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/export.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> /* __set_page_dirty_buffers */
  36#include <linux/pagevec.h>
  37#include <linux/timer.h>
  38#include <trace/events/writeback.h>
  39
  40/*
  41 * Sleep at most 200ms at a time in balance_dirty_pages().
  42 */
  43#define MAX_PAUSE               max(HZ/5, 1)
  44
  45/*
  46 * Try to keep balance_dirty_pages() call intervals higher than this many pages
  47 * by raising pause time to max_pause when falls below it.
  48 */
  49#define DIRTY_POLL_THRESH       (128 >> (PAGE_SHIFT - 10))
  50
  51/*
  52 * Estimate write bandwidth at 200ms intervals.
  53 */
  54#define BANDWIDTH_INTERVAL      max(HZ/5, 1)
  55
  56#define RATELIMIT_CALC_SHIFT    10
  57
  58/*
  59 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
  60 * will look to see if it needs to force writeback or throttling.
  61 */
  62static long ratelimit_pages = 32;
  63
  64/* The following parameters are exported via /proc/sys/vm */
  65
  66/*
  67 * Start background writeback (via writeback threads) 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
  99EXPORT_SYMBOL_GPL(dirty_writeback_interval);
 100
 101/*
 102 * The longest time for which data is allowed to remain dirty
 103 */
 104unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
 105
 106/*
 107 * Flag that makes the machine dump writes/reads and block dirtyings.
 108 */
 109int block_dump;
 110
 111/*
 112 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
 113 * a full sync is triggered after this time elapses without any disk activity.
 114 */
 115int laptop_mode;
 116
 117EXPORT_SYMBOL(laptop_mode);
 118
 119/* End of sysctl-exported parameters */
 120
 121unsigned long global_dirty_limit;
 122
 123/*
 124 * Scale the writeback cache size proportional to the relative writeout speeds.
 125 *
 126 * We do this by keeping a floating proportion between BDIs, based on page
 127 * writeback completions [end_page_writeback()]. Those devices that write out
 128 * pages fastest will get the larger share, while the slower will get a smaller
 129 * share.
 130 *
 131 * We use page writeout completions because we are interested in getting rid of
 132 * dirty pages. Having them written out is the primary goal.
 133 *
 134 * We introduce a concept of time, a period over which we measure these events,
 135 * because demand can/will vary over time. The length of this period itself is
 136 * measured in page writeback completions.
 137 *
 138 */
 139static struct fprop_global writeout_completions;
 140
 141static void writeout_period(unsigned long t);
 142/* Timer for aging of writeout_completions */
 143static struct timer_list writeout_period_timer =
 144                TIMER_DEFERRED_INITIALIZER(writeout_period, 0, 0);
 145static unsigned long writeout_period_time = 0;
 146
 147/*
 148 * Length of period for aging writeout fractions of bdis. This is an
 149 * arbitrarily chosen number. The longer the period, the slower fractions will
 150 * reflect changes in current writeout rate.
 151 */
 152#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
 153
 154/*
 155 * Work out the current dirty-memory clamping and background writeout
 156 * thresholds.
 157 *
 158 * The main aim here is to lower them aggressively if there is a lot of mapped
 159 * memory around.  To avoid stressing page reclaim with lots of unreclaimable
 160 * pages.  It is better to clamp down on writers than to start swapping, and
 161 * performing lots of scanning.
 162 *
 163 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
 164 *
 165 * We don't permit the clamping level to fall below 5% - that is getting rather
 166 * excessive.
 167 *
 168 * We make sure that the background writeout level is below the adjusted
 169 * clamping level.
 170 */
 171
 172/*
 173 * In a memory zone, there is a certain amount of pages we consider
 174 * available for the page cache, which is essentially the number of
 175 * free and reclaimable pages, minus some zone reserves to protect
 176 * lowmem and the ability to uphold the zone's watermarks without
 177 * requiring writeback.
 178 *
 179 * This number of dirtyable pages is the base value of which the
 180 * user-configurable dirty ratio is the effictive number of pages that
 181 * are allowed to be actually dirtied.  Per individual zone, or
 182 * globally by using the sum of dirtyable pages over all zones.
 183 *
 184 * Because the user is allowed to specify the dirty limit globally as
 185 * absolute number of bytes, calculating the per-zone dirty limit can
 186 * require translating the configured limit into a percentage of
 187 * global dirtyable memory first.
 188 */
 189
 190static unsigned long highmem_dirtyable_memory(unsigned long total)
 191{
 192#ifdef CONFIG_HIGHMEM
 193        int node;
 194        unsigned long x = 0;
 195
 196        for_each_node_state(node, N_HIGH_MEMORY) {
 197                struct zone *z =
 198                        &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
 199
 200                x += zone_page_state(z, NR_FREE_PAGES) +
 201                     zone_reclaimable_pages(z) - z->dirty_balance_reserve;
 202        }
 203        /*
 204         * Make sure that the number of highmem pages is never larger
 205         * than the number of the total dirtyable memory. This can only
 206         * occur in very strange VM situations but we want to make sure
 207         * that this does not occur.
 208         */
 209        return min(x, total);
 210#else
 211        return 0;
 212#endif
 213}
 214
 215/**
 216 * global_dirtyable_memory - number of globally dirtyable pages
 217 *
 218 * Returns the global number of pages potentially available for dirty
 219 * page cache.  This is the base value for the global dirty limits.
 220 */
 221static unsigned long global_dirtyable_memory(void)
 222{
 223        unsigned long x;
 224
 225        x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages() -
 226            dirty_balance_reserve;
 227
 228        if (!vm_highmem_is_dirtyable)
 229                x -= highmem_dirtyable_memory(x);
 230
 231        return x + 1;   /* Ensure that we never return 0 */
 232}
 233
 234/*
 235 * global_dirty_limits - background-writeback and dirty-throttling thresholds
 236 *
 237 * Calculate the dirty thresholds based on sysctl parameters
 238 * - vm.dirty_background_ratio  or  vm.dirty_background_bytes
 239 * - vm.dirty_ratio             or  vm.dirty_bytes
 240 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
 241 * real-time tasks.
 242 */
 243void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
 244{
 245        unsigned long background;
 246        unsigned long dirty;
 247        unsigned long uninitialized_var(available_memory);
 248        struct task_struct *tsk;
 249
 250        if (!vm_dirty_bytes || !dirty_background_bytes)
 251                available_memory = global_dirtyable_memory();
 252
 253        if (vm_dirty_bytes)
 254                dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
 255        else
 256                dirty = (vm_dirty_ratio * available_memory) / 100;
 257
 258        if (dirty_background_bytes)
 259                background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
 260        else
 261                background = (dirty_background_ratio * available_memory) / 100;
 262
 263        if (background >= dirty)
 264                background = dirty / 2;
 265        tsk = current;
 266        if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
 267                background += background / 4;
 268                dirty += dirty / 4;
 269        }
 270        *pbackground = background;
 271        *pdirty = dirty;
 272        trace_global_dirty_state(background, dirty);
 273}
 274
 275/**
 276 * zone_dirtyable_memory - number of dirtyable pages in a zone
 277 * @zone: the zone
 278 *
 279 * Returns the zone's number of pages potentially available for dirty
 280 * page cache.  This is the base value for the per-zone dirty limits.
 281 */
 282static unsigned long zone_dirtyable_memory(struct zone *zone)
 283{
 284        /*
 285         * The effective global number of dirtyable pages may exclude
 286         * highmem as a big-picture measure to keep the ratio between
 287         * dirty memory and lowmem reasonable.
 288         *
 289         * But this function is purely about the individual zone and a
 290         * highmem zone can hold its share of dirty pages, so we don't
 291         * care about vm_highmem_is_dirtyable here.
 292         */
 293        return zone_page_state(zone, NR_FREE_PAGES) +
 294               zone_reclaimable_pages(zone) -
 295               zone->dirty_balance_reserve;
 296}
 297
 298/**
 299 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
 300 * @zone: the zone
 301 *
 302 * Returns the maximum number of dirty pages allowed in a zone, based
 303 * on the zone's dirtyable memory.
 304 */
 305static unsigned long zone_dirty_limit(struct zone *zone)
 306{
 307        unsigned long zone_memory = zone_dirtyable_memory(zone);
 308        struct task_struct *tsk = current;
 309        unsigned long dirty;
 310
 311        if (vm_dirty_bytes)
 312                dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
 313                        zone_memory / global_dirtyable_memory();
 314        else
 315                dirty = vm_dirty_ratio * zone_memory / 100;
 316
 317        if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
 318                dirty += dirty / 4;
 319
 320        return dirty;
 321}
 322
 323/**
 324 * zone_dirty_ok - tells whether a zone is within its dirty limits
 325 * @zone: the zone to check
 326 *
 327 * Returns %true when the dirty pages in @zone are within the zone's
 328 * dirty limit, %false if the limit is exceeded.
