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 #include <linux/vs_base.h>
  23 #include <linux/vs_dlimit.h>
  24
  25 /*
  26  * balloc.c contains the blocks allocation and deallocation routines
  27  */
  28
  29 /*
  30  * The free blocks are managed by bitmaps.  A file system contains several
  31  * blocks groups.  Each group contains 1 bitmap block for blocks, 1 bitmap
  32  * block for inodes, N blocks for the inode table and data blocks.
  33  *
  34  * The file system contains group descriptors which are located after the
  35  * super block.  Each descriptor contains the number of the bitmap block and
  36  * the free blocks count in the block.  The descriptors are loaded in memory
  37  * when a file system is mounted (see ext3_read_super).
  38  */
  39
  40
  41 #define in_range(b, first, len) ((b) >= (first) && (b) <= (first) + (len) - 1)
  42
  43 struct ext3_group_desc * ext3_get_group_desc(struct super_block * sb,
  44                                              unsigned int block_group,
  45                                              struct buffer_head ** bh)
  46 {
  47         unsigned long group_desc;
  48         unsigned long offset;
  49         struct ext3_group_desc * desc;
  50         struct ext3_sb_info *sbi = EXT3_SB(sb);
  51
  52         if (block_group >= sbi->s_groups_count) {
  53                 ext3_error (sb, "ext3_get_group_desc",
  54                             "block_group >= groups_count - "
  55                             "block_group = %d, groups_count = %lu",
  56                             block_group, sbi->s_groups_count);
  57
  58                 return NULL;
  59         }
  60         smp_rmb();
  61
  62         group_desc = block_group >> EXT3_DESC_PER_BLOCK_BITS(sb);
  63         offset = block_group & (EXT3_DESC_PER_BLOCK(sb) - 1);
  64         if (!sbi->s_group_desc[group_desc]) {
  65                 ext3_error (sb, "ext3_get_group_desc",
  66                             "Group descriptor not loaded - "
  67                             "block_group = %d, group_desc = %lu, desc = %lu",
  68                              block_group, group_desc, offset);
  69                 return NULL;
  70         }
  71
  72         desc = (struct ext3_group_desc *) sbi->s_group_desc[group_desc]->b_data;
  73         if (bh)
  74                 *bh = sbi->s_group_desc[group_desc];
  75         return desc + offset;
  76 }
  77
  78 /*
  79  * Read the bitmap for a given block_group, reading into the specified
  80  * slot in the superblock's bitmap cache.
  81  *
  82  * Return buffer_head on success or NULL in case of failure.
  83  */
  84 static struct buffer_head *
  85 read_block_bitmap(struct super_block *sb, unsigned int block_group)
  86 {
  87         struct ext3_group_desc * desc;
  88         struct buffer_head * bh = NULL;
  89
  90         desc = ext3_get_group_desc (sb, block_group, NULL);
  91         if (!desc)
  92                 goto error_out;
  93         bh = sb_bread(sb, le32_to_cpu(desc->bg_block_bitmap));
  94         if (!bh)
  95                 ext3_error (sb, "read_block_bitmap",
  96                             "Cannot read block bitmap - "
  97                             "block_group = %d, block_bitmap = %u",
  98                             block_group, le32_to_cpu(desc->bg_block_bitmap));
  99 error_out:
 100         return bh;
 101 }
 102 /*
 103  * The reservation window structure operations
 104  * --------------------------------------------
 105  * Operations include:
 106  * dump, find, add, remove, is_empty, find_next_reservable_window, etc.
 107  *
 108  * We use sorted double linked list for the per-filesystem reservation
 109  * window list. (like in vm_region).
 110  *
 111  * Initially, we keep those small operations in the abstract functions,
 112  * so later if we need a better searching tree than double linked-list,
 113  * we could easily switch to that without changing too much
 114  * code.
 115  */
 116 #if 0
 117 static void __rsv_window_dump(struct rb_root *root, int verbose,
 118                               const char *fn)
 119 {
 120         struct rb_node *n;
 121         struct ext3_reserve_window_node *rsv, *prev;
 122         int bad;
 123
 124 restart:
 125         n = rb_first(root);
 126         bad = 0;
 127         prev = NULL;
 128
 129         printk("Block Allocation Reservation Windows Map (%s):\n", fn);
 130         while (n) {
 131                 rsv = list_entry(n, struct ext3_reserve_window_node, rsv_node);
 132                 if (verbose)
 133                         printk("reservation window 0x%p "
 134                                "start:  %d, end:  %d\n",
 135                                rsv, rsv->rsv_start, rsv->rsv_end);
 136                 if (rsv->rsv_start && rsv->rsv_start >= rsv->rsv_end) {
 137                         printk("Bad reservation %p (start >= end)\n",
 138                                rsv);
 139                         bad = 1;
 140                 }
 141                 if (prev && prev->rsv_end >= rsv->rsv_start) {
 142                         printk("Bad reservation %p (prev->end >= start)\n",
 143                                rsv);
 144                         bad = 1;
 145                 }
 146                 if (bad) {
 147                         if (!verbose) {
 148                                 printk("Restarting reservation walk in verbose mode\n");
 149                                 verbose = 1;
 150                                 goto restart;
 151                         }
 152                 }
 153                 n = rb_next(n);
 154                 prev = rsv;
 155         }
 156         printk("Window map complete.\n");
 157         if (bad)
 158                 BUG();
 159 }
 160 #define rsv_window_dump(root, verbose) \
 161         __rsv_window_dump((root), (verbose), __FUNCTION__)
 162 #else
 163 #define rsv_window_dump(root, verbose) do {} while (0)
 164 #endif
 165
 166 static int
 167 goal_in_my_reservation(struct ext3_reserve_window *rsv, ext3_grpblk_t grp_goal,
 168                         unsigned int group, struct super_block * sb)
 169 {
 170         ext3_fsblk_t group_first_block, group_last_block;
 171
 172         group_first_block = ext3_group_first_block_no(sb, group);
 173         group_last_block = group_first_block + (EXT3_BLOCKS_PER_GROUP(sb) - 1);
 174
 175         if ((rsv->_rsv_start > group_last_block) ||
 176             (rsv->_rsv_end < group_first_block))
 177                 return 0;
 178         if ((grp_goal >= 0) && ((grp_goal + group_first_block < rsv->_rsv_start)
 179                 || (grp_goal + group_first_block > rsv->_rsv_end)))
 180                 return 0;
 181         return 1;
 182 }
 183
 184 /*
 185  * Find the reserved window which includes the goal, or the previous one
 186  * if the goal is not in any window.
 187  * Returns NULL if there are no windows or if all windows start after the goal.
 188  */
 189 static struct ext3_reserve_window_node *
 190 search_reserve_window(struct rb_root *root, ext3_fsblk_t goal)
 191 {
 192         struct rb_node *n = root->rb_node;
 193         struct ext3_reserve_window_node *rsv;
 194
 195         if (!n)
 196                 return NULL;
 197
 198         do {
 199                 rsv = rb_entry(n, struct ext3_reserve_window_node, rsv_node);
 200
 201                 if (goal < rsv->rsv_start)
 202                         n = n->rb_left;
 203                 else if (goal > rsv->rsv_end)
 204                         n = n->rb_right;
 205                 else
 206                         return rsv;
 207         } while (n);
 208         /*
 209          * We've fallen off the end of the tree: the goal wasn't inside
 210          * any particular node.  OK, the previous node must be to one
 211          * side of the interval containing the goal.  If it's the RHS,
 212          * we need to back up one.
