Loading...
1/*
2 * Common Block IO controller cgroup interface
3 *
4 * Based on ideas and code from CFQ, CFS and BFQ:
5 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
6 *
7 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
8 * Paolo Valente <paolo.valente@unimore.it>
9 *
10 * Copyright (C) 2009 Vivek Goyal <vgoyal@redhat.com>
11 * Nauman Rafique <nauman@google.com>
12 */
13#include <linux/ioprio.h>
14#include <linux/kdev_t.h>
15#include <linux/module.h>
16#include <linux/err.h>
17#include <linux/blkdev.h>
18#include <linux/slab.h>
19#include <linux/genhd.h>
20#include <linux/delay.h>
21#include <linux/atomic.h>
22#include "blk-cgroup.h"
23#include "blk.h"
24
25#define MAX_KEY_LEN 100
26
27static DEFINE_MUTEX(blkcg_pol_mutex);
28
29struct blkcg blkcg_root = { .cfq_weight = 2 * CFQ_WEIGHT_DEFAULT,
30 .cfq_leaf_weight = 2 * CFQ_WEIGHT_DEFAULT, };
31EXPORT_SYMBOL_GPL(blkcg_root);
32
33static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
34
35static bool blkcg_policy_enabled(struct request_queue *q,
36 const struct blkcg_policy *pol)
37{
38 return pol && test_bit(pol->plid, q->blkcg_pols);
39}
40
41/**
42 * blkg_free - free a blkg
43 * @blkg: blkg to free
44 *
45 * Free @blkg which may be partially allocated.
46 */
47static void blkg_free(struct blkcg_gq *blkg)
48{
49 int i;
50
51 if (!blkg)
52 return;
53
54 for (i = 0; i < BLKCG_MAX_POLS; i++)
55 kfree(blkg->pd[i]);
56
57 blk_exit_rl(&blkg->rl);
58 kfree(blkg);
59}
60
61/**
62 * blkg_alloc - allocate a blkg
63 * @blkcg: block cgroup the new blkg is associated with
64 * @q: request_queue the new blkg is associated with
65 * @gfp_mask: allocation mask to use
66 *
67 * Allocate a new blkg assocating @blkcg and @q.
68 */
69static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
70 gfp_t gfp_mask)
71{
72 struct blkcg_gq *blkg;
73 int i;
74
75 /* alloc and init base part */
76 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
77 if (!blkg)
78 return NULL;
79
80 blkg->q = q;
81 INIT_LIST_HEAD(&blkg->q_node);
82 blkg->blkcg = blkcg;
83 blkg->refcnt = 1;
84
85 /* root blkg uses @q->root_rl, init rl only for !root blkgs */
86 if (blkcg != &blkcg_root) {
87 if (blk_init_rl(&blkg->rl, q, gfp_mask))
88 goto err_free;
89 blkg->rl.blkg = blkg;
90 }
91
92 for (i = 0; i < BLKCG_MAX_POLS; i++) {
93 struct blkcg_policy *pol = blkcg_policy[i];
94 struct blkg_policy_data *pd;
95
96 if (!blkcg_policy_enabled(q, pol))
97 continue;
98
99 /* alloc per-policy data and attach it to blkg */
100 pd = kzalloc_node(pol->pd_size, gfp_mask, q->node);
101 if (!pd)
102 goto err_free;
103
104 blkg->pd[i] = pd;
105 pd->blkg = blkg;
106 pd->plid = i;
107 }
108
109 return blkg;
110
111err_free:
112 blkg_free(blkg);
113 return NULL;
114}
115
116/**
117 * __blkg_lookup - internal version of blkg_lookup()
118 * @blkcg: blkcg of interest
119 * @q: request_queue of interest
120 * @update_hint: whether to update lookup hint with the result or not
121 *
122 * This is internal version and shouldn't be used by policy
123 * implementations. Looks up blkgs for the @blkcg - @q pair regardless of
124 * @q's bypass state. If @update_hint is %true, the caller should be
125 * holding @q->queue_lock and lookup hint is updated on success.
126 */
127struct blkcg_gq *__blkg_lookup(struct blkcg *blkcg, struct request_queue *q,
128 bool update_hint)
129{
130 struct blkcg_gq *blkg;
131
132 blkg = rcu_dereference(blkcg->blkg_hint);
133 if (blkg && blkg->q == q)
134 return blkg;
135
136 /*
137 * Hint didn't match. Look up from the radix tree. Note that the
138 * hint can only be updated under queue_lock as otherwise @blkg
139 * could have already been removed from blkg_tree. The caller is
140 * responsible for grabbing queue_lock if @update_hint.
141 */
142 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
143 if (blkg && blkg->q == q) {
144 if (update_hint) {
145 lockdep_assert_held(q->queue_lock);
146 rcu_assign_pointer(blkcg->blkg_hint, blkg);
147 }
148 return blkg;
149 }
150
151 return NULL;
152}
153
154/**
155 * blkg_lookup - lookup blkg for the specified blkcg - q pair
156 * @blkcg: blkcg of interest
157 * @q: request_queue of interest
158 *
159 * Lookup blkg for the @blkcg - @q pair. This function should be called
160 * under RCU read lock and is guaranteed to return %NULL if @q is bypassing
161 * - see blk_queue_bypass_start() for details.
162 */
163struct blkcg_gq *blkg_lookup(struct blkcg *blkcg, struct request_queue *q)
164{
165 WARN_ON_ONCE(!rcu_read_lock_held());
166
167 if (unlikely(blk_queue_bypass(q)))
168 return NULL;
169 return __blkg_lookup(blkcg, q, false);
170}
171EXPORT_SYMBOL_GPL(blkg_lookup);
172
173/*
174 * If @new_blkg is %NULL, this function tries to allocate a new one as
175 * necessary using %GFP_ATOMIC. @new_blkg is always consumed on return.
176 */
177static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
178 struct request_queue *q,
179 struct blkcg_gq *new_blkg)
180{
181 struct blkcg_gq *blkg;
182 int i, ret;
183
184 WARN_ON_ONCE(!rcu_read_lock_held());
185 lockdep_assert_held(q->queue_lock);
186
187 /* blkg holds a reference to blkcg */
188 if (!css_tryget(&blkcg->css)) {
189 ret = -EINVAL;
190 goto err_free_blkg;
191 }
192
193 /* allocate */
194 if (!new_blkg) {
195 new_blkg = blkg_alloc(blkcg, q, GFP_ATOMIC);
196 if (unlikely(!new_blkg)) {
197 ret = -ENOMEM;
198 goto err_put_css;
199 }
200 }
201 blkg = new_blkg;
202
203 /* link parent */
204 if (blkcg_parent(blkcg)) {
205 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
206 if (WARN_ON_ONCE(!blkg->parent)) {
207 ret = -EINVAL;
208 goto err_put_css;
209 }
210 blkg_get(blkg->parent);
211 }
212
213 /* invoke per-policy init */
214 for (i = 0; i < BLKCG_MAX_POLS; i++) {
215 struct blkcg_policy *pol = blkcg_policy[i];
216
217 if (blkg->pd[i] && pol->pd_init_fn)
218 pol->pd_init_fn(blkg);
219 }
220
221 /* insert */
222 spin_lock(&blkcg->lock);
223 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
224 if (likely(!ret)) {
225 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
226 list_add(&blkg->q_node, &q->blkg_list);
227
228 for (i = 0; i < BLKCG_MAX_POLS; i++) {
229 struct blkcg_policy *pol = blkcg_policy[i];
230
231 if (blkg->pd[i] && pol->pd_online_fn)
232 pol->pd_online_fn(blkg);
233 }
234 }
235 blkg->online = true;
236 spin_unlock(&blkcg->lock);
237
238 if (!ret) {
239 if (blkcg == &blkcg_root) {
240 q->root_blkg = blkg;
241 q->root_rl.blkg = blkg;
242 }
243 return blkg;
244 }
245
246 /* @blkg failed fully initialized, use the usual release path */
247 blkg_put(blkg);
248 return ERR_PTR(ret);
249
250err_put_css:
251 css_put(&blkcg->css);
252err_free_blkg:
253 blkg_free(new_blkg);
254 return ERR_PTR(ret);
255}
256
257/**
258 * blkg_lookup_create - lookup blkg, try to create one if not there
259 * @blkcg: blkcg of interest
260 * @q: request_queue of interest
261 *
262 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
263 * create one. blkg creation is performed recursively from blkcg_root such
264 * that all non-root blkg's have access to the parent blkg. This function
265 * should be called under RCU read lock and @q->queue_lock.
266 *
267 * Returns pointer to the looked up or created blkg on success, ERR_PTR()
268 * value on error. If @q is dead, returns ERR_PTR(-EINVAL). If @q is not
269 * dead and bypassing, returns ERR_PTR(-EBUSY).
270 */
271struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
272 struct request_queue *q)
273{
274 struct blkcg_gq *blkg;
275
276 WARN_ON_ONCE(!rcu_read_lock_held());
277 lockdep_assert_held(q->queue_lock);
278
279 /*
280 * This could be the first entry point of blkcg implementation and
281 * we shouldn't allow anything to go through for a bypassing queue.
282 */
283 if (unlikely(blk_queue_bypass(q)))
284 return ERR_PTR(blk_queue_dying(q) ? -EINVAL : -EBUSY);
285
286 blkg = __blkg_lookup(blkcg, q, true);
287 if (blkg)
288 return blkg;
289
290 /*
291 * Create blkgs walking down from blkcg_root to @blkcg, so that all
292 * non-root blkgs have access to their parents.
293 */
294 while (true) {
295 struct blkcg *pos = blkcg;
296 struct blkcg *parent = blkcg_parent(blkcg);
297
298 while (parent && !__blkg_lookup(parent, q, false)) {
299 pos = parent;
300 parent = blkcg_parent(parent);
301 }
302
303 blkg = blkg_create(pos, q, NULL);
304 if (pos == blkcg || IS_ERR(blkg))
305 return blkg;
306 }
307}
308EXPORT_SYMBOL_GPL(blkg_lookup_create);
309
310static void blkg_destroy(struct blkcg_gq *blkg)
311{
312 struct blkcg *blkcg = blkg->blkcg;
313 int i;
314
315 lockdep_assert_held(blkg->q->queue_lock);
316 lockdep_assert_held(&blkcg->lock);
317
318 /* Something wrong if we are trying to remove same group twice */
319 WARN_ON_ONCE(list_empty(&blkg->q_node));
320 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
321
322 for (i = 0; i < BLKCG_MAX_POLS; i++) {
323 struct blkcg_policy *pol = blkcg_policy[i];
324
325 if (blkg->pd[i] && pol->pd_offline_fn)
326 pol->pd_offline_fn(blkg);
327 }
328 blkg->online = false;
329
330 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
331 list_del_init(&blkg->q_node);
332 hlist_del_init_rcu(&blkg->blkcg_node);
333
334 /*
335 * Both setting lookup hint to and clearing it from @blkg are done
336 * under queue_lock. If it's not pointing to @blkg now, it never
337 * will. Hint assignment itself can race safely.
338 */
339 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
340 rcu_assign_pointer(blkcg->blkg_hint, NULL);
341
342 /*
343 * If root blkg is destroyed. Just clear the pointer since root_rl
344 * does not take reference on root blkg.
345 */
346 if (blkcg == &blkcg_root) {
347 blkg->q->root_blkg = NULL;
348 blkg->q->root_rl.blkg = NULL;
349 }
350
351 /*
352 * Put the reference taken at the time of creation so that when all
353 * queues are gone, group can be destroyed.
354 */
355 blkg_put(blkg);
356}
357
358/**
359 * blkg_destroy_all - destroy all blkgs associated with a request_queue
360 * @q: request_queue of interest
361 *
362 * Destroy all blkgs associated with @q.
363 */
364static void blkg_destroy_all(struct request_queue *q)
365{
366 struct blkcg_gq *blkg, *n;
367
368 lockdep_assert_held(q->queue_lock);
369
370 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
371 struct blkcg *blkcg = blkg->blkcg;
372
373 spin_lock(&blkcg->lock);
374 blkg_destroy(blkg);
375 spin_unlock(&blkcg->lock);
376 }
377}
378
379/*
380 * A group is RCU protected, but having an rcu lock does not mean that one
381 * can access all the fields of blkg and assume these are valid. For
382 * example, don't try to follow throtl_data and request queue links.
383 *
384 * Having a reference to blkg under an rcu allows accesses to only values
385 * local to groups like group stats and group rate limits.
