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, struct ext3_reserve_window *my_rsv)
 657 {
 658         int group_first_block, start, end;
 659
 660         /* we do allocation within the reservation window if we have a window */
 661         if (my_rsv) {
 662                 group_first_block =
 663                         le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block) +
 664                         group * EXT3_BLOCKS_PER_GROUP(sb);
 665                 if (my_rsv->_rsv_start >= group_first_block)
 666                         start = my_rsv->_rsv_start - group_first_block;
 667                 else
 668                         /* reservation window cross group boundary */
 669                         start = 0;
 670                 end = my_rsv->_rsv_end - group_first_block + 1;
 671                 if (end > EXT3_BLOCKS_PER_GROUP(sb))
 672                         /* reservation window crosses group boundary */
 673                         end = EXT3_BLOCKS_PER_GROUP(sb);
 674                 if ((start <= goal) && (goal < end))
 675                         start = goal;
 676                 else
 677                         goal = -1;
 678         } else {
 679                 if (goal > 0)
 680                         start = goal;
 681                 else
 682                         start = 0;
 683                 end = EXT3_BLOCKS_PER_GROUP(sb);
 684         }
 685
 686         BUG_ON(start > EXT3_BLOCKS_PER_GROUP(sb));
 687
 688 repeat:
 689         if (goal < 0 || !ext3_test_allocatable(goal, bitmap_bh)) {
 690                 goal = find_next_usable_block(start, bitmap_bh, end);
 691                 if (goal < 0)
 692                         goto fail_access;
 693                 if (!my_rsv) {
 694                         int i;
 695
 696                         for (i = 0; i < 7 && goal > start &&
 697                                         ext3_test_allocatable(goal - 1,
 698                                                                 bitmap_bh);
 699                                         i++, goal--)
 700                                 ;
 701                 }
 702         }
 703         start = goal;
 704
 705         if (!claim_block(sb_bgl_lock(EXT3_SB(sb), group), goal, bitmap_bh)) {
 706                 /*
 707                  * The block was allocated by another thread, or it was
 708                  * allocated and then freed by another thread
 709                  */
 710                 start++;
 711                 goal++;
 712                 if (start >= end)
 713                         goto fail_access;
 714                 goto repeat;
 715         }
 716         return goal;
 717 fail_access:
 718         return -1;
 719 }
 720
 721 /**
 722  *      find_next_reservable_window():
 723  *              find a reservable space within the given range.
 724  *              It does not allocate the reservation window for now:
 725  *              alloc_new_reservation() will do the work later.
 726  *
 727  *      @search_head: the head of the searching list;
 728  *              This is not necessarily the list head of the whole filesystem
 729  *
 730  *              We have both head and start_block to assist the search
 731  *              for the reservable space. The list starts from head,
 732  *              but we will shift to the place where start_block is,
 733  *              then start from there, when looking for a reservable space.
 734  *
 735  *      @size: the target new reservation window size
 736  *
 737  *      @group_first_block: the first block we consider to start
 738  *                      the real search from
 739  *
 740  *      @last_block:
 741  *              the maximum block number that our goal reservable space
 742  *              could start from. This is normally the last block in this
 743  *              group. The search will end when we found the start of next
 744  *              possible reservable space is out of this boundary.
 745  *              This could handle the cross boundary reservation window
 746  *              request.
 747  *
 748  *      basically we search from the given range, rather than the whole
 749  *      reservation double linked list, (start_block, last_block)
 750  *      to find a free region that is of my size and has not
 751  *      been reserved.
 752  *
 753  */
 754 static int find_next_reservable_window(
 755                                 struct ext3_reserve_window_node *search_head,
 756                                 struct ext3_reserve_window_node *my_rsv,
 757                                 struct super_block * sb, int start_block,
 758                                 int last_block)
 759 {
 760         struct rb_node *next;
 761         struct ext3_reserve_window_node *rsv, *prev;
 762         int cur;
 763         int size = my_rsv->rsv_goal_size;
 764
 765         /* TODO: make the start of the reservation window byte-aligned */
 766         /* cur = *start_block & ~7;*/
 767         cur = start_block;
 768         rsv = search_head;
 769         if (!rsv)
 770                 return -1;
 771
 772         while (1) {
 773                 if (cur <= rsv->rsv_end)
 774                         cur = rsv->rsv_end + 1;
 775
 776                 /* TODO?
