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 static ssize_t stable_pages_required_show(struct device *dev,
 225                                           struct device_attribute *attr,
 226                                           char *page)
 227 {
 228         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 229
 230         return snprintf(page, PAGE_SIZE-1, "%d\n",
 231                         bdi_cap_stable_pages_required(bdi) ? 1 : 0);
 232 }
 233
 234 #define __ATTR_RW(attr) __ATTR(attr, 0644, attr##_show, attr##_store)
 235
 236 static struct device_attribute bdi_dev_attrs[] = {
 237         __ATTR_RW(read_ahead_kb),
 238         __ATTR_RW(min_ratio),
 239         __ATTR_RW(max_ratio),
 240         __ATTR_RO(stable_pages_required),
 241         __ATTR_NULL,
 242 };
 243
 244 static __init int bdi_class_init(void)
 245 {
 246         bdi_class = class_create(THIS_MODULE, "bdi");
 247         if (IS_ERR(bdi_class))
 248                 return PTR_ERR(bdi_class);
 249
 250         bdi_class->dev_attrs = bdi_dev_attrs;
 251         bdi_debug_init();
 252         return 0;
 253 }
 254 postcore_initcall(bdi_class_init);
 255
 256 static int __init default_bdi_init(void)
 257 {
 258         int err;
 259
 260         err = bdi_init(&default_backing_dev_info);
 261         if (!err)
 262                 bdi_register(&default_backing_dev_info, NULL, "default");
 263         err = bdi_init(&noop_backing_dev_info);
 264
 265         return err;
 266 }
 267 subsys_initcall(default_bdi_init);
 268
 269 int bdi_has_dirty_io(struct backing_dev_info *bdi)
 270 {
 271         return wb_has_dirty_io(&bdi->wb);
 272 }
 273
 274 static void wakeup_timer_fn(unsigned long data)
 275 {
 276         struct backing_dev_info *bdi = (struct backing_dev_info *)data;
 277
 278         spin_lock_bh(&bdi->wb_lock);
 279         if (bdi->wb.task) {
 280                 trace_writeback_wake_thread(bdi);
 281                 wake_up_process(bdi->wb.task);
 282         } else if (bdi->dev) {
 283                 /*
 284                  * When bdi tasks are inactive for long time, they are killed.
 285                  * In this case we have to wake-up the forker thread which
 286                  * should create and run the bdi thread.
 287                  */
 288                 trace_writeback_wake_forker_thread(bdi);
 289                 wake_up_process(default_backing_dev_info.wb.task);
 290         }
 291         spin_unlock_bh(&bdi->wb_lock);
 292 }
 293
 294 /*
 295  * This function is used when the first inode for this bdi is marked dirty. It
 296  * wakes-up the corresponding bdi thread which should then take care of the
 297  * periodic background write-out of dirty inodes. Since the write-out would
 298  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 299  * set up a timer which wakes the bdi thread up later.
 300  *
 301  * Note, we wouldn't bother setting up the timer, but this function is on the
 302  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 303  * by delaying the wake-up.
 304  */
 305 void bdi_wakeup_thread_delayed(struct backing_dev_info *bdi)
 306 {
 307         unsigned long timeout;
 308
 309         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 310         mod_timer(&bdi->wb.wakeup_timer, jiffies + timeout);
 311 }
 312
 313 /*
 314  * Calculate the longest interval (jiffies) bdi threads are allowed to be
 315  * inactive.
