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 __cacheline_aligned_in_smp;   /* 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                 if (!dirty_exceeded)
 216                         dirty_exceeded = 1;
 217
 218                 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
 219                  * Unstable writes are a feature of certain networked
 220                  * filesystems (i.e. NFS) in which data may have been
 221                  * written to the server's write cache, but has not yet
 222                  * been flushed to permanent storage.
 223                  */
 224                 if (nr_reclaimable) {
 225                         writeback_inodes(&wbc);
 226                         get_dirty_limits(&wbs, &background_thresh,
 227                                         &dirty_thresh, mapping);
 228                         nr_reclaimable = wbs.nr_dirty + wbs.nr_unstable;
 229                         if (nr_reclaimable + wbs.nr_writeback <= dirty_thresh)
 230                                 break;
 231                         pages_written += write_chunk - wbc.nr_to_write;
 232                         if (pages_written >= write_chunk)
 233                                 break;          /* We've done our duty */
 234                 }
 235                 blk_congestion_wait(WRITE, HZ/10);
 236         }
 237
 238         if (nr_reclaimable + wbs.nr_writeback <= dirty_thresh && dirty_exceeded)
 239                 dirty_exceeded = 0;
 240
 241         if (writeback_in_progress(bdi))
 242                 return;         /* pdflush is already working this queue */
 243
 244         /*
 245          * In laptop mode, we wait until hitting the higher threshold before
 246          * starting background writeout, and then write out all the way down
 247          * to the lower threshold.  So slow writers cause minimal disk activity.
 248          *
 249          * In normal mode, we start background writeout at the lower
 250          * background_thresh, to keep the amount of dirty memory low.
 251          */
 252         if ((laptop_mode && pages_written) ||
 253              (!laptop_mode && (nr_reclaimable > background_thresh)))
 254                 pdflush_operation(background_writeout, 0);
 255 }
 256
 257 /**
 258  * balance_dirty_pages_ratelimited - balance dirty memory state
 259  * @mapping: address_space which was dirtied
 260  *
 261  * Processes which are dirtying memory should call in here once for each page
 262  * which was newly dirtied.  The function will periodically check the system's
 263  * dirty state and will initiate writeback if needed.
 264  *
 265  * On really big machines, get_writeback_state is expensive, so try to avoid
 266  * calling it too often (ratelimiting).  But once we're over the dirty memory
 267  * limit we decrease the ratelimiting by a lot, to prevent individual processes
 268  * from overshooting the limit by (ratelimit_pages) each.
 269  */
 270 void balance_dirty_pages_ratelimited(struct address_space *mapping)
 271 {
 272         static DEFINE_PER_CPU(int, ratelimits) = 0;
 273         long ratelimit;
 274
 275         ratelimit = ratelimit_pages;
 276         if (dirty_exceeded)
 277                 ratelimit = 8;
 278
 279         /*
 280          * Check the rate limiting. Also, we do not want to throttle real-time
 281          * tasks in balance_dirty_pages(). Period.
 282          */
 283         if (get_cpu_var(ratelimits)++ >= ratelimit) {
 284                 __get_cpu_var(ratelimits) = 0;
 285                 put_cpu_var(ratelimits);
 286                 balance_dirty_pages(mapping);
 287                 return;
 288         }
 289         put_cpu_var(ratelimits);
 290 }
 291 EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
 292
 293 void throttle_vm_writeout(void)
 294 {
 295         struct writeback_state wbs;
 296         long background_thresh;
 297         long dirty_thresh;
 298
 299         for ( ; ; ) {
 300                 get_dirty_limits(&wbs, &background_thresh, &dirty_thresh, NULL);
 301
 302                 /*
 303                  * Boost the allowable dirty threshold a bit for page
 304                  * allocators so they don't get DoS'ed by heavy writers
 305                  */
 306                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
 307
 308                 if (wbs.nr_unstable + wbs.nr_writeback <= dirty_thresh)
 309                         break;
 310                 blk_congestion_wait(WRITE, HZ/10);
 311         }
 312 }
 313
 314
 315 /*
 316  * writeback at least _min_pages, and keep writing until the amount of dirty
 317  * memory is less than the background threshold, or until we're all clean.
