_lock_init(&blkg->async_bio_lock);
329 bio_list_init(&blkg->async_bios);
330 INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
331#endif
332
333 u64_stats_init(&blkg->iostat.sync);
334 for_each_possible_cpu(cpu) {
335 u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync);
336 per_cpu_ptr(blkg->iostat_cpu, cpu)->blkg = blkg;
337 }
338
339 for (i = 0; i < BLKCG_MAX_POLS; i++) {
340 struct blkcg_policy *pol = blkcg_policy[i];
341 struct blkg_policy_data *pd;
342
343 if (!blkcg_policy_enabled(disk->queue, pol))
344 continue;
345
346 /* alloc per-policy data and attach it to blkg */
347 pd = pol->pd_alloc_fn(disk, blkcg, gfp_mask);
348 if (!pd)
349 goto out_free_pds;
350 blkg->pd[i] = pd;
351 pd->blkg = blkg;
352 pd->plid = i;
353 pd->online = false;
354 }
355
356 return blkg;
357
358out_free_pds:
359 while (--i >= 0)
360 if (blkg->pd[i])
361 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
362 blk_put_queue(disk->queue);
363out_free_iostat:
364 free_percpu(blkg->iostat_cpu);
365out_exit_refcnt:
366 percpu_ref_exit(&blkg->refcnt);
367out_free_blkg:
368 kfree(blkg);
369 return NULL;
370}
371
372/*
373 * If @new_blkg is %NULL, this function tries to allocate a new one as
374 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
375 */
376static struct blkcg_gq *blkg_create(struct blkcg *blkcg, struct gendisk *disk,
377 struct blkcg_gq *new_blkg)
378{
379 struct blkcg_gq *blkg;
380 int i, ret;
381
382 lockdep_assert_held(&disk->queue->queue_lock);
383
384 /* request_queue is dying, do not create/recreate a blkg */
385 if (blk_queue_dying(disk->queue)) {
386 ret = -ENODEV;
387 goto err_free_blkg;
388 }
389
390 /* blkg holds a reference to blkcg */
391 if (!css_tryget_online(&blkcg->css)) {
392 ret = -ENODEV;
393 goto err_free_blkg;
394 }
395
396 /* allocate */
397 if (!new_blkg) {
398 new_blkg = blkg_alloc(blkcg, disk, GFP_NOWAIT | __GFP_NOWARN);
399 if (unlikely(!new_blkg)) {
400 ret = -ENOMEM;
401 goto err_put_css;
402 }
403 }
404 blkg = new_blkg;
405
406 /* link parent */
407 if (blkcg_parent(blkcg)) {
408 blkg->parent = blkg_lookup(blkcg_parent(blkcg), disk->queue);
409 if (WARN_ON_ONCE(!blkg->parent)) {
410 ret = -ENODEV;
411 goto err_put_css;
412 }
413 blkg_get(blkg->parent);
414 }
415
416 /* invoke per-policy init */
417 for (i = 0; i < BLKCG_MAX_POLS; i++) {
418 struct blkcg_policy *pol = blkcg_policy[i];
419
420 if (blkg->pd[i] && pol->pd_init_fn)
421 pol->pd_init_fn(blkg->pd[i]);
422 }
423
424 /* insert */
425 spin_lock(&blkcg->lock);
426 ret = radix_tree_insert(&blkcg->blkg_tree, disk->queue->id, blkg);
427 if (likely(!ret)) {
428 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
429 list_add(&blkg->q_node, &disk->queue->blkg_list);
430
431 for (i = 0; i < BLKCG_MAX_POLS; i++) {
432 struct blkcg_policy *pol = blkcg_policy[i];
433
434 if (blkg->pd[i]) {
435 if (pol->pd_online_fn)
436 pol->pd_online_fn(blkg->pd[i]);
437 blkg->pd[i]->online = true;
438 }
439 }
440 }
441 blkg->online = true;
442 spin_unlock(&blkcg->lock);
443
444 if (!ret)
445 return blkg;
446
447 /* @blkg failed fully initialized, use the usual release path */
448 blkg_put(blkg);
449 return ERR_PTR(ret);
450
451err_put_css:
452 css_put(&blkcg->css);
453err_free_blkg:
454 if (new_blkg)
455 blkg_free(new_blkg);
456 return ERR_PTR(ret);
457}
458
459/**
460 * blkg_lookup_create - lookup blkg, try to create one if not there
461 * @blkcg: blkcg of interest
462 * @disk: gendisk of interest
463 *
464 * Lookup blkg for the @blkcg - @disk pair. If it doesn't exist, try to
465 * create one. blkg creation is performed recursively from blkcg_root such
466 * that all non-root blkg's have access to the parent blkg. This function
467 * should be called under RCU read lock and takes @disk->queue->queue_lock.
468 *
469 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
470 * down from root.
471 */
472static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
473 struct gendisk *disk)
474{
475 struct request_queue *q = disk->queue;
476 struct blkcg_gq *blkg;
477 unsigned long flags;
478
479 WARN_ON_ONCE(!rcu_read_lock_held());
480
481 blkg = blkg_lookup(blkcg, q);
482 if (blkg)
483 return blkg;
484
485 spin_lock_irqsave(&q->queue_lock, flags);
486 blkg = blkg_lookup(blkcg, q);
487 if (blkg) {
488 if (blkcg != &blkcg_root &&
489 blkg != rcu_dereference(blkcg->blkg_hint))
490 rcu_assign_pointer(blkcg->blkg_hint, blkg);
491 goto found;
492 }
493
494 /*
495 * Create blkgs walking down from blkcg_root to @blkcg, so that all
496 * non-root blkgs have access to their parents. Returns the closest
497 * blkg to the intended blkg should blkg_create() fail.
498 */
499 while (true) {
500 struct blkcg *pos = blkcg;
501 struct blkcg *parent = blkcg_parent(blkcg);
502 struct blkcg_gq *ret_blkg = q->root_blkg;
503
504 while (parent) {
505 blkg = blkg_lookup(parent, q);
506 if (blkg) {
507 /* remember closest blkg */
508 ret_blkg = blkg;
509 break;
510 }
511 pos = parent;
512 parent = blkcg_parent(parent);
513 }
514
515 blkg = blkg_create(pos, disk, NULL);
516 if (IS_ERR(blkg)) {
517 blkg = ret_blkg;
518 break;
519 }
520 if (pos == blkcg)
521 break;
522 }
523
524found:
525 spin_unlock_irqrestore(&q->queue_lock, flags);
526 return blkg;
527}
528
529static void blkg_destroy(struct blkcg_gq *blkg)
530{
531 struct blkcg *blkcg = blkg->blkcg;
532 int i;
533
534 lockdep_assert_held(&blkg->q->queue_lock);
535 lockdep_assert_held(&blkcg->lock);
536
537 /*
538 * blkg stays on the queue list until blkg_free_workfn(), see details in
539 * blkg_free_workfn(), hence this function can be called from
540 * blkcg_destroy_blkgs() first and again from blkg_destroy_all() before
541 * blkg_free_workfn().
542 */
543 if (hlist_unhashed(&blkg->blkcg_node))
544 return;
545
546 for (i = 0; i < BLKCG_MAX_POLS; i++) {
547 struct blkcg_policy *pol = blkcg_policy[i];
548
549 if (blkg->pd[i] && blkg->pd[i]->online) {
550 blkg->pd[i]->online = false;
551 if (pol->pd_offline_fn)
552 pol->pd_offline_fn(blkg->pd[i]);
553 }
554 }
555
556 blkg->online = false;
557
558 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
559 hlist_del_init_rcu(&blkg->blkcg_node);
560
561 /*
562 * Both setting lookup hint to and clearing it from @blkg are done
563 * under queue_lock. If it's not pointing to @blkg now, it never
564 * will. Hint assignment itself can race safely.
565 */
566 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
567 rcu_assign_pointer(blkcg->blkg_hint, NULL);
568
569 /*
570 * Put the reference taken at the time of creation so that when all
571 * queues are gone, group can be destroyed.
