mm/backing-dev.c

   1
   2 #include <linux/wait.h>
   3 #include <linux/backing-dev.h>
   4 #include <linux/kthread.h>
   5 #include <linux/freezer.h>
   6 #include <linux/fs.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/module.h>
  11 #include <linux/writeback.h>
  12 #include <linux/device.h>
  13 #include <trace/events/writeback.h>
  14
  15 static atomic_long_t bdi_seq = ATOMIC_LONG_INIT(0);
  16
  17 void default_unplug_io_fn(struct backing_dev_info *bdi, struct page *page)
  18 {
  19 }
  20 EXPORT_SYMBOL(default_unplug_io_fn);
  21
  22 struct backing_dev_info default_backing_dev_info = {
  23         .name           = "default",
  24         .ra_pages       = VM_MAX_READAHEAD * 1024 / PAGE_CACHE_SIZE,
  25         .state          = 0,
  26         .capabilities   = BDI_CAP_MAP_COPY,
  27         .unplug_io_fn   = default_unplug_io_fn,
  28 };
  29 EXPORT_SYMBOL_GPL(default_backing_dev_info);
  30
  31 struct backing_dev_info noop_backing_dev_info = {
  32         .name           = "noop",
  33 };
  34 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
  35
  36 static struct class *bdi_class;
  37
  38 /*
  39  * bdi_lock protects updates to bdi_list and bdi_pending_list, as well as
  40  * reader side protection for bdi_pending_list. bdi_list has RCU reader side
  41  * locking.
  42  */
  43 DEFINE_SPINLOCK(bdi_lock);
  44 LIST_HEAD(bdi_list);
  45 LIST_HEAD(bdi_pending_list);
  46
  47 static struct task_struct *sync_supers_tsk;
  48 static struct timer_list sync_supers_timer;
  49
  50 static int bdi_sync_supers(void *);
  51 static void sync_supers_timer_fn(unsigned long);
  52
  53 #ifdef CONFIG_DEBUG_FS
  54 #include <linux/debugfs.h>
  55 #include <linux/seq_file.h>
  56
  57 static struct dentry *bdi_debug_root;
  58
  59 static void bdi_debug_init(void)
  60 {
  61         bdi_debug_root = debugfs_create_dir("bdi", NULL);
  62 }
  63
  64 static int bdi_debug_stats_show(struct seq_file *m, void *v)
  65 {
  66         struct backing_dev_info *bdi = m->private;
  67         struct bdi_writeback *wb = &bdi->wb;
  68         unsigned long background_thresh;
  69         unsigned long dirty_thresh;
  70         unsigned long bdi_thresh;
  71         unsigned long nr_dirty, nr_io, nr_more_io, nr_wb;
  72         struct inode *inode;
  73
  74         nr_wb = nr_dirty = nr_io = nr_more_io = 0;
  75         spin_lock(&inode_lock);
  76         list_for_each_entry(inode, &wb->b_dirty, i_list)
  77                 nr_dirty++;
  78         list_for_each_entry(inode, &wb->b_io, i_list)
  79                 nr_io++;
  80         list_for_each_entry(inode, &wb->b_more_io, i_list)
  81                 nr_more_io++;
  82         spin_unlock(&inode_lock);
  83
  84         global_dirty_limits(&background_thresh, &dirty_thresh);
  85         bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
  86
  87 #define K(x) ((x) << (PAGE_SHIFT - 10))
  88         seq_printf(m,
  89                    "BdiWriteback:     %8lu kB\n"
  90                    "BdiReclaimable:   %8lu kB\n"
  91                    "BdiDirtyThresh:   %8lu kB\n"
  92                    "DirtyThresh:      %8lu kB\n"
  93                    "BackgroundThresh: %8lu kB\n"
  94                    "b_dirty:          %8lu\n"
  95                    "b_io:             %8lu\n"
  96                    "b_more_io:        %8lu\n"
  97                    "bdi_list:         %8u\n"
  98                    "state:            %8lx\n",
  99                    (unsigned long) K(bdi_stat(bdi, BDI_WRITEBACK)),
 100                    (unsigned long) K(bdi_stat(bdi, BDI_RECLAIMABLE)),
 101                    K(bdi_thresh), K(dirty_thresh),
 102                    K(background_thresh), nr_dirty, nr_io, nr_more_io,
 103                    !list_empty(&bdi->bdi_list), bdi->state);
 104 #undef K
 105
 106         return 0;
 107 }
 108
 109 static int bdi_debug_stats_open(struct inode *inode, struct file *file)
 110 {
 111         return single_open(file, bdi_debug_stats_show, inode->i_private);
