fs/ntfs/file.c

   1 /*
   2  * file.c - NTFS kernel file operations.  Part of the Linux-NTFS project.
   3  *
   4  * Copyright (c) 2001-2011 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/buffer_head.h>
  23 #include <linux/gfp.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/pagevec.h>
  26 #include <linux/sched.h>
  27 #include <linux/swap.h>
  28 #include <linux/uio.h>
  29 #include <linux/writeback.h>
  30 #include <linux/aio.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 /**
 332  * ntfs_fault_in_pages_readable -
 333  *
 334  * Fault a number of userspace pages into pagetables.
 335  *
 336  * Unlike include/linux/pagemap.h::fault_in_pages_readable(), this one copes
 337  * with more than two userspace pages as well as handling the single page case
 338  * elegantly.
 339  *
 340  * If you find this difficult to understand, then think of the while loop being
 341  * the following code, except that we do without the integer variable ret:
 342  *
 343  *      do {
 344  *              ret = __get_user(c, uaddr);
 345  *              uaddr += PAGE_SIZE;
 346  *      } while (!ret && uaddr < end);
 347  *
 348  * Note, the final __get_user() may well run out-of-bounds of the user buffer,
 349  * but _not_ out-of-bounds of the page the user buffer belongs to, and since
 350  * this is only a read and not a write, and since it is still in the same page,
 351  * it should not matter and this makes the code much simpler.
 352  */
 353 static inline void ntfs_fault_in_pages_readable(const char __user *uaddr,
 354                 int bytes)
 355 {
 356         const char __user *end;
 357         volatile char c;
 358
 359         /* Set @end to the first byte outside the last page we care about. */
 360         end = (const char __user*)PAGE_ALIGN((unsigned long)uaddr + bytes);
 361
 362         while (!__get_user(c, uaddr) && (uaddr += PAGE_SIZE, uaddr < end))
 363                 ;
 364 }
 365
 366 /**
 367  * ntfs_fault_in_pages_readable_iovec -
 368  *
 369  * Same as ntfs_fault_in_pages_readable() but operates on an array of iovecs.
 370  */
 371 static inline void ntfs_fault_in_pages_readable_iovec(const struct iovec *iov,
 372                 size_t iov_ofs, int bytes)
 373 {
 374         do {
 375                 const char __user *buf;
 376                 unsigned len;
 377
 378                 buf = iov->iov_base + iov_ofs;
 379                 len = iov->iov_len - iov_ofs;
 380                 if (len > bytes)
 381                         len = bytes;
 382                 ntfs_fault_in_pages_readable(buf, len);
 383                 bytes -= len;
 384                 iov++;
 385                 iov_ofs = 0;
 386         } while (bytes);
 387 }
 388
 389 /**
 390  * __ntfs_grab_cache_pages - obtain a number of locked pages
 391  * @mapping:    address space mapping from which to obtain page cache pages
 392  * @index:      starting index in @mapping at which to begin obtaining pages
 393  * @nr_pages:   number of page cache pages to obtain
 394  * @pages:      array of pages in which to return the obtained page cache pages
 395  * @cached_page: allocated but as yet unused page
 396  *
 397  * Obtain @nr_pages locked page cache pages from the mapping @mapping and
 398  * starting at index @index.
 399  *
 400  * If a page is newly created, add it to lru list
 401  *
 402  * Note, the page locks are obtained in ascending page index order.
 403  */
 404 static inline int __ntfs_grab_cache_pages(struct address_space *mapping,
 405                 pgoff_t index, const unsigned nr_pages, struct page **pages,
 406                 struct page **cached_page)
 407 {
 408         int err, nr;
 409
 410         BUG_ON(!nr_pages);
 411         err = nr = 0;
 412         do {
 413                 pages[nr] = find_lock_page(mapping, index);
 414                 if (!pages[nr]) {
 415                         if (!*cached_page) {
 416                                 *cached_page = page_cache_alloc(mapping);
 417                                 if (unlikely(!*cached_page)) {
 418                                         err = -ENOMEM;
 419                                         goto err_out;
 420                                 }
 421                         }
 422                         err = add_to_page_cache_lru(*cached_page, mapping, index,
 423                                         GFP_KERNEL);
 424                         if (unlikely(err)) {
 425                                 if (err == -EEXIST)
 426                                         continue;
 427                                 goto err_out;
 428                         }
 429                         pages[nr] = *cached_page;
 430                         *cached_page = NULL;
 431                 }
 432                 index++;
 433                 nr++;
 434         } while (nr < nr_pages);
 435 out:
 436         return err;
 437 err_out:
 438         while (nr > 0) {
 439                 unlock_page(pages[--nr]);
 440                 page_cache_release(pages[nr]);
 441         }
 442         goto out;
 443 }
 444
 445 static inline int ntfs_submit_bh_for_read(struct buffer_head *bh)
 446 {
 447         lock_buffer(bh);
 448         get_bh(bh);
 449         bh->b_end_io = end_buffer_read_sync;
 450         return submit_bh(READ, bh);
 451 }
 452
 453 /**
 454  * ntfs_prepare_pages_for_non_resident_write - prepare pages for receiving data
 455  * @pages:      array of destination pages
 456  * @nr_pages:   number of pages in @pages
 457  * @pos:        byte position in file at which the write begins
 458  * @bytes:      number of bytes to be written
 459  *
 460  * This is called for non-resident attributes from ntfs_file_buffered_write()
 461  * with i_mutex held on the inode (@pages[0]->mapping->host).  There are
 462  * @nr_pages pages in @pages which are locked but not kmap()ped.  The source
 463  * data has not yet been copied into the @pages.
 464  *
 465  * Need to fill any holes with actual clusters, allocate buffers if necessary,
 466  * ensure all the buffers are mapped, and bring uptodate any buffers that are
 467  * only partially being written to.
 468  *
 469  * If @nr_pages is greater than one, we are guaranteed that the cluster size is
 470  * greater than PAGE_CACHE_SIZE, that all pages in @pages are entirely inside
 471  * the same cluster and that they are the entirety of that cluster, and that
 472  * the cluster is sparse, i.e. we need to allocate a cluster to fill the hole.
 473  *
 474  * i_size is not to be modified yet.
 475  *
 476  * Return 0 on success or -errno on error.
 477  */
 478 static int ntfs_prepare_pages_for_non_resident_write(struct page **pages,
 479                 unsigned nr_pages, s64 pos, size_t bytes)
 480 {
 481         VCN vcn, highest_vcn = 0, cpos, cend, bh_cpos, bh_cend;
 482         LCN lcn;
 483         s64 bh_pos, vcn_len, end, initialized_size;
 484         sector_t lcn_block;
 485         struct page *page;
 486         struct inode *vi;
 487         ntfs_inode *ni, *base_ni = NULL;
 488         ntfs_volume *vol;
 489         runlist_element *rl, *rl2;
 490         struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
 491         ntfs_attr_search_ctx *ctx = NULL;
 492         MFT_RECORD *m = NULL;
 493         ATTR_RECORD *a = NULL;
 494         unsigned long flags;
 495         u32 attr_rec_len = 0;
 496         unsigned blocksize, u;
 497         int err, mp_size;
 498         bool rl_write_locked, was_hole, is_retry;
 499         unsigned char blocksize_bits;
 500         struct {
 501                 u8 runlist_merged:1;
 502                 u8 mft_attr_mapped:1;
 503                 u8 mp_rebuilt:1;
 504                 u8 attr_switched:1;
 505         } status = { 0, 0, 0, 0 };
 506
 507         BUG_ON(!nr_pages);
 508         BUG_ON(!pages);
 509         BUG_ON(!*pages);
 510         vi = pages[0]->mapping->host;
 511         ni = NTFS_I(vi);
 512         vol = ni->vol;
 513         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, start page "
 514                         "index 0x%lx, nr_pages 0x%x, pos 0x%llx, bytes 0x%zx.",
 515                         vi->i_ino, ni->type, pages[0]->index, nr_pages,
 516                         (long long)pos, bytes);
 517         blocksize = vol->sb->s_blocksize;
 518         blocksize_bits = vol->sb->s_blocksize_bits;
 519         u = 0;
 520         do {
 521                 page = pages[u];
 522                 BUG_ON(!page);
 523                 /*
 524                  * create_empty_buffers() will create uptodate/dirty buffers if
 525                  * the page is uptodate/dirty.
 526                  */
 527                 if (!page_has_buffers(page)) {
 528                         create_empty_buffers(page, blocksize, 0);
 529                         if (unlikely(!page_has_buffers(page)))
 530                                 return -ENOMEM;
 531                 }
 532         } while (++u < nr_pages);
 533         rl_write_locked = false;
 534         rl = NULL;
 535         err = 0;
 536         vcn = lcn = -1;
 537         vcn_len = 0;
 538         lcn_block = -1;
 539         was_hole = false;
 540         cpos = pos >> vol->cluster_size_bits;
 541         end = pos + bytes;
 542         cend = (end + vol->cluster_size - 1) >> vol->cluster_size_bits;
 543         /*
 544          * Loop over each page and for each page over each buffer.  Use goto to
 545          * reduce indentation.
 546          */
 547         u = 0;
 548 do_next_page:
 549         page = pages[u];
 550         bh_pos = (s64)page->index << PAGE_CACHE_SHIFT;
 551         bh = head = page_buffers(page);
 552         do {
 553                 VCN cdelta;
 554                 s64 bh_end;
 555                 unsigned bh_cofs;
 556
 557                 /* Clear buffer_new on all buffers to reinitialise state. */
 558                 if (buffer_new(bh))
 559                         clear_buffer_new(bh);
 560                 bh_end = bh_pos + blocksize;
 561                 bh_cpos = bh_pos >> vol->cluster_size_bits;
 562                 bh_cofs = bh_pos & vol->cluster_size_mask;
 563                 if (buffer_mapped(bh)) {
 564                         /*
 565                          * The buffer is already mapped.  If it is uptodate,
 566                          * ignore it.
