1/**
2 * compress.c - NTFS kernel compressed attributes handling.
3 *		Part of the Linux-NTFS project.
4 *
5 * Copyright (c) 2001-2004 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/fs.h>
25#include <linux/buffer_head.h>
26#include <linux/blkdev.h>
27#include <linux/vmalloc.h>
28
29#include "attrib.h"
30#include "inode.h"
31#include "debug.h"
32#include "ntfs.h"
33
34/**
35 * ntfs_compression_constants - enum of constants used in the compression code
36 */
37typedef enum {
38	/* Token types and access mask. */
39	NTFS_SYMBOL_TOKEN	=	0,
40	NTFS_PHRASE_TOKEN	=	1,
41	NTFS_TOKEN_MASK		=	1,
42
43	/* Compression sub-block constants. */
44	NTFS_SB_SIZE_MASK	=	0x0fff,
45	NTFS_SB_SIZE		=	0x1000,
46	NTFS_SB_IS_COMPRESSED	=	0x8000,
47
48	/*
49	 * The maximum compression block size is by definition 16 * the cluster
50	 * size, with the maximum supported cluster size being 4kiB. Thus the
51	 * maximum compression buffer size is 64kiB, so we use this when
52	 * initializing the compression buffer.
53	 */
54	NTFS_MAX_CB_SIZE	= 64 * 1024,
55} ntfs_compression_constants;
56
57/**
58 * ntfs_compression_buffer - one buffer for the decompression engine
59 */
60static u8 *ntfs_compression_buffer = NULL;
61
62/**
63 * ntfs_cb_lock - spinlock which protects ntfs_compression_buffer
64 */
65static DEFINE_SPINLOCK(ntfs_cb_lock);
66
67/**
68 * allocate_compression_buffers - allocate the decompression buffers
69 *
70 * Caller has to hold the ntfs_lock mutex.
71 *
72 * Return 0 on success or -ENOMEM if the allocations failed.
73 */
74int allocate_compression_buffers(void)
75{
76	BUG_ON(ntfs_compression_buffer);
77
78	ntfs_compression_buffer = vmalloc(NTFS_MAX_CB_SIZE);
79	if (!ntfs_compression_buffer)
80		return -ENOMEM;
81	return 0;
82}
83
84/**
85 * free_compression_buffers - free the decompression buffers
86 *
87 * Caller has to hold the ntfs_lock mutex.
88 */
89void free_compression_buffers(void)
90{
91	BUG_ON(!ntfs_compression_buffer);
92	vfree(ntfs_compression_buffer);
93	ntfs_compression_buffer = NULL;
94}
95
96/**
97 * zero_partial_compressed_page - zero out of bounds compressed page region
98 */
99static void zero_partial_compressed_page(struct page *page,
100		const s64 initialized_size)
101{
102	u8 *kp = page_address(page);
103	unsigned int kp_ofs;
104
105	ntfs_debug("Zeroing page region outside initialized size.");
106	if (((s64)page->index << PAGE_CACHE_SHIFT) >= initialized_size) {
107		clear_page(kp);
108		return;
109	}
110	kp_ofs = initialized_size & ~PAGE_CACHE_MASK;
111	memset(kp + kp_ofs, 0, PAGE_CACHE_SIZE - kp_ofs);
112	return;
113}
114
115/**
116 * handle_bounds_compressed_page - test for&handle out of bounds compressed page
117 */
118static inline void handle_bounds_compressed_page(struct page *page,
119		const loff_t i_size, const s64 initialized_size)
120{
121	if ((page->index >= (initialized_size >> PAGE_CACHE_SHIFT)) &&
122			(initialized_size < i_size))
123		zero_partial_compressed_page(page, initialized_size);
124	return;
125}
126
127/**
128 * ntfs_decompress - decompress a compression block into an array of pages
129 * @dest_pages:		destination array of pages
130 * @dest_index:		current index into @dest_pages (IN/OUT)
131 * @dest_ofs:		current offset within @dest_pages[@dest_index] (IN/OUT)
132 * @dest_max_index:	maximum index into @dest_pages (IN)
133 * @dest_max_ofs:	maximum offset within @dest_pages[@dest_max_index] (IN)
134 * @xpage:		the target page (-1 if none) (IN)
135 * @xpage_done:		set to 1 if xpage was completed successfully (IN/OUT)
136 * @cb_start:		compression block to decompress (IN)
137 * @cb_size:		size of compression block @cb_start in bytes (IN)
138 * @i_size:		file size when we started the read (IN)
139 * @initialized_size:	initialized file size when we started the read (IN)
140 *
141 * The caller must have disabled preemption. ntfs_decompress() reenables it when
142 * the critical section is finished.
