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         get_dirty_limits(&background_thresh, &dirty_thresh, &bdi_thresh, bdi);
  85
  86 #define K(x) ((x) << (PAGE_SHIFT - 10))
  87         seq_printf(m,
  88                    "BdiWriteback:     %8lu kB\n"
  89                    "BdiReclaimable:   %8lu kB\n"
  90                    "BdiDirtyThresh:   %8lu kB\n"
  91                    "DirtyThresh:      %8lu kB\n"
  92                    "BackgroundThresh: %8lu kB\n"
  93                    "b_dirty:          %8lu\n"
  94                    "b_io:             %8lu\n"
  95                    "b_more_io:        %8lu\n"
  96                    "bdi_list:         %8u\n"
  97                    "state:            %8lx\n",
  98                    (unsigned long) K(bdi_stat(bdi, BDI_WRITEBACK)),
  99                    (unsigned long) K(bdi_stat(bdi, BDI_RECLAIMABLE)),
 100                    K(bdi_thresh), K(dirty_thresh),
 101                    K(background_thresh), nr_dirty, nr_io, nr_more_io,
 102                    !list_empty(&bdi->bdi_list), bdi->state);
 103 #undef K
 104
 105         return 0;
 106 }
 107
 108 static int bdi_debug_stats_open(struct inode *inode, struct file *file)
 109 {
 110         return single_open(file, bdi_debug_stats_show, inode->i_private);
 111 }
 112
 113 static const struct file_operations bdi_debug_stats_fops = {
 114         .open           = bdi_debug_stats_open,
 115         .read           = seq_read,
 116         .llseek         = seq_lseek,
 117         .release        = single_release,
 118 };
 119
 120 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
 121 {
 122         bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
 123         bdi->debug_stats = debugfs_create_file("stats", 0444, bdi->debug_dir,
 124                                                bdi, &bdi_debug_stats_fops);
 125 }
 126
 127 static void bdi_debug_unregister(struct backing_dev_info *bdi)
 128 {
 129         debugfs_remove(bdi->debug_stats);
 130         debugfs_remove(bdi->debug_dir);
 131 }
 132 #else
 133 static inline void bdi_debug_init(void)
 134 {
 135 }
 136 static inline void bdi_debug_register(struct backing_dev_info *bdi,
 137                                       const char *name)
 138 {
 139 }
 140 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
 141 {
 142 }
 143 #endif
 144
 145 static ssize_t read_ahead_kb_store(struct device *dev,
 146                                   struct device_attribute *attr,
 147                                   const char *buf, size_t count)
 148 {
 149         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 150         char *end;
 151         unsigned long read_ahead_kb;
 152         ssize_t ret = -EINVAL;
 153
 154         read_ahead_kb = simple_strtoul(buf, &end, 10);
 155         if (*buf && (end[0] == '\0' || (end[0] == '\n' && end[1] == '\0'))) {
 156                 bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
 157                 ret = count;
 158         }
 159         return ret;
 160 }
 161
 162 #define K(pages) ((pages) << (PAGE_SHIFT - 10))
 163
 164 #define BDI_SHOW(name, expr)                                            \
 165 static ssize_t name##_show(struct device *dev,                          \
 166                            struct device_attribute *attr, char *page)   \
 167 {                                                                       \
 168         struct backing_dev_info *bdi = dev_get_drvdata(dev);            \
 169                                                                         \
 170         return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr);  \
 171 }
 172
 173 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
 174
 175 static ssize_t min_ratio_store(struct device *dev,
 176                 struct device_attribute *attr, const char *buf, size_t count)
 177 {
 178         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 179         char *end;
 180         unsigned int ratio;
 181         ssize_t ret = -EINVAL;
 182
 183         ratio = simple_strtoul(buf, &end, 10);
 184         if (*buf && (end[0] == '\0' || (end[0] == '\n' && end[1] == '\0'))) {
 185                 ret = bdi_set_min_ratio(bdi, ratio);