386 */
387void __blkg_release_rcu(struct rcu_head *rcu_head)
388{
389 struct blkcg_gq *blkg = container_of(rcu_head, struct blkcg_gq, rcu_head);
390 int i;
391
392 /* tell policies that this one is being freed */
393 for (i = 0; i < BLKCG_MAX_POLS; i++) {
394 struct blkcg_policy *pol = blkcg_policy[i];
395
396 if (blkg->pd[i] && pol->pd_exit_fn)
397 pol->pd_exit_fn(blkg);
398 }
399
400 /* release the blkcg and parent blkg refs this blkg has been holding */
401 css_put(&blkg->blkcg->css);
402 if (blkg->parent) {
403 spin_lock_irq(blkg->q->queue_lock);
404 blkg_put(blkg->parent);
405 spin_unlock_irq(blkg->q->queue_lock);
406 }
407
408 blkg_free(blkg);
409}
410EXPORT_SYMBOL_GPL(__blkg_release_rcu);
411
412/*
413 * The next function used by blk_queue_for_each_rl(). It's a bit tricky
414 * because the root blkg uses @q->root_rl instead of its own rl.
415 */
416struct request_list *__blk_queue_next_rl(struct request_list *rl,
417 struct request_queue *q)
418{
419 struct list_head *ent;
420 struct blkcg_gq *blkg;
421
422 /*
423 * Determine the current blkg list_head. The first entry is
424 * root_rl which is off @q->blkg_list and mapped to the head.
425 */
426 if (rl == &q->root_rl) {
427 ent = &q->blkg_list;
428 /* There are no more block groups, hence no request lists */
429 if (list_empty(ent))
430 return NULL;
431 } else {
432 blkg = container_of(rl, struct blkcg_gq, rl);
433 ent = &blkg->q_node;
434 }
435
436 /* walk to the next list_head, skip root blkcg */
437 ent = ent->next;
438 if (ent == &q->root_blkg->q_node)
439 ent = ent->next;
440 if (ent == &q->blkg_list)
441 return NULL;
442
443 blkg = container_of(ent, struct blkcg_gq, q_node);
444 return &blkg->rl;
445}
446
447static int blkcg_reset_stats(struct cgroup_subsys_state *css,
448 struct cftype *cftype, u64 val)
449{
450 struct blkcg *blkcg = css_to_blkcg(css);
451 struct blkcg_gq *blkg;
452 int i;
453
454 /*
455 * XXX: We invoke cgroup_add/rm_cftypes() under blkcg_pol_mutex
456 * which ends up putting cgroup's internal cgroup_tree_mutex under
457 * it; however, cgroup_tree_mutex is nested above cgroup file
458 * active protection and grabbing blkcg_pol_mutex from a cgroup
459 * file operation creates a possible circular dependency. cgroup
460 * internal locking is planned to go through further simplification
461 * and this issue should go away soon. For now, let's trylock
462 * blkcg_pol_mutex and restart the write on failure.
463 *
464 * http://lkml.kernel.org/g/5363C04B.4010400@oracle.com
465 */
466 if (!mutex_trylock(&blkcg_pol_mutex))
467 return restart_syscall();
468 spin_lock_irq(&blkcg->lock);
469
470 /*
471 * Note that stat reset is racy - it doesn't synchronize against
472 * stat updates. This is a debug feature which shouldn't exist
473 * anyway. If you get hit by a race, retry.
474 */
475 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
476 for (i = 0; i < BLKCG_MAX_POLS; i++) {
477 struct blkcg_policy *pol = blkcg_policy[i];
478
479 if (blkcg_policy_enabled(blkg->q, pol) &&
480 pol->pd_reset_stats_fn)
481 pol->pd_reset_stats_fn(blkg);
482 }
483 }
484
485 spin_unlock_irq(&blkcg->lock);
486 mutex_unlock(&blkcg_pol_mutex);
487 return 0;
488}
489
490static const char *blkg_dev_name(struct blkcg_gq *blkg)
491{
492 /* some drivers (floppy) instantiate a queue w/o disk registered */
493 if (blkg->q->backing_dev_info.dev)
494 return dev_name(blkg->q->backing_dev_info.dev);
495 return NULL;
496}
497
498/**
499 * blkcg_print_blkgs - helper for printing per-blkg data
500 * @sf: seq_file to print to
501 * @blkcg: blkcg of interest
502 * @prfill: fill function to print out a blkg
503 * @pol: policy in question
504 * @data: data to be passed to @prfill
505 * @show_total: to print out sum of prfill return values or not
506 *
507 * This function invokes @prfill on each blkg of @blkcg if pd for the
508 * policy specified by @pol exists. @prfill is invoked with @sf, the
509 * policy data and @data and the matching queue lock held. If @show_total
510 * is %true, the sum of the return values from @prfill is printed with
511 * "Total" label at the end.
512 *
513 * This is to be used to construct print functions for
514 * cftype->read_seq_string method.
515 */
516void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
517 u64 (*prfill)(struct seq_file *,
518 struct blkg_policy_data *, int),
519 const struct blkcg_policy *pol, int data,
520 bool show_total)
521{
522 struct blkcg_gq *blkg;
523 u64 total = 0;
524
525 rcu_read_lock();
526 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
527 spin_lock_irq(blkg->q->queue_lock);
528 if (blkcg_policy_enabled(blkg->q, pol))
529 total += prfill(sf, blkg->pd[pol->plid], data);
530 spin_unlock_irq(blkg->q->queue_lock);
531 }
532 rcu_read_unlock();
533
534 if (show_total)
535 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
536}
537EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
538
539/**
540 * __blkg_prfill_u64 - prfill helper for a single u64 value
541 * @sf: seq_file to print to
542 * @pd: policy private data of interest
543 * @v: value to print
544 *
545 * Print @v to @sf for the device assocaited with @pd.
546 */
547u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
548{
549 const char *dname = blkg_dev_name(pd->blkg);
550
551 if (!dname)
552 return 0;
553
554 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
555 return v;
556}
557EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
558
559/**
560 * __blkg_prfill_rwstat - prfill helper for a blkg_rwstat
561 * @sf: seq_file to print to
562 * @pd: policy private data of interest
563 * @rwstat: rwstat to print
564 *
565 * Print @rwstat to @sf for the device assocaited with @pd.
566 */
567u64 __blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
568 const struct blkg_rwstat *rwstat)
569{
570 static const char *rwstr[] = {
571 [BLKG_RWSTAT_READ] = "Read",
572 [BLKG_RWSTAT_WRITE] = "Write",
573 [BLKG_RWSTAT_SYNC] = "Sync",
574 [BLKG_RWSTAT_ASYNC] = "Async",
575 };
576 const char *dname = blkg_dev_name(pd->blkg);
577 u64 v;
578 int i;
579
580 if (!dname)
581 return 0;
582
583 for (i = 0; i < BLKG_RWSTAT_NR; i++)
584 seq_printf(sf, "%s %s %llu\n", dname, rwstr[i],
585 (unsigned long long)rwstat->cnt[i]);
586
587 v = rwstat->cnt[BLKG_RWSTAT_READ] + rwstat->cnt[BLKG_RWSTAT_WRITE];
588 seq_printf(sf, "%s Total %llu\n", dname, (unsigned long long)v);
589 return v;
590}
591EXPORT_SYMBOL_GPL(__blkg_prfill_rwstat);
592
593/**
594 * blkg_prfill_stat - prfill callback for blkg_stat
595 * @sf: seq_file to print to
596 * @pd: policy private data of interest
597 * @off: offset to the blkg_stat in @pd
598 *
599 * prfill callback for printing a blkg_stat.
600 */
601u64 blkg_prfill_stat(struct seq_file *sf, struct blkg_policy_data *pd, int off)
602{
603 return __blkg_prfill_u64(sf, pd, blkg_stat_read((void *)pd + off));
604}
605EXPORT_SYMBOL_GPL(blkg_prfill_stat);
606
607/**
608 * blkg_prfill_rwstat - prfill callback for blkg_rwstat
609 * @sf: seq_file to print to
610 * @pd: policy private data of interest
611 * @off: offset to the blkg_rwstat in @pd
612 *
613 * prfill callback for printing a blkg_rwstat.
614 */
615u64 blkg_prfill_rwstat(struct seq_file *sf, struct blkg_policy_data *pd,
616 int off)
617{
618 struct blkg_rwstat rwstat = blkg_rwstat_read((void *)pd + off);
619
620 return __blkg_prfill_rwstat(sf, pd, &rwstat);
621}
622EXPORT_SYMBOL_GPL(blkg_prfill_rwstat);
623
624/**
625 * blkg_stat_recursive_sum - collect hierarchical blkg_stat
626 * @pd: policy private data of interest
627 * @off: offset to the blkg_stat in @pd
628 *
629 * Collect the blkg_stat specified by @off from @pd and all its online
630 * descendants and return the sum. The caller must be holding the queue
631 * lock for online tests.
632 */
633u64 blkg_stat_recursive_sum(struct blkg_policy_data *pd, int off)
634{
635 struct blkcg_policy *pol = blkcg_policy[pd->plid];
636 struct blkcg_gq *pos_blkg;
637 struct cgroup_subsys_state *pos_css;
638 u64 sum = 0;
639
640 lockdep_assert_held(pd->blkg->q->queue_lock);
641
642 rcu_read_lock();
643 blkg_for_each_descendant_pre(pos_blkg, pos_css, pd_to_blkg(pd)) {
644 struct blkg_policy_data *pos_pd = blkg_to_pd(pos_blkg, pol);
645 struct blkg_stat *stat = (void *)pos_pd + off;
646
647 if (pos_blkg->online)
648 sum += blkg_stat_read(stat);
649 }
650 rcu_read_unlock();
651
652 return sum;
653}
654EXPORT_SYMBOL_GPL(blkg_stat_recursive_sum);
655
656/**
657 * blkg_rwstat_recursive_sum - collect hierarchical blkg_rwstat
658 * @pd: policy private data of interest
659 * @off: offset to the blkg_stat in @pd
660 *
661 * Collect the blkg_rwstat specified by @off from @pd and all its online
662 * descendants and return the sum. The caller must be holding the queue
663 * lock for online tests.
664 */
665struct blkg_rwstat blkg_rwstat_recursive_sum(struct blkg_policy_data *pd,
666 int off)
667{
668 struct blkcg_policy *pol = blkcg_policy[pd->plid];
669 struct blkcg_gq *pos_blkg;
670 struct cgroup_subsys_state *pos_css;
671 struct blkg_rwstat sum = { };
672 int i;
673
674 lockdep_assert_held(pd->blkg->q->queue_lock);
675
676 rcu_read_lock();
677 blkg_for_each_descendant_pre(pos_blkg, pos_css, pd_to_blkg(pd)) {
678 struct blkg_policy_data *pos_pd = blkg_to_pd(pos_blkg, pol);
679 struct blkg_rwstat *rwstat = (void *)pos_pd + off;
680 struct blkg_rwstat tmp;
681
682 if (!pos_blkg->online)
683 continue;
684
685 tmp = blkg_rwstat_read(rwstat);
686
687 for (i = 0; i < BLKG_RWSTAT_NR; i++)
688 sum.cnt[i] += tmp.cnt[i];
689 }
690 rcu_read_unlock();
691
692 return sum;
693}
694EXPORT_SYMBOL_GPL(blkg_rwstat_recursive_sum);
695
696/**
697 * blkg_conf_prep - parse and prepare for per-blkg config update
698 * @blkcg: target block cgroup
699 * @pol: target policy
700 * @input: input string
701 * @ctx: blkg_conf_ctx to be filled
702 *
703 * Parse per-blkg config update from @input and initialize @ctx with the
704 * result. @ctx->blkg points to the blkg to be updated and @ctx->v the new
705 * value. This function returns with RCU read lock and queue lock held and
706 * must be paired with blkg_conf_finish().
707 */
708int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
709 const char *input, struct blkg_conf_ctx *ctx)
710 __acquires(rcu) __acquires(disk->queue->queue_lock)
711{
712 struct gendisk *disk;
713 struct blkcg_gq *blkg;
714 unsigned int major, minor;
715 unsigned long long v;
716 int part, ret;
717
718 if (sscanf(input, "%u:%u %llu", &major, &minor, &v) != 3)
719 return -EINVAL;
720
721 disk = get_gendisk(MKDEV(major, minor), &part);
722 if (!disk || part)
723 return -EINVAL;
724
725 rcu_read_lock();
726 spin_lock_irq(disk->queue->queue_lock);
727
728 if (blkcg_policy_enabled(disk->queue, pol))
729 blkg = blkg_lookup_create(blkcg, disk->queue);
730 else
731 blkg = ERR_PTR(-EINVAL);
732
733 if (IS_ERR(blkg)) {
734 ret = PTR_ERR(blkg);
735 rcu_read_unlock();
736 spin_unlock_irq(disk->queue->queue_lock);
737 put_disk(disk);
738 /*
739 * If queue was bypassing, we should retry. Do so after a
740 * short msleep(). It isn't strictly necessary but queue
741 * can be bypassing for some time and it's always nice to
742 * avoid busy looping.
