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/delay.h> |
27 | #include <linux/atomic.h> |
28 | #include <linux/ctype.h> |
29 | #include <linux/resume_user_mode.h> |
30 | #include <linux/psi.h> |
31 | #include <linux/part_stat.h> |
32 | #include "blk.h" |
33 | #include "blk-cgroup.h" |
34 | #include "blk-ioprio.h" |
35 | #include "blk-throttle.h" |
36 | |
37 | static void __blkcg_rstat_flush(struct blkcg *blkcg, int cpu); |
38 | |
39 | /* |
40 | * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation. |
41 | * blkcg_pol_register_mutex nests outside of it and synchronizes entire |
42 | * policy [un]register operations including cgroup file additions / |
43 | * removals. Putting cgroup file registration outside blkcg_pol_mutex |
44 | * allows grabbing it from cgroup callbacks. |
45 | */ |
46 | static DEFINE_MUTEX(blkcg_pol_register_mutex); |
47 | static DEFINE_MUTEX(blkcg_pol_mutex); |
48 | |
49 | struct blkcg blkcg_root; |
50 | EXPORT_SYMBOL_GPL(blkcg_root); |
51 | |
52 | struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css; |
53 | EXPORT_SYMBOL_GPL(blkcg_root_css); |
54 | |
55 | static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS]; |
56 | |
57 | static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */ |
58 | |
59 | bool blkcg_debug_stats = false; |
60 | |
61 | static DEFINE_RAW_SPINLOCK(blkg_stat_lock); |
62 | |
63 | #define BLKG_DESTROY_BATCH_SIZE 64 |
64 | |
65 | /* |
66 | * Lockless lists for tracking IO stats update |
67 | * |
68 | * New IO stats are stored in the percpu iostat_cpu within blkcg_gq (blkg). |
69 | * There are multiple blkg's (one for each block device) attached to each |
70 | * blkcg. The rstat code keeps track of which cpu has IO stats updated, |
71 | * but it doesn't know which blkg has the updated stats. If there are many |
72 | * block devices in a system, the cost of iterating all the blkg's to flush |
73 | * out the IO stats can be high. To reduce such overhead, a set of percpu |
74 | * lockless lists (lhead) per blkcg are used to track the set of recently |
75 | * updated iostat_cpu's since the last flush. An iostat_cpu will be put |
76 | * onto the lockless list on the update side [blk_cgroup_bio_start()] if |
77 | * not there yet and then removed when being flushed [blkcg_rstat_flush()]. |
78 | * References to blkg are gotten and then put back in the process to |
79 | * protect against blkg removal. |
80 | * |
81 | * Return: 0 if successful or -ENOMEM if allocation fails. |
82 | */ |
83 | static int init_blkcg_llists(struct blkcg *blkcg) |
84 | { |
85 | int cpu; |
86 | |
87 | blkcg->lhead = alloc_percpu_gfp(struct llist_head, GFP_KERNEL); |
88 | if (!blkcg->lhead) |
89 | return -ENOMEM; |
90 | |
91 | for_each_possible_cpu(cpu) |
92 | init_llist_head(per_cpu_ptr(blkcg->lhead, cpu)); |
93 | return 0; |
94 | } |
95 | |
96 | /** |
97 | * blkcg_css - find the current css |
98 | * |
99 | * Find the css associated with either the kthread or the current task. |
100 | * This may return a dying css, so it is up to the caller to use tryget logic |
101 | * to confirm it is alive and well. |
102 | */ |
103 | static struct cgroup_subsys_state *blkcg_css(void) |
104 | { |
105 | struct cgroup_subsys_state *css; |
106 | |
107 | css = kthread_blkcg(); |
108 | if (css) |
109 | return css; |
110 | return task_css(current, subsys_id: io_cgrp_id); |
111 | } |
112 | |
113 | static bool blkcg_policy_enabled(struct request_queue *q, |
114 | const struct blkcg_policy *pol) |
115 | { |
116 | return pol && test_bit(pol->plid, q->blkcg_pols); |
117 | } |
118 | |
119 | static void blkg_free_workfn(struct work_struct *work) |
120 | { |
121 | struct blkcg_gq *blkg = container_of(work, struct blkcg_gq, |
122 | free_work); |
123 | struct request_queue *q = blkg->q; |
124 | int i; |
125 | |
126 | /* |
127 | * pd_free_fn() can also be called from blkcg_deactivate_policy(), |
128 | * in order to make sure pd_free_fn() is called in order, the deletion |
129 | * of the list blkg->q_node is delayed to here from blkg_destroy(), and |
130 | * blkcg_mutex is used to synchronize blkg_free_workfn() and |
131 | * blkcg_deactivate_policy(). |
132 | */ |
133 | mutex_lock(&q->blkcg_mutex); |
134 | for (i = 0; i < BLKCG_MAX_POLS; i++) |
135 | if (blkg->pd[i]) |
136 | blkcg_policy[i]->pd_free_fn(blkg->pd[i]); |
137 | if (blkg->parent) |
138 | blkg_put(blkg: blkg->parent); |
139 | spin_lock_irq(lock: &q->queue_lock); |
140 | list_del_init(entry: &blkg->q_node); |
141 | spin_unlock_irq(lock: &q->queue_lock); |
142 | mutex_unlock(lock: &q->blkcg_mutex); |
143 | |
144 | blk_put_queue(q); |
145 | free_percpu(pdata: blkg->iostat_cpu); |
146 | percpu_ref_exit(ref: &blkg->refcnt); |
147 | kfree(objp: blkg); |
148 | } |
149 | |
150 | /** |
151 | * blkg_free - free a blkg |
152 | * @blkg: blkg to free |
153 | * |
154 | * Free @blkg which may be partially allocated. |
155 | */ |
156 | static void blkg_free(struct blkcg_gq *blkg) |
157 | { |
158 | if (!blkg) |
159 | return; |
160 | |
161 | /* |
162 | * Both ->pd_free_fn() and request queue's release handler may |
163 | * sleep, so free us by scheduling one work func |
164 | */ |
165 | INIT_WORK(&blkg->free_work, blkg_free_workfn); |
166 | schedule_work(work: &blkg->free_work); |
167 | } |
168 | |
169 | static void __blkg_release(struct rcu_head *rcu) |
170 | { |
171 | struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head); |
172 | struct blkcg *blkcg = blkg->blkcg; |
173 | int cpu; |
174 | |
175 | #ifdef CONFIG_BLK_CGROUP_PUNT_BIO |
176 | WARN_ON(!bio_list_empty(&blkg->async_bios)); |
177 | #endif |
178 | /* |
179 | * Flush all the non-empty percpu lockless lists before releasing |
180 | * us, given these stat belongs to us. |
181 | * |
182 | * blkg_stat_lock is for serializing blkg stat update |
183 | */ |
184 | for_each_possible_cpu(cpu) |
185 | __blkcg_rstat_flush(blkcg, cpu); |
186 | |
187 | /* release the blkcg and parent blkg refs this blkg has been holding */ |
188 | css_put(css: &blkg->blkcg->css); |
189 | blkg_free(blkg); |
190 | } |
191 | |
192 | /* |
193 | * A group is RCU protected, but having an rcu lock does not mean that one |
194 | * can access all the fields of blkg and assume these are valid. For |
195 | * example, don't try to follow throtl_data and request queue links. |
196 | * |
197 | * Having a reference to blkg under an rcu allows accesses to only values |
198 | * local to groups like group stats and group rate limits. |
199 | */ |
200 | static void blkg_release(struct percpu_ref *ref) |
201 | { |
202 | struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt); |
203 | |
204 | call_rcu(head: &blkg->rcu_head, func: __blkg_release); |
205 | } |
206 | |
207 | #ifdef CONFIG_BLK_CGROUP_PUNT_BIO |
208 | static struct workqueue_struct *blkcg_punt_bio_wq; |
209 | |
210 | static void blkg_async_bio_workfn(struct work_struct *work) |
211 | { |
212 | struct blkcg_gq *blkg = container_of(work, struct blkcg_gq, |
213 | async_bio_work); |
214 | struct bio_list bios = BIO_EMPTY_LIST; |
215 | struct bio *bio; |
216 | struct blk_plug plug; |
217 | bool need_plug = false; |
218 | |
219 | /* as long as there are pending bios, @blkg can't go away */ |
220 | spin_lock(lock: &blkg->async_bio_lock); |
221 | bio_list_merge(bl: &bios, bl2: &blkg->async_bios); |
222 | bio_list_init(bl: &blkg->async_bios); |
223 | spin_unlock(lock: &blkg->async_bio_lock); |
224 | |
225 | /* start plug only when bio_list contains at least 2 bios */ |
226 | if (bios.head && bios.head->bi_next) { |
227 | need_plug = true; |
228 | blk_start_plug(&plug); |
229 | } |
230 | while ((bio = bio_list_pop(bl: &bios))) |
231 | submit_bio(bio); |
232 | if (need_plug) |
233 | blk_finish_plug(&plug); |
234 | } |
235 | |
236 | /* |
237 | * When a shared kthread issues a bio for a cgroup, doing so synchronously can |
238 | * lead to priority inversions as the kthread can be trapped waiting for that |
239 | * cgroup. Use this helper instead of submit_bio to punt the actual issuing to |
240 | * a dedicated per-blkcg work item to avoid such priority inversions. |
241 | */ |
242 | void blkcg_punt_bio_submit(struct bio *bio) |
243 | { |
244 | struct blkcg_gq *blkg = bio->bi_blkg; |
245 | |
246 | if (blkg->parent) { |
247 | spin_lock(lock: &blkg->async_bio_lock); |
248 | bio_list_add(bl: &blkg->async_bios, bio); |
249 | spin_unlock(lock: &blkg->async_bio_lock); |
250 | queue_work(wq: blkcg_punt_bio_wq, work: &blkg->async_bio_work); |
251 | } else { |
252 | /* never bounce for the root cgroup */ |
253 | submit_bio(bio); |
254 | } |
255 | } |
256 | EXPORT_SYMBOL_GPL(blkcg_punt_bio_submit); |
257 | |
258 | static int __init blkcg_punt_bio_init(void) |
259 | { |
260 | blkcg_punt_bio_wq = alloc_workqueue(fmt: "blkcg_punt_bio" , |
261 | flags: WQ_MEM_RECLAIM | WQ_FREEZABLE | |
262 | WQ_UNBOUND | WQ_SYSFS, max_active: 0); |
263 | if (!blkcg_punt_bio_wq) |
264 | return -ENOMEM; |
265 | return 0; |
266 | } |
267 | subsys_initcall(blkcg_punt_bio_init); |
268 | #endif /* CONFIG_BLK_CGROUP_PUNT_BIO */ |
269 | |
270 | /** |
271 | * bio_blkcg_css - return the blkcg CSS associated with a bio |
272 | * @bio: target bio |
273 | * |
274 | * This returns the CSS for the blkcg associated with a bio, or %NULL if not |