 112 }
 113
 114 static const struct file_operations bdi_debug_stats_fops = {
 115         .open           = bdi_debug_stats_open,
 116         .read           = seq_read,
 117         .llseek         = seq_lseek,
 118         .release        = single_release,
 119 };
 120
 121 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
 122 {
 123         bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
 124         bdi->debug_stats = debugfs_create_file("stats", 0444, bdi->debug_dir,
 125                                                bdi, &bdi_debug_stats_fops);
 126 }
 127
 128 static void bdi_debug_unregister(struct backing_dev_info *bdi)
 129 {
 130         debugfs_remove(bdi->debug_stats);
 131         debugfs_remove(bdi->debug_dir);
 132 }
 133 #else
 134 static inline void bdi_debug_init(void)
 135 {
 136 }
 137 static inline void bdi_debug_register(struct backing_dev_info *bdi,
 138                                       const char *name)
 139 {
 140 }
 141 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
 142 {
 143 }
 144 #endif
 145
 146 static ssize_t read_ahead_kb_store(struct device *dev,
 147                                   struct device_attribute *attr,
 148                                   const char *buf, size_t count)
 149 {
 150         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 151         char *end;
 152         unsigned long read_ahead_kb;
 153         ssize_t ret = -EINVAL;
 154
 155         read_ahead_kb = simple_strtoul(buf, &end, 10);
 156         if (*buf && (end[0] == '\0' || (end[0] == '\n' && end[1] == '\0'))) {
 157                 bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
 158                 ret = count;
 159         }
 160         return ret;
 161 }
 162
 163 #define K(pages) ((pages) << (PAGE_SHIFT - 10))
 164
 165 #define BDI_SHOW(name, expr)                                            \
 166 static ssize_t name##_show(struct device *dev,                          \
 167                            struct device_attribute *attr, char *page)   \
 168 {                                                                       \
 169         struct backing_dev_info *bdi = dev_get_drvdata(dev);            \
 170                                                                         \
 171         return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr);  \
 172 }
 173
 174 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
 175
 176 static ssize_t min_ratio_store(struct device *dev,
 177                 struct device_attribute *attr, const char *buf, size_t count)
 178 {
 179         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 180         char *end;
 181         unsigned int ratio;
 182         ssize_t ret = -EINVAL;
 183
 184         ratio = simple_strtoul(buf, &end, 10);
 185         if (*buf && (end[0] == '\0' || (end[0] == '\n' && end[1] == '\0'))) {
 186                 ret = bdi_set_min_ratio(bdi, ratio);
 187                 if (!ret)
 188                         ret = count;
 189         }
 190         return ret;
 191 }
 192 BDI_SHOW(min_ratio, bdi->min_ratio)
 193
 194 static ssize_t max_ratio_store(struct device *dev,
 195                 struct device_attribute *attr, const char *buf, size_t count)
 196 {
 197         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 198         char *end;
 199         unsigned int ratio;
 200         ssize_t ret = -EINVAL;
 201
 202         ratio = simple_strtoul(buf, &end, 10);
 203         if (*buf && (end[0] == '\0' || (end[0] == '\n' && end[1] == '\0'))) {
 204                 ret = bdi_set_max_ratio(bdi, ratio);
 205                 if (!ret)
 206                         ret = count;
 207         }
 208         return ret;
 209 }
 210 BDI_SHOW(max_ratio, bdi->max_ratio)
 211
 212 #define __ATTR_RW(attr) __ATTR(attr, 0644, attr##_show, attr##_store)
 213
 214 static struct device_attribute bdi_dev_attrs[] = {
 215         __ATTR_RW(read_ahead_kb),