 567                          */
 568                         if (buffer_uptodate(bh))
 569                                 continue;
 570                         /*
 571                          * The buffer is not uptodate.  If the page is uptodate
 572                          * set the buffer uptodate and otherwise ignore it.
 573                          */
 574                         if (PageUptodate(page)) {
 575                                 set_buffer_uptodate(bh);
 576                                 continue;
 577                         }
 578                         /*
 579                          * Neither the page nor the buffer are uptodate.  If
 580                          * the buffer is only partially being written to, we
 581                          * need to read it in before the write, i.e. now.
 582                          */
 583                         if ((bh_pos < pos && bh_end > pos) ||
 584                                         (bh_pos < end && bh_end > end)) {
 585                                 /*
 586                                  * If the buffer is fully or partially within
 587                                  * the initialized size, do an actual read.
 588                                  * Otherwise, simply zero the buffer.
 589                                  */
 590                                 read_lock_irqsave(&ni->size_lock, flags);
 591                                 initialized_size = ni->initialized_size;
 592                                 read_unlock_irqrestore(&ni->size_lock, flags);
 593                                 if (bh_pos < initialized_size) {
 594                                         ntfs_submit_bh_for_read(bh);
 595                                         *wait_bh++ = bh;
 596                                 } else {
 597                                         zero_user(page, bh_offset(bh),
 598                                                         blocksize);
 599                                         set_buffer_uptodate(bh);
 600                                 }
 601                         }
 602                         continue;
 603                 }
 604                 /* Unmapped buffer.  Need to map it. */
 605                 bh->b_bdev = vol->sb->s_bdev;
 606                 /*
 607                  * If the current buffer is in the same clusters as the map
 608                  * cache, there is no need to check the runlist again.  The
 609                  * map cache is made up of @vcn, which is the first cached file
 610                  * cluster, @vcn_len which is the number of cached file
 611                  * clusters, @lcn is the device cluster corresponding to @vcn,
 612                  * and @lcn_block is the block number corresponding to @lcn.
 613                  */
 614                 cdelta = bh_cpos - vcn;
 615                 if (likely(!cdelta || (cdelta > 0 && cdelta < vcn_len))) {
 616 map_buffer_cached:
 617                         BUG_ON(lcn < 0);
 618                         bh->b_blocknr = lcn_block +
 619                                         (cdelta << (vol->cluster_size_bits -
 620                                         blocksize_bits)) +
 621                                         (bh_cofs >> blocksize_bits);
 622                         set_buffer_mapped(bh);
 623                         /*
 624                          * If the page is uptodate so is the buffer.  If the
 625                          * buffer is fully outside the write, we ignore it if
 626                          * it was already allocated and we mark it dirty so it
 627                          * gets written out if we allocated it.  On the other
 628                          * hand, if we allocated the buffer but we are not
 629                          * marking it dirty we set buffer_new so we can do
 630                          * error recovery.
 631                          */
 632                         if (PageUptodate(page)) {
 633                                 if (!buffer_uptodate(bh))
 634                                         set_buffer_uptodate(bh);
 635                                 if (unlikely(was_hole)) {
 636                                         /* We allocated the buffer. */
 637                                         unmap_underlying_metadata(bh->b_bdev,
 638                                                         bh->b_blocknr);
 639                                         if (bh_end <= pos || bh_pos >= end)
 640                                                 mark_buffer_dirty(bh);
 641                                         else
 642                                                 set_buffer_new(bh);
 643                                 }
 644                                 continue;
 645                         }
 646                         /* Page is _not_ uptodate. */
 647                         if (likely(!was_hole)) {
 648                                 /*
 649                                  * Buffer was already allocated.  If it is not
 650                                  * uptodate and is only partially being written
 651                                  * to, we need to read it in before the write,
 652                                  * i.e. now.
 653                                  */
 654                                 if (!buffer_uptodate(bh) && bh_pos < end &&
 655                                                 bh_end > pos &&
 656                                                 (bh_pos < pos ||
 657                                                 bh_end > end)) {
 658                                         /*
 659                                          * If the buffer is fully or partially
 660                                          * within the initialized size, do an
 661                                          * actual read.  Otherwise, simply zero
 662                                          * the buffer.
 663                                          */
 664                                         read_lock_irqsave(&ni->size_lock,
 665                                                         flags);
 666                                         initialized_size = ni->initialized_size;
 667                                         read_unlock_irqrestore(&ni->size_lock,
 668                                                         flags);
 669                                         if (bh_pos < initialized_size) {
 670                                                 ntfs_submit_bh_for_read(bh);
 671                                                 *wait_bh++ = bh;
 672                                         } else {
 673                                                 zero_user(page, bh_offset(bh),
 674                                                                 blocksize);
 675                                                 set_buffer_uptodate(bh);
 676                                         }
 677                                 }
 678                                 continue;
 679                         }
 680                         /* We allocated the buffer. */
 681                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 682                         /*
 683                          * If the buffer is fully outside the write, zero it,
 684                          * set it uptodate, and mark it dirty so it gets
 685                          * written out.  If it is partially being written to,
 686                          * zero region surrounding the write but leave it to
 687                          * commit write to do anything else.  Finally, if the
 688                          * buffer is fully being overwritten, do nothing.
 689                          */
 690                         if (bh_end <= pos || bh_pos >= end) {
 691                                 if (!buffer_uptodate(bh)) {
 692                                         zero_user(page, bh_offset(bh),
 693                                                         blocksize);
 694                                         set_buffer_uptodate(bh);
 695                                 }
 696                                 mark_buffer_dirty(bh);
 697                                 continue;
 698                         }
 699                         set_buffer_new(bh);
 700                         if (!buffer_uptodate(bh) &&
 701                                         (bh_pos < pos || bh_end > end)) {
 702                                 u8 *kaddr;
 703                                 unsigned pofs;
 704
 705                                 kaddr = kmap_atomic(page);
 706                                 if (bh_pos < pos) {
 707                                         pofs = bh_pos & ~PAGE_CACHE_MASK;
 708                                         memset(kaddr + pofs, 0, pos - bh_pos);
 709                                 }
 710                                 if (bh_end > end) {
 711                                         pofs = end & ~PAGE_CACHE_MASK;
 712                                         memset(kaddr + pofs, 0, bh_end - end);
 713                                 }
 714                                 kunmap_atomic(kaddr);
 715                                 flush_dcache_page(page);
 716                         }
 717                         continue;
 718                 }
 719                 /*
 720                  * Slow path: this is the first buffer in the cluster.  If it
 721                  * is outside allocated size and is not uptodate, zero it and
 722                  * set it uptodate.
 723                  */
 724                 read_lock_irqsave(&ni->size_lock, flags);
 725                 initialized_size = ni->allocated_size;
 726                 read_unlock_irqrestore(&ni->size_lock, flags);
 727                 if (bh_pos > initialized_size) {
 728                         if (PageUptodate(page)) {
 729                                 if (!buffer_uptodate(bh))
 730                                         set_buffer_uptodate(bh);
 731                         } else if (!buffer_uptodate(bh)) {
 732                                 zero_user(page, bh_offset(bh), blocksize);
 733                                 set_buffer_uptodate(bh);
 734                         }
 735                         continue;
 736                 }
 737                 is_retry = false;
 738                 if (!rl) {
 739                         down_read(&ni->runlist.lock);
 740 retry_remap:
 741                         rl = ni->runlist.rl;
 742                 }
 743                 if (likely(rl != NULL)) {
 744                         /* Seek to element containing target cluster. */
 745                         while (rl->length && rl[1].vcn <= bh_cpos)
 746                                 rl++;
 747                         lcn = ntfs_rl_vcn_to_lcn(rl, bh_cpos);
 748                         if (likely(lcn >= 0)) {
 749                                 /*
 750                                  * Successful remap, setup the map cache and
 751                                  * use that to deal with the buffer.
 752                                  */
 753                                 was_hole = false;
 754                                 vcn = bh_cpos;
 755                                 vcn_len = rl[1].vcn - vcn;
 756                                 lcn_block = lcn << (vol->cluster_size_bits -
 757                                                 blocksize_bits);
 758                                 cdelta = 0;
 759                                 /*
 760                                  * If the number of remaining clusters touched
 761                                  * by the write is smaller or equal to the
 762                                  * number of cached clusters, unlock the
 763                                  * runlist as the map cache will be used from
 764                                  * now on.
 765                                  */
 766                                 if (likely(vcn + vcn_len >= cend)) {
 767                                         if (rl_write_locked) {
 768                                                 up_write(&ni->runlist.lock);
 769                                                 rl_write_locked = false;
 770                                         } else
 771                                                 up_read(&ni->runlist.lock);
 772                                         rl = NULL;
 773                                 }
 774                                 goto map_buffer_cached;
 775                         }
 776                 } else
 777                         lcn = LCN_RL_NOT_MAPPED;
 778                 /*
 779                  * If it is not a hole and not out of bounds, the runlist is
 780                  * probably unmapped so try to map it now.
 781                  */
 782                 if (unlikely(lcn != LCN_HOLE && lcn != LCN_ENOENT)) {
 783                         if (likely(!is_retry && lcn == LCN_RL_NOT_MAPPED)) {
 784                                 /* Attempt to map runlist. */
 785                                 if (!rl_write_locked) {
 786                                         /*
 787                                          * We need the runlist locked for
 788                                          * writing, so if it is locked for
 789                                          * reading relock it now and retry in
 790                                          * case it changed whilst we dropped
 791                                          * the lock.
 792                                          */
 793                                         up_read(&ni->runlist.lock);
 794                                         down_write(&ni->runlist.lock);
 795                                         rl_write_locked = true;
 796                                         goto retry_remap;
 797                                 }
 798                                 err = ntfs_map_runlist_nolock(ni, bh_cpos,
 799                                                 NULL);
 800                                 if (likely(!err)) {
 801                                         is_retry = true;
 802                                         goto retry_remap;
 803                                 }
 804                                 /*
 805                                  * If @vcn is out of bounds, pretend @lcn is
 806                                  * LCN_ENOENT.  As long as the buffer is out
 807                                  * of bounds this will work fine.
