NTFS: The big ntfs write(2) rewrite has arrived. We now implement our own
Anton Altaparmakov [Tue, 11 Oct 2005 14:40:40 +0000 (15:40 +0100)]
      file operations ->write(), ->aio_write(), and ->writev() for regular
      files.  This replaces the old use of generic_file_write(), et al and
      the address space operations ->prepare_write and ->commit_write.
      This means that both sparse and non-sparse (unencrypted and
      uncompressed) files can now be extended using the normal write(2)
      code path.  There are two limitations at present and these are that
      we never create sparse files and that we only have limited support
      for highly fragmented files, i.e. ones whose data attribute is split
      across multiple extents.   When such a case is encountered,
      EOPNOTSUPP is returned.

Signed-off-by: Anton Altaparmakov <aia21@cantab.net>

Documentation/filesystems/ntfs.txt
fs/ntfs/ChangeLog
fs/ntfs/Makefile
fs/ntfs/file.c

index a5fbc8e..614de31 100644 (file)
@@ -50,9 +50,14 @@ userspace utilities, etc.
 Features
 ========
 
-- This is a complete rewrite of the NTFS driver that used to be in the kernel.
-  This new driver implements NTFS read support and is functionally equivalent
-  to the old ntfs driver.
+- This is a complete rewrite of the NTFS driver that used to be in the 2.4 and
+  earlier kernels.  This new driver implements NTFS read support and is
+  functionally equivalent to the old ntfs driver and it also implements limited
+  write support.  The biggest limitation at present is that files/directories
+  cannot be created or deleted.  See below for the list of write features that
+  are so far supported.  Another limitation is that writing to compressed files
+  is not implemented at all.  Also, neither read nor write access to encrypted
+  files is so far implemented.
 - The new driver has full support for sparse files on NTFS 3.x volumes which
   the old driver isn't happy with.
 - The new driver supports execution of binaries due to mmap() now being
@@ -78,7 +83,20 @@ Features
 - The new driver supports fsync(2), fdatasync(2), and msync(2).
 - The new driver supports readv(2) and writev(2).
 - The new driver supports access time updates (including mtime and ctime).
-
+- The new driver supports truncate(2) and open(2) with O_TRUNC.  But at present
+  only very limited support for highly fragmented files, i.e. ones which have
+  their data attribute split across multiple extents, is included.  Another
+  limitation is that at present truncate(2) will never create sparse files,
+  since to mark a file sparse we need to modify the directory entry for the
+  file and we do not implement directory modifications yet.
+- The new driver supports write(2) which can both overwrite existing data and
+  extend the file size so that you can write beyond the existing data.  Also,
+  writing into sparse regions is supported and the holes are filled in with
+  clusters.  But at present only limited support for highly fragmented files,
+  i.e. ones which have their data attribute split across multiple extents, is
+  included.  Another limitation is that write(2) will never create sparse
+  files, since to mark a file sparse we need to modify the directory entry for
+  the file and we do not implement directory modifications yet.
 
 Supported mount options
 =======================
@@ -439,6 +457,22 @@ ChangeLog
 
 Note, a technical ChangeLog aimed at kernel hackers is in fs/ntfs/ChangeLog.
 
+2.1.25:
+       - Write support is now extended with write(2) being able to both
+         overwrite existing file data and to extend files.  Also, if a write
+         to a sparse region occurs, write(2) will fill in the hole.  Note,
+         mmap(2) based writes still do not support writing into holes or
+         writing beyond the initialized size.
+       - Write support has a new feature and that is that truncate(2) and
+         open(2) with O_TRUNC are now implemented thus files can be both made
+         smaller and larger.
+       - Note: Both write(2) and truncate(2)/open(2) with O_TRUNC still have
+         limitations in that they
+         - only provide limited support for highly fragmented files.
+         - only work on regular, i.e. uncompressed and unencrypted files.
+         - never create sparse files although this will change once directory
+           operations are implemented.
+       - Lots of bug fixes and enhancements across the board.
 2.1.24:
        - Support journals ($LogFile) which have been modified by chkdsk.  This
          means users can boot into Windows after we marked the volume dirty.
index 3b8ff23..03015c7 100644 (file)
@@ -1,16 +1,15 @@
 ToDo/Notes:
        - Find and fix bugs.
-       - In between ntfs_prepare/commit_write, need exclusion between
-         simultaneous file extensions.  This is given to us by holding i_sem
-         on the inode.  The only places in the kernel when a file is resized
-         are prepare/commit write and ntfs_truncate() for both of which i_sem
-         is held.  Just have to be careful in read-/writepage and other helpers
+       - The only places in the kernel where a file is resized are
+         ntfs_file_write*() and ntfs_truncate() for both of which i_sem is
+         held.  Just have to be careful in read-/writepage and other helpers
          not running under i_sem that we play nice...  Also need to be careful
-         with initialized_size extention in ntfs_prepare_write and writepage.
-         UPDATE: The only things that need to be checked are
-         prepare/commit_write as well as the compressed write and the other
-         attribute resize/write cases like index attributes, etc.  For now
-         none of these are implemented so are safe.
+         with initialized_size extension in ntfs_file_write*() and writepage.
+         UPDATE: The only things that need to be checked are the compressed
+         write and the other attribute resize/write cases like index
+         attributes, etc.  For now none of these are implemented so are safe.
+       - Implement filling in of holes in aops.c::ntfs_writepage() and its
+         helpers.
        - Implement mft.c::sync_mft_mirror_umount().  We currently will just
          leave the volume dirty on umount if the final iput(vol->mft_ino)
          causes a write of any mirrored mft records due to the mft mirror
@@ -20,7 +19,7 @@ ToDo/Notes:
        - Enable the code for setting the NT4 compatibility flag when we start
          making NTFS 1.2 specific modifications.
 
-2.1.25-WIP
+2.1.25 - (Almost) fully implement write(2) and truncate(2).
 
        - Change ntfs_map_runlist_nolock(), ntfs_attr_find_vcn_nolock() and
          {__,}ntfs_cluster_free() to also take an optional attribute search
@@ -49,7 +48,12 @@ ToDo/Notes:
          extend the allocation of an attributes.  Optionally, the data size,
          but not the initialized size can be extended, too.
        - Implement fs/ntfs/inode.[hc]::ntfs_truncate().  It only supports
-         uncompressed and unencrypted files.
+         uncompressed and unencrypted files and it never creates sparse files
+         at least for the moment (making a file sparse requires us to modify
+         its directory entries and we do not support directory operations at
+         the moment).  Also, support for highly fragmented files, i.e. ones
+         whose data attribute is split across multiple extents, is severly
+         limited.  When such a case is encountered, EOPNOTSUPP is returned.
        - Enable ATTR_SIZE attribute changes in ntfs_setattr().  This completes
          the initial implementation of file truncation.  Now both open(2)ing
          a file with the O_TRUNC flag and the {,f}truncate(2) system calls
@@ -61,6 +65,16 @@ ToDo/Notes:
          and cond_resched() in the main loop as we could be dirtying a lot of
          pages and this ensures we play nice with the VM and the system as a
          whole.
+       - Implement file operations ->write, ->aio_write, ->writev for regular
+         files.  This replaces the old use of generic_file_write(), et al and
+         the address space operations ->prepare_write and ->commit_write.
+         This means that both sparse and non-sparse (unencrypted and
+         uncompressed) files can now be extended using the normal write(2)
+         code path.  There are two limitations at present and these are that
+         we never create sparse files and that we only have limited support
+         for highly fragmented files, i.e. ones whose data attribute is split
+         across multiple extents.   When such a case is encountered,
+         EOPNOTSUPP is returned.
 
 2.1.24 - Lots of bug fixes and support more clean journal states.
 
index a3ce2c0..d0d45d1 100644 (file)
@@ -6,7 +6,7 @@ ntfs-objs := aops.o attrib.o collate.o compress.o debug.o dir.o file.o \
             index.o inode.o mft.o mst.o namei.o runlist.o super.o sysctl.o \
             unistr.o upcase.o
 
-EXTRA_CFLAGS = -DNTFS_VERSION=\"2.1.25-WIP\"
+EXTRA_CFLAGS = -DNTFS_VERSION=\"2.1.25\"
 
 ifeq ($(CONFIG_NTFS_DEBUG),y)
 EXTRA_CFLAGS += -DDEBUG
index be9fd1d..cf2a0e2 100644 (file)
  * Foundation,Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  */
 
-#include <linux/pagemap.h>
 #include <linux/buffer_head.h>
+#include <linux/pagemap.h>
+#include <linux/pagevec.h>
+#include <linux/sched.h>
+#include <linux/swap.h>
+#include <linux/uio.h>
+#include <linux/writeback.h>
 
+#include <asm/page.h>
+#include <asm/uaccess.h>
+
+#include "attrib.h"
+#include "bitmap.h"
 #include "inode.h"
 #include "debug.h"
+#include "lcnalloc.h"
+#include "malloc.h"
+#include "mft.h"
 #include "ntfs.h"
 
 /**
@@ -56,6 +69,2176 @@ static int ntfs_file_open(struct inode *vi, struct file *filp)
 #ifdef NTFS_RW
 
