]>
Commit | Line | Data |
---|---|---|
1 | // SPDX-License-Identifier: GPL-2.0 | |
2 | /* | |
3 | * Common Block IO controller cgroup interface | |
4 | * | |
5 | * Based on ideas and code from CFQ, CFS and BFQ: | |
6 | * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk> | |
7 | * | |
8 | * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it> | |
9 | * Paolo Valente <paolo.valente@unimore.it> | |
10 | * | |
11 | * Copyright (C) 2009 Vivek Goyal <vgoyal@redhat.com> | |
12 | * Nauman Rafique <nauman@google.com> | |
13 | * | |
14 | * For policy-specific per-blkcg data: | |
15 | * Copyright (C) 2015 Paolo Valente <paolo.valente@unimore.it> | |
16 | * Arianna Avanzini <avanzini.arianna@gmail.com> | |
17 | */ | |
18 | #include <linux/ioprio.h> | |
19 | #include <linux/kdev_t.h> | |
20 | #include <linux/module.h> | |
21 | #include <linux/sched/signal.h> | |
22 | #include <linux/err.h> | |
23 | #include <linux/blkdev.h> | |
24 | #include <linux/backing-dev.h> | |
25 | #include <linux/slab.h> | |
26 | #include <linux/genhd.h> | |
27 | #include <linux/delay.h> | |
28 | #include <linux/atomic.h> | |
29 | #include <linux/ctype.h> | |
30 | #include <linux/blk-cgroup.h> | |
31 | #include <linux/tracehook.h> | |
32 | #include <linux/psi.h> | |
33 | #include "blk.h" | |
34 | #include "blk-ioprio.h" | |
35 | ||
36 | /* | |
37 | * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation. | |
38 | * blkcg_pol_register_mutex nests outside of it and synchronizes entire | |
39 | * policy [un]register operations including cgroup file additions / | |
40 | * removals. Putting cgroup file registration outside blkcg_pol_mutex | |
41 | * allows grabbing it from cgroup callbacks. | |
42 | */ | |
43 | static DEFINE_MUTEX(blkcg_pol_register_mutex); | |
44 | static DEFINE_MUTEX(blkcg_pol_mutex); | |
45 | ||
46 | struct blkcg blkcg_root; | |
47 | EXPORT_SYMBOL_GPL(blkcg_root); | |
48 | ||
49 | struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css; | |
50 | EXPORT_SYMBOL_GPL(blkcg_root_css); | |
51 | ||
52 | static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS]; | |
53 | ||
54 | static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */ | |
55 | ||
56 | bool blkcg_debug_stats = false; | |
57 | static struct workqueue_struct *blkcg_punt_bio_wq; | |
58 | ||
59 | #define BLKG_DESTROY_BATCH_SIZE 64 | |
60 | ||
61 | static bool blkcg_policy_enabled(struct request_queue *q, | |
62 | const struct blkcg_policy *pol) | |
63 | { | |
64 | return pol && test_bit(pol->plid, q->blkcg_pols); | |
65 | } | |
66 | ||
67 | /** | |
68 | * blkg_free - free a blkg | |
69 | * @blkg: blkg to free | |
70 | * | |
71 | * Free @blkg which may be partially allocated. | |
72 | */ | |
73 | static void blkg_free(struct blkcg_gq *blkg) | |
74 | { | |
75 | int i; | |
76 | ||
77 | if (!blkg) | |
78 | return; | |
79 | ||
80 | for (i = 0; i < BLKCG_MAX_POLS; i++) | |
81 | if (blkg->pd[i]) | |
82 | blkcg_policy[i]->pd_free_fn(blkg->pd[i]); | |
83 | ||
84 | free_percpu(blkg->iostat_cpu); | |
85 | percpu_ref_exit(&blkg->refcnt); | |
86 | kfree(blkg); | |
87 | } | |
88 | ||
89 | static void __blkg_release(struct rcu_head *rcu) | |
90 | { | |
91 | struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head); | |
92 | ||
93 | WARN_ON(!bio_list_empty(&blkg->async_bios)); | |
94 | ||
95 | /* release the blkcg and parent blkg refs this blkg has been holding */ | |
96 | css_put(&blkg->blkcg->css); | |
97 | if (blkg->parent) | |
98 | blkg_put(blkg->parent); | |
99 | blkg_free(blkg); | |
100 | } | |
101 | ||
102 | /* | |
103 | * A group is RCU protected, but having an rcu lock does not mean that one | |
104 | * can access all the fields of blkg and assume these are valid. For | |
105 | * example, don't try to follow throtl_data and request queue links. | |
106 | * | |
107 | * Having a reference to blkg under an rcu allows accesses to only values | |
108 | * local to groups like group stats and group rate limits. | |
109 | */ | |
110 | static void blkg_release(struct percpu_ref *ref) | |
111 | { | |
112 | struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt); | |
113 | ||
114 | call_rcu(&blkg->rcu_head, __blkg_release); | |
115 | } | |
116 | ||
117 | static void blkg_async_bio_workfn(struct work_struct *work) | |
118 | { | |
119 | struct blkcg_gq *blkg = container_of(work, struct blkcg_gq, | |
120 | async_bio_work); | |
121 | struct bio_list bios = BIO_EMPTY_LIST; | |
122 | struct bio *bio; | |
123 | struct blk_plug plug; | |
124 | bool need_plug = false; | |
125 | ||
126 | /* as long as there are pending bios, @blkg can't go away */ | |
127 | spin_lock_bh(&blkg->async_bio_lock); | |
128 | bio_list_merge(&bios, &blkg->async_bios); | |
129 | bio_list_init(&blkg->async_bios); | |
130 | spin_unlock_bh(&blkg->async_bio_lock); | |
131 | ||
132 | /* start plug only when bio_list contains at least 2 bios */ | |
133 | if (bios.head && bios.head->bi_next) { | |
134 | need_plug = true; | |
135 | blk_start_plug(&plug); | |
136 | } | |
137 | while ((bio = bio_list_pop(&bios))) | |
138 | submit_bio(bio); | |
139 | if (need_plug) | |
140 | blk_finish_plug(&plug); | |
141 | } | |
142 | ||
143 | /** | |
144 | * blkg_alloc - allocate a blkg | |
145 | * @blkcg: block cgroup the new blkg is associated with | |
146 | * @q: request_queue the new blkg is associated with | |
147 | * @gfp_mask: allocation mask to use | |
148 | * | |
149 | * Allocate a new blkg assocating @blkcg and @q. | |
150 | */ | |
151 | static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q, | |
152 | gfp_t gfp_mask) | |
153 | { | |
154 | struct blkcg_gq *blkg; | |
155 | int i, cpu; | |
156 | ||
157 | /* alloc and init base part */ | |
158 | blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node); | |
159 | if (!blkg) | |
160 | return NULL; | |
161 | ||
162 | if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask)) | |
163 | goto err_free; | |
164 | ||
165 | blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask); | |
166 | if (!blkg->iostat_cpu) | |
167 | goto err_free; | |
168 | ||
169 | blkg->q = q; | |
170 | INIT_LIST_HEAD(&blkg->q_node); | |
171 | spin_lock_init(&blkg->async_bio_lock); | |
172 | bio_list_init(&blkg->async_bios); | |
173 | INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn); | |
174 | blkg->blkcg = blkcg; | |
175 | ||
176 | u64_stats_init(&blkg->iostat.sync); | |
177 | for_each_possible_cpu(cpu) | |
178 | u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync); | |
179 | ||
180 | for (i = 0; i < BLKCG_MAX_POLS; i++) { | |
181 | struct blkcg_policy *pol = blkcg_policy[i]; | |
182 | struct blkg_policy_data *pd; | |
183 | ||
184 | if (!blkcg_policy_enabled(q, pol)) | |
185 | continue; | |
186 | ||
187 | /* alloc per-policy data and attach it to blkg */ | |
188 | pd = pol->pd_alloc_fn(gfp_mask, q, blkcg); | |
189 | if (!pd) | |
190 | goto err_free; | |
191 | ||
192 | blkg->pd[i] = pd; | |
193 | pd->blkg = blkg; | |
194 | pd->plid = i; | |
195 | } | |
196 | ||
197 | return blkg; | |
198 | ||
199 | err_free: | |
200 | blkg_free(blkg); | |
201 | return NULL; | |
202 | } | |
203 | ||
204 | struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg, | |
205 | struct request_queue *q, bool update_hint) | |
206 | { | |
207 | struct blkcg_gq *blkg; | |
208 | ||
209 | /* | |
210 | * Hint didn't match. Look up from the radix tree. Note that the | |
211 | * hint can only be updated under queue_lock as otherwise @blkg | |
212 | * could have already been removed from blkg_tree. The caller is | |
213 | * responsible for grabbing queue_lock if @update_hint. | |
214 | */ | |
215 | blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id); | |
216 | if (blkg && blkg->q == q) { | |
217 | if (update_hint) { | |
218 | lockdep_assert_held(&q->queue_lock); | |
219 | rcu_assign_pointer(blkcg->blkg_hint, blkg); | |
220 | } | |
221 | return blkg; | |
222 | } | |
223 | ||
224 | return NULL; | |
225 | } | |
226 | EXPORT_SYMBOL_GPL(blkg_lookup_slowpath); | |
227 | ||
228 | /* | |
229 | * If @new_blkg is %NULL, this function tries to allocate a new one as | |
230 | * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return. | |
231 | */ | |
232 | static struct blkcg_gq *blkg_create(struct blkcg *blkcg, | |
233 | struct request_queue *q, | |
234 | struct blkcg_gq *new_blkg) | |
235 | { | |
236 | struct blkcg_gq *blkg; | |
237 | int i, ret; | |
238 | ||
239 | WARN_ON_ONCE(!rcu_read_lock_held()); | |
240 | lockdep_assert_held(&q->queue_lock); | |
241 | ||
242 | /* request_queue is dying, do not create/recreate a blkg */ | |