743 */
744 if (ret == -EBUSY) {
745 msleep(10);
746 ret = restart_syscall();
747 }
748 return ret;
749 }
750
751 ctx->disk = disk;
752 ctx->blkg = blkg;
753 ctx->v = v;
754 return 0;
755}
756EXPORT_SYMBOL_GPL(blkg_conf_prep);
757
758/**
759 * blkg_conf_finish - finish up per-blkg config update
760 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
761 *
762 * Finish up after per-blkg config update. This function must be paired
763 * with blkg_conf_prep().
764 */
765void blkg_conf_finish(struct blkg_conf_ctx *ctx)
766 __releases(ctx->disk->queue->queue_lock) __releases(rcu)
767{
768 spin_unlock_irq(ctx->disk->queue->queue_lock);
769 rcu_read_unlock();
770 put_disk(ctx->disk);
771}
772EXPORT_SYMBOL_GPL(blkg_conf_finish);
773
774struct cftype blkcg_files[] = {
775 {
776 .name = "reset_stats",
777 .write_u64 = blkcg_reset_stats,
778 },
779 { } /* terminate */
780};
781
782/**
783 * blkcg_css_offline - cgroup css_offline callback
784 * @css: css of interest
785 *
786 * This function is called when @css is about to go away and responsible
787 * for shooting down all blkgs associated with @css. blkgs should be
788 * removed while holding both q and blkcg locks. As blkcg lock is nested
789 * inside q lock, this function performs reverse double lock dancing.
790 *
791 * This is the blkcg counterpart of ioc_release_fn().
792 */
793static void blkcg_css_offline(struct cgroup_subsys_state *css)
794{
795 struct blkcg *blkcg = css_to_blkcg(css);
796
797 spin_lock_irq(&blkcg->lock);
798
799 while (!hlist_empty(&blkcg->blkg_list)) {
800 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
801 struct blkcg_gq, blkcg_node);
802 struct request_queue *q = blkg->q;
803
804 if (spin_trylock(q->queue_lock)) {
805 blkg_destroy(blkg);
806 spin_unlock(q->queue_lock);
807 } else {
808 spin_unlock_irq(&blkcg->lock);
809 cpu_relax();
810 spin_lock_irq(&blkcg->lock);
811 }
812 }
813
814 spin_unlock_irq(&blkcg->lock);
815}
816
817static void blkcg_css_free(struct cgroup_subsys_state *css)
818{
819 struct blkcg *blkcg = css_to_blkcg(css);
820
821 if (blkcg != &blkcg_root)
822 kfree(blkcg);
823}
824
825static struct cgroup_subsys_state *
826blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
827{
828 static atomic64_t id_seq = ATOMIC64_INIT(0);
829 struct blkcg *blkcg;
830
831 if (!parent_css) {
832 blkcg = &blkcg_root;
833 goto done;
834 }
835
836 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
837 if (!blkcg)
838 return ERR_PTR(-ENOMEM);
839
840 blkcg->cfq_weight = CFQ_WEIGHT_DEFAULT;
841 blkcg->cfq_leaf_weight = CFQ_WEIGHT_DEFAULT;
842 blkcg->id = atomic64_inc_return(&id_seq); /* root is 0, start from 1 */
843done:
844 spin_lock_init(&blkcg->lock);
845 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_ATOMIC);
846 INIT_HLIST_HEAD(&blkcg->blkg_list);
847
848 return &blkcg->css;
849}
850
851/**
852 * blkcg_init_queue - initialize blkcg part of request queue
853 * @q: request_queue to initialize
854 *
855 * Called from blk_alloc_queue_node(). Responsible for initializing blkcg
856 * part of new request_queue @q.
857 *
858 * RETURNS:
859 * 0 on success, -errno on failure.
860 */
861int blkcg_init_queue(struct request_queue *q)
862{
863 might_sleep();
864
865 return blk_throtl_init(q);
866}
867
868/**
869 * blkcg_drain_queue - drain blkcg part of request_queue
870 * @q: request_queue to drain
871 *
872 * Called from blk_drain_queue(). Responsible for draining blkcg part.
873 */
874void blkcg_drain_queue(struct request_queue *q)
875{
876 lockdep_assert_held(q->queue_lock);
877
878 blk_throtl_drain(q);
879}
880
881/**
882 * blkcg_exit_queue - exit and release blkcg part of request_queue
883 * @q: request_queue being released
884 *
885 * Called from blk_release_queue(). Responsible for exiting blkcg part.
886 */
887void blkcg_exit_queue(struct request_queue *q)
888{
889 spin_lock_irq(q->queue_lock);
890 blkg_destroy_all(q);
891 spin_unlock_irq(q->queue_lock);
892
893 blk_throtl_exit(q);
894}
895
896/*
897 * We cannot support shared io contexts, as we have no mean to support
898 * two tasks with the same ioc in two different groups without major rework
899 * of the main cic data structures. For now we allow a task to change
900 * its cgroup only if it's the only owner of its ioc.
901 */
902static int blkcg_can_attach(struct cgroup_subsys_state *css,
903 struct cgroup_taskset *tset)
904{
905 struct task_struct *task;
906 struct io_context *ioc;
907 int ret = 0;
908
909 /* task_lock() is needed to avoid races with exit_io_context() */
910 cgroup_taskset_for_each(task, tset) {
911 task_lock(task);
912 ioc = task->io_context;
913 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
914 ret = -EINVAL;
915 task_unlock(task);
916 if (ret)
917 break;
918 }
919 return ret;
920}
921
922struct cgroup_subsys blkio_cgrp_subsys = {
923 .css_alloc = blkcg_css_alloc,
924 .css_offline = blkcg_css_offline,
925 .css_free = blkcg_css_free,
926 .can_attach = blkcg_can_attach,
927 .base_cftypes = blkcg_files,
928};
929EXPORT_SYMBOL_GPL(blkio_cgrp_subsys);
930
931/**
932 * blkcg_activate_policy - activate a blkcg policy on a request_queue
933 * @q: request_queue of interest
934 * @pol: blkcg policy to activate
935 *
936 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
937 * bypass mode to populate its blkgs with policy_data for @pol.
938 *
939 * Activation happens with @q bypassed, so nobody would be accessing blkgs
940 * from IO path. Update of each blkg is protected by both queue and blkcg
941 * locks so that holding either lock and testing blkcg_policy_enabled() is
942 * always enough for dereferencing policy data.
943 *
944 * The caller is responsible for synchronizing [de]activations and policy
945 * [un]registerations. Returns 0 on success, -errno on failure.
946 */
947int blkcg_activate_policy(struct request_queue *q,
948 const struct blkcg_policy *pol)
949{
950 LIST_HEAD(pds);
951 struct blkcg_gq *blkg, *new_blkg;
952 struct blkg_policy_data *pd, *n;
953 int cnt = 0, ret;
954 bool preloaded;
955
956 if (blkcg_policy_enabled(q, pol))
957 return 0;
958
959 /* preallocations for root blkg */
960 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
961 if (!new_blkg)
962 return -ENOMEM;
963
964 blk_queue_bypass_start(q);
965
966 preloaded = !radix_tree_preload(GFP_KERNEL);
967
968 /*
969 * Make sure the root blkg exists and count the existing blkgs. As
970 * @q is bypassing at this point, blkg_lookup_create() can't be
971 * used. Open code it.
972 */
973 spin_lock_irq(q->queue_lock);
974
975 rcu_read_lock();
976 blkg = __blkg_lookup(&blkcg_root, q, false);
977 if (blkg)
978 blkg_free(new_blkg);
979 else
980 blkg = blkg_create(&blkcg_root, q, new_blkg);
981 rcu_read_unlock();
982
983 if (preloaded)
984 radix_tree_preload_end();
985
986 if (IS_ERR(blkg)) {
987 ret = PTR_ERR(blkg);
988 goto out_unlock;
989 }
990
991 list_for_each_entry(blkg, &q->blkg_list, q_node)
992 cnt++;
993
994 spin_unlock_irq(q->queue_lock);
995
996 /* allocate policy_data for all existing blkgs */
997 while (cnt--) {
998 pd = kzalloc_node(pol->pd_size, GFP_KERNEL, q->node);
999 if (!pd) {
1000 ret = -ENOMEM;
1001 goto out_free;
1002 }
1003 list_add_tail(&pd->alloc_node, &pds);
1004 }
1005
1006 /*
1007 * Install the allocated pds. With @q bypassing, no new blkg
1008 * should have been created while the queue lock was dropped.
1009 */
1010 spin_lock_irq(q->queue_lock);
1011
1012 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1013 if (WARN_ON(list_empty(&pds))) {
1014 /* umm... this shouldn't happen, just abort */
1015 ret = -ENOMEM;
1016 goto out_unlock;
1017 }
1018 pd = list_first_entry(&pds, struct blkg_policy_data, alloc_node);
1019 list_del_init(&pd->alloc_node);
1020
1021 /* grab blkcg lock too while installing @pd on @blkg */
1022 spin_lock(&blkg->blkcg->lock);
1023
1024 blkg->pd[pol->plid] = pd;
1025 pd->blkg = blkg;
1026 pd->plid = pol->plid;
1027 pol->pd_init_fn(blkg);
1028
1029 spin_unlock(&blkg->blkcg->lock);
1030 }
1031
1032 __set_bit(pol->plid, q->blkcg_pols);
1033 ret = 0;
1034out_unlock:
1035 spin_unlock_irq(q->queue_lock);
1036out_free:
1037 blk_queue_bypass_end(q);
1038 list_for_each_entry_safe(pd, n, &pds, alloc_node)
1039 kfree(pd);
1040 return ret;
1041}
1042EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1043
1044/**
1045 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1046 * @q: request_queue of interest
1047 * @pol: blkcg policy to deactivate
1048 *
1049 * Deactivate @pol on @q. Follows the same synchronization rules as
1050 * blkcg_activate_policy().
1051 */
1052void blkcg_deactivate_policy(struct request_queue *q,
1053 const struct blkcg_policy *pol)
1054{
1055 struct blkcg_gq *blkg;
1056
1057 if (!blkcg_policy_enabled(q, pol))
1058 return;
1059
1060 blk_queue_bypass_start(q);
1061 spin_lock_irq(q->queue_lock);
1062
1063 __clear_bit(pol->plid, q->blkcg_pols);
1064
1065 /* if no policy is left, no need for blkgs - shoot them down */
1066 if (bitmap_empty(q->blkcg_pols, BLKCG_MAX_POLS))
1067 blkg_destroy_all(q);
1068
1069 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1070 /* grab blkcg lock too while removing @pd from @blkg */
1071 spin_lock(&blkg->blkcg->lock);
1072
1073 if (pol->pd_offline_fn)
1074 pol->pd_offline_fn(blkg);
1075 if (pol->pd_exit_fn)
1076 pol->pd_exit_fn(blkg);
1077
1078 kfree(blkg->pd[pol->plid]);
1079 blkg->pd[pol->plid] = NULL;
1080
1081 spin_unlock(&blkg->blkcg->lock);
1082 }
1083
1084 spin_unlock_irq(q->queue_lock);
1085 blk_queue_bypass_end(q);
1086}
1087EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1088
1089/**
1090 * blkcg_policy_register - register a blkcg policy
1091 * @pol: blkcg policy to register
1092 *
1093 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1094 * successful registration. Returns 0 on success and -errno on failure.
1095 */
1096int blkcg_policy_register(struct blkcg_policy *pol)
1097{
1098 int i, ret;
1099
1100 if (WARN_ON(pol->pd_size < sizeof(struct blkg_policy_data)))
1101 return -EINVAL;
1102
1103 mutex_lock(&blkcg_pol_mutex);
1104
1105 /* find an empty slot */
1106 ret = -ENOSPC;
1107 for (i = 0; i < BLKCG_MAX_POLS; i++)
1108 if (!blkcg_policy[i])
1109 break;
1110 if (i >= BLKCG_MAX_POLS)
1111 goto out_unlock;
1112
1113 /* register and update blkgs */
1114 pol->plid = i;
1115 blkcg_policy[i] = pol;
1116
1117 /* everything is in place, add intf files for the new policy */
1118 if (pol->cftypes)
1119 WARN_ON(cgroup_add_cftypes(&blkio_cgrp_subsys, pol->cftypes));
1120 ret = 0;
1121out_unlock:
1122 mutex_unlock(&blkcg_pol_mutex);
1123 return ret;
1124}
1125EXPORT_SYMBOL_GPL(blkcg_policy_register);
1126
1127/**
1128 * blkcg_policy_unregister - unregister a blkcg policy
1129 * @pol: blkcg policy to unregister
1130 *
1131 * Undo blkcg_policy_register(@pol). Might sleep.