 213          */
 214         if (rsv->rsv_start > goal) {
 215                 n = rb_prev(&rsv->rsv_node);
 216                 rsv = rb_entry(n, struct ext3_reserve_window_node, rsv_node);
 217         }
 218         return rsv;
 219 }
 220
 221 void ext3_rsv_window_add(struct super_block *sb,
 222                     struct ext3_reserve_window_node *rsv)
 223 {
 224         struct rb_root *root = &EXT3_SB(sb)->s_rsv_window_root;
 225         struct rb_node *node = &rsv->rsv_node;
 226         ext3_fsblk_t start = rsv->rsv_start;
 227
 228         struct rb_node ** p = &root->rb_node;
 229         struct rb_node * parent = NULL;
 230         struct ext3_reserve_window_node *this;
 231
 232         while (*p)
 233         {
 234                 parent = *p;
 235                 this = rb_entry(parent, struct ext3_reserve_window_node, rsv_node);
 236
 237                 if (start < this->rsv_start)
 238                         p = &(*p)->rb_left;
 239                 else if (start > this->rsv_end)
 240                         p = &(*p)->rb_right;
 241                 else
 242                         BUG();
 243         }
 244
 245         rb_link_node(node, parent, p);
 246         rb_insert_color(node, root);
 247 }
 248
 249 static void rsv_window_remove(struct super_block *sb,
 250                               struct ext3_reserve_window_node *rsv)
 251 {
 252         rsv->rsv_start = EXT3_RESERVE_WINDOW_NOT_ALLOCATED;
 253         rsv->rsv_end = EXT3_RESERVE_WINDOW_NOT_ALLOCATED;
 254         rsv->rsv_alloc_hit = 0;
 255         rb_erase(&rsv->rsv_node, &EXT3_SB(sb)->s_rsv_window_root);
 256 }
 257
 258 static inline int rsv_is_empty(struct ext3_reserve_window *rsv)
 259 {
 260         /* a valid reservation end block could not be 0 */
 261         return (rsv->_rsv_end == EXT3_RESERVE_WINDOW_NOT_ALLOCATED);
 262 }
 263 void ext3_init_block_alloc_info(struct inode *inode)
 264 {
 265         struct ext3_inode_info *ei = EXT3_I(inode);
 266         struct ext3_block_alloc_info *block_i = ei->i_block_alloc_info;
 267         struct super_block *sb = inode->i_sb;
 268
 269         block_i = kmalloc(sizeof(*block_i), GFP_NOFS);
 270         if (block_i) {
 271                 struct ext3_reserve_window_node *rsv = &block_i->rsv_window_node;
 272
 273                 rsv->rsv_start = EXT3_RESERVE_WINDOW_NOT_ALLOCATED;
 274                 rsv->rsv_end = EXT3_RESERVE_WINDOW_NOT_ALLOCATED;
 275
 276                 /*
 277                  * if filesystem is mounted with NORESERVATION, the goal
 278                  * reservation window size is set to zero to indicate
 279                  * block reservation is off
 280                  */
 281                 if (!test_opt(sb, RESERVATION))
 282                         rsv->rsv_goal_size = 0;
 283                 else
 284                         rsv->rsv_goal_size = EXT3_DEFAULT_RESERVE_BLOCKS;
 285                 rsv->rsv_alloc_hit = 0;
 286                 block_i->last_alloc_logical_block = 0;
 287                 block_i->last_alloc_physical_block = 0;
 288         }
 289         ei->i_block_alloc_info = block_i;
 290 }
 291
 292 void ext3_discard_reservation(struct inode *inode)
 293 {
 294         struct ext3_inode_info *ei = EXT3_I(inode);
 295         struct ext3_block_alloc_info *block_i = ei->i_block_alloc_info;
 296         struct ext3_reserve_window_node *rsv;
 297         spinlock_t *rsv_lock = &EXT3_SB(inode->i_sb)->s_rsv_window_lock;
 298
 299         if (!block_i)
 300                 return;
 301
 302         rsv = &block_i->rsv_window_node;
 303         if (!rsv_is_empty(&rsv->rsv_window)) {
 304                 spin_lock(rsv_lock);
 305                 if (!rsv_is_empty(&rsv->rsv_window))
 306                         rsv_window_remove(inode->i_sb, rsv);
 307                 spin_unlock(rsv_lock);
 308         }
 309 }
 310
 311 /* Free given blocks, update quota and i_blocks field */
 312 void ext3_free_blocks_sb(handle_t *handle, struct super_block *sb,
 313                          ext3_fsblk_t block, unsigned long count,
 314                          unsigned long *pdquot_freed_blocks)
 315 {
 316         struct buffer_head *bitmap_bh = NULL;
 317         struct buffer_head *gd_bh;
 318         unsigned long block_group;
 319         ext3_grpblk_t bit;
 320         unsigned long i;
 321         unsigned long overflow;
 322         struct ext3_group_desc * desc;
 323         struct ext3_super_block * es;
 324         struct ext3_sb_info *sbi;
 325         int err = 0, ret;
 326         ext3_grpblk_t group_freed;
 327
 328         *pdquot_freed_blocks = 0;
 329         sbi = EXT3_SB(sb);
 330         es = sbi->s_es;
 331         if (block < le32_to_cpu(es->s_first_data_block) ||
 332             block + count < block ||
 333             block + count > le32_to_cpu(es->s_blocks_count)) {
 334                 ext3_error (sb, "ext3_free_blocks",
 335                             "Freeing blocks not in datazone - "
 336                             "block = "E3FSBLK", count = %lu", block, count);
 337                 goto error_return;
 338         }
 339
 340         ext3_debug ("freeing block(s) %lu-%lu\n", block, block + count - 1);
 341
 342 do_more:
 343         overflow = 0;
 344         block_group = (block - le32_to_cpu(es->s_first_data_block)) /
 345                       EXT3_BLOCKS_PER_GROUP(sb);
 346         bit = (block - le32_to_cpu(es->s_first_data_block)) %
 347                       EXT3_BLOCKS_PER_GROUP(sb);
 348         /*
 349          * Check to see if we are freeing blocks across a group
 350          * boundary.
 351          */
 352         if (bit + count > EXT3_BLOCKS_PER_GROUP(sb)) {
 353                 overflow = bit + count - EXT3_BLOCKS_PER_GROUP(sb);
 354                 count -= overflow;
 355         }
 356         brelse(bitmap_bh);
 357         bitmap_bh = read_block_bitmap(sb, block_group);
 358         if (!bitmap_bh)
 359                 goto error_return;
 360         desc = ext3_get_group_desc (sb, block_group, &gd_bh);
 361         if (!desc)
 362                 goto error_return;
 363
 364         if (in_range (le32_to_cpu(desc->bg_block_bitmap), block, count) ||
 365             in_range (le32_to_cpu(desc->bg_inode_bitmap), block, count) ||
 366             in_range (block, le32_to_cpu(desc->bg_inode_table),
 367                       sbi->s_itb_per_group) ||
 368             in_range (block + count - 1, le32_to_cpu(desc->bg_inode_table),
 369                       sbi->s_itb_per_group))
 370                 ext3_error (sb, "ext3_free_blocks",
 371                             "Freeing blocks in system zones - "
 372                             "Block = "E3FSBLK", count = %lu",
 373                             block, count);
 374
 375         /*
 376          * We are about to start releasing blocks in the bitmap,
 377          * so we need undo access.
 378          */
 379         /* @@@ check errors */
 380         BUFFER_TRACE(bitmap_bh, "getting undo access");
 381         err = ext3_journal_get_undo_access(handle, bitmap_bh);
 382         if (err)
 383                 goto error_return;
 384
 385         /*
 386          * We are about to modify some metadata.  Call the journal APIs
 387          * to unshare ->b_data if a currently-committing transaction is
 388          * using it
 389          */
 390         BUFFER_TRACE(gd_bh, "get_write_access");
 391         err = ext3_journal_get_write_access(handle, gd_bh);
 392         if (err)
 393                 goto error_return;
 394
 395         jbd_lock_bh_state(bitmap_bh);
 396
 397         for (i = 0, group_freed = 0; i < count; i++) {
 398                 /*
 399                  * An HJ special.  This is expensive...
 400                  */
 401 #ifdef CONFIG_JBD_DEBUG
 402                 jbd_unlock_bh_state(bitmap_bh);
 403                 {
 404                         struct buffer_head *debug_bh;
 405                         debug_bh = sb_find_get_block(sb, block + i);
 406                         if (debug_bh) {
 407                                 BUFFER_TRACE(debug_bh, "Deleted!");
 408                                 if (!bh2jh(bitmap_bh)->b_committed_data)
 409                                         BUFFER_TRACE(debug_bh,
 410                                                 "No commited data in bitmap");
 411                                 BUFFER_TRACE2(debug_bh, bitmap_bh, "bitmap");
 412                                 __brelse(debug_bh);
 413                         }
 414                 }
 415                 jbd_lock_bh_state(bitmap_bh);
 416 #endif
 417                 if (need_resched()) {
 418                         jbd_unlock_bh_state(bitmap_bh);
 419                         cond_resched();
 420                         jbd_lock_bh_state(bitmap_bh);
 421                 }
 422                 /* @@@ This prevents newly-allocated data from being
 423                  * freed and then reallocated within the same
 424                  * transaction.
