Merge git://git.kernel.org/pub/scm/linux/kernel/git/aia21/ntfs-2.6

14 files changed, 3841 insertions, 1043 deletions

diff --git a/Documentation/filesystems/ntfs.txt b/Documentation/filesystems/ntfs.txt
index a5fbc8e897fa..614de3124901 100644
--- a/Documentation/filesystems/ntfs.txt
+++ b/Documentation/filesystems/ntfs.txt
@@ -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.
diff --git a/fs/ntfs/ChangeLog b/fs/ntfs/ChangeLog
index de58579a1d0e..50a7749cfca1 100644
--- a/fs/ntfs/ChangeLog
+++ b/fs/ntfs/ChangeLog
@@ -1,18 +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 truncate for both of which i_sem is
-	  held.  Just have to be careful in readpage/writepage and all other
-	  helpers not running under i_sem that we play nice...
-	  Also need to be careful with initialized_size extention in
-	  ntfs_prepare_write. Basically, just be _very_ careful in this code...
-	  UPDATE: The only things that need to be checked are read/writepage
-	  which do not hold i_sem.  Note writepage cannot change i_size but it
-	  needs to cope with a concurrent i_size change, just like readpage.
-	  Also both need to cope with concurrent changes to the other sizes,
-	  i.e. initialized/allocated/compressed size, as well.
+	- 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 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
@@ -22,6 +19,68 @@ ToDo/Notes:
 	- Enable the code for setting the NT4 compatibility flag when we start
 	  making NTFS 1.2 specific modifications.
 
+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
+	  context as argument.  This allows calling these functions with the
+	  mft record mapped.  Update all callers.
+	- Fix potential deadlock in ntfs_mft_data_extend_allocation_nolock()
+	  error handling by passing in the active search context when calling
+	  ntfs_cluster_free().
+	- Change ntfs_cluster_alloc() to take an extra boolean parameter
+	  specifying whether the cluster are being allocated to extend an
+	  attribute or to fill a hole.
+	- Change ntfs_attr_make_non_resident() to call ntfs_cluster_alloc()
+	  with @is_extension set to TRUE and remove the runlist terminator
+	  fixup code as this is now done by ntfs_cluster_alloc().
+	- Change ntfs_attr_make_non_resident to take the attribute value size
+	  as an extra parameter.  This is needed since we need to know the size
+	  before we can map the mft record and our callers always know it.  The
+	  reason we cannot simply read the size from the vfs inode i_size is
+	  that this is not necessarily uptodate.  This happens when
+	  ntfs_attr_make_non_resident() is called in the ->truncate call path.
+	- Fix ntfs_attr_make_non_resident() to update the vfs inode i_blocks
+	  which is zero for a resident attribute but should no longer be zero
+	  once the attribute is non-resident as it then has real clusters
+	  allocated.
+	- Add fs/ntfs/attrib.[hc]::ntfs_attr_extend_allocation(), a function to
+	  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 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
+	  will resize a file appropriately.  The limitations are that only
+	  uncompressed and unencrypted files are supported.  Also, there is
+	  only very limited support for highly fragmented files (the ones whose
+	  $DATA attribute is split into multiple attribute extents).
+	- In attrib.c::ntfs_attr_set() call balance_dirty_pages_ratelimited()
+	  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.
+	- $EA attributes can be both resident and non-resident.
+	- Use %z for size_t to fix compilation warnings.  (Andrew Morton)
+	- Fix compilation warnings with gcc-4.0.2 on SUSE 10.0.
+	- Document extended attribute ($EA) NEED_EA flag.  (Based on libntfs
+	  patch by Yura Pakhuchiy.)
+
 2.1.24 - Lots of bug fixes and support more clean journal states.
 
 	- Support journals ($LogFile) which have been modified by chkdsk.  This
diff --git a/fs/ntfs/Makefile b/fs/ntfs/Makefile
index 894b2b876d35..d0d45d1c853a 100644
--- a/fs/ntfs/Makefile
+++ b/fs/ntfs/Makefile
@@ -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=\"2.1.24\"
+EXTRA_CFLAGS = -DNTFS_VERSION=\"2.1.25\"
 
 ifeq ($(CONFIG_NTFS_DEBUG),y)
 EXTRA_CFLAGS += -DDEBUG
diff --git a/fs/ntfs/aops.c b/fs/ntfs/aops.c
index 5e80c07c6a4d..1c0a4315876a 100644
--- a/fs/ntfs/aops.c
+++ b/fs/ntfs/aops.c
@@ -1391,8 +1391,7 @@ retry_writepage:
 		if (NInoEncrypted(ni)) {
 			unlock_page(page);
 			BUG_ON(ni->type != AT_DATA);
-			ntfs_debug("Denying write access to encrypted "
-					"file.");
+			ntfs_debug("Denying write access to encrypted file.");
 			return -EACCES;
 		}
 		/* Compressed data streams are handled in compress.c. */
@@ -1508,8 +1507,8 @@ retry_writepage:
 	/* Zero out of bounds area in the page cache page. */
 	memset(kaddr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
 	kunmap_atomic(kaddr, KM_USER0);
-	flush_dcache_mft_record_page(ctx->ntfs_ino);
 	flush_dcache_page(page);
+	flush_dcache_mft_record_page(ctx->ntfs_ino);
 	/* We are done with the page. */
 	end_page_writeback(page);
 	/* Finally, mark the mft record dirty, so it gets written back. */
@@ -1542,830 +1541,6 @@ err_out:
 	return err;
 }
 
