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 struct backing_dev_info default_backing_dev_info = {
  18         .name           = "default",
  19         .ra_pages       = VM_MAX_READAHEAD * 1024 / PAGE_CACHE_SIZE,
  20         .state          = 0,
  21         .capabilities   = BDI_CAP_MAP_COPY,
  22 };
  23 EXPORT_SYMBOL_GPL(default_backing_dev_info);
  24
  25 struct backing_dev_info noop_backing_dev_info = {
  26         .name           = "noop",
  27         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK,
  28 };
  29 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
  30
  31 static struct class *bdi_class;
  32
  33 /*
  34  * bdi_lock protects updates to bdi_list and bdi_pending_list, as well as
  35  * reader side protection for bdi_pending_list. bdi_list has RCU reader side
  36  * locking.
  37  */
  38 DEFINE_SPINLOCK(bdi_lock);
  39 LIST_HEAD(bdi_list);
  40 LIST_HEAD(bdi_pending_list);
  41
  42 void bdi_lock_two(struct bdi_writeback *wb1, struct bdi_writeback *wb2)
  43 {
  44         if (wb1 < wb2) {
  45                 spin_lock(&wb1->list_lock);
  46                 spin_lock_nested(&wb2->list_lock, 1);
  47         } else {
  48                 spin_lock(&wb2->list_lock);
  49                 spin_lock_nested(&wb1->list_lock, 1);
  50         }
  51 }
  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;
  72         struct inode *inode;
  73
  74         nr_dirty = nr_io = nr_more_io = 0;
  75         spin_lock(&wb->list_lock);
  76         list_for_each_entry(inode, &wb->b_dirty, i_wb_list)
  77                 nr_dirty++;
  78         list_for_each_entry(inode, &wb->b_io, i_wb_list)
  79                 nr_io++;
  80         list_for_each_entry(inode, &wb->b_more_io, i_wb_list)
  81                 nr_more_io++;
  82         spin_unlock(&wb->list_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:       %10lu kB\n"
  90                    "BdiReclaimable:     %10lu kB\n"
  91                    "BdiDirtyThresh:     %10lu kB\n"
  92                    "DirtyThresh:        %10lu kB\n"
  93                    "BackgroundThresh:   %10lu kB\n"
  94                    "BdiDirtied:         %10lu kB\n"
  95                    "BdiWritten:         %10lu kB\n"
  96                    "BdiWriteBandwidth:  %10lu kBps\n"
  97                    "b_dirty:            %10lu\n"
  98                    "b_io:               %10lu\n"
  99                    "b_more_io:          %10lu\n"
 100                    "bdi_list:           %10u\n"
 101                    "state:              %10lx\n",
 102                    (unsigned long) K(bdi_stat(bdi, BDI_WRITEBACK)),
 103                    (unsigned long) K(bdi_stat(bdi, BDI_RECLAIMABLE)),
 104                    K(bdi_thresh),
 105                    K(dirty_thresh),
 106                    K(background_thresh),
 107                    (unsigned long) K(bdi_stat(bdi, BDI_DIRTIED)),
 108                    (unsigned long) K(bdi_stat(bdi, BDI_WRITTEN)),
 109                    (unsigned long) K(bdi->write_bandwidth),
 110                    nr_dirty,
 111                    nr_io,
 112                    nr_more_io,
 113                    !list_empty(&bdi->bdi_list), bdi->state);
 114 #undef K
 115
 116         return 0;
 117 }
 118
 119 static int bdi_debug_stats_open(struct inode *inode, struct file *file)
 120 {
 121         return single_open(file, bdi_debug_stats_show, inode->i_private);
 122 }
 123
 124 static const struct file_operations bdi_debug_stats_fops = {
 125         .open           = bdi_debug_stats_open,
 126         .read           = seq_read,
 127         .llseek         = seq_lseek,
 128         .release        = single_release,
 129 };
 130
 131 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
 132 {
 133         bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
 134         bdi->debug_stats = debugfs_create_file("stats", 0444, bdi->debug_dir,
 135                                                bdi, &bdi_debug_stats_fops);
 136 }
 137
 138 static void bdi_debug_unregister(struct backing_dev_info *bdi)
 139 {
 140         debugfs_remove(bdi->debug_stats);
 141         debugfs_remove(bdi->debug_dir);
 142 }
 143 #else
 144 static inline void bdi_debug_init(void)
 145 {
 146 }
 147 static inline void bdi_debug_register(struct backing_dev_info *bdi,
