fs/ext3/balloc.c

   1 /*
   2  *  linux/fs/ext3/balloc.c
   3  *
   4  * Copyright (C) 1992, 1993, 1994, 1995
   5  * Remy Card (card@masi.ibp.fr)
   6  * Laboratoire MASI - Institut Blaise Pascal
   7  * Universite Pierre et Marie Curie (Paris VI)
   8  *
   9  *  Enhanced block allocation by Stephen Tweedie (sct@redhat.com), 1993
  10  *  Big-endian to little-endian byte-swapping/bitmaps by
  11  *        David S. Miller (davem@caip.rutgers.edu), 1995
  12  */
  13
  14 #include <linux/config.h>
  15 #include <linux/time.h>
  16 #include <linux/fs.h>
  17 #include <linux/jbd.h>
  18 #include <linux/ext3_fs.h>
  19 #include <linux/ext3_jbd.h>
  20 #include <linux/quotaops.h>
  21 #include <linux/buffer_head.h>
  22
  23 /*
  24  * balloc.c contains the blocks allocation and deallocation routines
  25  */
  26
  27 /*
  28  * The free blocks are managed by bitmaps.  A file system contains several
  29  * blocks groups.  Each group contains 1 bitmap block for blocks, 1 bitmap
  30  * block for inodes, N blocks for the inode table and data blocks.
  31  *
  32  * The file system contains group descriptors which are located after the
  33  * super block.  Each descriptor contains the number of the bitmap block and
  34  * the free blocks count in the block.  The descriptors are loaded in memory
  35  * when a file system is mounted (see ext3_read_super).
  36  */
  37
  38
  39 #define in_range(b, first, len) ((b) >= (first) && (b) <= (first) + (len) - 1)
  40
  41 struct ext3_group_desc * ext3_get_group_desc(struct super_block * sb,
  42                                              unsigned int block_group,
  43                                              struct buffer_head ** bh)
  44 {
  45         unsigned long group_desc;
  46         unsigned long desc;
  47         struct ext3_group_desc * gdp;
  48
  49         if (block_group >= EXT3_SB(sb)->s_groups_count) {
  50                 ext3_error (sb, "ext3_get_group_desc",
  51                             "block_group >= groups_count - "
  52                             "block_group = %d, groups_count = %lu",
  53                             block_group, EXT3_SB(sb)->s_groups_count);
  54
  55                 return NULL;
  56         }
  57
  58         group_desc = block_group / EXT3_DESC_PER_BLOCK(sb);
  59         desc = block_group % EXT3_DESC_PER_BLOCK(sb);
  60         if (!EXT3_SB(sb)->s_group_desc[group_desc]) {
  61                 ext3_error (sb, "ext3_get_group_desc",
  62                             "Group descriptor not loaded - "
  63                             "block_group = %d, group_desc = %lu, desc = %lu",
  64                              block_group, group_desc, desc);
  65                 return NULL;
  66         }
  67
  68         gdp = (struct ext3_group_desc *)
  69               EXT3_SB(sb)->s_group_desc[group_desc]->b_data;
  70         if (bh)
  71                 *bh = EXT3_SB(sb)->s_group_desc[group_desc];
  72         return gdp + desc;
  73 }
  74
  75 /*
  76  * Read the bitmap for a given block_group, reading into the specified
  77  * slot in the superblock's bitmap cache.
  78  *
  79  * Return buffer_head on success or NULL in case of failure.
