fs/ext3/balloc.c

   1 /*
   2  *  linux/fs/ext3/balloc.c
   3  *
   4  * Copyright (C) 1992, 1993, 1994, 1995
   5  * Remy Card (card@masi.ibp.fr)
   6  * Laboratoire MASI - Institut Blaise Pascal
   7  * Universite Pierre et Marie Curie (Paris VI)
   8  *
   9  *  Enhanced block allocation by Stephen Tweedie (sct@redhat.com), 1993
  10  *  Big-endian to little-endian byte-swapping/bitmaps by
  11  *        David S. Miller (davem@caip.rutgers.edu), 1995
  12  */
  13
  14 #include <linux/time.h>
  15 #include <linux/capability.h>
  16 #include <linux/fs.h>
  17 #include <linux/jbd.h>
  18 #include <linux/ext3_fs.h>
  19 #include <linux/ext3_jbd.h>
  20 #include <linux/quotaops.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/vs_dlimit.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, ext3_grpblk_t grp_goal,
 167                         unsigned int group, struct super_block * sb)
 168 {
 169         ext3_fsblk_t group_first_block, group_last_block;
 170
 171         group_first_block = ext3_group_first_block_no(sb, group);
 172         group_last_block = group_first_block + (EXT3_BLOCKS_PER_GROUP(sb) - 1);
 173
 174         if ((rsv->_rsv_start > group_last_block) ||
 175             (rsv->_rsv_end < group_first_block))
 176                 return 0;
 177         if ((grp_goal >= 0) && ((grp_goal + group_first_block < rsv->_rsv_start)
 178                 || (grp_goal + group_first_block > rsv->_rsv_end)))
 179                 return 0;
 180         return 1;
 181 }
 182
 183 /*
 184  * Find the reserved window which includes the goal, or the previous one
 185  * if the goal is not in any window.
 186  * Returns NULL if there are no windows or if all windows start after the goal.
 187  */
 188 static struct ext3_reserve_window_node *
 189 search_reserve_window(struct rb_root *root, ext3_fsblk_t goal)
 190 {
 191         struct rb_node *n = root->rb_node;
 192         struct ext3_reserve_window_node *rsv;
 193
 194         if (!n)
 195                 return NULL;
 196
 197         do {
 198                 rsv = rb_entry(n, struct ext3_reserve_window_node, rsv_node);
 199
 200                 if (goal < rsv->rsv_start)
 201                         n = n->rb_left;
 202                 else if (goal > rsv->rsv_end)
 203                         n = n->rb_right;
 204                 else
 205                         return rsv;
 206         } while (n);
 207         /*
 208          * We've fallen off the end of the tree: the goal wasn't inside
 209          * any particular node.  OK, the previous node must be to one
 210          * side of the interval containing the goal.  If it's the RHS,
 211          * we need to back up one.
 212          */
 213         if (rsv->rsv_start > goal) {
 214                 n = rb_prev(&rsv->rsv_node);
 215                 rsv = rb_entry(n, struct ext3_reserve_window_node, rsv_node);
 216         }
 217         return rsv;
 218 }
 219
 220 void ext3_rsv_window_add(struct super_block *sb,
 221                     struct ext3_reserve_window_node *rsv)
 222 {
 223         struct rb_root *root = &EXT3_SB(sb)->s_rsv_window_root;
 224         struct rb_node *node = &rsv->rsv_node;
 225         ext3_fsblk_t start = rsv->rsv_start;
 226
 227         struct rb_node ** p = &root->rb_node;
 228         struct rb_node * parent = NULL;
 229         struct ext3_reserve_window_node *this;
 230
 231         while (*p)
 232         {
 233                 parent = *p;
 234                 this = rb_entry(parent, struct ext3_reserve_window_node, rsv_node);
 235
 236                 if (start < this->rsv_start)
 237                         p = &(*p)->rb_left;
 238                 else if (start > this->rsv_end)
 239                         p = &(*p)->rb_right;
 240                 else
 241                         BUG();
 242         }
 243
 244         rb_link_node(node, parent, p);
 245         rb_insert_color(node, root);
 246 }
 247
 248 static void rsv_window_remove(struct super_block *sb,
 249                               struct ext3_reserve_window_node *rsv)
 250 {
 251         rsv->rsv_start = EXT3_RESERVE_WINDOW_NOT_ALLOCATED;
 252         rsv->rsv_end = EXT3_RESERVE_WINDOW_NOT_ALLOCATED;
 253         rsv->rsv_alloc_hit = 0;
 254         rb_erase(&rsv->rsv_node, &EXT3_SB(sb)->s_rsv_window_root);
 255 }
 256
 257 static inline int rsv_is_empty(struct ext3_reserve_window *rsv)
 258 {
 259         /* a valid reservation end block could not be 0 */
 260         return (rsv->_rsv_end == EXT3_RESERVE_WINDOW_NOT_ALLOCATED);
 261 }
 262 void ext3_init_block_alloc_info(struct inode *inode)
 263 {
 264         struct ext3_inode_info *ei = EXT3_I(inode);
 265         struct ext3_block_alloc_info *block_i = ei->i_block_alloc_info;
 266         struct super_block *sb = inode->i_sb;
 267
 268         block_i = kmalloc(sizeof(*block_i), GFP_NOFS);
 269         if (block_i) {
 270                 struct ext3_reserve_window_node *rsv = &block_i->rsv_window_node;
 271
 272                 rsv->rsv_start = EXT3_RESERVE_WINDOW_NOT_ALLOCATED;
 273                 rsv->rsv_end = EXT3_RESERVE_WINDOW_NOT_ALLOCATED;
 274
 275                 /*
 276                  * if filesystem is mounted with NORESERVATION, the goal
 277                  * reservation window size is set to zero to indicate
 278                  * block reservation is off
 279                  */
 280                 if (!test_opt(sb, RESERVATION))
 281                         rsv->rsv_goal_size = 0;
 282                 else
 283                         rsv->rsv_goal_size = EXT3_DEFAULT_RESERVE_BLOCKS;
 284                 rsv->rsv_alloc_hit = 0;
 285                 block_i->last_alloc_logical_block = 0;
 286                 block_i->last_alloc_physical_block = 0;
 287         }
 288         ei->i_block_alloc_info = block_i;
 289 }
 290
 291 void ext3_discard_reservation(struct inode *inode)
 292 {
 293         struct ext3_inode_info *ei = EXT3_I(inode);
 294         struct ext3_block_alloc_info *block_i = ei->i_block_alloc_info;
 295         struct ext3_reserve_window_node *rsv;
 296         spinlock_t *rsv_lock = &EXT3_SB(inode->i_sb)->s_rsv_window_lock;
 297
 298         if (!block_i)
 299                 return;
 300
 301         rsv = &block_i->rsv_window_node;
 302         if (!rsv_is_empty(&rsv->rsv_window)) {
 303                 spin_lock(rsv_lock);
 304                 if (!rsv_is_empty(&rsv->rsv_window))
 305                         rsv_window_remove(inode->i_sb, rsv);
 306                 spin_unlock(rsv_lock);
 307         }
 308 }
 309
 310 /* Free given blocks, update quota and i_blocks field */
 311 void ext3_free_blocks_sb(handle_t *handle, struct super_block *sb,
 312                          ext3_fsblk_t block, unsigned long count,
 313                          unsigned long *pdquot_freed_blocks)
 314 {
 315         struct buffer_head *bitmap_bh = NULL;
 316         struct buffer_head *gd_bh;
 317         unsigned long block_group;
 318         ext3_grpblk_t bit;
 319         unsigned long i;
 320         unsigned long overflow;
 321         struct ext3_group_desc * desc;
 322         struct ext3_super_block * es;
 323         struct ext3_sb_info *sbi;
 324         int err = 0, ret;
 325         ext3_grpblk_t group_freed;
 326
 327         *pdquot_freed_blocks = 0;
 328         sbi = EXT3_SB(sb);
 329         es = sbi->s_es;
 330         if (block < le32_to_cpu(es->s_first_data_block) ||
 331             block + count < block ||
 332             block + count > le32_to_cpu(es->s_blocks_count)) {
 333                 ext3_error (sb, "ext3_free_blocks",
 334                             "Freeing blocks not in datazone - "
 335                             "block = "E3FSBLK", count = %lu", block, count);
 336                 goto error_return;
 337         }
 338
 339         ext3_debug ("freeing block(s) %lu-%lu\n", block, block + count - 1);
 340
 341 do_more:
 342         overflow = 0;
 343         block_group = (block - le32_to_cpu(es->s_first_data_block)) /
 344                       EXT3_BLOCKS_PER_GROUP(sb);
 345         bit = (block - le32_to_cpu(es->s_first_data_block)) %
 346                       EXT3_BLOCKS_PER_GROUP(sb);
 347         /*
 348          * Check to see if we are freeing blocks across a group
 349          * boundary.
