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#include <linux/ioprio.h>
14#include <linux/kdev_t.h>
15#include <linux/module.h>
16#include <linux/err.h>
17#include <linux/blkdev.h>
18#include <linux/slab.h>
19#include <linux/genhd.h>
20#include <linux/delay.h>
21#include <linux/atomic.h>
22#include "blk-cgroup.h"
23#include "blk.h"
24
25#define MAX_KEY_LEN 100
26
27static DEFINE_MUTEX(blkcg_pol_mutex);
28
29struct blkcg blkcg_root = { .cfq_weight = 2 * CFQ_WEIGHT_DEFAULT,
30 .cfq_leaf_weight = 2 * CFQ_WEIGHT_DEFAULT, };
31EXPORT_SYMBOL_GPL(blkcg_root);
32
33static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
34
35static bool blkcg_policy_enabled(struct request_queue *q,
36 const struct blkcg_policy *pol)
37{
38 return pol && test_bit(pol->plid, q->blkcg_pols);
39}
40
41/**
42 * blkg_free - free a blkg
43 * @blkg: blkg to free
44 *
45 * Free @blkg which may be partially allocated.
46 */
47static void blkg_free(struct blkcg_gq *blkg)
48{
49 int i;
50
51 if (!blkg)
52 return;
53
54 for (i = 0; i < BLKCG_MAX_POLS; i++)
55 kfree(blkg->pd[i]);
56
57 blk_exit_rl(&blkg->rl);
58 kfree(blkg);
59}
60
61/**
62 * blkg_alloc - allocate a blkg
63 * @blkcg: block cgroup the new blkg is associated with
64 * @q: request_queue the new blkg is associated with
65 * @gfp_mask: allocation mask to use
66 *
67 * Allocate a new blkg assocating @blkcg and @q.
68 */
69static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
70 gfp_t gfp_mask)
71{
72 struct blkcg_gq *blkg;
73 int i;
74
75 /* alloc and init base part */
76 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
77 if (!blkg)
78 return NULL;
79
80 blkg->q = q;
81 INIT_LIST_HEAD(&blkg->q_node);
82 blkg->blkcg = blkcg;
83 blkg->refcnt = 1;
84
85 /* root blkg uses @q->root_rl, init rl only for !root blkgs */
86 if (blkcg != &blkcg_root) {
87 if (blk_init_rl(&blkg->rl, q, gfp_mask))
88 goto err_free;
89 blkg->rl.blkg = blkg;
90 }
91
92 for (i = 0; i < BLKCG_MAX_POLS; i++) {
93 struct blkcg_policy *pol = blkcg_policy[i];
94 struct blkg_policy_data *pd;
95
96 if (!blkcg_policy_enabled(q, pol))
97 continue;
98
99 /* alloc per-policy data and attach it to blkg */
100 pd = kzalloc_node(pol->pd_size, gfp_mask, q->node);
101 if (!pd)
102 goto err_free;
103
104 blkg->pd[i] = pd;
105 pd->blkg = blkg;
106 pd->plid = i;
107 }
108
109 return blkg;
110
111err_free:
112 blkg_free(blkg);
113 return NULL;
114}
115
116/**
117 * __blkg_lookup - internal version of blkg_lookup()
118 * @blkcg: blkcg of interest
119 * @q: request_queue of interest
120 * @update_hint: whether to update lookup hint with the result or not
121 *
122 * This is internal version and shouldn't be used by policy
123 * implementations. Looks up blkgs for the @blkcg - @q pair regardless of
124 * @q's bypass state. If @update_hint is %true, the caller should be
125 * holding @q->queue_lock and lookup hint is updated on success.
126 */
127struct blkcg_gq *__blkg_lookup(struct blkcg *blkcg, struct request_queue *q,
128 bool update_hint)
129{
130 struct blkcg_gq *blkg;
131
132 blkg = rcu_dereference(blkcg->blkg_hint);
133 if (blkg && blkg->q == q)
134 return blkg;
135
136 /*
137 * Hint didn't match. Look up from the radix tree. Note that the
138 * hint can only be updated under queue_lock as otherwise @blkg
139 * could have already been removed from blkg_tree. The caller is
140 * responsible for grabbing queue_lock if @update_hint.
141 */
142 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
143 if (blkg && blkg->q == q) {
144 if (update_hint) {
145 lockdep_assert_held(q->queue_lock);
146 rcu_assign_pointer(blkcg->blkg_hint, blkg);
147 }
148 return blkg;
149 }
150
151 return NULL;
152}
153
154/**
155 * blkg_lookup - lookup blkg for the specified blkcg - q pair
156 * @blkcg: blkcg of interest
157 * @q: request_queue of interest
158 *
159 * Lookup blkg for the @blkcg - @q pair. This function should be called
160 * under RCU read lock and is guaranteed to return %NULL if @q is bypassing
161 * - see blk_queue_bypass_start() for details.
162 */
163struct blkcg_gq *blkg_lookup(struct blkcg *blkcg, struct request_queue *q)
164{
165 WARN_ON_ONCE(!rcu_read_lock_held());
166
167 if (unlikely(blk_queue_bypass(q)))
168 return NULL;
169 return __blkg_lookup(blkcg, q, false);
170}
171EXPORT_SYMBOL_GPL(blkg_lookup);
172
173/*
174 * If @new_blkg is %NULL, this function tries to allocate a new one as
175 * necessary using %GFP_ATOMIC. @new_blkg is always consumed on return.
176 */
177static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
178 struct request_queue *q,
179 struct blkcg_gq *new_blkg)
180{
181 struct blkcg_gq *blkg;
182 int i, ret;
183
184 WARN_ON_ONCE(!rcu_read_lock_held());
185 lockdep_assert_held(q->queue_lock);
186
187 /* blkg holds a reference to blkcg */
188 if (!css_tryget(&blkcg->css)) {
189 ret = -EINVAL;
190 goto err_free_blkg;
191 }
192
193 /* allocate */
194 if (!new_blkg) {
195 new_blkg = blkg_alloc(blkcg, q, GFP_ATOMIC);
196 if (unlikely(!new_blkg)) {
197 ret = -ENOMEM;
198 goto err_put_css;
199 }
200 }
201 blkg = new_blkg;
202
203 /* link parent */
204 if (blkcg_parent(blkcg)) {
205 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
206 if (WARN_ON_ONCE(!blkg->parent)) {
207 ret = -EINVAL;
208 goto err_put_css;
209 }
210 blkg_get(blkg->parent);
211 }
212
213 /* invoke per-policy init */
214 for (i = 0; i < BLKCG_MAX_POLS; i++) {
215 struct blkcg_policy *pol = blkcg_policy[i];
216
217 if (blkg->pd[i] && pol->pd_init_fn)
218 pol->pd_init_fn(blkg);
219 }
220
221 /* insert */
222 spin_lock(&blkcg->lock);
223 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
224 if (likely(!ret)) {
225 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
226 list_add(&blkg->q_node, &q->blkg_list);
227
228 for (i = 0; i < BLKCG_MAX_POLS; i++) {
229 struct blkcg_policy *pol = blkcg_policy[i];
230
231 if (blkg->pd[i] && pol->pd_online_fn)
232 pol->pd_online_fn(blkg);
233 }
234 }
235 blkg->online = true;
236 spin_unlock(&blkcg->lock);
237
238 if (!ret) {
239 if (blkcg == &blkcg_root) {
240 q->root_blkg = blkg;
241 q->root_rl.blkg = blkg;
242 }
243 return blkg;
244 }
245
246 /* @blkg failed fully initialized, use the usual release path */
247 blkg_put(blkg);
248 return ERR_PTR(ret);
249
250err_put_css:
251 css_put(&blkcg->css);
252err_free_blkg:
253 blkg_free(new_blkg);
254 return ERR_PTR(ret);
255}
256
257/**
258 * blkg_lookup_create - lookup blkg, try to create one if not there
259 * @blkcg: blkcg of interest
260 * @q: request_queue of interest
261 *
262 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
263 * create one. blkg creation is performed recursively from blkcg_root such
264 * that all non-root blkg's have access to the parent blkg. This function
265 * should be called under RCU read lock and @q->queue_lock.
266 *
267 * Returns pointer to the looked up or created blkg on success, ERR_PTR()
268 * value on error. If @q is dead, returns ERR_PTR(-EINVAL). If @q is not
269 * dead and bypassing, returns ERR_PTR(-EBUSY).
270 */
271struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
272 struct request_queue *q)
273{
274 struct blkcg_gq *blkg;
275
276 WARN_ON_ONCE(!rcu_read_lock_held());
277 lockdep_assert_held(q->queue_lock);
278
279 /*
280 * This could be the first entry point of blkcg implementation and
281 * we shouldn't allow anything to go through for a bypassing queue.
282 */
283 if (unlikely(blk_queue_bypass(q)))
284 return ERR_PTR(blk_queue_dying(q) ? -EINVAL : -EBUSY);
285
286 blkg = __blkg_lookup(blkcg, q, true);
287 if (blkg)
288 return blkg;
289
290 /*
291 * Create blkgs walking down from blkcg_root to @blkcg, so that all
292 * non-root blkgs have access to their parents.
293 */
294 while (true) {
295 struct blkcg *pos = blkcg;
296 struct blkcg *parent = blkcg_parent(blkcg);
297
298 while (parent && !__blkg_lookup(parent, q, false)) {
299 pos = parent;
300 parent = blkcg_parent(parent);
301 }
302
303 blkg = blkg_create(pos, q, NULL);
304 if (pos == blkcg || IS_ERR(blkg))
305 return blkg;
306 }
307}
308EXPORT_SYMBOL_GPL(blkg_lookup_create);
309
310static void blkg_destroy(struct blkcg_gq *blkg)
311{
312 struct blkcg *blkcg = blkg->blkcg;
313 int i;
314
315 lockdep_assert_held(blkg->q->queue_lock);
316 lockdep_assert_held(&blkcg->lock);
317
318 /* Something wrong if we are trying to remove same group twice */
319 WARN_ON_ONCE(list_empty(&blkg->q_node));
320 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
321
322 for (i = 0; i < BLKCG_MAX_POLS; i++) {
323 struct blkcg_policy *pol = blkcg_policy[i];
324
325 if (blkg->pd[i] && pol->pd_offline_fn)
326 pol->pd_offline_fn(blkg);
327 }
328 blkg->online = false;
329
330 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
331 list_del_init(&blkg->q_node);
332 hlist_del_init_rcu(&blkg->blkcg_node);
333
334 /*
335 * Both setting lookup hint to and clearing it from @blkg are done
336 * under queue_lock. If it's not pointing to @blkg now, it never
337 * will. Hint assignment itself can race safely.
338 */
339 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
340 rcu_assign_pointer(blkcg->blkg_hint, NULL);
341
342 /*
343 * If root blkg is destroyed. Just clear the pointer since root_rl
344 * does not take reference on root blkg.
345 */
346 if (blkcg == &blkcg_root) {
347 blkg->q->root_blkg = NULL;
348 blkg->q->root_rl.blkg = NULL;
349 }
350
351 /*
352 * Put the reference taken at the time of creation so that when all
353 * queues are gone, group can be destroyed.
354 */
355 blkg_put(blkg);
356}
357
358/**
359 * blkg_destroy_all - destroy all blkgs associated with a request_queue
360 * @q: request_queue of interest
361 *
362 * Destroy all blkgs associated with @q.
363 */
364static void blkg_destroy_all(struct request_queue *q)
365{
366 struct blkcg_gq *blkg, *n;
367
368 lockdep_assert_held(q->queue_lock);
369
370 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
371 struct blkcg *blkcg = blkg->blkcg;
372
373 spin_lock(&blkcg->lock);
374 blkg_destroy(blkg);
375 spin_unlock(&blkcg->lock);
376 }
377}
378
379/*
380 * A group is RCU protected, but having an rcu lock does not mean that one
381 * can access all the fields of blkg and assume these are valid. For
382 * example, don't try to follow throtl_data and request queue links.
383 *
384 * Having a reference to blkg under an rcu allows accesses to only values
385 * local to groups like group stats and group rate limits.