 186                 if (!ret)
 187                         ret = count;
 188         }
 189         return ret;
 190 }
 191 BDI_SHOW(min_ratio, bdi->min_ratio)
 192
 193 static ssize_t max_ratio_store(struct device *dev,
 194                 struct device_attribute *attr, const char *buf, size_t count)
 195 {
 196         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 197         char *end;
 198         unsigned int ratio;
 199         ssize_t ret = -EINVAL;
 200
 201         ratio = simple_strtoul(buf, &end, 10);
 202         if (*buf && (end[0] == '\0' || (end[0] == '\n' && end[1] == '\0'))) {
 203                 ret = bdi_set_max_ratio(bdi, ratio);
 204                 if (!ret)
 205                         ret = count;
 206         }
 207         return ret;
 208 }
 209 BDI_SHOW(max_ratio, bdi->max_ratio)
 210
 211 #define __ATTR_RW(attr) __ATTR(attr, 0644, attr##_show, attr##_store)
 212
 213 static struct device_attribute bdi_dev_attrs[] = {
 214         __ATTR_RW(read_ahead_kb),
 215         __ATTR_RW(min_ratio),
 216         __ATTR_RW(max_ratio),
 217         __ATTR_NULL,
 218 };
 219
 220 static __init int bdi_class_init(void)
 221 {
 222         bdi_class = class_create(THIS_MODULE, "bdi");
 223         if (IS_ERR(bdi_class))
 224                 return PTR_ERR(bdi_class);
 225
 226         bdi_class->dev_attrs = bdi_dev_attrs;
 227         bdi_debug_init();
 228         return 0;
 229 }
 230 postcore_initcall(bdi_class_init);
 231
 232 static int __init default_bdi_init(void)
 233 {
 234         int err;
 235
 236         sync_supers_tsk = kthread_run(bdi_sync_supers, NULL, "sync_supers");
 237         BUG_ON(IS_ERR(sync_supers_tsk));
 238
 239         setup_timer(&sync_supers_timer, sync_supers_timer_fn, 0);
 240         bdi_arm_supers_timer();
 241
 242         err = bdi_init(&default_backing_dev_info);
 243         if (!err)
 244                 bdi_register(&default_backing_dev_info, NULL, "default");
 245
 246         return err;
 247 }
 248 subsys_initcall(default_bdi_init);
 249
 250 int bdi_has_dirty_io(struct backing_dev_info *bdi)
 251 {
 252         return wb_has_dirty_io(&bdi->wb);
 253 }
 254
 255 static void bdi_flush_io(struct backing_dev_info *bdi)
 256 {
 257         struct writeback_control wbc = {
 258                 .sync_mode              = WB_SYNC_NONE,
 259                 .older_than_this        = NULL,
 260                 .range_cyclic           = 1,
 261                 .nr_to_write            = 1024,
 262         };
 263
 264         writeback_inodes_wb(&bdi->wb, &wbc);
 265 }
 266
 267 /*
 268  * kupdated() used to do this. We cannot do it from the bdi_forker_thread()
 269  * or we risk deadlocking on ->s_umount. The longer term solution would be
 270  * to implement sync_supers_bdi() or similar and simply do it from the
 271  * bdi writeback thread individually.
 272  */
 273 static int bdi_sync_supers(void *unused)
 274 {
 275         set_user_nice(current, 0);
 276
 277         while (!kthread_should_stop()) {
 278                 set_current_state(TASK_INTERRUPTIBLE);
 279                 schedule();
 280
 281                 /*
 282                  * Do this periodically, like kupdated() did before.
 283                  */
 284                 sync_supers();
 285         }
 286
 287         return 0;
 288 }
 289
 290 void bdi_arm_supers_timer(void)
 291 {
 292         unsigned long next;
 293
 294         if (!dirty_writeback_interval)
 295                 return;
 296
 297         next = msecs_to_jiffies(dirty_writeback_interval * 10) + jiffies;
 298         mod_timer(&sync_supers_timer, round_jiffies_up(next));
 299 }
 300
 301 static void sync_supers_timer_fn(unsigned long unused)
 302 {
 303         wake_up_process(sync_supers_tsk);
 304         bdi_arm_supers_timer();
 305 }
 306
 307 static void wakeup_timer_fn(unsigned long data)
 308 {
 309         struct backing_dev_info *bdi = (struct backing_dev_info *)data;
 310
 311         spin_lock_bh(&bdi->wb_lock);
 312         if (bdi->wb.task) {
 313                 wake_up_process(bdi->wb.task);
 314         } else {
 315                 /*
 316                  * When bdi tasks are inactive for long time, they are killed.