 350          */
 351         if (bit + count > EXT3_BLOCKS_PER_GROUP(sb)) {
 352                 overflow = bit + count - EXT3_BLOCKS_PER_GROUP(sb);
 353                 count -= overflow;
 354         }
 355         brelse(bitmap_bh);
 356         bitmap_bh = read_block_bitmap(sb, block_group);
 357         if (!bitmap_bh)
 358                 goto error_return;
 359         desc = ext3_get_group_desc (sb, block_group, &gd_bh);
 360         if (!desc)
 361                 goto error_return;
 362
 363         if (in_range (le32_to_cpu(desc->bg_block_bitmap), block, count) ||
 364             in_range (le32_to_cpu(desc->bg_inode_bitmap), block, count) ||
 365             in_range (block, le32_to_cpu(desc->bg_inode_table),
 366                       sbi->s_itb_per_group) ||
 367             in_range (block + count - 1, le32_to_cpu(desc->bg_inode_table),
 368                       sbi->s_itb_per_group))
 369                 ext3_error (sb, "ext3_free_blocks",
 370                             "Freeing blocks in system zones - "
 371                             "Block = "E3FSBLK", count = %lu",
 372                             block, count);
 373
 374         /*
 375          * We are about to start releasing blocks in the bitmap,
 376          * so we need undo access.
 377          */
 378         /* @@@ check errors */
 379         BUFFER_TRACE(bitmap_bh, "getting undo access");
 380         err = ext3_journal_get_undo_access(handle, bitmap_bh);
 381         if (err)
 382                 goto error_return;
 383
 384         /*
 385          * We are about to modify some metadata.  Call the journal APIs
 386          * to unshare ->b_data if a currently-committing transaction is
 387          * using it
 388          */
 389         BUFFER_TRACE(gd_bh, "get_write_access");
 390         err = ext3_journal_get_write_access(handle, gd_bh);
 391         if (err)
 392                 goto error_return;
 393
 394         jbd_lock_bh_state(bitmap_bh);
 395
 396         for (i = 0, group_freed = 0; i < count; i++) {
 397                 /*
 398                  * An HJ special.  This is expensive...
 399                  */
 400 #ifdef CONFIG_JBD_DEBUG
 401                 jbd_unlock_bh_state(bitmap_bh);
 402                 {
 403                         struct buffer_head *debug_bh;
 404                         debug_bh = sb_find_get_block(sb, block + i);
 405                         if (debug_bh) {
 406                                 BUFFER_TRACE(debug_bh, "Deleted!");
 407                                 if (!bh2jh(bitmap_bh)->b_committed_data)
 408                                         BUFFER_TRACE(debug_bh,
 409                                                 "No commited data in bitmap");
 410                                 BUFFER_TRACE2(debug_bh, bitmap_bh, "bitmap");
 411                                 __brelse(debug_bh);
 412                         }
 413                 }
 414                 jbd_lock_bh_state(bitmap_bh);
 415 #endif
 416                 if (need_resched()) {
 417                         jbd_unlock_bh_state(bitmap_bh);
 418                         cond_resched();
 419                         jbd_lock_bh_state(bitmap_bh);
 420                 }
 421                 /* @@@ This prevents newly-allocated data from being
 422                  * freed and then reallocated within the same
 423                  * transaction.
 424                  *
 425                  * Ideally we would want to allow that to happen, but to
 426                  * do so requires making journal_forget() capable of
 427                  * revoking the queued write of a data block, which
 428                  * implies blocking on the journal lock.  *forget()
 429                  * cannot block due to truncate races.
 430                  *
 431                  * Eventually we can fix this by making journal_forget()
 432                  * return a status indicating whether or not it was able
 433                  * to revoke the buffer.  On successful revoke, it is
 434                  * safe not to set the allocation bit in the committed
 435                  * bitmap, because we know that there is no outstanding
 436                  * activity on the buffer any more and so it is safe to
 437                  * reallocate it.
 438                  */
 439                 BUFFER_TRACE(bitmap_bh, "set in b_committed_data");
 440                 J_ASSERT_BH(bitmap_bh,
 441                                 bh2jh(bitmap_bh)->b_committed_data != NULL);
 442                 ext3_set_bit_atomic(sb_bgl_lock(sbi, block_group), bit + i,
 443                                 bh2jh(bitmap_bh)->b_committed_data);
 444
 445                 /*
 446                  * We clear the bit in the bitmap after setting the committed
 447                  * data bit, because this is the reverse order to that which
 448                  * the allocator uses.
 449                  */
 450                 BUFFER_TRACE(bitmap_bh, "clear bit");
 451                 if (!ext3_clear_bit_atomic(sb_bgl_lock(sbi, block_group),
 452                                                 bit + i, bitmap_bh->b_data)) {
 453                         jbd_unlock_bh_state(bitmap_bh);
 454                         ext3_error(sb, __FUNCTION__,
 455                                 "bit already cleared for block "E3FSBLK,
 456                                  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                         ext3_fsblk_t block, unsigned long count)
 497 {
 498         struct super_block * sb;
 499         unsigned long 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                 DLIMIT_FREE_BLOCK(inode, dquot_freed_blocks);
 509                 DQUOT_FREE_BLOCK(inode, dquot_freed_blocks);
 510         }
 511         return;
 512 }
 513
 514 /*
 515  * For ext3 allocations, we must not reuse any blocks which are
 516  * allocated in the bitmap buffer's "last committed data" copy.  This
 517  * prevents deletes from freeing up the page for reuse until we have
 518  * committed the delete transaction.
