fs/ntfs/file.c

   1 /*
   2  * file.c - NTFS kernel file operations.  Part of the Linux-NTFS project.
   3  *
   4  * Copyright (c) 2001-2006 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 -EFBIG;
  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((ptrdiff_t __user)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                 struct page *page = pages[u];
 535                 /*
 536                  * create_empty_buffers() will create uptodate/dirty buffers if
 537                  * the page is uptodate/dirty.
 538                  */
 539                 if (!page_has_buffers(page)) {
 540                         create_empty_buffers(page, blocksize, 0);
 541                         if (unlikely(!page_has_buffers(page)))
 542                                 return -ENOMEM;
 543                 }
 544         } while (++u < nr_pages);
 545         rl_write_locked = false;
 546         rl = NULL;
 547         err = 0;
 548         vcn = lcn = -1;
 549         vcn_len = 0;
 550         lcn_block = -1;
 551         was_hole = false;
 552         cpos = pos >> vol->cluster_size_bits;
 553         end = pos + bytes;
 554         cend = (end + vol->cluster_size - 1) >> vol->cluster_size_bits;
 555         /*
 556          * Loop over each page and for each page over each buffer.  Use goto to
 557          * reduce indentation.
 558          */
 559         u = 0;
 560 do_next_page:
 561         page = pages[u];
 562         bh_pos = (s64)page->index << PAGE_CACHE_SHIFT;
 563         bh = head = page_buffers(page);
 564         do {
 565                 VCN cdelta;
 566                 s64 bh_end;
 567                 unsigned bh_cofs;
 568
 569                 /* Clear buffer_new on all buffers to reinitialise state. */
 570                 if (buffer_new(bh))
 571                         clear_buffer_new(bh);
 572                 bh_end = bh_pos + blocksize;
 573                 bh_cpos = bh_pos >> vol->cluster_size_bits;
 574                 bh_cofs = bh_pos & vol->cluster_size_mask;
 575                 if (buffer_mapped(bh)) {
 576                         /*
 577                          * The buffer is already mapped.  If it is uptodate,
 578                          * ignore it.
 579                          */
 580                         if (buffer_uptodate(bh))
 581                                 continue;
 582                         /*
 583                          * The buffer is not uptodate.  If the page is uptodate
 584                          * set the buffer uptodate and otherwise ignore it.
 585                          */
 586                         if (PageUptodate(page)) {
 587                                 set_buffer_uptodate(bh);
 588                                 continue;
 589                         }
 590                         /*
 591                          * Neither the page nor the buffer are uptodate.  If
 592                          * the buffer is only partially being written to, we
 593                          * need to read it in before the write, i.e. now.
 594                          */
 595                         if ((bh_pos < pos && bh_end > pos) ||
 596                                         (bh_pos < end && bh_end > end)) {
 597                                 /*
 598                                  * If the buffer is fully or partially within
 599                                  * the initialized size, do an actual read.
 600                                  * Otherwise, simply zero the buffer.
 601                                  */
 602                                 read_lock_irqsave(&ni->size_lock, flags);
 603                                 initialized_size = ni->initialized_size;
 604                                 read_unlock_irqrestore(&ni->size_lock, flags);
 605                                 if (bh_pos < initialized_size) {
 606                                         ntfs_submit_bh_for_read(bh);
 607                                         *wait_bh++ = bh;
 608                                 } else {
 609                                         zero_user_page(page, bh_offset(bh),
 610                                                         blocksize, KM_USER0);
 611                                         set_buffer_uptodate(bh);
 612                                 }
 613                         }
 614                         continue;
 615                 }
 616                 /* Unmapped buffer.  Need to map it. */
 617                 bh->b_bdev = vol->sb->s_bdev;
 618                 /*
 619                  * If the current buffer is in the same clusters as the map
 620                  * cache, there is no need to check the runlist again.  The
 621                  * map cache is made up of @vcn, which is the first cached file
 622                  * cluster, @vcn_len which is the number of cached file
 623                  * clusters, @lcn is the device cluster corresponding to @vcn,
 624                  * and @lcn_block is the block number corresponding to @lcn.
 625                  */
 626                 cdelta = bh_cpos - vcn;
 627                 if (likely(!cdelta || (cdelta > 0 && cdelta < vcn_len))) {
 628 map_buffer_cached:
 629                         BUG_ON(lcn < 0);
 630                         bh->b_blocknr = lcn_block +
 631                                         (cdelta << (vol->cluster_size_bits -
 632                                         blocksize_bits)) +
 633                                         (bh_cofs >> blocksize_bits);
 634                         set_buffer_mapped(bh);
 635                         /*
 636                          * If the page is uptodate so is the buffer.  If the
 637                          * buffer is fully outside the write, we ignore it if
 638                          * it was already allocated and we mark it dirty so it
 639                          * gets written out if we allocated it.  On the other
 640                          * hand, if we allocated the buffer but we are not
 641                          * marking it dirty we set buffer_new so we can do
 642                          * error recovery.
 643                          */
 644                         if (PageUptodate(page)) {
 645                                 if (!buffer_uptodate(bh))
 646                                         set_buffer_uptodate(bh);
 647                                 if (unlikely(was_hole)) {
 648                                         /* We allocated the buffer. */
 649                                         unmap_underlying_metadata(bh->b_bdev,
 650                                                         bh->b_blocknr);
 651                                         if (bh_end <= pos || bh_pos >= end)
 652                                                 mark_buffer_dirty(bh);
 653                                         else
 654                                                 set_buffer_new(bh);
 655                                 }
 656                                 continue;
 657                         }
 658                         /* Page is _not_ uptodate. */
 659                         if (likely(!was_hole)) {
 660                                 /*
 661                                  * Buffer was already allocated.  If it is not
 662                                  * uptodate and is only partially being written
 663                                  * to, we need to read it in before the write,
 664                                  * i.e. now.
 665                                  */
 666                                 if (!buffer_uptodate(bh) && bh_pos < end &&
 667                                                 bh_end > pos &&
 668                                                 (bh_pos < pos ||
 669                                                 bh_end > end)) {
 670                                         /*
 671                                          * If the buffer is fully or partially
 672                                          * within the initialized size, do an
 673                                          * actual read.  Otherwise, simply zero
 674                                          * the buffer.
 675                                          */
 676                                         read_lock_irqsave(&ni->size_lock,
 677                                                         flags);
 678                                         initialized_size = ni->initialized_size;
 679                                         read_unlock_irqrestore(&ni->size_lock,
 680                                                         flags);
 681                                         if (bh_pos < initialized_size) {
 682                                                 ntfs_submit_bh_for_read(bh);
 683                                                 *wait_bh++ = bh;
 684                                         } else {
 685                                                 zero_user_page(page,
 686                                                         bh_offset(bh),
 687                                                         blocksize, KM_USER0);
 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(page, bh_offset(bh),
 706                                                         blocksize, KM_USER0);
 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(page, bh_offset(bh), blocksize,
 746                                                 KM_USER0);
 747                                 set_buffer_uptodate(bh);
 748                         }
 749                         continue;
 750                 }
 751                 is_retry = false;
 752                 if (!rl) {
 753                         down_read(&ni->runlist.lock);
 754 retry_remap:
 755                         rl = ni->runlist.rl;
 756                 }
 757                 if (likely(rl != NULL)) {
 758                         /* Seek to element containing target cluster. */
 759                         while (rl->length && rl[1].vcn <= bh_cpos)
 760                                 rl++;
 761                         lcn = ntfs_rl_vcn_to_lcn(rl, bh_cpos);
 762                         if (likely(lcn >= 0)) {
 763                                 /*
 764                                  * Successful remap, setup the map cache and
 765                                  * use that to deal with the buffer.
 766                                  */
 767                                 was_hole = false;
 768                                 vcn = bh_cpos;
 769                                 vcn_len = rl[1].vcn - vcn;
 770                                 lcn_block = lcn << (vol->cluster_size_bits -
 771                                                 blocksize_bits);
 772                                 cdelta = 0;
 773                                 /*
 774                                  * If the number of remaining clusters touched
 775                                  * by the write is smaller or equal to the
 776                                  * number of cached clusters, unlock the
 777                                  * runlist as the map cache will be used from
 778                                  * now on.
 779                                  */
 780                                 if (likely(vcn + vcn_len >= cend)) {
 781                                         if (rl_write_locked) {
 782                                                 up_write(&ni->runlist.lock);
 783                                                 rl_write_locked = false;
 784                                         } else
 785                                                 up_read(&ni->runlist.lock);
 786                                         rl = NULL;
 787                                 }
 788                                 goto map_buffer_cached;
 789                         }
 790                 } else
 791                         lcn = LCN_RL_NOT_MAPPED;
 792                 /*
 793                  * If it is not a hole and not out of bounds, the runlist is
 794                  * probably unmapped so try to map it now.
