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/time.h>
  15 #include <linux/capability.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_fill_super).
  36  */
  37
  38
  39 #define in_range(b, first, len) ((b) >= (first) && (b) <= (first) + (len) - 1)
  40
  41 /**
  42  * ext3_get_group_desc() -- load group descriptor from disk
  43  * @sb:                 super block
  44  * @block_group:        given block group
  45  * @bh:                 pointer to the buffer head to store the block
  46  *                      group descriptor
  47  */
  48 struct ext3_group_desc * ext3_get_group_desc(struct super_block * sb,
  49                                              unsigned int block_group,
  50                                              struct buffer_head ** bh)
  51 {
  52         unsigned long group_desc;
  53         unsigned long offset;
  54         struct ext3_group_desc * desc;
  55         struct ext3_sb_info *sbi = EXT3_SB(sb);
  56
  57         if (block_group >= sbi->s_groups_count) {
  58                 ext3_error (sb, "ext3_get_group_desc",
  59                             "block_group >= groups_count - "
  60                             "block_group = %d, groups_count = %lu",
  61                             block_group, sbi->s_groups_count);
  62
  63                 return NULL;
  64         }
  65         smp_rmb();
  66
  67         group_desc = block_group >> EXT3_DESC_PER_BLOCK_BITS(sb);
  68         offset = block_group & (EXT3_DESC_PER_BLOCK(sb) - 1);
  69         if (!sbi->s_group_desc[group_desc]) {
  70                 ext3_error (sb, "ext3_get_group_desc",
  71                             "Group descriptor not loaded - "
  72                             "block_group = %d, group_desc = %lu, desc = %lu",
  73                              block_group, group_desc, offset);
  74                 return NULL;
  75         }
  76
  77         desc = (struct ext3_group_desc *) sbi->s_group_desc[group_desc]->b_data;
  78         if (bh)
  79                 *bh = sbi->s_group_desc[group_desc];
  80         return desc + offset;
  81 }
  82
  83 static int ext3_valid_block_bitmap(struct super_block *sb,
  84                                         struct ext3_group_desc *desc,
  85                                         unsigned int block_group,
  86                                         struct buffer_head *bh)
  87 {
  88         ext3_grpblk_t offset;
  89         ext3_grpblk_t next_zero_bit;
  90         ext3_fsblk_t bitmap_blk;
  91         ext3_fsblk_t group_first_block;
  92
  93         group_first_block = ext3_group_first_block_no(sb, block_group);
  94
  95         /* check whether block bitmap block number is set */
  96         bitmap_blk = le32_to_cpu(desc->bg_block_bitmap);
  97         offset = bitmap_blk - group_first_block;
  98         if (!ext3_test_bit(offset, bh->b_data))
  99                 /* bad block bitmap */
 100                 goto err_out;
 101
 102         /* check whether the inode bitmap block number is set */
 103         bitmap_blk = le32_to_cpu(desc->bg_inode_bitmap);
 104         offset = bitmap_blk - group_first_block;
 105         if (!ext3_test_bit(offset, bh->b_data))
 106                 /* bad block bitmap */
 107                 goto err_out;
 108
 109         /* check whether the inode table block number is set */
 110         bitmap_blk = le32_to_cpu(desc->bg_inode_table);
 111         offset = bitmap_blk - group_first_block;
 112         next_zero_bit = ext3_find_next_zero_bit(bh->b_data,
 113                                 offset + EXT3_SB(sb)->s_itb_per_group,
 114                                 offset);
 115         if (next_zero_bit >= offset + EXT3_SB(sb)->s_itb_per_group)
 116                 /* good bitmap for inode tables */
 117                 return 1;
 118
 119 err_out:
 120         ext3_error(sb, __FUNCTION__,
 121                         "Invalid block bitmap - "
 122                         "block_group = %d, block = %lu",
 123                         block_group, bitmap_blk);
 124         return 0;
 125 }
 126
 127 /**
 128  * read_block_bitmap()
 129  * @sb:                 super block
 130  * @block_group:        given block group
 131  *
 132  * Read the bitmap for a given block_group,and validate the
 133  * bits for block/inode/inode tables are set in the bitmaps
 134  *
 135  * Return buffer_head on success or NULL in case of failure.
 136  */
 137 static struct buffer_head *
 138 read_block_bitmap(struct super_block *sb, unsigned int block_group)
 139 {
 140         struct ext3_group_desc * desc;
 141         struct buffer_head * bh = NULL;
 142         ext3_fsblk_t bitmap_blk;
 143
 144         desc = ext3_get_group_desc(sb, block_group, NULL);
 145         if (!desc)
 146                 return NULL;
 147         bitmap_blk = le32_to_cpu(desc->bg_block_bitmap);
 148         bh = sb_getblk(sb, bitmap_blk);
 149         if (unlikely(!bh)) {
 150                 ext3_error(sb, __FUNCTION__,
 151                             "Cannot read block bitmap - "
 152                             "block_group = %d, block_bitmap = %u",
 153                             block_group, le32_to_cpu(desc->bg_block_bitmap));
 154                 return NULL;
 155         }
 156         if (likely(bh_uptodate_or_lock(bh)))
 157                 return bh;
 158
 159         if (bh_submit_read(bh) < 0) {
 160                 brelse(bh);
 161                 ext3_error(sb, __FUNCTION__,
 162                             "Cannot read block bitmap - "
 163                             "block_group = %d, block_bitmap = %u",
 164                             block_group, le32_to_cpu(desc->bg_block_bitmap));
 165                 return NULL;
 166         }
 167         if (!ext3_valid_block_bitmap(sb, desc, block_group, bh)) {
 168                 brelse(bh);
 169                 return NULL;
 170         }
 171         return bh;
 172 }
 173 /*
 174  * The reservation window structure operations
 175  * --------------------------------------------
 176  * Operations include:
 177  * dump, find, add, remove, is_empty, find_next_reservable_window, etc.
 178  *
 179  * We use a red-black tree to represent per-filesystem reservation
 180  * windows.
 181  *
 182  */
 183
 184 /**
 185  * __rsv_window_dump() -- Dump the filesystem block allocation reservation map
 186  * @rb_root:            root of per-filesystem reservation rb tree
 187  * @verbose:            verbose mode
 188  * @fn:                 function which wishes to dump the reservation map
 189  *
 190  * If verbose is turned on, it will print the whole block reservation
 191  * windows(start, end). Otherwise, it will only print out the "bad" windows,
 192  * those windows that overlap with their immediate neighbors.
 193  */
 194 #if 1
 195 static void __rsv_window_dump(struct rb_root *root, int verbose,
 196                               const char *fn)
 197 {
 198         struct rb_node *n;
 199         struct ext3_reserve_window_node *rsv, *prev;
 200         int bad;
 201
 202 restart:
 203         n = rb_first(root);
 204         bad = 0;
 205         prev = NULL;
 206
 207         printk("Block Allocation Reservation Windows Map (%s):\n", fn);
 208         while (n) {
 209                 rsv = rb_entry(n, struct ext3_reserve_window_node, rsv_node);
 210                 if (verbose)
 211                         printk("reservation window 0x%p "
 212                                "start:  %lu, end:  %lu\n",
 213                                rsv, rsv->rsv_start, rsv->rsv_end);
 214                 if (rsv->rsv_start && rsv->rsv_start >= rsv->rsv_end) {
 215                         printk("Bad reservation %p (start >= end)\n",
 216                                rsv);
 217                         bad = 1;
 218                 }
 219                 if (prev && prev->rsv_end >= rsv->rsv_start) {
 220                         printk("Bad reservation %p (prev->end >= start)\n",
 221                                rsv);
 222                         bad = 1;
 223                 }
 224                 if (bad) {
 225                         if (!verbose) {
 226                                 printk("Restarting reservation walk in verbose mode\n");
 227                                 verbose = 1;
 228                                 goto restart;
 229                         }
 230                 }
 231                 n = rb_next(n);
 232                 prev = rsv;
 233         }
 234         printk("Window map complete.\n");
 235         if (bad)
 236                 BUG();
 237 }
 238 #define rsv_window_dump(root, verbose) \
 239         __rsv_window_dump((root), (verbose), __FUNCTION__)
 240 #else
 241 #define rsv_window_dump(root, verbose) do {} while (0)
 242 #endif
 243
 244 /**
 245  * goal_in_my_reservation()
 246  * @rsv:                inode's reservation window
 247  * @grp_goal:           given goal block relative to the allocation block group
 248  * @group:              the current allocation block group
 249  * @sb:                 filesystem super block
 250  *
 251  * Test if the given goal block (group relative) is within the file's
 252  * own block reservation window range.
 253  *
 254  * If the reservation window is outside the goal allocation group, return 0;
 255  * grp_goal (given goal block) could be -1, which means no specific
 256  * goal block. In this case, always return 1.