 519  *
 520  * If we didn't do this, then deleting something and reallocating it as
 521  * data would allow the old block to be overwritten before the
 522  * transaction committed (because we force data to disk before commit).
 523  * This would lead to corruption if we crashed between overwriting the
 524  * data and committing the delete.
 525  *
 526  * @@@ We may want to make this allocation behaviour conditional on
 527  * data-writes at some point, and disable it for metadata allocations or
 528  * sync-data inodes.
 529  */
 530 static int ext3_test_allocatable(ext3_grpblk_t nr, struct buffer_head *bh)
 531 {
 532         int ret;
 533         struct journal_head *jh = bh2jh(bh);
 534
 535         if (ext3_test_bit(nr, bh->b_data))
 536                 return 0;
 537
 538         jbd_lock_bh_state(bh);
 539         if (!jh->b_committed_data)
 540                 ret = 1;
 541         else
 542                 ret = !ext3_test_bit(nr, jh->b_committed_data);
 543         jbd_unlock_bh_state(bh);
 544         return ret;
 545 }
 546
 547 static ext3_grpblk_t
 548 bitmap_search_next_usable_block(ext3_grpblk_t start, struct buffer_head *bh,
 549                                         ext3_grpblk_t maxblocks)
 550 {
 551         ext3_grpblk_t next;
 552         struct journal_head *jh = bh2jh(bh);
 553
 554         /*
 555          * The bitmap search --- search forward alternately through the actual
 556          * bitmap and the last-committed copy until we find a bit free in
 557          * both
 558          */
 559         while (start < maxblocks) {
 560                 next = ext3_find_next_zero_bit(bh->b_data, maxblocks, start);
 561                 if (next >= maxblocks)
 562                         return -1;
 563                 if (ext3_test_allocatable(next, bh))
 564                         return next;
 565                 jbd_lock_bh_state(bh);
 566                 if (jh->b_committed_data)
 567                         start = ext3_find_next_zero_bit(jh->b_committed_data,
 568                                                         maxblocks, next);
 569                 jbd_unlock_bh_state(bh);
 570         }
 571         return -1;
 572 }
 573
 574 /*
 575  * Find an allocatable block in a bitmap.  We honour both the bitmap and
 576  * its last-committed copy (if that exists), and perform the "most
 577  * appropriate allocation" algorithm of looking for a free block near
 578  * the initial goal; then for a free byte somewhere in the bitmap; then
 579  * for any free bit in the bitmap.
 580  */
 581 static ext3_grpblk_t
 582 find_next_usable_block(ext3_grpblk_t start, struct buffer_head *bh,
 583                         ext3_grpblk_t maxblocks)
 584 {
 585         ext3_grpblk_t 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                 ext3_grpblk_t 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, ext3_grpblk_t 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 ext3_grpblk_t
 658 ext3_try_to_allocate(struct super_block *sb, handle_t *handle, int group,
 659                         struct buffer_head *bitmap_bh, ext3_grpblk_t grp_goal,
 660                         unsigned long *count, struct ext3_reserve_window *my_rsv)
 661 {
 662         ext3_fsblk_t group_first_block;
 663         ext3_grpblk_t start, end;
 664         unsigned long num = 0;
 665
 666         /* we do allocation within the reservation window if we have a window */
 667         if (my_rsv) {
 668                 group_first_block = ext3_group_first_block_no(sb, group);
 669                 if (my_rsv->_rsv_start >= group_first_block)
 670                         start = my_rsv->_rsv_start - group_first_block;
 671                 else
 672                         /* reservation window cross group boundary */
 673                         start = 0;
 674                 end = my_rsv->_rsv_end - group_first_block + 1;
 675                 if (end > EXT3_BLOCKS_PER_GROUP(sb))
 676                         /* reservation window crosses group boundary */
 677                         end = EXT3_BLOCKS_PER_GROUP(sb);
 678                 if ((start <= grp_goal) && (grp_goal < end))
 679                         start = grp_goal;
 680                 else
 681                         grp_goal = -1;
 682         } else {
 683                 if (grp_goal > 0)
 684                         start = grp_goal;
 685                 else
 686                         start = 0;
 687                 end = EXT3_BLOCKS_PER_GROUP(sb);
 688         }
 689
 690         BUG_ON(start > EXT3_BLOCKS_PER_GROUP(sb));
 691
 692 repeat:
 693         if (grp_goal < 0 || !ext3_test_allocatable(grp_goal, bitmap_bh)) {
 694                 grp_goal = find_next_usable_block(start, bitmap_bh, end);
 695                 if (grp_goal < 0)
 696                         goto fail_access;
 697                 if (!my_rsv) {
 698                         int i;
 699
 700                         for (i = 0; i < 7 && grp_goal > start &&
 701                                         ext3_test_allocatable(grp_goal - 1,
 702                                                                 bitmap_bh);
 703                                         i++, grp_goal--)
 704                                 ;
 705                 }
 706         }
 707         start = grp_goal;
 708
 709         if (!claim_block(sb_bgl_lock(EXT3_SB(sb), group), grp_goal, bitmap_bh)) {
 710                 /*
 711                  * The block was allocated by another thread, or it was
 712                  * allocated and then freed by another thread
 713                  */
 714                 start++;
 715                 grp_goal++;
 716                 if (start >= end)
 717                         goto fail_access;
 718                 goto repeat;
 719         }
 720         num++;
 721         grp_goal++;
 722         while (num < *count && grp_goal < end
 723                 && ext3_test_allocatable(grp_goal, bitmap_bh)
 724                 && claim_block(sb_bgl_lock(EXT3_SB(sb), group), grp_goal, bitmap_bh)) {
 725                 num++;
 726                 grp_goal++;
 727         }
 728         *count = num;
 729         return grp_goal - num;
 730 fail_access:
 731         *count = num;
 732         return -1;
 733 }
 734
 735 /**
 736  *      find_next_reservable_window():
 737  *              find a reservable space within the given range.
 738  *              It does not allocate the reservation window for now:
 739  *              alloc_new_reservation() will do the work later.
 740  *
 741  *      @search_head: the head of the searching list;
 742  *              This is not necessarily the list head of the whole filesystem
 743  *
 744  *              We have both head and start_block to assist the search
 745  *              for the reservable space. The list starts from head,
 746  *              but we will shift to the place where start_block is,
 747  *              then start from there, when looking for a reservable space.
 748  *
 749  *      @size: the target new reservation window size
 750  *
 751  *      @group_first_block: the first block we consider to start
 752  *                      the real search from
 753  *
 754  *      @last_block:
 755  *              the maximum block number that our goal reservable space
 756  *              could start from. This is normally the last block in this
 757  *              group. The search will end when we found the start of next
 758  *              possible reservable space is out of this boundary.
 759  *              This could handle the cross boundary reservation window
 760  *              request.
 761  *
 762  *      basically we search from the given range, rather than the whole
 763  *      reservation double linked list, (start_block, last_block)
 764  *      to find a free region that is of my size and has not
 765  *      been reserved.