 318  */
 319 static void background_writeout(unsigned long _min_pages)
 320 {
 321         long min_pages = _min_pages;
 322         struct writeback_control wbc = {
 323                 .bdi            = NULL,
 324                 .sync_mode      = WB_SYNC_NONE,
 325                 .older_than_this = NULL,
 326                 .nr_to_write    = 0,
 327                 .nonblocking    = 1,
 328         };
 329
 330         for ( ; ; ) {
 331                 struct writeback_state wbs;
 332                 long background_thresh;
 333                 long dirty_thresh;
 334
 335                 get_dirty_limits(&wbs, &background_thresh, &dirty_thresh, NULL);
 336                 if (wbs.nr_dirty + wbs.nr_unstable < background_thresh
 337                                 && min_pages <= 0)
 338                         break;
 339                 wbc.encountered_congestion = 0;
 340                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
 341                 wbc.pages_skipped = 0;
 342                 writeback_inodes(&wbc);
 343                 min_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
 344                 if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
 345                         /* Wrote less than expected */
 346                         blk_congestion_wait(WRITE, HZ/10);
 347                         if (!wbc.encountered_congestion)
 348                                 break;
 349                 }
 350         }
 351 }
 352
 353 /*
 354  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
 355  * the whole world.  Returns 0 if a pdflush thread was dispatched.  Returns
 356  * -1 if all pdflush threads were busy.
 357  */
 358 int wakeup_pdflush(long nr_pages)
 359 {
 360         if (nr_pages == 0) {
 361                 struct writeback_state wbs;
 362
 363                 get_writeback_state(&wbs);
 364                 nr_pages = wbs.nr_dirty + wbs.nr_unstable;
 365         }
 366         return pdflush_operation(background_writeout, nr_pages);
 367 }
 368
 369 static void wb_timer_fn(unsigned long unused);
 370 static void laptop_timer_fn(unsigned long unused);
 371
 372 static DEFINE_TIMER(wb_timer, wb_timer_fn, 0, 0);
 373 static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
 374
 375 /*
 376  * Periodic writeback of "old" data.
 377  *
 378  * Define "old": the first time one of an inode's pages is dirtied, we mark the
 379  * dirtying-time in the inode's address_space.  So this periodic writeback code
 380  * just walks the superblock inode list, writing back any inodes which are
 381  * older than a specific point in time.
 382  *
 383  * Try to run once per dirty_writeback_centisecs.  But if a writeback event
 384  * takes longer than a dirty_writeback_centisecs interval, then leave a
 385  * one-second gap.
 386  *
 387  * older_than_this takes precedence over nr_to_write.  So we'll only write back
 388  * all dirty pages if they are all attached to "old" mappings.
 389  */
 390 static void wb_kupdate(unsigned long arg)
 391 {
 392         unsigned long oldest_jif;
 393         unsigned long start_jif;
 394         unsigned long next_jif;
 395         long nr_to_write;
 396         struct writeback_state wbs;
 397         struct writeback_control wbc = {
 398                 .bdi            = NULL,
 399                 .sync_mode      = WB_SYNC_NONE,
 400                 .older_than_this = &oldest_jif,
 401                 .nr_to_write    = 0,
 402                 .nonblocking    = 1,
 403                 .for_kupdate    = 1,
 404         };
 405
 406         sync_supers();
 407
 408         get_writeback_state(&wbs);
 409         oldest_jif = jiffies - (dirty_expire_centisecs * HZ) / 100;
 410         start_jif = jiffies;
 411         next_jif = start_jif + (dirty_writeback_centisecs * HZ) / 100;
 412         nr_to_write = wbs.nr_dirty + wbs.nr_unstable +
 413                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
 414         while (nr_to_write > 0) {
 415                 wbc.encountered_congestion = 0;
 416                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
 417                 writeback_inodes(&wbc);
 418                 if (wbc.nr_to_write > 0) {
 419                         if (wbc.encountered_congestion)
 420                                 blk_congestion_wait(WRITE, HZ/10);
 421                         else
 422                                 break;  /* All the old data is written */
 423                 }
 424                 nr_to_write -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
 425         }
 426         if (time_before(next_jif, jiffies + HZ))
 427                 next_jif = jiffies + HZ;
 428         if (dirty_writeback_centisecs)
 429                 mod_timer(&wb_timer, next_jif);
 430 }
 431
 432 /*
 433  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
 434  */
 435 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
 436                 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
 437 {
 438         proc_dointvec(table, write, file, buffer, length, ppos);
 439         if (dirty_writeback_centisecs) {
 440                 mod_timer(&wb_timer,
 441                         jiffies + (dirty_writeback_centisecs * HZ) / 100);
 442         } else {
 443                 del_timer(&wb_timer);
 444         }
 445         return 0;
 446 }
 447
 448 static void wb_timer_fn(unsigned long unused)
 449 {
 450         if (pdflush_operation(wb_kupdate, 0) < 0)
 451                 mod_timer(&wb_timer, jiffies + HZ); /* delay 1 second */
 452 }
 453
 454 static void laptop_flush(unsigned long unused)
 455 {
 456         sys_sync();
 457 }
 458
 459 static void laptop_timer_fn(unsigned long unused)
 460 {
 461         pdflush_operation(laptop_flush, 0);
 462 }
 463
 464 /*
 465  * We've spun up the disk and we're in laptop mode: schedule writeback
 466  * of all dirty data a few seconds from now.  If the flush is already scheduled
 467  * then push it back - the user is still using the disk.
