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/blkdev.h>
   9 #include <linux/backing-dev.h>
  10 #include <linux/buffer_head.h>
  11 #include <linux/gfp.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/pagevec.h>
  14 #include <linux/sched/signal.h>
  15 #include <linux/swap.h>
  16 #include <linux/uio.h>
  17 #include <linux/writeback.h>
  18
  19 #include <asm/page.h>
  20 #include <linux/uaccess.h>
  21
  22 #include "attrib.h"
  23 #include "bitmap.h"
  24 #include "inode.h"
  25 #include "debug.h"
  26 #include "lcnalloc.h"
  27 #include "malloc.h"
  28 #include "mft.h"
  29 #include "ntfs.h"
  30
  31 /**
  32  * ntfs_file_open - called when an inode is about to be opened
  33  * @vi:         inode to be opened
  34  * @filp:       file structure describing the inode
  35  *
  36  * Limit file size to the page cache limit on architectures where unsigned long
  37  * is 32-bits. This is the most we can do for now without overflowing the page
  38  * cache page index. Doing it this way means we don't run into problems because
  39  * of existing too large files. It would be better to allow the user to read
  40  * the beginning of the file but I doubt very much anyone is going to hit this
  41  * check on a 32-bit architecture, so there is no point in adding the extra
  42  * complexity required to support this.
  43  *
  44  * On 64-bit architectures, the check is hopefully optimized away by the
  45  * compiler.
  46  *
  47  * After the check passes, just call generic_file_open() to do its work.
  48  */
  49 static int ntfs_file_open(struct inode *vi, struct file *filp)
  50 {
  51         if (sizeof(unsigned long) < 8) {
  52                 if (i_size_read(vi) > MAX_LFS_FILESIZE)
  53                         return -EOVERFLOW;
  54         }
  55         return generic_file_open(vi, filp);
  56 }
  57
  58 #ifdef NTFS_RW
  59
  60 /**
  61  * ntfs_attr_extend_initialized - extend the initialized size of an attribute
  62  * @ni:                 ntfs inode of the attribute to extend
  63  * @new_init_size:      requested new initialized size in bytes
  64  *
  65  * Extend the initialized size of an attribute described by the ntfs inode @ni
  66  * to @new_init_size bytes.  This involves zeroing any non-sparse space between
  67  * the old initialized size and @new_init_size both in the page cache and on
  68  * disk (if relevant complete pages are already uptodate in the page cache then
  69  * these are simply marked dirty).
  70  *
  71  * As a side-effect, the file size (vfs inode->i_size) may be incremented as,
  72  * in the resident attribute case, it is tied to the initialized size and, in
  73  * the non-resident attribute case, it may not fall below the initialized size.
  74  *
  75  * Note that if the attribute is resident, we do not need to touch the page
  76  * cache at all.  This is because if the page cache page is not uptodate we
  77  * bring it uptodate later, when doing the write to the mft record since we
  78  * then already have the page mapped.  And if the page is uptodate, the
  79  * non-initialized region will already have been zeroed when the page was
  80  * brought uptodate and the region may in fact already have been overwritten
  81  * with new data via mmap() based writes, so we cannot just zero it.  And since
  82  * POSIX specifies that the behaviour of resizing a file whilst it is mmap()ped
  83  * is unspecified, we choose not to do zeroing and thus we do not need to touch
  84  * the page at all.  For a more detailed explanation see ntfs_truncate() in
  85  * fs/ntfs/inode.c.
  86  *
  87  * Return 0 on success and -errno on error.  In the case that an error is
  88  * encountered it is possible that the initialized size will already have been
  89  * incremented some way towards @new_init_size but it is guaranteed that if
  90  * this is the case, the necessary zeroing will also have happened and that all
  91  * metadata is self-consistent.
  92  *
  93  * Locking: i_mutex on the vfs inode corrseponsind to the ntfs inode @ni must be
  94  *          held by the caller.
  95  */
  96 static int ntfs_attr_extend_initialized(ntfs_inode *ni, const s64 new_init_size)
  97 {
  98         s64 old_init_size;
  99         loff_t old_i_size;
 100         pgoff_t index, end_index;
 101         unsigned long flags;
 102         struct inode *vi = VFS_I(ni);
 103         ntfs_inode *base_ni;
 104         MFT_RECORD *m = NULL;
 105         ATTR_RECORD *a;
 106         ntfs_attr_search_ctx *ctx = NULL;
 107         struct address_space *mapping;
 108         struct page *page = NULL;
 109         u8 *kattr;
 110         int err;
 111         u32 attr_len;
 112
 113         read_lock_irqsave(&ni->size_lock, flags);
 114         old_init_size = ni->initialized_size;
 115         old_i_size = i_size_read(vi);
 116         BUG_ON(new_init_size > ni->allocated_size);
 117         read_unlock_irqrestore(&ni->size_lock, flags);
 118         ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, "
 119                         "old_initialized_size 0x%llx, "
 120                         "new_initialized_size 0x%llx, i_size 0x%llx.",
 121                         vi->i_ino, (unsigned)le32_to_cpu(ni->type),
 122                         (unsigned long long)old_init_size,
 123                         (unsigned long long)new_init_size, old_i_size);
 124         if (!NInoAttr(ni))
 125                 base_ni = ni;
 126         else
 127                 base_ni = ni->ext.base_ntfs_ino;
 128         /* Use goto to reduce indentation and we need the label below anyway. */
 129         if (NInoNonResident(ni))
 130                 goto do_non_resident_extend;
 131         BUG_ON(old_init_size != old_i_size);
 132         m = map_mft_record(base_ni);
 133         if (IS_ERR(m)) {
 134                 err = PTR_ERR(m);
 135                 m = NULL;
 136                 goto err_out;
 137         }
 138         ctx = ntfs_attr_get_search_ctx(base_ni, m);
 139         if (unlikely(!ctx)) {
 140                 err = -ENOMEM;
 141                 goto err_out;
 142         }
 143         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
 144                         CASE_SENSITIVE, 0, NULL, 0, ctx);
 145         if (unlikely(err)) {
 146                 if (err == -ENOENT)
 147                         err = -EIO;
 148                 goto err_out;
 149         }
 150         m = ctx->mrec;
 151         a = ctx->attr;
 152         BUG_ON(a->non_resident);
 153         /* The total length of the attribute value. */
 154         attr_len = le32_to_cpu(a->data.resident.value_length);
 155         BUG_ON(old_i_size != (loff_t)attr_len);
 156         /*
 157          * Do the zeroing in the mft record and update the attribute size in
 158          * the mft record.
 159          */
 160         kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
 161         memset(kattr + attr_len, 0, new_init_size - attr_len);
 162         a->data.resident.value_length = cpu_to_le32((u32)new_init_size);
 163         /* Finally, update the sizes in the vfs and ntfs inodes. */
 164         write_lock_irqsave(&ni->size_lock, flags);
 165         i_size_write(vi, new_init_size);
 166         ni->initialized_size = new_init_size;
 167         write_unlock_irqrestore(&ni->size_lock, flags);
 168         goto done;
 169 do_non_resident_extend:
 170         /*
 171          * If the new initialized size @new_init_size exceeds the current file
 172          * size (vfs inode->i_size), we need to extend the file size to the
 173          * new initialized size.
 174          */
 175         if (new_init_size > old_i_size) {
 176                 m = map_mft_record(base_ni);
 177                 if (IS_ERR(m)) {
 178                         err = PTR_ERR(m);
 179                         m = NULL;
 180                         goto err_out;
 181                 }
 182                 ctx = ntfs_attr_get_search_ctx(base_ni, m);
 183                 if (unlikely(!ctx)) {
 184                         err = -ENOMEM;
 185                         goto err_out;
 186                 }
 187                 err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
 188                                 CASE_SENSITIVE, 0, NULL, 0, ctx);
 189                 if (unlikely(err)) {
 190                         if (err == -ENOENT)
 191                                 err = -EIO;
 192                         goto err_out;
 193                 }
 194                 m = ctx->mrec;
 195                 a = ctx->attr;
 196                 BUG_ON(!a->non_resident);
 197                 BUG_ON(old_i_size != (loff_t)
 198                                 sle64_to_cpu(a->data.non_resident.data_size));
 199                 a->data.non_resident.data_size = cpu_to_sle64(new_init_size);
 200                 flush_dcache_mft_record_page(ctx->ntfs_ino);
 201                 mark_mft_record_dirty(ctx->ntfs_ino);
 202                 /* Update the file size in the vfs inode. */
 203                 i_size_write(vi, new_init_size);
 204                 ntfs_attr_put_search_ctx(ctx);
 205                 ctx = NULL;
 206                 unmap_mft_record(base_ni);
 207                 m = NULL;
 208         }
 209         mapping = vi->i_mapping;
 210         index = old_init_size >> PAGE_SHIFT;
 211         end_index = (new_init_size + PAGE_SIZE - 1) >> PAGE_SHIFT;
 212         do {
 213                 /*
 214                  * Read the page.  If the page is not present, this will zero
 215                  * the uninitialized regions for us.
 216                  */
 217                 page = read_mapping_page(mapping, index, NULL);
 218                 if (IS_ERR(page)) {
 219                         err = PTR_ERR(page);
 220                         goto init_err_out;
 221                 }
 222                 if (unlikely(PageError(page))) {
 223                         put_page(page);
 224                         err = -EIO;
 225                         goto init_err_out;
 226                 }
 227                 /*
 228                  * Update the initialized size in the ntfs inode.  This is
 229                  * enough to make ntfs_writepage() work.
 230                  */
 231                 write_lock_irqsave(&ni->size_lock, flags);
 232                 ni->initialized_size = (s64)(index + 1) << PAGE_SHIFT;
 233                 if (ni->initialized_size > new_init_size)
 234                         ni->initialized_size = new_init_size;
 235                 write_unlock_irqrestore(&ni->size_lock, flags);
 236                 /* Set the page dirty so it gets written out. */
 237                 set_page_dirty(page);
 238                 put_page(page);
 239                 /*
 240                  * Play nice with the vm and the rest of the system.  This is
 241                  * very much needed as we can potentially be modifying the
 242                  * initialised size from a very small value to a really huge
 243                  * value, e.g.
 244                  *      f = open(somefile, O_TRUNC);
 245                  *      truncate(f, 10GiB);
 246                  *      seek(f, 10GiB);
 247                  *      write(f, 1);
 248                  * And this would mean we would be marking dirty hundreds of
 249                  * thousands of pages or as in the above example more than
 250                  * two and a half million pages!