 766  *
 767  */
 768 static int find_next_reservable_window(
 769                                 struct ext3_reserve_window_node *search_head,
 770                                 struct ext3_reserve_window_node *my_rsv,
 771                                 struct super_block * sb,
 772                                 ext3_fsblk_t start_block,
 773                                 ext3_fsblk_t last_block)
 774 {
 775         struct rb_node *next;
 776         struct ext3_reserve_window_node *rsv, *prev;
 777         ext3_fsblk_t cur;
 778         int size = my_rsv->rsv_goal_size;
 779
 780         /* TODO: make the start of the reservation window byte-aligned */
 781         /* cur = *start_block & ~7;*/
 782         cur = start_block;
 783         rsv = search_head;
 784         if (!rsv)
 785                 return -1;
 786
 787         while (1) {
 788                 if (cur <= rsv->rsv_end)
 789                         cur = rsv->rsv_end + 1;
 790
 791                 /* TODO?
 792                  * in the case we could not find a reservable space
 793                  * that is what is expected, during the re-search, we could
 794                  * remember what's the largest reservable space we could have
 795                  * and return that one.
 796                  *
 797                  * For now it will fail if we could not find the reservable
 798                  * space with expected-size (or more)...
 799                  */
 800                 if (cur > last_block)
 801                         return -1;              /* fail */
 802
 803                 prev = rsv;
 804                 next = rb_next(&rsv->rsv_node);
 805                 rsv = list_entry(next,struct ext3_reserve_window_node,rsv_node);
 806
 807                 /*
 808                  * Reached the last reservation, we can just append to the
 809                  * previous one.
 810                  */
 811                 if (!next)
 812                         break;
 813
 814                 if (cur + size <= rsv->rsv_start) {
 815                         /*
 816                          * Found a reserveable space big enough.  We could
 817                          * have a reservation across the group boundary here
 818                          */
 819                         break;
 820                 }
 821         }
 822         /*
 823          * we come here either :
 824          * when we reach the end of the whole list,
 825          * and there is empty reservable space after last entry in the list.
 826          * append it to the end of the list.
 827          *
 828          * or we found one reservable space in the middle of the list,
 829          * return the reservation window that we could append to.
 830          * succeed.
 831          */
 832
 833         if ((prev != my_rsv) && (!rsv_is_empty(&my_rsv->rsv_window)))
 834                 rsv_window_remove(sb, my_rsv);
 835
 836         /*
 837          * Let's book the whole avaliable window for now.  We will check the
 838          * disk bitmap later and then, if there are free blocks then we adjust
 839          * the window size if it's larger than requested.
 840          * Otherwise, we will remove this node from the tree next time
 841          * call find_next_reservable_window.
 842          */
 843         my_rsv->rsv_start = cur;
 844         my_rsv->rsv_end = cur + size - 1;
 845         my_rsv->rsv_alloc_hit = 0;
 846
 847         if (prev != my_rsv)
 848                 ext3_rsv_window_add(sb, my_rsv);
 849
 850         return 0;
 851 }
 852
 853 /**
 854  *      alloc_new_reservation()--allocate a new reservation window
 855  *
 856  *              To make a new reservation, we search part of the filesystem
 857  *              reservation list (the list that inside the group). We try to
 858  *              allocate a new reservation window near the allocation goal,
 859  *              or the beginning of the group, if there is no goal.
 860  *
 861  *              We first find a reservable space after the goal, then from
 862  *              there, we check the bitmap for the first free block after
 863  *              it. If there is no free block until the end of group, then the
 864  *              whole group is full, we failed. Otherwise, check if the free
 865  *              block is inside the expected reservable space, if so, we
 866  *              succeed.
 867  *              If the first free block is outside the reservable space, then
 868  *              start from the first free block, we search for next available
 869  *              space, and go on.
 870  *
 871  *      on succeed, a new reservation will be found and inserted into the list
 872  *      It contains at least one free block, and it does not overlap with other
 873  *      reservation windows.
 874  *
 875  *      failed: we failed to find a reservation window in this group
 876  *
 877  *      @rsv: the reservation
 878  *
 879  *      @grp_goal: The goal (group-relative).  It is where the search for a
 880  *              free reservable space should start from.
 881  *              if we have a grp_goal(grp_goal >0 ), then start from there,
 882  *              no grp_goal(grp_goal = -1), we start from the first block
 883  *              of the group.
 884  *
 885  *      @sb: the super block
 886  *      @group: the group we are trying to allocate in
 887  *      @bitmap_bh: the block group block bitmap
 888  *
 889  */
 890 static int alloc_new_reservation(struct ext3_reserve_window_node *my_rsv,
 891                 ext3_grpblk_t grp_goal, struct super_block *sb,
 892                 unsigned int group, struct buffer_head *bitmap_bh)
 893 {
 894         struct ext3_reserve_window_node *search_head;
 895         ext3_fsblk_t group_first_block, group_end_block, start_block;
 896         ext3_grpblk_t first_free_block;
 897         struct rb_root *fs_rsv_root = &EXT3_SB(sb)->s_rsv_window_root;
 898         unsigned long size;
 899         int ret;
 900         spinlock_t *rsv_lock = &EXT3_SB(sb)->s_rsv_window_lock;
 901
 902         group_first_block = ext3_group_first_block_no(sb, group);
 903         group_end_block = group_first_block + (EXT3_BLOCKS_PER_GROUP(sb) - 1);
 904
 905         if (grp_goal < 0)
 906                 start_block = group_first_block;
 907         else
 908                 start_block = grp_goal + group_first_block;
 909
 910         size = my_rsv->rsv_goal_size;
 911
 912         if (!rsv_is_empty(&my_rsv->rsv_window)) {
 913                 /*
 914                  * if the old reservation is cross group boundary
 915                  * and if the goal is inside the old reservation window,
 916                  * we will come here when we just failed to allocate from
 917                  * the first part of the window. We still have another part
 918                  * that belongs to the next group. In this case, there is no
 919                  * point to discard our window and try to allocate a new one
 920                  * in this group(which will fail). we should
 921                  * keep the reservation window, just simply move on.
 922                  *
 923                  * Maybe we could shift the start block of the reservation
 924                  * window to the first block of next group.
 925                  */
 926
 927                 if ((my_rsv->rsv_start <= group_end_block) &&
 928                                 (my_rsv->rsv_end > group_end_block) &&
 929                                 (start_block >= my_rsv->rsv_start))
 930                         return -1;
 931
 932                 if ((my_rsv->rsv_alloc_hit >
 933                      (my_rsv->rsv_end - my_rsv->rsv_start + 1) / 2)) {
 934                         /*
 935                          * if we previously allocation hit ration is greater than half
 936                          * we double the size of reservation window next time
 937                          * otherwise keep the same
 938                          */
 939                         size = size * 2;
 940                         if (size > EXT3_MAX_RESERVE_BLOCKS)
 941                                 size = EXT3_MAX_RESERVE_BLOCKS;
 942                         my_rsv->rsv_goal_size= size;
 943                 }
 944         }
 945
 946         spin_lock(rsv_lock);
 947         /*
 948          * shift the search start to the window near the goal block
 949          */
 950         search_head = search_reserve_window(fs_rsv_root, start_block);
 951
 952         /*
 953          * find_next_reservable_window() simply finds a reservable window
 954          * inside the given range(start_block, group_end_block).
