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 <linux/sched/rt.h>
  39#include <linux/mm_inline.h>
  40#include <trace/events/writeback.h>
  41
  42#include "internal.h"
  43
  44/*
  45 * Sleep at most 200ms at a time in balance_dirty_pages().
  46 */
  47#define MAX_PAUSE               max(HZ/5, 1)
  48
  49/*
  50 * Try to keep balance_dirty_pages() call intervals higher than this many pages
  51 * by raising pause time to max_pause when falls below it.
  52 */
  53#define DIRTY_POLL_THRESH       (128 >> (PAGE_SHIFT - 10))
  54
  55/*
  56 * Estimate write bandwidth at 200ms intervals.
  57 */
  58#define BANDWIDTH_INTERVAL      max(HZ/5, 1)
  59
  60#define RATELIMIT_CALC_SHIFT    10
  61
  62/*
  63 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
  64 * will look to see if it needs to force writeback or throttling.
  65 */
  66static long ratelimit_pages = 32;
  67
  68/* The following parameters are exported via /proc/sys/vm */
  69
  70/*
  71 * Start background writeback (via writeback threads) at this percentage
  72 */
  73int dirty_background_ratio = 10;
  74
  75/*
  76 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
  77 * dirty_background_ratio * the amount of dirtyable memory
  78 */
  79unsigned long dirty_background_bytes;
  80
  81/*
  82 * free highmem will not be subtracted from the total free memory
  83 * for calculating free ratios if vm_highmem_is_dirtyable is true
  84 */
  85int vm_highmem_is_dirtyable;
  86
  87/*
  88 * The generator of dirty data starts writeback at this percentage
  89 */
  90int vm_dirty_ratio = 20;
  91
  92/*
  93 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
  94 * vm_dirty_ratio * the amount of dirtyable memory
  95 */
  96unsigned long vm_dirty_bytes;
  97
  98/*
  99 * The interval between `kupdate'-style writebacks
 100 */
 101unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
 102
 103EXPORT_SYMBOL_GPL(dirty_writeback_interval);
 104
 105/*
 106 * The longest time for which data is allowed to remain dirty
 107 */
 108unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
 109
 110/*
 111 * Flag that makes the machine dump writes/reads and block dirtyings.
 112 */
 113int block_dump;
 114
 115/*
 116 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
 117 * a full sync is triggered after this time elapses without any disk activity.
 118 */
 119int laptop_mode;
 120
 121EXPORT_SYMBOL(laptop_mode);
 122
 123/* End of sysctl-exported parameters */
 124
 125unsigned long global_dirty_limit;
 126
 127/*
 128 * Scale the writeback cache size proportional to the relative writeout speeds.
 129 *
 130 * We do this by keeping a floating proportion between BDIs, based on page
 131 * writeback completions [end_page_writeback()]. Those devices that write out
 132 * pages fastest will get the larger share, while the slower will get a smaller
 133 * share.
 134 *
 135 * We use page writeout completions because we are interested in getting rid of
 136 * dirty pages. Having them written out is the primary goal.
 137 *
 138 * We introduce a concept of time, a period over which we measure these events,
 139 * because demand can/will vary over time. The length of this period itself is
 140 * measured in page writeback completions.
 141 *
 142 */
 143static struct fprop_global writeout_completions;
 144
 145static void writeout_period(unsigned long t);
 146/* Timer for aging of writeout_completions */
 147static struct timer_list writeout_period_timer =
 148                TIMER_DEFERRED_INITIALIZER(writeout_period, 0, 0);
 149static unsigned long writeout_period_time = 0;
 150
 151/*
 152 * Length of period for aging writeout fractions of bdis. This is an
 153 * arbitrarily chosen number. The longer the period, the slower fractions will
 154 * reflect changes in current writeout rate.
 155 */
 156#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
 157
 158/*
 159 * Work out the current dirty-memory clamping and background writeout
 160 * thresholds.
 161 *
 162 * The main aim here is to lower them aggressively if there is a lot of mapped
 163 * memory around.  To avoid stressing page reclaim with lots of unreclaimable
 164 * pages.  It is better to clamp down on writers than to start swapping, and
 165 * performing lots of scanning.
 166 *
 167 * We only allow 1/2 of the currently-unmapped memory to be dirtied.
 168 *
 169 * We don't permit the clamping level to fall below 5% - that is getting rather
 170 * excessive.
 171 *
 172 * We make sure that the background writeout level is below the adjusted
 173 * clamping level.
 174 */
 175
 176/*
 177 * In a memory zone, there is a certain amount of pages we consider
 178 * available for the page cache, which is essentially the number of
 179 * free and reclaimable pages, minus some zone reserves to protect
 180 * lowmem and the ability to uphold the zone's watermarks without
 181 * requiring writeback.
 182 *
 183 * This number of dirtyable pages is the base value of which the
 184 * user-configurable dirty ratio is the effictive number of pages that
 185 * are allowed to be actually dirtied.  Per individual zone, or
 186 * globally by using the sum of dirtyable pages over all zones.
 187 *
 188 * Because the user is allowed to specify the dirty limit globally as
 189 * absolute number of bytes, calculating the per-zone dirty limit can
 190 * require translating the configured limit into a percentage of
 191 * global dirtyable memory first.
 192 */
 193
 194static unsigned long highmem_dirtyable_memory(unsigned long total)
 195{
 196#ifdef CONFIG_HIGHMEM
 197        int node;
 198        unsigned long x = 0;
 199
 200        for_each_node_state(node, N_HIGH_MEMORY) {
 201                struct zone *z =
 202                        &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
 203
 204                x += zone_page_state(z, NR_FREE_PAGES) +
 205                     zone_reclaimable_pages(z) - z->dirty_balance_reserve;
 206        }
 207        /*
 208         * Unreclaimable memory (kernel memory or anonymous memory
 209         * without swap) can bring down the dirtyable pages below
 210         * the zone's dirty balance reserve and the above calculation
 211         * will underflow.  However we still want to add in nodes
 212         * which are below threshold (negative values) to get a more
 213         * accurate calculation but make sure that the total never
 214         * underflows.
 215         */
 216        if ((long)x < 0)
 217                x = 0;
 218
 219        /*
 220         * Make sure that the number of highmem pages is never larger
 221         * than the number of the total dirtyable memory. This can only
 222         * occur in very strange VM situations but we want to make sure
 223         * that this does not occur.
 224         */
 225        return min(x, total);
 226#else
 227        return 0;
 228#endif
 229}
 230
 231/**
 232 * global_dirtyable_memory - number of globally dirtyable pages
 233 *
 234 * Returns the global number of pages potentially available for dirty
 235 * page cache.  This is the base value for the global dirty limits.
 236 */
 237static unsigned long global_dirtyable_memory(void)
 238{
 239        unsigned long x;
 240
 241        x = global_page_state(NR_FREE_PAGES) + global_reclaimable_pages();
 242        x -= min(x, dirty_balance_reserve);
 243
 244        if (!vm_highmem_is_dirtyable)
 245                x -= highmem_dirtyable_memory(x);
 246
 247        return x + 1;   /* Ensure that we never return 0 */
 248}
 249
 250/*
 251 * global_dirty_limits - background-writeback and dirty-throttling thresholds
 252 *
 253 * Calculate the dirty thresholds based on sysctl parameters
 254 * - vm.dirty_background_ratio  or  vm.dirty_background_bytes
 255 * - vm.dirty_ratio             or  vm.dirty_bytes
 256 * The dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
 257 * real-time tasks.
 258 */
 259void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
 260{
 261        unsigned long background;
 262        unsigned long dirty;
 263        unsigned long uninitialized_var(available_memory);
 264        struct task_struct *tsk;
 265
 266        if (!vm_dirty_bytes || !dirty_background_bytes)
 267                available_memory = global_dirtyable_memory();
 268
 269        if (vm_dirty_bytes)
 270                dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE);
 271        else
 272                dirty = (vm_dirty_ratio * available_memory) / 100;
 273
 274        if (dirty_background_bytes)
 275                background = DIV_ROUND_UP(dirty_background_bytes, PAGE_SIZE);
 276        else
 277                background = (dirty_background_ratio * available_memory) / 100;
 278
 279        if (background >= dirty)
 280                background = dirty / 2;
 281        tsk = current;
 282        if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
 283                background += background / 4;
 284                dirty += dirty / 4;
 285        }
 286        *pbackground = background;
 287        *pdirty = dirty;
 288        trace_global_dirty_state(background, dirty);
 289}
 290
 291/**
 292 * zone_dirtyable_memory - number of dirtyable pages in a zone
 293 * @zone: the zone
 294 *
 295 * Returns the zone's number of pages potentially available for dirty
 296 * page cache.  This is the base value for the per-zone dirty limits.
 297 */
 298static unsigned long zone_dirtyable_memory(struct zone *zone)
 299{
 300        /*
 301         * The effective global number of dirtyable pages may exclude
 302         * highmem as a big-picture measure to keep the ratio between
 303         * dirty memory and lowmem reasonable.
 304         *
 305         * But this function is purely about the individual zone and a
 306         * highmem zone can hold its share of dirty pages, so we don't
 307         * care about vm_highmem_is_dirtyable here.
 308         */
 309        unsigned long nr_pages = zone_page_state(zone, NR_FREE_PAGES) +
 310                zone_reclaimable_pages(zone);
 311
 312        /* don't allow this to underflow */
 313        nr_pages -= min(nr_pages, zone->dirty_balance_reserve);
 314        return nr_pages;
 315}
 316
 317/**
 318 * zone_dirty_limit - maximum number of dirty pages allowed in a zone
 319 * @zone: the zone
 320 *
 321 * Returns the maximum number of dirty pages allowed in a zone, based
 322 * on the zone's dirtyable memory.
 323 */
 324static unsigned long zone_dirty_limit(struct zone *zone)
 325{
 326        unsigned long zone_memory = zone_dirtyable_memory(zone);
 327        struct task_struct *tsk = current;
 328        unsigned long dirty;
 329
 330        if (vm_dirty_bytes)
 331                dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
 332                        zone_memory / global_dirtyable_memory();
 333        else
 334                dirty = vm_dirty_ratio * zone_memory / 100;
 335
 336        if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
 337                dirty += dirty / 4;
 338
 339        return dirty;
 340}
 341
 342/**
 343 * zone_dirty_ok - tells whether a zone is within its dirty limits
 344 * @zone: the zone to check
 345 *
 346 * Returns %true when the dirty pages in @zone are within the zone's
 347 * dirty limit, %false if the limit is exceeded.
