Loading...
1/*
2 * Common Block IO controller cgroup interface
3 *
4 * Based on ideas and code from CFQ, CFS and BFQ:
5 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
6 *
7 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
8 * Paolo Valente <paolo.valente@unimore.it>
9 *
10 * Copyright (C) 2009 Vivek Goyal <vgoyal@redhat.com>
11 * Nauman Rafique <nauman@google.com>
12 *
13 * For policy-specific per-blkcg data:
14 * Copyright (C) 2015 Paolo Valente <paolo.valente@unimore.it>
15 * Arianna Avanzini <avanzini.arianna@gmail.com>
16 */
17#include <linux/ioprio.h>
18#include <linux/kdev_t.h>
19#include <linux/module.h>
20#include <linux/err.h>
21#include <linux/blkdev.h>
22#include <linux/backing-dev.h>
23#include <linux/slab.h>
24#include <linux/genhd.h>
25#include <linux/delay.h>
26#include <linux/atomic.h>
27#include <linux/ctype.h>
28#include <linux/blk-cgroup.h>
29#include "blk.h"
30
31#define MAX_KEY_LEN 100
32
33/*
34 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
35 * blkcg_pol_register_mutex nests outside of it and synchronizes entire
36 * policy [un]register operations including cgroup file additions /
37 * removals. Putting cgroup file registration outside blkcg_pol_mutex
38 * allows grabbing it from cgroup callbacks.
39 */
40static DEFINE_MUTEX(blkcg_pol_register_mutex);
41static DEFINE_MUTEX(blkcg_pol_mutex);
42
43struct blkcg blkcg_root;
44EXPORT_SYMBOL_GPL(blkcg_root);
45
46struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
47
48static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
49
50static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */
51
52static bool blkcg_policy_enabled(struct request_queue *q,
53 const struct blkcg_policy *pol)
54{
55 return pol && test_bit(pol->plid, q->blkcg_pols);
56}
57
58/**
59 * blkg_free - free a blkg
60 * @blkg: blkg to free
61 *
62 * Free @blkg which may be partially allocated.
63 */
64static void blkg_free(struct blkcg_gq *blkg)
65{
66 int i;
67
68 if (!blkg)
69 return;
70
71 for (i = 0; i < BLKCG_MAX_POLS; i++)
72 if (blkg->pd[i])
73 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
74
75 if (blkg->blkcg != &blkcg_root)
76 blk_exit_rl(&blkg->rl);
77
78 blkg_rwstat_exit(&blkg->stat_ios);
79 blkg_rwstat_exit(&blkg->stat_bytes);
80 kfree(blkg);
81}
82
83/**
84 * blkg_alloc - allocate a blkg
85 * @blkcg: block cgroup the new blkg is associated with
86 * @q: request_queue the new blkg is associated with
87 * @gfp_mask: allocation mask to use
88 *
89 * Allocate a new blkg assocating @blkcg and @q.
90 */
91static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
92 gfp_t gfp_mask)
93{
94 struct blkcg_gq *blkg;
95 int i;
96
97 /* alloc and init base part */
98 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
99 if (!blkg)
100 return NULL;
101
102 if (blkg_rwstat_init(&blkg->stat_bytes, gfp_mask) ||
103 blkg_rwstat_init(&blkg->stat_ios, gfp_mask))
104 goto err_free;
105
106 blkg->q = q;
107 INIT_LIST_HEAD(&blkg->q_node);
108 blkg->blkcg = blkcg;
109 atomic_set(&blkg->refcnt, 1);
110
111 /* root blkg uses @q->root_rl, init rl only for !root blkgs */
112 if (blkcg != &blkcg_root) {
113 if (blk_init_rl(&blkg->rl, q, gfp_mask))
114 goto err_free;
115 blkg->rl.blkg = blkg;
116 }
117
118 for (i = 0; i < BLKCG_MAX_POLS; i++) {
119 struct blkcg_policy *pol = blkcg_policy[i];
120 struct blkg_policy_data *pd;
121
122 if (!blkcg_policy_enabled(q, pol))
123 continue;
124
125 /* alloc per-policy data and attach it to blkg */
126 pd = pol->pd_alloc_fn(gfp_mask, q->node);
127 if (!pd)
128 goto err_free;
129
130 blkg->pd[i] = pd;
131 pd->blkg = blkg;
132 pd->plid = i;
133 }
134
135 return blkg;
136
137err_free:
138 blkg_free(blkg);
139 return NULL;
140}
141
142struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg,
143 struct request_queue *q, bool update_hint)
144{
145 struct blkcg_gq *blkg;
146
147 /*
148 * Hint didn't match. Look up from the radix tree. Note that the
149 * hint can only be updated under queue_lock as otherwise @blkg
150 * could have already been removed from blkg_tree. The caller is
151 * responsible for grabbing queue_lock if @update_hint.
