linux/fs/buffer.c
<<
>>
Prefs
   1/*
   2 *  linux/fs/buffer.c
   3 *
   4 *  Copyright (C) 1991, 1992, 2002  Linus Torvalds
   5 */
   6
   7/*
   8 * Start bdflush() with kernel_thread not syscall - Paul Gortmaker, 12/95
   9 *
  10 * Removed a lot of unnecessary code and simplified things now that
  11 * the buffer cache isn't our primary cache - Andrew Tridgell 12/96
  12 *
  13 * Speed up hash, lru, and free list operations.  Use gfp() for allocating
  14 * hash table, use SLAB cache for buffer heads. SMP threading.  -DaveM
  15 *
  16 * Added 32k buffer block sizes - these are required older ARM systems. - RMK
  17 *
  18 * async buffer flushing, 1999 Andrea Arcangeli <andrea@suse.de>
  19 */
  20
  21#include <linux/kernel.h>
  22#include <linux/syscalls.h>
  23#include <linux/fs.h>
  24#include <linux/mm.h>
  25#include <linux/percpu.h>
  26#include <linux/slab.h>
  27#include <linux/capability.h>
  28#include <linux/blkdev.h>
  29#include <linux/file.h>
  30#include <linux/quotaops.h>
  31#include <linux/highmem.h>
  32#include <linux/export.h>
  33#include <linux/writeback.h>
  34#include <linux/hash.h>
  35#include <linux/suspend.h>
  36#include <linux/buffer_head.h>
  37#include <linux/task_io_accounting_ops.h>
  38#include <linux/bio.h>
  39#include <linux/notifier.h>
  40#include <linux/cpu.h>
  41#include <linux/bitops.h>
  42#include <linux/mpage.h>
  43#include <linux/bit_spinlock.h>
  44
  45static int fsync_buffers_list(spinlock_t *lock, struct list_head *list);
  46
  47#define BH_ENTRY(list) list_entry((list), struct buffer_head, b_assoc_buffers)
  48
  49inline void
  50init_buffer(struct buffer_head *bh, bh_end_io_t *handler, void *private)
  51{
  52        bh->b_end_io = handler;
  53        bh->b_private = private;
  54}
  55EXPORT_SYMBOL(init_buffer);
  56
  57static int sleep_on_buffer(void *word)
  58{
  59        io_schedule();
  60        return 0;
  61}
  62
  63void __lock_buffer(struct buffer_head *bh)
  64{
  65        wait_on_bit_lock(&bh->b_state, BH_Lock, sleep_on_buffer,
  66                                                        TASK_UNINTERRUPTIBLE);
  67}
  68EXPORT_SYMBOL(__lock_buffer);
  69
  70void unlock_buffer(struct buffer_head *bh)
  71{
  72        clear_bit_unlock(BH_Lock, &bh->b_state);
  73        smp_mb__after_clear_bit();
  74        wake_up_bit(&bh->b_state, BH_Lock);
  75}
  76EXPORT_SYMBOL(unlock_buffer);
  77
  78/*
  79 * Block until a buffer comes unlocked.  This doesn't stop it
  80 * from becoming locked again - you have to lock it yourself
  81 * if you want to preserve its state.
  82 */
  83void __wait_on_buffer(struct buffer_head * bh)
  84{
  85        wait_on_bit(&bh->b_state, BH_Lock, sleep_on_buffer, TASK_UNINTERRUPTIBLE);
  86}
  87EXPORT_SYMBOL(__wait_on_buffer);
  88
  89static void
  90__clear_page_buffers(struct page *page)
  91{
  92        ClearPagePrivate(page);
  93        set_page_private(page, 0);
  94        page_cache_release(page);
  95}
  96
  97
  98static int quiet_error(struct buffer_head *bh)
  99{
 100        if (!test_bit(BH_Quiet, &bh->b_state) && printk_ratelimit())
 101                return 0;
 102        return 1;
 103}
 104
 105
 106static void buffer_io_error(struct buffer_head *bh)
 107{
 108        char b[BDEVNAME_SIZE];
 109        printk(KERN_ERR "Buffer I/O error on device %s, logical block %Lu\n",
 110                        bdevname(bh->b_bdev, b),
 111                        (unsigned long long)bh->b_blocknr);
 112}
 113
 114/*
 115 * End-of-IO handler helper function which does not touch the bh after
 116 * unlocking it.
 117 * Note: unlock_buffer() sort-of does touch the bh after unlocking it, but
 118 * a race there is benign: unlock_buffer() only use the bh's address for
 119 * hashing after unlocking the buffer, so it doesn't actually touch the bh
 120 * itself.
 121 */
 122static void __end_buffer_read_notouch(struct buffer_head *bh, int uptodate)
 123{
 124        if (uptodate) {
 125                set_buffer_uptodate(bh);
 126        } else {
 127                /* This happens, due to failed READA attempts. */
 128                clear_buffer_uptodate(bh);
 129        }
 130        unlock_buffer(bh);
 131}
 132
 133/*
 134 * Default synchronous end-of-IO handler..  Just mark it up-to-date and
 135 * unlock the buffer. This is what ll_rw_block uses too.
 136 */
 137void end_buffer_read_sync(struct buffer_head *bh, int uptodate)
 138{
 139        __end_buffer_read_notouch(bh, uptodate);
 140        put_bh(bh);
 141}
 142EXPORT_SYMBOL(end_buffer_read_sync);
 143
 144void end_buffer_write_sync(struct buffer_head *bh, int uptodate)
 145{
 146        char b[BDEVNAME_SIZE];
 147
 148        if (uptodate) {
 149                set_buffer_uptodate(bh);
 150        } else {
 151                if (!quiet_error(bh)) {
 152                        buffer_io_error(bh);
 153                        printk(KERN_WARNING "lost page write due to "
 154                                        "I/O error on %s\n",
 155                                       bdevname(bh->b_bdev, b));
 156                }
 157                set_buffer_write_io_error(bh);
 158                clear_buffer_uptodate(bh);
 159        }
 160        unlock_buffer(bh);
 161        put_bh(bh);
 162}
 163EXPORT_SYMBOL(end_buffer_write_sync);
 164
 165/*
 166 * Various filesystems appear to want __find_get_block to be non-blocking.
 167 * But it's the page lock which protects the buffers.  To get around this,
 168 * we get exclusion from try_to_free_buffers with the blockdev mapping's
 169 * private_lock.
 170 *
 171 * Hack idea: for the blockdev mapping, i_bufferlist_lock contention
 172 * may be quite high.  This code could TryLock the page, and if that
 173 * succeeds, there is no need to take private_lock. (But if
 174 * private_lock is contended then so is mapping->tree_lock).
 175 */
 176static struct buffer_head *
 177__find_get_block_slow(struct block_device *bdev, sector_t block)
 178{
 179        struct inode *bd_inode = bdev->bd_inode;
 180        struct address_space *bd_mapping = bd_inode->i_mapping;
 181        struct buffer_head *ret = NULL;
 182        pgoff_t index;
 183        struct buffer_head *bh;
 184        struct buffer_head *head;
 185        struct page *page;
 186        int all_mapped = 1;
 187
 188        index = block >> (PAGE_CACHE_SHIFT - bd_inode->i_blkbits);
 189        page = find_get_page(bd_mapping, index);
 190        if (!page)
 191                goto out;
 192
 193        spin_lock(&bd_mapping->private_lock);
 194        if (!page_has_buffers(page))
 195                goto out_unlock;
 196        head = page_buffers(page);
 197        bh = head;
 198        do {
 199                if (!buffer_mapped(bh))
 200                        all_mapped = 0;
 201                else if (bh->b_blocknr == block) {
 202                        ret = bh;
 203                        get_bh(bh);
 204                        goto out_unlock;
 205                }
 206                bh = bh->b_this_page;
 207        } while (bh != head);
 208
 209        /* we might be here because some of the buffers on this page are
 210         * not mapped.  This is due to various races between
 211         * file io on the block device and getblk.  It gets dealt with
 212         * elsewhere, don't buffer_error if we had some unmapped buffers
 213         */
 214        if (all_mapped) {
 215                char b[BDEVNAME_SIZE];
 216
 217                printk("__find_get_block_slow() failed. "
 218                        "block=%llu, b_blocknr=%llu\n",
 219                        (unsigned long long)block,
 220                        (unsigned long long)bh->b_blocknr);
 221                printk("b_state=0x%08lx, b_size=%zu\n",
 222                        bh->b_state, bh->b_size);
 223                printk("device %s blocksize: %d\n", bdevname(bdev, b),
 224                        1 << bd_inode->i_blkbits);
 225        }
 226out_unlock:
 227        spin_unlock(&bd_mapping->private_lock);
 228        page_cache_release(page);
 229out:
 230        return ret;
 231}
 232
 233/*
 234 * Kick the writeback threads then try to free up some ZONE_NORMAL memory.
 235 */
 236static void free_more_memory(void)
 237{
 238        struct zone *zone;
 239        int nid;
 240
 241        wakeup_flusher_threads(1024, WB_REASON_FREE_MORE_MEM);
 242        yield();
 243
 244        for_each_online_node(nid) {
 245                (void)first_zones_zonelist(node_zonelist(nid, GFP_NOFS),
 246                                                gfp_zone(GFP_NOFS), NULL,
 247                                                &zone);
 248                if (zone)
 249                        try_to_free_pages(node_zonelist(nid, GFP_NOFS), 0,
 250                                                GFP_NOFS, NULL);
 251        }
 252}
 253
 254/*
 255 * I/O completion handler for block_read_full_page() - pages
 256 * which come unlocked at the end of I/O.
 257 */
 258static void end_buffer_async_read(struct buffer_head *bh, int uptodate)
 259{
 260        unsigned long flags;
 261        struct buffer_head *first;
 262        struct buffer_head *tmp;
 263        struct page *page;
 264        int page_uptodate = 1;
 265
 266        BUG_ON(!buffer_async_read(bh));
 267
 268        page = bh->b_page;
 269        if (uptodate) {
 270                set_buffer_uptodate(bh);
 271        } else {
 272                clear_buffer_uptodate(bh);
 273                if (!quiet_error(bh))
 274                        buffer_io_error(bh);
 275                SetPageError(page);
 276        }
 277
 278        /*
 279         * Be _very_ careful from here on. Bad things can happen if
 280         * two buffer heads end IO at almost the same time and both
 281         * decide that the page is now completely done.
 282         */
 283        first = page_buffers(page);
 284        local_irq_save(flags);
 285        bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
 286        clear_buffer_async_read(bh);
 287        unlock_buffer(bh);
 288        tmp = bh;
 289        do {
 290                if (!buffer_uptodate(tmp))
 291                        page_uptodate = 0;
 292                if (buffer_async_read(tmp)) {
 293                        BUG_ON(!buffer_locked(tmp));
 294                        goto still_busy;
 295                }
 296                tmp = tmp->b_this_page;
 297        } while (tmp != bh);
 298        bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
 299        local_irq_restore(flags);
 300
 301        /*
 302         * If none of the buffers had errors and they are all
 303         * uptodate then we can set the page uptodate.
 304         */
 305        if (page_uptodate && !PageError(page))
 306                SetPageUptodate(page);
 307        unlock_page(page);
 308        return;
 309
 310still_busy:
 311        bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
 312        local_irq_restore(flags);
 313        return;
 314}
 315
 316/*
 317 * Completion handler for block_write_full_page() - pages which are unlocked
 318 * during I/O, and which have PageWriteback cleared upon I/O completion.
 319 */
 320void end_buffer_async_write(struct buffer_head *bh, int uptodate)
 321{
 322        char b[BDEVNAME_SIZE];
 323        unsigned long flags;
 324        struct buffer_head *first;
 325        struct buffer_head *tmp;
 326        struct page *page;
 327
 328        BUG_ON(!buffer_async_write(bh));
 329
 330        page = bh->b_page;
 331        if (uptodate) {
 332                set_buffer_uptodate(bh);
 333        } else {
 334                if (!quiet_error(bh)) {
 335                        buffer_io_error(bh);
 336                        printk(KERN_WARNING "lost page write due to "
 337                                        "I/O error on %s\n",
 338                               bdevname(bh->b_bdev, b));
 339                }
 340                set_bit(AS_EIO, &page->mapping->flags);
 341                set_buffer_write_io_error(bh);
 342                clear_buffer_uptodate(bh);
 343                SetPageError(page);
 344        }
 345
 346        first = page_buffers(page);
 347        local_irq_save(flags);
 348        bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
 349
 350        clear_buffer_async_write(bh);
 351        unlock_buffer(bh);
 352        tmp = bh->b_this_page;
 353        while (tmp != bh) {
 354                if (buffer_async_write(tmp)) {
 355                        BUG_ON(!buffer_locked(tmp));
 356                        goto still_busy;
 357                }
 358                tmp = tmp->b_this_page;
 359        }
 360        bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
 361        local_irq_restore(flags);
 362        end_page_writeback(page);
 363        return;
 364
 365still_busy:
 366        bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
 367        local_irq_restore(flags);
 368        return;
 369}
 370EXPORT_SYMBOL(end_buffer_async_write);
 371
 372/*
 373 * If a page's buffers are under async readin (end_buffer_async_read
 374 * completion) then there is a possibility that another thread of
 375 * control could lock one of the buffers after it has completed
 376 * but while some of the other buffers have not completed.  This
 377 * locked buffer would confuse end_buffer_async_read() into not unlocking
 378 * the page.  So the absence of BH_Async_Read tells end_buffer_async_read()
 379 * that this buffer is not under async I/O.
 380 *
 381 * The page comes unlocked when it has no locked buffer_async buffers
 382 * left.
 383 *
 384 * PageLocked prevents anyone starting new async I/O reads any of
 385 * the buffers.
 386 *
 387 * PageWriteback is used to prevent simultaneous writeout of the same
 388 * page.
 389 *
 390 * PageLocked prevents anyone from starting writeback of a page which is
 391 * under read I/O (PageWriteback is only ever set against a locked page).
 392 */
 393static void mark_buffer_async_read(struct buffer_head *bh)
 394{
 395        bh->b_end_io = end_buffer_async_read;
 396        set_buffer_async_read(bh);
 397}
 398
 399static void mark_buffer_async_write_endio(struct buffer_head *bh,
 400                                          bh_end_io_t *handler)
 401{
 402        bh->b_end_io = handler;
 403        set_buffer_async_write(bh);
 404}
 405
 406void mark_buffer_async_write(struct buffer_head *bh)
 407{
 408        mark_buffer_async_write_endio(bh, end_buffer_async_write);
 409}
 410EXPORT_SYMBOL(mark_buffer_async_write);
 411
 412
 413/*
 414 * fs/buffer.c contains helper functions for buffer-backed address space's
 415 * fsync functions.  A common requirement for buffer-based filesystems is
 416 * that certain data from the backing blockdev needs to be written out for
 417 * a successful fsync().  For example, ext2 indirect blocks need to be
 418 * written back and waited upon before fsync() returns.
