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#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 };
30EXPORT_SYMBOL_GPL(blkcg_root);
31
32static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
33
34struct blkcg *cgroup_to_blkcg(struct cgroup *cgroup)
35{
36 return container_of(cgroup_subsys_state(cgroup, blkio_subsys_id),
37 struct blkcg, css);
38}
39EXPORT_SYMBOL_GPL(cgroup_to_blkcg);
40
41static struct blkcg *task_blkcg(struct task_struct *tsk)
42{
43 return container_of(task_subsys_state(tsk, blkio_subsys_id),
44 struct blkcg, css);
45}
46
47struct blkcg *bio_blkcg(struct bio *bio)
48{
49 if (bio && bio->bi_css)
50 return container_of(bio->bi_css, struct blkcg, css);
51 return task_blkcg(current);
52}
53EXPORT_SYMBOL_GPL(bio_blkcg);
54
55static bool blkcg_policy_enabled(struct request_queue *q,
56 const struct blkcg_policy *pol)
57{
58 return pol && test_bit(pol->plid, q->blkcg_pols);
59}
60
61/**
62 * blkg_free - free a blkg
63 * @blkg: blkg to free
64 *
65 * Free @blkg which may be partially allocated.
66 */
67static void blkg_free(struct blkcg_gq *blkg)
68{
69 int i;
70
71 if (!blkg)
72 return;
73
74 for (i = 0; i < BLKCG_MAX_POLS; i++) {
75 struct blkcg_policy *pol = blkcg_policy[i];
76 struct blkg_policy_data *pd = blkg->pd[i];
77
78 if (!pd)
79 continue;
80
81 if (pol && pol->pd_exit_fn)
82 pol->pd_exit_fn(blkg);
83
84 kfree(pd);
85 }
86
87 kfree(blkg);
88}
89
90/**
91 * blkg_alloc - allocate a blkg
92 * @blkcg: block cgroup the new blkg is associated with
93 * @q: request_queue the new blkg is associated with
94 *
95 * Allocate a new blkg assocating @blkcg and @q.
96 */
97static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q)
98{
99 struct blkcg_gq *blkg;
100 int i;
101
102 /* alloc and init base part */
103 blkg = kzalloc_node(sizeof(*blkg), GFP_ATOMIC, q->node);
104 if (!blkg)
105 return NULL;
106
107 blkg->q = q;
108 INIT_LIST_HEAD(&blkg->q_node);
109 blkg->blkcg = blkcg;
110 blkg->refcnt = 1;
111
112 for (i = 0; i < BLKCG_MAX_POLS; i++) {
113 struct blkcg_policy *pol = blkcg_policy[i];
114 struct blkg_policy_data *pd;
115
116 if (!blkcg_policy_enabled(q, pol))
117 continue;
118
119 /* alloc per-policy data and attach it to blkg */
120 pd = kzalloc_node(pol->pd_size, GFP_ATOMIC, q->node);
121 if (!pd) {
122 blkg_free(blkg);
123 return NULL;
124 }
125
126 blkg->pd[i] = pd;
127 pd->blkg = blkg;
128
129 /* invoke per-policy init */
130 if (blkcg_policy_enabled(blkg->q, pol))
131 pol->pd_init_fn(blkg);
132 }
133
134 return blkg;
135}
136
137static struct blkcg_gq *__blkg_lookup(struct blkcg *blkcg,
138 struct request_queue *q)
139{
140 struct blkcg_gq *blkg;
141
142 blkg = rcu_dereference(blkcg->blkg_hint);
143 if (blkg && blkg->q == q)
144 return blkg;
145
146 /*
147 * Hint didn't match. Look up from the radix tree. Note that we
148 * may not be holding queue_lock and thus are not sure whether
149 * @blkg from blkg_tree has already been removed or not, so we
150 * can't update hint to the lookup result. Leave it to the caller.
151 */
152 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
153 if (blkg && blkg->q == q)
154 return blkg;
155
156 return NULL;
157}
158
159/**
160 * blkg_lookup - lookup blkg for the specified blkcg - q pair
161 * @blkcg: blkcg of interest
162 * @q: request_queue of interest
163 *
164 * Lookup blkg for the @blkcg - @q pair. This function should be called
165 * under RCU read lock and is guaranteed to return %NULL if @q is bypassing
166 * - see blk_queue_bypass_start() for details.
