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