 317                  * In this case we have to wake-up the forker thread which
 318                  * should create and run the bdi thread.
 319                  */
 320                 wake_up_process(default_backing_dev_info.wb.task);
 321         }
 322         spin_unlock_bh(&bdi->wb_lock);
 323 }
 324
 325 /*
 326  * This function is used when the first inode for this bdi is marked dirty. It
 327  * wakes-up the corresponding bdi thread which should then take care of the
 328  * periodic background write-out of dirty inodes. Since the write-out would
 329  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 330  * set up a timer which wakes the bdi thread up later.
 331  *
 332  * Note, we wouldn't bother setting up the timer, but this function is on the
 333  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 334  * by delaying the wake-up.
 335  */
 336 void bdi_wakeup_thread_delayed(struct backing_dev_info *bdi)
 337 {
 338         unsigned long timeout;
 339
 340         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 341         mod_timer(&bdi->wb.wakeup_timer, jiffies + timeout);
 342 }
 343
 344 /*
 345  * Calculate the longest interval (jiffies) bdi threads are allowed to be
 346  * inactive.
 347  */
 348 static unsigned long bdi_longest_inactive(void)
 349 {
 350         unsigned long interval;
 351
 352         interval = msecs_to_jiffies(dirty_writeback_interval * 10);
 353         return max(5UL * 60 * HZ, interval);
 354 }
 355
 356 static int bdi_forker_thread(void *ptr)
 357 {
 358         struct bdi_writeback *me = ptr;
 359
 360         current->flags |= PF_FLUSHER | PF_SWAPWRITE;
 361         set_freezable();
 362
 363         /*
 364          * Our parent may run at a different priority, just set us to normal
 365          */
 366         set_user_nice(current, 0);
 367
 368         for (;;) {
 369                 struct task_struct *task = NULL;
 370                 struct backing_dev_info *bdi;
 371                 enum {
 372                         NO_ACTION,   /* Nothing to do */
 373                         FORK_THREAD, /* Fork bdi thread */
 374                         KILL_THREAD, /* Kill inactive bdi thread */
 375                 } action = NO_ACTION;
 376
 377                 /*
 378                  * Temporary measure, we want to make sure we don't see
 379                  * dirty data on the default backing_dev_info
 380                  */
 381                 if (wb_has_dirty_io(me) || !list_empty(&me->bdi->work_list)) {
 382                         del_timer(&me->wakeup_timer);
 383                         wb_do_writeback(me, 0);
 384                 }
 385
 386                 spin_lock_bh(&bdi_lock);
 387                 set_current_state(TASK_INTERRUPTIBLE);
 388
 389                 list_for_each_entry(bdi, &bdi_list, bdi_list) {
 390                         bool have_dirty_io;
 391
 392                         if (!bdi_cap_writeback_dirty(bdi) ||
 393                              bdi_cap_flush_forker(bdi))
 394                                 continue;
 395
 396                         WARN(!test_bit(BDI_registered, &bdi->state),
 397                              "bdi %p/%s is not registered!\n", bdi, bdi->name);
 398
 399                         have_dirty_io = !list_empty(&bdi->work_list) ||
 400                                         wb_has_dirty_io(&bdi->wb);
 401
 402                         /*
 403                          * If the bdi has work to do, but the thread does not
 404                          * exist - create it.
 405                          */
 406                         if (!bdi->wb.task && have_dirty_io) {
 407                                 /*
 408                                  * Set the pending bit - if someone will try to
 409                                  * unregister this bdi - it'll wait on this bit.
 410                                  */
 411                                 set_bit(BDI_pending, &bdi->state);
 412                                 action = FORK_THREAD;
 413                                 break;
 414                         }
 415
 416                         spin_lock_bh(&bdi->wb_lock);
 417                         /*
 418                          * If there is no work to do and the bdi thread was
 419                          * inactive long enough - kill it. The wb_lock is taken
 420                          * to make sure no-one adds more work to this bdi and
 421                          * wakes the bdi thread up.