 468  */
 469 void laptop_io_completion(void)
 470 {
 471         mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode * HZ);
 472 }
 473
 474 /*
 475  * We're in laptop mode and we've just synced. The sync's writes will have
 476  * caused another writeback to be scheduled by laptop_io_completion.
 477  * Nothing needs to be written back anymore, so we unschedule the writeback.
 478  */
 479 void laptop_sync_completion(void)
 480 {
 481         del_timer(&laptop_mode_wb_timer);
 482 }
 483
 484 /*
 485  * If ratelimit_pages is too high then we can get into dirty-data overload
 486  * if a large number of processes all perform writes at the same time.
 487  * If it is too low then SMP machines will call the (expensive)
 488  * get_writeback_state too often.
 489  *
 490  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
 491  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
 492  * thresholds before writeback cuts in.
 493  *
 494  * But the limit should not be set too high.  Because it also controls the
 495  * amount of memory which the balance_dirty_pages() caller has to write back.
 496  * If this is too large then the caller will block on the IO queue all the
 497  * time.  So limit it to four megabytes - the balance_dirty_pages() caller
 498  * will write six megabyte chunks, max.
 499  */
 500
 501 static void set_ratelimit(void)
 502 {
 503         ratelimit_pages = total_pages / (num_online_cpus() * 32);
 504         if (ratelimit_pages < 16)
 505                 ratelimit_pages = 16;
 506         if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
 507                 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
 508 }
 509
 510 static int
 511 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
 512 {
 513         set_ratelimit();
 514         return 0;
 515 }
 516
 517 static struct notifier_block ratelimit_nb = {
 518         .notifier_call  = ratelimit_handler,
 519         .next           = NULL,
 520 };
 521
 522 /*
 523  * If the machine has a large highmem:lowmem ratio then scale back the default
 524  * dirty memory thresholds: allowing too much dirty highmem pins an excessive
 525  * number of buffer_heads.
 526  */
 527 void __init page_writeback_init(void)
 528 {
 529         long buffer_pages = nr_free_buffer_pages();
 530         long correction;
 531
 532         total_pages = nr_free_pagecache_pages();
 533
 534         correction = (100 * 4 * buffer_pages) / total_pages;
 535
 536         if (correction < 100) {
 537                 dirty_background_ratio *= correction;
 538                 dirty_background_ratio /= 100;
 539                 vm_dirty_ratio *= correction;
 540                 vm_dirty_ratio /= 100;
 541
 542                 if (dirty_background_ratio <= 0)
 543                         dirty_background_ratio = 1;
 544                 if (vm_dirty_ratio <= 0)
 545                         vm_dirty_ratio = 1;
 546         }
 547         mod_timer(&wb_timer, jiffies + (dirty_writeback_centisecs * HZ) / 100);
 548         set_ratelimit();
 549         register_cpu_notifier(&ratelimit_nb);
 550 }
 551
 552 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
 553 {
 554         int ret;
 555
 556         if (wbc->nr_to_write <= 0)
 557                 return 0;
 558         wbc->for_writepages = 1;
 559         if (mapping->a_ops->writepages)
 560                 ret =  mapping->a_ops->writepages(mapping, wbc);
 561         else
 562                 ret = generic_writepages(mapping, wbc);
 563         wbc->for_writepages = 0;
 564         return ret;
 565 }
 566
 567 /**
 568  * write_one_page - write out a single page and optionally wait on I/O
 569  *
 570  * @page: the page to write
 571  * @wait: if true, wait on writeout
 572  *
 573  * The page must be locked by the caller and will be unlocked upon return.
 574  *
 575  * write_one_page() returns a negative error code if I/O failed.