 419 *
 420 * The functions mark_buffer_inode_dirty(), fsync_inode_buffers(),
 421 * inode_has_buffers() and invalidate_inode_buffers() are provided for the
 422 * management of a list of dependent buffers at ->i_mapping->private_list.
 423 *
 424 * Locking is a little subtle: try_to_free_buffers() will remove buffers
 425 * from their controlling inode's queue when they are being freed.  But
 426 * try_to_free_buffers() will be operating against the *blockdev* mapping
 427 * at the time, not against the S_ISREG file which depends on those buffers.
 428 * So the locking for private_list is via the private_lock in the address_space
 429 * which backs the buffers.  Which is different from the address_space 
 430 * against which the buffers are listed.  So for a particular address_space,
 431 * mapping->private_lock does *not* protect mapping->private_list!  In fact,
 432 * mapping->private_list will always be protected by the backing blockdev's
 433 * ->private_lock.
 434 *
 435 * Which introduces a requirement: all buffers on an address_space's
 436 * ->private_list must be from the same address_space: the blockdev's.
 437 *
 438 * address_spaces which do not place buffers at ->private_list via these
 439 * utility functions are free to use private_lock and private_list for
 440 * whatever they want.  The only requirement is that list_empty(private_list)
 441 * be true at clear_inode() time.
 442 *
 443 * FIXME: clear_inode should not call invalidate_inode_buffers().  The
 444 * filesystems should do that.  invalidate_inode_buffers() should just go
 445 * BUG_ON(!list_empty).
 446 *
 447 * FIXME: mark_buffer_dirty_inode() is a data-plane operation.  It should
 448 * take an address_space, not an inode.  And it should be called
 449 * mark_buffer_dirty_fsync() to clearly define why those buffers are being
 450 * queued up.
 451 *
 452 * FIXME: mark_buffer_dirty_inode() doesn't need to add the buffer to the
 453 * list if it is already on a list.  Because if the buffer is on a list,
 454 * it *must* already be on the right one.  If not, the filesystem is being
 455 * silly.  This will save a ton of locking.  But first we have to ensure
 456 * that buffers are taken *off* the old inode's list when they are freed
 457 * (presumably in truncate).  That requires careful auditing of all
 458 * filesystems (do it inside bforget()).  It could also be done by bringing
 459 * b_inode back.
 460 */
 461
 462/*
 463 * The buffer's backing address_space's private_lock must be held
 464 */
 465static void __remove_assoc_queue(struct buffer_head *bh)
 466{
 467        list_del_init(&bh->b_assoc_buffers);
 468        WARN_ON(!bh->b_assoc_map);
 469        if (buffer_write_io_error(bh))
 470                set_bit(AS_EIO, &bh->b_assoc_map->flags);
 471        bh->b_assoc_map = NULL;
 472}
 473
 474int inode_has_buffers(struct inode *inode)
 475{
 476        return !list_empty(&inode->i_data.private_list);
 477}
 478
 479/*
 480 * osync is designed to support O_SYNC io.  It waits synchronously for
 481 * all already-submitted IO to complete, but does not queue any new
 482 * writes to the disk.
 483 *
 484 * To do O_SYNC writes, just queue the buffer writes with ll_rw_block as
 485 * you dirty the buffers, and then use osync_inode_buffers to wait for
 486 * completion.  Any other dirty buffers which are not yet queued for
 487 * write will not be flushed to disk by the osync.
 488 */
 489static int osync_buffers_list(spinlock_t *lock, struct list_head *list)
 490{
 491        struct buffer_head *bh;
 492        struct list_head *p;
 493        int err = 0;
 494
 495        spin_lock(lock);
 496repeat:
 497        list_for_each_prev(p, list) {
 498                bh = BH_ENTRY(p);
 499                if (buffer_locked(bh)) {
 500                        get_bh(bh);
 501                        spin_unlock(lock);
 502                        wait_on_buffer(bh);
 503                        if (!buffer_uptodate(bh))
 504                                err = -EIO;
 505                        brelse(bh);
 506                        spin_lock(lock);
 507                        goto repeat;
 508                }
 509        }
 510        spin_unlock(lock);
 511        return err;
 512}
 513
 514static void do_thaw_one(struct super_block *sb, void *unused)
 515{
 516        char b[BDEVNAME_SIZE];
 517        while (sb->s_bdev && !thaw_bdev(sb->s_bdev, sb))
 518                printk(KERN_WARNING "Emergency Thaw on %s\n",
 519                       bdevname(sb->s_bdev, b));
 520}
 521
 522static void do_thaw_all(struct work_struct *work)
 523{
 524        iterate_supers(do_thaw_one, NULL);
 525        kfree(work);
 526        printk(KERN_WARNING "Emergency Thaw complete\n");
 527}
 528
 529/**
 530 * emergency_thaw_all -- forcibly thaw every frozen filesystem
 531 *
 532 * Used for emergency unfreeze of all filesystems via SysRq
 533 */
 534void emergency_thaw_all(void)
 535{
 536        struct work_struct *work;
 537
 538        work = kmalloc(sizeof(*work), GFP_ATOMIC);
 539        if (work) {
 540                INIT_WORK(work, do_thaw_all);
 541                schedule_work(work);
 542        }
 543}
 544
 545/**
 546 * sync_mapping_buffers - write out & wait upon a mapping's "associated" buffers
 547 * @mapping: the mapping which wants those buffers written
 548 *
 549 * Starts I/O against the buffers at mapping->private_list, and waits upon
 550 * that I/O.
 551 *
 552 * Basically, this is a convenience function for fsync().
 553 * @mapping is a file or directory which needs those buffers to be written for
 554 * a successful fsync().
 555 */
 556int sync_mapping_buffers(struct address_space *mapping)
 557{
 558        struct address_space *buffer_mapping = mapping->assoc_mapping;
 559
 560        if (buffer_mapping == NULL || list_empty(&mapping->private_list))
 561                return 0;
 562
 563        return fsync_buffers_list(&buffer_mapping->private_lock,
 564                                        &mapping->private_list);
 565}
 566EXPORT_SYMBOL(sync_mapping_buffers);
 567
 568/*
 569 * Called when we've recently written block `bblock', and it is known that
 570 * `bblock' was for a buffer_boundary() buffer.  This means that the block at
 571 * `bblock + 1' is probably a dirty indirect block.  Hunt it down and, if it's
 572 * dirty, schedule it for IO.  So that indirects merge nicely with their data.
 573 */
 574void write_boundary_block(struct block_device *bdev,
 575                        sector_t bblock, unsigned blocksize)
 576{
 577        struct buffer_head *bh = __find_get_block(bdev, bblock + 1, blocksize);
 578        if (bh) {
 579                if (buffer_dirty(bh))
 580                        ll_rw_block(WRITE, 1, &bh);
 581                put_bh(bh);
 582        }
 583}
 584
 585void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
 586{
 587        struct address_space *mapping = inode->i_mapping;
 588        struct address_space *buffer_mapping = bh->b_page->mapping;
 589
 590        mark_buffer_dirty(bh);
 591        if (!mapping->assoc_mapping) {
 592                mapping->assoc_mapping = buffer_mapping;
 593        } else {
 594                BUG_ON(mapping->assoc_mapping != buffer_mapping);
 595        }
 596        if (!bh->b_assoc_map) {
 597                spin_lock(&buffer_mapping->private_lock);
 598                list_move_tail(&bh->b_assoc_buffers,
 599                                &mapping->private_list);
 600                bh->b_assoc_map = mapping;
 601                spin_unlock(&buffer_mapping->private_lock);
 602        }
 603}
 604EXPORT_SYMBOL(mark_buffer_dirty_inode);
 605
 606/*
 607 * Mark the page dirty, and set it dirty in the radix tree, and mark the inode
 608 * dirty.
 609 *
 610 * If warn is true, then emit a warning if the page is not uptodate and has
 611 * not been truncated.
 612 */
 613static void __set_page_dirty(struct page *page,
 614                struct address_space *mapping, int warn)
 615{
 616        spin_lock_irq(&mapping->tree_lock);
 617        if (page->mapping) {    /* Race with truncate? */
 618                WARN_ON_ONCE(warn && !PageUptodate(page));
 619                account_page_dirtied(page, mapping);
 620                radix_tree_tag_set(&mapping->page_tree,
 621                                page_index(page), PAGECACHE_TAG_DIRTY);
 622        }
 623        spin_unlock_irq(&mapping->tree_lock);
 624        __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
 625}
 626
 627/*
 628 * Add a page to the dirty page list.
 629 *
 630 * It is a sad fact of life that this function is called from several places
 631 * deeply under spinlocking.  It may not sleep.
 632 *
 633 * If the page has buffers, the uptodate buffers are set dirty, to preserve
 634 * dirty-state coherency between the page and the buffers.  It the page does
 635 * not have buffers then when they are later attached they will all be set
 636 * dirty.
 637 *
 638 * The buffers are dirtied before the page is dirtied.  There's a small race
 639 * window in which a writepage caller may see the page cleanness but not the
 640 * buffer dirtiness.  That's fine.  If this code were to set the page dirty
 641 * before the buffers, a concurrent writepage caller could clear the page dirty
 642 * bit, see a bunch of clean buffers and we'd end up with dirty buffers/clean
 643 * page on the dirty page list.
 644 *
 645 * We use private_lock to lock against try_to_free_buffers while using the
 646 * page's buffer list.  Also use this to protect against clean buffers being
 647 * added to the page after it was set dirty.
 648 *
 649 * FIXME: may need to call ->reservepage here as well.  That's rather up to the
 650 * address_space though.
 651 */
 652int __set_page_dirty_buffers(struct page *page)
 653{
 654        int newly_dirty;
 655        struct address_space *mapping = page_mapping(page);
 656
 657        if (unlikely(!mapping))
 658                return !TestSetPageDirty(page);
 659
 660        spin_lock(&mapping->private_lock);
 661        if (page_has_buffers(page)) {
 662                struct buffer_head *head = page_buffers(page);
 663                struct buffer_head *bh = head;
 664
 665                do {
 666                        set_buffer_dirty(bh);
 667                        bh = bh->b_this_page;
 668                } while (bh != head);
 669        }
 670        newly_dirty = !TestSetPageDirty(page);
 671        spin_unlock(&mapping->private_lock);
 672
 673        if (newly_dirty)
 674                __set_page_dirty(page, mapping, 1);
 675        return newly_dirty;
 676}
 677EXPORT_SYMBOL(__set_page_dirty_buffers);
 678
 679/*
 680 * Write out and wait upon a list of buffers.
 681 *
 682 * We have conflicting pressures: we want to make sure that all
 683 * initially dirty buffers get waited on, but that any subsequently
 684 * dirtied buffers don't.  After all, we don't want fsync to last
 685 * forever if somebody is actively writing to the file.
 686 *
 687 * Do this in two main stages: first we copy dirty buffers to a
 688 * temporary inode list, queueing the writes as we go.  Then we clean
 689 * up, waiting for those writes to complete.
 690 * 
 691 * During this second stage, any subsequent updates to the file may end
 692 * up refiling the buffer on the original inode's dirty list again, so
 693 * there is a chance we will end up with a buffer queued for write but
 694 * not yet completed on that list.  So, as a final cleanup we go through
 695 * the osync code to catch these locked, dirty buffers without requeuing
 696 * any newly dirty buffers for write.
 697 */
 698static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
 699{
 700        struct buffer_head *bh;
 701        struct list_head tmp;
 702        struct address_space *mapping;
 703        int err = 0, err2;
 704        struct blk_plug plug;
 705
 706        INIT_LIST_HEAD(&tmp);
 707        blk_start_plug(&plug);
 708
 709        spin_lock(lock);
 710        while (!list_empty(list)) {
 711                bh = BH_ENTRY(list->next);
 712                mapping = bh->b_assoc_map;
 713                __remove_assoc_queue(bh);
 714                /* Avoid race with mark_buffer_dirty_inode() which does
 715                 * a lockless check and we rely on seeing the dirty bit */
 716                smp_mb();
 717                if (buffer_dirty(bh) || buffer_locked(bh)) {
 718                        list_add(&bh->b_assoc_buffers, &tmp);
 719                        bh->b_assoc_map = mapping;
 720                        if (buffer_dirty(bh)) {
 721                                get_bh(bh);
 722                                spin_unlock(lock);
 723                                /*
 724                                 * Ensure any pending I/O completes so that
 725                                 * write_dirty_buffer() actually writes the
 726                                 * current contents - it is a noop if I/O is
 727                                 * still in flight on potentially older
 728                                 * contents.
 729                                 */
 730                                write_dirty_buffer(bh, WRITE_SYNC);
 731
 732                                /*
 733                                 * Kick off IO for the previous mapping. Note
 734                                 * that we will not run the very last mapping,
 735                                 * wait_on_buffer() will do that for us
 736                                 * through sync_buffer().
 737                                 */
 738                                brelse(bh);
 739                                spin_lock(lock);
 740                        }
 741                }
 742        }
 743
 744        spin_unlock(lock);
 745        blk_finish_plug(&plug);
 746        spin_lock(lock);
 747
 748        while (!list_empty(&tmp)) {
 749                bh = BH_ENTRY(tmp.prev);
 750                get_bh(bh);
 751                mapping = bh->b_assoc_map;
 752                __remove_assoc_queue(bh);
 753                /* Avoid race with mark_buffer_dirty_inode() which does
 754                 * a lockless check and we rely on seeing the dirty bit */
 755                smp_mb();
 756                if (buffer_dirty(bh)) {
 757                        list_add(&bh->b_assoc_buffers,
 758                                 &mapping->private_list);
 759                        bh->b_assoc_map = mapping;
 760                }
 761                spin_unlock(lock);
 762                wait_on_buffer(bh);
 763                if (!buffer_uptodate(bh))
 764                        err = -EIO;
 765                brelse(bh);
 766                spin_lock(lock);
 767        }
 768        
 769        spin_unlock(lock);
 770        err2 = osync_buffers_list(lock, list);
 771        if (err)
 772                return err;
 773        else
 774                return err2;
 775}
 776
 777/*
 778 * Invalidate any and all dirty buffers on a given inode.  We are
 779 * probably unmounting the fs, but that doesn't mean we have already
 780 * done a sync().  Just drop the buffers from the inode list.
 781 *
 782 * NOTE: we take the inode's blockdev's mapping's private_lock.  Which
 783 * assumes that all the buffers are against the blockdev.  Not true
 784 * for reiserfs.
 785 */
 786void invalidate_inode_buffers(struct inode *inode)
 787{
 788        if (inode_has_buffers(inode)) {
 789                struct address_space *mapping = &inode->i_data;
 790                struct list_head *list = &mapping->private_list;
 791                struct address_space *buffer_mapping = mapping->assoc_mapping;
 792
 793                spin_lock(&buffer_mapping->private_lock);
 794                while (!list_empty(list))
 795                        __remove_assoc_queue(BH_ENTRY(list->next));
 796                spin_unlock(&buffer_mapping->private_lock);
 797        }
 798}
 799EXPORT_SYMBOL(invalidate_inode_buffers);
 800
 801/*
 802 * Remove any clean buffers from the inode's buffer list.  This is called
 803 * when we're trying to free the inode itself.  Those buffers can pin it.