275 | * associated. Callers are expected to either handle %NULL or know association |
276 | * has been done prior to calling this. |
277 | */ |
278 | struct cgroup_subsys_state *bio_blkcg_css(struct bio *bio) |
279 | { |
280 | if (!bio || !bio->bi_blkg) |
281 | return NULL; |
282 | return &bio->bi_blkg->blkcg->css; |
283 | } |
284 | EXPORT_SYMBOL_GPL(bio_blkcg_css); |
285 | |
286 | /** |
287 | * blkcg_parent - get the parent of a blkcg |
288 | * @blkcg: blkcg of interest |
289 | * |
290 | * Return the parent blkcg of @blkcg. Can be called anytime. |
291 | */ |
292 | static inline struct blkcg *blkcg_parent(struct blkcg *blkcg) |
293 | { |
294 | return css_to_blkcg(css: blkcg->css.parent); |
295 | } |
296 | |
297 | /** |
298 | * blkg_alloc - allocate a blkg |
299 | * @blkcg: block cgroup the new blkg is associated with |
300 | * @disk: gendisk the new blkg is associated with |
301 | * @gfp_mask: allocation mask to use |
302 | * |
303 | * Allocate a new blkg assocating @blkcg and @q. |
304 | */ |
305 | static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct gendisk *disk, |
306 | gfp_t gfp_mask) |
307 | { |
308 | struct blkcg_gq *blkg; |
309 | int i, cpu; |
310 | |
311 | /* alloc and init base part */ |
312 | blkg = kzalloc_node(size: sizeof(*blkg), flags: gfp_mask, node: disk->queue->node); |
313 | if (!blkg) |
314 | return NULL; |
315 | if (percpu_ref_init(ref: &blkg->refcnt, release: blkg_release, flags: 0, gfp: gfp_mask)) |
316 | goto out_free_blkg; |
317 | blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask); |
318 | if (!blkg->iostat_cpu) |
319 | goto out_exit_refcnt; |
320 | if (!blk_get_queue(disk->queue)) |
321 | goto out_free_iostat; |
322 | |
323 | blkg->q = disk->queue; |
324 | INIT_LIST_HEAD(list: &blkg->q_node); |
325 | blkg->blkcg = blkcg; |
326 | #ifdef CONFIG_BLK_CGROUP_PUNT_BIO |
327 | spin_lock_init(&blkg->async_bio_lock); |
328 | bio_list_init(bl: &blkg->async_bios); |
329 | INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn); |
330 | #endif |
331 | |
332 | u64_stats_init(syncp: &blkg->iostat.sync); |
333 | for_each_possible_cpu(cpu) { |
334 | u64_stats_init(syncp: &per_cpu_ptr(blkg->iostat_cpu, cpu)->sync); |
335 | per_cpu_ptr(blkg->iostat_cpu, cpu)->blkg = blkg; |
336 | } |
337 | |
338 | for (i = 0; i < BLKCG_MAX_POLS; i++) { |
339 | struct blkcg_policy *pol = blkcg_policy[i]; |
340 | struct blkg_policy_data *pd; |
341 | |
342 | if (!blkcg_policy_enabled(q: disk->queue, pol)) |
343 | continue; |
344 | |
345 | /* alloc per-policy data and attach it to blkg */ |
346 | pd = pol->pd_alloc_fn(disk, blkcg, gfp_mask); |
347 | if (!pd) |
348 | goto out_free_pds; |
349 | blkg->pd[i] = pd; |
350 | pd->blkg = blkg; |
351 | pd->plid = i; |
352 | pd->online = false; |
353 | } |
354 | |
355 | return blkg; |
356 | |
357 | out_free_pds: |
358 | while (--i >= 0) |
359 | if (blkg->pd[i]) |
360 | blkcg_policy[i]->pd_free_fn(blkg->pd[i]); |
361 | blk_put_queue(disk->queue); |
362 | out_free_iostat: |
363 | free_percpu(pdata: blkg->iostat_cpu); |
364 | out_exit_refcnt: |
365 | percpu_ref_exit(ref: &blkg->refcnt); |
366 | out_free_blkg: |
367 | kfree(objp: blkg); |
368 | return NULL; |
369 | } |
370 | |
371 | /* |
372 | * If @new_blkg is %NULL, this function tries to allocate a new one as |
373 | * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return. |
374 | */ |
375 | static struct blkcg_gq *blkg_create(struct blkcg *blkcg, struct gendisk *disk, |
376 | struct blkcg_gq *new_blkg) |
377 | { |
378 | struct blkcg_gq *blkg; |
379 | int i, ret; |
380 | |
381 | lockdep_assert_held(&disk->queue->queue_lock); |
382 | |
383 | /* request_queue is dying, do not create/recreate a blkg */ |
384 | if (blk_queue_dying(disk->queue)) { |
385 | ret = -ENODEV; |
386 | goto err_free_blkg; |
387 | } |
388 | |
389 | /* blkg holds a reference to blkcg */ |
390 | if (!css_tryget_online(css: &blkcg->css)) { |
391 | ret = -ENODEV; |
392 | goto err_free_blkg; |
393 | } |
394 | |
395 | /* allocate */ |
396 | if (!new_blkg) { |
397 | new_blkg = blkg_alloc(blkcg, disk, GFP_NOWAIT | __GFP_NOWARN); |
398 | if (unlikely(!new_blkg)) { |
399 | ret = -ENOMEM; |
400 | goto err_put_css; |
401 | } |
402 | } |
403 | blkg = new_blkg; |
404 | |
405 | /* link parent */ |
406 | if (blkcg_parent(blkcg)) { |
407 | blkg->parent = blkg_lookup(blkcg: blkcg_parent(blkcg), q: disk->queue); |
408 | if (WARN_ON_ONCE(!blkg->parent)) { |
409 | ret = -ENODEV; |
410 | goto err_put_css; |
411 | } |
412 | blkg_get(blkg: blkg->parent); |
413 | } |
414 | |
415 | /* invoke per-policy init */ |
416 | for (i = 0; i < BLKCG_MAX_POLS; i++) { |
417 | struct blkcg_policy *pol = blkcg_policy[i]; |
418 | |
419 | if (blkg->pd[i] && pol->pd_init_fn) |
420 | pol->pd_init_fn(blkg->pd[i]); |
421 | } |
422 | |
423 | /* insert */ |
424 | spin_lock(lock: &blkcg->lock); |
425 | ret = radix_tree_insert(&blkcg->blkg_tree, index: disk->queue->id, blkg); |
426 | if (likely(!ret)) { |
427 | hlist_add_head_rcu(n: &blkg->blkcg_node, h: &blkcg->blkg_list); |
428 | list_add(new: &blkg->q_node, head: &disk->queue->blkg_list); |
429 | |
430 | for (i = 0; i < BLKCG_MAX_POLS; i++) { |
431 | struct blkcg_policy *pol = blkcg_policy[i]; |
432 | |
433 | if (blkg->pd[i]) { |
434 | if (pol->pd_online_fn) |
435 | pol->pd_online_fn(blkg->pd[i]); |
436 | blkg->pd[i]->online = true; |
437 | } |
438 | } |
439 | } |
440 | blkg->online = true; |
441 | spin_unlock(lock: &blkcg->lock); |
442 | |
443 | if (!ret) |
444 | return blkg; |
445 | |
446 | /* @blkg failed fully initialized, use the usual release path */ |
447 | blkg_put(blkg); |
448 | return ERR_PTR(error: ret); |
449 | |
450 | err_put_css: |
451 | css_put(css: &blkcg->css); |
452 | err_free_blkg: |
453 | if (new_blkg) |
454 | blkg_free(blkg: new_blkg); |
455 | return ERR_PTR(error: ret); |
456 | } |
457 | |
458 | /** |
459 | * blkg_lookup_create - lookup blkg, try to create one if not there |
460 | * @blkcg: blkcg of interest |
461 | * @disk: gendisk of interest |
462 | * |
463 | * Lookup blkg for the @blkcg - @disk pair. If it doesn't exist, try to |
464 | * create one. blkg creation is performed recursively from blkcg_root such |
465 | * that all non-root blkg's have access to the parent blkg. This function |
466 | * should be called under RCU read lock and takes @disk->queue->queue_lock. |
467 | * |
468 | * Returns the blkg or the closest blkg if blkg_create() fails as it walks |
469 | * down from root. |
470 | */ |
471 | static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg, |
472 | struct gendisk *disk) |
473 | { |
474 | struct request_queue *q = disk->queue; |
475 | struct blkcg_gq *blkg; |
476 | unsigned long flags; |
477 | |
478 | WARN_ON_ONCE(!rcu_read_lock_held()); |
479 | |
480 | blkg = blkg_lookup(blkcg, q); |
481 | if (blkg) |
482 | return blkg; |
483 | |
484 | spin_lock_irqsave(&q->queue_lock, flags); |
485 | blkg = blkg_lookup(blkcg, q); |
486 | if (blkg) { |
487 | if (blkcg != &blkcg_root && |
488 | blkg != rcu_dereference(blkcg->blkg_hint)) |
489 | rcu_assign_pointer(blkcg->blkg_hint, blkg); |
490 | goto found; |
491 | } |
492 | |
493 | /* |
494 | * Create blkgs walking down from blkcg_root to @blkcg, so that all |
495 | * non-root blkgs have access to their parents. Returns the closest |
496 | * blkg to the intended blkg should blkg_create() fail. |
497 | */ |
498 | while (true) { |
499 | struct blkcg *pos = blkcg; |
500 | struct blkcg *parent = blkcg_parent(blkcg); |
501 | struct blkcg_gq *ret_blkg = q->root_blkg; |
502 | |
503 | while (parent) { |
504 | blkg = blkg_lookup(blkcg: parent, q); |
505 | if (blkg) { |
506 | /* remember closest blkg */ |
507 | ret_blkg = blkg; |
508 | break; |
509 | } |
510 | pos = parent; |
511 | parent = blkcg_parent(blkcg: parent); |
512 | } |
513 | |
514 | blkg = blkg_create(blkcg: pos, disk, NULL); |
515 | if (IS_ERR(ptr: blkg)) { |
516 | blkg = ret_blkg; |
517 | break; |
518 | } |
519 | if (pos == blkcg) |
520 | break; |
521 | } |
522 | |
523 | found: |
524 | spin_unlock_irqrestore(lock: &q->queue_lock, flags); |
525 | return blkg; |
526 | } |
527 | |
528 | static void blkg_destroy(struct blkcg_gq *blkg) |
529 | { |
530 | struct blkcg *blkcg = blkg->blkcg; |
531 | int i; |
532 | |
533 | lockdep_assert_held(&blkg->q->queue_lock); |
534 | lockdep_assert_held(&blkcg->lock); |
535 | |
536 | /* |
537 | * blkg stays on the queue list until blkg_free_workfn(), see details in |
538 | * blkg_free_workfn(), hence this function can be called from |
539 | * blkcg_destroy_blkgs() first and again from blkg_destroy_all() before |
540 | * blkg_free_workfn(). |
541 | */ |
542 | if (hlist_unhashed(h: &blkg->blkcg_node)) |
543 | return; |
544 | |
545 | for (i = 0; i < BLKCG_MAX_POLS; i++) { |
546 | struct blkcg_policy *pol = blkcg_policy[i]; |
547 | |
548 | if (blkg->pd[i] && blkg->pd[i]->online) { |
549 | blkg->pd[i]->online = false; |
550 | if (pol->pd_offline_fn) |
551 | pol->pd_offline_fn(blkg->pd[i]); |
552 | } |
553 | } |
554 | |
555 | blkg->online = false; |
556 | |
557 | radix_tree_delete(&blkcg->blkg_tree, blkg->q->id); |
558 | hlist_del_init_rcu(n: &blkg->blkcg_node); |
559 | |
560 | /* |
561 | * Both setting lookup hint to and clearing it from @blkg are done |