 /**
+ * ntfs_attr_extend_initialized - extend the initialized size of an attribute
+ * @ni:                        ntfs inode of the attribute to extend
+ * @new_init_size:     requested new initialized size in bytes
+ * @cached_page:       store any allocated but unused page here
+ * @lru_pvec:          lru-buffering pagevec of the caller
+ *
+ * Extend the initialized size of an attribute described by the ntfs inode @ni
+ * to @new_init_size bytes.  This involves zeroing any non-sparse space between
+ * the old initialized size and @new_init_size both in the page cache and on
+ * disk (if relevant complete pages are zeroed in the page cache then these may
+ * simply be marked dirty for later writeout).  There is one caveat and that is
+ * that if any uptodate page cache pages between the old initialized size and
+ * the smaller of @new_init_size and the file size (vfs inode->i_size) are in
+ * memory, these need to be marked dirty without being zeroed since they could
+ * be non-zero due to mmap() based writes.
+ *
+ * As a side-effect, the file size (vfs inode->i_size) may be incremented as,
+ * in the resident attribute case, it is tied to the initialized size and, in
+ * the non-resident attribute case, it may not fall below the initialized size.
+ *
+ * Note that if the attribute is resident, we do not need to touch the page
+ * cache at all.  This is because if the page cache page is not uptodate we
+ * bring it uptodate later, when doing the write to the mft record since we
+ * then already have the page mapped.  And if the page is uptodate, the
+ * non-initialized region will already have been zeroed when the page was
+ * brought uptodate and the region may in fact already have been overwritten
+ * with new data via mmap() based writes, so we cannot just zero it.  And since
+ * POSIX specifies that the behaviour of resizing a file whilst it is mmap()ped
+ * is unspecified, we choose not to do zeroing and thus we do not need to touch
+ * the page at all.  For a more detailed explanation see ntfs_truncate() which
+ * is in fs/ntfs/inode.c.
+ *
+ * @cached_page and @lru_pvec are just optimisations for dealing with multiple
+ * pages.
+ *
+ * Return 0 on success and -errno on error.  In the case that an error is
+ * encountered it is possible that the initialized size will already have been
+ * incremented some way towards @new_init_size but it is guaranteed that if
+ * this is the case, the necessary zeroing will also have happened and that all
+ * metadata is self-consistent.
+ *
+ * Locking: This function locks the mft record of the base ntfs inode and
+ * maintains the lock throughout execution of the function.  This is required
+ * so that the initialized size of the attribute can be modified safely.
+ */
+static int ntfs_attr_extend_initialized(ntfs_inode *ni, const s64 new_init_size,
+               struct page **cached_page, struct pagevec *lru_pvec)
+{
+       s64 old_init_size;
+       loff_t old_i_size;
+       pgoff_t index, end_index;
+       unsigned long flags;
+       struct inode *vi = VFS_I(ni);
+       ntfs_inode *base_ni;
+       MFT_RECORD *m = NULL;
+       ATTR_RECORD *a;
+       ntfs_attr_search_ctx *ctx = NULL;
+       struct address_space *mapping;
+       struct page *page = NULL;
+       u8 *kattr;
+       int err;
+       u32 attr_len;
+
+       read_lock_irqsave(&ni->size_lock, flags);
+       old_init_size = ni->initialized_size;
+       old_i_size = i_size_read(vi);
+       BUG_ON(new_init_size > ni->allocated_size);
+       read_unlock_irqrestore(&ni->size_lock, flags);
+       ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, "
+                       "old_initialized_size 0x%llx, "
+                       "new_initialized_size 0x%llx, i_size 0x%llx.",
+                       vi->i_ino, (unsigned)le32_to_cpu(ni->type),
+                       (unsigned long long)old_init_size,
+                       (unsigned long long)new_init_size, old_i_size);
+       if (!NInoAttr(ni))
+               base_ni = ni;
+       else
+               base_ni = ni->ext.base_ntfs_ino;
+       /* Use goto to reduce indentation and we need the label below anyway. */
+       if (NInoNonResident(ni))
+               goto do_non_resident_extend;
+       BUG_ON(old_init_size != old_i_size);
+       m = map_mft_record(base_ni);
+       if (IS_ERR(m)) {
+               err = PTR_ERR(m);
+               m = NULL;
+               goto err_out;
+       }
+       ctx = ntfs_attr_get_search_ctx(base_ni, m);
+       if (unlikely(!ctx)) {
+               err = -ENOMEM;
+               goto err_out;
+       }
+       err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
+                       CASE_SENSITIVE, 0, NULL, 0, ctx);
+       if (unlikely(err)) {
+               if (err == -ENOENT)
+                       err = -EIO;
+               goto err_out;
+       }
+       m = ctx->mrec;
+       a = ctx->attr;
+       BUG_ON(a->non_resident);
+       /* The total length of the attribute value. */
+       attr_len = le32_to_cpu(a->data.resident.value_length);
+       BUG_ON(old_i_size != (loff_t)attr_len);
+       /*
+        * Do the zeroing in the mft record and update the attribute size in
+        * the mft record.
+        */
+       kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
+       memset(kattr + attr_len, 0, new_init_size - attr_len);
+       a->data.resident.value_length = cpu_to_le32((u32)new_init_size);
+       /* Finally, update the sizes in the vfs and ntfs inodes. */
+       write_lock_irqsave(&ni->size_lock, flags);
+       i_size_write(vi, new_init_size);
+       ni->initialized_size = new_init_size;
+       write_unlock_irqrestore(&ni->size_lock, flags);
+       goto done;
+do_non_resident_extend:
+       /*
+        * If the new initialized size @new_init_size exceeds the current file
+        * size (vfs inode->i_size), we need to extend the file size to the
+        * new initialized size.
+        */
+       if (new_init_size > old_i_size) {
+               m = map_mft_record(base_ni);
+               if (IS_ERR(m)) {
+                       err = PTR_ERR(m);
+                       m = NULL;
+                       goto err_out;
+               }
+               ctx = ntfs_attr_get_search_ctx(base_ni, m);
+               if (unlikely(!ctx)) {
+                       err = -ENOMEM;
+                       goto err_out;
+               }
+               err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
+                               CASE_SENSITIVE, 0, NULL, 0, ctx);
+               if (unlikely(err)) {
+                       if (err == -ENOENT)
+                               err = -EIO;
+                       goto err_out;
+               }
+               m = ctx->mrec;
+               a = ctx->attr;
+               BUG_ON(!a->non_resident);
+               BUG_ON(old_i_size != (loff_t)
+                               sle64_to_cpu(a->data.non_resident.data_size));
+               a->data.non_resident.data_size = cpu_to_sle64(new_init_size);
+               flush_dcache_mft_record_page(ctx->ntfs_ino);
+               mark_mft_record_dirty(ctx->ntfs_ino);
+               /* Update the file size in the vfs inode. */
+               i_size_write(vi, new_init_size);
+               ntfs_attr_put_search_ctx(ctx);
+               ctx = NULL;
+               unmap_mft_record(base_ni);
+               m = NULL;
+       }
+       mapping = vi->i_mapping;
+       index = old_init_size >> PAGE_CACHE_SHIFT;
+       end_index = (new_init_size + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
+       do {
+               /*
+                * Read the page.  If the page is not present, this will zero
+                * the uninitialized regions for us.
+                */
+               page = read_cache_page(mapping, index,
+                               (filler_t*)mapping->a_ops->readpage, NULL);
+               if (IS_ERR(page)) {
+                       err = PTR_ERR(page);
+                       goto init_err_out;
+               }
+               wait_on_page_locked(page);
+               if (unlikely(!PageUptodate(page) || PageError(page))) {
+                       page_cache_release(page);
+                       err = -EIO;
+                       goto init_err_out;
+               }
+               /*
+                * Update the initialized size in the ntfs inode.  This is
+                * enough to make ntfs_writepage() work.
+                */
+               write_lock_irqsave(&ni->size_lock, flags);
+               ni->initialized_size = (index + 1) << PAGE_CACHE_SHIFT;
+               if (ni->initialized_size > new_init_size)
+                       ni->initialized_size = new_init_size;
+               write_unlock_irqrestore(&ni->size_lock, flags);
+               /* Set the page dirty so it gets written out. */
+               set_page_dirty(page);
+               page_cache_release(page);
+               /*
+                * Play nice with the vm and the rest of the system.  This is
+                * very much needed as we can potentially be modifying the
+                * initialised size from a very small value to a really huge
+                * value, e.g.
+                *      f = open(somefile, O_TRUNC);
+                *      truncate(f, 10GiB);
+                *      seek(f, 10GiB);
+                *      write(f, 1);
+                * And this would mean we would be marking dirty hundreds of
+                * thousands of pages or as in the above example more than
+                * two and a half million pages!
+                *
+                * TODO: For sparse pages could optimize this workload by using
+                * the FsMisc / MiscFs page bit as a "PageIsSparse" bit.  This
+                * would be set in readpage for sparse pages and here we would
+                * not need to mark dirty any pages which have this bit set.
+                * The only caveat is that we have to clear the bit everywhere
+                * where we allocate any clusters that lie in the page or that
+                * contain the page.
+                *
+                * TODO: An even greater optimization would be for us to only
+                * call readpage() on pages which are not in sparse regions as
+                * determined from the runlist.  This would greatly reduce the
+                * number of pages we read and make dirty in the case of sparse
+                * files.
+                */
+               balance_dirty_pages_ratelimited(mapping);
+               cond_resched();
+       } while (++index < end_index);
+       read_lock_irqsave(&ni->size_lock, flags);
+       BUG_ON(ni->initialized_size != new_init_size);
+       read_unlock_irqrestore(&ni->size_lock, flags);
+       /* Now bring in sync the initialized_size in the mft record. */
+       m = map_mft_record(base_ni);
+       if (IS_ERR(m)) {
+               err = PTR_ERR(m);
+               m = NULL;
+               goto init_err_out;
+       }
+       ctx = ntfs_attr_get_search_ctx(base_ni, m);
+       if (unlikely(!ctx)) {
+               err = -ENOMEM;
+               goto init_err_out;
+       }
+       err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
+                       CASE_SENSITIVE, 0, NULL, 0, ctx);
+       if (unlikely(err)) {
+               if (err == -ENOENT)
+                       err = -EIO;
+               goto init_err_out;
+       }
+       m = ctx->mrec;
+       a = ctx->attr;
+       BUG_ON(!a->non_resident);
+       a->data.non_resident.initialized_size = cpu_to_sle64(new_init_size);
+done:
+       flush_dcache_mft_record_page(ctx->ntfs_ino);
+       mark_mft_record_dirty(ctx->ntfs_ino);
+       if (ctx)
+               ntfs_attr_put_search_ctx(ctx);
+       if (m)
+               unmap_mft_record(base_ni);
+       ntfs_debug("Done, initialized_size 0x%llx, i_size 0x%llx.",
+                       (unsigned long long)new_init_size, i_size_read(vi));
+       return 0;
+init_err_out:
+       write_lock_irqsave(&ni->size_lock, flags);
+       ni->initialized_size = old_init_size;
+       write_unlock_irqrestore(&ni->size_lock, flags);
+err_out:
+       if (ctx)
+               ntfs_attr_put_search_ctx(ctx);
+       if (m)
+               unmap_mft_record(base_ni);
+       ntfs_debug("Failed.  Returning error code %i.", err);
+       return err;
+}
+
+/**
+ * ntfs_fault_in_pages_readable -
+ *
+ * Fault a number of userspace pages into pagetables.
+ *
+ * Unlike include/linux/pagemap.h::fault_in_pages_readable(), this one copes
+ * with more than two userspace pages as well as handling the single page case
+ * elegantly.
+ *
+ * If you find this difficult to understand, then think of the while loop being
+ * the following code, except that we do without the integer variable ret:
+ *
+ *     do {
+ *             ret = __get_user(c, uaddr);
+ *             uaddr += PAGE_SIZE;
+ *     } while (!ret && uaddr < end);
+ *
+ * Note, the final __get_user() may well run out-of-bounds of the user buffer,
+ * but _not_ out-of-bounds of the page the user buffer belongs to, and since
+ * this is only a read and not a write, and since it is still in the same page,
+ * it should not matter and this makes the code much simpler.
+ */
+static inline void ntfs_fault_in_pages_readable(const char __user *uaddr,
+               int bytes)
+{
+       const char __user *end;
+       volatile char c;
+
+       /* Set @end to the first byte outside the last page we care about. */
+       end = (const char __user*)PAGE_ALIGN((ptrdiff_t __user)uaddr + bytes);
+
+       while (!__get_user(c, uaddr) && (uaddr += PAGE_SIZE, uaddr < end))
+               ;
+}
+
+/**
+ * ntfs_fault_in_pages_readable_iovec -
+ *
+ * Same as ntfs_fault_in_pages_readable() but operates on an array of iovecs.
+ */
+static inline void ntfs_fault_in_pages_readable_iovec(const struct iovec *iov,
+               size_t iov_ofs, int bytes)
+{
+       do {
+               const char __user *buf;
+               unsigned len;
+
+               buf = iov->iov_base + iov_ofs;
+               len = iov->iov_len - iov_ofs;
+               if (len > bytes)
+                       len = bytes;
+               ntfs_fault_in_pages_readable(buf, len);
+               bytes -= len;
+               iov++;
+               iov_ofs = 0;
+       } while (bytes);
+}
+
+/**
+ * __ntfs_grab_cache_pages - obtain a number of locked pages
+ * @mapping:   address space mapping from which to obtain page cache pages
+ * @index:     starting index in @mapping at which to begin obtaining pages
+ * @nr_pages:  number of page cache pages to obtain
+ * @pages:     array of pages in which to return the obtained page cache pages
+ * @cached_page: allocated but as yet unused page
+ * @lru_pvec:  lru-buffering pagevec of caller
+ *
+ * Obtain @nr_pages locked page cache pages from the mapping @maping and
+ * starting at index @index.
+ *
+ * If a page is newly created, increment its refcount and add it to the
+ * caller's lru-buffering pagevec @lru_pvec.
+ *
+ * This is the same as mm/filemap.c::__grab_cache_page(), except that @nr_pages
+ * are obtained at once instead of just one page and that 0 is returned on
+ * success and -errno on error.
+ *
+ * Note, the page locks are obtained in ascending page index order.
+ */
+static inline int __ntfs_grab_cache_pages(struct address_space *mapping,
+               pgoff_t index, const unsigned nr_pages, struct page **pages,
+               struct page **cached_page, struct pagevec *lru_pvec)
+{
+       int err, nr;
+
+       BUG_ON(!nr_pages);
+       err = nr = 0;
+       do {
+               pages[nr] = find_lock_page(mapping, index);
+               if (!pages[nr]) {
+                       if (!*cached_page) {
+                               *cached_page = page_cache_alloc(mapping);
+                               if (unlikely(!*cached_page)) {
+                                       err = -ENOMEM;
+                                       goto err_out;
+                               }
+                       }
+                       err = add_to_page_cache(*cached_page, mapping, index,
+                                       GFP_KERNEL);
+                       if (unlikely(err)) {
+                               if (err == -EEXIST)
+                                       continue;
+                               goto err_out;
+                       }
+                       pages[nr] = *cached_page;
