fs/ntfs/file.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * file.c - NTFS kernel file operations.  Part of the Linux-NTFS project.
   4  *
   5  * Copyright (c) 2001-2015 Anton Altaparmakov and Tuxera Inc.
   6  */
   7
   8 #include <linux/backing-dev.h>
   9 #include <linux/buffer_head.h>
  10 #include <linux/gfp.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/pagevec.h>
  13 #include <linux/sched/signal.h>
  14 #include <linux/swap.h>
  15 #include <linux/uio.h>
  16 #include <linux/writeback.h>
  17
  18 #include <asm/page.h>
  19 #include <linux/uaccess.h>
  20
  21 #include "attrib.h"
  22 #include "bitmap.h"
  23 #include "inode.h"
  24 #include "debug.h"
  25 #include "lcnalloc.h"
  26 #include "malloc.h"
  27 #include "mft.h"
  28 #include "ntfs.h"
  29
  30 /**
  31  * ntfs_file_open - called when an inode is about to be opened
  32  * @vi:         inode to be opened
  33  * @filp:       file structure describing the inode
  34  *
  35  * Limit file size to the page cache limit on architectures where unsigned long
  36  * is 32-bits. This is the most we can do for now without overflowing the page
  37  * cache page index. Doing it this way means we don't run into problems because
  38  * of existing too large files. It would be better to allow the user to read
  39  * the beginning of the file but I doubt very much anyone is going to hit this
  40  * check on a 32-bit architecture, so there is no point in adding the extra
  41  * complexity required to support this.
  42  *
  43  * On 64-bit architectures, the check is hopefully optimized away by the
  44  * compiler.
  45  *
  46  * After the check passes, just call generic_file_open() to do its work.
  47  */
  48 static int ntfs_file_open(struct inode *vi, struct file *filp)
  49 {
  50         if (sizeof(unsigned long) < 8) {
  51                 if (i_size_read(vi) > MAX_LFS_FILESIZE)
  52                         return -EOVERFLOW;
  53         }
  54         return generic_file_open(vi, filp);
  55 }
  56
  57 #ifdef NTFS_RW
  58
  59 /**
  60  * ntfs_attr_extend_initialized - extend the initialized size of an attribute
  61  * @ni:                 ntfs inode of the attribute to extend
  62  * @new_init_size:      requested new initialized size in bytes
  63  *
  64  * Extend the initialized size of an attribute described by the ntfs inode @ni
  65  * to @new_init_size bytes.  This involves zeroing any non-sparse space between
  66  * the old initialized size and @new_init_size both in the page cache and on
  67  * disk (if relevant complete pages are already uptodate in the page cache then
  68  * these are simply marked dirty).
  69  *
  70  * As a side-effect, the file size (vfs inode->i_size) may be incremented as,
  71  * in the resident attribute case, it is tied to the initialized size and, in
  72  * the non-resident attribute case, it may not fall below the initialized size.
  73  *
  74  * Note that if the attribute is resident, we do not need to touch the page
  75  * cache at all.  This is because if the page cache page is not uptodate we
  76  * bring it uptodate later, when doing the write to the mft record since we
  77  * then already have the page mapped.  And if the page is uptodate, the
  78  * non-initialized region will already have been zeroed when the page was
  79  * brought uptodate and the region may in fact already have been overwritten
  80  * with new data via mmap() based writes, so we cannot just zero it.  And since
  81  * POSIX specifies that the behaviour of resizing a file whilst it is mmap()ped
  82  * is unspecified, we choose not to do zeroing and thus we do not need to touch
  83  * the page at all.  For a more detailed explanation see ntfs_truncate() in
  84  * fs/ntfs/inode.c.
  85  *
  86  * Return 0 on success and -errno on error.  In the case that an error is
  87  * encountered it is possible that the initialized size will already have been
  88  * incremented some way towards @new_init_size but it is guaranteed that if
  89  * this is the case, the necessary zeroing will also have happened and that all
  90  * metadata is self-consistent.
  91  *
  92  * Locking: i_mutex on the vfs inode corrseponsind to the ntfs inode @ni must be
  93  *          held by the caller.
  94  */
  95 static int ntfs_attr_extend_initialized(ntfs_inode *ni, const s64 new_init_size)
  96 {
  97         s64 old_init_size;
  98         loff_t old_i_size;
  99         pgoff_t index, end_index;
 100         unsigned long flags;
 101         struct inode *vi = VFS_I(ni);
 102         ntfs_inode *base_ni;
 103         MFT_RECORD *m = NULL;
 104         ATTR_RECORD *a;
 105         ntfs_attr_search_ctx *ctx = NULL;
 106         struct address_space *mapping;
 107         struct page *page = NULL;
 108         u8 *kattr;
 109         int err;
 110         u32 attr_len;
 111
 112         read_lock_irqsave(&ni->size_lock, flags);
 113         old_init_size = ni->initialized_size;
 114         old_i_size = i_size_read(vi);
 115         BUG_ON(new_init_size > ni->allocated_size);
 116         read_unlock_irqrestore(&ni->size_lock, flags);
 117         ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, "
 118                         "old_initialized_size 0x%llx, "
 119                         "new_initialized_size 0x%llx, i_size 0x%llx.",
 120                         vi->i_ino, (unsigned)le32_to_cpu(ni->type),
 121                         (unsigned long long)old_init_size,
 122                         (unsigned long long)new_init_size, old_i_size);
 123         if (!NInoAttr(ni))
 124                 base_ni = ni;
 125         else
 126                 base_ni = ni->ext.base_ntfs_ino;
 127         /* Use goto to reduce indentation and we need the label below anyway. */
 128         if (NInoNonResident(ni))
 129                 goto do_non_resident_extend;
 130         BUG_ON(old_init_size != old_i_size);
 131         m = map_mft_record(base_ni);
 132         if (IS_ERR(m)) {
 133                 err = PTR_ERR(m);
 134                 m = NULL;
 135                 goto err_out;
 136         }
 137         ctx = ntfs_attr_get_search_ctx(base_ni, m);
 138         if (unlikely(!ctx)) {
 139                 err = -ENOMEM;
 140                 goto err_out;
 141         }
 142         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
 143                         CASE_SENSITIVE, 0, NULL, 0, ctx);
 144         if (unlikely(err)) {
 145                 if (err == -ENOENT)
 146                         err = -EIO;
 147                 goto err_out;
 148         }
 149         m = ctx->mrec;
 150         a = ctx->attr;
 151         BUG_ON(a->non_resident);
 152         /* The total length of the attribute value. */
 153         attr_len = le32_to_cpu(a->data.resident.value_length);
 154         BUG_ON(old_i_size != (loff_t)attr_len);
 155         /*
 156          * Do the zeroing in the mft record and update the attribute size in
 157          * the mft record.
 158          */
 159         kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
 160         memset(kattr + attr_len, 0, new_init_size - attr_len);
 161         a->data.resident.value_length = cpu_to_le32((u32)new_init_size);
 162         /* Finally, update the sizes in the vfs and ntfs inodes. */
 163         write_lock_irqsave(&ni->size_lock, flags);
 164         i_size_write(vi, new_init_size);
 165         ni->initialized_size = new_init_size;
 166         write_unlock_irqrestore(&ni->size_lock, flags);
 167         goto done;
 168 do_non_resident_extend:
 169         /*
 170          * If the new initialized size @new_init_size exceeds the current file
 171          * size (vfs inode->i_size), we need to extend the file size to the
 172          * new initialized size.
 173          */
 174         if (new_init_size > old_i_size) {
 175                 m = map_mft_record(base_ni);
 176                 if (IS_ERR(m)) {
 177                         err = PTR_ERR(m);
 178                         m = NULL;
 179                         goto err_out;
 180                 }
 181                 ctx = ntfs_attr_get_search_ctx(base_ni, m);
 182                 if (unlikely(!ctx)) {
 183                         err = -ENOMEM;
 184                         goto err_out;
 185                 }
 186                 err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
 187                                 CASE_SENSITIVE, 0, NULL, 0, ctx);
 188                 if (unlikely(err)) {
 189                         if (err == -ENOENT)
 190                                 err = -EIO;
 191                         goto err_out;
 192                 }
 193                 m = ctx->mrec;
 194                 a = ctx->attr;
 195                 BUG_ON(!a->non_resident);
 196                 BUG_ON(old_i_size != (loff_t)
 197                                 sle64_to_cpu(a->data.non_resident.data_size));
 198                 a->data.non_resident.data_size = cpu_to_sle64(new_init_size);
 199                 flush_dcache_mft_record_page(ctx->ntfs_ino);
 200                 mark_mft_record_dirty(ctx->ntfs_ino);
 201                 /* Update the file size in the vfs inode. */
 202                 i_size_write(vi, new_init_size);
 203                 ntfs_attr_put_search_ctx(ctx);
 204                 ctx = NULL;
 205                 unmap_mft_record(base_ni);
 206                 m = NULL;
 207         }
 208         mapping = vi->i_mapping;
 209         index = old_init_size >> PAGE_SHIFT;
 210         end_index = (new_init_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 211         do {
 212                 /*
 213                  * Read the page.  If the page is not present, this will zero
 214                  * the uninitialized regions for us.
 215                  */
 216                 page = read_mapping_page(mapping, index, NULL);
 217                 if (IS_ERR(page)) {
 218                         err = PTR_ERR(page);
 219                         goto init_err_out;
 220                 }
 221                 if (unlikely(PageError(page))) {
 222                         put_page(page);
 223                         err = -EIO;
 224                         goto init_err_out;
 225                 }
 226                 /*
 227                  * Update the initialized size in the ntfs inode.  This is
 228                  * enough to make ntfs_writepage() work.
 229                  */
 230                 write_lock_irqsave(&ni->size_lock, flags);
 231                 ni->initialized_size = (s64)(index + 1) << PAGE_SHIFT;
 232                 if (ni->initialized_size > new_init_size)
 233                         ni->initialized_size = new_init_size;
 234                 write_unlock_irqrestore(&ni->size_lock, flags);
 235                 /* Set the page dirty so it gets written out. */
 236                 set_page_dirty(page);
 237                 put_page(page);
 238                 /*
 239                  * Play nice with the vm and the rest of the system.  This is
 240                  * very much needed as we can potentially be modifying the
 241                  * initialised size from a very small value to a really huge
 242                  * value, e.g.
 243                  *      f = open(somefile, O_TRUNC);
 244                  *      truncate(f, 10GiB);
 245                  *      seek(f, 10GiB);
 246                  *      write(f, 1);
 247                  * And this would mean we would be marking dirty hundreds of
 248                  * thousands of pages or as in the above example more than
 249                  * two and a half million pages!
