linux/mm/page-writeback.c
<<
>>
Prefs
   1/*
   2 * mm/page-writeback.c
   3 *
   4 * Copyright (C) 2002, Linus Torvalds.
   5 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
   6 *
   7 * Contains functions related to writing back dirty pages at the
   8 * address_space level.
   9 *
  10 * 10Apr2002    Andrew Morton
  11 *              Initial version
  12 */
  13
  14#include <linux/kernel.h>
  15#include <linux/module.h>
  16#include <linux/spinlock.h>
  17#include <linux/fs.h>
  18#include <linux/mm.h>
  19#include <linux/swap.h>
  20#include <linux/slab.h>
  21#include <linux/pagemap.h>
  22#include <linux/writeback.h>
  23#include <linux/init.h>
  24#include <linux/backing-dev.h>
  25#include <linux/task_io_accounting_ops.h>
  26#include <linux/blkdev.h>
  27#include <linux/mpage.h>
  28#include <linux/rmap.h>
  29#include <linux/percpu.h>
  30#include <linux/notifier.h>
  31#include <linux/smp.h>
  32#include <linux/sysctl.h>
  33#include <linux/cpu.h>
  34#include <linux/syscalls.h>
  35#include <linux/buffer_head.h>
  36#include <linux/pagevec.h>
  37
  38/*
  39 * The maximum number of pages to writeout in a single bdflush/kupdate
  40 * operation.  We do this so we don't hold I_SYNC against an inode for
  41 * enormous amounts of time, which would block a userspace task which has
  42 * been forced to throttle against that inode.  Also, the code reevaluates
  43 * the dirty each time it has written this many pages.
  44 */
  45#define MAX_WRITEBACK_PAGES     1024
  46
  47/*
  48 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
  49 * will look to see if it needs to force writeback or throttling.
  50 */
  51static long ratelimit_pages = 32;
  52
  53/*
  54 * When balance_dirty_pages decides that the caller needs to perform some
  55 * non-background writeback, this is how many pages it will attempt to write.
  56 * It should be somewhat larger than RATELIMIT_PAGES to ensure that reasonably
  57 * large amounts of I/O are submitted.
  58 */
  59static inline long sync_writeback_pages(void)
  60{
  61        return ratelimit_pages + ratelimit_pages / 2;
  62}
  63
  64/* The following parameters are exported via /proc/sys/vm */
  65
  66/*
  67 * Start background writeback (via pdflush) at this percentage
  68 */
  69int dirty_background_ratio = 10;
  70
  71/*
  72 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
  73 * dirty_background_ratio * the amount of dirtyable memory
  74 */
  75unsigned long dirty_background_bytes;
  76
  77/*
  78 * free highmem will not be subtracted from the total free memory
  79 * for calculating free ratios if vm_highmem_is_dirtyable is true
  80 */
  81int vm_highmem_is_dirtyable;
  82
  83/*
  84 * The generator of dirty data starts writeback at this percentage
  85 */
  86int vm_dirty_ratio = 20;
  87
  88/*
  89 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
  90 * vm_dirty_ratio * the amount of dirtyable memory
  91 */
  92unsigned long vm_dirty_bytes;
  93
  94/*
  95 * The interval between `kupdate'-style writebacks
  96 */
  97unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
  98
  99/*
 100 * The longest time for which data is allowed to remain dirty
 101 */
 102unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
 103
 104/*
 105 * Flag that makes the machine dump writes/reads and block dirtyings.
 106 */
 107int block_dump;
 108
 109/*
 110 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
 111 * a full sync is triggered after this time elapses without any disk activity.
 112 */
 113int laptop_mode;
 114
 115EXPORT_SYMBOL(laptop_mode);
 116
 117/* End of sysctl-exported parameters */
 118
 119
 120static void background_writeout(unsigned long _min_pages);
 121
 122/*
 123 * Scale the writeback cache size proportional to the relative writeout speeds.
 124 *
 125 * We do this by keeping a floating proportion between BDIs, based on page
 126 * writeback completions [end_page_writeback()]. Those devices that write out
 127 * pages fastest will get the larger share, while the slower will get a smaller
 128 * share.
 129 *
 130 * We use page writeout completions because we are interested in getting rid of
 131 * dirty pages. Having them written out is the primary goal.
 132 *
 133 * We introduce a concept of time, a period over which we measure these events,
 134 * because demand can/will vary over time. The length of this period itself is
 135 * measured in page writeback completions.
 136 *
 137 */
 138static struct prop_descriptor vm_completions;
 139static struct prop_descriptor vm_dirties;
 140
 141/*
 142 * couple the period to the dirty_ratio:
 143 *
 144 *   period/2 ~ roundup_pow_of_two(dirty limit)
 145 */
 146static int calc_period_shift(void)
 147{
 148        unsigned long dirty_total;
 149
 150        if (vm_dirty_bytes)
 151                dirty_total = vm_dirty_bytes / PAGE_SIZE;
 152        else
 153                dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) /
 154                                100;
 155        return 2 + ilog2(dirty_total - 1);
 156}
 157
 158/*
 159 * update the period when the dirty threshold changes.
 160 */
 161static void update_completion_period(void)
 162{
 163        int shift = calc_period_shift();
 164        prop_change_shift(&vm_completions, shift);
 165        prop_change_shift(&vm_dirties, shift);
 166}
 167
 168int dirty_background_ratio_handler(struct ctl_table *table, int write,
 169                struct file *filp, void __user *buffer, size_t *lenp,
 170                loff_t *ppos)
 171{
 172        int ret;
 173
 174        ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
 175        if (ret == 0 && write)
 176                dirty_background_bytes = 0;
 177        return ret;
 178}
 179
 180int dirty_background_bytes_handler(struct ctl_table *table, int write,
 181                struct file *filp, void __user *buffer, size_t *lenp,
 182                loff_t *ppos)
 183{
 184        int ret;
 185
 186        ret = proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos);
 187        if (ret == 0 && write)
 188                dirty_background_ratio = 0;
 189        return ret;
 190}
 191
 192int dirty_ratio_handler(struct ctl_table *table, int write,
 193                struct file *filp, void __user *buffer, size_t *lenp,
 194                loff_t *ppos)
 195{
 196        int old_ratio = vm_dirty_ratio;
 197        int ret;
 198
 199        ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
 200        if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
 201                update_completion_period();
 202                vm_dirty_bytes = 0;
 203        }
 204        return ret;
 205}
 206
 207
 208int dirty_bytes_handler(struct ctl_table *table, int write,
 209                struct file *filp, void __user *buffer, size_t *lenp,
 210                loff_t *ppos)
 211{
 212        unsigned long old_bytes = vm_dirty_bytes;
 213        int ret;
 214
 215        ret = proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos);
 216        if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
 217                update_completion_period();
 218                vm_dirty_ratio = 0;
 219        }
 220        return ret;
 221}
 222
 223/*
 224 * Increment the BDI's writeout completion count and the global writeout
 225 * completion count. Called from test_clear_page_writeback().
