fs/ntfs/aops.c

   1 /**
   2  * aops.c - NTFS kernel address space operations and page cache handling.
   3  *          Part of the Linux-NTFS project.
   4  *
   5  * Copyright (c) 2001-2005 Anton Altaparmakov
   6  * Copyright (c) 2002 Richard Russon
   7  *
   8  * This program/include file is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU General Public License as published
  10  * by the Free Software Foundation; either version 2 of the License, or
  11  * (at your option) any later version.
  12  *
  13  * This program/include file is distributed in the hope that it will be
  14  * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  15  * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * along with this program (in the main directory of the Linux-NTFS
  20  * distribution in the file COPYING); if not, write to the Free Software
  21  * Foundation,Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  22  */
  23
  24 #include <linux/errno.h>
  25 #include <linux/mm.h>
  26 #include <linux/pagemap.h>
  27 #include <linux/swap.h>
  28 #include <linux/buffer_head.h>
  29 #include <linux/writeback.h>
  30
  31 #include "aops.h"
  32 #include "attrib.h"
  33 #include "debug.h"
  34 #include "inode.h"
  35 #include "mft.h"
  36 #include "runlist.h"
  37 #include "types.h"
  38 #include "ntfs.h"
  39
  40 /**
  41  * ntfs_end_buffer_async_read - async io completion for reading attributes
  42  * @bh:         buffer head on which io is completed
  43  * @uptodate:   whether @bh is now uptodate or not
  44  *
  45  * Asynchronous I/O completion handler for reading pages belonging to the
  46  * attribute address space of an inode.  The inodes can either be files or
  47  * directories or they can be fake inodes describing some attribute.
  48  *
  49  * If NInoMstProtected(), perform the post read mst fixups when all IO on the
  50  * page has been completed and mark the page uptodate or set the error bit on
  51  * the page.  To determine the size of the records that need fixing up, we
  52  * cheat a little bit by setting the index_block_size in ntfs_inode to the ntfs
  53  * record size, and index_block_size_bits, to the log(base 2) of the ntfs
  54  * record size.
  55  */
  56 static void ntfs_end_buffer_async_read(struct buffer_head *bh, int uptodate)
  57 {
  58         static DEFINE_SPINLOCK(page_uptodate_lock);
  59         unsigned long flags;
  60         struct buffer_head *tmp;
  61         struct page *page;
  62         ntfs_inode *ni;
  63         int page_uptodate = 1;
  64
  65         page = bh->b_page;
  66         ni = NTFS_I(page->mapping->host);
  67
  68         if (likely(uptodate)) {
  69                 s64 file_ofs, initialized_size;
  70
  71                 set_buffer_uptodate(bh);
  72
  73                 file_ofs = ((s64)page->index << PAGE_CACHE_SHIFT) +
  74                                 bh_offset(bh);
  75                 read_lock_irqsave(&ni->size_lock, flags);
  76                 initialized_size = ni->initialized_size;
  77                 read_unlock_irqrestore(&ni->size_lock, flags);
  78                 /* Check for the current buffer head overflowing. */
  79                 if (file_ofs + bh->b_size > initialized_size) {
  80                         char *addr;
  81                         int ofs = 0;
  82
  83                         if (file_ofs < initialized_size)
  84                                 ofs = initialized_size - file_ofs;
  85                         addr = kmap_atomic(page, KM_BIO_SRC_IRQ);
  86                         memset(addr + bh_offset(bh) + ofs, 0, bh->b_size - ofs);
  87                         flush_dcache_page(page);
  88                         kunmap_atomic(addr, KM_BIO_SRC_IRQ);
  89                 }
  90         } else {
  91                 clear_buffer_uptodate(bh);
  92                 ntfs_error(ni->vol->sb, "Buffer I/O error, logical block %llu.",
  93                                 (unsigned long long)bh->b_blocknr);
  94                 SetPageError(page);
  95         }
  96         spin_lock_irqsave(&page_uptodate_lock, flags);
  97         clear_buffer_async_read(bh);
  98         unlock_buffer(bh);
  99         tmp = bh;
 100         do {
 101                 if (!buffer_uptodate(tmp))
 102                         page_uptodate = 0;
 103                 if (buffer_async_read(tmp)) {
 104                         if (likely(buffer_locked(tmp)))
 105                                 goto still_busy;
 106                         /* Async buffers must be locked. */
 107                         BUG();
 108                 }
 109                 tmp = tmp->b_this_page;
 110         } while (tmp != bh);
 111         spin_unlock_irqrestore(&page_uptodate_lock, flags);
 112         /*
 113          * If none of the buffers had errors then we can set the page uptodate,
 114          * but we first have to perform the post read mst fixups, if the
 115          * attribute is mst protected, i.e. if NInoMstProteced(ni) is true.
 116          * Note we ignore fixup errors as those are detected when
 117          * map_mft_record() is called which gives us per record granularity
 118          * rather than per page granularity.
 119          */
 120         if (!NInoMstProtected(ni)) {
 121                 if (likely(page_uptodate && !PageError(page)))
 122                         SetPageUptodate(page);
 123         } else {
 124                 char *addr;
 125                 unsigned int i, recs;
 126                 u32 rec_size;
 127
 128                 rec_size = ni->itype.index.block_size;
 129                 recs = PAGE_CACHE_SIZE / rec_size;
 130                 /* Should have been verified before we got here... */
 131                 BUG_ON(!recs);
 132                 addr = kmap_atomic(page, KM_BIO_SRC_IRQ);
 133                 for (i = 0; i < recs; i++)
 134                         post_read_mst_fixup((NTFS_RECORD*)(addr +
 135                                         i * rec_size), rec_size);
 136                 flush_dcache_page(page);
 137                 kunmap_atomic(addr, KM_BIO_SRC_IRQ);
 138                 if (likely(page_uptodate && !PageError(page)))
 139                         SetPageUptodate(page);
 140         }
 141         unlock_page(page);
 142         return;
 143 still_busy:
 144         spin_unlock_irqrestore(&page_uptodate_lock, flags);
 145         return;
 146 }
 147
 148 /**
 149  * ntfs_read_block - fill a @page of an address space with data
 150  * @page:       page cache page to fill with data
 151  *
 152  * Fill the page @page of the address space belonging to the @page->host inode.
 153  * We read each buffer asynchronously and when all buffers are read in, our io
 154  * completion handler ntfs_end_buffer_read_async(), if required, automatically
 155  * applies the mst fixups to the page before finally marking it uptodate and
 156  * unlocking it.
 157  *
 158  * We only enforce allocated_size limit because i_size is checked for in
 159  * generic_file_read().
 160  *
 161  * Return 0 on success and -errno on error.
 162  *
 163  * Contains an adapted version of fs/buffer.c::block_read_full_page().
 164  */
 165 static int ntfs_read_block(struct page *page)
 166 {
 167         VCN vcn;
 168         LCN lcn;
 169         ntfs_inode *ni;
 170         ntfs_volume *vol;
 171         runlist_element *rl;
 172         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
 173         sector_t iblock, lblock, zblock;
 174         unsigned long flags;
 175         unsigned int blocksize, vcn_ofs;
 176         int i, nr;
 177         unsigned char blocksize_bits;
 178
 179         ni = NTFS_I(page->mapping->host);
 180         vol = ni->vol;
 181
 182         /* $MFT/$DATA must have its complete runlist in memory at all times. */
 183         BUG_ON(!ni->runlist.rl && !ni->mft_no && !NInoAttr(ni));
 184
 185         blocksize_bits = VFS_I(ni)->i_blkbits;
 186         blocksize = 1 << blocksize_bits;
 187
 188         if (!page_has_buffers(page))
 189                 create_empty_buffers(page, blocksize, 0);
 190         bh = head = page_buffers(page);
 191         if (unlikely(!bh)) {
 192                 unlock_page(page);
 193                 return -ENOMEM;
 194         }
 195
 196         iblock = (s64)page->index << (PAGE_CACHE_SHIFT - blocksize_bits);
 197         read_lock_irqsave(&ni->size_lock, flags);
 198         lblock = (ni->allocated_size + blocksize - 1) >> blocksize_bits;
 199         zblock = (ni->initialized_size + blocksize - 1) >> blocksize_bits;
 200         read_unlock_irqrestore(&ni->size_lock, flags);
 201
 202         /* Loop through all the buffers in the page. */
 203         rl = NULL;
 204         nr = i = 0;
 205         do {
 206                 u8 *kaddr;
 207                 int err;
 208
 209                 if (unlikely(buffer_uptodate(bh)))
 210                         continue;
 211                 if (unlikely(buffer_mapped(bh))) {
 212                         arr[nr++] = bh;
 213                         continue;
 214                 }
 215                 err = 0;
 216                 bh->b_bdev = vol->sb->s_bdev;
 217                 /* Is the block within the allowed limits? */
 218                 if (iblock < lblock) {
 219                         BOOL is_retry = FALSE;
 220
 221                         /* Convert iblock into corresponding vcn and offset. */
 222                         vcn = (VCN)iblock << blocksize_bits >>
 223                                         vol->cluster_size_bits;
 224                         vcn_ofs = ((VCN)iblock << blocksize_bits) &
 225                                         vol->cluster_size_mask;
 226                         if (!rl) {
 227 lock_retry_remap:
 228                                 down_read(&ni->runlist.lock);
 229                                 rl = ni->runlist.rl;
 230                         }
 231                         if (likely(rl != NULL)) {
 232                                 /* Seek to element containing target vcn. */
 233                                 while (rl->length && rl[1].vcn <= vcn)
 234                                         rl++;
 235                                 lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
 236                         } else
 237                                 lcn = LCN_RL_NOT_MAPPED;
 238                         /* Successful remap. */
 239                         if (lcn >= 0) {
 240                                 /* Setup buffer head to correct block. */
 241                                 bh->b_blocknr = ((lcn << vol->cluster_size_bits)
 242                                                 + vcn_ofs) >> blocksize_bits;
 243                                 set_buffer_mapped(bh);
 244                                 /* Only read initialized data blocks. */
 245                                 if (iblock < zblock) {
 246                                         arr[nr++] = bh;
 247                                         continue;
 248                                 }
 249                                 /* Fully non-initialized data block, zero it. */
 250                                 goto handle_zblock;
 251                         }
 252                         /* It is a hole, need to zero it. */
 253                         if (lcn == LCN_HOLE)
 254                                 goto handle_hole;
 255                         /* If first try and runlist unmapped, map and retry. */
 256                         if (!is_retry && lcn == LCN_RL_NOT_MAPPED) {
 257                                 is_retry = TRUE;
 258                                 /*
 259                                  * Attempt to map runlist, dropping lock for
 260                                  * the duration.
 261                                  */
 262                                 up_read(&ni->runlist.lock);
 263                                 err = ntfs_map_runlist(ni, vcn);
 264                                 if (likely(!err))
 265                                         goto lock_retry_remap;
 266                                 rl = NULL;
 267                         } else if (!rl)
 268                                 up_read(&ni->runlist.lock);
 269                         /*
 270                          * If buffer is outside the runlist, treat it as a
 271                          * hole.  This can happen due to concurrent truncate
 272                          * for example.
 273                          */
 274                         if (err == -ENOENT || lcn == LCN_ENOENT) {
 275                                 err = 0;
 276                                 goto handle_hole;
 277                         }
 278                         /* Hard error, zero out region. */
 279                         if (!err)
 280                                 err = -EIO;
 281                         bh->b_blocknr = -1;
 282                         SetPageError(page);
 283                         ntfs_error(vol->sb, "Failed to read from inode 0x%lx, "
 284                                         "attribute type 0x%x, vcn 0x%llx, "
 285                                         "offset 0x%x because its location on "
 286                                         "disk could not be determined%s "
 287                                         "(error code %i).", ni->mft_no,
 288                                         ni->type, (unsigned long long)vcn,
 289                                         vcn_ofs, is_retry ? " even after "
 290                                         "retrying" : "", err);
 291                 }
 292                 /*
 293                  * Either iblock was outside lblock limits or
 294                  * ntfs_rl_vcn_to_lcn() returned error.  Just zero that portion
 295                  * of the page and set the buffer uptodate.
