fs/ext3/balloc.c

   1 /*
   2  *  linux/fs/ext3/balloc.c
   3  *
   4  * Copyright (C) 1992, 1993, 1994, 1995
   5  * Remy Card (card@masi.ibp.fr)
   6  * Laboratoire MASI - Institut Blaise Pascal
   7  * Universite Pierre et Marie Curie (Paris VI)
   8  *
   9  *  Enhanced block allocation by Stephen Tweedie (sct@redhat.com), 1993
  10  *  Big-endian to little-endian byte-swapping/bitmaps by
  11  *        David S. Miller (davem@caip.rutgers.edu), 1995
  12  */
  13
  14 #include <linux/config.h>
  15 #include <linux/time.h>
  16 #include <linux/capability.h>
  17 #include <linux/fs.h>
  18 #include <linux/jbd.h>
  19 #include <linux/ext3_fs.h>
  20 #include <linux/ext3_jbd.h>
  21 #include <linux/quotaops.h>
  22 #include <linux/buffer_head.h>
  23
  24 /*
  25  * balloc.c contains the blocks allocation and deallocation routines
  26  */
  27
  28 /*
  29  * The free blocks are managed by bitmaps.  A file system contains several
  30  * blocks groups.  Each group contains 1 bitmap block for blocks, 1 bitmap
  31  * block for inodes, N blocks for the inode table and data blocks.
  32  *
  33  * The file system contains group descriptors which are located after the
  34  * super block.  Each descriptor contains the number of the bitmap block and
  35  * the free blocks count in the block.  The descriptors are loaded in memory
  36  * when a file system is mounted (see ext3_read_super).
  37  */
  38
  39
  40 #define in_range(b, first, len) ((b) >= (first) && (b) <= (first) + (len) - 1)
  41
  42 struct ext3_group_desc * ext3_get_group_desc(struct super_block * sb,
  43                                              unsigned int block_group,
  44                                              struct buffer_head ** bh)
  45 {
  46         unsigned long group_desc;
  47         unsigned long offset;
  48         struct ext3_group_desc * desc;
  49         struct ext3_sb_info *sbi = EXT3_SB(sb);
  50
  51         if (block_group >= sbi->s_groups_count) {
  52                 ext3_error (sb, "ext3_get_group_desc",
  53                             "block_group >= groups_count - "
  54                             "block_group = %d, groups_count = %lu",
  55                             block_group, sbi->s_groups_count);
  56
  57                 return NULL;
  58         }
  59         smp_rmb();
  60
  61         group_desc = block_group >> EXT3_DESC_PER_BLOCK_BITS(sb);
  62         offset = block_group & (EXT3_DESC_PER_BLOCK(sb) - 1);
  63         if (!sbi->s_group_desc[group_desc]) {
  64                 ext3_error (sb, "ext3_get_group_desc",
  65                             "Group descriptor not loaded - "
  66                             "block_group = %d, group_desc = %lu, desc = %lu",
  67                              block_group, group_desc, offset);
  68                 return NULL;
  69         }
  70
  71         desc = (struct ext3_group_desc *) sbi->s_group_desc[group_desc]->b_data;
  72         if (bh)
  73                 *bh = sbi->s_group_desc[group_desc];
  74         return desc + offset;
  75 }
  76
  77 /*
  78  * Read the bitmap for a given block_group, reading into the specified
  79  * slot in the superblock's bitmap cache.
  80  *
  81  * Return buffer_head on success or NULL in case of failure.
  82  */
  83 static struct buffer_head *
  84 read_block_bitmap(struct super_block *sb, unsigned int block_group)
  85 {
  86         struct ext3_group_desc * desc;
  87         struct buffer_head * bh = NULL;
  88
  89         desc = ext3_get_group_desc (sb, block_group, NULL);
  90         if (!desc)
  91                 goto error_out;
  92         bh = sb_bread(sb, le32_to_cpu(desc->bg_block_bitmap));
  93         if (!bh)
  94                 ext3_error (sb, "read_block_bitmap",
  95                             "Cannot read block bitmap - "
  96                             "block_group = %d, block_bitmap = %u",
  97                             block_group, le32_to_cpu(desc->bg_block_bitmap));
  98 error_out:
  99         return bh;
 100 }
 101 /*
 102  * The reservation window structure operations
 103  * --------------------------------------------
 104  * Operations include:
 105  * dump, find, add, remove, is_empty, find_next_reservable_window, etc.
 106  *
 107  * We use sorted double linked list for the per-filesystem reservation
 108  * window list. (like in vm_region).
 109  *
 110  * Initially, we keep those small operations in the abstract functions,
 111  * so later if we need a better searching tree than double linked-list,
 112  * we could easily switch to that without changing too much
 113  * code.
