xref: /openbmc/linux/fs/ntfs/compress.c (revision 565d76cb)
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 = NULL;
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  */
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  */
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  */
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_CACHE_SHIFT) >= initialized_size) {
108 		/*
109 		 * FIXME: Using clear_page() will become wrong when we get
110 		 * PAGE_CACHE_SIZE != PAGE_SIZE but for now there is no problem.
111 		 */
112 		clear_page(kp);
113 		return;
114 	}
115 	kp_ofs = initialized_size & ~PAGE_CACHE_MASK;
116 	memset(kp + kp_ofs, 0, PAGE_CACHE_SIZE - kp_ofs);
117 	return;
118 }
119 
120 /**
121  * handle_bounds_compressed_page - test for&handle out of bounds compressed page
122  */
123 static inline void handle_bounds_compressed_page(struct page *page,
124 		const loff_t i_size, const s64 initialized_size)
125 {
126 	if ((page->index >= (initialized_size >> PAGE_CACHE_SHIFT)) &&
127 			(initialized_size < i_size))
128 		zero_partial_compressed_page(page, initialized_size);
129 	return;
130 }
131 
132 /**
133  * ntfs_decompress - decompress a compression block into an array of pages
134  * @dest_pages:		destination array of pages
135  * @dest_index:		current index into @dest_pages (IN/OUT)
136  * @dest_ofs:		current offset within @dest_pages[@dest_index] (IN/OUT)
137  * @dest_max_index:	maximum index into @dest_pages (IN)
138  * @dest_max_ofs:	maximum offset within @dest_pages[@dest_max_index] (IN)
139  * @xpage:		the target page (-1 if none) (IN)
140  * @xpage_done:		set to 1 if xpage was completed successfully (IN/OUT)
141  * @cb_start:		compression block to decompress (IN)
142  * @cb_size:		size of compression block @cb_start in bytes (IN)
143  * @i_size:		file size when we started the read (IN)
144  * @initialized_size:	initialized file size when we started the read (IN)
145  *
146  * The caller must have disabled preemption. ntfs_decompress() reenables it when
147  * the critical section is finished.
148  *
149  * This decompresses the compression block @cb_start into the array of
150  * destination pages @dest_pages starting at index @dest_index into @dest_pages
151  * and at offset @dest_pos into the page @dest_pages[@dest_index].
152  *
153  * When the page @dest_pages[@xpage] is completed, @xpage_done is set to 1.
154  * If xpage is -1 or @xpage has not been completed, @xpage_done is not modified.
155  *
156  * @cb_start is a pointer to the compression block which needs decompressing
157  * and @cb_size is the size of @cb_start in bytes (8-64kiB).
158  *
159  * Return 0 if success or -EOVERFLOW on error in the compressed stream.
160  * @xpage_done indicates whether the target page (@dest_pages[@xpage]) was
161  * completed during the decompression of the compression block (@cb_start).
162  *
163  * Warning: This function *REQUIRES* PAGE_CACHE_SIZE >= 4096 or it will blow up
164  * unpredicatbly! You have been warned!
165  *
166  * Note to hackers: This function may not sleep until it has finished accessing
167  * the compression block @cb_start as it is a per-CPU buffer.
168  */
169 static int ntfs_decompress(struct page *dest_pages[], int *dest_index,
170 		int *dest_ofs, const int dest_max_index, const int dest_max_ofs,
171 		const int xpage, char *xpage_done, u8 *const cb_start,
172 		const u32 cb_size, const loff_t i_size,
173 		const s64 initialized_size)
174 {
175 	/*
176 	 * Pointers into the compressed data, i.e. the compression block (cb),
177 	 * and the therein contained sub-blocks (sb).