+                       page_cache_get(*cached_page);
+                       if (unlikely(!pagevec_add(lru_pvec, *cached_page)))
+                               __pagevec_lru_add(lru_pvec);
+                       *cached_page = NULL;
+               }
+               index++;
+               nr++;
+       } while (nr < nr_pages);
+out:
+       return err;
+err_out:
+       while (nr > 0) {
+               unlock_page(pages[--nr]);
+               page_cache_release(pages[nr]);
+       }
+       goto out;
+}
+
+static inline int ntfs_submit_bh_for_read(struct buffer_head *bh)
+{
+       lock_buffer(bh);
+       get_bh(bh);
+       bh->b_end_io = end_buffer_read_sync;
+       return submit_bh(READ, bh);
+}
+
+/**
+ * ntfs_prepare_pages_for_non_resident_write - prepare pages for receiving data
+ * @pages:     array of destination pages
+ * @nr_pages:  number of pages in @pages
+ * @pos:       byte position in file at which the write begins
+ * @bytes:     number of bytes to be written
+ *
+ * This is called for non-resident attributes from ntfs_file_buffered_write()
+ * with i_sem held on the inode (@pages[0]->mapping->host).  There are
+ * @nr_pages pages in @pages which are locked but not kmap()ped.  The source
+ * data has not yet been copied into the @pages.
+ * 
+ * Need to fill any holes with actual clusters, allocate buffers if necessary,
+ * ensure all the buffers are mapped, and bring uptodate any buffers that are
+ * only partially being written to.
+ *
+ * If @nr_pages is greater than one, we are guaranteed that the cluster size is
+ * greater than PAGE_CACHE_SIZE, that all pages in @pages are entirely inside
+ * the same cluster and that they are the entirety of that cluster, and that
+ * the cluster is sparse, i.e. we need to allocate a cluster to fill the hole.
+ *
+ * i_size is not to be modified yet.
+ *
+ * Return 0 on success or -errno on error.
+ */
+static int ntfs_prepare_pages_for_non_resident_write(struct page **pages,
+               unsigned nr_pages, s64 pos, size_t bytes)
+{
+       VCN vcn, highest_vcn = 0, cpos, cend, bh_cpos, bh_cend;
+       LCN lcn;
+       s64 bh_pos, vcn_len, end, initialized_size;
+       sector_t lcn_block;
+       struct page *page;
+       struct inode *vi;
+       ntfs_inode *ni, *base_ni = NULL;
+       ntfs_volume *vol;
+       runlist_element *rl, *rl2;
+       struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
+       ntfs_attr_search_ctx *ctx = NULL;
+       MFT_RECORD *m = NULL;
+       ATTR_RECORD *a = NULL;
+       unsigned long flags;
+       u32 attr_rec_len = 0;
+       unsigned blocksize, u;
+       int err, mp_size;
+       BOOL rl_write_locked, was_hole, is_retry;
+       unsigned char blocksize_bits;
+       struct {
+               u8 runlist_merged:1;
+               u8 mft_attr_mapped:1;
+               u8 mp_rebuilt:1;
+               u8 attr_switched:1;
+       } status = { 0, 0, 0, 0 };
+
+       BUG_ON(!nr_pages);
+       BUG_ON(!pages);
+       BUG_ON(!*pages);
+       vi = pages[0]->mapping->host;
+       ni = NTFS_I(vi);
+       vol = ni->vol;
+       ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, start page "
+                       "index 0x%lx, nr_pages 0x%x, pos 0x%llx, bytes 0x%x.",
+                       vi->i_ino, ni->type, pages[0]->index, nr_pages,
+                       (long long)pos, bytes);
+       blocksize_bits = vi->i_blkbits;
+       blocksize = 1 << blocksize_bits;
+       u = 0;
+       do {
+               struct page *page = pages[u];
+               /*
+                * create_empty_buffers() will create uptodate/dirty buffers if
+                * the page is uptodate/dirty.
+                */
+               if (!page_has_buffers(page)) {
+                       create_empty_buffers(page, blocksize, 0);
+                       if (unlikely(!page_has_buffers(page)))
+                               return -ENOMEM;
+               }
+       } while (++u < nr_pages);
+       rl_write_locked = FALSE;
+       rl = NULL;
+       err = 0;
+       vcn = lcn = -1;
+       vcn_len = 0;
+       lcn_block = -1;
+       was_hole = FALSE;
+       cpos = pos >> vol->cluster_size_bits;
+       end = pos + bytes;
+       cend = (end + vol->cluster_size - 1) >> vol->cluster_size_bits;
+       /*
+        * Loop over each page and for each page over each buffer.  Use goto to
+        * reduce indentation.
+        */
+       u = 0;
+do_next_page:
+       page = pages[u];
+       bh_pos = (s64)page->index << PAGE_CACHE_SHIFT;
+       bh = head = page_buffers(page);
+       do {
+               VCN cdelta;
+               s64 bh_end;
+               unsigned bh_cofs;
+
+               /* Clear buffer_new on all buffers to reinitialise state. */
+               if (buffer_new(bh))
+                       clear_buffer_new(bh);
+               bh_end = bh_pos + blocksize;
+               bh_cpos = bh_pos >> vol->cluster_size_bits;
+               bh_cofs = bh_pos & vol->cluster_size_mask;
+               if (buffer_mapped(bh)) {
+                       /*
+                        * The buffer is already mapped.  If it is uptodate,
+                        * ignore it.
+                        */
+                       if (buffer_uptodate(bh))
+                               continue;
+                       /*
+                        * The buffer is not uptodate.  If the page is uptodate
+                        * set the buffer uptodate and otherwise ignore it.
+                        */
+                       if (PageUptodate(page)) {
+                               set_buffer_uptodate(bh);
+                               continue;
+                       }
+                       /*
+                        * Neither the page nor the buffer are uptodate.  If
+                        * the buffer is only partially being written to, we
+                        * need to read it in before the write, i.e. now.
+                        */
+                       if ((bh_pos < pos && bh_end > pos) ||
+                                       (bh_pos < end && bh_end > end)) {
+                               /*
+                                * If the buffer is fully or partially within
+                                * the initialized size, do an actual read.
+                                * Otherwise, simply zero the buffer.
+                                */
+                               read_lock_irqsave(&ni->size_lock, flags);
+                               initialized_size = ni->initialized_size;
+                               read_unlock_irqrestore(&ni->size_lock, flags);
+                               if (bh_pos < initialized_size) {
+                                       ntfs_submit_bh_for_read(bh);
+                                       *wait_bh++ = bh;
+                               } else {
+                                       u8 *kaddr = kmap_atomic(page, KM_USER0);
+                                       memset(kaddr + bh_offset(bh), 0,
+                                                       blocksize);
+                                       kunmap_atomic(kaddr, KM_USER0);
+                                       flush_dcache_page(page);
+                                       set_buffer_uptodate(bh);
+                               }
+                       }
+                       continue;
+               }
+               /* Unmapped buffer.  Need to map it. */
+               bh->b_bdev = vol->sb->s_bdev;
+               /*
+                * If the current buffer is in the same clusters as the map
+                * cache, there is no need to check the runlist again.  The
+                * map cache is made up of @vcn, which is the first cached file
+                * cluster, @vcn_len which is the number of cached file
+                * clusters, @lcn is the device cluster corresponding to @vcn,
+                * and @lcn_block is the block number corresponding to @lcn.
+                */
+               cdelta = bh_cpos - vcn;
+               if (likely(!cdelta || (cdelta > 0 && cdelta < vcn_len))) {
+map_buffer_cached:
+                       BUG_ON(lcn < 0);
+                       bh->b_blocknr = lcn_block +
+                                       (cdelta << (vol->cluster_size_bits -
+                                       blocksize_bits)) +
+                                       (bh_cofs >> blocksize_bits);
+                       set_buffer_mapped(bh);
+                       /*
+                        * If the page is uptodate so is the buffer.  If the
+                        * buffer is fully outside the write, we ignore it if
+                        * it was already allocated and we mark it dirty so it
+                        * gets written out if we allocated it.  On the other
+                        * hand, if we allocated the buffer but we are not
+                        * marking it dirty we set buffer_new so we can do
+                        * error recovery.
+                        */
+                       if (PageUptodate(page)) {
+                               if (!buffer_uptodate(bh))
+                                       set_buffer_uptodate(bh);
+                               if (unlikely(was_hole)) {
+                                       /* We allocated the buffer. */
+                                       unmap_underlying_metadata(bh->b_bdev,
+                                                       bh->b_blocknr);
+                                       if (bh_end <= pos || bh_pos >= end)
+                                               mark_buffer_dirty(bh);
+                                       else
+                                               set_buffer_new(bh);
+                               }
+                               continue;
+                       }
+                       /* Page is _not_ uptodate. */
+                       if (likely(!was_hole)) {
+                               /*
+                                * Buffer was already allocated.  If it is not
+                                * uptodate and is only partially being written
+                                * to, we need to read it in before the write,
+                                * i.e. now.
+                                */
+                               if (!buffer_uptodate(bh) && ((bh_pos < pos &&
+                                               bh_end > pos) ||
+                                               (bh_end > end &&
+                                               bh_end > end))) {
+                                       /*
+                                        * If the buffer is fully or partially
+                                        * within the initialized size, do an
+                                        * actual read.  Otherwise, simply zero
+                                        * the buffer.
+                                        */
+                                       read_lock_irqsave(&ni->size_lock,
+                                                       flags);
+                                       initialized_size = ni->initialized_size;
+                                       read_unlock_irqrestore(&ni->size_lock,
+                                                       flags);
+                                       if (bh_pos < initialized_size) {
+                                               ntfs_submit_bh_for_read(bh);
+                                               *wait_bh++ = bh;
+                                       } else {
+                                               u8 *kaddr = kmap_atomic(page,
+                                                               KM_USER0);
+                                               memset(kaddr + bh_offset(bh),
+                                                               0, blocksize);
+                                               kunmap_atomic(kaddr, KM_USER0);
+                                               flush_dcache_page(page);
+                                               set_buffer_uptodate(bh);
+                                       }
+                               }
+                               continue;
+                       }
+                       /* We allocated the buffer. */
+                       unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
+                       /*
+                        * If the buffer is fully outside the write, zero it,
+                        * set it uptodate, and mark it dirty so it gets
+                        * written out.  If it is partially being written to,
+                        * zero region surrounding the write but leave it to
+                        * commit write to do anything else.  Finally, if the
+                        * buffer is fully being overwritten, do nothing.
+                        */
+                       if (bh_end <= pos || bh_pos >= end) {
+                               if (!buffer_uptodate(bh)) {
+                                       u8 *kaddr = kmap_atomic(page, KM_USER0);
+                                       memset(kaddr + bh_offset(bh), 0,
+                                                       blocksize);
+                                       kunmap_atomic(kaddr, KM_USER0);
+                                       flush_dcache_page(page);
+                                       set_buffer_uptodate(bh);
+                               }
+                               mark_buffer_dirty(bh);
+                               continue;
+                       }
+                       set_buffer_new(bh);
+                       if (!buffer_uptodate(bh) &&
+                                       (bh_pos < pos || bh_end > end)) {
+                               u8 *kaddr;
+                               unsigned pofs;
+                                       
+                               kaddr = kmap_atomic(page, KM_USER0);
+                               if (bh_pos < pos) {
+                                       pofs = bh_pos & ~PAGE_CACHE_MASK;
+                                       memset(kaddr + pofs, 0, pos - bh_pos);
+                               }
+                               if (bh_end > end) {
+                                       pofs = end & ~PAGE_CACHE_MASK;
+                                       memset(kaddr + pofs, 0, bh_end - end);
+                               }
+                               kunmap_atomic(kaddr, KM_USER0);
+                               flush_dcache_page(page);
+                       }
+                       continue;
+               }
+               /*
+                * Slow path: this is the first buffer in the cluster.  If it
+                * is outside allocated size and is not uptodate, zero it and
+                * set it uptodate.
+                */
+               read_lock_irqsave(&ni->size_lock, flags);
+               initialized_size = ni->allocated_size;
+               read_unlock_irqrestore(&ni->size_lock, flags);
+               if (bh_pos > initialized_size) {
+                       if (PageUptodate(page)) {
+                               if (!buffer_uptodate(bh))
+                                       set_buffer_uptodate(bh);
+                       } else if (!buffer_uptodate(bh)) {
+                               u8 *kaddr = kmap_atomic(page, KM_USER0);
+                               memset(kaddr + bh_offset(bh), 0, blocksize);
+                               kunmap_atomic(kaddr, KM_USER0);
+                               flush_dcache_page(page);
+                               set_buffer_uptodate(bh);
+                       }
+                       continue;
+               }
+               is_retry = FALSE;
+               if (!rl) {
+                       down_read(&ni->runlist.lock);
+retry_remap:
+                       rl = ni->runlist.rl;
+               }
+               if (likely(rl != NULL)) {
+                       /* Seek to element containing target cluster. */
+                       while (rl->length && rl[1].vcn <= bh_cpos)
+                               rl++;
+                       lcn = ntfs_rl_vcn_to_lcn(rl, bh_cpos);
+                       if (likely(lcn >= 0)) {
+                               /*
+                                * Successful remap, setup the map cache and
+                                * use that to deal with the buffer.
+                                */
