fs/ntfs/lcnalloc.c

   1 /*
   2  * lcnalloc.c - Cluster (de)allocation code.  Part of the Linux-NTFS project.
   3  *
   4  * Copyright (c) 2004 Anton Altaparmakov
   5  *
   6  * This program/include file is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License as published
   8  * by the Free Software Foundation; either version 2 of the License, or
   9  * (at your option) any later version.
  10  *
  11  * This program/include file is distributed in the hope that it will be
  12  * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  13  * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  * GNU General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program (in the main directory of the Linux-NTFS
  18  * distribution in the file COPYING); if not, write to the Free Software
  19  * Foundation,Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  20  */
  21
  22 #ifdef NTFS_RW
  23
  24 #include <linux/pagemap.h>
  25
  26 #include "lcnalloc.h"
  27 #include "debug.h"
  28 #include "bitmap.h"
  29 #include "inode.h"
  30 #include "volume.h"
  31 #include "attrib.h"
  32 #include "malloc.h"
  33 #include "ntfs.h"
  34
  35 /**
  36  * ntfs_cluster_free_from_rl_nolock - free clusters from runlist
  37  * @vol:        mounted ntfs volume on which to free the clusters
  38  * @rl:         runlist describing the clusters to free
  39  *
  40  * Free all the clusters described by the runlist @rl on the volume @vol.  In
  41  * the case of an error being returned, at least some of the clusters were not
  42  * freed.
  43  *
  44  * Return 0 on success and -errno on error.
  45  *
  46  * Locking: - The volume lcn bitmap must be locked for writing on entry and is
  47  *            left locked on return.
  48  */
  49 static int ntfs_cluster_free_from_rl_nolock(ntfs_volume *vol,
  50                 const runlist_element *rl)
  51 {
  52         struct inode *lcnbmp_vi = vol->lcnbmp_ino;
  53         int ret = 0;
  54
  55         ntfs_debug("Entering.");
  56         for (; rl->length; rl++) {
  57                 int err;
  58
  59                 if (rl->lcn < 0)
  60                         continue;
  61                 err = ntfs_bitmap_clear_run(lcnbmp_vi, rl->lcn, rl->length);
  62                 if (unlikely(err && (!ret || ret == ENOMEM) && ret != err))
  63                         ret = err;
  64         }
  65         ntfs_debug("Done.");
  66         return ret;
  67 }
  68
  69 /**
  70  * ntfs_cluster_alloc - allocate clusters on an ntfs volume
  71  * @vol:        mounted ntfs volume on which to allocate the clusters
  72  * @start_vcn:  vcn to use for the first allocated cluster
  73  * @count:      number of clusters to allocate
  74  * @start_lcn:  starting lcn at which to allocate the clusters (or -1 if none)
  75  * @zone:       zone from which to allocate the clusters
  76  *
  77  * Allocate @count clusters preferably starting at cluster @start_lcn or at the
  78  * current allocator position if @start_lcn is -1, on the mounted ntfs volume
  79  * @vol. @zone is either DATA_ZONE for allocation of normal clusters or
  80  * MFT_ZONE for allocation of clusters for the master file table, i.e. the
  81  * $MFT/$DATA attribute.
  82  *
  83  * @start_vcn specifies the vcn of the first allocated cluster.  This makes
  84  * merging the resulting runlist with the old runlist easier.
  85  *
  86  * You need to check the return value with IS_ERR().  If this is false, the
  87  * function was successful and the return value is a runlist describing the
  88  * allocated cluster(s).  If IS_ERR() is true, the function failed and
  89  * PTR_ERR() gives you the error code.
  90  *
  91  * Notes on the allocation algorithm
  92  * =================================
  93  *
  94  * There are two data zones.  First is the area between the end of the mft zone
  95  * and the end of the volume, and second is the area between the start of the
  96  * volume and the start of the mft zone.  On unmodified/standard NTFS 1.x
  97  * volumes, the second data zone does not exist due to the mft zone being
  98  * expanded to cover the start of the volume in order to reserve space for the
  99  * mft bitmap attribute.
 100  *
 101  * This is not the prettiest function but the complexity stems from the need of
 102  * implementing the mft vs data zoned approach and from the fact that we have
 103  * access to the lcn bitmap in portions of up to 8192 bytes at a time, so we
 104  * need to cope with crossing over boundaries of two buffers.  Further, the
 105  * fact that the allocator allows for caller supplied hints as to the location
 106  * of where allocation should begin and the fact that the allocator keeps track
 107  * of where in the data zones the next natural allocation should occur,
 108  * contribute to the complexity of the function.  But it should all be
 109  * worthwhile, because this allocator should: 1) be a full implementation of
 110  * the MFT zone approach used by Windows NT, 2) cause reduction in
 111  * fragmentation, and 3) be speedy in allocations (the code is not optimized
 112  * for speed, but the algorithm is, so further speed improvements are probably
 113  * possible).
 114  *
 115  * FIXME: We should be monitoring cluster allocation and increment the MFT zone
 116  * size dynamically but this is something for the future.  We will just cause
 117  * heavier fragmentation by not doing it and I am not even sure Windows would
 118  * grow the MFT zone dynamically, so it might even be correct not to do this.
 119  * The overhead in doing dynamic MFT zone expansion would be very large and
 120  * unlikely worth the effort. (AIA)
 121  *
 122  * TODO: I have added in double the required zone position pointer wrap around
 123  * logic which can be optimized to having only one of the two logic sets.
 124  * However, having the double logic will work fine, but if we have only one of
 125  * the sets and we get it wrong somewhere, then we get into trouble, so
 126  * removing the duplicate logic requires _very_ careful consideration of _all_
 127  * possible code paths.  So at least for now, I am leaving the double logic -
 128  * better safe than sorry... (AIA)
 129  *
 130  * Locking: - The volume lcn bitmap must be unlocked on entry and is unlocked
 131  *            on return.