 296                  */
 297 handle_hole:
 298                 bh->b_blocknr = -1UL;
 299                 clear_buffer_mapped(bh);
 300 handle_zblock:
 301                 kaddr = kmap_atomic(page, KM_USER0);
 302                 memset(kaddr + i * blocksize, 0, blocksize);
 303                 kunmap_atomic(kaddr, KM_USER0);
 304                 flush_dcache_page(page);
 305                 if (likely(!err))
 306                         set_buffer_uptodate(bh);
 307         } while (i++, iblock++, (bh = bh->b_this_page) != head);
 308
 309         /* Release the lock if we took it. */
 310         if (rl)
 311                 up_read(&ni->runlist.lock);
 312
 313         /* Check we have at least one buffer ready for i/o. */
 314         if (nr) {
 315                 struct buffer_head *tbh;
 316
 317                 /* Lock the buffers. */
 318                 for (i = 0; i < nr; i++) {
 319                         tbh = arr[i];
 320                         lock_buffer(tbh);
 321                         tbh->b_end_io = ntfs_end_buffer_async_read;
 322                         set_buffer_async_read(tbh);
 323                 }
 324                 /* Finally, start i/o on the buffers. */
 325                 for (i = 0; i < nr; i++) {
 326                         tbh = arr[i];
 327                         if (likely(!buffer_uptodate(tbh)))
 328                                 submit_bh(READ, tbh);
 329                         else
 330                                 ntfs_end_buffer_async_read(tbh, 1);
 331                 }
 332                 return 0;
 333         }
 334         /* No i/o was scheduled on any of the buffers. */
 335         if (likely(!PageError(page)))
 336                 SetPageUptodate(page);
 337         else /* Signal synchronous i/o error. */
 338                 nr = -EIO;
 339         unlock_page(page);
 340         return nr;
 341 }
 342
 343 /**
 344  * ntfs_readpage - fill a @page of a @file with data from the device
 345  * @file:       open file to which the page @page belongs or NULL
 346  * @page:       page cache page to fill with data
 347  *
 348  * For non-resident attributes, ntfs_readpage() fills the @page of the open
 349  * file @file by calling the ntfs version of the generic block_read_full_page()
 350  * function, ntfs_read_block(), which in turn creates and reads in the buffers
 351  * associated with the page asynchronously.
 352  *
 353  * For resident attributes, OTOH, ntfs_readpage() fills @page by copying the
 354  * data from the mft record (which at this stage is most likely in memory) and
 355  * fills the remainder with zeroes. Thus, in this case, I/O is synchronous, as
 356  * even if the mft record is not cached at this point in time, we need to wait
 357  * for it to be read in before we can do the copy.
 358  *
 359  * Return 0 on success and -errno on error.
 360  */
 361 static int ntfs_readpage(struct file *file, struct page *page)
 362 {
 363         ntfs_inode *ni, *base_ni;
 364         u8 *kaddr;
 365         ntfs_attr_search_ctx *ctx;
 366         MFT_RECORD *mrec;
 367         unsigned long flags;
 368         u32 attr_len;
 369         int err = 0;
 370
 371 retry_readpage:
 372         BUG_ON(!PageLocked(page));
 373         /*
 374          * This can potentially happen because we clear PageUptodate() during
 375          * ntfs_writepage() of MstProtected() attributes.
 376          */
 377         if (PageUptodate(page)) {
 378                 unlock_page(page);
 379                 return 0;
 380         }
 381         ni = NTFS_I(page->mapping->host);
 382         /*
 383          * Only $DATA attributes can be encrypted and only unnamed $DATA
 384          * attributes can be compressed.  Index root can have the flags set but
 385          * this means to create compressed/encrypted files, not that the
 386          * attribute is compressed/encrypted.
 387          */
 388         if (ni->type != AT_INDEX_ROOT) {
 389                 /* If attribute is encrypted, deny access, just like NT4. */
 390                 if (NInoEncrypted(ni)) {
 391                         BUG_ON(ni->type != AT_DATA);
 392                         err = -EACCES;
 393                         goto err_out;
 394                 }
 395                 /* Compressed data streams are handled in compress.c. */
 396                 if (NInoNonResident(ni) && NInoCompressed(ni)) {
 397                         BUG_ON(ni->type != AT_DATA);
 398                         BUG_ON(ni->name_len);
 399                         return ntfs_read_compressed_block(page);
 400                 }
 401         }
 402         /* NInoNonResident() == NInoIndexAllocPresent() */
 403         if (NInoNonResident(ni)) {
 404                 /* Normal, non-resident data stream. */
 405                 return ntfs_read_block(page);
 406         }
 407         /*
 408          * Attribute is resident, implying it is not compressed or encrypted.
 409          * This also means the attribute is smaller than an mft record and
 410          * hence smaller than a page, so can simply zero out any pages with
 411          * index above 0.  Note the attribute can actually be marked compressed
 412          * but if it is resident the actual data is not compressed so we are
 413          * ok to ignore the compressed flag here.
 414          */
 415         if (unlikely(page->index > 0)) {
 416                 kaddr = kmap_atomic(page, KM_USER0);
 417                 memset(kaddr, 0, PAGE_CACHE_SIZE);
 418                 flush_dcache_page(page);
 419                 kunmap_atomic(kaddr, KM_USER0);
 420                 goto done;
 421         }
 422         if (!NInoAttr(ni))
 423                 base_ni = ni;
 424         else
 425                 base_ni = ni->ext.base_ntfs_ino;
 426         /* Map, pin, and lock the mft record. */
 427         mrec = map_mft_record(base_ni);
 428         if (IS_ERR(mrec)) {
 429                 err = PTR_ERR(mrec);
 430                 goto err_out;
 431         }
 432         /*
 433          * If a parallel write made the attribute non-resident, drop the mft
 434          * record and retry the readpage.
 435          */
 436         if (unlikely(NInoNonResident(ni))) {
 437                 unmap_mft_record(base_ni);
 438                 goto retry_readpage;
 439         }
 440         ctx = ntfs_attr_get_search_ctx(base_ni, mrec);
 441         if (unlikely(!ctx)) {
 442                 err = -ENOMEM;
 443                 goto unm_err_out;
 444         }
 445         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
 446                         CASE_SENSITIVE, 0, NULL, 0, ctx);
 447         if (unlikely(err))
 448                 goto put_unm_err_out;
 449         attr_len = le32_to_cpu(ctx->attr->data.resident.value_length);
 450         read_lock_irqsave(&ni->size_lock, flags);
 451         if (unlikely(attr_len > ni->initialized_size))
 452                 attr_len = ni->initialized_size;
 453         read_unlock_irqrestore(&ni->size_lock, flags);
 454         kaddr = kmap_atomic(page, KM_USER0);
 455         /* Copy the data to the page. */
 456         memcpy(kaddr, (u8*)ctx->attr +
 457                         le16_to_cpu(ctx->attr->data.resident.value_offset),
 458                         attr_len);
 459         /* Zero the remainder of the page. */
 460         memset(kaddr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
 461         flush_dcache_page(page);
 462         kunmap_atomic(kaddr, KM_USER0);
 463 put_unm_err_out:
 464         ntfs_attr_put_search_ctx(ctx);
 465 unm_err_out:
 466         unmap_mft_record(base_ni);
 467 done:
 468         SetPageUptodate(page);
 469 err_out:
 470         unlock_page(page);
 471         return err;
 472 }
 473
 474 #ifdef NTFS_RW
 475
 476 /**
 477  * ntfs_write_block - write a @page to the backing store
 478  * @page:       page cache page to write out
 479  * @wbc:        writeback control structure
 480  *
 481  * This function is for writing pages belonging to non-resident, non-mst
 482  * protected attributes to their backing store.
 483  *
 484  * For a page with buffers, map and write the dirty buffers asynchronously
 485  * under page writeback. For a page without buffers, create buffers for the
 486  * page, then proceed as above.
 487  *
 488  * If a page doesn't have buffers the page dirty state is definitive. If a page
 489  * does have buffers, the page dirty state is just a hint, and the buffer dirty
 490  * state is definitive. (A hint which has rules: dirty buffers against a clean
 491  * page is illegal. Other combinations are legal and need to be handled. In
 492  * particular a dirty page containing clean buffers for example.)
 493  *
 494  * Return 0 on success and -errno on error.
 495  *
 496  * Based on ntfs_read_block() and __block_write_full_page().
 497  */
 498 static int ntfs_write_block(struct page *page, struct writeback_control *wbc)
 499 {
 500         VCN vcn;
 501         LCN lcn;
 502         s64 initialized_size;
 503         loff_t i_size;
 504         sector_t block, dblock, iblock;
 505         struct inode *vi;
 506         ntfs_inode *ni;
 507         ntfs_volume *vol;
 508         runlist_element *rl;
 509         struct buffer_head *bh, *head;
 510         unsigned long flags;
 511         unsigned int blocksize, vcn_ofs;
 512         int err;
 513         BOOL need_end_writeback;
 514         unsigned char blocksize_bits;
 515
 516         vi = page->mapping->host;
 517         ni = NTFS_I(vi);
 518         vol = ni->vol;
 519
 520         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
 521                         "0x%lx.", ni->mft_no, ni->type, page->index);
 522
 523         BUG_ON(!NInoNonResident(ni));
 524         BUG_ON(NInoMstProtected(ni));
 525
 526         blocksize_bits = vi->i_blkbits;
 527         blocksize = 1 << blocksize_bits;
 528
 529         if (!page_has_buffers(page)) {
 530                 BUG_ON(!PageUptodate(page));
 531                 create_empty_buffers(page, blocksize,
 532                                 (1 << BH_Uptodate) | (1 << BH_Dirty));
 533         }
 534         bh = head = page_buffers(page);
 535         if (unlikely(!bh)) {
 536                 ntfs_warning(vol->sb, "Error allocating page buffers. "
 537                                 "Redirtying page so we try again later.");
 538                 /*
 539                  * Put the page back on mapping->dirty_pages, but leave its
 540                  * buffer's dirty state as-is.
 541                  */
 542                 redirty_page_for_writepage(wbc, page);
 543                 unlock_page(page);
 544                 return 0;
 545         }
 546
 547         /* NOTE: Different naming scheme to ntfs_read_block()! */
 548
 549         /* The first block in the page. */
 550         block = (s64)page->index << (PAGE_CACHE_SHIFT - blocksize_bits);
 551
 552         read_lock_irqsave(&ni->size_lock, flags);
 553         i_size = i_size_read(vi);
 554         initialized_size = ni->initialized_size;
 555         read_unlock_irqrestore(&ni->size_lock, flags);
 556
 557         /* The first out of bounds block for the data size. */
 558         dblock = (i_size + blocksize - 1) >> blocksize_bits;
 559
 560         /* The last (fully or partially) initialized block. */
 561         iblock = initialized_size >> blocksize_bits;
 562
 563         /*
 564          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
 565          * here, and the (potentially unmapped) buffers may become dirty at
 566          * any time.  If a buffer becomes dirty here after we've inspected it
 567          * then we just miss that fact, and the page stays dirty.
 568          *
 569          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
 570          * handle that here by just cleaning them.
 571          */
 572
 573         /*
 574          * Loop through all the buffers in the page, mapping all the dirty
 575          * buffers to disk addresses and handling any aliases from the
 576          * underlying block device's mapping.
 577          */
 578         rl = NULL;
 579         err = 0;
 580         do {
 581                 BOOL is_retry = FALSE;
 582
 583                 if (unlikely(block >= dblock)) {
 584                         /*
 585                          * Mapped buffers outside i_size will occur, because
 586                          * this page can be outside i_size when there is a
 587                          * truncate in progress. The contents of such buffers
 588                          * were zeroed by ntfs_writepage().
 589                          *
 590                          * FIXME: What about the small race window where
 591                          * ntfs_writepage() has not done any clearing because
 592                          * the page was within i_size but before we get here,
 593                          * vmtruncate() modifies i_size?
 594                          */
 595                         clear_buffer_dirty(bh);
 596                         set_buffer_uptodate(bh);
 597                         continue;
 598                 }
 599
 600                 /* Clean buffers are not written out, so no need to map them. */
 601                 if (!buffer_dirty(bh))
 602                         continue;
 603
 604                 /* Make sure we have enough initialized size. */
 605                 if (unlikely((block >= iblock) &&
 606                                 (initialized_size < i_size))) {
 607                         /*
 608                          * If this page is fully outside initialized size, zero
 609                          * out all pages between the current initialized size
 610                          * and the current page. Just use ntfs_readpage() to do
 611                          * the zeroing transparently.