 348 */
 349bool zone_dirty_ok(struct zone *zone)
 350{
 351        unsigned long limit = zone_dirty_limit(zone);
 352
 353        return zone_page_state(zone, NR_FILE_DIRTY) +
 354               zone_page_state(zone, NR_UNSTABLE_NFS) +
 355               zone_page_state(zone, NR_WRITEBACK) <= limit;
 356}
 357
 358int dirty_background_ratio_handler(struct ctl_table *table, int write,
 359                void __user *buffer, size_t *lenp,
 360                loff_t *ppos)
 361{
 362        int ret;
 363
 364        ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 365        if (ret == 0 && write)
 366                dirty_background_bytes = 0;
 367        return ret;
 368}
 369
 370int dirty_background_bytes_handler(struct ctl_table *table, int write,
 371                void __user *buffer, size_t *lenp,
 372                loff_t *ppos)
 373{
 374        int ret;
 375
 376        ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
 377        if (ret == 0 && write)
 378                dirty_background_ratio = 0;
 379        return ret;
 380}
 381
 382int dirty_ratio_handler(struct ctl_table *table, int write,
 383                void __user *buffer, size_t *lenp,
 384                loff_t *ppos)
 385{
 386        int old_ratio = vm_dirty_ratio;
 387        int ret;
 388
 389        ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 390        if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
 391                writeback_set_ratelimit();
 392                vm_dirty_bytes = 0;
 393        }
 394        return ret;
 395}
 396
 397int dirty_bytes_handler(struct ctl_table *table, int write,
 398                void __user *buffer, size_t *lenp,
 399                loff_t *ppos)
 400{
 401        unsigned long old_bytes = vm_dirty_bytes;
 402        int ret;
 403
 404        ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
 405        if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
 406                writeback_set_ratelimit();
 407                vm_dirty_ratio = 0;
 408        }
 409        return ret;
 410}
 411
 412static unsigned long wp_next_time(unsigned long cur_time)
 413{
 414        cur_time += VM_COMPLETIONS_PERIOD_LEN;
 415        /* 0 has a special meaning... */
 416        if (!cur_time)
 417                return 1;
 418        return cur_time;
 419}
 420
 421/*
 422 * Increment the BDI's writeout completion count and the global writeout
 423 * completion count. Called from test_clear_page_writeback().
 424 */
 425static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
 426{
 427        __inc_bdi_stat(bdi, BDI_WRITTEN);
 428        __fprop_inc_percpu_max(&writeout_completions, &bdi->completions,
 429                               bdi->max_prop_frac);
 430        /* First event after period switching was turned off? */
 431        if (!unlikely(writeout_period_time)) {
 432                /*
 433                 * We can race with other __bdi_writeout_inc calls here but
 434                 * it does not cause any harm since the resulting time when
 435                 * timer will fire and what is in writeout_period_time will be
 436                 * roughly the same.
 437                 */
 438                writeout_period_time = wp_next_time(jiffies);
 439                mod_timer(&writeout_period_timer, writeout_period_time);
 440        }
 441}
 442
 443void bdi_writeout_inc(struct backing_dev_info *bdi)
 444{
 445        unsigned long flags;
 446
 447        local_irq_save(flags);
 448        __bdi_writeout_inc(bdi);
 449        local_irq_restore(flags);
 450}
 451EXPORT_SYMBOL_GPL(bdi_writeout_inc);
 452
 453/*
 454 * Obtain an accurate fraction of the BDI's portion.
 455 */
 456static void bdi_writeout_fraction(struct backing_dev_info *bdi,
 457                long *numerator, long *denominator)
 458{
 459        fprop_fraction_percpu(&writeout_completions, &bdi->completions,
 460                                numerator, denominator);
 461}
 462
 463/*
 464 * On idle system, we can be called long after we scheduled because we use
 465 * deferred timers so count with missed periods.
 466 */
 467static void writeout_period(unsigned long t)
 468{
 469        int miss_periods = (jiffies - writeout_period_time) /
 470                                                 VM_COMPLETIONS_PERIOD_LEN;
 471
 472        if (fprop_new_period(&writeout_completions, miss_periods + 1)) {
 473                writeout_period_time = wp_next_time(writeout_period_time +
 474                                miss_periods * VM_COMPLETIONS_PERIOD_LEN);
 475                mod_timer(&writeout_period_timer, writeout_period_time);
 476        } else {
 477                /*
 478                 * Aging has zeroed all fractions. Stop wasting CPU on period
 479                 * updates.
 480                 */
 481                writeout_period_time = 0;
 482        }
 483}
 484
 485/*
 486 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
 487 * registered backing devices, which, for obvious reasons, can not
 488 * exceed 100%.
 489 */
 490static unsigned int bdi_min_ratio;
 491
 492int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
 493{
 494        int ret = 0;
 495
 496        spin_lock_bh(&bdi_lock);
 497        if (min_ratio > bdi->max_ratio) {
 498                ret = -EINVAL;
 499        } else {
 500                min_ratio -= bdi->min_ratio;
 501                if (bdi_min_ratio + min_ratio < 100) {
 502                        bdi_min_ratio += min_ratio;
 503                        bdi->min_ratio += min_ratio;
 504                } else {
 505                        ret = -EINVAL;
 506                }
 507        }
 508        spin_unlock_bh(&bdi_lock);
 509
 510        return ret;
 511}
 512
 513int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
 514{
 515        int ret = 0;
 516
 517        if (max_ratio > 100)
 518                return -EINVAL;
 519
 520        spin_lock_bh(&bdi_lock);
 521        if (bdi->min_ratio > max_ratio) {
 522                ret = -EINVAL;
 523        } else {
 524                bdi->max_ratio = max_ratio;
 525                bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
 526        }
 527        spin_unlock_bh(&bdi_lock);
 528
 529        return ret;
 530}
 531EXPORT_SYMBOL(bdi_set_max_ratio);
 532
 533static unsigned long dirty_freerun_ceiling(unsigned long thresh,
 534                                           unsigned long bg_thresh)
 535{
 536        return (thresh + bg_thresh) / 2;
 537}
 538
 539static unsigned long hard_dirty_limit(unsigned long thresh)
 540{
 541        return max(thresh, global_dirty_limit);
 542}
 543
 544/**
 545 * bdi_dirty_limit - @bdi's share of dirty throttling threshold
 546 * @bdi: the backing_dev_info to query
 547 * @dirty: global dirty limit in pages
 548 *
 549 * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
 550 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
 551 *
 552 * Note that balance_dirty_pages() will only seriously take it as a hard limit
 553 * when sleeping max_pause per page is not enough to keep the dirty pages under
 554 * control. For example, when the device is completely stalled due to some error
 555 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
 556 * In the other normal situations, it acts more gently by throttling the tasks
 557 * more (rather than completely block them) when the bdi dirty pages go high.
 558 *
 559 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
 560 * - starving fast devices
 561 * - piling up dirty pages (that will take long time to sync) on slow devices
 562 *
 563 * The bdi's share of dirty limit will be adapting to its throughput and
 564 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
 565 */
 566unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
 567{
 568        u64 bdi_dirty;
 569        long numerator, denominator;
 570
 571        /*
 572         * Calculate this BDI's share of the dirty ratio.
 573         */
 574        bdi_writeout_fraction(bdi, &numerator, &denominator);
 575
 576        bdi_dirty = (dirty * (100 - bdi_min_ratio)) / 100;
 577        bdi_dirty *= numerator;
 578        do_div(bdi_dirty, denominator);
 579
 580        bdi_dirty += (dirty * bdi->min_ratio) / 100;
 581        if (bdi_dirty > (dirty * bdi->max_ratio) / 100)
 582                bdi_dirty = dirty * bdi->max_ratio / 100;
 583
 584        return bdi_dirty;
 585}
 586
 587/*
 588 *                           setpoint - dirty 3
 589 *        f(dirty) := 1.0 + (----------------)
 590 *                           limit - setpoint
 591 *
 592 * it's a 3rd order polynomial that subjects to
 593 *
 594 * (1) f(freerun)  = 2.0 => rampup dirty_ratelimit reasonably fast
 595 * (2) f(setpoint) = 1.0 => the balance point
 596 * (3) f(limit)    = 0   => the hard limit
 597 * (4) df/dx      <= 0   => negative feedback control
 598 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
 599 *     => fast response on large errors; small oscillation near setpoint
 600 */
 601static inline long long pos_ratio_polynom(unsigned long setpoint,
 602                                          unsigned long dirty,
 603                                          unsigned long limit)
 604{
 605        long long pos_ratio;
 606        long x;
 607
 608        x = div_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
 609                    limit - setpoint + 1);
 610        pos_ratio = x;
 611        pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
 612        pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
 613        pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
 614
 615        return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
 616}
 617
 618/*
 619 * Dirty position control.
 620 *
 621 * (o) global/bdi setpoints
 622 *
 623 * We want the dirty pages be balanced around the global/bdi setpoints.
 624 * When the number of dirty pages is higher/lower than the setpoint, the
 625 * dirty position control ratio (and hence task dirty ratelimit) will be
 626 * decreased/increased to bring the dirty pages back to the setpoint.
 627 *
 628 *     pos_ratio = 1 << RATELIMIT_CALC_SHIFT
 629 *
 630 *     if (dirty < setpoint) scale up   pos_ratio
 631 *     if (dirty > setpoint) scale down pos_ratio
 632 *
 633 *     if (bdi_dirty < bdi_setpoint) scale up   pos_ratio
 634 *     if (bdi_dirty > bdi_setpoint) scale down pos_ratio
 635 *
 636 *     task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
 637 *
 638 * (o) global control line
 639 *
 640 *     ^ pos_ratio
 641 *     |
 642 *     |            |<===== global dirty control scope ======>|
 643 * 2.0 .............*
 644 *     |            .*
 645 *     |            . *
 646 *     |            .   *
 647 *     |            .     *
 648 *     |            .        *
 649 *     |            .            *
 650 * 1.0 ................................*
 651 *     |            .                  .     *
 652 *     |            .                  .          *
 653 *     |            .                  .              *
 654 *     |            .                  .                 *
 655 *     |            .                  .                    *
 656 *   0 +------------.------------------.----------------------*------------->
 657 *           freerun^          setpoint^                 limit^   dirty pages
 658 *
 659 * (o) bdi control line
 660 *
 661 *     ^ pos_ratio
 662 *     |
 663 *     |            *
 664 *     |              *
 665 *     |                *
 666 *     |                  *
 667 *     |                    * |<=========== span ============>|
 668 * 1.0 .......................*
 669 *     |                      . *
 670 *     |                      .   *
 671 *     |                      .     *
 672 *     |                      .       *
 673 *     |                      .         *
 674 *     |                      .           *
 675 *     |                      .             *
 676 *     |                      .               *
 677 *     |                      .                 *
 678 *     |                      .                   *
 679 *     |                      .                     *
 680 * 1/4 ...............................................* * * * * * * * * * * *
 681 *     |                      .                         .
 682 *     |                      .                           .
 683 *     |                      .                             .
 684 *   0 +----------------------.-------------------------------.------------->
 685 *                bdi_setpoint^                    x_intercept^
 686 *
 687 * The bdi control line won't drop below pos_ratio=1/4, so that bdi_dirty can
 688 * be smoothly throttled down to normal if it starts high in situations like
 689 * - start writing to a slow SD card and a fast disk at the same time. The SD
 690 *   card's bdi_dirty may rush to many times higher than bdi_setpoint.