 955          *
 956          * To make sure the reservation window has a free bit inside it, we
 957          * need to check the bitmap after we found a reservable window.
 958          */
 959 retry:
 960         ret = find_next_reservable_window(search_head, my_rsv, sb,
 961                                                 start_block, group_end_block);
 962
 963         if (ret == -1) {
 964                 if (!rsv_is_empty(&my_rsv->rsv_window))
 965                         rsv_window_remove(sb, my_rsv);
 966                 spin_unlock(rsv_lock);
 967                 return -1;
 968         }
 969
 970         /*
 971          * On success, find_next_reservable_window() returns the
 972          * reservation window where there is a reservable space after it.
 973          * Before we reserve this reservable space, we need
 974          * to make sure there is at least a free block inside this region.
 975          *
 976          * searching the first free bit on the block bitmap and copy of
 977          * last committed bitmap alternatively, until we found a allocatable
 978          * block. Search start from the start block of the reservable space
 979          * we just found.
 980          */
 981         spin_unlock(rsv_lock);
 982         first_free_block = bitmap_search_next_usable_block(
 983                         my_rsv->rsv_start - group_first_block,
 984                         bitmap_bh, group_end_block - group_first_block + 1);
 985
 986         if (first_free_block < 0) {
 987                 /*
 988                  * no free block left on the bitmap, no point
 989                  * to reserve the space. return failed.
 990                  */
 991                 spin_lock(rsv_lock);
 992                 if (!rsv_is_empty(&my_rsv->rsv_window))
 993                         rsv_window_remove(sb, my_rsv);
 994                 spin_unlock(rsv_lock);
 995                 return -1;              /* failed */
 996         }
 997
 998         start_block = first_free_block + group_first_block;
 999         /*
1000          * check if the first free block is within the
1001          * free space we just reserved
1002          */
1003         if (start_block >= my_rsv->rsv_start && start_block < my_rsv->rsv_end)
1004                 return 0;               /* success */
1005         /*
1006          * if the first free bit we found is out of the reservable space
1007          * continue search for next reservable space,
1008          * start from where the free block is,
1009          * we also shift the list head to where we stopped last time
1010          */
1011         search_head = my_rsv;
1012         spin_lock(rsv_lock);
1013         goto retry;
1014 }
1015
1016 static void try_to_extend_reservation(struct ext3_reserve_window_node *my_rsv,
1017                         struct super_block *sb, int size)
1018 {
1019         struct ext3_reserve_window_node *next_rsv;
1020         struct rb_node *next;
1021         spinlock_t *rsv_lock = &EXT3_SB(sb)->s_rsv_window_lock;
1022
1023         if (!spin_trylock(rsv_lock))
1024                 return;
1025
1026         next = rb_next(&my_rsv->rsv_node);
1027
1028         if (!next)
1029                 my_rsv->rsv_end += size;
1030         else {
1031                 next_rsv = list_entry(next, struct ext3_reserve_window_node, rsv_node);
1032
1033                 if ((next_rsv->rsv_start - my_rsv->rsv_end - 1) >= size)
1034                         my_rsv->rsv_end += size;
1035                 else
1036                         my_rsv->rsv_end = next_rsv->rsv_start - 1;
1037         }
1038         spin_unlock(rsv_lock);
1039 }
1040
1041 /*
1042  * This is the main function used to allocate a new block and its reservation
1043  * window.
1044  *
1045  * Each time when a new block allocation is need, first try to allocate from
1046  * its own reservation.  If it does not have a reservation window, instead of
1047  * looking for a free bit on bitmap first, then look up the reservation list to
1048  * see if it is inside somebody else's reservation window, we try to allocate a
1049  * reservation window for it starting from the goal first. Then do the block
1050  * allocation within the reservation window.
1051  *
1052  * This will avoid keeping on searching the reservation list again and
1053  * again when somebody is looking for a free block (without
1054  * reservation), and there are lots of free blocks, but they are all
1055  * being reserved.
1056  *
1057  * We use a sorted double linked list for the per-filesystem reservation list.
1058  * The insert, remove and find a free space(non-reserved) operations for the
1059  * sorted double linked list should be fast.
1060  *
1061  */
1062 static ext3_grpblk_t
1063 ext3_try_to_allocate_with_rsv(struct super_block *sb, handle_t *handle,
1064                         unsigned int group, struct buffer_head *bitmap_bh,
1065                         ext3_grpblk_t grp_goal,
1066                         struct ext3_reserve_window_node * my_rsv,
1067                         unsigned long *count, int *errp)
1068 {
1069         ext3_fsblk_t group_first_block, group_last_block;
1070         ext3_grpblk_t ret = 0;
1071         int fatal;
1072         unsigned long num = *count;
1073
1074         *errp = 0;
1075
1076         /*
1077          * Make sure we use undo access for the bitmap, because it is critical
1078          * that we do the frozen_data COW on bitmap buffers in all cases even
1079          * if the buffer is in BJ_Forget state in the committing transaction.
1080          */
1081         BUFFER_TRACE(bitmap_bh, "get undo access for new block");
1082         fatal = ext3_journal_get_undo_access(handle, bitmap_bh);
1083         if (fatal) {
1084                 *errp = fatal;
1085                 return -1;
1086         }
1087
1088         /*
1089          * we don't deal with reservation when
1090          * filesystem is mounted without reservation
1091          * or the file is not a regular file
1092          * or last attempt to allocate a block with reservation turned on failed
1093          */
1094         if (my_rsv == NULL ) {
1095                 ret = ext3_try_to_allocate(sb, handle, group, bitmap_bh,
1096                                                 grp_goal, count, NULL);
1097                 goto out;
1098         }
1099         /*
1100          * grp_goal is a group relative block number (if there is a goal)
1101          * 0 < grp_goal < EXT3_BLOCKS_PER_GROUP(sb)
1102          * first block is a filesystem wide block number
1103          * first block is the block number of the first block in this group
1104          */
1105         group_first_block = ext3_group_first_block_no(sb, group);
1106         group_last_block = group_first_block + (EXT3_BLOCKS_PER_GROUP(sb) - 1);
1107
1108         /*
1109          * Basically we will allocate a new block from inode's reservation
1110          * window.
1111          *
1112          * We need to allocate a new reservation window, if:
1113          * a) inode does not have a reservation window; or
1114          * b) last attempt to allocate a block from existing reservation
1115          *    failed; or
1116          * c) we come here with a goal and with a reservation window
1117          *
1118          * We do not need to allocate a new reservation window if we come here
1119          * at the beginning with a goal and the goal is inside the window, or
1120          * we don't have a goal but already have a reservation window.
1121          * then we could go to allocate from the reservation window directly.