562 | * under queue_lock. If it's not pointing to @blkg now, it never |
563 | * will. Hint assignment itself can race safely. |
564 | */ |
565 | if (rcu_access_pointer(blkcg->blkg_hint) == blkg) |
566 | rcu_assign_pointer(blkcg->blkg_hint, NULL); |
567 | |
568 | /* |
569 | * Put the reference taken at the time of creation so that when all |
570 | * queues are gone, group can be destroyed. |
571 | */ |
572 | percpu_ref_kill(ref: &blkg->refcnt); |
573 | } |
574 | |
575 | static void blkg_destroy_all(struct gendisk *disk) |
576 | { |
577 | struct request_queue *q = disk->queue; |
578 | struct blkcg_gq *blkg, *n; |
579 | int count = BLKG_DESTROY_BATCH_SIZE; |
580 | |
581 | restart: |
582 | spin_lock_irq(lock: &q->queue_lock); |
583 | list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) { |
584 | struct blkcg *blkcg = blkg->blkcg; |
585 | |
586 | if (hlist_unhashed(h: &blkg->blkcg_node)) |
587 | continue; |
588 | |
589 | spin_lock(lock: &blkcg->lock); |
590 | blkg_destroy(blkg); |
591 | spin_unlock(lock: &blkcg->lock); |
592 | |
593 | /* |
594 | * in order to avoid holding the spin lock for too long, release |
595 | * it when a batch of blkgs are destroyed. |
596 | */ |
597 | if (!(--count)) { |
598 | count = BLKG_DESTROY_BATCH_SIZE; |
599 | spin_unlock_irq(lock: &q->queue_lock); |
600 | cond_resched(); |
601 | goto restart; |
602 | } |
603 | } |
604 | |
605 | q->root_blkg = NULL; |
606 | spin_unlock_irq(lock: &q->queue_lock); |
607 | } |
608 | |
609 | static int blkcg_reset_stats(struct cgroup_subsys_state *css, |
610 | struct cftype *cftype, u64 val) |
611 | { |
612 | struct blkcg *blkcg = css_to_blkcg(css); |
613 | struct blkcg_gq *blkg; |
614 | int i, cpu; |
615 | |
616 | mutex_lock(&blkcg_pol_mutex); |
617 | spin_lock_irq(lock: &blkcg->lock); |
618 | |
619 | /* |
620 | * Note that stat reset is racy - it doesn't synchronize against |
621 | * stat updates. This is a debug feature which shouldn't exist |
622 | * anyway. If you get hit by a race, retry. |
623 | */ |
624 | hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) { |
625 | for_each_possible_cpu(cpu) { |
626 | struct blkg_iostat_set *bis = |
627 | per_cpu_ptr(blkg->iostat_cpu, cpu); |
628 | memset(bis, 0, sizeof(*bis)); |
629 | |
630 | /* Re-initialize the cleared blkg_iostat_set */ |
631 | u64_stats_init(syncp: &bis->sync); |
632 | bis->blkg = blkg; |
633 | } |
634 | memset(&blkg->iostat, 0, sizeof(blkg->iostat)); |
635 | u64_stats_init(syncp: &blkg->iostat.sync); |
636 | |
637 | for (i = 0; i < BLKCG_MAX_POLS; i++) { |
638 | struct blkcg_policy *pol = blkcg_policy[i]; |
639 | |
640 | if (blkg->pd[i] && pol->pd_reset_stats_fn) |
641 | pol->pd_reset_stats_fn(blkg->pd[i]); |
642 | } |
643 | } |
644 | |
645 | spin_unlock_irq(lock: &blkcg->lock); |
646 | mutex_unlock(lock: &blkcg_pol_mutex); |
647 | return 0; |
648 | } |
649 | |
650 | const char *blkg_dev_name(struct blkcg_gq *blkg) |
651 | { |
652 | if (!blkg->q->disk) |
653 | return NULL; |
654 | return bdi_dev_name(bdi: blkg->q->disk->bdi); |
655 | } |
656 | |
657 | /** |
658 | * blkcg_print_blkgs - helper for printing per-blkg data |
659 | * @sf: seq_file to print to |
660 | * @blkcg: blkcg of interest |
661 | * @prfill: fill function to print out a blkg |
662 | * @pol: policy in question |
663 | * @data: data to be passed to @prfill |
664 | * @show_total: to print out sum of prfill return values or not |
665 | * |
666 | * This function invokes @prfill on each blkg of @blkcg if pd for the |
667 | * policy specified by @pol exists. @prfill is invoked with @sf, the |
668 | * policy data and @data and the matching queue lock held. If @show_total |
669 | * is %true, the sum of the return values from @prfill is printed with |
670 | * "Total" label at the end. |
671 | * |
672 | * This is to be used to construct print functions for |
673 | * cftype->read_seq_string method. |
674 | */ |
675 | void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg, |
676 | u64 (*prfill)(struct seq_file *, |
677 | struct blkg_policy_data *, int), |
678 | const struct blkcg_policy *pol, int data, |
679 | bool show_total) |
680 | { |
681 | struct blkcg_gq *blkg; |
682 | u64 total = 0; |
683 | |
684 | rcu_read_lock(); |
685 | hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) { |
686 | spin_lock_irq(lock: &blkg->q->queue_lock); |
687 | if (blkcg_policy_enabled(q: blkg->q, pol)) |
688 | total += prfill(sf, blkg->pd[pol->plid], data); |
689 | spin_unlock_irq(lock: &blkg->q->queue_lock); |
690 | } |
691 | rcu_read_unlock(); |
692 | |
693 | if (show_total) |
694 | seq_printf(m: sf, fmt: "Total %llu\n" , (unsigned long long)total); |
695 | } |
696 | EXPORT_SYMBOL_GPL(blkcg_print_blkgs); |
697 | |
698 | /** |
699 | * __blkg_prfill_u64 - prfill helper for a single u64 value |
700 | * @sf: seq_file to print to |
701 | * @pd: policy private data of interest |
702 | * @v: value to print |
703 | * |
704 | * Print @v to @sf for the device associated with @pd. |
705 | */ |
706 | u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v) |
707 | { |
708 | const char *dname = blkg_dev_name(blkg: pd->blkg); |
709 | |
710 | if (!dname) |
711 | return 0; |
712 | |
713 | seq_printf(m: sf, fmt: "%s %llu\n" , dname, (unsigned long long)v); |
714 | return v; |
715 | } |
716 | EXPORT_SYMBOL_GPL(__blkg_prfill_u64); |
717 | |
718 | /** |
719 | * blkg_conf_init - initialize a blkg_conf_ctx |
720 | * @ctx: blkg_conf_ctx to initialize |
721 | * @input: input string |
722 | * |
723 | * Initialize @ctx which can be used to parse blkg config input string @input. |
724 | * Once initialized, @ctx can be used with blkg_conf_open_bdev() and |
725 | * blkg_conf_prep(), and must be cleaned up with blkg_conf_exit(). |
726 | */ |
727 | void blkg_conf_init(struct blkg_conf_ctx *ctx, char *input) |
728 | { |
729 | *ctx = (struct blkg_conf_ctx){ .input = input }; |
730 | } |
731 | EXPORT_SYMBOL_GPL(blkg_conf_init); |
732 | |
733 | /** |
734 | * blkg_conf_open_bdev - parse and open bdev for per-blkg config update |
735 | * @ctx: blkg_conf_ctx initialized with blkg_conf_init() |
736 | * |
737 | * Parse the device node prefix part, MAJ:MIN, of per-blkg config update from |
738 | * @ctx->input and get and store the matching bdev in @ctx->bdev. @ctx->body is |
739 | * set to point past the device node prefix. |
740 | * |
741 | * This function may be called multiple times on @ctx and the extra calls become |
742 | * NOOPs. blkg_conf_prep() implicitly calls this function. Use this function |
743 | * explicitly if bdev access is needed without resolving the blkcg / policy part |
744 | * of @ctx->input. Returns -errno on error. |
745 | */ |
746 | int blkg_conf_open_bdev(struct blkg_conf_ctx *ctx) |
747 | { |
748 | char *input = ctx->input; |
749 | unsigned int major, minor; |
750 | struct block_device *bdev; |
751 | int key_len; |
752 | |
753 | if (ctx->bdev) |
754 | return 0; |
755 | |
756 | if (sscanf(input, "%u:%u%n" , &major, &minor, &key_len) != 2) |
757 | return -EINVAL; |
758 | |
759 | input += key_len; |
760 | if (!isspace(*input)) |
761 | return -EINVAL; |
762 | input = skip_spaces(input); |
763 | |
764 | bdev = blkdev_get_no_open(MKDEV(major, minor)); |
765 | if (!bdev) |
766 | return -ENODEV; |
767 | if (bdev_is_partition(bdev)) { |
768 | blkdev_put_no_open(bdev); |
769 | return -ENODEV; |
770 | } |
771 | |
772 | mutex_lock(&bdev->bd_queue->rq_qos_mutex); |
773 | if (!disk_live(disk: bdev->bd_disk)) { |
774 | blkdev_put_no_open(bdev); |
775 | mutex_unlock(lock: &bdev->bd_queue->rq_qos_mutex); |
776 | return -ENODEV; |
777 | } |
778 | |
779 | ctx->body = input; |
780 | ctx->bdev = bdev; |
781 | return 0; |
782 | } |
783 | |
784 | /** |
785 | * blkg_conf_prep - parse and prepare for per-blkg config update |
786 | * @blkcg: target block cgroup |
787 | * @pol: target policy |
788 | * @ctx: blkg_conf_ctx initialized with blkg_conf_init() |
789 | * |
790 | * Parse per-blkg config update from @ctx->input and initialize @ctx |
791 | * accordingly. On success, @ctx->body points to the part of @ctx->input |
792 | * following MAJ:MIN, @ctx->bdev points to the target block device and |
793 | * @ctx->blkg to the blkg being configured. |
794 | * |
795 | * blkg_conf_open_bdev() may be called on @ctx beforehand. On success, this |
796 | * function returns with queue lock held and must be followed by |
797 | * blkg_conf_exit(). |
798 | */ |
799 | int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol, |
800 | struct blkg_conf_ctx *ctx) |
801 | __acquires(&bdev->bd_queue->queue_lock) |
802 | { |
803 | struct gendisk *disk; |
804 | struct request_queue *q; |
805 | struct blkcg_gq *blkg; |
806 | int ret; |
807 | |
808 | ret = blkg_conf_open_bdev(ctx); |
809 | if (ret) |
810 | return ret; |
811 | |
812 | disk = ctx->bdev->bd_disk; |
813 | q = disk->queue; |
814 | |
815 | /* |
816 | * blkcg_deactivate_policy() requires queue to be frozen, we can grab |
817 | * q_usage_counter to prevent concurrent with blkcg_deactivate_policy(). |
818 | */ |
819 | ret = blk_queue_enter(q, flags: 0); |
820 | if (ret) |
821 | goto fail; |
822 | |
823 | spin_lock_irq(lock: &q->queue_lock); |
824 | |
825 | if (!blkcg_policy_enabled(q, pol)) { |
826 | ret = -EOPNOTSUPP; |
