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 (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                 spin_unlock(sb_bgl_lock(EXT4_SB(sb), block_group));
 132                 unlock_buffer(bh);
 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 = %u, 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, count;
 192         ext4_group_t flex_group;
 193
 194         if (atomic_read(&inode->i_count) > 1) {
 195                 printk(KERN_ERR "ext4_free_inode: inode has count=%d\n",
 196                        atomic_read(&inode->i_count));
 197                 return;
 198         }
 199         if (inode->i_nlink) {
 200                 printk(KERN_ERR "ext4_free_inode: inode has nlink=%d\n",
 201                        inode->i_nlink);
 202                 return;
 203         }
 204         if (!sb) {
 205                 printk(KERN_ERR "ext4_free_inode: inode on "
 206                        "nonexistent device\n");
 207                 return;
 208         }
 209         sbi = EXT4_SB(sb);
 210
 211         ino = inode->i_ino;
 212         ext4_debug("freeing inode %lu\n", ino);
 213
 214         /*
 215          * Note: we must free any quota before locking the superblock,
 216          * as writing the quota to disk may need the lock as well.
 217          */
 218         DQUOT_INIT(inode);
 219         ext4_xattr_delete_inode(handle, inode);
 220         DQUOT_FREE_INODE(inode);
 221         DQUOT_DROP(inode);
 222
 223         is_directory = S_ISDIR(inode->i_mode);
 224
 225         /* Do this BEFORE marking the inode not in use or returning an error */
 226         clear_inode(inode);
 227
 228         es = EXT4_SB(sb)->s_es;
 229         if (ino < EXT4_FIRST_INO(sb) || ino > le32_to_cpu(es->s_inodes_count)) {
 230                 ext4_error(sb, "ext4_free_inode",
 231                            "reserved or nonexistent inode %lu", ino);
 232                 goto error_return;
 233         }
 234         block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb);
 235         bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb);
 236         bitmap_bh = ext4_read_inode_bitmap(sb, block_group);
 237         if (!bitmap_bh)
 238                 goto error_return;
 239
 240         BUFFER_TRACE(bitmap_bh, "get_write_access");
 241         fatal = ext4_journal_get_write_access(handle, bitmap_bh);
 242         if (fatal)
 243                 goto error_return;
 244
 245         /* Ok, now we can actually update the inode bitmaps.. */
 246         if (!ext4_clear_bit_atomic(sb_bgl_lock(sbi, block_group),
 247                                         bit, bitmap_bh->b_data))
 248                 ext4_error(sb, "ext4_free_inode",
 249                            "bit already cleared for inode %lu", ino);
 250         else {
 251                 gdp = ext4_get_group_desc(sb, block_group, &bh2);
 252
 253                 BUFFER_TRACE(bh2, "get_write_access");
 254                 fatal = ext4_journal_get_write_access(handle, bh2);
 255                 if (fatal) goto error_return;
 256
 257                 if (gdp) {
 258                         spin_lock(sb_bgl_lock(sbi, block_group));
 259                         count = ext4_free_inodes_count(sb, gdp) + 1;
 260                         ext4_free_inodes_set(sb, gdp, count);
 261                         if (is_directory) {
 262                                 count = ext4_used_dirs_count(sb, gdp) - 1;
 263                                 ext4_used_dirs_set(sb, gdp, count);
 264                         }
 265                         gdp->bg_checksum = ext4_group_desc_csum(sbi,
 266                                                         block_group, gdp);
 267                         spin_unlock(sb_bgl_lock(sbi, block_group));
 268                         percpu_counter_inc(&sbi->s_freeinodes_counter);
 269                         if (is_directory)
 270                                 percpu_counter_dec(&sbi->s_dirs_counter);
 271
 272                         if (sbi->s_log_groups_per_flex) {
 273                                 flex_group = ext4_flex_group(sbi, block_group);
 274                                 spin_lock(sb_bgl_lock(sbi, flex_group));
 275                                 sbi->s_flex_groups[flex_group].free_inodes++;
 276                                 spin_unlock(sb_bgl_lock(sbi, flex_group));
 277                         }
 278                 }
 279                 BUFFER_TRACE(bh2, "call ext4_handle_dirty_metadata");
 280                 err = ext4_handle_dirty_metadata(handle, NULL, bh2);
 281                 if (!fatal) fatal = err;
 282         }
 283         BUFFER_TRACE(bitmap_bh, "call ext4_handle_dirty_metadata");
 284         err = ext4_handle_dirty_metadata(handle, NULL, bitmap_bh);
 285         if (!fatal)
 286                 fatal = err;
 287         sb->s_dirt = 1;
 288 error_return:
 289         brelse(bitmap_bh);
 290         ext4_std_error(sb, fatal);
 291 }
 292
 293 /*
 294  * There are two policies for allocating an inode.  If the new inode is
 295  * a directory, then a forward search is made for a block group with both
 296  * free space and a low directory-to-inode ratio; if that fails, then of
 297  * the groups with above-average free space, that group with the fewest
 298  * directories already is chosen.
