1/**
   2 * aops.c - NTFS kernel address space operations and page cache handling.
   3 *	    Part of the Linux-NTFS project.
   4 *
   5 * Copyright (c) 2001-2007 Anton Altaparmakov
   6 * Copyright (c) 2002 Richard Russon
   7 *
   8 * This program/include file is free software; you can redistribute it and/or
   9 * modify it under the terms of the GNU General Public License as published
  10 * by the Free Software Foundation; either version 2 of the License, or
  11 * (at your option) any later version.
  12 *
  13 * This program/include file is distributed in the hope that it will be
  14 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  15 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16 * GNU General Public License for more details.
  17 *
  18 * You should have received a copy of the GNU General Public License
  19 * along with this program (in the main directory of the Linux-NTFS
  20 * distribution in the file COPYING); if not, write to the Free Software
  21 * Foundation,Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  22 */
  23
  24#include <linux/errno.h>
  25#include <linux/fs.h>
  26#include <linux/gfp.h>
  27#include <linux/mm.h>
  28#include <linux/pagemap.h>
  29#include <linux/swap.h>
  30#include <linux/buffer_head.h>
  31#include <linux/writeback.h>
  32#include <linux/bit_spinlock.h>
 
  33
  34#include "aops.h"
  35#include "attrib.h"
  36#include "debug.h"
  37#include "inode.h"
  38#include "mft.h"
  39#include "runlist.h"
  40#include "types.h"
  41#include "ntfs.h"
  42
  43/**
  44 * ntfs_end_buffer_async_read - async io completion for reading attributes
  45 * @bh:		buffer head on which io is completed
  46 * @uptodate:	whether @bh is now uptodate or not
  47 *
  48 * Asynchronous I/O completion handler for reading pages belonging to the
  49 * attribute address space of an inode.  The inodes can either be files or
  50 * directories or they can be fake inodes describing some attribute.
  51 *
  52 * If NInoMstProtected(), perform the post read mst fixups when all IO on the
  53 * page has been completed and mark the page uptodate or set the error bit on
  54 * the page.  To determine the size of the records that need fixing up, we
  55 * cheat a little bit by setting the index_block_size in ntfs_inode to the ntfs
  56 * record size, and index_block_size_bits, to the log(base 2) of the ntfs
  57 * record size.
  58 */
  59static void ntfs_end_buffer_async_read(struct buffer_head *bh, int uptodate)
  60{
  61	unsigned long flags;
  62	struct buffer_head *first, *tmp;
  63	struct page *page;
  64	struct inode *vi;
  65	ntfs_inode *ni;
  66	int page_uptodate = 1;
  67
  68	page = bh->b_page;
  69	vi = page->mapping->host;
  70	ni = NTFS_I(vi);
  71
  72	if (likely(uptodate)) {
  73		loff_t i_size;
  74		s64 file_ofs, init_size;
  75
  76		set_buffer_uptodate(bh);
  77
  78		file_ofs = ((s64)page->index << PAGE_CACHE_SHIFT) +
  79				bh_offset(bh);
  80		read_lock_irqsave(&ni->size_lock, flags);
  81		init_size = ni->initialized_size;
  82		i_size = i_size_read(vi);
  83		read_unlock_irqrestore(&ni->size_lock, flags);
  84		if (unlikely(init_size > i_size)) {
  85			/* Race with shrinking truncate. */
  86			init_size = i_size;
  87		}
  88		/* Check for the current buffer head overflowing. */
  89		if (unlikely(file_ofs + bh->b_size > init_size)) {
  90			int ofs;
  91			void *kaddr;
  92
  93			ofs = 0;
  94			if (file_ofs < init_size)
  95				ofs = init_size - file_ofs;
  96			local_irq_save(flags);
  97			kaddr = kmap_atomic(page, KM_BIO_SRC_IRQ);
  98			memset(kaddr + bh_offset(bh) + ofs, 0,
  99					bh->b_size - ofs);
 100			flush_dcache_page(page);
 101			kunmap_atomic(kaddr, KM_BIO_SRC_IRQ);
 102			local_irq_restore(flags);
 103		}
 104	} else {
 105		clear_buffer_uptodate(bh);
 106		SetPageError(page);
 107		ntfs_error(ni->vol->sb, "Buffer I/O error, logical block "
 108				"0x%llx.", (unsigned long long)bh->b_blocknr);
 109	}
 110	first = page_buffers(page);
 111	local_irq_save(flags);
 112	bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
 113	clear_buffer_async_read(bh);
 114	unlock_buffer(bh);
 115	tmp = bh;
 116	do {
 117		if (!buffer_uptodate(tmp))
 118			page_uptodate = 0;
 119		if (buffer_async_read(tmp)) {
 120			if (likely(buffer_locked(tmp)))
 121				goto still_busy;
 122			/* Async buffers must be locked. */
 123			BUG();
 124		}
 125		tmp = tmp->b_this_page;
 126	} while (tmp != bh);
 127	bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
 128	local_irq_restore(flags);
 129	/*
 130	 * If none of the buffers had errors then we can set the page uptodate,
 131	 * but we first have to perform the post read mst fixups, if the
 132	 * attribute is mst protected, i.e. if NInoMstProteced(ni) is true.
 133	 * Note we ignore fixup errors as those are detected when
 134	 * map_mft_record() is called which gives us per record granularity
 135	 * rather than per page granularity.
 136	 */
 137	if (!NInoMstProtected(ni)) {
 138		if (likely(page_uptodate && !PageError(page)))
 139			SetPageUptodate(page);
 140	} else {
 141		u8 *kaddr;
 142		unsigned int i, recs;
 143		u32 rec_size;
 144
 145		rec_size = ni->itype.index.block_size;
 146		recs = PAGE_CACHE_SIZE / rec_size;
 147		/* Should have been verified before we got here... */
 148		BUG_ON(!recs);
 149		local_irq_save(flags);
 150		kaddr = kmap_atomic(page, KM_BIO_SRC_IRQ);
 151		for (i = 0; i < recs; i++)
 152			post_read_mst_fixup((NTFS_RECORD*)(kaddr +
 153					i * rec_size), rec_size);
 154		kunmap_atomic(kaddr, KM_BIO_SRC_IRQ);
 155		local_irq_restore(flags);
 156		flush_dcache_page(page);
 157		if (likely(page_uptodate && !PageError(page)))
 158			SetPageUptodate(page);
 159	}
 160	unlock_page(page);
 161	return;
 162still_busy:
 163	bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
 164	local_irq_restore(flags);
 165	return;
 166}
 167
 168/**
 169 * ntfs_read_block - fill a @page of an address space with data
 170 * @page:	page cache page to fill with data
 171 *
 172 * Fill the page @page of the address space belonging to the @page->host inode.
 173 * We read each buffer asynchronously and when all buffers are read in, our io
 174 * completion handler ntfs_end_buffer_read_async(), if required, automatically
 175 * applies the mst fixups to the page before finally marking it uptodate and
 176 * unlocking it.
 177 *
 178 * We only enforce allocated_size limit because i_size is checked for in
 179 * generic_file_read().
 180 *
 181 * Return 0 on success and -errno on error.
 182 *
 183 * Contains an adapted version of fs/buffer.c::block_read_full_page().
 184 */
 185static int ntfs_read_block(struct page *page)
 186{
 187	loff_t i_size;
 188	VCN vcn;
 189	LCN lcn;
 190	s64 init_size;
 191	struct inode *vi;
 192	ntfs_inode *ni;
 193	ntfs_volume *vol;
 194	runlist_element *rl;
 195	struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
 196	sector_t iblock, lblock, zblock;
 197	unsigned long flags;
 198	unsigned int blocksize, vcn_ofs;
 199	int i, nr;
 200	unsigned char blocksize_bits;
 201
 202	vi = page->mapping->host;
 203	ni = NTFS_I(vi);
 204	vol = ni->vol;
 205
 206	/* $MFT/$DATA must have its complete runlist in memory at all times. */
 207	BUG_ON(!ni->runlist.rl && !ni->mft_no && !NInoAttr(ni));
 208
 209	blocksize = vol->sb->s_blocksize;
 210	blocksize_bits = vol->sb->s_blocksize_bits;
 211
 212	if (!page_has_buffers(page)) {
 213		create_empty_buffers(page, blocksize, 0);
 214		if (unlikely(!page_has_buffers(page))) {
 215			unlock_page(page);
 216			return -ENOMEM;
 217		}
 218	}
 219	bh = head = page_buffers(page);
 220	BUG_ON(!bh);
 221
 222	/*
 223	 * We may be racing with truncate.  To avoid some of the problems we
 224	 * now take a snapshot of the various sizes and use those for the whole
 225	 * of the function.  In case of an extending truncate it just means we
 226	 * may leave some buffers unmapped which are now allocated.  This is
 227	 * not a problem since these buffers will just get mapped when a write
 228	 * occurs.  In case of a shrinking truncate, we will detect this later
 229	 * on due to the runlist being incomplete and if the page is being
 230	 * fully truncated, truncate will throw it away as soon as we unlock
 231	 * it so no need to worry what we do with it.