152 */
153 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
154 if (blkg && blkg->q == q) {
155 if (update_hint) {
156 lockdep_assert_held(q->queue_lock);
157 rcu_assign_pointer(blkcg->blkg_hint, blkg);
158 }
159 return blkg;
160 }
161
162 return NULL;
163}
164EXPORT_SYMBOL_GPL(blkg_lookup_slowpath);
165
166/*
167 * If @new_blkg is %NULL, this function tries to allocate a new one as
168 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
169 */
170static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
171 struct request_queue *q,
172 struct blkcg_gq *new_blkg)
173{
174 struct blkcg_gq *blkg;
175 struct bdi_writeback_congested *wb_congested;
176 int i, ret;
177
178 WARN_ON_ONCE(!rcu_read_lock_held());
179 lockdep_assert_held(q->queue_lock);
180
181 /* blkg holds a reference to blkcg */
182 if (!css_tryget_online(&blkcg->css)) {
183 ret = -ENODEV;
184 goto err_free_blkg;
185 }
186
187 wb_congested = wb_congested_get_create(&q->backing_dev_info,
188 blkcg->css.id, GFP_NOWAIT);
189 if (!wb_congested) {
190 ret = -ENOMEM;
191 goto err_put_css;
192 }
193
194 /* allocate */
195 if (!new_blkg) {
196 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT);
197 if (unlikely(!new_blkg)) {
198 ret = -ENOMEM;
199 goto err_put_congested;
200 }
201 }
202 blkg = new_blkg;
203 blkg->wb_congested = wb_congested;
204
205 /* link parent */
206 if (blkcg_parent(blkcg)) {
207 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
208 if (WARN_ON_ONCE(!blkg->parent)) {
209 ret = -ENODEV;
210 goto err_put_congested;
211 }
212 blkg_get(blkg->parent);
213 }
214
215 /* invoke per-policy init */
216 for (i = 0; i < BLKCG_MAX_POLS; i++) {
217 struct blkcg_policy *pol = blkcg_policy[i];
218
219 if (blkg->pd[i] && pol->pd_init_fn)
220 pol->pd_init_fn(blkg->pd[i]);
221 }
222
223 /* insert */
224 spin_lock(&blkcg->lock);
225 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
226 if (likely(!ret)) {
227 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
228 list_add(&blkg->q_node, &q->blkg_list);
229
230 for (i = 0; i < BLKCG_MAX_POLS; i++) {
231 struct blkcg_policy *pol = blkcg_policy[i];
232
233 if (blkg->pd[i] && pol->pd_online_fn)
234 pol->pd_online_fn(blkg->pd[i]);
235 }
236 }
237 blkg->online = true;
238 spin_unlock(&blkcg->lock);
239
240 if (!ret)
241 return blkg;
242
243 /* @blkg failed fully initialized, use the usual release path */
244 blkg_put(blkg);
245 return ERR_PTR(ret);
246
247err_put_congested:
248 wb_congested_put(wb_congested);
249err_put_css:
250 css_put(&blkcg->css);
251err_free_blkg:
252 blkg_free(new_blkg);
253 return ERR_PTR(ret);
254}
255
256/**
257 * blkg_lookup_create - lookup blkg, try to create one if not there
258 * @blkcg: blkcg of interest
259 * @q: request_queue of interest
260 *
261 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
262 * create one. blkg creation is performed recursively from blkcg_root such
263 * that all non-root blkg's have access to the parent blkg. This function
264 * should be called under RCU read lock and @q->queue_lock.
265 *
266 * Returns pointer to the looked up or created blkg on success, ERR_PTR()
267 * value on error. If @q is dead, returns ERR_PTR(-EINVAL). If @q is not
268 * dead and bypassing, returns ERR_PTR(-EBUSY).
269 */
270struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
271 struct request_queue *q)
272{
273 struct blkcg_gq *blkg;
274
275 WARN_ON_ONCE(!rcu_read_lock_held());
276 lockdep_assert_held(q->queue_lock);
277
278 /*
279 * This could be the first entry point of blkcg implementation and
280 * we shouldn't allow anything to go through for a bypassing queue.
281 */
282 if (unlikely(blk_queue_bypass(q)))
283 return ERR_PTR(blk_queue_dying(q) ? -ENODEV : -EBUSY);
284
285 blkg = __blkg_lookup(blkcg, q, true);
286 if (blkg)
287 return blkg;
288
289 /*
290 * Create blkgs walking down from blkcg_root to @blkcg, so that all
291 * non-root blkgs have access to their parents.
292 */
293 while (true) {
294 struct blkcg *pos = blkcg;
295 struct blkcg *parent = blkcg_parent(blkcg);
296
297 while (parent && !__blkg_lookup(parent, q, false)) {
298 pos = parent;
299 parent = blkcg_parent(parent);
300 }
301
302 blkg = blkg_create(pos, q, NULL);
303 if (pos == blkcg || IS_ERR(blkg))
304 return blkg;
305 }
306}
307
308static void blkg_destroy(struct blkcg_gq *blkg)
309{
310 struct blkcg *blkcg = blkg->blkcg;
311 struct blkcg_gq *parent = blkg->parent;
312 int i;
313
314 lockdep_assert_held(blkg->q->queue_lock);
315 lockdep_assert_held(&blkcg->lock);
316
317 /* Something wrong if we are trying to remove same group twice */
318 WARN_ON_ONCE(list_empty(&blkg->q_node));
319 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
320
321 for (i = 0; i < BLKCG_MAX_POLS; i++) {
322 struct blkcg_policy *pol = blkcg_policy[i];
323
324 if (blkg->pd[i] && pol->pd_offline_fn)
325 pol->pd_offline_fn(blkg->pd[i]);
326 }
327
328 if (parent) {
329 blkg_rwstat_add_aux(&parent->stat_bytes, &blkg->stat_bytes);
330 blkg_rwstat_add_aux(&parent->stat_ios, &blkg->stat_ios);
331 }
332
333 blkg->online = false;
334
335 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
336 list_del_init(&blkg->q_node);
337 hlist_del_init_rcu(&blkg->blkcg_node);
338
339 /*
340 * Both setting lookup hint to and clearing it from @blkg are done
341 * under queue_lock. If it's not pointing to @blkg now, it never
342 * will. Hint assignment itself can race safely.
343 */
344 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
345 rcu_assign_pointer(blkcg->blkg_hint, NULL);
346
347 /*
348 * Put the reference taken at the time of creation so that when all
349 * queues are gone, group can be destroyed.
350 */
351 blkg_put(blkg);
352}
353
354/**
355 * blkg_destroy_all - destroy all blkgs associated with a request_queue
356 * @q: request_queue of interest
357 *
358 * Destroy all blkgs associated with @q.
