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) 2001-2004 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
  30 #include "ntfs.h"
  31
  32 /**
  33  * ntfs_end_buffer_async_read - async io completion for reading attributes
  34  * @bh:         buffer head on which io is completed
  35  * @uptodate:   whether @bh is now uptodate or not
  36  *
  37  * Asynchronous I/O completion handler for reading pages belonging to the
  38  * attribute address space of an inode. The inodes can either be files or
  39  * directories or they can be fake inodes describing some attribute.
  40  *
  41  * If NInoMstProtected(), perform the post read mst fixups when all IO on the
  42  * page has been completed and mark the page uptodate or set the error bit on
  43  * the page. To determine the size of the records that need fixing up, we cheat
  44  * a little bit by setting the index_block_size in ntfs_inode to the ntfs
  45  * record size, and index_block_size_bits, to the log(base 2) of the ntfs
  46  * record size.
  47  */
  48 static void ntfs_end_buffer_async_read(struct buffer_head *bh, int uptodate)
  49 {
  50         static spinlock_t page_uptodate_lock = SPIN_LOCK_UNLOCKED;
  51         unsigned long flags;
  52         struct buffer_head *tmp;
  53         struct page *page;
  54         ntfs_inode *ni;
  55         int page_uptodate = 1;
  56
  57         page = bh->b_page;
  58         ni = NTFS_I(page->mapping->host);
  59
  60         if (likely(uptodate)) {
  61                 s64 file_ofs;
  62
  63                 set_buffer_uptodate(bh);
  64
  65                 file_ofs = (page->index << PAGE_CACHE_SHIFT) + bh_offset(bh);
  66                 /* Check for the current buffer head overflowing. */
  67                 if (file_ofs + bh->b_size > ni->initialized_size) {
  68                         char *addr;
  69                         int ofs = 0;
  70
  71                         if (file_ofs < ni->initialized_size)
  72                                 ofs = ni->initialized_size - file_ofs;
  73                         addr = kmap_atomic(page, KM_BIO_SRC_IRQ);
  74                         memset(addr + bh_offset(bh) + ofs, 0, bh->b_size - ofs);
  75                         flush_dcache_page(page);
  76                         kunmap_atomic(addr, KM_BIO_SRC_IRQ);
  77                 }
  78         } else {
  79                 clear_buffer_uptodate(bh);
  80                 ntfs_error(ni->vol->sb, "Buffer I/O error, logical block %llu.",
  81                                 (unsigned long long)bh->b_blocknr);
  82                 SetPageError(page);
  83         }
  84
  85         spin_lock_irqsave(&page_uptodate_lock, flags);
  86         clear_buffer_async_read(bh);
  87         unlock_buffer(bh);
  88         tmp = bh;
  89         do {
  90                 if (!buffer_uptodate(tmp))
  91                         page_uptodate = 0;
  92                 if (buffer_async_read(tmp)) {
  93                         if (likely(buffer_locked(tmp)))
  94                                 goto still_busy;
  95                         /* Async buffers must be locked. */
  96                         BUG();
  97                 }
  98                 tmp = tmp->b_this_page;
  99         } while (tmp != bh);
 100         spin_unlock_irqrestore(&page_uptodate_lock, flags);
 101         /*
 102          * If none of the buffers had errors then we can set the page uptodate,
 103          * but we first have to perform the post read mst fixups, if the
 104          * attribute is mst protected, i.e. if NInoMstProteced(ni) is true.
 105          */
 106         if (!NInoMstProtected(ni)) {
 107                 if (likely(page_uptodate && !PageError(page)))
 108                         SetPageUptodate(page);
 109         } else {
 110                 char *addr;
 111                 unsigned int i, recs, nr_err;
 112                 u32 rec_size;
 113
 114                 rec_size = ni->itype.index.block_size;
 115                 recs = PAGE_CACHE_SIZE / rec_size;
 116                 addr = kmap_atomic(page, KM_BIO_SRC_IRQ);
 117                 for (i = nr_err = 0; i < recs; i++) {
 118                         if (likely(!post_read_mst_fixup((NTFS_RECORD*)(addr +
 119                                         i * rec_size), rec_size)))
 120                                 continue;
 121                         nr_err++;
 122                         ntfs_error(ni->vol->sb, "post_read_mst_fixup() failed, "
 123                                         "corrupt %s record 0x%llx. Run chkdsk.",
 124                                         ni->mft_no ? "index" : "mft",
 125                                         (unsigned long long)(((s64)page->index
 126                                         << PAGE_CACHE_SHIFT >>
 127                                         ni->itype.index.block_size_bits) + i));
 128                 }
 129                 flush_dcache_page(page);
 130                 kunmap_atomic(addr, KM_BIO_SRC_IRQ);
 131                 if (likely(!PageError(page))) {
 132                         if (likely(!nr_err && recs)) {
 133                                 if (likely(page_uptodate))
 134                                         SetPageUptodate(page);
 135                         } else {
 136                                 ntfs_error(ni->vol->sb, "Setting page error, "
 137                                                 "index 0x%lx.", page->index);
 138                                 SetPageError(page);
 139                         }
 140                 }
 141         }
 142         unlock_page(page);
 143         return;
 144 still_busy:
 145         spin_unlock_irqrestore(&page_uptodate_lock, flags);
 146         return;
 147 }
 148
 149 /**
 150  * ntfs_read_block - fill a @page of an address space with data
 151  * @page:       page cache page to fill with data
 152  *
 153  * Fill the page @page of the address space belonging to the @page->host inode.
 154  * We read each buffer asynchronously and when all buffers are read in, our io
 155  * completion handler ntfs_end_buffer_read_async(), if required, automatically
 156  * applies the mst fixups to the page before finally marking it uptodate and
 157  * unlocking it.
 158  *
 159  * We only enforce allocated_size limit because i_size is checked for in
 160  * generic_file_read().
 161  *
 162  * Return 0 on success and -errno on error.
 163  *
 164  * Contains an adapted version of fs/buffer.c::block_read_full_page().
 165  */
 166 static int ntfs_read_block(struct page *page)
 167 {
 168         VCN vcn;
 169         LCN lcn;
 170         ntfs_inode *ni;
 171         ntfs_volume *vol;
 172         run_list_element *rl;
 173         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
 174         sector_t iblock, lblock, zblock;
 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         blocksize_bits = VFS_I(ni)->i_blkbits;
 183         blocksize = 1 << blocksize_bits;
 184
 185         if (!page_has_buffers(page))
 186                 create_empty_buffers(page, blocksize, 0);
 187         bh = head = page_buffers(page);
 188         if (unlikely(!bh)) {
 189                 unlock_page(page);
 190                 return -ENOMEM;
 191         }
 192
 193         iblock = page->index << (PAGE_CACHE_SHIFT - blocksize_bits);
 194         lblock = (ni->allocated_size + blocksize - 1) >> blocksize_bits;
 195         zblock = (ni->initialized_size + blocksize - 1) >> blocksize_bits;
 196
 197 #ifdef DEBUG
 198         if (unlikely(!ni->run_list.rl && !ni->mft_no && !NInoAttr(ni)))
 199                 panic("NTFS: $MFT/$DATA run list has been unmapped! This is a "
 200                                 "very serious bug! Cannot continue...");
 201 #endif
 202
 203         /* Loop through all the buffers in the page. */
 204         rl = NULL;
 205         nr = i = 0;
 206         do {
 207                 if (unlikely(buffer_uptodate(bh)))
 208                         continue;
 209                 if (unlikely(buffer_mapped(bh))) {
 210                         arr[nr++] = bh;
 211                         continue;
 212                 }
 213                 bh->b_bdev = vol->sb->s_bdev;
 214                 /* Is the block within the allowed limits? */
 215                 if (iblock < lblock) {
 216                         BOOL is_retry = FALSE;
 217
 218                         /* Convert iblock into corresponding vcn and offset. */
 219                         vcn = (VCN)iblock << blocksize_bits >>
 220                                         vol->cluster_size_bits;
 221                         vcn_ofs = ((VCN)iblock << blocksize_bits) &
 222                                         vol->cluster_size_mask;
 223                         if (!rl) {
 224 lock_retry_remap:
 225                                 down_read(&ni->run_list.lock);
 226                                 rl = ni->run_list.rl;
 227                         }
 228                         if (likely(rl != NULL)) {
 229                                 /* Seek to element containing target vcn. */
 230                                 while (rl->length && rl[1].vcn <= vcn)
 231                                         rl++;
 232                                 lcn = vcn_to_lcn(rl, vcn);
 233                         } else
 234                                 lcn = (LCN)LCN_RL_NOT_MAPPED;
 235                         /* Successful remap. */
 236                         if (lcn >= 0) {
 237                                 /* Setup buffer head to correct block. */
 238                                 bh->b_blocknr = ((lcn << vol->cluster_size_bits)
 239                                                 + vcn_ofs) >> blocksize_bits;
 240                                 set_buffer_mapped(bh);
 241                                 /* Only read initialized data blocks. */
 242                                 if (iblock < zblock) {
 243                                         arr[nr++] = bh;
 244                                         continue;
 245                                 }
 246                                 /* Fully non-initialized data block, zero it. */
 247                                 goto handle_zblock;
 248                         }
 249                         /* It is a hole, need to zero it. */
 250                         if (lcn == LCN_HOLE)
 251                                 goto handle_hole;
 252                         /* If first try and run list unmapped, map and retry. */
 253                         if (!is_retry && lcn == LCN_RL_NOT_MAPPED) {
 254                                 is_retry = TRUE;
 255                                 /*
 256                                  * Attempt to map run list, dropping lock for
 257                                  * the duration.