+                               was_hole = FALSE;
+                               vcn = bh_cpos;
+                               vcn_len = rl[1].vcn - vcn;
+                               lcn_block = lcn << (vol->cluster_size_bits -
+                                               blocksize_bits);
+                               /*
+                                * If the number of remaining clusters in the
+                                * @pages is smaller or equal to the number of
+                                * cached clusters, unlock the runlist as the
+                                * map cache will be used from now on.
+                                */
+                               if (likely(vcn + vcn_len >= cend)) {
+                                       if (rl_write_locked) {
+                                               up_write(&ni->runlist.lock);
+                                               rl_write_locked = FALSE;
+                                       } else
+                                               up_read(&ni->runlist.lock);
+                                       rl = NULL;
+                               }
+                               goto map_buffer_cached;
+                       }
+               } else
+                       lcn = LCN_RL_NOT_MAPPED;
+               /*
+                * If it is not a hole and not out of bounds, the runlist is
+                * probably unmapped so try to map it now.
+                */
+               if (unlikely(lcn != LCN_HOLE && lcn != LCN_ENOENT)) {
+                       if (likely(!is_retry && lcn == LCN_RL_NOT_MAPPED)) {
+                               /* Attempt to map runlist. */
+                               if (!rl_write_locked) {
+                                       /*
+                                        * We need the runlist locked for
+                                        * writing, so if it is locked for
+                                        * reading relock it now and retry in
+                                        * case it changed whilst we dropped
+                                        * the lock.
+                                        */
+                                       up_read(&ni->runlist.lock);
+                                       down_write(&ni->runlist.lock);
+                                       rl_write_locked = TRUE;
+                                       goto retry_remap;
+                               }
+                               err = ntfs_map_runlist_nolock(ni, bh_cpos,
+                                               NULL);
+                               if (likely(!err)) {
+                                       is_retry = TRUE;
+                                       goto retry_remap;
+                               }
+                               /*
+                                * If @vcn is out of bounds, pretend @lcn is
+                                * LCN_ENOENT.  As long as the buffer is out
+                                * of bounds this will work fine.
+                                */
+                               if (err == -ENOENT) {
+                                       lcn = LCN_ENOENT;
+                                       err = 0;
+                                       goto rl_not_mapped_enoent;
+                               }
+                       } else
+                               err = -EIO;
+                       /* Failed to map the buffer, even after retrying. */
+                       bh->b_blocknr = -1;
+                       ntfs_error(vol->sb, "Failed to write to inode 0x%lx, "
+                                       "attribute type 0x%x, vcn 0x%llx, "
+                                       "vcn offset 0x%x, because its "
+                                       "location on disk could not be "
+                                       "determined%s (error code %i).",
+                                       ni->mft_no, ni->type,
+                                       (unsigned long long)bh_cpos,
+                                       (unsigned)bh_pos &
+                                       vol->cluster_size_mask,
+                                       is_retry ? " even after retrying" : "",
+                                       err);
+                       break;
+               }
+rl_not_mapped_enoent:
+               /*
+                * The buffer is in a hole or out of bounds.  We need to fill
+                * the hole, unless the buffer is in a cluster which is not
+                * touched by the write, in which case we just leave the buffer
+                * unmapped.  This can only happen when the cluster size is
+                * less than the page cache size.
+                */
+               if (unlikely(vol->cluster_size < PAGE_CACHE_SIZE)) {
+                       bh_cend = (bh_end + vol->cluster_size - 1) >>
+                                       vol->cluster_size_bits;
+                       if ((bh_cend <= cpos || bh_cpos >= cend)) {
+                               bh->b_blocknr = -1;
+                               /*
+                                * If the buffer is uptodate we skip it.  If it
+                                * is not but the page is uptodate, we can set
+                                * the buffer uptodate.  If the page is not
+                                * uptodate, we can clear the buffer and set it
+                                * uptodate.  Whether this is worthwhile is
+                                * debatable and this could be removed.
+                                */
+                               if (PageUptodate(page)) {
+                                       if (!buffer_uptodate(bh))
+                                               set_buffer_uptodate(bh);
+                               } else if (!buffer_uptodate(bh)) {
+                                       u8 *kaddr = kmap_atomic(page, KM_USER0);
+                                       memset(kaddr + bh_offset(bh), 0,
+                                                       blocksize);
+                                       kunmap_atomic(kaddr, KM_USER0);
+                                       flush_dcache_page(page);
+                                       set_buffer_uptodate(bh);
+                               }
+                               continue;
+                       }
+               }
+               /*
+                * Out of bounds buffer is invalid if it was not really out of
+                * bounds.
+                */
+               BUG_ON(lcn != LCN_HOLE);
+               /*
+                * We need the runlist locked for writing, so if it is locked
+                * for reading relock it now and retry in case it changed
+                * whilst we dropped the lock.
+                */
+               BUG_ON(!rl);
+               if (!rl_write_locked) {
+                       up_read(&ni->runlist.lock);
+                       down_write(&ni->runlist.lock);
+                       rl_write_locked = TRUE;
+                       goto retry_remap;
+               }
+               /* Find the previous last allocated cluster. */
+               BUG_ON(rl->lcn != LCN_HOLE);
+               lcn = -1;
+               rl2 = rl;
+               while (--rl2 >= ni->runlist.rl) {
+                       if (rl2->lcn >= 0) {
+                               lcn = rl2->lcn + rl2->length;
+                               break;
+                       }
+               }
+               rl2 = ntfs_cluster_alloc(vol, bh_cpos, 1, lcn, DATA_ZONE,
+                               FALSE);
+               if (IS_ERR(rl2)) {
+                       err = PTR_ERR(rl2);
+                       ntfs_debug("Failed to allocate cluster, error code %i.",
+                                       err);
+                       break;
+               }
+               lcn = rl2->lcn;
+               rl = ntfs_runlists_merge(ni->runlist.rl, rl2);
+               if (IS_ERR(rl)) {
+                       err = PTR_ERR(rl);
+                       if (err != -ENOMEM)
+                               err = -EIO;
+                       if (ntfs_cluster_free_from_rl(vol, rl2)) {
+                               ntfs_error(vol->sb, "Failed to release "
+                                               "allocated cluster in error "
+                                               "code path.  Run chkdsk to "
+                                               "recover the lost cluster.");
+                               NVolSetErrors(vol);
+                       }
+                       ntfs_free(rl2);
+                       break;
+               }
+               ni->runlist.rl = rl;
+               status.runlist_merged = 1;
+               ntfs_debug("Allocated cluster, lcn 0x%llx.", lcn);
+               /* Map and lock the mft record and get the attribute record. */
+               if (!NInoAttr(ni))
+                       base_ni = ni;
+               else
+                       base_ni = ni->ext.base_ntfs_ino;
+               m = map_mft_record(base_ni);
+               if (IS_ERR(m)) {
+                       err = PTR_ERR(m);
+                       break;
+               }
+               ctx = ntfs_attr_get_search_ctx(base_ni, m);
+               if (unlikely(!ctx)) {
+                       err = -ENOMEM;
+                       unmap_mft_record(base_ni);
+                       break;
+               }
+               status.mft_attr_mapped = 1;
+               err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
+                               CASE_SENSITIVE, bh_cpos, NULL, 0, ctx);
+               if (unlikely(err)) {
+                       if (err == -ENOENT)
+                               err = -EIO;
+                       break;
+               }
+               m = ctx->mrec;
+               a = ctx->attr;
+               /*
+                * Find the runlist element with which the attribute extent
+                * starts.  Note, we cannot use the _attr_ version because we
+                * have mapped the mft record.  That is ok because we know the
+                * runlist fragment must be mapped already to have ever gotten
+                * here, so we can just use the _rl_ version.
+                */
+               vcn = sle64_to_cpu(a->data.non_resident.lowest_vcn);
+               rl2 = ntfs_rl_find_vcn_nolock(rl, vcn);
+               BUG_ON(!rl2);
+               BUG_ON(!rl2->length);
+               BUG_ON(rl2->lcn < LCN_HOLE);
+               highest_vcn = sle64_to_cpu(a->data.non_resident.highest_vcn);
+               /*
+                * If @highest_vcn is zero, calculate the real highest_vcn
+                * (which can really be zero).
+                */
+               if (!highest_vcn)
+                       highest_vcn = (sle64_to_cpu(
+                                       a->data.non_resident.allocated_size) >>
+                                       vol->cluster_size_bits) - 1;
+               /*
+                * Determine the size of the mapping pairs array for the new
+                * extent, i.e. the old extent with the hole filled.
+                */
+               mp_size = ntfs_get_size_for_mapping_pairs(vol, rl2, vcn,
+                               highest_vcn);
+               if (unlikely(mp_size <= 0)) {
+                       if (!(err = mp_size))
+                               err = -EIO;
+                       ntfs_debug("Failed to get size for mapping pairs "
+                                       "array, error code %i.", err);
+                       break;
+               }
+               /*
+                * Resize the attribute record to fit the new mapping pairs
+                * array.
+                */
+               attr_rec_len = le32_to_cpu(a->length);
+               err = ntfs_attr_record_resize(m, a, mp_size + le16_to_cpu(
+                               a->data.non_resident.mapping_pairs_offset));
+               if (unlikely(err)) {
+                       BUG_ON(err != -ENOSPC);
+                       // TODO: Deal with this by using the current attribute
+                       // and fill it with as much of the mapping pairs
+                       // array as possible.  Then loop over each attribute
+                       // extent rewriting the mapping pairs arrays as we go
+                       // along and if when we reach the end we have not
+                       // enough space, try to resize the last attribute
+                       // extent and if even that fails, add a new attribute
+                       // extent.
+                       // We could also try to resize at each step in the hope
+                       // that we will not need to rewrite every single extent.
+                       // Note, we may need to decompress some extents to fill
+                       // the runlist as we are walking the extents...
+                       ntfs_error(vol->sb, "Not enough space in the mft "
+                                       "record for the extended attribute "
+                                       "record.  This case is not "
+                                       "implemented yet.");
+                       err = -EOPNOTSUPP;
+                       break ;
+               }
+               status.mp_rebuilt = 1;
+               /*
+                * Generate the mapping pairs array directly into the attribute
+                * record.
+                */
+               err = ntfs_mapping_pairs_build(vol, (u8*)a + le16_to_cpu(
+                               a->data.non_resident.mapping_pairs_offset),
+                               mp_size, rl2, vcn, highest_vcn, NULL);
+               if (unlikely(err)) {
+                       ntfs_error(vol->sb, "Cannot fill hole in inode 0x%lx, "
+                                       "attribute type 0x%x, because building "
+                                       "the mapping pairs failed with error "
+                                       "code %i.", vi->i_ino,
+                                       (unsigned)le32_to_cpu(ni->type), err);
+                       err = -EIO;
+                       break;
+               }
+               /* Update the highest_vcn but only if it was not set. */
+               if (unlikely(!a->data.non_resident.highest_vcn))
+                       a->data.non_resident.highest_vcn =
+                                       cpu_to_sle64(highest_vcn);
+               /*
+                * If the attribute is sparse/compressed, update the compressed
+                * size in the ntfs_inode structure and the attribute record.
+                */
+               if (likely(NInoSparse(ni) || NInoCompressed(ni))) {
+                       /*
+                        * If we are not in the first attribute extent, switch
+                        * to it, but first ensure the changes will make it to
+                        * disk later.
+                        */
+                       if (a->data.non_resident.lowest_vcn) {
+                               flush_dcache_mft_record_page(ctx->ntfs_ino);
+                               mark_mft_record_dirty(ctx->ntfs_ino);
+                               ntfs_attr_reinit_search_ctx(ctx);
+                               err = ntfs_attr_lookup(ni->type, ni->name,
+                                               ni->name_len, CASE_SENSITIVE,
+                                               0, NULL, 0, ctx);
+                               if (unlikely(err)) {
+                                       status.attr_switched = 1;
+                                       break;
+                               }
+                               /* @m is not used any more so do not set it. */
+                               a = ctx->attr;
+                       }
+                       write_lock_irqsave(&ni->size_lock, flags);
+                       ni->itype.compressed.size += vol->cluster_size;
+                       a->data.non_resident.compressed_size =
+                                       cpu_to_sle64(ni->itype.compressed.size);
+                       write_unlock_irqrestore(&ni->size_lock, flags);
+               }
+               /* Ensure the changes make it to disk. */
+               flush_dcache_mft_record_page(ctx->ntfs_ino);