 216         __ATTR_RW(min_ratio),
 217         __ATTR_RW(max_ratio),
 218         __ATTR_NULL,
 219 };
 220
 221 static __init int bdi_class_init(void)
 222 {
 223         bdi_class = class_create(THIS_MODULE, "bdi");
 224         if (IS_ERR(bdi_class))
 225                 return PTR_ERR(bdi_class);
 226
 227         bdi_class->dev_attrs = bdi_dev_attrs;
 228         bdi_debug_init();
 229         return 0;
 230 }
 231 postcore_initcall(bdi_class_init);
 232
 233 static int __init default_bdi_init(void)
 234 {
 235         int err;
 236
 237         sync_supers_tsk = kthread_run(bdi_sync_supers, NULL, "sync_supers");
 238         BUG_ON(IS_ERR(sync_supers_tsk));
 239
 240         setup_timer(&sync_supers_timer, sync_supers_timer_fn, 0);
 241         bdi_arm_supers_timer();
 242
 243         err = bdi_init(&default_backing_dev_info);
 244         if (!err)
 245                 bdi_register(&default_backing_dev_info, NULL, "default");
 246
 247         return err;
 248 }
 249 subsys_initcall(default_bdi_init);
 250
 251 int bdi_has_dirty_io(struct backing_dev_info *bdi)
 252 {
 253         return wb_has_dirty_io(&bdi->wb);
 254 }
 255
 256 static void bdi_flush_io(struct backing_dev_info *bdi)
 257 {
 258         struct writeback_control wbc = {
 259                 .sync_mode              = WB_SYNC_NONE,
 260                 .older_than_this        = NULL,
 261                 .range_cyclic           = 1,
 262                 .nr_to_write            = 1024,
 263         };
 264
 265         writeback_inodes_wb(&bdi->wb, &wbc);
 266 }
 267
 268 /*
 269  * kupdated() used to do this. We cannot do it from the bdi_forker_thread()
 270  * or we risk deadlocking on ->s_umount. The longer term solution would be
 271  * to implement sync_supers_bdi() or similar and simply do it from the
 272  * bdi writeback thread individually.
 273  */
 274 static int bdi_sync_supers(void *unused)
 275 {
 276         set_user_nice(current, 0);
 277
 278         while (!kthread_should_stop()) {
 279                 set_current_state(TASK_INTERRUPTIBLE);
 280                 schedule();
 281
 282                 /*
 283                  * Do this periodically, like kupdated() did before.
 284                  */
 285                 sync_supers();
 286         }
 287
 288         return 0;
 289 }
 290
 291 void bdi_arm_supers_timer(void)
 292 {
 293         unsigned long next;
 294
 295         if (!dirty_writeback_interval)
 296                 return;
 297
 298         next = msecs_to_jiffies(dirty_writeback_interval * 10) + jiffies;
 299         mod_timer(&sync_supers_timer, round_jiffies_up(next));
 300 }
 301
 302 static void sync_supers_timer_fn(unsigned long unused)
 303 {
 304         wake_up_process(sync_supers_tsk);
 305         bdi_arm_supers_timer();
 306 }
 307
 308 static void wakeup_timer_fn(unsigned long data)
 309 {
 310         struct backing_dev_info *bdi = (struct backing_dev_info *)data;
 311
 312         spin_lock_bh(&bdi->wb_lock);
 313         if (bdi->wb.task) {
 314                 trace_writeback_wake_thread(bdi);
 315                 wake_up_process(bdi->wb.task);
 316         } else {
 317                 /*
 318                  * When bdi tasks are inactive for long time, they are killed.
 319                  * In this case we have to wake-up the forker thread which
 320                  * should create and run the bdi thread.
 321                  */
 322                 trace_writeback_wake_forker_thread(bdi);
 323                 wake_up_process(default_backing_dev_info.wb.task);
 324         }
 325         spin_unlock_bh(&bdi->wb_lock);
 326 }
 327
 328 /*
 329  * This function is used when the first inode for this bdi is marked dirty. It
 330  * wakes-up the corresponding bdi thread which should then take care of the
 331  * periodic background write-out of dirty inodes. Since the write-out would
 332  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 333  * set up a timer which wakes the bdi thread up later.