 258                                  */
 259                                 up_read(&ni->run_list.lock);
 260                                 if (!map_run_list(ni, vcn))
 261                                         goto lock_retry_remap;
 262                                 rl = NULL;
 263                         }
 264                         /* Hard error, zero out region. */
 265                         SetPageError(page);
 266                         ntfs_error(vol->sb, "vcn_to_lcn(vcn = 0x%llx) failed "
 267                                         "with error code 0x%llx%s.",
 268                                         (unsigned long long)vcn,
 269                                         (unsigned long long)-lcn,
 270                                         is_retry ? " even after retrying" : "");
 271                         // FIXME: Depending on vol->on_errors, do something.
 272                 }
 273                 /*
 274                  * Either iblock was outside lblock limits or vcn_to_lcn()
 275                  * returned error. Just zero that portion of the page and set
 276                  * the buffer uptodate.
 277                  */
 278 handle_hole:
 279                 bh->b_blocknr = -1UL;
 280                 clear_buffer_mapped(bh);
 281 handle_zblock:
 282                 memset(kmap(page) + i * blocksize, 0, blocksize);
 283                 flush_dcache_page(page);
 284                 kunmap(page);
 285                 set_buffer_uptodate(bh);
 286         } while (i++, iblock++, (bh = bh->b_this_page) != head);
 287
 288         /* Release the lock if we took it. */
 289         if (rl)
 290                 up_read(&ni->run_list.lock);
 291
 292         /* Check we have at least one buffer ready for i/o. */
 293         if (nr) {
 294                 struct buffer_head *tbh;
 295
 296                 /* Lock the buffers. */
 297                 for (i = 0; i < nr; i++) {
 298                         tbh = arr[i];
 299                         lock_buffer(tbh);
 300                         tbh->b_end_io = ntfs_end_buffer_async_read;
 301                         set_buffer_async_read(tbh);
 302                 }
 303                 /* Finally, start i/o on the buffers. */
 304                 for (i = 0; i < nr; i++) {
 305                         tbh = arr[i];
 306                         if (likely(!buffer_uptodate(tbh)))
 307                                 submit_bh(READ, tbh);
 308                         else
 309                                 ntfs_end_buffer_async_read(tbh, 1);
 310                 }
 311                 return 0;
 312         }
 313         /* No i/o was scheduled on any of the buffers. */
 314         if (likely(!PageError(page)))
 315                 SetPageUptodate(page);
 316         else /* Signal synchronous i/o error. */
 317                 nr = -EIO;
 318         unlock_page(page);
 319         return nr;
 320 }
 321
 322 /**
 323  * ntfs_readpage - fill a @page of a @file with data from the device
 324  * @file:       open file to which the page @page belongs or NULL
 325  * @page:       page cache page to fill with data
 326  *
 327  * For non-resident attributes, ntfs_readpage() fills the @page of the open
 328  * file @file by calling the ntfs version of the generic block_read_full_page()
 329  * function, ntfs_read_block(), which in turn creates and reads in the buffers
 330  * associated with the page asynchronously.
 331  *
 332  * For resident attributes, OTOH, ntfs_readpage() fills @page by copying the
 333  * data from the mft record (which at this stage is most likely in memory) and
 334  * fills the remainder with zeroes. Thus, in this case, I/O is synchronous, as
 335  * even if the mft record is not cached at this point in time, we need to wait
 336  * for it to be read in before we can do the copy.
 337  *
 338  * Return 0 on success and -errno on error.
 339  *
 340  * WARNING: Do not make this function static! It is used by mft.c!
 341  */
 342 int ntfs_readpage(struct file *file, struct page *page)
 343 {
 344         s64 attr_pos;
 345         ntfs_inode *ni, *base_ni;
 346         char *addr;
 347         attr_search_context *ctx;
 348         MFT_RECORD *mrec;
 349         u32 attr_len;
 350         int err = 0;
 351
 352         BUG_ON(!PageLocked(page));
 353
 354         /*
 355          * This can potentially happen because we clear PageUptodate() during
 356          * ntfs_writepage() of MstProtected() attributes.
 357          */
 358         if (PageUptodate(page)) {
 359                 unlock_page(page);
 360                 return 0;
 361         }
 362
 363         ni = NTFS_I(page->mapping->host);
 364
 365         if (NInoNonResident(ni)) {
 366                 /*
 367                  * Only unnamed $DATA attributes can be compressed or
 368                  * encrypted.
 369                  */
 370                 if (ni->type == AT_DATA && !ni->name_len) {
 371                         /* If file is encrypted, deny access, just like NT4. */
 372                         if (NInoEncrypted(ni)) {
 373                                 err = -EACCES;
 374                                 goto err_out;
 375                         }
 376                         /* Compressed data streams are handled in compress.c. */
 377                         if (NInoCompressed(ni))
 378                                 return ntfs_read_compressed_block(page);
 379                 }
 380                 /* Normal data stream. */
 381                 return ntfs_read_block(page);
 382         }
 383         /* Attribute is resident, implying it is not compressed or encrypted. */
 384         if (!NInoAttr(ni))
 385                 base_ni = ni;
 386         else
 387                 base_ni = ni->ext.base_ntfs_ino;
 388
 389         /* Map, pin, and lock the mft record. */
 390         mrec = map_mft_record(base_ni);
 391         if (unlikely(IS_ERR(mrec))) {
 392                 err = PTR_ERR(mrec);
 393                 goto err_out;
 394         }
 395         ctx = get_attr_search_ctx(base_ni, mrec);
 396         if (unlikely(!ctx)) {
 397                 err = -ENOMEM;
 398                 goto unm_err_out;
 399         }
 400         if (unlikely(!lookup_attr(ni->type, ni->name, ni->name_len,
 401                         IGNORE_CASE, 0, NULL, 0, ctx))) {
 402                 err = -ENOENT;
 403                 goto put_unm_err_out;
 404         }
 405
 406         /* Starting position of the page within the attribute value. */
 407         attr_pos = page->index << PAGE_CACHE_SHIFT;
 408
 409         /* The total length of the attribute value. */
 410         attr_len = le32_to_cpu(ctx->attr->data.resident.value_length);
 411
 412         addr = kmap(page);
 413         /* Copy over in bounds data, zeroing the remainder of the page. */
 414         if (attr_pos < attr_len) {
 415                 u32 bytes = attr_len - attr_pos;
 416                 if (bytes > PAGE_CACHE_SIZE)
 417                         bytes = PAGE_CACHE_SIZE;
 418                 else if (bytes < PAGE_CACHE_SIZE)
 419                         memset(addr + bytes, 0, PAGE_CACHE_SIZE - bytes);
 420                 /* Copy the data to the page. */
 421                 memcpy(addr, attr_pos + (char*)ctx->attr +
 422                                 le16_to_cpu(
 423                                 ctx->attr->data.resident.value_offset), bytes);
 424         } else
 425                 memset(addr, 0, PAGE_CACHE_SIZE);
 426         flush_dcache_page(page);
 427         kunmap(page);
 428
 429         SetPageUptodate(page);
 430 put_unm_err_out:
 431         put_attr_search_ctx(ctx);
 432 unm_err_out:
 433         unmap_mft_record(base_ni);
 434 err_out:
 435         unlock_page(page);
 436         return err;
 437 }
 438
 439 #ifdef NTFS_RW
 440
 441 /**
 442  * ntfs_write_block - write a @page to the backing store
 443  * @page:       page cache page to write out
 444  *
 445  * This function is for writing pages belonging to non-resident, non-mst
 446  * protected attributes to their backing store.
 447  *
 448  * For a page with buffers, map and write the dirty buffers asynchronously
 449  * under page writeback. For a page without buffers, create buffers for the
 450  * page, then proceed as above.
 451  *
 452  * If a page doesn't have buffers the page dirty state is definitive. If a page
 453  * does have buffers, the page dirty state is just a hint, and the buffer dirty
 454  * state is definitive. (A hint which has rules: dirty buffers against a clean
 455  * page is illegal. Other combinations are legal and need to be handled. In
 456  * particular a dirty page containing clean buffers for example.)
 457  *
 458  * Return 0 on success and -errno on error.
 459  *
 460  * Based on ntfs_read_block() and __block_write_full_page().
