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 %u",
  78                            block_group);
  79                 ext4_free_blks_set(sb, gdp, 0);
  80                 ext4_free_inodes_set(sb, gdp, 0);
  81                 ext4_itable_unused_set(sb, gdp, 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 = %u, inode_bitmap = %llu",
 115                             block_group, bitmap_blk);
 116                 return NULL;
 117         }
 118         if (buffer_uptodate(bh) &&
 119             !(desc->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)))
 120                 return bh;
 121
 122         lock_buffer(bh);
 123         spin_lock(sb_bgl_lock(EXT4_SB(sb), block_group));
 124         if (desc->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) {
 125                 ext4_init_inode_bitmap(sb, bh, block_group, desc);
 126                 set_buffer_uptodate(bh);
 127                 unlock_buffer(bh);
 128                 spin_unlock(sb_bgl_lock(EXT4_SB(sb), block_group));
 129                 return bh;
 130         }
 131         spin_unlock(sb_bgl_lock(EXT4_SB(sb), block_group));
 132         if (bh_submit_read(bh) < 0) {
 133                 put_bh(bh);
 134                 ext4_error(sb, __func__,
 135                             "Cannot read inode bitmap - "
 136                             "block_group = %u, inode_bitmap = %llu",
 137                             block_group, bitmap_blk);
 138                 return NULL;
 139         }
 140         return bh;
 141 }
 142
 143 /*
 144  * NOTE! When we get the inode, we're the only people
 145  * that have access to it, and as such there are no
 146  * race conditions we have to worry about. The inode
 147  * is not on the hash-lists, and it cannot be reached
 148  * through the filesystem because the directory entry
 149  * has been deleted earlier.
 150  *
 151  * HOWEVER: we must make sure that we get no aliases,
 152  * which means that we have to call "clear_inode()"
 153  * _before_ we mark the inode not in use in the inode
 154  * bitmaps. Otherwise a newly created file might use
 155  * the same inode number (not actually the same pointer
 156  * though), and then we'd have two inodes sharing the
 157  * same inode number and space on the harddisk.
 158  */
 159 void ext4_free_inode(handle_t *handle, struct inode *inode)
 160 {
 161         struct super_block *sb = inode->i_sb;
 162         int is_directory;
 163         unsigned long ino;
 164         struct buffer_head *bitmap_bh = NULL;
 165         struct buffer_head *bh2;
 166         ext4_group_t block_group;
 167         unsigned long bit;
 168         struct ext4_group_desc *gdp;
 169         struct ext4_super_block *es;
 170         struct ext4_sb_info *sbi;
 171         int fatal = 0, err, count;
 172         ext4_group_t flex_group;
 173
 174         if (atomic_read(&inode->i_count) > 1) {
 175                 printk(KERN_ERR "ext4_free_inode: inode has count=%d\n",
 176                        atomic_read(&inode->i_count));
 177                 return;
 178         }
 179         if (inode->i_nlink) {
 180                 printk(KERN_ERR "ext4_free_inode: inode has nlink=%d\n",
 181                        inode->i_nlink);
 182                 return;
 183         }
 184         if (!sb) {
 185                 printk(KERN_ERR "ext4_free_inode: inode on "
 186                        "nonexistent device\n");
 187                 return;
 188         }
 189         sbi = EXT4_SB(sb);
 190
 191         ino = inode->i_ino;
 192         ext4_debug("freeing inode %lu\n", ino);
 193
 194         /*
 195          * Note: we must free any quota before locking the superblock,
 196          * as writing the quota to disk may need the lock as well.