 808                                  */
 809                                 if (err == -ENOENT) {
 810                                         lcn = LCN_ENOENT;
 811                                         err = 0;
 812                                         goto rl_not_mapped_enoent;
 813                                 }
 814                         } else
 815                                 err = -EIO;
 816                         /* Failed to map the buffer, even after retrying. */
 817                         bh->b_blocknr = -1;
 818                         ntfs_error(vol->sb, "Failed to write to inode 0x%lx, "
 819                                         "attribute type 0x%x, vcn 0x%llx, "
 820                                         "vcn offset 0x%x, because its "
 821                                         "location on disk could not be "
 822                                         "determined%s (error code %i).",
 823                                         ni->mft_no, ni->type,
 824                                         (unsigned long long)bh_cpos,
 825                                         (unsigned)bh_pos &
 826                                         vol->cluster_size_mask,
 827                                         is_retry ? " even after retrying" : "",
 828                                         err);
 829                         break;
 830                 }
 831 rl_not_mapped_enoent:
 832                 /*
 833                  * The buffer is in a hole or out of bounds.  We need to fill
 834                  * the hole, unless the buffer is in a cluster which is not
 835                  * touched by the write, in which case we just leave the buffer
 836                  * unmapped.  This can only happen when the cluster size is
 837                  * less than the page cache size.
 838                  */
 839                 if (unlikely(vol->cluster_size < PAGE_CACHE_SIZE)) {
 840                         bh_cend = (bh_end + vol->cluster_size - 1) >>
 841                                         vol->cluster_size_bits;
 842                         if ((bh_cend <= cpos || bh_cpos >= cend)) {
 843                                 bh->b_blocknr = -1;
 844                                 /*
 845                                  * If the buffer is uptodate we skip it.  If it
 846                                  * is not but the page is uptodate, we can set
 847                                  * the buffer uptodate.  If the page is not
 848                                  * uptodate, we can clear the buffer and set it
 849                                  * uptodate.  Whether this is worthwhile is
 850                                  * debatable and this could be removed.
 851                                  */
 852                                 if (PageUptodate(page)) {
 853                                         if (!buffer_uptodate(bh))
 854                                                 set_buffer_uptodate(bh);
 855                                 } else if (!buffer_uptodate(bh)) {
 856                                         zero_user(page, bh_offset(bh),
 857                                                 blocksize);
 858                                         set_buffer_uptodate(bh);
 859                                 }
 860                                 continue;
 861                         }
 862                 }
 863                 /*
 864                  * Out of bounds buffer is invalid if it was not really out of
 865                  * bounds.
 866                  */
 867                 BUG_ON(lcn != LCN_HOLE);
 868                 /*
 869                  * We need the runlist locked for writing, so if it is locked
 870                  * for reading relock it now and retry in case it changed
 871                  * whilst we dropped the lock.
 872                  */
 873                 BUG_ON(!rl);
 874                 if (!rl_write_locked) {
 875                         up_read(&ni->runlist.lock);
 876                         down_write(&ni->runlist.lock);
 877                         rl_write_locked = true;
 878                         goto retry_remap;
 879                 }
 880                 /* Find the previous last allocated cluster. */
 881                 BUG_ON(rl->lcn != LCN_HOLE);
 882                 lcn = -1;
 883                 rl2 = rl;
 884                 while (--rl2 >= ni->runlist.rl) {
 885                         if (rl2->lcn >= 0) {
 886                                 lcn = rl2->lcn + rl2->length;
 887                                 break;
 888                         }
 889                 }
 890                 rl2 = ntfs_cluster_alloc(vol, bh_cpos, 1, lcn, DATA_ZONE,
 891                                 false);
 892                 if (IS_ERR(rl2)) {
 893                         err = PTR_ERR(rl2);
 894                         ntfs_debug("Failed to allocate cluster, error code %i.",
 895                                         err);
 896                         break;
 897                 }
 898                 lcn = rl2->lcn;
 899                 rl = ntfs_runlists_merge(ni->runlist.rl, rl2);
 900                 if (IS_ERR(rl)) {
 901                         err = PTR_ERR(rl);
 902                         if (err != -ENOMEM)
 903                                 err = -EIO;
 904                         if (ntfs_cluster_free_from_rl(vol, rl2)) {
 905                                 ntfs_error(vol->sb, "Failed to release "
 906                                                 "allocated cluster in error "
 907                                                 "code path.  Run chkdsk to "
 908                                                 "recover the lost cluster.");
 909                                 NVolSetErrors(vol);
 910                         }
 911                         ntfs_free(rl2);
 912                         break;
 913                 }
 914                 ni->runlist.rl = rl;
 915                 status.runlist_merged = 1;
 916                 ntfs_debug("Allocated cluster, lcn 0x%llx.",
 917                                 (unsigned long long)lcn);
 918                 /* Map and lock the mft record and get the attribute record. */
 919                 if (!NInoAttr(ni))
 920                         base_ni = ni;
 921                 else
 922                         base_ni = ni->ext.base_ntfs_ino;
 923                 m = map_mft_record(base_ni);
 924                 if (IS_ERR(m)) {
 925                         err = PTR_ERR(m);
 926                         break;
 927                 }
 928                 ctx = ntfs_attr_get_search_ctx(base_ni, m);
 929                 if (unlikely(!ctx)) {
 930                         err = -ENOMEM;
 931                         unmap_mft_record(base_ni);
 932                         break;
 933                 }
 934                 status.mft_attr_mapped = 1;
 935                 err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
 936                                 CASE_SENSITIVE, bh_cpos, NULL, 0, ctx);
 937                 if (unlikely(err)) {
 938                         if (err == -ENOENT)
 939                                 err = -EIO;
 940                         break;
 941                 }
 942                 m = ctx->mrec;
 943                 a = ctx->attr;
 944                 /*
 945                  * Find the runlist element with which the attribute extent
 946                  * starts.  Note, we cannot use the _attr_ version because we
 947                  * have mapped the mft record.  That is ok because we know the
 948                  * runlist fragment must be mapped already to have ever gotten
 949                  * here, so we can just use the _rl_ version.
 950                  */
 951                 vcn = sle64_to_cpu(a->data.non_resident.lowest_vcn);
 952                 rl2 = ntfs_rl_find_vcn_nolock(rl, vcn);
 953                 BUG_ON(!rl2);
 954                 BUG_ON(!rl2->length);
 955                 BUG_ON(rl2->lcn < LCN_HOLE);
 956                 highest_vcn = sle64_to_cpu(a->data.non_resident.highest_vcn);
 957                 /*
 958                  * If @highest_vcn is zero, calculate the real highest_vcn
 959                  * (which can really be zero).
 960                  */
 961                 if (!highest_vcn)
 962                         highest_vcn = (sle64_to_cpu(
 963                                         a->data.non_resident.allocated_size) >>
 964                                         vol->cluster_size_bits) - 1;
 965                 /*
 966                  * Determine the size of the mapping pairs array for the new
 967                  * extent, i.e. the old extent with the hole filled.
 968                  */
 969                 mp_size = ntfs_get_size_for_mapping_pairs(vol, rl2, vcn,
 970                                 highest_vcn);
 971                 if (unlikely(mp_size <= 0)) {
 972                         if (!(err = mp_size))
 973                                 err = -EIO;
 974                         ntfs_debug("Failed to get size for mapping pairs "
 975                                         "array, error code %i.", err);
 976                         break;
 977                 }
 978                 /*
 979                  * Resize the attribute record to fit the new mapping pairs
 980                  * array.
 981                  */
 982                 attr_rec_len = le32_to_cpu(a->length);
 983                 err = ntfs_attr_record_resize(m, a, mp_size + le16_to_cpu(
 984                                 a->data.non_resident.mapping_pairs_offset));
 985                 if (unlikely(err)) {
 986                         BUG_ON(err != -ENOSPC);
 987                         // TODO: Deal with this by using the current attribute
 988                         // and fill it with as much of the mapping pairs
 989                         // array as possible.  Then loop over each attribute
 990                         // extent rewriting the mapping pairs arrays as we go
 991                         // along and if when we reach the end we have not
 992                         // enough space, try to resize the last attribute
 993                         // extent and if even that fails, add a new attribute
 994                         // extent.
 995                         // We could also try to resize at each step in the hope
 996                         // that we will not need to rewrite every single extent.
 997                         // Note, we may need to decompress some extents to fill
 998                         // the runlist as we are walking the extents...
 999                         ntfs_error(vol->sb, "Not enough space in the mft "
1000                                         "record for the extended attribute "
1001                                         "record.  This case is not "
1002                                         "implemented yet.");
1003                         err = -EOPNOTSUPP;
1004                         break ;
1005                 }
1006                 status.mp_rebuilt = 1;
1007                 /*
1008                  * Generate the mapping pairs array directly into the attribute
1009                  * record.
1010                  */
1011                 err = ntfs_mapping_pairs_build(vol, (u8*)a + le16_to_cpu(
1012                                 a->data.non_resident.mapping_pairs_offset),
1013                                 mp_size, rl2, vcn, highest_vcn, NULL);
1014                 if (unlikely(err)) {
1015                         ntfs_error(vol->sb, "Cannot fill hole in inode 0x%lx, "
1016                                         "attribute type 0x%x, because building "
1017                                         "the mapping pairs failed with error "
1018                                         "code %i.", vi->i_ino,
1019                                         (unsigned)le32_to_cpu(ni->type), err);
1020                         err = -EIO;
1021                         break;
1022                 }
1023                 /* Update the highest_vcn but only if it was not set. */
1024                 if (unlikely(!a->data.non_resident.highest_vcn))
1025                         a->data.non_resident.highest_vcn =
1026                                         cpu_to_sle64(highest_vcn);
1027                 /*
1028                  * If the attribute is sparse/compressed, update the compressed
1029                  * size in the ntfs_inode structure and the attribute record.