+               mark_mft_record_dirty(ctx->ntfs_ino);
+               ntfs_attr_put_search_ctx(ctx);
+               unmap_mft_record(base_ni);
+               /* Successfully filled the hole. */
+               status.runlist_merged = 0;
+               status.mft_attr_mapped = 0;
+               status.mp_rebuilt = 0;
+               /* Setup the map cache and use that to deal with the buffer. */
+               was_hole = TRUE;
+               vcn = bh_cpos;
+               vcn_len = 1;
+               lcn_block = lcn << (vol->cluster_size_bits - blocksize_bits);
+               cdelta = 0;
+               /*
+                * If the number of remaining clusters in the @pages is smaller
+                * or equal to the number of cached clusters, unlock the
+                * runlist as the map cache will be used from now on.
+                */
+               if (likely(vcn + vcn_len >= cend)) {
+                       up_write(&ni->runlist.lock);
+                       rl_write_locked = FALSE;
+                       rl = NULL;
+               }
+               goto map_buffer_cached;
+       } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
+       /* If there are no errors, do the next page. */
+       if (likely(!err && ++u < nr_pages))
+               goto do_next_page;
+       /* If there are no errors, release the runlist lock if we took it. */
+       if (likely(!err)) {
+               if (unlikely(rl_write_locked)) {
+                       up_write(&ni->runlist.lock);
+                       rl_write_locked = FALSE;
+               } else if (unlikely(rl))
+                       up_read(&ni->runlist.lock);
+               rl = NULL;
+       }
+       /* If we issued read requests, let them complete. */
+       read_lock_irqsave(&ni->size_lock, flags);
+       initialized_size = ni->initialized_size;
+       read_unlock_irqrestore(&ni->size_lock, flags);
+       while (wait_bh > wait) {
+               bh = *--wait_bh;
+               wait_on_buffer(bh);
+               if (likely(buffer_uptodate(bh))) {
+                       page = bh->b_page;
+                       bh_pos = ((s64)page->index << PAGE_CACHE_SHIFT) +
+                                       bh_offset(bh);
+                       /*
+                        * If the buffer overflows the initialized size, need
+                        * to zero the overflowing region.
+                        */
+                       if (unlikely(bh_pos + blocksize > initialized_size)) {
+                               u8 *kaddr;
+                               int ofs = 0;
+
+                               if (likely(bh_pos < initialized_size))
+                                       ofs = initialized_size - bh_pos;
+                               kaddr = kmap_atomic(page, KM_USER0);
+                               memset(kaddr + bh_offset(bh) + ofs, 0,
+                                               blocksize - ofs);
+                               kunmap_atomic(kaddr, KM_USER0);
+                               flush_dcache_page(page);
+                       }
+               } else /* if (unlikely(!buffer_uptodate(bh))) */
+                       err = -EIO;
+       }
+       if (likely(!err)) {
+               /* Clear buffer_new on all buffers. */
+               u = 0;
+               do {
+                       bh = head = page_buffers(pages[u]);
+                       do {
+                               if (buffer_new(bh))
+                                       clear_buffer_new(bh);
+                       } while ((bh = bh->b_this_page) != head);
+               } while (++u < nr_pages);
+               ntfs_debug("Done.");
+               return err;
+       }
+       if (status.attr_switched) {
+               /* Get back to the attribute extent we modified. */
+               ntfs_attr_reinit_search_ctx(ctx);
+               if (ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
+                               CASE_SENSITIVE, bh_cpos, NULL, 0, ctx)) {
+                       ntfs_error(vol->sb, "Failed to find required "
+                                       "attribute extent of attribute in "
+                                       "error code path.  Run chkdsk to "
+                                       "recover.");
+                       write_lock_irqsave(&ni->size_lock, flags);
+                       ni->itype.compressed.size += vol->cluster_size;
+                       write_unlock_irqrestore(&ni->size_lock, flags);
+                       flush_dcache_mft_record_page(ctx->ntfs_ino);
+                       mark_mft_record_dirty(ctx->ntfs_ino);
+                       /*
+                        * The only thing that is now wrong is the compressed
+                        * size of the base attribute extent which chkdsk
+                        * should be able to fix.
+                        */
+                       NVolSetErrors(vol);
+               } else {
+                       m = ctx->mrec;
+                       a = ctx->attr;
+                       status.attr_switched = 0;
+               }
+       }
+       /*
+        * If the runlist has been modified, need to restore it by punching a
+        * hole into it and we then need to deallocate the on-disk cluster as
+        * well.  Note, we only modify the runlist if we are able to generate a
+        * new mapping pairs array, i.e. only when the mapped attribute extent
+        * is not switched.
+        */
+       if (status.runlist_merged && !status.attr_switched) {
+               BUG_ON(!rl_write_locked);
+               /* Make the file cluster we allocated sparse in the runlist. */
+               if (ntfs_rl_punch_nolock(vol, &ni->runlist, bh_cpos, 1)) {
+                       ntfs_error(vol->sb, "Failed to punch hole into "
+                                       "attribute runlist in error code "
+                                       "path.  Run chkdsk to recover the "
+                                       "lost cluster.");
+                       make_bad_inode(vi);
+                       make_bad_inode(VFS_I(base_ni));
+                       NVolSetErrors(vol);
+               } else /* if (success) */ {
+                       status.runlist_merged = 0;
+                       /*
+                        * Deallocate the on-disk cluster we allocated but only
+                        * if we succeeded in punching its vcn out of the
+                        * runlist.
+                        */
+                       down_write(&vol->lcnbmp_lock);
+                       if (ntfs_bitmap_clear_bit(vol->lcnbmp_ino, lcn)) {
+                               ntfs_error(vol->sb, "Failed to release "
+                                               "allocated cluster in error "
+                                               "code path.  Run chkdsk to "
+                                               "recover the lost cluster.");
+                               NVolSetErrors(vol);
+                       }
+                       up_write(&vol->lcnbmp_lock);
+               }
+       }
+       /*
+        * Resize the attribute record to its old size and rebuild the mapping
+        * pairs array.  Note, we only can do this if the runlist has been
+        * restored to its old state which also implies that the mapped
+        * attribute extent is not switched.
+        */
+       if (status.mp_rebuilt && !status.runlist_merged) {
+               if (ntfs_attr_record_resize(m, a, attr_rec_len)) {
+                       ntfs_error(vol->sb, "Failed to restore attribute "
+                                       "record in error code path.  Run "
+                                       "chkdsk to recover.");
+                       make_bad_inode(vi);
+                       make_bad_inode(VFS_I(base_ni));
+                       NVolSetErrors(vol);
+               } else /* if (success) */ {
+                       if (ntfs_mapping_pairs_build(vol, (u8*)a +
+                                       le16_to_cpu(a->data.non_resident.
+                                       mapping_pairs_offset), attr_rec_len -
+                                       le16_to_cpu(a->data.non_resident.
+                                       mapping_pairs_offset), ni->runlist.rl,
+                                       vcn, highest_vcn, NULL)) {
+                               ntfs_error(vol->sb, "Failed to restore "
+                                               "mapping pairs array in error "
+                                               "code path.  Run chkdsk to "
+                                               "recover.");
+                               make_bad_inode(vi);
+                               make_bad_inode(VFS_I(base_ni));
+                               NVolSetErrors(vol);
+                       }
+                       flush_dcache_mft_record_page(ctx->ntfs_ino);
+                       mark_mft_record_dirty(ctx->ntfs_ino);
+               }
+       }
+       /* Release the mft record and the attribute. */
+       if (status.mft_attr_mapped) {
+               ntfs_attr_put_search_ctx(ctx);
+               unmap_mft_record(base_ni);
+       }
+       /* Release the runlist lock. */
+       if (rl_write_locked)
+               up_write(&ni->runlist.lock);
+       else if (rl)
+               up_read(&ni->runlist.lock);
+       /*
+        * Zero out any newly allocated blocks to avoid exposing stale data.
+        * If BH_New is set, we know that the block was newly allocated above
+        * and that it has not been fully zeroed and marked dirty yet.
+        */
+       nr_pages = u;
+       u = 0;
+       end = bh_cpos << vol->cluster_size_bits;
+       do {
+               page = pages[u];
+               bh = head = page_buffers(page);
+               do {
+                       if (u == nr_pages &&
+                                       ((s64)page->index << PAGE_CACHE_SHIFT) +
+                                       bh_offset(bh) >= end)
+                               break;
+                       if (!buffer_new(bh))
+                               continue;
+                       clear_buffer_new(bh);
+                       if (!buffer_uptodate(bh)) {
+                               if (PageUptodate(page))
+                                       set_buffer_uptodate(bh);
+                               else {
+                                       u8 *kaddr = kmap_atomic(page, KM_USER0);
+                                       memset(kaddr + bh_offset(bh), 0,
+                                                       blocksize);
+                                       kunmap_atomic(kaddr, KM_USER0);
+                                       flush_dcache_page(page);
+                                       set_buffer_uptodate(bh);
+                               }
+                       }
+                       mark_buffer_dirty(bh);
+               } while ((bh = bh->b_this_page) != head);
+       } while (++u <= nr_pages);
+       ntfs_error(vol->sb, "Failed.  Returning error code %i.", err);
+       return err;
+}
+
+/*
+ * Copy as much as we can into the pages and return the number of bytes which
+ * were sucessfully copied.  If a fault is encountered then clear the pages
+ * out to (ofs + bytes) and return the number of bytes which were copied.
+ */
+static inline size_t ntfs_copy_from_user(struct page **pages,
+               unsigned nr_pages, unsigned ofs, const char __user *buf,
+               size_t bytes)
+{
+       struct page **last_page = pages + nr_pages;
+       char *kaddr;
+       size_t total = 0;
+       unsigned len;
+       int left;
+
+       do {
+               len = PAGE_CACHE_SIZE - ofs;
+               if (len > bytes)
+                       len = bytes;
+               kaddr = kmap_atomic(*pages, KM_USER0);
+               left = __copy_from_user_inatomic(kaddr + ofs, buf, len);
+               kunmap_atomic(kaddr, KM_USER0);
+               if (unlikely(left)) {
+                       /* Do it the slow way. */
+                       kaddr = kmap(*pages);
+                       left = __copy_from_user(kaddr + ofs, buf, len);
+                       kunmap(*pages);
+                       if (unlikely(left))
+                               goto err_out;
+               }
+               total += len;
+               bytes -= len;
+               if (!bytes)
+                       break;
+               buf += len;
+               ofs = 0;
+       } while (++pages < last_page);
+out:
+       return total;
+err_out:
+       total += len - left;
+       /* Zero the rest of the target like __copy_from_user(). */
+       while (++pages < last_page) {
+               bytes -= len;
+               if (!bytes)
+                       break;
+               len = PAGE_CACHE_SIZE;
+               if (len > bytes)
+                       len = bytes;
+               kaddr = kmap_atomic(*pages, KM_USER0);
+               memset(kaddr, 0, len);
+               kunmap_atomic(kaddr, KM_USER0);
+       }
+       goto out;
+}
+
+static size_t __ntfs_copy_from_user_iovec(char *vaddr,
+               const struct iovec *iov, size_t iov_ofs, size_t bytes)
+{
+       size_t total = 0;
+
+       while (1) {
+               const char __user *buf = iov->iov_base + iov_ofs;
+               unsigned len;
+               size_t left;
+
+               len = iov->iov_len - iov_ofs;
+               if (len > bytes)
+                       len = bytes;
+               left = __copy_from_user_inatomic(vaddr, buf, len);
+               total += len;
+               bytes -= len;
+               vaddr += len;
+               if (unlikely(left)) {
+                       /*
+                        * Zero the rest of the target like __copy_from_user().
+                        */
+                       memset(vaddr, 0, bytes);
+                       total -= left;
+                       break;
+               }
+               if (!bytes)
+                       break;
+               iov++;
+               iov_ofs = 0;
+       }
+       return total;
+}
+
+static inline void ntfs_set_next_iovec(const struct iovec **iovp,
+               size_t *iov_ofsp, size_t bytes)
+{
+       const struct iovec *iov = *iovp;
+       size_t iov_ofs = *iov_ofsp;
+
+       while (bytes) {
+               unsigned len;
+
+               len = iov->iov_len - iov_ofs;
+               if (len > bytes)
+                       len = bytes;
+               bytes -= len;
+               iov_ofs += len;
+               if (iov->iov_len == iov_ofs) {
+                       iov++;
+                       iov_ofs = 0;
+               }
+       }
+       *iovp = iov;
+       *iov_ofsp = iov_ofs;
+}
+
+/*
+ * This has the same side-effects and return value as ntfs_copy_from_user().
+ * The difference is that on a fault we need to memset the remainder of the
+ * pages (out to offset + bytes), to emulate ntfs_copy_from_user()'s
+ * single-segment behaviour.
+ *
+ * We call the same helper (__ntfs_copy_from_user_iovec()) both when atomic and
+ * when not atomic.  This is ok because __ntfs_copy_from_user_iovec() calls
+ * __copy_from_user_inatomic() and it is ok to call this when non-atomic.  In
+ * fact, the only difference between __copy_from_user_inatomic() and
+ * __copy_from_user() is that the latter calls might_sleep().  And on many
+ * architectures __copy_from_user_inatomic() is just defined to
+ * __copy_from_user() so it makes no difference at all on those architectures.
+ */
+static inline size_t ntfs_copy_from_user_iovec(struct page **pages,
+               unsigned nr_pages, unsigned ofs, const struct iovec **iov,
+               size_t *iov_ofs, size_t bytes)
+{
+       struct page **last_page = pages + nr_pages;
+       char *kaddr;
+       size_t copied, len, total = 0;
+
+       do {
+               len = PAGE_CACHE_SIZE - ofs;
+               if (len > bytes)
+                       len = bytes;
+               kaddr = kmap_atomic(*pages, KM_USER0);
+               copied = __ntfs_copy_from_user_iovec(kaddr + ofs,
+                               *iov, *iov_ofs, len);
+               kunmap_atomic(kaddr, KM_USER0);
+               if (unlikely(copied != len)) {
+                       /* Do it the slow way. */
+                       kaddr = kmap(*pages);
+                       copied = __ntfs_copy_from_user_iovec(kaddr + ofs,
+                                       *iov, *iov_ofs, len);