 257  * If the goal block is within the reservation window, return 1;
 258  * otherwise, return 0;
 259  */
 260 static int
 261 goal_in_my_reservation(struct ext3_reserve_window *rsv, ext3_grpblk_t grp_goal,
 262                         unsigned int group, struct super_block * sb)
 263 {
 264         ext3_fsblk_t group_first_block, group_last_block;
 265
 266         group_first_block = ext3_group_first_block_no(sb, group);
 267         group_last_block = group_first_block + (EXT3_BLOCKS_PER_GROUP(sb) - 1);
 268
 269         if ((rsv->_rsv_start > group_last_block) ||
 270             (rsv->_rsv_end < group_first_block))
 271                 return 0;
 272         if ((grp_goal >= 0) && ((grp_goal + group_first_block < rsv->_rsv_start)
 273                 || (grp_goal + group_first_block > rsv->_rsv_end)))
 274                 return 0;
 275         return 1;
 276 }
 277
 278 /**
 279  * search_reserve_window()
 280  * @rb_root:            root of reservation tree
 281  * @goal:               target allocation block
 282  *
 283  * Find the reserved window which includes the goal, or the previous one
 284  * if the goal is not in any window.
 285  * Returns NULL if there are no windows or if all windows start after the goal.
 286  */
 287 static struct ext3_reserve_window_node *
 288 search_reserve_window(struct rb_root *root, ext3_fsblk_t goal)
 289 {
 290         struct rb_node *n = root->rb_node;
 291         struct ext3_reserve_window_node *rsv;
 292
 293         if (!n)
 294                 return NULL;
 295
 296         do {
 297                 rsv = rb_entry(n, struct ext3_reserve_window_node, rsv_node);
 298
 299                 if (goal < rsv->rsv_start)
 300                         n = n->rb_left;
 301                 else if (goal > rsv->rsv_end)
 302                         n = n->rb_right;
 303                 else
 304                         return rsv;
 305         } while (n);
 306         /*
 307          * We've fallen off the end of the tree: the goal wasn't inside
 308          * any particular node.  OK, the previous node must be to one
 309          * side of the interval containing the goal.  If it's the RHS,
 310          * we need to back up one.
 311          */
 312         if (rsv->rsv_start > goal) {
 313                 n = rb_prev(&rsv->rsv_node);
 314                 rsv = rb_entry(n, struct ext3_reserve_window_node, rsv_node);
 315         }
 316         return rsv;
 317 }
 318
 319 /**
 320  * ext3_rsv_window_add() -- Insert a window to the block reservation rb tree.
 321  * @sb:                 super block
 322  * @rsv:                reservation window to add
 323  *
 324  * Must be called with rsv_lock hold.
 325  */
 326 void ext3_rsv_window_add(struct super_block *sb,
 327                     struct ext3_reserve_window_node *rsv)
 328 {
 329         struct rb_root *root = &EXT3_SB(sb)->s_rsv_window_root;
 330         struct rb_node *node = &rsv->rsv_node;
 331         ext3_fsblk_t start = rsv->rsv_start;
 332
 333         struct rb_node ** p = &root->rb_node;
 334         struct rb_node * parent = NULL;
 335         struct ext3_reserve_window_node *this;
 336
 337         while (*p)
 338         {
 339                 parent = *p;
 340                 this = rb_entry(parent, struct ext3_reserve_window_node, rsv_node);
 341
 342                 if (start < this->rsv_start)
 343                         p = &(*p)->rb_left;
 344                 else if (start > this->rsv_end)
 345                         p = &(*p)->rb_right;
 346                 else {
 347                         rsv_window_dump(root, 1);
 348                         BUG();
 349                 }
 350         }
 351
 352         rb_link_node(node, parent, p);
 353         rb_insert_color(node, root);
 354 }
 355
 356 /**
 357  * ext3_rsv_window_remove() -- unlink a window from the reservation rb tree
 358  * @sb:                 super block
 359  * @rsv:                reservation window to remove
 360  *
 361  * Mark the block reservation window as not allocated, and unlink it
 362  * from the filesystem reservation window rb tree. Must be called with
 363  * rsv_lock hold.
 364  */
 365 static void rsv_window_remove(struct super_block *sb,
 366                               struct ext3_reserve_window_node *rsv)
 367 {
 368         rsv->rsv_start = EXT3_RESERVE_WINDOW_NOT_ALLOCATED;
 369         rsv->rsv_end = EXT3_RESERVE_WINDOW_NOT_ALLOCATED;
 370         rsv->rsv_alloc_hit = 0;
 371         rb_erase(&rsv->rsv_node, &EXT3_SB(sb)->s_rsv_window_root);
 372 }
 373
 374 /*
 375  * rsv_is_empty() -- Check if the reservation window is allocated.
 376  * @rsv:                given reservation window to check
 377  *
 378  * returns 1 if the end block is EXT3_RESERVE_WINDOW_NOT_ALLOCATED.
 379  */
 380 static inline int rsv_is_empty(struct ext3_reserve_window *rsv)
 381 {
 382         /* a valid reservation end block could not be 0 */
 383         return rsv->_rsv_end == EXT3_RESERVE_WINDOW_NOT_ALLOCATED;
 384 }
 385
 386 /**
 387  * ext3_init_block_alloc_info()
 388  * @inode:              file inode structure
 389  *
 390  * Allocate and initialize the  reservation window structure, and
 391  * link the window to the ext3 inode structure at last
 392  *
 393  * The reservation window structure is only dynamically allocated
 394  * and linked to ext3 inode the first time the open file
 395  * needs a new block. So, before every ext3_new_block(s) call, for
 396  * regular files, we should check whether the reservation window
 397  * structure exists or not. In the latter case, this function is called.
 398  * Fail to do so will result in block reservation being turned off for that
 399  * open file.
 400  *
 401  * This function is called from ext3_get_blocks_handle(), also called
 402  * when setting the reservation window size through ioctl before the file
 403  * is open for write (needs block allocation).
 404  *
 405  * Needs truncate_mutex protection prior to call this function.
 406  */
 407 void ext3_init_block_alloc_info(struct inode *inode)
 408 {
 409         struct ext3_inode_info *ei = EXT3_I(inode);
 410         struct ext3_block_alloc_info *block_i = ei->i_block_alloc_info;
 411         struct super_block *sb = inode->i_sb;
 412
 413         block_i = kmalloc(sizeof(*block_i), GFP_NOFS);
 414         if (block_i) {
 415                 struct ext3_reserve_window_node *rsv = &block_i->rsv_window_node;
 416
 417                 rsv->rsv_start = EXT3_RESERVE_WINDOW_NOT_ALLOCATED;
 418                 rsv->rsv_end = EXT3_RESERVE_WINDOW_NOT_ALLOCATED;
 419
 420                 /*
 421                  * if filesystem is mounted with NORESERVATION, the goal
 422                  * reservation window size is set to zero to indicate
 423                  * block reservation is off
 424                  */
 425                 if (!test_opt(sb, RESERVATION))
 426                         rsv->rsv_goal_size = 0;
 427                 else
 428                         rsv->rsv_goal_size = EXT3_DEFAULT_RESERVE_BLOCKS;
 429                 rsv->rsv_alloc_hit = 0;
 430                 block_i->last_alloc_logical_block = 0;
 431                 block_i->last_alloc_physical_block = 0;
 432         }
 433         ei->i_block_alloc_info = block_i;
 434 }
 435
 436 /**
 437  * ext3_discard_reservation()
 438  * @inode:              inode
 439  *
 440  * Discard(free) block reservation window on last file close, or truncate
 441  * or at last iput().
 442  *
 443  * It is being called in three cases:
 444  *      ext3_release_file(): last writer close the file
 445  *      ext3_clear_inode(): last iput(), when nobody link to this file.
 446  *      ext3_truncate(): when the block indirect map is about to change.
 447  *
 448  */
 449 void ext3_discard_reservation(struct inode *inode)
 450 {
 451         struct ext3_inode_info *ei = EXT3_I(inode);
 452         struct ext3_block_alloc_info *block_i = ei->i_block_alloc_info;
 453         struct ext3_reserve_window_node *rsv;
 454         spinlock_t *rsv_lock = &EXT3_SB(inode->i_sb)->s_rsv_window_lock;
 455
 456         if (!block_i)
 457                 return;
 458
 459         rsv = &block_i->rsv_window_node;
 460         if (!rsv_is_empty(&rsv->rsv_window)) {
 461                 spin_lock(rsv_lock);
 462                 if (!rsv_is_empty(&rsv->rsv_window))
 463                         rsv_window_remove(inode->i_sb, rsv);
 464                 spin_unlock(rsv_lock);
 465         }
 466 }
 467
 468 /**
 469  * ext3_free_blocks_sb() -- Free given blocks and update quota
 470  * @handle:                     handle to this transaction
 471  * @sb:                         super block
 472  * @block:                      start physcial block to free
 473  * @count:                      number of blocks to free
 474  * @pdquot_freed_blocks:        pointer to quota
 475  */
 476 void ext3_free_blocks_sb(handle_t *handle, struct super_block *sb,
 477                          ext3_fsblk_t block, unsigned long count,
 478                          unsigned long *pdquot_freed_blocks)
 479 {
 480         struct buffer_head *bitmap_bh = NULL;
 481         struct buffer_head *gd_bh;
 482         unsigned long block_group;
 483         ext3_grpblk_t bit;
 484         unsigned long i;
 485         unsigned long overflow;
 486         struct ext3_group_desc * desc;
 487         struct ext3_super_block * es;
 488         struct ext3_sb_info *sbi;
 489         int err = 0, ret;
 490         ext3_grpblk_t group_freed;
 491
 492         *pdquot_freed_blocks = 0;
 493         sbi = EXT3_SB(sb);
 494         es = sbi->s_es;
 495         if (block < le32_to_cpu(es->s_first_data_block) ||
 496             block + count < block ||
 497             block + count > le32_to_cpu(es->s_blocks_count)) {
 498                 ext3_error (sb, "ext3_free_blocks",
 499                             "Freeing blocks not in datazone - "
 500                             "block = "E3FSBLK", count = %lu", block, count);
 501                 goto error_return;
 502         }
 503
 504         ext3_debug ("freeing block(s) %lu-%lu\n", block, block + count - 1);
 505
 506 do_more:
 507         overflow = 0;
 508         block_group = (block - le32_to_cpu(es->s_first_data_block)) /
 509                       EXT3_BLOCKS_PER_GROUP(sb);
 510         bit = (block - le32_to_cpu(es->s_first_data_block)) %
 511                       EXT3_BLOCKS_PER_GROUP(sb);
 512         /*
 513          * Check to see if we are freeing blocks across a group
 514          * boundary.
