fs/ext4/balloc.c

   1 /*
   2  *  linux/fs/ext4/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/quotaops.h>
  18 #include <linux/buffer_head.h>
  19 #include "ext4.h"
  20 #include "ext4_jbd2.h"
  21 #include "mballoc.h"
  22
  23 #include <trace/events/ext4.h>
  24
  25 static unsigned ext4_num_base_meta_clusters(struct super_block *sb,
  26                                             ext4_group_t block_group);
  27 /*
  28  * balloc.c contains the blocks allocation and deallocation routines
  29  */
  30
  31 /*
  32  * Calculate block group number for a given block number
  33  */
  34 ext4_group_t ext4_get_group_number(struct super_block *sb,
  35                                    ext4_fsblk_t block)
  36 {
  37         ext4_group_t group;
  38
  39         if (test_opt2(sb, STD_GROUP_SIZE))
  40                 group = (block -
  41                          le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block)) >>
  42                         (EXT4_BLOCK_SIZE_BITS(sb) + EXT4_CLUSTER_BITS(sb) + 3);
  43         else
  44                 ext4_get_group_no_and_offset(sb, block, &group, NULL);
  45         return group;
  46 }
  47
  48 /*
  49  * Calculate the block group number and offset into the block/cluster
  50  * allocation bitmap, given a block number
  51  */
  52 void ext4_get_group_no_and_offset(struct super_block *sb, ext4_fsblk_t blocknr,
  53                 ext4_group_t *blockgrpp, ext4_grpblk_t *offsetp)
  54 {
  55         struct ext4_super_block *es = EXT4_SB(sb)->s_es;
  56         ext4_grpblk_t offset;
  57
  58         blocknr = blocknr - le32_to_cpu(es->s_first_data_block);
  59         offset = do_div(blocknr, EXT4_BLOCKS_PER_GROUP(sb)) >>
  60                 EXT4_SB(sb)->s_cluster_bits;
  61         if (offsetp)
  62                 *offsetp = offset;
  63         if (blockgrpp)
  64                 *blockgrpp = blocknr;
  65
  66 }
  67
  68 /*
  69  * Check whether the 'block' lives within the 'block_group'. Returns 1 if so
  70  * and 0 otherwise.
  71  */
  72 static inline int ext4_block_in_group(struct super_block *sb,
  73                                       ext4_fsblk_t block,
  74                                       ext4_group_t block_group)
  75 {
  76         ext4_group_t actual_group;
  77
  78         actual_group = ext4_get_group_number(sb, block);
  79         return (actual_group == block_group) ? 1 : 0;
  80 }
  81
  82 /* Return the number of clusters used for file system metadata; this
  83  * represents the overhead needed by the file system.
  84  */
  85 static unsigned ext4_num_overhead_clusters(struct super_block *sb,
  86                                            ext4_group_t block_group,
  87                                            struct ext4_group_desc *gdp)
  88 {
  89         unsigned num_clusters;
  90         int block_cluster = -1, inode_cluster = -1, itbl_cluster = -1, i, c;
  91         ext4_fsblk_t start = ext4_group_first_block_no(sb, block_group);
  92         ext4_fsblk_t itbl_blk;
  93         struct ext4_sb_info *sbi = EXT4_SB(sb);
  94
  95         /* This is the number of clusters used by the superblock,
  96          * block group descriptors, and reserved block group
  97          * descriptor blocks */
  98         num_clusters = ext4_num_base_meta_clusters(sb, block_group);
  99
 100         /*
 101          * For the allocation bitmaps and inode table, we first need
 102          * to check to see if the block is in the block group.  If it
 103          * is, then check to see if the cluster is already accounted
 104          * for in the clusters used for the base metadata cluster, or
 105          * if we can increment the base metadata cluster to include
 106          * that block.  Otherwise, we will have to track the cluster
 107          * used for the allocation bitmap or inode table explicitly.
 108          * Normally all of these blocks are contiguous, so the special
 109          * case handling shouldn't be necessary except for *very*
 110          * unusual file system layouts.
