fs/ntfs/file.c

   1 /*
   2  * file.c - NTFS kernel file operations.  Part of the Linux-NTFS project.
   3  *
   4  * Copyright (c) 2001-2007 Anton Altaparmakov
   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/pagemap.h>
  24 #include <linux/pagevec.h>
  25 #include <linux/sched.h>
  26 #include <linux/swap.h>
  27 #include <linux/uio.h>
  28 #include <linux/writeback.h>
  29
  30 #include <asm/page.h>
  31 #include <asm/uaccess.h>
  32
  33 #include "attrib.h"
  34 #include "bitmap.h"
  35 #include "inode.h"
  36 #include "debug.h"
  37 #include "lcnalloc.h"
  38 #include "malloc.h"
  39 #include "mft.h"
  40 #include "ntfs.h"
  41
  42 /**
  43  * ntfs_file_open - called when an inode is about to be opened
  44  * @vi:         inode to be opened
  45  * @filp:       file structure describing the inode
  46  *
  47  * Limit file size to the page cache limit on architectures where unsigned long
  48  * is 32-bits. This is the most we can do for now without overflowing the page
  49  * cache page index. Doing it this way means we don't run into problems because
  50  * of existing too large files. It would be better to allow the user to read
  51  * the beginning of the file but I doubt very much anyone is going to hit this
  52  * check on a 32-bit architecture, so there is no point in adding the extra
  53  * complexity required to support this.
  54  *
  55  * On 64-bit architectures, the check is hopefully optimized away by the
  56  * compiler.
  57  *
  58  * After the check passes, just call generic_file_open() to do its work.
  59  */
  60 static int ntfs_file_open(struct inode *vi, struct file *filp)
  61 {
  62         if (sizeof(unsigned long) < 8) {
  63                 if (i_size_read(vi) > MAX_LFS_FILESIZE)
  64                         return -EOVERFLOW;
  65         }
  66         return generic_file_open(vi, filp);
  67 }
  68
  69 #ifdef NTFS_RW
  70
  71 /**
  72  * ntfs_attr_extend_initialized - extend the initialized size of an attribute
  73  * @ni:                 ntfs inode of the attribute to extend
  74  * @new_init_size:      requested new initialized size in bytes
  75  * @cached_page:        store any allocated but unused page here
  76  * @lru_pvec:           lru-buffering pagevec of the caller
  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  * @cached_page and @lru_pvec are just optimizations for dealing with multiple
 101  * pages.
 102  *
 103  * Return 0 on success and -errno on error.  In the case that an error is
 104  * encountered it is possible that the initialized size will already have been
 105  * incremented some way towards @new_init_size but it is guaranteed that if
 106  * this is the case, the necessary zeroing will also have happened and that all
 107  * metadata is self-consistent.
 108  *
 109  * Locking: i_mutex on the vfs inode corrseponsind to the ntfs inode @ni must be
 110  *          held by the caller.
 111  */
 112 static int ntfs_attr_extend_initialized(ntfs_inode *ni, const s64 new_init_size,
 113                 struct page **cached_page, struct pagevec *lru_pvec)
 114 {
 115         s64 old_init_size;
 116         loff_t old_i_size;
 117         pgoff_t index, end_index;
 118         unsigned long flags;
 119         struct inode *vi = VFS_I(ni);
 120         ntfs_inode *base_ni;
 121         MFT_RECORD *m = NULL;
 122         ATTR_RECORD *a;
 123         ntfs_attr_search_ctx *ctx = NULL;
 124         struct address_space *mapping;
 125         struct page *page = NULL;
 126         u8 *kattr;
 127         int err;
 128         u32 attr_len;
 129
 130         read_lock_irqsave(&ni->size_lock, flags);
 131         old_init_size = ni->initialized_size;
 132         old_i_size = i_size_read(vi);
 133         BUG_ON(new_init_size > ni->allocated_size);
 134         read_unlock_irqrestore(&ni->size_lock, flags);
 135         ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, "
 136                         "old_initialized_size 0x%llx, "
 137                         "new_initialized_size 0x%llx, i_size 0x%llx.",
 138                         vi->i_ino, (unsigned)le32_to_cpu(ni->type),
 139                         (unsigned long long)old_init_size,
 140                         (unsigned long long)new_init_size, old_i_size);
 141         if (!NInoAttr(ni))
 142                 base_ni = ni;
 143         else
 144                 base_ni = ni->ext.base_ntfs_ino;
 145         /* Use goto to reduce indentation and we need the label below anyway. */
 146         if (NInoNonResident(ni))
 147                 goto do_non_resident_extend;
 148         BUG_ON(old_init_size != old_i_size);
 149         m = map_mft_record(base_ni);
 150         if (IS_ERR(m)) {
 151                 err = PTR_ERR(m);
 152                 m = NULL;
 153                 goto err_out;
 154         }
 155         ctx = ntfs_attr_get_search_ctx(base_ni, m);
 156         if (unlikely(!ctx)) {
 157                 err = -ENOMEM;
 158                 goto err_out;
 159         }
 160         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
 161                         CASE_SENSITIVE, 0, NULL, 0, ctx);
 162         if (unlikely(err)) {
 163                 if (err == -ENOENT)
 164                         err = -EIO;
 165                 goto err_out;
 166         }
 167         m = ctx->mrec;
 168         a = ctx->attr;
 169         BUG_ON(a->non_resident);
 170         /* The total length of the attribute value. */
 171         attr_len = le32_to_cpu(a->data.resident.value_length);
 172         BUG_ON(old_i_size != (loff_t)attr_len);
 173         /*
 174          * Do the zeroing in the mft record and update the attribute size in
 175          * the mft record.
 176          */
 177         kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
 178         memset(kattr + attr_len, 0, new_init_size - attr_len);
 179         a->data.resident.value_length = cpu_to_le32((u32)new_init_size);
 180         /* Finally, update the sizes in the vfs and ntfs inodes. */
 181         write_lock_irqsave(&ni->size_lock, flags);
 182         i_size_write(vi, new_init_size);
 183         ni->initialized_size = new_init_size;
 184         write_unlock_irqrestore(&ni->size_lock, flags);
 185         goto done;
 186 do_non_resident_extend:
 187         /*
 188          * If the new initialized size @new_init_size exceeds the current file
 189          * size (vfs inode->i_size), we need to extend the file size to the
 190          * new initialized size.
 191          */
 192         if (new_init_size > old_i_size) {
 193                 m = map_mft_record(base_ni);
 194                 if (IS_ERR(m)) {
 195                         err = PTR_ERR(m);
 196                         m = NULL;
 197                         goto err_out;
 198                 }
 199                 ctx = ntfs_attr_get_search_ctx(base_ni, m);
 200                 if (unlikely(!ctx)) {
 201                         err = -ENOMEM;
 202                         goto err_out;
 203                 }
 204                 err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
 205                                 CASE_SENSITIVE, 0, NULL, 0, ctx);
 206                 if (unlikely(err)) {
 207                         if (err == -ENOENT)
 208                                 err = -EIO;
 209                         goto err_out;
 210                 }
 211                 m = ctx->mrec;
 212                 a = ctx->attr;
 213                 BUG_ON(!a->non_resident);
 214                 BUG_ON(old_i_size != (loff_t)
 215                                 sle64_to_cpu(a->data.non_resident.data_size));
 216                 a->data.non_resident.data_size = cpu_to_sle64(new_init_size);
 217                 flush_dcache_mft_record_page(ctx->ntfs_ino);
 218                 mark_mft_record_dirty(ctx->ntfs_ino);
 219                 /* Update the file size in the vfs inode. */
 220                 i_size_write(vi, new_init_size);
 221                 ntfs_attr_put_search_ctx(ctx);
 222                 ctx = NULL;
 223                 unmap_mft_record(base_ni);
 224                 m = NULL;
 225         }
 226         mapping = vi->i_mapping;
 227         index = old_init_size >> PAGE_CACHE_SHIFT;
 228         end_index = (new_init_size + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
 229         do {
 230                 /*
 231                  * Read the page.  If the page is not present, this will zero
 232                  * the uninitialized regions for us.
 233                  */
 234                 page = read_mapping_page(mapping, index, NULL);
 235                 if (IS_ERR(page)) {
 236                         err = PTR_ERR(page);
 237                         goto init_err_out;
 238                 }
 239                 if (unlikely(PageError(page))) {
 240                         page_cache_release(page);
 241                         err = -EIO;
 242                         goto init_err_out;
 243                 }
 244                 /*
 245                  * Update the initialized size in the ntfs inode.  This is
 246                  * enough to make ntfs_writepage() work.
 247                  */
 248                 write_lock_irqsave(&ni->size_lock, flags);
 249                 ni->initialized_size = (s64)(index + 1) << PAGE_CACHE_SHIFT;
 250                 if (ni->initialized_size > new_init_size)
 251                         ni->initialized_size = new_init_size;
 252                 write_unlock_irqrestore(&ni->size_lock, flags);
 253                 /* Set the page dirty so it gets written out. */
 254                 set_page_dirty(page);
 255                 page_cache_release(page);
 256                 /*
 257                  * Play nice with the vm and the rest of the system.  This is
 258                  * very much needed as we can potentially be modifying the
 259                  * initialised size from a very small value to a really huge
 260                  * value, e.g.
 261                  *      f = open(somefile, O_TRUNC);
 262                  *      truncate(f, 10GiB);
 263                  *      seek(f, 10GiB);
 264                  *      write(f, 1);
 265                  * And this would mean we would be marking dirty hundreds of
 266                  * thousands of pages or as in the above example more than
 267                  * two and a half million pages!
 268                  *
 269                  * TODO: For sparse pages could optimize this workload by using
 270                  * the FsMisc / MiscFs page bit as a "PageIsSparse" bit.  This
 271                  * would be set in readpage for sparse pages and here we would
 272                  * not need to mark dirty any pages which have this bit set.
 273                  * The only caveat is that we have to clear the bit everywhere
 274                  * where we allocate any clusters that lie in the page or that
 275                  * contain the page.
 276                  *
 277                  * TODO: An even greater optimization would be for us to only
 278                  * call readpage() on pages which are not in sparse regions as
 279                  * determined from the runlist.  This would greatly reduce the
 280                  * number of pages we read and make dirty in the case of sparse
 281                  * files.
 282                  */
 283                 balance_dirty_pages_ratelimited(mapping);
 284                 cond_resched();
 285         } while (++index < end_index);
 286         read_lock_irqsave(&ni->size_lock, flags);
 287         BUG_ON(ni->initialized_size != new_init_size);
 288         read_unlock_irqrestore(&ni->size_lock, flags);
 289         /* Now bring in sync the initialized_size in the mft record. */
 290         m = map_mft_record(base_ni);
 291         if (IS_ERR(m)) {
 292                 err = PTR_ERR(m);
 293                 m = NULL;
 294                 goto init_err_out;
 295         }
 296         ctx = ntfs_attr_get_search_ctx(base_ni, m);
 297         if (unlikely(!ctx)) {
 298                 err = -ENOMEM;
 299                 goto init_err_out;
 300         }
 301         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
 302                         CASE_SENSITIVE, 0, NULL, 0, ctx);
 303         if (unlikely(err)) {
 304                 if (err == -ENOENT)
 305                         err = -EIO;
 306                 goto init_err_out;
 307         }
 308         m = ctx->mrec;
 309         a = ctx->attr;
 310         BUG_ON(!a->non_resident);
 311         a->data.non_resident.initialized_size = cpu_to_sle64(new_init_size);
 312 done:
 313         flush_dcache_mft_record_page(ctx->ntfs_ino);
 314         mark_mft_record_dirty(ctx->ntfs_ino);
 315         if (ctx)
 316                 ntfs_attr_put_search_ctx(ctx);
 317         if (m)
 318                 unmap_mft_record(base_ni);
 319         ntfs_debug("Done, initialized_size 0x%llx, i_size 0x%llx.",
 320                         (unsigned long long)new_init_size, i_size_read(vi));
 321         return 0;
 322 init_err_out:
 323         write_lock_irqsave(&ni->size_lock, flags);
 324         ni->initialized_size = old_init_size;
 325         write_unlock_irqrestore(&ni->size_lock, flags);
 326 err_out:
 327         if (ctx)
 328                 ntfs_attr_put_search_ctx(ctx);
 329         if (m)
 330                 unmap_mft_record(base_ni);
 331         ntfs_debug("Failed.  Returning error code %i.", err);
 332         return err;
 333 }
 334
 335 /**
 336  * ntfs_fault_in_pages_readable -
 337  *
 338  * Fault a number of userspace pages into pagetables.