 148                                       const char *name)
 149 {
 150 }
 151 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
 152 {
 153 }
 154 #endif
 155
 156 static ssize_t read_ahead_kb_store(struct device *dev,
 157                                   struct device_attribute *attr,
 158                                   const char *buf, size_t count)
 159 {
 160         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 161         unsigned long read_ahead_kb;
 162         ssize_t ret;
 163
 164         ret = kstrtoul(buf, 10, &read_ahead_kb);
 165         if (ret < 0)
 166                 return ret;
 167
 168         bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
 169
 170         return count;
 171 }
 172
 173 #define K(pages) ((pages) << (PAGE_SHIFT - 10))
 174
 175 #define BDI_SHOW(name, expr)                                            \
 176 static ssize_t name##_show(struct device *dev,                          \
 177                            struct device_attribute *attr, char *page)   \
 178 {                                                                       \
 179         struct backing_dev_info *bdi = dev_get_drvdata(dev);            \
 180                                                                         \
 181         return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr);  \
 182 }
 183
 184 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
 185
 186 static ssize_t min_ratio_store(struct device *dev,
 187                 struct device_attribute *attr, const char *buf, size_t count)
 188 {
 189         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 190         unsigned int ratio;
 191         ssize_t ret;
 192
 193         ret = kstrtouint(buf, 10, &ratio);
 194         if (ret < 0)
 195                 return ret;
 196
 197         ret = bdi_set_min_ratio(bdi, ratio);
 198         if (!ret)
 199                 ret = count;
 200
 201         return ret;
 202 }
 203 BDI_SHOW(min_ratio, bdi->min_ratio)
 204
 205 static ssize_t max_ratio_store(struct device *dev,
 206                 struct device_attribute *attr, const char *buf, size_t count)
 207 {
 208         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 209         unsigned int ratio;
 210         ssize_t ret;
 211
 212         ret = kstrtouint(buf, 10, &ratio);
 213         if (ret < 0)
 214                 return ret;
 215
 216         ret = bdi_set_max_ratio(bdi, ratio);
 217         if (!ret)
 218                 ret = count;
 219
 220         return ret;
 221 }
 222 BDI_SHOW(max_ratio, bdi->max_ratio)
 223
 224 #define __ATTR_RW(attr) __ATTR(attr, 0644, attr##_show, attr##_store)
 225
 226 static struct device_attribute bdi_dev_attrs[] = {
 227         __ATTR_RW(read_ahead_kb),
 228         __ATTR_RW(min_ratio),
 229         __ATTR_RW(max_ratio),
 230         __ATTR_NULL,
 231 };
 232
 233 static __init int bdi_class_init(void)
 234 {
 235         bdi_class = class_create(THIS_MODULE, "bdi");
 236         if (IS_ERR(bdi_class))
 237                 return PTR_ERR(bdi_class);
 238
 239         bdi_class->dev_attrs = bdi_dev_attrs;
 240         bdi_debug_init();
 241         return 0;
 242 }
 243 postcore_initcall(bdi_class_init);
 244
 245 static int __init default_bdi_init(void)
 246 {
 247         int err;
 248
 249         err = bdi_init(&default_backing_dev_info);
 250         if (!err)
 251                 bdi_register(&default_backing_dev_info, NULL, "default");
 252         err = bdi_init(&noop_backing_dev_info);
 253
 254         return err;
 255 }
 256 subsys_initcall(default_bdi_init);
 257
 258 int bdi_has_dirty_io(struct backing_dev_info *bdi)
 259 {
 260         return wb_has_dirty_io(&bdi->wb);
 261 }
 262
 263 static void wakeup_timer_fn(unsigned long data)
 264 {
 265         struct backing_dev_info *bdi = (struct backing_dev_info *)data;
 266
 267         spin_lock_bh(&bdi->wb_lock);
 268         if (bdi->wb.task) {
 269                 trace_writeback_wake_thread(bdi);
 270                 wake_up_process(bdi->wb.task);
 271         } else if (bdi->dev) {
 272                 /*
 273                  * When bdi tasks are inactive for long time, they are killed.
 274                  * In this case we have to wake-up the forker thread which
 275                  * should create and run the bdi thread.