 111          */
 112         if (ext4_block_in_group(sb, ext4_block_bitmap(sb, gdp), block_group)) {
 113                 block_cluster = EXT4_B2C(sbi,
 114                                          ext4_block_bitmap(sb, gdp) - start);
 115                 if (block_cluster < num_clusters)
 116                         block_cluster = -1;
 117                 else if (block_cluster == num_clusters) {
 118                         num_clusters++;
 119                         block_cluster = -1;
 120                 }
 121         }
 122
 123         if (ext4_block_in_group(sb, ext4_inode_bitmap(sb, gdp), block_group)) {
 124                 inode_cluster = EXT4_B2C(sbi,
 125                                          ext4_inode_bitmap(sb, gdp) - start);
 126                 if (inode_cluster < num_clusters)
 127                         inode_cluster = -1;
 128                 else if (inode_cluster == num_clusters) {
 129                         num_clusters++;
 130                         inode_cluster = -1;
 131                 }
 132         }
 133
 134         itbl_blk = ext4_inode_table(sb, gdp);
 135         for (i = 0; i < sbi->s_itb_per_group; i++) {
 136                 if (ext4_block_in_group(sb, itbl_blk + i, block_group)) {
 137                         c = EXT4_B2C(sbi, itbl_blk + i - start);
 138                         if ((c < num_clusters) || (c == inode_cluster) ||
 139                             (c == block_cluster) || (c == itbl_cluster))
 140                                 continue;
 141                         if (c == num_clusters) {
 142                                 num_clusters++;
 143                                 continue;
 144                         }
 145                         num_clusters++;
 146                         itbl_cluster = c;
 147                 }
 148         }
 149
 150         if (block_cluster != -1)
 151                 num_clusters++;
 152         if (inode_cluster != -1)
 153                 num_clusters++;
 154
 155         return num_clusters;
 156 }
 157
 158 static unsigned int num_clusters_in_group(struct super_block *sb,
 159                                           ext4_group_t block_group)
 160 {
 161         unsigned int blocks;
 162
 163         if (block_group == ext4_get_groups_count(sb) - 1) {
 164                 /*
 165                  * Even though mke2fs always initializes the first and
 166                  * last group, just in case some other tool was used,
 167                  * we need to make sure we calculate the right free
 168                  * blocks.
 169                  */
 170                 blocks = ext4_blocks_count(EXT4_SB(sb)->s_es) -
 171                         ext4_group_first_block_no(sb, block_group);
 172         } else
 173                 blocks = EXT4_BLOCKS_PER_GROUP(sb);
 174         return EXT4_NUM_B2C(EXT4_SB(sb), blocks);
 175 }
 176
 177 /* Initializes an uninitialized block bitmap */
 178 static int ext4_init_block_bitmap(struct super_block *sb,
 179                                    struct buffer_head *bh,
 180                                    ext4_group_t block_group,
 181                                    struct ext4_group_desc *gdp)
 182 {
 183         unsigned int bit, bit_max;
 184         struct ext4_sb_info *sbi = EXT4_SB(sb);
 185         ext4_fsblk_t start, tmp;
 186         struct ext4_group_info *grp;
 187
 188         J_ASSERT_BH(bh, buffer_locked(bh));
 189
 190         /* If checksum is bad mark all blocks used to prevent allocation
 191          * essentially implementing a per-group read-only flag. */
 192         if (!ext4_group_desc_csum_verify(sb, block_group, gdp)) {
 193                 grp = ext4_get_group_info(sb, block_group);
 194                 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp))
 195                         percpu_counter_sub(&sbi->s_freeclusters_counter,
 196                                            grp->bb_free);
 197                 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state);
 198                 if (!EXT4_MB_GRP_IBITMAP_CORRUPT(grp)) {
 199                         int count;
 200                         count = ext4_free_inodes_count(sb, gdp);
 201                         percpu_counter_sub(&sbi->s_freeinodes_counter,
 202                                            count);
 203                 }
 204                 set_bit(EXT4_GROUP_INFO_IBITMAP_CORRUPT_BIT, &grp->bb_state);
 205                 return -EFSBADCRC;
 206         }
 207         memset(bh->b_data, 0, sb->s_blocksize);
 208
 209         bit_max = ext4_num_base_meta_clusters(sb, block_group);
 210         if ((bit_max >> 3) >= bh->b_size)
 211                 return -EFSCORRUPTED;
 212
 213         for (bit = 0; bit < bit_max; bit++)
 214                 ext4_set_bit(bit, bh->b_data);
 215
 216         start = ext4_group_first_block_no(sb, block_group);
 217
 218         /* Set bits for block and inode bitmaps, and inode table */
 219         tmp = ext4_block_bitmap(sb, gdp);
 220         if (ext4_block_in_group(sb, tmp, block_group))
 221                 ext4_set_bit(EXT4_B2C(sbi, tmp - start), bh->b_data);
 222
 223         tmp = ext4_inode_bitmap(sb, gdp);
 224         if (ext4_block_in_group(sb, tmp, block_group))
 225                 ext4_set_bit(EXT4_B2C(sbi, tmp - start), bh->b_data);
 226
 227         tmp = ext4_inode_table(sb, gdp);
 228         for (; tmp < ext4_inode_table(sb, gdp) +
 229                      sbi->s_itb_per_group; tmp++) {
 230                 if (ext4_block_in_group(sb, tmp, block_group))
 231                         ext4_set_bit(EXT4_B2C(sbi, tmp - start), bh->b_data);
 232         }
 233
 234         /*
 235          * Also if the number of blocks within the group is less than
 236          * the blocksize * 8 ( which is the size of bitmap ), set rest
 237          * of the block bitmap to 1
 238          */
 239         ext4_mark_bitmap_end(num_clusters_in_group(sb, block_group),
 240                              sb->s_blocksize * 8, bh->b_data);
 241         return 0;
 242 }
 243
 244 /* Return the number of free blocks in a block group.  It is used when
 245  * the block bitmap is uninitialized, so we can't just count the bits
 246  * in the bitmap. */
 247 unsigned ext4_free_clusters_after_init(struct super_block *sb,
 248                                        ext4_group_t block_group,
 249                                        struct ext4_group_desc *gdp)
 250 {
 251         return num_clusters_in_group(sb, block_group) -
 252                 ext4_num_overhead_clusters(sb, block_group, gdp);
 253 }
 254
 255 /*
 256  * The free blocks are managed by bitmaps.  A file system contains several
 257  * blocks groups.  Each group contains 1 bitmap block for blocks, 1 bitmap
 258  * block for inodes, N blocks for the inode table and data blocks.