 132  *          - This function takes the volume lcn bitmap lock for writing and
 133  *            modifies the bitmap contents.
 134  */
 135 runlist_element *ntfs_cluster_alloc(ntfs_volume *vol, const VCN start_vcn,
 136                 const s64 count, const LCN start_lcn,
 137                 const NTFS_CLUSTER_ALLOCATION_ZONES zone)
 138 {
 139         LCN zone_start, zone_end, bmp_pos, bmp_initial_pos, last_read_pos, lcn;
 140         LCN prev_lcn = 0, prev_run_len = 0, mft_zone_size;
 141         s64 clusters;
 142         struct inode *lcnbmp_vi;
 143         runlist_element *rl = NULL;
 144         struct address_space *mapping;
 145         struct page *page = NULL;
 146         u8 *buf, *byte;
 147         int err = 0, rlpos, rlsize, buf_size;
 148         u8 pass, done_zones, search_zone, need_writeback = 0, bit;
 149
 150         ntfs_debug("Entering for start_vcn 0x%llx, count 0x%llx, start_lcn "
 151                         "0x%llx, zone %s_ZONE.", (unsigned long long)start_vcn,
 152                         (unsigned long long)count,
 153                         (unsigned long long)start_lcn,
 154                         zone == MFT_ZONE ? "MFT" : "DATA");
 155         BUG_ON(!vol);
 156         lcnbmp_vi = vol->lcnbmp_ino;
 157         BUG_ON(!lcnbmp_vi);
 158         BUG_ON(start_vcn < 0);
 159         BUG_ON(count < 0);
 160         BUG_ON(start_lcn < -1);
 161         BUG_ON(zone < FIRST_ZONE);
 162         BUG_ON(zone > LAST_ZONE);
 163
 164         /* Return empty runlist if @count == 0 */
 165         // FIXME: Do we want to just return NULL instead? (AIA)
 166         if (!count) {
 167                 rl = ntfs_malloc_nofs(PAGE_SIZE);
 168                 if (!rl)
 169                         return ERR_PTR(-ENOMEM);
 170                 rl[0].vcn = start_vcn;
 171                 rl[0].lcn = LCN_RL_NOT_MAPPED;
 172                 rl[0].length = 0;
 173                 return rl;
 174         }
 175         /* Take the lcnbmp lock for writing. */
 176         down_write(&vol->lcnbmp_lock);
 177         /*
 178          * If no specific @start_lcn was requested, use the current data zone
 179          * position, otherwise use the requested @start_lcn but make sure it
 180          * lies outside the mft zone.  Also set done_zones to 0 (no zones done)
 181          * and pass depending on whether we are starting inside a zone (1) or
 182          * at the beginning of a zone (2).  If requesting from the MFT_ZONE,
 183          * we either start at the current position within the mft zone or at
 184          * the specified position.  If the latter is out of bounds then we start
 185          * at the beginning of the MFT_ZONE.
 186          */
 187         done_zones = 0;
 188         pass = 1;
 189         /*
 190          * zone_start and zone_end are the current search range.  search_zone
 191          * is 1 for mft zone, 2 for data zone 1 (end of mft zone till end of
 192          * volume) and 4 for data zone 2 (start of volume till start of mft
 193          * zone).
 194          */
 195         zone_start = start_lcn;
 196         if (zone_start < 0) {
 197                 if (zone == DATA_ZONE)
 198                         zone_start = vol->data1_zone_pos;
 199                 else
 200                         zone_start = vol->mft_zone_pos;
 201                 if (!zone_start) {
 202                         /*
 203                          * Zone starts at beginning of volume which means a
 204                          * single pass is sufficient.
 205                          */
 206                         pass = 2;
 207                 }
 208         } else if (zone == DATA_ZONE && zone_start >= vol->mft_zone_start &&
 209                         zone_start < vol->mft_zone_end) {
 210                 zone_start = vol->mft_zone_end;
 211                 /*
 212                  * Starting at beginning of data1_zone which means a single
 213                  * pass in this zone is sufficient.
 214                  */
 215                 pass = 2;
 216         } else if (zone == MFT_ZONE && (zone_start < vol->mft_zone_start ||
 217                         zone_start >= vol->mft_zone_end)) {
 218                 zone_start = vol->mft_lcn;
 219                 if (!vol->mft_zone_end)
 220                         zone_start = 0;
 221                 /*
 222                  * Starting at beginning of volume which means a single pass
 223                  * is sufficient.
 224                  */
 225                 pass = 2;
 226         }
 227         if (zone == MFT_ZONE) {
 228                 zone_end = vol->mft_zone_end;
 229                 search_zone = 1;
 230         } else /* if (zone == DATA_ZONE) */ {
 231                 /* Skip searching the mft zone. */
 232                 done_zones |= 1;
 233                 if (zone_start >= vol->mft_zone_end) {
 234                         zone_end = vol->nr_clusters;
 235                         search_zone = 2;
 236                 } else {
 237                         zone_end = vol->mft_zone_start;
 238                         search_zone = 4;
 239                 }
 240         }
 241         /*
 242          * bmp_pos is the current bit position inside the bitmap.  We use
 243          * bmp_initial_pos to determine whether or not to do a zone switch.