 299  *
 300  * For other inodes, search forward from the parent directory\'s block
 301  * group to find a free inode.
 302  */
 303 static int find_group_dir(struct super_block *sb, struct inode *parent,
 304                                 ext4_group_t *best_group)
 305 {
 306         ext4_group_t ngroups = EXT4_SB(sb)->s_groups_count;
 307         unsigned int freei, avefreei;
 308         struct ext4_group_desc *desc, *best_desc = NULL;
 309         ext4_group_t group;
 310         int ret = -1;
 311
 312         freei = percpu_counter_read_positive(&EXT4_SB(sb)->s_freeinodes_counter);
 313         avefreei = freei / ngroups;
 314
 315         for (group = 0; group < ngroups; group++) {
 316                 desc = ext4_get_group_desc(sb, group, NULL);
 317                 if (!desc || !ext4_free_inodes_count(sb, desc))
 318                         continue;
 319                 if (ext4_free_inodes_count(sb, desc) < avefreei)
 320                         continue;
 321                 if (!best_desc ||
 322                     (ext4_free_blks_count(sb, desc) >
 323                      ext4_free_blks_count(sb, best_desc))) {
 324                         *best_group = group;
 325                         best_desc = desc;
 326                         ret = 0;
 327                 }
 328         }
 329         return ret;
 330 }
 331
 332 #define free_block_ratio 10
 333
 334 static int find_group_flex(struct super_block *sb, struct inode *parent,
 335                            ext4_group_t *best_group)
 336 {
 337         struct ext4_sb_info *sbi = EXT4_SB(sb);
 338         struct ext4_group_desc *desc;
 339         struct buffer_head *bh;
 340         struct flex_groups *flex_group = sbi->s_flex_groups;
 341         ext4_group_t parent_group = EXT4_I(parent)->i_block_group;
 342         ext4_group_t parent_fbg_group = ext4_flex_group(sbi, parent_group);
 343         ext4_group_t ngroups = sbi->s_groups_count;
 344         int flex_size = ext4_flex_bg_size(sbi);
 345         ext4_group_t best_flex = parent_fbg_group;
 346         int blocks_per_flex = sbi->s_blocks_per_group * flex_size;
 347         int flexbg_free_blocks;
 348         int flex_freeb_ratio;
 349         ext4_group_t n_fbg_groups;
 350         ext4_group_t i;
 351
 352         n_fbg_groups = (sbi->s_groups_count + flex_size - 1) >>
 353                 sbi->s_log_groups_per_flex;
 354
 355 find_close_to_parent:
 356         flexbg_free_blocks = flex_group[best_flex].free_blocks;
 357         flex_freeb_ratio = flexbg_free_blocks * 100 / blocks_per_flex;
 358         if (flex_group[best_flex].free_inodes &&
 359             flex_freeb_ratio > free_block_ratio)
 360                 goto found_flexbg;
 361
 362         if (best_flex && best_flex == parent_fbg_group) {
 363                 best_flex--;
 364                 goto find_close_to_parent;
 365         }
 366
 367         for (i = 0; i < n_fbg_groups; i++) {
 368                 if (i == parent_fbg_group || i == parent_fbg_group - 1)
 369                         continue;
 370
 371                 flexbg_free_blocks = flex_group[i].free_blocks;
 372                 flex_freeb_ratio = flexbg_free_blocks * 100 / blocks_per_flex;
 373
 374                 if (flex_freeb_ratio > free_block_ratio &&
 375                     flex_group[i].free_inodes) {
 376                         best_flex = i;
 377                         goto found_flexbg;
 378                 }
 379
 380                 if (flex_group[best_flex].free_inodes == 0 ||
 381                     (flex_group[i].free_blocks >
 382                      flex_group[best_flex].free_blocks &&
 383                      flex_group[i].free_inodes))
 384                         best_flex = i;
 385         }
 386
 387         if (!flex_group[best_flex].free_inodes ||
 388             !flex_group[best_flex].free_blocks)
 389                 return -1;
 390
 391 found_flexbg:
 392         for (i = best_flex * flex_size; i < ngroups &&
 393                      i < (best_flex + 1) * flex_size; i++) {
 394                 desc = ext4_get_group_desc(sb, i, &bh);
 395                 if (ext4_free_inodes_count(sb, desc)) {
 396                         *best_group = i;
 397                         goto out;
 398                 }
 399         }
 400
 401         return -1;
 402 out:
 403         return 0;
 404 }
 405
 406 /*
 407  * Orlov's allocator for directories.