 250                  *
 251                  * TODO: For sparse pages could optimize this workload by using
 252                  * the FsMisc / MiscFs page bit as a "PageIsSparse" bit.  This
 253                  * would be set in readpage for sparse pages and here we would
 254                  * not need to mark dirty any pages which have this bit set.
 255                  * The only caveat is that we have to clear the bit everywhere
 256                  * where we allocate any clusters that lie in the page or that
 257                  * contain the page.
 258                  *
 259                  * TODO: An even greater optimization would be for us to only
 260                  * call readpage() on pages which are not in sparse regions as
 261                  * determined from the runlist.  This would greatly reduce the
 262                  * number of pages we read and make dirty in the case of sparse
 263                  * files.
 264                  */
 265                 balance_dirty_pages_ratelimited(mapping);
 266                 cond_resched();
 267         } while (++index < end_index);
 268         read_lock_irqsave(&ni->size_lock, flags);
 269         BUG_ON(ni->initialized_size != new_init_size);
 270         read_unlock_irqrestore(&ni->size_lock, flags);
 271         /* Now bring in sync the initialized_size in the mft record. */
 272         m = map_mft_record(base_ni);
 273         if (IS_ERR(m)) {
 274                 err = PTR_ERR(m);
 275                 m = NULL;
 276                 goto init_err_out;
 277         }
 278         ctx = ntfs_attr_get_search_ctx(base_ni, m);
 279         if (unlikely(!ctx)) {
 280                 err = -ENOMEM;
 281                 goto init_err_out;
 282         }
 283         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
 284                         CASE_SENSITIVE, 0, NULL, 0, ctx);
 285         if (unlikely(err)) {
 286                 if (err == -ENOENT)
 287                         err = -EIO;
 288                 goto init_err_out;
 289         }
 290         m = ctx->mrec;
 291         a = ctx->attr;
 292         BUG_ON(!a->non_resident);
 293         a->data.non_resident.initialized_size = cpu_to_sle64(new_init_size);
 294 done:
 295         flush_dcache_mft_record_page(ctx->ntfs_ino);
 296         mark_mft_record_dirty(ctx->ntfs_ino);
 297         if (ctx)
 298                 ntfs_attr_put_search_ctx(ctx);
 299         if (m)
 300                 unmap_mft_record(base_ni);
 301         ntfs_debug("Done, initialized_size 0x%llx, i_size 0x%llx.",
 302                         (unsigned long long)new_init_size, i_size_read(vi));
 303         return 0;
 304 init_err_out:
 305         write_lock_irqsave(&ni->size_lock, flags);
 306         ni->initialized_size = old_init_size;
 307         write_unlock_irqrestore(&ni->size_lock, flags);
 308 err_out:
 309         if (ctx)
 310                 ntfs_attr_put_search_ctx(ctx);
 311         if (m)
 312                 unmap_mft_record(base_ni);
 313         ntfs_debug("Failed.  Returning error code %i.", err);
 314         return err;
 315 }
 316
 317 static ssize_t ntfs_prepare_file_for_write(struct kiocb *iocb,
 318                 struct iov_iter *from)
 319 {
 320         loff_t pos;
 321         s64 end, ll;
 322         ssize_t err;
 323         unsigned long flags;
 324         struct file *file = iocb->ki_filp;
 325         struct inode *vi = file_inode(file);
 326         ntfs_inode *ni = NTFS_I(vi);
 327         ntfs_volume *vol = ni->vol;
 328
 329         ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, pos "
 330                         "0x%llx, count 0x%zx.", vi->i_ino,
 331                         (unsigned)le32_to_cpu(ni->type),
 332                         (unsigned long long)iocb->ki_pos,
 333                         iov_iter_count(from));
 334         err = generic_write_checks(iocb, from);
 335         if (unlikely(err <= 0))
 336                 goto out;
 337         /*
 338          * All checks have passed.  Before we start doing any writing we want
 339          * to abort any totally illegal writes.
 340          */
 341         BUG_ON(NInoMstProtected(ni));
 342         BUG_ON(ni->type != AT_DATA);
 343         /* If file is encrypted, deny access, just like NT4. */
 344         if (NInoEncrypted(ni)) {
 345                 /* Only $DATA attributes can be encrypted. */
 346                 /*
 347                  * Reminder for later: Encrypted files are _always_
 348                  * non-resident so that the content can always be encrypted.
 349                  */
 350                 ntfs_debug("Denying write access to encrypted file.");
 351                 err = -EACCES;
 352                 goto out;
 353         }
 354         if (NInoCompressed(ni)) {
 355                 /* Only unnamed $DATA attribute can be compressed. */
 356                 BUG_ON(ni->name_len);
 357                 /*
 358                  * Reminder for later: If resident, the data is not actually
 359                  * compressed.  Only on the switch to non-resident does
 360                  * compression kick in.  This is in contrast to encrypted files
 361                  * (see above).
 362                  */
 363                 ntfs_error(vi->i_sb, "Writing to compressed files is not "
 364                                 "implemented yet.  Sorry.");
 365                 err = -EOPNOTSUPP;
 366                 goto out;
 367         }
 368         err = file_remove_privs(file);
 369         if (unlikely(err))
 370                 goto out;
 371         /*
 372          * Our ->update_time method always succeeds thus file_update_time()
 373          * cannot fail either so there is no need to check the return code.
 374          */
 375         file_update_time(file);
 376         pos = iocb->ki_pos;
 377         /* The first byte after the last cluster being written to. */
 378         end = (pos + iov_iter_count(from) + vol->cluster_size_mask) &
 379                         ~(u64)vol->cluster_size_mask;
 380         /*
 381          * If the write goes beyond the allocated size, extend the allocation
 382          * to cover the whole of the write, rounded up to the nearest cluster.
 383          */
 384         read_lock_irqsave(&ni->size_lock, flags);
 385         ll = ni->allocated_size;
 386         read_unlock_irqrestore(&ni->size_lock, flags);
 387         if (end > ll) {
 388                 /*
 389                  * Extend the allocation without changing the data size.
 390                  *
 391                  * Note we ensure the allocation is big enough to at least
 392                  * write some data but we do not require the allocation to be
 393                  * complete, i.e. it may be partial.
 394                  */
 395                 ll = ntfs_attr_extend_allocation(ni, end, -1, pos);
 396                 if (likely(ll >= 0)) {
 397                         BUG_ON(pos >= ll);
 398                         /* If the extension was partial truncate the write. */
 399                         if (end > ll) {
 400                                 ntfs_debug("Truncating write to inode 0x%lx, "
 401                                                 "attribute type 0x%x, because "
 402                                                 "the allocation was only "
 403                                                 "partially extended.",
 404                                                 vi->i_ino, (unsigned)
 405                                                 le32_to_cpu(ni->type));
 406                                 iov_iter_truncate(from, ll - pos);
 407                         }
 408                 } else {
 409                         err = ll;
 410                         read_lock_irqsave(&ni->size_lock, flags);
 411                         ll = ni->allocated_size;
 412                         read_unlock_irqrestore(&ni->size_lock, flags);
 413                         /* Perform a partial write if possible or fail. */
 414                         if (pos < ll) {
 415                                 ntfs_debug("Truncating write to inode 0x%lx "
 416                                                 "attribute type 0x%x, because "
 417                                                 "extending the allocation "
 418                                                 "failed (error %d).",
 419                                                 vi->i_ino, (unsigned)
 420                                                 le32_to_cpu(ni->type),
 421                                                 (int)-err);
 422                                 iov_iter_truncate(from, ll - pos);
 423                         } else {
 424                                 if (err != -ENOSPC)
 425                                         ntfs_error(vi->i_sb, "Cannot perform "
 426                                                         "write to inode "
 427                                                         "0x%lx, attribute "
 428                                                         "type 0x%x, because "
 429                                                         "extending the "
 430                                                         "allocation failed "
 431                                                         "(error %ld).",
 432                                                         vi->i_ino, (unsigned)
 433                                                         le32_to_cpu(ni->type),
 434                                                         (long)-err);
 435                                 else
 436                                         ntfs_debug("Cannot perform write to "
 437                                                         "inode 0x%lx, "
 438                                                         "attribute type 0x%x, "
 439                                                         "because there is not "
 440                                                         "space left.",
 441                                                         vi->i_ino, (unsigned)
 442                                                         le32_to_cpu(ni->type));
 443                                 goto out;
 444                         }
 445                 }
 446         }
 447         /*
 448          * If the write starts beyond the initialized size, extend it up to the
 449          * beginning of the write and initialize all non-sparse space between
 450          * the old initialized size and the new one.  This automatically also
 451          * increments the vfs inode->i_size to keep it above or equal to the
 452          * initialized_size.
 453          */
 454         read_lock_irqsave(&ni->size_lock, flags);
 455         ll = ni->initialized_size;
 456         read_unlock_irqrestore(&ni->size_lock, flags);
 457         if (pos > ll) {
 458                 /*
 459                  * Wait for ongoing direct i/o to complete before proceeding.
 460                  * New direct i/o cannot start as we hold i_mutex.
 461                  */
 462                 inode_dio_wait(vi);
 463                 err = ntfs_attr_extend_initialized(ni, pos);
 464                 if (unlikely(err < 0))
 465                         ntfs_error(vi->i_sb, "Cannot perform write to inode "
 466                                         "0x%lx, attribute type 0x%x, because "
 467                                         "extending the initialized size "
 468                                         "failed (error %d).", vi->i_ino,
 469                                         (unsigned)le32_to_cpu(ni->type),
 470                                         (int)-err);
 471         }
 472 out:
 473         return err;
 474 }
 475
 476 /**
 477  * __ntfs_grab_cache_pages - obtain a number of locked pages
 478  * @mapping:    address space mapping from which to obtain page cache pages
 479  * @index:      starting index in @mapping at which to begin obtaining pages
 480  * @nr_pages:   number of page cache pages to obtain
 481  * @pages:      array of pages in which to return the obtained page cache pages
 482  * @cached_page: allocated but as yet unused page
 483  *
 484  * Obtain @nr_pages locked page cache pages from the mapping @mapping and
 485  * starting at index @index.
 486  *
 487  * If a page is newly created, add it to lru list
 488  *
 489  * Note, the page locks are obtained in ascending page index order.