 244          */
 245         bmp_pos = bmp_initial_pos = zone_start;
 246
 247         /* Loop until all clusters are allocated, i.e. clusters == 0. */
 248         clusters = count;
 249         rlpos = rlsize = 0;
 250         mapping = lcnbmp_vi->i_mapping;
 251         while (1) {
 252                 ntfs_debug("Start of outer while loop: done_zones 0x%x, "
 253                                 "search_zone %i, pass %i, zone_start 0x%llx, "
 254                                 "zone_end 0x%llx, bmp_initial_pos 0x%llx, "
 255                                 "bmp_pos 0x%llx, rlpos %i, rlsize %i.",
 256                                 done_zones, search_zone, pass,
 257                                 (unsigned long long)zone_start,
 258                                 (unsigned long long)zone_end,
 259                                 (unsigned long long)bmp_initial_pos,
 260                                 (unsigned long long)bmp_pos, rlpos, rlsize);
 261                 /* Loop until we run out of free clusters. */
 262                 last_read_pos = bmp_pos >> 3;
 263                 ntfs_debug("last_read_pos 0x%llx.",
 264                                 (unsigned long long)last_read_pos);
 265                 if (last_read_pos > lcnbmp_vi->i_size) {
 266                         ntfs_debug("End of attribute reached.  "
 267                                         "Skipping to zone_pass_done.");
 268                         goto zone_pass_done;
 269                 }
 270                 if (likely(page)) {
 271                         if (need_writeback) {
 272                                 ntfs_debug("Marking page dirty.");
 273                                 flush_dcache_page(page);
 274                                 set_page_dirty(page);
 275                                 need_writeback = 0;
 276                         }
 277                         ntfs_unmap_page(page);
 278                 }
 279                 page = ntfs_map_page(mapping, last_read_pos >>
 280                                 PAGE_CACHE_SHIFT);
 281                 if (IS_ERR(page)) {
 282                         err = PTR_ERR(page);
 283                         ntfs_error(vol->sb, "Failed to map page.");
 284                         goto out;
 285                 }
 286                 buf_size = last_read_pos & ~PAGE_CACHE_MASK;
 287                 buf = page_address(page) + buf_size;
 288                 buf_size = PAGE_CACHE_SIZE - buf_size;
 289                 if (unlikely(last_read_pos + buf_size > lcnbmp_vi->i_size))
 290                         buf_size = lcnbmp_vi->i_size - last_read_pos;
 291                 buf_size <<= 3;
 292                 lcn = bmp_pos & 7;
 293                 bmp_pos &= ~7;
 294                 ntfs_debug("Before inner while loop: buf_size %i, lcn 0x%llx, "
 295                                 "bmp_pos 0x%llx, need_writeback %i.", buf_size,
 296                                 (unsigned long long)lcn,
 297                                 (unsigned long long)bmp_pos, need_writeback);
 298                 while (lcn < buf_size && lcn + bmp_pos < zone_end) {
 299                         byte = buf + (lcn >> 3);
 300                         ntfs_debug("In inner while loop: buf_size %i, "
 301                                         "lcn 0x%llx, bmp_pos 0x%llx, "
 302                                         "need_writeback %i, byte ofs 0x%x, "
 303                                         "*byte 0x%x.", buf_size,
 304                                         (unsigned long long)lcn,
 305                                         (unsigned long long)bmp_pos,
 306                                         need_writeback,
 307                                         (unsigned int)(lcn >> 3),
 308                                         (unsigned int)*byte);
 309                         /* Skip full bytes. */
 310                         if (*byte == 0xff) {
 311                                 lcn = (lcn + 8) & ~7;
 312                                 ntfs_debug("Continuing while loop 1.");
 313                                 continue;
 314                         }
 315                         bit = 1 << (lcn & 7);
 316                         ntfs_debug("bit %i.", bit);
 317                         /* If the bit is already set, go onto the next one. */
 318                         if (*byte & bit) {
 319                                 lcn++;
 320                                 ntfs_debug("Continuing while loop 2.");
 321                                 continue;
 322                         }
 323                         /*
 324                          * Allocate more memory if needed, including space for
 325                          * the terminator element.
 326                          * ntfs_malloc_nofs() operates on whole pages only.
 327                          */
 328                         if ((rlpos + 2) * sizeof(*rl) > rlsize) {
 329                                 runlist_element *rl2;
 330
 331                                 ntfs_debug("Reallocating memory.");
 332                                 if (!rl)
 333                                         ntfs_debug("First free bit is at LCN "
 334                                                         "0x%llx.",
 335                                                         (unsigned long long)
 336                                                         (lcn + bmp_pos));
 337                                 rl2 = ntfs_malloc_nofs(rlsize + (int)PAGE_SIZE);
 338                                 if (unlikely(!rl2)) {
 339                                         err = -ENOMEM;
 340                                         ntfs_error(vol->sb, "Failed to "
 341                                                         "allocate memory.");
 342                                         goto out;
 343                                 }
 344                                 memcpy(rl2, rl, rlsize);
 345                                 ntfs_free(rl);
 346                                 rl = rl2;
 347                                 rlsize += PAGE_SIZE;
 348                                 ntfs_debug("Reallocated memory, rlsize 0x%x.",
 349                                                 rlsize);
 350                         }
 351                         /* Allocate the bitmap bit. */
 352                         *byte |= bit;
 353                         /* We need to write this bitmap page to disk. */
 354                         need_writeback = 1;
 355                         ntfs_debug("*byte 0x%x, need_writeback is set.",
 356                                         (unsigned int)*byte);
 357                         /*
 358                          * Coalesce with previous run if adjacent LCNs.
 359                          * Otherwise, append a new run.