 612                          */
 613                         if (block > iblock) {
 614                                 // TODO:
 615                                 // For each page do:
 616                                 // - read_cache_page()
 617                                 // Again for each page do:
 618                                 // - wait_on_page_locked()
 619                                 // - Check (PageUptodate(page) &&
 620                                 //                      !PageError(page))
 621                                 // Update initialized size in the attribute and
 622                                 // in the inode.
 623                                 // Again, for each page do:
 624                                 //      __set_page_dirty_buffers();
 625                                 // page_cache_release()
 626                                 // We don't need to wait on the writes.
 627                                 // Update iblock.
 628                         }
 629                         /*
 630                          * The current page straddles initialized size. Zero
 631                          * all non-uptodate buffers and set them uptodate (and
 632                          * dirty?). Note, there aren't any non-uptodate buffers
 633                          * if the page is uptodate.
 634                          * FIXME: For an uptodate page, the buffers may need to
 635                          * be written out because they were not initialized on
 636                          * disk before.
 637                          */
 638                         if (!PageUptodate(page)) {
 639                                 // TODO:
 640                                 // Zero any non-uptodate buffers up to i_size.
 641                                 // Set them uptodate and dirty.
 642                         }
 643                         // TODO:
 644                         // Update initialized size in the attribute and in the
 645                         // inode (up to i_size).
 646                         // Update iblock.
 647                         // FIXME: This is inefficient. Try to batch the two
 648                         // size changes to happen in one go.
 649                         ntfs_error(vol->sb, "Writing beyond initialized size "
 650                                         "is not supported yet. Sorry.");
 651                         err = -EOPNOTSUPP;
 652                         break;
 653                         // Do NOT set_buffer_new() BUT DO clear buffer range
 654                         // outside write request range.
 655                         // set_buffer_uptodate() on complete buffers as well as
 656                         // set_buffer_dirty().
 657                 }
 658
 659                 /* No need to map buffers that are already mapped. */
 660                 if (buffer_mapped(bh))
 661                         continue;
 662
 663                 /* Unmapped, dirty buffer. Need to map it. */
 664                 bh->b_bdev = vol->sb->s_bdev;
 665
 666                 /* Convert block into corresponding vcn and offset. */
 667                 vcn = (VCN)block << blocksize_bits;
 668                 vcn_ofs = vcn & vol->cluster_size_mask;
 669                 vcn >>= vol->cluster_size_bits;
 670                 if (!rl) {
 671 lock_retry_remap:
 672                         down_read(&ni->runlist.lock);
 673                         rl = ni->runlist.rl;
 674                 }
 675                 if (likely(rl != NULL)) {
 676                         /* Seek to element containing target vcn. */
 677                         while (rl->length && rl[1].vcn <= vcn)
 678                                 rl++;
 679                         lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
 680                 } else
 681                         lcn = LCN_RL_NOT_MAPPED;
 682                 /* Successful remap. */
 683                 if (lcn >= 0) {
 684                         /* Setup buffer head to point to correct block. */
 685                         bh->b_blocknr = ((lcn << vol->cluster_size_bits) +
 686                                         vcn_ofs) >> blocksize_bits;
 687                         set_buffer_mapped(bh);
 688                         continue;
 689                 }
 690                 /* It is a hole, need to instantiate it. */
 691                 if (lcn == LCN_HOLE) {
 692                         u8 *kaddr;
 693                         unsigned long *bpos, *bend;
 694
 695                         /* Check if the buffer is zero. */
 696                         kaddr = kmap_atomic(page, KM_USER0);
 697                         bpos = (unsigned long *)(kaddr + bh_offset(bh));
 698                         bend = (unsigned long *)((u8*)bpos + blocksize);
 699                         do {
 700                                 if (unlikely(*bpos))
 701                                         break;
 702                         } while (likely(++bpos < bend));
 703                         kunmap_atomic(kaddr, KM_USER0);
 704                         if (bpos == bend) {
 705                                 /*
 706                                  * Buffer is zero and sparse, no need to write
 707                                  * it.
 708                                  */
 709                                 bh->b_blocknr = -1;
 710                                 clear_buffer_dirty(bh);
 711                                 continue;
 712                         }
 713                         // TODO: Instantiate the hole.
 714                         // clear_buffer_new(bh);
 715                         // unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 716                         ntfs_error(vol->sb, "Writing into sparse regions is "
 717                                         "not supported yet. Sorry.");
 718                         err = -EOPNOTSUPP;
 719                         break;
 720                 }
 721                 /* If first try and runlist unmapped, map and retry. */
 722                 if (!is_retry && lcn == LCN_RL_NOT_MAPPED) {
 723                         is_retry = TRUE;
 724                         /*
 725                          * Attempt to map runlist, dropping lock for
 726                          * the duration.
 727                          */
 728                         up_read(&ni->runlist.lock);
 729                         err = ntfs_map_runlist(ni, vcn);
 730                         if (likely(!err))
 731                                 goto lock_retry_remap;
 732                         rl = NULL;
 733                 } else if (!rl)
 734                         up_read(&ni->runlist.lock);
 735                 /*
 736                  * If buffer is outside the runlist, truncate has cut it out
 737                  * of the runlist.  Just clean and clear the buffer and set it
 738                  * uptodate so it can get discarded by the VM.
 739                  */
 740                 if (err == -ENOENT || lcn == LCN_ENOENT) {
 741                         u8 *kaddr;
 742
 743                         bh->b_blocknr = -1;
 744                         clear_buffer_dirty(bh);
 745                         kaddr = kmap_atomic(page, KM_USER0);
 746                         memset(kaddr + bh_offset(bh), 0, blocksize);
 747                         kunmap_atomic(kaddr, KM_USER0);
 748                         flush_dcache_page(page);
 749                         set_buffer_uptodate(bh);
 750                         err = 0;
 751                         continue;
 752                 }
 753                 /* Failed to map the buffer, even after retrying. */
 754                 if (!err)
 755                         err = -EIO;
 756                 bh->b_blocknr = -1;
 757                 ntfs_error(vol->sb, "Failed to write to inode 0x%lx, "
 758                                 "attribute type 0x%x, vcn 0x%llx, offset 0x%x "
 759                                 "because its location on disk could not be "
 760                                 "determined%s (error code %i).", ni->mft_no,
 761                                 ni->type, (unsigned long long)vcn,
 762                                 vcn_ofs, is_retry ? " even after "
 763                                 "retrying" : "", err);
 764                 break;
 765         } while (block++, (bh = bh->b_this_page) != head);
 766
 767         /* Release the lock if we took it. */
 768         if (rl)
 769                 up_read(&ni->runlist.lock);
 770
 771         /* For the error case, need to reset bh to the beginning. */
 772         bh = head;
 773
 774         /* Just an optimization, so ->readpage() is not called later. */
 775         if (unlikely(!PageUptodate(page))) {
 776                 int uptodate = 1;
 777                 do {
 778                         if (!buffer_uptodate(bh)) {
 779                                 uptodate = 0;
 780                                 bh = head;
 781                                 break;
 782                         }
 783                 } while ((bh = bh->b_this_page) != head);
 784                 if (uptodate)
 785                         SetPageUptodate(page);
 786         }
 787
 788         /* Setup all mapped, dirty buffers for async write i/o. */
 789         do {
 790                 if (buffer_mapped(bh) && buffer_dirty(bh)) {
 791                         lock_buffer(bh);
 792                         if (test_clear_buffer_dirty(bh)) {
 793                                 BUG_ON(!buffer_uptodate(bh));
 794                                 mark_buffer_async_write(bh);
 795                         } else
 796                                 unlock_buffer(bh);
 797                 } else if (unlikely(err)) {
 798                         /*
 799                          * For the error case. The buffer may have been set
 800                          * dirty during attachment to a dirty page.
 801                          */
 802                         if (err != -ENOMEM)
 803                                 clear_buffer_dirty(bh);
 804                 }
 805         } while ((bh = bh->b_this_page) != head);
 806
 807         if (unlikely(err)) {
 808                 // TODO: Remove the -EOPNOTSUPP check later on...
 809                 if (unlikely(err == -EOPNOTSUPP))
 810                         err = 0;
 811                 else if (err == -ENOMEM) {
 812                         ntfs_warning(vol->sb, "Error allocating memory. "
 813                                         "Redirtying page so we try again "
 814                                         "later.");
 815                         /*
 816                          * Put the page back on mapping->dirty_pages, but
 817                          * leave its buffer's dirty state as-is.
 818                          */
 819                         redirty_page_for_writepage(wbc, page);
 820                         err = 0;
 821                 } else
 822                         SetPageError(page);
 823         }
 824
 825         BUG_ON(PageWriteback(page));
 826         set_page_writeback(page);       /* Keeps try_to_free_buffers() away. */
 827
 828         /* Submit the prepared buffers for i/o. */
 829         need_end_writeback = TRUE;
 830         do {
 831                 struct buffer_head *next = bh->b_this_page;
 832                 if (buffer_async_write(bh)) {
 833                         submit_bh(WRITE, bh);
 834                         need_end_writeback = FALSE;
 835                 }
 836                 bh = next;
 837         } while (bh != head);
 838         unlock_page(page);
 839
 840         /* If no i/o was started, need to end_page_writeback(). */
 841         if (unlikely(need_end_writeback))
 842                 end_page_writeback(page);
 843
 844         ntfs_debug("Done.");
 845         return err;
 846 }
 847
 848 /**
 849  * ntfs_write_mst_block - write a @page to the backing store
 850  * @page:       page cache page to write out
 851  * @wbc:        writeback control structure
 852  *
 853  * This function is for writing pages belonging to non-resident, mst protected
 854  * attributes to their backing store.  The only supported attributes are index
 855  * allocation and $MFT/$DATA.  Both directory inodes and index inodes are
 856  * supported for the index allocation case.
 857  *
 858  * The page must remain locked for the duration of the write because we apply
 859  * the mst fixups, write, and then undo the fixups, so if we were to unlock the
 860  * page before undoing the fixups, any other user of the page will see the
 861  * page contents as corrupt.
 862  *
 863  * We clear the page uptodate flag for the duration of the function to ensure
 864  * exclusion for the $MFT/$DATA case against someone mapping an mft record we
 865  * are about to apply the mst fixups to.
 866  *
 867  * Return 0 on success and -errno on error.
 868  *
 869  * Based on ntfs_write_block(), ntfs_mft_writepage(), and
 870  * write_mft_record_nolock().
 871  */
 872 static int ntfs_write_mst_block(struct page *page,
 873                 struct writeback_control *wbc)
 874 {
 875         sector_t block, dblock, rec_block;
 876         struct inode *vi = page->mapping->host;
 877         ntfs_inode *ni = NTFS_I(vi);
 878         ntfs_volume *vol = ni->vol;
 879         u8 *kaddr;
 880         unsigned int rec_size = ni->itype.index.block_size;
 881         ntfs_inode *locked_nis[PAGE_CACHE_SIZE / rec_size];
 882         struct buffer_head *bh, *head, *tbh, *rec_start_bh;
 883         struct buffer_head *bhs[MAX_BUF_PER_PAGE];
 884         runlist_element *rl;
 885         int i, nr_locked_nis, nr_recs, nr_bhs, max_bhs, bhs_per_rec, err, err2;
 886         unsigned bh_size, rec_size_bits;
 887         BOOL sync, is_mft, page_is_dirty, rec_is_dirty;
 888         unsigned char bh_size_bits;
 889
 890         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
 891                         "0x%lx.", vi->i_ino, ni->type, page->index);
 892         BUG_ON(!NInoNonResident(ni));
 893         BUG_ON(!NInoMstProtected(ni));
 894         is_mft = (S_ISREG(vi->i_mode) && !vi->i_ino);
 895         /*
 896          * NOTE: ntfs_write_mst_block() would be called for $MFTMirr if a page
 897          * in its page cache were to be marked dirty.  However this should
 898          * never happen with the current driver and considering we do not
 899          * handle this case here we do want to BUG(), at least for now.
