git.oblomov.eu Git - linux-2.6/blob - fs/ext4/ialloc.c

   1 /*
   2  *  linux/fs/ext4/ialloc.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  *  BSD ufs-inspired inode and directory allocation by
  10  *  Stephen Tweedie (sct@redhat.com), 1993
  11  *  Big-endian to little-endian byte-swapping/bitmaps by
  12  *        David S. Miller (davem@caip.rutgers.edu), 1995
  13  */
  14
  15 #include <linux/time.h>
  16 #include <linux/fs.h>
  17 #include <linux/jbd2.h>
  18 #include <linux/stat.h>
  19 #include <linux/string.h>
  20 #include <linux/quotaops.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/random.h>
  23 #include <linux/bitops.h>
  24 #include <linux/blkdev.h>
  25 #include <asm/byteorder.h>
  26 #include "ext4.h"
  27 #include "ext4_jbd2.h"
  28 #include "xattr.h"
  29 #include "acl.h"
  30 #include "group.h"
  31
  32 /*
  33  * ialloc.c contains the inodes allocation and deallocation routines
  34  */
  35
  36 /*
  37  * The free inodes are managed by bitmaps.  A file system contains several
  38  * blocks groups.  Each group contains 1 bitmap block for blocks, 1 bitmap
  39  * block for inodes, N blocks for the inode table and data blocks.
  40  *
  41  * The file system contains group descriptors which are located after the
  42  * super block.  Each descriptor contains the number of the bitmap block and
  43  * the free blocks count in the block.
  44  */
  45
  46 /*
  47  * To avoid calling the atomic setbit hundreds or thousands of times, we only
  48  * need to use it within a single byte (to ensure we get endianness right).
  49  * We can use memset for the rest of the bitmap as there are no other users.
  50  */
  51 void mark_bitmap_end(int start_bit, int end_bit, char *bitmap)
  52 {
  53         int i;
  54
  55         if (start_bit >= end_bit)
  56                 return;
  57
  58         ext4_debug("mark end bits +%d through +%d used\n", start_bit, end_bit);
  59         for (i = start_bit; i < ((start_bit + 7) & ~7UL); i++)
  60                 ext4_set_bit(i, bitmap);
  61         if (i < end_bit)
  62                 memset(bitmap + (i >> 3), 0xff, (end_bit - i) >> 3);
  63 }
  64
  65 /* Initializes an uninitialized inode bitmap */
  66 unsigned ext4_init_inode_bitmap(struct super_block *sb, struct buffer_head *bh,
  67                                 ext4_group_t block_group,
  68                                 struct ext4_group_desc *gdp)
  69 {
  70         struct ext4_sb_info *sbi = EXT4_SB(sb);
  71
  72         J_ASSERT_BH(bh, buffer_locked(bh));
  73
  74         /* If checksum is bad mark all blocks and inodes use to prevent
  75          * allocation, essentially implementing a per-group read-only flag. */
  76         if (!ext4_group_desc_csum_verify(sbi, block_group, gdp)) {
  77                 ext4_error(sb, __func__, "Checksum bad for group %lu\n",
  78                            block_group);
  79                 gdp->bg_free_blocks_count = 0;
  80                 gdp->bg_free_inodes_count = 0;
  81                 gdp->bg_itable_unused = 0;
  82                 memset(bh->b_data, 0xff, sb->s_blocksize);
  83                 return 0;
  84         }
  85
  86         memset(bh->b_data, 0, (EXT4_INODES_PER_GROUP(sb) + 7) / 8);
  87         mark_bitmap_end(EXT4_INODES_PER_GROUP(sb), EXT4_BLOCKS_PER_GROUP(sb),
  88                         bh->b_data);
  89
  90         return EXT4_INODES_PER_GROUP(sb);
  91 }
  92
  93 /*
  94  * Read the inode allocation bitmap for a given block_group, reading
  95  * into the specified slot in the superblock's bitmap cache.
  96  *
  97  * Return buffer_head of bitmap on success or NULL.
  98  */
  99 static struct buffer_head *
 100 read_inode_bitmap(struct super_block *sb, ext4_group_t block_group)
 101 {
 102         struct ext4_group_desc *desc;
 103         struct buffer_head *bh = NULL;
 104
 105         desc = ext4_get_group_desc(sb, block_group, NULL);
 106         if (!desc)
 107                 goto error_out;
 108         if (desc->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) {
 109                 bh = sb_getblk(sb, ext4_inode_bitmap(sb, desc));
 110                 if (!buffer_uptodate(bh)) {
 111                         lock_buffer(bh);
 112                         if (!buffer_uptodate(bh)) {
 113                                 ext4_init_inode_bitmap(sb, bh, block_group,
 114                                                        desc);
 115                                 set_buffer_uptodate(bh);
 116                         }
 117                         unlock_buffer(bh);
 118                 }
 119         } else {
 120                 bh = sb_bread(sb, ext4_inode_bitmap(sb, desc));
 121         }
 122         if (!bh)
 123                 ext4_error(sb, "read_inode_bitmap",
 124                             "Cannot read inode bitmap - "
 125                             "block_group = %lu, inode_bitmap = %llu",
 126                             block_group, ext4_inode_bitmap(sb, desc));
 127 error_out:
 128         return bh;
 129 }
 130
 131 /*
 132  * NOTE! When we get the inode, we're the only people
 133  * that have access to it, and as such there are no
 134  * race conditions we have to worry about. The inode
 135  * is not on the hash-lists, and it cannot be reached
 136  * through the filesystem because the directory entry
 137  * has been deleted earlier.
