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