 900          */
 901         BUG_ON(!(is_mft || S_ISDIR(vi->i_mode) ||
 902                         (NInoAttr(ni) && ni->type == AT_INDEX_ALLOCATION)));
 903         bh_size_bits = vi->i_blkbits;
 904         bh_size = 1 << bh_size_bits;
 905         max_bhs = PAGE_CACHE_SIZE / bh_size;
 906         BUG_ON(!max_bhs);
 907         BUG_ON(max_bhs > MAX_BUF_PER_PAGE);
 908
 909         /* Were we called for sync purposes? */
 910         sync = (wbc->sync_mode == WB_SYNC_ALL);
 911
 912         /* Make sure we have mapped buffers. */
 913         BUG_ON(!page_has_buffers(page));
 914         bh = head = page_buffers(page);
 915         BUG_ON(!bh);
 916
 917         rec_size_bits = ni->itype.index.block_size_bits;
 918         BUG_ON(!(PAGE_CACHE_SIZE >> rec_size_bits));
 919         bhs_per_rec = rec_size >> bh_size_bits;
 920         BUG_ON(!bhs_per_rec);
 921
 922         /* The first block in the page. */
 923         rec_block = block = (sector_t)page->index <<
 924                         (PAGE_CACHE_SHIFT - bh_size_bits);
 925
 926         /* The first out of bounds block for the data size. */
 927         dblock = (i_size_read(vi) + bh_size - 1) >> bh_size_bits;
 928
 929         rl = NULL;
 930         err = err2 = nr_bhs = nr_recs = nr_locked_nis = 0;
 931         page_is_dirty = rec_is_dirty = FALSE;
 932         rec_start_bh = NULL;
 933         do {
 934                 BOOL is_retry = FALSE;
 935
 936                 if (likely(block < rec_block)) {
 937                         if (unlikely(block >= dblock)) {
 938                                 clear_buffer_dirty(bh);
 939                                 set_buffer_uptodate(bh);
 940                                 continue;
 941                         }
 942                         /*
 943                          * This block is not the first one in the record.  We
 944                          * ignore the buffer's dirty state because we could
 945                          * have raced with a parallel mark_ntfs_record_dirty().
 946                          */
 947                         if (!rec_is_dirty)
 948                                 continue;
 949                         if (unlikely(err2)) {
 950                                 if (err2 != -ENOMEM)
 951                                         clear_buffer_dirty(bh);
 952                                 continue;
 953                         }
 954                 } else /* if (block == rec_block) */ {
 955                         BUG_ON(block > rec_block);
 956                         /* This block is the first one in the record. */
 957                         rec_block += bhs_per_rec;
 958                         err2 = 0;
 959                         if (unlikely(block >= dblock)) {
 960                                 clear_buffer_dirty(bh);
 961                                 continue;
 962                         }
 963                         if (!buffer_dirty(bh)) {
 964                                 /* Clean records are not written out. */
 965                                 rec_is_dirty = FALSE;
 966                                 continue;
 967                         }
 968                         rec_is_dirty = TRUE;
 969                         rec_start_bh = bh;
 970                 }
 971                 /* Need to map the buffer if it is not mapped already. */
 972                 if (unlikely(!buffer_mapped(bh))) {
 973                         VCN vcn;
 974                         LCN lcn;
 975                         unsigned int vcn_ofs;
 976
 977                         bh->b_bdev = vol->sb->s_bdev;
 978                         /* Obtain the vcn and offset of the current block. */
 979                         vcn = (VCN)block << bh_size_bits;
 980                         vcn_ofs = vcn & vol->cluster_size_mask;
 981                         vcn >>= vol->cluster_size_bits;
 982                         if (!rl) {
 983 lock_retry_remap:
 984                                 down_read(&ni->runlist.lock);
 985                                 rl = ni->runlist.rl;
 986                         }
 987                         if (likely(rl != NULL)) {
 988                                 /* Seek to element containing target vcn. */
 989                                 while (rl->length && rl[1].vcn <= vcn)
 990                                         rl++;
 991                                 lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
 992                         } else
 993                                 lcn = LCN_RL_NOT_MAPPED;
 994                         /* Successful remap. */
 995                         if (likely(lcn >= 0)) {
 996                                 /* Setup buffer head to correct block. */
 997                                 bh->b_blocknr = ((lcn <<
 998                                                 vol->cluster_size_bits) +
 999                                                 vcn_ofs) >> bh_size_bits;
1000                                 set_buffer_mapped(bh);
1001                         } else {
1002                                 /*
1003                                  * Remap failed.  Retry to map the runlist once
1004                                  * unless we are working on $MFT which always
1005                                  * has the whole of its runlist in memory.
1006                                  */
1007                                 if (!is_mft && !is_retry &&
1008                                                 lcn == LCN_RL_NOT_MAPPED) {
1009                                         is_retry = TRUE;
1010                                         /*
1011                                          * Attempt to map runlist, dropping
1012                                          * lock for the duration.
1013                                          */
1014                                         up_read(&ni->runlist.lock);
1015                                         err2 = ntfs_map_runlist(ni, vcn);
1016                                         if (likely(!err2))
1017                                                 goto lock_retry_remap;
1018                                         if (err2 == -ENOMEM)
1019                                                 page_is_dirty = TRUE;
1020                                         lcn = err2;
1021                                 } else {
1022                                         err2 = -EIO;
1023                                         if (!rl)
1024                                                 up_read(&ni->runlist.lock);
1025                                 }
1026                                 /* Hard error.  Abort writing this record. */
1027                                 if (!err || err == -ENOMEM)
1028                                         err = err2;
1029                                 bh->b_blocknr = -1;
1030                                 ntfs_error(vol->sb, "Cannot write ntfs record "
1031                                                 "0x%llx (inode 0x%lx, "
1032                                                 "attribute type 0x%x) because "
1033                                                 "its location on disk could "
1034                                                 "not be determined (error "
1035                                                 "code %lli).",
1036                                                 (long long)block <<
1037                                                 bh_size_bits >>
1038                                                 vol->mft_record_size_bits,
1039                                                 ni->mft_no, ni->type,
1040                                                 (long long)lcn);
1041                                 /*
1042                                  * If this is not the first buffer, remove the
1043                                  * buffers in this record from the list of
1044                                  * buffers to write and clear their dirty bit
1045                                  * if not error -ENOMEM.
1046                                  */
1047                                 if (rec_start_bh != bh) {
1048                                         while (bhs[--nr_bhs] != rec_start_bh)
1049                                                 ;
1050                                         if (err2 != -ENOMEM) {
1051                                                 do {
1052                                                         clear_buffer_dirty(
1053                                                                 rec_start_bh);
1054                                                 } while ((rec_start_bh =
1055                                                                 rec_start_bh->
1056                                                                 b_this_page) !=
1057                                                                 bh);
1058                                         }
1059                                 }
1060                                 continue;
1061                         }
1062                 }
1063                 BUG_ON(!buffer_uptodate(bh));
1064                 BUG_ON(nr_bhs >= max_bhs);
1065                 bhs[nr_bhs++] = bh;
1066         } while (block++, (bh = bh->b_this_page) != head);
1067         if (unlikely(rl))
1068                 up_read(&ni->runlist.lock);
1069         /* If there were no dirty buffers, we are done. */
1070         if (!nr_bhs)
1071                 goto done;
1072         /* Map the page so we can access its contents. */
1073         kaddr = kmap(page);
1074         /* Clear the page uptodate flag whilst the mst fixups are applied. */
1075         BUG_ON(!PageUptodate(page));
1076         ClearPageUptodate(page);
1077         for (i = 0; i < nr_bhs; i++) {
1078                 unsigned int ofs;
1079
1080                 /* Skip buffers which are not at the beginning of records. */
1081                 if (i % bhs_per_rec)
1082                         continue;
1083                 tbh = bhs[i];
1084                 ofs = bh_offset(tbh);
1085                 if (is_mft) {
1086                         ntfs_inode *tni;
1087                         unsigned long mft_no;
1088
1089                         /* Get the mft record number. */
1090                         mft_no = (((s64)page->index << PAGE_CACHE_SHIFT) + ofs)
1091                                         >> rec_size_bits;
1092                         /* Check whether to write this mft record. */
1093                         tni = NULL;
1094                         if (!ntfs_may_write_mft_record(vol, mft_no,
1095                                         (MFT_RECORD*)(kaddr + ofs), &tni)) {
1096                                 /*
1097                                  * The record should not be written.  This
1098                                  * means we need to redirty the page before
1099                                  * returning.
1100                                  */
1101                                 page_is_dirty = TRUE;
1102                                 /*
1103                                  * Remove the buffers in this mft record from
1104                                  * the list of buffers to write.
1105                                  */
1106                                 do {
1107                                         bhs[i] = NULL;
1108                                 } while (++i % bhs_per_rec);
1109                                 continue;
1110                         }
1111                         /*
1112                          * The record should be written.  If a locked ntfs
1113                          * inode was returned, add it to the array of locked
1114                          * ntfs inodes.
1115                          */
1116                         if (tni)
1117                                 locked_nis[nr_locked_nis++] = tni;
1118                 }
1119                 /* Apply the mst protection fixups. */
1120                 err2 = pre_write_mst_fixup((NTFS_RECORD*)(kaddr + ofs),
1121                                 rec_size);
1122                 if (unlikely(err2)) {
1123                         if (!err || err == -ENOMEM)
1124                                 err = -EIO;
1125                         ntfs_error(vol->sb, "Failed to apply mst fixups "
1126                                         "(inode 0x%lx, attribute type 0x%x, "
1127                                         "page index 0x%lx, page offset 0x%x)!"
1128                                         "  Unmount and run chkdsk.", vi->i_ino,
1129                                         ni->type, page->index, ofs);
1130                         /*
1131                          * Mark all the buffers in this record clean as we do
1132                          * not want to write corrupt data to disk.
1133                          */
1134                         do {
1135                                 clear_buffer_dirty(bhs[i]);
1136                                 bhs[i] = NULL;
1137                         } while (++i % bhs_per_rec);
1138                         continue;
1139                 }
1140                 nr_recs++;
1141         }
1142         /* If no records are to be written out, we are done. */
1143         if (!nr_recs)
1144                 goto unm_done;
1145         flush_dcache_page(page);
1146         /* Lock buffers and start synchronous write i/o on them. */
1147         for (i = 0; i < nr_bhs; i++) {
1148                 tbh = bhs[i];
1149                 if (!tbh)
1150                         continue;
1151                 if (unlikely(test_set_buffer_locked(tbh)))
1152                         BUG();
1153                 /* The buffer dirty state is now irrelevant, just clean it. */
1154                 clear_buffer_dirty(tbh);
1155                 BUG_ON(!buffer_uptodate(tbh));
1156                 BUG_ON(!buffer_mapped(tbh));
1157                 get_bh(tbh);
1158                 tbh->b_end_io = end_buffer_write_sync;
1159                 submit_bh(WRITE, tbh);
1160         }
1161         /* Synchronize the mft mirror now if not @sync. */
1162         if (is_mft && !sync)
1163                 goto do_mirror;
1164 do_wait:
1165         /* Wait on i/o completion of buffers. */
1166         for (i = 0; i < nr_bhs; i++) {
1167                 tbh = bhs[i];
1168                 if (!tbh)
1169                         continue;
1170                 wait_on_buffer(tbh);
1171                 if (unlikely(!buffer_uptodate(tbh))) {
1172                         ntfs_error(vol->sb, "I/O error while writing ntfs "
1173                                         "record buffer (inode 0x%lx, "
1174                                         "attribute type 0x%x, page index "
1175                                         "0x%lx, page offset 0x%lx)!  Unmount "
1176                                         "and run chkdsk.", vi->i_ino, ni->type,
1177                                         page->index, bh_offset(tbh));
1178                         if (!err || err == -ENOMEM)
1179                                 err = -EIO;
1180                         /*
1181                          * Set the buffer uptodate so the page and buffer
1182                          * states do not become out of sync.
1183                          */
1184                         set_buffer_uptodate(tbh);
1185                 }
1186         }
1187         /* If @sync, now synchronize the mft mirror. */
1188         if (is_mft && sync) {
1189 do_mirror:
1190                 for (i = 0; i < nr_bhs; i++) {
1191                         unsigned long mft_no;
1192                         unsigned int ofs;
1193
1194                         /*
1195                          * Skip buffers which are not at the beginning of
1196                          * records.