572 */
573 percpu_ref_kill(&blkg->refcnt);
574}
575
576static void blkg_destroy_all(struct gendisk *disk)
577{
578 struct request_queue *q = disk->queue;
579 struct blkcg_gq *blkg;
580 int count = BLKG_DESTROY_BATCH_SIZE;
581 int i;
582
583restart:
584 spin_lock_irq(&q->queue_lock);
585 list_for_each_entry(blkg, &q->blkg_list, q_node) {
586 struct blkcg *blkcg = blkg->blkcg;
587
588 if (hlist_unhashed(&blkg->blkcg_node))
589 continue;
590
591 spin_lock(&blkcg->lock);
592 blkg_destroy(blkg);
593 spin_unlock(&blkcg->lock);
594
595 /*
596 * in order to avoid holding the spin lock for too long, release
597 * it when a batch of blkgs are destroyed.
598 */
599 if (!(--count)) {
600 count = BLKG_DESTROY_BATCH_SIZE;
601 spin_unlock_irq(&q->queue_lock);
602 cond_resched();
603 goto restart;
604 }
605 }
606
607 /*
608 * Mark policy deactivated since policy offline has been done, and
609 * the free is scheduled, so future blkcg_deactivate_policy() can
610 * be bypassed
611 */
612 for (i = 0; i < BLKCG_MAX_POLS; i++) {
613 struct blkcg_policy *pol = blkcg_policy[i];
614
615 if (pol)
616 __clear_bit(pol->plid, q->blkcg_pols);
617 }
618
619 q->root_blkg = NULL;
620 spin_unlock_irq(&q->queue_lock);
621}
622
623static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
624{
625 int i;
626
627 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
628 dst->bytes[i] = src->bytes[i];
629 dst->ios[i] = src->ios[i];
630 }
631}
632
633static void __blkg_clear_stat(struct blkg_iostat_set *bis)
634{
635 struct blkg_iostat cur = {0};
636 unsigned long flags;
637
638 flags = u64_stats_update_begin_irqsave(&bis->sync);
639 blkg_iostat_set(&bis->cur, &cur);
640 blkg_iostat_set(&bis->last, &cur);
641 u64_stats_update_end_irqrestore(&bis->sync, flags);
642}
643
644static void blkg_clear_stat(struct blkcg_gq *blkg)
645{
646 int cpu;
647
648 for_each_possible_cpu(cpu) {
649 struct blkg_iostat_set *s = per_cpu_ptr(blkg->iostat_cpu, cpu);
650
651 __blkg_clear_stat(s);
652 }
653 __blkg_clear_stat(&blkg->iostat);
654}
655
656static int blkcg_reset_stats(struct cgroup_subsys_state *css,
657 struct cftype *cftype, u64 val)
658{
659 struct blkcg *blkcg = css_to_blkcg(css);
660 struct blkcg_gq *blkg;
661 int i;
662
663 mutex_lock(&blkcg_pol_mutex);
664 spin_lock_irq(&blkcg->lock);
665
666 /*
667 * Note that stat reset is racy - it doesn't synchronize against
668 * stat updates. This is a debug feature which shouldn't exist
669 * anyway. If you get hit by a race, retry.
670 */
671 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
672 blkg_clear_stat(blkg);
673 for (i = 0; i < BLKCG_MAX_POLS; i++) {
674 struct blkcg_policy *pol = blkcg_policy[i];
675
676 if (blkg->pd[i] && pol->pd_reset_stats_fn)
677 pol->pd_reset_stats_fn(blkg->pd[i]);
678 }
679 }
680
681 spin_unlock_irq(&blkcg->lock);
682 mutex_unlock(&blkcg_pol_mutex);
683 return 0;
684}
685
686const char *blkg_dev_name(struct blkcg_gq *blkg)
687{
688 if (!blkg->q->disk)
689 return NULL;
690 return bdi_dev_name(blkg->q->disk->bdi);
691}
692
693/**
694 * blkcg_print_blkgs - helper for printing per-blkg data
695 * @sf: seq_file to print to
696 * @blkcg: blkcg of interest
697 * @prfill: fill function to print out a blkg
698 * @pol: policy in question
699 * @data: data to be passed to @prfill
700 * @show_total: to print out sum of prfill return values or not
701 *
702 * This function invokes @prfill on each blkg of @blkcg if pd for the
703 * policy specified by @pol exists. @prfill is invoked with @sf, the
704 * policy data and @data and the matching queue lock held. If @show_total
705 * is %true, the sum of the return values from @prfill is printed with
706 * "Total" label at the end.
707 *
708 * This is to be used to construct print functions for
709 * cftype->read_seq_string method.
710 */
711void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
712 u64 (*prfill)(struct seq_file *,
713 struct blkg_policy_data *, int),
714 const struct blkcg_policy *pol, int data,
715 bool show_total)
716{
717 struct blkcg_gq *blkg;
718 u64 total = 0;
719
720 rcu_read_lock();
721 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
722 spin_lock_irq(&blkg->q->queue_lock);
723 if (blkcg_policy_enabled(blkg->q, pol))
724 total += prfill(sf, blkg->pd[pol->plid], data);
725 spin_unlock_irq(&blkg->q->queue_lock);
726 }
727 rcu_read_unlock();
728
729 if (show_total)
730 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
731}
732EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
733
734/**
735 * __blkg_prfill_u64 - prfill helper for a single u64 value
736 * @sf: seq_file to print to
737 * @pd: policy private data of interest
738 * @v: value to print
739 *
740 * Print @v to @sf for the device associated with @pd.
741 */
742u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
743{
744 const char *dname = blkg_dev_name(pd->blkg);
745
746 if (!dname)
747 return 0;
748
749 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
750 return v;
751}
752EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
753
754/**
755 * blkg_conf_init - initialize a blkg_conf_ctx
756 * @ctx: blkg_conf_ctx to initialize
757 * @input: input string
758 *
759 * Initialize @ctx which can be used to parse blkg config input string @input.
760 * Once initialized, @ctx can be used with blkg_conf_open_bdev() and
761 * blkg_conf_prep(), and must be cleaned up with blkg_conf_exit().
762 */
763void blkg_conf_init(struct blkg_conf_ctx *ctx, char *input)
764{
765 *ctx = (struct blkg_conf_ctx){ .input = input };
766}
767EXPORT_SYMBOL_GPL(blkg_conf_init);
768
769/**
770 * blkg_conf_open_bdev - parse and open bdev for per-blkg config update
771 * @ctx: blkg_conf_ctx initialized with blkg_conf_init()
772 *
773 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update from
774 * @ctx->input and get and store the matching bdev in @ctx->bdev. @ctx->body is
775 * set to point past the device node prefix.
776 *
777 * This function may be called multiple times on @ctx and the extra calls become
778 * NOOPs. blkg_conf_prep() implicitly calls this function. Use this function
779 * explicitly if bdev access is needed without resolving the blkcg / policy part
780 * of @ctx->input. Returns -errno on error.
781 */
782int blkg_conf_open_bdev(struct blkg_conf_ctx *ctx)
783{
784 char *input = ctx->input;
785 unsigned int major, minor;
786 struct block_device *bdev;
787 int key_len;
788
789 if (ctx->bdev)
790 return 0;
791
792 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
793 return -EINVAL;
794
795 input += key_len;
796 if (!isspace(*input))
797 return -EINVAL;
798 input = skip_spaces(input);
799
800 bdev = blkdev_get_no_open(MKDEV(major, minor));
801 if (!bdev)
802 return -ENODEV;
803 if (bdev_is_partition(bdev)) {
804 blkdev_put_no_open(bdev);
805 return -ENODEV;
806 }
807
808 mutex_lock(&bdev->bd_queue->rq_qos_mutex);
809 if (!disk_live(bdev->bd_disk)) {
810 blkdev_put_no_open(bdev);
811 mutex_unlock(&bdev->bd_queue->rq_qos_mutex);
812 return -ENODEV;
813 }
814
815 ctx->body = input;
816 ctx->bdev = bdev;
817 return 0;
818}
819
820/**
821 * blkg_conf_prep - parse and prepare for per-blkg config update
822 * @blkcg: target block cgroup
823 * @pol: target policy
824 * @ctx: blkg_conf_ctx initialized with blkg_conf_init()
825 *
826 * Parse per-blkg config update from @ctx->input and initialize @ctx
827 * accordingly. On success, @ctx->body points to the part of @ctx->input
828 * following MAJ:MIN, @ctx->bdev points to the target block device and
829 * @ctx->blkg to the blkg being configured.