 276                  */
 277                 trace_writeback_wake_forker_thread(bdi);
 278                 wake_up_process(default_backing_dev_info.wb.task);
 279         }
 280         spin_unlock_bh(&bdi->wb_lock);
 281 }
 282
 283 /*
 284  * This function is used when the first inode for this bdi is marked dirty. It
 285  * wakes-up the corresponding bdi thread which should then take care of the
 286  * periodic background write-out of dirty inodes. Since the write-out would
 287  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 288  * set up a timer which wakes the bdi thread up later.
 289  *
 290  * Note, we wouldn't bother setting up the timer, but this function is on the
 291  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 292  * by delaying the wake-up.
 293  */
 294 void bdi_wakeup_thread_delayed(struct backing_dev_info *bdi)
 295 {
 296         unsigned long timeout;
 297
 298         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 299         mod_timer(&bdi->wb.wakeup_timer, jiffies + timeout);
 300 }
 301
 302 /*
 303  * Calculate the longest interval (jiffies) bdi threads are allowed to be
 304  * inactive.
 305  */
 306 static unsigned long bdi_longest_inactive(void)
 307 {
 308         unsigned long interval;
 309
 310         interval = msecs_to_jiffies(dirty_writeback_interval * 10);
 311         return max(5UL * 60 * HZ, interval);
 312 }
 313
 314 /*
 315  * Clear pending bit and wakeup anybody waiting for flusher thread creation or
 316  * shutdown
 317  */
 318 static void bdi_clear_pending(struct backing_dev_info *bdi)
 319 {
 320         clear_bit(BDI_pending, &bdi->state);
 321         smp_mb__after_clear_bit();
 322         wake_up_bit(&bdi->state, BDI_pending);
 323 }
 324
 325 static int bdi_forker_thread(void *ptr)
 326 {
 327         struct bdi_writeback *me = ptr;
 328
 329         current->flags |= PF_SWAPWRITE;
 330         set_freezable();
 331
 332         /*
 333          * Our parent may run at a different priority, just set us to normal
 334          */
 335         set_user_nice(current, 0);
 336
 337         for (;;) {
 338                 struct task_struct *task = NULL;
 339                 struct backing_dev_info *bdi;
 340                 enum {
 341                         NO_ACTION,   /* Nothing to do */
 342                         FORK_THREAD, /* Fork bdi thread */
 343                         KILL_THREAD, /* Kill inactive bdi thread */
 344                 } action = NO_ACTION;
 345
 346                 /*
 347                  * Temporary measure, we want to make sure we don't see
 348                  * dirty data on the default backing_dev_info
 349                  */
 350                 if (wb_has_dirty_io(me) || !list_empty(&me->bdi->work_list)) {
 351                         del_timer(&me->wakeup_timer);
 352                         wb_do_writeback(me, 0);
 353                 }
 354
 355                 spin_lock_bh(&bdi_lock);
 356                 /*
 357                  * In the following loop we are going to check whether we have
 358                  * some work to do without any synchronization with tasks
 359                  * waking us up to do work for them. Set the task state here
 360                  * so that we don't miss wakeups after verifying conditions.
 361                  */
 362                 set_current_state(TASK_INTERRUPTIBLE);
 363
 364                 list_for_each_entry(bdi, &bdi_list, bdi_list) {
 365                         bool have_dirty_io;
 366
 367                         if (!bdi_cap_writeback_dirty(bdi) ||
 368                              bdi_cap_flush_forker(bdi))
 369                                 continue;
 370
 371                         WARN(!test_bit(BDI_registered, &bdi->state),
 372                              "bdi %p/%s is not registered!\n", bdi, bdi->name);
 373
 374                         have_dirty_io = !list_empty(&bdi->work_list) ||
 375                                         wb_has_dirty_io(&bdi->wb);
 376
 377                         /*
 378                          * If the bdi has work to do, but the thread does not
 379                          * exist - create it.
 380                          */
 381                         if (!bdi->wb.task && have_dirty_io) {
 382                                 /*
 383                                  * Set the pending bit - if someone will try to
 384                                  * unregister this bdi - it'll wait on this bit.