243 | if (blk_queue_dying(q)) { | |
244 | ret = -ENODEV; | |
245 | goto err_free_blkg; | |
246 | } | |
247 | ||
248 | /* blkg holds a reference to blkcg */ | |
249 | if (!css_tryget_online(&blkcg->css)) { | |
250 | ret = -ENODEV; | |
251 | goto err_free_blkg; | |
252 | } | |
253 | ||
254 | /* allocate */ | |
255 | if (!new_blkg) { | |
256 | new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN); | |
257 | if (unlikely(!new_blkg)) { | |
258 | ret = -ENOMEM; | |
259 | goto err_put_css; | |
260 | } | |
261 | } | |
262 | blkg = new_blkg; | |
263 | ||
264 | /* link parent */ | |
265 | if (blkcg_parent(blkcg)) { | |
266 | blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false); | |
267 | if (WARN_ON_ONCE(!blkg->parent)) { | |
268 | ret = -ENODEV; | |
269 | goto err_put_css; | |
270 | } | |
271 | blkg_get(blkg->parent); | |
272 | } | |
273 | ||
274 | /* invoke per-policy init */ | |
275 | for (i = 0; i < BLKCG_MAX_POLS; i++) { | |
276 | struct blkcg_policy *pol = blkcg_policy[i]; | |
277 | ||
278 | if (blkg->pd[i] && pol->pd_init_fn) | |
279 | pol->pd_init_fn(blkg->pd[i]); | |
280 | } | |
281 | ||
282 | /* insert */ | |
283 | spin_lock(&blkcg->lock); | |
284 | ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg); | |
285 | if (likely(!ret)) { | |
286 | hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list); | |
287 | list_add(&blkg->q_node, &q->blkg_list); | |
288 | ||
289 | for (i = 0; i < BLKCG_MAX_POLS; i++) { | |
290 | struct blkcg_policy *pol = blkcg_policy[i]; | |
291 | ||
292 | if (blkg->pd[i] && pol->pd_online_fn) | |
293 | pol->pd_online_fn(blkg->pd[i]); | |
294 | } | |
295 | } | |
296 | blkg->online = true; | |
297 | spin_unlock(&blkcg->lock); | |
298 | ||
299 | if (!ret) | |
300 | return blkg; | |
301 | ||
302 | /* @blkg failed fully initialized, use the usual release path */ | |
303 | blkg_put(blkg); | |
304 | return ERR_PTR(ret); | |
305 | ||
306 | err_put_css: | |
307 | css_put(&blkcg->css); | |
308 | err_free_blkg: | |
309 | blkg_free(new_blkg); | |
310 | return ERR_PTR(ret); | |
311 | } | |
312 | ||
313 | /** | |
314 | * blkg_lookup_create - lookup blkg, try to create one if not there | |
315 | * @blkcg: blkcg of interest | |
316 | * @q: request_queue of interest | |
317 | * | |
318 | * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to | |
319 | * create one. blkg creation is performed recursively from blkcg_root such | |
320 | * that all non-root blkg's have access to the parent blkg. This function | |
321 | * should be called under RCU read lock and takes @q->queue_lock. | |
322 | * | |
323 | * Returns the blkg or the closest blkg if blkg_create() fails as it walks | |
324 | * down from root. | |
325 | */ | |
326 | static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg, | |
327 | struct request_queue *q) | |
328 | { | |
329 | struct blkcg_gq *blkg; | |
330 | unsigned long flags; | |
331 | ||
332 | WARN_ON_ONCE(!rcu_read_lock_held()); | |
333 | ||
334 | blkg = blkg_lookup(blkcg, q); | |
335 | if (blkg) | |
336 | return blkg; | |
337 | ||
338 | spin_lock_irqsave(&q->queue_lock, flags); | |
339 | blkg = __blkg_lookup(blkcg, q, true); | |
340 | if (blkg) | |
341 | goto found; | |
342 | ||
343 | /* | |
344 | * Create blkgs walking down from blkcg_root to @blkcg, so that all | |
345 | * non-root blkgs have access to their parents. Returns the closest | |
346 | * blkg to the intended blkg should blkg_create() fail. | |
347 | */ | |
348 | while (true) { | |
349 | struct blkcg *pos = blkcg; | |
350 | struct blkcg *parent = blkcg_parent(blkcg); | |
351 | struct blkcg_gq *ret_blkg = q->root_blkg; | |
352 | ||
353 | while (parent) { | |
354 | blkg = __blkg_lookup(parent, q, false); | |
355 | if (blkg) { | |
356 | /* remember closest blkg */ | |
357 | ret_blkg = blkg; | |
358 | break; | |
359 | } | |
360 | pos = parent; | |
361 | parent = blkcg_parent(parent); | |
362 | } | |
363 | ||
364 | blkg = blkg_create(pos, q, NULL); | |
365 | if (IS_ERR(blkg)) { | |
366 | blkg = ret_blkg; | |
367 | break; | |
368 | } | |
369 | if (pos == blkcg) | |
370 | break; | |
371 | } | |
372 | ||
373 | found: | |
374 | spin_unlock_irqrestore(&q->queue_lock, flags); | |
375 | return blkg; | |
376 | } | |
377 | ||
378 | static void blkg_destroy(struct blkcg_gq *blkg) | |
379 | { | |
380 | struct blkcg *blkcg = blkg->blkcg; | |
381 | int i; | |
382 | ||
383 | lockdep_assert_held(&blkg->q->queue_lock); | |
384 | lockdep_assert_held(&blkcg->lock); | |
385 | ||
386 | /* Something wrong if we are trying to remove same group twice */ | |
387 | WARN_ON_ONCE(list_empty(&blkg->q_node)); | |
388 | WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node)); | |
389 | ||
390 | for (i = 0; i < BLKCG_MAX_POLS; i++) { | |
391 | struct blkcg_policy *pol = blkcg_policy[i]; | |
392 | ||
393 | if (blkg->pd[i] && pol->pd_offline_fn) | |
394 | pol->pd_offline_fn(blkg->pd[i]); | |
395 | } | |
396 | ||
397 | blkg->online = false; | |
398 | ||
399 | radix_tree_delete(&blkcg->blkg_tree, blkg->q->id); | |
400 | list_del_init(&blkg->q_node); | |
401 | hlist_del_init_rcu(&blkg->blkcg_node); | |
402 | ||
403 | /* | |
404 | * Both setting lookup hint to and clearing it from @blkg are done | |
405 | * under queue_lock. If it's not pointing to @blkg now, it never | |
406 | * will. Hint assignment itself can race safely. | |
407 | */ | |
408 | if (rcu_access_pointer(blkcg->blkg_hint) == blkg) | |
409 | rcu_assign_pointer(blkcg->blkg_hint, NULL); | |
410 | ||
411 | /* | |
412 | * Put the reference taken at the time of creation so that when all | |
413 | * queues are gone, group can be destroyed. | |
414 | */ | |
415 | percpu_ref_kill(&blkg->refcnt); | |
416 | } | |
417 | ||
418 | /** | |
419 | * blkg_destroy_all - destroy all blkgs associated with a request_queue | |
420 | * @q: request_queue of interest | |
421 | * | |
422 | * Destroy all blkgs associated with @q. | |
423 | */ | |
424 | static void blkg_destroy_all(struct request_queue *q) | |
425 | { | |
426 | struct blkcg_gq *blkg, *n; | |
427 | int count = BLKG_DESTROY_BATCH_SIZE; | |
428 | ||
429 | restart: | |
430 | spin_lock_irq(&q->queue_lock); | |
431 | list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) { | |
432 | struct blkcg *blkcg = blkg->blkcg; | |
433 | ||
434 | spin_lock(&blkcg->lock); | |
435 | blkg_destroy(blkg); | |
436 | spin_unlock(&blkcg->lock); | |
437 | ||
438 | /* | |
439 | * in order to avoid holding the spin lock for too long, release | |
440 | * it when a batch of blkgs are destroyed. | |
441 | */ | |
442 | if (!(--count)) { | |
443 | count = BLKG_DESTROY_BATCH_SIZE; | |
444 | spin_unlock_irq(&q->queue_lock); | |
445 | cond_resched(); | |
446 | goto restart; | |
447 | } | |
448 | } | |
449 | ||
450 | q->root_blkg = NULL; | |
451 | spin_unlock_irq(&q->queue_lock); | |
452 | } | |
453 | ||
454 | static int blkcg_reset_stats(struct cgroup_subsys_state *css, | |
455 | struct cftype *cftype, u64 val) | |
456 | { | |
457 | struct blkcg *blkcg = css_to_blkcg(css); | |
458 | struct blkcg_gq *blkg; | |
459 | int i, cpu; | |
460 | ||
461 | mutex_lock(&blkcg_pol_mutex); | |
462 | spin_lock_irq(&blkcg->lock); | |
463 | ||
464 | /* | |
465 | * Note that stat reset is racy - it doesn't synchronize against | |
466 | * stat updates. This is a debug feature which shouldn't exist | |
467 | * anyway. If you get hit by a race, retry. | |
468 | */ | |
469 | hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) { | |
470 | for_each_possible_cpu(cpu) { | |
471 | struct blkg_iostat_set *bis = | |
472 | per_cpu_ptr(blkg->iostat_cpu, cpu); | |
473 | memset(bis, 0, sizeof(*bis)); | |
474 | } | |
475 | memset(&blkg->iostat, 0, sizeof(blkg->iostat)); | |
476 | ||
477 | for (i = 0; i < BLKCG_MAX_POLS; i++) { | |
478 | struct blkcg_policy *pol = blkcg_policy[i]; | |
479 | ||
480 | if (blkg->pd[i] && pol->pd_reset_stats_fn) | |
481 | pol->pd_reset_stats_fn(blkg->pd[i]); | |
482 | } | |
483 | } | |
484 | ||
485 | spin_unlock_irq(&blkcg->lock); | |
486 | mutex_unlock(&blkcg_pol_mutex); | |
487 | return 0; | |
488 | } | |
489 | ||
490 | const char *blkg_dev_name(struct blkcg_gq *blkg) | |
491 | { | |
492 | if (!blkg->q->disk || !blkg->q->disk->bdi->dev) | |
493 | return NULL; | |
494 | return bdi_dev_name(blkg->q->disk->bdi); | |
495 | } | |
496 | ||
497 | /** | |
498 | * blkcg_print_blkgs - helper for printing per-blkg data | |
499 | * @sf: seq_file to print to | |
500 | * @blkcg: blkcg of interest | |
501 | * @prfill: fill function to print out a blkg | |