830 *
831 * blkg_conf_open_bdev() may be called on @ctx beforehand. On success, this
832 * function returns with queue lock held and must be followed by
833 * blkg_conf_exit().
834 */
835int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
836 struct blkg_conf_ctx *ctx)
837 __acquires(&bdev->bd_queue->queue_lock)
838{
839 struct gendisk *disk;
840 struct request_queue *q;
841 struct blkcg_gq *blkg;
842 int ret;
843
844 ret = blkg_conf_open_bdev(ctx);
845 if (ret)
846 return ret;
847
848 disk = ctx->bdev->bd_disk;
849 q = disk->queue;
850
851 /*
852 * blkcg_deactivate_policy() requires queue to be frozen, we can grab
853 * q_usage_counter to prevent concurrent with blkcg_deactivate_policy().
854 */
855 ret = blk_queue_enter(q, 0);
856 if (ret)
857 goto fail;
858
859 spin_lock_irq(&q->queue_lock);
860
861 if (!blkcg_policy_enabled(q, pol)) {
862 ret = -EOPNOTSUPP;
863 goto fail_unlock;
864 }
865
866 blkg = blkg_lookup(blkcg, q);
867 if (blkg)
868 goto success;
869
870 /*
871 * Create blkgs walking down from blkcg_root to @blkcg, so that all
872 * non-root blkgs have access to their parents.
873 */
874 while (true) {
875 struct blkcg *pos = blkcg;
876 struct blkcg *parent;
877 struct blkcg_gq *new_blkg;
878
879 parent = blkcg_parent(blkcg);
880 while (parent && !blkg_lookup(parent, q)) {
881 pos = parent;
882 parent = blkcg_parent(parent);
883 }
884
885 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
886 spin_unlock_irq(&q->queue_lock);
887
888 new_blkg = blkg_alloc(pos, disk, GFP_KERNEL);
889 if (unlikely(!new_blkg)) {
890 ret = -ENOMEM;
891 goto fail_exit_queue;
892 }
893
894 if (radix_tree_preload(GFP_KERNEL)) {
895 blkg_free(new_blkg);
896 ret = -ENOMEM;
897 goto fail_exit_queue;
898 }
899
900 spin_lock_irq(&q->queue_lock);
901
902 if (!blkcg_policy_enabled(q, pol)) {
903 blkg_free(new_blkg);
904 ret = -EOPNOTSUPP;
905 goto fail_preloaded;
906 }
907
908 blkg = blkg_lookup(pos, q);
909 if (blkg) {
910 blkg_free(new_blkg);
911 } else {
912 blkg = blkg_create(pos, disk, new_blkg);
913 if (IS_ERR(blkg)) {
914 ret = PTR_ERR(blkg);
915 goto fail_preloaded;
916 }
917 }
918
919 radix_tree_preload_end();
920
921 if (pos == blkcg)
922 goto success;
923 }
924success:
925 blk_queue_exit(q);
926 ctx->blkg = blkg;
927 return 0;
928
929fail_preloaded:
930 radix_tree_preload_end();
931fail_unlock:
932 spin_unlock_irq(&q->queue_lock);
933fail_exit_queue:
934 blk_queue_exit(q);
935fail:
936 /*
937 * If queue was bypassing, we should retry. Do so after a
938 * short msleep(). It isn't strictly necessary but queue
939 * can be bypassing for some time and it's always nice to
940 * avoid busy looping.
941 */
942 if (ret == -EBUSY) {
943 msleep(10);
944 ret = restart_syscall();
945 }
946 return ret;
947}
948EXPORT_SYMBOL_GPL(blkg_conf_prep);
949
950/**
951 * blkg_conf_exit - clean up per-blkg config update
952 * @ctx: blkg_conf_ctx initialized with blkg_conf_init()
953 *
954 * Clean up after per-blkg config update. This function must be called on all
955 * blkg_conf_ctx's initialized with blkg_conf_init().
956 */
957void blkg_conf_exit(struct blkg_conf_ctx *ctx)
958 __releases(&ctx->bdev->bd_queue->queue_lock)
959 __releases(&ctx->bdev->bd_queue->rq_qos_mutex)
960{
961 if (ctx->blkg) {
962 spin_unlock_irq(&bdev_get_queue(ctx->bdev)->queue_lock);
963 ctx->blkg = NULL;
964 }
965
966 if (ctx->bdev) {
967 mutex_unlock(&ctx->bdev->bd_queue->rq_qos_mutex);
968 blkdev_put_no_open(ctx->bdev);
969 ctx->body = NULL;
970 ctx->bdev = NULL;
971 }
972}
973EXPORT_SYMBOL_GPL(blkg_conf_exit);
974
975static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
976{
977 int i;
978
979 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
980 dst->bytes[i] += src->bytes[i];
981 dst->ios[i] += src->ios[i];
982 }
983}
984
985static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
986{
987 int i;
988
989 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
990 dst->bytes[i] -= src->bytes[i];
991 dst->ios[i] -= src->ios[i];
992 }
993}
994
995static void blkcg_iostat_update(struct blkcg_gq *blkg, struct blkg_iostat *cur,
996 struct blkg_iostat *last)
997{
998 struct blkg_iostat delta;
999 unsigned long flags;
1000
1001 /* propagate percpu delta to global */
1002 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
1003 blkg_iostat_set(&delta, cur);
1004 blkg_iostat_sub(&delta, last);
1005 blkg_iostat_add(&blkg->iostat.cur, &delta);
1006 blkg_iostat_add(last, &delta);
1007 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
1008}
1009
1010static void __blkcg_rstat_flush(struct blkcg *blkcg, int cpu)
1011{
1012 struct llist_head *lhead = per_cpu_ptr(blkcg->lhead, cpu);
1013 struct llist_node *lnode;
1014 struct blkg_iostat_set *bisc, *next_bisc;
1015 unsigned long flags;
1016
1017 rcu_read_lock();
1018
1019 lnode = llist_del_all(lhead);
1020 if (!lnode)
1021 goto out;
1022
1023 /*
1024 * For covering concurrent parent blkg update from blkg_release().
1025 *
1026 * When flushing from cgroup, cgroup_rstat_lock is always held, so
1027 * this lock won't cause contention most of time.
1028 */
1029 raw_spin_lock_irqsave(&blkg_stat_lock, flags);
1030
1031 /*
1032 * Iterate only the iostat_cpu's queued in the lockless list.
1033 */
1034 llist_for_each_entry_safe(bisc, next_bisc, lnode, lnode) {
1035 struct blkcg_gq *blkg = bisc->blkg;
1036 struct blkcg_gq *parent = blkg->parent;
1037 struct blkg_iostat cur;
1038 unsigned int seq;
1039
1040 /*
1041 * Order assignment of `next_bisc` from `bisc->lnode.next` in
1042 * llist_for_each_entry_safe and clearing `bisc->lqueued` for
1043 * avoiding to assign `next_bisc` with new next pointer added
1044 * in blk_cgroup_bio_start() in case of re-ordering.
1045 *
1046 * The pair barrier is implied in llist_add() in blk_cgroup_bio_start().
1047 */
1048 smp_mb();
1049
1050 WRITE_ONCE(bisc->lqueued, false);
1051 if (bisc == &blkg->iostat)
1052 goto propagate_up; /* propagate up to parent only */
1053
1054 /* fetch the current per-cpu values */
1055 do {
1056 seq = u64_stats_fetch_begin(&bisc->sync);
1057 blkg_iostat_set(&cur, &bisc->cur);
1058 } while (u64_stats_fetch_retry(&bisc->sync, seq));
1059
1060 blkcg_iostat_update(blkg, &cur, &bisc->last);
1061
1062propagate_up:
1063 /* propagate global delta to parent (unless that's root) */
1064 if (parent && parent->parent) {
1065 blkcg_iostat_update(parent, &blkg->iostat.cur,
1066 &blkg->iostat.last);
1067 /*
1068 * Queue parent->iostat to its blkcg's lockless
1069 * list to propagate up to the grandparent if the
1070 * iostat hasn't been queued yet.