 425                  *
 426                  * Ideally we would want to allow that to happen, but to
 427                  * do so requires making journal_forget() capable of
 428                  * revoking the queued write of a data block, which
 429                  * implies blocking on the journal lock.  *forget()
 430                  * cannot block due to truncate races.
 431                  *
 432                  * Eventually we can fix this by making journal_forget()
 433                  * return a status indicating whether or not it was able
 434                  * to revoke the buffer.  On successful revoke, it is
 435                  * safe not to set the allocation bit in the committed
 436                  * bitmap, because we know that there is no outstanding
 437                  * activity on the buffer any more and so it is safe to
 438                  * reallocate it.
 439                  */
 440                 BUFFER_TRACE(bitmap_bh, "set in b_committed_data");
 441                 J_ASSERT_BH(bitmap_bh,
 442                                 bh2jh(bitmap_bh)->b_committed_data != NULL);
 443                 ext3_set_bit_atomic(sb_bgl_lock(sbi, block_group), bit + i,
 444                                 bh2jh(bitmap_bh)->b_committed_data);
 445
 446                 /*
 447                  * We clear the bit in the bitmap after setting the committed
 448                  * data bit, because this is the reverse order to that which
 449                  * the allocator uses.
 450                  */
 451                 BUFFER_TRACE(bitmap_bh, "clear bit");
 452                 if (!ext3_clear_bit_atomic(sb_bgl_lock(sbi, block_group),
 453                                                 bit + i, bitmap_bh->b_data)) {
 454                         jbd_unlock_bh_state(bitmap_bh);
 455                         ext3_error(sb, __FUNCTION__,
 456                                 "bit already cleared for block "E3FSBLK,
 457                                  block + i);
 458                         jbd_lock_bh_state(bitmap_bh);
 459                         BUFFER_TRACE(bitmap_bh, "bit already cleared");
 460                 } else {
 461                         group_freed++;
 462                 }
 463         }
 464         jbd_unlock_bh_state(bitmap_bh);
 465
 466         spin_lock(sb_bgl_lock(sbi, block_group));
 467         desc->bg_free_blocks_count =
 468                 cpu_to_le16(le16_to_cpu(desc->bg_free_blocks_count) +
 469                         group_freed);
 470         spin_unlock(sb_bgl_lock(sbi, block_group));
 471         percpu_counter_mod(&sbi->s_freeblocks_counter, count);
 472
 473         /* We dirtied the bitmap block */
 474         BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
 475         err = ext3_journal_dirty_metadata(handle, bitmap_bh);
 476
 477         /* And the group descriptor block */
 478         BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
 479         ret = ext3_journal_dirty_metadata(handle, gd_bh);
 480         if (!err) err = ret;
 481         *pdquot_freed_blocks += group_freed;
 482
 483         if (overflow && !err) {
 484                 block += count;
 485                 count = overflow;
 486                 goto do_more;
 487         }
 488         sb->s_dirt = 1;
 489 error_return:
 490         brelse(bitmap_bh);
 491         ext3_std_error(sb, err);
 492         return;
 493 }
 494
 495 /* Free given blocks, update quota and i_blocks field */
 496 void ext3_free_blocks(handle_t *handle, struct inode *inode,
 497                         ext3_fsblk_t block, unsigned long count)
 498 {
 499         struct super_block * sb;
 500         unsigned long dquot_freed_blocks;
 501
 502         sb = inode->i_sb;
 503         if (!sb) {
 504                 printk ("ext3_free_blocks: nonexistent device");
 505                 return;
 506         }
 507         ext3_free_blocks_sb(handle, sb, block, count, &dquot_freed_blocks);
 508         if (dquot_freed_blocks) {
 509                 DLIMIT_FREE_BLOCK(inode, dquot_freed_blocks);
 510                 DQUOT_FREE_BLOCK(inode, dquot_freed_blocks);
 511         }
 512         return;
 513 }
 514
 515 /*
 516  * For ext3 allocations, we must not reuse any blocks which are
 517  * allocated in the bitmap buffer's "last committed data" copy.  This
 518  * prevents deletes from freeing up the page for reuse until we have
 519  * committed the delete transaction.
 520  *
 521  * If we didn't do this, then deleting something and reallocating it as
 522  * data would allow the old block to be overwritten before the
 523  * transaction committed (because we force data to disk before commit).
 524  * This would lead to corruption if we crashed between overwriting the
 525  * data and committing the delete.
 526  *
 527  * @@@ We may want to make this allocation behaviour conditional on
 528  * data-writes at some point, and disable it for metadata allocations or
 529  * sync-data inodes.
 530  */
 531 static int ext3_test_allocatable(ext3_grpblk_t nr, struct buffer_head *bh)
 532 {
 533         int ret;
 534         struct journal_head *jh = bh2jh(bh);
 535
 536         if (ext3_test_bit(nr, bh->b_data))
 537                 return 0;
 538
 539         jbd_lock_bh_state(bh);
 540         if (!jh->b_committed_data)
 541                 ret = 1;
 542         else
 543                 ret = !ext3_test_bit(nr, jh->b_committed_data);
 544         jbd_unlock_bh_state(bh);
 545         return ret;
 546 }
 547
 548 static ext3_grpblk_t
 549 bitmap_search_next_usable_block(ext3_grpblk_t start, struct buffer_head *bh,
 550                                         ext3_grpblk_t maxblocks)
 551 {
 552         ext3_grpblk_t next;
 553         struct journal_head *jh = bh2jh(bh);
 554
 555         /*
 556          * The bitmap search --- search forward alternately through the actual
 557          * bitmap and the last-committed copy until we find a bit free in
 558          * both
 559          */
 560         while (start < maxblocks) {
 561                 next = ext3_find_next_zero_bit(bh->b_data, maxblocks, start);
 562                 if (next >= maxblocks)
 563                         return -1;
 564                 if (ext3_test_allocatable(next, bh))
 565                         return next;
 566                 jbd_lock_bh_state(bh);
 567                 if (jh->b_committed_data)
 568                         start = ext3_find_next_zero_bit(jh->b_committed_data,
 569                                                         maxblocks, next);
 570                 jbd_unlock_bh_state(bh);
 571         }
 572         return -1;
 573 }
 574
 575 /*
 576  * Find an allocatable block in a bitmap.  We honour both the bitmap and
 577  * its last-committed copy (if that exists), and perform the "most
 578  * appropriate allocation" algorithm of looking for a free block near
 579  * the initial goal; then for a free byte somewhere in the bitmap; then
 580  * for any free bit in the bitmap.
 581  */
 582 static ext3_grpblk_t
 583 find_next_usable_block(ext3_grpblk_t start, struct buffer_head *bh,
 584                         ext3_grpblk_t maxblocks)
 585 {
 586         ext3_grpblk_t here, next;
 587         char *p, *r;
 588
 589         if (start > 0) {
 590                 /*
 591                  * The goal was occupied; search forward for a free
 592                  * block within the next XX blocks.
 593                  *
 594                  * end_goal is more or less random, but it has to be
 595                  * less than EXT3_BLOCKS_PER_GROUP. Aligning up to the
 596                  * next 64-bit boundary is simple..
 597                  */
 598                 ext3_grpblk_t end_goal = (start + 63) & ~63;
 599                 if (end_goal > maxblocks)
 600                         end_goal = maxblocks;
 601                 here = ext3_find_next_zero_bit(bh->b_data, end_goal, start);
 602                 if (here < end_goal && ext3_test_allocatable(here, bh))
 603                         return here;
 604                 ext3_debug("Bit not found near goal\n");
 605         }
 606
 607         here = start;
 608         if (here < 0)
 609                 here = 0;
 610
 611         p = ((char *)bh->b_data) + (here >> 3);
 612         r = memscan(p, 0, (maxblocks - here + 7) >> 3);
 613         next = (r - ((char *)bh->b_data)) << 3;
 614
 615         if (next < maxblocks && next >= start && ext3_test_allocatable(next, bh))
 616                 return next;
 617
 618         /*
 619          * The bitmap search --- search forward alternately through the actual
 620          * bitmap and the last-committed copy until we find a bit free in
 621          * both
 622          */
 623         here = bitmap_search_next_usable_block(here, bh, maxblocks);
 624         return here;
 625 }
 626
 627 /*
 628  * We think we can allocate this block in this bitmap.  Try to set the bit.
