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 #include <linux/slab.h>
29 
30 #include "attrib.h"
31 #include "inode.h"
32 #include "debug.h"
33 #include "ntfs.h"
34 
35 /**
36  * ntfs_compression_constants - enum of constants used in the compression code
37  */
38 typedef enum {
39 	/* Token types and access mask. */
40 	NTFS_SYMBOL_TOKEN	=	0,
41 	NTFS_PHRASE_TOKEN	=	1,
42 	NTFS_TOKEN_MASK		=	1,
43 
44 	/* Compression sub-block constants. */
45 	NTFS_SB_SIZE_MASK	=	0x0fff,
46 	NTFS_SB_SIZE		=	0x1000,
47 	NTFS_SB_IS_COMPRESSED	=	0x8000,
48 
49 	/*
50 	 * The maximum compression block size is by definition 16 * the cluster
51 	 * size, with the maximum supported cluster size being 4kiB. Thus the
52 	 * maximum compression buffer size is 64kiB, so we use this when
53 	 * initializing the compression buffer.
54 	 */
55 	NTFS_MAX_CB_SIZE	= 64 * 1024,
56 } ntfs_compression_constants;
57 
58 /**
59  * ntfs_compression_buffer - one buffer for the decompression engine
60  */
61 static u8 *ntfs_compression_buffer;
62 
63 /**
64  * ntfs_cb_lock - spinlock which protects ntfs_compression_buffer
65  */
66 static DEFINE_SPINLOCK(ntfs_cb_lock);
67 
68 /**
69  * allocate_compression_buffers - allocate the decompression buffers
70  *
71  * Caller has to hold the ntfs_lock mutex.
72  *
73  * Return 0 on success or -ENOMEM if the allocations failed.
74  */
allocate_compression_buffers(void)75 int allocate_compression_buffers(void)
76 {
77 	BUG_ON(ntfs_compression_buffer);
78 
79 	ntfs_compression_buffer = vmalloc(NTFS_MAX_CB_SIZE);
80 	if (!ntfs_compression_buffer)
81 		return -ENOMEM;
82 	return 0;
83 }
84 
85 /**
86  * free_compression_buffers - free the decompression buffers
87  *
88  * Caller has to hold the ntfs_lock mutex.
89  */
free_compression_buffers(void)90 void free_compression_buffers(void)
91 {
92 	BUG_ON(!ntfs_compression_buffer);
93 	vfree(ntfs_compression_buffer);
94 	ntfs_compression_buffer = NULL;
95 }
96 
97 /**
98  * zero_partial_compressed_page - zero out of bounds compressed page region
99  */
zero_partial_compressed_page(struct page * page,const s64 initialized_size)100 static void zero_partial_compressed_page(struct page *page,
101 		const s64 initialized_size)
102 {
103 	u8 *kp = page_address(page);
104 	unsigned int kp_ofs;
105 
106 	ntfs_debug("Zeroing page region outside initialized size.");
107 	if (((s64)page->index << PAGE_SHIFT) >= initialized_size) {
108 		clear_page(kp);
109 		return;
110 	}
111 	kp_ofs = initialized_size & ~PAGE_MASK;
112 	memset(kp + kp_ofs, 0, PAGE_SIZE - kp_ofs);
113 	return;
114 }
115 
116 /**
117  * handle_bounds_compressed_page - test for&handle out of bounds compressed page
118  */
handle_bounds_compressed_page(struct page * page,const loff_t i_size,const s64 initialized_size)119 static inline void handle_bounds_compressed_page(struct page *page,
120 		const loff_t i_size, const s64 initialized_size)
121 {
122 	if ((page->index >= (initialized_size >> PAGE_SHIFT)) &&
123 			(initialized_size < i_size))
124 		zero_partial_compressed_page(page, initialized_size);
125 	return;
126 }
127 
128 /**
129  * ntfs_decompress - decompress a compression block into an array of pages
130  * @dest_pages:		destination array of pages
131  * @completed_pages:	scratch space to track completed pages
132  * @dest_index:		current index into @dest_pages (IN/OUT)
133  * @dest_ofs:		current offset within @dest_pages[@dest_index] (IN/OUT)
134  * @dest_max_index:	maximum index into @dest_pages (IN)
135  * @dest_max_ofs:	maximum offset within @dest_pages[@dest_max_index] (IN)
136  * @xpage:		the target page (-1 if none) (IN)
137  * @xpage_done:		set to 1 if xpage was completed successfully (IN/OUT)
138  * @cb_start:		compression block to decompress (IN)
139  * @cb_size:		size of compression block @cb_start in bytes (IN)
140  * @i_size:		file size when we started the read (IN)
141  * @initialized_size:	initialized file size when we started the read (IN)
142  *
143  * The caller must have disabled preemption. ntfs_decompress() reenables it when
144  * the critical section is finished.