 339  *
 340  * Unlike include/linux/pagemap.h::fault_in_pages_readable(), this one copes
 341  * with more than two userspace pages as well as handling the single page case
 342  * elegantly.
 343  *
 344  * If you find this difficult to understand, then think of the while loop being
 345  * the following code, except that we do without the integer variable ret:
 346  *
 347  *      do {
 348  *              ret = __get_user(c, uaddr);
 349  *              uaddr += PAGE_SIZE;
 350  *      } while (!ret && uaddr < end);
 351  *
 352  * Note, the final __get_user() may well run out-of-bounds of the user buffer,
 353  * but _not_ out-of-bounds of the page the user buffer belongs to, and since
 354  * this is only a read and not a write, and since it is still in the same page,
 355  * it should not matter and this makes the code much simpler.
 356  */
 357 static inline void ntfs_fault_in_pages_readable(const char __user *uaddr,
 358                 int bytes)
 359 {
 360         const char __user *end;
 361         volatile char c;
 362
 363         /* Set @end to the first byte outside the last page we care about. */
 364         end = (const char __user*)PAGE_ALIGN((unsigned long)uaddr + bytes);
 365
 366         while (!__get_user(c, uaddr) && (uaddr += PAGE_SIZE, uaddr < end))
 367                 ;
 368 }
 369
 370 /**
 371  * ntfs_fault_in_pages_readable_iovec -
 372  *
 373  * Same as ntfs_fault_in_pages_readable() but operates on an array of iovecs.
 374  */
 375 static inline void ntfs_fault_in_pages_readable_iovec(const struct iovec *iov,
 376                 size_t iov_ofs, int bytes)
 377 {
 378         do {
 379                 const char __user *buf;
 380                 unsigned len;
 381
 382                 buf = iov->iov_base + iov_ofs;
 383                 len = iov->iov_len - iov_ofs;
 384                 if (len > bytes)
 385                         len = bytes;
 386                 ntfs_fault_in_pages_readable(buf, len);
 387                 bytes -= len;
 388                 iov++;
 389                 iov_ofs = 0;
 390         } while (bytes);
 391 }
 392
 393 /**
 394  * __ntfs_grab_cache_pages - obtain a number of locked pages
 395  * @mapping:    address space mapping from which to obtain page cache pages
 396  * @index:      starting index in @mapping at which to begin obtaining pages
 397  * @nr_pages:   number of page cache pages to obtain
 398  * @pages:      array of pages in which to return the obtained page cache pages
 399  * @cached_page: allocated but as yet unused page
 400  * @lru_pvec:   lru-buffering pagevec of caller
 401  *
 402  * Obtain @nr_pages locked page cache pages from the mapping @maping and
 403  * starting at index @index.
 404  *
 405  * If a page is newly created, increment its refcount and add it to the
 406  * caller's lru-buffering pagevec @lru_pvec.
 407  *
 408  * This is the same as mm/filemap.c::__grab_cache_page(), except that @nr_pages
 409  * are obtained at once instead of just one page and that 0 is returned on
 410  * success and -errno on error.
 411  *
 412  * Note, the page locks are obtained in ascending page index order.
 413  */
 414 static inline int __ntfs_grab_cache_pages(struct address_space *mapping,
 415                 pgoff_t index, const unsigned nr_pages, struct page **pages,
 416                 struct page **cached_page, struct pagevec *lru_pvec)
 417 {
 418         int err, nr;
 419
 420         BUG_ON(!nr_pages);
 421         err = nr = 0;
 422         do {
 423                 pages[nr] = find_lock_page(mapping, index);
 424                 if (!pages[nr]) {
 425                         if (!*cached_page) {
 426                                 *cached_page = page_cache_alloc(mapping);
 427                                 if (unlikely(!*cached_page)) {
 428                                         err = -ENOMEM;
 429                                         goto err_out;
 430                                 }
 431                         }
 432                         err = add_to_page_cache(*cached_page, mapping, index,
 433                                         GFP_KERNEL);
 434                         if (unlikely(err)) {
 435                                 if (err == -EEXIST)
 436                                         continue;
 437                                 goto err_out;
 438                         }
 439                         pages[nr] = *cached_page;
 440                         page_cache_get(*cached_page);
 441                         if (unlikely(!pagevec_add(lru_pvec, *cached_page)))
 442                                 __pagevec_lru_add(lru_pvec);
 443                         *cached_page = NULL;
 444                 }
 445                 index++;
 446                 nr++;
 447         } while (nr < nr_pages);
 448 out:
 449         return err;
 450 err_out:
 451         while (nr > 0) {
 452                 unlock_page(pages[--nr]);
 453                 page_cache_release(pages[nr]);
 454         }
 455         goto out;
 456 }
 457
 458 static inline int ntfs_submit_bh_for_read(struct buffer_head *bh)
 459 {
 460         lock_buffer(bh);
 461         get_bh(bh);
 462         bh->b_end_io = end_buffer_read_sync;
 463         return submit_bh(READ, bh);
 464 }
 465
 466 /**
 467  * ntfs_prepare_pages_for_non_resident_write - prepare pages for receiving data
 468  * @pages:      array of destination pages
 469  * @nr_pages:   number of pages in @pages
 470  * @pos:        byte position in file at which the write begins
 471  * @bytes:      number of bytes to be written
 472  *
 473  * This is called for non-resident attributes from ntfs_file_buffered_write()
 474  * with i_mutex held on the inode (@pages[0]->mapping->host).  There are
 475  * @nr_pages pages in @pages which are locked but not kmap()ped.  The source
 476  * data has not yet been copied into the @pages.
 477  *
 478  * Need to fill any holes with actual clusters, allocate buffers if necessary,
 479  * ensure all the buffers are mapped, and bring uptodate any buffers that are
 480  * only partially being written to.
 481  *
 482  * If @nr_pages is greater than one, we are guaranteed that the cluster size is
 483  * greater than PAGE_CACHE_SIZE, that all pages in @pages are entirely inside
 484  * the same cluster and that they are the entirety of that cluster, and that
 485  * the cluster is sparse, i.e. we need to allocate a cluster to fill the hole.
 486  *
 487  * i_size is not to be modified yet.
 488  *
 489  * Return 0 on success or -errno on error.
 490  */
 491 static int ntfs_prepare_pages_for_non_resident_write(struct page **pages,
 492                 unsigned nr_pages, s64 pos, size_t bytes)
 493 {
 494         VCN vcn, highest_vcn = 0, cpos, cend, bh_cpos, bh_cend;
 495         LCN lcn;
 496         s64 bh_pos, vcn_len, end, initialized_size;
 497         sector_t lcn_block;
 498         struct page *page;
 499         struct inode *vi;
 500         ntfs_inode *ni, *base_ni = NULL;
 501         ntfs_volume *vol;
 502         runlist_element *rl, *rl2;
 503         struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
 504         ntfs_attr_search_ctx *ctx = NULL;
 505         MFT_RECORD *m = NULL;
 506         ATTR_RECORD *a = NULL;
 507         unsigned long flags;
 508         u32 attr_rec_len = 0;
 509         unsigned blocksize, u;
 510         int err, mp_size;
 511         bool rl_write_locked, was_hole, is_retry;
 512         unsigned char blocksize_bits;
 513         struct {
 514                 u8 runlist_merged:1;
 515                 u8 mft_attr_mapped:1;
 516                 u8 mp_rebuilt:1;
 517                 u8 attr_switched:1;
 518         } status = { 0, 0, 0, 0 };
 519
 520         BUG_ON(!nr_pages);
 521         BUG_ON(!pages);
 522         BUG_ON(!*pages);
 523         vi = pages[0]->mapping->host;
 524         ni = NTFS_I(vi);
 525         vol = ni->vol;
 526         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, start page "
 527                         "index 0x%lx, nr_pages 0x%x, pos 0x%llx, bytes 0x%zx.",
 528                         vi->i_ino, ni->type, pages[0]->index, nr_pages,
 529                         (long long)pos, bytes);
 530         blocksize = vol->sb->s_blocksize;
 531         blocksize_bits = vol->sb->s_blocksize_bits;
 532         u = 0;
 533         do {
 534                 page = pages[u];
 535                 BUG_ON(!page);
 536                 /*
 537                  * create_empty_buffers() will create uptodate/dirty buffers if
 538                  * the page is uptodate/dirty.
 539                  */
 540                 if (!page_has_buffers(page)) {
 541                         create_empty_buffers(page, blocksize, 0);
 542                         if (unlikely(!page_has_buffers(page)))
 543                                 return -ENOMEM;
 544                 }
 545         } while (++u < nr_pages);
 546         rl_write_locked = false;
 547         rl = NULL;
 548         err = 0;
 549         vcn = lcn = -1;
 550         vcn_len = 0;
 551         lcn_block = -1;
 552         was_hole = false;
 553         cpos = pos >> vol->cluster_size_bits;
 554         end = pos + bytes;
 555         cend = (end + vol->cluster_size - 1) >> vol->cluster_size_bits;
 556         /*
 557          * Loop over each page and for each page over each buffer.  Use goto to
 558          * reduce indentation.
 559          */
 560         u = 0;
 561 do_next_page:
 562         page = pages[u];
 563         bh_pos = (s64)page->index << PAGE_CACHE_SHIFT;
 564         bh = head = page_buffers(page);
 565         do {
 566                 VCN cdelta;
 567                 s64 bh_end;
 568                 unsigned bh_cofs;
 569
 570                 /* Clear buffer_new on all buffers to reinitialise state. */
 571                 if (buffer_new(bh))
 572                         clear_buffer_new(bh);
 573                 bh_end = bh_pos + blocksize;
 574                 bh_cpos = bh_pos >> vol->cluster_size_bits;
 575                 bh_cofs = bh_pos & vol->cluster_size_mask;
 576                 if (buffer_mapped(bh)) {
 577                         /*
 578                          * The buffer is already mapped.  If it is uptodate,
 579                          * ignore it.
 580                          */
 581                         if (buffer_uptodate(bh))
 582                                 continue;
 583                         /*
 584                          * The buffer is not uptodate.  If the page is uptodate
 585                          * set the buffer uptodate and otherwise ignore it.
 586                          */
 587                         if (PageUptodate(page)) {
 588                                 set_buffer_uptodate(bh);
 589                                 continue;
 590                         }
 591                         /*
 592                          * Neither the page nor the buffer are uptodate.  If
 593                          * the buffer is only partially being written to, we
 594                          * need to read it in before the write, i.e. now.
 595                          */
 596                         if ((bh_pos < pos && bh_end > pos) ||
 597                                         (bh_pos < end && bh_end > end)) {
 598                                 /*
 599                                  * If the buffer is fully or partially within
 600                                  * the initialized size, do an actual read.
 601                                  * Otherwise, simply zero the buffer.