 777                  * in the case we could not find a reservable space
 778                  * that is what is expected, during the re-search, we could
 779                  * remember what's the largest reservable space we could have
 780                  * and return that one.
 781                  *
 782                  * For now it will fail if we could not find the reservable
 783                  * space with expected-size (or more)...
 784                  */
 785                 if (cur > last_block)
 786                         return -1;              /* fail */
 787
 788                 prev = rsv;
 789                 next = rb_next(&rsv->rsv_node);
 790                 rsv = list_entry(next,struct ext3_reserve_window_node,rsv_node);
 791
 792                 /*
 793                  * Reached the last reservation, we can just append to the
 794                  * previous one.
 795                  */
 796                 if (!next)
 797                         break;
 798
 799                 if (cur + size <= rsv->rsv_start) {
 800                         /*
 801                          * Found a reserveable space big enough.  We could
 802                          * have a reservation across the group boundary here
 803                          */
 804                         break;
 805                 }
 806         }
 807         /*
 808          * we come here either :
 809          * when we reach the end of the whole list,
 810          * and there is empty reservable space after last entry in the list.
 811          * append it to the end of the list.
 812          *
 813          * or we found one reservable space in the middle of the list,
 814          * return the reservation window that we could append to.
 815          * succeed.
 816          */
 817
 818         if ((prev != my_rsv) && (!rsv_is_empty(&my_rsv->rsv_window)))
 819                 rsv_window_remove(sb, my_rsv);
 820
 821         /*
 822          * Let's book the whole avaliable window for now.  We will check the
 823          * disk bitmap later and then, if there are free blocks then we adjust
 824          * the window size if it's larger than requested.
 825          * Otherwise, we will remove this node from the tree next time
 826          * call find_next_reservable_window.
 827          */
 828         my_rsv->rsv_start = cur;
 829         my_rsv->rsv_end = cur + size - 1;
 830         my_rsv->rsv_alloc_hit = 0;
 831
 832         if (prev != my_rsv)
 833                 ext3_rsv_window_add(sb, my_rsv);
 834
 835         return 0;
 836 }
 837
 838 /**
 839  *      alloc_new_reservation()--allocate a new reservation window
 840  *
 841  *              To make a new reservation, we search part of the filesystem
 842  *              reservation list (the list that inside the group). We try to
 843  *              allocate a new reservation window near the allocation goal,
 844  *              or the beginning of the group, if there is no goal.
 845  *
 846  *              We first find a reservable space after the goal, then from
 847  *              there, we check the bitmap for the first free block after
 848  *              it. If there is no free block until the end of group, then the
 849  *              whole group is full, we failed. Otherwise, check if the free
 850  *              block is inside the expected reservable space, if so, we
 851  *              succeed.
 852  *              If the first free block is outside the reservable space, then
 853  *              start from the first free block, we search for next available
 854  *              space, and go on.
 855  *
 856  *      on succeed, a new reservation will be found and inserted into the list
 857  *      It contains at least one free block, and it does not overlap with other
 858  *      reservation windows.
 859  *
 860  *      failed: we failed to find a reservation window in this group
 861  *
 862  *      @rsv: the reservation
 863  *
 864  *      @goal: The goal (group-relative).  It is where the search for a
 865  *              free reservable space should start from.
 866  *              if we have a goal(goal >0 ), then start from there,
 867  *              no goal(goal = -1), we start from the first block
 868  *              of the group.
 869  *
 870  *      @sb: the super block
 871  *      @group: the group we are trying to allocate in
 872  *      @bitmap_bh: the block group block bitmap
 873  *
 874  */
 875 static int alloc_new_reservation(struct ext3_reserve_window_node *my_rsv,
 876                 int goal, struct super_block *sb,
 877                 unsigned int group, struct buffer_head *bitmap_bh)
 878 {
 879         struct ext3_reserve_window_node *search_head;
 880         int group_first_block, group_end_block, start_block;
 881         int first_free_block;
 882         struct rb_root *fs_rsv_root = &EXT3_SB(sb)->s_rsv_window_root;
 883         unsigned long size;
 884         int ret;
 885         spinlock_t *rsv_lock = &EXT3_SB(sb)->s_rsv_window_lock;
 886
 887         group_first_block = le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block) +
 888                                 group * EXT3_BLOCKS_PER_GROUP(sb);
 889         group_end_block = group_first_block + EXT3_BLOCKS_PER_GROUP(sb) - 1;
 890
 891         if (goal < 0)
 892                 start_block = group_first_block;
 893         else
 894                 start_block = goal + group_first_block;
 895
 896         size = my_rsv->rsv_goal_size;
 897
 898         if (!rsv_is_empty(&my_rsv->rsv_window)) {
 899                 /*
 900                  * if the old reservation is cross group boundary
 901                  * and if the goal is inside the old reservation window,
 902                  * we will come here when we just failed to allocate from
 903                  * the first part of the window. We still have another part
 904                  * that belongs to the next group. In this case, there is no
 905                  * point to discard our window and try to allocate a new one
 906                  * in this group(which will fail). we should
 907                  * keep the reservation window, just simply move on.