145  *
146  * This decompresses the compression block @cb_start into the array of
147  * destination pages @dest_pages starting at index @dest_index into @dest_pages
148  * and at offset @dest_pos into the page @dest_pages[@dest_index].
149  *
150  * When the page @dest_pages[@xpage] is completed, @xpage_done is set to 1.
151  * If xpage is -1 or @xpage has not been completed, @xpage_done is not modified.
152  *
153  * @cb_start is a pointer to the compression block which needs decompressing
154  * and @cb_size is the size of @cb_start in bytes (8-64kiB).
155  *
156  * Return 0 if success or -EOVERFLOW on error in the compressed stream.
157  * @xpage_done indicates whether the target page (@dest_pages[@xpage]) was
158  * completed during the decompression of the compression block (@cb_start).
159  *
160  * Warning: This function *REQUIRES* PAGE_SIZE >= 4096 or it will blow up
161  * unpredicatbly! You have been warned!
162  *
163  * Note to hackers: This function may not sleep until it has finished accessing
164  * the compression block @cb_start as it is a per-CPU buffer.
165  */
ntfs_decompress(struct page * dest_pages[],int completed_pages[],int * dest_index,int * dest_ofs,const int dest_max_index,const int dest_max_ofs,const int xpage,char * xpage_done,u8 * const cb_start,const u32 cb_size,const loff_t i_size,const s64 initialized_size)166 static int ntfs_decompress(struct page *dest_pages[], int completed_pages[],
167 		int *dest_index, int *dest_ofs, const int dest_max_index,
168 		const int dest_max_ofs, const int xpage, char *xpage_done,
169 		u8 *const cb_start, const u32 cb_size, const loff_t i_size,
170 		const s64 initialized_size)
171 {
172 	/*
173 	 * Pointers into the compressed data, i.e. the compression block (cb),
174 	 * and the therein contained sub-blocks (sb).
175 	 */
176 	u8 *cb_end = cb_start + cb_size; /* End of cb. */
177 	u8 *cb = cb_start;	/* Current position in cb. */
178 	u8 *cb_sb_start = cb;	/* Beginning of the current sb in the cb. */
179 	u8 *cb_sb_end;		/* End of current sb / beginning of next sb. */
180 
181 	/* Variables for uncompressed data / destination. */
182 	struct page *dp;	/* Current destination page being worked on. */
183 	u8 *dp_addr;		/* Current pointer into dp. */
184 	u8 *dp_sb_start;	/* Start of current sub-block in dp. */
185 	u8 *dp_sb_end;		/* End of current sb in dp (dp_sb_start +
186 				   NTFS_SB_SIZE). */
187 	u16 do_sb_start;	/* @dest_ofs when starting this sub-block. */
188 	u16 do_sb_end;		/* @dest_ofs of end of this sb (do_sb_start +
189 				   NTFS_SB_SIZE). */
190 
191 	/* Variables for tag and token parsing. */
192 	u8 tag;			/* Current tag. */
193 	int token;		/* Loop counter for the eight tokens in tag. */
194 	int nr_completed_pages = 0;
195 
196 	/* Default error code. */
197 	int err = -EOVERFLOW;
198 
199 	ntfs_debug("Entering, cb_size = 0x%x.", cb_size);
200 do_next_sb:
201 	ntfs_debug("Beginning sub-block at offset = 0x%zx in the cb.",
202 			cb - cb_start);
203 	/*
204 	 * Have we reached the end of the compression block or the end of the
205 	 * decompressed data?  The latter can happen for example if the current
206 	 * position in the compression block is one byte before its end so the
207 	 * first two checks do not detect it.
208 	 */
209 	if (cb == cb_end || !le16_to_cpup((le16*)cb) ||
210 			(*dest_index == dest_max_index &&
211 			*dest_ofs == dest_max_ofs)) {
212 		int i;
213 
214 		ntfs_debug("Completed. Returning success (0).");
215 		err = 0;
216 return_error:
217 		/* We can sleep from now on, so we drop lock. */
218 		spin_unlock(&ntfs_cb_lock);
219 		/* Second stage: finalize completed pages. */
220 		if (nr_completed_pages > 0) {
221 			for (i = 0; i < nr_completed_pages; i++) {
222 				int di = completed_pages[i];
223 
224 				dp = dest_pages[di];
225 				/*
226 				 * If we are outside the initialized size, zero
227 				 * the out of bounds page range.