359 */
360static void blkg_destroy_all(struct request_queue *q)
361{
362 struct blkcg_gq *blkg, *n;
363
364 lockdep_assert_held(q->queue_lock);
365
366 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
367 struct blkcg *blkcg = blkg->blkcg;
368
369 spin_lock(&blkcg->lock);
370 blkg_destroy(blkg);
371 spin_unlock(&blkcg->lock);
372 }
373
374 q->root_blkg = NULL;
375 q->root_rl.blkg = NULL;
376}
377
378/*
379 * A group is RCU protected, but having an rcu lock does not mean that one
380 * can access all the fields of blkg and assume these are valid. For
381 * example, don't try to follow throtl_data and request queue links.
382 *
383 * Having a reference to blkg under an rcu allows accesses to only values
384 * local to groups like group stats and group rate limits.
385 */
386void __blkg_release_rcu(struct rcu_head *rcu_head)
387{
388 struct blkcg_gq *blkg = container_of(rcu_head, struct blkcg_gq, rcu_head);
389
390 /* release the blkcg and parent blkg refs this blkg has been holding */
391 css_put(&blkg->blkcg->css);
392 if (blkg->parent)
393 blkg_put(blkg->parent);
394
395 wb_congested_put(blkg->wb_congested);
396
397 blkg_free(blkg);
398}
399EXPORT_SYMBOL_GPL(__blkg_release_rcu);
400
401/*
402 * The next function used by blk_queue_for_each_rl(). It's a bit tricky
403 * because the root blkg uses @q->root_rl instead of its own rl.
404 */
405struct request_list *__blk_queue_next_rl(struct request_list *rl,
406 struct request_queue *q)
407{
408 struct list_head *ent;
409 struct blkcg_gq *blkg;
410
411 /*
412 * Determine the current blkg list_head. The first entry is
413 * root_rl which is off @q->blkg_list and mapped to the head.
414 */
415 if (rl == &q->root_rl) {
416 ent = &q->blkg_list;
417 /* There are no more block groups, hence no request lists */
418 if (list_empty(ent))
419 return NULL;
420 } else {
421 blkg = container_of(rl, struct blkcg_gq, rl);
422 ent = &blkg->q_node;
423 }
424
425 /* walk to the next list_head, skip root blkcg */
426 ent = ent->next;
427 if (ent == &q->root_blkg->q_node)
428 ent = ent->next;
429 if (ent == &q->blkg_list)
430 return NULL;
431
432 blkg = container_of(ent, struct blkcg_gq, q_node);
433 return &blkg->rl;
434}
435
436static int blkcg_reset_stats(struct cgroup_subsys_state *css,
437 struct cftype *cftype, u64 val)
438{
439 struct blkcg *blkcg = css_to_blkcg(css);
440 struct blkcg_gq *blkg;
441 int i;
442
443 mutex_lock(&blkcg_pol_mutex);
444 spin_lock_irq(&blkcg->lock);
445
446 /*
447 * Note that stat reset is racy - it doesn't synchronize against
448 * stat updates. This is a debug feature which shouldn't exist
449 * anyway. If you get hit by a race, retry.
450 */
451 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
452 blkg_rwstat_reset(&blkg->stat_bytes);
453 blkg_rwstat_reset(&blkg->stat_ios);
454
455 for (i = 0; i < BLKCG_MAX_POLS; i++) {
456 struct blkcg_policy *pol = blkcg_policy[i];
457
458 if (blkg->pd[i] && pol->pd_reset_stats_fn)
459 pol->pd_reset_stats_fn(blkg->pd[i]);
460 }
461 }
462
463 spin_unlock_irq(&blkcg->lock);
464 mutex_unlock(&blkcg_pol_mutex);
465 return 0;
466}
467
468const char *blkg_dev_name(struct blkcg_gq *blkg)
469{
470 /* some drivers (floppy) instantiate a queue w/o disk registered */
471 if (blkg->q->backing_dev_info.dev)
472 return dev_name(blkg->q->backing_dev_info.dev);
473 return NULL;
474}
475EXPORT_SYMBOL_GPL(blkg_dev_name);
476
477/**
478 * blkcg_print_blkgs - helper for printing per-blkg data
479 * @sf: seq_file to print to
480 * @blkcg: blkcg of interest
481 * @prfill: fill function to print out a blkg
482 * @pol: policy in question
483 * @data: data to be passed to @prfill
484 * @show_total: to print out sum of prfill return values or not
485 *
486 * This function invokes @prfill on each blkg of @blkcg if pd for the
487 * policy specified by @pol exists. @prfill is invoked with @sf, the
488 * policy data and @data and the matching queue lock held. If @show_total
489 * is %true, the sum of the return values from @prfill is printed with
490 * "Total" label at the end.
491 *
492 * This is to be used to construct print functions for
493 * cftype->read_seq_string method.
494 */
495void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
496 u64 (*prfill)(struct seq_file *,
497 struct blkg_policy_data *, int),
498 const struct blkcg_policy *pol, int data,
499 bool show_total)
500{
501 struct blkcg_gq *blkg;
502 u64 total = 0;
503
504 rcu_read_lock();
505 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
506 spin_lock_irq(blkg->q->queue_lock);
507 if (blkcg_policy_enabled(blkg->q, pol))
508 total += prfill(sf, blkg->pd[pol->plid], data);
509 spin_unlock_irq(blkg->q->queue_lock);
510 }
511 rcu_read_unlock();
512
513 if (show_total)
514 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
515}
516EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
517
518/**
519 * __blkg_prfill_u64 - prfill helper for a single u64 value
520 * @sf: seq_file to print to
521 * @pd: policy private data of interest
522 * @v: value to print
523 *
524 * Print @v to @sf for the device assocaited with @pd.
525 */
526u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
527{
528 const char *dname = blkg_dev_name(pd->blkg);
529
530 if (!dname)
531 return 0;
532
533 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
534 return v;
535}
536EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
537
538/**
539 * __blkg_prfill_rwstat - prfill helper for a blkg_rwstat
540 * @sf: seq_file to print to
541 * @pd: policy private data of interest
542 * @rwstat: rwstat to print
543 *
544 * Print @rwstat to @sf for the device assocaited with @pd.