 408  *
 409  * We always try to spread first-level directories.
 410  *
 411  * If there are blockgroups with both free inodes and free blocks counts
 412  * not worse than average we return one with smallest directory count.
 413  * Otherwise we simply return a random group.
 414  *
 415  * For the rest rules look so:
 416  *
 417  * It's OK to put directory into a group unless
 418  * it has too many directories already (max_dirs) or
 419  * it has too few free inodes left (min_inodes) or
 420  * it has too few free blocks left (min_blocks) or
 421  * it's already running too large debt (max_debt).
 422  * Parent's group is preferred, if it doesn't satisfy these
 423  * conditions we search cyclically through the rest. If none
 424  * of the groups look good we just look for a group with more
 425  * free inodes than average (starting at parent's group).
 426  *
 427  * Debt is incremented each time we allocate a directory and decremented
 428  * when we allocate an inode, within 0--255.
 429  */
 430
 431 #define INODE_COST 64
 432 #define BLOCK_COST 256
 433
 434 static int find_group_orlov(struct super_block *sb, struct inode *parent,
 435                                 ext4_group_t *group)
 436 {
 437         ext4_group_t parent_group = EXT4_I(parent)->i_block_group;
 438         struct ext4_sb_info *sbi = EXT4_SB(sb);
 439         struct ext4_super_block *es = sbi->s_es;
 440         ext4_group_t ngroups = sbi->s_groups_count;
 441         int inodes_per_group = EXT4_INODES_PER_GROUP(sb);
 442         unsigned int freei, avefreei;
 443         ext4_fsblk_t freeb, avefreeb;
 444         ext4_fsblk_t blocks_per_dir;
 445         unsigned int ndirs;
 446         int max_debt, max_dirs, min_inodes;
 447         ext4_grpblk_t min_blocks;
 448         ext4_group_t i;
 449         struct ext4_group_desc *desc;
 450
 451         freei = percpu_counter_read_positive(&sbi->s_freeinodes_counter);
 452         avefreei = freei / ngroups;
 453         freeb = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
 454         avefreeb = freeb;
 455         do_div(avefreeb, ngroups);
 456         ndirs = percpu_counter_read_positive(&sbi->s_dirs_counter);
 457
 458         if ((parent == sb->s_root->d_inode) ||
 459             (EXT4_I(parent)->i_flags & EXT4_TOPDIR_FL)) {
 460                 int best_ndir = inodes_per_group;
 461                 ext4_group_t grp;
 462                 int ret = -1;
 463
 464                 get_random_bytes(&grp, sizeof(grp));
 465                 parent_group = (unsigned)grp % ngroups;
 466                 for (i = 0; i < ngroups; i++) {
 467                         grp = (parent_group + i) % ngroups;
 468                         desc = ext4_get_group_desc(sb, grp, NULL);
 469                         if (!desc || !ext4_free_inodes_count(sb, desc))
 470                                 continue;
 471                         if (ext4_used_dirs_count(sb, desc) >= best_ndir)
 472                                 continue;
 473                         if (ext4_free_inodes_count(sb, desc) < avefreei)
 474                                 continue;
 475                         if (ext4_free_blks_count(sb, desc) < avefreeb)
 476                                 continue;
 477                         *group = grp;
 478                         ret = 0;
 479                         best_ndir = ext4_used_dirs_count(sb, desc);
 480                 }
 481                 if (ret == 0)
 482                         return ret;
 483                 goto fallback;
 484         }
 485
 486         blocks_per_dir = ext4_blocks_count(es) - freeb;
 487         do_div(blocks_per_dir, ndirs);
 488
 489         max_dirs = ndirs / ngroups + inodes_per_group / 16;
 490         min_inodes = avefreei - inodes_per_group / 4;
 491         min_blocks = avefreeb - EXT4_BLOCKS_PER_GROUP(sb) / 4;
 492
 493         max_debt = EXT4_BLOCKS_PER_GROUP(sb);
 494         max_debt /= max_t(int, blocks_per_dir, BLOCK_COST);
 495         if (max_debt * INODE_COST > inodes_per_group)
 496                 max_debt = inodes_per_group / INODE_COST;
 497         if (max_debt > 255)
 498                 max_debt = 255;
 499         if (max_debt == 0)
 500                 max_debt = 1;
 501
 502         for (i = 0; i < ngroups; i++) {
 503                 *group = (parent_group + i) % ngroups;
 504                 desc = ext4_get_group_desc(sb, *group, NULL);
 505                 if (!desc || !ext4_free_inodes_count(sb, desc))
 506                         continue;
 507                 if (ext4_used_dirs_count(sb, desc) >= max_dirs)
 508                         continue;
 509                 if (ext4_free_inodes_count(sb, desc) < min_inodes)
 510                         continue;
 511                 if (ext4_free_blks_count(sb, desc) < min_blocks)
 512                         continue;
 513                 return 0;
 514         }
 515
 516 fallback:
 517         for (i = 0; i < ngroups; i++) {
 518                 *group = (parent_group + i) % ngroups;
 519                 desc = ext4_get_group_desc(sb, *group, NULL);
 520                 if (desc && ext4_free_inodes_count(sb, desc) &&