 138  *
 139  * HOWEVER: we must make sure that we get no aliases,
 140  * which means that we have to call "clear_inode()"
 141  * _before_ we mark the inode not in use in the inode
 142  * bitmaps. Otherwise a newly created file might use
 143  * the same inode number (not actually the same pointer
 144  * though), and then we'd have two inodes sharing the
 145  * same inode number and space on the harddisk.
 146  */
 147 void ext4_free_inode (handle_t *handle, struct inode * inode)
 148 {
 149         struct super_block * sb = inode->i_sb;
 150         int is_directory;
 151         unsigned long ino;
 152         struct buffer_head *bitmap_bh = NULL;
 153         struct buffer_head *bh2;
 154         ext4_group_t block_group;
 155         unsigned long bit;
 156         struct ext4_group_desc * gdp;
 157         struct ext4_super_block * es;
 158         struct ext4_sb_info *sbi;
 159         int fatal = 0, err;
 160         ext4_group_t flex_group;
 161
 162         if (atomic_read(&inode->i_count) > 1) {
 163                 printk ("ext4_free_inode: inode has count=%d\n",
 164                                         atomic_read(&inode->i_count));
 165                 return;
 166         }
 167         if (inode->i_nlink) {
 168                 printk ("ext4_free_inode: inode has nlink=%d\n",
 169                         inode->i_nlink);
 170                 return;
 171         }
 172         if (!sb) {
 173                 printk("ext4_free_inode: inode on nonexistent device\n");
 174                 return;
 175         }
 176         sbi = EXT4_SB(sb);
 177
 178         ino = inode->i_ino;
 179         ext4_debug ("freeing inode %lu\n", ino);
 180
 181         /*
 182          * Note: we must free any quota before locking the superblock,
 183          * as writing the quota to disk may need the lock as well.
 184          */
 185         DQUOT_INIT(inode);
 186         ext4_xattr_delete_inode(handle, inode);
 187         DQUOT_FREE_INODE(inode);
 188         DQUOT_DROP(inode);
 189
 190         is_directory = S_ISDIR(inode->i_mode);
 191
 192         /* Do this BEFORE marking the inode not in use or returning an error */
 193         clear_inode (inode);
 194
 195         es = EXT4_SB(sb)->s_es;
 196         if (ino < EXT4_FIRST_INO(sb) || ino > le32_to_cpu(es->s_inodes_count)) {
 197                 ext4_error (sb, "ext4_free_inode",
 198                             "reserved or nonexistent inode %lu", ino);
 199                 goto error_return;
 200         }
 201         block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb);
 202         bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb);
 203         bitmap_bh = read_inode_bitmap(sb, block_group);
 204         if (!bitmap_bh)
 205                 goto error_return;
 206
 207         BUFFER_TRACE(bitmap_bh, "get_write_access");
 208         fatal = ext4_journal_get_write_access(handle, bitmap_bh);
 209         if (fatal)
 210                 goto error_return;
 211
 212         /* Ok, now we can actually update the inode bitmaps.. */
 213         if (!ext4_clear_bit_atomic(sb_bgl_lock(sbi, block_group),
 214                                         bit, bitmap_bh->b_data))
 215                 ext4_error (sb, "ext4_free_inode",
 216                               "bit already cleared for inode %lu", ino);
 217         else {
 218                 gdp = ext4_get_group_desc (sb, block_group, &bh2);
 219
 220                 BUFFER_TRACE(bh2, "get_write_access");
 221                 fatal = ext4_journal_get_write_access(handle, bh2);
 222                 if (fatal) goto error_return;
 223
 224                 if (gdp) {
 225                         spin_lock(sb_bgl_lock(sbi, block_group));
 226                         le16_add_cpu(&gdp->bg_free_inodes_count, 1);
 227                         if (is_directory)
 228                                 le16_add_cpu(&gdp->bg_used_dirs_count, -1);
 229                         gdp->bg_checksum = ext4_group_desc_csum(sbi,
 230                                                         block_group, gdp);
 231                         spin_unlock(sb_bgl_lock(sbi, block_group));
 232                         percpu_counter_inc(&sbi->s_freeinodes_counter);
 233                         if (is_directory)
 234                                 percpu_counter_dec(&sbi->s_dirs_counter);
 235
 236                         if (sbi->s_log_groups_per_flex) {
 237                                 flex_group = ext4_flex_group(sbi, block_group);
 238                                 spin_lock(sb_bgl_lock(sbi, flex_group));
 239                                 sbi->s_flex_groups[flex_group].free_inodes++;
 240                                 spin_unlock(sb_bgl_lock(sbi, flex_group));
 241                         }
 242                 }
 243                 BUFFER_TRACE(bh2, "call ext4_journal_dirty_metadata");
 244                 err = ext4_journal_dirty_metadata(handle, bh2);
 245                 if (!fatal) fatal = err;
 246         }
 247         BUFFER_TRACE(bitmap_bh, "call ext4_journal_dirty_metadata");
 248         err = ext4_journal_dirty_metadata(handle, bitmap_bh);
 249         if (!fatal)
 250                 fatal = err;
 251         sb->s_dirt = 1;
 252 error_return:
 253         brelse(bitmap_bh);
 254         ext4_std_error(sb, fatal);
 255 }
 256
 257 /*
 258  * There are two policies for allocating an inode.  If the new inode is
 259  * a directory, then a forward search is made for a block group with both
 260  * free space and a low directory-to-inode ratio; if that fails, then of
 261  * the groups with above-average free space, that group with the fewest
 262  * directories already is chosen.