 114  */
 115 #if 0
 116 static void __rsv_window_dump(struct rb_root *root, int verbose,
 117                               const char *fn)
 118 {
 119         struct rb_node *n;
 120         struct ext3_reserve_window_node *rsv, *prev;
 121         int bad;
 122
 123 restart:
 124         n = rb_first(root);
 125         bad = 0;
 126         prev = NULL;
 127
 128         printk("Block Allocation Reservation Windows Map (%s):\n", fn);
 129         while (n) {
 130                 rsv = list_entry(n, struct ext3_reserve_window_node, rsv_node);
 131                 if (verbose)
 132                         printk("reservation window 0x%p "
 133                                "start:  %d, end:  %d\n",
 134                                rsv, rsv->rsv_start, rsv->rsv_end);
 135                 if (rsv->rsv_start && rsv->rsv_start >= rsv->rsv_end) {
 136                         printk("Bad reservation %p (start >= end)\n",
 137                                rsv);
 138                         bad = 1;
 139                 }
 140                 if (prev && prev->rsv_end >= rsv->rsv_start) {
 141                         printk("Bad reservation %p (prev->end >= start)\n",
 142                                rsv);
 143                         bad = 1;
 144                 }
 145                 if (bad) {
 146                         if (!verbose) {
 147                                 printk("Restarting reservation walk in verbose mode\n");
 148                                 verbose = 1;
 149                                 goto restart;
 150                         }
 151                 }
 152                 n = rb_next(n);
 153                 prev = rsv;
 154         }
 155         printk("Window map complete.\n");
 156         if (bad)
 157                 BUG();
 158 }
 159 #define rsv_window_dump(root, verbose) \
 160         __rsv_window_dump((root), (verbose), __FUNCTION__)
 161 #else
 162 #define rsv_window_dump(root, verbose) do {} while (0)
 163 #endif
 164
 165 static int
 166 goal_in_my_reservation(struct ext3_reserve_window *rsv, int goal,
 167                         unsigned int group, struct super_block * sb)
 168 {
 169         unsigned long group_first_block, group_last_block;
 170
 171         group_first_block = le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block) +
 172                                 group * EXT3_BLOCKS_PER_GROUP(sb);
 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 ((goal >= 0) && ((goal + group_first_block < rsv->_rsv_start)
 179                 || (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, unsigned long 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         unsigned int 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                          unsigned long block, unsigned long count,
 314                          int *pdquot_freed_blocks)
 315 {
 316         struct buffer_head *bitmap_bh = NULL;
 317         struct buffer_head *gd_bh;
 318         unsigned long block_group;
 319         unsigned long 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         unsigned 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 = %lu, 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 = %lu, 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 %lu", block + i);
 457                         jbd_lock_bh_state(bitmap_bh);
 458                         BUFFER_TRACE(bitmap_bh, "bit already cleared");
 459                 } else {
 460                         group_freed++;
 461                 }
 462         }
 463         jbd_unlock_bh_state(bitmap_bh);
 464
 465         spin_lock(sb_bgl_lock(sbi, block_group));
 466         desc->bg_free_blocks_count =
 467                 cpu_to_le16(le16_to_cpu(desc->bg_free_blocks_count) +
 468                         group_freed);
 469         spin_unlock(sb_bgl_lock(sbi, block_group));
 470         percpu_counter_mod(&sbi->s_freeblocks_counter, count);
 471
 472         /* We dirtied the bitmap block */
 473         BUFFER_TRACE(bitmap_bh, "dirtied bitmap block");
 474         err = ext3_journal_dirty_metadata(handle, bitmap_bh);
 475
 476         /* And the group descriptor block */
 477         BUFFER_TRACE(gd_bh, "dirtied group descriptor block");
 478         ret = ext3_journal_dirty_metadata(handle, gd_bh);
 479         if (!err) err = ret;
 480         *pdquot_freed_blocks += group_freed;
 481
 482         if (overflow && !err) {
 483                 block += count;
 484                 count = overflow;
 485                 goto do_more;
 486         }
 487         sb->s_dirt = 1;
 488 error_return:
 489         brelse(bitmap_bh);
 490         ext3_std_error(sb, err);
 491         return;
 492 }
 493
 494 /* Free given blocks, update quota and i_blocks field */
 495 void ext3_free_blocks(handle_t *handle, struct inode *inode,
 496                         unsigned long block, unsigned long count)
 497 {
 498         struct super_block * sb;
 499         int dquot_freed_blocks;
 500
 501         sb = inode->i_sb;
 502         if (!sb) {
 503                 printk ("ext3_free_blocks: nonexistent device");
 504                 return;
 505         }
 506         ext3_free_blocks_sb(handle, sb, block, count, &dquot_freed_blocks);
 507         if (dquot_freed_blocks)
 508                 DQUOT_FREE_BLOCK(inode, dquot_freed_blocks);
 509         return;
 510 }
 511
 512 /*
 513  * For ext3 allocations, we must not reuse any blocks which are
 514  * allocated in the bitmap buffer's "last committed data" copy.  This
 515  * prevents deletes from freeing up the page for reuse until we have
 516  * committed the delete transaction.
 517  *
 518  * If we didn't do this, then deleting something and reallocating it as
 519  * data would allow the old block to be overwritten before the
 520  * transaction committed (because we force data to disk before commit).
 521  * This would lead to corruption if we crashed between overwriting the
 522  * data and committing the delete.
 523  *
 524  * @@@ We may want to make this allocation behaviour conditional on
 525  * data-writes at some point, and disable it for metadata allocations or
 526  * sync-data inodes.
 527  */
 528 static int ext3_test_allocatable(int nr, struct buffer_head *bh)
 529 {
 530         int ret;
 531         struct journal_head *jh = bh2jh(bh);
 532
 533         if (ext3_test_bit(nr, bh->b_data))
 534                 return 0;
 535
 536         jbd_lock_bh_state(bh);
 537         if (!jh->b_committed_data)
 538                 ret = 1;
 539         else
 540                 ret = !ext3_test_bit(nr, jh->b_committed_data);
 541         jbd_unlock_bh_state(bh);
 542         return ret;
 543 }
 544
 545 static int
 546 bitmap_search_next_usable_block(int start, struct buffer_head *bh,
 547                                         int maxblocks)
 548 {
 549         int next;
 550         struct journal_head *jh = bh2jh(bh);
 551
 552         /*
 553          * The bitmap search --- search forward alternately through the actual
 554          * bitmap and the last-committed copy until we find a bit free in
 555          * both
 556          */
 557         while (start < maxblocks) {
 558                 next = ext3_find_next_zero_bit(bh->b_data, maxblocks, start);
 559                 if (next >= maxblocks)
 560                         return -1;
 561                 if (ext3_test_allocatable(next, bh))
 562                         return next;
 563                 jbd_lock_bh_state(bh);
 564                 if (jh->b_committed_data)
 565                         start = ext3_find_next_zero_bit(jh->b_committed_data,
 566                                                         maxblocks, next);
 567                 jbd_unlock_bh_state(bh);
 568         }
 569         return -1;
 570 }
 571
 572 /*
 573  * Find an allocatable block in a bitmap.  We honour both the bitmap and
 574  * its last-committed copy (if that exists), and perform the "most
 575  * appropriate allocation" algorithm of looking for a free block near
 576  * the initial goal; then for a free byte somewhere in the bitmap; then
 577  * for any free bit in the bitmap.
 578  */
 579 static int
 580 find_next_usable_block(int start, struct buffer_head *bh, int maxblocks)
 581 {
 582         int here, next;
 583         char *p, *r;
 584
 585         if (start > 0) {
 586                 /*
 587                  * The goal was occupied; search forward for a free
 588                  * block within the next XX blocks.
 589                  *
 590                  * end_goal is more or less random, but it has to be
 591                  * less than EXT3_BLOCKS_PER_GROUP. Aligning up to the
 592                  * next 64-bit boundary is simple..