 521                         ext4_free_inodes_count(sb, desc) >= avefreei)
 522                         return 0;
 523         }
 524
 525         if (avefreei) {
 526                 /*
 527                  * The free-inodes counter is approximate, and for really small
 528                  * filesystems the above test can fail to find any blockgroups
 529                  */
 530                 avefreei = 0;
 531                 goto fallback;
 532         }
 533
 534         return -1;
 535 }
 536
 537 static int find_group_other(struct super_block *sb, struct inode *parent,
 538                                 ext4_group_t *group)
 539 {
 540         ext4_group_t parent_group = EXT4_I(parent)->i_block_group;
 541         ext4_group_t ngroups = EXT4_SB(sb)->s_groups_count;
 542         struct ext4_group_desc *desc;
 543         ext4_group_t i;
 544
 545         /*
 546          * Try to place the inode in its parent directory
 547          */
 548         *group = parent_group;
 549         desc = ext4_get_group_desc(sb, *group, NULL);
 550         if (desc && ext4_free_inodes_count(sb, desc) &&
 551                         ext4_free_blks_count(sb, desc))
 552                 return 0;
 553
 554         /*
 555          * We're going to place this inode in a different blockgroup from its
 556          * parent.  We want to cause files in a common directory to all land in
 557          * the same blockgroup.  But we want files which are in a different
 558          * directory which shares a blockgroup with our parent to land in a
 559          * different blockgroup.
 560          *
 561          * So add our directory's i_ino into the starting point for the hash.
 562          */
 563         *group = (*group + parent->i_ino) % ngroups;
 564
 565         /*
 566          * Use a quadratic hash to find a group with a free inode and some free
 567          * blocks.
 568          */
 569         for (i = 1; i < ngroups; i <<= 1) {
 570                 *group += i;
 571                 if (*group >= ngroups)
 572                         *group -= ngroups;
 573                 desc = ext4_get_group_desc(sb, *group, NULL);
 574                 if (desc && ext4_free_inodes_count(sb, desc) &&
 575                                 ext4_free_blks_count(sb, desc))
 576                         return 0;
 577         }
 578
 579         /*
 580          * That failed: try linear search for a free inode, even if that group
 581          * has no free blocks.
 582          */
 583         *group = parent_group;
 584         for (i = 0; i < ngroups; i++) {
 585                 if (++*group >= ngroups)
 586                         *group = 0;
 587                 desc = ext4_get_group_desc(sb, *group, NULL);
 588                 if (desc && ext4_free_inodes_count(sb, desc))
 589                         return 0;
 590         }
 591
 592         return -1;
 593 }
 594
 595 /*
 596  * claim the inode from the inode bitmap. If the group
 597  * is uninit we need to take the groups's sb_bgl_lock
 598  * and clear the uninit flag. The inode bitmap update
 599  * and group desc uninit flag clear should be done
 600  * after holding sb_bgl_lock so that ext4_read_inode_bitmap
 601  * doesn't race with the ext4_claim_inode
 602  */
 603 static int ext4_claim_inode(struct super_block *sb,
 604                         struct buffer_head *inode_bitmap_bh,
 605                         unsigned long ino, ext4_group_t group, int mode)
 606 {
 607         int free = 0, retval = 0, count;
 608         struct ext4_sb_info *sbi = EXT4_SB(sb);
 609         struct ext4_group_desc *gdp = ext4_get_group_desc(sb, group, NULL);
 610
 611         spin_lock(sb_bgl_lock(sbi, group));
 612         if (ext4_set_bit(ino, inode_bitmap_bh->b_data)) {
 613                 /* not a free inode */
 614                 retval = 1;
 615                 goto err_ret;
 616         }
 617         ino++;
 618         if ((group == 0 && ino < EXT4_FIRST_INO(sb)) ||
 619                         ino > EXT4_INODES_PER_GROUP(sb)) {
 620                 spin_unlock(sb_bgl_lock(sbi, group));
 621                 ext4_error(sb, __func__,
 622                            "reserved inode or inode > inodes count - "
 623                            "block_group = %u, inode=%lu", group,
 624                            ino + group * EXT4_INODES_PER_GROUP(sb));
 625                 return 1;
 626         }
 627         /* If we didn't allocate from within the initialized part of the inode
 628          * table then we need to initialize up to this inode. */
 629         if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM)) {
 630
 631                 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) {
 632                         gdp->bg_flags &= cpu_to_le16(~EXT4_BG_INODE_UNINIT);
 633                         /* When marking the block group with
 634                          * ~EXT4_BG_INODE_UNINIT we don't want to depend
 635                          * on the value of bg_itable_unused even though
 636                          * mke2fs could have initialized the same for us.