1197                          */
1198                         if (i % bhs_per_rec)
1199                                 continue;
1200                         tbh = bhs[i];
1201                         /* Skip removed buffers (and hence records). */
1202                         if (!tbh)
1203                                 continue;
1204                         ofs = bh_offset(tbh);
1205                         /* Get the mft record number. */
1206                         mft_no = (((s64)page->index << PAGE_CACHE_SHIFT) + ofs)
1207                                         >> rec_size_bits;
1208                         if (mft_no < vol->mftmirr_size)
1209                                 ntfs_sync_mft_mirror(vol, mft_no,
1210                                                 (MFT_RECORD*)(kaddr + ofs),
1211                                                 sync);
1212                 }
1213                 if (!sync)
1214                         goto do_wait;
1215         }
1216         /* Remove the mst protection fixups again. */
1217         for (i = 0; i < nr_bhs; i++) {
1218                 if (!(i % bhs_per_rec)) {
1219                         tbh = bhs[i];
1220                         if (!tbh)
1221                                 continue;
1222                         post_write_mst_fixup((NTFS_RECORD*)(kaddr +
1223                                         bh_offset(tbh)));
1224                 }
1225         }
1226         flush_dcache_page(page);
1227 unm_done:
1228         /* Unlock any locked inodes. */
1229         while (nr_locked_nis-- > 0) {
1230                 ntfs_inode *tni, *base_tni;
1231
1232                 tni = locked_nis[nr_locked_nis];
1233                 /* Get the base inode. */
1234                 down(&tni->extent_lock);
1235                 if (tni->nr_extents >= 0)
1236                         base_tni = tni;
1237                 else {
1238                         base_tni = tni->ext.base_ntfs_ino;
1239                         BUG_ON(!base_tni);
1240                 }
1241                 up(&tni->extent_lock);
1242                 ntfs_debug("Unlocking %s inode 0x%lx.",
1243                                 tni == base_tni ? "base" : "extent",
1244                                 tni->mft_no);
1245                 up(&tni->mrec_lock);
1246                 atomic_dec(&tni->count);
1247                 iput(VFS_I(base_tni));
1248         }
1249         SetPageUptodate(page);
1250         kunmap(page);
1251 done:
1252         if (unlikely(err && err != -ENOMEM)) {
1253                 /*
1254                  * Set page error if there is only one ntfs record in the page.
1255                  * Otherwise we would loose per-record granularity.
1256                  */
1257                 if (ni->itype.index.block_size == PAGE_CACHE_SIZE)
1258                         SetPageError(page);
1259                 NVolSetErrors(vol);
1260         }
1261         if (page_is_dirty) {
1262                 ntfs_debug("Page still contains one or more dirty ntfs "
1263                                 "records.  Redirtying the page starting at "
1264                                 "record 0x%lx.", page->index <<
1265                                 (PAGE_CACHE_SHIFT - rec_size_bits));
1266                 redirty_page_for_writepage(wbc, page);
1267                 unlock_page(page);
1268         } else {
1269                 /*
1270                  * Keep the VM happy.  This must be done otherwise the
1271                  * radix-tree tag PAGECACHE_TAG_DIRTY remains set even though
1272                  * the page is clean.
1273                  */
1274                 BUG_ON(PageWriteback(page));
1275                 set_page_writeback(page);
1276                 unlock_page(page);
1277                 end_page_writeback(page);
1278         }
1279         if (likely(!err))
1280                 ntfs_debug("Done.");
1281         return err;
1282 }
1283
1284 /**
1285  * ntfs_writepage - write a @page to the backing store
1286  * @page:       page cache page to write out
1287  * @wbc:        writeback control structure
1288  *
1289  * This is called from the VM when it wants to have a dirty ntfs page cache
1290  * page cleaned.  The VM has already locked the page and marked it clean.
1291  *
1292  * For non-resident attributes, ntfs_writepage() writes the @page by calling
1293  * the ntfs version of the generic block_write_full_page() function,
1294  * ntfs_write_block(), which in turn if necessary creates and writes the
1295  * buffers associated with the page asynchronously.
1296  *
1297  * For resident attributes, OTOH, ntfs_writepage() writes the @page by copying
1298  * the data to the mft record (which at this stage is most likely in memory).
1299  * The mft record is then marked dirty and written out asynchronously via the
1300  * vfs inode dirty code path for the inode the mft record belongs to or via the
1301  * vm page dirty code path for the page the mft record is in.
1302  *
1303  * Based on ntfs_readpage() and fs/buffer.c::block_write_full_page().
1304  *
1305  * Return 0 on success and -errno on error.
1306  */
1307 static int ntfs_writepage(struct page *page, struct writeback_control *wbc)
1308 {
1309         loff_t i_size;
1310         struct inode *vi = page->mapping->host;
1311         ntfs_inode *base_ni = NULL, *ni = NTFS_I(vi);
1312         char *kaddr;
1313         ntfs_attr_search_ctx *ctx = NULL;
1314         MFT_RECORD *m = NULL;
1315         u32 attr_len;
1316         int err;
1317
1318 retry_writepage:
1319         BUG_ON(!PageLocked(page));
1320         i_size = i_size_read(vi);
1321         /* Is the page fully outside i_size? (truncate in progress) */
1322         if (unlikely(page->index >= (i_size + PAGE_CACHE_SIZE - 1) >>
1323                         PAGE_CACHE_SHIFT)) {
1324                 /*
1325                  * The page may have dirty, unmapped buffers.  Make them
1326                  * freeable here, so the page does not leak.
1327                  */
1328                 block_invalidatepage(page, 0);
1329                 unlock_page(page);
1330                 ntfs_debug("Write outside i_size - truncated?");
1331                 return 0;
1332         }
1333         /*
1334          * Only $DATA attributes can be encrypted and only unnamed $DATA
1335          * attributes can be compressed.  Index root can have the flags set but
1336          * this means to create compressed/encrypted files, not that the
1337          * attribute is compressed/encrypted.
1338          */
1339         if (ni->type != AT_INDEX_ROOT) {
1340                 /* If file is encrypted, deny access, just like NT4. */
1341                 if (NInoEncrypted(ni)) {
1342                         unlock_page(page);
1343                         BUG_ON(ni->type != AT_DATA);
1344                         ntfs_debug("Denying write access to encrypted "
1345                                         "file.");
1346                         return -EACCES;
1347                 }
1348                 /* Compressed data streams are handled in compress.c. */
1349                 if (NInoNonResident(ni) && NInoCompressed(ni)) {
1350                         BUG_ON(ni->type != AT_DATA);
1351                         BUG_ON(ni->name_len);
1352                         // TODO: Implement and replace this with
1353                         // return ntfs_write_compressed_block(page);
1354                         unlock_page(page);
1355                         ntfs_error(vi->i_sb, "Writing to compressed files is "
1356                                         "not supported yet.  Sorry.");
1357                         return -EOPNOTSUPP;
1358                 }
1359                 // TODO: Implement and remove this check.
1360                 if (NInoNonResident(ni) && NInoSparse(ni)) {
1361                         unlock_page(page);
1362                         ntfs_error(vi->i_sb, "Writing to sparse files is not "
1363                                         "supported yet.  Sorry.");
1364                         return -EOPNOTSUPP;
1365                 }
1366         }
1367         /* NInoNonResident() == NInoIndexAllocPresent() */
1368         if (NInoNonResident(ni)) {
1369                 /* We have to zero every time due to mmap-at-end-of-file. */
1370                 if (page->index >= (i_size >> PAGE_CACHE_SHIFT)) {
1371                         /* The page straddles i_size. */
1372                         unsigned int ofs = i_size & ~PAGE_CACHE_MASK;
1373                         kaddr = kmap_atomic(page, KM_USER0);
1374                         memset(kaddr + ofs, 0, PAGE_CACHE_SIZE - ofs);
1375                         flush_dcache_page(page);
1376                         kunmap_atomic(kaddr, KM_USER0);
1377                 }
1378                 /* Handle mst protected attributes. */
1379                 if (NInoMstProtected(ni))
1380                         return ntfs_write_mst_block(page, wbc);
1381                 /* Normal, non-resident data stream. */
1382                 return ntfs_write_block(page, wbc);
1383         }
1384         /*
1385          * Attribute is resident, implying it is not compressed, encrypted, or
1386          * mst protected.  This also means the attribute is smaller than an mft
1387          * record and hence smaller than a page, so can simply return error on
1388          * any pages with index above 0.  Note the attribute can actually be
1389          * marked compressed but if it is resident the actual data is not
1390          * compressed so we are ok to ignore the compressed flag here.
1391          */
1392         BUG_ON(page_has_buffers(page));
1393         BUG_ON(!PageUptodate(page));
1394         if (unlikely(page->index > 0)) {
1395                 ntfs_error(vi->i_sb, "BUG()! page->index (0x%lx) > 0.  "
1396                                 "Aborting write.", page->index);
1397                 BUG_ON(PageWriteback(page));
1398                 set_page_writeback(page);
1399                 unlock_page(page);
1400                 end_page_writeback(page);
1401                 return -EIO;
1402         }
1403         if (!NInoAttr(ni))
1404                 base_ni = ni;
1405         else
1406                 base_ni = ni->ext.base_ntfs_ino;
1407         /* Map, pin, and lock the mft record. */
1408         m = map_mft_record(base_ni);
1409         if (IS_ERR(m)) {
1410                 err = PTR_ERR(m);
1411                 m = NULL;
1412                 ctx = NULL;
1413                 goto err_out;
1414         }
1415         /*
1416          * If a parallel write made the attribute non-resident, drop the mft
1417          * record and retry the writepage.
1418          */
1419         if (unlikely(NInoNonResident(ni))) {
1420                 unmap_mft_record(base_ni);
1421                 goto retry_writepage;
1422         }
1423         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1424         if (unlikely(!ctx)) {
1425                 err = -ENOMEM;
1426                 goto err_out;
1427         }
1428         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1429                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1430         if (unlikely(err))
1431                 goto err_out;
1432         /*
1433          * Keep the VM happy.  This must be done otherwise the radix-tree tag
1434          * PAGECACHE_TAG_DIRTY remains set even though the page is clean.
1435          */
1436         BUG_ON(PageWriteback(page));
1437         set_page_writeback(page);
1438         unlock_page(page);
1439         /*
1440          * Here, we do not need to zero the out of bounds area everytime
1441          * because the below memcpy() already takes care of the
1442          * mmap-at-end-of-file requirements.  If the file is converted to a
1443          * non-resident one, then the code path use is switched to the
1444          * non-resident one where the zeroing happens on each ntfs_writepage()
1445          * invocation.
1446          */
1447         attr_len = le32_to_cpu(ctx->attr->data.resident.value_length);
1448         i_size = i_size_read(vi);
1449         if (unlikely(attr_len > i_size)) {
1450                 attr_len = i_size;
1451                 ctx->attr->data.resident.value_length = cpu_to_le32(attr_len);
1452         }
1453         kaddr = kmap_atomic(page, KM_USER0);
1454         /* Copy the data from the page to the mft record. */
1455         memcpy((u8*)ctx->attr +
1456                         le16_to_cpu(ctx->attr->data.resident.value_offset),
1457                         kaddr, attr_len);
1458         flush_dcache_mft_record_page(ctx->ntfs_ino);
1459         /* Zero out of bounds area in the page cache page. */
1460         memset(kaddr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
1461         flush_dcache_page(page);
1462         kunmap_atomic(kaddr, KM_USER0);
1463
1464         end_page_writeback(page);
1465
1466         /* Mark the mft record dirty, so it gets written back. */
1467         mark_mft_record_dirty(ctx->ntfs_ino);
1468         ntfs_attr_put_search_ctx(ctx);
1469         unmap_mft_record(base_ni);
1470         return 0;
1471 err_out:
1472         if (err == -ENOMEM) {
1473                 ntfs_warning(vi->i_sb, "Error allocating memory. Redirtying "
1474                                 "page so we try again later.");
1475                 /*
1476                  * Put the page back on mapping->dirty_pages, but leave its
1477                  * buffers' dirty state as-is.
1478                  */
1479                 redirty_page_for_writepage(wbc, page);
1480                 err = 0;
1481         } else {
1482                 ntfs_error(vi->i_sb, "Resident attribute write failed with "
1483                                 "error %i.", err);
1484                 SetPageError(page);
1485                 NVolSetErrors(ni->vol);
1486                 make_bad_inode(vi);
1487         }
1488         unlock_page(page);
1489         if (ctx)
1490                 ntfs_attr_put_search_ctx(ctx);
1491         if (m)
1492                 unmap_mft_record(base_ni);
1493         return err;
1494 }
1495
1496 /**
1497  * ntfs_prepare_nonresident_write -
1498  *
1499  */
1500 static int ntfs_prepare_nonresident_write(struct page *page,
1501                 unsigned from, unsigned to)
1502 {
1503         VCN vcn;
1504         LCN lcn;
1505         s64 initialized_size;
1506         loff_t i_size;
1507         sector_t block, ablock, iblock;
1508         struct inode *vi;
1509         ntfs_inode *ni;
1510         ntfs_volume *vol;
1511         runlist_element *rl;
1512         struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
1513         unsigned long flags;
1514         unsigned int vcn_ofs, block_start, block_end, blocksize;
1515         int err;
1516         BOOL is_retry;
1517         unsigned char blocksize_bits;
1518
1519         vi = page->mapping->host;
1520         ni = NTFS_I(vi);
1521         vol = ni->vol;
1522
1523         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
1524                         "0x%lx, from = %u, to = %u.", ni->mft_no, ni->type,
1525                         page->index, from, to);
1526
1527         BUG_ON(!NInoNonResident(ni));
1528
1529         blocksize_bits = vi->i_blkbits;
1530         blocksize = 1 << blocksize_bits;
1531
1532         /*
1533          * create_empty_buffers() will create uptodate/dirty buffers if the
1534          * page is uptodate/dirty.
