fs/ntfs/compress.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /**
   3  * compress.c - NTFS kernel compressed attributes handling.
   4  *              Part of the Linux-NTFS project.
   5  *
   6  * Copyright (c) 2001-2004 Anton Altaparmakov
   7  * Copyright (c) 2002 Richard Russon
   8  */
   9
  10 #include <linux/fs.h>
  11 #include <linux/buffer_head.h>
  12 #include <linux/blkdev.h>
  13 #include <linux/vmalloc.h>
  14 #include <linux/slab.h>
  15
  16 #include "attrib.h"
  17 #include "inode.h"
  18 #include "debug.h"
  19 #include "ntfs.h"
  20
  21 /**
  22  * ntfs_compression_constants - enum of constants used in the compression code
  23  */
  24 typedef enum {
  25         /* Token types and access mask. */
  26         NTFS_SYMBOL_TOKEN       =       0,
  27         NTFS_PHRASE_TOKEN       =       1,
  28         NTFS_TOKEN_MASK         =       1,
  29
  30         /* Compression sub-block constants. */
  31         NTFS_SB_SIZE_MASK       =       0x0fff,
  32         NTFS_SB_SIZE            =       0x1000,
  33         NTFS_SB_IS_COMPRESSED   =       0x8000,
  34
  35         /*
  36          * The maximum compression block size is by definition 16 * the cluster
  37          * size, with the maximum supported cluster size being 4kiB. Thus the
  38          * maximum compression buffer size is 64kiB, so we use this when
  39          * initializing the compression buffer.
  40          */
  41         NTFS_MAX_CB_SIZE        = 64 * 1024,
  42 } ntfs_compression_constants;
  43
  44 /**
  45  * ntfs_compression_buffer - one buffer for the decompression engine
  46  */
  47 static u8 *ntfs_compression_buffer;
  48
  49 /**
  50  * ntfs_cb_lock - spinlock which protects ntfs_compression_buffer
  51  */
  52 static DEFINE_SPINLOCK(ntfs_cb_lock);
  53
  54 /**
  55  * allocate_compression_buffers - allocate the decompression buffers
  56  *
  57  * Caller has to hold the ntfs_lock mutex.
  58  *
  59  * Return 0 on success or -ENOMEM if the allocations failed.
  60  */
  61 int allocate_compression_buffers(void)
  62 {
  63         BUG_ON(ntfs_compression_buffer);
  64
  65         ntfs_compression_buffer = vmalloc(NTFS_MAX_CB_SIZE);
  66         if (!ntfs_compression_buffer)
  67                 return -ENOMEM;
  68         return 0;
  69 }
  70
  71 /**
  72  * free_compression_buffers - free the decompression buffers
  73  *
  74  * Caller has to hold the ntfs_lock mutex.
  75  */
  76 void free_compression_buffers(void)
  77 {
  78         BUG_ON(!ntfs_compression_buffer);
  79         vfree(ntfs_compression_buffer);
  80         ntfs_compression_buffer = NULL;
  81 }
  82
  83 /**
  84  * zero_partial_compressed_page - zero out of bounds compressed page region
  85  */
  86 static void zero_partial_compressed_page(struct page *page,
  87                 const s64 initialized_size)
  88 {
  89         u8 *kp = page_address(page);
  90         unsigned int kp_ofs;
  91
  92         ntfs_debug("Zeroing page region outside initialized size.");
  93         if (((s64)page->index << PAGE_SHIFT) >= initialized_size) {
  94                 clear_page(kp);
  95                 return;
  96         }
  97         kp_ofs = initialized_size & ~PAGE_MASK;
  98         memset(kp + kp_ofs, 0, PAGE_SIZE - kp_ofs);
  99         return;
 100 }
 101
 102 /**
 103  * handle_bounds_compressed_page - test for&handle out of bounds compressed page
 104  */
 105 static inline void handle_bounds_compressed_page(struct page *page,
 106                 const loff_t i_size, const s64 initialized_size)
 107 {
 108         if ((page->index >= (initialized_size >> PAGE_SHIFT)) &&
 109                         (initialized_size < i_size))
 110                 zero_partial_compressed_page(page, initialized_size);
 111         return;
 112 }
 113
 114 /**
 115  * ntfs_decompress - decompress a compression block into an array of pages
 116  * @dest_pages:         destination array of pages
 117  * @completed_pages:    scratch space to track completed pages
 118  * @dest_index:         current index into @dest_pages (IN/OUT)
 119  * @dest_ofs:           current offset within @dest_pages[@dest_index] (IN/OUT)
 120  * @dest_max_index:     maximum index into @dest_pages (IN)
 121  * @dest_max_ofs:       maximum offset within @dest_pages[@dest_max_index] (IN)
 122  * @xpage:              the target page (-1 if none) (IN)
 123  * @xpage_done:         set to 1 if xpage was completed successfully (IN/OUT)
 124  * @cb_start:           compression block to decompress (IN)
 125  * @cb_size:            size of compression block @cb_start in bytes (IN)
 126  * @i_size:             file size when we started the read (IN)
 127  * @initialized_size:   initialized file size when we started the read (IN)
 128  *
 129  * The caller must have disabled preemption. ntfs_decompress() reenables it when
 130  * the critical section is finished.
