block/blk-cgroup.c

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