 490  */
 491 static inline int __ntfs_grab_cache_pages(struct address_space *mapping,
 492                 pgoff_t index, const unsigned nr_pages, struct page **pages,
 493                 struct page **cached_page)
 494 {
 495         int err, nr;
 496
 497         BUG_ON(!nr_pages);
 498         err = nr = 0;
 499         do {
 500                 pages[nr] = find_get_page_flags(mapping, index, FGP_LOCK |
 501                                 FGP_ACCESSED);
 502                 if (!pages[nr]) {
 503                         if (!*cached_page) {
 504                                 *cached_page = page_cache_alloc(mapping);
 505                                 if (unlikely(!*cached_page)) {
 506                                         err = -ENOMEM;
 507                                         goto err_out;
 508                                 }
 509                         }
 510                         err = add_to_page_cache_lru(*cached_page, mapping,
 511                                    index,
 512                                    mapping_gfp_constraint(mapping, GFP_KERNEL));
 513                         if (unlikely(err)) {
 514                                 if (err == -EEXIST)
 515                                         continue;
 516                                 goto err_out;
 517                         }
 518                         pages[nr] = *cached_page;
 519                         *cached_page = NULL;
 520                 }
 521                 index++;
 522                 nr++;
 523         } while (nr < nr_pages);
 524 out:
 525         return err;
 526 err_out:
 527         while (nr > 0) {
 528                 unlock_page(pages[--nr]);
 529                 put_page(pages[nr]);
 530         }
 531         goto out;
 532 }
 533
 534 static inline int ntfs_submit_bh_for_read(struct buffer_head *bh)
 535 {
 536         lock_buffer(bh);
 537         get_bh(bh);
 538         bh->b_end_io = end_buffer_read_sync;
 539         return submit_bh(REQ_OP_READ, 0, bh);
 540 }
 541
 542 /**
 543  * ntfs_prepare_pages_for_non_resident_write - prepare pages for receiving data
 544  * @pages:      array of destination pages
 545  * @nr_pages:   number of pages in @pages
 546  * @pos:        byte position in file at which the write begins
 547  * @bytes:      number of bytes to be written
 548  *
 549  * This is called for non-resident attributes from ntfs_file_buffered_write()
 550  * with i_mutex held on the inode (@pages[0]->mapping->host).  There are
 551  * @nr_pages pages in @pages which are locked but not kmap()ped.  The source
 552  * data has not yet been copied into the @pages.
 553  *
 554  * Need to fill any holes with actual clusters, allocate buffers if necessary,
 555  * ensure all the buffers are mapped, and bring uptodate any buffers that are
 556  * only partially being written to.
 557  *
 558  * If @nr_pages is greater than one, we are guaranteed that the cluster size is
 559  * greater than PAGE_SIZE, that all pages in @pages are entirely inside
 560  * the same cluster and that they are the entirety of that cluster, and that
 561  * the cluster is sparse, i.e. we need to allocate a cluster to fill the hole.
 562  *
 563  * i_size is not to be modified yet.
 564  *
 565  * Return 0 on success or -errno on error.
 566  */
 567 static int ntfs_prepare_pages_for_non_resident_write(struct page **pages,
 568                 unsigned nr_pages, s64 pos, size_t bytes)
 569 {
 570         VCN vcn, highest_vcn = 0, cpos, cend, bh_cpos, bh_cend;
 571         LCN lcn;
 572         s64 bh_pos, vcn_len, end, initialized_size;
 573         sector_t lcn_block;
 574         struct page *page;
 575         struct inode *vi;
 576         ntfs_inode *ni, *base_ni = NULL;
 577         ntfs_volume *vol;
 578         runlist_element *rl, *rl2;
 579         struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
 580         ntfs_attr_search_ctx *ctx = NULL;
 581         MFT_RECORD *m = NULL;
 582         ATTR_RECORD *a = NULL;
 583         unsigned long flags;
 584         u32 attr_rec_len = 0;
 585         unsigned blocksize, u;
 586         int err, mp_size;
 587         bool rl_write_locked, was_hole, is_retry;
 588         unsigned char blocksize_bits;
 589         struct {
 590                 u8 runlist_merged:1;
 591                 u8 mft_attr_mapped:1;
 592                 u8 mp_rebuilt:1;
 593                 u8 attr_switched:1;
 594         } status = { 0, 0, 0, 0 };
 595
 596         BUG_ON(!nr_pages);
 597         BUG_ON(!pages);
 598         BUG_ON(!*pages);
 599         vi = pages[0]->mapping->host;
 600         ni = NTFS_I(vi);
 601         vol = ni->vol;
 602         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, start page "
 603                         "index 0x%lx, nr_pages 0x%x, pos 0x%llx, bytes 0x%zx.",
 604                         vi->i_ino, ni->type, pages[0]->index, nr_pages,
 605                         (long long)pos, bytes);
 606         blocksize = vol->sb->s_blocksize;
 607         blocksize_bits = vol->sb->s_blocksize_bits;
 608         u = 0;
 609         do {
 610                 page = pages[u];
 611                 BUG_ON(!page);
 612                 /*
 613                  * create_empty_buffers() will create uptodate/dirty buffers if
 614                  * the page is uptodate/dirty.
 615                  */
 616                 if (!page_has_buffers(page)) {
 617                         create_empty_buffers(page, blocksize, 0);
 618                         if (unlikely(!page_has_buffers(page)))
 619                                 return -ENOMEM;
 620                 }
 621         } while (++u < nr_pages);
 622         rl_write_locked = false;
 623         rl = NULL;
 624         err = 0;
 625         vcn = lcn = -1;
 626         vcn_len = 0;
 627         lcn_block = -1;
 628         was_hole = false;
 629         cpos = pos >> vol->cluster_size_bits;
 630         end = pos + bytes;
 631         cend = (end + vol->cluster_size - 1) >> vol->cluster_size_bits;
 632         /*
 633          * Loop over each page and for each page over each buffer.  Use goto to
 634          * reduce indentation.
 635          */
 636         u = 0;
 637 do_next_page:
 638         page = pages[u];
 639         bh_pos = (s64)page->index << PAGE_SHIFT;
 640         bh = head = page_buffers(page);
 641         do {
 642                 VCN cdelta;
 643                 s64 bh_end;
 644                 unsigned bh_cofs;
 645
 646                 /* Clear buffer_new on all buffers to reinitialise state. */
 647                 if (buffer_new(bh))
 648                         clear_buffer_new(bh);
 649                 bh_end = bh_pos + blocksize;
 650                 bh_cpos = bh_pos >> vol->cluster_size_bits;
 651                 bh_cofs = bh_pos & vol->cluster_size_mask;
 652                 if (buffer_mapped(bh)) {
 653                         /*
 654                          * The buffer is already mapped.  If it is uptodate,
 655                          * ignore it.
 656                          */
 657                         if (buffer_uptodate(bh))
 658                                 continue;
 659                         /*
 660                          * The buffer is not uptodate.  If the page is uptodate
 661                          * set the buffer uptodate and otherwise ignore it.
 662                          */
 663                         if (PageUptodate(page)) {
 664                                 set_buffer_uptodate(bh);
 665                                 continue;
 666                         }
 667                         /*
 668                          * Neither the page nor the buffer are uptodate.  If
 669                          * the buffer is only partially being written to, we
 670                          * need to read it in before the write, i.e. now.
 671                          */
 672                         if ((bh_pos < pos && bh_end > pos) ||
 673                                         (bh_pos < end && bh_end > end)) {
 674                                 /*
 675                                  * If the buffer is fully or partially within
 676                                  * the initialized size, do an actual read.
 677                                  * Otherwise, simply zero the buffer.
 678                                  */
 679                                 read_lock_irqsave(&ni->size_lock, flags);
 680                                 initialized_size = ni->initialized_size;
 681                                 read_unlock_irqrestore(&ni->size_lock, flags);
 682                                 if (bh_pos < initialized_size) {
 683                                         ntfs_submit_bh_for_read(bh);
 684                                         *wait_bh++ = bh;
 685                                 } else {
 686                                         zero_user(page, bh_offset(bh),
 687                                                         blocksize);
 688                                         set_buffer_uptodate(bh);
 689                                 }
 690                         }
 691                         continue;
 692                 }
 693                 /* Unmapped buffer.  Need to map it. */
 694                 bh->b_bdev = vol->sb->s_bdev;
 695                 /*
 696                  * If the current buffer is in the same clusters as the map
 697                  * cache, there is no need to check the runlist again.  The
 698                  * map cache is made up of @vcn, which is the first cached file
 699                  * cluster, @vcn_len which is the number of cached file
 700                  * clusters, @lcn is the device cluster corresponding to @vcn,
 701                  * and @lcn_block is the block number corresponding to @lcn.
 702                  */
 703                 cdelta = bh_cpos - vcn;
 704                 if (likely(!cdelta || (cdelta > 0 && cdelta < vcn_len))) {
 705 map_buffer_cached:
 706                         BUG_ON(lcn < 0);
 707                         bh->b_blocknr = lcn_block +
 708                                         (cdelta << (vol->cluster_size_bits -
 709                                         blocksize_bits)) +
 710                                         (bh_cofs >> blocksize_bits);
 711                         set_buffer_mapped(bh);
 712                         /*
 713                          * If the page is uptodate so is the buffer.  If the
 714                          * buffer is fully outside the write, we ignore it if
 715                          * it was already allocated and we mark it dirty so it
 716                          * gets written out if we allocated it.  On the other
 717                          * hand, if we allocated the buffer but we are not
 718                          * marking it dirty we set buffer_new so we can do
 719                          * error recovery.
 720                          */
 721                         if (PageUptodate(page)) {
 722                                 if (!buffer_uptodate(bh))
 723                                         set_buffer_uptodate(bh);
 724                                 if (unlikely(was_hole)) {
 725                                         /* We allocated the buffer. */
 726                                         clean_bdev_bh_alias(bh);
 727                                         if (bh_end <= pos || bh_pos >= end)
 728                                                 mark_buffer_dirty(bh);
 729                                         else
 730                                                 set_buffer_new(bh);
 731                                 }
 732                                 continue;
 733                         }
 734                         /* Page is _not_ uptodate. */
 735                         if (likely(!was_hole)) {
 736                                 /*
 737                                  * Buffer was already allocated.  If it is not
 738                                  * uptodate and is only partially being written
 739                                  * to, we need to read it in before the write,
 740                                  * i.e. now.
 741                                  */
 742                                 if (!buffer_uptodate(bh) && bh_pos < end &&
 743                                                 bh_end > pos &&
 744                                                 (bh_pos < pos ||
 745                                                 bh_end > end)) {
 746                                         /*
 747                                          * If the buffer is fully or partially
 748                                          * within the initialized size, do an
 749                                          * actual read.  Otherwise, simply zero
 750                                          * the buffer.
