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/sched/signal.h>
21#include <linux/err.h>
22#include <linux/blkdev.h>
23#include <linux/backing-dev.h>
24#include <linux/slab.h>
25#include <linux/genhd.h>
26#include <linux/delay.h>
27#include <linux/atomic.h>
28#include <linux/ctype.h>
29#include <linux/blk-cgroup.h>
30#include "blk.h"
31
32#define MAX_KEY_LEN 100
33
34/*
35 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
36 * blkcg_pol_register_mutex nests outside of it and synchronizes entire
37 * policy [un]register operations including cgroup file additions /
38 * removals. Putting cgroup file registration outside blkcg_pol_mutex
39 * allows grabbing it from cgroup callbacks.
40 */
41static DEFINE_MUTEX(blkcg_pol_register_mutex);
42static DEFINE_MUTEX(blkcg_pol_mutex);
43
44struct blkcg blkcg_root;
45EXPORT_SYMBOL_GPL(blkcg_root);
46
47struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
48
49static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
50
51static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */
52
53static bool blkcg_policy_enabled(struct request_queue *q,
54 const struct blkcg_policy *pol)
55{
56 return pol && test_bit(pol->plid, q->blkcg_pols);
57}
58
59/**
60 * blkg_free - free a blkg
61 * @blkg: blkg to free
62 *
63 * Free @blkg which may be partially allocated.
64 */
65static void blkg_free(struct blkcg_gq *blkg)
66{
67 int i;
68
69 if (!blkg)
70 return;
71
72 for (i = 0; i < BLKCG_MAX_POLS; i++)
73 if (blkg->pd[i])
74 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
75
76 if (blkg->blkcg != &blkcg_root)
77 blk_exit_rl(blkg->q, &blkg->rl);
78
79 blkg_rwstat_exit(&blkg->stat_ios);
80 blkg_rwstat_exit(&blkg->stat_bytes);
81 kfree(blkg);
82}
83
84/**
85 * blkg_alloc - allocate a blkg
86 * @blkcg: block cgroup the new blkg is associated with
87 * @q: request_queue the new blkg is associated with
88 * @gfp_mask: allocation mask to use
89 *
90 * Allocate a new blkg assocating @blkcg and @q.
91 */
92static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
93 gfp_t gfp_mask)
94{
95 struct blkcg_gq *blkg;
96 int i;
97
98 /* alloc and init base part */
99 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
100 if (!blkg)
101 return NULL;
102
103 if (blkg_rwstat_init(&blkg->stat_bytes, gfp_mask) ||
104 blkg_rwstat_init(&blkg->stat_ios, gfp_mask))
105 goto err_free;
106
107 blkg->q = q;
108 INIT_LIST_HEAD(&blkg->q_node);
109 blkg->blkcg = blkcg;
110 atomic_set(&blkg->refcnt, 1);
111
112 /* root blkg uses @q->root_rl, init rl only for !root blkgs */
113 if (blkcg != &blkcg_root) {
114 if (blk_init_rl(&blkg->rl, q, gfp_mask))
115 goto err_free;
116 blkg->rl.blkg = blkg;
117 }
118
119 for (i = 0; i < BLKCG_MAX_POLS; i++) {
120 struct blkcg_policy *pol = blkcg_policy[i];
121 struct blkg_policy_data *pd;
122
123 if (!blkcg_policy_enabled(q, pol))
124 continue;
125
126 /* alloc per-policy data and attach it to blkg */
127 pd = pol->pd_alloc_fn(gfp_mask, q->node);
128 if (!pd)
129 goto err_free;
130
131 blkg->pd[i] = pd;
132 pd->blkg = blkg;
133 pd->plid = i;
134 }
135
136 return blkg;
137
138err_free:
139 blkg_free(blkg);
140 return NULL;
141}
142
143struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg,
144 struct request_queue *q, bool update_hint)
145{
146 struct blkcg_gq *blkg;
147
148 /*
149 * Hint didn't match. Look up from the radix tree. Note that the
150 * hint can only be updated under queue_lock as otherwise @blkg
151 * could have already been removed from blkg_tree. The caller is
152 * responsible for grabbing queue_lock if @update_hint.
153 */
154 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
155 if (blkg && blkg->q == q) {
156 if (update_hint) {
157 lockdep_assert_held(q->queue_lock);
158 rcu_assign_pointer(blkcg->blkg_hint, blkg);
159 }
160 return blkg;
161 }
162
163 return NULL;
164}
165EXPORT_SYMBOL_GPL(blkg_lookup_slowpath);
166
167/*
168 * If @new_blkg is %NULL, this function tries to allocate a new one as
169 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
170 */
171static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
172 struct request_queue *q,
173 struct blkcg_gq *new_blkg)
174{
175 struct blkcg_gq *blkg;
176 struct bdi_writeback_congested *wb_congested;
177 int i, ret;
178
179 WARN_ON_ONCE(!rcu_read_lock_held());
180 lockdep_assert_held(q->queue_lock);
181
182 /* blkg holds a reference to blkcg */
183 if (!css_tryget_online(&blkcg->css)) {
184 ret = -ENODEV;
185 goto err_free_blkg;
186 }
187
188 wb_congested = wb_congested_get_create(q->backing_dev_info,
189 blkcg->css.id,
190 GFP_NOWAIT | __GFP_NOWARN);
191 if (!wb_congested) {
192 ret = -ENOMEM;
193 goto err_put_css;
194 }
195
196 /* allocate */
197 if (!new_blkg) {
198 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN);
199 if (unlikely(!new_blkg)) {
200 ret = -ENOMEM;
201 goto err_put_congested;
202 }
203 }
204 blkg = new_blkg;
205 blkg->wb_congested = wb_congested;
206
207 /* link parent */
208 if (blkcg_parent(blkcg)) {
209 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
210 if (WARN_ON_ONCE(!blkg->parent)) {
211 ret = -ENODEV;
212 goto err_put_congested;
213 }
214 blkg_get(blkg->parent);
215 }
216
217 /* invoke per-policy init */
218 for (i = 0; i < BLKCG_MAX_POLS; i++) {
219 struct blkcg_policy *pol = blkcg_policy[i];
220
221 if (blkg->pd[i] && pol->pd_init_fn)
222 pol->pd_init_fn(blkg->pd[i]);
223 }
224
225 /* insert */
226 spin_lock(&blkcg->lock);
227 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
228 if (likely(!ret)) {
229 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
230 list_add(&blkg->q_node, &q->blkg_list);
231
232 for (i = 0; i < BLKCG_MAX_POLS; i++) {
233 struct blkcg_policy *pol = blkcg_policy[i];
234
235 if (blkg->pd[i] && pol->pd_online_fn)
236 pol->pd_online_fn(blkg->pd[i]);
237 }
238 }
239 blkg->online = true;
240 spin_unlock(&blkcg->lock);
241
242 if (!ret)
243 return blkg;
244
245 /* @blkg failed fully initialized, use the usual release path */
246 blkg_put(blkg);
247 return ERR_PTR(ret);
248
249err_put_congested:
250 wb_congested_put(wb_congested);
251err_put_css:
252 css_put(&blkcg->css);
253err_free_blkg:
254 blkg_free(new_blkg);
255 return ERR_PTR(ret);
256}
257
258/**
259 * blkg_lookup_create - lookup blkg, try to create one if not there
260 * @blkcg: blkcg of interest
261 * @q: request_queue of interest
262 *
263 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
264 * create one. blkg creation is performed recursively from blkcg_root such
265 * that all non-root blkg's have access to the parent blkg. This function
266 * should be called under RCU read lock and @q->queue_lock.
267 *
268 * Returns pointer to the looked up or created blkg on success, ERR_PTR()
269 * value on error. If @q is dead, returns ERR_PTR(-EINVAL). If @q is not
270 * dead and bypassing, returns ERR_PTR(-EBUSY).
271 */
272struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
273 struct request_queue *q)
274{
275 struct blkcg_gq *blkg;
276
277 WARN_ON_ONCE(!rcu_read_lock_held());
278 lockdep_assert_held(q->queue_lock);
279
280 /*
281 * This could be the first entry point of blkcg implementation and
282 * we shouldn't allow anything to go through for a bypassing queue.
283 */
284 if (unlikely(blk_queue_bypass(q)))
285 return ERR_PTR(blk_queue_dying(q) ? -ENODEV : -EBUSY);
286
287 blkg = __blkg_lookup(blkcg, q, true);
288 if (blkg)
289 return blkg;
290
291 /*
292 * Create blkgs walking down from blkcg_root to @blkcg, so that all
293 * non-root blkgs have access to their parents.
294 */
295 while (true) {
296 struct blkcg *pos = blkcg;
297 struct blkcg *parent = blkcg_parent(blkcg);
298
299 while (parent && !__blkg_lookup(parent, q, false)) {
300 pos = parent;
301 parent = blkcg_parent(parent);
302 }
303
304 blkg = blkg_create(pos, q, NULL);
305 if (pos == blkcg || IS_ERR(blkg))
306 return blkg;
307 }
308}
309
310static void blkg_pd_offline(struct blkcg_gq *blkg)
311{
312 int i;
313
314 lockdep_assert_held(blkg->q->queue_lock);
315 lockdep_assert_held(&blkg->blkcg->lock);
316
317 for (i = 0; i < BLKCG_MAX_POLS; i++) {
318 struct blkcg_policy *pol = blkcg_policy[i];
319
320 if (blkg->pd[i] && !blkg->pd[i]->offline &&
321 pol->pd_offline_fn) {
322 pol->pd_offline_fn(blkg->pd[i]);
323 blkg->pd[i]->offline = true;
324 }
325 }
326}
327
328static void blkg_destroy(struct blkcg_gq *blkg)
329{
330 struct blkcg *blkcg = blkg->blkcg;
331 struct blkcg_gq *parent = blkg->parent;
332
333 lockdep_assert_held(blkg->q->queue_lock);
334 lockdep_assert_held(&blkcg->lock);
335
336 /* Something wrong if we are trying to remove same group twice */
337 WARN_ON_ONCE(list_empty(&blkg->q_node));
338 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
339
340 if (parent) {
341 blkg_rwstat_add_aux(&parent->stat_bytes, &blkg->stat_bytes);
342 blkg_rwstat_add_aux(&parent->stat_ios, &blkg->stat_ios);
343 }
344
345 blkg->online = false;
346
347 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
348 list_del_init(&blkg->q_node);
349 hlist_del_init_rcu(&blkg->blkcg_node);
350
351 /*
352 * Both setting lookup hint to and clearing it from @blkg are done
353 * under queue_lock. If it's not pointing to @blkg now, it never
354 * will. Hint assignment itself can race safely.
355 */
356 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
357 rcu_assign_pointer(blkcg->blkg_hint, NULL);
358
359 /*
360 * Put the reference taken at the time of creation so that when all
361 * queues are gone, group can be destroyed.