 197          */
 198         DQUOT_INIT(inode);
 199         ext4_xattr_delete_inode(handle, inode);
 200         DQUOT_FREE_INODE(inode);
 201         DQUOT_DROP(inode);
 202
 203         is_directory = S_ISDIR(inode->i_mode);
 204
 205         /* Do this BEFORE marking the inode not in use or returning an error */
 206         clear_inode(inode);
 207
 208         es = EXT4_SB(sb)->s_es;
 209         if (ino < EXT4_FIRST_INO(sb) || ino > le32_to_cpu(es->s_inodes_count)) {
 210                 ext4_error(sb, "ext4_free_inode",
 211                            "reserved or nonexistent inode %lu", ino);
 212                 goto error_return;
 213         }
 214         block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb);
 215         bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb);
 216         bitmap_bh = ext4_read_inode_bitmap(sb, block_group);
 217         if (!bitmap_bh)
 218                 goto error_return;
 219
 220         BUFFER_TRACE(bitmap_bh, "get_write_access");
 221         fatal = ext4_journal_get_write_access(handle, bitmap_bh);
 222         if (fatal)
 223                 goto error_return;
 224
 225         /* Ok, now we can actually update the inode bitmaps.. */
 226         if (!ext4_clear_bit_atomic(sb_bgl_lock(sbi, block_group),
 227                                         bit, bitmap_bh->b_data))
 228                 ext4_error(sb, "ext4_free_inode",
 229                            "bit already cleared for inode %lu", ino);
 230         else {
 231                 gdp = ext4_get_group_desc(sb, block_group, &bh2);
 232
 233                 BUFFER_TRACE(bh2, "get_write_access");
 234                 fatal = ext4_journal_get_write_access(handle, bh2);
 235                 if (fatal) goto error_return;
 236
 237                 if (gdp) {
 238                         spin_lock(sb_bgl_lock(sbi, block_group));
 239                         count = ext4_free_inodes_count(sb, gdp) + 1;
 240                         ext4_free_inodes_set(sb, gdp, count);
 241                         if (is_directory) {
 242                                 count = ext4_used_dirs_count(sb, gdp) - 1;
 243                                 ext4_used_dirs_set(sb, gdp, count);
 244                         }
 245                         gdp->bg_checksum = ext4_group_desc_csum(sbi,
 246                                                         block_group, gdp);
 247                         spin_unlock(sb_bgl_lock(sbi, block_group));
 248                         percpu_counter_inc(&sbi->s_freeinodes_counter);
 249                         if (is_directory)
 250                                 percpu_counter_dec(&sbi->s_dirs_counter);
 251
 252                         if (sbi->s_log_groups_per_flex) {
 253                                 flex_group = ext4_flex_group(sbi, block_group);
 254                                 spin_lock(sb_bgl_lock(sbi, flex_group));
 255                                 sbi->s_flex_groups[flex_group].free_inodes++;
 256                                 spin_unlock(sb_bgl_lock(sbi, flex_group));
 257                         }
 258                 }
 259                 BUFFER_TRACE(bh2, "call ext4_handle_dirty_metadata");
 260                 err = ext4_handle_dirty_metadata(handle, NULL, bh2);
 261                 if (!fatal) fatal = err;
 262         }
 263         BUFFER_TRACE(bitmap_bh, "call ext4_handle_dirty_metadata");
 264         err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
 265         if (!fatal)
 266                 fatal = err;
 267         sb->s_dirt = 1;
 268 error_return:
 269         brelse(bitmap_bh);
 270         ext4_std_error(sb, fatal);
 271 }
 272
 273 /*
 274  * There are two policies for allocating an inode.  If the new inode is
 275  * a directory, then a forward search is made for a block group with both
 276  * free space and a low directory-to-inode ratio; if that fails, then of
 277  * the groups with above-average free space, that group with the fewest
 278  * directories already is chosen.
 279  *
 280  * For other inodes, search forward from the parent directory\'s block
 281  * group to find a free inode.
 282  */
 283 static int find_group_dir(struct super_block *sb, struct inode *parent,
 284                                 ext4_group_t *best_group)
 285 {
 286         ext4_group_t ngroups = EXT4_SB(sb)->s_groups_count;
 287         unsigned int freei, avefreei;
 288         struct ext4_group_desc *desc, *best_desc = NULL;
 289         ext4_group_t group;
 290         int ret = -1;
 291
 292         freei = percpu_counter_read_positive(&EXT4_SB(sb)->s_freeinodes_counter);
 293         avefreei = freei / ngroups;
 294
 295         for (group = 0; group < ngroups; group++) {
 296                 desc = ext4_get_group_desc(sb, group, NULL);
 297                 if (!desc || !ext4_free_inodes_count(sb, desc))
 298                         continue;
 299                 if (ext4_free_inodes_count(sb, desc) < avefreei)
 300                         continue;
 301                 if (!best_desc ||
 302                     (ext4_free_blks_count(sb, desc) >
 303                      ext4_free_blks_count(sb, best_desc))) {
 304                         *best_group = group;
 305                         best_desc = desc;
 306                         ret = 0;
 307                 }
 308         }
 309         return ret;
