Loading...
1/*
2 * Common Block IO controller cgroup interface
3 *
4 * Based on ideas and code from CFQ, CFS and BFQ:
5 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
6 *
7 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
8 * Paolo Valente <paolo.valente@unimore.it>
9 *
10 * Copyright (C) 2009 Vivek Goyal <vgoyal@redhat.com>
11 * Nauman Rafique <nauman@google.com>
12 */
13#include <linux/ioprio.h>
14#include <linux/kdev_t.h>
15#include <linux/module.h>
16#include <linux/err.h>
17#include <linux/blkdev.h>
18#include <linux/slab.h>
19#include <linux/genhd.h>
20#include <linux/delay.h>
21#include <linux/atomic.h>
22#include "blk-cgroup.h"
23#include "blk.h"
24
25#define MAX_KEY_LEN 100
26
27static DEFINE_MUTEX(blkcg_pol_mutex);
28
29struct blkcg blkcg_root = { .cfq_weight = 2 * CFQ_WEIGHT_DEFAULT,
30 .cfq_leaf_weight = 2 * CFQ_WEIGHT_DEFAULT, };
31EXPORT_SYMBOL_GPL(blkcg_root);
32
33static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
34
35static bool blkcg_policy_enabled(struct request_queue *q,
36 const struct blkcg_policy *pol)
37{
38 return pol && test_bit(pol->plid, q->blkcg_pols);
39}
40
41/**
42 * blkg_free - free a blkg
43 * @blkg: blkg to free
44 *
45 * Free @blkg which may be partially allocated.
46 */
47static void blkg_free(struct blkcg_gq *blkg)
48{
49 int i;
50
51 if (!blkg)
52 return;
53
54 for (i = 0; i < BLKCG_MAX_POLS; i++)
55 kfree(blkg->pd[i]);
56
57 blk_exit_rl(&blkg->rl);
58 kfree(blkg);
59}
60
61/**
62 * blkg_alloc - allocate a blkg
63 * @blkcg: block cgroup the new blkg is associated with
64 * @q: request_queue the new blkg is associated with
65 * @gfp_mask: allocation mask to use
66 *
67 * Allocate a new blkg assocating @blkcg and @q.
68 */
69static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
70 gfp_t gfp_mask)
71{
72 struct blkcg_gq *blkg;
73 int i;
74
75 /* alloc and init base part */
76 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
77 if (!blkg)
78 return NULL;
79
80 blkg->q = q;
81 INIT_LIST_HEAD(&blkg->q_node);
82 blkg->blkcg = blkcg;
83 blkg->refcnt = 1;
84
85 /* root blkg uses @q->root_rl, init rl only for !root blkgs */
86 if (blkcg != &blkcg_root) {
87 if (blk_init_rl(&blkg->rl, q, gfp_mask))
88 goto err_free;
89 blkg->rl.blkg = blkg;
90 }
91
92 for (i = 0; i < BLKCG_MAX_POLS; i++) {
93 struct blkcg_policy *pol = blkcg_policy[i];
94 struct blkg_policy_data *pd;
95
96 if (!blkcg_policy_enabled(q, pol))
97 continue;
98
99 /* alloc per-policy data and attach it to blkg */
100 pd = kzalloc_node(pol->pd_size, gfp_mask, q->node);
101 if (!pd)
102 goto err_free;
103
104 blkg->pd[i] = pd;
105 pd->blkg = blkg;
106 pd->plid = i;
107 }
108
109 return blkg;
110
111err_free:
112 blkg_free(blkg);
113 return NULL;
114}
115
116/**
117 * __blkg_lookup - internal version of blkg_lookup()
118 * @blkcg: blkcg of interest
119 * @q: request_queue of interest
120 * @update_hint: whether to update lookup hint with the result or not
121 *
122 * This is internal version and shouldn't be used by policy
123 * implementations. Looks up blkgs for the @blkcg - @q pair regardless of
124 * @q's bypass state. If @update_hint is %true, the caller should be
125 * holding @q->queue_lock and lookup hint is updated on success.
126 */
127struct blkcg_gq *__blkg_lookup(struct blkcg *blkcg, struct request_queue *q,
128 bool update_hint)
129{
130 struct blkcg_gq *blkg;
131
132 blkg = rcu_dereference(blkcg->blkg_hint);
133 if (blkg && blkg->q == q)
134 return blkg;
135
136 /*
137 * Hint didn't match. Look up from the radix tree. Note that the
138 * hint can only be updated under queue_lock as otherwise @blkg
139 * could have already been removed from blkg_tree. The caller is
140 * responsible for grabbing queue_lock if @update_hint.
141 */
142 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
143 if (blkg && blkg->q == q) {
144 if (update_hint) {
145 lockdep_assert_held(q->queue_lock);
146 rcu_assign_pointer(blkcg->blkg_hint, blkg);
147 }
148 return blkg;
149 }
150
151 return NULL;
152}
153
154/**
155 * blkg_lookup - lookup blkg for the specified blkcg - q pair
156 * @blkcg: blkcg of interest
157 * @q: request_queue of interest
158 *
159 * Lookup blkg for the @blkcg - @q pair. This function should be called
160 * under RCU read lock and is guaranteed to return %NULL if @q is bypassing
161 * - see blk_queue_bypass_start() for details.
162 */
163struct blkcg_gq *blkg_lookup(struct blkcg *blkcg, struct request_queue *q)
164{
165 WARN_ON_ONCE(!rcu_read_lock_held());
166
167 if (unlikely(blk_queue_bypass(q)))
168 return NULL;
169 return __blkg_lookup(blkcg, q, false);
170}
171EXPORT_SYMBOL_GPL(blkg_lookup);
172
173/*
174 * If @new_blkg is %NULL, this function tries to allocate a new one as
175 * necessary using %GFP_ATOMIC. @new_blkg is always consumed on return.
176 */
177static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
178 struct request_queue *q,
179 struct blkcg_gq *new_blkg)
180{
181 struct blkcg_gq *blkg;
182 int i, ret;
183
184 WARN_ON_ONCE(!rcu_read_lock_held());
185 lockdep_assert_held(q->queue_lock);
186
187 /* blkg holds a reference to blkcg */
188 if (!css_tryget(&blkcg->css)) {
189 ret = -EINVAL;
190 goto err_free_blkg;
191 }
192
193 /* allocate */
194 if (!new_blkg) {
195 new_blkg = blkg_alloc(blkcg, q, GFP_ATOMIC);
196 if (unlikely(!new_blkg)) {
197 ret = -ENOMEM;
198 goto err_put_css;
199 }
200 }
201 blkg = new_blkg;
202
203 /* link parent */
204 if (blkcg_parent(blkcg)) {
205 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
206 if (WARN_ON_ONCE(!blkg->parent)) {
207 ret = -EINVAL;
208 goto err_put_css;
209 }
210 blkg_get(blkg->parent);
211 }
212
213 /* invoke per-policy init */
214 for (i = 0; i < BLKCG_MAX_POLS; i++) {
215 struct blkcg_policy *pol = blkcg_policy[i];
216
217 if (blkg->pd[i] && pol->pd_init_fn)
218 pol->pd_init_fn(blkg);
219 }
220
221 /* insert */
222 spin_lock(&blkcg->lock);
223 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
224 if (likely(!ret)) {
225 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
226 list_add(&blkg->q_node, &q->blkg_list);
227
228 for (i = 0; i < BLKCG_MAX_POLS; i++) {
229 struct blkcg_policy *pol = blkcg_policy[i];
230
231 if (blkg->pd[i] && pol->pd_online_fn)
232 pol->pd_online_fn(blkg);
233 }
234 }
235 blkg->online = true;
236 spin_unlock(&blkcg->lock);
237
238 if (!ret) {
239 if (blkcg == &blkcg_root) {
240 q->root_blkg = blkg;
241 q->root_rl.blkg = blkg;
242 }
243 return blkg;
244 }
245
246 /* @blkg failed fully initialized, use the usual release path */
247 blkg_put(blkg);
248 return ERR_PTR(ret);
249
250err_put_css:
251 css_put(&blkcg->css);
252err_free_blkg:
253 blkg_free(new_blkg);
254 return ERR_PTR(ret);
255}
256
257/**
258 * blkg_lookup_create - lookup blkg, try to create one if not there
259 * @blkcg: blkcg of interest
260 * @q: request_queue of interest
261 *
262 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
263 * create one. blkg creation is performed recursively from blkcg_root such
264 * that all non-root blkg's have access to the parent blkg. This function
265 * should be called under RCU read lock and @q->queue_lock.
266 *
267 * Returns pointer to the looked up or created blkg on success, ERR_PTR()
268 * value on error. If @q is dead, returns ERR_PTR(-EINVAL). If @q is not
269 * dead and bypassing, returns ERR_PTR(-EBUSY).
270 */
271struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
272 struct request_queue *q)
273{
274 struct blkcg_gq *blkg;
275
276 WARN_ON_ONCE(!rcu_read_lock_held());
277 lockdep_assert_held(q->queue_lock);
278
279 /*
280 * This could be the first entry point of blkcg implementation and
281 * we shouldn't allow anything to go through for a bypassing queue.
282 */
283 if (unlikely(blk_queue_bypass(q)))
284 return ERR_PTR(blk_queue_dying(q) ? -EINVAL : -EBUSY);
285
286 blkg = __blkg_lookup(blkcg, q, true);
287 if (blkg)
288 return blkg;
289
290 /*
291 * Create blkgs walking down from blkcg_root to @blkcg, so that all
292 * non-root blkgs have access to their parents.
293 */
294 while (true) {
295 struct blkcg *pos = blkcg;
296 struct blkcg *parent = blkcg_parent(blkcg);
297
298 while (parent && !__blkg_lookup(parent, q, false)) {
299 pos = parent;
300 parent = blkcg_parent(parent);
301 }
302
303 blkg = blkg_create(pos, q, NULL);
304 if (pos == blkcg || IS_ERR(blkg))
305 return blkg;
306 }
307}
308EXPORT_SYMBOL_GPL(blkg_lookup_create);
309
310static void blkg_destroy(struct blkcg_gq *blkg)
311{
312 struct blkcg *blkcg = blkg->blkcg;
313 int i;
314
315 lockdep_assert_held(blkg->q->queue_lock);
316 lockdep_assert_held(&blkcg->lock);
317
318 /* Something wrong if we are trying to remove same group twice */
319 WARN_ON_ONCE(list_empty(&blkg->q_node));
320 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
321
322 for (i = 0; i < BLKCG_MAX_POLS; i++) {
323 struct blkcg_policy *pol = blkcg_policy[i];
324
325 if (blkg->pd[i] && pol->pd_offline_fn)
326 pol->pd_offline_fn(blkg);
327 }
328 blkg->online = false;
329
330 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
331 list_del_init(&blkg->q_node);
332 hlist_del_init_rcu(&blkg->blkcg_node);
333
334 /*
335 * Both setting lookup hint to and clearing it from @blkg are done
336 * under queue_lock. If it's not pointing to @blkg now, it never
337 * will. Hint assignment itself can race safely.
338 */
339 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
340 rcu_assign_pointer(blkcg->blkg_hint, NULL);
341
342 /*
343 * If root blkg is destroyed. Just clear the pointer since root_rl
344 * does not take reference on root blkg.
345 */
346 if (blkcg == &blkcg_root) {
347 blkg->q->root_blkg = NULL;
348 blkg->q->root_rl.blkg = NULL;
349 }
350
351 /*
352 * Put the reference taken at the time of creation so that when all
353 * queues are gone, group can be destroyed.
354 */
355 blkg_put(blkg);
356}
357
358/**
359 * blkg_destroy_all - destroy all blkgs associated with a request_queue
360 * @q: request_queue of interest
361 *
362 * Destroy all blkgs associated with @q.
363 */
364static void blkg_destroy_all(struct request_queue *q)
365{
366 struct blkcg_gq *blkg, *n;
367
368 lockdep_assert_held(q->queue_lock);
369
370 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
371 struct blkcg *blkcg = blkg->blkcg;
372
373 spin_lock(&blkcg->lock);
374 blkg_destroy(blkg);
375 spin_unlock(&blkcg->lock);
376 }
377}
378
379/*
380 * A group is RCU protected, but having an rcu lock does not mean that one
381 * can access all the fields of blkg and assume these are valid. For
382 * example, don't try to follow throtl_data and request queue links.
383 *
384 * Having a reference to blkg under an rcu allows accesses to only values
385 * local to groups like group stats and group rate limits.