 251                  *
 252                  * TODO: For sparse pages could optimize this workload by using
 253                  * the FsMisc / MiscFs page bit as a "PageIsSparse" bit.  This
 254                  * would be set in readpage for sparse pages and here we would
 255                  * not need to mark dirty any pages which have this bit set.
 256                  * The only caveat is that we have to clear the bit everywhere
 257                  * where we allocate any clusters that lie in the page or that
 258                  * contain the page.
 259                  *
 260                  * TODO: An even greater optimization would be for us to only
 261                  * call readpage() on pages which are not in sparse regions as
 262                  * determined from the runlist.  This would greatly reduce the
 263                  * number of pages we read and make dirty in the case of sparse
 264                  * files.
 265                  */
 266                 balance_dirty_pages_ratelimited(mapping);
 267                 cond_resched();
 268         } while (++index < end_index);
 269         read_lock_irqsave(&ni->size_lock, flags);
 270         BUG_ON(ni->initialized_size != new_init_size);
 271         read_unlock_irqrestore(&ni->size_lock, flags);
 272         /* Now bring in sync the initialized_size in the mft record. */
 273         m = map_mft_record(base_ni);
 274         if (IS_ERR(m)) {
 275                 err = PTR_ERR(m);
 276                 m = NULL;
 277                 goto init_err_out;
 278         }
 279         ctx = ntfs_attr_get_search_ctx(base_ni, m);
 280         if (unlikely(!ctx)) {
 281                 err = -ENOMEM;
 282                 goto init_err_out;
 283         }
 284         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
 285                         CASE_SENSITIVE, 0, NULL, 0, ctx);
 286         if (unlikely(err)) {
 287                 if (err == -ENOENT)
 288                         err = -EIO;
 289                 goto init_err_out;
 290         }
 291         m = ctx->mrec;
 292         a = ctx->attr;
 293         BUG_ON(!a->non_resident);
 294         a->data.non_resident.initialized_size = cpu_to_sle64(new_init_size);
 295 done:
 296         flush_dcache_mft_record_page(ctx->ntfs_ino);
 297         mark_mft_record_dirty(ctx->ntfs_ino);
 298         if (ctx)
 299                 ntfs_attr_put_search_ctx(ctx);
 300         if (m)
 301                 unmap_mft_record(base_ni);
 302         ntfs_debug("Done, initialized_size 0x%llx, i_size 0x%llx.",
 303                         (unsigned long long)new_init_size, i_size_read(vi));
 304         return 0;
 305 init_err_out:
 306         write_lock_irqsave(&ni->size_lock, flags);
 307         ni->initialized_size = old_init_size;
 308         write_unlock_irqrestore(&ni->size_lock, flags);
 309 err_out:
 310         if (ctx)
 311                 ntfs_attr_put_search_ctx(ctx);
 312         if (m)
 313                 unmap_mft_record(base_ni);
 314         ntfs_debug("Failed.  Returning error code %i.", err);
 315         return err;
 316 }
 317
 318 static ssize_t ntfs_prepare_file_for_write(struct kiocb *iocb,
 319                 struct iov_iter *from)
 320 {
 321         loff_t pos;
 322         s64 end, ll;
 323         ssize_t err;
 324         unsigned long flags;
 325         struct file *file = iocb->ki_filp;
 326         struct inode *vi = file_inode(file);
 327         ntfs_inode *ni = NTFS_I(vi);
 328         ntfs_volume *vol = ni->vol;
 329
 330         ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, pos "
 331                         "0x%llx, count 0x%zx.", vi->i_ino,
 332                         (unsigned)le32_to_cpu(ni->type),
 333                         (unsigned long long)iocb->ki_pos,
 334                         iov_iter_count(from));
 335         err = generic_write_checks(iocb, from);
 336         if (unlikely(err <= 0))
 337                 goto out;
 338         /*
 339          * All checks have passed.  Before we start doing any writing we want
 340          * to abort any totally illegal writes.
 341          */
 342         BUG_ON(NInoMstProtected(ni));
 343         BUG_ON(ni->type != AT_DATA);
 344         /* If file is encrypted, deny access, just like NT4. */
 345         if (NInoEncrypted(ni)) {
 346                 /* Only $DATA attributes can be encrypted. */
 347                 /*
 348                  * Reminder for later: Encrypted files are _always_
 349                  * non-resident so that the content can always be encrypted.
 350                  */
 351                 ntfs_debug("Denying write access to encrypted file.");
 352                 err = -EACCES;
 353                 goto out;
 354         }
 355         if (NInoCompressed(ni)) {
 356                 /* Only unnamed $DATA attribute can be compressed. */
 357                 BUG_ON(ni->name_len);
 358                 /*
 359                  * Reminder for later: If resident, the data is not actually
 360                  * compressed.  Only on the switch to non-resident does
 361                  * compression kick in.  This is in contrast to encrypted files
 362                  * (see above).
 363                  */
 364                 ntfs_error(vi->i_sb, "Writing to compressed files is not "
 365                                 "implemented yet.  Sorry.");
 366                 err = -EOPNOTSUPP;
 367                 goto out;
 368         }
 369         err = file_remove_privs(file);
 370         if (unlikely(err))
 371                 goto out;
 372         /*
 373          * Our ->update_time method always succeeds thus file_update_time()
 374          * cannot fail either so there is no need to check the return code.
 375          */
 376         file_update_time(file);
 377         pos = iocb->ki_pos;
 378         /* The first byte after the last cluster being written to. */
 379         end = (pos + iov_iter_count(from) + vol->cluster_size_mask) &
 380                         ~(u64)vol->cluster_size_mask;
 381         /*
 382          * If the write goes beyond the allocated size, extend the allocation
 383          * to cover the whole of the write, rounded up to the nearest cluster.
 384          */
 385         read_lock_irqsave(&ni->size_lock, flags);
 386         ll = ni->allocated_size;
 387         read_unlock_irqrestore(&ni->size_lock, flags);
 388         if (end > ll) {
 389                 /*
 390                  * Extend the allocation without changing the data size.
 391                  *
 392                  * Note we ensure the allocation is big enough to at least
 393                  * write some data but we do not require the allocation to be
 394                  * complete, i.e. it may be partial.
 395                  */
 396                 ll = ntfs_attr_extend_allocation(ni, end, -1, pos);
 397                 if (likely(ll >= 0)) {
 398                         BUG_ON(pos >= ll);
 399                         /* If the extension was partial truncate the write. */
 400                         if (end > ll) {
 401                                 ntfs_debug("Truncating write to inode 0x%lx, "
 402                                                 "attribute type 0x%x, because "
 403                                                 "the allocation was only "
 404                                                 "partially extended.",
 405                                                 vi->i_ino, (unsigned)
 406                                                 le32_to_cpu(ni->type));
 407                                 iov_iter_truncate(from, ll - pos);
 408                         }
 409                 } else {
 410                         err = ll;
 411                         read_lock_irqsave(&ni->size_lock, flags);
 412                         ll = ni->allocated_size;
 413                         read_unlock_irqrestore(&ni->size_lock, flags);
 414                         /* Perform a partial write if possible or fail. */
 415                         if (pos < ll) {
 416                                 ntfs_debug("Truncating write to inode 0x%lx "
 417                                                 "attribute type 0x%x, because "
 418                                                 "extending the allocation "
 419                                                 "failed (error %d).",
 420                                                 vi->i_ino, (unsigned)
 421                                                 le32_to_cpu(ni->type),
 422                                                 (int)-err);
 423                                 iov_iter_truncate(from, ll - pos);
 424                         } else {
 425                                 if (err != -ENOSPC)
 426                                         ntfs_error(vi->i_sb, "Cannot perform "
 427                                                         "write to inode "
 428                                                         "0x%lx, attribute "
 429                                                         "type 0x%x, because "
 430                                                         "extending the "
 431                                                         "allocation failed "
 432                                                         "(error %ld).",
 433                                                         vi->i_ino, (unsigned)
 434                                                         le32_to_cpu(ni->type),
 435                                                         (long)-err);
 436                                 else
 437                                         ntfs_debug("Cannot perform write to "
 438                                                         "inode 0x%lx, "
 439                                                         "attribute type 0x%x, "
 440                                                         "because there is not "
 441                                                         "space left.",
 442                                                         vi->i_ino, (unsigned)
 443                                                         le32_to_cpu(ni->type));
 444                                 goto out;
 445                         }
 446                 }
 447         }
 448         /*
 449          * If the write starts beyond the initialized size, extend it up to the
 450          * beginning of the write and initialize all non-sparse space between
 451          * the old initialized size and the new one.  This automatically also
 452          * increments the vfs inode->i_size to keep it above or equal to the
 453          * initialized_size.
 454          */
 455         read_lock_irqsave(&ni->size_lock, flags);
 456         ll = ni->initialized_size;
 457         read_unlock_irqrestore(&ni->size_lock, flags);
 458         if (pos > ll) {
 459                 /*
 460                  * Wait for ongoing direct i/o to complete before proceeding.
 461                  * New direct i/o cannot start as we hold i_mutex.
 462                  */
 463                 inode_dio_wait(vi);
 464                 err = ntfs_attr_extend_initialized(ni, pos);
 465                 if (unlikely(err < 0))
 466                         ntfs_error(vi->i_sb, "Cannot perform write to inode "
 467                                         "0x%lx, attribute type 0x%x, because "
 468                                         "extending the initialized size "
 469                                         "failed (error %d).", vi->i_ino,
 470                                         (unsigned)le32_to_cpu(ni->type),
 471                                         (int)-err);
 472         }
 473 out:
 474         return err;
 475 }
 476
 477 /**
 478  * __ntfs_grab_cache_pages - obtain a number of locked pages
 479  * @mapping:    address space mapping from which to obtain page cache pages
 480  * @index:      starting index in @mapping at which to begin obtaining pages
 481  * @nr_pages:   number of page cache pages to obtain
 482  * @pages:      array of pages in which to return the obtained page cache pages
 483  * @cached_page: allocated but as yet unused page
 484  *
 485  * Obtain @nr_pages locked page cache pages from the mapping @mapping and
 486  * starting at index @index.
 487  *
 488  * If a page is newly created, add it to lru list
 489  *
 490  * Note, the page locks are obtained in ascending page index order.