 795                  */
 796                 if (unlikely(lcn != LCN_HOLE && lcn != LCN_ENOENT)) {
 797                         if (likely(!is_retry && lcn == LCN_RL_NOT_MAPPED)) {
 798                                 /* Attempt to map runlist. */
 799                                 if (!rl_write_locked) {
 800                                         /*
 801                                          * We need the runlist locked for
 802                                          * writing, so if it is locked for
 803                                          * reading relock it now and retry in
 804                                          * case it changed whilst we dropped
 805                                          * the lock.
 806                                          */
 807                                         up_read(&ni->runlist.lock);
 808                                         down_write(&ni->runlist.lock);
 809                                         rl_write_locked = true;
 810                                         goto retry_remap;
 811                                 }
 812                                 err = ntfs_map_runlist_nolock(ni, bh_cpos,
 813                                                 NULL);
 814                                 if (likely(!err)) {
 815                                         is_retry = true;
 816                                         goto retry_remap;
 817                                 }
 818                                 /*
 819                                  * If @vcn is out of bounds, pretend @lcn is
 820                                  * LCN_ENOENT.  As long as the buffer is out
 821                                  * of bounds this will work fine.
 822                                  */
 823                                 if (err == -ENOENT) {
 824                                         lcn = LCN_ENOENT;
 825                                         err = 0;
 826                                         goto rl_not_mapped_enoent;
 827                                 }
 828                         } else
 829                                 err = -EIO;
 830                         /* Failed to map the buffer, even after retrying. */
 831                         bh->b_blocknr = -1;
 832                         ntfs_error(vol->sb, "Failed to write to inode 0x%lx, "
 833                                         "attribute type 0x%x, vcn 0x%llx, "
 834                                         "vcn offset 0x%x, because its "
 835                                         "location on disk could not be "
 836                                         "determined%s (error code %i).",
 837                                         ni->mft_no, ni->type,
 838                                         (unsigned long long)bh_cpos,
 839                                         (unsigned)bh_pos &
 840                                         vol->cluster_size_mask,
 841                                         is_retry ? " even after retrying" : "",
 842                                         err);
 843                         break;
 844                 }
 845 rl_not_mapped_enoent:
 846                 /*
 847                  * The buffer is in a hole or out of bounds.  We need to fill
 848                  * the hole, unless the buffer is in a cluster which is not
 849                  * touched by the write, in which case we just leave the buffer
 850                  * unmapped.  This can only happen when the cluster size is
 851                  * less than the page cache size.
 852                  */
 853                 if (unlikely(vol->cluster_size < PAGE_CACHE_SIZE)) {
 854                         bh_cend = (bh_end + vol->cluster_size - 1) >>
 855                                         vol->cluster_size_bits;
 856                         if ((bh_cend <= cpos || bh_cpos >= cend)) {
 857                                 bh->b_blocknr = -1;
 858                                 /*
 859                                  * If the buffer is uptodate we skip it.  If it
 860                                  * is not but the page is uptodate, we can set
 861                                  * the buffer uptodate.  If the page is not
 862                                  * uptodate, we can clear the buffer and set it
 863                                  * uptodate.  Whether this is worthwhile is
 864                                  * debatable and this could be removed.
 865                                  */
 866                                 if (PageUptodate(page)) {
 867                                         if (!buffer_uptodate(bh))
 868                                                 set_buffer_uptodate(bh);
 869                                 } else if (!buffer_uptodate(bh)) {
 870                                         zero_user_page(page, bh_offset(bh),
 871                                                         blocksize, KM_USER0);
 872                                         set_buffer_uptodate(bh);
 873                                 }
 874                                 continue;
 875                         }
 876                 }
 877                 /*
 878                  * Out of bounds buffer is invalid if it was not really out of
 879                  * bounds.
 880                  */
 881                 BUG_ON(lcn != LCN_HOLE);
 882                 /*
 883                  * We need the runlist locked for writing, so if it is locked
 884                  * for reading relock it now and retry in case it changed
 885                  * whilst we dropped the lock.
 886                  */
 887                 BUG_ON(!rl);
 888                 if (!rl_write_locked) {
 889                         up_read(&ni->runlist.lock);
 890                         down_write(&ni->runlist.lock);
 891                         rl_write_locked = true;
 892                         goto retry_remap;
 893                 }
 894                 /* Find the previous last allocated cluster. */
 895                 BUG_ON(rl->lcn != LCN_HOLE);
 896                 lcn = -1;
 897                 rl2 = rl;
 898                 while (--rl2 >= ni->runlist.rl) {
 899                         if (rl2->lcn >= 0) {
 900                                 lcn = rl2->lcn + rl2->length;
 901                                 break;
 902                         }
 903                 }
 904                 rl2 = ntfs_cluster_alloc(vol, bh_cpos, 1, lcn, DATA_ZONE,
 905                                 false);
 906                 if (IS_ERR(rl2)) {
 907                         err = PTR_ERR(rl2);
 908                         ntfs_debug("Failed to allocate cluster, error code %i.",
 909                                         err);
 910                         break;
 911                 }
 912                 lcn = rl2->lcn;
 913                 rl = ntfs_runlists_merge(ni->runlist.rl, rl2);
 914                 if (IS_ERR(rl)) {
 915                         err = PTR_ERR(rl);
 916                         if (err != -ENOMEM)
 917                                 err = -EIO;
 918                         if (ntfs_cluster_free_from_rl(vol, rl2)) {
 919                                 ntfs_error(vol->sb, "Failed to release "
 920                                                 "allocated cluster in error "
 921                                                 "code path.  Run chkdsk to "
 922                                                 "recover the lost cluster.");
 923                                 NVolSetErrors(vol);
 924                         }
 925                         ntfs_free(rl2);
 926                         break;
 927                 }
 928                 ni->runlist.rl = rl;
 929                 status.runlist_merged = 1;
 930                 ntfs_debug("Allocated cluster, lcn 0x%llx.",
 931                                 (unsigned long long)lcn);
 932                 /* Map and lock the mft record and get the attribute record. */
 933                 if (!NInoAttr(ni))
 934                         base_ni = ni;
 935                 else
 936                         base_ni = ni->ext.base_ntfs_ino;
 937                 m = map_mft_record(base_ni);
 938                 if (IS_ERR(m)) {
 939                         err = PTR_ERR(m);
 940                         break;
 941                 }
 942                 ctx = ntfs_attr_get_search_ctx(base_ni, m);
 943                 if (unlikely(!ctx)) {
 944                         err = -ENOMEM;
 945                         unmap_mft_record(base_ni);
 946                         break;
 947                 }
 948                 status.mft_attr_mapped = 1;
 949                 err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
 950                                 CASE_SENSITIVE, bh_cpos, NULL, 0, ctx);
 951                 if (unlikely(err)) {
 952                         if (err == -ENOENT)
 953                                 err = -EIO;
 954                         break;
 955                 }
 956                 m = ctx->mrec;
 957                 a = ctx->attr;
 958                 /*
 959                  * Find the runlist element with which the attribute extent
 960                  * starts.  Note, we cannot use the _attr_ version because we
 961                  * have mapped the mft record.  That is ok because we know the
 962                  * runlist fragment must be mapped already to have ever gotten
 963                  * here, so we can just use the _rl_ version.
 964                  */
 965                 vcn = sle64_to_cpu(a->data.non_resident.lowest_vcn);
 966                 rl2 = ntfs_rl_find_vcn_nolock(rl, vcn);
 967                 BUG_ON(!rl2);
 968                 BUG_ON(!rl2->length);
 969                 BUG_ON(rl2->lcn < LCN_HOLE);
 970                 highest_vcn = sle64_to_cpu(a->data.non_resident.highest_vcn);
 971                 /*
 972                  * If @highest_vcn is zero, calculate the real highest_vcn
 973                  * (which can really be zero).
 974                  */
 975                 if (!highest_vcn)
 976                         highest_vcn = (sle64_to_cpu(
 977                                         a->data.non_resident.allocated_size) >>
 978                                         vol->cluster_size_bits) - 1;
 979                 /*
 980                  * Determine the size of the mapping pairs array for the new
 981                  * extent, i.e. the old extent with the hole filled.
 982                  */
 983                 mp_size = ntfs_get_size_for_mapping_pairs(vol, rl2, vcn,
 984                                 highest_vcn);
 985                 if (unlikely(mp_size <= 0)) {
 986                         if (!(err = mp_size))
 987                                 err = -EIO;
 988                         ntfs_debug("Failed to get size for mapping pairs "
 989                                         "array, error code %i.", err);
 990                         break;
 991                 }
 992                 /*
 993                  * Resize the attribute record to fit the new mapping pairs
 994                  * array.
 995                  */
 996                 attr_rec_len = le32_to_cpu(a->length);
 997                 err = ntfs_attr_record_resize(m, a, mp_size + le16_to_cpu(
 998                                 a->data.non_resident.mapping_pairs_offset));
 999                 if (unlikely(err)) {
1000                         BUG_ON(err != -ENOSPC);
1001                         // TODO: Deal with this by using the current attribute
1002                         // and fill it with as much of the mapping pairs
1003                         // array as possible.  Then loop over each attribute
1004                         // extent rewriting the mapping pairs arrays as we go
1005                         // along and if when we reach the end we have not
1006                         // enough space, try to resize the last attribute
1007                         // extent and if even that fails, add a new attribute
1008                         // extent.
