linux/fs/fs-writeback.c
<<
>>
Prefs
   1/*
   2 * fs/fs-writeback.c
   3 *
   4 * Copyright (C) 2002, Linus Torvalds.
   5 *
   6 * Contains all the functions related to writing back and waiting
   7 * upon dirty inodes against superblocks, and writing back dirty
   8 * pages against inodes.  ie: data writeback.  Writeout of the
   9 * inode itself is not handled here.
  10 *
  11 * 10Apr2002    Andrew Morton
  12 *              Split out of fs/inode.c
  13 *              Additions for address_space-based writeback
  14 */
  15
  16#include <linux/kernel.h>
  17#include <linux/module.h>
  18#include <linux/spinlock.h>
  19#include <linux/slab.h>
  20#include <linux/sched.h>
  21#include <linux/fs.h>
  22#include <linux/mm.h>
  23#include <linux/kthread.h>
  24#include <linux/freezer.h>
  25#include <linux/writeback.h>
  26#include <linux/blkdev.h>
  27#include <linux/backing-dev.h>
  28#include <linux/buffer_head.h>
  29#include <linux/tracepoint.h>
  30#include "internal.h"
  31
  32/*
  33 * Passed into wb_writeback(), essentially a subset of writeback_control
  34 */
  35struct wb_writeback_work {
  36        long nr_pages;
  37        struct super_block *sb;
  38        enum writeback_sync_modes sync_mode;
  39        unsigned int for_kupdate:1;
  40        unsigned int range_cyclic:1;
  41        unsigned int for_background:1;
  42
  43        struct list_head list;          /* pending work list */
  44        struct completion *done;        /* set if the caller waits */
  45};
  46
  47/*
  48 * Include the creation of the trace points after defining the
  49 * wb_writeback_work structure so that the definition remains local to this
  50 * file.
  51 */
  52#define CREATE_TRACE_POINTS
  53#include <trace/events/writeback.h>
  54
  55/*
  56 * We don't actually have pdflush, but this one is exported though /proc...
  57 */
  58int nr_pdflush_threads;
  59
  60/**
  61 * writeback_in_progress - determine whether there is writeback in progress
  62 * @bdi: the device's backing_dev_info structure.
  63 *
  64 * Determine whether there is writeback waiting to be handled against a
  65 * backing device.
  66 */
  67int writeback_in_progress(struct backing_dev_info *bdi)
  68{
  69        return test_bit(BDI_writeback_running, &bdi->state);
  70}
  71
  72static inline struct backing_dev_info *inode_to_bdi(struct inode *inode)
  73{
  74        struct super_block *sb = inode->i_sb;
  75
  76        if (strcmp(sb->s_type->name, "bdev") == 0)
  77                return inode->i_mapping->backing_dev_info;
  78
  79        return sb->s_bdi;
  80}
  81
  82static inline struct inode *wb_inode(struct list_head *head)
  83{
  84        return list_entry(head, struct inode, i_wb_list);
  85}
  86
  87/* Wakeup flusher thread or forker thread to fork it. Requires bdi->wb_lock. */
  88static void bdi_wakeup_flusher(struct backing_dev_info *bdi)
  89{
  90        if (bdi->wb.task) {
  91                wake_up_process(bdi->wb.task);
  92        } else {
  93                /*
  94                 * The bdi thread isn't there, wake up the forker thread which
  95                 * will create and run it.
  96                 */
  97                wake_up_process(default_backing_dev_info.wb.task);
  98        }
  99}
 100
 101static void bdi_queue_work(struct backing_dev_info *bdi,
 102                           struct wb_writeback_work *work)
 103{
 104        trace_writeback_queue(bdi, work);
 105
 106        spin_lock_bh(&bdi->wb_lock);
 107        list_add_tail(&work->list, &bdi->work_list);
 108        if (!bdi->wb.task)
 109                trace_writeback_nothread(bdi, work);
 110        bdi_wakeup_flusher(bdi);
 111        spin_unlock_bh(&bdi->wb_lock);
 112}
 113
 114static void
 115__bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
 116                      bool range_cyclic)
 117{
 118        struct wb_writeback_work *work;
 119
 120        /*
 121         * This is WB_SYNC_NONE writeback, so if allocation fails just
 122         * wakeup the thread for old dirty data writeback
 123         */
 124        work = kzalloc(sizeof(*work), GFP_ATOMIC);
 125        if (!work) {
 126                if (bdi->wb.task) {
 127                        trace_writeback_nowork(bdi);
 128                        wake_up_process(bdi->wb.task);
 129                }
 130                return;
 131        }
 132
 133        work->sync_mode = WB_SYNC_NONE;
 134        work->nr_pages  = nr_pages;
 135        work->range_cyclic = range_cyclic;
 136
 137        bdi_queue_work(bdi, work);
 138}
 139
 140/**
 141 * bdi_start_writeback - start writeback
 142 * @bdi: the backing device to write from
 143 * @nr_pages: the number of pages to write
 144 *
 145 * Description:
 146 *   This does WB_SYNC_NONE opportunistic writeback. The IO is only
 147 *   started when this function returns, we make no guarantees on
 148 *   completion. Caller need not hold sb s_umount semaphore.
 149 *
 150 */
 151void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages)
 152{
 153        __bdi_start_writeback(bdi, nr_pages, true);
 154}
 155
 156/**
 157 * bdi_start_background_writeback - start background writeback
 158 * @bdi: the backing device to write from
 159 *
 160 * Description:
 161 *   This makes sure WB_SYNC_NONE background writeback happens. When
 162 *   this function returns, it is only guaranteed that for given BDI
 163 *   some IO is happening if we are over background dirty threshold.
 164 *   Caller need not hold sb s_umount semaphore.
 165 */
 166void bdi_start_background_writeback(struct backing_dev_info *bdi)
 167{
 168        /*
 169         * We just wake up the flusher thread. It will perform background
 170         * writeback as soon as there is no other work to do.