 491  */
 492 static inline int __ntfs_grab_cache_pages(struct address_space *mapping,
 493                 pgoff_t index, const unsigned nr_pages, struct page **pages,
 494                 struct page **cached_page)
 495 {
 496         int err, nr;
 497
 498         BUG_ON(!nr_pages);
 499         err = nr = 0;
 500         do {
 501                 pages[nr] = find_get_page_flags(mapping, index, FGP_LOCK |
 502                                 FGP_ACCESSED);
 503                 if (!pages[nr]) {
 504                         if (!*cached_page) {
 505                                 *cached_page = page_cache_alloc(mapping);
 506                                 if (unlikely(!*cached_page)) {
 507                                         err = -ENOMEM;
 508                                         goto err_out;
 509                                 }
 510                         }
 511                         err = add_to_page_cache_lru(*cached_page, mapping,
 512                                    index,
 513                                    mapping_gfp_constraint(mapping, GFP_KERNEL));
 514                         if (unlikely(err)) {
 515                                 if (err == -EEXIST)
 516                                         continue;
 517                                 goto err_out;
 518                         }
 519                         pages[nr] = *cached_page;
 520                         *cached_page = NULL;
 521                 }
 522                 index++;
 523                 nr++;
 524         } while (nr < nr_pages);
 525 out:
 526         return err;
 527 err_out:
 528         while (nr > 0) {
 529                 unlock_page(pages[--nr]);
 530                 put_page(pages[nr]);
 531         }
 532         goto out;
 533 }
 534
 535 static inline int ntfs_submit_bh_for_read(struct buffer_head *bh)
 536 {
 537         lock_buffer(bh);
 538         get_bh(bh);
 539         bh->b_end_io = end_buffer_read_sync;
 540         return submit_bh(REQ_OP_READ, 0, bh);
 541 }
 542
 543 /**
 544  * ntfs_prepare_pages_for_non_resident_write - prepare pages for receiving data
 545  * @pages:      array of destination pages
 546  * @nr_pages:   number of pages in @pages
 547  * @pos:        byte position in file at which the write begins
 548  * @bytes:      number of bytes to be written
 549  *
 550  * This is called for non-resident attributes from ntfs_file_buffered_write()
 551  * with i_mutex held on the inode (@pages[0]->mapping->host).  There are
 552  * @nr_pages pages in @pages which are locked but not kmap()ped.  The source
 553  * data has not yet been copied into the @pages.
 554  *
 555  * Need to fill any holes with actual clusters, allocate buffers if necessary,
 556  * ensure all the buffers are mapped, and bring uptodate any buffers that are
 557  * only partially being written to.
 558  *
 559  * If @nr_pages is greater than one, we are guaranteed that the cluster size is
 560  * greater than PAGE_SIZE, that all pages in @pages are entirely inside
 561  * the same cluster and that they are the entirety of that cluster, and that
 562  * the cluster is sparse, i.e. we need to allocate a cluster to fill the hole.
 563  *
 564  * i_size is not to be modified yet.
 565  *
 566  * Return 0 on success or -errno on error.
 567  */
 568 static int ntfs_prepare_pages_for_non_resident_write(struct page **pages,
 569                 unsigned nr_pages, s64 pos, size_t bytes)
 570 {
 571         VCN vcn, highest_vcn = 0, cpos, cend, bh_cpos, bh_cend;
 572         LCN lcn;
 573         s64 bh_pos, vcn_len, end, initialized_size;
 574         sector_t lcn_block;
 575         struct page *page;
 576         struct inode *vi;
 577         ntfs_inode *ni, *base_ni = NULL;
 578         ntfs_volume *vol;
 579         runlist_element *rl, *rl2;
 580         struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
 581         ntfs_attr_search_ctx *ctx = NULL;
 582         MFT_RECORD *m = NULL;
 583         ATTR_RECORD *a = NULL;
 584         unsigned long flags;
 585         u32 attr_rec_len = 0;
 586         unsigned blocksize, u;
 587         int err, mp_size;
 588         bool rl_write_locked, was_hole, is_retry;
 589         unsigned char blocksize_bits;
 590         struct {
 591                 u8 runlist_merged:1;
 592                 u8 mft_attr_mapped:1;
 593                 u8 mp_rebuilt:1;
 594                 u8 attr_switched:1;
 595         } status = { 0, 0, 0, 0 };
 596
 597         BUG_ON(!nr_pages);
 598         BUG_ON(!pages);
 599         BUG_ON(!*pages);
 600         vi = pages[0]->mapping->host;
 601         ni = NTFS_I(vi);
 602         vol = ni->vol;
 603         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, start page "
 604                         "index 0x%lx, nr_pages 0x%x, pos 0x%llx, bytes 0x%zx.",
 605                         vi->i_ino, ni->type, pages[0]->index, nr_pages,
 606                         (long long)pos, bytes);
 607         blocksize = vol->sb->s_blocksize;
 608         blocksize_bits = vol->sb->s_blocksize_bits;
 609         u = 0;
 610         do {
 611                 page = pages[u];
 612                 BUG_ON(!page);
 613                 /*
 614                  * create_empty_buffers() will create uptodate/dirty buffers if
 615                  * the page is uptodate/dirty.
 616                  */
 617                 if (!page_has_buffers(page)) {
 618                         create_empty_buffers(page, blocksize, 0);
 619                         if (unlikely(!page_has_buffers(page)))
 620                                 return -ENOMEM;
 621                 }
 622         } while (++u < nr_pages);
 623         rl_write_locked = false;
 624         rl = NULL;
 625         err = 0;
 626         vcn = lcn = -1;
 627         vcn_len = 0;
 628         lcn_block = -1;
 629         was_hole = false;
 630         cpos = pos >> vol->cluster_size_bits;
 631         end = pos + bytes;
 632         cend = (end + vol->cluster_size - 1) >> vol->cluster_size_bits;
 633         /*
 634          * Loop over each page and for each page over each buffer.  Use goto to
 635          * reduce indentation.
 636          */
 637         u = 0;
 638 do_next_page:
 639         page = pages[u];
 640         bh_pos = (s64)page->index << PAGE_SHIFT;
 641         bh = head = page_buffers(page);
 642         do {
 643                 VCN cdelta;
 644                 s64 bh_end;
 645                 unsigned bh_cofs;
 646
 647                 /* Clear buffer_new on all buffers to reinitialise state. */
 648                 if (buffer_new(bh))
 649                         clear_buffer_new(bh);
 650                 bh_end = bh_pos + blocksize;
 651                 bh_cpos = bh_pos >> vol->cluster_size_bits;
 652                 bh_cofs = bh_pos & vol->cluster_size_mask;
 653                 if (buffer_mapped(bh)) {
 654                         /*
 655                          * The buffer is already mapped.  If it is uptodate,
 656                          * ignore it.
 657                          */
 658                         if (buffer_uptodate(bh))
 659                                 continue;
 660                         /*
 661                          * The buffer is not uptodate.  If the page is uptodate
 662                          * set the buffer uptodate and otherwise ignore it.
 663                          */
 664                         if (PageUptodate(page)) {
 665                                 set_buffer_uptodate(bh);
 666                                 continue;
 667                         }
 668                         /*
 669                          * Neither the page nor the buffer are uptodate.  If
 670                          * the buffer is only partially being written to, we
 671                          * need to read it in before the write, i.e. now.
 672                          */
 673                         if ((bh_pos < pos && bh_end > pos) ||
 674                                         (bh_pos < end && bh_end > end)) {
 675                                 /*
 676                                  * If the buffer is fully or partially within
 677                                  * the initialized size, do an actual read.
 678                                  * Otherwise, simply zero the buffer.
 679                                  */
 680                                 read_lock_irqsave(&ni->size_lock, flags);
 681                                 initialized_size = ni->initialized_size;
 682                                 read_unlock_irqrestore(&ni->size_lock, flags);
 683                                 if (bh_pos < initialized_size) {
 684                                         ntfs_submit_bh_for_read(bh);
 685                                         *wait_bh++ = bh;
 686                                 } else {
 687                                         zero_user(page, bh_offset(bh),
 688                                                         blocksize);
 689                                         set_buffer_uptodate(bh);
 690                                 }
 691                         }
 692                         continue;
 693                 }
 694                 /* Unmapped buffer.  Need to map it. */
 695                 bh->b_bdev = vol->sb->s_bdev;
 696                 /*
 697                  * If the current buffer is in the same clusters as the map
 698                  * cache, there is no need to check the runlist again.  The
 699                  * map cache is made up of @vcn, which is the first cached file
 700                  * cluster, @vcn_len which is the number of cached file
 701                  * clusters, @lcn is the device cluster corresponding to @vcn,
 702                  * and @lcn_block is the block number corresponding to @lcn.
 703                  */
 704                 cdelta = bh_cpos - vcn;
 705                 if (likely(!cdelta || (cdelta > 0 && cdelta < vcn_len))) {
 706 map_buffer_cached:
 707                         BUG_ON(lcn < 0);
 708                         bh->b_blocknr = lcn_block +
 709                                         (cdelta << (vol->cluster_size_bits -
 710                                         blocksize_bits)) +
 711                                         (bh_cofs >> blocksize_bits);
 712                         set_buffer_mapped(bh);
 713                         /*
 714                          * If the page is uptodate so is the buffer.  If the
 715                          * buffer is fully outside the write, we ignore it if
 716                          * it was already allocated and we mark it dirty so it
 717                          * gets written out if we allocated it.  On the other
 718                          * hand, if we allocated the buffer but we are not
 719                          * marking it dirty we set buffer_new so we can do
 720                          * error recovery.
 721                          */
 722                         if (PageUptodate(page)) {
 723                                 if (!buffer_uptodate(bh))
 724                                         set_buffer_uptodate(bh);
 725                                 if (unlikely(was_hole)) {
 726                                         /* We allocated the buffer. */
 727                                         clean_bdev_bh_alias(bh);
 728                                         if (bh_end <= pos || bh_pos >= end)
 729                                                 mark_buffer_dirty(bh);
 730                                         else
 731                                                 set_buffer_new(bh);
 732                                 }
 733                                 continue;
 734                         }
 735                         /* Page is _not_ uptodate. */
 736                         if (likely(!was_hole)) {
 737                                 /*
 738                                  * Buffer was already allocated.  If it is not
 739                                  * uptodate and is only partially being written
 740                                  * to, we need to read it in before the write,
 741                                  * i.e. now.
 742                                  */
 743                                 if (!buffer_uptodate(bh) && bh_pos < end &&
 744                                                 bh_end > pos &&
 745                                                 (bh_pos < pos ||
 746                                                 bh_end > end)) {
 747                                         /*
 748                                          * If the buffer is fully or partially
 749                                          * within the initialized size, do an
 750                                          * actual read.  Otherwise, simply zero
 751                                          * the buffer.