 804 *
 805 * Returns true if all buffers were removed.
 806 */
 807int remove_inode_buffers(struct inode *inode)
 808{
 809        int ret = 1;
 810
 811        if (inode_has_buffers(inode)) {
 812                struct address_space *mapping = &inode->i_data;
 813                struct list_head *list = &mapping->private_list;
 814                struct address_space *buffer_mapping = mapping->assoc_mapping;
 815
 816                spin_lock(&buffer_mapping->private_lock);
 817                while (!list_empty(list)) {
 818                        struct buffer_head *bh = BH_ENTRY(list->next);
 819                        if (buffer_dirty(bh)) {
 820                                ret = 0;
 821                                break;
 822                        }
 823                        __remove_assoc_queue(bh);
 824                }
 825                spin_unlock(&buffer_mapping->private_lock);
 826        }
 827        return ret;
 828}
 829
 830/*
 831 * Create the appropriate buffers when given a page for data area and
 832 * the size of each buffer.. Use the bh->b_this_page linked list to
 833 * follow the buffers created.  Return NULL if unable to create more
 834 * buffers.
 835 *
 836 * The retry flag is used to differentiate async IO (paging, swapping)
 837 * which may not fail from ordinary buffer allocations.
 838 */
 839struct buffer_head *alloc_page_buffers(struct page *page, unsigned long size,
 840                int retry)
 841{
 842        struct buffer_head *bh, *head;
 843        long offset;
 844
 845try_again:
 846        head = NULL;
 847        offset = PAGE_SIZE;
 848        while ((offset -= size) >= 0) {
 849                bh = alloc_buffer_head(GFP_NOFS);
 850                if (!bh)
 851                        goto no_grow;
 852
 853                bh->b_bdev = NULL;
 854                bh->b_this_page = head;
 855                bh->b_blocknr = -1;
 856                head = bh;
 857
 858                bh->b_state = 0;
 859                atomic_set(&bh->b_count, 0);
 860                bh->b_size = size;
 861
 862                /* Link the buffer to its page */
 863                set_bh_page(bh, page, offset);
 864
 865                init_buffer(bh, NULL, NULL);
 866        }
 867        return head;
 868/*
 869 * In case anything failed, we just free everything we got.
 870 */
 871no_grow:
 872        if (head) {
 873                do {
 874                        bh = head;
 875                        head = head->b_this_page;
 876                        free_buffer_head(bh);
 877                } while (head);
 878        }
 879
 880        /*
 881         * Return failure for non-async IO requests.  Async IO requests
 882         * are not allowed to fail, so we have to wait until buffer heads
 883         * become available.  But we don't want tasks sleeping with 
 884         * partially complete buffers, so all were released above.
 885         */
 886        if (!retry)
 887                return NULL;
 888
 889        /* We're _really_ low on memory. Now we just
 890         * wait for old buffer heads to become free due to
 891         * finishing IO.  Since this is an async request and
 892         * the reserve list is empty, we're sure there are 
 893         * async buffer heads in use.
 894         */
 895        free_more_memory();
 896        goto try_again;
 897}
 898EXPORT_SYMBOL_GPL(alloc_page_buffers);
 899
 900static inline void
 901link_dev_buffers(struct page *page, struct buffer_head *head)
 902{
 903        struct buffer_head *bh, *tail;
 904
 905        bh = head;
 906        do {
 907                tail = bh;
 908                bh = bh->b_this_page;
 909        } while (bh);
 910        tail->b_this_page = head;
 911        attach_page_buffers(page, head);
 912}
 913
 914static sector_t blkdev_max_block(struct block_device *bdev, unsigned int size)
 915{
 916        sector_t retval = ~((sector_t)0);
 917        loff_t sz = i_size_read(bdev->bd_inode);
 918
 919        if (sz) {
 920                unsigned int sizebits = blksize_bits(size);
 921                retval = (sz >> sizebits);
 922        }
 923        return retval;
 924}
 925
 926/*
 927 * Initialise the state of a blockdev page's buffers.
 928 */ 
 929static sector_t
 930init_page_buffers(struct page *page, struct block_device *bdev,
 931                        sector_t block, int size)
 932{
 933        struct buffer_head *head = page_buffers(page);
 934        struct buffer_head *bh = head;
 935        int uptodate = PageUptodate(page);
 936        sector_t end_block = blkdev_max_block(I_BDEV(bdev->bd_inode), size);
 937
 938        do {
 939                if (!buffer_mapped(bh)) {
 940                        init_buffer(bh, NULL, NULL);
 941                        bh->b_bdev = bdev;
 942                        bh->b_blocknr = block;
 943                        if (uptodate)
 944                                set_buffer_uptodate(bh);
 945                        if (block < end_block)
 946                                set_buffer_mapped(bh);
 947                }
 948                block++;
 949                bh = bh->b_this_page;
 950        } while (bh != head);
 951
 952        /*
 953         * Caller needs to validate requested block against end of device.
 954         */
 955        return end_block;
 956}
 957
 958/*
 959 * Create the page-cache page that contains the requested block.
 960 *
 961 * This is used purely for blockdev mappings.
 962 */
 963static int
 964grow_dev_page(struct block_device *bdev, sector_t block,
 965                pgoff_t index, int size, int sizebits)
 966{
 967        struct inode *inode = bdev->bd_inode;
 968        struct page *page;
 969        struct buffer_head *bh;
 970        sector_t end_block;
 971        int ret = 0;            /* Will call free_more_memory() */
 972
 973        page = find_or_create_page(inode->i_mapping, index,
 974                (mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS)|__GFP_MOVABLE);
 975        if (!page)
 976                return ret;
 977
 978        BUG_ON(!PageLocked(page));
 979
 980        if (page_has_buffers(page)) {
 981                bh = page_buffers(page);
 982                if (bh->b_size == size) {
 983                        end_block = init_page_buffers(page, bdev,
 984                                                index << sizebits, size);
 985                        goto done;
 986                }
 987                if (!try_to_free_buffers(page))
 988                        goto failed;
 989        }
 990
 991        /*
 992         * Allocate some buffers for this page
 993         */
 994        bh = alloc_page_buffers(page, size, 0);
 995        if (!bh)
 996                goto failed;
 997
 998        /*
 999         * Link the page to the buffers and initialise them.  Take the
1000         * lock to be atomic wrt __find_get_block(), which does not
1001         * run under the page lock.
1002         */
1003        spin_lock(&inode->i_mapping->private_lock);
1004        link_dev_buffers(page, bh);
1005        end_block = init_page_buffers(page, bdev, index << sizebits, size);
1006        spin_unlock(&inode->i_mapping->private_lock);
1007done:
1008        ret = (block < end_block) ? 1 : -ENXIO;
1009failed:
1010        unlock_page(page);
1011        page_cache_release(page);
1012        return ret;
1013}
1014
1015/*
1016 * Create buffers for the specified block device block's page.  If
1017 * that page was dirty, the buffers are set dirty also.
1018 */
1019static int
1020grow_buffers(struct block_device *bdev, sector_t block, int size)
1021{
1022        pgoff_t index;
1023        int sizebits;
1024
1025        sizebits = -1;
1026        do {
1027                sizebits++;
1028        } while ((size << sizebits) < PAGE_SIZE);
1029
1030        index = block >> sizebits;
1031
1032        /*
1033         * Check for a block which wants to lie outside our maximum possible
1034         * pagecache index.  (this comparison is done using sector_t types).
1035         */
1036        if (unlikely(index != block >> sizebits)) {
1037                char b[BDEVNAME_SIZE];
1038
1039                printk(KERN_ERR "%s: requested out-of-range block %llu for "
1040                        "device %s\n",
1041                        __func__, (unsigned long long)block,
1042                        bdevname(bdev, b));
1043                return -EIO;
1044        }
1045
1046        /* Create a page with the proper size buffers.. */
1047        return grow_dev_page(bdev, block, index, size, sizebits);
1048}
1049
1050static struct buffer_head *
1051__getblk_slow(struct block_device *bdev, sector_t block, int size)
1052{
1053        /* Size must be multiple of hard sectorsize */
1054        if (unlikely(size & (bdev_logical_block_size(bdev)-1) ||
1055                        (size < 512 || size > PAGE_SIZE))) {
1056                printk(KERN_ERR "getblk(): invalid block size %d requested\n",
1057                                        size);
1058                printk(KERN_ERR "logical block size: %d\n",
1059                                        bdev_logical_block_size(bdev));
1060
1061                dump_stack();
1062                return NULL;
1063        }
1064
1065        for (;;) {
1066                struct buffer_head *bh;
1067                int ret;
1068
1069                bh = __find_get_block(bdev, block, size);
1070                if (bh)
1071                        return bh;
1072
1073                ret = grow_buffers(bdev, block, size);
1074                if (ret < 0)
1075                        return NULL;
1076                if (ret == 0)
1077                        free_more_memory();
1078        }
1079}
1080
1081/*
1082 * The relationship between dirty buffers and dirty pages:
1083 *
1084 * Whenever a page has any dirty buffers, the page's dirty bit is set, and
1085 * the page is tagged dirty in its radix tree.
1086 *
1087 * At all times, the dirtiness of the buffers represents the dirtiness of
1088 * subsections of the page.  If the page has buffers, the page dirty bit is
1089 * merely a hint about the true dirty state.
1090 *
1091 * When a page is set dirty in its entirety, all its buffers are marked dirty
1092 * (if the page has buffers).
1093 *
1094 * When a buffer is marked dirty, its page is dirtied, but the page's other
1095 * buffers are not.
1096 *
1097 * Also.  When blockdev buffers are explicitly read with bread(), they
1098 * individually become uptodate.  But their backing page remains not
1099 * uptodate - even if all of its buffers are uptodate.  A subsequent
1100 * block_read_full_page() against that page will discover all the uptodate
1101 * buffers, will set the page uptodate and will perform no I/O.
1102 */
1103
1104/**
1105 * mark_buffer_dirty - mark a buffer_head as needing writeout
1106 * @bh: the buffer_head to mark dirty
1107 *
1108 * mark_buffer_dirty() will set the dirty bit against the buffer, then set its
1109 * backing page dirty, then tag the page as dirty in its address_space's radix
1110 * tree and then attach the address_space's inode to its superblock's dirty
1111 * inode list.
1112 *
1113 * mark_buffer_dirty() is atomic.  It takes bh->b_page->mapping->private_lock,
1114 * mapping->tree_lock and mapping->host->i_lock.
1115 */
1116void mark_buffer_dirty(struct buffer_head *bh)
1117{
1118        WARN_ON_ONCE(!buffer_uptodate(bh));
1119
1120        /*
1121         * Very *carefully* optimize the it-is-already-dirty case.
1122         *
1123         * Don't let the final "is it dirty" escape to before we
1124         * perhaps modified the buffer.
1125         */
1126        if (buffer_dirty(bh)) {
1127                smp_mb();
1128                if (buffer_dirty(bh))
1129                        return;
1130        }
1131
1132        if (!test_set_buffer_dirty(bh)) {
1133                struct page *page = bh->b_page;
1134                if (!TestSetPageDirty(page)) {
1135                        struct address_space *mapping = page_mapping(page);
1136                        if (mapping)
1137                                __set_page_dirty(page, mapping, 0);
1138                }
1139        }
1140}
1141EXPORT_SYMBOL(mark_buffer_dirty);
1142
1143/*
1144 * Decrement a buffer_head's reference count.  If all buffers against a page
1145 * have zero reference count, are clean and unlocked, and if the page is clean
1146 * and unlocked then try_to_free_buffers() may strip the buffers from the page
1147 * in preparation for freeing it (sometimes, rarely, buffers are removed from
1148 * a page but it ends up not being freed, and buffers may later be reattached).
1149 */
1150void __brelse(struct buffer_head * buf)
1151{
1152        if (atomic_read(&buf->b_count)) {
1153                put_bh(buf);
1154                return;
1155        }
1156        WARN(1, KERN_ERR "VFS: brelse: Trying to free free buffer\n");
1157}
1158EXPORT_SYMBOL(__brelse);
1159
1160/*
1161 * bforget() is like brelse(), except it discards any
1162 * potentially dirty data.
1163 */
1164void __bforget(struct buffer_head *bh)
1165{
1166        clear_buffer_dirty(bh);
1167        if (bh->b_assoc_map) {
1168                struct address_space *buffer_mapping = bh->b_page->mapping;
1169
1170                spin_lock(&buffer_mapping->private_lock);
1171                list_del_init(&bh->b_assoc_buffers);
1172                bh->b_assoc_map = NULL;
1173                spin_unlock(&buffer_mapping->private_lock);
1174        }
1175        __brelse(bh);
1176}
1177EXPORT_SYMBOL(__bforget);
1178
1179static struct buffer_head *__bread_slow(struct buffer_head *bh)
1180{
1181        lock_buffer(bh);
1182        if (buffer_uptodate(bh)) {
1183                unlock_buffer(bh);
1184                return bh;
1185        } else {
1186                get_bh(bh);
1187                bh->b_end_io = end_buffer_read_sync;
1188                submit_bh(READ, bh);
1189                wait_on_buffer(bh);
1190                if (buffer_uptodate(bh))
1191                        return bh;
1192        }
1193        brelse(bh);
1194        return NULL;
1195}
1196
1197/*
1198 * Per-cpu buffer LRU implementation.  To reduce the cost of __find_get_block().
1199 * The bhs[] array is sorted - newest buffer is at bhs[0].  Buffers have their
1200 * refcount elevated by one when they're in an LRU.  A buffer can only appear
1201 * once in a particular CPU's LRU.  A single buffer can be present in multiple
1202 * CPU's LRUs at the same time.
1203 *
1204 * This is a transparent caching front-end to sb_bread(), sb_getblk() and
1205 * sb_find_get_block().
1206 *
1207 * The LRUs themselves only need locking against invalidate_bh_lrus.  We use
1208 * a local interrupt disable for that.
1209 */
1210
1211#define BH_LRU_SIZE     8
1212
1213struct bh_lru {
1214        struct buffer_head *bhs[BH_LRU_SIZE];
1215};
1216
1217static DEFINE_PER_CPU(struct bh_lru, bh_lrus) = {{ NULL }};
1218
1219#ifdef CONFIG_SMP
1220#define bh_lru_lock()   local_irq_disable()
1221#define bh_lru_unlock() local_irq_enable()
1222#else
1223#define bh_lru_lock()   preempt_disable()
1224#define bh_lru_unlock() preempt_enable()
1225#endif
1226
1227static inline void check_irqs_on(void)
1228{
1229#ifdef irqs_disabled
1230        BUG_ON(irqs_disabled());
1231#endif
1232}
1233
1234/*
1235 * The LRU management algorithm is dopey-but-simple.  Sorry.
