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 #include <linux/vs_base.h>
  23 #include <linux/vs_dlimit.h>
  24
  25 /*
  26  * balloc.c contains the blocks allocation and deallocation routines
  27  */
  28
  29 /*
  30  * The free blocks are managed by bitmaps.  A file system contains several
  31  * blocks groups.  Each group contains 1 bitmap block for blocks, 1 bitmap
  32  * block for inodes, N blocks for the inode table and data blocks.
  33  *
  34  * The file system contains group descriptors which are located after the
  35  * super block.  Each descriptor contains the number of the bitmap block and
  36  * the free blocks count in the block.  The descriptors are loaded in memory
  37  * when a file system is mounted (see ext3_read_super).
  38  */
  39
  40
  41 #define in_range(b, first, len) ((b) >= (first) && (b) <= (first) + (len) - 1)
  42
  43 struct ext3_group_desc * ext3_get_group_desc(struct super_block * sb,
  44                                              unsigned int block_group,
  45                                              struct buffer_head ** bh)
  46 {
  47         unsigned long group_desc;
  48         unsigned long desc;
  49         struct ext3_group_desc * gdp;
  50
  51         if (block_group >= EXT3_SB(sb)->s_groups_count) {
  52                 ext3_error (sb, "ext3_get_group_desc",
  53                             "block_group >= groups_count - "
  54                             "block_group = %d, groups_count = %lu",
  55                             block_group, EXT3_SB(sb)->s_groups_count);
  56
  57                 return NULL;
  58         }
  59
  60         group_desc = block_group / EXT3_DESC_PER_BLOCK(sb);
  61         desc = block_group % EXT3_DESC_PER_BLOCK(sb);
  62         if (!EXT3_SB(sb)->s_group_desc[group_desc]) {
  63                 ext3_error (sb, "ext3_get_group_desc",
  64                             "Group descriptor not loaded - "
  65                             "block_group = %d, group_desc = %lu, desc = %lu",
  66                              block_group, group_desc, desc);
  67                 return NULL;
  68         }
  69
  70         gdp = (struct ext3_group_desc *)
  71               EXT3_SB(sb)->s_group_desc[group_desc]->b_data;
  72         if (bh)
  73                 *bh = EXT3_SB(sb)->s_group_desc[group_desc];
  74         return gdp + desc;
  75 }
  76
  77 /*
  78  * Read the bitmap for a given block_group, reading into the specified
  79  * slot in the superblock's bitmap cache.
  80  *
  81  * Return buffer_head on success or NULL in case of failure.
  82  */
  83 static struct buffer_head *
  84 read_block_bitmap(struct super_block *sb, unsigned int block_group)
  85 {
  86         struct ext3_group_desc * desc;
  87         struct buffer_head * bh = NULL;
  88
  89         desc = ext3_get_group_desc (sb, block_group, NULL);
  90         if (!desc)
  91                 goto error_out;
  92         bh = sb_bread(sb, le32_to_cpu(desc->bg_block_bitmap));
  93         if (!bh)
  94                 ext3_error (sb, "read_block_bitmap",
  95                             "Cannot read block bitmap - "
  96                             "block_group = %d, block_bitmap = %u",
  97                             block_group, le32_to_cpu(desc->bg_block_bitmap));
  98 error_out:
  99         return bh;
 100 }
 101
 102 /* Free given blocks, update quota and i_blocks field */
 103 void ext3_free_blocks (handle_t *handle, struct inode * inode,
 104                         unsigned long block, unsigned long count)
 105 {
 106         struct buffer_head *bitmap_bh = NULL;
 107         struct buffer_head *gd_bh;
 108         unsigned long block_group;
 109         unsigned long bit;
 110         unsigned long i;
 111         unsigned long overflow;
 112         struct super_block * sb;
 113         struct ext3_group_desc * gdp;
 114         struct ext3_super_block * es;
 115         struct ext3_sb_info *sbi;
 116         int err = 0, ret;
 117         int dquot_freed_blocks = 0;
 118
 119         sb = inode->i_sb;
 120         if (!sb) {
 121                 printk ("ext3_free_blocks: nonexistent device");
 122                 return;
 123         }
 124         sbi = EXT3_SB(sb);
 125         es = EXT3_SB(sb)->s_es;
 126         if (block < le32_to_cpu(es->s_first_data_block) ||
 127             block + count < block ||
 128             block + count > le32_to_cpu(es->s_blocks_count)) {
 129                 ext3_error (sb, "ext3_free_blocks",
 130                             "Freeing blocks not in datazone - "
 131                             "block = %lu, count = %lu", block, count);
 132                 goto error_return;
 133         }
 134
 135         ext3_debug ("freeing block %lu\n", block);
 136
 137 do_more:
 138         overflow = 0;
 139         block_group = (block - le32_to_cpu(es->s_first_data_block)) /
 140                       EXT3_BLOCKS_PER_GROUP(sb);
 141         bit = (block - le32_to_cpu(es->s_first_data_block)) %
 142                       EXT3_BLOCKS_PER_GROUP(sb);
 143         /*
 144          * Check to see if we are freeing blocks across a group
 145          * boundary.