 334  *
 335  * Note, we wouldn't bother setting up the timer, but this function is on the
 336  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 337  * by delaying the wake-up.
 338  */
 339 void bdi_wakeup_thread_delayed(struct backing_dev_info *bdi)
 340 {
 341         unsigned long timeout;
 342
 343         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 344         mod_timer(&bdi->wb.wakeup_timer, jiffies + timeout);
 345 }
 346
 347 /*
 348  * Calculate the longest interval (jiffies) bdi threads are allowed to be
 349  * inactive.
 350  */
 351 static unsigned long bdi_longest_inactive(void)
 352 {
 353         unsigned long interval;
 354
 355         interval = msecs_to_jiffies(dirty_writeback_interval * 10);
 356         return max(5UL * 60 * HZ, interval);
 357 }
 358
 359 static int bdi_forker_thread(void *ptr)
 360 {
 361         struct bdi_writeback *me = ptr;
 362
 363         current->flags |= PF_FLUSHER | PF_SWAPWRITE;
 364         set_freezable();
 365
 366         /*
 367          * Our parent may run at a different priority, just set us to normal
 368          */
 369         set_user_nice(current, 0);
 370
 371         for (;;) {
 372                 struct task_struct *task = NULL;
 373                 struct backing_dev_info *bdi;
 374                 enum {
 375                         NO_ACTION,   /* Nothing to do */
 376                         FORK_THREAD, /* Fork bdi thread */
 377                         KILL_THREAD, /* Kill inactive bdi thread */
 378                 } action = NO_ACTION;
 379
 380                 /*
 381                  * Temporary measure, we want to make sure we don't see
 382                  * dirty data on the default backing_dev_info
 383                  */
 384                 if (wb_has_dirty_io(me) || !list_empty(&me->bdi->work_list)) {
 385                         del_timer(&me->wakeup_timer);
 386                         wb_do_writeback(me, 0);
 387                 }
 388
 389                 spin_lock_bh(&bdi_lock);
 390                 set_current_state(TASK_INTERRUPTIBLE);
 391
 392                 list_for_each_entry(bdi, &bdi_list, bdi_list) {
 393                         bool have_dirty_io;
 394
 395                         if (!bdi_cap_writeback_dirty(bdi) ||
 396                              bdi_cap_flush_forker(bdi))
 397                                 continue;
 398
 399                         WARN(!test_bit(BDI_registered, &bdi->state),
 400                              "bdi %p/%s is not registered!\n", bdi, bdi->name);
 401
 402                         have_dirty_io = !list_empty(&bdi->work_list) ||
 403                                         wb_has_dirty_io(&bdi->wb);
 404
 405                         /*
 406                          * If the bdi has work to do, but the thread does not
 407                          * exist - create it.
 408                          */
 409                         if (!bdi->wb.task && have_dirty_io) {
 410                                 /*
 411                                  * Set the pending bit - if someone will try to
 412                                  * unregister this bdi - it'll wait on this bit.
 413                                  */
 414                                 set_bit(BDI_pending, &bdi->state);
 415                                 action = FORK_THREAD;
 416                                 break;
 417                         }
 418
 419                         spin_lock(&bdi->wb_lock);
 420
 421                         /*
 422                          * If there is no work to do and the bdi thread was
 423                          * inactive long enough - kill it. The wb_lock is taken
 424                          * to make sure no-one adds more work to this bdi and
 425                          * wakes the bdi thread up.