 752                                          */
 753                                         read_lock_irqsave(&ni->size_lock,
 754                                                         flags);
 755                                         initialized_size = ni->initialized_size;
 756                                         read_unlock_irqrestore(&ni->size_lock,
 757                                                         flags);
 758                                         if (bh_pos < initialized_size) {
 759                                                 ntfs_submit_bh_for_read(bh);
 760                                                 *wait_bh++ = bh;
 761                                         } else {
 762                                                 zero_user(page, bh_offset(bh),
 763                                                                 blocksize);
 764                                                 set_buffer_uptodate(bh);
 765                                         }
 766                                 }
 767                                 continue;
 768                         }
 769                         /* We allocated the buffer. */
 770                         clean_bdev_bh_alias(bh);
 771                         /*
 772                          * If the buffer is fully outside the write, zero it,
 773                          * set it uptodate, and mark it dirty so it gets
 774                          * written out.  If it is partially being written to,
 775                          * zero region surrounding the write but leave it to
 776                          * commit write to do anything else.  Finally, if the
 777                          * buffer is fully being overwritten, do nothing.
 778                          */
 779                         if (bh_end <= pos || bh_pos >= end) {
 780                                 if (!buffer_uptodate(bh)) {
 781                                         zero_user(page, bh_offset(bh),
 782                                                         blocksize);
 783                                         set_buffer_uptodate(bh);
 784                                 }
 785                                 mark_buffer_dirty(bh);
 786                                 continue;
 787                         }
 788                         set_buffer_new(bh);
 789                         if (!buffer_uptodate(bh) &&
 790                                         (bh_pos < pos || bh_end > end)) {
 791                                 u8 *kaddr;
 792                                 unsigned pofs;
 793
 794                                 kaddr = kmap_atomic(page);
 795                                 if (bh_pos < pos) {
 796                                         pofs = bh_pos & ~PAGE_MASK;
 797                                         memset(kaddr + pofs, 0, pos - bh_pos);
 798                                 }
 799                                 if (bh_end > end) {
 800                                         pofs = end & ~PAGE_MASK;
 801                                         memset(kaddr + pofs, 0, bh_end - end);
 802                                 }
 803                                 kunmap_atomic(kaddr);
 804                                 flush_dcache_page(page);
 805                         }
 806                         continue;
 807                 }
 808                 /*
 809                  * Slow path: this is the first buffer in the cluster.  If it
 810                  * is outside allocated size and is not uptodate, zero it and
 811                  * set it uptodate.
 812                  */
 813                 read_lock_irqsave(&ni->size_lock, flags);
 814                 initialized_size = ni->allocated_size;
 815                 read_unlock_irqrestore(&ni->size_lock, flags);
 816                 if (bh_pos > initialized_size) {
 817                         if (PageUptodate(page)) {
 818                                 if (!buffer_uptodate(bh))
 819                                         set_buffer_uptodate(bh);
 820                         } else if (!buffer_uptodate(bh)) {
 821                                 zero_user(page, bh_offset(bh), blocksize);
 822                                 set_buffer_uptodate(bh);
 823                         }
 824                         continue;
 825                 }
 826                 is_retry = false;
 827                 if (!rl) {
 828                         down_read(&ni->runlist.lock);
 829 retry_remap:
 830                         rl = ni->runlist.rl;
 831                 }
 832                 if (likely(rl != NULL)) {
 833                         /* Seek to element containing target cluster. */
 834                         while (rl->length && rl[1].vcn <= bh_cpos)
 835                                 rl++;
 836                         lcn = ntfs_rl_vcn_to_lcn(rl, bh_cpos);
 837                         if (likely(lcn >= 0)) {
 838                                 /*
 839                                  * Successful remap, setup the map cache and
 840                                  * use that to deal with the buffer.
 841                                  */
 842                                 was_hole = false;
 843                                 vcn = bh_cpos;
 844                                 vcn_len = rl[1].vcn - vcn;
 845                                 lcn_block = lcn << (vol->cluster_size_bits -
 846                                                 blocksize_bits);
 847                                 cdelta = 0;
 848                                 /*
 849                                  * If the number of remaining clusters touched
 850                                  * by the write is smaller or equal to the
 851                                  * number of cached clusters, unlock the
 852                                  * runlist as the map cache will be used from
 853                                  * now on.
 854                                  */
 855                                 if (likely(vcn + vcn_len >= cend)) {
 856                                         if (rl_write_locked) {
 857                                                 up_write(&ni->runlist.lock);
 858                                                 rl_write_locked = false;
 859                                         } else
 860                                                 up_read(&ni->runlist.lock);
 861                                         rl = NULL;
 862                                 }
 863                                 goto map_buffer_cached;
 864                         }
 865                 } else
 866                         lcn = LCN_RL_NOT_MAPPED;
 867                 /*
 868                  * If it is not a hole and not out of bounds, the runlist is
 869                  * probably unmapped so try to map it now.
 870                  */
 871                 if (unlikely(lcn != LCN_HOLE && lcn != LCN_ENOENT)) {
 872                         if (likely(!is_retry && lcn == LCN_RL_NOT_MAPPED)) {
 873                                 /* Attempt to map runlist. */
 874                                 if (!rl_write_locked) {
 875                                         /*
 876                                          * We need the runlist locked for
 877                                          * writing, so if it is locked for
 878                                          * reading relock it now and retry in
 879                                          * case it changed whilst we dropped
 880                                          * the lock.
 881                                          */
 882                                         up_read(&ni->runlist.lock);
 883                                         down_write(&ni->runlist.lock);
 884                                         rl_write_locked = true;
 885                                         goto retry_remap;
 886                                 }
 887                                 err = ntfs_map_runlist_nolock(ni, bh_cpos,
 888                                                 NULL);
 889                                 if (likely(!err)) {
 890                                         is_retry = true;
 891                                         goto retry_remap;
 892                                 }
 893                                 /*
 894                                  * If @vcn is out of bounds, pretend @lcn is
 895                                  * LCN_ENOENT.  As long as the buffer is out
 896                                  * of bounds this will work fine.
 897                                  */
 898                                 if (err == -ENOENT) {
 899                                         lcn = LCN_ENOENT;
 900                                         err = 0;
 901                                         goto rl_not_mapped_enoent;
 902                                 }
 903                         } else
 904                                 err = -EIO;
 905                         /* Failed to map the buffer, even after retrying. */
 906                         bh->b_blocknr = -1;
 907                         ntfs_error(vol->sb, "Failed to write to inode 0x%lx, "
 908                                         "attribute type 0x%x, vcn 0x%llx, "
 909                                         "vcn offset 0x%x, because its "
 910                                         "location on disk could not be "
 911                                         "determined%s (error code %i).",
 912                                         ni->mft_no, ni->type,
 913                                         (unsigned long long)bh_cpos,
 914                                         (unsigned)bh_pos &
 915                                         vol->cluster_size_mask,
 916                                         is_retry ? " even after retrying" : "",
 917                                         err);
 918                         break;
 919                 }
 920 rl_not_mapped_enoent:
 921                 /*
 922                  * The buffer is in a hole or out of bounds.  We need to fill
 923                  * the hole, unless the buffer is in a cluster which is not
 924                  * touched by the write, in which case we just leave the buffer
 925                  * unmapped.  This can only happen when the cluster size is
 926                  * less than the page cache size.
 927                  */
 928                 if (unlikely(vol->cluster_size < PAGE_SIZE)) {
 929                         bh_cend = (bh_end + vol->cluster_size - 1) >>
 930                                         vol->cluster_size_bits;
 931                         if ((bh_cend <= cpos || bh_cpos >= cend)) {
 932                                 bh->b_blocknr = -1;
 933                                 /*
 934                                  * If the buffer is uptodate we skip it.  If it
 935                                  * is not but the page is uptodate, we can set
 936                                  * the buffer uptodate.  If the page is not
 937                                  * uptodate, we can clear the buffer and set it
 938                                  * uptodate.  Whether this is worthwhile is
 939                                  * debatable and this could be removed.
 940                                  */
 941                                 if (PageUptodate(page)) {
 942                                         if (!buffer_uptodate(bh))
 943                                                 set_buffer_uptodate(bh);
 944                                 } else if (!buffer_uptodate(bh)) {
 945                                         zero_user(page, bh_offset(bh),
 946                                                 blocksize);
 947                                         set_buffer_uptodate(bh);
 948                                 }
 949                                 continue;
 950                         }
 951                 }
 952                 /*
 953                  * Out of bounds buffer is invalid if it was not really out of
 954                  * bounds.
 955                  */
 956                 BUG_ON(lcn != LCN_HOLE);
 957                 /*
 958                  * We need the runlist locked for writing, so if it is locked
 959                  * for reading relock it now and retry in case it changed
 960                  * whilst we dropped the lock.
 961                  */
 962                 BUG_ON(!rl);
 963                 if (!rl_write_locked) {
 964                         up_read(&ni->runlist.lock);
 965                         down_write(&ni->runlist.lock);
 966                         rl_write_locked = true;
 967                         goto retry_remap;
 968                 }
 969                 /* Find the previous last allocated cluster. */
 970                 BUG_ON(rl->lcn != LCN_HOLE);
 971                 lcn = -1;
 972                 rl2 = rl;
 973                 while (--rl2 >= ni->runlist.rl) {
 974                         if (rl2->lcn >= 0) {
 975                                 lcn = rl2->lcn + rl2->length;
 976                                 break;
 977                         }
 978                 }
 979                 rl2 = ntfs_cluster_alloc(vol, bh_cpos, 1, lcn, DATA_ZONE,
 980                                 false);
 981                 if (IS_ERR(rl2)) {
 982                         err = PTR_ERR(rl2);
 983                         ntfs_debug("Failed to allocate cluster, error code %i.",
 984                                         err);
 985                         break;
 986                 }
 987                 lcn = rl2->lcn;
 988                 rl = ntfs_runlists_merge(ni->runlist.rl, rl2);
 989                 if (IS_ERR(rl)) {
 990                         err = PTR_ERR(rl);
 991                         if (err != -ENOMEM)
 992                                 err = -EIO;
 993                         if (ntfs_cluster_free_from_rl(vol, rl2)) {
 994                                 ntfs_error(vol->sb, "Failed to release "
 995                                                 "allocated cluster in error "
 996                                                 "code path.  Run chkdsk to "
 997                                                 "recover the lost cluster.");
 998                                 NVolSetErrors(vol);
 999                         }
1000                         ntfs_free(rl2);
1001                         break;
1002                 }
1003                 ni->runlist.rl = rl;
1004                 status.runlist_merged = 1;
1005                 ntfs_debug("Allocated cluster, lcn 0x%llx.",
1006                                 (unsigned long long)lcn);
1007                 /* Map and lock the mft record and get the attribute record. */
1008                 if (!NInoAttr(ni))
1009                         base_ni = ni;
1010                 else
1011                         base_ni = ni->ext.base_ntfs_ino;
1012                 m = map_mft_record(base_ni);
1013                 if (IS_ERR(m)) {
1014                         err = PTR_ERR(m);
1015                         break;
1016                 }
1017                 ctx = ntfs_attr_get_search_ctx(base_ni, m);
1018                 if (unlikely(!ctx)) {
1019                         err = -ENOMEM;
1020                         unmap_mft_record(base_ni);
1021                         break;
1022                 }
1023                 status.mft_attr_mapped = 1;
1024                 err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1025                                 CASE_SENSITIVE, bh_cpos, NULL, 0, ctx);
1026                 if (unlikely(err)) {
1027                         if (err == -ENOENT)
1028                                 err = -EIO;
1029                         break;
1030                 }
1031                 m = ctx->mrec;
1032                 a = ctx->attr;
1033                 /*
1034                  * Find the runlist element with which the attribute extent
1035                  * starts.  Note, we cannot use the _attr_ version because we
1036                  * have mapped the mft record.  That is ok because we know the
1037                  * runlist fragment must be mapped already to have ever gotten
1038                  * here, so we can just use the _rl_ version.