 360                          */
 361                         ntfs_debug("Adding run (lcn 0x%llx, len 0x%llx), "
 362                                         "prev_lcn 0x%llx, lcn 0x%llx, "
 363                                         "bmp_pos 0x%llx, prev_run_len 0x%llx, "
 364                                         "rlpos %i.",
 365                                         (unsigned long long)(lcn + bmp_pos),
 366                                         1ULL, (unsigned long long)prev_lcn,
 367                                         (unsigned long long)lcn,
 368                                         (unsigned long long)bmp_pos,
 369                                         (unsigned long long)prev_run_len,
 370                                         rlpos);
 371                         if (prev_lcn == lcn + bmp_pos - prev_run_len && rlpos) {
 372                                 ntfs_debug("Coalescing to run (lcn 0x%llx, "
 373                                                 "len 0x%llx).",
 374                                                 (unsigned long long)
 375                                                 rl[rlpos - 1].lcn,
 376                                                 (unsigned long long)
 377                                                 rl[rlpos - 1].length);
 378                                 rl[rlpos - 1].length = ++prev_run_len;
 379                                 ntfs_debug("Run now (lcn 0x%llx, len 0x%llx), "
 380                                                 "prev_run_len 0x%llx.",
 381                                                 (unsigned long long)
 382                                                 rl[rlpos - 1].lcn,
 383                                                 (unsigned long long)
 384                                                 rl[rlpos - 1].length,
 385                                                 (unsigned long long)
 386                                                 prev_run_len);
 387                         } else {
 388                                 if (likely(rlpos)) {
 389                                         ntfs_debug("Adding new run, (previous "
 390                                                         "run lcn 0x%llx, "
 391                                                         "len 0x%llx).",
 392                                                         (unsigned long long)
 393                                                         rl[rlpos - 1].lcn,
 394                                                         (unsigned long long)
 395                                                         rl[rlpos - 1].length);
 396                                         rl[rlpos].vcn = rl[rlpos - 1].vcn +
 397                                                         prev_run_len;
 398                                 } else {
 399                                         ntfs_debug("Adding new run, is first "
 400                                                         "run.");
 401                                         rl[rlpos].vcn = start_vcn;
 402                                 }
 403                                 rl[rlpos].lcn = prev_lcn = lcn + bmp_pos;
 404                                 rl[rlpos].length = prev_run_len = 1;
 405                                 rlpos++;
 406                         }
 407                         /* Done? */
 408                         if (!--clusters) {
 409                                 LCN tc;
 410                                 /*
 411                                  * Update the current zone position.  Positions
 412                                  * of already scanned zones have been updated
 413                                  * during the respective zone switches.
 414                                  */
 415                                 tc = lcn + bmp_pos + 1;
 416                                 ntfs_debug("Done. Updating current zone "
 417                                                 "position, tc 0x%llx, "
 418                                                 "search_zone %i.",
 419                                                 (unsigned long long)tc,
 420                                                 search_zone);
 421                                 switch (search_zone) {
 422                                 case 1:
 423                                         ntfs_debug("Before checks, "
 424                                                         "vol->mft_zone_pos "
 425                                                         "0x%llx.",
 426                                                         (unsigned long long)
 427                                                         vol->mft_zone_pos);
 428                                         if (tc >= vol->mft_zone_end) {
 429                                                 vol->mft_zone_pos =
 430                                                                 vol->mft_lcn;
 431                                                 if (!vol->mft_zone_end)
 432                                                         vol->mft_zone_pos = 0;
 433                                         } else if ((bmp_initial_pos >=
 434                                                         vol->mft_zone_pos ||
 435                                                         tc > vol->mft_zone_pos)
 436                                                         && tc >= vol->mft_lcn)
 437                                                 vol->mft_zone_pos = tc;
 438                                         ntfs_debug("After checks, "
 439                                                         "vol->mft_zone_pos "
 440                                                         "0x%llx.",
 441                                                         (unsigned long long)
 442                                                         vol->mft_zone_pos);
 443                                         break;
 444                                 case 2:
 445                                         ntfs_debug("Before checks, "
 446                                                         "vol->data1_zone_pos "
 447                                                         "0x%llx.",
 448                                                         (unsigned long long)
 449                                                         vol->data1_zone_pos);
 450                                         if (tc >= vol->nr_clusters)
 451                                                 vol->data1_zone_pos =
 452                                                              vol->mft_zone_end;
 453                                         else if ((bmp_initial_pos >=
 454                                                     vol->data1_zone_pos ||
 455                                                     tc > vol->data1_zone_pos)
 456                                                     && tc >= vol->mft_zone_end)
 457                                                 vol->data1_zone_pos = tc;
 458                                         ntfs_debug("After checks, "
 459                                                         "vol->data1_zone_pos "
 460                                                         "0x%llx.",
 461                                                         (unsigned long long)
 462                                                         vol->data1_zone_pos);
 463                                         break;
 464                                 case 4:
 465                                         ntfs_debug("Before checks, "
 466                                                         "vol->data2_zone_pos "
 467                                                         "0x%llx.",
 468                                                         (unsigned long long)
 469                                                         vol->data2_zone_pos);
 470                                         if (tc >= vol->mft_zone_start)
 471                                                 vol->data2_zone_pos = 0;
 472                                         else if (bmp_initial_pos >=
 473                                                       vol->data2_zone_pos ||
 474                                                       tc > vol->data2_zone_pos)
 475                                                 vol->data2_zone_pos = tc;
 476                                         ntfs_debug("After checks, "
 477                                                         "vol->data2_zone_pos "
 478                                                         "0x%llx.",
 479                                                         (unsigned long long)
 480                                                         vol->data2_zone_pos);
 481                                         break;
 482                                 default:
 483                                         BUG();
 484                                 }
 485                                 ntfs_debug("Finished.  Going to out.");
 486                                 goto out;
 487                         }
 488                         lcn++;
 489                 }
 490                 bmp_pos += buf_size;
 491                 ntfs_debug("After inner while loop: buf_size 0x%x, lcn "
 492                                 "0x%llx, bmp_pos 0x%llx, need_writeback %i.",
 493                                 buf_size, (unsigned long long)lcn,
 494                                 (unsigned long long)bmp_pos, need_writeback);
 495                 if (bmp_pos < zone_end) {
 496                         ntfs_debug("Continuing outer while loop, "
 497                                         "bmp_pos 0x%llx, zone_end 0x%llx.",
 498                                         (unsigned long long)bmp_pos,
 499                                         (unsigned long long)zone_end);
 500                         continue;
 501                 }
 502 zone_pass_done: /* Finished with the current zone pass. */
 503                 ntfs_debug("At zone_pass_done, pass %i.", pass);
 504                 if (pass == 1) {
 505                         /*
 506                          * Now do pass 2, scanning the first part of the zone
 507                          * we omitted in pass 1.