1122          */
1123         while (1) {
1124                 if (rsv_is_empty(&my_rsv->rsv_window) || (ret < 0) ||
1125                         !goal_in_my_reservation(&my_rsv->rsv_window, grp_goal, group, sb)) {
1126                         if (my_rsv->rsv_goal_size < *count)
1127                                 my_rsv->rsv_goal_size = *count;
1128                         ret = alloc_new_reservation(my_rsv, grp_goal, sb,
1129                                                         group, bitmap_bh);
1130                         if (ret < 0)
1131                                 break;                  /* failed */
1132
1133                         if (!goal_in_my_reservation(&my_rsv->rsv_window, grp_goal, group, sb))
1134                                 grp_goal = -1;
1135                 } else if (grp_goal > 0 && (my_rsv->rsv_end-grp_goal+1) < *count)
1136                         try_to_extend_reservation(my_rsv, sb,
1137                                         *count-my_rsv->rsv_end + grp_goal - 1);
1138
1139                 if ((my_rsv->rsv_start > group_last_block) ||
1140                                 (my_rsv->rsv_end < group_first_block))
1141                         BUG();
1142                 ret = ext3_try_to_allocate(sb, handle, group, bitmap_bh, grp_goal,
1143                                            &num, &my_rsv->rsv_window);
1144                 if (ret >= 0) {
1145                         my_rsv->rsv_alloc_hit += num;
1146                         *count = num;
1147                         break;                          /* succeed */
1148                 }
1149                 num = *count;
1150         }
1151 out:
1152         if (ret >= 0) {
1153                 BUFFER_TRACE(bitmap_bh, "journal_dirty_metadata for "
1154                                         "bitmap block");
1155                 fatal = ext3_journal_dirty_metadata(handle, bitmap_bh);
1156                 if (fatal) {
1157                         *errp = fatal;
1158                         return -1;
1159                 }
1160                 return ret;
1161         }
1162
1163         BUFFER_TRACE(bitmap_bh, "journal_release_buffer");
1164         ext3_journal_release_buffer(handle, bitmap_bh);
1165         return ret;
1166 }
1167
1168 static int ext3_has_free_blocks(struct super_block *sb)
1169 {
1170         struct ext3_sb_info *sbi = EXT3_SB(sb);
1171         ext3_fsblk_t free_blocks, root_blocks;
1172         int cond;
1173
1174         free_blocks = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
1175         root_blocks = le32_to_cpu(sbi->s_es->s_r_blocks_count);
1176
1177         vxdprintk(VXD_CBIT(dlim, 3),
1178                 "ext3_has_free_blocks(%p): free=%lu, root=%lu",
1179                 sb, free_blocks, root_blocks);
1180
1181         DLIMIT_ADJUST_BLOCK(sb, vx_current_xid(), &free_blocks, &root_blocks);
1182
1183         cond = (free_blocks < root_blocks + 1 &&
1184                 !capable(CAP_SYS_RESOURCE) &&
1185                 sbi->s_resuid != current->fsuid &&
1186                 (sbi->s_resgid == 0 || !in_group_p (sbi->s_resgid)));
1187
1188         vxdprintk(VXD_CBIT(dlim, 3),
1189                 "ext3_has_free_blocks(%p): %lu<%lu+1, %c, %u!=%u r=%d",
1190                 sb, free_blocks, root_blocks,
1191                 !capable(CAP_SYS_RESOURCE)?'1':'0',
1192                 sbi->s_resuid, current->fsuid, cond?0:1);
1193
1194         return (cond ? 0 : 1);
1195 }
1196
1197 /*
1198  * ext3_should_retry_alloc() is called when ENOSPC is returned, and if
1199  * it is profitable to retry the operation, this function will wait
1200  * for the current or commiting transaction to complete, and then
1201  * return TRUE.
1202  */
1203 int ext3_should_retry_alloc(struct super_block *sb, int *retries)
1204 {
1205         if (!ext3_has_free_blocks(sb) || (*retries)++ > 3)
1206                 return 0;
1207
1208         jbd_debug(1, "%s: retrying operation after ENOSPC\n", sb->s_id);
1209
1210         return journal_force_commit_nested(EXT3_SB(sb)->s_journal);
1211 }
1212
1213 /*
1214  * ext3_new_block uses a goal block to assist allocation.  If the goal is
1215  * free, or there is a free block within 32 blocks of the goal, that block
1216  * is allocated.  Otherwise a forward search is made for a free block; within
1217  * each block group the search first looks for an entire free byte in the block
1218  * bitmap, and then for any free bit if that fails.
1219  * This function also updates quota and i_blocks field.
1220  */
1221 ext3_fsblk_t ext3_new_blocks(handle_t *handle, struct inode *inode,
1222                         ext3_fsblk_t goal, unsigned long *count, int *errp)
1223 {
1224         struct buffer_head *bitmap_bh = NULL;
1225         struct buffer_head *gdp_bh;
1226         int group_no;
1227         int goal_group;
1228         ext3_grpblk_t grp_target_blk;   /* blockgroup relative goal block */
1229         ext3_grpblk_t grp_alloc_blk;    /* blockgroup-relative allocated block*/
1230         ext3_fsblk_t ret_block;         /* filesyetem-wide allocated block */
1231         int bgi;                        /* blockgroup iteration index */
1232         int fatal = 0, err;
1233         int performed_allocation = 0;
1234         ext3_grpblk_t free_blocks;      /* number of free blocks in a group */
1235         struct super_block *sb;
1236         struct ext3_group_desc *gdp;
1237         struct ext3_super_block *es;
1238         struct ext3_sb_info *sbi;
1239         struct ext3_reserve_window_node *my_rsv = NULL;
1240         struct ext3_block_alloc_info *block_i;
1241         unsigned short windowsz = 0;
1242 #ifdef EXT3FS_DEBUG
1243         static int goal_hits, goal_attempts;
1244 #endif
1245         unsigned long ngroups;
1246         unsigned long num = *count;
1247
1248         *errp = -ENOSPC;
1249         sb = inode->i_sb;
1250         if (!sb) {
1251                 printk("ext3_new_block: nonexistent device");
1252                 return 0;
1253         }
1254
1255         /*
1256          * Check quota for allocation of this block.
1257          */
1258         if (DQUOT_ALLOC_BLOCK(inode, num)) {
1259                 *errp = -EDQUOT;
1260                 return 0;
1261         }
1262         if (DLIMIT_ALLOC_BLOCK(inode, 1))
1263             goto out_dlimit;
1264
1265         sbi = EXT3_SB(sb);
1266         es = EXT3_SB(sb)->s_es;
1267         ext3_debug("goal=%lu.\n", goal);
1268         /*
1269          * Allocate a block from reservation only when
1270          * filesystem is mounted with reservation(default,-o reservation), and
1271          * it's a regular file, and
1272          * the desired window size is greater than 0 (One could use ioctl
1273          * command EXT3_IOC_SETRSVSZ to set the window size to 0 to turn off
1274          * reservation on that particular file)
1275          */
1276         block_i = EXT3_I(inode)->i_block_alloc_info;
1277         if (block_i && ((windowsz = block_i->rsv_window_node.rsv_goal_size) > 0))
1278                 my_rsv = &block_i->rsv_window_node;
1279
1280         if (!ext3_has_free_blocks(sb)) {
1281                 *errp = -ENOSPC;
1282                 goto out;
1283         }
1284
1285         /*
1286          * First, test whether the goal block is free.