386 */
387void __blkg_release_rcu(struct rcu_head *rcu_head)
388{
389 struct blkcg_gq *blkg = container_of(rcu_head, struct blkcg_gq, rcu_head);
390 int i;
391
392 /* tell policies that this one is being freed */
393 for (i = 0; i < BLKCG_MAX_POLS; i++) {
394 struct blkcg_policy *pol = blkcg_policy[i];
395
396 if (blkg->pd[i] && pol->pd_exit_fn)
397 pol->pd_exit_fn(blkg);
398 }
399
400 /* release the blkcg and parent blkg refs this blkg has been holding */
401 css_put(&blkg->blkcg->css);
402 if (blkg->parent) {
403 spin_lock_irq(blkg->q->queue_lock);
404 blkg_put(blkg->parent);
405 spin_unlock_irq(blkg->q->queue_lock);
406 }
407
408 blkg_free(blkg);
409}
410EXPORT_SYMBOL_GPL(__blkg_release_rcu);
411
412/*
413 * The next function used by blk_queue_for_each_rl(). It's a bit tricky
414 * because the root blkg uses @q->root_rl instead of its own rl.
415 */
416struct request_list *__blk_queue_next_rl(struct request_list *rl,
417 struct request_queue *q)
418{
419 struct list_head *ent;
420 struct blkcg_gq *blkg;
421
422 /*
423 * Determine the current blkg list_head. The first entry is
424 * root_rl which is off @q->blkg_list and mapped to the head.
425 */
426 if (rl == &q->root_rl) {
427 ent = &q->blkg_list;
428 /* There are no more block groups, hence no request lists */
429 if (list_empty(ent))
430 return NULL;
431 } else {
432 blkg = container_of(rl, struct blkcg_gq, rl);
433 ent = &blkg->q_node;
434 }
435
436 /* walk to the next list_head, skip root blkcg */
437 ent = ent->next;
438 if (ent == &q->root_blkg->q_node)
439 ent = ent->next;
440 if (ent == &q->blkg_list)
441 return NULL;
442
443 blkg = container_of(ent, struct blkcg_gq, q_node);
444 return &blkg->rl;
445}
446
447static int blkcg_reset_stats(struct cgroup_subsys_state *css,
448 struct cftype *cftype, u64 val)
449{
450 struct blkcg *blkcg = css_to_blkcg(css);
451 struct blkcg_gq *blkg;
452 int i;
453
454 /*
455 * XXX: We invoke cgroup_add/rm_cftypes() under blkcg_pol_mutex
456 * which ends up putting cgroup's internal cgroup_tree_mutex under
457 * it; however, cgroup_tree_mutex is nested above cgroup file
458 * active protection and grabbing blkcg_pol_mutex from a cgroup
459 * file operation creates a possible circular dependency. cgroup
460 * internal locking is planned to go through further simplification
461 * and this issue should go away soon. For now, let's trylock
462 * blkcg_pol_mutex and restart the write on failure.
463 *
464 * http://lkml.kernel.org/g/5363C04B.4010400@oracle.com
465 */
466 if (!mutex_trylock(&blkcg_pol_mutex))
467 return restart_syscall();
468 spin_lock_irq(&blkcg->lock);
469
470 /*
471 * Note that stat reset is racy - it doesn't synchronize against
472 * stat updates. This is a debug feature which shouldn't exist
473 * anyway. If you get hit by a race, retry.
474 */
475 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
476 for (i = 0; i < BLKCG_MAX_POLS; i++) {
477 struct blkcg_policy *pol = blkcg_policy[i];
478
479 if (blkcg_policy_enabled(blkg->q, pol) &&
480 pol->pd_reset_stats_fn)
481 pol->pd_reset_stats_fn(blkg);
482 }
483 }
484
485 spin_unlock_irq(&blkcg->lock);
486 mutex_unlock(&blkcg_pol_mutex);
487 return 0;
488}
489
490static const char *blkg_dev_name(struct blkcg_gq *blkg)
491{
492 /* some drivers (floppy) instantiate a queue w/o disk registered */
493 if (blkg->q->backing_dev_info.dev)
494 return dev_name(blkg->q->backing_dev_info.dev);
495 return NULL;
496}
497
498/**
499 * blkcg_print_blkgs - helper for printing per-blkg data
500 * @sf: seq_file to print to
501 * @blkcg: blkcg of interest
502 * @prfill: fill function to print out a blkg
503 * @pol: policy in question
504 * @data: data to be passed to @prfill
505 * @show_total: to print out sum of prfill return values or not
506 *
507 * This function invokes @prfill on each blkg of @blkcg if pd for the
508 * policy specified by @pol exists. @prfill is invoked with @sf, the
509 * policy data and @data and the matching queue lock held. If @show_total
510 * is %true, the sum of the return values from @prfill is printed with
511 * "Total" label at the end.
512 *
513 * This is to be used to construct print functions for
514 * cftype->read_seq_string method.
515 */
516void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
517 u64 (*prfill)(struct seq_file *,
518 struct blkg_policy_data *, int),
519 const struct blkcg_policy *pol, int data,
520 bool show_total)
521{
522 struct blkcg_gq *blkg;
523 u64 total = 0;
524
525 rcu_read_lock();
526 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
527 spin_lock_irq(blkg->q->queue_lock);
528 if (blkcg_policy_enabled(blkg->q, pol))
529 total += prfill(sf, blkg->pd[pol->plid], data);
530 spin_unlock_irq(blkg->q->queue_lock);
531 }
532 rcu_read_unlock();
533
534 if (show_total)
535 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
536}
537EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
538
539/**
540 * __blkg_prfill_u64 - prfill helper for a single u64 value
541 * @sf: seq_file to print to
542 * @pd: policy private data of interest
543 * @v: value to print
544 *
545 * Print @v to @sf for the device assocaited with @pd.
546 */
547u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
548{
549 const char *dname = blkg_dev_name(pd->blkg);
550
551 if (!dname)
552 return 0;
553
554 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
555 return v;
556}
557EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
558
559/**
560 * __blkg_prfill_rwstat - prfill helper for a blkg_rwstat
561 * @sf: seq_file to print to
562 * @pd: policy private data of interest
563 * @rwstat: rwstat to print
564 *
565 * Print @rwstat to @sf for the device assocaited with @pd.
566 */
567u64 __blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
568 const struct blkg_rwstat *rwstat)
569{
570 static const char *rwstr[] = {
571 [BLKG_RWSTAT_READ] = "Read",
572 [BLKG_RWSTAT_WRITE] = "Write",
573 [BLKG_RWSTAT_SYNC] = "Sync",
574 [BLKG_RWSTAT_ASYNC] = "Async",
575 };
576 const char *dname = blkg_dev_name(pd->blkg);
577 u64 v;
578 int i;
579
580 if (!dname)
581 return 0;
582
583 for (i = 0; i < BLKG_RWSTAT_NR; i++)
584 seq_printf(sf, "%s %s %llu\n", dname, rwstr[i],
585 (unsigned long long)rwstat->cnt[i]);
586
587 v = rwstat->cnt[BLKG_RWSTAT_READ] + rwstat->cnt[BLKG_RWSTAT_WRITE];
588 seq_printf(sf, "%s Total %llu\n", dname, (unsigned long long)v);
589 return v;
590}
591EXPORT_SYMBOL_GPL(__blkg_prfill_rwstat);
592
593/**
594 * blkg_prfill_stat - prfill callback for blkg_stat
595 * @sf: seq_file to print to
596 * @pd: policy private data of interest
597 * @off: offset to the blkg_stat in @pd
598 *
599 * prfill callback for printing a blkg_stat.
600 */
601u64 blkg_prfill_stat(struct seq_file *sf, struct blkg_policy_data *pd, int off)
602{
603 return __blkg_prfill_u64(sf, pd, blkg_stat_read((void *)pd + off));
604}
605EXPORT_SYMBOL_GPL(blkg_prfill_stat);
606
607/**
608 * blkg_prfill_rwstat - prfill callback for blkg_rwstat
609 * @sf: seq_file to print to
610 * @pd: policy private data of interest
611 * @off: offset to the blkg_rwstat in @pd
612 *
613 * prfill callback for printing a blkg_rwstat.
614 */
615u64 blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
616 int off)
617{
618 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd + off);
619
620 return __blkg_prfill_rwstat(sf, pd, &rwstat);
621}
622EXPORT_SYMBOL_GPL(blkg_prfill_rwstat);
623
624/**
625 * blkg_stat_recursive_sum - collect hierarchical blkg_stat
626 * @pd: policy private data of interest
627 * @off: offset to the blkg_stat in @pd
628 *
629 * Collect the blkg_stat specified by @off from @pd and all its online
630 * descendants and return the sum. The caller must be holding the queue
631 * lock for online tests.
632 */
633u64 blkg_stat_recursive_sum(struct blkg_policy_data *pd, int off)
634{
635 struct blkcg_policy *pol = blkcg_policy[pd->plid];
636 struct blkcg_gq *pos_blkg;
637 struct cgroup_subsys_state *pos_css;
638 u64 sum = 0;
639
640 lockdep_assert_held(pd->blkg->q->queue_lock);
641
642 rcu_read_lock();
643 blkg_for_each_descendant_pre(pos_blkg, pos_css, pd_to_blkg(pd)) {
644 struct blkg_policy_data *pos_pd = blkg_to_pd(pos_blkg, pol);
645 struct blkg_stat *stat = (void *)pos_pd + off;
646
647 if (pos_blkg->online)
648 sum += blkg_stat_read(stat);
649 }
650 rcu_read_unlock();
651
652 return sum;
653}
654EXPORT_SYMBOL_GPL(blkg_stat_recursive_sum);
655
656/**
657 * blkg_rwstat_recursive_sum - collect hierarchical blkg_rwstat
658 * @pd: policy private data of interest
659 * @off: offset to the blkg_stat in @pd
660 *
661 * Collect the blkg_rwstat specified by @off from @pd and all its online
662 * descendants and return the sum. The caller must be holding the queue
663 * lock for online tests.
664 */
665struct blkg_rwstat blkg_rwstat_recursive_sum(struct blkg_policy_data *pd,
666 int off)
667{
668 struct blkcg_policy *pol = blkcg_policy[pd->plid];
669 struct blkcg_gq *pos_blkg;
670 struct cgroup_subsys_state *pos_css;
671 struct blkg_rwstat sum = { };
672 int i;
673
674 lockdep_assert_held(pd->blkg->q->queue_lock);
675
676 rcu_read_lock();
677 blkg_for_each_descendant_pre(pos_blkg, pos_css, pd_to_blkg(pd)) {
678 struct blkg_policy_data *pos_pd = blkg_to_pd(pos_blkg, pol);
679 struct blkg_rwstat *rwstat = (void *)pos_pd + off;
680 struct blkg_rwstat tmp;
681
682 if (!pos_blkg->online)
683 continue;
684
685 tmp = blkg_rwstat_read(rwstat);
686
687 for (i = 0; i < BLKG_RWSTAT_NR; i++)
688 sum.cnt[i] += tmp.cnt[i];
689 }
690 rcu_read_unlock();
691
692 return sum;
693}
694EXPORT_SYMBOL_GPL(blkg_rwstat_recursive_sum);
695
696/**
697 * blkg_conf_prep - parse and prepare for per-blkg config update
698 * @blkcg: target block cgroup
699 * @pol: target policy
700 * @input: input string
701 * @ctx: blkg_conf_ctx to be filled
702 *
703 * Parse per-blkg config update from @input and initialize @ctx with the
704 * result. @ctx->blkg points to the blkg to be updated and @ctx->v the new
705 * value. This function returns with RCU read lock and queue lock held and
706 * must be paired with blkg_conf_finish().
707 */
708int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
709 const char *input, struct blkg_conf_ctx *ctx)
710 __acquires(rcu) __acquires(disk->queue->queue_lock)
711{
712 struct gendisk *disk;
713 struct blkcg_gq *blkg;
714 unsigned int major, minor;
715 unsigned long long v;
716 int part, ret;
717
718 if (sscanf(input, "%u:%u %llu", &major, &minor, &v) != 3)
719 return -EINVAL;
720
721 disk = get_gendisk(MKDEV(major, minor), &part);
722 if (!disk || part)
723 return -EINVAL;
724
725 rcu_read_lock();
726 spin_lock_irq(disk->queue->queue_lock);
727
728 if (blkcg_policy_enabled(disk->queue, pol))
729 blkg = blkg_lookup_create(blkcg, disk->queue);
730 else
731 blkg = ERR_PTR(-EINVAL);
732
733 if (IS_ERR(blkg)) {
734 ret = PTR_ERR(blkg);
735 rcu_read_unlock();
736 spin_unlock_irq(disk->queue->queue_lock);
737 put_disk(disk);
738 /*
739 * If queue was bypassing, we should retry. Do so after a
740 * short msleep(). It isn't strictly necessary but queue
741 * can be bypassing for some time and it's always nice to
742 * avoid busy looping.