 629  * If that succeeds then check that nobody has allocated and then freed the
 630  * block since we saw that is was not marked in b_committed_data.  If it _was_
 631  * allocated and freed then clear the bit in the bitmap again and return
 632  * zero (failure).
 633  */
 634 static inline int
 635 claim_block(spinlock_t *lock, ext3_grpblk_t block, struct buffer_head *bh)
 636 {
 637         struct journal_head *jh = bh2jh(bh);
 638         int ret;
 639
 640         if (ext3_set_bit_atomic(lock, block, bh->b_data))
 641                 return 0;
 642         jbd_lock_bh_state(bh);
 643         if (jh->b_committed_data && ext3_test_bit(block,jh->b_committed_data)) {
 644                 ext3_clear_bit_atomic(lock, block, bh->b_data);
 645                 ret = 0;
 646         } else {
 647                 ret = 1;
 648         }
 649         jbd_unlock_bh_state(bh);
 650         return ret;
 651 }
 652
 653 /*
 654  * If we failed to allocate the desired block then we may end up crossing to a
 655  * new bitmap.  In that case we must release write access to the old one via
 656  * ext3_journal_release_buffer(), else we'll run out of credits.
 657  */
 658 static ext3_grpblk_t
 659 ext3_try_to_allocate(struct super_block *sb, handle_t *handle, int group,
 660                         struct buffer_head *bitmap_bh, ext3_grpblk_t grp_goal,
 661                         unsigned long *count, struct ext3_reserve_window *my_rsv)
 662 {
 663         ext3_fsblk_t group_first_block;
 664         ext3_grpblk_t start, end;
 665         unsigned long num = 0;
 666
 667         /* we do allocation within the reservation window if we have a window */
 668         if (my_rsv) {
 669                 group_first_block = ext3_group_first_block_no(sb, group);
 670                 if (my_rsv->_rsv_start >= group_first_block)
 671                         start = my_rsv->_rsv_start - group_first_block;
 672                 else
 673                         /* reservation window cross group boundary */
 674                         start = 0;
 675                 end = my_rsv->_rsv_end - group_first_block + 1;
 676                 if (end > EXT3_BLOCKS_PER_GROUP(sb))
 677                         /* reservation window crosses group boundary */
 678                         end = EXT3_BLOCKS_PER_GROUP(sb);
 679                 if ((start <= grp_goal) && (grp_goal < end))
 680                         start = grp_goal;
 681                 else
 682                         grp_goal = -1;
 683         } else {
 684                 if (grp_goal > 0)
 685                         start = grp_goal;
 686                 else
 687                         start = 0;
 688                 end = EXT3_BLOCKS_PER_GROUP(sb);
 689         }
 690
 691         BUG_ON(start > EXT3_BLOCKS_PER_GROUP(sb));
 692
 693 repeat:
 694         if (grp_goal < 0 || !ext3_test_allocatable(grp_goal, bitmap_bh)) {
 695                 grp_goal = find_next_usable_block(start, bitmap_bh, end);
 696                 if (grp_goal < 0)
 697                         goto fail_access;
 698                 if (!my_rsv) {
 699                         int i;
 700
 701                         for (i = 0; i < 7 && grp_goal > start &&
 702                                         ext3_test_allocatable(grp_goal - 1,
 703                                                                 bitmap_bh);
 704                                         i++, grp_goal--)
 705                                 ;
 706                 }
 707         }
 708         start = grp_goal;
 709
 710         if (!claim_block(sb_bgl_lock(EXT3_SB(sb), group), grp_goal, bitmap_bh)) {
 711                 /*
 712                  * The block was allocated by another thread, or it was
 713                  * allocated and then freed by another thread
 714                  */
 715                 start++;
 716                 grp_goal++;
 717                 if (start >= end)
 718                         goto fail_access;
 719                 goto repeat;
 720         }
 721         num++;
 722         grp_goal++;
 723         while (num < *count && grp_goal < end
 724                 && ext3_test_allocatable(grp_goal, bitmap_bh)
 725                 && claim_block(sb_bgl_lock(EXT3_SB(sb), group), grp_goal, bitmap_bh)) {
 726                 num++;
 727                 grp_goal++;
 728         }
 729         *count = num;
 730         return grp_goal - num;
 731 fail_access:
 732         *count = num;
 733         return -1;
 734 }
 735
 736 /**
 737  *      find_next_reservable_window():
 738  *              find a reservable space within the given range.
 739  *              It does not allocate the reservation window for now:
 740  *              alloc_new_reservation() will do the work later.
 741  *
 742  *      @search_head: the head of the searching list;
 743  *              This is not necessarily the list head of the whole filesystem
 744  *
 745  *              We have both head and start_block to assist the search
 746  *              for the reservable space. The list starts from head,
 747  *              but we will shift to the place where start_block is,
 748  *              then start from there, when looking for a reservable space.
 749  *
 750  *      @size: the target new reservation window size
 751  *
 752  *      @group_first_block: the first block we consider to start
 753  *                      the real search from
 754  *
 755  *      @last_block:
 756  *              the maximum block number that our goal reservable space
 757  *              could start from. This is normally the last block in this
 758  *              group. The search will end when we found the start of next
 759  *              possible reservable space is out of this boundary.
 760  *              This could handle the cross boundary reservation window
 761  *              request.
 762  *
 763  *      basically we search from the given range, rather than the whole
 764  *      reservation double linked list, (start_block, last_block)
 765  *      to find a free region that is of my size and has not
 766  *      been reserved.
 767  *
 768  */
 769 static int find_next_reservable_window(
 770                                 struct ext3_reserve_window_node *search_head,
 771                                 struct ext3_reserve_window_node *my_rsv,
 772                                 struct super_block * sb,
 773                                 ext3_fsblk_t start_block,
 774                                 ext3_fsblk_t last_block)
 775 {
 776         struct rb_node *next;
 777         struct ext3_reserve_window_node *rsv, *prev;
 778         ext3_fsblk_t cur;
 779         int size = my_rsv->rsv_goal_size;
 780
 781         /* TODO: make the start of the reservation window byte-aligned */
 782         /* cur = *start_block & ~7;*/
 783         cur = start_block;
 784         rsv = search_head;
 785         if (!rsv)
 786                 return -1;
 787
 788         while (1) {
 789                 if (cur <= rsv->rsv_end)
 790                         cur = rsv->rsv_end + 1;
 791
 792                 /* TODO?
 793                  * in the case we could not find a reservable space
 794                  * that is what is expected, during the re-search, we could
 795                  * remember what's the largest reservable space we could have
 796                  * and return that one.
 797                  *
 798                  * For now it will fail if we could not find the reservable
 799                  * space with expected-size (or more)...
 800                  */
 801                 if (cur > last_block)
 802                         return -1;              /* fail */
 803
 804                 prev = rsv;
 805                 next = rb_next(&rsv->rsv_node);
 806                 rsv = list_entry(next,struct ext3_reserve_window_node,rsv_node);
 807
 808                 /*
 809                  * Reached the last reservation, we can just append to the
 810                  * previous one.
 811                  */
 812                 if (!next)
 813                         break;
 814
 815                 if (cur + size <= rsv->rsv_start) {
 816                         /*
 817                          * Found a reserveable space big enough.  We could
 818                          * have a reservation across the group boundary here
 819                          */
 820                         break;
 821                 }
 822         }
 823         /*
 824          * we come here either :
 825          * when we reach the end of the whole list,
 826          * and there is empty reservable space after last entry in the list.
 827          * append it to the end of the list.
 828          *
 829          * or we found one reservable space in the middle of the list,
 830          * return the reservation window that we could append to.
 831          * succeed.
 832          */
 833
 834         if ((prev != my_rsv) && (!rsv_is_empty(&my_rsv->rsv_window)))
 835                 rsv_window_remove(sb, my_rsv);
 836
 837         /*
 838          * Let's book the whole avaliable window for now.  We will check the
 839          * disk bitmap later and then, if there are free blocks then we adjust
 840          * the window size if it's larger than requested.