 385                                  */
 386                                 set_bit(BDI_pending, &bdi->state);
 387                                 action = FORK_THREAD;
 388                                 break;
 389                         }
 390
 391                         spin_lock(&bdi->wb_lock);
 392
 393                         /*
 394                          * If there is no work to do and the bdi thread was
 395                          * inactive long enough - kill it. The wb_lock is taken
 396                          * to make sure no-one adds more work to this bdi and
 397                          * wakes the bdi thread up.
 398                          */
 399                         if (bdi->wb.task && !have_dirty_io &&
 400                             time_after(jiffies, bdi->wb.last_active +
 401                                                 bdi_longest_inactive())) {
 402                                 task = bdi->wb.task;
 403                                 bdi->wb.task = NULL;
 404                                 spin_unlock(&bdi->wb_lock);
 405                                 set_bit(BDI_pending, &bdi->state);
 406                                 action = KILL_THREAD;
 407                                 break;
 408                         }
 409                         spin_unlock(&bdi->wb_lock);
 410                 }
 411                 spin_unlock_bh(&bdi_lock);
 412
 413                 /* Keep working if default bdi still has things to do */
 414                 if (!list_empty(&me->bdi->work_list))
 415                         __set_current_state(TASK_RUNNING);
 416
 417                 switch (action) {
 418                 case FORK_THREAD:
 419                         __set_current_state(TASK_RUNNING);
 420                         task = kthread_create(bdi_writeback_thread, &bdi->wb,
 421                                               "flush-%s", dev_name(bdi->dev));
 422                         if (IS_ERR(task)) {
 423                                 /*
 424                                  * If thread creation fails, force writeout of
 425                                  * the bdi from the thread. Hopefully 1024 is
 426                                  * large enough for efficient IO.
 427                                  */
 428                                 writeback_inodes_wb(&bdi->wb, 1024,
 429                                                     WB_REASON_FORKER_THREAD);
 430                         } else {
 431                                 /*
 432                                  * The spinlock makes sure we do not lose
 433                                  * wake-ups when racing with 'bdi_queue_work()'.
 434                                  * And as soon as the bdi thread is visible, we
 435                                  * can start it.
 436                                  */
 437                                 spin_lock_bh(&bdi->wb_lock);
 438                                 bdi->wb.task = task;
 439                                 spin_unlock_bh(&bdi->wb_lock);
 440                                 wake_up_process(task);
 441                         }
 442                         bdi_clear_pending(bdi);
 443                         break;
 444
 445                 case KILL_THREAD:
 446                         __set_current_state(TASK_RUNNING);
 447                         kthread_stop(task);
 448                         bdi_clear_pending(bdi);
 449                         break;
 450
 451                 case NO_ACTION:
 452                         if (!wb_has_dirty_io(me) || !dirty_writeback_interval)
 453                                 /*
 454                                  * There are no dirty data. The only thing we
 455                                  * should now care about is checking for
 456                                  * inactive bdi threads and killing them. Thus,
 457                                  * let's sleep for longer time, save energy and
 458                                  * be friendly for battery-driven devices.
 459                                  */
 460                                 schedule_timeout(bdi_longest_inactive());
 461                         else
 462                                 schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
 463                         try_to_freeze();
 464                         break;
 465                 }
 466         }
 467
 468         return 0;
 469 }
 470
 471 /*
 472  * Remove bdi from bdi_list, and ensure that it is no longer visible
 473  */
 474 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 475 {
 476         spin_lock_bh(&bdi_lock);
 477         list_del_rcu(&bdi->bdi_list);
 478         spin_unlock_bh(&bdi_lock);
 479
 480         synchronize_rcu_expedited();
 481 }
 482
 483 int bdi_register(struct backing_dev_info *bdi, struct device *parent,
 484                 const char *fmt, ...)
 485 {
 486         va_list args;
 487         struct device *dev;
 488
 489         if (bdi->dev)   /* The driver needs to use separate queues per device */
 490                 return 0;
 491
 492         va_start(args, fmt);
 493         dev = device_create_vargs(bdi_class, parent, MKDEV(0, 0), bdi, fmt, args);
 494         va_end(args);
 495         if (IS_ERR(dev))
 496                 return PTR_ERR(dev);
 497
 498         bdi->dev = dev;
 499
 500         /*
 501          * Just start the forker thread for our default backing_dev_info,
 502          * and add other bdi's to the list. They will get a thread created
 503          * on-demand when they need it.