 461  */
 462 static int ntfs_write_block(struct writeback_control *wbc, struct page *page)
 463 {
 464         VCN vcn;
 465         LCN lcn;
 466         sector_t block, dblock, iblock;
 467         struct inode *vi;
 468         ntfs_inode *ni;
 469         ntfs_volume *vol;
 470         run_list_element *rl;
 471         struct buffer_head *bh, *head;
 472         unsigned int blocksize, vcn_ofs;
 473         int err;
 474         BOOL need_end_writeback;
 475         unsigned char blocksize_bits;
 476
 477         vi = page->mapping->host;
 478         ni = NTFS_I(vi);
 479         vol = ni->vol;
 480
 481         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
 482                         "0x%lx.", vi->i_ino, ni->type, page->index);
 483
 484         BUG_ON(!NInoNonResident(ni));
 485         BUG_ON(NInoMstProtected(ni));
 486
 487         blocksize_bits = vi->i_blkbits;
 488         blocksize = 1 << blocksize_bits;
 489
 490         if (!page_has_buffers(page)) {
 491                 BUG_ON(!PageUptodate(page));
 492                 create_empty_buffers(page, blocksize,
 493                                 (1 << BH_Uptodate) | (1 << BH_Dirty));
 494         }
 495         bh = head = page_buffers(page);
 496         if (unlikely(!bh)) {
 497                 ntfs_warning(vol->sb, "Error allocating page buffers. "
 498                                 "Redirtying page so we try again later.");
 499                 /*
 500                  * Put the page back on mapping->dirty_pages, but leave its
 501                  * buffer's dirty state as-is.
 502                  */
 503                 redirty_page_for_writepage(wbc, page);
 504                 unlock_page(page);
 505                 return 0;
 506         }
 507
 508         /* NOTE: Different naming scheme to ntfs_read_block()! */
 509
 510         /* The first block in the page. */
 511         block = page->index << (PAGE_CACHE_SHIFT - blocksize_bits);
 512
 513         /* The first out of bounds block for the data size. */
 514         dblock = (vi->i_size + blocksize - 1) >> blocksize_bits;
 515
 516         /* The last (fully or partially) initialized block. */
 517         iblock = ni->initialized_size >> blocksize_bits;
 518
 519         /*
 520          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
 521          * here, and the (potentially unmapped) buffers may become dirty at
 522          * any time.  If a buffer becomes dirty here after we've inspected it
 523          * then we just miss that fact, and the page stays dirty.
 524          *
 525          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
 526          * handle that here by just cleaning them.
 527          */
 528
 529         /*
 530          * Loop through all the buffers in the page, mapping all the dirty
 531          * buffers to disk addresses and handling any aliases from the
 532          * underlying block device's mapping.
 533          */
 534         rl = NULL;
 535         err = 0;
 536         do {
 537                 BOOL is_retry = FALSE;
 538
 539                 if (unlikely(block >= dblock)) {
 540                         /*
 541                          * Mapped buffers outside i_size will occur, because
 542                          * this page can be outside i_size when there is a
 543                          * truncate in progress. The contents of such buffers
 544                          * were zeroed by ntfs_writepage().
 545                          *
 546                          * FIXME: What about the small race window where
 547                          * ntfs_writepage() has not done any clearing because
 548                          * the page was within i_size but before we get here,
 549                          * vmtruncate() modifies i_size?
 550                          */
 551                         clear_buffer_dirty(bh);
 552                         set_buffer_uptodate(bh);
 553                         continue;
 554                 }
 555
 556                 /* Clean buffers are not written out, so no need to map them. */
 557                 if (!buffer_dirty(bh))
 558                         continue;
 559
 560                 /* Make sure we have enough initialized size. */
 561                 if (unlikely((block >= iblock) &&
 562                                 (ni->initialized_size < vi->i_size))) {
 563                         /*
 564                          * If this page is fully outside initialized size, zero
 565                          * out all pages between the current initialized size
 566                          * and the current page. Just use ntfs_readpage() to do
 567                          * the zeroing transparently.
 568                          */
 569                         if (block > iblock) {
 570                                 // TODO:
 571                                 // For each page do:
 572                                 // - read_cache_page()
 573                                 // Again for each page do:
 574                                 // - wait_on_page_locked()
 575                                 // - Check (PageUptodate(page) &&
 576                                 //                      !PageError(page))
 577                                 // Update initialized size in the attribute and
 578                                 // in the inode.
 579                                 // Again, for each page do:
 580                                 //      __set_page_dirty_buffers();
 581                                 // page_cache_release()
 582                                 // We don't need to wait on the writes.
 583                                 // Update iblock.
 584                         }
 585                         /*
 586                          * The current page straddles initialized size. Zero
 587                          * all non-uptodate buffers and set them uptodate (and
 588                          * dirty?). Note, there aren't any non-uptodate buffers
 589                          * if the page is uptodate.
 590                          * FIXME: For an uptodate page, the buffers may need to
 591                          * be written out because they were not initialized on
 592                          * disk before.
 593                          */
 594                         if (!PageUptodate(page)) {
 595                                 // TODO:
 596                                 // Zero any non-uptodate buffers up to i_size.
 597                                 // Set them uptodate and dirty.
 598                         }
 599                         // TODO:
 600                         // Update initialized size in the attribute and in the
 601                         // inode (up to i_size).
 602                         // Update iblock.
 603                         // FIXME: This is inefficient. Try to batch the two
 604                         // size changes to happen in one go.
 605                         ntfs_error(vol->sb, "Writing beyond initialized size "
 606                                         "is not supported yet. Sorry.");
 607                         err = -EOPNOTSUPP;
 608                         break;
 609                         // Do NOT set_buffer_new() BUT DO clear buffer range
 610                         // outside write request range.
 611                         // set_buffer_uptodate() on complete buffers as well as
 612                         // set_buffer_dirty().
 613                 }
 614
 615                 /* No need to map buffers that are already mapped. */
 616                 if (buffer_mapped(bh))
 617                         continue;
 618
 619                 /* Unmapped, dirty buffer. Need to map it. */
 620                 bh->b_bdev = vol->sb->s_bdev;
 621
 622                 /* Convert block into corresponding vcn and offset. */
 623                 vcn = (VCN)block << blocksize_bits >> vol->cluster_size_bits;
 624                 vcn_ofs = ((VCN)block << blocksize_bits) &
 625                                 vol->cluster_size_mask;
 626                 if (!rl) {
 627 lock_retry_remap:
 628                         down_read(&ni->run_list.lock);
 629                         rl = ni->run_list.rl;
 630                 }
 631                 if (likely(rl != NULL)) {
 632                         /* Seek to element containing target vcn. */
 633                         while (rl->length && rl[1].vcn <= vcn)
 634                                 rl++;
 635                         lcn = vcn_to_lcn(rl, vcn);
 636                 } else
 637                         lcn = (LCN)LCN_RL_NOT_MAPPED;
 638                 /* Successful remap. */
 639                 if (lcn >= 0) {
 640                         /* Setup buffer head to point to correct block. */
 641                         bh->b_blocknr = ((lcn << vol->cluster_size_bits) +
 642                                         vcn_ofs) >> blocksize_bits;
 643                         set_buffer_mapped(bh);
 644                         continue;
 645                 }
 646                 /* It is a hole, need to instantiate it. */
 647                 if (lcn == LCN_HOLE) {
 648                         // TODO: Instantiate the hole.
 649                         // clear_buffer_new(bh);
 650                         // unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 651                         ntfs_error(vol->sb, "Writing into sparse regions is "
 652                                         "not supported yet. Sorry.");
 653                         err = -EOPNOTSUPP;
 654                         break;
 655                 }
 656                 /* If first try and run list unmapped, map and retry. */
 657                 if (!is_retry && lcn == LCN_RL_NOT_MAPPED) {
 658                         is_retry = TRUE;
 659                         /*
 660                          * Attempt to map run list, dropping lock for
 661                          * the duration.
 662                          */
 663                         up_read(&ni->run_list.lock);
 664                         err = map_run_list(ni, vcn);
 665                         if (likely(!err))
 666                                 goto lock_retry_remap;
 667                         rl = NULL;
 668                 }
 669                 /* Failed to map the buffer, even after retrying. */
 670                 bh->b_blocknr = -1UL;
 671                 ntfs_error(vol->sb, "vcn_to_lcn(vcn = 0x%llx) failed "
 672                                 "with error code 0x%llx%s.",
 673                                 (unsigned long long)vcn,
 674                                 (unsigned long long)-lcn,
 675                                 is_retry ? " even after retrying" : "");
 676                 // FIXME: Depending on vol->on_errors, do something.
