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), sb->s_blocksize * 8,
  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 (bitmap_uptodate(bh))
 119                 return bh;
 120
 121         lock_buffer(bh);
 122         if (bitmap_uptodate(bh)) {
 123                 unlock_buffer(bh);
 124                 return bh;
 125         }
 126         spin_lock(sb_bgl_lock(EXT4_SB(sb), block_group));
 127         if (desc->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) {
 128                 ext4_init_inode_bitmap(sb, bh, block_group, desc);
 129                 set_bitmap_uptodate(bh);
 130                 set_buffer_uptodate(bh);
 131                 unlock_buffer(bh);
 132                 spin_unlock(sb_bgl_lock(EXT4_SB(sb), block_group));
 133                 return bh;
 134         }
 135         spin_unlock(sb_bgl_lock(EXT4_SB(sb), block_group));
 136         if (buffer_uptodate(bh)) {
 137                 /*
 138                  * if not uninit if bh is uptodate,
 139                  * bitmap is also uptodate
 140                  */
 141                 set_bitmap_uptodate(bh);
 142                 unlock_buffer(bh);
 143                 return bh;
 144         }
 145         /*
 146          * submit the buffer_head for read. We can
 147          * safely mark the bitmap as uptodate now.
 148          * We do it here so the bitmap uptodate bit
 149          * get set with buffer lock held.
 150          */
 151         set_bitmap_uptodate(bh);
 152         if (bh_submit_read(bh) < 0) {
 153                 put_bh(bh);
 154                 ext4_error(sb, __func__,
 155                             "Cannot read inode bitmap - "
 156                             "block_group = %lu, inode_bitmap = %llu",
 157                             block_group, bitmap_blk);
 158                 return NULL;
 159         }
 160         return bh;
 161 }
 162
 163 /*
 164  * NOTE! When we get the inode, we're the only people
 165  * that have access to it, and as such there are no
 166  * race conditions we have to worry about. The inode
 167  * is not on the hash-lists, and it cannot be reached
 168  * through the filesystem because the directory entry
 169  * has been deleted earlier.
 170  *
 171  * HOWEVER: we must make sure that we get no aliases,
 172  * which means that we have to call "clear_inode()"
 173  * _before_ we mark the inode not in use in the inode
 174  * bitmaps. Otherwise a newly created file might use
 175  * the same inode number (not actually the same pointer
 176  * though), and then we'd have two inodes sharing the
 177  * same inode number and space on the harddisk.
 178  */
 179 void ext4_free_inode (handle_t *handle, struct inode * inode)
 180 {
 181         struct super_block * sb = inode->i_sb;
 182         int is_directory;
 183         unsigned long ino;
 184         struct buffer_head *bitmap_bh = NULL;
 185         struct buffer_head *bh2;
 186         ext4_group_t block_group;
 187         unsigned long bit;
 188         struct ext4_group_desc * gdp;
 189         struct ext4_super_block * es;
 190         struct ext4_sb_info *sbi;
 191         int fatal = 0, err;
 192         ext4_group_t flex_group;
 193
 194         if (atomic_read(&inode->i_count) > 1) {
 195                 printk ("ext4_free_inode: inode has count=%d\n",
 196                                         atomic_read(&inode->i_count));
 197                 return;
 198         }
 199         if (inode->i_nlink) {
 200                 printk ("ext4_free_inode: inode has nlink=%d\n",
 201                         inode->i_nlink);
 202                 return;
 203         }
 204         if (!sb) {
 205                 printk("ext4_free_inode: inode on nonexistent device\n");
 206                 return;
 207         }
 208         sbi = EXT4_SB(sb);
 209
 210         ino = inode->i_ino;
 211         ext4_debug ("freeing inode %lu\n", ino);
 212
 213         /*
 214          * Note: we must free any quota before locking the superblock,
 215          * as writing the quota to disk may need the lock as well.