1535          */
1536         if (!page_has_buffers(page))
1537                 create_empty_buffers(page, blocksize, 0);
1538         bh = head = page_buffers(page);
1539         if (unlikely(!bh))
1540                 return -ENOMEM;
1541
1542         /* The first block in the page. */
1543         block = (s64)page->index << (PAGE_CACHE_SHIFT - blocksize_bits);
1544
1545         read_lock_irqsave(&ni->size_lock, flags);
1546         /*
1547          * The first out of bounds block for the allocated size.  No need to
1548          * round up as allocated_size is in multiples of cluster size and the
1549          * minimum cluster size is 512 bytes, which is equal to the smallest
1550          * blocksize.
1551          */
1552         ablock = ni->allocated_size >> blocksize_bits;
1553         i_size = i_size_read(vi);
1554         initialized_size = ni->initialized_size;
1555         read_unlock_irqrestore(&ni->size_lock, flags);
1556
1557         /* The last (fully or partially) initialized block. */
1558         iblock = initialized_size >> blocksize_bits;
1559
1560         /* Loop through all the buffers in the page. */
1561         block_start = 0;
1562         rl = NULL;
1563         err = 0;
1564         do {
1565                 block_end = block_start + blocksize;
1566                 /*
1567                  * If buffer @bh is outside the write, just mark it uptodate
1568                  * if the page is uptodate and continue with the next buffer.
1569                  */
1570                 if (block_end <= from || block_start >= to) {
1571                         if (PageUptodate(page)) {
1572                                 if (!buffer_uptodate(bh))
1573                                         set_buffer_uptodate(bh);
1574                         }
1575                         continue;
1576                 }
1577                 /*
1578                  * @bh is at least partially being written to.
1579                  * Make sure it is not marked as new.
1580                  */
1581                 //if (buffer_new(bh))
1582                 //      clear_buffer_new(bh);
1583
1584                 if (block >= ablock) {
1585                         // TODO: block is above allocated_size, need to
1586                         // allocate it. Best done in one go to accommodate not
1587                         // only block but all above blocks up to and including:
1588                         // ((page->index << PAGE_CACHE_SHIFT) + to + blocksize
1589                         // - 1) >> blobksize_bits. Obviously will need to round
1590                         // up to next cluster boundary, too. This should be
1591                         // done with a helper function, so it can be reused.
1592                         ntfs_error(vol->sb, "Writing beyond allocated size "
1593                                         "is not supported yet. Sorry.");
1594                         err = -EOPNOTSUPP;
1595                         goto err_out;
1596                         // Need to update ablock.
1597                         // Need to set_buffer_new() on all block bhs that are
1598                         // newly allocated.
1599                 }
1600                 /*
1601                  * Now we have enough allocated size to fulfill the whole
1602                  * request, i.e. block < ablock is true.
1603                  */
1604                 if (unlikely((block >= iblock) &&
1605                                 (initialized_size < i_size))) {
1606                         /*
1607                          * If this page is fully outside initialized size, zero
1608                          * out all pages between the current initialized size
1609                          * and the current page. Just use ntfs_readpage() to do
1610                          * the zeroing transparently.
1611                          */
1612                         if (block > iblock) {
1613                                 // TODO:
1614                                 // For each page do:
1615                                 // - read_cache_page()
1616                                 // Again for each page do:
1617                                 // - wait_on_page_locked()
1618                                 // - Check (PageUptodate(page) &&
1619                                 //                      !PageError(page))
1620                                 // Update initialized size in the attribute and
1621                                 // in the inode.
1622                                 // Again, for each page do:
1623                                 //      __set_page_dirty_buffers();
1624                                 // page_cache_release()
1625                                 // We don't need to wait on the writes.
1626                                 // Update iblock.
1627                         }
1628                         /*
1629                          * The current page straddles initialized size. Zero
1630                          * all non-uptodate buffers and set them uptodate (and
1631                          * dirty?). Note, there aren't any non-uptodate buffers
1632                          * if the page is uptodate.
1633                          * FIXME: For an uptodate page, the buffers may need to
1634                          * be written out because they were not initialized on
1635                          * disk before.
1636                          */
1637                         if (!PageUptodate(page)) {
1638                                 // TODO:
1639                                 // Zero any non-uptodate buffers up to i_size.
1640                                 // Set them uptodate and dirty.
1641                         }
1642                         // TODO:
1643                         // Update initialized size in the attribute and in the
1644                         // inode (up to i_size).
1645                         // Update iblock.
1646                         // FIXME: This is inefficient. Try to batch the two
1647                         // size changes to happen in one go.
1648                         ntfs_error(vol->sb, "Writing beyond initialized size "
1649                                         "is not supported yet. Sorry.");
1650                         err = -EOPNOTSUPP;
1651                         goto err_out;
1652                         // Do NOT set_buffer_new() BUT DO clear buffer range
1653                         // outside write request range.
1654                         // set_buffer_uptodate() on complete buffers as well as
1655                         // set_buffer_dirty().
1656                 }
1657
1658                 /* Need to map unmapped buffers. */
1659                 if (!buffer_mapped(bh)) {
1660                         /* Unmapped buffer. Need to map it. */
1661                         bh->b_bdev = vol->sb->s_bdev;
1662
1663                         /* Convert block into corresponding vcn and offset. */
1664                         vcn = (VCN)block << blocksize_bits >>
1665                                         vol->cluster_size_bits;
1666                         vcn_ofs = ((VCN)block << blocksize_bits) &
1667                                         vol->cluster_size_mask;
1668
1669                         is_retry = FALSE;
1670                         if (!rl) {
1671 lock_retry_remap:
1672                                 down_read(&ni->runlist.lock);
1673                                 rl = ni->runlist.rl;
1674                         }
1675                         if (likely(rl != NULL)) {
1676                                 /* Seek to element containing target vcn. */
1677                                 while (rl->length && rl[1].vcn <= vcn)
1678                                         rl++;
1679                                 lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
1680                         } else
1681                                 lcn = LCN_RL_NOT_MAPPED;
1682                         if (unlikely(lcn < 0)) {
1683                                 /*
1684                                  * We extended the attribute allocation above.
1685                                  * If we hit an ENOENT here it means that the
1686                                  * allocation was insufficient which is a bug.
1687                                  */
1688                                 BUG_ON(lcn == LCN_ENOENT);
1689
1690                                 /* It is a hole, need to instantiate it. */
1691                                 if (lcn == LCN_HOLE) {
1692                                         // TODO: Instantiate the hole.
1693                                         // clear_buffer_new(bh);
1694                                         // unmap_underlying_metadata(bh->b_bdev,
1695                                         //              bh->b_blocknr);
1696                                         // For non-uptodate buffers, need to
1697                                         // zero out the region outside the
1698                                         // request in this bh or all bhs,
1699                                         // depending on what we implemented
1700                                         // above.
1701                                         // Need to flush_dcache_page().
1702                                         // Or could use set_buffer_new()
1703                                         // instead?
1704                                         ntfs_error(vol->sb, "Writing into "
1705                                                         "sparse regions is "
1706                                                         "not supported yet. "
1707                                                         "Sorry.");
1708                                         err = -EOPNOTSUPP;
1709                                         if (!rl)
1710                                                 up_read(&ni->runlist.lock);
1711                                         goto err_out;
1712                                 } else if (!is_retry &&
1713                                                 lcn == LCN_RL_NOT_MAPPED) {
1714                                         is_retry = TRUE;
1715                                         /*
1716                                          * Attempt to map runlist, dropping
1717                                          * lock for the duration.
1718                                          */
1719                                         up_read(&ni->runlist.lock);
1720                                         err = ntfs_map_runlist(ni, vcn);
1721                                         if (likely(!err))
1722                                                 goto lock_retry_remap;
1723                                         rl = NULL;
1724                                         lcn = err;
1725                                 } else if (!rl)
1726                                         up_read(&ni->runlist.lock);
1727                                 /*
1728                                  * Failed to map the buffer, even after
1729                                  * retrying.
1730                                  */
1731                                 bh->b_blocknr = -1;
1732                                 ntfs_error(vol->sb, "Failed to write to inode "
1733                                                 "0x%lx, attribute type 0x%x, "
1734                                                 "vcn 0x%llx, offset 0x%x "
1735                                                 "because its location on disk "
1736                                                 "could not be determined%s "
1737                                                 "(error code %lli).",
1738                                                 ni->mft_no, ni->type,
1739                                                 (unsigned long long)vcn,
1740                                                 vcn_ofs, is_retry ? " even "
1741                                                 "after retrying" : "",
1742                                                 (long long)lcn);
1743                                 if (!err)
1744                                         err = -EIO;
1745                                 goto err_out;
1746                         }
1747                         /* We now have a successful remap, i.e. lcn >= 0. */
1748
1749                         /* Setup buffer head to correct block. */
1750                         bh->b_blocknr = ((lcn << vol->cluster_size_bits)
1751                                         + vcn_ofs) >> blocksize_bits;
1752                         set_buffer_mapped(bh);
1753
1754                         // FIXME: Something analogous to this is needed for
1755                         // each newly allocated block, i.e. BH_New.
1756                         // FIXME: Might need to take this out of the
1757                         // if (!buffer_mapped(bh)) {}, depending on how we
1758                         // implement things during the allocated_size and
1759                         // initialized_size extension code above.
1760                         if (buffer_new(bh)) {
1761                                 clear_buffer_new(bh);
1762                                 unmap_underlying_metadata(bh->b_bdev,
1763                                                 bh->b_blocknr);
1764                                 if (PageUptodate(page)) {
1765                                         set_buffer_uptodate(bh);
1766                                         continue;
1767                                 }
1768                                 /*
1769                                  * Page is _not_ uptodate, zero surrounding
1770                                  * region. NOTE: This is how we decide if to
1771                                  * zero or not!
1772                                  */
1773                                 if (block_end > to || block_start < from) {
1774                                         void *kaddr;
1775
1776                                         kaddr = kmap_atomic(page, KM_USER0);
1777                                         if (block_end > to)
1778                                                 memset(kaddr + to, 0,
1779                                                                 block_end - to);
1780                                         if (block_start < from)
1781                                                 memset(kaddr + block_start, 0,
1782                                                                 from -
1783                                                                 block_start);
1784                                         flush_dcache_page(page);
1785                                         kunmap_atomic(kaddr, KM_USER0);
1786                                 }
1787                                 continue;
1788                         }
1789                 }
1790                 /* @bh is mapped, set it uptodate if the page is uptodate. */
1791                 if (PageUptodate(page)) {
1792                         if (!buffer_uptodate(bh))
1793                                 set_buffer_uptodate(bh);
1794                         continue;
1795                 }
1796                 /*
1797                  * The page is not uptodate. The buffer is mapped. If it is not
1798                  * uptodate, and it is only partially being written to, we need
1799                  * to read the buffer in before the write, i.e. right now.
1800                  */
1801                 if (!buffer_uptodate(bh) &&
1802                                 (block_start < from || block_end > to)) {
1803                         ll_rw_block(READ, 1, &bh);
1804                         *wait_bh++ = bh;
1805                 }
1806         } while (block++, block_start = block_end,
1807                         (bh = bh->b_this_page) != head);
1808
1809         /* Release the lock if we took it. */
1810         if (rl) {
1811                 up_read(&ni->runlist.lock);
1812                 rl = NULL;
1813         }
1814
1815         /* If we issued read requests, let them complete. */
1816         while (wait_bh > wait) {
1817                 wait_on_buffer(*--wait_bh);
1818                 if (!buffer_uptodate(*wait_bh))
1819                         return -EIO;
1820         }
1821
1822         ntfs_debug("Done.");
1823         return 0;
1824 err_out:
1825         /*
1826          * Zero out any newly allocated blocks to avoid exposing stale data.