 908                  *
 909                  * Maybe we could shift the start block of the reservation
 910                  * window to the first block of next group.
 911                  */
 912
 913                 if ((my_rsv->rsv_start <= group_end_block) &&
 914                                 (my_rsv->rsv_end > group_end_block) &&
 915                                 (start_block >= my_rsv->rsv_start))
 916                         return -1;
 917
 918                 if ((my_rsv->rsv_alloc_hit >
 919                      (my_rsv->rsv_end - my_rsv->rsv_start + 1) / 2)) {
 920                         /*
 921                          * if we previously allocation hit ration is greater than half
 922                          * we double the size of reservation window next time
 923                          * otherwise keep the same
 924                          */
 925                         size = size * 2;
 926                         if (size > EXT3_MAX_RESERVE_BLOCKS)
 927                                 size = EXT3_MAX_RESERVE_BLOCKS;
 928                         my_rsv->rsv_goal_size= size;
 929                 }
 930         }
 931
 932         spin_lock(rsv_lock);
 933         /*
 934          * shift the search start to the window near the goal block
 935          */
 936         search_head = search_reserve_window(fs_rsv_root, start_block);
 937
 938         /*
 939          * find_next_reservable_window() simply finds a reservable window
 940          * inside the given range(start_block, group_end_block).
 941          *
 942          * To make sure the reservation window has a free bit inside it, we
 943          * need to check the bitmap after we found a reservable window.
 944          */
 945 retry:
 946         ret = find_next_reservable_window(search_head, my_rsv, sb,
 947                                                 start_block, group_end_block);
 948
 949         if (ret == -1) {
 950                 if (!rsv_is_empty(&my_rsv->rsv_window))
 951                         rsv_window_remove(sb, my_rsv);
 952                 spin_unlock(rsv_lock);
 953                 return -1;
 954         }
 955
 956         /*
 957          * On success, find_next_reservable_window() returns the
 958          * reservation window where there is a reservable space after it.
 959          * Before we reserve this reservable space, we need
 960          * to make sure there is at least a free block inside this region.
 961          *
 962          * searching the first free bit on the block bitmap and copy of
 963          * last committed bitmap alternatively, until we found a allocatable
 964          * block. Search start from the start block of the reservable space
 965          * we just found.
 966          */
 967         spin_unlock(rsv_lock);
 968         first_free_block = bitmap_search_next_usable_block(
 969                         my_rsv->rsv_start - group_first_block,
 970                         bitmap_bh, group_end_block - group_first_block + 1);
 971
 972         if (first_free_block < 0) {
 973                 /*
 974                  * no free block left on the bitmap, no point
 975                  * to reserve the space. return failed.
 976                  */
 977                 spin_lock(rsv_lock);
 978                 if (!rsv_is_empty(&my_rsv->rsv_window))
 979                         rsv_window_remove(sb, my_rsv);
 980                 spin_unlock(rsv_lock);
 981                 return -1;              /* failed */
 982         }
 983
 984         start_block = first_free_block + group_first_block;
 985         /*
 986          * check if the first free block is within the
 987          * free space we just reserved
 988          */
 989         if (start_block >= my_rsv->rsv_start && start_block < my_rsv->rsv_end)
 990                 return 0;               /* success */
 991         /*
 992          * if the first free bit we found is out of the reservable space
 993          * continue search for next reservable space,
 994          * start from where the free block is,
 995          * we also shift the list head to where we stopped last time
 996          */
 997         search_head = my_rsv;
 998         spin_lock(rsv_lock);
 999         goto retry;
1000 }
1001
1002 /*
1003  * This is the main function used to allocate a new block and its reservation
1004  * window.