 232	 */
 233	iblock = (s64)page->index << (PAGE_CACHE_SHIFT - blocksize_bits);
 234	read_lock_irqsave(&ni->size_lock, flags);
 235	lblock = (ni->allocated_size + blocksize - 1) >> blocksize_bits;
 236	init_size = ni->initialized_size;
 237	i_size = i_size_read(vi);
 238	read_unlock_irqrestore(&ni->size_lock, flags);
 239	if (unlikely(init_size > i_size)) {
 240		/* Race with shrinking truncate. */
 241		init_size = i_size;
 242	}
 243	zblock = (init_size + blocksize - 1) >> blocksize_bits;
 244
 245	/* Loop through all the buffers in the page. */
 246	rl = NULL;
 247	nr = i = 0;
 248	do {
 249		int err = 0;
 250
 251		if (unlikely(buffer_uptodate(bh)))
 252			continue;
 253		if (unlikely(buffer_mapped(bh))) {
 254			arr[nr++] = bh;
 255			continue;
 256		}
 257		bh->b_bdev = vol->sb->s_bdev;
 258		/* Is the block within the allowed limits? */
 259		if (iblock < lblock) {
 260			bool is_retry = false;
 261
 262			/* Convert iblock into corresponding vcn and offset. */
 263			vcn = (VCN)iblock << blocksize_bits >>
 264					vol->cluster_size_bits;
 265			vcn_ofs = ((VCN)iblock << blocksize_bits) &
 266					vol->cluster_size_mask;
 267			if (!rl) {
 268lock_retry_remap:
 269				down_read(&ni->runlist.lock);
 270				rl = ni->runlist.rl;
 271			}
 272			if (likely(rl != NULL)) {
 273				/* Seek to element containing target vcn. */
 274				while (rl->length && rl[1].vcn <= vcn)
 275					rl++;
 276				lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
 277			} else
 278				lcn = LCN_RL_NOT_MAPPED;
 279			/* Successful remap. */
 280			if (lcn >= 0) {
 281				/* Setup buffer head to correct block. */
 282				bh->b_blocknr = ((lcn << vol->cluster_size_bits)
 283						+ vcn_ofs) >> blocksize_bits;
 284				set_buffer_mapped(bh);
 285				/* Only read initialized data blocks. */
 286				if (iblock < zblock) {
 287					arr[nr++] = bh;
 288					continue;
 289				}
 290				/* Fully non-initialized data block, zero it. */
 291				goto handle_zblock;
 292			}
 293			/* It is a hole, need to zero it. */
 294			if (lcn == LCN_HOLE)
 295				goto handle_hole;
 296			/* If first try and runlist unmapped, map and retry. */
 297			if (!is_retry && lcn == LCN_RL_NOT_MAPPED) {
 298				is_retry = true;
 299				/*
 300				 * Attempt to map runlist, dropping lock for
 301				 * the duration.
 302				 */
 303				up_read(&ni->runlist.lock);
 304				err = ntfs_map_runlist(ni, vcn);
 305				if (likely(!err))
 306					goto lock_retry_remap;
 307				rl = NULL;
 308			} else if (!rl)
 309				up_read(&ni->runlist.lock);
 310			/*
 311			 * If buffer is outside the runlist, treat it as a
 312			 * hole.  This can happen due to concurrent truncate
 313			 * for example.
 314			 */
 315			if (err == -ENOENT || lcn == LCN_ENOENT) {
 316				err = 0;
 317				goto handle_hole;
 318			}
 319			/* Hard error, zero out region. */
 320			if (!err)
 321				err = -EIO;
 322			bh->b_blocknr = -1;
 323			SetPageError(page);
 324			ntfs_error(vol->sb, "Failed to read from inode 0x%lx, "
 325					"attribute type 0x%x, vcn 0x%llx, "
 326					"offset 0x%x because its location on "
 327					"disk could not be determined%s "
 328					"(error code %i).", ni->mft_no,
 329					ni->type, (unsigned long long)vcn,
 330					vcn_ofs, is_retry ? " even after "
 331					"retrying" : "", err);
 332		}
 333		/*
 334		 * Either iblock was outside lblock limits or
 335		 * ntfs_rl_vcn_to_lcn() returned error.  Just zero that portion
 336		 * of the page and set the buffer uptodate.
 337		 */
 338handle_hole:
 339		bh->b_blocknr = -1UL;
 340		clear_buffer_mapped(bh);
 341handle_zblock:
 342		zero_user(page, i * blocksize, blocksize);
 343		if (likely(!err))
 344			set_buffer_uptodate(bh);
 345	} while (i++, iblock++, (bh = bh->b_this_page) != head);
 346
 347	/* Release the lock if we took it. */
 348	if (rl)
 349		up_read(&ni->runlist.lock);
 350
 351	/* Check we have at least one buffer ready for i/o. */
 352	if (nr) {
 353		struct buffer_head *tbh;
 354
 355		/* Lock the buffers. */
 356		for (i = 0; i < nr; i++) {
 357			tbh = arr[i];
 358			lock_buffer(tbh);
 359			tbh->b_end_io = ntfs_end_buffer_async_read;
 360			set_buffer_async_read(tbh);
 361		}
 362		/* Finally, start i/o on the buffers. */
 363		for (i = 0; i < nr; i++) {
 364			tbh = arr[i];
 365			if (likely(!buffer_uptodate(tbh)))
 366				submit_bh(READ, tbh);
 367			else
 368				ntfs_end_buffer_async_read(tbh, 1);
 369		}
 370		return 0;
 371	}
 372	/* No i/o was scheduled on any of the buffers. */
 373	if (likely(!PageError(page)))
 374		SetPageUptodate(page);
 375	else /* Signal synchronous i/o error. */
 376		nr = -EIO;
 377	unlock_page(page);
 378	return nr;
 379}
 380
 381/**
 382 * ntfs_readpage - fill a @page of a @file with data from the device
 383 * @file:	open file to which the page @page belongs or NULL
 384 * @page:	page cache page to fill with data
 385 *
 386 * For non-resident attributes, ntfs_readpage() fills the @page of the open
 387 * file @file by calling the ntfs version of the generic block_read_full_page()
 388 * function, ntfs_read_block(), which in turn creates and reads in the buffers
 389 * associated with the page asynchronously.
 390 *
 391 * For resident attributes, OTOH, ntfs_readpage() fills @page by copying the
 392 * data from the mft record (which at this stage is most likely in memory) and
 393 * fills the remainder with zeroes. Thus, in this case, I/O is synchronous, as
 394 * even if the mft record is not cached at this point in time, we need to wait
 395 * for it to be read in before we can do the copy.
 396 *
 397 * Return 0 on success and -errno on error.
 398 */
 399static int ntfs_readpage(struct file *file, struct page *page)
 400{
 401	loff_t i_size;
 402	struct inode *vi;
 403	ntfs_inode *ni, *base_ni;
 404	u8 *addr;
 405	ntfs_attr_search_ctx *ctx;
 406	MFT_RECORD *mrec;
 407	unsigned long flags;
 408	u32 attr_len;
 409	int err = 0;
 410
 411retry_readpage:
 412	BUG_ON(!PageLocked(page));
 413	vi = page->mapping->host;
 414	i_size = i_size_read(vi);
 415	/* Is the page fully outside i_size? (truncate in progress) */
 416	if (unlikely(page->index >= (i_size + PAGE_CACHE_SIZE - 1) >>
 417			PAGE_CACHE_SHIFT)) {
 418		zero_user(page, 0, PAGE_CACHE_SIZE);
 419		ntfs_debug("Read outside i_size - truncated?");
 420		goto done;
 421	}
 422	/*
 423	 * This can potentially happen because we clear PageUptodate() during
 424	 * ntfs_writepage() of MstProtected() attributes.