 841          * Otherwise, we will remove this node from the tree next time
 842          * call find_next_reservable_window.
 843          */
 844         my_rsv->rsv_start = cur;
 845         my_rsv->rsv_end = cur + size - 1;
 846         my_rsv->rsv_alloc_hit = 0;
 847
 848         if (prev != my_rsv)
 849                 ext3_rsv_window_add(sb, my_rsv);
 850
 851         return 0;
 852 }
 853
 854 /**
 855  *      alloc_new_reservation()--allocate a new reservation window
 856  *
 857  *              To make a new reservation, we search part of the filesystem
 858  *              reservation list (the list that inside the group). We try to
 859  *              allocate a new reservation window near the allocation goal,
 860  *              or the beginning of the group, if there is no goal.
 861  *
 862  *              We first find a reservable space after the goal, then from
 863  *              there, we check the bitmap for the first free block after
 864  *              it. If there is no free block until the end of group, then the
 865  *              whole group is full, we failed. Otherwise, check if the free
 866  *              block is inside the expected reservable space, if so, we
 867  *              succeed.
 868  *              If the first free block is outside the reservable space, then
 869  *              start from the first free block, we search for next available
 870  *              space, and go on.
 871  *
 872  *      on succeed, a new reservation will be found and inserted into the list
 873  *      It contains at least one free block, and it does not overlap with other
 874  *      reservation windows.
 875  *
 876  *      failed: we failed to find a reservation window in this group
 877  *
 878  *      @rsv: the reservation
 879  *
 880  *      @grp_goal: The goal (group-relative).  It is where the search for a
 881  *              free reservable space should start from.
 882  *              if we have a grp_goal(grp_goal >0 ), then start from there,
 883  *              no grp_goal(grp_goal = -1), we start from the first block
 884  *              of the group.
 885  *
 886  *      @sb: the super block
 887  *      @group: the group we are trying to allocate in
 888  *      @bitmap_bh: the block group block bitmap
 889  *
 890  */
 891 static int alloc_new_reservation(struct ext3_reserve_window_node *my_rsv,
 892                 ext3_grpblk_t grp_goal, struct super_block *sb,
 893                 unsigned int group, struct buffer_head *bitmap_bh)
 894 {
 895         struct ext3_reserve_window_node *search_head;
 896         ext3_fsblk_t group_first_block, group_end_block, start_block;
 897         ext3_grpblk_t first_free_block;
 898         struct rb_root *fs_rsv_root = &EXT3_SB(sb)->s_rsv_window_root;
 899         unsigned long size;
 900         int ret;
 901         spinlock_t *rsv_lock = &EXT3_SB(sb)->s_rsv_window_lock;
 902
 903         group_first_block = ext3_group_first_block_no(sb, group);
 904         group_end_block = group_first_block + (EXT3_BLOCKS_PER_GROUP(sb) - 1);
 905
 906         if (grp_goal < 0)
 907                 start_block = group_first_block;
 908         else
 909                 start_block = grp_goal + group_first_block;
 910
 911         size = my_rsv->rsv_goal_size;
 912
 913         if (!rsv_is_empty(&my_rsv->rsv_window)) {
 914                 /*
 915                  * if the old reservation is cross group boundary
 916                  * and if the goal is inside the old reservation window,
 917                  * we will come here when we just failed to allocate from
 918                  * the first part of the window. We still have another part
 919                  * that belongs to the next group. In this case, there is no
 920                  * point to discard our window and try to allocate a new one
 921                  * in this group(which will fail). we should
 922                  * keep the reservation window, just simply move on.
 923                  *
 924                  * Maybe we could shift the start block of the reservation
 925                  * window to the first block of next group.
 926                  */
 927
 928                 if ((my_rsv->rsv_start <= group_end_block) &&
 929                                 (my_rsv->rsv_end > group_end_block) &&
 930                                 (start_block >= my_rsv->rsv_start))
 931                         return -1;
 932
 933                 if ((my_rsv->rsv_alloc_hit >
 934                      (my_rsv->rsv_end - my_rsv->rsv_start + 1) / 2)) {
 935                         /*
 936                          * if we previously allocation hit ration is greater than half
 937                          * we double the size of reservation window next time
 938                          * otherwise keep the same
 939                          */
 940                         size = size * 2;
 941                         if (size > EXT3_MAX_RESERVE_BLOCKS)
 942                                 size = EXT3_MAX_RESERVE_BLOCKS;
 943                         my_rsv->rsv_goal_size= size;
 944                 }
 945         }
 946
 947         spin_lock(rsv_lock);
 948         /*
 949          * shift the search start to the window near the goal block
 950          */
 951         search_head = search_reserve_window(fs_rsv_root, start_block);
 952
 953         /*
 954          * find_next_reservable_window() simply finds a reservable window
 955          * inside the given range(start_block, group_end_block).
 956          *
 957          * To make sure the reservation window has a free bit inside it, we
 958          * need to check the bitmap after we found a reservable window.
 959          */
 960 retry:
 961         ret = find_next_reservable_window(search_head, my_rsv, sb,
 962                                                 start_block, group_end_block);
 963
 964         if (ret == -1) {
 965                 if (!rsv_is_empty(&my_rsv->rsv_window))
 966                         rsv_window_remove(sb, my_rsv);
 967                 spin_unlock(rsv_lock);
 968                 return -1;
 969         }
 970
 971         /*
 972          * On success, find_next_reservable_window() returns the
 973          * reservation window where there is a reservable space after it.
 974          * Before we reserve this reservable space, we need
 975          * to make sure there is at least a free block inside this region.
 976          *
 977          * searching the first free bit on the block bitmap and copy of
 978          * last committed bitmap alternatively, until we found a allocatable
 979          * block. Search start from the start block of the reservable space
 980          * we just found.
 981          */
 982         spin_unlock(rsv_lock);
 983         first_free_block = bitmap_search_next_usable_block(
 984                         my_rsv->rsv_start - group_first_block,
 985                         bitmap_bh, group_end_block - group_first_block + 1);
 986
 987         if (first_free_block < 0) {
 988                 /*
 989                  * no free block left on the bitmap, no point
 990                  * to reserve the space. return failed.
 991                  */
 992                 spin_lock(rsv_lock);
 993                 if (!rsv_is_empty(&my_rsv->rsv_window))
 994                         rsv_window_remove(sb, my_rsv);
 995                 spin_unlock(rsv_lock);
 996                 return -1;              /* failed */
 997         }
 998
 999         start_block = first_free_block + group_first_block;
1000         /*
1001          * check if the first free block is within the
1002          * free space we just reserved
1003          */
1004         if (start_block >= my_rsv->rsv_start && start_block < my_rsv->rsv_end)
1005                 return 0;               /* success */
1006         /*
1007          * if the first free bit we found is out of the reservable space
1008          * continue search for next reservable space,
1009          * start from where the free block is,
1010          * we also shift the list head to where we stopped last time
1011          */
1012         search_head = my_rsv;
1013         spin_lock(rsv_lock);
1014         goto retry;
1015 }
1016
1017 static void try_to_extend_reservation(struct ext3_reserve_window_node *my_rsv,
1018                         struct super_block *sb, int size)
1019 {
1020         struct ext3_reserve_window_node *next_rsv;
1021         struct rb_node *next;
1022         spinlock_t *rsv_lock = &EXT3_SB(sb)->s_rsv_window_lock;
1023
1024         if (!spin_trylock(rsv_lock))
1025                 return;
1026
1027         next = rb_next(&my_rsv->rsv_node);
1028
1029         if (!next)
1030                 my_rsv->rsv_end += size;
1031         else {
1032                 next_rsv = list_entry(next, struct ext3_reserve_window_node, rsv_node);
1033
1034                 if ((next_rsv->rsv_start - my_rsv->rsv_end - 1) >= size)
1035                         my_rsv->rsv_end += size;
1036                 else
1037                         my_rsv->rsv_end = next_rsv->rsv_start - 1;
1038         }
1039         spin_unlock(rsv_lock);
1040 }
1041
1042 /*
1043  * This is the main function used to allocate a new block and its reservation
1044  * window.