362 */
363 blkg_put(blkg);
364}
365
366/**
367 * blkg_destroy_all - destroy all blkgs associated with a request_queue
368 * @q: request_queue of interest
369 *
370 * Destroy all blkgs associated with @q.
371 */
372static void blkg_destroy_all(struct request_queue *q)
373{
374 struct blkcg_gq *blkg, *n;
375
376 lockdep_assert_held(q->queue_lock);
377
378 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
379 struct blkcg *blkcg = blkg->blkcg;
380
381 spin_lock(&blkcg->lock);
382 blkg_pd_offline(blkg);
383 blkg_destroy(blkg);
384 spin_unlock(&blkcg->lock);
385 }
386
387 q->root_blkg = NULL;
388 q->root_rl.blkg = NULL;
389}
390
391/*
392 * A group is RCU protected, but having an rcu lock does not mean that one
393 * can access all the fields of blkg and assume these are valid. For
394 * example, don't try to follow throtl_data and request queue links.
395 *
396 * Having a reference to blkg under an rcu allows accesses to only values
397 * local to groups like group stats and group rate limits.
398 */
399void __blkg_release_rcu(struct rcu_head *rcu_head)
400{
401 struct blkcg_gq *blkg = container_of(rcu_head, struct blkcg_gq, rcu_head);
402
403 /* release the blkcg and parent blkg refs this blkg has been holding */
404 css_put(&blkg->blkcg->css);
405 if (blkg->parent)
406 blkg_put(blkg->parent);
407
408 wb_congested_put(blkg->wb_congested);
409
410 blkg_free(blkg);
411}
412EXPORT_SYMBOL_GPL(__blkg_release_rcu);
413
414/*
415 * The next function used by blk_queue_for_each_rl(). It's a bit tricky
416 * because the root blkg uses @q->root_rl instead of its own rl.
417 */
418struct request_list *__blk_queue_next_rl(struct request_list *rl,
419 struct request_queue *q)
420{
421 struct list_head *ent;
422 struct blkcg_gq *blkg;
423
424 /*
425 * Determine the current blkg list_head. The first entry is
426 * root_rl which is off @q->blkg_list and mapped to the head.
427 */
428 if (rl == &q->root_rl) {
429 ent = &q->blkg_list;
430 /* There are no more block groups, hence no request lists */
431 if (list_empty(ent))
432 return NULL;
433 } else {
434 blkg = container_of(rl, struct blkcg_gq, rl);
435 ent = &blkg->q_node;
436 }
437
438 /* walk to the next list_head, skip root blkcg */
439 ent = ent->next;
440 if (ent == &q->root_blkg->q_node)
441 ent = ent->next;
442 if (ent == &q->blkg_list)
443 return NULL;
444
445 blkg = container_of(ent, struct blkcg_gq, q_node);
446 return &blkg->rl;
447}
448
449static int blkcg_reset_stats(struct cgroup_subsys_state *css,
450 struct cftype *cftype, u64 val)
451{
452 struct blkcg *blkcg = css_to_blkcg(css);
453 struct blkcg_gq *blkg;
454 int i;
455
456 mutex_lock(&blkcg_pol_mutex);
457 spin_lock_irq(&blkcg->lock);
458
459 /*
460 * Note that stat reset is racy - it doesn't synchronize against
461 * stat updates. This is a debug feature which shouldn't exist
462 * anyway. If you get hit by a race, retry.
463 */
464 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
465 blkg_rwstat_reset(&blkg->stat_bytes);
466 blkg_rwstat_reset(&blkg->stat_ios);
467
468 for (i = 0; i < BLKCG_MAX_POLS; i++) {
469 struct blkcg_policy *pol = blkcg_policy[i];
470
471 if (blkg->pd[i] && pol->pd_reset_stats_fn)
472 pol->pd_reset_stats_fn(blkg->pd[i]);
473 }
474 }
475
476 spin_unlock_irq(&blkcg->lock);
477 mutex_unlock(&blkcg_pol_mutex);
478 return 0;
479}
480
481const char *blkg_dev_name(struct blkcg_gq *blkg)
482{
483 /* some drivers (floppy) instantiate a queue w/o disk registered */
484 if (blkg->q->backing_dev_info->dev)
485 return dev_name(blkg->q->backing_dev_info->dev);
486 return NULL;
487}
488EXPORT_SYMBOL_GPL(blkg_dev_name);
489
490/**
491 * blkcg_print_blkgs - helper for printing per-blkg data
492 * @sf: seq_file to print to
493 * @blkcg: blkcg of interest
494 * @prfill: fill function to print out a blkg
495 * @pol: policy in question
496 * @data: data to be passed to @prfill
497 * @show_total: to print out sum of prfill return values or not
498 *
499 * This function invokes @prfill on each blkg of @blkcg if pd for the
500 * policy specified by @pol exists. @prfill is invoked with @sf, the
501 * policy data and @data and the matching queue lock held. If @show_total
502 * is %true, the sum of the return values from @prfill is printed with
503 * "Total" label at the end.
504 *
505 * This is to be used to construct print functions for
506 * cftype->read_seq_string method.
507 */
508void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
509 u64 (*prfill)(struct seq_file *,
510 struct blkg_policy_data *, int),
511 const struct blkcg_policy *pol, int data,
512 bool show_total)
513{
514 struct blkcg_gq *blkg;
515 u64 total = 0;
516
517 rcu_read_lock();
518 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
519 spin_lock_irq(blkg->q->queue_lock);
520 if (blkcg_policy_enabled(blkg->q, pol))
521 total += prfill(sf, blkg->pd[pol->plid], data);
522 spin_unlock_irq(blkg->q->queue_lock);
523 }
524 rcu_read_unlock();
525
526 if (show_total)
527 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
528}
529EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
530
531/**
532 * __blkg_prfill_u64 - prfill helper for a single u64 value
533 * @sf: seq_file to print to
534 * @pd: policy private data of interest
535 * @v: value to print
536 *
537 * Print @v to @sf for the device assocaited with @pd.
538 */
539u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
540{
541 const char *dname = blkg_dev_name(pd->blkg);
542
543 if (!dname)
544 return 0;
545
546 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
547 return v;
548}
549EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
550
551/**
552 * __blkg_prfill_rwstat - prfill helper for a blkg_rwstat
553 * @sf: seq_file to print to
554 * @pd: policy private data of interest
555 * @rwstat: rwstat to print
556 *
557 * Print @rwstat to @sf for the device assocaited with @pd.
558 */
559u64 __blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
560 const struct blkg_rwstat *rwstat)
561{
562 static const char *rwstr[] = {
563 [BLKG_RWSTAT_READ] = "Read",
564 [BLKG_RWSTAT_WRITE] = "Write",
565 [BLKG_RWSTAT_SYNC] = "Sync",
566 [BLKG_RWSTAT_ASYNC] = "Async",
567 };
568 const char *dname = blkg_dev_name(pd->blkg);
569 u64 v;
570 int i;
571
572 if (!dname)
573 return 0;
574
575 for (i = 0; i < BLKG_RWSTAT_NR; i++)
576 seq_printf(sf, "%s %s %llu\n", dname, rwstr[i],
577 (unsigned long long)atomic64_read(&rwstat->aux_cnt[i]));
578
579 v = atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_READ]) +
580 atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_WRITE]);
581 seq_printf(sf, "%s Total %llu\n", dname, (unsigned long long)v);
582 return v;
583}
584EXPORT_SYMBOL_GPL(__blkg_prfill_rwstat);
585
586/**
587 * blkg_prfill_stat - prfill callback for blkg_stat
588 * @sf: seq_file to print to
589 * @pd: policy private data of interest
590 * @off: offset to the blkg_stat in @pd
591 *
592 * prfill callback for printing a blkg_stat.
593 */
594u64 blkg_prfill_stat(struct seq_file *sf, struct blkg_policy_data *pd, int off)
595{
596 return __blkg_prfill_u64(sf, pd, blkg_stat_read((void *)pd + off));
597}
598EXPORT_SYMBOL_GPL(blkg_prfill_stat);
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 rwstat = blkg_rwstat_read((void *)pd + off);
612
613 return __blkg_prfill_rwstat(sf, pd, &rwstat);
614}
615EXPORT_SYMBOL_GPL(blkg_prfill_rwstat);
616
617static u64 blkg_prfill_rwstat_field(struct seq_file *sf,
618 struct blkg_policy_data *pd, int off)
619{
620 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd->blkg + off);
621
622 return __blkg_prfill_rwstat(sf, pd, &rwstat);
623}
624
625/**
626 * blkg_print_stat_bytes - seq_show callback for blkg->stat_bytes
627 * @sf: seq_file to print to
628 * @v: unused
629 *
630 * To be used as cftype->seq_show to print blkg->stat_bytes.
631 * cftype->private must be set to the blkcg_policy.
632 */
633int blkg_print_stat_bytes(struct seq_file *sf, void *v)
634{
635 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
636 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
637 offsetof(struct blkcg_gq, stat_bytes), true);
638 return 0;
639}
640EXPORT_SYMBOL_GPL(blkg_print_stat_bytes);
641
642/**
643 * blkg_print_stat_bytes - seq_show callback for blkg->stat_ios
644 * @sf: seq_file to print to
645 * @v: unused
646 *
647 * To be used as cftype->seq_show to print blkg->stat_ios. cftype->private
648 * must be set to the blkcg_policy.
649 */
650int blkg_print_stat_ios(struct seq_file *sf, void *v)
651{
652 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
653 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
654 offsetof(struct blkcg_gq, stat_ios), true);
655 return 0;
656}
657EXPORT_SYMBOL_GPL(blkg_print_stat_ios);
658
659static u64 blkg_prfill_rwstat_field_recursive(struct seq_file *sf,
660 struct blkg_policy_data *pd,
661 int off)
662{
663 struct blkg_rwstat rwstat = blkg_rwstat_recursive_sum(pd->blkg,
664 NULL, off);
665 return __blkg_prfill_rwstat(sf, pd, &rwstat);
666}
667
668/**
669 * blkg_print_stat_bytes_recursive - recursive version of blkg_print_stat_bytes
670 * @sf: seq_file to print to
671 * @v: unused
672 */
673int blkg_print_stat_bytes_recursive(struct seq_file *sf, void *v)
674{
675 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
676 blkg_prfill_rwstat_field_recursive,
677 (void *)seq_cft(sf)->private,
678 offsetof(struct blkcg_gq, stat_bytes), true);
679 return 0;
680}
681EXPORT_SYMBOL_GPL(blkg_print_stat_bytes_recursive);
682
683/**
684 * blkg_print_stat_ios_recursive - recursive version of blkg_print_stat_ios
685 * @sf: seq_file to print to
686 * @v: unused
687 */
688int blkg_print_stat_ios_recursive(struct seq_file *sf, void *v)
689{
690 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
691 blkg_prfill_rwstat_field_recursive,
692 (void *)seq_cft(sf)->private,
693 offsetof(struct blkcg_gq, stat_ios), true);
694 return 0;
695}
696EXPORT_SYMBOL_GPL(blkg_print_stat_ios_recursive);
697
698/**
699 * blkg_stat_recursive_sum - collect hierarchical blkg_stat
700 * @blkg: blkg of interest
701 * @pol: blkcg_policy which contains the blkg_stat
702 * @off: offset to the blkg_stat in blkg_policy_data or @blkg
703 *
704 * Collect the blkg_stat specified by @blkg, @pol and @off and all its
705 * online descendants and their aux counts. The caller must be holding the
706 * queue lock for online tests.