502 | * @pol: policy in question | |
503 | * @data: data to be passed to @prfill | |
504 | * @show_total: to print out sum of prfill return values or not | |
505 | * | |
506 | * This function invokes @prfill on each blkg of @blkcg if pd for the | |
507 | * policy specified by @pol exists. @prfill is invoked with @sf, the | |
508 | * policy data and @data and the matching queue lock held. If @show_total | |
509 | * is %true, the sum of the return values from @prfill is printed with | |
510 | * "Total" label at the end. | |
511 | * | |
512 | * This is to be used to construct print functions for | |
513 | * cftype->read_seq_string method. | |
514 | */ | |
515 | void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg, | |
516 | u64 (*prfill)(struct seq_file *, | |
517 | struct blkg_policy_data *, int), | |
518 | const struct blkcg_policy *pol, int data, | |
519 | bool show_total) | |
520 | { | |
521 | struct blkcg_gq *blkg; | |
522 | u64 total = 0; | |
523 | ||
524 | rcu_read_lock(); | |
525 | hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) { | |
526 | spin_lock_irq(&blkg->q->queue_lock); | |
527 | if (blkcg_policy_enabled(blkg->q, pol)) | |
528 | total += prfill(sf, blkg->pd[pol->plid], data); | |
529 | spin_unlock_irq(&blkg->q->queue_lock); | |
530 | } | |
531 | rcu_read_unlock(); | |
532 | ||
533 | if (show_total) | |
534 | seq_printf(sf, "Total %llu\n", (unsigned long long)total); | |
535 | } | |
536 | EXPORT_SYMBOL_GPL(blkcg_print_blkgs); | |
537 | ||
538 | /** | |
539 | * __blkg_prfill_u64 - prfill helper for a single u64 value | |
540 | * @sf: seq_file to print to | |
541 | * @pd: policy private data of interest | |
542 | * @v: value to print | |
543 | * | |
544 | * Print @v to @sf for the device assocaited with @pd. | |
545 | */ | |
546 | u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v) | |
547 | { | |
548 | const char *dname = blkg_dev_name(pd->blkg); | |
549 | ||
550 | if (!dname) | |
551 | return 0; | |
552 | ||
553 | seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v); | |
554 | return v; | |
555 | } | |
556 | EXPORT_SYMBOL_GPL(__blkg_prfill_u64); | |
557 | ||
558 | /* Performs queue bypass and policy enabled checks then looks up blkg. */ | |
559 | static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg, | |
560 | const struct blkcg_policy *pol, | |
561 | struct request_queue *q) | |
562 | { | |
563 | WARN_ON_ONCE(!rcu_read_lock_held()); | |
564 | lockdep_assert_held(&q->queue_lock); | |
565 | ||
566 | if (!blkcg_policy_enabled(q, pol)) | |
567 | return ERR_PTR(-EOPNOTSUPP); | |
568 | return __blkg_lookup(blkcg, q, true /* update_hint */); | |
569 | } | |
570 | ||
571 | /** | |
572 | * blkcg_conf_open_bdev - parse and open bdev for per-blkg config update | |
573 | * @inputp: input string pointer | |
574 | * | |
575 | * Parse the device node prefix part, MAJ:MIN, of per-blkg config update | |
576 | * from @input and get and return the matching bdev. *@inputp is | |
577 | * updated to point past the device node prefix. Returns an ERR_PTR() | |
578 | * value on error. | |
579 | * | |
580 | * Use this function iff blkg_conf_prep() can't be used for some reason. | |
581 | */ | |
582 | struct block_device *blkcg_conf_open_bdev(char **inputp) | |
583 | { | |
584 | char *input = *inputp; | |
585 | unsigned int major, minor; | |
586 | struct block_device *bdev; | |
587 | int key_len; | |
588 | ||
589 | if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2) | |
590 | return ERR_PTR(-EINVAL); | |
591 | ||
592 | input += key_len; | |
593 | if (!isspace(*input)) | |
594 | return ERR_PTR(-EINVAL); | |
595 | input = skip_spaces(input); | |
596 | ||
597 | bdev = blkdev_get_no_open(MKDEV(major, minor)); | |
598 | if (!bdev) | |
599 | return ERR_PTR(-ENODEV); | |
600 | if (bdev_is_partition(bdev)) { | |
601 | blkdev_put_no_open(bdev); | |
602 | return ERR_PTR(-ENODEV); | |
603 | } | |
604 | ||
605 | *inputp = input; | |
606 | return bdev; | |
607 | } | |
608 | ||
609 | /** | |
610 | * blkg_conf_prep - parse and prepare for per-blkg config update | |
611 | * @blkcg: target block cgroup | |
612 | * @pol: target policy | |
613 | * @input: input string | |
614 | * @ctx: blkg_conf_ctx to be filled | |
615 | * | |
616 | * Parse per-blkg config update from @input and initialize @ctx with the | |
617 | * result. @ctx->blkg points to the blkg to be updated and @ctx->body the | |
618 | * part of @input following MAJ:MIN. This function returns with RCU read | |
619 | * lock and queue lock held and must be paired with blkg_conf_finish(). | |
620 | */ | |
621 | int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol, | |
622 | char *input, struct blkg_conf_ctx *ctx) | |
623 | __acquires(rcu) __acquires(&bdev->bd_disk->queue->queue_lock) | |
624 | { | |
625 | struct block_device *bdev; | |
626 | struct request_queue *q; | |
627 | struct blkcg_gq *blkg; | |
628 | int ret; | |
629 | ||
630 | bdev = blkcg_conf_open_bdev(&input); | |
631 | if (IS_ERR(bdev)) | |
632 | return PTR_ERR(bdev); | |
633 | ||
634 | q = bdev->bd_disk->queue; | |
635 | ||
636 | rcu_read_lock(); | |
637 | spin_lock_irq(&q->queue_lock); | |
638 | ||
639 | blkg = blkg_lookup_check(blkcg, pol, q); | |
640 | if (IS_ERR(blkg)) { | |
641 | ret = PTR_ERR(blkg); | |
642 | goto fail_unlock; | |
643 | } | |
644 | ||
645 | if (blkg) | |
646 | goto success; | |
647 | ||
648 | /* | |
649 | * Create blkgs walking down from blkcg_root to @blkcg, so that all | |
650 | * non-root blkgs have access to their parents. | |
651 | */ | |
652 | while (true) { | |
653 | struct blkcg *pos = blkcg; | |
654 | struct blkcg *parent; | |
655 | struct blkcg_gq *new_blkg; | |
656 | ||
657 | parent = blkcg_parent(blkcg); | |
658 | while (parent && !__blkg_lookup(parent, q, false)) { | |
659 | pos = parent; | |
660 | parent = blkcg_parent(parent); | |
661 | } | |
662 | ||
663 | /* Drop locks to do new blkg allocation with GFP_KERNEL. */ | |
664 | spin_unlock_irq(&q->queue_lock); | |
665 | rcu_read_unlock(); | |
666 | ||
667 | new_blkg = blkg_alloc(pos, q, GFP_KERNEL); | |
668 | if (unlikely(!new_blkg)) { | |
669 | ret = -ENOMEM; | |
670 | goto fail; | |
671 | } | |
672 | ||
673 | if (radix_tree_preload(GFP_KERNEL)) { | |
674 | blkg_free(new_blkg); | |
675 | ret = -ENOMEM; | |
676 | goto fail; | |
677 | } | |
678 | ||
679 | rcu_read_lock(); | |
680 | spin_lock_irq(&q->queue_lock); | |
681 | ||
682 | blkg = blkg_lookup_check(pos, pol, q); | |
683 | if (IS_ERR(blkg)) { | |
684 | ret = PTR_ERR(blkg); | |
685 | blkg_free(new_blkg); | |
686 | goto fail_preloaded; | |
687 | } | |
688 | ||
689 | if (blkg) { | |
690 | blkg_free(new_blkg); | |
691 | } else { | |
692 | blkg = blkg_create(pos, q, new_blkg); | |
693 | if (IS_ERR(blkg)) { | |
694 | ret = PTR_ERR(blkg); | |
695 | goto fail_preloaded; | |
696 | } | |
697 | } | |
698 | ||
699 | radix_tree_preload_end(); | |
700 | ||
701 | if (pos == blkcg) | |
702 | goto success; | |
703 | } | |
704 | success: | |
705 | ctx->bdev = bdev; | |
706 | ctx->blkg = blkg; | |
707 | ctx->body = input; | |
708 | return 0; | |
709 | ||
710 | fail_preloaded: | |
711 | radix_tree_preload_end(); | |
712 | fail_unlock: | |
713 | spin_unlock_irq(&q->queue_lock); | |
714 | rcu_read_unlock(); | |
715 | fail: | |
716 | blkdev_put_no_open(bdev); | |
717 | /* | |
718 | * If queue was bypassing, we should retry. Do so after a | |
719 | * short msleep(). It isn't strictly necessary but queue | |
720 | * can be bypassing for some time and it's always nice to | |
721 | * avoid busy looping. | |
722 | */ | |
723 | if (ret == -EBUSY) { | |
724 | msleep(10); | |
725 | ret = restart_syscall(); | |
726 | } | |
727 | return ret; | |
728 | } | |
729 | EXPORT_SYMBOL_GPL(blkg_conf_prep); | |
730 | ||
731 | /** | |
732 | * blkg_conf_finish - finish up per-blkg config update | |
733 | * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep() | |
734 | * | |
735 | * Finish up after per-blkg config update. This function must be paired | |
736 | * with blkg_conf_prep(). | |
737 | */ | |
738 | void blkg_conf_finish(struct blkg_conf_ctx *ctx) | |
739 | __releases(&ctx->bdev->bd_disk->queue->queue_lock) __releases(rcu) | |
740 | { | |
741 | spin_unlock_irq(&ctx->bdev->bd_disk->queue->queue_lock); | |
742 | rcu_read_unlock(); | |
743 | blkdev_put_no_open(ctx->bdev); | |
744 | } | |
745 | EXPORT_SYMBOL_GPL(blkg_conf_finish); | |
746 | ||
747 | static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src) | |
748 | { | |
749 | int i; | |
750 | ||