 593                  */
 594                 int end_goal = (start + 63) & ~63;
 595                 if (end_goal > maxblocks)
 596                         end_goal = maxblocks;
 597                 here = ext3_find_next_zero_bit(bh->b_data, end_goal, start);
 598                 if (here < end_goal && ext3_test_allocatable(here, bh))
 599                         return here;
 600                 ext3_debug("Bit not found near goal\n");
 601         }
 602
 603         here = start;
 604         if (here < 0)
 605                 here = 0;
 606
 607         p = ((char *)bh->b_data) + (here >> 3);
 608         r = memscan(p, 0, (maxblocks - here + 7) >> 3);
 609         next = (r - ((char *)bh->b_data)) << 3;
 610
 611         if (next < maxblocks && next >= start && ext3_test_allocatable(next, bh))
 612                 return next;
 613
 614         /*
 615          * The bitmap search --- search forward alternately through the actual
 616          * bitmap and the last-committed copy until we find a bit free in
 617          * both
 618          */
 619         here = bitmap_search_next_usable_block(here, bh, maxblocks);
 620         return here;
 621 }
 622
 623 /*
 624  * We think we can allocate this block in this bitmap.  Try to set the bit.
 625  * If that succeeds then check that nobody has allocated and then freed the
 626  * block since we saw that is was not marked in b_committed_data.  If it _was_
 627  * allocated and freed then clear the bit in the bitmap again and return
 628  * zero (failure).
 629  */
 630 static inline int
 631 claim_block(spinlock_t *lock, int block, struct buffer_head *bh)
 632 {
 633         struct journal_head *jh = bh2jh(bh);
 634         int ret;
 635
 636         if (ext3_set_bit_atomic(lock, block, bh->b_data))
 637                 return 0;
 638         jbd_lock_bh_state(bh);
 639         if (jh->b_committed_data && ext3_test_bit(block,jh->b_committed_data)) {
 640                 ext3_clear_bit_atomic(lock, block, bh->b_data);
 641                 ret = 0;
 642         } else {
 643                 ret = 1;
 644         }
 645         jbd_unlock_bh_state(bh);
 646         return ret;
 647 }
 648
 649 /*
 650  * If we failed to allocate the desired block then we may end up crossing to a
 651  * new bitmap.  In that case we must release write access to the old one via
 652  * ext3_journal_release_buffer(), else we'll run out of credits.
 653  */
 654 static int
 655 ext3_try_to_allocate(struct super_block *sb, handle_t *handle, int group,
 656                         struct buffer_head *bitmap_bh, int goal,
 657                         unsigned long *count, struct ext3_reserve_window *my_rsv)
 658 {
 659         int group_first_block, start, end;
 660         unsigned long num = 0;
 661
 662         /* we do allocation within the reservation window if we have a window */
 663         if (my_rsv) {
 664                 group_first_block =
 665                         le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block) +
 666                         group * EXT3_BLOCKS_PER_GROUP(sb);
 667                 if (my_rsv->_rsv_start >= group_first_block)
 668                         start = my_rsv->_rsv_start - group_first_block;
 669                 else
 670                         /* reservation window cross group boundary */
 671                         start = 0;
 672                 end = my_rsv->_rsv_end - group_first_block + 1;
 673                 if (end > EXT3_BLOCKS_PER_GROUP(sb))
 674                         /* reservation window crosses group boundary */
 675                         end = EXT3_BLOCKS_PER_GROUP(sb);
 676                 if ((start <= goal) && (goal < end))
 677                         start = goal;
 678                 else
 679                         goal = -1;
 680         } else {
 681                 if (goal > 0)
 682                         start = goal;
 683                 else
 684                         start = 0;
 685                 end = EXT3_BLOCKS_PER_GROUP(sb);
 686         }
 687
 688         BUG_ON(start > EXT3_BLOCKS_PER_GROUP(sb));
 689
 690 repeat:
 691         if (goal < 0 || !ext3_test_allocatable(goal, bitmap_bh)) {
 692                 goal = find_next_usable_block(start, bitmap_bh, end);
 693                 if (goal < 0)
 694                         goto fail_access;
 695                 if (!my_rsv) {
 696                         int i;
 697
 698                         for (i = 0; i < 7 && goal > start &&
 699                                         ext3_test_allocatable(goal - 1,
 700                                                                 bitmap_bh);
 701                                         i++, goal--)
 702                                 ;
 703                 }
 704         }
 705         start = goal;
 706
 707         if (!claim_block(sb_bgl_lock(EXT3_SB(sb), group), goal, bitmap_bh)) {
 708                 /*
 709                  * The block was allocated by another thread, or it was
 710                  * allocated and then freed by another thread
 711                  */
 712                 start++;
 713                 goal++;
 714                 if (start >= end)
 715                         goto fail_access;
 716                 goto repeat;
 717         }
 718         num++;
 719         goal++;
 720         while (num < *count && goal < end
 721                 && ext3_test_allocatable(goal, bitmap_bh)
 722                 && claim_block(sb_bgl_lock(EXT3_SB(sb), group), goal, bitmap_bh)) {
 723                 num++;
 724                 goal++;
 725         }
 726         *count = num;
 727         return goal - num;
 728 fail_access:
 729         *count = num;
 730         return -1;
 731 }
 732
 733 /**
 734  *      find_next_reservable_window():
 735  *              find a reservable space within the given range.
 736  *              It does not allocate the reservation window for now:
 737  *              alloc_new_reservation() will do the work later.
 738  *
 739  *      @search_head: the head of the searching list;
 740  *              This is not necessarily the list head of the whole filesystem
 741  *
 742  *              We have both head and start_block to assist the search
 743  *              for the reservable space. The list starts from head,
 744  *              but we will shift to the place where start_block is,
 745  *              then start from there, when looking for a reservable space.
 746  *
 747  *      @size: the target new reservation window size
 748  *
 749  *      @group_first_block: the first block we consider to start
 750  *                      the real search from
 751  *
 752  *      @last_block:
 753  *              the maximum block number that our goal reservable space
 754  *              could start from. This is normally the last block in this
 755  *              group. The search will end when we found the start of next
 756  *              possible reservable space is out of this boundary.
 757  *              This could handle the cross boundary reservation window
 758  *              request.
 759  *
 760  *      basically we search from the given range, rather than the whole
 761  *      reservation double linked list, (start_block, last_block)
 762  *      to find a free region that is of my size and has not
 763  *      been reserved.
 764  *
 765  */
 766 static int find_next_reservable_window(
 767                                 struct ext3_reserve_window_node *search_head,
 768                                 struct ext3_reserve_window_node *my_rsv,
 769                                 struct super_block * sb, int start_block,
 770                                 int last_block)
 771 {
 772         struct rb_node *next;
 773         struct ext3_reserve_window_node *rsv, *prev;
 774         int cur;
 775         int size = my_rsv->rsv_goal_size;
 776
 777         /* TODO: make the start of the reservation window byte-aligned */
 778         /* cur = *start_block & ~7;*/
 779         cur = start_block;
 780         rsv = search_head;
 781         if (!rsv)
 782                 return -1;
 783
 784         while (1) {
 785                 if (cur <= rsv->rsv_end)
 786                         cur = rsv->rsv_end + 1;
 787
 788                 /* TODO?