 263  *
 264  * For other inodes, search forward from the parent directory\'s block
 265  * group to find a free inode.
 266  */
 267 static int find_group_dir(struct super_block *sb, struct inode *parent,
 268                                 ext4_group_t *best_group)
 269 {
 270         ext4_group_t ngroups = EXT4_SB(sb)->s_groups_count;
 271         unsigned int freei, avefreei;
 272         struct ext4_group_desc *desc, *best_desc = NULL;
 273         ext4_group_t group;
 274         int ret = -1;
 275
 276         freei = percpu_counter_read_positive(&EXT4_SB(sb)->s_freeinodes_counter);
 277         avefreei = freei / ngroups;
 278
 279         for (group = 0; group < ngroups; group++) {
 280                 desc = ext4_get_group_desc (sb, group, NULL);
 281                 if (!desc || !desc->bg_free_inodes_count)
 282                         continue;
 283                 if (le16_to_cpu(desc->bg_free_inodes_count) < avefreei)
 284                         continue;
 285                 if (!best_desc ||
 286                     (le16_to_cpu(desc->bg_free_blocks_count) >
 287                      le16_to_cpu(best_desc->bg_free_blocks_count))) {
 288                         *best_group = group;
 289                         best_desc = desc;
 290                         ret = 0;
 291                 }
 292         }
 293         return ret;
 294 }
 295
 296 #define free_block_ratio 10
 297
 298 static int find_group_flex(struct super_block *sb, struct inode *parent,
 299                            ext4_group_t *best_group)
 300 {
 301         struct ext4_sb_info *sbi = EXT4_SB(sb);
 302         struct ext4_group_desc *desc;
 303         struct buffer_head *bh;
 304         struct flex_groups *flex_group = sbi->s_flex_groups;
 305         ext4_group_t parent_group = EXT4_I(parent)->i_block_group;
 306         ext4_group_t parent_fbg_group = ext4_flex_group(sbi, parent_group);
 307         ext4_group_t ngroups = sbi->s_groups_count;
 308         int flex_size = ext4_flex_bg_size(sbi);
 309         ext4_group_t best_flex = parent_fbg_group;
 310         int blocks_per_flex = sbi->s_blocks_per_group * flex_size;
 311         int flexbg_free_blocks;
 312         int flex_freeb_ratio;
 313         ext4_group_t n_fbg_groups;
 314         ext4_group_t i;
 315
 316         n_fbg_groups = (sbi->s_groups_count + flex_size - 1) >>
 317                 sbi->s_log_groups_per_flex;
 318
 319 find_close_to_parent:
 320         flexbg_free_blocks = flex_group[best_flex].free_blocks;
 321         flex_freeb_ratio = flexbg_free_blocks * 100 / blocks_per_flex;
 322         if (flex_group[best_flex].free_inodes &&
 323             flex_freeb_ratio > free_block_ratio)
 324                 goto found_flexbg;
 325
 326         if (best_flex && best_flex == parent_fbg_group) {
 327                 best_flex--;
 328                 goto find_close_to_parent;
 329         }
 330
 331         for (i = 0; i < n_fbg_groups; i++) {
 332                 if (i == parent_fbg_group || i == parent_fbg_group - 1)
 333                         continue;
 334
 335                 flexbg_free_blocks = flex_group[i].free_blocks;
 336                 flex_freeb_ratio = flexbg_free_blocks * 100 / blocks_per_flex;
 337
 338                 if (flex_freeb_ratio > free_block_ratio &&
 339                     flex_group[i].free_inodes) {
 340                         best_flex = i;
 341                         goto found_flexbg;
 342                 }
 343
 344                 if (best_flex < 0 ||
 345                     (flex_group[i].free_blocks >
 346                      flex_group[best_flex].free_blocks &&
 347                      flex_group[i].free_inodes))
 348                         best_flex = i;
 349         }
 350
 351         if (!flex_group[best_flex].free_inodes ||
 352             !flex_group[best_flex].free_blocks)
 353                 return -1;
 354
 355 found_flexbg:
 356         for (i = best_flex * flex_size; i < ngroups &&
 357                      i < (best_flex + 1) * flex_size; i++) {
 358                 desc = ext4_get_group_desc(sb, i, &bh);
 359                 if (le16_to_cpu(desc->bg_free_inodes_count)) {
 360                         *best_group = i;
 361                         goto out;
 362                 }
 363         }
 364
 365         return -1;
 366 out:
 367         return 0;
 368 }
 369
 370 /*
 371  * Orlov's allocator for directories.
 372  *
 373  * We always try to spread first-level directories.
 374  *
 375  * If there are blockgroups with both free inodes and free blocks counts
 376  * not worse than average we return one with smallest directory count.