 602                                  */
 603                                 read_lock_irqsave(&ni->size_lock, flags);
 604                                 initialized_size = ni->initialized_size;
 605                                 read_unlock_irqrestore(&ni->size_lock, flags);
 606                                 if (bh_pos < initialized_size) {
 607                                         ntfs_submit_bh_for_read(bh);
 608                                         *wait_bh++ = bh;
 609                                 } else {
 610                                         zero_user(page, bh_offset(bh),
 611                                                         blocksize);
 612                                         set_buffer_uptodate(bh);
 613                                 }
 614                         }
 615                         continue;
 616                 }
 617                 /* Unmapped buffer.  Need to map it. */
 618                 bh->b_bdev = vol->sb->s_bdev;
 619                 /*
 620                  * If the current buffer is in the same clusters as the map
 621                  * cache, there is no need to check the runlist again.  The
 622                  * map cache is made up of @vcn, which is the first cached file
 623                  * cluster, @vcn_len which is the number of cached file
 624                  * clusters, @lcn is the device cluster corresponding to @vcn,
 625                  * and @lcn_block is the block number corresponding to @lcn.
 626                  */
 627                 cdelta = bh_cpos - vcn;
 628                 if (likely(!cdelta || (cdelta > 0 && cdelta < vcn_len))) {
 629 map_buffer_cached:
 630                         BUG_ON(lcn < 0);
 631                         bh->b_blocknr = lcn_block +
 632                                         (cdelta << (vol->cluster_size_bits -
 633                                         blocksize_bits)) +
 634                                         (bh_cofs >> blocksize_bits);
 635                         set_buffer_mapped(bh);
 636                         /*
 637                          * If the page is uptodate so is the buffer.  If the
 638                          * buffer is fully outside the write, we ignore it if
 639                          * it was already allocated and we mark it dirty so it
 640                          * gets written out if we allocated it.  On the other
 641                          * hand, if we allocated the buffer but we are not
 642                          * marking it dirty we set buffer_new so we can do
 643                          * error recovery.
 644                          */
 645                         if (PageUptodate(page)) {
 646                                 if (!buffer_uptodate(bh))
 647                                         set_buffer_uptodate(bh);
 648                                 if (unlikely(was_hole)) {
 649                                         /* We allocated the buffer. */
 650                                         unmap_underlying_metadata(bh->b_bdev,
 651                                                         bh->b_blocknr);
 652                                         if (bh_end <= pos || bh_pos >= end)
 653                                                 mark_buffer_dirty(bh);
 654                                         else
 655                                                 set_buffer_new(bh);
 656                                 }
 657                                 continue;
 658                         }
 659                         /* Page is _not_ uptodate. */
 660                         if (likely(!was_hole)) {
 661                                 /*
 662                                  * Buffer was already allocated.  If it is not
 663                                  * uptodate and is only partially being written
 664                                  * to, we need to read it in before the write,
 665                                  * i.e. now.
 666                                  */
 667                                 if (!buffer_uptodate(bh) && bh_pos < end &&
 668                                                 bh_end > pos &&
 669                                                 (bh_pos < pos ||
 670                                                 bh_end > end)) {
 671                                         /*
 672                                          * If the buffer is fully or partially
 673                                          * within the initialized size, do an
 674                                          * actual read.  Otherwise, simply zero
 675                                          * the buffer.
 676                                          */
 677                                         read_lock_irqsave(&ni->size_lock,
 678                                                         flags);
 679                                         initialized_size = ni->initialized_size;
 680                                         read_unlock_irqrestore(&ni->size_lock,
 681                                                         flags);
 682                                         if (bh_pos < initialized_size) {
 683                                                 ntfs_submit_bh_for_read(bh);
 684                                                 *wait_bh++ = bh;
 685                                         } else {
 686                                                 zero_user(page, bh_offset(bh),
 687                                                                 blocksize);
 688                                                 set_buffer_uptodate(bh);
 689                                         }
 690                                 }
 691                                 continue;
 692                         }
 693                         /* We allocated the buffer. */
 694                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 695                         /*
 696                          * If the buffer is fully outside the write, zero it,
 697                          * set it uptodate, and mark it dirty so it gets
 698                          * written out.  If it is partially being written to,
 699                          * zero region surrounding the write but leave it to
 700                          * commit write to do anything else.  Finally, if the
 701                          * buffer is fully being overwritten, do nothing.
 702                          */
 703                         if (bh_end <= pos || bh_pos >= end) {
 704                                 if (!buffer_uptodate(bh)) {
 705                                         zero_user(page, bh_offset(bh),
 706                                                         blocksize);
 707                                         set_buffer_uptodate(bh);
 708                                 }
 709                                 mark_buffer_dirty(bh);
 710                                 continue;
 711                         }
 712                         set_buffer_new(bh);
 713                         if (!buffer_uptodate(bh) &&
 714                                         (bh_pos < pos || bh_end > end)) {
 715                                 u8 *kaddr;
 716                                 unsigned pofs;
 717
 718                                 kaddr = kmap_atomic(page, KM_USER0);
 719                                 if (bh_pos < pos) {
 720                                         pofs = bh_pos & ~PAGE_CACHE_MASK;
 721                                         memset(kaddr + pofs, 0, pos - bh_pos);
 722                                 }
 723                                 if (bh_end > end) {
 724                                         pofs = end & ~PAGE_CACHE_MASK;
 725                                         memset(kaddr + pofs, 0, bh_end - end);
 726                                 }
 727                                 kunmap_atomic(kaddr, KM_USER0);
 728                                 flush_dcache_page(page);
 729                         }
 730                         continue;
 731                 }
 732                 /*
 733                  * Slow path: this is the first buffer in the cluster.  If it
 734                  * is outside allocated size and is not uptodate, zero it and
 735                  * set it uptodate.
 736                  */
 737                 read_lock_irqsave(&ni->size_lock, flags);
 738                 initialized_size = ni->allocated_size;
 739                 read_unlock_irqrestore(&ni->size_lock, flags);
 740                 if (bh_pos > initialized_size) {
 741                         if (PageUptodate(page)) {
 742                                 if (!buffer_uptodate(bh))
 743                                         set_buffer_uptodate(bh);
 744                         } else if (!buffer_uptodate(bh)) {
 745                                 zero_user(page, bh_offset(bh), blocksize);
 746                                 set_buffer_uptodate(bh);
 747                         }
 748                         continue;
 749                 }
 750                 is_retry = false;
 751                 if (!rl) {
 752                         down_read(&ni->runlist.lock);
 753 retry_remap:
 754                         rl = ni->runlist.rl;
 755                 }
 756                 if (likely(rl != NULL)) {
 757                         /* Seek to element containing target cluster. */
 758                         while (rl->length && rl[1].vcn <= bh_cpos)
 759                                 rl++;
 760                         lcn = ntfs_rl_vcn_to_lcn(rl, bh_cpos);
 761                         if (likely(lcn >= 0)) {
 762                                 /*
 763                                  * Successful remap, setup the map cache and
 764                                  * use that to deal with the buffer.
 765                                  */
 766                                 was_hole = false;
 767                                 vcn = bh_cpos;
 768                                 vcn_len = rl[1].vcn - vcn;
 769                                 lcn_block = lcn << (vol->cluster_size_bits -
 770                                                 blocksize_bits);
 771                                 cdelta = 0;
 772                                 /*
 773                                  * If the number of remaining clusters touched
 774                                  * by the write is smaller or equal to the
 775                                  * number of cached clusters, unlock the
 776                                  * runlist as the map cache will be used from
 777                                  * now on.
 778                                  */
 779                                 if (likely(vcn + vcn_len >= cend)) {
 780                                         if (rl_write_locked) {
 781                                                 up_write(&ni->runlist.lock);
 782                                                 rl_write_locked = false;
 783                                         } else
 784                                                 up_read(&ni->runlist.lock);
 785                                         rl = NULL;
 786                                 }
 787                                 goto map_buffer_cached;
 788                         }
 789                 } else
 790                         lcn = LCN_RL_NOT_MAPPED;
 791                 /*
 792                  * If it is not a hole and not out of bounds, the runlist is
 793                  * probably unmapped so try to map it now.
 794                  */
 795                 if (unlikely(lcn != LCN_HOLE && lcn != LCN_ENOENT)) {
 796                         if (likely(!is_retry && lcn == LCN_RL_NOT_MAPPED)) {
 797                                 /* Attempt to map runlist. */
 798                                 if (!rl_write_locked) {
 799                                         /*
 800                                          * We need the runlist locked for
 801                                          * writing, so if it is locked for
 802                                          * reading relock it now and retry in
 803                                          * case it changed whilst we dropped
 804                                          * the lock.
 805                                          */
 806                                         up_read(&ni->runlist.lock);
 807                                         down_write(&ni->runlist.lock);
 808                                         rl_write_locked = true;
 809                                         goto retry_remap;
 810                                 }
 811                                 err = ntfs_map_runlist_nolock(ni, bh_cpos,
 812                                                 NULL);
 813                                 if (likely(!err)) {
 814                                         is_retry = true;
 815                                         goto retry_remap;
 816                                 }
 817                                 /*
 818                                  * If @vcn is out of bounds, pretend @lcn is
 819                                  * LCN_ENOENT.  As long as the buffer is out
 820                                  * of bounds this will work fine.
 821                                  */
 822                                 if (err == -ENOENT) {
 823                                         lcn = LCN_ENOENT;
 824                                         err = 0;
 825                                         goto rl_not_mapped_enoent;
 826                                 }
 827                         } else
 828                                 err = -EIO;
 829                         /* Failed to map the buffer, even after retrying. */
 830                         bh->b_blocknr = -1;
 831                         ntfs_error(vol->sb, "Failed to write to inode 0x%lx, "
 832                                         "attribute type 0x%x, vcn 0x%llx, "
 833                                         "vcn offset 0x%x, because its "
 834                                         "location on disk could not be "
 835                                         "determined%s (error code %i).",
 836                                         ni->mft_no, ni->type,
 837                                         (unsigned long long)bh_cpos,
 838                                         (unsigned)bh_pos &
 839                                         vol->cluster_size_mask,
 840                                         is_retry ? " even after retrying" : "",
 841                                         err);
 842                         break;
 843                 }
 844 rl_not_mapped_enoent:
 845                 /*
 846                  * The buffer is in a hole or out of bounds.  We need to fill
 847                  * the hole, unless the buffer is in a cluster which is not
 848                  * touched by the write, in which case we just leave the buffer
 849                  * unmapped.  This can only happen when the cluster size is
 850                  * less than the page cache size.
 851                  */
 852                 if (unlikely(vol->cluster_size < PAGE_CACHE_SIZE)) {
 853                         bh_cend = (bh_end + vol->cluster_size - 1) >>
 854                                         vol->cluster_size_bits;
 855                         if ((bh_cend <= cpos || bh_cpos >= cend)) {
 856                                 bh->b_blocknr = -1;
 857                                 /*
 858                                  * If the buffer is uptodate we skip it.  If it
 859                                  * is not but the page is uptodate, we can set
 860                                  * the buffer uptodate.  If the page is not
 861                                  * uptodate, we can clear the buffer and set it
 862                                  * uptodate.  Whether this is worthwhile is
 863                                  * debatable and this could be removed.
 864                                  */
 865                                 if (PageUptodate(page)) {
 866                                         if (!buffer_uptodate(bh))
 867                                                 set_buffer_uptodate(bh);
 868                                 } else if (!buffer_uptodate(bh)) {
 869                                         zero_user(page, bh_offset(bh),
 870                                                 blocksize);
 871                                         set_buffer_uptodate(bh);
 872                                 }
 873                                 continue;
 874                         }
 875                 }
 876                 /*
 877                  * Out of bounds buffer is invalid if it was not really out of
 878                  * bounds.
 879                  */
 880                 BUG_ON(lcn != LCN_HOLE);
 881                 /*
 882                  * We need the runlist locked for writing, so if it is locked
 883                  * for reading relock it now and retry in case it changed
 884                  * whilst we dropped the lock.