143 *
144 * This decompresses the compression block @cb_start into the array of
145 * destination pages @dest_pages starting at index @dest_index into @dest_pages
146 * and at offset @dest_pos into the page @dest_pages[@dest_index].
147 *
148 * When the page @dest_pages[@xpage] is completed, @xpage_done is set to 1.
149 * If xpage is -1 or @xpage has not been completed, @xpage_done is not modified.
150 *
151 * @cb_start is a pointer to the compression block which needs decompressing
152 * and @cb_size is the size of @cb_start in bytes (8-64kiB).
153 *
154 * Return 0 if success or -EOVERFLOW on error in the compressed stream.
155 * @xpage_done indicates whether the target page (@dest_pages[@xpage]) was
156 * completed during the decompression of the compression block (@cb_start).
157 *
158 * Warning: This function *REQUIRES* PAGE_CACHE_SIZE >= 4096 or it will blow up
159 * unpredicatbly! You have been warned!
160 *
161 * Note to hackers: This function may not sleep until it has finished accessing
162 * the compression block @cb_start as it is a per-CPU buffer.
163 */
164static int ntfs_decompress(struct page *dest_pages[], int *dest_index,
165		int *dest_ofs, const int dest_max_index, const int dest_max_ofs,
166		const int xpage, char *xpage_done, u8 *const cb_start,
167		const u32 cb_size, const loff_t i_size,
168		const s64 initialized_size)
169{
170	/*
171	 * Pointers into the compressed data, i.e. the compression block (cb),
172	 * and the therein contained sub-blocks (sb).
173	 */
174	u8 *cb_end = cb_start + cb_size; /* End of cb. */
175	u8 *cb = cb_start;	/* Current position in cb. */
176	u8 *cb_sb_start = cb;	/* Beginning of the current sb in the cb. */
177	u8 *cb_sb_end;		/* End of current sb / beginning of next sb. */
178
179	/* Variables for uncompressed data / destination. */
180	struct page *dp;	/* Current destination page being worked on. */
181	u8 *dp_addr;		/* Current pointer into dp. */
182	u8 *dp_sb_start;	/* Start of current sub-block in dp. */
183	u8 *dp_sb_end;		/* End of current sb in dp (dp_sb_start +
184				   NTFS_SB_SIZE). */
185	u16 do_sb_start;	/* @dest_ofs when starting this sub-block. */
186	u16 do_sb_end;		/* @dest_ofs of end of this sb (do_sb_start +
187				   NTFS_SB_SIZE). */
188
189	/* Variables for tag and token parsing. */
190	u8 tag;			/* Current tag. */
191	int token;		/* Loop counter for the eight tokens in tag. */
192
193	/* Need this because we can't sleep, so need two stages. */
194	int completed_pages[dest_max_index - *dest_index + 1];
195	int nr_completed_pages = 0;
196
197	/* Default error code. */
198	int err = -EOVERFLOW;
199
200	ntfs_debug("Entering, cb_size = 0x%x.", cb_size);
201do_next_sb:
202	ntfs_debug("Beginning sub-block at offset = 0x%zx in the cb.",
203			cb - cb_start);
204	/*
205	 * Have we reached the end of the compression block or the end of the
206	 * decompressed data?  The latter can happen for example if the current
207	 * position in the compression block is one byte before its end so the
208	 * first two checks do not detect it.
209	 */
210	if (cb == cb_end || !le16_to_cpup((le16*)cb) ||
211			(*dest_index == dest_max_index &&
212			*dest_ofs == dest_max_ofs)) {
213		int i;
214
215		ntfs_debug("Completed. Returning success (0).");
216		err = 0;
217return_error:
218		/* We can sleep from now on, so we drop lock. */
219		spin_unlock(&ntfs_cb_lock);
220		/* Second stage: finalize completed pages. */
221		if (nr_completed_pages > 0) {
222			for (i = 0; i < nr_completed_pages; i++) {
223				int di = completed_pages[i];
224
225				dp = dest_pages[di];
226				/*
227				 * If we are outside the initialized size, zero
228				 * the out of bounds page range.