 426                          */
 427                         if (bdi->wb.task && !have_dirty_io &&
 428                             time_after(jiffies, bdi->wb.last_active +
 429                                                 bdi_longest_inactive())) {
 430                                 task = bdi->wb.task;
 431                                 bdi->wb.task = NULL;
 432                                 spin_unlock(&bdi->wb_lock);
 433                                 set_bit(BDI_pending, &bdi->state);
 434                                 action = KILL_THREAD;
 435                                 break;
 436                         }
 437                         spin_unlock(&bdi->wb_lock);
 438                 }
 439                 spin_unlock_bh(&bdi_lock);
 440
 441                 /* Keep working if default bdi still has things to do */
 442                 if (!list_empty(&me->bdi->work_list))
 443                         __set_current_state(TASK_RUNNING);
 444
 445                 switch (action) {
 446                 case FORK_THREAD:
 447                         __set_current_state(TASK_RUNNING);
 448                         task = kthread_run(bdi_writeback_thread, &bdi->wb, "flush-%s",
 449                                            dev_name(bdi->dev));
 450                         if (IS_ERR(task)) {
 451                                 /*
 452                                  * If thread creation fails, force writeout of
 453                                  * the bdi from the thread.
 454                                  */
 455                                 bdi_flush_io(bdi);
 456                         } else {
 457                                 /*
 458                                  * The spinlock makes sure we do not lose
 459                                  * wake-ups when racing with 'bdi_queue_work()'.
 460                                  */
 461                                 spin_lock_bh(&bdi->wb_lock);
 462                                 bdi->wb.task = task;
 463                                 spin_unlock_bh(&bdi->wb_lock);
 464                         }
 465                         break;
 466
 467                 case KILL_THREAD:
 468                         __set_current_state(TASK_RUNNING);
 469                         kthread_stop(task);
 470                         break;
 471
 472                 case NO_ACTION:
 473                         if (!wb_has_dirty_io(me) || !dirty_writeback_interval)
 474                                 /*
 475                                  * There are no dirty data. The only thing we
 476                                  * should now care about is checking for
 477                                  * inactive bdi threads and killing them. Thus,
 478                                  * let's sleep for longer time, save energy and
 479                                  * be friendly for battery-driven devices.
 480                                  */
 481                                 schedule_timeout(bdi_longest_inactive());
 482                         else
 483                                 schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
 484                         try_to_freeze();
 485                         /* Back to the main loop */
 486                         continue;
 487                 }
 488
 489                 /*
 490                  * Clear pending bit and wakeup anybody waiting to tear us down.
 491                  */
 492                 clear_bit(BDI_pending, &bdi->state);
 493                 smp_mb__after_clear_bit();
 494                 wake_up_bit(&bdi->state, BDI_pending);
 495         }
 496
 497         return 0;
 498 }
 499
 500 /*
 501  * Remove bdi from bdi_list, and ensure that it is no longer visible
 502  */
 503 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 504 {
 505         spin_lock_bh(&bdi_lock);
 506         list_del_rcu(&bdi->bdi_list);
 507         spin_unlock_bh(&bdi_lock);
 508
 509         synchronize_rcu();
 510 }
 511
 512 int bdi_register(struct backing_dev_info *bdi, struct device *parent,
 513                 const char *fmt, ...)
 514 {
 515         va_list args;
 516         struct device *dev;
 517
 518         if (bdi->dev)   /* The driver needs to use separate queues per device */
 519                 return 0;
 520
 521         va_start(args, fmt);
 522         dev = device_create_vargs(bdi_class, parent, MKDEV(0, 0), bdi, fmt, args);
 523         va_end(args);
 524         if (IS_ERR(dev))
 525                 return PTR_ERR(dev);
 526
 527         bdi->dev = dev;
 528
 529         /*
 530          * Just start the forker thread for our default backing_dev_info,
 531          * and add other bdi's to the list. They will get a thread created
 532          * on-demand when they need it.
