1/**
   2 * aops.c - NTFS kernel address space operations and page cache handling.
   3 *
   4 * Copyright (c) 2001-2014 Anton Altaparmakov and Tuxera Inc.
   5 * Copyright (c) 2002 Richard Russon
   6 *
   7 * This program/include file is free software; you can redistribute it and/or
   8 * modify it under the terms of the GNU General Public License as published
   9 * by the Free Software Foundation; either version 2 of the License, or
  10 * (at your option) any later version.
  11 *
  12 * This program/include file is distributed in the hope that it will be
  13 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  14 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15 * GNU General Public License for more details.
  16 *
  17 * You should have received a copy of the GNU General Public License
  18 * along with this program (in the main directory of the Linux-NTFS
  19 * distribution in the file COPYING); if not, write to the Free Software
  20 * Foundation,Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  21 */
  22
  23#include <linux/errno.h>
  24#include <linux/fs.h>
  25#include <linux/gfp.h>
  26#include <linux/mm.h>
  27#include <linux/pagemap.h>
  28#include <linux/swap.h>
  29#include <linux/buffer_head.h>
  30#include <linux/writeback.h>
  31#include <linux/bit_spinlock.h>
  32
  33#include "aops.h"
  34#include "attrib.h"
  35#include "debug.h"
  36#include "inode.h"
  37#include "mft.h"
  38#include "runlist.h"
  39#include "types.h"
  40#include "ntfs.h"
  41
  42/**
  43 * ntfs_end_buffer_async_read - async io completion for reading attributes
  44 * @bh:		buffer head on which io is completed
  45 * @uptodate:	whether @bh is now uptodate or not
  46 *
  47 * Asynchronous I/O completion handler for reading pages belonging to the
  48 * attribute address space of an inode.  The inodes can either be files or
  49 * directories or they can be fake inodes describing some attribute.
  50 *
  51 * If NInoMstProtected(), perform the post read mst fixups when all IO on the
  52 * page has been completed and mark the page uptodate or set the error bit on
  53 * the page.  To determine the size of the records that need fixing up, we
  54 * cheat a little bit by setting the index_block_size in ntfs_inode to the ntfs
  55 * record size, and index_block_size_bits, to the log(base 2) of the ntfs
  56 * record size.
  57 */
  58static void ntfs_end_buffer_async_read(struct buffer_head *bh, int uptodate)
  59{
  60	unsigned long flags;
  61	struct buffer_head *first, *tmp;
  62	struct page *page;
  63	struct inode *vi;
  64	ntfs_inode *ni;
  65	int page_uptodate = 1;
  66
  67	page = bh->b_page;
  68	vi = page->mapping->host;
  69	ni = NTFS_I(vi);
  70
  71	if (likely(uptodate)) {
  72		loff_t i_size;
  73		s64 file_ofs, init_size;
  74
  75		set_buffer_uptodate(bh);
  76
  77		file_ofs = ((s64)page->index << PAGE_SHIFT) +
  78				bh_offset(bh);
  79		read_lock_irqsave(&ni->size_lock, flags);
  80		init_size = ni->initialized_size;
  81		i_size = i_size_read(vi);
  82		read_unlock_irqrestore(&ni->size_lock, flags);
  83		if (unlikely(init_size > i_size)) {
  84			/* Race with shrinking truncate. */
  85			init_size = i_size;
  86		}
  87		/* Check for the current buffer head overflowing. */
  88		if (unlikely(file_ofs + bh->b_size > init_size)) {
  89			int ofs;
  90			void *kaddr;
  91
  92			ofs = 0;
  93			if (file_ofs < init_size)
  94				ofs = init_size - file_ofs;
  95			local_irq_save(flags);
  96			kaddr = kmap_atomic(page);
  97			memset(kaddr + bh_offset(bh) + ofs, 0,
  98					bh->b_size - ofs);
  99			flush_dcache_page(page);
 100			kunmap_atomic(kaddr);
 101			local_irq_restore(flags);
 102		}
 103	} else {
 104		clear_buffer_uptodate(bh);
 105		SetPageError(page);
 106		ntfs_error(ni->vol->sb, "Buffer I/O error, logical block "
 107				"0x%llx.", (unsigned long long)bh->b_blocknr);
 108	}
 109	first = page_buffers(page);
 110	local_irq_save(flags);
 111	bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
 112	clear_buffer_async_read(bh);
 113	unlock_buffer(bh);
 114	tmp = bh;
 115	do {
 116		if (!buffer_uptodate(tmp))
 117			page_uptodate = 0;
 118		if (buffer_async_read(tmp)) {
 119			if (likely(buffer_locked(tmp)))
 120				goto still_busy;
 121			/* Async buffers must be locked. */
 122			BUG();
 123		}
 124		tmp = tmp->b_this_page;
 125	} while (tmp != bh);
 126	bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
 127	local_irq_restore(flags);
 128	/*
 129	 * If none of the buffers had errors then we can set the page uptodate,
 130	 * but we first have to perform the post read mst fixups, if the
 131	 * attribute is mst protected, i.e. if NInoMstProteced(ni) is true.
 132	 * Note we ignore fixup errors as those are detected when
 133	 * map_mft_record() is called which gives us per record granularity
 134	 * rather than per page granularity.
 135	 */
 136	if (!NInoMstProtected(ni)) {
 137		if (likely(page_uptodate && !PageError(page)))
 138			SetPageUptodate(page);
 139	} else {
 140		u8 *kaddr;
 141		unsigned int i, recs;
 142		u32 rec_size;
 143
 144		rec_size = ni->itype.index.block_size;
 145		recs = PAGE_SIZE / rec_size;
 146		/* Should have been verified before we got here... */
 147		BUG_ON(!recs);
 148		local_irq_save(flags);
 149		kaddr = kmap_atomic(page);
 150		for (i = 0; i < recs; i++)
 151			post_read_mst_fixup((NTFS_RECORD*)(kaddr +
 152					i * rec_size), rec_size);
 153		kunmap_atomic(kaddr);
 154		local_irq_restore(flags);
 155		flush_dcache_page(page);
 156		if (likely(page_uptodate && !PageError(page)))
 157			SetPageUptodate(page);
 158	}
 159	unlock_page(page);
 160	return;
 161still_busy:
 162	bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
 163	local_irq_restore(flags);
 164	return;
 165}
 166
 167/**
 168 * ntfs_read_block - fill a @page of an address space with data
 169 * @page:	page cache page to fill with data
 170 *
 171 * Fill the page @page of the address space belonging to the @page->host inode.
 172 * We read each buffer asynchronously and when all buffers are read in, our io
 173 * completion handler ntfs_end_buffer_read_async(), if required, automatically
 174 * applies the mst fixups to the page before finally marking it uptodate and
 175 * unlocking it.
 176 *
 177 * We only enforce allocated_size limit because i_size is checked for in
 178 * generic_file_read().
 179 *
 180 * Return 0 on success and -errno on error.
 181 *
 182 * Contains an adapted version of fs/buffer.c::block_read_full_page().
 183 */
 184static int ntfs_read_block(struct page *page)
 185{
 186	loff_t i_size;
 187	VCN vcn;
 188	LCN lcn;
 189	s64 init_size;
 190	struct inode *vi;
 191	ntfs_inode *ni;
 192	ntfs_volume *vol;
 193	runlist_element *rl;
 194	struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
 195	sector_t iblock, lblock, zblock;
 196	unsigned long flags;
 197	unsigned int blocksize, vcn_ofs;
 198	int i, nr;
 199	unsigned char blocksize_bits;
 200
 201	vi = page->mapping->host;
 202	ni = NTFS_I(vi);
 203	vol = ni->vol;
 204
 205	/* $MFT/$DATA must have its complete runlist in memory at all times. */
 206	BUG_ON(!ni->runlist.rl && !ni->mft_no && !NInoAttr(ni));
 207
 208	blocksize = vol->sb->s_blocksize;
 209	blocksize_bits = vol->sb->s_blocksize_bits;
 210
 211	if (!page_has_buffers(page)) {
 212		create_empty_buffers(page, blocksize, 0);
 213		if (unlikely(!page_has_buffers(page))) {
 214			unlock_page(page);
 215			return -ENOMEM;
 216		}
 217	}
 218	bh = head = page_buffers(page);
 219	BUG_ON(!bh);
 220
 221	/*
 222	 * We may be racing with truncate.  To avoid some of the problems we
 223	 * now take a snapshot of the various sizes and use those for the whole
 224	 * of the function.  In case of an extending truncate it just means we
 225	 * may leave some buffers unmapped which are now allocated.  This is
 226	 * not a problem since these buffers will just get mapped when a write
 227	 * occurs.  In case of a shrinking truncate, we will detect this later
 228	 * on due to the runlist being incomplete and if the page is being
 229	 * fully truncated, truncate will throw it away as soon as we unlock
 230	 * it so no need to worry what we do with it.