 226 */
 227static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
 228{
 229        __prop_inc_percpu_max(&vm_completions, &bdi->completions,
 230                              bdi->max_prop_frac);
 231}
 232
 233void bdi_writeout_inc(struct backing_dev_info *bdi)
 234{
 235        unsigned long flags;
 236
 237        local_irq_save(flags);
 238        __bdi_writeout_inc(bdi);
 239        local_irq_restore(flags);
 240}
 241EXPORT_SYMBOL_GPL(bdi_writeout_inc);
 242
 243void task_dirty_inc(struct task_struct *tsk)
 244{
 245        prop_inc_single(&vm_dirties, &tsk->dirties);
 246}
 247
 248/*
 249 * Obtain an accurate fraction of the BDI's portion.
 250 */
 251static void bdi_writeout_fraction(struct backing_dev_info *bdi,
 252                long *numerator, long *denominator)
 253{
 254        if (bdi_cap_writeback_dirty(bdi)) {
 255                prop_fraction_percpu(&vm_completions, &bdi->completions,
 256                                numerator, denominator);
 257        } else {
 258                *numerator = 0;
 259                *denominator = 1;
 260        }
 261}
 262
 263/*
 264 * Clip the earned share of dirty pages to that which is actually available.
 265 * This avoids exceeding the total dirty_limit when the floating averages
 266 * fluctuate too quickly.
 267 */
 268static void
 269clip_bdi_dirty_limit(struct backing_dev_info *bdi, long dirty, long *pbdi_dirty)
 270{
 271        long avail_dirty;
 272
 273        avail_dirty = dirty -
 274                (global_page_state(NR_FILE_DIRTY) +
 275                 global_page_state(NR_WRITEBACK) +
 276                 global_page_state(NR_UNSTABLE_NFS) +
 277                 global_page_state(NR_WRITEBACK_TEMP));
 278
 279        if (avail_dirty < 0)
 280                avail_dirty = 0;
 281
 282        avail_dirty += bdi_stat(bdi, BDI_RECLAIMABLE) +
 283                bdi_stat(bdi, BDI_WRITEBACK);
 284
 285        *pbdi_dirty = min(*pbdi_dirty, avail_dirty);
 286}
 287
 288static inline void task_dirties_fraction(struct task_struct *tsk,
 289                long *numerator, long *denominator)
 290{
 291        prop_fraction_single(&vm_dirties, &tsk->dirties,
 292                                numerator, denominator);
 293}
 294
 295/*
 296 * scale the dirty limit
 297 *
 298 * task specific dirty limit:
 299 *
 300 *   dirty -= (dirty/8) * p_{t}
 301 */
 302static void task_dirty_limit(struct task_struct *tsk, long *pdirty)
 303{
 304        long numerator, denominator;
 305        long dirty = *pdirty;
 306        u64 inv = dirty >> 3;
 307
 308        task_dirties_fraction(tsk, &numerator, &denominator);
 309        inv *= numerator;
 310        do_div(inv, denominator);
 311
 312        dirty -= inv;
 313        if (dirty < *pdirty/2)
 314                dirty = *pdirty/2;
 315
 316        *pdirty = dirty;
 317}
 318
 319/*
 320 *
 321 */
 322static DEFINE_SPINLOCK(bdi_lock);
 323static unsigned int bdi_min_ratio;
 324
 325int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
 326{
 327        int ret = 0;
 328        unsigned long flags;
 329
 330        spin_lock_irqsave(&bdi_lock, flags);
 331        if (min_ratio > bdi->max_ratio) {
 332                ret = -EINVAL;
 333        } else {
 334                min_ratio -= bdi->min_ratio;
 335                if (bdi_min_ratio + min_ratio < 100) {
 336                        bdi_min_ratio += min_ratio;
 337                        bdi->min_ratio += min_ratio;
 338                } else {
 339                        ret = -EINVAL;
 340                }
 341        }
 342        spin_unlock_irqrestore(&bdi_lock, flags);
 343
 344        return ret;
 345}
 346
 347int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
 348{
 349        unsigned long flags;
 350        int ret = 0;
 351
 352        if (max_ratio > 100)
 353                return -EINVAL;
 354
 355        spin_lock_irqsave(&bdi_lock, flags);
 356        if (bdi->min_ratio > max_ratio) {
 357                ret = -EINVAL;
 358        } else {
 359                bdi->max_ratio = max_ratio;
 360                bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
 361        }
 362        spin_unlock_irqrestore(&bdi_lock, flags);
 363
 364        return ret;
 365}
 366EXPORT_SYMBOL(bdi_set_max_ratio);
 367
 368/*
 369 * Work out the current dirty-memory clamping and background writeout
 370 * thresholds.
 371 *
 372 * The main aim here is to lower them aggressively if there is a lot of mapped
 373 * memory around.  To avoid stressing page reclaim with lots of unreclaimable
 374 * pages.  It is better to clamp down on writers than to start swapping, and
 375 * performing lots of scanning.
 376 *
 377 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
 378 *
 379 * We don't permit the clamping level to fall below 5% - that is getting rather
 380 * excessive.
 381 *
 382 * We make sure that the background writeout level is below the adjusted
 383 * clamping level.
 384 */
 385
 386static unsigned long highmem_dirtyable_memory(unsigned long total)
 387{
 388#ifdef CONFIG_HIGHMEM
 389        int node;
 390        unsigned long x = 0;
 391
 392        for_each_node_state(node, N_HIGH_MEMORY) {
 393                struct zone *z =
 394                        &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
 395
 396                x += zone_page_state(z, NR_FREE_PAGES) + zone_lru_pages(z);
 397        }
 398        /*
 399         * Make sure that the number of highmem pages is never larger
 400         * than the number of the total dirtyable memory. This can only
 401         * occur in very strange VM situations but we want to make sure
 402         * that this does not occur.