827 | goto fail_unlock; |
828 | } |
829 | |
830 | blkg = blkg_lookup(blkcg, q); |
831 | if (blkg) |
832 | goto success; |
833 | |
834 | /* |
835 | * Create blkgs walking down from blkcg_root to @blkcg, so that all |
836 | * non-root blkgs have access to their parents. |
837 | */ |
838 | while (true) { |
839 | struct blkcg *pos = blkcg; |
840 | struct blkcg *parent; |
841 | struct blkcg_gq *new_blkg; |
842 | |
843 | parent = blkcg_parent(blkcg); |
844 | while (parent && !blkg_lookup(blkcg: parent, q)) { |
845 | pos = parent; |
846 | parent = blkcg_parent(blkcg: parent); |
847 | } |
848 | |
849 | /* Drop locks to do new blkg allocation with GFP_KERNEL. */ |
850 | spin_unlock_irq(lock: &q->queue_lock); |
851 | |
852 | new_blkg = blkg_alloc(blkcg: pos, disk, GFP_KERNEL); |
853 | if (unlikely(!new_blkg)) { |
854 | ret = -ENOMEM; |
855 | goto fail_exit_queue; |
856 | } |
857 | |
858 | if (radix_tree_preload(GFP_KERNEL)) { |
859 | blkg_free(blkg: new_blkg); |
860 | ret = -ENOMEM; |
861 | goto fail_exit_queue; |
862 | } |
863 | |
864 | spin_lock_irq(lock: &q->queue_lock); |
865 | |
866 | if (!blkcg_policy_enabled(q, pol)) { |
867 | blkg_free(blkg: new_blkg); |
868 | ret = -EOPNOTSUPP; |
869 | goto fail_preloaded; |
870 | } |
871 | |
872 | blkg = blkg_lookup(blkcg: pos, q); |
873 | if (blkg) { |
874 | blkg_free(blkg: new_blkg); |
875 | } else { |
876 | blkg = blkg_create(blkcg: pos, disk, new_blkg); |
877 | if (IS_ERR(ptr: blkg)) { |
878 | ret = PTR_ERR(ptr: blkg); |
879 | goto fail_preloaded; |
880 | } |
881 | } |
882 | |
883 | radix_tree_preload_end(); |
884 | |
885 | if (pos == blkcg) |
886 | goto success; |
887 | } |
888 | success: |
889 | blk_queue_exit(q); |
890 | ctx->blkg = blkg; |
891 | return 0; |
892 | |
893 | fail_preloaded: |
894 | radix_tree_preload_end(); |
895 | fail_unlock: |
896 | spin_unlock_irq(lock: &q->queue_lock); |
897 | fail_exit_queue: |
898 | blk_queue_exit(q); |
899 | fail: |
900 | /* |
901 | * If queue was bypassing, we should retry. Do so after a |
902 | * short msleep(). It isn't strictly necessary but queue |
903 | * can be bypassing for some time and it's always nice to |
904 | * avoid busy looping. |
905 | */ |
906 | if (ret == -EBUSY) { |
907 | msleep(msecs: 10); |
908 | ret = restart_syscall(); |
909 | } |
910 | return ret; |
911 | } |
912 | EXPORT_SYMBOL_GPL(blkg_conf_prep); |
913 | |
914 | /** |
915 | * blkg_conf_exit - clean up per-blkg config update |
916 | * @ctx: blkg_conf_ctx initialized with blkg_conf_init() |
917 | * |
918 | * Clean up after per-blkg config update. This function must be called on all |
919 | * blkg_conf_ctx's initialized with blkg_conf_init(). |
920 | */ |
921 | void blkg_conf_exit(struct blkg_conf_ctx *ctx) |
922 | __releases(&ctx->bdev->bd_queue->queue_lock) |
923 | __releases(&ctx->bdev->bd_queue->rq_qos_mutex) |
924 | { |
925 | if (ctx->blkg) { |
926 | spin_unlock_irq(lock: &bdev_get_queue(bdev: ctx->bdev)->queue_lock); |
927 | ctx->blkg = NULL; |
928 | } |
929 | |
930 | if (ctx->bdev) { |
931 | mutex_unlock(lock: &ctx->bdev->bd_queue->rq_qos_mutex); |
932 | blkdev_put_no_open(bdev: ctx->bdev); |
933 | ctx->body = NULL; |
934 | ctx->bdev = NULL; |
935 | } |
936 | } |
937 | EXPORT_SYMBOL_GPL(blkg_conf_exit); |
938 | |
939 | static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src) |
940 | { |
941 | int i; |
942 | |
943 | for (i = 0; i < BLKG_IOSTAT_NR; i++) { |
944 | dst->bytes[i] = src->bytes[i]; |
945 | dst->ios[i] = src->ios[i]; |
946 | } |
947 | } |
948 | |
949 | static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src) |
950 | { |
951 | int i; |
952 | |
953 | for (i = 0; i < BLKG_IOSTAT_NR; i++) { |
954 | dst->bytes[i] += src->bytes[i]; |
955 | dst->ios[i] += src->ios[i]; |
956 | } |
957 | } |
958 | |
959 | static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src) |
960 | { |
961 | int i; |
962 | |
963 | for (i = 0; i < BLKG_IOSTAT_NR; i++) { |
964 | dst->bytes[i] -= src->bytes[i]; |
965 | dst->ios[i] -= src->ios[i]; |
966 | } |
967 | } |
968 | |
969 | static void blkcg_iostat_update(struct blkcg_gq *blkg, struct blkg_iostat *cur, |
970 | struct blkg_iostat *last) |
971 | { |
972 | struct blkg_iostat delta; |
973 | unsigned long flags; |
974 | |
975 | /* propagate percpu delta to global */ |
976 | flags = u64_stats_update_begin_irqsave(syncp: &blkg->iostat.sync); |
977 | blkg_iostat_set(dst: &delta, src: cur); |
978 | blkg_iostat_sub(dst: &delta, src: last); |
979 | blkg_iostat_add(dst: &blkg->iostat.cur, src: &delta); |
980 | blkg_iostat_add(dst: last, src: &delta); |
981 | u64_stats_update_end_irqrestore(syncp: &blkg->iostat.sync, flags); |
982 | } |
983 | |
984 | static void __blkcg_rstat_flush(struct blkcg *blkcg, int cpu) |
985 | { |
986 | struct llist_head *lhead = per_cpu_ptr(blkcg->lhead, cpu); |
987 | struct llist_node *lnode; |
988 | struct blkg_iostat_set *bisc, *next_bisc; |
989 | unsigned long flags; |
990 | |
991 | rcu_read_lock(); |
992 | |
993 | lnode = llist_del_all(head: lhead); |
994 | if (!lnode) |
995 | goto out; |
996 | |
997 | /* |
998 | * For covering concurrent parent blkg update from blkg_release(). |
999 | * |
1000 | * When flushing from cgroup, cgroup_rstat_lock is always held, so |
1001 | * this lock won't cause contention most of time. |
1002 | */ |
1003 | raw_spin_lock_irqsave(&blkg_stat_lock, flags); |
1004 | |
1005 | /* |
1006 | * Iterate only the iostat_cpu's queued in the lockless list. |
1007 | */ |
1008 | llist_for_each_entry_safe(bisc, next_bisc, lnode, lnode) { |
1009 | struct blkcg_gq *blkg = bisc->blkg; |
1010 | struct blkcg_gq *parent = blkg->parent; |
1011 | struct blkg_iostat cur; |
1012 | unsigned int seq; |
1013 | |
1014 | WRITE_ONCE(bisc->lqueued, false); |
1015 | |
1016 | /* fetch the current per-cpu values */ |
1017 | do { |
1018 | seq = u64_stats_fetch_begin(syncp: &bisc->sync); |
1019 | blkg_iostat_set(dst: &cur, src: &bisc->cur); |
1020 | } while (u64_stats_fetch_retry(syncp: &bisc->sync, start: seq)); |
1021 | |
1022 | blkcg_iostat_update(blkg, cur: &cur, last: &bisc->last); |
1023 | |
1024 | /* propagate global delta to parent (unless that's root) */ |
1025 | if (parent && parent->parent) |
1026 | blkcg_iostat_update(blkg: parent, cur: &blkg->iostat.cur, |
1027 | last: &blkg->iostat.last); |
1028 | } |
1029 | raw_spin_unlock_irqrestore(&blkg_stat_lock, flags); |
1030 | out: |
1031 | rcu_read_unlock(); |
1032 | } |
1033 | |
1034 | static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu) |
1035 | { |
1036 | /* Root-level stats are sourced from system-wide IO stats */ |
1037 | if (cgroup_parent(cgrp: css->cgroup)) |
1038 | __blkcg_rstat_flush(blkcg: css_to_blkcg(css), cpu); |
1039 | } |
1040 | |
1041 | /* |
1042 | * We source root cgroup stats from the system-wide stats to avoid |
1043 | * tracking the same information twice and incurring overhead when no |
1044 | * cgroups are defined. For that reason, cgroup_rstat_flush in |
1045 | * blkcg_print_stat does not actually fill out the iostat in the root |
1046 | * cgroup's blkcg_gq. |
1047 | * |
1048 | * However, we would like to re-use the printing code between the root and |
1049 | * non-root cgroups to the extent possible. For that reason, we simulate |
1050 | * flushing the root cgroup's stats by explicitly filling in the iostat |
1051 | * with disk level statistics. |
1052 | */ |
1053 | static void blkcg_fill_root_iostats(void) |
1054 | { |
1055 | struct class_dev_iter iter; |
1056 | struct device *dev; |
1057 | |
1058 | class_dev_iter_init(iter: &iter, class: &block_class, NULL, type: &disk_type); |
1059 | while ((dev = class_dev_iter_next(iter: &iter))) { |
1060 | struct block_device *bdev = dev_to_bdev(dev); |
1061 | struct blkcg_gq *blkg = bdev->bd_disk->queue->root_blkg; |
1062 | struct blkg_iostat tmp; |
1063 | int cpu; |
1064 | unsigned long flags; |
1065 | |
1066 | memset(&tmp, 0, sizeof(tmp)); |
1067 | for_each_possible_cpu(cpu) { |
1068 | struct disk_stats *cpu_dkstats; |
1069 | |
1070 | cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu); |
1071 | tmp.ios[BLKG_IOSTAT_READ] += |
1072 | cpu_dkstats->ios[STAT_READ]; |
1073 | tmp.ios[BLKG_IOSTAT_WRITE] += |
1074 | cpu_dkstats->ios[STAT_WRITE]; |
1075 | tmp.ios[BLKG_IOSTAT_DISCARD] += |
1076 | cpu_dkstats->ios[STAT_DISCARD]; |
1077 | // convert sectors to bytes |
1078 | tmp.bytes[BLKG_IOSTAT_READ] += |
1079 | cpu_dkstats->sectors[STAT_READ] << 9; |
1080 | tmp.bytes[BLKG_IOSTAT_WRITE] += |
1081 | cpu_dkstats->sectors[STAT_WRITE] << 9; |
1082 | tmp.bytes[BLKG_IOSTAT_DISCARD] += |
1083 | cpu_dkstats->sectors[STAT_DISCARD] << 9; |
1084 | } |
1085 | |
1086 | flags = u64_stats_update_begin_irqsave(syncp: &blkg->iostat.sync); |
1087 | blkg_iostat_set(dst: &blkg->iostat.cur, src: &tmp); |
1088 | u64_stats_update_end_irqrestore(syncp: &blkg->iostat.sync, flags); |
1089 | } |
1090 | } |
1091 | |
1092 | static void blkcg_print_one_stat(struct blkcg_gq *blkg, struct seq_file *s) |
1093 | { |
1094 | struct blkg_iostat_set *bis = &blkg->iostat; |
1095 | u64 rbytes, wbytes, rios, wios, dbytes, dios; |
1096 | const char *dname; |
1097 | unsigned seq; |
1098 | int i; |
1099 | |
1100 | if (!blkg->online) |
1101 | return; |
1102 | |
1103 | dname = blkg_dev_name(blkg); |