-/**
- * ntfs_prepare_nonresident_write -
- *
- */
-static int ntfs_prepare_nonresident_write(struct page *page,
-		unsigned from, unsigned to)
-{
-	VCN vcn;
-	LCN lcn;
-	s64 initialized_size;
-	loff_t i_size;
-	sector_t block, ablock, iblock;
-	struct inode *vi;
-	ntfs_inode *ni;
-	ntfs_volume *vol;
-	runlist_element *rl;
-	struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
-	unsigned long flags;
-	unsigned int vcn_ofs, block_start, block_end, blocksize;
-	int err;
-	BOOL is_retry;
-	unsigned char blocksize_bits;
-
-	vi = page->mapping->host;
-	ni = NTFS_I(vi);
-	vol = ni->vol;
-
-	ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
-			"0x%lx, from = %u, to = %u.", ni->mft_no, ni->type,
-			page->index, from, to);
-
-	BUG_ON(!NInoNonResident(ni));
-
-	blocksize_bits = vi->i_blkbits;
-	blocksize = 1 << blocksize_bits;
-
-	/*
-	 * 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);
-	bh = head = page_buffers(page);
-	if (unlikely(!bh))
-		return -ENOMEM;
-
-	/* The first block in the page. */
-	block = (s64)page->index << (PAGE_CACHE_SHIFT - blocksize_bits);
-
-	read_lock_irqsave(&ni->size_lock, flags);
-	/*
-	 * The first out of bounds block for the allocated size.  No need to
-	 * round up as allocated_size is in multiples of cluster size and the
-	 * minimum cluster size is 512 bytes, which is equal to the smallest
-	 * blocksize.
-	 */
-	ablock = ni->allocated_size >> blocksize_bits;
-	i_size = i_size_read(vi);
-	initialized_size = ni->initialized_size;
-	read_unlock_irqrestore(&ni->size_lock, flags);
-
-	/* The last (fully or partially) initialized block. */
-	iblock = initialized_size >> blocksize_bits;
-
-	/* Loop through all the buffers in the page. */
-	block_start = 0;
-	rl = NULL;
-	err = 0;
-	do {
-		block_end = block_start + blocksize;
-		/*
-		 * If buffer @bh is outside the write, just mark it uptodate
-		 * if the page is uptodate and continue with the next buffer.
-		 */
-		if (block_end <= from || block_start >= to) {
-			if (PageUptodate(page)) {
-				if (!buffer_uptodate(bh))
-					set_buffer_uptodate(bh);
-			}
-			continue;
-		}
-		/*
-		 * @bh is at least partially being written to.
-		 * Make sure it is not marked as new.
-		 */
-		//if (buffer_new(bh))
-		//	clear_buffer_new(bh);
-
-		if (block >= ablock) {
-			// TODO: block is above allocated_size, need to
-			// allocate it. Best done in one go to accommodate not
-			// only block but all above blocks up to and including:
-			// ((page->index << PAGE_CACHE_SHIFT) + to + blocksize
-			// - 1) >> blobksize_bits. Obviously will need to round
-			// up to next cluster boundary, too. This should be
-			// done with a helper function, so it can be reused.
-			ntfs_error(vol->sb, "Writing beyond allocated size "
-					"is not supported yet. Sorry.");
-			err = -EOPNOTSUPP;
-			goto err_out;
-			// Need to update ablock.
-			// Need to set_buffer_new() on all block bhs that are
-			// newly allocated.
-		}
-		/*
-		 * Now we have enough allocated size to fulfill the whole
-		 * request, i.e. block < ablock is true.
-		 */
-		if (unlikely((block >= iblock) &&
-				(initialized_size < i_size))) {
-			/*
-			 * If this page is fully outside initialized size, zero
-			 * out all pages between the current initialized size
-			 * and the current page. Just use ntfs_readpage() to do
-			 * the zeroing transparently.
-			 */
-			if (block > iblock) {
-				// TODO:
-				// For each page do:
-				// - read_cache_page()
-				// Again for each page do:
-				// - wait_on_page_locked()
-				// - Check (PageUptodate(page) &&
-				//			!PageError(page))
-				// Update initialized size in the attribute and
-				// in the inode.
-				// Again, for each page do:
-				//	__set_page_dirty_buffers();
-				// page_cache_release()
-				// We don't need to wait on the writes.
-				// Update iblock.
-			}
-			/*
-			 * The current page straddles initialized size. Zero
-			 * all non-uptodate buffers and set them uptodate (and
-			 * dirty?). Note, there aren't any non-uptodate buffers
-			 * if the page is uptodate.
-			 * FIXME: For an uptodate page, the buffers may need to
-			 * be written out because they were not initialized on
-			 * disk before.
-			 */
-			if (!PageUptodate(page)) {
-				// TODO:
-				// Zero any non-uptodate buffers up to i_size.
-				// Set them uptodate and dirty.
-			}
-			// TODO:
-			// Update initialized size in the attribute and in the
-			// inode (up to i_size).
-			// Update iblock.
-			// FIXME: This is inefficient. Try to batch the two
-			// size changes to happen in one go.
-			ntfs_error(vol->sb, "Writing beyond initialized size "
-					"is not supported yet. Sorry.");
-			err = -EOPNOTSUPP;
-			goto err_out;
-			// Do NOT set_buffer_new() BUT DO clear buffer range
-			// outside write request range.
-			// set_buffer_uptodate() on complete buffers as well as
-			// set_buffer_dirty().
-		}
-
-		/* Need to map unmapped buffers. */
-		if (!buffer_mapped(bh)) {
-			/* Unmapped buffer. Need to map it. */
-			bh->b_bdev = vol->sb->s_bdev;
-
-			/* Convert block into corresponding vcn and offset. */
-			vcn = (VCN)block << blocksize_bits >>
-					vol->cluster_size_bits;
-			vcn_ofs = ((VCN)block << blocksize_bits) &
-					vol->cluster_size_mask;
-
-			is_retry = FALSE;
-			if (!rl) {
-lock_retry_remap:
-				down_read(&ni->runlist.lock);
-				rl = ni->runlist.rl;
-			}
-			if (likely(rl != NULL)) {
-				/* Seek to element containing target vcn. */
-				while (rl->length && rl[1].vcn <= vcn)
-					rl++;
-				lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
-			} else
-				lcn = LCN_RL_NOT_MAPPED;
-			if (unlikely(lcn < 0)) {
-				/*
-				 * We extended the attribute allocation above.
-				 * If we hit an ENOENT here it means that the
-				 * allocation was insufficient which is a bug.
-				 */
-				BUG_ON(lcn == LCN_ENOENT);
-
-				/* It is a hole, need to instantiate it. */
-				if (lcn == LCN_HOLE) {
-					// TODO: Instantiate the hole.
-					// clear_buffer_new(bh);
-					// unmap_underlying_metadata(bh->b_bdev,
-					//		bh->b_blocknr);
-					// For non-uptodate buffers, need to
-					// zero out the region outside the
-					// request in this bh or all bhs,
-					// depending on what we implemented
-					// above.
-					// Need to flush_dcache_page().
-					// Or could use set_buffer_new()
-					// instead?
-					ntfs_error(vol->sb, "Writing into "
-							"sparse regions is "
-							"not supported yet. "
-							"Sorry.");
-					err = -EOPNOTSUPP;
-					if (!rl)
-						up_read(&ni->runlist.lock);
-					goto err_out;
-				} else if (!is_retry &&
-						lcn == LCN_RL_NOT_MAPPED) {
-					is_retry = TRUE;
-					/*
-					 * Attempt to map runlist, dropping
-					 * lock for the duration.
-					 */
-					up_read(&ni->runlist.lock);
-					err = ntfs_map_runlist(ni, vcn);
-					if (likely(!err))
-						goto lock_retry_remap;
-					rl = NULL;
-				} else if (!rl)
-					up_read(&ni->runlist.lock);
-				/*
-				 * Failed to map the buffer, even after
-				 * retrying.
-				 */
-				if (!err)
-					err = -EIO;
-				bh->b_blocknr = -1;
-				ntfs_error(vol->sb, "Failed to write to inode "
-						"0x%lx, attribute type 0x%x, "
-						"vcn 0x%llx, offset 0x%x "
-						"because its location on disk "
-						"could not be determined%s "
-						"(error code %i).",
-						ni->mft_no, ni->type,
-						(unsigned long long)vcn,
-						vcn_ofs, is_retry ? " even "
-						"after retrying" : "", err);
-				goto err_out;
-			}
-			/* We now have a successful remap, i.e. lcn >= 0. */
-
-			/* Setup buffer head to correct block. */
-			bh->b_blocknr = ((lcn << vol->cluster_size_bits)
-					+ vcn_ofs) >> blocksize_bits;
-			set_buffer_mapped(bh);
-
-			// FIXME: Something analogous to this is needed for
-			// each newly allocated block, i.e. BH_New.
-			// FIXME: Might need to take this out of the
-			// if (!buffer_mapped(bh)) {}, depending on how we
-			// implement things during the allocated_size and
-			// initialized_size extension code above.
-			if (buffer_new(bh)) {
-				clear_buffer_new(bh);
-				unmap_underlying_metadata(bh->b_bdev,
-						bh->b_blocknr);
-				if (PageUptodate(page)) {
-					set_buffer_uptodate(bh);
-					continue;
-				}
-				/*
-				 * Page is _not_ uptodate, zero surrounding
-				 * region. NOTE: This is how we decide if to
-				 * zero or not!
-				 */
-				if (block_end > to || block_start < from) {
-					void *kaddr;
-
-					kaddr = kmap_atomic(page, KM_USER0);
-					if (block_end > to)
-						memset(kaddr + to, 0,
-								block_end - to);
-					if (block_start < from)
-						memset(kaddr + block_start, 0,
-								from -
-								block_start);
-					flush_dcache_page(page);
-					kunmap_atomic(kaddr, KM_USER0);
-				}
-				continue;
-			}
-		}
-		/* @bh is mapped, set it uptodate if the page is uptodate. */
-		if (PageUptodate(page)) {
-			if (!buffer_uptodate(bh))
-				set_buffer_uptodate(bh);
-			continue;
-		}
-		/*
-		 * The page is not uptodate. The buffer is mapped. If it is not
-		 * uptodate, and it is only partially being written to, we need
-		 * to read the buffer in before the write, i.e. right now.
-		 */
-		if (!buffer_uptodate(bh) &&
-				(block_start < from || block_end > to)) {
-			ll_rw_block(READ, 1, &bh);
-			*wait_bh++ = bh;
-		}
-	} while (block++, block_start = block_end,
-			(bh = bh->b_this_page) != head);
-
-	/* Release the lock if we took it. */
-	if (rl) {
-		up_read(&ni->runlist.lock);
-		rl = NULL;
-	}
-
-	/* If we issued read requests, let them complete. */
-	while (wait_bh > wait) {
-		wait_on_buffer(*--wait_bh);
-		if (!buffer_uptodate(*wait_bh))
-			return -EIO;
-	}
-
-	ntfs_debug("Done.");
-	return 0;
-err_out:
-	/*
-	 * Zero out any newly allocated blocks to avoid exposing stale data.
-	 * If BH_New is set, we know that the block was newly allocated in the
-	 * above loop.
-	 * FIXME: What about initialized_size increments? Have we done all the
-	 * required zeroing above? If not this error handling is broken, and
-	 * in particular the if (block_end <= from) check is completely bogus.
-	 */
-	bh = head;
-	block_start = 0;
-	is_retry = FALSE;
-	do {
-		block_end = block_start + blocksize;
-		if (block_end <= from)
-			continue;
-		if (block_start >= to)
-			break;
-		if (buffer_new(bh)) {
-			void *kaddr;
-
-			clear_buffer_new(bh);
-			kaddr = kmap_atomic(page, KM_USER0);
-			memset(kaddr + block_start, 0, bh->b_size);
-			kunmap_atomic(kaddr, KM_USER0);
-			set_buffer_uptodate(bh);
-			mark_buffer_dirty(bh);
-			is_retry = TRUE;
-		}
-	} while (block_start = block_end, (bh = bh->b_this_page) != head);
-	if (is_retry)
-		flush_dcache_page(page);
-	if (rl)
-		up_read(&ni->runlist.lock);
-	return err;
-}
-
-/**
- * ntfs_prepare_write - prepare a page for receiving data
- *
- * This is called from generic_file_write() with i_sem held on the inode
- * (@page->mapping->host).  The @page is locked but not kmap()ped.  The source
- * data has not yet been copied into the @page.
- *
- * Need to extend the attribute/fill in holes if necessary, create blocks and
- * make partially overwritten blocks uptodate,
- *
- * i_size is not to be modified yet.
- *
- * Return 0 on success or -errno on error.
- *
- * Should be using block_prepare_write() [support for sparse files] or
- * cont_prepare_write() [no support for sparse files].  Cannot do that due to
- * ntfs specifics but can look at them for implementation guidance.
- *
- * Note: In the range, @from is inclusive and @to is exclusive, i.e. @from is
- * the first byte in the page that will be written to and @to is the first byte
- * after the last byte that will be written to.
- */
-static int ntfs_prepare_write(struct file *file, struct page *page,
-		unsigned from, unsigned to)
-{
-	s64 new_size;
-	loff_t i_size;
-	struct inode *vi = page->mapping->host;
-	ntfs_inode *base_ni = NULL, *ni = NTFS_I(vi);
-	ntfs_volume *vol = ni->vol;
-	ntfs_attr_search_ctx *ctx = NULL;
-	MFT_RECORD *m = NULL;
-	ATTR_RECORD *a;
-	u8 *kaddr;
-	u32 attr_len;
-	int err;
-
-	ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
-			"0x%lx, from = %u, to = %u.", vi->i_ino, ni->type,
-			page->index, from, to);
-	BUG_ON(!PageLocked(page));
-	BUG_ON(from > PAGE_CACHE_SIZE);
-	BUG_ON(to > PAGE_CACHE_SIZE);
-	BUG_ON(from > to);
-	BUG_ON(NInoMstProtected(ni));
-	/*
-	 * 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;
-			goto err_out;
-		}
-	}
-	/* If the attribute is not resident, deal with it elsewhere. */
-	if (NInoNonResident(ni)) {
-		/*
-		 * Only unnamed $DATA attributes can be compressed, encrypted,
-		 * and/or sparse.
-		 */
-		if (ni->type == AT_DATA && !ni->name_len) {
-			/* If file is encrypted, deny access, just like NT4. */
-			if (NInoEncrypted(ni)) {
-				ntfs_debug("Denying write access to encrypted "
-						"file.");
-				return -EACCES;
-			}
-			/* Compressed data streams are handled in compress.c. */
-			if (NInoCompressed(ni)) {
-				// TODO: Implement and replace this check with
-				// return ntfs_write_compressed_block(page);
-				ntfs_error(vi->i_sb, "Writing to compressed "
-						"files is not supported yet. "
-						"Sorry.");
-				return -EOPNOTSUPP;
-			}
-			// TODO: Implement and remove this check.
-			if (NInoSparse(ni)) {
-				ntfs_error(vi->i_sb, "Writing to sparse files "
-						"is not supported yet. Sorry.");
-				return -EOPNOTSUPP;
-			}
-		}
-		/* Normal data stream. */
-		return ntfs_prepare_nonresident_write(page, from, to);
-	}
-	/*
-	 * Attribute is resident, implying it is not compressed, encrypted, or
-	 * sparse.
-	 */
-	BUG_ON(page_has_buffers(page));
-	new_size = ((s64)page->index << PAGE_CACHE_SHIFT) + to;
-	/* If we do not need to resize the attribute allocation we are done. */
-	if (new_size <= i_size_read(vi))
-		goto done;
-	/* Map, pin, and lock the (base) mft 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);
-		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;
-	}
-	m = ctx->mrec;
-	a = ctx->attr;
-	/* The total length of the attribute value. */
-	attr_len = le32_to_cpu(a->data.resident.value_length);
-	/* Fix an eventual previous failure of ntfs_commit_write(). */
-	i_size = i_size_read(vi);
-	if (unlikely(attr_len > i_size)) {
-		attr_len = i_size;
-		a->data.resident.value_length = cpu_to_le32(attr_len);
-	}
-	/* If we do not need to resize the attribute allocation we are done. */
-	if (new_size <= attr_len)
-		goto done_unm;
-	/* Check if new size is allowed in $AttrDef. */
-	err = ntfs_attr_size_bounds_check(vol, ni->type, new_size);
-	if (unlikely(err)) {
-		if (err == -ERANGE) {
-			ntfs_error(vol->sb, "Write would cause the inode "
-					"0x%lx to exceed the maximum size for "
-					"its attribute type (0x%x).  Aborting "
-					"write.", vi->i_ino,
-					le32_to_cpu(ni->type));
-		} else {
-			ntfs_error(vol->sb, "Inode 0x%lx has unknown "
-					"attribute type 0x%x.  Aborting "
-					"write.", vi->i_ino,
-					le32_to_cpu(ni->type));
-			err = -EIO;
-		}
-		goto err_out2;
-	}
-	/*
-	 * Extend the attribute record to be able to store the new attribute
-	 * size.
-	 */
-	if (new_size >= vol->mft_record_size || ntfs_attr_record_resize(m, a,
-			le16_to_cpu(a->data.resident.value_offset) +
-			new_size)) {
-		/* Not enough space in the mft record. */
-		ntfs_error(vol->sb, "Not enough space in the mft record for "
-				"the resized attribute value.  This is not "
-				"supported yet.  Aborting write.");
-		err = -EOPNOTSUPP;
-		goto err_out2;
-	}
-	/*
-	 * We have enough space in the mft record to fit the write.  This
-	 * implies the attribute is smaller than the mft record and hence the
-	 * attribute must be in a single page and hence page->index must be 0.
-	 */
-	BUG_ON(page->index);
-	/*
-	 * If the beginning of the write is past the old size, enlarge the
-	 * attribute value up to the beginning of the write and fill it with
-	 * zeroes.
-	 */
-	if (from > attr_len) {
-		memset((u8*)a + le16_to_cpu(a->data.resident.value_offset) +
-				attr_len, 0, from - attr_len);
-		a->data.resident.value_length = cpu_to_le32(from);
-		/* Zero the corresponding area in the page as well. */
-		if (PageUptodate(page)) {
-			kaddr = kmap_atomic(page, KM_USER0);
-			memset(kaddr + attr_len, 0, from - attr_len);
-			kunmap_atomic(kaddr, KM_USER0);
-			flush_dcache_page(page);
-		}
-	}
-	flush_dcache_mft_record_page(ctx->ntfs_ino);
-	mark_mft_record_dirty(ctx->ntfs_ino);
-done_unm:
-	ntfs_attr_put_search_ctx(ctx);
-	unmap_mft_record(base_ni);
-	/*
-	 * Because resident attributes are handled by memcpy() to/from the
-	 * corresponding MFT record, and because this form of i/o is byte
-	 * aligned rather than block aligned, there is no need to bring the
-	 * page uptodate here as in the non-resident case where we need to
-	 * bring the buffers straddled by the write uptodate before
-	 * generic_file_write() does the copying from userspace.
-	 *
-	 * We thus defer the uptodate bringing of the page region outside the
-	 * region written to to ntfs_commit_write(), which makes the code
-	 * simpler and saves one atomic kmap which is good.
-	 */
-done:
-	ntfs_debug("Done.");
-	return 0;
-err_out:
-	if (err == -ENOMEM)
-		ntfs_warning(vi->i_sb, "Error allocating memory required to "
-				"prepare the write.");
-	else {
-		ntfs_error(vi->i_sb, "Resident attribute prepare write failed "
-				"with error %i.", err);
-		NVolSetErrors(vol);
-		make_bad_inode(vi);
-	}
-err_out2:
-	if (ctx)
-		ntfs_attr_put_search_ctx(ctx);
-	if (m)
-		unmap_mft_record(base_ni);
-	return err;
-}
-
-/**
- * ntfs_commit_nonresident_write -
- *
- */
-static int ntfs_commit_nonresident_write(struct page *page,
-		unsigned from, unsigned to)
-{
-	s64 pos = ((s64)page->index << PAGE_CACHE_SHIFT) + to;
-	struct inode *vi = page->mapping->host;
-	struct buffer_head *bh, *head;
-	unsigned int block_start, block_end, blocksize;
-	BOOL partial;
-
-	ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
-			"0x%lx, from = %u, to = %u.", vi->i_ino,
-			NTFS_I(vi)->type, page->index, from, to);
-	blocksize = 1 << vi->i_blkbits;
-
-	// FIXME: We need a whole slew of special cases in here for compressed
-	// files for example...
-	// For now, we know ntfs_prepare_write() would have failed so we can't
-	// get here in any of the cases which we have to special case, so we
-	// are just a ripped off, unrolled generic_commit_write().
-
-	bh = head = page_buffers(page);
-	block_start = 0;
-	partial = FALSE;
-	do {
-		block_end = block_start + blocksize;
-		if (block_end <= from || block_start >= to) {
-			if (!buffer_uptodate(bh))
-				partial = TRUE;
-		} else {
-			set_buffer_uptodate(bh);
-			mark_buffer_dirty(bh);
-		}
-	} while (block_start = block_end, (bh = bh->b_this_page) != head);
-	/*
-	 * If this is a partial write which happened to make all buffers
-	 * uptodate then we can optimize away a bogus ->readpage() for the next
-	 * read().  Here we 'discover' whether the page went uptodate as a
-	 * result of this (potentially partial) write.
-	 */
-	if (!partial)
-		SetPageUptodate(page);
-	/*
-	 * Not convinced about this at all.  See disparity comment above.  For
-	 * now we know ntfs_prepare_write() would have failed in the write
-	 * exceeds i_size case, so this will never trigger which is fine.
-	 */
-	if (pos > i_size_read(vi)) {
-		ntfs_error(vi->i_sb, "Writing beyond the existing file size is "
-				"not supported yet.  Sorry.");
-		return -EOPNOTSUPP;
-		// vi->i_size = pos;
-		// mark_inode_dirty(vi);
-	}
-	ntfs_debug("Done.");
-	return 0;
-}
-
-/**
- * ntfs_commit_write - commit the received data
- *
- * This is called from generic_file_write() with i_sem held on the inode
- * (@page->mapping->host).  The @page is locked but not kmap()ped.  The source
- * data has already been copied into the @page.  ntfs_prepare_write() has been
- * called before the data copied and it returned success so we can take the
- * results of various BUG checks and some error handling for granted.
- *
- * Need to mark modified blocks dirty so they get written out later when
- * ntfs_writepage() is invoked by the VM.
- *
- * Return 0 on success or -errno on error.
- *
- * Should be using generic_commit_write().  This 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.
- *
- * Cannot use generic_commit_write() due to ntfs specialities but can look at
- * it for implementation guidance.
- *
- * If things have gone as outlined in ntfs_prepare_write(), then we do not
- * need to do any page content modifications here at all, except in the write
- * to resident attribute case, where 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.
- */
-static int ntfs_commit_write(struct file *file, struct page *page,
-		unsigned from, unsigned to)
-{
-	struct inode *vi = page->mapping->host;
-	ntfs_inode *base_ni, *ni = NTFS_I(vi);
-	char *kaddr, *kattr;
-	ntfs_attr_search_ctx *ctx;
-	MFT_RECORD *m;
-	ATTR_RECORD *a;
-	u32 attr_len;
-	int err;
-
-	ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
-			"0x%lx, from = %u, to = %u.", vi->i_ino, ni->type,
-			page->index, from, to);
-	/* If the attribute is not resident, deal with it elsewhere. */
-	if (NInoNonResident(ni)) {
-		/* Only unnamed $DATA attributes can be compressed/encrypted. */
-		if (ni->type == AT_DATA && !ni->name_len) {
-			/* Encrypted files need separate handling. */
-			if (NInoEncrypted(ni)) {
-				// We never get here at present!
-				BUG();
-			}
-			/* Compressed data streams are handled in compress.c. */
-			if (NInoCompressed(ni)) {
-				// TODO: Implement this!
-				// return ntfs_write_compressed_block(page);
-				// We never get here at present!
-				BUG();
-			}
-		}
-		/* Normal data stream. */
-		return ntfs_commit_nonresident_write(page, from, to);
-	}
-	/*
-	 * 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;
-	/* 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;
-	/* The total length of the attribute value. */
-	attr_len = le32_to_cpu(a->data.resident.value_length);
-	BUG_ON(from > attr_len);
-	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 + from, kaddr + from, to - from);
-	/* Update the attribute length if necessary. */
-	if (to > attr_len) {
-		attr_len = to;
-		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 (from > 0)
-			memcpy(kaddr, kattr, from);
-		if (to < attr_len)
-			memcpy(kaddr + to, kattr + to, attr_len - to);
-		/* Zero the region outside the end of the attribute value. */
-		if (attr_len < PAGE_CACHE_SIZE)
-			memset(kaddr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
-		/*
-		 * The probability of not having done any of the above is
-		 * extremely small, so we just flush unconditionally.
-		 */
-		flush_dcache_page(page);
-		SetPageUptodate(page);
-	}
-	kunmap_atomic(kaddr, KM_USER0);
-	/* Update i_size if necessary. */
-	if (i_size_read(vi) < attr_len) {
-		unsigned long flags;
-
-		write_lock_irqsave(&ni->size_lock, flags);
-		ni->allocated_size = ni->initialized_size = attr_len;
-		i_size_write(vi, attr_len);
-		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(vi);
-	}
-	if (ctx)
-		ntfs_attr_put_search_ctx(ctx);
-	if (m)
-		unmap_mft_record(base_ni);
-	return err;
-}
-
 #endif	/* NTFS_RW */
 
 /**
@@ -2377,9 +1552,6 @@ struct address_space_operations ntfs_aops = {
 						   disk request queue. */
 #ifdef NTFS_RW
 	.writepage	= ntfs_writepage,	/* Write dirty page to disk. */
-	.prepare_write	= ntfs_prepare_write,	/* Prepare page and buffers
-						   ready to receive data. */
-	.commit_write	= ntfs_commit_write,	/* Commit received data. */
 #endif /* NTFS_RW */
 };
 
diff --git a/fs/ntfs/attrib.c b/fs/ntfs/attrib.c
index 3f9a4ff42ee5..eda056bac256 100644
--- a/fs/ntfs/attrib.c
+++ b/fs/ntfs/attrib.c
@@ -21,7 +21,9 @@
  */
 
 #include <linux/buffer_head.h>
+#include <linux/sched.h>
 #include <linux/swap.h>
+#include <linux/writeback.h>
 