1009                         // We could also try to resize at each step in the hope
1010                         // that we will not need to rewrite every single extent.
1011                         // Note, we may need to decompress some extents to fill
1012                         // the runlist as we are walking the extents...
1013                         ntfs_error(vol->sb, "Not enough space in the mft "
1014                                         "record for the extended attribute "
1015                                         "record.  This case is not "
1016                                         "implemented yet.");
1017                         err = -EOPNOTSUPP;
1018                         break ;
1019                 }
1020                 status.mp_rebuilt = 1;
1021                 /*
1022                  * Generate the mapping pairs array directly into the attribute
1023                  * record.
1024                  */
1025                 err = ntfs_mapping_pairs_build(vol, (u8*)a + le16_to_cpu(
1026                                 a->data.non_resident.mapping_pairs_offset),
1027                                 mp_size, rl2, vcn, highest_vcn, NULL);
1028                 if (unlikely(err)) {
1029                         ntfs_error(vol->sb, "Cannot fill hole in inode 0x%lx, "
1030                                         "attribute type 0x%x, because building "
1031                                         "the mapping pairs failed with error "
1032                                         "code %i.", vi->i_ino,
1033                                         (unsigned)le32_to_cpu(ni->type), err);
1034                         err = -EIO;
1035                         break;
1036                 }
1037                 /* Update the highest_vcn but only if it was not set. */
1038                 if (unlikely(!a->data.non_resident.highest_vcn))
1039                         a->data.non_resident.highest_vcn =
1040                                         cpu_to_sle64(highest_vcn);
1041                 /*
1042                  * If the attribute is sparse/compressed, update the compressed
1043                  * size in the ntfs_inode structure and the attribute record.
1044                  */
1045                 if (likely(NInoSparse(ni) || NInoCompressed(ni))) {
1046                         /*
1047                          * If we are not in the first attribute extent, switch
1048                          * to it, but first ensure the changes will make it to
1049                          * disk later.
1050                          */
1051                         if (a->data.non_resident.lowest_vcn) {
1052                                 flush_dcache_mft_record_page(ctx->ntfs_ino);
1053                                 mark_mft_record_dirty(ctx->ntfs_ino);
1054                                 ntfs_attr_reinit_search_ctx(ctx);
1055                                 err = ntfs_attr_lookup(ni->type, ni->name,
1056                                                 ni->name_len, CASE_SENSITIVE,
1057                                                 0, NULL, 0, ctx);
1058                                 if (unlikely(err)) {
1059                                         status.attr_switched = 1;
1060                                         break;
1061                                 }
1062                                 /* @m is not used any more so do not set it. */
1063                                 a = ctx->attr;
1064                         }
1065                         write_lock_irqsave(&ni->size_lock, flags);
1066                         ni->itype.compressed.size += vol->cluster_size;
1067                         a->data.non_resident.compressed_size =
1068                                         cpu_to_sle64(ni->itype.compressed.size);
1069                         write_unlock_irqrestore(&ni->size_lock, flags);
1070                 }
1071                 /* Ensure the changes make it to disk. */
1072                 flush_dcache_mft_record_page(ctx->ntfs_ino);
1073                 mark_mft_record_dirty(ctx->ntfs_ino);
1074                 ntfs_attr_put_search_ctx(ctx);
1075                 unmap_mft_record(base_ni);
1076                 /* Successfully filled the hole. */
1077                 status.runlist_merged = 0;
1078                 status.mft_attr_mapped = 0;
1079                 status.mp_rebuilt = 0;
1080                 /* Setup the map cache and use that to deal with the buffer. */
1081                 was_hole = true;
1082                 vcn = bh_cpos;
1083                 vcn_len = 1;
1084                 lcn_block = lcn << (vol->cluster_size_bits - blocksize_bits);
1085                 cdelta = 0;
1086                 /*
1087                  * If the number of remaining clusters in the @pages is smaller
1088                  * or equal to the number of cached clusters, unlock the
1089                  * runlist as the map cache will be used from now on.
1090                  */
1091                 if (likely(vcn + vcn_len >= cend)) {
1092                         up_write(&ni->runlist.lock);
1093                         rl_write_locked = false;
1094                         rl = NULL;
1095                 }
1096                 goto map_buffer_cached;
1097         } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
1098         /* If there are no errors, do the next page. */
1099         if (likely(!err && ++u < nr_pages))
1100                 goto do_next_page;
1101         /* If there are no errors, release the runlist lock if we took it. */
1102         if (likely(!err)) {
1103                 if (unlikely(rl_write_locked)) {
1104                         up_write(&ni->runlist.lock);
1105                         rl_write_locked = false;
1106                 } else if (unlikely(rl))
1107                         up_read(&ni->runlist.lock);
1108                 rl = NULL;
1109         }
1110         /* If we issued read requests, let them complete. */
1111         read_lock_irqsave(&ni->size_lock, flags);
1112         initialized_size = ni->initialized_size;
1113         read_unlock_irqrestore(&ni->size_lock, flags);
1114         while (wait_bh > wait) {
1115                 bh = *--wait_bh;
1116                 wait_on_buffer(bh);
1117                 if (likely(buffer_uptodate(bh))) {
1118                         page = bh->b_page;
1119                         bh_pos = ((s64)page->index << PAGE_CACHE_SHIFT) +
1120                                         bh_offset(bh);
1121                         /*
1122                          * If the buffer overflows the initialized size, need
1123                          * to zero the overflowing region.
1124                          */
1125                         if (unlikely(bh_pos + blocksize > initialized_size)) {
1126                                 int ofs = 0;
1127
1128                                 if (likely(bh_pos < initialized_size))
1129                                         ofs = initialized_size - bh_pos;
1130                                 zero_user_page(page, bh_offset(bh) + ofs,
1131                                                 blocksize - ofs, KM_USER0);
1132                         }
1133                 } else /* if (unlikely(!buffer_uptodate(bh))) */
1134                         err = -EIO;
1135         }
1136         if (likely(!err)) {
1137                 /* Clear buffer_new on all buffers. */
1138                 u = 0;
1139                 do {
1140                         bh = head = page_buffers(pages[u]);
1141                         do {
1142                                 if (buffer_new(bh))
1143                                         clear_buffer_new(bh);
1144                         } while ((bh = bh->b_this_page) != head);
1145                 } while (++u < nr_pages);
1146                 ntfs_debug("Done.");
1147                 return err;
1148         }
1149         if (status.attr_switched) {
1150                 /* Get back to the attribute extent we modified. */
1151                 ntfs_attr_reinit_search_ctx(ctx);
1152                 if (ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1153                                 CASE_SENSITIVE, bh_cpos, NULL, 0, ctx)) {
1154                         ntfs_error(vol->sb, "Failed to find required "
1155                                         "attribute extent of attribute in "
1156                                         "error code path.  Run chkdsk to "
1157                                         "recover.");
1158                         write_lock_irqsave(&ni->size_lock, flags);
1159                         ni->itype.compressed.size += vol->cluster_size;
1160                         write_unlock_irqrestore(&ni->size_lock, flags);
1161                         flush_dcache_mft_record_page(ctx->ntfs_ino);
1162                         mark_mft_record_dirty(ctx->ntfs_ino);
1163                         /*
1164                          * The only thing that is now wrong is the compressed
1165                          * size of the base attribute extent which chkdsk
1166                          * should be able to fix.
1167                          */
1168                         NVolSetErrors(vol);
1169                 } else {
1170                         m = ctx->mrec;
1171                         a = ctx->attr;
1172                         status.attr_switched = 0;
1173                 }
1174         }
1175         /*
1176          * If the runlist has been modified, need to restore it by punching a
1177          * hole into it and we then need to deallocate the on-disk cluster as
1178          * well.  Note, we only modify the runlist if we are able to generate a
1179          * new mapping pairs array, i.e. only when the mapped attribute extent
1180          * is not switched.
1181          */
1182         if (status.runlist_merged && !status.attr_switched) {
1183                 BUG_ON(!rl_write_locked);
1184                 /* Make the file cluster we allocated sparse in the runlist. */
1185                 if (ntfs_rl_punch_nolock(vol, &ni->runlist, bh_cpos, 1)) {
1186                         ntfs_error(vol->sb, "Failed to punch hole into "
1187                                         "attribute runlist in error code "
1188                                         "path.  Run chkdsk to recover the "
1189                                         "lost cluster.");
1190                         NVolSetErrors(vol);
1191                 } else /* if (success) */ {
1192                         status.runlist_merged = 0;
1193                         /*
1194                          * Deallocate the on-disk cluster we allocated but only
1195                          * if we succeeded in punching its vcn out of the
1196                          * runlist.
