block/blk-cgroup.c

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