386 */
387void __blkg_release_rcu(struct rcu_head *rcu_head)
388{
389 struct blkcg_gq *blkg = container_of(rcu_head, struct blkcg_gq, rcu_head);
390 int i;
391
392 /* tell policies that this one is being freed */
393 for (i = 0; i < BLKCG_MAX_POLS; i++) {
394 struct blkcg_policy *pol = blkcg_policy[i];
395
396 if (blkg->pd[i] && pol->pd_exit_fn)
397 pol->pd_exit_fn(blkg);
398 }
399
400 /* release the blkcg and parent blkg refs this blkg has been holding */
401 css_put(&blkg->blkcg->css);
402 if (blkg->parent) {
403 spin_lock_irq(blkg->q->queue_lock);
404 blkg_put(blkg->parent);
405 spin_unlock_irq(blkg->q->queue_lock);
406 }
407
408 blkg_free(blkg);
409}
410EXPORT_SYMBOL_GPL(__blkg_release_rcu);
411
412/*
413 * The next function used by blk_queue_for_each_rl(). It's a bit tricky
414 * because the root blkg uses @q->root_rl instead of its own rl.
415 */
416struct request_list *__blk_queue_next_rl(struct request_list *rl,
417 struct request_queue *q)
418{
419 struct list_head *ent;
420 struct blkcg_gq *blkg;
421
422 /*
423 * Determine the current blkg list_head. The first entry is
424 * root_rl which is off @q->blkg_list and mapped to the head.
425 */
426 if (rl == &q->root_rl) {
427 ent = &q->blkg_list;
428 /* There are no more block groups, hence no request lists */
429 if (list_empty(ent))
430 return NULL;
431 } else {
432 blkg = container_of(rl, struct blkcg_gq, rl);
433 ent = &blkg->q_node;
434 }
435
436 /* walk to the next list_head, skip root blkcg */
437 ent = ent->next;
438 if (ent == &q->root_blkg->q_node)
439 ent = ent->next;
440 if (ent == &q->blkg_list)
441 return NULL;
442
443 blkg = container_of(ent, struct blkcg_gq, q_node);
444 return &blkg->rl;
445}
446
447static int blkcg_reset_stats(struct cgroup_subsys_state *css,
448 struct cftype *cftype, u64 val)
449{
450 struct blkcg *blkcg = css_to_blkcg(css);
451 struct blkcg_gq *blkg;
452 int i;
453
454 /*
455 * XXX: We invoke cgroup_add/rm_cftypes() under blkcg_pol_mutex
456 * which ends up putting cgroup's internal cgroup_tree_mutex under
457 * it; however, cgroup_tree_mutex is nested above cgroup file
458 * active protection and grabbing blkcg_pol_mutex from a cgroup
459 * file operation creates a possible circular dependency. cgroup
460 * internal locking is planned to go through further simplification
461 * and this issue should go away soon. For now, let's trylock
462 * blkcg_pol_mutex and restart the write on failure.
463 *
464 * http://lkml.kernel.org/g/5363C04B.4010400@oracle.com
465 */
466 if (!mutex_trylock(&blkcg_pol_mutex))
467 return restart_syscall();
468 spin_lock_irq(&blkcg->lock);
469
470 /*
471 * Note that stat reset is racy - it doesn't synchronize against
472 * stat updates. This is a debug feature which shouldn't exist
473 * anyway. If you get hit by a race, retry.
474 */
475 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
476 for (i = 0; i < BLKCG_MAX_POLS; i++) {
477 struct blkcg_policy *pol = blkcg_policy[i];
478
479 if (blkcg_policy_enabled(blkg->q, pol) &&
480 pol->pd_reset_stats_fn)
481 pol->pd_reset_stats_fn(blkg);
482 }
483 }
484
485 spin_unlock_irq(&blkcg->lock);
486 mutex_unlock(&blkcg_pol_mutex);
487 return 0;
488}
489
490static const char *blkg_dev_name(struct blkcg_gq *blkg)
491{
492 /* some drivers (floppy) instantiate a queue w/o disk registered */
493 if (blkg->q->backing_dev_info.dev)
494 return dev_name(blkg->q->backing_dev_info.dev);
495 return NULL;
496}
497
498/**
499 * blkcg_print_blkgs - helper for printing per-blkg data
500 * @sf: seq_file to print to
501 * @blkcg: blkcg of interest
502 * @prfill: fill function to print out a blkg
503 * @pol: policy in question
504 * @data: data to be passed to @prfill
505 * @show_total: to print out sum of prfill return values or not
506 *
507 * This function invokes @prfill on each blkg of @blkcg if pd for the
508 * policy specified by @pol exists. @prfill is invoked with @sf, the
509 * policy data and @data and the matching queue lock held. If @show_total
510 * is %true, the sum of the return values from @prfill is printed with
511 * "Total" label at the end.
512 *
513 * This is to be used to construct print functions for
514 * cftype->read_seq_string method.
515 */
516void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
517 u64 (*prfill)(struct seq_file *,
518 struct blkg_policy_data *, int),
519 const struct blkcg_policy *pol, int data,
520 bool show_total)
521{
522 struct blkcg_gq *blkg;
523 u64 total = 0;
524
525 rcu_read_lock();
526 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
527 spin_lock_irq(blkg->q->queue_lock);
528 if (blkcg_policy_enabled(blkg->q, pol))
529 total += prfill(sf, blkg->pd[pol->plid], data);
530 spin_unlock_irq(blkg->q->queue_lock);
531 }
532 rcu_read_unlock();
533
534 if (show_total)
535 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
536}
537EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
538
539/**
540 * __blkg_prfill_u64 - prfill helper for a single u64 value
541 * @sf: seq_file to print to
542 * @pd: policy private data of interest
543 * @v: value to print
544 *
545 * Print @v to @sf for the device assocaited with @pd.
546 */
547u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
548{
549 const char *dname = blkg_dev_name(pd->blkg);
550
551 if (!dname)
552 return 0;
553
554 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
555 return v;
556}
557EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
558
559/**
560 * __blkg_prfill_rwstat - prfill helper for a blkg_rwstat
561 * @sf: seq_file to print to
562 * @pd: policy private data of interest
563 * @rwstat: rwstat to print
564 *
565 * Print @rwstat to @sf for the device assocaited with @pd.
566 */
567u64 __blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
568 const struct blkg_rwstat *rwstat)
569{
570 static const char *rwstr[] = {
571 [BLKG_RWSTAT_READ] = "Read",
572 [BLKG_RWSTAT_WRITE] = "Write",
573 [BLKG_RWSTAT_SYNC] = "Sync",
574 [BLKG_RWSTAT_ASYNC] = "Async",
575 };
576 const char *dname = blkg_dev_name(pd->blkg);
577 u64 v;
578 int i;
579
580 if (!dname)
581 return 0;
582
583 for (i = 0; i < BLKG_RWSTAT_NR; i++)
584 seq_printf(sf, "%s %s %llu\n", dname, rwstr[i],
585 (unsigned long long)rwstat->cnt[i]);
586
587 v = rwstat->cnt[BLKG_RWSTAT_READ] + rwstat->cnt[BLKG_RWSTAT_WRITE];
588 seq_printf(sf, "%s Total %llu\n", dname, (unsigned long long)v);
589 return v;
590}
591EXPORT_SYMBOL_GPL(__blkg_prfill_rwstat);
592
593/**
594 * blkg_prfill_stat - prfill callback for blkg_stat
595 * @sf: seq_file to print to
596 * @pd: policy private data of interest
597 * @off: offset to the blkg_stat in @pd
598 *
599 * prfill callback for printing a blkg_stat.
600 */
601u64 blkg_prfill_stat(struct seq_file *sf, struct blkg_policy_data *pd, int off)
602{
603 return __blkg_prfill_u64(sf, pd, blkg_stat_read((void *)pd + off));
604}
605EXPORT_SYMBOL_GPL(blkg_prfill_stat);
606
607/**
608 * blkg_prfill_rwstat - prfill callback for blkg_rwstat
609 * @sf: seq_file to print to
610 * @pd: policy private data of interest
611 * @off: offset to the blkg_rwstat in @pd
612 *
613 * prfill callback for printing a blkg_rwstat.
614 */
615u64 blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
616 int off)
617{
618 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd + off);
619
620 return __blkg_prfill_rwstat(sf, pd, &rwstat);
621}
622EXPORT_SYMBOL_GPL(blkg_prfill_rwstat);
623
624/**
625 * blkg_stat_recursive_sum - collect hierarchical blkg_stat
626 * @pd: policy private data of interest
627 * @off: offset to the blkg_stat in @pd
628 *
629 * Collect the blkg_stat specified by @off from @pd and all its online
630 * descendants and return the sum. The caller must be holding the queue
631 * lock for online tests.
632 */
633u64 blkg_stat_recursive_sum(struct blkg_policy_data *pd, int off)
634{
635 struct blkcg_policy *pol = blkcg_policy[pd->plid];
636 struct blkcg_gq *pos_blkg;
637 struct cgroup_subsys_state *pos_css;
638 u64 sum = 0;
639
640 lockdep_assert_held(pd->blkg->q->queue_lock);
641
642 rcu_read_lock();
643 blkg_for_each_descendant_pre(pos_blkg, pos_css, pd_to_blkg(pd)) {
644 struct blkg_policy_data *pos_pd = blkg_to_pd(pos_blkg, pol);
645 struct blkg_stat *stat = (void *)pos_pd + off;
646
647 if (pos_blkg->online)
648 sum += blkg_stat_read(stat);
649 }
650 rcu_read_unlock();
651
652 return sum;
653}
654EXPORT_SYMBOL_GPL(blkg_stat_recursive_sum);
655
656/**
657 * blkg_rwstat_recursive_sum - collect hierarchical blkg_rwstat
658 * @pd: policy private data of interest
659 * @off: offset to the blkg_stat in @pd
660 *
661 * Collect the blkg_rwstat specified by @off from @pd and all its online
662 * descendants and return the sum. The caller must be holding the queue
663 * lock for online tests.
664 */
665struct blkg_rwstat blkg_rwstat_recursive_sum(struct blkg_policy_data *pd,
666 int off)
667{
668 struct blkcg_policy *pol = blkcg_policy[pd->plid];
669 struct blkcg_gq *pos_blkg;
670 struct cgroup_subsys_state *pos_css;
671 struct blkg_rwstat sum = { };
672 int i;
673
674 lockdep_assert_held(pd->blkg->q->queue_lock);
675
676 rcu_read_lock();
677 blkg_for_each_descendant_pre(pos_blkg, pos_css, pd_to_blkg(pd)) {
678 struct blkg_policy_data *pos_pd = blkg_to_pd(pos_blkg, pol);
679 struct blkg_rwstat *rwstat = (void *)pos_pd + off;
680 struct blkg_rwstat tmp;
681
682 if (!pos_blkg->online)
683 continue;
684
685 tmp = blkg_rwstat_read(rwstat);
686
687 for (i = 0; i < BLKG_RWSTAT_NR; i++)
688 sum.cnt[i] += tmp.cnt[i];
689 }
690 rcu_read_unlock();
691
692 return sum;
693}
694EXPORT_SYMBOL_GPL(blkg_rwstat_recursive_sum);
695
696/**
697 * blkg_conf_prep - parse and prepare for per-blkg config update
698 * @blkcg: target block cgroup
699 * @pol: target policy
700 * @input: input string
701 * @ctx: blkg_conf_ctx to be filled
702 *
703 * Parse per-blkg config update from @input and initialize @ctx with the
704 * result. @ctx->blkg points to the blkg to be updated and @ctx->v the new
705 * value. This function returns with RCU read lock and queue lock held and
706 * must be paired with blkg_conf_finish().
707 */
708int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
709 const char *input, struct blkg_conf_ctx *ctx)
710 __acquires(rcu) __acquires(disk->queue->queue_lock)
711{
712 struct gendisk *disk;
713 struct blkcg_gq *blkg;
714 unsigned int major, minor;
715 unsigned long long v;
716 int part, ret;
717
718 if (sscanf(input, "%u:%u %llu", &major, &minor, &v) != 3)
719 return -EINVAL;
720
721 disk = get_gendisk(MKDEV(major, minor), &part);
722 if (!disk || part)
723 return -EINVAL;
724
725 rcu_read_lock();
726 spin_lock_irq(disk->queue->queue_lock);
727
728 if (blkcg_policy_enabled(disk->queue, pol))
729 blkg = blkg_lookup_create(blkcg, disk->queue);
730 else
731 blkg = ERR_PTR(-EINVAL);
732
733 if (IS_ERR(blkg)) {
734 ret = PTR_ERR(blkg);
735 rcu_read_unlock();
736 spin_unlock_irq(disk->queue->queue_lock);
737 put_disk(disk);
738 /*
739 * If queue was bypassing, we should retry. Do so after a
740 * short msleep(). It isn't strictly necessary but queue
741 * can be bypassing for some time and it's always nice to
742 * avoid busy looping.