1132 */
1133void blkcg_policy_unregister(struct blkcg_policy *pol)
1134{
1135 mutex_lock(&blkcg_pol_mutex);
1136
1137 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1138 goto out_unlock;
1139
1140 /* kill the intf files first */
1141 if (pol->cftypes)
1142 cgroup_rm_cftypes(pol->cftypes);
1143
1144 /* unregister and update blkgs */
1145 blkcg_policy[pol->plid] = NULL;
1146out_unlock:
1147 mutex_unlock(&blkcg_pol_mutex);
1148}
1149EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Common Block IO controller cgroup interface
4 *
5 * Based on ideas and code from CFQ, CFS and BFQ:
6 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
7 *
8 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
9 * Paolo Valente <paolo.valente@unimore.it>
10 *
11 * Copyright (C) 2009 Vivek Goyal <vgoyal@redhat.com>
12 * Nauman Rafique <nauman@google.com>
13 *
14 * For policy-specific per-blkcg data:
15 * Copyright (C) 2015 Paolo Valente <paolo.valente@unimore.it>
16 * Arianna Avanzini <avanzini.arianna@gmail.com>
17 */
18#include <linux/ioprio.h>
19#include <linux/kdev_t.h>
20#include <linux/module.h>
21#include <linux/sched/signal.h>
22#include <linux/err.h>
23#include <linux/blkdev.h>
24#include <linux/backing-dev.h>
25#include <linux/slab.h>
26#include <linux/genhd.h>
27#include <linux/delay.h>
28#include <linux/atomic.h>
29#include <linux/ctype.h>
30#include <linux/blk-cgroup.h>
31#include <linux/tracehook.h>
32#include <linux/psi.h>
33#include "blk.h"
34
35#define MAX_KEY_LEN 100
36
37/*
38 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation.
39 * blkcg_pol_register_mutex nests outside of it and synchronizes entire
40 * policy [un]register operations including cgroup file additions /
41 * removals. Putting cgroup file registration outside blkcg_pol_mutex
42 * allows grabbing it from cgroup callbacks.
43 */
44static DEFINE_MUTEX(blkcg_pol_register_mutex);
45static DEFINE_MUTEX(blkcg_pol_mutex);
46
47struct blkcg blkcg_root;
48EXPORT_SYMBOL_GPL(blkcg_root);
49
50struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css;
51EXPORT_SYMBOL_GPL(blkcg_root_css);
52
53static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS];
54
55static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */
56
57bool blkcg_debug_stats = false;
58static struct workqueue_struct *blkcg_punt_bio_wq;
59
60static bool blkcg_policy_enabled(struct request_queue *q,
61 const struct blkcg_policy *pol)
62{
63 return pol && test_bit(pol->plid, q->blkcg_pols);
64}
65
66/**
67 * blkg_free - free a blkg
68 * @blkg: blkg to free
69 *
70 * Free @blkg which may be partially allocated.
71 */
72static void blkg_free(struct blkcg_gq *blkg)
73{
74 int i;
75
76 if (!blkg)
77 return;
78
79 for (i = 0; i < BLKCG_MAX_POLS; i++)
80 if (blkg->pd[i])
81 blkcg_policy[i]->pd_free_fn(blkg->pd[i]);
82
83 free_percpu(blkg->iostat_cpu);
84 percpu_ref_exit(&blkg->refcnt);
85 kfree(blkg);
86}
87
88static void __blkg_release(struct rcu_head *rcu)
89{
90 struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head);
91
92 WARN_ON(!bio_list_empty(&blkg->async_bios));
93
94 /* release the blkcg and parent blkg refs this blkg has been holding */
95 css_put(&blkg->blkcg->css);
96 if (blkg->parent)
97 blkg_put(blkg->parent);
98 blkg_free(blkg);
99}
100
101/*
102 * A group is RCU protected, but having an rcu lock does not mean that one
103 * can access all the fields of blkg and assume these are valid. For
104 * example, don't try to follow throtl_data and request queue links.
105 *
106 * Having a reference to blkg under an rcu allows accesses to only values
107 * local to groups like group stats and group rate limits.
108 */
109static void blkg_release(struct percpu_ref *ref)
110{
111 struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt);
112
113 call_rcu(&blkg->rcu_head, __blkg_release);
114}
115
116static void blkg_async_bio_workfn(struct work_struct *work)
117{
118 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq,
119 async_bio_work);
120 struct bio_list bios = BIO_EMPTY_LIST;
121 struct bio *bio;
122
123 /* as long as there are pending bios, @blkg can't go away */
124 spin_lock_bh(&blkg->async_bio_lock);
125 bio_list_merge(&bios, &blkg->async_bios);
126 bio_list_init(&blkg->async_bios);
127 spin_unlock_bh(&blkg->async_bio_lock);
128
129 while ((bio = bio_list_pop(&bios)))
130 submit_bio(bio);
131}
132
133/**
134 * blkg_alloc - allocate a blkg
135 * @blkcg: block cgroup the new blkg is associated with
136 * @q: request_queue the new blkg is associated with
137 * @gfp_mask: allocation mask to use
138 *
139 * Allocate a new blkg assocating @blkcg and @q.
140 */
141static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q,
142 gfp_t gfp_mask)
143{
144 struct blkcg_gq *blkg;
145 int i, cpu;
146
147 /* alloc and init base part */
148 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node);
149 if (!blkg)
150 return NULL;
151
152 if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask))
153 goto err_free;
154
155 blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask);
156 if (!blkg->iostat_cpu)
157 goto err_free;
158
159 blkg->q = q;
160 INIT_LIST_HEAD(&blkg->q_node);
161 spin_lock_init(&blkg->async_bio_lock);
162 bio_list_init(&blkg->async_bios);
163 INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn);
164 blkg->blkcg = blkcg;
165
166 u64_stats_init(&blkg->iostat.sync);
167 for_each_possible_cpu(cpu)
168 u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync);
169
170 for (i = 0; i < BLKCG_MAX_POLS; i++) {
171 struct blkcg_policy *pol = blkcg_policy[i];
172 struct blkg_policy_data *pd;
173
174 if (!blkcg_policy_enabled(q, pol))
175 continue;
176
177 /* alloc per-policy data and attach it to blkg */
178 pd = pol->pd_alloc_fn(gfp_mask, q, blkcg);
179 if (!pd)
180 goto err_free;
181
182 blkg->pd[i] = pd;
183 pd->blkg = blkg;
184 pd->plid = i;
185 }
186
187 return blkg;
188
189err_free:
190 blkg_free(blkg);
191 return NULL;
192}
193
194struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg,
195 struct request_queue *q, bool update_hint)
196{
197 struct blkcg_gq *blkg;
198
199 /*
200 * Hint didn't match. Look up from the radix tree. Note that the
201 * hint can only be updated under queue_lock as otherwise @blkg
202 * could have already been removed from blkg_tree. The caller is
203 * responsible for grabbing queue_lock if @update_hint.
204 */
205 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id);
206 if (blkg && blkg->q == q) {
207 if (update_hint) {
208 lockdep_assert_held(&q->queue_lock);
209 rcu_assign_pointer(blkcg->blkg_hint, blkg);
210 }
211 return blkg;
212 }
213
214 return NULL;
215}
216EXPORT_SYMBOL_GPL(blkg_lookup_slowpath);
217
218/*
219 * If @new_blkg is %NULL, this function tries to allocate a new one as
220 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return.
221 */
222static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
223 struct request_queue *q,
224 struct blkcg_gq *new_blkg)
225{
226 struct blkcg_gq *blkg;
227 int i, ret;
228
229 WARN_ON_ONCE(!rcu_read_lock_held());
230 lockdep_assert_held(&q->queue_lock);
231
232 /* request_queue is dying, do not create/recreate a blkg */
233 if (blk_queue_dying(q)) {
234 ret = -ENODEV;
235 goto err_free_blkg;
236 }
237
238 /* blkg holds a reference to blkcg */
239 if (!css_tryget_online(&blkcg->css)) {
240 ret = -ENODEV;
241 goto err_free_blkg;
242 }
243
244 /* allocate */
245 if (!new_blkg) {
246 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN);
247 if (unlikely(!new_blkg)) {
248 ret = -ENOMEM;
249 goto err_put_css;
250 }
251 }
252 blkg = new_blkg;
253
254 /* link parent */
255 if (blkcg_parent(blkcg)) {
256 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false);
257 if (WARN_ON_ONCE(!blkg->parent)) {
258 ret = -ENODEV;
259 goto err_put_css;
260 }
261 blkg_get(blkg->parent);
262 }
263
264 /* invoke per-policy init */
265 for (i = 0; i < BLKCG_MAX_POLS; i++) {
266 struct blkcg_policy *pol = blkcg_policy[i];
267
268 if (blkg->pd[i] && pol->pd_init_fn)
269 pol->pd_init_fn(blkg->pd[i]);
270 }
271
272 /* insert */
273 spin_lock(&blkcg->lock);
274 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg);
275 if (likely(!ret)) {
276 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list);
277 list_add(&blkg->q_node, &q->blkg_list);
278
279 for (i = 0; i < BLKCG_MAX_POLS; i++) {
280 struct blkcg_policy *pol = blkcg_policy[i];
281
282 if (blkg->pd[i] && pol->pd_online_fn)
283 pol->pd_online_fn(blkg->pd[i]);
284 }
285 }
286 blkg->online = true;
287 spin_unlock(&blkcg->lock);
288
289 if (!ret)
290 return blkg;
291
292 /* @blkg failed fully initialized, use the usual release path */
293 blkg_put(blkg);
294 return ERR_PTR(ret);
295
296err_put_css:
297 css_put(&blkcg->css);
298err_free_blkg:
299 blkg_free(new_blkg);
300 return ERR_PTR(ret);
301}
302
303/**
304 * blkg_lookup_create - lookup blkg, try to create one if not there
305 * @blkcg: blkcg of interest
306 * @q: request_queue of interest
307 *
308 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to
309 * create one. blkg creation is performed recursively from blkcg_root such
310 * that all non-root blkg's have access to the parent blkg. This function
311 * should be called under RCU read lock and takes @q->queue_lock.
312 *
313 * Returns the blkg or the closest blkg if blkg_create() fails as it walks
314 * down from root.
315 */
316static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg,
317 struct request_queue *q)
318{
319 struct blkcg_gq *blkg;
320 unsigned long flags;
321
322 WARN_ON_ONCE(!rcu_read_lock_held());
323
324 blkg = blkg_lookup(blkcg, q);
325 if (blkg)
326 return blkg;
327
328 spin_lock_irqsave(&q->queue_lock, flags);
329 blkg = __blkg_lookup(blkcg, q, true);
330 if (blkg)
331 goto found;
332
333 /*
334 * Create blkgs walking down from blkcg_root to @blkcg, so that all
335 * non-root blkgs have access to their parents. Returns the closest
336 * blkg to the intended blkg should blkg_create() fail.
337 */
338 while (true) {
339 struct blkcg *pos = blkcg;
340 struct blkcg *parent = blkcg_parent(blkcg);
341 struct blkcg_gq *ret_blkg = q->root_blkg;
342
343 while (parent) {
344 blkg = __blkg_lookup(parent, q, false);
345 if (blkg) {
346 /* remember closest blkg */
347 ret_blkg = blkg;
348 break;
349 }
350 pos = parent;
351 parent = blkcg_parent(parent);
352 }
353
354 blkg = blkg_create(pos, q, NULL);
355 if (IS_ERR(blkg)) {
356 blkg = ret_blkg;
357 break;
358 }
359 if (pos == blkcg)
360 break;
361 }
362
363found:
364 spin_unlock_irqrestore(&q->queue_lock, flags);
365 return blkg;
366}
367
368static void blkg_destroy(struct blkcg_gq *blkg)
369{
370 struct blkcg *blkcg = blkg->blkcg;
371 int i;
372
373 lockdep_assert_held(&blkg->q->queue_lock);
374 lockdep_assert_held(&blkcg->lock);
375
376 /* Something wrong if we are trying to remove same group twice */
377 WARN_ON_ONCE(list_empty(&blkg->q_node));
378 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node));
379
380 for (i = 0; i < BLKCG_MAX_POLS; i++) {
381 struct blkcg_policy *pol = blkcg_policy[i];
382
383 if (blkg->pd[i] && pol->pd_offline_fn)
384 pol->pd_offline_fn(blkg->pd[i]);
385 }
386
387 blkg->online = false;
388
389 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id);
390 list_del_init(&blkg->q_node);
391 hlist_del_init_rcu(&blkg->blkcg_node);
392
393 /*
394 * Both setting lookup hint to and clearing it from @blkg are done
395 * under queue_lock. If it's not pointing to @blkg now, it never
396 * will. Hint assignment itself can race safely.
397 */
398 if (rcu_access_pointer(blkcg->blkg_hint) == blkg)
399 rcu_assign_pointer(blkcg->blkg_hint, NULL);
400
401 /*
402 * Put the reference taken at the time of creation so that when all
403 * queues are gone, group can be destroyed.
404 */
405 percpu_ref_kill(&blkg->refcnt);
406}
407
408/**
409 * blkg_destroy_all - destroy all blkgs associated with a request_queue
410 * @q: request_queue of interest
411 *
412 * Destroy all blkgs associated with @q.