+                       kunmap(*pages);
+                       if (unlikely(copied != len))
+                               goto err_out;
+               }
+               total += len;
+               bytes -= len;
+               if (!bytes)
+                       break;
+               ntfs_set_next_iovec(iov, iov_ofs, len);
+               ofs = 0;
+       } while (++pages < last_page);
+out:
+       return total;
+err_out:
+       total += copied;
+       /* Zero the rest of the target like __copy_from_user(). */
+       while (++pages < last_page) {
+               bytes -= len;
+               if (!bytes)
+                       break;
+               len = PAGE_CACHE_SIZE;
+               if (len > bytes)
+                       len = bytes;
+               kaddr = kmap_atomic(*pages, KM_USER0);
+               memset(kaddr, 0, len);
+               kunmap_atomic(kaddr, KM_USER0);
+       }
+       goto out;
+}
+
+static inline void ntfs_flush_dcache_pages(struct page **pages,
+               unsigned nr_pages)
+{
+       BUG_ON(!nr_pages);
+       do {
+               /*
+                * Warning: Do not do the decrement at the same time as the
+                * call because flush_dcache_page() is a NULL macro on i386
+                * and hence the decrement never happens.
+                */
+               flush_dcache_page(pages[nr_pages]);
+       } while (--nr_pages > 0);
+}
+
+/**
+ * ntfs_commit_pages_after_non_resident_write - commit the received data
+ * @pages:     array of destination pages
+ * @nr_pages:  number of pages in @pages
+ * @pos:       byte position in file at which the write begins
+ * @bytes:     number of bytes to be written
+ *
+ * See description of ntfs_commit_pages_after_write(), below.
+ */
+static inline int ntfs_commit_pages_after_non_resident_write(
+               struct page **pages, const unsigned nr_pages,
+               s64 pos, size_t bytes)
+{
+       s64 end, initialized_size;
+       struct inode *vi;
+       ntfs_inode *ni, *base_ni;
+       struct buffer_head *bh, *head;
+       ntfs_attr_search_ctx *ctx;
+       MFT_RECORD *m;
+       ATTR_RECORD *a;
+       unsigned long flags;
+       unsigned blocksize, u;
+       int err;
+
+       vi = pages[0]->mapping->host;
+       ni = NTFS_I(vi);
+       blocksize = 1 << vi->i_blkbits;
+       end = pos + bytes;
+       u = 0;
+       do {
+               s64 bh_pos;
+               struct page *page;
+               BOOL partial;
+
+               page = pages[u];
+               bh_pos = (s64)page->index << PAGE_CACHE_SHIFT;
+               bh = head = page_buffers(page);
+               partial = FALSE;
+               do {
+                       s64 bh_end;
+
+                       bh_end = bh_pos + blocksize;
+                       if (bh_end <= pos || bh_pos >= end) {
+                               if (!buffer_uptodate(bh))
+                                       partial = TRUE;
+                       } else {
+                               set_buffer_uptodate(bh);
+                               mark_buffer_dirty(bh);
+                       }
+               } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
+               /*
+                * If all buffers are now uptodate but the page is not, set the
+                * page uptodate.
+                */
+               if (!partial && !PageUptodate(page))
+                       SetPageUptodate(page);
+       } while (++u < nr_pages);
+       /*
+        * Finally, if we do not need to update initialized_size or i_size we
+        * are finished.
+        */
+       read_lock_irqsave(&ni->size_lock, flags);
+       initialized_size = ni->initialized_size;
+       read_unlock_irqrestore(&ni->size_lock, flags);
+       if (end <= initialized_size) {
+               ntfs_debug("Done.");
+               return 0;
+       }
+       /*
+        * Update initialized_size/i_size as appropriate, both in the inode and
+        * the mft record.
+        */
+       if (!NInoAttr(ni))
+               base_ni = ni;
+       else
+               base_ni = ni->ext.base_ntfs_ino;
+       /* Map, pin, and lock the mft record. */
+       m = map_mft_record(base_ni);
+       if (IS_ERR(m)) {
+               err = PTR_ERR(m);
+               m = NULL;
+               ctx = NULL;
+               goto err_out;
+       }
+       BUG_ON(!NInoNonResident(ni));
+       ctx = ntfs_attr_get_search_ctx(base_ni, m);
+       if (unlikely(!ctx)) {
+               err = -ENOMEM;
+               goto err_out;
+       }
+       err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
+                       CASE_SENSITIVE, 0, NULL, 0, ctx);
+       if (unlikely(err)) {
+               if (err == -ENOENT)
+                       err = -EIO;
+               goto err_out;
+       }
+       a = ctx->attr;
+       BUG_ON(!a->non_resident);
+       write_lock_irqsave(&ni->size_lock, flags);
+       BUG_ON(end > ni->allocated_size);
+       ni->initialized_size = end;
+       a->data.non_resident.initialized_size = cpu_to_sle64(end);
+       if (end > i_size_read(vi)) {
+               i_size_write(vi, end);
+               a->data.non_resident.data_size =
+                               a->data.non_resident.initialized_size;
+       }
+       write_unlock_irqrestore(&ni->size_lock, flags);
+       /* Mark the mft record dirty, so it gets written back. */
+       flush_dcache_mft_record_page(ctx->ntfs_ino);
+       mark_mft_record_dirty(ctx->ntfs_ino);
+       ntfs_attr_put_search_ctx(ctx);
+       unmap_mft_record(base_ni);
+       ntfs_debug("Done.");
+       return 0;
+err_out:
+       if (ctx)
+               ntfs_attr_put_search_ctx(ctx);
+       if (m)
+               unmap_mft_record(base_ni);
+       ntfs_error(vi->i_sb, "Failed to update initialized_size/i_size (error "
+                       "code %i).", err);
+       if (err != -ENOMEM) {
+               NVolSetErrors(ni->vol);
+               make_bad_inode(VFS_I(base_ni));
+               make_bad_inode(vi);
+       }
+       return err;
+}
+
+/**
+ * ntfs_commit_pages_after_write - commit the received data
+ * @pages:     array of destination pages
+ * @nr_pages:  number of pages in @pages
+ * @pos:       byte position in file at which the write begins
+ * @bytes:     number of bytes to be written
+ *
+ * This is called from ntfs_file_buffered_write() with i_sem held on the inode
+ * (@pages[0]->mapping->host).  There are @nr_pages pages in @pages which are
+ * locked but not kmap()ped.  The source data has already been copied into the
+ * @page.  ntfs_prepare_pages_for_non_resident_write() has been called before
+ * the data was copied (for non-resident attributes only) and it returned
+ * success.
+ *
+ * Need to set uptodate and mark dirty all buffers within the boundary of the
+ * write.  If all buffers in a page are uptodate we set the page uptodate, too.
+ *
+ * Setting the buffers dirty ensures that they get written out later when
+ * ntfs_writepage() is invoked by the VM.
+ *
+ * Finally, we need to update i_size and initialized_size as appropriate both
+ * in the inode and the mft record.
+ *
+ * This is modelled after fs/buffer.c::generic_commit_write(), which marks
+ * buffers uptodate and dirty, sets the page uptodate if all buffers in the
+ * page are uptodate, and updates i_size if the end of io is beyond i_size.  In
+ * that case, it also marks the inode dirty.
+ *
+ * If things have gone as outlined in
+ * ntfs_prepare_pages_for_non_resident_write(), we do not need to do any page
+ * content modifications here for non-resident attributes.  For resident
+ * attributes we need to do the uptodate bringing here which we combine with
+ * the copying into the mft record which means we save one atomic kmap.
+ *
+ * Return 0 on success or -errno on error.
+ */
+static int ntfs_commit_pages_after_write(struct page **pages,
+               const unsigned nr_pages, s64 pos, size_t bytes)
+{
+       s64 end, initialized_size;
+       loff_t i_size;
+       struct inode *vi;
+       ntfs_inode *ni, *base_ni;
+       struct page *page;
+       ntfs_attr_search_ctx *ctx;
+       MFT_RECORD *m;
+       ATTR_RECORD *a;
+       char *kattr, *kaddr;
+       unsigned long flags;
+       u32 attr_len;
+       int err;
+
+       BUG_ON(!nr_pages);
+       BUG_ON(!pages);
+       page = pages[0];
+       BUG_ON(!page);
+       vi = page->mapping->host;
+       ni = NTFS_I(vi);
+       ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, start page "
+                       "index 0x%lx, nr_pages 0x%x, pos 0x%llx, bytes 0x%x.",
+                       vi->i_ino, ni->type, page->index, nr_pages,
+                       (long long)pos, bytes);
+       if (NInoNonResident(ni))
+               return ntfs_commit_pages_after_non_resident_write(pages,
+                               nr_pages, pos, bytes);
+       BUG_ON(nr_pages > 1);
+       /*
+        * Attribute is resident, implying it is not compressed, encrypted, or
+        * sparse.
+        */
+       if (!NInoAttr(ni))
+               base_ni = ni;
+       else
+               base_ni = ni->ext.base_ntfs_ino;
+       BUG_ON(NInoNonResident(ni));
+       /* Map, pin, and lock the mft record. */
+       m = map_mft_record(base_ni);
+       if (IS_ERR(m)) {
+               err = PTR_ERR(m);
+               m = NULL;
+               ctx = NULL;
+               goto err_out;
+       }
+       ctx = ntfs_attr_get_search_ctx(base_ni, m);
+       if (unlikely(!ctx)) {
+               err = -ENOMEM;
+               goto err_out;
+       }
+       err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
+                       CASE_SENSITIVE, 0, NULL, 0, ctx);
+       if (unlikely(err)) {
+               if (err == -ENOENT)
+                       err = -EIO;
+               goto err_out;
+       }
+       a = ctx->attr;
+       BUG_ON(a->non_resident);
+       /* The total length of the attribute value. */
+       attr_len = le32_to_cpu(a->data.resident.value_length);
+       i_size = i_size_read(vi);
+       BUG_ON(attr_len != i_size);
+       BUG_ON(pos > attr_len);
+       end = pos + bytes;
+       BUG_ON(end > le32_to_cpu(a->length) -
+                       le16_to_cpu(a->data.resident.value_offset));
+       kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
+       kaddr = kmap_atomic(page, KM_USER0);
+       /* Copy the received data from the page to the mft record. */
+       memcpy(kattr + pos, kaddr + pos, bytes);
+       /* Update the attribute length if necessary. */
+       if (end > attr_len) {
+               attr_len = end;
+               a->data.resident.value_length = cpu_to_le32(attr_len);
+       }
+       /*
+        * If the page is not uptodate, bring the out of bounds area(s)
+        * uptodate by copying data from the mft record to the page.
+        */
+       if (!PageUptodate(page)) {
+               if (pos > 0)
+                       memcpy(kaddr, kattr, pos);
+               if (end < attr_len)
+                       memcpy(kaddr + end, kattr + end, attr_len - end);
+               /* Zero the region outside the end of the attribute value. */
+               memset(kaddr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
+               flush_dcache_page(page);
+               SetPageUptodate(page);
+       }
+       kunmap_atomic(kaddr, KM_USER0);
+       /* Update initialized_size/i_size if necessary. */
+       read_lock_irqsave(&ni->size_lock, flags);
+       initialized_size = ni->initialized_size;
+       BUG_ON(end > ni->allocated_size);
+       read_unlock_irqrestore(&ni->size_lock, flags);
+       BUG_ON(initialized_size != i_size);
+       if (end > initialized_size) {
+               unsigned long flags;
+
+               write_lock_irqsave(&ni->size_lock, flags);
+               ni->initialized_size = end;
+               i_size_write(vi, end);
+               write_unlock_irqrestore(&ni->size_lock, flags);
+       }
+       /* Mark the mft record dirty, so it gets written back. */
+       flush_dcache_mft_record_page(ctx->ntfs_ino);
+       mark_mft_record_dirty(ctx->ntfs_ino);
+       ntfs_attr_put_search_ctx(ctx);
+       unmap_mft_record(base_ni);
+       ntfs_debug("Done.");
+       return 0;
+err_out:
+       if (err == -ENOMEM) {
+               ntfs_warning(vi->i_sb, "Error allocating memory required to "
+                               "commit the write.");
+               if (PageUptodate(page)) {
+                       ntfs_warning(vi->i_sb, "Page is uptodate, setting "
+                                       "dirty so the write will be retried "
+                                       "later on by the VM.");
+                       /*
+                        * Put the page on mapping->dirty_pages, but leave its
+                        * buffers' dirty state as-is.
+                        */
+                       __set_page_dirty_nobuffers(page);
+                       err = 0;
+               } else
+                       ntfs_error(vi->i_sb, "Page is not uptodate.  Written "
+                                       "data has been lost.");
+       } else {
+               ntfs_error(vi->i_sb, "Resident attribute commit write failed "
+                               "with error %i.", err);
+               NVolSetErrors(ni->vol);
+               make_bad_inode(VFS_I(base_ni));
+               make_bad_inode(vi);
+       }
+       if (ctx)
+               ntfs_attr_put_search_ctx(ctx);
+       if (m)
+               unmap_mft_record(base_ni);
+       return err;
+}
+
+/**
+ * ntfs_file_buffered_write -
+ *
+ * Locking: The vfs is holding ->i_sem on the inode.
+ */
+static ssize_t ntfs_file_buffered_write(struct kiocb *iocb,
+               const struct iovec *iov, unsigned long nr_segs,
+               loff_t pos, loff_t *ppos, size_t count)
+{
+       struct file *file = iocb->ki_filp;
+       struct address_space *mapping = file->f_mapping;
+       struct inode *vi = mapping->host;
+       ntfs_inode *ni = NTFS_I(vi);
+       ntfs_volume *vol = ni->vol;
+       struct page *pages[NTFS_MAX_PAGES_PER_CLUSTER];
+       struct page *cached_page = NULL;
+       char __user *buf = NULL;
+       s64 end, ll;
+       VCN last_vcn;
+       LCN lcn;
+       unsigned long flags;
+       size_t bytes, iov_ofs;
+       ssize_t status, written;
+       unsigned nr_pages;
+       int err;
+       struct pagevec lru_pvec;
+
+       ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, "
+                       "pos 0x%llx, count 0x%lx.",
+                       vi->i_ino, (unsigned)le32_to_cpu(ni->type),
+                       (unsigned long long)pos, (unsigned long)count);
+       if (unlikely(!count))
+               return 0;
+       BUG_ON(NInoMstProtected(ni));
+       /*
+        * If the attribute is not an index root and it is encrypted or
+        * compressed, we cannot write to it yet.  Note we need to check for
+        * AT_INDEX_ALLOCATION since this is the type of both directory and
+        * index inodes.
+        */
+       if (ni->type != AT_INDEX_ALLOCATION) {
+               /* If file is encrypted, deny access, just like NT4. */
+               if (NInoEncrypted(ni)) {
+                       ntfs_debug("Denying write access to encrypted file.");
+                       return -EACCES;
+               }
+               if (NInoCompressed(ni)) {
+                       ntfs_error(vi->i_sb, "Writing to compressed files is "
+                                       "not implemented yet.  Sorry.");
+                       return -EOPNOTSUPP;
+               }
+       }
+       /*
+        * If a previous ntfs_truncate() failed, repeat it and abort if it
+        * fails again.
+        */
+       if (unlikely(NInoTruncateFailed(ni))) {
+               down_write(&vi->i_alloc_sem);
+               err = ntfs_truncate(vi);
+               up_write(&vi->i_alloc_sem);
+               if (err || NInoTruncateFailed(ni)) {
+                       if (!err)
+                               err = -EIO;
+                       ntfs_error(vol->sb, "Cannot perform write to inode "
+                                       "0x%lx, attribute type 0x%x, because "
+                                       "ntfs_truncate() failed (error code "
+                                       "%i).", vi->i_ino,
+                                       (unsigned)le32_to_cpu(ni->type), err);
+                       return err;
+               }
+       }