  80  */
  81 static struct buffer_head *
  82 read_block_bitmap(struct super_block *sb, unsigned int block_group)
  83 {
  84         struct ext3_group_desc * desc;
  85         struct buffer_head * bh = NULL;
  86
  87         desc = ext3_get_group_desc (sb, block_group, NULL);
  88         if (!desc)
  89                 goto error_out;
  90         bh = sb_bread(sb, le32_to_cpu(desc->bg_block_bitmap));
  91         if (!bh)
  92                 ext3_error (sb, "read_block_bitmap",
  93                             "Cannot read block bitmap - "
  94                             "block_group = %d, block_bitmap = %lu",
  95                             block_group, (unsigned long) desc->bg_block_bitmap);
  96 error_out:
  97         return bh;
  98 }
  99
 100 /* Free given blocks, update quota and i_blocks field */
 101 void ext3_free_blocks (handle_t *handle, struct inode * inode,
 102                         unsigned long block, unsigned long count)
 103 {
 104         struct buffer_head *bitmap_bh = NULL;
 105         struct buffer_head *gd_bh;
 106         unsigned long block_group;
 107         unsigned long bit;
 108         unsigned long i;
 109         unsigned long overflow;
 110         struct super_block * sb;
 111         struct ext3_group_desc * gdp;
 112         struct ext3_super_block * es;
 113         struct ext3_sb_info *sbi;
 114         int err = 0, ret;
 115         int dquot_freed_blocks = 0;
 116
 117         sb = inode->i_sb;
 118         if (!sb) {
 119                 printk ("ext3_free_blocks: nonexistent device");
 120                 return;
 121         }
 122         sbi = EXT3_SB(sb);
 123         es = EXT3_SB(sb)->s_es;
 124         if (block < le32_to_cpu(es->s_first_data_block) ||
 125             block + count < block ||
 126             block + count > le32_to_cpu(es->s_blocks_count)) {
 127                 ext3_error (sb, "ext3_free_blocks",
 128                             "Freeing blocks not in datazone - "
 129                             "block = %lu, count = %lu", block, count);
 130                 goto error_return;
 131         }
 132
 133         ext3_debug ("freeing block %lu\n", block);
 134
 135 do_more:
 136         overflow = 0;
 137         block_group = (block - le32_to_cpu(es->s_first_data_block)) /
 138                       EXT3_BLOCKS_PER_GROUP(sb);
 139         bit = (block - le32_to_cpu(es->s_first_data_block)) %
 140                       EXT3_BLOCKS_PER_GROUP(sb);
 141         /*
 142          * Check to see if we are freeing blocks across a group
 143          * boundary.
 144          */
 145         if (bit + count > EXT3_BLOCKS_PER_GROUP(sb)) {
 146                 overflow = bit + count - EXT3_BLOCKS_PER_GROUP(sb);
 147                 count -= overflow;
 148         }
 149         brelse(bitmap_bh);
 150         bitmap_bh = read_block_bitmap(sb, block_group);
 151         if (!bitmap_bh)
 152                 goto error_return;
 153         gdp = ext3_get_group_desc (sb, block_group, &gd_bh);
 154         if (!gdp)
 155                 goto error_return;
 156
 157         if (in_range (le32_to_cpu(gdp->bg_block_bitmap), block, count) ||
 158             in_range (le32_to_cpu(gdp->bg_inode_bitmap), block, count) ||
 159             in_range (block, le32_to_cpu(gdp->bg_inode_table),
 160                       EXT3_SB(sb)->s_itb_per_group) ||
 161             in_range (block + count - 1, le32_to_cpu(gdp->bg_inode_table),
 162                       EXT3_SB(sb)->s_itb_per_group))
 163                 ext3_error (sb, "ext3_free_blocks",
 164                             "Freeing blocks in system zones - "
 165                             "Block = %lu, count = %lu",
 166                             block, count);
 167
 168         /*
 169          * We are about to start releasing blocks in the bitmap,
 170          * so we need undo access.
 171          */
 172         /* @@@ check errors */
 173         BUFFER_TRACE(bitmap_bh, "getting undo access");
 174         err = ext3_journal_get_undo_access(handle, bitmap_bh, NULL);
 175         if (err)
 176                 goto error_return;
 177
 178         /*
 179          * We are about to modify some metadata.  Call the journal APIs
 180          * to unshare ->b_data if a currently-committing transaction is
 181          * using it
 182          */
 183         BUFFER_TRACE(gd_bh, "get_write_access");
 184         err = ext3_journal_get_write_access(handle, gd_bh);
 185         if (err)
 186                 goto error_return;
 187
 188         jbd_lock_bh_state(bitmap_bh);
 189
 190         for (i = 0; i < count; i++) {
 191                 /*
 192                  * An HJ special.  This is expensive...
 193                  */
 194 #ifdef CONFIG_JBD_DEBUG
 195                 jbd_unlock_bh_state(bitmap_bh);
 196                 {
 197                         struct buffer_head *debug_bh;
 198                         debug_bh = sb_find_get_block(sb, block + i);
 199                         if (debug_bh) {
 200                                 BUFFER_TRACE(debug_bh, "Deleted!");
 201                                 if (!bh2jh(bitmap_bh)->b_committed_data)
 202                                         BUFFER_TRACE(debug_bh,
 203                                                 "No commited data in bitmap");
 204                                 BUFFER_TRACE2(debug_bh, bitmap_bh, "bitmap");
 205                                 __brelse(debug_bh);
 206                         }
 207                 }
 208                 jbd_lock_bh_state(bitmap_bh);
 209 #endif
 210                 /* @@@ This prevents newly-allocated data from being
 211                  * freed and then reallocated within the same
 212                  * transaction.
 213                  *
 214                  * Ideally we would want to allow that to happen, but to
 215                  * do so requires making journal_forget() capable of
 216                  * revoking the queued write of a data block, which
 217                  * implies blocking on the journal lock.  *forget()
 218                  * cannot block due to truncate races.