 171         */
 172        trace_writeback_wake_background(bdi);
 173        spin_lock_bh(&bdi->wb_lock);
 174        bdi_wakeup_flusher(bdi);
 175        spin_unlock_bh(&bdi->wb_lock);
 176}
 177
 178/*
 179 * Remove the inode from the writeback list it is on.
 180 */
 181void inode_wb_list_del(struct inode *inode)
 182{
 183        spin_lock(&inode_wb_list_lock);
 184        list_del_init(&inode->i_wb_list);
 185        spin_unlock(&inode_wb_list_lock);
 186}
 187
 188
 189/*
 190 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
 191 * furthest end of its superblock's dirty-inode list.
 192 *
 193 * Before stamping the inode's ->dirtied_when, we check to see whether it is
 194 * already the most-recently-dirtied inode on the b_dirty list.  If that is
 195 * the case then the inode must have been redirtied while it was being written
 196 * out and we don't reset its dirtied_when.
 197 */
 198static void redirty_tail(struct inode *inode)
 199{
 200        struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
 201
 202        assert_spin_locked(&inode_wb_list_lock);
 203        if (!list_empty(&wb->b_dirty)) {
 204                struct inode *tail;
 205
 206                tail = wb_inode(wb->b_dirty.next);
 207                if (time_before(inode->dirtied_when, tail->dirtied_when))
 208                        inode->dirtied_when = jiffies;
 209        }
 210        list_move(&inode->i_wb_list, &wb->b_dirty);
 211}
 212
 213/*
 214 * requeue inode for re-scanning after bdi->b_io list is exhausted.
 215 */
 216static void requeue_io(struct inode *inode)
 217{
 218        struct bdi_writeback *wb = &inode_to_bdi(inode)->wb;
 219
 220        assert_spin_locked(&inode_wb_list_lock);
 221        list_move(&inode->i_wb_list, &wb->b_more_io);
 222}
 223
 224static void inode_sync_complete(struct inode *inode)
 225{
 226        /*
 227         * Prevent speculative execution through
 228         * spin_unlock(&inode_wb_list_lock);
 229         */
 230
 231        smp_mb();
 232        wake_up_bit(&inode->i_state, __I_SYNC);
 233}
 234
 235static bool inode_dirtied_after(struct inode *inode, unsigned long t)
 236{
 237        bool ret = time_after(inode->dirtied_when, t);
 238#ifndef CONFIG_64BIT
 239        /*
 240         * For inodes being constantly redirtied, dirtied_when can get stuck.
 241         * It _appears_ to be in the future, but is actually in distant past.
 242         * This test is necessary to prevent such wrapped-around relative times
 243         * from permanently stopping the whole bdi writeback.
 244         */
 245        ret = ret && time_before_eq(inode->dirtied_when, jiffies);
 246#endif
 247        return ret;
 248}
 249
 250/*
 251 * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
 252 */
 253static void move_expired_inodes(struct list_head *delaying_queue,
 254                               struct list_head *dispatch_queue,
 255                                unsigned long *older_than_this)
 256{
 257        LIST_HEAD(tmp);
 258        struct list_head *pos, *node;
 259        struct super_block *sb = NULL;
 260        struct inode *inode;
 261        int do_sb_sort = 0;
 262
 263        while (!list_empty(delaying_queue)) {
 264                inode = wb_inode(delaying_queue->prev);
 265                if (older_than_this &&
 266                    inode_dirtied_after(inode, *older_than_this))
 267                        break;
 268                if (sb && sb != inode->i_sb)
 269                        do_sb_sort = 1;
 270                sb = inode->i_sb;
 271                list_move(&inode->i_wb_list, &tmp);
 272        }
 273
 274        /* just one sb in list, splice to dispatch_queue and we're done */
 275        if (!do_sb_sort) {
 276                list_splice(&tmp, dispatch_queue);
 277                return;
 278        }
 279
 280        /* Move inodes from one superblock together */
 281        while (!list_empty(&tmp)) {
 282                sb = wb_inode(tmp.prev)->i_sb;
 283                list_for_each_prev_safe(pos, node, &tmp) {
 284                        inode = wb_inode(pos);
 285                        if (inode->i_sb == sb)
 286                                list_move(&inode->i_wb_list, dispatch_queue);
 287                }
 288        }
 289}
 290
 291/*
 292 * Queue all expired dirty inodes for io, eldest first.
 293 * Before
 294 *         newly dirtied     b_dirty    b_io    b_more_io
 295 *         =============>    gf         edc     BA
 296 * After
 297 *         newly dirtied     b_dirty    b_io    b_more_io
 298 *         =============>    g          fBAedc
 299 *                                           |
 300 *                                           +--> dequeue for IO
 301 */
 302static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
 303{
 304        assert_spin_locked(&inode_wb_list_lock);
 305        list_splice_init(&wb->b_more_io, &wb->b_io);
 306        move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
 307}
 308
 309static int write_inode(struct inode *inode, struct writeback_control *wbc)
 310{
 311        if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
 312                return inode->i_sb->s_op->write_inode(inode, wbc);
 313        return 0;
 314}
 315
 316/*
 317 * Wait for writeback on an inode to complete.
 318 */
 319static void inode_wait_for_writeback(struct inode *inode)
 320{
 321        DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
 322        wait_queue_head_t *wqh;
 323
 324        wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
 325        while (inode->i_state & I_SYNC) {
 326                spin_unlock(&inode->i_lock);
 327                spin_unlock(&inode_wb_list_lock);
 328                __wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
 329                spin_lock(&inode_wb_list_lock);
 330                spin_lock(&inode->i_lock);
 331        }
 332}
 333
 334/*
 335 * Write out an inode's dirty pages.  Called under inode_wb_list_lock and
 336 * inode->i_lock.  Either the caller has an active reference on the inode or
 337 * the inode has I_WILL_FREE set.
 338 *
 339 * If `wait' is set, wait on the writeout.
 340 *
 341 * The whole writeout design is quite complex and fragile.  We want to avoid
 342 * starvation of particular inodes when others are being redirtied, prevent
 343 * livelocks, etc.