 316  */
 317 static unsigned long bdi_longest_inactive(void)
 318 {
 319         unsigned long interval;
 320
 321         interval = msecs_to_jiffies(dirty_writeback_interval * 10);
 322         return max(5UL * 60 * HZ, interval);
 323 }
 324
 325 /*
 326  * Clear pending bit and wakeup anybody waiting for flusher thread creation or
 327  * shutdown
 328  */
 329 static void bdi_clear_pending(struct backing_dev_info *bdi)
 330 {
 331         clear_bit(BDI_pending, &bdi->state);
 332         smp_mb__after_clear_bit();
 333         wake_up_bit(&bdi->state, BDI_pending);
 334 }
 335
 336 static int bdi_forker_thread(void *ptr)
 337 {
 338         struct bdi_writeback *me = ptr;
 339
 340         current->flags |= PF_SWAPWRITE;
 341         set_freezable();
 342
 343         /*
 344          * Our parent may run at a different priority, just set us to normal
 345          */
 346         set_user_nice(current, 0);
 347
 348         for (;;) {
 349                 struct task_struct *task = NULL;
 350                 struct backing_dev_info *bdi;
 351                 enum {
 352                         NO_ACTION,   /* Nothing to do */
 353                         FORK_THREAD, /* Fork bdi thread */
 354                         KILL_THREAD, /* Kill inactive bdi thread */
 355                 } action = NO_ACTION;
 356
 357                 /*
 358                  * Temporary measure, we want to make sure we don't see
 359                  * dirty data on the default backing_dev_info
 360                  */
 361                 if (wb_has_dirty_io(me) || !list_empty(&me->bdi->work_list)) {
 362                         del_timer(&me->wakeup_timer);
 363                         wb_do_writeback(me, 0);
 364                 }
 365
 366                 spin_lock_bh(&bdi_lock);
 367                 /*
 368                  * In the following loop we are going to check whether we have
 369                  * some work to do without any synchronization with tasks
 370                  * waking us up to do work for them. Set the task state here
 371                  * so that we don't miss wakeups after verifying conditions.
 372                  */
 373                 set_current_state(TASK_INTERRUPTIBLE);
 374
 375                 list_for_each_entry(bdi, &bdi_list, bdi_list) {
 376                         bool have_dirty_io;
 377
 378                         if (!bdi_cap_writeback_dirty(bdi) ||
 379                              bdi_cap_flush_forker(bdi))
 380                                 continue;
 381
 382                         WARN(!test_bit(BDI_registered, &bdi->state),
 383                              "bdi %p/%s is not registered!\n", bdi, bdi->name);
 384
 385                         have_dirty_io = !list_empty(&bdi->work_list) ||
 386                                         wb_has_dirty_io(&bdi->wb);
 387
 388                         /*
 389                          * If the bdi has work to do, but the thread does not
 390                          * exist - create it.
 391                          */
 392                         if (!bdi->wb.task && have_dirty_io) {
 393                                 /*
 394                                  * Set the pending bit - if someone will try to
 395                                  * unregister this bdi - it'll wait on this bit.
 396                                  */
 397                                 set_bit(BDI_pending, &bdi->state);
 398                                 action = FORK_THREAD;
 399                                 break;
 400                         }
 401
 402                         spin_lock(&bdi->wb_lock);
 403
 404                         /*
 405                          * If there is no work to do and the bdi thread was
 406                          * inactive long enough - kill it. The wb_lock is taken
 407                          * to make sure no-one adds more work to this bdi and
 408                          * wakes the bdi thread up.
 409                          */
 410                         if (bdi->wb.task && !have_dirty_io &&
 411                             time_after(jiffies, bdi->wb.last_active +
 412                                                 bdi_longest_inactive())) {
 413                                 task = bdi->wb.task;
 414                                 bdi->wb.task = NULL;
 415                                 spin_unlock(&bdi->wb_lock);
 416                                 set_bit(BDI_pending, &bdi->state);
 417                                 action = KILL_THREAD;
 418                                 break;
 419                         }
 420                         spin_unlock(&bdi->wb_lock);
 421                 }
 422                 spin_unlock_bh(&bdi_lock);
 423
 424                 /* Keep working if default bdi still has things to do */
 425                 if (!list_empty(&me->bdi->work_list))
 426                         __set_current_state(TASK_RUNNING);
 427
 428                 switch (action) {
 429                 case FORK_THREAD:
 430                         __set_current_state(TASK_RUNNING);
 431                         task = kthread_create(bdi_writeback_thread, &bdi->wb,
 432                                               "flush-%s", dev_name(bdi->dev));
 433                         if (IS_ERR(task)) {
 434                                 /*
 435                                  * If thread creation fails, force writeout of
 436                                  * the bdi from the thread. Hopefully 1024 is
 437                                  * large enough for efficient IO.