707 *
708 * If @pol is NULL, blkg_stat is at @off bytes into @blkg; otherwise, it is
709 * at @off bytes into @blkg's blkg_policy_data of the policy.
710 */
711u64 blkg_stat_recursive_sum(struct blkcg_gq *blkg,
712 struct blkcg_policy *pol, int off)
713{
714 struct blkcg_gq *pos_blkg;
715 struct cgroup_subsys_state *pos_css;
716 u64 sum = 0;
717
718 lockdep_assert_held(blkg->q->queue_lock);
719
720 rcu_read_lock();
721 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
722 struct blkg_stat *stat;
723
724 if (!pos_blkg->online)
725 continue;
726
727 if (pol)
728 stat = (void *)blkg_to_pd(pos_blkg, pol) + off;
729 else
730 stat = (void *)blkg + off;
731
732 sum += blkg_stat_read(stat) + atomic64_read(&stat->aux_cnt);
733 }
734 rcu_read_unlock();
735
736 return sum;
737}
738EXPORT_SYMBOL_GPL(blkg_stat_recursive_sum);
739
740/**
741 * blkg_rwstat_recursive_sum - collect hierarchical blkg_rwstat
742 * @blkg: blkg of interest
743 * @pol: blkcg_policy which contains the blkg_rwstat
744 * @off: offset to the blkg_rwstat in blkg_policy_data or @blkg
745 *
746 * Collect the blkg_rwstat specified by @blkg, @pol and @off and all its
747 * online descendants and their aux counts. The caller must be holding the
748 * queue lock for online tests.
749 *
750 * If @pol is NULL, blkg_rwstat is at @off bytes into @blkg; otherwise, it
751 * is at @off bytes into @blkg's blkg_policy_data of the policy.
752 */
753struct blkg_rwstat blkg_rwstat_recursive_sum(struct blkcg_gq *blkg,
754 struct blkcg_policy *pol, int off)
755{
756 struct blkcg_gq *pos_blkg;
757 struct cgroup_subsys_state *pos_css;
758 struct blkg_rwstat sum = { };
759 int i;
760
761 lockdep_assert_held(blkg->q->queue_lock);
762
763 rcu_read_lock();
764 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
765 struct blkg_rwstat *rwstat;
766
767 if (!pos_blkg->online)
768 continue;
769
770 if (pol)
771 rwstat = (void *)blkg_to_pd(pos_blkg, pol) + off;
772 else
773 rwstat = (void *)pos_blkg + off;
774
775 for (i = 0; i < BLKG_RWSTAT_NR; i++)
776 atomic64_add(atomic64_read(&rwstat->aux_cnt[i]) +
777 percpu_counter_sum_positive(&rwstat->cpu_cnt[i]),
778 &sum.aux_cnt[i]);
779 }
780 rcu_read_unlock();
781
782 return sum;
783}
784EXPORT_SYMBOL_GPL(blkg_rwstat_recursive_sum);
785
786/* Performs queue bypass and policy enabled checks then looks up blkg. */
787static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg,
788 const struct blkcg_policy *pol,
789 struct request_queue *q)
790{
791 WARN_ON_ONCE(!rcu_read_lock_held());
792 lockdep_assert_held(q->queue_lock);
793
794 if (!blkcg_policy_enabled(q, pol))
795 return ERR_PTR(-EOPNOTSUPP);
796
797 /*
798 * This could be the first entry point of blkcg implementation and
799 * we shouldn't allow anything to go through for a bypassing queue.
800 */
801 if (unlikely(blk_queue_bypass(q)))
802 return ERR_PTR(blk_queue_dying(q) ? -ENODEV : -EBUSY);
803
804 return __blkg_lookup(blkcg, q, true /* update_hint */);
805}
806
807/**
808 * blkg_conf_prep - parse and prepare for per-blkg config update
809 * @blkcg: target block cgroup
810 * @pol: target policy
811 * @input: input string
812 * @ctx: blkg_conf_ctx to be filled
813 *
814 * Parse per-blkg config update from @input and initialize @ctx with the
815 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
816 * part of @input following MAJ:MIN. This function returns with RCU read
817 * lock and queue lock held and must be paired with blkg_conf_finish().
818 */
819int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
820 char *input, struct blkg_conf_ctx *ctx)
821 __acquires(rcu) __acquires(disk->queue->queue_lock)
822{
823 struct gendisk *disk;
824 struct request_queue *q;
825 struct blkcg_gq *blkg;
826 unsigned int major, minor;
827 int key_len, part, ret;
828 char *body;
829
830 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
831 return -EINVAL;
832
833 body = input + key_len;
834 if (!isspace(*body))
835 return -EINVAL;
836 body = skip_spaces(body);
837
838 disk = get_gendisk(MKDEV(major, minor), &part);
839 if (!disk)
840 return -ENODEV;
841 if (part) {
842 ret = -ENODEV;
843 goto fail;
844 }
845
846 q = disk->queue;
847
848 rcu_read_lock();
849 spin_lock_irq(q->queue_lock);
850
851 blkg = blkg_lookup_check(blkcg, pol, q);
852 if (IS_ERR(blkg)) {
853 ret = PTR_ERR(blkg);
854 goto fail_unlock;
855 }
856
857 if (blkg)
858 goto success;
859
860 /*
861 * Create blkgs walking down from blkcg_root to @blkcg, so that all
862 * non-root blkgs have access to their parents.
863 */
864 while (true) {
865 struct blkcg *pos = blkcg;
866 struct blkcg *parent;
867 struct blkcg_gq *new_blkg;
868
869 parent = blkcg_parent(blkcg);
870 while (parent && !__blkg_lookup(parent, q, false)) {
871 pos = parent;
872 parent = blkcg_parent(parent);
873 }
874
875 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
876 spin_unlock_irq(q->queue_lock);
877 rcu_read_unlock();
878
879 new_blkg = blkg_alloc(pos, q, GFP_KERNEL);
880 if (unlikely(!new_blkg)) {
881 ret = -ENOMEM;
882 goto fail;
883 }
884
885 rcu_read_lock();
886 spin_lock_irq(q->queue_lock);
887
888 blkg = blkg_lookup_check(pos, pol, q);
889 if (IS_ERR(blkg)) {
890 ret = PTR_ERR(blkg);
891 goto fail_unlock;
892 }
893
894 if (blkg) {
895 blkg_free(new_blkg);
896 } else {
897 blkg = blkg_create(pos, q, new_blkg);
898 if (unlikely(IS_ERR(blkg))) {
899 ret = PTR_ERR(blkg);
900 goto fail_unlock;
901 }
902 }
903
904 if (pos == blkcg)
905 goto success;
906 }
907success:
908 ctx->disk = disk;
909 ctx->blkg = blkg;
910 ctx->body = body;
911 return 0;
912
913fail_unlock:
914 spin_unlock_irq(q->queue_lock);
915 rcu_read_unlock();
916fail:
917 put_disk_and_module(disk);
918 /*
919 * If queue was bypassing, we should retry. Do so after a
920 * short msleep(). It isn't strictly necessary but queue
921 * can be bypassing for some time and it's always nice to
922 * avoid busy looping.
923 */
924 if (ret == -EBUSY) {
925 msleep(10);
926 ret = restart_syscall();
927 }
928 return ret;
929}
930EXPORT_SYMBOL_GPL(blkg_conf_prep);
931
932/**
933 * blkg_conf_finish - finish up per-blkg config update
934 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
935 *
936 * Finish up after per-blkg config update. This function must be paired
937 * with blkg_conf_prep().
938 */
939void blkg_conf_finish(struct blkg_conf_ctx *ctx)
940 __releases(ctx->disk->queue->queue_lock) __releases(rcu)
941{
942 spin_unlock_irq(ctx->disk->queue->queue_lock);
943 rcu_read_unlock();
944 put_disk_and_module(ctx->disk);
945}
946EXPORT_SYMBOL_GPL(blkg_conf_finish);
947
948static int blkcg_print_stat(struct seq_file *sf, void *v)
949{
950 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
951 struct blkcg_gq *blkg;
952
953 rcu_read_lock();
954
955 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
956 const char *dname;
957 struct blkg_rwstat rwstat;
958 u64 rbytes, wbytes, rios, wios;
959
960 dname = blkg_dev_name(blkg);
961 if (!dname)
962 continue;
963
964 spin_lock_irq(blkg->q->queue_lock);
965
966 rwstat = blkg_rwstat_recursive_sum(blkg, NULL,
967 offsetof(struct blkcg_gq, stat_bytes));
968 rbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_READ]);
969 wbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_WRITE]);
970
971 rwstat = blkg_rwstat_recursive_sum(blkg, NULL,
972 offsetof(struct blkcg_gq, stat_ios));
973 rios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_READ]);
974 wios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_WRITE]);
975
976 spin_unlock_irq(blkg->q->queue_lock);
977
978 if (rbytes || wbytes || rios || wios)
979 seq_printf(sf, "%s rbytes=%llu wbytes=%llu rios=%llu wios=%llu\n",
980 dname, rbytes, wbytes, rios, wios);
981 }
982
983 rcu_read_unlock();
984 return 0;
985}
986
987static struct cftype blkcg_files[] = {
988 {
989 .name = "stat",
990 .flags = CFTYPE_NOT_ON_ROOT,
991 .seq_show = blkcg_print_stat,
992 },
993 { } /* terminate */
994};
995
996static struct cftype blkcg_legacy_files[] = {
997 {
998 .name = "reset_stats",
999 .write_u64 = blkcg_reset_stats,
1000 },
1001 { } /* terminate */
1002};
1003
1004/**
1005 * blkcg_css_offline - cgroup css_offline callback
1006 * @css: css of interest
1007 *
1008 * This function is called when @css is about to go away and responsible
1009 * for offlining all blkgs pd and killing all wbs associated with @css.
1010 * blkgs pd offline should be done while holding both q and blkcg locks.
1011 * As blkcg lock is nested inside q lock, this function performs reverse
1012 * double lock dancing.
1013 *
1014 * This is the blkcg counterpart of ioc_release_fn().