 146          */
 147         if (bit + count > EXT3_BLOCKS_PER_GROUP(sb)) {
 148                 overflow = bit + count - EXT3_BLOCKS_PER_GROUP(sb);
 149                 count -= overflow;
 150         }
 151         brelse(bitmap_bh);
 152         bitmap_bh = read_block_bitmap(sb, block_group);
 153         if (!bitmap_bh)
 154                 goto error_return;
 155         gdp = ext3_get_group_desc (sb, block_group, &gd_bh);
 156         if (!gdp)
 157                 goto error_return;
 158
 159         if (in_range (le32_to_cpu(gdp->bg_block_bitmap), block, count) ||
 160             in_range (le32_to_cpu(gdp->bg_inode_bitmap), block, count) ||
 161             in_range (block, le32_to_cpu(gdp->bg_inode_table),
 162                       EXT3_SB(sb)->s_itb_per_group) ||
 163             in_range (block + count - 1, le32_to_cpu(gdp->bg_inode_table),
 164                       EXT3_SB(sb)->s_itb_per_group))
 165                 ext3_error (sb, "ext3_free_blocks",
 166                             "Freeing blocks in system zones - "
 167                             "Block = %lu, count = %lu",
 168                             block, count);
 169
 170         /*
 171          * We are about to start releasing blocks in the bitmap,
 172          * so we need undo access.
 173          */
 174         /* @@@ check errors */
 175         BUFFER_TRACE(bitmap_bh, "getting undo access");
 176         err = ext3_journal_get_undo_access(handle, bitmap_bh, NULL);
 177         if (err)
 178                 goto error_return;
 179
 180         /*
 181          * We are about to modify some metadata.  Call the journal APIs
 182          * to unshare ->b_data if a currently-committing transaction is
 183          * using it
 184          */
 185         BUFFER_TRACE(gd_bh, "get_write_access");
 186         err = ext3_journal_get_write_access(handle, gd_bh);
 187         if (err)
 188                 goto error_return;
 189
 190         jbd_lock_bh_state(bitmap_bh);
 191
 192         for (i = 0; i < count; i++) {
 193                 /*
 194                  * An HJ special.  This is expensive...
 195                  */
 196 #ifdef CONFIG_JBD_DEBUG
 197                 jbd_unlock_bh_state(bitmap_bh);
 198                 {
 199                         struct buffer_head *debug_bh;
 200                         debug_bh = sb_find_get_block(sb, block + i);
 201                         if (debug_bh) {
 202                                 BUFFER_TRACE(debug_bh, "Deleted!");
 203                                 if (!bh2jh(bitmap_bh)->b_committed_data)
 204                                         BUFFER_TRACE(debug_bh,
 205                                                 "No commited data in bitmap");
 206                                 BUFFER_TRACE2(debug_bh, bitmap_bh, "bitmap");
 207                                 __brelse(debug_bh);
 208                         }
 209                 }
 210                 jbd_lock_bh_state(bitmap_bh);
 211 #endif
 212                 /* @@@ This prevents newly-allocated data from being
 213                  * freed and then reallocated within the same
 214                  * transaction.
 215                  *
 216                  * Ideally we would want to allow that to happen, but to
 217                  * do so requires making journal_forget() capable of
 218                  * revoking the queued write of a data block, which
 219                  * implies blocking on the journal lock.  *forget()
 220                  * cannot block due to truncate races.
 221                  *
 222                  * Eventually we can fix this by making journal_forget()
 223                  * return a status indicating whether or not it was able
 224                  * to revoke the buffer.  On successful revoke, it is
 225                  * safe not to set the allocation bit in the committed
 226                  * bitmap, because we know that there is no outstanding
 227                  * activity on the buffer any more and so it is safe to
 228                  * reallocate it.
 229                  */
 230                 BUFFER_TRACE(bitmap_bh, "set in b_committed_data");
 231                 J_ASSERT_BH(bitmap_bh,
 232                                 bh2jh(bitmap_bh)->b_committed_data != NULL);
 233                 ext3_set_bit_atomic(sb_bgl_lock(sbi, block_group), bit + i,
 234                                 bh2jh(bitmap_bh)->b_committed_data);
 235
 236                 /*
 237                  * We clear the bit in the bitmap after setting the committed
 238                  * data bit, because this is the reverse order to that which
 239                  * the allocator uses.