+       /* The first byte after the write. */
+       end = pos + count;
+       /*
+        * If the write goes beyond the allocated size, extend the allocation
+        * to cover the whole of the write, rounded up to the nearest cluster.
+        */
+       read_lock_irqsave(&ni->size_lock, flags);
+       ll = ni->allocated_size;
+       read_unlock_irqrestore(&ni->size_lock, flags);
+       if (end > ll) {
+               /* Extend the allocation without changing the data size. */
+               ll = ntfs_attr_extend_allocation(ni, end, -1, pos);
+               if (likely(ll >= 0)) {
+                       BUG_ON(pos >= ll);
+                       /* If the extension was partial truncate the write. */
+                       if (end > ll) {
+                               ntfs_debug("Truncating write to inode 0x%lx, "
+                                               "attribute type 0x%x, because "
+                                               "the allocation was only "
+                                               "partially extended.",
+                                               vi->i_ino, (unsigned)
+                                               le32_to_cpu(ni->type));
+                               end = ll;
+                               count = ll - pos;
+                       }
+               } else {
+                       err = ll;
+                       read_lock_irqsave(&ni->size_lock, flags);
+                       ll = ni->allocated_size;
+                       read_unlock_irqrestore(&ni->size_lock, flags);
+                       /* Perform a partial write if possible or fail. */
+                       if (pos < ll) {
+                               ntfs_debug("Truncating write to inode 0x%lx, "
+                                               "attribute type 0x%x, because "
+                                               "extending the allocation "
+                                               "failed (error code %i).",
+                                               vi->i_ino, (unsigned)
+                                               le32_to_cpu(ni->type), err);
+                               end = ll;
+                               count = ll - pos;
+                       } else {
+                               ntfs_error(vol->sb, "Cannot perform write to "
+                                               "inode 0x%lx, attribute type "
+                                               "0x%x, because extending the "
+                                               "allocation failed (error "
+                                               "code %i).", vi->i_ino,
+                                               (unsigned)
+                                               le32_to_cpu(ni->type), err);
+                               return err;
+                       }
+               }
+       }
+       pagevec_init(&lru_pvec, 0);
+       written = 0;
+       /*
+        * If the write starts beyond the initialized size, extend it up to the
+        * beginning of the write and initialize all non-sparse space between
+        * the old initialized size and the new one.  This automatically also
+        * increments the vfs inode->i_size to keep it above or equal to the
+        * initialized_size.
+        */
+       read_lock_irqsave(&ni->size_lock, flags);
+       ll = ni->initialized_size;
+       read_unlock_irqrestore(&ni->size_lock, flags);
+       if (pos > ll) {
+               err = ntfs_attr_extend_initialized(ni, pos, &cached_page,
+                               &lru_pvec);
+               if (err < 0) {
+                       ntfs_error(vol->sb, "Cannot perform write to inode "
+                                       "0x%lx, attribute type 0x%x, because "
+                                       "extending the initialized size "
+                                       "failed (error code %i).", vi->i_ino,
+                                       (unsigned)le32_to_cpu(ni->type), err);
+                       status = err;
+                       goto err_out;
+               }
+       }
+       /*
+        * Determine the number of pages per cluster for non-resident
+        * attributes.
+        */
+       nr_pages = 1;
+       if (vol->cluster_size > PAGE_CACHE_SIZE && NInoNonResident(ni))
+               nr_pages = vol->cluster_size >> PAGE_CACHE_SHIFT;
+       /* Finally, perform the actual write. */
+       last_vcn = -1;
+       if (likely(nr_segs == 1))
+               buf = iov->iov_base;
+       else
+               iov_ofs = 0;    /* Offset in the current iovec. */
+       do {
+               VCN vcn;
+               pgoff_t idx, start_idx;
+               unsigned ofs, do_pages, u;
+               size_t copied;
+
+               start_idx = idx = pos >> PAGE_CACHE_SHIFT;
+               ofs = pos & ~PAGE_CACHE_MASK;
+               bytes = PAGE_CACHE_SIZE - ofs;
+               do_pages = 1;
+               if (nr_pages > 1) {
+                       vcn = pos >> vol->cluster_size_bits;
+                       if (vcn != last_vcn) {
+                               last_vcn = vcn;
+                               /*
+                                * Get the lcn of the vcn the write is in.  If
+                                * it is a hole, need to lock down all pages in
+                                * the cluster.
+                                */
+                               down_read(&ni->runlist.lock);
+                               lcn = ntfs_attr_vcn_to_lcn_nolock(ni, pos >>
+                                               vol->cluster_size_bits, FALSE);
+                               up_read(&ni->runlist.lock);
+                               if (unlikely(lcn < LCN_HOLE)) {
+                                       status = -EIO;
+                                       if (lcn == LCN_ENOMEM)
+                                               status = -ENOMEM;
+                                       else
+                                               ntfs_error(vol->sb, "Cannot "
+                                                       "perform write to "
+                                                       "inode 0x%lx, "
+                                                       "attribute type 0x%x, "
+                                                       "because the attribute "
+                                                       "is corrupt.",
+                                                       vi->i_ino, (unsigned)
+                                                       le32_to_cpu(ni->type));
+                                       break;
+                               }
+                               if (lcn == LCN_HOLE) {
+                                       start_idx = (pos & ~(s64)
+                                                       vol->cluster_size_mask)
+                                                       >> PAGE_CACHE_SHIFT;
+                                       bytes = vol->cluster_size - (pos &
+                                                       vol->cluster_size_mask);
+                                       do_pages = nr_pages;
+                               }
+                       }
+               }
+               if (bytes > count)
+                       bytes = count;
+               /*
+                * Bring in the user page(s) that we will copy from _first_.
+                * Otherwise there is a nasty deadlock on copying from the same
+                * page(s) as we are writing to, without it/them being marked
+                * up-to-date.  Note, at present there is nothing to stop the
+                * pages being swapped out between us bringing them into memory
+                * and doing the actual copying.
+                */
+               if (likely(nr_segs == 1))
+                       ntfs_fault_in_pages_readable(buf, bytes);
+               else
+                       ntfs_fault_in_pages_readable_iovec(iov, iov_ofs, bytes);
+               /* Get and lock @do_pages starting at index @start_idx. */
+               status = __ntfs_grab_cache_pages(mapping, start_idx, do_pages,
+                               pages, &cached_page, &lru_pvec);
+               if (unlikely(status))
+                       break;
+               /*
+                * For non-resident attributes, we need to fill any holes with
+                * actual clusters and ensure all bufferes are mapped.  We also
+                * need to bring uptodate any buffers that are only partially
+                * being written to.
+                */
+               if (NInoNonResident(ni)) {
+                       status = ntfs_prepare_pages_for_non_resident_write(
+                                       pages, do_pages, pos, bytes);
+                       if (unlikely(status)) {
+                               loff_t i_size;
+
+                               do {
+                                       unlock_page(pages[--do_pages]);
+                                       page_cache_release(pages[do_pages]);
+                               } while (do_pages);
+                               /*
+                                * The write preparation may have instantiated
+                                * allocated space outside i_size.  Trim this
+                                * off again.  We can ignore any errors in this
+                                * case as we will just be waisting a bit of
+                                * allocated space, which is not a disaster.
+                                */
+                               i_size = i_size_read(vi);
+                               if (pos + bytes > i_size)
+                                       vmtruncate(vi, i_size);
+                               break;
+                       }
+               }
+               u = (pos >> PAGE_CACHE_SHIFT) - pages[0]->index;
+               if (likely(nr_segs == 1)) {
+                       copied = ntfs_copy_from_user(pages + u, do_pages - u,
+                                       ofs, buf, bytes);
+                       buf += copied;
+               } else
+                       copied = ntfs_copy_from_user_iovec(pages + u,
+                                       do_pages - u, ofs, &iov, &iov_ofs,
+                                       bytes);
+               ntfs_flush_dcache_pages(pages + u, do_pages - u);
+               status = ntfs_commit_pages_after_write(pages, do_pages, pos,
+                               bytes);
+               if (likely(!status)) {
+                       written += copied;
+                       count -= copied;
+                       pos += copied;
+                       if (unlikely(copied != bytes))
+                               status = -EFAULT;
+               }
+               do {
+                       unlock_page(pages[--do_pages]);
+                       mark_page_accessed(pages[do_pages]);
+                       page_cache_release(pages[do_pages]);
+               } while (do_pages);
+               if (unlikely(status))
+                       break;
+               balance_dirty_pages_ratelimited(mapping);
+               cond_resched();
+       } while (count);
+err_out:
+       *ppos = pos;
+       if (cached_page)
+               page_cache_release(cached_page);
+       /* For now, when the user asks for O_SYNC, we actually give O_DSYNC. */
+       if (likely(!status)) {
+               if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(vi))) {
+                       if (!mapping->a_ops->writepage || !is_sync_kiocb(iocb))
+                               status = generic_osync_inode(vi, mapping,
+                                               OSYNC_METADATA|OSYNC_DATA);
+               }
+       }
+       pagevec_lru_add(&lru_pvec);
+       ntfs_debug("Done.  Returning %s (written 0x%lx, status %li).",
+                       written ? "written" : "status", (unsigned long)written,
+                       (long)status);
+       return written ? written : status;
+}
+
+/**
+ * ntfs_file_aio_write_nolock -
+ */
+static ssize_t ntfs_file_aio_write_nolock(struct kiocb *iocb,
+               const struct iovec *iov, unsigned long nr_segs, loff_t *ppos)
+{
+       struct file *file = iocb->ki_filp;
+       struct address_space *mapping = file->f_mapping;
+       struct inode *inode = mapping->host;
+       loff_t pos;
+       unsigned long seg;
+       size_t count;           /* after file limit checks */
+       ssize_t written, err;
+
+       count = 0;
+       for (seg = 0; seg < nr_segs; seg++) {
+               const struct iovec *iv = &iov[seg];
+               /*
+                * If any segment has a negative length, or the cumulative
+                * length ever wraps negative then return -EINVAL.
+                */
+               count += iv->iov_len;
+               if (unlikely((ssize_t)(count|iv->iov_len) < 0))
+                       return -EINVAL;
+               if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
+                       continue;
+               if (!seg)
+                       return -EFAULT;
+               nr_segs = seg;
+               count -= iv->iov_len;   /* This segment is no good */
+               break;
+       }
+       pos = *ppos;
+       vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
+       /* We can write back this queue in page reclaim. */
+       current->backing_dev_info = mapping->backing_dev_info;
+       written = 0;
+       err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
+       if (err)
+               goto out;
+       if (!count)
+               goto out;
+       err = remove_suid(file->f_dentry);
+       if (err)
+               goto out;
+       inode_update_time(inode, 1);
+       written = ntfs_file_buffered_write(iocb, iov, nr_segs, pos, ppos,
+                       count);
+out:
+       current->backing_dev_info = NULL;
+       return written ? written : err;
+}
+
+/**
+ * ntfs_file_aio_write -
+ */
+static ssize_t ntfs_file_aio_write(struct kiocb *iocb, const char __user *buf,
+               size_t count, loff_t pos)
+{
+       struct file *file = iocb->ki_filp;
+       struct address_space *mapping = file->f_mapping;
+       struct inode *inode = mapping->host;
+       ssize_t ret;
+       struct iovec local_iov = { .iov_base = (void __user *)buf,
+                                  .iov_len = count };
+
+       BUG_ON(iocb->ki_pos != pos);
+
+       down(&inode->i_sem);
+       ret = ntfs_file_aio_write_nolock(iocb, &local_iov, 1, &iocb->ki_pos);
+       up(&inode->i_sem);
+       if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
+               int err = sync_page_range(inode, mapping, pos, ret);
+               if (err < 0)
+                       ret = err;
+       }
+       return ret;
+}
+
+/**
+ * ntfs_file_writev -
+ *
+ * Basically the same as generic_file_writev() except that it ends up calling
+ * ntfs_file_aio_write_nolock() instead of __generic_file_aio_write_nolock().
+ */
+static ssize_t ntfs_file_writev(struct file *file, const struct iovec *iov,
+               unsigned long nr_segs, loff_t *ppos)
+{
+       struct address_space *mapping = file->f_mapping;
+       struct inode *inode = mapping->host;
+       struct kiocb kiocb;
+       ssize_t ret;
+
+       down(&inode->i_sem);
+       init_sync_kiocb(&kiocb, file);
+       ret = ntfs_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
+       if (ret == -EIOCBQUEUED)
+               ret = wait_on_sync_kiocb(&kiocb);
+       up(&inode->i_sem);
+       if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
+               int err = sync_page_range(inode, mapping, *ppos - ret, ret);
+               if (err < 0)
+                       ret = err;
+       }
+       return ret;
+}
+
+/**
+ * ntfs_file_write - simple wrapper for ntfs_file_writev()
+ */
+static ssize_t ntfs_file_write(struct file *file, const char __user *buf,
+               size_t count, loff_t *ppos)
+{
+       struct iovec local_iov = { .iov_base = (void __user *)buf,
+                                  .iov_len = count };
+
+       return ntfs_file_writev(file, &local_iov, 1, ppos);
+}
+
+/**
  * ntfs_file_fsync - sync a file to disk
  * @filp:      file to be synced
  * @dentry:    dentry describing the file to sync
@@ -113,39 +2296,39 @@ static int ntfs_file_fsync(struct file *filp, struct dentry *dentry,
 #endif /* NTFS_RW */
 