751 | for (i = 0; i < BLKG_IOSTAT_NR; i++) { | |
752 | dst->bytes[i] = src->bytes[i]; | |
753 | dst->ios[i] = src->ios[i]; | |
754 | } | |
755 | } | |
756 | ||
757 | static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src) | |
758 | { | |
759 | int i; | |
760 | ||
761 | for (i = 0; i < BLKG_IOSTAT_NR; i++) { | |
762 | dst->bytes[i] += src->bytes[i]; | |
763 | dst->ios[i] += src->ios[i]; | |
764 | } | |
765 | } | |
766 | ||
767 | static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src) | |
768 | { | |
769 | int i; | |
770 | ||
771 | for (i = 0; i < BLKG_IOSTAT_NR; i++) { | |
772 | dst->bytes[i] -= src->bytes[i]; | |
773 | dst->ios[i] -= src->ios[i]; | |
774 | } | |
775 | } | |
776 | ||
777 | static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu) | |
778 | { | |
779 | struct blkcg *blkcg = css_to_blkcg(css); | |
780 | struct blkcg_gq *blkg; | |
781 | ||
782 | /* Root-level stats are sourced from system-wide IO stats */ | |
783 | if (!cgroup_parent(css->cgroup)) | |
784 | return; | |
785 | ||
786 | rcu_read_lock(); | |
787 | ||
788 | hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) { | |
789 | struct blkcg_gq *parent = blkg->parent; | |
790 | struct blkg_iostat_set *bisc = per_cpu_ptr(blkg->iostat_cpu, cpu); | |
791 | struct blkg_iostat cur, delta; | |
792 | unsigned long flags; | |
793 | unsigned int seq; | |
794 | ||
795 | /* fetch the current per-cpu values */ | |
796 | do { | |
797 | seq = u64_stats_fetch_begin(&bisc->sync); | |
798 | blkg_iostat_set(&cur, &bisc->cur); | |
799 | } while (u64_stats_fetch_retry(&bisc->sync, seq)); | |
800 | ||
801 | /* propagate percpu delta to global */ | |
802 | flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync); | |
803 | blkg_iostat_set(&delta, &cur); | |
804 | blkg_iostat_sub(&delta, &bisc->last); | |
805 | blkg_iostat_add(&blkg->iostat.cur, &delta); | |
806 | blkg_iostat_add(&bisc->last, &delta); | |
807 | u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags); | |
808 | ||
809 | /* propagate global delta to parent (unless that's root) */ | |
810 | if (parent && parent->parent) { | |
811 | flags = u64_stats_update_begin_irqsave(&parent->iostat.sync); | |
812 | blkg_iostat_set(&delta, &blkg->iostat.cur); | |
813 | blkg_iostat_sub(&delta, &blkg->iostat.last); | |
814 | blkg_iostat_add(&parent->iostat.cur, &delta); | |
815 | blkg_iostat_add(&blkg->iostat.last, &delta); | |
816 | u64_stats_update_end_irqrestore(&parent->iostat.sync, flags); | |
817 | } | |
818 | } | |
819 | ||
820 | rcu_read_unlock(); | |
821 | } | |
822 | ||
823 | /* | |
824 | * We source root cgroup stats from the system-wide stats to avoid | |
825 | * tracking the same information twice and incurring overhead when no | |
826 | * cgroups are defined. For that reason, cgroup_rstat_flush in | |
827 | * blkcg_print_stat does not actually fill out the iostat in the root | |
828 | * cgroup's blkcg_gq. | |
829 | * | |
830 | * However, we would like to re-use the printing code between the root and | |
831 | * non-root cgroups to the extent possible. For that reason, we simulate | |
832 | * flushing the root cgroup's stats by explicitly filling in the iostat | |
833 | * with disk level statistics. | |
834 | */ | |
835 | static void blkcg_fill_root_iostats(void) | |
836 | { | |
837 | struct class_dev_iter iter; | |
838 | struct device *dev; | |
839 | ||
840 | class_dev_iter_init(&iter, &block_class, NULL, &disk_type); | |
841 | while ((dev = class_dev_iter_next(&iter))) { | |
842 | struct block_device *bdev = dev_to_bdev(dev); | |
843 | struct blkcg_gq *blkg = | |
844 | blk_queue_root_blkg(bdev->bd_disk->queue); | |
845 | struct blkg_iostat tmp; | |
846 | int cpu; | |
847 | ||
848 | memset(&tmp, 0, sizeof(tmp)); | |
849 | for_each_possible_cpu(cpu) { | |
850 | struct disk_stats *cpu_dkstats; | |
851 | unsigned long flags; | |
852 | ||
853 | cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu); | |
854 | tmp.ios[BLKG_IOSTAT_READ] += | |
855 | cpu_dkstats->ios[STAT_READ]; | |
856 | tmp.ios[BLKG_IOSTAT_WRITE] += | |
857 | cpu_dkstats->ios[STAT_WRITE]; | |
858 | tmp.ios[BLKG_IOSTAT_DISCARD] += | |
859 | cpu_dkstats->ios[STAT_DISCARD]; | |
860 | // convert sectors to bytes | |
861 | tmp.bytes[BLKG_IOSTAT_READ] += | |
862 | cpu_dkstats->sectors[STAT_READ] << 9; | |
863 | tmp.bytes[BLKG_IOSTAT_WRITE] += | |
864 | cpu_dkstats->sectors[STAT_WRITE] << 9; | |
865 | tmp.bytes[BLKG_IOSTAT_DISCARD] += | |
866 | cpu_dkstats->sectors[STAT_DISCARD] << 9; | |
867 | ||
868 | flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync); | |
869 | blkg_iostat_set(&blkg->iostat.cur, &tmp); | |
870 | u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags); | |
871 | } | |
872 | } | |
873 | } | |
874 | ||
875 | static void blkcg_print_one_stat(struct blkcg_gq *blkg, struct seq_file *s) | |
876 | { | |
877 | struct blkg_iostat_set *bis = &blkg->iostat; | |
878 | u64 rbytes, wbytes, rios, wios, dbytes, dios; | |
879 | bool has_stats = false; | |
880 | const char *dname; | |
881 | unsigned seq; | |
882 | int i; | |
883 | ||
884 | if (!blkg->online) | |
885 | return; | |
886 | ||
887 | dname = blkg_dev_name(blkg); | |
888 | if (!dname) | |
889 | return; | |
890 | ||
891 | seq_printf(s, "%s ", dname); | |
892 | ||
893 | do { | |
894 | seq = u64_stats_fetch_begin(&bis->sync); | |
895 | ||
896 | rbytes = bis->cur.bytes[BLKG_IOSTAT_READ]; | |
897 | wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE]; | |
898 | dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD]; | |
899 | rios = bis->cur.ios[BLKG_IOSTAT_READ]; | |
900 | wios = bis->cur.ios[BLKG_IOSTAT_WRITE]; | |
901 | dios = bis->cur.ios[BLKG_IOSTAT_DISCARD]; | |
902 | } while (u64_stats_fetch_retry(&bis->sync, seq)); | |
903 | ||
904 | if (rbytes || wbytes || rios || wios) { | |
905 | has_stats = true; | |
906 | seq_printf(s, "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu", | |
907 | rbytes, wbytes, rios, wios, | |
908 | dbytes, dios); | |
909 | } | |
910 | ||
911 | if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) { | |
912 | has_stats = true; | |
913 | seq_printf(s, " use_delay=%d delay_nsec=%llu", | |
914 | atomic_read(&blkg->use_delay), | |
915 | atomic64_read(&blkg->delay_nsec)); | |
916 | } | |
917 | ||
918 | for (i = 0; i < BLKCG_MAX_POLS; i++) { | |
919 | struct blkcg_policy *pol = blkcg_policy[i]; | |
920 | ||
921 | if (!blkg->pd[i] || !pol->pd_stat_fn) | |
922 | continue; | |
923 | ||
924 | if (pol->pd_stat_fn(blkg->pd[i], s)) | |
925 | has_stats = true; | |
926 | } | |
927 | ||
928 | if (has_stats) | |
929 | seq_printf(s, "\n"); | |
930 | } | |
931 | ||
932 | static int blkcg_print_stat(struct seq_file *sf, void *v) | |
933 | { | |
934 | struct blkcg *blkcg = css_to_blkcg(seq_css(sf)); | |
935 | struct blkcg_gq *blkg; | |
936 | ||
937 | if (!seq_css(sf)->parent) | |
938 | blkcg_fill_root_iostats(); | |
939 | else | |
940 | cgroup_rstat_flush(blkcg->css.cgroup); | |
941 | ||
942 | rcu_read_lock(); | |
943 | hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) { | |
944 | spin_lock_irq(&blkg->q->queue_lock); | |
945 | blkcg_print_one_stat(blkg, sf); | |
946 | spin_unlock_irq(&blkg->q->queue_lock); | |
947 | } | |
948 | rcu_read_unlock(); | |
949 | return 0; | |
950 | } | |
951 | ||
952 | static struct cftype blkcg_files[] = { | |
953 | { | |
954 | .name = "stat", | |
955 | .seq_show = blkcg_print_stat, | |
956 | }, | |
957 | { } /* terminate */ | |
958 | }; | |
959 | ||
960 | static struct cftype blkcg_legacy_files[] = { | |
961 | { | |
962 | .name = "reset_stats", | |
963 | .write_u64 = blkcg_reset_stats, | |
964 | }, | |
965 | { } /* terminate */ | |
966 | }; | |
967 | ||
968 | /* | |
969 | * blkcg destruction is a three-stage process. | |
970 | * | |
971 | * 1. Destruction starts. The blkcg_css_offline() callback is invoked | |
972 | * which offlines writeback. Here we tie the next stage of blkg destruction | |
973 | * to the completion of writeback associated with the blkcg. This lets us | |
974 | * avoid punting potentially large amounts of outstanding writeback to root | |
975 | * while maintaining any ongoing policies. The next stage is triggered when | |
976 | * the nr_cgwbs count goes to zero. | |
977 | * | |
978 | * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called | |
979 | * and handles the destruction of blkgs. Here the css reference held by | |
980 | * the blkg is put back eventually allowing blkcg_css_free() to be called. | |
981 | * This work may occur in cgwb_release_workfn() on the cgwb_release | |