 310 }
 311
 312 #define free_block_ratio 10
 313
 314 static int find_group_flex(struct super_block *sb, struct inode *parent,
 315                            ext4_group_t *best_group)
 316 {
 317         struct ext4_sb_info *sbi = EXT4_SB(sb);
 318         struct ext4_group_desc *desc;
 319         struct buffer_head *bh;
 320         struct flex_groups *flex_group = sbi->s_flex_groups;
 321         ext4_group_t parent_group = EXT4_I(parent)->i_block_group;
 322         ext4_group_t parent_fbg_group = ext4_flex_group(sbi, parent_group);
 323         ext4_group_t ngroups = sbi->s_groups_count;
 324         int flex_size = ext4_flex_bg_size(sbi);
 325         ext4_group_t best_flex = parent_fbg_group;
 326         int blocks_per_flex = sbi->s_blocks_per_group * flex_size;
 327         int flexbg_free_blocks;
 328         int flex_freeb_ratio;
 329         ext4_group_t n_fbg_groups;
 330         ext4_group_t i;
 331
 332         n_fbg_groups = (sbi->s_groups_count + flex_size - 1) >>
 333                 sbi->s_log_groups_per_flex;
 334
 335 find_close_to_parent:
 336         flexbg_free_blocks = flex_group[best_flex].free_blocks;
 337         flex_freeb_ratio = flexbg_free_blocks * 100 / blocks_per_flex;
 338         if (flex_group[best_flex].free_inodes &&
 339             flex_freeb_ratio > free_block_ratio)
 340                 goto found_flexbg;
 341
 342         if (best_flex && best_flex == parent_fbg_group) {
 343                 best_flex--;
 344                 goto find_close_to_parent;
 345         }
 346
 347         for (i = 0; i < n_fbg_groups; i++) {
 348                 if (i == parent_fbg_group || i == parent_fbg_group - 1)
 349                         continue;
 350
 351                 flexbg_free_blocks = flex_group[i].free_blocks;
 352                 flex_freeb_ratio = flexbg_free_blocks * 100 / blocks_per_flex;
 353
 354                 if (flex_freeb_ratio > free_block_ratio &&
 355                     flex_group[i].free_inodes) {
 356                         best_flex = i;
 357                         goto found_flexbg;
 358                 }
 359
 360                 if (flex_group[best_flex].free_inodes == 0 ||
 361                     (flex_group[i].free_blocks >
 362                      flex_group[best_flex].free_blocks &&
 363                      flex_group[i].free_inodes))
 364                         best_flex = i;
 365         }
 366
 367         if (!flex_group[best_flex].free_inodes ||
 368             !flex_group[best_flex].free_blocks)
 369                 return -1;
 370
 371 found_flexbg:
 372         for (i = best_flex * flex_size; i < ngroups &&
 373                      i < (best_flex + 1) * flex_size; i++) {
 374                 desc = ext4_get_group_desc(sb, i, &bh);
 375                 if (ext4_free_inodes_count(sb, desc)) {
 376                         *best_group = i;
 377                         goto out;
 378                 }
 379         }
 380
 381         return -1;
 382 out:
 383         return 0;
 384 }
 385
 386 /*
 387  * Orlov's allocator for directories.
 388  *
 389  * We always try to spread first-level directories.
 390  *
 391  * If there are blockgroups with both free inodes and free blocks counts
 392  * not worse than average we return one with smallest directory count.
 393  * Otherwise we simply return a random group.
 394  *
 395  * For the rest rules look so:
 396  *
 397  * It's OK to put directory into a group unless
 398  * it has too many directories already (max_dirs) or
 399  * it has too few free inodes left (min_inodes) or
 400  * it has too few free blocks left (min_blocks) or
 401  * it's already running too large debt (max_debt).
 402  * Parent's group is preferred, if it doesn't satisfy these
 403  * conditions we search cyclically through the rest. If none
 404  * of the groups look good we just look for a group with more
 405  * free inodes than average (starting at parent's group).
 406  *
 407  * Debt is incremented each time we allocate a directory and decremented
 408  * when we allocate an inode, within 0--255.
 409  */
 410
 411 #define INODE_COST 64
 412 #define BLOCK_COST 256
 413
 414 static int find_group_orlov(struct super_block *sb, struct inode *parent,
 415                                 ext4_group_t *group)
 416 {
 417         ext4_group_t parent_group = EXT4_I(parent)->i_block_group;
 418         struct ext4_sb_info *sbi = EXT4_SB(sb);
 419         struct ext4_super_block *es = sbi->s_es;
 420         ext4_group_t ngroups = sbi->s_groups_count;
 421         int inodes_per_group = EXT4_INODES_PER_GROUP(sb);
 422         unsigned int freei, avefreei;
 423         ext4_fsblk_t freeb, avefreeb;
 424         ext4_fsblk_t blocks_per_dir;
 425         unsigned int ndirs;
 426         int max_debt, max_dirs, min_inodes;
 427         ext4_grpblk_t min_blocks;
 428         ext4_group_t i;
 429         struct ext4_group_desc *desc;
 430
 431         freei = percpu_counter_read_positive(&sbi->s_freeinodes_counter);
 432         avefreei = freei / ngroups;
 433         freeb = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
 434         avefreeb = freeb;
 435         do_div(avefreeb, ngroups);
 436         ndirs = percpu_counter_read_positive(&sbi->s_dirs_counter);
 437
 438         if ((parent == sb->s_root->d_inode) ||
 439             (EXT4_I(parent)->i_flags & EXT4_TOPDIR_FL)) {