 425	 */
 426	if (PageUptodate(page)) {
 427		unlock_page(page);
 428		return 0;
 429	}
 430	ni = NTFS_I(vi);
 431	/*
 432	 * Only $DATA attributes can be encrypted and only unnamed $DATA
 433	 * attributes can be compressed.  Index root can have the flags set but
 434	 * this means to create compressed/encrypted files, not that the
 435	 * attribute is compressed/encrypted.  Note we need to check for
 436	 * AT_INDEX_ALLOCATION since this is the type of both directory and
 437	 * index inodes.
 438	 */
 439	if (ni->type != AT_INDEX_ALLOCATION) {
 440		/* If attribute is encrypted, deny access, just like NT4. */
 441		if (NInoEncrypted(ni)) {
 442			BUG_ON(ni->type != AT_DATA);
 443			err = -EACCES;
 444			goto err_out;
 445		}
 446		/* Compressed data streams are handled in compress.c. */
 447		if (NInoNonResident(ni) && NInoCompressed(ni)) {
 448			BUG_ON(ni->type != AT_DATA);
 449			BUG_ON(ni->name_len);
 450			return ntfs_read_compressed_block(page);
 451		}
 452	}
 453	/* NInoNonResident() == NInoIndexAllocPresent() */
 454	if (NInoNonResident(ni)) {
 455		/* Normal, non-resident data stream. */
 456		return ntfs_read_block(page);
 457	}
 458	/*
 459	 * Attribute is resident, implying it is not compressed or encrypted.
 460	 * This also means the attribute is smaller than an mft record and
 461	 * hence smaller than a page, so can simply zero out any pages with
 462	 * index above 0.  Note the attribute can actually be marked compressed
 463	 * but if it is resident the actual data is not compressed so we are
 464	 * ok to ignore the compressed flag here.
 465	 */
 466	if (unlikely(page->index > 0)) {
 467		zero_user(page, 0, PAGE_CACHE_SIZE);
 468		goto done;
 469	}
 470	if (!NInoAttr(ni))
 471		base_ni = ni;
 472	else
 473		base_ni = ni->ext.base_ntfs_ino;
 474	/* Map, pin, and lock the mft record. */
 475	mrec = map_mft_record(base_ni);
 476	if (IS_ERR(mrec)) {
 477		err = PTR_ERR(mrec);
 478		goto err_out;
 479	}
 480	/*
 481	 * If a parallel write made the attribute non-resident, drop the mft
 482	 * record and retry the readpage.
 483	 */
 484	if (unlikely(NInoNonResident(ni))) {
 485		unmap_mft_record(base_ni);
 486		goto retry_readpage;
 487	}
 488	ctx = ntfs_attr_get_search_ctx(base_ni, mrec);
 489	if (unlikely(!ctx)) {
 490		err = -ENOMEM;
 491		goto unm_err_out;
 492	}
 493	err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
 494			CASE_SENSITIVE, 0, NULL, 0, ctx);
 495	if (unlikely(err))
 496		goto put_unm_err_out;
 497	attr_len = le32_to_cpu(ctx->attr->data.resident.value_length);
 498	read_lock_irqsave(&ni->size_lock, flags);
 499	if (unlikely(attr_len > ni->initialized_size))
 500		attr_len = ni->initialized_size;
 501	i_size = i_size_read(vi);
 502	read_unlock_irqrestore(&ni->size_lock, flags);
 503	if (unlikely(attr_len > i_size)) {
 504		/* Race with shrinking truncate. */
 505		attr_len = i_size;
 506	}
 507	addr = kmap_atomic(page, KM_USER0);
 508	/* Copy the data to the page. */
 509	memcpy(addr, (u8*)ctx->attr +
 510			le16_to_cpu(ctx->attr->data.resident.value_offset),
 511			attr_len);
 512	/* Zero the remainder of the page. */
 513	memset(addr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
 514	flush_dcache_page(page);
 515	kunmap_atomic(addr, KM_USER0);
 516put_unm_err_out:
 517	ntfs_attr_put_search_ctx(ctx);
 518unm_err_out:
 519	unmap_mft_record(base_ni);
 520done:
 521	SetPageUptodate(page);
 522err_out:
 523	unlock_page(page);
 524	return err;
 525}
 526
 527#ifdef NTFS_RW
 528
 529/**
 530 * ntfs_write_block - write a @page to the backing store
 531 * @page:	page cache page to write out
 532 * @wbc:	writeback control structure
 533 *
 534 * This function is for writing pages belonging to non-resident, non-mst
 535 * protected attributes to their backing store.
 536 *
 537 * For a page with buffers, map and write the dirty buffers asynchronously
 538 * under page writeback. For a page without buffers, create buffers for the
 539 * page, then proceed as above.
 540 *
 541 * If a page doesn't have buffers the page dirty state is definitive. If a page
 542 * does have buffers, the page dirty state is just a hint, and the buffer dirty
 543 * state is definitive. (A hint which has rules: dirty buffers against a clean
 544 * page is illegal. Other combinations are legal and need to be handled. In
 545 * particular a dirty page containing clean buffers for example.)
 546 *
 547 * Return 0 on success and -errno on error.
 548 *
 549 * Based on ntfs_read_block() and __block_write_full_page().
 550 */
 551static int ntfs_write_block(struct page *page, struct writeback_control *wbc)
 552{
 553	VCN vcn;
 554	LCN lcn;
 555	s64 initialized_size;
 556	loff_t i_size;
 557	sector_t block, dblock, iblock;
 558	struct inode *vi;
 559	ntfs_inode *ni;
 560	ntfs_volume *vol;
 561	runlist_element *rl;
 562	struct buffer_head *bh, *head;
 563	unsigned long flags;
 564	unsigned int blocksize, vcn_ofs;
 565	int err;
 566	bool need_end_writeback;
 567	unsigned char blocksize_bits;
 568
 569	vi = page->mapping->host;
 570	ni = NTFS_I(vi);
 571	vol = ni->vol;
 572
 573	ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
 574			"0x%lx.", ni->mft_no, ni->type, page->index);
 575
 576	BUG_ON(!NInoNonResident(ni));
 577	BUG_ON(NInoMstProtected(ni));
 578	blocksize = vol->sb->s_blocksize;
 579	blocksize_bits = vol->sb->s_blocksize_bits;
 580	if (!page_has_buffers(page)) {
 581		BUG_ON(!PageUptodate(page));
 582		create_empty_buffers(page, blocksize,
 583				(1 << BH_Uptodate) | (1 << BH_Dirty));
 584		if (unlikely(!page_has_buffers(page))) {
 585			ntfs_warning(vol->sb, "Error allocating page "
 586					"buffers.  Redirtying page so we try "
 587					"again later.");
 588			/*
 589			 * Put the page back on mapping->dirty_pages, but leave
 590			 * its buffers' dirty state as-is.
 591			 */
 592			redirty_page_for_writepage(wbc, page);
 593			unlock_page(page);
 594			return 0;
 595		}
 596	}
 597	bh = head = page_buffers(page);
 598	BUG_ON(!bh);
 599
 600	/* NOTE: Different naming scheme to ntfs_read_block()! */
 601
 602	/* The first block in the page. */
 603	block = (s64)page->index << (PAGE_CACHE_SHIFT - blocksize_bits);
 604
 605	read_lock_irqsave(&ni->size_lock, flags);
 606	i_size = i_size_read(vi);
 607	initialized_size = ni->initialized_size;
 608	read_unlock_irqrestore(&ni->size_lock, flags);
 609
 610	/* The first out of bounds block for the data size. */
 611	dblock = (i_size + blocksize - 1) >> blocksize_bits;
 612
 613	/* The last (fully or partially) initialized block. */
 614	iblock = initialized_size >> blocksize_bits;
 615
 616	/*
 617	 * Be very careful.  We have no exclusion from __set_page_dirty_buffers
 618	 * here, and the (potentially unmapped) buffers may become dirty at
 619	 * any time.  If a buffer becomes dirty here after we've inspected it
 620	 * then we just miss that fact, and the page stays dirty.