 885                  */
 886                 BUG_ON(!rl);
 887                 if (!rl_write_locked) {
 888                         up_read(&ni->runlist.lock);
 889                         down_write(&ni->runlist.lock);
 890                         rl_write_locked = true;
 891                         goto retry_remap;
 892                 }
 893                 /* Find the previous last allocated cluster. */
 894                 BUG_ON(rl->lcn != LCN_HOLE);
 895                 lcn = -1;
 896                 rl2 = rl;
 897                 while (--rl2 >= ni->runlist.rl) {
 898                         if (rl2->lcn >= 0) {
 899                                 lcn = rl2->lcn + rl2->length;
 900                                 break;
 901                         }
 902                 }
 903                 rl2 = ntfs_cluster_alloc(vol, bh_cpos, 1, lcn, DATA_ZONE,
 904                                 false);
 905                 if (IS_ERR(rl2)) {
 906                         err = PTR_ERR(rl2);
 907                         ntfs_debug("Failed to allocate cluster, error code %i.",
 908                                         err);
 909                         break;
 910                 }
 911                 lcn = rl2->lcn;
 912                 rl = ntfs_runlists_merge(ni->runlist.rl, rl2);
 913                 if (IS_ERR(rl)) {
 914                         err = PTR_ERR(rl);
 915                         if (err != -ENOMEM)
 916                                 err = -EIO;
 917                         if (ntfs_cluster_free_from_rl(vol, rl2)) {
 918                                 ntfs_error(vol->sb, "Failed to release "
 919                                                 "allocated cluster in error "
 920                                                 "code path.  Run chkdsk to "
 921                                                 "recover the lost cluster.");
 922                                 NVolSetErrors(vol);
 923                         }
 924                         ntfs_free(rl2);
 925                         break;
 926                 }
 927                 ni->runlist.rl = rl;
 928                 status.runlist_merged = 1;
 929                 ntfs_debug("Allocated cluster, lcn 0x%llx.",
 930                                 (unsigned long long)lcn);
 931                 /* Map and lock the mft record and get the attribute record. */
 932                 if (!NInoAttr(ni))
 933                         base_ni = ni;
 934                 else
 935                         base_ni = ni->ext.base_ntfs_ino;
 936                 m = map_mft_record(base_ni);
 937                 if (IS_ERR(m)) {
 938                         err = PTR_ERR(m);
 939                         break;
 940                 }
 941                 ctx = ntfs_attr_get_search_ctx(base_ni, m);
 942                 if (unlikely(!ctx)) {
 943                         err = -ENOMEM;
 944                         unmap_mft_record(base_ni);
 945                         break;
 946                 }
 947                 status.mft_attr_mapped = 1;
 948                 err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
 949                                 CASE_SENSITIVE, bh_cpos, NULL, 0, ctx);
 950                 if (unlikely(err)) {
 951                         if (err == -ENOENT)
 952                                 err = -EIO;
 953                         break;
 954                 }
 955                 m = ctx->mrec;
 956                 a = ctx->attr;
 957                 /*
 958                  * Find the runlist element with which the attribute extent
 959                  * starts.  Note, we cannot use the _attr_ version because we
 960                  * have mapped the mft record.  That is ok because we know the
 961                  * runlist fragment must be mapped already to have ever gotten
 962                  * here, so we can just use the _rl_ version.
 963                  */
 964                 vcn = sle64_to_cpu(a->data.non_resident.lowest_vcn);
 965                 rl2 = ntfs_rl_find_vcn_nolock(rl, vcn);
 966                 BUG_ON(!rl2);
 967                 BUG_ON(!rl2->length);
 968                 BUG_ON(rl2->lcn < LCN_HOLE);
 969                 highest_vcn = sle64_to_cpu(a->data.non_resident.highest_vcn);
 970                 /*
 971                  * If @highest_vcn is zero, calculate the real highest_vcn
 972                  * (which can really be zero).
 973                  */
 974                 if (!highest_vcn)
 975                         highest_vcn = (sle64_to_cpu(
 976                                         a->data.non_resident.allocated_size) >>
 977                                         vol->cluster_size_bits) - 1;
 978                 /*
 979                  * Determine the size of the mapping pairs array for the new
 980                  * extent, i.e. the old extent with the hole filled.
 981                  */
 982                 mp_size = ntfs_get_size_for_mapping_pairs(vol, rl2, vcn,
 983                                 highest_vcn);
 984                 if (unlikely(mp_size <= 0)) {
 985                         if (!(err = mp_size))
 986                                 err = -EIO;
 987                         ntfs_debug("Failed to get size for mapping pairs "
 988                                         "array, error code %i.", err);
 989                         break;
 990                 }
 991                 /*
 992                  * Resize the attribute record to fit the new mapping pairs
 993                  * array.
 994                  */
 995                 attr_rec_len = le32_to_cpu(a->length);
 996                 err = ntfs_attr_record_resize(m, a, mp_size + le16_to_cpu(
 997                                 a->data.non_resident.mapping_pairs_offset));
 998                 if (unlikely(err)) {
 999                         BUG_ON(err != -ENOSPC);
1000                         // TODO: Deal with this by using the current attribute
1001                         // and fill it with as much of the mapping pairs
1002                         // array as possible.  Then loop over each attribute
1003                         // extent rewriting the mapping pairs arrays as we go
1004                         // along and if when we reach the end we have not
1005                         // enough space, try to resize the last attribute
1006                         // extent and if even that fails, add a new attribute
1007                         // extent.
1008                         // We could also try to resize at each step in the hope
1009                         // that we will not need to rewrite every single extent.
1010                         // Note, we may need to decompress some extents to fill
1011                         // the runlist as we are walking the extents...
1012                         ntfs_error(vol->sb, "Not enough space in the mft "
1013                                         "record for the extended attribute "
1014                                         "record.  This case is not "
1015                                         "implemented yet.");
1016                         err = -EOPNOTSUPP;
1017                         break ;
1018                 }
1019                 status.mp_rebuilt = 1;
1020                 /*
1021                  * Generate the mapping pairs array directly into the attribute
1022                  * record.
1023                  */
1024                 err = ntfs_mapping_pairs_build(vol, (u8*)a + le16_to_cpu(
1025                                 a->data.non_resident.mapping_pairs_offset),
1026                                 mp_size, rl2, vcn, highest_vcn, NULL);
1027                 if (unlikely(err)) {
1028                         ntfs_error(vol->sb, "Cannot fill hole in inode 0x%lx, "
1029                                         "attribute type 0x%x, because building "
1030                                         "the mapping pairs failed with error "
1031                                         "code %i.", vi->i_ino,
1032                                         (unsigned)le32_to_cpu(ni->type), err);
1033                         err = -EIO;
1034                         break;
1035                 }
1036                 /* Update the highest_vcn but only if it was not set. */
1037                 if (unlikely(!a->data.non_resident.highest_vcn))
1038                         a->data.non_resident.highest_vcn =
1039                                         cpu_to_sle64(highest_vcn);
1040                 /*
1041                  * If the attribute is sparse/compressed, update the compressed
1042                  * size in the ntfs_inode structure and the attribute record.
1043                  */
1044                 if (likely(NInoSparse(ni) || NInoCompressed(ni))) {
1045                         /*
1046                          * If we are not in the first attribute extent, switch
1047                          * to it, but first ensure the changes will make it to
1048                          * disk later.
1049                          */
1050                         if (a->data.non_resident.lowest_vcn) {
1051                                 flush_dcache_mft_record_page(ctx->ntfs_ino);
1052                                 mark_mft_record_dirty(ctx->ntfs_ino);
1053                                 ntfs_attr_reinit_search_ctx(ctx);
1054                                 err = ntfs_attr_lookup(ni->type, ni->name,
1055                                                 ni->name_len, CASE_SENSITIVE,
1056                                                 0, NULL, 0, ctx);
1057                                 if (unlikely(err)) {
1058                                         status.attr_switched = 1;
1059                                         break;
1060                                 }
1061                                 /* @m is not used any more so do not set it. */
1062                                 a = ctx->attr;
1063                         }
1064                         write_lock_irqsave(&ni->size_lock, flags);
1065                         ni->itype.compressed.size += vol->cluster_size;
1066                         a->data.non_resident.compressed_size =
1067                                         cpu_to_sle64(ni->itype.compressed.size);
1068                         write_unlock_irqrestore(&ni->size_lock, flags);
1069                 }
1070                 /* Ensure the changes make it to disk. */
1071                 flush_dcache_mft_record_page(ctx->ntfs_ino);
1072                 mark_mft_record_dirty(ctx->ntfs_ino);
1073                 ntfs_attr_put_search_ctx(ctx);
1074                 unmap_mft_record(base_ni);
1075                 /* Successfully filled the hole. */
1076                 status.runlist_merged = 0;
1077                 status.mft_attr_mapped = 0;
1078                 status.mp_rebuilt = 0;
1079                 /* Setup the map cache and use that to deal with the buffer. */
1080                 was_hole = true;
1081                 vcn = bh_cpos;
1082                 vcn_len = 1;
1083                 lcn_block = lcn << (vol->cluster_size_bits - blocksize_bits);
1084                 cdelta = 0;
1085                 /*
1086                  * If the number of remaining clusters in the @pages is smaller
1087                  * or equal to the number of cached clusters, unlock the
1088                  * runlist as the map cache will be used from now on.
1089                  */
1090                 if (likely(vcn + vcn_len >= cend)) {
1091                         up_write(&ni->runlist.lock);
1092                         rl_write_locked = false;
1093                         rl = NULL;
1094                 }
1095                 goto map_buffer_cached;
1096         } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
1097         /* If there are no errors, do the next page. */
1098         if (likely(!err && ++u < nr_pages))
1099                 goto do_next_page;
1100         /* If there are no errors, release the runlist lock if we took it. */
1101         if (likely(!err)) {
1102                 if (unlikely(rl_write_locked)) {
1103                         up_write(&ni->runlist.lock);
1104                         rl_write_locked = false;
1105                 } else if (unlikely(rl))
1106                         up_read(&ni->runlist.lock);
1107                 rl = NULL;
1108         }
1109         /* If we issued read requests, let them complete. */
1110         read_lock_irqsave(&ni->size_lock, flags);
1111         initialized_size = ni->initialized_size;
1112         read_unlock_irqrestore(&ni->size_lock, flags);
1113         while (wait_bh > wait) {
1114                 bh = *--wait_bh;
1115                 wait_on_buffer(bh);
1116                 if (likely(buffer_uptodate(bh))) {
1117                         page = bh->b_page;
1118                         bh_pos = ((s64)page->index << PAGE_CACHE_SHIFT) +
1119                                         bh_offset(bh);
1120                         /*
1121                          * If the buffer overflows the initialized size, need
1122                          * to zero the overflowing region.