1039                  */
1040                 vcn = sle64_to_cpu(a->data.non_resident.lowest_vcn);
1041                 rl2 = ntfs_rl_find_vcn_nolock(rl, vcn);
1042                 BUG_ON(!rl2);
1043                 BUG_ON(!rl2->length);
1044                 BUG_ON(rl2->lcn < LCN_HOLE);
1045                 highest_vcn = sle64_to_cpu(a->data.non_resident.highest_vcn);
1046                 /*
1047                  * If @highest_vcn is zero, calculate the real highest_vcn
1048                  * (which can really be zero).
1049                  */
1050                 if (!highest_vcn)
1051                         highest_vcn = (sle64_to_cpu(
1052                                         a->data.non_resident.allocated_size) >>
1053                                         vol->cluster_size_bits) - 1;
1054                 /*
1055                  * Determine the size of the mapping pairs array for the new
1056                  * extent, i.e. the old extent with the hole filled.
1057                  */
1058                 mp_size = ntfs_get_size_for_mapping_pairs(vol, rl2, vcn,
1059                                 highest_vcn);
1060                 if (unlikely(mp_size <= 0)) {
1061                         if (!(err = mp_size))
1062                                 err = -EIO;
1063                         ntfs_debug("Failed to get size for mapping pairs "
1064                                         "array, error code %i.", err);
1065                         break;
1066                 }
1067                 /*
1068                  * Resize the attribute record to fit the new mapping pairs
1069                  * array.
1070                  */
1071                 attr_rec_len = le32_to_cpu(a->length);
1072                 err = ntfs_attr_record_resize(m, a, mp_size + le16_to_cpu(
1073                                 a->data.non_resident.mapping_pairs_offset));
1074                 if (unlikely(err)) {
1075                         BUG_ON(err != -ENOSPC);
1076                         // TODO: Deal with this by using the current attribute
1077                         // and fill it with as much of the mapping pairs
1078                         // array as possible.  Then loop over each attribute
1079                         // extent rewriting the mapping pairs arrays as we go
1080                         // along and if when we reach the end we have not
1081                         // enough space, try to resize the last attribute
1082                         // extent and if even that fails, add a new attribute
1083                         // extent.
1084                         // We could also try to resize at each step in the hope
1085                         // that we will not need to rewrite every single extent.
1086                         // Note, we may need to decompress some extents to fill
1087                         // the runlist as we are walking the extents...
1088                         ntfs_error(vol->sb, "Not enough space in the mft "
1089                                         "record for the extended attribute "
1090                                         "record.  This case is not "
1091                                         "implemented yet.");
1092                         err = -EOPNOTSUPP;
1093                         break ;
1094                 }
1095                 status.mp_rebuilt = 1;
1096                 /*
1097                  * Generate the mapping pairs array directly into the attribute
1098                  * record.
1099                  */
1100                 err = ntfs_mapping_pairs_build(vol, (u8*)a + le16_to_cpu(
1101                                 a->data.non_resident.mapping_pairs_offset),
1102                                 mp_size, rl2, vcn, highest_vcn, NULL);
1103                 if (unlikely(err)) {
1104                         ntfs_error(vol->sb, "Cannot fill hole in inode 0x%lx, "
1105                                         "attribute type 0x%x, because building "
1106                                         "the mapping pairs failed with error "
1107                                         "code %i.", vi->i_ino,
1108                                         (unsigned)le32_to_cpu(ni->type), err);
1109                         err = -EIO;
1110                         break;
1111                 }
1112                 /* Update the highest_vcn but only if it was not set. */
1113                 if (unlikely(!a->data.non_resident.highest_vcn))
1114                         a->data.non_resident.highest_vcn =
1115                                         cpu_to_sle64(highest_vcn);
1116                 /*
1117                  * If the attribute is sparse/compressed, update the compressed
1118                  * size in the ntfs_inode structure and the attribute record.
1119                  */
1120                 if (likely(NInoSparse(ni) || NInoCompressed(ni))) {
1121                         /*
1122                          * If we are not in the first attribute extent, switch
1123                          * to it, but first ensure the changes will make it to
1124                          * disk later.
1125                          */
1126                         if (a->data.non_resident.lowest_vcn) {
1127                                 flush_dcache_mft_record_page(ctx->ntfs_ino);
1128                                 mark_mft_record_dirty(ctx->ntfs_ino);
1129                                 ntfs_attr_reinit_search_ctx(ctx);
1130                                 err = ntfs_attr_lookup(ni->type, ni->name,
1131                                                 ni->name_len, CASE_SENSITIVE,
1132                                                 0, NULL, 0, ctx);
1133                                 if (unlikely(err)) {
1134                                         status.attr_switched = 1;
1135                                         break;
1136                                 }
1137                                 /* @m is not used any more so do not set it. */
1138                                 a = ctx->attr;
1139                         }
1140                         write_lock_irqsave(&ni->size_lock, flags);
1141                         ni->itype.compressed.size += vol->cluster_size;
1142                         a->data.non_resident.compressed_size =
1143                                         cpu_to_sle64(ni->itype.compressed.size);
1144                         write_unlock_irqrestore(&ni->size_lock, flags);
1145                 }
1146                 /* Ensure the changes make it to disk. */
1147                 flush_dcache_mft_record_page(ctx->ntfs_ino);
1148                 mark_mft_record_dirty(ctx->ntfs_ino);
1149                 ntfs_attr_put_search_ctx(ctx);
1150                 unmap_mft_record(base_ni);
1151                 /* Successfully filled the hole. */
1152                 status.runlist_merged = 0;
1153                 status.mft_attr_mapped = 0;
1154                 status.mp_rebuilt = 0;
1155                 /* Setup the map cache and use that to deal with the buffer. */
1156                 was_hole = true;
1157                 vcn = bh_cpos;
1158                 vcn_len = 1;
1159                 lcn_block = lcn << (vol->cluster_size_bits - blocksize_bits);
1160                 cdelta = 0;
1161                 /*
1162                  * If the number of remaining clusters in the @pages is smaller
1163                  * or equal to the number of cached clusters, unlock the
1164                  * runlist as the map cache will be used from now on.
1165                  */
1166                 if (likely(vcn + vcn_len >= cend)) {
1167                         up_write(&ni->runlist.lock);
1168                         rl_write_locked = false;
1169                         rl = NULL;
1170                 }
1171                 goto map_buffer_cached;
1172         } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
1173         /* If there are no errors, do the next page. */
1174         if (likely(!err && ++u < nr_pages))
1175                 goto do_next_page;
1176         /* If there are no errors, release the runlist lock if we took it. */
1177         if (likely(!err)) {
1178                 if (unlikely(rl_write_locked)) {
1179                         up_write(&ni->runlist.lock);
1180                         rl_write_locked = false;
1181                 } else if (unlikely(rl))
1182                         up_read(&ni->runlist.lock);
1183                 rl = NULL;
1184         }
1185         /* If we issued read requests, let them complete. */
1186         read_lock_irqsave(&ni->size_lock, flags);
1187         initialized_size = ni->initialized_size;
1188         read_unlock_irqrestore(&ni->size_lock, flags);
1189         while (wait_bh > wait) {
1190                 bh = *--wait_bh;
1191                 wait_on_buffer(bh);
1192                 if (likely(buffer_uptodate(bh))) {
1193                         page = bh->b_page;
1194                         bh_pos = ((s64)page->index << PAGE_SHIFT) +
1195                                         bh_offset(bh);
1196                         /*
1197                          * If the buffer overflows the initialized size, need
1198                          * to zero the overflowing region.
1199                          */
1200                         if (unlikely(bh_pos + blocksize > initialized_size)) {
1201                                 int ofs = 0;
1202
1203                                 if (likely(bh_pos < initialized_size))
1204                                         ofs = initialized_size - bh_pos;
1205                                 zero_user_segment(page, bh_offset(bh) + ofs,
1206                                                 blocksize);
1207                         }
1208                 } else /* if (unlikely(!buffer_uptodate(bh))) */
1209                         err = -EIO;
1210         }
1211         if (likely(!err)) {
1212                 /* Clear buffer_new on all buffers. */
1213                 u = 0;
1214                 do {
1215                         bh = head = page_buffers(pages[u]);
1216                         do {
1217                                 if (buffer_new(bh))
1218                                         clear_buffer_new(bh);
1219                         } while ((bh = bh->b_this_page) != head);
1220                 } while (++u < nr_pages);
1221                 ntfs_debug("Done.");
1222                 return err;
1223         }
1224         if (status.attr_switched) {
1225                 /* Get back to the attribute extent we modified. */
1226                 ntfs_attr_reinit_search_ctx(ctx);
1227                 if (ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1228                                 CASE_SENSITIVE, bh_cpos, NULL, 0, ctx)) {
1229                         ntfs_error(vol->sb, "Failed to find required "
1230                                         "attribute extent of attribute in "
1231                                         "error code path.  Run chkdsk to "
1232                                         "recover.");
1233                         write_lock_irqsave(&ni->size_lock, flags);
1234                         ni->itype.compressed.size += vol->cluster_size;
1235                         write_unlock_irqrestore(&ni->size_lock, flags);
1236                         flush_dcache_mft_record_page(ctx->ntfs_ino);
1237                         mark_mft_record_dirty(ctx->ntfs_ino);
1238                         /*
1239                          * The only thing that is now wrong is the compressed
1240                          * size of the base attribute extent which chkdsk
1241                          * should be able to fix.