229				 */
230				handle_bounds_compressed_page(dp, i_size,
231						initialized_size);
232				flush_dcache_page(dp);
233				kunmap(dp);
234				SetPageUptodate(dp);
235				unlock_page(dp);
236				if (di == xpage)
237					*xpage_done = 1;
238				else
239					page_cache_release(dp);
240				dest_pages[di] = NULL;
241			}
242		}
243		return err;
244	}
245
246	/* Setup offsets for the current sub-block destination. */
247	do_sb_start = *dest_ofs;
248	do_sb_end = do_sb_start + NTFS_SB_SIZE;
249
250	/* Check that we are still within allowed boundaries. */
251	if (*dest_index == dest_max_index && do_sb_end > dest_max_ofs)
252		goto return_overflow;
253
254	/* Does the minimum size of a compressed sb overflow valid range? */
255	if (cb + 6 > cb_end)
256		goto return_overflow;
257
258	/* Setup the current sub-block source pointers and validate range. */
259	cb_sb_start = cb;
260	cb_sb_end = cb_sb_start + (le16_to_cpup((le16*)cb) & NTFS_SB_SIZE_MASK)
261			+ 3;
262	if (cb_sb_end > cb_end)
263		goto return_overflow;
264
265	/* Get the current destination page. */
266	dp = dest_pages[*dest_index];
267	if (!dp) {
268		/* No page present. Skip decompression of this sub-block. */
269		cb = cb_sb_end;
270
271		/* Advance destination position to next sub-block. */
272		*dest_ofs = (*dest_ofs + NTFS_SB_SIZE) & ~PAGE_CACHE_MASK;
273		if (!*dest_ofs && (++*dest_index > dest_max_index))
274			goto return_overflow;
275		goto do_next_sb;
276	}
277
278	/* We have a valid destination page. Setup the destination pointers. */
279	dp_addr = (u8*)page_address(dp) + do_sb_start;
280
281	/* Now, we are ready to process the current sub-block (sb). */
282	if (!(le16_to_cpup((le16*)cb) & NTFS_SB_IS_COMPRESSED)) {
283		ntfs_debug("Found uncompressed sub-block.");
284		/* This sb is not compressed, just copy it into destination. */
285
286		/* Advance source position to first data byte. */
287		cb += 2;
288
289		/* An uncompressed sb must be full size. */
290		if (cb_sb_end - cb != NTFS_SB_SIZE)
291			goto return_overflow;
292
293		/* Copy the block and advance the source position. */
294		memcpy(dp_addr, cb, NTFS_SB_SIZE);
295		cb += NTFS_SB_SIZE;
296
297		/* Advance destination position to next sub-block. */
298		*dest_ofs += NTFS_SB_SIZE;
299		if (!(*dest_ofs &= ~PAGE_CACHE_MASK)) {
300finalize_page:
301			/*
302			 * First stage: add current page index to array of
303			 * completed pages.
304			 */
305			completed_pages[nr_completed_pages++] = *dest_index;
306			if (++*dest_index > dest_max_index)
307				goto return_overflow;
308		}
309		goto do_next_sb;
310	}
311	ntfs_debug("Found compressed sub-block.");
312	/* This sb is compressed, decompress it into destination. */
313
314	/* Setup destination pointers. */
315	dp_sb_start = dp_addr;
316	dp_sb_end = dp_sb_start + NTFS_SB_SIZE;
317
318	/* Forward to the first tag in the sub-block. */
319	cb += 2;
320do_next_tag:
321	if (cb == cb_sb_end) {
322		/* Check if the decompressed sub-block was not full-length. */
323		if (dp_addr < dp_sb_end) {
324			int nr_bytes = do_sb_end - *dest_ofs;
325
326			ntfs_debug("Filling incomplete sub-block with "
327					"zeroes.");
328			/* Zero remainder and update destination position. */
329			memset(dp_addr, 0, nr_bytes);
330			*dest_ofs += nr_bytes;
331		}
332		/* We have finished the current sub-block. */
333		if (!(*dest_ofs &= ~PAGE_CACHE_MASK))
334			goto finalize_page;
335		goto do_next_sb;
336	}
337
338	/* Check we are still in range. */
339	if (cb > cb_sb_end || dp_addr > dp_sb_end)
340		goto return_overflow;
341
342	/* Get the next tag and advance to first token. */
343	tag = *cb++;
344
345	/* Parse the eight tokens described by the tag. */
346	for (token = 0; token < 8; token++, tag >>= 1) {
347		u16 lg, pt, length, max_non_overlap;
348		register u16 i;
349		u8 *dp_back_addr;
350
351		/* Check if we are done / still in range. */
352		if (cb >= cb_sb_end || dp_addr > dp_sb_end)
353			break;
354
355		/* Determine token type and parse appropriately.*/
356		if ((tag & NTFS_TOKEN_MASK) == NTFS_SYMBOL_TOKEN) {
357			/*
358			 * We have a symbol token, copy the symbol across, and
359			 * advance the source and destination positions.