 259  *
 260  * The file system contains group descriptors which are located after the
 261  * super block.  Each descriptor contains the number of the bitmap block and
 262  * the free blocks count in the block.  The descriptors are loaded in memory
 263  * when a file system is mounted (see ext4_fill_super).
 264  */
 265
 266 /**
 267  * ext4_get_group_desc() -- load group descriptor from disk
 268  * @sb:                 super block
 269  * @block_group:        given block group
 270  * @bh:                 pointer to the buffer head to store the block
 271  *                      group descriptor
 272  */
 273 struct ext4_group_desc * ext4_get_group_desc(struct super_block *sb,
 274                                              ext4_group_t block_group,
 275                                              struct buffer_head **bh)
 276 {
 277         unsigned int group_desc;
 278         unsigned int offset;
 279         ext4_group_t ngroups = ext4_get_groups_count(sb);
 280         struct ext4_group_desc *desc;
 281         struct ext4_sb_info *sbi = EXT4_SB(sb);
 282         struct buffer_head *bh_p;
 283
 284         if (block_group >= ngroups) {
 285                 ext4_error(sb, "block_group >= groups_count - block_group = %u,"
 286                            " groups_count = %u", block_group, ngroups);
 287
 288                 return NULL;
 289         }
 290
 291         group_desc = block_group >> EXT4_DESC_PER_BLOCK_BITS(sb);
 292         offset = block_group & (EXT4_DESC_PER_BLOCK(sb) - 1);
 293         bh_p = sbi_array_rcu_deref(sbi, s_group_desc, group_desc);
 294         /*
 295          * sbi_array_rcu_deref returns with rcu unlocked, this is ok since
 296          * the pointer being dereferenced won't be dereferenced again. By
 297          * looking at the usage in add_new_gdb() the value isn't modified,
 298          * just the pointer, and so it remains valid.
 299          */
 300         if (!bh_p) {
 301                 ext4_error(sb, "Group descriptor not loaded - "
 302                            "block_group = %u, group_desc = %u, desc = %u",
 303                            block_group, group_desc, offset);
 304                 return NULL;
 305         }
 306
 307         desc = (struct ext4_group_desc *)(
 308                 (__u8 *)bh_p->b_data +
 309                 offset * EXT4_DESC_SIZE(sb));
 310         if (bh)
 311                 *bh = bh_p;
 312         return desc;
 313 }
 314
 315 /*
 316  * Return the block number which was discovered to be invalid, or 0 if
 317  * the block bitmap is valid.
 318  */
 319 static ext4_fsblk_t ext4_valid_block_bitmap(struct super_block *sb,
 320                                             struct ext4_group_desc *desc,
 321                                             ext4_group_t block_group,
 322                                             struct buffer_head *bh)
 323 {
 324         struct ext4_sb_info *sbi = EXT4_SB(sb);
 325         ext4_grpblk_t offset;
 326         ext4_grpblk_t next_zero_bit;
 327         ext4_grpblk_t max_bit = EXT4_CLUSTERS_PER_GROUP(sb);
 328         ext4_fsblk_t blk;
 329         ext4_fsblk_t group_first_block;
 330
 331         if (ext4_has_feature_flex_bg(sb)) {
 332                 /* with FLEX_BG, the inode/block bitmaps and itable
 333                  * blocks may not be in the group at all
 334                  * so the bitmap validation will be skipped for those groups
 335                  * or it has to also read the block group where the bitmaps
 336                  * are located to verify they are set.