1015 */
1016static void blkcg_css_offline(struct cgroup_subsys_state *css)
1017{
1018 struct blkcg *blkcg = css_to_blkcg(css);
1019 struct blkcg_gq *blkg;
1020
1021 spin_lock_irq(&blkcg->lock);
1022
1023 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1024 struct request_queue *q = blkg->q;
1025
1026 if (spin_trylock(q->queue_lock)) {
1027 blkg_pd_offline(blkg);
1028 spin_unlock(q->queue_lock);
1029 } else {
1030 spin_unlock_irq(&blkcg->lock);
1031 cpu_relax();
1032 spin_lock_irq(&blkcg->lock);
1033 }
1034 }
1035
1036 spin_unlock_irq(&blkcg->lock);
1037
1038 wb_blkcg_offline(blkcg);
1039}
1040
1041/**
1042 * blkcg_destroy_all_blkgs - destroy all blkgs associated with a blkcg
1043 * @blkcg: blkcg of interest
1044 *
1045 * This function is called when blkcg css is about to free and responsible for
1046 * destroying all blkgs associated with @blkcg.
1047 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1048 * is nested inside q lock, this function performs reverse double lock dancing.
1049 */
1050static void blkcg_destroy_all_blkgs(struct blkcg *blkcg)
1051{
1052 spin_lock_irq(&blkcg->lock);
1053 while (!hlist_empty(&blkcg->blkg_list)) {
1054 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1055 struct blkcg_gq,
1056 blkcg_node);
1057 struct request_queue *q = blkg->q;
1058
1059 if (spin_trylock(q->queue_lock)) {
1060 blkg_destroy(blkg);
1061 spin_unlock(q->queue_lock);
1062 } else {
1063 spin_unlock_irq(&blkcg->lock);
1064 cpu_relax();
1065 spin_lock_irq(&blkcg->lock);
1066 }
1067 }
1068 spin_unlock_irq(&blkcg->lock);
1069}
1070
1071static void blkcg_css_free(struct cgroup_subsys_state *css)
1072{
1073 struct blkcg *blkcg = css_to_blkcg(css);
1074 int i;
1075
1076 blkcg_destroy_all_blkgs(blkcg);
1077
1078 mutex_lock(&blkcg_pol_mutex);
1079
1080 list_del(&blkcg->all_blkcgs_node);
1081
1082 for (i = 0; i < BLKCG_MAX_POLS; i++)
1083 if (blkcg->cpd[i])
1084 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1085
1086 mutex_unlock(&blkcg_pol_mutex);
1087
1088 kfree(blkcg);
1089}
1090
1091static struct cgroup_subsys_state *
1092blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1093{
1094 struct blkcg *blkcg;
1095 struct cgroup_subsys_state *ret;
1096 int i;
1097
1098 mutex_lock(&blkcg_pol_mutex);
1099
1100 if (!parent_css) {
1101 blkcg = &blkcg_root;
1102 } else {
1103 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1104 if (!blkcg) {
1105 ret = ERR_PTR(-ENOMEM);
1106 goto unlock;
1107 }
1108 }
1109
1110 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1111 struct blkcg_policy *pol = blkcg_policy[i];
1112 struct blkcg_policy_data *cpd;
1113
1114 /*
1115 * If the policy hasn't been attached yet, wait for it
1116 * to be attached before doing anything else. Otherwise,
1117 * check if the policy requires any specific per-cgroup
1118 * data: if it does, allocate and initialize it.
1119 */
1120 if (!pol || !pol->cpd_alloc_fn)
1121 continue;
1122
1123 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1124 if (!cpd) {
1125 ret = ERR_PTR(-ENOMEM);
1126 goto free_pd_blkcg;
1127 }
1128 blkcg->cpd[i] = cpd;
1129 cpd->blkcg = blkcg;
1130 cpd->plid = i;
1131 if (pol->cpd_init_fn)
1132 pol->cpd_init_fn(cpd);
1133 }
1134
1135 spin_lock_init(&blkcg->lock);
1136 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1137 INIT_HLIST_HEAD(&blkcg->blkg_list);
1138#ifdef CONFIG_CGROUP_WRITEBACK
1139 INIT_LIST_HEAD(&blkcg->cgwb_list);
1140#endif
1141 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1142
1143 mutex_unlock(&blkcg_pol_mutex);
1144 return &blkcg->css;
1145
1146free_pd_blkcg:
1147 for (i--; i >= 0; i--)
1148 if (blkcg->cpd[i])
1149 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1150
1151 if (blkcg != &blkcg_root)
1152 kfree(blkcg);
1153unlock:
1154 mutex_unlock(&blkcg_pol_mutex);
1155 return ret;
1156}
1157
1158/**
1159 * blkcg_init_queue - initialize blkcg part of request queue
1160 * @q: request_queue to initialize
1161 *
1162 * Called from blk_alloc_queue_node(). Responsible for initializing blkcg
1163 * part of new request_queue @q.
1164 *
1165 * RETURNS:
1166 * 0 on success, -errno on failure.
1167 */
1168int blkcg_init_queue(struct request_queue *q)
1169{
1170 struct blkcg_gq *new_blkg, *blkg;
1171 bool preloaded;
1172 int ret;
1173
1174 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1175 if (!new_blkg)
1176 return -ENOMEM;
1177
1178 preloaded = !radix_tree_preload(GFP_KERNEL);
1179
1180 /* Make sure the root blkg exists. */
1181 rcu_read_lock();
1182 spin_lock_irq(q->queue_lock);
1183 blkg = blkg_create(&blkcg_root, q, new_blkg);
1184 if (IS_ERR(blkg))
1185 goto err_unlock;
1186 q->root_blkg = blkg;
1187 q->root_rl.blkg = blkg;
1188 spin_unlock_irq(q->queue_lock);
1189 rcu_read_unlock();
1190
1191 if (preloaded)
1192 radix_tree_preload_end();
1193
1194 ret = blk_throtl_init(q);
1195 if (ret) {
1196 spin_lock_irq(q->queue_lock);
1197 blkg_destroy_all(q);
1198 spin_unlock_irq(q->queue_lock);
1199 }
1200 return ret;
1201
1202err_unlock:
1203 spin_unlock_irq(q->queue_lock);
1204 rcu_read_unlock();
1205 if (preloaded)
1206 radix_tree_preload_end();
1207 return PTR_ERR(blkg);
1208}
1209
1210/**
1211 * blkcg_drain_queue - drain blkcg part of request_queue
1212 * @q: request_queue to drain
1213 *
1214 * Called from blk_drain_queue(). Responsible for draining blkcg part.
1215 */
1216void blkcg_drain_queue(struct request_queue *q)
1217{
1218 lockdep_assert_held(q->queue_lock);
1219
1220 /*
1221 * @q could be exiting and already have destroyed all blkgs as
1222 * indicated by NULL root_blkg. If so, don't confuse policies.
1223 */
1224 if (!q->root_blkg)
1225 return;
1226
1227 blk_throtl_drain(q);
1228}
1229
1230/**
1231 * blkcg_exit_queue - exit and release blkcg part of request_queue
1232 * @q: request_queue being released
1233 *
1234 * Called from blk_release_queue(). Responsible for exiting blkcg part.
1235 */
1236void blkcg_exit_queue(struct request_queue *q)
1237{
1238 spin_lock_irq(q->queue_lock);
1239 blkg_destroy_all(q);
1240 spin_unlock_irq(q->queue_lock);
1241
1242 blk_throtl_exit(q);
1243}
1244
1245/*
1246 * We cannot support shared io contexts, as we have no mean to support
1247 * two tasks with the same ioc in two different groups without major rework
1248 * of the main cic data structures. For now we allow a task to change
1249 * its cgroup only if it's the only owner of its ioc.
1250 */
1251static int blkcg_can_attach(struct cgroup_taskset *tset)
1252{
1253 struct task_struct *task;
1254 struct cgroup_subsys_state *dst_css;
1255 struct io_context *ioc;
1256 int ret = 0;
1257
1258 /* task_lock() is needed to avoid races with exit_io_context() */
1259 cgroup_taskset_for_each(task, dst_css, tset) {
1260 task_lock(task);
1261 ioc = task->io_context;
1262 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
1263 ret = -EINVAL;
1264 task_unlock(task);
1265 if (ret)
1266 break;
1267 }
1268 return ret;
1269}
1270
1271static void blkcg_bind(struct cgroup_subsys_state *root_css)
1272{
1273 int i;
1274
1275 mutex_lock(&blkcg_pol_mutex);
1276
1277 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1278 struct blkcg_policy *pol = blkcg_policy[i];
1279 struct blkcg *blkcg;
1280
1281 if (!pol || !pol->cpd_bind_fn)
1282 continue;
1283
1284 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1285 if (blkcg->cpd[pol->plid])
1286 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1287 }
1288 mutex_unlock(&blkcg_pol_mutex);
1289}
1290
1291struct cgroup_subsys io_cgrp_subsys = {
1292 .css_alloc = blkcg_css_alloc,
1293 .css_offline = blkcg_css_offline,
1294 .css_free = blkcg_css_free,
1295 .can_attach = blkcg_can_attach,
1296 .bind = blkcg_bind,
1297 .dfl_cftypes = blkcg_files,
1298 .legacy_cftypes = blkcg_legacy_files,
1299 .legacy_name = "blkio",
1300#ifdef CONFIG_MEMCG
1301 /*
1302 * This ensures that, if available, memcg is automatically enabled
1303 * together on the default hierarchy so that the owner cgroup can
1304 * be retrieved from writeback pages.
1305 */
1306 .depends_on = 1 << memory_cgrp_id,
1307#endif
1308};
1309EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1310
1311/**
1312 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1313 * @q: request_queue of interest
1314 * @pol: blkcg policy to activate
1315 *
1316 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1317 * bypass mode to populate its blkgs with policy_data for @pol.
1318 *
1319 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1320 * from IO path. Update of each blkg is protected by both queue and blkcg
1321 * locks so that holding either lock and testing blkcg_policy_enabled() is
1322 * always enough for dereferencing policy data.
1323 *
1324 * The caller is responsible for synchronizing [de]activations and policy
1325 * [un]registerations. Returns 0 on success, -errno on failure.