 216          */
 217         DQUOT_INIT(inode);
 218         ext4_xattr_delete_inode(handle, inode);
 219         DQUOT_FREE_INODE(inode);
 220         DQUOT_DROP(inode);
 221
 222         is_directory = S_ISDIR(inode->i_mode);
 223
 224         /* Do this BEFORE marking the inode not in use or returning an error */
 225         clear_inode (inode);
 226
 227         es = EXT4_SB(sb)->s_es;
 228         if (ino < EXT4_FIRST_INO(sb) || ino > le32_to_cpu(es->s_inodes_count)) {
 229                 ext4_error (sb, "ext4_free_inode",
 230                             "reserved or nonexistent inode %lu", ino);
 231                 goto error_return;
 232         }
 233         block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb);
 234         bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb);
 235         bitmap_bh = ext4_read_inode_bitmap(sb, block_group);
 236         if (!bitmap_bh)
 237                 goto error_return;
 238
 239         BUFFER_TRACE(bitmap_bh, "get_write_access");
 240         fatal = ext4_journal_get_write_access(handle, bitmap_bh);
 241         if (fatal)
 242                 goto error_return;
 243
 244         /* Ok, now we can actually update the inode bitmaps.. */
 245         if (!ext4_clear_bit_atomic(sb_bgl_lock(sbi, block_group),
 246                                         bit, bitmap_bh->b_data))
 247                 ext4_error (sb, "ext4_free_inode",
 248                               "bit already cleared for inode %lu", ino);
 249         else {
 250                 gdp = ext4_get_group_desc (sb, block_group, &bh2);
 251
 252                 BUFFER_TRACE(bh2, "get_write_access");
 253                 fatal = ext4_journal_get_write_access(handle, bh2);
 254                 if (fatal) goto error_return;
 255
 256                 if (gdp) {
 257                         spin_lock(sb_bgl_lock(sbi, block_group));
 258                         le16_add_cpu(&gdp->bg_free_inodes_count, 1);
 259                         if (is_directory)
 260                                 le16_add_cpu(&gdp->bg_used_dirs_count, -1);
 261                         gdp->bg_checksum = ext4_group_desc_csum(sbi,
 262                                                         block_group, gdp);
 263                         spin_unlock(sb_bgl_lock(sbi, block_group));
 264                         percpu_counter_inc(&sbi->s_freeinodes_counter);
 265                         if (is_directory)
 266                                 percpu_counter_dec(&sbi->s_dirs_counter);
 267
 268                         if (sbi->s_log_groups_per_flex) {
 269                                 flex_group = ext4_flex_group(sbi, block_group);
 270                                 spin_lock(sb_bgl_lock(sbi, flex_group));
 271                                 sbi->s_flex_groups[flex_group].free_inodes++;
 272                                 spin_unlock(sb_bgl_lock(sbi, flex_group));
 273                         }
 274                 }
 275                 BUFFER_TRACE(bh2, "call ext4_journal_dirty_metadata");
 276                 err = ext4_journal_dirty_metadata(handle, bh2);
 277                 if (!fatal) fatal = err;
 278         }
 279         BUFFER_TRACE(bitmap_bh, "call ext4_journal_dirty_metadata");
 280         err = ext4_journal_dirty_metadata(handle, bitmap_bh);
 281         if (!fatal)
 282                 fatal = err;
 283         sb->s_dirt = 1;
 284 error_return:
 285         brelse(bitmap_bh);
 286         ext4_std_error(sb, fatal);
 287 }
 288
 289 /*
 290  * There are two policies for allocating an inode.  If the new inode is
 291  * a directory, then a forward search is made for a block group with both
 292  * free space and a low directory-to-inode ratio; if that fails, then of
 293  * the groups with above-average free space, that group with the fewest
 294  * directories already is chosen.
 295  *
 296  * For other inodes, search forward from the parent directory\'s block
 297  * group to find a free inode.
 298  */
 299 static int find_group_dir(struct super_block *sb, struct inode *parent,
 300                                 ext4_group_t *best_group)
 301 {
 302         ext4_group_t ngroups = EXT4_SB(sb)->s_groups_count;
 303         unsigned int freei, avefreei;
 304         struct ext4_group_desc *desc, *best_desc = NULL;
 305         ext4_group_t group;
 306         int ret = -1;
 307
 308         freei = percpu_counter_read_positive(&EXT4_SB(sb)->s_freeinodes_counter);
 309         avefreei = freei / ngroups;
 310
 311         for (group = 0; group < ngroups; group++) {
 312                 desc = ext4_get_group_desc (sb, group, NULL);
 313                 if (!desc || !desc->bg_free_inodes_count)
 314                         continue;
 315                 if (le16_to_cpu(desc->bg_free_inodes_count) < avefreei)
 316                         continue;
 317                 if (!best_desc ||
 318                     (le16_to_cpu(desc->bg_free_blocks_count) >
 319                      le16_to_cpu(best_desc->bg_free_blocks_count))) {
 320                         *best_group = group;
 321                         best_desc = desc;
 322                         ret = 0;
 323                 }
 324         }
 325         return ret;
 326 }
 327
 328 #define free_block_ratio 10
 329
 330 static int find_group_flex(struct super_block *sb, struct inode *parent,