 789                  * in the case we could not find a reservable space
 790                  * that is what is expected, during the re-search, we could
 791                  * remember what's the largest reservable space we could have
 792                  * and return that one.
 793                  *
 794                  * For now it will fail if we could not find the reservable
 795                  * space with expected-size (or more)...
 796                  */
 797                 if (cur > last_block)
 798                         return -1;              /* fail */
 799
 800                 prev = rsv;
 801                 next = rb_next(&rsv->rsv_node);
 802                 rsv = list_entry(next,struct ext3_reserve_window_node,rsv_node);
 803
 804                 /*
 805                  * Reached the last reservation, we can just append to the
 806                  * previous one.
 807                  */
 808                 if (!next)
 809                         break;
 810
 811                 if (cur + size <= rsv->rsv_start) {
 812                         /*
 813                          * Found a reserveable space big enough.  We could
 814                          * have a reservation across the group boundary here
 815                          */
 816                         break;
 817                 }
 818         }
 819         /*
 820          * we come here either :
 821          * when we reach the end of the whole list,
 822          * and there is empty reservable space after last entry in the list.
 823          * append it to the end of the list.
 824          *
 825          * or we found one reservable space in the middle of the list,
 826          * return the reservation window that we could append to.
 827          * succeed.
 828          */
 829
 830         if ((prev != my_rsv) && (!rsv_is_empty(&my_rsv->rsv_window)))
 831                 rsv_window_remove(sb, my_rsv);
 832
 833         /*
 834          * Let's book the whole avaliable window for now.  We will check the
 835          * disk bitmap later and then, if there are free blocks then we adjust
 836          * the window size if it's larger than requested.
 837          * Otherwise, we will remove this node from the tree next time
 838          * call find_next_reservable_window.
 839          */
 840         my_rsv->rsv_start = cur;
 841         my_rsv->rsv_end = cur + size - 1;
 842         my_rsv->rsv_alloc_hit = 0;
 843
 844         if (prev != my_rsv)
 845                 ext3_rsv_window_add(sb, my_rsv);
 846
 847         return 0;
 848 }
 849
 850 /**
 851  *      alloc_new_reservation()--allocate a new reservation window
 852  *
 853  *              To make a new reservation, we search part of the filesystem
 854  *              reservation list (the list that inside the group). We try to
 855  *              allocate a new reservation window near the allocation goal,
 856  *              or the beginning of the group, if there is no goal.
 857  *
 858  *              We first find a reservable space after the goal, then from
 859  *              there, we check the bitmap for the first free block after
 860  *              it. If there is no free block until the end of group, then the
 861  *              whole group is full, we failed. Otherwise, check if the free
 862  *              block is inside the expected reservable space, if so, we
 863  *              succeed.
 864  *              If the first free block is outside the reservable space, then
 865  *              start from the first free block, we search for next available
 866  *              space, and go on.
 867  *
 868  *      on succeed, a new reservation will be found and inserted into the list
 869  *      It contains at least one free block, and it does not overlap with other
 870  *      reservation windows.
 871  *
 872  *      failed: we failed to find a reservation window in this group
 873  *
 874  *      @rsv: the reservation
 875  *
 876  *      @goal: The goal (group-relative).  It is where the search for a
 877  *              free reservable space should start from.
 878  *              if we have a goal(goal >0 ), then start from there,
 879  *              no goal(goal = -1), we start from the first block
 880  *              of the group.
 881  *
 882  *      @sb: the super block
 883  *      @group: the group we are trying to allocate in
 884  *      @bitmap_bh: the block group block bitmap
 885  *
 886  */
 887 static int alloc_new_reservation(struct ext3_reserve_window_node *my_rsv,
 888                 int goal, struct super_block *sb,
 889                 unsigned int group, struct buffer_head *bitmap_bh)
 890 {
 891         struct ext3_reserve_window_node *search_head;
 892         int group_first_block, group_end_block, start_block;
 893         int first_free_block;
 894         struct rb_root *fs_rsv_root = &EXT3_SB(sb)->s_rsv_window_root;
 895         unsigned long size;
 896         int ret;
 897         spinlock_t *rsv_lock = &EXT3_SB(sb)->s_rsv_window_lock;
 898
 899         group_first_block = le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block) +
 900                                 group * EXT3_BLOCKS_PER_GROUP(sb);
 901         group_end_block = group_first_block + EXT3_BLOCKS_PER_GROUP(sb) - 1;
 902
 903         if (goal < 0)
 904                 start_block = group_first_block;
 905         else
 906                 start_block = goal + group_first_block;
 907
 908         size = my_rsv->rsv_goal_size;
 909
 910         if (!rsv_is_empty(&my_rsv->rsv_window)) {
 911                 /*
 912                  * if the old reservation is cross group boundary
 913                  * and if the goal is inside the old reservation window,
 914                  * we will come here when we just failed to allocate from
 915                  * the first part of the window. We still have another part
 916                  * that belongs to the next group. In this case, there is no
 917                  * point to discard our window and try to allocate a new one
 918                  * in this group(which will fail). we should
 919                  * keep the reservation window, just simply move on.
 920                  *
 921                  * Maybe we could shift the start block of the reservation
 922                  * window to the first block of next group.
 923                  */
 924
 925                 if ((my_rsv->rsv_start <= group_end_block) &&
 926                                 (my_rsv->rsv_end > group_end_block) &&
 927                                 (start_block >= my_rsv->rsv_start))
 928                         return -1;
 929
 930                 if ((my_rsv->rsv_alloc_hit >
 931                      (my_rsv->rsv_end - my_rsv->rsv_start + 1) / 2)) {
 932                         /*
 933                          * if we previously allocation hit ration is greater than half
 934                          * we double the size of reservation window next time
 935                          * otherwise keep the same
 936                          */
 937                         size = size * 2;
 938                         if (size > EXT3_MAX_RESERVE_BLOCKS)
 939                                 size = EXT3_MAX_RESERVE_BLOCKS;
 940                         my_rsv->rsv_goal_size= size;
 941                 }
 942         }
 943
 944         spin_lock(rsv_lock);
 945         /*
 946          * shift the search start to the window near the goal block
 947          */
 948         search_head = search_reserve_window(fs_rsv_root, start_block);
 949
 950         /*
 951          * find_next_reservable_window() simply finds a reservable window
 952          * inside the given range(start_block, group_end_block).
 953          *
 954          * To make sure the reservation window has a free bit inside it, we
 955          * need to check the bitmap after we found a reservable window.