178 	 */
179 	u8 *cb_end = cb_start + cb_size; /* End of cb. */
180 	u8 *cb = cb_start;	/* Current position in cb. */
181 	u8 *cb_sb_start = cb;	/* Beginning of the current sb in the cb. */
182 	u8 *cb_sb_end;		/* End of current sb / beginning of next sb. */
183 
184 	/* Variables for uncompressed data / destination. */
185 	struct page *dp;	/* Current destination page being worked on. */
186 	u8 *dp_addr;		/* Current pointer into dp. */
187 	u8 *dp_sb_start;	/* Start of current sub-block in dp. */
188 	u8 *dp_sb_end;		/* End of current sb in dp (dp_sb_start +
189 				   NTFS_SB_SIZE). */
190 	u16 do_sb_start;	/* @dest_ofs when starting this sub-block. */
191 	u16 do_sb_end;		/* @dest_ofs of end of this sb (do_sb_start +
192 				   NTFS_SB_SIZE). */
193 
194 	/* Variables for tag and token parsing. */
195 	u8 tag;			/* Current tag. */
196 	int token;		/* Loop counter for the eight tokens in tag. */
197 
198 	/* Need this because we can't sleep, so need two stages. */
199 	int completed_pages[dest_max_index - *dest_index + 1];
200 	int nr_completed_pages = 0;
201 
202 	/* Default error code. */
203 	int err = -EOVERFLOW;
204 
205 	ntfs_debug("Entering, cb_size = 0x%x.", cb_size);
206 do_next_sb:
207 	ntfs_debug("Beginning sub-block at offset = 0x%zx in the cb.",
208 			cb - cb_start);
209 	/*
210 	 * Have we reached the end of the compression block or the end of the
211 	 * decompressed data?  The latter can happen for example if the current
212 	 * position in the compression block is one byte before its end so the
213 	 * first two checks do not detect it.
214 	 */
215 	if (cb == cb_end || !le16_to_cpup((le16*)cb) ||
216 			(*dest_index == dest_max_index &&
217 			*dest_ofs == dest_max_ofs)) {
218 		int i;
219 
220 		ntfs_debug("Completed. Returning success (0).");
221 		err = 0;
222 return_error:
223 		/* We can sleep from now on, so we drop lock. */
224 		spin_unlock(&ntfs_cb_lock);
225 		/* Second stage: finalize completed pages. */
226 		if (nr_completed_pages > 0) {
227 			for (i = 0; i < nr_completed_pages; i++) {
228 				int di = completed_pages[i];
229 
230 				dp = dest_pages[di];
231 				/*
232 				 * If we are outside the initialized size, zero
233 				 * the out of bounds page range.
234 				 */
235 				handle_bounds_compressed_page(dp, i_size,
236 						initialized_size);
237 				flush_dcache_page(dp);
238 				kunmap(dp);
239 				SetPageUptodate(dp);
240 				unlock_page(dp);
241 				if (di == xpage)
242 					*xpage_done = 1;
243 				else
244 					page_cache_release(dp);
245 				dest_pages[di] = NULL;
246 			}
247 		}
248 		return err;
249 	}
250 
251 	/* Setup offsets for the current sub-block destination. */
252 	do_sb_start = *dest_ofs;
253 	do_sb_end = do_sb_start + NTFS_SB_SIZE;
254 
255 	/* Check that we are still within allowed boundaries. */
256 	if (*dest_index == dest_max_index && do_sb_end > dest_max_ofs)
257 		goto return_overflow;
258 
259 	/* Does the minimum size of a compressed sb overflow valid range? */
260 	if (cb + 6 > cb_end)
261 		goto return_overflow;
262 
263 	/* Setup the current sub-block source pointers and validate range. */
264 	cb_sb_start = cb;
265 	cb_sb_end = cb_sb_start + (le16_to_cpup((le16*)cb) & NTFS_SB_SIZE_MASK)
266 			+ 3;
267 	if (cb_sb_end > cb_end)
268 		goto return_overflow;
269 
270 	/* Get the current destination page. */
271 	dp = dest_pages[*dest_index];
272 	if (!dp) {
273 		/* No page present. Skip decompression of this sub-block. */
274 		cb = cb_sb_end;
275 
276 		/* Advance destination position to next sub-block. */
277 		*dest_ofs = (*dest_ofs + NTFS_SB_SIZE) & ~PAGE_CACHE_MASK;
278 		if (!*dest_ofs && (++*dest_index > dest_max_index))
279 			goto return_overflow;
280 		goto do_next_sb;
281 	}
282 
283 	/* We have a valid destination page. Setup the destination pointers. */
284 	dp_addr = (u8*)page_address(dp) + do_sb_start;
285 
286 	/* Now, we are ready to process the current sub-block (sb). */
287 	if (!(le16_to_cpup((le16*)cb) & NTFS_SB_IS_COMPRESSED)) {
288 		ntfs_debug("Found uncompressed sub-block.");
289 		/* This sb is not compressed, just copy it into destination. */
290 
291 		/* Advance source position to first data byte. */
292 		cb += 2;
293 
294 		/* An uncompressed sb must be full size. */
295 		if (cb_sb_end - cb != NTFS_SB_SIZE)
296 			goto return_overflow;
297 
298 		/* Copy the block and advance the source position. */
299 		memcpy(dp_addr, cb, NTFS_SB_SIZE);
300 		cb += NTFS_SB_SIZE;
301 
302 		/* Advance destination position to next sub-block. */
303 		*dest_ofs += NTFS_SB_SIZE;
304 		if (!(*dest_ofs &= ~PAGE_CACHE_MASK)) {
305 finalize_page:
306 			/*
307 			 * First stage: add current page index to array of
308 			 * completed pages.