545 */
546u64 __blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
547 const struct blkg_rwstat *rwstat)
548{
549 static const char *rwstr[] = {
550 [BLKG_RWSTAT_READ] = "Read",
551 [BLKG_RWSTAT_WRITE] = "Write",
552 [BLKG_RWSTAT_SYNC] = "Sync",
553 [BLKG_RWSTAT_ASYNC] = "Async",
554 };
555 const char *dname = blkg_dev_name(pd->blkg);
556 u64 v;
557 int i;
558
559 if (!dname)
560 return 0;
561
562 for (i = 0; i < BLKG_RWSTAT_NR; i++)
563 seq_printf(sf, "%s %s %llu\n", dname, rwstr[i],
564 (unsigned long long)atomic64_read(&rwstat->aux_cnt[i]));
565
566 v = atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_READ]) +
567 atomic64_read(&rwstat->aux_cnt[BLKG_RWSTAT_WRITE]);
568 seq_printf(sf, "%s Total %llu\n", dname, (unsigned long long)v);
569 return v;
570}
571EXPORT_SYMBOL_GPL(__blkg_prfill_rwstat);
572
573/**
574 * blkg_prfill_stat - prfill callback for blkg_stat
575 * @sf: seq_file to print to
576 * @pd: policy private data of interest
577 * @off: offset to the blkg_stat in @pd
578 *
579 * prfill callback for printing a blkg_stat.
580 */
581u64 blkg_prfill_stat(struct seq_file *sf, struct blkg_policy_data *pd, int off)
582{
583 return __blkg_prfill_u64(sf, pd, blkg_stat_read((void *)pd + off));
584}
585EXPORT_SYMBOL_GPL(blkg_prfill_stat);
586
587/**
588 * blkg_prfill_rwstat - prfill callback for blkg_rwstat
589 * @sf: seq_file to print to
590 * @pd: policy private data of interest
591 * @off: offset to the blkg_rwstat in @pd
592 *
593 * prfill callback for printing a blkg_rwstat.
594 */
595u64 blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
596 int off)
597{
598 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd + off);
599
600 return __blkg_prfill_rwstat(sf, pd, &rwstat);
601}
602EXPORT_SYMBOL_GPL(blkg_prfill_rwstat);
603
604static u64 blkg_prfill_rwstat_field(struct seq_file *sf,
605 struct blkg_policy_data *pd, int off)
606{
607 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd->blkg + off);
608
609 return __blkg_prfill_rwstat(sf, pd, &rwstat);
610}
611
612/**
613 * blkg_print_stat_bytes - seq_show callback for blkg->stat_bytes
614 * @sf: seq_file to print to
615 * @v: unused
616 *
617 * To be used as cftype->seq_show to print blkg->stat_bytes.
618 * cftype->private must be set to the blkcg_policy.
619 */
620int blkg_print_stat_bytes(struct seq_file *sf, void *v)
621{
622 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
623 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
624 offsetof(struct blkcg_gq, stat_bytes), true);
625 return 0;
626}
627EXPORT_SYMBOL_GPL(blkg_print_stat_bytes);
628
629/**
630 * blkg_print_stat_bytes - seq_show callback for blkg->stat_ios
631 * @sf: seq_file to print to
632 * @v: unused
633 *
634 * To be used as cftype->seq_show to print blkg->stat_ios. cftype->private
635 * must be set to the blkcg_policy.
636 */
637int blkg_print_stat_ios(struct seq_file *sf, void *v)
638{
639 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
640 blkg_prfill_rwstat_field, (void *)seq_cft(sf)->private,
641 offsetof(struct blkcg_gq, stat_ios), true);
642 return 0;
643}
644EXPORT_SYMBOL_GPL(blkg_print_stat_ios);
645
646static u64 blkg_prfill_rwstat_field_recursive(struct seq_file *sf,
647 struct blkg_policy_data *pd,
648 int off)
649{
650 struct blkg_rwstat rwstat = blkg_rwstat_recursive_sum(pd->blkg,
651 NULL, off);
652 return __blkg_prfill_rwstat(sf, pd, &rwstat);
653}
654
655/**
656 * blkg_print_stat_bytes_recursive - recursive version of blkg_print_stat_bytes
657 * @sf: seq_file to print to
658 * @v: unused
659 */
660int blkg_print_stat_bytes_recursive(struct seq_file *sf, void *v)
661{
662 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
663 blkg_prfill_rwstat_field_recursive,
664 (void *)seq_cft(sf)->private,
665 offsetof(struct blkcg_gq, stat_bytes), true);
666 return 0;
667}
668EXPORT_SYMBOL_GPL(blkg_print_stat_bytes_recursive);
669
670/**
671 * blkg_print_stat_ios_recursive - recursive version of blkg_print_stat_ios
672 * @sf: seq_file to print to
673 * @v: unused
674 */
675int blkg_print_stat_ios_recursive(struct seq_file *sf, void *v)
676{
677 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
678 blkg_prfill_rwstat_field_recursive,
679 (void *)seq_cft(sf)->private,
680 offsetof(struct blkcg_gq, stat_ios), true);
681 return 0;
682}
683EXPORT_SYMBOL_GPL(blkg_print_stat_ios_recursive);
684
685/**
686 * blkg_stat_recursive_sum - collect hierarchical blkg_stat
687 * @blkg: blkg of interest
688 * @pol: blkcg_policy which contains the blkg_stat
689 * @off: offset to the blkg_stat in blkg_policy_data or @blkg
690 *
691 * Collect the blkg_stat specified by @blkg, @pol and @off and all its
692 * online descendants and their aux counts. The caller must be holding the
693 * queue lock for online tests.
694 *
695 * If @pol is NULL, blkg_stat is at @off bytes into @blkg; otherwise, it is
696 * at @off bytes into @blkg's blkg_policy_data of the policy.