1197                          */
1198                         down_write(&vol->lcnbmp_lock);
1199                         if (ntfs_bitmap_clear_bit(vol->lcnbmp_ino, lcn)) {
1200                                 ntfs_error(vol->sb, "Failed to release "
1201                                                 "allocated cluster in error "
1202                                                 "code path.  Run chkdsk to "
1203                                                 "recover the lost cluster.");
1204                                 NVolSetErrors(vol);
1205                         }
1206                         up_write(&vol->lcnbmp_lock);
1207                 }
1208         }
1209         /*
1210          * Resize the attribute record to its old size and rebuild the mapping
1211          * pairs array.  Note, we only can do this if the runlist has been
1212          * restored to its old state which also implies that the mapped
1213          * attribute extent is not switched.
1214          */
1215         if (status.mp_rebuilt && !status.runlist_merged) {
1216                 if (ntfs_attr_record_resize(m, a, attr_rec_len)) {
1217                         ntfs_error(vol->sb, "Failed to restore attribute "
1218                                         "record in error code path.  Run "
1219                                         "chkdsk to recover.");
1220                         NVolSetErrors(vol);
1221                 } else /* if (success) */ {
1222                         if (ntfs_mapping_pairs_build(vol, (u8*)a +
1223                                         le16_to_cpu(a->data.non_resident.
1224                                         mapping_pairs_offset), attr_rec_len -
1225                                         le16_to_cpu(a->data.non_resident.
1226                                         mapping_pairs_offset), ni->runlist.rl,
1227                                         vcn, highest_vcn, NULL)) {
1228                                 ntfs_error(vol->sb, "Failed to restore "
1229                                                 "mapping pairs array in error "
1230                                                 "code path.  Run chkdsk to "
1231                                                 "recover.");
1232                                 NVolSetErrors(vol);
1233                         }
1234                         flush_dcache_mft_record_page(ctx->ntfs_ino);
1235                         mark_mft_record_dirty(ctx->ntfs_ino);
1236                 }
1237         }
1238         /* Release the mft record and the attribute. */
1239         if (status.mft_attr_mapped) {
1240                 ntfs_attr_put_search_ctx(ctx);
1241                 unmap_mft_record(base_ni);
1242         }
1243         /* Release the runlist lock. */
1244         if (rl_write_locked)
1245                 up_write(&ni->runlist.lock);
1246         else if (rl)
1247                 up_read(&ni->runlist.lock);
1248         /*
1249          * Zero out any newly allocated blocks to avoid exposing stale data.
1250          * If BH_New is set, we know that the block was newly allocated above
1251          * and that it has not been fully zeroed and marked dirty yet.
1252          */
1253         nr_pages = u;
1254         u = 0;
1255         end = bh_cpos << vol->cluster_size_bits;
1256         do {
1257                 page = pages[u];
1258                 bh = head = page_buffers(page);
1259                 do {
1260                         if (u == nr_pages &&
1261                                         ((s64)page->index << PAGE_CACHE_SHIFT) +
1262                                         bh_offset(bh) >= end)
1263                                 break;
1264                         if (!buffer_new(bh))
1265                                 continue;
1266                         clear_buffer_new(bh);
1267                         if (!buffer_uptodate(bh)) {
1268                                 if (PageUptodate(page))
1269                                         set_buffer_uptodate(bh);
1270                                 else {
1271                                         zero_user_page(page, bh_offset(bh),
1272                                                         blocksize, KM_USER0);
1273                                         set_buffer_uptodate(bh);
1274                                 }
1275                         }
1276                         mark_buffer_dirty(bh);
1277                 } while ((bh = bh->b_this_page) != head);
1278         } while (++u <= nr_pages);
1279         ntfs_error(vol->sb, "Failed.  Returning error code %i.", err);
1280         return err;
1281 }
1282
1283 /*
1284  * Copy as much as we can into the pages and return the number of bytes which
1285  * were sucessfully copied.  If a fault is encountered then clear the pages
1286  * out to (ofs + bytes) and return the number of bytes which were copied.
1287  */
1288 static inline size_t ntfs_copy_from_user(struct page **pages,
1289                 unsigned nr_pages, unsigned ofs, const char __user *buf,
1290                 size_t bytes)
1291 {
1292         struct page **last_page = pages + nr_pages;
1293         char *kaddr;
1294         size_t total = 0;
1295         unsigned len;
1296         int left;
1297
1298         do {
1299                 len = PAGE_CACHE_SIZE - ofs;
1300                 if (len > bytes)
1301                         len = bytes;
1302                 kaddr = kmap_atomic(*pages, KM_USER0);
1303                 left = __copy_from_user_inatomic(kaddr + ofs, buf, len);
1304                 kunmap_atomic(kaddr, KM_USER0);
1305                 if (unlikely(left)) {
1306                         /* Do it the slow way. */
1307                         kaddr = kmap(*pages);
1308                         left = __copy_from_user(kaddr + ofs, buf, len);
1309                         kunmap(*pages);
1310                         if (unlikely(left))
1311                                 goto err_out;
1312                 }
1313                 total += len;
1314                 bytes -= len;
1315                 if (!bytes)
1316                         break;
1317                 buf += len;
1318                 ofs = 0;
1319         } while (++pages < last_page);
1320 out:
1321         return total;
1322 err_out:
1323         total += len - left;
1324         /* Zero the rest of the target like __copy_from_user(). */
1325         while (++pages < last_page) {
1326                 bytes -= len;
1327                 if (!bytes)
1328                         break;
1329                 len = PAGE_CACHE_SIZE;
1330                 if (len > bytes)
1331                         len = bytes;
1332                 zero_user_page(*pages, 0, len, KM_USER0);
1333         }
1334         goto out;
1335 }
1336
1337 static size_t __ntfs_copy_from_user_iovec_inatomic(char *vaddr,
1338                 const struct iovec *iov, size_t iov_ofs, size_t bytes)
1339 {
1340         size_t total = 0;
1341
1342         while (1) {
1343                 const char __user *buf = iov->iov_base + iov_ofs;
1344                 unsigned len;
1345                 size_t left;
1346
1347                 len = iov->iov_len - iov_ofs;
1348                 if (len > bytes)
1349                         len = bytes;
1350                 left = __copy_from_user_inatomic(vaddr, buf, len);
1351                 total += len;
1352                 bytes -= len;
1353                 vaddr += len;
1354                 if (unlikely(left)) {
1355                         total -= left;
1356                         break;
1357                 }
1358                 if (!bytes)
1359                         break;
1360                 iov++;
1361                 iov_ofs = 0;
1362         }
1363         return total;
1364 }
1365
1366 static inline void ntfs_set_next_iovec(const struct iovec **iovp,
1367                 size_t *iov_ofsp, size_t bytes)
1368 {
1369         const struct iovec *iov = *iovp;
1370         size_t iov_ofs = *iov_ofsp;
1371
1372         while (bytes) {
1373                 unsigned len;
1374
1375                 len = iov->iov_len - iov_ofs;
1376                 if (len > bytes)
1377                         len = bytes;
1378                 bytes -= len;
1379                 iov_ofs += len;
1380                 if (iov->iov_len == iov_ofs) {
1381                         iov++;
1382                         iov_ofs = 0;
1383                 }
1384         }
1385         *iovp = iov;
1386         *iov_ofsp = iov_ofs;
1387 }
1388
1389 /*
1390  * This has the same side-effects and return value as ntfs_copy_from_user().
1391  * The difference is that on a fault we need to memset the remainder of the
1392  * pages (out to offset + bytes), to emulate ntfs_copy_from_user()'s
1393  * single-segment behaviour.
1394  *
1395  * We call the same helper (__ntfs_copy_from_user_iovec_inatomic()) both
1396  * when atomic and when not atomic.  This is ok because
1397  * __ntfs_copy_from_user_iovec_inatomic() calls __copy_from_user_inatomic()
1398  * and it is ok to call this when non-atomic.
1399  * Infact, the only difference between __copy_from_user_inatomic() and
1400  * __copy_from_user() is that the latter calls might_sleep() and the former
1401  * should not zero the tail of the buffer on error.  And on many
1402  * architectures __copy_from_user_inatomic() is just defined to
1403  * __copy_from_user() so it makes no difference at all on those architectures.
1404  */
1405 static inline size_t ntfs_copy_from_user_iovec(struct page **pages,
1406                 unsigned nr_pages, unsigned ofs, const struct iovec **iov,
1407                 size_t *iov_ofs, size_t bytes)
1408 {
1409         struct page **last_page = pages + nr_pages;
1410         char *kaddr;
1411         size_t copied, len, total = 0;
1412
1413         do {
1414                 len = PAGE_CACHE_SIZE - ofs;
1415                 if (len > bytes)
1416                         len = bytes;
1417                 kaddr = kmap_atomic(*pages, KM_USER0);
1418                 copied = __ntfs_copy_from_user_iovec_inatomic(kaddr + ofs,
1419                                 *iov, *iov_ofs, len);
1420                 kunmap_atomic(kaddr, KM_USER0);
1421                 if (unlikely(copied != len)) {
1422                         /* Do it the slow way. */
1423                         kaddr = kmap(*pages);
1424                         copied = __ntfs_copy_from_user_iovec_inatomic(kaddr + ofs,
1425                                         *iov, *iov_ofs, len);
1426                         /*
1427                          * Zero the rest of the target like __copy_from_user().
