SAFE public projects git trees. - safe/jmp/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 ext4_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         ext4_fsblk_t bitmap_blk;
 105
 106         desc = ext4_get_group_desc(sb, block_group, NULL);
 107         if (!desc)
 108                 return NULL;
 109         bitmap_blk = ext4_inode_bitmap(sb, desc);
 110         bh = sb_getblk(sb, bitmap_blk);
 111         if (unlikely(!bh)) {
 112                 ext4_error(sb, __func__,
 113                             "Cannot read inode bitmap - "
 114                             "block_group = %lu, inode_bitmap = %llu",
 115                             block_group, bitmap_blk);
 116                 return NULL;
 117         }
 118         if (bh_uptodate_or_lock(bh))
 119                 return bh;
 120
 121         if (desc->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) {
 122                 ext4_init_inode_bitmap(sb, bh, block_group, desc);
 123                 set_buffer_uptodate(bh);
 124                 unlock_buffer(bh);
 125                 return bh;
 126         }
 127         if (bh_submit_read(bh) < 0) {
 128                 put_bh(bh);
 129                 ext4_error(sb, __func__,
 130                             "Cannot read inode bitmap - "
 131                             "block_group = %lu, inode_bitmap = %llu",
 132                             block_group, bitmap_blk);
 133                 return NULL;
 134         }
 135         return bh;
 136 }
 137
 138 /*
 139  * NOTE! When we get the inode, we're the only people
 140  * that have access to it, and as such there are no
 141  * race conditions we have to worry about. The inode
 142  * is not on the hash-lists, and it cannot be reached
 143  * through the filesystem because the directory entry
 144  * has been deleted earlier.
 145  *
 146  * HOWEVER: we must make sure that we get no aliases,
 147  * which means that we have to call "clear_inode()"
 148  * _before_ we mark the inode not in use in the inode
 149  * bitmaps. Otherwise a newly created file might use
 150  * the same inode number (not actually the same pointer
 151  * though), and then we'd have two inodes sharing the
 152  * same inode number and space on the harddisk.
 153  */
 154 void ext4_free_inode (handle_t *handle, struct inode * inode)
 155 {
 156         struct super_block * sb = inode->i_sb;
 157         int is_directory;
 158         unsigned long ino;
 159         struct buffer_head *bitmap_bh = NULL;
 160         struct buffer_head *bh2;
 161         ext4_group_t block_group;
 162         unsigned long bit;
 163         struct ext4_group_desc * gdp;
 164         struct ext4_super_block * es;
 165         struct ext4_sb_info *sbi;
 166         int fatal = 0, err;
 167         ext4_group_t flex_group;
 168
 169         if (atomic_read(&inode->i_count) > 1) {
 170                 printk ("ext4_free_inode: inode has count=%d\n",
 171                                         atomic_read(&inode->i_count));
 172                 return;
 173         }
 174         if (inode->i_nlink) {
 175                 printk ("ext4_free_inode: inode has nlink=%d\n",
 176                         inode->i_nlink);
 177                 return;
 178         }
 179         if (!sb) {
 180                 printk("ext4_free_inode: inode on nonexistent device\n");
 181                 return;
 182         }
 183         sbi = EXT4_SB(sb);
 184
 185         ino = inode->i_ino;
 186         ext4_debug ("freeing inode %lu\n", ino);
 187
 188         /*
 189          * Note: we must free any quota before locking the superblock,
 190          * as writing the quota to disk may need the lock as well.