1071 */
1072 if (!parent->iostat.lqueued) {
1073 struct llist_head *plhead;
1074
1075 plhead = per_cpu_ptr(parent->blkcg->lhead, cpu);
1076 llist_add(&parent->iostat.lnode, plhead);
1077 parent->iostat.lqueued = true;
1078 }
1079 }
1080 }
1081 raw_spin_unlock_irqrestore(&blkg_stat_lock, flags);
1082out:
1083 rcu_read_unlock();
1084}
1085
1086static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
1087{
1088 /* Root-level stats are sourced from system-wide IO stats */
1089 if (cgroup_parent(css->cgroup))
1090 __blkcg_rstat_flush(css_to_blkcg(css), cpu);
1091}
1092
1093/*
1094 * We source root cgroup stats from the system-wide stats to avoid
1095 * tracking the same information twice and incurring overhead when no
1096 * cgroups are defined. For that reason, cgroup_rstat_flush in
1097 * blkcg_print_stat does not actually fill out the iostat in the root
1098 * cgroup's blkcg_gq.
1099 *
1100 * However, we would like to re-use the printing code between the root and
1101 * non-root cgroups to the extent possible. For that reason, we simulate
1102 * flushing the root cgroup's stats by explicitly filling in the iostat
1103 * with disk level statistics.
1104 */
1105static void blkcg_fill_root_iostats(void)
1106{
1107 struct class_dev_iter iter;
1108 struct device *dev;
1109
1110 class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
1111 while ((dev = class_dev_iter_next(&iter))) {
1112 struct block_device *bdev = dev_to_bdev(dev);
1113 struct blkcg_gq *blkg = bdev->bd_disk->queue->root_blkg;
1114 struct blkg_iostat tmp;
1115 int cpu;
1116 unsigned long flags;
1117
1118 memset(&tmp, 0, sizeof(tmp));
1119 for_each_possible_cpu(cpu) {
1120 struct disk_stats *cpu_dkstats;
1121
1122 cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu);
1123 tmp.ios[BLKG_IOSTAT_READ] +=
1124 cpu_dkstats->ios[STAT_READ];
1125 tmp.ios[BLKG_IOSTAT_WRITE] +=
1126 cpu_dkstats->ios[STAT_WRITE];
1127 tmp.ios[BLKG_IOSTAT_DISCARD] +=
1128 cpu_dkstats->ios[STAT_DISCARD];
1129 // convert sectors to bytes
1130 tmp.bytes[BLKG_IOSTAT_READ] +=
1131 cpu_dkstats->sectors[STAT_READ] << 9;
1132 tmp.bytes[BLKG_IOSTAT_WRITE] +=
1133 cpu_dkstats->sectors[STAT_WRITE] << 9;
1134 tmp.bytes[BLKG_IOSTAT_DISCARD] +=
1135 cpu_dkstats->sectors[STAT_DISCARD] << 9;
1136 }
1137
1138 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
1139 blkg_iostat_set(&blkg->iostat.cur, &tmp);
1140 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
1141 }
1142}
1143
1144static void blkcg_print_one_stat(struct blkcg_gq *blkg, struct seq_file *s)
1145{
1146 struct blkg_iostat_set *bis = &blkg->iostat;
1147 u64 rbytes, wbytes, rios, wios, dbytes, dios;
1148 const char *dname;
1149 unsigned seq;
1150 int i;
1151
1152 if (!blkg->online)
1153 return;
1154
1155 dname = blkg_dev_name(blkg);
1156 if (!dname)
1157 return;
1158
1159 seq_printf(s, "%s ", dname);
1160
1161 do {
1162 seq = u64_stats_fetch_begin(&bis->sync);
1163
1164 rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
1165 wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
1166 dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
1167 rios = bis->cur.ios[BLKG_IOSTAT_READ];
1168 wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
1169 dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
1170 } while (u64_stats_fetch_retry(&bis->sync, seq));
1171
1172 if (rbytes || wbytes || rios || wios) {
1173 seq_printf(s, "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
1174 rbytes, wbytes, rios, wios,
1175 dbytes, dios);
1176 }
1177
1178 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
1179 seq_printf(s, " use_delay=%d delay_nsec=%llu",
1180 atomic_read(&blkg->use_delay),
1181 atomic64_read(&blkg->delay_nsec));
1182 }
1183
1184 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1185 struct blkcg_policy *pol = blkcg_policy[i];
1186
1187 if (!blkg->pd[i] || !pol->pd_stat_fn)
1188 continue;
1189
1190 pol->pd_stat_fn(blkg->pd[i], s);
1191 }
1192
1193 seq_puts(s, "\n");
1194}
1195
1196static int blkcg_print_stat(struct seq_file *sf, void *v)
1197{
1198 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1199 struct blkcg_gq *blkg;
1200
1201 if (!seq_css(sf)->parent)
1202 blkcg_fill_root_iostats();
1203 else
1204 cgroup_rstat_flush(blkcg->css.cgroup);
1205
1206 rcu_read_lock();
1207 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
1208 spin_lock_irq(&blkg->q->queue_lock);
1209 blkcg_print_one_stat(blkg, sf);
1210 spin_unlock_irq(&blkg->q->queue_lock);
1211 }
1212 rcu_read_unlock();
1213 return 0;
1214}
1215
1216static struct cftype blkcg_files[] = {
1217 {
1218 .name = "stat",
1219 .seq_show = blkcg_print_stat,
1220 },
1221 { } /* terminate */
1222};
1223
1224static struct cftype blkcg_legacy_files[] = {
1225 {
1226 .name = "reset_stats",
1227 .write_u64 = blkcg_reset_stats,
1228 },
1229 { } /* terminate */
1230};
1231
1232#ifdef CONFIG_CGROUP_WRITEBACK
1233struct list_head *blkcg_get_cgwb_list(struct cgroup_subsys_state *css)
1234{
1235 return &css_to_blkcg(css)->cgwb_list;
1236}
1237#endif
1238
1239/*
1240 * blkcg destruction is a three-stage process.
1241 *
1242 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
1243 * which offlines writeback. Here we tie the next stage of blkg destruction
1244 * to the completion of writeback associated with the blkcg. This lets us
1245 * avoid punting potentially large amounts of outstanding writeback to root
1246 * while maintaining any ongoing policies. The next stage is triggered when
1247 * the nr_cgwbs count goes to zero.
1248 *
1249 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
1250 * and handles the destruction of blkgs. Here the css reference held by
1251 * the blkg is put back eventually allowing blkcg_css_free() to be called.
1252 * This work may occur in cgwb_release_workfn() on the cgwb_release
1253 * workqueue. Any submitted ios that fail to get the blkg ref will be
1254 * punted to the root_blkg.
1255 *
1256 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1257 * This finally frees the blkcg.
1258 */
1259
1260/**
1261 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1262 * @blkcg: blkcg of interest
1263 *
1264 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1265 * is nested inside q lock, this function performs reverse double lock dancing.
1266 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1267 * blkcg_css_free to eventually be called.
1268 *
1269 * This is the blkcg counterpart of ioc_release_fn().
1270 */
1271static void blkcg_destroy_blkgs(struct blkcg *blkcg)
1272{
1273 might_sleep();
1274
1275 spin_lock_irq(&blkcg->lock);
1276
1277 while (!hlist_empty(&blkcg->blkg_list)) {
1278 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1279 struct blkcg_gq, blkcg_node);
1280 struct request_queue *q = blkg->q;
1281
1282 if (need_resched() || !spin_trylock(&q->queue_lock)) {
1283 /*
1284 * Given that the system can accumulate a huge number
1285 * of blkgs in pathological cases, check to see if we
1286 * need to rescheduling to avoid softlockup.