1104 | if (!dname) |
1105 | return; |
1106 | |
1107 | seq_printf(m: s, fmt: "%s " , dname); |
1108 | |
1109 | do { |
1110 | seq = u64_stats_fetch_begin(syncp: &bis->sync); |
1111 | |
1112 | rbytes = bis->cur.bytes[BLKG_IOSTAT_READ]; |
1113 | wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE]; |
1114 | dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD]; |
1115 | rios = bis->cur.ios[BLKG_IOSTAT_READ]; |
1116 | wios = bis->cur.ios[BLKG_IOSTAT_WRITE]; |
1117 | dios = bis->cur.ios[BLKG_IOSTAT_DISCARD]; |
1118 | } while (u64_stats_fetch_retry(syncp: &bis->sync, start: seq)); |
1119 | |
1120 | if (rbytes || wbytes || rios || wios) { |
1121 | seq_printf(m: s, fmt: "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu" , |
1122 | rbytes, wbytes, rios, wios, |
1123 | dbytes, dios); |
1124 | } |
1125 | |
1126 | if (blkcg_debug_stats && atomic_read(v: &blkg->use_delay)) { |
1127 | seq_printf(m: s, fmt: " use_delay=%d delay_nsec=%llu" , |
1128 | atomic_read(v: &blkg->use_delay), |
1129 | atomic64_read(v: &blkg->delay_nsec)); |
1130 | } |
1131 | |
1132 | for (i = 0; i < BLKCG_MAX_POLS; i++) { |
1133 | struct blkcg_policy *pol = blkcg_policy[i]; |
1134 | |
1135 | if (!blkg->pd[i] || !pol->pd_stat_fn) |
1136 | continue; |
1137 | |
1138 | pol->pd_stat_fn(blkg->pd[i], s); |
1139 | } |
1140 | |
1141 | seq_puts(m: s, s: "\n" ); |
1142 | } |
1143 | |
1144 | static int blkcg_print_stat(struct seq_file *sf, void *v) |
1145 | { |
1146 | struct blkcg *blkcg = css_to_blkcg(css: seq_css(seq: sf)); |
1147 | struct blkcg_gq *blkg; |
1148 | |
1149 | if (!seq_css(seq: sf)->parent) |
1150 | blkcg_fill_root_iostats(); |
1151 | else |
1152 | cgroup_rstat_flush(cgrp: blkcg->css.cgroup); |
1153 | |
1154 | rcu_read_lock(); |
1155 | hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) { |
1156 | spin_lock_irq(lock: &blkg->q->queue_lock); |
1157 | blkcg_print_one_stat(blkg, s: sf); |
1158 | spin_unlock_irq(lock: &blkg->q->queue_lock); |
1159 | } |
1160 | rcu_read_unlock(); |
1161 | return 0; |
1162 | } |
1163 | |
1164 | static struct cftype blkcg_files[] = { |
1165 | { |
1166 | .name = "stat" , |
1167 | .seq_show = blkcg_print_stat, |
1168 | }, |
1169 | { } /* terminate */ |
1170 | }; |
1171 | |
1172 | static struct cftype blkcg_legacy_files[] = { |
1173 | { |
1174 | .name = "reset_stats" , |
1175 | .write_u64 = blkcg_reset_stats, |
1176 | }, |
1177 | { } /* terminate */ |
1178 | }; |
1179 | |
1180 | #ifdef CONFIG_CGROUP_WRITEBACK |
1181 | struct list_head *blkcg_get_cgwb_list(struct cgroup_subsys_state *css) |
1182 | { |
1183 | return &css_to_blkcg(css)->cgwb_list; |
1184 | } |
1185 | #endif |
1186 | |
1187 | /* |
1188 | * blkcg destruction is a three-stage process. |
1189 | * |
1190 | * 1. Destruction starts. The blkcg_css_offline() callback is invoked |
1191 | * which offlines writeback. Here we tie the next stage of blkg destruction |
1192 | * to the completion of writeback associated with the blkcg. This lets us |
1193 | * avoid punting potentially large amounts of outstanding writeback to root |
1194 | * while maintaining any ongoing policies. The next stage is triggered when |
1195 | * the nr_cgwbs count goes to zero. |
1196 | * |
1197 | * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called |
1198 | * and handles the destruction of blkgs. Here the css reference held by |
1199 | * the blkg is put back eventually allowing blkcg_css_free() to be called. |
1200 | * This work may occur in cgwb_release_workfn() on the cgwb_release |
1201 | * workqueue. Any submitted ios that fail to get the blkg ref will be |
1202 | * punted to the root_blkg. |
1203 | * |
1204 | * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called. |
1205 | * This finally frees the blkcg. |
1206 | */ |
1207 | |
1208 | /** |
1209 | * blkcg_destroy_blkgs - responsible for shooting down blkgs |
1210 | * @blkcg: blkcg of interest |
1211 | * |
1212 | * blkgs should be removed while holding both q and blkcg locks. As blkcg lock |
1213 | * is nested inside q lock, this function performs reverse double lock dancing. |
1214 | * Destroying the blkgs releases the reference held on the blkcg's css allowing |
1215 | * blkcg_css_free to eventually be called. |
1216 | * |
1217 | * This is the blkcg counterpart of ioc_release_fn(). |
1218 | */ |
1219 | static void blkcg_destroy_blkgs(struct blkcg *blkcg) |
1220 | { |
1221 | might_sleep(); |
1222 | |
1223 | spin_lock_irq(lock: &blkcg->lock); |
1224 | |
1225 | while (!hlist_empty(h: &blkcg->blkg_list)) { |
1226 | struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first, |
1227 | struct blkcg_gq, blkcg_node); |
1228 | struct request_queue *q = blkg->q; |
1229 | |
1230 | if (need_resched() || !spin_trylock(lock: &q->queue_lock)) { |
1231 | /* |
1232 | * Given that the system can accumulate a huge number |
1233 | * of blkgs in pathological cases, check to see if we |
1234 | * need to rescheduling to avoid softlockup. |
1235 | */ |
1236 | spin_unlock_irq(lock: &blkcg->lock); |
1237 | cond_resched(); |
1238 | spin_lock_irq(lock: &blkcg->lock); |
1239 | continue; |
1240 | } |
1241 | |
1242 | blkg_destroy(blkg); |
1243 | spin_unlock(lock: &q->queue_lock); |
1244 | } |
1245 | |
1246 | spin_unlock_irq(lock: &blkcg->lock); |
1247 | } |
1248 | |
1249 | /** |
1250 | * blkcg_pin_online - pin online state |
1251 | * @blkcg_css: blkcg of interest |
1252 | * |
1253 | * While pinned, a blkcg is kept online. This is primarily used to |
1254 | * impedance-match blkg and cgwb lifetimes so that blkg doesn't go offline |
1255 | * while an associated cgwb is still active. |
1256 | */ |
1257 | void blkcg_pin_online(struct cgroup_subsys_state *blkcg_css) |
1258 | { |
1259 | refcount_inc(r: &css_to_blkcg(css: blkcg_css)->online_pin); |
1260 | } |
1261 | |
1262 | /** |
1263 | * blkcg_unpin_online - unpin online state |
1264 | * @blkcg_css: blkcg of interest |
1265 | * |
1266 | * This is primarily used to impedance-match blkg and cgwb lifetimes so |
1267 | * that blkg doesn't go offline while an associated cgwb is still active. |
1268 | * When this count goes to zero, all active cgwbs have finished so the |
1269 | * blkcg can continue destruction by calling blkcg_destroy_blkgs(). |
1270 | */ |
1271 | void blkcg_unpin_online(struct cgroup_subsys_state *blkcg_css) |
1272 | { |
1273 | struct blkcg *blkcg = css_to_blkcg(css: blkcg_css); |
1274 | |
1275 | do { |
1276 | if (!refcount_dec_and_test(r: &blkcg->online_pin)) |
1277 | break; |
1278 | blkcg_destroy_blkgs(blkcg); |
1279 | blkcg = blkcg_parent(blkcg); |
1280 | } while (blkcg); |
1281 | } |
1282 | |
1283 | /** |
1284 | * blkcg_css_offline - cgroup css_offline callback |
1285 | * @css: css of interest |
1286 | * |
1287 | * This function is called when @css is about to go away. Here the cgwbs are |
1288 | * offlined first and only once writeback associated with the blkcg has |
1289 | * finished do we start step 2 (see above). |
1290 | */ |
1291 | static void blkcg_css_offline(struct cgroup_subsys_state *css) |
1292 | { |
1293 | /* this prevents anyone from attaching or migrating to this blkcg */ |
1294 | wb_blkcg_offline(css); |
1295 | |
1296 | /* put the base online pin allowing step 2 to be triggered */ |
1297 | blkcg_unpin_online(blkcg_css: css); |
1298 | } |
1299 | |
1300 | static void blkcg_css_free(struct cgroup_subsys_state *css) |
1301 | { |
1302 | struct blkcg *blkcg = css_to_blkcg(css); |
1303 | int i; |
1304 | |
1305 | mutex_lock(&blkcg_pol_mutex); |
1306 | |
1307 | list_del(entry: &blkcg->all_blkcgs_node); |
1308 | |
1309 | for (i = 0; i < BLKCG_MAX_POLS; i++) |
1310 | if (blkcg->cpd[i]) |
1311 | blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]); |
1312 | |
1313 | mutex_unlock(lock: &blkcg_pol_mutex); |
1314 | |
1315 | free_percpu(pdata: blkcg->lhead); |
1316 | kfree(objp: blkcg); |
1317 | } |
1318 | |
1319 | static struct cgroup_subsys_state * |
1320 | blkcg_css_alloc(struct cgroup_subsys_state *parent_css) |
1321 | { |
1322 | struct blkcg *blkcg; |
1323 | int i; |
1324 | |
1325 | mutex_lock(&blkcg_pol_mutex); |
1326 | |
1327 | if (!parent_css) { |
1328 | blkcg = &blkcg_root; |
1329 | } else { |
1330 | blkcg = kzalloc(size: sizeof(*blkcg), GFP_KERNEL); |
1331 | if (!blkcg) |
1332 | goto unlock; |
1333 | } |
1334 | |
1335 | if (init_blkcg_llists(blkcg)) |
1336 | goto free_blkcg; |
1337 | |
1338 | for (i = 0; i < BLKCG_MAX_POLS ; i++) { |
1339 | struct blkcg_policy *pol = blkcg_policy[i]; |
1340 | struct blkcg_policy_data *cpd; |
1341 | |
1342 | /* |
1343 | * If the policy hasn't been attached yet, wait for it |
1344 | * to be attached before doing anything else. Otherwise, |
1345 | * check if the policy requires any specific per-cgroup |
1346 | * data: if it does, allocate and initialize it. |
1347 | */ |
1348 | if (!pol || !pol->cpd_alloc_fn) |
1349 | continue; |
1350 | |
1351 | cpd = pol->cpd_alloc_fn(GFP_KERNEL); |
1352 | if (!cpd) |
1353 | goto free_pd_blkcg; |
1354 | |
1355 | blkcg->cpd[i] = cpd; |
1356 | cpd->blkcg = blkcg; |
1357 | cpd->plid = i; |
1358 | } |
1359 | |
1360 | spin_lock_init(&blkcg->lock); |
1361 | refcount_set(r: &blkcg->online_pin, n: 1); |
1362 | INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN); |