 438                                  */
 439                                 writeback_inodes_wb(&bdi->wb, 1024,
 440                                                     WB_REASON_FORKER_THREAD);
 441                         } else {
 442                                 /*
 443                                  * The spinlock makes sure we do not lose
 444                                  * wake-ups when racing with 'bdi_queue_work()'.
 445                                  * And as soon as the bdi thread is visible, we
 446                                  * can start it.
 447                                  */
 448                                 spin_lock_bh(&bdi->wb_lock);
 449                                 bdi->wb.task = task;
 450                                 spin_unlock_bh(&bdi->wb_lock);
 451                                 wake_up_process(task);
 452                         }
 453                         bdi_clear_pending(bdi);
 454                         break;
 455
 456                 case KILL_THREAD:
 457                         __set_current_state(TASK_RUNNING);
 458                         kthread_stop(task);
 459                         bdi_clear_pending(bdi);
 460                         break;
 461
 462                 case NO_ACTION:
 463                         if (!wb_has_dirty_io(me) || !dirty_writeback_interval)
 464                                 /*
 465                                  * There are no dirty data. The only thing we
 466                                  * should now care about is checking for
 467                                  * inactive bdi threads and killing them. Thus,
 468                                  * let's sleep for longer time, save energy and
 469                                  * be friendly for battery-driven devices.
 470                                  */
 471                                 schedule_timeout(bdi_longest_inactive());
 472                         else
 473                                 schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
 474                         try_to_freeze();
 475                         break;
 476                 }
 477         }
 478
 479         return 0;
 480 }
 481
 482 /*
 483  * Remove bdi from bdi_list, and ensure that it is no longer visible
 484  */
 485 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 486 {
 487         spin_lock_bh(&bdi_lock);
 488         list_del_rcu(&bdi->bdi_list);
 489         spin_unlock_bh(&bdi_lock);
 490
 491         synchronize_rcu_expedited();
 492 }
 493
 494 int bdi_register(struct backing_dev_info *bdi, struct device *parent,
 495                 const char *fmt, ...)
 496 {
 497         va_list args;
 498         struct device *dev;
 499
 500         if (bdi->dev)   /* The driver needs to use separate queues per device */
 501                 return 0;
 502
 503         va_start(args, fmt);
 504         dev = device_create_vargs(bdi_class, parent, MKDEV(0, 0), bdi, fmt, args);
 505         va_end(args);
 506         if (IS_ERR(dev))
 507                 return PTR_ERR(dev);
 508
 509         bdi->dev = dev;
 510
 511         /*
 512          * Just start the forker thread for our default backing_dev_info,
 513          * and add other bdi's to the list. They will get a thread created
 514          * on-demand when they need it.
 515          */
 516         if (bdi_cap_flush_forker(bdi)) {
 517                 struct bdi_writeback *wb = &bdi->wb;
 518
 519                 wb->task = kthread_run(bdi_forker_thread, wb, "bdi-%s",
 520                                                 dev_name(dev));
 521                 if (IS_ERR(wb->task))
 522                         return PTR_ERR(wb->task);
 523         }
 524
 525         bdi_debug_register(bdi, dev_name(dev));
 526         set_bit(BDI_registered, &bdi->state);
 527
 528         spin_lock_bh(&bdi_lock);
 529         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 530         spin_unlock_bh(&bdi_lock);
 531
 532         trace_writeback_bdi_register(bdi);
 533         return 0;
 534 }
 535 EXPORT_SYMBOL(bdi_register);
 536
 537 int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev)
 538 {
 539         return bdi_register(bdi, NULL, "%u:%u", MAJOR(dev), MINOR(dev));
 540 }
 541 EXPORT_SYMBOL(bdi_register_dev);
 542
 543 /*
 544  * Remove bdi from the global list and shutdown any threads we have running
 545  */
 546 static void bdi_wb_shutdown(struct backing_dev_info *bdi)
 547 {
 548         struct task_struct *task;
 549
 550         if (!bdi_cap_writeback_dirty(bdi))
 551                 return;
 552
 553         /*
 554          * Make sure nobody finds us on the bdi_list anymore
 555          */
 556         bdi_remove_from_list(bdi);
 557
 558         /*
 559          * If setup is pending, wait for that to complete first
 560          */
 561         wait_on_bit(&bdi->state, BDI_pending, bdi_sched_wait,
 562                         TASK_UNINTERRUPTIBLE);
 563
 564         /*
 565          * Finally, kill the kernel thread. We don't need to be RCU
 566          * safe anymore, since the bdi is gone from visibility.