 621	 *
 622	 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
 623	 * handle that here by just cleaning them.
 624	 */
 625
 626	/*
 627	 * Loop through all the buffers in the page, mapping all the dirty
 628	 * buffers to disk addresses and handling any aliases from the
 629	 * underlying block device's mapping.
 630	 */
 631	rl = NULL;
 632	err = 0;
 633	do {
 634		bool is_retry = false;
 635
 636		if (unlikely(block >= dblock)) {
 637			/*
 638			 * Mapped buffers outside i_size will occur, because
 639			 * this page can be outside i_size when there is a
 640			 * truncate in progress. The contents of such buffers
 641			 * were zeroed by ntfs_writepage().
 642			 *
 643			 * FIXME: What about the small race window where
 644			 * ntfs_writepage() has not done any clearing because
 645			 * the page was within i_size but before we get here,
 646			 * vmtruncate() modifies i_size?
 647			 */
 648			clear_buffer_dirty(bh);
 649			set_buffer_uptodate(bh);
 650			continue;
 651		}
 652
 653		/* Clean buffers are not written out, so no need to map them. */
 654		if (!buffer_dirty(bh))
 655			continue;
 656
 657		/* Make sure we have enough initialized size. */
 658		if (unlikely((block >= iblock) &&
 659				(initialized_size < i_size))) {
 660			/*
 661			 * If this page is fully outside initialized size, zero
 662			 * out all pages between the current initialized size
 663			 * and the current page. Just use ntfs_readpage() to do
 664			 * the zeroing transparently.
 665			 */
 666			if (block > iblock) {
 667				// TODO:
 668				// For each page do:
 669				// - read_cache_page()
 670				// Again for each page do:
 671				// - wait_on_page_locked()
 672				// - Check (PageUptodate(page) &&
 673				//			!PageError(page))
 674				// Update initialized size in the attribute and
 675				// in the inode.
 676				// Again, for each page do:
 677				//	__set_page_dirty_buffers();
 678				// page_cache_release()
 679				// We don't need to wait on the writes.
 680				// Update iblock.
 681			}
 682			/*
 683			 * The current page straddles initialized size. Zero
 684			 * all non-uptodate buffers and set them uptodate (and
 685			 * dirty?). Note, there aren't any non-uptodate buffers
 686			 * if the page is uptodate.
 687			 * FIXME: For an uptodate page, the buffers may need to
 688			 * be written out because they were not initialized on
 689			 * disk before.
 690			 */
 691			if (!PageUptodate(page)) {
 692				// TODO:
 693				// Zero any non-uptodate buffers up to i_size.
 694				// Set them uptodate and dirty.
 695			}
 696			// TODO:
 697			// Update initialized size in the attribute and in the
 698			// inode (up to i_size).
 699			// Update iblock.
 700			// FIXME: This is inefficient. Try to batch the two
 701			// size changes to happen in one go.
 702			ntfs_error(vol->sb, "Writing beyond initialized size "
 703					"is not supported yet. Sorry.");
 704			err = -EOPNOTSUPP;
 705			break;
 706			// Do NOT set_buffer_new() BUT DO clear buffer range
 707			// outside write request range.
 708			// set_buffer_uptodate() on complete buffers as well as
 709			// set_buffer_dirty().
 710		}
 711
 712		/* No need to map buffers that are already mapped. */
 713		if (buffer_mapped(bh))
 714			continue;
 715
 716		/* Unmapped, dirty buffer. Need to map it. */
 717		bh->b_bdev = vol->sb->s_bdev;
 718
 719		/* Convert block into corresponding vcn and offset. */
 720		vcn = (VCN)block << blocksize_bits;
 721		vcn_ofs = vcn & vol->cluster_size_mask;
 722		vcn >>= vol->cluster_size_bits;
 723		if (!rl) {
 724lock_retry_remap:
 725			down_read(&ni->runlist.lock);
 726			rl = ni->runlist.rl;
 727		}
 728		if (likely(rl != NULL)) {
 729			/* Seek to element containing target vcn. */
 730			while (rl->length && rl[1].vcn <= vcn)
 731				rl++;
 732			lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
 733		} else
 734			lcn = LCN_RL_NOT_MAPPED;
 735		/* Successful remap. */
 736		if (lcn >= 0) {
 737			/* Setup buffer head to point to correct block. */
 738			bh->b_blocknr = ((lcn << vol->cluster_size_bits) +
 739					vcn_ofs) >> blocksize_bits;
 740			set_buffer_mapped(bh);
 741			continue;
 742		}
 743		/* It is a hole, need to instantiate it. */
 744		if (lcn == LCN_HOLE) {
 745			u8 *kaddr;
 746			unsigned long *bpos, *bend;
 747
 748			/* Check if the buffer is zero. */
 749			kaddr = kmap_atomic(page, KM_USER0);
 750			bpos = (unsigned long *)(kaddr + bh_offset(bh));
 751			bend = (unsigned long *)((u8*)bpos + blocksize);
 752			do {
 753				if (unlikely(*bpos))
 754					break;
 755			} while (likely(++bpos < bend));
 756			kunmap_atomic(kaddr, KM_USER0);
 757			if (bpos == bend) {
 758				/*
 759				 * Buffer is zero and sparse, no need to write
 760				 * it.
 761				 */
 762				bh->b_blocknr = -1;
 763				clear_buffer_dirty(bh);
 764				continue;
 765			}
 766			// TODO: Instantiate the hole.
 767			// clear_buffer_new(bh);
 768			// unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
 769			ntfs_error(vol->sb, "Writing into sparse regions is "
 770					"not supported yet. Sorry.");
 771			err = -EOPNOTSUPP;
 772			break;
 773		}
 774		/* If first try and runlist unmapped, map and retry. */
 775		if (!is_retry && lcn == LCN_RL_NOT_MAPPED) {
 776			is_retry = true;
 777			/*
 778			 * Attempt to map runlist, dropping lock for
 779			 * the duration.
 780			 */
 781			up_read(&ni->runlist.lock);
 782			err = ntfs_map_runlist(ni, vcn);
 783			if (likely(!err))
 784				goto lock_retry_remap;
 785			rl = NULL;
 786		} else if (!rl)
 787			up_read(&ni->runlist.lock);
 788		/*
 789		 * If buffer is outside the runlist, truncate has cut it out
 790		 * of the runlist.  Just clean and clear the buffer and set it
 791		 * uptodate so it can get discarded by the VM.
 792		 */
 793		if (err == -ENOENT || lcn == LCN_ENOENT) {
 794			bh->b_blocknr = -1;
 795			clear_buffer_dirty(bh);
 796			zero_user(page, bh_offset(bh), blocksize);
 797			set_buffer_uptodate(bh);
 798			err = 0;
 799			continue;
 800		}
 801		/* Failed to map the buffer, even after retrying. */
 802		if (!err)
 803			err = -EIO;
 804		bh->b_blocknr = -1;
 805		ntfs_error(vol->sb, "Failed to write to inode 0x%lx, "
 806				"attribute type 0x%x, vcn 0x%llx, offset 0x%x "
 807				"because its location on disk could not be "
 808				"determined%s (error code %i).", ni->mft_no,
 809				ni->type, (unsigned long long)vcn,
 810				vcn_ofs, is_retry ? " even after "
 811				"retrying" : "", err);
 812		break;
 813	} while (block++, (bh = bh->b_this_page) != head);
 814
 815	/* Release the lock if we took it. */
 816	if (rl)
 817		up_read(&ni->runlist.lock);
 818
 819	/* For the error case, need to reset bh to the beginning. */
 820	bh = head;
 821
 822	/* Just an optimization, so ->readpage() is not called later. */
 823	if (unlikely(!PageUptodate(page))) {
 824		int uptodate = 1;
 825		do {
 826			if (!buffer_uptodate(bh)) {
 827				uptodate = 0;
 828				bh = head;
 829				break;
 830			}
 831		} while ((bh = bh->b_this_page) != head);
 832		if (uptodate)
 833			SetPageUptodate(page);
 834	}
 835
 836	/* Setup all mapped, dirty buffers for async write i/o. */
 837	do {
 838		if (buffer_mapped(bh) && buffer_dirty(bh)) {
 839			lock_buffer(bh);
 840			if (test_clear_buffer_dirty(bh)) {
 841				BUG_ON(!buffer_uptodate(bh));
 842				mark_buffer_async_write(bh);
 843			} else
 844				unlock_buffer(bh);
 845		} else if (unlikely(err)) {
 846			/*
 847			 * For the error case. The buffer may have been set
 848			 * dirty during attachment to a dirty page.