982 | * workqueue. Any submitted ios that fail to get the blkg ref will be | |
983 | * punted to the root_blkg. | |
984 | * | |
985 | * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called. | |
986 | * This finally frees the blkcg. | |
987 | */ | |
988 | ||
989 | /** | |
990 | * blkcg_css_offline - cgroup css_offline callback | |
991 | * @css: css of interest | |
992 | * | |
993 | * This function is called when @css is about to go away. Here the cgwbs are | |
994 | * offlined first and only once writeback associated with the blkcg has | |
995 | * finished do we start step 2 (see above). | |
996 | */ | |
997 | static void blkcg_css_offline(struct cgroup_subsys_state *css) | |
998 | { | |
999 | struct blkcg *blkcg = css_to_blkcg(css); | |
1000 | ||
1001 | /* this prevents anyone from attaching or migrating to this blkcg */ | |
1002 | wb_blkcg_offline(blkcg); | |
1003 | ||
1004 | /* put the base online pin allowing step 2 to be triggered */ | |
1005 | blkcg_unpin_online(blkcg); | |
1006 | } | |
1007 | ||
1008 | /** | |
1009 | * blkcg_destroy_blkgs - responsible for shooting down blkgs | |
1010 | * @blkcg: blkcg of interest | |
1011 | * | |
1012 | * blkgs should be removed while holding both q and blkcg locks. As blkcg lock | |
1013 | * is nested inside q lock, this function performs reverse double lock dancing. | |
1014 | * Destroying the blkgs releases the reference held on the blkcg's css allowing | |
1015 | * blkcg_css_free to eventually be called. | |
1016 | * | |
1017 | * This is the blkcg counterpart of ioc_release_fn(). | |
1018 | */ | |
1019 | void blkcg_destroy_blkgs(struct blkcg *blkcg) | |
1020 | { | |
1021 | might_sleep(); | |
1022 | ||
1023 | spin_lock_irq(&blkcg->lock); | |
1024 | ||
1025 | while (!hlist_empty(&blkcg->blkg_list)) { | |
1026 | struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first, | |
1027 | struct blkcg_gq, blkcg_node); | |
1028 | struct request_queue *q = blkg->q; | |
1029 | ||
1030 | if (need_resched() || !spin_trylock(&q->queue_lock)) { | |
1031 | /* | |
1032 | * Given that the system can accumulate a huge number | |
1033 | * of blkgs in pathological cases, check to see if we | |
1034 | * need to rescheduling to avoid softlockup. | |
1035 | */ | |
1036 | spin_unlock_irq(&blkcg->lock); | |
1037 | cond_resched(); | |
1038 | spin_lock_irq(&blkcg->lock); | |
1039 | continue; | |
1040 | } | |
1041 | ||
1042 | blkg_destroy(blkg); | |
1043 | spin_unlock(&q->queue_lock); | |
1044 | } | |
1045 | ||
1046 | spin_unlock_irq(&blkcg->lock); | |
1047 | } | |
1048 | ||
1049 | static void blkcg_css_free(struct cgroup_subsys_state *css) | |
1050 | { | |
1051 | struct blkcg *blkcg = css_to_blkcg(css); | |
1052 | int i; | |
1053 | ||
1054 | mutex_lock(&blkcg_pol_mutex); | |
1055 | ||
1056 | list_del(&blkcg->all_blkcgs_node); | |
1057 | ||
1058 | for (i = 0; i < BLKCG_MAX_POLS; i++) | |
1059 | if (blkcg->cpd[i]) | |
1060 | blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]); | |
1061 | ||
1062 | mutex_unlock(&blkcg_pol_mutex); | |
1063 | ||
1064 | kfree(blkcg); | |
1065 | } | |
1066 | ||
1067 | static struct cgroup_subsys_state * | |
1068 | blkcg_css_alloc(struct cgroup_subsys_state *parent_css) | |
1069 | { | |
1070 | struct blkcg *blkcg; | |
1071 | struct cgroup_subsys_state *ret; | |
1072 | int i; | |
1073 | ||
1074 | mutex_lock(&blkcg_pol_mutex); | |
1075 | ||
1076 | if (!parent_css) { | |
1077 | blkcg = &blkcg_root; | |
1078 | } else { | |
1079 | blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL); | |
1080 | if (!blkcg) { | |
1081 | ret = ERR_PTR(-ENOMEM); | |
1082 | goto unlock; | |
1083 | } | |
1084 | } | |
1085 | ||
1086 | for (i = 0; i < BLKCG_MAX_POLS ; i++) { | |
1087 | struct blkcg_policy *pol = blkcg_policy[i]; | |
1088 | struct blkcg_policy_data *cpd; | |
1089 | ||
1090 | /* | |
1091 | * If the policy hasn't been attached yet, wait for it | |
1092 | * to be attached before doing anything else. Otherwise, | |
1093 | * check if the policy requires any specific per-cgroup | |
1094 | * data: if it does, allocate and initialize it. | |
1095 | */ | |
1096 | if (!pol || !pol->cpd_alloc_fn) | |
1097 | continue; | |
1098 | ||
1099 | cpd = pol->cpd_alloc_fn(GFP_KERNEL); | |
1100 | if (!cpd) { | |
1101 | ret = ERR_PTR(-ENOMEM); | |
1102 | goto free_pd_blkcg; | |
1103 | } | |
1104 | blkcg->cpd[i] = cpd; | |
1105 | cpd->blkcg = blkcg; | |
1106 | cpd->plid = i; | |
1107 | if (pol->cpd_init_fn) | |
1108 | pol->cpd_init_fn(cpd); | |
1109 | } | |
1110 | ||
1111 | spin_lock_init(&blkcg->lock); | |
1112 | refcount_set(&blkcg->online_pin, 1); | |
1113 | INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN); | |
1114 | INIT_HLIST_HEAD(&blkcg->blkg_list); | |
1115 | #ifdef CONFIG_CGROUP_WRITEBACK | |
1116 | INIT_LIST_HEAD(&blkcg->cgwb_list); | |
1117 | #endif | |
1118 | list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs); | |
1119 | ||
1120 | mutex_unlock(&blkcg_pol_mutex); | |
1121 | return &blkcg->css; | |
1122 | ||
1123 | free_pd_blkcg: | |
1124 | for (i--; i >= 0; i--) | |
1125 | if (blkcg->cpd[i]) | |
1126 | blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]); | |
1127 | ||
1128 | if (blkcg != &blkcg_root) | |
1129 | kfree(blkcg); | |
1130 | unlock: | |
1131 | mutex_unlock(&blkcg_pol_mutex); | |
1132 | return ret; | |
1133 | } | |
1134 | ||
1135 | static int blkcg_css_online(struct cgroup_subsys_state *css) | |
1136 | { | |
1137 | struct blkcg *blkcg = css_to_blkcg(css); | |
1138 | struct blkcg *parent = blkcg_parent(blkcg); | |
1139 | ||
1140 | /* | |
1141 | * blkcg_pin_online() is used to delay blkcg offline so that blkgs | |
1142 | * don't go offline while cgwbs are still active on them. Pin the | |
1143 | * parent so that offline always happens towards the root. | |
1144 | */ | |
1145 | if (parent) | |
1146 | blkcg_pin_online(parent); | |
1147 | return 0; | |
1148 | } | |
1149 | ||
1150 | /** | |
1151 | * blkcg_init_queue - initialize blkcg part of request queue | |
1152 | * @q: request_queue to initialize | |
1153 | * | |
1154 | * Called from blk_alloc_queue(). Responsible for initializing blkcg | |
1155 | * part of new request_queue @q. | |
1156 | * | |
1157 | * RETURNS: | |
1158 | * 0 on success, -errno on failure. | |
1159 | */ | |
1160 | int blkcg_init_queue(struct request_queue *q) | |
1161 | { | |
1162 | struct blkcg_gq *new_blkg, *blkg; | |
1163 | bool preloaded; | |
1164 | int ret; | |
1165 | ||
1166 | new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL); | |
1167 | if (!new_blkg) | |
1168 | return -ENOMEM; | |
1169 | ||
1170 | preloaded = !radix_tree_preload(GFP_KERNEL); | |
1171 | ||
1172 | /* Make sure the root blkg exists. */ | |
1173 | rcu_read_lock(); | |
1174 | spin_lock_irq(&q->queue_lock); | |
1175 | blkg = blkg_create(&blkcg_root, q, new_blkg); | |
1176 | if (IS_ERR(blkg)) | |
1177 | goto err_unlock; | |
1178 | q->root_blkg = blkg; | |
1179 | spin_unlock_irq(&q->queue_lock); | |
1180 | rcu_read_unlock(); | |
1181 | ||
1182 | if (preloaded) | |
1183 | radix_tree_preload_end(); | |
1184 | ||
1185 | ret = blk_iolatency_init(q); | |
1186 | if (ret) | |
1187 | goto err_destroy_all; | |
1188 | ||
1189 | ret = blk_ioprio_init(q); | |
1190 | if (ret) | |
1191 | goto err_destroy_all; | |
1192 | ||
1193 | ret = blk_throtl_init(q); | |
1194 | if (ret) | |
1195 | goto err_destroy_all; | |
1196 | ||
1197 | return 0; | |
1198 | ||
1199 | err_destroy_all: | |
1200 | blkg_destroy_all(q); | |
1201 | return ret; | |
1202 | err_unlock: | |
1203 | spin_unlock_irq(&q->queue_lock); | |
1204 | rcu_read_unlock(); | |
1205 | if (preloaded) | |
1206 | radix_tree_preload_end(); | |
1207 | return PTR_ERR(blkg); | |
1208 | } | |
1209 | ||
1210 | /** | |
1211 | * blkcg_exit_queue - exit and release blkcg part of request_queue | |
1212 | * @q: request_queue being released | |
1213 | * | |
1214 | * Called from blk_exit_queue(). Responsible for exiting blkcg part. | |
1215 | */ | |
1216 | void blkcg_exit_queue(struct request_queue *q) | |
1217 | { | |
1218 | blkg_destroy_all(q); | |
1219 | blk_throtl_exit(q); | |
1220 | } | |
1221 | ||
1222 | static void blkcg_bind(struct cgroup_subsys_state *root_css) | |
1223 | { | |
1224 | int i; | |
1225 | ||
1226 | mutex_lock(&blkcg_pol_mutex); | |
1227 | ||
1228 | for (i = 0; i < BLKCG_MAX_POLS; i++) { | |
1229 | struct blkcg_policy *pol = blkcg_policy[i]; | |
1230 | struct blkcg *blkcg; | |
1231 | ||
1232 | if (!pol || !pol->cpd_bind_fn) | |
1233 | continue; | |
1234 | ||
1235 | list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) | |