1030                  */
1031                 if (likely(NInoSparse(ni) || NInoCompressed(ni))) {
1032                         /*
1033                          * If we are not in the first attribute extent, switch
1034                          * to it, but first ensure the changes will make it to
1035                          * disk later.
1036                          */
1037                         if (a->data.non_resident.lowest_vcn) {
1038                                 flush_dcache_mft_record_page(ctx->ntfs_ino);
1039                                 mark_mft_record_dirty(ctx->ntfs_ino);
1040                                 ntfs_attr_reinit_search_ctx(ctx);
1041                                 err = ntfs_attr_lookup(ni->type, ni->name,
1042                                                 ni->name_len, CASE_SENSITIVE,
1043                                                 0, NULL, 0, ctx);
1044                                 if (unlikely(err)) {
1045                                         status.attr_switched = 1;
1046                                         break;
1047                                 }
1048                                 /* @m is not used any more so do not set it. */
1049                                 a = ctx->attr;
1050                         }
1051                         write_lock_irqsave(&ni->size_lock, flags);
1052                         ni->itype.compressed.size += vol->cluster_size;
1053                         a->data.non_resident.compressed_size =
1054                                         cpu_to_sle64(ni->itype.compressed.size);
1055                         write_unlock_irqrestore(&ni->size_lock, flags);
1056                 }
1057                 /* Ensure the changes make it to disk. */
1058                 flush_dcache_mft_record_page(ctx->ntfs_ino);
1059                 mark_mft_record_dirty(ctx->ntfs_ino);
1060                 ntfs_attr_put_search_ctx(ctx);
1061                 unmap_mft_record(base_ni);
1062                 /* Successfully filled the hole. */
1063                 status.runlist_merged = 0;
1064                 status.mft_attr_mapped = 0;
1065                 status.mp_rebuilt = 0;
1066                 /* Setup the map cache and use that to deal with the buffer. */
1067                 was_hole = true;
1068                 vcn = bh_cpos;
1069                 vcn_len = 1;
1070                 lcn_block = lcn << (vol->cluster_size_bits - blocksize_bits);
1071                 cdelta = 0;
1072                 /*
1073                  * If the number of remaining clusters in the @pages is smaller
1074                  * or equal to the number of cached clusters, unlock the
1075                  * runlist as the map cache will be used from now on.
1076                  */
1077                 if (likely(vcn + vcn_len >= cend)) {
1078                         up_write(&ni->runlist.lock);
1079                         rl_write_locked = false;
1080                         rl = NULL;
1081                 }
1082                 goto map_buffer_cached;
1083         } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
1084         /* If there are no errors, do the next page. */
1085         if (likely(!err && ++u < nr_pages))
1086                 goto do_next_page;
1087         /* If there are no errors, release the runlist lock if we took it. */
1088         if (likely(!err)) {
1089                 if (unlikely(rl_write_locked)) {
1090                         up_write(&ni->runlist.lock);
1091                         rl_write_locked = false;
1092                 } else if (unlikely(rl))
1093                         up_read(&ni->runlist.lock);
1094                 rl = NULL;
1095         }
1096         /* If we issued read requests, let them complete. */
1097         read_lock_irqsave(&ni->size_lock, flags);
1098         initialized_size = ni->initialized_size;
1099         read_unlock_irqrestore(&ni->size_lock, flags);
1100         while (wait_bh > wait) {
1101                 bh = *--wait_bh;
1102                 wait_on_buffer(bh);
1103                 if (likely(buffer_uptodate(bh))) {
1104                         page = bh->b_page;
1105                         bh_pos = ((s64)page->index << PAGE_CACHE_SHIFT) +
1106                                         bh_offset(bh);
1107                         /*
1108                          * If the buffer overflows the initialized size, need
1109                          * to zero the overflowing region.
1110                          */
1111                         if (unlikely(bh_pos + blocksize > initialized_size)) {
1112                                 int ofs = 0;
1113
1114                                 if (likely(bh_pos < initialized_size))
1115                                         ofs = initialized_size - bh_pos;
1116                                 zero_user_segment(page, bh_offset(bh) + ofs,
1117                                                 blocksize);
1118                         }
1119                 } else /* if (unlikely(!buffer_uptodate(bh))) */
1120                         err = -EIO;
1121         }
1122         if (likely(!err)) {
1123                 /* Clear buffer_new on all buffers. */
1124                 u = 0;
1125                 do {
1126                         bh = head = page_buffers(pages[u]);
1127                         do {
1128                                 if (buffer_new(bh))
1129                                         clear_buffer_new(bh);
1130                         } while ((bh = bh->b_this_page) != head);
1131                 } while (++u < nr_pages);
1132                 ntfs_debug("Done.");
1133                 return err;
1134         }
1135         if (status.attr_switched) {
1136                 /* Get back to the attribute extent we modified. */
1137                 ntfs_attr_reinit_search_ctx(ctx);
1138                 if (ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1139                                 CASE_SENSITIVE, bh_cpos, NULL, 0, ctx)) {
1140                         ntfs_error(vol->sb, "Failed to find required "
1141                                         "attribute extent of attribute in "
1142                                         "error code path.  Run chkdsk to "
1143                                         "recover.");
1144                         write_lock_irqsave(&ni->size_lock, flags);
1145                         ni->itype.compressed.size += vol->cluster_size;
1146                         write_unlock_irqrestore(&ni->size_lock, flags);
1147                         flush_dcache_mft_record_page(ctx->ntfs_ino);
1148                         mark_mft_record_dirty(ctx->ntfs_ino);
1149                         /*
1150                          * The only thing that is now wrong is the compressed
1151                          * size of the base attribute extent which chkdsk
1152                          * should be able to fix.
1153                          */
1154                         NVolSetErrors(vol);
1155                 } else {
1156                         m = ctx->mrec;
1157                         a = ctx->attr;
1158                         status.attr_switched = 0;
1159                 }
1160         }
1161         /*
1162          * If the runlist has been modified, need to restore it by punching a
1163          * hole into it and we then need to deallocate the on-disk cluster as
1164          * well.  Note, we only modify the runlist if we are able to generate a
1165          * new mapping pairs array, i.e. only when the mapped attribute extent
1166          * is not switched.
1167          */
1168         if (status.runlist_merged && !status.attr_switched) {
1169                 BUG_ON(!rl_write_locked);
1170                 /* Make the file cluster we allocated sparse in the runlist. */
1171                 if (ntfs_rl_punch_nolock(vol, &ni->runlist, bh_cpos, 1)) {
1172                         ntfs_error(vol->sb, "Failed to punch hole into "
1173                                         "attribute runlist in error code "
1174                                         "path.  Run chkdsk to recover the "
1175                                         "lost cluster.");
1176                         NVolSetErrors(vol);
1177                 } else /* if (success) */ {
1178                         status.runlist_merged = 0;
1179                         /*
1180                          * Deallocate the on-disk cluster we allocated but only
1181                          * if we succeeded in punching its vcn out of the
1182                          * runlist.
1183                          */
1184                         down_write(&vol->lcnbmp_lock);
1185                         if (ntfs_bitmap_clear_bit(vol->lcnbmp_ino, lcn)) {
1186                                 ntfs_error(vol->sb, "Failed to release "
1187                                                 "allocated cluster in error "
1188                                                 "code path.  Run chkdsk to "
1189                                                 "recover the lost cluster.");
1190                                 NVolSetErrors(vol);
1191                         }
1192                         up_write(&vol->lcnbmp_lock);
1193                 }
1194         }
1195         /*
1196          * Resize the attribute record to its old size and rebuild the mapping
1197          * pairs array.  Note, we only can do this if the runlist has been
1198          * restored to its old state which also implies that the mapped
1199          * attribute extent is not switched.
1200          */
1201         if (status.mp_rebuilt && !status.runlist_merged) {
1202                 if (ntfs_attr_record_resize(m, a, attr_rec_len)) {
1203                         ntfs_error(vol->sb, "Failed to restore attribute "
1204                                         "record in error code path.  Run "
1205                                         "chkdsk to recover.");
1206                         NVolSetErrors(vol);
1207                 } else /* if (success) */ {
1208                         if (ntfs_mapping_pairs_build(vol, (u8*)a +
1209                                         le16_to_cpu(a->data.non_resident.
1210                                         mapping_pairs_offset), attr_rec_len -
1211                                         le16_to_cpu(a->data.non_resident.
1212                                         mapping_pairs_offset), ni->runlist.rl,
1213                                         vcn, highest_vcn, NULL)) {
1214                                 ntfs_error(vol->sb, "Failed to restore "
1215                                                 "mapping pairs array in error "
1216                                                 "code path.  Run chkdsk to "
1217                                                 "recover.");
1218                                 NVolSetErrors(vol);
1219                         }
1220                         flush_dcache_mft_record_page(ctx->ntfs_ino);
1221                         mark_mft_record_dirty(ctx->ntfs_ino);
1222                 }
1223         }
1224         /* Release the mft record and the attribute. */
1225         if (status.mft_attr_mapped) {
1226                 ntfs_attr_put_search_ctx(ctx);
1227                 unmap_mft_record(base_ni);
1228         }
1229         /* Release the runlist lock. */
1230         if (rl_write_locked)
1231                 up_write(&ni->runlist.lock);
1232         else if (rl)
1233                 up_read(&ni->runlist.lock);
1234         /*
1235          * Zero out any newly allocated blocks to avoid exposing stale data.
1236          * If BH_New is set, we know that the block was newly allocated above
1237          * and that it has not been fully zeroed and marked dirty yet.