 508                          */
 509                         pass = 2;
 510                         zone_end = zone_start;
 511                         switch (search_zone) {
 512                         case 1: /* mft_zone */
 513                                 zone_start = vol->mft_zone_start;
 514                                 break;
 515                         case 2: /* data1_zone */
 516                                 zone_start = vol->mft_zone_end;
 517                                 break;
 518                         case 4: /* data2_zone */
 519                                 zone_start = 0;
 520                                 break;
 521                         default:
 522                                 BUG();
 523                         }
 524                         /* Sanity check. */
 525                         if (zone_end < zone_start)
 526                                 zone_end = zone_start;
 527                         bmp_pos = zone_start;
 528                         ntfs_debug("Continuing outer while loop, pass 2, "
 529                                         "zone_start 0x%llx, zone_end 0x%llx, "
 530                                         "bmp_pos 0x%llx.",
 531                                         (unsigned long long)zone_start,
 532                                         (unsigned long long)zone_end,
 533                                         (unsigned long long)bmp_pos);
 534                         continue;
 535                 } /* pass == 2 */
 536 done_zones_check:
 537                 ntfs_debug("At done_zones_check, search_zone %i, done_zones "
 538                                 "before 0x%x, done_zones after 0x%x.",
 539                                 search_zone, done_zones,
 540                                 done_zones | search_zone);
 541                 done_zones |= search_zone;
 542                 if (done_zones < 7) {
 543                         ntfs_debug("Switching zone.");
 544                         /* Now switch to the next zone we haven't done yet. */
 545                         pass = 1;
 546                         switch (search_zone) {
 547                         case 1:
 548                                 ntfs_debug("Switching from mft zone to data1 "
 549                                                 "zone.");
 550                                 /* Update mft zone position. */
 551                                 if (rlpos) {
 552                                         LCN tc;
 553
 554                                         ntfs_debug("Before checks, "
 555                                                         "vol->mft_zone_pos "
 556                                                         "0x%llx.",
 557                                                         (unsigned long long)
 558                                                         vol->mft_zone_pos);
 559                                         tc = rl[rlpos - 1].lcn +
 560                                                         rl[rlpos - 1].length;
 561                                         if (tc >= vol->mft_zone_end) {
 562                                                 vol->mft_zone_pos =
 563                                                                 vol->mft_lcn;
 564                                                 if (!vol->mft_zone_end)
 565                                                         vol->mft_zone_pos = 0;
 566                                         } else if ((bmp_initial_pos >=
 567                                                         vol->mft_zone_pos ||
 568                                                         tc > vol->mft_zone_pos)
 569                                                         && tc >= vol->mft_lcn)
 570                                                 vol->mft_zone_pos = tc;
 571                                         ntfs_debug("After checks, "
 572                                                         "vol->mft_zone_pos "
 573                                                         "0x%llx.",
 574                                                         (unsigned long long)
 575                                                         vol->mft_zone_pos);
 576                                 }
 577                                 /* Switch from mft zone to data1 zone. */
 578 switch_to_data1_zone:           search_zone = 2;
 579                                 zone_start = bmp_initial_pos =
 580                                                 vol->data1_zone_pos;
 581                                 zone_end = vol->nr_clusters;
 582                                 if (zone_start == vol->mft_zone_end)
 583                                         pass = 2;
 584                                 if (zone_start >= zone_end) {
 585                                         vol->data1_zone_pos = zone_start =
 586                                                         vol->mft_zone_end;
 587                                         pass = 2;
 588                                 }
 589                                 break;
 590                         case 2:
 591                                 ntfs_debug("Switching from data1 zone to "
 592                                                 "data2 zone.");
 593                                 /* Update data1 zone position. */
 594                                 if (rlpos) {
 595                                         LCN tc;
 596
 597                                         ntfs_debug("Before checks, "
 598                                                         "vol->data1_zone_pos "
 599                                                         "0x%llx.",
 600                                                         (unsigned long long)
 601                                                         vol->data1_zone_pos);
 602                                         tc = rl[rlpos - 1].lcn +
 603                                                         rl[rlpos - 1].length;
 604                                         if (tc >= vol->nr_clusters)
 605                                                 vol->data1_zone_pos =
 606                                                              vol->mft_zone_end;
 607                                         else if ((bmp_initial_pos >=
 608                                                     vol->data1_zone_pos ||
 609                                                     tc > vol->data1_zone_pos)
 610                                                     && tc >= vol->mft_zone_end)