 515          */
 516         if (bit + count > EXT3_BLOCKS_PER_GROUP(sb)) {
 517                 overflow = bit + count - EXT3_BLOCKS_PER_GROUP(sb);
 518                 count -= overflow;
 519         }
 520         brelse(bitmap_bh);
 521         bitmap_bh = read_block_bitmap(sb, block_group);
 522         if (!bitmap_bh)
 523                 goto error_return;
 524         desc = ext3_get_group_desc (sb, block_group, &gd_bh);
 525         if (!desc)
 526                 goto error_return;
 527
 528         if (in_range (le32_to_cpu(desc->bg_block_bitmap), block, count) ||
 529             in_range (le32_to_cpu(desc->bg_inode_bitmap), block, count) ||
 530             in_range (block, le32_to_cpu(desc->bg_inode_table),
 531                       sbi->s_itb_per_group) ||
 532             in_range (block + count - 1, le32_to_cpu(desc->bg_inode_table),
 533                       sbi->s_itb_per_group)) {
 534                 ext3_error (sb, "ext3_free_blocks",
 535                             "Freeing blocks in system zones - "
 536                             "Block = "E3FSBLK", count = %lu",
 537                             block, count);
 538                 goto error_return;
 539         }
 540
 541         /*
 542          * We are about to start releasing blocks in the bitmap,
 543          * so we need undo access.
 544          */
 545         /* @@@ check errors */
 546         BUFFER_TRACE(bitmap_bh, "getting undo access");
 547         err = ext3_journal_get_undo_access(handle, bitmap_bh);
 548         if (err)
 549                 goto error_return;
 550
 551         /*
 552          * We are about to modify some metadata.  Call the journal APIs
 553          * to unshare ->b_data if a currently-committing transaction is
 554          * using it
 555          */
 556         BUFFER_TRACE(gd_bh, "get_write_access");
 557         err = ext3_journal_get_write_access(handle, gd_bh);
 558         if (err)
 559                 goto error_return;
 560
 561         jbd_lock_bh_state(bitmap_bh);
 562
 563         for (i = 0, group_freed = 0; i < count; i++) {
 564                 /*
 565                  * An HJ special.  This is expensive...
 566                  */
 567 #ifdef CONFIG_JBD_DEBUG
 568                 jbd_unlock_bh_state(bitmap_bh);
 569                 {
 570                         struct buffer_head *debug_bh;
 571                         debug_bh = sb_find_get_block(sb, block + i);
 572                         if (debug_bh) {
 573                                 BUFFER_TRACE(debug_bh, "Deleted!");
 574                                 if (!bh2jh(bitmap_bh)->b_committed_data)
 575                                         BUFFER_TRACE(debug_bh,
 576                                                 "No commited data in bitmap");
 577                                 BUFFER_TRACE2(debug_bh, bitmap_bh, "bitmap");
 578                                 __brelse(debug_bh);
 579                         }
 580                 }
 581                 jbd_lock_bh_state(bitmap_bh);
 582 #endif
 583                 if (need_resched()) {
 584                         jbd_unlock_bh_state(bitmap_bh);
 585                         cond_resched();
 586                         jbd_lock_bh_state(bitmap_bh);
 587                 }
 588                 /* @@@ This prevents newly-allocated data from being
 589                  * freed and then reallocated within the same
 590                  * transaction.
 591                  *
 592                  * Ideally we would want to allow that to happen, but to
 593                  * do so requires making journal_forget() capable of
 594                  * revoking the queued write of a data block, which
 595                  * implies blocking on the journal lock.  *forget()
 596                  * cannot block due to truncate races.
 597                  *
 598                  * Eventually we can fix this by making journal_forget()
 599                  * return a status indicating whether or not it was able
 600                  * to revoke the buffer.  On successful revoke, it is
 601                  * safe not to set the allocation bit in the committed
 602                  * bitmap, because we know that there is no outstanding
 603                  * activity on the buffer any more and so it is safe to
 604                  * reallocate it.
 605                  */
 606                 BUFFER_TRACE(bitmap_bh, "set in b_committed_data");
 607                 J_ASSERT_BH(bitmap_bh,
 608                                 bh2jh(bitmap_bh)->b_committed_data != NULL);
 609                 ext3_set_bit_atomic(sb_bgl_lock(sbi, block_group), bit + i,
 610                                 bh2jh(bitmap_bh)->b_committed_data);
 611
 612                 /*
 613                  * We clear the bit in the bitmap after setting the committed
 614                  * data bit, because this is the reverse order to that which
 615                  * the allocator uses.
 616                  */
 617                 BUFFER_TRACE(bitmap_bh, "clear bit");
 618                 if (!ext3_clear_bit_atomic(sb_bgl_lock(sbi, block_group),
 619                                                 bit + i, bitmap_bh->b_data)) {
 620                         jbd_unlock_bh_state(bitmap_bh);
 621                         ext3_error(sb, __FUNCTION__,
 622                                 "bit already cleared for block "E3FSBLK,
 623                                  block + i);
 624                         jbd_lock_bh_state(bitmap_bh);
 625                         BUFFER_TRACE(bitmap_bh, "bit already cleared");
 626                 } else {
 627                         group_freed++;
 628                 }
 629         }
 630         jbd_unlock_bh_state(bitmap_bh);
 631
 632         spin_lock(sb_bgl_lock(sbi, block_group));
 633         desc->bg_free_blocks_count =
 634                 cpu_to_le16(le16_to_cpu(desc->bg_free_blocks_count) +
 635                         group_freed);
 636         spin_unlock(sb_bgl_lock(sbi, block_group));
 637         percpu_counter_add(&sbi->s_freeblocks_counter, count);
 638
 639         /* We dirtied the bitmap block */
 640         BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
 641         err = ext3_journal_dirty_metadata(handle, bitmap_bh);
 642
 643         /* And the group descriptor block */
 644         BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
 645         ret = ext3_journal_dirty_metadata(handle, gd_bh);
 646         if (!err) err = ret;
 647         *pdquot_freed_blocks += group_freed;
 648
 649         if (overflow && !err) {
 650                 block += count;
 651                 count = overflow;
 652                 goto do_more;
 653         }
 654         sb->s_dirt = 1;
 655 error_return:
 656         brelse(bitmap_bh);
 657         ext3_std_error(sb, err);
 658         return;
 659 }
 660
 661 /**
 662  * ext3_free_blocks() -- Free given blocks and update quota
 663  * @handle:             handle for this transaction
 664  * @inode:              inode
 665  * @block:              start physical block to free
 666  * @count:              number of blocks to count
 667  */
 668 void ext3_free_blocks(handle_t *handle, struct inode *inode,
 669                         ext3_fsblk_t block, unsigned long count)
 670 {
 671         struct super_block * sb;
 672         unsigned long dquot_freed_blocks;
 673
 674         sb = inode->i_sb;
 675         if (!sb) {
 676                 printk ("ext3_free_blocks: nonexistent device");
 677                 return;
 678         }
 679         ext3_free_blocks_sb(handle, sb, block, count, &dquot_freed_blocks);
 680         if (dquot_freed_blocks)
 681                 DQUOT_FREE_BLOCK(inode, dquot_freed_blocks);
 682         return;
 683 }
 684
 685 /**
 686  * ext3_test_allocatable()
 687  * @nr:                 given allocation block group
 688  * @bh:                 bufferhead contains the bitmap of the given block group
 689  *
 690  * For ext3 allocations, we must not reuse any blocks which are
 691  * allocated in the bitmap buffer's "last committed data" copy.  This
 692  * prevents deletes from freeing up the page for reuse until we have
 693  * committed the delete transaction.
 694  *
 695  * If we didn't do this, then deleting something and reallocating it as
 696  * data would allow the old block to be overwritten before the
 697  * transaction committed (because we force data to disk before commit).
 698  * This would lead to corruption if we crashed between overwriting the
 699  * data and committing the delete.
 700  *
 701  * @@@ We may want to make this allocation behaviour conditional on
 702  * data-writes at some point, and disable it for metadata allocations or
 703  * sync-data inodes.
