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