 331                            ext4_group_t *best_group)
 332 {
 333         struct ext4_sb_info *sbi = EXT4_SB(sb);
 334         struct ext4_group_desc *desc;
 335         struct buffer_head *bh;
 336         struct flex_groups *flex_group = sbi->s_flex_groups;
 337         ext4_group_t parent_group = EXT4_I(parent)->i_block_group;
 338         ext4_group_t parent_fbg_group = ext4_flex_group(sbi, parent_group);
 339         ext4_group_t ngroups = sbi->s_groups_count;
 340         int flex_size = ext4_flex_bg_size(sbi);
 341         ext4_group_t best_flex = parent_fbg_group;
 342         int blocks_per_flex = sbi->s_blocks_per_group * flex_size;
 343         int flexbg_free_blocks;
 344         int flex_freeb_ratio;
 345         ext4_group_t n_fbg_groups;
 346         ext4_group_t i;
 347
 348         n_fbg_groups = (sbi->s_groups_count + flex_size - 1) >>
 349                 sbi->s_log_groups_per_flex;
 350
 351 find_close_to_parent:
 352         flexbg_free_blocks = flex_group[best_flex].free_blocks;
 353         flex_freeb_ratio = flexbg_free_blocks * 100 / blocks_per_flex;
 354         if (flex_group[best_flex].free_inodes &&
 355             flex_freeb_ratio > free_block_ratio)
 356                 goto found_flexbg;
 357
 358         if (best_flex && best_flex == parent_fbg_group) {
 359                 best_flex--;
 360                 goto find_close_to_parent;
 361         }
 362
 363         for (i = 0; i < n_fbg_groups; i++) {
 364                 if (i == parent_fbg_group || i == parent_fbg_group - 1)
 365                         continue;
 366
 367                 flexbg_free_blocks = flex_group[i].free_blocks;
 368                 flex_freeb_ratio = flexbg_free_blocks * 100 / blocks_per_flex;
 369
 370                 if (flex_freeb_ratio > free_block_ratio &&
 371                     flex_group[i].free_inodes) {
 372                         best_flex = i;
 373                         goto found_flexbg;
 374                 }
 375
 376                 if (flex_group[best_flex].free_inodes == 0 ||
 377                     (flex_group[i].free_blocks >
 378                      flex_group[best_flex].free_blocks &&
 379                      flex_group[i].free_inodes))
 380                         best_flex = i;
 381         }
 382
 383         if (!flex_group[best_flex].free_inodes ||
 384             !flex_group[best_flex].free_blocks)
 385                 return -1;
 386
 387 found_flexbg:
 388         for (i = best_flex * flex_size; i < ngroups &&
 389                      i < (best_flex + 1) * flex_size; i++) {
 390                 desc = ext4_get_group_desc(sb, i, &bh);
 391                 if (le16_to_cpu(desc->bg_free_inodes_count)) {
 392                         *best_group = i;
 393                         goto out;
 394                 }
 395         }
 396
 397         return -1;
 398 out:
 399         return 0;
 400 }
 401
 402 /*
 403  * Orlov's allocator for directories.
 404  *
 405  * We always try to spread first-level directories.
 406  *
 407  * If there are blockgroups with both free inodes and free blocks counts
 408  * not worse than average we return one with smallest directory count.
 409  * Otherwise we simply return a random group.
 410  *
 411  * For the rest rules look so:
 412  *
 413  * It's OK to put directory into a group unless
 414  * it has too many directories already (max_dirs) or
 415  * it has too few free inodes left (min_inodes) or
 416  * it has too few free blocks left (min_blocks) or
 417  * it's already running too large debt (max_debt).
 418  * Parent's group is preferred, if it doesn't satisfy these
 419  * conditions we search cyclically through the rest. If none
 420  * of the groups look good we just look for a group with more
 421  * free inodes than average (starting at parent's group).
 422  *
 423  * Debt is incremented each time we allocate a directory and decremented
 424  * when we allocate an inode, within 0--255.
 425  */
 426
 427 #define INODE_COST 64
 428 #define BLOCK_COST 256
 429
 430 static int find_group_orlov(struct super_block *sb, struct inode *parent,
 431                                 ext4_group_t *group)
 432 {
 433         ext4_group_t parent_group = EXT4_I(parent)->i_block_group;
 434         struct ext4_sb_info *sbi = EXT4_SB(sb);
 435         struct ext4_super_block *es = sbi->s_es;
 436         ext4_group_t ngroups = sbi->s_groups_count;
 437         int inodes_per_group = EXT4_INODES_PER_GROUP(sb);
 438         unsigned int freei, avefreei;
 439         ext4_fsblk_t freeb, avefreeb;
 440         ext4_fsblk_t blocks_per_dir;
 441         unsigned int ndirs;
 442         int max_debt, max_dirs, min_inodes;
 443         ext4_grpblk_t min_blocks;
 444         ext4_group_t i;
 445         struct ext4_group_desc *desc;
 446
 447         freei = percpu_counter_read_positive(&sbi->s_freeinodes_counter);
 448         avefreei = freei / ngroups;
 449         freeb = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
 450         avefreeb = freeb;
 451         do_div(avefreeb, ngroups);
 452         ndirs = percpu_counter_read_positive(&sbi->s_dirs_counter);
 453
 454         if ((parent == sb->s_root->d_inode) ||
 455             (EXT4_I(parent)->i_flags & EXT4_TOPDIR_FL)) {
 456                 int best_ndir = inodes_per_group;
 457                 ext4_group_t grp;