1238          */
1239         nr_pages = u;
1240         u = 0;
1241         end = bh_cpos << vol->cluster_size_bits;
1242         do {
1243                 page = pages[u];
1244                 bh = head = page_buffers(page);
1245                 do {
1246                         if (u == nr_pages &&
1247                                         ((s64)page->index << PAGE_CACHE_SHIFT) +
1248                                         bh_offset(bh) >= end)
1249                                 break;
1250                         if (!buffer_new(bh))
1251                                 continue;
1252                         clear_buffer_new(bh);
1253                         if (!buffer_uptodate(bh)) {
1254                                 if (PageUptodate(page))
1255                                         set_buffer_uptodate(bh);
1256                                 else {
1257                                         zero_user(page, bh_offset(bh),
1258                                                         blocksize);
1259                                         set_buffer_uptodate(bh);
1260                                 }
1261                         }
1262                         mark_buffer_dirty(bh);
1263                 } while ((bh = bh->b_this_page) != head);
1264         } while (++u <= nr_pages);
1265         ntfs_error(vol->sb, "Failed.  Returning error code %i.", err);
1266         return err;
1267 }
1268
1269 /*
1270  * Copy as much as we can into the pages and return the number of bytes which
1271  * were successfully copied.  If a fault is encountered then clear the pages
1272  * out to (ofs + bytes) and return the number of bytes which were copied.
1273  */
1274 static inline size_t ntfs_copy_from_user(struct page **pages,
1275                 unsigned nr_pages, unsigned ofs, const char __user *buf,
1276                 size_t bytes)
1277 {
1278         struct page **last_page = pages + nr_pages;
1279         char *addr;
1280         size_t total = 0;
1281         unsigned len;
1282         int left;
1283
1284         do {
1285                 len = PAGE_CACHE_SIZE - ofs;
1286                 if (len > bytes)
1287                         len = bytes;
1288                 addr = kmap_atomic(*pages);
1289                 left = __copy_from_user_inatomic(addr + ofs, buf, len);
1290                 kunmap_atomic(addr);
1291                 if (unlikely(left)) {
1292                         /* Do it the slow way. */
1293                         addr = kmap(*pages);
1294                         left = __copy_from_user(addr + ofs, buf, len);
1295                         kunmap(*pages);
1296                         if (unlikely(left))
1297                                 goto err_out;
1298                 }
1299                 total += len;
1300                 bytes -= len;
1301                 if (!bytes)
1302                         break;
1303                 buf += len;
1304                 ofs = 0;
1305         } while (++pages < last_page);
1306 out:
1307         return total;
1308 err_out:
1309         total += len - left;
1310         /* Zero the rest of the target like __copy_from_user(). */
1311         while (++pages < last_page) {
1312                 bytes -= len;
1313                 if (!bytes)
1314                         break;
1315                 len = PAGE_CACHE_SIZE;
1316                 if (len > bytes)
1317                         len = bytes;
1318                 zero_user(*pages, 0, len);
1319         }
1320         goto out;
1321 }
1322
1323 static size_t __ntfs_copy_from_user_iovec_inatomic(char *vaddr,
1324                 const struct iovec *iov, size_t iov_ofs, size_t bytes)
1325 {
1326         size_t total = 0;
1327
1328         while (1) {
1329                 const char __user *buf = iov->iov_base + iov_ofs;
1330                 unsigned len;
1331                 size_t left;
1332
1333                 len = iov->iov_len - iov_ofs;
1334                 if (len > bytes)
1335                         len = bytes;
1336                 left = __copy_from_user_inatomic(vaddr, buf, len);
1337                 total += len;
1338                 bytes -= len;
1339                 vaddr += len;
1340                 if (unlikely(left)) {
1341                         total -= left;
1342                         break;
1343                 }
1344                 if (!bytes)
1345                         break;
1346                 iov++;
1347                 iov_ofs = 0;
1348         }
1349         return total;
1350 }
1351
1352 static inline void ntfs_set_next_iovec(const struct iovec **iovp,
1353                 size_t *iov_ofsp, size_t bytes)
1354 {
1355         const struct iovec *iov = *iovp;
1356         size_t iov_ofs = *iov_ofsp;
1357
1358         while (bytes) {
1359                 unsigned len;
1360
1361                 len = iov->iov_len - iov_ofs;
1362                 if (len > bytes)
1363                         len = bytes;
1364                 bytes -= len;
1365                 iov_ofs += len;
1366                 if (iov->iov_len == iov_ofs) {
1367                         iov++;
1368                         iov_ofs = 0;
1369                 }
1370         }
1371         *iovp = iov;
1372         *iov_ofsp = iov_ofs;
1373 }
1374
1375 /*
1376  * This has the same side-effects and return value as ntfs_copy_from_user().
1377  * The difference is that on a fault we need to memset the remainder of the
1378  * pages (out to offset + bytes), to emulate ntfs_copy_from_user()'s
1379  * single-segment behaviour.
1380  *
1381  * We call the same helper (__ntfs_copy_from_user_iovec_inatomic()) both when
1382  * atomic and when not atomic.  This is ok because it calls
1383  * __copy_from_user_inatomic() and it is ok to call this when non-atomic.  In
1384  * fact, the only difference between __copy_from_user_inatomic() and
1385  * __copy_from_user() is that the latter calls might_sleep() and the former
1386  * should not zero the tail of the buffer on error.  And on many architectures
1387  * __copy_from_user_inatomic() is just defined to __copy_from_user() so it
1388  * makes no difference at all on those architectures.
1389  */
1390 static inline size_t ntfs_copy_from_user_iovec(struct page **pages,
1391                 unsigned nr_pages, unsigned ofs, const struct iovec **iov,
1392                 size_t *iov_ofs, size_t bytes)
1393 {
1394         struct page **last_page = pages + nr_pages;
1395         char *addr;
1396         size_t copied, len, total = 0;
1397
1398         do {
1399                 len = PAGE_CACHE_SIZE - ofs;
1400                 if (len > bytes)
1401                         len = bytes;
1402                 addr = kmap_atomic(*pages);
1403                 copied = __ntfs_copy_from_user_iovec_inatomic(addr + ofs,
1404                                 *iov, *iov_ofs, len);
1405                 kunmap_atomic(addr);
1406                 if (unlikely(copied != len)) {
1407                         /* Do it the slow way. */
1408                         addr = kmap(*pages);
1409                         copied = __ntfs_copy_from_user_iovec_inatomic(addr +
1410                                         ofs, *iov, *iov_ofs, len);
1411                         if (unlikely(copied != len))
1412                                 goto err_out;
1413                         kunmap(*pages);
1414                 }
1415                 total += len;
1416                 ntfs_set_next_iovec(iov, iov_ofs, len);
1417                 bytes -= len;
1418                 if (!bytes)
1419                         break;
1420                 ofs = 0;
1421         } while (++pages < last_page);
1422 out:
1423         return total;
1424 err_out:
1425         BUG_ON(copied > len);
1426         /* Zero the rest of the target like __copy_from_user(). */
1427         memset(addr + ofs + copied, 0, len - copied);
1428         kunmap(*pages);
1429         total += copied;
1430         ntfs_set_next_iovec(iov, iov_ofs, copied);
1431         while (++pages < last_page) {
1432                 bytes -= len;
1433                 if (!bytes)
1434                         break;
1435                 len = PAGE_CACHE_SIZE;
1436                 if (len > bytes)
1437                         len = bytes;
1438                 zero_user(*pages, 0, len);
1439         }
1440         goto out;
1441 }
1442
1443 static inline void ntfs_flush_dcache_pages(struct page **pages,
1444                 unsigned nr_pages)
1445 {
1446         BUG_ON(!nr_pages);
1447         /*
1448          * Warning: Do not do the decrement at the same time as the call to
1449          * flush_dcache_page() because it is a NULL macro on i386 and hence the
1450          * decrement never happens so the loop never terminates.
1451          */
1452         do {
1453                 --nr_pages;
1454                 flush_dcache_page(pages[nr_pages]);
1455         } while (nr_pages > 0);
1456 }
1457
1458 /**
1459  * ntfs_commit_pages_after_non_resident_write - commit the received data
1460  * @pages:      array of destination pages
1461  * @nr_pages:   number of pages in @pages
1462  * @pos:        byte position in file at which the write begins
1463  * @bytes:      number of bytes to be written
1464  *
1465  * See description of ntfs_commit_pages_after_write(), below.
1466  */
1467 static inline int ntfs_commit_pages_after_non_resident_write(
1468                 struct page **pages, const unsigned nr_pages,
1469                 s64 pos, size_t bytes)
1470 {
1471         s64 end, initialized_size;
1472         struct inode *vi;
1473         ntfs_inode *ni, *base_ni;
1474         struct buffer_head *bh, *head;
1475         ntfs_attr_search_ctx *ctx;
1476         MFT_RECORD *m;
1477         ATTR_RECORD *a;
1478         unsigned long flags;
1479         unsigned blocksize, u;
1480         int err;
1481
1482         vi = pages[0]->mapping->host;
1483         ni = NTFS_I(vi);
1484         blocksize = vi->i_sb->s_blocksize;
1485         end = pos + bytes;
1486         u = 0;
1487         do {
1488                 s64 bh_pos;
1489                 struct page *page;
1490                 bool partial;
1491
1492                 page = pages[u];
1493                 bh_pos = (s64)page->index << PAGE_CACHE_SHIFT;
1494                 bh = head = page_buffers(page);
1495                 partial = false;
1496                 do {
1497                         s64 bh_end;
1498
1499                         bh_end = bh_pos + blocksize;
1500                         if (bh_end <= pos || bh_pos >= end) {
1501                                 if (!buffer_uptodate(bh))
1502                                         partial = true;
1503                         } else {
1504                                 set_buffer_uptodate(bh);
1505                                 mark_buffer_dirty(bh);
1506                         }
1507                 } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
1508                 /*
1509                  * If all buffers are now uptodate but the page is not, set the
1510                  * page uptodate.
1511                  */
1512                 if (!partial && !PageUptodate(page))
1513                         SetPageUptodate(page);
1514         } while (++u < nr_pages);
1515         /*
1516          * Finally, if we do not need to update initialized_size or i_size we
1517          * are finished.