 329 */
 330bool zone_dirty_ok(struct zone *zone)
 331{
 332        unsigned long limit = zone_dirty_limit(zone);
 333
 334        return zone_page_state(zone, NR_FILE_DIRTY) +
 335               zone_page_state(zone, NR_UNSTABLE_NFS) +
 336               zone_page_state(zone, NR_WRITEBACK) <= limit;
 337}
 338
 339int dirty_background_ratio_handler(struct ctl_table *table, int write,
 340                void __user *buffer, size_t *lenp,
 341                loff_t *ppos)
 342{
 343        int ret;
 344
 345        ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 346        if (ret == 0 && write)
 347                dirty_background_bytes = 0;
 348        return ret;
 349}
 350
 351int dirty_background_bytes_handler(struct ctl_table *table, int write,
 352                void __user *buffer, size_t *lenp,
 353                loff_t *ppos)
 354{
 355        int ret;
 356
 357        ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
 358        if (ret == 0 && write)
 359                dirty_background_ratio = 0;
 360        return ret;
 361}
 362
 363int dirty_ratio_handler(struct ctl_table *table, int write,
 364                void __user *buffer, size_t *lenp,
 365                loff_t *ppos)
 366{
 367        int old_ratio = vm_dirty_ratio;
 368        int ret;
 369
 370        ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 371        if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
 372                writeback_set_ratelimit();
 373                vm_dirty_bytes = 0;
 374        }
 375        return ret;
 376}
 377
 378int dirty_bytes_handler(struct ctl_table *table, int write,
 379                void __user *buffer, size_t *lenp,
 380                loff_t *ppos)
 381{
 382        unsigned long old_bytes = vm_dirty_bytes;
 383        int ret;
 384
 385        ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
 386        if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
 387                writeback_set_ratelimit();
 388                vm_dirty_ratio = 0;
 389        }
 390        return ret;
 391}
 392
 393static unsigned long wp_next_time(unsigned long cur_time)
 394{
 395        cur_time += VM_COMPLETIONS_PERIOD_LEN;
 396        /* 0 has a special meaning... */
 397        if (!cur_time)
 398                return 1;
 399        return cur_time;
 400}
 401
 402/*
 403 * Increment the BDI's writeout completion count and the global writeout
 404 * completion count. Called from test_clear_page_writeback().
 405 */
 406static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
 407{
 408        __inc_bdi_stat(bdi, BDI_WRITTEN);
 409        __fprop_inc_percpu_max(&writeout_completions, &bdi->completions,
 410                               bdi->max_prop_frac);
 411        /* First event after period switching was turned off? */
 412        if (!unlikely(writeout_period_time)) {
 413                /*
 414                 * We can race with other __bdi_writeout_inc calls here but
 415                 * it does not cause any harm since the resulting time when
 416                 * timer will fire and what is in writeout_period_time will be
 417                 * roughly the same.
 418                 */
 419                writeout_period_time = wp_next_time(jiffies);
 420                mod_timer(&writeout_period_timer, writeout_period_time);
 421        }
 422}
 423
 424void bdi_writeout_inc(struct backing_dev_info *bdi)
 425{
 426        unsigned long flags;
 427
 428        local_irq_save(flags);
 429        __bdi_writeout_inc(bdi);
 430        local_irq_restore(flags);
 431}
 432EXPORT_SYMBOL_GPL(bdi_writeout_inc);
 433
 434/*
 435 * Obtain an accurate fraction of the BDI's portion.
 436 */
 437static void bdi_writeout_fraction(struct backing_dev_info *bdi,
 438                long *numerator, long *denominator)
 439{
 440        fprop_fraction_percpu(&writeout_completions, &bdi->completions,
 441                                numerator, denominator);
 442}
 443
 444/*
 445 * On idle system, we can be called long after we scheduled because we use
 446 * deferred timers so count with missed periods.
 447 */
 448static void writeout_period(unsigned long t)
 449{
 450        int miss_periods = (jiffies - writeout_period_time) /
 451                                                 VM_COMPLETIONS_PERIOD_LEN;
 452
 453        if (fprop_new_period(&writeout_completions, miss_periods + 1)) {
 454                writeout_period_time = wp_next_time(writeout_period_time +
 455                                miss_periods * VM_COMPLETIONS_PERIOD_LEN);
 456                mod_timer(&writeout_period_timer, writeout_period_time);
 457        } else {
 458                /*
 459                 * Aging has zeroed all fractions. Stop wasting CPU on period
 460                 * updates.
 461                 */
 462                writeout_period_time = 0;
 463        }
 464}
 465
 466/*
 467 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
 468 * registered backing devices, which, for obvious reasons, can not
 469 * exceed 100%.
 470 */
 471static unsigned int bdi_min_ratio;
 472
 473int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
 474{
 475        int ret = 0;
 476
 477        spin_lock_bh(&bdi_lock);
 478        if (min_ratio > bdi->max_ratio) {
 479                ret = -EINVAL;
 480        } else {
 481                min_ratio -= bdi->min_ratio;
 482                if (bdi_min_ratio + min_ratio < 100) {
 483                        bdi_min_ratio += min_ratio;
 484                        bdi->min_ratio += min_ratio;
 485                } else {
 486                        ret = -EINVAL;
 487                }
 488        }
 489        spin_unlock_bh(&bdi_lock);
 490
 491        return ret;
 492}
 493
 494int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
 495{
 496        int ret = 0;
 497
 498        if (max_ratio > 100)
 499                return -EINVAL;
 500
 501        spin_lock_bh(&bdi_lock);
 502        if (bdi->min_ratio > max_ratio) {
 503                ret = -EINVAL;
 504        } else {
 505                bdi->max_ratio = max_ratio;
 506                bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
 507        }
 508        spin_unlock_bh(&bdi_lock);
 509
 510        return ret;
 511}
 512EXPORT_SYMBOL(bdi_set_max_ratio);
 513
 514static unsigned long dirty_freerun_ceiling(unsigned long thresh,
 515                                           unsigned long bg_thresh)
 516{
 517        return (thresh + bg_thresh) / 2;
 518}
 519
 520static unsigned long hard_dirty_limit(unsigned long thresh)
 521{
 522        return max(thresh, global_dirty_limit);
 523}
 524
 525/**
 526 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
 527 * @bdi: the backing_dev_info to query
 528 * @dirty: global dirty limit in pages
 529 *
 530 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
 531 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
 532 *
 533 * Note that balance_dirty_pages() will only seriously take it as a hard limit
 534 * when sleeping max_pause per page is not enough to keep the dirty pages under
 535 * control. For example, when the device is completely stalled due to some error
 536 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
 537 * In the other normal situations, it acts more gently by throttling the tasks
 538 * more (rather than completely block them) when the bdi dirty pages go high.
 539 *
 540 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
 541 * - starving fast devices
 542 * - piling up dirty pages (that will take long time to sync) on slow devices
 543 *
 544 * The bdi's share of dirty limit will be adapting to its throughput and
 545 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
 546 */
 547unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
 548{
 549        u64 bdi_dirty;
 550        long numerator, denominator;
 551
 552        /*
 553         * Calculate this BDI's share of the dirty ratio.
 554         */
 555        bdi_writeout_fraction(bdi, &numerator, &denominator);
 556
 557        bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
 558        bdi_dirty *= numerator;
 559        do_div(bdi_dirty, denominator);
 560
 561        bdi_dirty += (dirty * bdi->min_ratio) / 100;
 562        if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
 563                bdi_dirty = dirty * bdi->max_ratio / 100;
 564
 565        return bdi_dirty;
 566}
 567
 568/*
 569 * Dirty position control.
 570 *
 571 * (o) global/bdi setpoints
 572 *
 573 * We want the dirty pages be balanced around the global/bdi setpoints.
 574 * When the number of dirty pages is higher/lower than the setpoint, the
 575 * dirty position control ratio (and hence task dirty ratelimit) will be
 576 * decreased/increased to bring the dirty pages back to the setpoint.
 577 *
 578 *     pos_ratio = 1 << RATELIMIT_CALC_SHIFT
 579 *
 580 *     if (dirty < setpoint) scale up   pos_ratio
 581 *     if (dirty > setpoint) scale down pos_ratio
 582 *
 583 *     if (bdi_dirty < bdi_setpoint) scale up   pos_ratio
 584 *     if (bdi_dirty > bdi_setpoint) scale down pos_ratio
 585 *
 586 *     task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
 587 *
 588 * (o) global control line
 589 *
 590 *     ^ pos_ratio
 591 *     |
 592 *     |            |<===== global dirty control scope ======>|
 593 * 2.0 .............*
 594 *     |            .*
 595 *     |            . *
 596 *     |            .   *
 597 *     |            .     *
 598 *     |            .        *
 599 *     |            .            *
 600 * 1.0 ................................*
 601 *     |            .                  .     *
 602 *     |            .                  .          *
 603 *     |            .                  .              *
 604 *     |            .                  .                 *
 605 *     |            .                  .                    *
 606 *   0 +------------.------------------.----------------------*------------->
 607 *           freerun^          setpoint^                 limit^   dirty pages
 608 *
 609 * (o) bdi control line
 610 *
 611 *     ^ pos_ratio
 612 *     |
 613 *     |            *
 614 *     |              *
 615 *     |                *
 616 *     |                  *
 617 *     |                    * |<=========== span ============>|
 618 * 1.0 .......................*
 619 *     |                      . *
 620 *     |                      .   *
 621 *     |                      .     *
 622 *     |                      .       *
 623 *     |                      .         *
 624 *     |                      .           *
 625 *     |                      .             *
 626 *     |                      .               *
 627 *     |                      .                 *
 628 *     |                      .                   *
 629 *     |                      .                     *
 630 * 1/4 ...............................................* * * * * * * * * * * *
 631 *     |                      .                         .
 632 *     |                      .                           .
 633 *     |                      .                             .
 634 *   0 +----------------------.-------------------------------.------------->
 635 *                bdi_setpoint^                    x_intercept^
 636 *
 637 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
 638 * be smoothly throttled down to normal if it starts high in situations like
 639 * - start writing to a slow SD card and a fast disk at the same time. The SD
 640 *   card's bdi_dirty may rush to many times higher than bdi_setpoint.