1242                          */
1243                         NVolSetErrors(vol);
1244                 } else {
1245                         m = ctx->mrec;
1246                         a = ctx->attr;
1247                         status.attr_switched = 0;
1248                 }
1249         }
1250         /*
1251          * If the runlist has been modified, need to restore it by punching a
1252          * hole into it and we then need to deallocate the on-disk cluster as
1253          * well.  Note, we only modify the runlist if we are able to generate a
1254          * new mapping pairs array, i.e. only when the mapped attribute extent
1255          * is not switched.
1256          */
1257         if (status.runlist_merged && !status.attr_switched) {
1258                 BUG_ON(!rl_write_locked);
1259                 /* Make the file cluster we allocated sparse in the runlist. */
1260                 if (ntfs_rl_punch_nolock(vol, &ni->runlist, bh_cpos, 1)) {
1261                         ntfs_error(vol->sb, "Failed to punch hole into "
1262                                         "attribute runlist in error code "
1263                                         "path.  Run chkdsk to recover the "
1264                                         "lost cluster.");
1265                         NVolSetErrors(vol);
1266                 } else /* if (success) */ {
1267                         status.runlist_merged = 0;
1268                         /*
1269                          * Deallocate the on-disk cluster we allocated but only
1270                          * if we succeeded in punching its vcn out of the
1271                          * runlist.
1272                          */
1273                         down_write(&vol->lcnbmp_lock);
1274                         if (ntfs_bitmap_clear_bit(vol->lcnbmp_ino, lcn)) {
1275                                 ntfs_error(vol->sb, "Failed to release "
1276                                                 "allocated cluster in error "
1277                                                 "code path.  Run chkdsk to "
1278                                                 "recover the lost cluster.");
1279                                 NVolSetErrors(vol);
1280                         }
1281                         up_write(&vol->lcnbmp_lock);
1282                 }
1283         }
1284         /*
1285          * Resize the attribute record to its old size and rebuild the mapping
1286          * pairs array.  Note, we only can do this if the runlist has been
1287          * restored to its old state which also implies that the mapped
1288          * attribute extent is not switched.
1289          */
1290         if (status.mp_rebuilt && !status.runlist_merged) {
1291                 if (ntfs_attr_record_resize(m, a, attr_rec_len)) {
1292                         ntfs_error(vol->sb, "Failed to restore attribute "
1293                                         "record in error code path.  Run "
1294                                         "chkdsk to recover.");
1295                         NVolSetErrors(vol);
1296                 } else /* if (success) */ {
1297                         if (ntfs_mapping_pairs_build(vol, (u8*)a +
1298                                         le16_to_cpu(a->data.non_resident.
1299                                         mapping_pairs_offset), attr_rec_len -
1300                                         le16_to_cpu(a->data.non_resident.
1301                                         mapping_pairs_offset), ni->runlist.rl,
1302                                         vcn, highest_vcn, NULL)) {
1303                                 ntfs_error(vol->sb, "Failed to restore "
1304                                                 "mapping pairs array in error "
1305                                                 "code path.  Run chkdsk to "
1306                                                 "recover.");
1307                                 NVolSetErrors(vol);
1308                         }
1309                         flush_dcache_mft_record_page(ctx->ntfs_ino);
1310                         mark_mft_record_dirty(ctx->ntfs_ino);
1311                 }
1312         }
1313         /* Release the mft record and the attribute. */
1314         if (status.mft_attr_mapped) {
1315                 ntfs_attr_put_search_ctx(ctx);
1316                 unmap_mft_record(base_ni);
1317         }
1318         /* Release the runlist lock. */
1319         if (rl_write_locked)
1320                 up_write(&ni->runlist.lock);
1321         else if (rl)
1322                 up_read(&ni->runlist.lock);
1323         /*
1324          * Zero out any newly allocated blocks to avoid exposing stale data.
1325          * If BH_New is set, we know that the block was newly allocated above
1326          * and that it has not been fully zeroed and marked dirty yet.
1327          */
1328         nr_pages = u;
1329         u = 0;
1330         end = bh_cpos << vol->cluster_size_bits;
1331         do {
1332                 page = pages[u];
1333                 bh = head = page_buffers(page);
1334                 do {
1335                         if (u == nr_pages &&
1336                                         ((s64)page->index << PAGE_SHIFT) +
1337                                         bh_offset(bh) >= end)
1338                                 break;
1339                         if (!buffer_new(bh))
1340                                 continue;
1341                         clear_buffer_new(bh);
1342                         if (!buffer_uptodate(bh)) {
1343                                 if (PageUptodate(page))
1344                                         set_buffer_uptodate(bh);
1345                                 else {
1346                                         zero_user(page, bh_offset(bh),
1347                                                         blocksize);
1348                                         set_buffer_uptodate(bh);
1349                                 }
1350                         }
1351                         mark_buffer_dirty(bh);
1352                 } while ((bh = bh->b_this_page) != head);
1353         } while (++u <= nr_pages);
1354         ntfs_error(vol->sb, "Failed.  Returning error code %i.", err);
1355         return err;
1356 }
1357
1358 static inline void ntfs_flush_dcache_pages(struct page **pages,
1359                 unsigned nr_pages)
1360 {
1361         BUG_ON(!nr_pages);
1362         /*
1363          * Warning: Do not do the decrement at the same time as the call to
1364          * flush_dcache_page() because it is a NULL macro on i386 and hence the
1365          * decrement never happens so the loop never terminates.
1366          */
1367         do {
1368                 --nr_pages;
1369                 flush_dcache_page(pages[nr_pages]);
1370         } while (nr_pages > 0);
1371 }
1372
1373 /**
1374  * ntfs_commit_pages_after_non_resident_write - commit the received data
1375  * @pages:      array of destination pages
1376  * @nr_pages:   number of pages in @pages
1377  * @pos:        byte position in file at which the write begins
1378  * @bytes:      number of bytes to be written
1379  *
1380  * See description of ntfs_commit_pages_after_write(), below.
1381  */
1382 static inline int ntfs_commit_pages_after_non_resident_write(
1383                 struct page **pages, const unsigned nr_pages,
1384                 s64 pos, size_t bytes)
1385 {
1386         s64 end, initialized_size;
1387         struct inode *vi;
1388         ntfs_inode *ni, *base_ni;
1389         struct buffer_head *bh, *head;
1390         ntfs_attr_search_ctx *ctx;
1391         MFT_RECORD *m;
1392         ATTR_RECORD *a;
1393         unsigned long flags;
1394         unsigned blocksize, u;
1395         int err;
1396
1397         vi = pages[0]->mapping->host;
1398         ni = NTFS_I(vi);
1399         blocksize = vi->i_sb->s_blocksize;
1400         end = pos + bytes;
1401         u = 0;
1402         do {
1403                 s64 bh_pos;
1404                 struct page *page;
1405                 bool partial;
1406
1407                 page = pages[u];
1408                 bh_pos = (s64)page->index << PAGE_SHIFT;
1409                 bh = head = page_buffers(page);
1410                 partial = false;
1411                 do {
1412                         s64 bh_end;
1413
1414                         bh_end = bh_pos + blocksize;
1415                         if (bh_end <= pos || bh_pos >= end) {
1416                                 if (!buffer_uptodate(bh))
1417                                         partial = true;
1418                         } else {
1419                                 set_buffer_uptodate(bh);
1420                                 mark_buffer_dirty(bh);
1421                         }
1422                 } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
1423                 /*
1424                  * If all buffers are now uptodate but the page is not, set the
1425                  * page uptodate.
1426                  */
1427                 if (!partial && !PageUptodate(page))
1428                         SetPageUptodate(page);
1429         } while (++u < nr_pages);
1430         /*
1431          * Finally, if we do not need to update initialized_size or i_size we
1432          * are finished.
1433          */
1434         read_lock_irqsave(&ni->size_lock, flags);
1435         initialized_size = ni->initialized_size;
1436         read_unlock_irqrestore(&ni->size_lock, flags);
1437         if (end <= initialized_size) {
1438                 ntfs_debug("Done.");
1439                 return 0;
1440         }
1441         /*
1442          * Update initialized_size/i_size as appropriate, both in the inode and
1443          * the mft record.
1444          */
1445         if (!NInoAttr(ni))
1446                 base_ni = ni;
1447         else
1448                 base_ni = ni->ext.base_ntfs_ino;
1449         /* Map, pin, and lock the mft record. */
1450         m = map_mft_record(base_ni);
1451         if (IS_ERR(m)) {
1452                 err = PTR_ERR(m);
1453                 m = NULL;
1454                 ctx = NULL;
1455                 goto err_out;
1456         }
1457         BUG_ON(!NInoNonResident(ni));
1458         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1459         if (unlikely(!ctx)) {
1460                 err = -ENOMEM;
1461                 goto err_out;
1462         }
1463         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1464                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1465         if (unlikely(err)) {
1466                 if (err == -ENOENT)
1467                         err = -EIO;
1468                 goto err_out;
1469         }
1470         a = ctx->attr;
1471         BUG_ON(!a->non_resident);
1472         write_lock_irqsave(&ni->size_lock, flags);
1473         BUG_ON(end > ni->allocated_size);
1474         ni->initialized_size = end;
1475         a->data.non_resident.initialized_size = cpu_to_sle64(end);
1476         if (end > i_size_read(vi)) {
1477                 i_size_write(vi, end);
1478                 a->data.non_resident.data_size =
1479                                 a->data.non_resident.initialized_size;
1480         }
1481         write_unlock_irqrestore(&ni->size_lock, flags);
1482         /* Mark the mft record dirty, so it gets written back. */
1483         flush_dcache_mft_record_page(ctx->ntfs_ino);
1484         mark_mft_record_dirty(ctx->ntfs_ino);
1485         ntfs_attr_put_search_ctx(ctx);
1486         unmap_mft_record(base_ni);
1487         ntfs_debug("Done.");
1488         return 0;
1489 err_out:
1490         if (ctx)
1491                 ntfs_attr_put_search_ctx(ctx);
1492         if (m)
1493                 unmap_mft_record(base_ni);
1494         ntfs_error(vi->i_sb, "Failed to update initialized_size/i_size (error "
1495                         "code %i).", err);
1496         if (err != -ENOMEM)
1497                 NVolSetErrors(ni->vol);
1498         return err;
1499 }
1500
1501 /**
1502  * ntfs_commit_pages_after_write - commit the received data
1503  * @pages:      array of destination pages
1504  * @nr_pages:   number of pages in @pages
1505  * @pos:        byte position in file at which the write begins
1506  * @bytes:      number of bytes to be written
1507  *
1508  * This is called from ntfs_file_buffered_write() with i_mutex held on the inode
1509  * (@pages[0]->mapping->host).  There are @nr_pages pages in @pages which are
1510  * locked but not kmap()ped.  The source data has already been copied into the
1511  * @page.  ntfs_prepare_pages_for_non_resident_write() has been called before
1512  * the data was copied (for non-resident attributes only) and it returned
1513  * success.