167 */
168struct blkcg_gq *blkg_lookup(struct blkcg *blkcg, struct request_queue *q)
169{
170 WARN_ON_ONCE(!rcu_read_lock_held());
171
172 if (unlikely(blk_queue_bypass(q)))
173 return NULL;
174 return __blkg_lookup(blkcg, q);
175}
176EXPORT_SYMBOL_GPL(blkg_lookup);
177
178static struct blkcg_gq *__blkg_lookup_create(struct blkcg *blkcg,
179 struct request_queue *q)
180 __releases(q->queue_lock) __acquires(q->queue_lock)
181{
182 struct blkcg_gq *blkg;
183 int ret;
184
185 WARN_ON_ONCE(!rcu_read_lock_held());
186 lockdep_assert_held(q->queue_lock);
187
188 /* lookup and update hint on success, see __blkg_lookup() for details */
189 blkg = __blkg_lookup(blkcg, q);
190 if (blkg) {
191 rcu_assign_pointer(blkcg->blkg_hint, blkg);
192 return blkg;
193 }
194
195 /* blkg holds a reference to blkcg */
196 if (!css_tryget(&blkcg->css))
197 return ERR_PTR(-EINVAL);
198
199 /* allocate */
200 ret = -ENOMEM;
201 blkg = blkg_alloc(blkcg, q);
202 if (unlikely(!blkg))
203 goto err_put;
204
205 /* insert */
206 ret = radix_tree_preload(GFP_ATOMIC);
207 if (ret)
208 goto err_free;
209
210 spin_lock(&blkcg->lock);
211 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
212 if (likely(!ret)) {
213 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
214 list_add(&blkg->q_node, &q->blkg_list);
215 }
216 spin_unlock(&blkcg->lock);
217
218 radix_tree_preload_end();
219
220 if (!ret)
221 return blkg;
222err_free:
223 blkg_free(blkg);
224err_put:
225 css_put(&blkcg->css);
226 return ERR_PTR(ret);
227}
228
229struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
230 struct request_queue *q)
231{
232 /*
233 * This could be the first entry point of blkcg implementation and
234 * we shouldn't allow anything to go through for a bypassing queue.
235 */
236 if (unlikely(blk_queue_bypass(q)))
237 return ERR_PTR(blk_queue_dead(q) ? -EINVAL : -EBUSY);
238 return __blkg_lookup_create(blkcg, q);
239}
240EXPORT_SYMBOL_GPL(blkg_lookup_create);
241
242static void blkg_destroy(struct blkcg_gq *blkg)
243{
244 struct blkcg *blkcg = blkg->blkcg;
245
246 lockdep_assert_held(blkg->q->queue_lock);
247 lockdep_assert_held(&blkcg->lock);
248
249 /* Something wrong if we are trying to remove same group twice */
250 WARN_ON_ONCE(list_empty(&blkg->q_node));
251 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
252
253 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
254 list_del_init(&blkg->q_node);
255 hlist_del_init_rcu(&blkg->blkcg_node);
256
257 /*
258 * Both setting lookup hint to and clearing it from @blkg are done
259 * under queue_lock. If it's not pointing to @blkg now, it never
260 * will. Hint assignment itself can race safely.
261 */
262 if (rcu_dereference_raw(blkcg->blkg_hint) == blkg)
263 rcu_assign_pointer(blkcg->blkg_hint, NULL);
264
265 /*
266 * Put the reference taken at the time of creation so that when all
267 * queues are gone, group can be destroyed.
268 */
269 blkg_put(blkg);
270}
271
272/**
273 * blkg_destroy_all - destroy all blkgs associated with a request_queue
274 * @q: request_queue of interest
275 *
276 * Destroy all blkgs associated with @q.