 377  * Otherwise we simply return a random group.
 378  *
 379  * For the rest rules look so:
 380  *
 381  * It's OK to put directory into a group unless
 382  * it has too many directories already (max_dirs) or
 383  * it has too few free inodes left (min_inodes) or
 384  * it has too few free blocks left (min_blocks) or
 385  * it's already running too large debt (max_debt).
 386  * Parent's group is preferred, if it doesn't satisfy these
 387  * conditions we search cyclically through the rest. If none
 388  * of the groups look good we just look for a group with more
 389  * free inodes than average (starting at parent's group).
 390  *
 391  * Debt is incremented each time we allocate a directory and decremented
 392  * when we allocate an inode, within 0--255.
 393  */
 394
 395 #define INODE_COST 64
 396 #define BLOCK_COST 256
 397
 398 static int find_group_orlov(struct super_block *sb, struct inode *parent,
 399                                 ext4_group_t *group)
 400 {
 401         ext4_group_t parent_group = EXT4_I(parent)->i_block_group;
 402         struct ext4_sb_info *sbi = EXT4_SB(sb);
 403         struct ext4_super_block *es = sbi->s_es;
 404         ext4_group_t ngroups = sbi->s_groups_count;
 405         int inodes_per_group = EXT4_INODES_PER_GROUP(sb);
 406         unsigned int freei, avefreei;
 407         ext4_fsblk_t freeb, avefreeb;
 408         ext4_fsblk_t blocks_per_dir;
 409         unsigned int ndirs;
 410         int max_debt, max_dirs, min_inodes;
 411         ext4_grpblk_t min_blocks;
 412         ext4_group_t i;
 413         struct ext4_group_desc *desc;
 414
 415         freei = percpu_counter_read_positive(&sbi->s_freeinodes_counter);
 416         avefreei = freei / ngroups;
 417         freeb = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
 418         avefreeb = freeb;
 419         do_div(avefreeb, ngroups);
 420         ndirs = percpu_counter_read_positive(&sbi->s_dirs_counter);
 421
 422         if ((parent == sb->s_root->d_inode) ||
 423             (EXT4_I(parent)->i_flags & EXT4_TOPDIR_FL)) {
 424                 int best_ndir = inodes_per_group;
 425                 ext4_group_t grp;
 426                 int ret = -1;
 427
 428                 get_random_bytes(&grp, sizeof(grp));
 429                 parent_group = (unsigned)grp % ngroups;
 430                 for (i = 0; i < ngroups; i++) {
 431                         grp = (parent_group + i) % ngroups;
 432                         desc = ext4_get_group_desc(sb, grp, NULL);
 433                         if (!desc || !desc->bg_free_inodes_count)
 434                                 continue;
 435                         if (le16_to_cpu(desc->bg_used_dirs_count) >= best_ndir)
 436                                 continue;
 437                         if (le16_to_cpu(desc->bg_free_inodes_count) < avefreei)
 438                                 continue;
 439                         if (le16_to_cpu(desc->bg_free_blocks_count) < avefreeb)
 440                                 continue;
 441                         *group = grp;
 442                         ret = 0;
 443                         best_ndir = le16_to_cpu(desc->bg_used_dirs_count);
 444                 }
 445                 if (ret == 0)
 446                         return ret;
 447                 goto fallback;
 448         }
 449
 450         blocks_per_dir = ext4_blocks_count(es) - freeb;
 451         do_div(blocks_per_dir, ndirs);
 452
 453         max_dirs = ndirs / ngroups + inodes_per_group / 16;
 454         min_inodes = avefreei - inodes_per_group / 4;
 455         min_blocks = avefreeb - EXT4_BLOCKS_PER_GROUP(sb) / 4;
 456
 457         max_debt = EXT4_BLOCKS_PER_GROUP(sb);
 458         max_debt /= max_t(int, blocks_per_dir, BLOCK_COST);
 459         if (max_debt * INODE_COST > inodes_per_group)
 460                 max_debt = inodes_per_group / INODE_COST;
 461         if (max_debt > 255)
 462                 max_debt = 255;
 463         if (max_debt == 0)
 464                 max_debt = 1;
 465
 466         for (i = 0; i < ngroups; i++) {
 467                 *group = (parent_group + i) % ngroups;
 468                 desc = ext4_get_group_desc(sb, *group, NULL);
 469                 if (!desc || !desc->bg_free_inodes_count)
 470                         continue;
 471                 if (le16_to_cpu(desc->bg_used_dirs_count) >= max_dirs)
 472                         continue;
 473                 if (le16_to_cpu(desc->bg_free_inodes_count) < min_inodes)
 474                         continue;
 475                 if (le16_to_cpu(desc->bg_free_blocks_count) < min_blocks)
 476                         continue;
 477                 return 0;
 478         }
 479
 480 fallback:
 481         for (i = 0; i < ngroups; i++) {
 482                 *group = (parent_group + i) % ngroups;
 483                 desc = ext4_get_group_desc(sb, *group, NULL);
 484                 if (desc && desc->bg_free_inodes_count &&
 485                         le16_to_cpu(desc->bg_free_inodes_count) >= avefreei)
 486                         return 0;
 487         }
 488
 489         if (avefreei) {
 490                 /*
 491                  * The free-inodes counter is approximate, and for really small
 492                  * filesystems the above test can fail to find any blockgroups
 493                  */
 494                 avefreei = 0;
 495                 goto fallback;
 496         }
 497
 498         return -1;
 499 }
 500
 501 static int find_group_other(struct super_block *sb, struct inode *parent,
 502                                 ext4_group_t *group)
 503 {
 504         ext4_group_t parent_group = EXT4_I(parent)->i_block_group;
 505         ext4_group_t ngroups = EXT4_SB(sb)->s_groups_count;
 506         struct ext4_group_desc *desc;
 507         ext4_group_t i;
 508
 509         /*
 510          * Try to place the inode in its parent directory
 511          */
 512         *group = parent_group;
 513         desc = ext4_get_group_desc(sb, *group, NULL);
 514         if (desc && le16_to_cpu(desc->bg_free_inodes_count) &&
 515                         le16_to_cpu(desc->bg_free_blocks_count))
 516                 return 0;
 517
 518         /*
 519          * We're going to place this inode in a different blockgroup from its
 520          * parent.  We want to cause files in a common directory to all land in
 521          * the same blockgroup.  But we want files which are in a different
 522          * directory which shares a blockgroup with our parent to land in a
 523          * different blockgroup.