1123                          */
1124                         if (unlikely(bh_pos + blocksize > initialized_size)) {
1125                                 int ofs = 0;
1126
1127                                 if (likely(bh_pos < initialized_size))
1128                                         ofs = initialized_size - bh_pos;
1129                                 zero_user_segment(page, bh_offset(bh) + ofs,
1130                                                 blocksize);
1131                         }
1132                 } else /* if (unlikely(!buffer_uptodate(bh))) */
1133                         err = -EIO;
1134         }
1135         if (likely(!err)) {
1136                 /* Clear buffer_new on all buffers. */
1137                 u = 0;
1138                 do {
1139                         bh = head = page_buffers(pages[u]);
1140                         do {
1141                                 if (buffer_new(bh))
1142                                         clear_buffer_new(bh);
1143                         } while ((bh = bh->b_this_page) != head);
1144                 } while (++u < nr_pages);
1145                 ntfs_debug("Done.");
1146                 return err;
1147         }
1148         if (status.attr_switched) {
1149                 /* Get back to the attribute extent we modified. */
1150                 ntfs_attr_reinit_search_ctx(ctx);
1151                 if (ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1152                                 CASE_SENSITIVE, bh_cpos, NULL, 0, ctx)) {
1153                         ntfs_error(vol->sb, "Failed to find required "
1154                                         "attribute extent of attribute in "
1155                                         "error code path.  Run chkdsk to "
1156                                         "recover.");
1157                         write_lock_irqsave(&ni->size_lock, flags);
1158                         ni->itype.compressed.size += vol->cluster_size;
1159                         write_unlock_irqrestore(&ni->size_lock, flags);
1160                         flush_dcache_mft_record_page(ctx->ntfs_ino);
1161                         mark_mft_record_dirty(ctx->ntfs_ino);
1162                         /*
1163                          * The only thing that is now wrong is the compressed
1164                          * size of the base attribute extent which chkdsk
1165                          * should be able to fix.
1166                          */
1167                         NVolSetErrors(vol);
1168                 } else {
1169                         m = ctx->mrec;
1170                         a = ctx->attr;
1171                         status.attr_switched = 0;
1172                 }
1173         }
1174         /*
1175          * If the runlist has been modified, need to restore it by punching a
1176          * hole into it and we then need to deallocate the on-disk cluster as
1177          * well.  Note, we only modify the runlist if we are able to generate a
1178          * new mapping pairs array, i.e. only when the mapped attribute extent
1179          * is not switched.
1180          */
1181         if (status.runlist_merged && !status.attr_switched) {
1182                 BUG_ON(!rl_write_locked);
1183                 /* Make the file cluster we allocated sparse in the runlist. */
1184                 if (ntfs_rl_punch_nolock(vol, &ni->runlist, bh_cpos, 1)) {
1185                         ntfs_error(vol->sb, "Failed to punch hole into "
1186                                         "attribute runlist in error code "
1187                                         "path.  Run chkdsk to recover the "
1188                                         "lost cluster.");
1189                         NVolSetErrors(vol);
1190                 } else /* if (success) */ {
1191                         status.runlist_merged = 0;
1192                         /*
1193                          * Deallocate the on-disk cluster we allocated but only
1194                          * if we succeeded in punching its vcn out of the
1195                          * runlist.
1196                          */
1197                         down_write(&vol->lcnbmp_lock);
1198                         if (ntfs_bitmap_clear_bit(vol->lcnbmp_ino, lcn)) {
1199                                 ntfs_error(vol->sb, "Failed to release "
1200                                                 "allocated cluster in error "
1201                                                 "code path.  Run chkdsk to "
1202                                                 "recover the lost cluster.");
1203                                 NVolSetErrors(vol);
1204                         }
1205                         up_write(&vol->lcnbmp_lock);
1206                 }
1207         }
1208         /*
1209          * Resize the attribute record to its old size and rebuild the mapping
1210          * pairs array.  Note, we only can do this if the runlist has been
1211          * restored to its old state which also implies that the mapped
1212          * attribute extent is not switched.
1213          */
1214         if (status.mp_rebuilt && !status.runlist_merged) {
1215                 if (ntfs_attr_record_resize(m, a, attr_rec_len)) {
1216                         ntfs_error(vol->sb, "Failed to restore attribute "
1217                                         "record in error code path.  Run "
1218                                         "chkdsk to recover.");
1219                         NVolSetErrors(vol);
1220                 } else /* if (success) */ {
1221                         if (ntfs_mapping_pairs_build(vol, (u8*)a +
1222                                         le16_to_cpu(a->data.non_resident.
1223                                         mapping_pairs_offset), attr_rec_len -
1224                                         le16_to_cpu(a->data.non_resident.
1225                                         mapping_pairs_offset), ni->runlist.rl,
1226                                         vcn, highest_vcn, NULL)) {
1227                                 ntfs_error(vol->sb, "Failed to restore "
1228                                                 "mapping pairs array in error "
1229                                                 "code path.  Run chkdsk to "
1230                                                 "recover.");
1231                                 NVolSetErrors(vol);
1232                         }
1233                         flush_dcache_mft_record_page(ctx->ntfs_ino);
1234                         mark_mft_record_dirty(ctx->ntfs_ino);
1235                 }
1236         }
1237         /* Release the mft record and the attribute. */
1238         if (status.mft_attr_mapped) {
1239                 ntfs_attr_put_search_ctx(ctx);
1240                 unmap_mft_record(base_ni);
1241         }
1242         /* Release the runlist lock. */
1243         if (rl_write_locked)
1244                 up_write(&ni->runlist.lock);
1245         else if (rl)
1246                 up_read(&ni->runlist.lock);
1247         /*
1248          * Zero out any newly allocated blocks to avoid exposing stale data.
1249          * If BH_New is set, we know that the block was newly allocated above
1250          * and that it has not been fully zeroed and marked dirty yet.
1251          */
1252         nr_pages = u;
1253         u = 0;
1254         end = bh_cpos << vol->cluster_size_bits;
1255         do {
1256                 page = pages[u];
1257                 bh = head = page_buffers(page);
1258                 do {
1259                         if (u == nr_pages &&
1260                                         ((s64)page->index << PAGE_CACHE_SHIFT) +
1261                                         bh_offset(bh) >= end)
1262                                 break;
1263                         if (!buffer_new(bh))
1264                                 continue;
1265                         clear_buffer_new(bh);
1266                         if (!buffer_uptodate(bh)) {
1267                                 if (PageUptodate(page))
1268                                         set_buffer_uptodate(bh);
1269                                 else {
1270                                         zero_user(page, bh_offset(bh),
1271                                                         blocksize);
1272                                         set_buffer_uptodate(bh);
1273                                 }
1274                         }
1275                         mark_buffer_dirty(bh);
1276                 } while ((bh = bh->b_this_page) != head);
1277         } while (++u <= nr_pages);
1278         ntfs_error(vol->sb, "Failed.  Returning error code %i.", err);
1279         return err;
1280 }
1281
1282 /*
1283  * Copy as much as we can into the pages and return the number of bytes which
1284  * were sucessfully copied.  If a fault is encountered then clear the pages
1285  * out to (ofs + bytes) and return the number of bytes which were copied.
1286  */
1287 static inline size_t ntfs_copy_from_user(struct page **pages,
1288                 unsigned nr_pages, unsigned ofs, const char __user *buf,
1289                 size_t bytes)
1290 {
1291         struct page **last_page = pages + nr_pages;
1292         char *addr;
1293         size_t total = 0;
1294         unsigned len;
1295         int left;
1296
1297         do {
1298                 len = PAGE_CACHE_SIZE - ofs;
1299                 if (len > bytes)
1300                         len = bytes;
1301                 addr = kmap_atomic(*pages, KM_USER0);
1302                 left = __copy_from_user_inatomic(addr + ofs, buf, len);
1303                 kunmap_atomic(addr, KM_USER0);
1304                 if (unlikely(left)) {
1305                         /* Do it the slow way. */
1306                         addr = kmap(*pages);
1307                         left = __copy_from_user(addr + ofs, buf, len);
1308                         kunmap(*pages);
1309                         if (unlikely(left))
1310                                 goto err_out;
1311                 }
1312                 total += len;
1313                 bytes -= len;
1314                 if (!bytes)
1315                         break;
1316                 buf += len;
1317                 ofs = 0;
1318         } while (++pages < last_page);
1319 out:
1320         return total;
1321 err_out:
1322         total += len - left;
1323         /* Zero the rest of the target like __copy_from_user(). */
1324         while (++pages < last_page) {
1325                 bytes -= len;
1326                 if (!bytes)
1327                         break;
1328                 len = PAGE_CACHE_SIZE;
1329                 if (len > bytes)
1330                         len = bytes;
1331                 zero_user(*pages, 0, len);
1332         }
1333         goto out;
1334 }
1335
1336 static size_t __ntfs_copy_from_user_iovec_inatomic(char *vaddr,
1337                 const struct iovec *iov, size_t iov_ofs, size_t bytes)
1338 {
1339         size_t total = 0;
1340
1341         while (1) {
1342                 const char __user *buf = iov->iov_base + iov_ofs;
1343                 unsigned len;
1344                 size_t left;
1345
1346                 len = iov->iov_len - iov_ofs;
1347                 if (len > bytes)
1348                         len = bytes;
1349                 left = __copy_from_user_inatomic(vaddr, buf, len);
1350                 total += len;
1351                 bytes -= len;
1352                 vaddr += len;
1353                 if (unlikely(left)) {
1354                         total -= left;
1355                         break;
1356                 }
1357                 if (!bytes)
1358                         break;
1359                 iov++;
1360                 iov_ofs = 0;
1361         }
1362         return total;
1363 }
1364
1365 static inline void ntfs_set_next_iovec(const struct iovec **iovp,
1366                 size_t *iov_ofsp, size_t bytes)
1367 {
1368         const struct iovec *iov = *iovp;
1369         size_t iov_ofs = *iov_ofsp;
1370
1371         while (bytes) {
1372                 unsigned len;
1373
1374                 len = iov->iov_len - iov_ofs;
1375                 if (len > bytes)
1376                         len = bytes;
1377                 bytes -= len;
1378                 iov_ofs += len;
1379                 if (iov->iov_len == iov_ofs) {
1380                         iov++;
1381                         iov_ofs = 0;
1382                 }
1383         }
1384         *iovp = iov;
1385         *iov_ofsp = iov_ofs;
1386 }
1387
1388 /*
1389  * This has the same side-effects and return value as ntfs_copy_from_user().
1390  * The difference is that on a fault we need to memset the remainder of the
1391  * pages (out to offset + bytes), to emulate ntfs_copy_from_user()'s
1392  * single-segment behaviour.
1393  *
1394  * We call the same helper (__ntfs_copy_from_user_iovec_inatomic()) both
1395  * when atomic and when not atomic.  This is ok because
1396  * __ntfs_copy_from_user_iovec_inatomic() calls __copy_from_user_inatomic()
1397  * and it is ok to call this when non-atomic.
1398  * Infact, the only difference between __copy_from_user_inatomic() and
1399  * __copy_from_user() is that the latter calls might_sleep() and the former
1400  * should not zero the tail of the buffer on error.  And on many
1401  * architectures __copy_from_user_inatomic() is just defined to
1402  * __copy_from_user() so it makes no difference at all on those architectures.
1403  */
1404 static inline size_t ntfs_copy_from_user_iovec(struct page **pages,
1405                 unsigned nr_pages, unsigned ofs, const struct iovec **iov,
1406                 size_t *iov_ofs, size_t bytes)
1407 {
1408         struct page **last_page = pages + nr_pages;
1409         char *addr;
1410         size_t copied, len, total = 0;
1411
1412         do {
1413                 len = PAGE_CACHE_SIZE - ofs;
1414                 if (len > bytes)
1415                         len = bytes;
1416                 addr = kmap_atomic(*pages, KM_USER0);
1417                 copied = __ntfs_copy_from_user_iovec_inatomic(addr + ofs,
1418                                 *iov, *iov_ofs, len);
1419                 kunmap_atomic(addr, KM_USER0);
1420                 if (unlikely(copied != len)) {
1421                         /* Do it the slow way. */
1422                         addr = kmap(*pages);
1423                         copied = __ntfs_copy_from_user_iovec_inatomic(addr + ofs,
1424                                         *iov, *iov_ofs, len);
1425                         /*
1426                          * Zero the rest of the target like __copy_from_user().