 344 */
 345static int
 346writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
 347{
 348        struct address_space *mapping = inode->i_mapping;
 349        unsigned dirty;
 350        int ret;
 351
 352        assert_spin_locked(&inode_wb_list_lock);
 353        assert_spin_locked(&inode->i_lock);
 354
 355        if (!atomic_read(&inode->i_count))
 356                WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
 357        else
 358                WARN_ON(inode->i_state & I_WILL_FREE);
 359
 360        if (inode->i_state & I_SYNC) {
 361                /*
 362                 * If this inode is locked for writeback and we are not doing
 363                 * writeback-for-data-integrity, move it to b_more_io so that
 364                 * writeback can proceed with the other inodes on s_io.
 365                 *
 366                 * We'll have another go at writing back this inode when we
 367                 * completed a full scan of b_io.
 368                 */
 369                if (wbc->sync_mode != WB_SYNC_ALL) {
 370                        requeue_io(inode);
 371                        return 0;
 372                }
 373
 374                /*
 375                 * It's a data-integrity sync.  We must wait.
 376                 */
 377                inode_wait_for_writeback(inode);
 378        }
 379
 380        BUG_ON(inode->i_state & I_SYNC);
 381
 382        /* Set I_SYNC, reset I_DIRTY_PAGES */
 383        inode->i_state |= I_SYNC;
 384        inode->i_state &= ~I_DIRTY_PAGES;
 385        spin_unlock(&inode->i_lock);
 386        spin_unlock(&inode_wb_list_lock);
 387
 388        ret = do_writepages(mapping, wbc);
 389
 390        /*
 391         * Make sure to wait on the data before writing out the metadata.
 392         * This is important for filesystems that modify metadata on data
 393         * I/O completion.
 394         */
 395        if (wbc->sync_mode == WB_SYNC_ALL) {
 396                int err = filemap_fdatawait(mapping);
 397                if (ret == 0)
 398                        ret = err;
 399        }
 400
 401        /*
 402         * Some filesystems may redirty the inode during the writeback
 403         * due to delalloc, clear dirty metadata flags right before
 404         * write_inode()
 405         */
 406        spin_lock(&inode->i_lock);
 407        dirty = inode->i_state & I_DIRTY;
 408        inode->i_state &= ~(I_DIRTY_SYNC | I_DIRTY_DATASYNC);
 409        spin_unlock(&inode->i_lock);
 410        /* Don't write the inode if only I_DIRTY_PAGES was set */
 411        if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
 412                int err = write_inode(inode, wbc);
 413                if (ret == 0)
 414                        ret = err;
 415        }
 416
 417        spin_lock(&inode_wb_list_lock);
 418        spin_lock(&inode->i_lock);
 419        inode->i_state &= ~I_SYNC;
 420        if (!(inode->i_state & I_FREEING)) {
 421                if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
 422                        /*
 423                         * We didn't write back all the pages.  nfs_writepages()
 424                         * sometimes bales out without doing anything.
 425                         */
 426                        inode->i_state |= I_DIRTY_PAGES;
 427                        if (wbc->nr_to_write <= 0) {
 428                                /*
 429                                 * slice used up: queue for next turn
 430                                 */
 431                                requeue_io(inode);
 432                        } else {
 433                                /*
 434                                 * Writeback blocked by something other than
 435                                 * congestion. Delay the inode for some time to
 436                                 * avoid spinning on the CPU (100% iowait)
 437                                 * retrying writeback of the dirty page/inode
 438                                 * that cannot be performed immediately.
 439                                 */
 440                                redirty_tail(inode);
 441                        }
 442                } else if (inode->i_state & I_DIRTY) {
 443                        /*
 444                         * Filesystems can dirty the inode during writeback
 445                         * operations, such as delayed allocation during
 446                         * submission or metadata updates after data IO
 447                         * completion.
 448                         */
 449                        redirty_tail(inode);
 450                } else {
 451                        /*
 452                         * The inode is clean.  At this point we either have
 453                         * a reference to the inode or it's on it's way out.
 454                         * No need to add it back to the LRU.
 455                         */
 456                        list_del_init(&inode->i_wb_list);
 457                }
 458        }
 459        inode_sync_complete(inode);
 460        return ret;
 461}
 462
 463/*
 464 * For background writeback the caller does not have the sb pinned
 465 * before calling writeback. So make sure that we do pin it, so it doesn't
 466 * go away while we are writing inodes from it.
 467 */
 468static bool pin_sb_for_writeback(struct super_block *sb)
 469{
 470        spin_lock(&sb_lock);
 471        if (list_empty(&sb->s_instances)) {
 472                spin_unlock(&sb_lock);
 473                return false;
 474        }
 475
 476        sb->s_count++;
 477        spin_unlock(&sb_lock);
 478
 479        if (down_read_trylock(&sb->s_umount)) {
 480                if (sb->s_root)
 481                        return true;
 482                up_read(&sb->s_umount);
 483        }
 484
 485        put_super(sb);
 486        return false;
 487}
 488
 489/*
 490 * Write a portion of b_io inodes which belong to @sb.
 491 *
 492 * If @only_this_sb is true, then find and write all such
 493 * inodes. Otherwise write only ones which go sequentially
 494 * in reverse order.
 495 *
 496 * Return 1, if the caller writeback routine should be
 497 * interrupted. Otherwise return 0.
 498 */
 499static int writeback_sb_inodes(struct super_block *sb, struct bdi_writeback *wb,
 500                struct writeback_control *wbc, bool only_this_sb)
 501{
 502        while (!list_empty(&wb->b_io)) {
 503                long pages_skipped;
 504                struct inode *inode = wb_inode(wb->b_io.prev);
 505
 506                if (inode->i_sb != sb) {
 507                        if (only_this_sb) {
 508                                /*
 509                                 * We only want to write back data for this
 510                                 * superblock, move all inodes not belonging
 511                                 * to it back onto the dirty list.
 512                                 */
 513                                redirty_tail(inode);
 514                                continue;
 515                        }
 516
 517                        /*
 518                         * The inode belongs to a different superblock.