1514  *
1515  * Need to set uptodate and mark dirty all buffers within the boundary of the
1516  * write.  If all buffers in a page are uptodate we set the page uptodate, too.
1517  *
1518  * Setting the buffers dirty ensures that they get written out later when
1519  * ntfs_writepage() is invoked by the VM.
1520  *
1521  * Finally, we need to update i_size and initialized_size as appropriate both
1522  * in the inode and the mft record.
1523  *
1524  * This is modelled after fs/buffer.c::generic_commit_write(), which marks
1525  * buffers uptodate and dirty, sets the page uptodate if all buffers in the
1526  * page are uptodate, and updates i_size if the end of io is beyond i_size.  In
1527  * that case, it also marks the inode dirty.
1528  *
1529  * If things have gone as outlined in
1530  * ntfs_prepare_pages_for_non_resident_write(), we do not need to do any page
1531  * content modifications here for non-resident attributes.  For resident
1532  * attributes we need to do the uptodate bringing here which we combine with
1533  * the copying into the mft record which means we save one atomic kmap.
1534  *
1535  * Return 0 on success or -errno on error.
1536  */
1537 static int ntfs_commit_pages_after_write(struct page **pages,
1538                 const unsigned nr_pages, s64 pos, size_t bytes)
1539 {
1540         s64 end, initialized_size;
1541         loff_t i_size;
1542         struct inode *vi;
1543         ntfs_inode *ni, *base_ni;
1544         struct page *page;
1545         ntfs_attr_search_ctx *ctx;
1546         MFT_RECORD *m;
1547         ATTR_RECORD *a;
1548         char *kattr, *kaddr;
1549         unsigned long flags;
1550         u32 attr_len;
1551         int err;
1552
1553         BUG_ON(!nr_pages);
1554         BUG_ON(!pages);
1555         page = pages[0];
1556         BUG_ON(!page);
1557         vi = page->mapping->host;
1558         ni = NTFS_I(vi);
1559         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, start page "
1560                         "index 0x%lx, nr_pages 0x%x, pos 0x%llx, bytes 0x%zx.",
1561                         vi->i_ino, ni->type, page->index, nr_pages,
1562                         (long long)pos, bytes);
1563         if (NInoNonResident(ni))
1564                 return ntfs_commit_pages_after_non_resident_write(pages,
1565                                 nr_pages, pos, bytes);
1566         BUG_ON(nr_pages > 1);
1567         /*
1568          * Attribute is resident, implying it is not compressed, encrypted, or
1569          * sparse.
1570          */
1571         if (!NInoAttr(ni))
1572                 base_ni = ni;
1573         else
1574                 base_ni = ni->ext.base_ntfs_ino;
1575         BUG_ON(NInoNonResident(ni));
1576         /* Map, pin, and lock the mft record. */
1577         m = map_mft_record(base_ni);
1578         if (IS_ERR(m)) {
1579                 err = PTR_ERR(m);
1580                 m = NULL;
1581                 ctx = NULL;
1582                 goto err_out;
1583         }
1584         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1585         if (unlikely(!ctx)) {
1586                 err = -ENOMEM;
1587                 goto err_out;
1588         }
1589         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1590                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1591         if (unlikely(err)) {
1592                 if (err == -ENOENT)
1593                         err = -EIO;
1594                 goto err_out;
1595         }
1596         a = ctx->attr;
1597         BUG_ON(a->non_resident);
1598         /* The total length of the attribute value. */
1599         attr_len = le32_to_cpu(a->data.resident.value_length);
1600         i_size = i_size_read(vi);
1601         BUG_ON(attr_len != i_size);
1602         BUG_ON(pos > attr_len);
1603         end = pos + bytes;
1604         BUG_ON(end > le32_to_cpu(a->length) -
1605                         le16_to_cpu(a->data.resident.value_offset));
1606         kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
1607         kaddr = kmap_atomic(page);
1608         /* Copy the received data from the page to the mft record. */
1609         memcpy(kattr + pos, kaddr + pos, bytes);
1610         /* Update the attribute length if necessary. */
1611         if (end > attr_len) {
1612                 attr_len = end;
1613                 a->data.resident.value_length = cpu_to_le32(attr_len);
1614         }
1615         /*
1616          * If the page is not uptodate, bring the out of bounds area(s)
1617          * uptodate by copying data from the mft record to the page.
1618          */
1619         if (!PageUptodate(page)) {
1620                 if (pos > 0)
1621                         memcpy(kaddr, kattr, pos);
1622                 if (end < attr_len)
1623                         memcpy(kaddr + end, kattr + end, attr_len - end);
1624                 /* Zero the region outside the end of the attribute value. */
1625                 memset(kaddr + attr_len, 0, PAGE_SIZE - attr_len);
1626                 flush_dcache_page(page);
1627                 SetPageUptodate(page);
1628         }
1629         kunmap_atomic(kaddr);
1630         /* Update initialized_size/i_size if necessary. */
1631         read_lock_irqsave(&ni->size_lock, flags);
1632         initialized_size = ni->initialized_size;
1633         BUG_ON(end > ni->allocated_size);
1634         read_unlock_irqrestore(&ni->size_lock, flags);
1635         BUG_ON(initialized_size != i_size);
1636         if (end > initialized_size) {
1637                 write_lock_irqsave(&ni->size_lock, flags);
1638                 ni->initialized_size = end;
1639                 i_size_write(vi, end);
1640                 write_unlock_irqrestore(&ni->size_lock, flags);
1641         }
1642         /* Mark the mft record dirty, so it gets written back. */
1643         flush_dcache_mft_record_page(ctx->ntfs_ino);
1644         mark_mft_record_dirty(ctx->ntfs_ino);
1645         ntfs_attr_put_search_ctx(ctx);
1646         unmap_mft_record(base_ni);
1647         ntfs_debug("Done.");
1648         return 0;
1649 err_out:
1650         if (err == -ENOMEM) {
1651                 ntfs_warning(vi->i_sb, "Error allocating memory required to "
1652                                 "commit the write.");
1653                 if (PageUptodate(page)) {
1654                         ntfs_warning(vi->i_sb, "Page is uptodate, setting "
1655                                         "dirty so the write will be retried "
1656                                         "later on by the VM.");
1657                         /*
1658                          * Put the page on mapping->dirty_pages, but leave its
1659                          * buffers' dirty state as-is.
1660                          */
1661                         __set_page_dirty_nobuffers(page);
1662                         err = 0;
1663                 } else
1664                         ntfs_error(vi->i_sb, "Page is not uptodate.  Written "
1665                                         "data has been lost.");
1666         } else {
1667                 ntfs_error(vi->i_sb, "Resident attribute commit write failed "
1668                                 "with error %i.", err);
1669                 NVolSetErrors(ni->vol);
1670         }
1671         if (ctx)
1672                 ntfs_attr_put_search_ctx(ctx);
1673         if (m)
1674                 unmap_mft_record(base_ni);
1675         return err;
1676 }
1677
1678 /*
1679  * Copy as much as we can into the pages and return the number of bytes which
1680  * were successfully copied.  If a fault is encountered then clear the pages
1681  * out to (ofs + bytes) and return the number of bytes which were copied.
1682  */
1683 static size_t ntfs_copy_from_user_iter(struct page **pages, unsigned nr_pages,
1684                 unsigned ofs, struct iov_iter *i, size_t bytes)
1685 {
1686         struct page **last_page = pages + nr_pages;
1687         size_t total = 0;
1688         unsigned len, copied;
1689
1690         do {
1691                 len = PAGE_SIZE - ofs;
1692                 if (len > bytes)
1693                         len = bytes;
1694                 copied = copy_page_from_iter_atomic(*pages, ofs, len, i);
1695                 total += copied;
1696                 bytes -= copied;
1697                 if (!bytes)
1698                         break;
1699                 if (copied < len)
1700                         goto err;
1701                 ofs = 0;
1702         } while (++pages < last_page);
1703 out:
1704         return total;
1705 err:
1706         /* Zero the rest of the target like __copy_from_user(). */
1707         len = PAGE_SIZE - copied;
1708         do {
1709                 if (len > bytes)
1710                         len = bytes;
1711                 zero_user(*pages, copied, len);
1712                 bytes -= len;
1713                 copied = 0;
1714                 len = PAGE_SIZE;
1715         } while (++pages < last_page);
1716         goto out;
1717 }
1718
1719 /**
1720  * ntfs_perform_write - perform buffered write to a file
1721  * @file:       file to write to
1722  * @i:          iov_iter with data to write
1723  * @pos:        byte offset in file at which to begin writing to
1724  */
1725 static ssize_t ntfs_perform_write(struct file *file, struct iov_iter *i,
1726                 loff_t pos)
1727 {
1728         struct address_space *mapping = file->f_mapping;
1729         struct inode *vi = mapping->host;
1730         ntfs_inode *ni = NTFS_I(vi);
1731         ntfs_volume *vol = ni->vol;
1732         struct page *pages[NTFS_MAX_PAGES_PER_CLUSTER];
1733         struct page *cached_page = NULL;
1734         VCN last_vcn;
1735         LCN lcn;
1736         size_t bytes;
1737         ssize_t status, written = 0;
1738         unsigned nr_pages;
1739
1740         ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, pos "
1741                         "0x%llx, count 0x%lx.", vi->i_ino,
1742                         (unsigned)le32_to_cpu(ni->type),
1743                         (unsigned long long)pos,
1744                         (unsigned long)iov_iter_count(i));
1745         /*
1746          * If a previous ntfs_truncate() failed, repeat it and abort if it
1747          * fails again.