1236 */
1237static void bh_lru_install(struct buffer_head *bh)
1238{
1239        struct buffer_head *evictee = NULL;
1240
1241        check_irqs_on();
1242        bh_lru_lock();
1243        if (__this_cpu_read(bh_lrus.bhs[0]) != bh) {
1244                struct buffer_head *bhs[BH_LRU_SIZE];
1245                int in;
1246                int out = 0;
1247
1248                get_bh(bh);
1249                bhs[out++] = bh;
1250                for (in = 0; in < BH_LRU_SIZE; in++) {
1251                        struct buffer_head *bh2 =
1252                                __this_cpu_read(bh_lrus.bhs[in]);
1253
1254                        if (bh2 == bh) {
1255                                __brelse(bh2);
1256                        } else {
1257                                if (out >= BH_LRU_SIZE) {
1258                                        BUG_ON(evictee != NULL);
1259                                        evictee = bh2;
1260                                } else {
1261                                        bhs[out++] = bh2;
1262                                }
1263                        }
1264                }
1265                while (out < BH_LRU_SIZE)
1266                        bhs[out++] = NULL;
1267                memcpy(__this_cpu_ptr(&bh_lrus.bhs), bhs, sizeof(bhs));
1268        }
1269        bh_lru_unlock();
1270
1271        if (evictee)
1272                __brelse(evictee);
1273}
1274
1275/*
1276 * Look up the bh in this cpu's LRU.  If it's there, move it to the head.
1277 */
1278static struct buffer_head *
1279lookup_bh_lru(struct block_device *bdev, sector_t block, unsigned size)
1280{
1281        struct buffer_head *ret = NULL;
1282        unsigned int i;
1283
1284        check_irqs_on();
1285        bh_lru_lock();
1286        for (i = 0; i < BH_LRU_SIZE; i++) {
1287                struct buffer_head *bh = __this_cpu_read(bh_lrus.bhs[i]);
1288
1289                if (bh && bh->b_bdev == bdev &&
1290                                bh->b_blocknr == block && bh->b_size == size) {
1291                        if (i) {
1292                                while (i) {
1293                                        __this_cpu_write(bh_lrus.bhs[i],
1294                                                __this_cpu_read(bh_lrus.bhs[i - 1]));
1295                                        i--;
1296                                }
1297                                __this_cpu_write(bh_lrus.bhs[0], bh);
1298                        }
1299                        get_bh(bh);
1300                        ret = bh;
1301                        break;
1302                }
1303        }
1304        bh_lru_unlock();
1305        return ret;
1306}
1307
1308/*
1309 * Perform a pagecache lookup for the matching buffer.  If it's there, refresh
1310 * it in the LRU and mark it as accessed.  If it is not present then return
1311 * NULL
1312 */
1313struct buffer_head *
1314__find_get_block(struct block_device *bdev, sector_t block, unsigned size)
1315{
1316        struct buffer_head *bh = lookup_bh_lru(bdev, block, size);
1317
1318        if (bh == NULL) {
1319                bh = __find_get_block_slow(bdev, block);
1320                if (bh)
1321                        bh_lru_install(bh);
1322        }
1323        if (bh)
1324                touch_buffer(bh);
1325        return bh;
1326}
1327EXPORT_SYMBOL(__find_get_block);
1328
1329/*
1330 * __getblk will locate (and, if necessary, create) the buffer_head
1331 * which corresponds to the passed block_device, block and size. The
1332 * returned buffer has its reference count incremented.
1333 *
1334 * __getblk() will lock up the machine if grow_dev_page's try_to_free_buffers()
1335 * attempt is failing.  FIXME, perhaps?
1336 */
1337struct buffer_head *
1338__getblk(struct block_device *bdev, sector_t block, unsigned size)
1339{
1340        struct buffer_head *bh = __find_get_block(bdev, block, size);
1341
1342        might_sleep();
1343        if (bh == NULL)
1344                bh = __getblk_slow(bdev, block, size);
1345        return bh;
1346}
1347EXPORT_SYMBOL(__getblk);
1348
1349/*
1350 * Do async read-ahead on a buffer..
1351 */
1352void __breadahead(struct block_device *bdev, sector_t block, unsigned size)
1353{
1354        struct buffer_head *bh = __getblk(bdev, block, size);
1355        if (likely(bh)) {
1356                ll_rw_block(READA, 1, &bh);
1357                brelse(bh);
1358        }
1359}
1360EXPORT_SYMBOL(__breadahead);
1361
1362/**
1363 *  __bread() - reads a specified block and returns the bh
1364 *  @bdev: the block_device to read from
1365 *  @block: number of block
1366 *  @size: size (in bytes) to read
1367 * 
1368 *  Reads a specified block, and returns buffer head that contains it.
1369 *  It returns NULL if the block was unreadable.
1370 */
1371struct buffer_head *
1372__bread(struct block_device *bdev, sector_t block, unsigned size)
1373{
1374        struct buffer_head *bh = __getblk(bdev, block, size);
1375
1376        if (likely(bh) && !buffer_uptodate(bh))
1377                bh = __bread_slow(bh);
1378        return bh;
1379}
1380EXPORT_SYMBOL(__bread);
1381
1382/*
1383 * invalidate_bh_lrus() is called rarely - but not only at unmount.
1384 * This doesn't race because it runs in each cpu either in irq
1385 * or with preempt disabled.
1386 */
1387static void invalidate_bh_lru(void *arg)
1388{
1389        struct bh_lru *b = &get_cpu_var(bh_lrus);
1390        int i;
1391
1392        for (i = 0; i < BH_LRU_SIZE; i++) {
1393                brelse(b->bhs[i]);
1394                b->bhs[i] = NULL;
1395        }
1396        put_cpu_var(bh_lrus);
1397}
1398
1399static bool has_bh_in_lru(int cpu, void *dummy)
1400{
1401        struct bh_lru *b = per_cpu_ptr(&bh_lrus, cpu);
1402        int i;
1403        
1404        for (i = 0; i < BH_LRU_SIZE; i++) {
1405                if (b->bhs[i])
1406                        return 1;
1407        }
1408
1409        return 0;
1410}
1411
1412void invalidate_bh_lrus(void)
1413{
1414        on_each_cpu_cond(has_bh_in_lru, invalidate_bh_lru, NULL, 1, GFP_KERNEL);
1415}
1416EXPORT_SYMBOL_GPL(invalidate_bh_lrus);
1417
1418void set_bh_page(struct buffer_head *bh,
1419                struct page *page, unsigned long offset)
1420{
1421        bh->b_page = page;
1422        BUG_ON(offset >= PAGE_SIZE);
1423        if (PageHighMem(page))
1424                /*
1425                 * This catches illegal uses and preserves the offset:
1426                 */
1427                bh->b_data = (char *)(0 + offset);
1428        else
1429                bh->b_data = page_address(page) + offset;
1430}
1431EXPORT_SYMBOL(set_bh_page);
1432
1433/*
1434 * Called when truncating a buffer on a page completely.
1435 */
1436static void discard_buffer(struct buffer_head * bh)
1437{
1438        lock_buffer(bh);
1439        clear_buffer_dirty(bh);
1440        bh->b_bdev = NULL;
1441        clear_buffer_mapped(bh);
1442        clear_buffer_req(bh);
1443        clear_buffer_new(bh);
1444        clear_buffer_delay(bh);
1445        clear_buffer_unwritten(bh);
1446        unlock_buffer(bh);
1447}
1448
1449/**
1450 * block_invalidatepage - invalidate part or all of a buffer-backed page
1451 *
1452 * @page: the page which is affected
1453 * @offset: the index of the truncation point
1454 *
1455 * block_invalidatepage() is called when all or part of the page has become
1456 * invalidated by a truncate operation.
1457 *
1458 * block_invalidatepage() does not have to release all buffers, but it must
1459 * ensure that no dirty buffer is left outside @offset and that no I/O
1460 * is underway against any of the blocks which are outside the truncation
1461 * point.  Because the caller is about to free (and possibly reuse) those
1462 * blocks on-disk.
1463 */
1464void block_invalidatepage(struct page *page, unsigned long offset)
1465{
1466        struct buffer_head *head, *bh, *next;
1467        unsigned int curr_off = 0;
1468
1469        BUG_ON(!PageLocked(page));
1470        if (!page_has_buffers(page))
1471                goto out;
1472
1473        head = page_buffers(page);
1474        bh = head;
1475        do {
1476                unsigned int next_off = curr_off + bh->b_size;
1477                next = bh->b_this_page;
1478
1479                /*
1480                 * is this block fully invalidated?
1481                 */
1482                if (offset <= curr_off)
1483                        discard_buffer(bh);
1484                curr_off = next_off;
1485                bh = next;
1486        } while (bh != head);
1487
1488        /*
1489         * We release buffers only if the entire page is being invalidated.
1490         * The get_block cached value has been unconditionally invalidated,
1491         * so real IO is not possible anymore.
1492         */
1493        if (offset == 0)
1494                try_to_release_page(page, 0);
1495out:
1496        return;
1497}
1498EXPORT_SYMBOL(block_invalidatepage);
1499
1500/*
1501 * We attach and possibly dirty the buffers atomically wrt
1502 * __set_page_dirty_buffers() via private_lock.  try_to_free_buffers
1503 * is already excluded via the page lock.
1504 */
1505void create_empty_buffers(struct page *page,
1506                        unsigned long blocksize, unsigned long b_state)
1507{
1508        struct buffer_head *bh, *head, *tail;
1509
1510        head = alloc_page_buffers(page, blocksize, 1);
1511        bh = head;
1512        do {
1513                bh->b_state |= b_state;
1514                tail = bh;
1515                bh = bh->b_this_page;
1516        } while (bh);
1517        tail->b_this_page = head;
1518
1519        spin_lock(&page->mapping->private_lock);
1520        if (PageUptodate(page) || PageDirty(page)) {
1521                bh = head;
1522                do {
1523                        if (PageDirty(page))
1524                                set_buffer_dirty(bh);
1525                        if (PageUptodate(page))
1526                                set_buffer_uptodate(bh);
1527                        bh = bh->b_this_page;
1528                } while (bh != head);
1529        }
1530        attach_page_buffers(page, head);
1531        spin_unlock(&page->mapping->private_lock);
1532}
1533EXPORT_SYMBOL(create_empty_buffers);
1534
1535/*
1536 * We are taking a block for data and we don't want any output from any
1537 * buffer-cache aliases starting from return from that function and
1538 * until the moment when something will explicitly mark the buffer
1539 * dirty (hopefully that will not happen until we will free that block ;-)
1540 * We don't even need to mark it not-uptodate - nobody can expect
1541 * anything from a newly allocated buffer anyway. We used to used
1542 * unmap_buffer() for such invalidation, but that was wrong. We definitely
1543 * don't want to mark the alias unmapped, for example - it would confuse
1544 * anyone who might pick it with bread() afterwards...
1545 *
1546 * Also..  Note that bforget() doesn't lock the buffer.  So there can
1547 * be writeout I/O going on against recently-freed buffers.  We don't
1548 * wait on that I/O in bforget() - it's more efficient to wait on the I/O
1549 * only if we really need to.  That happens here.
1550 */
1551void unmap_underlying_metadata(struct block_device *bdev, sector_t block)
1552{
1553        struct buffer_head *old_bh;
1554
1555        might_sleep();
1556
1557        old_bh = __find_get_block_slow(bdev, block);
1558        if (old_bh) {
1559                clear_buffer_dirty(old_bh);
1560                wait_on_buffer(old_bh);
1561                clear_buffer_req(old_bh);
1562                __brelse(old_bh);
1563        }
1564}
1565EXPORT_SYMBOL(unmap_underlying_metadata);
1566
1567/*
1568 * Size is a power-of-two in the range 512..PAGE_SIZE,
1569 * and the case we care about most is PAGE_SIZE.
1570 *
1571 * So this *could* possibly be written with those
1572 * constraints in mind (relevant mostly if some
1573 * architecture has a slow bit-scan instruction)
1574 */
1575static inline int block_size_bits(unsigned int blocksize)
1576{
1577        return ilog2(blocksize);
1578}
1579
1580static struct buffer_head *create_page_buffers(struct page *page, struct inode *inode, unsigned int b_state)
1581{
1582        BUG_ON(!PageLocked(page));
1583
1584        if (!page_has_buffers(page))
1585                create_empty_buffers(page, 1 << ACCESS_ONCE(inode->i_blkbits), b_state);
1586        return page_buffers(page);
1587}
1588
1589/*
1590 * NOTE! All mapped/uptodate combinations are valid:
1591 *
1592 *      Mapped  Uptodate        Meaning
1593 *
1594 *      No      No              "unknown" - must do get_block()
1595 *      No      Yes             "hole" - zero-filled
1596 *      Yes     No              "allocated" - allocated on disk, not read in
1597 *      Yes     Yes             "valid" - allocated and up-to-date in memory.
1598 *
1599 * "Dirty" is valid only with the last case (mapped+uptodate).
1600 */
1601
1602/*
1603 * While block_write_full_page is writing back the dirty buffers under
1604 * the page lock, whoever dirtied the buffers may decide to clean them
1605 * again at any time.  We handle that by only looking at the buffer
1606 * state inside lock_buffer().
1607 *
1608 * If block_write_full_page() is called for regular writeback
1609 * (wbc->sync_mode == WB_SYNC_NONE) then it will redirty a page which has a
1610 * locked buffer.   This only can happen if someone has written the buffer
1611 * directly, with submit_bh().  At the address_space level PageWriteback
1612 * prevents this contention from occurring.
1613 *
1614 * If block_write_full_page() is called with wbc->sync_mode ==
1615 * WB_SYNC_ALL, the writes are posted using WRITE_SYNC; this
1616 * causes the writes to be flagged as synchronous writes.
1617 */
1618static int __block_write_full_page(struct inode *inode, struct page *page,
1619                        get_block_t *get_block, struct writeback_control *wbc,
1620                        bh_end_io_t *handler)
1621{
1622        int err;
1623        sector_t block;
1624        sector_t last_block;
1625        struct buffer_head *bh, *head;
1626        unsigned int blocksize, bbits;
1627        int nr_underway = 0;
1628        int write_op = (wbc->sync_mode == WB_SYNC_ALL ?
1629                        WRITE_SYNC : WRITE);
1630
1631        head = create_page_buffers(page, inode,
1632                                        (1 << BH_Dirty)|(1 << BH_Uptodate));
1633
1634        /*
1635         * Be very careful.  We have no exclusion from __set_page_dirty_buffers
1636         * here, and the (potentially unmapped) buffers may become dirty at
1637         * any time.  If a buffer becomes dirty here after we've inspected it
1638         * then we just miss that fact, and the page stays dirty.
1639         *
1640         * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
1641         * handle that here by just cleaning them.