309 			 */
310 			completed_pages[nr_completed_pages++] = *dest_index;
311 			if (++*dest_index > dest_max_index)
312 				goto return_overflow;
313 		}
314 		goto do_next_sb;
315 	}
316 	ntfs_debug("Found compressed sub-block.");
317 	/* This sb is compressed, decompress it into destination. */
318 
319 	/* Setup destination pointers. */
320 	dp_sb_start = dp_addr;
321 	dp_sb_end = dp_sb_start + NTFS_SB_SIZE;
322 
323 	/* Forward to the first tag in the sub-block. */
324 	cb += 2;
325 do_next_tag:
326 	if (cb == cb_sb_end) {
327 		/* Check if the decompressed sub-block was not full-length. */
328 		if (dp_addr < dp_sb_end) {
329 			int nr_bytes = do_sb_end - *dest_ofs;
330 
331 			ntfs_debug("Filling incomplete sub-block with "
332 					"zeroes.");
333 			/* Zero remainder and update destination position. */
334 			memset(dp_addr, 0, nr_bytes);
335 			*dest_ofs += nr_bytes;
336 		}
337 		/* We have finished the current sub-block. */
338 		if (!(*dest_ofs &= ~PAGE_CACHE_MASK))
339 			goto finalize_page;
340 		goto do_next_sb;
341 	}
342 
343 	/* Check we are still in range. */
344 	if (cb > cb_sb_end || dp_addr > dp_sb_end)
345 		goto return_overflow;
346 
347 	/* Get the next tag and advance to first token. */
348 	tag = *cb++;
349 
350 	/* Parse the eight tokens described by the tag. */
351 	for (token = 0; token < 8; token++, tag >>= 1) {
352 		u16 lg, pt, length, max_non_overlap;
353 		register u16 i;
354 		u8 *dp_back_addr;
355 
356 		/* Check if we are done / still in range. */
357 		if (cb >= cb_sb_end || dp_addr > dp_sb_end)
358 			break;
359 
360 		/* Determine token type and parse appropriately.*/
361 		if ((tag & NTFS_TOKEN_MASK) == NTFS_SYMBOL_TOKEN) {
362 			/*
363 			 * We have a symbol token, copy the symbol across, and
364 			 * advance the source and destination positions.
365 			 */
366 			*dp_addr++ = *cb++;
367 			++*dest_ofs;
368 
369 			/* Continue with the next token. */
370 			continue;
371 		}
372 
373 		/*
374 		 * We have a phrase token. Make sure it is not the first tag in
375 		 * the sb as this is illegal and would confuse the code below.
376 		 */
377 		if (dp_addr == dp_sb_start)
378 			goto return_overflow;
379 
380 		/*
381 		 * Determine the number of bytes to go back (p) and the number
382 		 * of bytes to copy (l). We use an optimized algorithm in which
383 		 * we first calculate log2(current destination position in sb),
384 		 * which allows determination of l and p in O(1) rather than
385 		 * O(n). We just need an arch-optimized log2() function now.
386 		 */
387 		lg = 0;
388 		for (i = *dest_ofs - do_sb_start - 1; i >= 0x10; i >>= 1)
389 			lg++;
390 
391 		/* Get the phrase token into i. */
392 		pt = le16_to_cpup((le16*)cb);
393 
394 		/*
395 		 * Calculate starting position of the byte sequence in
396 		 * the destination using the fact that p = (pt >> (12 - lg)) + 1
397 		 * and make sure we don't go too far back.
398 		 */
399 		dp_back_addr = dp_addr - (pt >> (12 - lg)) - 1;
400 		if (dp_back_addr < dp_sb_start)
401 			goto return_overflow;
402 
403 		/* Now calculate the length of the byte sequence. */
404 		length = (pt & (0xfff >> lg)) + 3;
405 
406 		/* Advance destination position and verify it is in range. */
407 		*dest_ofs += length;
408 		if (*dest_ofs > do_sb_end)
409 			goto return_overflow;
410 
411 		/* The number of non-overlapping bytes. */
412 		max_non_overlap = dp_addr - dp_back_addr;
413 
414 		if (length <= max_non_overlap) {
415 			/* The byte sequence doesn't overlap, just copy it. */
416 			memcpy(dp_addr, dp_back_addr, length);
417 
418 			/* Advance destination pointer. */
419 			dp_addr += length;
420 		} else {
421 			/*
422 			 * The byte sequence does overlap, copy non-overlapping
423 			 * part and then do a slow byte by byte copy for the
424 			 * overlapping part. Also, advance the destination
425 			 * pointer.