 337                  */
 338                 return 0;
 339         }
 340         group_first_block = ext4_group_first_block_no(sb, block_group);
 341
 342         /* check whether block bitmap block number is set */
 343         blk = ext4_block_bitmap(sb, desc);
 344         offset = blk - group_first_block;
 345         if (offset < 0 || EXT4_B2C(sbi, offset) >= max_bit ||
 346             !ext4_test_bit(EXT4_B2C(sbi, offset), bh->b_data))
 347                 /* bad block bitmap */
 348                 return blk;
 349
 350         /* check whether the inode bitmap block number is set */
 351         blk = ext4_inode_bitmap(sb, desc);
 352         offset = blk - group_first_block;
 353         if (offset < 0 || EXT4_B2C(sbi, offset) >= max_bit ||
 354             !ext4_test_bit(EXT4_B2C(sbi, offset), bh->b_data))
 355                 /* bad block bitmap */
 356                 return blk;
 357
 358         /* check whether the inode table block number is set */
 359         blk = ext4_inode_table(sb, desc);
 360         offset = blk - group_first_block;
 361         if (offset < 0 || EXT4_B2C(sbi, offset) >= max_bit ||
 362             EXT4_B2C(sbi, offset + sbi->s_itb_per_group) >= max_bit)
 363                 return blk;
 364         next_zero_bit = ext4_find_next_zero_bit(bh->b_data,
 365                         EXT4_B2C(sbi, offset + EXT4_SB(sb)->s_itb_per_group),
 366                         EXT4_B2C(sbi, offset));
 367         if (next_zero_bit <
 368             EXT4_B2C(sbi, offset + EXT4_SB(sb)->s_itb_per_group))
 369                 /* bad bitmap for inode tables */
 370                 return blk;
 371         return 0;
 372 }
 373
 374 static int ext4_validate_block_bitmap(struct super_block *sb,
 375                                       struct ext4_group_desc *desc,
 376                                       ext4_group_t block_group,
 377                                       struct buffer_head *bh)
 378 {
 379         ext4_fsblk_t    blk;
 380         struct ext4_group_info *grp = ext4_get_group_info(sb, block_group);
 381         struct ext4_sb_info *sbi = EXT4_SB(sb);
 382
 383         if (buffer_verified(bh))
 384                 return 0;
 385         if (EXT4_MB_GRP_BBITMAP_CORRUPT(grp))
 386                 return -EFSCORRUPTED;
 387
 388         ext4_lock_group(sb, block_group);
 389         if (buffer_verified(bh))
 390                 goto verified;
 391         if (unlikely(!ext4_block_bitmap_csum_verify(sb, block_group,
 392                         desc, bh))) {
 393                 ext4_unlock_group(sb, block_group);
 394                 ext4_error(sb, "bg %u: bad block bitmap checksum", block_group);
 395                 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp))
 396                         percpu_counter_sub(&sbi->s_freeclusters_counter,
 397                                            grp->bb_free);
 398                 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state);
 399                 return -EFSBADCRC;
 400         }
 401         blk = ext4_valid_block_bitmap(sb, desc, block_group, bh);
 402         if (unlikely(blk != 0)) {
 403                 ext4_unlock_group(sb, block_group);
 404                 ext4_error(sb, "bg %u: block %llu: invalid block bitmap",
 405                            block_group, blk);
 406                 if (!EXT4_MB_GRP_BBITMAP_CORRUPT(grp))
 407                         percpu_counter_sub(&sbi->s_freeclusters_counter,
 408                                            grp->bb_free);
 409                 set_bit(EXT4_GROUP_INFO_BBITMAP_CORRUPT_BIT, &grp->bb_state);
 410                 return -EFSCORRUPTED;
 411         }
 412         set_buffer_verified(bh);
 413 verified:
 414         ext4_unlock_group(sb, block_group);
 415         return 0;
 416 }
 417
 418 /**
 419  * ext4_read_block_bitmap_nowait()
 420  * @sb:                 super block
 421  * @block_group:        given block group
 422  *
 423  * Read the bitmap for a given block_group,and validate the
 424  * bits for block/inode/inode tables are set in the bitmaps
 425  *
 426  * Return buffer_head on success or NULL in case of failure.