 422                          */
 423                         if (bdi->wb.task && !have_dirty_io &&
 424                             time_after(jiffies, bdi->wb.last_active +
 425                                                 bdi_longest_inactive())) {
 426                                 task = bdi->wb.task;
 427                                 bdi->wb.task = NULL;
 428                                 spin_unlock(&bdi->wb_lock);
 429                                 set_bit(BDI_pending, &bdi->state);
 430                                 action = KILL_THREAD;
 431                                 break;
 432                         }
 433                         spin_unlock_bh(&bdi->wb_lock);
 434                 }
 435                 spin_unlock_bh(&bdi_lock);
 436
 437                 /* Keep working if default bdi still has things to do */
 438                 if (!list_empty(&me->bdi->work_list))
 439                         __set_current_state(TASK_RUNNING);
 440
 441                 switch (action) {
 442                 case FORK_THREAD:
 443                         __set_current_state(TASK_RUNNING);
 444                         task = kthread_run(bdi_writeback_thread, &bdi->wb, "flush-%s",
 445                                            dev_name(bdi->dev));
 446                         if (IS_ERR(task)) {
 447                                 /*
 448                                  * If thread creation fails, force writeout of
 449                                  * the bdi from the thread.
 450                                  */
 451                                 bdi_flush_io(bdi);
 452                         } else {
 453                                 /*
 454                                  * The spinlock makes sure we do not lose
 455                                  * wake-ups when racing with 'bdi_queue_work()'.
 456                                  */
 457                                 spin_lock_bh(&bdi->wb_lock);
 458                                 bdi->wb.task = task;
 459                                 spin_unlock_bh(&bdi->wb_lock);
 460                         }
 461                         break;
 462
 463                 case KILL_THREAD:
 464                         __set_current_state(TASK_RUNNING);
 465                         kthread_stop(task);
 466                         break;
 467
 468                 case NO_ACTION:
 469                         if (!wb_has_dirty_io(me) || !dirty_writeback_interval)
 470                                 /*
 471                                  * There are no dirty data. The only thing we
 472                                  * should now care about is checking for
 473                                  * inactive bdi threads and killing them. Thus,
 474                                  * let's sleep for longer time, save energy and
 475                                  * be friendly for battery-driven devices.
 476                                  */
 477                                 schedule_timeout(bdi_longest_inactive());
 478                         else
 479                                 schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
 480                         try_to_freeze();
 481                         /* Back to the main loop */
 482                         continue;
 483                 }
 484
 485                 /*
 486                  * Clear pending bit and wakeup anybody waiting to tear us down.
 487                  */
 488                 clear_bit(BDI_pending, &bdi->state);
 489                 smp_mb__after_clear_bit();
 490                 wake_up_bit(&bdi->state, BDI_pending);
 491         }
 492
 493         return 0;
 494 }
 495
 496 /*
 497  * Remove bdi from bdi_list, and ensure that it is no longer visible
 498  */
 499 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 500 {
 501         spin_lock_bh(&bdi_lock);
 502         list_del_rcu(&bdi->bdi_list);
 503         spin_unlock_bh(&bdi_lock);
 504
 505         synchronize_rcu();
 506 }
 507
 508 int bdi_register(struct backing_dev_info *bdi, struct device *parent,
 509                 const char *fmt, ...)
 510 {
 511         va_list args;
 512         int ret = 0;
 513         struct device *dev;
 514
 515         if (bdi->dev)   /* The driver needs to use separate queues per device */
 516                 goto exit;
 517
 518         va_start(args, fmt);
 519         dev = device_create_vargs(bdi_class, parent, MKDEV(0, 0), bdi, fmt, args);
 520         va_end(args);
 521         if (IS_ERR(dev)) {
 522                 ret = PTR_ERR(dev);
 523                 goto exit;
 524         }
 525
 526         spin_lock_bh(&bdi_lock);
 527         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 528         spin_unlock_bh(&bdi_lock);
 529
 530         bdi->dev = dev;
 531
 532         /*
 533          * Just start the forker thread for our default backing_dev_info,
 534          * and add other bdi's to the list. They will get a thread created
 535          * on-demand when they need it.