1363 | INIT_HLIST_HEAD(&blkcg->blkg_list); |
1364 | #ifdef CONFIG_CGROUP_WRITEBACK |
1365 | INIT_LIST_HEAD(list: &blkcg->cgwb_list); |
1366 | #endif |
1367 | list_add_tail(new: &blkcg->all_blkcgs_node, head: &all_blkcgs); |
1368 | |
1369 | mutex_unlock(lock: &blkcg_pol_mutex); |
1370 | return &blkcg->css; |
1371 | |
1372 | free_pd_blkcg: |
1373 | for (i--; i >= 0; i--) |
1374 | if (blkcg->cpd[i]) |
1375 | blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]); |
1376 | free_percpu(pdata: blkcg->lhead); |
1377 | free_blkcg: |
1378 | if (blkcg != &blkcg_root) |
1379 | kfree(objp: blkcg); |
1380 | unlock: |
1381 | mutex_unlock(lock: &blkcg_pol_mutex); |
1382 | return ERR_PTR(error: -ENOMEM); |
1383 | } |
1384 | |
1385 | static int blkcg_css_online(struct cgroup_subsys_state *css) |
1386 | { |
1387 | struct blkcg *parent = blkcg_parent(blkcg: css_to_blkcg(css)); |
1388 | |
1389 | /* |
1390 | * blkcg_pin_online() is used to delay blkcg offline so that blkgs |
1391 | * don't go offline while cgwbs are still active on them. Pin the |
1392 | * parent so that offline always happens towards the root. |
1393 | */ |
1394 | if (parent) |
1395 | blkcg_pin_online(blkcg_css: &parent->css); |
1396 | return 0; |
1397 | } |
1398 | |
1399 | int blkcg_init_disk(struct gendisk *disk) |
1400 | { |
1401 | struct request_queue *q = disk->queue; |
1402 | struct blkcg_gq *new_blkg, *blkg; |
1403 | bool preloaded; |
1404 | int ret; |
1405 | |
1406 | INIT_LIST_HEAD(list: &q->blkg_list); |
1407 | mutex_init(&q->blkcg_mutex); |
1408 | |
1409 | new_blkg = blkg_alloc(blkcg: &blkcg_root, disk, GFP_KERNEL); |
1410 | if (!new_blkg) |
1411 | return -ENOMEM; |
1412 | |
1413 | preloaded = !radix_tree_preload(GFP_KERNEL); |
1414 | |
1415 | /* Make sure the root blkg exists. */ |
1416 | /* spin_lock_irq can serve as RCU read-side critical section. */ |
1417 | spin_lock_irq(lock: &q->queue_lock); |
1418 | blkg = blkg_create(blkcg: &blkcg_root, disk, new_blkg); |
1419 | if (IS_ERR(ptr: blkg)) |
1420 | goto err_unlock; |
1421 | q->root_blkg = blkg; |
1422 | spin_unlock_irq(lock: &q->queue_lock); |
1423 | |
1424 | if (preloaded) |
1425 | radix_tree_preload_end(); |
1426 | |
1427 | ret = blk_ioprio_init(disk); |
1428 | if (ret) |
1429 | goto err_destroy_all; |
1430 | |
1431 | ret = blk_throtl_init(disk); |
1432 | if (ret) |
1433 | goto err_ioprio_exit; |
1434 | |
1435 | return 0; |
1436 | |
1437 | err_ioprio_exit: |
1438 | blk_ioprio_exit(disk); |
1439 | err_destroy_all: |
1440 | blkg_destroy_all(disk); |
1441 | return ret; |
1442 | err_unlock: |
1443 | spin_unlock_irq(lock: &q->queue_lock); |
1444 | if (preloaded) |
1445 | radix_tree_preload_end(); |
1446 | return PTR_ERR(ptr: blkg); |
1447 | } |
1448 | |
1449 | void blkcg_exit_disk(struct gendisk *disk) |
1450 | { |
1451 | blkg_destroy_all(disk); |
1452 | blk_throtl_exit(disk); |
1453 | } |
1454 | |
1455 | static void blkcg_exit(struct task_struct *tsk) |
1456 | { |
1457 | if (tsk->throttle_disk) |
1458 | put_disk(disk: tsk->throttle_disk); |
1459 | tsk->throttle_disk = NULL; |
1460 | } |
1461 | |
1462 | struct cgroup_subsys io_cgrp_subsys = { |
1463 | .css_alloc = blkcg_css_alloc, |
1464 | .css_online = blkcg_css_online, |
1465 | .css_offline = blkcg_css_offline, |
1466 | .css_free = blkcg_css_free, |
1467 | .css_rstat_flush = blkcg_rstat_flush, |
1468 | .dfl_cftypes = blkcg_files, |
1469 | .legacy_cftypes = blkcg_legacy_files, |
1470 | .legacy_name = "blkio" , |
1471 | .exit = blkcg_exit, |
1472 | #ifdef CONFIG_MEMCG |
1473 | /* |
1474 | * This ensures that, if available, memcg is automatically enabled |
1475 | * together on the default hierarchy so that the owner cgroup can |
1476 | * be retrieved from writeback pages. |
1477 | */ |
1478 | .depends_on = 1 << memory_cgrp_id, |
1479 | #endif |
1480 | }; |
1481 | EXPORT_SYMBOL_GPL(io_cgrp_subsys); |
1482 | |
1483 | /** |
1484 | * blkcg_activate_policy - activate a blkcg policy on a gendisk |
1485 | * @disk: gendisk of interest |
1486 | * @pol: blkcg policy to activate |
1487 | * |
1488 | * Activate @pol on @disk. Requires %GFP_KERNEL context. @disk goes through |
1489 | * bypass mode to populate its blkgs with policy_data for @pol. |
1490 | * |
1491 | * Activation happens with @disk bypassed, so nobody would be accessing blkgs |
1492 | * from IO path. Update of each blkg is protected by both queue and blkcg |
1493 | * locks so that holding either lock and testing blkcg_policy_enabled() is |
1494 | * always enough for dereferencing policy data. |
1495 | * |
1496 | * The caller is responsible for synchronizing [de]activations and policy |
1497 | * [un]registerations. Returns 0 on success, -errno on failure. |
1498 | */ |
1499 | int blkcg_activate_policy(struct gendisk *disk, const struct blkcg_policy *pol) |
1500 | { |
1501 | struct request_queue *q = disk->queue; |
1502 | struct blkg_policy_data *pd_prealloc = NULL; |
1503 | struct blkcg_gq *blkg, *pinned_blkg = NULL; |
1504 | int ret; |
1505 | |
1506 | if (blkcg_policy_enabled(q, pol)) |
1507 | return 0; |
1508 | |
1509 | if (queue_is_mq(q)) |
1510 | blk_mq_freeze_queue(q); |
1511 | retry: |
1512 | spin_lock_irq(lock: &q->queue_lock); |
1513 | |
1514 | /* blkg_list is pushed at the head, reverse walk to initialize parents first */ |
1515 | list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) { |
1516 | struct blkg_policy_data *pd; |
1517 | |
1518 | if (blkg->pd[pol->plid]) |
1519 | continue; |
1520 | |
1521 | /* If prealloc matches, use it; otherwise try GFP_NOWAIT */ |
1522 | if (blkg == pinned_blkg) { |
1523 | pd = pd_prealloc; |
1524 | pd_prealloc = NULL; |
1525 | } else { |
1526 | pd = pol->pd_alloc_fn(disk, blkg->blkcg, |
1527 | GFP_NOWAIT | __GFP_NOWARN); |
1528 | } |
1529 | |
1530 | if (!pd) { |
1531 | /* |
1532 | * GFP_NOWAIT failed. Free the existing one and |
1533 | * prealloc for @blkg w/ GFP_KERNEL. |
1534 | */ |
1535 | if (pinned_blkg) |
1536 | blkg_put(blkg: pinned_blkg); |
1537 | blkg_get(blkg); |
1538 | pinned_blkg = blkg; |
1539 | |
1540 | spin_unlock_irq(lock: &q->queue_lock); |
1541 | |
1542 | if (pd_prealloc) |
1543 | pol->pd_free_fn(pd_prealloc); |
1544 | pd_prealloc = pol->pd_alloc_fn(disk, blkg->blkcg, |
1545 | GFP_KERNEL); |
1546 | if (pd_prealloc) |
1547 | goto retry; |
1548 | else |
1549 | goto enomem; |
1550 | } |
1551 | |
1552 | spin_lock(lock: &blkg->blkcg->lock); |
1553 | |
1554 | pd->blkg = blkg; |
1555 | pd->plid = pol->plid; |
1556 | blkg->pd[pol->plid] = pd; |
1557 | |
1558 | if (pol->pd_init_fn) |
1559 | pol->pd_init_fn(pd); |
1560 | |
1561 | if (pol->pd_online_fn) |
1562 | pol->pd_online_fn(pd); |
1563 | pd->online = true; |
1564 | |
1565 | spin_unlock(lock: &blkg->blkcg->lock); |
1566 | } |
1567 | |
1568 | __set_bit(pol->plid, q->blkcg_pols); |
1569 | ret = 0; |
1570 | |
1571 | spin_unlock_irq(lock: &q->queue_lock); |
1572 | out: |
1573 | if (queue_is_mq(q)) |
1574 | blk_mq_unfreeze_queue(q); |
1575 | if (pinned_blkg) |
1576 | blkg_put(blkg: pinned_blkg); |
1577 | if (pd_prealloc) |
1578 | pol->pd_free_fn(pd_prealloc); |
1579 | return ret; |
1580 | |
1581 | enomem: |
1582 | /* alloc failed, take down everything */ |
1583 | spin_lock_irq(lock: &q->queue_lock); |
1584 | list_for_each_entry(blkg, &q->blkg_list, q_node) { |
1585 | struct blkcg *blkcg = blkg->blkcg; |
1586 | struct blkg_policy_data *pd; |
1587 | |
1588 | spin_lock(lock: &blkcg->lock); |
1589 | pd = blkg->pd[pol->plid]; |
1590 | if (pd) { |
1591 | if (pd->online && pol->pd_offline_fn) |
1592 | pol->pd_offline_fn(pd); |
1593 | pd->online = false; |
1594 | pol->pd_free_fn(pd); |
1595 | blkg->pd[pol->plid] = NULL; |
1596 | } |
1597 | spin_unlock(lock: &blkcg->lock); |
1598 | } |
1599 | spin_unlock_irq(lock: &q->queue_lock); |
1600 | ret = -ENOMEM; |
1601 | goto out; |
1602 | } |
1603 | EXPORT_SYMBOL_GPL(blkcg_activate_policy); |
1604 | |
1605 | /** |
1606 | * blkcg_deactivate_policy - deactivate a blkcg policy on a gendisk |
1607 | * @disk: gendisk of interest |
1608 | * @pol: blkcg policy to deactivate |
1609 | * |
1610 | * Deactivate @pol on @disk. Follows the same synchronization rules as |
1611 | * blkcg_activate_policy(). |
1612 | */ |
1613 | void blkcg_deactivate_policy(struct gendisk *disk, |
1614 | const struct blkcg_policy *pol) |
1615 | { |
1616 | struct request_queue *q = disk->queue; |
1617 | struct blkcg_gq *blkg; |
1618 | |
1619 | if (!blkcg_policy_enabled(q, pol)) |
1620 | return; |
1621 | |
1622 | if (queue_is_mq(q)) |
1623 | blk_mq_freeze_queue(q); |
1624 | |
1625 | mutex_lock(&q->blkcg_mutex); |
1626 | spin_lock_irq(lock: &q->queue_lock); |
1627 | |
1628 | __clear_bit(pol->plid, q->blkcg_pols); |
1629 | |
1630 | list_for_each_entry(blkg, &q->blkg_list, q_node) { |
1631 | struct blkcg *blkcg = blkg->blkcg; |
1632 | |
1633 | spin_lock(lock: &blkcg->lock); |
1634 | if (blkg->pd[pol->plid]) { |
1635 | if (blkg->pd[pol->plid]->online && pol->pd_offline_fn) |
1636 | pol->pd_offline_fn(blkg->pd[pol->plid]); |
1637 | pol->pd_free_fn(blkg->pd[pol->plid]); |
1638 | blkg->pd[pol->plid] = NULL; |
1639 | } |
1640 | spin_unlock(lock: &blkcg->lock); |
1641 | } |
1642 | |
1643 | spin_unlock_irq(lock: &q->queue_lock); |
1644 | mutex_unlock(lock: &q->blkcg_mutex); |
1645 | |
1646 | if (queue_is_mq(q)) |