 403         */
 404        return min(x, total);
 405#else
 406        return 0;
 407#endif
 408}
 409
 410/**
 411 * determine_dirtyable_memory - amount of memory that may be used
 412 *
 413 * Returns the numebr of pages that can currently be freed and used
 414 * by the kernel for direct mappings.
 415 */
 416unsigned long determine_dirtyable_memory(void)
 417{
 418        unsigned long x;
 419
 420        x = global_page_state(NR_FREE_PAGES) + global_lru_pages();
 421
 422        if (!vm_highmem_is_dirtyable)
 423                x -= highmem_dirtyable_memory(x);
 424
 425        return x + 1;   /* Ensure that we never return 0 */
 426}
 427
 428void
 429get_dirty_limits(unsigned long *pbackground, unsigned long *pdirty,
 430                 unsigned long *pbdi_dirty, struct backing_dev_info *bdi)
 431{
 432        unsigned long background;
 433        unsigned long dirty;
 434        unsigned long available_memory = determine_dirtyable_memory();
 435        struct task_struct *tsk;
 436
 437        if (vm_dirty_bytes)
 438                dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
 439        else {
 440                int dirty_ratio;
 441
 442                dirty_ratio = vm_dirty_ratio;
 443                if (dirty_ratio < 5)
 444                        dirty_ratio = 5;
 445                dirty = (dirty_ratio * available_memory) / 100;
 446        }
 447
 448        if (dirty_background_bytes)
 449                background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
 450        else
 451                background = (dirty_background_ratio * available_memory) / 100;
 452
 453        if (background >= dirty)
 454                background = dirty / 2;
 455        tsk = current;
 456        if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
 457                background += background / 4;
 458                dirty += dirty / 4;
 459        }
 460        *pbackground = background;
 461        *pdirty = dirty;
 462
 463        if (bdi) {
 464                u64 bdi_dirty;
 465                long numerator, denominator;
 466
 467                /*
 468                 * Calculate this BDI's share of the dirty ratio.
 469                 */
 470                bdi_writeout_fraction(bdi, &numerator, &denominator);
 471
 472                bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
 473                bdi_dirty *= numerator;
 474                do_div(bdi_dirty, denominator);
 475                bdi_dirty += (dirty * bdi->min_ratio) / 100;
 476                if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
 477                        bdi_dirty = dirty * bdi->max_ratio / 100;
 478
 479                *pbdi_dirty = bdi_dirty;
 480                clip_bdi_dirty_limit(bdi, dirty, pbdi_dirty);
 481                task_dirty_limit(current, pbdi_dirty);
 482        }
 483}
 484
 485/*
 486 * balance_dirty_pages() must be called by processes which are generating dirty
 487 * data.  It looks at the number of dirty pages in the machine and will force
 488 * the caller to perform writeback if the system is over `vm_dirty_ratio'.
 489 * If we're over `background_thresh' then pdflush is woken to perform some
 490 * writeout.
 491 */
 492static void balance_dirty_pages(struct address_space *mapping)
 493{
 494        long nr_reclaimable, bdi_nr_reclaimable;
 495        long nr_writeback, bdi_nr_writeback;
 496        unsigned long background_thresh;
 497        unsigned long dirty_thresh;
 498        unsigned long bdi_thresh;
 499        unsigned long pages_written = 0;
 500        unsigned long write_chunk = sync_writeback_pages();
 501
 502        struct backing_dev_info *bdi = mapping->backing_dev_info;
 503
 504        for (;;) {
 505                struct writeback_control wbc = {
 506                        .bdi            = bdi,
 507                        .sync_mode      = WB_SYNC_NONE,
 508                        .older_than_this = NULL,
 509                        .nr_to_write    = write_chunk,
 510                        .range_cyclic   = 1,
 511                };
 512
 513                get_dirty_limits(&background_thresh, &dirty_thresh,
 514                                &bdi_thresh, bdi);
 515
 516                nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
 517                                        global_page_state(NR_UNSTABLE_NFS);
 518                nr_writeback = global_page_state(NR_WRITEBACK);
 519
 520                bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
 521                bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
 522
 523                if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
 524                        break;
 525
 526                /*
 527                 * Throttle it only when the background writeback cannot
 528                 * catch-up. This avoids (excessively) small writeouts
 529                 * when the bdi limits are ramping up.
 530                 */
 531                if (nr_reclaimable + nr_writeback <
 532                                (background_thresh + dirty_thresh) / 2)
 533                        break;
 534
 535                if (!bdi->dirty_exceeded)
 536                        bdi->dirty_exceeded = 1;
 537
 538                /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
 539                 * Unstable writes are a feature of certain networked
 540                 * filesystems (i.e. NFS) in which data may have been
 541                 * written to the server's write cache, but has not yet
 542                 * been flushed to permanent storage.
 543                 */
 544                if (bdi_nr_reclaimable) {
 545                        writeback_inodes(&wbc);
 546                        pages_written += write_chunk - wbc.nr_to_write;
 547                        get_dirty_limits(&background_thresh, &dirty_thresh,
 548                                       &bdi_thresh, bdi);
 549                }
 550
 551                /*
 552                 * In order to avoid the stacked BDI deadlock we need
 553                 * to ensure we accurately count the 'dirty' pages when
 554                 * the threshold is low.
 555                 *
 556                 * Otherwise it would be possible to get thresh+n pages
 557                 * reported dirty, even though there are thresh-m pages
 558                 * actually dirty; with m+n sitting in the percpu
 559                 * deltas.
 560                 */
 561                if (bdi_thresh < 2*bdi_stat_error(bdi)) {
 562                        bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
 563                        bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
 564                } else if (bdi_nr_reclaimable) {
 565                        bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
 566                        bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
 567                }
 568
 569                if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
 570                        break;
 571                if (pages_written >= write_chunk)
 572                        break;          /* We've done our duty */
 573
 574                congestion_wait(WRITE, HZ/10);
 575        }
 576
 577        if (bdi_nr_reclaimable + bdi_nr_writeback < bdi_thresh &&
 578                        bdi->dirty_exceeded)
 579                bdi->dirty_exceeded = 0;
 580
 581        if (writeback_in_progress(bdi))
 582                return;         /* pdflush is already working this queue */
 583
 584        /*
 585         * In laptop mode, we wait until hitting the higher threshold before
 586         * starting background writeout, and then write out all the way down
 587         * to the lower threshold.  So slow writers cause minimal disk activity.
