fs/ntfs/file.c

   1 /*
   2  * file.c - NTFS kernel file operations.  Part of the Linux-NTFS project.
   3  *
   4  * Copyright (c) 2001-2015 Anton Altaparmakov and Tuxera Inc.
   5  *
   6  * This program/include file is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License as published
   8  * by the Free Software Foundation; either version 2 of the License, or
   9  * (at your option) any later version.
  10  *
  11  * This program/include file is distributed in the hope that it will be
  12  * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  13  * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  * GNU General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program (in the main directory of the Linux-NTFS
  18  * distribution in the file COPYING); if not, write to the Free Software
  19  * Foundation,Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  20  */
  21
  22 #include <linux/backing-dev.h>
  23 #include <linux/buffer_head.h>
  24 #include <linux/gfp.h>
  25 #include <linux/pagemap.h>
  26 #include <linux/pagevec.h>
  27 #include <linux/sched.h>
  28 #include <linux/swap.h>
  29 #include <linux/uio.h>
  30 #include <linux/writeback.h>
  31
  32 #include <asm/page.h>
  33 #include <asm/uaccess.h>
  34
  35 #include "attrib.h"
  36 #include "bitmap.h"
  37 #include "inode.h"
  38 #include "debug.h"
  39 #include "lcnalloc.h"
  40 #include "malloc.h"
  41 #include "mft.h"
  42 #include "ntfs.h"
  43
  44 /**
  45  * ntfs_file_open - called when an inode is about to be opened
  46  * @vi:         inode to be opened
  47  * @filp:       file structure describing the inode
  48  *
  49  * Limit file size to the page cache limit on architectures where unsigned long
  50  * is 32-bits. This is the most we can do for now without overflowing the page
  51  * cache page index. Doing it this way means we don't run into problems because
  52  * of existing too large files. It would be better to allow the user to read
  53  * the beginning of the file but I doubt very much anyone is going to hit this
  54  * check on a 32-bit architecture, so there is no point in adding the extra
  55  * complexity required to support this.
  56  *
  57  * On 64-bit architectures, the check is hopefully optimized away by the
  58  * compiler.
  59  *
  60  * After the check passes, just call generic_file_open() to do its work.
  61  */
  62 static int ntfs_file_open(struct inode *vi, struct file *filp)
  63 {
  64         if (sizeof(unsigned long) < 8) {
  65                 if (i_size_read(vi) > MAX_LFS_FILESIZE)
  66                         return -EOVERFLOW;
  67         }
  68         return generic_file_open(vi, filp);
  69 }
  70
  71 #ifdef NTFS_RW
  72
  73 /**
  74  * ntfs_attr_extend_initialized - extend the initialized size of an attribute
  75  * @ni:                 ntfs inode of the attribute to extend
  76  * @new_init_size:      requested new initialized size in bytes
  77  *
  78  * Extend the initialized size of an attribute described by the ntfs inode @ni
  79  * to @new_init_size bytes.  This involves zeroing any non-sparse space between
  80  * the old initialized size and @new_init_size both in the page cache and on
  81  * disk (if relevant complete pages are already uptodate in the page cache then
  82  * these are simply marked dirty).
  83  *
  84  * As a side-effect, the file size (vfs inode->i_size) may be incremented as,
  85  * in the resident attribute case, it is tied to the initialized size and, in
  86  * the non-resident attribute case, it may not fall below the initialized size.
  87  *
  88  * Note that if the attribute is resident, we do not need to touch the page
  89  * cache at all.  This is because if the page cache page is not uptodate we
  90  * bring it uptodate later, when doing the write to the mft record since we
  91  * then already have the page mapped.  And if the page is uptodate, the
  92  * non-initialized region will already have been zeroed when the page was
  93  * brought uptodate and the region may in fact already have been overwritten
  94  * with new data via mmap() based writes, so we cannot just zero it.  And since
  95  * POSIX specifies that the behaviour of resizing a file whilst it is mmap()ped
  96  * is unspecified, we choose not to do zeroing and thus we do not need to touch
  97  * the page at all.  For a more detailed explanation see ntfs_truncate() in
  98  * fs/ntfs/inode.c.
  99  *
 100  * Return 0 on success and -errno on error.  In the case that an error is
 101  * encountered it is possible that the initialized size will already have been
 102  * incremented some way towards @new_init_size but it is guaranteed that if
 103  * this is the case, the necessary zeroing will also have happened and that all
 104  * metadata is self-consistent.
 105  *
 106  * Locking: i_mutex on the vfs inode corrseponsind to the ntfs inode @ni must be
 107  *          held by the caller.
 108  */
 109 static int ntfs_attr_extend_initialized(ntfs_inode *ni, const s64 new_init_size)
 110 {
 111         s64 old_init_size;
 112         loff_t old_i_size;
 113         pgoff_t index, end_index;
 114         unsigned long flags;
 115         struct inode *vi = VFS_I(ni);
 116         ntfs_inode *base_ni;
 117         MFT_RECORD *m = NULL;
 118         ATTR_RECORD *a;
 119         ntfs_attr_search_ctx *ctx = NULL;
 120         struct address_space *mapping;
 121         struct page *page = NULL;
 122         u8 *kattr;
 123         int err;
 124         u32 attr_len;
 125
 126         read_lock_irqsave(&ni->size_lock, flags);
 127         old_init_size = ni->initialized_size;
 128         old_i_size = i_size_read(vi);
 129         BUG_ON(new_init_size > ni->allocated_size);
 130         read_unlock_irqrestore(&ni->size_lock, flags);
 131         ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, "
 132                         "old_initialized_size 0x%llx, "
 133                         "new_initialized_size 0x%llx, i_size 0x%llx.",
 134                         vi->i_ino, (unsigned)le32_to_cpu(ni->type),
 135                         (unsigned long long)old_init_size,
 136                         (unsigned long long)new_init_size, old_i_size);
 137         if (!NInoAttr(ni))
 138                 base_ni = ni;
 139         else
 140                 base_ni = ni->ext.base_ntfs_ino;
 141         /* Use goto to reduce indentation and we need the label below anyway. */
 142         if (NInoNonResident(ni))
 143                 goto do_non_resident_extend;
 144         BUG_ON(old_init_size != old_i_size);
 145         m = map_mft_record(base_ni);
 146         if (IS_ERR(m)) {
 147                 err = PTR_ERR(m);
 148                 m = NULL;
 149                 goto err_out;
 150         }
 151         ctx = ntfs_attr_get_search_ctx(base_ni, m);
 152         if (unlikely(!ctx)) {
 153                 err = -ENOMEM;
 154                 goto err_out;
 155         }
 156         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
 157                         CASE_SENSITIVE, 0, NULL, 0, ctx);
 158         if (unlikely(err)) {
 159                 if (err == -ENOENT)
 160                         err = -EIO;
 161                 goto err_out;
 162         }
 163         m = ctx->mrec;
 164         a = ctx->attr;
 165         BUG_ON(a->non_resident);
 166         /* The total length of the attribute value. */
 167         attr_len = le32_to_cpu(a->data.resident.value_length);
 168         BUG_ON(old_i_size != (loff_t)attr_len);
 169         /*
 170          * Do the zeroing in the mft record and update the attribute size in
 171          * the mft record.
 172          */
 173         kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
 174         memset(kattr + attr_len, 0, new_init_size - attr_len);
 175         a->data.resident.value_length = cpu_to_le32((u32)new_init_size);
 176         /* Finally, update the sizes in the vfs and ntfs inodes. */
 177         write_lock_irqsave(&ni->size_lock, flags);
 178         i_size_write(vi, new_init_size);
 179         ni->initialized_size = new_init_size;
 180         write_unlock_irqrestore(&ni->size_lock, flags);
 181         goto done;
 182 do_non_resident_extend:
 183         /*
 184          * If the new initialized size @new_init_size exceeds the current file
 185          * size (vfs inode->i_size), we need to extend the file size to the
 186          * new initialized size.
 187          */
 188         if (new_init_size > old_i_size) {
 189                 m = map_mft_record(base_ni);
 190                 if (IS_ERR(m)) {
 191                         err = PTR_ERR(m);
 192                         m = NULL;
 193                         goto err_out;
 194                 }
 195                 ctx = ntfs_attr_get_search_ctx(base_ni, m);
 196                 if (unlikely(!ctx)) {
 197                         err = -ENOMEM;
 198                         goto err_out;
 199                 }
 200                 err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
 201                                 CASE_SENSITIVE, 0, NULL, 0, ctx);
 202                 if (unlikely(err)) {
 203                         if (err == -ENOENT)
 204                                 err = -EIO;
 205                         goto err_out;
 206                 }
 207                 m = ctx->mrec;
 208                 a = ctx->attr;
 209                 BUG_ON(!a->non_resident);
 210                 BUG_ON(old_i_size != (loff_t)
 211                                 sle64_to_cpu(a->data.non_resident.data_size));
 212                 a->data.non_resident.data_size = cpu_to_sle64(new_init_size);
 213                 flush_dcache_mft_record_page(ctx->ntfs_ino);
 214                 mark_mft_record_dirty(ctx->ntfs_ino);
 215                 /* Update the file size in the vfs inode. */
 216                 i_size_write(vi, new_init_size);
 217                 ntfs_attr_put_search_ctx(ctx);
 218                 ctx = NULL;
 219                 unmap_mft_record(base_ni);
 220                 m = NULL;
 221         }
 222         mapping = vi->i_mapping;
 223         index = old_init_size >> PAGE_CACHE_SHIFT;
 224         end_index = (new_init_size + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 225         do {
 226                 /*
 227                  * Read the page.  If the page is not present, this will zero
 228                  * the uninitialized regions for us.
 229                  */
 230                 page = read_mapping_page(mapping, index, NULL);
 231                 if (IS_ERR(page)) {
 232                         err = PTR_ERR(page);
 233                         goto init_err_out;
 234                 }
 235                 if (unlikely(PageError(page))) {
 236                         page_cache_release(page);
 237                         err = -EIO;
 238                         goto init_err_out;
 239                 }
 240                 /*
 241                  * Update the initialized size in the ntfs inode.  This is
 242                  * enough to make ntfs_writepage() work.
 243                  */
 244                 write_lock_irqsave(&ni->size_lock, flags);
 245                 ni->initialized_size = (s64)(index + 1) << PAGE_CACHE_SHIFT;
 246                 if (ni->initialized_size > new_init_size)
 247                         ni->initialized_size = new_init_size;
 248                 write_unlock_irqrestore(&ni->size_lock, flags);
 249                 /* Set the page dirty so it gets written out. */
 250                 set_page_dirty(page);
 251                 page_cache_release(page);
 252                 /*
 253                  * Play nice with the vm and the rest of the system.  This is
 254                  * very much needed as we can potentially be modifying the
 255                  * initialised size from a very small value to a really huge
 256                  * value, e.g.