 191          */
 192         DQUOT_INIT(inode);
 193         ext4_xattr_delete_inode(handle, inode);
 194         DQUOT_FREE_INODE(inode);
 195         DQUOT_DROP(inode);
 196
 197         is_directory = S_ISDIR(inode->i_mode);
 198
 199         /* Do this BEFORE marking the inode not in use or returning an error */
 200         clear_inode (inode);
 201
 202         es = EXT4_SB(sb)->s_es;
 203         if (ino < EXT4_FIRST_INO(sb) || ino > le32_to_cpu(es->s_inodes_count)) {
 204                 ext4_error (sb, "ext4_free_inode",
 205                             "reserved or nonexistent inode %lu", ino);
 206                 goto error_return;
 207         }
 208         block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb);
 209         bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb);
 210         bitmap_bh = ext4_read_inode_bitmap(sb, block_group);
 211         if (!bitmap_bh)
 212                 goto error_return;
 213
 214         BUFFER_TRACE(bitmap_bh, "get_write_access");
 215         fatal = ext4_journal_get_write_access(handle, bitmap_bh);
 216         if (fatal)
 217                 goto error_return;
 218
 219         /* Ok, now we can actually update the inode bitmaps.. */
 220         if (!ext4_clear_bit_atomic(sb_bgl_lock(sbi, block_group),
 221                                         bit, bitmap_bh->b_data))
 222                 ext4_error (sb, "ext4_free_inode",
 223                               "bit already cleared for inode %lu", ino);
 224         else {
 225                 gdp = ext4_get_group_desc (sb, block_group, &bh2);
 226
 227                 BUFFER_TRACE(bh2, "get_write_access");
 228                 fatal = ext4_journal_get_write_access(handle, bh2);
 229                 if (fatal) goto error_return;
 230
 231                 if (gdp) {
 232                         spin_lock(sb_bgl_lock(sbi, block_group));
 233                         le16_add_cpu(&gdp->bg_free_inodes_count, 1);
 234                         if (is_directory)
 235                                 le16_add_cpu(&gdp->bg_used_dirs_count, -1);
 236                         gdp->bg_checksum = ext4_group_desc_csum(sbi,
 237                                                         block_group, gdp);
 238                         spin_unlock(sb_bgl_lock(sbi, block_group));
 239                         percpu_counter_inc(&sbi->s_freeinodes_counter);
 240                         if (is_directory)
 241                                 percpu_counter_dec(&sbi->s_dirs_counter);
 242
 243                         if (sbi->s_log_groups_per_flex) {
 244                                 flex_group = ext4_flex_group(sbi, block_group);
 245                                 spin_lock(sb_bgl_lock(sbi, flex_group));
 246                                 sbi->s_flex_groups[flex_group].free_inodes++;
 247                                 spin_unlock(sb_bgl_lock(sbi, flex_group));
 248                         }
 249                 }
 250                 BUFFER_TRACE(bh2, "call ext4_journal_dirty_metadata");
 251                 err = ext4_journal_dirty_metadata(handle, bh2);
 252                 if (!fatal) fatal = err;
 253         }
 254         BUFFER_TRACE(bitmap_bh, "call ext4_journal_dirty_metadata");
 255         err = ext4_journal_dirty_metadata(handle, bitmap_bh);
 256         if (!fatal)
 257                 fatal = err;
 258         sb->s_dirt = 1;
 259 error_return:
 260         brelse(bitmap_bh);
 261         ext4_std_error(sb, fatal);
 262 }
 263
 264 /*
 265  * There are two policies for allocating an inode.  If the new inode is
 266  * a directory, then a forward search is made for a block group with both
 267  * free space and a low directory-to-inode ratio; if that fails, then of
 268  * the groups with above-average free space, that group with the fewest
 269  * directories already is chosen.
 270  *
 271  * For other inodes, search forward from the parent directory\'s block
 272  * group to find a free inode.
 273  */
 274 static int find_group_dir(struct super_block *sb, struct inode *parent,
 275                                 ext4_group_t *best_group)
 276 {
 277         ext4_group_t ngroups = EXT4_SB(sb)->s_groups_count;
 278         unsigned int freei, avefreei;
 279         struct ext4_group_desc *desc, *best_desc = NULL;
 280         ext4_group_t group;
 281         int ret = -1;
 282
 283         freei = percpu_counter_read_positive(&EXT4_SB(sb)->s_freeinodes_counter);
 284         avefreei = freei / ngroups;
 285
 286         for (group = 0; group < ngroups; group++) {
 287                 desc = ext4_get_group_desc (sb, group, NULL);
 288                 if (!desc || !desc->bg_free_inodes_count)
 289                         continue;
 290                 if (le16_to_cpu(desc->bg_free_inodes_count) < avefreei)
 291                         continue;
 292                 if (!best_desc ||
 293                     (le16_to_cpu(desc->bg_free_blocks_count) >
 294                      le16_to_cpu(best_desc->bg_free_blocks_count))) {
 295                         *best_group = group;
 296                         best_desc = desc;
 297                         ret = 0;
 298                 }
 299         }
 300         return ret;
 301 }
 302
 303 #define free_block_ratio 10
 304
 305 static int find_group_flex(struct super_block *sb, struct inode *parent,