 849			 */
 850			if (err != -ENOMEM)
 851				clear_buffer_dirty(bh);
 852		}
 853	} while ((bh = bh->b_this_page) != head);
 854
 855	if (unlikely(err)) {
 856		// TODO: Remove the -EOPNOTSUPP check later on...
 857		if (unlikely(err == -EOPNOTSUPP))
 858			err = 0;
 859		else if (err == -ENOMEM) {
 860			ntfs_warning(vol->sb, "Error allocating memory. "
 861					"Redirtying page so we try again "
 862					"later.");
 863			/*
 864			 * Put the page back on mapping->dirty_pages, but
 865			 * leave its buffer's dirty state as-is.
 866			 */
 867			redirty_page_for_writepage(wbc, page);
 868			err = 0;
 869		} else
 870			SetPageError(page);
 871	}
 872
 873	BUG_ON(PageWriteback(page));
 874	set_page_writeback(page);	/* Keeps try_to_free_buffers() away. */
 875
 876	/* Submit the prepared buffers for i/o. */
 877	need_end_writeback = true;
 878	do {
 879		struct buffer_head *next = bh->b_this_page;
 880		if (buffer_async_write(bh)) {
 881			submit_bh(WRITE, bh);
 882			need_end_writeback = false;
 883		}
 884		bh = next;
 885	} while (bh != head);
 886	unlock_page(page);
 887
 888	/* If no i/o was started, need to end_page_writeback(). */
 889	if (unlikely(need_end_writeback))
 890		end_page_writeback(page);
 891
 892	ntfs_debug("Done.");
 893	return err;
 894}
 895
 896/**
 897 * ntfs_write_mst_block - write a @page to the backing store
 898 * @page:	page cache page to write out
 899 * @wbc:	writeback control structure
 900 *
 901 * This function is for writing pages belonging to non-resident, mst protected
 902 * attributes to their backing store.  The only supported attributes are index
 903 * allocation and $MFT/$DATA.  Both directory inodes and index inodes are
 904 * supported for the index allocation case.
 905 *
 906 * The page must remain locked for the duration of the write because we apply
 907 * the mst fixups, write, and then undo the fixups, so if we were to unlock the
 908 * page before undoing the fixups, any other user of the page will see the
 909 * page contents as corrupt.
 910 *
 911 * We clear the page uptodate flag for the duration of the function to ensure
 912 * exclusion for the $MFT/$DATA case against someone mapping an mft record we
 913 * are about to apply the mst fixups to.
 914 *
 915 * Return 0 on success and -errno on error.
 916 *
 917 * Based on ntfs_write_block(), ntfs_mft_writepage(), and
 918 * write_mft_record_nolock().
 919 */
 920static int ntfs_write_mst_block(struct page *page,
 921		struct writeback_control *wbc)
 922{
 923	sector_t block, dblock, rec_block;
 924	struct inode *vi = page->mapping->host;
 925	ntfs_inode *ni = NTFS_I(vi);
 926	ntfs_volume *vol = ni->vol;
 927	u8 *kaddr;
 928	unsigned int rec_size = ni->itype.index.block_size;
 929	ntfs_inode *locked_nis[PAGE_CACHE_SIZE / rec_size];
 930	struct buffer_head *bh, *head, *tbh, *rec_start_bh;
 931	struct buffer_head *bhs[MAX_BUF_PER_PAGE];
 932	runlist_element *rl;
 933	int i, nr_locked_nis, nr_recs, nr_bhs, max_bhs, bhs_per_rec, err, err2;
 934	unsigned bh_size, rec_size_bits;
 935	bool sync, is_mft, page_is_dirty, rec_is_dirty;
 936	unsigned char bh_size_bits;
 937
 
 
 
 938	ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
 939			"0x%lx.", vi->i_ino, ni->type, page->index);
 940	BUG_ON(!NInoNonResident(ni));
 941	BUG_ON(!NInoMstProtected(ni));
 942	is_mft = (S_ISREG(vi->i_mode) && !vi->i_ino);
 943	/*
 944	 * NOTE: ntfs_write_mst_block() would be called for $MFTMirr if a page
 945	 * in its page cache were to be marked dirty.  However this should
 946	 * never happen with the current driver and considering we do not
 947	 * handle this case here we do want to BUG(), at least for now.
 948	 */
 949	BUG_ON(!(is_mft || S_ISDIR(vi->i_mode) ||
 950			(NInoAttr(ni) && ni->type == AT_INDEX_ALLOCATION)));
 951	bh_size = vol->sb->s_blocksize;
 952	bh_size_bits = vol->sb->s_blocksize_bits;
 953	max_bhs = PAGE_CACHE_SIZE / bh_size;
 954	BUG_ON(!max_bhs);
 955	BUG_ON(max_bhs > MAX_BUF_PER_PAGE);
 956
 957	/* Were we called for sync purposes? */
 958	sync = (wbc->sync_mode == WB_SYNC_ALL);
 959
 960	/* Make sure we have mapped buffers. */
 961	bh = head = page_buffers(page);
 962	BUG_ON(!bh);
 963
 964	rec_size_bits = ni->itype.index.block_size_bits;
 965	BUG_ON(!(PAGE_CACHE_SIZE >> rec_size_bits));
 966	bhs_per_rec = rec_size >> bh_size_bits;
 967	BUG_ON(!bhs_per_rec);
 968
 969	/* The first block in the page. */
 970	rec_block = block = (sector_t)page->index <<
 971			(PAGE_CACHE_SHIFT - bh_size_bits);
 972
 973	/* The first out of bounds block for the data size. */
 974	dblock = (i_size_read(vi) + bh_size - 1) >> bh_size_bits;
 975
 976	rl = NULL;
 977	err = err2 = nr_bhs = nr_recs = nr_locked_nis = 0;
 978	page_is_dirty = rec_is_dirty = false;
 979	rec_start_bh = NULL;
 980	do {
 981		bool is_retry = false;
 982
 983		if (likely(block < rec_block)) {
 984			if (unlikely(block >= dblock)) {
 985				clear_buffer_dirty(bh);
 986				set_buffer_uptodate(bh);
 987				continue;
 988			}
 989			/*
 990			 * This block is not the first one in the record.  We
 991			 * ignore the buffer's dirty state because we could
 992			 * have raced with a parallel mark_ntfs_record_dirty().
 993			 */
 994			if (!rec_is_dirty)
 995				continue;
 996			if (unlikely(err2)) {
 997				if (err2 != -ENOMEM)
 998					clear_buffer_dirty(bh);
 999				continue;
1000			}
1001		} else /* if (block == rec_block) */ {
1002			BUG_ON(block > rec_block);
1003			/* This block is the first one in the record. */
1004			rec_block += bhs_per_rec;
1005			err2 = 0;
1006			if (unlikely(block >= dblock)) {
1007				clear_buffer_dirty(bh);
1008				continue;
1009			}
1010			if (!buffer_dirty(bh)) {
1011				/* Clean records are not written out. */
1012				rec_is_dirty = false;
1013				continue;
1014			}
1015			rec_is_dirty = true;
1016			rec_start_bh = bh;
1017		}
1018		/* Need to map the buffer if it is not mapped already. */
1019		if (unlikely(!buffer_mapped(bh))) {
1020			VCN vcn;
1021			LCN lcn;
1022			unsigned int vcn_ofs;
1023
1024			bh->b_bdev = vol->sb->s_bdev;
1025			/* Obtain the vcn and offset of the current block. */
1026			vcn = (VCN)block << bh_size_bits;
1027			vcn_ofs = vcn & vol->cluster_size_mask;
1028			vcn >>= vol->cluster_size_bits;
1029			if (!rl) {
1030lock_retry_remap:
1031				down_read(&ni->runlist.lock);
1032				rl = ni->runlist.rl;
1033			}
1034			if (likely(rl != NULL)) {
1035				/* Seek to element containing target vcn. */
1036				while (rl->length && rl[1].vcn <= vcn)
1037					rl++;
1038				lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
1039			} else
1040				lcn = LCN_RL_NOT_MAPPED;
1041			/* Successful remap. */
1042			if (likely(lcn >= 0)) {
1043				/* Setup buffer head to correct block. */
1044				bh->b_blocknr = ((lcn <<
1045						vol->cluster_size_bits) +
1046						vcn_ofs) >> bh_size_bits;
1047				set_buffer_mapped(bh);
1048			} else {
1049				/*
1050				 * Remap failed.  Retry to map the runlist once
1051				 * unless we are working on $MFT which always
1052				 * has the whole of its runlist in memory.