1005  *
1006  * Each time when a new block allocation is need, first try to allocate from
1007  * its own reservation.  If it does not have a reservation window, instead of
1008  * looking for a free bit on bitmap first, then look up the reservation list to
1009  * see if it is inside somebody else's reservation window, we try to allocate a
1010  * reservation window for it starting from the goal first. Then do the block
1011  * allocation within the reservation window.
1012  *
1013  * This will avoid keeping on searching the reservation list again and
1014  * again when somebody is looking for a free block (without
1015  * reservation), and there are lots of free blocks, but they are all
1016  * being reserved.
1017  *
1018  * We use a sorted double linked list for the per-filesystem reservation list.
1019  * The insert, remove and find a free space(non-reserved) operations for the
1020  * sorted double linked list should be fast.
1021  *
1022  */
1023 static int
1024 ext3_try_to_allocate_with_rsv(struct super_block *sb, handle_t *handle,
1025                         unsigned int group, struct buffer_head *bitmap_bh,
1026                         int goal, struct ext3_reserve_window_node * my_rsv,
1027                         int *errp)
1028 {
1029         unsigned long group_first_block;
1030         int ret = 0;
1031         int fatal;
1032
1033         *errp = 0;
1034
1035         /*
1036          * Make sure we use undo access for the bitmap, because it is critical
1037          * that we do the frozen_data COW on bitmap buffers in all cases even
1038          * if the buffer is in BJ_Forget state in the committing transaction.
1039          */
1040         BUFFER_TRACE(bitmap_bh, "get undo access for new block");
1041         fatal = ext3_journal_get_undo_access(handle, bitmap_bh);
1042         if (fatal) {
1043                 *errp = fatal;
1044                 return -1;
1045         }
1046
1047         /*
1048          * we don't deal with reservation when
1049          * filesystem is mounted without reservation
1050          * or the file is not a regular file
1051          * or last attempt to allocate a block with reservation turned on failed
1052          */
1053         if (my_rsv == NULL ) {
1054                 ret = ext3_try_to_allocate(sb, handle, group, bitmap_bh, goal, NULL);
1055                 goto out;
1056         }
1057         /*
1058          * goal is a group relative block number (if there is a goal)
1059          * 0 < goal < EXT3_BLOCKS_PER_GROUP(sb)
1060          * first block is a filesystem wide block number
1061          * first block is the block number of the first block in this group
1062          */
1063         group_first_block = le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block) +
1064                         group * EXT3_BLOCKS_PER_GROUP(sb);
1065
1066         /*
1067          * Basically we will allocate a new block from inode's reservation
1068          * window.
1069          *
1070          * We need to allocate a new reservation window, if:
1071          * a) inode does not have a reservation window; or
1072          * b) last attempt to allocate a block from existing reservation
1073          *    failed; or
1074          * c) we come here with a goal and with a reservation window
1075          *
1076          * We do not need to allocate a new reservation window if we come here
1077          * at the beginning with a goal and the goal is inside the window, or
1078          * we don't have a goal but already have a reservation window.
1079          * then we could go to allocate from the reservation window directly.
1080          */
1081         while (1) {
1082                 if (rsv_is_empty(&my_rsv->rsv_window) || (ret < 0) ||
1083                         !goal_in_my_reservation(&my_rsv->rsv_window, goal, group, sb)) {
1084                         ret = alloc_new_reservation(my_rsv, goal, sb,
1085                                                         group, bitmap_bh);
1086                         if (ret < 0)
1087                                 break;                  /* failed */
1088
1089                         if (!goal_in_my_reservation(&my_rsv->rsv_window, goal, group, sb))
1090                                 goal = -1;
1091                 }
1092                 if ((my_rsv->rsv_start >= group_first_block + EXT3_BLOCKS_PER_GROUP(sb))
1093                     || (my_rsv->rsv_end < group_first_block))
1094                         BUG();
1095                 ret = ext3_try_to_allocate(sb, handle, group, bitmap_bh, goal,
1096                                            &my_rsv->rsv_window);
1097                 if (ret >= 0) {
1098                         my_rsv->rsv_alloc_hit++;
1099                         break;                          /* succeed */
1100                 }
1101         }
1102 out:
1103         if (ret >= 0) {
1104                 BUFFER_TRACE(bitmap_bh, "journal_dirty_metadata for "
1105                                         "bitmap block");
1106                 fatal = ext3_journal_dirty_metadata(handle, bitmap_bh);
1107                 if (fatal) {
1108                         *errp = fatal;
1109                         return -1;
1110                 }
1111                 return ret;
1112         }
1113
1114         BUFFER_TRACE(bitmap_bh, "journal_release_buffer");
1115         ext3_journal_release_buffer(handle, bitmap_bh);
1116         return ret;
1117 }
1118
1119 static int ext3_has_free_blocks(struct ext3_sb_info *sbi)
1120 {
1121         int free_blocks, root_blocks;
1122
1123         free_blocks = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
1124         root_blocks = le32_to_cpu(sbi->s_es->s_r_blocks_count);
1125         if (free_blocks < root_blocks + 1 && !capable(CAP_SYS_RESOURCE) &&
1126                 sbi->s_resuid != current->fsuid &&
1127                 (sbi->s_resgid == 0 || !in_group_p (sbi->s_resgid))) {
1128                 return 0;
1129         }
1130         return 1;
1131 }
1132
1133 /*
1134  * ext3_should_retry_alloc() is called when ENOSPC is returned, and if
1135  * it is profitable to retry the operation, this function will wait
1136  * for the current or commiting transaction to complete, and then
1137  * return TRUE.