 677                 if (!err)
 678                         err = -EIO;
 679                 break;
 680         } while (block++, (bh = bh->b_this_page) != head);
 681
 682         /* Release the lock if we took it. */
 683         if (rl)
 684                 up_read(&ni->run_list.lock);
 685
 686         /* For the error case, need to reset bh to the beginning. */
 687         bh = head;
 688
 689         /* Just an optimization, so ->readpage() isn't called later. */
 690         if (unlikely(!PageUptodate(page))) {
 691                 int uptodate = 1;
 692                 do {
 693                         if (!buffer_uptodate(bh)) {
 694                                 uptodate = 0;
 695                                 bh = head;
 696                                 break;
 697                         }
 698                 } while ((bh = bh->b_this_page) != head);
 699                 if (uptodate)
 700                         SetPageUptodate(page);
 701         }
 702
 703         /* Setup all mapped, dirty buffers for async write i/o. */
 704         do {
 705                 get_bh(bh);
 706                 if (buffer_mapped(bh) && buffer_dirty(bh)) {
 707                         lock_buffer(bh);
 708                         if (test_clear_buffer_dirty(bh)) {
 709                                 BUG_ON(!buffer_uptodate(bh));
 710                                 mark_buffer_async_write(bh);
 711                         } else
 712                                 unlock_buffer(bh);
 713                 } else if (unlikely(err)) {
 714                         /*
 715                          * For the error case. The buffer may have been set
 716                          * dirty during attachment to a dirty page.
 717                          */
 718                         if (err != -ENOMEM)
 719                                 clear_buffer_dirty(bh);
 720                 }
 721         } while ((bh = bh->b_this_page) != head);
 722
 723         if (unlikely(err)) {
 724                 // TODO: Remove the -EOPNOTSUPP check later on...
 725                 if (unlikely(err == -EOPNOTSUPP))
 726                         err = 0;
 727                 else if (err == -ENOMEM) {
 728                         ntfs_warning(vol->sb, "Error allocating memory. "
 729                                         "Redirtying page so we try again "
 730                                         "later.");
 731                         /*
 732                          * Put the page back on mapping->dirty_pages, but
 733                          * leave its buffer's dirty state as-is.
 734                          */
 735                         redirty_page_for_writepage(wbc, page);
 736                         err = 0;
 737                 } else
 738                         SetPageError(page);
 739         }
 740
 741         BUG_ON(PageWriteback(page));
 742         set_page_writeback(page);       /* Keeps try_to_free_buffers() away. */
 743         unlock_page(page);
 744
 745         /*
 746          * Submit the prepared buffers for i/o. Note the page is unlocked,
 747          * and the async write i/o completion handler can end_page_writeback()
 748          * at any time after the *first* submit_bh(). So the buffers can then
 749          * disappear...
 750          */
 751         need_end_writeback = TRUE;
 752         do {
 753                 struct buffer_head *next = bh->b_this_page;
 754                 if (buffer_async_write(bh)) {
 755                         submit_bh(WRITE, bh);
 756                         need_end_writeback = FALSE;
 757                 }
 758                 put_bh(bh);
 759                 bh = next;
 760         } while (bh != head);
 761
 762         /* If no i/o was started, need to end_page_writeback(). */
 763         if (unlikely(need_end_writeback))
 764                 end_page_writeback(page);
 765
 766         ntfs_debug("Done.");
 767         return err;
 768 }
 769
 770 /**
 771  * ntfs_writepage - write a @page to the backing store
 772  * @page:       page cache page to write out
 773  *
 774  * For non-resident attributes, ntfs_writepage() writes the @page by calling
 775  * the ntfs version of the generic block_write_full_page() function,
 776  * ntfs_write_block(), which in turn if necessary creates and writes the
 777  * buffers associated with the page asynchronously.
 778  *
 779  * For resident attributes, OTOH, ntfs_writepage() writes the @page by copying
 780  * the data to the mft record (which at this stage is most likely in memory).
 781  * The mft record is then marked dirty and written out asynchronously via the
 782  * vfs inode dirty code path.
 783  *
 784  * Note the caller clears the page dirty flag before calling ntfs_writepage().
 785  *
 786  * Based on ntfs_readpage() and fs/buffer.c::block_write_full_page().
 787  *
 788  * Return 0 on success and -errno on error.
 789  */
 790 static int ntfs_writepage(struct page *page, struct writeback_control *wbc)
 791 {
 792         s64 attr_pos;
 793         struct inode *vi;
 794         ntfs_inode *ni, *base_ni;
 795         char *kaddr;
 796         attr_search_context *ctx;
 797         MFT_RECORD *m;
 798         u32 attr_len, bytes;
 799         int err;
 800
 801         BUG_ON(!PageLocked(page));
 802
 803         vi = page->mapping->host;
 804
 805         /* Is the page fully outside i_size? (truncate in progress) */
 806         if (unlikely(page->index >= (vi->i_size + PAGE_CACHE_SIZE - 1) >>
 807                         PAGE_CACHE_SHIFT)) {
 808                 unlock_page(page);
 809                 ntfs_debug("Write outside i_size - truncated?");
 810                 return 0;
 811         }
 812
 813         ni = NTFS_I(vi);
 814
 815         if (NInoNonResident(ni)) {
 816                 /*
 817                  * Only unnamed $DATA attributes can be compressed, encrypted,
 818                  * and/or sparse.
 819                  */
 820                 if (ni->type == AT_DATA && !ni->name_len) {
 821                         /* If file is encrypted, deny access, just like NT4. */
 822                         if (NInoEncrypted(ni)) {
 823                                 unlock_page(page);
 824                                 ntfs_debug("Denying write access to encrypted "
 825                                                 "file.");
 826                                 return -EACCES;
 827                         }
 828                         /* Compressed data streams are handled in compress.c. */
 829                         if (NInoCompressed(ni)) {
 830                                 // TODO: Implement and replace this check with
 831                                 // return ntfs_write_compressed_block(page);
 832                                 unlock_page(page);
 833                                 ntfs_error(vi->i_sb, "Writing to compressed "
 834                                                 "files is not supported yet. "
 835                                                 "Sorry.");
 836                                 return -EOPNOTSUPP;
 837                         }
 838                         // TODO: Implement and remove this check.
 839                         if (NInoSparse(ni)) {
 840                                 unlock_page(page);
 841                                 ntfs_error(vi->i_sb, "Writing to sparse files "
 842                                                 "is not supported yet. Sorry.");
 843                                 return -EOPNOTSUPP;
 844                         }
 845                 }
 846
 847                 /* We have to zero every time due to mmap-at-end-of-file. */
 848                 if (page->index >= (vi->i_size >> PAGE_CACHE_SHIFT)) {
 849                         /* The page straddles i_size. */
 850                         unsigned int ofs = vi->i_size & ~PAGE_CACHE_MASK;
 851                         kaddr = kmap_atomic(page, KM_USER0);
 852                         memset(kaddr + ofs, 0, PAGE_CACHE_SIZE - ofs);
 853                         flush_dcache_page(page);
 854                         kunmap_atomic(kaddr, KM_USER0);
 855                 }
 856
 857                 // TODO: Implement and remove this check.
 858                 if (NInoMstProtected(ni)) {
 859                         unlock_page(page);
 860                         ntfs_error(vi->i_sb, "Writing to MST protected "
 861                                         "attributes is not supported yet. "
 862                                         "Sorry.");
 863                         return -EOPNOTSUPP;
 864                 }
 865
 866                 /* Normal data stream. */
 867                 return ntfs_write_block(wbc, page);
 868         }
 869
 870         /*
 871          * Attribute is resident, implying it is not compressed, encrypted, or
 872          * mst protected.
 873          */
 874         BUG_ON(page_has_buffers(page));
 875         BUG_ON(!PageUptodate(page));
 876
 877         if (!NInoAttr(ni))
 878                 base_ni = ni;
 879         else
 880                 base_ni = ni->ext.base_ntfs_ino;
 881
 882         /* Map, pin, and lock the mft record. */
 883         m = map_mft_record(base_ni);
 884         if (unlikely(IS_ERR(m))) {
 885                 err = PTR_ERR(m);
 886                 m = NULL;
 887                 ctx = NULL;
 888                 goto err_out;
 889         }
 890         ctx = get_attr_search_ctx(base_ni, m);
 891         if (unlikely(!ctx)) {
 892                 err = -ENOMEM;
 893                 goto err_out;
 894         }
 895         if (unlikely(!lookup_attr(ni->type, ni->name, ni->name_len,
 896                         IGNORE_CASE, 0, NULL, 0, ctx))) {
 897                 err = -ENOENT;
 898                 goto err_out;
 899         }
 900
 901         /* Starting position of the page within the attribute value. */
 902         attr_pos = page->index << PAGE_CACHE_SHIFT;
 903
 904         /* The total length of the attribute value. */
 905         attr_len = le32_to_cpu(ctx->attr->data.resident.value_length);
 906
 907         if (unlikely(vi->i_size != attr_len)) {
 908                 ntfs_error(vi->i_sb, "BUG()! i_size (0x%llx) doesn't match "
 909                                 "attr_len (0x%x). Aborting write.", vi->i_size,
 910                                 attr_len);
 911                 err = -EIO;
 912                 goto err_out;
 913         }
 914         if (unlikely(attr_pos >= attr_len)) {
 915                 ntfs_error(vi->i_sb, "BUG()! attr_pos (0x%llx) > attr_len "
 916                                 "(0x%x). Aborting write.",
 917                                 (unsigned long long)attr_pos, attr_len);
 918                 err = -EIO;
 919                 goto err_out;
 920         }
 921
 922         bytes = attr_len - attr_pos;
 923         if (unlikely(bytes > PAGE_CACHE_SIZE))
 924                 bytes = PAGE_CACHE_SIZE;
 925
 926         /*
 927          * Keep the VM happy.  This must be done otherwise the radix-tree tag
 928          * PAGECACHE_TAG_DIRTY remains set even though the page is clean.