1428                          */
1429                         memset(kaddr + ofs + copied, 0, len - copied);
1430                         kunmap(*pages);
1431                         if (unlikely(copied != len))
1432                                 goto err_out;
1433                 }
1434                 total += len;
1435                 bytes -= len;
1436                 if (!bytes)
1437                         break;
1438                 ntfs_set_next_iovec(iov, iov_ofs, len);
1439                 ofs = 0;
1440         } while (++pages < last_page);
1441 out:
1442         return total;
1443 err_out:
1444         total += copied;
1445         /* Zero the rest of the target like __copy_from_user(). */
1446         while (++pages < last_page) {
1447                 bytes -= len;
1448                 if (!bytes)
1449                         break;
1450                 len = PAGE_CACHE_SIZE;
1451                 if (len > bytes)
1452                         len = bytes;
1453                 zero_user_page(*pages, 0, len, KM_USER0);
1454         }
1455         goto out;
1456 }
1457
1458 static inline void ntfs_flush_dcache_pages(struct page **pages,
1459                 unsigned nr_pages)
1460 {
1461         BUG_ON(!nr_pages);
1462         /*
1463          * Warning: Do not do the decrement at the same time as the call to
1464          * flush_dcache_page() because it is a NULL macro on i386 and hence the
1465          * decrement never happens so the loop never terminates.
1466          */
1467         do {
1468                 --nr_pages;
1469                 flush_dcache_page(pages[nr_pages]);
1470         } while (nr_pages > 0);
1471 }
1472
1473 /**
1474  * ntfs_commit_pages_after_non_resident_write - commit the received data
1475  * @pages:      array of destination pages
1476  * @nr_pages:   number of pages in @pages
1477  * @pos:        byte position in file at which the write begins
1478  * @bytes:      number of bytes to be written
1479  *
1480  * See description of ntfs_commit_pages_after_write(), below.
1481  */
1482 static inline int ntfs_commit_pages_after_non_resident_write(
1483                 struct page **pages, const unsigned nr_pages,
1484                 s64 pos, size_t bytes)
1485 {
1486         s64 end, initialized_size;
1487         struct inode *vi;
1488         ntfs_inode *ni, *base_ni;
1489         struct buffer_head *bh, *head;
1490         ntfs_attr_search_ctx *ctx;
1491         MFT_RECORD *m;
1492         ATTR_RECORD *a;
1493         unsigned long flags;
1494         unsigned blocksize, u;
1495         int err;
1496
1497         vi = pages[0]->mapping->host;
1498         ni = NTFS_I(vi);
1499         blocksize = vi->i_sb->s_blocksize;
1500         end = pos + bytes;
1501         u = 0;
1502         do {
1503                 s64 bh_pos;
1504                 struct page *page;
1505                 bool partial;
1506
1507                 page = pages[u];
1508                 bh_pos = (s64)page->index << PAGE_CACHE_SHIFT;
1509                 bh = head = page_buffers(page);
1510                 partial = false;
1511                 do {
1512                         s64 bh_end;
1513
1514                         bh_end = bh_pos + blocksize;
1515                         if (bh_end <= pos || bh_pos >= end) {
1516                                 if (!buffer_uptodate(bh))
1517                                         partial = true;
1518                         } else {
1519                                 set_buffer_uptodate(bh);
1520                                 mark_buffer_dirty(bh);
1521                         }
1522                 } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
1523                 /*
1524                  * If all buffers are now uptodate but the page is not, set the
1525                  * page uptodate.
1526                  */
1527                 if (!partial && !PageUptodate(page))
1528                         SetPageUptodate(page);
1529         } while (++u < nr_pages);
1530         /*
1531          * Finally, if we do not need to update initialized_size or i_size we
1532          * are finished.
1533          */
1534         read_lock_irqsave(&ni->size_lock, flags);
1535         initialized_size = ni->initialized_size;
1536         read_unlock_irqrestore(&ni->size_lock, flags);
1537         if (end <= initialized_size) {
1538                 ntfs_debug("Done.");
1539                 return 0;
1540         }
1541         /*
1542          * Update initialized_size/i_size as appropriate, both in the inode and
1543          * the mft record.
1544          */
1545         if (!NInoAttr(ni))
1546                 base_ni = ni;
1547         else
1548                 base_ni = ni->ext.base_ntfs_ino;
1549         /* Map, pin, and lock the mft record. */
1550         m = map_mft_record(base_ni);
1551         if (IS_ERR(m)) {
1552                 err = PTR_ERR(m);
1553                 m = NULL;
1554                 ctx = NULL;
1555                 goto err_out;
1556         }
1557         BUG_ON(!NInoNonResident(ni));
1558         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1559         if (unlikely(!ctx)) {
1560                 err = -ENOMEM;
1561                 goto err_out;
1562         }
1563         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1564                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1565         if (unlikely(err)) {
1566                 if (err == -ENOENT)
1567                         err = -EIO;
1568                 goto err_out;
1569         }
1570         a = ctx->attr;
1571         BUG_ON(!a->non_resident);
1572         write_lock_irqsave(&ni->size_lock, flags);
1573         BUG_ON(end > ni->allocated_size);
1574         ni->initialized_size = end;
1575         a->data.non_resident.initialized_size = cpu_to_sle64(end);
1576         if (end > i_size_read(vi)) {
1577                 i_size_write(vi, end);
1578                 a->data.non_resident.data_size =
1579                                 a->data.non_resident.initialized_size;
1580         }
1581         write_unlock_irqrestore(&ni->size_lock, flags);
1582         /* Mark the mft record dirty, so it gets written back. */
1583         flush_dcache_mft_record_page(ctx->ntfs_ino);
1584         mark_mft_record_dirty(ctx->ntfs_ino);
1585         ntfs_attr_put_search_ctx(ctx);
1586         unmap_mft_record(base_ni);
1587         ntfs_debug("Done.");
1588         return 0;
1589 err_out:
1590         if (ctx)
1591                 ntfs_attr_put_search_ctx(ctx);
1592         if (m)
1593                 unmap_mft_record(base_ni);
1594         ntfs_error(vi->i_sb, "Failed to update initialized_size/i_size (error "
1595                         "code %i).", err);
1596         if (err != -ENOMEM)
1597                 NVolSetErrors(ni->vol);
1598         return err;
1599 }
1600
1601 /**
1602  * ntfs_commit_pages_after_write - commit the received data
1603  * @pages:      array of destination pages
1604  * @nr_pages:   number of pages in @pages
1605  * @pos:        byte position in file at which the write begins
1606  * @bytes:      number of bytes to be written
1607  *
1608  * This is called from ntfs_file_buffered_write() with i_mutex held on the inode
1609  * (@pages[0]->mapping->host).  There are @nr_pages pages in @pages which are
1610  * locked but not kmap()ped.  The source data has already been copied into the
1611  * @page.  ntfs_prepare_pages_for_non_resident_write() has been called before
1612  * the data was copied (for non-resident attributes only) and it returned
1613  * success.
1614  *
1615  * Need to set uptodate and mark dirty all buffers within the boundary of the
1616  * write.  If all buffers in a page are uptodate we set the page uptodate, too.
1617  *
1618  * Setting the buffers dirty ensures that they get written out later when
1619  * ntfs_writepage() is invoked by the VM.
1620  *
1621  * Finally, we need to update i_size and initialized_size as appropriate both
1622  * in the inode and the mft record.
1623  *
1624  * This is modelled after fs/buffer.c::generic_commit_write(), which marks
1625  * buffers uptodate and dirty, sets the page uptodate if all buffers in the
1626  * page are uptodate, and updates i_size if the end of io is beyond i_size.  In
1627  * that case, it also marks the inode dirty.
1628  *
1629  * If things have gone as outlined in
1630  * ntfs_prepare_pages_for_non_resident_write(), we do not need to do any page
1631  * content modifications here for non-resident attributes.  For resident
1632  * attributes we need to do the uptodate bringing here which we combine with
1633  * the copying into the mft record which means we save one atomic kmap.
1634  *
1635  * Return 0 on success or -errno on error.
1636  */
1637 static int ntfs_commit_pages_after_write(struct page **pages,
1638                 const unsigned nr_pages, s64 pos, size_t bytes)
1639 {
1640         s64 end, initialized_size;
1641         loff_t i_size;
1642         struct inode *vi;
1643         ntfs_inode *ni, *base_ni;
1644         struct page *page;
1645         ntfs_attr_search_ctx *ctx;
1646         MFT_RECORD *m;
1647         ATTR_RECORD *a;
1648         char *kattr, *kaddr;
1649         unsigned long flags;
1650         u32 attr_len;
1651         int err;
1652
1653         BUG_ON(!nr_pages);
1654         BUG_ON(!pages);
1655         page = pages[0];
1656         BUG_ON(!page);
1657         vi = page->mapping->host;
1658         ni = NTFS_I(vi);
1659         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, start page "
1660                         "index 0x%lx, nr_pages 0x%x, pos 0x%llx, bytes 0x%zx.",
1661                         vi->i_ino, ni->type, page->index, nr_pages,
1662                         (long long)pos, bytes);
1663         if (NInoNonResident(ni))
1664                 return ntfs_commit_pages_after_non_resident_write(pages,
1665                                 nr_pages, pos, bytes);
1666         BUG_ON(nr_pages > 1);
1667         /*
1668          * Attribute is resident, implying it is not compressed, encrypted, or
1669          * sparse.