1287 */
1288 spin_unlock_irq(&blkcg->lock);
1289 cond_resched();
1290 spin_lock_irq(&blkcg->lock);
1291 continue;
1292 }
1293
1294 blkg_destroy(blkg);
1295 spin_unlock(&q->queue_lock);
1296 }
1297
1298 spin_unlock_irq(&blkcg->lock);
1299}
1300
1301/**
1302 * blkcg_pin_online - pin online state
1303 * @blkcg_css: blkcg of interest
1304 *
1305 * While pinned, a blkcg is kept online. This is primarily used to
1306 * impedance-match blkg and cgwb lifetimes so that blkg doesn't go offline
1307 * while an associated cgwb is still active.
1308 */
1309void blkcg_pin_online(struct cgroup_subsys_state *blkcg_css)
1310{
1311 refcount_inc(&css_to_blkcg(blkcg_css)->online_pin);
1312}
1313
1314/**
1315 * blkcg_unpin_online - unpin online state
1316 * @blkcg_css: blkcg of interest
1317 *
1318 * This is primarily used to impedance-match blkg and cgwb lifetimes so
1319 * that blkg doesn't go offline while an associated cgwb is still active.
1320 * When this count goes to zero, all active cgwbs have finished so the
1321 * blkcg can continue destruction by calling blkcg_destroy_blkgs().
1322 */
1323void blkcg_unpin_online(struct cgroup_subsys_state *blkcg_css)
1324{
1325 struct blkcg *blkcg = css_to_blkcg(blkcg_css);
1326
1327 do {
1328 if (!refcount_dec_and_test(&blkcg->online_pin))
1329 break;
1330 blkcg_destroy_blkgs(blkcg);
1331 blkcg = blkcg_parent(blkcg);
1332 } while (blkcg);
1333}
1334
1335/**
1336 * blkcg_css_offline - cgroup css_offline callback
1337 * @css: css of interest
1338 *
1339 * This function is called when @css is about to go away. Here the cgwbs are
1340 * offlined first and only once writeback associated with the blkcg has
1341 * finished do we start step 2 (see above).
1342 */
1343static void blkcg_css_offline(struct cgroup_subsys_state *css)
1344{
1345 /* this prevents anyone from attaching or migrating to this blkcg */
1346 wb_blkcg_offline(css);
1347
1348 /* put the base online pin allowing step 2 to be triggered */
1349 blkcg_unpin_online(css);
1350}
1351
1352static void blkcg_css_free(struct cgroup_subsys_state *css)
1353{
1354 struct blkcg *blkcg = css_to_blkcg(css);
1355 int i;
1356
1357 mutex_lock(&blkcg_pol_mutex);
1358
1359 list_del(&blkcg->all_blkcgs_node);
1360
1361 for (i = 0; i < BLKCG_MAX_POLS; i++)
1362 if (blkcg->cpd[i])
1363 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1364
1365 mutex_unlock(&blkcg_pol_mutex);
1366
1367 free_percpu(blkcg->lhead);
1368 kfree(blkcg);
1369}
1370
1371static struct cgroup_subsys_state *
1372blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1373{
1374 struct blkcg *blkcg;
1375 int i;
1376
1377 mutex_lock(&blkcg_pol_mutex);
1378
1379 if (!parent_css) {
1380 blkcg = &blkcg_root;
1381 } else {
1382 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1383 if (!blkcg)
1384 goto unlock;
1385 }
1386
1387 if (init_blkcg_llists(blkcg))
1388 goto free_blkcg;
1389
1390 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1391 struct blkcg_policy *pol = blkcg_policy[i];
1392 struct blkcg_policy_data *cpd;
1393
1394 /*
1395 * If the policy hasn't been attached yet, wait for it
1396 * to be attached before doing anything else. Otherwise,
1397 * check if the policy requires any specific per-cgroup
1398 * data: if it does, allocate and initialize it.
1399 */
1400 if (!pol || !pol->cpd_alloc_fn)
1401 continue;
1402
1403 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1404 if (!cpd)
1405 goto free_pd_blkcg;
1406
1407 blkcg->cpd[i] = cpd;
1408 cpd->blkcg = blkcg;
1409 cpd->plid = i;
1410 }
1411
1412 spin_lock_init(&blkcg->lock);
1413 refcount_set(&blkcg->online_pin, 1);
1414 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1415 INIT_HLIST_HEAD(&blkcg->blkg_list);
1416#ifdef CONFIG_CGROUP_WRITEBACK
1417 INIT_LIST_HEAD(&blkcg->cgwb_list);
1418#endif
1419 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1420
1421 mutex_unlock(&blkcg_pol_mutex);
1422 return &blkcg->css;
1423
1424free_pd_blkcg:
1425 for (i--; i >= 0; i--)
1426 if (blkcg->cpd[i])
1427 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1428 free_percpu(blkcg->lhead);
1429free_blkcg:
1430 if (blkcg != &blkcg_root)
1431 kfree(blkcg);
1432unlock:
1433 mutex_unlock(&blkcg_pol_mutex);
1434 return ERR_PTR(-ENOMEM);
1435}
1436
1437static int blkcg_css_online(struct cgroup_subsys_state *css)
1438{
1439 struct blkcg *parent = blkcg_parent(css_to_blkcg(css));
1440
1441 /*
1442 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1443 * don't go offline while cgwbs are still active on them. Pin the
1444 * parent so that offline always happens towards the root.
1445 */
1446 if (parent)
1447 blkcg_pin_online(&parent->css);
1448 return 0;
1449}
1450
1451void blkg_init_queue(struct request_queue *q)
1452{
1453 INIT_LIST_HEAD(&q->blkg_list);
1454 mutex_init(&q->blkcg_mutex);
1455}
1456
1457int blkcg_init_disk(struct gendisk *disk)
1458{
1459 struct request_queue *q = disk->queue;
1460 struct blkcg_gq *new_blkg, *blkg;
1461 bool preloaded;
1462 int ret;
1463
1464 new_blkg = blkg_alloc(&blkcg_root, disk, GFP_KERNEL);
1465 if (!new_blkg)
1466 return -ENOMEM;
1467
1468 preloaded = !radix_tree_preload(GFP_KERNEL);
1469
1470 /* Make sure the root blkg exists. */
1471 /* spin_lock_irq can serve as RCU read-side critical section. */
1472 spin_lock_irq(&q->queue_lock);
1473 blkg = blkg_create(&blkcg_root, disk, new_blkg);
1474 if (IS_ERR(blkg))
1475 goto err_unlock;
1476 q->root_blkg = blkg;
1477 spin_unlock_irq(&q->queue_lock);
1478
1479 if (preloaded)
1480 radix_tree_preload_end();
1481
1482 ret = blk_ioprio_init(disk);
1483 if (ret)
1484 goto err_destroy_all;
1485
1486 ret = blk_throtl_init(disk);
1487 if (ret)
1488 goto err_ioprio_exit;
1489
1490 return 0;
1491
1492err_ioprio_exit:
1493 blk_ioprio_exit(disk);
1494err_destroy_all:
1495 blkg_destroy_all(disk);
1496 return ret;
1497err_unlock:
1498 spin_unlock_irq(&q->queue_lock);
1499 if (preloaded)
1500 radix_tree_preload_end();
1501 return PTR_ERR(blkg);
1502}
1503
1504void blkcg_exit_disk(struct gendisk *disk)
1505{
1506 blkg_destroy_all(disk);
1507 blk_throtl_exit(disk);
1508}
1509
1510static void blkcg_exit(struct task_struct *tsk)
1511{
1512 if (tsk->throttle_disk)
1513 put_disk(tsk->throttle_disk);
1514 tsk->throttle_disk = NULL;
1515}
1516
1517struct cgroup_subsys io_cgrp_subsys = {
1518 .css_alloc = blkcg_css_alloc,
1519 .css_online = blkcg_css_online,
1520 .css_offline = blkcg_css_offline,
1521 .css_free = blkcg_css_free,
1522 .css_rstat_flush = blkcg_rstat_flush,
1523 .dfl_cftypes = blkcg_files,
1524 .legacy_cftypes = blkcg_legacy_files,
1525 .legacy_name = "blkio",
1526 .exit = blkcg_exit,
1527#ifdef CONFIG_MEMCG
1528 /*
1529 * This ensures that, if available, memcg is automatically enabled
1530 * together on the default hierarchy so that the owner cgroup can
1531 * be retrieved from writeback pages.