 588         *
 589         * In normal mode, we start background writeout at the lower
 590         * background_thresh, to keep the amount of dirty memory low.
 591         */
 592        if ((laptop_mode && pages_written) ||
 593                        (!laptop_mode && (global_page_state(NR_FILE_DIRTY)
 594                                          + global_page_state(NR_UNSTABLE_NFS)
 595                                          > background_thresh)))
 596                pdflush_operation(background_writeout, 0);
 597}
 598
 599void set_page_dirty_balance(struct page *page, int page_mkwrite)
 600{
 601        if (set_page_dirty(page) || page_mkwrite) {
 602                struct address_space *mapping = page_mapping(page);
 603
 604                if (mapping)
 605                        balance_dirty_pages_ratelimited(mapping);
 606        }
 607}
 608
 609/**
 610 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
 611 * @mapping: address_space which was dirtied
 612 * @nr_pages_dirtied: number of pages which the caller has just dirtied
 613 *
 614 * Processes which are dirtying memory should call in here once for each page
 615 * which was newly dirtied.  The function will periodically check the system's
 616 * dirty state and will initiate writeback if needed.
 617 *
 618 * On really big machines, get_writeback_state is expensive, so try to avoid
 619 * calling it too often (ratelimiting).  But once we're over the dirty memory
 620 * limit we decrease the ratelimiting by a lot, to prevent individual processes
 621 * from overshooting the limit by (ratelimit_pages) each.
 622 */
 623void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
 624                                        unsigned long nr_pages_dirtied)
 625{
 626        static DEFINE_PER_CPU(unsigned long, ratelimits) = 0;
 627        unsigned long ratelimit;
 628        unsigned long *p;
 629
 630        ratelimit = ratelimit_pages;
 631        if (mapping->backing_dev_info->dirty_exceeded)
 632                ratelimit = 8;
 633
 634        /*
 635         * Check the rate limiting. Also, we do not want to throttle real-time
 636         * tasks in balance_dirty_pages(). Period.
 637         */
 638        preempt_disable();
 639        p =  &__get_cpu_var(ratelimits);
 640        *p += nr_pages_dirtied;
 641        if (unlikely(*p >= ratelimit)) {
 642                *p = 0;
 643                preempt_enable();
 644                balance_dirty_pages(mapping);
 645                return;
 646        }
 647        preempt_enable();
 648}
 649EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
 650
 651void throttle_vm_writeout(gfp_t gfp_mask)
 652{
 653        unsigned long background_thresh;
 654        unsigned long dirty_thresh;
 655
 656        for ( ; ; ) {
 657                get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
 658
 659                /*
 660                 * Boost the allowable dirty threshold a bit for page
 661                 * allocators so they don't get DoS'ed by heavy writers
 662                 */
 663                dirty_thresh += dirty_thresh / 10;      /* wheeee... */
 664
 665                if (global_page_state(NR_UNSTABLE_NFS) +
 666                        global_page_state(NR_WRITEBACK) <= dirty_thresh)
 667                                break;
 668                congestion_wait(WRITE, HZ/10);
 669
 670                /*
 671                 * The caller might hold locks which can prevent IO completion
 672                 * or progress in the filesystem.  So we cannot just sit here
 673                 * waiting for IO to complete.
 674                 */
 675                if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
 676                        break;
 677        }
 678}
 679
 680/*
 681 * writeback at least _min_pages, and keep writing until the amount of dirty
 682 * memory is less than the background threshold, or until we're all clean.
 683 */
 684static void background_writeout(unsigned long _min_pages)
 685{
 686        long min_pages = _min_pages;
 687        struct writeback_control wbc = {
 688                .bdi            = NULL,
 689                .sync_mode      = WB_SYNC_NONE,
 690                .older_than_this = NULL,
 691                .nr_to_write    = 0,
 692                .nonblocking    = 1,
 693                .range_cyclic   = 1,
 694        };
 695
 696        for ( ; ; ) {
 697                unsigned long background_thresh;
 698                unsigned long dirty_thresh;
 699
 700                get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
 701                if (global_page_state(NR_FILE_DIRTY) +
 702                        global_page_state(NR_UNSTABLE_NFS) < background_thresh
 703                                && min_pages <= 0)
 704                        break;
 705                wbc.more_io = 0;
 706                wbc.encountered_congestion = 0;
 707                wbc.nr_to_write = MAX_WRITEBACK_PAGES;
 708                wbc.pages_skipped = 0;
 709                writeback_inodes(&wbc);
 710                min_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
 711                if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
 712                        /* Wrote less than expected */
 713                        if (wbc.encountered_congestion || wbc.more_io)
 714                                congestion_wait(WRITE, HZ/10);
 715                        else
 716                                break;
 717                }
 718        }
 719}
 720
 721/*
 722 * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
 723 * the whole world.  Returns 0 if a pdflush thread was dispatched.  Returns
 724 * -1 if all pdflush threads were busy.
 725 */
 726int wakeup_pdflush(long nr_pages)
 727{
 728        if (nr_pages == 0)
 729                nr_pages = global_page_state(NR_FILE_DIRTY) +
 730                                global_page_state(NR_UNSTABLE_NFS);
 731        return pdflush_operation(background_writeout, nr_pages);
 732}
 733
 734static void wb_timer_fn(unsigned long unused);
 735static void laptop_timer_fn(unsigned long unused);
 736
 737static DEFINE_TIMER(wb_timer, wb_timer_fn, 0, 0);
 738static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
 739
 740/*
 741 * Periodic writeback of "old" data.
 742 *
 743 * Define "old": the first time one of an inode's pages is dirtied, we mark the
 744 * dirtying-time in the inode's address_space.  So this periodic writeback code
 745 * just walks the superblock inode list, writing back any inodes which are
 746 * older than a specific point in time.