1670          */
1671         if (!NInoAttr(ni))
1672                 base_ni = ni;
1673         else
1674                 base_ni = ni->ext.base_ntfs_ino;
1675         BUG_ON(NInoNonResident(ni));
1676         /* Map, pin, and lock the mft record. */
1677         m = map_mft_record(base_ni);
1678         if (IS_ERR(m)) {
1679                 err = PTR_ERR(m);
1680                 m = NULL;
1681                 ctx = NULL;
1682                 goto err_out;
1683         }
1684         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1685         if (unlikely(!ctx)) {
1686                 err = -ENOMEM;
1687                 goto err_out;
1688         }
1689         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1690                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1691         if (unlikely(err)) {
1692                 if (err == -ENOENT)
1693                         err = -EIO;
1694                 goto err_out;
1695         }
1696         a = ctx->attr;
1697         BUG_ON(a->non_resident);
1698         /* The total length of the attribute value. */
1699         attr_len = le32_to_cpu(a->data.resident.value_length);
1700         i_size = i_size_read(vi);
1701         BUG_ON(attr_len != i_size);
1702         BUG_ON(pos > attr_len);
1703         end = pos + bytes;
1704         BUG_ON(end > le32_to_cpu(a->length) -
1705                         le16_to_cpu(a->data.resident.value_offset));
1706         kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
1707         kaddr = kmap_atomic(page, KM_USER0);
1708         /* Copy the received data from the page to the mft record. */
1709         memcpy(kattr + pos, kaddr + pos, bytes);
1710         /* Update the attribute length if necessary. */
1711         if (end > attr_len) {
1712                 attr_len = end;
1713                 a->data.resident.value_length = cpu_to_le32(attr_len);
1714         }
1715         /*
1716          * If the page is not uptodate, bring the out of bounds area(s)
1717          * uptodate by copying data from the mft record to the page.
1718          */
1719         if (!PageUptodate(page)) {
1720                 if (pos > 0)
1721                         memcpy(kaddr, kattr, pos);
1722                 if (end < attr_len)
1723                         memcpy(kaddr + end, kattr + end, attr_len - end);
1724                 /* Zero the region outside the end of the attribute value. */
1725                 memset(kaddr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
1726                 flush_dcache_page(page);
1727                 SetPageUptodate(page);
1728         }
1729         kunmap_atomic(kaddr, KM_USER0);
1730         /* Update initialized_size/i_size if necessary. */
1731         read_lock_irqsave(&ni->size_lock, flags);
1732         initialized_size = ni->initialized_size;
1733         BUG_ON(end > ni->allocated_size);
1734         read_unlock_irqrestore(&ni->size_lock, flags);
1735         BUG_ON(initialized_size != i_size);
1736         if (end > initialized_size) {
1737                 unsigned long flags;
1738
1739                 write_lock_irqsave(&ni->size_lock, flags);
1740                 ni->initialized_size = end;
1741                 i_size_write(vi, end);
1742                 write_unlock_irqrestore(&ni->size_lock, flags);
1743         }
1744         /* Mark the mft record dirty, so it gets written back. */
1745         flush_dcache_mft_record_page(ctx->ntfs_ino);
1746         mark_mft_record_dirty(ctx->ntfs_ino);
1747         ntfs_attr_put_search_ctx(ctx);
1748         unmap_mft_record(base_ni);
1749         ntfs_debug("Done.");
1750         return 0;
1751 err_out:
1752         if (err == -ENOMEM) {
1753                 ntfs_warning(vi->i_sb, "Error allocating memory required to "
1754                                 "commit the write.");
1755                 if (PageUptodate(page)) {
1756                         ntfs_warning(vi->i_sb, "Page is uptodate, setting "
1757                                         "dirty so the write will be retried "
1758                                         "later on by the VM.");
1759                         /*
1760                          * Put the page on mapping->dirty_pages, but leave its
1761                          * buffers' dirty state as-is.
1762                          */
1763                         __set_page_dirty_nobuffers(page);
1764                         err = 0;
1765                 } else
1766                         ntfs_error(vi->i_sb, "Page is not uptodate.  Written "
1767                                         "data has been lost.");
1768         } else {
1769                 ntfs_error(vi->i_sb, "Resident attribute commit write failed "
1770                                 "with error %i.", err);
1771                 NVolSetErrors(ni->vol);
1772         }
1773         if (ctx)
1774                 ntfs_attr_put_search_ctx(ctx);
1775         if (m)
1776                 unmap_mft_record(base_ni);
1777         return err;
1778 }
1779
1780 /**
1781  * ntfs_file_buffered_write -
1782  *
1783  * Locking: The vfs is holding ->i_mutex on the inode.
1784  */
1785 static ssize_t ntfs_file_buffered_write(struct kiocb *iocb,
1786                 const struct iovec *iov, unsigned long nr_segs,
1787                 loff_t pos, loff_t *ppos, size_t count)
1788 {
1789         struct file *file = iocb->ki_filp;
1790         struct address_space *mapping = file->f_mapping;
1791         struct inode *vi = mapping->host;
1792         ntfs_inode *ni = NTFS_I(vi);
1793         ntfs_volume *vol = ni->vol;
1794         struct page *pages[NTFS_MAX_PAGES_PER_CLUSTER];
1795         struct page *cached_page = NULL;
1796         char __user *buf = NULL;
1797         s64 end, ll;
1798         VCN last_vcn;
1799         LCN lcn;
1800         unsigned long flags;
1801         size_t bytes, iov_ofs = 0;      /* Offset in the current iovec. */
1802         ssize_t status, written;
1803         unsigned nr_pages;
1804         int err;
1805         struct pagevec lru_pvec;
1806
1807         ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, "
1808                         "pos 0x%llx, count 0x%lx.",
1809                         vi->i_ino, (unsigned)le32_to_cpu(ni->type),
1810                         (unsigned long long)pos, (unsigned long)count);
1811         if (unlikely(!count))
1812                 return 0;
1813         BUG_ON(NInoMstProtected(ni));
1814         /*
1815          * If the attribute is not an index root and it is encrypted or
1816          * compressed, we cannot write to it yet.  Note we need to check for
1817          * AT_INDEX_ALLOCATION since this is the type of both directory and
1818          * index inodes.
1819          */
1820         if (ni->type != AT_INDEX_ALLOCATION) {
1821                 /* If file is encrypted, deny access, just like NT4. */
1822                 if (NInoEncrypted(ni)) {
1823                         /*
1824                          * Reminder for later: Encrypted files are _always_
1825                          * non-resident so that the content can always be
1826                          * encrypted.
1827                          */
1828                         ntfs_debug("Denying write access to encrypted file.");
1829                         return -EACCES;
1830                 }
1831                 if (NInoCompressed(ni)) {
1832                         /* Only unnamed $DATA attribute can be compressed. */
1833                         BUG_ON(ni->type != AT_DATA);
1834                         BUG_ON(ni->name_len);
1835                         /*
1836                          * Reminder for later: If resident, the data is not
1837                          * actually compressed.  Only on the switch to non-
1838                          * resident does compression kick in.  This is in
1839                          * contrast to encrypted files (see above).
1840                          */
1841                         ntfs_error(vi->i_sb, "Writing to compressed files is "
1842                                         "not implemented yet.  Sorry.");
1843                         return -EOPNOTSUPP;
1844                 }
1845         }
1846         /*
1847          * If a previous ntfs_truncate() failed, repeat it and abort if it
1848          * fails again.
1849          */
1850         if (unlikely(NInoTruncateFailed(ni))) {
1851                 down_write(&vi->i_alloc_sem);
1852                 err = ntfs_truncate(vi);
1853                 up_write(&vi->i_alloc_sem);
1854                 if (err || NInoTruncateFailed(ni)) {
1855                         if (!err)
1856                                 err = -EIO;
1857                         ntfs_error(vol->sb, "Cannot perform write to inode "
1858                                         "0x%lx, attribute type 0x%x, because "
1859                                         "ntfs_truncate() failed (error code "
1860                                         "%i).", vi->i_ino,
1861                                         (unsigned)le32_to_cpu(ni->type), err);
1862                         return err;
1863                 }
1864         }
1865         /* The first byte after the write. */
1866         end = pos + count;
1867         /*
1868          * If the write goes beyond the allocated size, extend the allocation
1869          * to cover the whole of the write, rounded up to the nearest cluster.