1532 */
1533 .depends_on = 1 << memory_cgrp_id,
1534#endif
1535};
1536EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1537
1538/**
1539 * blkcg_activate_policy - activate a blkcg policy on a gendisk
1540 * @disk: gendisk of interest
1541 * @pol: blkcg policy to activate
1542 *
1543 * Activate @pol on @disk. Requires %GFP_KERNEL context. @disk goes through
1544 * bypass mode to populate its blkgs with policy_data for @pol.
1545 *
1546 * Activation happens with @disk bypassed, so nobody would be accessing blkgs
1547 * from IO path. Update of each blkg is protected by both queue and blkcg
1548 * locks so that holding either lock and testing blkcg_policy_enabled() is
1549 * always enough for dereferencing policy data.
1550 *
1551 * The caller is responsible for synchronizing [de]activations and policy
1552 * [un]registerations. Returns 0 on success, -errno on failure.
1553 */
1554int blkcg_activate_policy(struct gendisk *disk, const struct blkcg_policy *pol)
1555{
1556 struct request_queue *q = disk->queue;
1557 struct blkg_policy_data *pd_prealloc = NULL;
1558 struct blkcg_gq *blkg, *pinned_blkg = NULL;
1559 int ret;
1560
1561 if (blkcg_policy_enabled(q, pol))
1562 return 0;
1563
1564 if (queue_is_mq(q))
1565 blk_mq_freeze_queue(q);
1566retry:
1567 spin_lock_irq(&q->queue_lock);
1568
1569 /* blkg_list is pushed at the head, reverse walk to initialize parents first */
1570 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1571 struct blkg_policy_data *pd;
1572
1573 if (blkg->pd[pol->plid])
1574 continue;
1575
1576 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1577 if (blkg == pinned_blkg) {
1578 pd = pd_prealloc;
1579 pd_prealloc = NULL;
1580 } else {
1581 pd = pol->pd_alloc_fn(disk, blkg->blkcg,
1582 GFP_NOWAIT | __GFP_NOWARN);
1583 }
1584
1585 if (!pd) {
1586 /*
1587 * GFP_NOWAIT failed. Free the existing one and
1588 * prealloc for @blkg w/ GFP_KERNEL.
1589 */
1590 if (pinned_blkg)
1591 blkg_put(pinned_blkg);
1592 blkg_get(blkg);
1593 pinned_blkg = blkg;
1594
1595 spin_unlock_irq(&q->queue_lock);
1596
1597 if (pd_prealloc)
1598 pol->pd_free_fn(pd_prealloc);
1599 pd_prealloc = pol->pd_alloc_fn(disk, blkg->blkcg,
1600 GFP_KERNEL);
1601 if (pd_prealloc)
1602 goto retry;
1603 else
1604 goto enomem;
1605 }
1606
1607 spin_lock(&blkg->blkcg->lock);
1608
1609 pd->blkg = blkg;
1610 pd->plid = pol->plid;
1611 blkg->pd[pol->plid] = pd;
1612
1613 if (pol->pd_init_fn)
1614 pol->pd_init_fn(pd);
1615
1616 if (pol->pd_online_fn)
1617 pol->pd_online_fn(pd);
1618 pd->online = true;
1619
1620 spin_unlock(&blkg->blkcg->lock);
1621 }
1622
1623 __set_bit(pol->plid, q->blkcg_pols);
1624 ret = 0;
1625
1626 spin_unlock_irq(&q->queue_lock);
1627out:
1628 if (queue_is_mq(q))
1629 blk_mq_unfreeze_queue(q);
1630 if (pinned_blkg)
1631 blkg_put(pinned_blkg);
1632 if (pd_prealloc)
1633 pol->pd_free_fn(pd_prealloc);
1634 return ret;
1635
1636enomem:
1637 /* alloc failed, take down everything */
1638 spin_lock_irq(&q->queue_lock);
1639 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1640 struct blkcg *blkcg = blkg->blkcg;
1641 struct blkg_policy_data *pd;
1642
1643 spin_lock(&blkcg->lock);
1644 pd = blkg->pd[pol->plid];
1645 if (pd) {
1646 if (pd->online && pol->pd_offline_fn)
1647 pol->pd_offline_fn(pd);
1648 pd->online = false;
1649 pol->pd_free_fn(pd);
1650 blkg->pd[pol->plid] = NULL;
1651 }
1652 spin_unlock(&blkcg->lock);
1653 }
1654 spin_unlock_irq(&q->queue_lock);
1655 ret = -ENOMEM;
1656 goto out;
1657}
1658EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1659
1660/**
1661 * blkcg_deactivate_policy - deactivate a blkcg policy on a gendisk
1662 * @disk: gendisk of interest
1663 * @pol: blkcg policy to deactivate
1664 *
1665 * Deactivate @pol on @disk. Follows the same synchronization rules as
1666 * blkcg_activate_policy().
1667 */
1668void blkcg_deactivate_policy(struct gendisk *disk,
1669 const struct blkcg_policy *pol)
1670{
1671 struct request_queue *q = disk->queue;
1672 struct blkcg_gq *blkg;
1673
1674 if (!blkcg_policy_enabled(q, pol))
1675 return;
1676
1677 if (queue_is_mq(q))
1678 blk_mq_freeze_queue(q);
1679
1680 mutex_lock(&q->blkcg_mutex);
1681 spin_lock_irq(&q->queue_lock);
1682
1683 __clear_bit(pol->plid, q->blkcg_pols);
1684
1685 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1686 struct blkcg *blkcg = blkg->blkcg;
1687
1688 spin_lock(&blkcg->lock);
1689 if (blkg->pd[pol->plid]) {
1690 if (blkg->pd[pol->plid]->online && pol->pd_offline_fn)
1691 pol->pd_offline_fn(blkg->pd[pol->plid]);
1692 pol->pd_free_fn(blkg->pd[pol->plid]);
1693 blkg->pd[pol->plid] = NULL;
1694 }
1695 spin_unlock(&blkcg->lock);
1696 }
1697
1698 spin_unlock_irq(&q->queue_lock);
1699 mutex_unlock(&q->blkcg_mutex);
1700
1701 if (queue_is_mq(q))
1702 blk_mq_unfreeze_queue(q);
1703}
1704EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1705
1706static void blkcg_free_all_cpd(struct blkcg_policy *pol)
1707{
1708 struct blkcg *blkcg;
1709
1710 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1711 if (blkcg->cpd[pol->plid]) {
1712 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1713 blkcg->cpd[pol->plid] = NULL;
1714 }
1715 }
1716}
1717
1718/**
1719 * blkcg_policy_register - register a blkcg policy
1720 * @pol: blkcg policy to register
1721 *
1722 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1723 * successful registration. Returns 0 on success and -errno on failure.
