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