1053				 */
1054				if (!is_mft && !is_retry &&
1055						lcn == LCN_RL_NOT_MAPPED) {
1056					is_retry = true;
1057					/*
1058					 * Attempt to map runlist, dropping
1059					 * lock for the duration.
1060					 */
1061					up_read(&ni->runlist.lock);
1062					err2 = ntfs_map_runlist(ni, vcn);
1063					if (likely(!err2))
1064						goto lock_retry_remap;
1065					if (err2 == -ENOMEM)
1066						page_is_dirty = true;
1067					lcn = err2;
1068				} else {
1069					err2 = -EIO;
1070					if (!rl)
1071						up_read(&ni->runlist.lock);
1072				}
1073				/* Hard error.  Abort writing this record. */
1074				if (!err || err == -ENOMEM)
1075					err = err2;
1076				bh->b_blocknr = -1;
1077				ntfs_error(vol->sb, "Cannot write ntfs record "
1078						"0x%llx (inode 0x%lx, "
1079						"attribute type 0x%x) because "
1080						"its location on disk could "
1081						"not be determined (error "
1082						"code %lli).",
1083						(long long)block <<
1084						bh_size_bits >>
1085						vol->mft_record_size_bits,
1086						ni->mft_no, ni->type,
1087						(long long)lcn);
1088				/*
1089				 * If this is not the first buffer, remove the
1090				 * buffers in this record from the list of
1091				 * buffers to write and clear their dirty bit
1092				 * if not error -ENOMEM.
1093				 */
1094				if (rec_start_bh != bh) {
1095					while (bhs[--nr_bhs] != rec_start_bh)
1096						;
1097					if (err2 != -ENOMEM) {
1098						do {
1099							clear_buffer_dirty(
1100								rec_start_bh);
1101						} while ((rec_start_bh =
1102								rec_start_bh->
1103								b_this_page) !=
1104								bh);
1105					}
1106				}
1107				continue;
1108			}
1109		}
1110		BUG_ON(!buffer_uptodate(bh));
1111		BUG_ON(nr_bhs >= max_bhs);
1112		bhs[nr_bhs++] = bh;
1113	} while (block++, (bh = bh->b_this_page) != head);
1114	if (unlikely(rl))
1115		up_read(&ni->runlist.lock);
1116	/* If there were no dirty buffers, we are done. */
1117	if (!nr_bhs)
1118		goto done;
1119	/* Map the page so we can access its contents. */
1120	kaddr = kmap(page);
1121	/* Clear the page uptodate flag whilst the mst fixups are applied. */
1122	BUG_ON(!PageUptodate(page));
1123	ClearPageUptodate(page);
1124	for (i = 0; i < nr_bhs; i++) {
1125		unsigned int ofs;
1126
1127		/* Skip buffers which are not at the beginning of records. */
1128		if (i % bhs_per_rec)
1129			continue;
1130		tbh = bhs[i];
1131		ofs = bh_offset(tbh);
1132		if (is_mft) {
1133			ntfs_inode *tni;
1134			unsigned long mft_no;
1135
1136			/* Get the mft record number. */
1137			mft_no = (((s64)page->index << PAGE_CACHE_SHIFT) + ofs)
1138					>> rec_size_bits;
1139			/* Check whether to write this mft record. */
1140			tni = NULL;
1141			if (!ntfs_may_write_mft_record(vol, mft_no,
1142					(MFT_RECORD*)(kaddr + ofs), &tni)) {
1143				/*
1144				 * The record should not be written.  This
1145				 * means we need to redirty the page before
1146				 * returning.
1147				 */
1148				page_is_dirty = true;
1149				/*
1150				 * Remove the buffers in this mft record from
1151				 * the list of buffers to write.
1152				 */
1153				do {
1154					bhs[i] = NULL;
1155				} while (++i % bhs_per_rec);
1156				continue;
1157			}
1158			/*
1159			 * The record should be written.  If a locked ntfs
1160			 * inode was returned, add it to the array of locked
1161			 * ntfs inodes.
1162			 */
1163			if (tni)
1164				locked_nis[nr_locked_nis++] = tni;
1165		}
1166		/* Apply the mst protection fixups. */
1167		err2 = pre_write_mst_fixup((NTFS_RECORD*)(kaddr + ofs),
1168				rec_size);
1169		if (unlikely(err2)) {
1170			if (!err || err == -ENOMEM)
1171				err = -EIO;
1172			ntfs_error(vol->sb, "Failed to apply mst fixups "
1173					"(inode 0x%lx, attribute type 0x%x, "
1174					"page index 0x%lx, page offset 0x%x)!"
1175					"  Unmount and run chkdsk.", vi->i_ino,
1176					ni->type, page->index, ofs);
1177			/*
1178			 * Mark all the buffers in this record clean as we do
1179			 * not want to write corrupt data to disk.
1180			 */
1181			do {
1182				clear_buffer_dirty(bhs[i]);
1183				bhs[i] = NULL;
1184			} while (++i % bhs_per_rec);
1185			continue;
1186		}
1187		nr_recs++;
1188	}
1189	/* If no records are to be written out, we are done. */
1190	if (!nr_recs)
1191		goto unm_done;
1192	flush_dcache_page(page);
1193	/* Lock buffers and start synchronous write i/o on them. */
1194	for (i = 0; i < nr_bhs; i++) {
1195		tbh = bhs[i];
1196		if (!tbh)
1197			continue;
1198		if (!trylock_buffer(tbh))
1199			BUG();
1200		/* The buffer dirty state is now irrelevant, just clean it. */
1201		clear_buffer_dirty(tbh);
1202		BUG_ON(!buffer_uptodate(tbh));
1203		BUG_ON(!buffer_mapped(tbh));
1204		get_bh(tbh);
1205		tbh->b_end_io = end_buffer_write_sync;
1206		submit_bh(WRITE, tbh);
1207	}
1208	/* Synchronize the mft mirror now if not @sync. */
1209	if (is_mft && !sync)
1210		goto do_mirror;
1211do_wait:
1212	/* Wait on i/o completion of buffers. */
1213	for (i = 0; i < nr_bhs; i++) {
1214		tbh = bhs[i];
1215		if (!tbh)
1216			continue;
1217		wait_on_buffer(tbh);
1218		if (unlikely(!buffer_uptodate(tbh))) {
1219			ntfs_error(vol->sb, "I/O error while writing ntfs "
1220					"record buffer (inode 0x%lx, "
1221					"attribute type 0x%x, page index "
1222					"0x%lx, page offset 0x%lx)!  Unmount "
1223					"and run chkdsk.", vi->i_ino, ni->type,
1224					page->index, bh_offset(tbh));
1225			if (!err || err == -ENOMEM)
1226				err = -EIO;
1227			/*
1228			 * Set the buffer uptodate so the page and buffer
1229			 * states do not become out of sync.
1230			 */
1231			set_buffer_uptodate(tbh);
1232		}
1233	}
1234	/* If @sync, now synchronize the mft mirror. */
1235	if (is_mft && sync) {
1236do_mirror:
1237		for (i = 0; i < nr_bhs; i++) {
1238			unsigned long mft_no;
1239			unsigned int ofs;
1240
1241			/*
1242			 * Skip buffers which are not at the beginning of
1243			 * records.