1427                          */
1428                         memset(addr + ofs + copied, 0, len - copied);
1429                         kunmap(*pages);
1430                         if (unlikely(copied != len))
1431                                 goto err_out;
1432                 }
1433                 total += len;
1434                 bytes -= len;
1435                 if (!bytes)
1436                         break;
1437                 ntfs_set_next_iovec(iov, iov_ofs, len);
1438                 ofs = 0;
1439         } while (++pages < last_page);
1440 out:
1441         return total;
1442 err_out:
1443         total += copied;
1444         /* Zero the rest of the target like __copy_from_user(). */
1445         while (++pages < last_page) {
1446                 bytes -= len;
1447                 if (!bytes)
1448                         break;
1449                 len = PAGE_CACHE_SIZE;
1450                 if (len > bytes)
1451                         len = bytes;
1452                 zero_user(*pages, 0, len);
1453         }
1454         goto out;
1455 }
1456
1457 static inline void ntfs_flush_dcache_pages(struct page **pages,
1458                 unsigned nr_pages)
1459 {
1460         BUG_ON(!nr_pages);
1461         /*
1462          * Warning: Do not do the decrement at the same time as the call to
1463          * flush_dcache_page() because it is a NULL macro on i386 and hence the
1464          * decrement never happens so the loop never terminates.
1465          */
1466         do {
1467                 --nr_pages;
1468                 flush_dcache_page(pages[nr_pages]);
1469         } while (nr_pages > 0);
1470 }
1471
1472 /**
1473  * ntfs_commit_pages_after_non_resident_write - commit the received data
1474  * @pages:      array of destination pages
1475  * @nr_pages:   number of pages in @pages
1476  * @pos:        byte position in file at which the write begins
1477  * @bytes:      number of bytes to be written
1478  *
1479  * See description of ntfs_commit_pages_after_write(), below.
1480  */
1481 static inline int ntfs_commit_pages_after_non_resident_write(
1482                 struct page **pages, const unsigned nr_pages,
1483                 s64 pos, size_t bytes)
1484 {
1485         s64 end, initialized_size;
1486         struct inode *vi;
1487         ntfs_inode *ni, *base_ni;
1488         struct buffer_head *bh, *head;
1489         ntfs_attr_search_ctx *ctx;
1490         MFT_RECORD *m;
1491         ATTR_RECORD *a;
1492         unsigned long flags;
1493         unsigned blocksize, u;
1494         int err;
1495
1496         vi = pages[0]->mapping->host;
1497         ni = NTFS_I(vi);
1498         blocksize = vi->i_sb->s_blocksize;
1499         end = pos + bytes;
1500         u = 0;
1501         do {
1502                 s64 bh_pos;
1503                 struct page *page;
1504                 bool partial;
1505
1506                 page = pages[u];
1507                 bh_pos = (s64)page->index << PAGE_CACHE_SHIFT;
1508                 bh = head = page_buffers(page);
1509                 partial = false;
1510                 do {
1511                         s64 bh_end;
1512
1513                         bh_end = bh_pos + blocksize;
1514                         if (bh_end <= pos || bh_pos >= end) {
1515                                 if (!buffer_uptodate(bh))
1516                                         partial = true;
1517                         } else {
1518                                 set_buffer_uptodate(bh);
1519                                 mark_buffer_dirty(bh);
1520                         }
1521                 } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
1522                 /*
1523                  * If all buffers are now uptodate but the page is not, set the
1524                  * page uptodate.
1525                  */
1526                 if (!partial && !PageUptodate(page))
1527                         SetPageUptodate(page);
1528         } while (++u < nr_pages);
1529         /*
1530          * Finally, if we do not need to update initialized_size or i_size we
1531          * are finished.
1532          */
1533         read_lock_irqsave(&ni->size_lock, flags);
1534         initialized_size = ni->initialized_size;
1535         read_unlock_irqrestore(&ni->size_lock, flags);
1536         if (end <= initialized_size) {
1537                 ntfs_debug("Done.");
1538                 return 0;
1539         }
1540         /*
1541          * Update initialized_size/i_size as appropriate, both in the inode and
1542          * the mft record.
1543          */
1544         if (!NInoAttr(ni))
1545                 base_ni = ni;
1546         else
1547                 base_ni = ni->ext.base_ntfs_ino;
1548         /* Map, pin, and lock the mft record. */
1549         m = map_mft_record(base_ni);
1550         if (IS_ERR(m)) {
1551                 err = PTR_ERR(m);
1552                 m = NULL;
1553                 ctx = NULL;
1554                 goto err_out;
1555         }
1556         BUG_ON(!NInoNonResident(ni));
1557         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1558         if (unlikely(!ctx)) {
1559                 err = -ENOMEM;
1560                 goto err_out;
1561         }
1562         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1563                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1564         if (unlikely(err)) {
1565                 if (err == -ENOENT)
1566                         err = -EIO;
1567                 goto err_out;
1568         }
1569         a = ctx->attr;
1570         BUG_ON(!a->non_resident);
1571         write_lock_irqsave(&ni->size_lock, flags);
1572         BUG_ON(end > ni->allocated_size);
1573         ni->initialized_size = end;
1574         a->data.non_resident.initialized_size = cpu_to_sle64(end);
1575         if (end > i_size_read(vi)) {
1576                 i_size_write(vi, end);
1577                 a->data.non_resident.data_size =
1578                                 a->data.non_resident.initialized_size;
1579         }
1580         write_unlock_irqrestore(&ni->size_lock, flags);
1581         /* Mark the mft record dirty, so it gets written back. */
1582         flush_dcache_mft_record_page(ctx->ntfs_ino);
1583         mark_mft_record_dirty(ctx->ntfs_ino);
1584         ntfs_attr_put_search_ctx(ctx);
1585         unmap_mft_record(base_ni);
1586         ntfs_debug("Done.");
1587         return 0;
1588 err_out:
1589         if (ctx)
1590                 ntfs_attr_put_search_ctx(ctx);
1591         if (m)
1592                 unmap_mft_record(base_ni);
1593         ntfs_error(vi->i_sb, "Failed to update initialized_size/i_size (error "
1594                         "code %i).", err);
1595         if (err != -ENOMEM)
1596                 NVolSetErrors(ni->vol);
1597         return err;
1598 }
1599
1600 /**
1601  * ntfs_commit_pages_after_write - commit the received data
1602  * @pages:      array of destination pages
1603  * @nr_pages:   number of pages in @pages
1604  * @pos:        byte position in file at which the write begins
1605  * @bytes:      number of bytes to be written
1606  *
1607  * This is called from ntfs_file_buffered_write() with i_mutex held on the inode
1608  * (@pages[0]->mapping->host).  There are @nr_pages pages in @pages which are
1609  * locked but not kmap()ped.  The source data has already been copied into the
1610  * @page.  ntfs_prepare_pages_for_non_resident_write() has been called before
1611  * the data was copied (for non-resident attributes only) and it returned
1612  * success.
1613  *
1614  * Need to set uptodate and mark dirty all buffers within the boundary of the
1615  * write.  If all buffers in a page are uptodate we set the page uptodate, too.
1616  *
1617  * Setting the buffers dirty ensures that they get written out later when
1618  * ntfs_writepage() is invoked by the VM.
1619  *
1620  * Finally, we need to update i_size and initialized_size as appropriate both
1621  * in the inode and the mft record.
1622  *
1623  * This is modelled after fs/buffer.c::generic_commit_write(), which marks
1624  * buffers uptodate and dirty, sets the page uptodate if all buffers in the
1625  * page are uptodate, and updates i_size if the end of io is beyond i_size.  In
1626  * that case, it also marks the inode dirty.
1627  *
1628  * If things have gone as outlined in
1629  * ntfs_prepare_pages_for_non_resident_write(), we do not need to do any page
1630  * content modifications here for non-resident attributes.  For resident
1631  * attributes we need to do the uptodate bringing here which we combine with
1632  * the copying into the mft record which means we save one atomic kmap.
1633  *
1634  * Return 0 on success or -errno on error.
1635  */
1636 static int ntfs_commit_pages_after_write(struct page **pages,
1637                 const unsigned nr_pages, s64 pos, size_t bytes)
1638 {
1639         s64 end, initialized_size;
1640         loff_t i_size;
1641         struct inode *vi;
1642         ntfs_inode *ni, *base_ni;
1643         struct page *page;
1644         ntfs_attr_search_ctx *ctx;
1645         MFT_RECORD *m;
1646         ATTR_RECORD *a;
1647         char *kattr, *kaddr;
1648         unsigned long flags;
1649         u32 attr_len;
1650         int err;
1651
1652         BUG_ON(!nr_pages);
1653         BUG_ON(!pages);
1654         page = pages[0];
1655         BUG_ON(!page);
1656         vi = page->mapping->host;
1657         ni = NTFS_I(vi);
1658         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, start page "
1659                         "index 0x%lx, nr_pages 0x%x, pos 0x%llx, bytes 0x%zx.",
1660                         vi->i_ino, ni->type, page->index, nr_pages,
1661                         (long long)pos, bytes);
1662         if (NInoNonResident(ni))
1663                 return ntfs_commit_pages_after_non_resident_write(pages,
1664                                 nr_pages, pos, bytes);
1665         BUG_ON(nr_pages > 1);
1666         /*
1667          * Attribute is resident, implying it is not compressed, encrypted, or
1668          * sparse.
1669          */
1670         if (!NInoAttr(ni))
1671                 base_ni = ni;
1672         else
1673                 base_ni = ni->ext.base_ntfs_ino;
1674         BUG_ON(NInoNonResident(ni));
1675         /* Map, pin, and lock the mft record. */
1676         m = map_mft_record(base_ni);
1677         if (IS_ERR(m)) {
1678                 err = PTR_ERR(m);
1679                 m = NULL;
1680                 ctx = NULL;
1681                 goto err_out;
1682         }
1683         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1684         if (unlikely(!ctx)) {
1685                 err = -ENOMEM;
1686                 goto err_out;
1687         }
1688         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1689                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1690         if (unlikely(err)) {
1691                 if (err == -ENOENT)
1692                         err = -EIO;
1693                 goto err_out;
1694         }
1695         a = ctx->attr;
1696         BUG_ON(a->non_resident);
1697         /* The total length of the attribute value. */
1698         attr_len = le32_to_cpu(a->data.resident.value_length);
1699         i_size = i_size_read(vi);
1700         BUG_ON(attr_len != i_size);
1701         BUG_ON(pos > attr_len);
1702         end = pos + bytes;
1703         BUG_ON(end > le32_to_cpu(a->length) -
1704                         le16_to_cpu(a->data.resident.value_offset));
1705         kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
1706         kaddr = kmap_atomic(page, KM_USER0);
1707         /* Copy the received data from the page to the mft record. */
1708         memcpy(kattr + pos, kaddr + pos, bytes);
1709         /* Update the attribute length if necessary. */
1710         if (end > attr_len) {
1711                 attr_len = end;
1712                 a->data.resident.value_length = cpu_to_le32(attr_len);
1713         }
1714         /*
1715          * If the page is not uptodate, bring the out of bounds area(s)
1716          * uptodate by copying data from the mft record to the page.
