mm/page-writeback.c

   1 /*
   2  * mm/page-writeback.c.
   3  *
   4  * Copyright (C) 2002, Linus Torvalds.
   5  *
   6  * Contains functions related to writing back dirty pages at the
   7  * address_space level.
   8  *
   9  * 10Apr2002    akpm@zip.com.au
  10  *              Initial version
  11  */
  12
  13 #include <linux/kernel.h>
  14 #include <linux/module.h>
  15 #include <linux/spinlock.h>
  16 #include <linux/fs.h>
  17 #include <linux/mm.h>
  18 #include <linux/swap.h>
  19 #include <linux/slab.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/writeback.h>
  22 #include <linux/init.h>
  23 #include <linux/backing-dev.h>
  24 #include <linux/blkdev.h>
  25 #include <linux/mpage.h>
  26 #include <linux/percpu.h>
  27 #include <linux/notifier.h>
  28 #include <linux/smp.h>
  29 #include <linux/sysctl.h>
  30 #include <linux/cpu.h>
  31 #include <linux/syscalls.h>
  32
  33 /*
  34  * The maximum number of pages to writeout in a single bdflush/kupdate
  35  * operation.  We do this so we don't hold I_LOCK against an inode for
  36  * enormous amounts of time, which would block a userspace task which has
  37  * been forced to throttle against that inode.  Also, the code reevaluates
  38  * the dirty each time it has written this many pages.
  39  */
  40 #define MAX_WRITEBACK_PAGES     1024
  41
  42 /*
  43  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
  44  * will look to see if it needs to force writeback or throttling.
  45  */
  46 static long ratelimit_pages = 32;
  47
  48 static long total_pages;        /* The total number of pages in the machine. */
  49 static int dirty_exceeded;      /* Dirty mem may be over limit */
  50
  51 /*
  52  * When balance_dirty_pages decides that the caller needs to perform some
  53  * non-background writeback, this is how many pages it will attempt to write.
  54  * It should be somewhat larger than RATELIMIT_PAGES to ensure that reasonably
  55  * large amounts of I/O are submitted.
  56  */
  57 static inline long sync_writeback_pages(void)
  58 {
  59         return ratelimit_pages + ratelimit_pages / 2;
  60 }
  61
  62 /* The following parameters are exported via /proc/sys/vm */
  63
  64 /*
  65  * Start background writeback (via pdflush) at this percentage
  66  */
  67 int dirty_background_ratio = 10;
  68
  69 /*
  70  * The generator of dirty data starts writeback at this percentage
  71  */
  72 int vm_dirty_ratio = 40;
  73
  74 /*
  75  * The interval between `kupdate'-style writebacks, in centiseconds
  76  * (hundredths of a second)
  77  */
  78 int dirty_writeback_centisecs = 5 * 100;
  79
  80 /*
  81  * The longest number of centiseconds for which data is allowed to remain dirty
  82  */
  83 int dirty_expire_centisecs = 30 * 100;
  84
  85 /*
  86  * Flag that makes the machine dump writes/reads and block dirtyings.
  87  */
  88 int block_dump;
  89
  90 /*
  91  * Flag that puts the machine in "laptop mode".
  92  */
  93 int laptop_mode;
  94
  95 EXPORT_SYMBOL(laptop_mode);
  96
  97 /* End of sysctl-exported parameters */
  98
  99
 100 static void background_writeout(unsigned long _min_pages);
 101
 102 struct writeback_state
 103 {
 104         unsigned long nr_dirty;
 105         unsigned long nr_unstable;
 106         unsigned long nr_mapped;
 107         unsigned long nr_writeback;
 108 };
 109
 110 static void get_writeback_state(struct writeback_state *wbs)
 111 {
 112         wbs->nr_dirty = read_page_state(nr_dirty);
 113         wbs->nr_unstable = read_page_state(nr_unstable);
 114         wbs->nr_mapped = read_page_state(nr_mapped);
 115         wbs->nr_writeback = read_page_state(nr_writeback);
 116 }
 117
 118 /*
 119  * Work out the current dirty-memory clamping and background writeout
 120  * thresholds.
 121  *
 122  * The main aim here is to lower them aggressively if there is a lot of mapped
 123  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
 124  * pages.  It is better to clamp down on writers than to start swapping, and
 125  * performing lots of scanning.