 929          */
 930         BUG_ON(PageWriteback(page));
 931         set_page_writeback(page);
 932         unlock_page(page);
 933
 934         /*
 935          * Here, we don't need to zero the out of bounds area everytime because
 936          * the below memcpy() already takes care of the mmap-at-end-of-file
 937          * requirements. If the file is converted to a non-resident one, then
 938          * the code path use is switched to the non-resident one where the
 939          * zeroing happens on each ntfs_writepage() invocation.
 940          *
 941          * The above also applies nicely when i_size is decreased.
 942          *
 943          * When i_size is increased, the memory between the old and new i_size
 944          * _must_ be zeroed (or overwritten with new data). Otherwise we will
 945          * expose data to userspace/disk which should never have been exposed.
 946          *
 947          * FIXME: Ensure that i_size increases do the zeroing/overwriting and
 948          * if we cannot guarantee that, then enable the zeroing below.  If the
 949          * zeroing below is enabled, we MUST move the unlock_page() from above
 950          * to after the kunmap_atomic(), i.e. just before the
 951          * end_page_writeback().
 952          */
 953
 954         kaddr = kmap_atomic(page, KM_USER0);
 955         /* Copy the data from the page to the mft record. */
 956         memcpy((u8*)ctx->attr + le16_to_cpu(
 957                         ctx->attr->data.resident.value_offset) + attr_pos,
 958                         kaddr, bytes);
 959         flush_dcache_mft_record_page(ctx->ntfs_ino);
 960 #if 0
 961         /* Zero out of bounds area. */
 962         if (likely(bytes < PAGE_CACHE_SIZE)) {
 963                 memset(kaddr + bytes, 0, PAGE_CACHE_SIZE - bytes);
 964                 flush_dcache_page(page);
 965         }
 966 #endif
 967         kunmap_atomic(kaddr, KM_USER0);
 968
 969         end_page_writeback(page);
 970
 971         /* Mark the mft record dirty, so it gets written back. */
 972         mark_mft_record_dirty(ctx->ntfs_ino);
 973
 974         put_attr_search_ctx(ctx);
 975         unmap_mft_record(base_ni);
 976         return 0;
 977 err_out:
 978         if (err == -ENOMEM) {
 979                 ntfs_warning(vi->i_sb, "Error allocating memory. Redirtying "
 980                                 "page so we try again later.");
 981                 /*
 982                  * Put the page back on mapping->dirty_pages, but leave its
 983                  * buffer's dirty state as-is.
 984                  */
 985                 redirty_page_for_writepage(wbc, page);
 986                 err = 0;
 987         } else {
 988                 ntfs_error(vi->i_sb, "Resident attribute write failed with "
 989                                 "error %i. Setting page error flag.", -err);
 990                 SetPageError(page);
 991         }
 992         unlock_page(page);
 993         if (ctx)
 994                 put_attr_search_ctx(ctx);
 995         if (m)
 996                 unmap_mft_record(base_ni);
 997         return err;
 998 }
 999
1000 /**
1001  * ntfs_prepare_nonresident_write -
1002  *
1003  */
1004 static int ntfs_prepare_nonresident_write(struct page *page,
1005                 unsigned from, unsigned to)
1006 {
1007         VCN vcn;
1008         LCN lcn;
1009         sector_t block, ablock, iblock;
1010         struct inode *vi;
1011         ntfs_inode *ni;
1012         ntfs_volume *vol;
1013         run_list_element *rl;
1014         struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
1015         unsigned int vcn_ofs, block_start, block_end, blocksize;
1016         int err;
1017         BOOL is_retry;
1018         unsigned char blocksize_bits;
1019
1020         vi = page->mapping->host;
1021         ni = NTFS_I(vi);
1022         vol = ni->vol;
1023
1024         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
1025                         "0x%lx, from = %u, to = %u.", vi->i_ino, ni->type,
1026                         page->index, from, to);
1027
1028         BUG_ON(!NInoNonResident(ni));
1029         BUG_ON(NInoMstProtected(ni));
1030
1031         blocksize_bits = vi->i_blkbits;
1032         blocksize = 1 << blocksize_bits;
1033
1034         /*
1035          * create_empty_buffers() will create uptodate/dirty buffers if the
1036          * page is uptodate/dirty.
1037          */
1038         if (!page_has_buffers(page))
1039                 create_empty_buffers(page, blocksize, 0);
1040         bh = head = page_buffers(page);
1041         if (unlikely(!bh))
1042                 return -ENOMEM;
1043
1044         /* The first block in the page. */
1045         block = page->index << (PAGE_CACHE_SHIFT - blocksize_bits);
1046
1047         /*
1048          * The first out of bounds block for the allocated size. No need to
1049          * round up as allocated_size is in multiples of cluster size and the
1050          * minimum cluster size is 512 bytes, which is equal to the smallest
1051          * blocksize.
1052          */
1053         ablock = ni->allocated_size >> blocksize_bits;
1054
1055         /* The last (fully or partially) initialized block. */
1056         iblock = ni->initialized_size >> blocksize_bits;
1057
1058         /* Loop through all the buffers in the page. */
1059         block_start = 0;
1060         rl = NULL;
1061         err = 0;
1062         do {
1063                 block_end = block_start + blocksize;
1064                 /*
1065                  * If buffer @bh is outside the write, just mark it uptodate
1066                  * if the page is uptodate and continue with the next buffer.
1067                  */
1068                 if (block_end <= from || block_start >= to) {
1069                         if (PageUptodate(page)) {
1070                                 if (!buffer_uptodate(bh))
1071                                         set_buffer_uptodate(bh);
1072                         }
1073                         continue;
1074                 }
1075                 /*
1076                  * @bh is at least partially being written to.
1077                  * Make sure it is not marked as new.
1078                  */
1079                 //if (buffer_new(bh))
1080                 //      clear_buffer_new(bh);
1081
1082                 if (block >= ablock) {
1083                         // TODO: block is above allocated_size, need to
1084                         // allocate it. Best done in one go to accommodate not
1085                         // only block but all above blocks up to and including:
1086                         // ((page->index << PAGE_CACHE_SHIFT) + to + blocksize
1087                         // - 1) >> blobksize_bits. Obviously will need to round
1088                         // up to next cluster boundary, too. This should be
1089                         // done with a helper function, so it can be reused.
1090                         ntfs_error(vol->sb, "Writing beyond allocated size "
1091                                         "is not supported yet. Sorry.");
1092                         err = -EOPNOTSUPP;
1093                         goto err_out;
1094                         // Need to update ablock.
1095                         // Need to set_buffer_new() on all block bhs that are
1096                         // newly allocated.
1097                 }
1098                 /*
1099                  * Now we have enough allocated size to fulfill the whole
1100                  * request, i.e. block < ablock is true.
1101                  */
1102                 if (unlikely((block >= iblock) &&
1103                                 (ni->initialized_size < vi->i_size))) {
1104                         /*
1105                          * If this page is fully outside initialized size, zero
1106                          * out all pages between the current initialized size
1107                          * and the current page. Just use ntfs_readpage() to do
1108                          * the zeroing transparently.
1109                          */
1110                         if (block > iblock) {
1111                                 // TODO:
1112                                 // For each page do:
1113                                 // - read_cache_page()
1114                                 // Again for each page do:
1115                                 // - wait_on_page_locked()
1116                                 // - Check (PageUptodate(page) &&
1117                                 //                      !PageError(page))
1118                                 // Update initialized size in the attribute and
1119                                 // in the inode.
1120                                 // Again, for each page do:
1121                                 //      __set_page_dirty_buffers();
1122                                 // page_cache_release()
1123                                 // We don't need to wait on the writes.
1124                                 // Update iblock.
1125                         }
1126                         /*
1127                          * The current page straddles initialized size. Zero
1128                          * all non-uptodate buffers and set them uptodate (and
1129                          * dirty?). Note, there aren't any non-uptodate buffers
1130                          * if the page is uptodate.
1131                          * FIXME: For an uptodate page, the buffers may need to
1132                          * be written out because they were not initialized on
1133                          * disk before.
1134                          */
1135                         if (!PageUptodate(page)) {
1136                                 // TODO:
1137                                 // Zero any non-uptodate buffers up to i_size.
1138                                 // Set them uptodate and dirty.