1045  *
1046  * Each time when a new block allocation is need, first try to allocate from
1047  * its own reservation.  If it does not have a reservation window, instead of
1048  * looking for a free bit on bitmap first, then look up the reservation list to
1049  * see if it is inside somebody else's reservation window, we try to allocate a
1050  * reservation window for it starting from the goal first. Then do the block
1051  * allocation within the reservation window.
1052  *
1053  * This will avoid keeping on searching the reservation list again and
1054  * again when somebody is looking for a free block (without
1055  * reservation), and there are lots of free blocks, but they are all
1056  * being reserved.
1057  *
1058  * We use a sorted double linked list for the per-filesystem reservation list.
1059  * The insert, remove and find a free space(non-reserved) operations for the
1060  * sorted double linked list should be fast.
1061  *
1062  */
1063 static ext3_grpblk_t
1064 ext3_try_to_allocate_with_rsv(struct super_block *sb, handle_t *handle,
1065                         unsigned int group, struct buffer_head *bitmap_bh,
1066                         ext3_grpblk_t grp_goal,
1067                         struct ext3_reserve_window_node * my_rsv,
1068                         unsigned long *count, int *errp)
1069 {
1070         ext3_fsblk_t group_first_block, group_last_block;
1071         ext3_grpblk_t ret = 0;
1072         int fatal;
1073         unsigned long num = *count;
1074
1075         *errp = 0;
1076
1077         /*
1078          * Make sure we use undo access for the bitmap, because it is critical
1079          * that we do the frozen_data COW on bitmap buffers in all cases even
1080          * if the buffer is in BJ_Forget state in the committing transaction.
1081          */
1082         BUFFER_TRACE(bitmap_bh, "get undo access for new block");
1083         fatal = ext3_journal_get_undo_access(handle, bitmap_bh);
1084         if (fatal) {
1085                 *errp = fatal;
1086                 return -1;
1087         }
1088
1089         /*
1090          * we don't deal with reservation when
1091          * filesystem is mounted without reservation
1092          * or the file is not a regular file
1093          * or last attempt to allocate a block with reservation turned on failed
1094          */
1095         if (my_rsv == NULL ) {
1096                 ret = ext3_try_to_allocate(sb, handle, group, bitmap_bh,
1097                                                 grp_goal, count, NULL);
1098                 goto out;
1099         }
1100         /*
1101          * grp_goal is a group relative block number (if there is a goal)
1102          * 0 < grp_goal < EXT3_BLOCKS_PER_GROUP(sb)
1103          * first block is a filesystem wide block number
1104          * first block is the block number of the first block in this group
1105          */
1106         group_first_block = ext3_group_first_block_no(sb, group);
1107         group_last_block = group_first_block + (EXT3_BLOCKS_PER_GROUP(sb) - 1);
1108
1109         /*
1110          * Basically we will allocate a new block from inode's reservation
1111          * window.
1112          *
1113          * We need to allocate a new reservation window, if:
1114          * a) inode does not have a reservation window; or
1115          * b) last attempt to allocate a block from existing reservation
1116          *    failed; or
1117          * c) we come here with a goal and with a reservation window
1118          *
1119          * We do not need to allocate a new reservation window if we come here
1120          * at the beginning with a goal and the goal is inside the window, or
1121          * we don't have a goal but already have a reservation window.
1122          * then we could go to allocate from the reservation window directly.
1123          */
1124         while (1) {
1125                 if (rsv_is_empty(&my_rsv->rsv_window) || (ret < 0) ||
1126                         !goal_in_my_reservation(&my_rsv->rsv_window, grp_goal, group, sb)) {
1127                         if (my_rsv->rsv_goal_size < *count)
1128                                 my_rsv->rsv_goal_size = *count;
1129                         ret = alloc_new_reservation(my_rsv, grp_goal, sb,
1130                                                         group, bitmap_bh);
1131                         if (ret < 0)
1132                                 break;                  /* failed */
1133
1134                         if (!goal_in_my_reservation(&my_rsv->rsv_window, grp_goal, group, sb))
1135                                 grp_goal = -1;
1136                 } else if (grp_goal > 0 && (my_rsv->rsv_end-grp_goal+1) < *count)
1137                         try_to_extend_reservation(my_rsv, sb,
1138                                         *count-my_rsv->rsv_end + grp_goal - 1);
1139
1140                 if ((my_rsv->rsv_start > group_last_block) ||
1141                                 (my_rsv->rsv_end < group_first_block))
1142                         BUG();
1143                 ret = ext3_try_to_allocate(sb, handle, group, bitmap_bh, grp_goal,
1144                                            &num, &my_rsv->rsv_window);
1145                 if (ret >= 0) {
1146                         my_rsv->rsv_alloc_hit += num;
1147                         *count = num;
1148                         break;                          /* succeed */
1149                 }
1150                 num = *count;
1151         }
1152 out:
1153         if (ret >= 0) {
1154                 BUFFER_TRACE(bitmap_bh, "journal_dirty_metadata for "
1155                                         "bitmap block");
1156                 fatal = ext3_journal_dirty_metadata(handle, bitmap_bh);
1157                 if (fatal) {
1158                         *errp = fatal;
1159                         return -1;
1160                 }
1161                 return ret;
1162         }
1163
1164         BUFFER_TRACE(bitmap_bh, "journal_release_buffer");
1165         ext3_journal_release_buffer(handle, bitmap_bh);
1166         return ret;
1167 }
1168
1169 static int ext3_has_free_blocks(struct super_block *sb)
1170 {
1171         struct ext3_sb_info *sbi = EXT3_SB(sb);
1172         ext3_fsblk_t free_blocks, root_blocks;
1173         int cond;
1174
1175         free_blocks = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
1176         root_blocks = le32_to_cpu(sbi->s_es->s_r_blocks_count);
1177
1178         vxdprintk(VXD_CBIT(dlim, 3),
1179                 "ext3_has_free_blocks(%p): free=%lu, root=%lu",
1180                 sb, free_blocks, root_blocks);
1181
1182         DLIMIT_ADJUST_BLOCK(sb, vx_current_xid(), &free_blocks, &root_blocks);
1183
1184         cond = (free_blocks < root_blocks + 1 &&
1185                 !capable(CAP_SYS_RESOURCE) &&
1186                 sbi->s_resuid != current->fsuid &&
1187                 (sbi->s_resgid == 0 || !in_group_p (sbi->s_resgid)));
1188
1189         vxdprintk(VXD_CBIT(dlim, 3),
1190                 "ext3_has_free_blocks(%p): %lu<%lu+1, %c, %u!=%u r=%d",
1191                 sb, free_blocks, root_blocks,
1192                 !capable(CAP_SYS_RESOURCE)?'1':'0',
1193                 sbi->s_resuid, current->fsuid, cond?0:1);
1194
1195         return (cond ? 0 : 1);
1196 }
1197
1198 /*
1199  * ext3_should_retry_alloc() is called when ENOSPC is returned, and if
1200  * it is profitable to retry the operation, this function will wait
1201  * for the current or commiting transaction to complete, and then
1202  * return TRUE.
1203  */
1204 int ext3_should_retry_alloc(struct super_block *sb, int *retries)
1205 {
1206         if (!ext3_has_free_blocks(sb) || (*retries)++ > 3)
1207                 return 0;
1208
1209         jbd_debug(1, "%s: retrying operation after ENOSPC\n", sb->s_id);
1210
1211         return journal_force_commit_nested(EXT3_SB(sb)->s_journal);
1212 }
1213
1214 /*
1215  * ext3_new_block uses a goal block to assist allocation.  If the goal is
1216  * free, or there is a free block within 32 blocks of the goal, that block
1217  * is allocated.  Otherwise a forward search is made for a free block; within
1218  * each block group the search first looks for an entire free byte in the block
1219  * bitmap, and then for any free bit if that fails.
1220  * This function also updates quota and i_blocks field.