 751                                          */
 752                                         read_lock_irqsave(&ni->size_lock,
 753                                                         flags);
 754                                         initialized_size = ni->initialized_size;
 755                                         read_unlock_irqrestore(&ni->size_lock,
 756                                                         flags);
 757                                         if (bh_pos < initialized_size) {
 758                                                 ntfs_submit_bh_for_read(bh);
 759                                                 *wait_bh++ = bh;
 760                                         } else {
 761                                                 zero_user(page, bh_offset(bh),
 762                                                                 blocksize);
 763                                                 set_buffer_uptodate(bh);
 764                                         }
 765                                 }
 766                                 continue;
 767                         }
 768                         /* We allocated the buffer. */
 769                         clean_bdev_bh_alias(bh);
 770                         /*
 771                          * If the buffer is fully outside the write, zero it,
 772                          * set it uptodate, and mark it dirty so it gets
 773                          * written out.  If it is partially being written to,
 774                          * zero region surrounding the write but leave it to
 775                          * commit write to do anything else.  Finally, if the
 776                          * buffer is fully being overwritten, do nothing.
 777                          */
 778                         if (bh_end <= pos || bh_pos >= end) {
 779                                 if (!buffer_uptodate(bh)) {
 780                                         zero_user(page, bh_offset(bh),
 781                                                         blocksize);
 782                                         set_buffer_uptodate(bh);
 783                                 }
 784                                 mark_buffer_dirty(bh);
 785                                 continue;
 786                         }
 787                         set_buffer_new(bh);
 788                         if (!buffer_uptodate(bh) &&
 789                                         (bh_pos < pos || bh_end > end)) {
 790                                 u8 *kaddr;
 791                                 unsigned pofs;
 792
 793                                 kaddr = kmap_atomic(page);
 794                                 if (bh_pos < pos) {
 795                                         pofs = bh_pos & ~PAGE_MASK;
 796                                         memset(kaddr + pofs, 0, pos - bh_pos);
 797                                 }
 798                                 if (bh_end > end) {
 799                                         pofs = end & ~PAGE_MASK;
 800                                         memset(kaddr + pofs, 0, bh_end - end);
 801                                 }
 802                                 kunmap_atomic(kaddr);
 803                                 flush_dcache_page(page);
 804                         }
 805                         continue;
 806                 }
 807                 /*
 808                  * Slow path: this is the first buffer in the cluster.  If it
 809                  * is outside allocated size and is not uptodate, zero it and
 810                  * set it uptodate.
 811                  */
 812                 read_lock_irqsave(&ni->size_lock, flags);
 813                 initialized_size = ni->allocated_size;
 814                 read_unlock_irqrestore(&ni->size_lock, flags);
 815                 if (bh_pos > initialized_size) {
 816                         if (PageUptodate(page)) {
 817                                 if (!buffer_uptodate(bh))
 818                                         set_buffer_uptodate(bh);
 819                         } else if (!buffer_uptodate(bh)) {
 820                                 zero_user(page, bh_offset(bh), blocksize);
 821                                 set_buffer_uptodate(bh);
 822                         }
 823                         continue;
 824                 }
 825                 is_retry = false;
 826                 if (!rl) {
 827                         down_read(&ni->runlist.lock);
 828 retry_remap:
 829                         rl = ni->runlist.rl;
 830                 }
 831                 if (likely(rl != NULL)) {
 832                         /* Seek to element containing target cluster. */
 833                         while (rl->length && rl[1].vcn <= bh_cpos)
 834                                 rl++;
 835                         lcn = ntfs_rl_vcn_to_lcn(rl, bh_cpos);
 836                         if (likely(lcn >= 0)) {
 837                                 /*
 838                                  * Successful remap, setup the map cache and
 839                                  * use that to deal with the buffer.
 840                                  */
 841                                 was_hole = false;
 842                                 vcn = bh_cpos;
 843                                 vcn_len = rl[1].vcn - vcn;
 844                                 lcn_block = lcn << (vol->cluster_size_bits -
 845                                                 blocksize_bits);
 846                                 cdelta = 0;
 847                                 /*
 848                                  * If the number of remaining clusters touched
 849                                  * by the write is smaller or equal to the
 850                                  * number of cached clusters, unlock the
 851                                  * runlist as the map cache will be used from
 852                                  * now on.
 853                                  */
 854                                 if (likely(vcn + vcn_len >= cend)) {
 855                                         if (rl_write_locked) {
 856                                                 up_write(&ni->runlist.lock);
 857                                                 rl_write_locked = false;
 858                                         } else
 859                                                 up_read(&ni->runlist.lock);
 860                                         rl = NULL;
 861                                 }
 862                                 goto map_buffer_cached;
 863                         }
 864                 } else
 865                         lcn = LCN_RL_NOT_MAPPED;
 866                 /*
 867                  * If it is not a hole and not out of bounds, the runlist is
 868                  * probably unmapped so try to map it now.
 869                  */
 870                 if (unlikely(lcn != LCN_HOLE && lcn != LCN_ENOENT)) {
 871                         if (likely(!is_retry && lcn == LCN_RL_NOT_MAPPED)) {
 872                                 /* Attempt to map runlist. */
 873                                 if (!rl_write_locked) {
 874                                         /*
 875                                          * We need the runlist locked for
 876                                          * writing, so if it is locked for
 877                                          * reading relock it now and retry in
 878                                          * case it changed whilst we dropped
 879                                          * the lock.
 880                                          */
 881                                         up_read(&ni->runlist.lock);
 882                                         down_write(&ni->runlist.lock);
 883                                         rl_write_locked = true;
 884                                         goto retry_remap;
 885                                 }
 886                                 err = ntfs_map_runlist_nolock(ni, bh_cpos,
 887                                                 NULL);
 888                                 if (likely(!err)) {
 889                                         is_retry = true;
 890                                         goto retry_remap;
 891                                 }
 892                                 /*
 893                                  * If @vcn is out of bounds, pretend @lcn is
 894                                  * LCN_ENOENT.  As long as the buffer is out
 895                                  * of bounds this will work fine.
 896                                  */
 897                                 if (err == -ENOENT) {
 898                                         lcn = LCN_ENOENT;
 899                                         err = 0;
 900                                         goto rl_not_mapped_enoent;
 901                                 }
 902                         } else
 903                                 err = -EIO;
 904                         /* Failed to map the buffer, even after retrying. */
 905                         bh->b_blocknr = -1;
 906                         ntfs_error(vol->sb, "Failed to write to inode 0x%lx, "
 907                                         "attribute type 0x%x, vcn 0x%llx, "
 908                                         "vcn offset 0x%x, because its "
 909                                         "location on disk could not be "
 910                                         "determined%s (error code %i).",
 911                                         ni->mft_no, ni->type,
 912                                         (unsigned long long)bh_cpos,
 913                                         (unsigned)bh_pos &
 914                                         vol->cluster_size_mask,
 915                                         is_retry ? " even after retrying" : "",
 916                                         err);
 917                         break;
 918                 }
 919 rl_not_mapped_enoent:
 920                 /*
 921                  * The buffer is in a hole or out of bounds.  We need to fill
 922                  * the hole, unless the buffer is in a cluster which is not
 923                  * touched by the write, in which case we just leave the buffer
 924                  * unmapped.  This can only happen when the cluster size is
 925                  * less than the page cache size.
 926                  */
 927                 if (unlikely(vol->cluster_size < PAGE_SIZE)) {
 928                         bh_cend = (bh_end + vol->cluster_size - 1) >>
 929                                         vol->cluster_size_bits;
 930                         if ((bh_cend <= cpos || bh_cpos >= cend)) {
 931                                 bh->b_blocknr = -1;
 932                                 /*
 933                                  * If the buffer is uptodate we skip it.  If it
 934                                  * is not but the page is uptodate, we can set
 935                                  * the buffer uptodate.  If the page is not
 936                                  * uptodate, we can clear the buffer and set it
 937                                  * uptodate.  Whether this is worthwhile is
 938                                  * debatable and this could be removed.
 939                                  */
 940                                 if (PageUptodate(page)) {
 941                                         if (!buffer_uptodate(bh))
 942                                                 set_buffer_uptodate(bh);
 943                                 } else if (!buffer_uptodate(bh)) {
 944                                         zero_user(page, bh_offset(bh),
 945                                                 blocksize);
 946                                         set_buffer_uptodate(bh);
 947                                 }
 948                                 continue;
 949                         }
 950                 }
 951                 /*
 952                  * Out of bounds buffer is invalid if it was not really out of
 953                  * bounds.
 954                  */
 955                 BUG_ON(lcn != LCN_HOLE);
 956                 /*
 957                  * We need the runlist locked for writing, so if it is locked
 958                  * for reading relock it now and retry in case it changed
 959                  * whilst we dropped the lock.
 960                  */
 961                 BUG_ON(!rl);
 962                 if (!rl_write_locked) {
 963                         up_read(&ni->runlist.lock);
 964                         down_write(&ni->runlist.lock);
 965                         rl_write_locked = true;
 966                         goto retry_remap;
 967                 }
 968                 /* Find the previous last allocated cluster. */
 969                 BUG_ON(rl->lcn != LCN_HOLE);
 970                 lcn = -1;
 971                 rl2 = rl;
 972                 while (--rl2 >= ni->runlist.rl) {
 973                         if (rl2->lcn >= 0) {
 974                                 lcn = rl2->lcn + rl2->length;
 975                                 break;
 976                         }
 977                 }
 978                 rl2 = ntfs_cluster_alloc(vol, bh_cpos, 1, lcn, DATA_ZONE,
 979                                 false);
 980                 if (IS_ERR(rl2)) {
 981                         err = PTR_ERR(rl2);
 982                         ntfs_debug("Failed to allocate cluster, error code %i.",
 983                                         err);
 984                         break;
 985                 }
 986                 lcn = rl2->lcn;
 987                 rl = ntfs_runlists_merge(ni->runlist.rl, rl2);
 988                 if (IS_ERR(rl)) {
 989                         err = PTR_ERR(rl);
 990                         if (err != -ENOMEM)
 991                                 err = -EIO;
 992                         if (ntfs_cluster_free_from_rl(vol, rl2)) {
 993                                 ntfs_error(vol->sb, "Failed to release "
 994                                                 "allocated cluster in error "
 995                                                 "code path.  Run chkdsk to "
 996                                                 "recover the lost cluster.");
 997                                 NVolSetErrors(vol);
 998                         }
 999                         ntfs_free(rl2);
1000                         break;
1001                 }
1002                 ni->runlist.rl = rl;
1003                 status.runlist_merged = 1;
1004                 ntfs_debug("Allocated cluster, lcn 0x%llx.",
1005                                 (unsigned long long)lcn);
1006                 /* Map and lock the mft record and get the attribute record. */
1007                 if (!NInoAttr(ni))
1008                         base_ni = ni;
1009                 else
1010                         base_ni = ni->ext.base_ntfs_ino;
1011                 m = map_mft_record(base_ni);
1012                 if (IS_ERR(m)) {
1013                         err = PTR_ERR(m);
1014                         break;
1015                 }
1016                 ctx = ntfs_attr_get_search_ctx(base_ni, m);
1017                 if (unlikely(!ctx)) {
1018                         err = -ENOMEM;
1019                         unmap_mft_record(base_ni);
1020                         break;
1021                 }
1022                 status.mft_attr_mapped = 1;
1023                 err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1024                                 CASE_SENSITIVE, bh_cpos, NULL, 0, ctx);
1025                 if (unlikely(err)) {
1026                         if (err == -ENOENT)
1027                                 err = -EIO;
1028                         break;
1029                 }
1030                 m = ctx->mrec;
1031                 a = ctx->attr;
1032                 /*
1033                  * Find the runlist element with which the attribute extent
1034                  * starts.  Note, we cannot use the _attr_ version because we
1035                  * have mapped the mft record.  That is ok because we know the
1036                  * runlist fragment must be mapped already to have ever gotten
1037                  * here, so we can just use the _rl_ version.