 427  */
 428 struct buffer_head *
 429 ext4_read_block_bitmap_nowait(struct super_block *sb, ext4_group_t block_group)
 430 {
 431         struct ext4_group_desc *desc;
 432         struct ext4_sb_info *sbi = EXT4_SB(sb);
 433         struct buffer_head *bh;
 434         ext4_fsblk_t bitmap_blk;
 435         int err;
 436
 437         desc = ext4_get_group_desc(sb, block_group, NULL);
 438         if (!desc)
 439                 return ERR_PTR(-EFSCORRUPTED);
 440         bitmap_blk = ext4_block_bitmap(sb, desc);
 441         if ((bitmap_blk <= le32_to_cpu(sbi->s_es->s_first_data_block)) ||
 442             (bitmap_blk >= ext4_blocks_count(sbi->s_es))) {
 443                 ext4_error(sb, "Invalid block bitmap block %llu in "
 444                            "block_group %u", bitmap_blk, block_group);
 445                 return ERR_PTR(-EFSCORRUPTED);
 446         }
 447         bh = sb_getblk(sb, bitmap_blk);
 448         if (unlikely(!bh)) {
 449                 ext4_error(sb, "Cannot get buffer for block bitmap - "
 450                            "block_group = %u, block_bitmap = %llu",
 451                            block_group, bitmap_blk);
 452                 return ERR_PTR(-ENOMEM);
 453         }
 454
 455         if (bitmap_uptodate(bh))
 456                 goto verify;
 457
 458         lock_buffer(bh);
 459         if (bitmap_uptodate(bh)) {
 460                 unlock_buffer(bh);
 461                 goto verify;
 462         }
 463         ext4_lock_group(sb, block_group);
 464         if (ext4_has_group_desc_csum(sb) &&
 465             (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT))) {
 466                 if (block_group == 0) {
 467                         ext4_unlock_group(sb, block_group);
 468                         unlock_buffer(bh);
 469                         ext4_error(sb, "Block bitmap for bg 0 marked "
 470                                    "uninitialized");
 471                         err = -EFSCORRUPTED;
 472                         goto out;
 473                 }
 474                 err = ext4_init_block_bitmap(sb, bh, block_group, desc);
 475                 set_bitmap_uptodate(bh);
 476                 set_buffer_uptodate(bh);
 477                 set_buffer_verified(bh);
 478                 ext4_unlock_group(sb, block_group);
 479                 unlock_buffer(bh);
 480                 if (err) {
 481                         ext4_error(sb, "Failed to init block bitmap for group "
 482                                    "%u: %d", block_group, err);
 483                         goto out;
 484                 }
 485                 goto verify;
 486         }
 487         ext4_unlock_group(sb, block_group);
 488         if (buffer_uptodate(bh)) {
 489                 /*
 490                  * if not uninit if bh is uptodate,
 491                  * bitmap is also uptodate
 492                  */
 493                 set_bitmap_uptodate(bh);
 494                 unlock_buffer(bh);
 495                 goto verify;
 496         }
 497         /*
 498          * submit the buffer_head for reading
 499          */
 500         set_buffer_new(bh);
 501         trace_ext4_read_block_bitmap_load(sb, block_group);
 502         bh->b_end_io = ext4_end_bitmap_read;
 503         get_bh(bh);
 504         submit_bh(REQ_OP_READ, REQ_META | REQ_PRIO, bh);
 505         return bh;
 506 verify:
 507         err = ext4_validate_block_bitmap(sb, desc, block_group, bh);
 508         if (err)
 509                 goto out;
 510         return bh;
 511 out:
 512         put_bh(bh);
 513         return ERR_PTR(err);
 514 }
 515
 516 /* Returns 0 on success, 1 on error */
 517 int ext4_wait_block_bitmap(struct super_block *sb, ext4_group_t block_group,
 518                            struct buffer_head *bh)
 519 {
 520         struct ext4_group_desc *desc;
 521
 522         if (!buffer_new(bh))
 523                 return 0;
 524         desc = ext4_get_group_desc(sb, block_group, NULL);
 525         if (!desc)
 526                 return -EFSCORRUPTED;
 527         wait_on_buffer(bh);
 528         if (!buffer_uptodate(bh)) {
 529                 ext4_error(sb, "Cannot read block bitmap - "
 530                            "block_group = %u, block_bitmap = %llu",
 531                            block_group, (unsigned long long) bh->b_blocknr);
 532                 return -EIO;
 533         }
 534         clear_buffer_new(bh);
 535         /* Panic or remount fs read-only if block bitmap is invalid */
 536         return ext4_validate_block_bitmap(sb, desc, block_group, bh);
 537 }
 538
 539 struct buffer_head *
 540 ext4_read_block_bitmap(struct super_block *sb, ext4_group_t block_group)
 541 {
 542         struct buffer_head *bh;
 543         int err;
 544
 545         bh = ext4_read_block_bitmap_nowait(sb, block_group);
 546         if (IS_ERR(bh))
 547                 return bh;
 548         err = ext4_wait_block_bitmap(sb, block_group, bh);
 549         if (err) {
 550                 put_bh(bh);
 551                 return ERR_PTR(err);
 552         }
 553         return bh;
 554 }
 555
 556 /**
 557  * ext4_has_free_clusters()
 558  * @sbi:        in-core super block structure.
 559  * @nclusters:  number of needed blocks
 560  * @flags:      flags from ext4_mb_new_blocks()
 561  *
 562  * Check if filesystem has nclusters free & available for allocation.
 563  * On success return 1, return 0 on failure.