 704  */
 705 static int ext3_test_allocatable(ext3_grpblk_t nr, struct buffer_head *bh)
 706 {
 707         int ret;
 708         struct journal_head *jh = bh2jh(bh);
 709
 710         if (ext3_test_bit(nr, bh->b_data))
 711                 return 0;
 712
 713         jbd_lock_bh_state(bh);
 714         if (!jh->b_committed_data)
 715                 ret = 1;
 716         else
 717                 ret = !ext3_test_bit(nr, jh->b_committed_data);
 718         jbd_unlock_bh_state(bh);
 719         return ret;
 720 }
 721
 722 /**
 723  * bitmap_search_next_usable_block()
 724  * @start:              the starting block (group relative) of the search
 725  * @bh:                 bufferhead contains the block group bitmap
 726  * @maxblocks:          the ending block (group relative) of the reservation
 727  *
 728  * The bitmap search --- search forward alternately through the actual
 729  * bitmap on disk and the last-committed copy in journal, until we find a
 730  * bit free in both bitmaps.
 731  */
 732 static ext3_grpblk_t
 733 bitmap_search_next_usable_block(ext3_grpblk_t start, struct buffer_head *bh,
 734                                         ext3_grpblk_t maxblocks)
 735 {
 736         ext3_grpblk_t next;
 737         struct journal_head *jh = bh2jh(bh);
 738
 739         while (start < maxblocks) {
 740                 next = ext3_find_next_zero_bit(bh->b_data, maxblocks, start);
 741                 if (next >= maxblocks)
 742                         return -1;
 743                 if (ext3_test_allocatable(next, bh))
 744                         return next;
 745                 jbd_lock_bh_state(bh);
 746                 if (jh->b_committed_data)
 747                         start = ext3_find_next_zero_bit(jh->b_committed_data,
 748                                                         maxblocks, next);
 749                 jbd_unlock_bh_state(bh);
 750         }
 751         return -1;
 752 }
 753
 754 /**
 755  * find_next_usable_block()
 756  * @start:              the starting block (group relative) to find next
 757  *                      allocatable block in bitmap.
 758  * @bh:                 bufferhead contains the block group bitmap
 759  * @maxblocks:          the ending block (group relative) for the search
 760  *
 761  * Find an allocatable block in a bitmap.  We honor both the bitmap and
 762  * its last-committed copy (if that exists), and perform the "most
 763  * appropriate allocation" algorithm of looking for a free block near
 764  * the initial goal; then for a free byte somewhere in the bitmap; then
 765  * for any free bit in the bitmap.
 766  */
 767 static ext3_grpblk_t
 768 find_next_usable_block(ext3_grpblk_t start, struct buffer_head *bh,
 769                         ext3_grpblk_t maxblocks)
 770 {
 771         ext3_grpblk_t here, next;
 772         char *p, *r;
 773
 774         if (start > 0) {
 775                 /*
 776                  * The goal was occupied; search forward for a free
 777                  * block within the next XX blocks.
 778                  *
 779                  * end_goal is more or less random, but it has to be
 780                  * less than EXT3_BLOCKS_PER_GROUP. Aligning up to the
 781                  * next 64-bit boundary is simple..
 782                  */
 783                 ext3_grpblk_t end_goal = (start + 63) & ~63;
 784                 if (end_goal > maxblocks)
 785                         end_goal = maxblocks;
 786                 here = ext3_find_next_zero_bit(bh->b_data, end_goal, start);
 787                 if (here < end_goal && ext3_test_allocatable(here, bh))
 788                         return here;
 789                 ext3_debug("Bit not found near goal\n");
 790         }
 791
 792         here = start;
 793         if (here < 0)
 794                 here = 0;
 795
 796         p = ((char *)bh->b_data) + (here >> 3);
 797         r = memscan(p, 0, ((maxblocks + 7) >> 3) - (here >> 3));
 798         next = (r - ((char *)bh->b_data)) << 3;
 799
 800         if (next < maxblocks && next >= start && ext3_test_allocatable(next, bh))
 801                 return next;
 802
 803         /*
 804          * The bitmap search --- search forward alternately through the actual
 805          * bitmap and the last-committed copy until we find a bit free in
 806          * both
 807          */
 808         here = bitmap_search_next_usable_block(here, bh, maxblocks);
 809         return here;
 810 }
 811
 812 /**
 813  * claim_block()
 814  * @block:              the free block (group relative) to allocate
 815  * @bh:                 the bufferhead containts the block group bitmap
 816  *
 817  * We think we can allocate this block in this bitmap.  Try to set the bit.
 818  * If that succeeds then check that nobody has allocated and then freed the
 819  * block since we saw that is was not marked in b_committed_data.  If it _was_
 820  * allocated and freed then clear the bit in the bitmap again and return
 821  * zero (failure).
 822  */
 823 static inline int
 824 claim_block(spinlock_t *lock, ext3_grpblk_t block, struct buffer_head *bh)
 825 {
 826         struct journal_head *jh = bh2jh(bh);
 827         int ret;
 828
 829         if (ext3_set_bit_atomic(lock, block, bh->b_data))
 830                 return 0;
 831         jbd_lock_bh_state(bh);
 832         if (jh->b_committed_data && ext3_test_bit(block,jh->b_committed_data)) {
 833                 ext3_clear_bit_atomic(lock, block, bh->b_data);
 834                 ret = 0;
 835         } else {
 836                 ret = 1;
 837         }
 838         jbd_unlock_bh_state(bh);
 839         return ret;
 840 }
 841
 842 /**
 843  * ext3_try_to_allocate()
 844  * @sb:                 superblock
 845  * @handle:             handle to this transaction
 846  * @group:              given allocation block group
 847  * @bitmap_bh:          bufferhead holds the block bitmap
 848  * @grp_goal:           given target block within the group
 849  * @count:              target number of blocks to allocate
 850  * @my_rsv:             reservation window
 851  *
 852  * Attempt to allocate blocks within a give range. Set the range of allocation
 853  * first, then find the first free bit(s) from the bitmap (within the range),
 854  * and at last, allocate the blocks by claiming the found free bit as allocated.
 855  *
 856  * To set the range of this allocation:
 857  *      if there is a reservation window, only try to allocate block(s) from the
 858  *      file's own reservation window;
 859  *      Otherwise, the allocation range starts from the give goal block, ends at
 860  *      the block group's last block.
 861  *
 862  * If we failed to allocate the desired block then we may end up crossing to a
 863  * new bitmap.  In that case we must release write access to the old one via
 864  * ext3_journal_release_buffer(), else we'll run out of credits.
 865  */
 866 static ext3_grpblk_t
 867 ext3_try_to_allocate(struct super_block *sb, handle_t *handle, int group,
 868                         struct buffer_head *bitmap_bh, ext3_grpblk_t grp_goal,
 869                         unsigned long *count, struct ext3_reserve_window *my_rsv)
 870 {
 871         ext3_fsblk_t group_first_block;
 872         ext3_grpblk_t start, end;
 873         unsigned long num = 0;
 874
 875         /* we do allocation within the reservation window if we have a window */
 876         if (my_rsv) {
 877                 group_first_block = ext3_group_first_block_no(sb, group);
 878                 if (my_rsv->_rsv_start >= group_first_block)
 879                         start = my_rsv->_rsv_start - group_first_block;
 880                 else
 881                         /* reservation window cross group boundary */
 882                         start = 0;
 883                 end = my_rsv->_rsv_end - group_first_block + 1;
 884                 if (end > EXT3_BLOCKS_PER_GROUP(sb))
 885                         /* reservation window crosses group boundary */
 886                         end = EXT3_BLOCKS_PER_GROUP(sb);
 887                 if ((start <= grp_goal) && (grp_goal < end))
 888                         start = grp_goal;
 889                 else
 890                         grp_goal = -1;
 891         } else {
 892                 if (grp_goal > 0)
 893                         start = grp_goal;
 894                 else
 895                         start = 0;
 896                 end = EXT3_BLOCKS_PER_GROUP(sb);
 897         }
 898
 899         BUG_ON(start > EXT3_BLOCKS_PER_GROUP(sb));
 900
 901 repeat:
 902         if (grp_goal < 0 || !ext3_test_allocatable(grp_goal, bitmap_bh)) {
 903                 grp_goal = find_next_usable_block(start, bitmap_bh, end);
 904                 if (grp_goal < 0)
 905                         goto fail_access;
 906                 if (!my_rsv) {
 907                         int i;
 908
 909                         for (i = 0; i < 7 && grp_goal > start &&
 910                                         ext3_test_allocatable(grp_goal - 1,
 911                                                                 bitmap_bh);
 912                                         i++, grp_goal--)
 913                                 ;
 914                 }
 915         }
 916         start = grp_goal;
 917
 918         if (!claim_block(sb_bgl_lock(EXT3_SB(sb), group),
 919                 grp_goal, bitmap_bh)) {
 920                 /*
 921                  * The block was allocated by another thread, or it was
 922                  * allocated and then freed by another thread
 923                  */
 924                 start++;
 925                 grp_goal++;
 926                 if (start >= end)
 927                         goto fail_access;
 928                 goto repeat;
 929         }
 930         num++;
 931         grp_goal++;
 932         while (num < *count && grp_goal < end
 933                 && ext3_test_allocatable(grp_goal, bitmap_bh)
 934                 && claim_block(sb_bgl_lock(EXT3_SB(sb), group),
 935                                 grp_goal, bitmap_bh)) {
 936                 num++;
 937                 grp_goal++;
 938         }
 939         *count = num;
 940         return grp_goal - num;
 941 fail_access:
 942         *count = num;
 943         return -1;
 944 }
 945
 946 /**
 947  *      find_next_reservable_window():
 948  *              find a reservable space within the given range.