 240                  */
 241                 BUFFER_TRACE(bitmap_bh, "clear bit");
 242                 if (!ext3_clear_bit_atomic(sb_bgl_lock(sbi, block_group),
 243                                                 bit + i, bitmap_bh->b_data)) {
 244                         jbd_unlock_bh_state(bitmap_bh);
 245                         ext3_error(sb, __FUNCTION__,
 246                                 "bit already cleared for block %lu", block + i);
 247                         jbd_lock_bh_state(bitmap_bh);
 248                         BUFFER_TRACE(bitmap_bh, "bit already cleared");
 249                 } else {
 250                         dquot_freed_blocks++;
 251                 }
 252         }
 253         jbd_unlock_bh_state(bitmap_bh);
 254
 255         spin_lock(sb_bgl_lock(sbi, block_group));
 256         gdp->bg_free_blocks_count =
 257                 cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count) +
 258                         dquot_freed_blocks);
 259         spin_unlock(sb_bgl_lock(sbi, block_group));
 260         percpu_counter_mod(&sbi->s_freeblocks_counter, count);
 261
 262         /* We dirtied the bitmap block */
 263         BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
 264         err = ext3_journal_dirty_metadata(handle, bitmap_bh);
 265
 266         /* And the group descriptor block */
 267         BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
 268         ret = ext3_journal_dirty_metadata(handle, gd_bh);
 269         if (!err) err = ret;
 270
 271         if (overflow && !err) {
 272                 block += count;
 273                 count = overflow;
 274                 goto do_more;
 275         }
 276         sb->s_dirt = 1;
 277 error_return:
 278         brelse(bitmap_bh);
 279         ext3_std_error(sb, err);
 280         if (dquot_freed_blocks) {
 281                 DLIMIT_FREE_BLOCK(sb, inode->i_xid, dquot_freed_blocks);
 282                 DQUOT_FREE_BLOCK(inode, dquot_freed_blocks);
 283         }
 284         return;
 285 }
 286
 287 /*
 288  * For ext3 allocations, we must not reuse any blocks which are
 289  * allocated in the bitmap buffer's "last committed data" copy.  This
 290  * prevents deletes from freeing up the page for reuse until we have
 291  * committed the delete transaction.
 292  *
 293  * If we didn't do this, then deleting something and reallocating it as
 294  * data would allow the old block to be overwritten before the
 295  * transaction committed (because we force data to disk before commit).
 296  * This would lead to corruption if we crashed between overwriting the
 297  * data and committing the delete.
 298  *
 299  * @@@ We may want to make this allocation behaviour conditional on
 300  * data-writes at some point, and disable it for metadata allocations or
 301  * sync-data inodes.
 302  */
 303 static inline int ext3_test_allocatable(int nr, struct buffer_head *bh)
 304 {
 305         int ret;
 306         struct journal_head *jh = bh2jh(bh);
 307
 308         if (ext3_test_bit(nr, bh->b_data))
 309                 return 0;
 310
 311         jbd_lock_bh_state(bh);
 312         if (!jh->b_committed_data)
 313                 ret = 1;
 314         else
 315                 ret = !ext3_test_bit(nr, jh->b_committed_data);
 316         jbd_unlock_bh_state(bh);
 317         return ret;
 318 }
 319
 320 /*
 321  * Find an allocatable block in a bitmap.  We honour both the bitmap and
 322  * its last-committed copy (if that exists), and perform the "most
 323  * appropriate allocation" algorithm of looking for a free block near
 324  * the initial goal; then for a free byte somewhere in the bitmap; then
 325  * for any free bit in the bitmap.
 326  */
 327 static int
 328 find_next_usable_block(int start, struct buffer_head *bh, int maxblocks)
 329 {
 330         int here, next;
 331         char *p, *r;
 332         struct journal_head *jh = bh2jh(bh);
 333
 334         if (start > 0) {
 335                 /*
 336                  * The goal was occupied; search forward for a free
 337                  * block within the next XX blocks.
 338                  *
 339                  * end_goal is more or less random, but it has to be
 340                  * less than EXT3_BLOCKS_PER_GROUP. Aligning up to the
 341                  * next 64-bit boundary is simple..