413 */
414static void blkg_destroy_all(struct request_queue *q)
415{
416 struct blkcg_gq *blkg, *n;
417
418 spin_lock_irq(&q->queue_lock);
419 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
420 struct blkcg *blkcg = blkg->blkcg;
421
422 spin_lock(&blkcg->lock);
423 blkg_destroy(blkg);
424 spin_unlock(&blkcg->lock);
425 }
426
427 q->root_blkg = NULL;
428 spin_unlock_irq(&q->queue_lock);
429}
430
431static int blkcg_reset_stats(struct cgroup_subsys_state *css,
432 struct cftype *cftype, u64 val)
433{
434 struct blkcg *blkcg = css_to_blkcg(css);
435 struct blkcg_gq *blkg;
436 int i, cpu;
437
438 mutex_lock(&blkcg_pol_mutex);
439 spin_lock_irq(&blkcg->lock);
440
441 /*
442 * Note that stat reset is racy - it doesn't synchronize against
443 * stat updates. This is a debug feature which shouldn't exist
444 * anyway. If you get hit by a race, retry.
445 */
446 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
447 for_each_possible_cpu(cpu) {
448 struct blkg_iostat_set *bis =
449 per_cpu_ptr(blkg->iostat_cpu, cpu);
450 memset(bis, 0, sizeof(*bis));
451 }
452 memset(&blkg->iostat, 0, sizeof(blkg->iostat));
453
454 for (i = 0; i < BLKCG_MAX_POLS; i++) {
455 struct blkcg_policy *pol = blkcg_policy[i];
456
457 if (blkg->pd[i] && pol->pd_reset_stats_fn)
458 pol->pd_reset_stats_fn(blkg->pd[i]);
459 }
460 }
461
462 spin_unlock_irq(&blkcg->lock);
463 mutex_unlock(&blkcg_pol_mutex);
464 return 0;
465}
466
467const char *blkg_dev_name(struct blkcg_gq *blkg)
468{
469 /* some drivers (floppy) instantiate a queue w/o disk registered */
470 if (blkg->q->backing_dev_info->dev)
471 return bdi_dev_name(blkg->q->backing_dev_info);
472 return NULL;
473}
474
475/**
476 * blkcg_print_blkgs - helper for printing per-blkg data
477 * @sf: seq_file to print to
478 * @blkcg: blkcg of interest
479 * @prfill: fill function to print out a blkg
480 * @pol: policy in question
481 * @data: data to be passed to @prfill
482 * @show_total: to print out sum of prfill return values or not
483 *
484 * This function invokes @prfill on each blkg of @blkcg if pd for the
485 * policy specified by @pol exists. @prfill is invoked with @sf, the
486 * policy data and @data and the matching queue lock held. If @show_total
487 * is %true, the sum of the return values from @prfill is printed with
488 * "Total" label at the end.
489 *
490 * This is to be used to construct print functions for
491 * cftype->read_seq_string method.
492 */
493void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg,
494 u64 (*prfill)(struct seq_file *,
495 struct blkg_policy_data *, int),
496 const struct blkcg_policy *pol, int data,
497 bool show_total)
498{
499 struct blkcg_gq *blkg;
500 u64 total = 0;
501
502 rcu_read_lock();
503 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
504 spin_lock_irq(&blkg->q->queue_lock);
505 if (blkcg_policy_enabled(blkg->q, pol))
506 total += prfill(sf, blkg->pd[pol->plid], data);
507 spin_unlock_irq(&blkg->q->queue_lock);
508 }
509 rcu_read_unlock();
510
511 if (show_total)
512 seq_printf(sf, "Total %llu\n", (unsigned long long)total);
513}
514EXPORT_SYMBOL_GPL(blkcg_print_blkgs);
515
516/**
517 * __blkg_prfill_u64 - prfill helper for a single u64 value
518 * @sf: seq_file to print to
519 * @pd: policy private data of interest
520 * @v: value to print
521 *
522 * Print @v to @sf for the device assocaited with @pd.
523 */
524u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v)
525{
526 const char *dname = blkg_dev_name(pd->blkg);
527
528 if (!dname)
529 return 0;
530
531 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v);
532 return v;
533}
534EXPORT_SYMBOL_GPL(__blkg_prfill_u64);
535
536/* Performs queue bypass and policy enabled checks then looks up blkg. */
537static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg,
538 const struct blkcg_policy *pol,
539 struct request_queue *q)
540{
541 WARN_ON_ONCE(!rcu_read_lock_held());
542 lockdep_assert_held(&q->queue_lock);
543
544 if (!blkcg_policy_enabled(q, pol))
545 return ERR_PTR(-EOPNOTSUPP);
546 return __blkg_lookup(blkcg, q, true /* update_hint */);
547}
548
549/**
550 * blkg_conf_prep - parse and prepare for per-blkg config update
551 * @inputp: input string pointer
552 *
553 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update
554 * from @input and get and return the matching gendisk. *@inputp is
555 * updated to point past the device node prefix. Returns an ERR_PTR()
556 * value on error.
557 *
558 * Use this function iff blkg_conf_prep() can't be used for some reason.
559 */
560struct gendisk *blkcg_conf_get_disk(char **inputp)
561{
562 char *input = *inputp;
563 unsigned int major, minor;
564 struct gendisk *disk;
565 int key_len, part;
566
567 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2)
568 return ERR_PTR(-EINVAL);
569
570 input += key_len;
571 if (!isspace(*input))
572 return ERR_PTR(-EINVAL);
573 input = skip_spaces(input);
574
575 disk = get_gendisk(MKDEV(major, minor), &part);
576 if (!disk)
577 return ERR_PTR(-ENODEV);
578 if (part) {
579 put_disk_and_module(disk);
580 return ERR_PTR(-ENODEV);
581 }
582
583 *inputp = input;
584 return disk;
585}
586
587/**
588 * blkg_conf_prep - parse and prepare for per-blkg config update
589 * @blkcg: target block cgroup
590 * @pol: target policy
591 * @input: input string
592 * @ctx: blkg_conf_ctx to be filled
593 *
594 * Parse per-blkg config update from @input and initialize @ctx with the
595 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the
596 * part of @input following MAJ:MIN. This function returns with RCU read
597 * lock and queue lock held and must be paired with blkg_conf_finish().
598 */
599int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol,
600 char *input, struct blkg_conf_ctx *ctx)
601 __acquires(rcu) __acquires(&disk->queue->queue_lock)
602{
603 struct gendisk *disk;
604 struct request_queue *q;
605 struct blkcg_gq *blkg;
606 int ret;
607
608 disk = blkcg_conf_get_disk(&input);
609 if (IS_ERR(disk))
610 return PTR_ERR(disk);
611
612 q = disk->queue;
613
614 rcu_read_lock();
615 spin_lock_irq(&q->queue_lock);
616
617 blkg = blkg_lookup_check(blkcg, pol, q);
618 if (IS_ERR(blkg)) {
619 ret = PTR_ERR(blkg);
620 goto fail_unlock;
621 }
622
623 if (blkg)
624 goto success;
625
626 /*
627 * Create blkgs walking down from blkcg_root to @blkcg, so that all
628 * non-root blkgs have access to their parents.
629 */
630 while (true) {
631 struct blkcg *pos = blkcg;
632 struct blkcg *parent;
633 struct blkcg_gq *new_blkg;
634
635 parent = blkcg_parent(blkcg);
636 while (parent && !__blkg_lookup(parent, q, false)) {
637 pos = parent;
638 parent = blkcg_parent(parent);
639 }
640
641 /* Drop locks to do new blkg allocation with GFP_KERNEL. */
642 spin_unlock_irq(&q->queue_lock);
643 rcu_read_unlock();
644
645 new_blkg = blkg_alloc(pos, q, GFP_KERNEL);
646 if (unlikely(!new_blkg)) {
647 ret = -ENOMEM;
648 goto fail;
649 }
650
651 rcu_read_lock();
652 spin_lock_irq(&q->queue_lock);
653
654 blkg = blkg_lookup_check(pos, pol, q);
655 if (IS_ERR(blkg)) {
656 ret = PTR_ERR(blkg);
657 goto fail_unlock;
658 }
659
660 if (blkg) {
661 blkg_free(new_blkg);
662 } else {
663 blkg = blkg_create(pos, q, new_blkg);
664 if (IS_ERR(blkg)) {
665 ret = PTR_ERR(blkg);
666 goto fail_unlock;
667 }
668 }
669
670 if (pos == blkcg)
671 goto success;
672 }
673success:
674 ctx->disk = disk;
675 ctx->blkg = blkg;
676 ctx->body = input;
677 return 0;
678
679fail_unlock:
680 spin_unlock_irq(&q->queue_lock);
681 rcu_read_unlock();
682fail:
683 put_disk_and_module(disk);
684 /*
685 * If queue was bypassing, we should retry. Do so after a
686 * short msleep(). It isn't strictly necessary but queue
687 * can be bypassing for some time and it's always nice to
688 * avoid busy looping.
689 */
690 if (ret == -EBUSY) {
691 msleep(10);
692 ret = restart_syscall();
693 }
694 return ret;
695}
696EXPORT_SYMBOL_GPL(blkg_conf_prep);
697
698/**
699 * blkg_conf_finish - finish up per-blkg config update
700 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep()
701 *
702 * Finish up after per-blkg config update. This function must be paired
703 * with blkg_conf_prep().
704 */
705void blkg_conf_finish(struct blkg_conf_ctx *ctx)
706 __releases(&ctx->disk->queue->queue_lock) __releases(rcu)
707{
708 spin_unlock_irq(&ctx->disk->queue->queue_lock);
709 rcu_read_unlock();
710 put_disk_and_module(ctx->disk);
711}
712EXPORT_SYMBOL_GPL(blkg_conf_finish);
713
714static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src)
715{
716 int i;
717
718 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
719 dst->bytes[i] = src->bytes[i];
720 dst->ios[i] = src->ios[i];
721 }
722}
723
724static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src)
725{
726 int i;
727
728 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
729 dst->bytes[i] += src->bytes[i];
730 dst->ios[i] += src->ios[i];
731 }
732}
733
734static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src)
735{
736 int i;
737
738 for (i = 0; i < BLKG_IOSTAT_NR; i++) {
739 dst->bytes[i] -= src->bytes[i];
740 dst->ios[i] -= src->ios[i];
741 }
742}
743
744static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu)
745{
746 struct blkcg *blkcg = css_to_blkcg(css);
747 struct blkcg_gq *blkg;
748
749 rcu_read_lock();
750
751 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
752 struct blkcg_gq *parent = blkg->parent;
753 struct blkg_iostat_set *bisc = per_cpu_ptr(blkg->iostat_cpu, cpu);
754 struct blkg_iostat cur, delta;
755 unsigned int seq;
756
757 /* fetch the current per-cpu values */
758 do {
759 seq = u64_stats_fetch_begin(&bisc->sync);
760 blkg_iostat_set(&cur, &bisc->cur);
761 } while (u64_stats_fetch_retry(&bisc->sync, seq));
762
763 /* propagate percpu delta to global */
764 u64_stats_update_begin(&blkg->iostat.sync);
765 blkg_iostat_set(&delta, &cur);
766 blkg_iostat_sub(&delta, &bisc->last);
767 blkg_iostat_add(&blkg->iostat.cur, &delta);
768 blkg_iostat_add(&bisc->last, &delta);
769 u64_stats_update_end(&blkg->iostat.sync);
770
771 /* propagate global delta to parent */
772 if (parent) {
773 u64_stats_update_begin(&parent->iostat.sync);
774 blkg_iostat_set(&delta, &blkg->iostat.cur);
775 blkg_iostat_sub(&delta, &blkg->iostat.last);
776 blkg_iostat_add(&parent->iostat.cur, &delta);
777 blkg_iostat_add(&blkg->iostat.last, &delta);
778 u64_stats_update_end(&parent->iostat.sync);
779 }
780 }
781
782 rcu_read_unlock();
783}
784
785/*
786 * The rstat algorithms intentionally don't handle the root cgroup to avoid
787 * incurring overhead when no cgroups are defined. For that reason,
788 * cgroup_rstat_flush in blkcg_print_stat does not actually fill out the
789 * iostat in the root cgroup's blkcg_gq.
790 *
791 * However, we would like to re-use the printing code between the root and
792 * non-root cgroups to the extent possible. For that reason, we simulate
793 * flushing the root cgroup's stats by explicitly filling in the iostat
794 * with disk level statistics.