 131  *
 132  * This decompresses the compression block @cb_start into the array of
 133  * destination pages @dest_pages starting at index @dest_index into @dest_pages
 134  * and at offset @dest_pos into the page @dest_pages[@dest_index].
 135  *
 136  * When the page @dest_pages[@xpage] is completed, @xpage_done is set to 1.
 137  * If xpage is -1 or @xpage has not been completed, @xpage_done is not modified.
 138  *
 139  * @cb_start is a pointer to the compression block which needs decompressing
 140  * and @cb_size is the size of @cb_start in bytes (8-64kiB).
 141  *
 142  * Return 0 if success or -EOVERFLOW on error in the compressed stream.
 143  * @xpage_done indicates whether the target page (@dest_pages[@xpage]) was
 144  * completed during the decompression of the compression block (@cb_start).
 145  *
 146  * Warning: This function *REQUIRES* PAGE_SIZE >= 4096 or it will blow up
 147  * unpredicatbly! You have been warned!
 148  *
 149  * Note to hackers: This function may not sleep until it has finished accessing
 150  * the compression block @cb_start as it is a per-CPU buffer.
 151  */
 152 static int ntfs_decompress(struct page *dest_pages[], int completed_pages[],
 153                 int *dest_index, int *dest_ofs, const int dest_max_index,
 154                 const int dest_max_ofs, const int xpage, char *xpage_done,
 155                 u8 *const cb_start, const u32 cb_size, const loff_t i_size,
 156                 const s64 initialized_size)
 157 {
 158         /*
 159          * Pointers into the compressed data, i.e. the compression block (cb),
 160          * and the therein contained sub-blocks (sb).
 161          */
 162         u8 *cb_end = cb_start + cb_size; /* End of cb. */
 163         u8 *cb = cb_start;      /* Current position in cb. */
 164         u8 *cb_sb_start = cb;   /* Beginning of the current sb in the cb. */
 165         u8 *cb_sb_end;          /* End of current sb / beginning of next sb. */
 166
 167         /* Variables for uncompressed data / destination. */
 168         struct page *dp;        /* Current destination page being worked on. */
 169         u8 *dp_addr;            /* Current pointer into dp. */
 170         u8 *dp_sb_start;        /* Start of current sub-block in dp. */
 171         u8 *dp_sb_end;          /* End of current sb in dp (dp_sb_start +
 172                                    NTFS_SB_SIZE). */
 173         u16 do_sb_start;        /* @dest_ofs when starting this sub-block. */
 174         u16 do_sb_end;          /* @dest_ofs of end of this sb (do_sb_start +
 175                                    NTFS_SB_SIZE). */
 176
 177         /* Variables for tag and token parsing. */
 178         u8 tag;                 /* Current tag. */
 179         int token;              /* Loop counter for the eight tokens in tag. */
 180         int nr_completed_pages = 0;
 181
 182         /* Default error code. */
 183         int err = -EOVERFLOW;
 184
 185         ntfs_debug("Entering, cb_size = 0x%x.", cb_size);
 186 do_next_sb:
 187         ntfs_debug("Beginning sub-block at offset = 0x%zx in the cb.",
 188                         cb - cb_start);
 189         /*
 190          * Have we reached the end of the compression block or the end of the
 191          * decompressed data?  The latter can happen for example if the current
 192          * position in the compression block is one byte before its end so the
 193          * first two checks do not detect it.
 194          */
 195         if (cb == cb_end || !le16_to_cpup((le16*)cb) ||
 196                         (*dest_index == dest_max_index &&
 197                         *dest_ofs == dest_max_ofs)) {
 198                 int i;
 199
 200                 ntfs_debug("Completed. Returning success (0).");
 201                 err = 0;
 202 return_error:
 203                 /* We can sleep from now on, so we drop lock. */
 204                 spin_unlock(&ntfs_cb_lock);
 205                 /* Second stage: finalize completed pages. */
 206                 if (nr_completed_pages > 0) {
 207                         for (i = 0; i < nr_completed_pages; i++) {
 208                                 int di = completed_pages[i];
 209
 210                                 dp = dest_pages[di];
 211                                 /*
 212                                  * If we are outside the initialized size, zero
 213                                  * the out of bounds page range.