 257                  *      f = open(somefile, O_TRUNC);
 258                  *      truncate(f, 10GiB);
 259                  *      seek(f, 10GiB);
 260                  *      write(f, 1);
 261                  * And this would mean we would be marking dirty hundreds of
 262                  * thousands of pages or as in the above example more than
 263                  * two and a half million pages!
 264                  *
 265                  * TODO: For sparse pages could optimize this workload by using
 266                  * the FsMisc / MiscFs page bit as a "PageIsSparse" bit.  This
 267                  * would be set in readpage for sparse pages and here we would
 268                  * not need to mark dirty any pages which have this bit set.
 269                  * The only caveat is that we have to clear the bit everywhere
 270                  * where we allocate any clusters that lie in the page or that
 271                  * contain the page.
 272                  *
 273                  * TODO: An even greater optimization would be for us to only
 274                  * call readpage() on pages which are not in sparse regions as
 275                  * determined from the runlist.  This would greatly reduce the
 276                  * number of pages we read and make dirty in the case of sparse
 277                  * files.
 278                  */
 279                 balance_dirty_pages_ratelimited(mapping);
 280                 cond_resched();
 281         } while (++index < end_index);
 282         read_lock_irqsave(&ni->size_lock, flags);
 283         BUG_ON(ni->initialized_size != new_init_size);
 284         read_unlock_irqrestore(&ni->size_lock, flags);
 285         /* Now bring in sync the initialized_size in the mft record. */
 286         m = map_mft_record(base_ni);
 287         if (IS_ERR(m)) {
 288                 err = PTR_ERR(m);
 289                 m = NULL;
 290                 goto init_err_out;
 291         }
 292         ctx = ntfs_attr_get_search_ctx(base_ni, m);
 293         if (unlikely(!ctx)) {
 294                 err = -ENOMEM;
 295                 goto init_err_out;
 296         }
 297         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
 298                         CASE_SENSITIVE, 0, NULL, 0, ctx);
 299         if (unlikely(err)) {
 300                 if (err == -ENOENT)
 301                         err = -EIO;
 302                 goto init_err_out;
 303         }
 304         m = ctx->mrec;
 305         a = ctx->attr;
 306         BUG_ON(!a->non_resident);
 307         a->data.non_resident.initialized_size = cpu_to_sle64(new_init_size);
 308 done:
 309         flush_dcache_mft_record_page(ctx->ntfs_ino);
 310         mark_mft_record_dirty(ctx->ntfs_ino);
 311         if (ctx)
 312                 ntfs_attr_put_search_ctx(ctx);
 313         if (m)
 314                 unmap_mft_record(base_ni);
 315         ntfs_debug("Done, initialized_size 0x%llx, i_size 0x%llx.",
 316                         (unsigned long long)new_init_size, i_size_read(vi));
 317         return 0;
 318 init_err_out:
 319         write_lock_irqsave(&ni->size_lock, flags);
 320         ni->initialized_size = old_init_size;
 321         write_unlock_irqrestore(&ni->size_lock, flags);
 322 err_out:
 323         if (ctx)
 324                 ntfs_attr_put_search_ctx(ctx);
 325         if (m)
 326                 unmap_mft_record(base_ni);
 327         ntfs_debug("Failed.  Returning error code %i.", err);
 328         return err;
 329 }
 330
 331 static ssize_t ntfs_prepare_file_for_write(struct file *file, loff_t *ppos,
 332                 size_t *count)
 333 {
 334         loff_t pos;
 335         s64 end, ll;
 336         ssize_t err;
 337         unsigned long flags;
 338         struct inode *vi = file_inode(file);
 339         ntfs_inode *base_ni, *ni = NTFS_I(vi);
 340         ntfs_volume *vol = ni->vol;
 341
 342         ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, pos "
 343                         "0x%llx, count 0x%lx.", vi->i_ino,
 344                         (unsigned)le32_to_cpu(ni->type),
 345                         (unsigned long long)*ppos, (unsigned long)*count);
 346         /* We can write back this queue in page reclaim. */
 347         current->backing_dev_info = inode_to_bdi(vi);
 348         err = generic_write_checks(file, ppos, count, S_ISBLK(vi->i_mode));
 349         if (unlikely(err))
 350                 goto out;
 351         /*
 352          * All checks have passed.  Before we start doing any writing we want
 353          * to abort any totally illegal writes.
 354          */
 355         BUG_ON(NInoMstProtected(ni));
 356         BUG_ON(ni->type != AT_DATA);
 357         /* If file is encrypted, deny access, just like NT4. */
 358         if (NInoEncrypted(ni)) {
 359                 /* Only $DATA attributes can be encrypted. */
 360                 /*
 361                  * Reminder for later: Encrypted files are _always_
 362                  * non-resident so that the content can always be encrypted.
 363                  */
 364                 ntfs_debug("Denying write access to encrypted file.");
 365                 err = -EACCES;
 366                 goto out;
 367         }
 368         if (NInoCompressed(ni)) {
 369                 /* Only unnamed $DATA attribute can be compressed. */
 370                 BUG_ON(ni->name_len);
 371                 /*
 372                  * Reminder for later: If resident, the data is not actually
 373                  * compressed.  Only on the switch to non-resident does
 374                  * compression kick in.  This is in contrast to encrypted files
 375                  * (see above).
 376                  */
 377                 ntfs_error(vi->i_sb, "Writing to compressed files is not "
 378                                 "implemented yet.  Sorry.");
 379                 err = -EOPNOTSUPP;
 380                 goto out;
 381         }
 382         if (*count == 0)
 383                 goto out;
 384         base_ni = ni;
 385         if (NInoAttr(ni))
 386                 base_ni = ni->ext.base_ntfs_ino;
 387         err = file_remove_suid(file);
 388         if (unlikely(err))
 389                 goto out;
 390         /*
 391          * Our ->update_time method always succeeds thus file_update_time()
 392          * cannot fail either so there is no need to check the return code.
 393          */
 394         file_update_time(file);
 395         pos = *ppos;
 396         /* The first byte after the last cluster being written to. */
 397         end = (pos + *count + vol->cluster_size_mask) &
 398                         ~(u64)vol->cluster_size_mask;
 399         /*
 400          * If the write goes beyond the allocated size, extend the allocation
 401          * to cover the whole of the write, rounded up to the nearest cluster.
 402          */
 403         read_lock_irqsave(&ni->size_lock, flags);
 404         ll = ni->allocated_size;
 405         read_unlock_irqrestore(&ni->size_lock, flags);
 406         if (end > ll) {
 407                 /*
 408                  * Extend the allocation without changing the data size.
 409                  *
 410                  * Note we ensure the allocation is big enough to at least
 411                  * write some data but we do not require the allocation to be
 412                  * complete, i.e. it may be partial.
 413                  */
 414                 ll = ntfs_attr_extend_allocation(ni, end, -1, pos);
 415                 if (likely(ll >= 0)) {
 416                         BUG_ON(pos >= ll);
 417                         /* If the extension was partial truncate the write. */
 418                         if (end > ll) {
 419                                 ntfs_debug("Truncating write to inode 0x%lx, "
 420                                                 "attribute type 0x%x, because "
 421                                                 "the allocation was only "
 422                                                 "partially extended.",
 423                                                 vi->i_ino, (unsigned)
 424                                                 le32_to_cpu(ni->type));
 425                                 *count = ll - pos;
 426                         }
 427                 } else {
 428                         err = ll;
 429                         read_lock_irqsave(&ni->size_lock, flags);
 430                         ll = ni->allocated_size;
 431                         read_unlock_irqrestore(&ni->size_lock, flags);
 432                         /* Perform a partial write if possible or fail. */
 433                         if (pos < ll) {
 434                                 ntfs_debug("Truncating write to inode 0x%lx "
 435                                                 "attribute type 0x%x, because "
 436                                                 "extending the allocation "
 437                                                 "failed (error %d).",
 438                                                 vi->i_ino, (unsigned)
 439                                                 le32_to_cpu(ni->type),
 440                                                 (int)-err);
 441                                 *count = ll - pos;
 442                         } else {
 443                                 if (err != -ENOSPC)
 444                                         ntfs_error(vi->i_sb, "Cannot perform "
 445                                                         "write to inode "
 446                                                         "0x%lx, attribute "
 447                                                         "type 0x%x, because "
 448                                                         "extending the "
 449                                                         "allocation failed "
 450                                                         "(error %ld).",
 451                                                         vi->i_ino, (unsigned)
 452                                                         le32_to_cpu(ni->type),
 453                                                         (long)-err);
 454                                 else
 455                                         ntfs_debug("Cannot perform write to "
 456                                                         "inode 0x%lx, "
 457                                                         "attribute type 0x%x, "
 458                                                         "because there is not "
 459                                                         "space left.",
 460                                                         vi->i_ino, (unsigned)
 461                                                         le32_to_cpu(ni->type));
 462                                 goto out;
 463                         }
 464                 }
 465         }
 466         /*
 467          * If the write starts beyond the initialized size, extend it up to the
 468          * beginning of the write and initialize all non-sparse space between
 469          * the old initialized size and the new one.  This automatically also
 470          * increments the vfs inode->i_size to keep it above or equal to the
 471          * initialized_size.
 472          */
 473         read_lock_irqsave(&ni->size_lock, flags);
 474         ll = ni->initialized_size;
 475         read_unlock_irqrestore(&ni->size_lock, flags);
 476         if (pos > ll) {
 477                 /*
 478                  * Wait for ongoing direct i/o to complete before proceeding.
 479                  * New direct i/o cannot start as we hold i_mutex.
 480                  */
 481                 inode_dio_wait(vi);
 482                 err = ntfs_attr_extend_initialized(ni, pos);
 483                 if (unlikely(err < 0))
 484                         ntfs_error(vi->i_sb, "Cannot perform write to inode "
 485                                         "0x%lx, attribute type 0x%x, because "
 486                                         "extending the initialized size "
 487                                         "failed (error %d).", vi->i_ino,
 488                                         (unsigned)le32_to_cpu(ni->type),
 489                                         (int)-err);
 490         }
 491 out:
 492         return err;
 493 }
 494
 495 /**
 496  * __ntfs_grab_cache_pages - obtain a number of locked pages
 497  * @mapping:    address space mapping from which to obtain page cache pages
 498  * @index:      starting index in @mapping at which to begin obtaining pages
 499  * @nr_pages:   number of page cache pages to obtain
 500  * @pages:      array of pages in which to return the obtained page cache pages
 501  * @cached_page: allocated but as yet unused page
 502  *
 503  * Obtain @nr_pages locked page cache pages from the mapping @mapping and
 504  * starting at index @index.
 505  *
 506  * If a page is newly created, add it to lru list
 507  *
 508  * Note, the page locks are obtained in ascending page index order.