277 */
278static void blkg_destroy_all(struct request_queue *q)
279{
280 struct blkcg_gq *blkg, *n;
281
282 lockdep_assert_held(q->queue_lock);
283
284 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
285 struct blkcg *blkcg = blkg->blkcg;
286
287 spin_lock(&blkcg->lock);
288 blkg_destroy(blkg);
289 spin_unlock(&blkcg->lock);
290 }
291}
292
293static void blkg_rcu_free(struct rcu_head *rcu_head)
294{
295 blkg_free(container_of(rcu_head, struct blkcg_gq, rcu_head));
296}
297
298void __blkg_release(struct blkcg_gq *blkg)
299{
300 /* release the extra blkcg reference this blkg has been holding */
301 css_put(&blkg->blkcg->css);
302
303 /*
304 * A group is freed in rcu manner. But having an rcu lock does not
305 * mean that one can access all the fields of blkg and assume these
306 * are valid. For example, don't try to follow throtl_data and
307 * request queue links.
308 *
309 * Having a reference to blkg under an rcu allows acess to only
310 * values local to groups like group stats and group rate limits
311 */
312 call_rcu(&blkg->rcu_head, blkg_rcu_free);
313}
314EXPORT_SYMBOL_GPL(__blkg_release);
315
316static int blkcg_reset_stats(struct cgroup *cgroup, struct cftype *cftype,
317 u64 val)
318{
319 struct blkcg *blkcg = cgroup_to_blkcg(cgroup);
320 struct blkcg_gq *blkg;
321 struct hlist_node *n;
322 int i;
323
324 mutex_lock(&blkcg_pol_mutex);
325 spin_lock_irq(&blkcg->lock);
326
327 /*
328 * Note that stat reset is racy - it doesn't synchronize against
329 * stat updates. This is a debug feature which shouldn't exist
330 * anyway. If you get hit by a race, retry.
331 */
332 hlist_for_each_entry(blkg, n, &blkcg->blkg_list, blkcg_node) {
333 for (i = 0; i < BLKCG_MAX_POLS; i++) {
334 struct blkcg_policy *pol = blkcg_policy[i];
335
336 if (blkcg_policy_enabled(blkg->q, pol) &&
337 pol->pd_reset_stats_fn)
338 pol->pd_reset_stats_fn(blkg);
339 }
340 }
341
342 spin_unlock_irq(&blkcg->lock);
343 mutex_unlock(&blkcg_pol_mutex);
344 return 0;
345}
346
347static const char *blkg_dev_name(struct blkcg_gq *blkg)
348{
349 /* some drivers (floppy) instantiate a queue w/o disk registered */
350 if (blkg->q->backing_dev_info.dev)
351 return dev_name(blkg->q->backing_dev_info.dev);
352 return NULL;
353}
354
355/**
356 * blkcg_print_blkgs - helper for printing per-blkg data
357 * @sf: seq_file to print to
358 * @blkcg: blkcg of interest
359 * @prfill: fill function to print out a blkg
360 * @pol: policy in question
361 * @data: data to be passed to @prfill
362 * @show_total: to print out sum of prfill return values or not
363 *
364 * This function invokes @prfill on each blkg of @blkcg if pd for the
365 * policy specified by @pol exists. @prfill is invoked with @sf, the
366 * policy data and @data. If @show_total is %true, the sum of the return
367 * values from @prfill is printed with "Total" label at the end.
368 *
369 * This is to be used to construct print functions for
370 * cftype->read_seq_string method.
371 */
372void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
373 u64 (*prfill)(struct seq_file *,
374 struct blkg_policy_data *, int),
375 const struct blkcg_policy *pol, int data,
376 bool show_total)
377{
378 struct blkcg_gq *blkg;
379 struct hlist_node *n;
380 u64 total = 0;
381
382 spin_lock_irq(&blkcg->lock);
383 hlist_for_each_entry(blkg, n, &blkcg->blkg_list, blkcg_node)
384 if (blkcg_policy_enabled(blkg->q, pol))
385 total += prfill(sf, blkg->pd[pol->plid], data);
386 spin_unlock_irq(&blkcg->lock);
387
388 if (show_total)
389 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
390}
391EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
392
393/**
394 * __blkg_prfill_u64 - prfill helper for a single u64 value
395 * @sf: seq_file to print to
396 * @pd: policy private data of interest
397 * @v: value to print
398 *
399 * Print @v to @sf for the device assocaited with @pd.
400 */
401u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
402{
403 const char *dname = blkg_dev_name(pd->blkg);
404
405 if (!dname)
406 return 0;
407
408 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
409 return v;
410}
411EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
412
413/**
414 * __blkg_prfill_rwstat - prfill helper for a blkg_rwstat
415 * @sf: seq_file to print to
416 * @pd: policy private data of interest
417 * @rwstat: rwstat to print
418 *
419 * Print @rwstat to @sf for the device assocaited with @pd.