 576  */
 577 int write_one_page(struct page *page, int wait)
 578 {
 579         struct address_space *mapping = page->mapping;
 580         int ret = 0;
 581         struct writeback_control wbc = {
 582                 .sync_mode = WB_SYNC_ALL,
 583                 .nr_to_write = 1,
 584         };
 585
 586         BUG_ON(!PageLocked(page));
 587
 588         if (wait)
 589                 wait_on_page_writeback(page);
 590
 591         if (clear_page_dirty_for_io(page)) {
 592                 page_cache_get(page);
 593                 ret = mapping->a_ops->writepage(page, &wbc);
 594                 if (ret == 0 && wait) {
 595                         wait_on_page_writeback(page);
 596                         if (PageError(page))
 597                                 ret = -EIO;
 598                 }
 599                 page_cache_release(page);
 600         } else {
 601                 unlock_page(page);
 602         }
 603         return ret;
 604 }
 605 EXPORT_SYMBOL(write_one_page);
 606
 607 /*
 608  * For address_spaces which do not use buffers.  Just tag the page as dirty in
 609  * its radix tree.
 610  *
 611  * This is also used when a single buffer is being dirtied: we want to set the
 612  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
 613  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
 614  *
 615  * Most callers have locked the page, which pins the address_space in memory.
 616  * But zap_pte_range() does not lock the page, however in that case the
 617  * mapping is pinned by the vma's ->vm_file reference.
 618  *
 619  * We take care to handle the case where the page was truncated from the
 620  * mapping by re-checking page_mapping() insode tree_lock.
 621  */
 622 int __set_page_dirty_nobuffers(struct page *page)
 623 {
 624         int ret = 0;
 625
 626         if (!TestSetPageDirty(page)) {
 627                 struct address_space *mapping = page_mapping(page);
 628                 struct address_space *mapping2;
 629
 630                 if (mapping) {
 631                         write_lock_irq(&mapping->tree_lock);
 632                         mapping2 = page_mapping(page);
 633                         if (mapping2) { /* Race with truncate? */
 634                                 BUG_ON(mapping2 != mapping);
 635                                 if (mapping_cap_account_dirty(mapping))
 636                                         inc_page_state(nr_dirty);
 637                                 radix_tree_tag_set(&mapping->page_tree,
 638                                         page_index(page), PAGECACHE_TAG_DIRTY);
 639                         }
 640                         write_unlock_irq(&mapping->tree_lock);
 641                         if (mapping->host) {
 642                                 /* !PageAnon && !swapper_space */
 643                                 __mark_inode_dirty(mapping->host,
 644                                                         I_DIRTY_PAGES);
 645                         }
 646                 }
 647         }
 648         return ret;
 649 }
 650 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
 651
 652 /*
 653  * When a writepage implementation decides that it doesn't want to write this
 654  * page for some reason, it should redirty the locked page via
 655  * redirty_page_for_writepage() and it should then unlock the page and return 0
 656  */
 657 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
 658 {
 659         wbc->pages_skipped++;
 660         return __set_page_dirty_nobuffers(page);
 661 }
 662 EXPORT_SYMBOL(redirty_page_for_writepage);
 663
 664 /*
 665  * If the mapping doesn't provide a set_page_dirty a_op, then
 666  * just fall through and assume that it wants buffer_heads.
 667  */
 668 int fastcall set_page_dirty(struct page *page)
 669 {
 670         struct address_space *mapping = page_mapping(page);
 671
 672         if (likely(mapping)) {
 673                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
 674                 if (spd)
 675                         return (*spd)(page);
 676                 return __set_page_dirty_buffers(page);
 677         }
 678         if (!PageDirty(page))
 679                 SetPageDirty(page);
 680         return 0;
 681 }
 682 EXPORT_SYMBOL(set_page_dirty);
 683
 684 /*
 685  * set_page_dirty() is racy if the caller has no reference against
 686  * page->mapping->host, and if the page is unlocked.  This is because another
 687  * CPU could truncate the page off the mapping and then free the mapping.
 688  *
 689  * Usually, the page _is_ locked, or the caller is a user-space process which
 690  * holds a reference on the inode by having an open file.
 691  *
 692  * In other cases, the page should be locked before running set_page_dirty().
 693  */
 694 int set_page_dirty_lock(struct page *page)
 695 {
 696         int ret;
 697
 698         lock_page(page);
 699         ret = set_page_dirty(page);
 700         unlock_page(page);
 701         return ret;
 702 }
 703 EXPORT_SYMBOL(set_page_dirty_lock);
 704
 705 /*
 706  * Clear a page's dirty flag, while caring for dirty memory accounting.