1038                  */
1039                 vcn = sle64_to_cpu(a->data.non_resident.lowest_vcn);
1040                 rl2 = ntfs_rl_find_vcn_nolock(rl, vcn);
1041                 BUG_ON(!rl2);
1042                 BUG_ON(!rl2->length);
1043                 BUG_ON(rl2->lcn < LCN_HOLE);
1044                 highest_vcn = sle64_to_cpu(a->data.non_resident.highest_vcn);
1045                 /*
1046                  * If @highest_vcn is zero, calculate the real highest_vcn
1047                  * (which can really be zero).
1048                  */
1049                 if (!highest_vcn)
1050                         highest_vcn = (sle64_to_cpu(
1051                                         a->data.non_resident.allocated_size) >>
1052                                         vol->cluster_size_bits) - 1;
1053                 /*
1054                  * Determine the size of the mapping pairs array for the new
1055                  * extent, i.e. the old extent with the hole filled.
1056                  */
1057                 mp_size = ntfs_get_size_for_mapping_pairs(vol, rl2, vcn,
1058                                 highest_vcn);
1059                 if (unlikely(mp_size <= 0)) {
1060                         if (!(err = mp_size))
1061                                 err = -EIO;
1062                         ntfs_debug("Failed to get size for mapping pairs "
1063                                         "array, error code %i.", err);
1064                         break;
1065                 }
1066                 /*
1067                  * Resize the attribute record to fit the new mapping pairs
1068                  * array.
1069                  */
1070                 attr_rec_len = le32_to_cpu(a->length);
1071                 err = ntfs_attr_record_resize(m, a, mp_size + le16_to_cpu(
1072                                 a->data.non_resident.mapping_pairs_offset));
1073                 if (unlikely(err)) {
1074                         BUG_ON(err != -ENOSPC);
1075                         // TODO: Deal with this by using the current attribute
1076                         // and fill it with as much of the mapping pairs
1077                         // array as possible.  Then loop over each attribute
1078                         // extent rewriting the mapping pairs arrays as we go
1079                         // along and if when we reach the end we have not
1080                         // enough space, try to resize the last attribute
1081                         // extent and if even that fails, add a new attribute
1082                         // extent.
1083                         // We could also try to resize at each step in the hope
1084                         // that we will not need to rewrite every single extent.
1085                         // Note, we may need to decompress some extents to fill
1086                         // the runlist as we are walking the extents...
1087                         ntfs_error(vol->sb, "Not enough space in the mft "
1088                                         "record for the extended attribute "
1089                                         "record.  This case is not "
1090                                         "implemented yet.");
1091                         err = -EOPNOTSUPP;
1092                         break ;
1093                 }
1094                 status.mp_rebuilt = 1;
1095                 /*
1096                  * Generate the mapping pairs array directly into the attribute
1097                  * record.
1098                  */
1099                 err = ntfs_mapping_pairs_build(vol, (u8*)a + le16_to_cpu(
1100                                 a->data.non_resident.mapping_pairs_offset),
1101                                 mp_size, rl2, vcn, highest_vcn, NULL);
1102                 if (unlikely(err)) {
1103                         ntfs_error(vol->sb, "Cannot fill hole in inode 0x%lx, "
1104                                         "attribute type 0x%x, because building "
1105                                         "the mapping pairs failed with error "
1106                                         "code %i.", vi->i_ino,
1107                                         (unsigned)le32_to_cpu(ni->type), err);
1108                         err = -EIO;
1109                         break;
1110                 }
1111                 /* Update the highest_vcn but only if it was not set. */
1112                 if (unlikely(!a->data.non_resident.highest_vcn))
1113                         a->data.non_resident.highest_vcn =
1114                                         cpu_to_sle64(highest_vcn);
1115                 /*
1116                  * If the attribute is sparse/compressed, update the compressed
1117                  * size in the ntfs_inode structure and the attribute record.
1118                  */
1119                 if (likely(NInoSparse(ni) || NInoCompressed(ni))) {
1120                         /*
1121                          * If we are not in the first attribute extent, switch
1122                          * to it, but first ensure the changes will make it to
1123                          * disk later.
1124                          */
1125                         if (a->data.non_resident.lowest_vcn) {
1126                                 flush_dcache_mft_record_page(ctx->ntfs_ino);
1127                                 mark_mft_record_dirty(ctx->ntfs_ino);
1128                                 ntfs_attr_reinit_search_ctx(ctx);
1129                                 err = ntfs_attr_lookup(ni->type, ni->name,
1130                                                 ni->name_len, CASE_SENSITIVE,
1131                                                 0, NULL, 0, ctx);
1132                                 if (unlikely(err)) {
1133                                         status.attr_switched = 1;
1134                                         break;
1135                                 }
1136                                 /* @m is not used any more so do not set it. */
1137                                 a = ctx->attr;
1138                         }
1139                         write_lock_irqsave(&ni->size_lock, flags);
1140                         ni->itype.compressed.size += vol->cluster_size;
1141                         a->data.non_resident.compressed_size =
1142                                         cpu_to_sle64(ni->itype.compressed.size);
1143                         write_unlock_irqrestore(&ni->size_lock, flags);
1144                 }
1145                 /* Ensure the changes make it to disk. */
1146                 flush_dcache_mft_record_page(ctx->ntfs_ino);
1147                 mark_mft_record_dirty(ctx->ntfs_ino);
1148                 ntfs_attr_put_search_ctx(ctx);
1149                 unmap_mft_record(base_ni);
1150                 /* Successfully filled the hole. */
1151                 status.runlist_merged = 0;
1152                 status.mft_attr_mapped = 0;
1153                 status.mp_rebuilt = 0;
1154                 /* Setup the map cache and use that to deal with the buffer. */
1155                 was_hole = true;
1156                 vcn = bh_cpos;
1157                 vcn_len = 1;
1158                 lcn_block = lcn << (vol->cluster_size_bits - blocksize_bits);
1159                 cdelta = 0;
1160                 /*
1161                  * If the number of remaining clusters in the @pages is smaller
1162                  * or equal to the number of cached clusters, unlock the
1163                  * runlist as the map cache will be used from now on.
1164                  */
1165                 if (likely(vcn + vcn_len >= cend)) {
1166                         up_write(&ni->runlist.lock);
1167                         rl_write_locked = false;
1168                         rl = NULL;
1169                 }
1170                 goto map_buffer_cached;
1171         } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
1172         /* If there are no errors, do the next page. */
1173         if (likely(!err && ++u < nr_pages))
1174                 goto do_next_page;
1175         /* If there are no errors, release the runlist lock if we took it. */
1176         if (likely(!err)) {
1177                 if (unlikely(rl_write_locked)) {
1178                         up_write(&ni->runlist.lock);
1179                         rl_write_locked = false;
1180                 } else if (unlikely(rl))
1181                         up_read(&ni->runlist.lock);
1182                 rl = NULL;
1183         }
1184         /* If we issued read requests, let them complete. */
1185         read_lock_irqsave(&ni->size_lock, flags);
1186         initialized_size = ni->initialized_size;
1187         read_unlock_irqrestore(&ni->size_lock, flags);
1188         while (wait_bh > wait) {
1189                 bh = *--wait_bh;
1190                 wait_on_buffer(bh);
1191                 if (likely(buffer_uptodate(bh))) {
1192                         page = bh->b_page;
1193                         bh_pos = ((s64)page->index << PAGE_SHIFT) +
1194                                         bh_offset(bh);
1195                         /*
1196                          * If the buffer overflows the initialized size, need
1197                          * to zero the overflowing region.
1198                          */
1199                         if (unlikely(bh_pos + blocksize > initialized_size)) {
1200                                 int ofs = 0;
1201
1202                                 if (likely(bh_pos < initialized_size))
1203                                         ofs = initialized_size - bh_pos;
1204                                 zero_user_segment(page, bh_offset(bh) + ofs,
1205                                                 blocksize);
1206                         }
1207                 } else /* if (unlikely(!buffer_uptodate(bh))) */
1208                         err = -EIO;
1209         }
1210         if (likely(!err)) {
1211                 /* Clear buffer_new on all buffers. */
1212                 u = 0;
1213                 do {
1214                         bh = head = page_buffers(pages[u]);
1215                         do {
1216                                 if (buffer_new(bh))
1217                                         clear_buffer_new(bh);
1218                         } while ((bh = bh->b_this_page) != head);
1219                 } while (++u < nr_pages);
1220                 ntfs_debug("Done.");
1221                 return err;
1222         }
1223         if (status.attr_switched) {
1224                 /* Get back to the attribute extent we modified. */
1225                 ntfs_attr_reinit_search_ctx(ctx);
1226                 if (ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1227                                 CASE_SENSITIVE, bh_cpos, NULL, 0, ctx)) {
1228                         ntfs_error(vol->sb, "Failed to find required "
1229                                         "attribute extent of attribute in "
1230                                         "error code path.  Run chkdsk to "
1231                                         "recover.");
1232                         write_lock_irqsave(&ni->size_lock, flags);
1233                         ni->itype.compressed.size += vol->cluster_size;
1234                         write_unlock_irqrestore(&ni->size_lock, flags);
1235                         flush_dcache_mft_record_page(ctx->ntfs_ino);
1236                         mark_mft_record_dirty(ctx->ntfs_ino);
1237                         /*
1238                          * The only thing that is now wrong is the compressed
1239                          * size of the base attribute extent which chkdsk
1240                          * should be able to fix.