 536          */
 537         if (bdi_cap_flush_forker(bdi)) {
 538                 struct bdi_writeback *wb = &bdi->wb;
 539
 540                 wb->task = kthread_run(bdi_forker_thread, wb, "bdi-%s",
 541                                                 dev_name(dev));
 542                 if (IS_ERR(wb->task)) {
 543                         wb->task = NULL;
 544                         ret = -ENOMEM;
 545
 546                         bdi_remove_from_list(bdi);
 547                         goto exit;
 548                 }
 549         }
 550
 551         bdi_debug_register(bdi, dev_name(dev));
 552         set_bit(BDI_registered, &bdi->state);
 553         trace_writeback_bdi_register(bdi);
 554 exit:
 555         return ret;
 556 }
 557 EXPORT_SYMBOL(bdi_register);
 558
 559 int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev)
 560 {
 561         return bdi_register(bdi, NULL, "%u:%u", MAJOR(dev), MINOR(dev));
 562 }
 563 EXPORT_SYMBOL(bdi_register_dev);
 564
 565 /*
 566  * Remove bdi from the global list and shutdown any threads we have running
 567  */
 568 static void bdi_wb_shutdown(struct backing_dev_info *bdi)
 569 {
 570         if (!bdi_cap_writeback_dirty(bdi))
 571                 return;
 572
 573         /*
 574          * Make sure nobody finds us on the bdi_list anymore
 575          */
 576         bdi_remove_from_list(bdi);
 577
 578         /*
 579          * If setup is pending, wait for that to complete first
 580          */
 581         wait_on_bit(&bdi->state, BDI_pending, bdi_sched_wait,
 582                         TASK_UNINTERRUPTIBLE);
 583
 584         /*
 585          * Finally, kill the kernel thread. We don't need to be RCU
 586          * safe anymore, since the bdi is gone from visibility. Force
 587          * unfreeze of the thread before calling kthread_stop(), otherwise
 588          * it would never exet if it is currently stuck in the refrigerator.
 589          */
 590         if (bdi->wb.task) {
 591                 thaw_process(bdi->wb.task);
 592                 kthread_stop(bdi->wb.task);
 593         }
 594 }
 595
 596 /*
 597  * This bdi is going away now, make sure that no super_blocks point to it
 598  */
 599 static void bdi_prune_sb(struct backing_dev_info *bdi)
 600 {
 601         struct super_block *sb;
 602
 603         spin_lock(&sb_lock);
 604         list_for_each_entry(sb, &super_blocks, s_list) {
 605                 if (sb->s_bdi == bdi)
 606                         sb->s_bdi = NULL;
 607         }
 608         spin_unlock(&sb_lock);
 609 }
 610
 611 void bdi_unregister(struct backing_dev_info *bdi)
 612 {
 613         if (bdi->dev) {
 614                 trace_writeback_bdi_unregister(bdi);
 615                 bdi_prune_sb(bdi);
 616                 del_timer_sync(&bdi->wb.wakeup_timer);
 617
 618                 if (!bdi_cap_flush_forker(bdi))
 619                         bdi_wb_shutdown(bdi);
 620                 bdi_debug_unregister(bdi);
 621                 device_unregister(bdi->dev);
 622                 bdi->dev = NULL;
 623         }
 624 }
 625 EXPORT_SYMBOL(bdi_unregister);
 626
 627 static void bdi_wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi)
 628 {
 629         memset(wb, 0, sizeof(*wb));
 630
 631         wb->bdi = bdi;
 632         wb->last_old_flush = jiffies;
 633         INIT_LIST_HEAD(&wb->b_dirty);
 634         INIT_LIST_HEAD(&wb->b_io);
 635         INIT_LIST_HEAD(&wb->b_more_io);
 636         setup_timer(&wb->wakeup_timer, wakeup_timer_fn, (unsigned long)bdi);
 637 }
 638
 639 int bdi_init(struct backing_dev_info *bdi)
 640 {
 641         int i, err;
 642
 643         bdi->dev = NULL;
 644
 645         bdi->min_ratio = 0;
 646         bdi->max_ratio = 100;
 647         bdi->max_prop_frac = PROP_FRAC_BASE;
 648         spin_lock_init(&bdi->wb_lock);
 649         INIT_LIST_HEAD(&bdi->bdi_list);
 650         INIT_LIST_HEAD(&bdi->work_list);
 651
 652         bdi_wb_init(&bdi->wb, bdi);
 653