743 */
744 if (ret == -EBUSY) {
745 msleep(10);
746 ret = restart_syscall();
747 }
748 return ret;
749 }
750
751 ctx->disk = disk;
752 ctx->blkg = blkg;
753 ctx->v = v;
754 return 0;
755}
756EXPORT_SYMBOL_GPL(blkg_conf_prep);
757
758/**
759 * blkg_conf_finish - finish up per-blkg config update
760 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
761 *
762 * Finish up after per-blkg config update. This function must be paired
763 * with blkg_conf_prep().
764 */
765void blkg_conf_finish(struct blkg_conf_ctx *ctx)
766 __releases(ctx->disk->queue->queue_lock) __releases(rcu)
767{
768 spin_unlock_irq(ctx->disk->queue->queue_lock);
769 rcu_read_unlock();
770 put_disk(ctx->disk);
771}
772EXPORT_SYMBOL_GPL(blkg_conf_finish);
773
774struct cftype blkcg_files[] = {
775 {
776 .name = "reset_stats",
777 .write_u64 = blkcg_reset_stats,
778 },
779 { } /* terminate */
780};
781
782/**
783 * blkcg_css_offline - cgroup css_offline callback
784 * @css: css of interest
785 *
786 * This function is called when @css is about to go away and responsible
787 * for shooting down all blkgs associated with @css. blkgs should be
788 * removed while holding both q and blkcg locks. As blkcg lock is nested
789 * inside q lock, this function performs reverse double lock dancing.
790 *
791 * This is the blkcg counterpart of ioc_release_fn().
792 */
793static void blkcg_css_offline(struct cgroup_subsys_state *css)
794{
795 struct blkcg *blkcg = css_to_blkcg(css);
796
797 spin_lock_irq(&blkcg->lock);
798
799 while (!hlist_empty(&blkcg->blkg_list)) {
800 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
801 struct blkcg_gq, blkcg_node);
802 struct request_queue *q = blkg->q;
803
804 if (spin_trylock(q->queue_lock)) {
805 blkg_destroy(blkg);
806 spin_unlock(q->queue_lock);
807 } else {
808 spin_unlock_irq(&blkcg->lock);
809 cpu_relax();
810 spin_lock_irq(&blkcg->lock);
811 }
812 }
813
814 spin_unlock_irq(&blkcg->lock);
815}
816
817static void blkcg_css_free(struct cgroup_subsys_state *css)
818{
819 struct blkcg *blkcg = css_to_blkcg(css);
820
821 if (blkcg != &blkcg_root)
822 kfree(blkcg);
823}
824
825static struct cgroup_subsys_state *
826blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
827{
828 static atomic64_t id_seq = ATOMIC64_INIT(0);
829 struct blkcg *blkcg;
830
831 if (!parent_css) {
832 blkcg = &blkcg_root;
833 goto done;
834 }
835
836 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
837 if (!blkcg)
838 return ERR_PTR(-ENOMEM);
839
840 blkcg->cfq_weight = CFQ_WEIGHT_DEFAULT;
841 blkcg->cfq_leaf_weight = CFQ_WEIGHT_DEFAULT;
842 blkcg->id = atomic64_inc_return(&id_seq); /* root is 0, start from 1 */
843done:
844 spin_lock_init(&blkcg->lock);
845 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_ATOMIC);
846 INIT_HLIST_HEAD(&blkcg->blkg_list);
847
848 return &blkcg->css;
849}
850
851/**
852 * blkcg_init_queue - initialize blkcg part of request queue
853 * @q: request_queue to initialize
854 *
855 * Called from blk_alloc_queue_node(). Responsible for initializing blkcg
856 * part of new request_queue @q.
857 *
858 * RETURNS:
859 * 0 on success, -errno on failure.
860 */
861int blkcg_init_queue(struct request_queue *q)
862{
863 might_sleep();
864
865 return blk_throtl_init(q);
866}
867
868/**
869 * blkcg_drain_queue - drain blkcg part of request_queue
870 * @q: request_queue to drain
871 *
872 * Called from blk_drain_queue(). Responsible for draining blkcg part.
873 */
874void blkcg_drain_queue(struct request_queue *q)
875{
876 lockdep_assert_held(q->queue_lock);
877
878 blk_throtl_drain(q);
879}
880
881/**
882 * blkcg_exit_queue - exit and release blkcg part of request_queue
883 * @q: request_queue being released
884 *
885 * Called from blk_release_queue(). Responsible for exiting blkcg part.
886 */
887void blkcg_exit_queue(struct request_queue *q)
888{
889 spin_lock_irq(q->queue_lock);
890 blkg_destroy_all(q);
891 spin_unlock_irq(q->queue_lock);
892
893 blk_throtl_exit(q);
894}
895
896/*
897 * We cannot support shared io contexts, as we have no mean to support
898 * two tasks with the same ioc in two different groups without major rework
899 * of the main cic data structures. For now we allow a task to change
900 * its cgroup only if it's the only owner of its ioc.
901 */
902static int blkcg_can_attach(struct cgroup_subsys_state *css,
903 struct cgroup_taskset *tset)
904{
905 struct task_struct *task;
906 struct io_context *ioc;
907 int ret = 0;
908
909 /* task_lock() is needed to avoid races with exit_io_context() */
910 cgroup_taskset_for_each(task, tset) {
911 task_lock(task);
912 ioc = task->io_context;
913 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
914 ret = -EINVAL;
915 task_unlock(task);
916 if (ret)
917 break;
918 }
919 return ret;
920}
921
922struct cgroup_subsys blkio_cgrp_subsys = {
923 .css_alloc = blkcg_css_alloc,
924 .css_offline = blkcg_css_offline,
925 .css_free = blkcg_css_free,
926 .can_attach = blkcg_can_attach,
927 .base_cftypes = blkcg_files,
928};
929EXPORT_SYMBOL_GPL(blkio_cgrp_subsys);
930
931/**
932 * blkcg_activate_policy - activate a blkcg policy on a request_queue
933 * @q: request_queue of interest
934 * @pol: blkcg policy to activate
935 *
936 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
937 * bypass mode to populate its blkgs with policy_data for @pol.
938 *
939 * Activation happens with @q bypassed, so nobody would be accessing blkgs
940 * from IO path. Update of each blkg is protected by both queue and blkcg
941 * locks so that holding either lock and testing blkcg_policy_enabled() is
942 * always enough for dereferencing policy data.
943 *
944 * The caller is responsible for synchronizing [de]activations and policy
945 * [un]registerations. Returns 0 on success, -errno on failure.
946 */
947int blkcg_activate_policy(struct request_queue *q,
948 const struct blkcg_policy *pol)
949{
950 LIST_HEAD(pds);
951 struct blkcg_gq *blkg, *new_blkg;
952 struct blkg_policy_data *pd, *n;
953 int cnt = 0, ret;
954 bool preloaded;
955
956 if (blkcg_policy_enabled(q, pol))
957 return 0;
958
959 /* preallocations for root blkg */
960 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
961 if (!new_blkg)
962 return -ENOMEM;
963
964 blk_queue_bypass_start(q);
965
966 preloaded = !radix_tree_preload(GFP_KERNEL);
967
968 /*
969 * Make sure the root blkg exists and count the existing blkgs. As
970 * @q is bypassing at this point, blkg_lookup_create() can't be
971 * used. Open code it.
972 */
973 spin_lock_irq(q->queue_lock);
974
975 rcu_read_lock();
976 blkg = __blkg_lookup(&blkcg_root, q, false);
977 if (blkg)
978 blkg_free(new_blkg);
979 else
980 blkg = blkg_create(&blkcg_root, q, new_blkg);
981 rcu_read_unlock();
982
983 if (preloaded)
984 radix_tree_preload_end();
985
986 if (IS_ERR(blkg)) {
987 ret = PTR_ERR(blkg);
988 goto out_unlock;
989 }
990
991 list_for_each_entry(blkg, &q->blkg_list, q_node)
992 cnt++;
993
994 spin_unlock_irq(q->queue_lock);
995
996 /* allocate policy_data for all existing blkgs */
997 while (cnt--) {
998 pd = kzalloc_node(pol->pd_size, GFP_KERNEL, q->node);
999 if (!pd) {
1000 ret = -ENOMEM;
1001 goto out_free;
1002 }
1003 list_add_tail(&pd->alloc_node, &pds);
1004 }
1005
1006 /*
1007 * Install the allocated pds. With @q bypassing, no new blkg
1008 * should have been created while the queue lock was dropped.
1009 */
1010 spin_lock_irq(q->queue_lock);
1011
1012 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1013 if (WARN_ON(list_empty(&pds))) {
1014 /* umm... this shouldn't happen, just abort */
1015 ret = -ENOMEM;
1016 goto out_unlock;
1017 }
1018 pd = list_first_entry(&pds, struct blkg_policy_data, alloc_node);
1019 list_del_init(&pd->alloc_node);
1020
1021 /* grab blkcg lock too while installing @pd on @blkg */
1022 spin_lock(&blkg->blkcg->lock);
1023
1024 blkg->pd[pol->plid] = pd;
1025 pd->blkg = blkg;
1026 pd->plid = pol->plid;
1027 pol->pd_init_fn(blkg);
1028
1029 spin_unlock(&blkg->blkcg->lock);
1030 }
1031
1032 __set_bit(pol->plid, q->blkcg_pols);
1033 ret = 0;
1034out_unlock:
1035 spin_unlock_irq(q->queue_lock);
1036out_free:
1037 blk_queue_bypass_end(q);
1038 list_for_each_entry_safe(pd, n, &pds, alloc_node)
1039 kfree(pd);
1040 return ret;
1041}
1042EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1043
1044/**
1045 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1046 * @q: request_queue of interest
1047 * @pol: blkcg policy to deactivate
1048 *
1049 * Deactivate @pol on @q. Follows the same synchronization rules as
1050 * blkcg_activate_policy().
1051 */
1052void blkcg_deactivate_policy(struct request_queue *q,
1053 const struct blkcg_policy *pol)
1054{
1055 struct blkcg_gq *blkg;
1056
1057 if (!blkcg_policy_enabled(q, pol))
1058 return;
1059
1060 blk_queue_bypass_start(q);
1061 spin_lock_irq(q->queue_lock);
1062
1063 __clear_bit(pol->plid, q->blkcg_pols);
1064
1065 /* if no policy is left, no need for blkgs - shoot them down */
1066 if (bitmap_empty(q->blkcg_pols, BLKCG_MAX_POLS))
1067 blkg_destroy_all(q);
1068
1069 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1070 /* grab blkcg lock too while removing @pd from @blkg */
1071 spin_lock(&blkg->blkcg->lock);
1072
1073 if (pol->pd_offline_fn)
1074 pol->pd_offline_fn(blkg);
1075 if (pol->pd_exit_fn)
1076 pol->pd_exit_fn(blkg);
1077
1078 kfree(blkg->pd[pol->plid]);
1079 blkg->pd[pol->plid] = NULL;
1080
1081 spin_unlock(&blkg->blkcg->lock);
1082 }
1083
1084 spin_unlock_irq(q->queue_lock);
1085 blk_queue_bypass_end(q);
1086}
1087EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1088
1089/**
1090 * blkcg_policy_register - register a blkcg policy
1091 * @pol: blkcg policy to register
1092 *
1093 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1094 * successful registration. Returns 0 on success and -errno on failure.