 564  */
 565 static int ext4_has_free_clusters(struct ext4_sb_info *sbi,
 566                                   s64 nclusters, unsigned int flags)
 567 {
 568         s64 free_clusters, dirty_clusters, rsv, resv_clusters;
 569         struct percpu_counter *fcc = &sbi->s_freeclusters_counter;
 570         struct percpu_counter *dcc = &sbi->s_dirtyclusters_counter;
 571
 572         free_clusters  = percpu_counter_read_positive(fcc);
 573         dirty_clusters = percpu_counter_read_positive(dcc);
 574         resv_clusters = atomic64_read(&sbi->s_resv_clusters);
 575
 576         /*
 577          * r_blocks_count should always be multiple of the cluster ratio so
 578          * we are safe to do a plane bit shift only.
 579          */
 580         rsv = (ext4_r_blocks_count(sbi->s_es) >> sbi->s_cluster_bits) +
 581               resv_clusters;
 582
 583         if (free_clusters - (nclusters + rsv + dirty_clusters) <
 584                                         EXT4_FREECLUSTERS_WATERMARK) {
 585                 free_clusters  = percpu_counter_sum_positive(fcc);
 586                 dirty_clusters = percpu_counter_sum_positive(dcc);
 587         }
 588         /* Check whether we have space after accounting for current
 589          * dirty clusters & root reserved clusters.
 590          */
 591         if (free_clusters >= (rsv + nclusters + dirty_clusters))
 592                 return 1;
 593
 594         /* Hm, nope.  Are (enough) root reserved clusters available? */
 595         if (uid_eq(sbi->s_resuid, current_fsuid()) ||
 596             (!gid_eq(sbi->s_resgid, GLOBAL_ROOT_GID) && in_group_p(sbi->s_resgid)) ||
 597             capable(CAP_SYS_RESOURCE) ||
 598             (flags & EXT4_MB_USE_ROOT_BLOCKS)) {
 599
 600                 if (free_clusters >= (nclusters + dirty_clusters +
 601                                       resv_clusters))
 602                         return 1;
 603         }
 604         /* No free blocks. Let's see if we can dip into reserved pool */
 605         if (flags & EXT4_MB_USE_RESERVED) {
 606                 if (free_clusters >= (nclusters + dirty_clusters))
 607                         return 1;
 608         }
 609
 610         return 0;
 611 }
 612
 613 int ext4_claim_free_clusters(struct ext4_sb_info *sbi,
 614                              s64 nclusters, unsigned int flags)
 615 {
 616         if (ext4_has_free_clusters(sbi, nclusters, flags)) {
 617                 percpu_counter_add(&sbi->s_dirtyclusters_counter, nclusters);
 618                 return 0;
 619         } else
 620                 return -ENOSPC;
 621 }
 622
 623 /**
 624  * ext4_should_retry_alloc()
 625  * @sb:                 super block
 626  * @retries             number of attemps has been made
 627  *
 628  * ext4_should_retry_alloc() is called when ENOSPC is returned, and if
 629  * it is profitable to retry the operation, this function will wait
 630  * for the current or committing transaction to complete, and then
 631  * return TRUE.
 632  *
 633  * if the total number of retries exceed three times, return FALSE.
 634  */
 635 int ext4_should_retry_alloc(struct super_block *sb, int *retries)
 636 {
 637         if (!ext4_has_free_clusters(EXT4_SB(sb), 1, 0) ||
 638             (*retries)++ > 3 ||
 639             !EXT4_SB(sb)->s_journal)
 640                 return 0;
 641
 642         jbd_debug(1, "%s: retrying operation after ENOSPC\n", sb->s_id);
 643
 644         smp_mb();
 645         if (EXT4_SB(sb)->s_mb_free_pending)
 646                 jbd2_journal_force_commit_nested(EXT4_SB(sb)->s_journal);
 647         return 1;
 648 }
 649
 650 /*
 651  * ext4_new_meta_blocks() -- allocate block for meta data (indexing) blocks
 652  *
 653  * @handle:             handle to this transaction
 654  * @inode:              file inode
 655  * @goal:               given target block(filesystem wide)
 656  * @count:              pointer to total number of clusters needed
 657  * @errp:               error code
 658  *
 659  * Return 1st allocated block number on success, *count stores total account
 660  * error stores in errp pointer
 661  */
 662 ext4_fsblk_t ext4_new_meta_blocks(handle_t *handle, struct inode *inode,
 663                                   ext4_fsblk_t goal, unsigned int flags,
 664                                   unsigned long *count, int *errp)
 665 {
 666         struct ext4_allocation_request ar;
 667         ext4_fsblk_t ret;
 668
 669         memset(&ar, 0, sizeof(ar));
 670         /* Fill with neighbour allocated blocks */
 671         ar.inode = inode;
 672         ar.goal = goal;
 673         ar.len = count ? *count : 1;
 674         ar.flags = flags;
 675
 676         ret = ext4_mb_new_blocks(handle, &ar, errp);
 677         if (count)
 678                 *count = ar.len;
 679         /*
 680          * Account for the allocated meta blocks.  We will never
 681          * fail EDQUOT for metdata, but we do account for it.