 641 * - the bdi dirty thresh drops quickly due to change of JBOD workload
 642 */
 643static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
 644                                        unsigned long thresh,
 645                                        unsigned long bg_thresh,
 646                                        unsigned long dirty,
 647                                        unsigned long bdi_thresh,
 648                                        unsigned long bdi_dirty)
 649{
 650        unsigned long write_bw = bdi->avg_write_bandwidth;
 651        unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
 652        unsigned long limit = hard_dirty_limit(thresh);
 653        unsigned long x_intercept;
 654        unsigned long setpoint;         /* dirty pages' target balance point */
 655        unsigned long bdi_setpoint;
 656        unsigned long span;
 657        long long pos_ratio;            /* for scaling up/down the rate limit */
 658        long x;
 659
 660        if (unlikely(dirty >= limit))
 661                return 0;
 662
 663        /*
 664         * global setpoint
 665         *
 666         *                           setpoint - dirty 3
 667         *        f(dirty) := 1.0 + (----------------)
 668         *                           limit - setpoint
 669         *
 670         * it's a 3rd order polynomial that subjects to
 671         *
 672         * (1) f(freerun)  = 2.0 => rampup dirty_ratelimit reasonably fast
 673         * (2) f(setpoint) = 1.0 => the balance point
 674         * (3) f(limit)    = 0   => the hard limit
 675         * (4) df/dx      <= 0   => negative feedback control
 676         * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
 677         *     => fast response on large errors; small oscillation near setpoint
 678         */
 679        setpoint = (freerun + limit) / 2;
 680        x = div_s64((setpoint - dirty) << RATELIMIT_CALC_SHIFT,
 681                    limit - setpoint + 1);
 682        pos_ratio = x;
 683        pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
 684        pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
 685        pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
 686
 687        /*
 688         * We have computed basic pos_ratio above based on global situation. If
 689         * the bdi is over/under its share of dirty pages, we want to scale
 690         * pos_ratio further down/up. That is done by the following mechanism.
 691         */
 692
 693        /*
 694         * bdi setpoint
 695         *
 696         *        f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
 697         *
 698         *                        x_intercept - bdi_dirty
 699         *                     := --------------------------
 700         *                        x_intercept - bdi_setpoint
 701         *
 702         * The main bdi control line is a linear function that subjects to
 703         *
 704         * (1) f(bdi_setpoint) = 1.0
 705         * (2) k = - 1 / (8 * write_bw)  (in single bdi case)
 706         *     or equally: x_intercept = bdi_setpoint + 8 * write_bw
 707         *
 708         * For single bdi case, the dirty pages are observed to fluctuate
 709         * regularly within range
 710         *        [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
 711         * for various filesystems, where (2) can yield in a reasonable 12.5%
 712         * fluctuation range for pos_ratio.
 713         *
 714         * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
 715         * own size, so move the slope over accordingly and choose a slope that
 716         * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
 717         */
 718        if (unlikely(bdi_thresh > thresh))
 719                bdi_thresh = thresh;
 720        /*
 721         * It's very possible that bdi_thresh is close to 0 not because the
 722         * device is slow, but that it has remained inactive for long time.
 723         * Honour such devices a reasonable good (hopefully IO efficient)
 724         * threshold, so that the occasional writes won't be blocked and active
 725         * writes can rampup the threshold quickly.
 726         */
 727        bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
 728        /*
 729         * scale global setpoint to bdi's:
 730         *      bdi_setpoint = setpoint * bdi_thresh / thresh
 731         */
 732        x = div_u64((u64)bdi_thresh << 16, thresh + 1);
 733        bdi_setpoint = setpoint * (u64)x >> 16;
 734        /*
 735         * Use span=(8*write_bw) in single bdi case as indicated by
 736         * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
 737         *
 738         *        bdi_thresh                    thresh - bdi_thresh
 739         * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
 740         *          thresh                            thresh
 741         */
 742        span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
 743        x_intercept = bdi_setpoint + span;
 744
 745        if (bdi_dirty < x_intercept - span / 4) {
 746                pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
 747                                    x_intercept - bdi_setpoint + 1);
 748        } else
 749                pos_ratio /= 4;
 750
 751        /*
 752         * bdi reserve area, safeguard against dirty pool underrun and disk idle
 753         * It may push the desired control point of global dirty pages higher
 754         * than setpoint.
 755         */
 756        x_intercept = bdi_thresh / 2;
 757        if (bdi_dirty < x_intercept) {
 758                if (bdi_dirty > x_intercept / 8)
 759                        pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
 760                else
 761                        pos_ratio *= 8;
 762        }
 763
 764        return pos_ratio;
 765}
 766
 767static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
 768                                       unsigned long elapsed,
 769                                       unsigned long written)
 770{
 771        const unsigned long period = roundup_pow_of_two(3 * HZ);
 772        unsigned long avg = bdi->avg_write_bandwidth;
 773        unsigned long old = bdi->write_bandwidth;
 774        u64 bw;
 775
 776        /*
 777         * bw = written * HZ / elapsed
 778         *
 779         *                   bw * elapsed + write_bandwidth * (period - elapsed)
 780         * write_bandwidth = ---------------------------------------------------
 781         *                                          period
 782         */
 783        bw = written - bdi->written_stamp;
 784        bw *= HZ;
 785        if (unlikely(elapsed > period)) {
 786                do_div(bw, elapsed);
 787                avg = bw;
 788                goto out;
 789        }
 790        bw += (u64)bdi->write_bandwidth * (period - elapsed);
 791        bw >>= ilog2(period);
 792
 793        /*
 794         * one more level of smoothing, for filtering out sudden spikes
 795         */
 796        if (avg > old && old >= (unsigned long)bw)
 797                avg -= (avg - old) >> 3;
 798
 799        if (avg < old && old <= (unsigned long)bw)
 800                avg += (old - avg) >> 3;
 801
 802out:
 803        bdi->write_bandwidth = bw;
 804        bdi->avg_write_bandwidth = avg;
 805}
 806
 807/*
 808 * The global dirtyable memory and dirty threshold could be suddenly knocked
 809 * down by a large amount (eg. on the startup of KVM in a swapless system).
 810 * This may throw the system into deep dirty exceeded state and throttle
 811 * heavy/light dirtiers alike. To retain good responsiveness, maintain
 812 * global_dirty_limit for tracking slowly down to the knocked down dirty
 813 * threshold.
 814 */
 815static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
 816{
 817        unsigned long limit = global_dirty_limit;
 818
 819        /*
 820         * Follow up in one step.
 821         */
 822        if (limit < thresh) {
 823                limit = thresh;
 824                goto update;
 825        }
 826
 827        /*
 828         * Follow down slowly. Use the higher one as the target, because thresh
 829         * may drop below dirty. This is exactly the reason to introduce
 830         * global_dirty_limit which is guaranteed to lie above the dirty pages.
 831         */
 832        thresh = max(thresh, dirty);
 833        if (limit > thresh) {
 834                limit -= (limit - thresh) >> 5;
 835                goto update;
 836        }
 837        return;
 838update:
 839        global_dirty_limit = limit;
 840}
 841
 842static void global_update_bandwidth(unsigned long thresh,
 843                                    unsigned long dirty,
 844                                    unsigned long now)
 845{
 846        static DEFINE_SPINLOCK(dirty_lock);
 847        static unsigned long update_time;
 848
 849        /*
 850         * check locklessly first to optimize away locking for the most time
 851         */
 852        if (time_before(now, update_time + BANDWIDTH_INTERVAL))
 853                return;
 854
 855        spin_lock(&dirty_lock);
 856        if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
 857                update_dirty_limit(thresh, dirty);
 858                update_time = now;
 859        }
 860        spin_unlock(&dirty_lock);
 861}
 862
 863/*
 864 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
 865 *
 866 * Normal bdi tasks will be curbed at or below it in long term.
 867 * Obviously it should be around (write_bw / N) when there are N dd tasks.
 868 */
 869static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
 870                                       unsigned long thresh,
 871                                       unsigned long bg_thresh,
 872                                       unsigned long dirty,
 873                                       unsigned long bdi_thresh,
 874                                       unsigned long bdi_dirty,
 875                                       unsigned long dirtied,
 876                                       unsigned long elapsed)
 877{
 878        unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
 879        unsigned long limit = hard_dirty_limit(thresh);
 880        unsigned long setpoint = (freerun + limit) / 2;
 881        unsigned long write_bw = bdi->avg_write_bandwidth;
 882        unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
 883        unsigned long dirty_rate;
 884        unsigned long task_ratelimit;
 885        unsigned long balanced_dirty_ratelimit;
 886        unsigned long pos_ratio;
 887        unsigned long step;
 888        unsigned long x;
 889
 890        /*
 891         * The dirty rate will match the writeout rate in long term, except
 892         * when dirty pages are truncated by userspace or re-dirtied by FS.
 893         */
 894        dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
 895
 896        pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
 897                                       bdi_thresh, bdi_dirty);
 898        /*
 899         * task_ratelimit reflects each dd's dirty rate for the past 200ms.
 900         */
 901        task_ratelimit = (u64)dirty_ratelimit *
 902                                        pos_ratio >> RATELIMIT_CALC_SHIFT;
 903        task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
 904
 905        /*
 906         * A linear estimation of the "balanced" throttle rate. The theory is,
 907         * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
 908         * dirty_rate will be measured to be (N * task_ratelimit). So the below
 909         * formula will yield the balanced rate limit (write_bw / N).