795 */
796static void blkcg_fill_root_iostats(void)
797{
798 struct class_dev_iter iter;
799 struct device *dev;
800
801 class_dev_iter_init(&iter, &block_class, NULL, &disk_type);
802 while ((dev = class_dev_iter_next(&iter))) {
803 struct gendisk *disk = dev_to_disk(dev);
804 struct hd_struct *part = disk_get_part(disk, 0);
805 struct blkcg_gq *blkg = blk_queue_root_blkg(disk->queue);
806 struct blkg_iostat tmp;
807 int cpu;
808
809 memset(&tmp, 0, sizeof(tmp));
810 for_each_possible_cpu(cpu) {
811 struct disk_stats *cpu_dkstats;
812
813 cpu_dkstats = per_cpu_ptr(part->dkstats, cpu);
814 tmp.ios[BLKG_IOSTAT_READ] +=
815 cpu_dkstats->ios[STAT_READ];
816 tmp.ios[BLKG_IOSTAT_WRITE] +=
817 cpu_dkstats->ios[STAT_WRITE];
818 tmp.ios[BLKG_IOSTAT_DISCARD] +=
819 cpu_dkstats->ios[STAT_DISCARD];
820 // convert sectors to bytes
821 tmp.bytes[BLKG_IOSTAT_READ] +=
822 cpu_dkstats->sectors[STAT_READ] << 9;
823 tmp.bytes[BLKG_IOSTAT_WRITE] +=
824 cpu_dkstats->sectors[STAT_WRITE] << 9;
825 tmp.bytes[BLKG_IOSTAT_DISCARD] +=
826 cpu_dkstats->sectors[STAT_DISCARD] << 9;
827
828 u64_stats_update_begin(&blkg->iostat.sync);
829 blkg_iostat_set(&blkg->iostat.cur, &tmp);
830 u64_stats_update_end(&blkg->iostat.sync);
831 }
832 }
833}
834
835static int blkcg_print_stat(struct seq_file *sf, void *v)
836{
837 struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
838 struct blkcg_gq *blkg;
839
840 if (!seq_css(sf)->parent)
841 blkcg_fill_root_iostats();
842 else
843 cgroup_rstat_flush(blkcg->css.cgroup);
844
845 rcu_read_lock();
846
847 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) {
848 struct blkg_iostat_set *bis = &blkg->iostat;
849 const char *dname;
850 char *buf;
851 u64 rbytes, wbytes, rios, wios, dbytes, dios;
852 size_t size = seq_get_buf(sf, &buf), off = 0;
853 int i;
854 bool has_stats = false;
855 unsigned seq;
856
857 spin_lock_irq(&blkg->q->queue_lock);
858
859 if (!blkg->online)
860 goto skip;
861
862 dname = blkg_dev_name(blkg);
863 if (!dname)
864 goto skip;
865
866 /*
867 * Hooray string manipulation, count is the size written NOT
868 * INCLUDING THE \0, so size is now count+1 less than what we
869 * had before, but we want to start writing the next bit from
870 * the \0 so we only add count to buf.
871 */
872 off += scnprintf(buf+off, size-off, "%s ", dname);
873
874 do {
875 seq = u64_stats_fetch_begin(&bis->sync);
876
877 rbytes = bis->cur.bytes[BLKG_IOSTAT_READ];
878 wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE];
879 dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD];
880 rios = bis->cur.ios[BLKG_IOSTAT_READ];
881 wios = bis->cur.ios[BLKG_IOSTAT_WRITE];
882 dios = bis->cur.ios[BLKG_IOSTAT_DISCARD];
883 } while (u64_stats_fetch_retry(&bis->sync, seq));
884
885 if (rbytes || wbytes || rios || wios) {
886 has_stats = true;
887 off += scnprintf(buf+off, size-off,
888 "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu",
889 rbytes, wbytes, rios, wios,
890 dbytes, dios);
891 }
892
893 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) {
894 has_stats = true;
895 off += scnprintf(buf+off, size-off,
896 " use_delay=%d delay_nsec=%llu",
897 atomic_read(&blkg->use_delay),
898 (unsigned long long)atomic64_read(&blkg->delay_nsec));
899 }
900
901 for (i = 0; i < BLKCG_MAX_POLS; i++) {
902 struct blkcg_policy *pol = blkcg_policy[i];
903 size_t written;
904
905 if (!blkg->pd[i] || !pol->pd_stat_fn)
906 continue;
907
908 written = pol->pd_stat_fn(blkg->pd[i], buf+off, size-off);
909 if (written)
910 has_stats = true;
911 off += written;
912 }
913
914 if (has_stats) {
915 if (off < size - 1) {
916 off += scnprintf(buf+off, size-off, "\n");
917 seq_commit(sf, off);
918 } else {
919 seq_commit(sf, -1);
920 }
921 }
922 skip:
923 spin_unlock_irq(&blkg->q->queue_lock);
924 }
925
926 rcu_read_unlock();
927 return 0;
928}
929
930static struct cftype blkcg_files[] = {
931 {
932 .name = "stat",
933 .seq_show = blkcg_print_stat,
934 },
935 { } /* terminate */
936};
937
938static struct cftype blkcg_legacy_files[] = {
939 {
940 .name = "reset_stats",
941 .write_u64 = blkcg_reset_stats,
942 },
943 { } /* terminate */
944};
945
946/*
947 * blkcg destruction is a three-stage process.
948 *
949 * 1. Destruction starts. The blkcg_css_offline() callback is invoked
950 * which offlines writeback. Here we tie the next stage of blkg destruction
951 * to the completion of writeback associated with the blkcg. This lets us
952 * avoid punting potentially large amounts of outstanding writeback to root
953 * while maintaining any ongoing policies. The next stage is triggered when
954 * the nr_cgwbs count goes to zero.
955 *
956 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called
957 * and handles the destruction of blkgs. Here the css reference held by
958 * the blkg is put back eventually allowing blkcg_css_free() to be called.
959 * This work may occur in cgwb_release_workfn() on the cgwb_release
960 * workqueue. Any submitted ios that fail to get the blkg ref will be
961 * punted to the root_blkg.
962 *
963 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called.
964 * This finally frees the blkcg.
965 */
966
967/**
968 * blkcg_css_offline - cgroup css_offline callback
969 * @css: css of interest
970 *
971 * This function is called when @css is about to go away. Here the cgwbs are
972 * offlined first and only once writeback associated with the blkcg has
973 * finished do we start step 2 (see above).
974 */
975static void blkcg_css_offline(struct cgroup_subsys_state *css)
976{
977 struct blkcg *blkcg = css_to_blkcg(css);
978
979 /* this prevents anyone from attaching or migrating to this blkcg */
980 wb_blkcg_offline(blkcg);
981
982 /* put the base online pin allowing step 2 to be triggered */
983 blkcg_unpin_online(blkcg);
984}
985
986/**
987 * blkcg_destroy_blkgs - responsible for shooting down blkgs
988 * @blkcg: blkcg of interest
989 *
990 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock
991 * is nested inside q lock, this function performs reverse double lock dancing.
992 * Destroying the blkgs releases the reference held on the blkcg's css allowing
993 * blkcg_css_free to eventually be called.
994 *
995 * This is the blkcg counterpart of ioc_release_fn().
996 */
997void blkcg_destroy_blkgs(struct blkcg *blkcg)
998{
999 spin_lock_irq(&blkcg->lock);
1000
1001 while (!hlist_empty(&blkcg->blkg_list)) {
1002 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first,
1003 struct blkcg_gq, blkcg_node);
1004 struct request_queue *q = blkg->q;
1005
1006 if (spin_trylock(&q->queue_lock)) {
1007 blkg_destroy(blkg);
1008 spin_unlock(&q->queue_lock);
1009 } else {
1010 spin_unlock_irq(&blkcg->lock);
1011 cpu_relax();
1012 spin_lock_irq(&blkcg->lock);
1013 }
1014 }
1015
1016 spin_unlock_irq(&blkcg->lock);
1017}
1018
1019static void blkcg_css_free(struct cgroup_subsys_state *css)
1020{
1021 struct blkcg *blkcg = css_to_blkcg(css);
1022 int i;
1023
1024 mutex_lock(&blkcg_pol_mutex);
1025
1026 list_del(&blkcg->all_blkcgs_node);
1027
1028 for (i = 0; i < BLKCG_MAX_POLS; i++)
1029 if (blkcg->cpd[i])
1030 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1031
1032 mutex_unlock(&blkcg_pol_mutex);
1033
1034 kfree(blkcg);
1035}
1036
1037static struct cgroup_subsys_state *
1038blkcg_css_alloc(struct cgroup_subsys_state *parent_css)
1039{
1040 struct blkcg *blkcg;
1041 struct cgroup_subsys_state *ret;
1042 int i;
1043
1044 mutex_lock(&blkcg_pol_mutex);
1045
1046 if (!parent_css) {
1047 blkcg = &blkcg_root;
1048 } else {
1049 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL);
1050 if (!blkcg) {
1051 ret = ERR_PTR(-ENOMEM);
1052 goto unlock;
1053 }
1054 }
1055
1056 for (i = 0; i < BLKCG_MAX_POLS ; i++) {
1057 struct blkcg_policy *pol = blkcg_policy[i];
1058 struct blkcg_policy_data *cpd;
1059
1060 /*
1061 * If the policy hasn't been attached yet, wait for it
1062 * to be attached before doing anything else. Otherwise,
1063 * check if the policy requires any specific per-cgroup
1064 * data: if it does, allocate and initialize it.
1065 */
1066 if (!pol || !pol->cpd_alloc_fn)
1067 continue;
1068
1069 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1070 if (!cpd) {
1071 ret = ERR_PTR(-ENOMEM);
1072 goto free_pd_blkcg;
1073 }
1074 blkcg->cpd[i] = cpd;
1075 cpd->blkcg = blkcg;
1076 cpd->plid = i;
1077 if (pol->cpd_init_fn)
1078 pol->cpd_init_fn(cpd);
1079 }
1080
1081 spin_lock_init(&blkcg->lock);
1082 refcount_set(&blkcg->online_pin, 1);
1083 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN);
1084 INIT_HLIST_HEAD(&blkcg->blkg_list);
1085#ifdef CONFIG_CGROUP_WRITEBACK
1086 INIT_LIST_HEAD(&blkcg->cgwb_list);
1087#endif
1088 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs);
1089
1090 mutex_unlock(&blkcg_pol_mutex);
1091 return &blkcg->css;
1092
1093free_pd_blkcg:
1094 for (i--; i >= 0; i--)
1095 if (blkcg->cpd[i])
1096 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]);
1097
1098 if (blkcg != &blkcg_root)
1099 kfree(blkcg);
1100unlock:
1101 mutex_unlock(&blkcg_pol_mutex);
1102 return ret;
1103}
1104
1105static int blkcg_css_online(struct cgroup_subsys_state *css)
1106{
1107 struct blkcg *blkcg = css_to_blkcg(css);
1108 struct blkcg *parent = blkcg_parent(blkcg);
1109
1110 /*
1111 * blkcg_pin_online() is used to delay blkcg offline so that blkgs
1112 * don't go offline while cgwbs are still active on them. Pin the
1113 * parent so that offline always happens towards the root.
1114 */
1115 if (parent)
1116 blkcg_pin_online(parent);
1117 return 0;
1118}
1119
1120/**
1121 * blkcg_init_queue - initialize blkcg part of request queue
1122 * @q: request_queue to initialize
1123 *
1124 * Called from blk_alloc_queue(). Responsible for initializing blkcg
1125 * part of new request_queue @q.
1126 *
1127 * RETURNS:
1128 * 0 on success, -errno on failure.
1129 */
1130int blkcg_init_queue(struct request_queue *q)
1131{
1132 struct blkcg_gq *new_blkg, *blkg;
1133 bool preloaded;
1134 int ret;
1135
1136 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL);
1137 if (!new_blkg)
1138 return -ENOMEM;
1139
1140 preloaded = !radix_tree_preload(GFP_KERNEL);
1141
1142 /* Make sure the root blkg exists. */
1143 rcu_read_lock();
1144 spin_lock_irq(&q->queue_lock);
1145 blkg = blkg_create(&blkcg_root, q, new_blkg);
1146 if (IS_ERR(blkg))
1147 goto err_unlock;
1148 q->root_blkg = blkg;
1149 spin_unlock_irq(&q->queue_lock);
1150 rcu_read_unlock();
1151
1152 if (preloaded)
1153 radix_tree_preload_end();
1154
1155 ret = blk_throtl_init(q);
1156 if (ret)
1157 goto err_destroy_all;
1158
1159 ret = blk_iolatency_init(q);
1160 if (ret) {
1161 blk_throtl_exit(q);
1162 goto err_destroy_all;
1163 }
1164 return 0;
1165
1166err_destroy_all:
1167 blkg_destroy_all(q);
1168 return ret;
1169err_unlock:
1170 spin_unlock_irq(&q->queue_lock);
1171 rcu_read_unlock();
1172 if (preloaded)
1173 radix_tree_preload_end();
1174 return PTR_ERR(blkg);
1175}
1176
1177/**
1178 * blkcg_exit_queue - exit and release blkcg part of request_queue
1179 * @q: request_queue being released
1180 *
1181 * Called from blk_exit_queue(). Responsible for exiting blkcg part.