 219                  *
 220                  * Eventually we can fix this by making journal_forget()
 221                  * return a status indicating whether or not it was able
 222                  * to revoke the buffer.  On successful revoke, it is
 223                  * safe not to set the allocation bit in the committed
 224                  * bitmap, because we know that there is no outstanding
 225                  * activity on the buffer any more and so it is safe to
 226                  * reallocate it.
 227                  */
 228                 BUFFER_TRACE(bitmap_bh, "set in b_committed_data");
 229                 J_ASSERT_BH(bitmap_bh,
 230                                 bh2jh(bitmap_bh)->b_committed_data != NULL);
 231                 ext3_set_bit_atomic(sb_bgl_lock(sbi, block_group), bit + i,
 232                                 bh2jh(bitmap_bh)->b_committed_data);
 233
 234                 /*
 235                  * We clear the bit in the bitmap after setting the committed
 236                  * data bit, because this is the reverse order to that which
 237                  * the allocator uses.
 238                  */
 239                 BUFFER_TRACE(bitmap_bh, "clear bit");
 240                 if (!ext3_clear_bit_atomic(sb_bgl_lock(sbi, block_group),
 241                                                 bit + i, bitmap_bh->b_data)) {
 242                         jbd_unlock_bh_state(bitmap_bh);
 243                         ext3_error(sb, __FUNCTION__,
 244                                 "bit already cleared for block %lu", block + i);
 245                         jbd_lock_bh_state(bitmap_bh);
 246                         BUFFER_TRACE(bitmap_bh, "bit already cleared");
 247                 } else {
 248                         dquot_freed_blocks++;
 249                 }
 250         }
 251         jbd_unlock_bh_state(bitmap_bh);
 252
 253         spin_lock(sb_bgl_lock(sbi, block_group));
 254         gdp->bg_free_blocks_count =
 255                 cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count) +
 256                         dquot_freed_blocks);
 257         spin_unlock(sb_bgl_lock(sbi, block_group));
 258         percpu_counter_mod(&sbi->s_freeblocks_counter, count);
 259
 260         /* We dirtied the bitmap block */
 261         BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
 262         err = ext3_journal_dirty_metadata(handle, bitmap_bh);
 263
 264         /* And the group descriptor block */
 265         BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
 266         ret = ext3_journal_dirty_metadata(handle, gd_bh);
 267         if (!err) err = ret;
 268
 269         if (overflow && !err) {
 270                 block += count;
 271                 count = overflow;
 272                 goto do_more;
 273         }
 274         sb->s_dirt = 1;
 275 error_return:
 276         brelse(bitmap_bh);
 277         ext3_std_error(sb, err);
 278         if (dquot_freed_blocks)
 279                 DQUOT_FREE_BLOCK(inode, dquot_freed_blocks);
 280         return;
 281 }
 282
 283 /*
 284  * For ext3 allocations, we must not reuse any blocks which are
 285  * allocated in the bitmap buffer's "last committed data" copy.  This
 286  * prevents deletes from freeing up the page for reuse until we have
 287  * committed the delete transaction.
 288  *
 289  * If we didn't do this, then deleting something and reallocating it as
 290  * data would allow the old block to be overwritten before the
 291  * transaction committed (because we force data to disk before commit).
 292  * This would lead to corruption if we crashed between overwriting the
 293  * data and committing the delete.
 294  *
 295  * @@@ We may want to make this allocation behaviour conditional on
 296  * data-writes at some point, and disable it for metadata allocations or
 297  * sync-data inodes.
 298  */
 299 static inline int ext3_test_allocatable(int nr, struct buffer_head *bh)
 300 {
 301         int ret;
 302         struct journal_head *jh = bh2jh(bh);
 303
 304         if (ext3_test_bit(nr, bh->b_data))
 305                 return 0;
 306
 307         jbd_lock_bh_state(bh);
 308         if (!jh->b_committed_data)
 309                 ret = 1;
 310         else
 311                 ret = !ext3_test_bit(nr, jh->b_committed_data);
 312         jbd_unlock_bh_state(bh);
 313         return ret;
 314 }
 315
 316 /*
 317  * Find an allocatable block in a bitmap.  We honour both the bitmap and
 318  * its last-committed copy (if that exists), and perform the "most
 319  * appropriate allocation" algorithm of looking for a free block near
 320  * the initial goal; then for a free byte somewhere in the bitmap; then
 321  * for any free bit in the bitmap.
 322  */
 323 static int
 324 find_next_usable_block(int start, struct buffer_head *bh, int maxblocks)
 325 {
 326         int here, next;
 327         char *p, *r;
 328         struct journal_head *jh = bh2jh(bh);
 329
 330         if (start > 0) {
 331                 /*
 332                  * The goal was occupied; search forward for a free
 333                  * block within the next XX blocks.