 214                                  */
 215                                 handle_bounds_compressed_page(dp, i_size,
 216                                                 initialized_size);
 217                                 flush_dcache_page(dp);
 218                                 kunmap(dp);
 219                                 SetPageUptodate(dp);
 220                                 unlock_page(dp);
 221                                 if (di == xpage)
 222                                         *xpage_done = 1;
 223                                 else
 224                                         put_page(dp);
 225                                 dest_pages[di] = NULL;
 226                         }
 227                 }
 228                 return err;
 229         }
 230
 231         /* Setup offsets for the current sub-block destination. */
 232         do_sb_start = *dest_ofs;
 233         do_sb_end = do_sb_start + NTFS_SB_SIZE;
 234
 235         /* Check that we are still within allowed boundaries. */
 236         if (*dest_index == dest_max_index && do_sb_end > dest_max_ofs)
 237                 goto return_overflow;
 238
 239         /* Does the minimum size of a compressed sb overflow valid range? */
 240         if (cb + 6 > cb_end)
 241                 goto return_overflow;
 242
 243         /* Setup the current sub-block source pointers and validate range. */
 244         cb_sb_start = cb;
 245         cb_sb_end = cb_sb_start + (le16_to_cpup((le16*)cb) & NTFS_SB_SIZE_MASK)
 246                         + 3;
 247         if (cb_sb_end > cb_end)
 248                 goto return_overflow;
 249
 250         /* Get the current destination page. */
 251         dp = dest_pages[*dest_index];
 252         if (!dp) {
 253                 /* No page present. Skip decompression of this sub-block. */
 254                 cb = cb_sb_end;
 255
 256                 /* Advance destination position to next sub-block. */
 257                 *dest_ofs = (*dest_ofs + NTFS_SB_SIZE) & ~PAGE_MASK;
 258                 if (!*dest_ofs && (++*dest_index > dest_max_index))
 259                         goto return_overflow;
 260                 goto do_next_sb;
 261         }
 262
 263         /* We have a valid destination page. Setup the destination pointers. */
 264         dp_addr = (u8*)page_address(dp) + do_sb_start;
 265
 266         /* Now, we are ready to process the current sub-block (sb). */
 267         if (!(le16_to_cpup((le16*)cb) & NTFS_SB_IS_COMPRESSED)) {
 268                 ntfs_debug("Found uncompressed sub-block.");
 269                 /* This sb is not compressed, just copy it into destination. */
 270
 271                 /* Advance source position to first data byte. */
 272                 cb += 2;
 273
 274                 /* An uncompressed sb must be full size. */
 275                 if (cb_sb_end - cb != NTFS_SB_SIZE)
 276                         goto return_overflow;
 277
 278                 /* Copy the block and advance the source position. */
 279                 memcpy(dp_addr, cb, NTFS_SB_SIZE);
 280                 cb += NTFS_SB_SIZE;
 281
 282                 /* Advance destination position to next sub-block. */
 283                 *dest_ofs += NTFS_SB_SIZE;
 284                 if (!(*dest_ofs &= ~PAGE_MASK)) {
 285 finalize_page:
 286                         /*
 287                          * First stage: add current page index to array of
 288                          * completed pages.
 289                          */
 290                         completed_pages[nr_completed_pages++] = *dest_index;
 291                         if (++*dest_index > dest_max_index)
 292                                 goto return_overflow;
 293                 }
 294                 goto do_next_sb;
 295         }
 296         ntfs_debug("Found compressed sub-block.");
 297         /* This sb is compressed, decompress it into destination. */
 298
 299         /* Setup destination pointers. */
 300         dp_sb_start = dp_addr;
 301         dp_sb_end = dp_sb_start + NTFS_SB_SIZE;
 302
 303         /* Forward to the first tag in the sub-block. */
 304         cb += 2;
 305 do_next_tag:
 306         if (cb == cb_sb_end) {
 307                 /* Check if the decompressed sub-block was not full-length. */
 308                 if (dp_addr < dp_sb_end) {
 309                         int nr_bytes = do_sb_end - *dest_ofs;
 310
 311                         ntfs_debug("Filling incomplete sub-block with "
 312                                         "zeroes.");
 313                         /* Zero remainder and update destination position. */
 314                         memset(dp_addr, 0, nr_bytes);
 315                         *dest_ofs += nr_bytes;
 316                 }
 317                 /* We have finished the current sub-block. */
 318                 if (!(*dest_ofs &= ~PAGE_MASK))
 319                         goto finalize_page;
 320                 goto do_next_sb;
 321         }
 322
 323         /* Check we are still in range. */
 324         if (cb > cb_sb_end || dp_addr > dp_sb_end)
 325                 goto return_overflow;
 326
 327         /* Get the next tag and advance to first token. */
 328         tag = *cb++;
 329
 330         /* Parse the eight tokens described by the tag. */
 331         for (token = 0; token < 8; token++, tag >>= 1) {
 332                 u16 lg, pt, length, max_non_overlap;
 333                 register u16 i;
 334                 u8 *dp_back_addr;
 335
 336                 /* Check if we are done / still in range. */
 337                 if (cb >= cb_sb_end || dp_addr > dp_sb_end)
 338                         break;
 339
 340                 /* Determine token type and parse appropriately.*/
 341                 if ((tag & NTFS_TOKEN_MASK) == NTFS_SYMBOL_TOKEN) {
 342                         /*
 343                          * We have a symbol token, copy the symbol across, and
 344                          * advance the source and destination positions.
 345                          */
 346                         *dp_addr++ = *cb++;
 347                         ++*dest_ofs;
 348
 349                         /* Continue with the next token. */
 350                         continue;
 351                 }
 352
 353                 /*
 354                  * We have a phrase token. Make sure it is not the first tag in
 355                  * the sb as this is illegal and would confuse the code below.