1642         */
1643
1644        bh = head;
1645        blocksize = bh->b_size;
1646        bbits = block_size_bits(blocksize);
1647
1648        block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1649        last_block = (i_size_read(inode) - 1) >> bbits;
1650
1651        /*
1652         * Get all the dirty buffers mapped to disk addresses and
1653         * handle any aliases from the underlying blockdev's mapping.
1654         */
1655        do {
1656                if (block > last_block) {
1657                        /*
1658                         * mapped buffers outside i_size will occur, because
1659                         * this page can be outside i_size when there is a
1660                         * truncate in progress.
1661                         */
1662                        /*
1663                         * The buffer was zeroed by block_write_full_page()
1664                         */
1665                        clear_buffer_dirty(bh);
1666                        set_buffer_uptodate(bh);
1667                } else if ((!buffer_mapped(bh) || buffer_delay(bh)) &&
1668                           buffer_dirty(bh)) {
1669                        WARN_ON(bh->b_size != blocksize);
1670                        err = get_block(inode, block, bh, 1);
1671                        if (err)
1672                                goto recover;
1673                        clear_buffer_delay(bh);
1674                        if (buffer_new(bh)) {
1675                                /* blockdev mappings never come here */
1676                                clear_buffer_new(bh);
1677                                unmap_underlying_metadata(bh->b_bdev,
1678                                                        bh->b_blocknr);
1679                        }
1680                }
1681                bh = bh->b_this_page;
1682                block++;
1683        } while (bh != head);
1684
1685        do {
1686                if (!buffer_mapped(bh))
1687                        continue;
1688                /*
1689                 * If it's a fully non-blocking write attempt and we cannot
1690                 * lock the buffer then redirty the page.  Note that this can
1691                 * potentially cause a busy-wait loop from writeback threads
1692                 * and kswapd activity, but those code paths have their own
1693                 * higher-level throttling.
1694                 */
1695                if (wbc->sync_mode != WB_SYNC_NONE) {
1696                        lock_buffer(bh);
1697                } else if (!trylock_buffer(bh)) {
1698                        redirty_page_for_writepage(wbc, page);
1699                        continue;
1700                }
1701                if (test_clear_buffer_dirty(bh)) {
1702                        mark_buffer_async_write_endio(bh, handler);
1703                } else {
1704                        unlock_buffer(bh);
1705                }
1706        } while ((bh = bh->b_this_page) != head);
1707
1708        /*
1709         * The page and its buffers are protected by PageWriteback(), so we can
1710         * drop the bh refcounts early.
1711         */
1712        BUG_ON(PageWriteback(page));
1713        set_page_writeback(page);
1714
1715        do {
1716                struct buffer_head *next = bh->b_this_page;
1717                if (buffer_async_write(bh)) {
1718                        submit_bh(write_op, bh);
1719                        nr_underway++;
1720                }
1721                bh = next;
1722        } while (bh != head);
1723        unlock_page(page);
1724
1725        err = 0;
1726done:
1727        if (nr_underway == 0) {
1728                /*
1729                 * The page was marked dirty, but the buffers were
1730                 * clean.  Someone wrote them back by hand with
1731                 * ll_rw_block/submit_bh.  A rare case.
1732                 */
1733                end_page_writeback(page);
1734
1735                /*
1736                 * The page and buffer_heads can be released at any time from
1737                 * here on.
1738                 */
1739        }
1740        return err;
1741
1742recover:
1743        /*
1744         * ENOSPC, or some other error.  We may already have added some
1745         * blocks to the file, so we need to write these out to avoid
1746         * exposing stale data.
1747         * The page is currently locked and not marked for writeback
1748         */
1749        bh = head;
1750        /* Recovery: lock and submit the mapped buffers */
1751        do {
1752                if (buffer_mapped(bh) && buffer_dirty(bh) &&
1753                    !buffer_delay(bh)) {
1754                        lock_buffer(bh);
1755                        mark_buffer_async_write_endio(bh, handler);
1756                } else {
1757                        /*
1758                         * The buffer may have been set dirty during
1759                         * attachment to a dirty page.
1760                         */
1761                        clear_buffer_dirty(bh);
1762                }
1763        } while ((bh = bh->b_this_page) != head);
1764        SetPageError(page);
1765        BUG_ON(PageWriteback(page));
1766        mapping_set_error(page->mapping, err);
1767        set_page_writeback(page);
1768        do {
1769                struct buffer_head *next = bh->b_this_page;
1770                if (buffer_async_write(bh)) {
1771                        clear_buffer_dirty(bh);
1772                        submit_bh(write_op, bh);
1773                        nr_underway++;
1774                }
1775                bh = next;
1776        } while (bh != head);
1777        unlock_page(page);
1778        goto done;
1779}
1780
1781/*
1782 * If a page has any new buffers, zero them out here, and mark them uptodate
1783 * and dirty so they'll be written out (in order to prevent uninitialised
1784 * block data from leaking). And clear the new bit.
1785 */
1786void page_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1787{
1788        unsigned int block_start, block_end;
1789        struct buffer_head *head, *bh;
1790
1791        BUG_ON(!PageLocked(page));
1792        if (!page_has_buffers(page))
1793                return;
1794
1795        bh = head = page_buffers(page);
1796        block_start = 0;
1797        do {
1798                block_end = block_start + bh->b_size;
1799
1800                if (buffer_new(bh)) {
1801                        if (block_end > from && block_start < to) {
1802                                if (!PageUptodate(page)) {
1803                                        unsigned start, size;
1804
1805                                        start = max(from, block_start);
1806                                        size = min(to, block_end) - start;
1807
1808                                        zero_user(page, start, size);
1809                                        set_buffer_uptodate(bh);
1810                                }
1811
1812                                clear_buffer_new(bh);
1813                                mark_buffer_dirty(bh);
1814                        }
1815                }
1816
1817                block_start = block_end;
1818                bh = bh->b_this_page;
1819        } while (bh != head);
1820}
1821EXPORT_SYMBOL(page_zero_new_buffers);
1822
1823int __block_write_begin(struct page *page, loff_t pos, unsigned len,
1824                get_block_t *get_block)
1825{
1826        unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1827        unsigned to = from + len;
1828        struct inode *inode = page->mapping->host;
1829        unsigned block_start, block_end;
1830        sector_t block;
1831        int err = 0;
1832        unsigned blocksize, bbits;
1833        struct buffer_head *bh, *head, *wait[2], **wait_bh=wait;
1834
1835        BUG_ON(!PageLocked(page));
1836        BUG_ON(from > PAGE_CACHE_SIZE);
1837        BUG_ON(to > PAGE_CACHE_SIZE);
1838        BUG_ON(from > to);
1839
1840        head = create_page_buffers(page, inode, 0);
1841        blocksize = head->b_size;
1842        bbits = block_size_bits(blocksize);
1843
1844        block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
1845
1846        for(bh = head, block_start = 0; bh != head || !block_start;
1847            block++, block_start=block_end, bh = bh->b_this_page) {
1848                block_end = block_start + blocksize;
1849                if (block_end <= from || block_start >= to) {
1850                        if (PageUptodate(page)) {
1851                                if (!buffer_uptodate(bh))
1852                                        set_buffer_uptodate(bh);
1853                        }
1854                        continue;
1855                }
1856                if (buffer_new(bh))
1857                        clear_buffer_new(bh);
1858                if (!buffer_mapped(bh)) {
1859                        WARN_ON(bh->b_size != blocksize);
1860                        err = get_block(inode, block, bh, 1);
1861                        if (err)
1862                                break;
1863                        if (buffer_new(bh)) {
1864                                unmap_underlying_metadata(bh->b_bdev,
1865                                                        bh->b_blocknr);
1866                                if (PageUptodate(page)) {
1867                                        clear_buffer_new(bh);
1868                                        set_buffer_uptodate(bh);
1869                                        mark_buffer_dirty(bh);
1870                                        continue;
1871                                }
1872                                if (block_end > to || block_start < from)
1873                                        zero_user_segments(page,
1874                                                to, block_end,
1875                                                block_start, from);
1876                                continue;
1877                        }
1878                }
1879                if (PageUptodate(page)) {
1880                        if (!buffer_uptodate(bh))
1881                                set_buffer_uptodate(bh);
1882                        continue; 
1883                }
1884                if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1885                    !buffer_unwritten(bh) &&
1886                     (block_start < from || block_end > to)) {
1887                        ll_rw_block(READ, 1, &bh);
1888                        *wait_bh++=bh;
1889                }
1890        }
1891        /*
1892         * If we issued read requests - let them complete.
1893         */
1894        while(wait_bh > wait) {
1895                wait_on_buffer(*--wait_bh);
1896                if (!buffer_uptodate(*wait_bh))
1897                        err = -EIO;
1898        }
1899        if (unlikely(err))
1900                page_zero_new_buffers(page, from, to);
1901        return err;
1902}
1903EXPORT_SYMBOL(__block_write_begin);
1904
1905static int __block_commit_write(struct inode *inode, struct page *page,
1906                unsigned from, unsigned to)
1907{
1908        unsigned block_start, block_end;
1909        int partial = 0;
1910        unsigned blocksize;
1911        struct buffer_head *bh, *head;
1912
1913        bh = head = page_buffers(page);
1914        blocksize = bh->b_size;
1915
1916        block_start = 0;
1917        do {
1918                block_end = block_start + blocksize;
1919                if (block_end <= from || block_start >= to) {
1920                        if (!buffer_uptodate(bh))
1921                                partial = 1;
1922                } else {
1923                        set_buffer_uptodate(bh);
1924                        mark_buffer_dirty(bh);
1925                }
1926                clear_buffer_new(bh);
1927
1928                block_start = block_end;
1929                bh = bh->b_this_page;
1930        } while (bh != head);
1931
1932        /*
1933         * If this is a partial write which happened to make all buffers
1934         * uptodate then we can optimize away a bogus readpage() for
1935         * the next read(). Here we 'discover' whether the page went
1936         * uptodate as a result of this (potentially partial) write.
1937         */
1938        if (!partial)
1939                SetPageUptodate(page);
1940        return 0;
1941}
1942
1943/*
1944 * block_write_begin takes care of the basic task of block allocation and
1945 * bringing partial write blocks uptodate first.
1946 *
1947 * The filesystem needs to handle block truncation upon failure.
1948 */
1949int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
1950                unsigned flags, struct page **pagep, get_block_t *get_block)
1951{
1952        pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1953        struct page *page;
1954        int status;
1955
1956        page = grab_cache_page_write_begin(mapping, index, flags);
1957        if (!page)
1958                return -ENOMEM;
1959
1960        status = __block_write_begin(page, pos, len, get_block);
1961        if (unlikely(status)) {
1962                unlock_page(page);
1963                page_cache_release(page);
1964                page = NULL;
1965        }
1966
1967        *pagep = page;
1968        return status;
1969}
1970EXPORT_SYMBOL(block_write_begin);
1971
1972int block_write_end(struct file *file, struct address_space *mapping,
1973                        loff_t pos, unsigned len, unsigned copied,
1974                        struct page *page, void *fsdata)
1975{
1976        struct inode *inode = mapping->host;
1977        unsigned start;
1978
1979        start = pos & (PAGE_CACHE_SIZE - 1);
1980
1981        if (unlikely(copied < len)) {
1982                /*
1983                 * The buffers that were written will now be uptodate, so we
1984                 * don't have to worry about a readpage reading them and
1985                 * overwriting a partial write. However if we have encountered
1986                 * a short write and only partially written into a buffer, it
1987                 * will not be marked uptodate, so a readpage might come in and
1988                 * destroy our partial write.
1989                 *
1990                 * Do the simplest thing, and just treat any short write to a
1991                 * non uptodate page as a zero-length write, and force the
1992                 * caller to redo the whole thing.
1993                 */
1994                if (!PageUptodate(page))
1995                        copied = 0;
1996
1997                page_zero_new_buffers(page, start+copied, start+len);
1998        }
1999        flush_dcache_page(page);
2000
2001        /* This could be a short (even 0-length) commit */
2002        __block_commit_write(inode, page, start, start+copied);
2003
2004        return copied;
2005}
2006EXPORT_SYMBOL(block_write_end);
2007
2008int generic_write_end(struct file *file, struct address_space *mapping,
2009                        loff_t pos, unsigned len, unsigned copied,
2010                        struct page *page, void *fsdata)
2011{
2012        struct inode *inode = mapping->host;
2013        int i_size_changed = 0;
2014
2015        copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
2016
2017        /*
2018         * No need to use i_size_read() here, the i_size
2019         * cannot change under us because we hold i_mutex.
2020         *
2021         * But it's important to update i_size while still holding page lock:
2022         * page writeout could otherwise come in and zero beyond i_size.
2023         */
2024        if (pos+copied > inode->i_size) {
2025                i_size_write(inode, pos+copied);
2026                i_size_changed = 1;
2027        }
2028
2029        unlock_page(page);
2030        page_cache_release(page);
2031
2032        /*
2033         * Don't mark the inode dirty under page lock. First, it unnecessarily
2034         * makes the holding time of page lock longer. Second, it forces lock
2035         * ordering of page lock and transaction start for journaling
2036         * filesystems.
2037         */
2038        if (i_size_changed)
2039                mark_inode_dirty(inode);
2040
2041        return copied;
2042}
2043EXPORT_SYMBOL(generic_write_end);
2044
2045/*
2046 * block_is_partially_uptodate checks whether buffers within a page are
2047 * uptodate or not.
2048 *
2049 * Returns true if all buffers which correspond to a file portion
2050 * we want to read are uptodate.
2051 */
2052int block_is_partially_uptodate(struct page *page, read_descriptor_t *desc,
2053                                        unsigned long from)
2054{
2055        unsigned block_start, block_end, blocksize;
2056        unsigned to;
2057        struct buffer_head *bh, *head;
2058        int ret = 1;
2059
2060        if (!page_has_buffers(page))
2061                return 0;
2062
2063        head = page_buffers(page);
2064        blocksize = head->b_size;
2065        to = min_t(unsigned, PAGE_CACHE_SIZE - from, desc->count);
2066        to = from + to;
2067        if (from < blocksize && to > PAGE_CACHE_SIZE - blocksize)
2068                return 0;
2069
2070        bh = head;
2071        block_start = 0;
2072        do {
2073                block_end = block_start + blocksize;
2074                if (block_end > from && block_start < to) {
2075                        if (!buffer_uptodate(bh)) {
2076                                ret = 0;
2077                                break;
2078                        }
2079                        if (block_end >= to)
2080                                break;
2081                }
2082                block_start = block_end;
2083                bh = bh->b_this_page;
2084        } while (bh != head);
2085
2086        return ret;
2087}
2088EXPORT_SYMBOL(block_is_partially_uptodate);
2089
2090/*
2091 * Generic "read page" function for block devices that have the normal
2092 * get_block functionality. This is most of the block device filesystems.
2093 * Reads the page asynchronously --- the unlock_buffer() and
2094 * set/clear_buffer_uptodate() functions propagate buffer state into the
2095 * page struct once IO has completed.