 956          */
 957 retry:
 958         ret = find_next_reservable_window(search_head, my_rsv, sb,
 959                                                 start_block, group_end_block);
 960
 961         if (ret == -1) {
 962                 if (!rsv_is_empty(&my_rsv->rsv_window))
 963                         rsv_window_remove(sb, my_rsv);
 964                 spin_unlock(rsv_lock);
 965                 return -1;
 966         }
 967
 968         /*
 969          * On success, find_next_reservable_window() returns the
 970          * reservation window where there is a reservable space after it.
 971          * Before we reserve this reservable space, we need
 972          * to make sure there is at least a free block inside this region.
 973          *
 974          * searching the first free bit on the block bitmap and copy of
 975          * last committed bitmap alternatively, until we found a allocatable
 976          * block. Search start from the start block of the reservable space
 977          * we just found.
 978          */
 979         spin_unlock(rsv_lock);
 980         first_free_block = bitmap_search_next_usable_block(
 981                         my_rsv->rsv_start - group_first_block,
 982                         bitmap_bh, group_end_block - group_first_block + 1);
 983
 984         if (first_free_block < 0) {
 985                 /*
 986                  * no free block left on the bitmap, no point
 987                  * to reserve the space. return failed.
 988                  */
 989                 spin_lock(rsv_lock);
 990                 if (!rsv_is_empty(&my_rsv->rsv_window))
 991                         rsv_window_remove(sb, my_rsv);
 992                 spin_unlock(rsv_lock);
 993                 return -1;              /* failed */
 994         }
 995
 996         start_block = first_free_block + group_first_block;
 997         /*
 998          * check if the first free block is within the
 999          * free space we just reserved
1000          */
1001         if (start_block >= my_rsv->rsv_start && start_block < my_rsv->rsv_end)
1002                 return 0;               /* success */
1003         /*
1004          * if the first free bit we found is out of the reservable space
1005          * continue search for next reservable space,
1006          * start from where the free block is,
1007          * we also shift the list head to where we stopped last time
1008          */
1009         search_head = my_rsv;
1010         spin_lock(rsv_lock);
1011         goto retry;
1012 }
1013
1014 static void try_to_extend_reservation(struct ext3_reserve_window_node *my_rsv,
1015                         struct super_block *sb, int size)
1016 {
1017         struct ext3_reserve_window_node *next_rsv;
1018         struct rb_node *next;
1019         spinlock_t *rsv_lock = &EXT3_SB(sb)->s_rsv_window_lock;
1020
1021         if (!spin_trylock(rsv_lock))
1022                 return;
1023
1024         next = rb_next(&my_rsv->rsv_node);
1025
1026         if (!next)
1027                 my_rsv->rsv_end += size;
1028         else {
1029                 next_rsv = list_entry(next, struct ext3_reserve_window_node, rsv_node);
1030
1031                 if ((next_rsv->rsv_start - my_rsv->rsv_end - 1) >= size)
1032                         my_rsv->rsv_end += size;
1033                 else
1034                         my_rsv->rsv_end = next_rsv->rsv_start - 1;
1035         }
1036         spin_unlock(rsv_lock);
1037 }
1038
1039 /*
1040  * This is the main function used to allocate a new block and its reservation
1041  * window.
1042  *
1043  * Each time when a new block allocation is need, first try to allocate from
1044  * its own reservation.  If it does not have a reservation window, instead of
1045  * looking for a free bit on bitmap first, then look up the reservation list to
1046  * see if it is inside somebody else's reservation window, we try to allocate a
1047  * reservation window for it starting from the goal first. Then do the block
1048  * allocation within the reservation window.
1049  *
1050  * This will avoid keeping on searching the reservation list again and
1051  * again when somebody is looking for a free block (without
1052  * reservation), and there are lots of free blocks, but they are all
1053  * being reserved.
1054  *
1055  * We use a sorted double linked list for the per-filesystem reservation list.
1056  * The insert, remove and find a free space(non-reserved) operations for the
1057  * sorted double linked list should be fast.
1058  *
1059  */
1060 static int
1061 ext3_try_to_allocate_with_rsv(struct super_block *sb, handle_t *handle,
1062                         unsigned int group, struct buffer_head *bitmap_bh,
1063                         int goal, struct ext3_reserve_window_node * my_rsv,
1064                         unsigned long *count, int *errp)
1065 {
1066         unsigned long group_first_block;
1067         int ret = 0;
1068         int fatal;
1069         unsigned long num = *count;
1070
1071         *errp = 0;
1072
1073         /*
1074          * Make sure we use undo access for the bitmap, because it is critical
1075          * that we do the frozen_data COW on bitmap buffers in all cases even
1076          * if the buffer is in BJ_Forget state in the committing transaction.
1077          */
1078         BUFFER_TRACE(bitmap_bh, "get undo access for new block");
1079         fatal = ext3_journal_get_undo_access(handle, bitmap_bh);
1080         if (fatal) {
1081                 *errp = fatal;
1082                 return -1;
1083         }
1084
1085         /*
1086          * we don't deal with reservation when
1087          * filesystem is mounted without reservation
1088          * or the file is not a regular file
1089          * or last attempt to allocate a block with reservation turned on failed
1090          */
1091         if (my_rsv == NULL ) {
1092                 ret = ext3_try_to_allocate(sb, handle, group, bitmap_bh,
1093                                                 goal, count, NULL);
1094                 goto out;
1095         }
1096         /*
1097          * goal is a group relative block number (if there is a goal)
1098          * 0 < goal < EXT3_BLOCKS_PER_GROUP(sb)
1099          * first block is a filesystem wide block number
1100          * first block is the block number of the first block in this group
1101          */
1102         group_first_block = le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block) +
1103                         group * EXT3_BLOCKS_PER_GROUP(sb);
1104
1105         /*
1106          * Basically we will allocate a new block from inode's reservation
1107          * window.
1108          *
1109          * We need to allocate a new reservation window, if:
1110          * a) inode does not have a reservation window; or
1111          * b) last attempt to allocate a block from existing reservation
1112          *    failed; or
1113          * c) we come here with a goal and with a reservation window
1114          *
1115          * We do not need to allocate a new reservation window if we come here
1116          * at the beginning with a goal and the goal is inside the window, or
1117          * we don't have a goal but already have a reservation window.
1118          * then we could go to allocate from the reservation window directly.