 910         *
 911         * Note that the expanded form is not a pure rate feedback:
 912         *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate)              (1)
 913         * but also takes pos_ratio into account:
 914         *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio  (2)
 915         *
 916         * (1) is not realistic because pos_ratio also takes part in balancing
 917         * the dirty rate.  Consider the state
 918         *      pos_ratio = 0.5                                              (3)
 919         *      rate = 2 * (write_bw / N)                                    (4)
 920         * If (1) is used, it will stuck in that state! Because each dd will
 921         * be throttled at
 922         *      task_ratelimit = pos_ratio * rate = (write_bw / N)           (5)
 923         * yielding
 924         *      dirty_rate = N * task_ratelimit = write_bw                   (6)
 925         * put (6) into (1) we get
 926         *      rate_(i+1) = rate_(i)                                        (7)
 927         *
 928         * So we end up using (2) to always keep
 929         *      rate_(i+1) ~= (write_bw / N)                                 (8)
 930         * regardless of the value of pos_ratio. As long as (8) is satisfied,
 931         * pos_ratio is able to drive itself to 1.0, which is not only where
 932         * the dirty count meet the setpoint, but also where the slope of
 933         * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
 934         */
 935        balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
 936                                           dirty_rate | 1);
 937        /*
 938         * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
 939         */
 940        if (unlikely(balanced_dirty_ratelimit > write_bw))
 941                balanced_dirty_ratelimit = write_bw;
 942
 943        /*
 944         * We could safely do this and return immediately:
 945         *
 946         *      bdi->dirty_ratelimit = balanced_dirty_ratelimit;
 947         *
 948         * However to get a more stable dirty_ratelimit, the below elaborated
 949         * code makes use of task_ratelimit to filter out singular points and
 950         * limit the step size.
 951         *
 952         * The below code essentially only uses the relative value of
 953         *
 954         *      task_ratelimit - dirty_ratelimit
 955         *      = (pos_ratio - 1) * dirty_ratelimit
 956         *
 957         * which reflects the direction and size of dirty position error.
 958         */
 959
 960        /*
 961         * dirty_ratelimit will follow balanced_dirty_ratelimit iff
 962         * task_ratelimit is on the same side of dirty_ratelimit, too.
 963         * For example, when
 964         * - dirty_ratelimit > balanced_dirty_ratelimit
 965         * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
 966         * lowering dirty_ratelimit will help meet both the position and rate
 967         * control targets. Otherwise, don't update dirty_ratelimit if it will
 968         * only help meet the rate target. After all, what the users ultimately
 969         * feel and care are stable dirty rate and small position error.
 970         *
 971         * |task_ratelimit - dirty_ratelimit| is used to limit the step size
 972         * and filter out the singular points of balanced_dirty_ratelimit. Which
 973         * keeps jumping around randomly and can even leap far away at times
 974         * due to the small 200ms estimation period of dirty_rate (we want to
 975         * keep that period small to reduce time lags).
 976         */
 977        step = 0;
 978        if (dirty < setpoint) {
 979                x = min(bdi->balanced_dirty_ratelimit,
 980                         min(balanced_dirty_ratelimit, task_ratelimit));
 981                if (dirty_ratelimit < x)
 982                        step = x - dirty_ratelimit;
 983        } else {
 984                x = max(bdi->balanced_dirty_ratelimit,
 985                         max(balanced_dirty_ratelimit, task_ratelimit));
 986                if (dirty_ratelimit > x)
 987                        step = dirty_ratelimit - x;
 988        }
 989
 990        /*
 991         * Don't pursue 100% rate matching. It's impossible since the balanced
 992         * rate itself is constantly fluctuating. So decrease the track speed
 993         * when it gets close to the target. Helps eliminate pointless tremors.
 994         */
 995        step >>= dirty_ratelimit / (2 * step + 1);
 996        /*
 997         * Limit the tracking speed to avoid overshooting.
 998         */
 999        step = (step + 7) / 8;
1000
1001        if (dirty_ratelimit < balanced_dirty_ratelimit)
1002                dirty_ratelimit += step;
1003        else
1004                dirty_ratelimit -= step;
1005
1006        bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1007        bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1008
1009        trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
1010}
1011
1012void __bdi_update_bandwidth(struct backing_dev_info *bdi,
1013                            unsigned long thresh,
1014                            unsigned long bg_thresh,
1015                            unsigned long dirty,
1016                            unsigned long bdi_thresh,
1017                            unsigned long bdi_dirty,
1018                            unsigned long start_time)
1019{
1020        unsigned long now = jiffies;
1021        unsigned long elapsed = now - bdi->bw_time_stamp;
1022        unsigned long dirtied;
1023        unsigned long written;
1024
1025        /*
1026         * rate-limit, only update once every 200ms.
1027         */
1028        if (elapsed < BANDWIDTH_INTERVAL)
1029                return;
1030
1031        dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
1032        written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
1033
1034        /*
1035         * Skip quiet periods when disk bandwidth is under-utilized.
1036         * (at least 1s idle time between two flusher runs)
1037         */
1038        if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
1039                goto snapshot;
1040
1041        if (thresh) {
1042                global_update_bandwidth(thresh, dirty, now);
1043                bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
1044                                           bdi_thresh, bdi_dirty,
1045                                           dirtied, elapsed);
1046        }
1047        bdi_update_write_bandwidth(bdi, elapsed, written);
1048
1049snapshot:
1050        bdi->dirtied_stamp = dirtied;
1051        bdi->written_stamp = written;
1052        bdi->bw_time_stamp = now;
1053}
1054
1055static void bdi_update_bandwidth(struct backing_dev_info *bdi,
1056                                 unsigned long thresh,
1057                                 unsigned long bg_thresh,
1058                                 unsigned long dirty,
1059                                 unsigned long bdi_thresh,
1060                                 unsigned long bdi_dirty,
1061                                 unsigned long start_time)
1062{
1063        if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
1064                return;
1065        spin_lock(&bdi->wb.list_lock);
1066        __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
1067                               bdi_thresh, bdi_dirty, start_time);
1068        spin_unlock(&bdi->wb.list_lock);
1069}
1070
1071/*
1072 * After a task dirtied this many pages, balance_dirty_pages_ratelimited_nr()
1073 * will look to see if it needs to start dirty throttling.
1074 *
1075 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1076 * global_page_state() too often. So scale it near-sqrt to the safety margin
1077 * (the number of pages we may dirty without exceeding the dirty limits).
1078 */
1079static unsigned long dirty_poll_interval(unsigned long dirty,
1080                                         unsigned long thresh)
1081{
1082        if (thresh > dirty)
1083                return 1UL << (ilog2(thresh - dirty) >> 1);
1084
1085        return 1;
1086}
1087
1088static long bdi_max_pause(struct backing_dev_info *bdi,
1089                          unsigned long bdi_dirty)
1090{
1091        long bw = bdi->avg_write_bandwidth;
1092        long t;
1093
1094        /*
1095         * Limit pause time for small memory systems. If sleeping for too long
1096         * time, a small pool of dirty/writeback pages may go empty and disk go
1097         * idle.
1098         *
1099         * 8 serves as the safety ratio.
1100         */
1101        t = bdi_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1102        t++;
1103
1104        return min_t(long, t, MAX_PAUSE);
1105}
1106
1107static long bdi_min_pause(struct backing_dev_info *bdi,
1108                          long max_pause,
1109                          unsigned long task_ratelimit,
1110                          unsigned long dirty_ratelimit,
1111                          int *nr_dirtied_pause)
1112{
1113        long hi = ilog2(bdi->avg_write_bandwidth);
1114        long lo = ilog2(bdi->dirty_ratelimit);
1115        long t;         /* target pause */
1116        long pause;     /* estimated next pause */
1117        int pages;      /* target nr_dirtied_pause */
1118
1119        /* target for 10ms pause on 1-dd case */
1120        t = max(1, HZ / 100);
1121
1122        /*
1123         * Scale up pause time for concurrent dirtiers in order to reduce CPU
1124         * overheads.
1125         *
1126         * (N * 10ms) on 2^N concurrent tasks.
1127         */
1128        if (hi > lo)
1129                t += (hi - lo) * (10 * HZ) / 1024;
1130
1131        /*
1132         * This is a bit convoluted. We try to base the next nr_dirtied_pause
1133         * on the much more stable dirty_ratelimit. However the next pause time
1134         * will be computed based on task_ratelimit and the two rate limits may
1135         * depart considerably at some time. Especially if task_ratelimit goes
1136         * below dirty_ratelimit/2 and the target pause is max_pause, the next
1137         * pause time will be max_pause*2 _trimmed down_ to max_pause.  As a
1138         * result task_ratelimit won't be executed faithfully, which could
1139         * eventually bring down dirty_ratelimit.
1140         *
1141         * We apply two rules to fix it up:
1142         * 1) try to estimate the next pause time and if necessary, use a lower
1143         *    nr_dirtied_pause so as not to exceed max_pause. When this happens,
1144         *    nr_dirtied_pause will be "dancing" with task_ratelimit.
1145         * 2) limit the target pause time to max_pause/2, so that the normal
1146         *    small fluctuations of task_ratelimit won't trigger rule (1) and
1147         *    nr_dirtied_pause will remain as stable as dirty_ratelimit.
1148         */
1149        t = min(t, 1 + max_pause / 2);
1150        pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1151
1152        /*
1153         * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1154         * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1155         * When the 16 consecutive reads are often interrupted by some dirty
1156         * throttling pause during the async writes, cfq will go into idles
1157         * (deadline is fine). So push nr_dirtied_pause as high as possible
1158         * until reaches DIRTY_POLL_THRESH=32 pages.
1159         */
1160        if (pages < DIRTY_POLL_THRESH) {
1161                t = max_pause;
1162                pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1163                if (pages > DIRTY_POLL_THRESH) {
1164                        pages = DIRTY_POLL_THRESH;
1165                        t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1166                }
1167        }
1168
1169        pause = HZ * pages / (task_ratelimit + 1);
1170        if (pause > max_pause) {
1171                t = max_pause;
1172                pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1173        }
1174
1175        *nr_dirtied_pause = pages;
1176        /*
1177         * The minimal pause time will normally be half the target pause time.