 691 * - the bdi dirty thresh drops quickly due to change of JBOD workload
 692 */
 693static unsigned long bdi_position_ratio(struct backing_dev_info *bdi,
 694                                        unsigned long thresh,
 695                                        unsigned long bg_thresh,
 696                                        unsigned long dirty,
 697                                        unsigned long bdi_thresh,
 698                                        unsigned long bdi_dirty)
 699{
 700        unsigned long write_bw = bdi->avg_write_bandwidth;
 701        unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
 702        unsigned long limit = hard_dirty_limit(thresh);
 703        unsigned long x_intercept;
 704        unsigned long setpoint;         /* dirty pages' target balance point */
 705        unsigned long bdi_setpoint;
 706        unsigned long span;
 707        long long pos_ratio;            /* for scaling up/down the rate limit */
 708        long x;
 709
 710        if (unlikely(dirty >= limit))
 711                return 0;
 712
 713        /*
 714         * global setpoint
 715         *
 716         * See comment for pos_ratio_polynom().
 717         */
 718        setpoint = (freerun + limit) / 2;
 719        pos_ratio = pos_ratio_polynom(setpoint, dirty, limit);
 720
 721        /*
 722         * The strictlimit feature is a tool preventing mistrusted filesystems
 723         * from growing a large number of dirty pages before throttling. For
 724         * such filesystems balance_dirty_pages always checks bdi counters
 725         * against bdi limits. Even if global "nr_dirty" is under "freerun".
 726         * This is especially important for fuse which sets bdi->max_ratio to
 727         * 1% by default. Without strictlimit feature, fuse writeback may
 728         * consume arbitrary amount of RAM because it is accounted in
 729         * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
 730         *
 731         * Here, in bdi_position_ratio(), we calculate pos_ratio based on
 732         * two values: bdi_dirty and bdi_thresh. Let's consider an example:
 733         * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
 734         * limits are set by default to 10% and 20% (background and throttle).
 735         * Then bdi_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
 736         * bdi_dirty_limit(bdi, bg_thresh) is about ~4K pages. bdi_setpoint is
 737         * about ~6K pages (as the average of background and throttle bdi
 738         * limits). The 3rd order polynomial will provide positive feedback if
 739         * bdi_dirty is under bdi_setpoint and vice versa.
 740         *
 741         * Note, that we cannot use global counters in these calculations
 742         * because we want to throttle process writing to a strictlimit BDI
 743         * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
 744         * in the example above).
 745         */
 746        if (unlikely(bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
 747                long long bdi_pos_ratio;
 748                unsigned long bdi_bg_thresh;
 749
 750                if (bdi_dirty < 8)
 751                        return min_t(long long, pos_ratio * 2,
 752                                     2 << RATELIMIT_CALC_SHIFT);
 753
 754                if (bdi_dirty >= bdi_thresh)
 755                        return 0;
 756
 757                bdi_bg_thresh = div_u64((u64)bdi_thresh * bg_thresh, thresh);
 758                bdi_setpoint = dirty_freerun_ceiling(bdi_thresh,
 759                                                     bdi_bg_thresh);
 760
 761                if (bdi_setpoint == 0 || bdi_setpoint == bdi_thresh)
 762                        return 0;
 763
 764                bdi_pos_ratio = pos_ratio_polynom(bdi_setpoint, bdi_dirty,
 765                                                  bdi_thresh);
 766
 767                /*
 768                 * Typically, for strictlimit case, bdi_setpoint << setpoint
 769                 * and pos_ratio >> bdi_pos_ratio. In the other words global
 770                 * state ("dirty") is not limiting factor and we have to
 771                 * make decision based on bdi counters. But there is an
 772                 * important case when global pos_ratio should get precedence:
 773                 * global limits are exceeded (e.g. due to activities on other
 774                 * BDIs) while given strictlimit BDI is below limit.
 775                 *
 776                 * "pos_ratio * bdi_pos_ratio" would work for the case above,
 777                 * but it would look too non-natural for the case of all
 778                 * activity in the system coming from a single strictlimit BDI
 779                 * with bdi->max_ratio == 100%.
 780                 *
 781                 * Note that min() below somewhat changes the dynamics of the
 782                 * control system. Normally, pos_ratio value can be well over 3
 783                 * (when globally we are at freerun and bdi is well below bdi
 784                 * setpoint). Now the maximum pos_ratio in the same situation
 785                 * is 2. We might want to tweak this if we observe the control
 786                 * system is too slow to adapt.
 787                 */
 788                return min(pos_ratio, bdi_pos_ratio);
 789        }
 790
 791        /*
 792         * We have computed basic pos_ratio above based on global situation. If
 793         * the bdi is over/under its share of dirty pages, we want to scale
 794         * pos_ratio further down/up. That is done by the following mechanism.
 795         */
 796
 797        /*
 798         * bdi setpoint
 799         *
 800         *        f(bdi_dirty) := 1.0 + k * (bdi_dirty - bdi_setpoint)
 801         *
 802         *                        x_intercept - bdi_dirty
 803         *                     := --------------------------
 804         *                        x_intercept - bdi_setpoint
 805         *
 806         * The main bdi control line is a linear function that subjects to
 807         *
 808         * (1) f(bdi_setpoint) = 1.0
 809         * (2) k = - 1 / (8 * write_bw)  (in single bdi case)
 810         *     or equally: x_intercept = bdi_setpoint + 8 * write_bw
 811         *
 812         * For single bdi case, the dirty pages are observed to fluctuate
 813         * regularly within range
 814         *        [bdi_setpoint - write_bw/2, bdi_setpoint + write_bw/2]
 815         * for various filesystems, where (2) can yield in a reasonable 12.5%
 816         * fluctuation range for pos_ratio.
 817         *
 818         * For JBOD case, bdi_thresh (not bdi_dirty!) could fluctuate up to its
 819         * own size, so move the slope over accordingly and choose a slope that
 820         * yields 100% pos_ratio fluctuation on suddenly doubled bdi_thresh.
 821         */
 822        if (unlikely(bdi_thresh > thresh))
 823                bdi_thresh = thresh;
 824        /*
 825         * It's very possible that bdi_thresh is close to 0 not because the
 826         * device is slow, but that it has remained inactive for long time.
 827         * Honour such devices a reasonable good (hopefully IO efficient)
 828         * threshold, so that the occasional writes won't be blocked and active
 829         * writes can rampup the threshold quickly.
 830         */
 831        bdi_thresh = max(bdi_thresh, (limit - dirty) / 8);
 832        /*
 833         * scale global setpoint to bdi's:
 834         *      bdi_setpoint = setpoint * bdi_thresh / thresh
 835         */
 836        x = div_u64((u64)bdi_thresh << 16, thresh + 1);
 837        bdi_setpoint = setpoint * (u64)x >> 16;
 838        /*
 839         * Use span=(8*write_bw) in single bdi case as indicated by
 840         * (thresh - bdi_thresh ~= 0) and transit to bdi_thresh in JBOD case.
 841         *
 842         *        bdi_thresh                    thresh - bdi_thresh
 843         * span = ---------- * (8 * write_bw) + ------------------- * bdi_thresh
 844         *          thresh                            thresh
 845         */
 846        span = (thresh - bdi_thresh + 8 * write_bw) * (u64)x >> 16;
 847        x_intercept = bdi_setpoint + span;
 848
 849        if (bdi_dirty < x_intercept - span / 4) {
 850                pos_ratio = div_u64(pos_ratio * (x_intercept - bdi_dirty),
 851                                    x_intercept - bdi_setpoint + 1);
 852        } else
 853                pos_ratio /= 4;
 854
 855        /*
 856         * bdi reserve area, safeguard against dirty pool underrun and disk idle
 857         * It may push the desired control point of global dirty pages higher
 858         * than setpoint.
 859         */
 860        x_intercept = bdi_thresh / 2;
 861        if (bdi_dirty < x_intercept) {
 862                if (bdi_dirty > x_intercept / 8)
 863                        pos_ratio = div_u64(pos_ratio * x_intercept, bdi_dirty);
 864                else
 865                        pos_ratio *= 8;
 866        }
 867
 868        return pos_ratio;
 869}
 870
 871static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
 872                                       unsigned long elapsed,
 873                                       unsigned long written)
 874{
 875        const unsigned long period = roundup_pow_of_two(3 * HZ);
 876        unsigned long avg = bdi->avg_write_bandwidth;
 877        unsigned long old = bdi->write_bandwidth;
 878        u64 bw;
 879
 880        /*
 881         * bw = written * HZ / elapsed
 882         *
 883         *                   bw * elapsed + write_bandwidth * (period - elapsed)
 884         * write_bandwidth = ---------------------------------------------------
 885         *                                          period
 886         */
 887        bw = written - bdi->written_stamp;
 888        bw *= HZ;
 889        if (unlikely(elapsed > period)) {
 890                do_div(bw, elapsed);
 891                avg = bw;
 892                goto out;
 893        }
 894        bw += (u64)bdi->write_bandwidth * (period - elapsed);
 895        bw >>= ilog2(period);
 896
 897        /*
 898         * one more level of smoothing, for filtering out sudden spikes
 899         */
 900        if (avg > old && old >= (unsigned long)bw)
 901                avg -= (avg - old) >> 3;
 902
 903        if (avg < old && old <= (unsigned long)bw)
 904                avg += (old - avg) >> 3;
 905
 906out:
 907        bdi->write_bandwidth = bw;
 908        bdi->avg_write_bandwidth = avg;
 909}
 910
 911/*
 912 * The global dirtyable memory and dirty threshold could be suddenly knocked
 913 * down by a large amount (eg. on the startup of KVM in a swapless system).
 914 * This may throw the system into deep dirty exceeded state and throttle
 915 * heavy/light dirtiers alike. To retain good responsiveness, maintain
 916 * global_dirty_limit for tracking slowly down to the knocked down dirty
 917 * threshold.
 918 */
 919static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
 920{
 921        unsigned long limit = global_dirty_limit;
 922
 923        /*
 924         * Follow up in one step.
 925         */
 926        if (limit < thresh) {
 927                limit = thresh;
 928                goto update;
 929        }
 930
 931        /*
 932         * Follow down slowly. Use the higher one as the target, because thresh
 933         * may drop below dirty. This is exactly the reason to introduce
 934         * global_dirty_limit which is guaranteed to lie above the dirty pages.
 935         */
 936        thresh = max(thresh, dirty);
 937        if (limit > thresh) {
 938                limit -= (limit - thresh) >> 5;
 939                goto update;
 940        }
 941        return;
 942update:
 943        global_dirty_limit = limit;
 944}
 945
 946static void global_update_bandwidth(unsigned long thresh,
 947                                    unsigned long dirty,
 948                                    unsigned long now)
 949{
 950        static DEFINE_SPINLOCK(dirty_lock);
 951        static unsigned long update_time;
 952
 953        /*
 954         * check locklessly first to optimize away locking for the most time
 955         */
 956        if (time_before(now, update_time + BANDWIDTH_INTERVAL))
 957                return;
 958
 959        spin_lock(&dirty_lock);
 960        if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
 961                update_dirty_limit(thresh, dirty);
 962                update_time = now;
 963        }
 964        spin_unlock(&dirty_lock);
 965}
 966
 967/*
 968 * Maintain bdi->dirty_ratelimit, the base dirty throttle rate.
 969 *
 970 * Normal bdi tasks will be curbed at or below it in long term.
 971 * Obviously it should be around (write_bw / N) when there are N dd tasks.