 306                            ext4_group_t *best_group)
 307 {
 308         struct ext4_sb_info *sbi = EXT4_SB(sb);
 309         struct ext4_group_desc *desc;
 310         struct buffer_head *bh;
 311         struct flex_groups *flex_group = sbi->s_flex_groups;
 312         ext4_group_t parent_group = EXT4_I(parent)->i_block_group;
 313         ext4_group_t parent_fbg_group = ext4_flex_group(sbi, parent_group);
 314         ext4_group_t ngroups = sbi->s_groups_count;
 315         int flex_size = ext4_flex_bg_size(sbi);
 316         ext4_group_t best_flex = parent_fbg_group;
 317         int blocks_per_flex = sbi->s_blocks_per_group * flex_size;
 318         int flexbg_free_blocks;
 319         int flex_freeb_ratio;
 320         ext4_group_t n_fbg_groups;
 321         ext4_group_t i;
 322
 323         n_fbg_groups = (sbi->s_groups_count + flex_size - 1) >>
 324                 sbi->s_log_groups_per_flex;
 325
 326 find_close_to_parent:
 327         flexbg_free_blocks = flex_group[best_flex].free_blocks;
 328         flex_freeb_ratio = flexbg_free_blocks * 100 / blocks_per_flex;
 329         if (flex_group[best_flex].free_inodes &&
 330             flex_freeb_ratio > free_block_ratio)
 331                 goto found_flexbg;
 332
 333         if (best_flex && best_flex == parent_fbg_group) {
 334                 best_flex--;
 335                 goto find_close_to_parent;
 336         }
 337
 338         for (i = 0; i < n_fbg_groups; i++) {
 339                 if (i == parent_fbg_group || i == parent_fbg_group - 1)
 340                         continue;
 341
 342                 flexbg_free_blocks = flex_group[i].free_blocks;
 343                 flex_freeb_ratio = flexbg_free_blocks * 100 / blocks_per_flex;
 344
 345                 if (flex_freeb_ratio > free_block_ratio &&
 346                     flex_group[i].free_inodes) {
 347                         best_flex = i;
 348                         goto found_flexbg;
 349                 }
 350
 351                 if (best_flex < 0 ||
 352                     (flex_group[i].free_blocks >
 353                      flex_group[best_flex].free_blocks &&
 354                      flex_group[i].free_inodes))
 355                         best_flex = i;
 356         }
 357
 358         if (!flex_group[best_flex].free_inodes ||
 359             !flex_group[best_flex].free_blocks)
 360                 return -1;
 361
 362 found_flexbg:
 363         for (i = best_flex * flex_size; i < ngroups &&
 364                      i < (best_flex + 1) * flex_size; i++) {
 365                 desc = ext4_get_group_desc(sb, i, &bh);
 366                 if (le16_to_cpu(desc->bg_free_inodes_count)) {
 367                         *best_group = i;
 368                         goto out;
 369                 }
 370         }
 371
 372         return -1;
 373 out:
 374         return 0;
 375 }
 376
 377 /*
 378  * Orlov's allocator for directories.
 379  *
 380  * We always try to spread first-level directories.
 381  *
 382  * If there are blockgroups with both free inodes and free blocks counts
 383  * not worse than average we return one with smallest directory count.
 384  * Otherwise we simply return a random group.
 385  *
 386  * For the rest rules look so:
 387  *
 388  * It's OK to put directory into a group unless
 389  * it has too many directories already (max_dirs) or
 390  * it has too few free inodes left (min_inodes) or
 391  * it has too few free blocks left (min_blocks) or
 392  * it's already running too large debt (max_debt).
 393  * Parent's group is preferred, if it doesn't satisfy these
 394  * conditions we search cyclically through the rest. If none
 395  * of the groups look good we just look for a group with more
 396  * free inodes than average (starting at parent's group).
 397  *
 398  * Debt is incremented each time we allocate a directory and decremented
 399  * when we allocate an inode, within 0--255.
 400  */
 401
 402 #define INODE_COST 64
 403 #define BLOCK_COST 256
 404
 405 static int find_group_orlov(struct super_block *sb, struct inode *parent,
 406                                 ext4_group_t *group)
 407 {
 408         ext4_group_t parent_group = EXT4_I(parent)->i_block_group;
 409         struct ext4_sb_info *sbi = EXT4_SB(sb);
 410         struct ext4_super_block *es = sbi->s_es;
 411         ext4_group_t ngroups = sbi->s_groups_count;
 412         int inodes_per_group = EXT4_INODES_PER_GROUP(sb);
 413         unsigned int freei, avefreei;
 414         ext4_fsblk_t freeb, avefreeb;
 415         ext4_fsblk_t blocks_per_dir;
 416         unsigned int ndirs;
 417         int max_debt, max_dirs, min_inodes;
 418         ext4_grpblk_t min_blocks;
 419         ext4_group_t i;
 420         struct ext4_group_desc *desc;
 421
 422         freei = percpu_counter_read_positive(&sbi->s_freeinodes_counter);
 423         avefreei = freei / ngroups;
 424         freeb = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
 425         avefreeb = freeb;
 426         do_div(avefreeb, ngroups);
 427         ndirs = percpu_counter_read_positive(&sbi->s_dirs_counter);
 428
 429         if ((parent == sb->s_root->d_inode) ||
 430             (EXT4_I(parent)->i_flags & EXT4_TOPDIR_FL)) {
 431                 int best_ndir = inodes_per_group;
 432                 ext4_group_t grp;
