~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/block/blk-cgroup.c

Version: ~ [ linux-5.8 ] ~ [ linux-5.7.14 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.57 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.138 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.193 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.232 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.232 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.85 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

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

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp