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