 126  *
 127  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
 128  *
 129  * We don't permit the clamping level to fall below 5% - that is getting rather
 130  * excessive.
 131  *
 132  * We make sure that the background writeout level is below the adjusted
 133  * clamping level.
 134  */
 135 static void
 136 get_dirty_limits(struct writeback_state *wbs, long *pbackground, long *pdirty,
 137                 struct address_space *mapping)
 138 {
 139         int background_ratio;           /* Percentages */
 140         int dirty_ratio;
 141         int unmapped_ratio;
 142         long background;
 143         long dirty;
 144         unsigned long available_memory = total_pages;
 145         struct task_struct *tsk;
 146
 147         get_writeback_state(wbs);
 148
 149 #ifdef CONFIG_HIGHMEM
 150         /*
 151          * If this mapping can only allocate from low memory,
 152          * we exclude high memory from our count.
 153          */
 154         if (mapping && !(mapping_gfp_mask(mapping) & __GFP_HIGHMEM))
 155                 available_memory -= totalhigh_pages;
 156 #endif
 157
 158
 159         unmapped_ratio = 100 - (wbs->nr_mapped * 100) / total_pages;
 160
 161         dirty_ratio = vm_dirty_ratio;
 162         if (dirty_ratio > unmapped_ratio / 2)
 163                 dirty_ratio = unmapped_ratio / 2;
 164
 165         if (dirty_ratio < 5)
 166                 dirty_ratio = 5;
 167
 168         background_ratio = dirty_background_ratio;
 169         if (background_ratio >= dirty_ratio)
 170                 background_ratio = dirty_ratio / 2;
 171
 172         background = (background_ratio * available_memory) / 100;
 173         dirty = (dirty_ratio * available_memory) / 100;
 174         tsk = current;
 175         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
 176                 background += background / 4;
 177                 dirty += dirty / 4;
 178         }
 179         *pbackground = background;
 180         *pdirty = dirty;
 181 }
 182
 183 /*
 184  * balance_dirty_pages() must be called by processes which are generating dirty
 185  * data.  It looks at the number of dirty pages in the machine and will force
 186  * the caller to perform writeback if the system is over `vm_dirty_ratio'.
 187  * If we're over `background_thresh' then pdflush is woken to perform some
 188  * writeout.
 189  */
 190 static void balance_dirty_pages(struct address_space *mapping)
 191 {
 192         struct writeback_state wbs;
 193         long nr_reclaimable;
 194         long background_thresh;
 195         long dirty_thresh;
 196         unsigned long pages_written = 0;
 197         unsigned long write_chunk = sync_writeback_pages();
 198
 199         struct backing_dev_info *bdi = mapping->backing_dev_info;
 200
 201         for (;;) {
 202                 struct writeback_control wbc = {
 203                         .bdi            = bdi,
 204                         .sync_mode      = WB_SYNC_NONE,
 205                         .older_than_this = NULL,
 206                         .nr_to_write    = write_chunk,
 207                 };
 208
 209                 get_dirty_limits(&wbs, &background_thresh,
 210                                         &dirty_thresh, mapping);
 211                 nr_reclaimable = wbs.nr_dirty + wbs.nr_unstable;
 212                 if (nr_reclaimable + wbs.nr_writeback <= dirty_thresh)
 213                         break;
 214
 215                 dirty_exceeded = 1;
 216
 217                 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
 218                  * Unstable writes are a feature of certain networked
 219                  * filesystems (i.e. NFS) in which data may have been
 220                  * written to the server's write cache, but has not yet
 221                  * been flushed to permanent storage.
 222                  */
 223                 if (nr_reclaimable) {
 224                         writeback_inodes(&wbc);
 225                         get_dirty_limits(&wbs, &background_thresh,
 226                                         &dirty_thresh, mapping);
 227                         nr_reclaimable = wbs.nr_dirty + wbs.nr_unstable;
 228                         if (nr_reclaimable + wbs.nr_writeback <= dirty_thresh)
 229                                 break;
 230                         pages_written += write_chunk - wbc.nr_to_write;
 231                         if (pages_written >= write_chunk)
 232                                 break;          /* We've done our duty */
 233                 }
 234                 blk_congestion_wait(WRITE, HZ/10);
 235         }
 236
 237         if (nr_reclaimable + wbs.nr_writeback <= dirty_thresh)
 238                 dirty_exceeded = 0;
 239
 240         if (writeback_in_progress(bdi))
 241                 return;         /* pdflush is already working this queue */
 242
 243         /*
 244          * In laptop mode, we wait until hitting the higher threshold before
 245          * starting background writeout, and then write out all the way down
 246          * to the lower threshold.  So slow writers cause minimal disk activity.