228 				 */
229 				handle_bounds_compressed_page(dp, i_size,
230 						initialized_size);
231 				flush_dcache_page(dp);
232 				kunmap(dp);
233 				SetPageUptodate(dp);
234 				unlock_page(dp);
235 				if (di == xpage)
236 					*xpage_done = 1;
237 				else
238 					put_page(dp);
239 				dest_pages[di] = NULL;
240 			}
241 		}
242 		return err;
243 	}
244 
245 	/* Setup offsets for the current sub-block destination. */
246 	do_sb_start = *dest_ofs;
247 	do_sb_end = do_sb_start + NTFS_SB_SIZE;
248 
249 	/* Check that we are still within allowed boundaries. */
250 	if (*dest_index == dest_max_index && do_sb_end > dest_max_ofs)
251 		goto return_overflow;
252 
253 	/* Does the minimum size of a compressed sb overflow valid range? */
254 	if (cb + 6 > cb_end)
255 		goto return_overflow;
256 
257 	/* Setup the current sub-block source pointers and validate range. */
258 	cb_sb_start = cb;
259 	cb_sb_end = cb_sb_start + (le16_to_cpup((le16*)cb) & NTFS_SB_SIZE_MASK)
260 			+ 3;
261 	if (cb_sb_end > cb_end)
262 		goto return_overflow;
263 
264 	/* Get the current destination page. */
265 	dp = dest_pages[*dest_index];
266 	if (!dp) {
267 		/* No page present. Skip decompression of this sub-block. */
268 		cb = cb_sb_end;
269 
270 		/* Advance destination position to next sub-block. */
271 		*dest_ofs = (*dest_ofs + NTFS_SB_SIZE) & ~PAGE_MASK;
272 		if (!*dest_ofs && (++*dest_index > dest_max_index))
273 			goto return_overflow;
274 		goto do_next_sb;
275 	}
276 
277 	/* We have a valid destination page. Setup the destination pointers. */
278 	dp_addr = (u8*)page_address(dp) + do_sb_start;
279 
280 	/* Now, we are ready to process the current sub-block (sb). */
281 	if (!(le16_to_cpup((le16*)cb) & NTFS_SB_IS_COMPRESSED)) {
282 		ntfs_debug("Found uncompressed sub-block.");
283 		/* This sb is not compressed, just copy it into destination. */
284 
285 		/* Advance source position to first data byte. */
286 		cb += 2;
287 
288 		/* An uncompressed sb must be full size. */
289 		if (cb_sb_end - cb != NTFS_SB_SIZE)
290 			goto return_overflow;
291 
292 		/* Copy the block and advance the source position. */
293 		memcpy(dp_addr, cb, NTFS_SB_SIZE);
294 		cb += NTFS_SB_SIZE;
295 
296 		/* Advance destination position to next sub-block. */
297 		*dest_ofs += NTFS_SB_SIZE;
298 		if (!(*dest_ofs &= ~PAGE_MASK)) {
299 finalize_page:
300 			/*
301 			 * First stage: add current page index to array of
302 			 * completed pages.
303 			 */
304 			completed_pages[nr_completed_pages++] = *dest_index;
305 			if (++*dest_index > dest_max_index)
306 				goto return_overflow;
307 		}
308 		goto do_next_sb;
309 	}
310 	ntfs_debug("Found compressed sub-block.");
311 	/* This sb is compressed, decompress it into destination. */
312 
313 	/* Setup destination pointers. */
314 	dp_sb_start = dp_addr;
315 	dp_sb_end = dp_sb_start + NTFS_SB_SIZE;
316 
317 	/* Forward to the first tag in the sub-block. */
318 	cb += 2;
319 do_next_tag:
320 	if (cb == cb_sb_end) {
321 		/* Check if the decompressed sub-block was not full-length. */
322 		if (dp_addr < dp_sb_end) {
323 			int nr_bytes = do_sb_end - *dest_ofs;
324 
325 			ntfs_debug("Filling incomplete sub-block with "
326 					"zeroes.");
327 			/* Zero remainder and update destination position. */
328 			memset(dp_addr, 0, nr_bytes);
329 			*dest_ofs += nr_bytes;
330 		}
331 		/* We have finished the current sub-block. */
332 		if (!(*dest_ofs &= ~PAGE_MASK))
333 			goto finalize_page;
334 		goto do_next_sb;
335 	}
336 
337 	/* Check we are still in range. */
338 	if (cb > cb_sb_end || dp_addr > dp_sb_end)
339 		goto return_overflow;
340 
341 	/* Get the next tag and advance to first token. */
342 	tag = *cb++;
343 
344 	/* Parse the eight tokens described by the tag. */
345 	for (token = 0; token < 8; token++, tag >>= 1) {
346 		u16 lg, pt, length, max_non_overlap;
347 		register u16 i;
348 		u8 *dp_back_addr;
349 
350 		/* Check if we are done / still in range. */
351 		if (cb >= cb_sb_end || dp_addr > dp_sb_end)
352 			break;
353 
354 		/* Determine token type and parse appropriately.*/
355 		if ((tag & NTFS_TOKEN_MASK) == NTFS_SYMBOL_TOKEN) {
356 			/*
357 			 * We have a symbol token, copy the symbol across, and
358 			 * advance the source and destination positions.
359 			 */
360 			*dp_addr++ = *cb++;
361 			++*dest_ofs;
362 
363 			/* Continue with the next token. */
364 			continue;
365 		}
366 
367 		/*
368 		 * We have a phrase token. Make sure it is not the first tag in
369 		 * the sb as this is illegal and would confuse the code below.