1095 */
1096int blkcg_policy_register(struct blkcg_policy *pol)
1097{
1098 int i, ret;
1099
1100 if (WARN_ON(pol->pd_size < sizeof(struct blkg_policy_data)))
1101 return -EINVAL;
1102
1103 mutex_lock(&blkcg_pol_mutex);
1104
1105 /* find an empty slot */
1106 ret = -ENOSPC;
1107 for (i = 0; i < BLKCG_MAX_POLS; i++)
1108 if (!blkcg_policy[i])
1109 break;
1110 if (i >= BLKCG_MAX_POLS)
1111 goto out_unlock;
1112
1113 /* register and update blkgs */
1114 pol->plid = i;
1115 blkcg_policy[i] = pol;
1116
1117 /* everything is in place, add intf files for the new policy */
1118 if (pol->cftypes)
1119 WARN_ON(cgroup_add_cftypes(&blkio_cgrp_subsys, pol->cftypes));
1120 ret = 0;
1121out_unlock:
1122 mutex_unlock(&blkcg_pol_mutex);
1123 return ret;
1124}
1125EXPORT_SYMBOL_GPL(blkcg_policy_register);
1126
1127/**
1128 * blkcg_policy_unregister - unregister a blkcg policy
1129 * @pol: blkcg policy to unregister
1130 *
1131 * Undo blkcg_policy_register(@pol). Might sleep.
1132 */
1133void blkcg_policy_unregister(struct blkcg_policy *pol)
1134{
1135 mutex_lock(&blkcg_pol_mutex);
1136
1137 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1138 goto out_unlock;
1139
1140 /* kill the intf files first */
1141 if (pol->cftypes)
1142 cgroup_rm_cftypes(pol->cftypes);
1143
1144 /* unregister and update blkgs */
1145 blkcg_policy[pol->plid] = NULL;
1146out_unlock:
1147 mutex_unlock(&blkcg_pol_mutex);
1148}
1149EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Common Block IO controller cgroup interface
4 *
5 * Based on ideas and code from CFQ, CFS and BFQ:
6 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
7 *
8 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
9 * Paolo Valente <paolo.valente@unimore.it>
10 *
11 * Copyright (C) 2009 Vivek Goyal <vgoyal@redhat.com>
12 * Nauman Rafique <nauman@google.com>
13 *
14 * For policy-specific per-blkcg data:
15 * Copyright (C) 2015 Paolo Valente <paolo.valente@unimore.it>
16 * Arianna Avanzini <avanzini.arianna@gmail.com>
17 */
18#include <linux/ioprio.h>
19#include <linux/kdev_t.h>
20#include <linux/module.h>
21#include <linux/sched/signal.h>
22#include <linux/err.h>
23#include <linux/blkdev.h>
24#include <linux/backing-dev.h>
25#include <linux/slab.h>
26#include <linux/genhd.h>
27#include <linux/delay.h>
28#include <linux/atomic.h>
29#include <linux/ctype.h>
30#include <linux/blk-cgroup.h>
31#include <linux/tracehook.h>
32#include <linux/psi.h>
33#include "blk.h"
34#include "blk-ioprio.h"
35
36/*
37 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
38 * blkcg_pol_register_mutex nests outside of it and synchronizes entire
39 * policy [un]register operations including cgroup file additions /
40 * removals. Putting cgroup file registration outside blkcg_pol_mutex
41 * allows grabbing it from cgroup callbacks.
42 */
43static DEFINE_MUTEX(blkcg_pol_register_mutex);
44static DEFINE_MUTEX(blkcg_pol_mutex);
45
46struct blkcg blkcg_root;
47EXPORT_SYMBOL_GPL(blkcg_root);
48
49struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
50EXPORT_SYMBOL_GPL(blkcg_root_css);
51
52static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
53
54static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */
55
56bool blkcg_debug_stats = false;
57static struct workqueue_struct *blkcg_punt_bio_wq;
58
59#define BLKG_DESTROY_BATCH_SIZE 64
60
61static bool blkcg_policy_enabled(struct request_queue *q,
62 const struct blkcg_policy *pol)
63{
64 return pol && test_bit(pol->plid, q->blkcg_pols);
65}
66
67/**
68 * blkg_free - free a blkg
69 * @blkg: blkg to free
70 *
71 * Free @blkg which may be partially allocated.
72 */
73static void blkg_free(struct blkcg_gq *blkg)
74{
75 int i;
76
77 if (!blkg)
78 return;
79
80 for (i = 0; i < BLKCG_MAX_POLS; i++)
81 if (blkg->pd[i])
82 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
83
84 free_percpu(blkg->iostat_cpu);
85 percpu_ref_exit(&blkg->refcnt);
86 kfree(blkg);
87}
88
89static void __blkg_release(struct rcu_head *rcu)
90{
91 struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head);
92
93 WARN_ON(!bio_list_empty(&blkg->async_bios));
94
95 /* release the blkcg and parent blkg refs this blkg has been holding */
96 css_put(&blkg->blkcg->css);
97 if (blkg->parent)
98 blkg_put(blkg->parent);
99 blkg_free(blkg);
100}
101
102/*
103 * A group is RCU protected, but having an rcu lock does not mean that one
104 * can access all the fields of blkg and assume these are valid. For
105 * example, don't try to follow throtl_data and request queue links.
106 *
107 * Having a reference to blkg under an rcu allows accesses to only values
108 * local to groups like group stats and group rate limits.
109 */
110static void blkg_release(struct percpu_ref *ref)
111{
112 struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt);
113
114 call_rcu(&blkg->rcu_head, __blkg_release);
115}
116
117static void blkg_async_bio_workfn(struct work_struct *work)
118{
119 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
120 async_bio_work);
121 struct bio_list bios = BIO_EMPTY_LIST;
122 struct bio *bio;
123 struct blk_plug plug;
124 bool need_plug = false;
125
126 /* as long as there are pending bios, @blkg can't go away */
127 spin_lock_bh(&blkg->async_bio_lock);
128 bio_list_merge(&bios, &blkg->async_bios);
129 bio_list_init(&blkg->async_bios);
130 spin_unlock_bh(&blkg->async_bio_lock);
131
132 /* start plug only when bio_list contains at least 2 bios */
133 if (bios.head && bios.head->bi_next) {
134 need_plug = true;
135 blk_start_plug(&plug);
136 }
137 while ((bio = bio_list_pop(&bios)))
138 submit_bio(bio);
139 if (need_plug)
140 blk_finish_plug(&plug);
141}
142
143/**
144 * blkg_alloc - allocate a blkg
145 * @blkcg: block cgroup the new blkg is associated with
146 * @q: request_queue the new blkg is associated with
147 * @gfp_mask: allocation mask to use
148 *
149 * Allocate a new blkg assocating @blkcg and @q.
150 */
151static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
152 gfp_t gfp_mask)
153{
154 struct blkcg_gq *blkg;
155 int i, cpu;
156
157 /* alloc and init base part */
158 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
159 if (!blkg)
160 return NULL;
161
162 if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask))
163 goto err_free;
164
165 blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask);
166 if (!blkg->iostat_cpu)
167 goto err_free;
168
169 blkg->q = q;
170 INIT_LIST_HEAD(&blkg->q_node);
171 spin_lock_init(&blkg->async_bio_lock);
172 bio_list_init(&blkg->async_bios);
173 INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
174 blkg->blkcg = blkcg;
175
176 u64_stats_init(&blkg->iostat.sync);
177 for_each_possible_cpu(cpu)
178 u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync);
179
180 for (i = 0; i < BLKCG_MAX_POLS; i++) {
181 struct blkcg_policy *pol = blkcg_policy[i];
182 struct blkg_policy_data *pd;
183
184 if (!blkcg_policy_enabled(q, pol))
185 continue;
186
187 /* alloc per-policy data and attach it to blkg */
188 pd = pol->pd_alloc_fn(gfp_mask, q, blkcg);
189 if (!pd)
190 goto err_free;
191
192 blkg->pd[i] = pd;
193 pd->blkg = blkg;
194 pd->plid = i;
195 }
196
197 return blkg;
198
199err_free:
200 blkg_free(blkg);
201 return NULL;
202}
203
204struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg,
205 struct request_queue *q, bool update_hint)
206{
207 struct blkcg_gq *blkg;
208
209 /*
210 * Hint didn't match. Look up from the radix tree. Note that the
211 * hint can only be updated under queue_lock as otherwise @blkg
212 * could have already been removed from blkg_tree. The caller is
213 * responsible for grabbing queue_lock if @update_hint.
214 */
215 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
216 if (blkg && blkg->q == q) {
217 if (update_hint) {
218 lockdep_assert_held(&q->queue_lock);
219 rcu_assign_pointer(blkcg->blkg_hint, blkg);
220 }
221 return blkg;
222 }
223
224 return NULL;
225}
226EXPORT_SYMBOL_GPL(blkg_lookup_slowpath);
227
228/*
229 * If @new_blkg is %NULL, this function tries to allocate a new one as
230 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
231 */
232static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
233 struct request_queue *q,
234 struct blkcg_gq *new_blkg)
235{
236 struct blkcg_gq *blkg;
237 int i, ret;
238
239 WARN_ON_ONCE(!rcu_read_lock_held());
240 lockdep_assert_held(&q->queue_lock);
241
242 /* request_queue is dying, do not create/recreate a blkg */
243 if (blk_queue_dying(q)) {
244 ret = -ENODEV;
245 goto err_free_blkg;
246 }
247
248 /* blkg holds a reference to blkcg */
249 if (!css_tryget_online(&blkcg->css)) {
250 ret = -ENODEV;
251 goto err_free_blkg;
252 }
253
254 /* allocate */
255 if (!new_blkg) {
256 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN);
257 if (unlikely(!new_blkg)) {
258 ret = -ENOMEM;
259 goto err_put_css;
260 }
261 }
262 blkg = new_blkg;
263
264 /* link parent */
265 if (blkcg_parent(blkcg)) {
266 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
267 if (WARN_ON_ONCE(!blkg->parent)) {
268 ret = -ENODEV;
269 goto err_put_css;
270 }
271 blkg_get(blkg->parent);
272 }
273
274 /* invoke per-policy init */
275 for (i = 0; i < BLKCG_MAX_POLS; i++) {
276 struct blkcg_policy *pol = blkcg_policy[i];
277
278 if (blkg->pd[i] && pol->pd_init_fn)
279 pol->pd_init_fn(blkg->pd[i]);
280 }
281
282 /* insert */
283 spin_lock(&blkcg->lock);
284 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
285 if (likely(!ret)) {
286 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
287 list_add(&blkg->q_node, &q->blkg_list);
288
289 for (i = 0; i < BLKCG_MAX_POLS; i++) {
290 struct blkcg_policy *pol = blkcg_policy[i];
291
292 if (blkg->pd[i] && pol->pd_online_fn)
293 pol->pd_online_fn(blkg->pd[i]);
294 }
295 }
296 blkg->online = true;
297 spin_unlock(&blkcg->lock);
298
299 if (!ret)
300 return blkg;
301
302 /* @blkg failed fully initialized, use the usual release path */
303 blkg_put(blkg);
304 return ERR_PTR(ret);
305
306err_put_css:
307 css_put(&blkcg->css);
308err_free_blkg:
309 blkg_free(new_blkg);
310 return ERR_PTR(ret);
311}
312
313/**
314 * blkg_lookup_create - lookup blkg, try to create one if not there
315 * @blkcg: blkcg of interest
316 * @q: request_queue of interest
317 *
318 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
319 * create one. blkg creation is performed recursively from blkcg_root such
320 * that all non-root blkg's have access to the parent blkg. This function
321 * should be called under RCU read lock and takes @q->queue_lock.
322 *
323 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
324 * down from root.
325 */
326static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
327 struct request_queue *q)
328{
329 struct blkcg_gq *blkg;
330 unsigned long flags;
331
332 WARN_ON_ONCE(!rcu_read_lock_held());
333
334 blkg = blkg_lookup(blkcg, q);
335 if (blkg)
336 return blkg;
337
338 spin_lock_irqsave(&q->queue_lock, flags);
339 blkg = __blkg_lookup(blkcg, q, true);
340 if (blkg)
341 goto found;
342
343 /*
344 * Create blkgs walking down from blkcg_root to @blkcg, so that all
345 * non-root blkgs have access to their parents. Returns the closest
346 * blkg to the intended blkg should blkg_create() fail.
