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