 972 */
 973static void bdi_update_dirty_ratelimit(struct backing_dev_info *bdi,
 974                                       unsigned long thresh,
 975                                       unsigned long bg_thresh,
 976                                       unsigned long dirty,
 977                                       unsigned long bdi_thresh,
 978                                       unsigned long bdi_dirty,
 979                                       unsigned long dirtied,
 980                                       unsigned long elapsed)
 981{
 982        unsigned long freerun = dirty_freerun_ceiling(thresh, bg_thresh);
 983        unsigned long limit = hard_dirty_limit(thresh);
 984        unsigned long setpoint = (freerun + limit) / 2;
 985        unsigned long write_bw = bdi->avg_write_bandwidth;
 986        unsigned long dirty_ratelimit = bdi->dirty_ratelimit;
 987        unsigned long dirty_rate;
 988        unsigned long task_ratelimit;
 989        unsigned long balanced_dirty_ratelimit;
 990        unsigned long pos_ratio;
 991        unsigned long step;
 992        unsigned long x;
 993
 994        /*
 995         * The dirty rate will match the writeout rate in long term, except
 996         * when dirty pages are truncated by userspace or re-dirtied by FS.
 997         */
 998        dirty_rate = (dirtied - bdi->dirtied_stamp) * HZ / elapsed;
 999
1000        pos_ratio = bdi_position_ratio(bdi, thresh, bg_thresh, dirty,
1001                                       bdi_thresh, bdi_dirty);
1002        /*
1003         * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1004         */
1005        task_ratelimit = (u64)dirty_ratelimit *
1006                                        pos_ratio >> RATELIMIT_CALC_SHIFT;
1007        task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1008
1009        /*
1010         * A linear estimation of the "balanced" throttle rate. The theory is,
1011         * if there are N dd tasks, each throttled at task_ratelimit, the bdi's
1012         * dirty_rate will be measured to be (N * task_ratelimit). So the below
1013         * formula will yield the balanced rate limit (write_bw / N).
1014         *
1015         * Note that the expanded form is not a pure rate feedback:
1016         *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate)              (1)
1017         * but also takes pos_ratio into account:
1018         *      rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio  (2)
1019         *
1020         * (1) is not realistic because pos_ratio also takes part in balancing
1021         * the dirty rate.  Consider the state
1022         *      pos_ratio = 0.5                                              (3)
1023         *      rate = 2 * (write_bw / N)                                    (4)
1024         * If (1) is used, it will stuck in that state! Because each dd will
1025         * be throttled at
1026         *      task_ratelimit = pos_ratio * rate = (write_bw / N)           (5)
1027         * yielding
1028         *      dirty_rate = N * task_ratelimit = write_bw                   (6)
1029         * put (6) into (1) we get
1030         *      rate_(i+1) = rate_(i)                                        (7)
1031         *
1032         * So we end up using (2) to always keep
1033         *      rate_(i+1) ~= (write_bw / N)                                 (8)
1034         * regardless of the value of pos_ratio. As long as (8) is satisfied,
1035         * pos_ratio is able to drive itself to 1.0, which is not only where
1036         * the dirty count meet the setpoint, but also where the slope of
1037         * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1038         */
1039        balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1040                                           dirty_rate | 1);
1041        /*
1042         * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1043         */
1044        if (unlikely(balanced_dirty_ratelimit > write_bw))
1045                balanced_dirty_ratelimit = write_bw;
1046
1047        /*
1048         * We could safely do this and return immediately:
1049         *
1050         *      bdi->dirty_ratelimit = balanced_dirty_ratelimit;
1051         *
1052         * However to get a more stable dirty_ratelimit, the below elaborated
1053         * code makes use of task_ratelimit to filter out singular points and
1054         * limit the step size.
1055         *
1056         * The below code essentially only uses the relative value of
1057         *
1058         *      task_ratelimit - dirty_ratelimit
1059         *      = (pos_ratio - 1) * dirty_ratelimit
1060         *
1061         * which reflects the direction and size of dirty position error.
1062         */
1063
1064        /*
1065         * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1066         * task_ratelimit is on the same side of dirty_ratelimit, too.
1067         * For example, when
1068         * - dirty_ratelimit > balanced_dirty_ratelimit
1069         * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1070         * lowering dirty_ratelimit will help meet both the position and rate
1071         * control targets. Otherwise, don't update dirty_ratelimit if it will
1072         * only help meet the rate target. After all, what the users ultimately
1073         * feel and care are stable dirty rate and small position error.
1074         *
1075         * |task_ratelimit - dirty_ratelimit| is used to limit the step size
1076         * and filter out the singular points of balanced_dirty_ratelimit. Which
1077         * keeps jumping around randomly and can even leap far away at times
1078         * due to the small 200ms estimation period of dirty_rate (we want to
1079         * keep that period small to reduce time lags).
1080         */
1081        step = 0;
1082
1083        /*
1084         * For strictlimit case, calculations above were based on bdi counters
1085         * and limits (starting from pos_ratio = bdi_position_ratio() and up to
1086         * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
1087         * Hence, to calculate "step" properly, we have to use bdi_dirty as
1088         * "dirty" and bdi_setpoint as "setpoint".
1089         *
1090         * We rampup dirty_ratelimit forcibly if bdi_dirty is low because
1091         * it's possible that bdi_thresh is close to zero due to inactivity
1092         * of backing device (see the implementation of bdi_dirty_limit()).
1093         */
1094        if (unlikely(bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
1095                dirty = bdi_dirty;
1096                if (bdi_dirty < 8)
1097                        setpoint = bdi_dirty + 1;
1098                else
1099                        setpoint = (bdi_thresh +
1100                                    bdi_dirty_limit(bdi, bg_thresh)) / 2;
1101        }
1102
1103        if (dirty < setpoint) {
1104                x = min(bdi->balanced_dirty_ratelimit,
1105                         min(balanced_dirty_ratelimit, task_ratelimit));
1106                if (dirty_ratelimit < x)
1107                        step = x - dirty_ratelimit;
1108        } else {
1109                x = max(bdi->balanced_dirty_ratelimit,
1110                         max(balanced_dirty_ratelimit, task_ratelimit));
1111                if (dirty_ratelimit > x)
1112                        step = dirty_ratelimit - x;
1113        }
1114
1115        /*
1116         * Don't pursue 100% rate matching. It's impossible since the balanced
1117         * rate itself is constantly fluctuating. So decrease the track speed
1118         * when it gets close to the target. Helps eliminate pointless tremors.
1119         */
1120        step >>= dirty_ratelimit / (2 * step + 1);
1121        /*
1122         * Limit the tracking speed to avoid overshooting.
1123         */
1124        step = (step + 7) / 8;
1125
1126        if (dirty_ratelimit < balanced_dirty_ratelimit)
1127                dirty_ratelimit += step;
1128        else
1129                dirty_ratelimit -= step;
1130
1131        bdi->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1132        bdi->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
1133
1134        trace_bdi_dirty_ratelimit(bdi, dirty_rate, task_ratelimit);
1135}
1136
1137void __bdi_update_bandwidth(struct backing_dev_info *bdi,
1138                            unsigned long thresh,
1139                            unsigned long bg_thresh,
1140                            unsigned long dirty,
1141                            unsigned long bdi_thresh,
1142                            unsigned long bdi_dirty,
1143                            unsigned long start_time)
1144{
1145        unsigned long now = jiffies;
1146        unsigned long elapsed = now - bdi->bw_time_stamp;
1147        unsigned long dirtied;
1148        unsigned long written;
1149
1150        /*
1151         * rate-limit, only update once every 200ms.
1152         */
1153        if (elapsed < BANDWIDTH_INTERVAL)
1154                return;
1155
1156        dirtied = percpu_counter_read(&bdi->bdi_stat[BDI_DIRTIED]);
1157        written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
1158
1159        /*
1160         * Skip quiet periods when disk bandwidth is under-utilized.
1161         * (at least 1s idle time between two flusher runs)
1162         */
1163        if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
1164                goto snapshot;
1165
1166        if (thresh) {
1167                global_update_bandwidth(thresh, dirty, now);
1168                bdi_update_dirty_ratelimit(bdi, thresh, bg_thresh, dirty,
1169                                           bdi_thresh, bdi_dirty,
1170                                           dirtied, elapsed);
1171        }
1172        bdi_update_write_bandwidth(bdi, elapsed, written);
1173
1174snapshot:
1175        bdi->dirtied_stamp = dirtied;
1176        bdi->written_stamp = written;
1177        bdi->bw_time_stamp = now;
1178}
1179
1180static void bdi_update_bandwidth(struct backing_dev_info *bdi,
1181                                 unsigned long thresh,
1182                                 unsigned long bg_thresh,
1183                                 unsigned long dirty,
1184                                 unsigned long bdi_thresh,
1185                                 unsigned long bdi_dirty,
1186                                 unsigned long start_time)
1187{
1188        if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
1189                return;
1190        spin_lock(&bdi->wb.list_lock);
1191        __bdi_update_bandwidth(bdi, thresh, bg_thresh, dirty,
1192                               bdi_thresh, bdi_dirty, start_time);
1193        spin_unlock(&bdi->wb.list_lock);
1194}
1195
1196/*
1197 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
1198 * will look to see if it needs to start dirty throttling.
1199 *
1200 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
1201 * global_page_state() too often. So scale it near-sqrt to the safety margin
1202 * (the number of pages we may dirty without exceeding the dirty limits).
1203 */
1204static unsigned long dirty_poll_interval(unsigned long dirty,
1205                                         unsigned long thresh)
1206{
1207        if (thresh > dirty)
1208                return 1UL << (ilog2(thresh - dirty) >> 1);
1209
1210        return 1;
1211}
1212
1213static unsigned long bdi_max_pause(struct backing_dev_info *bdi,
1214                                   unsigned long bdi_dirty)
1215{
1216        unsigned long bw = bdi->avg_write_bandwidth;
1217        unsigned long t;
1218
1219        /*
1220         * Limit pause time for small memory systems. If sleeping for too long
1221         * time, a small pool of dirty/writeback pages may go empty and disk go
1222         * idle.
1223         *
1224         * 8 serves as the safety ratio.
1225         */
1226        t = bdi_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
1227        t++;
1228
1229        return min_t(unsigned long, t, MAX_PAUSE);
1230}
1231
1232static long bdi_min_pause(struct backing_dev_info *bdi,
1233                          long max_pause,
1234                          unsigned long task_ratelimit,
1235                          unsigned long dirty_ratelimit,
1236                          int *nr_dirtied_pause)
1237{
1238        long hi = ilog2(bdi->avg_write_bandwidth);
1239        long lo = ilog2(bdi->dirty_ratelimit);
1240        long t;         /* target pause */
1241        long pause;     /* estimated next pause */
1242        int pages;      /* target nr_dirtied_pause */
1243
1244        /* target for 10ms pause on 1-dd case */
1245        t = max(1, HZ / 100);
1246
1247        /*
1248         * Scale up pause time for concurrent dirtiers in order to reduce CPU
1249         * overheads.
1250         *
1251         * (N * 10ms) on 2^N concurrent tasks.