1178         */
1179        return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1180}
1181
1182/*
1183 * balance_dirty_pages() must be called by processes which are generating dirty
1184 * data.  It looks at the number of dirty pages in the machine and will force
1185 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1186 * If we're over `background_thresh' then the writeback threads are woken to
1187 * perform some writeout.
1188 */
1189static void balance_dirty_pages(struct address_space *mapping,
1190                                unsigned long pages_dirtied)
1191{
1192        unsigned long nr_reclaimable;   /* = file_dirty + unstable_nfs */
1193        unsigned long bdi_reclaimable;
1194        unsigned long nr_dirty;  /* = file_dirty + writeback + unstable_nfs */
1195        unsigned long bdi_dirty;
1196        unsigned long freerun;
1197        unsigned long background_thresh;
1198        unsigned long dirty_thresh;
1199        unsigned long bdi_thresh;
1200        long period;
1201        long pause;
1202        long max_pause;
1203        long min_pause;
1204        int nr_dirtied_pause;
1205        bool dirty_exceeded = false;
1206        unsigned long task_ratelimit;
1207        unsigned long dirty_ratelimit;
1208        unsigned long pos_ratio;
1209        struct backing_dev_info *bdi = mapping->backing_dev_info;
1210        unsigned long start_time = jiffies;
1211
1212        for (;;) {
1213                unsigned long now = jiffies;
1214
1215                /*
1216                 * Unstable writes are a feature of certain networked
1217                 * filesystems (i.e. NFS) in which data may have been
1218                 * written to the server's write cache, but has not yet
1219                 * been flushed to permanent storage.
1220                 */
1221                nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1222                                        global_page_state(NR_UNSTABLE_NFS);
1223                nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1224
1225                global_dirty_limits(&background_thresh, &dirty_thresh);
1226
1227                /*
1228                 * Throttle it only when the background writeback cannot
1229                 * catch-up. This avoids (excessively) small writeouts
1230                 * when the bdi limits are ramping up.
1231                 */
1232                freerun = dirty_freerun_ceiling(dirty_thresh,
1233                                                background_thresh);
1234                if (nr_dirty <= freerun) {
1235                        current->dirty_paused_when = now;
1236                        current->nr_dirtied = 0;
1237                        current->nr_dirtied_pause =
1238                                dirty_poll_interval(nr_dirty, dirty_thresh);
1239                        break;
1240                }
1241
1242                if (unlikely(!writeback_in_progress(bdi)))
1243                        bdi_start_background_writeback(bdi);
1244
1245                /*
1246                 * bdi_thresh is not treated as some limiting factor as
1247                 * dirty_thresh, due to reasons
1248                 * - in JBOD setup, bdi_thresh can fluctuate a lot
1249                 * - in a system with HDD and USB key, the USB key may somehow
1250                 *   go into state (bdi_dirty >> bdi_thresh) either because
1251                 *   bdi_dirty starts high, or because bdi_thresh drops low.
1252                 *   In this case we don't want to hard throttle the USB key
1253                 *   dirtiers for 100 seconds until bdi_dirty drops under
1254                 *   bdi_thresh. Instead the auxiliary bdi control line in
1255                 *   bdi_position_ratio() will let the dirtier task progress
1256                 *   at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1257                 */
1258                bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
1259
1260                /*
1261                 * In order to avoid the stacked BDI deadlock we need
1262                 * to ensure we accurately count the 'dirty' pages when
1263                 * the threshold is low.
1264                 *
1265                 * Otherwise it would be possible to get thresh+n pages
1266                 * reported dirty, even though there are thresh-m pages
1267                 * actually dirty; with m+n sitting in the percpu
1268                 * deltas.
1269                 */
1270                if (bdi_thresh < 2 * bdi_stat_error(bdi)) {
1271                        bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1272                        bdi_dirty = bdi_reclaimable +
1273                                    bdi_stat_sum(bdi, BDI_WRITEBACK);
1274                } else {
1275                        bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1276                        bdi_dirty = bdi_reclaimable +
1277                                    bdi_stat(bdi, BDI_WRITEBACK);
1278                }
1279
1280                dirty_exceeded = (bdi_dirty > bdi_thresh) &&
1281                                  (nr_dirty > dirty_thresh);
1282                if (dirty_exceeded && !bdi->dirty_exceeded)
1283                        bdi->dirty_exceeded = 1;
1284
1285                bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1286                                     nr_dirty, bdi_thresh, bdi_dirty,
1287                                     start_time);
1288
1289                dirty_ratelimit = bdi->dirty_ratelimit;
1290                pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1291                                               background_thresh, nr_dirty,
1292                                               bdi_thresh, bdi_dirty);
1293                task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1294                                                        RATELIMIT_CALC_SHIFT;
1295                max_pause = bdi_max_pause(bdi, bdi_dirty);
1296                min_pause = bdi_min_pause(bdi, max_pause,
1297                                          task_ratelimit, dirty_ratelimit,
1298                                          &nr_dirtied_pause);
1299
1300                if (unlikely(task_ratelimit == 0)) {
1301                        period = max_pause;
1302                        pause = max_pause;
1303                        goto pause;
1304                }
1305                period = HZ * pages_dirtied / task_ratelimit;
1306                pause = period;
1307                if (current->dirty_paused_when)
1308                        pause -= now - current->dirty_paused_when;
1309                /*
1310                 * For less than 1s think time (ext3/4 may block the dirtier
1311                 * for up to 800ms from time to time on 1-HDD; so does xfs,
1312                 * however at much less frequency), try to compensate it in
1313                 * future periods by updating the virtual time; otherwise just
1314                 * do a reset, as it may be a light dirtier.
1315                 */
1316                if (pause < min_pause) {
1317                        trace_balance_dirty_pages(bdi,
1318                                                  dirty_thresh,
1319                                                  background_thresh,
1320                                                  nr_dirty,
1321                                                  bdi_thresh,
1322                                                  bdi_dirty,
1323                                                  dirty_ratelimit,
1324                                                  task_ratelimit,
1325                                                  pages_dirtied,
1326                                                  period,
1327                                                  min(pause, 0L),
1328                                                  start_time);
1329                        if (pause < -HZ) {
1330                                current->dirty_paused_when = now;
1331                                current->nr_dirtied = 0;
1332                        } else if (period) {
1333                                current->dirty_paused_when += period;
1334                                current->nr_dirtied = 0;
1335                        } else if (current->nr_dirtied_pause <= pages_dirtied)
1336                                current->nr_dirtied_pause += pages_dirtied;
1337                        break;
1338                }
1339                if (unlikely(pause > max_pause)) {
1340                        /* for occasional dropped task_ratelimit */
1341                        now += min(pause - max_pause, max_pause);
1342                        pause = max_pause;
1343                }
1344
1345pause:
1346                trace_balance_dirty_pages(bdi,
1347                                          dirty_thresh,
1348                                          background_thresh,
1349                                          nr_dirty,
1350                                          bdi_thresh,
1351                                          bdi_dirty,
1352                                          dirty_ratelimit,
1353                                          task_ratelimit,
1354                                          pages_dirtied,
1355                                          period,
1356                                          pause,
1357                                          start_time);
1358                __set_current_state(TASK_KILLABLE);
1359                io_schedule_timeout(pause);
1360
1361                current->dirty_paused_when = now + pause;
1362                current->nr_dirtied = 0;
1363                current->nr_dirtied_pause = nr_dirtied_pause;
1364
1365                /*
1366                 * This is typically equal to (nr_dirty < dirty_thresh) and can
1367                 * also keep "1000+ dd on a slow USB stick" under control.
1368                 */
1369                if (task_ratelimit)
1370                        break;
1371
1372                /*
1373                 * In the case of an unresponding NFS server and the NFS dirty
1374                 * pages exceeds dirty_thresh, give the other good bdi's a pipe
1375                 * to go through, so that tasks on them still remain responsive.
1376                 *
1377                 * In theory 1 page is enough to keep the comsumer-producer
1378                 * pipe going: the flusher cleans 1 page => the task dirties 1
1379                 * more page. However bdi_dirty has accounting errors.  So use
1380                 * the larger and more IO friendly bdi_stat_error.
1381                 */
1382                if (bdi_dirty <= bdi_stat_error(bdi))
1383                        break;
1384
1385                if (fatal_signal_pending(current))
1386                        break;
1387        }
1388
1389        if (!dirty_exceeded && bdi->dirty_exceeded)
1390                bdi->dirty_exceeded = 0;
1391
1392        if (writeback_in_progress(bdi))
1393                return;
1394
1395        /*
1396         * In laptop mode, we wait until hitting the higher threshold before
1397         * starting background writeout, and then write out all the way down
1398         * to the lower threshold.  So slow writers cause minimal disk activity.
1399         *
1400         * In normal mode, we start background writeout at the lower
1401         * background_thresh, to keep the amount of dirty memory low.
1402         */
1403        if (laptop_mode)
1404                return;
1405
1406        if (nr_reclaimable > background_thresh)
1407                bdi_start_background_writeback(bdi);
1408}
1409
1410void set_page_dirty_balance(struct page *page, int page_mkwrite)
1411{
1412        if (set_page_dirty(page) || page_mkwrite) {
1413                struct address_space *mapping = page_mapping(page);
1414
1415                if (mapping)
1416                        balance_dirty_pages_ratelimited(mapping);
1417        }
1418}
1419
1420static DEFINE_PER_CPU(int, bdp_ratelimits);
1421
1422/*
1423 * Normal tasks are throttled by
1424 *      loop {
1425 *              dirty tsk->nr_dirtied_pause pages;
1426 *              take a snap in balance_dirty_pages();
1427 *      }
1428 * However there is a worst case. If every task exit immediately when dirtied
1429 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1430 * called to throttle the page dirties. The solution is to save the not yet
1431 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1432 * randomly into the running tasks. This works well for the above worst case,
1433 * as the new task will pick up and accumulate the old task's leaked dirty
1434 * count and eventually get throttled.