697 */
698u64 blkg_stat_recursive_sum(struct blkcg_gq *blkg,
699 struct blkcg_policy *pol, int off)
700{
701 struct blkcg_gq *pos_blkg;
702 struct cgroup_subsys_state *pos_css;
703 u64 sum = 0;
704
705 lockdep_assert_held(blkg->q->queue_lock);
706
707 rcu_read_lock();
708 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
709 struct blkg_stat *stat;
710
711 if (!pos_blkg->online)
712 continue;
713
714 if (pol)
715 stat = (void *)blkg_to_pd(pos_blkg, pol) + off;
716 else
717 stat = (void *)blkg + off;
718
719 sum += blkg_stat_read(stat) + atomic64_read(&stat->aux_cnt);
720 }
721 rcu_read_unlock();
722
723 return sum;
724}
725EXPORT_SYMBOL_GPL(blkg_stat_recursive_sum);
726
727/**
728 * blkg_rwstat_recursive_sum - collect hierarchical blkg_rwstat
729 * @blkg: blkg of interest
730 * @pol: blkcg_policy which contains the blkg_rwstat
731 * @off: offset to the blkg_rwstat in blkg_policy_data or @blkg
732 *
733 * Collect the blkg_rwstat specified by @blkg, @pol and @off and all its
734 * online descendants and their aux counts. The caller must be holding the
735 * queue lock for online tests.
736 *
737 * If @pol is NULL, blkg_rwstat is at @off bytes into @blkg; otherwise, it
738 * is at @off bytes into @blkg's blkg_policy_data of the policy.
739 */
740struct blkg_rwstat blkg_rwstat_recursive_sum(struct blkcg_gq *blkg,
741 struct blkcg_policy *pol, int off)
742{
743 struct blkcg_gq *pos_blkg;
744 struct cgroup_subsys_state *pos_css;
745 struct blkg_rwstat sum = { };
746 int i;
747
748 lockdep_assert_held(blkg->q->queue_lock);
749
750 rcu_read_lock();
751 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
752 struct blkg_rwstat *rwstat;
753
754 if (!pos_blkg->online)
755 continue;
756
757 if (pol)
758 rwstat = (void *)blkg_to_pd(pos_blkg, pol) + off;
759 else
760 rwstat = (void *)pos_blkg + off;
761
762 for (i = 0; i < BLKG_RWSTAT_NR; i++)
763 atomic64_add(atomic64_read(&rwstat->aux_cnt[i]) +
764 percpu_counter_sum_positive(&rwstat->cpu_cnt[i]),
765 &sum.aux_cnt[i]);
766 }
767 rcu_read_unlock();
768
769 return sum;
770}
771EXPORT_SYMBOL_GPL(blkg_rwstat_recursive_sum);
772
773/**
774 * blkg_conf_prep - parse and prepare for per-blkg config update
775 * @blkcg: target block cgroup
776 * @pol: target policy
777 * @input: input string
778 * @ctx: blkg_conf_ctx to be filled
779 *
780 * Parse per-blkg config update from @input and initialize @ctx with the
781 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
782 * part of @input following MAJ:MIN. This function returns with RCU read
783 * lock and queue lock held and must be paired with blkg_conf_finish().
784 */
785int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
786 char *input, struct blkg_conf_ctx *ctx)
787 __acquires(rcu) __acquires(disk->queue->queue_lock)
788{
789 struct gendisk *disk;
790 struct blkcg_gq *blkg;
791 struct module *owner;
792 unsigned int major, minor;
793 int key_len, part, ret;
794 char *body;
795
796 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
797 return -EINVAL;
798
799 body = input + key_len;
800 if (!isspace(*body))
801 return -EINVAL;
802 body = skip_spaces(body);
803
804 disk = get_gendisk(MKDEV(major, minor), &part);
805 if (!disk)
806 return -ENODEV;
807 if (part) {
808 owner = disk->fops->owner;
809 put_disk(disk);
810 module_put(owner);
811 return -ENODEV;
812 }
813
814 rcu_read_lock();
815 spin_lock_irq(disk->queue->queue_lock);
816
817 if (blkcg_policy_enabled(disk->queue, pol))
818 blkg = blkg_lookup_create(blkcg, disk->queue);
819 else
820 blkg = ERR_PTR(-EOPNOTSUPP);
821
822 if (IS_ERR(blkg)) {
823 ret = PTR_ERR(blkg);
824 rcu_read_unlock();
825 spin_unlock_irq(disk->queue->queue_lock);
826 owner = disk->fops->owner;
827 put_disk(disk);
828 module_put(owner);
829 /*
830 * If queue was bypassing, we should retry. Do so after a
831 * short msleep(). It isn't strictly necessary but queue
832 * can be bypassing for some time and it's always nice to
833 * avoid busy looping.
834 */
835 if (ret == -EBUSY) {
836 msleep(10);
837 ret = restart_syscall();
838 }
839 return ret;
840 }
841
842 ctx->disk = disk;
843 ctx->blkg = blkg;
844 ctx->body = body;
845 return 0;
846}
847EXPORT_SYMBOL_GPL(blkg_conf_prep);
848
849/**
850 * blkg_conf_finish - finish up per-blkg config update
851 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
852 *
853 * Finish up after per-blkg config update. This function must be paired
854 * with blkg_conf_prep().