 334                  *
 335                  * end_goal is more or less random, but it has to be
 336                  * less than EXT3_BLOCKS_PER_GROUP. Aligning up to the
 337                  * next 64-bit boundary is simple..
 338                  */
 339                 int end_goal = (start + 63) & ~63;
 340                 here = ext3_find_next_zero_bit(bh->b_data, end_goal, start);
 341                 if (here < end_goal && ext3_test_allocatable(here, bh))
 342                         return here;
 343                 ext3_debug("Bit not found near goal\n");
 344         }
 345
 346         here = start;
 347         if (here < 0)
 348                 here = 0;
 349
 350         p = ((char *)bh->b_data) + (here >> 3);
 351         r = memscan(p, 0, (maxblocks - here + 7) >> 3);
 352         next = (r - ((char *)bh->b_data)) << 3;
 353
 354         if (next < maxblocks && ext3_test_allocatable(next, bh))
 355                 return next;
 356
 357         /*
 358          * The bitmap search --- search forward alternately through the actual
 359          * bitmap and the last-committed copy until we find a bit free in
 360          * both
 361          */
 362         while (here < maxblocks) {
 363                 next = ext3_find_next_zero_bit(bh->b_data, maxblocks, here);
 364                 if (next >= maxblocks)
 365                         return -1;
 366                 if (ext3_test_allocatable(next, bh))
 367                         return next;
 368                 jbd_lock_bh_state(bh);
 369                 if (jh->b_committed_data)
 370                         here = ext3_find_next_zero_bit(jh->b_committed_data,
 371                                                         maxblocks, next);
 372                 jbd_unlock_bh_state(bh);
 373         }
 374         return -1;
 375 }
 376
 377 /*
 378  * We think we can allocate this block in this bitmap.  Try to set the bit.
 379  * If that succeeds then check that nobody has allocated and then freed the
 380  * block since we saw that is was not marked in b_committed_data.  If it _was_
 381  * allocated and freed then clear the bit in the bitmap again and return
 382  * zero (failure).
 383  */
 384 static inline int
 385 claim_block(spinlock_t *lock, int block, struct buffer_head *bh)
 386 {
 387         struct journal_head *jh = bh2jh(bh);
 388         int ret;
 389
 390         if (ext3_set_bit_atomic(lock, block, bh->b_data))
 391                 return 0;
 392         jbd_lock_bh_state(bh);
 393         if (jh->b_committed_data && ext3_test_bit(block,jh->b_committed_data)) {
 394                 ext3_clear_bit_atomic(lock, block, bh->b_data);
 395                 ret = 0;
 396         } else {
 397                 ret = 1;
 398         }
 399         jbd_unlock_bh_state(bh);
 400         return ret;
 401 }
 402
 403 /*
 404  * If we failed to allocate the desired block then we may end up crossing to a
 405  * new bitmap.  In that case we must release write access to the old one via
 406  * ext3_journal_release_buffer(), else we'll run out of credits.
 407  */
 408 static int
 409 ext3_try_to_allocate(struct super_block *sb, handle_t *handle, int group,
 410                 struct buffer_head *bitmap_bh, int goal, int *errp)
 411 {
 412         int i;
 413         int fatal;
 414         int credits = 0;
 415
 416         *errp = 0;
 417
 418         /*
 419          * Make sure we use undo access for the bitmap, because it is critical
 420          * that we do the frozen_data COW on bitmap buffers in all cases even
 421          * if the buffer is in BJ_Forget state in the committing transaction.