 519                         * Bounce back to the caller to unpin this and
 520                         * pin the next superblock.
 521                         */
 522                        return 0;
 523                }
 524
 525                /*
 526                 * Don't bother with new inodes or inodes beeing freed, first
 527                 * kind does not need peridic writeout yet, and for the latter
 528                 * kind writeout is handled by the freer.
 529                 */
 530                spin_lock(&inode->i_lock);
 531                if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
 532                        spin_unlock(&inode->i_lock);
 533                        requeue_io(inode);
 534                        continue;
 535                }
 536
 537                /*
 538                 * Was this inode dirtied after sync_sb_inodes was called?
 539                 * This keeps sync from extra jobs and livelock.
 540                 */
 541                if (inode_dirtied_after(inode, wbc->wb_start)) {
 542                        spin_unlock(&inode->i_lock);
 543                        return 1;
 544                }
 545
 546                __iget(inode);
 547
 548                pages_skipped = wbc->pages_skipped;
 549                writeback_single_inode(inode, wbc);
 550                if (wbc->pages_skipped != pages_skipped) {
 551                        /*
 552                         * writeback is not making progress due to locked
 553                         * buffers.  Skip this inode for now.
 554                         */
 555                        redirty_tail(inode);
 556                }
 557                spin_unlock(&inode->i_lock);
 558                spin_unlock(&inode_wb_list_lock);
 559                iput(inode);
 560                cond_resched();
 561                spin_lock(&inode_wb_list_lock);
 562                if (wbc->nr_to_write <= 0) {
 563                        wbc->more_io = 1;
 564                        return 1;
 565                }
 566                if (!list_empty(&wb->b_more_io))
 567                        wbc->more_io = 1;
 568        }
 569        /* b_io is empty */
 570        return 1;
 571}
 572
 573void writeback_inodes_wb(struct bdi_writeback *wb,
 574                struct writeback_control *wbc)
 575{
 576        int ret = 0;
 577
 578        if (!wbc->wb_start)
 579                wbc->wb_start = jiffies; /* livelock avoidance */
 580        spin_lock(&inode_wb_list_lock);
 581        if (!wbc->for_kupdate || list_empty(&wb->b_io))
 582                queue_io(wb, wbc->older_than_this);
 583
 584        while (!list_empty(&wb->b_io)) {
 585                struct inode *inode = wb_inode(wb->b_io.prev);
 586                struct super_block *sb = inode->i_sb;
 587
 588                if (!pin_sb_for_writeback(sb)) {
 589                        requeue_io(inode);
 590                        continue;
 591                }
 592                ret = writeback_sb_inodes(sb, wb, wbc, false);
 593                drop_super(sb);
 594
 595                if (ret)
 596                        break;
 597        }
 598        spin_unlock(&inode_wb_list_lock);
 599        /* Leave any unwritten inodes on b_io */
 600}
 601
 602static void __writeback_inodes_sb(struct super_block *sb,
 603                struct bdi_writeback *wb, struct writeback_control *wbc)
 604{
 605        WARN_ON(!rwsem_is_locked(&sb->s_umount));
 606
 607        spin_lock(&inode_wb_list_lock);
 608        if (!wbc->for_kupdate || list_empty(&wb->b_io))
 609                queue_io(wb, wbc->older_than_this);
 610        writeback_sb_inodes(sb, wb, wbc, true);
 611        spin_unlock(&inode_wb_list_lock);
 612}
 613
 614/*
 615 * The maximum number of pages to writeout in a single bdi flush/kupdate
 616 * operation.  We do this so we don't hold I_SYNC against an inode for
 617 * enormous amounts of time, which would block a userspace task which has
 618 * been forced to throttle against that inode.  Also, the code reevaluates
 619 * the dirty each time it has written this many pages.
 620 */
 621#define MAX_WRITEBACK_PAGES     1024
 622
 623static inline bool over_bground_thresh(void)
 624{
 625        unsigned long background_thresh, dirty_thresh;
 626
 627        global_dirty_limits(&background_thresh, &dirty_thresh);
 628
 629        return (global_page_state(NR_FILE_DIRTY) +
 630                global_page_state(NR_UNSTABLE_NFS) > background_thresh);
 631}
 632
 633/*
 634 * Explicit flushing or periodic writeback of "old" data.
 635 *
 636 * Define "old": the first time one of an inode's pages is dirtied, we mark the
 637 * dirtying-time in the inode's address_space.  So this periodic writeback code
 638 * just walks the superblock inode list, writing back any inodes which are
 639 * older than a specific point in time.
 640 *
 641 * Try to run once per dirty_writeback_interval.  But if a writeback event
 642 * takes longer than a dirty_writeback_interval interval, then leave a
 643 * one-second gap.
 644 *
 645 * older_than_this takes precedence over nr_to_write.  So we'll only write back
 646 * all dirty pages if they are all attached to "old" mappings.
 647 */
 648static long wb_writeback(struct bdi_writeback *wb,
 649                         struct wb_writeback_work *work)
 650{
 651        struct writeback_control wbc = {
 652                .sync_mode              = work->sync_mode,
 653                .older_than_this        = NULL,
 654                .for_kupdate            = work->for_kupdate,
 655                .for_background         = work->for_background,
 656                .range_cyclic           = work->range_cyclic,
 657        };
 658        unsigned long oldest_jif;
 659        long wrote = 0;
 660        long write_chunk;
 661        struct inode *inode;
 662
 663        if (wbc.for_kupdate) {
 664                wbc.older_than_this = &oldest_jif;
 665                oldest_jif = jiffies -
 666                                msecs_to_jiffies(dirty_expire_interval * 10);
 667        }
 668        if (!wbc.range_cyclic) {
 669                wbc.range_start = 0;
 670                wbc.range_end = LLONG_MAX;
 671        }
 672
 673        /*
 674         * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
 675         * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
 676         * here avoids calling into writeback_inodes_wb() more than once.