 342                  */
 343                 int end_goal = (start + 63) & ~63;
 344                 here = ext3_find_next_zero_bit(bh->b_data, end_goal, start);
 345                 if (here < end_goal && ext3_test_allocatable(here, bh))
 346                         return here;
 347                 ext3_debug("Bit not found near goal\n");
 348         }
 349
 350         here = start;
 351         if (here < 0)
 352                 here = 0;
 353
 354         p = ((char *)bh->b_data) + (here >> 3);
 355         r = memscan(p, 0, (maxblocks - here + 7) >> 3);
 356         next = (r - ((char *)bh->b_data)) << 3;
 357
 358         if (next < maxblocks && ext3_test_allocatable(next, bh))
 359                 return next;
 360
 361         /*
 362          * The bitmap search --- search forward alternately through the actual
 363          * bitmap and the last-committed copy until we find a bit free in
 364          * both
 365          */
 366         while (here < maxblocks) {
 367                 next = ext3_find_next_zero_bit(bh->b_data, maxblocks, here);
 368                 if (next >= maxblocks)
 369                         return -1;
 370                 if (ext3_test_allocatable(next, bh))
 371                         return next;
 372                 jbd_lock_bh_state(bh);
 373                 if (jh->b_committed_data)
 374                         here = ext3_find_next_zero_bit(jh->b_committed_data,
 375                                                         maxblocks, next);
 376                 jbd_unlock_bh_state(bh);
 377         }
 378         return -1;
 379 }
 380
 381 /*
 382  * We think we can allocate this block in this bitmap.  Try to set the bit.
 383  * If that succeeds then check that nobody has allocated and then freed the
 384  * block since we saw that is was not marked in b_committed_data.  If it _was_
 385  * allocated and freed then clear the bit in the bitmap again and return
 386  * zero (failure).
 387  */
 388 static inline int
 389 claim_block(spinlock_t *lock, int block, struct buffer_head *bh)
 390 {
 391         struct journal_head *jh = bh2jh(bh);
 392         int ret;
 393
 394         if (ext3_set_bit_atomic(lock, block, bh->b_data))
 395                 return 0;
 396         jbd_lock_bh_state(bh);
 397         if (jh->b_committed_data && ext3_test_bit(block,jh->b_committed_data)) {
 398                 ext3_clear_bit_atomic(lock, block, bh->b_data);
 399                 ret = 0;
 400         } else {
 401                 ret = 1;
 402         }
 403         jbd_unlock_bh_state(bh);
 404         return ret;
 405 }
 406
 407 /*
 408  * If we failed to allocate the desired block then we may end up crossing to a
 409  * new bitmap.  In that case we must release write access to the old one via
 410  * ext3_journal_release_buffer(), else we'll run out of credits.
 411  */
 412 static int
 413 ext3_try_to_allocate(struct super_block *sb, handle_t *handle, int group,
 414                 struct buffer_head *bitmap_bh, int goal, int *errp)
 415 {
 416         int i;
 417         int fatal;
 418         int credits = 0;
 419
 420         *errp = 0;
 421
 422         /*
 423          * Make sure we use undo access for the bitmap, because it is critical
 424          * that we do the frozen_data COW on bitmap buffers in all cases even
 425          * if the buffer is in BJ_Forget state in the committing transaction.
 426          */
 427         BUFFER_TRACE(bitmap_bh, "get undo access for new block");
 428         fatal = ext3_journal_get_undo_access(handle, bitmap_bh, &credits);
 429         if (fatal) {
 430                 *errp = fatal;
 431                 goto fail;
 432         }
 433
 434 repeat:
 435         if (goal < 0 || !ext3_test_allocatable(goal, bitmap_bh)) {
 436                 goal = find_next_usable_block(goal, bitmap_bh,
 437                                         EXT3_BLOCKS_PER_GROUP(sb));
 438                 if (goal < 0)
 439                         goto fail_access;
 440
 441                 for (i = 0; i < 7 && goal > 0 &&
 442                                 ext3_test_allocatable(goal - 1, bitmap_bh);
 443                         i++, goal--);
 444         }
 445
 446         if (!claim_block(sb_bgl_lock(EXT3_SB(sb), group), goal, bitmap_bh)) {
 447                 /*
 448                  * The block was allocated by another thread, or it was
 449                  * allocated and then freed by another thread
 450                  */
 451                 goal++;
 452                 if (goal >= EXT3_BLOCKS_PER_GROUP(sb))
 453                         goto fail_access;
 454                 goto repeat;
 455         }
 456
 457         BUFFER_TRACE(bitmap_bh, "journal_dirty_metadata for bitmap block");
 458         fatal = ext3_journal_dirty_metadata(handle, bitmap_bh);
 459         if (fatal) {
 460                 *errp = fatal;
 461                 goto fail;
 462         }
 463         return goal;
 464
 465 fail_access:
 466         BUFFER_TRACE(bitmap_bh, "journal_release_buffer");
 467         ext3_journal_release_buffer(handle, bitmap_bh, credits);
 468 fail:
 469         return -1;
 470 }
 471
 472 static int ext3_has_free_blocks(struct super_block *sb)
 473 {
 474         struct ext3_sb_info *sbi = EXT3_SB(sb);
 475         int free_blocks, root_blocks, cond;
 476
 477         free_blocks = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
 478         root_blocks = le32_to_cpu(sbi->s_es->s_r_blocks_count);
 479
 480         vxdprintk(VXD_CBIT(dlim, 3),
 481                 "ext3_has_free_blocks(%p): free=%u, root=%u",
 482                 sb, free_blocks, root_blocks);
 483
 484         DLIMIT_ADJUST_BLOCK(sb, vx_current_xid(), &free_blocks, &root_blocks);
 485
 486         cond = (free_blocks < root_blocks + 1 &&
 487                 !capable(CAP_SYS_RESOURCE) &&
 488                 sbi->s_resuid != current->fsuid &&
 489                 (sbi->s_resgid == 0 || !in_group_p (sbi->s_resgid)));
 490
 491         vxdprintk(VXD_CBIT(dlim, 3),
 492                 "ext3_has_free_blocks(%p): %u<%u+1, %c, %u!=%u r=%d",
 493                 sb, free_blocks, root_blocks,
 494                 !capable(CAP_SYS_RESOURCE)?'1':'0',
 495                 sbi->s_resuid, current->fsuid, cond?0:1);
 496
 497         return (cond ? 0 : 1);
 498 }
 499
 500 /*
 501  * ext3_should_retry_alloc() is called when ENOSPC is returned, and if
 502  * it is profitable to retry the operation, this function will wait
 503  * for the current or commiting transaction to complete, and then
 504  * return TRUE.