1724 */
1725int blkcg_policy_register(struct blkcg_policy *pol)
1726{
1727 struct blkcg *blkcg;
1728 int i, ret;
1729
1730 mutex_lock(&blkcg_pol_register_mutex);
1731 mutex_lock(&blkcg_pol_mutex);
1732
1733 /* find an empty slot */
1734 ret = -ENOSPC;
1735 for (i = 0; i < BLKCG_MAX_POLS; i++)
1736 if (!blkcg_policy[i])
1737 break;
1738 if (i >= BLKCG_MAX_POLS) {
1739 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1740 goto err_unlock;
1741 }
1742
1743 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1744 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1745 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1746 goto err_unlock;
1747
1748 /* register @pol */
1749 pol->plid = i;
1750 blkcg_policy[pol->plid] = pol;
1751
1752 /* allocate and install cpd's */
1753 if (pol->cpd_alloc_fn) {
1754 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1755 struct blkcg_policy_data *cpd;
1756
1757 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1758 if (!cpd)
1759 goto err_free_cpds;
1760
1761 blkcg->cpd[pol->plid] = cpd;
1762 cpd->blkcg = blkcg;
1763 cpd->plid = pol->plid;
1764 }
1765 }
1766
1767 mutex_unlock(&blkcg_pol_mutex);
1768
1769 /* everything is in place, add intf files for the new policy */
1770 if (pol->dfl_cftypes)
1771 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1772 pol->dfl_cftypes));
1773 if (pol->legacy_cftypes)
1774 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1775 pol->legacy_cftypes));
1776 mutex_unlock(&blkcg_pol_register_mutex);
1777 return 0;
1778
1779err_free_cpds:
1780 if (pol->cpd_free_fn)
1781 blkcg_free_all_cpd(pol);
1782
1783 blkcg_policy[pol->plid] = NULL;
1784err_unlock:
1785 mutex_unlock(&blkcg_pol_mutex);
1786 mutex_unlock(&blkcg_pol_register_mutex);
1787 return ret;
1788}
1789EXPORT_SYMBOL_GPL(blkcg_policy_register);
1790
1791/**
1792 * blkcg_policy_unregister - unregister a blkcg policy
1793 * @pol: blkcg policy to unregister
1794 *
1795 * Undo blkcg_policy_register(@pol). Might sleep.
1796 */
1797void blkcg_policy_unregister(struct blkcg_policy *pol)
1798{
1799 mutex_lock(&blkcg_pol_register_mutex);
1800
1801 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1802 goto out_unlock;
1803
1804 /* kill the intf files first */
1805 if (pol->dfl_cftypes)
1806 cgroup_rm_cftypes(pol->dfl_cftypes);
1807 if (pol->legacy_cftypes)
1808 cgroup_rm_cftypes(pol->legacy_cftypes);
1809
1810 /* remove cpds and unregister */
1811 mutex_lock(&blkcg_pol_mutex);
1812
1813 if (pol->cpd_free_fn)
1814 blkcg_free_all_cpd(pol);
1815
1816 blkcg_policy[pol->plid] = NULL;
1817
1818 mutex_unlock(&blkcg_pol_mutex);
1819out_unlock:
1820 mutex_unlock(&blkcg_pol_register_mutex);
1821}
1822EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1823
1824/*
1825 * Scale the accumulated delay based on how long it has been since we updated
1826 * the delay. We only call this when we are adding delay, in case it's been a
1827 * while since we added delay, and when we are checking to see if we need to
1828 * delay a task, to account for any delays that may have occurred.
1829 */
1830static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1831{
1832 u64 old = atomic64_read(&blkg->delay_start);
1833
1834 /* negative use_delay means no scaling, see blkcg_set_delay() */
1835 if (atomic_read(&blkg->use_delay) < 0)
1836 return;
1837
1838 /*
1839 * We only want to scale down every second. The idea here is that we
1840 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1841 * time window. We only want to throttle tasks for recent delay that
1842 * has occurred, in 1 second time windows since that's the maximum
1843 * things can be throttled. We save the current delay window in
1844 * blkg->last_delay so we know what amount is still left to be charged
1845 * to the blkg from this point onward. blkg->last_use keeps track of
1846 * the use_delay counter. The idea is if we're unthrottling the blkg we
1847 * are ok with whatever is happening now, and we can take away more of
1848 * the accumulated delay as we've already throttled enough that
1849 * everybody is happy with their IO latencies.
1850 */
1851 if (time_before64(old + NSEC_PER_SEC, now) &&
1852 atomic64_try_cmpxchg(&blkg->delay_start, &old, now)) {
1853 u64 cur = atomic64_read(&blkg->delay_nsec);
1854 u64 sub = min_t(u64, blkg->last_delay, now - old);
1855 int cur_use = atomic_read(&blkg->use_delay);
1856
1857 /*
1858 * We've been unthrottled, subtract a larger chunk of our
1859 * accumulated delay.
1860 */
1861 if (cur_use < blkg->last_use)
1862 sub = max_t(u64, sub, blkg->last_delay >> 1);
1863
1864 /*
1865 * This shouldn't happen, but handle it anyway. Our delay_nsec
1866 * should only ever be growing except here where we subtract out
1867 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1868 * rather not end up with negative numbers.
1869 */
1870 if (unlikely(cur < sub)) {
1871 atomic64_set(&blkg->delay_nsec, 0);
1872 blkg->last_delay = 0;
1873 } else {
1874 atomic64_sub(sub, &blkg->delay_nsec);
1875 blkg->last_delay = cur - sub;
1876 }
1877 blkg->last_use = cur_use;
1878 }
1879}
1880
1881/*
1882 * This is called when we want to actually walk up the hierarchy and check to
1883 * see if we need to throttle, and then actually throttle if there is some
1884 * accumulated delay. This should only be called upon return to user space so
1885 * we're not holding some lock that would induce a priority inversion.
1886 */
1887static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1888{
1889 unsigned long pflags;
1890 bool clamp;
1891 u64 now = blk_time_get_ns();
1892 u64 exp;
1893 u64 delay_nsec = 0;
1894 int tok;
1895
1896 while (blkg->parent) {
1897 int use_delay = atomic_read(&blkg->use_delay);
1898
1899 if (use_delay) {
1900 u64 this_delay;
1901
1902 blkcg_scale_delay(blkg, now);
1903 this_delay = atomic64_read(&blkg->delay_nsec);
1904 if (this_delay > delay_nsec) {
1905 delay_nsec = this_delay;
1906 clamp = use_delay > 0;
1907 }
1908 }
1909 blkg = blkg->parent;
1910 }
1911
1912 if (!delay_nsec)
1913 return;
1914
1915 /*
1916 * Let's not sleep for all eternity if we've amassed a huge delay.
1917 * Swapping or metadata IO can accumulate 10's of seconds worth of
1918 * delay, and we want userspace to be able to do _something_ so cap the
1919 * delays at 0.25s. If there's 10's of seconds worth of delay then the
1920 * tasks will be delayed for 0.25 second for every syscall. If
1921 * blkcg_set_delay() was used as indicated by negative use_delay, the
1922 * caller is responsible for regulating the range.
1923 */
1924 if (clamp)
1925 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1926
1927 if (use_memdelay)
1928 psi_memstall_enter(&pflags);
1929
1930 exp = ktime_add_ns(now, delay_nsec);
1931 tok = io_schedule_prepare();
1932 do {
1933 __set_current_state(TASK_KILLABLE);
1934 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1935 break;
1936 } while (!fatal_signal_pending(current));
1937 io_schedule_finish(tok);
1938
1939 if (use_memdelay)
1940 psi_memstall_leave(&pflags);
1941}
1942
1943/**
1944 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1945 *
1946 * This is only called if we've been marked with set_notify_resume(). Obviously
1947 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1948 * check to see if current->throttle_disk is set and if not this doesn't do
1949 * anything. This should only ever be called by the resume code, it's not meant
1950 * to be called by people willy-nilly as it will actually do the work to
1951 * throttle the task if it is setup for throttling.
1952 */
1953void blkcg_maybe_throttle_current(void)
1954{
1955 struct gendisk *disk = current->throttle_disk;
1956 struct blkcg *blkcg;
1957 struct blkcg_gq *blkg;
1958 bool use_memdelay = current->use_memdelay;
1959
1960 if (!disk)
1961 return;
1962
1963 current->throttle_disk = NULL;
1964 current->use_memdelay = false;
1965
1966 rcu_read_lock();
1967 blkcg = css_to_blkcg(blkcg_css());
1968 if (!blkcg)
1969 goto out;
1970 blkg = blkg_lookup(blkcg, disk->queue);
1971 if (!blkg)
1972 goto out;
1973 if (!blkg_tryget(blkg))
1974 goto out;
1975 rcu_read_unlock();
1976
1977 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1978 blkg_put(blkg);
1979 put_disk(disk);
1980 return;
1981out:
1982 rcu_read_unlock();
1983}
1984
1985/**
1986 * blkcg_schedule_throttle - this task needs to check for throttling
1987 * @disk: disk to throttle
1988 * @use_memdelay: do we charge this to memory delay for PSI
1989 *
1990 * This is called by the IO controller when we know there's delay accumulated
1991 * for the blkg for this task. We do not pass the blkg because there are places
1992 * we call this that may not have that information, the swapping code for
1993 * instance will only have a block_device at that point. This set's the
1994 * notify_resume for the task to check and see if it requires throttling before
1995 * returning to user space.