 509  */
 510 static inline int __ntfs_grab_cache_pages(struct address_space *mapping,
 511                 pgoff_t index, const unsigned nr_pages, struct page **pages,
 512                 struct page **cached_page)
 513 {
 514         int err, nr;
 515
 516         BUG_ON(!nr_pages);
 517         err = nr = 0;
 518         do {
 519                 pages[nr] = find_get_page_flags(mapping, index, FGP_LOCK |
 520                                 FGP_ACCESSED);
 521                 if (!pages[nr]) {
 522                         if (!*cached_page) {
 523                                 *cached_page = page_cache_alloc(mapping);
 524                                 if (unlikely(!*cached_page)) {
 525                                         err = -ENOMEM;
 526                                         goto err_out;
 527                                 }
 528                         }
 529                         err = add_to_page_cache_lru(*cached_page, mapping,
 530                                         index, GFP_KERNEL);
 531                         if (unlikely(err)) {
 532                                 if (err == -EEXIST)
 533                                         continue;
 534                                 goto err_out;
 535                         }
 536                         pages[nr] = *cached_page;
 537                         *cached_page = NULL;
 538                 }
 539                 index++;
 540                 nr++;
 541         } while (nr < nr_pages);
 542 out:
 543         return err;
 544 err_out:
 545         while (nr > 0) {
 546                 unlock_page(pages[--nr]);
 547                 page_cache_release(pages[nr]);
 548         }
 549         goto out;
 550 }
 551
 552 static inline int ntfs_submit_bh_for_read(struct buffer_head *bh)
 553 {
 554         lock_buffer(bh);
 555         get_bh(bh);
 556         bh->b_end_io = end_buffer_read_sync;
 557         return submit_bh(READ, bh);
 558 }
 559
 560 /**
 561  * ntfs_prepare_pages_for_non_resident_write - prepare pages for receiving data
 562  * @pages:      array of destination pages
 563  * @nr_pages:   number of pages in @pages
 564  * @pos:        byte position in file at which the write begins
 565  * @bytes:      number of bytes to be written
 566  *
 567  * This is called for non-resident attributes from ntfs_file_buffered_write()
 568  * with i_mutex held on the inode (@pages[0]->mapping->host).  There are
 569  * @nr_pages pages in @pages which are locked but not kmap()ped.  The source
 570  * data has not yet been copied into the @pages.
 571  *
 572  * Need to fill any holes with actual clusters, allocate buffers if necessary,
 573  * ensure all the buffers are mapped, and bring uptodate any buffers that are
 574  * only partially being written to.
 575  *
 576  * If @nr_pages is greater than one, we are guaranteed that the cluster size is
 577  * greater than PAGE_CACHE_SIZE, that all pages in @pages are entirely inside
 578  * the same cluster and that they are the entirety of that cluster, and that
 579  * the cluster is sparse, i.e. we need to allocate a cluster to fill the hole.
 580  *
 581  * i_size is not to be modified yet.
 582  *
 583  * Return 0 on success or -errno on error.
 584  */
 585 static int ntfs_prepare_pages_for_non_resident_write(struct page **pages,
 586                 unsigned nr_pages, s64 pos, size_t bytes)
 587 {
 588         VCN vcn, highest_vcn = 0, cpos, cend, bh_cpos, bh_cend;
 589         LCN lcn;
 590         s64 bh_pos, vcn_len, end, initialized_size;
 591         sector_t lcn_block;
 592         struct page *page;
 593         struct inode *vi;
 594         ntfs_inode *ni, *base_ni = NULL;
 595         ntfs_volume *vol;
 596         runlist_element *rl, *rl2;
 597         struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
 598         ntfs_attr_search_ctx *ctx = NULL;
 599         MFT_RECORD *m = NULL;
 600         ATTR_RECORD *a = NULL;
 601         unsigned long flags;
 602         u32 attr_rec_len = 0;
 603         unsigned blocksize, u;
 604         int err, mp_size;
 605         bool rl_write_locked, was_hole, is_retry;
 606         unsigned char blocksize_bits;
 607         struct {
 608                 u8 runlist_merged:1;
 609                 u8 mft_attr_mapped:1;
 610                 u8 mp_rebuilt:1;
 611                 u8 attr_switched:1;
 612         } status = { 0, 0, 0, 0 };
 613
 614         BUG_ON(!nr_pages);
 615         BUG_ON(!pages);
 616         BUG_ON(!*pages);
 617         vi = pages[0]->mapping->host;
 618         ni = NTFS_I(vi);
 619         vol = ni->vol;
 620         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, start page "
 621                         "index 0x%lx, nr_pages 0x%x, pos 0x%llx, bytes 0x%zx.",
 622                         vi->i_ino, ni->type, pages[0]->index, nr_pages,
 623                         (long long)pos, bytes);
 624         blocksize = vol->sb->s_blocksize;
 625         blocksize_bits = vol->sb->s_blocksize_bits;
 626         u = 0;
 627         do {
 628                 page = pages[u];
 629                 BUG_ON(!page);
 630                 /*
 631                  * create_empty_buffers() will create uptodate/dirty buffers if
 632                  * the page is uptodate/dirty.
 633                  */
 634                 if (!page_has_buffers(page)) {
 635                         create_empty_buffers(page, blocksize, 0);
 636                         if (unlikely(!page_has_buffers(page)))
 637                                 return -ENOMEM;
 638                 }
 639         } while (++u < nr_pages);
 640         rl_write_locked = false;
 641         rl = NULL;
 642         err = 0;
 643         vcn = lcn = -1;
 644         vcn_len = 0;
 645         lcn_block = -1;
 646         was_hole = false;
 647         cpos = pos >> vol->cluster_size_bits;
 648         end = pos + bytes;
 649         cend = (end + vol->cluster_size - 1) >> vol->cluster_size_bits;
 650         /*
 651          * Loop over each page and for each page over each buffer.  Use goto to
 652          * reduce indentation.
 653          */
 654         u = 0;
 655 do_next_page:
 656         page = pages[u];
 657         bh_pos = (s64)page->index << PAGE_CACHE_SHIFT;
 658         bh = head = page_buffers(page);
 659         do {
 660                 VCN cdelta;
 661                 s64 bh_end;
 662                 unsigned bh_cofs;
 663
 664                 /* Clear buffer_new on all buffers to reinitialise state. */
 665                 if (buffer_new(bh))
 666                         clear_buffer_new(bh);
 667                 bh_end = bh_pos + blocksize;
 668                 bh_cpos = bh_pos >> vol->cluster_size_bits;
 669                 bh_cofs = bh_pos & vol->cluster_size_mask;
 670                 if (buffer_mapped(bh)) {
 671                         /*
 672                          * The buffer is already mapped.  If it is uptodate,
 673                          * ignore it.
 674                          */
 675                         if (buffer_uptodate(bh))
 676                                 continue;
 677                         /*
 678                          * The buffer is not uptodate.  If the page is uptodate
 679                          * set the buffer uptodate and otherwise ignore it.
 680                          */
 681                         if (PageUptodate(page)) {
 682                                 set_buffer_uptodate(bh);
 683                                 continue;
 684                         }
 685                         /*
 686                          * Neither the page nor the buffer are uptodate.  If
 687                          * the buffer is only partially being written to, we
 688                          * need to read it in before the write, i.e. now.
 689                          */
 690                         if ((bh_pos < pos && bh_end > pos) ||
 691                                         (bh_pos < end && bh_end > end)) {
 692                                 /*
 693                                  * If the buffer is fully or partially within
 694                                  * the initialized size, do an actual read.
 695                                  * Otherwise, simply zero the buffer.
 696                                  */
 697                                 read_lock_irqsave(&ni->size_lock, flags);
 698                                 initialized_size = ni->initialized_size;
 699                                 read_unlock_irqrestore(&ni->size_lock, flags);
 700                                 if (bh_pos < initialized_size) {
 701                                         ntfs_submit_bh_for_read(bh);
 702                                         *wait_bh++ = bh;
 703                                 } else {
 704                                         zero_user(page, bh_offset(bh),
 705                                                         blocksize);
 706                                         set_buffer_uptodate(bh);
 707                                 }
 708                         }
 709                         continue;
 710                 }
 711                 /* Unmapped buffer.  Need to map it. */
 712                 bh->b_bdev = vol->sb->s_bdev;
 713                 /*
 714                  * If the current buffer is in the same clusters as the map
 715                  * cache, there is no need to check the runlist again.  The
 716                  * map cache is made up of @vcn, which is the first cached file
 717                  * cluster, @vcn_len which is the number of cached file
 718                  * clusters, @lcn is the device cluster corresponding to @vcn,
 719                  * and @lcn_block is the block number corresponding to @lcn.
 720                  */
 721                 cdelta = bh_cpos - vcn;
 722                 if (likely(!cdelta || (cdelta > 0 && cdelta < vcn_len))) {
 723 map_buffer_cached:
 724                         BUG_ON(lcn < 0);
 725                         bh->b_blocknr = lcn_block +
 726                                         (cdelta << (vol->cluster_size_bits -
 727                                         blocksize_bits)) +
 728                                         (bh_cofs >> blocksize_bits);
 729                         set_buffer_mapped(bh);
 730                         /*
 731                          * If the page is uptodate so is the buffer.  If the
 732                          * buffer is fully outside the write, we ignore it if
 733                          * it was already allocated and we mark it dirty so it
 734                          * gets written out if we allocated it.  On the other
 735                          * hand, if we allocated the buffer but we are not
 736                          * marking it dirty we set buffer_new so we can do
 737                          * error recovery.
 738                          */
 739                         if (PageUptodate(page)) {
 740                                 if (!buffer_uptodate(bh))
 741                                         set_buffer_uptodate(bh);
 742                                 if (unlikely(was_hole)) {
 743                                         /* We allocated the buffer. */
 744                                         unmap_underlying_metadata(bh->b_bdev,
 745                                                         bh->b_blocknr);
 746                                         if (bh_end <= pos || bh_pos >= end)
 747                                                 mark_buffer_dirty(bh);
 748                                         else
 749                                                 set_buffer_new(bh);
 750                                 }
 751                                 continue;
 752                         }
 753                         /* Page is _not_ uptodate. */
 754                         if (likely(!was_hole)) {
 755                                 /*
 756                                  * Buffer was already allocated.  If it is not
 757                                  * uptodate and is only partially being written
 758                                  * to, we need to read it in before the write,
 759                                  * i.e. now.
 760                                  */
 761                                 if (!buffer_uptodate(bh) && bh_pos < end &&
 762                                                 bh_end > pos &&
 763                                                 (bh_pos < pos ||
 764                                                 bh_end > end)) {
 765                                         /*
 766                                          * If the buffer is fully or partially
 767                                          * within the initialized size, do an
 768                                          * actual read.  Otherwise, simply zero
 769                                          * the buffer.