1435 */
1436DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1437
1438/**
1439 * balance_dirty_pages_ratelimited_nr - balance dirty memory state
1440 * @mapping: address_space which was dirtied
1441 * @nr_pages_dirtied: number of pages which the caller has just dirtied
1442 *
1443 * Processes which are dirtying memory should call in here once for each page
1444 * which was newly dirtied.  The function will periodically check the system's
1445 * dirty state and will initiate writeback if needed.
1446 *
1447 * On really big machines, get_writeback_state is expensive, so try to avoid
1448 * calling it too often (ratelimiting).  But once we're over the dirty memory
1449 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1450 * from overshooting the limit by (ratelimit_pages) each.
1451 */
1452void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
1453                                        unsigned long nr_pages_dirtied)
1454{
1455        struct backing_dev_info *bdi = mapping->backing_dev_info;
1456        int ratelimit;
1457        int *p;
1458
1459        if (!bdi_cap_account_dirty(bdi))
1460                return;
1461
1462        ratelimit = current->nr_dirtied_pause;
1463        if (bdi->dirty_exceeded)
1464                ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1465
1466        preempt_disable();
1467        /*
1468         * This prevents one CPU to accumulate too many dirtied pages without
1469         * calling into balance_dirty_pages(), which can happen when there are
1470         * 1000+ tasks, all of them start dirtying pages at exactly the same
1471         * time, hence all honoured too large initial task->nr_dirtied_pause.
1472         */
1473        p =  &__get_cpu_var(bdp_ratelimits);
1474        if (unlikely(current->nr_dirtied >= ratelimit))
1475                *p = 0;
1476        else if (unlikely(*p >= ratelimit_pages)) {
1477                *p = 0;
1478                ratelimit = 0;
1479        }
1480        /*
1481         * Pick up the dirtied pages by the exited tasks. This avoids lots of
1482         * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1483         * the dirty throttling and livelock other long-run dirtiers.
1484         */
1485        p = &__get_cpu_var(dirty_throttle_leaks);
1486        if (*p > 0 && current->nr_dirtied < ratelimit) {
1487                nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1488                *p -= nr_pages_dirtied;
1489                current->nr_dirtied += nr_pages_dirtied;
1490        }
1491        preempt_enable();
1492
1493        if (unlikely(current->nr_dirtied >= ratelimit))
1494                balance_dirty_pages(mapping, current->nr_dirtied);
1495}
1496EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
1497
1498void throttle_vm_writeout(gfp_t gfp_mask)
1499{
1500        unsigned long background_thresh;
1501        unsigned long dirty_thresh;
1502
1503        for ( ; ; ) {
1504                global_dirty_limits(&background_thresh, &dirty_thresh);
1505                dirty_thresh = hard_dirty_limit(dirty_thresh);
1506
1507                /*
1508                 * Boost the allowable dirty threshold a bit for page
1509                 * allocators so they don't get DoS'ed by heavy writers
1510                 */
1511                dirty_thresh += dirty_thresh / 10;      /* wheeee... */
1512
1513                if (global_page_state(NR_UNSTABLE_NFS) +
1514                        global_page_state(NR_WRITEBACK) <= dirty_thresh)
1515                                break;
1516                congestion_wait(BLK_RW_ASYNC, HZ/10);
1517
1518                /*
1519                 * The caller might hold locks which can prevent IO completion
1520                 * or progress in the filesystem.  So we cannot just sit here
1521                 * waiting for IO to complete.
1522                 */
1523                if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1524                        break;
1525        }
1526}
1527
1528/*
1529 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1530 */
1531int dirty_writeback_centisecs_handler(ctl_table *table, int write,
1532        void __user *buffer, size_t *length, loff_t *ppos)
1533{
1534        proc_dointvec(table, write, buffer, length, ppos);
1535        return 0;
1536}
1537
1538#ifdef CONFIG_BLOCK
1539void laptop_mode_timer_fn(unsigned long data)
1540{
1541        struct request_queue *q = (struct request_queue *)data;
1542        int nr_pages = global_page_state(NR_FILE_DIRTY) +
1543                global_page_state(NR_UNSTABLE_NFS);
1544
1545        /*
1546         * We want to write everything out, not just down to the dirty
1547         * threshold
1548         */
1549        if (bdi_has_dirty_io(&q->backing_dev_info))
1550                bdi_start_writeback(&q->backing_dev_info, nr_pages,
1551                                        WB_REASON_LAPTOP_TIMER);
1552}
1553
1554/*
1555 * We've spun up the disk and we're in laptop mode: schedule writeback
1556 * of all dirty data a few seconds from now.  If the flush is already scheduled
1557 * then push it back - the user is still using the disk.
1558 */
1559void laptop_io_completion(struct backing_dev_info *info)
1560{
1561        mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1562}
1563
1564/*
1565 * We're in laptop mode and we've just synced. The sync's writes will have
1566 * caused another writeback to be scheduled by laptop_io_completion.
1567 * Nothing needs to be written back anymore, so we unschedule the writeback.
1568 */
1569void laptop_sync_completion(void)
1570{
1571        struct backing_dev_info *bdi;
1572
1573        rcu_read_lock();
1574
1575        list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1576                del_timer(&bdi->laptop_mode_wb_timer);
1577
1578        rcu_read_unlock();
1579}
1580#endif
1581
1582/*
1583 * If ratelimit_pages is too high then we can get into dirty-data overload
1584 * if a large number of processes all perform writes at the same time.
1585 * If it is too low then SMP machines will call the (expensive)
1586 * get_writeback_state too often.
1587 *
1588 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1589 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1590 * thresholds.
1591 */
1592
1593void writeback_set_ratelimit(void)
1594{
1595        unsigned long background_thresh;
1596        unsigned long dirty_thresh;
1597        global_dirty_limits(&background_thresh, &dirty_thresh);
1598        global_dirty_limit = dirty_thresh;
1599        ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1600        if (ratelimit_pages < 16)
1601                ratelimit_pages = 16;
1602}
1603
1604static int __cpuinit
1605ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
1606{
1607        writeback_set_ratelimit();
1608        return NOTIFY_DONE;
1609}
1610
1611static struct notifier_block __cpuinitdata ratelimit_nb = {
1612        .notifier_call  = ratelimit_handler,
1613        .next           = NULL,
1614};
1615
1616/*
1617 * Called early on to tune the page writeback dirty limits.
1618 *
1619 * We used to scale dirty pages according to how total memory
1620 * related to pages that could be allocated for buffers (by
1621 * comparing nr_free_buffer_pages() to vm_total_pages.
1622 *
1623 * However, that was when we used "dirty_ratio" to scale with
1624 * all memory, and we don't do that any more. "dirty_ratio"
1625 * is now applied to total non-HIGHPAGE memory (by subtracting
1626 * totalhigh_pages from vm_total_pages), and as such we can't
1627 * get into the old insane situation any more where we had
1628 * large amounts of dirty pages compared to a small amount of
1629 * non-HIGHMEM memory.
1630 *
1631 * But we might still want to scale the dirty_ratio by how
1632 * much memory the box has..
1633 */
1634void __init page_writeback_init(void)
1635{
1636        writeback_set_ratelimit();
1637        register_cpu_notifier(&ratelimit_nb);
1638
1639        fprop_global_init(&writeout_completions);
1640}
1641
1642/**
1643 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1644 * @mapping: address space structure to write
1645 * @start: starting page index
1646 * @end: ending page index (inclusive)
1647 *
1648 * This function scans the page range from @start to @end (inclusive) and tags
1649 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1650 * that write_cache_pages (or whoever calls this function) will then use
1651 * TOWRITE tag to identify pages eligible for writeback.  This mechanism is
1652 * used to avoid livelocking of writeback by a process steadily creating new
1653 * dirty pages in the file (thus it is important for this function to be quick
1654 * so that it can tag pages faster than a dirtying process can create them).
1655 */
1656/*
1657 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1658 */
1659void tag_pages_for_writeback(struct address_space *mapping,
1660                             pgoff_t start, pgoff_t end)
1661{
1662#define WRITEBACK_TAG_BATCH 4096
1663        unsigned long tagged;
1664
1665        do {
1666                spin_lock_irq(&mapping->tree_lock);
1667                tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1668                                &start, end, WRITEBACK_TAG_BATCH,
1669                                PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1670                spin_unlock_irq(&mapping->tree_lock);
1671                WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1672                cond_resched();
1673                /* We check 'start' to handle wrapping when end == ~0UL */
1674        } while (tagged >= WRITEBACK_TAG_BATCH && start);
1675}
1676EXPORT_SYMBOL(tag_pages_for_writeback);
1677
1678/**
1679 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1680 * @mapping: address space structure to write
1681 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1682 * @writepage: function called for each page
1683 * @data: data passed to writepage function
1684 *
1685 * If a page is already under I/O, write_cache_pages() skips it, even
1686 * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
1687 * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
1688 * and msync() need to guarantee that all the data which was dirty at the time
1689 * the call was made get new I/O started against them.  If wbc->sync_mode is
1690 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1691 * existing IO to complete.
1692 *
1693 * To avoid livelocks (when other process dirties new pages), we first tag
1694 * pages which should be written back with TOWRITE tag and only then start
1695 * writing them. For data-integrity sync we have to be careful so that we do
1696 * not miss some pages (e.g., because some other process has cleared TOWRITE
1697 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1698 * by the process clearing the DIRTY tag (and submitting the page for IO).