 231	 */
 232	iblock = (s64)page->index << (PAGE_SHIFT - blocksize_bits);
 233	read_lock_irqsave(&ni->size_lock, flags);
 234	lblock = (ni->allocated_size + blocksize - 1) >> blocksize_bits;
 235	init_size = ni->initialized_size;
 236	i_size = i_size_read(vi);
 237	read_unlock_irqrestore(&ni->size_lock, flags);
 238	if (unlikely(init_size > i_size)) {
 239		/* Race with shrinking truncate. */
 240		init_size = i_size;
 241	}
 242	zblock = (init_size + blocksize - 1) >> blocksize_bits;
 243
 244	/* Loop through all the buffers in the page. */
 245	rl = NULL;
 246	nr = i = 0;
 247	do {
 248		int err = 0;
 249
 250		if (unlikely(buffer_uptodate(bh)))
 251			continue;
 252		if (unlikely(buffer_mapped(bh))) {
 253			arr[nr++] = bh;
 254			continue;
 255		}
 256		bh->b_bdev = vol->sb->s_bdev;
 257		/* Is the block within the allowed limits? */
 258		if (iblock < lblock) {
 259			bool is_retry = false;
 260
 261			/* Convert iblock into corresponding vcn and offset. */
 262			vcn = (VCN)iblock << blocksize_bits >>
 263					vol->cluster_size_bits;
 264			vcn_ofs = ((VCN)iblock << blocksize_bits) &
 265					vol->cluster_size_mask;
 266			if (!rl) {
 267lock_retry_remap:
 268				down_read(&ni->runlist.lock);
 269				rl = ni->runlist.rl;
 270			}
 271			if (likely(rl != NULL)) {
 272				/* Seek to element containing target vcn. */
 273				while (rl->length && rl[1].vcn <= vcn)
 274					rl++;
 275				lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
 276			} else
 277				lcn = LCN_RL_NOT_MAPPED;
 278			/* Successful remap. */
 279			if (lcn >= 0) {
 280				/* Setup buffer head to correct block. */
 281				bh->b_blocknr = ((lcn << vol->cluster_size_bits)
 282						+ vcn_ofs) >> blocksize_bits;
 283				set_buffer_mapped(bh);
 284				/* Only read initialized data blocks. */
 285				if (iblock < zblock) {
 286					arr[nr++] = bh;
 287					continue;
 288				}
 289				/* Fully non-initialized data block, zero it. */
 290				goto handle_zblock;
 291			}
 292			/* It is a hole, need to zero it. */
 293			if (lcn == LCN_HOLE)
 294				goto handle_hole;
 295			/* If first try and runlist unmapped, map and retry. */
 296			if (!is_retry && lcn == LCN_RL_NOT_MAPPED) {
 297				is_retry = true;
 298				/*
 299				 * Attempt to map runlist, dropping lock for
 300				 * the duration.
 301				 */
 302				up_read(&ni->runlist.lock);
 303				err = ntfs_map_runlist(ni, vcn);
 304				if (likely(!err))
 305					goto lock_retry_remap;
 306				rl = NULL;
 307			} else if (!rl)
 308				up_read(&ni->runlist.lock);
 309			/*
 310			 * If buffer is outside the runlist, treat it as a
 311			 * hole.  This can happen due to concurrent truncate
 312			 * for example.
 313			 */
 314			if (err == -ENOENT || lcn == LCN_ENOENT) {
 315				err = 0;
 316				goto handle_hole;
 317			}
 318			/* Hard error, zero out region. */
 319			if (!err)
 320				err = -EIO;
 321			bh->b_blocknr = -1;
 322			SetPageError(page);
 323			ntfs_error(vol->sb, "Failed to read from inode 0x%lx, "
 324					"attribute type 0x%x, vcn 0x%llx, "
 325					"offset 0x%x because its location on "
 326					"disk could not be determined%s "
 327					"(error code %i).", ni->mft_no,
 328					ni->type, (unsigned long long)vcn,
 329					vcn_ofs, is_retry ? " even after "
 330					"retrying" : "", err);
 331		}
 332		/*
 333		 * Either iblock was outside lblock limits or
 334		 * ntfs_rl_vcn_to_lcn() returned error.  Just zero that portion
 335		 * of the page and set the buffer uptodate.
 336		 */
 337handle_hole:
 338		bh->b_blocknr = -1UL;
 339		clear_buffer_mapped(bh);
 340handle_zblock:
 341		zero_user(page, i * blocksize, blocksize);
 342		if (likely(!err))
 343			set_buffer_uptodate(bh);
 344	} while (i++, iblock++, (bh = bh->b_this_page) != head);
 345
 346	/* Release the lock if we took it. */
 347	if (rl)
 348		up_read(&ni->runlist.lock);
 349
 350	/* Check we have at least one buffer ready for i/o. */
 351	if (nr) {
 352		struct buffer_head *tbh;
 353
 354		/* Lock the buffers. */
 355		for (i = 0; i < nr; i++) {
 356			tbh = arr[i];
 357			lock_buffer(tbh);
 358			tbh->b_end_io = ntfs_end_buffer_async_read;
 359			set_buffer_async_read(tbh);
 360		}
 361		/* Finally, start i/o on the buffers. */
 362		for (i = 0; i < nr; i++) {
 363			tbh = arr[i];
 364			if (likely(!buffer_uptodate(tbh)))
 365				submit_bh(READ, tbh);
 366			else
 367				ntfs_end_buffer_async_read(tbh, 1);
 368		}
 369		return 0;
 370	}
 371	/* No i/o was scheduled on any of the buffers. */
 372	if (likely(!PageError(page)))
 373		SetPageUptodate(page);
 374	else /* Signal synchronous i/o error. */
 375		nr = -EIO;
 376	unlock_page(page);
 377	return nr;
 378}
 379
 380/**
 381 * ntfs_readpage - fill a @page of a @file with data from the device
 382 * @file:	open file to which the page @page belongs or NULL
 383 * @page:	page cache page to fill with data
 384 *
 385 * For non-resident attributes, ntfs_readpage() fills the @page of the open
 386 * file @file by calling the ntfs version of the generic block_read_full_page()
 387 * function, ntfs_read_block(), which in turn creates and reads in the buffers
 388 * associated with the page asynchronously.
 389 *
 390 * For resident attributes, OTOH, ntfs_readpage() fills @page by copying the
 391 * data from the mft record (which at this stage is most likely in memory) and
 392 * fills the remainder with zeroes. Thus, in this case, I/O is synchronous, as
 393 * even if the mft record is not cached at this point in time, we need to wait
 394 * for it to be read in before we can do the copy.
 395 *
 396 * Return 0 on success and -errno on error.
 397 */
 398static int ntfs_readpage(struct file *file, struct page *page)
 399{
 400	loff_t i_size;
 401	struct inode *vi;
 402	ntfs_inode *ni, *base_ni;
 403	u8 *addr;
 404	ntfs_attr_search_ctx *ctx;
 405	MFT_RECORD *mrec;
 406	unsigned long flags;
 407	u32 attr_len;
 408	int err = 0;
 409
 410retry_readpage:
 411	BUG_ON(!PageLocked(page));
 412	vi = page->mapping->host;
 413	i_size = i_size_read(vi);
 414	/* Is the page fully outside i_size? (truncate in progress) */
 415	if (unlikely(page->index >= (i_size + PAGE_SIZE - 1) >>
 416			PAGE_SHIFT)) {
 417		zero_user(page, 0, PAGE_SIZE);
 418		ntfs_debug("Read outside i_size - truncated?");
 419		goto done;
 420	}
 421	/*
 422	 * This can potentially happen because we clear PageUptodate() during
 423	 * ntfs_writepage() of MstProtected() attributes.
 424	 */
 425	if (PageUptodate(page)) {
 426		unlock_page(page);
 427		return 0;
 428	}
 429	ni = NTFS_I(vi);
 430	/*
 431	 * Only $DATA attributes can be encrypted and only unnamed $DATA
 432	 * attributes can be compressed.  Index root can have the flags set but
 433	 * this means to create compressed/encrypted files, not that the
 434	 * attribute is compressed/encrypted.  Note we need to check for
 435	 * AT_INDEX_ALLOCATION since this is the type of both directory and
 436	 * index inodes.
 437	 */
 438	if (ni->type != AT_INDEX_ALLOCATION) {
 439		/* If attribute is encrypted, deny access, just like NT4. */
 440		if (NInoEncrypted(ni)) {
 441			BUG_ON(ni->type != AT_DATA);
 442			err = -EACCES;
 443			goto err_out;
 444		}
 445		/* Compressed data streams are handled in compress.c. */
 446		if (NInoNonResident(ni) && NInoCompressed(ni)) {
 447			BUG_ON(ni->type != AT_DATA);
 448			BUG_ON(ni->name_len);
 449			return ntfs_read_compressed_block(page);
 450		}
 451	}
 452	/* NInoNonResident() == NInoIndexAllocPresent() */
 453	if (NInoNonResident(ni)) {
 454		/* Normal, non-resident data stream. */
 455		return ntfs_read_block(page);
 456	}
 457	/*
 458	 * Attribute is resident, implying it is not compressed or encrypted.