370 		 */
371 		if (dp_addr == dp_sb_start)
372 			goto return_overflow;
373 
374 		/*
375 		 * Determine the number of bytes to go back (p) and the number
376 		 * of bytes to copy (l). We use an optimized algorithm in which
377 		 * we first calculate log2(current destination position in sb),
378 		 * which allows determination of l and p in O(1) rather than
379 		 * O(n). We just need an arch-optimized log2() function now.
380 		 */
381 		lg = 0;
382 		for (i = *dest_ofs - do_sb_start - 1; i >= 0x10; i >>= 1)
383 			lg++;
384 
385 		/* Get the phrase token into i. */
386 		pt = le16_to_cpup((le16*)cb);
387 
388 		/*
389 		 * Calculate starting position of the byte sequence in
390 		 * the destination using the fact that p = (pt >> (12 - lg)) + 1
391 		 * and make sure we don't go too far back.
392 		 */
393 		dp_back_addr = dp_addr - (pt >> (12 - lg)) - 1;
394 		if (dp_back_addr < dp_sb_start)
395 			goto return_overflow;
396 
397 		/* Now calculate the length of the byte sequence. */
398 		length = (pt & (0xfff >> lg)) + 3;
399 
400 		/* Advance destination position and verify it is in range. */
401 		*dest_ofs += length;
402 		if (*dest_ofs > do_sb_end)
403 			goto return_overflow;
404 
405 		/* The number of non-overlapping bytes. */
406 		max_non_overlap = dp_addr - dp_back_addr;
407 
408 		if (length <= max_non_overlap) {
409 			/* The byte sequence doesn't overlap, just copy it. */
410 			memcpy(dp_addr, dp_back_addr, length);
411 
412 			/* Advance destination pointer. */
413 			dp_addr += length;
414 		} else {
415 			/*
416 			 * The byte sequence does overlap, copy non-overlapping
417 			 * part and then do a slow byte by byte copy for the
418 			 * overlapping part. Also, advance the destination
419 			 * pointer.
420 			 */
421 			memcpy(dp_addr, dp_back_addr, max_non_overlap);
422 			dp_addr += max_non_overlap;
423 			dp_back_addr += max_non_overlap;
424 			length -= max_non_overlap;
425 			while (length--)
426 				*dp_addr++ = *dp_back_addr++;
427 		}
428 
429 		/* Advance source position and continue with the next token. */
430 		cb += 2;
431 	}
432 
433 	/* No tokens left in the current tag. Continue with the next tag. */
434 	goto do_next_tag;
435 
436 return_overflow:
437 	ntfs_error(NULL, "Failed. Returning -EOVERFLOW.");
438 	goto return_error;
439 }
440 
441 /**
442  * ntfs_read_compressed_block - read a compressed block into the page cache
443  * @page:	locked page in the compression block(s) we need to read
444  *
445  * When we are called the page has already been verified to be locked and the
446  * attribute is known to be non-resident, not encrypted, but compressed.
447  *
448  * 1. Determine which compression block(s) @page is in.
449  * 2. Get hold of all pages corresponding to this/these compression block(s).
450  * 3. Read the (first) compression block.
451  * 4. Decompress it into the corresponding pages.
452  * 5. Throw the compressed data away and proceed to 3. for the next compression
453  *    block or return success if no more compression blocks left.
454  *
455  * Warning: We have to be careful what we do about existing pages. They might
456  * have been written to so that we would lose data if we were to just overwrite
457  * them with the out-of-date uncompressed data.
458  *
459  * FIXME: For PAGE_SIZE > cb_size we are not doing the Right Thing(TM) at
460  * the end of the file I think. We need to detect this case and zero the out
461  * of bounds remainder of the page in question and mark it as handled. At the
462  * moment we would just return -EIO on such a page. This bug will only become
463  * apparent if pages are above 8kiB and the NTFS volume only uses 512 byte
464  * clusters so is probably not going to be seen by anyone. Still this should
465  * be fixed. (AIA)
466  *
467  * FIXME: Again for PAGE_SIZE > cb_size we are screwing up both in
468  * handling sparse and compressed cbs. (AIA)
469  *
470  * FIXME: At the moment we don't do any zeroing out in the case that
471  * initialized_size is less than data_size. This should be safe because of the
472  * nature of the compression algorithm used. Just in case we check and output
473  * an error message in read inode if the two sizes are not equal for a
474  * compressed file. (AIA)
475  */
ntfs_read_compressed_block(struct page * page)476 int ntfs_read_compressed_block(struct page *page)
477 {
478 	loff_t i_size;
479 	s64 initialized_size;
480 	struct address_space *mapping = page->mapping;
481 	ntfs_inode *ni = NTFS_I(mapping->host);
482 	ntfs_volume *vol = ni->vol;
483 	struct super_block *sb = vol->sb;
484 	runlist_element *rl;
485 	unsigned long flags, block_size = sb->s_blocksize;
486 	unsigned char block_size_bits = sb->s_blocksize_bits;
487 	u8 *cb, *cb_pos, *cb_end;
488 	struct buffer_head **bhs;
489 	unsigned long offset, index = page->index;
490 	u32 cb_size = ni->itype.compressed.block_size;
491 	u64 cb_size_mask = cb_size - 1UL;
492 	VCN vcn;
493 	LCN lcn;
494 	/* The first wanted vcn (minimum alignment is PAGE_SIZE). */
495 	VCN start_vcn = (((s64)index << PAGE_SHIFT) & ~cb_size_mask) >>
496 			vol->cluster_size_bits;
497 	/*
498 	 * The first vcn after the last wanted vcn (minimum alignment is again
499 	 * PAGE_SIZE.