 533          */
 534         if (bdi_cap_flush_forker(bdi)) {
 535                 struct bdi_writeback *wb = &bdi->wb;
 536
 537                 wb->task = kthread_run(bdi_forker_thread, wb, "bdi-%s",
 538                                                 dev_name(dev));
 539                 if (IS_ERR(wb->task))
 540                         return PTR_ERR(wb->task);
 541         }
 542
 543         bdi_debug_register(bdi, dev_name(dev));
 544         set_bit(BDI_registered, &bdi->state);
 545
 546         spin_lock_bh(&bdi_lock);
 547         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 548         spin_unlock_bh(&bdi_lock);
 549
 550         trace_writeback_bdi_register(bdi);
 551         return 0;
 552 }
 553 EXPORT_SYMBOL(bdi_register);
 554
 555 int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev)
 556 {
 557         return bdi_register(bdi, NULL, "%u:%u", MAJOR(dev), MINOR(dev));
 558 }
 559 EXPORT_SYMBOL(bdi_register_dev);
 560
 561 /*
 562  * Remove bdi from the global list and shutdown any threads we have running
 563  */
 564 static void bdi_wb_shutdown(struct backing_dev_info *bdi)
 565 {
 566         if (!bdi_cap_writeback_dirty(bdi))
 567                 return;
 568
 569         /*
 570          * Make sure nobody finds us on the bdi_list anymore
 571          */
 572         bdi_remove_from_list(bdi);
 573
 574         /*
 575          * If setup is pending, wait for that to complete first
 576          */
 577         wait_on_bit(&bdi->state, BDI_pending, bdi_sched_wait,
 578                         TASK_UNINTERRUPTIBLE);
 579
 580         /*
 581          * Finally, kill the kernel thread. We don't need to be RCU
 582          * safe anymore, since the bdi is gone from visibility. Force
 583          * unfreeze of the thread before calling kthread_stop(), otherwise
 584          * it would never exet if it is currently stuck in the refrigerator.
 585          */
 586         if (bdi->wb.task) {
 587                 thaw_process(bdi->wb.task);
 588                 kthread_stop(bdi->wb.task);
 589         }
 590 }
 591
 592 /*
 593  * This bdi is going away now, make sure that no super_blocks point to it
 594  */
 595 static void bdi_prune_sb(struct backing_dev_info *bdi)
 596 {
 597         struct super_block *sb;
 598
 599         spin_lock(&sb_lock);
 600         list_for_each_entry(sb, &super_blocks, s_list) {
 601                 if (sb->s_bdi == bdi)
 602                         sb->s_bdi = NULL;
 603         }
 604         spin_unlock(&sb_lock);
 605 }
 606
 607 void bdi_unregister(struct backing_dev_info *bdi)
 608 {
 609         if (bdi->dev) {
 610                 trace_writeback_bdi_unregister(bdi);
 611                 bdi_prune_sb(bdi);
 612                 del_timer_sync(&bdi->wb.wakeup_timer);
 613
 614                 if (!bdi_cap_flush_forker(bdi))
 615                         bdi_wb_shutdown(bdi);
 616                 bdi_debug_unregister(bdi);
 617                 device_unregister(bdi->dev);
 618                 bdi->dev = NULL;
 619         }
 620 }
 621 EXPORT_SYMBOL(bdi_unregister);
 622
 623 static void bdi_wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi)
 624 {
 625         memset(wb, 0, sizeof(*wb));
 626
 627         wb->bdi = bdi;
 628         wb->last_old_flush = jiffies;
 629         INIT_LIST_HEAD(&wb->b_dirty);
 630         INIT_LIST_HEAD(&wb->b_io);
 631         INIT_LIST_HEAD(&wb->b_more_io);
 632         setup_timer(&wb->wakeup_timer, wakeup_timer_fn, (unsigned long)bdi);
 633 }
 634
 635 int bdi_init(struct backing_dev_info *bdi)
 636 {
 637         int i, err;
 638
 639         bdi->dev = NULL;
 640
 641         bdi->min_ratio = 0;
 642         bdi->max_ratio = 100;
 643         bdi->max_prop_frac = PROP_FRAC_BASE;
 644         spin_lock_init(&bdi->wb_lock);
 645         INIT_LIST_HEAD(&bdi->bdi_list);
 646         INIT_LIST_HEAD(&bdi->work_list);
 647
 648         bdi_wb_init(&bdi->wb, bdi);
 649