 707  * Returns true if the page was previously dirty.
 708  */
 709 int test_clear_page_dirty(struct page *page)
 710 {
 711         struct address_space *mapping = page_mapping(page);
 712         unsigned long flags;
 713
 714         if (mapping) {
 715                 write_lock_irqsave(&mapping->tree_lock, flags);
 716                 if (TestClearPageDirty(page)) {
 717                         radix_tree_tag_clear(&mapping->page_tree,
 718                                                 page_index(page),
 719                                                 PAGECACHE_TAG_DIRTY);
 720                         write_unlock_irqrestore(&mapping->tree_lock, flags);
 721                         if (mapping_cap_account_dirty(mapping))
 722                                 dec_page_state(nr_dirty);
 723                         return 1;
 724                 }
 725                 write_unlock_irqrestore(&mapping->tree_lock, flags);
 726                 return 0;
 727         }
 728         return TestClearPageDirty(page);
 729 }
 730 EXPORT_SYMBOL(test_clear_page_dirty);
 731
 732 /*
 733  * Clear a page's dirty flag, while caring for dirty memory accounting.
 734  * Returns true if the page was previously dirty.
 735  *
 736  * This is for preparing to put the page under writeout.  We leave the page
 737  * tagged as dirty in the radix tree so that a concurrent write-for-sync
 738  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
 739  * implementation will run either set_page_writeback() or set_page_dirty(),
 740  * at which stage we bring the page's dirty flag and radix-tree dirty tag
 741  * back into sync.
 742  *
 743  * This incoherency between the page's dirty flag and radix-tree tag is
 744  * unfortunate, but it only exists while the page is locked.
 745  */
 746 int clear_page_dirty_for_io(struct page *page)
 747 {
 748         struct address_space *mapping = page_mapping(page);
 749
 750         if (mapping) {
 751                 if (TestClearPageDirty(page)) {
 752                         if (mapping_cap_account_dirty(mapping))
 753                                 dec_page_state(nr_dirty);
 754                         return 1;
 755                 }
 756                 return 0;
 757         }
 758         return TestClearPageDirty(page);
 759 }
 760 EXPORT_SYMBOL(clear_page_dirty_for_io);
 761
 762 int test_clear_page_writeback(struct page *page)
 763 {
 764         struct address_space *mapping = page_mapping(page);
 765         int ret;
 766
 767         if (mapping) {
 768                 unsigned long flags;
 769
 770                 write_lock_irqsave(&mapping->tree_lock, flags);
 771                 ret = TestClearPageWriteback(page);
 772                 if (ret)
 773                         radix_tree_tag_clear(&mapping->page_tree,
 774                                                 page_index(page),
 775                                                 PAGECACHE_TAG_WRITEBACK);
 776                 write_unlock_irqrestore(&mapping->tree_lock, flags);
 777         } else {
 778                 ret = TestClearPageWriteback(page);
 779         }
 780         return ret;
 781 }
 782
 783 int test_set_page_writeback(struct page *page)
 784 {
 785         struct address_space *mapping = page_mapping(page);
 786         int ret;
 787
 788         if (mapping) {
 789                 unsigned long flags;
 790
 791                 write_lock_irqsave(&mapping->tree_lock, flags);
 792                 ret = TestSetPageWriteback(page);
 793                 if (!ret)
 794                         radix_tree_tag_set(&mapping->page_tree,
 795                                                 page_index(page),
 796                                                 PAGECACHE_TAG_WRITEBACK);
 797                 if (!PageDirty(page))
 798                         radix_tree_tag_clear(&mapping->page_tree,
 799                                                 page_index(page),
 800                                                 PAGECACHE_TAG_DIRTY);
 801                 write_unlock_irqrestore(&mapping->tree_lock, flags);
 802         } else {
 803                 ret = TestSetPageWriteback(page);
 804         }
 805         return ret;
 806
 807 }
 808 EXPORT_SYMBOL(test_set_page_writeback);
 809
 810 /*
 811  * Return true if any of the pages in the mapping are marged with the
 812  * passed tag.
 813  */
 814 int mapping_tagged(struct address_space *mapping, int tag)
 815 {
 816         unsigned long flags;
 817         int ret;
 818
 819         read_lock_irqsave(&mapping->tree_lock, flags);
 820         ret = radix_tree_tagged(&mapping->page_tree, tag);
 821         read_unlock_irqrestore(&mapping->tree_lock, flags);
 822         return ret;
 823 }
 824 EXPORT_SYMBOL(mapping_tagged);