 struct file_operations ntfs_file_ops = {
-       .llseek         = generic_file_llseek,    /* Seek inside file. */
-       .read           = generic_file_read,      /* Read from file. */
-       .aio_read       = generic_file_aio_read,  /* Async read from file. */
-       .readv          = generic_file_readv,     /* Read from file. */
+       .llseek         = generic_file_llseek,   /* Seek inside file. */
+       .read           = generic_file_read,     /* Read from file. */
+       .aio_read       = generic_file_aio_read, /* Async read from file. */
+       .readv          = generic_file_readv,    /* Read from file. */
 #ifdef NTFS_RW
-       .write          = generic_file_write,     /* Write to file. */
-       .aio_write      = generic_file_aio_write, /* Async write to file. */
-       .writev         = generic_file_writev,    /* Write to file. */
-       /*.release      = ,*/                     /* Last file is closed.  See
-                                                    fs/ext2/file.c::
-                                                    ext2_release_file() for
-                                                    how to use this to discard
-                                                    preallocated space for
-                                                    write opened files. */
-       .fsync          = ntfs_file_fsync,        /* Sync a file to disk. */
-       /*.aio_fsync    = ,*/                     /* Sync all outstanding async
-                                                    i/o operations on a
-                                                    kiocb. */
+       .write          = ntfs_file_write,       /* Write to file. */
+       .aio_write      = ntfs_file_aio_write,   /* Async write to file. */
+       .writev         = ntfs_file_writev,      /* Write to file. */
+       /*.release      = ,*/                    /* Last file is closed.  See
+                                                   fs/ext2/file.c::
+                                                   ext2_release_file() for
+                                                   how to use this to discard
+                                                   preallocated space for
+                                                   write opened files. */
+       .fsync          = ntfs_file_fsync,       /* Sync a file to disk. */
+       /*.aio_fsync    = ,*/                    /* Sync all outstanding async
+                                                   i/o operations on a
+                                                   kiocb. */
 #endif /* NTFS_RW */
-       /*.ioctl        = ,*/                     /* Perform function on the
-                                                    mounted filesystem. */
-       .mmap           = generic_file_mmap,      /* Mmap file. */
-       .open           = ntfs_file_open,         /* Open file. */
-       .sendfile       = generic_file_sendfile,  /* Zero-copy data send with
-                                                    the data source being on
-                                                    the ntfs partition.  We
-                                                    do not need to care about
-                                                    the data destination. */
-       /*.sendpage     = ,*/                     /* Zero-copy data send with
-                                                    the data destination being
-                                                    on the ntfs partition.  We
-                                                    do not need to care about
-                                                    the data source. */
+       /*.ioctl        = ,*/                    /* Perform function on the
+                                                   mounted filesystem. */
+       .mmap           = generic_file_mmap,     /* Mmap file. */
+       .open           = ntfs_file_open,        /* Open file. */
+       .sendfile       = generic_file_sendfile, /* Zero-copy data send with
+                                                   the data source being on
+                                                   the ntfs partition.  We do
+                                                   not need to care about the
+                                                   data destination. */
+       /*.sendpage     = ,*/                    /* Zero-copy data send with
+                                                   the data destination being
+                                                   on the ntfs partition.  We
+                                                   do not need to care about
+                                                   the data source. */
 };
 
 struct inode_operations ntfs_file_inode_ops = {