426 			 */
427 			memcpy(dp_addr, dp_back_addr, max_non_overlap);
428 			dp_addr += max_non_overlap;
429 			dp_back_addr += max_non_overlap;
430 			length -= max_non_overlap;
431 			while (length--)
432 				*dp_addr++ = *dp_back_addr++;
433 		}
434 
435 		/* Advance source position and continue with the next token. */
436 		cb += 2;
437 	}
438 
439 	/* No tokens left in the current tag. Continue with the next tag. */
440 	goto do_next_tag;
441 
442 return_overflow:
443 	ntfs_error(NULL, "Failed. Returning -EOVERFLOW.");
444 	goto return_error;
445 }
446 
447 /**
448  * ntfs_read_compressed_block - read a compressed block into the page cache
449  * @page:	locked page in the compression block(s) we need to read
450  *
451  * When we are called the page has already been verified to be locked and the
452  * attribute is known to be non-resident, not encrypted, but compressed.
453  *
454  * 1. Determine which compression block(s) @page is in.
455  * 2. Get hold of all pages corresponding to this/these compression block(s).
456  * 3. Read the (first) compression block.
457  * 4. Decompress it into the corresponding pages.
458  * 5. Throw the compressed data away and proceed to 3. for the next compression
459  *    block or return success if no more compression blocks left.
460  *
461  * Warning: We have to be careful what we do about existing pages. They might
462  * have been written to so that we would lose data if we were to just overwrite
463  * them with the out-of-date uncompressed data.
464  *
465  * FIXME: For PAGE_CACHE_SIZE > cb_size we are not doing the Right Thing(TM) at
466  * the end of the file I think. We need to detect this case and zero the out
467  * of bounds remainder of the page in question and mark it as handled. At the
468  * moment we would just return -EIO on such a page. This bug will only become
469  * apparent if pages are above 8kiB and the NTFS volume only uses 512 byte
470  * clusters so is probably not going to be seen by anyone. Still this should
471  * be fixed. (AIA)
472  *
473  * FIXME: Again for PAGE_CACHE_SIZE > cb_size we are screwing up both in
474  * handling sparse and compressed cbs. (AIA)
475  *
476  * FIXME: At the moment we don't do any zeroing out in the case that
477  * initialized_size is less than data_size. This should be safe because of the
478  * nature of the compression algorithm used. Just in case we check and output
479  * an error message in read inode if the two sizes are not equal for a
480  * compressed file. (AIA)
481  */
482 int ntfs_read_compressed_block(struct page *page)
483 {
484 	loff_t i_size;
485 	s64 initialized_size;
486 	struct address_space *mapping = page->mapping;
487 	ntfs_inode *ni = NTFS_I(mapping->host);
488 	ntfs_volume *vol = ni->vol;
489 	struct super_block *sb = vol->sb;
490 	runlist_element *rl;
491 	unsigned long flags, block_size = sb->s_blocksize;
492 	unsigned char block_size_bits = sb->s_blocksize_bits;
493 	u8 *cb, *cb_pos, *cb_end;
494 	struct buffer_head **bhs;
495 	unsigned long offset, index = page->index;
496 	u32 cb_size = ni->itype.compressed.block_size;
497 	u64 cb_size_mask = cb_size - 1UL;
498 	VCN vcn;
499 	LCN lcn;
500 	/* The first wanted vcn (minimum alignment is PAGE_CACHE_SIZE). */
501 	VCN start_vcn = (((s64)index << PAGE_CACHE_SHIFT) & ~cb_size_mask) >>
502 			vol->cluster_size_bits;
503 	/*
504 	 * The first vcn after the last wanted vcn (minumum alignment is again
505 	 * PAGE_CACHE_SIZE.
506 	 */
507 	VCN end_vcn = ((((s64)(index + 1UL) << PAGE_CACHE_SHIFT) + cb_size - 1)
508 			& ~cb_size_mask) >> vol->cluster_size_bits;
509 	/* Number of compression blocks (cbs) in the wanted vcn range. */
510 	unsigned int nr_cbs = (end_vcn - start_vcn) << vol->cluster_size_bits
511 			>> ni->itype.compressed.block_size_bits;
512 	/*
513 	 * Number of pages required to store the uncompressed data from all
514 	 * compression blocks (cbs) overlapping @page. Due to alignment
515 	 * guarantees of start_vcn and end_vcn, no need to round up here.