 433                 int ret = -1;
 434
 435                 get_random_bytes(&grp, sizeof(grp));
 436                 parent_group = (unsigned)grp % ngroups;
 437                 for (i = 0; i < ngroups; i++) {
 438                         grp = (parent_group + i) % ngroups;
 439                         desc = ext4_get_group_desc(sb, grp, NULL);
 440                         if (!desc || !desc->bg_free_inodes_count)
 441                                 continue;
 442                         if (le16_to_cpu(desc->bg_used_dirs_count) >= best_ndir)
 443                                 continue;
 444                         if (le16_to_cpu(desc->bg_free_inodes_count) < avefreei)
 445                                 continue;
 446                         if (le16_to_cpu(desc->bg_free_blocks_count) < avefreeb)
 447                                 continue;
 448                         *group = grp;
 449                         ret = 0;
 450                         best_ndir = le16_to_cpu(desc->bg_used_dirs_count);
 451                 }
 452                 if (ret == 0)
 453                         return ret;
 454                 goto fallback;
 455         }
 456
 457         blocks_per_dir = ext4_blocks_count(es) - freeb;
 458         do_div(blocks_per_dir, ndirs);
 459
 460         max_dirs = ndirs / ngroups + inodes_per_group / 16;
 461         min_inodes = avefreei - inodes_per_group / 4;
 462         min_blocks = avefreeb - EXT4_BLOCKS_PER_GROUP(sb) / 4;
 463
 464         max_debt = EXT4_BLOCKS_PER_GROUP(sb);
 465         max_debt /= max_t(int, blocks_per_dir, BLOCK_COST);
 466         if (max_debt * INODE_COST > inodes_per_group)
 467                 max_debt = inodes_per_group / INODE_COST;
 468         if (max_debt > 255)
 469                 max_debt = 255;
 470         if (max_debt == 0)
 471                 max_debt = 1;
 472
 473         for (i = 0; i < ngroups; i++) {
 474                 *group = (parent_group + i) % ngroups;
 475                 desc = ext4_get_group_desc(sb, *group, NULL);
 476                 if (!desc || !desc->bg_free_inodes_count)
 477                         continue;
 478                 if (le16_to_cpu(desc->bg_used_dirs_count) >= max_dirs)
 479                         continue;
 480                 if (le16_to_cpu(desc->bg_free_inodes_count) < min_inodes)
 481                         continue;
 482                 if (le16_to_cpu(desc->bg_free_blocks_count) < min_blocks)
 483                         continue;
 484                 return 0;
 485         }
 486
 487 fallback:
 488         for (i = 0; i < ngroups; i++) {
 489                 *group = (parent_group + i) % ngroups;
 490                 desc = ext4_get_group_desc(sb, *group, NULL);
 491                 if (desc && desc->bg_free_inodes_count &&
 492                         le16_to_cpu(desc->bg_free_inodes_count) >= avefreei)
 493                         return 0;
 494         }
 495
 496         if (avefreei) {
 497                 /*
 498                  * The free-inodes counter is approximate, and for really small
 499                  * filesystems the above test can fail to find any blockgroups
 500                  */
 501                 avefreei = 0;
 502                 goto fallback;
 503         }
 504
 505         return -1;
 506 }
 507
 508 static int find_group_other(struct super_block *sb, struct inode *parent,
 509                                 ext4_group_t *group)
 510 {
 511         ext4_group_t parent_group = EXT4_I(parent)->i_block_group;
 512         ext4_group_t ngroups = EXT4_SB(sb)->s_groups_count;
 513         struct ext4_group_desc *desc;
 514         ext4_group_t i;
 515
 516         /*
 517          * Try to place the inode in its parent directory
 518          */
 519         *group = parent_group;
 520         desc = ext4_get_group_desc(sb, *group, NULL);
 521         if (desc && le16_to_cpu(desc->bg_free_inodes_count) &&
 522                         le16_to_cpu(desc->bg_free_blocks_count))
 523                 return 0;
 524
 525         /*
 526          * We're going to place this inode in a different blockgroup from its
 527          * parent.  We want to cause files in a common directory to all land in
 528          * the same blockgroup.  But we want files which are in a different
 529          * directory which shares a blockgroup with our parent to land in a
 530          * different blockgroup.
 531          *
 532          * So add our directory's i_ino into the starting point for the hash.
 533          */
 534         *group = (*group + parent->i_ino) % ngroups;
 535
 536         /*
 537          * Use a quadratic hash to find a group with a free inode and some free
 538          * blocks.
 539          */
 540         for (i = 1; i < ngroups; i <<= 1) {
 541                 *group += i;
 542                 if (*group >= ngroups)
 543                         *group -= ngroups;
 544                 desc = ext4_get_group_desc(sb, *group, NULL);
 545                 if (desc && le16_to_cpu(desc->bg_free_inodes_count) &&
 546                                 le16_to_cpu(desc->bg_free_blocks_count))
 547                         return 0;
 548         }
 549
 550         /*
 551          * That failed: try linear search for a free inode, even if that group
 552          * has no free blocks.