 440                 int best_ndir = inodes_per_group;
 441                 ext4_group_t grp;
 442                 int ret = -1;
 443
 444                 get_random_bytes(&grp, sizeof(grp));
 445                 parent_group = (unsigned)grp % ngroups;
 446                 for (i = 0; i < ngroups; i++) {
 447                         grp = (parent_group + i) % ngroups;
 448                         desc = ext4_get_group_desc(sb, grp, NULL);
 449                         if (!desc || !ext4_free_inodes_count(sb, desc))
 450                                 continue;
 451                         if (ext4_used_dirs_count(sb, desc) >= best_ndir)
 452                                 continue;
 453                         if (ext4_free_inodes_count(sb, desc) < avefreei)
 454                                 continue;
 455                         if (ext4_free_blks_count(sb, desc) < avefreeb)
 456                                 continue;
 457                         *group = grp;
 458                         ret = 0;
 459                         best_ndir = ext4_used_dirs_count(sb, desc);
 460                 }
 461                 if (ret == 0)
 462                         return ret;
 463                 goto fallback;
 464         }
 465
 466         blocks_per_dir = ext4_blocks_count(es) - freeb;
 467         do_div(blocks_per_dir, ndirs);
 468
 469         max_dirs = ndirs / ngroups + inodes_per_group / 16;
 470         min_inodes = avefreei - inodes_per_group / 4;
 471         min_blocks = avefreeb - EXT4_BLOCKS_PER_GROUP(sb) / 4;
 472
 473         max_debt = EXT4_BLOCKS_PER_GROUP(sb);
 474         max_debt /= max_t(int, blocks_per_dir, BLOCK_COST);
 475         if (max_debt * INODE_COST > inodes_per_group)
 476                 max_debt = inodes_per_group / INODE_COST;
 477         if (max_debt > 255)
 478                 max_debt = 255;
 479         if (max_debt == 0)
 480                 max_debt = 1;
 481
 482         for (i = 0; i < ngroups; i++) {
 483                 *group = (parent_group + i) % ngroups;
 484                 desc = ext4_get_group_desc(sb, *group, NULL);
 485                 if (!desc || !ext4_free_inodes_count(sb, desc))
 486                         continue;
 487                 if (ext4_used_dirs_count(sb, desc) >= max_dirs)
 488                         continue;
 489                 if (ext4_free_inodes_count(sb, desc) < min_inodes)
 490                         continue;
 491                 if (ext4_free_blks_count(sb, desc) < min_blocks)
 492                         continue;
 493                 return 0;
 494         }
 495
 496 fallback:
 497         for (i = 0; i < ngroups; i++) {
 498                 *group = (parent_group + i) % ngroups;
 499                 desc = ext4_get_group_desc(sb, *group, NULL);
 500                 if (desc && ext4_free_inodes_count(sb, desc) &&
 501                         ext4_free_inodes_count(sb, desc) >= avefreei)
 502                         return 0;
 503         }
 504
 505         if (avefreei) {
 506                 /*
 507                  * The free-inodes counter is approximate, and for really small
 508                  * filesystems the above test can fail to find any blockgroups
 509                  */
 510                 avefreei = 0;
 511                 goto fallback;
 512         }
 513
 514         return -1;
 515 }
 516
 517 static int find_group_other(struct super_block *sb, struct inode *parent,
 518                                 ext4_group_t *group)
 519 {
 520         ext4_group_t parent_group = EXT4_I(parent)->i_block_group;
 521         ext4_group_t ngroups = EXT4_SB(sb)->s_groups_count;
 522         struct ext4_group_desc *desc;
 523         ext4_group_t i;
 524
 525         /*
 526          * Try to place the inode in its parent directory
 527          */
 528         *group = parent_group;
 529         desc = ext4_get_group_desc(sb, *group, NULL);
 530         if (desc && ext4_free_inodes_count(sb, desc) &&
 531                         ext4_free_blks_count(sb, desc))
 532                 return 0;
 533
 534         /*
 535          * We're going to place this inode in a different blockgroup from its
 536          * parent.  We want to cause files in a common directory to all land in
 537          * the same blockgroup.  But we want files which are in a different
 538          * directory which shares a blockgroup with our parent to land in a
 539          * different blockgroup.
 540          *
 541          * So add our directory's i_ino into the starting point for the hash.
 542          */
 543         *group = (*group + parent->i_ino) % ngroups;
 544
 545         /*
 546          * Use a quadratic hash to find a group with a free inode and some free
 547          * blocks.
 548          */
 549         for (i = 1; i < ngroups; i <<= 1) {
 550                 *group += i;
 551                 if (*group >= ngroups)
 552                         *group -= ngroups;
 553                 desc = ext4_get_group_desc(sb, *group, NULL);
 554                 if (desc && ext4_free_inodes_count(sb, desc) &&
 555                                 ext4_free_blks_count(sb, desc))
 556                         return 0;
 557         }
 558
 559         /*
 560          * That failed: try linear search for a free inode, even if that group
 561          * has no free blocks.