1326 */
1327int blkcg_activate_policy(struct request_queue *q,
1328 const struct blkcg_policy *pol)
1329{
1330 struct blkg_policy_data *pd_prealloc = NULL;
1331 struct blkcg_gq *blkg;
1332 int ret;
1333
1334 if (blkcg_policy_enabled(q, pol))
1335 return 0;
1336
1337 if (q->mq_ops)
1338 blk_mq_freeze_queue(q);
1339 else
1340 blk_queue_bypass_start(q);
1341pd_prealloc:
1342 if (!pd_prealloc) {
1343 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q->node);
1344 if (!pd_prealloc) {
1345 ret = -ENOMEM;
1346 goto out_bypass_end;
1347 }
1348 }
1349
1350 spin_lock_irq(q->queue_lock);
1351
1352 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1353 struct blkg_policy_data *pd;
1354
1355 if (blkg->pd[pol->plid])
1356 continue;
1357
1358 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q->node);
1359 if (!pd)
1360 swap(pd, pd_prealloc);
1361 if (!pd) {
1362 spin_unlock_irq(q->queue_lock);
1363 goto pd_prealloc;
1364 }
1365
1366 blkg->pd[pol->plid] = pd;
1367 pd->blkg = blkg;
1368 pd->plid = pol->plid;
1369 if (pol->pd_init_fn)
1370 pol->pd_init_fn(pd);
1371 }
1372
1373 __set_bit(pol->plid, q->blkcg_pols);
1374 ret = 0;
1375
1376 spin_unlock_irq(q->queue_lock);
1377out_bypass_end:
1378 if (q->mq_ops)
1379 blk_mq_unfreeze_queue(q);
1380 else
1381 blk_queue_bypass_end(q);
1382 if (pd_prealloc)
1383 pol->pd_free_fn(pd_prealloc);
1384 return ret;
1385}
1386EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1387
1388/**
1389 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1390 * @q: request_queue of interest
1391 * @pol: blkcg policy to deactivate
1392 *
1393 * Deactivate @pol on @q. Follows the same synchronization rules as
1394 * blkcg_activate_policy().
1395 */
1396void blkcg_deactivate_policy(struct request_queue *q,
1397 const struct blkcg_policy *pol)
1398{
1399 struct blkcg_gq *blkg;
1400
1401 if (!blkcg_policy_enabled(q, pol))
1402 return;
1403
1404 if (q->mq_ops)
1405 blk_mq_freeze_queue(q);
1406 else
1407 blk_queue_bypass_start(q);
1408
1409 spin_lock_irq(q->queue_lock);
1410
1411 __clear_bit(pol->plid, q->blkcg_pols);
1412
1413 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1414 if (blkg->pd[pol->plid]) {
1415 if (!blkg->pd[pol->plid]->offline &&
1416 pol->pd_offline_fn) {
1417 pol->pd_offline_fn(blkg->pd[pol->plid]);
1418 blkg->pd[pol->plid]->offline = true;
1419 }
1420 pol->pd_free_fn(blkg->pd[pol->plid]);
1421 blkg->pd[pol->plid] = NULL;
1422 }
1423 }
1424
1425 spin_unlock_irq(q->queue_lock);
1426
1427 if (q->mq_ops)
1428 blk_mq_unfreeze_queue(q);
1429 else
1430 blk_queue_bypass_end(q);
1431}
1432EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1433
1434/**
1435 * blkcg_policy_register - register a blkcg policy
1436 * @pol: blkcg policy to register
1437 *
1438 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1439 * successful registration. Returns 0 on success and -errno on failure.
1440 */
1441int blkcg_policy_register(struct blkcg_policy *pol)
1442{
1443 struct blkcg *blkcg;
1444 int i, ret;
1445
1446 mutex_lock(&blkcg_pol_register_mutex);
1447 mutex_lock(&blkcg_pol_mutex);
1448
1449 /* find an empty slot */
1450 ret = -ENOSPC;
1451 for (i = 0; i < BLKCG_MAX_POLS; i++)
1452 if (!blkcg_policy[i])
1453 break;
1454 if (i >= BLKCG_MAX_POLS)
1455 goto err_unlock;
1456
1457 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1458 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1459 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1460 goto err_unlock;
1461
1462 /* register @pol */
1463 pol->plid = i;
1464 blkcg_policy[pol->plid] = pol;
1465
1466 /* allocate and install cpd's */
1467 if (pol->cpd_alloc_fn) {
1468 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1469 struct blkcg_policy_data *cpd;
1470
1471 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1472 if (!cpd)
1473 goto err_free_cpds;
1474
1475 blkcg->cpd[pol->plid] = cpd;
1476 cpd->blkcg = blkcg;
1477 cpd->plid = pol->plid;
1478 pol->cpd_init_fn(cpd);
1479 }
1480 }
1481
1482 mutex_unlock(&blkcg_pol_mutex);
1483
1484 /* everything is in place, add intf files for the new policy */
1485 if (pol->dfl_cftypes)
1486 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1487 pol->dfl_cftypes));
1488 if (pol->legacy_cftypes)
1489 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1490 pol->legacy_cftypes));
1491 mutex_unlock(&blkcg_pol_register_mutex);
1492 return 0;
1493
1494err_free_cpds:
1495 if (pol->cpd_free_fn) {
1496 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1497 if (blkcg->cpd[pol->plid]) {
1498 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1499 blkcg->cpd[pol->plid] = NULL;
1500 }
1501 }
1502 }
1503 blkcg_policy[pol->plid] = NULL;
1504err_unlock:
1505 mutex_unlock(&blkcg_pol_mutex);
1506 mutex_unlock(&blkcg_pol_register_mutex);
1507 return ret;
1508}
1509EXPORT_SYMBOL_GPL(blkcg_policy_register);
1510
1511/**
1512 * blkcg_policy_unregister - unregister a blkcg policy
1513 * @pol: blkcg policy to unregister
1514 *
1515 * Undo blkcg_policy_register(@pol). Might sleep.
1516 */
1517void blkcg_policy_unregister(struct blkcg_policy *pol)
1518{
1519 struct blkcg *blkcg;
1520
1521 mutex_lock(&blkcg_pol_register_mutex);
1522
1523 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1524 goto out_unlock;
1525
1526 /* kill the intf files first */
1527 if (pol->dfl_cftypes)
1528 cgroup_rm_cftypes(pol->dfl_cftypes);
1529 if (pol->legacy_cftypes)
1530 cgroup_rm_cftypes(pol->legacy_cftypes);
1531
1532 /* remove cpds and unregister */
1533 mutex_lock(&blkcg_pol_mutex);
1534
1535 if (pol->cpd_free_fn) {
1536 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1537 if (blkcg->cpd[pol->plid]) {
1538 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1539 blkcg->cpd[pol->plid] = NULL;
1540 }
1541 }
1542 }
1543 blkcg_policy[pol->plid] = NULL;
1544
1545 mutex_unlock(&blkcg_pol_mutex);
1546out_unlock:
1547 mutex_unlock(&blkcg_pol_register_mutex);
1548}
1549EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
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,
189 GFP_NOWAIT | __GFP_NOWARN);
190 if (!wb_congested) {
191 ret = -ENOMEM;
192 goto err_put_css;
193 }
194
195 /* allocate */
196 if (!new_blkg) {
197 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN);
198 if (unlikely(!new_blkg)) {
199 ret = -ENOMEM;
200 goto err_put_congested;
201 }
202 }
203 blkg = new_blkg;
204 blkg->wb_congested = wb_congested;
205
206 /* link parent */
207 if (blkcg_parent(blkcg)) {
208 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
209 if (WARN_ON_ONCE(!blkg->parent)) {
210 ret = -ENODEV;
211 goto err_put_congested;
212 }
213 blkg_get(blkg->parent);
214 }
215
216 /* invoke per-policy init */
217 for (i = 0; i < BLKCG_MAX_POLS; i++) {
218 struct blkcg_policy *pol = blkcg_policy[i];
219
220 if (blkg->pd[i] && pol->pd_init_fn)
221 pol->pd_init_fn(blkg->pd[i]);
222 }
223
224 /* insert */
225 spin_lock(&blkcg->lock);
226 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
227 if (likely(!ret)) {
228 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
229 list_add(&blkg->q_node, &q->blkg_list);
230
231 for (i = 0; i < BLKCG_MAX_POLS; i++) {
232 struct blkcg_policy *pol = blkcg_policy[i];
233
234 if (blkg->pd[i] && pol->pd_online_fn)
235 pol->pd_online_fn(blkg->pd[i]);
236 }
237 }
238 blkg->online = true;
239 spin_unlock(&blkcg->lock);
240
241 if (!ret)
242 return blkg;
243
244 /* @blkg failed fully initialized, use the usual release path */
245 blkg_put(blkg);
246 return ERR_PTR(ret);
247
248err_put_congested:
249 wb_congested_put(wb_congested);
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) ? -ENODEV : -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}
308
309static void blkg_destroy(struct blkcg_gq *blkg)
310{
311 struct blkcg *blkcg = blkg->blkcg;
312 struct blkcg_gq *parent = blkg->parent;
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->pd[i]);
327 }
328
329 if (parent) {
330 blkg_rwstat_add_aux(&parent->stat_bytes, &blkg->stat_bytes);
331 blkg_rwstat_add_aux(&parent->stat_ios, &blkg->stat_ios);
332 }
333
334 blkg->online = false;
335
336 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
337 list_del_init(&blkg->q_node);
338 hlist_del_init_rcu(&blkg->blkcg_node);
339
340 /*
341 * Both setting lookup hint to and clearing it from @blkg are done
342 * under queue_lock. If it's not pointing to @blkg now, it never
343 * will. Hint assignment itself can race safely.
344 */
345 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
346 rcu_assign_pointer(blkcg->blkg_hint, NULL);
347
348 /*
349 * Put the reference taken at the time of creation so that when all
350 * queues are gone, group can be destroyed.
351 */
352 blkg_put(blkg);
353}
354
355/**
356 * blkg_destroy_all - destroy all blkgs associated with a request_queue
357 * @q: request_queue of interest
358 *
359 * Destroy all blkgs associated with @q.
360 */
361static void blkg_destroy_all(struct request_queue *q)
362{
363 struct blkcg_gq *blkg, *n;
364
365 lockdep_assert_held(q->queue_lock);
366
367 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
368 struct blkcg *blkcg = blkg->blkcg;
369
370 spin_lock(&blkcg->lock);
371 blkg_destroy(blkg);
372 spin_unlock(&blkcg->lock);
373 }
374
375 q->root_blkg = NULL;
376 q->root_rl.blkg = NULL;
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
391 /* release the blkcg and parent blkg refs this blkg has been holding */
392 css_put(&blkg->blkcg->css);
393 if (blkg->parent)
394 blkg_put(blkg->parent);
395
396 wb_congested_put(blkg->wb_congested);
397
398 blkg_free(blkg);
399}
400EXPORT_SYMBOL_GPL(__blkg_release_rcu);
401
402/*
403 * The next function used by blk_queue_for_each_rl(). It's a bit tricky
404 * because the root blkg uses @q->root_rl instead of its own rl.
405 */
406struct request_list *__blk_queue_next_rl(struct request_list *rl,
407 struct request_queue *q)
408{
409 struct list_head *ent;
410 struct blkcg_gq *blkg;
411
412 /*
413 * Determine the current blkg list_head. The first entry is
414 * root_rl which is off @q->blkg_list and mapped to the head.