1138  */
1139 int ext3_should_retry_alloc(struct super_block *sb, int *retries)
1140 {
1141         if (!ext3_has_free_blocks(EXT3_SB(sb)) || (*retries)++ > 3)
1142                 return 0;
1143
1144         jbd_debug(1, "%s: retrying operation after ENOSPC\n", sb->s_id);
1145
1146         return journal_force_commit_nested(EXT3_SB(sb)->s_journal);
1147 }
1148
1149 /*
1150  * ext3_new_block uses a goal block to assist allocation.  If the goal is
1151  * free, or there is a free block within 32 blocks of the goal, that block
1152  * is allocated.  Otherwise a forward search is made for a free block; within
1153  * each block group the search first looks for an entire free byte in the block
1154  * bitmap, and then for any free bit if that fails.
1155  * This function also updates quota and i_blocks field.
1156  */
1157 int ext3_new_block(handle_t *handle, struct inode *inode,
1158                         unsigned long goal, int *errp)
1159 {
1160         struct buffer_head *bitmap_bh = NULL;
1161         struct buffer_head *gdp_bh;
1162         int group_no;
1163         int goal_group;
1164         int ret_block;
1165         int bgi;                        /* blockgroup iteration index */
1166         int target_block;
1167         int fatal = 0, err;
1168         int performed_allocation = 0;
1169         int free_blocks;
1170         struct super_block *sb;
1171         struct ext3_group_desc *gdp;
1172         struct ext3_super_block *es;
1173         struct ext3_sb_info *sbi;
1174         struct ext3_reserve_window_node *my_rsv = NULL;
1175         struct ext3_block_alloc_info *block_i;
1176         unsigned short windowsz = 0;
1177 #ifdef EXT3FS_DEBUG
1178         static int goal_hits, goal_attempts;
1179 #endif
1180         unsigned long ngroups;
1181
1182         *errp = -ENOSPC;
1183         sb = inode->i_sb;
1184         if (!sb) {
1185                 printk("ext3_new_block: nonexistent device");
1186                 return 0;
1187         }
1188
1189         /*
1190          * Check quota for allocation of this block.
1191          */
1192         if (DQUOT_ALLOC_BLOCK(inode, 1)) {
1193                 *errp = -EDQUOT;
1194                 return 0;
1195         }
1196
1197         sbi = EXT3_SB(sb);
1198         es = EXT3_SB(sb)->s_es;
1199         ext3_debug("goal=%lu.\n", goal);
1200         /*
1201          * Allocate a block from reservation only when
1202          * filesystem is mounted with reservation(default,-o reservation), and
1203          * it's a regular file, and
1204          * the desired window size is greater than 0 (One could use ioctl
1205          * command EXT3_IOC_SETRSVSZ to set the window size to 0 to turn off
1206          * reservation on that particular file)
1207          */
1208         block_i = EXT3_I(inode)->i_block_alloc_info;
1209         if (block_i && ((windowsz = block_i->rsv_window_node.rsv_goal_size) > 0))
1210                 my_rsv = &block_i->rsv_window_node;
1211
1212         if (!ext3_has_free_blocks(sbi)) {
1213                 *errp = -ENOSPC;
1214                 goto out;
1215         }
1216
1217         /*
1218          * First, test whether the goal block is free.