855 */
856void blkg_conf_finish(struct blkg_conf_ctx *ctx)
857 __releases(ctx->disk->queue->queue_lock) __releases(rcu)
858{
859 struct module *owner;
860
861 spin_unlock_irq(ctx->disk->queue->queue_lock);
862 rcu_read_unlock();
863 owner = ctx->disk->fops->owner;
864 put_disk(ctx->disk);
865 module_put(owner);
866}
867EXPORT_SYMBOL_GPL(blkg_conf_finish);
868
869static int blkcg_print_stat(struct seq_file *sf, void *v)
870{
871 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
872 struct blkcg_gq *blkg;
873
874 rcu_read_lock();
875
876 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
877 const char *dname;
878 struct blkg_rwstat rwstat;
879 u64 rbytes, wbytes, rios, wios;
880
881 dname = blkg_dev_name(blkg);
882 if (!dname)
883 continue;
884
885 spin_lock_irq(blkg->q->queue_lock);
886
887 rwstat = blkg_rwstat_recursive_sum(blkg, NULL,
888 offsetof(struct blkcg_gq, stat_bytes));
889 rbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_READ]);
890 wbytes = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_WRITE]);
891
892 rwstat = blkg_rwstat_recursive_sum(blkg, NULL,
893 offsetof(struct blkcg_gq, stat_ios));
894 rios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_READ]);
895 wios = atomic64_read(&rwstat.aux_cnt[BLKG_RWSTAT_WRITE]);
896
897 spin_unlock_irq(blkg->q->queue_lock);
898
899 if (rbytes || wbytes || rios || wios)
900 seq_printf(sf, "%s rbytes=%llu wbytes=%llu rios=%llu wios=%llu\n",
901 dname, rbytes, wbytes, rios, wios);
902 }
903
904 rcu_read_unlock();
905 return 0;
906}
907
908struct cftype blkcg_files[] = {
909 {
910 .name = "stat",
911 .flags = CFTYPE_NOT_ON_ROOT,
912 .seq_show = blkcg_print_stat,
913 },
914 { } /* terminate */
915};
916
917struct cftype blkcg_legacy_files[] = {
918 {
919 .name = "reset_stats",
920 .write_u64 = blkcg_reset_stats,
921 },
922 { } /* terminate */
923};
924
925/**
926 * blkcg_css_offline - cgroup css_offline callback
927 * @css: css of interest
928 *
929 * This function is called when @css is about to go away and responsible
930 * for shooting down all blkgs associated with @css. blkgs should be
931 * removed while holding both q and blkcg locks. As blkcg lock is nested
932 * inside q lock, this function performs reverse double lock dancing.
933 *
934 * This is the blkcg counterpart of ioc_release_fn().
935 */
936static void blkcg_css_offline(struct cgroup_subsys_state *css)
937{
938 struct blkcg *blkcg = css_to_blkcg(css);
939
940 spin_lock_irq(&blkcg->lock);
941
942 while (!hlist_empty(&blkcg->blkg_list)) {
943 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
944 struct blkcg_gq, blkcg_node);
945 struct request_queue *q = blkg->q;
946
947 if (spin_trylock(q->queue_lock)) {
948 blkg_destroy(blkg);
949 spin_unlock(q->queue_lock);
950 } else {
951 spin_unlock_irq(&blkcg->lock);
952 cpu_relax();
953 spin_lock_irq(&blkcg->lock);
954 }
955 }
956
957 spin_unlock_irq(&blkcg->lock);
958
959 wb_blkcg_offline(blkcg);
960}
961
962static void blkcg_css_free(struct cgroup_subsys_state *css)
963{
964 struct blkcg *blkcg = css_to_blkcg(css);
965 int i;
966
967 mutex_lock(&blkcg_pol_mutex);
968
969 list_del(&blkcg->all_blkcgs_node);
970
971 for (i = 0; i < BLKCG_MAX_POLS; i++)
972 if (blkcg->cpd[i])
973 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
974
975 mutex_unlock(&blkcg_pol_mutex);
976
977 kfree(blkcg);
978}
979
980static struct cgroup_subsys_state *
981blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
982{
983 struct blkcg *blkcg;
984 struct cgroup_subsys_state *ret;
985 int i;
986
987 mutex_lock(&blkcg_pol_mutex);
988
989 if (!parent_css) {
990 blkcg = &blkcg_root;
991 } else {
992 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
993 if (!blkcg) {
994 ret = ERR_PTR(-ENOMEM);
995 goto free_blkcg;
996 }
997 }
998
999 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1000 struct blkcg_policy *pol = blkcg_policy[i];
1001 struct blkcg_policy_data *cpd;
1002
1003 /*
1004 * If the policy hasn't been attached yet, wait for it
1005 * to be attached before doing anything else. Otherwise,
1006 * check if the policy requires any specific per-cgroup
1007 * data: if it does, allocate and initialize it.
1008 */
1009 if (!pol || !pol->cpd_alloc_fn)
1010 continue;
1011
1012 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1013 if (!cpd) {
1014 ret = ERR_PTR(-ENOMEM);
1015 goto free_pd_blkcg;
1016 }
1017 blkcg->cpd[i] = cpd;
1018 cpd->blkcg = blkcg;
1019 cpd->plid = i;
1020 if (pol->cpd_init_fn)
1021 pol->cpd_init_fn(cpd);
1022 }
1023
1024 spin_lock_init(&blkcg->lock);
1025 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT);
1026 INIT_HLIST_HEAD(&blkcg->blkg_list);
1027#ifdef CONFIG_CGROUP_WRITEBACK
1028 INIT_LIST_HEAD(&blkcg->cgwb_list);
1029#endif
1030 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1031
1032 mutex_unlock(&blkcg_pol_mutex);
1033 return &blkcg->css;
1034
1035free_pd_blkcg:
1036 for (i--; i >= 0; i--)
1037 if (blkcg->cpd[i])
1038 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1039free_blkcg:
1040 kfree(blkcg);
1041 mutex_unlock(&blkcg_pol_mutex);
1042 return ret;
1043}
1044
1045/**
1046 * blkcg_init_queue - initialize blkcg part of request queue
1047 * @q: request_queue to initialize
1048 *
1049 * Called from blk_alloc_queue_node(). Responsible for initializing blkcg
1050 * part of new request_queue @q.
1051 *
1052 * RETURNS:
1053 * 0 on success, -errno on failure.
1054 */
1055int blkcg_init_queue(struct request_queue *q)
1056{
1057 struct blkcg_gq *new_blkg, *blkg;
1058 bool preloaded;
1059 int ret;
1060
1061 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1062 if (!new_blkg)
1063 return -ENOMEM;
1064
1065 preloaded = !radix_tree_preload(GFP_KERNEL);
1066
1067 /*
1068 * Make sure the root blkg exists and count the existing blkgs. As
1069 * @q is bypassing at this point, blkg_lookup_create() can't be
1070 * used. Open code insertion.