 654         for (i = 0; i < NR_BDI_STAT_ITEMS; i++) {
 655                 err = percpu_counter_init(&bdi->bdi_stat[i], 0);
 656                 if (err)
 657                         goto err;
 658         }
 659
 660         bdi->dirty_exceeded = 0;
 661         err = prop_local_init_percpu(&bdi->completions);
 662
 663         if (err) {
 664 err:
 665                 while (i--)
 666                         percpu_counter_destroy(&bdi->bdi_stat[i]);
 667         }
 668
 669         return err;
 670 }
 671 EXPORT_SYMBOL(bdi_init);
 672
 673 void bdi_destroy(struct backing_dev_info *bdi)
 674 {
 675         int i;
 676
 677         /*
 678          * Splice our entries to the default_backing_dev_info, if this
 679          * bdi disappears
 680          */
 681         if (bdi_has_dirty_io(bdi)) {
 682                 struct bdi_writeback *dst = &default_backing_dev_info.wb;
 683
 684                 spin_lock(&inode_lock);
 685                 list_splice(&bdi->wb.b_dirty, &dst->b_dirty);
 686                 list_splice(&bdi->wb.b_io, &dst->b_io);
 687                 list_splice(&bdi->wb.b_more_io, &dst->b_more_io);
 688                 spin_unlock(&inode_lock);
 689         }
 690
 691         bdi_unregister(bdi);
 692
 693         for (i = 0; i < NR_BDI_STAT_ITEMS; i++)
 694                 percpu_counter_destroy(&bdi->bdi_stat[i]);
 695
 696         prop_local_destroy_percpu(&bdi->completions);
 697 }
 698 EXPORT_SYMBOL(bdi_destroy);
 699
 700 /*
 701  * For use from filesystems to quickly init and register a bdi associated
 702  * with dirty writeback
 703  */
 704 int bdi_setup_and_register(struct backing_dev_info *bdi, char *name,
 705                            unsigned int cap)
 706 {
 707         char tmp[32];
 708         int err;
 709
 710         bdi->name = name;
 711         bdi->capabilities = cap;
 712         err = bdi_init(bdi);
 713         if (err)
 714                 return err;
 715
 716         sprintf(tmp, "%.28s%s", name, "-%d");
 717         err = bdi_register(bdi, NULL, tmp, atomic_long_inc_return(&bdi_seq));
 718         if (err) {
 719                 bdi_destroy(bdi);
 720                 return err;
 721         }
 722
 723         return 0;
 724 }
 725 EXPORT_SYMBOL(bdi_setup_and_register);
 726
 727 static wait_queue_head_t congestion_wqh[2] = {
 728                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
 729                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
 730         };
 731
 732 void clear_bdi_congested(struct backing_dev_info *bdi, int sync)
 733 {
 734         enum bdi_state bit;
 735         wait_queue_head_t *wqh = &congestion_wqh[sync];
 736
 737         bit = sync ? BDI_sync_congested : BDI_async_congested;
 738         clear_bit(bit, &bdi->state);
 739         smp_mb__after_clear_bit();
 740         if (waitqueue_active(wqh))
 741                 wake_up(wqh);
 742 }
 743 EXPORT_SYMBOL(clear_bdi_congested);
 744
 745 void set_bdi_congested(struct backing_dev_info *bdi, int sync)
 746 {
 747         enum bdi_state bit;
 748
 749         bit = sync ? BDI_sync_congested : BDI_async_congested;
 750         set_bit(bit, &bdi->state);
 751 }
 752 EXPORT_SYMBOL(set_bdi_congested);
 753
 754 /**
 755  * congestion_wait - wait for a backing_dev to become uncongested
 756  * @sync: SYNC or ASYNC IO
 757  * @timeout: timeout in jiffies
 758  *
 759  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
 760  * write congestion.  If no backing_devs are congested then just wait for the
 761  * next write to be completed.
 762  */
 763 long congestion_wait(int sync, long timeout)
 764 {
 765         long ret;
 766         DEFINE_WAIT(wait);
 767         wait_queue_head_t *wqh = &congestion_wqh[sync];
 768
 769         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
 770         ret = io_schedule_timeout(timeout);
 771         finish_wait(wqh, &wait);
 772         return ret;
 773 }
 774 EXPORT_SYMBOL(congestion_wait);
 775