1236 | if (blkcg->cpd[pol->plid]) | |
1237 | pol->cpd_bind_fn(blkcg->cpd[pol->plid]); | |
1238 | } | |
1239 | mutex_unlock(&blkcg_pol_mutex); | |
1240 | } | |
1241 | ||
1242 | static void blkcg_exit(struct task_struct *tsk) | |
1243 | { | |
1244 | if (tsk->throttle_queue) | |
1245 | blk_put_queue(tsk->throttle_queue); | |
1246 | tsk->throttle_queue = NULL; | |
1247 | } | |
1248 | ||
1249 | struct cgroup_subsys io_cgrp_subsys = { | |
1250 | .css_alloc = blkcg_css_alloc, | |
1251 | .css_online = blkcg_css_online, | |
1252 | .css_offline = blkcg_css_offline, | |
1253 | .css_free = blkcg_css_free, | |
1254 | .css_rstat_flush = blkcg_rstat_flush, | |
1255 | .bind = blkcg_bind, | |
1256 | .dfl_cftypes = blkcg_files, | |
1257 | .legacy_cftypes = blkcg_legacy_files, | |
1258 | .legacy_name = "blkio", | |
1259 | .exit = blkcg_exit, | |
1260 | #ifdef CONFIG_MEMCG | |
1261 | /* | |
1262 | * This ensures that, if available, memcg is automatically enabled | |
1263 | * together on the default hierarchy so that the owner cgroup can | |
1264 | * be retrieved from writeback pages. | |
1265 | */ | |
1266 | .depends_on = 1 << memory_cgrp_id, | |
1267 | #endif | |
1268 | }; | |
1269 | EXPORT_SYMBOL_GPL(io_cgrp_subsys); | |
1270 | ||
1271 | /** | |
1272 | * blkcg_activate_policy - activate a blkcg policy on a request_queue | |
1273 | * @q: request_queue of interest | |
1274 | * @pol: blkcg policy to activate | |
1275 | * | |
1276 | * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through | |
1277 | * bypass mode to populate its blkgs with policy_data for @pol. | |
1278 | * | |
1279 | * Activation happens with @q bypassed, so nobody would be accessing blkgs | |
1280 | * from IO path. Update of each blkg is protected by both queue and blkcg | |
1281 | * locks so that holding either lock and testing blkcg_policy_enabled() is | |
1282 | * always enough for dereferencing policy data. | |
1283 | * | |
1284 | * The caller is responsible for synchronizing [de]activations and policy | |
1285 | * [un]registerations. Returns 0 on success, -errno on failure. | |
1286 | */ | |
1287 | int blkcg_activate_policy(struct request_queue *q, | |
1288 | const struct blkcg_policy *pol) | |
1289 | { | |
1290 | struct blkg_policy_data *pd_prealloc = NULL; | |
1291 | struct blkcg_gq *blkg, *pinned_blkg = NULL; | |
1292 | int ret; | |
1293 | ||
1294 | if (blkcg_policy_enabled(q, pol)) | |
1295 | return 0; | |
1296 | ||
1297 | if (queue_is_mq(q)) | |
1298 | blk_mq_freeze_queue(q); | |
1299 | retry: | |
1300 | spin_lock_irq(&q->queue_lock); | |
1301 | ||
1302 | /* blkg_list is pushed at the head, reverse walk to allocate parents first */ | |
1303 | list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) { | |
1304 | struct blkg_policy_data *pd; | |
1305 | ||
1306 | if (blkg->pd[pol->plid]) | |
1307 | continue; | |
1308 | ||
1309 | /* If prealloc matches, use it; otherwise try GFP_NOWAIT */ | |
1310 | if (blkg == pinned_blkg) { | |
1311 | pd = pd_prealloc; | |
1312 | pd_prealloc = NULL; | |
1313 | } else { | |
1314 | pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q, | |
1315 | blkg->blkcg); | |
1316 | } | |
1317 | ||
1318 | if (!pd) { | |
1319 | /* | |
1320 | * GFP_NOWAIT failed. Free the existing one and | |
1321 | * prealloc for @blkg w/ GFP_KERNEL. | |
1322 | */ | |
1323 | if (pinned_blkg) | |
1324 | blkg_put(pinned_blkg); | |
1325 | blkg_get(blkg); | |
1326 | pinned_blkg = blkg; | |
1327 | ||
1328 | spin_unlock_irq(&q->queue_lock); | |
1329 | ||
1330 | if (pd_prealloc) | |
1331 | pol->pd_free_fn(pd_prealloc); | |
1332 | pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q, | |
1333 | blkg->blkcg); | |
1334 | if (pd_prealloc) | |
1335 | goto retry; | |
1336 | else | |
1337 | goto enomem; | |
1338 | } | |
1339 | ||
1340 | blkg->pd[pol->plid] = pd; | |
1341 | pd->blkg = blkg; | |
1342 | pd->plid = pol->plid; | |
1343 | } | |
1344 | ||
1345 | /* all allocated, init in the same order */ | |
1346 | if (pol->pd_init_fn) | |
1347 | list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) | |
1348 | pol->pd_init_fn(blkg->pd[pol->plid]); | |
1349 | ||
1350 | __set_bit(pol->plid, q->blkcg_pols); | |
1351 | ret = 0; | |
1352 | ||
1353 | spin_unlock_irq(&q->queue_lock); | |
1354 | out: | |
1355 | if (queue_is_mq(q)) | |
1356 | blk_mq_unfreeze_queue(q); | |
1357 | if (pinned_blkg) | |
1358 | blkg_put(pinned_blkg); | |
1359 | if (pd_prealloc) | |
1360 | pol->pd_free_fn(pd_prealloc); | |
1361 | return ret; | |
1362 | ||
1363 | enomem: | |
1364 | /* alloc failed, nothing's initialized yet, free everything */ | |
1365 | spin_lock_irq(&q->queue_lock); | |
1366 | list_for_each_entry(blkg, &q->blkg_list, q_node) { | |
1367 | if (blkg->pd[pol->plid]) { | |
1368 | pol->pd_free_fn(blkg->pd[pol->plid]); | |
1369 | blkg->pd[pol->plid] = NULL; | |
1370 | } | |
1371 | } | |
1372 | spin_unlock_irq(&q->queue_lock); | |
1373 | ret = -ENOMEM; | |
1374 | goto out; | |
1375 | } | |
1376 | EXPORT_SYMBOL_GPL(blkcg_activate_policy); | |
1377 | ||
1378 | /** | |
1379 | * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue | |
1380 | * @q: request_queue of interest | |
1381 | * @pol: blkcg policy to deactivate | |
1382 | * | |
1383 | * Deactivate @pol on @q. Follows the same synchronization rules as | |
1384 | * blkcg_activate_policy(). | |
1385 | */ | |
1386 | void blkcg_deactivate_policy(struct request_queue *q, | |
1387 | const struct blkcg_policy *pol) | |
1388 | { | |
1389 | struct blkcg_gq *blkg; | |
1390 | ||
1391 | if (!blkcg_policy_enabled(q, pol)) | |
1392 | return; | |
1393 | ||
1394 | if (queue_is_mq(q)) | |
1395 | blk_mq_freeze_queue(q); | |
1396 | ||
1397 | spin_lock_irq(&q->queue_lock); | |
1398 | ||
1399 | __clear_bit(pol->plid, q->blkcg_pols); | |
1400 | ||
1401 | list_for_each_entry(blkg, &q->blkg_list, q_node) { | |
1402 | if (blkg->pd[pol->plid]) { | |
1403 | if (pol->pd_offline_fn) | |
1404 | pol->pd_offline_fn(blkg->pd[pol->plid]); | |
1405 | pol->pd_free_fn(blkg->pd[pol->plid]); | |
1406 | blkg->pd[pol->plid] = NULL; | |
1407 | } | |
1408 | } | |
1409 | ||
1410 | spin_unlock_irq(&q->queue_lock); | |
1411 | ||
1412 | if (queue_is_mq(q)) | |
1413 | blk_mq_unfreeze_queue(q); | |
1414 | } | |
1415 | EXPORT_SYMBOL_GPL(blkcg_deactivate_policy); | |
1416 | ||
1417 | /** | |
1418 | * blkcg_policy_register - register a blkcg policy | |
1419 | * @pol: blkcg policy to register | |
1420 | * | |
1421 | * Register @pol with blkcg core. Might sleep and @pol may be modified on | |
1422 | * successful registration. Returns 0 on success and -errno on failure. | |
1423 | */ | |
1424 | int blkcg_policy_register(struct blkcg_policy *pol) | |
1425 | { | |
1426 | struct blkcg *blkcg; | |
1427 | int i, ret; | |
1428 | ||
1429 | mutex_lock(&blkcg_pol_register_mutex); | |
1430 | mutex_lock(&blkcg_pol_mutex); | |
1431 | ||
1432 | /* find an empty slot */ | |
1433 | ret = -ENOSPC; | |
1434 | for (i = 0; i < BLKCG_MAX_POLS; i++) | |
1435 | if (!blkcg_policy[i]) | |
1436 | break; | |
1437 | if (i >= BLKCG_MAX_POLS) { | |
1438 | pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n"); | |
1439 | goto err_unlock; | |
1440 | } | |
1441 | ||
1442 | /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */ | |
1443 | if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) || | |
1444 | (!pol->pd_alloc_fn ^ !pol->pd_free_fn)) | |
1445 | goto err_unlock; | |
1446 | ||
1447 | /* register @pol */ | |
1448 | pol->plid = i; | |
1449 | blkcg_policy[pol->plid] = pol; | |
1450 | ||
1451 | /* allocate and install cpd's */ | |
1452 | if (pol->cpd_alloc_fn) { | |
1453 | list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) { | |
1454 | struct blkcg_policy_data *cpd; | |
1455 | ||
1456 | cpd = pol->cpd_alloc_fn(GFP_KERNEL); | |
1457 | if (!cpd) | |
1458 | goto err_free_cpds; | |
1459 | ||
1460 | blkcg->cpd[pol->plid] = cpd; | |
1461 | cpd->blkcg = blkcg; | |
1462 | cpd->plid = pol->plid; | |
1463 | if (pol->cpd_init_fn) | |
1464 | pol->cpd_init_fn(cpd); | |
1465 | } | |
1466 | } | |
1467 | ||
1468 | mutex_unlock(&blkcg_pol_mutex); | |
1469 | ||
1470 | /* everything is in place, add intf files for the new policy */ | |
1471 | if (pol->dfl_cftypes) | |
1472 | WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys, | |
1473 | pol->dfl_cftypes)); | |
1474 | if (pol->legacy_cftypes) | |
1475 | WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys, | |
1476 | pol->legacy_cftypes)); | |
1477 | mutex_unlock(&blkcg_pol_register_mutex); | |
1478 | return 0; | |
1479 | ||