1870          */
1871         read_lock_irqsave(&ni->size_lock, flags);
1872         ll = ni->allocated_size;
1873         read_unlock_irqrestore(&ni->size_lock, flags);
1874         if (end > ll) {
1875                 /* Extend the allocation without changing the data size. */
1876                 ll = ntfs_attr_extend_allocation(ni, end, -1, pos);
1877                 if (likely(ll >= 0)) {
1878                         BUG_ON(pos >= ll);
1879                         /* If the extension was partial truncate the write. */
1880                         if (end > ll) {
1881                                 ntfs_debug("Truncating write to inode 0x%lx, "
1882                                                 "attribute type 0x%x, because "
1883                                                 "the allocation was only "
1884                                                 "partially extended.",
1885                                                 vi->i_ino, (unsigned)
1886                                                 le32_to_cpu(ni->type));
1887                                 end = ll;
1888                                 count = ll - pos;
1889                         }
1890                 } else {
1891                         err = ll;
1892                         read_lock_irqsave(&ni->size_lock, flags);
1893                         ll = ni->allocated_size;
1894                         read_unlock_irqrestore(&ni->size_lock, flags);
1895                         /* Perform a partial write if possible or fail. */
1896                         if (pos < ll) {
1897                                 ntfs_debug("Truncating write to inode 0x%lx, "
1898                                                 "attribute type 0x%x, because "
1899                                                 "extending the allocation "
1900                                                 "failed (error code %i).",
1901                                                 vi->i_ino, (unsigned)
1902                                                 le32_to_cpu(ni->type), err);
1903                                 end = ll;
1904                                 count = ll - pos;
1905                         } else {
1906                                 ntfs_error(vol->sb, "Cannot perform write to "
1907                                                 "inode 0x%lx, attribute type "
1908                                                 "0x%x, because extending the "
1909                                                 "allocation failed (error "
1910                                                 "code %i).", vi->i_ino,
1911                                                 (unsigned)
1912                                                 le32_to_cpu(ni->type), err);
1913                                 return err;
1914                         }
1915                 }
1916         }
1917         pagevec_init(&lru_pvec, 0);
1918         written = 0;
1919         /*
1920          * If the write starts beyond the initialized size, extend it up to the
1921          * beginning of the write and initialize all non-sparse space between
1922          * the old initialized size and the new one.  This automatically also
1923          * increments the vfs inode->i_size to keep it above or equal to the
1924          * initialized_size.
1925          */
1926         read_lock_irqsave(&ni->size_lock, flags);
1927         ll = ni->initialized_size;
1928         read_unlock_irqrestore(&ni->size_lock, flags);
1929         if (pos > ll) {
1930                 err = ntfs_attr_extend_initialized(ni, pos, &cached_page,
1931                                 &lru_pvec);
1932                 if (err < 0) {
1933                         ntfs_error(vol->sb, "Cannot perform write to inode "
1934                                         "0x%lx, attribute type 0x%x, because "
1935                                         "extending the initialized size "
1936                                         "failed (error code %i).", vi->i_ino,
1937                                         (unsigned)le32_to_cpu(ni->type), err);
1938                         status = err;
1939                         goto err_out;
1940                 }
1941         }
1942         /*
1943          * Determine the number of pages per cluster for non-resident
1944          * attributes.
1945          */
1946         nr_pages = 1;
1947         if (vol->cluster_size > PAGE_CACHE_SIZE && NInoNonResident(ni))
1948                 nr_pages = vol->cluster_size >> PAGE_CACHE_SHIFT;
1949         /* Finally, perform the actual write. */
1950         last_vcn = -1;
1951         if (likely(nr_segs == 1))
1952                 buf = iov->iov_base;
1953         do {
1954                 VCN vcn;
1955                 pgoff_t idx, start_idx;
1956                 unsigned ofs, do_pages, u;
1957                 size_t copied;
1958
1959                 start_idx = idx = pos >> PAGE_CACHE_SHIFT;
1960                 ofs = pos & ~PAGE_CACHE_MASK;
1961                 bytes = PAGE_CACHE_SIZE - ofs;
1962                 do_pages = 1;
1963                 if (nr_pages > 1) {
1964                         vcn = pos >> vol->cluster_size_bits;
1965                         if (vcn != last_vcn) {
1966                                 last_vcn = vcn;
1967                                 /*
1968                                  * Get the lcn of the vcn the write is in.  If
1969                                  * it is a hole, need to lock down all pages in
1970                                  * the cluster.
1971                                  */
1972                                 down_read(&ni->runlist.lock);
1973                                 lcn = ntfs_attr_vcn_to_lcn_nolock(ni, pos >>
1974                                                 vol->cluster_size_bits, false);
1975                                 up_read(&ni->runlist.lock);
1976                                 if (unlikely(lcn < LCN_HOLE)) {
1977                                         status = -EIO;
1978                                         if (lcn == LCN_ENOMEM)
1979                                                 status = -ENOMEM;
1980                                         else
1981                                                 ntfs_error(vol->sb, "Cannot "
1982                                                         "perform write to "
1983                                                         "inode 0x%lx, "
1984                                                         "attribute type 0x%x, "
1985                                                         "because the attribute "
1986                                                         "is corrupt.",
1987                                                         vi->i_ino, (unsigned)
1988                                                         le32_to_cpu(ni->type));
1989                                         break;
1990                                 }
1991                                 if (lcn == LCN_HOLE) {
1992                                         start_idx = (pos & ~(s64)
1993                                                         vol->cluster_size_mask)
1994                                                         >> PAGE_CACHE_SHIFT;
1995                                         bytes = vol->cluster_size - (pos &
1996                                                         vol->cluster_size_mask);
1997                                         do_pages = nr_pages;
1998                                 }
1999                         }
2000                 }
2001                 if (bytes > count)
2002                         bytes = count;
2003                 /*
2004                  * Bring in the user page(s) that we will copy from _first_.
2005                  * Otherwise there is a nasty deadlock on copying from the same
2006                  * page(s) as we are writing to, without it/them being marked
2007                  * up-to-date.  Note, at present there is nothing to stop the
2008                  * pages being swapped out between us bringing them into memory
2009                  * and doing the actual copying.
2010                  */
2011                 if (likely(nr_segs == 1))
2012                         ntfs_fault_in_pages_readable(buf, bytes);
2013                 else
2014                         ntfs_fault_in_pages_readable_iovec(iov, iov_ofs, bytes);
2015                 /* Get and lock @do_pages starting at index @start_idx. */
2016                 status = __ntfs_grab_cache_pages(mapping, start_idx, do_pages,
2017                                 pages, &cached_page, &lru_pvec);
2018                 if (unlikely(status))
2019                         break;
2020                 /*
2021                  * For non-resident attributes, we need to fill any holes with
2022                  * actual clusters and ensure all bufferes are mapped.  We also
2023                  * need to bring uptodate any buffers that are only partially
2024                  * being written to.
2025                  */
2026                 if (NInoNonResident(ni)) {
2027                         status = ntfs_prepare_pages_for_non_resident_write(
2028                                         pages, do_pages, pos, bytes);
2029                         if (unlikely(status)) {
2030                                 loff_t i_size;
2031
2032                                 do {
2033                                         unlock_page(pages[--do_pages]);
2034                                         page_cache_release(pages[do_pages]);
2035                                 } while (do_pages);
2036                                 /*
2037                                  * The write preparation may have instantiated
2038                                  * allocated space outside i_size.  Trim this
2039                                  * off again.  We can ignore any errors in this
2040                                  * case as we will just be waisting a bit of
2041                                  * allocated space, which is not a disaster.