415 */
416 if (rl == &q->root_rl) {
417 ent = &q->blkg_list;
418 /* There are no more block groups, hence no request lists */
419 if (list_empty(ent))
420 return NULL;
421 } else {
422 blkg = container_of(rl, struct blkcg_gq, rl);
423 ent = &blkg->q_node;
424 }
425
426 /* walk to the next list_head, skip root blkcg */
427 ent = ent->next;
428 if (ent == &q->root_blkg->q_node)
429 ent = ent->next;
430 if (ent == &q->blkg_list)
431 return NULL;
432
433 blkg = container_of(ent, struct blkcg_gq, q_node);
434 return &blkg->rl;
435}
436
437static int blkcg_reset_stats(struct cgroup_subsys_state *css,
438 struct cftype *cftype, u64 val)
439{
440 struct blkcg *blkcg = css_to_blkcg(css);
441 struct blkcg_gq *blkg;
442 int i;
443
444 mutex_lock(&blkcg_pol_mutex);
445 spin_lock_irq(&blkcg->lock);
446
447 /*
448 * Note that stat reset is racy - it doesn't synchronize against
449 * stat updates. This is a debug feature which shouldn't exist
450 * anyway. If you get hit by a race, retry.
451 */
452 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
453 blkg_rwstat_reset(&blkg->stat_bytes);
454 blkg_rwstat_reset(&blkg->stat_ios);
455
456 for (i = 0; i < BLKCG_MAX_POLS; i++) {
457 struct blkcg_policy *pol = blkcg_policy[i];
458
459 if (blkg->pd[i] && pol->pd_reset_stats_fn)
460 pol->pd_reset_stats_fn(blkg->pd[i]);
461 }
462 }
463
464 spin_unlock_irq(&blkcg->lock);
465 mutex_unlock(&blkcg_pol_mutex);
466 return 0;
467}
468
469const char *blkg_dev_name(struct blkcg_gq *blkg)
470{
471 /* some drivers (floppy) instantiate a queue w/o disk registered */
472 if (blkg->q->backing_dev_info.dev)
473 return dev_name(blkg->q->backing_dev_info.dev);
474 return NULL;
475}
476EXPORT_SYMBOL_GPL(blkg_dev_name);
477
478/**
479 * blkcg_print_blkgs - helper for printing per-blkg data
480 * @sf: seq_file to print to
481 * @blkcg: blkcg of interest
482 * @prfill: fill function to print out a blkg
483 * @pol: policy in question
484 * @data: data to be passed to @prfill
485 * @show_total: to print out sum of prfill return values or not
486 *
487 * This function invokes @prfill on each blkg of @blkcg if pd for the
488 * policy specified by @pol exists. @prfill is invoked with @sf, the
489 * policy data and @data and the matching queue lock held. If @show_total
490 * is %true, the sum of the return values from @prfill is printed with
491 * "Total" label at the end.
492 *
493 * This is to be used to construct print functions for
494 * cftype->read_seq_string method.
495 */
496void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
497 u64 (*prfill)(struct seq_file *,
498 struct blkg_policy_data *, int),
499 const struct blkcg_policy *pol, int data,
500 bool show_total)
501{
502 struct blkcg_gq *blkg;
503 u64 total = 0;
504
505 rcu_read_lock();
506 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
507 spin_lock_irq(blkg->q->queue_lock);
508 if (blkcg_policy_enabled(blkg->q, pol))
509 total += prfill(sf, blkg->pd[pol->plid], data);
510 spin_unlock_irq(blkg->q->queue_lock);
511 }
512 rcu_read_unlock();
513
514 if (show_total)
515 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
516}
517EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
518
519/**
520 * __blkg_prfill_u64 - prfill helper for a single u64 value
521 * @sf: seq_file to print to
522 * @pd: policy private data of interest
523 * @v: value to print
524 *
525 * Print @v to @sf for the device assocaited with @pd.
526 */
527u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
528{
529 const char *dname = blkg_dev_name(pd->blkg);
530
531 if (!dname)
532 return 0;
533
534 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
535 return v;
536}
537EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
538
539/**
540 * __blkg_prfill_rwstat - prfill helper for a blkg_rwstat
541 * @sf: seq_file to print to
542 * @pd: policy private data of interest
543 * @rwstat: rwstat to print
544 *
545 * Print @rwstat to @sf for the device assocaited with @pd.
546 */
547u64 __blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
548 const struct blkg_rwstat *rwstat)
549{
550 static const char *rwstr[] = {
551 [BLKG_RWSTAT_READ] = "Read",
552 [BLKG_RWSTAT_WRITE] = "Write",
553 [BLKG_RWSTAT_SYNC] = "Sync",
554 [BLKG_RWSTAT_ASYNC] = "Async",
555 };
556 const char *dname = blkg_dev_name(pd->blkg);
557 u64 v;
558 int i;
559
560 if (!dname)
561 return 0;
562
563 for (i = 0; i < BLKG_RWSTAT_NR; i++)
564 seq_printf(sf, "%s %s %llu\n", dname, rwstr[i],
565 (unsigned long long)atomic64_read(&rwstat->aux_cnt[i]));
566
567 v = atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_READ]) +
568 atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_WRITE]);
569 seq_printf(sf, "%s Total %llu\n", dname, (unsigned long long)v);
570 return v;
571}
572EXPORT_SYMBOL_GPL(__blkg_prfill_rwstat);
573
574/**
575 * blkg_prfill_stat - prfill callback for blkg_stat
576 * @sf: seq_file to print to
577 * @pd: policy private data of interest
578 * @off: offset to the blkg_stat in @pd
579 *
580 * prfill callback for printing a blkg_stat.
581 */
582u64 blkg_prfill_stat(struct seq_file *sf, struct blkg_policy_data *pd, int off)
583{
584 return __blkg_prfill_u64(sf, pd, blkg_stat_read((void *)pd + off));
585}
586EXPORT_SYMBOL_GPL(blkg_prfill_stat);
587
588/**
589 * blkg_prfill_rwstat - prfill callback for blkg_rwstat
590 * @sf: seq_file to print to
591 * @pd: policy private data of interest
592 * @off: offset to the blkg_rwstat in @pd
593 *
594 * prfill callback for printing a blkg_rwstat.
595 */
596u64 blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
597 int off)
598{
599 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd + off);
600
601 return __blkg_prfill_rwstat(sf, pd, &rwstat);
602}
603EXPORT_SYMBOL_GPL(blkg_prfill_rwstat);
604
605static u64 blkg_prfill_rwstat_field(struct seq_file *sf,
606 struct blkg_policy_data *pd, int off)
607{
608 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd->blkg + off);
609
610 return __blkg_prfill_rwstat(sf, pd, &rwstat);
611}
612
613/**
614 * blkg_print_stat_bytes - seq_show callback for blkg->stat_bytes
615 * @sf: seq_file to print to
616 * @v: unused
617 *
618 * To be used as cftype->seq_show to print blkg->stat_bytes.
619 * cftype->private must be set to the blkcg_policy.
620 */
621int blkg_print_stat_bytes(struct seq_file *sf, void *v)
622{
623 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
624 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
625 offsetof(struct blkcg_gq, stat_bytes), true);
626 return 0;
627}
628EXPORT_SYMBOL_GPL(blkg_print_stat_bytes);
629
630/**
631 * blkg_print_stat_bytes - seq_show callback for blkg->stat_ios
632 * @sf: seq_file to print to
633 * @v: unused
634 *
635 * To be used as cftype->seq_show to print blkg->stat_ios. cftype->private
636 * must be set to the blkcg_policy.
637 */
638int blkg_print_stat_ios(struct seq_file *sf, void *v)
639{
640 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
641 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
642 offsetof(struct blkcg_gq, stat_ios), true);
643 return 0;
644}
645EXPORT_SYMBOL_GPL(blkg_print_stat_ios);
646
647static u64 blkg_prfill_rwstat_field_recursive(struct seq_file *sf,
648 struct blkg_policy_data *pd,
649 int off)
650{
651 struct blkg_rwstat rwstat = blkg_rwstat_recursive_sum(pd->blkg,
652 NULL, off);
653 return __blkg_prfill_rwstat(sf, pd, &rwstat);
654}
655
656/**
657 * blkg_print_stat_bytes_recursive - recursive version of blkg_print_stat_bytes
658 * @sf: seq_file to print to
659 * @v: unused
660 */
661int blkg_print_stat_bytes_recursive(struct seq_file *sf, void *v)
662{
663 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
664 blkg_prfill_rwstat_field_recursive,
665 (void *)seq_cft(sf)->private,
666 offsetof(struct blkcg_gq, stat_bytes), true);
667 return 0;
668}
669EXPORT_SYMBOL_GPL(blkg_print_stat_bytes_recursive);
670
671/**
672 * blkg_print_stat_ios_recursive - recursive version of blkg_print_stat_ios
673 * @sf: seq_file to print to
674 * @v: unused
675 */
676int blkg_print_stat_ios_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_ios), true);
682 return 0;
683}
684EXPORT_SYMBOL_GPL(blkg_print_stat_ios_recursive);
685
686/**
687 * blkg_stat_recursive_sum - collect hierarchical blkg_stat
688 * @blkg: blkg of interest
689 * @pol: blkcg_policy which contains the blkg_stat
690 * @off: offset to the blkg_stat in blkg_policy_data or @blkg
691 *
692 * Collect the blkg_stat specified by @blkg, @pol and @off and all its
693 * online descendants and their aux counts. The caller must be holding the
694 * queue lock for online tests.
695 *
696 * If @pol is NULL, blkg_stat is at @off bytes into @blkg; otherwise, it is
697 * at @off bytes into @blkg's blkg_policy_data of the policy.
698 */
699u64 blkg_stat_recursive_sum(struct blkcg_gq *blkg,
700 struct blkcg_policy *pol, int off)
701{
702 struct blkcg_gq *pos_blkg;
703 struct cgroup_subsys_state *pos_css;
704 u64 sum = 0;
705
706 lockdep_assert_held(blkg->q->queue_lock);
707
708 rcu_read_lock();
709 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
710 struct blkg_stat *stat;
711
712 if (!pos_blkg->online)
713 continue;
714
715 if (pol)
716 stat = (void *)blkg_to_pd(pos_blkg, pol) + off;
717 else
718 stat = (void *)blkg + off;
719
720 sum += blkg_stat_read(stat) + atomic64_read(&stat->aux_cnt);
721 }
722 rcu_read_unlock();
723
724 return sum;
725}
726EXPORT_SYMBOL_GPL(blkg_stat_recursive_sum);
727
728/**
729 * blkg_rwstat_recursive_sum - collect hierarchical blkg_rwstat
730 * @blkg: blkg of interest
731 * @pol: blkcg_policy which contains the blkg_rwstat
732 * @off: offset to the blkg_rwstat in blkg_policy_data or @blkg
733 *
734 * Collect the blkg_rwstat specified by @blkg, @pol and @off and all its
735 * online descendants and their aux counts. The caller must be holding the
736 * queue lock for online tests.
737 *
738 * If @pol is NULL, blkg_rwstat is at @off bytes into @blkg; otherwise, it
739 * is at @off bytes into @blkg's blkg_policy_data of the policy.