360			 */
361			*dp_addr++ = *cb++;
362			++*dest_ofs;
363
364			/* Continue with the next token. */
365			continue;
366		}
367
368		/*
369		 * We have a phrase token. Make sure it is not the first tag in
370		 * the sb as this is illegal and would confuse the code below.
371		 */
372		if (dp_addr == dp_sb_start)
373			goto return_overflow;
374
375		/*
376		 * Determine the number of bytes to go back (p) and the number
377		 * of bytes to copy (l). We use an optimized algorithm in which
378		 * we first calculate log2(current destination position in sb),
379		 * which allows determination of l and p in O(1) rather than
380		 * O(n). We just need an arch-optimized log2() function now.
381		 */
382		lg = 0;
383		for (i = *dest_ofs - do_sb_start - 1; i >= 0x10; i >>= 1)
384			lg++;
385
386		/* Get the phrase token into i. */
387		pt = le16_to_cpup((le16*)cb);
388
389		/*
390		 * Calculate starting position of the byte sequence in
391		 * the destination using the fact that p = (pt >> (12 - lg)) + 1
392		 * and make sure we don't go too far back.
393		 */
394		dp_back_addr = dp_addr - (pt >> (12 - lg)) - 1;
395		if (dp_back_addr < dp_sb_start)
396			goto return_overflow;
397
398		/* Now calculate the length of the byte sequence. */
399		length = (pt & (0xfff >> lg)) + 3;
400
401		/* Advance destination position and verify it is in range. */
402		*dest_ofs += length;
403		if (*dest_ofs > do_sb_end)
404			goto return_overflow;
405
406		/* The number of non-overlapping bytes. */
407		max_non_overlap = dp_addr - dp_back_addr;
408
409		if (length <= max_non_overlap) {
410			/* The byte sequence doesn't overlap, just copy it. */
411			memcpy(dp_addr, dp_back_addr, length);
412
413			/* Advance destination pointer. */
414			dp_addr += length;
415		} else {
416			/*
417			 * The byte sequence does overlap, copy non-overlapping
418			 * part and then do a slow byte by byte copy for the
419			 * overlapping part. Also, advance the destination
420			 * pointer.
421			 */
422			memcpy(dp_addr, dp_back_addr, max_non_overlap);
423			dp_addr += max_non_overlap;
424			dp_back_addr += max_non_overlap;
425			length -= max_non_overlap;
426			while (length--)
427				*dp_addr++ = *dp_back_addr++;
428		}
429
430		/* Advance source position and continue with the next token. */
431		cb += 2;
432	}
433
434	/* No tokens left in the current tag. Continue with the next tag. */
435	goto do_next_tag;
436
437return_overflow:
438	ntfs_error(NULL, "Failed. Returning -EOVERFLOW.");
439	goto return_error;
440}
441
442int ntfs_read_compressed_block(struct page *page)
443{
444	loff_t i_size;
445	s64 initialized_size;
446	struct address_space *mapping = page->mapping;
447	ntfs_inode *ni = NTFS_I(mapping->host);
448	ntfs_volume *vol = ni->vol;
449	struct super_block *sb = vol->sb;
450	runlist_element *rl;
451	unsigned long flags, block_size = sb->s_blocksize;
452	unsigned char block_size_bits = sb->s_blocksize_bits;
453	u8 *cb, *cb_pos, *cb_end;
454	struct buffer_head **bhs;
455	unsigned long offset, index = page->index;
456	u32 cb_size = ni->itype.compressed.block_size;
457	u64 cb_size_mask = cb_size - 1UL;
458	VCN vcn;
459	LCN lcn;
460	/* The first wanted vcn (minimum alignment is PAGE_CACHE_SIZE). */
461	VCN start_vcn = (((s64)index << PAGE_CACHE_SHIFT) & ~cb_size_mask) >>
462			vol->cluster_size_bits;
463	/*
464	 * The first vcn after the last wanted vcn (minumum alignment is again
465	 * PAGE_CACHE_SIZE.