 677         *
 678         * The intended call sequence for WB_SYNC_ALL writeback is:
 679         *
 680         *      wb_writeback()
 681         *          __writeback_inodes_sb()     <== called only once
 682         *              write_cache_pages()     <== called once for each inode
 683         *                   (quickly) tag currently dirty pages
 684         *                   (maybe slowly) sync all tagged pages
 685         */
 686        if (wbc.sync_mode == WB_SYNC_NONE)
 687                write_chunk = MAX_WRITEBACK_PAGES;
 688        else
 689                write_chunk = LONG_MAX;
 690
 691        wbc.wb_start = jiffies; /* livelock avoidance */
 692        for (;;) {
 693                /*
 694                 * Stop writeback when nr_pages has been consumed
 695                 */
 696                if (work->nr_pages <= 0)
 697                        break;
 698
 699                /*
 700                 * Background writeout and kupdate-style writeback may
 701                 * run forever. Stop them if there is other work to do
 702                 * so that e.g. sync can proceed. They'll be restarted
 703                 * after the other works are all done.
 704                 */
 705                if ((work->for_background || work->for_kupdate) &&
 706                    !list_empty(&wb->bdi->work_list))
 707                        break;
 708
 709                /*
 710                 * For background writeout, stop when we are below the
 711                 * background dirty threshold
 712                 */
 713                if (work->for_background && !over_bground_thresh())
 714                        break;
 715
 716                wbc.more_io = 0;
 717                wbc.nr_to_write = write_chunk;
 718                wbc.pages_skipped = 0;
 719
 720                trace_wbc_writeback_start(&wbc, wb->bdi);
 721                if (work->sb)
 722                        __writeback_inodes_sb(work->sb, wb, &wbc);
 723                else
 724                        writeback_inodes_wb(wb, &wbc);
 725                trace_wbc_writeback_written(&wbc, wb->bdi);
 726
 727                work->nr_pages -= write_chunk - wbc.nr_to_write;
 728                wrote += write_chunk - wbc.nr_to_write;
 729
 730                /*
 731                 * If we consumed everything, see if we have more
 732                 */
 733                if (wbc.nr_to_write <= 0)
 734                        continue;
 735                /*
 736                 * Didn't write everything and we don't have more IO, bail
 737                 */
 738                if (!wbc.more_io)
 739                        break;
 740                /*
 741                 * Did we write something? Try for more
 742                 */
 743                if (wbc.nr_to_write < write_chunk)
 744                        continue;
 745                /*
 746                 * Nothing written. Wait for some inode to
 747                 * become available for writeback. Otherwise
 748                 * we'll just busyloop.
 749                 */
 750                spin_lock(&inode_wb_list_lock);
 751                if (!list_empty(&wb->b_more_io))  {
 752                        inode = wb_inode(wb->b_more_io.prev);
 753                        trace_wbc_writeback_wait(&wbc, wb->bdi);
 754                        spin_lock(&inode->i_lock);
 755                        inode_wait_for_writeback(inode);
 756                        spin_unlock(&inode->i_lock);
 757                }
 758                spin_unlock(&inode_wb_list_lock);
 759        }
 760
 761        return wrote;
 762}
 763
 764/*
 765 * Return the next wb_writeback_work struct that hasn't been processed yet.
 766 */
 767static struct wb_writeback_work *
 768get_next_work_item(struct backing_dev_info *bdi)
 769{
 770        struct wb_writeback_work *work = NULL;
 771
 772        spin_lock_bh(&bdi->wb_lock);
 773        if (!list_empty(&bdi->work_list)) {
 774                work = list_entry(bdi->work_list.next,
 775                                  struct wb_writeback_work, list);
 776                list_del_init(&work->list);
 777        }
 778        spin_unlock_bh(&bdi->wb_lock);
 779        return work;
 780}
 781
 782/*
 783 * Add in the number of potentially dirty inodes, because each inode
 784 * write can dirty pagecache in the underlying blockdev.
 785 */
 786static unsigned long get_nr_dirty_pages(void)
 787{
 788        return global_page_state(NR_FILE_DIRTY) +
 789                global_page_state(NR_UNSTABLE_NFS) +
 790                get_nr_dirty_inodes();
 791}
 792
 793static long wb_check_background_flush(struct bdi_writeback *wb)
 794{
 795        if (over_bground_thresh()) {
 796
 797                struct wb_writeback_work work = {
 798                        .nr_pages       = LONG_MAX,
 799                        .sync_mode      = WB_SYNC_NONE,
 800                        .for_background = 1,
 801                        .range_cyclic   = 1,
 802                };
 803
 804                return wb_writeback(wb, &work);
 805        }
 806
 807        return 0;
 808}
 809
 810static long wb_check_old_data_flush(struct bdi_writeback *wb)
 811{
 812        unsigned long expired;
 813        long nr_pages;
 814
 815        /*
 816         * When set to zero, disable periodic writeback
 817         */
 818        if (!dirty_writeback_interval)
 819                return 0;
 820
 821        expired = wb->last_old_flush +
 822                        msecs_to_jiffies(dirty_writeback_interval * 10);
 823        if (time_before(jiffies, expired))
 824                return 0;
 825
 826        wb->last_old_flush = jiffies;
 827        nr_pages = get_nr_dirty_pages();
 828
 829        if (nr_pages) {
 830                struct wb_writeback_work work = {
 831                        .nr_pages       = nr_pages,
 832                        .sync_mode      = WB_SYNC_NONE,
 833                        .for_kupdate    = 1,
 834                        .range_cyclic   = 1,
 835                };
 836
 837                return wb_writeback(wb, &work);
 838        }
 839
 840        return 0;
 841}
 842
 843/*
 844 * Retrieve work items and do the writeback they describe
 845 */
 846long wb_do_writeback(struct bdi_writeback *wb, int force_wait)
 847{
 848        struct backing_dev_info *bdi = wb->bdi;
 849        struct wb_writeback_work *work;
 850        long wrote = 0;
 851
 852        set_bit(BDI_writeback_running, &wb->bdi->state);
 853        while ((work = get_next_work_item(bdi)) != NULL) {
 854                /*
 855                 * Override sync mode, in case we must wait for completion
 856                 * because this thread is exiting now.