 422          */
 423         BUFFER_TRACE(bitmap_bh, "get undo access for new block");
 424         fatal = ext3_journal_get_undo_access(handle, bitmap_bh, &credits);
 425         if (fatal) {
 426                 *errp = fatal;
 427                 goto fail;
 428         }
 429
 430 repeat:
 431         if (goal < 0 || !ext3_test_allocatable(goal, bitmap_bh)) {
 432                 goal = find_next_usable_block(goal, bitmap_bh,
 433                                         EXT3_BLOCKS_PER_GROUP(sb));
 434                 if (goal < 0)
 435                         goto fail_access;
 436
 437                 for (i = 0; i < 7 && goal > 0 &&
 438                                 ext3_test_allocatable(goal - 1, bitmap_bh);
 439                         i++, goal--);
 440         }
 441
 442         if (!claim_block(sb_bgl_lock(EXT3_SB(sb), group), goal, bitmap_bh)) {
 443                 /*
 444                  * The block was allocated by another thread, or it was
 445                  * allocated and then freed by another thread
 446                  */
 447                 goal++;
 448                 if (goal >= EXT3_BLOCKS_PER_GROUP(sb))
 449                         goto fail_access;
 450                 goto repeat;
 451         }
 452
 453         BUFFER_TRACE(bitmap_bh, "journal_dirty_metadata for bitmap block");
 454         fatal = ext3_journal_dirty_metadata(handle, bitmap_bh);
 455         if (fatal) {
 456                 *errp = fatal;
 457                 goto fail;
 458         }
 459         return goal;
 460
 461 fail_access:
 462         BUFFER_TRACE(bitmap_bh, "journal_release_buffer");
 463         ext3_journal_release_buffer(handle, bitmap_bh, credits);
 464 fail:
 465         return -1;
 466 }
 467
 468 static int ext3_has_free_blocks(struct ext3_sb_info *sbi)
 469 {
 470         int free_blocks, root_blocks;
 471
 472         free_blocks = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
 473         root_blocks = le32_to_cpu(sbi->s_es->s_r_blocks_count);
 474         if (free_blocks < root_blocks + 1 && !capable(CAP_SYS_RESOURCE) &&
 475                 sbi->s_resuid != current->fsuid &&
 476                 (sbi->s_resgid == 0 || !in_group_p (sbi->s_resgid))) {
 477                 return 0;
 478         }
 479         return 1;
 480 }
 481
 482 /*
 483  * ext3_should_retry_alloc() is called when ENOSPC is returned, and if
 484  * it is profitable to retry the operation, this function will wait
 485  * for the current or commiting transaction to complete, and then
 486  * return TRUE.
 487  */
 488 int ext3_should_retry_alloc(struct super_block *sb, int *retries)
 489 {
 490         if (!ext3_has_free_blocks(EXT3_SB(sb)) || (*retries)++ > 3)
 491                 return 0;
 492
 493         jbd_debug(1, "%s: retrying operation after ENOSPC\n", sb->s_id);
 494
 495         return journal_force_commit_nested(EXT3_SB(sb)->s_journal);
 496 }
 497
 498 /*
 499  * ext3_new_block uses a goal block to assist allocation.  If the goal is
 500  * free, or there is a free block within 32 blocks of the goal, that block
 501  * is allocated.  Otherwise a forward search is made for a free block; within
 502  * each block group the search first looks for an entire free byte in the block
 503  * bitmap, and then for any free bit if that fails.
 504  * This function also updates quota and i_blocks field.
 505  */
 506 int
 507 ext3_new_block(handle_t *handle, struct inode *inode, unsigned long goal,
 508                 u32 *prealloc_count, u32 *prealloc_block, int *errp)
 509 {
 510         struct buffer_head *bitmap_bh = NULL;   /* bh */
 511         struct buffer_head *gdp_bh;             /* bh2 */
 512         int group_no;                           /* i */
 513         int ret_block;                          /* j */
 514         int bgi;                                /* blockgroup iteration index */
 515         int target_block;                       /* tmp */
 516         int fatal = 0, err;
 517         int performed_allocation = 0;
 518         int free_blocks;
 519         struct super_block *sb;
 520         struct ext3_group_desc *gdp;
 521         struct ext3_super_block *es;
 522         struct ext3_sb_info *sbi;
 523 #ifdef EXT3FS_DEBUG
 524         static int goal_hits, goal_attempts;
 525 #endif
 526         *errp = -ENOSPC;
 527         sb = inode->i_sb;
 528         if (!sb) {
 529                 printk("ext3_new_block: nonexistent device");
 530                 return 0;
 531         }
 532
 533         /*
 534          * Check quota for allocation of this block.
 535          */
 536         if (DQUOT_ALLOC_BLOCK(inode, 1)) {
 537                 *errp = -EDQUOT;
 538                 return 0;
 539         }
 540
 541         sbi = EXT3_SB(sb);
 542         es = EXT3_SB(sb)->s_es;
 543         ext3_debug("goal=%lu.\n", goal);
 544
 545         if (!ext3_has_free_blocks(sbi)) {
 546                 *errp = -ENOSPC;
 547                 goto out;
 548         }
 549
 550         /*
 551          * First, test whether the goal block is free.