743 */
744 if (ret == -EBUSY) {
745 msleep(10);
746 ret = restart_syscall();
747 }
748 return ret;
749 }
750
751 ctx->disk = disk;
752 ctx->blkg = blkg;
753 ctx->v = v;
754 return 0;
755}
756EXPORT_SYMBOL_GPL(blkg_conf_prep);
757
758/**
759 * blkg_conf_finish - finish up per-blkg config update
760 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
761 *
762 * Finish up after per-blkg config update. This function must be paired
763 * with blkg_conf_prep().
764 */
765void blkg_conf_finish(struct blkg_conf_ctx *ctx)
766 __releases(ctx->disk->queue->queue_lock) __releases(rcu)
767{
768 spin_unlock_irq(ctx->disk->queue->queue_lock);
769 rcu_read_unlock();
770 put_disk(ctx->disk);
771}
772EXPORT_SYMBOL_GPL(blkg_conf_finish);
773
774struct cftype blkcg_files[] = {
775 {
776 .name = "reset_stats",
777 .write_u64 = blkcg_reset_stats,
778 },
779 { } /* terminate */
780};
781
782/**
783 * blkcg_css_offline - cgroup css_offline callback
784 * @css: css of interest
785 *
786 * This function is called when @css is about to go away and responsible
787 * for shooting down all blkgs associated with @css. blkgs should be
788 * removed while holding both q and blkcg locks. As blkcg lock is nested
789 * inside q lock, this function performs reverse double lock dancing.
790 *
791 * This is the blkcg counterpart of ioc_release_fn().
792 */
793static void blkcg_css_offline(struct cgroup_subsys_state *css)
794{
795 struct blkcg *blkcg = css_to_blkcg(css);
796
797 spin_lock_irq(&blkcg->lock);
798
799 while (!hlist_empty(&blkcg->blkg_list)) {
800 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
801 struct blkcg_gq, blkcg_node);
802 struct request_queue *q = blkg->q;
803
804 if (spin_trylock(q->queue_lock)) {
805 blkg_destroy(blkg);
806 spin_unlock(q->queue_lock);
807 } else {
808 spin_unlock_irq(&blkcg->lock);
809 cpu_relax();
810 spin_lock_irq(&blkcg->lock);
811 }
812 }
813
814 spin_unlock_irq(&blkcg->lock);
815}
816
817static void blkcg_css_free(struct cgroup_subsys_state *css)
818{
819 struct blkcg *blkcg = css_to_blkcg(css);
820
821 if (blkcg != &blkcg_root)
822 kfree(blkcg);
823}
824
825static struct cgroup_subsys_state *
826blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
827{
828 static atomic64_t id_seq = ATOMIC64_INIT(0);
829 struct blkcg *blkcg;
830
831 if (!parent_css) {
832 blkcg = &blkcg_root;
833 goto done;
834 }
835
836 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
837 if (!blkcg)
838 return ERR_PTR(-ENOMEM);
839
840 blkcg->cfq_weight = CFQ_WEIGHT_DEFAULT;
841 blkcg->cfq_leaf_weight = CFQ_WEIGHT_DEFAULT;
842 blkcg->id = atomic64_inc_return(&id_seq); /* root is 0, start from 1 */
843done:
844 spin_lock_init(&blkcg->lock);
845 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_ATOMIC);
846 INIT_HLIST_HEAD(&blkcg->blkg_list);
847
848 return &blkcg->css;
849}
850
851/**
852 * blkcg_init_queue - initialize blkcg part of request queue
853 * @q: request_queue to initialize
854 *
855 * Called from blk_alloc_queue_node(). Responsible for initializing blkcg
856 * part of new request_queue @q.
857 *
858 * RETURNS:
859 * 0 on success, -errno on failure.
860 */
861int blkcg_init_queue(struct request_queue *q)
862{
863 might_sleep();
864
865 return blk_throtl_init(q);
866}
867
868/**
869 * blkcg_drain_queue - drain blkcg part of request_queue
870 * @q: request_queue to drain
871 *
872 * Called from blk_drain_queue(). Responsible for draining blkcg part.
873 */
874void blkcg_drain_queue(struct request_queue *q)
875{
876 lockdep_assert_held(q->queue_lock);
877
878 blk_throtl_drain(q);
879}
880
881/**
882 * blkcg_exit_queue - exit and release blkcg part of request_queue
883 * @q: request_queue being released
884 *
885 * Called from blk_release_queue(). Responsible for exiting blkcg part.
886 */
887void blkcg_exit_queue(struct request_queue *q)
888{
889 spin_lock_irq(q->queue_lock);
890 blkg_destroy_all(q);
891 spin_unlock_irq(q->queue_lock);
892
893 blk_throtl_exit(q);
894}
895
896/*
897 * We cannot support shared io contexts, as we have no mean to support
898 * two tasks with the same ioc in two different groups without major rework
899 * of the main cic data structures. For now we allow a task to change
900 * its cgroup only if it's the only owner of its ioc.
901 */
902static int blkcg_can_attach(struct cgroup_subsys_state *css,
903 struct cgroup_taskset *tset)
904{
905 struct task_struct *task;
906 struct io_context *ioc;
907 int ret = 0;
908
909 /* task_lock() is needed to avoid races with exit_io_context() */
910 cgroup_taskset_for_each(task, tset) {
911 task_lock(task);
912 ioc = task->io_context;
913 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
914 ret = -EINVAL;
915 task_unlock(task);
916 if (ret)
917 break;
918 }
919 return ret;
920}
921
922struct cgroup_subsys blkio_cgrp_subsys = {
923 .css_alloc = blkcg_css_alloc,
924 .css_offline = blkcg_css_offline,
925 .css_free = blkcg_css_free,
926 .can_attach = blkcg_can_attach,
927 .base_cftypes = blkcg_files,
928};
929EXPORT_SYMBOL_GPL(blkio_cgrp_subsys);
930
931/**
932 * blkcg_activate_policy - activate a blkcg policy on a request_queue
933 * @q: request_queue of interest
934 * @pol: blkcg policy to activate
935 *
936 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
937 * bypass mode to populate its blkgs with policy_data for @pol.
938 *
939 * Activation happens with @q bypassed, so nobody would be accessing blkgs
940 * from IO path. Update of each blkg is protected by both queue and blkcg
941 * locks so that holding either lock and testing blkcg_policy_enabled() is
942 * always enough for dereferencing policy data.
943 *
944 * The caller is responsible for synchronizing [de]activations and policy
945 * [un]registerations. Returns 0 on success, -errno on failure.
946 */
947int blkcg_activate_policy(struct request_queue *q,
948 const struct blkcg_policy *pol)
949{
950 LIST_HEAD(pds);
951 struct blkcg_gq *blkg, *new_blkg;
952 struct blkg_policy_data *pd, *n;
953 int cnt = 0, ret;
954 bool preloaded;
955
956 if (blkcg_policy_enabled(q, pol))
957 return 0;
958
959 /* preallocations for root blkg */
960 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
961 if (!new_blkg)
962 return -ENOMEM;
963
964 blk_queue_bypass_start(q);
965
966 preloaded = !radix_tree_preload(GFP_KERNEL);
967
968 /*
969 * Make sure the root blkg exists and count the existing blkgs. As
970 * @q is bypassing at this point, blkg_lookup_create() can't be
971 * used. Open code it.
972 */
973 spin_lock_irq(q->queue_lock);
974
975 rcu_read_lock();
976 blkg = __blkg_lookup(&blkcg_root, q, false);
977 if (blkg)
978 blkg_free(new_blkg);
979 else
980 blkg = blkg_create(&blkcg_root, q, new_blkg);
981 rcu_read_unlock();
982
983 if (preloaded)
984 radix_tree_preload_end();
985
986 if (IS_ERR(blkg)) {
987 ret = PTR_ERR(blkg);
988 goto out_unlock;
989 }
990
991 list_for_each_entry(blkg, &q->blkg_list, q_node)
992 cnt++;
993
994 spin_unlock_irq(q->queue_lock);
995
996 /* allocate policy_data for all existing blkgs */
997 while (cnt--) {
998 pd = kzalloc_node(pol->pd_size, GFP_KERNEL, q->node);
999 if (!pd) {
1000 ret = -ENOMEM;
1001 goto out_free;
1002 }
1003 list_add_tail(&pd->alloc_node, &pds);
1004 }
1005
1006 /*
1007 * Install the allocated pds. With @q bypassing, no new blkg
1008 * should have been created while the queue lock was dropped.
1009 */
1010 spin_lock_irq(q->queue_lock);
1011
1012 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1013 if (WARN_ON(list_empty(&pds))) {
1014 /* umm... this shouldn't happen, just abort */
1015 ret = -ENOMEM;
1016 goto out_unlock;
1017 }
1018 pd = list_first_entry(&pds, struct blkg_policy_data, alloc_node);
1019 list_del_init(&pd->alloc_node);
1020
1021 /* grab blkcg lock too while installing @pd on @blkg */
1022 spin_lock(&blkg->blkcg->lock);
1023
1024 blkg->pd[pol->plid] = pd;
1025 pd->blkg = blkg;
1026 pd->plid = pol->plid;
1027 pol->pd_init_fn(blkg);
1028
1029 spin_unlock(&blkg->blkcg->lock);
1030 }
1031
1032 __set_bit(pol->plid, q->blkcg_pols);
1033 ret = 0;
1034out_unlock:
1035 spin_unlock_irq(q->queue_lock);
1036out_free:
1037 blk_queue_bypass_end(q);
1038 list_for_each_entry_safe(pd, n, &pds, alloc_node)
1039 kfree(pd);
1040 return ret;
1041}
1042EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1043
1044/**
1045 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1046 * @q: request_queue of interest
1047 * @pol: blkcg policy to deactivate
1048 *
1049 * Deactivate @pol on @q. Follows the same synchronization rules as
1050 * blkcg_activate_policy().
1051 */
1052void blkcg_deactivate_policy(struct request_queue *q,
1053 const struct blkcg_policy *pol)
1054{
1055 struct blkcg_gq *blkg;
1056
1057 if (!blkcg_policy_enabled(q, pol))
1058 return;
1059
1060 blk_queue_bypass_start(q);
1061 spin_lock_irq(q->queue_lock);
1062
1063 __clear_bit(pol->plid, q->blkcg_pols);
1064
1065 /* if no policy is left, no need for blkgs - shoot them down */
1066 if (bitmap_empty(q->blkcg_pols, BLKCG_MAX_POLS))
1067 blkg_destroy_all(q);
1068
1069 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1070 /* grab blkcg lock too while removing @pd from @blkg */
1071 spin_lock(&blkg->blkcg->lock);
1072
1073 if (pol->pd_offline_fn)
1074 pol->pd_offline_fn(blkg);
1075 if (pol->pd_exit_fn)
1076 pol->pd_exit_fn(blkg);
1077
1078 kfree(blkg->pd[pol->plid]);
1079 blkg->pd[pol->plid] = NULL;
1080
1081 spin_unlock(&blkg->blkcg->lock);
1082 }
1083
1084 spin_unlock_irq(q->queue_lock);
1085 blk_queue_bypass_end(q);
1086}
1087EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1088
1089/**
1090 * blkcg_policy_register - register a blkcg policy
1091 * @pol: blkcg policy to register
1092 *
1093 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1094 * successful registration. Returns 0 on success and -errno on failure.