500 	 */
501 	VCN end_vcn = ((((s64)(index + 1UL) << PAGE_SHIFT) + cb_size - 1)
502 			& ~cb_size_mask) >> vol->cluster_size_bits;
503 	/* Number of compression blocks (cbs) in the wanted vcn range. */
504 	unsigned int nr_cbs = (end_vcn - start_vcn) << vol->cluster_size_bits
505 			>> ni->itype.compressed.block_size_bits;
506 	/*
507 	 * Number of pages required to store the uncompressed data from all
508 	 * compression blocks (cbs) overlapping @page. Due to alignment
509 	 * guarantees of start_vcn and end_vcn, no need to round up here.
510 	 */
511 	unsigned int nr_pages = (end_vcn - start_vcn) <<
512 			vol->cluster_size_bits >> PAGE_SHIFT;
513 	unsigned int xpage, max_page, cur_page, cur_ofs, i;
514 	unsigned int cb_clusters, cb_max_ofs;
515 	int block, max_block, cb_max_page, bhs_size, nr_bhs, err = 0;
516 	struct page **pages;
517 	int *completed_pages;
518 	unsigned char xpage_done = 0;
519 
520 	ntfs_debug("Entering, page->index = 0x%lx, cb_size = 0x%x, nr_pages = "
521 			"%i.", index, cb_size, nr_pages);
522 	/*
523 	 * Bad things happen if we get here for anything that is not an
524 	 * unnamed $DATA attribute.
525 	 */
526 	BUG_ON(ni->type != AT_DATA);
527 	BUG_ON(ni->name_len);
528 
529 	pages = kmalloc_array(nr_pages, sizeof(struct page *), GFP_NOFS);
530 	completed_pages = kmalloc_array(nr_pages + 1, sizeof(int), GFP_NOFS);
531 
532 	/* Allocate memory to store the buffer heads we need. */
533 	bhs_size = cb_size / block_size * sizeof(struct buffer_head *);
534 	bhs = kmalloc(bhs_size, GFP_NOFS);
535 
536 	if (unlikely(!pages || !bhs || !completed_pages)) {
537 		kfree(bhs);
538 		kfree(pages);
539 		kfree(completed_pages);
540 		unlock_page(page);
541 		ntfs_error(vol->sb, "Failed to allocate internal buffers.");
542 		return -ENOMEM;
543 	}
544 
545 	/*
546 	 * We have already been given one page, this is the one we must do.
547 	 * Once again, the alignment guarantees keep it simple.
548 	 */
549 	offset = start_vcn << vol->cluster_size_bits >> PAGE_SHIFT;
550 	xpage = index - offset;
551 	pages[xpage] = page;
552 	/*
553 	 * The remaining pages need to be allocated and inserted into the page
554 	 * cache, alignment guarantees keep all the below much simpler. (-8
555 	 */
556 	read_lock_irqsave(&ni->size_lock, flags);
557 	i_size = i_size_read(VFS_I(ni));
558 	initialized_size = ni->initialized_size;
559 	read_unlock_irqrestore(&ni->size_lock, flags);
560 	max_page = ((i_size + PAGE_SIZE - 1) >> PAGE_SHIFT) -
561 			offset;
562 	/* Is the page fully outside i_size? (truncate in progress) */
563 	if (xpage >= max_page) {
564 		kfree(bhs);
565 		kfree(pages);
566 		kfree(completed_pages);
567 		zero_user(page, 0, PAGE_SIZE);
568 		ntfs_debug("Compressed read outside i_size - truncated?");
569 		SetPageUptodate(page);
570 		unlock_page(page);
571 		return 0;
572 	}
573 	if (nr_pages < max_page)
574 		max_page = nr_pages;
575 	for (i = 0; i < max_page; i++, offset++) {
576 		if (i != xpage)
577 			pages[i] = grab_cache_page_nowait(mapping, offset);
578 		page = pages[i];
579 		if (page) {
580 			/*
581 			 * We only (re)read the page if it isn't already read
582 			 * in and/or dirty or we would be losing data or at
583 			 * least wasting our time.