 505  */
 506 int ext3_should_retry_alloc(struct super_block *sb, int *retries)
 507 {
 508         if (!ext3_has_free_blocks(sb) || (*retries)++ > 3)
 509                 return 0;
 510
 511         jbd_debug(1, "%s: retrying operation after ENOSPC\n", sb->s_id);
 512
 513         return journal_force_commit_nested(EXT3_SB(sb)->s_journal);
 514 }
 515
 516 /*
 517  * ext3_new_block uses a goal block to assist allocation.  If the goal is
 518  * free, or there is a free block within 32 blocks of the goal, that block
 519  * is allocated.  Otherwise a forward search is made for a free block; within
 520  * each block group the search first looks for an entire free byte in the block
 521  * bitmap, and then for any free bit if that fails.
 522  * This function also updates quota and i_blocks field.
 523  */
 524 int
 525 ext3_new_block(handle_t *handle, struct inode *inode, unsigned long goal,
 526                 u32 *prealloc_count, u32 *prealloc_block, int *errp)
 527 {
 528         struct buffer_head *bitmap_bh = NULL;   /* bh */
 529         struct buffer_head *gdp_bh;             /* bh2 */
 530         int group_no;                           /* i */
 531         int ret_block;                          /* j */
 532         int bgi;                                /* blockgroup iteration index */
 533         int target_block;                       /* tmp */
 534         int fatal = 0, err;
 535         int performed_allocation = 0;
 536         int free_blocks;
 537         struct super_block *sb;
 538         struct ext3_group_desc *gdp;
 539         struct ext3_super_block *es;
 540         struct ext3_sb_info *sbi;
 541 #ifdef EXT3FS_DEBUG
 542         static int goal_hits, goal_attempts;
 543 #endif
 544         *errp = -ENOSPC;
 545         sb = inode->i_sb;
 546         if (!sb) {
 547                 printk("ext3_new_block: nonexistent device");
 548                 return 0;
 549         }
 550
 551         /*
 552          * Check quota for allocation of this block.
 553          */
 554         if (DQUOT_ALLOC_BLOCK(inode, 1)) {
 555                 *errp = -EDQUOT;
 556                 return 0;
 557         }
 558         if (DLIMIT_ALLOC_BLOCK(sb, inode->i_xid, 1))
 559                 goto out_dlimit;
 560
 561         sbi = EXT3_SB(sb);
 562         es = EXT3_SB(sb)->s_es;
 563         ext3_debug("goal=%lu.\n", goal);
 564
 565         if (!ext3_has_free_blocks(sb)) {
 566                 *errp = -ENOSPC;
 567                 goto out;
 568         }
 569
 570         /*
 571          * First, test whether the goal block is free.