1119          */
1120         while (1) {
1121                 if (rsv_is_empty(&my_rsv->rsv_window) || (ret < 0) ||
1122                         !goal_in_my_reservation(&my_rsv->rsv_window, goal, group, sb)) {
1123                         if (my_rsv->rsv_goal_size < *count)
1124                                 my_rsv->rsv_goal_size = *count;
1125                         ret = alloc_new_reservation(my_rsv, goal, sb,
1126                                                         group, bitmap_bh);
1127                         if (ret < 0)
1128                                 break;                  /* failed */
1129
1130                         if (!goal_in_my_reservation(&my_rsv->rsv_window, goal, group, sb))
1131                                 goal = -1;
1132                 } else if (goal > 0 && (my_rsv->rsv_end-goal+1) < *count)
1133                         try_to_extend_reservation(my_rsv, sb,
1134                                         *count-my_rsv->rsv_end + goal - 1);
1135
1136                 if ((my_rsv->rsv_start >= group_first_block + EXT3_BLOCKS_PER_GROUP(sb))
1137                     || (my_rsv->rsv_end < group_first_block))
1138                         BUG();
1139                 ret = ext3_try_to_allocate(sb, handle, group, bitmap_bh, goal,
1140                                            &num, &my_rsv->rsv_window);
1141                 if (ret >= 0) {
1142                         my_rsv->rsv_alloc_hit += num;
1143                         *count = num;
1144                         break;                          /* succeed */
1145                 }
1146                 num = *count;
1147         }
1148 out:
1149         if (ret >= 0) {
1150                 BUFFER_TRACE(bitmap_bh, "journal_dirty_metadata for "
1151                                         "bitmap block");
1152                 fatal = ext3_journal_dirty_metadata(handle, bitmap_bh);
1153                 if (fatal) {
1154                         *errp = fatal;
1155                         return -1;
1156                 }
1157                 return ret;
1158         }
1159
1160         BUFFER_TRACE(bitmap_bh, "journal_release_buffer");
1161         ext3_journal_release_buffer(handle, bitmap_bh);
1162         return ret;
1163 }
1164
1165 static int ext3_has_free_blocks(struct ext3_sb_info *sbi)
1166 {
1167         int free_blocks, root_blocks;
1168
1169         free_blocks = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
1170         root_blocks = le32_to_cpu(sbi->s_es->s_r_blocks_count);
1171         if (free_blocks < root_blocks + 1 && !capable(CAP_SYS_RESOURCE) &&
1172                 sbi->s_resuid != current->fsuid &&
1173                 (sbi->s_resgid == 0 || !in_group_p (sbi->s_resgid))) {
1174                 return 0;
1175         }
1176         return 1;
1177 }
1178
1179 /*
1180  * ext3_should_retry_alloc() is called when ENOSPC is returned, and if
1181  * it is profitable to retry the operation, this function will wait
1182  * for the current or commiting transaction to complete, and then
1183  * return TRUE.
1184  */
1185 int ext3_should_retry_alloc(struct super_block *sb, int *retries)
1186 {
1187         if (!ext3_has_free_blocks(EXT3_SB(sb)) || (*retries)++ > 3)
1188                 return 0;
1189
1190         jbd_debug(1, "%s: retrying operation after ENOSPC\n", sb->s_id);
1191
1192         return journal_force_commit_nested(EXT3_SB(sb)->s_journal);
1193 }
1194
1195 /*
1196  * ext3_new_block uses a goal block to assist allocation.  If the goal is
1197  * free, or there is a free block within 32 blocks of the goal, that block
1198  * is allocated.  Otherwise a forward search is made for a free block; within
1199  * each block group the search first looks for an entire free byte in the block
1200  * bitmap, and then for any free bit if that fails.
1201  * This function also updates quota and i_blocks field.
1202  */
1203 int ext3_new_blocks(handle_t *handle, struct inode *inode,
1204                         unsigned long goal, unsigned long *count, int *errp)
1205 {
1206         struct buffer_head *bitmap_bh = NULL;
1207         struct buffer_head *gdp_bh;
1208         int group_no;
1209         int goal_group;
1210         int ret_block;
1211         int bgi;                        /* blockgroup iteration index */
1212         int target_block;
1213         int fatal = 0, err;
1214         int performed_allocation = 0;
1215         int free_blocks;
1216         struct super_block *sb;
1217         struct ext3_group_desc *gdp;
1218         struct ext3_super_block *es;
1219         struct ext3_sb_info *sbi;
1220         struct ext3_reserve_window_node *my_rsv = NULL;
1221         struct ext3_block_alloc_info *block_i;
1222         unsigned short windowsz = 0;
1223 #ifdef EXT3FS_DEBUG
1224         static int goal_hits, goal_attempts;
1225 #endif
1226         unsigned long ngroups;
1227         unsigned long num = *count;
1228
1229         *errp = -ENOSPC;
1230         sb = inode->i_sb;
1231         if (!sb) {
1232                 printk("ext3_new_block: nonexistent device");
1233                 return 0;
1234         }
1235
1236         /*
1237          * Check quota for allocation of this block.
1238          */
1239         if (DQUOT_ALLOC_BLOCK(inode, num)) {
1240                 *errp = -EDQUOT;
1241                 return 0;
1242         }
1243
1244         sbi = EXT3_SB(sb);
1245         es = EXT3_SB(sb)->s_es;
1246         ext3_debug("goal=%lu.\n", goal);
1247         /*
1248          * Allocate a block from reservation only when
1249          * filesystem is mounted with reservation(default,-o reservation), and
1250          * it's a regular file, and
1251          * the desired window size is greater than 0 (One could use ioctl
1252          * command EXT3_IOC_SETRSVSZ to set the window size to 0 to turn off
1253          * reservation on that particular file)
1254          */
1255         block_i = EXT3_I(inode)->i_block_alloc_info;
1256         if (block_i && ((windowsz = block_i->rsv_window_node.rsv_goal_size) > 0))
1257                 my_rsv = &block_i->rsv_window_node;
1258
1259         if (!ext3_has_free_blocks(sbi)) {
1260                 *errp = -ENOSPC;
1261                 goto out;
1262         }
1263
1264         /*
1265          * First, test whether the goal block is free.