 770                                          */
 771                                         read_lock_irqsave(&ni->size_lock,
 772                                                         flags);
 773                                         initialized_size = ni->initialized_size;
 774                                         read_unlock_irqrestore(&ni->size_lock,
 775                                                         flags);
 776                                         if (bh_pos < initialized_size) {
 777                                                 ntfs_submit_bh_for_read(bh);
 778                                                 *wait_bh++ = bh;
 779                                         } else {
 780                                                 zero_user(page, bh_offset(bh),
 781                                                                 blocksize);
 782                                                 set_buffer_uptodate(bh);
 783                                         }
 784                                 }
 785                                 continue;
 786                         }
 787                         /* We allocated the buffer. */
 788                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 789                         /*
 790                          * If the buffer is fully outside the write, zero it,
 791                          * set it uptodate, and mark it dirty so it gets
 792                          * written out.  If it is partially being written to,
 793                          * zero region surrounding the write but leave it to
 794                          * commit write to do anything else.  Finally, if the
 795                          * buffer is fully being overwritten, do nothing.
 796                          */
 797                         if (bh_end <= pos || bh_pos >= end) {
 798                                 if (!buffer_uptodate(bh)) {
 799                                         zero_user(page, bh_offset(bh),
 800                                                         blocksize);
 801                                         set_buffer_uptodate(bh);
 802                                 }
 803                                 mark_buffer_dirty(bh);
 804                                 continue;
 805                         }
 806                         set_buffer_new(bh);
 807                         if (!buffer_uptodate(bh) &&
 808                                         (bh_pos < pos || bh_end > end)) {
 809                                 u8 *kaddr;
 810                                 unsigned pofs;
 811
 812                                 kaddr = kmap_atomic(page);
 813                                 if (bh_pos < pos) {
 814                                         pofs = bh_pos & ~PAGE_CACHE_MASK;
 815                                         memset(kaddr + pofs, 0, pos - bh_pos);
 816                                 }
 817                                 if (bh_end > end) {
 818                                         pofs = end & ~PAGE_CACHE_MASK;
 819                                         memset(kaddr + pofs, 0, bh_end - end);
 820                                 }
 821                                 kunmap_atomic(kaddr);
 822                                 flush_dcache_page(page);
 823                         }
 824                         continue;
 825                 }
 826                 /*
 827                  * Slow path: this is the first buffer in the cluster.  If it
 828                  * is outside allocated size and is not uptodate, zero it and
 829                  * set it uptodate.
 830                  */
 831                 read_lock_irqsave(&ni->size_lock, flags);
 832                 initialized_size = ni->allocated_size;
 833                 read_unlock_irqrestore(&ni->size_lock, flags);
 834                 if (bh_pos > initialized_size) {
 835                         if (PageUptodate(page)) {
 836                                 if (!buffer_uptodate(bh))
 837                                         set_buffer_uptodate(bh);
 838                         } else if (!buffer_uptodate(bh)) {
 839                                 zero_user(page, bh_offset(bh), blocksize);
 840                                 set_buffer_uptodate(bh);
 841                         }
 842                         continue;
 843                 }
 844                 is_retry = false;
 845                 if (!rl) {
 846                         down_read(&ni->runlist.lock);
 847 retry_remap:
 848                         rl = ni->runlist.rl;
 849                 }
 850                 if (likely(rl != NULL)) {
 851                         /* Seek to element containing target cluster. */
 852                         while (rl->length && rl[1].vcn <= bh_cpos)
 853                                 rl++;
 854                         lcn = ntfs_rl_vcn_to_lcn(rl, bh_cpos);
 855                         if (likely(lcn >= 0)) {
 856                                 /*
 857                                  * Successful remap, setup the map cache and
 858                                  * use that to deal with the buffer.
 859                                  */
 860                                 was_hole = false;
 861                                 vcn = bh_cpos;
 862                                 vcn_len = rl[1].vcn - vcn;
 863                                 lcn_block = lcn << (vol->cluster_size_bits -
 864                                                 blocksize_bits);
 865                                 cdelta = 0;
 866                                 /*
 867                                  * If the number of remaining clusters touched
 868                                  * by the write is smaller or equal to the
 869                                  * number of cached clusters, unlock the
 870                                  * runlist as the map cache will be used from
 871                                  * now on.
 872                                  */
 873                                 if (likely(vcn + vcn_len >= cend)) {
 874                                         if (rl_write_locked) {
 875                                                 up_write(&ni->runlist.lock);
 876                                                 rl_write_locked = false;
 877                                         } else
 878                                                 up_read(&ni->runlist.lock);
 879                                         rl = NULL;
 880                                 }
 881                                 goto map_buffer_cached;
 882                         }
 883                 } else
 884                         lcn = LCN_RL_NOT_MAPPED;
 885                 /*
 886                  * If it is not a hole and not out of bounds, the runlist is
 887                  * probably unmapped so try to map it now.
 888                  */
 889                 if (unlikely(lcn != LCN_HOLE && lcn != LCN_ENOENT)) {
 890                         if (likely(!is_retry && lcn == LCN_RL_NOT_MAPPED)) {
 891                                 /* Attempt to map runlist. */
 892                                 if (!rl_write_locked) {
 893                                         /*
 894                                          * We need the runlist locked for
 895                                          * writing, so if it is locked for
 896                                          * reading relock it now and retry in
 897                                          * case it changed whilst we dropped
 898                                          * the lock.
 899                                          */
 900                                         up_read(&ni->runlist.lock);
 901                                         down_write(&ni->runlist.lock);
 902                                         rl_write_locked = true;
 903                                         goto retry_remap;
 904                                 }
 905                                 err = ntfs_map_runlist_nolock(ni, bh_cpos,
 906                                                 NULL);
 907                                 if (likely(!err)) {
 908                                         is_retry = true;
 909                                         goto retry_remap;
 910                                 }
 911                                 /*
 912                                  * If @vcn is out of bounds, pretend @lcn is
 913                                  * LCN_ENOENT.  As long as the buffer is out
 914                                  * of bounds this will work fine.
 915                                  */
 916                                 if (err == -ENOENT) {
 917                                         lcn = LCN_ENOENT;
 918                                         err = 0;
 919                                         goto rl_not_mapped_enoent;
 920                                 }
 921                         } else
 922                                 err = -EIO;
 923                         /* Failed to map the buffer, even after retrying. */
 924                         bh->b_blocknr = -1;
 925                         ntfs_error(vol->sb, "Failed to write to inode 0x%lx, "
 926                                         "attribute type 0x%x, vcn 0x%llx, "
 927                                         "vcn offset 0x%x, because its "
 928                                         "location on disk could not be "
 929                                         "determined%s (error code %i).",
 930                                         ni->mft_no, ni->type,
 931                                         (unsigned long long)bh_cpos,
 932                                         (unsigned)bh_pos &
 933                                         vol->cluster_size_mask,
 934                                         is_retry ? " even after retrying" : "",
 935                                         err);
 936                         break;
 937                 }
 938 rl_not_mapped_enoent:
 939                 /*
 940                  * The buffer is in a hole or out of bounds.  We need to fill
 941                  * the hole, unless the buffer is in a cluster which is not
 942                  * touched by the write, in which case we just leave the buffer
 943                  * unmapped.  This can only happen when the cluster size is
 944                  * less than the page cache size.
 945                  */
 946                 if (unlikely(vol->cluster_size < PAGE_CACHE_SIZE)) {
 947                         bh_cend = (bh_end + vol->cluster_size - 1) >>
 948                                         vol->cluster_size_bits;
 949                         if ((bh_cend <= cpos || bh_cpos >= cend)) {
 950                                 bh->b_blocknr = -1;
 951                                 /*
 952                                  * If the buffer is uptodate we skip it.  If it
 953                                  * is not but the page is uptodate, we can set
 954                                  * the buffer uptodate.  If the page is not
 955                                  * uptodate, we can clear the buffer and set it
 956                                  * uptodate.  Whether this is worthwhile is
 957                                  * debatable and this could be removed.
 958                                  */
 959                                 if (PageUptodate(page)) {
 960                                         if (!buffer_uptodate(bh))
 961                                                 set_buffer_uptodate(bh);
 962                                 } else if (!buffer_uptodate(bh)) {
 963                                         zero_user(page, bh_offset(bh),
 964                                                 blocksize);
 965                                         set_buffer_uptodate(bh);
 966                                 }
 967                                 continue;
 968                         }
 969                 }
 970                 /*
 971                  * Out of bounds buffer is invalid if it was not really out of
 972                  * bounds.
 973                  */
 974                 BUG_ON(lcn != LCN_HOLE);
 975                 /*
 976                  * We need the runlist locked for writing, so if it is locked
 977                  * for reading relock it now and retry in case it changed
 978                  * whilst we dropped the lock.
 979                  */
 980                 BUG_ON(!rl);
 981                 if (!rl_write_locked) {
 982                         up_read(&ni->runlist.lock);
 983                         down_write(&ni->runlist.lock);
 984                         rl_write_locked = true;
 985                         goto retry_remap;
 986                 }
 987                 /* Find the previous last allocated cluster. */
 988                 BUG_ON(rl->lcn != LCN_HOLE);
 989                 lcn = -1;
 990                 rl2 = rl;
 991                 while (--rl2 >= ni->runlist.rl) {
 992                         if (rl2->lcn >= 0) {
 993                                 lcn = rl2->lcn + rl2->length;
 994                                 break;
 995                         }
 996                 }
 997                 rl2 = ntfs_cluster_alloc(vol, bh_cpos, 1, lcn, DATA_ZONE,
 998                                 false);
 999                 if (IS_ERR(rl2)) {
1000                         err = PTR_ERR(rl2);
1001                         ntfs_debug("Failed to allocate cluster, error code %i.",
1002                                         err);
1003                         break;
1004                 }
1005                 lcn = rl2->lcn;
1006                 rl = ntfs_runlists_merge(ni->runlist.rl, rl2);
1007                 if (IS_ERR(rl)) {
1008                         err = PTR_ERR(rl);
1009                         if (err != -ENOMEM)
1010                                 err = -EIO;
1011                         if (ntfs_cluster_free_from_rl(vol, rl2)) {
1012                                 ntfs_error(vol->sb, "Failed to release "
1013                                                 "allocated cluster in error "
1014                                                 "code path.  Run chkdsk to "
1015                                                 "recover the lost cluster.");
1016                                 NVolSetErrors(vol);
1017                         }
1018                         ntfs_free(rl2);
1019                         break;
1020                 }
1021                 ni->runlist.rl = rl;
1022                 status.runlist_merged = 1;
1023                 ntfs_debug("Allocated cluster, lcn 0x%llx.",
1024                                 (unsigned long long)lcn);
1025                 /* Map and lock the mft record and get the attribute record. */
1026                 if (!NInoAttr(ni))
1027                         base_ni = ni;
1028                 else
1029                         base_ni = ni->ext.base_ntfs_ino;
1030                 m = map_mft_record(base_ni);
1031                 if (IS_ERR(m)) {
1032                         err = PTR_ERR(m);
1033                         break;
1034                 }
1035                 ctx = ntfs_attr_get_search_ctx(base_ni, m);
1036                 if (unlikely(!ctx)) {
1037                         err = -ENOMEM;
1038                         unmap_mft_record(base_ni);
1039                         break;
1040                 }
1041                 status.mft_attr_mapped = 1;
1042                 err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1043                                 CASE_SENSITIVE, bh_cpos, NULL, 0, ctx);
1044                 if (unlikely(err)) {
1045                         if (err == -ENOENT)
1046                                 err = -EIO;
1047                         break;
1048                 }
1049                 m = ctx->mrec;
1050                 a = ctx->attr;
1051                 /*
1052                  * Find the runlist element with which the attribute extent
1053                  * starts.  Note, we cannot use the _attr_ version because we
1054                  * have mapped the mft record.  That is ok because we know the
1055                  * runlist fragment must be mapped already to have ever gotten
1056                  * here, so we can just use the _rl_ version.