 458                 int ret = -1;
 459
 460                 get_random_bytes(&grp, sizeof(grp));
 461                 parent_group = (unsigned)grp % ngroups;
 462                 for (i = 0; i < ngroups; i++) {
 463                         grp = (parent_group + i) % ngroups;
 464                         desc = ext4_get_group_desc(sb, grp, NULL);
 465                         if (!desc || !desc->bg_free_inodes_count)
 466                                 continue;
 467                         if (le16_to_cpu(desc->bg_used_dirs_count) >= best_ndir)
 468                                 continue;
 469                         if (le16_to_cpu(desc->bg_free_inodes_count) < avefreei)
 470                                 continue;
 471                         if (le16_to_cpu(desc->bg_free_blocks_count) < avefreeb)
 472                                 continue;
 473                         *group = grp;
 474                         ret = 0;
 475                         best_ndir = le16_to_cpu(desc->bg_used_dirs_count);
 476                 }
 477                 if (ret == 0)
 478                         return ret;
 479                 goto fallback;
 480         }
 481
 482         blocks_per_dir = ext4_blocks_count(es) - freeb;
 483         do_div(blocks_per_dir, ndirs);
 484
 485         max_dirs = ndirs / ngroups + inodes_per_group / 16;
 486         min_inodes = avefreei - inodes_per_group / 4;
 487         min_blocks = avefreeb - EXT4_BLOCKS_PER_GROUP(sb) / 4;
 488
 489         max_debt = EXT4_BLOCKS_PER_GROUP(sb);
 490         max_debt /= max_t(int, blocks_per_dir, BLOCK_COST);
 491         if (max_debt * INODE_COST > inodes_per_group)
 492                 max_debt = inodes_per_group / INODE_COST;
 493         if (max_debt > 255)
 494                 max_debt = 255;
 495         if (max_debt == 0)
 496                 max_debt = 1;
 497
 498         for (i = 0; i < ngroups; i++) {
 499                 *group = (parent_group + i) % ngroups;
 500                 desc = ext4_get_group_desc(sb, *group, NULL);
 501                 if (!desc || !desc->bg_free_inodes_count)
 502                         continue;
 503                 if (le16_to_cpu(desc->bg_used_dirs_count) >= max_dirs)
 504                         continue;
 505                 if (le16_to_cpu(desc->bg_free_inodes_count) < min_inodes)
 506                         continue;
 507                 if (le16_to_cpu(desc->bg_free_blocks_count) < min_blocks)
 508                         continue;
 509                 return 0;
 510         }
 511
 512 fallback:
 513         for (i = 0; i < ngroups; i++) {
 514                 *group = (parent_group + i) % ngroups;
 515                 desc = ext4_get_group_desc(sb, *group, NULL);
 516                 if (desc && desc->bg_free_inodes_count &&
 517                         le16_to_cpu(desc->bg_free_inodes_count) >= avefreei)
 518                         return 0;
 519         }
 520
 521         if (avefreei) {
 522                 /*
 523                  * The free-inodes counter is approximate, and for really small
 524                  * filesystems the above test can fail to find any blockgroups
 525                  */
 526                 avefreei = 0;
 527                 goto fallback;
 528         }
 529
 530         return -1;
 531 }
 532
 533 static int find_group_other(struct super_block *sb, struct inode *parent,
 534                                 ext4_group_t *group)
 535 {
 536         ext4_group_t parent_group = EXT4_I(parent)->i_block_group;
 537         ext4_group_t ngroups = EXT4_SB(sb)->s_groups_count;
 538         struct ext4_group_desc *desc;
 539         ext4_group_t i;
 540
 541         /*
 542          * Try to place the inode in its parent directory
 543          */
 544         *group = parent_group;
 545         desc = ext4_get_group_desc(sb, *group, NULL);
 546         if (desc && le16_to_cpu(desc->bg_free_inodes_count) &&
 547                         le16_to_cpu(desc->bg_free_blocks_count))
 548                 return 0;
 549
 550         /*
 551          * We're going to place this inode in a different blockgroup from its
 552          * parent.  We want to cause files in a common directory to all land in
 553          * the same blockgroup.  But we want files which are in a different
 554          * directory which shares a blockgroup with our parent to land in a
 555          * different blockgroup.
 556          *
 557          * So add our directory's i_ino into the starting point for the hash.
 558          */
 559         *group = (*group + parent->i_ino) % ngroups;
 560
 561         /*
 562          * Use a quadratic hash to find a group with a free inode and some free
 563          * blocks.
 564          */
 565         for (i = 1; i < ngroups; i <<= 1) {
 566                 *group += i;
 567                 if (*group >= ngroups)
 568                         *group -= ngroups;
 569                 desc = ext4_get_group_desc(sb, *group, NULL);
 570                 if (desc && le16_to_cpu(desc->bg_free_inodes_count) &&
 571                                 le16_to_cpu(desc->bg_free_blocks_count))
 572                         return 0;
 573         }
 574
 575         /*
 576          * That failed: try linear search for a free inode, even if that group
 577          * has no free blocks.