1518          */
1519         read_lock_irqsave(&ni->size_lock, flags);
1520         initialized_size = ni->initialized_size;
1521         read_unlock_irqrestore(&ni->size_lock, flags);
1522         if (end <= initialized_size) {
1523                 ntfs_debug("Done.");
1524                 return 0;
1525         }
1526         /*
1527          * Update initialized_size/i_size as appropriate, both in the inode and
1528          * the mft record.
1529          */
1530         if (!NInoAttr(ni))
1531                 base_ni = ni;
1532         else
1533                 base_ni = ni->ext.base_ntfs_ino;
1534         /* Map, pin, and lock the mft record. */
1535         m = map_mft_record(base_ni);
1536         if (IS_ERR(m)) {
1537                 err = PTR_ERR(m);
1538                 m = NULL;
1539                 ctx = NULL;
1540                 goto err_out;
1541         }
1542         BUG_ON(!NInoNonResident(ni));
1543         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1544         if (unlikely(!ctx)) {
1545                 err = -ENOMEM;
1546                 goto err_out;
1547         }
1548         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1549                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1550         if (unlikely(err)) {
1551                 if (err == -ENOENT)
1552                         err = -EIO;
1553                 goto err_out;
1554         }
1555         a = ctx->attr;
1556         BUG_ON(!a->non_resident);
1557         write_lock_irqsave(&ni->size_lock, flags);
1558         BUG_ON(end > ni->allocated_size);
1559         ni->initialized_size = end;
1560         a->data.non_resident.initialized_size = cpu_to_sle64(end);
1561         if (end > i_size_read(vi)) {
1562                 i_size_write(vi, end);
1563                 a->data.non_resident.data_size =
1564                                 a->data.non_resident.initialized_size;
1565         }
1566         write_unlock_irqrestore(&ni->size_lock, flags);
1567         /* Mark the mft record dirty, so it gets written back. */
1568         flush_dcache_mft_record_page(ctx->ntfs_ino);
1569         mark_mft_record_dirty(ctx->ntfs_ino);
1570         ntfs_attr_put_search_ctx(ctx);
1571         unmap_mft_record(base_ni);
1572         ntfs_debug("Done.");
1573         return 0;
1574 err_out:
1575         if (ctx)
1576                 ntfs_attr_put_search_ctx(ctx);
1577         if (m)
1578                 unmap_mft_record(base_ni);
1579         ntfs_error(vi->i_sb, "Failed to update initialized_size/i_size (error "
1580                         "code %i).", err);
1581         if (err != -ENOMEM)
1582                 NVolSetErrors(ni->vol);
1583         return err;
1584 }
1585
1586 /**
1587  * ntfs_commit_pages_after_write - commit the received data
1588  * @pages:      array of destination pages
1589  * @nr_pages:   number of pages in @pages
1590  * @pos:        byte position in file at which the write begins
1591  * @bytes:      number of bytes to be written
1592  *
1593  * This is called from ntfs_file_buffered_write() with i_mutex held on the inode
1594  * (@pages[0]->mapping->host).  There are @nr_pages pages in @pages which are
1595  * locked but not kmap()ped.  The source data has already been copied into the
1596  * @page.  ntfs_prepare_pages_for_non_resident_write() has been called before
1597  * the data was copied (for non-resident attributes only) and it returned
1598  * success.
1599  *
1600  * Need to set uptodate and mark dirty all buffers within the boundary of the
1601  * write.  If all buffers in a page are uptodate we set the page uptodate, too.
1602  *
1603  * Setting the buffers dirty ensures that they get written out later when
1604  * ntfs_writepage() is invoked by the VM.
1605  *
1606  * Finally, we need to update i_size and initialized_size as appropriate both
1607  * in the inode and the mft record.
1608  *
1609  * This is modelled after fs/buffer.c::generic_commit_write(), which marks
1610  * buffers uptodate and dirty, sets the page uptodate if all buffers in the
1611  * page are uptodate, and updates i_size if the end of io is beyond i_size.  In
1612  * that case, it also marks the inode dirty.
1613  *
1614  * If things have gone as outlined in
1615  * ntfs_prepare_pages_for_non_resident_write(), we do not need to do any page
1616  * content modifications here for non-resident attributes.  For resident
1617  * attributes we need to do the uptodate bringing here which we combine with
1618  * the copying into the mft record which means we save one atomic kmap.
1619  *
1620  * Return 0 on success or -errno on error.
1621  */
1622 static int ntfs_commit_pages_after_write(struct page **pages,
1623                 const unsigned nr_pages, s64 pos, size_t bytes)
1624 {
1625         s64 end, initialized_size;
1626         loff_t i_size;
1627         struct inode *vi;
1628         ntfs_inode *ni, *base_ni;
1629         struct page *page;
1630         ntfs_attr_search_ctx *ctx;
1631         MFT_RECORD *m;
1632         ATTR_RECORD *a;
1633         char *kattr, *kaddr;
1634         unsigned long flags;
1635         u32 attr_len;
1636         int err;
1637
1638         BUG_ON(!nr_pages);
1639         BUG_ON(!pages);
1640         page = pages[0];
1641         BUG_ON(!page);
1642         vi = page->mapping->host;
1643         ni = NTFS_I(vi);
1644         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, start page "
1645                         "index 0x%lx, nr_pages 0x%x, pos 0x%llx, bytes 0x%zx.",
1646                         vi->i_ino, ni->type, page->index, nr_pages,
1647                         (long long)pos, bytes);
1648         if (NInoNonResident(ni))
1649                 return ntfs_commit_pages_after_non_resident_write(pages,
1650                                 nr_pages, pos, bytes);
1651         BUG_ON(nr_pages > 1);
1652         /*
1653          * Attribute is resident, implying it is not compressed, encrypted, or
1654          * sparse.
1655          */
1656         if (!NInoAttr(ni))
1657                 base_ni = ni;
1658         else
1659                 base_ni = ni->ext.base_ntfs_ino;
1660         BUG_ON(NInoNonResident(ni));
1661         /* Map, pin, and lock the mft record. */
1662         m = map_mft_record(base_ni);
1663         if (IS_ERR(m)) {
1664                 err = PTR_ERR(m);
1665                 m = NULL;
1666                 ctx = NULL;
1667                 goto err_out;
1668         }
1669         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1670         if (unlikely(!ctx)) {
1671                 err = -ENOMEM;
1672                 goto err_out;
1673         }
1674         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1675                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1676         if (unlikely(err)) {
1677                 if (err == -ENOENT)
1678                         err = -EIO;
1679                 goto err_out;
1680         }
1681         a = ctx->attr;
1682         BUG_ON(a->non_resident);
1683         /* The total length of the attribute value. */
1684         attr_len = le32_to_cpu(a->data.resident.value_length);
1685         i_size = i_size_read(vi);
1686         BUG_ON(attr_len != i_size);
1687         BUG_ON(pos > attr_len);
1688         end = pos + bytes;
1689         BUG_ON(end > le32_to_cpu(a->length) -
1690                         le16_to_cpu(a->data.resident.value_offset));
1691         kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
1692         kaddr = kmap_atomic(page);
1693         /* Copy the received data from the page to the mft record. */
1694         memcpy(kattr + pos, kaddr + pos, bytes);
1695         /* Update the attribute length if necessary. */
1696         if (end > attr_len) {
1697                 attr_len = end;
1698                 a->data.resident.value_length = cpu_to_le32(attr_len);
1699         }
1700         /*
1701          * If the page is not uptodate, bring the out of bounds area(s)
1702          * uptodate by copying data from the mft record to the page.
1703          */
1704         if (!PageUptodate(page)) {
1705                 if (pos > 0)
1706                         memcpy(kaddr, kattr, pos);
1707                 if (end < attr_len)
1708                         memcpy(kaddr + end, kattr + end, attr_len - end);
1709                 /* Zero the region outside the end of the attribute value. */
1710                 memset(kaddr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
1711                 flush_dcache_page(page);
1712                 SetPageUptodate(page);
1713         }
1714         kunmap_atomic(kaddr);
1715         /* Update initialized_size/i_size if necessary. */
1716         read_lock_irqsave(&ni->size_lock, flags);
1717         initialized_size = ni->initialized_size;
1718         BUG_ON(end > ni->allocated_size);
1719         read_unlock_irqrestore(&ni->size_lock, flags);
1720         BUG_ON(initialized_size != i_size);
1721         if (end > initialized_size) {
1722                 write_lock_irqsave(&ni->size_lock, flags);
1723                 ni->initialized_size = end;
1724                 i_size_write(vi, end);
1725                 write_unlock_irqrestore(&ni->size_lock, flags);
1726         }
1727         /* Mark the mft record dirty, so it gets written back. */
1728         flush_dcache_mft_record_page(ctx->ntfs_ino);
1729         mark_mft_record_dirty(ctx->ntfs_ino);
1730         ntfs_attr_put_search_ctx(ctx);
1731         unmap_mft_record(base_ni);
1732         ntfs_debug("Done.");
1733         return 0;
1734 err_out:
1735         if (err == -ENOMEM) {
1736                 ntfs_warning(vi->i_sb, "Error allocating memory required to "
1737                                 "commit the write.");
1738                 if (PageUptodate(page)) {
1739                         ntfs_warning(vi->i_sb, "Page is uptodate, setting "
1740                                         "dirty so the write will be retried "
1741                                         "later on by the VM.");
1742                         /*
1743                          * Put the page on mapping->dirty_pages, but leave its
1744                          * buffers' dirty state as-is.
1745                          */
1746                         __set_page_dirty_nobuffers(page);
1747                         err = 0;
1748                 } else
1749                         ntfs_error(vi->i_sb, "Page is not uptodate.  Written "
1750                                         "data has been lost.");
1751         } else {
1752                 ntfs_error(vi->i_sb, "Resident attribute commit write failed "
1753                                 "with error %i.", err);
1754                 NVolSetErrors(ni->vol);
1755         }
1756         if (ctx)
1757                 ntfs_attr_put_search_ctx(ctx);
1758         if (m)
1759                 unmap_mft_record(base_ni);
1760         return err;
1761 }
1762
1763 static void ntfs_write_failed(struct address_space *mapping, loff_t to)
1764 {
1765         struct inode *inode = mapping->host;
1766
1767         if (to > inode->i_size) {
1768                 truncate_pagecache(inode, inode->i_size);
1769                 ntfs_truncate_vfs(inode);
1770         }
1771 }
1772
1773 /**
1774  * ntfs_file_buffered_write -
1775  *
1776  * Locking: The vfs is holding ->i_mutex on the inode.