1748          */
1749         if (unlikely(NInoTruncateFailed(ni))) {
1750                 int err;
1751
1752                 inode_dio_wait(vi);
1753                 err = ntfs_truncate(vi);
1754                 if (err || NInoTruncateFailed(ni)) {
1755                         if (!err)
1756                                 err = -EIO;
1757                         ntfs_error(vol->sb, "Cannot perform write to inode "
1758                                         "0x%lx, attribute type 0x%x, because "
1759                                         "ntfs_truncate() failed (error code "
1760                                         "%i).", vi->i_ino,
1761                                         (unsigned)le32_to_cpu(ni->type), err);
1762                         return err;
1763                 }
1764         }
1765         /*
1766          * Determine the number of pages per cluster for non-resident
1767          * attributes.
1768          */
1769         nr_pages = 1;
1770         if (vol->cluster_size > PAGE_SIZE && NInoNonResident(ni))
1771                 nr_pages = vol->cluster_size >> PAGE_SHIFT;
1772         last_vcn = -1;
1773         do {
1774                 VCN vcn;
1775                 pgoff_t idx, start_idx;
1776                 unsigned ofs, do_pages, u;
1777                 size_t copied;
1778
1779                 start_idx = idx = pos >> PAGE_SHIFT;
1780                 ofs = pos & ~PAGE_MASK;
1781                 bytes = PAGE_SIZE - ofs;
1782                 do_pages = 1;
1783                 if (nr_pages > 1) {
1784                         vcn = pos >> vol->cluster_size_bits;
1785                         if (vcn != last_vcn) {
1786                                 last_vcn = vcn;
1787                                 /*
1788                                  * Get the lcn of the vcn the write is in.  If
1789                                  * it is a hole, need to lock down all pages in
1790                                  * the cluster.
1791                                  */
1792                                 down_read(&ni->runlist.lock);
1793                                 lcn = ntfs_attr_vcn_to_lcn_nolock(ni, pos >>
1794                                                 vol->cluster_size_bits, false);
1795                                 up_read(&ni->runlist.lock);
1796                                 if (unlikely(lcn < LCN_HOLE)) {
1797                                         if (lcn == LCN_ENOMEM)
1798                                                 status = -ENOMEM;
1799                                         else {
1800                                                 status = -EIO;
1801                                                 ntfs_error(vol->sb, "Cannot "
1802                                                         "perform write to "
1803                                                         "inode 0x%lx, "
1804                                                         "attribute type 0x%x, "
1805                                                         "because the attribute "
1806                                                         "is corrupt.",
1807                                                         vi->i_ino, (unsigned)
1808                                                         le32_to_cpu(ni->type));
1809                                         }
1810                                         break;
1811                                 }
1812                                 if (lcn == LCN_HOLE) {
1813                                         start_idx = (pos & ~(s64)
1814                                                         vol->cluster_size_mask)
1815                                                         >> PAGE_SHIFT;
1816                                         bytes = vol->cluster_size - (pos &
1817                                                         vol->cluster_size_mask);
1818                                         do_pages = nr_pages;
1819                                 }
1820                         }
1821                 }
1822                 if (bytes > iov_iter_count(i))
1823                         bytes = iov_iter_count(i);
1824 again:
1825                 /*
1826                  * Bring in the user page(s) that we will copy from _first_.
1827                  * Otherwise there is a nasty deadlock on copying from the same
1828                  * page(s) as we are writing to, without it/them being marked
1829                  * up-to-date.  Note, at present there is nothing to stop the
1830                  * pages being swapped out between us bringing them into memory
1831                  * and doing the actual copying.
1832                  */
1833                 if (unlikely(fault_in_iov_iter_readable(i, bytes))) {
1834                         status = -EFAULT;
1835                         break;
1836                 }
1837                 /* Get and lock @do_pages starting at index @start_idx. */
1838                 status = __ntfs_grab_cache_pages(mapping, start_idx, do_pages,
1839                                 pages, &cached_page);
1840                 if (unlikely(status))
1841                         break;
1842                 /*
1843                  * For non-resident attributes, we need to fill any holes with
1844                  * actual clusters and ensure all bufferes are mapped.  We also
1845                  * need to bring uptodate any buffers that are only partially
1846                  * being written to.
1847                  */
1848                 if (NInoNonResident(ni)) {
1849                         status = ntfs_prepare_pages_for_non_resident_write(
1850                                         pages, do_pages, pos, bytes);
1851                         if (unlikely(status)) {
1852                                 do {
1853                                         unlock_page(pages[--do_pages]);
1854                                         put_page(pages[do_pages]);
1855                                 } while (do_pages);
1856                                 break;
1857                         }
1858                 }
1859                 u = (pos >> PAGE_SHIFT) - pages[0]->index;
1860                 copied = ntfs_copy_from_user_iter(pages + u, do_pages - u, ofs,
1861                                         i, bytes);
1862                 ntfs_flush_dcache_pages(pages + u, do_pages - u);
1863                 status = 0;
1864                 if (likely(copied == bytes)) {
1865                         status = ntfs_commit_pages_after_write(pages, do_pages,
1866                                         pos, bytes);
1867                 }
1868                 do {
1869                         unlock_page(pages[--do_pages]);
1870                         put_page(pages[do_pages]);
1871                 } while (do_pages);
1872                 if (unlikely(status < 0)) {
1873                         iov_iter_revert(i, copied);
1874                         break;
1875                 }
1876                 cond_resched();
1877                 if (unlikely(copied < bytes)) {
1878                         iov_iter_revert(i, copied);
1879                         if (copied)
1880                                 bytes = copied;
1881                         else if (bytes > PAGE_SIZE - ofs)
1882                                 bytes = PAGE_SIZE - ofs;
1883                         goto again;
1884                 }
1885                 pos += copied;
1886                 written += copied;
1887                 balance_dirty_pages_ratelimited(mapping);
1888                 if (fatal_signal_pending(current)) {
1889                         status = -EINTR;
1890                         break;
1891                 }
1892         } while (iov_iter_count(i));
1893         if (cached_page)
1894                 put_page(cached_page);
1895         ntfs_debug("Done.  Returning %s (written 0x%lx, status %li).",
1896                         written ? "written" : "status", (unsigned long)written,
1897                         (long)status);
1898         return written ? written : status;
1899 }
1900
1901 /**
1902  * ntfs_file_write_iter - simple wrapper for ntfs_file_write_iter_nolock()
1903  * @iocb:       IO state structure
1904  * @from:       iov_iter with data to write
1905  *
1906  * Basically the same as generic_file_write_iter() except that it ends up
1907  * up calling ntfs_perform_write() instead of generic_perform_write() and that
1908  * O_DIRECT is not implemented.
1909  */
1910 static ssize_t ntfs_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1911 {
1912         struct file *file = iocb->ki_filp;
1913         struct inode *vi = file_inode(file);
1914         ssize_t written = 0;
1915         ssize_t err;
1916
1917         inode_lock(vi);
1918         /* We can write back this queue in page reclaim. */
1919         current->backing_dev_info = inode_to_bdi(vi);
1920         err = ntfs_prepare_file_for_write(iocb, from);
1921         if (iov_iter_count(from) && !err)
1922                 written = ntfs_perform_write(file, from, iocb->ki_pos);
1923         current->backing_dev_info = NULL;
1924         inode_unlock(vi);
1925         iocb->ki_pos += written;
1926         if (likely(written > 0))
1927                 written = generic_write_sync(iocb, written);
1928         return written ? written : err;
1929 }
1930
1931 /**
1932  * ntfs_file_fsync - sync a file to disk
1933  * @filp:       file to be synced
1934  * @datasync:   if non-zero only flush user data and not metadata
1935  *
1936  * Data integrity sync of a file to disk.  Used for fsync, fdatasync, and msync
1937  * system calls.  This function is inspired by fs/buffer.c::file_fsync().
1938  *
1939  * If @datasync is false, write the mft record and all associated extent mft
1940  * records as well as the $DATA attribute and then sync the block device.
1941  *
1942  * If @datasync is true and the attribute is non-resident, we skip the writing
1943  * of the mft record and all associated extent mft records (this might still
1944  * happen due to the write_inode_now() call).
1945  *
1946  * Also, if @datasync is true, we do not wait on the inode to be written out
1947  * but we always wait on the page cache pages to be written out.
1948  *
1949  * Locking: Caller must hold i_mutex on the inode.
1950  *
1951  * TODO: We should probably also write all attribute/index inodes associated
1952  * with this inode but since we have no simple way of getting to them we ignore
1953  * this problem for now.
1954  */
1955 static int ntfs_file_fsync(struct file *filp, loff_t start, loff_t end,
1956                            int datasync)
1957 {
1958         struct inode *vi = filp->f_mapping->host;
1959         int err, ret = 0;
1960
1961         ntfs_debug("Entering for inode 0x%lx.", vi->i_ino);
1962
1963         err = file_write_and_wait_range(filp, start, end);
1964         if (err)
1965                 return err;
1966         inode_lock(vi);
1967
1968         BUG_ON(S_ISDIR(vi->i_mode));
1969         if (!datasync || !NInoNonResident(NTFS_I(vi)))
1970                 ret = __ntfs_write_inode(vi, 1);
1971         write_inode_now(vi, !datasync);
1972         /*
1973          * NOTE: If we were to use mapping->private_list (see ext2 and
1974          * fs/buffer.c) for dirty blocks then we could optimize the below to be
1975          * sync_mapping_buffers(vi->i_mapping).
1976          */
1977         err = sync_blockdev(vi->i_sb->s_bdev);
1978         if (unlikely(err && !ret))
1979                 ret = err;
1980         if (likely(!ret))
1981                 ntfs_debug("Done.");
1982         else
1983                 ntfs_warning(vi->i_sb, "Failed to f%ssync inode 0x%lx.  Error "
1984                                 "%u.", datasync ? "data" : "", vi->i_ino, -ret);
1985         inode_unlock(vi);
1986         return ret;
1987 }
1988
1989 #endif /* NTFS_RW */
1990
1991 const struct file_operations ntfs_file_ops = {
1992         .llseek         = generic_file_llseek,
1993         .read_iter      = generic_file_read_iter,
1994 #ifdef NTFS_RW
1995         .write_iter     = ntfs_file_write_iter,
1996         .fsync          = ntfs_file_fsync,
1997 #endif /* NTFS_RW */
1998         .mmap           = generic_file_mmap,
1999         .open           = ntfs_file_open,
2000         .splice_read    = generic_file_splice_read,
2001 };
2002
2003 const struct inode_operations ntfs_file_inode_ops = {
2004 #ifdef NTFS_RW
2005         .setattr        = ntfs_setattr,
2006 #endif /* NTFS_RW */
2007 };
2008
2009 const struct file_operations ntfs_empty_file_ops = {};
2010
2011 const struct inode_operations ntfs_empty_inode_ops = {};