1221  */
1222 ext3_fsblk_t ext3_new_blocks(handle_t *handle, struct inode *inode,
1223                         ext3_fsblk_t goal, unsigned long *count, int *errp)
1224 {
1225         struct buffer_head *bitmap_bh = NULL;
1226         struct buffer_head *gdp_bh;
1227         int group_no;
1228         int goal_group;
1229         ext3_grpblk_t grp_target_blk;   /* blockgroup relative goal block */
1230         ext3_grpblk_t grp_alloc_blk;    /* blockgroup-relative allocated block*/
1231         ext3_fsblk_t ret_block;         /* filesyetem-wide allocated block */
1232         int bgi;                        /* blockgroup iteration index */
1233         int fatal = 0, err;
1234         int performed_allocation = 0;
1235         ext3_grpblk_t free_blocks;      /* number of free blocks in a group */
1236         struct super_block *sb;
1237         struct ext3_group_desc *gdp;
1238         struct ext3_super_block *es;
1239         struct ext3_sb_info *sbi;
1240         struct ext3_reserve_window_node *my_rsv = NULL;
1241         struct ext3_block_alloc_info *block_i;
1242         unsigned short windowsz = 0;
1243 #ifdef EXT3FS_DEBUG
1244         static int goal_hits, goal_attempts;
1245 #endif
1246         unsigned long ngroups;
1247         unsigned long num = *count;
1248
1249         *errp = -ENOSPC;
1250         sb = inode->i_sb;
1251         if (!sb) {
1252                 printk("ext3_new_block: nonexistent device");
1253                 return 0;
1254         }
1255
1256         /*
1257          * Check quota for allocation of this block.
1258          */
1259         if (DQUOT_ALLOC_BLOCK(inode, num)) {
1260                 *errp = -EDQUOT;
1261                 return 0;
1262         }
1263         if (DLIMIT_ALLOC_BLOCK(inode, 1))
1264             goto out_dlimit;
1265
1266         sbi = EXT3_SB(sb);
1267         es = EXT3_SB(sb)->s_es;
1268         ext3_debug("goal=%lu.\n", goal);
1269         /*
1270          * Allocate a block from reservation only when
1271          * filesystem is mounted with reservation(default,-o reservation), and
1272          * it's a regular file, and
1273          * the desired window size is greater than 0 (One could use ioctl
1274          * command EXT3_IOC_SETRSVSZ to set the window size to 0 to turn off
1275          * reservation on that particular file)
1276          */
1277         block_i = EXT3_I(inode)->i_block_alloc_info;
1278         if (block_i && ((windowsz = block_i->rsv_window_node.rsv_goal_size) > 0))
1279                 my_rsv = &block_i->rsv_window_node;
1280
1281         if (!ext3_has_free_blocks(sb)) {
1282                 *errp = -ENOSPC;
1283                 goto out;
1284         }
1285
1286         /*
1287          * First, test whether the goal block is free.
1288          */
1289         if (goal < le32_to_cpu(es->s_first_data_block) ||
1290             goal >= le32_to_cpu(es->s_blocks_count))
1291                 goal = le32_to_cpu(es->s_first_data_block);
1292         group_no = (goal - le32_to_cpu(es->s_first_data_block)) /
1293                         EXT3_BLOCKS_PER_GROUP(sb);
1294         goal_group = group_no;
1295 retry_alloc:
1296         gdp = ext3_get_group_desc(sb, group_no, &gdp_bh);
1297         if (!gdp)
1298                 goto io_error;
1299
1300         free_blocks = le16_to_cpu(gdp->bg_free_blocks_count);
1301         /*
1302          * if there is not enough free blocks to make a new resevation
1303          * turn off reservation for this allocation
1304          */
1305         if (my_rsv && (free_blocks < windowsz)
1306                 && (rsv_is_empty(&my_rsv->rsv_window)))
1307                 my_rsv = NULL;
1308
1309         if (free_blocks > 0) {
1310                 grp_target_blk = ((goal - le32_to_cpu(es->s_first_data_block)) %
1311                                 EXT3_BLOCKS_PER_GROUP(sb));
1312                 bitmap_bh = read_block_bitmap(sb, group_no);
1313                 if (!bitmap_bh)
1314                         goto io_error;
1315                 grp_alloc_blk = ext3_try_to_allocate_with_rsv(sb, handle,
1316                                         group_no, bitmap_bh, grp_target_blk,
1317                                         my_rsv, &num, &fatal);
1318                 if (fatal)
1319                         goto out;
1320                 if (grp_alloc_blk >= 0)
1321                         goto allocated;
1322         }
1323
1324         ngroups = EXT3_SB(sb)->s_groups_count;
1325         smp_rmb();
1326
1327         /*
1328          * Now search the rest of the groups.  We assume that
1329          * i and gdp correctly point to the last group visited.
1330          */
1331         for (bgi = 0; bgi < ngroups; bgi++) {
1332                 group_no++;
1333                 if (group_no >= ngroups)
1334                         group_no = 0;
1335                 gdp = ext3_get_group_desc(sb, group_no, &gdp_bh);
1336                 if (!gdp) {
1337                         *errp = -EIO;
1338                         goto out;
1339                 }
1340                 free_blocks = le16_to_cpu(gdp->bg_free_blocks_count);
1341                 /*
1342                  * skip this group if the number of
1343                  * free blocks is less than half of the reservation
1344                  * window size.
1345                  */
1346                 if (free_blocks <= (windowsz/2))
1347                         continue;
1348
1349                 brelse(bitmap_bh);
1350                 bitmap_bh = read_block_bitmap(sb, group_no);
1351                 if (!bitmap_bh)
1352                         goto io_error;
1353                 /*
1354                  * try to allocate block(s) from this group, without a goal(-1).
1355                  */
1356                 grp_alloc_blk = ext3_try_to_allocate_with_rsv(sb, handle,
1357                                         group_no, bitmap_bh, -1, my_rsv,
1358                                         &num, &fatal);
1359                 if (fatal)
1360                         goto out;
1361                 if (grp_alloc_blk >= 0)
1362                         goto allocated;
1363         }
1364         /*
1365          * We may end up a bogus ealier ENOSPC error due to
1366          * filesystem is "full" of reservations, but
1367          * there maybe indeed free blocks avaliable on disk
1368          * In this case, we just forget about the reservations
1369          * just do block allocation as without reservations.
1370          */
1371         if (my_rsv) {
1372                 my_rsv = NULL;
1373                 group_no = goal_group;
1374                 goto retry_alloc;
1375         }
1376         /* No space left on the device */
1377         *errp = -ENOSPC;
1378         goto out;
1379
1380 allocated:
1381
1382         ext3_debug("using block group %d(%d)\n",
1383                         group_no, gdp->bg_free_blocks_count);
1384
1385         BUFFER_TRACE(gdp_bh, "get_write_access");
1386         fatal = ext3_journal_get_write_access(handle, gdp_bh);
1387         if (fatal)
1388                 goto out;
1389
1390         ret_block = grp_alloc_blk + ext3_group_first_block_no(sb, group_no);
1391
1392         if (in_range(le32_to_cpu(gdp->bg_block_bitmap), ret_block, num) ||
1393             in_range(le32_to_cpu(gdp->bg_inode_bitmap), ret_block, num) ||
1394             in_range(ret_block, le32_to_cpu(gdp->bg_inode_table),
1395                       EXT3_SB(sb)->s_itb_per_group) ||
1396             in_range(ret_block + num - 1, le32_to_cpu(gdp->bg_inode_table),
1397                       EXT3_SB(sb)->s_itb_per_group))
1398                 ext3_error(sb, "ext3_new_block",
1399                             "Allocating block in system zone - "
1400                             "blocks from "E3FSBLK", length %lu",
1401                              ret_block, num);
1402
1403         performed_allocation = 1;
1404
1405 #ifdef CONFIG_JBD_DEBUG
1406         {
1407                 struct buffer_head *debug_bh;
1408
1409                 /* Record bitmap buffer state in the newly allocated block */
1410                 debug_bh = sb_find_get_block(sb, ret_block);
1411                 if (debug_bh) {
1412                         BUFFER_TRACE(debug_bh, "state when allocated");
1413                         BUFFER_TRACE2(debug_bh, bitmap_bh, "bitmap state");
1414                         brelse(debug_bh);
1415                 }
1416         }
1417         jbd_lock_bh_state(bitmap_bh);
1418         spin_lock(sb_bgl_lock(sbi, group_no));
1419         if (buffer_jbd(bitmap_bh) && bh2jh(bitmap_bh)->b_committed_data) {
1420                 int i;
1421
1422                 for (i = 0; i < num; i++) {
1423                         if (ext3_test_bit(grp_alloc_blk+i,
1424                                         bh2jh(bitmap_bh)->b_committed_data)) {
1425                                 printk("%s: block was unexpectedly set in "
1426                                         "b_committed_data\n", __FUNCTION__);
1427                         }
1428                 }
1429         }
1430         ext3_debug("found bit %d\n", grp_alloc_blk);
1431         spin_unlock(sb_bgl_lock(sbi, group_no));
1432         jbd_unlock_bh_state(bitmap_bh);
1433 #endif
1434
1435         if (ret_block + num - 1 >= le32_to_cpu(es->s_blocks_count)) {
1436                 ext3_error(sb, "ext3_new_block",
1437                             "block("E3FSBLK") >= blocks count(%d) - "
1438                             "block_group = %d, es == %p ", ret_block,
1439                         le32_to_cpu(es->s_blocks_count), group_no, es);
1440                 goto out;
1441         }
1442
1443         /*
1444          * It is up to the caller to add the new buffer to a journal
1445          * list of some description.  We don't know in advance whether
1446          * the caller wants to use it as metadata or data.