 356                  */
 357                 if (dp_addr == dp_sb_start)
 358                         goto return_overflow;
 359
 360                 /*
 361                  * Determine the number of bytes to go back (p) and the number
 362                  * of bytes to copy (l). We use an optimized algorithm in which
 363                  * we first calculate log2(current destination position in sb),
 364                  * which allows determination of l and p in O(1) rather than
 365                  * O(n). We just need an arch-optimized log2() function now.
 366                  */
 367                 lg = 0;
 368                 for (i = *dest_ofs - do_sb_start - 1; i >= 0x10; i >>= 1)
 369                         lg++;
 370
 371                 /* Get the phrase token into i. */
 372                 pt = le16_to_cpup((le16*)cb);
 373
 374                 /*
 375                  * Calculate starting position of the byte sequence in
 376                  * the destination using the fact that p = (pt >> (12 - lg)) + 1
 377                  * and make sure we don't go too far back.
 378                  */
 379                 dp_back_addr = dp_addr - (pt >> (12 - lg)) - 1;
 380                 if (dp_back_addr < dp_sb_start)
 381                         goto return_overflow;
 382
 383                 /* Now calculate the length of the byte sequence. */
 384                 length = (pt & (0xfff >> lg)) + 3;
 385
 386                 /* Advance destination position and verify it is in range. */
 387                 *dest_ofs += length;
 388                 if (*dest_ofs > do_sb_end)
 389                         goto return_overflow;
 390
 391                 /* The number of non-overlapping bytes. */
 392                 max_non_overlap = dp_addr - dp_back_addr;
 393
 394                 if (length <= max_non_overlap) {
 395                         /* The byte sequence doesn't overlap, just copy it. */
 396                         memcpy(dp_addr, dp_back_addr, length);
 397
 398                         /* Advance destination pointer. */
 399                         dp_addr += length;
 400                 } else {
 401                         /*
 402                          * The byte sequence does overlap, copy non-overlapping
 403                          * part and then do a slow byte by byte copy for the
 404                          * overlapping part. Also, advance the destination
 405                          * pointer.
 406                          */
 407                         memcpy(dp_addr, dp_back_addr, max_non_overlap);
 408                         dp_addr += max_non_overlap;
 409                         dp_back_addr += max_non_overlap;
 410                         length -= max_non_overlap;
 411                         while (length--)
 412                                 *dp_addr++ = *dp_back_addr++;
 413                 }
 414
 415                 /* Advance source position and continue with the next token. */
 416                 cb += 2;
 417         }
 418
 419         /* No tokens left in the current tag. Continue with the next tag. */
 420         goto do_next_tag;
 421
 422 return_overflow:
 423         ntfs_error(NULL, "Failed. Returning -EOVERFLOW.");
 424         goto return_error;
 425 }
 426
 427 /**
 428  * ntfs_read_compressed_block - read a compressed block into the page cache
 429  * @page:       locked page in the compression block(s) we need to read
 430  *
 431  * When we are called the page has already been verified to be locked and the
 432  * attribute is known to be non-resident, not encrypted, but compressed.
 433  *
 434  * 1. Determine which compression block(s) @page is in.
 435  * 2. Get hold of all pages corresponding to this/these compression block(s).
 436  * 3. Read the (first) compression block.
 437  * 4. Decompress it into the corresponding pages.
 438  * 5. Throw the compressed data away and proceed to 3. for the next compression
 439  *    block or return success if no more compression blocks left.
 440  *
 441  * Warning: We have to be careful what we do about existing pages. They might
 442  * have been written to so that we would lose data if we were to just overwrite
 443  * them with the out-of-date uncompressed data.
 444  *
 445  * FIXME: For PAGE_SIZE > cb_size we are not doing the Right Thing(TM) at
 446  * the end of the file I think. We need to detect this case and zero the out
 447  * of bounds remainder of the page in question and mark it as handled. At the
 448  * moment we would just return -EIO on such a page. This bug will only become
 449  * apparent if pages are above 8kiB and the NTFS volume only uses 512 byte
 450  * clusters so is probably not going to be seen by anyone. Still this should
 451  * be fixed. (AIA)
 452  *
 453  * FIXME: Again for PAGE_SIZE > cb_size we are screwing up both in
 454  * handling sparse and compressed cbs. (AIA)
 455  *
 456  * FIXME: At the moment we don't do any zeroing out in the case that
 457  * initialized_size is less than data_size. This should be safe because of the
 458  * nature of the compression algorithm used. Just in case we check and output
 459  * an error message in read inode if the two sizes are not equal for a
 460  * compressed file. (AIA)
 461  */
 462 int ntfs_read_compressed_block(struct page *page)
 463 {
 464         loff_t i_size;
 465         s64 initialized_size;
 466         struct address_space *mapping = page->mapping;
 467         ntfs_inode *ni = NTFS_I(mapping->host);
 468         ntfs_volume *vol = ni->vol;
 469         struct super_block *sb = vol->sb;
 470         runlist_element *rl;
 471         unsigned long flags, block_size = sb->s_blocksize;
 472         unsigned char block_size_bits = sb->s_blocksize_bits;
 473         u8 *cb, *cb_pos, *cb_end;
 474         struct buffer_head **bhs;
 475         unsigned long offset, index = page->index;
 476         u32 cb_size = ni->itype.compressed.block_size;
 477         u64 cb_size_mask = cb_size - 1UL;
 478         VCN vcn;
 479         LCN lcn;
 480         /* The first wanted vcn (minimum alignment is PAGE_SIZE). */
 481         VCN start_vcn = (((s64)index << PAGE_SHIFT) & ~cb_size_mask) >>
 482                         vol->cluster_size_bits;
 483         /*
 484          * The first vcn after the last wanted vcn (minimum alignment is again
 485          * PAGE_SIZE.