 459	 * This also means the attribute is smaller than an mft record and
 460	 * hence smaller than a page, so can simply zero out any pages with
 461	 * index above 0.  Note the attribute can actually be marked compressed
 462	 * but if it is resident the actual data is not compressed so we are
 463	 * ok to ignore the compressed flag here.
 464	 */
 465	if (unlikely(page->index > 0)) {
 466		zero_user(page, 0, PAGE_SIZE);
 467		goto done;
 468	}
 469	if (!NInoAttr(ni))
 470		base_ni = ni;
 471	else
 472		base_ni = ni->ext.base_ntfs_ino;
 473	/* Map, pin, and lock the mft record. */
 474	mrec = map_mft_record(base_ni);
 475	if (IS_ERR(mrec)) {
 476		err = PTR_ERR(mrec);
 477		goto err_out;
 478	}
 479	/*
 480	 * If a parallel write made the attribute non-resident, drop the mft
 481	 * record and retry the readpage.
 482	 */
 483	if (unlikely(NInoNonResident(ni))) {
 484		unmap_mft_record(base_ni);
 485		goto retry_readpage;
 486	}
 487	ctx = ntfs_attr_get_search_ctx(base_ni, mrec);
 488	if (unlikely(!ctx)) {
 489		err = -ENOMEM;
 490		goto unm_err_out;
 491	}
 492	err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
 493			CASE_SENSITIVE, 0, NULL, 0, ctx);
 494	if (unlikely(err))
 495		goto put_unm_err_out;
 496	attr_len = le32_to_cpu(ctx->attr->data.resident.value_length);
 497	read_lock_irqsave(&ni->size_lock, flags);
 498	if (unlikely(attr_len > ni->initialized_size))
 499		attr_len = ni->initialized_size;
 500	i_size = i_size_read(vi);
 501	read_unlock_irqrestore(&ni->size_lock, flags);
 502	if (unlikely(attr_len > i_size)) {
 503		/* Race with shrinking truncate. */
 504		attr_len = i_size;
 505	}
 506	addr = kmap_atomic(page);
 507	/* Copy the data to the page. */
 508	memcpy(addr, (u8*)ctx->attr +
 509			le16_to_cpu(ctx->attr->data.resident.value_offset),
 510			attr_len);
 511	/* Zero the remainder of the page. */
 512	memset(addr + attr_len, 0, PAGE_SIZE - attr_len);
 513	flush_dcache_page(page);
 514	kunmap_atomic(addr);
 515put_unm_err_out:
 516	ntfs_attr_put_search_ctx(ctx);
 517unm_err_out:
 518	unmap_mft_record(base_ni);
 519done:
 520	SetPageUptodate(page);
 521err_out:
 522	unlock_page(page);
 523	return err;
 524}
 525
 526#ifdef NTFS_RW
 527
 528/**
 529 * ntfs_write_block - write a @page to the backing store
 530 * @page:	page cache page to write out
 531 * @wbc:	writeback control structure
 532 *
 533 * This function is for writing pages belonging to non-resident, non-mst
 534 * protected attributes to their backing store.
 535 *
 536 * For a page with buffers, map and write the dirty buffers asynchronously
 537 * under page writeback. For a page without buffers, create buffers for the
 538 * page, then proceed as above.
 539 *
 540 * If a page doesn't have buffers the page dirty state is definitive. If a page
 541 * does have buffers, the page dirty state is just a hint, and the buffer dirty
 542 * state is definitive. (A hint which has rules: dirty buffers against a clean
 543 * page is illegal. Other combinations are legal and need to be handled. In
 544 * particular a dirty page containing clean buffers for example.)
 545 *
 546 * Return 0 on success and -errno on error.
 547 *
 548 * Based on ntfs_read_block() and __block_write_full_page().
 549 */
 550static int ntfs_write_block(struct page *page, struct writeback_control *wbc)
 551{
 552	VCN vcn;
 553	LCN lcn;
 554	s64 initialized_size;
 555	loff_t i_size;
 556	sector_t block, dblock, iblock;
 557	struct inode *vi;
 558	ntfs_inode *ni;
 559	ntfs_volume *vol;
 560	runlist_element *rl;
 561	struct buffer_head *bh, *head;
 562	unsigned long flags;
 563	unsigned int blocksize, vcn_ofs;
 564	int err;
 565	bool need_end_writeback;
 566	unsigned char blocksize_bits;
 567
 568	vi = page->mapping->host;
 569	ni = NTFS_I(vi);
 570	vol = ni->vol;
 571
 572	ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
 573			"0x%lx.", ni->mft_no, ni->type, page->index);
 574
 575	BUG_ON(!NInoNonResident(ni));
 576	BUG_ON(NInoMstProtected(ni));
 577	blocksize = vol->sb->s_blocksize;
 578	blocksize_bits = vol->sb->s_blocksize_bits;
 579	if (!page_has_buffers(page)) {
 580		BUG_ON(!PageUptodate(page));
 581		create_empty_buffers(page, blocksize,
 582				(1 << BH_Uptodate) | (1 << BH_Dirty));
 583		if (unlikely(!page_has_buffers(page))) {
 584			ntfs_warning(vol->sb, "Error allocating page "
 585					"buffers.  Redirtying page so we try "
 586					"again later.");
 587			/*
 588			 * Put the page back on mapping->dirty_pages, but leave
 589			 * its buffers' dirty state as-is.
 590			 */
 591			redirty_page_for_writepage(wbc, page);
 592			unlock_page(page);
 593			return 0;
 594		}
 595	}
 596	bh = head = page_buffers(page);
 597	BUG_ON(!bh);
 598
 599	/* NOTE: Different naming scheme to ntfs_read_block()! */
 600
 601	/* The first block in the page. */
 602	block = (s64)page->index << (PAGE_SHIFT - blocksize_bits);
 603
 604	read_lock_irqsave(&ni->size_lock, flags);
 605	i_size = i_size_read(vi);
 606	initialized_size = ni->initialized_size;
 607	read_unlock_irqrestore(&ni->size_lock, flags);
 608
 609	/* The first out of bounds block for the data size. */
 610	dblock = (i_size + blocksize - 1) >> blocksize_bits;
 611
 612	/* The last (fully or partially) initialized block. */
 613	iblock = initialized_size >> blocksize_bits;
 614
 615	/*
 616	 * Be very careful.  We have no exclusion from __set_page_dirty_buffers
 617	 * here, and the (potentially unmapped) buffers may become dirty at
 618	 * any time.  If a buffer becomes dirty here after we've inspected it
 619	 * then we just miss that fact, and the page stays dirty.
 620	 *
 621	 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
 622	 * handle that here by just cleaning them.
 623	 */
 624
 625	/*
 626	 * Loop through all the buffers in the page, mapping all the dirty
 627	 * buffers to disk addresses and handling any aliases from the
 628	 * underlying block device's mapping.
 629	 */
 630	rl = NULL;
 631	err = 0;
 632	do {
 633		bool is_retry = false;
 634
 635		if (unlikely(block >= dblock)) {
 636			/*
 637			 * Mapped buffers outside i_size will occur, because
 638			 * this page can be outside i_size when there is a
 639			 * truncate in progress. The contents of such buffers
 640			 * were zeroed by ntfs_writepage().
 641			 *
 642			 * FIXME: What about the small race window where
 643			 * ntfs_writepage() has not done any clearing because
 644			 * the page was within i_size but before we get here,
 645			 * vmtruncate() modifies i_size?
 646			 */
 647			clear_buffer_dirty(bh);
 648			set_buffer_uptodate(bh);
 649			continue;
 650		}
 651
 652		/* Clean buffers are not written out, so no need to map them. */
 653		if (!buffer_dirty(bh))
 654			continue;
 655
 656		/* Make sure we have enough initialized size. */
 657		if (unlikely((block >= iblock) &&
 658				(initialized_size < i_size))) {
 659			/*
 660			 * If this page is fully outside initialized size, zero
 661			 * out all pages between the current initialized size
 662			 * and the current page. Just use ntfs_readpage() to do
 663			 * the zeroing transparently.
 664			 */
 665			if (block > iblock) {
 666				// TODO:
 667				// For each page do:
 668				// - read_cache_page()
 669				// Again for each page do:
 670				// - wait_on_page_locked()
 671				// - Check (PageUptodate(page) &&
 672				//			!PageError(page))
 673				// Update initialized size in the attribute and
 674				// in the inode.
 675				// Again, for each page do:
 676				//	__set_page_dirty_buffers();
 677				// put_page()
 678				// We don't need to wait on the writes.
 679				// Update iblock.
 680			}
 681			/*
 682			 * The current page straddles initialized size. Zero
 683			 * all non-uptodate buffers and set them uptodate (and
 684			 * dirty?). Note, there aren't any non-uptodate buffers
 685			 * if the page is uptodate.