 578          */
 579         *group = parent_group;
 580         for (i = 0; i < ngroups; i++) {
 581                 if (++*group >= ngroups)
 582                         *group = 0;
 583                 desc = ext4_get_group_desc(sb, *group, NULL);
 584                 if (desc && le16_to_cpu(desc->bg_free_inodes_count))
 585                         return 0;
 586         }
 587
 588         return -1;
 589 }
 590
 591 /*
 592  * claim the inode from the inode bitmap. If the group
 593  * is uninit we need to take the groups's sb_bgl_lock
 594  * and clear the uninit flag. The inode bitmap update
 595  * and group desc uninit flag clear should be done
 596  * after holding sb_bgl_lock so that ext4_read_inode_bitmap
 597  * doesn't race with the ext4_claim_inode
 598  */
 599 static int ext4_claim_inode(struct super_block *sb,
 600                         struct buffer_head *inode_bitmap_bh,
 601                         unsigned long ino, ext4_group_t group, int mode)
 602 {
 603         int free = 0, retval = 0;
 604         struct ext4_sb_info *sbi = EXT4_SB(sb);
 605         struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, NULL);
 606
 607         spin_lock(sb_bgl_lock(sbi, group));
 608         if (ext4_set_bit(ino, inode_bitmap_bh->b_data)) {
 609                 /* not a free inode */
 610                 retval = 1;
 611                 goto err_ret;
 612         }
 613         ino++;
 614         if ((group == 0 && ino < EXT4_FIRST_INO(sb)) ||
 615                         ino > EXT4_INODES_PER_GROUP(sb)) {
 616                 spin_unlock(sb_bgl_lock(sbi, group));
 617                 ext4_error(sb, __func__,
 618                            "reserved inode or inode > inodes count - "
 619                            "block_group = %lu, inode=%lu", group,
 620                            ino + group * EXT4_INODES_PER_GROUP(sb));
 621                 return 1;
 622         }
 623         /* If we didn't allocate from within the initialized part of the inode
 624          * table then we need to initialize up to this inode. */
 625         if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM)) {
 626
 627                 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) {
 628                         gdp->bg_flags &= cpu_to_le16(~EXT4_BG_INODE_UNINIT);
 629                         /* When marking the block group with
 630                          * ~EXT4_BG_INODE_UNINIT we don't want to depend
 631                          * on the value of bg_itable_unused even though
 632                          * mke2fs could have initialized the same for us.
 633                          * Instead we calculated the value below
 634                          */
 635
 636                         free = 0;
 637                 } else {
 638                         free = EXT4_INODES_PER_GROUP(sb) -
 639                                 le16_to_cpu(gdp->bg_itable_unused);
 640                 }
 641
 642                 /*
 643                  * Check the relative inode number against the last used
 644                  * relative inode number in this group. if it is greater
 645                  * we need to  update the bg_itable_unused count
 646                  *
 647                  */
 648                 if (ino > free)
 649                         gdp->bg_itable_unused =
 650                                 cpu_to_le16(EXT4_INODES_PER_GROUP(sb) - ino);
 651         }
 652         le16_add_cpu(&gdp->bg_free_inodes_count, -1);
 653         if (S_ISDIR(mode)) {
 654                 le16_add_cpu(&gdp->bg_used_dirs_count, 1);
 655         }
 656         gdp->bg_checksum = ext4_group_desc_csum(sbi, group, gdp);
 657 err_ret:
 658         spin_unlock(sb_bgl_lock(sbi, group));
 659         return retval;
 660 }
 661
 662 /*
 663  * There are two policies for allocating an inode.  If the new inode is
 664  * a directory, then a forward search is made for a block group with both
 665  * free space and a low directory-to-inode ratio; if that fails, then of
 666  * the groups with above-average free space, that group with the fewest
 667  * directories already is chosen.
 668  *
 669  * For other inodes, search forward from the parent directory's block
 670  * group to find a free inode.
 671  */
 672 struct inode *ext4_new_inode(handle_t *handle, struct inode * dir, int mode)
 673 {
 674         struct super_block *sb;
 675         struct buffer_head *bitmap_bh = NULL;
 676         struct buffer_head *bh2;
 677         ext4_group_t group = 0;
 678         unsigned long ino = 0;
 679         struct inode * inode;
 680         struct ext4_group_desc * gdp = NULL;
 681         struct ext4_super_block * es;
 682         struct ext4_inode_info *ei;
 683         struct ext4_sb_info *sbi;
 684         int ret2, err = 0;
 685         struct inode *ret;
 686         ext4_group_t i;
 687         int free = 0;
 688         ext4_group_t flex_group;
 689
 690         /* Cannot create files in a deleted directory */
 691         if (!dir || !dir->i_nlink)
 692                 return ERR_PTR(-EPERM);
 693
 694         sb = dir->i_sb;
 695         inode = new_inode(sb);
 696         if (!inode)
 697                 return ERR_PTR(-ENOMEM);
 698         ei = EXT4_I(inode);
 699
 700         sbi = EXT4_SB(sb);
 701         es = sbi->s_es;
 702
 703         if (sbi->s_log_groups_per_flex) {
 704                 ret2 = find_group_flex(sb, dir, &group);
 705                 if (ret2 == -1) {
 706                         ret2 = find_group_other(sb, dir, &group);
 707                         if (ret2 == 0 && printk_ratelimit())