584 			 */
585 			if (!PageDirty(page) && (!PageUptodate(page) ||
586 					PageError(page))) {
587 				ClearPageError(page);
588 				kmap(page);
589 				continue;
590 			}
591 			unlock_page(page);
592 			put_page(page);
593 			pages[i] = NULL;
594 		}
595 	}
596 
597 	/*
598 	 * We have the runlist, and all the destination pages we need to fill.
599 	 * Now read the first compression block.
600 	 */
601 	cur_page = 0;
602 	cur_ofs = 0;
603 	cb_clusters = ni->itype.compressed.block_clusters;
604 do_next_cb:
605 	nr_cbs--;
606 	nr_bhs = 0;
607 
608 	/* Read all cb buffer heads one cluster at a time. */
609 	rl = NULL;
610 	for (vcn = start_vcn, start_vcn += cb_clusters; vcn < start_vcn;
611 			vcn++) {
612 		bool is_retry = false;
613 
614 		if (!rl) {
615 lock_retry_remap:
616 			down_read(&ni->runlist.lock);
617 			rl = ni->runlist.rl;
618 		}
619 		if (likely(rl != NULL)) {
620 			/* Seek to element containing target vcn. */
621 			while (rl->length && rl[1].vcn <= vcn)
622 				rl++;
623 			lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
624 		} else
625 			lcn = LCN_RL_NOT_MAPPED;
626 		ntfs_debug("Reading vcn = 0x%llx, lcn = 0x%llx.",
627 				(unsigned long long)vcn,
628 				(unsigned long long)lcn);
629 		if (lcn < 0) {
630 			/*
631 			 * When we reach the first sparse cluster we have
632 			 * finished with the cb.
633 			 */
634 			if (lcn == LCN_HOLE)
635 				break;
636 			if (is_retry || lcn != LCN_RL_NOT_MAPPED)
637 				goto rl_err;
638 			is_retry = true;
639 			/*
640 			 * Attempt to map runlist, dropping lock for the
641 			 * duration.
642 			 */
643 			up_read(&ni->runlist.lock);
644 			if (!ntfs_map_runlist(ni, vcn))
645 				goto lock_retry_remap;
646 			goto map_rl_err;
647 		}
648 		block = lcn << vol->cluster_size_bits >> block_size_bits;
649 		/* Read the lcn from device in chunks of block_size bytes. */
650 		max_block = block + (vol->cluster_size >> block_size_bits);
651 		do {
652 			ntfs_debug("block = 0x%x.", block);
653 			if (unlikely(!(bhs[nr_bhs] = sb_getblk(sb, block))))
654 				goto getblk_err;
655 			nr_bhs++;
656 		} while (++block < max_block);
657 	}
658 
659 	/* Release the lock if we took it. */
660 	if (rl)
661 		up_read(&ni->runlist.lock);
662 
663 	/* Setup and initiate io on all buffer heads. */
664 	for (i = 0; i < nr_bhs; i++) {
665 		struct buffer_head *tbh = bhs[i];
666 
667 		if (!trylock_buffer(tbh))
668 			continue;
669 		if (unlikely(buffer_uptodate(tbh))) {
670 			unlock_buffer(tbh);
671 			continue;
672 		}
673 		get_bh(tbh);
674 		tbh->b_end_io = end_buffer_read_sync;
675 		submit_bh(REQ_OP_READ, 0, tbh);
676 	}
677 
678 	/* Wait for io completion on all buffer heads. */
679 	for (i = 0; i < nr_bhs; i++) {
680 		struct buffer_head *tbh = bhs[i];
681 
682 		if (buffer_uptodate(tbh))
683 			continue;
684 		wait_on_buffer(tbh);
685 		/*
686 		 * We need an optimization barrier here, otherwise we start
687 		 * hitting the below fixup code when accessing a loopback
688 		 * mounted ntfs partition. This indicates either there is a
689 		 * race condition in the loop driver or, more likely, gcc
690 		 * overoptimises the code without the barrier and it doesn't
691 		 * do the Right Thing(TM).
692 		 */
693 		barrier();
694 		if (unlikely(!buffer_uptodate(tbh))) {
695 			ntfs_warning(vol->sb, "Buffer is unlocked but not "
696 					"uptodate! Unplugging the disk queue "
697 					"and rescheduling.");
698 			get_bh(tbh);
699 			io_schedule();
700 			put_bh(tbh);
701 			if (unlikely(!buffer_uptodate(tbh)))
702 				goto read_err;
703 			ntfs_warning(vol->sb, "Buffer is now uptodate. Good.");
704 		}
705 	}
706 
707 	/*
708 	 * Get the compression buffer. We must not sleep any more
709 	 * until we are finished with it.