 #include "attrib.h"
 #include "debug.h"
@@ -36,9 +38,27 @@
  * ntfs_map_runlist_nolock - map (a part of) a runlist of an ntfs inode
  * @ni:		ntfs inode for which to map (part of) a runlist
  * @vcn:	map runlist part containing this vcn
+ * @ctx:	active attribute search context if present or NULL if not
  *
  * Map the part of a runlist containing the @vcn of the ntfs inode @ni.
  *
+ * If @ctx is specified, it is an active search context of @ni and its base mft
+ * record.  This is needed when ntfs_map_runlist_nolock() encounters unmapped
+ * runlist fragments and allows their mapping.  If you do not have the mft
+ * record mapped, you can specify @ctx as NULL and ntfs_map_runlist_nolock()
+ * will perform the necessary mapping and unmapping.
+ *
+ * Note, ntfs_map_runlist_nolock() saves the state of @ctx on entry and
+ * restores it before returning.  Thus, @ctx will be left pointing to the same
+ * attribute on return as on entry.  However, the actual pointers in @ctx may
+ * point to different memory locations on return, so you must remember to reset
+ * any cached pointers from the @ctx, i.e. after the call to
+ * ntfs_map_runlist_nolock(), you will probably want to do:
+ *	m = ctx->mrec;
+ *	a = ctx->attr;
+ * Assuming you cache ctx->attr in a variable @a of type ATTR_RECORD * and that
+ * you cache ctx->mrec in a variable @m of type MFT_RECORD *.
+ *
  * Return 0 on success and -errno on error.  There is one special error code
  * which is not an error as such.  This is -ENOENT.  It means that @vcn is out
  * of bounds of the runlist.
@@ -46,19 +66,32 @@
  * Note the runlist can be NULL after this function returns if @vcn is zero and
  * the attribute has zero allocated size, i.e. there simply is no runlist.
  *
- * Locking: - The runlist must be locked for writing.
- *	    - This function modifies the runlist.
+ * WARNING: If @ctx is supplied, regardless of whether success or failure is
+ *	    returned, you need to check IS_ERR(@ctx->mrec) and if TRUE the @ctx
+ *	    is no longer valid, i.e. you need to either call
+ *	    ntfs_attr_reinit_search_ctx() or ntfs_attr_put_search_ctx() on it.
+ *	    In that case PTR_ERR(@ctx->mrec) will give you the error code for
+ *	    why the mapping of the old inode failed.
+ *
+ * Locking: - The runlist described by @ni must be locked for writing on entry
+ *	      and is locked on return.  Note the runlist will be modified.
+ *	    - If @ctx is NULL, the base mft record of @ni must not be mapped on
+ *	      entry and it will be left unmapped on return.
+ *	    - If @ctx is not NULL, the base mft record must be mapped on entry
+ *	      and it will be left mapped on return.
  */
-int ntfs_map_runlist_nolock(ntfs_inode *ni, VCN vcn)
+int ntfs_map_runlist_nolock(ntfs_inode *ni, VCN vcn, ntfs_attr_search_ctx *ctx)
 {
 	VCN end_vcn;
+	unsigned long flags;
 	ntfs_inode *base_ni;
 	MFT_RECORD *m;
 	ATTR_RECORD *a;
-	ntfs_attr_search_ctx *ctx;
 	runlist_element *rl;
-	unsigned long flags;
+	struct page *put_this_page = NULL;
 	int err = 0;
+	BOOL ctx_is_temporary, ctx_needs_reset;
+	ntfs_attr_search_ctx old_ctx = { NULL, };
 
 	ntfs_debug("Mapping runlist part containing vcn 0x%llx.",
 			(unsigned long long)vcn);
@@ -66,20 +99,77 @@ int ntfs_map_runlist_nolock(ntfs_inode *ni, VCN vcn)
 		base_ni = ni;
 	else
 		base_ni = ni->ext.base_ntfs_ino;
-	m = map_mft_record(base_ni);
-	if (IS_ERR(m))
-		return PTR_ERR(m);
-	ctx = ntfs_attr_get_search_ctx(base_ni, m);
-	if (unlikely(!ctx)) {
-		err = -ENOMEM;
-		goto err_out;
+	if (!ctx) {
+		ctx_is_temporary = ctx_needs_reset = TRUE;
+		m = map_mft_record(base_ni);
+		if (IS_ERR(m))
+			return PTR_ERR(m);
+		ctx = ntfs_attr_get_search_ctx(base_ni, m);
+		if (unlikely(!ctx)) {
+			err = -ENOMEM;
+			goto err_out;
+		}
+	} else {
+		VCN allocated_size_vcn;
+
+		BUG_ON(IS_ERR(ctx->mrec));
+		a = ctx->attr;
+		BUG_ON(!a->non_resident);
+		ctx_is_temporary = FALSE;
+		end_vcn = sle64_to_cpu(a->data.non_resident.highest_vcn);
+		read_lock_irqsave(&ni->size_lock, flags);
+		allocated_size_vcn = ni->allocated_size >>
+				ni->vol->cluster_size_bits;
+		read_unlock_irqrestore(&ni->size_lock, flags);
+		if (!a->data.non_resident.lowest_vcn && end_vcn <= 0)
+			end_vcn = allocated_size_vcn - 1;
+		/*
+		 * If we already have the attribute extent containing @vcn in
+		 * @ctx, no need to look it up again.  We slightly cheat in
+		 * that if vcn exceeds the allocated size, we will refuse to
+		 * map the runlist below, so there is definitely no need to get
+		 * the right attribute extent.
+		 */
+		if (vcn >= allocated_size_vcn || (a->type == ni->type &&
+				a->name_length == ni->name_len &&
+				!memcmp((u8*)a + le16_to_cpu(a->name_offset),
+				ni->name, ni->name_len) &&
+				sle64_to_cpu(a->data.non_resident.lowest_vcn)
+				<= vcn && end_vcn >= vcn))
+			ctx_needs_reset = FALSE;
+		else {
+			/* Save the old search context. */
+			old_ctx = *ctx;
+			/*
+			 * If the currently mapped (extent) inode is not the
+			 * base inode we will unmap it when we reinitialize the
+			 * search context which means we need to get a
+			 * reference to the page containing the mapped mft
+			 * record so we do not accidentally drop changes to the
+			 * mft record when it has not been marked dirty yet.
+			 */
+			if (old_ctx.base_ntfs_ino && old_ctx.ntfs_ino !=
+					old_ctx.base_ntfs_ino) {
+				put_this_page = old_ctx.ntfs_ino->page;
+				page_cache_get(put_this_page);
+			}
+			/*
+			 * Reinitialize the search context so we can lookup the
+			 * needed attribute extent.
+			 */
+			ntfs_attr_reinit_search_ctx(ctx);
+			ctx_needs_reset = TRUE;
+		}
 	}
-	err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
-			CASE_SENSITIVE, vcn, NULL, 0, ctx);
-	if (unlikely(err)) {
-		if (err == -ENOENT)
-			err = -EIO;
-		goto err_out;
+	if (ctx_needs_reset) {
+		err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
+				CASE_SENSITIVE, vcn, NULL, 0, ctx);
+		if (unlikely(err)) {
+			if (err == -ENOENT)
+				err = -EIO;
+			goto err_out;
+		}
+		BUG_ON(!ctx->attr->non_resident);
 	}
 	a = ctx->attr;
 	/*
@@ -89,11 +179,9 @@ int ntfs_map_runlist_nolock(ntfs_inode *ni, VCN vcn)
 	 * ntfs_mapping_pairs_decompress() fails.
 	 */
 	end_vcn = sle64_to_cpu(a->data.non_resident.highest_vcn) + 1;
-	if (unlikely(!a->data.non_resident.lowest_vcn && end_vcn <= 1)) {
-		read_lock_irqsave(&ni->size_lock, flags);
-		end_vcn = ni->allocated_size >> ni->vol->cluster_size_bits;
-		read_unlock_irqrestore(&ni->size_lock, flags);
-	}
+	if (!a->data.non_resident.lowest_vcn && end_vcn == 1)
+		end_vcn = sle64_to_cpu(a->data.non_resident.allocated_size) >>
+				ni->vol->cluster_size_bits;
 	if (unlikely(vcn >= end_vcn)) {
 		err = -ENOENT;
 		goto err_out;
@@ -104,9 +192,93 @@ int ntfs_map_runlist_nolock(ntfs_inode *ni, VCN vcn)
 	else
 		ni->runlist.rl = rl;
 err_out:
-	if (likely(ctx))
-		ntfs_attr_put_search_ctx(ctx);
-	unmap_mft_record(base_ni);
+	if (ctx_is_temporary) {
+		if (likely(ctx))
+			ntfs_attr_put_search_ctx(ctx);
+		unmap_mft_record(base_ni);
+	} else if (ctx_needs_reset) {
+		/*
+		 * If there is no attribute list, restoring the search context
+		 * is acomplished simply by copying the saved context back over
+		 * the caller supplied context.  If there is an attribute list,
+		 * things are more complicated as we need to deal with mapping
+		 * of mft records and resulting potential changes in pointers.
+		 */
+		if (NInoAttrList(base_ni)) {
+			/*
+			 * If the currently mapped (extent) inode is not the
+			 * one we had before, we need to unmap it and map the
+			 * old one.
+			 */
+			if (ctx->ntfs_ino != old_ctx.ntfs_ino) {
+				/*
+				 * If the currently mapped inode is not the
+				 * base inode, unmap it.
+				 */
+				if (ctx->base_ntfs_ino && ctx->ntfs_ino !=
+						ctx->base_ntfs_ino) {
+					unmap_extent_mft_record(ctx->ntfs_ino);
+					ctx->mrec = ctx->base_mrec;
+					BUG_ON(!ctx->mrec);
+				}
+				/*
+				 * If the old mapped inode is not the base
+				 * inode, map it.
+				 */
+				if (old_ctx.base_ntfs_ino &&
+						old_ctx.ntfs_ino !=
+						old_ctx.base_ntfs_ino) {
+retry_map:
+					ctx->mrec = map_mft_record(
+							old_ctx.ntfs_ino);
+					/*
+					 * Something bad has happened.  If out
+					 * of memory retry till it succeeds.
+					 * Any other errors are fatal and we
+					 * return the error code in ctx->mrec.
+					 * Let the caller deal with it...  We
+					 * just need to fudge things so the
+					 * caller can reinit and/or put the
+					 * search context safely.
+					 */
+					if (IS_ERR(ctx->mrec)) {
+						if (PTR_ERR(ctx->mrec) ==
+								-ENOMEM) {
+							schedule();
+							goto retry_map;
+						} else
+							old_ctx.ntfs_ino =
+								old_ctx.
+								base_ntfs_ino;
+					}
+				}
+			}
+			/* Update the changed pointers in the saved context. */
+			if (ctx->mrec != old_ctx.mrec) {
+				if (!IS_ERR(ctx->mrec))
+					old_ctx.attr = (ATTR_RECORD*)(
+							(u8*)ctx->mrec +
+							((u8*)old_ctx.attr -
+							(u8*)old_ctx.mrec));
+				old_ctx.mrec = ctx->mrec;
+			}
+		}
+		/* Restore the search context to the saved one. */
+		*ctx = old_ctx;
+		/*
+		 * We drop the reference on the page we took earlier.  In the
+		 * case that IS_ERR(ctx->mrec) is true this means we might lose
+		 * some changes to the mft record that had been made between
+		 * the last time it was marked dirty/written out and now.  This
+		 * at this stage is not a problem as the mapping error is fatal
+		 * enough that the mft record cannot be written out anyway and
+		 * the caller is very likely to shutdown the whole inode
+		 * immediately and mark the volume dirty for chkdsk to pick up
+		 * the pieces anyway.
+		 */
+		if (put_this_page)
+			page_cache_release(put_this_page);
+	}
 	return err;
 }
 
@@ -122,8 +294,8 @@ err_out:
  * of bounds of the runlist.
  *
  * Locking: - The runlist must be unlocked on entry and is unlocked on return.
- *	    - This function takes the runlist lock for writing and modifies the
- *	      runlist.
+ *	    - This function takes the runlist lock for writing and may modify
+ *	      the runlist.
  */
 int ntfs_map_runlist(ntfs_inode *ni, VCN vcn)
 {
@@ -133,7 +305,7 @@ int ntfs_map_runlist(ntfs_inode *ni, VCN vcn)
 	/* Make sure someone else didn't do the work while we were sleeping. */
 	if (likely(ntfs_rl_vcn_to_lcn(ni->runlist.rl, vcn) <=
 			LCN_RL_NOT_MAPPED))
-		err = ntfs_map_runlist_nolock(ni, vcn);
+		err = ntfs_map_runlist_nolock(ni, vcn, NULL);
 	up_write(&ni->runlist.lock);
 	return err;
 }
@@ -212,7 +384,7 @@ retry_remap:
 				goto retry_remap;
 			}
 		}
-		err = ntfs_map_runlist_nolock(ni, vcn);
+		err = ntfs_map_runlist_nolock(ni, vcn, NULL);
 		if (!write_locked) {
 			up_write(&ni->runlist.lock);
 			down_read(&ni->runlist.lock);
@@ -236,9 +408,9 @@ retry_remap:
 
 /**
  * ntfs_attr_find_vcn_nolock - find a vcn in the runlist of an ntfs inode
- * @ni:			ntfs inode describing the runlist to search
- * @vcn:		vcn to find
- * @write_locked:	true if the runlist is locked for writing
+ * @ni:		ntfs inode describing the runlist to search
+ * @vcn:	vcn to find
+ * @ctx:	active attribute search context if present or NULL if not
  *
  * Find the virtual cluster number @vcn in the runlist described by the ntfs
  * inode @ni and return the address of the runlist element containing the @vcn.
@@ -246,9 +418,22 @@ retry_remap:
  * If the @vcn is not mapped yet, the attempt is made to map the attribute
  * extent containing the @vcn and the vcn to lcn conversion is retried.
  *
- * If @write_locked is true the caller has locked the runlist for writing and
- * if false for reading.
- *
+ * If @ctx is specified, it is an active search context of @ni and its base mft
+ * record.  This is needed when ntfs_attr_find_vcn_nolock() encounters unmapped
+ * runlist fragments and allows their mapping.  If you do not have the mft
+ * record mapped, you can specify @ctx as NULL and ntfs_attr_find_vcn_nolock()
+ * will perform the necessary mapping and unmapping.
+ *
+ * Note, ntfs_attr_find_vcn_nolock() saves the state of @ctx on entry and
+ * restores it before returning.  Thus, @ctx will be left pointing to the same
+ * attribute on return as on entry.  However, the actual pointers in @ctx may
+ * point to different memory locations on return, so you must remember to reset
+ * any cached pointers from the @ctx, i.e. after the call to
+ * ntfs_attr_find_vcn_nolock(), you will probably want to do:
+ *	m = ctx->mrec;
+ *	a = ctx->attr;
+ * Assuming you cache ctx->attr in a variable @a of type ATTR_RECORD * and that
+ * you cache ctx->mrec in a variable @m of type MFT_RECORD *.
  * Note you need to distinguish between the lcn of the returned runlist element
  * being >= 0 and LCN_HOLE.  In the later case you have to return zeroes on
  * read and allocate clusters on write.
@@ -263,22 +448,31 @@ retry_remap:
  *	-ENOMEM - Not enough memory to map runlist.
  *	-EIO	- Critical error (runlist/file is corrupt, i/o error, etc).
  *
- * Locking: - The runlist must be locked on entry and is left locked on return.
- *	    - If @write_locked is FALSE, i.e. the runlist is locked for reading,
- *	      the lock may be dropped inside the function so you cannot rely on
- *	      the runlist still being the same when this function returns.
+ * WARNING: If @ctx is supplied, regardless of whether success or failure is
+ *	    returned, you need to check IS_ERR(@ctx->mrec) and if TRUE the @ctx
+ *	    is no longer valid, i.e. you need to either call
+ *	    ntfs_attr_reinit_search_ctx() or ntfs_attr_put_search_ctx() on it.
+ *	    In that case PTR_ERR(@ctx->mrec) will give you the error code for
+ *	    why the mapping of the old inode failed.
+ *
+ * Locking: - The runlist described by @ni must be locked for writing on entry
+ *	      and is locked on return.  Note the runlist may be modified when
+ *	      needed runlist fragments need to be mapped.
+ *	    - If @ctx is NULL, the base mft record of @ni must not be mapped on
+ *	      entry and it will be left unmapped on return.
+ *	    - If @ctx is not NULL, the base mft record must be mapped on entry
+ *	      and it will be left mapped on return.
  */
 runlist_element *ntfs_attr_find_vcn_nolock(ntfs_inode *ni, const VCN vcn,
-		const BOOL write_locked)
+		ntfs_attr_search_ctx *ctx)
 {
 	unsigned long flags;
 	runlist_element *rl;
 	int err = 0;
 	BOOL is_retry = FALSE;
 