 686			 * FIXME: For an uptodate page, the buffers may need to
 687			 * be written out because they were not initialized on
 688			 * disk before.
 689			 */
 690			if (!PageUptodate(page)) {
 691				// TODO:
 692				// Zero any non-uptodate buffers up to i_size.
 693				// Set them uptodate and dirty.
 694			}
 695			// TODO:
 696			// Update initialized size in the attribute and in the
 697			// inode (up to i_size).
 698			// Update iblock.
 699			// FIXME: This is inefficient. Try to batch the two
 700			// size changes to happen in one go.
 701			ntfs_error(vol->sb, "Writing beyond initialized size "
 702					"is not supported yet. Sorry.");
 703			err = -EOPNOTSUPP;
 704			break;
 705			// Do NOT set_buffer_new() BUT DO clear buffer range
 706			// outside write request range.
 707			// set_buffer_uptodate() on complete buffers as well as
 708			// set_buffer_dirty().
 709		}
 710
 711		/* No need to map buffers that are already mapped. */
 712		if (buffer_mapped(bh))
 713			continue;
 714
 715		/* Unmapped, dirty buffer. Need to map it. */
 716		bh->b_bdev = vol->sb->s_bdev;
 717
 718		/* Convert block into corresponding vcn and offset. */
 719		vcn = (VCN)block << blocksize_bits;
 720		vcn_ofs = vcn & vol->cluster_size_mask;
 721		vcn >>= vol->cluster_size_bits;
 722		if (!rl) {
 723lock_retry_remap:
 724			down_read(&ni->runlist.lock);
 725			rl = ni->runlist.rl;
 726		}
 727		if (likely(rl != NULL)) {
 728			/* Seek to element containing target vcn. */
 729			while (rl->length && rl[1].vcn <= vcn)
 730				rl++;
 731			lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
 732		} else
 733			lcn = LCN_RL_NOT_MAPPED;
 734		/* Successful remap. */
 735		if (lcn >= 0) {
 736			/* Setup buffer head to point to correct block. */
 737			bh->b_blocknr = ((lcn << vol->cluster_size_bits) +
 738					vcn_ofs) >> blocksize_bits;
 739			set_buffer_mapped(bh);
 740			continue;
 741		}
 742		/* It is a hole, need to instantiate it. */
 743		if (lcn == LCN_HOLE) {
 744			u8 *kaddr;
 745			unsigned long *bpos, *bend;
 746
 747			/* Check if the buffer is zero. */
 748			kaddr = kmap_atomic(page);
 749			bpos = (unsigned long *)(kaddr + bh_offset(bh));
 750			bend = (unsigned long *)((u8*)bpos + blocksize);
 751			do {
 752				if (unlikely(*bpos))
 753					break;
 754			} while (likely(++bpos < bend));
 755			kunmap_atomic(kaddr);
 756			if (bpos == bend) {
 757				/*
 758				 * Buffer is zero and sparse, no need to write
 759				 * it.
 760				 */
 761				bh->b_blocknr = -1;
 762				clear_buffer_dirty(bh);
 763				continue;
 764			}
 765			// TODO: Instantiate the hole.
 766			// clear_buffer_new(bh);
 767			// unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 768			ntfs_error(vol->sb, "Writing into sparse regions is "
 769					"not supported yet. Sorry.");
 770			err = -EOPNOTSUPP;
 771			break;
 772		}
 773		/* If first try and runlist unmapped, map and retry. */
 774		if (!is_retry && lcn == LCN_RL_NOT_MAPPED) {
 775			is_retry = true;
 776			/*
 777			 * Attempt to map runlist, dropping lock for
 778			 * the duration.
 779			 */
 780			up_read(&ni->runlist.lock);
 781			err = ntfs_map_runlist(ni, vcn);
 782			if (likely(!err))
 783				goto lock_retry_remap;
 784			rl = NULL;
 785		} else if (!rl)
 786			up_read(&ni->runlist.lock);
 787		/*
 788		 * If buffer is outside the runlist, truncate has cut it out
 789		 * of the runlist.  Just clean and clear the buffer and set it
 790		 * uptodate so it can get discarded by the VM.
 791		 */
 792		if (err == -ENOENT || lcn == LCN_ENOENT) {
 793			bh->b_blocknr = -1;
 794			clear_buffer_dirty(bh);
 795			zero_user(page, bh_offset(bh), blocksize);
 796			set_buffer_uptodate(bh);
 797			err = 0;
 798			continue;
 799		}
 800		/* Failed to map the buffer, even after retrying. */
 801		if (!err)
 802			err = -EIO;
 803		bh->b_blocknr = -1;
 804		ntfs_error(vol->sb, "Failed to write to inode 0x%lx, "
 805				"attribute type 0x%x, vcn 0x%llx, offset 0x%x "
 806				"because its location on disk could not be "
 807				"determined%s (error code %i).", ni->mft_no,
 808				ni->type, (unsigned long long)vcn,
 809				vcn_ofs, is_retry ? " even after "
 810				"retrying" : "", err);
 811		break;
 812	} while (block++, (bh = bh->b_this_page) != head);
 813
 814	/* Release the lock if we took it. */
 815	if (rl)
 816		up_read(&ni->runlist.lock);
 817
 818	/* For the error case, need to reset bh to the beginning. */
 819	bh = head;
 820
 821	/* Just an optimization, so ->readpage() is not called later. */
 822	if (unlikely(!PageUptodate(page))) {
 823		int uptodate = 1;
 824		do {
 825			if (!buffer_uptodate(bh)) {
 826				uptodate = 0;
 827				bh = head;
 828				break;
 829			}
 830		} while ((bh = bh->b_this_page) != head);
 831		if (uptodate)
 832			SetPageUptodate(page);
 833	}
 834
 835	/* Setup all mapped, dirty buffers for async write i/o. */
 836	do {
 837		if (buffer_mapped(bh) && buffer_dirty(bh)) {
 838			lock_buffer(bh);
 839			if (test_clear_buffer_dirty(bh)) {
 840				BUG_ON(!buffer_uptodate(bh));
 841				mark_buffer_async_write(bh);
 842			} else
 843				unlock_buffer(bh);
 844		} else if (unlikely(err)) {
 845			/*
 846			 * For the error case. The buffer may have been set
 847			 * dirty during attachment to a dirty page.
 848			 */
 849			if (err != -ENOMEM)
 850				clear_buffer_dirty(bh);
 851		}
 852	} while ((bh = bh->b_this_page) != head);
 853
 854	if (unlikely(err)) {
 855		// TODO: Remove the -EOPNOTSUPP check later on...
 856		if (unlikely(err == -EOPNOTSUPP))
 857			err = 0;
 858		else if (err == -ENOMEM) {
 859			ntfs_warning(vol->sb, "Error allocating memory. "
 860					"Redirtying page so we try again "
 861					"later.");
 862			/*
 863			 * Put the page back on mapping->dirty_pages, but
 864			 * leave its buffer's dirty state as-is.
 865			 */
 866			redirty_page_for_writepage(wbc, page);
 867			err = 0;
 868		} else
 869			SetPageError(page);
 870	}
 871
 872	BUG_ON(PageWriteback(page));
 873	set_page_writeback(page);	/* Keeps try_to_free_buffers() away. */
 874
 875	/* Submit the prepared buffers for i/o. */
 876	need_end_writeback = true;
 877	do {
 878		struct buffer_head *next = bh->b_this_page;
 879		if (buffer_async_write(bh)) {
 880			submit_bh(WRITE, bh);
 881			need_end_writeback = false;
 882		}
 883		bh = next;
 884	} while (bh != head);
 885	unlock_page(page);
 886
 887	/* If no i/o was started, need to end_page_writeback(). */
 888	if (unlikely(need_end_writeback))
 889		end_page_writeback(page);
 890
 891	ntfs_debug("Done.");
 892	return err;
 893}
 894
 895/**
 896 * ntfs_write_mst_block - write a @page to the backing store
 897 * @page:	page cache page to write out
 898 * @wbc:	writeback control structure
 899 *
 900 * This function is for writing pages belonging to non-resident, mst protected
 901 * attributes to their backing store.  The only supported attributes are index
 902 * allocation and $MFT/$DATA.  Both directory inodes and index inodes are
 903 * supported for the index allocation case.
 904 *
 905 * The page must remain locked for the duration of the write because we apply
 906 * the mst fixups, write, and then undo the fixups, so if we were to unlock the
 907 * page before undoing the fixups, any other user of the page will see the
 908 * page contents as corrupt.
 909 *
 910 * We clear the page uptodate flag for the duration of the function to ensure
 911 * exclusion for the $MFT/$DATA case against someone mapping an mft record we
 912 * are about to apply the mst fixups to.
 913 *
 914 * Return 0 on success and -errno on error.
 915 *
 916 * Based on ntfs_write_block(), ntfs_mft_writepage(), and
 917 * write_mft_record_nolock().