 247          *
 248          * In normal mode, we start background writeout at the lower
 249          * background_thresh, to keep the amount of dirty memory low.
 250          */
 251         if ((laptop_mode && pages_written) ||
 252              (!laptop_mode && (nr_reclaimable > background_thresh)))
 253                 pdflush_operation(background_writeout, 0);
 254 }
 255
 256 /**
 257  * balance_dirty_pages_ratelimited - balance dirty memory state
 258  * @mapping: address_space which was dirtied
 259  *
 260  * Processes which are dirtying memory should call in here once for each page
 261  * which was newly dirtied.  The function will periodically check the system's
 262  * dirty state and will initiate writeback if needed.
 263  *
 264  * On really big machines, get_writeback_state is expensive, so try to avoid
 265  * calling it too often (ratelimiting).  But once we're over the dirty memory
 266  * limit we decrease the ratelimiting by a lot, to prevent individual processes
 267  * from overshooting the limit by (ratelimit_pages) each.
 268  */
 269 void balance_dirty_pages_ratelimited(struct address_space *mapping)
 270 {
 271         static DEFINE_PER_CPU(int, ratelimits) = 0;
 272         long ratelimit;
 273
 274         ratelimit = ratelimit_pages;
 275         if (dirty_exceeded)
 276                 ratelimit = 8;
 277
 278         /*
 279          * Check the rate limiting. Also, we do not want to throttle real-time
 280          * tasks in balance_dirty_pages(). Period.
 281          */
 282         if (get_cpu_var(ratelimits)++ >= ratelimit) {
 283                 __get_cpu_var(ratelimits) = 0;
 284                 put_cpu_var(ratelimits);
 285                 balance_dirty_pages(mapping);
 286                 return;
 287         }
 288         put_cpu_var(ratelimits);
 289 }
 290 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
 291
 292 void throttle_vm_writeout(void)
 293 {
 294         struct writeback_state wbs;
 295         long background_thresh;
 296         long dirty_thresh;
 297
 298         for ( ; ; ) {
 299                 get_dirty_limits(&wbs, &background_thresh, &dirty_thresh, NULL);
 300
 301                 /*
 302                  * Boost the allowable dirty threshold a bit for page
 303                  * allocators so they don't get DoS'ed by heavy writers
 304                  */
 305                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
 306
 307                 if (wbs.nr_unstable + wbs.nr_writeback <= dirty_thresh)
 308                         break;
 309                 blk_congestion_wait(WRITE, HZ/10);
 310         }
 311 }
 312
 313
 314 /*
 315  * writeback at least _min_pages, and keep writing until the amount of dirty
 316  * memory is less than the background threshold, or until we're all clean.
 317  */
 318 static void background_writeout(unsigned long _min_pages)
 319 {
 320         long min_pages = _min_pages;
 321         struct writeback_control wbc = {
 322                 .bdi            = NULL,
 323                 .sync_mode      = WB_SYNC_NONE,
 324                 .older_than_this = NULL,
 325                 .nr_to_write    = 0,
 326                 .nonblocking    = 1,
 327         };
 328
 329         for ( ; ; ) {
 330                 struct writeback_state wbs;
 331                 long background_thresh;
 332                 long dirty_thresh;
 333
 334                 get_dirty_limits(&wbs, &background_thresh, &dirty_thresh, NULL);
 335                 if (wbs.nr_dirty + wbs.nr_unstable < background_thresh
 336                                 && min_pages <= 0)
 337                         break;
 338                 wbc.encountered_congestion = 0;
 339                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
 340                 wbc.pages_skipped = 0;
 341                 writeback_inodes(&wbc);
 342                 min_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
 343                 if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
 344                         /* Wrote less than expected */
 345                         blk_congestion_wait(WRITE, HZ/10);
 346                         if (!wbc.encountered_congestion)
 347                                 break;
 348                 }
 349         }
 350 }
 351
 352 /*
 353  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
 354  * the whole world.  Returns 0 if a pdflush thread was dispatched.  Returns
 355  * -1 if all pdflush threads were busy.