-	ntfs_debug("Entering for i_ino 0x%lx, vcn 0x%llx, %s_locked.",
-			ni->mft_no, (unsigned long long)vcn,
-			write_locked ? "write" : "read");
+	ntfs_debug("Entering for i_ino 0x%lx, vcn 0x%llx, with%s ctx.",
+			ni->mft_no, (unsigned long long)vcn, ctx ? "" : "out");
 	BUG_ON(!ni);
 	BUG_ON(!NInoNonResident(ni));
 	BUG_ON(vcn < 0);
@@ -312,33 +506,22 @@ retry_remap:
 	}
 	if (!err && !is_retry) {
 		/*
-		 * The @vcn is in an unmapped region, map the runlist and
-		 * retry.
+		 * If the search context is invalid we cannot map the unmapped
+		 * region.
 		 */
-		if (!write_locked) {
-			up_read(&ni->runlist.lock);
-			down_write(&ni->runlist.lock);
-			if (unlikely(ntfs_rl_vcn_to_lcn(ni->runlist.rl, vcn) !=
-					LCN_RL_NOT_MAPPED)) {
-				up_write(&ni->runlist.lock);
-				down_read(&ni->runlist.lock);
+		if (IS_ERR(ctx->mrec))
+			err = PTR_ERR(ctx->mrec);
+		else {
+			/*
+			 * The @vcn is in an unmapped region, map the runlist
+			 * and retry.
+			 */
+			err = ntfs_map_runlist_nolock(ni, vcn, ctx);
+			if (likely(!err)) {
+				is_retry = TRUE;
 				goto retry_remap;
 			}
 		}
-		err = ntfs_map_runlist_nolock(ni, vcn);
-		if (!write_locked) {
-			up_write(&ni->runlist.lock);
-			down_read(&ni->runlist.lock);
-		}
-		if (likely(!err)) {
-			is_retry = TRUE;
-			goto retry_remap;
-		}
-		/*
-		 * -EINVAL coming from a failed mapping attempt is equivalent
-		 * to i/o error for us as it should not happen in our code
-		 * paths.
-		 */
 		if (err == -EINVAL)
 			err = -EIO;
 	} else if (!err)
@@ -1011,6 +1194,7 @@ int ntfs_attr_lookup(const ATTR_TYPE type, const ntfschar *name,
 	ntfs_inode *base_ni;
 
 	ntfs_debug("Entering.");
+	BUG_ON(IS_ERR(ctx->mrec));
 	if (ctx->base_ntfs_ino)
 		base_ni = ctx->base_ntfs_ino;
 	else
@@ -1227,7 +1411,7 @@ int ntfs_attr_can_be_non_resident(const ntfs_volume *vol, const ATTR_TYPE type)
  */
 int ntfs_attr_can_be_resident(const ntfs_volume *vol, const ATTR_TYPE type)
 {
-	if (type == AT_INDEX_ALLOCATION || type == AT_EA)
+	if (type == AT_INDEX_ALLOCATION)
 		return -EPERM;
 	return 0;
 }
@@ -1319,10 +1503,17 @@ int ntfs_resident_attr_value_resize(MFT_RECORD *m, ATTR_RECORD *a,
 /**
  * ntfs_attr_make_non_resident - convert a resident to a non-resident attribute
  * @ni:		ntfs inode describing the attribute to convert
+ * @data_size:	size of the resident data to copy to the non-resident attribute
  *
  * Convert the resident ntfs attribute described by the ntfs inode @ni to a
  * non-resident one.
  *
+ * @data_size must be equal to the attribute value size.  This is needed since
+ * we need to know the size before we can map the mft record and our callers
+ * always know it.  The reason we cannot simply read the size from the vfs
+ * inode i_size is that this is not necessarily uptodate.  This happens when
+ * ntfs_attr_make_non_resident() is called in the ->truncate call path(s).
+ *
  * Return 0 on success and -errno on error.  The following error return codes
  * are defined:
  *	-EPERM	- The attribute is not allowed to be non-resident.
@@ -1343,7 +1534,7 @@ int ntfs_resident_attr_value_resize(MFT_RECORD *m, ATTR_RECORD *a,
  *
  * Locking: - The caller must hold i_sem on the inode.
  */
-int ntfs_attr_make_non_resident(ntfs_inode *ni)
+int ntfs_attr_make_non_resident(ntfs_inode *ni, const u32 data_size)
 {
 	s64 new_size;
 	struct inode *vi = VFS_I(ni);
@@ -1381,11 +1572,9 @@ int ntfs_attr_make_non_resident(ntfs_inode *ni)
 	 * The size needs to be aligned to a cluster boundary for allocation
 	 * purposes.
 	 */
-	new_size = (i_size_read(vi) + vol->cluster_size - 1) &
+	new_size = (data_size + vol->cluster_size - 1) &
 			~(vol->cluster_size - 1);
 	if (new_size > 0) {
-		runlist_element *rl2;
-
 		/*
 		 * Will need the page later and since the page lock nests
 		 * outside all ntfs locks, we need to get the page now.
@@ -1396,7 +1585,7 @@ int ntfs_attr_make_non_resident(ntfs_inode *ni)
 			return -ENOMEM;
 		/* Start by allocating clusters to hold the attribute value. */
 		rl = ntfs_cluster_alloc(vol, 0, new_size >>
-				vol->cluster_size_bits, -1, DATA_ZONE);
+				vol->cluster_size_bits, -1, DATA_ZONE, TRUE);
 		if (IS_ERR(rl)) {
 			err = PTR_ERR(rl);
 			ntfs_debug("Failed to allocate cluster%s, error code "
@@ -1405,12 +1594,6 @@ int ntfs_attr_make_non_resident(ntfs_inode *ni)
 					err);
 			goto page_err_out;
 		}
-		/* Change the runlist terminator to LCN_ENOENT. */
-		rl2 = rl;
-		while (rl2->length)
-			rl2++;
-		BUG_ON(rl2->lcn != LCN_RL_NOT_MAPPED);
-		rl2->lcn = LCN_ENOENT;
 	} else {
 		rl = NULL;
 		page = NULL;
@@ -1473,7 +1656,7 @@ int ntfs_attr_make_non_resident(ntfs_inode *ni)
 	 * attribute value.
 	 */
 	attr_size = le32_to_cpu(a->data.resident.value_length);
-	BUG_ON(attr_size != i_size_read(vi));
+	BUG_ON(attr_size != data_size);
 	if (page && !PageUptodate(page)) {
 		kaddr = kmap_atomic(page, KM_USER0);
 		memcpy(kaddr, (u8*)a +
@@ -1538,7 +1721,9 @@ int ntfs_attr_make_non_resident(ntfs_inode *ni)
 				ffs(ni->itype.compressed.block_size) - 1;
 		ni->itype.compressed.block_clusters = 1U <<
 				a->data.non_resident.compression_unit;
-	}
+		vi->i_blocks = ni->itype.compressed.size >> 9;
+	} else
+		vi->i_blocks = ni->allocated_size >> 9;
 	write_unlock_irqrestore(&ni->size_lock, flags);
 	/*
 	 * This needs to be last since the address space operations ->readpage
@@ -1652,6 +1837,640 @@ page_err_out:
 }
 