 857                 */
 858                if (force_wait)
 859                        work->sync_mode = WB_SYNC_ALL;
 860
 861                trace_writeback_exec(bdi, work);
 862
 863                wrote += wb_writeback(wb, work);
 864
 865                /*
 866                 * Notify the caller of completion if this is a synchronous
 867                 * work item, otherwise just free it.
 868                 */
 869                if (work->done)
 870                        complete(work->done);
 871                else
 872                        kfree(work);
 873        }
 874
 875        /*
 876         * Check for periodic writeback, kupdated() style
 877         */
 878        wrote += wb_check_old_data_flush(wb);
 879        wrote += wb_check_background_flush(wb);
 880        clear_bit(BDI_writeback_running, &wb->bdi->state);
 881
 882        return wrote;
 883}
 884
 885/*
 886 * Handle writeback of dirty data for the device backed by this bdi. Also
 887 * wakes up periodically and does kupdated style flushing.
 888 */
 889int bdi_writeback_thread(void *data)
 890{
 891        struct bdi_writeback *wb = data;
 892        struct backing_dev_info *bdi = wb->bdi;
 893        long pages_written;
 894
 895        current->flags |= PF_SWAPWRITE;
 896        set_freezable();
 897        wb->last_active = jiffies;
 898
 899        /*
 900         * Our parent may run at a different priority, just set us to normal
 901         */
 902        set_user_nice(current, 0);
 903
 904        trace_writeback_thread_start(bdi);
 905
 906        while (!kthread_should_stop()) {
 907                /*
 908                 * Remove own delayed wake-up timer, since we are already awake
 909                 * and we'll take care of the preriodic write-back.
 910                 */
 911                del_timer(&wb->wakeup_timer);
 912
 913                pages_written = wb_do_writeback(wb, 0);
 914
 915                trace_writeback_pages_written(pages_written);
 916
 917                if (pages_written)
 918                        wb->last_active = jiffies;
 919
 920                set_current_state(TASK_INTERRUPTIBLE);
 921                if (!list_empty(&bdi->work_list) || kthread_should_stop()) {
 922                        __set_current_state(TASK_RUNNING);
 923                        continue;
 924                }
 925
 926                if (wb_has_dirty_io(wb) && dirty_writeback_interval)
 927                        schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
 928                else {
 929                        /*
 930                         * We have nothing to do, so can go sleep without any
 931                         * timeout and save power. When a work is queued or
 932                         * something is made dirty - we will be woken up.
 933                         */
 934                        schedule();
 935                }
 936
 937                try_to_freeze();
 938        }
 939
 940        /* Flush any work that raced with us exiting */
 941        if (!list_empty(&bdi->work_list))
 942                wb_do_writeback(wb, 1);
 943
 944        trace_writeback_thread_stop(bdi);
 945        return 0;
 946}
 947
 948
 949/*
 950 * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
 951 * the whole world.
 952 */
 953void wakeup_flusher_threads(long nr_pages)
 954{
 955        struct backing_dev_info *bdi;
 956
 957        if (!nr_pages) {
 958                nr_pages = global_page_state(NR_FILE_DIRTY) +
 959                                global_page_state(NR_UNSTABLE_NFS);
 960        }
 961
 962        rcu_read_lock();
 963        list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
 964                if (!bdi_has_dirty_io(bdi))
 965                        continue;
 966                __bdi_start_writeback(bdi, nr_pages, false);
 967        }
 968        rcu_read_unlock();
 969}
 970
 971static noinline void block_dump___mark_inode_dirty(struct inode *inode)
 972{
 973        if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) {
 974                struct dentry *dentry;
 975                const char *name = "?";
 976
 977                dentry = d_find_alias(inode);
 978                if (dentry) {
 979                        spin_lock(&dentry->d_lock);
 980                        name = (const char *) dentry->d_name.name;
 981                }
 982                printk(KERN_DEBUG
 983                       "%s(%d): dirtied inode %lu (%s) on %s\n",
 984                       current->comm, task_pid_nr(current), inode->i_ino,
 985                       name, inode->i_sb->s_id);
 986                if (dentry) {
 987                        spin_unlock(&dentry->d_lock);
 988                        dput(dentry);
 989                }
 990        }
 991}
 992
 993/**
 994 *      __mark_inode_dirty -    internal function
 995 *      @inode: inode to mark
 996 *      @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
 997 *      Mark an inode as dirty. Callers should use mark_inode_dirty or
 998 *      mark_inode_dirty_sync.
 999 *
1000 * Put the inode on the super block's dirty list.
1001 *
1002 * CAREFUL! We mark it dirty unconditionally, but move it onto the
1003 * dirty list only if it is hashed or if it refers to a blockdev.
1004 * If it was not hashed, it will never be added to the dirty list
1005 * even if it is later hashed, as it will have been marked dirty already.
1006 *
1007 * In short, make sure you hash any inodes _before_ you start marking
1008 * them dirty.
1009 *
1010 * This function *must* be atomic for the I_DIRTY_PAGES case -
1011 * set_page_dirty() is called under spinlock in several places.
1012 *
1013 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
1014 * the block-special inode (/dev/hda1) itself.  And the ->dirtied_when field of
1015 * the kernel-internal blockdev inode represents the dirtying time of the
1016 * blockdev's pages.  This is why for I_DIRTY_PAGES we always use
1017 * page->mapping->host, so the page-dirtying time is recorded in the internal
1018 * blockdev inode.
1019 */
1020void __mark_inode_dirty(struct inode *inode, int flags)
1021{
1022        struct super_block *sb = inode->i_sb;
1023        struct backing_dev_info *bdi = NULL;
1024
1025        /*
1026         * Don't do this for I_DIRTY_PAGES - that doesn't actually
1027         * dirty the inode itself
1028         */
1029        if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
1030                if (sb->s_op->dirty_inode)
1031                        sb->s_op->dirty_inode(inode);
1032        }
1033
1034        /*
1035         * make sure that changes are seen by all cpus before we test i_state
1036         * -- mikulas
1037         */
1038        smp_mb();
1039
1040        /* avoid the locking if we can */
1041        if ((inode->i_state & flags) == flags)
1042                return;
1043
1044        if (unlikely(block_dump))
1045                block_dump___mark_inode_dirty(inode);
1046
1047        spin_lock(&inode->i_lock);
1048        if ((inode->i_state & flags) != flags) {
1049                const int was_dirty = inode->i_state & I_DIRTY;
1050
1051                inode->i_state |= flags;
1052
1053                /*
1054                 * If the inode is being synced, just update its dirty state.