 553          */
 554         *group = parent_group;
 555         for (i = 0; i < ngroups; i++) {
 556                 if (++*group >= ngroups)
 557                         *group = 0;
 558                 desc = ext4_get_group_desc(sb, *group, NULL);
 559                 if (desc && le16_to_cpu(desc->bg_free_inodes_count))
 560                         return 0;
 561         }
 562
 563         return -1;
 564 }
 565
 566 /*
 567  * There are two policies for allocating an inode.  If the new inode is
 568  * a directory, then a forward search is made for a block group with both
 569  * free space and a low directory-to-inode ratio; if that fails, then of
 570  * the groups with above-average free space, that group with the fewest
 571  * directories already is chosen.
 572  *
 573  * For other inodes, search forward from the parent directory's block
 574  * group to find a free inode.
 575  */
 576 struct inode *ext4_new_inode(handle_t *handle, struct inode * dir, int mode)
 577 {
 578         struct super_block *sb;
 579         struct buffer_head *bitmap_bh = NULL;
 580         struct buffer_head *bh2;
 581         ext4_group_t group = 0;
 582         unsigned long ino = 0;
 583         struct inode * inode;
 584         struct ext4_group_desc * gdp = NULL;
 585         struct ext4_super_block * es;
 586         struct ext4_inode_info *ei;
 587         struct ext4_sb_info *sbi;
 588         int ret2, err = 0;
 589         struct inode *ret;
 590         ext4_group_t i;
 591         int free = 0;
 592         ext4_group_t flex_group;
 593
 594         /* Cannot create files in a deleted directory */
 595         if (!dir || !dir->i_nlink)
 596                 return ERR_PTR(-EPERM);
 597
 598         sb = dir->i_sb;
 599         inode = new_inode(sb);
 600         if (!inode)
 601                 return ERR_PTR(-ENOMEM);
 602         ei = EXT4_I(inode);
 603
 604         sbi = EXT4_SB(sb);
 605         es = sbi->s_es;
 606
 607         if (sbi->s_log_groups_per_flex) {
 608                 ret2 = find_group_flex(sb, dir, &group);
 609                 goto got_group;
 610         }
 611
 612         if (S_ISDIR(mode)) {
 613                 if (test_opt (sb, OLDALLOC))
 614                         ret2 = find_group_dir(sb, dir, &group);
 615                 else
 616                         ret2 = find_group_orlov(sb, dir, &group);
 617         } else
 618                 ret2 = find_group_other(sb, dir, &group);
 619
 620 got_group:
 621         err = -ENOSPC;
 622         if (ret2 == -1)
 623                 goto out;
 624
 625         for (i = 0; i < sbi->s_groups_count; i++) {
 626                 err = -EIO;
 627
 628                 gdp = ext4_get_group_desc(sb, group, &bh2);
 629                 if (!gdp)
 630                         goto fail;
 631
 632                 brelse(bitmap_bh);
 633                 bitmap_bh = ext4_read_inode_bitmap(sb, group);
 634                 if (!bitmap_bh)
 635                         goto fail;
 636
 637                 ino = 0;
 638
 639 repeat_in_this_group:
 640                 ino = ext4_find_next_zero_bit((unsigned long *)
 641                                 bitmap_bh->b_data, EXT4_INODES_PER_GROUP(sb), ino);
 642                 if (ino < EXT4_INODES_PER_GROUP(sb)) {
 643
 644                         BUFFER_TRACE(bitmap_bh, "get_write_access");
 645                         err = ext4_journal_get_write_access(handle, bitmap_bh);
 646                         if (err)
 647                                 goto fail;
 648
 649                         if (!ext4_set_bit_atomic(sb_bgl_lock(sbi, group),
 650                                                 ino, bitmap_bh->b_data)) {
 651                                 /* we won it */
 652                                 BUFFER_TRACE(bitmap_bh,
 653                                         "call ext4_journal_dirty_metadata");
 654                                 err = ext4_journal_dirty_metadata(handle,
 655                                                                 bitmap_bh);
 656                                 if (err)
 657                                         goto fail;
 658                                 goto got;
 659                         }
 660                         /* we lost it */
 661                         jbd2_journal_release_buffer(handle, bitmap_bh);
 662
 663                         if (++ino < EXT4_INODES_PER_GROUP(sb))
 664                                 goto repeat_in_this_group;
 665                 }
 666
 667                 /*
 668                  * This case is possible in concurrent environment.  It is very
 669                  * rare.  We cannot repeat the find_group_xxx() call because
 670                  * that will simply return the same blockgroup, because the
 671                  * group descriptor metadata has not yet been updated.