516 	 */
517 	unsigned int nr_pages = (end_vcn - start_vcn) <<
518 			vol->cluster_size_bits >> PAGE_CACHE_SHIFT;
519 	unsigned int xpage, max_page, cur_page, cur_ofs, i;
520 	unsigned int cb_clusters, cb_max_ofs;
521 	int block, max_block, cb_max_page, bhs_size, nr_bhs, err = 0;
522 	struct page **pages;
523 	unsigned char xpage_done = 0;
524 
525 	ntfs_debug("Entering, page->index = 0x%lx, cb_size = 0x%x, nr_pages = "
526 			"%i.", index, cb_size, nr_pages);
527 	/*
528 	 * Bad things happen if we get here for anything that is not an
529 	 * unnamed $DATA attribute.
530 	 */
531 	BUG_ON(ni->type != AT_DATA);
532 	BUG_ON(ni->name_len);
533 
534 	pages = kmalloc(nr_pages * sizeof(struct page *), GFP_NOFS);
535 
536 	/* Allocate memory to store the buffer heads we need. */
537 	bhs_size = cb_size / block_size * sizeof(struct buffer_head *);
538 	bhs = kmalloc(bhs_size, GFP_NOFS);
539 
540 	if (unlikely(!pages || !bhs)) {
541 		kfree(bhs);
542 		kfree(pages);
543 		unlock_page(page);
544 		ntfs_error(vol->sb, "Failed to allocate internal buffers.");
545 		return -ENOMEM;
546 	}
547 
548 	/*
549 	 * We have already been given one page, this is the one we must do.
550 	 * Once again, the alignment guarantees keep it simple.
551 	 */
552 	offset = start_vcn << vol->cluster_size_bits >> PAGE_CACHE_SHIFT;
553 	xpage = index - offset;
554 	pages[xpage] = page;
555 	/*
556 	 * The remaining pages need to be allocated and inserted into the page
557 	 * cache, alignment guarantees keep all the below much simpler. (-8
558 	 */
559 	read_lock_irqsave(&ni->size_lock, flags);
560 	i_size = i_size_read(VFS_I(ni));
561 	initialized_size = ni->initialized_size;
562 	read_unlock_irqrestore(&ni->size_lock, flags);
563 	max_page = ((i_size + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) -
564 			offset;
565 	/* Is the page fully outside i_size? (truncate in progress) */
566 	if (xpage >= max_page) {
567 		kfree(bhs);
568 		kfree(pages);
569 		zero_user(page, 0, PAGE_CACHE_SIZE);
570 		ntfs_debug("Compressed read outside i_size - truncated?");
571 		SetPageUptodate(page);
572 		unlock_page(page);
573 		return 0;
574 	}
575 	if (nr_pages < max_page)
576 		max_page = nr_pages;
577 	for (i = 0; i < max_page; i++, offset++) {
578 		if (i != xpage)
579 			pages[i] = grab_cache_page_nowait(mapping, offset);
580 		page = pages[i];
581 		if (page) {
582 			/*
583 			 * We only (re)read the page if it isn't already read
584 			 * in and/or dirty or we would be losing data or at
585 			 * least wasting our time.
586 			 */
587 			if (!PageDirty(page) && (!PageUptodate(page) ||
588 					PageError(page))) {
589 				ClearPageError(page);
590 				kmap(page);
591 				continue;
592 			}
593 			unlock_page(page);
594 			page_cache_release(page);
595 			pages[i] = NULL;
596 		}
597 	}
598 
599 	/*
600 	 * We have the runlist, and all the destination pages we need to fill.
601 	 * Now read the first compression block.
602 	 */
603 	cur_page = 0;
604 	cur_ofs = 0;
605 	cb_clusters = ni->itype.compressed.block_clusters;
606 do_next_cb:
607 	nr_cbs--;
608 	nr_bhs = 0;
609 
610 	/* Read all cb buffer heads one cluster at a time. */
611 	rl = NULL;
612 	for (vcn = start_vcn, start_vcn += cb_clusters; vcn < start_vcn;
613 			vcn++) {
614 		bool is_retry = false;
615 
616 		if (!rl) {
617 lock_retry_remap:
618 			down_read(&ni->runlist.lock);
619 			rl = ni->runlist.rl;
620 		}
621 		if (likely(rl != NULL)) {
622 			/* Seek to element containing target vcn. */
623 			while (rl->length && rl[1].vcn <= vcn)
624 				rl++;
625 			lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
626 		} else
627 			lcn = LCN_RL_NOT_MAPPED;
628 		ntfs_debug("Reading vcn = 0x%llx, lcn = 0x%llx.",
629 				(unsigned long long)vcn,
630 				(unsigned long long)lcn);
631 		if (lcn < 0) {
632 			/*
633 			 * When we reach the first sparse cluster we have
634 			 * finished with the cb.