1219          */
1220         if (goal < le32_to_cpu(es->s_first_data_block) ||
1221             goal >= le32_to_cpu(es->s_blocks_count))
1222                 goal = le32_to_cpu(es->s_first_data_block);
1223         group_no = (goal - le32_to_cpu(es->s_first_data_block)) /
1224                         EXT3_BLOCKS_PER_GROUP(sb);
1225         gdp = ext3_get_group_desc(sb, group_no, &gdp_bh);
1226         if (!gdp)
1227                 goto io_error;
1228
1229         goal_group = group_no;
1230 retry:
1231         free_blocks = le16_to_cpu(gdp->bg_free_blocks_count);
1232         /*
1233          * if there is not enough free blocks to make a new resevation
1234          * turn off reservation for this allocation
1235          */
1236         if (my_rsv && (free_blocks < windowsz)
1237                 && (rsv_is_empty(&my_rsv->rsv_window)))
1238                 my_rsv = NULL;
1239
1240         if (free_blocks > 0) {
1241                 ret_block = ((goal - le32_to_cpu(es->s_first_data_block)) %
1242                                 EXT3_BLOCKS_PER_GROUP(sb));
1243                 bitmap_bh = read_block_bitmap(sb, group_no);
1244                 if (!bitmap_bh)
1245                         goto io_error;
1246                 ret_block = ext3_try_to_allocate_with_rsv(sb, handle, group_no,
1247                                         bitmap_bh, ret_block, my_rsv, &fatal);
1248                 if (fatal)
1249                         goto out;
1250                 if (ret_block >= 0)
1251                         goto allocated;
1252         }
1253
1254         ngroups = EXT3_SB(sb)->s_groups_count;
1255         smp_rmb();
1256
1257         /*
1258          * Now search the rest of the groups.  We assume that
1259          * i and gdp correctly point to the last group visited.
1260          */
1261         for (bgi = 0; bgi < ngroups; bgi++) {
1262                 group_no++;
1263                 if (group_no >= ngroups)
1264                         group_no = 0;
1265                 gdp = ext3_get_group_desc(sb, group_no, &gdp_bh);
1266                 if (!gdp) {
1267                         *errp = -EIO;
1268                         goto out;
1269                 }
1270                 free_blocks = le16_to_cpu(gdp->bg_free_blocks_count);
1271                 /*
1272                  * skip this group if the number of
1273                  * free blocks is less than half of the reservation
1274                  * window size.
1275                  */
1276                 if (free_blocks <= (windowsz/2))
1277                         continue;
1278
1279                 brelse(bitmap_bh);
1280                 bitmap_bh = read_block_bitmap(sb, group_no);
1281                 if (!bitmap_bh)
1282                         goto io_error;
1283                 ret_block = ext3_try_to_allocate_with_rsv(sb, handle, group_no,
1284                                         bitmap_bh, -1, my_rsv, &fatal);
1285                 if (fatal)
1286                         goto out;
1287                 if (ret_block >= 0)
1288                         goto allocated;
1289         }
1290         /*
1291          * We may end up a bogus ealier ENOSPC error due to
1292          * filesystem is "full" of reservations, but
1293          * there maybe indeed free blocks avaliable on disk
1294          * In this case, we just forget about the reservations
1295          * just do block allocation as without reservations.