1139                         }
1140                         // TODO:
1141                         // Update initialized size in the attribute and in the
1142                         // inode (up to i_size).
1143                         // Update iblock.
1144                         // FIXME: This is inefficient. Try to batch the two
1145                         // size changes to happen in one go.
1146                         ntfs_error(vol->sb, "Writing beyond initialized size "
1147                                         "is not supported yet. Sorry.");
1148                         err = -EOPNOTSUPP;
1149                         goto err_out;
1150                         // Do NOT set_buffer_new() BUT DO clear buffer range
1151                         // outside write request range.
1152                         // set_buffer_uptodate() on complete buffers as well as
1153                         // set_buffer_dirty().
1154                 }
1155
1156                 /* Need to map unmapped buffers. */
1157                 if (!buffer_mapped(bh)) {
1158                         /* Unmapped buffer. Need to map it. */
1159                         bh->b_bdev = vol->sb->s_bdev;
1160
1161                         /* Convert block into corresponding vcn and offset. */
1162                         vcn = (VCN)block << blocksize_bits >>
1163                                         vol->cluster_size_bits;
1164                         vcn_ofs = ((VCN)block << blocksize_bits) &
1165                                         vol->cluster_size_mask;
1166
1167                         is_retry = FALSE;
1168                         if (!rl) {
1169 lock_retry_remap:
1170                                 down_read(&ni->run_list.lock);
1171                                 rl = ni->run_list.rl;
1172                         }
1173                         if (likely(rl != NULL)) {
1174                                 /* Seek to element containing target vcn. */
1175                                 while (rl->length && rl[1].vcn <= vcn)
1176                                         rl++;
1177                                 lcn = vcn_to_lcn(rl, vcn);
1178                         } else
1179                                 lcn = (LCN)LCN_RL_NOT_MAPPED;
1180                         if (unlikely(lcn < 0)) {
1181                                 /*
1182                                  * We extended the attribute allocation above.
1183                                  * If we hit an ENOENT here it means that the
1184                                  * allocation was insufficient which is a bug.
1185                                  */
1186                                 BUG_ON(lcn == LCN_ENOENT);
1187
1188                                 /* It is a hole, need to instantiate it. */
1189                                 if (lcn == LCN_HOLE) {
1190                                         // TODO: Instantiate the hole.
1191                                         // clear_buffer_new(bh);
1192                                         // unmap_underlying_metadata(bh->b_bdev,
1193                                         //              bh->b_blocknr);
1194                                         // For non-uptodate buffers, need to
1195                                         // zero out the region outside the
1196                                         // request in this bh or all bhs,
1197                                         // depending on what we implemented
1198                                         // above.
1199                                         // Need to flush_dcache_page().
1200                                         // Or could use set_buffer_new()
1201                                         // instead?
1202                                         ntfs_error(vol->sb, "Writing into "
1203                                                         "sparse regions is "
1204                                                         "not supported yet. "
1205                                                         "Sorry.");
1206                                         err = -EOPNOTSUPP;
1207                                         goto err_out;
1208                                 } else if (!is_retry &&
1209                                                 lcn == LCN_RL_NOT_MAPPED) {
1210                                         is_retry = TRUE;
1211                                         /*
1212                                          * Attempt to map run list, dropping
1213                                          * lock for the duration.
1214                                          */
1215                                         up_read(&ni->run_list.lock);
1216                                         err = map_run_list(ni, vcn);
1217                                         if (likely(!err))
1218                                                 goto lock_retry_remap;
1219                                         rl = NULL;
1220                                 }
1221                                 /*
1222                                  * Failed to map the buffer, even after
1223                                  * retrying.
1224                                  */
1225                                 bh->b_blocknr = -1UL;
1226                                 ntfs_error(vol->sb, "vcn_to_lcn(vcn = 0x%llx) "
1227                                                 "failed with error code "
1228                                                 "0x%llx%s.",
1229                                                 (unsigned long long)vcn,
1230                                                 (unsigned long long)-lcn,
1231                                                 is_retry ? " even after "
1232                                                 "retrying" : "");
1233                                 // FIXME: Depending on vol->on_errors, do
1234                                 // something.
1235                                 if (!err)
1236                                         err = -EIO;
1237                                 goto err_out;
1238                         }
1239                         /* We now have a successful remap, i.e. lcn >= 0. */
1240
1241                         /* Setup buffer head to correct block. */
1242                         bh->b_blocknr = ((lcn << vol->cluster_size_bits)
1243                                         + vcn_ofs) >> blocksize_bits;
1244                         set_buffer_mapped(bh);
1245
1246                         // FIXME: Something analogous to this is needed for
1247                         // each newly allocated block, i.e. BH_New.
1248                         // FIXME: Might need to take this out of the
1249                         // if (!buffer_mapped(bh)) {}, depending on how we
1250                         // implement things during the allocated_size and
1251                         // initialized_size extension code above.
1252                         if (buffer_new(bh)) {
1253                                 clear_buffer_new(bh);
1254                                 unmap_underlying_metadata(bh->b_bdev,
1255                                                 bh->b_blocknr);
1256                                 if (PageUptodate(page)) {
1257                                         set_buffer_uptodate(bh);
1258                                         continue;
1259                                 }
1260                                 /*
1261                                  * Page is _not_ uptodate, zero surrounding
1262                                  * region. NOTE: This is how we decide if to
1263                                  * zero or not!
1264                                  */
1265                                 if (block_end > to || block_start < from) {
1266                                         void *kaddr;
1267
1268                                         kaddr = kmap_atomic(page, KM_USER0);
1269                                         if (block_end > to)
1270                                                 memset(kaddr + to, 0,
1271                                                                 block_end - to);
1272                                         if (block_start < from)
1273                                                 memset(kaddr + block_start, 0,
1274                                                                 from -
1275                                                                 block_start);
1276                                         flush_dcache_page(page);
1277                                         kunmap_atomic(kaddr, KM_USER0);
1278                                 }
1279                                 continue;
1280                         }
1281                 }
1282                 /* @bh is mapped, set it uptodate if the page is uptodate. */
1283                 if (PageUptodate(page)) {
1284                         if (!buffer_uptodate(bh))
1285                                 set_buffer_uptodate(bh);
1286                         continue;
1287                 }
1288                 /*
1289                  * The page is not uptodate. The buffer is mapped. If it is not
1290                  * uptodate, and it is only partially being written to, we need
1291                  * to read the buffer in before the write, i.e. right now.
1292                  */
1293                 if (!buffer_uptodate(bh) &&
1294                                 (block_start < from || block_end > to)) {
1295                         ll_rw_block(READ, 1, &bh);
1296                         *wait_bh++ = bh;
1297                 }
1298         } while (block++, block_start = block_end,
1299                         (bh = bh->b_this_page) != head);
1300
1301         /* Release the lock if we took it. */
1302         if (rl) {
1303                 up_read(&ni->run_list.lock);
1304                 rl = NULL;
1305         }
1306
1307         /* If we issued read requests, let them complete. */
1308         while (wait_bh > wait) {
1309                 wait_on_buffer(*--wait_bh);
1310                 if (!buffer_uptodate(*wait_bh))
1311                         return -EIO;
1312         }
1313
1314         ntfs_debug("Done.");
1315         return 0;
1316 err_out:
1317         /*
1318          * Zero out any newly allocated blocks to avoid exposing stale data.
1319          * If BH_New is set, we know that the block was newly allocated in the
1320          * above loop.
1321          * FIXME: What about initialized_size increments? Have we done all the
1322          * required zeroing above? If not this error handling is broken, and
1323          * in particular the if (block_end <= from) check is completely bogus.
1324          */
1325         bh = head;
1326         block_start = 0;
1327         is_retry = FALSE;
1328         do {
1329                 block_end = block_start + blocksize;
1330                 if (block_end <= from)
1331                         continue;
1332                 if (block_start >= to)
1333                         break;
1334                 if (buffer_new(bh)) {
1335                         void *kaddr;
1336
1337                         clear_buffer_new(bh);
1338                         kaddr = kmap_atomic(page, KM_USER0);
1339                         memset(kaddr + block_start, 0, bh->b_size);
1340                         kunmap_atomic(kaddr, KM_USER0);
1341                         set_buffer_uptodate(bh);
1342                         mark_buffer_dirty(bh);
1343                         is_retry = TRUE;
1344                 }
1345         } while (block_start = block_end, (bh = bh->b_this_page) != head);
1346         if (is_retry)
1347                 flush_dcache_page(page);
1348         if (rl)
1349                 up_read(&ni->run_list.lock);
1350         return err;
1351 }
1352
1353 /**
1354  * ntfs_prepare_write - prepare a page for receiving data
1355  *
1356  * This is called from generic_file_write() with i_sem held on the inode
1357  * (@page->mapping->host). The @page is locked and kmap()ped so page_address()
1358  * can simply be used. The source data has not yet been copied into the @page.