1266          */
1267         if (goal < le32_to_cpu(es->s_first_data_block) ||
1268             goal >= le32_to_cpu(es->s_blocks_count))
1269                 goal = le32_to_cpu(es->s_first_data_block);
1270         group_no = (goal - le32_to_cpu(es->s_first_data_block)) /
1271                         EXT3_BLOCKS_PER_GROUP(sb);
1272         gdp = ext3_get_group_desc(sb, group_no, &gdp_bh);
1273         if (!gdp)
1274                 goto io_error;
1275
1276         goal_group = group_no;
1277 retry:
1278         free_blocks = le16_to_cpu(gdp->bg_free_blocks_count);
1279         /*
1280          * if there is not enough free blocks to make a new resevation
1281          * turn off reservation for this allocation
1282          */
1283         if (my_rsv && (free_blocks < windowsz)
1284                 && (rsv_is_empty(&my_rsv->rsv_window)))
1285                 my_rsv = NULL;
1286
1287         if (free_blocks > 0) {
1288                 ret_block = ((goal - le32_to_cpu(es->s_first_data_block)) %
1289                                 EXT3_BLOCKS_PER_GROUP(sb));
1290                 bitmap_bh = read_block_bitmap(sb, group_no);
1291                 if (!bitmap_bh)
1292                         goto io_error;
1293                 ret_block = ext3_try_to_allocate_with_rsv(sb, handle, group_no,
1294                                         bitmap_bh, ret_block, my_rsv, &num, &fatal);
1295                 if (fatal)
1296                         goto out;
1297                 if (ret_block >= 0)
1298                         goto allocated;
1299         }
1300
1301         ngroups = EXT3_SB(sb)->s_groups_count;
1302         smp_rmb();
1303
1304         /*
1305          * Now search the rest of the groups.  We assume that
1306          * i and gdp correctly point to the last group visited.
1307          */
1308         for (bgi = 0; bgi < ngroups; bgi++) {
1309                 group_no++;
1310                 if (group_no >= ngroups)
1311                         group_no = 0;
1312                 gdp = ext3_get_group_desc(sb, group_no, &gdp_bh);
1313                 if (!gdp) {
1314                         *errp = -EIO;
1315                         goto out;
1316                 }
1317                 free_blocks = le16_to_cpu(gdp->bg_free_blocks_count);
1318                 /*
1319                  * skip this group if the number of
1320                  * free blocks is less than half of the reservation
1321                  * window size.
1322                  */
1323                 if (free_blocks <= (windowsz/2))
1324                         continue;
1325
1326                 brelse(bitmap_bh);
1327                 bitmap_bh = read_block_bitmap(sb, group_no);
1328                 if (!bitmap_bh)
1329                         goto io_error;
1330                 ret_block = ext3_try_to_allocate_with_rsv(sb, handle, group_no,
1331                                         bitmap_bh, -1, my_rsv, &num, &fatal);
1332                 if (fatal)
1333                         goto out;
1334                 if (ret_block >= 0)
1335                         goto allocated;
1336         }
1337         /*
1338          * We may end up a bogus ealier ENOSPC error due to
1339          * filesystem is "full" of reservations, but
1340          * there maybe indeed free blocks avaliable on disk
1341          * In this case, we just forget about the reservations
1342          * just do block allocation as without reservations.
1343          */
1344         if (my_rsv) {
1345                 my_rsv = NULL;
1346                 group_no = goal_group;
1347                 goto retry;
1348         }
1349         /* No space left on the device */
1350         *errp = -ENOSPC;
1351         goto out;
1352
1353 allocated:
1354
1355         ext3_debug("using block group %d(%d)\n",
1356                         group_no, gdp->bg_free_blocks_count);
1357
1358         BUFFER_TRACE(gdp_bh, "get_write_access");
1359         fatal = ext3_journal_get_write_access(handle, gdp_bh);
1360         if (fatal)
1361                 goto out;
1362
1363         target_block = ret_block + group_no * EXT3_BLOCKS_PER_GROUP(sb)
1364                                 + le32_to_cpu(es->s_first_data_block);
1365
1366         if (in_range(le32_to_cpu(gdp->bg_block_bitmap), target_block, num) ||
1367             in_range(le32_to_cpu(gdp->bg_inode_bitmap), target_block, num) ||
1368             in_range(target_block, le32_to_cpu(gdp->bg_inode_table),
1369                       EXT3_SB(sb)->s_itb_per_group) ||
1370             in_range(target_block + num - 1, le32_to_cpu(gdp->bg_inode_table),
1371                       EXT3_SB(sb)->s_itb_per_group))
1372                 ext3_error(sb, "ext3_new_block",
1373                             "Allocating block in system zone - "
1374                             "blocks from %u, length %lu", target_block, num);
1375
1376         performed_allocation = 1;
1377
1378 #ifdef CONFIG_JBD_DEBUG
1379         {
1380                 struct buffer_head *debug_bh;
1381
1382                 /* Record bitmap buffer state in the newly allocated block */
1383                 debug_bh = sb_find_get_block(sb, target_block);
1384                 if (debug_bh) {
1385                         BUFFER_TRACE(debug_bh, "state when allocated");
1386                         BUFFER_TRACE2(debug_bh, bitmap_bh, "bitmap state");
1387                         brelse(debug_bh);
1388                 }
1389         }
1390         jbd_lock_bh_state(bitmap_bh);
1391         spin_lock(sb_bgl_lock(sbi, group_no));
1392         if (buffer_jbd(bitmap_bh) && bh2jh(bitmap_bh)->b_committed_data) {
1393                 int i;
1394
1395                 for (i = 0; i < num; i++) {
1396                         if (ext3_test_bit(ret_block,
1397                                         bh2jh(bitmap_bh)->b_committed_data)) {
1398                                 printk("%s: block was unexpectedly set in "
1399                                         "b_committed_data\n", __FUNCTION__);
1400                         }
1401                 }
1402         }
1403         ext3_debug("found bit %d\n", ret_block);
1404         spin_unlock(sb_bgl_lock(sbi, group_no));
1405         jbd_unlock_bh_state(bitmap_bh);
1406 #endif
1407
1408         /* ret_block was blockgroup-relative.  Now it becomes fs-relative */
1409         ret_block = target_block;
1410
1411         if (ret_block + num - 1 >= le32_to_cpu(es->s_blocks_count)) {
1412                 ext3_error(sb, "ext3_new_block",
1413                             "block(%d) >= blocks count(%d) - "
1414                             "block_group = %d, es == %p ", ret_block,
1415                         le32_to_cpu(es->s_blocks_count), group_no, es);
1416                 goto out;
1417         }
1418
1419         /*
1420          * It is up to the caller to add the new buffer to a journal
1421          * list of some description.  We don't know in advance whether
1422          * the caller wants to use it as metadata or data.