1095 */
1096int blkcg_policy_register(struct blkcg_policy *pol)
1097{
1098 int i, ret;
1099
1100 if (WARN_ON(pol->pd_size < sizeof(struct blkg_policy_data)))
1101 return -EINVAL;
1102
1103 mutex_lock(&blkcg_pol_mutex);
1104
1105 /* find an empty slot */
1106 ret = -ENOSPC;
1107 for (i = 0; i < BLKCG_MAX_POLS; i++)
1108 if (!blkcg_policy[i])
1109 break;
1110 if (i >= BLKCG_MAX_POLS)
1111 goto out_unlock;
1112
1113 /* register and update blkgs */
1114 pol->plid = i;
1115 blkcg_policy[i] = pol;
1116
1117 /* everything is in place, add intf files for the new policy */
1118 if (pol->cftypes)
1119 WARN_ON(cgroup_add_cftypes(&blkio_cgrp_subsys, pol->cftypes));
1120 ret = 0;
1121out_unlock:
1122 mutex_unlock(&blkcg_pol_mutex);
1123 return ret;
1124}
1125EXPORT_SYMBOL_GPL(blkcg_policy_register);
1126
1127/**
1128 * blkcg_policy_unregister - unregister a blkcg policy
1129 * @pol: blkcg policy to unregister
1130 *
1131 * Undo blkcg_policy_register(@pol). Might sleep.
1132 */
1133void blkcg_policy_unregister(struct blkcg_policy *pol)
1134{
1135 mutex_lock(&blkcg_pol_mutex);
1136
1137 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1138 goto out_unlock;
1139
1140 /* kill the intf files first */
1141 if (pol->cftypes)
1142 cgroup_rm_cftypes(pol->cftypes);
1143
1144 /* unregister and update blkgs */
1145 blkcg_policy[pol->plid] = NULL;
1146out_unlock:
1147 mutex_unlock(&blkcg_pol_mutex);
1148}
1149EXPORT_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/genhd.h>
27#include <linux/delay.h>
28#include <linux/atomic.h>
29#include <linux/ctype.h>
30#include <linux/blk-cgroup.h>
31#include <linux/tracehook.h>
32#include <linux/psi.h>
33#include "blk.h"
34
35#define MAX_KEY_LEN 100
36
37/*
38 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
39 * blkcg_pol_register_mutex nests outside of it and synchronizes entire
40 * policy [un]register operations including cgroup file additions /
41 * removals. Putting cgroup file registration outside blkcg_pol_mutex
42 * allows grabbing it from cgroup callbacks.
43 */
44static DEFINE_MUTEX(blkcg_pol_register_mutex);
45static DEFINE_MUTEX(blkcg_pol_mutex);
46
47struct blkcg blkcg_root;
48EXPORT_SYMBOL_GPL(blkcg_root);
49
50struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
51EXPORT_SYMBOL_GPL(blkcg_root_css);
52
53static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
54
55static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */
56
57bool blkcg_debug_stats = false;
58static struct workqueue_struct *blkcg_punt_bio_wq;
59
60static bool blkcg_policy_enabled(struct request_queue *q,
61 const struct blkcg_policy *pol)
62{
63 return pol && test_bit(pol->plid, q->blkcg_pols);
64}
65
66/**
67 * blkg_free - free a blkg
68 * @blkg: blkg to free
69 *
70 * Free @blkg which may be partially allocated.
71 */
72static void blkg_free(struct blkcg_gq *blkg)
73{
74 int i;
75
76 if (!blkg)
77 return;
78
79 for (i = 0; i < BLKCG_MAX_POLS; i++)
80 if (blkg->pd[i])
81 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
82
83 blkg_rwstat_exit(&blkg->stat_ios);
84 blkg_rwstat_exit(&blkg->stat_bytes);
85 percpu_ref_exit(&blkg->refcnt);
86 kfree(blkg);
87}
88
89static void __blkg_release(struct rcu_head *rcu)
90{
91 struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head);
92
93 WARN_ON(!bio_list_empty(&blkg->async_bios));
94
95 /* release the blkcg and parent blkg refs this blkg has been holding */
96 css_put(&blkg->blkcg->css);
97 if (blkg->parent)
98 blkg_put(blkg->parent);
99
100 wb_congested_put(blkg->wb_congested);
101
102 blkg_free(blkg);
103}
104
105/*
106 * A group is RCU protected, but having an rcu lock does not mean that one
107 * can access all the fields of blkg and assume these are valid. For
108 * example, don't try to follow throtl_data and request queue links.
109 *
110 * Having a reference to blkg under an rcu allows accesses to only values
111 * local to groups like group stats and group rate limits.
112 */
113static void blkg_release(struct percpu_ref *ref)
114{
115 struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt);
116
117 call_rcu(&blkg->rcu_head, __blkg_release);
118}
119
120static void blkg_async_bio_workfn(struct work_struct *work)
121{
122 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
123 async_bio_work);
124 struct bio_list bios = BIO_EMPTY_LIST;
125 struct bio *bio;
126
127 /* as long as there are pending bios, @blkg can't go away */
128 spin_lock_bh(&blkg->async_bio_lock);
129 bio_list_merge(&bios, &blkg->async_bios);
130 bio_list_init(&blkg->async_bios);
131 spin_unlock_bh(&blkg->async_bio_lock);
132
133 while ((bio = bio_list_pop(&bios)))
134 submit_bio(bio);
135}
136
137/**
138 * blkg_alloc - allocate a blkg
139 * @blkcg: block cgroup the new blkg is associated with
140 * @q: request_queue the new blkg is associated with
141 * @gfp_mask: allocation mask to use
142 *
143 * Allocate a new blkg assocating @blkcg and @q.
144 */
145static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
146 gfp_t gfp_mask)
147{
148 struct blkcg_gq *blkg;
149 int i;
150
151 /* alloc and init base part */
152 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
153 if (!blkg)
154 return NULL;
155
156 if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask))
157 goto err_free;
158
159 if (blkg_rwstat_init(&blkg->stat_bytes, gfp_mask) ||
160 blkg_rwstat_init(&blkg->stat_ios, gfp_mask))
161 goto err_free;
162
163 blkg->q = q;
164 INIT_LIST_HEAD(&blkg->q_node);
165 spin_lock_init(&blkg->async_bio_lock);
166 bio_list_init(&blkg->async_bios);
167 INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
168 blkg->blkcg = blkcg;
169
170 for (i = 0; i < BLKCG_MAX_POLS; i++) {
171 struct blkcg_policy *pol = blkcg_policy[i];
172 struct blkg_policy_data *pd;
173
174 if (!blkcg_policy_enabled(q, pol))
175 continue;
176
177 /* alloc per-policy data and attach it to blkg */
178 pd = pol->pd_alloc_fn(gfp_mask, q, blkcg);
179 if (!pd)
180 goto err_free;
181
182 blkg->pd[i] = pd;
183 pd->blkg = blkg;
184 pd->plid = i;
185 }
186
187 return blkg;
188
189err_free:
190 blkg_free(blkg);
191 return NULL;
192}
193
194struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg,
195 struct request_queue *q, bool update_hint)
196{
197 struct blkcg_gq *blkg;
198
199 /*
200 * Hint didn't match. Look up from the radix tree. Note that the
201 * hint can only be updated under queue_lock as otherwise @blkg
202 * could have already been removed from blkg_tree. The caller is
203 * responsible for grabbing queue_lock if @update_hint.
204 */
205 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
206 if (blkg && blkg->q == q) {
207 if (update_hint) {
208 lockdep_assert_held(&q->queue_lock);
209 rcu_assign_pointer(blkcg->blkg_hint, blkg);
210 }
211 return blkg;
212 }
213
214 return NULL;
215}
216EXPORT_SYMBOL_GPL(blkg_lookup_slowpath);
217
218/*
219 * If @new_blkg is %NULL, this function tries to allocate a new one as
220 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
221 */
222static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
223 struct request_queue *q,
224 struct blkcg_gq *new_blkg)
225{
226 struct blkcg_gq *blkg;
227 struct bdi_writeback_congested *wb_congested;
228 int i, ret;
229
230 WARN_ON_ONCE(!rcu_read_lock_held());
231 lockdep_assert_held(&q->queue_lock);
232
233 /* request_queue is dying, do not create/recreate a blkg */
234 if (blk_queue_dying(q)) {
235 ret = -ENODEV;
236 goto err_free_blkg;
237 }
238
239 /* blkg holds a reference to blkcg */
240 if (!css_tryget_online(&blkcg->css)) {
241 ret = -ENODEV;
242 goto err_free_blkg;
243 }
244
245 wb_congested = wb_congested_get_create(q->backing_dev_info,
246 blkcg->css.id,
247 GFP_NOWAIT | __GFP_NOWARN);
248 if (!wb_congested) {
249 ret = -ENOMEM;
250 goto err_put_css;
251 }
252
253 /* allocate */
254 if (!new_blkg) {
255 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN);
256 if (unlikely(!new_blkg)) {
257 ret = -ENOMEM;
258 goto err_put_congested;
259 }
260 }
261 blkg = new_blkg;
262 blkg->wb_congested = wb_congested;
263
264 /* link parent */
265 if (blkcg_parent(blkcg)) {
266 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
267 if (WARN_ON_ONCE(!blkg->parent)) {
268 ret = -ENODEV;
269 goto err_put_congested;
270 }
271 blkg_get(blkg->parent);
272 }
273
274 /* invoke per-policy init */
275 for (i = 0; i < BLKCG_MAX_POLS; i++) {
276 struct blkcg_policy *pol = blkcg_policy[i];
277
278 if (blkg->pd[i] && pol->pd_init_fn)
279 pol->pd_init_fn(blkg->pd[i]);
280 }
281
282 /* insert */
283 spin_lock(&blkcg->lock);
284 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
285 if (likely(!ret)) {
286 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
287 list_add(&blkg->q_node, &q->blkg_list);
288
289 for (i = 0; i < BLKCG_MAX_POLS; i++) {
290 struct blkcg_policy *pol = blkcg_policy[i];
291
292 if (blkg->pd[i] && pol->pd_online_fn)
293 pol->pd_online_fn(blkg->pd[i]);
294 }
295 }
296 blkg->online = true;
297 spin_unlock(&blkcg->lock);
298
299 if (!ret)
300 return blkg;
301
302 /* @blkg failed fully initialized, use the usual release path */
303 blkg_put(blkg);
304 return ERR_PTR(ret);
305
306err_put_congested:
307 wb_congested_put(wb_congested);
308err_put_css:
309 css_put(&blkcg->css);
310err_free_blkg:
311 blkg_free(new_blkg);
312 return ERR_PTR(ret);
313}
314
315/**
316 * __blkg_lookup_create - lookup blkg, try to create one if not there
317 * @blkcg: blkcg of interest
318 * @q: request_queue of interest
319 *
320 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
321 * create one. blkg creation is performed recursively from blkcg_root such
322 * that all non-root blkg's have access to the parent blkg. This function
323 * should be called under RCU read lock and @q->queue_lock.
324 *
325 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
326 * down from root.
327 */
328struct blkcg_gq *__blkg_lookup_create(struct blkcg *blkcg,
329 struct request_queue *q)
330{
331 struct blkcg_gq *blkg;
332
333 WARN_ON_ONCE(!rcu_read_lock_held());
334 lockdep_assert_held(&q->queue_lock);
335
336 blkg = __blkg_lookup(blkcg, q, true);
337 if (blkg)
338 return blkg;
339
340 /*
341 * Create blkgs walking down from blkcg_root to @blkcg, so that all
342 * non-root blkgs have access to their parents. Returns the closest
343 * blkg to the intended blkg should blkg_create() fail.
344 */
345 while (true) {
346 struct blkcg *pos = blkcg;
347 struct blkcg *parent = blkcg_parent(blkcg);
348 struct blkcg_gq *ret_blkg = q->root_blkg;
349
350 while (parent) {
351 blkg = __blkg_lookup(parent, q, false);
352 if (blkg) {
353 /* remember closest blkg */
354 ret_blkg = blkg;
355 break;
356 }
357 pos = parent;
358 parent = blkcg_parent(parent);
359 }
360
361 blkg = blkg_create(pos, q, NULL);
362 if (IS_ERR(blkg))
363 return ret_blkg;
364 if (pos == blkcg)
365 return blkg;
366 }
367}
368
369/**
370 * blkg_lookup_create - find or create a blkg
371 * @blkcg: target block cgroup
372 * @q: target request_queue
373 *
374 * This looks up or creates the blkg representing the unique pair
375 * of the blkcg and the request_queue.