 672                  * So we just go onto the next blockgroup.
 673                  */
 674                 if (++group == sbi->s_groups_count)
 675                         group = 0;
 676         }
 677         err = -ENOSPC;
 678         goto out;
 679
 680 got:
 681         ino++;
 682         if ((group == 0 && ino < EXT4_FIRST_INO(sb)) ||
 683             ino > EXT4_INODES_PER_GROUP(sb)) {
 684                 ext4_error(sb, __func__,
 685                            "reserved inode or inode > inodes count - "
 686                            "block_group = %lu, inode=%lu", group,
 687                            ino + group * EXT4_INODES_PER_GROUP(sb));
 688                 err = -EIO;
 689                 goto fail;
 690         }
 691
 692         BUFFER_TRACE(bh2, "get_write_access");
 693         err = ext4_journal_get_write_access(handle, bh2);
 694         if (err) goto fail;
 695
 696         /* We may have to initialize the block bitmap if it isn't already */
 697         if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM) &&
 698             gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
 699                 struct buffer_head *block_bh = ext4_read_block_bitmap(sb, group);
 700
 701                 BUFFER_TRACE(block_bh, "get block bitmap access");
 702                 err = ext4_journal_get_write_access(handle, block_bh);
 703                 if (err) {
 704                         brelse(block_bh);
 705                         goto fail;
 706                 }
 707
 708                 free = 0;
 709                 spin_lock(sb_bgl_lock(sbi, group));
 710                 /* recheck and clear flag under lock if we still need to */
 711                 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
 712                         gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
 713                         free = ext4_free_blocks_after_init(sb, group, gdp);
 714                         gdp->bg_free_blocks_count = cpu_to_le16(free);
 715                 }
 716                 spin_unlock(sb_bgl_lock(sbi, group));
 717
 718                 /* Don't need to dirty bitmap block if we didn't change it */
 719                 if (free) {
 720                         BUFFER_TRACE(block_bh, "dirty block bitmap");
 721                         err = ext4_journal_dirty_metadata(handle, block_bh);
 722                 }
 723
 724                 brelse(block_bh);
 725                 if (err)
 726                         goto fail;
 727         }
 728
 729         spin_lock(sb_bgl_lock(sbi, group));
 730         /* If we didn't allocate from within the initialized part of the inode
 731          * table then we need to initialize up to this inode. */
 732         if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM)) {
 733                 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) {
 734                         gdp->bg_flags &= cpu_to_le16(~EXT4_BG_INODE_UNINIT);
 735
 736                         /* When marking the block group with
 737                          * ~EXT4_BG_INODE_UNINIT we don't want to depend
 738                          * on the value of bg_itable_unsed even though
 739                          * mke2fs could have initialized the same for us.
 740                          * Instead we calculated the value below
 741                          */
 742
 743                         free = 0;
 744                 } else {
 745                         free = EXT4_INODES_PER_GROUP(sb) -
 746                                 le16_to_cpu(gdp->bg_itable_unused);
 747                 }
 748
 749                 /*
 750                  * Check the relative inode number against the last used
 751                  * relative inode number in this group. if it is greater
 752                  * we need to  update the bg_itable_unused count
 753                  *
 754                  */
 755                 if (ino > free)
 756                         gdp->bg_itable_unused =
 757                                 cpu_to_le16(EXT4_INODES_PER_GROUP(sb) - ino);
 758         }
 759
 760         le16_add_cpu(&gdp->bg_free_inodes_count, -1);
 761         if (S_ISDIR(mode)) {
 762                 le16_add_cpu(&gdp->bg_used_dirs_count, 1);
 763         }
 764         gdp->bg_checksum = ext4_group_desc_csum(sbi, group, gdp);
 765         spin_unlock(sb_bgl_lock(sbi, group));
 766         BUFFER_TRACE(bh2, "call ext4_journal_dirty_metadata");
 767         err = ext4_journal_dirty_metadata(handle, bh2);
 768         if (err) goto fail;
 769
 770         percpu_counter_dec(&sbi->s_freeinodes_counter);
 771         if (S_ISDIR(mode))
 772                 percpu_counter_inc(&sbi->s_dirs_counter);
 773         sb->s_dirt = 1;
 774
 775         if (sbi->s_log_groups_per_flex) {
 776                 flex_group = ext4_flex_group(sbi, group);
 777                 spin_lock(sb_bgl_lock(sbi, flex_group));
 778                 sbi->s_flex_groups[flex_group].free_inodes--;
 779                 spin_unlock(sb_bgl_lock(sbi, flex_group));