 486          */
 487         VCN end_vcn = ((((s64)(index + 1UL) << PAGE_SHIFT) + cb_size - 1)
 488                         & ~cb_size_mask) >> vol->cluster_size_bits;
 489         /* Number of compression blocks (cbs) in the wanted vcn range. */
 490         unsigned int nr_cbs = (end_vcn - start_vcn) << vol->cluster_size_bits
 491                         >> ni->itype.compressed.block_size_bits;
 492         /*
 493          * Number of pages required to store the uncompressed data from all
 494          * compression blocks (cbs) overlapping @page. Due to alignment
 495          * guarantees of start_vcn and end_vcn, no need to round up here.
 496          */
 497         unsigned int nr_pages = (end_vcn - start_vcn) <<
 498                         vol->cluster_size_bits >> PAGE_SHIFT;
 499         unsigned int xpage, max_page, cur_page, cur_ofs, i;
 500         unsigned int cb_clusters, cb_max_ofs;
 501         int block, max_block, cb_max_page, bhs_size, nr_bhs, err = 0;
 502         struct page **pages;
 503         int *completed_pages;
 504         unsigned char xpage_done = 0;
 505
 506         ntfs_debug("Entering, page->index = 0x%lx, cb_size = 0x%x, nr_pages = "
 507                         "%i.", index, cb_size, nr_pages);
 508         /*
 509          * Bad things happen if we get here for anything that is not an
 510          * unnamed $DATA attribute.
 511          */
 512         BUG_ON(ni->type != AT_DATA);
 513         BUG_ON(ni->name_len);
 514
 515         pages = kmalloc_array(nr_pages, sizeof(struct page *), GFP_NOFS);
 516         completed_pages = kmalloc_array(nr_pages + 1, sizeof(int), GFP_NOFS);
 517
 518         /* Allocate memory to store the buffer heads we need. */
 519         bhs_size = cb_size / block_size * sizeof(struct buffer_head *);
 520         bhs = kmalloc(bhs_size, GFP_NOFS);
 521
 522         if (unlikely(!pages || !bhs || !completed_pages)) {
 523                 kfree(bhs);
 524                 kfree(pages);
 525                 kfree(completed_pages);
 526                 unlock_page(page);
 527                 ntfs_error(vol->sb, "Failed to allocate internal buffers.");
 528                 return -ENOMEM;
 529         }
 530
 531         /*
 532          * We have already been given one page, this is the one we must do.
 533          * Once again, the alignment guarantees keep it simple.
 534          */
 535         offset = start_vcn << vol->cluster_size_bits >> PAGE_SHIFT;
 536         xpage = index - offset;
 537         pages[xpage] = page;
 538         /*
 539          * The remaining pages need to be allocated and inserted into the page
 540          * cache, alignment guarantees keep all the below much simpler. (-8
 541          */
 542         read_lock_irqsave(&ni->size_lock, flags);
 543         i_size = i_size_read(VFS_I(ni));
 544         initialized_size = ni->initialized_size;
 545         read_unlock_irqrestore(&ni->size_lock, flags);
 546         max_page = ((i_size + PAGE_SIZE - 1) >> PAGE_SHIFT) -
 547                         offset;
 548         /* Is the page fully outside i_size? (truncate in progress) */
 549         if (xpage >= max_page) {
 550                 kfree(bhs);
 551                 kfree(pages);
 552                 kfree(completed_pages);
 553                 zero_user(page, 0, PAGE_SIZE);
 554                 ntfs_debug("Compressed read outside i_size - truncated?");
 555                 SetPageUptodate(page);
 556                 unlock_page(page);
 557                 return 0;
 558         }
 559         if (nr_pages < max_page)
 560                 max_page = nr_pages;
 561         for (i = 0; i < max_page; i++, offset++) {
 562                 if (i != xpage)
 563                         pages[i] = grab_cache_page_nowait(mapping, offset);
 564                 page = pages[i];
 565                 if (page) {
 566                         /*
 567                          * We only (re)read the page if it isn't already read
 568                          * in and/or dirty or we would be losing data or at
 569                          * least wasting our time.