1057                  */
1058                 vcn = sle64_to_cpu(a->data.non_resident.lowest_vcn);
1059                 rl2 = ntfs_rl_find_vcn_nolock(rl, vcn);
1060                 BUG_ON(!rl2);
1061                 BUG_ON(!rl2->length);
1062                 BUG_ON(rl2->lcn < LCN_HOLE);
1063                 highest_vcn = sle64_to_cpu(a->data.non_resident.highest_vcn);
1064                 /*
1065                  * If @highest_vcn is zero, calculate the real highest_vcn
1066                  * (which can really be zero).
1067                  */
1068                 if (!highest_vcn)
1069                         highest_vcn = (sle64_to_cpu(
1070                                         a->data.non_resident.allocated_size) >>
1071                                         vol->cluster_size_bits) - 1;
1072                 /*
1073                  * Determine the size of the mapping pairs array for the new
1074                  * extent, i.e. the old extent with the hole filled.
1075                  */
1076                 mp_size = ntfs_get_size_for_mapping_pairs(vol, rl2, vcn,
1077                                 highest_vcn);
1078                 if (unlikely(mp_size <= 0)) {
1079                         if (!(err = mp_size))
1080                                 err = -EIO;
1081                         ntfs_debug("Failed to get size for mapping pairs "
1082                                         "array, error code %i.", err);
1083                         break;
1084                 }
1085                 /*
1086                  * Resize the attribute record to fit the new mapping pairs
1087                  * array.
1088                  */
1089                 attr_rec_len = le32_to_cpu(a->length);
1090                 err = ntfs_attr_record_resize(m, a, mp_size + le16_to_cpu(
1091                                 a->data.non_resident.mapping_pairs_offset));
1092                 if (unlikely(err)) {
1093                         BUG_ON(err != -ENOSPC);
1094                         // TODO: Deal with this by using the current attribute
1095                         // and fill it with as much of the mapping pairs
1096                         // array as possible.  Then loop over each attribute
1097                         // extent rewriting the mapping pairs arrays as we go
1098                         // along and if when we reach the end we have not
1099                         // enough space, try to resize the last attribute
1100                         // extent and if even that fails, add a new attribute
1101                         // extent.
1102                         // We could also try to resize at each step in the hope
1103                         // that we will not need to rewrite every single extent.
1104                         // Note, we may need to decompress some extents to fill
1105                         // the runlist as we are walking the extents...
1106                         ntfs_error(vol->sb, "Not enough space in the mft "
1107                                         "record for the extended attribute "
1108                                         "record.  This case is not "
1109                                         "implemented yet.");
1110                         err = -EOPNOTSUPP;
1111                         break ;
1112                 }
1113                 status.mp_rebuilt = 1;
1114                 /*
1115                  * Generate the mapping pairs array directly into the attribute
1116                  * record.
1117                  */
1118                 err = ntfs_mapping_pairs_build(vol, (u8*)a + le16_to_cpu(
1119                                 a->data.non_resident.mapping_pairs_offset),
1120                                 mp_size, rl2, vcn, highest_vcn, NULL);
1121                 if (unlikely(err)) {
1122                         ntfs_error(vol->sb, "Cannot fill hole in inode 0x%lx, "
1123                                         "attribute type 0x%x, because building "
1124                                         "the mapping pairs failed with error "
1125                                         "code %i.", vi->i_ino,
1126                                         (unsigned)le32_to_cpu(ni->type), err);
1127                         err = -EIO;
1128                         break;
1129                 }
1130                 /* Update the highest_vcn but only if it was not set. */
1131                 if (unlikely(!a->data.non_resident.highest_vcn))
1132                         a->data.non_resident.highest_vcn =
1133                                         cpu_to_sle64(highest_vcn);
1134                 /*
1135                  * If the attribute is sparse/compressed, update the compressed
1136                  * size in the ntfs_inode structure and the attribute record.
1137                  */
1138                 if (likely(NInoSparse(ni) || NInoCompressed(ni))) {
1139                         /*
1140                          * If we are not in the first attribute extent, switch
1141                          * to it, but first ensure the changes will make it to
1142                          * disk later.
1143                          */
1144                         if (a->data.non_resident.lowest_vcn) {
1145                                 flush_dcache_mft_record_page(ctx->ntfs_ino);
1146                                 mark_mft_record_dirty(ctx->ntfs_ino);
1147                                 ntfs_attr_reinit_search_ctx(ctx);
1148                                 err = ntfs_attr_lookup(ni->type, ni->name,
1149                                                 ni->name_len, CASE_SENSITIVE,
1150                                                 0, NULL, 0, ctx);
1151                                 if (unlikely(err)) {
1152                                         status.attr_switched = 1;
1153                                         break;
1154                                 }
1155                                 /* @m is not used any more so do not set it. */
1156                                 a = ctx->attr;
1157                         }
1158                         write_lock_irqsave(&ni->size_lock, flags);
1159                         ni->itype.compressed.size += vol->cluster_size;
1160                         a->data.non_resident.compressed_size =
1161                                         cpu_to_sle64(ni->itype.compressed.size);
1162                         write_unlock_irqrestore(&ni->size_lock, flags);
1163                 }
1164                 /* Ensure the changes make it to disk. */
1165                 flush_dcache_mft_record_page(ctx->ntfs_ino);
1166                 mark_mft_record_dirty(ctx->ntfs_ino);
1167                 ntfs_attr_put_search_ctx(ctx);
1168                 unmap_mft_record(base_ni);
1169                 /* Successfully filled the hole. */
1170                 status.runlist_merged = 0;
1171                 status.mft_attr_mapped = 0;
1172                 status.mp_rebuilt = 0;
1173                 /* Setup the map cache and use that to deal with the buffer. */
1174                 was_hole = true;
1175                 vcn = bh_cpos;
1176                 vcn_len = 1;
1177                 lcn_block = lcn << (vol->cluster_size_bits - blocksize_bits);
1178                 cdelta = 0;
1179                 /*
1180                  * If the number of remaining clusters in the @pages is smaller
1181                  * or equal to the number of cached clusters, unlock the
1182                  * runlist as the map cache will be used from now on.
1183                  */
1184                 if (likely(vcn + vcn_len >= cend)) {
1185                         up_write(&ni->runlist.lock);
1186                         rl_write_locked = false;
1187                         rl = NULL;
1188                 }
1189                 goto map_buffer_cached;
1190         } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
1191         /* If there are no errors, do the next page. */
1192         if (likely(!err && ++u < nr_pages))
1193                 goto do_next_page;
1194         /* If there are no errors, release the runlist lock if we took it. */
1195         if (likely(!err)) {
1196                 if (unlikely(rl_write_locked)) {
1197                         up_write(&ni->runlist.lock);
1198                         rl_write_locked = false;
1199                 } else if (unlikely(rl))
1200                         up_read(&ni->runlist.lock);
1201                 rl = NULL;
1202         }
1203         /* If we issued read requests, let them complete. */
1204         read_lock_irqsave(&ni->size_lock, flags);
1205         initialized_size = ni->initialized_size;
1206         read_unlock_irqrestore(&ni->size_lock, flags);
1207         while (wait_bh > wait) {
1208                 bh = *--wait_bh;
1209                 wait_on_buffer(bh);
1210                 if (likely(buffer_uptodate(bh))) {
1211                         page = bh->b_page;
1212                         bh_pos = ((s64)page->index << PAGE_CACHE_SHIFT) +
1213                                         bh_offset(bh);
1214                         /*
1215                          * If the buffer overflows the initialized size, need
1216                          * to zero the overflowing region.
1217                          */
1218                         if (unlikely(bh_pos + blocksize > initialized_size)) {
1219                                 int ofs = 0;
1220
1221                                 if (likely(bh_pos < initialized_size))
1222                                         ofs = initialized_size - bh_pos;
1223                                 zero_user_segment(page, bh_offset(bh) + ofs,
1224                                                 blocksize);
1225                         }
1226                 } else /* if (unlikely(!buffer_uptodate(bh))) */
1227                         err = -EIO;
1228         }
1229         if (likely(!err)) {
1230                 /* Clear buffer_new on all buffers. */
1231                 u = 0;
1232                 do {
1233                         bh = head = page_buffers(pages[u]);
1234                         do {
1235                                 if (buffer_new(bh))
1236                                         clear_buffer_new(bh);
1237                         } while ((bh = bh->b_this_page) != head);
1238                 } while (++u < nr_pages);
1239                 ntfs_debug("Done.");
1240                 return err;
1241         }
1242         if (status.attr_switched) {
1243                 /* Get back to the attribute extent we modified. */
1244                 ntfs_attr_reinit_search_ctx(ctx);
1245                 if (ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1246                                 CASE_SENSITIVE, bh_cpos, NULL, 0, ctx)) {
1247                         ntfs_error(vol->sb, "Failed to find required "
1248                                         "attribute extent of attribute in "
1249                                         "error code path.  Run chkdsk to "
1250                                         "recover.");
1251                         write_lock_irqsave(&ni->size_lock, flags);
1252                         ni->itype.compressed.size += vol->cluster_size;
1253                         write_unlock_irqrestore(&ni->size_lock, flags);
1254                         flush_dcache_mft_record_page(ctx->ntfs_ino);
1255                         mark_mft_record_dirty(ctx->ntfs_ino);
1256                         /*
1257                          * The only thing that is now wrong is the compressed
1258                          * size of the base attribute extent which chkdsk
1259                          * should be able to fix.
1260                          */
1261                         NVolSetErrors(vol);
1262                 } else {
1263                         m = ctx->mrec;
1264                         a = ctx->attr;
1265                         status.attr_switched = 0;
1266                 }
1267         }
1268         /*
1269          * If the runlist has been modified, need to restore it by punching a
1270          * hole into it and we then need to deallocate the on-disk cluster as
1271          * well.  Note, we only modify the runlist if we are able to generate a
1272          * new mapping pairs array, i.e. only when the mapped attribute extent
1273          * is not switched.