2096 */
2097int block_read_full_page(struct page *page, get_block_t *get_block)
2098{
2099        struct inode *inode = page->mapping->host;
2100        sector_t iblock, lblock;
2101        struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
2102        unsigned int blocksize, bbits;
2103        int nr, i;
2104        int fully_mapped = 1;
2105
2106        head = create_page_buffers(page, inode, 0);
2107        blocksize = head->b_size;
2108        bbits = block_size_bits(blocksize);
2109
2110        iblock = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
2111        lblock = (i_size_read(inode)+blocksize-1) >> bbits;
2112        bh = head;
2113        nr = 0;
2114        i = 0;
2115
2116        do {
2117                if (buffer_uptodate(bh))
2118                        continue;
2119
2120                if (!buffer_mapped(bh)) {
2121                        int err = 0;
2122
2123                        fully_mapped = 0;
2124                        if (iblock < lblock) {
2125                                WARN_ON(bh->b_size != blocksize);
2126                                err = get_block(inode, iblock, bh, 0);
2127                                if (err)
2128                                        SetPageError(page);
2129                        }
2130                        if (!buffer_mapped(bh)) {
2131                                zero_user(page, i * blocksize, blocksize);
2132                                if (!err)
2133                                        set_buffer_uptodate(bh);
2134                                continue;
2135                        }
2136                        /*
2137                         * get_block() might have updated the buffer
2138                         * synchronously
2139                         */
2140                        if (buffer_uptodate(bh))
2141                                continue;
2142                }
2143                arr[nr++] = bh;
2144        } while (i++, iblock++, (bh = bh->b_this_page) != head);
2145
2146        if (fully_mapped)
2147                SetPageMappedToDisk(page);
2148
2149        if (!nr) {
2150                /*
2151                 * All buffers are uptodate - we can set the page uptodate
2152                 * as well. But not if get_block() returned an error.
2153                 */
2154                if (!PageError(page))
2155                        SetPageUptodate(page);
2156                unlock_page(page);
2157                return 0;
2158        }
2159
2160        /* Stage two: lock the buffers */
2161        for (i = 0; i < nr; i++) {
2162                bh = arr[i];
2163                lock_buffer(bh);
2164                mark_buffer_async_read(bh);
2165        }
2166
2167        /*
2168         * Stage 3: start the IO.  Check for uptodateness
2169         * inside the buffer lock in case another process reading
2170         * the underlying blockdev brought it uptodate (the sct fix).
2171         */
2172        for (i = 0; i < nr; i++) {
2173                bh = arr[i];
2174                if (buffer_uptodate(bh))
2175                        end_buffer_async_read(bh, 1);
2176                else
2177                        submit_bh(READ, bh);
2178        }
2179        return 0;
2180}
2181EXPORT_SYMBOL(block_read_full_page);
2182
2183/* utility function for filesystems that need to do work on expanding
2184 * truncates.  Uses filesystem pagecache writes to allow the filesystem to
2185 * deal with the hole.  
2186 */
2187int generic_cont_expand_simple(struct inode *inode, loff_t size)
2188{
2189        struct address_space *mapping = inode->i_mapping;
2190        struct page *page;
2191        void *fsdata;
2192        int err;
2193
2194        err = inode_newsize_ok(inode, size);
2195        if (err)
2196                goto out;
2197
2198        err = pagecache_write_begin(NULL, mapping, size, 0,
2199                                AOP_FLAG_UNINTERRUPTIBLE|AOP_FLAG_CONT_EXPAND,
2200                                &page, &fsdata);
2201        if (err)
2202                goto out;
2203
2204        err = pagecache_write_end(NULL, mapping, size, 0, 0, page, fsdata);
2205        BUG_ON(err > 0);
2206
2207out:
2208        return err;
2209}
2210EXPORT_SYMBOL(generic_cont_expand_simple);
2211
2212static int cont_expand_zero(struct file *file, struct address_space *mapping,
2213                            loff_t pos, loff_t *bytes)
2214{
2215        struct inode *inode = mapping->host;
2216        unsigned blocksize = 1 << inode->i_blkbits;
2217        struct page *page;
2218        void *fsdata;
2219        pgoff_t index, curidx;
2220        loff_t curpos;
2221        unsigned zerofrom, offset, len;
2222        int err = 0;
2223
2224        index = pos >> PAGE_CACHE_SHIFT;
2225        offset = pos & ~PAGE_CACHE_MASK;
2226
2227        while (index > (curidx = (curpos = *bytes)>>PAGE_CACHE_SHIFT)) {
2228                zerofrom = curpos & ~PAGE_CACHE_MASK;
2229                if (zerofrom & (blocksize-1)) {
2230                        *bytes |= (blocksize-1);
2231                        (*bytes)++;
2232                }
2233                len = PAGE_CACHE_SIZE - zerofrom;
2234
2235                err = pagecache_write_begin(file, mapping, curpos, len,
2236                                                AOP_FLAG_UNINTERRUPTIBLE,
2237                                                &page, &fsdata);
2238                if (err)
2239                        goto out;
2240                zero_user(page, zerofrom, len);
2241                err = pagecache_write_end(file, mapping, curpos, len, len,
2242                                                page, fsdata);
2243                if (err < 0)
2244                        goto out;
2245                BUG_ON(err != len);
2246                err = 0;
2247
2248                balance_dirty_pages_ratelimited(mapping);
2249        }
2250
2251        /* page covers the boundary, find the boundary offset */
2252        if (index == curidx) {
2253                zerofrom = curpos & ~PAGE_CACHE_MASK;
2254                /* if we will expand the thing last block will be filled */
2255                if (offset <= zerofrom) {
2256                        goto out;
2257                }
2258                if (zerofrom & (blocksize-1)) {
2259                        *bytes |= (blocksize-1);
2260                        (*bytes)++;
2261                }
2262                len = offset - zerofrom;
2263
2264                err = pagecache_write_begin(file, mapping, curpos, len,
2265                                                AOP_FLAG_UNINTERRUPTIBLE,
2266                                                &page, &fsdata);
2267                if (err)
2268                        goto out;
2269                zero_user(page, zerofrom, len);
2270                err = pagecache_write_end(file, mapping, curpos, len, len,
2271                                                page, fsdata);
2272                if (err < 0)
2273                        goto out;
2274                BUG_ON(err != len);
2275                err = 0;
2276        }
2277out:
2278        return err;
2279}
2280
2281/*
2282 * For moronic filesystems that do not allow holes in file.
2283 * We may have to extend the file.
2284 */
2285int cont_write_begin(struct file *file, struct address_space *mapping,
2286                        loff_t pos, unsigned len, unsigned flags,
2287                        struct page **pagep, void **fsdata,
2288                        get_block_t *get_block, loff_t *bytes)
2289{
2290        struct inode *inode = mapping->host;
2291        unsigned blocksize = 1 << inode->i_blkbits;
2292        unsigned zerofrom;
2293        int err;
2294
2295        err = cont_expand_zero(file, mapping, pos, bytes);
2296        if (err)
2297                return err;
2298
2299        zerofrom = *bytes & ~PAGE_CACHE_MASK;
2300        if (pos+len > *bytes && zerofrom & (blocksize-1)) {
2301                *bytes |= (blocksize-1);
2302                (*bytes)++;
2303        }
2304
2305        return block_write_begin(mapping, pos, len, flags, pagep, get_block);
2306}
2307EXPORT_SYMBOL(cont_write_begin);
2308
2309int block_commit_write(struct page *page, unsigned from, unsigned to)
2310{
2311        struct inode *inode = page->mapping->host;
2312        __block_commit_write(inode,page,from,to);
2313        return 0;
2314}
2315EXPORT_SYMBOL(block_commit_write);
2316
2317/*
2318 * block_page_mkwrite() is not allowed to change the file size as it gets
2319 * called from a page fault handler when a page is first dirtied. Hence we must
2320 * be careful to check for EOF conditions here. We set the page up correctly
2321 * for a written page which means we get ENOSPC checking when writing into
2322 * holes and correct delalloc and unwritten extent mapping on filesystems that
2323 * support these features.
2324 *
2325 * We are not allowed to take the i_mutex here so we have to play games to
2326 * protect against truncate races as the page could now be beyond EOF.  Because
2327 * truncate writes the inode size before removing pages, once we have the
2328 * page lock we can determine safely if the page is beyond EOF. If it is not
2329 * beyond EOF, then the page is guaranteed safe against truncation until we
2330 * unlock the page.
2331 *
2332 * Direct callers of this function should protect against filesystem freezing
2333 * using sb_start_write() - sb_end_write() functions.
2334 */
2335int __block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2336                         get_block_t get_block)
2337{
2338        struct page *page = vmf->page;
2339        struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
2340        unsigned long end;
2341        loff_t size;
2342        int ret;
2343
2344        lock_page(page);
2345        size = i_size_read(inode);
2346        if ((page->mapping != inode->i_mapping) ||
2347            (page_offset(page) > size)) {
2348                /* We overload EFAULT to mean page got truncated */
2349                ret = -EFAULT;
2350                goto out_unlock;
2351        }
2352
2353        /* page is wholly or partially inside EOF */
2354        if (((page->index + 1) << PAGE_CACHE_SHIFT) > size)
2355                end = size & ~PAGE_CACHE_MASK;
2356        else
2357                end = PAGE_CACHE_SIZE;
2358
2359        ret = __block_write_begin(page, 0, end, get_block);
2360        if (!ret)
2361                ret = block_commit_write(page, 0, end);
2362
2363        if (unlikely(ret < 0))
2364                goto out_unlock;
2365        set_page_dirty(page);
2366        wait_on_page_writeback(page);
2367        return 0;
2368out_unlock:
2369        unlock_page(page);
2370        return ret;
2371}
2372EXPORT_SYMBOL(__block_page_mkwrite);
2373
2374int block_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf,
2375                   get_block_t get_block)
2376{
2377        int ret;
2378        struct super_block *sb = vma->vm_file->f_path.dentry->d_inode->i_sb;
2379
2380        sb_start_pagefault(sb);
2381
2382        /*
2383         * Update file times before taking page lock. We may end up failing the
2384         * fault so this update may be superfluous but who really cares...
2385         */
2386        file_update_time(vma->vm_file);
2387
2388        ret = __block_page_mkwrite(vma, vmf, get_block);
2389        sb_end_pagefault(sb);
2390        return block_page_mkwrite_return(ret);
2391}
2392EXPORT_SYMBOL(block_page_mkwrite);
2393
2394/*
2395 * nobh_write_begin()'s prereads are special: the buffer_heads are freed
2396 * immediately, while under the page lock.  So it needs a special end_io
2397 * handler which does not touch the bh after unlocking it.
2398 */
2399static void end_buffer_read_nobh(struct buffer_head *bh, int uptodate)
2400{
2401        __end_buffer_read_notouch(bh, uptodate);
2402}
2403
2404/*
2405 * Attach the singly-linked list of buffers created by nobh_write_begin, to
2406 * the page (converting it to circular linked list and taking care of page
2407 * dirty races).
2408 */
2409static void attach_nobh_buffers(struct page *page, struct buffer_head *head)
2410{
2411        struct buffer_head *bh;
2412
2413        BUG_ON(!PageLocked(page));
2414
2415        spin_lock(&page->mapping->private_lock);
2416        bh = head;
2417        do {
2418                if (PageDirty(page))
2419                        set_buffer_dirty(bh);
2420                if (!bh->b_this_page)
2421                        bh->b_this_page = head;
2422                bh = bh->b_this_page;
2423        } while (bh != head);
2424        attach_page_buffers(page, head);
2425        spin_unlock(&page->mapping->private_lock);
2426}
2427
2428/*
2429 * On entry, the page is fully not uptodate.
2430 * On exit the page is fully uptodate in the areas outside (from,to)
2431 * The filesystem needs to handle block truncation upon failure.
2432 */
2433int nobh_write_begin(struct address_space *mapping,
2434                        loff_t pos, unsigned len, unsigned flags,
2435                        struct page **pagep, void **fsdata,
2436                        get_block_t *get_block)
2437{
2438        struct inode *inode = mapping->host;
2439        const unsigned blkbits = inode->i_blkbits;
2440        const unsigned blocksize = 1 << blkbits;
2441        struct buffer_head *head, *bh;
2442        struct page *page;
2443        pgoff_t index;
2444        unsigned from, to;
2445        unsigned block_in_page;
2446        unsigned block_start, block_end;
2447        sector_t block_in_file;
2448        int nr_reads = 0;
2449        int ret = 0;
2450        int is_mapped_to_disk = 1;
2451
2452        index = pos >> PAGE_CACHE_SHIFT;
2453        from = pos & (PAGE_CACHE_SIZE - 1);
2454        to = from + len;
2455
2456        page = grab_cache_page_write_begin(mapping, index, flags);
2457        if (!page)
2458                return -ENOMEM;
2459        *pagep = page;
2460        *fsdata = NULL;
2461
2462        if (page_has_buffers(page)) {
2463                ret = __block_write_begin(page, pos, len, get_block);
2464                if (unlikely(ret))
2465                        goto out_release;
2466                return ret;
2467        }
2468
2469        if (PageMappedToDisk(page))
2470                return 0;
2471
2472        /*
2473         * Allocate buffers so that we can keep track of state, and potentially
2474         * attach them to the page if an error occurs. In the common case of
2475         * no error, they will just be freed again without ever being attached
2476         * to the page (which is all OK, because we're under the page lock).
2477         *
2478         * Be careful: the buffer linked list is a NULL terminated one, rather
2479         * than the circular one we're used to.
2480         */
2481        head = alloc_page_buffers(page, blocksize, 0);
2482        if (!head) {
2483                ret = -ENOMEM;
2484                goto out_release;
2485        }
2486
2487        block_in_file = (sector_t)page->index << (PAGE_CACHE_SHIFT - blkbits);
2488
2489        /*
2490         * We loop across all blocks in the page, whether or not they are
2491         * part of the affected region.  This is so we can discover if the
2492         * page is fully mapped-to-disk.
2493         */
2494        for (block_start = 0, block_in_page = 0, bh = head;
2495                  block_start < PAGE_CACHE_SIZE;
2496                  block_in_page++, block_start += blocksize, bh = bh->b_this_page) {
2497                int create;
2498
2499                block_end = block_start + blocksize;
2500                bh->b_state = 0;
2501                create = 1;
2502                if (block_start >= to)
2503                        create = 0;
2504                ret = get_block(inode, block_in_file + block_in_page,
2505                                        bh, create);
2506                if (ret)
2507                        goto failed;
2508                if (!buffer_mapped(bh))
2509                        is_mapped_to_disk = 0;
2510                if (buffer_new(bh))
2511                        unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
2512                if (PageUptodate(page)) {
2513                        set_buffer_uptodate(bh);
2514                        continue;
2515                }
2516                if (buffer_new(bh) || !buffer_mapped(bh)) {
2517                        zero_user_segments(page, block_start, from,
2518                                                        to, block_end);
2519                        continue;
2520                }
2521                if (buffer_uptodate(bh))
2522                        continue;       /* reiserfs does this */
2523                if (block_start < from || block_end > to) {
2524                        lock_buffer(bh);
2525                        bh->b_end_io = end_buffer_read_nobh;
2526                        submit_bh(READ, bh);
2527                        nr_reads++;
2528                }
2529        }
2530
2531        if (nr_reads) {
2532                /*
2533                 * The page is locked, so these buffers are protected from
2534                 * any VM or truncate activity.  Hence we don't need to care
2535                 * for the buffer_head refcounts.