 570                          */
 571                         if (!PageDirty(page) && (!PageUptodate(page) ||
 572                                         PageError(page))) {
 573                                 ClearPageError(page);
 574                                 kmap(page);
 575                                 continue;
 576                         }
 577                         unlock_page(page);
 578                         put_page(page);
 579                         pages[i] = NULL;
 580                 }
 581         }
 582
 583         /*
 584          * We have the runlist, and all the destination pages we need to fill.
 585          * Now read the first compression block.
 586          */
 587         cur_page = 0;
 588         cur_ofs = 0;
 589         cb_clusters = ni->itype.compressed.block_clusters;
 590 do_next_cb:
 591         nr_cbs--;
 592         nr_bhs = 0;
 593
 594         /* Read all cb buffer heads one cluster at a time. */
 595         rl = NULL;
 596         for (vcn = start_vcn, start_vcn += cb_clusters; vcn < start_vcn;
 597                         vcn++) {
 598                 bool is_retry = false;
 599
 600                 if (!rl) {
 601 lock_retry_remap:
 602                         down_read(&ni->runlist.lock);
 603                         rl = ni->runlist.rl;
 604                 }
 605                 if (likely(rl != NULL)) {
 606                         /* Seek to element containing target vcn. */
 607                         while (rl->length && rl[1].vcn <= vcn)
 608                                 rl++;
 609                         lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
 610                 } else
 611                         lcn = LCN_RL_NOT_MAPPED;
 612                 ntfs_debug("Reading vcn = 0x%llx, lcn = 0x%llx.",
 613                                 (unsigned long long)vcn,
 614                                 (unsigned long long)lcn);
 615                 if (lcn < 0) {
 616                         /*
 617                          * When we reach the first sparse cluster we have
 618                          * finished with the cb.
 619                          */
 620                         if (lcn == LCN_HOLE)
 621                                 break;
 622                         if (is_retry || lcn != LCN_RL_NOT_MAPPED)
 623                                 goto rl_err;
 624                         is_retry = true;
 625                         /*
 626                          * Attempt to map runlist, dropping lock for the
 627                          * duration.
 628                          */
 629                         up_read(&ni->runlist.lock);
 630                         if (!ntfs_map_runlist(ni, vcn))
 631                                 goto lock_retry_remap;
 632                         goto map_rl_err;
 633                 }
 634                 block = lcn << vol->cluster_size_bits >> block_size_bits;
 635                 /* Read the lcn from device in chunks of block_size bytes. */
 636                 max_block = block + (vol->cluster_size >> block_size_bits);
 637                 do {
 638                         ntfs_debug("block = 0x%x.", block);
 639                         if (unlikely(!(bhs[nr_bhs] = sb_getblk(sb, block))))
 640                                 goto getblk_err;
 641                         nr_bhs++;
 642                 } while (++block < max_block);
 643         }
 644
 645         /* Release the lock if we took it. */
 646         if (rl)
 647                 up_read(&ni->runlist.lock);
 648
 649         /* Setup and initiate io on all buffer heads. */
 650         for (i = 0; i < nr_bhs; i++) {
 651                 struct buffer_head *tbh = bhs[i];
 652
 653                 if (!trylock_buffer(tbh))
 654                         continue;
 655                 if (unlikely(buffer_uptodate(tbh))) {
 656                         unlock_buffer(tbh);
 657                         continue;
 658                 }
 659                 get_bh(tbh);
 660                 tbh->b_end_io = end_buffer_read_sync;
 661                 submit_bh(REQ_OP_READ, 0, tbh);
 662         }
 663
 664         /* Wait for io completion on all buffer heads. */
 665         for (i = 0; i < nr_bhs; i++) {
 666                 struct buffer_head *tbh = bhs[i];
 667
 668                 if (buffer_uptodate(tbh))
 669                         continue;
 670                 wait_on_buffer(tbh);
 671                 /*
 672                  * We need an optimization barrier here, otherwise we start
 673                  * hitting the below fixup code when accessing a loopback
 674                  * mounted ntfs partition. This indicates either there is a
 675                  * race condition in the loop driver or, more likely, gcc
 676                  * overoptimises the code without the barrier and it doesn't
 677                  * do the Right Thing(TM).
 678                  */
 679                 barrier();
 680                 if (unlikely(!buffer_uptodate(tbh))) {
 681                         ntfs_warning(vol->sb, "Buffer is unlocked but not "
 682                                         "uptodate! Unplugging the disk queue "
 683                                         "and rescheduling.");
 684                         get_bh(tbh);
 685                         io_schedule();
 686                         put_bh(tbh);
 687                         if (unlikely(!buffer_uptodate(tbh)))
 688                                 goto read_err;
 689                         ntfs_warning(vol->sb, "Buffer is now uptodate. Good.");
 690                 }
 691         }
 692
 693         /*
 694          * Get the compression buffer. We must not sleep any more
 695          * until we are finished with it.