 572          */
 573         if (goal < le32_to_cpu(es->s_first_data_block) ||
 574             goal >= le32_to_cpu(es->s_blocks_count))
 575                 goal = le32_to_cpu(es->s_first_data_block);
 576         group_no = (goal - le32_to_cpu(es->s_first_data_block)) /
 577                         EXT3_BLOCKS_PER_GROUP(sb);
 578         gdp = ext3_get_group_desc(sb, group_no, &gdp_bh);
 579         if (!gdp)
 580                 goto io_error;
 581
 582         free_blocks = le16_to_cpu(gdp->bg_free_blocks_count);
 583         if (free_blocks > 0) {
 584                 ret_block = ((goal - le32_to_cpu(es->s_first_data_block)) %
 585                                 EXT3_BLOCKS_PER_GROUP(sb));
 586                 bitmap_bh = read_block_bitmap(sb, group_no);
 587                 if (!bitmap_bh)
 588                         goto io_error;
 589                 ret_block = ext3_try_to_allocate(sb, handle, group_no,
 590                                         bitmap_bh, ret_block, &fatal);
 591                 if (fatal)
 592                         goto out;
 593                 if (ret_block >= 0)
 594                         goto allocated;
 595         }
 596
 597         /*
 598          * Now search the rest of the groups.  We assume that
 599          * i and gdp correctly point to the last group visited.
 600          */
 601         for (bgi = 0; bgi < EXT3_SB(sb)->s_groups_count; bgi++) {
 602                 group_no++;
 603                 if (group_no >= EXT3_SB(sb)->s_groups_count)
 604                         group_no = 0;
 605                 gdp = ext3_get_group_desc(sb, group_no, &gdp_bh);
 606                 if (!gdp) {
 607                         *errp = -EIO;
 608                         goto out;
 609                 }
 610                 free_blocks = le16_to_cpu(gdp->bg_free_blocks_count);
 611                 if (free_blocks <= 0)
 612                         continue;
 613
 614                 brelse(bitmap_bh);
 615                 bitmap_bh = read_block_bitmap(sb, group_no);
 616                 if (!bitmap_bh)
 617                         goto io_error;
 618                 ret_block = ext3_try_to_allocate(sb, handle, group_no,
 619                                                 bitmap_bh, -1, &fatal);
 620                 if (fatal)
 621                         goto out;
 622                 if (ret_block >= 0)
 623                         goto allocated;
 624         }
 625
 626         /* No space left on the device */
 627         *errp = -ENOSPC;
 628         goto out;
 629
 630 allocated:
 631
 632         ext3_debug("using block group %d(%d)\n",
 633                         group_no, gdp->bg_free_blocks_count);
 634
 635         BUFFER_TRACE(gdp_bh, "get_write_access");
 636         fatal = ext3_journal_get_write_access(handle, gdp_bh);
 637         if (fatal)
 638                 goto out;
 639
 640         target_block = ret_block + group_no * EXT3_BLOCKS_PER_GROUP(sb)
 641                                 + le32_to_cpu(es->s_first_data_block);
 642
 643         if (target_block == le32_to_cpu(gdp->bg_block_bitmap) ||
 644             target_block == le32_to_cpu(gdp->bg_inode_bitmap) ||
 645             in_range(target_block, le32_to_cpu(gdp->bg_inode_table),
 646                       EXT3_SB(sb)->s_itb_per_group))
 647                 ext3_error(sb, "ext3_new_block",
 648                             "Allocating block in system zone - "
 649                             "block = %u", target_block);
 650
 651         performed_allocation = 1;
 652
 653 #ifdef CONFIG_JBD_DEBUG
 654         {
 655                 struct buffer_head *debug_bh;
 656
 657                 /* Record bitmap buffer state in the newly allocated block */
 658                 debug_bh = sb_find_get_block(sb, target_block);
 659                 if (debug_bh) {
 660                         BUFFER_TRACE(debug_bh, "state when allocated");
 661                         BUFFER_TRACE2(debug_bh, bitmap_bh, "bitmap state");
 662                         brelse(debug_bh);
 663                 }
 664         }
 665         jbd_lock_bh_state(bitmap_bh);
 666         spin_lock(sb_bgl_lock(sbi, group_no));
 667         if (buffer_jbd(bitmap_bh) && bh2jh(bitmap_bh)->b_committed_data) {
 668                 if (ext3_test_bit(ret_block,
 669                                 bh2jh(bitmap_bh)->b_committed_data)) {
 670                         printk("%s: block was unexpectedly set in "
 671                                 "b_committed_data\n", __FUNCTION__);
 672                 }
 673         }
 674         ext3_debug("found bit %d\n", ret_block);
 675         spin_unlock(sb_bgl_lock(sbi, group_no));
 676         jbd_unlock_bh_state(bitmap_bh);
 677 #endif
 678
 679         /* ret_block was blockgroup-relative.  Now it becomes fs-relative */
 680         ret_block = target_block;
 681
 682         if (ret_block >= le32_to_cpu(es->s_blocks_count)) {
 683                 ext3_error(sb, "ext3_new_block",
 684                             "block(%d) >= blocks count(%d) - "
 685                             "block_group = %d, es == %p ", ret_block,
 686                         le32_to_cpu(es->s_blocks_count), group_no, es);
 687                 goto out;
 688         }
 689
 690         /*
 691          * It is up to the caller to add the new buffer to a journal
 692          * list of some description.  We don't know in advance whether
 693          * the caller wants to use it as metadata or data.