1182 */
1183void blkcg_exit_queue(struct request_queue *q)
1184{
1185 blkg_destroy_all(q);
1186 blk_throtl_exit(q);
1187}
1188
1189/*
1190 * We cannot support shared io contexts, as we have no mean to support
1191 * two tasks with the same ioc in two different groups without major rework
1192 * of the main cic data structures. For now we allow a task to change
1193 * its cgroup only if it's the only owner of its ioc.
1194 */
1195static int blkcg_can_attach(struct cgroup_taskset *tset)
1196{
1197 struct task_struct *task;
1198 struct cgroup_subsys_state *dst_css;
1199 struct io_context *ioc;
1200 int ret = 0;
1201
1202 /* task_lock() is needed to avoid races with exit_io_context() */
1203 cgroup_taskset_for_each(task, dst_css, tset) {
1204 task_lock(task);
1205 ioc = task->io_context;
1206 if (ioc && atomic_read(&ioc->nr_tasks) > 1)
1207 ret = -EINVAL;
1208 task_unlock(task);
1209 if (ret)
1210 break;
1211 }
1212 return ret;
1213}
1214
1215static void blkcg_bind(struct cgroup_subsys_state *root_css)
1216{
1217 int i;
1218
1219 mutex_lock(&blkcg_pol_mutex);
1220
1221 for (i = 0; i < BLKCG_MAX_POLS; i++) {
1222 struct blkcg_policy *pol = blkcg_policy[i];
1223 struct blkcg *blkcg;
1224
1225 if (!pol || !pol->cpd_bind_fn)
1226 continue;
1227
1228 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node)
1229 if (blkcg->cpd[pol->plid])
1230 pol->cpd_bind_fn(blkcg->cpd[pol->plid]);
1231 }
1232 mutex_unlock(&blkcg_pol_mutex);
1233}
1234
1235static void blkcg_exit(struct task_struct *tsk)
1236{
1237 if (tsk->throttle_queue)
1238 blk_put_queue(tsk->throttle_queue);
1239 tsk->throttle_queue = NULL;
1240}
1241
1242struct cgroup_subsys io_cgrp_subsys = {
1243 .css_alloc = blkcg_css_alloc,
1244 .css_online = blkcg_css_online,
1245 .css_offline = blkcg_css_offline,
1246 .css_free = blkcg_css_free,
1247 .can_attach = blkcg_can_attach,
1248 .css_rstat_flush = blkcg_rstat_flush,
1249 .bind = blkcg_bind,
1250 .dfl_cftypes = blkcg_files,
1251 .legacy_cftypes = blkcg_legacy_files,
1252 .legacy_name = "blkio",
1253 .exit = blkcg_exit,
1254#ifdef CONFIG_MEMCG
1255 /*
1256 * This ensures that, if available, memcg is automatically enabled
1257 * together on the default hierarchy so that the owner cgroup can
1258 * be retrieved from writeback pages.
1259 */
1260 .depends_on = 1 << memory_cgrp_id,
1261#endif
1262};
1263EXPORT_SYMBOL_GPL(io_cgrp_subsys);
1264
1265/**
1266 * blkcg_activate_policy - activate a blkcg policy on a request_queue
1267 * @q: request_queue of interest
1268 * @pol: blkcg policy to activate
1269 *
1270 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through
1271 * bypass mode to populate its blkgs with policy_data for @pol.
1272 *
1273 * Activation happens with @q bypassed, so nobody would be accessing blkgs
1274 * from IO path. Update of each blkg is protected by both queue and blkcg
1275 * locks so that holding either lock and testing blkcg_policy_enabled() is
1276 * always enough for dereferencing policy data.
1277 *
1278 * The caller is responsible for synchronizing [de]activations and policy
1279 * [un]registerations. Returns 0 on success, -errno on failure.
1280 */
1281int blkcg_activate_policy(struct request_queue *q,
1282 const struct blkcg_policy *pol)
1283{
1284 struct blkg_policy_data *pd_prealloc = NULL;
1285 struct blkcg_gq *blkg, *pinned_blkg = NULL;
1286 int ret;
1287
1288 if (blkcg_policy_enabled(q, pol))
1289 return 0;
1290
1291 if (queue_is_mq(q))
1292 blk_mq_freeze_queue(q);
1293retry:
1294 spin_lock_irq(&q->queue_lock);
1295
1296 /* blkg_list is pushed at the head, reverse walk to allocate parents first */
1297 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) {
1298 struct blkg_policy_data *pd;
1299
1300 if (blkg->pd[pol->plid])
1301 continue;
1302
1303 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */
1304 if (blkg == pinned_blkg) {
1305 pd = pd_prealloc;
1306 pd_prealloc = NULL;
1307 } else {
1308 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q,
1309 blkg->blkcg);
1310 }
1311
1312 if (!pd) {
1313 /*
1314 * GFP_NOWAIT failed. Free the existing one and
1315 * prealloc for @blkg w/ GFP_KERNEL.
1316 */
1317 if (pinned_blkg)
1318 blkg_put(pinned_blkg);
1319 blkg_get(blkg);
1320 pinned_blkg = blkg;
1321
1322 spin_unlock_irq(&q->queue_lock);
1323
1324 if (pd_prealloc)
1325 pol->pd_free_fn(pd_prealloc);
1326 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q,
1327 blkg->blkcg);
1328 if (pd_prealloc)
1329 goto retry;
1330 else
1331 goto enomem;
1332 }
1333
1334 blkg->pd[pol->plid] = pd;
1335 pd->blkg = blkg;
1336 pd->plid = pol->plid;
1337 }
1338
1339 /* all allocated, init in the same order */
1340 if (pol->pd_init_fn)
1341 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node)
1342 pol->pd_init_fn(blkg->pd[pol->plid]);
1343
1344 __set_bit(pol->plid, q->blkcg_pols);
1345 ret = 0;
1346
1347 spin_unlock_irq(&q->queue_lock);
1348out:
1349 if (queue_is_mq(q))
1350 blk_mq_unfreeze_queue(q);
1351 if (pinned_blkg)
1352 blkg_put(pinned_blkg);
1353 if (pd_prealloc)
1354 pol->pd_free_fn(pd_prealloc);
1355 return ret;
1356
1357enomem:
1358 /* alloc failed, nothing's initialized yet, free everything */
1359 spin_lock_irq(&q->queue_lock);
1360 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1361 if (blkg->pd[pol->plid]) {
1362 pol->pd_free_fn(blkg->pd[pol->plid]);
1363 blkg->pd[pol->plid] = NULL;
1364 }
1365 }
1366 spin_unlock_irq(&q->queue_lock);
1367 ret = -ENOMEM;
1368 goto out;
1369}
1370EXPORT_SYMBOL_GPL(blkcg_activate_policy);
1371
1372/**
1373 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue
1374 * @q: request_queue of interest
1375 * @pol: blkcg policy to deactivate
1376 *
1377 * Deactivate @pol on @q. Follows the same synchronization rules as
1378 * blkcg_activate_policy().
1379 */
1380void blkcg_deactivate_policy(struct request_queue *q,
1381 const struct blkcg_policy *pol)
1382{
1383 struct blkcg_gq *blkg;
1384
1385 if (!blkcg_policy_enabled(q, pol))
1386 return;
1387
1388 if (queue_is_mq(q))
1389 blk_mq_freeze_queue(q);
1390
1391 spin_lock_irq(&q->queue_lock);
1392
1393 __clear_bit(pol->plid, q->blkcg_pols);
1394
1395 list_for_each_entry(blkg, &q->blkg_list, q_node) {
1396 if (blkg->pd[pol->plid]) {
1397 if (pol->pd_offline_fn)
1398 pol->pd_offline_fn(blkg->pd[pol->plid]);
1399 pol->pd_free_fn(blkg->pd[pol->plid]);
1400 blkg->pd[pol->plid] = NULL;
1401 }
1402 }
1403
1404 spin_unlock_irq(&q->queue_lock);
1405
1406 if (queue_is_mq(q))
1407 blk_mq_unfreeze_queue(q);
1408}
1409EXPORT_SYMBOL_GPL(blkcg_deactivate_policy);
1410
1411/**
1412 * blkcg_policy_register - register a blkcg policy
1413 * @pol: blkcg policy to register
1414 *
1415 * Register @pol with blkcg core. Might sleep and @pol may be modified on
1416 * successful registration. Returns 0 on success and -errno on failure.
1417 */
1418int blkcg_policy_register(struct blkcg_policy *pol)
1419{
1420 struct blkcg *blkcg;
1421 int i, ret;
1422
1423 mutex_lock(&blkcg_pol_register_mutex);
1424 mutex_lock(&blkcg_pol_mutex);
1425
1426 /* find an empty slot */
1427 ret = -ENOSPC;
1428 for (i = 0; i < BLKCG_MAX_POLS; i++)
1429 if (!blkcg_policy[i])
1430 break;
1431 if (i >= BLKCG_MAX_POLS) {
1432 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n");
1433 goto err_unlock;
1434 }
1435
1436 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */
1437 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) ||
1438 (!pol->pd_alloc_fn ^ !pol->pd_free_fn))
1439 goto err_unlock;
1440
1441 /* register @pol */
1442 pol->plid = i;
1443 blkcg_policy[pol->plid] = pol;
1444
1445 /* allocate and install cpd's */
1446 if (pol->cpd_alloc_fn) {
1447 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1448 struct blkcg_policy_data *cpd;
1449
1450 cpd = pol->cpd_alloc_fn(GFP_KERNEL);
1451 if (!cpd)
1452 goto err_free_cpds;
1453
1454 blkcg->cpd[pol->plid] = cpd;
1455 cpd->blkcg = blkcg;
1456 cpd->plid = pol->plid;
1457 if (pol->cpd_init_fn)
1458 pol->cpd_init_fn(cpd);
1459 }
1460 }
1461
1462 mutex_unlock(&blkcg_pol_mutex);
1463
1464 /* everything is in place, add intf files for the new policy */
1465 if (pol->dfl_cftypes)
1466 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys,
1467 pol->dfl_cftypes));
1468 if (pol->legacy_cftypes)
1469 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys,
1470 pol->legacy_cftypes));
1471 mutex_unlock(&blkcg_pol_register_mutex);
1472 return 0;
1473
1474err_free_cpds:
1475 if (pol->cpd_free_fn) {
1476 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1477 if (blkcg->cpd[pol->plid]) {
1478 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1479 blkcg->cpd[pol->plid] = NULL;
1480 }
1481 }
1482 }
1483 blkcg_policy[pol->plid] = NULL;
1484err_unlock:
1485 mutex_unlock(&blkcg_pol_mutex);
1486 mutex_unlock(&blkcg_pol_register_mutex);
1487 return ret;
1488}
1489EXPORT_SYMBOL_GPL(blkcg_policy_register);
1490
1491/**
1492 * blkcg_policy_unregister - unregister a blkcg policy
1493 * @pol: blkcg policy to unregister
1494 *
1495 * Undo blkcg_policy_register(@pol). Might sleep.
1496 */
1497void blkcg_policy_unregister(struct blkcg_policy *pol)
1498{
1499 struct blkcg *blkcg;
1500
1501 mutex_lock(&blkcg_pol_register_mutex);
1502
1503 if (WARN_ON(blkcg_policy[pol->plid] != pol))
1504 goto out_unlock;
1505
1506 /* kill the intf files first */
1507 if (pol->dfl_cftypes)
1508 cgroup_rm_cftypes(pol->dfl_cftypes);
1509 if (pol->legacy_cftypes)
1510 cgroup_rm_cftypes(pol->legacy_cftypes);
1511
1512 /* remove cpds and unregister */
1513 mutex_lock(&blkcg_pol_mutex);
1514
1515 if (pol->cpd_free_fn) {
1516 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) {
1517 if (blkcg->cpd[pol->plid]) {
1518 pol->cpd_free_fn(blkcg->cpd[pol->plid]);
1519 blkcg->cpd[pol->plid] = NULL;
1520 }
1521 }
1522 }
1523 blkcg_policy[pol->plid] = NULL;
1524
1525 mutex_unlock(&blkcg_pol_mutex);
1526out_unlock:
1527 mutex_unlock(&blkcg_pol_register_mutex);
1528}
1529EXPORT_SYMBOL_GPL(blkcg_policy_unregister);
1530
1531bool __blkcg_punt_bio_submit(struct bio *bio)
1532{
1533 struct blkcg_gq *blkg = bio->bi_blkg;
1534
1535 /* consume the flag first */
1536 bio->bi_opf &= ~REQ_CGROUP_PUNT;
1537
1538 /* never bounce for the root cgroup */
1539 if (!blkg->parent)
1540 return false;
1541
1542 spin_lock_bh(&blkg->async_bio_lock);
1543 bio_list_add(&blkg->async_bios, bio);
1544 spin_unlock_bh(&blkg->async_bio_lock);
1545
1546 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work);
1547 return true;
1548}
1549
1550/*
1551 * Scale the accumulated delay based on how long it has been since we updated
1552 * the delay. We only call this when we are adding delay, in case it's been a
1553 * while since we added delay, and when we are checking to see if we need to
1554 * delay a task, to account for any delays that may have occurred.