376 */
377struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
378 struct request_queue *q)
379{
380 struct blkcg_gq *blkg = blkg_lookup(blkcg, q);
381
382 if (unlikely(!blkg)) {
383 unsigned long flags;
384
385 spin_lock_irqsave(&q->queue_lock, flags);
386 blkg = __blkg_lookup_create(blkcg, q);
387 spin_unlock_irqrestore(&q->queue_lock, flags);
388 }
389
390 return blkg;
391}
392
393static void blkg_destroy(struct blkcg_gq *blkg)
394{
395 struct blkcg *blkcg = blkg->blkcg;
396 struct blkcg_gq *parent = blkg->parent;
397 int i;
398
399 lockdep_assert_held(&blkg->q->queue_lock);
400 lockdep_assert_held(&blkcg->lock);
401
402 /* Something wrong if we are trying to remove same group twice */
403 WARN_ON_ONCE(list_empty(&blkg->q_node));
404 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
405
406 for (i = 0; i < BLKCG_MAX_POLS; i++) {
407 struct blkcg_policy *pol = blkcg_policy[i];
408
409 if (blkg->pd[i] && pol->pd_offline_fn)
410 pol->pd_offline_fn(blkg->pd[i]);
411 }
412
413 if (parent) {
414 blkg_rwstat_add_aux(&parent->stat_bytes, &blkg->stat_bytes);
415 blkg_rwstat_add_aux(&parent->stat_ios, &blkg->stat_ios);
416 }
417
418 blkg->online = false;
419
420 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
421 list_del_init(&blkg->q_node);
422 hlist_del_init_rcu(&blkg->blkcg_node);
423
424 /*
425 * Both setting lookup hint to and clearing it from @blkg are done
426 * under queue_lock. If it's not pointing to @blkg now, it never
427 * will. Hint assignment itself can race safely.
428 */
429 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
430 rcu_assign_pointer(blkcg->blkg_hint, NULL);
431
432 /*
433 * Put the reference taken at the time of creation so that when all
434 * queues are gone, group can be destroyed.
435 */
436 percpu_ref_kill(&blkg->refcnt);
437}
438
439/**
440 * blkg_destroy_all - destroy all blkgs associated with a request_queue
441 * @q: request_queue of interest
442 *
443 * Destroy all blkgs associated with @q.
444 */
445static void blkg_destroy_all(struct request_queue *q)
446{
447 struct blkcg_gq *blkg, *n;
448
449 spin_lock_irq(&q->queue_lock);
450 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
451 struct blkcg *blkcg = blkg->blkcg;
452
453 spin_lock(&blkcg->lock);
454 blkg_destroy(blkg);
455 spin_unlock(&blkcg->lock);
456 }
457
458 q->root_blkg = NULL;
459 spin_unlock_irq(&q->queue_lock);
460}
461
462static int blkcg_reset_stats(struct cgroup_subsys_state *css,
463 struct cftype *cftype, u64 val)
464{
465 struct blkcg *blkcg = css_to_blkcg(css);
466 struct blkcg_gq *blkg;
467 int i;
468
469 mutex_lock(&blkcg_pol_mutex);
470 spin_lock_irq(&blkcg->lock);
471
472 /*
473 * Note that stat reset is racy - it doesn't synchronize against
474 * stat updates. This is a debug feature which shouldn't exist
475 * anyway. If you get hit by a race, retry.
476 */
477 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
478 blkg_rwstat_reset(&blkg->stat_bytes);
479 blkg_rwstat_reset(&blkg->stat_ios);
480
481 for (i = 0; i < BLKCG_MAX_POLS; i++) {
482 struct blkcg_policy *pol = blkcg_policy[i];
483
484 if (blkg->pd[i] && pol->pd_reset_stats_fn)
485 pol->pd_reset_stats_fn(blkg->pd[i]);
486 }
487 }
488
489 spin_unlock_irq(&blkcg->lock);
490 mutex_unlock(&blkcg_pol_mutex);
491 return 0;
492}
493
494const char *blkg_dev_name(struct blkcg_gq *blkg)
495{
496 /* some drivers (floppy) instantiate a queue w/o disk registered */
497 if (blkg->q->backing_dev_info->dev)
498 return dev_name(blkg->q->backing_dev_info->dev);
499 return NULL;
500}
501
502/**
503 * blkcg_print_blkgs - helper for printing per-blkg data
504 * @sf: seq_file to print to
505 * @blkcg: blkcg of interest
506 * @prfill: fill function to print out a blkg
507 * @pol: policy in question
508 * @data: data to be passed to @prfill
509 * @show_total: to print out sum of prfill return values or not
510 *
511 * This function invokes @prfill on each blkg of @blkcg if pd for the
512 * policy specified by @pol exists. @prfill is invoked with @sf, the
513 * policy data and @data and the matching queue lock held. If @show_total
514 * is %true, the sum of the return values from @prfill is printed with
515 * "Total" label at the end.
516 *
517 * This is to be used to construct print functions for
518 * cftype->read_seq_string method.
519 */
520void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
521 u64 (*prfill)(struct seq_file *,
522 struct blkg_policy_data *, int),
523 const struct blkcg_policy *pol, int data,
524 bool show_total)
525{
526 struct blkcg_gq *blkg;
527 u64 total = 0;
528
529 rcu_read_lock();
530 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
531 spin_lock_irq(&blkg->q->queue_lock);
532 if (blkcg_policy_enabled(blkg->q, pol))
533 total += prfill(sf, blkg->pd[pol->plid], data);
534 spin_unlock_irq(&blkg->q->queue_lock);
535 }
536 rcu_read_unlock();
537
538 if (show_total)
539 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
540}
541EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
542
543/**
544 * __blkg_prfill_u64 - prfill helper for a single u64 value
545 * @sf: seq_file to print to
546 * @pd: policy private data of interest
547 * @v: value to print
548 *
549 * Print @v to @sf for the device assocaited with @pd.
550 */
551u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
552{
553 const char *dname = blkg_dev_name(pd->blkg);
554
555 if (!dname)
556 return 0;
557
558 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
559 return v;
560}
561EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
562
563/**
564 * __blkg_prfill_rwstat - prfill helper for a blkg_rwstat
565 * @sf: seq_file to print to
566 * @pd: policy private data of interest
567 * @rwstat: rwstat to print
568 *
569 * Print @rwstat to @sf for the device assocaited with @pd.
570 */
571u64 __blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
572 const struct blkg_rwstat_sample *rwstat)
573{
574 static const char *rwstr[] = {
575 [BLKG_RWSTAT_READ] = "Read",
576 [BLKG_RWSTAT_WRITE] = "Write",
577 [BLKG_RWSTAT_SYNC] = "Sync",
578 [BLKG_RWSTAT_ASYNC] = "Async",
579 [BLKG_RWSTAT_DISCARD] = "Discard",
580 };
581 const char *dname = blkg_dev_name(pd->blkg);
582 u64 v;
583 int i;
584
585 if (!dname)
586 return 0;
587
588 for (i = 0; i < BLKG_RWSTAT_NR; i++)
589 seq_printf(sf, "%s %s %llu\n", dname, rwstr[i],
590 rwstat->cnt[i]);
591
592 v = rwstat->cnt[BLKG_RWSTAT_READ] +
593 rwstat->cnt[BLKG_RWSTAT_WRITE] +
594 rwstat->cnt[BLKG_RWSTAT_DISCARD];
595 seq_printf(sf, "%s Total %llu\n", dname, v);
596 return v;
597}
598EXPORT_SYMBOL_GPL(__blkg_prfill_rwstat);
599
600/**
601 * blkg_prfill_rwstat - prfill callback for blkg_rwstat
602 * @sf: seq_file to print to
603 * @pd: policy private data of interest
604 * @off: offset to the blkg_rwstat in @pd
605 *
606 * prfill callback for printing a blkg_rwstat.
607 */
608u64 blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
609 int off)
610{
611 struct blkg_rwstat_sample rwstat = { };
612
613 blkg_rwstat_read((void *)pd + off, &rwstat);
614 return __blkg_prfill_rwstat(sf, pd, &rwstat);
615}
616EXPORT_SYMBOL_GPL(blkg_prfill_rwstat);
617
618static u64 blkg_prfill_rwstat_field(struct seq_file *sf,
619 struct blkg_policy_data *pd, int off)
620{
621 struct blkg_rwstat_sample rwstat = { };
622
623 blkg_rwstat_read((void *)pd->blkg + off, &rwstat);
624 return __blkg_prfill_rwstat(sf, pd, &rwstat);
625}
626
627/**
628 * blkg_print_stat_bytes - seq_show callback for blkg->stat_bytes
629 * @sf: seq_file to print to
630 * @v: unused
631 *
632 * To be used as cftype->seq_show to print blkg->stat_bytes.
633 * cftype->private must be set to the blkcg_policy.
634 */
635int blkg_print_stat_bytes(struct seq_file *sf, void *v)
636{
637 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
638 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
639 offsetof(struct blkcg_gq, stat_bytes), true);
640 return 0;
641}
642EXPORT_SYMBOL_GPL(blkg_print_stat_bytes);
643
644/**
645 * blkg_print_stat_bytes - seq_show callback for blkg->stat_ios
646 * @sf: seq_file to print to
647 * @v: unused
648 *
649 * To be used as cftype->seq_show to print blkg->stat_ios. cftype->private
650 * must be set to the blkcg_policy.
651 */
652int blkg_print_stat_ios(struct seq_file *sf, void *v)
653{
654 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
655 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
656 offsetof(struct blkcg_gq, stat_ios), true);
657 return 0;
658}
659EXPORT_SYMBOL_GPL(blkg_print_stat_ios);
660
661static u64 blkg_prfill_rwstat_field_recursive(struct seq_file *sf,
662 struct blkg_policy_data *pd,
663 int off)
664{
665 struct blkg_rwstat_sample rwstat;
666
667 blkg_rwstat_recursive_sum(pd->blkg, NULL, off, &rwstat);
668 return __blkg_prfill_rwstat(sf, pd, &rwstat);
669}
670
671/**
672 * blkg_print_stat_bytes_recursive - recursive version of blkg_print_stat_bytes
673 * @sf: seq_file to print to
674 * @v: unused
675 */
676int blkg_print_stat_bytes_recursive(struct seq_file *sf, void *v)
677{
678 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
679 blkg_prfill_rwstat_field_recursive,
680 (void *)seq_cft(sf)->private,
681 offsetof(struct blkcg_gq, stat_bytes), true);
682 return 0;
683}
684EXPORT_SYMBOL_GPL(blkg_print_stat_bytes_recursive);
685
686/**
687 * blkg_print_stat_ios_recursive - recursive version of blkg_print_stat_ios
688 * @sf: seq_file to print to
689 * @v: unused
690 */
691int blkg_print_stat_ios_recursive(struct seq_file *sf, void *v)
692{
693 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
694 blkg_prfill_rwstat_field_recursive,
695 (void *)seq_cft(sf)->private,
696 offsetof(struct blkcg_gq, stat_ios), true);
697 return 0;
698}
699EXPORT_SYMBOL_GPL(blkg_print_stat_ios_recursive);
700
701/**
702 * blkg_rwstat_recursive_sum - collect hierarchical blkg_rwstat
703 * @blkg: blkg of interest
704 * @pol: blkcg_policy which contains the blkg_rwstat
705 * @off: offset to the blkg_rwstat in blkg_policy_data or @blkg
706 * @sum: blkg_rwstat_sample structure containing the results
707 *
708 * Collect the blkg_rwstat specified by @blkg, @pol and @off and all its
709 * online descendants and their aux counts. The caller must be holding the
710 * queue lock for online tests.
711 *
712 * If @pol is NULL, blkg_rwstat is at @off bytes into @blkg; otherwise, it
713 * is at @off bytes into @blkg's blkg_policy_data of the policy.