1996 *
1997 * We will only schedule once per syscall. You can call this over and over
1998 * again and it will only do the check once upon return to user space, and only
1999 * throttle once. If the task needs to be throttled again it'll need to be
2000 * re-set at the next time we see the task.
2001 */
2002void blkcg_schedule_throttle(struct gendisk *disk, bool use_memdelay)
2003{
2004 if (unlikely(current->flags & PF_KTHREAD))
2005 return;
2006
2007 if (current->throttle_disk != disk) {
2008 if (test_bit(GD_DEAD, &disk->state))
2009 return;
2010 get_device(disk_to_dev(disk));
2011
2012 if (current->throttle_disk)
2013 put_disk(current->throttle_disk);
2014 current->throttle_disk = disk;
2015 }
2016
2017 if (use_memdelay)
2018 current->use_memdelay = use_memdelay;
2019 set_notify_resume(current);
2020}
2021
2022/**
2023 * blkcg_add_delay - add delay to this blkg
2024 * @blkg: blkg of interest
2025 * @now: the current time in nanoseconds
2026 * @delta: how many nanoseconds of delay to add
2027 *
2028 * Charge @delta to the blkg's current delay accumulation. This is used to
2029 * throttle tasks if an IO controller thinks we need more throttling.
2030 */
2031void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
2032{
2033 if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
2034 return;
2035 blkcg_scale_delay(blkg, now);
2036 atomic64_add(delta, &blkg->delay_nsec);
2037}
2038
2039/**
2040 * blkg_tryget_closest - try and get a blkg ref on the closet blkg
2041 * @bio: target bio
2042 * @css: target css
2043 *
2044 * As the failure mode here is to walk up the blkg tree, this ensure that the
2045 * blkg->parent pointers are always valid. This returns the blkg that it ended
2046 * up taking a reference on or %NULL if no reference was taken.
2047 */
2048static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio,
2049 struct cgroup_subsys_state *css)
2050{
2051 struct blkcg_gq *blkg, *ret_blkg = NULL;
2052
2053 rcu_read_lock();
2054 blkg = blkg_lookup_create(css_to_blkcg(css), bio->bi_bdev->bd_disk);
2055 while (blkg) {
2056 if (blkg_tryget(blkg)) {
2057 ret_blkg = blkg;
2058 break;
2059 }
2060 blkg = blkg->parent;
2061 }
2062 rcu_read_unlock();
2063
2064 return ret_blkg;
2065}
2066
2067/**
2068 * bio_associate_blkg_from_css - associate a bio with a specified css
2069 * @bio: target bio
2070 * @css: target css
2071 *
2072 * Associate @bio with the blkg found by combining the css's blkg and the
2073 * request_queue of the @bio. An association failure is handled by walking up
2074 * the blkg tree. Therefore, the blkg associated can be anything between @blkg
2075 * and q->root_blkg. This situation only happens when a cgroup is dying and
2076 * then the remaining bios will spill to the closest alive blkg.
2077 *
2078 * A reference will be taken on the blkg and will be released when @bio is
2079 * freed.
2080 */
2081void bio_associate_blkg_from_css(struct bio *bio,
2082 struct cgroup_subsys_state *css)
2083{
2084 if (bio->bi_blkg)
2085 blkg_put(bio->bi_blkg);
2086
2087 if (css && css->parent) {
2088 bio->bi_blkg = blkg_tryget_closest(bio, css);
2089 } else {
2090 blkg_get(bdev_get_queue(bio->bi_bdev)->root_blkg);
2091 bio->bi_blkg = bdev_get_queue(bio->bi_bdev)->root_blkg;
2092 }
2093}
2094EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css);
2095
2096/**
2097 * bio_associate_blkg - associate a bio with a blkg
2098 * @bio: target bio
2099 *
2100 * Associate @bio with the blkg found from the bio's css and request_queue.
2101 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is
2102 * already associated, the css is reused and association redone as the
2103 * request_queue may have changed.
2104 */
2105void bio_associate_blkg(struct bio *bio)
2106{
2107 struct cgroup_subsys_state *css;
2108
2109 if (blk_op_is_passthrough(bio->bi_opf))
2110 return;
2111
2112 rcu_read_lock();
2113
2114 if (bio->bi_blkg)
2115 css = bio_blkcg_css(bio);
2116 else
2117 css = blkcg_css();
2118
2119 bio_associate_blkg_from_css(bio, css);
2120
2121 rcu_read_unlock();
2122}
2123EXPORT_SYMBOL_GPL(bio_associate_blkg);
2124
2125/**
2126 * bio_clone_blkg_association - clone blkg association from src to dst bio
2127 * @dst: destination bio
2128 * @src: source bio
2129 */
2130void bio_clone_blkg_association(struct bio *dst, struct bio *src)
2131{
2132 if (src->bi_blkg)
2133 bio_associate_blkg_from_css(dst, bio_blkcg_css(src));
2134}
2135EXPORT_SYMBOL_GPL(bio_clone_blkg_association);
2136
2137static int blk_cgroup_io_type(struct bio *bio)
2138{
2139 if (op_is_discard(bio->bi_opf))
2140 return BLKG_IOSTAT_DISCARD;
2141 if (op_is_write(bio->bi_opf))
2142 return BLKG_IOSTAT_WRITE;
2143 return BLKG_IOSTAT_READ;
2144}
2145
2146void blk_cgroup_bio_start(struct bio *bio)
2147{
2148 struct blkcg *blkcg = bio->bi_blkg->blkcg;
2149 int rwd = blk_cgroup_io_type(bio), cpu;
2150 struct blkg_iostat_set *bis;
2151 unsigned long flags;
2152
2153 if (!cgroup_subsys_on_dfl(io_cgrp_subsys))
2154 return;
2155
2156 /* Root-level stats are sourced from system-wide IO stats */
2157 if (!cgroup_parent(blkcg->css.cgroup))
2158 return;
2159
2160 cpu = get_cpu();
2161 bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu);
2162 flags = u64_stats_update_begin_irqsave(&bis->sync);
2163
2164 /*
2165 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
2166 * bio and we would have already accounted for the size of the bio.
2167 */
2168 if (!bio_flagged(bio, BIO_CGROUP_ACCT)) {
2169 bio_set_flag(bio, BIO_CGROUP_ACCT);
2170 bis->cur.bytes[rwd] += bio->bi_iter.bi_size;
2171 }
2172 bis->cur.ios[rwd]++;
2173
2174 /*
2175 * If the iostat_cpu isn't in a lockless list, put it into the
2176 * list to indicate that a stat update is pending.
2177 */
2178 if (!READ_ONCE(bis->lqueued)) {
2179 struct llist_head *lhead = this_cpu_ptr(blkcg->lhead);
2180
2181 llist_add(&bis->lnode, lhead);
2182 WRITE_ONCE(bis->lqueued, true);
2183 }
2184
2185 u64_stats_update_end_irqrestore(&bis->sync, flags);
2186 cgroup_rstat_updated(blkcg->css.cgroup, cpu);
2187 put_cpu();
2188}
2189
2190bool blk_cgroup_congested(void)
2191{
2192 struct cgroup_subsys_state *css;
2193 bool ret = false;
2194
2195 rcu_read_lock();
2196 for (css = blkcg_css(); css; css = css->parent) {
2197 if (atomic_read(&css->cgroup->congestion_count)) {
2198 ret = true;
2199 break;
2200 }
2201 }
2202 rcu_read_unlock();
2203 return ret;
2204}
2205
2206module_param(blkcg_debug_stats, bool, 0644);
2207MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");