 918 */
 919static int ntfs_write_mst_block(struct page *page,
 920		struct writeback_control *wbc)
 921{
 922	sector_t block, dblock, rec_block;
 923	struct inode *vi = page->mapping->host;
 924	ntfs_inode *ni = NTFS_I(vi);
 925	ntfs_volume *vol = ni->vol;
 926	u8 *kaddr;
 927	unsigned int rec_size = ni->itype.index.block_size;
 928	ntfs_inode *locked_nis[PAGE_SIZE / rec_size];
 929	struct buffer_head *bh, *head, *tbh, *rec_start_bh;
 930	struct buffer_head *bhs[MAX_BUF_PER_PAGE];
 931	runlist_element *rl;
 932	int i, nr_locked_nis, nr_recs, nr_bhs, max_bhs, bhs_per_rec, err, err2;
 933	unsigned bh_size, rec_size_bits;
 934	bool sync, is_mft, page_is_dirty, rec_is_dirty;
 935	unsigned char bh_size_bits;
 936
 937	ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
 938			"0x%lx.", vi->i_ino, ni->type, page->index);
 939	BUG_ON(!NInoNonResident(ni));
 940	BUG_ON(!NInoMstProtected(ni));
 941	is_mft = (S_ISREG(vi->i_mode) && !vi->i_ino);
 942	/*
 943	 * NOTE: ntfs_write_mst_block() would be called for $MFTMirr if a page
 944	 * in its page cache were to be marked dirty.  However this should
 945	 * never happen with the current driver and considering we do not
 946	 * handle this case here we do want to BUG(), at least for now.
 947	 */
 948	BUG_ON(!(is_mft || S_ISDIR(vi->i_mode) ||
 949			(NInoAttr(ni) && ni->type == AT_INDEX_ALLOCATION)));
 950	bh_size = vol->sb->s_blocksize;
 951	bh_size_bits = vol->sb->s_blocksize_bits;
 952	max_bhs = PAGE_SIZE / bh_size;
 953	BUG_ON(!max_bhs);
 954	BUG_ON(max_bhs > MAX_BUF_PER_PAGE);
 955
 956	/* Were we called for sync purposes? */
 957	sync = (wbc->sync_mode == WB_SYNC_ALL);
 958
 959	/* Make sure we have mapped buffers. */
 960	bh = head = page_buffers(page);
 961	BUG_ON(!bh);
 962
 963	rec_size_bits = ni->itype.index.block_size_bits;
 964	BUG_ON(!(PAGE_SIZE >> rec_size_bits));
 965	bhs_per_rec = rec_size >> bh_size_bits;
 966	BUG_ON(!bhs_per_rec);
 967
 968	/* The first block in the page. */
 969	rec_block = block = (sector_t)page->index <<
 970			(PAGE_SHIFT - bh_size_bits);
 971
 972	/* The first out of bounds block for the data size. */
 973	dblock = (i_size_read(vi) + bh_size - 1) >> bh_size_bits;
 974
 975	rl = NULL;
 976	err = err2 = nr_bhs = nr_recs = nr_locked_nis = 0;
 977	page_is_dirty = rec_is_dirty = false;
 978	rec_start_bh = NULL;
 979	do {
 980		bool is_retry = false;
 981
 982		if (likely(block < rec_block)) {
 983			if (unlikely(block >= dblock)) {
 984				clear_buffer_dirty(bh);
 985				set_buffer_uptodate(bh);
 986				continue;
 987			}
 988			/*
 989			 * This block is not the first one in the record.  We
 990			 * ignore the buffer's dirty state because we could
 991			 * have raced with a parallel mark_ntfs_record_dirty().
 992			 */
 993			if (!rec_is_dirty)
 994				continue;
 995			if (unlikely(err2)) {
 996				if (err2 != -ENOMEM)
 997					clear_buffer_dirty(bh);
 998				continue;
 999			}
1000		} else /* if (block == rec_block) */ {
1001			BUG_ON(block > rec_block);
1002			/* This block is the first one in the record. */
1003			rec_block += bhs_per_rec;
1004			err2 = 0;
1005			if (unlikely(block >= dblock)) {
1006				clear_buffer_dirty(bh);
1007				continue;
1008			}
1009			if (!buffer_dirty(bh)) {
1010				/* Clean records are not written out. */
1011				rec_is_dirty = false;
1012				continue;
1013			}
1014			rec_is_dirty = true;
1015			rec_start_bh = bh;
1016		}
1017		/* Need to map the buffer if it is not mapped already. */
1018		if (unlikely(!buffer_mapped(bh))) {
1019			VCN vcn;
1020			LCN lcn;
1021			unsigned int vcn_ofs;
1022
1023			bh->b_bdev = vol->sb->s_bdev;
1024			/* Obtain the vcn and offset of the current block. */
1025			vcn = (VCN)block << bh_size_bits;
1026			vcn_ofs = vcn & vol->cluster_size_mask;
1027			vcn >>= vol->cluster_size_bits;
1028			if (!rl) {
1029lock_retry_remap:
1030				down_read(&ni->runlist.lock);
1031				rl = ni->runlist.rl;
1032			}
1033			if (likely(rl != NULL)) {
1034				/* Seek to element containing target vcn. */
1035				while (rl->length && rl[1].vcn <= vcn)
1036					rl++;
1037				lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
1038			} else
1039				lcn = LCN_RL_NOT_MAPPED;
1040			/* Successful remap. */
1041			if (likely(lcn >= 0)) {
1042				/* Setup buffer head to correct block. */
1043				bh->b_blocknr = ((lcn <<
1044						vol->cluster_size_bits) +
1045						vcn_ofs) >> bh_size_bits;
1046				set_buffer_mapped(bh);
1047			} else {
1048				/*
1049				 * Remap failed.  Retry to map the runlist once
1050				 * unless we are working on $MFT which always
1051				 * has the whole of its runlist in memory.
1052				 */
1053				if (!is_mft && !is_retry &&
1054						lcn == LCN_RL_NOT_MAPPED) {
1055					is_retry = true;
1056					/*
1057					 * Attempt to map runlist, dropping
1058					 * lock for the duration.
1059					 */
1060					up_read(&ni->runlist.lock);
1061					err2 = ntfs_map_runlist(ni, vcn);
1062					if (likely(!err2))
1063						goto lock_retry_remap;
1064					if (err2 == -ENOMEM)
1065						page_is_dirty = true;
1066					lcn = err2;
1067				} else {
1068					err2 = -EIO;
1069					if (!rl)
1070						up_read(&ni->runlist.lock);
1071				}
1072				/* Hard error.  Abort writing this record. */
1073				if (!err || err == -ENOMEM)
1074					err = err2;
1075				bh->b_blocknr = -1;
1076				ntfs_error(vol->sb, "Cannot write ntfs record "
1077						"0x%llx (inode 0x%lx, "
1078						"attribute type 0x%x) because "
1079						"its location on disk could "
1080						"not be determined (error "
1081						"code %lli).",
1082						(long long)block <<
1083						bh_size_bits >>
1084						vol->mft_record_size_bits,
1085						ni->mft_no, ni->type,
1086						(long long)lcn);
1087				/*
1088				 * If this is not the first buffer, remove the
1089				 * buffers in this record from the list of
1090				 * buffers to write and clear their dirty bit
1091				 * if not error -ENOMEM.
1092				 */
1093				if (rec_start_bh != bh) {
1094					while (bhs[--nr_bhs] != rec_start_bh)
1095						;
1096					if (err2 != -ENOMEM) {
1097						do {
1098							clear_buffer_dirty(
1099								rec_start_bh);
1100						} while ((rec_start_bh =
1101								rec_start_bh->
1102								b_this_page) !=
1103								bh);
1104					}
1105				}
1106				continue;
1107			}
1108		}
1109		BUG_ON(!buffer_uptodate(bh));
1110		BUG_ON(nr_bhs >= max_bhs);
1111		bhs[nr_bhs++] = bh;
1112	} while (block++, (bh = bh->b_this_page) != head);
1113	if (unlikely(rl))
1114		up_read(&ni->runlist.lock);
1115	/* If there were no dirty buffers, we are done. */
1116	if (!nr_bhs)
1117		goto done;
1118	/* Map the page so we can access its contents. */
1119	kaddr = kmap(page);
1120	/* Clear the page uptodate flag whilst the mst fixups are applied. */
1121	BUG_ON(!PageUptodate(page));
1122	ClearPageUptodate(page);
1123	for (i = 0; i < nr_bhs; i++) {
1124		unsigned int ofs;
1125
1126		/* Skip buffers which are not at the beginning of records. */
1127		if (i % bhs_per_rec)
1128			continue;
1129		tbh = bhs[i];
1130		ofs = bh_offset(tbh);
1131		if (is_mft) {
1132			ntfs_inode *tni;
1133			unsigned long mft_no;
1134
1135			/* Get the mft record number. */
1136			mft_no = (((s64)page->index << PAGE_SHIFT) + ofs)
1137					>> rec_size_bits;
1138			/* Check whether to write this mft record. */
1139			tni = NULL;
1140			if (!ntfs_may_write_mft_record(vol, mft_no,
1141					(MFT_RECORD*)(kaddr + ofs), &tni)) {
1142				/*
1143				 * The record should not be written.  This
1144				 * means we need to redirty the page before
1145				 * returning.