420 */
421u64 __blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
422 const struct blkg_rwstat *rwstat)
423{
424 static const char *rwstr[] = {
425 [BLKG_RWSTAT_READ] = "Read",
426 [BLKG_RWSTAT_WRITE] = "Write",
427 [BLKG_RWSTAT_SYNC] = "Sync",
428 [BLKG_RWSTAT_ASYNC] = "Async",
429 };
430 const char *dname = blkg_dev_name(pd->blkg);
431 u64 v;
432 int i;
433
434 if (!dname)
435 return 0;
436
437 for (i = 0; i < BLKG_RWSTAT_NR; i++)
438 seq_printf(sf, "%s %s %llu\n", dname, rwstr[i],
439 (unsigned long long)rwstat->cnt[i]);
440
441 v = rwstat->cnt[BLKG_RWSTAT_READ] + rwstat->cnt[BLKG_RWSTAT_WRITE];
442 seq_printf(sf, "%s Total %llu\n", dname, (unsigned long long)v);
443 return v;
444}
445
446/**
447 * blkg_prfill_stat - prfill callback for blkg_stat
448 * @sf: seq_file to print to
449 * @pd: policy private data of interest
450 * @off: offset to the blkg_stat in @pd
451 *
452 * prfill callback for printing a blkg_stat.
453 */
454u64 blkg_prfill_stat(struct seq_file *sf, struct blkg_policy_data *pd, int off)
455{
456 return __blkg_prfill_u64(sf, pd, blkg_stat_read((void *)pd + off));
457}
458EXPORT_SYMBOL_GPL(blkg_prfill_stat);
459
460/**
461 * blkg_prfill_rwstat - prfill callback for blkg_rwstat
462 * @sf: seq_file to print to
463 * @pd: policy private data of interest
464 * @off: offset to the blkg_rwstat in @pd
465 *
466 * prfill callback for printing a blkg_rwstat.
467 */
468u64 blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
469 int off)
470{
471 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd + off);
472
473 return __blkg_prfill_rwstat(sf, pd, &rwstat);
474}
475EXPORT_SYMBOL_GPL(blkg_prfill_rwstat);
476
477/**
478 * blkg_conf_prep - parse and prepare for per-blkg config update
479 * @blkcg: target block cgroup
480 * @pol: target policy
481 * @input: input string
482 * @ctx: blkg_conf_ctx to be filled
483 *
484 * Parse per-blkg config update from @input and initialize @ctx with the
485 * result. @ctx->blkg points to the blkg to be updated and @ctx->v the new
486 * value. This function returns with RCU read lock and queue lock held and
487 * must be paired with blkg_conf_finish().
488 */
489int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
490 const char *input, struct blkg_conf_ctx *ctx)
491 __acquires(rcu) __acquires(disk->queue->queue_lock)
492{
493 struct gendisk *disk;
494 struct blkcg_gq *blkg;
495 unsigned int major, minor;
496 unsigned long long v;
497 int part, ret;
498
499 if (sscanf(input, "%u:%u %llu", &major, &minor, &v) != 3)
500 return -EINVAL;
501
502 disk = get_gendisk(MKDEV(major, minor), &part);
503 if (!disk || part)
504 return -EINVAL;
505
506 rcu_read_lock();
507 spin_lock_irq(disk->queue->queue_lock);
508
509 if (blkcg_policy_enabled(disk->queue, pol))
510 blkg = blkg_lookup_create(blkcg, disk->queue);
511 else
512 blkg = ERR_PTR(-EINVAL);
513
514 if (IS_ERR(blkg)) {
515 ret = PTR_ERR(blkg);
516 rcu_read_unlock();
517 spin_unlock_irq(disk->queue->queue_lock);
518 put_disk(disk);
519 /*
520 * If queue was bypassing, we should retry. Do so after a
521 * short msleep(). It isn't strictly necessary but queue
522 * can be bypassing for some time and it's always nice to
523 * avoid busy looping.