 682          */
 683         if (!(*errp) && (flags & EXT4_MB_DELALLOC_RESERVED)) {
 684                 dquot_alloc_block_nofail(inode,
 685                                 EXT4_C2B(EXT4_SB(inode->i_sb), ar.len));
 686         }
 687         return ret;
 688 }
 689
 690 /**
 691  * ext4_count_free_clusters() -- count filesystem free clusters
 692  * @sb:         superblock
 693  *
 694  * Adds up the number of free clusters from each block group.
 695  */
 696 ext4_fsblk_t ext4_count_free_clusters(struct super_block *sb)
 697 {
 698         ext4_fsblk_t desc_count;
 699         struct ext4_group_desc *gdp;
 700         ext4_group_t i;
 701         ext4_group_t ngroups = ext4_get_groups_count(sb);
 702         struct ext4_group_info *grp;
 703 #ifdef EXT4FS_DEBUG
 704         struct ext4_super_block *es;
 705         ext4_fsblk_t bitmap_count;
 706         unsigned int x;
 707         struct buffer_head *bitmap_bh = NULL;
 708
 709         es = EXT4_SB(sb)->s_es;
 710         desc_count = 0;
 711         bitmap_count = 0;
 712         gdp = NULL;
 713
 714         for (i = 0; i < ngroups; i++) {
 715                 gdp = ext4_get_group_desc(sb, i, NULL);
 716                 if (!gdp)
 717                         continue;
 718                 grp = NULL;
 719                 if (EXT4_SB(sb)->s_group_info)
 720                         grp = ext4_get_group_info(sb, i);
 721                 if (!grp || !EXT4_MB_GRP_BBITMAP_CORRUPT(grp))
 722                         desc_count += ext4_free_group_clusters(sb, gdp);
 723                 brelse(bitmap_bh);
 724                 bitmap_bh = ext4_read_block_bitmap(sb, i);
 725                 if (IS_ERR(bitmap_bh)) {
 726                         bitmap_bh = NULL;
 727                         continue;
 728                 }
 729
 730                 x = ext4_count_free(bitmap_bh->b_data,
 731                                     EXT4_CLUSTERS_PER_GROUP(sb) / 8);
 732                 printk(KERN_DEBUG "group %u: stored = %d, counted = %u\n",
 733                         i, ext4_free_group_clusters(sb, gdp), x);
 734                 bitmap_count += x;
 735         }
 736         brelse(bitmap_bh);
 737         printk(KERN_DEBUG "ext4_count_free_clusters: stored = %llu"
 738                ", computed = %llu, %llu\n",
 739                EXT4_NUM_B2C(EXT4_SB(sb), ext4_free_blocks_count(es)),
 740                desc_count, bitmap_count);
 741         return bitmap_count;
 742 #else
 743         desc_count = 0;
 744         for (i = 0; i < ngroups; i++) {
 745                 gdp = ext4_get_group_desc(sb, i, NULL);
 746                 if (!gdp)
 747                         continue;
 748                 grp = NULL;
 749                 if (EXT4_SB(sb)->s_group_info)
 750                         grp = ext4_get_group_info(sb, i);
 751                 if (!grp || !EXT4_MB_GRP_BBITMAP_CORRUPT(grp))
 752                         desc_count += ext4_free_group_clusters(sb, gdp);
 753         }
 754
 755         return desc_count;
 756 #endif
 757 }
 758
 759 static inline int test_root(ext4_group_t a, int b)
 760 {
 761         while (1) {
 762                 if (a < b)
 763                         return 0;
 764                 if (a == b)
 765                         return 1;
 766                 if ((a % b) != 0)
 767                         return 0;
 768                 a = a / b;
 769         }
 770 }
 771
 772 /**
 773  *      ext4_bg_has_super - number of blocks used by the superblock in group
 774  *      @sb: superblock for filesystem
 775  *      @group: group number to check
 776  *
 777  *      Return the number of blocks used by the superblock (primary or backup)
 778  *      in this group.  Currently this will be only 0 or 1.