 780         }
 781
 782         inode->i_uid = current->fsuid;
 783         if (test_opt (sb, GRPID))
 784                 inode->i_gid = dir->i_gid;
 785         else if (dir->i_mode & S_ISGID) {
 786                 inode->i_gid = dir->i_gid;
 787                 if (S_ISDIR(mode))
 788                         mode |= S_ISGID;
 789         } else
 790                 inode->i_gid = current->fsgid;
 791         inode->i_mode = mode;
 792
 793         inode->i_ino = ino + group * EXT4_INODES_PER_GROUP(sb);
 794         /* This is the optimal IO size (for stat), not the fs block size */
 795         inode->i_blocks = 0;
 796         inode->i_mtime = inode->i_atime = inode->i_ctime = ei->i_crtime =
 797                                                        ext4_current_time(inode);
 798
 799         memset(ei->i_data, 0, sizeof(ei->i_data));
 800         ei->i_dir_start_lookup = 0;
 801         ei->i_disksize = 0;
 802
 803         /*
 804          * Don't inherit extent flag from directory. We set extent flag on
 805          * newly created directory and file only if -o extent mount option is
 806          * specified
 807          */
 808         ei->i_flags = EXT4_I(dir)->i_flags & ~(EXT4_INDEX_FL|EXT4_EXTENTS_FL);
 809         if (S_ISLNK(mode))
 810                 ei->i_flags &= ~(EXT4_IMMUTABLE_FL|EXT4_APPEND_FL);
 811         /* dirsync only applies to directories */
 812         if (!S_ISDIR(mode))
 813                 ei->i_flags &= ~EXT4_DIRSYNC_FL;
 814         ei->i_file_acl = 0;
 815         ei->i_dtime = 0;
 816         ei->i_block_alloc_info = NULL;
 817         ei->i_block_group = group;
 818
 819         ext4_set_inode_flags(inode);
 820         if (IS_DIRSYNC(inode))
 821                 handle->h_sync = 1;
 822         insert_inode_hash(inode);
 823         spin_lock(&sbi->s_next_gen_lock);
 824         inode->i_generation = sbi->s_next_generation++;
 825         spin_unlock(&sbi->s_next_gen_lock);
 826
 827         ei->i_state = EXT4_STATE_NEW;
 828
 829         ei->i_extra_isize = EXT4_SB(sb)->s_want_extra_isize;
 830
 831         ret = inode;
 832         if(DQUOT_ALLOC_INODE(inode)) {
 833                 err = -EDQUOT;
 834                 goto fail_drop;
 835         }
 836
 837         err = ext4_init_acl(handle, inode, dir);
 838         if (err)
 839                 goto fail_free_drop;
 840
 841         err = ext4_init_security(handle,inode, dir);
 842         if (err)
 843                 goto fail_free_drop;
 844
 845         if (test_opt(sb, EXTENTS)) {
 846                 /* set extent flag only for directory, file and normal symlink*/
 847                 if (S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode)) {
 848                         EXT4_I(inode)->i_flags |= EXT4_EXTENTS_FL;
 849                         ext4_ext_tree_init(handle, inode);
 850                 }
 851         }
 852
 853         err = ext4_mark_inode_dirty(handle, inode);
 854         if (err) {
 855                 ext4_std_error(sb, err);
 856                 goto fail_free_drop;
 857         }
 858
 859         ext4_debug("allocating inode %lu\n", inode->i_ino);
 860         goto really_out;
 861 fail:
 862         ext4_std_error(sb, err);
 863 out:
 864         iput(inode);
 865         ret = ERR_PTR(err);
 866 really_out:
 867         brelse(bitmap_bh);
 868         return ret;
 869
 870 fail_free_drop:
 871         DQUOT_FREE_INODE(inode);
 872
 873 fail_drop:
 874         DQUOT_DROP(inode);
 875         inode->i_flags |= S_NOQUOTA;
 876         inode->i_nlink = 0;
 877         iput(inode);
 878         brelse(bitmap_bh);
 879         return ERR_PTR(err);
 880 }
 881
 882 /* Verify that we are loading a valid orphan from disk */
 883 struct inode *ext4_orphan_get(struct super_block *sb, unsigned long ino)
 884 {
 885         unsigned long max_ino = le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count);
 886         ext4_group_t block_group;
 887         int bit;
 888         struct buffer_head *bitmap_bh;
 889         struct inode *inode = NULL;
 890         long err = -EIO;
 891
 892         /* Error cases - e2fsck has already cleaned up for us */
 893         if (ino > max_ino) {
 894                 ext4_warning(sb, __func__,
 895                              "bad orphan ino %lu!  e2fsck was run?", ino);
 896                 goto error;
 897         }
 898
 899         block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb);
 900         bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb);
 901         bitmap_bh = ext4_read_inode_bitmap(sb, block_group);
 902         if (!bitmap_bh) {
 903                 ext4_warning(sb, __func__,
 904                              "inode bitmap error for orphan %lu", ino);
 905                 goto error;
 906         }
 907
 908         /* Having the inode bit set should be a 100% indicator that this
 909          * is a valid orphan (no e2fsck run on fs).  Orphans also include
 910          * inodes that were being truncated, so we can't check i_nlink==0.