1287          */
1288         if (goal < le32_to_cpu(es->s_first_data_block) ||
1289             goal >= le32_to_cpu(es->s_blocks_count))
1290                 goal = le32_to_cpu(es->s_first_data_block);
1291         group_no = (goal - le32_to_cpu(es->s_first_data_block)) /
1292                         EXT3_BLOCKS_PER_GROUP(sb);
1293         goal_group = group_no;
1294 retry_alloc:
1295         gdp = ext3_get_group_desc(sb, group_no, &gdp_bh);
1296         if (!gdp)
1297                 goto io_error;
1298
1299         free_blocks = le16_to_cpu(gdp->bg_free_blocks_count);
1300         /*
1301          * if there is not enough free blocks to make a new resevation
1302          * turn off reservation for this allocation
1303          */
1304         if (my_rsv && (free_blocks < windowsz)
1305                 && (rsv_is_empty(&my_rsv->rsv_window)))
1306                 my_rsv = NULL;
1307
1308         if (free_blocks > 0) {
1309                 grp_target_blk = ((goal - le32_to_cpu(es->s_first_data_block)) %
1310                                 EXT3_BLOCKS_PER_GROUP(sb));
1311                 bitmap_bh = read_block_bitmap(sb, group_no);
1312                 if (!bitmap_bh)
1313                         goto io_error;
1314                 grp_alloc_blk = ext3_try_to_allocate_with_rsv(sb, handle,
1315                                         group_no, bitmap_bh, grp_target_blk,
1316                                         my_rsv, &num, &fatal);
1317                 if (fatal)
1318                         goto out;
1319                 if (grp_alloc_blk >= 0)
1320                         goto allocated;
1321         }
1322
1323         ngroups = EXT3_SB(sb)->s_groups_count;
1324         smp_rmb();
1325
1326         /*
1327          * Now search the rest of the groups.  We assume that
1328          * i and gdp correctly point to the last group visited.
1329          */
1330         for (bgi = 0; bgi < ngroups; bgi++) {
1331                 group_no++;
1332                 if (group_no >= ngroups)
1333                         group_no = 0;
1334                 gdp = ext3_get_group_desc(sb, group_no, &gdp_bh);
1335                 if (!gdp) {
1336                         *errp = -EIO;
1337                         goto out;
1338                 }
1339                 free_blocks = le16_to_cpu(gdp->bg_free_blocks_count);
1340                 /*
1341                  * skip this group if the number of
1342                  * free blocks is less than half of the reservation
1343                  * window size.
1344                  */
1345                 if (free_blocks <= (windowsz/2))
1346                         continue;
1347
1348                 brelse(bitmap_bh);
1349                 bitmap_bh = read_block_bitmap(sb, group_no);
1350                 if (!bitmap_bh)
1351                         goto io_error;
1352                 /*
1353                  * try to allocate block(s) from this group, without a goal(-1).
1354                  */
1355                 grp_alloc_blk = ext3_try_to_allocate_with_rsv(sb, handle,
1356                                         group_no, bitmap_bh, -1, my_rsv,
1357                                         &num, &fatal);
1358                 if (fatal)
1359                         goto out;
1360                 if (grp_alloc_blk >= 0)
1361                         goto allocated;
1362         }
1363         /*
1364          * We may end up a bogus ealier ENOSPC error due to
1365          * filesystem is "full" of reservations, but
1366          * there maybe indeed free blocks avaliable on disk
1367          * In this case, we just forget about the reservations
1368          * just do block allocation as without reservations.
1369          */
1370         if (my_rsv) {
1371                 my_rsv = NULL;
1372                 group_no = goal_group;
1373                 goto retry_alloc;
1374         }
1375         /* No space left on the device */
1376         *errp = -ENOSPC;
1377         goto out;
1378
1379 allocated:
1380
1381         ext3_debug("using block group %d(%d)\n",
1382                         group_no, gdp->bg_free_blocks_count);
1383
1384         BUFFER_TRACE(gdp_bh, "get_write_access");
1385         fatal = ext3_journal_get_write_access(handle, gdp_bh);
1386         if (fatal)
1387                 goto out;
1388
1389         ret_block = grp_alloc_blk + ext3_group_first_block_no(sb, group_no);
1390
1391         if (in_range(le32_to_cpu(gdp->bg_block_bitmap), ret_block, num) ||
1392             in_range(le32_to_cpu(gdp->bg_inode_bitmap), ret_block, num) ||
1393             in_range(ret_block, le32_to_cpu(gdp->bg_inode_table),
1394                       EXT3_SB(sb)->s_itb_per_group) ||
1395             in_range(ret_block + num - 1, le32_to_cpu(gdp->bg_inode_table),
1396                       EXT3_SB(sb)->s_itb_per_group))
1397                 ext3_error(sb, "ext3_new_block",
1398                             "Allocating block in system zone - "
1399                             "blocks from "E3FSBLK", length %lu",
1400                              ret_block, num);
1401
1402         performed_allocation = 1;
1403
1404 #ifdef CONFIG_JBD_DEBUG
1405         {
1406                 struct buffer_head *debug_bh;
1407
1408                 /* Record bitmap buffer state in the newly allocated block */
1409                 debug_bh = sb_find_get_block(sb, ret_block);
1410                 if (debug_bh) {
1411                         BUFFER_TRACE(debug_bh, "state when allocated");
1412                         BUFFER_TRACE2(debug_bh, bitmap_bh, "bitmap state");
1413                         brelse(debug_bh);
1414                 }
1415         }
1416         jbd_lock_bh_state(bitmap_bh);
1417         spin_lock(sb_bgl_lock(sbi, group_no));
1418         if (buffer_jbd(bitmap_bh) && bh2jh(bitmap_bh)->b_committed_data) {
1419                 int i;
1420
1421                 for (i = 0; i < num; i++) {
1422                         if (ext3_test_bit(grp_alloc_blk+i,
1423                                         bh2jh(bitmap_bh)->b_committed_data)) {
1424                                 printk("%s: block was unexpectedly set in "
1425                                         "b_committed_data\n", __FUNCTION__);
1426                         }
1427                 }
1428         }
1429         ext3_debug("found bit %d\n", grp_alloc_blk);
1430         spin_unlock(sb_bgl_lock(sbi, group_no));
1431         jbd_unlock_bh_state(bitmap_bh);
1432 #endif
1433
1434         if (ret_block + num - 1 >= le32_to_cpu(es->s_blocks_count)) {
1435                 ext3_error(sb, "ext3_new_block",
1436                             "block("E3FSBLK") >= blocks count(%d) - "
1437                             "block_group = %d, es == %p ", ret_block,
1438                         le32_to_cpu(es->s_blocks_count), group_no, es);
1439                 goto out;
1440         }
1441
1442         /*
1443          * It is up to the caller to add the new buffer to a journal
1444          * list of some description.  We don't know in advance whether
1445          * the caller wants to use it as metadata or data.