1241                          */
1242                         NVolSetErrors(vol);
1243                 } else {
1244                         m = ctx->mrec;
1245                         a = ctx->attr;
1246                         status.attr_switched = 0;
1247                 }
1248         }
1249         /*
1250          * If the runlist has been modified, need to restore it by punching a
1251          * hole into it and we then need to deallocate the on-disk cluster as
1252          * well.  Note, we only modify the runlist if we are able to generate a
1253          * new mapping pairs array, i.e. only when the mapped attribute extent
1254          * is not switched.
1255          */
1256         if (status.runlist_merged && !status.attr_switched) {
1257                 BUG_ON(!rl_write_locked);
1258                 /* Make the file cluster we allocated sparse in the runlist. */
1259                 if (ntfs_rl_punch_nolock(vol, &ni->runlist, bh_cpos, 1)) {
1260                         ntfs_error(vol->sb, "Failed to punch hole into "
1261                                         "attribute runlist in error code "
1262                                         "path.  Run chkdsk to recover the "
1263                                         "lost cluster.");
1264                         NVolSetErrors(vol);
1265                 } else /* if (success) */ {
1266                         status.runlist_merged = 0;
1267                         /*
1268                          * Deallocate the on-disk cluster we allocated but only
1269                          * if we succeeded in punching its vcn out of the
1270                          * runlist.
1271                          */
1272                         down_write(&vol->lcnbmp_lock);
1273                         if (ntfs_bitmap_clear_bit(vol->lcnbmp_ino, lcn)) {
1274                                 ntfs_error(vol->sb, "Failed to release "
1275                                                 "allocated cluster in error "
1276                                                 "code path.  Run chkdsk to "
1277                                                 "recover the lost cluster.");
1278                                 NVolSetErrors(vol);
1279                         }
1280                         up_write(&vol->lcnbmp_lock);
1281                 }
1282         }
1283         /*
1284          * Resize the attribute record to its old size and rebuild the mapping
1285          * pairs array.  Note, we only can do this if the runlist has been
1286          * restored to its old state which also implies that the mapped
1287          * attribute extent is not switched.
1288          */
1289         if (status.mp_rebuilt && !status.runlist_merged) {
1290                 if (ntfs_attr_record_resize(m, a, attr_rec_len)) {
1291                         ntfs_error(vol->sb, "Failed to restore attribute "
1292                                         "record in error code path.  Run "
1293                                         "chkdsk to recover.");
1294                         NVolSetErrors(vol);
1295                 } else /* if (success) */ {
1296                         if (ntfs_mapping_pairs_build(vol, (u8*)a +
1297                                         le16_to_cpu(a->data.non_resident.
1298                                         mapping_pairs_offset), attr_rec_len -
1299                                         le16_to_cpu(a->data.non_resident.
1300                                         mapping_pairs_offset), ni->runlist.rl,
1301                                         vcn, highest_vcn, NULL)) {
1302                                 ntfs_error(vol->sb, "Failed to restore "
1303                                                 "mapping pairs array in error "
1304                                                 "code path.  Run chkdsk to "
1305                                                 "recover.");
1306                                 NVolSetErrors(vol);
1307                         }
1308                         flush_dcache_mft_record_page(ctx->ntfs_ino);
1309                         mark_mft_record_dirty(ctx->ntfs_ino);
1310                 }
1311         }
1312         /* Release the mft record and the attribute. */
1313         if (status.mft_attr_mapped) {
1314                 ntfs_attr_put_search_ctx(ctx);
1315                 unmap_mft_record(base_ni);
1316         }
1317         /* Release the runlist lock. */
1318         if (rl_write_locked)
1319                 up_write(&ni->runlist.lock);
1320         else if (rl)
1321                 up_read(&ni->runlist.lock);
1322         /*
1323          * Zero out any newly allocated blocks to avoid exposing stale data.
1324          * If BH_New is set, we know that the block was newly allocated above
1325          * and that it has not been fully zeroed and marked dirty yet.
1326          */
1327         nr_pages = u;
1328         u = 0;
1329         end = bh_cpos << vol->cluster_size_bits;
1330         do {
1331                 page = pages[u];
1332                 bh = head = page_buffers(page);
1333                 do {
1334                         if (u == nr_pages &&
1335                                         ((s64)page->index << PAGE_SHIFT) +
1336                                         bh_offset(bh) >= end)
1337                                 break;
1338                         if (!buffer_new(bh))
1339                                 continue;
1340                         clear_buffer_new(bh);
1341                         if (!buffer_uptodate(bh)) {
1342                                 if (PageUptodate(page))
1343                                         set_buffer_uptodate(bh);
1344                                 else {
1345                                         zero_user(page, bh_offset(bh),
1346                                                         blocksize);
1347                                         set_buffer_uptodate(bh);
1348                                 }
1349                         }
1350                         mark_buffer_dirty(bh);
1351                 } while ((bh = bh->b_this_page) != head);
1352         } while (++u <= nr_pages);
1353         ntfs_error(vol->sb, "Failed.  Returning error code %i.", err);
1354         return err;
1355 }
1356
1357 static inline void ntfs_flush_dcache_pages(struct page **pages,
1358                 unsigned nr_pages)
1359 {
1360         BUG_ON(!nr_pages);
1361         /*
1362          * Warning: Do not do the decrement at the same time as the call to
1363          * flush_dcache_page() because it is a NULL macro on i386 and hence the
1364          * decrement never happens so the loop never terminates.
1365          */
1366         do {
1367                 --nr_pages;
1368                 flush_dcache_page(pages[nr_pages]);
1369         } while (nr_pages > 0);
1370 }
1371
1372 /**
1373  * ntfs_commit_pages_after_non_resident_write - commit the received data
1374  * @pages:      array of destination pages
1375  * @nr_pages:   number of pages in @pages
1376  * @pos:        byte position in file at which the write begins
1377  * @bytes:      number of bytes to be written
1378  *
1379  * See description of ntfs_commit_pages_after_write(), below.
1380  */
1381 static inline int ntfs_commit_pages_after_non_resident_write(
1382                 struct page **pages, const unsigned nr_pages,
1383                 s64 pos, size_t bytes)
1384 {
1385         s64 end, initialized_size;
1386         struct inode *vi;
1387         ntfs_inode *ni, *base_ni;
1388         struct buffer_head *bh, *head;
1389         ntfs_attr_search_ctx *ctx;
1390         MFT_RECORD *m;
1391         ATTR_RECORD *a;
1392         unsigned long flags;
1393         unsigned blocksize, u;
1394         int err;
1395
1396         vi = pages[0]->mapping->host;
1397         ni = NTFS_I(vi);
1398         blocksize = vi->i_sb->s_blocksize;
1399         end = pos + bytes;
1400         u = 0;
1401         do {
1402                 s64 bh_pos;
1403                 struct page *page;
1404                 bool partial;
1405
1406                 page = pages[u];
1407                 bh_pos = (s64)page->index << PAGE_SHIFT;
1408                 bh = head = page_buffers(page);
1409                 partial = false;
1410                 do {
1411                         s64 bh_end;
1412
1413                         bh_end = bh_pos + blocksize;
1414                         if (bh_end <= pos || bh_pos >= end) {
1415                                 if (!buffer_uptodate(bh))
1416                                         partial = true;
1417                         } else {
1418                                 set_buffer_uptodate(bh);
1419                                 mark_buffer_dirty(bh);
1420                         }
1421                 } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
1422                 /*
1423                  * If all buffers are now uptodate but the page is not, set the
1424                  * page uptodate.
1425                  */
1426                 if (!partial && !PageUptodate(page))
1427                         SetPageUptodate(page);
1428         } while (++u < nr_pages);
1429         /*
1430          * Finally, if we do not need to update initialized_size or i_size we
1431          * are finished.
1432          */
1433         read_lock_irqsave(&ni->size_lock, flags);
1434         initialized_size = ni->initialized_size;
1435         read_unlock_irqrestore(&ni->size_lock, flags);
1436         if (end <= initialized_size) {
1437                 ntfs_debug("Done.");
1438                 return 0;
1439         }
1440         /*
1441          * Update initialized_size/i_size as appropriate, both in the inode and
1442          * the mft record.
1443          */
1444         if (!NInoAttr(ni))
1445                 base_ni = ni;
1446         else
1447                 base_ni = ni->ext.base_ntfs_ino;
1448         /* Map, pin, and lock the mft record. */
1449         m = map_mft_record(base_ni);
1450         if (IS_ERR(m)) {
1451                 err = PTR_ERR(m);
1452                 m = NULL;
1453                 ctx = NULL;
1454                 goto err_out;
1455         }
1456         BUG_ON(!NInoNonResident(ni));
1457         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1458         if (unlikely(!ctx)) {
1459                 err = -ENOMEM;
1460                 goto err_out;
1461         }
1462         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1463                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1464         if (unlikely(err)) {
1465                 if (err == -ENOENT)
1466                         err = -EIO;
1467                 goto err_out;
1468         }
1469         a = ctx->attr;
1470         BUG_ON(!a->non_resident);
1471         write_lock_irqsave(&ni->size_lock, flags);
1472         BUG_ON(end > ni->allocated_size);
1473         ni->initialized_size = end;
1474         a->data.non_resident.initialized_size = cpu_to_sle64(end);
1475         if (end > i_size_read(vi)) {
1476                 i_size_write(vi, end);
1477                 a->data.non_resident.data_size =
1478                                 a->data.non_resident.initialized_size;
1479         }
1480         write_unlock_irqrestore(&ni->size_lock, flags);
1481         /* Mark the mft record dirty, so it gets written back. */
1482         flush_dcache_mft_record_page(ctx->ntfs_ino);
1483         mark_mft_record_dirty(ctx->ntfs_ino);
1484         ntfs_attr_put_search_ctx(ctx);
1485         unmap_mft_record(base_ni);
1486         ntfs_debug("Done.");
1487         return 0;
1488 err_out:
1489         if (ctx)
1490                 ntfs_attr_put_search_ctx(ctx);
1491         if (m)
1492                 unmap_mft_record(base_ni);
1493         ntfs_error(vi->i_sb, "Failed to update initialized_size/i_size (error "
1494                         "code %i).", err);
1495         if (err != -ENOMEM)
1496                 NVolSetErrors(ni->vol);
1497         return err;
1498 }
1499
1500 /**
1501  * ntfs_commit_pages_after_write - commit the received data
1502  * @pages:      array of destination pages
1503  * @nr_pages:   number of pages in @pages
1504  * @pos:        byte position in file at which the write begins
1505  * @bytes:      number of bytes to be written
1506  *
1507  * This is called from ntfs_file_buffered_write() with i_mutex held on the inode
1508  * (@pages[0]->mapping->host).  There are @nr_pages pages in @pages which are
1509  * locked but not kmap()ped.  The source data has already been copied into the
1510  * @page.  ntfs_prepare_pages_for_non_resident_write() has been called before
1511  * the data was copied (for non-resident attributes only) and it returned
1512  * success.