 562          */
 563         *group = parent_group;
 564         for (i = 0; i < ngroups; i++) {
 565                 if (++*group >= ngroups)
 566                         *group = 0;
 567                 desc = ext4_get_group_desc(sb, *group, NULL);
 568                 if (desc && ext4_free_inodes_count(sb, desc))
 569                         return 0;
 570         }
 571
 572         return -1;
 573 }
 574
 575 /*
 576  * There are two policies for allocating an inode.  If the new inode is
 577  * a directory, then a forward search is made for a block group with both
 578  * free space and a low directory-to-inode ratio; if that fails, then of
 579  * the groups with above-average free space, that group with the fewest
 580  * directories already is chosen.
 581  *
 582  * For other inodes, search forward from the parent directory's block
 583  * group to find a free inode.
 584  */
 585 struct inode *ext4_new_inode(handle_t *handle, struct inode *dir, int mode)
 586 {
 587         struct super_block *sb;
 588         struct buffer_head *bitmap_bh = NULL;
 589         struct buffer_head *bh2;
 590         ext4_group_t group = 0;
 591         unsigned long ino = 0;
 592         struct inode *inode;
 593         struct ext4_group_desc *gdp = NULL;
 594         struct ext4_super_block *es;
 595         struct ext4_inode_info *ei;
 596         struct ext4_sb_info *sbi;
 597         int ret2, err = 0, count;
 598         struct inode *ret;
 599         ext4_group_t i;
 600         int free = 0;
 601         ext4_group_t flex_group;
 602
 603         /* Cannot create files in a deleted directory */
 604         if (!dir || !dir->i_nlink)
 605                 return ERR_PTR(-EPERM);
 606
 607         sb = dir->i_sb;
 608         inode = new_inode(sb);
 609         if (!inode)
 610                 return ERR_PTR(-ENOMEM);
 611         ei = EXT4_I(inode);
 612
 613         sbi = EXT4_SB(sb);
 614         es = sbi->s_es;
 615
 616         if (sbi->s_log_groups_per_flex) {
 617                 ret2 = find_group_flex(sb, dir, &group);
 618                 goto got_group;
 619         }
 620
 621         if (S_ISDIR(mode)) {
 622                 if (test_opt(sb, OLDALLOC))
 623                         ret2 = find_group_dir(sb, dir, &group);
 624                 else
 625                         ret2 = find_group_orlov(sb, dir, &group);
 626         } else
 627                 ret2 = find_group_other(sb, dir, &group);
 628
 629 got_group:
 630         err = -ENOSPC;
 631         if (ret2 == -1)
 632                 goto out;
 633
 634         for (i = 0; i < sbi->s_groups_count; i++) {
 635                 err = -EIO;
 636
 637                 gdp = ext4_get_group_desc(sb, group, &bh2);
 638                 if (!gdp)
 639                         goto fail;
 640
 641                 brelse(bitmap_bh);
 642                 bitmap_bh = ext4_read_inode_bitmap(sb, group);
 643                 if (!bitmap_bh)
 644                         goto fail;
 645
 646                 ino = 0;
 647
 648 repeat_in_this_group:
 649                 ino = ext4_find_next_zero_bit((unsigned long *)
 650                                 bitmap_bh->b_data, EXT4_INODES_PER_GROUP(sb), ino);
 651                 if (ino < EXT4_INODES_PER_GROUP(sb)) {
 652
 653                         BUFFER_TRACE(bitmap_bh, "get_write_access");
 654                         err = ext4_journal_get_write_access(handle, bitmap_bh);
 655                         if (err)
 656                                 goto fail;
 657
 658                         if (!ext4_set_bit_atomic(sb_bgl_lock(sbi, group),
 659                                                 ino, bitmap_bh->b_data)) {
 660                                 /* we won it */
 661                                 BUFFER_TRACE(bitmap_bh,
 662                                         "call ext4_handle_dirty_metadata");
 663                                 err = ext4_handle_dirty_metadata(handle,
 664                                                                 inode,
 665                                                                 bitmap_bh);
 666                                 if (err)
 667                                         goto fail;
 668                                 goto got;
 669                         }
 670                         /* we lost it */
 671                         ext4_handle_release_buffer(handle, bitmap_bh);
 672
 673                         if (++ino < EXT4_INODES_PER_GROUP(sb))
 674                                 goto repeat_in_this_group;
 675                 }
 676
 677                 /*
 678                  * This case is possible in concurrent environment.  It is very
 679                  * rare.  We cannot repeat the find_group_xxx() call because
 680                  * that will simply return the same blockgroup, because the
 681                  * group descriptor metadata has not yet been updated.
 682                  * So we just go onto the next blockgroup.