2536                 */
2537                for (bh = head; bh; bh = bh->b_this_page) {
2538                        wait_on_buffer(bh);
2539                        if (!buffer_uptodate(bh))
2540                                ret = -EIO;
2541                }
2542                if (ret)
2543                        goto failed;
2544        }
2545
2546        if (is_mapped_to_disk)
2547                SetPageMappedToDisk(page);
2548
2549        *fsdata = head; /* to be released by nobh_write_end */
2550
2551        return 0;
2552
2553failed:
2554        BUG_ON(!ret);
2555        /*
2556         * Error recovery is a bit difficult. We need to zero out blocks that
2557         * were newly allocated, and dirty them to ensure they get written out.
2558         * Buffers need to be attached to the page at this point, otherwise
2559         * the handling of potential IO errors during writeout would be hard
2560         * (could try doing synchronous writeout, but what if that fails too?)
2561         */
2562        attach_nobh_buffers(page, head);
2563        page_zero_new_buffers(page, from, to);
2564
2565out_release:
2566        unlock_page(page);
2567        page_cache_release(page);
2568        *pagep = NULL;
2569
2570        return ret;
2571}
2572EXPORT_SYMBOL(nobh_write_begin);
2573
2574int nobh_write_end(struct file *file, struct address_space *mapping,
2575                        loff_t pos, unsigned len, unsigned copied,
2576                        struct page *page, void *fsdata)
2577{
2578        struct inode *inode = page->mapping->host;
2579        struct buffer_head *head = fsdata;
2580        struct buffer_head *bh;
2581        BUG_ON(fsdata != NULL && page_has_buffers(page));
2582
2583        if (unlikely(copied < len) && head)
2584                attach_nobh_buffers(page, head);
2585        if (page_has_buffers(page))
2586                return generic_write_end(file, mapping, pos, len,
2587                                        copied, page, fsdata);
2588
2589        SetPageUptodate(page);
2590        set_page_dirty(page);
2591        if (pos+copied > inode->i_size) {
2592                i_size_write(inode, pos+copied);
2593                mark_inode_dirty(inode);
2594        }
2595
2596        unlock_page(page);
2597        page_cache_release(page);
2598
2599        while (head) {
2600                bh = head;
2601                head = head->b_this_page;
2602                free_buffer_head(bh);
2603        }
2604
2605        return copied;
2606}
2607EXPORT_SYMBOL(nobh_write_end);
2608
2609/*
2610 * nobh_writepage() - based on block_full_write_page() except
2611 * that it tries to operate without attaching bufferheads to
2612 * the page.
2613 */
2614int nobh_writepage(struct page *page, get_block_t *get_block,
2615                        struct writeback_control *wbc)
2616{
2617        struct inode * const inode = page->mapping->host;
2618        loff_t i_size = i_size_read(inode);
2619        const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2620        unsigned offset;
2621        int ret;
2622
2623        /* Is the page fully inside i_size? */
2624        if (page->index < end_index)
2625                goto out;
2626
2627        /* Is the page fully outside i_size? (truncate in progress) */
2628        offset = i_size & (PAGE_CACHE_SIZE-1);
2629        if (page->index >= end_index+1 || !offset) {
2630                /*
2631                 * The page may have dirty, unmapped buffers.  For example,
2632                 * they may have been added in ext3_writepage().  Make them
2633                 * freeable here, so the page does not leak.
2634                 */
2635#if 0
2636                /* Not really sure about this  - do we need this ? */
2637                if (page->mapping->a_ops->invalidatepage)
2638                        page->mapping->a_ops->invalidatepage(page, offset);
2639#endif
2640                unlock_page(page);
2641                return 0; /* don't care */
2642        }
2643
2644        /*
2645         * The page straddles i_size.  It must be zeroed out on each and every
2646         * writepage invocation because it may be mmapped.  "A file is mapped
2647         * in multiples of the page size.  For a file that is not a multiple of
2648         * the  page size, the remaining memory is zeroed when mapped, and
2649         * writes to that region are not written out to the file."
2650         */
2651        zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2652out:
2653        ret = mpage_writepage(page, get_block, wbc);
2654        if (ret == -EAGAIN)
2655                ret = __block_write_full_page(inode, page, get_block, wbc,
2656                                              end_buffer_async_write);
2657        return ret;
2658}
2659EXPORT_SYMBOL(nobh_writepage);
2660
2661int nobh_truncate_page(struct address_space *mapping,
2662                        loff_t from, get_block_t *get_block)
2663{
2664        pgoff_t index = from >> PAGE_CACHE_SHIFT;
2665        unsigned offset = from & (PAGE_CACHE_SIZE-1);
2666        unsigned blocksize;
2667        sector_t iblock;
2668        unsigned length, pos;
2669        struct inode *inode = mapping->host;
2670        struct page *page;
2671        struct buffer_head map_bh;
2672        int err;
2673
2674        blocksize = 1 << inode->i_blkbits;
2675        length = offset & (blocksize - 1);
2676
2677        /* Block boundary? Nothing to do */
2678        if (!length)
2679                return 0;
2680
2681        length = blocksize - length;
2682        iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2683
2684        page = grab_cache_page(mapping, index);
2685        err = -ENOMEM;
2686        if (!page)
2687                goto out;
2688
2689        if (page_has_buffers(page)) {
2690has_buffers:
2691                unlock_page(page);
2692                page_cache_release(page);
2693                return block_truncate_page(mapping, from, get_block);
2694        }
2695
2696        /* Find the buffer that contains "offset" */
2697        pos = blocksize;
2698        while (offset >= pos) {
2699                iblock++;
2700                pos += blocksize;
2701        }
2702
2703        map_bh.b_size = blocksize;
2704        map_bh.b_state = 0;
2705        err = get_block(inode, iblock, &map_bh, 0);
2706        if (err)
2707                goto unlock;
2708        /* unmapped? It's a hole - nothing to do */
2709        if (!buffer_mapped(&map_bh))
2710                goto unlock;
2711
2712        /* Ok, it's mapped. Make sure it's up-to-date */
2713        if (!PageUptodate(page)) {
2714                err = mapping->a_ops->readpage(NULL, page);
2715                if (err) {
2716                        page_cache_release(page);
2717                        goto out;
2718                }
2719                lock_page(page);
2720                if (!PageUptodate(page)) {
2721                        err = -EIO;
2722                        goto unlock;
2723                }
2724                if (page_has_buffers(page))
2725                        goto has_buffers;
2726        }
2727        zero_user(page, offset, length);
2728        set_page_dirty(page);
2729        err = 0;
2730
2731unlock:
2732        unlock_page(page);
2733        page_cache_release(page);
2734out:
2735        return err;
2736}
2737EXPORT_SYMBOL(nobh_truncate_page);
2738
2739int block_truncate_page(struct address_space *mapping,
2740                        loff_t from, get_block_t *get_block)
2741{
2742        pgoff_t index = from >> PAGE_CACHE_SHIFT;
2743        unsigned offset = from & (PAGE_CACHE_SIZE-1);
2744        unsigned blocksize;
2745        sector_t iblock;
2746        unsigned length, pos;
2747        struct inode *inode = mapping->host;
2748        struct page *page;
2749        struct buffer_head *bh;
2750        int err;
2751
2752        blocksize = 1 << inode->i_blkbits;
2753        length = offset & (blocksize - 1);
2754
2755        /* Block boundary? Nothing to do */
2756        if (!length)
2757                return 0;
2758
2759        length = blocksize - length;
2760        iblock = (sector_t)index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2761        
2762        page = grab_cache_page(mapping, index);
2763        err = -ENOMEM;
2764        if (!page)
2765                goto out;
2766
2767        if (!page_has_buffers(page))
2768                create_empty_buffers(page, blocksize, 0);
2769
2770        /* Find the buffer that contains "offset" */
2771        bh = page_buffers(page);
2772        pos = blocksize;
2773        while (offset >= pos) {
2774                bh = bh->b_this_page;
2775                iblock++;
2776                pos += blocksize;
2777        }
2778
2779        err = 0;
2780        if (!buffer_mapped(bh)) {
2781                WARN_ON(bh->b_size != blocksize);
2782                err = get_block(inode, iblock, bh, 0);
2783                if (err)
2784                        goto unlock;
2785                /* unmapped? It's a hole - nothing to do */
2786                if (!buffer_mapped(bh))
2787                        goto unlock;
2788        }
2789
2790        /* Ok, it's mapped. Make sure it's up-to-date */
2791        if (PageUptodate(page))
2792                set_buffer_uptodate(bh);
2793
2794        if (!buffer_uptodate(bh) && !buffer_delay(bh) && !buffer_unwritten(bh)) {
2795                err = -EIO;
2796                ll_rw_block(READ, 1, &bh);
2797                wait_on_buffer(bh);
2798                /* Uhhuh. Read error. Complain and punt. */
2799                if (!buffer_uptodate(bh))
2800                        goto unlock;
2801        }
2802
2803        zero_user(page, offset, length);
2804        mark_buffer_dirty(bh);
2805        err = 0;
2806
2807unlock:
2808        unlock_page(page);
2809        page_cache_release(page);
2810out:
2811        return err;
2812}
2813EXPORT_SYMBOL(block_truncate_page);
2814
2815/*
2816 * The generic ->writepage function for buffer-backed address_spaces
2817 * this form passes in the end_io handler used to finish the IO.
2818 */
2819int block_write_full_page_endio(struct page *page, get_block_t *get_block,
2820                        struct writeback_control *wbc, bh_end_io_t *handler)
2821{
2822        struct inode * const inode = page->mapping->host;
2823        loff_t i_size = i_size_read(inode);
2824        const pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
2825        unsigned offset;
2826
2827        /* Is the page fully inside i_size? */
2828        if (page->index < end_index)
2829                return __block_write_full_page(inode, page, get_block, wbc,
2830                                               handler);
2831
2832        /* Is the page fully outside i_size? (truncate in progress) */
2833        offset = i_size & (PAGE_CACHE_SIZE-1);
2834        if (page->index >= end_index+1 || !offset) {
2835                /*
2836                 * The page may have dirty, unmapped buffers.  For example,
2837                 * they may have been added in ext3_writepage().  Make them
2838                 * freeable here, so the page does not leak.
2839                 */
2840                do_invalidatepage(page, 0);
2841                unlock_page(page);
2842                return 0; /* don't care */
2843        }
2844
2845        /*
2846         * The page straddles i_size.  It must be zeroed out on each and every
2847         * writepage invocation because it may be mmapped.  "A file is mapped
2848         * in multiples of the page size.  For a file that is not a multiple of
2849         * the  page size, the remaining memory is zeroed when mapped, and
2850         * writes to that region are not written out to the file."
2851         */
2852        zero_user_segment(page, offset, PAGE_CACHE_SIZE);
2853        return __block_write_full_page(inode, page, get_block, wbc, handler);
2854}
2855EXPORT_SYMBOL(block_write_full_page_endio);
2856
2857/*
2858 * The generic ->writepage function for buffer-backed address_spaces
2859 */
2860int block_write_full_page(struct page *page, get_block_t *get_block,
2861                        struct writeback_control *wbc)
2862{
2863        return block_write_full_page_endio(page, get_block, wbc,
2864                                           end_buffer_async_write);
2865}
2866EXPORT_SYMBOL(block_write_full_page);
2867
2868sector_t generic_block_bmap(struct address_space *mapping, sector_t block,
2869                            get_block_t *get_block)
2870{
2871        struct buffer_head tmp;
2872        struct inode *inode = mapping->host;
2873        tmp.b_state = 0;
2874        tmp.b_blocknr = 0;
2875        tmp.b_size = 1 << inode->i_blkbits;
2876        get_block(inode, block, &tmp, 0);
2877        return tmp.b_blocknr;
2878}
2879EXPORT_SYMBOL(generic_block_bmap);
2880
2881static void end_bio_bh_io_sync(struct bio *bio, int err)
2882{
2883        struct buffer_head *bh = bio->bi_private;
2884
2885        if (err == -EOPNOTSUPP) {
2886                set_bit(BIO_EOPNOTSUPP, &bio->bi_flags);
2887        }
2888
2889        if (unlikely (test_bit(BIO_QUIET,&bio->bi_flags)))
2890                set_bit(BH_Quiet, &bh->b_state);
2891
2892        bh->b_end_io(bh, test_bit(BIO_UPTODATE, &bio->bi_flags));
2893        bio_put(bio);
2894}
2895
2896/*
2897 * This allows us to do IO even on the odd last sectors
2898 * of a device, even if the bh block size is some multiple
2899 * of the physical sector size.
2900 *
2901 * We'll just truncate the bio to the size of the device,
2902 * and clear the end of the buffer head manually.
2903 *
2904 * Truly out-of-range accesses will turn into actual IO
2905 * errors, this only handles the "we need to be able to
2906 * do IO at the final sector" case.
2907 */
2908static void guard_bh_eod(int rw, struct bio *bio, struct buffer_head *bh)
2909{
2910        sector_t maxsector;
2911        unsigned bytes;
2912
2913        maxsector = i_size_read(bio->bi_bdev->bd_inode) >> 9;
2914        if (!maxsector)
2915                return;
2916
2917        /*
2918         * If the *whole* IO is past the end of the device,
2919         * let it through, and the IO layer will turn it into
2920         * an EIO.