1827          * If BH_New is set, we know that the block was newly allocated in the
1828          * above loop.
1829          * FIXME: What about initialized_size increments? Have we done all the
1830          * required zeroing above? If not this error handling is broken, and
1831          * in particular the if (block_end <= from) check is completely bogus.
1832          */
1833         bh = head;
1834         block_start = 0;
1835         is_retry = FALSE;
1836         do {
1837                 block_end = block_start + blocksize;
1838                 if (block_end <= from)
1839                         continue;
1840                 if (block_start >= to)
1841                         break;
1842                 if (buffer_new(bh)) {
1843                         void *kaddr;
1844
1845                         clear_buffer_new(bh);
1846                         kaddr = kmap_atomic(page, KM_USER0);
1847                         memset(kaddr + block_start, 0, bh->b_size);
1848                         kunmap_atomic(kaddr, KM_USER0);
1849                         set_buffer_uptodate(bh);
1850                         mark_buffer_dirty(bh);
1851                         is_retry = TRUE;
1852                 }
1853         } while (block_start = block_end, (bh = bh->b_this_page) != head);
1854         if (is_retry)
1855                 flush_dcache_page(page);
1856         if (rl)
1857                 up_read(&ni->runlist.lock);
1858         return err;
1859 }
1860
1861 /**
1862  * ntfs_prepare_write - prepare a page for receiving data
1863  *
1864  * This is called from generic_file_write() with i_sem held on the inode
1865  * (@page->mapping->host).  The @page is locked but not kmap()ped.  The source
1866  * data has not yet been copied into the @page.
1867  *
1868  * Need to extend the attribute/fill in holes if necessary, create blocks and
1869  * make partially overwritten blocks uptodate,
1870  *
1871  * i_size is not to be modified yet.
1872  *
1873  * Return 0 on success or -errno on error.
1874  *
1875  * Should be using block_prepare_write() [support for sparse files] or
1876  * cont_prepare_write() [no support for sparse files].  Cannot do that due to
1877  * ntfs specifics but can look at them for implementation guidance.
1878  *
1879  * Note: In the range, @from is inclusive and @to is exclusive, i.e. @from is
1880  * the first byte in the page that will be written to and @to is the first byte
1881  * after the last byte that will be written to.
1882  */
1883 static int ntfs_prepare_write(struct file *file, struct page *page,
1884                 unsigned from, unsigned to)
1885 {
1886         s64 new_size;
1887         loff_t i_size;
1888         struct inode *vi = page->mapping->host;
1889         ntfs_inode *base_ni = NULL, *ni = NTFS_I(vi);
1890         ntfs_volume *vol = ni->vol;
1891         ntfs_attr_search_ctx *ctx = NULL;
1892         MFT_RECORD *m = NULL;
1893         ATTR_RECORD *a;
1894         u8 *kaddr;
1895         u32 attr_len;
1896         int err;
1897
1898         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
1899                         "0x%lx, from = %u, to = %u.", vi->i_ino, ni->type,
1900                         page->index, from, to);
1901         BUG_ON(!PageLocked(page));
1902         BUG_ON(from > PAGE_CACHE_SIZE);
1903         BUG_ON(to > PAGE_CACHE_SIZE);
1904         BUG_ON(from > to);
1905         BUG_ON(NInoMstProtected(ni));
1906         /*
1907          * If a previous ntfs_truncate() failed, repeat it and abort if it
1908          * fails again.
1909          */
1910         if (unlikely(NInoTruncateFailed(ni))) {
1911                 down_write(&vi->i_alloc_sem);
1912                 err = ntfs_truncate(vi);
1913                 up_write(&vi->i_alloc_sem);
1914                 if (err || NInoTruncateFailed(ni)) {
1915                         if (!err)
1916                                 err = -EIO;
1917                         goto err_out;
1918                 }
1919         }
1920         /* If the attribute is not resident, deal with it elsewhere. */
1921         if (NInoNonResident(ni)) {
1922                 /*
1923                  * Only unnamed $DATA attributes can be compressed, encrypted,
1924                  * and/or sparse.
1925                  */
1926                 if (ni->type == AT_DATA && !ni->name_len) {
1927                         /* If file is encrypted, deny access, just like NT4. */
1928                         if (NInoEncrypted(ni)) {
1929                                 ntfs_debug("Denying write access to encrypted "
1930                                                 "file.");
1931                                 return -EACCES;
1932                         }
1933                         /* Compressed data streams are handled in compress.c. */
1934                         if (NInoCompressed(ni)) {
1935                                 // TODO: Implement and replace this check with
1936                                 // return ntfs_write_compressed_block(page);
1937                                 ntfs_error(vi->i_sb, "Writing to compressed "
1938                                                 "files is not supported yet. "
1939                                                 "Sorry.");
1940                                 return -EOPNOTSUPP;
1941                         }
1942                         // TODO: Implement and remove this check.
1943                         if (NInoSparse(ni)) {
1944                                 ntfs_error(vi->i_sb, "Writing to sparse files "
1945                                                 "is not supported yet. Sorry.");
1946                                 return -EOPNOTSUPP;
1947                         }
1948                 }
1949                 /* Normal data stream. */
1950                 return ntfs_prepare_nonresident_write(page, from, to);
1951         }
1952         /*
1953          * Attribute is resident, implying it is not compressed, encrypted, or
1954          * sparse.
1955          */
1956         BUG_ON(page_has_buffers(page));
1957         new_size = ((s64)page->index << PAGE_CACHE_SHIFT) + to;
1958         /* If we do not need to resize the attribute allocation we are done. */
1959         if (new_size <= i_size_read(vi))
1960                 goto done;
1961         /* Map, pin, and lock the (base) mft record. */
1962         if (!NInoAttr(ni))
1963                 base_ni = ni;
1964         else
1965                 base_ni = ni->ext.base_ntfs_ino;
1966         m = map_mft_record(base_ni);
1967         if (IS_ERR(m)) {
1968                 err = PTR_ERR(m);
1969                 m = NULL;
1970                 ctx = NULL;
1971                 goto err_out;
1972         }
1973         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1974         if (unlikely(!ctx)) {
1975                 err = -ENOMEM;
1976                 goto err_out;
1977         }
1978         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1979                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1980         if (unlikely(err)) {
1981                 if (err == -ENOENT)
1982                         err = -EIO;
1983                 goto err_out;
1984         }
1985         m = ctx->mrec;
1986         a = ctx->attr;
1987         /* The total length of the attribute value. */
1988         attr_len = le32_to_cpu(a->data.resident.value_length);
1989         /* Fix an eventual previous failure of ntfs_commit_write(). */
1990         i_size = i_size_read(vi);
1991         if (unlikely(attr_len > i_size)) {
1992                 attr_len = i_size;
1993                 a->data.resident.value_length = cpu_to_le32(attr_len);
1994         }
1995         /* If we do not need to resize the attribute allocation we are done. */
1996         if (new_size <= attr_len)
1997                 goto done_unm;
1998         /* Check if new size is allowed in $AttrDef. */
1999         err = ntfs_attr_size_bounds_check(vol, ni->type, new_size);
2000         if (unlikely(err)) {
2001                 if (err == -ERANGE) {
2002                         ntfs_error(vol->sb, "Write would cause the inode "
2003                                         "0x%lx to exceed the maximum size for "
2004                                         "its attribute type (0x%x).  Aborting "
2005                                         "write.", vi->i_ino,
2006                                         le32_to_cpu(ni->type));
2007                 } else {
2008                         ntfs_error(vol->sb, "Inode 0x%lx has unknown "
2009                                         "attribute type 0x%x.  Aborting "
2010                                         "write.", vi->i_ino,
2011                                         le32_to_cpu(ni->type));
2012                         err = -EIO;
2013                 }
2014                 goto err_out2;
2015         }
2016         /*
2017          * Extend the attribute record to be able to store the new attribute
2018          * size.
2019          */
2020         if (new_size >= vol->mft_record_size || ntfs_attr_record_resize(m, a,
2021                         le16_to_cpu(a->data.resident.value_offset) +
2022                         new_size)) {
2023                 /* Not enough space in the mft record. */
2024                 ntfs_error(vol->sb, "Not enough space in the mft record for "
2025                                 "the resized attribute value.  This is not "
2026                                 "supported yet.  Aborting write.");
2027                 err = -EOPNOTSUPP;
2028                 goto err_out2;
2029         }
2030         /*
2031          * We have enough space in the mft record to fit the write.  This
2032          * implies the attribute is smaller than the mft record and hence the
2033          * attribute must be in a single page and hence page->index must be 0.
2034          */
2035         BUG_ON(page->index);
2036         /*
2037          * If the beginning of the write is past the old size, enlarge the
2038          * attribute value up to the beginning of the write and fill it with
2039          * zeroes.
2040          */
2041         if (from > attr_len) {
2042                 memset((u8*)a + le16_to_cpu(a->data.resident.value_offset) +
2043                                 attr_len, 0, from - attr_len);
2044                 a->data.resident.value_length = cpu_to_le32(from);
2045                 /* Zero the corresponding area in the page as well. */
2046                 if (PageUptodate(page)) {
2047                         kaddr = kmap_atomic(page, KM_USER0);
2048                         memset(kaddr + attr_len, 0, from - attr_len);
2049                         kunmap_atomic(kaddr, KM_USER0);
2050                         flush_dcache_page(page);
2051                 }
2052         }
2053         flush_dcache_mft_record_page(ctx->ntfs_ino);
2054         mark_mft_record_dirty(ctx->ntfs_ino);
2055 done_unm:
2056         ntfs_attr_put_search_ctx(ctx);
2057         unmap_mft_record(base_ni);
2058         /*
2059          * Because resident attributes are handled by memcpy() to/from the
2060          * corresponding MFT record, and because this form of i/o is byte
2061          * aligned rather than block aligned, there is no need to bring the
2062          * page uptodate here as in the non-resident case where we need to
2063          * bring the buffers straddled by the write uptodate before
2064          * generic_file_write() does the copying from userspace.
2065          *
2066          * We thus defer the uptodate bringing of the page region outside the
2067          * region written to to ntfs_commit_write(), which makes the code
2068          * simpler and saves one atomic kmap which is good.
2069          */
2070 done:
2071         ntfs_debug("Done.");
2072         return 0;
2073 err_out:
2074         if (err == -ENOMEM)
2075                 ntfs_warning(vi->i_sb, "Error allocating memory required to "
2076                                 "prepare the write.");
2077         else {
2078                 ntfs_error(vi->i_sb, "Resident attribute prepare write failed "
2079                                 "with error %i.", err);
2080                 NVolSetErrors(vol);
2081                 make_bad_inode(vi);
2082         }
2083 err_out2:
2084         if (ctx)
2085                 ntfs_attr_put_search_ctx(ctx);
2086         if (m)
2087                 unmap_mft_record(base_ni);
2088         return err;
2089 }
2090
2091 /**
2092  * ntfs_commit_nonresident_write -
2093  *
2094  */
2095 static int ntfs_commit_nonresident_write(struct page *page,
2096                 unsigned from, unsigned to)
2097 {
2098         s64 pos = ((s64)page->index << PAGE_CACHE_SHIFT) + to;
2099         struct inode *vi = page->mapping->host;
2100         struct buffer_head *bh, *head;
2101         unsigned int block_start, block_end, blocksize;
2102         BOOL partial;
2103
2104         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
2105                         "0x%lx, from = %u, to = %u.", vi->i_ino,
2106                         NTFS_I(vi)->type, page->index, from, to);
2107         blocksize = 1 << vi->i_blkbits;
2108
2109         // FIXME: We need a whole slew of special cases in here for compressed
2110         // files for example...
2111         // For now, we know ntfs_prepare_write() would have failed so we can't
2112         // get here in any of the cases which we have to special case, so we
2113         // are just a ripped off, unrolled generic_commit_write().