 949  *              It does not allocate the reservation window for now:
 950  *              alloc_new_reservation() will do the work later.
 951  *
 952  *      @search_head: the head of the searching list;
 953  *              This is not necessarily the list head of the whole filesystem
 954  *
 955  *              We have both head and start_block to assist the search
 956  *              for the reservable space. The list starts from head,
 957  *              but we will shift to the place where start_block is,
 958  *              then start from there, when looking for a reservable space.
 959  *
 960  *      @size: the target new reservation window size
 961  *
 962  *      @group_first_block: the first block we consider to start
 963  *                      the real search from
 964  *
 965  *      @last_block:
 966  *              the maximum block number that our goal reservable space
 967  *              could start from. This is normally the last block in this
 968  *              group. The search will end when we found the start of next
 969  *              possible reservable space is out of this boundary.
 970  *              This could handle the cross boundary reservation window
 971  *              request.
 972  *
 973  *      basically we search from the given range, rather than the whole
 974  *      reservation double linked list, (start_block, last_block)
 975  *      to find a free region that is of my size and has not
 976  *      been reserved.
 977  *
 978  */
 979 static int find_next_reservable_window(
 980                                 struct ext3_reserve_window_node *search_head,
 981                                 struct ext3_reserve_window_node *my_rsv,
 982                                 struct super_block * sb,
 983                                 ext3_fsblk_t start_block,
 984                                 ext3_fsblk_t last_block)
 985 {
 986         struct rb_node *next;
 987         struct ext3_reserve_window_node *rsv, *prev;
 988         ext3_fsblk_t cur;
 989         int size = my_rsv->rsv_goal_size;
 990
 991         /* TODO: make the start of the reservation window byte-aligned */
 992         /* cur = *start_block & ~7;*/
 993         cur = start_block;
 994         rsv = search_head;
 995         if (!rsv)
 996                 return -1;
 997
 998         while (1) {
 999                 if (cur <= rsv->rsv_end)
1000                         cur = rsv->rsv_end + 1;
1001
1002                 /* TODO?
1003                  * in the case we could not find a reservable space
1004                  * that is what is expected, during the re-search, we could
1005                  * remember what's the largest reservable space we could have
1006                  * and return that one.
1007                  *
1008                  * For now it will fail if we could not find the reservable
1009                  * space with expected-size (or more)...
1010                  */
1011                 if (cur > last_block)
1012                         return -1;              /* fail */
1013
1014                 prev = rsv;
1015                 next = rb_next(&rsv->rsv_node);
1016                 rsv = rb_entry(next,struct ext3_reserve_window_node,rsv_node);
1017
1018                 /*
1019                  * Reached the last reservation, we can just append to the
1020                  * previous one.
1021                  */
1022                 if (!next)
1023                         break;
1024
1025                 if (cur + size <= rsv->rsv_start) {
1026                         /*
1027                          * Found a reserveable space big enough.  We could
1028                          * have a reservation across the group boundary here
1029                          */
1030                         break;
1031                 }
1032         }
1033         /*
1034          * we come here either :
1035          * when we reach the end of the whole list,
1036          * and there is empty reservable space after last entry in the list.
1037          * append it to the end of the list.
1038          *
1039          * or we found one reservable space in the middle of the list,
1040          * return the reservation window that we could append to.
1041          * succeed.
1042          */
1043
1044         if ((prev != my_rsv) && (!rsv_is_empty(&my_rsv->rsv_window)))
1045                 rsv_window_remove(sb, my_rsv);
1046
1047         /*
1048          * Let's book the whole avaliable window for now.  We will check the
1049          * disk bitmap later and then, if there are free blocks then we adjust
1050          * the window size if it's larger than requested.
1051          * Otherwise, we will remove this node from the tree next time
1052          * call find_next_reservable_window.
1053          */
1054         my_rsv->rsv_start = cur;
1055         my_rsv->rsv_end = cur + size - 1;
1056         my_rsv->rsv_alloc_hit = 0;
1057
1058         if (prev != my_rsv)
1059                 ext3_rsv_window_add(sb, my_rsv);
1060
1061         return 0;
1062 }
1063
1064 /**
1065  *      alloc_new_reservation()--allocate a new reservation window
1066  *
1067  *              To make a new reservation, we search part of the filesystem
1068  *              reservation list (the list that inside the group). We try to
1069  *              allocate a new reservation window near the allocation goal,
1070  *              or the beginning of the group, if there is no goal.
1071  *
1072  *              We first find a reservable space after the goal, then from
1073  *              there, we check the bitmap for the first free block after
1074  *              it. If there is no free block until the end of group, then the
1075  *              whole group is full, we failed. Otherwise, check if the free
1076  *              block is inside the expected reservable space, if so, we
1077  *              succeed.
1078  *              If the first free block is outside the reservable space, then
1079  *              start from the first free block, we search for next available
1080  *              space, and go on.
1081  *
1082  *      on succeed, a new reservation will be found and inserted into the list
1083  *      It contains at least one free block, and it does not overlap with other
1084  *      reservation windows.
1085  *
1086  *      failed: we failed to find a reservation window in this group
1087  *
1088  *      @rsv: the reservation
1089  *
1090  *      @grp_goal: The goal (group-relative).  It is where the search for a
1091  *              free reservable space should start from.
1092  *              if we have a grp_goal(grp_goal >0 ), then start from there,
1093  *              no grp_goal(grp_goal = -1), we start from the first block
1094  *              of the group.
1095  *
1096  *      @sb: the super block
1097  *      @group: the group we are trying to allocate in
1098  *      @bitmap_bh: the block group block bitmap
1099  *
1100  */
1101 static int alloc_new_reservation(struct ext3_reserve_window_node *my_rsv,
1102                 ext3_grpblk_t grp_goal, struct super_block *sb,
1103                 unsigned int group, struct buffer_head *bitmap_bh)
1104 {
1105         struct ext3_reserve_window_node *search_head;
1106         ext3_fsblk_t group_first_block, group_end_block, start_block;
1107         ext3_grpblk_t first_free_block;
1108         struct rb_root *fs_rsv_root = &EXT3_SB(sb)->s_rsv_window_root;
1109         unsigned long size;
1110         int ret;
1111         spinlock_t *rsv_lock = &EXT3_SB(sb)->s_rsv_window_lock;
1112
1113         group_first_block = ext3_group_first_block_no(sb, group);
1114         group_end_block = group_first_block + (EXT3_BLOCKS_PER_GROUP(sb) - 1);
1115
1116         if (grp_goal < 0)
1117                 start_block = group_first_block;
1118         else
1119                 start_block = grp_goal + group_first_block;
1120
1121         size = my_rsv->rsv_goal_size;
1122
1123         if (!rsv_is_empty(&my_rsv->rsv_window)) {
1124                 /*
1125                  * if the old reservation is cross group boundary
1126                  * and if the goal is inside the old reservation window,
1127                  * we will come here when we just failed to allocate from
1128                  * the first part of the window. We still have another part
1129                  * that belongs to the next group. In this case, there is no
1130                  * point to discard our window and try to allocate a new one
1131                  * in this group(which will fail). we should
1132                  * keep the reservation window, just simply move on.
1133                  *
1134                  * Maybe we could shift the start block of the reservation
1135                  * window to the first block of next group.
1136                  */
1137
1138                 if ((my_rsv->rsv_start <= group_end_block) &&
1139                                 (my_rsv->rsv_end > group_end_block) &&
1140                                 (start_block >= my_rsv->rsv_start))
1141                         return -1;
1142
1143                 if ((my_rsv->rsv_alloc_hit >
1144                      (my_rsv->rsv_end - my_rsv->rsv_start + 1) / 2)) {
1145                         /*
1146                          * if the previously allocation hit ratio is
1147                          * greater than 1/2, then we double the size of
1148                          * the reservation window the next time,
1149                          * otherwise we keep the same size window
1150                          */
1151                         size = size * 2;
1152                         if (size > EXT3_MAX_RESERVE_BLOCKS)
1153                                 size = EXT3_MAX_RESERVE_BLOCKS;
1154                         my_rsv->rsv_goal_size= size;
1155                 }
1156         }
1157
1158         spin_lock(rsv_lock);
1159         /*
1160          * shift the search start to the window near the goal block
1161          */
1162         search_head = search_reserve_window(fs_rsv_root, start_block);
1163
1164         /*
1165          * find_next_reservable_window() simply finds a reservable window
1166          * inside the given range(start_block, group_end_block).
1167          *
1168          * To make sure the reservation window has a free bit inside it, we
1169          * need to check the bitmap after we found a reservable window.
1170          */
1171 retry:
1172         ret = find_next_reservable_window(search_head, my_rsv, sb,
1173                                                 start_block, group_end_block);
1174
1175         if (ret == -1) {
1176                 if (!rsv_is_empty(&my_rsv->rsv_window))
1177                         rsv_window_remove(sb, my_rsv);
1178                 spin_unlock(rsv_lock);
1179                 return -1;
1180         }
1181
1182         /*
1183          * On success, find_next_reservable_window() returns the
1184          * reservation window where there is a reservable space after it.
1185          * Before we reserve this reservable space, we need
1186          * to make sure there is at least a free block inside this region.
1187          *
1188          * searching the first free bit on the block bitmap and copy of
1189          * last committed bitmap alternatively, until we found a allocatable
1190          * block. Search start from the start block of the reservable space
1191          * we just found.