 637                          * Instead we calculated the value below
 638                          */
 639
 640                         free = 0;
 641                 } else {
 642                         free = EXT4_INODES_PER_GROUP(sb) -
 643                                 ext4_itable_unused_count(sb, gdp);
 644                 }
 645
 646                 /*
 647                  * Check the relative inode number against the last used
 648                  * relative inode number in this group. if it is greater
 649                  * we need to  update the bg_itable_unused count
 650                  *
 651                  */
 652                 if (ino > free)
 653                         ext4_itable_unused_set(sb, gdp,
 654                                         (EXT4_INODES_PER_GROUP(sb) - ino));
 655         }
 656         count = ext4_free_inodes_count(sb, gdp) - 1;
 657         ext4_free_inodes_set(sb, gdp, count);
 658         if (S_ISDIR(mode)) {
 659                 count = ext4_used_dirs_count(sb, gdp) + 1;
 660                 ext4_used_dirs_set(sb, gdp, count);
 661         }
 662         gdp->bg_checksum = ext4_group_desc_csum(sbi, group, gdp);
 663 err_ret:
 664         spin_unlock(sb_bgl_lock(sbi, group));
 665         return retval;
 666 }
 667
 668 /*
 669  * There are two policies for allocating an inode.  If the new inode is
 670  * a directory, then a forward search is made for a block group with both
 671  * free space and a low directory-to-inode ratio; if that fails, then of
 672  * the groups with above-average free space, that group with the fewest
 673  * directories already is chosen.
 674  *
 675  * For other inodes, search forward from the parent directory's block
 676  * group to find a free inode.
 677  */
 678 struct inode *ext4_new_inode(handle_t *handle, struct inode *dir, int mode)
 679 {
 680         struct super_block *sb;
 681         struct buffer_head *inode_bitmap_bh = NULL;
 682         struct buffer_head *group_desc_bh;
 683         ext4_group_t group = 0;
 684         unsigned long ino = 0;
 685         struct inode *inode;
 686         struct ext4_group_desc *gdp = NULL;
 687         struct ext4_super_block *es;
 688         struct ext4_inode_info *ei;
 689         struct ext4_sb_info *sbi;
 690         int ret2, err = 0;
 691         struct inode *ret;
 692         ext4_group_t i;
 693         int free = 0;
 694         ext4_group_t flex_group;
 695
 696         /* Cannot create files in a deleted directory */
 697         if (!dir || !dir->i_nlink)
 698                 return ERR_PTR(-EPERM);
 699
 700         sb = dir->i_sb;
 701         inode = new_inode(sb);
 702         if (!inode)
 703                 return ERR_PTR(-ENOMEM);
 704         ei = EXT4_I(inode);
 705
 706         sbi = EXT4_SB(sb);
 707         es = sbi->s_es;
 708
 709         if (sbi->s_log_groups_per_flex) {
 710                 ret2 = find_group_flex(sb, dir, &group);
 711                 goto got_group;
 712         }
 713
 714         if (S_ISDIR(mode)) {
 715                 if (test_opt(sb, OLDALLOC))
 716                         ret2 = find_group_dir(sb, dir, &group);
 717                 else
 718                         ret2 = find_group_orlov(sb, dir, &group);
 719         } else
 720                 ret2 = find_group_other(sb, dir, &group);
 721
 722 got_group:
 723         err = -ENOSPC;
 724         if (ret2 == -1)
 725                 goto out;
 726
 727         for (i = 0; i < sbi->s_groups_count; i++) {
 728                 err = -EIO;
 729
 730                 gdp = ext4_get_group_desc(sb, group, &group_desc_bh);
 731                 if (!gdp)
 732                         goto fail;
 733
 734                 brelse(inode_bitmap_bh);
 735                 inode_bitmap_bh = ext4_read_inode_bitmap(sb, group);
 736                 if (!inode_bitmap_bh)
 737                         goto fail;
 738
 739                 ino = 0;
 740
 741 repeat_in_this_group:
 742                 ino = ext4_find_next_zero_bit((unsigned long *)
 743                                               inode_bitmap_bh->b_data,
 744                                               EXT4_INODES_PER_GROUP(sb), ino);
 745
 746                 if (ino < EXT4_INODES_PER_GROUP(sb)) {
 747
 748                         BUFFER_TRACE(inode_bitmap_bh, "get_write_access");
 749                         err = ext4_journal_get_write_access(handle,
 750                                                             inode_bitmap_bh);
 751                         if (err)
 752                                 goto fail;
 753
 754                         BUFFER_TRACE(group_desc_bh, "get_write_access");
 755                         err = ext4_journal_get_write_access(handle,
 756                                                                 group_desc_bh);