635 			 */
636 			if (lcn == LCN_HOLE)
637 				break;
638 			if (is_retry || lcn != LCN_RL_NOT_MAPPED)
639 				goto rl_err;
640 			is_retry = true;
641 			/*
642 			 * Attempt to map runlist, dropping lock for the
643 			 * duration.
644 			 */
645 			up_read(&ni->runlist.lock);
646 			if (!ntfs_map_runlist(ni, vcn))
647 				goto lock_retry_remap;
648 			goto map_rl_err;
649 		}
650 		block = lcn << vol->cluster_size_bits >> block_size_bits;
651 		/* Read the lcn from device in chunks of block_size bytes. */
652 		max_block = block + (vol->cluster_size >> block_size_bits);
653 		do {
654 			ntfs_debug("block = 0x%x.", block);
655 			if (unlikely(!(bhs[nr_bhs] = sb_getblk(sb, block))))
656 				goto getblk_err;
657 			nr_bhs++;
658 		} while (++block < max_block);
659 	}
660 
661 	/* Release the lock if we took it. */
662 	if (rl)
663 		up_read(&ni->runlist.lock);
664 
665 	/* Setup and initiate io on all buffer heads. */
666 	for (i = 0; i < nr_bhs; i++) {
667 		struct buffer_head *tbh = bhs[i];
668 
669 		if (!trylock_buffer(tbh))
670 			continue;
671 		if (unlikely(buffer_uptodate(tbh))) {
672 			unlock_buffer(tbh);
673 			continue;
674 		}
675 		get_bh(tbh);
676 		tbh->b_end_io = end_buffer_read_sync;
677 		submit_bh(READ, tbh);
678 	}
679 
680 	/* Wait for io completion on all buffer heads. */
681 	for (i = 0; i < nr_bhs; i++) {
682 		struct buffer_head *tbh = bhs[i];
683 
684 		if (buffer_uptodate(tbh))
685 			continue;
686 		wait_on_buffer(tbh);
687 		/*
688 		 * We need an optimization barrier here, otherwise we start
689 		 * hitting the below fixup code when accessing a loopback
690 		 * mounted ntfs partition. This indicates either there is a
691 		 * race condition in the loop driver or, more likely, gcc
692 		 * overoptimises the code without the barrier and it doesn't
693 		 * do the Right Thing(TM).
694 		 */
695 		barrier();
696 		if (unlikely(!buffer_uptodate(tbh))) {
697 			ntfs_warning(vol->sb, "Buffer is unlocked but not "
698 					"uptodate! Unplugging the disk queue "
699 					"and rescheduling.");
700 			get_bh(tbh);
701 			blk_run_address_space(mapping);
702 			schedule();
703 			put_bh(tbh);
704 			if (unlikely(!buffer_uptodate(tbh)))
705 				goto read_err;
706 			ntfs_warning(vol->sb, "Buffer is now uptodate. Good.");
707 		}
708 	}
709 
710 	/*
711 	 * Get the compression buffer. We must not sleep any more
712 	 * until we are finished with it.
713 	 */
714 	spin_lock(&ntfs_cb_lock);
715 	cb = ntfs_compression_buffer;
716 
717 	BUG_ON(!cb);
718 
719 	cb_pos = cb;
720 	cb_end = cb + cb_size;
721 
722 	/* Copy the buffer heads into the contiguous buffer. */
723 	for (i = 0; i < nr_bhs; i++) {
724 		memcpy(cb_pos, bhs[i]->b_data, block_size);
725 		cb_pos += block_size;
726 	}
727 
728 	/* Just a precaution. */
729 	if (cb_pos + 2 <= cb + cb_size)
730 		*(u16*)cb_pos = 0;
731 
732 	/* Reset cb_pos back to the beginning. */
733 	cb_pos = cb;
734 
735 	/* We now have both source (if present) and destination. */
736 	ntfs_debug("Successfully read the compression block.");
737 
738 	/* The last page and maximum offset within it for the current cb. */
739 	cb_max_page = (cur_page << PAGE_CACHE_SHIFT) + cur_ofs + cb_size;
740 	cb_max_ofs = cb_max_page & ~PAGE_CACHE_MASK;
741 	cb_max_page >>= PAGE_CACHE_SHIFT;
742 
743 	/* Catch end of file inside a compression block. */
744 	if (cb_max_page > max_page)
745 		cb_max_page = max_page;
746 
747 	if (vcn == start_vcn - cb_clusters) {
748 		/* Sparse cb, zero out page range overlapping the cb. */
749 		ntfs_debug("Found sparse compression block.");
750 		/* We can sleep from now on, so we drop lock. */
751 		spin_unlock(&ntfs_cb_lock);
752 		if (cb_max_ofs)
753 			cb_max_page--;
754 		for (; cur_page < cb_max_page; cur_page++) {
755 			page = pages[cur_page];
756 			if (page) {
757 				/*
758 				 * FIXME: Using clear_page() will become wrong
759 				 * when we get PAGE_CACHE_SIZE != PAGE_SIZE but
760 				 * for now there is no problem.