1192          */
1193         spin_unlock(rsv_lock);
1194         first_free_block = bitmap_search_next_usable_block(
1195                         my_rsv->rsv_start - group_first_block,
1196                         bitmap_bh, group_end_block - group_first_block + 1);
1197
1198         if (first_free_block < 0) {
1199                 /*
1200                  * no free block left on the bitmap, no point
1201                  * to reserve the space. return failed.
1202                  */
1203                 spin_lock(rsv_lock);
1204                 if (!rsv_is_empty(&my_rsv->rsv_window))
1205                         rsv_window_remove(sb, my_rsv);
1206                 spin_unlock(rsv_lock);
1207                 return -1;              /* failed */
1208         }
1209
1210         start_block = first_free_block + group_first_block;
1211         /*
1212          * check if the first free block is within the
1213          * free space we just reserved
1214          */
1215         if (start_block >= my_rsv->rsv_start && start_block <= my_rsv->rsv_end)
1216                 return 0;               /* success */
1217         /*
1218          * if the first free bit we found is out of the reservable space
1219          * continue search for next reservable space,
1220          * start from where the free block is,
1221          * we also shift the list head to where we stopped last time
1222          */
1223         search_head = my_rsv;
1224         spin_lock(rsv_lock);
1225         goto retry;
1226 }
1227
1228 /**
1229  * try_to_extend_reservation()
1230  * @my_rsv:             given reservation window
1231  * @sb:                 super block
1232  * @size:               the delta to extend
1233  *
1234  * Attempt to expand the reservation window large enough to have
1235  * required number of free blocks
1236  *
1237  * Since ext3_try_to_allocate() will always allocate blocks within
1238  * the reservation window range, if the window size is too small,
1239  * multiple blocks allocation has to stop at the end of the reservation
1240  * window. To make this more efficient, given the total number of
1241  * blocks needed and the current size of the window, we try to
1242  * expand the reservation window size if necessary on a best-effort
1243  * basis before ext3_new_blocks() tries to allocate blocks,
1244  */
1245 static void try_to_extend_reservation(struct ext3_reserve_window_node *my_rsv,
1246                         struct super_block *sb, int size)
1247 {
1248         struct ext3_reserve_window_node *next_rsv;
1249         struct rb_node *next;
1250         spinlock_t *rsv_lock = &EXT3_SB(sb)->s_rsv_window_lock;
1251
1252         if (!spin_trylock(rsv_lock))
1253                 return;
1254
1255         next = rb_next(&my_rsv->rsv_node);
1256
1257         if (!next)
1258                 my_rsv->rsv_end += size;
1259         else {
1260                 next_rsv = rb_entry(next, struct ext3_reserve_window_node, rsv_node);
1261
1262                 if ((next_rsv->rsv_start - my_rsv->rsv_end - 1) >= size)
1263                         my_rsv->rsv_end += size;
1264                 else
1265                         my_rsv->rsv_end = next_rsv->rsv_start - 1;
1266         }
1267         spin_unlock(rsv_lock);
1268 }
1269
1270 /**
1271  * ext3_try_to_allocate_with_rsv()
1272  * @sb:                 superblock
1273  * @handle:             handle to this transaction
1274  * @group:              given allocation block group
1275  * @bitmap_bh:          bufferhead holds the block bitmap
1276  * @grp_goal:           given target block within the group
1277  * @count:              target number of blocks to allocate
1278  * @my_rsv:             reservation window
1279  * @errp:               pointer to store the error code
1280  *
1281  * This is the main function used to allocate a new block and its reservation
1282  * window.
1283  *
1284  * Each time when a new block allocation is need, first try to allocate from
1285  * its own reservation.  If it does not have a reservation window, instead of
1286  * looking for a free bit on bitmap first, then look up the reservation list to
1287  * see if it is inside somebody else's reservation window, we try to allocate a
1288  * reservation window for it starting from the goal first. Then do the block
1289  * allocation within the reservation window.
1290  *
1291  * This will avoid keeping on searching the reservation list again and
1292  * again when somebody is looking for a free block (without
1293  * reservation), and there are lots of free blocks, but they are all
1294  * being reserved.
1295  *
1296  * We use a red-black tree for the per-filesystem reservation list.
1297  *
1298  */
1299 static ext3_grpblk_t
1300 ext3_try_to_allocate_with_rsv(struct super_block *sb, handle_t *handle,
1301                         unsigned int group, struct buffer_head *bitmap_bh,
1302                         ext3_grpblk_t grp_goal,
1303                         struct ext3_reserve_window_node * my_rsv,
1304                         unsigned long *count, int *errp)
1305 {
1306         ext3_fsblk_t group_first_block, group_last_block;
1307         ext3_grpblk_t ret = 0;
1308         int fatal;
1309         unsigned long num = *count;
1310
1311         *errp = 0;
1312
1313         /*
1314          * Make sure we use undo access for the bitmap, because it is critical
1315          * that we do the frozen_data COW on bitmap buffers in all cases even
1316          * if the buffer is in BJ_Forget state in the committing transaction.
1317          */
1318         BUFFER_TRACE(bitmap_bh, "get undo access for new block");
1319         fatal = ext3_journal_get_undo_access(handle, bitmap_bh);
1320         if (fatal) {
1321                 *errp = fatal;
1322                 return -1;
1323         }
1324
1325         /*
1326          * we don't deal with reservation when
1327          * filesystem is mounted without reservation
1328          * or the file is not a regular file
1329          * or last attempt to allocate a block with reservation turned on failed
1330          */
1331         if (my_rsv == NULL ) {
1332                 ret = ext3_try_to_allocate(sb, handle, group, bitmap_bh,
1333                                                 grp_goal, count, NULL);
1334                 goto out;
1335         }
1336         /*
1337          * grp_goal is a group relative block number (if there is a goal)
1338          * 0 <= grp_goal < EXT3_BLOCKS_PER_GROUP(sb)
1339          * first block is a filesystem wide block number
1340          * first block is the block number of the first block in this group
1341          */
1342         group_first_block = ext3_group_first_block_no(sb, group);
1343         group_last_block = group_first_block + (EXT3_BLOCKS_PER_GROUP(sb) - 1);
1344
1345         /*
1346          * Basically we will allocate a new block from inode's reservation
1347          * window.
1348          *
1349          * We need to allocate a new reservation window, if:
1350          * a) inode does not have a reservation window; or
1351          * b) last attempt to allocate a block from existing reservation
1352          *    failed; or
1353          * c) we come here with a goal and with a reservation window
1354          *
1355          * We do not need to allocate a new reservation window if we come here
1356          * at the beginning with a goal and the goal is inside the window, or
1357          * we don't have a goal but already have a reservation window.
1358          * then we could go to allocate from the reservation window directly.
1359          */
1360         while (1) {
1361                 if (rsv_is_empty(&my_rsv->rsv_window) || (ret < 0) ||
1362                         !goal_in_my_reservation(&my_rsv->rsv_window,
1363                                                 grp_goal, group, sb)) {
1364                         if (my_rsv->rsv_goal_size < *count)
1365                                 my_rsv->rsv_goal_size = *count;
1366                         ret = alloc_new_reservation(my_rsv, grp_goal, sb,
1367                                                         group, bitmap_bh);
1368                         if (ret < 0)
1369                                 break;                  /* failed */
1370
1371                         if (!goal_in_my_reservation(&my_rsv->rsv_window,
1372                                                         grp_goal, group, sb))
1373                                 grp_goal = -1;
1374                 } else if (grp_goal >= 0) {
1375                         int curr = my_rsv->rsv_end -
1376                                         (grp_goal + group_first_block) + 1;
1377
1378                         if (curr < *count)
1379                                 try_to_extend_reservation(my_rsv, sb,
1380                                                         *count - curr);
1381                 }
1382
1383                 if ((my_rsv->rsv_start > group_last_block) ||
1384                                 (my_rsv->rsv_end < group_first_block)) {
1385                         rsv_window_dump(&EXT3_SB(sb)->s_rsv_window_root, 1);
1386                         BUG();
1387                 }
1388                 ret = ext3_try_to_allocate(sb, handle, group, bitmap_bh,
1389                                            grp_goal, &num, &my_rsv->rsv_window);
1390                 if (ret >= 0) {
1391                         my_rsv->rsv_alloc_hit += num;
1392                         *count = num;
1393                         break;                          /* succeed */
1394                 }
1395                 num = *count;
1396         }
1397 out:
1398         if (ret >= 0) {
1399                 BUFFER_TRACE(bitmap_bh, "journal_dirty_metadata for "
1400                                         "bitmap block");
1401                 fatal = ext3_journal_dirty_metadata(handle, bitmap_bh);
1402                 if (fatal) {
1403                         *errp = fatal;
1404                         return -1;
1405                 }
1406                 return ret;
1407         }
1408
1409         BUFFER_TRACE(bitmap_bh, "journal_release_buffer");
1410         ext3_journal_release_buffer(handle, bitmap_bh);
1411         return ret;
1412 }
1413
1414 /**
1415  * ext3_has_free_blocks()
1416  * @sbi:                in-core super block structure.
1417  *
1418  * Check if filesystem has at least 1 free block available for allocation.