710 	 */
711 	spin_lock(&ntfs_cb_lock);
712 	cb = ntfs_compression_buffer;
713 
714 	BUG_ON(!cb);
715 
716 	cb_pos = cb;
717 	cb_end = cb + cb_size;
718 
719 	/* Copy the buffer heads into the contiguous buffer. */
720 	for (i = 0; i < nr_bhs; i++) {
721 		memcpy(cb_pos, bhs[i]->b_data, block_size);
722 		cb_pos += block_size;
723 	}
724 
725 	/* Just a precaution. */
726 	if (cb_pos + 2 <= cb + cb_size)
727 		*(u16*)cb_pos = 0;
728 
729 	/* Reset cb_pos back to the beginning. */
730 	cb_pos = cb;
731 
732 	/* We now have both source (if present) and destination. */
733 	ntfs_debug("Successfully read the compression block.");
734 
735 	/* The last page and maximum offset within it for the current cb. */
736 	cb_max_page = (cur_page << PAGE_SHIFT) + cur_ofs + cb_size;
737 	cb_max_ofs = cb_max_page & ~PAGE_MASK;
738 	cb_max_page >>= PAGE_SHIFT;
739 
740 	/* Catch end of file inside a compression block. */
741 	if (cb_max_page > max_page)
742 		cb_max_page = max_page;
743 
744 	if (vcn == start_vcn - cb_clusters) {
745 		/* Sparse cb, zero out page range overlapping the cb. */
746 		ntfs_debug("Found sparse compression block.");
747 		/* We can sleep from now on, so we drop lock. */
748 		spin_unlock(&ntfs_cb_lock);
749 		if (cb_max_ofs)
750 			cb_max_page--;
751 		for (; cur_page < cb_max_page; cur_page++) {
752 			page = pages[cur_page];
753 			if (page) {
754 				if (likely(!cur_ofs))
755 					clear_page(page_address(page));
756 				else
757 					memset(page_address(page) + cur_ofs, 0,
758 							PAGE_SIZE -
759 							cur_ofs);
760 				flush_dcache_page(page);
761 				kunmap(page);
762 				SetPageUptodate(page);
763 				unlock_page(page);
764 				if (cur_page == xpage)
765 					xpage_done = 1;
766 				else
767 					put_page(page);
768 				pages[cur_page] = NULL;
769 			}
770 			cb_pos += PAGE_SIZE - cur_ofs;
771 			cur_ofs = 0;
772 			if (cb_pos >= cb_end)
773 				break;
774 		}
775 		/* If we have a partial final page, deal with it now. */
776 		if (cb_max_ofs && cb_pos < cb_end) {
777 			page = pages[cur_page];
778 			if (page)
779 				memset(page_address(page) + cur_ofs, 0,
780 						cb_max_ofs - cur_ofs);
781 			/*
782 			 * No need to update cb_pos at this stage:
783 			 *	cb_pos += cb_max_ofs - cur_ofs;
784 			 */
785 			cur_ofs = cb_max_ofs;
786 		}
787 	} else if (vcn == start_vcn) {
788 		/* We can't sleep so we need two stages. */
789 		unsigned int cur2_page = cur_page;
790 		unsigned int cur_ofs2 = cur_ofs;
791 		u8 *cb_pos2 = cb_pos;
792 
793 		ntfs_debug("Found uncompressed compression block.");
794 		/* Uncompressed cb, copy it to the destination pages. */
795 		/*
796 		 * TODO: As a big optimization, we could detect this case
797 		 * before we read all the pages and use block_read_full_page()
798 		 * on all full pages instead (we still have to treat partial
799 		 * pages especially but at least we are getting rid of the
800 		 * synchronous io for the majority of pages.
801 		 * Or if we choose not to do the read-ahead/-behind stuff, we
802 		 * could just return block_read_full_page(pages[xpage]) as long
803 		 * as PAGE_SIZE <= cb_size.
804 		 */
805 		if (cb_max_ofs)
806 			cb_max_page--;
807 		/* First stage: copy data into destination pages. */
808 		for (; cur_page < cb_max_page; cur_page++) {
809 			page = pages[cur_page];
810 			if (page)
811 				memcpy(page_address(page) + cur_ofs, cb_pos,
812 						PAGE_SIZE - cur_ofs);
813 			cb_pos += PAGE_SIZE - cur_ofs;
814 			cur_ofs = 0;
815 			if (cb_pos >= cb_end)
816 				break;
817 		}
818 		/* If we have a partial final page, deal with it now. */
819 		if (cb_max_ofs && cb_pos < cb_end) {
820 			page = pages[cur_page];
821 			if (page)
822 				memcpy(page_address(page) + cur_ofs, cb_pos,
823 						cb_max_ofs - cur_ofs);
824 			cb_pos += cb_max_ofs - cur_ofs;
825 			cur_ofs = cb_max_ofs;
826 		}
827 		/* We can sleep from now on, so drop lock. */
828 		spin_unlock(&ntfs_cb_lock);
829 		/* Second stage: finalize pages. */
830 		for (; cur2_page < cb_max_page; cur2_page++) {
831 			page = pages[cur2_page];
832 			if (page) {
833 				/*
834 				 * If we are outside the initialized size, zero
835 				 * the out of bounds page range.