1252         */
1253        if (hi > lo)
1254                t += (hi - lo) * (10 * HZ) / 1024;
1255
1256        /*
1257         * This is a bit convoluted. We try to base the next nr_dirtied_pause
1258         * on the much more stable dirty_ratelimit. However the next pause time
1259         * will be computed based on task_ratelimit and the two rate limits may
1260         * depart considerably at some time. Especially if task_ratelimit goes
1261         * below dirty_ratelimit/2 and the target pause is max_pause, the next
1262         * pause time will be max_pause*2 _trimmed down_ to max_pause.  As a
1263         * result task_ratelimit won't be executed faithfully, which could
1264         * eventually bring down dirty_ratelimit.
1265         *
1266         * We apply two rules to fix it up:
1267         * 1) try to estimate the next pause time and if necessary, use a lower
1268         *    nr_dirtied_pause so as not to exceed max_pause. When this happens,
1269         *    nr_dirtied_pause will be "dancing" with task_ratelimit.
1270         * 2) limit the target pause time to max_pause/2, so that the normal
1271         *    small fluctuations of task_ratelimit won't trigger rule (1) and
1272         *    nr_dirtied_pause will remain as stable as dirty_ratelimit.
1273         */
1274        t = min(t, 1 + max_pause / 2);
1275        pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1276
1277        /*
1278         * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1279         * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1280         * When the 16 consecutive reads are often interrupted by some dirty
1281         * throttling pause during the async writes, cfq will go into idles
1282         * (deadline is fine). So push nr_dirtied_pause as high as possible
1283         * until reaches DIRTY_POLL_THRESH=32 pages.
1284         */
1285        if (pages < DIRTY_POLL_THRESH) {
1286                t = max_pause;
1287                pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1288                if (pages > DIRTY_POLL_THRESH) {
1289                        pages = DIRTY_POLL_THRESH;
1290                        t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1291                }
1292        }
1293
1294        pause = HZ * pages / (task_ratelimit + 1);
1295        if (pause > max_pause) {
1296                t = max_pause;
1297                pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1298        }
1299
1300        *nr_dirtied_pause = pages;
1301        /*
1302         * The minimal pause time will normally be half the target pause time.
1303         */
1304        return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
1305}
1306
1307static inline void bdi_dirty_limits(struct backing_dev_info *bdi,
1308                                    unsigned long dirty_thresh,
1309                                    unsigned long background_thresh,
1310                                    unsigned long *bdi_dirty,
1311                                    unsigned long *bdi_thresh,
1312                                    unsigned long *bdi_bg_thresh)
1313{
1314        unsigned long bdi_reclaimable;
1315
1316        /*
1317         * bdi_thresh is not treated as some limiting factor as
1318         * dirty_thresh, due to reasons
1319         * - in JBOD setup, bdi_thresh can fluctuate a lot
1320         * - in a system with HDD and USB key, the USB key may somehow
1321         *   go into state (bdi_dirty >> bdi_thresh) either because
1322         *   bdi_dirty starts high, or because bdi_thresh drops low.
1323         *   In this case we don't want to hard throttle the USB key
1324         *   dirtiers for 100 seconds until bdi_dirty drops under
1325         *   bdi_thresh. Instead the auxiliary bdi control line in
1326         *   bdi_position_ratio() will let the dirtier task progress
1327         *   at some rate <= (write_bw / 2) for bringing down bdi_dirty.
1328         */
1329        *bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
1330
1331        if (bdi_bg_thresh)
1332                *bdi_bg_thresh = div_u64((u64)*bdi_thresh *
1333                                         background_thresh,
1334                                         dirty_thresh);
1335
1336        /*
1337         * In order to avoid the stacked BDI deadlock we need
1338         * to ensure we accurately count the 'dirty' pages when
1339         * the threshold is low.
1340         *
1341         * Otherwise it would be possible to get thresh+n pages
1342         * reported dirty, even though there are thresh-m pages
1343         * actually dirty; with m+n sitting in the percpu
1344         * deltas.
1345         */
1346        if (*bdi_thresh < 2 * bdi_stat_error(bdi)) {
1347                bdi_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
1348                *bdi_dirty = bdi_reclaimable +
1349                        bdi_stat_sum(bdi, BDI_WRITEBACK);
1350        } else {
1351                bdi_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
1352                *bdi_dirty = bdi_reclaimable +
1353                        bdi_stat(bdi, BDI_WRITEBACK);
1354        }
1355}
1356
1357/*
1358 * balance_dirty_pages() must be called by processes which are generating dirty
1359 * data.  It looks at the number of dirty pages in the machine and will force
1360 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
1361 * If we're over `background_thresh' then the writeback threads are woken to
1362 * perform some writeout.
1363 */
1364static void balance_dirty_pages(struct address_space *mapping,
1365                                unsigned long pages_dirtied)
1366{
1367        unsigned long nr_reclaimable;   /* = file_dirty + unstable_nfs */
1368        unsigned long nr_dirty;  /* = file_dirty + writeback + unstable_nfs */
1369        unsigned long background_thresh;
1370        unsigned long dirty_thresh;
1371        long period;
1372        long pause;
1373        long max_pause;
1374        long min_pause;
1375        int nr_dirtied_pause;
1376        bool dirty_exceeded = false;
1377        unsigned long task_ratelimit;
1378        unsigned long dirty_ratelimit;
1379        unsigned long pos_ratio;
1380        struct backing_dev_info *bdi = mapping->backing_dev_info;
1381        bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
1382        unsigned long start_time = jiffies;
1383
1384        for (;;) {
1385                unsigned long now = jiffies;
1386                unsigned long uninitialized_var(bdi_thresh);
1387                unsigned long thresh;
1388                unsigned long uninitialized_var(bdi_dirty);
1389                unsigned long dirty;
1390                unsigned long bg_thresh;
1391
1392                /*
1393                 * Unstable writes are a feature of certain networked
1394                 * filesystems (i.e. NFS) in which data may have been
1395                 * written to the server's write cache, but has not yet
1396                 * been flushed to permanent storage.
1397                 */
1398                nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
1399                                        global_page_state(NR_UNSTABLE_NFS);
1400                nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
1401
1402                global_dirty_limits(&background_thresh, &dirty_thresh);
1403
1404                if (unlikely(strictlimit)) {
1405                        bdi_dirty_limits(bdi, dirty_thresh, background_thresh,
1406                                         &bdi_dirty, &bdi_thresh, &bg_thresh);
1407
1408                        dirty = bdi_dirty;
1409                        thresh = bdi_thresh;
1410                } else {
1411                        dirty = nr_dirty;
1412                        thresh = dirty_thresh;
1413                        bg_thresh = background_thresh;
1414                }
1415
1416                /*
1417                 * Throttle it only when the background writeback cannot
1418                 * catch-up. This avoids (excessively) small writeouts
1419                 * when the bdi limits are ramping up in case of !strictlimit.
1420                 *
1421                 * In strictlimit case make decision based on the bdi counters
1422                 * and limits. Small writeouts when the bdi limits are ramping
1423                 * up are the price we consciously pay for strictlimit-ing.
1424                 */
1425                if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh)) {
1426                        current->dirty_paused_when = now;
1427                        current->nr_dirtied = 0;
1428                        current->nr_dirtied_pause =
1429                                dirty_poll_interval(dirty, thresh);
1430                        break;
1431                }
1432
1433                if (unlikely(!writeback_in_progress(bdi)))
1434                        bdi_start_background_writeback(bdi);
1435
1436                if (!strictlimit)
1437                        bdi_dirty_limits(bdi, dirty_thresh, background_thresh,
1438                                         &bdi_dirty, &bdi_thresh, NULL);
1439
1440                dirty_exceeded = (bdi_dirty > bdi_thresh) &&
1441                                 ((nr_dirty > dirty_thresh) || strictlimit);
1442                if (dirty_exceeded && !bdi->dirty_exceeded)
1443                        bdi->dirty_exceeded = 1;
1444
1445                bdi_update_bandwidth(bdi, dirty_thresh, background_thresh,
1446                                     nr_dirty, bdi_thresh, bdi_dirty,
1447                                     start_time);
1448
1449                dirty_ratelimit = bdi->dirty_ratelimit;
1450                pos_ratio = bdi_position_ratio(bdi, dirty_thresh,
1451                                               background_thresh, nr_dirty,
1452                                               bdi_thresh, bdi_dirty);
1453                task_ratelimit = ((u64)dirty_ratelimit * pos_ratio) >>
1454                                                        RATELIMIT_CALC_SHIFT;
1455                max_pause = bdi_max_pause(bdi, bdi_dirty);
1456                min_pause = bdi_min_pause(bdi, max_pause,
1457                                          task_ratelimit, dirty_ratelimit,
1458                                          &nr_dirtied_pause);
1459
1460                if (unlikely(task_ratelimit == 0)) {
1461                        period = max_pause;
1462                        pause = max_pause;
1463                        goto pause;
1464                }
1465                period = HZ * pages_dirtied / task_ratelimit;
1466                pause = period;
1467                if (current->dirty_paused_when)
1468                        pause -= now - current->dirty_paused_when;
1469                /*
1470                 * For less than 1s think time (ext3/4 may block the dirtier
1471                 * for up to 800ms from time to time on 1-HDD; so does xfs,
1472                 * however at much less frequency), try to compensate it in
1473                 * future periods by updating the virtual time; otherwise just
1474                 * do a reset, as it may be a light dirtier.
1475                 */
1476                if (pause < min_pause) {
1477                        trace_balance_dirty_pages(bdi,
1478                                                  dirty_thresh,
1479                                                  background_thresh,
1480                                                  nr_dirty,
1481                                                  bdi_thresh,
1482                                                  bdi_dirty,
1483                                                  dirty_ratelimit,
1484                                                  task_ratelimit,
1485                                                  pages_dirtied,
1486                                                  period,
1487                                                  min(pause, 0L),
1488                                                  start_time);
1489                        if (pause < -HZ) {
1490                                current->dirty_paused_when = now;
1491                                current->nr_dirtied = 0;
1492                        } else if (period) {
1493                                current->dirty_paused_when += period;
1494                                current->nr_dirtied = 0;
1495                        } else if (current->nr_dirtied_pause <= pages_dirtied)
1496                                current->nr_dirtied_pause += pages_dirtied;
1497                        break;
1498                }
1499                if (unlikely(pause > max_pause)) {
1500                        /* for occasional dropped task_ratelimit */
1501                        now += min(pause - max_pause, max_pause);
1502                        pause = max_pause;
1503                }
1504
1505pause:
1506                trace_balance_dirty_pages(bdi,
1507                                          dirty_thresh,
1508                                          background_thresh,
1509                                          nr_dirty,
1510                                          bdi_thresh,
1511                                          bdi_dirty,
1512                                          dirty_ratelimit,
1513                                          task_ratelimit,
1514                                          pages_dirtied,
1515                                          period,
1516                                          pause,
1517                                          start_time);
1518                __set_current_state(TASK_KILLABLE);
1519                io_schedule_timeout(pause);
1520
1521                current->dirty_paused_when = now + pause;
1522                current->nr_dirtied = 0;
1523                current->nr_dirtied_pause = nr_dirtied_pause;
1524
1525                /*
1526                 * This is typically equal to (nr_dirty < dirty_thresh) and can
1527                 * also keep "1000+ dd on a slow USB stick" under control.