 747 *
 748 * Try to run once per dirty_writeback_interval.  But if a writeback event
 749 * takes longer than a dirty_writeback_interval interval, then leave a
 750 * one-second gap.
 751 *
 752 * older_than_this takes precedence over nr_to_write.  So we'll only write back
 753 * all dirty pages if they are all attached to "old" mappings.
 754 */
 755static void wb_kupdate(unsigned long arg)
 756{
 757        unsigned long oldest_jif;
 758        unsigned long start_jif;
 759        unsigned long next_jif;
 760        long nr_to_write;
 761        struct writeback_control wbc = {
 762                .bdi            = NULL,
 763                .sync_mode      = WB_SYNC_NONE,
 764                .older_than_this = &oldest_jif,
 765                .nr_to_write    = 0,
 766                .nonblocking    = 1,
 767                .for_kupdate    = 1,
 768                .range_cyclic   = 1,
 769        };
 770
 771        sync_supers();
 772
 773        oldest_jif = jiffies - msecs_to_jiffies(dirty_expire_interval * 10);
 774        start_jif = jiffies;
 775        next_jif = start_jif + msecs_to_jiffies(dirty_writeback_interval * 10);
 776        nr_to_write = global_page_state(NR_FILE_DIRTY) +
 777                        global_page_state(NR_UNSTABLE_NFS) +
 778                        (inodes_stat.nr_inodes - inodes_stat.nr_unused);
 779        while (nr_to_write > 0) {
 780                wbc.more_io = 0;
 781                wbc.encountered_congestion = 0;
 782                wbc.nr_to_write = MAX_WRITEBACK_PAGES;
 783                writeback_inodes(&wbc);
 784                if (wbc.nr_to_write > 0) {
 785                        if (wbc.encountered_congestion || wbc.more_io)
 786                                congestion_wait(WRITE, HZ/10);
 787                        else
 788                                break;  /* All the old data is written */
 789                }
 790                nr_to_write -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
 791        }
 792        if (time_before(next_jif, jiffies + HZ))
 793                next_jif = jiffies + HZ;
 794        if (dirty_writeback_interval)
 795                mod_timer(&wb_timer, next_jif);
 796}
 797
 798/*
 799 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
 800 */
 801int dirty_writeback_centisecs_handler(ctl_table *table, int write,
 802        struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
 803{
 804        proc_dointvec(table, write, file, buffer, length, ppos);
 805        if (dirty_writeback_interval)
 806                mod_timer(&wb_timer, jiffies +
 807                        msecs_to_jiffies(dirty_writeback_interval * 10));
 808        else
 809                del_timer(&wb_timer);
 810        return 0;
 811}
 812
 813static void wb_timer_fn(unsigned long unused)
 814{
 815        if (pdflush_operation(wb_kupdate, 0) < 0)
 816                mod_timer(&wb_timer, jiffies + HZ); /* delay 1 second */
 817}
 818
 819static void laptop_flush(unsigned long unused)
 820{
 821        sys_sync();
 822}
 823
 824static void laptop_timer_fn(unsigned long unused)
 825{
 826        pdflush_operation(laptop_flush, 0);
 827}
 828
 829/*
 830 * We've spun up the disk and we're in laptop mode: schedule writeback
 831 * of all dirty data a few seconds from now.  If the flush is already scheduled
 832 * then push it back - the user is still using the disk.
 833 */
 834void laptop_io_completion(void)
 835{
 836        mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
 837}
 838
 839/*
 840 * We're in laptop mode and we've just synced. The sync's writes will have
 841 * caused another writeback to be scheduled by laptop_io_completion.
 842 * Nothing needs to be written back anymore, so we unschedule the writeback.
 843 */
 844void laptop_sync_completion(void)
 845{
 846        del_timer(&laptop_mode_wb_timer);
 847}
 848
 849/*
 850 * If ratelimit_pages is too high then we can get into dirty-data overload
 851 * if a large number of processes all perform writes at the same time.
 852 * If it is too low then SMP machines will call the (expensive)
 853 * get_writeback_state too often.
 854 *
 855 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
 856 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
 857 * thresholds before writeback cuts in.
 858 *
 859 * But the limit should not be set too high.  Because it also controls the
 860 * amount of memory which the balance_dirty_pages() caller has to write back.
 861 * If this is too large then the caller will block on the IO queue all the
 862 * time.  So limit it to four megabytes - the balance_dirty_pages() caller
 863 * will write six megabyte chunks, max.
 864 */
 865
 866void writeback_set_ratelimit(void)
 867{
 868        ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
 869        if (ratelimit_pages < 16)
 870                ratelimit_pages = 16;
 871        if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
 872                ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
 873}
 874
 875static int __cpuinit
 876ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
 877{
 878        writeback_set_ratelimit();
 879        return NOTIFY_DONE;
 880}
 881
 882static struct notifier_block __cpuinitdata ratelimit_nb = {
 883        .notifier_call  = ratelimit_handler,
 884        .next           = NULL,
 885};
 886
 887/*
 888 * Called early on to tune the page writeback dirty limits.
 889 *
 890 * We used to scale dirty pages according to how total memory
 891 * related to pages that could be allocated for buffers (by
 892 * comparing nr_free_buffer_pages() to vm_total_pages.
 893 *
 894 * However, that was when we used "dirty_ratio" to scale with
 895 * all memory, and we don't do that any more. "dirty_ratio"
 896 * is now applied to total non-HIGHPAGE memory (by subtracting
 897 * totalhigh_pages from vm_total_pages), and as such we can't
 898 * get into the old insane situation any more where we had
 899 * large amounts of dirty pages compared to a small amount of
 900 * non-HIGHMEM memory.
 901 *
 902 * But we might still want to scale the dirty_ratio by how
 903 * much memory the box has..
 904 */
 905void __init page_writeback_init(void)
 906{
 907        int shift;
 908
 909        mod_timer(&wb_timer,
 910                  jiffies + msecs_to_jiffies(dirty_writeback_interval * 10));
 911        writeback_set_ratelimit();
 912        register_cpu_notifier(&ratelimit_nb);
 913
 914        shift = calc_period_shift();
 915        prop_descriptor_init(&vm_completions, shift);
 916        prop_descriptor_init(&vm_dirties, shift);
 917}
 918
 919/**
 920 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
 921 * @mapping: address space structure to write
 922 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
 923 * @writepage: function called for each page
 924 * @data: data passed to writepage function
 925 *
 926 * If a page is already under I/O, write_cache_pages() skips it, even
 927 * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
 928 * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
 929 * and msync() need to guarantee that all the data which was dirty at the time
 930 * the call was made get new I/O started against them.  If wbc->sync_mode is
 931 * WB_SYNC_ALL then we were called for data integrity and we must wait for
 932 * existing IO to complete.