1446          */
1447         ext3_debug("allocating block %lu. Goal hits %d of %d.\n",
1448                         ret_block, goal_hits, goal_attempts);
1449
1450         spin_lock(sb_bgl_lock(sbi, group_no));
1451         gdp->bg_free_blocks_count =
1452                         cpu_to_le16(le16_to_cpu(gdp->bg_free_blocks_count) - num);
1453         spin_unlock(sb_bgl_lock(sbi, group_no));
1454         percpu_counter_mod(&sbi->s_freeblocks_counter, -num);
1455
1456         BUFFER_TRACE(gdp_bh, "journal_dirty_metadata for group descriptor");
1457         err = ext3_journal_dirty_metadata(handle, gdp_bh);
1458         if (!fatal)
1459                 fatal = err;
1460
1461         sb->s_dirt = 1;
1462         if (fatal)
1463                 goto out;
1464
1465         *errp = 0;
1466         brelse(bitmap_bh);
1467         DQUOT_FREE_BLOCK(inode, *count-num);
1468         *count = num;
1469         return ret_block;
1470
1471 io_error:
1472         *errp = -EIO;
1473 out:
1474         if (!performed_allocation)
1475                 DLIMIT_FREE_BLOCK(inode, 1);
1476 out_dlimit:
1477         if (fatal) {
1478                 *errp = fatal;
1479                 ext3_std_error(sb, fatal);
1480         }
1481         /*
1482          * Undo the block allocation
1483          */
1484         if (!performed_allocation)
1485                 DQUOT_FREE_BLOCK(inode, *count);
1486         brelse(bitmap_bh);
1487         return 0;
1488 }
1489
1490 ext3_fsblk_t ext3_new_block(handle_t *handle, struct inode *inode,
1491                         ext3_fsblk_t goal, int *errp)
1492 {
1493         unsigned long count = 1;
1494
1495         return ext3_new_blocks(handle, inode, goal, &count, errp);
1496 }
1497
1498 ext3_fsblk_t ext3_count_free_blocks(struct super_block *sb)
1499 {
1500         ext3_fsblk_t desc_count;
1501         struct ext3_group_desc *gdp;
1502         int i;
1503         unsigned long ngroups = EXT3_SB(sb)->s_groups_count;
1504 #ifdef EXT3FS_DEBUG
1505         struct ext3_super_block *es;
1506         ext3_fsblk_t bitmap_count;
1507         unsigned long x;
1508         struct buffer_head *bitmap_bh = NULL;
1509
1510         es = EXT3_SB(sb)->s_es;
1511         desc_count = 0;
1512         bitmap_count = 0;
1513         gdp = NULL;
1514
1515         smp_rmb();
1516         for (i = 0; i < ngroups; i++) {
1517                 gdp = ext3_get_group_desc(sb, i, NULL);
1518                 if (!gdp)
1519                         continue;
1520                 desc_count += le16_to_cpu(gdp->bg_free_blocks_count);
1521                 brelse(bitmap_bh);
1522                 bitmap_bh = read_block_bitmap(sb, i);
1523                 if (bitmap_bh == NULL)
1524                         continue;
1525
1526                 x = ext3_count_free(bitmap_bh, sb->s_blocksize);
1527                 printk("group %d: stored = %d, counted = %lu\n",
1528                         i, le16_to_cpu(gdp->bg_free_blocks_count), x);
1529                 bitmap_count += x;
1530         }
1531         brelse(bitmap_bh);
1532         printk("ext3_count_free_blocks: stored = "E3FSBLK
1533                 ", computed = "E3FSBLK", "E3FSBLK"\n",
1534                le32_to_cpu(es->s_free_blocks_count),
1535                 desc_count, bitmap_count);
1536         return bitmap_count;
1537 #else
1538         desc_count = 0;
1539         smp_rmb();
1540         for (i = 0; i < ngroups; i++) {
1541                 gdp = ext3_get_group_desc(sb, i, NULL);
1542                 if (!gdp)
1543                         continue;
1544                 desc_count += le16_to_cpu(gdp->bg_free_blocks_count);
1545         }
1546
1547         return desc_count;
1548 #endif
1549 }
1550
1551 static inline int
1552 block_in_use(ext3_fsblk_t block, struct super_block *sb, unsigned char *map)
1553 {
1554         return ext3_test_bit ((block -
1555                 le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block)) %
1556                          EXT3_BLOCKS_PER_GROUP(sb), map);
1557 }
1558
1559 static inline int test_root(int a, int b)
1560 {
1561         int num = b;
1562
1563         while (a > num)
1564                 num *= b;
1565         return num == a;
1566 }
1567
1568 static int ext3_group_sparse(int group)
1569 {
1570         if (group <= 1)
1571                 return 1;
1572         if (!(group & 1))
1573                 return 0;
1574         return (test_root(group, 7) || test_root(group, 5) ||
1575                 test_root(group, 3));
1576 }
1577
1578 /**
1579  *      ext3_bg_has_super - number of blocks used by the superblock in group
1580  *      @sb: superblock for filesystem
1581  *      @group: group number to check
1582  *
1583  *      Return the number of blocks used by the superblock (primary or backup)
1584  *      in this group.  Currently this will be only 0 or 1.
1585  */
1586 int ext3_bg_has_super(struct super_block *sb, int group)
1587 {
1588         if (EXT3_HAS_RO_COMPAT_FEATURE(sb,
1589                                 EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER) &&
1590                         !ext3_group_sparse(group))
1591                 return 0;
1592         return 1;
1593 }
1594
1595 static unsigned long ext3_bg_num_gdb_meta(struct super_block *sb, int group)
1596 {
1597         unsigned long metagroup = group / EXT3_DESC_PER_BLOCK(sb);
1598         unsigned long first = metagroup * EXT3_DESC_PER_BLOCK(sb);
1599         unsigned long last = first + EXT3_DESC_PER_BLOCK(sb) - 1;
1600
1601         if (group == first || group == first + 1 || group == last)
1602                 return 1;
1603         return 0;
1604 }
1605
1606 static unsigned long ext3_bg_num_gdb_nometa(struct super_block *sb, int group)
1607 {
1608         if (EXT3_HAS_RO_COMPAT_FEATURE(sb,
1609                                 EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER) &&
1610                         !ext3_group_sparse(group))
1611                 return 0;
1612         return EXT3_SB(sb)->s_gdb_count;
1613 }
1614
1615 /**
1616  *      ext3_bg_num_gdb - number of blocks used by the group table in group
1617  *      @sb: superblock for filesystem
1618  *      @group: group number to check
1619  *
1620  *      Return the number of blocks used by the group descriptor table
1621  *      (primary or backup) in this group.  In the future there may be a
1622  *      different number of descriptor blocks in each group.
1623  */
1624 unsigned long ext3_bg_num_gdb(struct super_block *sb, int group)
1625 {
1626         unsigned long first_meta_bg =
1627                         le32_to_cpu(EXT3_SB(sb)->s_es->s_first_meta_bg);
1628         unsigned long metagroup = group / EXT3_DESC_PER_BLOCK(sb);
1629
1630         if (!EXT3_HAS_INCOMPAT_FEATURE(sb,EXT3_FEATURE_INCOMPAT_META_BG) ||
1631                         metagroup < first_meta_bg)
1632                 return ext3_bg_num_gdb_nometa(sb,group);
1633
1634         return ext3_bg_num_gdb_meta(sb,group);
1635
1636 }