1528                 */
1529                if (task_ratelimit)
1530                        break;
1531
1532                /*
1533                 * In the case of an unresponding NFS server and the NFS dirty
1534                 * pages exceeds dirty_thresh, give the other good bdi's a pipe
1535                 * to go through, so that tasks on them still remain responsive.
1536                 *
1537                 * In theory 1 page is enough to keep the comsumer-producer
1538                 * pipe going: the flusher cleans 1 page => the task dirties 1
1539                 * more page. However bdi_dirty has accounting errors.  So use
1540                 * the larger and more IO friendly bdi_stat_error.
1541                 */
1542                if (bdi_dirty <= bdi_stat_error(bdi))
1543                        break;
1544
1545                if (fatal_signal_pending(current))
1546                        break;
1547        }
1548
1549        if (!dirty_exceeded && bdi->dirty_exceeded)
1550                bdi->dirty_exceeded = 0;
1551
1552        if (writeback_in_progress(bdi))
1553                return;
1554
1555        /*
1556         * In laptop mode, we wait until hitting the higher threshold before
1557         * starting background writeout, and then write out all the way down
1558         * to the lower threshold.  So slow writers cause minimal disk activity.
1559         *
1560         * In normal mode, we start background writeout at the lower
1561         * background_thresh, to keep the amount of dirty memory low.
1562         */
1563        if (laptop_mode)
1564                return;
1565
1566        if (nr_reclaimable > background_thresh)
1567                bdi_start_background_writeback(bdi);
1568}
1569
1570void set_page_dirty_balance(struct page *page, int page_mkwrite)
1571{
1572        if (set_page_dirty(page) || page_mkwrite) {
1573                struct address_space *mapping = page_mapping(page);
1574
1575                if (mapping)
1576                        balance_dirty_pages_ratelimited(mapping);
1577        }
1578}
1579
1580static DEFINE_PER_CPU(int, bdp_ratelimits);
1581
1582/*
1583 * Normal tasks are throttled by
1584 *      loop {
1585 *              dirty tsk->nr_dirtied_pause pages;
1586 *              take a snap in balance_dirty_pages();
1587 *      }
1588 * However there is a worst case. If every task exit immediately when dirtied
1589 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1590 * called to throttle the page dirties. The solution is to save the not yet
1591 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1592 * randomly into the running tasks. This works well for the above worst case,
1593 * as the new task will pick up and accumulate the old task's leaked dirty
1594 * count and eventually get throttled.
1595 */
1596DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1597
1598/**
1599 * balance_dirty_pages_ratelimited - balance dirty memory state
1600 * @mapping: address_space which was dirtied
1601 *
1602 * Processes which are dirtying memory should call in here once for each page
1603 * which was newly dirtied.  The function will periodically check the system's
1604 * dirty state and will initiate writeback if needed.
1605 *
1606 * On really big machines, get_writeback_state is expensive, so try to avoid
1607 * calling it too often (ratelimiting).  But once we're over the dirty memory
1608 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1609 * from overshooting the limit by (ratelimit_pages) each.
1610 */
1611void balance_dirty_pages_ratelimited(struct address_space *mapping)
1612{
1613        struct backing_dev_info *bdi = mapping->backing_dev_info;
1614        int ratelimit;
1615        int *p;
1616
1617        if (!bdi_cap_account_dirty(bdi))
1618                return;
1619
1620        ratelimit = current->nr_dirtied_pause;
1621        if (bdi->dirty_exceeded)
1622                ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
1623
1624        preempt_disable();
1625        /*
1626         * This prevents one CPU to accumulate too many dirtied pages without
1627         * calling into balance_dirty_pages(), which can happen when there are
1628         * 1000+ tasks, all of them start dirtying pages at exactly the same
1629         * time, hence all honoured too large initial task->nr_dirtied_pause.
1630         */
1631        p =  &__get_cpu_var(bdp_ratelimits);
1632        if (unlikely(current->nr_dirtied >= ratelimit))
1633                *p = 0;
1634        else if (unlikely(*p >= ratelimit_pages)) {
1635                *p = 0;
1636                ratelimit = 0;
1637        }
1638        /*
1639         * Pick up the dirtied pages by the exited tasks. This avoids lots of
1640         * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1641         * the dirty throttling and livelock other long-run dirtiers.
1642         */
1643        p = &__get_cpu_var(dirty_throttle_leaks);
1644        if (*p > 0 && current->nr_dirtied < ratelimit) {
1645                unsigned long nr_pages_dirtied;
1646                nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1647                *p -= nr_pages_dirtied;
1648                current->nr_dirtied += nr_pages_dirtied;
1649        }
1650        preempt_enable();
1651
1652        if (unlikely(current->nr_dirtied >= ratelimit))
1653                balance_dirty_pages(mapping, current->nr_dirtied);
1654}
1655EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
1656
1657void throttle_vm_writeout(gfp_t gfp_mask)
1658{
1659        unsigned long background_thresh;
1660        unsigned long dirty_thresh;
1661
1662        for ( ; ; ) {
1663                global_dirty_limits(&background_thresh, &dirty_thresh);
1664                dirty_thresh = hard_dirty_limit(dirty_thresh);
1665
1666                /*
1667                 * Boost the allowable dirty threshold a bit for page
1668                 * allocators so they don't get DoS'ed by heavy writers
1669                 */
1670                dirty_thresh += dirty_thresh / 10;      /* wheeee... */
1671
1672                if (global_page_state(NR_UNSTABLE_NFS) +
1673                        global_page_state(NR_WRITEBACK) <= dirty_thresh)
1674                                break;
1675                congestion_wait(BLK_RW_ASYNC, HZ/10);
1676
1677                /*
1678                 * The caller might hold locks which can prevent IO completion
1679                 * or progress in the filesystem.  So we cannot just sit here
1680                 * waiting for IO to complete.
1681                 */
1682                if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
1683                        break;
1684        }
1685}
1686
1687/*
1688 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1689 */
1690int dirty_writeback_centisecs_handler(ctl_table *table, int write,
1691        void __user *buffer, size_t *length, loff_t *ppos)
1692{
1693        proc_dointvec(table, write, buffer, length, ppos);
1694        return 0;
1695}
1696
1697#ifdef CONFIG_BLOCK
1698void laptop_mode_timer_fn(unsigned long data)
1699{
1700        struct request_queue *q = (struct request_queue *)data;
1701        int nr_pages = global_page_state(NR_FILE_DIRTY) +
1702                global_page_state(NR_UNSTABLE_NFS);
1703
1704        /*
1705         * We want to write everything out, not just down to the dirty
1706         * threshold
1707         */
1708        if (bdi_has_dirty_io(&q->backing_dev_info))
1709                bdi_start_writeback(&q->backing_dev_info, nr_pages,
1710                                        WB_REASON_LAPTOP_TIMER);
1711}
1712
1713/*
1714 * We've spun up the disk and we're in laptop mode: schedule writeback
1715 * of all dirty data a few seconds from now.  If the flush is already scheduled
1716 * then push it back - the user is still using the disk.
1717 */
1718void laptop_io_completion(struct backing_dev_info *info)
1719{
1720        mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
1721}
1722
1723/*
1724 * We're in laptop mode and we've just synced. The sync's writes will have
1725 * caused another writeback to be scheduled by laptop_io_completion.
1726 * Nothing needs to be written back anymore, so we unschedule the writeback.
1727 */
1728void laptop_sync_completion(void)
1729{
1730        struct backing_dev_info *bdi;
1731
1732        rcu_read_lock();
1733
1734        list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
1735                del_timer(&bdi->laptop_mode_wb_timer);
1736
1737        rcu_read_unlock();
1738}
1739#endif
1740
1741/*
1742 * If ratelimit_pages is too high then we can get into dirty-data overload
1743 * if a large number of processes all perform writes at the same time.
1744 * If it is too low then SMP machines will call the (expensive)
1745 * get_writeback_state too often.
1746 *
1747 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
1748 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
1749 * thresholds.
1750 */
1751
1752void writeback_set_ratelimit(void)
1753{
1754        unsigned long background_thresh;
1755        unsigned long dirty_thresh;
1756        global_dirty_limits(&background_thresh, &dirty_thresh);
1757        global_dirty_limit = dirty_thresh;
1758        ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
1759        if (ratelimit_pages < 16)
1760                ratelimit_pages = 16;
1761}
1762
1763static int
1764ratelimit_handler(struct notifier_block *self, unsigned long action,
1765                  void *hcpu)
1766{
1767
1768        switch (action & ~CPU_TASKS_FROZEN) {
1769        case CPU_ONLINE:
1770        case CPU_DEAD:
1771                writeback_set_ratelimit();
1772                return NOTIFY_OK;
1773        default:
1774                return NOTIFY_DONE;
1775        }
1776}
1777
1778static struct notifier_block ratelimit_nb = {
1779        .notifier_call  = ratelimit_handler,
1780        .next           = NULL,
1781};
1782
1783/*
1784 * Called early on to tune the page writeback dirty limits.
1785 *
1786 * We used to scale dirty pages according to how total memory
1787 * related to pages that could be allocated for buffers (by
1788 * comparing nr_free_buffer_pages() to vm_total_pages.
1789 *
1790 * However, that was when we used "dirty_ratio" to scale with
1791 * all memory, and we don't do that any more. "dirty_ratio"
1792 * is now applied to total non-HIGHPAGE memory (by subtracting
1793 * totalhigh_pages from vm_total_pages), and as such we can't
1794 * get into the old insane situation any more where we had
1795 * large amounts of dirty pages compared to a small amount of
1796 * non-HIGHMEM memory.
1797 *
1798 * But we might still want to scale the dirty_ratio by how
1799 * much memory the box has..
1800 */
1801void __init page_writeback_init(void)
1802{
1803        writeback_set_ratelimit();
1804        register_cpu_notifier(&ratelimit_nb);
1805
1806        fprop_global_init(&writeout_completions);
1807}
1808
1809/**
1810 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
1811 * @mapping: address space structure to write
1812 * @start: starting page index
1813 * @end: ending page index (inclusive)
1814 *
1815 * This function scans the page range from @start to @end (inclusive) and tags
1816 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
1817 * that write_cache_pages (or whoever calls this function) will then use
1818 * TOWRITE tag to identify pages eligible for writeback.  This mechanism is
1819 * used to avoid livelocking of writeback by a process steadily creating new
1820 * dirty pages in the file (thus it is important for this function to be quick
1821 * so that it can tag pages faster than a dirtying process can create them).
1822 */
1823/*
1824 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce tree_lock latency.