714 */
715void blkg_rwstat_recursive_sum(struct blkcg_gq *blkg, struct blkcg_policy *pol,
716 int off, struct blkg_rwstat_sample *sum)
717{
718 struct blkcg_gq *pos_blkg;
719 struct cgroup_subsys_state *pos_css;
720 unsigned int i;
721
722 lockdep_assert_held(&blkg->q->queue_lock);
723
724 rcu_read_lock();
725 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
726 struct blkg_rwstat *rwstat;
727
728 if (!pos_blkg->online)
729 continue;
730
731 if (pol)
732 rwstat = (void *)blkg_to_pd(pos_blkg, pol) + off;
733 else
734 rwstat = (void *)pos_blkg + off;
735
736 for (i = 0; i < BLKG_RWSTAT_NR; i++)
737 sum->cnt[i] = blkg_rwstat_read_counter(rwstat, i);
738 }
739 rcu_read_unlock();
740}
741EXPORT_SYMBOL_GPL(blkg_rwstat_recursive_sum);
742
743/* Performs queue bypass and policy enabled checks then looks up blkg. */
744static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg,
745 const struct blkcg_policy *pol,
746 struct request_queue *q)
747{
748 WARN_ON_ONCE(!rcu_read_lock_held());
749 lockdep_assert_held(&q->queue_lock);
750
751 if (!blkcg_policy_enabled(q, pol))
752 return ERR_PTR(-EOPNOTSUPP);
753 return __blkg_lookup(blkcg, q, true /* update_hint */);
754}
755
756/**
757 * blkg_conf_prep - parse and prepare for per-blkg config update
758 * @inputp: input string pointer
759 *
760 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update
761 * from @input and get and return the matching gendisk. *@inputp is
762 * updated to point past the device node prefix. Returns an ERR_PTR()
763 * value on error.
764 *
765 * Use this function iff blkg_conf_prep() can't be used for some reason.
766 */
767struct gendisk *blkcg_conf_get_disk(char **inputp)
768{
769 char *input = *inputp;
770 unsigned int major, minor;
771 struct gendisk *disk;
772 int key_len, part;
773
774 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
775 return ERR_PTR(-EINVAL);
776
777 input += key_len;
778 if (!isspace(*input))
779 return ERR_PTR(-EINVAL);
780 input = skip_spaces(input);
781
782 disk = get_gendisk(MKDEV(major, minor), &part);
783 if (!disk)
784 return ERR_PTR(-ENODEV);
785 if (part) {
786 put_disk_and_module(disk);
787 return ERR_PTR(-ENODEV);
788 }
789
790 *inputp = input;
791 return disk;
792}
793
794/**
795 * blkg_conf_prep - parse and prepare for per-blkg config update
796 * @blkcg: target block cgroup
797 * @pol: target policy
798 * @input: input string
799 * @ctx: blkg_conf_ctx to be filled
800 *
801 * Parse per-blkg config update from @input and initialize @ctx with the
802 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
803 * part of @input following MAJ:MIN. This function returns with RCU read
804 * lock and queue lock held and must be paired with blkg_conf_finish().
805 */
806int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
807 char *input, struct blkg_conf_ctx *ctx)
808 __acquires(rcu) __acquires(&disk->queue->queue_lock)
809{
810 struct gendisk *disk;
811 struct request_queue *q;
812 struct blkcg_gq *blkg;
813 int ret;
814
815 disk = blkcg_conf_get_disk(&input);
816 if (IS_ERR(disk))
817 return PTR_ERR(disk);
818
819 q = disk->queue;
820
821 rcu_read_lock();
822 spin_lock_irq(&q->queue_lock);
823
824 blkg = blkg_lookup_check(blkcg, pol, q);
825 if (IS_ERR(blkg)) {
826 ret = PTR_ERR(blkg);
827 goto fail_unlock;
828 }
829
830 if (blkg)
831 goto success;
832
833 /*
834 * Create blkgs walking down from blkcg_root to @blkcg, so that all
835 * non-root blkgs have access to their parents.
836 */
837 while (true) {
838 struct blkcg *pos = blkcg;
839 struct blkcg *parent;
840 struct blkcg_gq *new_blkg;
841
842 parent = blkcg_parent(blkcg);
843 while (parent && !__blkg_lookup(parent, q, false)) {
844 pos = parent;
845 parent = blkcg_parent(parent);
846 }
847
848 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
849 spin_unlock_irq(&q->queue_lock);
850 rcu_read_unlock();
851
852 new_blkg = blkg_alloc(pos, q, GFP_KERNEL);
853 if (unlikely(!new_blkg)) {
854 ret = -ENOMEM;
855 goto fail;
856 }
857
858 rcu_read_lock();
859 spin_lock_irq(&q->queue_lock);
860
861 blkg = blkg_lookup_check(pos, pol, q);
862 if (IS_ERR(blkg)) {
863 ret = PTR_ERR(blkg);
864 goto fail_unlock;
865 }
866
867 if (blkg) {
868 blkg_free(new_blkg);
869 } else {
870 blkg = blkg_create(pos, q, new_blkg);
871 if (IS_ERR(blkg)) {
872 ret = PTR_ERR(blkg);
873 goto fail_unlock;
874 }
875 }
876
877 if (pos == blkcg)
878 goto success;
879 }
880success:
881 ctx->disk = disk;
882 ctx->blkg = blkg;
883 ctx->body = input;
884 return 0;
885
886fail_unlock:
887 spin_unlock_irq(&q->queue_lock);
888 rcu_read_unlock();
889fail:
890 put_disk_and_module(disk);
891 /*
892 * If queue was bypassing, we should retry. Do so after a
893 * short msleep(). It isn't strictly necessary but queue
894 * can be bypassing for some time and it's always nice to
895 * avoid busy looping.
896 */
897 if (ret == -EBUSY) {
898 msleep(10);
899 ret = restart_syscall();
900 }
901 return ret;
902}
903EXPORT_SYMBOL_GPL(blkg_conf_prep);
904
905/**
906 * blkg_conf_finish - finish up per-blkg config update
907 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
908 *
909 * Finish up after per-blkg config update. This function must be paired
910 * with blkg_conf_prep().
911 */
912void blkg_conf_finish(struct blkg_conf_ctx *ctx)
913 __releases(&ctx->disk->queue->queue_lock) __releases(rcu)
914{
915 spin_unlock_irq(&ctx->disk->queue->queue_lock);
916 rcu_read_unlock();
917 put_disk_and_module(ctx->disk);
918}
919EXPORT_SYMBOL_GPL(blkg_conf_finish);
920
921static int blkcg_print_stat(struct seq_file *sf, void *v)
922{
923 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
924 struct blkcg_gq *blkg;
925
926 rcu_read_lock();
927
928 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
929 const char *dname;
930 char *buf;
931 struct blkg_rwstat_sample rwstat;
932 u64 rbytes, wbytes, rios, wios, dbytes, dios;
933 size_t size = seq_get_buf(sf, &buf), off = 0;
934 int i;
935 bool has_stats = false;
936
937 spin_lock_irq(&blkg->q->queue_lock);
938
939 if (!blkg->online)
940 goto skip;
941
942 dname = blkg_dev_name(blkg);
943 if (!dname)
944 goto skip;
945
946 /*
947 * Hooray string manipulation, count is the size written NOT
948 * INCLUDING THE \0, so size is now count+1 less than what we
949 * had before, but we want to start writing the next bit from
950 * the \0 so we only add count to buf.
951 */
952 off += scnprintf(buf+off, size-off, "%s ", dname);
953
954 blkg_rwstat_recursive_sum(blkg, NULL,
955 offsetof(struct blkcg_gq, stat_bytes), &rwstat);
956 rbytes = rwstat.cnt[BLKG_RWSTAT_READ];
957 wbytes = rwstat.cnt[BLKG_RWSTAT_WRITE];
958 dbytes = rwstat.cnt[BLKG_RWSTAT_DISCARD];
959
960 blkg_rwstat_recursive_sum(blkg, NULL,
961 offsetof(struct blkcg_gq, stat_ios), &rwstat);
962 rios = rwstat.cnt[BLKG_RWSTAT_READ];
963 wios = rwstat.cnt[BLKG_RWSTAT_WRITE];
964 dios = rwstat.cnt[BLKG_RWSTAT_DISCARD];
965
966 if (rbytes || wbytes || rios || wios) {
967 has_stats = true;
968 off += scnprintf(buf+off, size-off,
969 "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
970 rbytes, wbytes, rios, wios,
971 dbytes, dios);
972 }
973
974 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
975 has_stats = true;
976 off += scnprintf(buf+off, size-off,
977 " use_delay=%d delay_nsec=%llu",
978 atomic_read(&blkg->use_delay),
979 (unsigned long long)atomic64_read(&blkg->delay_nsec));
980 }
981
982 for (i = 0; i < BLKCG_MAX_POLS; i++) {
983 struct blkcg_policy *pol = blkcg_policy[i];
984 size_t written;
985
986 if (!blkg->pd[i] || !pol->pd_stat_fn)
987 continue;
988
989 written = pol->pd_stat_fn(blkg->pd[i], buf+off, size-off);
990 if (written)
991 has_stats = true;
992 off += written;
993 }
994
995 if (has_stats) {
996 if (off < size - 1) {
997 off += scnprintf(buf+off, size-off, "\n");
998 seq_commit(sf, off);
999 } else {
1000 seq_commit(sf, -1);
1001 }
1002 }
1003 skip:
1004 spin_unlock_irq(&blkg->q->queue_lock);
1005 }
1006
1007 rcu_read_unlock();
1008 return 0;
1009}
1010
1011static struct cftype blkcg_files[] = {
1012 {
1013 .name = "stat",
1014 .flags = CFTYPE_NOT_ON_ROOT,
1015 .seq_show = blkcg_print_stat,
1016 },
1017 { } /* terminate */
1018};
1019
1020static struct cftype blkcg_legacy_files[] = {
1021 {
1022 .name = "reset_stats",
1023 .write_u64 = blkcg_reset_stats,
1024 },
1025 { } /* terminate */
1026};
1027
1028/*
1029 * blkcg destruction is a three-stage process.
1030 *
1031 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
1032 * which offlines writeback. Here we tie the next stage of blkg destruction
1033 * to the completion of writeback associated with the blkcg. This lets us
1034 * avoid punting potentially large amounts of outstanding writeback to root
1035 * while maintaining any ongoing policies. The next stage is triggered when
1036 * the nr_cgwbs count goes to zero.
1037 *
1038 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
1039 * and handles the destruction of blkgs. Here the css reference held by
1040 * the blkg is put back eventually allowing blkcg_css_free() to be called.
1041 * This work may occur in cgwb_release_workfn() on the cgwb_release
1042 * workqueue. Any submitted ios that fail to get the blkg ref will be
1043 * punted to the root_blkg.
1044 *
1045 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1046 * This finally frees the blkcg.
1047 */
1048
1049/**
1050 * blkcg_css_offline - cgroup css_offline callback
1051 * @css: css of interest
1052 *
1053 * This function is called when @css is about to go away. Here the cgwbs are
1054 * offlined first and only once writeback associated with the blkcg has
1055 * finished do we start step 2 (see above).
1056 */
1057static void blkcg_css_offline(struct cgroup_subsys_state *css)
1058{
1059 struct blkcg *blkcg = css_to_blkcg(css);
1060
1061 /* this prevents anyone from attaching or migrating to this blkcg */
1062 wb_blkcg_offline(blkcg);
1063
1064 /* put the base cgwb reference allowing step 2 to be triggered */
1065 blkcg_cgwb_put(blkcg);
1066}
1067
1068/**
1069 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1070 * @blkcg: blkcg of interest
1071 *
1072 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1073 * is nested inside q lock, this function performs reverse double lock dancing.
1074 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1075 * blkcg_css_free to eventually be called.
1076 *
1077 * This is the blkcg counterpart of ioc_release_fn().