 911          */
 912         if (!ext4_test_bit(bit, bitmap_bh->b_data))
 913                 goto bad_orphan;
 914
 915         inode = ext4_iget(sb, ino);
 916         if (IS_ERR(inode))
 917                 goto iget_failed;
 918
 919         /*
 920          * If the orphans has i_nlinks > 0 then it should be able to be
 921          * truncated, otherwise it won't be removed from the orphan list
 922          * during processing and an infinite loop will result.
 923          */
 924         if (inode->i_nlink && !ext4_can_truncate(inode))
 925                 goto bad_orphan;
 926
 927         if (NEXT_ORPHAN(inode) > max_ino)
 928                 goto bad_orphan;
 929         brelse(bitmap_bh);
 930         return inode;
 931
 932 iget_failed:
 933         err = PTR_ERR(inode);
 934         inode = NULL;
 935 bad_orphan:
 936         ext4_warning(sb, __func__,
 937                      "bad orphan inode %lu!  e2fsck was run?", ino);
 938         printk(KERN_NOTICE "ext4_test_bit(bit=%d, block=%llu) = %d\n",
 939                bit, (unsigned long long)bitmap_bh->b_blocknr,
 940                ext4_test_bit(bit, bitmap_bh->b_data));
 941         printk(KERN_NOTICE "inode=%p\n", inode);
 942         if (inode) {
 943                 printk(KERN_NOTICE "is_bad_inode(inode)=%d\n",
 944                        is_bad_inode(inode));
 945                 printk(KERN_NOTICE "NEXT_ORPHAN(inode)=%u\n",
 946                        NEXT_ORPHAN(inode));
 947                 printk(KERN_NOTICE "max_ino=%lu\n", max_ino);
 948                 printk(KERN_NOTICE "i_nlink=%u\n", inode->i_nlink);
 949                 /* Avoid freeing blocks if we got a bad deleted inode */
 950                 if (inode->i_nlink == 0)
 951                         inode->i_blocks = 0;
 952                 iput(inode);
 953         }
 954         brelse(bitmap_bh);
 955 error:
 956         return ERR_PTR(err);
 957 }
 958
 959 unsigned long ext4_count_free_inodes (struct super_block * sb)
 960 {
 961         unsigned long desc_count;
 962         struct ext4_group_desc *gdp;
 963         ext4_group_t i;
 964 #ifdef EXT4FS_DEBUG
 965         struct ext4_super_block *es;
 966         unsigned long bitmap_count, x;
 967         struct buffer_head *bitmap_bh = NULL;
 968
 969         es = EXT4_SB(sb)->s_es;
 970         desc_count = 0;
 971         bitmap_count = 0;
 972         gdp = NULL;
 973         for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
 974                 gdp = ext4_get_group_desc (sb, i, NULL);
 975                 if (!gdp)
 976                         continue;
 977                 desc_count += le16_to_cpu(gdp->bg_free_inodes_count);
 978                 brelse(bitmap_bh);
 979                 bitmap_bh = ext4_read_inode_bitmap(sb, i);
 980                 if (!bitmap_bh)
 981                         continue;
 982
 983                 x = ext4_count_free(bitmap_bh, EXT4_INODES_PER_GROUP(sb) / 8);
 984                 printk(KERN_DEBUG "group %lu: stored = %d, counted = %lu\n",
 985                         i, le16_to_cpu(gdp->bg_free_inodes_count), x);
 986                 bitmap_count += x;
 987         }
 988         brelse(bitmap_bh);
 989         printk("ext4_count_free_inodes: stored = %u, computed = %lu, %lu\n",
 990                 le32_to_cpu(es->s_free_inodes_count), desc_count, bitmap_count);
 991         return desc_count;
 992 #else
 993         desc_count = 0;
 994         for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
 995                 gdp = ext4_get_group_desc (sb, i, NULL);
 996                 if (!gdp)
 997                         continue;
 998                 desc_count += le16_to_cpu(gdp->bg_free_inodes_count);
 999                 cond_resched();
1000         }
1001         return desc_count;
1002 #endif
1003 }
1004
1005 /* Called at mount-time, super-block is locked */
1006 unsigned long ext4_count_dirs (struct super_block * sb)
1007 {
1008         unsigned long count = 0;
1009         ext4_group_t i;
1010
1011         for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
1012                 struct ext4_group_desc *gdp = ext4_get_group_desc (sb, i, NULL);
1013                 if (!gdp)
1014                         continue;
1015                 count += le16_to_cpu(gdp->bg_used_dirs_count);
1016         }
1017         return count;
1018 }
1019