 567          */
 568         spin_lock_bh(&bdi->wb_lock);
 569         task = bdi->wb.task;
 570         bdi->wb.task = NULL;
 571         spin_unlock_bh(&bdi->wb_lock);
 572
 573         if (task)
 574                 kthread_stop(task);
 575 }
 576
 577 /*
 578  * This bdi is going away now, make sure that no super_blocks point to it
 579  */
 580 static void bdi_prune_sb(struct backing_dev_info *bdi)
 581 {
 582         struct super_block *sb;
 583
 584         spin_lock(&sb_lock);
 585         list_for_each_entry(sb, &super_blocks, s_list) {
 586                 if (sb->s_bdi == bdi)
 587                         sb->s_bdi = &default_backing_dev_info;
 588         }
 589         spin_unlock(&sb_lock);
 590 }
 591
 592 void bdi_unregister(struct backing_dev_info *bdi)
 593 {
 594         struct device *dev = bdi->dev;
 595
 596         if (dev) {
 597                 bdi_set_min_ratio(bdi, 0);
 598                 trace_writeback_bdi_unregister(bdi);
 599                 bdi_prune_sb(bdi);
 600                 del_timer_sync(&bdi->wb.wakeup_timer);
 601
 602                 if (!bdi_cap_flush_forker(bdi))
 603                         bdi_wb_shutdown(bdi);
 604                 bdi_debug_unregister(bdi);
 605
 606                 spin_lock_bh(&bdi->wb_lock);
 607                 bdi->dev = NULL;
 608                 spin_unlock_bh(&bdi->wb_lock);
 609
 610                 device_unregister(dev);
 611         }
 612 }
 613 EXPORT_SYMBOL(bdi_unregister);
 614
 615 static void bdi_wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi)
 616 {
 617         memset(wb, 0, sizeof(*wb));
 618
 619         wb->bdi = bdi;
 620         wb->last_old_flush = jiffies;
 621         INIT_LIST_HEAD(&wb->b_dirty);
 622         INIT_LIST_HEAD(&wb->b_io);
 623         INIT_LIST_HEAD(&wb->b_more_io);
 624         spin_lock_init(&wb->list_lock);
 625         setup_timer(&wb->wakeup_timer, wakeup_timer_fn, (unsigned long)bdi);
 626 }
 627
 628 /*
 629  * Initial write bandwidth: 100 MB/s
 630  */
 631 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
 632
 633 int bdi_init(struct backing_dev_info *bdi)
 634 {
 635         int i, err;
 636
 637         bdi->dev = NULL;
 638
 639         bdi->min_ratio = 0;
 640         bdi->max_ratio = 100;
 641         bdi->max_prop_frac = FPROP_FRAC_BASE;
 642         spin_lock_init(&bdi->wb_lock);
 643         INIT_LIST_HEAD(&bdi->bdi_list);
 644         INIT_LIST_HEAD(&bdi->work_list);
 645
 646         bdi_wb_init(&bdi->wb, bdi);
 647
 648         for (i = 0; i < NR_BDI_STAT_ITEMS; i++) {
 649                 err = percpu_counter_init(&bdi->bdi_stat[i], 0);
 650                 if (err)
 651                         goto err;
 652         }
 653
 654         bdi->dirty_exceeded = 0;
 655
 656         bdi->bw_time_stamp = jiffies;
 657         bdi->written_stamp = 0;
 658
 659         bdi->balanced_dirty_ratelimit = INIT_BW;
 660         bdi->dirty_ratelimit = INIT_BW;
 661         bdi->write_bandwidth = INIT_BW;
 662         bdi->avg_write_bandwidth = INIT_BW;
 663
 664         err = fprop_local_init_percpu(&bdi->completions);
 665
 666         if (err) {
 667 err:
 668                 while (i--)
 669                         percpu_counter_destroy(&bdi->bdi_stat[i]);
 670         }
 671
 672         return err;
 673 }
 674 EXPORT_SYMBOL(bdi_init);
 675
 676 void bdi_destroy(struct backing_dev_info *bdi)
 677 {
 678         int i;
 679
 680         /*