1419  */
1420 static int ext3_has_free_blocks(struct ext3_sb_info *sbi)
1421 {
1422         ext3_fsblk_t free_blocks, root_blocks;
1423
1424         free_blocks = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
1425         root_blocks = le32_to_cpu(sbi->s_es->s_r_blocks_count);
1426         if (free_blocks < root_blocks + 1 && !capable(CAP_SYS_RESOURCE) &&
1427                 sbi->s_resuid != current->fsuid &&
1428                 (sbi->s_resgid == 0 || !in_group_p (sbi->s_resgid))) {
1429                 return 0;
1430         }
1431         return 1;
1432 }
1433
1434 /**
1435  * ext3_should_retry_alloc()
1436  * @sb:                 super block
1437  * @retries             number of attemps has been made
1438  *
1439  * ext3_should_retry_alloc() is called when ENOSPC is returned, and if
1440  * it is profitable to retry the operation, this function will wait
1441  * for the current or commiting transaction to complete, and then
1442  * return TRUE.
1443  *
1444  * if the total number of retries exceed three times, return FALSE.
1445  */
1446 int ext3_should_retry_alloc(struct super_block *sb, int *retries)
1447 {
1448         if (!ext3_has_free_blocks(EXT3_SB(sb)) || (*retries)++ > 3)
1449                 return 0;
1450
1451         jbd_debug(1, "%s: retrying operation after ENOSPC\n", sb->s_id);
1452
1453         return journal_force_commit_nested(EXT3_SB(sb)->s_journal);
1454 }
1455
1456 /**
1457  * ext3_new_blocks() -- core block(s) allocation function
1458  * @handle:             handle to this transaction
1459  * @inode:              file inode
1460  * @goal:               given target block(filesystem wide)
1461  * @count:              target number of blocks to allocate
1462  * @errp:               error code
1463  *
1464  * ext3_new_blocks uses a goal block to assist allocation.  It tries to
1465  * allocate block(s) from the block group contains the goal block first. If that
1466  * fails, it will try to allocate block(s) from other block groups without
1467  * any specific goal block.
1468  *
1469  */
1470 ext3_fsblk_t ext3_new_blocks(handle_t *handle, struct inode *inode,
1471                         ext3_fsblk_t goal, unsigned long *count, int *errp)
1472 {
1473         struct buffer_head *bitmap_bh = NULL;
1474         struct buffer_head *gdp_bh;
1475         int group_no;
1476         int goal_group;
1477         ext3_grpblk_t grp_target_blk;   /* blockgroup relative goal block */
1478         ext3_grpblk_t grp_alloc_blk;    /* blockgroup-relative allocated block*/
1479         ext3_fsblk_t ret_block;         /* filesyetem-wide allocated block */
1480         int bgi;                        /* blockgroup iteration index */
1481         int fatal = 0, err;
1482         int performed_allocation = 0;
1483         ext3_grpblk_t free_blocks;      /* number of free blocks in a group */
1484         struct super_block *sb;
1485         struct ext3_group_desc *gdp;
1486         struct ext3_super_block *es;
1487         struct ext3_sb_info *sbi;
1488         struct ext3_reserve_window_node *my_rsv = NULL;
1489         struct ext3_block_alloc_info *block_i;
1490         unsigned short windowsz = 0;
1491 #ifdef EXT3FS_DEBUG
1492         static int goal_hits, goal_attempts;
1493 #endif
1494         unsigned long ngroups;
1495         unsigned long num = *count;
1496
1497         *errp = -ENOSPC;
1498         sb = inode->i_sb;
1499         if (!sb) {
1500                 printk("ext3_new_block: nonexistent device");
1501                 return 0;
1502         }
1503
1504         /*
1505          * Check quota for allocation of this block.
1506          */
1507         if (DQUOT_ALLOC_BLOCK(inode, num)) {
1508                 *errp = -EDQUOT;
1509                 return 0;
1510         }
1511
1512         sbi = EXT3_SB(sb);
1513         es = EXT3_SB(sb)->s_es;
1514         ext3_debug("goal=%lu.\n", goal);
1515         /*
1516          * Allocate a block from reservation only when
1517          * filesystem is mounted with reservation(default,-o reservation), and
1518          * it's a regular file, and
1519          * the desired window size is greater than 0 (One could use ioctl
1520          * command EXT3_IOC_SETRSVSZ to set the window size to 0 to turn off
1521          * reservation on that particular file)
1522          */
1523         block_i = EXT3_I(inode)->i_block_alloc_info;
1524         if (block_i && ((windowsz = block_i->rsv_window_node.rsv_goal_size) > 0))
1525                 my_rsv = &block_i->rsv_window_node;
1526
1527         if (!ext3_has_free_blocks(sbi)) {
1528                 *errp = -ENOSPC;
1529                 goto out;
1530         }
1531
1532         /*
1533          * First, test whether the goal block is free.
1534          */
1535         if (goal < le32_to_cpu(es->s_first_data_block) ||
1536             goal >= le32_to_cpu(es->s_blocks_count))
1537                 goal = le32_to_cpu(es->s_first_data_block);
1538         group_no = (goal - le32_to_cpu(es->s_first_data_block)) /
1539                         EXT3_BLOCKS_PER_GROUP(sb);
1540         goal_group = group_no;
1541 retry_alloc:
1542         gdp = ext3_get_group_desc(sb, group_no, &gdp_bh);
1543         if (!gdp)
1544                 goto io_error;
1545
1546         free_blocks = le16_to_cpu(gdp->bg_free_blocks_count);
1547         /*
1548          * if there is not enough free blocks to make a new resevation
1549          * turn off reservation for this allocation
1550          */
1551         if (my_rsv && (free_blocks < windowsz)
1552                 && (rsv_is_empty(&my_rsv->rsv_window)))
1553                 my_rsv = NULL;
1554
1555         if (free_blocks > 0) {
1556                 grp_target_blk = ((goal - le32_to_cpu(es->s_first_data_block)) %
1557                                 EXT3_BLOCKS_PER_GROUP(sb));
1558                 bitmap_bh = read_block_bitmap(sb, group_no);
1559                 if (!bitmap_bh)
1560                         goto io_error;
1561                 grp_alloc_blk = ext3_try_to_allocate_with_rsv(sb, handle,
1562                                         group_no, bitmap_bh, grp_target_blk,
1563                                         my_rsv, &num, &fatal);
1564                 if (fatal)
1565                         goto out;
1566                 if (grp_alloc_blk >= 0)
1567                         goto allocated;
1568         }
1569
1570         ngroups = EXT3_SB(sb)->s_groups_count;
1571         smp_rmb();
1572
1573         /*
1574          * Now search the rest of the groups.  We assume that
1575          * i and gdp correctly point to the last group visited.
1576          */
1577         for (bgi = 0; bgi < ngroups; bgi++) {
1578                 group_no++;
1579                 if (group_no >= ngroups)
1580                         group_no = 0;
1581                 gdp = ext3_get_group_desc(sb, group_no, &gdp_bh);
1582                 if (!gdp)
1583                         goto io_error;
1584                 free_blocks = le16_to_cpu(gdp->bg_free_blocks_count);
1585                 /*
1586                  * skip this group if the number of
1587                  * free blocks is less than half of the reservation
1588                  * window size.
1589                  */
1590                 if (free_blocks <= (windowsz/2))
1591                         continue;
1592
1593                 brelse(bitmap_bh);
1594                 bitmap_bh = read_block_bitmap(sb, group_no);
1595                 if (!bitmap_bh)
1596                         goto io_error;
1597                 /*
1598                  * try to allocate block(s) from this group, without a goal(-1).
1599                  */
1600                 grp_alloc_blk = ext3_try_to_allocate_with_rsv(sb, handle,
1601                                         group_no, bitmap_bh, -1, my_rsv,
1602                                         &num, &fatal);
1603                 if (fatal)
1604                         goto out;
1605                 if (grp_alloc_blk >= 0)
1606                         goto allocated;
1607         }
1608         /*
1609          * We may end up a bogus ealier ENOSPC error due to
1610          * filesystem is "full" of reservations, but
1611          * there maybe indeed free blocks avaliable on disk
1612          * In this case, we just forget about the reservations
1613          * just do block allocation as without reservations.