 524          *
 525          * So add our directory's i_ino into the starting point for the hash.
 526          */
 527         *group = (*group + parent->i_ino) % ngroups;
 528
 529         /*
 530          * Use a quadratic hash to find a group with a free inode and some free
 531          * blocks.
 532          */
 533         for (i = 1; i < ngroups; i <<= 1) {
 534                 *group += i;
 535                 if (*group >= ngroups)
 536                         *group -= ngroups;
 537                 desc = ext4_get_group_desc(sb, *group, NULL);
 538                 if (desc && le16_to_cpu(desc->bg_free_inodes_count) &&
 539                                 le16_to_cpu(desc->bg_free_blocks_count))
 540                         return 0;
 541         }
 542
 543         /*
 544          * That failed: try linear search for a free inode, even if that group
 545          * has no free blocks.
 546          */
 547         *group = parent_group;
 548         for (i = 0; i < ngroups; i++) {
 549                 if (++*group >= ngroups)
 550                         *group = 0;
 551                 desc = ext4_get_group_desc(sb, *group, NULL);
 552                 if (desc && le16_to_cpu(desc->bg_free_inodes_count))
 553                         return 0;
 554         }
 555
 556         return -1;
 557 }
 558
 559 /*
 560  * There are two policies for allocating an inode.  If the new inode is
 561  * a directory, then a forward search is made for a block group with both
 562  * free space and a low directory-to-inode ratio; if that fails, then of
 563  * the groups with above-average free space, that group with the fewest
 564  * directories already is chosen.
 565  *
 566  * For other inodes, search forward from the parent directory's block
 567  * group to find a free inode.
 568  */
 569 struct inode *ext4_new_inode(handle_t *handle, struct inode * dir, int mode)
 570 {
 571         struct super_block *sb;
 572         struct buffer_head *bitmap_bh = NULL;
 573         struct buffer_head *bh2;
 574         ext4_group_t group = 0;
 575         unsigned long ino = 0;
 576         struct inode * inode;
 577         struct ext4_group_desc * gdp = NULL;
 578         struct ext4_super_block * es;
 579         struct ext4_inode_info *ei;
 580         struct ext4_sb_info *sbi;
 581         int ret2, err = 0;
 582         struct inode *ret;
 583         ext4_group_t i;
 584         int free = 0;
 585         ext4_group_t flex_group;
 586
 587         /* Cannot create files in a deleted directory */
 588         if (!dir || !dir->i_nlink)
 589                 return ERR_PTR(-EPERM);
 590
 591         sb = dir->i_sb;
 592         inode = new_inode(sb);
 593         if (!inode)
 594                 return ERR_PTR(-ENOMEM);
 595         ei = EXT4_I(inode);
 596
 597         sbi = EXT4_SB(sb);
 598         es = sbi->s_es;
 599
 600         if (sbi->s_log_groups_per_flex) {
 601                 ret2 = find_group_flex(sb, dir, &group);
 602                 goto got_group;
 603         }
 604
 605         if (S_ISDIR(mode)) {
 606                 if (test_opt (sb, OLDALLOC))
 607                         ret2 = find_group_dir(sb, dir, &group);
 608                 else
 609                         ret2 = find_group_orlov(sb, dir, &group);
 610         } else
 611                 ret2 = find_group_other(sb, dir, &group);
 612
 613 got_group:
 614         err = -ENOSPC;
 615         if (ret2 == -1)
 616                 goto out;
 617
 618         for (i = 0; i < sbi->s_groups_count; i++) {
 619                 err = -EIO;
 620
 621                 gdp = ext4_get_group_desc(sb, group, &bh2);
 622                 if (!gdp)
 623                         goto fail;
 624
 625                 brelse(bitmap_bh);
 626                 bitmap_bh = read_inode_bitmap(sb, group);
 627                 if (!bitmap_bh)
 628                         goto fail;
 629
 630                 ino = 0;
 631
 632 repeat_in_this_group:
 633                 ino = ext4_find_next_zero_bit((unsigned long *)
 634                                 bitmap_bh->b_data, EXT4_INODES_PER_GROUP(sb), ino);
 635                 if (ino < EXT4_INODES_PER_GROUP(sb)) {
 636
 637                         BUFFER_TRACE(bitmap_bh, "get_write_access");
 638                         err = ext4_journal_get_write_access(handle, bitmap_bh);
 639                         if (err)
 640                                 goto fail;
 641
 642                         if (!ext4_set_bit_atomic(sb_bgl_lock(sbi, group),
 643                                                 ino, bitmap_bh->b_data)) {
 644                                 /* we won it */
 645                                 BUFFER_TRACE(bitmap_bh,
 646                                         "call ext4_journal_dirty_metadata");
 647                                 err = ext4_journal_dirty_metadata(handle,
 648                                                                 bitmap_bh);
 649                                 if (err)
 650                                         goto fail;
 651                                 goto got;
 652                         }
 653                         /* we lost it */
 654                         jbd2_journal_release_buffer(handle, bitmap_bh);
 655
 656                         if (++ino < EXT4_INODES_PER_GROUP(sb))
 657                                 goto repeat_in_this_group;
 658                 }
 659
 660                 /*
 661                  * This case is possible in concurrent environment.  It is very
 662                  * rare.  We cannot repeat the find_group_xxx() call because
 663                  * that will simply return the same blockgroup, because the
 664                  * group descriptor metadata has not yet been updated.