 681          * Splice our entries to the default_backing_dev_info, if this
 682          * bdi disappears
 683          */
 684         if (bdi_has_dirty_io(bdi)) {
 685                 struct bdi_writeback *dst = &default_backing_dev_info.wb;
 686
 687                 bdi_lock_two(&bdi->wb, dst);
 688                 list_splice(&bdi->wb.b_dirty, &dst->b_dirty);
 689                 list_splice(&bdi->wb.b_io, &dst->b_io);
 690                 list_splice(&bdi->wb.b_more_io, &dst->b_more_io);
 691                 spin_unlock(&bdi->wb.list_lock);
 692                 spin_unlock(&dst->list_lock);
 693         }
 694
 695         bdi_unregister(bdi);
 696
 697         /*
 698          * If bdi_unregister() had already been called earlier, the
 699          * wakeup_timer could still be armed because bdi_prune_sb()
 700          * can race with the bdi_wakeup_thread_delayed() calls from
 701          * __mark_inode_dirty().
 702          */
 703         del_timer_sync(&bdi->wb.wakeup_timer);
 704
 705         for (i = 0; i < NR_BDI_STAT_ITEMS; i++)
 706                 percpu_counter_destroy(&bdi->bdi_stat[i]);
 707
 708         fprop_local_destroy_percpu(&bdi->completions);
 709 }
 710 EXPORT_SYMBOL(bdi_destroy);
 711
 712 /*
 713  * For use from filesystems to quickly init and register a bdi associated
 714  * with dirty writeback
 715  */
 716 int bdi_setup_and_register(struct backing_dev_info *bdi, char *name,
 717                            unsigned int cap)
 718 {
 719         char tmp[32];
 720         int err;
 721
 722         bdi->name = name;
 723         bdi->capabilities = cap;
 724         err = bdi_init(bdi);
 725         if (err)
 726                 return err;
 727
 728         sprintf(tmp, "%.28s%s", name, "-%d");
 729         err = bdi_register(bdi, NULL, tmp, atomic_long_inc_return(&bdi_seq));
 730         if (err) {
 731                 bdi_destroy(bdi);
 732                 return err;
 733         }
 734
 735         return 0;
 736 }
 737 EXPORT_SYMBOL(bdi_setup_and_register);
 738
 739 static wait_queue_head_t congestion_wqh[2] = {
 740                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
 741                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
 742         };
 743 static atomic_t nr_bdi_congested[2];
 744
 745 void clear_bdi_congested(struct backing_dev_info *bdi, int sync)
 746 {
 747         enum bdi_state bit;
 748         wait_queue_head_t *wqh = &congestion_wqh[sync];
 749
 750         bit = sync ? BDI_sync_congested : BDI_async_congested;
 751         if (test_and_clear_bit(bit, &bdi->state))
 752                 atomic_dec(&nr_bdi_congested[sync]);
 753         smp_mb__after_clear_bit();
 754         if (waitqueue_active(wqh))
 755                 wake_up(wqh);
 756 }
 757 EXPORT_SYMBOL(clear_bdi_congested);
 758
 759 void set_bdi_congested(struct backing_dev_info *bdi, int sync)
 760 {
 761         enum bdi_state bit;
 762
 763         bit = sync ? BDI_sync_congested : BDI_async_congested;
 764         if (!test_and_set_bit(bit, &bdi->state))
 765                 atomic_inc(&nr_bdi_congested[sync]);
 766 }
 767 EXPORT_SYMBOL(set_bdi_congested);
 768
 769 /**
 770  * congestion_wait - wait for a backing_dev to become uncongested
 771  * @sync: SYNC or ASYNC IO
 772  * @timeout: timeout in jiffies
 773  *
 774  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
 775  * write congestion.  If no backing_devs are congested then just wait for the
 776  * next write to be completed.