1078 */
1079void blkcg_destroy_blkgs(struct blkcg *blkcg)
1080{
1081 spin_lock_irq(&blkcg->lock);
1082
1083 while (!hlist_empty(&blkcg->blkg_list)) {
1084 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1085 struct blkcg_gq, blkcg_node);
1086 struct request_queue *q = blkg->q;
1087
1088 if (spin_trylock(&q->queue_lock)) {
1089 blkg_destroy(blkg);
1090 spin_unlock(&q->queue_lock);
1091 } else {
1092 spin_unlock_irq(&blkcg->lock);
1093 cpu_relax();
1094 spin_lock_irq(&blkcg->lock);
1095 }
1096 }
1097
1098 spin_unlock_irq(&blkcg->lock);
1099}
1100
1101static void blkcg_css_free(struct cgroup_subsys_state *css)
1102{
1103 struct blkcg *blkcg = css_to_blkcg(css);
1104 int i;
1105
1106 mutex_lock(&blkcg_pol_mutex);
1107
1108 list_del(&blkcg->all_blkcgs_node);
1109
1110 for (i = 0; i < BLKCG_MAX_POLS; i++)
1111 if (blkcg->cpd[i])
1112 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1113
1114 mutex_unlock(&blkcg_pol_mutex);
1115
1116 kfree(blkcg);
1117}
1118
1119static struct cgroup_subsys_state *
1120blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1121{
1122 struct blkcg *blkcg;
1123 struct cgroup_subsys_state *ret;
1124 int i;
1125
1126 mutex_lock(&blkcg_pol_mutex);
1127
1128 if (!parent_css) {
1129 blkcg = &blkcg_root;
1130 } else {
1131 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1132 if (!blkcg) {
1133 ret = ERR_PTR(-ENOMEM);
1134 goto unlock;
1135 }
1136 }
1137
1138 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1139 struct blkcg_policy *pol = blkcg_policy[i];
1140 struct blkcg_policy_data *cpd;
1141
1142 /*
1143 * If the policy hasn't been attached yet, wait for it
1144 * to be attached before doing anything else. Otherwise,
1145 * check if the policy requires any specific per-cgroup
1146 * data: if it does, allocate and initialize it.
1147 */
1148 if (!pol || !pol->cpd_alloc_fn)
1149 continue;
1150
1151 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1152 if (!cpd) {
1153 ret = ERR_PTR(-ENOMEM);
1154 goto free_pd_blkcg;
1155 }
1156 blkcg->cpd[i] = cpd;
1157 cpd->blkcg = blkcg;
1158 cpd->plid = i;
1159 if (pol->cpd_init_fn)
1160 pol->cpd_init_fn(cpd);
1161 }
1162
1163 spin_lock_init(&blkcg->lock);
1164 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1165 INIT_HLIST_HEAD(&blkcg->blkg_list);
1166#ifdef CONFIG_CGROUP_WRITEBACK
1167 INIT_LIST_HEAD(&blkcg->cgwb_list);
1168 refcount_set(&blkcg->cgwb_refcnt, 1);
1169#endif
1170 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1171
1172 mutex_unlock(&blkcg_pol_mutex);
1173 return &blkcg->css;
1174
1175free_pd_blkcg:
1176 for (i--; i >= 0; i--)
1177 if (blkcg->cpd[i])
1178 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1179
1180 if (blkcg != &blkcg_root)
1181 kfree(blkcg);
1182unlock:
1183 mutex_unlock(&blkcg_pol_mutex);
1184 return ret;
1185}
1186
1187/**
1188 * blkcg_init_queue - initialize blkcg part of request queue
1189 * @q: request_queue to initialize
1190 *
1191 * Called from blk_alloc_queue_node(). Responsible for initializing blkcg
1192 * part of new request_queue @q.
1193 *
1194 * RETURNS:
1195 * 0 on success, -errno on failure.
1196 */
1197int blkcg_init_queue(struct request_queue *q)
1198{
1199 struct blkcg_gq *new_blkg, *blkg;
1200 bool preloaded;
1201 int ret;
1202
1203 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1204 if (!new_blkg)
1205 return -ENOMEM;
1206
1207 preloaded = !radix_tree_preload(GFP_KERNEL);
1208
1209 /* Make sure the root blkg exists. */
1210 rcu_read_lock();
1211 spin_lock_irq(&q->queue_lock);
1212 blkg = blkg_create(&blkcg_root, q, new_blkg);
1213 if (IS_ERR(blkg))
1214 goto err_unlock;
1215 q->root_blkg = blkg;
1216 spin_unlock_irq(&q->queue_lock);
1217 rcu_read_unlock();
1218
1219 if (preloaded)
1220 radix_tree_preload_end();
1221
1222 ret = blk_iolatency_init(q);
1223 if (ret)
1224 goto err_destroy_all;
1225
1226 ret = blk_throtl_init(q);
1227 if (ret)
1228 goto err_destroy_all;
1229 return 0;
1230
1231err_destroy_all:
1232 blkg_destroy_all(q);
1233 return ret;
1234err_unlock:
1235 spin_unlock_irq(&q->queue_lock);
1236 rcu_read_unlock();
1237 if (preloaded)
1238 radix_tree_preload_end();
1239 return PTR_ERR(blkg);
1240}
1241
1242/**
1243 * blkcg_drain_queue - drain blkcg part of request_queue
1244 * @q: request_queue to drain
1245 *
1246 * Called from blk_drain_queue(). Responsible for draining blkcg part.
1247 */
1248void blkcg_drain_queue(struct request_queue *q)
1249{
1250 lockdep_assert_held(&q->queue_lock);
1251
1252 /*
1253 * @q could be exiting and already have destroyed all blkgs as
1254 * indicated by NULL root_blkg. If so, don't confuse policies.
1255 */
1256 if (!q->root_blkg)
1257 return;
1258
1259 blk_throtl_drain(q);
1260}
1261
1262/**
1263 * blkcg_exit_queue - exit and release blkcg part of request_queue
1264 * @q: request_queue being released
1265 *
1266 * Called from blk_exit_queue(). Responsible for exiting blkcg part.
1267 */
1268void blkcg_exit_queue(struct request_queue *q)
1269{
1270 blkg_destroy_all(q);
1271 blk_throtl_exit(q);
1272}
1273
1274/*
1275 * We cannot support shared io contexts, as we have no mean to support
1276 * two tasks with the same ioc in two different groups without major rework
1277 * of the main cic data structures. For now we allow a task to change
1278 * its cgroup only if it's the only owner of its ioc.
1279 */
1280static int blkcg_can_attach(struct cgroup_taskset *tset)
1281{
1282 struct task_struct *task;
1283 struct cgroup_subsys_state *dst_css;
1284 struct io_context *ioc;
1285 int ret = 0;
1286
1287 /* task_lock() is needed to avoid races with exit_io_context() */
1288 cgroup_taskset_for_each(task, dst_css, tset) {
1289 task_lock(task);
1290 ioc = task->io_context;
1291 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
1292 ret = -EINVAL;
1293 task_unlock(task);
1294 if (ret)
1295 break;
1296 }
1297 return ret;
1298}
1299
1300static void blkcg_bind(struct cgroup_subsys_state *root_css)
1301{
1302 int i;
1303
1304 mutex_lock(&blkcg_pol_mutex);
1305
1306 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1307 struct blkcg_policy *pol = blkcg_policy[i];
1308 struct blkcg *blkcg;
1309
1310 if (!pol || !pol->cpd_bind_fn)
1311 continue;
1312
1313 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1314 if (blkcg->cpd[pol->plid])
1315 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1316 }
1317 mutex_unlock(&blkcg_pol_mutex);
1318}
1319
1320static void blkcg_exit(struct task_struct *tsk)
1321{
1322 if (tsk->throttle_queue)
1323 blk_put_queue(tsk->throttle_queue);
1324 tsk->throttle_queue = NULL;
1325}
1326
1327struct cgroup_subsys io_cgrp_subsys = {
1328 .css_alloc = blkcg_css_alloc,
1329 .css_offline = blkcg_css_offline,
1330 .css_free = blkcg_css_free,
1331 .can_attach = blkcg_can_attach,
1332 .bind = blkcg_bind,
1333 .dfl_cftypes = blkcg_files,
1334 .legacy_cftypes = blkcg_legacy_files,
1335 .legacy_name = "blkio",
1336 .exit = blkcg_exit,
1337#ifdef CONFIG_MEMCG
1338 /*
1339 * This ensures that, if available, memcg is automatically enabled
1340 * together on the default hierarchy so that the owner cgroup can
1341 * be retrieved from writeback pages.
1342 */
1343 .depends_on = 1 << memory_cgrp_id,
1344#endif
1345};
1346EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1347
1348/**
1349 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1350 * @q: request_queue of interest
1351 * @pol: blkcg policy to activate
1352 *
1353 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1354 * bypass mode to populate its blkgs with policy_data for @pol.
1355 *
1356 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1357 * from IO path. Update of each blkg is protected by both queue and blkcg
1358 * locks so that holding either lock and testing blkcg_policy_enabled() is
1359 * always enough for dereferencing policy data.
1360 *
1361 * The caller is responsible for synchronizing [de]activations and policy
1362 * [un]registerations. Returns 0 on success, -errno on failure.
1363 */
1364int blkcg_activate_policy(struct request_queue *q,
1365 const struct blkcg_policy *pol)
1366{
1367 struct blkg_policy_data *pd_prealloc = NULL;
1368 struct blkcg_gq *blkg, *pinned_blkg = NULL;
1369 int ret;
1370
1371 if (blkcg_policy_enabled(q, pol))
1372 return 0;
1373
1374 if (queue_is_mq(q))
1375 blk_mq_freeze_queue(q);
1376retry:
1377 spin_lock_irq(&q->queue_lock);
1378
1379 /* blkg_list is pushed at the head, reverse walk to allocate parents first */
1380 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1381 struct blkg_policy_data *pd;
1382
1383 if (blkg->pd[pol->plid])
1384 continue;
1385
1386 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1387 if (blkg == pinned_blkg) {
1388 pd = pd_prealloc;
1389 pd_prealloc = NULL;
1390 } else {
1391 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q,
1392 blkg->blkcg);
1393 }
1394
1395 if (!pd) {
1396 /*
1397 * GFP_NOWAIT failed. Free the existing one and
1398 * prealloc for @blkg w/ GFP_KERNEL.
1399 */
1400 if (pinned_blkg)
1401 blkg_put(pinned_blkg);
1402 blkg_get(blkg);
1403 pinned_blkg = blkg;
1404
1405 spin_unlock_irq(&q->queue_lock);
1406
1407 if (pd_prealloc)
1408 pol->pd_free_fn(pd_prealloc);
1409 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q,
1410 blkg->blkcg);
1411 if (pd_prealloc)
1412 goto retry;
1413 else
1414 goto enomem;
1415 }
1416
1417 blkg->pd[pol->plid] = pd;
1418 pd->blkg = blkg;
1419 pd->plid = pol->plid;
1420 }
1421
1422 /* all allocated, init in the same order */
1423 if (pol->pd_init_fn)
1424 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1425 pol->pd_init_fn(blkg->pd[pol->plid]);
1426
1427 __set_bit(pol->plid, q->blkcg_pols);
1428 ret = 0;
1429
1430 spin_unlock_irq(&q->queue_lock);
1431out:
1432 if (queue_is_mq(q))
1433 blk_mq_unfreeze_queue(q);
1434 if (pinned_blkg)
1435 blkg_put(pinned_blkg);
1436 if (pd_prealloc)
1437 pol->pd_free_fn(pd_prealloc);
1438 return ret;
1439
1440enomem:
1441 /* alloc failed, nothing's initialized yet, free everything */
1442 spin_lock_irq(&q->queue_lock);
1443 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1444 if (blkg->pd[pol->plid]) {
1445 pol->pd_free_fn(blkg->pd[pol->plid]);
1446 blkg->pd[pol->plid] = NULL;
1447 }
1448 }
1449 spin_unlock_irq(&q->queue_lock);
1450 ret = -ENOMEM;
1451 goto out;
1452}
1453EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1454
1455/**
1456 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1457 * @q: request_queue of interest
1458 * @pol: blkcg policy to deactivate
1459 *
1460 * Deactivate @pol on @q. Follows the same synchronization rules as
1461 * blkcg_activate_policy().
1462 */
1463void blkcg_deactivate_policy(struct request_queue *q,
1464 const struct blkcg_policy *pol)
1465{
1466 struct blkcg_gq *blkg;
1467
1468 if (!blkcg_policy_enabled(q, pol))
1469 return;
1470
1471 if (queue_is_mq(q))
1472 blk_mq_freeze_queue(q);
1473
1474 spin_lock_irq(&q->queue_lock);
1475
1476 __clear_bit(pol->plid, q->blkcg_pols);
1477
1478 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1479 if (blkg->pd[pol->plid]) {
1480 if (pol->pd_offline_fn)
1481 pol->pd_offline_fn(blkg->pd[pol->plid]);
1482 pol->pd_free_fn(blkg->pd[pol->plid]);
1483 blkg->pd[pol->plid] = NULL;
1484 }
1485 }
1486
1487 spin_unlock_irq(&q->queue_lock);
1488
1489 if (queue_is_mq(q))
1490 blk_mq_unfreeze_queue(q);
1491}
1492EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1493
1494/**
1495 * blkcg_policy_register - register a blkcg policy
1496 * @pol: blkcg policy to register
1497 *
1498 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1499 * successful registration. Returns 0 on success and -errno on failure.