1055                 * The unlocker will place the inode on the appropriate
1056                 * superblock list, based upon its state.
1057                 */
1058                if (inode->i_state & I_SYNC)
1059                        goto out_unlock_inode;
1060
1061                /*
1062                 * Only add valid (hashed) inodes to the superblock's
1063                 * dirty list.  Add blockdev inodes as well.
1064                 */
1065                if (!S_ISBLK(inode->i_mode)) {
1066                        if (inode_unhashed(inode))
1067                                goto out_unlock_inode;
1068                }
1069                if (inode->i_state & I_FREEING)
1070                        goto out_unlock_inode;
1071
1072                /*
1073                 * If the inode was already on b_dirty/b_io/b_more_io, don't
1074                 * reposition it (that would break b_dirty time-ordering).
1075                 */
1076                if (!was_dirty) {
1077                        bool wakeup_bdi = false;
1078                        bdi = inode_to_bdi(inode);
1079
1080                        if (bdi_cap_writeback_dirty(bdi)) {
1081                                WARN(!test_bit(BDI_registered, &bdi->state),
1082                                     "bdi-%s not registered\n", bdi->name);
1083
1084                                /*
1085                                 * If this is the first dirty inode for this
1086                                 * bdi, we have to wake-up the corresponding
1087                                 * bdi thread to make sure background
1088                                 * write-back happens later.
1089                                 */
1090                                if (!wb_has_dirty_io(&bdi->wb))
1091                                        wakeup_bdi = true;
1092                        }
1093
1094                        spin_unlock(&inode->i_lock);
1095                        spin_lock(&inode_wb_list_lock);
1096                        inode->dirtied_when = jiffies;
1097                        list_move(&inode->i_wb_list, &bdi->wb.b_dirty);
1098                        spin_unlock(&inode_wb_list_lock);
1099
1100                        if (wakeup_bdi)
1101                                bdi_wakeup_thread_delayed(bdi);
1102                        return;
1103                }
1104        }
1105out_unlock_inode:
1106        spin_unlock(&inode->i_lock);
1107
1108}
1109EXPORT_SYMBOL(__mark_inode_dirty);
1110
1111/*
1112 * Write out a superblock's list of dirty inodes.  A wait will be performed
1113 * upon no inodes, all inodes or the final one, depending upon sync_mode.
1114 *
1115 * If older_than_this is non-NULL, then only write out inodes which
1116 * had their first dirtying at a time earlier than *older_than_this.
1117 *
1118 * If `bdi' is non-zero then we're being asked to writeback a specific queue.
1119 * This function assumes that the blockdev superblock's inodes are backed by
1120 * a variety of queues, so all inodes are searched.  For other superblocks,
1121 * assume that all inodes are backed by the same queue.
1122 *
1123 * The inodes to be written are parked on bdi->b_io.  They are moved back onto
1124 * bdi->b_dirty as they are selected for writing.  This way, none can be missed
1125 * on the writer throttling path, and we get decent balancing between many
1126 * throttled threads: we don't want them all piling up on inode_sync_wait.
1127 */
1128static void wait_sb_inodes(struct super_block *sb)
1129{
1130        struct inode *inode, *old_inode = NULL;
1131
1132        /*
1133         * We need to be protected against the filesystem going from
1134         * r/o to r/w or vice versa.
1135         */
1136        WARN_ON(!rwsem_is_locked(&sb->s_umount));
1137
1138        spin_lock(&inode_sb_list_lock);
1139
1140        /*
1141         * Data integrity sync. Must wait for all pages under writeback,
1142         * because there may have been pages dirtied before our sync
1143         * call, but which had writeout started before we write it out.
1144         * In which case, the inode may not be on the dirty list, but
1145         * we still have to wait for that writeout.
1146         */
1147        list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
1148                struct address_space *mapping = inode->i_mapping;
1149
1150                spin_lock(&inode->i_lock);
1151                if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) ||
1152                    (mapping->nrpages == 0)) {
1153                        spin_unlock(&inode->i_lock);
1154                        continue;
1155                }
1156                __iget(inode);
1157                spin_unlock(&inode->i_lock);
1158                spin_unlock(&inode_sb_list_lock);
1159
1160                /*
1161                 * We hold a reference to 'inode' so it couldn't have been
1162                 * removed from s_inodes list while we dropped the
1163                 * inode_sb_list_lock.  We cannot iput the inode now as we can
1164                 * be holding the last reference and we cannot iput it under
1165                 * inode_sb_list_lock. So we keep the reference and iput it
1166                 * later.
1167                 */
1168                iput(old_inode);
1169                old_inode = inode;
1170
1171                filemap_fdatawait(mapping);
1172
1173                cond_resched();
1174
1175                spin_lock(&inode_sb_list_lock);
1176        }
1177        spin_unlock(&inode_sb_list_lock);
1178        iput(old_inode);
1179}
1180
1181/**
1182 * writeback_inodes_sb_nr -     writeback dirty inodes from given super_block
1183 * @sb: the superblock
1184 * @nr: the number of pages to write
1185 *
1186 * Start writeback on some inodes on this super_block. No guarantees are made
1187 * on how many (if any) will be written, and this function does not wait
1188 * for IO completion of submitted IO.