1146				 */
1147				page_is_dirty = true;
1148				/*
1149				 * Remove the buffers in this mft record from
1150				 * the list of buffers to write.
1151				 */
1152				do {
1153					bhs[i] = NULL;
1154				} while (++i % bhs_per_rec);
1155				continue;
1156			}
1157			/*
1158			 * The record should be written.  If a locked ntfs
1159			 * inode was returned, add it to the array of locked
1160			 * ntfs inodes.
1161			 */
1162			if (tni)
1163				locked_nis[nr_locked_nis++] = tni;
1164		}
1165		/* Apply the mst protection fixups. */
1166		err2 = pre_write_mst_fixup((NTFS_RECORD*)(kaddr + ofs),
1167				rec_size);
1168		if (unlikely(err2)) {
1169			if (!err || err == -ENOMEM)
1170				err = -EIO;
1171			ntfs_error(vol->sb, "Failed to apply mst fixups "
1172					"(inode 0x%lx, attribute type 0x%x, "
1173					"page index 0x%lx, page offset 0x%x)!"
1174					"  Unmount and run chkdsk.", vi->i_ino,
1175					ni->type, page->index, ofs);
1176			/*
1177			 * Mark all the buffers in this record clean as we do
1178			 * not want to write corrupt data to disk.
1179			 */
1180			do {
1181				clear_buffer_dirty(bhs[i]);
1182				bhs[i] = NULL;
1183			} while (++i % bhs_per_rec);
1184			continue;
1185		}
1186		nr_recs++;
1187	}
1188	/* If no records are to be written out, we are done. */
1189	if (!nr_recs)
1190		goto unm_done;
1191	flush_dcache_page(page);
1192	/* Lock buffers and start synchronous write i/o on them. */
1193	for (i = 0; i < nr_bhs; i++) {
1194		tbh = bhs[i];
1195		if (!tbh)
1196			continue;
1197		if (!trylock_buffer(tbh))
1198			BUG();
1199		/* The buffer dirty state is now irrelevant, just clean it. */
1200		clear_buffer_dirty(tbh);
1201		BUG_ON(!buffer_uptodate(tbh));
1202		BUG_ON(!buffer_mapped(tbh));
1203		get_bh(tbh);
1204		tbh->b_end_io = end_buffer_write_sync;
1205		submit_bh(WRITE, tbh);
1206	}
1207	/* Synchronize the mft mirror now if not @sync. */
1208	if (is_mft && !sync)
1209		goto do_mirror;
1210do_wait:
1211	/* Wait on i/o completion of buffers. */
1212	for (i = 0; i < nr_bhs; i++) {
1213		tbh = bhs[i];
1214		if (!tbh)
1215			continue;
1216		wait_on_buffer(tbh);
1217		if (unlikely(!buffer_uptodate(tbh))) {
1218			ntfs_error(vol->sb, "I/O error while writing ntfs "
1219					"record buffer (inode 0x%lx, "
1220					"attribute type 0x%x, page index "
1221					"0x%lx, page offset 0x%lx)!  Unmount "
1222					"and run chkdsk.", vi->i_ino, ni->type,
1223					page->index, bh_offset(tbh));
1224			if (!err || err == -ENOMEM)
1225				err = -EIO;
1226			/*
1227			 * Set the buffer uptodate so the page and buffer
1228			 * states do not become out of sync.
1229			 */
1230			set_buffer_uptodate(tbh);
1231		}
1232	}
1233	/* If @sync, now synchronize the mft mirror. */
1234	if (is_mft && sync) {
1235do_mirror:
1236		for (i = 0; i < nr_bhs; i++) {
1237			unsigned long mft_no;
1238			unsigned int ofs;
1239
1240			/*
1241			 * Skip buffers which are not at the beginning of
1242			 * records.
1243			 */
1244			if (i % bhs_per_rec)
1245				continue;
1246			tbh = bhs[i];
1247			/* Skip removed buffers (and hence records). */
1248			if (!tbh)
1249				continue;
1250			ofs = bh_offset(tbh);
1251			/* Get the mft record number. */
1252			mft_no = (((s64)page->index << PAGE_SHIFT) + ofs)
1253					>> rec_size_bits;
1254			if (mft_no < vol->mftmirr_size)
1255				ntfs_sync_mft_mirror(vol, mft_no,
1256						(MFT_RECORD*)(kaddr + ofs),
1257						sync);
1258		}
1259		if (!sync)
1260			goto do_wait;
1261	}
1262	/* Remove the mst protection fixups again. */
1263	for (i = 0; i < nr_bhs; i++) {
1264		if (!(i % bhs_per_rec)) {
1265			tbh = bhs[i];
1266			if (!tbh)
1267				continue;
1268			post_write_mst_fixup((NTFS_RECORD*)(kaddr +
1269					bh_offset(tbh)));
1270		}
1271	}
1272	flush_dcache_page(page);
1273unm_done:
1274	/* Unlock any locked inodes. */
1275	while (nr_locked_nis-- > 0) {
1276		ntfs_inode *tni, *base_tni;
1277		
1278		tni = locked_nis[nr_locked_nis];
1279		/* Get the base inode. */
1280		mutex_lock(&tni->extent_lock);
1281		if (tni->nr_extents >= 0)
1282			base_tni = tni;
1283		else {
1284			base_tni = tni->ext.base_ntfs_ino;
1285			BUG_ON(!base_tni);
1286		}
1287		mutex_unlock(&tni->extent_lock);
1288		ntfs_debug("Unlocking %s inode 0x%lx.",
1289				tni == base_tni ? "base" : "extent",
1290				tni->mft_no);
1291		mutex_unlock(&tni->mrec_lock);
1292		atomic_dec(&tni->count);
1293		iput(VFS_I(base_tni));
1294	}
1295	SetPageUptodate(page);
1296	kunmap(page);
1297done:
1298	if (unlikely(err && err != -ENOMEM)) {
1299		/*
1300		 * Set page error if there is only one ntfs record in the page.
1301		 * Otherwise we would loose per-record granularity.
1302		 */
1303		if (ni->itype.index.block_size == PAGE_SIZE)
1304			SetPageError(page);
1305		NVolSetErrors(vol);
1306	}
1307	if (page_is_dirty) {
1308		ntfs_debug("Page still contains one or more dirty ntfs "
1309				"records.  Redirtying the page starting at "
1310				"record 0x%lx.", page->index <<
1311				(PAGE_SHIFT - rec_size_bits));
1312		redirty_page_for_writepage(wbc, page);
1313		unlock_page(page);
1314	} else {
1315		/*
1316		 * Keep the VM happy.  This must be done otherwise the
1317		 * radix-tree tag PAGECACHE_TAG_DIRTY remains set even though
1318		 * the page is clean.
1319		 */
1320		BUG_ON(PageWriteback(page));
1321		set_page_writeback(page);
1322		unlock_page(page);
1323		end_page_writeback(page);
1324	}
1325	if (likely(!err))
1326		ntfs_debug("Done.");
1327	return err;
1328}
1329
1330/**
1331 * ntfs_writepage - write a @page to the backing store
1332 * @page:	page cache page to write out
1333 * @wbc:	writeback control structure
1334 *
1335 * This is called from the VM when it wants to have a dirty ntfs page cache
1336 * page cleaned.  The VM has already locked the page and marked it clean.
1337 *
1338 * For non-resident attributes, ntfs_writepage() writes the @page by calling
1339 * the ntfs version of the generic block_write_full_page() function,
1340 * ntfs_write_block(), which in turn if necessary creates and writes the
1341 * buffers associated with the page asynchronously.
1342 *
1343 * For resident attributes, OTOH, ntfs_writepage() writes the @page by copying
1344 * the data to the mft record (which at this stage is most likely in memory).
1345 * The mft record is then marked dirty and written out asynchronously via the
1346 * vfs inode dirty code path for the inode the mft record belongs to or via the
1347 * vm page dirty code path for the page the mft record is in.
1348 *
1349 * Based on ntfs_readpage() and fs/buffer.c::block_write_full_page().
1350 *
1351 * Return 0 on success and -errno on error.
1352 */
1353static int ntfs_writepage(struct page *page, struct writeback_control *wbc)
1354{
1355	loff_t i_size;
1356	struct inode *vi = page->mapping->host;
1357	ntfs_inode *base_ni = NULL, *ni = NTFS_I(vi);
1358	char *addr;
1359	ntfs_attr_search_ctx *ctx = NULL;
1360	MFT_RECORD *m = NULL;
1361	u32 attr_len;
1362	int err;
1363
1364retry_writepage:
1365	BUG_ON(!PageLocked(page));
1366	i_size = i_size_read(vi);
1367	/* Is the page fully outside i_size? (truncate in progress) */
1368	if (unlikely(page->index >= (i_size + PAGE_SIZE - 1) >>
1369			PAGE_SHIFT)) {
1370		/*
1371		 * The page may have dirty, unmapped buffers.  Make them
1372		 * freeable here, so the page does not leak.