 611                                                 vol->data1_zone_pos = tc;
 612                                         ntfs_debug("After checks, "
 613                                                         "vol->data1_zone_pos "
 614                                                         "0x%llx.",
 615                                                         (unsigned long long)
 616                                                         vol->data1_zone_pos);
 617                                 }
 618                                 /* Switch from data1 zone to data2 zone. */
 619                                 search_zone = 4;
 620                                 zone_start = bmp_initial_pos =
 621                                                 vol->data2_zone_pos;
 622                                 zone_end = vol->mft_zone_start;
 623                                 if (!zone_start)
 624                                         pass = 2;
 625                                 if (zone_start >= zone_end) {
 626                                         vol->data2_zone_pos = zone_start =
 627                                                         bmp_initial_pos = 0;
 628                                         pass = 2;
 629                                 }
 630                                 break;
 631                         case 4:
 632                                 ntfs_debug("Switching from data2 zone to "
 633                                                 "data1 zone.");
 634                                 /* Update data2 zone position. */
 635                                 if (rlpos) {
 636                                         LCN tc;
 637
 638                                         ntfs_debug("Before checks, "
 639                                                         "vol->data2_zone_pos "
 640                                                         "0x%llx.",
 641                                                         (unsigned long long)
 642                                                         vol->data2_zone_pos);
 643                                         tc = rl[rlpos - 1].lcn +
 644                                                         rl[rlpos - 1].length;
 645                                         if (tc >= vol->mft_zone_start)
 646                                                 vol->data2_zone_pos = 0;
 647                                         else if (bmp_initial_pos >=
 648                                                       vol->data2_zone_pos ||
 649                                                       tc > vol->data2_zone_pos)
 650                                                 vol->data2_zone_pos = tc;
 651                                         ntfs_debug("After checks, "
 652                                                         "vol->data2_zone_pos "
 653                                                         "0x%llx.",
 654                                                         (unsigned long long)
 655                                                         vol->data2_zone_pos);
 656                                 }
 657                                 /* Switch from data2 zone to data1 zone. */
 658                                 goto switch_to_data1_zone;
 659                         default:
 660                                 BUG();
 661                         }
 662                         ntfs_debug("After zone switch, search_zone %i, "
 663                                         "pass %i, bmp_initial_pos 0x%llx, "
 664                                         "zone_start 0x%llx, zone_end 0x%llx.",
 665                                         search_zone, pass,
 666                                         (unsigned long long)bmp_initial_pos,
 667                                         (unsigned long long)zone_start,
 668                                         (unsigned long long)zone_end);
 669                         bmp_pos = zone_start;
 670                         if (zone_start == zone_end) {
 671                                 ntfs_debug("Empty zone, going to "
 672                                                 "done_zones_check.");
 673                                 /* Empty zone. Don't bother searching it. */
 674                                 goto done_zones_check;
 675                         }
 676                         ntfs_debug("Continuing outer while loop.");
 677                         continue;
 678                 } /* done_zones == 7 */
 679                 ntfs_debug("All zones are finished.");
 680                 /*
 681                  * All zones are finished!  If DATA_ZONE, shrink mft zone.  If
 682                  * MFT_ZONE, we have really run out of space.
 683                  */
 684                 mft_zone_size = vol->mft_zone_end - vol->mft_zone_start;
 685                 ntfs_debug("vol->mft_zone_start 0x%llx, vol->mft_zone_end "
 686                                 "0x%llx, mft_zone_size 0x%llx.",
 687                                 (unsigned long long)vol->mft_zone_start,
 688                                 (unsigned long long)vol->mft_zone_end,
 689                                 (unsigned long long)mft_zone_size);
 690                 if (zone == MFT_ZONE || mft_zone_size <= 0) {
 691                         ntfs_debug("No free clusters left, going to out.");
 692                         /* Really no more space left on device. */
 693                         err = ENOSPC;
 694                         goto out;
 695                 } /* zone == DATA_ZONE && mft_zone_size > 0 */
 696                 ntfs_debug("Shrinking mft zone.");
 697                 zone_end = vol->mft_zone_end;
 698                 mft_zone_size >>= 1;
 699                 if (mft_zone_size > 0)
 700                         vol->mft_zone_end = vol->mft_zone_start + mft_zone_size;
 701                 else /* mft zone and data2 zone no longer exist. */
 702                         vol->data2_zone_pos = vol->mft_zone_start =
 703                                         vol->mft_zone_end = 0;
 704                 if (vol->mft_zone_pos >= vol->mft_zone_end) {
 705                         vol->mft_zone_pos = vol->mft_lcn;
 706                         if (!vol->mft_zone_end)
 707                                 vol->mft_zone_pos = 0;
 708                 }
 709                 bmp_pos = zone_start = bmp_initial_pos =
 710                                 vol->data1_zone_pos = vol->mft_zone_end;
 711                 search_zone = 2;
 712                 pass = 2;