1699 */
1700int write_cache_pages(struct address_space *mapping,
1701                      struct writeback_control *wbc, writepage_t writepage,
1702                      void *data)
1703{
1704        int ret = 0;
1705        int done = 0;
1706        struct pagevec pvec;
1707        int nr_pages;
1708        pgoff_t uninitialized_var(writeback_index);
1709        pgoff_t index;
1710        pgoff_t end;            /* Inclusive */
1711        pgoff_t done_index;
1712        int cycled;
1713        int range_whole = 0;
1714        int tag;
1715
1716        pagevec_init(&pvec, 0);
1717        if (wbc->range_cyclic) {
1718                writeback_index = mapping->writeback_index; /* prev offset */
1719                index = writeback_index;
1720                if (index == 0)
1721                        cycled = 1;
1722                else
1723                        cycled = 0;
1724                end = -1;
1725        } else {
1726                index = wbc->range_start >> PAGE_CACHE_SHIFT;
1727                end = wbc->range_end >> PAGE_CACHE_SHIFT;
1728                if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1729                        range_whole = 1;
1730                cycled = 1; /* ignore range_cyclic tests */
1731        }
1732        if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1733                tag = PAGECACHE_TAG_TOWRITE;
1734        else
1735                tag = PAGECACHE_TAG_DIRTY;
1736retry:
1737        if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1738                tag_pages_for_writeback(mapping, index, end);
1739        done_index = index;
1740        while (!done && (index <= end)) {
1741                int i;
1742
1743                nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1744                              min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1745                if (nr_pages == 0)
1746                        break;
1747
1748                for (i = 0; i < nr_pages; i++) {
1749                        struct page *page = pvec.pages[i];
1750
1751                        /*
1752                         * At this point, the page may be truncated or
1753                         * invalidated (changing page->mapping to NULL), or
1754                         * even swizzled back from swapper_space to tmpfs file
1755                         * mapping. However, page->index will not change
1756                         * because we have a reference on the page.
1757                         */
1758                        if (page->index > end) {
1759                                /*
1760                                 * can't be range_cyclic (1st pass) because
1761                                 * end == -1 in that case.
1762                                 */
1763                                done = 1;
1764                                break;
1765                        }
1766
1767                        done_index = page->index;
1768
1769                        lock_page(page);
1770
1771                        /*
1772                         * Page truncated or invalidated. We can freely skip it
1773                         * then, even for data integrity operations: the page
1774                         * has disappeared concurrently, so there could be no
1775                         * real expectation of this data interity operation
1776                         * even if there is now a new, dirty page at the same
1777                         * pagecache address.
1778                         */
1779                        if (unlikely(page->mapping != mapping)) {
1780continue_unlock:
1781                                unlock_page(page);
1782                                continue;
1783                        }
1784
1785                        if (!PageDirty(page)) {
1786                                /* someone wrote it for us */
1787                                goto continue_unlock;
1788                        }
1789
1790                        if (PageWriteback(page)) {
1791                                if (wbc->sync_mode != WB_SYNC_NONE)
1792                                        wait_on_page_writeback(page);
1793                                else
1794                                        goto continue_unlock;
1795                        }
1796
1797                        BUG_ON(PageWriteback(page));
1798                        if (!clear_page_dirty_for_io(page))
1799                                goto continue_unlock;
1800
1801                        trace_wbc_writepage(wbc, mapping->backing_dev_info);
1802                        ret = (*writepage)(page, wbc, data);
1803                        if (unlikely(ret)) {
1804                                if (ret == AOP_WRITEPAGE_ACTIVATE) {
1805                                        unlock_page(page);
1806                                        ret = 0;
1807                                } else {
1808                                        /*
1809                                         * done_index is set past this page,
1810                                         * so media errors will not choke
1811                                         * background writeout for the entire
1812                                         * file. This has consequences for
1813                                         * range_cyclic semantics (ie. it may
1814                                         * not be suitable for data integrity
1815                                         * writeout).
1816                                         */
1817                                        done_index = page->index + 1;
1818                                        done = 1;
1819                                        break;
1820                                }
1821                        }
1822
1823                        /*
1824                         * We stop writing back only if we are not doing
1825                         * integrity sync. In case of integrity sync we have to
1826                         * keep going until we have written all the pages
1827                         * we tagged for writeback prior to entering this loop.
1828                         */
1829                        if (--wbc->nr_to_write <= 0 &&
1830                            wbc->sync_mode == WB_SYNC_NONE) {
1831                                done = 1;
1832                                break;
1833                        }
1834                }
1835                pagevec_release(&pvec);
1836                cond_resched();
1837        }
1838        if (!cycled && !done) {
1839                /*
1840                 * range_cyclic:
1841                 * We hit the last page and there is more work to be done: wrap
1842                 * back to the start of the file
1843                 */
1844                cycled = 1;
1845                index = 0;
1846                end = writeback_index - 1;
1847                goto retry;
1848        }
1849        if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
1850                mapping->writeback_index = done_index;
1851
1852        return ret;
1853}
1854EXPORT_SYMBOL(write_cache_pages);
1855
1856/*
1857 * Function used by generic_writepages to call the real writepage
1858 * function and set the mapping flags on error
1859 */
1860static int __writepage(struct page *page, struct writeback_control *wbc,
1861                       void *data)
1862{
1863        struct address_space *mapping = data;
1864        int ret = mapping->a_ops->writepage(page, wbc);
1865        mapping_set_error(mapping, ret);
1866        return ret;
1867}
1868
1869/**
1870 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
1871 * @mapping: address space structure to write
1872 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1873 *
1874 * This is a library function, which implements the writepages()
1875 * address_space_operation.
1876 */
1877int generic_writepages(struct address_space *mapping,
1878                       struct writeback_control *wbc)
1879{
1880        struct blk_plug plug;
1881        int ret;
1882
1883        /* deal with chardevs and other special file */
1884        if (!mapping->a_ops->writepage)
1885                return 0;
1886
1887        blk_start_plug(&plug);
1888        ret = write_cache_pages(mapping, wbc, __writepage, mapping);
1889        blk_finish_plug(&plug);
1890        return ret;
1891}
1892
1893EXPORT_SYMBOL(generic_writepages);
1894
1895int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
1896{
1897        int ret;
1898
1899        if (wbc->nr_to_write <= 0)
1900                return 0;
1901        if (mapping->a_ops->writepages)
1902                ret = mapping->a_ops->writepages(mapping, wbc);
1903        else
1904                ret = generic_writepages(mapping, wbc);
1905        return ret;
1906}
1907
1908/**
1909 * write_one_page - write out a single page and optionally wait on I/O
1910 * @page: the page to write
1911 * @wait: if true, wait on writeout
1912 *
1913 * The page must be locked by the caller and will be unlocked upon return.
1914 *
1915 * write_one_page() returns a negative error code if I/O failed.
1916 */
1917int write_one_page(struct page *page, int wait)
1918{
1919        struct address_space *mapping = page->mapping;
1920        int ret = 0;
1921        struct writeback_control wbc = {
1922                .sync_mode = WB_SYNC_ALL,
1923                .nr_to_write = 1,
1924        };
1925
1926        BUG_ON(!PageLocked(page));
1927
1928        if (wait)
1929                wait_on_page_writeback(page);
1930
1931        if (clear_page_dirty_for_io(page)) {
1932                page_cache_get(page);
1933                ret = mapping->a_ops->writepage(page, &wbc);
1934                if (ret == 0 && wait) {
1935                        wait_on_page_writeback(page);
1936                        if (PageError(page))
1937                                ret = -EIO;
1938                }
1939                page_cache_release(page);
1940        } else {
1941                unlock_page(page);
1942        }
1943        return ret;
1944}
1945EXPORT_SYMBOL(write_one_page);
1946
1947/*
1948 * For address_spaces which do not use buffers nor write back.
1949 */
1950int __set_page_dirty_no_writeback(struct page *page)
1951{
1952        if (!PageDirty(page))
1953                return !TestSetPageDirty(page);
1954        return 0;
1955}
1956
1957/*
1958 * Helper function for set_page_dirty family.
1959 * NOTE: This relies on being atomic wrt interrupts.
1960 */
1961void account_page_dirtied(struct page *page, struct address_space *mapping)
1962{
1963        if (mapping_cap_account_dirty(mapping)) {
1964                __inc_zone_page_state(page, NR_FILE_DIRTY);
1965                __inc_zone_page_state(page, NR_DIRTIED);
1966                __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
1967                __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
1968                task_io_account_write(PAGE_CACHE_SIZE);
1969                current->nr_dirtied++;
1970                this_cpu_inc(bdp_ratelimits);
1971        }
1972}
1973EXPORT_SYMBOL(account_page_dirtied);
1974
1975/*
1976 * Helper function for set_page_writeback family.
1977 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
1978 * wrt interrupts.
1979 */
1980void account_page_writeback(struct page *page)
1981{
1982        inc_zone_page_state(page, NR_WRITEBACK);
1983}
1984EXPORT_SYMBOL(account_page_writeback);
1985
1986/*
1987 * For address_spaces which do not use buffers.  Just tag the page as dirty in
1988 * its radix tree.
1989 *
1990 * This is also used when a single buffer is being dirtied: we want to set the
1991 * page dirty in that case, but not all the buffers.  This is a "bottom-up"
1992 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
1993 *
1994 * Most callers have locked the page, which pins the address_space in memory.
1995 * But zap_pte_range() does not lock the page, however in that case the
1996 * mapping is pinned by the vma's ->vm_file reference.
1997 *
1998 * We take care to handle the case where the page was truncated from the
1999 * mapping by re-checking page_mapping() inside tree_lock.