2114
2115         bh = head = page_buffers(page);
2116         block_start = 0;
2117         partial = FALSE;
2118         do {
2119                 block_end = block_start + blocksize;
2120                 if (block_end <= from || block_start >= to) {
2121                         if (!buffer_uptodate(bh))
2122                                 partial = TRUE;
2123                 } else {
2124                         set_buffer_uptodate(bh);
2125                         mark_buffer_dirty(bh);
2126                 }
2127         } while (block_start = block_end, (bh = bh->b_this_page) != head);
2128         /*
2129          * If this is a partial write which happened to make all buffers
2130          * uptodate then we can optimize away a bogus ->readpage() for the next
2131          * read().  Here we 'discover' whether the page went uptodate as a
2132          * result of this (potentially partial) write.
2133          */
2134         if (!partial)
2135                 SetPageUptodate(page);
2136         /*
2137          * Not convinced about this at all.  See disparity comment above.  For
2138          * now we know ntfs_prepare_write() would have failed in the write
2139          * exceeds i_size case, so this will never trigger which is fine.
2140          */
2141         if (pos > i_size_read(vi)) {
2142                 ntfs_error(vi->i_sb, "Writing beyond the existing file size is "
2143                                 "not supported yet.  Sorry.");
2144                 return -EOPNOTSUPP;
2145                 // vi->i_size = pos;
2146                 // mark_inode_dirty(vi);
2147         }
2148         ntfs_debug("Done.");
2149         return 0;
2150 }
2151
2152 /**
2153  * ntfs_commit_write - commit the received data
2154  *
2155  * This is called from generic_file_write() with i_sem held on the inode
2156  * (@page->mapping->host).  The @page is locked but not kmap()ped.  The source
2157  * data has already been copied into the @page.  ntfs_prepare_write() has been
2158  * called before the data copied and it returned success so we can take the
2159  * results of various BUG checks and some error handling for granted.
2160  *
2161  * Need to mark modified blocks dirty so they get written out later when
2162  * ntfs_writepage() is invoked by the VM.
2163  *
2164  * Return 0 on success or -errno on error.
2165  *
2166  * Should be using generic_commit_write().  This marks buffers uptodate and
2167  * dirty, sets the page uptodate if all buffers in the page are uptodate, and
2168  * updates i_size if the end of io is beyond i_size.  In that case, it also
2169  * marks the inode dirty.
2170  *
2171  * Cannot use generic_commit_write() due to ntfs specialities but can look at
2172  * it for implementation guidance.
2173  *
2174  * If things have gone as outlined in ntfs_prepare_write(), then we do not
2175  * need to do any page content modifications here at all, except in the write
2176  * to resident attribute case, where we need to do the uptodate bringing here
2177  * which we combine with the copying into the mft record which means we save
2178  * one atomic kmap.
2179  */
2180 static int ntfs_commit_write(struct file *file, struct page *page,
2181                 unsigned from, unsigned to)
2182 {
2183         struct inode *vi = page->mapping->host;
2184         ntfs_inode *base_ni, *ni = NTFS_I(vi);
2185         char *kaddr, *kattr;
2186         ntfs_attr_search_ctx *ctx;
2187         MFT_RECORD *m;
2188         ATTR_RECORD *a;
2189         u32 attr_len;
2190         int err;
2191
2192         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
2193                         "0x%lx, from = %u, to = %u.", vi->i_ino, ni->type,
2194                         page->index, from, to);
2195         /* If the attribute is not resident, deal with it elsewhere. */
2196         if (NInoNonResident(ni)) {
2197                 /* Only unnamed $DATA attributes can be compressed/encrypted. */
2198                 if (ni->type == AT_DATA && !ni->name_len) {
2199                         /* Encrypted files need separate handling. */
2200                         if (NInoEncrypted(ni)) {
2201                                 // We never get here at present!
2202                                 BUG();
2203                         }
2204                         /* Compressed data streams are handled in compress.c. */
2205                         if (NInoCompressed(ni)) {
2206                                 // TODO: Implement this!
2207                                 // return ntfs_write_compressed_block(page);
2208                                 // We never get here at present!
2209                                 BUG();
2210                         }
2211                 }
2212                 /* Normal data stream. */
2213                 return ntfs_commit_nonresident_write(page, from, to);
2214         }
2215         /*
2216          * Attribute is resident, implying it is not compressed, encrypted, or
2217          * sparse.
2218          */
2219         if (!NInoAttr(ni))
2220                 base_ni = ni;
2221         else
2222                 base_ni = ni->ext.base_ntfs_ino;
2223         /* Map, pin, and lock the mft record. */
2224         m = map_mft_record(base_ni);
2225         if (IS_ERR(m)) {
2226                 err = PTR_ERR(m);
2227                 m = NULL;
2228                 ctx = NULL;
2229                 goto err_out;
2230         }
2231         ctx = ntfs_attr_get_search_ctx(base_ni, m);
2232         if (unlikely(!ctx)) {
2233                 err = -ENOMEM;
2234                 goto err_out;
2235         }
2236         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
2237                         CASE_SENSITIVE, 0, NULL, 0, ctx);
2238         if (unlikely(err)) {
2239                 if (err == -ENOENT)
2240                         err = -EIO;
2241                 goto err_out;
2242         }
2243         a = ctx->attr;
2244         /* The total length of the attribute value. */
2245         attr_len = le32_to_cpu(a->data.resident.value_length);
2246         BUG_ON(from > attr_len);
2247         kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
2248         kaddr = kmap_atomic(page, KM_USER0);
2249         /* Copy the received data from the page to the mft record. */
2250         memcpy(kattr + from, kaddr + from, to - from);
2251         /* Update the attribute length if necessary. */
2252         if (to > attr_len) {
2253                 attr_len = to;
2254                 a->data.resident.value_length = cpu_to_le32(attr_len);
2255         }
2256         /*
2257          * If the page is not uptodate, bring the out of bounds area(s)
2258          * uptodate by copying data from the mft record to the page.
2259          */
2260         if (!PageUptodate(page)) {
2261                 if (from > 0)
2262                         memcpy(kaddr, kattr, from);
2263                 if (to < attr_len)
2264                         memcpy(kaddr + to, kattr + to, attr_len - to);
2265                 /* Zero the region outside the end of the attribute value. */
2266                 if (attr_len < PAGE_CACHE_SIZE)
2267                         memset(kaddr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
2268                 /*
2269                  * The probability of not having done any of the above is
2270                  * extremely small, so we just flush unconditionally.
2271                  */
2272                 flush_dcache_page(page);
2273                 SetPageUptodate(page);
2274         }
2275         kunmap_atomic(kaddr, KM_USER0);
2276         /* Update i_size if necessary. */
2277         if (i_size_read(vi) < attr_len) {
2278                 unsigned long flags;
2279
2280                 write_lock_irqsave(&ni->size_lock, flags);
2281                 ni->allocated_size = ni->initialized_size = attr_len;
2282                 i_size_write(vi, attr_len);
2283                 write_unlock_irqrestore(&ni->size_lock, flags);
2284         }
2285         /* Mark the mft record dirty, so it gets written back. */
2286         flush_dcache_mft_record_page(ctx->ntfs_ino);
2287         mark_mft_record_dirty(ctx->ntfs_ino);
2288         ntfs_attr_put_search_ctx(ctx);
2289         unmap_mft_record(base_ni);
2290         ntfs_debug("Done.");
2291         return 0;
2292 err_out:
2293         if (err == -ENOMEM) {
2294                 ntfs_warning(vi->i_sb, "Error allocating memory required to "
2295                                 "commit the write.");
2296                 if (PageUptodate(page)) {
2297                         ntfs_warning(vi->i_sb, "Page is uptodate, setting "
2298                                         "dirty so the write will be retried "
2299                                         "later on by the VM.");
2300                         /*
2301                          * Put the page on mapping->dirty_pages, but leave its
2302                          * buffers' dirty state as-is.
2303                          */
2304                         __set_page_dirty_nobuffers(page);
2305                         err = 0;
2306                 } else
2307                         ntfs_error(vi->i_sb, "Page is not uptodate.  Written "
2308                                         "data has been lost.");
2309         } else {
2310                 ntfs_error(vi->i_sb, "Resident attribute commit write failed "
2311                                 "with error %i.", err);
2312                 NVolSetErrors(ni->vol);
2313                 make_bad_inode(vi);
2314         }
2315         if (ctx)
2316                 ntfs_attr_put_search_ctx(ctx);
2317         if (m)
2318                 unmap_mft_record(base_ni);
2319         return err;
2320 }
2321
2322 #endif  /* NTFS_RW */
2323
2324 /**
2325  * ntfs_aops - general address space operations for inodes and attributes
2326  */
2327 struct address_space_operations ntfs_aops = {
2328         .readpage       = ntfs_readpage,        /* Fill page with data. */
2329         .sync_page      = block_sync_page,      /* Currently, just unplugs the
2330                                                    disk request queue. */
2331 #ifdef NTFS_RW
2332         .writepage      = ntfs_writepage,       /* Write dirty page to disk. */
2333         .prepare_write  = ntfs_prepare_write,   /* Prepare page and buffers
2334                                                    ready to receive data. */
2335         .commit_write   = ntfs_commit_write,    /* Commit received data. */
2336 #endif /* NTFS_RW */
2337 };
2338
2339 /**
2340  * ntfs_mst_aops - general address space operations for mst protecteed inodes
2341  *                 and attributes
2342  */
2343 struct address_space_operations ntfs_mst_aops = {
2344         .readpage       = ntfs_readpage,        /* Fill page with data. */
2345         .sync_page      = block_sync_page,      /* Currently, just unplugs the
2346                                                    disk request queue. */
2347 #ifdef NTFS_RW
2348         .writepage      = ntfs_writepage,       /* Write dirty page to disk. */
2349         .set_page_dirty = __set_page_dirty_nobuffers,   /* Set the page dirty
2350                                                    without touching the buffers
2351                                                    belonging to the page. */
2352 #endif /* NTFS_RW */
2353 };
2354
2355 #ifdef NTFS_RW
2356
2357 /**
2358  * mark_ntfs_record_dirty - mark an ntfs record dirty
2359  * @page:       page containing the ntfs record to mark dirty
2360  * @ofs:        byte offset within @page at which the ntfs record begins
2361  *
2362  * Set the buffers and the page in which the ntfs record is located dirty.
2363  *
2364  * The latter also marks the vfs inode the ntfs record belongs to dirty
2365  * (I_DIRTY_PAGES only).
2366  *
2367  * If the page does not have buffers, we create them and set them uptodate.
2368  * The page may not be locked which is why we need to handle the buffers under
2369  * the mapping->private_lock.  Once the buffers are marked dirty we no longer
2370  * need the lock since try_to_free_buffers() does not free dirty buffers.
2371  */
2372 void mark_ntfs_record_dirty(struct page *page, const unsigned int ofs) {
2373         struct address_space *mapping = page->mapping;
2374         ntfs_inode *ni = NTFS_I(mapping->host);
2375         struct buffer_head *bh, *head, *buffers_to_free = NULL;
2376         unsigned int end, bh_size, bh_ofs;
2377
2378         BUG_ON(!PageUptodate(page));
2379         end = ofs + ni->itype.index.block_size;
2380         bh_size = 1 << VFS_I(ni)->i_blkbits;
2381         spin_lock(&mapping->private_lock);
2382         if (unlikely(!page_has_buffers(page))) {
2383                 spin_unlock(&mapping->private_lock);
2384                 bh = head = alloc_page_buffers(page, bh_size, 1);
2385                 spin_lock(&mapping->private_lock);
2386                 if (likely(!page_has_buffers(page))) {
2387                         struct buffer_head *tail;
2388
2389                         do {
2390                                 set_buffer_uptodate(bh);
2391                                 tail = bh;
2392                                 bh = bh->b_this_page;
2393                         } while (bh);
2394                         tail->b_this_page = head;
2395                         attach_page_buffers(page, head);
2396                 } else
2397                         buffers_to_free = bh;
2398         }
2399         bh = head = page_buffers(page);
2400         do {
2401                 bh_ofs = bh_offset(bh);
2402                 if (bh_ofs + bh_size <= ofs)
2403                         continue;
2404                 if (unlikely(bh_ofs >= end))
2405                         break;
2406                 set_buffer_dirty(bh);
2407         } while ((bh = bh->b_this_page) != head);
2408         spin_unlock(&mapping->private_lock);
2409         __set_page_dirty_nobuffers(page);
2410         if (unlikely(buffers_to_free)) {
2411                 do {
2412                         bh = buffers_to_free->b_this_page;
2413                         free_buffer_head(buffers_to_free);
2414                         buffers_to_free = bh;
2415                 } while (buffers_to_free);
2416         }
2417 }
2418
2419 #endif /* NTFS_RW */