 504          */
 505         if (bdi_cap_flush_forker(bdi)) {
 506                 struct bdi_writeback *wb = &bdi->wb;
 507
 508                 wb->task = kthread_run(bdi_forker_thread, wb, "bdi-%s",
 509                                                 dev_name(dev));
 510                 if (IS_ERR(wb->task))
 511                         return PTR_ERR(wb->task);
 512         }
 513
 514         bdi_debug_register(bdi, dev_name(dev));
 515         set_bit(BDI_registered, &bdi->state);
 516
 517         spin_lock_bh(&bdi_lock);
 518         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 519         spin_unlock_bh(&bdi_lock);
 520
 521         trace_writeback_bdi_register(bdi);
 522         return 0;
 523 }
 524 EXPORT_SYMBOL(bdi_register);
 525
 526 int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev)
 527 {
 528         return bdi_register(bdi, NULL, "%u:%u", MAJOR(dev), MINOR(dev));
 529 }
 530 EXPORT_SYMBOL(bdi_register_dev);
 531
 532 /*
 533  * Remove bdi from the global list and shutdown any threads we have running
 534  */
 535 static void bdi_wb_shutdown(struct backing_dev_info *bdi)
 536 {
 537         struct task_struct *task;
 538
 539         if (!bdi_cap_writeback_dirty(bdi))
 540                 return;
 541
 542         /*
 543          * Make sure nobody finds us on the bdi_list anymore
 544          */
 545         bdi_remove_from_list(bdi);
 546
 547         /*
 548          * If setup is pending, wait for that to complete first
 549          */
 550         wait_on_bit(&bdi->state, BDI_pending, bdi_sched_wait,
 551                         TASK_UNINTERRUPTIBLE);
 552
 553         /*
 554          * Finally, kill the kernel thread. We don't need to be RCU
 555          * safe anymore, since the bdi is gone from visibility.
 556          */
 557         spin_lock_bh(&bdi->wb_lock);
 558         task = bdi->wb.task;
 559         bdi->wb.task = NULL;
 560         spin_unlock_bh(&bdi->wb_lock);
 561
 562         if (task)
 563                 kthread_stop(task);
 564 }
 565
 566 /*
 567  * This bdi is going away now, make sure that no super_blocks point to it
 568  */
 569 static void bdi_prune_sb(struct backing_dev_info *bdi)
 570 {
 571         struct super_block *sb;
 572
 573         spin_lock(&sb_lock);
 574         list_for_each_entry(sb, &super_blocks, s_list) {
 575                 if (sb->s_bdi == bdi)
 576                         sb->s_bdi = &default_backing_dev_info;
 577         }
 578         spin_unlock(&sb_lock);
 579 }
 580
 581 void bdi_unregister(struct backing_dev_info *bdi)
 582 {
 583         struct device *dev = bdi->dev;
 584
 585         if (dev) {
 586                 bdi_set_min_ratio(bdi, 0);
 587                 trace_writeback_bdi_unregister(bdi);
 588                 bdi_prune_sb(bdi);
 589                 del_timer_sync(&bdi->wb.wakeup_timer);
 590
 591                 if (!bdi_cap_flush_forker(bdi))
 592                         bdi_wb_shutdown(bdi);
 593                 bdi_debug_unregister(bdi);
 594
 595                 spin_lock_bh(&bdi->wb_lock);
 596                 bdi->dev = NULL;
 597                 spin_unlock_bh(&bdi->wb_lock);
 598
 599                 device_unregister(dev);
 600         }
 601 }
 602 EXPORT_SYMBOL(bdi_unregister);
 603
 604 static void bdi_wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi)
 605 {
 606         memset(wb, 0, sizeof(*wb));
 607
 608         wb->bdi = bdi;
 609         wb->last_old_flush = jiffies;
 610         INIT_LIST_HEAD(&wb->b_dirty);
 611         INIT_LIST_HEAD(&wb->b_io);
 612         INIT_LIST_HEAD(&wb->b_more_io);
 613         spin_lock_init(&wb->list_lock);
 614         setup_timer(&wb->wakeup_timer, wakeup_timer_fn, (unsigned long)bdi);
 615 }
 616
 617 /*
 618  * Initial write bandwidth: 100 MB/s
 619  */
 620 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
 621
 622 int bdi_init(struct backing_dev_info *bdi)
 623 {
 624         int i, err;
 625
 626         bdi->dev = NULL;
 627
 628         bdi->min_ratio = 0;
 629         bdi->max_ratio = 100;
 630         bdi->max_prop_frac = FPROP_FRAC_BASE;
 631         spin_lock_init(&bdi->wb_lock);
 632         INIT_LIST_HEAD(&bdi->bdi_list);
 633         INIT_LIST_HEAD(&bdi->work_list);
 634
 635         bdi_wb_init(&bdi->wb, bdi);
 636
 637         for (i = 0; i < NR_BDI_STAT_ITEMS; i++) {