1274          */
1275         if (status.runlist_merged && !status.attr_switched) {
1276                 BUG_ON(!rl_write_locked);
1277                 /* Make the file cluster we allocated sparse in the runlist. */
1278                 if (ntfs_rl_punch_nolock(vol, &ni->runlist, bh_cpos, 1)) {
1279                         ntfs_error(vol->sb, "Failed to punch hole into "
1280                                         "attribute runlist in error code "
1281                                         "path.  Run chkdsk to recover the "
1282                                         "lost cluster.");
1283                         NVolSetErrors(vol);
1284                 } else /* if (success) */ {
1285                         status.runlist_merged = 0;
1286                         /*
1287                          * Deallocate the on-disk cluster we allocated but only
1288                          * if we succeeded in punching its vcn out of the
1289                          * runlist.
1290                          */
1291                         down_write(&vol->lcnbmp_lock);
1292                         if (ntfs_bitmap_clear_bit(vol->lcnbmp_ino, lcn)) {
1293                                 ntfs_error(vol->sb, "Failed to release "
1294                                                 "allocated cluster in error "
1295                                                 "code path.  Run chkdsk to "
1296                                                 "recover the lost cluster.");
1297                                 NVolSetErrors(vol);
1298                         }
1299                         up_write(&vol->lcnbmp_lock);
1300                 }
1301         }
1302         /*
1303          * Resize the attribute record to its old size and rebuild the mapping
1304          * pairs array.  Note, we only can do this if the runlist has been
1305          * restored to its old state which also implies that the mapped
1306          * attribute extent is not switched.
1307          */
1308         if (status.mp_rebuilt && !status.runlist_merged) {
1309                 if (ntfs_attr_record_resize(m, a, attr_rec_len)) {
1310                         ntfs_error(vol->sb, "Failed to restore attribute "
1311                                         "record in error code path.  Run "
1312                                         "chkdsk to recover.");
1313                         NVolSetErrors(vol);
1314                 } else /* if (success) */ {
1315                         if (ntfs_mapping_pairs_build(vol, (u8*)a +
1316                                         le16_to_cpu(a->data.non_resident.
1317                                         mapping_pairs_offset), attr_rec_len -
1318                                         le16_to_cpu(a->data.non_resident.
1319                                         mapping_pairs_offset), ni->runlist.rl,
1320                                         vcn, highest_vcn, NULL)) {
1321                                 ntfs_error(vol->sb, "Failed to restore "
1322                                                 "mapping pairs array in error "
1323                                                 "code path.  Run chkdsk to "
1324                                                 "recover.");
1325                                 NVolSetErrors(vol);
1326                         }
1327                         flush_dcache_mft_record_page(ctx->ntfs_ino);
1328                         mark_mft_record_dirty(ctx->ntfs_ino);
1329                 }
1330         }
1331         /* Release the mft record and the attribute. */
1332         if (status.mft_attr_mapped) {
1333                 ntfs_attr_put_search_ctx(ctx);
1334                 unmap_mft_record(base_ni);
1335         }
1336         /* Release the runlist lock. */
1337         if (rl_write_locked)
1338                 up_write(&ni->runlist.lock);
1339         else if (rl)
1340                 up_read(&ni->runlist.lock);
1341         /*
1342          * Zero out any newly allocated blocks to avoid exposing stale data.
1343          * If BH_New is set, we know that the block was newly allocated above
1344          * and that it has not been fully zeroed and marked dirty yet.
1345          */
1346         nr_pages = u;
1347         u = 0;
1348         end = bh_cpos << vol->cluster_size_bits;
1349         do {
1350                 page = pages[u];
1351                 bh = head = page_buffers(page);
1352                 do {
1353                         if (u == nr_pages &&
1354                                         ((s64)page->index << PAGE_CACHE_SHIFT) +
1355                                         bh_offset(bh) >= end)
1356                                 break;
1357                         if (!buffer_new(bh))
1358                                 continue;
1359                         clear_buffer_new(bh);
1360                         if (!buffer_uptodate(bh)) {
1361                                 if (PageUptodate(page))
1362                                         set_buffer_uptodate(bh);
1363                                 else {
1364                                         zero_user(page, bh_offset(bh),
1365                                                         blocksize);
1366                                         set_buffer_uptodate(bh);
1367                                 }
1368                         }
1369                         mark_buffer_dirty(bh);
1370                 } while ((bh = bh->b_this_page) != head);
1371         } while (++u <= nr_pages);
1372         ntfs_error(vol->sb, "Failed.  Returning error code %i.", err);
1373         return err;
1374 }
1375
1376 static inline void ntfs_flush_dcache_pages(struct page **pages,
1377                 unsigned nr_pages)
1378 {
1379         BUG_ON(!nr_pages);
1380         /*
1381          * Warning: Do not do the decrement at the same time as the call to
1382          * flush_dcache_page() because it is a NULL macro on i386 and hence the
1383          * decrement never happens so the loop never terminates.
1384          */
1385         do {
1386                 --nr_pages;
1387                 flush_dcache_page(pages[nr_pages]);
1388         } while (nr_pages > 0);
1389 }
1390
1391 /**
1392  * ntfs_commit_pages_after_non_resident_write - commit the received data
1393  * @pages:      array of destination pages
1394  * @nr_pages:   number of pages in @pages
1395  * @pos:        byte position in file at which the write begins
1396  * @bytes:      number of bytes to be written
1397  *
1398  * See description of ntfs_commit_pages_after_write(), below.
1399  */
1400 static inline int ntfs_commit_pages_after_non_resident_write(
1401                 struct page **pages, const unsigned nr_pages,
1402                 s64 pos, size_t bytes)
1403 {
1404         s64 end, initialized_size;
1405         struct inode *vi;
1406         ntfs_inode *ni, *base_ni;
1407         struct buffer_head *bh, *head;
1408         ntfs_attr_search_ctx *ctx;
1409         MFT_RECORD *m;
1410         ATTR_RECORD *a;
1411         unsigned long flags;
1412         unsigned blocksize, u;
1413         int err;
1414
1415         vi = pages[0]->mapping->host;
1416         ni = NTFS_I(vi);
1417         blocksize = vi->i_sb->s_blocksize;
1418         end = pos + bytes;
1419         u = 0;
1420         do {
1421                 s64 bh_pos;
1422                 struct page *page;
1423                 bool partial;
1424
1425                 page = pages[u];
1426                 bh_pos = (s64)page->index << PAGE_CACHE_SHIFT;
1427                 bh = head = page_buffers(page);
1428                 partial = false;
1429                 do {
1430                         s64 bh_end;
1431
1432                         bh_end = bh_pos + blocksize;
1433                         if (bh_end <= pos || bh_pos >= end) {
1434                                 if (!buffer_uptodate(bh))
1435                                         partial = true;
1436                         } else {
1437                                 set_buffer_uptodate(bh);
1438                                 mark_buffer_dirty(bh);
1439                         }
1440                 } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
1441                 /*
1442                  * If all buffers are now uptodate but the page is not, set the
1443                  * page uptodate.
1444                  */
1445                 if (!partial && !PageUptodate(page))
1446                         SetPageUptodate(page);
1447         } while (++u < nr_pages);
1448         /*
1449          * Finally, if we do not need to update initialized_size or i_size we
1450          * are finished.
1451          */
1452         read_lock_irqsave(&ni->size_lock, flags);
1453         initialized_size = ni->initialized_size;
1454         read_unlock_irqrestore(&ni->size_lock, flags);
1455         if (end <= initialized_size) {
1456                 ntfs_debug("Done.");
1457                 return 0;
1458         }
1459         /*
1460          * Update initialized_size/i_size as appropriate, both in the inode and
1461          * the mft record.
1462          */
1463         if (!NInoAttr(ni))
1464                 base_ni = ni;
1465         else
1466                 base_ni = ni->ext.base_ntfs_ino;
1467         /* Map, pin, and lock the mft record. */
1468         m = map_mft_record(base_ni);
1469         if (IS_ERR(m)) {
1470                 err = PTR_ERR(m);
1471                 m = NULL;
1472                 ctx = NULL;
1473                 goto err_out;
1474         }
1475         BUG_ON(!NInoNonResident(ni));
1476         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1477         if (unlikely(!ctx)) {
1478                 err = -ENOMEM;
1479                 goto err_out;
1480         }
1481         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1482                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1483         if (unlikely(err)) {
1484                 if (err == -ENOENT)
1485                         err = -EIO;
1486                 goto err_out;
1487         }
1488         a = ctx->attr;
1489         BUG_ON(!a->non_resident);
1490         write_lock_irqsave(&ni->size_lock, flags);
1491         BUG_ON(end > ni->allocated_size);
1492         ni->initialized_size = end;
1493         a->data.non_resident.initialized_size = cpu_to_sle64(end);
1494         if (end > i_size_read(vi)) {
1495                 i_size_write(vi, end);
1496                 a->data.non_resident.data_size =
1497                                 a->data.non_resident.initialized_size;
1498         }
1499         write_unlock_irqrestore(&ni->size_lock, flags);
1500         /* Mark the mft record dirty, so it gets written back. */
1501         flush_dcache_mft_record_page(ctx->ntfs_ino);
1502         mark_mft_record_dirty(ctx->ntfs_ino);
1503         ntfs_attr_put_search_ctx(ctx);
1504         unmap_mft_record(base_ni);
1505         ntfs_debug("Done.");
1506         return 0;
1507 err_out:
1508         if (ctx)
1509                 ntfs_attr_put_search_ctx(ctx);
1510         if (m)
1511                 unmap_mft_record(base_ni);
1512         ntfs_error(vi->i_sb, "Failed to update initialized_size/i_size (error "
1513                         "code %i).", err);
1514         if (err != -ENOMEM)
1515                 NVolSetErrors(ni->vol);
1516         return err;
1517 }
1518
1519 /**
1520  * ntfs_commit_pages_after_write - commit the received data
1521  * @pages:      array of destination pages
1522  * @nr_pages:   number of pages in @pages
1523  * @pos:        byte position in file at which the write begins
1524  * @bytes:      number of bytes to be written
1525  *
1526  * This is called from ntfs_file_buffered_write() with i_mutex held on the inode
1527  * (@pages[0]->mapping->host).  There are @nr_pages pages in @pages which are
1528  * locked but not kmap()ped.  The source data has already been copied into the
1529  * @page.  ntfs_prepare_pages_for_non_resident_write() has been called before
1530  * the data was copied (for non-resident attributes only) and it returned
1531  * success.
1532  *
1533  * Need to set uptodate and mark dirty all buffers within the boundary of the
1534  * write.  If all buffers in a page are uptodate we set the page uptodate, too.
1535  *
1536  * Setting the buffers dirty ensures that they get written out later when
1537  * ntfs_writepage() is invoked by the VM.