1423          */
1424         ext3_debug("allocating block %d. Goal hits %d of %d.\n",
1425                         ret_block, goal_hits, goal_attempts);
1426
1427         spin_lock(sb_bgl_lock(sbi, group_no));
1428         gdp->bg_free_blocks_count =
1429                         cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count) - num);
1430         spin_unlock(sb_bgl_lock(sbi, group_no));
1431         percpu_counter_mod(&sbi->s_freeblocks_counter, -num);
1432
1433         BUFFER_TRACE(gdp_bh, "journal_dirty_metadata for group descriptor");
1434         err = ext3_journal_dirty_metadata(handle, gdp_bh);
1435         if (!fatal)
1436                 fatal = err;
1437
1438         sb->s_dirt = 1;
1439         if (fatal)
1440                 goto out;
1441
1442         *errp = 0;
1443         brelse(bitmap_bh);
1444         DQUOT_FREE_BLOCK(inode, *count-num);
1445         *count = num;
1446         return ret_block;
1447
1448 io_error:
1449         *errp = -EIO;
1450 out:
1451         if (fatal) {
1452                 *errp = fatal;
1453                 ext3_std_error(sb, fatal);
1454         }
1455         /*
1456          * Undo the block allocation
1457          */
1458         if (!performed_allocation)
1459                 DQUOT_FREE_BLOCK(inode, *count);
1460         brelse(bitmap_bh);
1461         return 0;
1462 }
1463
1464 int ext3_new_block(handle_t *handle, struct inode *inode,
1465                         unsigned long goal, int *errp)
1466 {
1467         unsigned long count = 1;
1468
1469         return ext3_new_blocks(handle, inode, goal, &count, errp);
1470 }
1471
1472 unsigned long ext3_count_free_blocks(struct super_block *sb)
1473 {
1474         unsigned long desc_count;
1475         struct ext3_group_desc *gdp;
1476         int i;
1477         unsigned long ngroups = EXT3_SB(sb)->s_groups_count;
1478 #ifdef EXT3FS_DEBUG
1479         struct ext3_super_block *es;
1480         unsigned long bitmap_count, x;
1481         struct buffer_head *bitmap_bh = NULL;
1482
1483         es = EXT3_SB(sb)->s_es;
1484         desc_count = 0;
1485         bitmap_count = 0;
1486         gdp = NULL;
1487
1488         smp_rmb();
1489         for (i = 0; i < ngroups; i++) {
1490                 gdp = ext3_get_group_desc(sb, i, NULL);
1491                 if (!gdp)
1492                         continue;
1493                 desc_count += le16_to_cpu(gdp->bg_free_blocks_count);
1494                 brelse(bitmap_bh);
1495                 bitmap_bh = read_block_bitmap(sb, i);
1496                 if (bitmap_bh == NULL)
1497                         continue;
1498
1499                 x = ext3_count_free(bitmap_bh, sb->s_blocksize);
1500                 printk("group %d: stored = %d, counted = %lu\n",
1501                         i, le16_to_cpu(gdp->bg_free_blocks_count), x);
1502                 bitmap_count += x;
1503         }
1504         brelse(bitmap_bh);
1505         printk("ext3_count_free_blocks: stored = %u, computed = %lu, %lu\n",
1506                le32_to_cpu(es->s_free_blocks_count), desc_count, bitmap_count);
1507         return bitmap_count;
1508 #else
1509         desc_count = 0;
1510         smp_rmb();
1511         for (i = 0; i < ngroups; i++) {
1512                 gdp = ext3_get_group_desc(sb, i, NULL);
1513                 if (!gdp)
1514                         continue;
1515                 desc_count += le16_to_cpu(gdp->bg_free_blocks_count);
1516         }
1517
1518         return desc_count;
1519 #endif
1520 }
1521
1522 static inline int
1523 block_in_use(unsigned long block, struct super_block *sb, unsigned char *map)
1524 {
1525         return ext3_test_bit ((block -
1526                 le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block)) %
1527                          EXT3_BLOCKS_PER_GROUP(sb), map);
1528 }
1529
1530 static inline int test_root(int a, int b)
1531 {
1532         int num = b;
1533
1534         while (a > num)
1535                 num *= b;
1536         return num == a;
1537 }
1538
1539 static int ext3_group_sparse(int group)
1540 {
1541         if (group <= 1)
1542                 return 1;
1543         if (!(group & 1))
1544                 return 0;
1545         return (test_root(group, 7) || test_root(group, 5) ||
1546                 test_root(group, 3));
1547 }
1548
1549 /**
1550  *      ext3_bg_has_super - number of blocks used by the superblock in group
1551  *      @sb: superblock for filesystem
1552  *      @group: group number to check
1553  *
1554  *      Return the number of blocks used by the superblock (primary or backup)
1555  *      in this group.  Currently this will be only 0 or 1.
1556  */
1557 int ext3_bg_has_super(struct super_block *sb, int group)
1558 {
1559         if (EXT3_HAS_RO_COMPAT_FEATURE(sb,
1560                                 EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER) &&
1561                         !ext3_group_sparse(group))
1562                 return 0;
1563         return 1;
1564 }
1565
1566 static unsigned long ext3_bg_num_gdb_meta(struct super_block *sb, int group)
1567 {
1568         unsigned long metagroup = group / EXT3_DESC_PER_BLOCK(sb);
1569         unsigned long first = metagroup * EXT3_DESC_PER_BLOCK(sb);
1570         unsigned long last = first + EXT3_DESC_PER_BLOCK(sb) - 1;
1571
1572         if (group == first || group == first + 1 || group == last)
1573                 return 1;
1574         return 0;
1575 }
1576
1577 static unsigned long ext3_bg_num_gdb_nometa(struct super_block *sb, int group)
1578 {
1579         if (EXT3_HAS_RO_COMPAT_FEATURE(sb,
1580                                 EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER) &&
1581                         !ext3_group_sparse(group))
1582                 return 0;
1583         return EXT3_SB(sb)->s_gdb_count;
1584 }
1585
1586 /**
1587  *      ext3_bg_num_gdb - number of blocks used by the group table in group
1588  *      @sb: superblock for filesystem
1589  *      @group: group number to check
1590  *
1591  *      Return the number of blocks used by the group descriptor table
1592  *      (primary or backup) in this group.  In the future there may be a
1593  *      different number of descriptor blocks in each group.
1594  */
1595 unsigned long ext3_bg_num_gdb(struct super_block *sb, int group)
1596 {
1597         unsigned long first_meta_bg =
1598                         le32_to_cpu(EXT3_SB(sb)->s_es->s_first_meta_bg);
1599         unsigned long metagroup = group / EXT3_DESC_PER_BLOCK(sb);
1600
1601         if (!EXT3_HAS_INCOMPAT_FEATURE(sb,EXT3_FEATURE_INCOMPAT_META_BG) ||
1602                         metagroup < first_meta_bg)
1603                 return ext3_bg_num_gdb_nometa(sb,group);
1604
1605         return ext3_bg_num_gdb_meta(sb,group);
1606
1607 }