 /**
+ * ntfs_attr_extend_allocation - extend the allocated space of an attribute
+ * @ni:			ntfs inode of the attribute whose allocation to extend
+ * @new_alloc_size:	new size in bytes to which to extend the allocation to
+ * @new_data_size:	new size in bytes to which to extend the data to
+ * @data_start:		beginning of region which is required to be non-sparse
+ *
+ * Extend the allocated space of an attribute described by the ntfs inode @ni
+ * to @new_alloc_size bytes.  If @data_start is -1, the whole extension may be
+ * implemented as a hole in the file (as long as both the volume and the ntfs
+ * inode @ni have sparse support enabled).  If @data_start is >= 0, then the
+ * region between the old allocated size and @data_start - 1 may be made sparse
+ * but the regions between @data_start and @new_alloc_size must be backed by
+ * actual clusters.
+ *
+ * If @new_data_size is -1, it is ignored.  If it is >= 0, then the data size
+ * of the attribute is extended to @new_data_size.  Note that the i_size of the
+ * vfs inode is not updated.  Only the data size in the base attribute record
+ * is updated.  The caller has to update i_size separately if this is required.
+ * WARNING: It is a BUG() for @new_data_size to be smaller than the old data
+ * size as well as for @new_data_size to be greater than @new_alloc_size.
+ *
+ * For resident attributes this involves resizing the attribute record and if
+ * necessary moving it and/or other attributes into extent mft records and/or
+ * converting the attribute to a non-resident attribute which in turn involves
+ * extending the allocation of a non-resident attribute as described below.
+ *
+ * For non-resident attributes this involves allocating clusters in the data
+ * zone on the volume (except for regions that are being made sparse) and
+ * extending the run list to describe the allocated clusters as well as
+ * updating the mapping pairs array of the attribute.  This in turn involves
+ * resizing the attribute record and if necessary moving it and/or other
+ * attributes into extent mft records and/or splitting the attribute record
+ * into multiple extent attribute records.
+ *
+ * Also, the attribute list attribute is updated if present and in some of the
+ * above cases (the ones where extent mft records/attributes come into play),
+ * an attribute list attribute is created if not already present.
+ *
+ * Return the new allocated size on success and -errno on error.  In the case
+ * that an error is encountered but a partial extension at least up to
+ * @data_start (if present) is possible, the allocation is partially extended
+ * and this is returned.  This means the caller must check the returned size to
+ * determine if the extension was partial.  If @data_start is -1 then partial
+ * allocations are not performed.
+ *
+ * WARNING: Do not call ntfs_attr_extend_allocation() for $MFT/$DATA.
+ *
+ * Locking: This function takes the runlist lock of @ni for writing as well as
+ * locking the mft record of the base ntfs inode.  These locks are maintained
+ * throughout execution of the function.  These locks are required so that the
+ * attribute can be resized safely and so that it can for example be converted
+ * from resident to non-resident safely.
+ *
+ * TODO: At present attribute list attribute handling is not implemented.
+ *
+ * TODO: At present it is not safe to call this function for anything other
+ * than the $DATA attribute(s) of an uncompressed and unencrypted file.
+ */
+s64 ntfs_attr_extend_allocation(ntfs_inode *ni, s64 new_alloc_size,
+		const s64 new_data_size, const s64 data_start)
+{
+	VCN vcn;
+	s64 ll, allocated_size, start = data_start;
+	struct inode *vi = VFS_I(ni);
+	ntfs_volume *vol = ni->vol;
+	ntfs_inode *base_ni;
+	MFT_RECORD *m;
+	ATTR_RECORD *a;
+	ntfs_attr_search_ctx *ctx;
+	runlist_element *rl, *rl2;
+	unsigned long flags;
+	int err, mp_size;
+	u32 attr_len = 0; /* Silence stupid gcc warning. */
+	BOOL mp_rebuilt;
+
+#ifdef NTFS_DEBUG
+	read_lock_irqsave(&ni->size_lock, flags);
+	allocated_size = ni->allocated_size;
+	read_unlock_irqrestore(&ni->size_lock, flags);
+	ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, "
+			"old_allocated_size 0x%llx, "
+			"new_allocated_size 0x%llx, new_data_size 0x%llx, "
+			"data_start 0x%llx.", vi->i_ino,
+			(unsigned)le32_to_cpu(ni->type),
+			(unsigned long long)allocated_size,
+			(unsigned long long)new_alloc_size,
+			(unsigned long long)new_data_size,
+			(unsigned long long)start);
+#endif
+retry_extend:
+	/*
+	 * For non-resident attributes, @start and @new_size need to be aligned
+	 * to cluster boundaries for allocation purposes.
+	 */
+	if (NInoNonResident(ni)) {
+		if (start > 0)
+			start &= ~(s64)vol->cluster_size_mask;
+		new_alloc_size = (new_alloc_size + vol->cluster_size - 1) &
+				~(s64)vol->cluster_size_mask;
+	}
+	BUG_ON(new_data_size >= 0 && new_data_size > new_alloc_size);
+	/* Check if new size is allowed in $AttrDef. */
+	err = ntfs_attr_size_bounds_check(vol, ni->type, new_alloc_size);
+	if (unlikely(err)) {
+		/* Only emit errors when the write will fail completely. */
+		read_lock_irqsave(&ni->size_lock, flags);
+		allocated_size = ni->allocated_size;
+		read_unlock_irqrestore(&ni->size_lock, flags);
+		if (start < 0 || start >= allocated_size) {
+			if (err == -ERANGE) {
+				ntfs_error(vol->sb, "Cannot extend allocation "
+						"of inode 0x%lx, attribute "
+						"type 0x%x, because the new "
+						"allocation would exceed the "
+						"maximum allowed size for "
+						"this attribute type.",
+						vi->i_ino, (unsigned)
+						le32_to_cpu(ni->type));
+			} else {
+				ntfs_error(vol->sb, "Cannot extend allocation "
+						"of inode 0x%lx, attribute "
+						"type 0x%x, because this "
+						"attribute type is not "
+						"defined on the NTFS volume.  "
+						"Possible corruption!  You "
+						"should run chkdsk!",
+						vi->i_ino, (unsigned)
+						le32_to_cpu(ni->type));
+			}
+		}
+		/* Translate error code to be POSIX conformant for write(2). */
+		if (err == -ERANGE)
+			err = -EFBIG;
+		else
+			err = -EIO;
+		return err;
+	}
+	if (!NInoAttr(ni))
+		base_ni = ni;
+	else
+		base_ni = ni->ext.base_ntfs_ino;
+	/*
+	 * We will be modifying both the runlist (if non-resident) and the mft
+	 * record so lock them both down.
+	 */
+	down_write(&ni->runlist.lock);
+	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;
+	}
+	read_lock_irqsave(&ni->size_lock, flags);
+	allocated_size = ni->allocated_size;
+	read_unlock_irqrestore(&ni->size_lock, flags);
+	/*
+	 * If non-resident, seek to the last extent.  If resident, there is
+	 * only one extent, so seek to that.
+	 */
+	vcn = NInoNonResident(ni) ? allocated_size >> vol->cluster_size_bits :
+			0;
+	/*
+	 * Abort if someone did the work whilst we waited for the locks.  If we
+	 * just converted the attribute from resident to non-resident it is
+	 * likely that exactly this has happened already.  We cannot quite
+	 * abort if we need to update the data size.
+	 */
+	if (unlikely(new_alloc_size <= allocated_size)) {
+		ntfs_debug("Allocated size already exceeds requested size.");
+		new_alloc_size = allocated_size;
+		if (new_data_size < 0)
+			goto done;
+		/*
+		 * We want the first attribute extent so that we can update the
+		 * data size.
+		 */
+		vcn = 0;
+	}
+	err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
+			CASE_SENSITIVE, vcn, NULL, 0, ctx);
+	if (unlikely(err)) {
+		if (err == -ENOENT)
+			err = -EIO;
+		goto err_out;
+	}
+	m = ctx->mrec;
+	a = ctx->attr;
+	/* Use goto to reduce indentation. */
+	if (a->non_resident)
+		goto do_non_resident_extend;
+	BUG_ON(NInoNonResident(ni));
+	/* The total length of the attribute value. */
+	attr_len = le32_to_cpu(a->data.resident.value_length);
+	/*
+	 * Extend the attribute record to be able to store the new attribute
+	 * size.  ntfs_attr_record_resize() will not do anything if the size is
+	 * not changing.
+	 */
+	if (new_alloc_size < vol->mft_record_size &&
+			!ntfs_attr_record_resize(m, a,
+			le16_to_cpu(a->data.resident.value_offset) +
+			new_alloc_size)) {
+		/* The resize succeeded! */
+		write_lock_irqsave(&ni->size_lock, flags);
+		ni->allocated_size = le32_to_cpu(a->length) -
+				le16_to_cpu(a->data.resident.value_offset);
+		write_unlock_irqrestore(&ni->size_lock, flags);
+		if (new_data_size >= 0) {
+			BUG_ON(new_data_size < attr_len);
+			a->data.resident.value_length =
+					cpu_to_le32((u32)new_data_size);
+		}
+		goto flush_done;
+	}
+	/*
+	 * We have to drop all the locks so we can call
+	 * ntfs_attr_make_non_resident().  This could be optimised by try-
+	 * locking the first page cache page and only if that fails dropping
+	 * the locks, locking the page, and redoing all the locking and
+	 * lookups.  While this would be a huge optimisation, it is not worth
+	 * it as this is definitely a slow code path.
+	 */
+	ntfs_attr_put_search_ctx(ctx);
+	unmap_mft_record(base_ni);
+	up_write(&ni->runlist.lock);
+	/*
+	 * Not enough space in the mft record, try to make the attribute
+	 * non-resident and if successful restart the extension process.
+	 */
+	err = ntfs_attr_make_non_resident(ni, attr_len);
+	if (likely(!err))
+		goto retry_extend;
+	/*
+	 * Could not make non-resident.  If this is due to this not being
+	 * permitted for this attribute type or there not being enough space,
+	 * try to make other attributes non-resident.  Otherwise fail.
+	 */
+	if (unlikely(err != -EPERM && err != -ENOSPC)) {
+		/* Only emit errors when the write will fail completely. */
+		read_lock_irqsave(&ni->size_lock, flags);
+		allocated_size = ni->allocated_size;
+		read_unlock_irqrestore(&ni->size_lock, flags);
+		if (start < 0 || start >= allocated_size)
+			ntfs_error(vol->sb, "Cannot extend allocation of "
+					"inode 0x%lx, attribute type 0x%x, "
+					"because the conversion from resident "
+					"to non-resident attribute failed "
+					"with error code %i.", vi->i_ino,
+					(unsigned)le32_to_cpu(ni->type), err);
+		if (err != -ENOMEM)
+			err = -EIO;
+		goto conv_err_out;
+	}
+	/* TODO: Not implemented from here, abort. */
+	read_lock_irqsave(&ni->size_lock, flags);
+	allocated_size = ni->allocated_size;
+	read_unlock_irqrestore(&ni->size_lock, flags);
+	if (start < 0 || start >= allocated_size) {
+		if (err == -ENOSPC)
+			ntfs_error(vol->sb, "Not enough space in the mft "
+					"record/on disk for the non-resident "
+					"attribute value.  This case is not "
+					"implemented yet.");
+		else /* if (err == -EPERM) */
+			ntfs_error(vol->sb, "This attribute type may not be "
+					"non-resident.  This case is not "
+					"implemented yet.");
+	}
+	err = -EOPNOTSUPP;
+	goto conv_err_out;
+#if 0
+	// TODO: Attempt to make other attributes non-resident.
+	if (!err)
+		goto do_resident_extend;
+	/*
+	 * Both the attribute list attribute and the standard information
+	 * attribute must remain in the base inode.  Thus, if this is one of
+	 * these attributes, we have to try to move other attributes out into
+	 * extent mft records instead.
+	 */
+	if (ni->type == AT_ATTRIBUTE_LIST ||
+			ni->type == AT_STANDARD_INFORMATION) {
+		// TODO: Attempt to move other attributes into extent mft
+		// records.
+		err = -EOPNOTSUPP;
+		if (!err)
+			goto do_resident_extend;
+		goto err_out;
+	}
+	// TODO: Attempt to move this attribute to an extent mft record, but
+	// only if it is not already the only attribute in an mft record in
+	// which case there would be nothing to gain.
+	err = -EOPNOTSUPP;
+	if (!err)
+		goto do_resident_extend;
+	/* There is nothing we can do to make enough space. )-: */
+	goto err_out;
+#endif
+do_non_resident_extend:
+	BUG_ON(!NInoNonResident(ni));
+	if (new_alloc_size == allocated_size) {
+		BUG_ON(vcn);
+		goto alloc_done;
+	}
+	/*
+	 * If the data starts after the end of the old allocation, this is a
+	 * $DATA attribute and sparse attributes are enabled on the volume and
+	 * for this inode, then create a sparse region between the old
+	 * allocated size and the start of the data.  Otherwise simply proceed
+	 * with filling the whole space between the old allocated size and the
+	 * new allocated size with clusters.
+	 */
+	if ((start >= 0 && start <= allocated_size) || ni->type != AT_DATA ||
+			!NVolSparseEnabled(vol) || NInoSparseDisabled(ni))
+		goto skip_sparse;
+	// TODO: This is not implemented yet.  We just fill in with real
+	// clusters for now...
+	ntfs_debug("Inserting holes is not-implemented yet.  Falling back to "
+			"allocating real clusters instead.");
+skip_sparse:
+	rl = ni->runlist.rl;
+	if (likely(rl)) {
+		/* Seek to the end of the runlist. */
+		while (rl->length)
+			rl++;
+	}
+	/* If this attribute extent is not mapped, map it now. */
+	if (unlikely(!rl || rl->lcn == LCN_RL_NOT_MAPPED ||
+			(rl->lcn == LCN_ENOENT && rl > ni->runlist.rl &&
+			(rl-1)->lcn == LCN_RL_NOT_MAPPED))) {
+		if (!rl && !allocated_size)
+			goto first_alloc;
+		rl = ntfs_mapping_pairs_decompress(vol, a, ni->runlist.rl);
+		if (IS_ERR(rl)) {
+			err = PTR_ERR(rl);
+			if (start < 0 || start >= allocated_size)
+				ntfs_error(vol->sb, "Cannot extend allocation "
+						"of inode 0x%lx, attribute "
+						"type 0x%x, because the "
+						"mapping of a runlist "
+						"fragment failed with error "
+						"code %i.", vi->i_ino,
+						(unsigned)le32_to_cpu(ni->type),
+						err);
+			if (err != -ENOMEM)
+				err = -EIO;
+			goto err_out;
+		}
+		ni->runlist.rl = rl;
+		/* Seek to the end of the runlist. */
+		while (rl->length)
+			rl++;
+	}
+	/*
+	 * We now know the runlist of the last extent is mapped and @rl is at
+	 * the end of the runlist.  We want to begin allocating clusters
+	 * starting at the last allocated cluster to reduce fragmentation.  If
+	 * there are no valid LCNs in the attribute we let the cluster
+	 * allocator choose the starting cluster.
+	 */
+	/* If the last LCN is a hole or simillar seek back to last real LCN. */
+	while (rl->lcn < 0 && rl > ni->runlist.rl)
+		rl--;
+first_alloc:
+	// FIXME: Need to implement partial allocations so at least part of the
+	// write can be performed when start >= 0.  (Needed for POSIX write(2)
+	// conformance.)
+	rl2 = ntfs_cluster_alloc(vol, allocated_size >> vol->cluster_size_bits,
+			(new_alloc_size - allocated_size) >>
+			vol->cluster_size_bits, (rl && (rl->lcn >= 0)) ?
+			rl->lcn + rl->length : -1, DATA_ZONE, TRUE);
+	if (IS_ERR(rl2)) {
+		err = PTR_ERR(rl2);
+		if (start < 0 || start >= allocated_size)
+			ntfs_error(vol->sb, "Cannot extend allocation of "
+					"inode 0x%lx, attribute type 0x%x, "
+					"because the allocation of clusters "
+					"failed with error code %i.", vi->i_ino,
+					(unsigned)le32_to_cpu(ni->type), err);
+		if (err != -ENOMEM && err != -ENOSPC)
+			err = -EIO;
+		goto err_out;
+	}
+	rl = ntfs_runlists_merge(ni->runlist.rl, rl2);
+	if (IS_ERR(rl)) {
+		err = PTR_ERR(rl);
+		if (start < 0 || start >= allocated_size)
+			ntfs_error(vol->sb, "Cannot extend allocation of "
+					"inode 0x%lx, attribute type 0x%x, "
+					"because the runlist merge failed "
+					"with error code %i.", vi->i_ino,
+					(unsigned)le32_to_cpu(ni->type), err);
+		if (err != -ENOMEM)
+			err = -EIO;
+		if (ntfs_cluster_free_from_rl(vol, rl2)) {
+			ntfs_error(vol->sb, "Failed to release allocated "
+					"cluster(s) in error code path.  Run "
+					"chkdsk to recover the lost "
+					"cluster(s).");
+			NVolSetErrors(vol);
+		}
+		ntfs_free(rl2);
+		goto err_out;
+	}
+	ni->runlist.rl = rl;
+	ntfs_debug("Allocated 0x%llx clusters.", (long long)(new_alloc_size -
+			allocated_size) >> vol->cluster_size_bits);
+	/* Find the runlist element with which the attribute extent starts. */
+	ll = sle64_to_cpu(a->data.non_resident.lowest_vcn);
+	rl2 = ntfs_rl_find_vcn_nolock(rl, ll);
+	BUG_ON(!rl2);
+	BUG_ON(!rl2->length);
+	BUG_ON(rl2->lcn < LCN_HOLE);
+	mp_rebuilt = FALSE;
+	/* Get the size for the new mapping pairs array for this extent. */
+	mp_size = ntfs_get_size_for_mapping_pairs(vol, rl2, ll, -1);
+	if (unlikely(mp_size <= 0)) {
+		err = mp_size;
+		if (start < 0 || start >= allocated_size)
+			ntfs_error(vol->sb, "Cannot extend allocation of "
+					"inode 0x%lx, attribute type 0x%x, "
+					"because determining the size for the "
+					"mapping pairs failed with error code "
+					"%i.", vi->i_ino,
+					(unsigned)le32_to_cpu(ni->type), err);
+		err = -EIO;
+		goto undo_alloc;
+	}
+	/* Extend the attribute record to fit the bigger mapping pairs array. */
+	attr_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 moving this extent to a new mft
+		// record or by starting a new extent in a new mft record,
+		// possibly by extending this extent partially and filling it
+		// and creating a new extent for the remainder, or by making
+		// other attributes non-resident and/or by moving other
+		// attributes out of this mft record.
+		if (start < 0 || start >= allocated_size)
+			ntfs_error(vol->sb, "Not enough space in the mft "
+					"record for the extended attribute "
+					"record.  This case is not "
+					"implemented yet.");
+		err = -EOPNOTSUPP;
+		goto undo_alloc;
+	}
+	mp_rebuilt = TRUE;
+	/* Generate the mapping pairs array directly into the attr record. */
+	err = ntfs_mapping_pairs_build(vol, (u8*)a +
+			le16_to_cpu(a->data.non_resident.mapping_pairs_offset),
+			mp_size, rl2, ll, -1, NULL);
+	if (unlikely(err)) {
+		if (start < 0 || start >= allocated_size)
+			ntfs_error(vol->sb, "Cannot extend allocation of "
+					"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;
+		goto undo_alloc;
+	}
+	/* Update the highest_vcn. */
+	a->data.non_resident.highest_vcn = cpu_to_sle64((new_alloc_size >>
+			vol->cluster_size_bits) - 1);
+	/*
+	 * We now have extended the allocated size of the attribute.  Reflect
+	 * this in the ntfs_inode structure and the attribute record.
+	 */
+	if (a->data.non_resident.lowest_vcn) {
+		/*
+		 * We are not in the first attribute extent, switch to it, but
+		 * first ensure the changes will make it to disk later.
+		 */
+		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))
+			goto restore_undo_alloc;
+		/* @m is not used any more so no need to set it. */
+		a = ctx->attr;
+	}
+	write_lock_irqsave(&ni->size_lock, flags);
+	ni->allocated_size = new_alloc_size;
+	a->data.non_resident.allocated_size = cpu_to_sle64(new_alloc_size);
+	/*
+	 * FIXME: This would fail if @ni is a directory, $MFT, or an index,
+	 * since those can have sparse/compressed set.  For example can be
+	 * set compressed even though it is not compressed itself and in that
+	 * case the bit means that files are to be created compressed in the
+	 * directory...  At present this is ok as this code is only called for
+	 * regular files, and only for their $DATA attribute(s).
+	 * FIXME: The calculation is wrong if we created a hole above.  For now
+	 * it does not matter as we never create holes.
+	 */
+	if (NInoSparse(ni) || NInoCompressed(ni)) {
+		ni->itype.compressed.size += new_alloc_size - allocated_size;
+		a->data.non_resident.compressed_size =
+				cpu_to_sle64(ni->itype.compressed.size);
+		vi->i_blocks = ni->itype.compressed.size >> 9;
+	} else
+		vi->i_blocks = new_alloc_size >> 9;
+	write_unlock_irqrestore(&ni->size_lock, flags);
+alloc_done:
+	if (new_data_size >= 0) {
+		BUG_ON(new_data_size <
+				sle64_to_cpu(a->data.non_resident.data_size));
+		a->data.non_resident.data_size = cpu_to_sle64(new_data_size);
+	}
+flush_done:
+	/* Ensure the changes make it to disk. */
+	flush_dcache_mft_record_page(ctx->ntfs_ino);
+	mark_mft_record_dirty(ctx->ntfs_ino);
+done:
+	ntfs_attr_put_search_ctx(ctx);
+	unmap_mft_record(base_ni);
+	up_write(&ni->runlist.lock);
+	ntfs_debug("Done, new_allocated_size 0x%llx.",
+			(unsigned long long)new_alloc_size);
+	return new_alloc_size;
+restore_undo_alloc:
+	if (start < 0 || start >= allocated_size)
+		ntfs_error(vol->sb, "Cannot complete extension of allocation "
+				"of inode 0x%lx, attribute type 0x%x, because "
+				"lookup of first attribute extent failed with "
+				"error code %i.", vi->i_ino,
+				(unsigned)le32_to_cpu(ni->type), err);
+	if (err == -ENOENT)
+		err = -EIO;
+	ntfs_attr_reinit_search_ctx(ctx);
+	if (ntfs_attr_lookup(ni->type, ni->name, ni->name_len, CASE_SENSITIVE,
+			allocated_size >> vol->cluster_size_bits, NULL, 0,
+			ctx)) {
+		ntfs_error(vol->sb, "Failed to find last attribute extent of "
+				"attribute in error code path.  Run chkdsk to "
+				"recover.");
+		write_lock_irqsave(&ni->size_lock, flags);
+		ni->allocated_size = new_alloc_size;
+		/*
+		 * FIXME: This would fail if @ni is a directory...  See above.
+		 * FIXME: The calculation is wrong if we created a hole above.
+		 * For now it does not matter as we never create holes.
+		 */
+		if (NInoSparse(ni) || NInoCompressed(ni)) {
+			ni->itype.compressed.size += new_alloc_size -
+					allocated_size;
+			vi->i_blocks = ni->itype.compressed.size >> 9;
+		} else
+			vi->i_blocks = new_alloc_size >> 9;
+		write_unlock_irqrestore(&ni->size_lock, flags);
+		ntfs_attr_put_search_ctx(ctx);
+		unmap_mft_record(base_ni);
+		up_write(&ni->runlist.lock);
+		/*
+		 * The only thing that is now wrong is the allocated size of the
+		 * base attribute extent which chkdsk should be able to fix.
+		 */
+		NVolSetErrors(vol);
+		return err;
+	}
+	ctx->attr->data.non_resident.highest_vcn = cpu_to_sle64(
+			(allocated_size >> vol->cluster_size_bits) - 1);
+undo_alloc:
+	ll = allocated_size >> vol->cluster_size_bits;
+	if (ntfs_cluster_free(ni, ll, -1, ctx) < 0) {
+		ntfs_error(vol->sb, "Failed to release allocated cluster(s) "
+				"in error code path.  Run chkdsk to recover "
+				"the lost cluster(s).");
+		NVolSetErrors(vol);
+	}
+	m = ctx->mrec;
+	a = ctx->attr;
+	/*
+	 * If the runlist truncation fails and/or the search context is no
+	 * longer valid, we cannot resize the attribute record or build the
+	 * mapping pairs array thus we mark the inode bad so that no access to
+	 * the freed clusters can happen.
+	 */
+	if (ntfs_rl_truncate_nolock(vol, &ni->runlist, ll) || IS_ERR(m)) {
+		ntfs_error(vol->sb, "Failed to %s in error code path.  Run "
+				"chkdsk to recover.", IS_ERR(m) ?
+				"restore attribute search context" :
+				"truncate attribute runlist");
+		make_bad_inode(vi);
+		make_bad_inode(VFS_I(base_ni));
+		NVolSetErrors(vol);
+	} else if (mp_rebuilt) {
+		if (ntfs_attr_record_resize(m, a, attr_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_len -
+					le16_to_cpu(a->data.non_resident.
+					mapping_pairs_offset), rl2, ll, -1,
+					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);
+		}
+	}
+err_out:
+	if (ctx)
+		ntfs_attr_put_search_ctx(ctx);
+	if (m)
+		unmap_mft_record(base_ni);
+	up_write(&ni->runlist.lock);
+conv_err_out:
+	ntfs_debug("Failed.  Returning error code %i.", err);
+	return err;
+}
+
+/**
  * ntfs_attr_set - fill (a part of) an attribute with a byte
  * @ni:		ntfs inode describing the attribute to fill
  * @ofs:	offset inside the attribute at which to start to fill
@@ -1773,6 +2592,8 @@ int ntfs_attr_set(ntfs_inode *ni, const s64 ofs, const s64 cnt, const u8 val)
 		/* Finally unlock and release the page. */
 		unlock_page(page);
 		page_cache_release(page);
+		balance_dirty_pages_ratelimited(mapping);
+		cond_resched();
 	}
 	/* If there is a last partial page, need to do it the slow way. */
 	if (end_ofs) {
diff --git a/fs/ntfs/attrib.h b/fs/ntfs/attrib.h
index 0618ed6fd7b3..9074886b44ba 100644
--- a/fs/ntfs/attrib.h
+++ b/fs/ntfs/attrib.h
@@ -60,14 +60,15 @@ typedef struct {
 	ATTR_RECORD *base_attr;
 } ntfs_attr_search_ctx;
 