 650         for (i = 0; i < NR_BDI_STAT_ITEMS; i++) {
 651                 err = percpu_counter_init(&bdi->bdi_stat[i], 0);
 652                 if (err)
 653                         goto err;
 654         }
 655
 656         bdi->dirty_exceeded = 0;
 657         err = prop_local_init_percpu(&bdi->completions);
 658
 659         if (err) {
 660 err:
 661                 while (i--)
 662                         percpu_counter_destroy(&bdi->bdi_stat[i]);
 663         }
 664
 665         return err;
 666 }
 667 EXPORT_SYMBOL(bdi_init);
 668
 669 void bdi_destroy(struct backing_dev_info *bdi)
 670 {
 671         int i;
 672
 673         /*
 674          * Splice our entries to the default_backing_dev_info, if this
 675          * bdi disappears
 676          */
 677         if (bdi_has_dirty_io(bdi)) {
 678                 struct bdi_writeback *dst = &default_backing_dev_info.wb;
 679
 680                 spin_lock(&inode_lock);
 681                 list_splice(&bdi->wb.b_dirty, &dst->b_dirty);
 682                 list_splice(&bdi->wb.b_io, &dst->b_io);
 683                 list_splice(&bdi->wb.b_more_io, &dst->b_more_io);
 684                 spin_unlock(&inode_lock);
 685         }
 686
 687         bdi_unregister(bdi);
 688
 689         for (i = 0; i < NR_BDI_STAT_ITEMS; i++)
 690                 percpu_counter_destroy(&bdi->bdi_stat[i]);
 691
 692         prop_local_destroy_percpu(&bdi->completions);
 693 }
 694 EXPORT_SYMBOL(bdi_destroy);
 695
 696 /*
 697  * For use from filesystems to quickly init and register a bdi associated
 698  * with dirty writeback
 699  */
 700 int bdi_setup_and_register(struct backing_dev_info *bdi, char *name,
 701                            unsigned int cap)
 702 {
 703         char tmp[32];
 704         int err;
 705
 706         bdi->name = name;
 707         bdi->capabilities = cap;
 708         err = bdi_init(bdi);
 709         if (err)
 710                 return err;
 711
 712         sprintf(tmp, "%.28s%s", name, "-%d");
 713         err = bdi_register(bdi, NULL, tmp, atomic_long_inc_return(&bdi_seq));
 714         if (err) {
 715                 bdi_destroy(bdi);
 716                 return err;
 717         }
 718
 719         return 0;
 720 }
 721 EXPORT_SYMBOL(bdi_setup_and_register);
 722
 723 static wait_queue_head_t congestion_wqh[2] = {
 724                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
 725                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
 726         };
 727
 728 void clear_bdi_congested(struct backing_dev_info *bdi, int sync)
 729 {
 730         enum bdi_state bit;
 731         wait_queue_head_t *wqh = &congestion_wqh[sync];
 732
 733         bit = sync ? BDI_sync_congested : BDI_async_congested;
 734         clear_bit(bit, &bdi->state);
 735         smp_mb__after_clear_bit();
 736         if (waitqueue_active(wqh))
 737                 wake_up(wqh);
 738 }
 739 EXPORT_SYMBOL(clear_bdi_congested);
 740
 741 void set_bdi_congested(struct backing_dev_info *bdi, int sync)
 742 {
 743         enum bdi_state bit;
 744
 745         bit = sync ? BDI_sync_congested : BDI_async_congested;
 746         set_bit(bit, &bdi->state);
 747 }
 748 EXPORT_SYMBOL(set_bdi_congested);
 749
 750 /**
 751  * congestion_wait - wait for a backing_dev to become uncongested
 752  * @sync: SYNC or ASYNC IO
 753  * @timeout: timeout in jiffies
 754  *
 755  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
 756  * write congestion.  If no backing_devs are congested then just wait for the
 757  * next write to be completed.
 758  */
 759 long congestion_wait(int sync, long timeout)
 760 {
 761         long ret;
 762         DEFINE_WAIT(wait);
 763         wait_queue_head_t *wqh = &congestion_wqh[sync];
 764
 765         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
 766         ret = io_schedule_timeout(timeout);
 767         finish_wait(wqh, &wait);
 768         return ret;
 769 }
 770 EXPORT_SYMBOL(congestion_wait);
 771