347 */
348 while (true) {
349 struct blkcg *pos = blkcg;
350 struct blkcg *parent = blkcg_parent(blkcg);
351 struct blkcg_gq *ret_blkg = q->root_blkg;
352
353 while (parent) {
354 blkg = __blkg_lookup(parent, q, false);
355 if (blkg) {
356 /* remember closest blkg */
357 ret_blkg = blkg;
358 break;
359 }
360 pos = parent;
361 parent = blkcg_parent(parent);
362 }
363
364 blkg = blkg_create(pos, q, NULL);
365 if (IS_ERR(blkg)) {
366 blkg = ret_blkg;
367 break;
368 }
369 if (pos == blkcg)
370 break;
371 }
372
373found:
374 spin_unlock_irqrestore(&q->queue_lock, flags);
375 return blkg;
376}
377
378static void blkg_destroy(struct blkcg_gq *blkg)
379{
380 struct blkcg *blkcg = blkg->blkcg;
381 int i;
382
383 lockdep_assert_held(&blkg->q->queue_lock);
384 lockdep_assert_held(&blkcg->lock);
385
386 /* Something wrong if we are trying to remove same group twice */
387 WARN_ON_ONCE(list_empty(&blkg->q_node));
388 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
389
390 for (i = 0; i < BLKCG_MAX_POLS; i++) {
391 struct blkcg_policy *pol = blkcg_policy[i];
392
393 if (blkg->pd[i] && pol->pd_offline_fn)
394 pol->pd_offline_fn(blkg->pd[i]);
395 }
396
397 blkg->online = false;
398
399 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
400 list_del_init(&blkg->q_node);
401 hlist_del_init_rcu(&blkg->blkcg_node);
402
403 /*
404 * Both setting lookup hint to and clearing it from @blkg are done
405 * under queue_lock. If it's not pointing to @blkg now, it never
406 * will. Hint assignment itself can race safely.
407 */
408 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
409 rcu_assign_pointer(blkcg->blkg_hint, NULL);
410
411 /*
412 * Put the reference taken at the time of creation so that when all
413 * queues are gone, group can be destroyed.
414 */
415 percpu_ref_kill(&blkg->refcnt);
416}
417
418/**
419 * blkg_destroy_all - destroy all blkgs associated with a request_queue
420 * @q: request_queue of interest
421 *
422 * Destroy all blkgs associated with @q.
423 */
424static void blkg_destroy_all(struct request_queue *q)
425{
426 struct blkcg_gq *blkg, *n;
427 int count = BLKG_DESTROY_BATCH_SIZE;
428
429restart:
430 spin_lock_irq(&q->queue_lock);
431 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
432 struct blkcg *blkcg = blkg->blkcg;
433
434 spin_lock(&blkcg->lock);
435 blkg_destroy(blkg);
436 spin_unlock(&blkcg->lock);
437
438 /*
439 * in order to avoid holding the spin lock for too long, release
440 * it when a batch of blkgs are destroyed.
441 */
442 if (!(--count)) {
443 count = BLKG_DESTROY_BATCH_SIZE;
444 spin_unlock_irq(&q->queue_lock);
445 cond_resched();
446 goto restart;
447 }
448 }
449
450 q->root_blkg = NULL;
451 spin_unlock_irq(&q->queue_lock);
452}
453
454static int blkcg_reset_stats(struct cgroup_subsys_state *css,
455 struct cftype *cftype, u64 val)
456{
457 struct blkcg *blkcg = css_to_blkcg(css);
458 struct blkcg_gq *blkg;
459 int i, cpu;
460
461 mutex_lock(&blkcg_pol_mutex);
462 spin_lock_irq(&blkcg->lock);
463
464 /*
465 * Note that stat reset is racy - it doesn't synchronize against
466 * stat updates. This is a debug feature which shouldn't exist
467 * anyway. If you get hit by a race, retry.
468 */
469 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
470 for_each_possible_cpu(cpu) {
471 struct blkg_iostat_set *bis =
472 per_cpu_ptr(blkg->iostat_cpu, cpu);
473 memset(bis, 0, sizeof(*bis));
474 }
475 memset(&blkg->iostat, 0, sizeof(blkg->iostat));
476
477 for (i = 0; i < BLKCG_MAX_POLS; i++) {
478 struct blkcg_policy *pol = blkcg_policy[i];
479
480 if (blkg->pd[i] && pol->pd_reset_stats_fn)
481 pol->pd_reset_stats_fn(blkg->pd[i]);
482 }
483 }
484
485 spin_unlock_irq(&blkcg->lock);
486 mutex_unlock(&blkcg_pol_mutex);
487 return 0;
488}
489
490const 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 bdi_dev_name(blkg->q->backing_dev_info);
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/* Performs queue bypass and policy enabled checks then looks up blkg. */
560static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg,
561 const struct blkcg_policy *pol,
562 struct request_queue *q)
563{
564 WARN_ON_ONCE(!rcu_read_lock_held());
565 lockdep_assert_held(&q->queue_lock);
566
567 if (!blkcg_policy_enabled(q, pol))
568 return ERR_PTR(-EOPNOTSUPP);
569 return __blkg_lookup(blkcg, q, true /* update_hint */);
570}
571
572/**
573 * blkcg_conf_open_bdev - parse and open bdev for per-blkg config update
574 * @inputp: input string pointer
575 *
576 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update
577 * from @input and get and return the matching bdev. *@inputp is
578 * updated to point past the device node prefix. Returns an ERR_PTR()
579 * value on error.
580 *
581 * Use this function iff blkg_conf_prep() can't be used for some reason.
582 */
583struct block_device *blkcg_conf_open_bdev(char **inputp)
584{
585 char *input = *inputp;
586 unsigned int major, minor;
587 struct block_device *bdev;
588 int key_len;
589
590 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
591 return ERR_PTR(-EINVAL);
592
593 input += key_len;
594 if (!isspace(*input))
595 return ERR_PTR(-EINVAL);
596 input = skip_spaces(input);
597
598 bdev = blkdev_get_no_open(MKDEV(major, minor));
599 if (!bdev)
600 return ERR_PTR(-ENODEV);
601 if (bdev_is_partition(bdev)) {
602 blkdev_put_no_open(bdev);
603 return ERR_PTR(-ENODEV);
604 }
605
606 *inputp = input;
607 return bdev;
608}
609
610/**
611 * blkg_conf_prep - parse and prepare for per-blkg config update
612 * @blkcg: target block cgroup
613 * @pol: target policy
614 * @input: input string
615 * @ctx: blkg_conf_ctx to be filled
616 *
617 * Parse per-blkg config update from @input and initialize @ctx with the
618 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
619 * part of @input following MAJ:MIN. This function returns with RCU read
620 * lock and queue lock held and must be paired with blkg_conf_finish().
621 */
622int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
623 char *input, struct blkg_conf_ctx *ctx)
624 __acquires(rcu) __acquires(&bdev->bd_disk->queue->queue_lock)
625{
626 struct block_device *bdev;
627 struct request_queue *q;
628 struct blkcg_gq *blkg;
629 int ret;
630
631 bdev = blkcg_conf_open_bdev(&input);
632 if (IS_ERR(bdev))
633 return PTR_ERR(bdev);
634
635 q = bdev->bd_disk->queue;
636
637 rcu_read_lock();
638 spin_lock_irq(&q->queue_lock);
639
640 blkg = blkg_lookup_check(blkcg, pol, q);
641 if (IS_ERR(blkg)) {
642 ret = PTR_ERR(blkg);
643 goto fail_unlock;
644 }
645
646 if (blkg)
647 goto success;
648
649 /*
650 * Create blkgs walking down from blkcg_root to @blkcg, so that all
651 * non-root blkgs have access to their parents.
652 */
653 while (true) {
654 struct blkcg *pos = blkcg;
655 struct blkcg *parent;
656 struct blkcg_gq *new_blkg;
657
658 parent = blkcg_parent(blkcg);
659 while (parent && !__blkg_lookup(parent, q, false)) {
660 pos = parent;
661 parent = blkcg_parent(parent);
662 }
663
664 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
665 spin_unlock_irq(&q->queue_lock);
666 rcu_read_unlock();
667
668 new_blkg = blkg_alloc(pos, q, GFP_KERNEL);
669 if (unlikely(!new_blkg)) {
670 ret = -ENOMEM;
671 goto fail;
672 }
673
674 if (radix_tree_preload(GFP_KERNEL)) {
675 blkg_free(new_blkg);
676 ret = -ENOMEM;
677 goto fail;
678 }
679
680 rcu_read_lock();
681 spin_lock_irq(&q->queue_lock);
682
683 blkg = blkg_lookup_check(pos, pol, q);
684 if (IS_ERR(blkg)) {
685 ret = PTR_ERR(blkg);
686 blkg_free(new_blkg);
687 goto fail_preloaded;
688 }
689
690 if (blkg) {
691 blkg_free(new_blkg);
692 } else {
693 blkg = blkg_create(pos, q, new_blkg);
694 if (IS_ERR(blkg)) {
695 ret = PTR_ERR(blkg);
696 goto fail_preloaded;
697 }
698 }
699
700 radix_tree_preload_end();
701
702 if (pos == blkcg)
703 goto success;
704 }
705success:
706 ctx->bdev = bdev;
707 ctx->blkg = blkg;
708 ctx->body = input;
709 return 0;
710
711fail_preloaded:
712 radix_tree_preload_end();
713fail_unlock:
714 spin_unlock_irq(&q->queue_lock);
715 rcu_read_unlock();
716fail:
717 blkdev_put_no_open(bdev);
718 /*
719 * If queue was bypassing, we should retry. Do so after a
720 * short msleep(). It isn't strictly necessary but queue
721 * can be bypassing for some time and it's always nice to
722 * avoid busy looping.
723 */
724 if (ret == -EBUSY) {
725 msleep(10);
726 ret = restart_syscall();
727 }
728 return ret;
729}
730EXPORT_SYMBOL_GPL(blkg_conf_prep);
731
732/**
733 * blkg_conf_finish - finish up per-blkg config update
734 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
735 *
736 * Finish up after per-blkg config update. This function must be paired
737 * with blkg_conf_prep().
738 */
739void blkg_conf_finish(struct blkg_conf_ctx *ctx)
740 __releases(&ctx->bdev->bd_disk->queue->queue_lock) __releases(rcu)
741{
742 spin_unlock_irq(&ctx->bdev->bd_disk->queue->queue_lock);
743 rcu_read_unlock();
744 blkdev_put_no_open(ctx->bdev);
745}
746EXPORT_SYMBOL_GPL(blkg_conf_finish);
747
748static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
749{
750 int i;
751
752 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
753 dst->bytes[i] = src->bytes[i];
754 dst->ios[i] = src->ios[i];
755 }
756}
757
758static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
759{
760 int i;
761
762 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
763 dst->bytes[i] += src->bytes[i];
764 dst->ios[i] += src->ios[i];
765 }
766}
767
768static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
769{
770 int i;
771
772 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
773 dst->bytes[i] -= src->bytes[i];
774 dst->ios[i] -= src->ios[i];
775 }
776}
777
778static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
779{
780 struct blkcg *blkcg = css_to_blkcg(css);
781 struct blkcg_gq *blkg;
782
783 /* Root-level stats are sourced from system-wide IO stats */
784 if (!cgroup_parent(css->cgroup))
785 return;
786
787 rcu_read_lock();
788
789 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
790 struct blkcg_gq *parent = blkg->parent;
791 struct blkg_iostat_set *bisc = per_cpu_ptr(blkg->iostat_cpu, cpu);
792 struct blkg_iostat cur, delta;
793 unsigned long flags;
794 unsigned int seq;
795
796 /* fetch the current per-cpu values */
797 do {
798 seq = u64_stats_fetch_begin(&bisc->sync);
799 blkg_iostat_set(&cur, &bisc->cur);
800 } while (u64_stats_fetch_retry(&bisc->sync, seq));
801
802 /* propagate percpu delta to global */
803 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
804 blkg_iostat_set(&delta, &cur);
805 blkg_iostat_sub(&delta, &bisc->last);
806 blkg_iostat_add(&blkg->iostat.cur, &delta);
807 blkg_iostat_add(&bisc->last, &delta);
808 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
809
810 /* propagate global delta to parent (unless that's root) */
811 if (parent && parent->parent) {
812 flags = u64_stats_update_begin_irqsave(&parent->iostat.sync);
813 blkg_iostat_set(&delta, &blkg->iostat.cur);
814 blkg_iostat_sub(&delta, &blkg->iostat.last);
815 blkg_iostat_add(&parent->iostat.cur, &delta);
816 blkg_iostat_add(&blkg->iostat.last, &delta);
817 u64_stats_update_end_irqrestore(&parent->iostat.sync, flags);
818 }
819 }
820
821 rcu_read_unlock();
822}
823
824/*
825 * We source root cgroup stats from the system-wide stats to avoid
826 * tracking the same information twice and incurring overhead when no
827 * cgroups are defined. For that reason, cgroup_rstat_flush in
828 * blkcg_print_stat does not actually fill out the iostat in the root
829 * cgroup's blkcg_gq.