 757                         if (err)
 758                                 goto fail;
 759                         if (!ext4_claim_inode(sb, inode_bitmap_bh,
 760                                                 ino, group, mode)) {
 761                                 /* we won it */
 762                                 BUFFER_TRACE(inode_bitmap_bh,
 763                                         "call ext4_handle_dirty_metadata");
 764                                 err = ext4_handle_dirty_metadata(handle,
 765                                                                  inode,
 766                                                         inode_bitmap_bh);
 767                                 if (err)
 768                                         goto fail;
 769                                 /* zero bit is inode number 1*/
 770                                 ino++;
 771                                 goto got;
 772                         }
 773                         /* we lost it */
 774                         ext4_handle_release_buffer(handle, inode_bitmap_bh);
 775                         ext4_handle_release_buffer(handle, group_desc_bh);
 776
 777                         if (++ino < EXT4_INODES_PER_GROUP(sb))
 778                                 goto repeat_in_this_group;
 779                 }
 780
 781                 /*
 782                  * This case is possible in concurrent environment.  It is very
 783                  * rare.  We cannot repeat the find_group_xxx() call because
 784                  * that will simply return the same blockgroup, because the
 785                  * group descriptor metadata has not yet been updated.
 786                  * So we just go onto the next blockgroup.
 787                  */
 788                 if (++group == sbi->s_groups_count)
 789                         group = 0;
 790         }
 791         err = -ENOSPC;
 792         goto out;
 793
 794 got:
 795         /* We may have to initialize the block bitmap if it isn't already */
 796         if (EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_GDT_CSUM) &&
 797             gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
 798                 struct buffer_head *block_bitmap_bh;
 799
 800                 block_bitmap_bh = ext4_read_block_bitmap(sb, group);
 801                 BUFFER_TRACE(block_bitmap_bh, "get block bitmap access");
 802                 err = ext4_journal_get_write_access(handle, block_bitmap_bh);
 803                 if (err) {
 804                         brelse(block_bitmap_bh);
 805                         goto fail;
 806                 }
 807
 808                 free = 0;
 809                 spin_lock(sb_bgl_lock(sbi, group));
 810                 /* recheck and clear flag under lock if we still need to */
 811                 if (gdp->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
 812                         free = ext4_free_blocks_after_init(sb, group, gdp);
 813                         gdp->bg_flags &= cpu_to_le16(~EXT4_BG_BLOCK_UNINIT);
 814                         ext4_free_blks_set(sb, gdp, free);
 815                         gdp->bg_checksum = ext4_group_desc_csum(sbi, group,
 816                                                                 gdp);
 817                 }
 818                 spin_unlock(sb_bgl_lock(sbi, group));
 819
 820                 /* Don't need to dirty bitmap block if we didn't change it */
 821                 if (free) {
 822                         BUFFER_TRACE(block_bitmap_bh, "dirty block bitmap");
 823                         err = ext4_handle_dirty_metadata(handle,
 824                                                         NULL, block_bitmap_bh);
 825                 }
 826
 827                 brelse(block_bitmap_bh);
 828                 if (err)
 829                         goto fail;
 830         }
 831         BUFFER_TRACE(group_desc_bh, "call ext4_handle_dirty_metadata");
 832         err = ext4_handle_dirty_metadata(handle, NULL, group_desc_bh);
 833         if (err)
 834                 goto fail;
 835
 836         percpu_counter_dec(&sbi->s_freeinodes_counter);
 837         if (S_ISDIR(mode))
 838                 percpu_counter_inc(&sbi->s_dirs_counter);
 839         sb->s_dirt = 1;
 840
 841         if (sbi->s_log_groups_per_flex) {
 842                 flex_group = ext4_flex_group(sbi, group);
 843                 spin_lock(sb_bgl_lock(sbi, flex_group));
 844                 sbi->s_flex_groups[flex_group].free_inodes--;
 845                 spin_unlock(sb_bgl_lock(sbi, flex_group));
 846         }
 847
 848         inode->i_uid = current_fsuid();
 849         if (test_opt(sb, GRPID))
 850                 inode->i_gid = dir->i_gid;
 851         else if (dir->i_mode & S_ISGID) {
 852                 inode->i_gid = dir->i_gid;