2000 */
2001int __set_page_dirty_nobuffers(struct page *page)
2002{
2003        if (!TestSetPageDirty(page)) {
2004                struct address_space *mapping = page_mapping(page);
2005                struct address_space *mapping2;
2006
2007                if (!mapping)
2008                        return 1;
2009
2010                spin_lock_irq(&mapping->tree_lock);
2011                mapping2 = page_mapping(page);
2012                if (mapping2) { /* Race with truncate? */
2013                        BUG_ON(mapping2 != mapping);
2014                        WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2015                        account_page_dirtied(page, mapping);
2016                        radix_tree_tag_set(&mapping->page_tree,
2017                                page_index(page), PAGECACHE_TAG_DIRTY);
2018                }
2019                spin_unlock_irq(&mapping->tree_lock);
2020                if (mapping->host) {
2021                        /* !PageAnon && !swapper_space */
2022                        __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2023                }
2024                return 1;
2025        }
2026        return 0;
2027}
2028EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2029
2030/*
2031 * Call this whenever redirtying a page, to de-account the dirty counters
2032 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2033 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2034 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2035 * control.
2036 */
2037void account_page_redirty(struct page *page)
2038{
2039        struct address_space *mapping = page->mapping;
2040        if (mapping && mapping_cap_account_dirty(mapping)) {
2041                current->nr_dirtied--;
2042                dec_zone_page_state(page, NR_DIRTIED);
2043                dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
2044        }
2045}
2046EXPORT_SYMBOL(account_page_redirty);
2047
2048/*
2049 * When a writepage implementation decides that it doesn't want to write this
2050 * page for some reason, it should redirty the locked page via
2051 * redirty_page_for_writepage() and it should then unlock the page and return 0
2052 */
2053int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2054{
2055        wbc->pages_skipped++;
2056        account_page_redirty(page);
2057        return __set_page_dirty_nobuffers(page);
2058}
2059EXPORT_SYMBOL(redirty_page_for_writepage);
2060
2061/*
2062 * Dirty a page.
2063 *
2064 * For pages with a mapping this should be done under the page lock
2065 * for the benefit of asynchronous memory errors who prefer a consistent
2066 * dirty state. This rule can be broken in some special cases,
2067 * but should be better not to.
2068 *
2069 * If the mapping doesn't provide a set_page_dirty a_op, then
2070 * just fall through and assume that it wants buffer_heads.
2071 */
2072int set_page_dirty(struct page *page)
2073{
2074        struct address_space *mapping = page_mapping(page);
2075
2076        if (likely(mapping)) {
2077                int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2078                /*
2079                 * readahead/lru_deactivate_page could remain
2080                 * PG_readahead/PG_reclaim due to race with end_page_writeback
2081                 * About readahead, if the page is written, the flags would be
2082                 * reset. So no problem.
2083                 * About lru_deactivate_page, if the page is redirty, the flag
2084                 * will be reset. So no problem. but if the page is used by readahead
2085                 * it will confuse readahead and make it restart the size rampup
2086                 * process. But it's a trivial problem.
2087                 */
2088                ClearPageReclaim(page);
2089#ifdef CONFIG_BLOCK
2090                if (!spd)
2091                        spd = __set_page_dirty_buffers;
2092#endif
2093                return (*spd)(page);
2094        }
2095        if (!PageDirty(page)) {
2096                if (!TestSetPageDirty(page))
2097                        return 1;
2098        }
2099        return 0;
2100}
2101EXPORT_SYMBOL(set_page_dirty);
2102
2103/*
2104 * set_page_dirty() is racy if the caller has no reference against
2105 * page->mapping->host, and if the page is unlocked.  This is because another
2106 * CPU could truncate the page off the mapping and then free the mapping.
2107 *
2108 * Usually, the page _is_ locked, or the caller is a user-space process which
2109 * holds a reference on the inode by having an open file.
2110 *
2111 * In other cases, the page should be locked before running set_page_dirty().
2112 */
2113int set_page_dirty_lock(struct page *page)
2114{
2115        int ret;
2116
2117        lock_page(page);
2118        ret = set_page_dirty(page);
2119        unlock_page(page);
2120        return ret;
2121}
2122EXPORT_SYMBOL(set_page_dirty_lock);
2123
2124/*
2125 * Clear a page's dirty flag, while caring for dirty memory accounting.
2126 * Returns true if the page was previously dirty.
2127 *
2128 * This is for preparing to put the page under writeout.  We leave the page
2129 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2130 * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
2131 * implementation will run either set_page_writeback() or set_page_dirty(),
2132 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2133 * back into sync.
2134 *
2135 * This incoherency between the page's dirty flag and radix-tree tag is
2136 * unfortunate, but it only exists while the page is locked.
2137 */
2138int clear_page_dirty_for_io(struct page *page)
2139{
2140        struct address_space *mapping = page_mapping(page);
2141
2142        BUG_ON(!PageLocked(page));
2143
2144        if (mapping && mapping_cap_account_dirty(mapping)) {
2145                /*
2146                 * Yes, Virginia, this is indeed insane.
2147                 *
2148                 * We use this sequence to make sure that
2149                 *  (a) we account for dirty stats properly
2150                 *  (b) we tell the low-level filesystem to
2151                 *      mark the whole page dirty if it was
2152                 *      dirty in a pagetable. Only to then
2153                 *  (c) clean the page again and return 1 to
2154                 *      cause the writeback.
2155                 *
2156                 * This way we avoid all nasty races with the
2157                 * dirty bit in multiple places and clearing
2158                 * them concurrently from different threads.
2159                 *
2160                 * Note! Normally the "set_page_dirty(page)"
2161                 * has no effect on the actual dirty bit - since
2162                 * that will already usually be set. But we
2163                 * need the side effects, and it can help us
2164                 * avoid races.
2165                 *
2166                 * We basically use the page "master dirty bit"
2167                 * as a serialization point for all the different
2168                 * threads doing their things.
2169                 */
2170                if (page_mkclean(page))
2171                        set_page_dirty(page);
2172                /*
2173                 * We carefully synchronise fault handlers against
2174                 * installing a dirty pte and marking the page dirty
2175                 * at this point. We do this by having them hold the
2176                 * page lock at some point after installing their
2177                 * pte, but before marking the page dirty.
2178                 * Pages are always locked coming in here, so we get
2179                 * the desired exclusion. See mm/memory.c:do_wp_page()
2180                 * for more comments.
2181                 */
2182                if (TestClearPageDirty(page)) {
2183                        dec_zone_page_state(page, NR_FILE_DIRTY);
2184                        dec_bdi_stat(mapping->backing_dev_info,
2185                                        BDI_RECLAIMABLE);
2186                        return 1;
2187                }
2188                return 0;
2189        }
2190        return TestClearPageDirty(page);
2191}
2192EXPORT_SYMBOL(clear_page_dirty_for_io);
2193
2194int test_clear_page_writeback(struct page *page)
2195{
2196        struct address_space *mapping = page_mapping(page);
2197        int ret;
2198
2199        if (mapping) {
2200                struct backing_dev_info *bdi = mapping->backing_dev_info;
2201                unsigned long flags;
2202
2203                spin_lock_irqsave(&mapping->tree_lock, flags);
2204                ret = TestClearPageWriteback(page);
2205                if (ret) {
2206                        radix_tree_tag_clear(&mapping->page_tree,
2207                                                page_index(page),
2208                                                PAGECACHE_TAG_WRITEBACK);
2209                        if (bdi_cap_account_writeback(bdi)) {
2210                                __dec_bdi_stat(bdi, BDI_WRITEBACK);
2211                                __bdi_writeout_inc(bdi);
2212                        }
2213                }
2214                spin_unlock_irqrestore(&mapping->tree_lock, flags);
2215        } else {
2216                ret = TestClearPageWriteback(page);
2217        }
2218        if (ret) {
2219                dec_zone_page_state(page, NR_WRITEBACK);
2220                inc_zone_page_state(page, NR_WRITTEN);
2221        }
2222        return ret;
2223}
2224
2225int test_set_page_writeback(struct page *page)
2226{
2227        struct address_space *mapping = page_mapping(page);
2228        int ret;
2229
2230        if (mapping) {
2231                struct backing_dev_info *bdi = mapping->backing_dev_info;
2232                unsigned long flags;
2233
2234                spin_lock_irqsave(&mapping->tree_lock, flags);
2235                ret = TestSetPageWriteback(page);
2236                if (!ret) {
2237                        radix_tree_tag_set(&mapping->page_tree,
2238                                                page_index(page),
2239                                                PAGECACHE_TAG_WRITEBACK);
2240                        if (bdi_cap_account_writeback(bdi))
2241                                __inc_bdi_stat(bdi, BDI_WRITEBACK);
2242                }
2243                if (!PageDirty(page))
2244                        radix_tree_tag_clear(&mapping->page_tree,
2245                                                page_index(page),
2246                                                PAGECACHE_TAG_DIRTY);
2247                radix_tree_tag_clear(&mapping->page_tree,
2248                                     page_index(page),
2249                                     PAGECACHE_TAG_TOWRITE);
2250                spin_unlock_irqrestore(&mapping->tree_lock, flags);
2251        } else {
2252                ret = TestSetPageWriteback(page);
2253        }
2254        if (!ret)
2255                account_page_writeback(page);
2256        return ret;
2257
2258}
2259EXPORT_SYMBOL(test_set_page_writeback);
2260
2261/*
2262 * Return true if any of the pages in the mapping are marked with the
2263 * passed tag.
2264 */
2265int mapping_tagged(struct address_space *mapping, int tag)
2266{
2267        return radix_tree_tagged(&mapping->page_tree, tag);
2268}
2269EXPORT_SYMBOL(mapping_tagged);
2270
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.