 665                  * So we just go onto the next blockgroup.
 666                  */
 667                 if (++group == sbi->s_groups_count)
 668                         group = 0;
 669         }
 670         err = -ENOSPC;
 671         goto out;
 672
 673 got:
 674         ino++;
 675         if ((group == 0 && ino < EXT4_FIRST_INO(sb)) ||
 676             ino > EXT4_INODES_PER_GROUP(sb)) {
 677                 ext4_error(sb, __func__,
 678                            "reserved inode or inode > inodes count - "
 679                            "block_group = %lu, inode=%lu", group,
 680                            ino + group * EXT4_INODES_PER_GROUP(sb));
 681                 err = -EIO;
 682                 goto fail;
 683         }
 684
 685         BUFFER_TRACE(bh2, "get_write_access");
 686         err = ext4_journal_get_write_access(handle, bh2);
 687         if (err) goto fail;
 688
 689         /* We may have to initialize the block bitmap if it isn't already */
 690         if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM) &&
 691             gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
 692                 struct buffer_head *block_bh = ext4_read_block_bitmap(sb, group);
 693
 694                 BUFFER_TRACE(block_bh, "get block bitmap access");
 695                 err = ext4_journal_get_write_access(handle, block_bh);
 696                 if (err) {
 697                         brelse(block_bh);
 698                         goto fail;
 699                 }
 700
 701                 free = 0;
 702                 spin_lock(sb_bgl_lock(sbi, group));
 703                 /* recheck and clear flag under lock if we still need to */
 704                 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
 705                         gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
 706                         free = ext4_free_blocks_after_init(sb, group, gdp);
 707                         gdp->bg_free_blocks_count = cpu_to_le16(free);
 708                 }
 709                 spin_unlock(sb_bgl_lock(sbi, group));
 710
 711                 /* Don't need to dirty bitmap block if we didn't change it */
 712                 if (free) {
 713                         BUFFER_TRACE(block_bh, "dirty block bitmap");
 714                         err = ext4_journal_dirty_metadata(handle, block_bh);
 715                 }
 716
 717                 brelse(block_bh);
 718                 if (err)
 719                         goto fail;
 720         }
 721
 722         spin_lock(sb_bgl_lock(sbi, group));
 723         /* If we didn't allocate from within the initialized part of the inode
 724          * table then we need to initialize up to this inode. */
 725         if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM)) {
 726                 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) {
 727                         gdp->bg_flags &= cpu_to_le16(~EXT4_BG_INODE_UNINIT);
 728
 729                         /* When marking the block group with
 730                          * ~EXT4_BG_INODE_UNINIT we don't want to depend
 731                          * on the value of bg_itable_unsed even though
 732                          * mke2fs could have initialized the same for us.