 853                 if (S_ISDIR(mode))
 854                         mode |= S_ISGID;
 855         } else
 856                 inode->i_gid = current_fsgid();
 857         inode->i_mode = mode;
 858
 859         inode->i_ino = ino + group * EXT4_INODES_PER_GROUP(sb);
 860         /* This is the optimal IO size (for stat), not the fs block size */
 861         inode->i_blocks = 0;
 862         inode->i_mtime = inode->i_atime = inode->i_ctime = ei->i_crtime =
 863                                                        ext4_current_time(inode);
 864
 865         memset(ei->i_data, 0, sizeof(ei->i_data));
 866         ei->i_dir_start_lookup = 0;
 867         ei->i_disksize = 0;
 868
 869         /*
 870          * Don't inherit extent flag from directory. We set extent flag on
 871          * newly created directory and file only if -o extent mount option is
 872          * specified
 873          */
 874         ei->i_flags = EXT4_I(dir)->i_flags & ~(EXT4_INDEX_FL|EXT4_EXTENTS_FL);
 875         if (S_ISLNK(mode))
 876                 ei->i_flags &= ~(EXT4_IMMUTABLE_FL|EXT4_APPEND_FL);
 877         /* dirsync only applies to directories */
 878         if (!S_ISDIR(mode))
 879                 ei->i_flags &= ~EXT4_DIRSYNC_FL;
 880         ei->i_file_acl = 0;
 881         ei->i_dtime = 0;
 882         ei->i_block_group = group;
 883
 884         ext4_set_inode_flags(inode);
 885         if (IS_DIRSYNC(inode))
 886                 ext4_handle_sync(handle);
 887         if (insert_inode_locked(inode) < 0) {
 888                 err = -EINVAL;
 889                 goto fail_drop;
 890         }
 891         spin_lock(&sbi->s_next_gen_lock);
 892         inode->i_generation = sbi->s_next_generation++;
 893         spin_unlock(&sbi->s_next_gen_lock);
 894
 895         ei->i_state = EXT4_STATE_NEW;
 896
 897         ei->i_extra_isize = EXT4_SB(sb)->s_want_extra_isize;
 898
 899         ret = inode;
 900         if (DQUOT_ALLOC_INODE(inode)) {
 901                 err = -EDQUOT;
 902                 goto fail_drop;
 903         }
 904
 905         err = ext4_init_acl(handle, inode, dir);
 906         if (err)
 907                 goto fail_free_drop;
 908
 909         err = ext4_init_security(handle, inode, dir);
 910         if (err)
 911                 goto fail_free_drop;
 912
 913         if (test_opt(sb, EXTENTS)) {
 914                 /* set extent flag only for directory, file and normal symlink*/
 915                 if (S_ISDIR(mode) || S_ISREG(mode) || S_ISLNK(mode)) {
 916                         EXT4_I(inode)->i_flags |= EXT4_EXTENTS_FL;
 917                         ext4_ext_tree_init(handle, inode);
 918                 }
 919         }
 920
 921         err = ext4_mark_inode_dirty(handle, inode);
 922         if (err) {
 923                 ext4_std_error(sb, err);
 924                 goto fail_free_drop;
 925         }
 926
 927         ext4_debug("allocating inode %lu\n", inode->i_ino);
 928         goto really_out;
 929 fail:
 930         ext4_std_error(sb, err);
 931 out:
 932         iput(inode);
 933         ret = ERR_PTR(err);
 934 really_out:
 935         brelse(inode_bitmap_bh);
 936         return ret;
 937
 938 fail_free_drop:
 939         DQUOT_FREE_INODE(inode);
 940
 941 fail_drop:
 942         DQUOT_DROP(inode);
 943         inode->i_flags |= S_NOQUOTA;
 944         inode->i_nlink = 0;
 945         unlock_new_inode(inode);
 946         iput(inode);
 947         brelse(inode_bitmap_bh);
 948         return ERR_PTR(err);
 949 }
 950
 951 /* Verify that we are loading a valid orphan from disk */
 952 struct inode *ext4_orphan_get(struct super_block *sb, unsigned long ino)
 953 {
 954         unsigned long max_ino = le32_to_cpu(EXT4_SB(sb)->s_es->s_inodes_count);
 955         ext4_group_t block_group;
 956         int bit;
 957         struct buffer_head *bitmap_bh;
 958         struct inode *inode = NULL;
 959         long err = -EIO;
 960
 961         /* Error cases - e2fsck has already cleaned up for us */
 962         if (ino > max_ino) {
 963                 ext4_warning(sb, __func__,
 964                              "bad orphan ino %lu!  e2fsck was run?", ino);
 965                 goto error;
 966         }
 967
 968         block_group = (ino - 1) / EXT4_INODES_PER_GROUP(sb);
 969         bit = (ino - 1) % EXT4_INODES_PER_GROUP(sb);
 970         bitmap_bh = ext4_read_inode_bitmap(sb, block_group);
 971         if (!bitmap_bh) {
 972                 ext4_warning(sb, __func__,
 973                              "inode bitmap error for orphan %lu", ino);
 974                 goto error;
 975         }
 976
 977         /* Having the inode bit set should be a 100% indicator that this
 978          * is a valid orphan (no e2fsck run on fs).  Orphans also include
 979          * inodes that were being truncated, so we can't check i_nlink==0.