 779  */
 780 int ext4_bg_has_super(struct super_block *sb, ext4_group_t group)
 781 {
 782         struct ext4_super_block *es = EXT4_SB(sb)->s_es;
 783
 784         if (group == 0)
 785                 return 1;
 786         if (ext4_has_feature_sparse_super2(sb)) {
 787                 if (group == le32_to_cpu(es->s_backup_bgs[0]) ||
 788                     group == le32_to_cpu(es->s_backup_bgs[1]))
 789                         return 1;
 790                 return 0;
 791         }
 792         if ((group <= 1) || !ext4_has_feature_sparse_super(sb))
 793                 return 1;
 794         if (!(group & 1))
 795                 return 0;
 796         if (test_root(group, 3) || (test_root(group, 5)) ||
 797             test_root(group, 7))
 798                 return 1;
 799
 800         return 0;
 801 }
 802
 803 static unsigned long ext4_bg_num_gdb_meta(struct super_block *sb,
 804                                         ext4_group_t group)
 805 {
 806         unsigned long metagroup = group / EXT4_DESC_PER_BLOCK(sb);
 807         ext4_group_t first = metagroup * EXT4_DESC_PER_BLOCK(sb);
 808         ext4_group_t last = first + EXT4_DESC_PER_BLOCK(sb) - 1;
 809
 810         if (group == first || group == first + 1 || group == last)
 811                 return 1;
 812         return 0;
 813 }
 814
 815 static unsigned long ext4_bg_num_gdb_nometa(struct super_block *sb,
 816                                         ext4_group_t group)
 817 {
 818         if (!ext4_bg_has_super(sb, group))
 819                 return 0;
 820
 821         if (ext4_has_feature_meta_bg(sb))
 822                 return le32_to_cpu(EXT4_SB(sb)->s_es->s_first_meta_bg);
 823         else
 824                 return EXT4_SB(sb)->s_gdb_count;
 825 }
 826
 827 /**
 828  *      ext4_bg_num_gdb - number of blocks used by the group table in group
 829  *      @sb: superblock for filesystem
 830  *      @group: group number to check
 831  *
 832  *      Return the number of blocks used by the group descriptor table
 833  *      (primary or backup) in this group.  In the future there may be a
 834  *      different number of descriptor blocks in each group.
 835  */
 836 unsigned long ext4_bg_num_gdb(struct super_block *sb, ext4_group_t group)
 837 {
 838         unsigned long first_meta_bg =
 839                         le32_to_cpu(EXT4_SB(sb)->s_es->s_first_meta_bg);
 840         unsigned long metagroup = group / EXT4_DESC_PER_BLOCK(sb);
 841
 842         if (!ext4_has_feature_meta_bg(sb) || metagroup < first_meta_bg)
 843                 return ext4_bg_num_gdb_nometa(sb, group);
 844
 845         return ext4_bg_num_gdb_meta(sb,group);
 846
 847 }
 848
 849 /*
 850  * This function returns the number of file system metadata clusters at
 851  * the beginning of a block group, including the reserved gdt blocks.
 852  */
 853 static unsigned ext4_num_base_meta_clusters(struct super_block *sb,
 854                                      ext4_group_t block_group)
 855 {
 856         struct ext4_sb_info *sbi = EXT4_SB(sb);
 857         unsigned num;
 858
 859         /* Check for superblock and gdt backups in this group */
 860         num = ext4_bg_has_super(sb, block_group);
 861
 862         if (!ext4_has_feature_meta_bg(sb) ||
 863             block_group < le32_to_cpu(sbi->s_es->s_first_meta_bg) *
 864                           sbi->s_desc_per_block) {
 865                 if (num) {
 866                         num += ext4_bg_num_gdb(sb, block_group);
 867                         num += le16_to_cpu(sbi->s_es->s_reserved_gdt_blocks);
 868                 }
 869         } else { /* For META_BG_BLOCK_GROUPS */
 870                 num += ext4_bg_num_gdb(sb, block_group);
 871         }
 872         return EXT4_NUM_B2C(sbi, num);
 873 }
 874 /**
 875  *      ext4_inode_to_goal_block - return a hint for block allocation
 876  *      @inode: inode for block allocation
 877  *
 878  *      Return the ideal location to start allocating blocks for a
 879  *      newly created inode.
 880  */
 881 ext4_fsblk_t ext4_inode_to_goal_block(struct inode *inode)
 882 {
 883         struct ext4_inode_info *ei = EXT4_I(inode);
 884         ext4_group_t block_group;
 885         ext4_grpblk_t colour;
 886         int flex_size = ext4_flex_bg_size(EXT4_SB(inode->i_sb));
 887         ext4_fsblk_t bg_start;
 888         ext4_fsblk_t last_block;
 889
 890         block_group = ei->i_block_group;
 891         if (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) {
 892                 /*
 893                  * If there are at least EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME
 894                  * block groups per flexgroup, reserve the first block
 895                  * group for directories and special files.  Regular
 896                  * files will start at the second block group.  This
 897                  * tends to speed up directory access and improves
 898                  * fsck times.
 899                  */
 900                 block_group &= ~(flex_size-1);
 901                 if (S_ISREG(inode->i_mode))
 902                         block_group++;
 903         }
 904         bg_start = ext4_group_first_block_no(inode->i_sb, block_group);
 905         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
 906
 907         /*
 908          * If we are doing delayed allocation, we don't need take
 909          * colour into account.
 910          */
 911         if (test_opt(inode->i_sb, DELALLOC))
 912                 return bg_start;
 913
 914         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
 915                 colour = (current->pid % 16) *
 916                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
 917         else
 918                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
 919         return bg_start + colour;
 920 }
 921