2042                                  */
2043                                 i_size = i_size_read(vi);
2044                                 if (pos + bytes > i_size)
2045                                         vmtruncate(vi, i_size);
2046                                 break;
2047                         }
2048                 }
2049                 u = (pos >> PAGE_CACHE_SHIFT) - pages[0]->index;
2050                 if (likely(nr_segs == 1)) {
2051                         copied = ntfs_copy_from_user(pages + u, do_pages - u,
2052                                         ofs, buf, bytes);
2053                         buf += copied;
2054                 } else
2055                         copied = ntfs_copy_from_user_iovec(pages + u,
2056                                         do_pages - u, ofs, &iov, &iov_ofs,
2057                                         bytes);
2058                 ntfs_flush_dcache_pages(pages + u, do_pages - u);
2059                 status = ntfs_commit_pages_after_write(pages, do_pages, pos,
2060                                 bytes);
2061                 if (likely(!status)) {
2062                         written += copied;
2063                         count -= copied;
2064                         pos += copied;
2065                         if (unlikely(copied != bytes))
2066                                 status = -EFAULT;
2067                 }
2068                 do {
2069                         unlock_page(pages[--do_pages]);
2070                         mark_page_accessed(pages[do_pages]);
2071                         page_cache_release(pages[do_pages]);
2072                 } while (do_pages);
2073                 if (unlikely(status))
2074                         break;
2075                 balance_dirty_pages_ratelimited(mapping);
2076                 cond_resched();
2077         } while (count);
2078 err_out:
2079         *ppos = pos;
2080         if (cached_page)
2081                 page_cache_release(cached_page);
2082         /* For now, when the user asks for O_SYNC, we actually give O_DSYNC. */
2083         if (likely(!status)) {
2084                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(vi))) {
2085                         if (!mapping->a_ops->writepage || !is_sync_kiocb(iocb))
2086                                 status = generic_osync_inode(vi, mapping,
2087                                                 OSYNC_METADATA|OSYNC_DATA);
2088                 }
2089         }
2090         pagevec_lru_add(&lru_pvec);
2091         ntfs_debug("Done.  Returning %s (written 0x%lx, status %li).",
2092                         written ? "written" : "status", (unsigned long)written,
2093                         (long)status);
2094         return written ? written : status;
2095 }
2096
2097 /**
2098  * ntfs_file_aio_write_nolock -
2099  */
2100 static ssize_t ntfs_file_aio_write_nolock(struct kiocb *iocb,
2101                 const struct iovec *iov, unsigned long nr_segs, loff_t *ppos)
2102 {
2103         struct file *file = iocb->ki_filp;
2104         struct address_space *mapping = file->f_mapping;
2105         struct inode *inode = mapping->host;
2106         loff_t pos;
2107         size_t count;           /* after file limit checks */
2108         ssize_t written, err;
2109
2110         count = 0;
2111         err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
2112         if (err)
2113                 return err;
2114         pos = *ppos;
2115         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2116         /* We can write back this queue in page reclaim. */
2117         current->backing_dev_info = mapping->backing_dev_info;
2118         written = 0;
2119         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2120         if (err)
2121                 goto out;
2122         if (!count)
2123                 goto out;
2124         err = remove_suid(file->f_path.dentry);
2125         if (err)
2126                 goto out;
2127         file_update_time(file);
2128         written = ntfs_file_buffered_write(iocb, iov, nr_segs, pos, ppos,
2129                         count);
2130 out:
2131         current->backing_dev_info = NULL;
2132         return written ? written : err;
2133 }
2134
2135 /**
2136  * ntfs_file_aio_write -
2137  */
2138 static ssize_t ntfs_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2139                 unsigned long nr_segs, loff_t pos)
2140 {
2141         struct file *file = iocb->ki_filp;
2142         struct address_space *mapping = file->f_mapping;
2143         struct inode *inode = mapping->host;
2144         ssize_t ret;
2145
2146         BUG_ON(iocb->ki_pos != pos);
2147
2148         mutex_lock(&inode->i_mutex);
2149         ret = ntfs_file_aio_write_nolock(iocb, iov, nr_segs, &iocb->ki_pos);
2150         mutex_unlock(&inode->i_mutex);
2151         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2152                 int err = sync_page_range(inode, mapping, pos, ret);
2153                 if (err < 0)
2154                         ret = err;
2155         }
2156         return ret;
2157 }
2158
2159 /**
2160  * ntfs_file_writev -
2161  *
2162  * Basically the same as generic_file_writev() except that it ends up calling
2163  * ntfs_file_aio_write_nolock() instead of __generic_file_aio_write_nolock().
2164  */
2165 static ssize_t ntfs_file_writev(struct file *file, const struct iovec *iov,
2166                 unsigned long nr_segs, loff_t *ppos)
2167 {
2168         struct address_space *mapping = file->f_mapping;
2169         struct inode *inode = mapping->host;
2170         struct kiocb kiocb;
2171         ssize_t ret;
2172
2173         mutex_lock(&inode->i_mutex);
2174         init_sync_kiocb(&kiocb, file);
2175         ret = ntfs_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2176         if (ret == -EIOCBQUEUED)
2177                 ret = wait_on_sync_kiocb(&kiocb);
2178         mutex_unlock(&inode->i_mutex);
2179         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2180                 int err = sync_page_range(inode, mapping, *ppos - ret, ret);
2181                 if (err < 0)
2182                         ret = err;
2183         }
2184         return ret;
2185 }
2186
2187 /**
2188  * ntfs_file_write - simple wrapper for ntfs_file_writev()
2189  */
2190 static ssize_t ntfs_file_write(struct file *file, const char __user *buf,
2191                 size_t count, loff_t *ppos)
2192 {
2193         struct iovec local_iov = { .iov_base = (void __user *)buf,
2194                                    .iov_len = count };
2195
2196         return ntfs_file_writev(file, &local_iov, 1, ppos);
2197 }
2198
2199 /**
2200  * ntfs_file_fsync - sync a file to disk
2201  * @filp:       file to be synced
2202  * @dentry:     dentry describing the file to sync
2203  * @datasync:   if non-zero only flush user data and not metadata
2204  *
2205  * Data integrity sync of a file to disk.  Used for fsync, fdatasync, and msync
2206  * system calls.  This function is inspired by fs/buffer.c::file_fsync().
2207  *
2208  * If @datasync is false, write the mft record and all associated extent mft
2209  * records as well as the $DATA attribute and then sync the block device.
2210  *
2211  * If @datasync is true and the attribute is non-resident, we skip the writing
2212  * of the mft record and all associated extent mft records (this might still
2213  * happen due to the write_inode_now() call).
2214  *
2215  * Also, if @datasync is true, we do not wait on the inode to be written out
2216  * but we always wait on the page cache pages to be written out.
2217  *
2218  * Note: In the past @filp could be NULL so we ignore it as we don't need it
2219  * anyway.
2220  *
2221  * Locking: Caller must hold i_mutex on the inode.
2222  *
2223  * TODO: We should probably also write all attribute/index inodes associated
2224  * with this inode but since we have no simple way of getting to them we ignore
2225  * this problem for now.
2226  */
2227 static int ntfs_file_fsync(struct file *filp, struct dentry *dentry,
2228                 int datasync)
2229 {
2230         struct inode *vi = dentry->d_inode;
2231         int err, ret = 0;
2232
2233         ntfs_debug("Entering for inode 0x%lx.", vi->i_ino);
2234         BUG_ON(S_ISDIR(vi->i_mode));
2235         if (!datasync || !NInoNonResident(NTFS_I(vi)))
2236                 ret = ntfs_write_inode(vi, 1);
2237         write_inode_now(vi, !datasync);
2238         /*
2239          * NOTE: If we were to use mapping->private_list (see ext2 and
2240          * fs/buffer.c) for dirty blocks then we could optimize the below to be
2241          * sync_mapping_buffers(vi->i_mapping).
2242          */
2243         err = sync_blockdev(vi->i_sb->s_bdev);
2244         if (unlikely(err && !ret))
2245                 ret = err;
2246         if (likely(!ret))
2247                 ntfs_debug("Done.");
2248         else
2249                 ntfs_warning(vi->i_sb, "Failed to f%ssync inode 0x%lx.  Error "
2250                                 "%u.", datasync ? "data" : "", vi->i_ino, -ret);
2251         return ret;
2252 }
2253
2254 #endif /* NTFS_RW */
2255
2256 const struct file_operations ntfs_file_ops = {
2257         .llseek         = generic_file_llseek,   /* Seek inside file. */
2258         .read           = do_sync_read,          /* Read from file. */
2259         .aio_read       = generic_file_aio_read, /* Async read from file. */
2260 #ifdef NTFS_RW
2261         .write          = ntfs_file_write,       /* Write to file. */
2262         .aio_write      = ntfs_file_aio_write,   /* Async write to file. */
2263         /*.release      = ,*/                    /* Last file is closed.  See
2264                                                     fs/ext2/file.c::
2265                                                     ext2_release_file() for
2266                                                     how to use this to discard
2267                                                     preallocated space for
2268                                                     write opened files. */
2269         .fsync          = ntfs_file_fsync,       /* Sync a file to disk. */
2270         /*.aio_fsync    = ,*/                    /* Sync all outstanding async
2271                                                     i/o operations on a
2272                                                     kiocb. */
2273 #endif /* NTFS_RW */
2274         /*.ioctl        = ,*/                    /* Perform function on the
2275                                                     mounted filesystem. */
2276         .mmap           = generic_file_mmap,     /* Mmap file. */
2277         .open           = ntfs_file_open,        /* Open file. */
2278         .sendfile       = generic_file_sendfile, /* Zero-copy data send with
2279                                                     the data source being on
2280                                                     the ntfs partition.  We do
2281                                                     not need to care about the
2282                                                     data destination. */
2283         /*.sendpage     = ,*/                    /* Zero-copy data send with
2284                                                     the data destination being
2285                                                     on the ntfs partition.  We
2286                                                     do not need to care about
2287                                                     the data source. */
2288 };
2289
2290 const struct inode_operations ntfs_file_inode_ops = {
2291 #ifdef NTFS_RW
2292         .truncate       = ntfs_truncate_vfs,
2293         .setattr        = ntfs_setattr,
2294 #endif /* NTFS_RW */
2295 };
2296
2297 const struct file_operations ntfs_empty_file_ops = {};
2298
2299 const struct inode_operations ntfs_empty_inode_ops = {};