 694          */
 695         ext3_debug("allocating block %d. Goal hits %d of %d.\n",
 696                         ret_block, goal_hits, goal_attempts);
 697
 698         spin_lock(sb_bgl_lock(sbi, group_no));
 699         gdp->bg_free_blocks_count =
 700                         cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count) - 1);
 701         spin_unlock(sb_bgl_lock(sbi, group_no));
 702         percpu_counter_mod(&sbi->s_freeblocks_counter, -1);
 703
 704         BUFFER_TRACE(gdp_bh, "journal_dirty_metadata for group descriptor");
 705         err = ext3_journal_dirty_metadata(handle, gdp_bh);
 706         if (!fatal)
 707                 fatal = err;
 708
 709         sb->s_dirt = 1;
 710         if (fatal)
 711                 goto out;
 712
 713         *errp = 0;
 714         brelse(bitmap_bh);
 715         return ret_block;
 716
 717 io_error:
 718         *errp = -EIO;
 719 out:
 720         if (!performed_allocation)
 721                 DLIMIT_FREE_BLOCK(sb, inode->i_xid, 1);
 722 out_dlimit:
 723         if (fatal) {
 724                 *errp = fatal;
 725                 ext3_std_error(sb, fatal);
 726         }
 727         /*
 728          * Undo the block allocation
 729          */
 730         if (!performed_allocation)
 731                 DQUOT_FREE_BLOCK(inode, 1);
 732         brelse(bitmap_bh);
 733         return 0;
 734 }
 735
 736 unsigned long ext3_count_free_blocks(struct super_block *sb)
 737 {
 738         unsigned long desc_count;
 739         struct ext3_group_desc *gdp;
 740         int i;
 741 #ifdef EXT3FS_DEBUG
 742         struct ext3_super_block *es;
 743         unsigned long bitmap_count, x;
 744         struct buffer_head *bitmap_bh = NULL;
 745
 746         lock_super(sb);
 747         es = EXT3_SB(sb)->s_es;
 748         desc_count = 0;
 749         bitmap_count = 0;
 750         gdp = NULL;
 751         for (i = 0; i < EXT3_SB(sb)->s_groups_count; i++) {
 752                 gdp = ext3_get_group_desc(sb, i, NULL);
 753                 if (!gdp)
 754                         continue;
 755                 desc_count += le16_to_cpu(gdp->bg_free_blocks_count);
 756                 brelse(bitmap_bh);
 757                 bitmap_bh = read_block_bitmap(sb, i);
 758                 if (bitmap_bh == NULL)
 759                         continue;
 760
 761                 x = ext3_count_free(bitmap_bh, sb->s_blocksize);
 762                 printk("group %d: stored = %d, counted = %lu\n",
 763                         i, le16_to_cpu(gdp->bg_free_blocks_count), x);
 764                 bitmap_count += x;
 765         }
 766         brelse(bitmap_bh);
 767         printk("ext3_count_free_blocks: stored = %u, computed = %lu, %lu\n",
 768                le32_to_cpu(es->s_free_blocks_count), desc_count, bitmap_count);
 769         unlock_super(sb);
 770         return bitmap_count;
 771 #else
 772         desc_count = 0;
 773         for (i = 0; i < EXT3_SB(sb)->s_groups_count; i++) {
 774                 gdp = ext3_get_group_desc(sb, i, NULL);
 775                 if (!gdp)
 776                         continue;
 777                 desc_count += le16_to_cpu(gdp->bg_free_blocks_count);
 778         }
 779
 780         return desc_count;
 781 #endif
 782 }
 783
 784 static inline int block_in_use(unsigned long block,
 785                                 struct super_block * sb,
 786                                 unsigned char * map)
 787 {
 788         return ext3_test_bit ((block -
 789                 le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block)) %
 790                          EXT3_BLOCKS_PER_GROUP(sb), map);
 791 }
 792
 793 static inline int test_root(int a, int b)
 794 {
 795         if (a == 0)
 796                 return 1;
 797         while (1) {
 798                 if (a == 1)
 799                         return 1;
 800                 if (a % b)
 801                         return 0;
 802                 a = a / b;
 803         }
 804 }
 805
 806 int ext3_group_sparse(int group)
 807 {
 808         return (test_root(group, 3) || test_root(group, 5) ||
 809                 test_root(group, 7));
 810 }
 811
 812 /**
 813  *      ext3_bg_has_super - number of blocks used by the superblock in group
 814  *      @sb: superblock for filesystem
 815  *      @group: group number to check
 816  *
 817  *      Return the number of blocks used by the superblock (primary or backup)
 818  *      in this group.  Currently this will be only 0 or 1.