 696          */
 697         spin_lock(&ntfs_cb_lock);
 698         cb = ntfs_compression_buffer;
 699
 700         BUG_ON(!cb);
 701
 702         cb_pos = cb;
 703         cb_end = cb + cb_size;
 704
 705         /* Copy the buffer heads into the contiguous buffer. */
 706         for (i = 0; i < nr_bhs; i++) {
 707                 memcpy(cb_pos, bhs[i]->b_data, block_size);
 708                 cb_pos += block_size;
 709         }
 710
 711         /* Just a precaution. */
 712         if (cb_pos + 2 <= cb + cb_size)
 713                 *(u16*)cb_pos = 0;
 714
 715         /* Reset cb_pos back to the beginning. */
 716         cb_pos = cb;
 717
 718         /* We now have both source (if present) and destination. */
 719         ntfs_debug("Successfully read the compression block.");
 720
 721         /* The last page and maximum offset within it for the current cb. */
 722         cb_max_page = (cur_page << PAGE_SHIFT) + cur_ofs + cb_size;
 723         cb_max_ofs = cb_max_page & ~PAGE_MASK;
 724         cb_max_page >>= PAGE_SHIFT;
 725
 726         /* Catch end of file inside a compression block. */
 727         if (cb_max_page > max_page)
 728                 cb_max_page = max_page;
 729
 730         if (vcn == start_vcn - cb_clusters) {
 731                 /* Sparse cb, zero out page range overlapping the cb. */
 732                 ntfs_debug("Found sparse compression block.");
 733                 /* We can sleep from now on, so we drop lock. */
 734                 spin_unlock(&ntfs_cb_lock);
 735                 if (cb_max_ofs)
 736                         cb_max_page--;
 737                 for (; cur_page < cb_max_page; cur_page++) {
 738                         page = pages[cur_page];
 739                         if (page) {
 740                                 if (likely(!cur_ofs))
 741                                         clear_page(page_address(page));
 742                                 else
 743                                         memset(page_address(page) + cur_ofs, 0,
 744                                                         PAGE_SIZE -
 745                                                         cur_ofs);
 746                                 flush_dcache_page(page);
 747                                 kunmap(page);
 748                                 SetPageUptodate(page);
 749                                 unlock_page(page);
 750                                 if (cur_page == xpage)
 751                                         xpage_done = 1;
 752                                 else
 753                                         put_page(page);
 754                                 pages[cur_page] = NULL;
 755                         }
 756                         cb_pos += PAGE_SIZE - cur_ofs;
 757                         cur_ofs = 0;
 758                         if (cb_pos >= cb_end)
 759                                 break;
 760                 }
 761                 /* If we have a partial final page, deal with it now. */
 762                 if (cb_max_ofs && cb_pos < cb_end) {
 763                         page = pages[cur_page];
 764                         if (page)
 765                                 memset(page_address(page) + cur_ofs, 0,
 766                                                 cb_max_ofs - cur_ofs);
 767                         /*
 768                          * No need to update cb_pos at this stage:
 769                          *      cb_pos += cb_max_ofs - cur_ofs;
 770                          */
 771                         cur_ofs = cb_max_ofs;
 772                 }
 773         } else if (vcn == start_vcn) {
 774                 /* We can't sleep so we need two stages. */
 775                 unsigned int cur2_page = cur_page;
 776                 unsigned int cur_ofs2 = cur_ofs;
 777                 u8 *cb_pos2 = cb_pos;
 778
 779                 ntfs_debug("Found uncompressed compression block.");
 780                 /* Uncompressed cb, copy it to the destination pages. */
 781                 /*
 782                  * TODO: As a big optimization, we could detect this case
 783                  * before we read all the pages and use block_read_full_page()
 784                  * on all full pages instead (we still have to treat partial
 785                  * pages especially but at least we are getting rid of the
 786                  * synchronous io for the majority of pages.
 787                  * Or if we choose not to do the read-ahead/-behind stuff, we
 788                  * could just return block_read_full_page(pages[xpage]) as long
 789                  * as PAGE_SIZE <= cb_size.
 790                  */
 791                 if (cb_max_ofs)
 792                         cb_max_page--;
 793                 /* First stage: copy data into destination pages. */
 794                 for (; cur_page < cb_max_page; cur_page++) {
 795                         page = pages[cur_page];
 796                         if (page)
 797                                 memcpy(page_address(page) + cur_ofs, cb_pos,
 798                                                 PAGE_SIZE - cur_ofs);
 799                         cb_pos += PAGE_SIZE - cur_ofs;
 800                         cur_ofs = 0;
 801                         if (cb_pos >= cb_end)
 802                                 break;
 803                 }
 804                 /* If we have a partial final page, deal with it now. */
 805                 if (cb_max_ofs && cb_pos < cb_end) {
 806                         page = pages[cur_page];
 807                         if (page)
 808                                 memcpy(page_address(page) + cur_ofs, cb_pos,
 809                                                 cb_max_ofs - cur_ofs);
 810                         cb_pos += cb_max_ofs - cur_ofs;
 811                         cur_ofs = cb_max_ofs;
 812                 }
 813                 /* We can sleep from now on, so drop lock. */
 814                 spin_unlock(&ntfs_cb_lock);
 815                 /* Second stage: finalize pages. */
 816                 for (; cur2_page < cb_max_page; cur2_page++) {
 817                         page = pages[cur2_page];
 818                         if (page) {
 819                                 /*
 820                                  * If we are outside the initialized size, zero
 821                                  * the out of bounds page range.