1373		 */
1374		block_invalidatepage(page, 0, PAGE_SIZE);
1375		unlock_page(page);
1376		ntfs_debug("Write outside i_size - truncated?");
1377		return 0;
1378	}
1379	/*
1380	 * Only $DATA attributes can be encrypted and only unnamed $DATA
1381	 * attributes can be compressed.  Index root can have the flags set but
1382	 * this means to create compressed/encrypted files, not that the
1383	 * attribute is compressed/encrypted.  Note we need to check for
1384	 * AT_INDEX_ALLOCATION since this is the type of both directory and
1385	 * index inodes.
1386	 */
1387	if (ni->type != AT_INDEX_ALLOCATION) {
1388		/* If file is encrypted, deny access, just like NT4. */
1389		if (NInoEncrypted(ni)) {
1390			unlock_page(page);
1391			BUG_ON(ni->type != AT_DATA);
1392			ntfs_debug("Denying write access to encrypted file.");
1393			return -EACCES;
1394		}
1395		/* Compressed data streams are handled in compress.c. */
1396		if (NInoNonResident(ni) && NInoCompressed(ni)) {
1397			BUG_ON(ni->type != AT_DATA);
1398			BUG_ON(ni->name_len);
1399			// TODO: Implement and replace this with
1400			// return ntfs_write_compressed_block(page);
1401			unlock_page(page);
1402			ntfs_error(vi->i_sb, "Writing to compressed files is "
1403					"not supported yet.  Sorry.");
1404			return -EOPNOTSUPP;
1405		}
1406		// TODO: Implement and remove this check.
1407		if (NInoNonResident(ni) && NInoSparse(ni)) {
1408			unlock_page(page);
1409			ntfs_error(vi->i_sb, "Writing to sparse files is not "
1410					"supported yet.  Sorry.");
1411			return -EOPNOTSUPP;
1412		}
1413	}
1414	/* NInoNonResident() == NInoIndexAllocPresent() */
1415	if (NInoNonResident(ni)) {
1416		/* We have to zero every time due to mmap-at-end-of-file. */
1417		if (page->index >= (i_size >> PAGE_SHIFT)) {
1418			/* The page straddles i_size. */
1419			unsigned int ofs = i_size & ~PAGE_MASK;
1420			zero_user_segment(page, ofs, PAGE_SIZE);
1421		}
1422		/* Handle mst protected attributes. */
1423		if (NInoMstProtected(ni))
1424			return ntfs_write_mst_block(page, wbc);
1425		/* Normal, non-resident data stream. */
1426		return ntfs_write_block(page, wbc);
1427	}
1428	/*
1429	 * Attribute is resident, implying it is not compressed, encrypted, or
1430	 * mst protected.  This also means the attribute is smaller than an mft
1431	 * record and hence smaller than a page, so can simply return error on
1432	 * any pages with index above 0.  Note the attribute can actually be
1433	 * marked compressed but if it is resident the actual data is not
1434	 * compressed so we are ok to ignore the compressed flag here.
1435	 */
1436	BUG_ON(page_has_buffers(page));
1437	BUG_ON(!PageUptodate(page));
1438	if (unlikely(page->index > 0)) {
1439		ntfs_error(vi->i_sb, "BUG()! page->index (0x%lx) > 0.  "
1440				"Aborting write.", page->index);
1441		BUG_ON(PageWriteback(page));
1442		set_page_writeback(page);
1443		unlock_page(page);
1444		end_page_writeback(page);
1445		return -EIO;
1446	}
1447	if (!NInoAttr(ni))
1448		base_ni = ni;
1449	else
1450		base_ni = ni->ext.base_ntfs_ino;
1451	/* Map, pin, and lock the mft record. */
1452	m = map_mft_record(base_ni);
1453	if (IS_ERR(m)) {
1454		err = PTR_ERR(m);
1455		m = NULL;
1456		ctx = NULL;
1457		goto err_out;
1458	}
1459	/*
1460	 * If a parallel write made the attribute non-resident, drop the mft
1461	 * record and retry the writepage.
1462	 */
1463	if (unlikely(NInoNonResident(ni))) {
1464		unmap_mft_record(base_ni);
1465		goto retry_writepage;
1466	}
1467	ctx = ntfs_attr_get_search_ctx(base_ni, m);
1468	if (unlikely(!ctx)) {
1469		err = -ENOMEM;
1470		goto err_out;
1471	}
1472	err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1473			CASE_SENSITIVE, 0, NULL, 0, ctx);
1474	if (unlikely(err))
1475		goto err_out;
1476	/*
1477	 * Keep the VM happy.  This must be done otherwise the radix-tree tag
1478	 * PAGECACHE_TAG_DIRTY remains set even though the page is clean.
1479	 */
1480	BUG_ON(PageWriteback(page));
1481	set_page_writeback(page);
1482	unlock_page(page);
1483	attr_len = le32_to_cpu(ctx->attr->data.resident.value_length);
1484	i_size = i_size_read(vi);
1485	if (unlikely(attr_len > i_size)) {
1486		/* Race with shrinking truncate or a failed truncate. */
1487		attr_len = i_size;
1488		/*
1489		 * If the truncate failed, fix it up now.  If a concurrent
1490		 * truncate, we do its job, so it does not have to do anything.
1491		 */
1492		err = ntfs_resident_attr_value_resize(ctx->mrec, ctx->attr,
1493				attr_len);
1494		/* Shrinking cannot fail. */
1495		BUG_ON(err);
1496	}
1497	addr = kmap_atomic(page);
1498	/* Copy the data from the page to the mft record. */
1499	memcpy((u8*)ctx->attr +
1500			le16_to_cpu(ctx->attr->data.resident.value_offset),
1501			addr, attr_len);
1502	/* Zero out of bounds area in the page cache page. */
1503	memset(addr + attr_len, 0, PAGE_SIZE - attr_len);
1504	kunmap_atomic(addr);
1505	flush_dcache_page(page);
1506	flush_dcache_mft_record_page(ctx->ntfs_ino);
1507	/* We are done with the page. */
1508	end_page_writeback(page);
1509	/* Finally, mark the mft record dirty, so it gets written back. */
1510	mark_mft_record_dirty(ctx->ntfs_ino);
1511	ntfs_attr_put_search_ctx(ctx);
1512	unmap_mft_record(base_ni);
1513	return 0;
1514err_out:
1515	if (err == -ENOMEM) {
1516		ntfs_warning(vi->i_sb, "Error allocating memory. Redirtying "
1517				"page so we try again later.");
1518		/*
1519		 * Put the page back on mapping->dirty_pages, but leave its
1520		 * buffers' dirty state as-is.
1521		 */
1522		redirty_page_for_writepage(wbc, page);
1523		err = 0;
1524	} else {
1525		ntfs_error(vi->i_sb, "Resident attribute write failed with "
1526				"error %i.", err);
1527		SetPageError(page);
1528		NVolSetErrors(ni->vol);
1529	}
1530	unlock_page(page);
1531	if (ctx)
1532		ntfs_attr_put_search_ctx(ctx);
1533	if (m)
1534		unmap_mft_record(base_ni);
1535	return err;
1536}
1537
1538#endif	/* NTFS_RW */
1539
1540/**
1541 * ntfs_bmap - map logical file block to physical device block
1542 * @mapping:	address space mapping to which the block to be mapped belongs
1543 * @block:	logical block to map to its physical device block
1544 *
1545 * For regular, non-resident files (i.e. not compressed and not encrypted), map
1546 * the logical @block belonging to the file described by the address space
1547 * mapping @mapping to its physical device block.
1548 *
1549 * The size of the block is equal to the @s_blocksize field of the super block
1550 * of the mounted file system which is guaranteed to be smaller than or equal
1551 * to the cluster size thus the block is guaranteed to fit entirely inside the
1552 * cluster which means we do not need to care how many contiguous bytes are
1553 * available after the beginning of the block.
1554 *
1555 * Return the physical device block if the mapping succeeded or 0 if the block
1556 * is sparse or there was an error.
1557 *
1558 * Note: This is a problem if someone tries to run bmap() on $Boot system file
1559 * as that really is in block zero but there is nothing we can do.  bmap() is
1560 * just broken in that respect (just like it cannot distinguish sparse from
1561 * not available or error).