-extern int ntfs_map_runlist_nolock(ntfs_inode *ni, VCN vcn);
+extern int ntfs_map_runlist_nolock(ntfs_inode *ni, VCN vcn,
+		ntfs_attr_search_ctx *ctx);
 extern int ntfs_map_runlist(ntfs_inode *ni, VCN vcn);
 
 extern LCN ntfs_attr_vcn_to_lcn_nolock(ntfs_inode *ni, const VCN vcn,
 		const BOOL write_locked);
 
 extern runlist_element *ntfs_attr_find_vcn_nolock(ntfs_inode *ni,
-		const VCN vcn, const BOOL write_locked);
+		const VCN vcn, ntfs_attr_search_ctx *ctx);
 
 int ntfs_attr_lookup(const ATTR_TYPE type, const ntfschar *name,
 		const u32 name_len, const IGNORE_CASE_BOOL ic,
@@ -102,7 +103,10 @@ extern int ntfs_attr_record_resize(MFT_RECORD *m, ATTR_RECORD *a, u32 new_size);
 extern int ntfs_resident_attr_value_resize(MFT_RECORD *m, ATTR_RECORD *a,
 		const u32 new_size);
 
-extern int ntfs_attr_make_non_resident(ntfs_inode *ni);
+extern int ntfs_attr_make_non_resident(ntfs_inode *ni, const u32 data_size);
+
+extern s64 ntfs_attr_extend_allocation(ntfs_inode *ni, s64 new_alloc_size,
+		const s64 new_data_size, const s64 data_start);
 
 extern int ntfs_attr_set(ntfs_inode *ni, const s64 ofs, const s64 cnt,
 		const u8 val);
diff --git a/fs/ntfs/file.c b/fs/ntfs/file.c
index be9fd1dd423d..cf3e6ced2d01 100644
--- a/fs/ntfs/file.c
+++ b/fs/ntfs/file.c
@@ -19,11 +19,24 @@
  * 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 +69,2184 @@ 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 already uptodate in the page cache then
+ * these are simply marked dirty).
+ *
+ * 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() in
+ * fs/ntfs/inode.c.
+ *
+ * @cached_page and @lru_pvec are just optimizations 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: i_sem on the vfs inode corrseponsind to the ntfs inode @ni must be
+ *	    held by the caller.
+ */
+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%zx.",
+			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);
+				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)) {
+					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%zx.",
+			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 = 0;	/* Offset in the current iovec. */
+	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)) {
+			/*
+			 * Reminder for later: Encrypted files are _always_
+			 * non-resident so that the content can always be
+			 * encrypted.
+			 */
+			ntfs_debug("Denying write access to encrypted file.");
+			return -EACCES;
+		}
+		if (NInoCompressed(ni)) {
+			/* Only unnamed $DATA attribute can be compressed. */
+			BUG_ON(ni->type != AT_DATA);
+			BUG_ON(ni->name_len);
+			/*
+			 * Reminder for later: If resident, the data is not
+			 * actually compressed.  Only on the switch to non-
+			 * resident does compression kick in.  This is in
+			 * contrast to encrypted files (see above).
+			 */
+			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;
+	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 +2304,39 @@ 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 = {
diff --git a/fs/ntfs/inode.c b/fs/ntfs/inode.c
index 7ec045131808..b24f4c4b2c5c 100644
--- a/fs/ntfs/inode.c
+++ b/fs/ntfs/inode.c
@@ -30,6 +30,7 @@
 #include "debug.h"
 #include "inode.h"
 #include "attrib.h"
+#include "lcnalloc.h"
 #include "malloc.h"
 #include "mft.h"
 #include "time.h"
@@ -2291,11 +2292,16 @@ int ntfs_show_options(struct seq_file *sf, struct vfsmount *mnt)
 
 #ifdef NTFS_RW
 
+static const char *es = "  Leaving inconsistent metadata.  Unmount and run "
+		"chkdsk.";
+
 /**
  * ntfs_truncate - called when the i_size of an ntfs inode is changed
  * @vi:		inode for which the i_size was changed
  *
- * We do not support i_size changes yet.
+ * We only support i_size changes for normal files at present, i.e. not
+ * compressed and not encrypted.  This is enforced in ntfs_setattr(), see
+ * below.
  *
  * The kernel guarantees that @vi is a regular file (S_ISREG() is true) and
  * that the change is allowed.
@@ -2306,80 +2312,499 @@ int ntfs_show_options(struct seq_file *sf, struct vfsmount *mnt)
  * Returns 0 on success or -errno on error.
  *
  * Called with ->i_sem held.  In all but one case ->i_alloc_sem is held for
- * writing.  The only case where ->i_alloc_sem is not held is
+ * writing.  The only case in the kernel where ->i_alloc_sem is not held is
  * mm/filemap.c::generic_file_buffered_write() where vmtruncate() is called
- * with the current i_size as the offset which means that it is a noop as far
- * as ntfs_truncate() is concerned.
+ * with the current i_size as the offset.  The analogous place in NTFS is in
+ * fs/ntfs/file.c::ntfs_file_buffered_write() where we call vmtruncate() again
+ * without holding ->i_alloc_sem.
  */
 int ntfs_truncate(struct inode *vi)
 {
-	ntfs_inode *ni = NTFS_I(vi);
+	s64 new_size, old_size, nr_freed, new_alloc_size, old_alloc_size;
+	VCN highest_vcn;
+	unsigned long flags;
+	ntfs_inode *base_ni, *ni = NTFS_I(vi);
 	ntfs_volume *vol = ni->vol;
 	ntfs_attr_search_ctx *ctx;
 	MFT_RECORD *m;
 	ATTR_RECORD *a;
 	const char *te = "  Leaving file length out of sync with i_size.";
-	int err;
+	int err, mp_size, size_change, alloc_change;
+	u32 attr_len;
 
 	ntfs_debug("Entering for inode 0x%lx.", vi->i_ino);
 	BUG_ON(NInoAttr(ni));
+	BUG_ON(S_ISDIR(vi->i_mode));
+	BUG_ON(NInoMstProtected(ni));
 	BUG_ON(ni->nr_extents < 0);
-	m = map_mft_record(ni);
+retry_truncate:
+	/*
+	 * Lock the runlist for writing and map the mft record to ensure it is
+	 * safe to mess with the attribute runlist and sizes.
+	 */
+	down_write(&ni->runlist.lock);
+	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);
 		ntfs_error(vi->i_sb, "Failed to map mft record for inode 0x%lx "
 				"(error code %d).%s", vi->i_ino, err, te);
 		ctx = NULL;
 		m = NULL;
-		goto err_out;
+		goto old_bad_out;
 	}
-	ctx = ntfs_attr_get_search_ctx(ni, m);
+	ctx = ntfs_attr_get_search_ctx(base_ni, m);
 	if (unlikely(!ctx)) {
 		ntfs_error(vi->i_sb, "Failed to allocate a search context for "
 				"inode 0x%lx (not enough memory).%s",
 				vi->i_ino, te);
 		err = -ENOMEM;
-		goto err_out;
+		goto old_bad_out;
 	}
 	err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
 			CASE_SENSITIVE, 0, NULL, 0, ctx);
 	if (unlikely(err)) {
-		if (err == -ENOENT)
+		if (err == -ENOENT) {
 			ntfs_error(vi->i_sb, "Open attribute is missing from "
 					"mft record.  Inode 0x%lx is corrupt.  "
-					"Run chkdsk.", vi->i_ino);
-		else
+					"Run chkdsk.%s", vi->i_ino, te);
+			err = -EIO;
+		} else
 			ntfs_error(vi->i_sb, "Failed to lookup attribute in "
-					"inode 0x%lx (error code %d).",
-					vi->i_ino, err);
-		goto err_out;
+					"inode 0x%lx (error code %d).%s",
+					vi->i_ino, err, te);
+		goto old_bad_out;
 	}
+	m = ctx->mrec;
 	a = ctx->attr;
-	/* If the size has not changed there is nothing to do. */
-	if (ntfs_attr_size(a) == i_size_read(vi))
-		goto done;
-	// TODO: Implement the truncate...
-	ntfs_error(vi->i_sb, "Inode size has changed but this is not "
-			"implemented yet.  Resetting inode size to old value. "
-			" This is most likely a bug in the ntfs driver!");
-	i_size_write(vi, ntfs_attr_size(a)); 
-done:
+	/*
+	 * The i_size of the vfs inode is the new size for the attribute value.
+	 */
+	new_size = i_size_read(vi);
+	/* The current size of the attribute value is the old size. */
+	old_size = ntfs_attr_size(a);
+	/* Calculate the new allocated size. */
+	if (NInoNonResident(ni))
+		new_alloc_size = (new_size + vol->cluster_size - 1) &
+				~(s64)vol->cluster_size_mask;
+	else
+		new_alloc_size = (new_size + 7) & ~7;
+	/* The current allocated size is the old allocated size. */
+	read_lock_irqsave(&ni->size_lock, flags);
+	old_alloc_size = ni->allocated_size;
+	read_unlock_irqrestore(&ni->size_lock, flags);
+	/*
+	 * The change in the file size.  This will be 0 if no change, >0 if the
+	 * size is growing, and <0 if the size is shrinking.
+	 */
+	size_change = -1;
+	if (new_size - old_size >= 0) {
+		size_change = 1;
+		if (new_size == old_size)
+			size_change = 0;
+	}
+	/* As above for the allocated size. */
+	alloc_change = -1;
+	if (new_alloc_size - old_alloc_size >= 0) {
+		alloc_change = 1;
+		if (new_alloc_size == old_alloc_size)
+			alloc_change = 0;
+	}
+	/*
+	 * If neither the size nor the allocation are being changed there is
+	 * nothing to do.
+	 */
+	if (!size_change && !alloc_change)
+		goto unm_done;
+	/* If the size is changing, check if new size is allowed in $AttrDef. */
+	if (size_change) {
+		err = ntfs_attr_size_bounds_check(vol, ni->type, new_size);
+		if (unlikely(err)) {
+			if (err == -ERANGE) {
+				ntfs_error(vol->sb, "Truncate would cause the "
+						"inode 0x%lx to %simum size "
+						"for its attribute type "
+						"(0x%x).  Aborting truncate.",
+						vi->i_ino,
+						new_size > old_size ? "exceed "
+						"the max" : "go under the min",
+						le32_to_cpu(ni->type));
+				err = -EFBIG;
+			} else {
+				ntfs_error(vol->sb, "Inode 0x%lx has unknown "
+						"attribute type 0x%x.  "
+						"Aborting truncate.",
+						vi->i_ino,
+						le32_to_cpu(ni->type));
+				err = -EIO;
+			}
+			/* Reset the vfs inode size to the old size. */
+			i_size_write(vi, old_size);
+			goto err_out;
+		}
+	}
+	if (NInoCompressed(ni) || NInoEncrypted(ni)) {
+		ntfs_warning(vi->i_sb, "Changes in inode size are not "
+				"supported yet for %s files, ignoring.",
+				NInoCompressed(ni) ? "compressed" :
+				"encrypted");
+		err = -EOPNOTSUPP;
+		goto bad_out;
+	}
+	if (a->non_resident)
+		goto do_non_resident_truncate;
+	BUG_ON(NInoNonResident(ni));
+	/* Resize the attribute record to best fit the new attribute size. */
+	if (new_size < vol->mft_record_size &&
+			!ntfs_resident_attr_value_resize(m, a, new_size)) {
+		unsigned long flags;
+
+		/* The resize succeeded! */
+		flush_dcache_mft_record_page(ctx->ntfs_ino);
+		mark_mft_record_dirty(ctx->ntfs_ino);
+		write_lock_irqsave(&ni->size_lock, flags);
+		/* Update the sizes in the ntfs inode and all is done. */
+		ni->allocated_size = le32_to_cpu(a->length) -
+				le16_to_cpu(a->data.resident.value_offset);
+		/*
+		 * Note ntfs_resident_attr_value_resize() has already done any
+		 * necessary data clearing in the attribute record.  When the
+		 * file is being shrunk vmtruncate() will already have cleared
+		 * the top part of the last partial page, i.e. since this is
+		 * the resident case this is the page with index 0.  However,
+		 * when the file is being expanded, the page cache page data
+		 * between the old data_size, i.e. old_size, and the new_size
+		 * has not been zeroed.  Fortunately, we do not need to zero it
+		 * either since on one hand it will either already be zero due
+		 * to both readpage and writepage clearing partial page data
+		 * beyond i_size in which case there is nothing to do or in the
+		 * case of the file being mmap()ped at the same time, POSIX
+		 * specifies that the behaviour is unspecified thus we do not
+		 * have to do anything.  This means that in our implementation
+		 * in the rare case that the file is mmap()ped and a write
+		 * occured into the mmap()ped region just beyond the file size
+		 * and writepage has not yet been called to write out the page
+		 * (which would clear the area beyond the file size) and we now
+		 * extend the file size to incorporate this dirty region
+		 * outside the file size, a write of the page would result in
+		 * this data being written to disk instead of being cleared.
+		 * Given both POSIX and the Linux mmap(2) man page specify that
+		 * this corner case is undefined, we choose to leave it like
+		 * that as this is much simpler for us as we cannot lock the
+		 * relevant page now since we are holding too many ntfs locks
+		 * which would result in a lock reversal deadlock.
+		 */
+		ni->initialized_size = new_size;
+		write_unlock_irqrestore(&ni->size_lock, flags);
+		goto unm_done;
+	}
+	/* If the above resize failed, this must be an attribute extension. */
+	BUG_ON(size_change < 0);
+	/*
+	 * We have to drop all the locks so we can call
+	 * ntfs_attr_make_non_resident().  This could be optimised by try-
+	 * locking the first page cache page and only if that fails dropping
+	 * the locks, locking the page, and redoing all the locking and
+	 * lookups.  While this would be a huge optimisation, it is not worth
+	 * it as this is definitely a slow code path as it only ever can happen
+	 * once for any given file.
+	 */
 	ntfs_attr_put_search_ctx(ctx);
-	unmap_mft_record(ni);
-	NInoClearTruncateFailed(ni);
-	ntfs_debug("Done.");
-	return 0;
-err_out:
-	if (err != -ENOMEM) {
+	unmap_mft_record(base_ni);
+	up_write(&ni->runlist.lock);
+	/*
+	 * Not enough space in the mft record, try to make the attribute
+	 * non-resident and if successful restart the truncation process.
+	 */
+	err = ntfs_attr_make_non_resident(ni, old_size);
+	if (likely(!err))
+		goto retry_truncate;
+	/*
+	 * Could not make non-resident.  If this is due to this not being
+	 * permitted for this attribute type or there not being enough space,
+	 * try to make other attributes non-resident.  Otherwise fail.
+	 */
+	if (unlikely(err != -EPERM && err != -ENOSPC)) {
+		ntfs_error(vol->sb, "Cannot truncate inode 0x%lx, attribute "
+				"type 0x%x, because the conversion from "
+				"resident to non-resident attribute failed "
+				"with error code %i.", vi->i_ino,
+				(unsigned)le32_to_cpu(ni->type), err);
+		if (err != -ENOMEM)
+			err = -EIO;
+		goto conv_err_out;
+	}
+	/* TODO: Not implemented from here, abort. */
+	if (err == -ENOSPC)
+		ntfs_error(vol->sb, "Not enough space in the mft record/on "
+				"disk for the non-resident attribute value.  "
+				"This case is not implemented yet.");
+	else /* if (err == -EPERM) */
+		ntfs_error(vol->sb, "This attribute type may not be "
+				"non-resident.  This case is not implemented "
+				"yet.");
+	err = -EOPNOTSUPP;
+	goto conv_err_out;
+#if 0
+	// TODO: Attempt to make other attributes non-resident.
+	if (!err)
+		goto do_resident_extend;
+	/*
+	 * Both the attribute list attribute and the standard information
+	 * attribute must remain in the base inode.  Thus, if this is one of
+	 * these attributes, we have to try to move other attributes out into
+	 * extent mft records instead.
+	 */
+	if (ni->type == AT_ATTRIBUTE_LIST ||
+			ni->type == AT_STANDARD_INFORMATION) {
+		// TODO: Attempt to move other attributes into extent mft
+		// records.
+		err = -EOPNOTSUPP;
+		if (!err)
+			goto do_resident_extend;
+		goto err_out;
+	}
+	// TODO: Attempt to move this attribute to an extent mft record, but
+	// only if it is not already the only attribute in an mft record in
+	// which case there would be nothing to gain.
+	err = -EOPNOTSUPP;
+	if (!err)
+		goto do_resident_extend;
+	/* There is nothing we can do to make enough space. )-: */
+	goto err_out;
+#endif
+do_non_resident_truncate:
+	BUG_ON(!NInoNonResident(ni));
+	if (alloc_change < 0) {
+		highest_vcn = sle64_to_cpu(a->data.non_resident.highest_vcn);
+		if (highest_vcn > 0 &&
+				old_alloc_size >> vol->cluster_size_bits >
+				highest_vcn + 1) {
+			/*
+			 * This attribute has multiple extents.  Not yet
+			 * supported.
+			 */
+			ntfs_error(vol->sb, "Cannot truncate inode 0x%lx, "
+					"attribute type 0x%x, because the "
+					"attribute is highly fragmented (it "
+					"consists of multiple extents) and "
+					"this case is not implemented yet.",
+					vi->i_ino,
+					(unsigned)le32_to_cpu(ni->type));
+			err = -EOPNOTSUPP;
+			goto bad_out;
+		}
+	}
+	/*
+	 * If the size is shrinking, need to reduce the initialized_size and
+	 * the data_size before reducing the allocation.
+	 */
+	if (size_change < 0) {
+		/*
+		 * Make the valid size smaller (i_size is already up-to-date).
+		 */
+		write_lock_irqsave(&ni->size_lock, flags);
+		if (new_size < ni->initialized_size) {
+			ni->initialized_size = new_size;
+			a->data.non_resident.initialized_size =
+					cpu_to_sle64(new_size);
+		}
+		a->data.non_resident.data_size = cpu_to_sle64(new_size);
+		write_unlock_irqrestore(&ni->size_lock, flags);
+		flush_dcache_mft_record_page(ctx->ntfs_ino);
+		mark_mft_record_dirty(ctx->ntfs_ino);
+		/* If the allocated size is not changing, we are done. */
+		if (!alloc_change)
+			goto unm_done;
+		/*
+		 * If the size is shrinking it makes no sense for the
+		 * allocation to be growing.
+		 */
+		BUG_ON(alloc_change > 0);
+	} else /* if (size_change >= 0) */ {
+		/*
+		 * The file size is growing or staying the same but the
+		 * allocation can be shrinking, growing or staying the same.
+		 */
+		if (alloc_change > 0) {
+			/*
+			 * We need to extend the allocation and possibly update
+			 * the data size.  If we are updating the data size,
+			 * since we are not touching the initialized_size we do
+			 * not need to worry about the actual data on disk.
+			 * And as far as the page cache is concerned, there
+			 * will be no pages beyond the old data size and any
+			 * partial region in the last page between the old and
+			 * new data size (or the end of the page if the new
+			 * data size is outside the page) does not need to be
+			 * modified as explained above for the resident
+			 * attribute truncate case.  To do this, we simply drop
+			 * the locks we hold and leave all the work to our
+			 * friendly helper ntfs_attr_extend_allocation().
+			 */
+			ntfs_attr_put_search_ctx(ctx);
+			unmap_mft_record(base_ni);
+			up_write(&ni->runlist.lock);
+			err = ntfs_attr_extend_allocation(ni, new_size,
+					size_change > 0 ? new_size : -1, -1);
+			/*
+			 * ntfs_attr_extend_allocation() will have done error
+			 * output already.
+			 */
+			goto done;
+		}
+		if (!alloc_change)
+			goto alloc_done;
+	}
+	/* alloc_change < 0 */
+	/* Free the clusters. */
+	nr_freed = ntfs_cluster_free(ni, new_alloc_size >>
+			vol->cluster_size_bits, -1, ctx);
+	m = ctx->mrec;
+	a = ctx->attr;
+	if (unlikely(nr_freed < 0)) {
+		ntfs_error(vol->sb, "Failed to release cluster(s) (error code "
+				"%lli).  Unmount and run chkdsk to recover "
+				"the lost cluster(s).", (long long)nr_freed);
 		NVolSetErrors(vol);
+		nr_freed = 0;
+	}
+	/* Truncate the runlist. */
+	err = ntfs_rl_truncate_nolock(vol, &ni->runlist,
+			new_alloc_size >> vol->cluster_size_bits);
+	/*
+	 * If the runlist truncation failed and/or the search context is no
+	 * longer valid, we cannot resize the attribute record or build the
+	 * mapping pairs array thus we mark the inode bad so that no access to
+	 * the freed clusters can happen.
+	 */
+	if (unlikely(err || IS_ERR(m))) {
+		ntfs_error(vol->sb, "Failed to %s (error code %li).%s",
+				IS_ERR(m) ?
+				"restore attribute search context" :
+				"truncate attribute runlist",
+				IS_ERR(m) ? PTR_ERR(m) : err, es);
+		err = -EIO;
+		goto bad_out;
+	}
+	/* Get the size for the shrunk mapping pairs array for the runlist. */
+	mp_size = ntfs_get_size_for_mapping_pairs(vol, ni->runlist.rl, 0, -1);
+	if (unlikely(mp_size <= 0)) {
+		ntfs_error(vol->sb, "Cannot shrink allocation of inode 0x%lx, "
+				"attribute type 0x%x, because determining the "
+				"size for the mapping pairs failed with error "
+				"code %i.%s", vi->i_ino,
+				(unsigned)le32_to_cpu(ni->type), mp_size, es);
+		err = -EIO;
+		goto bad_out;
+	}
+	/*
+	 * Shrink the attribute record for the new mapping pairs array.  Note,
+	 * this cannot fail since we are making the attribute smaller thus by
+	 * definition there is enough space to do so.
+	 */
+	attr_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));
+	BUG_ON(err);
+	/*
+	 * 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, ni->runlist.rl, 0, -1, NULL);
+	if (unlikely(err)) {
+		ntfs_error(vol->sb, "Cannot shrink allocation of inode 0x%lx, "
+				"attribute type 0x%x, because building the "
+				"mapping pairs failed with error code %i.%s",
+				vi->i_ino, (unsigned)le32_to_cpu(ni->type),
+				err, es);
+		err = -EIO;
+		goto bad_out;
+	}
+	/* Update the allocated/compressed size as well as the highest vcn. */
+	a->data.non_resident.highest_vcn = cpu_to_sle64((new_alloc_size >>
+			vol->cluster_size_bits) - 1);
+	write_lock_irqsave(&ni->size_lock, flags);
+	ni->allocated_size = new_alloc_size;
+	a->data.non_resident.allocated_size = cpu_to_sle64(new_alloc_size);
+	if (NInoSparse(ni) || NInoCompressed(ni)) {
+		if (nr_freed) {
+			ni->itype.compressed.size -= nr_freed <<
+					vol->cluster_size_bits;
+			BUG_ON(ni->itype.compressed.size < 0);
+			a->data.non_resident.compressed_size = cpu_to_sle64(
+					ni->itype.compressed.size);
+			vi->i_blocks = ni->itype.compressed.size >> 9;
+		}
+	} else
+		vi->i_blocks = new_alloc_size >> 9;
+	write_unlock_irqrestore(&ni->size_lock, flags);
+	/*
+	 * We have shrunk the allocation.  If this is a shrinking truncate we
+	 * have already dealt with the initialized_size and the data_size above
+	 * and we are done.  If the truncate is only changing the allocation
+	 * and not the data_size, we are also done.  If this is an extending
+	 * truncate, need to extend the data_size now which is ensured by the
+	 * fact that @size_change is positive.
+	 */
+alloc_done:
+	/*
+	 * If the size is growing, need to update it now.  If it is shrinking,
+	 * we have already updated it above (before the allocation change).
+	 */
+	if (size_change > 0)
+		a->data.non_resident.data_size = cpu_to_sle64(new_size);
+	/* Ensure the modified mft record is written out. */
+	flush_dcache_mft_record_page(ctx->ntfs_ino);
+	mark_mft_record_dirty(ctx->ntfs_ino);
+unm_done:
+	ntfs_attr_put_search_ctx(ctx);
+	unmap_mft_record(base_ni);
+	up_write(&ni->runlist.lock);
+done:
+	/* Update the mtime and ctime on the base inode. */
+	inode_update_time(VFS_I(base_ni), 1);
+	if (likely(!err)) {
+		NInoClearTruncateFailed(ni);
+		ntfs_debug("Done.");
+	}
+	return err;
+old_bad_out:
+	old_size = -1;
+bad_out:
+	if (err != -ENOMEM && err != -EOPNOTSUPP) {
 		make_bad_inode(vi);
+		make_bad_inode(VFS_I(base_ni));
+		NVolSetErrors(vol);
 	}
+	if (err != -EOPNOTSUPP)
+		NInoSetTruncateFailed(ni);
+	else if (old_size >= 0)
+		i_size_write(vi, old_size);
+err_out:
 	if (ctx)
 		ntfs_attr_put_search_ctx(ctx);
 	if (m)
-		unmap_mft_record(ni);
-	NInoSetTruncateFailed(ni);
+		unmap_mft_record(base_ni);
+	up_write(&ni->runlist.lock);
+out:
+	ntfs_debug("Failed.  Returning error code %i.", err);
 	return err;
+conv_err_out:
+	if (err != -ENOMEM && err != -EOPNOTSUPP) {
+		make_bad_inode(vi);
+		make_bad_inode(VFS_I(base_ni));
+		NVolSetErrors(vol);
+	}
+	if (err != -EOPNOTSUPP)
+		NInoSetTruncateFailed(ni);
+	else
+		i_size_write(vi, old_size);
+	goto out;
 }
 