1480 | err_free_cpds: | |
1481 | if (pol->cpd_free_fn) { | |
1482 | list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) { | |
1483 | if (blkcg->cpd[pol->plid]) { | |
1484 | pol->cpd_free_fn(blkcg->cpd[pol->plid]); | |
1485 | blkcg->cpd[pol->plid] = NULL; | |
1486 | } | |
1487 | } | |
1488 | } | |
1489 | blkcg_policy[pol->plid] = NULL; | |
1490 | err_unlock: | |
1491 | mutex_unlock(&blkcg_pol_mutex); | |
1492 | mutex_unlock(&blkcg_pol_register_mutex); | |
1493 | return ret; | |
1494 | } | |
1495 | EXPORT_SYMBOL_GPL(blkcg_policy_register); | |
1496 | ||
1497 | /** | |
1498 | * blkcg_policy_unregister - unregister a blkcg policy | |
1499 | * @pol: blkcg policy to unregister | |
1500 | * | |
1501 | * Undo blkcg_policy_register(@pol). Might sleep. | |
1502 | */ | |
1503 | void blkcg_policy_unregister(struct blkcg_policy *pol) | |
1504 | { | |
1505 | struct blkcg *blkcg; | |
1506 | ||
1507 | mutex_lock(&blkcg_pol_register_mutex); | |
1508 | ||
1509 | if (WARN_ON(blkcg_policy[pol->plid] != pol)) | |
1510 | goto out_unlock; | |
1511 | ||
1512 | /* kill the intf files first */ | |
1513 | if (pol->dfl_cftypes) | |
1514 | cgroup_rm_cftypes(pol->dfl_cftypes); | |
1515 | if (pol->legacy_cftypes) | |
1516 | cgroup_rm_cftypes(pol->legacy_cftypes); | |
1517 | ||
1518 | /* remove cpds and unregister */ | |
1519 | mutex_lock(&blkcg_pol_mutex); | |
1520 | ||
1521 | if (pol->cpd_free_fn) { | |
1522 | list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) { | |
1523 | if (blkcg->cpd[pol->plid]) { | |
1524 | pol->cpd_free_fn(blkcg->cpd[pol->plid]); | |
1525 | blkcg->cpd[pol->plid] = NULL; | |
1526 | } | |
1527 | } | |
1528 | } | |
1529 | blkcg_policy[pol->plid] = NULL; | |
1530 | ||
1531 | mutex_unlock(&blkcg_pol_mutex); | |
1532 | out_unlock: | |
1533 | mutex_unlock(&blkcg_pol_register_mutex); | |
1534 | } | |
1535 | EXPORT_SYMBOL_GPL(blkcg_policy_unregister); | |
1536 | ||
1537 | bool __blkcg_punt_bio_submit(struct bio *bio) | |
1538 | { | |
1539 | struct blkcg_gq *blkg = bio->bi_blkg; | |
1540 | ||
1541 | /* consume the flag first */ | |
1542 | bio->bi_opf &= ~REQ_CGROUP_PUNT; | |
1543 | ||
1544 | /* never bounce for the root cgroup */ | |
1545 | if (!blkg->parent) | |
1546 | return false; | |
1547 | ||
1548 | spin_lock_bh(&blkg->async_bio_lock); | |
1549 | bio_list_add(&blkg->async_bios, bio); | |
1550 | spin_unlock_bh(&blkg->async_bio_lock); | |
1551 | ||
1552 | queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work); | |
1553 | return true; | |
1554 | } | |
1555 | ||
1556 | /* | |
1557 | * Scale the accumulated delay based on how long it has been since we updated | |
1558 | * the delay. We only call this when we are adding delay, in case it's been a | |
1559 | * while since we added delay, and when we are checking to see if we need to | |
1560 | * delay a task, to account for any delays that may have occurred. | |
1561 | */ | |
1562 | static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now) | |
1563 | { | |
1564 | u64 old = atomic64_read(&blkg->delay_start); | |
1565 | ||
1566 | /* negative use_delay means no scaling, see blkcg_set_delay() */ | |
1567 | if (atomic_read(&blkg->use_delay) < 0) | |
1568 | return; | |
1569 | ||
1570 | /* | |
1571 | * We only want to scale down every second. The idea here is that we | |
1572 | * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain | |
1573 | * time window. We only want to throttle tasks for recent delay that | |
1574 | * has occurred, in 1 second time windows since that's the maximum | |
1575 | * things can be throttled. We save the current delay window in | |
1576 | * blkg->last_delay so we know what amount is still left to be charged | |
1577 | * to the blkg from this point onward. blkg->last_use keeps track of | |
1578 | * the use_delay counter. The idea is if we're unthrottling the blkg we | |
1579 | * are ok with whatever is happening now, and we can take away more of | |
1580 | * the accumulated delay as we've already throttled enough that | |
1581 | * everybody is happy with their IO latencies. | |
1582 | */ | |
1583 | if (time_before64(old + NSEC_PER_SEC, now) && | |
1584 | atomic64_cmpxchg(&blkg->delay_start, old, now) == old) { | |
1585 | u64 cur = atomic64_read(&blkg->delay_nsec); | |
1586 | u64 sub = min_t(u64, blkg->last_delay, now - old); | |
1587 | int cur_use = atomic_read(&blkg->use_delay); | |
1588 | ||
1589 | /* | |
1590 | * We've been unthrottled, subtract a larger chunk of our | |
1591 | * accumulated delay. | |
1592 | */ | |
1593 | if (cur_use < blkg->last_use) | |
1594 | sub = max_t(u64, sub, blkg->last_delay >> 1); | |
1595 | ||
1596 | /* | |
1597 | * This shouldn't happen, but handle it anyway. Our delay_nsec | |
1598 | * should only ever be growing except here where we subtract out | |
1599 | * min(last_delay, 1 second), but lord knows bugs happen and I'd | |
1600 | * rather not end up with negative numbers. | |
1601 | */ | |
1602 | if (unlikely(cur < sub)) { | |
1603 | atomic64_set(&blkg->delay_nsec, 0); | |
1604 | blkg->last_delay = 0; | |
1605 | } else { | |
1606 | atomic64_sub(sub, &blkg->delay_nsec); | |
1607 | blkg->last_delay = cur - sub; | |
1608 | } | |
1609 | blkg->last_use = cur_use; | |
1610 | } | |
1611 | } | |
1612 | ||
1613 | /* | |
1614 | * This is called when we want to actually walk up the hierarchy and check to | |
1615 | * see if we need to throttle, and then actually throttle if there is some | |
1616 | * accumulated delay. This should only be called upon return to user space so | |
1617 | * we're not holding some lock that would induce a priority inversion. | |
1618 | */ | |
1619 | static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay) | |
1620 | { | |
1621 | unsigned long pflags; | |
1622 | bool clamp; | |
1623 | u64 now = ktime_to_ns(ktime_get()); | |
1624 | u64 exp; | |
1625 | u64 delay_nsec = 0; | |
1626 | int tok; | |
1627 | ||
1628 | while (blkg->parent) { | |
1629 | int use_delay = atomic_read(&blkg->use_delay); | |
1630 | ||
1631 | if (use_delay) { | |
1632 | u64 this_delay; | |
1633 | ||
1634 | blkcg_scale_delay(blkg, now); | |
1635 | this_delay = atomic64_read(&blkg->delay_nsec); | |
1636 | if (this_delay > delay_nsec) { | |
1637 | delay_nsec = this_delay; | |
1638 | clamp = use_delay > 0; | |
1639 | } | |
1640 | } | |
1641 | blkg = blkg->parent; | |
1642 | } | |
1643 | ||
1644 | if (!delay_nsec) | |
1645 | return; | |
1646 | ||
1647 | /* | |
1648 | * Let's not sleep for all eternity if we've amassed a huge delay. | |
1649 | * Swapping or metadata IO can accumulate 10's of seconds worth of | |
1650 | * delay, and we want userspace to be able to do _something_ so cap the | |
1651 | * delays at 0.25s. If there's 10's of seconds worth of delay then the | |
1652 | * tasks will be delayed for 0.25 second for every syscall. If | |
1653 | * blkcg_set_delay() was used as indicated by negative use_delay, the | |
1654 | * caller is responsible for regulating the range. | |
1655 | */ | |
1656 | if (clamp) | |
1657 | delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC); | |
1658 | ||
1659 | if (use_memdelay) | |
1660 | psi_memstall_enter(&pflags); | |
1661 | ||
1662 | exp = ktime_add_ns(now, delay_nsec); | |
1663 | tok = io_schedule_prepare(); | |
1664 | do { | |
1665 | __set_current_state(TASK_KILLABLE); | |
1666 | if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS)) | |
1667 | break; | |
1668 | } while (!fatal_signal_pending(current)); | |
1669 | io_schedule_finish(tok); | |
1670 | ||
1671 | if (use_memdelay) | |
1672 | psi_memstall_leave(&pflags); | |
1673 | } | |
1674 | ||
1675 | /** | |
1676 | * blkcg_maybe_throttle_current - throttle the current task if it has been marked | |
1677 | * | |
1678 | * This is only called if we've been marked with set_notify_resume(). Obviously | |
1679 | * we can be set_notify_resume() for reasons other than blkcg throttling, so we | |
1680 | * check to see if current->throttle_queue is set and if not this doesn't do | |
1681 | * anything. This should only ever be called by the resume code, it's not meant | |
1682 | * to be called by people willy-nilly as it will actually do the work to | |
1683 | * throttle the task if it is setup for throttling. | |
1684 | */ | |
1685 | void blkcg_maybe_throttle_current(void) | |
1686 | { | |
1687 | struct request_queue *q = current->throttle_queue; | |
1688 | struct cgroup_subsys_state *css; | |
1689 | struct blkcg *blkcg; | |
1690 | struct blkcg_gq *blkg; | |
1691 | bool use_memdelay = current->use_memdelay; | |
1692 | ||