524 */
525 if (ret == -EBUSY) {
526 msleep(10);
527 ret = restart_syscall();
528 }
529 return ret;
530 }
531
532 ctx->disk = disk;
533 ctx->blkg = blkg;
534 ctx->v = v;
535 return 0;
536}
537EXPORT_SYMBOL_GPL(blkg_conf_prep);
538
539/**
540 * blkg_conf_finish - finish up per-blkg config update
541 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
542 *
543 * Finish up after per-blkg config update. This function must be paired
544 * with blkg_conf_prep().
545 */
546void blkg_conf_finish(struct blkg_conf_ctx *ctx)
547 __releases(ctx->disk->queue->queue_lock) __releases(rcu)
548{
549 spin_unlock_irq(ctx->disk->queue->queue_lock);
550 rcu_read_unlock();
551 put_disk(ctx->disk);
552}
553EXPORT_SYMBOL_GPL(blkg_conf_finish);
554
555struct cftype blkcg_files[] = {
556 {
557 .name = "reset_stats",
558 .write_u64 = blkcg_reset_stats,
559 },
560 { } /* terminate */
561};
562
563/**
564 * blkcg_pre_destroy - cgroup pre_destroy callback
565 * @cgroup: cgroup of interest
566 *
567 * This function is called when @cgroup is about to go away and responsible
568 * for shooting down all blkgs associated with @cgroup. blkgs should be
569 * removed while holding both q and blkcg locks. As blkcg lock is nested
570 * inside q lock, this function performs reverse double lock dancing.
571 *
572 * This is the blkcg counterpart of ioc_release_fn().
573 */
574static int blkcg_pre_destroy(struct cgroup *cgroup)
575{
576 struct blkcg *blkcg = cgroup_to_blkcg(cgroup);
577
578 spin_lock_irq(&blkcg->lock);
579
580 while (!hlist_empty(&blkcg->blkg_list)) {
581 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
582 struct blkcg_gq, blkcg_node);
583 struct request_queue *q = blkg->q;
584
585 if (spin_trylock(q->queue_lock)) {
586 blkg_destroy(blkg);
587 spin_unlock(q->queue_lock);
588 } else {
589 spin_unlock_irq(&blkcg->lock);
590 cpu_relax();
591 spin_lock_irq(&blkcg->lock);
592 }
593 }
594
595 spin_unlock_irq(&blkcg->lock);
596 return 0;
597}
598
599static void blkcg_destroy(struct cgroup *cgroup)
600{
601 struct blkcg *blkcg = cgroup_to_blkcg(cgroup);
602
603 if (blkcg != &blkcg_root)
604 kfree(blkcg);
605}
606
607static struct cgroup_subsys_state *blkcg_create(struct cgroup *cgroup)
608{
609 static atomic64_t id_seq = ATOMIC64_INIT(0);
610 struct blkcg *blkcg;
611 struct cgroup *parent = cgroup->parent;
612
613 if (!parent) {
614 blkcg = &blkcg_root;
615 goto done;
616 }
617
618 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
619 if (!blkcg)
620 return ERR_PTR(-ENOMEM);
621
622 blkcg->cfq_weight = CFQ_WEIGHT_DEFAULT;
623 blkcg->id = atomic64_inc_return(&id_seq); /* root is 0, start from 1 */
624done:
625 spin_lock_init(&blkcg->lock);
626 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_ATOMIC);
627 INIT_HLIST_HEAD(&blkcg->blkg_list);
628
629 return &blkcg->css;
630}
631
632/**
633 * blkcg_init_queue - initialize blkcg part of request queue
634 * @q: request_queue to initialize
635 *
636 * Called from blk_alloc_queue_node(). Responsible for initializing blkcg
637 * part of new request_queue @q.
638 *
639 * RETURNS:
640 * 0 on success, -errno on failure.
641 */
642int blkcg_init_queue(struct request_queue *q)
643{
644 might_sleep();
645
646 return blk_throtl_init(q);
647}
648
649/**
650 * blkcg_drain_queue - drain blkcg part of request_queue
651 * @q: request_queue to drain
652 *
653 * Called from blk_drain_queue(). Responsible for draining blkcg part.
654 */
655void blkcg_drain_queue(struct request_queue *q)
656{
657 lockdep_assert_held(q->queue_lock);
658
659 blk_throtl_drain(q);
660}
661
662/**
663 * blkcg_exit_queue - exit and release blkcg part of request_queue
664 * @q: request_queue being released
665 *
666 * Called from blk_release_queue(). Responsible for exiting blkcg part.