1071 */
1072 rcu_read_lock();
1073 spin_lock_irq(q->queue_lock);
1074 blkg = blkg_create(&blkcg_root, q, new_blkg);
1075 spin_unlock_irq(q->queue_lock);
1076 rcu_read_unlock();
1077
1078 if (preloaded)
1079 radix_tree_preload_end();
1080
1081 if (IS_ERR(blkg)) {
1082 blkg_free(new_blkg);
1083 return PTR_ERR(blkg);
1084 }
1085
1086 q->root_blkg = blkg;
1087 q->root_rl.blkg = blkg;
1088
1089 ret = blk_throtl_init(q);
1090 if (ret) {
1091 spin_lock_irq(q->queue_lock);
1092 blkg_destroy_all(q);
1093 spin_unlock_irq(q->queue_lock);
1094 }
1095 return ret;
1096}
1097
1098/**
1099 * blkcg_drain_queue - drain blkcg part of request_queue
1100 * @q: request_queue to drain
1101 *
1102 * Called from blk_drain_queue(). Responsible for draining blkcg part.
1103 */
1104void blkcg_drain_queue(struct request_queue *q)
1105{
1106 lockdep_assert_held(q->queue_lock);
1107
1108 /*
1109 * @q could be exiting and already have destroyed all blkgs as
1110 * indicated by NULL root_blkg. If so, don't confuse policies.
1111 */
1112 if (!q->root_blkg)
1113 return;
1114
1115 blk_throtl_drain(q);
1116}
1117
1118/**
1119 * blkcg_exit_queue - exit and release blkcg part of request_queue
1120 * @q: request_queue being released
1121 *
1122 * Called from blk_release_queue(). Responsible for exiting blkcg part.
1123 */
1124void blkcg_exit_queue(struct request_queue *q)
1125{
1126 spin_lock_irq(q->queue_lock);
1127 blkg_destroy_all(q);
1128 spin_unlock_irq(q->queue_lock);
1129
1130 blk_throtl_exit(q);
1131}
1132
1133/*
1134 * We cannot support shared io contexts, as we have no mean to support
1135 * two tasks with the same ioc in two different groups without major rework
1136 * of the main cic data structures. For now we allow a task to change
1137 * its cgroup only if it's the only owner of its ioc.
1138 */
1139static int blkcg_can_attach(struct cgroup_taskset *tset)
1140{
1141 struct task_struct *task;
1142 struct cgroup_subsys_state *dst_css;
1143 struct io_context *ioc;
1144 int ret = 0;
1145
1146 /* task_lock() is needed to avoid races with exit_io_context() */
1147 cgroup_taskset_for_each(task, dst_css, tset) {
1148 task_lock(task);
1149 ioc = task->io_context;
1150 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
1151 ret = -EINVAL;
1152 task_unlock(task);
1153 if (ret)
1154 break;
1155 }
1156 return ret;
1157}
1158
1159static void blkcg_bind(struct cgroup_subsys_state *root_css)
1160{
1161 int i;
1162
1163 mutex_lock(&blkcg_pol_mutex);
1164
1165 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1166 struct blkcg_policy *pol = blkcg_policy[i];
1167 struct blkcg *blkcg;
1168
1169 if (!pol || !pol->cpd_bind_fn)
1170 continue;
1171
1172 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1173 if (blkcg->cpd[pol->plid])
1174 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1175 }
1176 mutex_unlock(&blkcg_pol_mutex);
1177}
1178
1179struct cgroup_subsys io_cgrp_subsys = {
1180 .css_alloc = blkcg_css_alloc,
1181 .css_offline = blkcg_css_offline,
1182 .css_free = blkcg_css_free,
1183 .can_attach = blkcg_can_attach,
1184 .bind = blkcg_bind,
1185 .dfl_cftypes = blkcg_files,
1186 .legacy_cftypes = blkcg_legacy_files,
1187 .legacy_name = "blkio",
1188#ifdef CONFIG_MEMCG
1189 /*
1190 * This ensures that, if available, memcg is automatically enabled
1191 * together on the default hierarchy so that the owner cgroup can
1192 * be retrieved from writeback pages.
1193 */
1194 .depends_on = 1 << memory_cgrp_id,
1195#endif
1196};
1197EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1198
1199/**
1200 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1201 * @q: request_queue of interest
1202 * @pol: blkcg policy to activate
1203 *
1204 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1205 * bypass mode to populate its blkgs with policy_data for @pol.
1206 *
1207 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1208 * from IO path. Update of each blkg is protected by both queue and blkcg
1209 * locks so that holding either lock and testing blkcg_policy_enabled() is
1210 * always enough for dereferencing policy data.
1211 *
1212 * The caller is responsible for synchronizing [de]activations and policy
1213 * [un]registerations. Returns 0 on success, -errno on failure.