 733                          * Instead we calculated the value below
 734                          */
 735
 736                         free = 0;
 737                 } else {
 738                         free = EXT4_INODES_PER_GROUP(sb) -
 739                                 le16_to_cpu(gdp->bg_itable_unused);
 740                 }
 741
 742                 /*
 743                  * Check the relative inode number against the last used
 744                  * relative inode number in this group. if it is greater
 745                  * we need to  update the bg_itable_unused count
 746                  *
 747                  */
 748                 if (ino > free)
 749                         gdp->bg_itable_unused =
 750                                 cpu_to_le16(EXT4_INODES_PER_GROUP(sb) - ino);
 751         }
 752
 753         le16_add_cpu(&gdp->bg_free_inodes_count, -1);
 754         if (S_ISDIR(mode)) {
 755                 le16_add_cpu(&gdp->bg_used_dirs_count, 1);
 756         }
 757         gdp->bg_checksum = ext4_group_desc_csum(sbi, group, gdp);
 758         spin_unlock(sb_bgl_lock(sbi, group));
 759         BUFFER_TRACE(bh2, "call ext4_journal_dirty_metadata");
 760         err = ext4_journal_dirty_metadata(handle, bh2);
 761         if (err) goto fail;
 762
 763         percpu_counter_dec(&sbi->s_freeinodes_counter);
 764         if (S_ISDIR(mode))
 765                 percpu_counter_inc(&sbi->s_dirs_counter);
 766         sb->s_dirt = 1;
 767
 768         if (sbi->s_log_groups_per_flex) {
 769                 flex_group = ext4_flex_group(sbi, group);
 770                 spin_lock(sb_bgl_lock(sbi, flex_group));
 771                 sbi->s_flex_groups[flex_group].free_inodes--;
 772                 spin_unlock(sb_bgl_lock(sbi, flex_group));
 773         }
 774
 775         inode->i_uid = current->fsuid;
 776         if (test_opt (sb, GRPID))
 777                 inode->i_gid = dir->i_gid;
 778         else if (dir->i_mode & S_ISGID) {
 779                 inode->i_gid = dir->i_gid;
 780                 if (S_ISDIR(mode))
 781                         mode |= S_ISGID;
 782         } else
 783                 inode->i_gid = current->fsgid;
 784         inode->i_mode = mode;
 785
 786         inode->i_ino = ino + group * EXT4_INODES_PER_GROUP(sb);
 787         /* This is the optimal IO size (for stat), not the fs block size */
 788         inode->i_blocks = 0;
 789         inode->i_mtime = inode->i_atime = inode->i_ctime = ei->i_crtime =
 790                                                        ext4_current_time(inode);
 791
 792         memset(ei->i_data, 0, sizeof(ei->i_data));
 793         ei->i_dir_start_lookup = 0;
 794         ei->i_disksize = 0;
 795
 796         /*
 797          * Don't inherit extent flag from directory. We set extent flag on
 798          * newly created directory and file only if -o extent mount option is
 799          * specified
 800          */
 801         ei->i_flags = EXT4_I(dir)->i_flags & ~(EXT4_INDEX_FL|EXT4_EXTENTS_FL);
 802         if (S_ISLNK(mode))
 803                 ei->i_flags &= ~(EXT4_IMMUTABLE_FL|EXT4_APPEND_FL);
 804         /* dirsync only applies to directories */
 805         if (!S_ISDIR(mode))
 806                 ei->i_flags &= ~EXT4_DIRSYNC_FL;
 807         ei->i_file_acl = 0;
 808         ei->i_dtime = 0;
 809         ei->i_block_alloc_info = NULL;
 810         ei->i_block_group = group;
 811
 812         ext4_set_inode_flags(inode);
 813         if (IS_DIRSYNC(inode))
 814                 handle->h_sync = 1;
 815         insert_inode_hash(inode);
 816         spin_lock(&sbi->s_next_gen_lock);
 817         inode->i_generation = sbi->s_next_generation++;
 818         spin_unlock(&sbi->s_next_gen_lock);
 819
 820         ei->i_state = EXT4_STATE_NEW;
 821
 822         ei->i_extra_isize = EXT4_SB(sb)->s_want_extra_isize;
 823
 824         ret = inode;
 825         if(DQUOT_ALLOC_INODE(inode)) {
 826                 err = -EDQUOT;
 827                 goto fail_drop;
 828         }
 829
 830         err = ext4_init_acl(handle, inode, dir);
 831         if (err)
 832                 goto fail_free_drop;
 833
 834         err = ext4_init_security(handle,inode, dir);
 835         if (err)
 836                 goto fail_free_drop;
 837
 838         if (test_opt(sb, EXTENTS)) {
 839                 /* set extent flag only for directory, file and normal symlink*/
 840                 if (S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode)) {
 841                         EXT4_I(inode)->i_flags |= EXT4_EXTENTS_FL;
 842                         ext4_ext_tree_init(handle, inode);
 843                 }
 844         }
 845
 846         err = ext4_mark_inode_dirty(handle, inode);
 847         if (err) {
 848                 ext4_std_error(sb, err);
 849                 goto fail_free_drop;
 850         }
 851
 852         ext4_debug("allocating inode %lu\n", inode->i_ino);
 853         goto really_out;
 854 fail:
 855         ext4_std_error(sb, err);
 856 out:
 857         iput(inode);
 858         ret = ERR_PTR(err);
 859 really_out:
 860         brelse(bitmap_bh);
 861         return ret;
 862
 863 fail_free_drop:
 864         DQUOT_FREE_INODE(inode);
 865
 866 fail_drop:
 867         DQUOT_DROP(inode);
 868         inode->i_flags |= S_NOQUOTA;
 869         inode->i_nlink = 0;
 870         iput(inode);
 871         brelse(bitmap_bh);
 872         return ERR_PTR(err);
 873 }
 874
 875 /* Verify that we are loading a valid orphan from disk */
 876 struct inode *ext4_orphan_get(struct super_block *sb, unsigned long ino)
 877 {
 878         unsigned long max_ino = le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count);
 879         ext4_group_t block_group;
 880         int bit;
 881         struct buffer_head *bitmap_bh;
 882         struct inode *inode = NULL;
 883         long err = -EIO;
 884
 885         /* Error cases - e2fsck has already cleaned up for us */
 886         if (ino > max_ino) {
 887                 ext4_warning(sb, __func__,
 888                              "bad orphan ino %lu!  e2fsck was run?", ino);
 889                 goto error;
 890         }
 891
 892         block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb);
 893         bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb);
 894         bitmap_bh = read_inode_bitmap(sb, block_group);
 895         if (!bitmap_bh) {
 896                 ext4_warning(sb, __func__,
 897                              "inode bitmap error for orphan %lu", ino);
 898                 goto error;
 899         }
 900
 901         /* Having the inode bit set should be a 100% indicator that this
 902          * is a valid orphan (no e2fsck run on fs).  Orphans also include
 903          * inodes that were being truncated, so we can't check i_nlink==0.