 683                  */
 684                 if (++group == sbi->s_groups_count)
 685                         group = 0;
 686         }
 687         err = -ENOSPC;
 688         goto out;
 689
 690 got:
 691         ino++;
 692         if ((group == 0 && ino < EXT4_FIRST_INO(sb)) ||
 693             ino > EXT4_INODES_PER_GROUP(sb)) {
 694                 ext4_error(sb, __func__,
 695                            "reserved inode or inode > inodes count - "
 696                            "block_group = %u, inode=%lu", group,
 697                            ino + group * EXT4_INODES_PER_GROUP(sb));
 698                 err = -EIO;
 699                 goto fail;
 700         }
 701
 702         BUFFER_TRACE(bh2, "get_write_access");
 703         err = ext4_journal_get_write_access(handle, bh2);
 704         if (err) goto fail;
 705
 706         /* We may have to initialize the block bitmap if it isn't already */
 707         if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM) &&
 708             gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
 709                 struct buffer_head *block_bh = ext4_read_block_bitmap(sb, group);
 710
 711                 BUFFER_TRACE(block_bh, "get block bitmap access");
 712                 err = ext4_journal_get_write_access(handle, block_bh);
 713                 if (err) {
 714                         brelse(block_bh);
 715                         goto fail;
 716                 }
 717
 718                 free = 0;
 719                 spin_lock(sb_bgl_lock(sbi, group));
 720                 /* recheck and clear flag under lock if we still need to */
 721                 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
 722                         gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
 723                         free = ext4_free_blocks_after_init(sb, group, gdp);
 724                         ext4_free_blks_set(sb, gdp, free);
 725                         gdp->bg_checksum = ext4_group_desc_csum(sbi, group,
 726                                                                 gdp);
 727                 }
 728                 spin_unlock(sb_bgl_lock(sbi, group));
 729
 730                 /* Don't need to dirty bitmap block if we didn't change it */
 731                 if (free) {
 732                         BUFFER_TRACE(block_bh, "dirty block bitmap");
 733                         err = ext4_handle_dirty_metadata(handle,
 734                                                         NULL, block_bh);
 735                 }
 736
 737                 brelse(block_bh);
 738                 if (err)
 739                         goto fail;
 740         }
 741
 742         spin_lock(sb_bgl_lock(sbi, group));
 743         /* If we didn't allocate from within the initialized part of the inode
 744          * table then we need to initialize up to this inode. */
 745         if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM)) {
 746                 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) {
 747                         gdp->bg_flags &= cpu_to_le16(~EXT4_BG_INODE_UNINIT);
 748
 749                         /* When marking the block group with
 750                          * ~EXT4_BG_INODE_UNINIT we don't want to depend
 751                          * on the value of bg_itable_unused even though
 752                          * mke2fs could have initialized the same for us.
 753                          * Instead we calculated the value below
 754                          */
 755
 756                         free = 0;
 757                 } else {
 758                         free = EXT4_INODES_PER_GROUP(sb) -
 759                                 ext4_itable_unused_count(sb, gdp);
 760                 }
 761
 762                 /*
 763                  * Check the relative inode number against the last used
 764                  * relative inode number in this group. if it is greater
 765                  * we need to  update the bg_itable_unused count
 766                  *
 767                  */
 768                 if (ino > free)
 769                         ext4_itable_unused_set(sb, gdp,
 770                                         (EXT4_INODES_PER_GROUP(sb) - ino));
 771         }
 772
 773         count = ext4_free_inodes_count(sb, gdp) - 1;
 774         ext4_free_inodes_set(sb, gdp, count);
 775         if (S_ISDIR(mode)) {
 776                 count = ext4_used_dirs_count(sb, gdp) + 1;
 777                 ext4_used_dirs_set(sb, gdp, count);
 778         }
 779         gdp->bg_checksum = ext4_group_desc_csum(sbi, group, gdp);
 780         spin_unlock(sb_bgl_lock(sbi, group));
 781         BUFFER_TRACE(bh2, "call ext4_handle_dirty_metadata");
 782         err = ext4_handle_dirty_metadata(handle, NULL, bh2);
 783         if (err) goto fail;
 784
 785         percpu_counter_dec(&sbi->s_freeinodes_counter);
 786         if (S_ISDIR(mode))
 787                 percpu_counter_inc(&sbi->s_dirs_counter);
 788         sb->s_dirt = 1;
 789
 790         if (sbi->s_log_groups_per_flex) {
 791                 flex_group = ext4_flex_group(sbi, group);
 792                 spin_lock(sb_bgl_lock(sbi, flex_group));
 793                 sbi->s_flex_groups[flex_group].free_inodes--;