830 *
831 * However, we would like to re-use the printing code between the root and
832 * non-root cgroups to the extent possible. For that reason, we simulate
833 * flushing the root cgroup's stats by explicitly filling in the iostat
834 * with disk level statistics.
835 */
836static void blkcg_fill_root_iostats(void)
837{
838 struct class_dev_iter iter;
839 struct device *dev;
840
841 class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
842 while ((dev = class_dev_iter_next(&iter))) {
843 struct block_device *bdev = dev_to_bdev(dev);
844 struct blkcg_gq *blkg =
845 blk_queue_root_blkg(bdev->bd_disk->queue);
846 struct blkg_iostat tmp;
847 int cpu;
848
849 memset(&tmp, 0, sizeof(tmp));
850 for_each_possible_cpu(cpu) {
851 struct disk_stats *cpu_dkstats;
852 unsigned long flags;
853
854 cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu);
855 tmp.ios[BLKG_IOSTAT_READ] +=
856 cpu_dkstats->ios[STAT_READ];
857 tmp.ios[BLKG_IOSTAT_WRITE] +=
858 cpu_dkstats->ios[STAT_WRITE];
859 tmp.ios[BLKG_IOSTAT_DISCARD] +=
860 cpu_dkstats->ios[STAT_DISCARD];
861 // convert sectors to bytes
862 tmp.bytes[BLKG_IOSTAT_READ] +=
863 cpu_dkstats->sectors[STAT_READ] << 9;
864 tmp.bytes[BLKG_IOSTAT_WRITE] +=
865 cpu_dkstats->sectors[STAT_WRITE] << 9;
866 tmp.bytes[BLKG_IOSTAT_DISCARD] +=
867 cpu_dkstats->sectors[STAT_DISCARD] << 9;
868
869 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync);
870 blkg_iostat_set(&blkg->iostat.cur, &tmp);
871 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags);
872 }
873 }
874}
875
876static int blkcg_print_stat(struct seq_file *sf, void *v)
877{
878 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
879 struct blkcg_gq *blkg;
880
881 if (!seq_css(sf)->parent)
882 blkcg_fill_root_iostats();
883 else
884 cgroup_rstat_flush(blkcg->css.cgroup);
885
886 rcu_read_lock();
887
888 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
889 struct blkg_iostat_set *bis = &blkg->iostat;
890 const char *dname;
891 char *buf;
892 u64 rbytes, wbytes, rios, wios, dbytes, dios;
893 size_t size = seq_get_buf(sf, &buf), off = 0;
894 int i;
895 bool has_stats = false;
896 unsigned seq;
897
898 spin_lock_irq(&blkg->q->queue_lock);
899
900 if (!blkg->online)
901 goto skip;
902
903 dname = blkg_dev_name(blkg);
904 if (!dname)
905 goto skip;
906
907 /*
908 * Hooray string manipulation, count is the size written NOT
909 * INCLUDING THE \0, so size is now count+1 less than what we
910 * had before, but we want to start writing the next bit from
911 * the \0 so we only add count to buf.
912 */
913 off += scnprintf(buf+off, size-off, "%s ", dname);
914
915 do {
916 seq = u64_stats_fetch_begin(&bis->sync);
917
918 rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
919 wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
920 dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
921 rios = bis->cur.ios[BLKG_IOSTAT_READ];
922 wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
923 dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
924 } while (u64_stats_fetch_retry(&bis->sync, seq));
925
926 if (rbytes || wbytes || rios || wios) {
927 has_stats = true;
928 off += scnprintf(buf+off, size-off,
929 "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
930 rbytes, wbytes, rios, wios,
931 dbytes, dios);
932 }
933
934 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
935 has_stats = true;
936 off += scnprintf(buf+off, size-off,
937 " use_delay=%d delay_nsec=%llu",
938 atomic_read(&blkg->use_delay),
939 (unsigned long long)atomic64_read(&blkg->delay_nsec));
940 }
941
942 for (i = 0; i < BLKCG_MAX_POLS; i++) {
943 struct blkcg_policy *pol = blkcg_policy[i];
944 size_t written;
945
946 if (!blkg->pd[i] || !pol->pd_stat_fn)
947 continue;
948
949 written = pol->pd_stat_fn(blkg->pd[i], buf+off, size-off);
950 if (written)
951 has_stats = true;
952 off += written;
953 }
954
955 if (has_stats) {
956 if (off < size - 1) {
957 off += scnprintf(buf+off, size-off, "\n");
958 seq_commit(sf, off);
959 } else {
960 seq_commit(sf, -1);
961 }
962 }
963 skip:
964 spin_unlock_irq(&blkg->q->queue_lock);
965 }
966
967 rcu_read_unlock();
968 return 0;
969}
970
971static struct cftype blkcg_files[] = {
972 {
973 .name = "stat",
974 .seq_show = blkcg_print_stat,
975 },
976 { } /* terminate */
977};
978
979static struct cftype blkcg_legacy_files[] = {
980 {
981 .name = "reset_stats",
982 .write_u64 = blkcg_reset_stats,
983 },
984 { } /* terminate */
985};
986
987/*
988 * blkcg destruction is a three-stage process.
989 *
990 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
991 * which offlines writeback. Here we tie the next stage of blkg destruction
992 * to the completion of writeback associated with the blkcg. This lets us
993 * avoid punting potentially large amounts of outstanding writeback to root
994 * while maintaining any ongoing policies. The next stage is triggered when
995 * the nr_cgwbs count goes to zero.
996 *
997 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
998 * and handles the destruction of blkgs. Here the css reference held by
999 * the blkg is put back eventually allowing blkcg_css_free() to be called.
1000 * This work may occur in cgwb_release_workfn() on the cgwb_release
1001 * workqueue. Any submitted ios that fail to get the blkg ref will be
1002 * punted to the root_blkg.
1003 *
1004 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
1005 * This finally frees the blkcg.
1006 */
1007
1008/**
1009 * blkcg_css_offline - cgroup css_offline callback
1010 * @css: css of interest
1011 *
1012 * This function is called when @css is about to go away. Here the cgwbs are
1013 * offlined first and only once writeback associated with the blkcg has
1014 * finished do we start step 2 (see above).
1015 */
1016static void blkcg_css_offline(struct cgroup_subsys_state *css)
1017{
1018 struct blkcg *blkcg = css_to_blkcg(css);
1019
1020 /* this prevents anyone from attaching or migrating to this blkcg */
1021 wb_blkcg_offline(blkcg);
1022
1023 /* put the base online pin allowing step 2 to be triggered */
1024 blkcg_unpin_online(blkcg);
1025}
1026
1027/**
1028 * blkcg_destroy_blkgs - responsible for shooting down blkgs
1029 * @blkcg: blkcg of interest
1030 *
1031 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
1032 * is nested inside q lock, this function performs reverse double lock dancing.
1033 * Destroying the blkgs releases the reference held on the blkcg's css allowing
1034 * blkcg_css_free to eventually be called.
1035 *
1036 * This is the blkcg counterpart of ioc_release_fn().
1037 */
1038void blkcg_destroy_blkgs(struct blkcg *blkcg)
1039{
1040 might_sleep();
1041
1042 spin_lock_irq(&blkcg->lock);
1043
1044 while (!hlist_empty(&blkcg->blkg_list)) {
1045 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1046 struct blkcg_gq, blkcg_node);
1047 struct request_queue *q = blkg->q;
1048
1049 if (need_resched() || !spin_trylock(&q->queue_lock)) {
1050 /*
1051 * Given that the system can accumulate a huge number
1052 * of blkgs in pathological cases, check to see if we
1053 * need to rescheduling to avoid softlockup.
1054 */
1055 spin_unlock_irq(&blkcg->lock);
1056 cond_resched();
1057 spin_lock_irq(&blkcg->lock);
1058 continue;
1059 }
1060
1061 blkg_destroy(blkg);
1062 spin_unlock(&q->queue_lock);
1063 }
1064
1065 spin_unlock_irq(&blkcg->lock);
1066}
1067
1068static void blkcg_css_free(struct cgroup_subsys_state *css)
1069{
1070 struct blkcg *blkcg = css_to_blkcg(css);
1071 int i;
1072
1073 mutex_lock(&blkcg_pol_mutex);
1074
1075 list_del(&blkcg->all_blkcgs_node);
1076
1077 for (i = 0; i < BLKCG_MAX_POLS; i++)
1078 if (blkcg->cpd[i])
1079 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1080
1081 mutex_unlock(&blkcg_pol_mutex);
1082
1083 kfree(blkcg);
1084}
1085
1086static struct cgroup_subsys_state *
1087blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1088{
1089 struct blkcg *blkcg;
1090 struct cgroup_subsys_state *ret;
1091 int i;
1092
1093 mutex_lock(&blkcg_pol_mutex);
1094
1095 if (!parent_css) {
1096 blkcg = &blkcg_root;
1097 } else {
1098 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1099 if (!blkcg) {
1100 ret = ERR_PTR(-ENOMEM);
1101 goto unlock;
1102 }
1103 }
1104
1105 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1106 struct blkcg_policy *pol = blkcg_policy[i];
1107 struct blkcg_policy_data *cpd;
1108
1109 /*
1110 * If the policy hasn't been attached yet, wait for it
1111 * to be attached before doing anything else. Otherwise,
1112 * check if the policy requires any specific per-cgroup
1113 * data: if it does, allocate and initialize it.
1114 */
1115 if (!pol || !pol->cpd_alloc_fn)
1116 continue;
1117
1118 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1119 if (!cpd) {
1120 ret = ERR_PTR(-ENOMEM);
1121 goto free_pd_blkcg;
1122 }
1123 blkcg->cpd[i] = cpd;
1124 cpd->blkcg = blkcg;
1125 cpd->plid = i;
1126 if (pol->cpd_init_fn)
1127 pol->cpd_init_fn(cpd);
1128 }
1129
1130 spin_lock_init(&blkcg->lock);
1131 refcount_set(&blkcg->online_pin, 1);
1132 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1133 INIT_HLIST_HEAD(&blkcg->blkg_list);
1134#ifdef CONFIG_CGROUP_WRITEBACK
1135 INIT_LIST_HEAD(&blkcg->cgwb_list);
1136#endif
1137 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1138
1139 mutex_unlock(&blkcg_pol_mutex);
1140 return &blkcg->css;
1141
1142free_pd_blkcg:
1143 for (i--; i >= 0; i--)
1144 if (blkcg->cpd[i])
1145 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1146
1147 if (blkcg != &blkcg_root)
1148 kfree(blkcg);
1149unlock:
1150 mutex_unlock(&blkcg_pol_mutex);
1151 return ret;
1152}
1153
1154static int blkcg_css_online(struct cgroup_subsys_state *css)
1155{
1156 struct blkcg *blkcg = css_to_blkcg(css);
1157 struct blkcg *parent = blkcg_parent(blkcg);
1158
1159 /*
1160 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1161 * don't go offline while cgwbs are still active on them. Pin the
1162 * parent so that offline always happens towards the root.
1163 */
1164 if (parent)
1165 blkcg_pin_online(parent);
1166 return 0;
1167}
1168
1169/**
1170 * blkcg_init_queue - initialize blkcg part of request queue
1171 * @q: request_queue to initialize
1172 *
1173 * Called from blk_alloc_queue(). Responsible for initializing blkcg
1174 * part of new request_queue @q.
1175 *
1176 * RETURNS:
1177 * 0 on success, -errno on failure.
1178 */
1179int blkcg_init_queue(struct request_queue *q)
1180{
1181 struct blkcg_gq *new_blkg, *blkg;
1182 bool preloaded;
1183 int ret;
1184
1185 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1186 if (!new_blkg)
1187 return -ENOMEM;
1188
1189 preloaded = !radix_tree_preload(GFP_KERNEL);
1190
1191 /* Make sure the root blkg exists. */
1192 rcu_read_lock();
1193 spin_lock_irq(&q->queue_lock);
1194 blkg = blkg_create(&blkcg_root, q, new_blkg);
1195 if (IS_ERR(blkg))
1196 goto err_unlock;
1197 q->root_blkg = blkg;
1198 spin_unlock_irq(&q->queue_lock);
1199 rcu_read_unlock();
1200
1201 if (preloaded)
1202 radix_tree_preload_end();
1203
1204 ret = blk_ioprio_init(q);
1205 if (ret)
1206 goto err_destroy_all;
1207
1208 ret = blk_throtl_init(q);
1209 if (ret)
1210 goto err_destroy_all;
1211
1212 ret = blk_iolatency_init(q);
1213 if (ret) {
1214 blk_throtl_exit(q);
1215 goto err_destroy_all;
1216 }
1217
1218 return 0;
1219
1220err_destroy_all:
1221 blkg_destroy_all(q);
1222 return ret;
1223err_unlock:
1224 spin_unlock_irq(&q->queue_lock);
1225 rcu_read_unlock();
1226 if (preloaded)
1227 radix_tree_preload_end();
1228 return PTR_ERR(blkg);
1229}
1230
1231/**
1232 * blkcg_exit_queue - exit and release blkcg part of request_queue
1233 * @q: request_queue being released
1234 *
1235 * Called from blk_exit_queue(). Responsible for exiting blkcg part.