2921         */
2922        if (unlikely(bio->bi_sector >= maxsector))
2923                return;
2924
2925        maxsector -= bio->bi_sector;
2926        bytes = bio->bi_size;
2927        if (likely((bytes >> 9) <= maxsector))
2928                return;
2929
2930        /* Uhhuh. We've got a bh that straddles the device size! */
2931        bytes = maxsector << 9;
2932
2933        /* Truncate the bio.. */
2934        bio->bi_size = bytes;
2935        bio->bi_io_vec[0].bv_len = bytes;
2936
2937        /* ..and clear the end of the buffer for reads */
2938        if ((rw & RW_MASK) == READ) {
2939                void *kaddr = kmap_atomic(bh->b_page);
2940                memset(kaddr + bh_offset(bh) + bytes, 0, bh->b_size - bytes);
2941                kunmap_atomic(kaddr);
2942                flush_dcache_page(bh->b_page);
2943        }
2944}
2945
2946int submit_bh(int rw, struct buffer_head * bh)
2947{
2948        struct bio *bio;
2949        int ret = 0;
2950
2951        BUG_ON(!buffer_locked(bh));
2952        BUG_ON(!buffer_mapped(bh));
2953        BUG_ON(!bh->b_end_io);
2954        BUG_ON(buffer_delay(bh));
2955        BUG_ON(buffer_unwritten(bh));
2956
2957        /*
2958         * Only clear out a write error when rewriting
2959         */
2960        if (test_set_buffer_req(bh) && (rw & WRITE))
2961                clear_buffer_write_io_error(bh);
2962
2963        /*
2964         * from here on down, it's all bio -- do the initial mapping,
2965         * submit_bio -> generic_make_request may further map this bio around
2966         */
2967        bio = bio_alloc(GFP_NOIO, 1);
2968
2969        bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
2970        bio->bi_bdev = bh->b_bdev;
2971        bio->bi_io_vec[0].bv_page = bh->b_page;
2972        bio->bi_io_vec[0].bv_len = bh->b_size;
2973        bio->bi_io_vec[0].bv_offset = bh_offset(bh);
2974
2975        bio->bi_vcnt = 1;
2976        bio->bi_idx = 0;
2977        bio->bi_size = bh->b_size;
2978
2979        bio->bi_end_io = end_bio_bh_io_sync;
2980        bio->bi_private = bh;
2981
2982        /* Take care of bh's that straddle the end of the device */
2983        guard_bh_eod(rw, bio, bh);
2984
2985        bio_get(bio);
2986        submit_bio(rw, bio);
2987
2988        if (bio_flagged(bio, BIO_EOPNOTSUPP))
2989                ret = -EOPNOTSUPP;
2990
2991        bio_put(bio);
2992        return ret;
2993}
2994EXPORT_SYMBOL(submit_bh);
2995
2996/**
2997 * ll_rw_block: low-level access to block devices (DEPRECATED)
2998 * @rw: whether to %READ or %WRITE or maybe %READA (readahead)
2999 * @nr: number of &struct buffer_heads in the array
3000 * @bhs: array of pointers to &struct buffer_head
3001 *
3002 * ll_rw_block() takes an array of pointers to &struct buffer_heads, and
3003 * requests an I/O operation on them, either a %READ or a %WRITE.  The third
3004 * %READA option is described in the documentation for generic_make_request()
3005 * which ll_rw_block() calls.
3006 *
3007 * This function drops any buffer that it cannot get a lock on (with the
3008 * BH_Lock state bit), any buffer that appears to be clean when doing a write
3009 * request, and any buffer that appears to be up-to-date when doing read
3010 * request.  Further it marks as clean buffers that are processed for
3011 * writing (the buffer cache won't assume that they are actually clean
3012 * until the buffer gets unlocked).
3013 *
3014 * ll_rw_block sets b_end_io to simple completion handler that marks
3015 * the buffer up-to-date (if approriate), unlocks the buffer and wakes
3016 * any waiters. 
3017 *
3018 * All of the buffers must be for the same device, and must also be a
3019 * multiple of the current approved size for the device.
3020 */
3021void ll_rw_block(int rw, int nr, struct buffer_head *bhs[])
3022{
3023        int i;
3024
3025        for (i = 0; i < nr; i++) {
3026                struct buffer_head *bh = bhs[i];
3027
3028                if (!trylock_buffer(bh))
3029                        continue;
3030                if (rw == WRITE) {
3031                        if (test_clear_buffer_dirty(bh)) {
3032                                bh->b_end_io = end_buffer_write_sync;
3033                                get_bh(bh);
3034                                submit_bh(WRITE, bh);
3035                                continue;
3036                        }
3037                } else {
3038                        if (!buffer_uptodate(bh)) {
3039                                bh->b_end_io = end_buffer_read_sync;
3040                                get_bh(bh);
3041                                submit_bh(rw, bh);
3042                                continue;
3043                        }
3044                }
3045                unlock_buffer(bh);
3046        }
3047}
3048EXPORT_SYMBOL(ll_rw_block);
3049
3050void write_dirty_buffer(struct buffer_head *bh, int rw)
3051{
3052        lock_buffer(bh);
3053        if (!test_clear_buffer_dirty(bh)) {
3054                unlock_buffer(bh);
3055                return;
3056        }
3057        bh->b_end_io = end_buffer_write_sync;
3058        get_bh(bh);
3059        submit_bh(rw, bh);
3060}
3061EXPORT_SYMBOL(write_dirty_buffer);
3062
3063/*
3064 * For a data-integrity writeout, we need to wait upon any in-progress I/O
3065 * and then start new I/O and then wait upon it.  The caller must have a ref on
3066 * the buffer_head.
3067 */
3068int __sync_dirty_buffer(struct buffer_head *bh, int rw)
3069{
3070        int ret = 0;
3071
3072        WARN_ON(atomic_read(&bh->b_count) < 1);
3073        lock_buffer(bh);
3074        if (test_clear_buffer_dirty(bh)) {
3075                get_bh(bh);
3076                bh->b_end_io = end_buffer_write_sync;
3077                ret = submit_bh(rw, bh);
3078                wait_on_buffer(bh);
3079                if (!ret && !buffer_uptodate(bh))
3080                        ret = -EIO;
3081        } else {
3082                unlock_buffer(bh);
3083        }
3084        return ret;
3085}
3086EXPORT_SYMBOL(__sync_dirty_buffer);
3087
3088int sync_dirty_buffer(struct buffer_head *bh)
3089{
3090        return __sync_dirty_buffer(bh, WRITE_SYNC);
3091}
3092EXPORT_SYMBOL(sync_dirty_buffer);
3093
3094/*
3095 * try_to_free_buffers() checks if all the buffers on this particular page
3096 * are unused, and releases them if so.
3097 *
3098 * Exclusion against try_to_free_buffers may be obtained by either
3099 * locking the page or by holding its mapping's private_lock.
3100 *
3101 * If the page is dirty but all the buffers are clean then we need to
3102 * be sure to mark the page clean as well.  This is because the page
3103 * may be against a block device, and a later reattachment of buffers
3104 * to a dirty page will set *all* buffers dirty.  Which would corrupt
3105 * filesystem data on the same device.
3106 *
3107 * The same applies to regular filesystem pages: if all the buffers are
3108 * clean then we set the page clean and proceed.  To do that, we require
3109 * total exclusion from __set_page_dirty_buffers().  That is obtained with
3110 * private_lock.
3111 *
3112 * try_to_free_buffers() is non-blocking.
3113 */
3114static inline int buffer_busy(struct buffer_head *bh)
3115{
3116        return atomic_read(&bh->b_count) |
3117                (bh->b_state & ((1 << BH_Dirty) | (1 << BH_Lock)));
3118}
3119
3120static int
3121drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
3122{
3123        struct buffer_head *head = page_buffers(page);
3124        struct buffer_head *bh;
3125
3126        bh = head;
3127        do {
3128                if (buffer_write_io_error(bh) && page->mapping)
3129                        set_bit(AS_EIO, &page->mapping->flags);
3130                if (buffer_busy(bh))
3131                        goto failed;
3132                bh = bh->b_this_page;
3133        } while (bh != head);
3134
3135        do {
3136                struct buffer_head *next = bh->b_this_page;
3137
3138                if (bh->b_assoc_map)
3139                        __remove_assoc_queue(bh);
3140                bh = next;
3141        } while (bh != head);
3142        *buffers_to_free = head;
3143        __clear_page_buffers(page);
3144        return 1;
3145failed:
3146        return 0;
3147}
3148
3149int try_to_free_buffers(struct page *page)
3150{
3151        struct address_space * const mapping = page->mapping;
3152        struct buffer_head *buffers_to_free = NULL;
3153        int ret = 0;
3154
3155        BUG_ON(!PageLocked(page));
3156        if (PageWriteback(page))
3157                return 0;
3158
3159        if (mapping == NULL) {          /* can this still happen? */
3160                ret = drop_buffers(page, &buffers_to_free);
3161                goto out;
3162        }
3163
3164        spin_lock(&mapping->private_lock);
3165        ret = drop_buffers(page, &buffers_to_free);
3166
3167        /*
3168         * If the filesystem writes its buffers by hand (eg ext3)
3169         * then we can have clean buffers against a dirty page.  We
3170         * clean the page here; otherwise the VM will never notice
3171         * that the filesystem did any IO at all.
3172         *
3173         * Also, during truncate, discard_buffer will have marked all
3174         * the page's buffers clean.  We discover that here and clean
3175         * the page also.
3176         *
3177         * private_lock must be held over this entire operation in order
3178         * to synchronise against __set_page_dirty_buffers and prevent the
3179         * dirty bit from being lost.
3180         */
3181        if (ret)
3182                cancel_dirty_page(page, PAGE_CACHE_SIZE);
3183        spin_unlock(&mapping->private_lock);
3184out:
3185        if (buffers_to_free) {
3186                struct buffer_head *bh = buffers_to_free;
3187
3188                do {
3189                        struct buffer_head *next = bh->b_this_page;
3190                        free_buffer_head(bh);
3191                        bh = next;
3192                } while (bh != buffers_to_free);
3193        }
3194        return ret;
3195}
3196EXPORT_SYMBOL(try_to_free_buffers);
3197
3198/*
3199 * There are no bdflush tunables left.  But distributions are
3200 * still running obsolete flush daemons, so we terminate them here.
3201 *
3202 * Use of bdflush() is deprecated and will be removed in a future kernel.
3203 * The `flush-X' kernel threads fully replace bdflush daemons and this call.
3204 */
3205SYSCALL_DEFINE2(bdflush, int, func, long, data)
3206{
3207        static int msg_count;
3208
3209        if (!capable(CAP_SYS_ADMIN))
3210                return -EPERM;
3211
3212        if (msg_count < 5) {
3213                msg_count++;
3214                printk(KERN_INFO
3215                        "warning: process `%s' used the obsolete bdflush"
3216                        " system call\n", current->comm);
3217                printk(KERN_INFO "Fix your initscripts?\n");
3218        }
3219
3220        if (func == 1)
3221                do_exit(0);
3222        return 0;
3223}
3224
3225/*
3226 * Buffer-head allocation
3227 */
3228static struct kmem_cache *bh_cachep __read_mostly;
3229
3230/*
3231 * Once the number of bh's in the machine exceeds this level, we start
3232 * stripping them in writeback.
3233 */
3234static int max_buffer_heads;
3235
3236int buffer_heads_over_limit;
3237
3238struct bh_accounting {
3239        int nr;                 /* Number of live bh's */
3240        int ratelimit;          /* Limit cacheline bouncing */
3241};
3242
3243static DEFINE_PER_CPU(struct bh_accounting, bh_accounting) = {0, 0};
3244
3245static void recalc_bh_state(void)
3246{
3247        int i;
3248        int tot = 0;
3249
3250        if (__this_cpu_inc_return(bh_accounting.ratelimit) - 1 < 4096)
3251                return;
3252        __this_cpu_write(bh_accounting.ratelimit, 0);
3253        for_each_online_cpu(i)
3254                tot += per_cpu(bh_accounting, i).nr;
3255        buffer_heads_over_limit = (tot > max_buffer_heads);
3256}
3257
3258struct buffer_head *alloc_buffer_head(gfp_t gfp_flags)
3259{
3260        struct buffer_head *ret = kmem_cache_zalloc(bh_cachep, gfp_flags);
3261        if (ret) {
3262                INIT_LIST_HEAD(&ret->b_assoc_buffers);
3263                preempt_disable();
3264                __this_cpu_inc(bh_accounting.nr);
3265                recalc_bh_state();
3266                preempt_enable();
3267        }
3268        return ret;
3269}
3270EXPORT_SYMBOL(alloc_buffer_head);
3271
3272void free_buffer_head(struct buffer_head *bh)
3273{
3274        BUG_ON(!list_empty(&bh->b_assoc_buffers));
3275        kmem_cache_free(bh_cachep, bh);
3276        preempt_disable();
3277        __this_cpu_dec(bh_accounting.nr);
3278        recalc_bh_state();
3279        preempt_enable();
3280}
3281EXPORT_SYMBOL(free_buffer_head);
3282
3283static void buffer_exit_cpu(int cpu)
3284{
3285        int i;
3286        struct bh_lru *b = &per_cpu(bh_lrus, cpu);
3287
3288        for (i = 0; i < BH_LRU_SIZE; i++) {
3289                brelse(b->bhs[i]);
3290                b->bhs[i] = NULL;
3291        }
3292        this_cpu_add(bh_accounting.nr, per_cpu(bh_accounting, cpu).nr);
3293        per_cpu(bh_accounting, cpu).nr = 0;
3294}
3295
3296static int buffer_cpu_notify(struct notifier_block *self,
3297                              unsigned long action, void *hcpu)
3298{
3299        if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
3300                buffer_exit_cpu((unsigned long)hcpu);
3301        return NOTIFY_OK;
3302}
3303
3304/**
3305 * bh_uptodate_or_lock - Test whether the buffer is uptodate
3306 * @bh: struct buffer_head
3307 *
3308 * Return true if the buffer is up-to-date and false,
3309 * with the buffer locked, if not.
3310 */
3311int bh_uptodate_or_lock(struct buffer_head *bh)
3312{
3313        if (!buffer_uptodate(bh)) {
3314                lock_buffer(bh);
3315                if (!buffer_uptodate(bh))
3316                        return 0;
3317                unlock_buffer(bh);
3318        }
3319        return 1;
3320}
3321EXPORT_SYMBOL(bh_uptodate_or_lock);
3322
3323/**
3324 * bh_submit_read - Submit a locked buffer for reading
3325 * @bh: struct buffer_head
3326 *
3327 * Returns zero on success and -EIO on error.
3328 */
3329int bh_submit_read(struct buffer_head *bh)
3330{
3331        BUG_ON(!buffer_locked(bh));
3332
3333        if (buffer_uptodate(bh)) {
3334                unlock_buffer(bh);
3335                return 0;
3336        }
3337
3338        get_bh(bh);
3339        bh->b_end_io = end_buffer_read_sync;
3340        submit_bh(READ, bh);
3341        wait_on_buffer(bh);
3342        if (buffer_uptodate(bh))
3343                return 0;
3344        return -EIO;
3345}
3346EXPORT_SYMBOL(bh_submit_read);
3347
3348void __init buffer_init(void)
3349{
3350        int nrpages;
3351
3352        bh_cachep = kmem_cache_create("buffer_head",
3353                        sizeof(struct buffer_head), 0,
3354                                (SLAB_RECLAIM_ACCOUNT|SLAB_PANIC|
3355                                SLAB_MEM_SPREAD),
3356                                NULL);
3357
3358        /*
3359         * Limit the bh occupancy to 10% of ZONE_NORMAL
3360         */
3361        nrpages = (nr_free_buffer_pages() * 10) / 100;
3362        max_buffer_heads = nrpages * (PAGE_SIZE / sizeof(struct buffer_head));
3363        hotcpu_notifier(buffer_cpu_notify, 0);
3364}
3365
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.