 933 */
 934int write_cache_pages(struct address_space *mapping,
 935                      struct writeback_control *wbc, writepage_t writepage,
 936                      void *data)
 937{
 938        struct backing_dev_info *bdi = mapping->backing_dev_info;
 939        int ret = 0;
 940        int done = 0;
 941        struct pagevec pvec;
 942        int nr_pages;
 943        pgoff_t uninitialized_var(writeback_index);
 944        pgoff_t index;
 945        pgoff_t end;            /* Inclusive */
 946        pgoff_t done_index;
 947        int cycled;
 948        int range_whole = 0;
 949        long nr_to_write = wbc->nr_to_write;
 950
 951        if (wbc->nonblocking && bdi_write_congested(bdi)) {
 952                wbc->encountered_congestion = 1;
 953                return 0;
 954        }
 955
 956        pagevec_init(&pvec, 0);
 957        if (wbc->range_cyclic) {
 958                writeback_index = mapping->writeback_index; /* prev offset */
 959                index = writeback_index;
 960                if (index == 0)
 961                        cycled = 1;
 962                else
 963                        cycled = 0;
 964                end = -1;
 965        } else {
 966                index = wbc->range_start >> PAGE_CACHE_SHIFT;
 967                end = wbc->range_end >> PAGE_CACHE_SHIFT;
 968                if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 969                        range_whole = 1;
 970                cycled = 1; /* ignore range_cyclic tests */
 971        }
 972retry:
 973        done_index = index;
 974        while (!done && (index <= end)) {
 975                int i;
 976
 977                nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 978                              PAGECACHE_TAG_DIRTY,
 979                              min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
 980                if (nr_pages == 0)
 981                        break;
 982
 983                for (i = 0; i < nr_pages; i++) {
 984                        struct page *page = pvec.pages[i];
 985
 986                        /*
 987                         * At this point, the page may be truncated or
 988                         * invalidated (changing page->mapping to NULL), or
 989                         * even swizzled back from swapper_space to tmpfs file
 990                         * mapping. However, page->index will not change
 991                         * because we have a reference on the page.
 992                         */
 993                        if (page->index > end) {
 994                                /*
 995                                 * can't be range_cyclic (1st pass) because
 996                                 * end == -1 in that case.
 997                                 */
 998                                done = 1;
 999                                break;
1000                        }
1001
1002                        done_index = page->index + 1;
1003
1004                        lock_page(page);
1005
1006                        /*
1007                         * Page truncated or invalidated. We can freely skip it
1008                         * then, even for data integrity operations: the page
1009                         * has disappeared concurrently, so there could be no
1010                         * real expectation of this data interity operation
1011                         * even if there is now a new, dirty page at the same
1012                         * pagecache address.
1013                         */
1014                        if (unlikely(page->mapping != mapping)) {
1015continue_unlock:
1016                                unlock_page(page);
1017                                continue;
1018                        }
1019
1020                        if (!PageDirty(page)) {
1021                                /* someone wrote it for us */
1022                                goto continue_unlock;
1023                        }
1024
1025                        if (PageWriteback(page)) {
1026                                if (wbc->sync_mode != WB_SYNC_NONE)
1027                                        wait_on_page_writeback(page);
1028                                else
1029                                        goto continue_unlock;
1030                        }
1031
1032                        BUG_ON(PageWriteback(page));
1033                        if (!clear_page_dirty_for_io(page))
1034                                goto continue_unlock;
1035
1036                        ret = (*writepage)(page, wbc, data);
1037                        if (unlikely(ret)) {
1038                                if (ret == AOP_WRITEPAGE_ACTIVATE) {
1039                                        unlock_page(page);
1040                                        ret = 0;
1041                                } else {
1042                                        /*
1043                                         * done_index is set past this page,
1044                                         * so media errors will not choke
1045                                         * background writeout for the entire
1046                                         * file. This has consequences for
1047                                         * range_cyclic semantics (ie. it may
1048                                         * not be suitable for data integrity
1049                                         * writeout).
1050                                         */
1051                                        done = 1;
1052                                        break;
1053                                }
1054                        }
1055
1056                        if (nr_to_write > 0) {
1057                                nr_to_write--;
1058                                if (nr_to_write == 0 &&
1059                                    wbc->sync_mode == WB_SYNC_NONE) {
1060                                        /*
1061                                         * We stop writing back only if we are
1062                                         * not doing integrity sync. In case of
1063                                         * integrity sync we have to keep going
1064                                         * because someone may be concurrently
1065                                         * dirtying pages, and we might have
1066                                         * synced a lot of newly appeared dirty
1067                                         * pages, but have not synced all of the
1068                                         * old dirty pages.
1069                                         */
1070                                        done = 1;
1071                                        break;
1072                                }
1073                        }
1074
1075                        if (wbc->nonblocking && bdi_write_congested(bdi)) {
1076                                wbc->encountered_congestion = 1;
1077                                done = 1;
1078                                break;
1079                        }
1080                }
1081                pagevec_release(&pvec);
1082                cond_resched();
1083        }
1084        if (!cycled && !done) {
1085                /*
1086                 * range_cyclic:
1087                 * We hit the last page and there is more work to be done: wrap
1088                 * back to the start of the file
1089                 */
1090                cycled = 1;
1091                index = 0;
1092                end = writeback_index - 1;
1093                goto retry;
1094        }
1095        if (!wbc->no_nrwrite_index_update) {
1096                if (wbc->range_cyclic || (range_whole && nr_to_write > 0))
1097                        mapping->writeback_index = done_index;
1098                wbc->nr_to_write = nr_to_write;
1099        }
1100
1101        return ret;
1102}
1103EXPORT_SYMBOL(write_cache_pages);
1104
1105/*
1106 * Function used by generic_writepages to call the real writepage
1107 * function and set the mapping flags on error
1108 */
1109static int __writepage(struct page *page, struct writeback_control *wbc,
1110                       void *data)
1111{
1112        struct address_space *mapping = data;
1113        int ret = mapping->a_ops->writepage(page, wbc);
1114        mapping_set_error(mapping, ret);
1115        return ret;
1116}
1117
1118/**
1119 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1120 * @mapping: address space structure to write
1121 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1122 *
1123 * This is a library function, which implements the writepages()
1124 * address_space_operation.