1647 | blk_mq_unfreeze_queue(q); |
1648 | } |
1649 | EXPORT_SYMBOL_GPL(blkcg_deactivate_policy); |
1650 | |
1651 | static void blkcg_free_all_cpd(struct blkcg_policy *pol) |
1652 | { |
1653 | struct blkcg *blkcg; |
1654 | |
1655 | list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) { |
1656 | if (blkcg->cpd[pol->plid]) { |
1657 | pol->cpd_free_fn(blkcg->cpd[pol->plid]); |
1658 | blkcg->cpd[pol->plid] = NULL; |
1659 | } |
1660 | } |
1661 | } |
1662 | |
1663 | /** |
1664 | * blkcg_policy_register - register a blkcg policy |
1665 | * @pol: blkcg policy to register |
1666 | * |
1667 | * Register @pol with blkcg core. Might sleep and @pol may be modified on |
1668 | * successful registration. Returns 0 on success and -errno on failure. |
1669 | */ |
1670 | int blkcg_policy_register(struct blkcg_policy *pol) |
1671 | { |
1672 | struct blkcg *blkcg; |
1673 | int i, ret; |
1674 | |
1675 | mutex_lock(&blkcg_pol_register_mutex); |
1676 | mutex_lock(&blkcg_pol_mutex); |
1677 | |
1678 | /* find an empty slot */ |
1679 | ret = -ENOSPC; |
1680 | for (i = 0; i < BLKCG_MAX_POLS; i++) |
1681 | if (!blkcg_policy[i]) |
1682 | break; |
1683 | if (i >= BLKCG_MAX_POLS) { |
1684 | pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n" ); |
1685 | goto err_unlock; |
1686 | } |
1687 | |
1688 | /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */ |
1689 | if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) || |
1690 | (!pol->pd_alloc_fn ^ !pol->pd_free_fn)) |
1691 | goto err_unlock; |
1692 | |
1693 | /* register @pol */ |
1694 | pol->plid = i; |
1695 | blkcg_policy[pol->plid] = pol; |
1696 | |
1697 | /* allocate and install cpd's */ |
1698 | if (pol->cpd_alloc_fn) { |
1699 | list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) { |
1700 | struct blkcg_policy_data *cpd; |
1701 | |
1702 | cpd = pol->cpd_alloc_fn(GFP_KERNEL); |
1703 | if (!cpd) |
1704 | goto err_free_cpds; |
1705 | |
1706 | blkcg->cpd[pol->plid] = cpd; |
1707 | cpd->blkcg = blkcg; |
1708 | cpd->plid = pol->plid; |
1709 | } |
1710 | } |
1711 | |
1712 | mutex_unlock(lock: &blkcg_pol_mutex); |
1713 | |
1714 | /* everything is in place, add intf files for the new policy */ |
1715 | if (pol->dfl_cftypes) |
1716 | WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys, |
1717 | pol->dfl_cftypes)); |
1718 | if (pol->legacy_cftypes) |
1719 | WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys, |
1720 | pol->legacy_cftypes)); |
1721 | mutex_unlock(lock: &blkcg_pol_register_mutex); |
1722 | return 0; |
1723 | |
1724 | err_free_cpds: |
1725 | if (pol->cpd_free_fn) |
1726 | blkcg_free_all_cpd(pol); |
1727 | |
1728 | blkcg_policy[pol->plid] = NULL; |
1729 | err_unlock: |
1730 | mutex_unlock(lock: &blkcg_pol_mutex); |
1731 | mutex_unlock(lock: &blkcg_pol_register_mutex); |
1732 | return ret; |
1733 | } |
1734 | EXPORT_SYMBOL_GPL(blkcg_policy_register); |
1735 | |
1736 | /** |
1737 | * blkcg_policy_unregister - unregister a blkcg policy |
1738 | * @pol: blkcg policy to unregister |
1739 | * |
1740 | * Undo blkcg_policy_register(@pol). Might sleep. |
1741 | */ |
1742 | void blkcg_policy_unregister(struct blkcg_policy *pol) |
1743 | { |
1744 | mutex_lock(&blkcg_pol_register_mutex); |
1745 | |
1746 | if (WARN_ON(blkcg_policy[pol->plid] != pol)) |
1747 | goto out_unlock; |
1748 | |
1749 | /* kill the intf files first */ |
1750 | if (pol->dfl_cftypes) |
1751 | cgroup_rm_cftypes(cfts: pol->dfl_cftypes); |
1752 | if (pol->legacy_cftypes) |
1753 | cgroup_rm_cftypes(cfts: pol->legacy_cftypes); |
1754 | |
1755 | /* remove cpds and unregister */ |
1756 | mutex_lock(&blkcg_pol_mutex); |
1757 | |
1758 | if (pol->cpd_free_fn) |
1759 | blkcg_free_all_cpd(pol); |
1760 | |
1761 | blkcg_policy[pol->plid] = NULL; |
1762 | |
1763 | mutex_unlock(lock: &blkcg_pol_mutex); |
1764 | out_unlock: |
1765 | mutex_unlock(lock: &blkcg_pol_register_mutex); |
1766 | } |
1767 | EXPORT_SYMBOL_GPL(blkcg_policy_unregister); |
1768 | |
1769 | /* |
1770 | * Scale the accumulated delay based on how long it has been since we updated |
1771 | * the delay. We only call this when we are adding delay, in case it's been a |
1772 | * while since we added delay, and when we are checking to see if we need to |
1773 | * delay a task, to account for any delays that may have occurred. |
1774 | */ |
1775 | static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now) |
1776 | { |
1777 | u64 old = atomic64_read(v: &blkg->delay_start); |
1778 | |
1779 | /* negative use_delay means no scaling, see blkcg_set_delay() */ |
1780 | if (atomic_read(v: &blkg->use_delay) < 0) |
1781 | return; |
1782 | |
1783 | /* |
1784 | * We only want to scale down every second. The idea here is that we |
1785 | * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain |
1786 | * time window. We only want to throttle tasks for recent delay that |
1787 | * has occurred, in 1 second time windows since that's the maximum |
1788 | * things can be throttled. We save the current delay window in |
1789 | * blkg->last_delay so we know what amount is still left to be charged |
1790 | * to the blkg from this point onward. blkg->last_use keeps track of |
1791 | * the use_delay counter. The idea is if we're unthrottling the blkg we |
1792 | * are ok with whatever is happening now, and we can take away more of |
1793 | * the accumulated delay as we've already throttled enough that |
1794 | * everybody is happy with their IO latencies. |
1795 | */ |
1796 | if (time_before64(old + NSEC_PER_SEC, now) && |
1797 | atomic64_try_cmpxchg(v: &blkg->delay_start, old: &old, new: now)) { |
1798 | u64 cur = atomic64_read(v: &blkg->delay_nsec); |
1799 | u64 sub = min_t(u64, blkg->last_delay, now - old); |
1800 | int cur_use = atomic_read(v: &blkg->use_delay); |
1801 | |
1802 | /* |
1803 | * We've been unthrottled, subtract a larger chunk of our |
1804 | * accumulated delay. |
1805 | */ |
1806 | if (cur_use < blkg->last_use) |
1807 | sub = max_t(u64, sub, blkg->last_delay >> 1); |
1808 | |
1809 | /* |
1810 | * This shouldn't happen, but handle it anyway. Our delay_nsec |
1811 | * should only ever be growing except here where we subtract out |
1812 | * min(last_delay, 1 second), but lord knows bugs happen and I'd |
1813 | * rather not end up with negative numbers. |
1814 | */ |
1815 | if (unlikely(cur < sub)) { |
1816 | atomic64_set(v: &blkg->delay_nsec, i: 0); |
1817 | blkg->last_delay = 0; |
1818 | } else { |
1819 | atomic64_sub(i: sub, v: &blkg->delay_nsec); |
1820 | blkg->last_delay = cur - sub; |
1821 | } |
1822 | blkg->last_use = cur_use; |
1823 | } |
1824 | } |
1825 | |
1826 | /* |
1827 | * This is called when we want to actually walk up the hierarchy and check to |
1828 | * see if we need to throttle, and then actually throttle if there is some |
1829 | * accumulated delay. This should only be called upon return to user space so |
1830 | * we're not holding some lock that would induce a priority inversion. |
1831 | */ |
1832 | static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay) |
1833 | { |
1834 | unsigned long pflags; |
1835 | bool clamp; |
1836 | u64 now = ktime_to_ns(kt: ktime_get()); |
1837 | u64 exp; |
1838 | u64 delay_nsec = 0; |
1839 | int tok; |
1840 | |
1841 | while (blkg->parent) { |
1842 | int use_delay = atomic_read(v: &blkg->use_delay); |
1843 | |
1844 | if (use_delay) { |
1845 | u64 this_delay; |
1846 | |
1847 | blkcg_scale_delay(blkg, now); |
1848 | this_delay = atomic64_read(v: &blkg->delay_nsec); |
1849 | if (this_delay > delay_nsec) { |
1850 | delay_nsec = this_delay; |
1851 | clamp = use_delay > 0; |
1852 | } |
1853 | } |
1854 | blkg = blkg->parent; |
1855 | } |
1856 | |
1857 | if (!delay_nsec) |
1858 | return; |
1859 | |
1860 | /* |
1861 | * Let's not sleep for all eternity if we've amassed a huge delay. |
1862 | * Swapping or metadata IO can accumulate 10's of seconds worth of |
1863 | * delay, and we want userspace to be able to do _something_ so cap the |
1864 | * delays at 0.25s. If there's 10's of seconds worth of delay then the |
1865 | * tasks will be delayed for 0.25 second for every syscall. If |
1866 | * blkcg_set_delay() was used as indicated by negative use_delay, the |
1867 | * caller is responsible for regulating the range. |
1868 | */ |
1869 | if (clamp) |
1870 | delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC); |
1871 | |
1872 | if (use_memdelay) |
1873 | psi_memstall_enter(flags: &pflags); |
1874 | |
1875 | exp = ktime_add_ns(now, delay_nsec); |
1876 | tok = io_schedule_prepare(); |
1877 | do { |
1878 | __set_current_state(TASK_KILLABLE); |
1879 | if (!schedule_hrtimeout(expires: &exp, mode: HRTIMER_MODE_ABS)) |
1880 | break; |
1881 | } while (!fatal_signal_pending(current)); |
1882 | io_schedule_finish(token: tok); |
1883 | |
1884 | if (use_memdelay) |
1885 | psi_memstall_leave(flags: &pflags); |
1886 | } |
1887 | |
1888 | /** |
1889 | * blkcg_maybe_throttle_current - throttle the current task if it has been marked |
1890 | * |
1891 | * This is only called if we've been marked with set_notify_resume(). Obviously |
1892 | * we can be set_notify_resume() for reasons other than blkcg throttling, so we |
1893 | * check to see if current->throttle_disk is set and if not this doesn't do |