1236 */
1237void blkcg_exit_queue(struct request_queue *q)
1238{
1239 blkg_destroy_all(q);
1240 blk_throtl_exit(q);
1241}
1242
1243static void blkcg_bind(struct cgroup_subsys_state *root_css)
1244{
1245 int i;
1246
1247 mutex_lock(&blkcg_pol_mutex);
1248
1249 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1250 struct blkcg_policy *pol = blkcg_policy[i];
1251 struct blkcg *blkcg;
1252
1253 if (!pol || !pol->cpd_bind_fn)
1254 continue;
1255
1256 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1257 if (blkcg->cpd[pol->plid])
1258 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1259 }
1260 mutex_unlock(&blkcg_pol_mutex);
1261}
1262
1263static void blkcg_exit(struct task_struct *tsk)
1264{
1265 if (tsk->throttle_queue)
1266 blk_put_queue(tsk->throttle_queue);
1267 tsk->throttle_queue = NULL;
1268}
1269
1270struct cgroup_subsys io_cgrp_subsys = {
1271 .css_alloc = blkcg_css_alloc,
1272 .css_online = blkcg_css_online,
1273 .css_offline = blkcg_css_offline,
1274 .css_free = blkcg_css_free,
1275 .css_rstat_flush = blkcg_rstat_flush,
1276 .bind = blkcg_bind,
1277 .dfl_cftypes = blkcg_files,
1278 .legacy_cftypes = blkcg_legacy_files,
1279 .legacy_name = "blkio",
1280 .exit = blkcg_exit,
1281#ifdef CONFIG_MEMCG
1282 /*
1283 * This ensures that, if available, memcg is automatically enabled
1284 * together on the default hierarchy so that the owner cgroup can
1285 * be retrieved from writeback pages.
1286 */
1287 .depends_on = 1 << memory_cgrp_id,
1288#endif
1289};
1290EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1291
1292/**
1293 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1294 * @q: request_queue of interest
1295 * @pol: blkcg policy to activate
1296 *
1297 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1298 * bypass mode to populate its blkgs with policy_data for @pol.
1299 *
1300 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1301 * from IO path. Update of each blkg is protected by both queue and blkcg
1302 * locks so that holding either lock and testing blkcg_policy_enabled() is
1303 * always enough for dereferencing policy data.
1304 *
1305 * The caller is responsible for synchronizing [de]activations and policy
1306 * [un]registerations. Returns 0 on success, -errno on failure.
1307 */
1308int blkcg_activate_policy(struct request_queue *q,
1309 const struct blkcg_policy *pol)
1310{
1311 struct blkg_policy_data *pd_prealloc = NULL;
1312 struct blkcg_gq *blkg, *pinned_blkg = NULL;
1313 int ret;
1314
1315 if (blkcg_policy_enabled(q, pol))
1316 return 0;
1317
1318 if (queue_is_mq(q))
1319 blk_mq_freeze_queue(q);
1320retry:
1321 spin_lock_irq(&q->queue_lock);
1322
1323 /* blkg_list is pushed at the head, reverse walk to allocate parents first */
1324 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1325 struct blkg_policy_data *pd;
1326
1327 if (blkg->pd[pol->plid])
1328 continue;
1329
1330 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1331 if (blkg == pinned_blkg) {
1332 pd = pd_prealloc;
1333 pd_prealloc = NULL;
1334 } else {
1335 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q,
1336 blkg->blkcg);
1337 }
1338
1339 if (!pd) {
1340 /*
1341 * GFP_NOWAIT failed. Free the existing one and
1342 * prealloc for @blkg w/ GFP_KERNEL.
1343 */
1344 if (pinned_blkg)
1345 blkg_put(pinned_blkg);
1346 blkg_get(blkg);
1347 pinned_blkg = blkg;
1348
1349 spin_unlock_irq(&q->queue_lock);
1350
1351 if (pd_prealloc)
1352 pol->pd_free_fn(pd_prealloc);
1353 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q,
1354 blkg->blkcg);
1355 if (pd_prealloc)
1356 goto retry;
1357 else
1358 goto enomem;
1359 }
1360
1361 blkg->pd[pol->plid] = pd;
1362 pd->blkg = blkg;
1363 pd->plid = pol->plid;
1364 }
1365
1366 /* all allocated, init in the same order */
1367 if (pol->pd_init_fn)
1368 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1369 pol->pd_init_fn(blkg->pd[pol->plid]);
1370
1371 __set_bit(pol->plid, q->blkcg_pols);
1372 ret = 0;
1373
1374 spin_unlock_irq(&q->queue_lock);
1375out:
1376 if (queue_is_mq(q))
1377 blk_mq_unfreeze_queue(q);
1378 if (pinned_blkg)
1379 blkg_put(pinned_blkg);
1380 if (pd_prealloc)
1381 pol->pd_free_fn(pd_prealloc);
1382 return ret;
1383
1384enomem:
1385 /* alloc failed, nothing's initialized yet, free everything */
1386 spin_lock_irq(&q->queue_lock);
1387 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1388 struct blkcg *blkcg = blkg->blkcg;
1389
1390 spin_lock(&blkcg->lock);
1391 if (blkg->pd[pol->plid]) {
1392 pol->pd_free_fn(blkg->pd[pol->plid]);
1393 blkg->pd[pol->plid] = NULL;
1394 }
1395 spin_unlock(&blkcg->lock);
1396 }
1397 spin_unlock_irq(&q->queue_lock);
1398 ret = -ENOMEM;
1399 goto out;
1400}
1401EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1402
1403/**
1404 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1405 * @q: request_queue of interest
1406 * @pol: blkcg policy to deactivate
1407 *
1408 * Deactivate @pol on @q. Follows the same synchronization rules as
1409 * blkcg_activate_policy().
1410 */
1411void blkcg_deactivate_policy(struct request_queue *q,
1412 const struct blkcg_policy *pol)
1413{
1414 struct blkcg_gq *blkg;
1415
1416 if (!blkcg_policy_enabled(q, pol))
1417 return;
1418
1419 if (queue_is_mq(q))
1420 blk_mq_freeze_queue(q);
1421
1422 spin_lock_irq(&q->queue_lock);
1423
1424 __clear_bit(pol->plid, q->blkcg_pols);
1425
1426 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1427 struct blkcg *blkcg = blkg->blkcg;
1428
1429 spin_lock(&blkcg->lock);
1430 if (blkg->pd[pol->plid]) {
1431 if (pol->pd_offline_fn)
1432 pol->pd_offline_fn(blkg->pd[pol->plid]);
1433 pol->pd_free_fn(blkg->pd[pol->plid]);
1434 blkg->pd[pol->plid] = NULL;
1435 }
1436 spin_unlock(&blkcg->lock);
1437 }
1438
1439 spin_unlock_irq(&q->queue_lock);
1440
1441 if (queue_is_mq(q))
1442 blk_mq_unfreeze_queue(q);
1443}
1444EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1445
1446/**
1447 * blkcg_policy_register - register a blkcg policy
1448 * @pol: blkcg policy to register
1449 *
1450 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1451 * successful registration. Returns 0 on success and -errno on failure.
1452 */
1453int blkcg_policy_register(struct blkcg_policy *pol)
1454{
1455 struct blkcg *blkcg;
1456 int i, ret;
1457
1458 mutex_lock(&blkcg_pol_register_mutex);
1459 mutex_lock(&blkcg_pol_mutex);
1460
1461 /* find an empty slot */
1462 ret = -ENOSPC;
1463 for (i = 0; i < BLKCG_MAX_POLS; i++)
1464 if (!blkcg_policy[i])
1465 break;
1466 if (i >= BLKCG_MAX_POLS) {
1467 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1468 goto err_unlock;
1469 }
1470
1471 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1472 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1473 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1474 goto err_unlock;
1475
1476 /* register @pol */
1477 pol->plid = i;
1478 blkcg_policy[pol->plid] = pol;
1479
1480 /* allocate and install cpd's */
1481 if (pol->cpd_alloc_fn) {
1482 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1483 struct blkcg_policy_data *cpd;
1484
1485 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1486 if (!cpd)
1487 goto err_free_cpds;
1488
1489 blkcg->cpd[pol->plid] = cpd;
1490 cpd->blkcg = blkcg;
1491 cpd->plid = pol->plid;
1492 if (pol->cpd_init_fn)
1493 pol->cpd_init_fn(cpd);
1494 }
1495 }
1496
1497 mutex_unlock(&blkcg_pol_mutex);
1498
1499 /* everything is in place, add intf files for the new policy */
1500 if (pol->dfl_cftypes)
1501 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1502 pol->dfl_cftypes));
1503 if (pol->legacy_cftypes)
1504 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1505 pol->legacy_cftypes));
1506 mutex_unlock(&blkcg_pol_register_mutex);
1507 return 0;
1508
1509err_free_cpds:
1510 if (pol->cpd_free_fn) {
1511 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1512 if (blkcg->cpd[pol->plid]) {
1513 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1514 blkcg->cpd[pol->plid] = NULL;
1515 }
1516 }
1517 }
1518 blkcg_policy[pol->plid] = NULL;
1519err_unlock:
1520 mutex_unlock(&blkcg_pol_mutex);
1521 mutex_unlock(&blkcg_pol_register_mutex);
1522 return ret;
1523}
1524EXPORT_SYMBOL_GPL(blkcg_policy_register);
1525
1526/**
1527 * blkcg_policy_unregister - unregister a blkcg policy
1528 * @pol: blkcg policy to unregister
1529 *
1530 * Undo blkcg_policy_register(@pol). Might sleep.
1531 */
1532void blkcg_policy_unregister(struct blkcg_policy *pol)
1533{
1534 struct blkcg *blkcg;
1535
1536 mutex_lock(&blkcg_pol_register_mutex);
1537
1538 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1539 goto out_unlock;
1540
1541 /* kill the intf files first */
1542 if (pol->dfl_cftypes)
1543 cgroup_rm_cftypes(pol->dfl_cftypes);
1544 if (pol->legacy_cftypes)
1545 cgroup_rm_cftypes(pol->legacy_cftypes);
1546
1547 /* remove cpds and unregister */
1548 mutex_lock(&blkcg_pol_mutex);
1549
1550 if (pol->cpd_free_fn) {
1551 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1552 if (blkcg->cpd[pol->plid]) {
1553 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1554 blkcg->cpd[pol->plid] = NULL;
1555 }
1556 }
1557 }
1558 blkcg_policy[pol->plid] = NULL;
1559
1560 mutex_unlock(&blkcg_pol_mutex);
1561out_unlock:
1562 mutex_unlock(&blkcg_pol_register_mutex);
1563}
1564EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1565
1566bool __blkcg_punt_bio_submit(struct bio *bio)
1567{
1568 struct blkcg_gq *blkg = bio->bi_blkg;
1569
1570 /* consume the flag first */
1571 bio->bi_opf &= ~REQ_CGROUP_PUNT;
1572
1573 /* never bounce for the root cgroup */
1574 if (!blkg->parent)
1575 return false;
1576
1577 spin_lock_bh(&blkg->async_bio_lock);
1578 bio_list_add(&blkg->async_bios, bio);
1579 spin_unlock_bh(&blkg->async_bio_lock);
1580
1581 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
1582 return true;
1583}
1584
1585/*
1586 * Scale the accumulated delay based on how long it has been since we updated
1587 * the delay. We only call this when we are adding delay, in case it's been a
1588 * while since we added delay, and when we are checking to see if we need to
1589 * delay a task, to account for any delays that may have occurred.