1296          */
1297         if (my_rsv) {
1298                 my_rsv = NULL;
1299                 group_no = goal_group;
1300                 goto retry;
1301         }
1302         /* No space left on the device */
1303         *errp = -ENOSPC;
1304         goto out;
1305
1306 allocated:
1307
1308         ext3_debug("using block group %d(%d)\n",
1309                         group_no, gdp->bg_free_blocks_count);
1310
1311         BUFFER_TRACE(gdp_bh, "get_write_access");
1312         fatal = ext3_journal_get_write_access(handle, gdp_bh);
1313         if (fatal)
1314                 goto out;
1315
1316         target_block = ret_block + group_no * EXT3_BLOCKS_PER_GROUP(sb)
1317                                 + le32_to_cpu(es->s_first_data_block);
1318
1319         if (target_block == le32_to_cpu(gdp->bg_block_bitmap) ||
1320             target_block == le32_to_cpu(gdp->bg_inode_bitmap) ||
1321             in_range(target_block, le32_to_cpu(gdp->bg_inode_table),
1322                       EXT3_SB(sb)->s_itb_per_group))
1323                 ext3_error(sb, "ext3_new_block",
1324                             "Allocating block in system zone - "
1325                             "block = %u", target_block);
1326
1327         performed_allocation = 1;
1328
1329 #ifdef CONFIG_JBD_DEBUG
1330         {
1331                 struct buffer_head *debug_bh;
1332
1333                 /* Record bitmap buffer state in the newly allocated block */
1334                 debug_bh = sb_find_get_block(sb, target_block);
1335                 if (debug_bh) {
1336                         BUFFER_TRACE(debug_bh, "state when allocated");
1337                         BUFFER_TRACE2(debug_bh, bitmap_bh, "bitmap state");
1338                         brelse(debug_bh);
1339                 }
1340         }
1341         jbd_lock_bh_state(bitmap_bh);
1342         spin_lock(sb_bgl_lock(sbi, group_no));
1343         if (buffer_jbd(bitmap_bh) && bh2jh(bitmap_bh)->b_committed_data) {
1344                 if (ext3_test_bit(ret_block,
1345                                 bh2jh(bitmap_bh)->b_committed_data)) {
1346                         printk("%s: block was unexpectedly set in "
1347                                 "b_committed_data\n", __FUNCTION__);
1348                 }
1349         }
1350         ext3_debug("found bit %d\n", ret_block);
1351         spin_unlock(sb_bgl_lock(sbi, group_no));
1352         jbd_unlock_bh_state(bitmap_bh);
1353 #endif
1354
1355         /* ret_block was blockgroup-relative.  Now it becomes fs-relative */
1356         ret_block = target_block;
1357
1358         if (ret_block >= le32_to_cpu(es->s_blocks_count)) {
1359                 ext3_error(sb, "ext3_new_block",
1360                             "block(%d) >= blocks count(%d) - "
1361                             "block_group = %d, es == %p ", ret_block,
1362                         le32_to_cpu(es->s_blocks_count), group_no, es);
1363                 goto out;
1364         }
1365
1366         /*
1367          * It is up to the caller to add the new buffer to a journal
1368          * list of some description.  We don't know in advance whether
1369          * the caller wants to use it as metadata or data.
1370          */
1371         ext3_debug("allocating block %d. Goal hits %d of %d.\n",
1372                         ret_block, goal_hits, goal_attempts);
1373
1374         spin_lock(sb_bgl_lock(sbi, group_no));
1375         gdp->bg_free_blocks_count =
1376                         cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count) - 1);
1377         spin_unlock(sb_bgl_lock(sbi, group_no));
1378         percpu_counter_mod(&sbi->s_freeblocks_counter, -1);
1379
1380         BUFFER_TRACE(gdp_bh, "journal_dirty_metadata for group descriptor");
1381         err = ext3_journal_dirty_metadata(handle, gdp_bh);
1382         if (!fatal)
1383                 fatal = err;
1384
1385         sb->s_dirt = 1;
1386         if (fatal)
1387                 goto out;
1388
1389         *errp = 0;
1390         brelse(bitmap_bh);
1391         return ret_block;
1392
1393 io_error:
1394         *errp = -EIO;
1395 out:
1396         if (fatal) {
1397                 *errp = fatal;
1398                 ext3_std_error(sb, fatal);
1399         }
1400         /*
1401          * Undo the block allocation
1402          */
1403         if (!performed_allocation)
1404                 DQUOT_FREE_BLOCK(inode, 1);
1405         brelse(bitmap_bh);
1406         return 0;
1407 }
1408
1409 unsigned long ext3_count_free_blocks(struct super_block *sb)
1410 {
1411         unsigned long desc_count;
1412         struct ext3_group_desc *gdp;
1413         int i;
1414         unsigned long ngroups = EXT3_SB(sb)->s_groups_count;
1415 #ifdef EXT3FS_DEBUG
1416         struct ext3_super_block *es;
1417         unsigned long bitmap_count, x;