 794                 spin_unlock(sb_bgl_lock(sbi, flex_group));
 795         }
 796
 797         inode->i_uid = current_fsuid();
 798         if (test_opt(sb, GRPID))
 799                 inode->i_gid = dir->i_gid;
 800         else if (dir->i_mode & S_ISGID) {
 801                 inode->i_gid = dir->i_gid;
 802                 if (S_ISDIR(mode))
 803                         mode |= S_ISGID;
 804         } else
 805                 inode->i_gid = current_fsgid();
 806         inode->i_mode = mode;
 807
 808         inode->i_ino = ino + group * EXT4_INODES_PER_GROUP(sb);
 809         /* This is the optimal IO size (for stat), not the fs block size */
 810         inode->i_blocks = 0;
 811         inode->i_mtime = inode->i_atime = inode->i_ctime = ei->i_crtime =
 812                                                        ext4_current_time(inode);
 813
 814         memset(ei->i_data, 0, sizeof(ei->i_data));
 815         ei->i_dir_start_lookup = 0;
 816         ei->i_disksize = 0;
 817
 818         /*
 819          * Don't inherit extent flag from directory. We set extent flag on
 820          * newly created directory and file only if -o extent mount option is
 821          * specified
 822          */
 823         ei->i_flags = EXT4_I(dir)->i_flags & ~(EXT4_INDEX_FL|EXT4_EXTENTS_FL);
 824         if (S_ISLNK(mode))
 825                 ei->i_flags &= ~(EXT4_IMMUTABLE_FL|EXT4_APPEND_FL);
 826         /* dirsync only applies to directories */
 827         if (!S_ISDIR(mode))
 828                 ei->i_flags &= ~EXT4_DIRSYNC_FL;
 829         ei->i_file_acl = 0;
 830         ei->i_dtime = 0;
 831         ei->i_block_group = group;
 832
 833         ext4_set_inode_flags(inode);
 834         if (IS_DIRSYNC(inode))
 835                 ext4_handle_sync(handle);
 836         if (insert_inode_locked(inode) < 0) {
 837                 err = -EINVAL;
 838                 goto fail_drop;
 839         }
 840         spin_lock(&sbi->s_next_gen_lock);
 841         inode->i_generation = sbi->s_next_generation++;
 842         spin_unlock(&sbi->s_next_gen_lock);
 843
 844         ei->i_state = EXT4_STATE_NEW;
 845
 846         ei->i_extra_isize = EXT4_SB(sb)->s_want_extra_isize;
 847
 848         ret = inode;
 849         if (DQUOT_ALLOC_INODE(inode)) {
 850                 err = -EDQUOT;
 851                 goto fail_drop;
 852         }
 853
 854         err = ext4_init_acl(handle, inode, dir);
 855         if (err)
 856                 goto fail_free_drop;
 857
 858         err = ext4_init_security(handle, inode, dir);
 859         if (err)
 860                 goto fail_free_drop;
 861
 862         if (test_opt(sb, EXTENTS)) {
 863                 /* set extent flag only for directory, file and normal symlink*/
 864                 if (S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode)) {
 865                         EXT4_I(inode)->i_flags |= EXT4_EXTENTS_FL;
 866                         ext4_ext_tree_init(handle, inode);
 867                 }
 868         }
 869
 870         err = ext4_mark_inode_dirty(handle, inode);
 871         if (err) {
 872                 ext4_std_error(sb, err);
 873                 goto fail_free_drop;
 874         }
 875
 876         ext4_debug("allocating inode %lu\n", inode->i_ino);
 877         goto really_out;
 878 fail:
 879         ext4_std_error(sb, err);
 880 out:
 881         iput(inode);
 882         ret = ERR_PTR(err);
 883 really_out:
 884         brelse(bitmap_bh);
 885         return ret;
 886
 887 fail_free_drop:
 888         DQUOT_FREE_INODE(inode);
 889
 890 fail_drop:
 891         DQUOT_DROP(inode);
 892         inode->i_flags |= S_NOQUOTA;
 893         inode->i_nlink = 0;
 894         unlock_new_inode(inode);
 895         iput(inode);
 896         brelse(bitmap_bh);
 897         return ERR_PTR(err);
 898 }
 899
 900 /* Verify that we are loading a valid orphan from disk */
 901 struct inode *ext4_orphan_get(struct super_block *sb, unsigned long ino)
 902 {
 903         unsigned long max_ino = le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count);
 904         ext4_group_t block_group;
 905         int bit;
 906         struct buffer_head *bitmap_bh;
 907         struct inode *inode = NULL;
 908         long err = -EIO;
 909
 910         /* Error cases - e2fsck has already cleaned up for us */
 911         if (ino > max_ino) {
 912                 ext4_warning(sb, __func__,
 913                              "bad orphan ino %lu!  e2fsck was run?", ino);
 914                 goto error;
 915         }
 916
 917         block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb);
 918         bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb);
 919         bitmap_bh = ext4_read_inode_bitmap(sb, block_group);
 920         if (!bitmap_bh) {
 921                 ext4_warning(sb, __func__,
 922                              "inode bitmap error for orphan %lu", ino);
 923                 goto error;
 924         }
 925
 926         /* Having the inode bit set should be a 100% indicator that this
 927          * is a valid orphan (no e2fsck run on fs).  Orphans also include
 928          * inodes that were being truncated, so we can't check i_nlink==0.