 708                                 printk(KERN_NOTICE "ext4: find_group_flex "
 709                                        "failed, fallback succeeded dir %lu\n",
 710                                        dir->i_ino);
 711                 }
 712                 goto got_group;
 713         }
 714
 715         if (S_ISDIR(mode)) {
 716                 if (test_opt (sb, OLDALLOC))
 717                         ret2 = find_group_dir(sb, dir, &group);
 718                 else
 719                         ret2 = find_group_orlov(sb, dir, &group);
 720         } else
 721                 ret2 = find_group_other(sb, dir, &group);
 722
 723 got_group:
 724         err = -ENOSPC;
 725         if (ret2 == -1)
 726                 goto out;
 727
 728         for (i = 0; i < sbi->s_groups_count; i++) {
 729                 err = -EIO;
 730
 731                 gdp = ext4_get_group_desc(sb, group, &bh2);
 732                 if (!gdp)
 733                         goto fail;
 734
 735                 brelse(bitmap_bh);
 736                 bitmap_bh = ext4_read_inode_bitmap(sb, group);
 737                 if (!bitmap_bh)
 738                         goto fail;
 739
 740                 ino = 0;
 741
 742 repeat_in_this_group:
 743                 ino = ext4_find_next_zero_bit((unsigned long *)
 744                                 bitmap_bh->b_data, EXT4_INODES_PER_GROUP(sb), ino);
 745                 if (ino < EXT4_INODES_PER_GROUP(sb)) {
 746
 747                         BUFFER_TRACE(bitmap_bh, "get_write_access");
 748                         err = ext4_journal_get_write_access(handle, bitmap_bh);
 749                         if (err)
 750                                 goto fail;
 751
 752                         BUFFER_TRACE(bh2, "get_write_access");
 753                         err = ext4_journal_get_write_access(handle, bh2);
 754                         if (err)
 755                                 goto fail;
 756                         if (!ext4_claim_inode(sb, bitmap_bh,
 757                                                 ino, group, mode)) {
 758                                 /* we won it */
 759                                 BUFFER_TRACE(bitmap_bh,
 760                                         "call ext4_journal_dirty_metadata");
 761                                 err = ext4_journal_dirty_metadata(handle,
 762                                                                 bitmap_bh);
 763                                 if (err)
 764                                         goto fail;
 765                                 /* zero bit is inode number 1*/
 766                                 ino++;
 767                                 goto got;
 768                         }
 769                         /* we lost it */
 770                         jbd2_journal_release_buffer(handle, bitmap_bh);
 771                         jbd2_journal_release_buffer(handle, bh2);
 772
 773                         if (++ino < EXT4_INODES_PER_GROUP(sb))
 774                                 goto repeat_in_this_group;
 775                 }
 776
 777                 /*
 778                  * This case is possible in concurrent environment.  It is very
 779                  * rare.  We cannot repeat the find_group_xxx() call because
 780                  * that will simply return the same blockgroup, because the
 781                  * group descriptor metadata has not yet been updated.
 782                  * So we just go onto the next blockgroup.
 783                  */
 784                 if (++group == sbi->s_groups_count)
 785                         group = 0;
 786         }
 787         err = -ENOSPC;
 788         goto out;
 789
 790 got:
 791         /* We may have to initialize the block bitmap if it isn't already */
 792         if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM) &&
 793             gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
 794                 struct buffer_head *block_bh = ext4_read_block_bitmap(sb, group);
 795
 796                 BUFFER_TRACE(block_bh, "get block bitmap access");
 797                 err = ext4_journal_get_write_access(handle, block_bh);
 798                 if (err) {
 799                         brelse(block_bh);
 800                         goto fail;
 801                 }
 802
 803                 free = 0;
 804                 spin_lock(sb_bgl_lock(sbi, group));
 805                 /* recheck and clear flag under lock if we still need to */
 806                 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
 807                         gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
 808                         free = ext4_free_blocks_after_init(sb, group, gdp);
 809                         gdp->bg_free_blocks_count = cpu_to_le16(free);
 810                         gdp->bg_checksum = ext4_group_desc_csum(sbi, group,
 811                                                                 gdp);
 812                 }
 813                 spin_unlock(sb_bgl_lock(sbi, group));
 814
 815                 /* Don't need to dirty bitmap block if we didn't change it */
 816                 if (free) {
 817                         BUFFER_TRACE(block_bh, "dirty block bitmap");
 818                         err = ext4_journal_dirty_metadata(handle, block_bh);
 819                 }
 820
 821                 brelse(block_bh);
 822                 if (err)
 823                         goto fail;
 824         }
 825         BUFFER_TRACE(bh2, "call ext4_handle_dirty_metadata");
 826         err = ext4_journal_dirty_metadata(handle, bh2);