1418         struct buffer_head *bitmap_bh = NULL;
1419
1420         es = EXT3_SB(sb)->s_es;
1421         desc_count = 0;
1422         bitmap_count = 0;
1423         gdp = NULL;
1424
1425         smp_rmb();
1426         for (i = 0; i < ngroups; i++) {
1427                 gdp = ext3_get_group_desc(sb, i, NULL);
1428                 if (!gdp)
1429                         continue;
1430                 desc_count += le16_to_cpu(gdp->bg_free_blocks_count);
1431                 brelse(bitmap_bh);
1432                 bitmap_bh = read_block_bitmap(sb, i);
1433                 if (bitmap_bh == NULL)
1434                         continue;
1435
1436                 x = ext3_count_free(bitmap_bh, sb->s_blocksize);
1437                 printk("group %d: stored = %d, counted = %lu\n",
1438                         i, le16_to_cpu(gdp->bg_free_blocks_count), x);
1439                 bitmap_count += x;
1440         }
1441         brelse(bitmap_bh);
1442         printk("ext3_count_free_blocks: stored = %u, computed = %lu, %lu\n",
1443                le32_to_cpu(es->s_free_blocks_count), desc_count, bitmap_count);
1444         return bitmap_count;
1445 #else
1446         desc_count = 0;
1447         smp_rmb();
1448         for (i = 0; i < ngroups; i++) {
1449                 gdp = ext3_get_group_desc(sb, i, NULL);
1450                 if (!gdp)
1451                         continue;
1452                 desc_count += le16_to_cpu(gdp->bg_free_blocks_count);
1453         }
1454
1455         return desc_count;
1456 #endif
1457 }
1458
1459 static inline int
1460 block_in_use(unsigned long block, struct super_block *sb, unsigned char *map)
1461 {
1462         return ext3_test_bit ((block -
1463                 le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block)) %
1464                          EXT3_BLOCKS_PER_GROUP(sb), map);
1465 }
1466
1467 static inline int test_root(int a, int b)
1468 {
1469         int num = b;
1470
1471         while (a > num)
1472                 num *= b;
1473         return num == a;
1474 }
1475
1476 static int ext3_group_sparse(int group)
1477 {
1478         if (group <= 1)
1479                 return 1;
1480         if (!(group & 1))
1481                 return 0;
1482         return (test_root(group, 7) || test_root(group, 5) ||
1483                 test_root(group, 3));
1484 }
1485
1486 /**
1487  *      ext3_bg_has_super - number of blocks used by the superblock in group
1488  *      @sb: superblock for filesystem
1489  *      @group: group number to check
1490  *
1491  *      Return the number of blocks used by the superblock (primary or backup)
1492  *      in this group.  Currently this will be only 0 or 1.
1493  */
1494 int ext3_bg_has_super(struct super_block *sb, int group)
1495 {
1496         if (EXT3_HAS_RO_COMPAT_FEATURE(sb,
1497                                 EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER) &&
1498                         !ext3_group_sparse(group))
1499                 return 0;
1500         return 1;
1501 }
1502
1503 static unsigned long ext3_bg_num_gdb_meta(struct super_block *sb, int group)
1504 {
1505         unsigned long metagroup = group / EXT3_DESC_PER_BLOCK(sb);
1506         unsigned long first = metagroup * EXT3_DESC_PER_BLOCK(sb);
1507         unsigned long last = first + EXT3_DESC_PER_BLOCK(sb) - 1;
1508
1509         if (group == first || group == first + 1 || group == last)
1510                 return 1;
1511         return 0;
1512 }
1513
1514 static unsigned long ext3_bg_num_gdb_nometa(struct super_block *sb, int group)
1515 {
1516         if (EXT3_HAS_RO_COMPAT_FEATURE(sb,
1517                                 EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER) &&
1518                         !ext3_group_sparse(group))
1519                 return 0;
1520         return EXT3_SB(sb)->s_gdb_count;
1521 }
1522
1523 /**
1524  *      ext3_bg_num_gdb - number of blocks used by the group table in group
1525  *      @sb: superblock for filesystem
1526  *      @group: group number to check
1527  *
1528  *      Return the number of blocks used by the group descriptor table
1529  *      (primary or backup) in this group.  In the future there may be a
1530  *      different number of descriptor blocks in each group.
1531  */
1532 unsigned long ext3_bg_num_gdb(struct super_block *sb, int group)
1533 {
1534         unsigned long first_meta_bg =
1535                         le32_to_cpu(EXT3_SB(sb)->s_es->s_first_meta_bg);
1536         unsigned long metagroup = group / EXT3_DESC_PER_BLOCK(sb);
1537
1538         if (!EXT3_HAS_INCOMPAT_FEATURE(sb,EXT3_FEATURE_INCOMPAT_META_BG) ||
1539                         metagroup < first_meta_bg)
1540                 return ext3_bg_num_gdb_nometa(sb,group);
1541
1542         return ext3_bg_num_gdb_meta(sb,group);
1543
1544 }