 777  */
 778 long congestion_wait(int sync, long timeout)
 779 {
 780         long ret;
 781         unsigned long start = jiffies;
 782         DEFINE_WAIT(wait);
 783         wait_queue_head_t *wqh = &congestion_wqh[sync];
 784
 785         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
 786         ret = io_schedule_timeout(timeout);
 787         finish_wait(wqh, &wait);
 788
 789         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
 790                                         jiffies_to_usecs(jiffies - start));
 791
 792         return ret;
 793 }
 794 EXPORT_SYMBOL(congestion_wait);
 795
 796 /**
 797  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a zone to complete writes
 798  * @zone: A zone to check if it is heavily congested
 799  * @sync: SYNC or ASYNC IO
 800  * @timeout: timeout in jiffies
 801  *
 802  * In the event of a congested backing_dev (any backing_dev) and the given
 803  * @zone has experienced recent congestion, this waits for up to @timeout
 804  * jiffies for either a BDI to exit congestion of the given @sync queue
 805  * or a write to complete.
 806  *
 807  * In the absence of zone congestion, cond_resched() is called to yield
 808  * the processor if necessary but otherwise does not sleep.
 809  *
 810  * The return value is 0 if the sleep is for the full timeout. Otherwise,
 811  * it is the number of jiffies that were still remaining when the function
 812  * returned. return_value == timeout implies the function did not sleep.
 813  */
 814 long wait_iff_congested(struct zone *zone, int sync, long timeout)
 815 {
 816         long ret;
 817         unsigned long start = jiffies;
 818         DEFINE_WAIT(wait);
 819         wait_queue_head_t *wqh = &congestion_wqh[sync];
 820
 821         /*
 822          * If there is no congestion, or heavy congestion is not being
 823          * encountered in the current zone, yield if necessary instead
 824          * of sleeping on the congestion queue
 825          */
 826         if (atomic_read(&nr_bdi_congested[sync]) == 0 ||
 827                         !zone_is_reclaim_congested(zone)) {
 828                 cond_resched();
 829
 830                 /* In case we scheduled, work out time remaining */
 831                 ret = timeout - (jiffies - start);
 832                 if (ret < 0)
 833                         ret = 0;
 834
 835                 goto out;
 836         }
 837
 838         /* Sleep until uncongested or a write happens */
 839         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
 840         ret = io_schedule_timeout(timeout);
 841         finish_wait(wqh, &wait);
 842
 843 out:
 844         trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
 845                                         jiffies_to_usecs(jiffies - start));
 846
 847         return ret;
 848 }
 849 EXPORT_SYMBOL(wait_iff_congested);
 850
 851 int pdflush_proc_obsolete(struct ctl_table *table, int write,
 852                         void __user *buffer, size_t *lenp, loff_t *ppos)
 853 {
 854         char kbuf[] = "0\n";
 855
 856         if (*ppos) {
 857                 *lenp = 0;
 858                 return 0;
 859         }
 860
 861         if (copy_to_user(buffer, kbuf, sizeof(kbuf)))
 862                 return -EFAULT;
 863         printk_once(KERN_WARNING "%s exported in /proc is scheduled for removal\n",
 864                         table->procname);
 865
 866         *lenp = 2;
 867         *ppos += *lenp;
 868         return 2;
 869 }