1894 | * anything. This should only ever be called by the resume code, it's not meant |
1895 | * to be called by people willy-nilly as it will actually do the work to |
1896 | * throttle the task if it is setup for throttling. |
1897 | */ |
1898 | void blkcg_maybe_throttle_current(void) |
1899 | { |
1900 | struct gendisk *disk = current->throttle_disk; |
1901 | struct blkcg *blkcg; |
1902 | struct blkcg_gq *blkg; |
1903 | bool use_memdelay = current->use_memdelay; |
1904 | |
1905 | if (!disk) |
1906 | return; |
1907 | |
1908 | current->throttle_disk = NULL; |
1909 | current->use_memdelay = false; |
1910 | |
1911 | rcu_read_lock(); |
1912 | blkcg = css_to_blkcg(css: blkcg_css()); |
1913 | if (!blkcg) |
1914 | goto out; |
1915 | blkg = blkg_lookup(blkcg, q: disk->queue); |
1916 | if (!blkg) |
1917 | goto out; |
1918 | if (!blkg_tryget(blkg)) |
1919 | goto out; |
1920 | rcu_read_unlock(); |
1921 | |
1922 | blkcg_maybe_throttle_blkg(blkg, use_memdelay); |
1923 | blkg_put(blkg); |
1924 | put_disk(disk); |
1925 | return; |
1926 | out: |
1927 | rcu_read_unlock(); |
1928 | } |
1929 | |
1930 | /** |
1931 | * blkcg_schedule_throttle - this task needs to check for throttling |
1932 | * @disk: disk to throttle |
1933 | * @use_memdelay: do we charge this to memory delay for PSI |
1934 | * |
1935 | * This is called by the IO controller when we know there's delay accumulated |
1936 | * for the blkg for this task. We do not pass the blkg because there are places |
1937 | * we call this that may not have that information, the swapping code for |
1938 | * instance will only have a block_device at that point. This set's the |
1939 | * notify_resume for the task to check and see if it requires throttling before |
1940 | * returning to user space. |
1941 | * |
1942 | * We will only schedule once per syscall. You can call this over and over |
1943 | * again and it will only do the check once upon return to user space, and only |
1944 | * throttle once. If the task needs to be throttled again it'll need to be |
1945 | * re-set at the next time we see the task. |
1946 | */ |
1947 | void blkcg_schedule_throttle(struct gendisk *disk, bool use_memdelay) |
1948 | { |
1949 | if (unlikely(current->flags & PF_KTHREAD)) |
1950 | return; |
1951 | |
1952 | if (current->throttle_disk != disk) { |
1953 | if (test_bit(GD_DEAD, &disk->state)) |
1954 | return; |
1955 | get_device(disk_to_dev(disk)); |
1956 | |
1957 | if (current->throttle_disk) |
1958 | put_disk(current->throttle_disk); |
1959 | current->throttle_disk = disk; |
1960 | } |
1961 | |
1962 | if (use_memdelay) |
1963 | current->use_memdelay = use_memdelay; |
1964 | set_notify_resume(current); |
1965 | } |
1966 | |
1967 | /** |
1968 | * blkcg_add_delay - add delay to this blkg |
1969 | * @blkg: blkg of interest |
1970 | * @now: the current time in nanoseconds |
1971 | * @delta: how many nanoseconds of delay to add |
1972 | * |
1973 | * Charge @delta to the blkg's current delay accumulation. This is used to |
1974 | * throttle tasks if an IO controller thinks we need more throttling. |
1975 | */ |
1976 | void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta) |
1977 | { |
1978 | if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0)) |
1979 | return; |
1980 | blkcg_scale_delay(blkg, now); |
1981 | atomic64_add(i: delta, v: &blkg->delay_nsec); |
1982 | } |
1983 | |
1984 | /** |
1985 | * blkg_tryget_closest - try and get a blkg ref on the closet blkg |
1986 | * @bio: target bio |
1987 | * @css: target css |
1988 | * |
1989 | * As the failure mode here is to walk up the blkg tree, this ensure that the |
1990 | * blkg->parent pointers are always valid. This returns the blkg that it ended |
1991 | * up taking a reference on or %NULL if no reference was taken. |
1992 | */ |
1993 | static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio, |
1994 | struct cgroup_subsys_state *css) |
1995 | { |
1996 | struct blkcg_gq *blkg, *ret_blkg = NULL; |
1997 | |
1998 | rcu_read_lock(); |
1999 | blkg = blkg_lookup_create(blkcg: css_to_blkcg(css), disk: bio->bi_bdev->bd_disk); |
2000 | while (blkg) { |
2001 | if (blkg_tryget(blkg)) { |
2002 | ret_blkg = blkg; |
2003 | break; |
2004 | } |
2005 | blkg = blkg->parent; |
2006 | } |
2007 | rcu_read_unlock(); |
2008 | |
2009 | return ret_blkg; |
2010 | } |
2011 | |
2012 | /** |
2013 | * bio_associate_blkg_from_css - associate a bio with a specified css |
2014 | * @bio: target bio |
2015 | * @css: target css |
2016 | * |
2017 | * Associate @bio with the blkg found by combining the css's blkg and the |
2018 | * request_queue of the @bio. An association failure is handled by walking up |
2019 | * the blkg tree. Therefore, the blkg associated can be anything between @blkg |
2020 | * and q->root_blkg. This situation only happens when a cgroup is dying and |
2021 | * then the remaining bios will spill to the closest alive blkg. |
2022 | * |
2023 | * A reference will be taken on the blkg and will be released when @bio is |
2024 | * freed. |
2025 | */ |
2026 | void bio_associate_blkg_from_css(struct bio *bio, |
2027 | struct cgroup_subsys_state *css) |
2028 | { |
2029 | if (bio->bi_blkg) |
2030 | blkg_put(blkg: bio->bi_blkg); |
2031 | |
2032 | if (css && css->parent) { |
2033 | bio->bi_blkg = blkg_tryget_closest(bio, css); |
2034 | } else { |
2035 | blkg_get(blkg: bdev_get_queue(bdev: bio->bi_bdev)->root_blkg); |
2036 | bio->bi_blkg = bdev_get_queue(bdev: bio->bi_bdev)->root_blkg; |
2037 | } |
2038 | } |
2039 | EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css); |
2040 | |
2041 | /** |
2042 | * bio_associate_blkg - associate a bio with a blkg |
2043 | * @bio: target bio |
2044 | * |
2045 | * Associate @bio with the blkg found from the bio's css and request_queue. |
2046 | * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is |
2047 | * already associated, the css is reused and association redone as the |
2048 | * request_queue may have changed. |
2049 | */ |
2050 | void bio_associate_blkg(struct bio *bio) |
2051 | { |
2052 | struct cgroup_subsys_state *css; |
2053 | |
2054 | rcu_read_lock(); |
2055 | |
2056 | if (bio->bi_blkg) |
2057 | css = bio_blkcg_css(bio); |
2058 | else |
2059 | css = blkcg_css(); |
2060 | |
2061 | bio_associate_blkg_from_css(bio, css); |
2062 | |
2063 | rcu_read_unlock(); |
2064 | } |
2065 | EXPORT_SYMBOL_GPL(bio_associate_blkg); |
2066 | |
2067 | /** |
2068 | * bio_clone_blkg_association - clone blkg association from src to dst bio |
2069 | * @dst: destination bio |
2070 | * @src: source bio |
2071 | */ |
2072 | void bio_clone_blkg_association(struct bio *dst, struct bio *src) |
2073 | { |
2074 | if (src->bi_blkg) |
2075 | bio_associate_blkg_from_css(dst, bio_blkcg_css(src)); |
2076 | } |
2077 | EXPORT_SYMBOL_GPL(bio_clone_blkg_association); |
2078 | |
2079 | static int blk_cgroup_io_type(struct bio *bio) |
2080 | { |
2081 | if (op_is_discard(op: bio->bi_opf)) |
2082 | return BLKG_IOSTAT_DISCARD; |
2083 | if (op_is_write(op: bio->bi_opf)) |
2084 | return BLKG_IOSTAT_WRITE; |
2085 | return BLKG_IOSTAT_READ; |
2086 | } |
2087 | |
2088 | void blk_cgroup_bio_start(struct bio *bio) |
2089 | { |
2090 | struct blkcg *blkcg = bio->bi_blkg->blkcg; |
2091 | int rwd = blk_cgroup_io_type(bio), cpu; |
2092 | struct blkg_iostat_set *bis; |
2093 | unsigned long flags; |
2094 | |
2095 | if (!cgroup_subsys_on_dfl(io_cgrp_subsys)) |
2096 | return; |
2097 | |
2098 | /* Root-level stats are sourced from system-wide IO stats */ |
2099 | if (!cgroup_parent(cgrp: blkcg->css.cgroup)) |
2100 | return; |
2101 | |
2102 | cpu = get_cpu(); |
2103 | bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu); |
2104 | flags = u64_stats_update_begin_irqsave(syncp: &bis->sync); |
2105 | |
2106 | /* |
2107 | * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split |
2108 | * bio and we would have already accounted for the size of the bio. |
2109 | */ |
2110 | if (!bio_flagged(bio, bit: BIO_CGROUP_ACCT)) { |
2111 | bio_set_flag(bio, bit: BIO_CGROUP_ACCT); |
2112 | bis->cur.bytes[rwd] += bio->bi_iter.bi_size; |
2113 | } |
2114 | bis->cur.ios[rwd]++; |
2115 | |
2116 | /* |
2117 | * If the iostat_cpu isn't in a lockless list, put it into the |
2118 | * list to indicate that a stat update is pending. |
2119 | */ |
2120 | if (!READ_ONCE(bis->lqueued)) { |
2121 | struct llist_head *lhead = this_cpu_ptr(blkcg->lhead); |
2122 | |
2123 | llist_add(new: &bis->lnode, head: lhead); |
2124 | WRITE_ONCE(bis->lqueued, true); |
2125 | } |
2126 | |
2127 | u64_stats_update_end_irqrestore(syncp: &bis->sync, flags); |
2128 | cgroup_rstat_updated(cgrp: blkcg->css.cgroup, cpu); |
2129 | put_cpu(); |
2130 | } |
2131 | |
2132 | bool blk_cgroup_congested(void) |
2133 | { |
2134 | struct cgroup_subsys_state *css; |
2135 | bool ret = false; |
2136 | |
2137 | rcu_read_lock(); |
2138 | for (css = blkcg_css(); css; css = css->parent) { |
2139 | if (atomic_read(v: &css->cgroup->congestion_count)) { |
2140 | ret = true; |
2141 | break; |
2142 | } |
2143 | } |
2144 | rcu_read_unlock(); |
2145 | return ret; |
2146 | } |
2147 | |
2148 | module_param(blkcg_debug_stats, bool, 0644); |
2149 | MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not" ); |
2150 | |