 713                 done_zones &= ~2;
 714                 ntfs_debug("After shrinking mft zone, mft_zone_size 0x%llx, "
 715                                 "vol->mft_zone_start 0x%llx, "
 716                                 "vol->mft_zone_end 0x%llx, "
 717                                 "vol->mft_zone_pos 0x%llx, search_zone 2, "
 718                                 "pass 2, dones_zones 0x%x, zone_start 0x%llx, "
 719                                 "zone_end 0x%llx, vol->data1_zone_pos 0x%llx, "
 720                                 "continuing outer while loop.",
 721                                 (unsigned long long)mft_zone_size,
 722                                 (unsigned long long)vol->mft_zone_start,
 723                                 (unsigned long long)vol->mft_zone_end,
 724                                 (unsigned long long)vol->mft_zone_pos,
 725                                 done_zones, (unsigned long long)zone_start,
 726                                 (unsigned long long)zone_end,
 727                                 (unsigned long long)vol->data1_zone_pos);
 728         }
 729         ntfs_debug("After outer while loop.");
 730 out:
 731         ntfs_debug("At out.");
 732         /* Add runlist terminator element. */
 733         if (likely(rl)) {
 734                 rl[rlpos].vcn = rl[rlpos - 1].vcn + rl[rlpos - 1].length;
 735                 rl[rlpos].lcn = LCN_RL_NOT_MAPPED;
 736                 rl[rlpos].length = 0;
 737         }
 738         if (likely(page && !IS_ERR(page))) {
 739                 if (need_writeback) {
 740                         ntfs_debug("Marking page dirty.");
 741                         flush_dcache_page(page);
 742                         set_page_dirty(page);
 743                         need_writeback = 0;
 744                 }
 745                 ntfs_unmap_page(page);
 746         }
 747         if (likely(!err)) {
 748                 up_write(&vol->lcnbmp_lock);
 749                 ntfs_debug("Done.");
 750                 return rl;
 751         }
 752         ntfs_error(vol->sb, "Failed to allocate clusters, aborting "
 753                         "(error %i).", err);
 754         if (rl) {
 755                 int err2;
 756
 757                 if (err == ENOSPC)
 758                         ntfs_debug("Not enough space to complete allocation, "
 759                                         "err ENOSPC, first free lcn 0x%llx, "
 760                                         "could allocate up to 0x%llx "
 761                                         "clusters.",
 762                                         (unsigned long long)rl[0].lcn,
 763                                         (unsigned long long)count - clusters);
 764                 /* Deallocate all allocated clusters. */
 765                 ntfs_debug("Attempting rollback...");
 766                 err2 = ntfs_cluster_free_from_rl_nolock(vol, rl);
 767                 if (err2) {
 768                         ntfs_error(vol->sb, "Failed to rollback (error %i).  "
 769                                         "Leaving inconsistent metadata!  "
 770                                         "Unmount and run chkdsk.", err2);
 771                         NVolSetErrors(vol);
 772                 }
 773                 /* Free the runlist. */
 774                 ntfs_free(rl);
 775         } else if (err == ENOSPC)
 776                 ntfs_debug("No space left at all, err = ENOSPC, "
 777                                 "first free lcn = 0x%llx.",
 778                                 (unsigned long long)vol->data1_zone_pos);
 779         up_write(&vol->lcnbmp_lock);
 780         return ERR_PTR(err);
 781 }
 782
 783 /**
 784  * __ntfs_cluster_free - free clusters on an ntfs volume
 785  * @vi:         vfs inode whose runlist describes the clusters to free
 786  * @start_vcn:  vcn in the runlist of @vi at which to start freeing clusters
 787  * @count:      number of clusters to free or -1 for all clusters
 788  * @is_rollback:        if TRUE this is a rollback operation
 789  *
 790  * Free @count clusters starting at the cluster @start_vcn in the runlist
 791  * described by the vfs inode @vi.
 792  *
 793  * If @count is -1, all clusters from @start_vcn to the end of the runlist are
 794  * deallocated.  Thus, to completely free all clusters in a runlist, use
 795  * @start_vcn = 0 and @count = -1.
 796  *
 797  * @is_rollback should always be FALSE, it is for internal use to rollback
 798  * errors.  You probably want to use ntfs_cluster_free() instead.
 799  *
 800  * Note, ntfs_cluster_free() does not modify the runlist at all, so the caller
 801  * has to deal with it later.
 802  *
 803  * Return the number of deallocated clusters (not counting sparse ones) on
 804  * success and -errno on error.
 805  *
 806  * Locking: - The runlist described by @vi must be unlocked on entry and is
 807  *            unlocked on return.
 808  *          - This function takes the runlist lock of @vi for reading and
 809  *            sometimes for writing and sometimes modifies the runlist.
 810  *          - The volume lcn bitmap must be unlocked on entry and is unlocked
 811  *            on return.
 812  *          - This function takes the volume lcn bitmap lock for writing and
 813  *            modifies the bitmap contents.
 814  */
 815 s64 __ntfs_cluster_free(struct inode *vi, const VCN start_vcn, s64 count,
 816                 const BOOL is_rollback)
 817 {
 818         s64 delta, to_free, total_freed, real_freed;
 819         ntfs_inode *ni;
 820         ntfs_volume *vol;
 821         struct inode *lcnbmp_vi;
 822         runlist_element *rl;
 823         int err;
 824
 825         BUG_ON(!vi);
 826         ntfs_debug("Entering for i_ino 0x%lx, start_vcn 0x%llx, count "
 827                         "0x%llx.%s", vi->i_ino, (unsigned long long)start_vcn,
 828                         (unsigned long long)count,
 829                         is_rollback ? " (rollback)" : "");
 830         ni = NTFS_I(vi);
 831         vol = ni->vol;
 832         lcnbmp_vi = vol->lcnbmp_ino;
 833         BUG_ON(!lcnbmp_vi);
 834         BUG_ON(start_vcn < 0);
 835         BUG_ON(count < -1);
 836         /*
 837          * Lock the lcn bitmap for writing but only if not rolling back.  We
 838          * must hold the lock all the way including through rollback otherwise
 839          * rollback is not possible because once we have cleared a bit and
 840          * dropped the lock, anyone could have set the bit again, thus
 841          * allocating the cluster for another use.