667 */
668void blkcg_exit_queue(struct request_queue *q)
669{
670 spin_lock_irq(q->queue_lock);
671 blkg_destroy_all(q);
672 spin_unlock_irq(q->queue_lock);
673
674 blk_throtl_exit(q);
675}
676
677/*
678 * We cannot support shared io contexts, as we have no mean to support
679 * two tasks with the same ioc in two different groups without major rework
680 * of the main cic data structures. For now we allow a task to change
681 * its cgroup only if it's the only owner of its ioc.
682 */
683static int blkcg_can_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
684{
685 struct task_struct *task;
686 struct io_context *ioc;
687 int ret = 0;
688
689 /* task_lock() is needed to avoid races with exit_io_context() */
690 cgroup_taskset_for_each(task, cgrp, tset) {
691 task_lock(task);
692 ioc = task->io_context;
693 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
694 ret = -EINVAL;
695 task_unlock(task);
696 if (ret)
697 break;
698 }
699 return ret;
700}
701
702struct cgroup_subsys blkio_subsys = {
703 .name = "blkio",
704 .create = blkcg_create,
705 .can_attach = blkcg_can_attach,
706 .pre_destroy = blkcg_pre_destroy,
707 .destroy = blkcg_destroy,
708 .subsys_id = blkio_subsys_id,
709 .base_cftypes = blkcg_files,
710 .module = THIS_MODULE,
711};
712EXPORT_SYMBOL_GPL(blkio_subsys);
713
714/**
715 * blkcg_activate_policy - activate a blkcg policy on a request_queue
716 * @q: request_queue of interest
717 * @pol: blkcg policy to activate
718 *
719 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
720 * bypass mode to populate its blkgs with policy_data for @pol.
721 *
722 * Activation happens with @q bypassed, so nobody would be accessing blkgs
723 * from IO path. Update of each blkg is protected by both queue and blkcg
724 * locks so that holding either lock and testing blkcg_policy_enabled() is
725 * always enough for dereferencing policy data.
726 *
727 * The caller is responsible for synchronizing [de]activations and policy
728 * [un]registerations. Returns 0 on success, -errno on failure.
729 */
730int blkcg_activate_policy(struct request_queue *q,
731 const struct blkcg_policy *pol)
732{
733 LIST_HEAD(pds);
734 struct blkcg_gq *blkg;
735 struct blkg_policy_data *pd, *n;
736 int cnt = 0, ret;
737
738 if (blkcg_policy_enabled(q, pol))
739 return 0;
740
741 blk_queue_bypass_start(q);
742
743 /* make sure the root blkg exists and count the existing blkgs */
744 spin_lock_irq(q->queue_lock);
745
746 rcu_read_lock();
747 blkg = __blkg_lookup_create(&blkcg_root, q);
748 rcu_read_unlock();
749
750 if (IS_ERR(blkg)) {
751 ret = PTR_ERR(blkg);
752 goto out_unlock;
753 }
754 q->root_blkg = blkg;
755
756 list_for_each_entry(blkg, &q->blkg_list, q_node)
757 cnt++;
758
759 spin_unlock_irq(q->queue_lock);
760
761 /* allocate policy_data for all existing blkgs */
762 while (cnt--) {
763 pd = kzalloc_node(pol->pd_size, GFP_KERNEL, q->node);
764 if (!pd) {
765 ret = -ENOMEM;
766 goto out_free;
767 }
768 list_add_tail(&pd->alloc_node, &pds);
769 }
770
771 /*
772 * Install the allocated pds. With @q bypassing, no new blkg
773 * should have been created while the queue lock was dropped.