 552          */
 553         if (goal < le32_to_cpu(es->s_first_data_block) ||
 554             goal >= le32_to_cpu(es->s_blocks_count))
 555                 goal = le32_to_cpu(es->s_first_data_block);
 556         group_no = (goal - le32_to_cpu(es->s_first_data_block)) /
 557                         EXT3_BLOCKS_PER_GROUP(sb);
 558         gdp = ext3_get_group_desc(sb, group_no, &gdp_bh);
 559         if (!gdp)
 560                 goto io_error;
 561
 562         free_blocks = le16_to_cpu(gdp->bg_free_blocks_count);
 563         if (free_blocks > 0) {
 564                 ret_block = ((goal - le32_to_cpu(es->s_first_data_block)) %
 565                                 EXT3_BLOCKS_PER_GROUP(sb));
 566                 bitmap_bh = read_block_bitmap(sb, group_no);
 567                 if (!bitmap_bh)
 568                         goto io_error;
 569                 ret_block = ext3_try_to_allocate(sb, handle, group_no,
 570                                         bitmap_bh, ret_block, &fatal);
 571                 if (fatal)
 572                         goto out;
 573                 if (ret_block >= 0)
 574                         goto allocated;
 575         }
 576
 577         /*
 578          * Now search the rest of the groups.  We assume that
 579          * i and gdp correctly point to the last group visited.
 580          */
 581         for (bgi = 0; bgi < EXT3_SB(sb)->s_groups_count; bgi++) {
 582                 group_no++;
 583                 if (group_no >= EXT3_SB(sb)->s_groups_count)
 584                         group_no = 0;
 585                 gdp = ext3_get_group_desc(sb, group_no, &gdp_bh);
 586                 if (!gdp) {
 587                         *errp = -EIO;
 588                         goto out;
 589                 }
 590                 free_blocks = le16_to_cpu(gdp->bg_free_blocks_count);
 591                 if (free_blocks <= 0)
 592                         continue;
 593
 594                 brelse(bitmap_bh);
 595                 bitmap_bh = read_block_bitmap(sb, group_no);
 596                 if (!bitmap_bh)
 597                         goto io_error;
 598                 ret_block = ext3_try_to_allocate(sb, handle, group_no,
 599                                                 bitmap_bh, -1, &fatal);
 600                 if (fatal)
 601                         goto out;
 602                 if (ret_block >= 0)
 603                         goto allocated;
 604         }
 605
 606         /* No space left on the device */
 607         *errp = -ENOSPC;
 608         goto out;
 609
 610 allocated:
 611
 612         ext3_debug("using block group %d(%d)\n",
 613                         group_no, gdp->bg_free_blocks_count);
 614
 615         BUFFER_TRACE(gdp_bh, "get_write_access");
 616         fatal = ext3_journal_get_write_access(handle, gdp_bh);
 617         if (fatal)
 618                 goto out;
 619
 620         target_block = ret_block + group_no * EXT3_BLOCKS_PER_GROUP(sb)
 621                                 + le32_to_cpu(es->s_first_data_block);
 622
 623         if (target_block == le32_to_cpu(gdp->bg_block_bitmap) ||
 624             target_block == le32_to_cpu(gdp->bg_inode_bitmap) ||
 625             in_range(target_block, le32_to_cpu(gdp->bg_inode_table),
 626                       EXT3_SB(sb)->s_itb_per_group))
 627                 ext3_error(sb, "ext3_new_block",
 628                             "Allocating block in system zone - "
 629                             "block = %u", target_block);
 630
 631         performed_allocation = 1;
 632
 633 #ifdef CONFIG_JBD_DEBUG
 634         {
 635                 struct buffer_head *debug_bh;
 636
 637                 /* Record bitmap buffer state in the newly allocated block */
 638                 debug_bh = sb_find_get_block(sb, target_block);
 639                 if (debug_bh) {
 640                         BUFFER_TRACE(debug_bh, "state when allocated");
 641                         BUFFER_TRACE2(debug_bh, bitmap_bh, "bitmap state");
 642                         brelse(debug_bh);
 643                 }
 644         }
 645         jbd_lock_bh_state(bitmap_bh);
 646         spin_lock(sb_bgl_lock(sbi, group_no));
 647         if (buffer_jbd(bitmap_bh) && bh2jh(bitmap_bh)->b_committed_data) {
 648                 if (ext3_test_bit(ret_block,
 649                                 bh2jh(bitmap_bh)->b_committed_data)) {
 650                         printk("%s: block was unexpectedly set in "
 651                                 "b_committed_data\n", __FUNCTION__);
 652                 }
 653         }
 654         ext3_debug("found bit %d\n", ret_block);
 655         spin_unlock(sb_bgl_lock(sbi, group_no));
 656         jbd_unlock_bh_state(bitmap_bh);
 657 #endif
 658
 659         /* ret_block was blockgroup-relative.  Now it becomes fs-relative */
 660         ret_block = target_block;
 661
 662         if (ret_block >= le32_to_cpu(es->s_blocks_count)) {
 663                 ext3_error(sb, "ext3_new_block",
 664                             "block(%d) >= blocks count(%d) - "
 665                             "block_group = %d, es == %p ", ret_block,
 666                         le32_to_cpu(es->s_blocks_count), group_no, es);
 667                 goto out;
 668         }
 669
 670         /*
 671          * It is up to the caller to add the new buffer to a journal
 672          * list of some description.  We don't know in advance whether
 673          * the caller wants to use it as metadata or data.