 356  */
 357 int wakeup_pdflush(long nr_pages)
 358 {
 359         if (nr_pages == 0) {
 360                 struct writeback_state wbs;
 361
 362                 get_writeback_state(&wbs);
 363                 nr_pages = wbs.nr_dirty + wbs.nr_unstable;
 364         }
 365         return pdflush_operation(background_writeout, nr_pages);
 366 }
 367
 368 static void wb_timer_fn(unsigned long unused);
 369 static void laptop_timer_fn(unsigned long unused);
 370
 371 static DEFINE_TIMER(wb_timer, wb_timer_fn, 0, 0);
 372 static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
 373
 374 /*
 375  * Periodic writeback of "old" data.
 376  *
 377  * Define "old": the first time one of an inode's pages is dirtied, we mark the
 378  * dirtying-time in the inode's address_space.  So this periodic writeback code
 379  * just walks the superblock inode list, writing back any inodes which are
 380  * older than a specific point in time.
 381  *
 382  * Try to run once per dirty_writeback_centisecs.  But if a writeback event
 383  * takes longer than a dirty_writeback_centisecs interval, then leave a
 384  * one-second gap.
 385  *
 386  * older_than_this takes precedence over nr_to_write.  So we'll only write back
 387  * all dirty pages if they are all attached to "old" mappings.
 388  */
 389 static void wb_kupdate(unsigned long arg)
 390 {
 391         unsigned long oldest_jif;
 392         unsigned long start_jif;
 393         unsigned long next_jif;
 394         long nr_to_write;
 395         struct writeback_state wbs;
 396         struct writeback_control wbc = {
 397                 .bdi            = NULL,
 398                 .sync_mode      = WB_SYNC_NONE,
 399                 .older_than_this = &oldest_jif,
 400                 .nr_to_write    = 0,
 401                 .nonblocking    = 1,
 402                 .for_kupdate    = 1,
 403         };
 404
 405         sync_supers();
 406
 407         get_writeback_state(&wbs);
 408         oldest_jif = jiffies - (dirty_expire_centisecs * HZ) / 100;
 409         start_jif = jiffies;
 410         next_jif = start_jif + (dirty_writeback_centisecs * HZ) / 100;
 411         nr_to_write = wbs.nr_dirty + wbs.nr_unstable +
 412                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
 413         while (nr_to_write > 0) {
 414                 wbc.encountered_congestion = 0;
 415                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
 416                 writeback_inodes(&wbc);
 417                 if (wbc.nr_to_write > 0) {
 418                         if (wbc.encountered_congestion)
 419                                 blk_congestion_wait(WRITE, HZ/10);
 420                         else
 421                                 break;  /* All the old data is written */
 422                 }
 423                 nr_to_write -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
 424         }
 425         if (time_before(next_jif, jiffies + HZ))
 426                 next_jif = jiffies + HZ;
 427         if (dirty_writeback_centisecs)
 428                 mod_timer(&wb_timer, next_jif);
 429 }
 430
 431 /*
 432  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
 433  */
 434 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
 435                 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
 436 {
 437         proc_dointvec(table, write, file, buffer, length, ppos);
 438         if (dirty_writeback_centisecs) {
 439                 mod_timer(&wb_timer,
 440                         jiffies + (dirty_writeback_centisecs * HZ) / 100);
 441         } else {
 442                 del_timer(&wb_timer);
 443         }
 444         return 0;
 445 }
 446
 447 static void wb_timer_fn(unsigned long unused)
 448 {
 449         if (pdflush_operation(wb_kupdate, 0) < 0)
 450                 mod_timer(&wb_timer, jiffies + HZ); /* delay 1 second */
 451 }
 452
 453 static void laptop_flush(unsigned long unused)
 454 {
 455         sys_sync();
 456 }
 457
 458 static void laptop_timer_fn(unsigned long unused)
 459 {
 460         pdflush_operation(laptop_flush, 0);
 461 }
 462
 463 /*
 464  * We've spun up the disk and we're in laptop mode: schedule writeback
 465  * of all dirty data a few seconds from now.  If the flush is already scheduled
 466  * then push it back - the user is still using the disk.