 904          */
 905         if (!ext4_test_bit(bit, bitmap_bh->b_data))
 906                 goto bad_orphan;
 907
 908         inode = ext4_iget(sb, ino);
 909         if (IS_ERR(inode))
 910                 goto iget_failed;
 911
 912         /*
 913          * If the orphans has i_nlinks > 0 then it should be able to be
 914          * truncated, otherwise it won't be removed from the orphan list
 915          * during processing and an infinite loop will result.
 916          */
 917         if (inode->i_nlink && !ext4_can_truncate(inode))
 918                 goto bad_orphan;
 919
 920         if (NEXT_ORPHAN(inode) > max_ino)
 921                 goto bad_orphan;
 922         brelse(bitmap_bh);
 923         return inode;
 924
 925 iget_failed:
 926         err = PTR_ERR(inode);
 927         inode = NULL;
 928 bad_orphan:
 929         ext4_warning(sb, __func__,
 930                      "bad orphan inode %lu!  e2fsck was run?", ino);
 931         printk(KERN_NOTICE "ext4_test_bit(bit=%d, block=%llu) = %d\n",
 932                bit, (unsigned long long)bitmap_bh->b_blocknr,
 933                ext4_test_bit(bit, bitmap_bh->b_data));
 934         printk(KERN_NOTICE "inode=%p\n", inode);
 935         if (inode) {
 936                 printk(KERN_NOTICE "is_bad_inode(inode)=%d\n",
 937                        is_bad_inode(inode));
 938                 printk(KERN_NOTICE "NEXT_ORPHAN(inode)=%u\n",
 939                        NEXT_ORPHAN(inode));
 940                 printk(KERN_NOTICE "max_ino=%lu\n", max_ino);
 941                 printk(KERN_NOTICE "i_nlink=%u\n", inode->i_nlink);
 942                 /* Avoid freeing blocks if we got a bad deleted inode */
 943                 if (inode->i_nlink == 0)
 944                         inode->i_blocks = 0;
 945                 iput(inode);
 946         }
 947         brelse(bitmap_bh);
 948 error:
 949         return ERR_PTR(err);
 950 }
 951
 952 unsigned long ext4_count_free_inodes (struct super_block * sb)
 953 {
 954         unsigned long desc_count;
 955         struct ext4_group_desc *gdp;
 956         ext4_group_t i;
 957 #ifdef EXT4FS_DEBUG
 958         struct ext4_super_block *es;
 959         unsigned long bitmap_count, x;
 960         struct buffer_head *bitmap_bh = NULL;
 961
 962         es = EXT4_SB(sb)->s_es;
 963         desc_count = 0;
 964         bitmap_count = 0;
 965         gdp = NULL;
 966         for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
 967                 gdp = ext4_get_group_desc (sb, i, NULL);
 968                 if (!gdp)
 969                         continue;
 970                 desc_count += le16_to_cpu(gdp->bg_free_inodes_count);
 971                 brelse(bitmap_bh);
 972                 bitmap_bh = read_inode_bitmap(sb, i);
 973                 if (!bitmap_bh)
 974                         continue;
 975
 976                 x = ext4_count_free(bitmap_bh, EXT4_INODES_PER_GROUP(sb) / 8);
 977                 printk(KERN_DEBUG "group %lu: stored = %d, counted = %lu\n",
 978                         i, le16_to_cpu(gdp->bg_free_inodes_count), x);
 979                 bitmap_count += x;
 980         }
 981         brelse(bitmap_bh);
 982         printk("ext4_count_free_inodes: stored = %u, computed = %lu, %lu\n",
 983                 le32_to_cpu(es->s_free_inodes_count), desc_count, bitmap_count);
 984         return desc_count;
 985 #else
 986         desc_count = 0;
 987         for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
 988                 gdp = ext4_get_group_desc (sb, i, NULL);
 989                 if (!gdp)
 990                         continue;
 991                 desc_count += le16_to_cpu(gdp->bg_free_inodes_count);
 992                 cond_resched();
 993         }
 994         return desc_count;
 995 #endif
 996 }
 997
 998 /* Called at mount-time, super-block is locked */
 999 unsigned long ext4_count_dirs (struct super_block * sb)
1000 {
1001         unsigned long count = 0;
1002         ext4_group_t i;
1003
1004         for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
1005                 struct ext4_group_desc *gdp = ext4_get_group_desc (sb, i, NULL);
1006                 if (!gdp)
1007                         continue;
1008                 count += le16_to_cpu(gdp->bg_used_dirs_count);
1009         }
1010         return count;
1011 }
1012