1359  *
1360  * Need to extend the attribute/fill in holes if necessary, create blocks and
1361  * make partially overwritten blocks uptodate,
1362  *
1363  * i_size is not to be modified yet.
1364  *
1365  * Return 0 on success or -errno on error.
1366  *
1367  * Should be using block_prepare_write() [support for sparse files] or
1368  * cont_prepare_write() [no support for sparse files]. Can't do that due to
1369  * ntfs specifics but can look at them for implementation guidancea.
1370  *
1371  * Note: In the range, @from is inclusive and @to is exclusive, i.e. @from is
1372  * the first byte in the page that will be written to and @to is the first byte
1373  * after the last byte that will be written to.
1374  */
1375 static int ntfs_prepare_write(struct file *file, struct page *page,
1376                 unsigned from, unsigned to)
1377 {
1378         struct inode *vi = page->mapping->host;
1379         ntfs_inode   *ni = NTFS_I(vi);
1380
1381         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
1382                         "0x%lx, from = %u, to = %u.", vi->i_ino, ni->type,
1383                         page->index, from, to);
1384
1385         BUG_ON(!PageLocked(page));
1386         BUG_ON(from > PAGE_CACHE_SIZE);
1387         BUG_ON(to > PAGE_CACHE_SIZE);
1388         BUG_ON(from > to);
1389
1390         if (NInoNonResident(ni)) {
1391                 /*
1392                  * Only unnamed $DATA attributes can be compressed, encrypted,
1393                  * and/or sparse.
1394                  */
1395                 if (ni->type == AT_DATA && !ni->name_len) {
1396                         /* If file is encrypted, deny access, just like NT4. */
1397                         if (NInoEncrypted(ni)) {
1398                                 ntfs_debug("Denying write access to encrypted "
1399                                                 "file.");
1400                                 return -EACCES;
1401                         }
1402                         /* Compressed data streams are handled in compress.c. */
1403                         if (NInoCompressed(ni)) {
1404                                 // TODO: Implement and replace this check with
1405                                 // return ntfs_write_compressed_block(page);
1406                                 ntfs_error(vi->i_sb, "Writing to compressed "
1407                                                 "files is not supported yet. "
1408                                                 "Sorry.");
1409                                 return -EOPNOTSUPP;
1410                         }
1411                         // TODO: Implement and remove this check.
1412                         if (NInoSparse(ni)) {
1413                                 ntfs_error(vi->i_sb, "Writing to sparse files "
1414                                                 "is not supported yet. Sorry.");
1415                                 return -EOPNOTSUPP;
1416                         }
1417                 }
1418
1419                 // TODO: Implement and remove this check.
1420                 if (NInoMstProtected(ni)) {
1421                         ntfs_error(vi->i_sb, "Writing to MST protected "
1422                                         "attributes is not supported yet. "
1423                                         "Sorry.");
1424                         return -EOPNOTSUPP;
1425                 }
1426
1427                 /* Normal data stream. */
1428                 return ntfs_prepare_nonresident_write(page, from, to);
1429         }
1430
1431         /*
1432          * Attribute is resident, implying it is not compressed, encrypted, or
1433          * mst protected.
1434          */
1435         BUG_ON(page_has_buffers(page));
1436
1437         /* Do we need to resize the attribute? */
1438         if (((s64)page->index << PAGE_CACHE_SHIFT) + to > vi->i_size) {
1439                 // TODO: Implement resize...
1440                 ntfs_error(vi->i_sb, "Writing beyond the existing file size is "
1441                                 "not supported yet. Sorry.");
1442                 return -EOPNOTSUPP;
1443         }
1444
1445         /*
1446          * Because resident attributes are handled by memcpy() to/from the
1447          * corresponding MFT record, and because this form of i/o is byte
1448          * aligned rather than block aligned, there is no need to bring the
1449          * page uptodate here as in the non-resident case where we need to
1450          * bring the buffers straddled by the write uptodate before
1451          * generic_file_write() does the copying from userspace.
1452          *
1453          * We thus defer the uptodate bringing of the page region outside the
1454          * region written to to ntfs_commit_write(). The reason for doing this
1455          * is that we save one round of:
1456          *      map_mft_record(), get_attr_search_ctx(), lookup_attr(),
1457          *      kmap_atomic(), kunmap_atomic(), put_attr_search_ctx(),
1458          *      unmap_mft_record().
1459          * Which is obviously a very worthwhile save.
1460          *
1461          * Thus we just return success now...
1462          */
1463         ntfs_debug("Done.");
1464         return 0;
1465 }
1466
1467 /*
1468  * NOTES: There is a disparity between the apparent need to extend the
1469  * attribute in prepare write but to update i_size only in commit write.
1470  * Need to make sure i_sem protection is sufficient. And if not will need to
1471  * handle this in some way or another.
1472  */
1473
1474 /**
1475  * ntfs_commit_nonresident_write -
1476  *
1477  */
1478 static int ntfs_commit_nonresident_write(struct page *page,
1479                 unsigned from, unsigned to)
1480 {
1481         s64 pos = ((s64)page->index << PAGE_CACHE_SHIFT) + to;
1482         struct inode *vi;
1483         struct buffer_head *bh, *head;
1484         unsigned int block_start, block_end, blocksize;
1485         BOOL partial;
1486
1487         vi = page->mapping->host;
1488
1489         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
1490                         "0x%lx, from = %u, to = %u.", vi->i_ino,
1491                         NTFS_I(vi)->type, page->index, from, to);
1492
1493         blocksize = 1 << vi->i_blkbits;
1494
1495         // FIXME: We need a whole slew of special cases in here for MST
1496         // protected attributes for example. For compressed files, too...
1497         // For now, we know ntfs_prepare_write() would have failed so we can't
1498         // get here in any of the cases which we have to special case, so we
1499         // are just a ripped off unrolled generic_commit_write() at present.
1500
1501         bh = head = page_buffers(page);
1502         block_start = 0;
1503         partial = FALSE;
1504         do {
1505                 block_end = block_start + blocksize;
1506                 if (block_end <= from || block_start >= to) {
1507                         if (!buffer_uptodate(bh))
1508                                 partial = TRUE;
1509                 } else {
1510                         set_buffer_uptodate(bh);
1511                         mark_buffer_dirty(bh);
1512                 }
1513         } while (block_start = block_end, (bh = bh->b_this_page) != head);
1514
1515         /*
1516          * If this is a partial write which happened to make all buffers
1517          * uptodate then we can optimize away a bogus ->readpage() for the next
1518          * read(). Here we 'discover' whether the page went uptodate as a
1519          * result of this (potentially partial) write.
1520          */
1521         if (!partial)
1522                 SetPageUptodate(page);
1523
1524         /*
1525          * Not convinced about this at all. See disparity comment above. For
1526          * now we know ntfs_prepare_write() would have failed in the write
1527          * exceeds i_size case, so this will never trigger which is fine.
1528          */
1529         if (pos > vi->i_size) {
1530                 ntfs_error(vi->i_sb, "Writing beyond the existing file size is "
1531                                 "not supported yet. Sorry.");
1532                 return -EOPNOTSUPP;
1533                 // vi->i_size = pos;
1534                 // mark_inode_dirty(vi);
1535         }
1536         ntfs_debug("Done.");
1537         return 0;
1538 }
1539
1540 /**
1541  * ntfs_commit_write - commit the received data
1542  *
1543  * This is called from generic_file_write() with i_sem held on the inode
1544  * (@page->mapping->host). The @page is locked and kmap()ped so page_address()
1545  * can simply be used. The source data has already been copied into the @page.
1546  *
1547  * Need to mark modified blocks dirty so they get written out later when
1548  * ntfs_writepage() is invoked by the VM.
1549  *
1550  * Return 0 on success or -errno on error.
1551  *
1552  * Should be using generic_commit_write(). This marks buffers uptodate and
1553  * dirty, sets the page uptodate if all buffers in the page are uptodate, and
1554  * updates i_size if the end of io is beyond i_size. In that case, it also
1555  * marks the inode dirty. - We could still use this (obviously except for
1556  * NInoMstProtected() attributes, where we will need to duplicate the core code
1557  * because we need our own async_io completion handler) but we could just do
1558  * the i_size update in prepare write, when we resize the attribute. Then
1559  * we would avoid the i_size update and mark_inode_dirty() happening here.
1560  *
1561  * Can't use generic_commit_write() due to ntfs specialities but can look at
1562  * it for implementation guidance.
1563  *
1564  * If things have gone as outlined in ntfs_prepare_write(), then we do not
1565  * need to do any page content modifications here at all, except in the write
1566  * to resident attribute case, where we need to do the uptodate bringing here
1567  * which we combine with the copying into the mft record which means we only
1568  * need to map the mft record and find the attribute record in it only once.