 467  */
 468 void laptop_io_completion(void)
 469 {
 470         mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode * HZ);
 471 }
 472
 473 /*
 474  * We're in laptop mode and we've just synced. The sync's writes will have
 475  * caused another writeback to be scheduled by laptop_io_completion.
 476  * Nothing needs to be written back anymore, so we unschedule the writeback.
 477  */
 478 void laptop_sync_completion(void)
 479 {
 480         del_timer(&laptop_mode_wb_timer);
 481 }
 482
 483 /*
 484  * If ratelimit_pages is too high then we can get into dirty-data overload
 485  * if a large number of processes all perform writes at the same time.
 486  * If it is too low then SMP machines will call the (expensive)
 487  * get_writeback_state too often.
 488  *
 489  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
 490  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
 491  * thresholds before writeback cuts in.
 492  *
 493  * But the limit should not be set too high.  Because it also controls the
 494  * amount of memory which the balance_dirty_pages() caller has to write back.
 495  * If this is too large then the caller will block on the IO queue all the
 496  * time.  So limit it to four megabytes - the balance_dirty_pages() caller
 497  * will write six megabyte chunks, max.
 498  */
 499
 500 static void set_ratelimit(void)
 501 {
 502         ratelimit_pages = total_pages / (num_online_cpus() * 32);
 503         if (ratelimit_pages < 16)
 504                 ratelimit_pages = 16;
 505         if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
 506                 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
 507 }
 508
 509 static int
 510 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
 511 {
 512         set_ratelimit();
 513         return 0;
 514 }
 515
 516 static struct notifier_block ratelimit_nb = {
 517         .notifier_call  = ratelimit_handler,
 518         .next           = NULL,
 519 };
 520
 521 /*
 522  * If the machine has a large highmem:lowmem ratio then scale back the default
 523  * dirty memory thresholds: allowing too much dirty highmem pins an excessive
 524  * number of buffer_heads.
 525  */
 526 void __init page_writeback_init(void)
 527 {
 528         long buffer_pages = nr_free_buffer_pages();
 529         long correction;
 530
 531         total_pages = nr_free_pagecache_pages();
 532
 533         correction = (100 * 4 * buffer_pages) / total_pages;
 534
 535         if (correction < 100) {
 536                 dirty_background_ratio *= correction;
 537                 dirty_background_ratio /= 100;
 538                 vm_dirty_ratio *= correction;
 539                 vm_dirty_ratio /= 100;
 540
 541                 if (dirty_background_ratio <= 0)
 542                         dirty_background_ratio = 1;
 543                 if (vm_dirty_ratio <= 0)
 544                         vm_dirty_ratio = 1;
 545         }
 546         mod_timer(&wb_timer, jiffies + (dirty_writeback_centisecs * HZ) / 100);
 547         set_ratelimit();
 548         register_cpu_notifier(&ratelimit_nb);
 549 }
 550
 551 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
 552 {
 553         int ret;
 554
 555         if (wbc->nr_to_write <= 0)
 556                 return 0;
 557         wbc->for_writepages = 1;
 558         if (mapping->a_ops->writepages)
 559                 ret =  mapping->a_ops->writepages(mapping, wbc);
 560         else
 561                 ret = generic_writepages(mapping, wbc);
 562         wbc->for_writepages = 0;
 563         return ret;
 564 }
 565
 566 /**
 567  * write_one_page - write out a single page and optionally wait on I/O
 568  *
 569  * @page: the page to write
 570  * @wait: if true, wait on writeout
 571  *
 572  * The page must be locked by the caller and will be unlocked upon return.
 573  *
 574  * write_one_page() returns a negative error code if I/O failed.
 575  */
 576 int write_one_page(struct page *page, int wait)
 577 {
 578         struct address_space *mapping = page->mapping;
 579         int ret = 0;
 580         struct writeback_control wbc = {
 581                 .sync_mode = WB_SYNC_ALL,
 582                 .nr_to_write = 1,
 583         };
 584
 585         BUG_ON(!PageLocked(page));
 586
 587         if (wait)
 588                 wait_on_page_writeback(page);
 589
 590         if (clear_page_dirty_for_io(page)) {
 591                 page_cache_get(page);
 592                 ret = mapping->a_ops->writepage(page, &wbc);
 593                 if (ret == 0 && wait) {
 594                         wait_on_page_writeback(page);
 595                         if (PageError(page))
 596                                 ret = -EIO;
 597                 }
 598                 page_cache_release(page);
 599         } else {
 600                 unlock_page(page);
 601         }
 602         return ret;
 603 }
 604 EXPORT_SYMBOL(write_one_page);
 605
 606 /*
 607  * For address_spaces which do not use buffers.  Just tag the page as dirty in
 608  * its radix tree.