1777  */
1778 static ssize_t ntfs_file_buffered_write(struct kiocb *iocb,
1779                 const struct iovec *iov, unsigned long nr_segs,
1780                 loff_t pos, loff_t *ppos, size_t count)
1781 {
1782         struct file *file = iocb->ki_filp;
1783         struct address_space *mapping = file->f_mapping;
1784         struct inode *vi = mapping->host;
1785         ntfs_inode *ni = NTFS_I(vi);
1786         ntfs_volume *vol = ni->vol;
1787         struct page *pages[NTFS_MAX_PAGES_PER_CLUSTER];
1788         struct page *cached_page = NULL;
1789         char __user *buf = NULL;
1790         s64 end, ll;
1791         VCN last_vcn;
1792         LCN lcn;
1793         unsigned long flags;
1794         size_t bytes, iov_ofs = 0;      /* Offset in the current iovec. */
1795         ssize_t status, written;
1796         unsigned nr_pages;
1797         int err;
1798
1799         ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, "
1800                         "pos 0x%llx, count 0x%lx.",
1801                         vi->i_ino, (unsigned)le32_to_cpu(ni->type),
1802                         (unsigned long long)pos, (unsigned long)count);
1803         if (unlikely(!count))
1804                 return 0;
1805         BUG_ON(NInoMstProtected(ni));
1806         /*
1807          * If the attribute is not an index root and it is encrypted or
1808          * compressed, we cannot write to it yet.  Note we need to check for
1809          * AT_INDEX_ALLOCATION since this is the type of both directory and
1810          * index inodes.
1811          */
1812         if (ni->type != AT_INDEX_ALLOCATION) {
1813                 /* If file is encrypted, deny access, just like NT4. */
1814                 if (NInoEncrypted(ni)) {
1815                         /*
1816                          * Reminder for later: Encrypted files are _always_
1817                          * non-resident so that the content can always be
1818                          * encrypted.
1819                          */
1820                         ntfs_debug("Denying write access to encrypted file.");
1821                         return -EACCES;
1822                 }
1823                 if (NInoCompressed(ni)) {
1824                         /* Only unnamed $DATA attribute can be compressed. */
1825                         BUG_ON(ni->type != AT_DATA);
1826                         BUG_ON(ni->name_len);
1827                         /*
1828                          * Reminder for later: If resident, the data is not
1829                          * actually compressed.  Only on the switch to non-
1830                          * resident does compression kick in.  This is in
1831                          * contrast to encrypted files (see above).
1832                          */
1833                         ntfs_error(vi->i_sb, "Writing to compressed files is "
1834                                         "not implemented yet.  Sorry.");
1835                         return -EOPNOTSUPP;
1836                 }
1837         }
1838         /*
1839          * If a previous ntfs_truncate() failed, repeat it and abort if it
1840          * fails again.
1841          */
1842         if (unlikely(NInoTruncateFailed(ni))) {
1843                 inode_dio_wait(vi);
1844                 err = ntfs_truncate(vi);
1845                 if (err || NInoTruncateFailed(ni)) {
1846                         if (!err)
1847                                 err = -EIO;
1848                         ntfs_error(vol->sb, "Cannot perform write to inode "
1849                                         "0x%lx, attribute type 0x%x, because "
1850                                         "ntfs_truncate() failed (error code "
1851                                         "%i).", vi->i_ino,
1852                                         (unsigned)le32_to_cpu(ni->type), err);
1853                         return err;
1854                 }
1855         }
1856         /* The first byte after the write. */
1857         end = pos + count;
1858         /*
1859          * If the write goes beyond the allocated size, extend the allocation
1860          * to cover the whole of the write, rounded up to the nearest cluster.
1861          */
1862         read_lock_irqsave(&ni->size_lock, flags);
1863         ll = ni->allocated_size;
1864         read_unlock_irqrestore(&ni->size_lock, flags);
1865         if (end > ll) {
1866                 /* Extend the allocation without changing the data size. */
1867                 ll = ntfs_attr_extend_allocation(ni, end, -1, pos);
1868                 if (likely(ll >= 0)) {
1869                         BUG_ON(pos >= ll);
1870                         /* If the extension was partial truncate the write. */
1871                         if (end > ll) {
1872                                 ntfs_debug("Truncating write to inode 0x%lx, "
1873                                                 "attribute type 0x%x, because "
1874                                                 "the allocation was only "
1875                                                 "partially extended.",
1876                                                 vi->i_ino, (unsigned)
1877                                                 le32_to_cpu(ni->type));
1878                                 end = ll;
1879                                 count = ll - pos;
1880                         }
1881                 } else {
1882                         err = ll;
1883                         read_lock_irqsave(&ni->size_lock, flags);
1884                         ll = ni->allocated_size;
1885                         read_unlock_irqrestore(&ni->size_lock, flags);
1886                         /* Perform a partial write if possible or fail. */
1887                         if (pos < ll) {
1888                                 ntfs_debug("Truncating write to inode 0x%lx, "
1889                                                 "attribute type 0x%x, because "
1890                                                 "extending the allocation "
1891                                                 "failed (error code %i).",
1892                                                 vi->i_ino, (unsigned)
1893                                                 le32_to_cpu(ni->type), err);
1894                                 end = ll;
1895                                 count = ll - pos;
1896                         } else {
1897                                 ntfs_error(vol->sb, "Cannot perform write to "
1898                                                 "inode 0x%lx, attribute type "
1899                                                 "0x%x, because extending the "
1900                                                 "allocation failed (error "
1901                                                 "code %i).", vi->i_ino,
1902                                                 (unsigned)
1903                                                 le32_to_cpu(ni->type), err);
1904                                 return err;
1905                         }
1906                 }
1907         }
1908         written = 0;
1909         /*
1910          * If the write starts beyond the initialized size, extend it up to the
1911          * beginning of the write and initialize all non-sparse space between
1912          * the old initialized size and the new one.  This automatically also
1913          * increments the vfs inode->i_size to keep it above or equal to the
1914          * initialized_size.
1915          */
1916         read_lock_irqsave(&ni->size_lock, flags);
1917         ll = ni->initialized_size;
1918         read_unlock_irqrestore(&ni->size_lock, flags);
1919         if (pos > ll) {
1920                 err = ntfs_attr_extend_initialized(ni, pos);
1921                 if (err < 0) {
1922                         ntfs_error(vol->sb, "Cannot perform write to inode "
1923                                         "0x%lx, attribute type 0x%x, because "
1924                                         "extending the initialized size "
1925                                         "failed (error code %i).", vi->i_ino,
1926                                         (unsigned)le32_to_cpu(ni->type), err);
1927                         status = err;
1928                         goto err_out;
1929                 }
1930         }
1931         /*
1932          * Determine the number of pages per cluster for non-resident
1933          * attributes.
1934          */
1935         nr_pages = 1;
1936         if (vol->cluster_size > PAGE_CACHE_SIZE && NInoNonResident(ni))
1937                 nr_pages = vol->cluster_size >> PAGE_CACHE_SHIFT;
1938         /* Finally, perform the actual write. */
1939         last_vcn = -1;
1940         if (likely(nr_segs == 1))
1941                 buf = iov->iov_base;
1942         do {
1943                 VCN vcn;
1944                 pgoff_t idx, start_idx;
1945                 unsigned ofs, do_pages, u;
1946                 size_t copied;
1947
1948                 start_idx = idx = pos >> PAGE_CACHE_SHIFT;
1949                 ofs = pos & ~PAGE_CACHE_MASK;
1950                 bytes = PAGE_CACHE_SIZE - ofs;
1951                 do_pages = 1;
1952                 if (nr_pages > 1) {
1953                         vcn = pos >> vol->cluster_size_bits;
1954                         if (vcn != last_vcn) {
1955                                 last_vcn = vcn;
1956                                 /*
1957                                  * Get the lcn of the vcn the write is in.  If
1958                                  * it is a hole, need to lock down all pages in
1959                                  * the cluster.
1960                                  */
1961                                 down_read(&ni->runlist.lock);
1962                                 lcn = ntfs_attr_vcn_to_lcn_nolock(ni, pos >>
1963                                                 vol->cluster_size_bits, false);
1964                                 up_read(&ni->runlist.lock);
1965                                 if (unlikely(lcn < LCN_HOLE)) {
1966                                         status = -EIO;
1967                                         if (lcn == LCN_ENOMEM)
1968                                                 status = -ENOMEM;
1969                                         else
1970                                                 ntfs_error(vol->sb, "Cannot "
1971                                                         "perform write to "
1972                                                         "inode 0x%lx, "
1973                                                         "attribute type 0x%x, "
1974                                                         "because the attribute "
1975                                                         "is corrupt.",
1976                                                         vi->i_ino, (unsigned)
1977                                                         le32_to_cpu(ni->type));
1978                                         break;
1979                                 }
1980                                 if (lcn == LCN_HOLE) {
1981                                         start_idx = (pos & ~(s64)
1982                                                         vol->cluster_size_mask)
1983                                                         >> PAGE_CACHE_SHIFT;
1984                                         bytes = vol->cluster_size - (pos &
1985                                                         vol->cluster_size_mask);
1986                                         do_pages = nr_pages;
1987                                 }
1988                         }
1989                 }
1990                 if (bytes > count)
1991                         bytes = count;
1992                 /*
1993                  * Bring in the user page(s) that we will copy from _first_.
1994                  * Otherwise there is a nasty deadlock on copying from the same
1995                  * page(s) as we are writing to, without it/them being marked
1996                  * up-to-date.  Note, at present there is nothing to stop the
1997                  * pages being swapped out between us bringing them into memory
1998                  * and doing the actual copying.