 827         if (err)
 828                 goto fail;
 829
 830         percpu_counter_dec(&sbi->s_freeinodes_counter);
 831         if (S_ISDIR(mode))
 832                 percpu_counter_inc(&sbi->s_dirs_counter);
 833         sb->s_dirt = 1;
 834
 835         if (sbi->s_log_groups_per_flex) {
 836                 flex_group = ext4_flex_group(sbi, group);
 837                 spin_lock(sb_bgl_lock(sbi, flex_group));
 838                 sbi->s_flex_groups[flex_group].free_inodes--;
 839                 spin_unlock(sb_bgl_lock(sbi, flex_group));
 840         }
 841
 842         inode->i_uid = current->fsuid;
 843         if (test_opt (sb, GRPID))
 844                 inode->i_gid = dir->i_gid;
 845         else if (dir->i_mode & S_ISGID) {
 846                 inode->i_gid = dir->i_gid;
 847                 if (S_ISDIR(mode))
 848                         mode |= S_ISGID;
 849         } else
 850                 inode->i_gid = current->fsgid;
 851         inode->i_mode = mode;
 852
 853         inode->i_ino = ino + group * EXT4_INODES_PER_GROUP(sb);
 854         /* This is the optimal IO size (for stat), not the fs block size */
 855         inode->i_blocks = 0;
 856         inode->i_mtime = inode->i_atime = inode->i_ctime = ei->i_crtime =
 857                                                        ext4_current_time(inode);
 858
 859         memset(ei->i_data, 0, sizeof(ei->i_data));
 860         ei->i_dir_start_lookup = 0;
 861         ei->i_disksize = 0;
 862
 863         /*
 864          * Don't inherit extent flag from directory. We set extent flag on
 865          * newly created directory and file only if -o extent mount option is
 866          * specified
 867          */
 868         ei->i_flags = EXT4_I(dir)->i_flags & ~(EXT4_INDEX_FL|EXT4_EXTENTS_FL);
 869         if (S_ISLNK(mode))
 870                 ei->i_flags &= ~(EXT4_IMMUTABLE_FL|EXT4_APPEND_FL);
 871         /* dirsync only applies to directories */
 872         if (!S_ISDIR(mode))
 873                 ei->i_flags &= ~EXT4_DIRSYNC_FL;
 874         ei->i_file_acl = 0;
 875         ei->i_dtime = 0;
 876         ei->i_block_alloc_info = NULL;
 877         ei->i_block_group = group;
 878
 879         ext4_set_inode_flags(inode);
 880         if (IS_DIRSYNC(inode))
 881                 handle->h_sync = 1;
 882         insert_inode_hash(inode);
 883         spin_lock(&sbi->s_next_gen_lock);
 884         inode->i_generation = sbi->s_next_generation++;
 885         spin_unlock(&sbi->s_next_gen_lock);
 886
 887         ei->i_state = EXT4_STATE_NEW;
 888
 889         ei->i_extra_isize = EXT4_SB(sb)->s_want_extra_isize;
 890
 891         ret = inode;
 892         if(DQUOT_ALLOC_INODE(inode)) {
 893                 err = -EDQUOT;
 894                 goto fail_drop;
 895         }
 896
 897         err = ext4_init_acl(handle, inode, dir);
 898         if (err)
 899                 goto fail_free_drop;
 900
 901         err = ext4_init_security(handle,inode, dir);
 902         if (err)
 903                 goto fail_free_drop;
 904
 905         if (test_opt(sb, EXTENTS)) {
 906                 /* set extent flag only for directory, file and normal symlink*/
 907                 if (S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode)) {
 908                         EXT4_I(inode)->i_flags |= EXT4_EXTENTS_FL;
 909                         ext4_ext_tree_init(handle, inode);
 910                 }
 911         }
 912
 913         err = ext4_mark_inode_dirty(handle, inode);
 914         if (err) {
 915                 ext4_std_error(sb, err);
 916                 goto fail_free_drop;
 917         }
 918
 919         ext4_debug("allocating inode %lu\n", inode->i_ino);
 920         goto really_out;
 921 fail:
 922         ext4_std_error(sb, err);
 923 out:
 924         iput(inode);
 925         ret = ERR_PTR(err);
 926 really_out:
 927         brelse(bitmap_bh);
 928         return ret;
 929
 930 fail_free_drop:
 931         DQUOT_FREE_INODE(inode);
 932
 933 fail_drop:
 934         DQUOT_DROP(inode);
 935         inode->i_flags |= S_NOQUOTA;
 936         inode->i_nlink = 0;
 937         iput(inode);
 938         brelse(bitmap_bh);
 939         return ERR_PTR(err);
 940 }
 941
 942 /* Verify that we are loading a valid orphan from disk */
 943 struct inode *ext4_orphan_get(struct super_block *sb, unsigned long ino)
 944 {
 945         unsigned long max_ino = le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count);
 946         ext4_group_t block_group;
 947         int bit;
 948         struct buffer_head *bitmap_bh;
 949         struct inode *inode = NULL;
 950         long err = -EIO;
 951
 952         /* Error cases - e2fsck has already cleaned up for us */
 953         if (ino > max_ino) {
 954                 ext4_warning(sb, __func__,
 955                              "bad orphan ino %lu!  e2fsck was run?", ino);
 956                 goto error;
 957         }
 958
 959         block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb);
 960         bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb);
 961         bitmap_bh = ext4_read_inode_bitmap(sb, block_group);
 962         if (!bitmap_bh) {
 963                 ext4_warning(sb, __func__,
 964                              "inode bitmap error for orphan %lu", ino);
 965                 goto error;
 966         }
 967
 968         /* Having the inode bit set should be a 100% indicator that this
 969          * is a valid orphan (no e2fsck run on fs).  Orphans also include
 970          * inodes that were being truncated, so we can't check i_nlink==0.