1717          */
1718         if (!PageUptodate(page)) {
1719                 if (pos > 0)
1720                         memcpy(kaddr, kattr, pos);
1721                 if (end < attr_len)
1722                         memcpy(kaddr + end, kattr + end, attr_len - end);
1723                 /* Zero the region outside the end of the attribute value. */
1724                 memset(kaddr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
1725                 flush_dcache_page(page);
1726                 SetPageUptodate(page);
1727         }
1728         kunmap_atomic(kaddr, KM_USER0);
1729         /* Update initialized_size/i_size if necessary. */
1730         read_lock_irqsave(&ni->size_lock, flags);
1731         initialized_size = ni->initialized_size;
1732         BUG_ON(end > ni->allocated_size);
1733         read_unlock_irqrestore(&ni->size_lock, flags);
1734         BUG_ON(initialized_size != i_size);
1735         if (end > initialized_size) {
1736                 write_lock_irqsave(&ni->size_lock, flags);
1737                 ni->initialized_size = end;
1738                 i_size_write(vi, end);
1739                 write_unlock_irqrestore(&ni->size_lock, flags);
1740         }
1741         /* Mark the mft record dirty, so it gets written back. */
1742         flush_dcache_mft_record_page(ctx->ntfs_ino);
1743         mark_mft_record_dirty(ctx->ntfs_ino);
1744         ntfs_attr_put_search_ctx(ctx);
1745         unmap_mft_record(base_ni);
1746         ntfs_debug("Done.");
1747         return 0;
1748 err_out:
1749         if (err == -ENOMEM) {
1750                 ntfs_warning(vi->i_sb, "Error allocating memory required to "
1751                                 "commit the write.");
1752                 if (PageUptodate(page)) {
1753                         ntfs_warning(vi->i_sb, "Page is uptodate, setting "
1754                                         "dirty so the write will be retried "
1755                                         "later on by the VM.");
1756                         /*
1757                          * Put the page on mapping->dirty_pages, but leave its
1758                          * buffers' dirty state as-is.
1759                          */
1760                         __set_page_dirty_nobuffers(page);
1761                         err = 0;
1762                 } else
1763                         ntfs_error(vi->i_sb, "Page is not uptodate.  Written "
1764                                         "data has been lost.");
1765         } else {
1766                 ntfs_error(vi->i_sb, "Resident attribute commit write failed "
1767                                 "with error %i.", err);
1768                 NVolSetErrors(ni->vol);
1769         }
1770         if (ctx)
1771                 ntfs_attr_put_search_ctx(ctx);
1772         if (m)
1773                 unmap_mft_record(base_ni);
1774         return err;
1775 }
1776
1777 /**
1778  * ntfs_file_buffered_write -
1779  *
1780  * Locking: The vfs is holding ->i_mutex on the inode.
1781  */
1782 static ssize_t ntfs_file_buffered_write(struct kiocb *iocb,
1783                 const struct iovec *iov, unsigned long nr_segs,
1784                 loff_t pos, loff_t *ppos, size_t count)
1785 {
1786         struct file *file = iocb->ki_filp;
1787         struct address_space *mapping = file->f_mapping;
1788         struct inode *vi = mapping->host;
1789         ntfs_inode *ni = NTFS_I(vi);
1790         ntfs_volume *vol = ni->vol;
1791         struct page *pages[NTFS_MAX_PAGES_PER_CLUSTER];
1792         struct page *cached_page = NULL;
1793         char __user *buf = NULL;
1794         s64 end, ll;
1795         VCN last_vcn;
1796         LCN lcn;
1797         unsigned long flags;
1798         size_t bytes, iov_ofs = 0;      /* Offset in the current iovec. */
1799         ssize_t status, written;
1800         unsigned nr_pages;
1801         int err;
1802         struct pagevec lru_pvec;
1803
1804         ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, "
1805                         "pos 0x%llx, count 0x%lx.",
1806                         vi->i_ino, (unsigned)le32_to_cpu(ni->type),
1807                         (unsigned long long)pos, (unsigned long)count);
1808         if (unlikely(!count))
1809                 return 0;
1810         BUG_ON(NInoMstProtected(ni));
1811         /*
1812          * If the attribute is not an index root and it is encrypted or
1813          * compressed, we cannot write to it yet.  Note we need to check for
1814          * AT_INDEX_ALLOCATION since this is the type of both directory and
1815          * index inodes.
1816          */
1817         if (ni->type != AT_INDEX_ALLOCATION) {
1818                 /* If file is encrypted, deny access, just like NT4. */
1819                 if (NInoEncrypted(ni)) {
1820                         /*
1821                          * Reminder for later: Encrypted files are _always_
1822                          * non-resident so that the content can always be
1823                          * encrypted.
1824                          */
1825                         ntfs_debug("Denying write access to encrypted file.");
1826                         return -EACCES;
1827                 }
1828                 if (NInoCompressed(ni)) {
1829                         /* Only unnamed $DATA attribute can be compressed. */
1830                         BUG_ON(ni->type != AT_DATA);
1831                         BUG_ON(ni->name_len);
1832                         /*
1833                          * Reminder for later: If resident, the data is not
1834                          * actually compressed.  Only on the switch to non-
1835                          * resident does compression kick in.  This is in
1836                          * contrast to encrypted files (see above).
1837                          */
1838                         ntfs_error(vi->i_sb, "Writing to compressed files is "
1839                                         "not implemented yet.  Sorry.");
1840                         return -EOPNOTSUPP;
1841                 }
1842         }
1843         /*
1844          * If a previous ntfs_truncate() failed, repeat it and abort if it
1845          * fails again.
1846          */
1847         if (unlikely(NInoTruncateFailed(ni))) {
1848                 down_write(&vi->i_alloc_sem);
1849                 err = ntfs_truncate(vi);
1850                 up_write(&vi->i_alloc_sem);
1851                 if (err || NInoTruncateFailed(ni)) {
1852                         if (!err)
1853                                 err = -EIO;
1854                         ntfs_error(vol->sb, "Cannot perform write to inode "
1855                                         "0x%lx, attribute type 0x%x, because "
1856                                         "ntfs_truncate() failed (error code "
1857                                         "%i).", vi->i_ino,
1858                                         (unsigned)le32_to_cpu(ni->type), err);
1859                         return err;
1860                 }
1861         }
1862         /* The first byte after the write. */
1863         end = pos + count;
1864         /*
1865          * If the write goes beyond the allocated size, extend the allocation
1866          * to cover the whole of the write, rounded up to the nearest cluster.
1867          */
1868         read_lock_irqsave(&ni->size_lock, flags);
1869         ll = ni->allocated_size;
1870         read_unlock_irqrestore(&ni->size_lock, flags);
1871         if (end > ll) {
1872                 /* Extend the allocation without changing the data size. */
1873                 ll = ntfs_attr_extend_allocation(ni, end, -1, pos);
1874                 if (likely(ll >= 0)) {
1875                         BUG_ON(pos >= ll);
1876                         /* If the extension was partial truncate the write. */
1877                         if (end > ll) {
1878                                 ntfs_debug("Truncating write to inode 0x%lx, "
1879                                                 "attribute type 0x%x, because "
1880                                                 "the allocation was only "
1881                                                 "partially extended.",
1882                                                 vi->i_ino, (unsigned)
1883                                                 le32_to_cpu(ni->type));
1884                                 end = ll;
1885                                 count = ll - pos;
1886                         }
1887                 } else {
1888                         err = ll;
1889                         read_lock_irqsave(&ni->size_lock, flags);
1890                         ll = ni->allocated_size;
1891                         read_unlock_irqrestore(&ni->size_lock, flags);
1892                         /* Perform a partial write if possible or fail. */
1893                         if (pos < ll) {
1894                                 ntfs_debug("Truncating write to inode 0x%lx, "
1895                                                 "attribute type 0x%x, because "
1896                                                 "extending the allocation "
1897                                                 "failed (error code %i).",
1898                                                 vi->i_ino, (unsigned)
1899                                                 le32_to_cpu(ni->type), err);
1900                                 end = ll;
1901                                 count = ll - pos;
1902                         } else {
1903                                 ntfs_error(vol->sb, "Cannot perform write to "
1904                                                 "inode 0x%lx, attribute type "
1905                                                 "0x%x, because extending the "
1906                                                 "allocation failed (error "
1907                                                 "code %i).", vi->i_ino,
1908                                                 (unsigned)
1909                                                 le32_to_cpu(ni->type), err);
1910                                 return err;
1911                         }
1912                 }
1913         }
1914         pagevec_init(&lru_pvec, 0);
1915         written = 0;
1916         /*
1917          * If the write starts beyond the initialized size, extend it up to the
1918          * beginning of the write and initialize all non-sparse space between
1919          * the old initialized size and the new one.  This automatically also
1920          * increments the vfs inode->i_size to keep it above or equal to the
1921          * initialized_size.
1922          */
1923         read_lock_irqsave(&ni->size_lock, flags);
1924         ll = ni->initialized_size;
1925         read_unlock_irqrestore(&ni->size_lock, flags);
1926         if (pos > ll) {
1927                 err = ntfs_attr_extend_initialized(ni, pos, &cached_page,
1928                                 &lru_pvec);
1929                 if (err < 0) {
1930                         ntfs_error(vol->sb, "Cannot perform write to inode "
1931                                         "0x%lx, attribute type 0x%x, because "
1932                                         "extending the initialized size "
1933                                         "failed (error code %i).", vi->i_ino,
1934                                         (unsigned)le32_to_cpu(ni->type), err);
1935                         status = err;
1936                         goto err_out;
1937                 }
1938         }
1939         /*
1940          * Determine the number of pages per cluster for non-resident
1941          * attributes.
1942          */
1943         nr_pages = 1;
1944         if (vol->cluster_size > PAGE_CACHE_SIZE && NInoNonResident(ni))
1945                 nr_pages = vol->cluster_size >> PAGE_CACHE_SHIFT;
1946         /* Finally, perform the actual write. */
1947         last_vcn = -1;
1948         if (likely(nr_segs == 1))
1949                 buf = iov->iov_base;
1950         do {
1951                 VCN vcn;
1952                 pgoff_t idx, start_idx;
1953                 unsigned ofs, do_pages, u;
1954                 size_t copied;
1955
1956                 start_idx = idx = pos >> PAGE_CACHE_SHIFT;
1957                 ofs = pos & ~PAGE_CACHE_MASK;
1958                 bytes = PAGE_CACHE_SIZE - ofs;
1959                 do_pages = 1;
1960                 if (nr_pages > 1) {
1961                         vcn = pos >> vol->cluster_size_bits;
1962                         if (vcn != last_vcn) {
1963                                 last_vcn = vcn;
1964                                 /*
1965                                  * Get the lcn of the vcn the write is in.  If
1966                                  * it is a hole, need to lock down all pages in
1967                                  * the cluster.
1968                                  */
1969                                 down_read(&ni->runlist.lock);
1970                                 lcn = ntfs_attr_vcn_to_lcn_nolock(ni, pos >>
1971                                                 vol->cluster_size_bits, false);
1972                                 up_read(&ni->runlist.lock);
1973                                 if (unlikely(lcn < LCN_HOLE)) {
1974                                         status = -EIO;
1975                                         if (lcn == LCN_ENOMEM)
1976                                                 status = -ENOMEM;
1977                                         else
1978                                                 ntfs_error(vol->sb, "Cannot "
1979                                                         "perform write to "
1980                                                         "inode 0x%lx, "
1981                                                         "attribute type 0x%x, "
1982                                                         "because the attribute "
1983                                                         "is corrupt.",
1984                                                         vi->i_ino, (unsigned)
1985                                                         le32_to_cpu(ni->type));
1986                                         break;
1987                                 }
1988                                 if (lcn == LCN_HOLE) {
1989                                         start_idx = (pos & ~(s64)
1990                                                         vol->cluster_size_mask)
1991                                                         >> PAGE_CACHE_SHIFT;
1992                                         bytes = vol->cluster_size - (pos &
1993                                                         vol->cluster_size_mask);
1994                                         do_pages = nr_pages;
1995                                 }
1996                         }
1997                 }
1998                 if (bytes > count)
1999                         bytes = count;
2000                 /*
2001                  * Bring in the user page(s) that we will copy from _first_.
2002                  * Otherwise there is a nasty deadlock on copying from the same
2003                  * page(s) as we are writing to, without it/them being marked
2004                  * up-to-date.  Note, at present there is nothing to stop the
2005                  * pages being swapped out between us bringing them into memory
2006                  * and doing the actual copying.