1999                  */
2000                 if (likely(nr_segs == 1))
2001                         ntfs_fault_in_pages_readable(buf, bytes);
2002                 else
2003                         ntfs_fault_in_pages_readable_iovec(iov, iov_ofs, bytes);
2004                 /* Get and lock @do_pages starting at index @start_idx. */
2005                 status = __ntfs_grab_cache_pages(mapping, start_idx, do_pages,
2006                                 pages, &cached_page);
2007                 if (unlikely(status))
2008                         break;
2009                 /*
2010                  * For non-resident attributes, we need to fill any holes with
2011                  * actual clusters and ensure all bufferes are mapped.  We also
2012                  * need to bring uptodate any buffers that are only partially
2013                  * being written to.
2014                  */
2015                 if (NInoNonResident(ni)) {
2016                         status = ntfs_prepare_pages_for_non_resident_write(
2017                                         pages, do_pages, pos, bytes);
2018                         if (unlikely(status)) {
2019                                 loff_t i_size;
2020
2021                                 do {
2022                                         unlock_page(pages[--do_pages]);
2023                                         page_cache_release(pages[do_pages]);
2024                                 } while (do_pages);
2025                                 /*
2026                                  * The write preparation may have instantiated
2027                                  * allocated space outside i_size.  Trim this
2028                                  * off again.  We can ignore any errors in this
2029                                  * case as we will just be waisting a bit of
2030                                  * allocated space, which is not a disaster.
2031                                  */
2032                                 i_size = i_size_read(vi);
2033                                 if (pos + bytes > i_size) {
2034                                         ntfs_write_failed(mapping, pos + bytes);
2035                                 }
2036                                 break;
2037                         }
2038                 }
2039                 u = (pos >> PAGE_CACHE_SHIFT) - pages[0]->index;
2040                 if (likely(nr_segs == 1)) {
2041                         copied = ntfs_copy_from_user(pages + u, do_pages - u,
2042                                         ofs, buf, bytes);
2043                         buf += copied;
2044                 } else
2045                         copied = ntfs_copy_from_user_iovec(pages + u,
2046                                         do_pages - u, ofs, &iov, &iov_ofs,
2047                                         bytes);
2048                 ntfs_flush_dcache_pages(pages + u, do_pages - u);
2049                 status = ntfs_commit_pages_after_write(pages, do_pages, pos,
2050                                 bytes);
2051                 if (likely(!status)) {
2052                         written += copied;
2053                         count -= copied;
2054                         pos += copied;
2055                         if (unlikely(copied != bytes))
2056                                 status = -EFAULT;
2057                 }
2058                 do {
2059                         unlock_page(pages[--do_pages]);
2060                         page_cache_release(pages[do_pages]);
2061                 } while (do_pages);
2062                 if (unlikely(status))
2063                         break;
2064                 balance_dirty_pages_ratelimited(mapping);
2065                 cond_resched();
2066         } while (count);
2067 err_out:
2068         *ppos = pos;
2069         if (cached_page)
2070                 page_cache_release(cached_page);
2071         ntfs_debug("Done.  Returning %s (written 0x%lx, status %li).",
2072                         written ? "written" : "status", (unsigned long)written,
2073                         (long)status);
2074         return written ? written : status;
2075 }
2076
2077 /**
2078  * ntfs_file_aio_write_nolock -
2079  */
2080 static ssize_t ntfs_file_aio_write_nolock(struct kiocb *iocb,
2081                 const struct iovec *iov, unsigned long nr_segs, loff_t *ppos)
2082 {
2083         struct file *file = iocb->ki_filp;
2084         struct address_space *mapping = file->f_mapping;
2085         struct inode *inode = mapping->host;
2086         loff_t pos;
2087         size_t count;           /* after file limit checks */
2088         ssize_t written, err;
2089
2090         count = iov_length(iov, nr_segs);
2091         pos = *ppos;
2092         /* We can write back this queue in page reclaim. */
2093         current->backing_dev_info = mapping->backing_dev_info;
2094         written = 0;
2095         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2096         if (err)
2097                 goto out;
2098         if (!count)
2099                 goto out;
2100         err = file_remove_suid(file);
2101         if (err)
2102                 goto out;
2103         err = file_update_time(file);
2104         if (err)
2105                 goto out;
2106         written = ntfs_file_buffered_write(iocb, iov, nr_segs, pos, ppos,
2107                         count);
2108 out:
2109         current->backing_dev_info = NULL;
2110         return written ? written : err;
2111 }
2112
2113 /**
2114  * ntfs_file_aio_write -
2115  */
2116 static ssize_t ntfs_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2117                 unsigned long nr_segs, loff_t pos)
2118 {
2119         struct file *file = iocb->ki_filp;
2120         struct address_space *mapping = file->f_mapping;
2121         struct inode *inode = mapping->host;
2122         ssize_t ret;
2123
2124         BUG_ON(iocb->ki_pos != pos);
2125
2126         mutex_lock(&inode->i_mutex);
2127         ret = ntfs_file_aio_write_nolock(iocb, iov, nr_segs, &iocb->ki_pos);
2128         mutex_unlock(&inode->i_mutex);
2129         if (ret > 0) {
2130                 int err = generic_write_sync(file, iocb->ki_pos - ret, ret);
2131                 if (err < 0)
2132                         ret = err;
2133         }
2134         return ret;
2135 }
2136
2137 /**
2138  * ntfs_file_fsync - sync a file to disk
2139  * @filp:       file to be synced
2140  * @datasync:   if non-zero only flush user data and not metadata
2141  *
2142  * Data integrity sync of a file to disk.  Used for fsync, fdatasync, and msync
2143  * system calls.  This function is inspired by fs/buffer.c::file_fsync().
2144  *
2145  * If @datasync is false, write the mft record and all associated extent mft
2146  * records as well as the $DATA attribute and then sync the block device.
2147  *
2148  * If @datasync is true and the attribute is non-resident, we skip the writing
2149  * of the mft record and all associated extent mft records (this might still
2150  * happen due to the write_inode_now() call).
2151  *
2152  * Also, if @datasync is true, we do not wait on the inode to be written out
2153  * but we always wait on the page cache pages to be written out.
2154  *
2155  * Locking: Caller must hold i_mutex on the inode.
2156  *
2157  * TODO: We should probably also write all attribute/index inodes associated
2158  * with this inode but since we have no simple way of getting to them we ignore
2159  * this problem for now.
2160  */
2161 static int ntfs_file_fsync(struct file *filp, loff_t start, loff_t end,
2162                            int datasync)
2163 {
2164         struct inode *vi = filp->f_mapping->host;
2165         int err, ret = 0;
2166
2167         ntfs_debug("Entering for inode 0x%lx.", vi->i_ino);
2168
2169         err = filemap_write_and_wait_range(vi->i_mapping, start, end);
2170         if (err)
2171                 return err;
2172         mutex_lock(&vi->i_mutex);
2173
2174         BUG_ON(S_ISDIR(vi->i_mode));
2175         if (!datasync || !NInoNonResident(NTFS_I(vi)))
2176                 ret = __ntfs_write_inode(vi, 1);
2177         write_inode_now(vi, !datasync);
2178         /*
2179          * NOTE: If we were to use mapping->private_list (see ext2 and
2180          * fs/buffer.c) for dirty blocks then we could optimize the below to be
2181          * sync_mapping_buffers(vi->i_mapping).
2182          */
2183         err = sync_blockdev(vi->i_sb->s_bdev);
2184         if (unlikely(err && !ret))
2185                 ret = err;
2186         if (likely(!ret))
2187                 ntfs_debug("Done.");
2188         else
2189                 ntfs_warning(vi->i_sb, "Failed to f%ssync inode 0x%lx.  Error "
2190                                 "%u.", datasync ? "data" : "", vi->i_ino, -ret);
2191         mutex_unlock(&vi->i_mutex);
2192         return ret;
2193 }
2194
2195 #endif /* NTFS_RW */
2196
2197 const struct file_operations ntfs_file_ops = {
2198         .llseek         = generic_file_llseek,   /* Seek inside file. */
2199         .read           = new_sync_read,         /* Read from file. */
2200         .read_iter      = generic_file_read_iter, /* Async read from file. */
2201 #ifdef NTFS_RW
2202         .write          = do_sync_write,         /* Write to file. */
2203         .aio_write      = ntfs_file_aio_write,   /* Async write to file. */
2204         /*.release      = ,*/                    /* Last file is closed.  See
2205                                                     fs/ext2/file.c::
2206                                                     ext2_release_file() for
2207                                                     how to use this to discard
2208                                                     preallocated space for
2209                                                     write opened files. */
2210         .fsync          = ntfs_file_fsync,       /* Sync a file to disk. */
2211         /*.aio_fsync    = ,*/                    /* Sync all outstanding async
2212                                                     i/o operations on a
2213                                                     kiocb. */
2214 #endif /* NTFS_RW */
2215         /*.ioctl        = ,*/                    /* Perform function on the
2216                                                     mounted filesystem. */
2217         .mmap           = generic_file_mmap,     /* Mmap file. */
2218         .open           = ntfs_file_open,        /* Open file. */
2219         .splice_read    = generic_file_splice_read /* Zero-copy data send with
2220                                                     the data source being on
2221                                                     the ntfs partition.  We do
2222                                                     not need to care about the
2223                                                     data destination. */
2224         /*.sendpage     = ,*/                    /* Zero-copy data send with
2225                                                     the data destination being
2226                                                     on the ntfs partition.  We
2227                                                     do not need to care about
2228                                                     the data source. */
2229 };
2230
2231 const struct inode_operations ntfs_file_inode_ops = {
2232 #ifdef NTFS_RW
2233         .setattr        = ntfs_setattr,
2234 #endif /* NTFS_RW */
2235 };
2236
2237 const struct file_operations ntfs_empty_file_ops = {};
2238
2239 const struct inode_operations ntfs_empty_inode_ops = {};