 638                 err = percpu_counter_init(&bdi->bdi_stat[i], 0);
 639                 if (err)
 640                         goto err;
 641         }
 642
 643         bdi->dirty_exceeded = 0;
 644
 645         bdi->bw_time_stamp = jiffies;
 646         bdi->written_stamp = 0;
 647
 648         bdi->balanced_dirty_ratelimit = INIT_BW;
 649         bdi->dirty_ratelimit = INIT_BW;
 650         bdi->write_bandwidth = INIT_BW;
 651         bdi->avg_write_bandwidth = INIT_BW;
 652
 653         err = fprop_local_init_percpu(&bdi->completions);
 654
 655         if (err) {
 656 err:
 657                 while (i--)
 658                         percpu_counter_destroy(&bdi->bdi_stat[i]);
 659         }
 660
 661         return err;
 662 }
 663 EXPORT_SYMBOL(bdi_init);
 664
 665 void bdi_destroy(struct backing_dev_info *bdi)
 666 {
 667         int i;
 668
 669         /*
 670          * Splice our entries to the default_backing_dev_info, if this
 671          * bdi disappears
 672          */
 673         if (bdi_has_dirty_io(bdi)) {
 674                 struct bdi_writeback *dst = &default_backing_dev_info.wb;
 675
 676                 bdi_lock_two(&bdi->wb, dst);
 677                 list_splice(&bdi->wb.b_dirty, &dst->b_dirty);
 678                 list_splice(&bdi->wb.b_io, &dst->b_io);
 679                 list_splice(&bdi->wb.b_more_io, &dst->b_more_io);
 680                 spin_unlock(&bdi->wb.list_lock);
 681                 spin_unlock(&dst->list_lock);
 682         }
 683
 684         bdi_unregister(bdi);
 685
 686         /*
 687          * If bdi_unregister() had already been called earlier, the
 688          * wakeup_timer could still be armed because bdi_prune_sb()
 689          * can race with the bdi_wakeup_thread_delayed() calls from
 690          * __mark_inode_dirty().
 691          */
 692         del_timer_sync(&bdi->wb.wakeup_timer);
 693
 694         for (i = 0; i < NR_BDI_STAT_ITEMS; i++)
 695                 percpu_counter_destroy(&bdi->bdi_stat[i]);
 696
 697         fprop_local_destroy_percpu(&bdi->completions);
 698 }
 699 EXPORT_SYMBOL(bdi_destroy);
 700
 701 /*
 702  * For use from filesystems to quickly init and register a bdi associated
 703  * with dirty writeback
 704  */
 705 int bdi_setup_and_register(struct backing_dev_info *bdi, char *name,
 706                            unsigned int cap)
 707 {
 708         char tmp[32];
 709         int err;
 710
 711         bdi->name = name;
 712         bdi->capabilities = cap;
 713         err = bdi_init(bdi);
 714         if (err)
 715                 return err;
 716
 717         sprintf(tmp, "%.28s%s", name, "-%d");
 718         err = bdi_register(bdi, NULL, tmp, atomic_long_inc_return(&bdi_seq));
 719         if (err) {
 720                 bdi_destroy(bdi);
 721                 return err;
 722         }
 723
 724         return 0;
 725 }
 726 EXPORT_SYMBOL(bdi_setup_and_register);
 727
 728 static wait_queue_head_t congestion_wqh[2] = {
 729                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
 730                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
 731         };
 732 static atomic_t nr_bdi_congested[2];
 733
 734 void clear_bdi_congested(struct backing_dev_info *bdi, int sync)
 735 {
 736         enum bdi_state bit;
 737         wait_queue_head_t *wqh = &congestion_wqh[sync];
 738
 739         bit = sync ? BDI_sync_congested : BDI_async_congested;
 740         if (test_and_clear_bit(bit, &bdi->state))
 741                 atomic_dec(&nr_bdi_congested[sync]);
 742         smp_mb__after_clear_bit();
 743         if (waitqueue_active(wqh))
 744                 wake_up(wqh);
 745 }
 746 EXPORT_SYMBOL(clear_bdi_congested);
 747
 748 void set_bdi_congested(struct backing_dev_info *bdi, int sync)
 749 {
 750         enum bdi_state bit;
 751
 752         bit = sync ? BDI_sync_congested : BDI_async_congested;
 753         if (!test_and_set_bit(bit, &bdi->state))
 754                 atomic_inc(&nr_bdi_congested[sync]);
 755 }
 756 EXPORT_SYMBOL(set_bdi_congested);
 757
 758 /**
 759  * congestion_wait - wait for a backing_dev to become uncongested
 760  * @sync: SYNC or ASYNC IO
 761  * @timeout: timeout in jiffies
 762  *
 763  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
 764  * write congestion.  If no backing_devs are congested then just wait for the
 765  * next write to be completed.