1513  *
1514  * Need to set uptodate and mark dirty all buffers within the boundary of the
1515  * write.  If all buffers in a page are uptodate we set the page uptodate, too.
1516  *
1517  * Setting the buffers dirty ensures that they get written out later when
1518  * ntfs_writepage() is invoked by the VM.
1519  *
1520  * Finally, we need to update i_size and initialized_size as appropriate both
1521  * in the inode and the mft record.
1522  *
1523  * This is modelled after fs/buffer.c::generic_commit_write(), which marks
1524  * buffers uptodate and dirty, sets the page uptodate if all buffers in the
1525  * page are uptodate, and updates i_size if the end of io is beyond i_size.  In
1526  * that case, it also marks the inode dirty.
1527  *
1528  * If things have gone as outlined in
1529  * ntfs_prepare_pages_for_non_resident_write(), we do not need to do any page
1530  * content modifications here for non-resident attributes.  For resident
1531  * attributes we need to do the uptodate bringing here which we combine with
1532  * the copying into the mft record which means we save one atomic kmap.
1533  *
1534  * Return 0 on success or -errno on error.
1535  */
1536 static int ntfs_commit_pages_after_write(struct page **pages,
1537                 const unsigned nr_pages, s64 pos, size_t bytes)
1538 {
1539         s64 end, initialized_size;
1540         loff_t i_size;
1541         struct inode *vi;
1542         ntfs_inode *ni, *base_ni;
1543         struct page *page;
1544         ntfs_attr_search_ctx *ctx;
1545         MFT_RECORD *m;
1546         ATTR_RECORD *a;
1547         char *kattr, *kaddr;
1548         unsigned long flags;
1549         u32 attr_len;
1550         int err;
1551
1552         BUG_ON(!nr_pages);
1553         BUG_ON(!pages);
1554         page = pages[0];
1555         BUG_ON(!page);
1556         vi = page->mapping->host;
1557         ni = NTFS_I(vi);
1558         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, start page "
1559                         "index 0x%lx, nr_pages 0x%x, pos 0x%llx, bytes 0x%zx.",
1560                         vi->i_ino, ni->type, page->index, nr_pages,
1561                         (long long)pos, bytes);
1562         if (NInoNonResident(ni))
1563                 return ntfs_commit_pages_after_non_resident_write(pages,
1564                                 nr_pages, pos, bytes);
1565         BUG_ON(nr_pages > 1);
1566         /*
1567          * Attribute is resident, implying it is not compressed, encrypted, or
1568          * sparse.
1569          */
1570         if (!NInoAttr(ni))
1571                 base_ni = ni;
1572         else
1573                 base_ni = ni->ext.base_ntfs_ino;
1574         BUG_ON(NInoNonResident(ni));
1575         /* Map, pin, and lock the mft record. */
1576         m = map_mft_record(base_ni);
1577         if (IS_ERR(m)) {
1578                 err = PTR_ERR(m);
1579                 m = NULL;
1580                 ctx = NULL;
1581                 goto err_out;
1582         }
1583         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1584         if (unlikely(!ctx)) {
1585                 err = -ENOMEM;
1586                 goto err_out;
1587         }
1588         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1589                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1590         if (unlikely(err)) {
1591                 if (err == -ENOENT)
1592                         err = -EIO;
1593                 goto err_out;
1594         }
1595         a = ctx->attr;
1596         BUG_ON(a->non_resident);
1597         /* The total length of the attribute value. */
1598         attr_len = le32_to_cpu(a->data.resident.value_length);
1599         i_size = i_size_read(vi);
1600         BUG_ON(attr_len != i_size);
1601         BUG_ON(pos > attr_len);
1602         end = pos + bytes;
1603         BUG_ON(end > le32_to_cpu(a->length) -
1604                         le16_to_cpu(a->data.resident.value_offset));
1605         kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
1606         kaddr = kmap_atomic(page);
1607         /* Copy the received data from the page to the mft record. */
1608         memcpy(kattr + pos, kaddr + pos, bytes);
1609         /* Update the attribute length if necessary. */
1610         if (end > attr_len) {
1611                 attr_len = end;
1612                 a->data.resident.value_length = cpu_to_le32(attr_len);
1613         }
1614         /*
1615          * If the page is not uptodate, bring the out of bounds area(s)
1616          * uptodate by copying data from the mft record to the page.
1617          */
1618         if (!PageUptodate(page)) {
1619                 if (pos > 0)
1620                         memcpy(kaddr, kattr, pos);
1621                 if (end < attr_len)
1622                         memcpy(kaddr + end, kattr + end, attr_len - end);
1623                 /* Zero the region outside the end of the attribute value. */
1624                 memset(kaddr + attr_len, 0, PAGE_SIZE - attr_len);
1625                 flush_dcache_page(page);
1626                 SetPageUptodate(page);
1627         }
1628         kunmap_atomic(kaddr);
1629         /* Update initialized_size/i_size if necessary. */
1630         read_lock_irqsave(&ni->size_lock, flags);
1631         initialized_size = ni->initialized_size;
1632         BUG_ON(end > ni->allocated_size);
1633         read_unlock_irqrestore(&ni->size_lock, flags);
1634         BUG_ON(initialized_size != i_size);
1635         if (end > initialized_size) {
1636                 write_lock_irqsave(&ni->size_lock, flags);
1637                 ni->initialized_size = end;
1638                 i_size_write(vi, end);
1639                 write_unlock_irqrestore(&ni->size_lock, flags);
1640         }
1641         /* Mark the mft record dirty, so it gets written back. */
1642         flush_dcache_mft_record_page(ctx->ntfs_ino);
1643         mark_mft_record_dirty(ctx->ntfs_ino);
1644         ntfs_attr_put_search_ctx(ctx);
1645         unmap_mft_record(base_ni);
1646         ntfs_debug("Done.");
1647         return 0;
1648 err_out:
1649         if (err == -ENOMEM) {
1650                 ntfs_warning(vi->i_sb, "Error allocating memory required to "
1651                                 "commit the write.");
1652                 if (PageUptodate(page)) {
1653                         ntfs_warning(vi->i_sb, "Page is uptodate, setting "
1654                                         "dirty so the write will be retried "
1655                                         "later on by the VM.");
1656                         /*
1657                          * Put the page on mapping->dirty_pages, but leave its
1658                          * buffers' dirty state as-is.
1659                          */
1660                         __set_page_dirty_nobuffers(page);
1661                         err = 0;
1662                 } else
1663                         ntfs_error(vi->i_sb, "Page is not uptodate.  Written "
1664                                         "data has been lost.");
1665         } else {
1666                 ntfs_error(vi->i_sb, "Resident attribute commit write failed "
1667                                 "with error %i.", err);
1668                 NVolSetErrors(ni->vol);
1669         }
1670         if (ctx)
1671                 ntfs_attr_put_search_ctx(ctx);
1672         if (m)
1673                 unmap_mft_record(base_ni);
1674         return err;
1675 }
1676
1677 /*
1678  * Copy as much as we can into the pages and return the number of bytes which
1679  * were successfully copied.  If a fault is encountered then clear the pages
1680  * out to (ofs + bytes) and return the number of bytes which were copied.
1681  */
1682 static size_t ntfs_copy_from_user_iter(struct page **pages, unsigned nr_pages,
1683                 unsigned ofs, struct iov_iter *i, size_t bytes)
1684 {
1685         struct page **last_page = pages + nr_pages;
1686         size_t total = 0;
1687         unsigned len, copied;
1688
1689         do {
1690                 len = PAGE_SIZE - ofs;
1691                 if (len > bytes)
1692                         len = bytes;
1693                 copied = copy_page_from_iter_atomic(*pages, ofs, len, i);
1694                 total += copied;
1695                 bytes -= copied;
1696                 if (!bytes)
1697                         break;
1698                 if (copied < len)
1699                         goto err;
1700                 ofs = 0;
1701         } while (++pages < last_page);
1702 out:
1703         return total;
1704 err:
1705         /* Zero the rest of the target like __copy_from_user(). */
1706         len = PAGE_SIZE - copied;
1707         do {
1708                 if (len > bytes)
1709                         len = bytes;
1710                 zero_user(*pages, copied, len);
1711                 bytes -= len;
1712                 copied = 0;
1713                 len = PAGE_SIZE;
1714         } while (++pages < last_page);
1715         goto out;
1716 }
1717
1718 /**
1719  * ntfs_perform_write - perform buffered write to a file
1720  * @file:       file to write to
1721  * @i:          iov_iter with data to write
1722  * @pos:        byte offset in file at which to begin writing to
1723  */
1724 static ssize_t ntfs_perform_write(struct file *file, struct iov_iter *i,
1725                 loff_t pos)
1726 {
1727         struct address_space *mapping = file->f_mapping;
1728         struct inode *vi = mapping->host;
1729         ntfs_inode *ni = NTFS_I(vi);
1730         ntfs_volume *vol = ni->vol;
1731         struct page *pages[NTFS_MAX_PAGES_PER_CLUSTER];
1732         struct page *cached_page = NULL;
1733         VCN last_vcn;
1734         LCN lcn;
1735         size_t bytes;
1736         ssize_t status, written = 0;
1737         unsigned nr_pages;
1738
1739         ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, pos "
1740                         "0x%llx, count 0x%lx.", vi->i_ino,
1741                         (unsigned)le32_to_cpu(ni->type),
1742                         (unsigned long long)pos,
1743                         (unsigned long)iov_iter_count(i));
1744         /*
1745          * If a previous ntfs_truncate() failed, repeat it and abort if it
1746          * fails again.