1214 */
1215int blkcg_activate_policy(struct request_queue *q,
1216 const struct blkcg_policy *pol)
1217{
1218 struct blkg_policy_data *pd_prealloc = NULL;
1219 struct blkcg_gq *blkg;
1220 int ret;
1221
1222 if (blkcg_policy_enabled(q, pol))
1223 return 0;
1224
1225 blk_queue_bypass_start(q);
1226pd_prealloc:
1227 if (!pd_prealloc) {
1228 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q->node);
1229 if (!pd_prealloc) {
1230 ret = -ENOMEM;
1231 goto out_bypass_end;
1232 }
1233 }
1234
1235 spin_lock_irq(q->queue_lock);
1236
1237 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1238 struct blkg_policy_data *pd;
1239
1240 if (blkg->pd[pol->plid])
1241 continue;
1242
1243 pd = pol->pd_alloc_fn(GFP_NOWAIT, q->node);
1244 if (!pd)
1245 swap(pd, pd_prealloc);
1246 if (!pd) {
1247 spin_unlock_irq(q->queue_lock);
1248 goto pd_prealloc;
1249 }
1250
1251 blkg->pd[pol->plid] = pd;
1252 pd->blkg = blkg;
1253 pd->plid = pol->plid;
1254 if (pol->pd_init_fn)
1255 pol->pd_init_fn(pd);
1256 }
1257
1258 __set_bit(pol->plid, q->blkcg_pols);
1259 ret = 0;
1260
1261 spin_unlock_irq(q->queue_lock);
1262out_bypass_end:
1263 blk_queue_bypass_end(q);
1264 if (pd_prealloc)
1265 pol->pd_free_fn(pd_prealloc);
1266 return ret;
1267}
1268EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1269
1270/**
1271 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1272 * @q: request_queue of interest
1273 * @pol: blkcg policy to deactivate
1274 *
1275 * Deactivate @pol on @q. Follows the same synchronization rules as
1276 * blkcg_activate_policy().
1277 */
1278void blkcg_deactivate_policy(struct request_queue *q,
1279 const struct blkcg_policy *pol)
1280{
1281 struct blkcg_gq *blkg;
1282
1283 if (!blkcg_policy_enabled(q, pol))
1284 return;
1285
1286 blk_queue_bypass_start(q);
1287 spin_lock_irq(q->queue_lock);
1288
1289 __clear_bit(pol->plid, q->blkcg_pols);
1290
1291 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1292 /* grab blkcg lock too while removing @pd from @blkg */
1293 spin_lock(&blkg->blkcg->lock);
1294
1295 if (blkg->pd[pol->plid]) {
1296 if (pol->pd_offline_fn)
1297 pol->pd_offline_fn(blkg->pd[pol->plid]);
1298 pol->pd_free_fn(blkg->pd[pol->plid]);
1299 blkg->pd[pol->plid] = NULL;
1300 }
1301
1302 spin_unlock(&blkg->blkcg->lock);
1303 }
1304
1305 spin_unlock_irq(q->queue_lock);
1306 blk_queue_bypass_end(q);
1307}
1308EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1309
1310/**
1311 * blkcg_policy_register - register a blkcg policy
1312 * @pol: blkcg policy to register
1313 *
1314 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1315 * successful registration. Returns 0 on success and -errno on failure.
1316 */
1317int blkcg_policy_register(struct blkcg_policy *pol)
1318{
1319 struct blkcg *blkcg;
1320 int i, ret;
1321
1322 mutex_lock(&blkcg_pol_register_mutex);
1323 mutex_lock(&blkcg_pol_mutex);
1324
1325 /* find an empty slot */
1326 ret = -ENOSPC;
1327 for (i = 0; i < BLKCG_MAX_POLS; i++)
1328 if (!blkcg_policy[i])
1329 break;
1330 if (i >= BLKCG_MAX_POLS)
1331 goto err_unlock;
1332
1333 /* register @pol */
1334 pol->plid = i;
1335 blkcg_policy[pol->plid] = pol;
1336
1337 /* allocate and install cpd's */
1338 if (pol->cpd_alloc_fn) {
1339 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1340 struct blkcg_policy_data *cpd;
1341
1342 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1343 if (!cpd) {
1344 mutex_unlock(&blkcg_pol_mutex);
1345 goto err_free_cpds;
1346 }
1347
1348 blkcg->cpd[pol->plid] = cpd;
1349 cpd->blkcg = blkcg;
1350 cpd->plid = pol->plid;
1351 pol->cpd_init_fn(cpd);
1352 }
1353 }
1354
1355 mutex_unlock(&blkcg_pol_mutex);
1356
1357 /* everything is in place, add intf files for the new policy */
1358 if (pol->dfl_cftypes)
1359 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1360 pol->dfl_cftypes));
1361 if (pol->legacy_cftypes)
1362 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1363 pol->legacy_cftypes));
1364 mutex_unlock(&blkcg_pol_register_mutex);
1365 return 0;
1366
1367err_free_cpds:
1368 if (pol->cpd_alloc_fn) {
1369 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1370 if (blkcg->cpd[pol->plid]) {
1371 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1372 blkcg->cpd[pol->plid] = NULL;
1373 }
1374 }
1375 }
1376 blkcg_policy[pol->plid] = NULL;
1377err_unlock:
1378 mutex_unlock(&blkcg_pol_mutex);
1379 mutex_unlock(&blkcg_pol_register_mutex);
1380 return ret;
1381}
1382EXPORT_SYMBOL_GPL(blkcg_policy_register);
1383
1384/**
1385 * blkcg_policy_unregister - unregister a blkcg policy
1386 * @pol: blkcg policy to unregister
1387 *
1388 * Undo blkcg_policy_register(@pol). Might sleep.
1389 */
1390void blkcg_policy_unregister(struct blkcg_policy *pol)
1391{
1392 struct blkcg *blkcg;
1393
1394 mutex_lock(&blkcg_pol_register_mutex);
1395
1396 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1397 goto out_unlock;
1398
1399 /* kill the intf files first */
1400 if (pol->dfl_cftypes)
1401 cgroup_rm_cftypes(pol->dfl_cftypes);
1402 if (pol->legacy_cftypes)
1403 cgroup_rm_cftypes(pol->legacy_cftypes);
1404
1405 /* remove cpds and unregister */
1406 mutex_lock(&blkcg_pol_mutex);
1407
1408 if (pol->cpd_alloc_fn) {
1409 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1410 if (blkcg->cpd[pol->plid]) {
1411 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1412 blkcg->cpd[pol->plid] = NULL;
1413 }
1414 }
1415 }
1416 blkcg_policy[pol->plid] = NULL;
1417
1418 mutex_unlock(&blkcg_pol_mutex);
1419out_unlock:
1420 mutex_unlock(&blkcg_pol_register_mutex);
1421}
1422EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1/*
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);