1189 */
1190void writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr)
1191{
1192        DECLARE_COMPLETION_ONSTACK(done);
1193        struct wb_writeback_work work = {
1194                .sb             = sb,
1195                .sync_mode      = WB_SYNC_NONE,
1196                .done           = &done,
1197                .nr_pages       = nr,
1198        };
1199
1200        WARN_ON(!rwsem_is_locked(&sb->s_umount));
1201        bdi_queue_work(sb->s_bdi, &work);
1202        wait_for_completion(&done);
1203}
1204EXPORT_SYMBOL(writeback_inodes_sb_nr);
1205
1206/**
1207 * writeback_inodes_sb  -       writeback dirty inodes from given super_block
1208 * @sb: the superblock
1209 *
1210 * Start writeback on some inodes on this super_block. No guarantees are made
1211 * on how many (if any) will be written, and this function does not wait
1212 * for IO completion of submitted IO.
1213 */
1214void writeback_inodes_sb(struct super_block *sb)
1215{
1216        return writeback_inodes_sb_nr(sb, get_nr_dirty_pages());
1217}
1218EXPORT_SYMBOL(writeback_inodes_sb);
1219
1220/**
1221 * writeback_inodes_sb_if_idle  -       start writeback if none underway
1222 * @sb: the superblock
1223 *
1224 * Invoke writeback_inodes_sb if no writeback is currently underway.
1225 * Returns 1 if writeback was started, 0 if not.
1226 */
1227int writeback_inodes_sb_if_idle(struct super_block *sb)
1228{
1229        if (!writeback_in_progress(sb->s_bdi)) {
1230                down_read(&sb->s_umount);
1231                writeback_inodes_sb(sb);
1232                up_read(&sb->s_umount);
1233                return 1;
1234        } else
1235                return 0;
1236}
1237EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
1238
1239/**
1240 * writeback_inodes_sb_if_idle  -       start writeback if none underway
1241 * @sb: the superblock
1242 * @nr: the number of pages to write
1243 *
1244 * Invoke writeback_inodes_sb if no writeback is currently underway.
1245 * Returns 1 if writeback was started, 0 if not.
1246 */
1247int writeback_inodes_sb_nr_if_idle(struct super_block *sb,
1248                                   unsigned long nr)
1249{
1250        if (!writeback_in_progress(sb->s_bdi)) {
1251                down_read(&sb->s_umount);
1252                writeback_inodes_sb_nr(sb, nr);
1253                up_read(&sb->s_umount);
1254                return 1;
1255        } else
1256                return 0;
1257}
1258EXPORT_SYMBOL(writeback_inodes_sb_nr_if_idle);
1259
1260/**
1261 * sync_inodes_sb       -       sync sb inode pages
1262 * @sb: the superblock
1263 *
1264 * This function writes and waits on any dirty inode belonging to this
1265 * super_block.
1266 */
1267void sync_inodes_sb(struct super_block *sb)
1268{
1269        DECLARE_COMPLETION_ONSTACK(done);
1270        struct wb_writeback_work work = {
1271                .sb             = sb,
1272                .sync_mode      = WB_SYNC_ALL,
1273                .nr_pages       = LONG_MAX,
1274                .range_cyclic   = 0,
1275                .done           = &done,
1276        };
1277
1278        WARN_ON(!rwsem_is_locked(&sb->s_umount));
1279
1280        bdi_queue_work(sb->s_bdi, &work);
1281        wait_for_completion(&done);
1282
1283        wait_sb_inodes(sb);
1284}
1285EXPORT_SYMBOL(sync_inodes_sb);
1286
1287/**
1288 * write_inode_now      -       write an inode to disk
1289 * @inode: inode to write to disk
1290 * @sync: whether the write should be synchronous or not
1291 *
1292 * This function commits an inode to disk immediately if it is dirty. This is
1293 * primarily needed by knfsd.
1294 *
1295 * The caller must either have a ref on the inode or must have set I_WILL_FREE.
1296 */
1297int write_inode_now(struct inode *inode, int sync)
1298{
1299        int ret;
1300        struct writeback_control wbc = {
1301                .nr_to_write = LONG_MAX,
1302                .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
1303                .range_start = 0,
1304                .range_end = LLONG_MAX,
1305        };
1306
1307        if (!mapping_cap_writeback_dirty(inode->i_mapping))
1308                wbc.nr_to_write = 0;
1309
1310        might_sleep();
1311        spin_lock(&inode_wb_list_lock);
1312        spin_lock(&inode->i_lock);
1313        ret = writeback_single_inode(inode, &wbc);
1314        spin_unlock(&inode->i_lock);
1315        spin_unlock(&inode_wb_list_lock);
1316        if (sync)
1317                inode_sync_wait(inode);
1318        return ret;
1319}
1320EXPORT_SYMBOL(write_inode_now);
1321
1322/**
1323 * sync_inode - write an inode and its pages to disk.
1324 * @inode: the inode to sync
1325 * @wbc: controls the writeback mode
1326 *
1327 * sync_inode() will write an inode and its pages to disk.  It will also
1328 * correctly update the inode on its superblock's dirty inode lists and will
1329 * update inode->i_state.
1330 *
1331 * The caller must have a ref on the inode.
1332 */
1333int sync_inode(struct inode *inode, struct writeback_control *wbc)
1334{
1335        int ret;
1336
1337        spin_lock(&inode_wb_list_lock);
1338        spin_lock(&inode->i_lock);
1339        ret = writeback_single_inode(inode, wbc);
1340        spin_unlock(&inode->i_lock);
1341        spin_unlock(&inode_wb_list_lock);
1342        return ret;
1343}
1344EXPORT_SYMBOL(sync_inode);
1345
1346/**
1347 * sync_inode_metadata - write an inode to disk
1348 * @inode: the inode to sync
1349 * @wait: wait for I/O to complete.
1350 *
1351 * Write an inode to disk and adjust its dirty state after completion.
1352 *
1353 * Note: only writes the actual inode, no associated data or other metadata.
1354 */
1355int sync_inode_metadata(struct inode *inode, int wait)
1356{
1357        struct writeback_control wbc = {
1358                .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
1359                .nr_to_write = 0, /* metadata-only */
1360        };
1361
1362        return sync_inode(inode, &wbc);
1363}
1364EXPORT_SYMBOL(sync_inode_metadata);
1365