466	 */
467	VCN end_vcn = ((((s64)(index + 1UL) << PAGE_CACHE_SHIFT) + cb_size - 1)
468			& ~cb_size_mask) >> vol->cluster_size_bits;
469	/* Number of compression blocks (cbs) in the wanted vcn range. */
470	unsigned int nr_cbs = (end_vcn - start_vcn) << vol->cluster_size_bits
471			>> ni->itype.compressed.block_size_bits;
472	/*
473	 * Number of pages required to store the uncompressed data from all
474	 * compression blocks (cbs) overlapping @page. Due to alignment
475	 * guarantees of start_vcn and end_vcn, no need to round up here.
476	 */
477	unsigned int nr_pages = (end_vcn - start_vcn) <<
478			vol->cluster_size_bits >> PAGE_CACHE_SHIFT;
479	unsigned int xpage, max_page, cur_page, cur_ofs, i;
480	unsigned int cb_clusters, cb_max_ofs;
481	int block, max_block, cb_max_page, bhs_size, nr_bhs, err = 0;
482	struct page **pages;
483	unsigned char xpage_done = 0;
484
485	ntfs_debug("Entering, page->index = 0x%lx, cb_size = 0x%x, nr_pages = "
486			"%i.", index, cb_size, nr_pages);
487	/*
488	 * Bad things happen if we get here for anything that is not an
489	 * unnamed $DATA attribute.
490	 */
491	BUG_ON(ni->type != AT_DATA);
492	BUG_ON(ni->name_len);
493
494	pages = kmalloc(nr_pages * sizeof(struct page *), GFP_NOFS);
495
496	/* Allocate memory to store the buffer heads we need. */
497	bhs_size = cb_size / block_size * sizeof(struct buffer_head *);
498	bhs = kmalloc(bhs_size, GFP_NOFS);
499
500	if (unlikely(!pages || !bhs)) {
501		kfree(bhs);
502		kfree(pages);
503		unlock_page(page);
504		ntfs_error(vol->sb, "Failed to allocate internal buffers.");
505		return -ENOMEM;
506	}
507
508	/*
509	 * We have already been given one page, this is the one we must do.
510	 * Once again, the alignment guarantees keep it simple.
511	 */
512	offset = start_vcn << vol->cluster_size_bits >> PAGE_CACHE_SHIFT;
513	xpage = index - offset;
514	pages[xpage] = page;
515	/*
516	 * The remaining pages need to be allocated and inserted into the page
517	 * cache, alignment guarantees keep all the below much simpler. (-8
518	 */
519	read_lock_irqsave(&ni->size_lock, flags);
520	i_size = i_size_read(VFS_I(ni));
521	initialized_size = ni->initialized_size;
522	read_unlock_irqrestore(&ni->size_lock, flags);
523	max_page = ((i_size + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
524			offset;
525	if (nr_pages < max_page)
526		max_page = nr_pages;
527	for (i = 0; i < max_page; i++, offset++) {
528		if (i != xpage)
529			pages[i] = grab_cache_page_nowait(mapping, offset);
530		page = pages[i];
531		if (page) {
532			/*
533			 * We only (re)read the page if it isn't already read
534			 * in and/or dirty or we would be losing data or at
535			 * least wasting our time.
536			 */
537			if (!PageDirty(page) && (!PageUptodate(page) ||
538					PageError(page))) {
539				ClearPageError(page);
540				kmap(page);
541				continue;
542			}
543			unlock_page(page);
544			page_cache_release(page);
545			pages[i] = NULL;
546		}
547	}
548
549	/*
550	 * We have the runlist, and all the destination pages we need to fill.
551	 * Now read the first compression block.