1500 */
1501int blkcg_policy_register(struct blkcg_policy *pol)
1502{
1503 struct blkcg *blkcg;
1504 int i, ret;
1505
1506 mutex_lock(&blkcg_pol_register_mutex);
1507 mutex_lock(&blkcg_pol_mutex);
1508
1509 /* find an empty slot */
1510 ret = -ENOSPC;
1511 for (i = 0; i < BLKCG_MAX_POLS; i++)
1512 if (!blkcg_policy[i])
1513 break;
1514 if (i >= BLKCG_MAX_POLS) {
1515 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1516 goto err_unlock;
1517 }
1518
1519 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1520 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1521 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1522 goto err_unlock;
1523
1524 /* register @pol */
1525 pol->plid = i;
1526 blkcg_policy[pol->plid] = pol;
1527
1528 /* allocate and install cpd's */
1529 if (pol->cpd_alloc_fn) {
1530 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1531 struct blkcg_policy_data *cpd;
1532
1533 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1534 if (!cpd)
1535 goto err_free_cpds;
1536
1537 blkcg->cpd[pol->plid] = cpd;
1538 cpd->blkcg = blkcg;
1539 cpd->plid = pol->plid;
1540 if (pol->cpd_init_fn)
1541 pol->cpd_init_fn(cpd);
1542 }
1543 }
1544
1545 mutex_unlock(&blkcg_pol_mutex);
1546
1547 /* everything is in place, add intf files for the new policy */
1548 if (pol->dfl_cftypes)
1549 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1550 pol->dfl_cftypes));
1551 if (pol->legacy_cftypes)
1552 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1553 pol->legacy_cftypes));
1554 mutex_unlock(&blkcg_pol_register_mutex);
1555 return 0;
1556
1557err_free_cpds:
1558 if (pol->cpd_free_fn) {
1559 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1560 if (blkcg->cpd[pol->plid]) {
1561 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1562 blkcg->cpd[pol->plid] = NULL;
1563 }
1564 }
1565 }
1566 blkcg_policy[pol->plid] = NULL;
1567err_unlock:
1568 mutex_unlock(&blkcg_pol_mutex);
1569 mutex_unlock(&blkcg_pol_register_mutex);
1570 return ret;
1571}
1572EXPORT_SYMBOL_GPL(blkcg_policy_register);
1573
1574/**
1575 * blkcg_policy_unregister - unregister a blkcg policy
1576 * @pol: blkcg policy to unregister
1577 *
1578 * Undo blkcg_policy_register(@pol). Might sleep.
1579 */
1580void blkcg_policy_unregister(struct blkcg_policy *pol)
1581{
1582 struct blkcg *blkcg;
1583
1584 mutex_lock(&blkcg_pol_register_mutex);
1585
1586 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1587 goto out_unlock;
1588
1589 /* kill the intf files first */
1590 if (pol->dfl_cftypes)
1591 cgroup_rm_cftypes(pol->dfl_cftypes);
1592 if (pol->legacy_cftypes)
1593 cgroup_rm_cftypes(pol->legacy_cftypes);
1594
1595 /* remove cpds and unregister */
1596 mutex_lock(&blkcg_pol_mutex);
1597
1598 if (pol->cpd_free_fn) {
1599 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1600 if (blkcg->cpd[pol->plid]) {
1601 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1602 blkcg->cpd[pol->plid] = NULL;
1603 }
1604 }
1605 }
1606 blkcg_policy[pol->plid] = NULL;
1607
1608 mutex_unlock(&blkcg_pol_mutex);
1609out_unlock:
1610 mutex_unlock(&blkcg_pol_register_mutex);
1611}
1612EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1613
1614bool __blkcg_punt_bio_submit(struct bio *bio)
1615{
1616 struct blkcg_gq *blkg = bio->bi_blkg;
1617
1618 /* consume the flag first */
1619 bio->bi_opf &= ~REQ_CGROUP_PUNT;
1620
1621 /* never bounce for the root cgroup */
1622 if (!blkg->parent)
1623 return false;
1624
1625 spin_lock_bh(&blkg->async_bio_lock);
1626 bio_list_add(&blkg->async_bios, bio);
1627 spin_unlock_bh(&blkg->async_bio_lock);
1628
1629 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
1630 return true;
1631}
1632
1633/*
1634 * Scale the accumulated delay based on how long it has been since we updated
1635 * the delay. We only call this when we are adding delay, in case it's been a
1636 * while since we added delay, and when we are checking to see if we need to
1637 * delay a task, to account for any delays that may have occurred.
1638 */
1639static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1640{
1641 u64 old = atomic64_read(&blkg->delay_start);
1642
1643 /*
1644 * We only want to scale down every second. The idea here is that we
1645 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1646 * time window. We only want to throttle tasks for recent delay that
1647 * has occurred, in 1 second time windows since that's the maximum
1648 * things can be throttled. We save the current delay window in
1649 * blkg->last_delay so we know what amount is still left to be charged
1650 * to the blkg from this point onward. blkg->last_use keeps track of
1651 * the use_delay counter. The idea is if we're unthrottling the blkg we
1652 * are ok with whatever is happening now, and we can take away more of
1653 * the accumulated delay as we've already throttled enough that
1654 * everybody is happy with their IO latencies.
1655 */
1656 if (time_before64(old + NSEC_PER_SEC, now) &&
1657 atomic64_cmpxchg(&blkg->delay_start, old, now) == old) {
1658 u64 cur = atomic64_read(&blkg->delay_nsec);
1659 u64 sub = min_t(u64, blkg->last_delay, now - old);
1660 int cur_use = atomic_read(&blkg->use_delay);
1661
1662 /*
1663 * We've been unthrottled, subtract a larger chunk of our
1664 * accumulated delay.
1665 */
1666 if (cur_use < blkg->last_use)
1667 sub = max_t(u64, sub, blkg->last_delay >> 1);
1668
1669 /*
1670 * This shouldn't happen, but handle it anyway. Our delay_nsec
1671 * should only ever be growing except here where we subtract out
1672 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1673 * rather not end up with negative numbers.
1674 */
1675 if (unlikely(cur < sub)) {
1676 atomic64_set(&blkg->delay_nsec, 0);
1677 blkg->last_delay = 0;
1678 } else {
1679 atomic64_sub(sub, &blkg->delay_nsec);
1680 blkg->last_delay = cur - sub;
1681 }
1682 blkg->last_use = cur_use;
1683 }
1684}
1685
1686/*
1687 * This is called when we want to actually walk up the hierarchy and check to
1688 * see if we need to throttle, and then actually throttle if there is some
1689 * accumulated delay. This should only be called upon return to user space so
1690 * we're not holding some lock that would induce a priority inversion.
1691 */
1692static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1693{
1694 unsigned long pflags;
1695 u64 now = ktime_to_ns(ktime_get());
1696 u64 exp;
1697 u64 delay_nsec = 0;
1698 int tok;
1699
1700 while (blkg->parent) {
1701 if (atomic_read(&blkg->use_delay)) {
1702 blkcg_scale_delay(blkg, now);
1703 delay_nsec = max_t(u64, delay_nsec,
1704 atomic64_read(&blkg->delay_nsec));
1705 }
1706 blkg = blkg->parent;
1707 }
1708
1709 if (!delay_nsec)
1710 return;
1711
1712 /*
1713 * Let's not sleep for all eternity if we've amassed a huge delay.
1714 * Swapping or metadata IO can accumulate 10's of seconds worth of
1715 * delay, and we want userspace to be able to do _something_ so cap the
1716 * delays at 1 second. If there's 10's of seconds worth of delay then
1717 * the tasks will be delayed for 1 second for every syscall.
1718 */
1719 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1720
1721 if (use_memdelay)
1722 psi_memstall_enter(&pflags);
1723
1724 exp = ktime_add_ns(now, delay_nsec);
1725 tok = io_schedule_prepare();
1726 do {
1727 __set_current_state(TASK_KILLABLE);
1728 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1729 break;
1730 } while (!fatal_signal_pending(current));
1731 io_schedule_finish(tok);
1732
1733 if (use_memdelay)
1734 psi_memstall_leave(&pflags);
1735}
1736
1737/**
1738 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1739 *
1740 * This is only called if we've been marked with set_notify_resume(). Obviously
1741 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1742 * check to see if current->throttle_queue is set and if not this doesn't do
1743 * anything. This should only ever be called by the resume code, it's not meant
1744 * to be called by people willy-nilly as it will actually do the work to
1745 * throttle the task if it is setup for throttling.
1746 */
1747void blkcg_maybe_throttle_current(void)
1748{
1749 struct request_queue *q = current->throttle_queue;
1750 struct cgroup_subsys_state *css;
1751 struct blkcg *blkcg;
1752 struct blkcg_gq *blkg;
1753 bool use_memdelay = current->use_memdelay;
1754
1755 if (!q)
1756 return;
1757
1758 current->throttle_queue = NULL;
1759 current->use_memdelay = false;
1760
1761 rcu_read_lock();
1762 css = kthread_blkcg();
1763 if (css)
1764 blkcg = css_to_blkcg(css);
1765 else
1766 blkcg = css_to_blkcg(task_css(current, io_cgrp_id));
1767
1768 if (!blkcg)
1769 goto out;
1770 blkg = blkg_lookup(blkcg, q);
1771 if (!blkg)
1772 goto out;
1773 if (!blkg_tryget(blkg))
1774 goto out;
1775 rcu_read_unlock();
1776
1777 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1778 blkg_put(blkg);
1779 blk_put_queue(q);
1780 return;
1781out:
1782 rcu_read_unlock();
1783 blk_put_queue(q);
1784}
1785
1786/**
1787 * blkcg_schedule_throttle - this task needs to check for throttling
1788 * @q: the request queue IO was submitted on
1789 * @use_memdelay: do we charge this to memory delay for PSI
1790 *
1791 * This is called by the IO controller when we know there's delay accumulated
1792 * for the blkg for this task. We do not pass the blkg because there are places
1793 * we call this that may not have that information, the swapping code for
1794 * instance will only have a request_queue at that point. This set's the
1795 * notify_resume for the task to check and see if it requires throttling before
1796 * returning to user space.
1797 *
1798 * We will only schedule once per syscall. You can call this over and over
1799 * again and it will only do the check once upon return to user space, and only
1800 * throttle once. If the task needs to be throttled again it'll need to be
1801 * re-set at the next time we see the task.
1802 */
1803void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay)
1804{
1805 if (unlikely(current->flags & PF_KTHREAD))
1806 return;
1807
1808 if (!blk_get_queue(q))
1809 return;
1810
1811 if (current->throttle_queue)
1812 blk_put_queue(current->throttle_queue);
1813 current->throttle_queue = q;
1814 if (use_memdelay)
1815 current->use_memdelay = use_memdelay;
1816 set_notify_resume(current);
1817}
1818
1819/**
1820 * blkcg_add_delay - add delay to this blkg
1821 * @blkg: blkg of interest
1822 * @now: the current time in nanoseconds
1823 * @delta: how many nanoseconds of delay to add
1824 *
1825 * Charge @delta to the blkg's current delay accumulation. This is used to
1826 * throttle tasks if an IO controller thinks we need more throttling.
1827 */
1828void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1829{
1830 blkcg_scale_delay(blkg, now);
1831 atomic64_add(delta, &blkg->delay_nsec);
1832}
1833
1834static int __init blkcg_init(void)
1835{
1836 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
1837 WQ_MEM_RECLAIM | WQ_FREEZABLE |
1838 WQ_UNBOUND | WQ_SYSFS, 0);
1839 if (!blkcg_punt_bio_wq)
1840 return -ENOMEM;
1841 return 0;
1842}
1843subsys_initcall(blkcg_init);
1844
1845module_param(blkcg_debug_stats, bool, 0644);
1846MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");