 674          */
 675         ext3_debug("allocating block %d. Goal hits %d of %d.\n",
 676                         ret_block, goal_hits, goal_attempts);
 677
 678         spin_lock(sb_bgl_lock(sbi, group_no));
 679         gdp->bg_free_blocks_count =
 680                         cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count) - 1);
 681         spin_unlock(sb_bgl_lock(sbi, group_no));
 682         percpu_counter_mod(&sbi->s_freeblocks_counter, -1);
 683
 684         BUFFER_TRACE(gdp_bh, "journal_dirty_metadata for group descriptor");
 685         err = ext3_journal_dirty_metadata(handle, gdp_bh);
 686         if (!fatal)
 687                 fatal = err;
 688
 689         sb->s_dirt = 1;
 690         if (fatal)
 691                 goto out;
 692
 693         *errp = 0;
 694         brelse(bitmap_bh);
 695         return ret_block;
 696
 697 io_error:
 698         *errp = -EIO;
 699 out:
 700         if (fatal) {
 701                 *errp = fatal;
 702                 ext3_std_error(sb, fatal);
 703         }
 704         /*
 705          * Undo the block allocation
 706          */
 707         if (!performed_allocation)
 708                 DQUOT_FREE_BLOCK(inode, 1);
 709         brelse(bitmap_bh);
 710         return 0;
 711 }
 712
 713 unsigned long ext3_count_free_blocks(struct super_block *sb)
 714 {
 715         unsigned long desc_count;
 716         struct ext3_group_desc *gdp;
 717         int i;
 718 #ifdef EXT3FS_DEBUG
 719         struct ext3_super_block *es;
 720         unsigned long bitmap_count, x;
 721         struct buffer_head *bitmap_bh = NULL;
 722
 723         lock_super(sb);
 724         es = EXT3_SB(sb)->s_es;
 725         desc_count = 0;
 726         bitmap_count = 0;
 727         gdp = NULL;
 728         for (i = 0; i < EXT3_SB(sb)->s_groups_count; i++) {
 729                 gdp = ext3_get_group_desc(sb, i, NULL);
 730                 if (!gdp)
 731                         continue;
 732                 desc_count += le16_to_cpu(gdp->bg_free_blocks_count);
 733                 brelse(bitmap_bh);
 734                 bitmap_bh = read_block_bitmap(sb, i);
 735                 if (bitmap_bh == NULL)
 736                         continue;
 737
 738                 x = ext3_count_free(bitmap_bh, sb->s_blocksize);
 739                 printk("group %d: stored = %d, counted = %lu\n",
 740                         i, le16_to_cpu(gdp->bg_free_blocks_count), x);
 741                 bitmap_count += x;
 742         }
 743         brelse(bitmap_bh);
 744         printk("ext3_count_free_blocks: stored = %u, computed = %lu, %lu\n",
 745                le32_to_cpu(es->s_free_blocks_count), desc_count, bitmap_count);
 746         unlock_super(sb);
 747         return bitmap_count;
 748 #else
 749         desc_count = 0;
 750         for (i = 0; i < EXT3_SB(sb)->s_groups_count; i++) {
 751                 gdp = ext3_get_group_desc(sb, i, NULL);
 752                 if (!gdp)
 753                         continue;
 754                 desc_count += le16_to_cpu(gdp->bg_free_blocks_count);
 755         }
 756
 757         return desc_count;
 758 #endif
 759 }
 760
 761 static inline int block_in_use(unsigned long block,
 762                                 struct super_block * sb,
 763                                 unsigned char * map)
 764 {
 765         return ext3_test_bit ((block -
 766                 le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block)) %
 767                          EXT3_BLOCKS_PER_GROUP(sb), map);
 768 }
 769
 770 static inline int test_root(int a, int b)
 771 {
 772         if (a == 0)
 773                 return 1;
 774         while (1) {
 775                 if (a == 1)
 776                         return 1;
 777                 if (a % b)
 778                         return 0;
 779                 a = a / b;
 780         }
 781 }
 782
 783 int ext3_group_sparse(int group)
 784 {
 785         return (test_root(group, 3) || test_root(group, 5) ||
 786                 test_root(group, 7));
 787 }
 788
 789 /**
 790  *      ext3_bg_has_super - number of blocks used by the superblock in group
 791  *      @sb: superblock for filesystem
 792  *      @group: group number to check
 793  *
 794  *      Return the number of blocks used by the superblock (primary or backup)
 795  *      in this group.  Currently this will be only 0 or 1.