 971          */
 972         if (!ext4_test_bit(bit, bitmap_bh->b_data))
 973                 goto bad_orphan;
 974
 975         inode = ext4_iget(sb, ino);
 976         if (IS_ERR(inode))
 977                 goto iget_failed;
 978
 979         /*
 980          * If the orphans has i_nlinks > 0 then it should be able to be
 981          * truncated, otherwise it won't be removed from the orphan list
 982          * during processing and an infinite loop will result.
 983          */
 984         if (inode->i_nlink && !ext4_can_truncate(inode))
 985                 goto bad_orphan;
 986
 987         if (NEXT_ORPHAN(inode) > max_ino)
 988                 goto bad_orphan;
 989         brelse(bitmap_bh);
 990         return inode;
 991
 992 iget_failed:
 993         err = PTR_ERR(inode);
 994         inode = NULL;
 995 bad_orphan:
 996         ext4_warning(sb, __func__,
 997                      "bad orphan inode %lu!  e2fsck was run?", ino);
 998         printk(KERN_NOTICE "ext4_test_bit(bit=%d, block=%llu) = %d\n",
 999                bit, (unsigned long long)bitmap_bh->b_blocknr,
1000                ext4_test_bit(bit, bitmap_bh->b_data));
1001         printk(KERN_NOTICE "inode=%p\n", inode);
1002         if (inode) {
1003                 printk(KERN_NOTICE "is_bad_inode(inode)=%d\n",
1004                        is_bad_inode(inode));
1005                 printk(KERN_NOTICE "NEXT_ORPHAN(inode)=%u\n",
1006                        NEXT_ORPHAN(inode));
1007                 printk(KERN_NOTICE "max_ino=%lu\n", max_ino);
1008                 printk(KERN_NOTICE "i_nlink=%u\n", inode->i_nlink);
1009                 /* Avoid freeing blocks if we got a bad deleted inode */
1010                 if (inode->i_nlink == 0)
1011                         inode->i_blocks = 0;
1012                 iput(inode);
1013         }
1014         brelse(bitmap_bh);
1015 error:
1016         return ERR_PTR(err);
1017 }
1018
1019 unsigned long ext4_count_free_inodes (struct super_block * sb)
1020 {
1021         unsigned long desc_count;
1022         struct ext4_group_desc *gdp;
1023         ext4_group_t i;
1024 #ifdef EXT4FS_DEBUG
1025         struct ext4_super_block *es;
1026         unsigned long bitmap_count, x;
1027         struct buffer_head *bitmap_bh = NULL;
1028
1029         es = EXT4_SB(sb)->s_es;
1030         desc_count = 0;
1031         bitmap_count = 0;
1032         gdp = NULL;
1033         for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
1034                 gdp = ext4_get_group_desc (sb, i, NULL);
1035                 if (!gdp)
1036                         continue;
1037                 desc_count += le16_to_cpu(gdp->bg_free_inodes_count);
1038                 brelse(bitmap_bh);
1039                 bitmap_bh = ext4_read_inode_bitmap(sb, i);
1040                 if (!bitmap_bh)
1041                         continue;
1042
1043                 x = ext4_count_free(bitmap_bh, EXT4_INODES_PER_GROUP(sb) / 8);
1044                 printk(KERN_DEBUG "group %lu: stored = %d, counted = %lu\n",
1045                         i, le16_to_cpu(gdp->bg_free_inodes_count), x);
1046                 bitmap_count += x;
1047         }
1048         brelse(bitmap_bh);
1049         printk("ext4_count_free_inodes: stored = %u, computed = %lu, %lu\n",
1050                 le32_to_cpu(es->s_free_inodes_count), desc_count, bitmap_count);
1051         return desc_count;
1052 #else
1053         desc_count = 0;
1054         for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
1055                 gdp = ext4_get_group_desc (sb, i, NULL);
1056                 if (!gdp)
1057                         continue;
1058                 desc_count += le16_to_cpu(gdp->bg_free_inodes_count);
1059                 cond_resched();
1060         }
1061         return desc_count;
1062 #endif
1063 }
1064
1065 /* Called at mount-time, super-block is locked */
1066 unsigned long ext4_count_dirs (struct super_block * sb)
1067 {
1068         unsigned long count = 0;
1069         ext4_group_t i;
1070
1071         for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
1072                 struct ext4_group_desc *gdp = ext4_get_group_desc (sb, i, NULL);
1073                 if (!gdp)
1074                         continue;
1075                 count += le16_to_cpu(gdp->bg_used_dirs_count);
1076         }
1077         return count;
1078 }
1079