2007                  */
2008                 if (likely(nr_segs == 1))
2009                         ntfs_fault_in_pages_readable(buf, bytes);
2010                 else
2011                         ntfs_fault_in_pages_readable_iovec(iov, iov_ofs, bytes);
2012                 /* Get and lock @do_pages starting at index @start_idx. */
2013                 status = __ntfs_grab_cache_pages(mapping, start_idx, do_pages,
2014                                 pages, &cached_page, &lru_pvec);
2015                 if (unlikely(status))
2016                         break;
2017                 /*
2018                  * For non-resident attributes, we need to fill any holes with
2019                  * actual clusters and ensure all bufferes are mapped.  We also
2020                  * need to bring uptodate any buffers that are only partially
2021                  * being written to.
2022                  */
2023                 if (NInoNonResident(ni)) {
2024                         status = ntfs_prepare_pages_for_non_resident_write(
2025                                         pages, do_pages, pos, bytes);
2026                         if (unlikely(status)) {
2027                                 loff_t i_size;
2028
2029                                 do {
2030                                         unlock_page(pages[--do_pages]);
2031                                         page_cache_release(pages[do_pages]);
2032                                 } while (do_pages);
2033                                 /*
2034                                  * The write preparation may have instantiated
2035                                  * allocated space outside i_size.  Trim this
2036                                  * off again.  We can ignore any errors in this
2037                                  * case as we will just be waisting a bit of
2038                                  * allocated space, which is not a disaster.
2039                                  */
2040                                 i_size = i_size_read(vi);
2041                                 if (pos + bytes > i_size)
2042                                         vmtruncate(vi, i_size);
2043                                 break;
2044                         }
2045                 }
2046                 u = (pos >> PAGE_CACHE_SHIFT) - pages[0]->index;
2047                 if (likely(nr_segs == 1)) {
2048                         copied = ntfs_copy_from_user(pages + u, do_pages - u,
2049                                         ofs, buf, bytes);
2050                         buf += copied;
2051                 } else
2052                         copied = ntfs_copy_from_user_iovec(pages + u,
2053                                         do_pages - u, ofs, &iov, &iov_ofs,
2054                                         bytes);
2055                 ntfs_flush_dcache_pages(pages + u, do_pages - u);
2056                 status = ntfs_commit_pages_after_write(pages, do_pages, pos,
2057                                 bytes);
2058                 if (likely(!status)) {
2059                         written += copied;
2060                         count -= copied;
2061                         pos += copied;
2062                         if (unlikely(copied != bytes))
2063                                 status = -EFAULT;
2064                 }
2065                 do {
2066                         unlock_page(pages[--do_pages]);
2067                         mark_page_accessed(pages[do_pages]);
2068                         page_cache_release(pages[do_pages]);
2069                 } while (do_pages);
2070                 if (unlikely(status))
2071                         break;
2072                 balance_dirty_pages_ratelimited(mapping);
2073                 cond_resched();
2074         } while (count);
2075 err_out:
2076         *ppos = pos;
2077         if (cached_page)
2078                 page_cache_release(cached_page);
2079         /* For now, when the user asks for O_SYNC, we actually give O_DSYNC. */
2080         if (likely(!status)) {
2081                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(vi))) {
2082                         if (!mapping->a_ops->writepage || !is_sync_kiocb(iocb))
2083                                 status = generic_osync_inode(vi, mapping,
2084                                                 OSYNC_METADATA|OSYNC_DATA);
2085                 }
2086         }
2087         pagevec_lru_add(&lru_pvec);
2088         ntfs_debug("Done.  Returning %s (written 0x%lx, status %li).",
2089                         written ? "written" : "status", (unsigned long)written,
2090                         (long)status);
2091         return written ? written : status;
2092 }
2093
2094 /**
2095  * ntfs_file_aio_write_nolock -
2096  */
2097 static ssize_t ntfs_file_aio_write_nolock(struct kiocb *iocb,
2098                 const struct iovec *iov, unsigned long nr_segs, loff_t *ppos)
2099 {
2100         struct file *file = iocb->ki_filp;
2101         struct address_space *mapping = file->f_mapping;
2102         struct inode *inode = mapping->host;
2103         loff_t pos;
2104         size_t count;           /* after file limit checks */
2105         ssize_t written, err;
2106
2107         count = 0;
2108         err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
2109         if (err)
2110                 return err;
2111         pos = *ppos;
2112         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2113         /* We can write back this queue in page reclaim. */
2114         current->backing_dev_info = mapping->backing_dev_info;
2115         written = 0;
2116         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2117         if (err)
2118                 goto out;
2119         if (!count)
2120                 goto out;
2121         err = file_remove_suid(file);
2122         if (err)
2123                 goto out;
2124         file_update_time(file);
2125         written = ntfs_file_buffered_write(iocb, iov, nr_segs, pos, ppos,
2126                         count);
2127 out:
2128         current->backing_dev_info = NULL;
2129         return written ? written : err;
2130 }
2131
2132 /**
2133  * ntfs_file_aio_write -
2134  */
2135 static ssize_t ntfs_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2136                 unsigned long nr_segs, loff_t pos)
2137 {
2138         struct file *file = iocb->ki_filp;
2139         struct address_space *mapping = file->f_mapping;
2140         struct inode *inode = mapping->host;
2141         ssize_t ret;
2142
2143         BUG_ON(iocb->ki_pos != pos);
2144
2145         mutex_lock(&inode->i_mutex);
2146         ret = ntfs_file_aio_write_nolock(iocb, iov, nr_segs, &iocb->ki_pos);
2147         mutex_unlock(&inode->i_mutex);
2148         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2149                 int err = sync_page_range(inode, mapping, pos, ret);
2150                 if (err < 0)
2151                         ret = err;
2152         }
2153         return ret;
2154 }
2155
2156 /**
2157  * ntfs_file_writev -
2158  *
2159  * Basically the same as generic_file_writev() except that it ends up calling
2160  * ntfs_file_aio_write_nolock() instead of __generic_file_aio_write_nolock().
2161  */
2162 static ssize_t ntfs_file_writev(struct file *file, const struct iovec *iov,
2163                 unsigned long nr_segs, loff_t *ppos)
2164 {
2165         struct address_space *mapping = file->f_mapping;
2166         struct inode *inode = mapping->host;
2167         struct kiocb kiocb;
2168         ssize_t ret;
2169
2170         mutex_lock(&inode->i_mutex);
2171         init_sync_kiocb(&kiocb, file);
2172         ret = ntfs_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2173         if (ret == -EIOCBQUEUED)
2174                 ret = wait_on_sync_kiocb(&kiocb);
2175         mutex_unlock(&inode->i_mutex);
2176         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2177                 int err = sync_page_range(inode, mapping, *ppos - ret, ret);
2178                 if (err < 0)
2179                         ret = err;
2180         }
2181         return ret;
2182 }
2183
2184 /**
2185  * ntfs_file_write - simple wrapper for ntfs_file_writev()
2186  */
2187 static ssize_t ntfs_file_write(struct file *file, const char __user *buf,
2188                 size_t count, loff_t *ppos)
2189 {
2190         struct iovec local_iov = { .iov_base = (void __user *)buf,
2191                                    .iov_len = count };
2192
2193         return ntfs_file_writev(file, &local_iov, 1, ppos);
2194 }
2195
2196 /**
2197  * ntfs_file_fsync - sync a file to disk
2198  * @filp:       file to be synced
2199  * @dentry:     dentry describing the file to sync
2200  * @datasync:   if non-zero only flush user data and not metadata
2201  *
2202  * Data integrity sync of a file to disk.  Used for fsync, fdatasync, and msync
2203  * system calls.  This function is inspired by fs/buffer.c::file_fsync().
2204  *
2205  * If @datasync is false, write the mft record and all associated extent mft
2206  * records as well as the $DATA attribute and then sync the block device.
2207  *
2208  * If @datasync is true and the attribute is non-resident, we skip the writing
2209  * of the mft record and all associated extent mft records (this might still
2210  * happen due to the write_inode_now() call).
2211  *
2212  * Also, if @datasync is true, we do not wait on the inode to be written out
2213  * but we always wait on the page cache pages to be written out.
2214  *
2215  * Note: In the past @filp could be NULL so we ignore it as we don't need it
2216  * anyway.
2217  *
2218  * Locking: Caller must hold i_mutex on the inode.
2219  *
2220  * TODO: We should probably also write all attribute/index inodes associated
2221  * with this inode but since we have no simple way of getting to them we ignore
2222  * this problem for now.
2223  */
2224 static int ntfs_file_fsync(struct file *filp, struct dentry *dentry,
2225                 int datasync)
2226 {
2227         struct inode *vi = dentry->d_inode;
2228         int err, ret = 0;
2229
2230         ntfs_debug("Entering for inode 0x%lx.", vi->i_ino);
2231         BUG_ON(S_ISDIR(vi->i_mode));
2232         if (!datasync || !NInoNonResident(NTFS_I(vi)))
2233                 ret = ntfs_write_inode(vi, 1);
2234         write_inode_now(vi, !datasync);
2235         /*
2236          * NOTE: If we were to use mapping->private_list (see ext2 and
2237          * fs/buffer.c) for dirty blocks then we could optimize the below to be
2238          * sync_mapping_buffers(vi->i_mapping).
2239          */
2240         err = sync_blockdev(vi->i_sb->s_bdev);
2241         if (unlikely(err && !ret))
2242                 ret = err;
2243         if (likely(!ret))
2244                 ntfs_debug("Done.");
2245         else
2246                 ntfs_warning(vi->i_sb, "Failed to f%ssync inode 0x%lx.  Error "
2247                                 "%u.", datasync ? "data" : "", vi->i_ino, -ret);
2248         return ret;
2249 }
2250
2251 #endif /* NTFS_RW */
2252
2253 const struct file_operations ntfs_file_ops = {
2254         .llseek         = generic_file_llseek,   /* Seek inside file. */
2255         .read           = do_sync_read,          /* Read from file. */
2256         .aio_read       = generic_file_aio_read, /* Async read from file. */
2257 #ifdef NTFS_RW
2258         .write          = ntfs_file_write,       /* Write to file. */
2259         .aio_write      = ntfs_file_aio_write,   /* Async write to file. */
2260         /*.release      = ,*/                    /* Last file is closed.  See
2261                                                     fs/ext2/file.c::
2262                                                     ext2_release_file() for
2263                                                     how to use this to discard
2264                                                     preallocated space for
2265                                                     write opened files. */
2266         .fsync          = ntfs_file_fsync,       /* Sync a file to disk. */
2267         /*.aio_fsync    = ,*/                    /* Sync all outstanding async
2268                                                     i/o operations on a
2269                                                     kiocb. */
2270 #endif /* NTFS_RW */
2271         /*.ioctl        = ,*/                    /* Perform function on the
2272                                                     mounted filesystem. */
2273         .mmap           = generic_file_mmap,     /* Mmap file. */
2274         .open           = ntfs_file_open,        /* Open file. */
2275         .splice_read    = generic_file_splice_read /* Zero-copy data send with
2276                                                     the data source being on
2277                                                     the ntfs partition.  We do
2278                                                     not need to care about the
2279                                                     data destination. */
2280         /*.sendpage     = ,*/                    /* Zero-copy data send with
2281                                                     the data destination being
2282                                                     on the ntfs partition.  We
2283                                                     do not need to care about
2284                                                     the data source. */
2285 };
2286
2287 const struct inode_operations ntfs_file_inode_ops = {
2288 #ifdef NTFS_RW
2289         .truncate       = ntfs_truncate_vfs,
2290         .setattr        = ntfs_setattr,
2291 #endif /* NTFS_RW */
2292 };
2293
2294 const struct file_operations ntfs_empty_file_ops = {};
2295
2296 const struct inode_operations ntfs_empty_inode_ops = {};