 /**
@@ -2420,8 +2845,7 @@ int ntfs_setattr(struct dentry *dentry, struct iattr *attr)
 
 	err = inode_change_ok(vi, attr);
 	if (err)
-		return err;
-
+		goto out;
 	/* We do not support NTFS ACLs yet. */
 	if (ia_valid & (ATTR_UID | ATTR_GID | ATTR_MODE)) {
 		ntfs_warning(vi->i_sb, "Changes in user/group/mode are not "
@@ -2429,14 +2853,22 @@ int ntfs_setattr(struct dentry *dentry, struct iattr *attr)
 		err = -EOPNOTSUPP;
 		goto out;
 	}
-
 	if (ia_valid & ATTR_SIZE) {
 		if (attr->ia_size != i_size_read(vi)) {
-			ntfs_warning(vi->i_sb, "Changes in inode size are not "
-					"supported yet, ignoring.");
-			err = -EOPNOTSUPP;
-			// TODO: Implement...
-			// err = vmtruncate(vi, attr->ia_size);
+			ntfs_inode *ni = NTFS_I(vi);
+			/*
+			 * FIXME: For now we do not support resizing of
+			 * compressed or encrypted files yet.
+			 */
+			if (NInoCompressed(ni) || NInoEncrypted(ni)) {
+				ntfs_warning(vi->i_sb, "Changes in inode size "
+						"are not supported yet for "
+						"%s files, ignoring.",
+						NInoCompressed(ni) ?
+						"compressed" : "encrypted");
+				err = -EOPNOTSUPP;
+			} else
+				err = vmtruncate(vi, attr->ia_size);
 			if (err || ia_valid == ATTR_SIZE)
 				goto out;
 		} else {
diff --git a/fs/ntfs/layout.h b/fs/ntfs/layout.h
index 5c248d404f05..f5678d5d7919 100644
--- a/fs/ntfs/layout.h
+++ b/fs/ntfs/layout.h
@@ -1021,10 +1021,17 @@ enum {
 	FILE_NAME_POSIX		= 0x00,
 	/* This is the largest namespace. It is case sensitive and allows all
 	   Unicode characters except for: '\0' and '/'.  Beware that in
-	   WinNT/2k files which eg have the same name except for their case
-	   will not be distinguished by the standard utilities and thus a "del
-	   filename" will delete both "filename" and "fileName" without
-	   warning. */
+	   WinNT/2k/2003 by default files which eg have the same name except
+	   for their case will not be distinguished by the standard utilities
+	   and thus a "del filename" will delete both "filename" and "fileName"
+	   without warning.  However if for example Services For Unix (SFU) are
+	   installed and the case sensitive option was enabled at installation
+	   time, then you can create/access/delete such files.
+	   Note that even SFU places restrictions on the filenames beyond the
+	   '\0' and '/' and in particular the following set of characters is
+	   not allowed: '"', '/', '<', '>', '\'.  All other characters,
+	   including the ones no allowed in WIN32 namespace are allowed.
+	   Tested with SFU 3.5 (this is now free) running on Windows XP. */
 	FILE_NAME_WIN32		= 0x01,
 	/* The standard WinNT/2k NTFS long filenames. Case insensitive.  All
 	   Unicode chars except: '\0', '"', '*', '/', ':', '<', '>', '?', '\',
@@ -2367,7 +2374,9 @@ typedef struct {
  * Extended attribute flags (8-bit).
  */
 enum {
-	NEED_EA	= 0x80
+	NEED_EA	= 0x80		/* If set the file to which the EA belongs
+				   cannot be interpreted without understanding
+				   the associates extended attributes. */
 } __attribute__ ((__packed__));
 
 typedef u8 EA_FLAGS;
@@ -2375,20 +2384,20 @@ typedef u8 EA_FLAGS;
 /*
  * Attribute: Extended attribute (EA) (0xe0).
  *
- * NOTE: Always non-resident. (Is this true?)
+ * NOTE: Can be resident or non-resident.
  *
  * Like the attribute list and the index buffer list, the EA attribute value is
  * a sequence of EA_ATTR variable length records.
- *
- * FIXME: It appears weird that the EA name is not unicode. Is it true?
  */
 typedef struct {
 	le32 next_entry_offset;	/* Offset to the next EA_ATTR. */
 	EA_FLAGS flags;		/* Flags describing the EA. */
-	u8 ea_name_length;	/* Length of the name of the EA in bytes. */
+	u8 ea_name_length;	/* Length of the name of the EA in bytes
+				   excluding the '\0' byte terminator. */
 	le16 ea_value_length;	/* Byte size of the EA's value. */
-	u8 ea_name[0];		/* Name of the EA. */
-	u8 ea_value[0];		/* The value of the EA. Immediately follows
+	u8 ea_name[0];		/* Name of the EA.  Note this is ASCII, not
+				   Unicode and it is zero terminated. */
+	u8 ea_value[0];		/* The value of the EA.  Immediately follows
 				   the name. */
 } __attribute__ ((__packed__)) EA_ATTR;
 