836 				 */
837 				handle_bounds_compressed_page(page, i_size,
838 						initialized_size);
839 				flush_dcache_page(page);
840 				kunmap(page);
841 				SetPageUptodate(page);
842 				unlock_page(page);
843 				if (cur2_page == xpage)
844 					xpage_done = 1;
845 				else
846 					put_page(page);
847 				pages[cur2_page] = NULL;
848 			}
849 			cb_pos2 += PAGE_SIZE - cur_ofs2;
850 			cur_ofs2 = 0;
851 			if (cb_pos2 >= cb_end)
852 				break;
853 		}
854 	} else {
855 		/* Compressed cb, decompress it into the destination page(s). */
856 		unsigned int prev_cur_page = cur_page;
857 
858 		ntfs_debug("Found compressed compression block.");
859 		err = ntfs_decompress(pages, completed_pages, &cur_page,
860 				&cur_ofs, cb_max_page, cb_max_ofs, xpage,
861 				&xpage_done, cb_pos, cb_size - (cb_pos - cb),
862 				i_size, initialized_size);
863 		/*
864 		 * We can sleep from now on, lock already dropped by
865 		 * ntfs_decompress().
866 		 */
867 		if (err) {
868 			ntfs_error(vol->sb, "ntfs_decompress() failed in inode "
869 					"0x%lx with error code %i. Skipping "
870 					"this compression block.",
871 					ni->mft_no, -err);
872 			/* Release the unfinished pages. */
873 			for (; prev_cur_page < cur_page; prev_cur_page++) {
874 				page = pages[prev_cur_page];
875 				if (page) {
876 					flush_dcache_page(page);
877 					kunmap(page);
878 					unlock_page(page);
879 					if (prev_cur_page != xpage)
880 						put_page(page);
881 					pages[prev_cur_page] = NULL;
882 				}
883 			}
884 		}
885 	}
886 
887 	/* Release the buffer heads. */
888 	for (i = 0; i < nr_bhs; i++)
889 		brelse(bhs[i]);
890 
891 	/* Do we have more work to do? */
892 	if (nr_cbs)
893 		goto do_next_cb;
894 
895 	/* We no longer need the list of buffer heads. */
896 	kfree(bhs);
897 
898 	/* Clean up if we have any pages left. Should never happen. */
899 	for (cur_page = 0; cur_page < max_page; cur_page++) {
900 		page = pages[cur_page];
901 		if (page) {
902 			ntfs_error(vol->sb, "Still have pages left! "
903 					"Terminating them with extreme "
904 					"prejudice.  Inode 0x%lx, page index "
905 					"0x%lx.", ni->mft_no, page->index);
906 			flush_dcache_page(page);
907 			kunmap(page);
908 			unlock_page(page);
909 			if (cur_page != xpage)
910 				put_page(page);
911 			pages[cur_page] = NULL;
912 		}
913 	}
914 
915 	/* We no longer need the list of pages. */
916 	kfree(pages);
917 	kfree(completed_pages);
918 
919 	/* If we have completed the requested page, we return success. */
920 	if (likely(xpage_done))
921 		return 0;
922 
923 	ntfs_debug("Failed. Returning error code %s.", err == -EOVERFLOW ?
924 			"EOVERFLOW" : (!err ? "EIO" : "unknown error"));
925 	return err < 0 ? err : -EIO;
926 
927 read_err:
928 	ntfs_error(vol->sb, "IO error while reading compressed data.");
929 	/* Release the buffer heads. */
930 	for (i = 0; i < nr_bhs; i++)
931 		brelse(bhs[i]);
932 	goto err_out;
933 
934 map_rl_err:
935 	ntfs_error(vol->sb, "ntfs_map_runlist() failed. Cannot read "
936 			"compression block.");
937 	goto err_out;
938 
939 rl_err:
940 	up_read(&ni->runlist.lock);
941 	ntfs_error(vol->sb, "ntfs_rl_vcn_to_lcn() failed. Cannot read "
942 			"compression block.");
943 	goto err_out;
944 
945 getblk_err:
946 	up_read(&ni->runlist.lock);
947 	ntfs_error(vol->sb, "getblk() failed. Cannot read compression block.");
948 
949 err_out:
950 	kfree(bhs);
951 	for (i = cur_page; i < max_page; i++) {
952 		page = pages[i];
953 		if (page) {
954 			flush_dcache_page(page);
955 			kunmap(page);
956 			unlock_page(page);
957 			if (i != xpage)
958 				put_page(page);
959 		}
960 	}
961 	kfree(pages);
962 	kfree(completed_pages);
963 	return -EIO;
964 }
965