1125 */
1126int generic_writepages(struct address_space *mapping,
1127                       struct writeback_control *wbc)
1128{
1129        /* deal with chardevs and other special file */
1130        if (!mapping->a_ops->writepage)
1131                return 0;
1132
1133        return write_cache_pages(mapping, wbc, __writepage, mapping);
1134}
1135
1136EXPORT_SYMBOL(generic_writepages);
1137
1138int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1139{
1140        int ret;
1141
1142        if (wbc->nr_to_write <= 0)
1143                return 0;
1144        wbc->for_writepages = 1;
1145        if (mapping->a_ops->writepages)
1146                ret = mapping->a_ops->writepages(mapping, wbc);
1147        else
1148                ret = generic_writepages(mapping, wbc);
1149        wbc->for_writepages = 0;
1150        return ret;
1151}
1152
1153/**
1154 * write_one_page - write out a single page and optionally wait on I/O
1155 * @page: the page to write
1156 * @wait: if true, wait on writeout
1157 *
1158 * The page must be locked by the caller and will be unlocked upon return.
1159 *
1160 * write_one_page() returns a negative error code if I/O failed.
1161 */
1162int write_one_page(struct page *page, int wait)
1163{
1164        struct address_space *mapping = page->mapping;
1165        int ret = 0;
1166        struct writeback_control wbc = {
1167                .sync_mode = WB_SYNC_ALL,
1168                .nr_to_write = 1,
1169        };
1170
1171        BUG_ON(!PageLocked(page));
1172
1173        if (wait)
1174                wait_on_page_writeback(page);
1175
1176        if (clear_page_dirty_for_io(page)) {
1177                page_cache_get(page);
1178                ret = mapping->a_ops->writepage(page, &wbc);
1179                if (ret == 0 && wait) {
1180                        wait_on_page_writeback(page);
1181                        if (PageError(page))
1182                                ret = -EIO;
1183                }
1184                page_cache_release(page);
1185        } else {
1186                unlock_page(page);
1187        }
1188        return ret;
1189}
1190EXPORT_SYMBOL(write_one_page);
1191
1192/*
1193 * For address_spaces which do not use buffers nor write back.
1194 */
1195int __set_page_dirty_no_writeback(struct page *page)
1196{
1197        if (!PageDirty(page))
1198                SetPageDirty(page);
1199        return 0;
1200}
1201
1202/*
1203 * Helper function for set_page_dirty family.
1204 * NOTE: This relies on being atomic wrt interrupts.
1205 */
1206void account_page_dirtied(struct page *page, struct address_space *mapping)
1207{
1208        if (mapping_cap_account_dirty(mapping)) {
1209                __inc_zone_page_state(page, NR_FILE_DIRTY);
1210                __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1211                task_dirty_inc(current);
1212                task_io_account_write(PAGE_CACHE_SIZE);
1213        }
1214}
1215
1216/*
1217 * For address_spaces which do not use buffers.  Just tag the page as dirty in
1218 * its radix tree.
1219 *
1220 * This is also used when a single buffer is being dirtied: we want to set the
1221 * page dirty in that case, but not all the buffers.  This is a "bottom-up"
1222 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1223 *
1224 * Most callers have locked the page, which pins the address_space in memory.
1225 * But zap_pte_range() does not lock the page, however in that case the
1226 * mapping is pinned by the vma's ->vm_file reference.
1227 *
1228 * We take care to handle the case where the page was truncated from the
1229 * mapping by re-checking page_mapping() inside tree_lock.
1230 */
1231int __set_page_dirty_nobuffers(struct page *page)
1232{
1233        if (!TestSetPageDirty(page)) {
1234                struct address_space *mapping = page_mapping(page);
1235                struct address_space *mapping2;
1236
1237                if (!mapping)
1238                        return 1;
1239
1240                spin_lock_irq(&mapping->tree_lock);
1241                mapping2 = page_mapping(page);
1242                if (mapping2) { /* Race with truncate? */
1243                        BUG_ON(mapping2 != mapping);
1244                        WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
1245                        account_page_dirtied(page, mapping);
1246                        radix_tree_tag_set(&mapping->page_tree,
1247                                page_index(page), PAGECACHE_TAG_DIRTY);
1248                }
1249                spin_unlock_irq(&mapping->tree_lock);
1250                if (mapping->host) {
1251                        /* !PageAnon && !swapper_space */
1252                        __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1253                }
1254                return 1;
1255        }
1256        return 0;
1257}
1258EXPORT_SYMBOL(__set_page_dirty_nobuffers);
1259
1260/*
1261 * When a writepage implementation decides that it doesn't want to write this
1262 * page for some reason, it should redirty the locked page via
1263 * redirty_page_for_writepage() and it should then unlock the page and return 0
1264 */
1265int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
1266{
1267        wbc->pages_skipped++;
1268        return __set_page_dirty_nobuffers(page);
1269}
1270EXPORT_SYMBOL(redirty_page_for_writepage);
1271
1272/*
1273 * If the mapping doesn't provide a set_page_dirty a_op, then
1274 * just fall through and assume that it wants buffer_heads.
1275 */
1276int set_page_dirty(struct page *page)
1277{
1278        struct address_space *mapping = page_mapping(page);
1279
1280        if (likely(mapping)) {
1281                int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
1282#ifdef CONFIG_BLOCK
1283                if (!spd)
1284                        spd = __set_page_dirty_buffers;
1285#endif
1286                return (*spd)(page);
1287        }
1288        if (!PageDirty(page)) {
1289                if (!TestSetPageDirty(page))
1290                        return 1;
1291        }
1292        return 0;
1293}
1294EXPORT_SYMBOL(set_page_dirty);
1295
1296/*
1297 * set_page_dirty() is racy if the caller has no reference against
1298 * page->mapping->host, and if the page is unlocked.  This is because another
1299 * CPU could truncate the page off the mapping and then free the mapping.