 819  */
 820 int ext3_bg_has_super(struct super_block *sb, int group)
 821 {
 822         if (EXT3_HAS_RO_COMPAT_FEATURE(sb,EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER)&&
 823             !ext3_group_sparse(group))
 824                 return 0;
 825         return 1;
 826 }
 827
 828 /**
 829  *      ext3_bg_num_gdb - number of blocks used by the group table in group
 830  *      @sb: superblock for filesystem
 831  *      @group: group number to check
 832  *
 833  *      Return the number of blocks used by the group descriptor table
 834  *      (primary or backup) in this group.  In the future there may be a
 835  *      different number of descriptor blocks in each group.
 836  */
 837 unsigned long ext3_bg_num_gdb(struct super_block *sb, int group)
 838 {
 839         if (EXT3_HAS_RO_COMPAT_FEATURE(sb,EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER)&&
 840             !ext3_group_sparse(group))
 841                 return 0;
 842         return EXT3_SB(sb)->s_gdb_count;
 843 }
 844
 845 #ifdef CONFIG_EXT3_CHECK
 846 /* Called at mount-time, super-block is locked */
 847 void ext3_check_blocks_bitmap (struct super_block * sb)
 848 {
 849         struct ext3_super_block *es;
 850         unsigned long desc_count, bitmap_count, x, j;
 851         unsigned long desc_blocks;
 852         struct buffer_head *bitmap_bh = NULL;
 853         struct ext3_group_desc *gdp;
 854         int i;
 855
 856         es = EXT3_SB(sb)->s_es;
 857         desc_count = 0;
 858         bitmap_count = 0;
 859         gdp = NULL;
 860         for (i = 0; i < EXT3_SB(sb)->s_groups_count; i++) {
 861                 gdp = ext3_get_group_desc (sb, i, NULL);
 862                 if (!gdp)
 863                         continue;
 864                 desc_count += le16_to_cpu(gdp->bg_free_blocks_count);
 865                 brelse(bitmap_bh);
 866                 bitmap_bh = read_block_bitmap(sb, i);
 867                 if (bitmap_bh == NULL)
 868                         continue;
 869
 870                 if (ext3_bg_has_super(sb, i) &&
 871                                 !ext3_test_bit(0, bitmap_bh->b_data))
 872                         ext3_error(sb, __FUNCTION__,
 873                                    "Superblock in group %d is marked free", i);
 874
 875                 desc_blocks = ext3_bg_num_gdb(sb, i);
 876                 for (j = 0; j < desc_blocks; j++)
 877                         if (!ext3_test_bit(j + 1, bitmap_bh->b_data))
 878                                 ext3_error(sb, __FUNCTION__,
 879                                            "Descriptor block #%ld in group "
 880                                            "%d is marked free", j, i);
 881
 882                 if (!block_in_use (le32_to_cpu(gdp->bg_block_bitmap),
 883                                                 sb, bitmap_bh->b_data))
 884                         ext3_error (sb, "ext3_check_blocks_bitmap",
 885                                     "Block bitmap for group %d is marked free",
 886                                     i);
 887
 888                 if (!block_in_use (le32_to_cpu(gdp->bg_inode_bitmap),
 889                                                 sb, bitmap_bh->b_data))
 890                         ext3_error (sb, "ext3_check_blocks_bitmap",
 891                                     "Inode bitmap for group %d is marked free",
 892                                     i);
 893
 894                 for (j = 0; j < EXT3_SB(sb)->s_itb_per_group; j++)
 895                         if (!block_in_use (le32_to_cpu(gdp->bg_inode_table) + j,
 896                                                         sb, bitmap_bh->b_data))
 897                                 ext3_error (sb, "ext3_check_blocks_bitmap",
 898                                             "Block #%d of the inode table in "
 899                                             "group %d is marked free", j, i);
 900
 901                 x = ext3_count_free(bitmap_bh, sb->s_blocksize);
 902                 if (le16_to_cpu(gdp->bg_free_blocks_count) != x)
 903                         ext3_error (sb, "ext3_check_blocks_bitmap",
 904                                     "Wrong free blocks count for group %d, "
 905                                     "stored = %d, counted = %lu", i,
 906                                     le16_to_cpu(gdp->bg_free_blocks_count), x);
 907                 bitmap_count += x;
 908         }
 909         brelse(bitmap_bh);
 910         if (le32_to_cpu(es->s_free_blocks_count) != bitmap_count)
 911                 ext3_error (sb, "ext3_check_blocks_bitmap",
 912                         "Wrong free blocks count in super block, "
 913                         "stored = %lu, counted = %lu",
 914                         (unsigned long)le32_to_cpu(es->s_free_blocks_count),
 915                         bitmap_count);
 916 }
 917 #endif