 796  */
 797 int ext3_bg_has_super(struct super_block *sb, int group)
 798 {
 799         if (EXT3_HAS_RO_COMPAT_FEATURE(sb,EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER)&&
 800             !ext3_group_sparse(group))
 801                 return 0;
 802         return 1;
 803 }
 804
 805 /**
 806  *      ext3_bg_num_gdb - number of blocks used by the group table in group
 807  *      @sb: superblock for filesystem
 808  *      @group: group number to check
 809  *
 810  *      Return the number of blocks used by the group descriptor table
 811  *      (primary or backup) in this group.  In the future there may be a
 812  *      different number of descriptor blocks in each group.
 813  */
 814 unsigned long ext3_bg_num_gdb(struct super_block *sb, int group)
 815 {
 816         if (EXT3_HAS_RO_COMPAT_FEATURE(sb,EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER)&&
 817             !ext3_group_sparse(group))
 818                 return 0;
 819         return EXT3_SB(sb)->s_gdb_count;
 820 }
 821
 822 #ifdef CONFIG_EXT3_CHECK
 823 /* Called at mount-time, super-block is locked */
 824 void ext3_check_blocks_bitmap (struct super_block * sb)
 825 {
 826         struct ext3_super_block *es;
 827         unsigned long desc_count, bitmap_count, x, j;
 828         unsigned long desc_blocks;
 829         struct buffer_head *bitmap_bh = NULL;
 830         struct ext3_group_desc *gdp;
 831         int i;
 832
 833         es = EXT3_SB(sb)->s_es;
 834         desc_count = 0;
 835         bitmap_count = 0;
 836         gdp = NULL;
 837         for (i = 0; i < EXT3_SB(sb)->s_groups_count; i++) {
 838                 gdp = ext3_get_group_desc (sb, i, NULL);
 839                 if (!gdp)
 840                         continue;
 841                 desc_count += le16_to_cpu(gdp->bg_free_blocks_count);
 842                 brelse(bitmap_bh);
 843                 bitmap_bh = read_block_bitmap(sb, i);
 844                 if (bitmap_bh == NULL)
 845                         continue;
 846
 847                 if (ext3_bg_has_super(sb, i) &&
 848                                 !ext3_test_bit(0, bitmap_bh->b_data))
 849                         ext3_error(sb, __FUNCTION__,
 850                                    "Superblock in group %d is marked free", i);
 851
 852                 desc_blocks = ext3_bg_num_gdb(sb, i);
 853                 for (j = 0; j < desc_blocks; j++)
 854                         if (!ext3_test_bit(j + 1, bitmap_bh->b_data))
 855                                 ext3_error(sb, __FUNCTION__,
 856                                            "Descriptor block #%ld in group "
 857                                            "%d is marked free", j, i);
 858
 859                 if (!block_in_use (le32_to_cpu(gdp->bg_block_bitmap),
 860                                                 sb, bitmap_bh->b_data))
 861                         ext3_error (sb, "ext3_check_blocks_bitmap",
 862                                     "Block bitmap for group %d is marked free",
 863                                     i);
 864
 865                 if (!block_in_use (le32_to_cpu(gdp->bg_inode_bitmap),
 866                                                 sb, bitmap_bh->b_data))
 867                         ext3_error (sb, "ext3_check_blocks_bitmap",
 868                                     "Inode bitmap for group %d is marked free",
 869                                     i);
 870
 871                 for (j = 0; j < EXT3_SB(sb)->s_itb_per_group; j++)
 872                         if (!block_in_use (le32_to_cpu(gdp->bg_inode_table) + j,
 873                                                         sb, bitmap_bh->b_data))
 874                                 ext3_error (sb, "ext3_check_blocks_bitmap",
 875                                             "Block #%d of the inode table in "
 876                                             "group %d is marked free", j, i);
 877
 878                 x = ext3_count_free(bitmap_bh, sb->s_blocksize);
 879                 if (le16_to_cpu(gdp->bg_free_blocks_count) != x)
 880                         ext3_error (sb, "ext3_check_blocks_bitmap",
 881                                     "Wrong free blocks count for group %d, "
 882                                     "stored = %d, counted = %lu", i,
 883                                     le16_to_cpu(gdp->bg_free_blocks_count), x);
 884                 bitmap_count += x;
 885         }
 886         brelse(bitmap_bh);
 887         if (le32_to_cpu(es->s_free_blocks_count) != bitmap_count)
 888                 ext3_error (sb, "ext3_check_blocks_bitmap",
 889                         "Wrong free blocks count in super block, "
 890                         "stored = %lu, counted = %lu",
 891                         (unsigned long)le32_to_cpu(es->s_free_blocks_count),
 892                         bitmap_count);
 893 }
 894 #endif