1538  *
1539  * Finally, we need to update i_size and initialized_size as appropriate both
1540  * in the inode and the mft record.
1541  *
1542  * This is modelled after fs/buffer.c::generic_commit_write(), which marks
1543  * buffers uptodate and dirty, sets the page uptodate if all buffers in the
1544  * page are uptodate, and updates i_size if the end of io is beyond i_size.  In
1545  * that case, it also marks the inode dirty.
1546  *
1547  * If things have gone as outlined in
1548  * ntfs_prepare_pages_for_non_resident_write(), we do not need to do any page
1549  * content modifications here for non-resident attributes.  For resident
1550  * attributes we need to do the uptodate bringing here which we combine with
1551  * the copying into the mft record which means we save one atomic kmap.
1552  *
1553  * Return 0 on success or -errno on error.
1554  */
1555 static int ntfs_commit_pages_after_write(struct page **pages,
1556                 const unsigned nr_pages, s64 pos, size_t bytes)
1557 {
1558         s64 end, initialized_size;
1559         loff_t i_size;
1560         struct inode *vi;
1561         ntfs_inode *ni, *base_ni;
1562         struct page *page;
1563         ntfs_attr_search_ctx *ctx;
1564         MFT_RECORD *m;
1565         ATTR_RECORD *a;
1566         char *kattr, *kaddr;
1567         unsigned long flags;
1568         u32 attr_len;
1569         int err;
1570
1571         BUG_ON(!nr_pages);
1572         BUG_ON(!pages);
1573         page = pages[0];
1574         BUG_ON(!page);
1575         vi = page->mapping->host;
1576         ni = NTFS_I(vi);
1577         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, start page "
1578                         "index 0x%lx, nr_pages 0x%x, pos 0x%llx, bytes 0x%zx.",
1579                         vi->i_ino, ni->type, page->index, nr_pages,
1580                         (long long)pos, bytes);
1581         if (NInoNonResident(ni))
1582                 return ntfs_commit_pages_after_non_resident_write(pages,
1583                                 nr_pages, pos, bytes);
1584         BUG_ON(nr_pages > 1);
1585         /*
1586          * Attribute is resident, implying it is not compressed, encrypted, or
1587          * sparse.
1588          */
1589         if (!NInoAttr(ni))
1590                 base_ni = ni;
1591         else
1592                 base_ni = ni->ext.base_ntfs_ino;
1593         BUG_ON(NInoNonResident(ni));
1594         /* Map, pin, and lock the mft record. */
1595         m = map_mft_record(base_ni);
1596         if (IS_ERR(m)) {
1597                 err = PTR_ERR(m);
1598                 m = NULL;
1599                 ctx = NULL;
1600                 goto err_out;
1601         }
1602         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1603         if (unlikely(!ctx)) {
1604                 err = -ENOMEM;
1605                 goto err_out;
1606         }
1607         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1608                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1609         if (unlikely(err)) {
1610                 if (err == -ENOENT)
1611                         err = -EIO;
1612                 goto err_out;
1613         }
1614         a = ctx->attr;
1615         BUG_ON(a->non_resident);
1616         /* The total length of the attribute value. */
1617         attr_len = le32_to_cpu(a->data.resident.value_length);
1618         i_size = i_size_read(vi);
1619         BUG_ON(attr_len != i_size);
1620         BUG_ON(pos > attr_len);
1621         end = pos + bytes;
1622         BUG_ON(end > le32_to_cpu(a->length) -
1623                         le16_to_cpu(a->data.resident.value_offset));
1624         kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
1625         kaddr = kmap_atomic(page);
1626         /* Copy the received data from the page to the mft record. */
1627         memcpy(kattr + pos, kaddr + pos, bytes);
1628         /* Update the attribute length if necessary. */
1629         if (end > attr_len) {
1630                 attr_len = end;
1631                 a->data.resident.value_length = cpu_to_le32(attr_len);
1632         }
1633         /*
1634          * If the page is not uptodate, bring the out of bounds area(s)
1635          * uptodate by copying data from the mft record to the page.
1636          */
1637         if (!PageUptodate(page)) {
1638                 if (pos > 0)
1639                         memcpy(kaddr, kattr, pos);
1640                 if (end < attr_len)
1641                         memcpy(kaddr + end, kattr + end, attr_len - end);
1642                 /* Zero the region outside the end of the attribute value. */
1643                 memset(kaddr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
1644                 flush_dcache_page(page);
1645                 SetPageUptodate(page);
1646         }
1647         kunmap_atomic(kaddr);
1648         /* Update initialized_size/i_size if necessary. */
1649         read_lock_irqsave(&ni->size_lock, flags);
1650         initialized_size = ni->initialized_size;
1651         BUG_ON(end > ni->allocated_size);
1652         read_unlock_irqrestore(&ni->size_lock, flags);
1653         BUG_ON(initialized_size != i_size);
1654         if (end > initialized_size) {
1655                 write_lock_irqsave(&ni->size_lock, flags);
1656                 ni->initialized_size = end;
1657                 i_size_write(vi, end);
1658                 write_unlock_irqrestore(&ni->size_lock, flags);
1659         }
1660         /* Mark the mft record dirty, so it gets written back. */
1661         flush_dcache_mft_record_page(ctx->ntfs_ino);
1662         mark_mft_record_dirty(ctx->ntfs_ino);
1663         ntfs_attr_put_search_ctx(ctx);
1664         unmap_mft_record(base_ni);
1665         ntfs_debug("Done.");
1666         return 0;
1667 err_out:
1668         if (err == -ENOMEM) {
1669                 ntfs_warning(vi->i_sb, "Error allocating memory required to "
1670                                 "commit the write.");
1671                 if (PageUptodate(page)) {
1672                         ntfs_warning(vi->i_sb, "Page is uptodate, setting "
1673                                         "dirty so the write will be retried "
1674                                         "later on by the VM.");
1675                         /*
1676                          * Put the page on mapping->dirty_pages, but leave its
1677                          * buffers' dirty state as-is.
1678                          */
1679                         __set_page_dirty_nobuffers(page);
1680                         err = 0;
1681                 } else
1682                         ntfs_error(vi->i_sb, "Page is not uptodate.  Written "
1683                                         "data has been lost.");
1684         } else {
1685                 ntfs_error(vi->i_sb, "Resident attribute commit write failed "
1686                                 "with error %i.", err);
1687                 NVolSetErrors(ni->vol);
1688         }
1689         if (ctx)
1690                 ntfs_attr_put_search_ctx(ctx);
1691         if (m)
1692                 unmap_mft_record(base_ni);
1693         return err;
1694 }
1695
1696 /*
1697  * Copy as much as we can into the pages and return the number of bytes which
1698  * were successfully copied.  If a fault is encountered then clear the pages
1699  * out to (ofs + bytes) and return the number of bytes which were copied.
1700  */
1701 static size_t ntfs_copy_from_user_iter(struct page **pages, unsigned nr_pages,
1702                 unsigned ofs, struct iov_iter *i, size_t bytes)
1703 {
1704         struct page **last_page = pages + nr_pages;
1705         size_t total = 0;
1706         struct iov_iter data = *i;
1707         unsigned len, copied;
1708
1709         do {
1710                 len = PAGE_CACHE_SIZE - ofs;
1711                 if (len > bytes)
1712                         len = bytes;
1713                 copied = iov_iter_copy_from_user_atomic(*pages, &data, ofs,
1714                                 len);
1715                 total += copied;
1716                 bytes -= copied;
1717                 if (!bytes)
1718                         break;
1719                 iov_iter_advance(&data, copied);
1720                 if (copied < len)
1721                         goto err;
1722                 ofs = 0;
1723         } while (++pages < last_page);
1724 out:
1725         return total;
1726 err:
1727         /* Zero the rest of the target like __copy_from_user(). */
1728         len = PAGE_CACHE_SIZE - copied;
1729         do {
1730                 if (len > bytes)
1731                         len = bytes;
1732                 zero_user(*pages, copied, len);
1733                 bytes -= len;
1734                 copied = 0;
1735                 len = PAGE_CACHE_SIZE;
1736         } while (++pages < last_page);
1737         goto out;
1738 }
1739
1740 /**
1741  * ntfs_perform_write - perform buffered write to a file
1742  * @file:       file to write to
1743  * @i:          iov_iter with data to write
1744  * @pos:        byte offset in file at which to begin writing to
1745  */
1746 static ssize_t ntfs_perform_write(struct file *file, struct iov_iter *i,
1747                 loff_t pos)
1748 {
1749         struct address_space *mapping = file->f_mapping;
1750         struct inode *vi = mapping->host;
1751         ntfs_inode *ni = NTFS_I(vi);
1752         ntfs_volume *vol = ni->vol;
1753         struct page *pages[NTFS_MAX_PAGES_PER_CLUSTER];
1754         struct page *cached_page = NULL;
1755         VCN last_vcn;
1756         LCN lcn;
1757         size_t bytes;
1758         ssize_t status, written = 0;
1759         unsigned nr_pages;
1760
1761         ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, pos "
1762                         "0x%llx, count 0x%lx.", vi->i_ino,
1763                         (unsigned)le32_to_cpu(ni->type),
1764                         (unsigned long long)pos,
1765                         (unsigned long)iov_iter_count(i));
1766         /*
1767          * If a previous ntfs_truncate() failed, repeat it and abort if it
1768          * fails again.
1769          */
1770         if (unlikely(NInoTruncateFailed(ni))) {
1771                 int err;
1772
1773                 inode_dio_wait(vi);
1774                 err = ntfs_truncate(vi);
1775                 if (err || NInoTruncateFailed(ni)) {
1776                         if (!err)
1777                                 err = -EIO;
1778                         ntfs_error(vol->sb, "Cannot perform write to inode "
1779                                         "0x%lx, attribute type 0x%x, because "
1780                                         "ntfs_truncate() failed (error code "
1781                                         "%i).", vi->i_ino,
1782                                         (unsigned)le32_to_cpu(ni->type), err);
1783                         return err;
1784                 }
1785         }
1786         /*
1787          * Determine the number of pages per cluster for non-resident
1788          * attributes.
1789          */
1790         nr_pages = 1;
1791         if (vol->cluster_size > PAGE_CACHE_SIZE && NInoNonResident(ni))
1792                 nr_pages = vol->cluster_size >> PAGE_CACHE_SHIFT;
1793         last_vcn = -1;
1794         do {
1795                 VCN vcn;
1796                 pgoff_t idx, start_idx;
1797                 unsigned ofs, do_pages, u;
1798                 size_t copied;
1799
1800                 start_idx = idx = pos >> PAGE_CACHE_SHIFT;
1801                 ofs = pos & ~PAGE_CACHE_MASK;
1802                 bytes = PAGE_CACHE_SIZE - ofs;
1803                 do_pages = 1;
1804                 if (nr_pages > 1) {
1805                         vcn = pos >> vol->cluster_size_bits;
1806                         if (vcn != last_vcn) {
1807                                 last_vcn = vcn;
1808                                 /*
1809                                  * Get the lcn of the vcn the write is in.  If
1810                                  * it is a hole, need to lock down all pages in
1811                                  * the cluster.