1555 */
1556static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now)
1557{
1558 u64 old = atomic64_read(&blkg->delay_start);
1559
1560 /* negative use_delay means no scaling, see blkcg_set_delay() */
1561 if (atomic_read(&blkg->use_delay) < 0)
1562 return;
1563
1564 /*
1565 * We only want to scale down every second. The idea here is that we
1566 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain
1567 * time window. We only want to throttle tasks for recent delay that
1568 * has occurred, in 1 second time windows since that's the maximum
1569 * things can be throttled. We save the current delay window in
1570 * blkg->last_delay so we know what amount is still left to be charged
1571 * to the blkg from this point onward. blkg->last_use keeps track of
1572 * the use_delay counter. The idea is if we're unthrottling the blkg we
1573 * are ok with whatever is happening now, and we can take away more of
1574 * the accumulated delay as we've already throttled enough that
1575 * everybody is happy with their IO latencies.
1576 */
1577 if (time_before64(old + NSEC_PER_SEC, now) &&
1578 atomic64_cmpxchg(&blkg->delay_start, old, now) == old) {
1579 u64 cur = atomic64_read(&blkg->delay_nsec);
1580 u64 sub = min_t(u64, blkg->last_delay, now - old);
1581 int cur_use = atomic_read(&blkg->use_delay);
1582
1583 /*
1584 * We've been unthrottled, subtract a larger chunk of our
1585 * accumulated delay.
1586 */
1587 if (cur_use < blkg->last_use)
1588 sub = max_t(u64, sub, blkg->last_delay >> 1);
1589
1590 /*
1591 * This shouldn't happen, but handle it anyway. Our delay_nsec
1592 * should only ever be growing except here where we subtract out
1593 * min(last_delay, 1 second), but lord knows bugs happen and I'd
1594 * rather not end up with negative numbers.
1595 */
1596 if (unlikely(cur < sub)) {
1597 atomic64_set(&blkg->delay_nsec, 0);
1598 blkg->last_delay = 0;
1599 } else {
1600 atomic64_sub(sub, &blkg->delay_nsec);
1601 blkg->last_delay = cur - sub;
1602 }
1603 blkg->last_use = cur_use;
1604 }
1605}
1606
1607/*
1608 * This is called when we want to actually walk up the hierarchy and check to
1609 * see if we need to throttle, and then actually throttle if there is some
1610 * accumulated delay. This should only be called upon return to user space so
1611 * we're not holding some lock that would induce a priority inversion.
1612 */
1613static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay)
1614{
1615 unsigned long pflags;
1616 u64 now = ktime_to_ns(ktime_get());
1617 u64 exp;
1618 u64 delay_nsec = 0;
1619 int tok;
1620
1621 while (blkg->parent) {
1622 if (atomic_read(&blkg->use_delay)) {
1623 blkcg_scale_delay(blkg, now);
1624 delay_nsec = max_t(u64, delay_nsec,
1625 atomic64_read(&blkg->delay_nsec));
1626 }
1627 blkg = blkg->parent;
1628 }
1629
1630 if (!delay_nsec)
1631 return;
1632
1633 /*
1634 * Let's not sleep for all eternity if we've amassed a huge delay.
1635 * Swapping or metadata IO can accumulate 10's of seconds worth of
1636 * delay, and we want userspace to be able to do _something_ so cap the
1637 * delays at 1 second. If there's 10's of seconds worth of delay then
1638 * the tasks will be delayed for 1 second for every syscall.
1639 */
1640 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC);
1641
1642 if (use_memdelay)
1643 psi_memstall_enter(&pflags);
1644
1645 exp = ktime_add_ns(now, delay_nsec);
1646 tok = io_schedule_prepare();
1647 do {
1648 __set_current_state(TASK_KILLABLE);
1649 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS))
1650 break;
1651 } while (!fatal_signal_pending(current));
1652 io_schedule_finish(tok);
1653
1654 if (use_memdelay)
1655 psi_memstall_leave(&pflags);
1656}
1657
1658/**
1659 * blkcg_maybe_throttle_current - throttle the current task if it has been marked
1660 *
1661 * This is only called if we've been marked with set_notify_resume(). Obviously
1662 * we can be set_notify_resume() for reasons other than blkcg throttling, so we
1663 * check to see if current->throttle_queue is set and if not this doesn't do
1664 * anything. This should only ever be called by the resume code, it's not meant
1665 * to be called by people willy-nilly as it will actually do the work to
1666 * throttle the task if it is setup for throttling.
1667 */
1668void blkcg_maybe_throttle_current(void)
1669{
1670 struct request_queue *q = current->throttle_queue;
1671 struct cgroup_subsys_state *css;
1672 struct blkcg *blkcg;
1673 struct blkcg_gq *blkg;
1674 bool use_memdelay = current->use_memdelay;
1675
1676 if (!q)
1677 return;
1678
1679 current->throttle_queue = NULL;
1680 current->use_memdelay = false;
1681
1682 rcu_read_lock();
1683 css = kthread_blkcg();
1684 if (css)
1685 blkcg = css_to_blkcg(css);
1686 else
1687 blkcg = css_to_blkcg(task_css(current, io_cgrp_id));
1688
1689 if (!blkcg)
1690 goto out;
1691 blkg = blkg_lookup(blkcg, q);
1692 if (!blkg)
1693 goto out;
1694 if (!blkg_tryget(blkg))
1695 goto out;
1696 rcu_read_unlock();
1697
1698 blkcg_maybe_throttle_blkg(blkg, use_memdelay);
1699 blkg_put(blkg);
1700 blk_put_queue(q);
1701 return;
1702out:
1703 rcu_read_unlock();
1704 blk_put_queue(q);
1705}
1706
1707/**
1708 * blkcg_schedule_throttle - this task needs to check for throttling
1709 * @q: the request queue IO was submitted on
1710 * @use_memdelay: do we charge this to memory delay for PSI
1711 *
1712 * This is called by the IO controller when we know there's delay accumulated
1713 * for the blkg for this task. We do not pass the blkg because there are places
1714 * we call this that may not have that information, the swapping code for
1715 * instance will only have a request_queue at that point. This set's the
1716 * notify_resume for the task to check and see if it requires throttling before
1717 * returning to user space.
1718 *
1719 * We will only schedule once per syscall. You can call this over and over
1720 * again and it will only do the check once upon return to user space, and only
1721 * throttle once. If the task needs to be throttled again it'll need to be
1722 * re-set at the next time we see the task.
1723 */
1724void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay)
1725{
1726 if (unlikely(current->flags & PF_KTHREAD))
1727 return;
1728
1729 if (!blk_get_queue(q))
1730 return;
1731
1732 if (current->throttle_queue)
1733 blk_put_queue(current->throttle_queue);
1734 current->throttle_queue = q;
1735 if (use_memdelay)
1736 current->use_memdelay = use_memdelay;
1737 set_notify_resume(current);
1738}
1739
1740/**
1741 * blkcg_add_delay - add delay to this blkg
1742 * @blkg: blkg of interest
1743 * @now: the current time in nanoseconds
1744 * @delta: how many nanoseconds of delay to add
1745 *
1746 * Charge @delta to the blkg's current delay accumulation. This is used to
1747 * throttle tasks if an IO controller thinks we need more throttling.
1748 */
1749void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta)
1750{
1751 if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0))
1752 return;
1753 blkcg_scale_delay(blkg, now);
1754 atomic64_add(delta, &blkg->delay_nsec);
1755}
1756
1757/**
1758 * blkg_tryget_closest - try and get a blkg ref on the closet blkg
1759 * @bio: target bio
1760 * @css: target css
1761 *
1762 * As the failure mode here is to walk up the blkg tree, this ensure that the
1763 * blkg->parent pointers are always valid. This returns the blkg that it ended
1764 * up taking a reference on or %NULL if no reference was taken.
1765 */
1766static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio,
1767 struct cgroup_subsys_state *css)
1768{
1769 struct blkcg_gq *blkg, *ret_blkg = NULL;
1770
1771 rcu_read_lock();
1772 blkg = blkg_lookup_create(css_to_blkcg(css), bio->bi_disk->queue);
1773 while (blkg) {
1774 if (blkg_tryget(blkg)) {
1775 ret_blkg = blkg;
1776 break;
1777 }
1778 blkg = blkg->parent;
1779 }
1780 rcu_read_unlock();
1781
1782 return ret_blkg;
1783}
1784
1785/**
1786 * bio_associate_blkg_from_css - associate a bio with a specified css
1787 * @bio: target bio
1788 * @css: target css
1789 *
1790 * Associate @bio with the blkg found by combining the css's blkg and the
1791 * request_queue of the @bio. An association failure is handled by walking up
1792 * the blkg tree. Therefore, the blkg associated can be anything between @blkg
1793 * and q->root_blkg. This situation only happens when a cgroup is dying and
1794 * then the remaining bios will spill to the closest alive blkg.
1795 *
1796 * A reference will be taken on the blkg and will be released when @bio is
1797 * freed.
1798 */
1799void bio_associate_blkg_from_css(struct bio *bio,
1800 struct cgroup_subsys_state *css)
1801{
1802 if (bio->bi_blkg)
1803 blkg_put(bio->bi_blkg);
1804
1805 if (css && css->parent) {
1806 bio->bi_blkg = blkg_tryget_closest(bio, css);
1807 } else {
1808 blkg_get(bio->bi_disk->queue->root_blkg);
1809 bio->bi_blkg = bio->bi_disk->queue->root_blkg;
1810 }
1811}
1812EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css);
1813
1814/**
1815 * bio_associate_blkg - associate a bio with a blkg
1816 * @bio: target bio
1817 *
1818 * Associate @bio with the blkg found from the bio's css and request_queue.
1819 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is
1820 * already associated, the css is reused and association redone as the
1821 * request_queue may have changed.
1822 */
1823void bio_associate_blkg(struct bio *bio)
1824{
1825 struct cgroup_subsys_state *css;
1826
1827 rcu_read_lock();
1828
1829 if (bio->bi_blkg)
1830 css = &bio_blkcg(bio)->css;
1831 else
1832 css = blkcg_css();
1833
1834 bio_associate_blkg_from_css(bio, css);
1835
1836 rcu_read_unlock();
1837}
1838EXPORT_SYMBOL_GPL(bio_associate_blkg);
1839
1840/**
1841 * bio_clone_blkg_association - clone blkg association from src to dst bio
1842 * @dst: destination bio
1843 * @src: source bio
1844 */
1845void bio_clone_blkg_association(struct bio *dst, struct bio *src)
1846{
1847 if (src->bi_blkg) {
1848 if (dst->bi_blkg)
1849 blkg_put(dst->bi_blkg);
1850 blkg_get(src->bi_blkg);
1851 dst->bi_blkg = src->bi_blkg;
1852 }
1853}
1854EXPORT_SYMBOL_GPL(bio_clone_blkg_association);
1855
1856static int blk_cgroup_io_type(struct bio *bio)
1857{
1858 if (op_is_discard(bio->bi_opf))
1859 return BLKG_IOSTAT_DISCARD;
1860 if (op_is_write(bio->bi_opf))
1861 return BLKG_IOSTAT_WRITE;
1862 return BLKG_IOSTAT_READ;
1863}
1864
1865void blk_cgroup_bio_start(struct bio *bio)
1866{
1867 int rwd = blk_cgroup_io_type(bio), cpu;
1868 struct blkg_iostat_set *bis;
1869
1870 cpu = get_cpu();
1871 bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu);
1872 u64_stats_update_begin(&bis->sync);
1873
1874 /*
1875 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split
1876 * bio and we would have already accounted for the size of the bio.
1877 */
1878 if (!bio_flagged(bio, BIO_CGROUP_ACCT)) {
1879 bio_set_flag(bio, BIO_CGROUP_ACCT);
1880 bis->cur.bytes[rwd] += bio->bi_iter.bi_size;
1881 }
1882 bis->cur.ios[rwd]++;
1883
1884 u64_stats_update_end(&bis->sync);
1885 if (cgroup_subsys_on_dfl(io_cgrp_subsys))
1886 cgroup_rstat_updated(bio->bi_blkg->blkcg->css.cgroup, cpu);
1887 put_cpu();
1888}
1889
1890static int __init blkcg_init(void)
1891{
1892 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio",
1893 WQ_MEM_RECLAIM | WQ_FREEZABLE |
1894 WQ_UNBOUND | WQ_SYSFS, 0);
1895 if (!blkcg_punt_bio_wq)
1896 return -ENOMEM;
1897 return 0;
1898}
1899subsys_initcall(blkcg_init);
1900
1901module_param(blkcg_debug_stats, bool, 0644);
1902MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not");