 609  *
 610  * This is also used when a single buffer is being dirtied: we want to set the
 611  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
 612  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
 613  *
 614  * Most callers have locked the page, which pins the address_space in memory.
 615  * But zap_pte_range() does not lock the page, however in that case the
 616  * mapping is pinned by the vma's ->vm_file reference.
 617  *
 618  * We take care to handle the case where the page was truncated from the
 619  * mapping by re-checking page_mapping() insode tree_lock.
 620  */
 621 int __set_page_dirty_nobuffers(struct page *page)
 622 {
 623         int ret = 0;
 624
 625         if (!TestSetPageDirty(page)) {
 626                 struct address_space *mapping = page_mapping(page);
 627                 struct address_space *mapping2;
 628
 629                 if (mapping) {
 630                         write_lock_irq(&mapping->tree_lock);
 631                         mapping2 = page_mapping(page);
 632                         if (mapping2) { /* Race with truncate? */
 633                                 BUG_ON(mapping2 != mapping);
 634                                 if (mapping_cap_account_dirty(mapping))
 635                                         inc_page_state(nr_dirty);
 636                                 radix_tree_tag_set(&mapping->page_tree,
 637                                         page_index(page), PAGECACHE_TAG_DIRTY);
 638                         }
 639                         write_unlock_irq(&mapping->tree_lock);
 640                         if (mapping->host) {
 641                                 /* !PageAnon && !swapper_space */
 642                                 __mark_inode_dirty(mapping->host,
 643                                                         I_DIRTY_PAGES);
 644                         }
 645                 }
 646         }
 647         return ret;
 648 }
 649 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
 650
 651 /*
 652  * When a writepage implementation decides that it doesn't want to write this
 653  * page for some reason, it should redirty the locked page via
 654  * redirty_page_for_writepage() and it should then unlock the page and return 0
 655  */
 656 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
 657 {
 658         wbc->pages_skipped++;
 659         return __set_page_dirty_nobuffers(page);
 660 }
 661 EXPORT_SYMBOL(redirty_page_for_writepage);
 662
 663 /*
 664  * If the mapping doesn't provide a set_page_dirty a_op, then
 665  * just fall through and assume that it wants buffer_heads.
 666  */
 667 int fastcall set_page_dirty(struct page *page)
 668 {
 669         struct address_space *mapping = page_mapping(page);
 670
 671         if (likely(mapping)) {
 672                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
 673                 if (spd)
 674                         return (*spd)(page);
 675                 return __set_page_dirty_buffers(page);
 676         }
 677         if (!PageDirty(page))
 678                 SetPageDirty(page);
 679         return 0;
 680 }
 681 EXPORT_SYMBOL(set_page_dirty);
 682
 683 /*
 684  * set_page_dirty() is racy if the caller has no reference against
 685  * page->mapping->host, and if the page is unlocked.  This is because another
 686  * CPU could truncate the page off the mapping and then free the mapping.
 687  *
 688  * Usually, the page _is_ locked, or the caller is a user-space process which
 689  * holds a reference on the inode by having an open file.
 690  *
 691  * In other cases, the page should be locked before running set_page_dirty().
 692  */
 693 int set_page_dirty_lock(struct page *page)
 694 {
 695         int ret;
 696
 697         lock_page(page);
 698         ret = set_page_dirty(page);
 699         unlock_page(page);
 700         return ret;
 701 }
 702 EXPORT_SYMBOL(set_page_dirty_lock);
 703
 704 /*
 705  * Clear a page's dirty flag, while caring for dirty memory accounting.
 706  * Returns true if the page was previously dirty.