 822                                  */
 823                                 handle_bounds_compressed_page(page, i_size,
 824                                                 initialized_size);
 825                                 flush_dcache_page(page);
 826                                 kunmap(page);
 827                                 SetPageUptodate(page);
 828                                 unlock_page(page);
 829                                 if (cur2_page == xpage)
 830                                         xpage_done = 1;
 831                                 else
 832                                         put_page(page);
 833                                 pages[cur2_page] = NULL;
 834                         }
 835                         cb_pos2 += PAGE_SIZE - cur_ofs2;
 836                         cur_ofs2 = 0;
 837                         if (cb_pos2 >= cb_end)
 838                                 break;
 839                 }
 840         } else {
 841                 /* Compressed cb, decompress it into the destination page(s). */
 842                 unsigned int prev_cur_page = cur_page;
 843
 844                 ntfs_debug("Found compressed compression block.");
 845                 err = ntfs_decompress(pages, completed_pages, &cur_page,
 846                                 &cur_ofs, cb_max_page, cb_max_ofs, xpage,
 847                                 &xpage_done, cb_pos, cb_size - (cb_pos - cb),
 848                                 i_size, initialized_size);
 849                 /*
 850                  * We can sleep from now on, lock already dropped by
 851                  * ntfs_decompress().
 852                  */
 853                 if (err) {
 854                         ntfs_error(vol->sb, "ntfs_decompress() failed in inode "
 855                                         "0x%lx with error code %i. Skipping "
 856                                         "this compression block.",
 857                                         ni->mft_no, -err);
 858                         /* Release the unfinished pages. */
 859                         for (; prev_cur_page < cur_page; prev_cur_page++) {
 860                                 page = pages[prev_cur_page];
 861                                 if (page) {
 862                                         flush_dcache_page(page);
 863                                         kunmap(page);
 864                                         unlock_page(page);
 865                                         if (prev_cur_page != xpage)
 866                                                 put_page(page);
 867                                         pages[prev_cur_page] = NULL;
 868                                 }
 869                         }
 870                 }
 871         }
 872
 873         /* Release the buffer heads. */
 874         for (i = 0; i < nr_bhs; i++)
 875                 brelse(bhs[i]);
 876
 877         /* Do we have more work to do? */
 878         if (nr_cbs)
 879                 goto do_next_cb;
 880
 881         /* We no longer need the list of buffer heads. */
 882         kfree(bhs);
 883
 884         /* Clean up if we have any pages left. Should never happen. */
 885         for (cur_page = 0; cur_page < max_page; cur_page++) {
 886                 page = pages[cur_page];
 887                 if (page) {
 888                         ntfs_error(vol->sb, "Still have pages left! "
 889                                         "Terminating them with extreme "
 890                                         "prejudice.  Inode 0x%lx, page index "
 891                                         "0x%lx.", ni->mft_no, page->index);
 892                         flush_dcache_page(page);
 893                         kunmap(page);
 894                         unlock_page(page);
 895                         if (cur_page != xpage)
 896                                 put_page(page);
 897                         pages[cur_page] = NULL;
 898                 }
 899         }
 900
 901         /* We no longer need the list of pages. */
 902         kfree(pages);
 903         kfree(completed_pages);
 904
 905         /* If we have completed the requested page, we return success. */
 906         if (likely(xpage_done))
 907                 return 0;
 908
 909         ntfs_debug("Failed. Returning error code %s.", err == -EOVERFLOW ?
 910                         "EOVERFLOW" : (!err ? "EIO" : "unknown error"));
 911         return err < 0 ? err : -EIO;
 912
 913 read_err:
 914         ntfs_error(vol->sb, "IO error while reading compressed data.");
 915         /* Release the buffer heads. */
 916         for (i = 0; i < nr_bhs; i++)
 917                 brelse(bhs[i]);
 918         goto err_out;
 919
 920 map_rl_err:
 921         ntfs_error(vol->sb, "ntfs_map_runlist() failed. Cannot read "
 922                         "compression block.");
 923         goto err_out;
 924
 925 rl_err:
 926         up_read(&ni->runlist.lock);
 927         ntfs_error(vol->sb, "ntfs_rl_vcn_to_lcn() failed. Cannot read "
 928                         "compression block.");
 929         goto err_out;
 930
 931 getblk_err:
 932         up_read(&ni->runlist.lock);
 933         ntfs_error(vol->sb, "getblk() failed. Cannot read compression block.");
 934
 935 err_out:
 936         kfree(bhs);
 937         for (i = cur_page; i < max_page; i++) {
 938                 page = pages[i];
 939                 if (page) {
 940                         flush_dcache_page(page);
 941                         kunmap(page);
 942                         unlock_page(page);
 943                         if (i != xpage)
 944                                 put_page(page);
 945                 }
 946         }
 947         kfree(pages);
 948         kfree(completed_pages);
 949         return -EIO;
 950 }