552	 */
553	cur_page = 0;
554	cur_ofs = 0;
555	cb_clusters = ni->itype.compressed.block_clusters;
556do_next_cb:
557	nr_cbs--;
558	nr_bhs = 0;
559
560	/* Read all cb buffer heads one cluster at a time. */
561	rl = NULL;
562	for (vcn = start_vcn, start_vcn += cb_clusters; vcn < start_vcn;
563			vcn++) {
564		bool is_retry = false;
565
566		if (!rl) {
567lock_retry_remap:
568			down_read(&ni->runlist.lock);
569			rl = ni->runlist.rl;
570		}
571		if (likely(rl != NULL)) {
572			/* Seek to element containing target vcn. */
573			while (rl->length && rl[1].vcn <= vcn)
574				rl++;
575			lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
576		} else
577			lcn = LCN_RL_NOT_MAPPED;
578		ntfs_debug("Reading vcn = 0x%llx, lcn = 0x%llx.",
579				(unsigned long long)vcn,
580				(unsigned long long)lcn);
581		if (lcn < 0) {
582			/*
583			 * When we reach the first sparse cluster we have
584			 * finished with the cb.
585			 */
586			if (lcn == LCN_HOLE)
587				break;
588			if (is_retry || lcn != LCN_RL_NOT_MAPPED)
589				goto rl_err;
590			is_retry = true;
591			/*
592			 * Attempt to map runlist, dropping lock for the
593			 * duration.
594			 */
595			up_read(&ni->runlist.lock);
596			if (!ntfs_map_runlist(ni, vcn))
597				goto lock_retry_remap;
598			goto map_rl_err;
599		}
600		block = lcn << vol->cluster_size_bits >> block_size_bits;
601		/* Read the lcn from device in chunks of block_size bytes. */
602		max_block = block + (vol->cluster_size >> block_size_bits);
603		do {
604			ntfs_debug("block = 0x%x.", block);
605			if (unlikely(!(bhs[nr_bhs] = sb_getblk(sb, block))))
606				goto getblk_err;
607			nr_bhs++;
608		} while (++block < max_block);
609	}
610
611	/* Release the lock if we took it. */
612	if (rl)
613		up_read(&ni->runlist.lock);
614
615	/* Setup and initiate io on all buffer heads. */
616	for (i = 0; i < nr_bhs; i++) {
617		struct buffer_head *tbh = bhs[i];
618
619		if (unlikely(test_set_buffer_locked(tbh)))
620			continue;
621		if (unlikely(buffer_uptodate(tbh))) {
622			unlock_buffer(tbh);
623			continue;
624		}
625		get_bh(tbh);
626		tbh->b_end_io = end_buffer_read_sync;
627		submit_bh(READ, tbh);
628	}
629
630	/* Wait for io completion on all buffer heads. */
631	for (i = 0; i < nr_bhs; i++) {
632		struct buffer_head *tbh = bhs[i];
633
634		if (buffer_uptodate(tbh))
635			continue;
636		wait_on_buffer(tbh);
637		/*
638		 * We need an optimization barrier here, otherwise we start
639		 * hitting the below fixup code when accessing a loopback
640		 * mounted ntfs partition. This indicates either there is a
641		 * race condition in the loop driver or, more likely, gcc
642		 * overoptimises the code without the barrier and it doesn't
643		 * do the Right Thing(TM).
644		 */
645		barrier();
646		if (unlikely(!buffer_uptodate(tbh))) {
647			ntfs_warning(vol->sb, "Buffer is unlocked but not "
648					"uptodate! Unplugging the disk queue "
649					"and rescheduling.");
650			get_bh(tbh);
651			blk_run_address_space(mapping);
652			schedule();
653			put_bh(tbh);
654			if (unlikely(!buffer_uptodate(tbh)))
655				goto read_err;
656			ntfs_warning(vol->sb, "Buffer is now uptodate. Good.");
657		}
658	}
659
660	/*
661	 * Get the compression buffer. We must not sleep any more
662	 * until we are finished with it.