761 				 */
762 				if (likely(!cur_ofs))
763 					clear_page(page_address(page));
764 				else
765 					memset(page_address(page) + cur_ofs, 0,
766 							PAGE_CACHE_SIZE -
767 							cur_ofs);
768 				flush_dcache_page(page);
769 				kunmap(page);
770 				SetPageUptodate(page);
771 				unlock_page(page);
772 				if (cur_page == xpage)
773 					xpage_done = 1;
774 				else
775 					page_cache_release(page);
776 				pages[cur_page] = NULL;
777 			}
778 			cb_pos += PAGE_CACHE_SIZE - cur_ofs;
779 			cur_ofs = 0;
780 			if (cb_pos >= cb_end)
781 				break;
782 		}
783 		/* If we have a partial final page, deal with it now. */
784 		if (cb_max_ofs && cb_pos < cb_end) {
785 			page = pages[cur_page];
786 			if (page)
787 				memset(page_address(page) + cur_ofs, 0,
788 						cb_max_ofs - cur_ofs);
789 			/*
790 			 * No need to update cb_pos at this stage:
791 			 *	cb_pos += cb_max_ofs - cur_ofs;
792 			 */
793 			cur_ofs = cb_max_ofs;
794 		}
795 	} else if (vcn == start_vcn) {
796 		/* We can't sleep so we need two stages. */
797 		unsigned int cur2_page = cur_page;
798 		unsigned int cur_ofs2 = cur_ofs;
799 		u8 *cb_pos2 = cb_pos;
800 
801 		ntfs_debug("Found uncompressed compression block.");
802 		/* Uncompressed cb, copy it to the destination pages. */
803 		/*
804 		 * TODO: As a big optimization, we could detect this case
805 		 * before we read all the pages and use block_read_full_page()
806 		 * on all full pages instead (we still have to treat partial
807 		 * pages especially but at least we are getting rid of the
808 		 * synchronous io for the majority of pages.
809 		 * Or if we choose not to do the read-ahead/-behind stuff, we
810 		 * could just return block_read_full_page(pages[xpage]) as long
811 		 * as PAGE_CACHE_SIZE <= cb_size.
812 		 */
813 		if (cb_max_ofs)
814 			cb_max_page--;
815 		/* First stage: copy data into destination pages. */
816 		for (; cur_page < cb_max_page; cur_page++) {
817 			page = pages[cur_page];
818 			if (page)
819 				memcpy(page_address(page) + cur_ofs, cb_pos,
820 						PAGE_CACHE_SIZE - cur_ofs);
821 			cb_pos += PAGE_CACHE_SIZE - cur_ofs;
822 			cur_ofs = 0;
823 			if (cb_pos >= cb_end)
824 				break;
825 		}
826 		/* If we have a partial final page, deal with it now. */
827 		if (cb_max_ofs && cb_pos < cb_end) {
828 			page = pages[cur_page];
829 			if (page)
830 				memcpy(page_address(page) + cur_ofs, cb_pos,
831 						cb_max_ofs - cur_ofs);
832 			cb_pos += cb_max_ofs - cur_ofs;
833 			cur_ofs = cb_max_ofs;
834 		}
835 		/* We can sleep from now on, so drop lock. */
836 		spin_unlock(&ntfs_cb_lock);
837 		/* Second stage: finalize pages. */
838 		for (; cur2_page < cb_max_page; cur2_page++) {
839 			page = pages[cur2_page];
840 			if (page) {
841 				/*
842 				 * If we are outside the initialized size, zero
843 				 * the out of bounds page range.