1569  */
1570 static int ntfs_commit_write(struct file *file, struct page *page,
1571                 unsigned from, unsigned to)
1572 {
1573         s64 attr_pos;
1574         struct inode *vi;
1575         ntfs_inode *ni, *base_ni;
1576         char *kaddr, *kattr;
1577         attr_search_context *ctx;
1578         MFT_RECORD *m;
1579         u32 attr_len, bytes;
1580         int err;
1581
1582         vi = page->mapping->host;
1583         ni = NTFS_I(vi);
1584
1585         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
1586                         "0x%lx, from = %u, to = %u.", vi->i_ino, ni->type,
1587                         page->index, from, to);
1588
1589         if (NInoNonResident(ni)) {
1590                 /*
1591                  * Only unnamed $DATA attributes can be compressed, encrypted,
1592                  * and/or sparse.
1593                  */
1594                 if (ni->type == AT_DATA && !ni->name_len) {
1595                         /* If file is encrypted, deny access, just like NT4. */
1596                         if (NInoEncrypted(ni)) {
1597                                 // Should never get here!
1598                                 ntfs_debug("Denying write access to encrypted "
1599                                                 "file.");
1600                                 return -EACCES;
1601                         }
1602                         /* Compressed data streams are handled in compress.c. */
1603                         if (NInoCompressed(ni)) {
1604                                 // TODO: Implement and replace this check with
1605                                 // return ntfs_write_compressed_block(page);
1606                                 // Should never get here!
1607                                 ntfs_error(vi->i_sb, "Writing to compressed "
1608                                                 "files is not supported yet. "
1609                                                 "Sorry.");
1610                                 return -EOPNOTSUPP;
1611                         }
1612                         // TODO: Implement and remove this check.
1613                         if (NInoSparse(ni)) {
1614                                 // Should never get here!
1615                                 ntfs_error(vi->i_sb, "Writing to sparse files "
1616                                                 "is not supported yet. Sorry.");
1617                                 return -EOPNOTSUPP;
1618                         }
1619                 }
1620
1621                 // TODO: Implement and remove this check.
1622                 if (NInoMstProtected(ni)) {
1623                         // Should never get here!
1624                         ntfs_error(vi->i_sb, "Writing to MST protected "
1625                                         "attributes is not supported yet. "
1626                                         "Sorry.");
1627                         return -EOPNOTSUPP;
1628                 }
1629
1630                 /* Normal data stream. */
1631                 return ntfs_commit_nonresident_write(page, from, to);
1632         }
1633
1634         /*
1635          * Attribute is resident, implying it is not compressed, encrypted, or
1636          * mst protected.
1637          */
1638
1639         /* Do we need to resize the attribute? */
1640         if (((s64)page->index << PAGE_CACHE_SHIFT) + to > vi->i_size) {
1641                 // TODO: Implement resize...
1642                 // pos = ((s64)page->index << PAGE_CACHE_SHIFT) + to;
1643                 // vi->i_size = pos;
1644                 // mark_inode_dirty(vi);
1645                 // Should never get here!
1646                 ntfs_error(vi->i_sb, "Writing beyond the existing file size is "
1647                                 "not supported yet. Sorry.");
1648                 return -EOPNOTSUPP;
1649         }
1650
1651         if (!NInoAttr(ni))
1652                 base_ni = ni;
1653         else
1654                 base_ni = ni->ext.base_ntfs_ino;
1655
1656         /* Map, pin, and lock the mft record. */
1657         m = map_mft_record(base_ni);
1658         if (unlikely(IS_ERR(m))) {
1659                 err = PTR_ERR(m);
1660                 m = NULL;
1661                 ctx = NULL;
1662                 goto err_out;
1663         }
1664         ctx = get_attr_search_ctx(base_ni, m);
1665         if (unlikely(!ctx)) {
1666                 err = -ENOMEM;
1667                 goto err_out;
1668         }
1669         if (unlikely(!lookup_attr(ni->type, ni->name, ni->name_len,
1670                         IGNORE_CASE, 0, NULL, 0, ctx))) {
1671                 err = -ENOENT;
1672                 goto err_out;
1673         }
1674
1675         /* Starting position of the page within the attribute value. */
1676         attr_pos = page->index << PAGE_CACHE_SHIFT;
1677
1678         /* The total length of the attribute value. */
1679         attr_len = le32_to_cpu(ctx->attr->data.resident.value_length);
1680
1681         if (unlikely(vi->i_size != attr_len)) {
1682                 ntfs_error(vi->i_sb, "BUG()! i_size (0x%llx) doesn't match "
1683                                 "attr_len (0x%x). Aborting write.", vi->i_size,
1684                                 attr_len);
1685                 err = -EIO;
1686                 goto err_out;
1687         }
1688         if (unlikely(attr_pos >= attr_len)) {
1689                 ntfs_error(vi->i_sb, "BUG()! attr_pos (0x%llx) > attr_len "
1690                                 "(0x%x). Aborting write.",
1691                                 (unsigned long long)attr_pos, attr_len);
1692                 err = -EIO;
1693                 goto err_out;
1694         }
1695
1696         bytes = attr_len - attr_pos;
1697         if (unlikely(bytes > PAGE_CACHE_SIZE))
1698                 bytes = PAGE_CACHE_SIZE;
1699
1700         /*
1701          * Calculate the address of the attribute value corresponding to the
1702          * beginning of the current data @page.
1703          */
1704         kattr = (u8*)ctx->attr + le16_to_cpu(
1705                         ctx->attr->data.resident.value_offset) + attr_pos;
1706
1707         kaddr = kmap_atomic(page, KM_USER0);
1708
1709         /* Copy the received data from the page to the mft record. */
1710         memcpy(kattr + from, kaddr + from, to - from);
1711         flush_dcache_mft_record_page(ctx->ntfs_ino);
1712
1713         if (!PageUptodate(page)) {
1714                 /*
1715                  * Bring the out of bounds area(s) uptodate by copying data
1716                  * from the mft record to the page.
1717                  */
1718                 if (from > 0)
1719                         memcpy(kaddr, kattr, from);
1720                 if (to < bytes)
1721                         memcpy(kaddr + to, kattr + to, bytes - to);
1722
1723                 /* Zero the region outside the end of the attribute value. */
1724                 if (likely(bytes < PAGE_CACHE_SIZE))
1725                         memset(kaddr + bytes, 0, PAGE_CACHE_SIZE - bytes);
1726
1727                 /*
1728                  * The probability of not having done any of the above is
1729                  * extremely small, so we just flush unconditionally.
1730                  */
1731                 flush_dcache_page(page);
1732                 SetPageUptodate(page);
1733         }
1734         kunmap_atomic(kaddr, KM_USER0);
1735
1736         /* Mark the mft record dirty, so it gets written back. */
1737         mark_mft_record_dirty(ctx->ntfs_ino);
1738
1739         put_attr_search_ctx(ctx);
1740         unmap_mft_record(base_ni);
1741         ntfs_debug("Done.");
1742         return 0;
1743 err_out:
1744         if (err == -ENOMEM) {
1745                 ntfs_warning(vi->i_sb, "Error allocating memory required to "
1746                                 "commit the write.");
1747                 if (PageUptodate(page)) {
1748                         ntfs_warning(vi->i_sb, "Page is uptodate, setting "
1749                                         "dirty so the write will be retried "
1750                                         "later on by the VM.");
1751                         /*
1752                          * Put the page on mapping->dirty_pages, but leave its
1753                          * buffer's dirty state as-is.
1754                          */
1755                         __set_page_dirty_nobuffers(page);
1756                         err = 0;
1757                 } else
1758                         ntfs_error(vi->i_sb, "Page is not uptodate. Written "
1759                                         "data has been lost. )-:");
1760         } else {
1761                 ntfs_error(vi->i_sb, "Resident attribute write failed with "
1762                                 "error %i. Setting page error flag.", -err);
1763                 SetPageError(page);
1764         }
1765         if (ctx)
1766                 put_attr_search_ctx(ctx);
1767         if (m)
1768                 unmap_mft_record(base_ni);
1769         return err;
1770 }
1771
1772 #endif  /* NTFS_RW */
1773
1774 /**
1775  * ntfs_aops - general address space operations for inodes and attributes
1776  */
1777 struct address_space_operations ntfs_aops = {
1778         .readpage       = ntfs_readpage,        /* Fill page with data. */
1779         .sync_page      = block_sync_page,      /* Currently, just unplugs the
1780                                                    disk request queue. */
1781 #ifdef NTFS_RW
1782         .writepage      = ntfs_writepage,       /* Write dirty page to disk. */
1783         .prepare_write  = ntfs_prepare_write,   /* Prepare page and buffers
1784                                                    ready to receive data. */
1785         .commit_write   = ntfs_commit_write,    /* Commit received data. */
1786 #endif
1787 };
1788
1789 /**
1790  * ntfs_mst_aops - general address space operations for mst protecteed inodes
1791  *                 and attributes
1792  */
1793 struct address_space_operations ntfs_mst_aops = {
1794         .readpage       = ntfs_readpage,        /* Fill page with data. */
1795         .sync_page      = block_sync_page,      /* Currently, just unplugs the
1796                                                    disk request queue. */
1797 };