740 */
741struct blkg_rwstat blkg_rwstat_recursive_sum(struct blkcg_gq *blkg,
742 struct blkcg_policy *pol, int off)
743{
744 struct blkcg_gq *pos_blkg;
745 struct cgroup_subsys_state *pos_css;
746 struct blkg_rwstat sum = { };
747 int i;
748
749 lockdep_assert_held(blkg->q->queue_lock);
750
751 rcu_read_lock();
752 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
753 struct blkg_rwstat *rwstat;
754
755 if (!pos_blkg->online)
756 continue;
757
758 if (pol)
759 rwstat = (void *)blkg_to_pd(pos_blkg, pol) + off;
760 else
761 rwstat = (void *)pos_blkg + off;
762
763 for (i = 0; i < BLKG_RWSTAT_NR; i++)
764 atomic64_add(atomic64_read(&rwstat->aux_cnt[i]) +
765 percpu_counter_sum_positive(&rwstat->cpu_cnt[i]),
766 &sum.aux_cnt[i]);
767 }
768 rcu_read_unlock();
769
770 return sum;
771}
772EXPORT_SYMBOL_GPL(blkg_rwstat_recursive_sum);
773
774/**
775 * blkg_conf_prep - parse and prepare for per-blkg config update
776 * @blkcg: target block cgroup
777 * @pol: target policy
778 * @input: input string
779 * @ctx: blkg_conf_ctx to be filled
780 *
781 * Parse per-blkg config update from @input and initialize @ctx with the
782 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
783 * part of @input following MAJ:MIN. This function returns with RCU read
784 * lock and queue lock held and must be paired with blkg_conf_finish().
785 */
786int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
787 char *input, struct blkg_conf_ctx *ctx)
788 __acquires(rcu) __acquires(disk->queue->queue_lock)
789{
790 struct gendisk *disk;
791 struct blkcg_gq *blkg;
792 struct module *owner;
793 unsigned int major, minor;
794 int key_len, part, ret;
795 char *body;
796
797 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
798 return -EINVAL;
799
800 body = input + key_len;
801 if (!isspace(*body))
802 return -EINVAL;
803 body = skip_spaces(body);
804
805 disk = get_gendisk(MKDEV(major, minor), &part);
806 if (!disk)
807 return -ENODEV;
808 if (part) {
809 owner = disk->fops->owner;
810 put_disk(disk);
811 module_put(owner);
812 return -ENODEV;
813 }
814
815 rcu_read_lock();
816 spin_lock_irq(disk->queue->queue_lock);
817
818 if (blkcg_policy_enabled(disk->queue, pol))
819 blkg = blkg_lookup_create(blkcg, disk->queue);
820 else
821 blkg = ERR_PTR(-EOPNOTSUPP);
822
823 if (IS_ERR(blkg)) {
824 ret = PTR_ERR(blkg);
825 rcu_read_unlock();
826 spin_unlock_irq(disk->queue->queue_lock);
827 owner = disk->fops->owner;
828 put_disk(disk);
829 module_put(owner);
830 /*
831 * If queue was bypassing, we should retry. Do so after a
832 * short msleep(). It isn't strictly necessary but queue
833 * can be bypassing for some time and it's always nice to
834 * avoid busy looping.
835 */
836 if (ret == -EBUSY) {
837 msleep(10);
838 ret = restart_syscall();
839 }
840 return ret;
841 }
842
843 ctx->disk = disk;
844 ctx->blkg = blkg;
845 ctx->body = body;
846 return 0;
847}
848EXPORT_SYMBOL_GPL(blkg_conf_prep);
849
850/**
851 * blkg_conf_finish - finish up per-blkg config update
852 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
853 *
854 * Finish up after per-blkg config update. This function must be paired
855 * with blkg_conf_prep().
856 */
857void blkg_conf_finish(struct blkg_conf_ctx *ctx)
858 __releases(ctx->disk->queue->queue_lock) __releases(rcu)
859{
860 struct module *owner;
861
862 spin_unlock_irq(ctx->disk->queue->queue_lock);
863 rcu_read_unlock();
864 owner = ctx->disk->fops->owner;
865 put_disk(ctx->disk);
866 module_put(owner);
867}
868EXPORT_SYMBOL_GPL(blkg_conf_finish);
869
870static int blkcg_print_stat(struct seq_file *sf, void *v)
871{
872 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
873 struct blkcg_gq *blkg;
874
875 rcu_read_lock();
876
877 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
878 const char *dname;
879 struct blkg_rwstat rwstat;
880 u64 rbytes, wbytes, rios, wios;
881
882 dname = blkg_dev_name(blkg);
883 if (!dname)
884 continue;
885
886 spin_lock_irq(blkg->q->queue_lock);
887
888 rwstat = blkg_rwstat_recursive_sum(blkg, NULL,
889 offsetof(struct blkcg_gq, stat_bytes));
890 rbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_READ]);
891 wbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_WRITE]);
892
893 rwstat = blkg_rwstat_recursive_sum(blkg, NULL,
894 offsetof(struct blkcg_gq, stat_ios));
895 rios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_READ]);
896 wios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_WRITE]);
897
898 spin_unlock_irq(blkg->q->queue_lock);
899
900 if (rbytes || wbytes || rios || wios)
901 seq_printf(sf, "%s rbytes=%llu wbytes=%llu rios=%llu wios=%llu\n",
902 dname, rbytes, wbytes, rios, wios);
903 }
904
905 rcu_read_unlock();
906 return 0;
907}
908
909static struct cftype blkcg_files[] = {
910 {
911 .name = "stat",
912 .flags = CFTYPE_NOT_ON_ROOT,
913 .seq_show = blkcg_print_stat,
914 },
915 { } /* terminate */
916};
917
918static struct cftype blkcg_legacy_files[] = {
919 {
920 .name = "reset_stats",
921 .write_u64 = blkcg_reset_stats,
922 },
923 { } /* terminate */
924};
925
926/**
927 * blkcg_css_offline - cgroup css_offline callback
928 * @css: css of interest
929 *
930 * This function is called when @css is about to go away and responsible
931 * for shooting down all blkgs associated with @css. blkgs should be
932 * removed while holding both q and blkcg locks. As blkcg lock is nested
933 * inside q lock, this function performs reverse double lock dancing.
934 *
935 * This is the blkcg counterpart of ioc_release_fn().
936 */
937static void blkcg_css_offline(struct cgroup_subsys_state *css)
938{
939 struct blkcg *blkcg = css_to_blkcg(css);
940
941 spin_lock_irq(&blkcg->lock);
942
943 while (!hlist_empty(&blkcg->blkg_list)) {
944 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
945 struct blkcg_gq, blkcg_node);
946 struct request_queue *q = blkg->q;
947
948 if (spin_trylock(q->queue_lock)) {
949 blkg_destroy(blkg);
950 spin_unlock(q->queue_lock);
951 } else {
952 spin_unlock_irq(&blkcg->lock);
953 cpu_relax();
954 spin_lock_irq(&blkcg->lock);
955 }
956 }
957
958 spin_unlock_irq(&blkcg->lock);
959
960 wb_blkcg_offline(blkcg);
961}
962
963static void blkcg_css_free(struct cgroup_subsys_state *css)
964{
965 struct blkcg *blkcg = css_to_blkcg(css);
966 int i;
967
968 mutex_lock(&blkcg_pol_mutex);
969
970 list_del(&blkcg->all_blkcgs_node);
971
972 for (i = 0; i < BLKCG_MAX_POLS; i++)
973 if (blkcg->cpd[i])
974 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
975
976 mutex_unlock(&blkcg_pol_mutex);
977
978 kfree(blkcg);
979}
980
981static struct cgroup_subsys_state *
982blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
983{
984 struct blkcg *blkcg;
985 struct cgroup_subsys_state *ret;
986 int i;
987
988 mutex_lock(&blkcg_pol_mutex);
989
990 if (!parent_css) {
991 blkcg = &blkcg_root;
992 } else {
993 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
994 if (!blkcg) {
995 ret = ERR_PTR(-ENOMEM);
996 goto free_blkcg;
997 }
998 }
999
1000 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1001 struct blkcg_policy *pol = blkcg_policy[i];
1002 struct blkcg_policy_data *cpd;
1003
1004 /*
1005 * If the policy hasn't been attached yet, wait for it
1006 * to be attached before doing anything else. Otherwise,
1007 * check if the policy requires any specific per-cgroup
1008 * data: if it does, allocate and initialize it.
1009 */
1010 if (!pol || !pol->cpd_alloc_fn)
1011 continue;
1012
1013 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1014 if (!cpd) {
1015 ret = ERR_PTR(-ENOMEM);
1016 goto free_pd_blkcg;
1017 }
1018 blkcg->cpd[i] = cpd;
1019 cpd->blkcg = blkcg;
1020 cpd->plid = i;
1021 if (pol->cpd_init_fn)
1022 pol->cpd_init_fn(cpd);
1023 }
1024
1025 spin_lock_init(&blkcg->lock);
1026 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1027 INIT_HLIST_HEAD(&blkcg->blkg_list);
1028#ifdef CONFIG_CGROUP_WRITEBACK
1029 INIT_LIST_HEAD(&blkcg->cgwb_list);
1030#endif
1031 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1032
1033 mutex_unlock(&blkcg_pol_mutex);
1034 return &blkcg->css;
1035
1036free_pd_blkcg:
1037 for (i--; i >= 0; i--)
1038 if (blkcg->cpd[i])
1039 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1040free_blkcg:
1041 kfree(blkcg);
1042 mutex_unlock(&blkcg_pol_mutex);
1043 return ret;
1044}
1045
1046/**
1047 * blkcg_init_queue - initialize blkcg part of request queue
1048 * @q: request_queue to initialize
1049 *
1050 * Called from blk_alloc_queue_node(). Responsible for initializing blkcg
1051 * part of new request_queue @q.
1052 *
1053 * RETURNS:
1054 * 0 on success, -errno on failure.
1055 */
1056int blkcg_init_queue(struct request_queue *q)
1057{
1058 struct blkcg_gq *new_blkg, *blkg;
1059 bool preloaded;
1060 int ret;
1061
1062 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1063 if (!new_blkg)
1064 return -ENOMEM;
1065
1066 preloaded = !radix_tree_preload(GFP_KERNEL);
1067
1068 /*
1069 * Make sure the root blkg exists and count the existing blkgs. As
1070 * @q is bypassing at this point, blkg_lookup_create() can't be
1071 * used. Open code insertion.