844 				 */
845 				handle_bounds_compressed_page(page, i_size,
846 						initialized_size);
847 				flush_dcache_page(page);
848 				kunmap(page);
849 				SetPageUptodate(page);
850 				unlock_page(page);
851 				if (cur2_page == xpage)
852 					xpage_done = 1;
853 				else
854 					page_cache_release(page);
855 				pages[cur2_page] = NULL;
856 			}
857 			cb_pos2 += PAGE_CACHE_SIZE - cur_ofs2;
858 			cur_ofs2 = 0;
859 			if (cb_pos2 >= cb_end)
860 				break;
861 		}
862 	} else {
863 		/* Compressed cb, decompress it into the destination page(s). */
864 		unsigned int prev_cur_page = cur_page;
865 
866 		ntfs_debug("Found compressed compression block.");
867 		err = ntfs_decompress(pages, &cur_page, &cur_ofs,
868 				cb_max_page, cb_max_ofs, xpage, &xpage_done,
869 				cb_pos,	cb_size - (cb_pos - cb), i_size,
870 				initialized_size);
871 		/*
872 		 * We can sleep from now on, lock already dropped by
873 		 * ntfs_decompress().
874 		 */
875 		if (err) {
876 			ntfs_error(vol->sb, "ntfs_decompress() failed in inode "
877 					"0x%lx with error code %i. Skipping "
878 					"this compression block.",
879 					ni->mft_no, -err);
880 			/* Release the unfinished pages. */
881 			for (; prev_cur_page < cur_page; prev_cur_page++) {
882 				page = pages[prev_cur_page];
883 				if (page) {
884 					flush_dcache_page(page);
885 					kunmap(page);
886 					unlock_page(page);
887 					if (prev_cur_page != xpage)
888 						page_cache_release(page);
889 					pages[prev_cur_page] = NULL;
890 				}
891 			}
892 		}
893 	}
894 
895 	/* Release the buffer heads. */
896 	for (i = 0; i < nr_bhs; i++)
897 		brelse(bhs[i]);
898 
899 	/* Do we have more work to do? */
900 	if (nr_cbs)
901 		goto do_next_cb;
902 
903 	/* We no longer need the list of buffer heads. */
904 	kfree(bhs);
905 
906 	/* Clean up if we have any pages left. Should never happen. */
907 	for (cur_page = 0; cur_page < max_page; cur_page++) {
908 		page = pages[cur_page];
909 		if (page) {
910 			ntfs_error(vol->sb, "Still have pages left! "
911 					"Terminating them with extreme "
912 					"prejudice.  Inode 0x%lx, page index "
913 					"0x%lx.", ni->mft_no, page->index);
914 			flush_dcache_page(page);
915 			kunmap(page);
916 			unlock_page(page);
917 			if (cur_page != xpage)
918 				page_cache_release(page);
919 			pages[cur_page] = NULL;
920 		}
921 	}
922 
923 	/* We no longer need the list of pages. */
924 	kfree(pages);
925 
926 	/* If we have completed the requested page, we return success. */
927 	if (likely(xpage_done))
928 		return 0;
929 
930 	ntfs_debug("Failed. Returning error code %s.", err == -EOVERFLOW ?
931 			"EOVERFLOW" : (!err ? "EIO" : "unknown error"));
932 	return err < 0 ? err : -EIO;
933 
934 read_err:
935 	ntfs_error(vol->sb, "IO error while reading compressed data.");
936 	/* Release the buffer heads. */
937 	for (i = 0; i < nr_bhs; i++)
938 		brelse(bhs[i]);
939 	goto err_out;
940 
941 map_rl_err:
942 	ntfs_error(vol->sb, "ntfs_map_runlist() failed. Cannot read "
943 			"compression block.");
944 	goto err_out;
945 
946 rl_err:
947 	up_read(&ni->runlist.lock);
948 	ntfs_error(vol->sb, "ntfs_rl_vcn_to_lcn() failed. Cannot read "
949 			"compression block.");
950 	goto err_out;
951 
952 getblk_err:
953 	up_read(&ni->runlist.lock);
954 	ntfs_error(vol->sb, "getblk() failed. Cannot read compression block.");
955 
956 err_out:
957 	kfree(bhs);
958 	for (i = cur_page; i < max_page; i++) {
959 		page = pages[i];
960 		if (page) {
961 			flush_dcache_page(page);
962 			kunmap(page);
963 			unlock_page(page);
964 			if (i != xpage)
965 				page_cache_release(page);
966 		}
967 	}
968 	kfree(pages);
969 	return -EIO;
970 }
971