1747          */
1748         if (unlikely(NInoTruncateFailed(ni))) {
1749                 int err;
1750
1751                 inode_dio_wait(vi);
1752                 err = ntfs_truncate(vi);
1753                 if (err || NInoTruncateFailed(ni)) {
1754                         if (!err)
1755                                 err = -EIO;
1756                         ntfs_error(vol->sb, "Cannot perform write to inode "
1757                                         "0x%lx, attribute type 0x%x, because "
1758                                         "ntfs_truncate() failed (error code "
1759                                         "%i).", vi->i_ino,
1760                                         (unsigned)le32_to_cpu(ni->type), err);
1761                         return err;
1762                 }
1763         }
1764         /*
1765          * Determine the number of pages per cluster for non-resident
1766          * attributes.
1767          */
1768         nr_pages = 1;
1769         if (vol->cluster_size > PAGE_SIZE && NInoNonResident(ni))
1770                 nr_pages = vol->cluster_size >> PAGE_SHIFT;
1771         last_vcn = -1;
1772         do {
1773                 VCN vcn;
1774                 pgoff_t idx, start_idx;
1775                 unsigned ofs, do_pages, u;
1776                 size_t copied;
1777
1778                 start_idx = idx = pos >> PAGE_SHIFT;
1779                 ofs = pos & ~PAGE_MASK;
1780                 bytes = PAGE_SIZE - ofs;
1781                 do_pages = 1;
1782                 if (nr_pages > 1) {
1783                         vcn = pos >> vol->cluster_size_bits;
1784                         if (vcn != last_vcn) {
1785                                 last_vcn = vcn;
1786                                 /*
1787                                  * Get the lcn of the vcn the write is in.  If
1788                                  * it is a hole, need to lock down all pages in
1789                                  * the cluster.
1790                                  */
1791                                 down_read(&ni->runlist.lock);
1792                                 lcn = ntfs_attr_vcn_to_lcn_nolock(ni, pos >>
1793                                                 vol->cluster_size_bits, false);
1794                                 up_read(&ni->runlist.lock);
1795                                 if (unlikely(lcn < LCN_HOLE)) {
1796                                         if (lcn == LCN_ENOMEM)
1797                                                 status = -ENOMEM;
1798                                         else {
1799                                                 status = -EIO;
1800                                                 ntfs_error(vol->sb, "Cannot "
1801                                                         "perform write to "
1802                                                         "inode 0x%lx, "
1803                                                         "attribute type 0x%x, "
1804                                                         "because the attribute "
1805                                                         "is corrupt.",
1806                                                         vi->i_ino, (unsigned)
1807                                                         le32_to_cpu(ni->type));
1808                                         }
1809                                         break;
1810                                 }
1811                                 if (lcn == LCN_HOLE) {
1812                                         start_idx = (pos & ~(s64)
1813                                                         vol->cluster_size_mask)
1814                                                         >> PAGE_SHIFT;
1815                                         bytes = vol->cluster_size - (pos &
1816                                                         vol->cluster_size_mask);
1817                                         do_pages = nr_pages;
1818                                 }
1819                         }
1820                 }
1821                 if (bytes > iov_iter_count(i))
1822                         bytes = iov_iter_count(i);
1823 again:
1824                 /*
1825                  * Bring in the user page(s) that we will copy from _first_.
1826                  * Otherwise there is a nasty deadlock on copying from the same
1827                  * page(s) as we are writing to, without it/them being marked
1828                  * up-to-date.  Note, at present there is nothing to stop the
1829                  * pages being swapped out between us bringing them into memory
1830                  * and doing the actual copying.
1831                  */
1832                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
1833                         status = -EFAULT;
1834                         break;
1835                 }
1836                 /* Get and lock @do_pages starting at index @start_idx. */
1837                 status = __ntfs_grab_cache_pages(mapping, start_idx, do_pages,
1838                                 pages, &cached_page);
1839                 if (unlikely(status))
1840                         break;
1841                 /*
1842                  * For non-resident attributes, we need to fill any holes with
1843                  * actual clusters and ensure all bufferes are mapped.  We also
1844                  * need to bring uptodate any buffers that are only partially
1845                  * being written to.
1846                  */
1847                 if (NInoNonResident(ni)) {
1848                         status = ntfs_prepare_pages_for_non_resident_write(
1849                                         pages, do_pages, pos, bytes);
1850                         if (unlikely(status)) {
1851                                 do {
1852                                         unlock_page(pages[--do_pages]);
1853                                         put_page(pages[do_pages]);
1854                                 } while (do_pages);
1855                                 break;
1856                         }
1857                 }
1858                 u = (pos >> PAGE_SHIFT) - pages[0]->index;
1859                 copied = ntfs_copy_from_user_iter(pages + u, do_pages - u, ofs,
1860                                         i, bytes);
1861                 ntfs_flush_dcache_pages(pages + u, do_pages - u);
1862                 status = 0;
1863                 if (likely(copied == bytes)) {
1864                         status = ntfs_commit_pages_after_write(pages, do_pages,
1865                                         pos, bytes);
1866                 }
1867                 do {
1868                         unlock_page(pages[--do_pages]);
1869                         put_page(pages[do_pages]);
1870                 } while (do_pages);
1871                 if (unlikely(status < 0)) {
1872                         iov_iter_revert(i, copied);
1873                         break;
1874                 }
1875                 cond_resched();
1876                 if (unlikely(copied < bytes)) {
1877                         iov_iter_revert(i, copied);
1878                         if (copied)
1879                                 bytes = copied;
1880                         else if (bytes > PAGE_SIZE - ofs)
1881                                 bytes = PAGE_SIZE - ofs;
1882                         goto again;
1883                 }
1884                 pos += copied;
1885                 written += copied;
1886                 balance_dirty_pages_ratelimited(mapping);
1887                 if (fatal_signal_pending(current)) {
1888                         status = -EINTR;
1889                         break;
1890                 }
1891         } while (iov_iter_count(i));
1892         if (cached_page)
1893                 put_page(cached_page);
1894         ntfs_debug("Done.  Returning %s (written 0x%lx, status %li).",
1895                         written ? "written" : "status", (unsigned long)written,
1896                         (long)status);
1897         return written ? written : status;
1898 }
1899
1900 /**
1901  * ntfs_file_write_iter - simple wrapper for ntfs_file_write_iter_nolock()
1902  * @iocb:       IO state structure
1903  * @from:       iov_iter with data to write
1904  *
1905  * Basically the same as generic_file_write_iter() except that it ends up
1906  * up calling ntfs_perform_write() instead of generic_perform_write() and that
1907  * O_DIRECT is not implemented.
1908  */
1909 static ssize_t ntfs_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1910 {
1911         struct file *file = iocb->ki_filp;
1912         struct inode *vi = file_inode(file);
1913         ssize_t written = 0;
1914         ssize_t err;
1915
1916         inode_lock(vi);
1917         /* We can write back this queue in page reclaim. */
1918         current->backing_dev_info = inode_to_bdi(vi);
1919         err = ntfs_prepare_file_for_write(iocb, from);
1920         if (iov_iter_count(from) && !err)
1921                 written = ntfs_perform_write(file, from, iocb->ki_pos);
1922         current->backing_dev_info = NULL;
1923         inode_unlock(vi);
1924         iocb->ki_pos += written;
1925         if (likely(written > 0))
1926                 written = generic_write_sync(iocb, written);
1927         return written ? written : err;
1928 }
1929
1930 /**
1931  * ntfs_file_fsync - sync a file to disk
1932  * @filp:       file to be synced
1933  * @datasync:   if non-zero only flush user data and not metadata
1934  *
1935  * Data integrity sync of a file to disk.  Used for fsync, fdatasync, and msync
1936  * system calls.  This function is inspired by fs/buffer.c::file_fsync().
1937  *
1938  * If @datasync is false, write the mft record and all associated extent mft
1939  * records as well as the $DATA attribute and then sync the block device.
1940  *
1941  * If @datasync is true and the attribute is non-resident, we skip the writing
1942  * of the mft record and all associated extent mft records (this might still
1943  * happen due to the write_inode_now() call).
1944  *
1945  * Also, if @datasync is true, we do not wait on the inode to be written out
1946  * but we always wait on the page cache pages to be written out.
1947  *
1948  * Locking: Caller must hold i_mutex on the inode.
1949  *
1950  * TODO: We should probably also write all attribute/index inodes associated
1951  * with this inode but since we have no simple way of getting to them we ignore
1952  * this problem for now.
1953  */
1954 static int ntfs_file_fsync(struct file *filp, loff_t start, loff_t end,
1955                            int datasync)
1956 {
1957         struct inode *vi = filp->f_mapping->host;
1958         int err, ret = 0;
1959
1960         ntfs_debug("Entering for inode 0x%lx.", vi->i_ino);
1961
1962         err = file_write_and_wait_range(filp, start, end);
1963         if (err)
1964                 return err;
1965         inode_lock(vi);
1966
1967         BUG_ON(S_ISDIR(vi->i_mode));
1968         if (!datasync || !NInoNonResident(NTFS_I(vi)))
1969                 ret = __ntfs_write_inode(vi, 1);
1970         write_inode_now(vi, !datasync);
1971         /*
1972          * NOTE: If we were to use mapping->private_list (see ext2 and
1973          * fs/buffer.c) for dirty blocks then we could optimize the below to be
1974          * sync_mapping_buffers(vi->i_mapping).
1975          */
1976         err = sync_blockdev(vi->i_sb->s_bdev);
1977         if (unlikely(err && !ret))
1978                 ret = err;
1979         if (likely(!ret))
1980                 ntfs_debug("Done.");
1981         else
1982                 ntfs_warning(vi->i_sb, "Failed to f%ssync inode 0x%lx.  Error "
1983                                 "%u.", datasync ? "data" : "", vi->i_ino, -ret);
1984         inode_unlock(vi);
1985         return ret;
1986 }
1987
1988 #endif /* NTFS_RW */
1989
1990 const struct file_operations ntfs_file_ops = {
1991         .llseek         = generic_file_llseek,
1992         .read_iter      = generic_file_read_iter,
1993 #ifdef NTFS_RW
1994         .write_iter     = ntfs_file_write_iter,
1995         .fsync          = ntfs_file_fsync,
1996 #endif /* NTFS_RW */
1997         .mmap           = generic_file_mmap,
1998         .open           = ntfs_file_open,
1999         .splice_read    = generic_file_splice_read,
2000 };
2001
2002 const struct inode_operations ntfs_file_inode_ops = {
2003 #ifdef NTFS_RW
2004         .setattr        = ntfs_setattr,
2005 #endif /* NTFS_RW */
2006 };
2007
2008 const struct file_operations ntfs_empty_file_ops = {};
2009
2010 const struct inode_operations ntfs_empty_inode_ops = {};