1447          */
1448         ext3_debug("allocating block %lu. Goal hits %d of %d.\n",
1449                         ret_block, goal_hits, goal_attempts);
1450
1451         spin_lock(sb_bgl_lock(sbi, group_no));
1452         gdp->bg_free_blocks_count =
1453                         cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count) - num);
1454         spin_unlock(sb_bgl_lock(sbi, group_no));
1455         percpu_counter_mod(&sbi->s_freeblocks_counter, -num);
1456
1457         BUFFER_TRACE(gdp_bh, "journal_dirty_metadata for group descriptor");
1458         err = ext3_journal_dirty_metadata(handle, gdp_bh);
1459         if (!fatal)
1460                 fatal = err;
1461
1462         sb->s_dirt = 1;
1463         if (fatal)
1464                 goto out;
1465
1466         *errp = 0;
1467         brelse(bitmap_bh);
1468         DQUOT_FREE_BLOCK(inode, *count-num);
1469         *count = num;
1470         return ret_block;
1471
1472 io_error:
1473         *errp = -EIO;
1474 out:
1475         if (!performed_allocation)
1476                 DLIMIT_FREE_BLOCK(inode, 1);
1477 out_dlimit:
1478         if (fatal) {
1479                 *errp = fatal;
1480                 ext3_std_error(sb, fatal);
1481         }
1482         /*
1483          * Undo the block allocation
1484          */
1485         if (!performed_allocation)
1486                 DQUOT_FREE_BLOCK(inode, *count);
1487         brelse(bitmap_bh);
1488         return 0;
1489 }
1490
1491 ext3_fsblk_t ext3_new_block(handle_t *handle, struct inode *inode,
1492                         ext3_fsblk_t goal, int *errp)
1493 {
1494         unsigned long count = 1;
1495
1496         return ext3_new_blocks(handle, inode, goal, &count, errp);
1497 }
1498
1499 ext3_fsblk_t ext3_count_free_blocks(struct super_block *sb)
1500 {
1501         ext3_fsblk_t desc_count;
1502         struct ext3_group_desc *gdp;
1503         int i;
1504         unsigned long ngroups = EXT3_SB(sb)->s_groups_count;
1505 #ifdef EXT3FS_DEBUG
1506         struct ext3_super_block *es;
1507         ext3_fsblk_t bitmap_count;
1508         unsigned long x;
1509         struct buffer_head *bitmap_bh = NULL;
1510
1511         es = EXT3_SB(sb)->s_es;
1512         desc_count = 0;
1513         bitmap_count = 0;
1514         gdp = NULL;
1515
1516         smp_rmb();
1517         for (i = 0; i < ngroups; i++) {
1518                 gdp = ext3_get_group_desc(sb, i, NULL);
1519                 if (!gdp)
1520                         continue;
1521                 desc_count += le16_to_cpu(gdp->bg_free_blocks_count);
1522                 brelse(bitmap_bh);
1523                 bitmap_bh = read_block_bitmap(sb, i);
1524                 if (bitmap_bh == NULL)
1525                         continue;
1526
1527                 x = ext3_count_free(bitmap_bh, sb->s_blocksize);
1528                 printk("group %d: stored = %d, counted = %lu\n",
1529                         i, le16_to_cpu(gdp->bg_free_blocks_count), x);
1530                 bitmap_count += x;
1531         }
1532         brelse(bitmap_bh);
1533         printk("ext3_count_free_blocks: stored = "E3FSBLK
1534                 ", computed = "E3FSBLK", "E3FSBLK"\n",
1535                le32_to_cpu(es->s_free_blocks_count),
1536                 desc_count, bitmap_count);
1537         return bitmap_count;
1538 #else
1539         desc_count = 0;
1540         smp_rmb();
1541         for (i = 0; i < ngroups; i++) {
1542                 gdp = ext3_get_group_desc(sb, i, NULL);
1543                 if (!gdp)
1544                         continue;
1545                 desc_count += le16_to_cpu(gdp->bg_free_blocks_count);
1546         }
1547
1548         return desc_count;
1549 #endif
1550 }
1551
1552 static inline int
1553 block_in_use(ext3_fsblk_t block, struct super_block *sb, unsigned char *map)
1554 {
1555         return ext3_test_bit ((block -
1556                 le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block)) %
1557                          EXT3_BLOCKS_PER_GROUP(sb), map);
1558 }
1559
1560 static inline int test_root(int a, int b)
1561 {
1562         int num = b;
1563
1564         while (a > num)
1565                 num *= b;
1566         return num == a;
1567 }
1568
1569 static int ext3_group_sparse(int group)
1570 {
1571         if (group <= 1)
1572                 return 1;
1573         if (!(group & 1))
1574                 return 0;
1575         return (test_root(group, 7) || test_root(group, 5) ||
1576                 test_root(group, 3));
1577 }
1578
1579 /**
1580  *      ext3_bg_has_super - number of blocks used by the superblock in group
1581  *      @sb: superblock for filesystem
1582  *      @group: group number to check
1583  *
1584  *      Return the number of blocks used by the superblock (primary or backup)
1585  *      in this group.  Currently this will be only 0 or 1.
1586  */
1587 int ext3_bg_has_super(struct super_block *sb, int group)
1588 {
1589         if (EXT3_HAS_RO_COMPAT_FEATURE(sb,
1590                                 EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER) &&
1591                         !ext3_group_sparse(group))
1592                 return 0;
1593         return 1;
1594 }
1595
1596 static unsigned long ext3_bg_num_gdb_meta(struct super_block *sb, int group)
1597 {
1598         unsigned long metagroup = group / EXT3_DESC_PER_BLOCK(sb);
1599         unsigned long first = metagroup * EXT3_DESC_PER_BLOCK(sb);
1600         unsigned long last = first + EXT3_DESC_PER_BLOCK(sb) - 1;
1601
1602         if (group == first || group == first + 1 || group == last)
1603                 return 1;
1604         return 0;
1605 }
1606
1607 static unsigned long ext3_bg_num_gdb_nometa(struct super_block *sb, int group)
1608 {
1609         if (EXT3_HAS_RO_COMPAT_FEATURE(sb,
1610                                 EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER) &&
1611                         !ext3_group_sparse(group))
1612                 return 0;
1613         return EXT3_SB(sb)->s_gdb_count;
1614 }
1615
1616 /**
1617  *      ext3_bg_num_gdb - number of blocks used by the group table in group
1618  *      @sb: superblock for filesystem
1619  *      @group: group number to check
1620  *
1621  *      Return the number of blocks used by the group descriptor table
1622  *      (primary or backup) in this group.  In the future there may be a
1623  *      different number of descriptor blocks in each group.
1624  */
1625 unsigned long ext3_bg_num_gdb(struct super_block *sb, int group)
1626 {
1627         unsigned long first_meta_bg =
1628                         le32_to_cpu(EXT3_SB(sb)->s_es->s_first_meta_bg);
1629         unsigned long metagroup = group / EXT3_DESC_PER_BLOCK(sb);
1630
1631         if (!EXT3_HAS_INCOMPAT_FEATURE(sb,EXT3_FEATURE_INCOMPAT_META_BG) ||
1632                         metagroup < first_meta_bg)
1633                 return ext3_bg_num_gdb_nometa(sb,group);
1634
1635         return ext3_bg_num_gdb_meta(sb,group);
1636
1637 }