774 */
775 spin_lock_irq(q->queue_lock);
776
777 list_for_each_entry(blkg, &q->blkg_list, q_node) {
778 if (WARN_ON(list_empty(&pds))) {
779 /* umm... this shouldn't happen, just abort */
780 ret = -ENOMEM;
781 goto out_unlock;
782 }
783 pd = list_first_entry(&pds, struct blkg_policy_data, alloc_node);
784 list_del_init(&pd->alloc_node);
785
786 /* grab blkcg lock too while installing @pd on @blkg */
787 spin_lock(&blkg->blkcg->lock);
788
789 blkg->pd[pol->plid] = pd;
790 pd->blkg = blkg;
791 pol->pd_init_fn(blkg);
792
793 spin_unlock(&blkg->blkcg->lock);
794 }
795
796 __set_bit(pol->plid, q->blkcg_pols);
797 ret = 0;
798out_unlock:
799 spin_unlock_irq(q->queue_lock);
800out_free:
801 blk_queue_bypass_end(q);
802 list_for_each_entry_safe(pd, n, &pds, alloc_node)
803 kfree(pd);
804 return ret;
805}
806EXPORT_SYMBOL_GPL(blkcg_activate_policy);
807
808/**
809 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
810 * @q: request_queue of interest
811 * @pol: blkcg policy to deactivate
812 *
813 * Deactivate @pol on @q. Follows the same synchronization rules as
814 * blkcg_activate_policy().
815 */
816void blkcg_deactivate_policy(struct request_queue *q,
817 const struct blkcg_policy *pol)
818{
819 struct blkcg_gq *blkg;
820
821 if (!blkcg_policy_enabled(q, pol))
822 return;
823
824 blk_queue_bypass_start(q);
825 spin_lock_irq(q->queue_lock);
826
827 __clear_bit(pol->plid, q->blkcg_pols);
828
829 /* if no policy is left, no need for blkgs - shoot them down */
830 if (bitmap_empty(q->blkcg_pols, BLKCG_MAX_POLS))
831 blkg_destroy_all(q);
832
833 list_for_each_entry(blkg, &q->blkg_list, q_node) {
834 /* grab blkcg lock too while removing @pd from @blkg */
835 spin_lock(&blkg->blkcg->lock);
836
837 if (pol->pd_exit_fn)
838 pol->pd_exit_fn(blkg);
839
840 kfree(blkg->pd[pol->plid]);
841 blkg->pd[pol->plid] = NULL;
842
843 spin_unlock(&blkg->blkcg->lock);
844 }
845
846 spin_unlock_irq(q->queue_lock);
847 blk_queue_bypass_end(q);
848}
849EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
850
851/**
852 * blkcg_policy_register - register a blkcg policy
853 * @pol: blkcg policy to register
854 *
855 * Register @pol with blkcg core. Might sleep and @pol may be modified on
856 * successful registration. Returns 0 on success and -errno on failure.
857 */
858int blkcg_policy_register(struct blkcg_policy *pol)
859{
860 int i, ret;
861
862 if (WARN_ON(pol->pd_size < sizeof(struct blkg_policy_data)))
863 return -EINVAL;
864
865 mutex_lock(&blkcg_pol_mutex);
866
867 /* find an empty slot */
868 ret = -ENOSPC;
869 for (i = 0; i < BLKCG_MAX_POLS; i++)
870 if (!blkcg_policy[i])
871 break;
872 if (i >= BLKCG_MAX_POLS)
873 goto out_unlock;
874
875 /* register and update blkgs */
876 pol->plid = i;
877 blkcg_policy[i] = pol;
878
879 /* everything is in place, add intf files for the new policy */
880 if (pol->cftypes)
881 WARN_ON(cgroup_add_cftypes(&blkio_subsys, pol->cftypes));
882 ret = 0;
883out_unlock:
884 mutex_unlock(&blkcg_pol_mutex);
885 return ret;
886}
887EXPORT_SYMBOL_GPL(blkcg_policy_register);
888
889/**
890 * blkcg_policy_unregister - unregister a blkcg policy
891 * @pol: blkcg policy to unregister
892 *
893 * Undo blkcg_policy_register(@pol). Might sleep.
894 */
895void blkcg_policy_unregister(struct blkcg_policy *pol)
896{
897 mutex_lock(&blkcg_pol_mutex);
898
899 if (WARN_ON(blkcg_policy[pol->plid] != pol))
900 goto out_unlock;
901
902 /* kill the intf files first */
903 if (pol->cftypes)
904 cgroup_rm_cftypes(&blkio_subsys, pol->cftypes);
905
906 /* unregister and update blkgs */
907 blkcg_policy[pol->plid] = NULL;
908out_unlock:
909 mutex_unlock(&blkcg_pol_mutex);
910}
911EXPORT_SYMBOL_GPL(blkcg_policy_unregister);