1825 */
1826void tag_pages_for_writeback(struct address_space *mapping,
1827                             pgoff_t start, pgoff_t end)
1828{
1829#define WRITEBACK_TAG_BATCH 4096
1830        unsigned long tagged;
1831
1832        do {
1833                spin_lock_irq(&mapping->tree_lock);
1834                tagged = radix_tree_range_tag_if_tagged(&mapping->page_tree,
1835                                &start, end, WRITEBACK_TAG_BATCH,
1836                                PAGECACHE_TAG_DIRTY, PAGECACHE_TAG_TOWRITE);
1837                spin_unlock_irq(&mapping->tree_lock);
1838                WARN_ON_ONCE(tagged > WRITEBACK_TAG_BATCH);
1839                cond_resched();
1840                /* We check 'start' to handle wrapping when end == ~0UL */
1841        } while (tagged >= WRITEBACK_TAG_BATCH && start);
1842}
1843EXPORT_SYMBOL(tag_pages_for_writeback);
1844
1845/**
1846 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
1847 * @mapping: address space structure to write
1848 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
1849 * @writepage: function called for each page
1850 * @data: data passed to writepage function
1851 *
1852 * If a page is already under I/O, write_cache_pages() skips it, even
1853 * if it's dirty.  This is desirable behaviour for memory-cleaning writeback,
1854 * but it is INCORRECT for data-integrity system calls such as fsync().  fsync()
1855 * and msync() need to guarantee that all the data which was dirty at the time
1856 * the call was made get new I/O started against them.  If wbc->sync_mode is
1857 * WB_SYNC_ALL then we were called for data integrity and we must wait for
1858 * existing IO to complete.
1859 *
1860 * To avoid livelocks (when other process dirties new pages), we first tag
1861 * pages which should be written back with TOWRITE tag and only then start
1862 * writing them. For data-integrity sync we have to be careful so that we do
1863 * not miss some pages (e.g., because some other process has cleared TOWRITE
1864 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
1865 * by the process clearing the DIRTY tag (and submitting the page for IO).
1866 */
1867int write_cache_pages(struct address_space *mapping,
1868                      struct writeback_control *wbc, writepage_t writepage,
1869                      void *data)
1870{
1871        int ret = 0;
1872        int done = 0;
1873        struct pagevec pvec;
1874        int nr_pages;
1875        pgoff_t uninitialized_var(writeback_index);
1876        pgoff_t index;
1877        pgoff_t end;            /* Inclusive */
1878        pgoff_t done_index;
1879        int cycled;
1880        int range_whole = 0;
1881        int tag;
1882
1883        pagevec_init(&pvec, 0);
1884        if (wbc->range_cyclic) {
1885                writeback_index = mapping->writeback_index; /* prev offset */
1886                index = writeback_index;
1887                if (index == 0)
1888                        cycled = 1;
1889                else
1890                        cycled = 0;
1891                end = -1;
1892        } else {
1893                index = wbc->range_start >> PAGE_CACHE_SHIFT;
1894                end = wbc->range_end >> PAGE_CACHE_SHIFT;
1895                if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
1896                        range_whole = 1;
1897                cycled = 1; /* ignore range_cyclic tests */
1898        }
1899        if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1900                tag = PAGECACHE_TAG_TOWRITE;
1901        else
1902                tag = PAGECACHE_TAG_DIRTY;
1903retry:
1904        if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1905                tag_pages_for_writeback(mapping, index, end);
1906        done_index = index;
1907        while (!done && (index <= end)) {
1908                int i;
1909
1910                nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1911                              min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1912                if (nr_pages == 0)
1913                        break;
1914
1915                for (i = 0; i < nr_pages; i++) {
1916                        struct page *page = pvec.pages[i];
1917
1918                        /*
1919                         * At this point, the page may be truncated or
1920                         * invalidated (changing page->mapping to NULL), or
1921                         * even swizzled back from swapper_space to tmpfs file
1922                         * mapping. However, page->index will not change
1923                         * because we have a reference on the page.
1924                         */
1925                        if (page->index > end) {
1926                                /*
1927                                 * can't be range_cyclic (1st pass) because
1928                                 * end == -1 in that case.
1929                                 */
1930                                done = 1;
1931                                break;
1932                        }
1933
1934                        done_index = page->index;
1935
1936                        lock_page(page);
1937
1938                        /*
1939                         * Page truncated or invalidated. We can freely skip it
1940                         * then, even for data integrity operations: the page
1941                         * has disappeared concurrently, so there could be no
1942                         * real expectation of this data interity operation
1943                         * even if there is now a new, dirty page at the same
1944                         * pagecache address.
1945                         */
1946                        if (unlikely(page->mapping != mapping)) {
1947continue_unlock:
1948                                unlock_page(page);
1949                                continue;
1950                        }
1951
1952                        if (!PageDirty(page)) {
1953                                /* someone wrote it for us */
1954                                goto continue_unlock;
1955                        }
1956
1957                        if (PageWriteback(page)) {
1958                                if (wbc->sync_mode != WB_SYNC_NONE)
1959                                        wait_on_page_writeback(page);
1960                                else
1961                                        goto continue_unlock;
1962                        }
1963
1964                        BUG_ON(PageWriteback(page));
1965                        if (!clear_page_dirty_for_io(page))
1966                                goto continue_unlock;
1967
1968                        trace_wbc_writepage(wbc, mapping->backing_dev_info);
1969                        ret = (*writepage)(page, wbc, data);
1970                        if (unlikely(ret)) {
1971                                if (ret == AOP_WRITEPAGE_ACTIVATE) {
1972                                        unlock_page(page);
1973                                        ret = 0;
1974                                } else {
1975                                        /*
1976                                         * done_index is set past this page,
1977                                         * so media errors will not choke
1978                                         * background writeout for the entire
1979                                         * file. This has consequences for
1980                                         * range_cyclic semantics (ie. it may
1981                                         * not be suitable for data integrity
1982                                         * writeout).
1983                                         */
1984                                        done_index = page->index + 1;
1985                                        done = 1;
1986                                        break;
1987                                }
1988                        }
1989
1990                        /*
1991                         * We stop writing back only if we are not doing
1992                         * integrity sync. In case of integrity sync we have to
1993                         * keep going until we have written all the pages
1994                         * we tagged for writeback prior to entering this loop.
1995                         */
1996                        if (--wbc->nr_to_write <= 0 &&
1997                            wbc->sync_mode == WB_SYNC_NONE) {
1998                                done = 1;
1999                                break;
2000                        }
2001                }
2002                pagevec_release(&pvec);
2003                cond_resched();
2004        }
2005        if (!cycled && !done) {
2006                /*
2007                 * range_cyclic:
2008                 * We hit the last page and there is more work to be done: wrap
2009                 * back to the start of the file
2010                 */
2011                cycled = 1;
2012                index = 0;
2013                end = writeback_index - 1;
2014                goto retry;
2015        }
2016        if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2017                mapping->writeback_index = done_index;
2018
2019        return ret;
2020}
2021EXPORT_SYMBOL(write_cache_pages);
2022
2023/*
2024 * Function used by generic_writepages to call the real writepage
2025 * function and set the mapping flags on error
2026 */
2027static int __writepage(struct page *page, struct writeback_control *wbc,
2028                       void *data)
2029{
2030        struct address_space *mapping = data;
2031        int ret = mapping->a_ops->writepage(page, wbc);
2032        mapping_set_error(mapping, ret);
2033        return ret;
2034}
2035
2036/**
2037 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2038 * @mapping: address space structure to write
2039 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2040 *
2041 * This is a library function, which implements the writepages()
2042 * address_space_operation.
2043 */
2044int generic_writepages(struct address_space *mapping,
2045                       struct writeback_control *wbc)
2046{
2047        struct blk_plug plug;
2048        int ret;
2049
2050        /* deal with chardevs and other special file */
2051        if (!mapping->a_ops->writepage)
2052                return 0;
2053
2054        blk_start_plug(&plug);
2055        ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2056        blk_finish_plug(&plug);
2057        return ret;
2058}
2059
2060EXPORT_SYMBOL(generic_writepages);
2061
2062int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2063{
2064        int ret;
2065
2066        if (wbc->nr_to_write <= 0)
2067                return 0;
2068        if (mapping->a_ops->writepages)
2069                ret = mapping->a_ops->writepages(mapping, wbc);
2070        else
2071                ret = generic_writepages(mapping, wbc);
2072        return ret;
2073}
2074
2075/**
2076 * write_one_page - write out a single page and optionally wait on I/O
2077 * @page: the page to write
2078 * @wait: if true, wait on writeout
2079 *
2080 * The page must be locked by the caller and will be unlocked upon return.
2081 *
2082 * write_one_page() returns a negative error code if I/O failed.
2083 */
2084int write_one_page(struct page *page, int wait)
2085{
2086        struct address_space *mapping = page->mapping;
2087        int ret = 0;
2088        struct writeback_control wbc = {
2089                .sync_mode = WB_SYNC_ALL,
2090                .nr_to_write = 1,
2091        };
2092
2093        BUG_ON(!PageLocked(page));
2094
2095        if (wait)
2096                wait_on_page_writeback(page);
2097
2098        if (clear_page_dirty_for_io(page)) {
2099                page_cache_get(page);
2100                ret = mapping->a_ops->writepage(page, &wbc);
2101                if (ret == 0 && wait) {
2102                        wait_on_page_writeback(page);
2103                        if (PageError(page))
2104                                ret = -EIO;
2105                }
2106                page_cache_release(page);
2107        } else {
2108                unlock_page(page);
2109        }
2110        return ret;
2111}
2112EXPORT_SYMBOL(write_one_page);
2113
2114/*
2115 * For address_spaces which do not use buffers nor write back.
2116 */
2117int __set_page_dirty_no_writeback(struct page *page)
2118{
2119        if (!PageDirty(page))
2120                return !TestSetPageDirty(page);
2121        return 0;
2122}
2123
2124/*
2125 * Helper function for set_page_dirty family.
2126 * NOTE: This relies on being atomic wrt interrupts.
2127 */
2128void account_page_dirtied(struct page *page, struct address_space *mapping)
2129{
2130        trace_writeback_dirty_page(page, mapping);
2131
2132        if (mapping_cap_account_dirty(mapping)) {
2133                __inc_zone_page_state(page, NR_FILE_DIRTY);
2134                __inc_zone_page_state(page, NR_DIRTIED);
2135                __inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
2136                __inc_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
2137                task_io_account_write(PAGE_CACHE_SIZE);
2138                current->nr_dirtied++;
2139                this_cpu_inc(bdp_ratelimits);
2140        }
2141}
2142EXPORT_SYMBOL(account_page_dirtied);
2143
2144/*
2145 * Helper function for set_page_writeback family.
2146 *
2147 * The caller must hold mem_cgroup_begin/end_update_page_stat() lock
2148 * while calling this function.
2149 * See test_set_page_writeback for example.
2150 *
2151 * NOTE: Unlike account_page_dirtied this does not rely on being atomic
2152 * wrt interrupts.
2153 */
2154void account_page_writeback(struct page *page)
2155{
2156        mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_WRITEBACK);
2157        inc_zone_page_state(page, NR_WRITEBACK);
2158}
2159EXPORT_SYMBOL(account_page_writeback);
2160
2161/*
2162 * For address_spaces which do not use buffers.  Just tag the page as dirty in
2163 * its radix tree.
2164 *
2165 * This is also used when a single buffer is being dirtied: we want to set the
2166 * page dirty in that case, but not all the buffers.  This is a "bottom-up"
2167 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2168 *
2169 * Most callers have locked the page, which pins the address_space in memory.
2170 * But zap_pte_range() does not lock the page, however in that case the
2171 * mapping is pinned by the vma's ->vm_file reference.