 929          */
 930         if (!ext4_test_bit(bit, bitmap_bh->b_data))
 931                 goto bad_orphan;
 932
 933         inode = ext4_iget(sb, ino);
 934         if (IS_ERR(inode))
 935                 goto iget_failed;
 936
 937         /*
 938          * If the orphans has i_nlinks > 0 then it should be able to be
 939          * truncated, otherwise it won't be removed from the orphan list
 940          * during processing and an infinite loop will result.
 941          */
 942         if (inode->i_nlink && !ext4_can_truncate(inode))
 943                 goto bad_orphan;
 944
 945         if (NEXT_ORPHAN(inode) > max_ino)
 946                 goto bad_orphan;
 947         brelse(bitmap_bh);
 948         return inode;
 949
 950 iget_failed:
 951         err = PTR_ERR(inode);
 952         inode = NULL;
 953 bad_orphan:
 954         ext4_warning(sb, __func__,
 955                      "bad orphan inode %lu!  e2fsck was run?", ino);
 956         printk(KERN_NOTICE "ext4_test_bit(bit=%d, block=%llu) = %d\n",
 957                bit, (unsigned long long)bitmap_bh->b_blocknr,
 958                ext4_test_bit(bit, bitmap_bh->b_data));
 959         printk(KERN_NOTICE "inode=%p\n", inode);
 960         if (inode) {
 961                 printk(KERN_NOTICE "is_bad_inode(inode)=%d\n",
 962                        is_bad_inode(inode));
 963                 printk(KERN_NOTICE "NEXT_ORPHAN(inode)=%u\n",
 964                        NEXT_ORPHAN(inode));
 965                 printk(KERN_NOTICE "max_ino=%lu\n", max_ino);
 966                 printk(KERN_NOTICE "i_nlink=%u\n", inode->i_nlink);
 967                 /* Avoid freeing blocks if we got a bad deleted inode */
 968                 if (inode->i_nlink == 0)
 969                         inode->i_blocks = 0;
 970                 iput(inode);
 971         }
 972         brelse(bitmap_bh);
 973 error:
 974         return ERR_PTR(err);
 975 }
 976
 977 unsigned long ext4_count_free_inodes(struct super_block *sb)
 978 {
 979         unsigned long desc_count;
 980         struct ext4_group_desc *gdp;
 981         ext4_group_t i;
 982 #ifdef EXT4FS_DEBUG
 983         struct ext4_super_block *es;
 984         unsigned long bitmap_count, x;
 985         struct buffer_head *bitmap_bh = NULL;
 986
 987         es = EXT4_SB(sb)->s_es;
 988         desc_count = 0;
 989         bitmap_count = 0;
 990         gdp = NULL;
 991         for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
 992                 gdp = ext4_get_group_desc(sb, i, NULL);
 993                 if (!gdp)
 994                         continue;
 995                 desc_count += ext4_free_inodes_count(sb, gdp);
 996                 brelse(bitmap_bh);
 997                 bitmap_bh = ext4_read_inode_bitmap(sb, i);
 998                 if (!bitmap_bh)
 999                         continue;
1000
1001                 x = ext4_count_free(bitmap_bh, EXT4_INODES_PER_GROUP(sb) / 8);
1002                 printk(KERN_DEBUG "group %lu: stored = %d, counted = %lu\n",
1003                         i, ext4_free_inodes_count(sb, gdp), x);
1004                 bitmap_count += x;
1005         }
1006         brelse(bitmap_bh);
1007         printk(KERN_DEBUG "ext4_count_free_inodes: "
1008                "stored = %u, computed = %lu, %lu\n",
1009                le32_to_cpu(es->s_free_inodes_count), desc_count, bitmap_count);
1010         return desc_count;
1011 #else
1012         desc_count = 0;
1013         for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
1014                 gdp = ext4_get_group_desc(sb, i, NULL);
1015                 if (!gdp)
1016                         continue;
1017                 desc_count += ext4_free_inodes_count(sb, gdp);
1018                 cond_resched();
1019         }
1020         return desc_count;
1021 #endif
1022 }
1023
1024 /* Called at mount-time, super-block is locked */
1025 unsigned long ext4_count_dirs(struct super_block * sb)
1026 {
1027         unsigned long count = 0;
1028         ext4_group_t i;
1029
1030         for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
1031                 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, i, NULL);
1032                 if (!gdp)
1033                         continue;
1034                 count += ext4_used_dirs_count(sb, gdp);
1035         }
1036         return count;
1037 }