1244			 */
1245			if (i % bhs_per_rec)
1246				continue;
1247			tbh = bhs[i];
1248			/* Skip removed buffers (and hence records). */
1249			if (!tbh)
1250				continue;
1251			ofs = bh_offset(tbh);
1252			/* Get the mft record number. */
1253			mft_no = (((s64)page->index << PAGE_CACHE_SHIFT) + ofs)
1254					>> rec_size_bits;
1255			if (mft_no < vol->mftmirr_size)
1256				ntfs_sync_mft_mirror(vol, mft_no,
1257						(MFT_RECORD*)(kaddr + ofs),
1258						sync);
1259		}
1260		if (!sync)
1261			goto do_wait;
1262	}
1263	/* Remove the mst protection fixups again. */
1264	for (i = 0; i < nr_bhs; i++) {
1265		if (!(i % bhs_per_rec)) {
1266			tbh = bhs[i];
1267			if (!tbh)
1268				continue;
1269			post_write_mst_fixup((NTFS_RECORD*)(kaddr +
1270					bh_offset(tbh)));
1271		}
1272	}
1273	flush_dcache_page(page);
1274unm_done:
1275	/* Unlock any locked inodes. */
1276	while (nr_locked_nis-- > 0) {
1277		ntfs_inode *tni, *base_tni;
1278		
1279		tni = locked_nis[nr_locked_nis];
1280		/* Get the base inode. */
1281		mutex_lock(&tni->extent_lock);
1282		if (tni->nr_extents >= 0)
1283			base_tni = tni;
1284		else {
1285			base_tni = tni->ext.base_ntfs_ino;
1286			BUG_ON(!base_tni);
1287		}
1288		mutex_unlock(&tni->extent_lock);
1289		ntfs_debug("Unlocking %s inode 0x%lx.",
1290				tni == base_tni ? "base" : "extent",
1291				tni->mft_no);
1292		mutex_unlock(&tni->mrec_lock);
1293		atomic_dec(&tni->count);
1294		iput(VFS_I(base_tni));
1295	}
1296	SetPageUptodate(page);
1297	kunmap(page);
1298done:
1299	if (unlikely(err && err != -ENOMEM)) {
1300		/*
1301		 * Set page error if there is only one ntfs record in the page.
1302		 * Otherwise we would loose per-record granularity.
1303		 */
1304		if (ni->itype.index.block_size == PAGE_CACHE_SIZE)
1305			SetPageError(page);
1306		NVolSetErrors(vol);
1307	}
1308	if (page_is_dirty) {
1309		ntfs_debug("Page still contains one or more dirty ntfs "
1310				"records.  Redirtying the page starting at "
1311				"record 0x%lx.", page->index <<
1312				(PAGE_CACHE_SHIFT - rec_size_bits));
1313		redirty_page_for_writepage(wbc, page);
1314		unlock_page(page);
1315	} else {
1316		/*
1317		 * Keep the VM happy.  This must be done otherwise the
1318		 * radix-tree tag PAGECACHE_TAG_DIRTY remains set even though
1319		 * the page is clean.
1320		 */
1321		BUG_ON(PageWriteback(page));
1322		set_page_writeback(page);
1323		unlock_page(page);
1324		end_page_writeback(page);
1325	}
1326	if (likely(!err))
1327		ntfs_debug("Done.");
1328	return err;
1329}
1330
1331/**
1332 * ntfs_writepage - write a @page to the backing store
1333 * @page:	page cache page to write out
1334 * @wbc:	writeback control structure
1335 *
1336 * This is called from the VM when it wants to have a dirty ntfs page cache
1337 * page cleaned.  The VM has already locked the page and marked it clean.
1338 *
1339 * For non-resident attributes, ntfs_writepage() writes the @page by calling
1340 * the ntfs version of the generic block_write_full_page() function,
1341 * ntfs_write_block(), which in turn if necessary creates and writes the
1342 * buffers associated with the page asynchronously.
1343 *
1344 * For resident attributes, OTOH, ntfs_writepage() writes the @page by copying
1345 * the data to the mft record (which at this stage is most likely in memory).
1346 * The mft record is then marked dirty and written out asynchronously via the
1347 * vfs inode dirty code path for the inode the mft record belongs to or via the
1348 * vm page dirty code path for the page the mft record is in.
1349 *
1350 * Based on ntfs_readpage() and fs/buffer.c::block_write_full_page().
1351 *
1352 * Return 0 on success and -errno on error.
1353 */
1354static int ntfs_writepage(struct page *page, struct writeback_control *wbc)
1355{
1356	loff_t i_size;
1357	struct inode *vi = page->mapping->host;
1358	ntfs_inode *base_ni = NULL, *ni = NTFS_I(vi);
1359	char *addr;
1360	ntfs_attr_search_ctx *ctx = NULL;
1361	MFT_RECORD *m = NULL;
1362	u32 attr_len;
1363	int err;
1364
1365retry_writepage:
1366	BUG_ON(!PageLocked(page));
1367	i_size = i_size_read(vi);
1368	/* Is the page fully outside i_size? (truncate in progress) */
1369	if (unlikely(page->index >= (i_size + PAGE_CACHE_SIZE - 1) >>
1370			PAGE_CACHE_SHIFT)) {
1371		/*
1372		 * The page may have dirty, unmapped buffers.  Make them
1373		 * freeable here, so the page does not leak.
1374		 */
1375		block_invalidatepage(page, 0);
1376		unlock_page(page);
1377		ntfs_debug("Write outside i_size - truncated?");
1378		return 0;
1379	}
1380	/*
1381	 * Only $DATA attributes can be encrypted and only unnamed $DATA
1382	 * attributes can be compressed.  Index root can have the flags set but
1383	 * this means to create compressed/encrypted files, not that the
1384	 * attribute is compressed/encrypted.  Note we need to check for
1385	 * AT_INDEX_ALLOCATION since this is the type of both directory and
1386	 * index inodes.
1387	 */
1388	if (ni->type != AT_INDEX_ALLOCATION) {
1389		/* If file is encrypted, deny access, just like NT4. */
1390		if (NInoEncrypted(ni)) {
1391			unlock_page(page);
1392			BUG_ON(ni->type != AT_DATA);
1393			ntfs_debug("Denying write access to encrypted file.");
1394			return -EACCES;
1395		}
1396		/* Compressed data streams are handled in compress.c. */
1397		if (NInoNonResident(ni) && NInoCompressed(ni)) {
1398			BUG_ON(ni->type != AT_DATA);
1399			BUG_ON(ni->name_len);
1400			// TODO: Implement and replace this with
1401			// return ntfs_write_compressed_block(page);
1402			unlock_page(page);
1403			ntfs_error(vi->i_sb, "Writing to compressed files is "
1404					"not supported yet.  Sorry.");
1405			return -EOPNOTSUPP;
1406		}
1407		// TODO: Implement and remove this check.
1408		if (NInoNonResident(ni) && NInoSparse(ni)) {
1409			unlock_page(page);
1410			ntfs_error(vi->i_sb, "Writing to sparse files is not "
1411					"supported yet.  Sorry.");
1412			return -EOPNOTSUPP;
1413		}
1414	}
1415	/* NInoNonResident() == NInoIndexAllocPresent() */
1416	if (NInoNonResident(ni)) {
1417		/* We have to zero every time due to mmap-at-end-of-file. */
1418		if (page->index >= (i_size >> PAGE_CACHE_SHIFT)) {
1419			/* The page straddles i_size. */
1420			unsigned int ofs = i_size & ~PAGE_CACHE_MASK;
1421			zero_user_segment(page, ofs, PAGE_CACHE_SIZE);
1422		}
1423		/* Handle mst protected attributes. */
1424		if (NInoMstProtected(ni))
1425			return ntfs_write_mst_block(page, wbc);
1426		/* Normal, non-resident data stream. */
1427		return ntfs_write_block(page, wbc);
1428	}
1429	/*
1430	 * Attribute is resident, implying it is not compressed, encrypted, or
1431	 * mst protected.  This also means the attribute is smaller than an mft
1432	 * record and hence smaller than a page, so can simply return error on
1433	 * any pages with index above 0.  Note the attribute can actually be
1434	 * marked compressed but if it is resident the actual data is not
1435	 * compressed so we are ok to ignore the compressed flag here.