2172 *
2173 * We take care to handle the case where the page was truncated from the
2174 * mapping by re-checking page_mapping() inside tree_lock.
2175 */
2176int __set_page_dirty_nobuffers(struct page *page)
2177{
2178        if (!TestSetPageDirty(page)) {
2179                struct address_space *mapping = page_mapping(page);
2180                struct address_space *mapping2;
2181
2182                if (!mapping)
2183                        return 1;
2184
2185                spin_lock_irq(&mapping->tree_lock);
2186                mapping2 = page_mapping(page);
2187                if (mapping2) { /* Race with truncate? */
2188                        BUG_ON(mapping2 != mapping);
2189                        WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
2190                        account_page_dirtied(page, mapping);
2191                        radix_tree_tag_set(&mapping->page_tree,
2192                                page_index(page), PAGECACHE_TAG_DIRTY);
2193                }
2194                spin_unlock_irq(&mapping->tree_lock);
2195                if (mapping->host) {
2196                        /* !PageAnon && !swapper_space */
2197                        __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
2198                }
2199                return 1;
2200        }
2201        return 0;
2202}
2203EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2204
2205/*
2206 * Call this whenever redirtying a page, to de-account the dirty counters
2207 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2208 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2209 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2210 * control.
2211 */
2212void account_page_redirty(struct page *page)
2213{
2214        struct address_space *mapping = page->mapping;
2215        if (mapping && mapping_cap_account_dirty(mapping)) {
2216                current->nr_dirtied--;
2217                dec_zone_page_state(page, NR_DIRTIED);
2218                dec_bdi_stat(mapping->backing_dev_info, BDI_DIRTIED);
2219        }
2220}
2221EXPORT_SYMBOL(account_page_redirty);
2222
2223/*
2224 * When a writepage implementation decides that it doesn't want to write this
2225 * page for some reason, it should redirty the locked page via
2226 * redirty_page_for_writepage() and it should then unlock the page and return 0
2227 */
2228int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2229{
2230        wbc->pages_skipped++;
2231        account_page_redirty(page);
2232        return __set_page_dirty_nobuffers(page);
2233}
2234EXPORT_SYMBOL(redirty_page_for_writepage);
2235
2236/*
2237 * Dirty a page.
2238 *
2239 * For pages with a mapping this should be done under the page lock
2240 * for the benefit of asynchronous memory errors who prefer a consistent
2241 * dirty state. This rule can be broken in some special cases,
2242 * but should be better not to.
2243 *
2244 * If the mapping doesn't provide a set_page_dirty a_op, then
2245 * just fall through and assume that it wants buffer_heads.
2246 */
2247int set_page_dirty(struct page *page)
2248{
2249        struct address_space *mapping = page_mapping(page);
2250
2251        if (likely(mapping)) {
2252                int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
2253                /*
2254                 * readahead/lru_deactivate_page could remain
2255                 * PG_readahead/PG_reclaim due to race with end_page_writeback
2256                 * About readahead, if the page is written, the flags would be
2257                 * reset. So no problem.
2258                 * About lru_deactivate_page, if the page is redirty, the flag
2259                 * will be reset. So no problem. but if the page is used by readahead
2260                 * it will confuse readahead and make it restart the size rampup
2261                 * process. But it's a trivial problem.
2262                 */
2263                ClearPageReclaim(page);
2264#ifdef CONFIG_BLOCK
2265                if (!spd)
2266                        spd = __set_page_dirty_buffers;
2267#endif
2268                return (*spd)(page);
2269        }
2270        if (!PageDirty(page)) {
2271                if (!TestSetPageDirty(page))
2272                        return 1;
2273        }
2274        return 0;
2275}
2276EXPORT_SYMBOL(set_page_dirty);
2277
2278/*
2279 * set_page_dirty() is racy if the caller has no reference against
2280 * page->mapping->host, and if the page is unlocked.  This is because another
2281 * CPU could truncate the page off the mapping and then free the mapping.
2282 *
2283 * Usually, the page _is_ locked, or the caller is a user-space process which
2284 * holds a reference on the inode by having an open file.
2285 *
2286 * In other cases, the page should be locked before running set_page_dirty().
2287 */
2288int set_page_dirty_lock(struct page *page)
2289{
2290        int ret;
2291
2292        lock_page(page);
2293        ret = set_page_dirty(page);
2294        unlock_page(page);
2295        return ret;
2296}
2297EXPORT_SYMBOL(set_page_dirty_lock);
2298
2299/*
2300 * Clear a page's dirty flag, while caring for dirty memory accounting.
2301 * Returns true if the page was previously dirty.
2302 *
2303 * This is for preparing to put the page under writeout.  We leave the page
2304 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2305 * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
2306 * implementation will run either set_page_writeback() or set_page_dirty(),
2307 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2308 * back into sync.
2309 *
2310 * This incoherency between the page's dirty flag and radix-tree tag is
2311 * unfortunate, but it only exists while the page is locked.
2312 */
2313int clear_page_dirty_for_io(struct page *page)
2314{
2315        struct address_space *mapping = page_mapping(page);
2316
2317        BUG_ON(!PageLocked(page));
2318
2319        if (mapping && mapping_cap_account_dirty(mapping)) {
2320                /*
2321                 * Yes, Virginia, this is indeed insane.
2322                 *
2323                 * We use this sequence to make sure that
2324                 *  (a) we account for dirty stats properly
2325                 *  (b) we tell the low-level filesystem to
2326                 *      mark the whole page dirty if it was
2327                 *      dirty in a pagetable. Only to then
2328                 *  (c) clean the page again and return 1 to
2329                 *      cause the writeback.
2330                 *
2331                 * This way we avoid all nasty races with the
2332                 * dirty bit in multiple places and clearing
2333                 * them concurrently from different threads.
2334                 *
2335                 * Note! Normally the "set_page_dirty(page)"
2336                 * has no effect on the actual dirty bit - since
2337                 * that will already usually be set. But we
2338                 * need the side effects, and it can help us
2339                 * avoid races.
2340                 *
2341                 * We basically use the page "master dirty bit"
2342                 * as a serialization point for all the different
2343                 * threads doing their things.
2344                 */
2345                if (page_mkclean(page))
2346                        set_page_dirty(page);
2347                /*
2348                 * We carefully synchronise fault handlers against
2349                 * installing a dirty pte and marking the page dirty
2350                 * at this point. We do this by having them hold the
2351                 * page lock at some point after installing their
2352                 * pte, but before marking the page dirty.
2353                 * Pages are always locked coming in here, so we get
2354                 * the desired exclusion. See mm/memory.c:do_wp_page()
2355                 * for more comments.
2356                 */
2357                if (TestClearPageDirty(page)) {
2358                        dec_zone_page_state(page, NR_FILE_DIRTY);
2359                        dec_bdi_stat(mapping->backing_dev_info,
2360                                        BDI_RECLAIMABLE);
2361                        return 1;
2362                }
2363                return 0;
2364        }
2365        return TestClearPageDirty(page);
2366}
2367EXPORT_SYMBOL(clear_page_dirty_for_io);
2368
2369int test_clear_page_writeback(struct page *page)
2370{
2371        struct address_space *mapping = page_mapping(page);
2372        int ret;
2373        bool locked;
2374        unsigned long memcg_flags;
2375
2376        mem_cgroup_begin_update_page_stat(page, &locked, &memcg_flags);
2377        if (mapping) {
2378                struct backing_dev_info *bdi = mapping->backing_dev_info;
2379                unsigned long flags;
2380
2381                spin_lock_irqsave(&mapping->tree_lock, flags);
2382                ret = TestClearPageWriteback(page);
2383                if (ret) {
2384                        radix_tree_tag_clear(&mapping->page_tree,
2385                                                page_index(page),
2386                                                PAGECACHE_TAG_WRITEBACK);
2387                        if (bdi_cap_account_writeback(bdi)) {
2388                                __dec_bdi_stat(bdi, BDI_WRITEBACK);
2389                                __bdi_writeout_inc(bdi);
2390                        }
2391                }
2392                spin_unlock_irqrestore(&mapping->tree_lock, flags);
2393        } else {
2394                ret = TestClearPageWriteback(page);
2395        }
2396        if (ret) {
2397                mem_cgroup_dec_page_stat(page, MEM_CGROUP_STAT_WRITEBACK);
2398                dec_zone_page_state(page, NR_WRITEBACK);
2399                inc_zone_page_state(page, NR_WRITTEN);
2400        }
2401        mem_cgroup_end_update_page_stat(page, &locked, &memcg_flags);
2402        return ret;
2403}
2404
2405int test_set_page_writeback(struct page *page)
2406{
2407        struct address_space *mapping = page_mapping(page);
2408        int ret;
2409        bool locked;
2410        unsigned long memcg_flags;
2411
2412        mem_cgroup_begin_update_page_stat(page, &locked, &memcg_flags);
2413        if (mapping) {
2414                struct backing_dev_info *bdi = mapping->backing_dev_info;
2415                unsigned long flags;
2416
2417                spin_lock_irqsave(&mapping->tree_lock, flags);
2418                ret = TestSetPageWriteback(page);
2419                if (!ret) {
2420                        radix_tree_tag_set(&mapping->page_tree,
2421                                                page_index(page),
2422                                                PAGECACHE_TAG_WRITEBACK);
2423                        if (bdi_cap_account_writeback(bdi))
2424                                __inc_bdi_stat(bdi, BDI_WRITEBACK);
2425                }
2426                if (!PageDirty(page))
2427                        radix_tree_tag_clear(&mapping->page_tree,
2428                                                page_index(page),
2429                                                PAGECACHE_TAG_DIRTY);
2430                radix_tree_tag_clear(&mapping->page_tree,
2431                                     page_index(page),
2432                                     PAGECACHE_TAG_TOWRITE);
2433                spin_unlock_irqrestore(&mapping->tree_lock, flags);
2434        } else {
2435                ret = TestSetPageWriteback(page);
2436        }
2437        if (!ret)
2438                account_page_writeback(page);
2439        mem_cgroup_end_update_page_stat(page, &locked, &memcg_flags);
2440        return ret;
2441
2442}
2443EXPORT_SYMBOL(test_set_page_writeback);
2444
2445/*
2446 * Return true if any of the pages in the mapping are marked with the
2447 * passed tag.
2448 */
2449int mapping_tagged(struct address_space *mapping, int tag)
2450{
2451        return radix_tree_tagged(&mapping->page_tree, tag);
2452}
2453EXPORT_SYMBOL(mapping_tagged);
2454
2455/**
2456 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2457 * @page:       The page to wait on.
2458 *
2459 * This function determines if the given page is related to a backing device
2460 * that requires page contents to be held stable during writeback.  If so, then
2461 * it will wait for any pending writeback to complete.
2462 */
2463void wait_for_stable_page(struct page *page)
2464{
2465        struct address_space *mapping = page_mapping(page);
2466        struct backing_dev_info *bdi = mapping->backing_dev_info;
2467
2468        if (!bdi_cap_stable_pages_required(bdi))
2469                return;
2470
2471        wait_on_page_writeback(page);
2472}
2473EXPORT_SYMBOL_GPL(wait_for_stable_page);
2474
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.