1614          */
1615         if (my_rsv) {
1616                 my_rsv = NULL;
1617                 windowsz = 0;
1618                 group_no = goal_group;
1619                 goto retry_alloc;
1620         }
1621         /* No space left on the device */
1622         *errp = -ENOSPC;
1623         goto out;
1624
1625 allocated:
1626
1627         ext3_debug("using block group %d(%d)\n",
1628                         group_no, gdp->bg_free_blocks_count);
1629
1630         BUFFER_TRACE(gdp_bh, "get_write_access");
1631         fatal = ext3_journal_get_write_access(handle, gdp_bh);
1632         if (fatal)
1633                 goto out;
1634
1635         ret_block = grp_alloc_blk + ext3_group_first_block_no(sb, group_no);
1636
1637         if (in_range(le32_to_cpu(gdp->bg_block_bitmap), ret_block, num) ||
1638             in_range(le32_to_cpu(gdp->bg_inode_bitmap), ret_block, num) ||
1639             in_range(ret_block, le32_to_cpu(gdp->bg_inode_table),
1640                       EXT3_SB(sb)->s_itb_per_group) ||
1641             in_range(ret_block + num - 1, le32_to_cpu(gdp->bg_inode_table),
1642                       EXT3_SB(sb)->s_itb_per_group)) {
1643                 ext3_error(sb, "ext3_new_block",
1644                             "Allocating block in system zone - "
1645                             "blocks from "E3FSBLK", length %lu",
1646                              ret_block, num);
1647                 goto out;
1648         }
1649
1650         performed_allocation = 1;
1651
1652 #ifdef CONFIG_JBD_DEBUG
1653         {
1654                 struct buffer_head *debug_bh;
1655
1656                 /* Record bitmap buffer state in the newly allocated block */
1657                 debug_bh = sb_find_get_block(sb, ret_block);
1658                 if (debug_bh) {
1659                         BUFFER_TRACE(debug_bh, "state when allocated");
1660                         BUFFER_TRACE2(debug_bh, bitmap_bh, "bitmap state");
1661                         brelse(debug_bh);
1662                 }
1663         }
1664         jbd_lock_bh_state(bitmap_bh);
1665         spin_lock(sb_bgl_lock(sbi, group_no));
1666         if (buffer_jbd(bitmap_bh) && bh2jh(bitmap_bh)->b_committed_data) {
1667                 int i;
1668
1669                 for (i = 0; i < num; i++) {
1670                         if (ext3_test_bit(grp_alloc_blk+i,
1671                                         bh2jh(bitmap_bh)->b_committed_data)) {
1672                                 printk("%s: block was unexpectedly set in "
1673                                         "b_committed_data\n", __FUNCTION__);
1674                         }
1675                 }
1676         }
1677         ext3_debug("found bit %d\n", grp_alloc_blk);
1678         spin_unlock(sb_bgl_lock(sbi, group_no));
1679         jbd_unlock_bh_state(bitmap_bh);
1680 #endif
1681
1682         if (ret_block + num - 1 >= le32_to_cpu(es->s_blocks_count)) {
1683                 ext3_error(sb, "ext3_new_block",
1684                             "block("E3FSBLK") >= blocks count(%d) - "
1685                             "block_group = %d, es == %p ", ret_block,
1686                         le32_to_cpu(es->s_blocks_count), group_no, es);
1687                 goto out;
1688         }
1689
1690         /*
1691          * It is up to the caller to add the new buffer to a journal
1692          * list of some description.  We don't know in advance whether
1693          * the caller wants to use it as metadata or data.
1694          */
1695         ext3_debug("allocating block %lu. Goal hits %d of %d.\n",
1696                         ret_block, goal_hits, goal_attempts);
1697
1698         spin_lock(sb_bgl_lock(sbi, group_no));
1699         gdp->bg_free_blocks_count =
1700                         cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count)-num);
1701         spin_unlock(sb_bgl_lock(sbi, group_no));
1702         percpu_counter_sub(&sbi->s_freeblocks_counter, num);
1703
1704         BUFFER_TRACE(gdp_bh, "journal_dirty_metadata for group descriptor");
1705         err = ext3_journal_dirty_metadata(handle, gdp_bh);
1706         if (!fatal)
1707                 fatal = err;
1708
1709         sb->s_dirt = 1;
1710         if (fatal)
1711                 goto out;
1712
1713         *errp = 0;
1714         brelse(bitmap_bh);
1715         DQUOT_FREE_BLOCK(inode, *count-num);
1716         *count = num;
1717         return ret_block;
1718
1719 io_error:
1720         *errp = -EIO;
1721 out:
1722         if (fatal) {
1723                 *errp = fatal;
1724                 ext3_std_error(sb, fatal);
1725         }
1726         /*
1727          * Undo the block allocation
1728          */
1729         if (!performed_allocation)
1730                 DQUOT_FREE_BLOCK(inode, *count);
1731         brelse(bitmap_bh);
1732         return 0;
1733 }
1734
1735 ext3_fsblk_t ext3_new_block(handle_t *handle, struct inode *inode,
1736                         ext3_fsblk_t goal, int *errp)
1737 {
1738         unsigned long count = 1;
1739
1740         return ext3_new_blocks(handle, inode, goal, &count, errp);
1741 }
1742
1743 /**
1744  * ext3_count_free_blocks() -- count filesystem free blocks
1745  * @sb:         superblock
1746  *
1747  * Adds up the number of free blocks from each block group.
1748  */
1749 ext3_fsblk_t ext3_count_free_blocks(struct super_block *sb)
1750 {
1751         ext3_fsblk_t desc_count;
1752         struct ext3_group_desc *gdp;
1753         int i;
1754         unsigned long ngroups = EXT3_SB(sb)->s_groups_count;
1755 #ifdef EXT3FS_DEBUG
1756         struct ext3_super_block *es;
1757         ext3_fsblk_t bitmap_count;
1758         unsigned long x;
1759         struct buffer_head *bitmap_bh = NULL;
1760
1761         es = EXT3_SB(sb)->s_es;
1762         desc_count = 0;
1763         bitmap_count = 0;
1764         gdp = NULL;
1765
1766         smp_rmb();
1767         for (i = 0; i < ngroups; i++) {
1768                 gdp = ext3_get_group_desc(sb, i, NULL);
1769                 if (!gdp)
1770                         continue;
1771                 desc_count += le16_to_cpu(gdp->bg_free_blocks_count);
1772                 brelse(bitmap_bh);
1773                 bitmap_bh = read_block_bitmap(sb, i);
1774                 if (bitmap_bh == NULL)
1775                         continue;
1776
1777                 x = ext3_count_free(bitmap_bh, sb->s_blocksize);
1778                 printk("group %d: stored = %d, counted = %lu\n",
1779                         i, le16_to_cpu(gdp->bg_free_blocks_count), x);
1780                 bitmap_count += x;
1781         }
1782         brelse(bitmap_bh);
1783         printk("ext3_count_free_blocks: stored = "E3FSBLK
1784                 ", computed = "E3FSBLK", "E3FSBLK"\n",
1785                le32_to_cpu(es->s_free_blocks_count),
1786                 desc_count, bitmap_count);
1787         return bitmap_count;
1788 #else
1789         desc_count = 0;
1790         smp_rmb();
1791         for (i = 0; i < ngroups; i++) {
1792                 gdp = ext3_get_group_desc(sb, i, NULL);
1793                 if (!gdp)
1794                         continue;
1795                 desc_count += le16_to_cpu(gdp->bg_free_blocks_count);
1796         }
1797
1798         return desc_count;
1799 #endif
1800 }
1801
1802 static inline int test_root(int a, int b)
1803 {
1804         int num = b;
1805
1806         while (a > num)
1807                 num *= b;
1808         return num == a;
1809 }
1810
1811 static int ext3_group_sparse(int group)
1812 {
1813         if (group <= 1)
1814                 return 1;
1815         if (!(group & 1))
1816                 return 0;
1817         return (test_root(group, 7) || test_root(group, 5) ||
1818                 test_root(group, 3));
1819 }
1820
1821 /**
1822  *      ext3_bg_has_super - number of blocks used by the superblock in group
1823  *      @sb: superblock for filesystem
1824  *      @group: group number to check
1825  *
1826  *      Return the number of blocks used by the superblock (primary or backup)
1827  *      in this group.  Currently this will be only 0 or 1.
1828  */
1829 int ext3_bg_has_super(struct super_block *sb, int group)
1830 {
1831         if (EXT3_HAS_RO_COMPAT_FEATURE(sb,
1832                                 EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER) &&
1833                         !ext3_group_sparse(group))
1834                 return 0;
1835         return 1;
1836 }
1837
1838 static unsigned long ext3_bg_num_gdb_meta(struct super_block *sb, int group)
1839 {
1840         unsigned long metagroup = group / EXT3_DESC_PER_BLOCK(sb);
1841         unsigned long first = metagroup * EXT3_DESC_PER_BLOCK(sb);
1842         unsigned long last = first + EXT3_DESC_PER_BLOCK(sb) - 1;
1843
1844         if (group == first || group == first + 1 || group == last)
1845                 return 1;
1846         return 0;
1847 }
1848
1849 static unsigned long ext3_bg_num_gdb_nometa(struct super_block *sb, int group)
1850 {
1851         if (EXT3_HAS_RO_COMPAT_FEATURE(sb,
1852                                 EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER) &&
1853                         !ext3_group_sparse(group))
1854                 return 0;
1855         return EXT3_SB(sb)->s_gdb_count;
1856 }
1857
1858 /**
1859  *      ext3_bg_num_gdb - number of blocks used by the group table in group
1860  *      @sb: superblock for filesystem
1861  *      @group: group number to check
1862  *
1863  *      Return the number of blocks used by the group descriptor table
1864  *      (primary or backup) in this group.  In the future there may be a
1865  *      different number of descriptor blocks in each group.
1866  */
1867 unsigned long ext3_bg_num_gdb(struct super_block *sb, int group)
1868 {
1869         unsigned long first_meta_bg =
1870                         le32_to_cpu(EXT3_SB(sb)->s_es->s_first_meta_bg);
1871         unsigned long metagroup = group / EXT3_DESC_PER_BLOCK(sb);
1872
1873         if (!EXT3_HAS_INCOMPAT_FEATURE(sb,EXT3_FEATURE_INCOMPAT_META_BG) ||
1874                         metagroup < first_meta_bg)
1875                 return ext3_bg_num_gdb_nometa(sb,group);
1876
1877         return ext3_bg_num_gdb_meta(sb,group);
1878
1879 }