663	 */
664	spin_lock(&ntfs_cb_lock);
665	cb = ntfs_compression_buffer;
666
667	BUG_ON(!cb);
668
669	cb_pos = cb;
670	cb_end = cb + cb_size;
671
672	/* Copy the buffer heads into the contiguous buffer. */
673	for (i = 0; i < nr_bhs; i++) {
674		memcpy(cb_pos, bhs[i]->b_data, block_size);
675		cb_pos += block_size;
676	}
677
678	/* Just a precaution. */
679	if (cb_pos + 2 <= cb + cb_size)
680		*(u16*)cb_pos = 0;
681
682	/* Reset cb_pos back to the beginning. */
683	cb_pos = cb;
684
685	/* We now have both source (if present) and destination. */
686	ntfs_debug("Successfully read the compression block.");
687
688	/* The last page and maximum offset within it for the current cb. */
689	cb_max_page = (cur_page << PAGE_CACHE_SHIFT) + cur_ofs + cb_size;
690	cb_max_ofs = cb_max_page & ~PAGE_CACHE_MASK;
691	cb_max_page >>= PAGE_CACHE_SHIFT;
692
693	/* Catch end of file inside a compression block. */
694	if (cb_max_page > max_page)
695		cb_max_page = max_page;
696
697	if (vcn == start_vcn - cb_clusters) {
698		/* Sparse cb, zero out page range overlapping the cb. */
699		ntfs_debug("Found sparse compression block.");
700		/* We can sleep from now on, so we drop lock. */
701		spin_unlock(&ntfs_cb_lock);
702		if (cb_max_ofs)
703			cb_max_page--;
704		for (; cur_page < cb_max_page; cur_page++) {
705			page = pages[cur_page];
706			if (page) {
707				if (likely(!cur_ofs))
708					clear_page(page_address(page));
709				else
710					memset(page_address(page) + cur_ofs, 0,
711							PAGE_CACHE_SIZE -
712							cur_ofs);
713				flush_dcache_page(page);
714				kunmap(page);
715				SetPageUptodate(page);
716				unlock_page(page);
717				if (cur_page == xpage)
718					xpage_done = 1;
719				else
720					page_cache_release(page);
721				pages[cur_page] = NULL;
722			}
723			cb_pos += PAGE_CACHE_SIZE - cur_ofs;
724			cur_ofs = 0;
725			if (cb_pos >= cb_end)
726				break;
727		}
728		/* If we have a partial final page, deal with it now. */
729		if (cb_max_ofs && cb_pos < cb_end) {
730			page = pages[cur_page];
731			if (page)
732				memset(page_address(page) + cur_ofs, 0,
733						cb_max_ofs - cur_ofs);
734			/*
735			 * No need to update cb_pos at this stage:
736			 *	cb_pos += cb_max_ofs - cur_ofs;
737			 */
738			cur_ofs = cb_max_ofs;
739		}
740	} else if (vcn == start_vcn) {
741		/* We can't sleep so we need two stages. */
742		unsigned int cur2_page = cur_page;
743		unsigned int cur_ofs2 = cur_ofs;
744		u8 *cb_pos2 = cb_pos;
745
746		ntfs_debug("Found uncompressed compression block.");
747		/* Uncompressed cb, copy it to the destination pages. */
748		/*
749		 * TODO: As a big optimization, we could detect this case
750		 * before we read all the pages and use block_read_full_page()
751		 * on all full pages instead (we still have to treat partial
752		 * pages especially but at least we are getting rid of the
753		 * synchronous io for the majority of pages.
754		 * Or if we choose not to do the read-ahead/-behind stuff, we
755		 * could just return block_read_full_page(pages[xpage]) as long
756		 * as PAGE_CACHE_SIZE <= cb_size.
757		 */
758		if (cb_max_ofs)
759			cb_max_page--;
760		/* First stage: copy data into destination pages. */
761		for (; cur_page < cb_max_page; cur_page++) {
762			page = pages[cur_page];
763			if (page)
764				memcpy(page_address(page) + cur_ofs, cb_pos,
765						PAGE_CACHE_SIZE - cur_ofs);
766			cb_pos += PAGE_CACHE_SIZE - cur_ofs;
767			cur_ofs = 0;
768			if (cb_pos >= cb_end)
769				break;
770		}
771		/* If we have a partial final page, deal with it now. */
772		if (cb_max_ofs && cb_pos < cb_end) {
773			page = pages[cur_page];
774			if (page)
775				memcpy(page_address(page) + cur_ofs, cb_pos,
776						cb_max_ofs - cur_ofs);
777			cb_pos += cb_max_ofs - cur_ofs;
778			cur_ofs = cb_max_ofs;
779		}
780		/* We can sleep from now on, so drop lock. */
781		spin_unlock(&ntfs_cb_lock);
782		/* Second stage: finalize pages. */
783		for (; cur2_page < cb_max_page; cur2_page++) {
784			page = pages[cur2_page];
785			if (page) {
786				/*
787				 * If we are outside the initialized size, zero
788				 * the out of bounds page range.