1812                                  */
1813                                 down_read(&ni->runlist.lock);
1814                                 lcn = ntfs_attr_vcn_to_lcn_nolock(ni, pos >>
1815                                                 vol->cluster_size_bits, false);
1816                                 up_read(&ni->runlist.lock);
1817                                 if (unlikely(lcn < LCN_HOLE)) {
1818                                         if (lcn == LCN_ENOMEM)
1819                                                 status = -ENOMEM;
1820                                         else {
1821                                                 status = -EIO;
1822                                                 ntfs_error(vol->sb, "Cannot "
1823                                                         "perform write to "
1824                                                         "inode 0x%lx, "
1825                                                         "attribute type 0x%x, "
1826                                                         "because the attribute "
1827                                                         "is corrupt.",
1828                                                         vi->i_ino, (unsigned)
1829                                                         le32_to_cpu(ni->type));
1830                                         }
1831                                         break;
1832                                 }
1833                                 if (lcn == LCN_HOLE) {
1834                                         start_idx = (pos & ~(s64)
1835                                                         vol->cluster_size_mask)
1836                                                         >> PAGE_CACHE_SHIFT;
1837                                         bytes = vol->cluster_size - (pos &
1838                                                         vol->cluster_size_mask);
1839                                         do_pages = nr_pages;
1840                                 }
1841                         }
1842                 }
1843                 if (bytes > iov_iter_count(i))
1844                         bytes = iov_iter_count(i);
1845 again:
1846                 /*
1847                  * Bring in the user page(s) that we will copy from _first_.
1848                  * Otherwise there is a nasty deadlock on copying from the same
1849                  * page(s) as we are writing to, without it/them being marked
1850                  * up-to-date.  Note, at present there is nothing to stop the
1851                  * pages being swapped out between us bringing them into memory
1852                  * and doing the actual copying.
1853                  */
1854                 if (unlikely(iov_iter_fault_in_multipages_readable(i, bytes))) {
1855                         status = -EFAULT;
1856                         break;
1857                 }
1858                 /* Get and lock @do_pages starting at index @start_idx. */
1859                 status = __ntfs_grab_cache_pages(mapping, start_idx, do_pages,
1860                                 pages, &cached_page);
1861                 if (unlikely(status))
1862                         break;
1863                 /*
1864                  * For non-resident attributes, we need to fill any holes with
1865                  * actual clusters and ensure all bufferes are mapped.  We also
1866                  * need to bring uptodate any buffers that are only partially
1867                  * being written to.
1868                  */
1869                 if (NInoNonResident(ni)) {
1870                         status = ntfs_prepare_pages_for_non_resident_write(
1871                                         pages, do_pages, pos, bytes);
1872                         if (unlikely(status)) {
1873                                 do {
1874                                         unlock_page(pages[--do_pages]);
1875                                         page_cache_release(pages[do_pages]);
1876                                 } while (do_pages);
1877                                 break;
1878                         }
1879                 }
1880                 u = (pos >> PAGE_CACHE_SHIFT) - pages[0]->index;
1881                 copied = ntfs_copy_from_user_iter(pages + u, do_pages - u, ofs,
1882                                         i, bytes);
1883                 ntfs_flush_dcache_pages(pages + u, do_pages - u);
1884                 status = 0;
1885                 if (likely(copied == bytes)) {
1886                         status = ntfs_commit_pages_after_write(pages, do_pages,
1887                                         pos, bytes);
1888                         if (!status)
1889                                 status = bytes;
1890                 }
1891                 do {
1892                         unlock_page(pages[--do_pages]);
1893                         page_cache_release(pages[do_pages]);
1894                 } while (do_pages);
1895                 if (unlikely(status < 0))
1896                         break;
1897                 copied = status;
1898                 cond_resched();
1899                 if (unlikely(!copied)) {
1900                         size_t sc;
1901
1902                         /*
1903                          * We failed to copy anything.  Fall back to single
1904                          * segment length write.
1905                          *
1906                          * This is needed to avoid possible livelock in the
1907                          * case that all segments in the iov cannot be copied
1908                          * at once without a pagefault.
1909                          */
1910                         sc = iov_iter_single_seg_count(i);
1911                         if (bytes > sc)
1912                                 bytes = sc;
1913                         goto again;
1914                 }
1915                 iov_iter_advance(i, copied);
1916                 pos += copied;
1917                 written += copied;
1918                 balance_dirty_pages_ratelimited(mapping);
1919                 if (fatal_signal_pending(current)) {
1920                         status = -EINTR;
1921                         break;
1922                 }
1923         } while (iov_iter_count(i));
1924         if (cached_page)
1925                 page_cache_release(cached_page);
1926         ntfs_debug("Done.  Returning %s (written 0x%lx, status %li).",
1927                         written ? "written" : "status", (unsigned long)written,
1928                         (long)status);
1929         return written ? written : status;
1930 }
1931
1932 /**
1933  * ntfs_file_write_iter_nolock - write data to a file
1934  * @iocb:       IO state structure (file, offset, etc.)
1935  * @from:       iov_iter with data to write
1936  *
1937  * Basically the same as __generic_file_write_iter() except that it ends
1938  * up calling ntfs_perform_write() instead of generic_perform_write() and that
1939  * O_DIRECT is not implemented.
1940  */
1941 static ssize_t ntfs_file_write_iter_nolock(struct kiocb *iocb,
1942                 struct iov_iter *from)
1943 {
1944         struct file *file = iocb->ki_filp;
1945         loff_t pos = iocb->ki_pos;
1946         ssize_t written = 0;
1947         ssize_t err;
1948         size_t count = iov_iter_count(from);
1949
1950         err = ntfs_prepare_file_for_write(file, &pos, &count);
1951         if (count && !err) {
1952                 iov_iter_truncate(from, count);
1953                 written = ntfs_perform_write(file, from, pos);
1954                 if (likely(written >= 0))
1955                         iocb->ki_pos = pos + written;
1956         }
1957         current->backing_dev_info = NULL;
1958         return written ? written : err;
1959 }
1960
1961 /**
1962  * ntfs_file_write_iter - simple wrapper for ntfs_file_write_iter_nolock()
1963  * @iocb:       IO state structure
1964  * @from:       iov_iter with data to write
1965  *
1966  * Basically the same as generic_file_write_iter() except that it ends up
1967  * calling ntfs_file_write_iter_nolock() instead of
1968  * __generic_file_write_iter().
1969  */
1970 static ssize_t ntfs_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1971 {
1972         struct file *file = iocb->ki_filp;
1973         struct inode *vi = file_inode(file);
1974         ssize_t ret;
1975
1976         mutex_lock(&vi->i_mutex);
1977         ret = ntfs_file_write_iter_nolock(iocb, from);
1978         mutex_unlock(&vi->i_mutex);
1979         if (ret > 0) {
1980                 ssize_t err;
1981
1982                 err = generic_write_sync(file, iocb->ki_pos - ret, ret);
1983                 if (err < 0)
1984                         ret = err;
1985         }
1986         return ret;
1987 }
1988
1989 /**
1990  * ntfs_file_fsync - sync a file to disk
1991  * @filp:       file to be synced
1992  * @datasync:   if non-zero only flush user data and not metadata
1993  *
1994  * Data integrity sync of a file to disk.  Used for fsync, fdatasync, and msync
1995  * system calls.  This function is inspired by fs/buffer.c::file_fsync().
1996  *
1997  * If @datasync is false, write the mft record and all associated extent mft
1998  * records as well as the $DATA attribute and then sync the block device.
1999  *
2000  * If @datasync is true and the attribute is non-resident, we skip the writing
2001  * of the mft record and all associated extent mft records (this might still
2002  * happen due to the write_inode_now() call).
2003  *
2004  * Also, if @datasync is true, we do not wait on the inode to be written out
2005  * but we always wait on the page cache pages to be written out.
2006  *
2007  * Locking: Caller must hold i_mutex on the inode.
2008  *
2009  * TODO: We should probably also write all attribute/index inodes associated
2010  * with this inode but since we have no simple way of getting to them we ignore
2011  * this problem for now.
2012  */
2013 static int ntfs_file_fsync(struct file *filp, loff_t start, loff_t end,
2014                            int datasync)
2015 {
2016         struct inode *vi = filp->f_mapping->host;
2017         int err, ret = 0;
2018
2019         ntfs_debug("Entering for inode 0x%lx.", vi->i_ino);
2020
2021         err = filemap_write_and_wait_range(vi->i_mapping, start, end);
2022         if (err)
2023                 return err;
2024         mutex_lock(&vi->i_mutex);
2025
2026         BUG_ON(S_ISDIR(vi->i_mode));
2027         if (!datasync || !NInoNonResident(NTFS_I(vi)))
2028                 ret = __ntfs_write_inode(vi, 1);
2029         write_inode_now(vi, !datasync);
2030         /*
2031          * NOTE: If we were to use mapping->private_list (see ext2 and
2032          * fs/buffer.c) for dirty blocks then we could optimize the below to be
2033          * sync_mapping_buffers(vi->i_mapping).
2034          */
2035         err = sync_blockdev(vi->i_sb->s_bdev);
2036         if (unlikely(err && !ret))
2037                 ret = err;
2038         if (likely(!ret))
2039                 ntfs_debug("Done.");
2040         else
2041                 ntfs_warning(vi->i_sb, "Failed to f%ssync inode 0x%lx.  Error "
2042                                 "%u.", datasync ? "data" : "", vi->i_ino, -ret);
2043         mutex_unlock(&vi->i_mutex);
2044         return ret;
2045 }
2046
2047 #endif /* NTFS_RW */
2048
2049 const struct file_operations ntfs_file_ops = {
2050         .llseek         = generic_file_llseek,
2051         .read           = new_sync_read,
2052         .read_iter      = generic_file_read_iter,
2053 #ifdef NTFS_RW
2054         .write          = new_sync_write,
2055         .write_iter     = ntfs_file_write_iter,
2056         .fsync          = ntfs_file_fsync,
2057 #endif /* NTFS_RW */
2058         .mmap           = generic_file_mmap,
2059         .open           = ntfs_file_open,
2060         .splice_read    = generic_file_splice_read,
2061 };
2062
2063 const struct inode_operations ntfs_file_inode_ops = {
2064 #ifdef NTFS_RW
2065         .setattr        = ntfs_setattr,
2066 #endif /* NTFS_RW */
2067 };
2068
2069 const struct file_operations ntfs_empty_file_ops = {};
2070
2071 const struct inode_operations ntfs_empty_inode_ops = {};