1300 *
1301 * Usually, the page _is_ locked, or the caller is a user-space process which
1302 * holds a reference on the inode by having an open file.
1303 *
1304 * In other cases, the page should be locked before running set_page_dirty().
1305 */
1306int set_page_dirty_lock(struct page *page)
1307{
1308        int ret;
1309
1310        lock_page_nosync(page);
1311        ret = set_page_dirty(page);
1312        unlock_page(page);
1313        return ret;
1314}
1315EXPORT_SYMBOL(set_page_dirty_lock);
1316
1317/*
1318 * Clear a page's dirty flag, while caring for dirty memory accounting.
1319 * Returns true if the page was previously dirty.
1320 *
1321 * This is for preparing to put the page under writeout.  We leave the page
1322 * tagged as dirty in the radix tree so that a concurrent write-for-sync
1323 * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
1324 * implementation will run either set_page_writeback() or set_page_dirty(),
1325 * at which stage we bring the page's dirty flag and radix-tree dirty tag
1326 * back into sync.
1327 *
1328 * This incoherency between the page's dirty flag and radix-tree tag is
1329 * unfortunate, but it only exists while the page is locked.
1330 */
1331int clear_page_dirty_for_io(struct page *page)
1332{
1333        struct address_space *mapping = page_mapping(page);
1334
1335        BUG_ON(!PageLocked(page));
1336
1337        ClearPageReclaim(page);
1338        if (mapping && mapping_cap_account_dirty(mapping)) {
1339                /*
1340                 * Yes, Virginia, this is indeed insane.
1341                 *
1342                 * We use this sequence to make sure that
1343                 *  (a) we account for dirty stats properly
1344                 *  (b) we tell the low-level filesystem to
1345                 *      mark the whole page dirty if it was
1346                 *      dirty in a pagetable. Only to then
1347                 *  (c) clean the page again and return 1 to
1348                 *      cause the writeback.
1349                 *
1350                 * This way we avoid all nasty races with the
1351                 * dirty bit in multiple places and clearing
1352                 * them concurrently from different threads.
1353                 *
1354                 * Note! Normally the "set_page_dirty(page)"
1355                 * has no effect on the actual dirty bit - since
1356                 * that will already usually be set. But we
1357                 * need the side effects, and it can help us
1358                 * avoid races.
1359                 *
1360                 * We basically use the page "master dirty bit"
1361                 * as a serialization point for all the different
1362                 * threads doing their things.
1363                 */
1364                if (page_mkclean(page))
1365                        set_page_dirty(page);
1366                /*
1367                 * We carefully synchronise fault handlers against
1368                 * installing a dirty pte and marking the page dirty
1369                 * at this point. We do this by having them hold the
1370                 * page lock at some point after installing their
1371                 * pte, but before marking the page dirty.
1372                 * Pages are always locked coming in here, so we get
1373                 * the desired exclusion. See mm/memory.c:do_wp_page()
1374                 * for more comments.
1375                 */
1376                if (TestClearPageDirty(page)) {
1377                        dec_zone_page_state(page, NR_FILE_DIRTY);
1378                        dec_bdi_stat(mapping->backing_dev_info,
1379                                        BDI_RECLAIMABLE);
1380                        return 1;
1381                }
1382                return 0;
1383        }
1384        return TestClearPageDirty(page);
1385}
1386EXPORT_SYMBOL(clear_page_dirty_for_io);
1387
1388int test_clear_page_writeback(struct page *page)
1389{
1390        struct address_space *mapping = page_mapping(page);
1391        int ret;
1392
1393        if (mapping) {
1394                struct backing_dev_info *bdi = mapping->backing_dev_info;
1395                unsigned long flags;
1396
1397                spin_lock_irqsave(&mapping->tree_lock, flags);
1398                ret = TestClearPageWriteback(page);
1399                if (ret) {
1400                        radix_tree_tag_clear(&mapping->page_tree,
1401                                                page_index(page),
1402                                                PAGECACHE_TAG_WRITEBACK);
1403                        if (bdi_cap_account_writeback(bdi)) {
1404                                __dec_bdi_stat(bdi, BDI_WRITEBACK);
1405                                __bdi_writeout_inc(bdi);
1406                        }
1407                }
1408                spin_unlock_irqrestore(&mapping->tree_lock, flags);
1409        } else {
1410                ret = TestClearPageWriteback(page);
1411        }
1412        if (ret)
1413                dec_zone_page_state(page, NR_WRITEBACK);
1414        return ret;
1415}
1416
1417int test_set_page_writeback(struct page *page)
1418{
1419        struct address_space *mapping = page_mapping(page);
1420        int ret;
1421
1422        if (mapping) {
1423                struct backing_dev_info *bdi = mapping->backing_dev_info;
1424                unsigned long flags;
1425
1426                spin_lock_irqsave(&mapping->tree_lock, flags);
1427                ret = TestSetPageWriteback(page);
1428                if (!ret) {
1429                        radix_tree_tag_set(&mapping->page_tree,
1430                                                page_index(page),
1431                                                PAGECACHE_TAG_WRITEBACK);
1432                        if (bdi_cap_account_writeback(bdi))
1433                                __inc_bdi_stat(bdi, BDI_WRITEBACK);
1434                }
1435                if (!PageDirty(page))
1436                        radix_tree_tag_clear(&mapping->page_tree,
1437                                                page_index(page),
1438                                                PAGECACHE_TAG_DIRTY);
1439                spin_unlock_irqrestore(&mapping->tree_lock, flags);
1440        } else {
1441                ret = TestSetPageWriteback(page);
1442        }
1443        if (!ret)
1444                inc_zone_page_state(page, NR_WRITEBACK);
1445        return ret;
1446
1447}
1448EXPORT_SYMBOL(test_set_page_writeback);
1449
1450/*
1451 * Return true if any of the pages in the mapping are marked with the
1452 * passed tag.
1453 */
1454int mapping_tagged(struct address_space *mapping, int tag)
1455{
1456        int ret;
1457        rcu_read_lock();
1458        ret = radix_tree_tagged(&mapping->page_tree, tag);
1459        rcu_read_unlock();
1460        return ret;
1461}
1462EXPORT_SYMBOL(mapping_tagged);
1463