 842          */
 843         if (likely(!is_rollback))
 844                 down_write(&vol->lcnbmp_lock);
 845
 846         total_freed = real_freed = 0;
 847
 848         /* This returns with ni->runlist locked for reading on success. */
 849         rl = ntfs_find_vcn(ni, start_vcn, FALSE);
 850         if (IS_ERR(rl)) {
 851                 if (!is_rollback)
 852                         ntfs_error(vol->sb, "Failed to find first runlist "
 853                                         "element (error %li), aborting.",
 854                                         PTR_ERR(rl));
 855                 err = PTR_ERR(rl);
 856                 goto err_out;
 857         }
 858         if (unlikely(rl->lcn < (LCN)LCN_HOLE)) {
 859                 if (!is_rollback)
 860                         ntfs_error(vol->sb, "First runlist element has "
 861                                         "invalid lcn, aborting.");
 862                 err = -EIO;
 863                 goto unl_err_out;
 864         }
 865         /* Find the starting cluster inside the run that needs freeing. */
 866         delta = start_vcn - rl->vcn;
 867
 868         /* The number of clusters in this run that need freeing. */
 869         to_free = rl->length - delta;
 870         if (count >= 0 && to_free > count)
 871                 to_free = count;
 872
 873         if (likely(rl->lcn >= 0)) {
 874                 /* Do the actual freeing of the clusters in this run. */
 875                 err = ntfs_bitmap_set_bits_in_run(lcnbmp_vi, rl->lcn + delta,
 876                                 to_free, likely(!is_rollback) ? 0 : 1);
 877                 if (unlikely(err)) {
 878                         if (!is_rollback)
 879                                 ntfs_error(vol->sb, "Failed to clear first run "
 880                                                 "(error %i), aborting.", err);
 881                         goto unl_err_out;
 882                 }
 883                 /* We have freed @to_free real clusters. */
 884                 real_freed = to_free;
 885         };
 886         /* Go to the next run and adjust the number of clusters left to free. */
 887         ++rl;
 888         if (count >= 0)
 889                 count -= to_free;
 890
 891         /* Keep track of the total "freed" clusters, including sparse ones. */
 892         total_freed = to_free;
 893         /*
 894          * Loop over the remaining runs, using @count as a capping value, and
 895          * free them.
 896          */
 897         for (; rl->length && count != 0; ++rl) {
 898                 if (unlikely(rl->lcn < (LCN)LCN_HOLE)) {
 899                         VCN vcn;
 900
 901                         /*
 902                          * Attempt to map runlist, dropping runlist lock for
 903                          * the duration.
 904                          */
 905                         up_read(&ni->runlist.lock);
 906                         vcn = rl->vcn;
 907                         err = ntfs_map_runlist(ni, vcn);
 908                         if (err) {
 909                                 if (!is_rollback)
 910                                         ntfs_error(vol->sb, "Failed to map "
 911                                                         "runlist fragment.");
 912                                 if (err == -EINVAL || err == -ENOENT)
 913                                         err = -EIO;
 914                                 goto err_out;
 915                         }
 916                         /*
 917                          * This returns with ni->runlist locked for reading on
 918                          * success.
 919                          */
 920                         rl = ntfs_find_vcn(ni, vcn, FALSE);
 921                         if (IS_ERR(rl)) {
 922                                 err = PTR_ERR(rl);
 923                                 if (!is_rollback)
 924                                         ntfs_error(vol->sb, "Failed to find "
 925                                                         "subsequent runlist "
 926                                                         "element.");
 927                                 goto err_out;
 928                         }
 929                         if (unlikely(rl->lcn < (LCN)LCN_HOLE)) {
 930                                 if (!is_rollback)
 931                                         ntfs_error(vol->sb, "Runlist element "
 932                                                         "has invalid lcn "
 933                                                         "(0x%llx).",
 934                                                         (unsigned long long)
 935                                                         rl->lcn);
 936                                 err = -EIO;
 937                                 goto unl_err_out;
 938                         }
 939                 }
 940                 /* The number of clusters in this run that need freeing. */
 941                 to_free = rl->length;
 942                 if (count >= 0 && to_free > count)
 943                         to_free = count;
 944
 945                 if (likely(rl->lcn >= 0)) {
 946                         /* Do the actual freeing of the clusters in the run. */
 947                         err = ntfs_bitmap_set_bits_in_run(lcnbmp_vi, rl->lcn,
 948                                         to_free, likely(!is_rollback) ? 0 : 1);
 949                         if (unlikely(err)) {
 950                                 if (!is_rollback)
 951                                         ntfs_error(vol->sb, "Failed to clear "
 952                                                         "subsequent run.");
 953                                 goto unl_err_out;
 954                         }
 955                         /* We have freed @to_free real clusters. */
 956                         real_freed += to_free;
 957                 }
 958                 /* Adjust the number of clusters left to free. */
 959                 if (count >= 0)
 960                         count -= to_free;
 961
 962                 /* Update the total done clusters. */
 963                 total_freed += to_free;
 964         }
 965         up_read(&ni->runlist.lock);
 966         if (likely(!is_rollback))
 967                 up_write(&vol->lcnbmp_lock);
 968
 969         BUG_ON(count > 0);
 970
 971         /* We are done.  Return the number of actually freed clusters. */
 972         ntfs_debug("Done.");
 973         return real_freed;
 974 unl_err_out:
 975         up_read(&ni->runlist.lock);
 976 err_out:
 977         if (is_rollback)
 978                 return err;
 979         /* If no real clusters were freed, no need to rollback. */
 980         if (!real_freed) {
 981                 up_write(&vol->lcnbmp_lock);
 982                 return err;
 983         }
 984         /*
 985          * Attempt to rollback and if that succeeds just return the error code.
 986          * If rollback fails, set the volume errors flag, emit an error
 987          * message, and return the error code.
 988          */
 989         delta = __ntfs_cluster_free(vi, start_vcn, total_freed, TRUE);
 990         if (delta < 0) {
 991                 ntfs_error(vol->sb, "Failed to rollback (error %i).  Leaving "
 992                                 "inconsistent metadata!  Unmount and run "
 993                                 "chkdsk.", (int)delta);
 994                 NVolSetErrors(vol);
 995         }
 996         up_write(&vol->lcnbmp_lock);
 997         ntfs_error(vol->sb, "Aborting (error %i).", err);
 998         return err;
 999 }
1000
1001 #endif /* NTFS_RW */