1072 */
1073 rcu_read_lock();
1074 spin_lock_irq(q->queue_lock);
1075 blkg = blkg_create(&blkcg_root, q, new_blkg);
1076 spin_unlock_irq(q->queue_lock);
1077 rcu_read_unlock();
1078
1079 if (preloaded)
1080 radix_tree_preload_end();
1081
1082 if (IS_ERR(blkg)) {
1083 blkg_free(new_blkg);
1084 return PTR_ERR(blkg);
1085 }
1086
1087 q->root_blkg = blkg;
1088 q->root_rl.blkg = blkg;
1089
1090 ret = blk_throtl_init(q);
1091 if (ret) {
1092 spin_lock_irq(q->queue_lock);
1093 blkg_destroy_all(q);
1094 spin_unlock_irq(q->queue_lock);
1095 }
1096 return ret;
1097}
1098
1099/**
1100 * blkcg_drain_queue - drain blkcg part of request_queue
1101 * @q: request_queue to drain
1102 *
1103 * Called from blk_drain_queue(). Responsible for draining blkcg part.
1104 */
1105void blkcg_drain_queue(struct request_queue *q)
1106{
1107 lockdep_assert_held(q->queue_lock);
1108
1109 /*
1110 * @q could be exiting and already have destroyed all blkgs as
1111 * indicated by NULL root_blkg. If so, don't confuse policies.
1112 */
1113 if (!q->root_blkg)
1114 return;
1115
1116 blk_throtl_drain(q);
1117}
1118
1119/**
1120 * blkcg_exit_queue - exit and release blkcg part of request_queue
1121 * @q: request_queue being released
1122 *
1123 * Called from blk_release_queue(). Responsible for exiting blkcg part.
1124 */
1125void blkcg_exit_queue(struct request_queue *q)
1126{
1127 spin_lock_irq(q->queue_lock);
1128 blkg_destroy_all(q);
1129 spin_unlock_irq(q->queue_lock);
1130
1131 blk_throtl_exit(q);
1132}
1133
1134/*
1135 * We cannot support shared io contexts, as we have no mean to support
1136 * two tasks with the same ioc in two different groups without major rework
1137 * of the main cic data structures. For now we allow a task to change
1138 * its cgroup only if it's the only owner of its ioc.
1139 */
1140static int blkcg_can_attach(struct cgroup_taskset *tset)
1141{
1142 struct task_struct *task;
1143 struct cgroup_subsys_state *dst_css;
1144 struct io_context *ioc;
1145 int ret = 0;
1146
1147 /* task_lock() is needed to avoid races with exit_io_context() */
1148 cgroup_taskset_for_each(task, dst_css, tset) {
1149 task_lock(task);
1150 ioc = task->io_context;
1151 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
1152 ret = -EINVAL;
1153 task_unlock(task);
1154 if (ret)
1155 break;
1156 }
1157 return ret;
1158}
1159
1160static void blkcg_bind(struct cgroup_subsys_state *root_css)
1161{
1162 int i;
1163
1164 mutex_lock(&blkcg_pol_mutex);
1165
1166 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1167 struct blkcg_policy *pol = blkcg_policy[i];
1168 struct blkcg *blkcg;
1169
1170 if (!pol || !pol->cpd_bind_fn)
1171 continue;
1172
1173 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1174 if (blkcg->cpd[pol->plid])
1175 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1176 }
1177 mutex_unlock(&blkcg_pol_mutex);
1178}
1179
1180struct cgroup_subsys io_cgrp_subsys = {
1181 .css_alloc = blkcg_css_alloc,
1182 .css_offline = blkcg_css_offline,
1183 .css_free = blkcg_css_free,
1184 .can_attach = blkcg_can_attach,
1185 .bind = blkcg_bind,
1186 .dfl_cftypes = blkcg_files,
1187 .legacy_cftypes = blkcg_legacy_files,
1188 .legacy_name = "blkio",
1189#ifdef CONFIG_MEMCG
1190 /*
1191 * This ensures that, if available, memcg is automatically enabled
1192 * together on the default hierarchy so that the owner cgroup can
1193 * be retrieved from writeback pages.
1194 */
1195 .depends_on = 1 << memory_cgrp_id,
1196#endif
1197};
1198EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1199
1200/**
1201 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1202 * @q: request_queue of interest
1203 * @pol: blkcg policy to activate
1204 *
1205 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1206 * bypass mode to populate its blkgs with policy_data for @pol.
1207 *
1208 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1209 * from IO path. Update of each blkg is protected by both queue and blkcg
1210 * locks so that holding either lock and testing blkcg_policy_enabled() is
1211 * always enough for dereferencing policy data.
1212 *
1213 * The caller is responsible for synchronizing [de]activations and policy
1214 * [un]registerations. Returns 0 on success, -errno on failure.
1215 */
1216int blkcg_activate_policy(struct request_queue *q,
1217 const struct blkcg_policy *pol)
1218{
1219 struct blkg_policy_data *pd_prealloc = NULL;
1220 struct blkcg_gq *blkg;
1221 int ret;
1222
1223 if (blkcg_policy_enabled(q, pol))
1224 return 0;
1225
1226 blk_queue_bypass_start(q);
1227pd_prealloc:
1228 if (!pd_prealloc) {
1229 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q->node);
1230 if (!pd_prealloc) {
1231 ret = -ENOMEM;
1232 goto out_bypass_end;
1233 }
1234 }
1235
1236 spin_lock_irq(q->queue_lock);
1237
1238 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1239 struct blkg_policy_data *pd;
1240
1241 if (blkg->pd[pol->plid])
1242 continue;
1243
1244 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q->node);
1245 if (!pd)
1246 swap(pd, pd_prealloc);
1247 if (!pd) {
1248 spin_unlock_irq(q->queue_lock);
1249 goto pd_prealloc;
1250 }
1251
1252 blkg->pd[pol->plid] = pd;
1253 pd->blkg = blkg;
1254 pd->plid = pol->plid;
1255 if (pol->pd_init_fn)
1256 pol->pd_init_fn(pd);
1257 }
1258
1259 __set_bit(pol->plid, q->blkcg_pols);
1260 ret = 0;
1261
1262 spin_unlock_irq(q->queue_lock);
1263out_bypass_end:
1264 blk_queue_bypass_end(q);
1265 if (pd_prealloc)
1266 pol->pd_free_fn(pd_prealloc);
1267 return ret;
1268}
1269EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1270
1271/**
1272 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1273 * @q: request_queue of interest
1274 * @pol: blkcg policy to deactivate
1275 *
1276 * Deactivate @pol on @q. Follows the same synchronization rules as
1277 * blkcg_activate_policy().
1278 */
1279void blkcg_deactivate_policy(struct request_queue *q,
1280 const struct blkcg_policy *pol)
1281{
1282 struct blkcg_gq *blkg;
1283
1284 if (!blkcg_policy_enabled(q, pol))
1285 return;
1286
1287 blk_queue_bypass_start(q);
1288 spin_lock_irq(q->queue_lock);
1289
1290 __clear_bit(pol->plid, q->blkcg_pols);
1291
1292 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1293 /* grab blkcg lock too while removing @pd from @blkg */
1294 spin_lock(&blkg->blkcg->lock);
1295
1296 if (blkg->pd[pol->plid]) {
1297 if (pol->pd_offline_fn)
1298 pol->pd_offline_fn(blkg->pd[pol->plid]);
1299 pol->pd_free_fn(blkg->pd[pol->plid]);
1300 blkg->pd[pol->plid] = NULL;
1301 }
1302
1303 spin_unlock(&blkg->blkcg->lock);
1304 }
1305
1306 spin_unlock_irq(q->queue_lock);
1307 blk_queue_bypass_end(q);
1308}
1309EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1310
1311/**
1312 * blkcg_policy_register - register a blkcg policy
1313 * @pol: blkcg policy to register
1314 *
1315 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1316 * successful registration. Returns 0 on success and -errno on failure.
1317 */
1318int blkcg_policy_register(struct blkcg_policy *pol)
1319{
1320 struct blkcg *blkcg;
1321 int i, ret;
1322
1323 mutex_lock(&blkcg_pol_register_mutex);
1324 mutex_lock(&blkcg_pol_mutex);
1325
1326 /* find an empty slot */
1327 ret = -ENOSPC;
1328 for (i = 0; i < BLKCG_MAX_POLS; i++)
1329 if (!blkcg_policy[i])
1330 break;
1331 if (i >= BLKCG_MAX_POLS)
1332 goto err_unlock;
1333
1334 /* register @pol */
1335 pol->plid = i;
1336 blkcg_policy[pol->plid] = pol;
1337
1338 /* allocate and install cpd's */
1339 if (pol->cpd_alloc_fn) {
1340 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1341 struct blkcg_policy_data *cpd;
1342
1343 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1344 if (!cpd)
1345 goto err_free_cpds;
1346
1347 blkcg->cpd[pol->plid] = cpd;
1348 cpd->blkcg = blkcg;
1349 cpd->plid = pol->plid;
1350 pol->cpd_init_fn(cpd);
1351 }
1352 }
1353
1354 mutex_unlock(&blkcg_pol_mutex);
1355
1356 /* everything is in place, add intf files for the new policy */
1357 if (pol->dfl_cftypes)
1358 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1359 pol->dfl_cftypes));
1360 if (pol->legacy_cftypes)
1361 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1362 pol->legacy_cftypes));
1363 mutex_unlock(&blkcg_pol_register_mutex);
1364 return 0;
1365
1366err_free_cpds:
1367 if (pol->cpd_alloc_fn) {
1368 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1369 if (blkcg->cpd[pol->plid]) {
1370 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1371 blkcg->cpd[pol->plid] = NULL;
1372 }
1373 }
1374 }
1375 blkcg_policy[pol->plid] = NULL;
1376err_unlock:
1377 mutex_unlock(&blkcg_pol_mutex);
1378 mutex_unlock(&blkcg_pol_register_mutex);
1379 return ret;
1380}
1381EXPORT_SYMBOL_GPL(blkcg_policy_register);
1382
1383/**
1384 * blkcg_policy_unregister - unregister a blkcg policy
1385 * @pol: blkcg policy to unregister
1386 *
1387 * Undo blkcg_policy_register(@pol). Might sleep.
1388 */
1389void blkcg_policy_unregister(struct blkcg_policy *pol)
1390{
1391 struct blkcg *blkcg;
1392
1393 mutex_lock(&blkcg_pol_register_mutex);
1394
1395 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1396 goto out_unlock;
1397
1398 /* kill the intf files first */
1399 if (pol->dfl_cftypes)
1400 cgroup_rm_cftypes(pol->dfl_cftypes);
1401 if (pol->legacy_cftypes)
1402 cgroup_rm_cftypes(pol->legacy_cftypes);
1403
1404 /* remove cpds and unregister */
1405 mutex_lock(&blkcg_pol_mutex);
1406
1407 if (pol->cpd_alloc_fn) {
1408 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1409 if (blkcg->cpd[pol->plid]) {
1410 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1411 blkcg->cpd[pol->plid] = NULL;
1412 }
1413 }
1414 }
1415 blkcg_policy[pol->plid] = NULL;
1416
1417 mutex_unlock(&blkcg_pol_mutex);
1418out_unlock:
1419 mutex_unlock(&blkcg_pol_register_mutex);
1420}
1421EXPORT_SYMBOL_GPL(blkcg_policy_unregister);