789				 */
790				handle_bounds_compressed_page(page, i_size,
791						initialized_size);
792				flush_dcache_page(page);
793				kunmap(page);
794				SetPageUptodate(page);
795				unlock_page(page);
796				if (cur2_page == xpage)
797					xpage_done = 1;
798				else
799					page_cache_release(page);
800				pages[cur2_page] = NULL;
801			}
802			cb_pos2 += PAGE_CACHE_SIZE - cur_ofs2;
803			cur_ofs2 = 0;
804			if (cb_pos2 >= cb_end)
805				break;
806		}
807	} else {
808		/* Compressed cb, decompress it into the destination page(s). */
809		unsigned int prev_cur_page = cur_page;
810
811		ntfs_debug("Found compressed compression block.");
812		err = ntfs_decompress(pages, &cur_page, &cur_ofs,
813				cb_max_page, cb_max_ofs, xpage, &xpage_done,
814				cb_pos,	cb_size - (cb_pos - cb), i_size,
815				initialized_size);
816		/*
817		 * We can sleep from now on, lock already dropped by
818		 * ntfs_decompress().
819		 */
820		if (err) {
821			ntfs_error(vol->sb, "ntfs_decompress() failed in inode "
822					"0x%lx with error code %i. Skipping "
823					"this compression block.",
824					ni->mft_no, -err);
825			/* Release the unfinished pages. */
826			for (; prev_cur_page < cur_page; prev_cur_page++) {
827				page = pages[prev_cur_page];
828				if (page) {
829					flush_dcache_page(page);
830					kunmap(page);
831					unlock_page(page);
832					if (prev_cur_page != xpage)
833						page_cache_release(page);
834					pages[prev_cur_page] = NULL;
835				}
836			}
837		}
838	}
839
840	/* Release the buffer heads. */
841	for (i = 0; i < nr_bhs; i++)
842		brelse(bhs[i]);
843
844	/* Do we have more work to do? */
845	if (nr_cbs)
846		goto do_next_cb;
847
848	/* We no longer need the list of buffer heads. */
849	kfree(bhs);
850
851	/* Clean up if we have any pages left. Should never happen. */
852	for (cur_page = 0; cur_page < max_page; cur_page++) {
853		page = pages[cur_page];
854		if (page) {
855			ntfs_error(vol->sb, "Still have pages left! "
856					"Terminating them with extreme "
857					"prejudice.  Inode 0x%lx, page index "
858					"0x%lx.", ni->mft_no, page->index);
859			flush_dcache_page(page);
860			kunmap(page);
861			unlock_page(page);
862			if (cur_page != xpage)
863				page_cache_release(page);
864			pages[cur_page] = NULL;
865		}
866	}
867
868	/* We no longer need the list of pages. */
869	kfree(pages);
870
871	/* If we have completed the requested page, we return success. */
872	if (likely(xpage_done))
873		return 0;
874
875	ntfs_debug("Failed. Returning error code %s.", err == -EOVERFLOW ?
876			"EOVERFLOW" : (!err ? "EIO" : "unkown error"));
877	return err < 0 ? err : -EIO;
878
879read_err:
880	ntfs_error(vol->sb, "IO error while reading compressed data.");
881	/* Release the buffer heads. */
882	for (i = 0; i < nr_bhs; i++)
883		brelse(bhs[i]);
884	goto err_out;
885
886map_rl_err:
887	ntfs_error(vol->sb, "ntfs_map_runlist() failed. Cannot read "
888			"compression block.");
889	goto err_out;
890
891rl_err:
892	up_read(&ni->runlist.lock);
893	ntfs_error(vol->sb, "ntfs_rl_vcn_to_lcn() failed. Cannot read "
894			"compression block.");
895	goto err_out;
896
897getblk_err:
898	up_read(&ni->runlist.lock);
899	ntfs_error(vol->sb, "getblk() failed. Cannot read compression block.");
900
901err_out:
902	kfree(bhs);
903	for (i = cur_page; i < max_page; i++) {
904		page = pages[i];
905		if (page) {
906			flush_dcache_page(page);
907			kunmap(page);
908			unlock_page(page);
909			if (i != xpage)
910				page_cache_release(page);
911		}
912	}
913	kfree(pages);
914	return -EIO;
915}
916