 707  */
 708 int test_clear_page_dirty(struct page *page)
 709 {
 710         struct address_space *mapping = page_mapping(page);
 711         unsigned long flags;
 712
 713         if (mapping) {
 714                 write_lock_irqsave(&mapping->tree_lock, flags);
 715                 if (TestClearPageDirty(page)) {
 716                         radix_tree_tag_clear(&mapping->page_tree,
 717                                                 page_index(page),
 718                                                 PAGECACHE_TAG_DIRTY);
 719                         write_unlock_irqrestore(&mapping->tree_lock, flags);
 720                         if (mapping_cap_account_dirty(mapping))
 721                                 dec_page_state(nr_dirty);
 722                         return 1;
 723                 }
 724                 write_unlock_irqrestore(&mapping->tree_lock, flags);
 725                 return 0;
 726         }
 727         return TestClearPageDirty(page);
 728 }
 729 EXPORT_SYMBOL(test_clear_page_dirty);
 730
 731 /*
 732  * Clear a page's dirty flag, while caring for dirty memory accounting.
 733  * Returns true if the page was previously dirty.
 734  *
 735  * This is for preparing to put the page under writeout.  We leave the page
 736  * tagged as dirty in the radix tree so that a concurrent write-for-sync
 737  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
 738  * implementation will run either set_page_writeback() or set_page_dirty(),
 739  * at which stage we bring the page's dirty flag and radix-tree dirty tag
 740  * back into sync.
 741  *
 742  * This incoherency between the page's dirty flag and radix-tree tag is
 743  * unfortunate, but it only exists while the page is locked.
 744  */
 745 int clear_page_dirty_for_io(struct page *page)
 746 {
 747         struct address_space *mapping = page_mapping(page);
 748
 749         if (mapping) {
 750                 if (TestClearPageDirty(page)) {
 751                         if (mapping_cap_account_dirty(mapping))
 752                                 dec_page_state(nr_dirty);
 753                         return 1;
 754                 }
 755                 return 0;
 756         }
 757         return TestClearPageDirty(page);
 758 }
 759 EXPORT_SYMBOL(clear_page_dirty_for_io);
 760
 761 int test_clear_page_writeback(struct page *page)
 762 {
 763         struct address_space *mapping = page_mapping(page);
 764         int ret;
 765
 766         if (mapping) {
 767                 unsigned long flags;
 768
 769                 write_lock_irqsave(&mapping->tree_lock, flags);
 770                 ret = TestClearPageWriteback(page);
 771                 if (ret)
 772                         radix_tree_tag_clear(&mapping->page_tree,
 773                                                 page_index(page),
 774                                                 PAGECACHE_TAG_WRITEBACK);
 775                 write_unlock_irqrestore(&mapping->tree_lock, flags);
 776         } else {
 777                 ret = TestClearPageWriteback(page);
 778         }
 779         return ret;
 780 }
 781
 782 int test_set_page_writeback(struct page *page)
 783 {
 784         struct address_space *mapping = page_mapping(page);
 785         int ret;
 786
 787         if (mapping) {
 788                 unsigned long flags;
 789
 790                 write_lock_irqsave(&mapping->tree_lock, flags);
 791                 ret = TestSetPageWriteback(page);
 792                 if (!ret)
 793                         radix_tree_tag_set(&mapping->page_tree,
 794                                                 page_index(page),
 795                                                 PAGECACHE_TAG_WRITEBACK);
 796                 if (!PageDirty(page))
 797                         radix_tree_tag_clear(&mapping->page_tree,
 798                                                 page_index(page),
 799                                                 PAGECACHE_TAG_DIRTY);
 800                 write_unlock_irqrestore(&mapping->tree_lock, flags);
 801         } else {
 802                 ret = TestSetPageWriteback(page);
 803         }
 804         return ret;
 805
 806 }
 807 EXPORT_SYMBOL(test_set_page_writeback);
 808
 809 /*
 810  * Return true if any of the pages in the mapping are marged with the
 811  * passed tag.
 812  */
 813 int mapping_tagged(struct address_space *mapping, int tag)
 814 {
 815         unsigned long flags;
 816         int ret;
 817
 818         read_lock_irqsave(&mapping->tree_lock, flags);
 819         ret = radix_tree_tagged(&mapping->page_tree, tag);
 820         read_unlock_irqrestore(&mapping->tree_lock, flags);
 821         return ret;
 822 }
 823 EXPORT_SYMBOL(mapping_tagged);