 766  */
 767 long congestion_wait(int sync, long timeout)
 768 {
 769         long ret;
 770         unsigned long start = jiffies;
 771         DEFINE_WAIT(wait);
 772         wait_queue_head_t *wqh = &congestion_wqh[sync];
 773
 774         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
 775         ret = io_schedule_timeout(timeout);
 776         finish_wait(wqh, &wait);
 777
 778         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
 779                                         jiffies_to_usecs(jiffies - start));
 780
 781         return ret;
 782 }
 783 EXPORT_SYMBOL(congestion_wait);
 784
 785 /**
 786  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a zone to complete writes
 787  * @zone: A zone to check if it is heavily congested
 788  * @sync: SYNC or ASYNC IO
 789  * @timeout: timeout in jiffies
 790  *
 791  * In the event of a congested backing_dev (any backing_dev) and the given
 792  * @zone has experienced recent congestion, this waits for up to @timeout
 793  * jiffies for either a BDI to exit congestion of the given @sync queue
 794  * or a write to complete.
 795  *
 796  * In the absence of zone congestion, cond_resched() is called to yield
 797  * the processor if necessary but otherwise does not sleep.
 798  *
 799  * The return value is 0 if the sleep is for the full timeout. Otherwise,
 800  * it is the number of jiffies that were still remaining when the function
 801  * returned. return_value == timeout implies the function did not sleep.
 802  */
 803 long wait_iff_congested(struct zone *zone, int sync, long timeout)
 804 {
 805         long ret;
 806         unsigned long start = jiffies;
 807         DEFINE_WAIT(wait);
 808         wait_queue_head_t *wqh = &congestion_wqh[sync];
 809
 810         /*
 811          * If there is no congestion, or heavy congestion is not being
 812          * encountered in the current zone, yield if necessary instead
 813          * of sleeping on the congestion queue
 814          */
 815         if (atomic_read(&nr_bdi_congested[sync]) == 0 ||
 816                         !zone_is_reclaim_congested(zone)) {
 817                 cond_resched();
 818
 819                 /* In case we scheduled, work out time remaining */
 820                 ret = timeout - (jiffies - start);
 821                 if (ret < 0)
 822                         ret = 0;
 823
 824                 goto out;
 825         }
 826
 827         /* Sleep until uncongested or a write happens */
 828         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
 829         ret = io_schedule_timeout(timeout);
 830         finish_wait(wqh, &wait);
 831
 832 out:
 833         trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
 834                                         jiffies_to_usecs(jiffies - start));
 835
 836         return ret;
 837 }
 838 EXPORT_SYMBOL(wait_iff_congested);
 839
 840 int pdflush_proc_obsolete(struct ctl_table *table, int write,
 841                         void __user *buffer, size_t *lenp, loff_t *ppos)
 842 {
 843         char kbuf[] = "0\n";
 844
 845         if (*ppos) {
 846                 *lenp = 0;
 847                 return 0;
 848         }
 849
 850         if (copy_to_user(buffer, kbuf, sizeof(kbuf)))
 851                 return -EFAULT;
 852         printk_once(KERN_WARNING "%s exported in /proc is scheduled for removal\n",
 853                         table->procname);
 854
 855         *lenp = 2;
 856         *ppos += *lenp;
 857         return 2;
 858 }