diff --git a/fs/ntfs/lcnalloc.c b/fs/ntfs/lcnalloc.c
index 5af3bf0b7eee..29cabf93d2d2 100644
--- a/fs/ntfs/lcnalloc.c
+++ b/fs/ntfs/lcnalloc.c
@@ -76,6 +76,7 @@ int ntfs_cluster_free_from_rl_nolock(ntfs_volume *vol,
  * @count:	number of clusters to allocate
  * @start_lcn:	starting lcn at which to allocate the clusters (or -1 if none)
  * @zone:	zone from which to allocate the clusters
+ * @is_extension:	if TRUE, this is an attribute extension
  *
  * Allocate @count clusters preferably starting at cluster @start_lcn or at the
  * current allocator position if @start_lcn is -1, on the mounted ntfs volume
@@ -86,6 +87,13 @@ int ntfs_cluster_free_from_rl_nolock(ntfs_volume *vol,
  * @start_vcn specifies the vcn of the first allocated cluster.  This makes
  * merging the resulting runlist with the old runlist easier.
  *
+ * If @is_extension is TRUE, the caller is allocating clusters to extend an
+ * attribute and if it is FALSE, the caller is allocating clusters to fill a
+ * hole in an attribute.  Practically the difference is that if @is_extension
+ * is TRUE the returned runlist will be terminated with LCN_ENOENT and if
+ * @is_extension is FALSE the runlist will be terminated with
+ * LCN_RL_NOT_MAPPED.
+ *
  * You need to check the return value with IS_ERR().  If this is false, the
  * function was successful and the return value is a runlist describing the
  * allocated cluster(s).  If IS_ERR() is true, the function failed and
@@ -137,7 +145,8 @@ int ntfs_cluster_free_from_rl_nolock(ntfs_volume *vol,
  */
 runlist_element *ntfs_cluster_alloc(ntfs_volume *vol, const VCN start_vcn,
 		const s64 count, const LCN start_lcn,
-		const NTFS_CLUSTER_ALLOCATION_ZONES zone)
+		const NTFS_CLUSTER_ALLOCATION_ZONES zone,
+		const BOOL is_extension)
 {
 	LCN zone_start, zone_end, bmp_pos, bmp_initial_pos, last_read_pos, lcn;
 	LCN prev_lcn = 0, prev_run_len = 0, mft_zone_size;
@@ -310,7 +319,7 @@ runlist_element *ntfs_cluster_alloc(ntfs_volume *vol, const VCN start_vcn,
 				continue;
 			}
 			bit = 1 << (lcn & 7);
-			ntfs_debug("bit %i.", bit);
+			ntfs_debug("bit 0x%x.", bit);
 			/* If the bit is already set, go onto the next one. */
 			if (*byte & bit) {
 				lcn++;
@@ -729,7 +738,7 @@ out:
 	/* Add runlist terminator element. */
 	if (likely(rl)) {
 		rl[rlpos].vcn = rl[rlpos - 1].vcn + rl[rlpos - 1].length;
-		rl[rlpos].lcn = LCN_RL_NOT_MAPPED;
+		rl[rlpos].lcn = is_extension ? LCN_ENOENT : LCN_RL_NOT_MAPPED;
 		rl[rlpos].length = 0;
 	}
 	if (likely(page && !IS_ERR(page))) {
@@ -782,6 +791,7 @@ out:
  * @ni:		ntfs inode whose runlist describes the clusters to free
  * @start_vcn:	vcn in the runlist of @ni at which to start freeing clusters
  * @count:	number of clusters to free or -1 for all clusters
+ * @ctx:	active attribute search context if present or NULL if not
  * @is_rollback:	true if this is a rollback operation
  *
  * Free @count clusters starting at the cluster @start_vcn in the runlist
@@ -791,15 +801,39 @@ out:
  * deallocated.  Thus, to completely free all clusters in a runlist, use
  * @start_vcn = 0 and @count = -1.
  *
+ * If @ctx is specified, it is an active search context of @ni and its base mft
+ * record.  This is needed when __ntfs_cluster_free() encounters unmapped
+ * runlist fragments and allows their mapping.  If you do not have the mft
+ * record mapped, you can specify @ctx as NULL and __ntfs_cluster_free() will
+ * perform the necessary mapping and unmapping.
+ *
+ * Note, __ntfs_cluster_free() saves the state of @ctx on entry and restores it
+ * before returning.  Thus, @ctx will be left pointing to the same attribute on
+ * return as on entry.  However, the actual pointers in @ctx may point to
+ * different memory locations on return, so you must remember to reset any
+ * cached pointers from the @ctx, i.e. after the call to __ntfs_cluster_free(),
+ * you will probably want to do:
+ *	m = ctx->mrec;
+ *	a = ctx->attr;
+ * Assuming you cache ctx->attr in a variable @a of type ATTR_RECORD * and that
+ * you cache ctx->mrec in a variable @m of type MFT_RECORD *.
+ *
  * @is_rollback should always be FALSE, it is for internal use to rollback
  * errors.  You probably want to use ntfs_cluster_free() instead.
  *
- * Note, ntfs_cluster_free() does not modify the runlist at all, so the caller
- * has to deal with it later.
+ * Note, __ntfs_cluster_free() does not modify the runlist, so you have to
+ * remove from the runlist or mark sparse the freed runs later.
  *
  * Return the number of deallocated clusters (not counting sparse ones) on
  * success and -errno on error.
  *
+ * WARNING: If @ctx is supplied, regardless of whether success or failure is
+ *	    returned, you need to check IS_ERR(@ctx->mrec) and if TRUE the @ctx
+ *	    is no longer valid, i.e. you need to either call
+ *	    ntfs_attr_reinit_search_ctx() or ntfs_attr_put_search_ctx() on it.
+ *	    In that case PTR_ERR(@ctx->mrec) will give you the error code for
+ *	    why the mapping of the old inode failed.
+ *
  * Locking: - The runlist described by @ni must be locked for writing on entry
  *	      and is locked on return.  Note the runlist may be modified when
  *	      needed runlist fragments need to be mapped.
@@ -807,9 +841,13 @@ out:
  *	      on return.
  *	    - This function takes the volume lcn bitmap lock for writing and
  *	      modifies the bitmap contents.
+ *	    - If @ctx is NULL, the base mft record of @ni must not be mapped on
+ *	      entry and it will be left unmapped on return.
+ *	    - If @ctx is not NULL, the base mft record must be mapped on entry
+ *	      and it will be left mapped on return.
  */
 s64 __ntfs_cluster_free(ntfs_inode *ni, const VCN start_vcn, s64 count,
-		const BOOL is_rollback)
+		ntfs_attr_search_ctx *ctx, const BOOL is_rollback)
 {
 	s64 delta, to_free, total_freed, real_freed;
 	ntfs_volume *vol;
@@ -839,7 +877,7 @@ s64 __ntfs_cluster_free(ntfs_inode *ni, const VCN start_vcn, s64 count,
 
 	total_freed = real_freed = 0;
 
-	rl = ntfs_attr_find_vcn_nolock(ni, start_vcn, TRUE);
+	rl = ntfs_attr_find_vcn_nolock(ni, start_vcn, ctx);
 	if (IS_ERR(rl)) {
 		if (!is_rollback)
 			ntfs_error(vol->sb, "Failed to find first runlist "
@@ -893,7 +931,7 @@ s64 __ntfs_cluster_free(ntfs_inode *ni, const VCN start_vcn, s64 count,
 
 			/* Attempt to map runlist. */
 			vcn = rl->vcn;
-			rl = ntfs_attr_find_vcn_nolock(ni, vcn, TRUE);
+			rl = ntfs_attr_find_vcn_nolock(ni, vcn, ctx);
 			if (IS_ERR(rl)) {
 				err = PTR_ERR(rl);
 				if (!is_rollback)
@@ -961,7 +999,7 @@ err_out:
 	 * If rollback fails, set the volume errors flag, emit an error
 	 * message, and return the error code.
 	 */
-	delta = __ntfs_cluster_free(ni, start_vcn, total_freed, TRUE);
+	delta = __ntfs_cluster_free(ni, start_vcn, total_freed, ctx, TRUE);
 	if (delta < 0) {
 		ntfs_error(vol->sb, "Failed to rollback (error %i).  Leaving "
 				"inconsistent metadata!  Unmount and run "
diff --git a/fs/ntfs/lcnalloc.h b/fs/ntfs/lcnalloc.h
index a6a8827882e7..72cbca7003b2 100644
--- a/fs/ntfs/lcnalloc.h
+++ b/fs/ntfs/lcnalloc.h
@@ -27,6 +27,7 @@
 
 #include <linux/fs.h>
 
+#include "attrib.h"
 #include "types.h"
 #include "inode.h"
 #include "runlist.h"
@@ -41,16 +42,18 @@ typedef enum {
 
 extern runlist_element *ntfs_cluster_alloc(ntfs_volume *vol,
 		const VCN start_vcn, const s64 count, const LCN start_lcn,
-		const NTFS_CLUSTER_ALLOCATION_ZONES zone);
+		const NTFS_CLUSTER_ALLOCATION_ZONES zone,
+		const BOOL is_extension);
 
 extern s64 __ntfs_cluster_free(ntfs_inode *ni, const VCN start_vcn,
-		s64 count, const BOOL is_rollback);
+		s64 count, ntfs_attr_search_ctx *ctx, const BOOL is_rollback);
 
 /**
  * ntfs_cluster_free - free clusters on an ntfs volume
  * @ni:		ntfs inode whose runlist describes the clusters to free
  * @start_vcn:	vcn in the runlist of @ni at which to start freeing clusters
  * @count:	number of clusters to free or -1 for all clusters
+ * @ctx:	active attribute search context if present or NULL if not
  *
  * Free @count clusters starting at the cluster @start_vcn in the runlist
  * described by the ntfs inode @ni.
@@ -59,12 +62,36 @@ extern s64 __ntfs_cluster_free(ntfs_inode *ni, const VCN start_vcn,
  * deallocated.  Thus, to completely free all clusters in a runlist, use
  * @start_vcn = 0 and @count = -1.
  *
- * Note, ntfs_cluster_free() does not modify the runlist at all, so the caller
- * has to deal with it later.
+ * If @ctx is specified, it is an active search context of @ni and its base mft
+ * record.  This is needed when ntfs_cluster_free() encounters unmapped runlist
+ * fragments and allows their mapping.  If you do not have the mft record
+ * mapped, you can specify @ctx as NULL and ntfs_cluster_free() will perform
+ * the necessary mapping and unmapping.
+ *
+ * Note, ntfs_cluster_free() saves the state of @ctx on entry and restores it
+ * before returning.  Thus, @ctx will be left pointing to the same attribute on
+ * return as on entry.  However, the actual pointers in @ctx may point to
+ * different memory locations on return, so you must remember to reset any
+ * cached pointers from the @ctx, i.e. after the call to ntfs_cluster_free(),
+ * you will probably want to do:
+ *	m = ctx->mrec;
+ *	a = ctx->attr;
+ * Assuming you cache ctx->attr in a variable @a of type ATTR_RECORD * and that
+ * you cache ctx->mrec in a variable @m of type MFT_RECORD *.
+ *
+ * Note, ntfs_cluster_free() does not modify the runlist, so you have to remove
+ * from the runlist or mark sparse the freed runs later.
  *
  * Return the number of deallocated clusters (not counting sparse ones) on
  * success and -errno on error.
  *
+ * WARNING: If @ctx is supplied, regardless of whether success or failure is
+ *	    returned, you need to check IS_ERR(@ctx->mrec) and if TRUE the @ctx
+ *	    is no longer valid, i.e. you need to either call
+ *	    ntfs_attr_reinit_search_ctx() or ntfs_attr_put_search_ctx() on it.
+ *	    In that case PTR_ERR(@ctx->mrec) will give you the error code for
+ *	    why the mapping of the old inode failed.
+ *
  * Locking: - The runlist described by @ni must be locked for writing on entry
  *	      and is locked on return.  Note the runlist may be modified when
  *	      needed runlist fragments need to be mapped.
@@ -72,11 +99,15 @@ extern s64 __ntfs_cluster_free(ntfs_inode *ni, const VCN start_vcn,
  *	      on return.
  *	    - This function takes the volume lcn bitmap lock for writing and
  *	      modifies the bitmap contents.
+ *	    - If @ctx is NULL, the base mft record of @ni must not be mapped on
+ *	      entry and it will be left unmapped on return.
+ *	    - If @ctx is not NULL, the base mft record must be mapped on entry
+ *	      and it will be left mapped on return.
  */
 static inline s64 ntfs_cluster_free(ntfs_inode *ni, const VCN start_vcn,
-		s64 count)
+		s64 count, ntfs_attr_search_ctx *ctx)
 {
-	return __ntfs_cluster_free(ni, start_vcn, count, FALSE);
+	return __ntfs_cluster_free(ni, start_vcn, count, ctx, FALSE);
 }
 
 extern int ntfs_cluster_free_from_rl_nolock(ntfs_volume *vol,
diff --git a/fs/ntfs/malloc.h b/fs/ntfs/malloc.h
index 590887b943f5..e38e402e4103 100644
--- a/fs/ntfs/malloc.h
+++ b/fs/ntfs/malloc.h
@@ -39,8 +39,7 @@
  * If there was insufficient memory to complete the request, return NULL.
  * Depending on @gfp_mask the allocation may be guaranteed to succeed.
  */
-static inline void *__ntfs_malloc(unsigned long size,
-		gfp_t gfp_mask)
+static inline void *__ntfs_malloc(unsigned long size, gfp_t gfp_mask)
 {
 	if (likely(size <= PAGE_SIZE)) {
 		BUG_ON(!size);
diff --git a/fs/ntfs/mft.c b/fs/ntfs/mft.c
index b011369b5956..0c65cbb8c5cf 100644
--- a/fs/ntfs/mft.c
+++ b/fs/ntfs/mft.c
@@ -49,7 +49,8 @@ static inline MFT_RECORD *map_mft_record_page(ntfs_inode *ni)
 	ntfs_volume *vol = ni->vol;
 	struct inode *mft_vi = vol->mft_ino;
 	struct page *page;
-	unsigned long index, ofs, end_index;
+	unsigned long index, end_index;
+	unsigned ofs;
 
 	BUG_ON(ni->page);
 	/*
@@ -1308,7 +1309,7 @@ static int ntfs_mft_bitmap_extend_allocation_nolock(ntfs_volume *vol)
 	ll = mftbmp_ni->allocated_size;
 	read_unlock_irqrestore(&mftbmp_ni->size_lock, flags);
 	rl = ntfs_attr_find_vcn_nolock(mftbmp_ni,
-			(ll - 1) >> vol->cluster_size_bits, TRUE);
+			(ll - 1) >> vol->cluster_size_bits, NULL);
 	if (unlikely(IS_ERR(rl) || !rl->length || rl->lcn < 0)) {
 		up_write(&mftbmp_ni->runlist.lock);
 		ntfs_error(vol->sb, "Failed to determine last allocated "
@@ -1354,7 +1355,8 @@ static int ntfs_mft_bitmap_extend_allocation_nolock(ntfs_volume *vol)
 		up_write(&vol->lcnbmp_lock);
 		ntfs_unmap_page(page);
 		/* Allocate a cluster from the DATA_ZONE. */
-		rl2 = ntfs_cluster_alloc(vol, rl[1].vcn, 1, lcn, DATA_ZONE);
+		rl2 = ntfs_cluster_alloc(vol, rl[1].vcn, 1, lcn, DATA_ZONE,
+				TRUE);
 		if (IS_ERR(rl2)) {
 			up_write(&mftbmp_ni->runlist.lock);
 			ntfs_error(vol->sb, "Failed to allocate a cluster for "
@@ -1738,7 +1740,7 @@ static int ntfs_mft_data_extend_allocation_nolock(ntfs_volume *vol)
 	ll = mft_ni->allocated_size;
 	read_unlock_irqrestore(&mft_ni->size_lock, flags);
 	rl = ntfs_attr_find_vcn_nolock(mft_ni,
-			(ll - 1) >> vol->cluster_size_bits, TRUE);
+			(ll - 1) >> vol->cluster_size_bits, NULL);
 	if (unlikely(IS_ERR(rl) || !rl->length || rl->lcn < 0)) {
 		up_write(&mft_ni->runlist.lock);
 		ntfs_error(vol->sb, "Failed to determine last allocated "
@@ -1779,7 +1781,8 @@ static int ntfs_mft_data_extend_allocation_nolock(ntfs_volume *vol)
 			nr > min_nr ? "default" : "minimal", (long long)nr);
 	old_last_vcn = rl[1].vcn;
 	do {
-		rl2 = ntfs_cluster_alloc(vol, old_last_vcn, nr, lcn, MFT_ZONE);
+		rl2 = ntfs_cluster_alloc(vol, old_last_vcn, nr, lcn, MFT_ZONE,
+				TRUE);
 		if (likely(!IS_ERR(rl2)))
 			break;
 		if (PTR_ERR(rl2) != -ENOSPC || nr == min_nr) {
@@ -1951,20 +1954,21 @@ restore_undo_alloc:
 		NVolSetErrors(vol);
 		return ret;
 	}
-	a = ctx->attr;
-	a->data.non_resident.highest_vcn = cpu_to_sle64(old_last_vcn - 1);
+	ctx->attr->data.non_resident.highest_vcn =
+			cpu_to_sle64(old_last_vcn - 1);
 undo_alloc:
-	if (ntfs_cluster_free(mft_ni, old_last_vcn, -1) < 0) {
+	if (ntfs_cluster_free(mft_ni, old_last_vcn, -1, ctx) < 0) {
 		ntfs_error(vol->sb, "Failed to free clusters from mft data "
 				"attribute.%s", es);
 		NVolSetErrors(vol);
 	}
+	a = ctx->attr;
 	if (ntfs_rl_truncate_nolock(vol, &mft_ni->runlist, old_last_vcn)) {
 		ntfs_error(vol->sb, "Failed to truncate mft data attribute "
 				"runlist.%s", es);
 		NVolSetErrors(vol);
 	}
-	if (mp_rebuilt) {
+	if (mp_rebuilt && !IS_ERR(ctx->mrec)) {
 		if (ntfs_mapping_pairs_build(vol, (u8*)a + le16_to_cpu(
 				a->data.non_resident.mapping_pairs_offset),
 				old_alen - le16_to_cpu(
@@ -1981,6 +1985,10 @@ undo_alloc:
 		}
 		flush_dcache_mft_record_page(ctx->ntfs_ino);
 		mark_mft_record_dirty(ctx->ntfs_ino);
+	} else if (IS_ERR(ctx->mrec)) {
+		ntfs_error(vol->sb, "Failed to restore attribute search "
+				"context.%s", es);
+		NVolSetErrors(vol);
 	}
 	if (ctx)
 		ntfs_attr_put_search_ctx(ctx);
diff --git a/fs/ntfs/super.c b/fs/ntfs/super.c
index 453d0d51ea4b..6c16db9e1a8a 100644
--- a/fs/ntfs/super.c
+++ b/fs/ntfs/super.c
@@ -1447,7 +1447,7 @@ not_enabled:
 	if (unlikely(i_size_read(tmp_ino) < sizeof(USN_HEADER))) {
 		ntfs_error(vol->sb, "Found corrupt $UsnJrnl/$DATA/$Max "
 				"attribute (size is 0x%llx but should be at "
-				"least 0x%x bytes).", i_size_read(tmp_ino),
+				"least 0x%zx bytes).", i_size_read(tmp_ino),
 				sizeof(USN_HEADER));
 		return FALSE;
 	}