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