1590 */
1591static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1592{
1593 u64 old = atomic64_read(&blkg->delay_start);
1594
1595 /* negative use_delay means no scaling, see blkcg_set_delay() */
1596 if (atomic_read(&blkg->use_delay) < 0)
1597 return;
1598
1599 /*
1600 * We only want to scale down every second. The idea here is that we
1601 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1602 * time window. We only want to throttle tasks for recent delay that
1603 * has occurred, in 1 second time windows since that's the maximum
1604 * things can be throttled. We save the current delay window in
1605 * blkg->last_delay so we know what amount is still left to be charged
1606 * to the blkg from this point onward. blkg->last_use keeps track of
1607 * the use_delay counter. The idea is if we're unthrottling the blkg we
1608 * are ok with whatever is happening now, and we can take away more of
1609 * the accumulated delay as we've already throttled enough that
1610 * everybody is happy with their IO latencies.
1611 */
1612 if (time_before64(old + NSEC_PER_SEC, now) &&
1613 atomic64_cmpxchg(&blkg->delay_start, old, now) == old) {
1614 u64 cur = atomic64_read(&blkg->delay_nsec);
1615 u64 sub = min_t(u64, blkg->last_delay, now - old);
1616 int cur_use = atomic_read(&blkg->use_delay);
1617
1618 /*
1619 * We've been unthrottled, subtract a larger chunk of our
1620 * accumulated delay.
1621 */
1622 if (cur_use < blkg->last_use)
1623 sub = max_t(u64, sub, blkg->last_delay >> 1);
1624
1625 /*
1626 * This shouldn't happen, but handle it anyway. Our delay_nsec
1627 * should only ever be growing except here where we subtract out
1628 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1629 * rather not end up with negative numbers.
1630 */
1631 if (unlikely(cur < sub)) {
1632 atomic64_set(&blkg->delay_nsec, 0);
1633 blkg->last_delay = 0;
1634 } else {
1635 atomic64_sub(sub, &blkg->delay_nsec);
1636 blkg->last_delay = cur - sub;
1637 }
1638 blkg->last_use = cur_use;
1639 }
1640}
1641
1642/*
1643 * This is called when we want to actually walk up the hierarchy and check to
1644 * see if we need to throttle, and then actually throttle if there is some
1645 * accumulated delay. This should only be called upon return to user space so
1646 * we're not holding some lock that would induce a priority inversion.
1647 */
1648static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1649{
1650 unsigned long pflags;
1651 bool clamp;
1652 u64 now = ktime_to_ns(ktime_get());
1653 u64 exp;
1654 u64 delay_nsec = 0;
1655 int tok;
1656
1657 while (blkg->parent) {
1658 int use_delay = atomic_read(&blkg->use_delay);
1659
1660 if (use_delay) {
1661 u64 this_delay;
1662
1663 blkcg_scale_delay(blkg, now);
1664 this_delay = atomic64_read(&blkg->delay_nsec);
1665 if (this_delay > delay_nsec) {
1666 delay_nsec = this_delay;
1667 clamp = use_delay > 0;
1668 }
1669 }
1670 blkg = blkg->parent;
1671 }
1672
1673 if (!delay_nsec)
1674 return;
1675
1676 /*
1677 * Let's not sleep for all eternity if we've amassed a huge delay.
1678 * Swapping or metadata IO can accumulate 10's of seconds worth of
1679 * delay, and we want userspace to be able to do _something_ so cap the
1680 * delays at 0.25s. If there's 10's of seconds worth of delay then the
1681 * tasks will be delayed for 0.25 second for every syscall. If
1682 * blkcg_set_delay() was used as indicated by negative use_delay, the
1683 * caller is responsible for regulating the range.
1684 */
1685 if (clamp)
1686 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1687
1688 if (use_memdelay)
1689 psi_memstall_enter(&pflags);
1690
1691 exp = ktime_add_ns(now, delay_nsec);
1692 tok = io_schedule_prepare();
1693 do {
1694 __set_current_state(TASK_KILLABLE);
1695 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1696 break;
1697 } while (!fatal_signal_pending(current));
1698 io_schedule_finish(tok);
1699
1700 if (use_memdelay)
1701 psi_memstall_leave(&pflags);
1702}
1703
1704/**
1705 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1706 *
1707 * This is only called if we've been marked with set_notify_resume(). Obviously
1708 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1709 * check to see if current->throttle_queue is set and if not this doesn't do
1710 * anything. This should only ever be called by the resume code, it's not meant
1711 * to be called by people willy-nilly as it will actually do the work to
1712 * throttle the task if it is setup for throttling.
1713 */
1714void blkcg_maybe_throttle_current(void)
1715{
1716 struct request_queue *q = current->throttle_queue;
1717 struct cgroup_subsys_state *css;
1718 struct blkcg *blkcg;
1719 struct blkcg_gq *blkg;
1720 bool use_memdelay = current->use_memdelay;
1721
1722 if (!q)
1723 return;
1724
1725 current->throttle_queue = NULL;
1726 current->use_memdelay = false;
1727
1728 rcu_read_lock();
1729 css = kthread_blkcg();
1730 if (css)
1731 blkcg = css_to_blkcg(css);
1732 else
1733 blkcg = css_to_blkcg(task_css(current, io_cgrp_id));
1734
1735 if (!blkcg)
1736 goto out;
1737 blkg = blkg_lookup(blkcg, q);
1738 if (!blkg)
1739 goto out;
1740 if (!blkg_tryget(blkg))
1741 goto out;
1742 rcu_read_unlock();
1743
1744 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1745 blkg_put(blkg);
1746 blk_put_queue(q);
1747 return;
1748out:
1749 rcu_read_unlock();
1750 blk_put_queue(q);
1751}
1752
1753/**
1754 * blkcg_schedule_throttle - this task needs to check for throttling
1755 * @q: the request queue IO was submitted on
1756 * @use_memdelay: do we charge this to memory delay for PSI
1757 *
1758 * This is called by the IO controller when we know there's delay accumulated
1759 * for the blkg for this task. We do not pass the blkg because there are places
1760 * we call this that may not have that information, the swapping code for
1761 * instance will only have a request_queue at that point. This set's the
1762 * notify_resume for the task to check and see if it requires throttling before
1763 * returning to user space.
1764 *
1765 * We will only schedule once per syscall. You can call this over and over
1766 * again and it will only do the check once upon return to user space, and only
1767 * throttle once. If the task needs to be throttled again it'll need to be
1768 * re-set at the next time we see the task.
1769 */
1770void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay)
1771{
1772 if (unlikely(current->flags & PF_KTHREAD))
1773 return;
1774
1775 if (current->throttle_queue != q) {
1776 if (!blk_get_queue(q))
1777 return;
1778
1779 if (current->throttle_queue)
1780 blk_put_queue(current->throttle_queue);
1781 current->throttle_queue = q;
1782 }
1783
1784 if (use_memdelay)
1785 current->use_memdelay = use_memdelay;
1786 set_notify_resume(current);
1787}
1788
1789/**
1790 * blkcg_add_delay - add delay to this blkg
1791 * @blkg: blkg of interest
1792 * @now: the current time in nanoseconds
1793 * @delta: how many nanoseconds of delay to add
1794 *
1795 * Charge @delta to the blkg's current delay accumulation. This is used to
1796 * throttle tasks if an IO controller thinks we need more throttling.
1797 */
1798void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1799{
1800 if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
1801 return;
1802 blkcg_scale_delay(blkg, now);
1803 atomic64_add(delta, &blkg->delay_nsec);
1804}
1805
1806/**
1807 * blkg_tryget_closest - try and get a blkg ref on the closet blkg
1808 * @bio: target bio
1809 * @css: target css
1810 *
1811 * As the failure mode here is to walk up the blkg tree, this ensure that the
1812 * blkg->parent pointers are always valid. This returns the blkg that it ended
1813 * up taking a reference on or %NULL if no reference was taken.
1814 */
1815static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio,
1816 struct cgroup_subsys_state *css)
1817{
1818 struct blkcg_gq *blkg, *ret_blkg = NULL;
1819
1820 rcu_read_lock();
1821 blkg = blkg_lookup_create(css_to_blkcg(css),
1822 bio->bi_bdev->bd_disk->queue);
1823 while (blkg) {
1824 if (blkg_tryget(blkg)) {
1825 ret_blkg = blkg;
1826 break;
1827 }
1828 blkg = blkg->parent;
1829 }
1830 rcu_read_unlock();
1831
1832 return ret_blkg;
1833}
1834
1835/**
1836 * bio_associate_blkg_from_css - associate a bio with a specified css
1837 * @bio: target bio
1838 * @css: target css
1839 *
1840 * Associate @bio with the blkg found by combining the css's blkg and the
1841 * request_queue of the @bio. An association failure is handled by walking up
1842 * the blkg tree. Therefore, the blkg associated can be anything between @blkg
1843 * and q->root_blkg. This situation only happens when a cgroup is dying and
1844 * then the remaining bios will spill to the closest alive blkg.
1845 *
1846 * A reference will be taken on the blkg and will be released when @bio is
1847 * freed.
1848 */
1849void bio_associate_blkg_from_css(struct bio *bio,
1850 struct cgroup_subsys_state *css)
1851{
1852 if (bio->bi_blkg)
1853 blkg_put(bio->bi_blkg);
1854
1855 if (css && css->parent) {
1856 bio->bi_blkg = blkg_tryget_closest(bio, css);
1857 } else {
1858 blkg_get(bio->bi_bdev->bd_disk->queue->root_blkg);
1859 bio->bi_blkg = bio->bi_bdev->bd_disk->queue->root_blkg;
1860 }
1861}
1862EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css);
1863
1864/**
1865 * bio_associate_blkg - associate a bio with a blkg
1866 * @bio: target bio
1867 *
1868 * Associate @bio with the blkg found from the bio's css and request_queue.
1869 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is
1870 * already associated, the css is reused and association redone as the
1871 * request_queue may have changed.
1872 */
1873void bio_associate_blkg(struct bio *bio)
1874{
1875 struct cgroup_subsys_state *css;
1876
1877 rcu_read_lock();
1878
1879 if (bio->bi_blkg)
1880 css = &bio_blkcg(bio)->css;
1881 else
1882 css = blkcg_css();
1883
1884 bio_associate_blkg_from_css(bio, css);
1885
1886 rcu_read_unlock();
1887}
1888EXPORT_SYMBOL_GPL(bio_associate_blkg);
1889
1890/**
1891 * bio_clone_blkg_association - clone blkg association from src to dst bio
1892 * @dst: destination bio
1893 * @src: source bio
1894 */
1895void bio_clone_blkg_association(struct bio *dst, struct bio *src)
1896{
1897 if (src->bi_blkg) {
1898 if (dst->bi_blkg)
1899 blkg_put(dst->bi_blkg);
1900 blkg_get(src->bi_blkg);
1901 dst->bi_blkg = src->bi_blkg;
1902 }
1903}
1904EXPORT_SYMBOL_GPL(bio_clone_blkg_association);
1905
1906static int blk_cgroup_io_type(struct bio *bio)
1907{
1908 if (op_is_discard(bio->bi_opf))
1909 return BLKG_IOSTAT_DISCARD;
1910 if (op_is_write(bio->bi_opf))
1911 return BLKG_IOSTAT_WRITE;
1912 return BLKG_IOSTAT_READ;
1913}
1914
1915void blk_cgroup_bio_start(struct bio *bio)
1916{
1917 int rwd = blk_cgroup_io_type(bio), cpu;
1918 struct blkg_iostat_set *bis;
1919 unsigned long flags;
1920
1921 cpu = get_cpu();
1922 bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu);
1923 flags = u64_stats_update_begin_irqsave(&bis->sync);
1924
1925 /*
1926 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
1927 * bio and we would have already accounted for the size of the bio.
1928 */
1929 if (!bio_flagged(bio, BIO_CGROUP_ACCT)) {
1930 bio_set_flag(bio, BIO_CGROUP_ACCT);
1931 bis->cur.bytes[rwd] += bio->bi_iter.bi_size;
1932 }
1933 bis->cur.ios[rwd]++;
1934
1935 u64_stats_update_end_irqrestore(&bis->sync, flags);
1936 if (cgroup_subsys_on_dfl(io_cgrp_subsys))
1937 cgroup_rstat_updated(bio->bi_blkg->blkcg->css.cgroup, cpu);
1938 put_cpu();
1939}
1940
1941static int __init blkcg_init(void)
1942{
1943 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
1944 WQ_MEM_RECLAIM | WQ_FREEZABLE |
1945 WQ_UNBOUND | WQ_SYSFS, 0);
1946 if (!blkcg_punt_bio_wq)
1947 return -ENOMEM;
1948 return 0;
1949}
1950subsys_initcall(blkcg_init);
1951
1952module_param(blkcg_debug_stats, bool, 0644);
1953MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");