1693 | if (!q) | |
1694 | return; | |
1695 | ||
1696 | current->throttle_queue = NULL; | |
1697 | current->use_memdelay = false; | |
1698 | ||
1699 | rcu_read_lock(); | |
1700 | css = kthread_blkcg(); | |
1701 | if (css) | |
1702 | blkcg = css_to_blkcg(css); | |
1703 | else | |
1704 | blkcg = css_to_blkcg(task_css(current, io_cgrp_id)); | |
1705 | ||
1706 | if (!blkcg) | |
1707 | goto out; | |
1708 | blkg = blkg_lookup(blkcg, q); | |
1709 | if (!blkg) | |
1710 | goto out; | |
1711 | if (!blkg_tryget(blkg)) | |
1712 | goto out; | |
1713 | rcu_read_unlock(); | |
1714 | ||
1715 | blkcg_maybe_throttle_blkg(blkg, use_memdelay); | |
1716 | blkg_put(blkg); | |
1717 | blk_put_queue(q); | |
1718 | return; | |
1719 | out: | |
1720 | rcu_read_unlock(); | |
1721 | blk_put_queue(q); | |
1722 | } | |
1723 | ||
1724 | /** | |
1725 | * blkcg_schedule_throttle - this task needs to check for throttling | |
1726 | * @q: the request queue IO was submitted on | |
1727 | * @use_memdelay: do we charge this to memory delay for PSI | |
1728 | * | |
1729 | * This is called by the IO controller when we know there's delay accumulated | |
1730 | * for the blkg for this task. We do not pass the blkg because there are places | |
1731 | * we call this that may not have that information, the swapping code for | |
1732 | * instance will only have a request_queue at that point. This set's the | |
1733 | * notify_resume for the task to check and see if it requires throttling before | |
1734 | * returning to user space. | |
1735 | * | |
1736 | * We will only schedule once per syscall. You can call this over and over | |
1737 | * again and it will only do the check once upon return to user space, and only | |
1738 | * throttle once. If the task needs to be throttled again it'll need to be | |
1739 | * re-set at the next time we see the task. | |
1740 | */ | |
1741 | void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay) | |
1742 | { | |
1743 | if (unlikely(current->flags & PF_KTHREAD)) | |
1744 | return; | |
1745 | ||
1746 | if (current->throttle_queue != q) { | |
1747 | if (!blk_get_queue(q)) | |
1748 | return; | |
1749 | ||
1750 | if (current->throttle_queue) | |
1751 | blk_put_queue(current->throttle_queue); | |
1752 | current->throttle_queue = q; | |
1753 | } | |
1754 | ||
1755 | if (use_memdelay) | |
1756 | current->use_memdelay = use_memdelay; | |
1757 | set_notify_resume(current); | |
1758 | } | |
1759 | ||
1760 | /** | |
1761 | * blkcg_add_delay - add delay to this blkg | |
1762 | * @blkg: blkg of interest | |
1763 | * @now: the current time in nanoseconds | |
1764 | * @delta: how many nanoseconds of delay to add | |
1765 | * | |
1766 | * Charge @delta to the blkg's current delay accumulation. This is used to | |
1767 | * throttle tasks if an IO controller thinks we need more throttling. | |
1768 | */ | |
1769 | void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta) | |
1770 | { | |
1771 | if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0)) | |
1772 | return; | |
1773 | blkcg_scale_delay(blkg, now); | |
1774 | atomic64_add(delta, &blkg->delay_nsec); | |
1775 | } | |
1776 | ||
1777 | /** | |
1778 | * blkg_tryget_closest - try and get a blkg ref on the closet blkg | |
1779 | * @bio: target bio | |
1780 | * @css: target css | |
1781 | * | |
1782 | * As the failure mode here is to walk up the blkg tree, this ensure that the | |
1783 | * blkg->parent pointers are always valid. This returns the blkg that it ended | |
1784 | * up taking a reference on or %NULL if no reference was taken. | |
1785 | */ | |
1786 | static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio, | |
1787 | struct cgroup_subsys_state *css) | |
1788 | { | |
1789 | struct blkcg_gq *blkg, *ret_blkg = NULL; | |
1790 | ||
1791 | rcu_read_lock(); | |
1792 | blkg = blkg_lookup_create(css_to_blkcg(css), | |
1793 | bio->bi_bdev->bd_disk->queue); | |
1794 | while (blkg) { | |
1795 | if (blkg_tryget(blkg)) { | |
1796 | ret_blkg = blkg; | |
1797 | break; | |
1798 | } | |
1799 | blkg = blkg->parent; | |
1800 | } | |
1801 | rcu_read_unlock(); | |
1802 | ||
1803 | return ret_blkg; | |
1804 | } | |
1805 | ||
1806 | /** | |
1807 | * bio_associate_blkg_from_css - associate a bio with a specified css | |
1808 | * @bio: target bio | |
1809 | * @css: target css | |
1810 | * | |
1811 | * Associate @bio with the blkg found by combining the css's blkg and the | |
1812 | * request_queue of the @bio. An association failure is handled by walking up | |
1813 | * the blkg tree. Therefore, the blkg associated can be anything between @blkg | |
1814 | * and q->root_blkg. This situation only happens when a cgroup is dying and | |
1815 | * then the remaining bios will spill to the closest alive blkg. | |
1816 | * | |
1817 | * A reference will be taken on the blkg and will be released when @bio is | |
1818 | * freed. | |
1819 | */ | |
1820 | void bio_associate_blkg_from_css(struct bio *bio, | |
1821 | struct cgroup_subsys_state *css) | |
1822 | { | |
1823 | if (bio->bi_blkg) | |
1824 | blkg_put(bio->bi_blkg); | |
1825 | ||
1826 | if (css && css->parent) { | |
1827 | bio->bi_blkg = blkg_tryget_closest(bio, css); | |
1828 | } else { | |
1829 | blkg_get(bio->bi_bdev->bd_disk->queue->root_blkg); | |
1830 | bio->bi_blkg = bio->bi_bdev->bd_disk->queue->root_blkg; | |
1831 | } | |
1832 | } | |
1833 | EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css); | |
1834 | ||
1835 | /** | |
1836 | * bio_associate_blkg - associate a bio with a blkg | |
1837 | * @bio: target bio | |
1838 | * | |
1839 | * Associate @bio with the blkg found from the bio's css and request_queue. | |
1840 | * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is | |
1841 | * already associated, the css is reused and association redone as the | |
1842 | * request_queue may have changed. | |
1843 | */ | |
1844 | void bio_associate_blkg(struct bio *bio) | |
1845 | { | |
1846 | struct cgroup_subsys_state *css; | |
1847 | ||
1848 | rcu_read_lock(); | |
1849 | ||
1850 | if (bio->bi_blkg) | |
1851 | css = &bio_blkcg(bio)->css; | |
1852 | else | |
1853 | css = blkcg_css(); | |
1854 | ||
1855 | bio_associate_blkg_from_css(bio, css); | |
1856 | ||
1857 | rcu_read_unlock(); | |
1858 | } | |
1859 | EXPORT_SYMBOL_GPL(bio_associate_blkg); | |
1860 | ||
1861 | /** | |
1862 | * bio_clone_blkg_association - clone blkg association from src to dst bio | |
1863 | * @dst: destination bio | |
1864 | * @src: source bio | |
1865 | */ | |
1866 | void bio_clone_blkg_association(struct bio *dst, struct bio *src) | |
1867 | { | |
1868 | if (src->bi_blkg) { | |
1869 | if (dst->bi_blkg) | |
1870 | blkg_put(dst->bi_blkg); | |
1871 | blkg_get(src->bi_blkg); | |
1872 | dst->bi_blkg = src->bi_blkg; | |
1873 | } | |
1874 | } | |
1875 | EXPORT_SYMBOL_GPL(bio_clone_blkg_association); | |
1876 | ||
1877 | static int blk_cgroup_io_type(struct bio *bio) | |
1878 | { | |
1879 | if (op_is_discard(bio->bi_opf)) | |
1880 | return BLKG_IOSTAT_DISCARD; | |
1881 | if (op_is_write(bio->bi_opf)) | |
1882 | return BLKG_IOSTAT_WRITE; | |
1883 | return BLKG_IOSTAT_READ; | |
1884 | } | |
1885 | ||
1886 | void blk_cgroup_bio_start(struct bio *bio) | |
1887 | { | |
1888 | int rwd = blk_cgroup_io_type(bio), cpu; | |
1889 | struct blkg_iostat_set *bis; | |
1890 | ||
1891 | cpu = get_cpu(); | |
1892 | bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu); | |
1893 | u64_stats_update_begin(&bis->sync); | |
1894 | ||
1895 | /* | |
1896 | * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split | |
1897 | * bio and we would have already accounted for the size of the bio. | |
1898 | */ | |
1899 | if (!bio_flagged(bio, BIO_CGROUP_ACCT)) { | |
1900 | bio_set_flag(bio, BIO_CGROUP_ACCT); | |
1901 | bis->cur.bytes[rwd] += bio->bi_iter.bi_size; | |
1902 | } | |
1903 | bis->cur.ios[rwd]++; | |
1904 | ||
1905 | u64_stats_update_end(&bis->sync); | |
1906 | if (cgroup_subsys_on_dfl(io_cgrp_subsys)) | |
1907 | cgroup_rstat_updated(bio->bi_blkg->blkcg->css.cgroup, cpu); | |
1908 | put_cpu(); | |
1909 | } | |
1910 | ||
1911 | static int __init blkcg_init(void) | |
1912 | { | |
1913 | blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio", | |
1914 | WQ_MEM_RECLAIM | WQ_FREEZABLE | | |
1915 | WQ_UNBOUND | WQ_SYSFS, 0); | |
1916 | if (!blkcg_punt_bio_wq) | |
1917 | return -ENOMEM; | |
1918 | return 0; | |
1919 | } | |
1920 | subsys_initcall(blkcg_init); | |
1921 | ||
1922 | module_param(blkcg_debug_stats, bool, 0644); | |
1923 | MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not"); |