1436	 */
1437	BUG_ON(page_has_buffers(page));
1438	BUG_ON(!PageUptodate(page));
1439	if (unlikely(page->index > 0)) {
1440		ntfs_error(vi->i_sb, "BUG()! page->index (0x%lx) > 0.  "
1441				"Aborting write.", page->index);
1442		BUG_ON(PageWriteback(page));
1443		set_page_writeback(page);
1444		unlock_page(page);
1445		end_page_writeback(page);
1446		return -EIO;
1447	}
1448	if (!NInoAttr(ni))
1449		base_ni = ni;
1450	else
1451		base_ni = ni->ext.base_ntfs_ino;
1452	/* Map, pin, and lock the mft record. */
1453	m = map_mft_record(base_ni);
1454	if (IS_ERR(m)) {
1455		err = PTR_ERR(m);
1456		m = NULL;
1457		ctx = NULL;
1458		goto err_out;
1459	}
1460	/*
1461	 * If a parallel write made the attribute non-resident, drop the mft
1462	 * record and retry the writepage.
1463	 */
1464	if (unlikely(NInoNonResident(ni))) {
1465		unmap_mft_record(base_ni);
1466		goto retry_writepage;
1467	}
1468	ctx = ntfs_attr_get_search_ctx(base_ni, m);
1469	if (unlikely(!ctx)) {
1470		err = -ENOMEM;
1471		goto err_out;
1472	}
1473	err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1474			CASE_SENSITIVE, 0, NULL, 0, ctx);
1475	if (unlikely(err))
1476		goto err_out;
1477	/*
1478	 * Keep the VM happy.  This must be done otherwise the radix-tree tag
1479	 * PAGECACHE_TAG_DIRTY remains set even though the page is clean.
1480	 */
1481	BUG_ON(PageWriteback(page));
1482	set_page_writeback(page);
1483	unlock_page(page);
1484	attr_len = le32_to_cpu(ctx->attr->data.resident.value_length);
1485	i_size = i_size_read(vi);
1486	if (unlikely(attr_len > i_size)) {
1487		/* Race with shrinking truncate or a failed truncate. */
1488		attr_len = i_size;
1489		/*
1490		 * If the truncate failed, fix it up now.  If a concurrent
1491		 * truncate, we do its job, so it does not have to do anything.
1492		 */
1493		err = ntfs_resident_attr_value_resize(ctx->mrec, ctx->attr,
1494				attr_len);
1495		/* Shrinking cannot fail. */
1496		BUG_ON(err);
1497	}
1498	addr = kmap_atomic(page, KM_USER0);
1499	/* Copy the data from the page to the mft record. */
1500	memcpy((u8*)ctx->attr +
1501			le16_to_cpu(ctx->attr->data.resident.value_offset),
1502			addr, attr_len);
1503	/* Zero out of bounds area in the page cache page. */
1504	memset(addr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
1505	kunmap_atomic(addr, KM_USER0);
1506	flush_dcache_page(page);
1507	flush_dcache_mft_record_page(ctx->ntfs_ino);
1508	/* We are done with the page. */
1509	end_page_writeback(page);
1510	/* Finally, mark the mft record dirty, so it gets written back. */
1511	mark_mft_record_dirty(ctx->ntfs_ino);
1512	ntfs_attr_put_search_ctx(ctx);
1513	unmap_mft_record(base_ni);
1514	return 0;
1515err_out:
1516	if (err == -ENOMEM) {
1517		ntfs_warning(vi->i_sb, "Error allocating memory. Redirtying "
1518				"page so we try again later.");
1519		/*
1520		 * Put the page back on mapping->dirty_pages, but leave its
1521		 * buffers' dirty state as-is.
1522		 */
1523		redirty_page_for_writepage(wbc, page);
1524		err = 0;
1525	} else {
1526		ntfs_error(vi->i_sb, "Resident attribute write failed with "
1527				"error %i.", err);
1528		SetPageError(page);
1529		NVolSetErrors(ni->vol);
1530	}
1531	unlock_page(page);
1532	if (ctx)
1533		ntfs_attr_put_search_ctx(ctx);
1534	if (m)
1535		unmap_mft_record(base_ni);
1536	return err;
1537}
1538
1539#endif	/* NTFS_RW */
1540
1541/**
1542 * ntfs_aops - general address space operations for inodes and attributes
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1543 */
1544const struct address_space_operations ntfs_aops = {
1545	.readpage	= ntfs_readpage,	/* Fill page with data. */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1546#ifdef NTFS_RW
1547	.writepage	= ntfs_writepage,	/* Write dirty page to disk. */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1548#endif /* NTFS_RW */
1549	.migratepage	= buffer_migrate_page,	/* Move a page cache page from
1550						   one physical page to an
1551						   other. */
1552	.error_remove_page = generic_error_remove_page,
1553};
1554
1555/**
1556 * ntfs_mst_aops - general address space operations for mst protecteed inodes
1557 *		   and attributes
1558 */
1559const struct address_space_operations ntfs_mst_aops = {
1560	.readpage	= ntfs_readpage,	/* Fill page with data. */
1561#ifdef NTFS_RW
1562	.writepage	= ntfs_writepage,	/* Write dirty page to disk. */
1563	.set_page_dirty	= __set_page_dirty_nobuffers,	/* Set the page dirty
1564						   without touching the buffers
1565						   belonging to the page. */
1566#endif /* NTFS_RW */
1567	.migratepage	= buffer_migrate_page,	/* Move a page cache page from
1568						   one physical page to an
1569						   other. */
1570	.error_remove_page = generic_error_remove_page,
1571};
1572
1573#ifdef NTFS_RW
1574
1575/**
1576 * mark_ntfs_record_dirty - mark an ntfs record dirty
1577 * @page:	page containing the ntfs record to mark dirty
1578 * @ofs:	byte offset within @page at which the ntfs record begins
1579 *
1580 * Set the buffers and the page in which the ntfs record is located dirty.
1581 *
1582 * The latter also marks the vfs inode the ntfs record belongs to dirty
1583 * (I_DIRTY_PAGES only).
1584 *
1585 * If the page does not have buffers, we create them and set them uptodate.
1586 * The page may not be locked which is why we need to handle the buffers under
1587 * the mapping->private_lock.  Once the buffers are marked dirty we no longer
1588 * need the lock since try_to_free_buffers() does not free dirty buffers.
1589 */
1590void mark_ntfs_record_dirty(struct page *page, const unsigned int ofs) {
1591	struct address_space *mapping = page->mapping;
1592	ntfs_inode *ni = NTFS_I(mapping->host);
1593	struct buffer_head *bh, *head, *buffers_to_free = NULL;
1594	unsigned int end, bh_size, bh_ofs;
1595
1596	BUG_ON(!PageUptodate(page));
1597	end = ofs + ni->itype.index.block_size;
1598	bh_size = VFS_I(ni)->i_sb->s_blocksize;
1599	spin_lock(&mapping->private_lock);
1600	if (unlikely(!page_has_buffers(page))) {
1601		spin_unlock(&mapping->private_lock);
1602		bh = head = alloc_page_buffers(page, bh_size, 1);
1603		spin_lock(&mapping->private_lock);
1604		if (likely(!page_has_buffers(page))) {
1605			struct buffer_head *tail;
1606
1607			do {
1608				set_buffer_uptodate(bh);
1609				tail = bh;
1610				bh = bh->b_this_page;
1611			} while (bh);
1612			tail->b_this_page = head;
1613			attach_page_buffers(page, head);
1614		} else
1615			buffers_to_free = bh;
1616	}
1617	bh = head = page_buffers(page);
1618	BUG_ON(!bh);
1619	do {
1620		bh_ofs = bh_offset(bh);
1621		if (bh_ofs + bh_size <= ofs)
1622			continue;
1623		if (unlikely(bh_ofs >= end))
1624			break;
1625		set_buffer_dirty(bh);
1626	} while ((bh = bh->b_this_page) != head);
1627	spin_unlock(&mapping->private_lock);
1628	__set_page_dirty_nobuffers(page);
1629	if (unlikely(buffers_to_free)) {
1630		do {
1631			bh = buffers_to_free->b_this_page;
1632			free_buffer_head(buffers_to_free);
1633			buffers_to_free = bh;
1634		} while (buffers_to_free);
1635	}
1636}
1637
1638#endif /* NTFS_RW */