1562 */
1563static sector_t ntfs_bmap(struct address_space *mapping, sector_t block)
1564{
1565	s64 ofs, size;
1566	loff_t i_size;
1567	LCN lcn;
1568	unsigned long blocksize, flags;
1569	ntfs_inode *ni = NTFS_I(mapping->host);
1570	ntfs_volume *vol = ni->vol;
1571	unsigned delta;
1572	unsigned char blocksize_bits, cluster_size_shift;
1573
1574	ntfs_debug("Entering for mft_no 0x%lx, logical block 0x%llx.",
1575			ni->mft_no, (unsigned long long)block);
1576	if (ni->type != AT_DATA || !NInoNonResident(ni) || NInoEncrypted(ni)) {
1577		ntfs_error(vol->sb, "BMAP does not make sense for %s "
1578				"attributes, returning 0.",
1579				(ni->type != AT_DATA) ? "non-data" :
1580				(!NInoNonResident(ni) ? "resident" :
1581				"encrypted"));
1582		return 0;
1583	}
1584	/* None of these can happen. */
1585	BUG_ON(NInoCompressed(ni));
1586	BUG_ON(NInoMstProtected(ni));
1587	blocksize = vol->sb->s_blocksize;
1588	blocksize_bits = vol->sb->s_blocksize_bits;
1589	ofs = (s64)block << blocksize_bits;
1590	read_lock_irqsave(&ni->size_lock, flags);
1591	size = ni->initialized_size;
1592	i_size = i_size_read(VFS_I(ni));
1593	read_unlock_irqrestore(&ni->size_lock, flags);
1594	/*
1595	 * If the offset is outside the initialized size or the block straddles
1596	 * the initialized size then pretend it is a hole unless the
1597	 * initialized size equals the file size.
1598	 */
1599	if (unlikely(ofs >= size || (ofs + blocksize > size && size < i_size)))
1600		goto hole;
1601	cluster_size_shift = vol->cluster_size_bits;
1602	down_read(&ni->runlist.lock);
1603	lcn = ntfs_attr_vcn_to_lcn_nolock(ni, ofs >> cluster_size_shift, false);
1604	up_read(&ni->runlist.lock);
1605	if (unlikely(lcn < LCN_HOLE)) {
1606		/*
1607		 * Step down to an integer to avoid gcc doing a long long
1608		 * comparision in the switch when we know @lcn is between
1609		 * LCN_HOLE and LCN_EIO (i.e. -1 to -5).
1610		 *
1611		 * Otherwise older gcc (at least on some architectures) will
1612		 * try to use __cmpdi2() which is of course not available in
1613		 * the kernel.
1614		 */
1615		switch ((int)lcn) {
1616		case LCN_ENOENT:
1617			/*
1618			 * If the offset is out of bounds then pretend it is a
1619			 * hole.
1620			 */
1621			goto hole;
1622		case LCN_ENOMEM:
1623			ntfs_error(vol->sb, "Not enough memory to complete "
1624					"mapping for inode 0x%lx.  "
1625					"Returning 0.", ni->mft_no);
1626			break;
1627		default:
1628			ntfs_error(vol->sb, "Failed to complete mapping for "
1629					"inode 0x%lx.  Run chkdsk.  "
1630					"Returning 0.", ni->mft_no);
1631			break;
1632		}
1633		return 0;
1634	}
1635	if (lcn < 0) {
1636		/* It is a hole. */
1637hole:
1638		ntfs_debug("Done (returning hole).");
1639		return 0;
1640	}
1641	/*
1642	 * The block is really allocated and fullfils all our criteria.
1643	 * Convert the cluster to units of block size and return the result.
1644	 */
1645	delta = ofs & vol->cluster_size_mask;
1646	if (unlikely(sizeof(block) < sizeof(lcn))) {
1647		block = lcn = ((lcn << cluster_size_shift) + delta) >>
1648				blocksize_bits;
1649		/* If the block number was truncated return 0. */
1650		if (unlikely(block != lcn)) {
1651			ntfs_error(vol->sb, "Physical block 0x%llx is too "
1652					"large to be returned, returning 0.",
1653					(long long)lcn);
1654			return 0;
1655		}
1656	} else
1657		block = ((lcn << cluster_size_shift) + delta) >>
1658				blocksize_bits;
1659	ntfs_debug("Done (returning block 0x%llx).", (unsigned long long)lcn);
1660	return block;
1661}
1662
1663/**
1664 * ntfs_normal_aops - address space operations for normal inodes and attributes
1665 *
1666 * Note these are not used for compressed or mst protected inodes and
1667 * attributes.
1668 */
1669const struct address_space_operations ntfs_normal_aops = {
1670	.readpage	= ntfs_readpage,
1671#ifdef NTFS_RW
1672	.writepage	= ntfs_writepage,
1673	.set_page_dirty	= __set_page_dirty_buffers,
1674#endif /* NTFS_RW */
1675	.bmap		= ntfs_bmap,
1676	.migratepage	= buffer_migrate_page,
1677	.is_partially_uptodate = block_is_partially_uptodate,
1678	.error_remove_page = generic_error_remove_page,
1679};
1680
1681/**
1682 * ntfs_compressed_aops - address space operations for compressed inodes
1683 */
1684const struct address_space_operations ntfs_compressed_aops = {
1685	.readpage	= ntfs_readpage,
1686#ifdef NTFS_RW
1687	.writepage	= ntfs_writepage,
1688	.set_page_dirty	= __set_page_dirty_buffers,
1689#endif /* NTFS_RW */
1690	.migratepage	= buffer_migrate_page,
1691	.is_partially_uptodate = block_is_partially_uptodate,
1692	.error_remove_page = generic_error_remove_page,
1693};
1694
1695/**
1696 * ntfs_mst_aops - general address space operations for mst protecteed inodes
1697 *		   and attributes
1698 */
1699const struct address_space_operations ntfs_mst_aops = {
1700	.readpage	= ntfs_readpage,	/* Fill page with data. */
1701#ifdef NTFS_RW
1702	.writepage	= ntfs_writepage,	/* Write dirty page to disk. */
1703	.set_page_dirty	= __set_page_dirty_nobuffers,	/* Set the page dirty
1704						   without touching the buffers
1705						   belonging to the page. */
1706#endif /* NTFS_RW */
1707	.migratepage	= buffer_migrate_page,
1708	.is_partially_uptodate	= block_is_partially_uptodate,
1709	.error_remove_page = generic_error_remove_page,
1710};
1711
1712#ifdef NTFS_RW
1713
1714/**
1715 * mark_ntfs_record_dirty - mark an ntfs record dirty
1716 * @page:	page containing the ntfs record to mark dirty
1717 * @ofs:	byte offset within @page at which the ntfs record begins
1718 *
1719 * Set the buffers and the page in which the ntfs record is located dirty.
1720 *
1721 * The latter also marks the vfs inode the ntfs record belongs to dirty
1722 * (I_DIRTY_PAGES only).
1723 *
1724 * If the page does not have buffers, we create them and set them uptodate.
1725 * The page may not be locked which is why we need to handle the buffers under
1726 * the mapping->private_lock.  Once the buffers are marked dirty we no longer
1727 * need the lock since try_to_free_buffers() does not free dirty buffers.
1728 */
1729void mark_ntfs_record_dirty(struct page *page, const unsigned int ofs) {
1730	struct address_space *mapping = page->mapping;
1731	ntfs_inode *ni = NTFS_I(mapping->host);
1732	struct buffer_head *bh, *head, *buffers_to_free = NULL;
1733	unsigned int end, bh_size, bh_ofs;
1734
1735	BUG_ON(!PageUptodate(page));
1736	end = ofs + ni->itype.index.block_size;
1737	bh_size = VFS_I(ni)->i_sb->s_blocksize;
1738	spin_lock(&mapping->private_lock);
1739	if (unlikely(!page_has_buffers(page))) {
1740		spin_unlock(&mapping->private_lock);
1741		bh = head = alloc_page_buffers(page, bh_size, 1);
1742		spin_lock(&mapping->private_lock);
1743		if (likely(!page_has_buffers(page))) {
1744			struct buffer_head *tail;
1745
1746			do {
1747				set_buffer_uptodate(bh);
1748				tail = bh;
1749				bh = bh->b_this_page;
1750			} while (bh);
1751			tail->b_this_page = head;
1752			attach_page_buffers(page, head);
1753		} else
1754			buffers_to_free = bh;
1755	}
1756	bh = head = page_buffers(page);
1757	BUG_ON(!bh);
1758	do {
1759		bh_ofs = bh_offset(bh);
1760		if (bh_ofs + bh_size <= ofs)
1761			continue;
1762		if (unlikely(bh_ofs >= end))
1763			break;
1764		set_buffer_dirty(bh);
1765	} while ((bh = bh->b_this_page) != head);
1766	spin_unlock(&mapping->private_lock);
1767	__set_page_dirty_nobuffers(page);
1768	if (unlikely(buffers_to_free)) {
1769		do {
1770			bh = buffers_to_free->b_this_page;
1771			free_buffer_head(buffers_to_free);
1772			buffers_to_free = bh;
1773		} while (buffers_to_free);
1774	}
1775}
1776
1777#endif /* NTFS_RW */