 980          */
 981         if (!ext4_test_bit(bit, bitmap_bh->b_data))
 982                 goto bad_orphan;
 983
 984         inode = ext4_iget(sb, ino);
 985         if (IS_ERR(inode))
 986                 goto iget_failed;
 987
 988         /*
 989          * If the orphans has i_nlinks > 0 then it should be able to be
 990          * truncated, otherwise it won't be removed from the orphan list
 991          * during processing and an infinite loop will result.
 992          */
 993         if (inode->i_nlink && !ext4_can_truncate(inode))
 994                 goto bad_orphan;
 995
 996         if (NEXT_ORPHAN(inode) > max_ino)
 997                 goto bad_orphan;
 998         brelse(bitmap_bh);
 999         return inode;
1000
1001 iget_failed:
1002         err = PTR_ERR(inode);
1003         inode = NULL;
1004 bad_orphan:
1005         ext4_warning(sb, __func__,
1006                      "bad orphan inode %lu!  e2fsck was run?", ino);
1007         printk(KERN_NOTICE "ext4_test_bit(bit=%d, block=%llu) = %d\n",
1008                bit, (unsigned long long)bitmap_bh->b_blocknr,
1009                ext4_test_bit(bit, bitmap_bh->b_data));
1010         printk(KERN_NOTICE "inode=%p\n", inode);
1011         if (inode) {
1012                 printk(KERN_NOTICE "is_bad_inode(inode)=%d\n",
1013                        is_bad_inode(inode));
1014                 printk(KERN_NOTICE "NEXT_ORPHAN(inode)=%u\n",
1015                        NEXT_ORPHAN(inode));
1016                 printk(KERN_NOTICE "max_ino=%lu\n", max_ino);
1017                 printk(KERN_NOTICE "i_nlink=%u\n", inode->i_nlink);
1018                 /* Avoid freeing blocks if we got a bad deleted inode */
1019                 if (inode->i_nlink == 0)
1020                         inode->i_blocks = 0;
1021                 iput(inode);
1022         }
1023         brelse(bitmap_bh);
1024 error:
1025         return ERR_PTR(err);
1026 }
1027
1028 unsigned long ext4_count_free_inodes(struct super_block *sb)
1029 {
1030         unsigned long desc_count;
1031         struct ext4_group_desc *gdp;
1032         ext4_group_t i;
1033 #ifdef EXT4FS_DEBUG
1034         struct ext4_super_block *es;
1035         unsigned long bitmap_count, x;
1036         struct buffer_head *bitmap_bh = NULL;
1037
1038         es = EXT4_SB(sb)->s_es;
1039         desc_count = 0;
1040         bitmap_count = 0;
1041         gdp = NULL;
1042         for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
1043                 gdp = ext4_get_group_desc(sb, i, NULL);
1044                 if (!gdp)
1045                         continue;
1046                 desc_count += ext4_free_inodes_count(sb, gdp);
1047                 brelse(bitmap_bh);
1048                 bitmap_bh = ext4_read_inode_bitmap(sb, i);
1049                 if (!bitmap_bh)
1050                         continue;
1051
1052                 x = ext4_count_free(bitmap_bh, EXT4_INODES_PER_GROUP(sb) / 8);
1053                 printk(KERN_DEBUG "group %lu: stored = %d, counted = %lu\n",
1054                         i, ext4_free_inodes_count(sb, gdp), x);
1055                 bitmap_count += x;
1056         }
1057         brelse(bitmap_bh);
1058         printk(KERN_DEBUG "ext4_count_free_inodes: "
1059                "stored = %u, computed = %lu, %lu\n",
1060                le32_to_cpu(es->s_free_inodes_count), desc_count, bitmap_count);
1061         return desc_count;
1062 #else
1063         desc_count = 0;
1064         for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
1065                 gdp = ext4_get_group_desc(sb, i, NULL);
1066                 if (!gdp)
1067                         continue;
1068                 desc_count += ext4_free_inodes_count(sb, gdp);
1069                 cond_resched();
1070         }
1071         return desc_count;
1072 #endif
1073 }
1074
1075 /* Called at mount-time, super-block is locked */
1076 unsigned long ext4_count_dirs(struct super_block * sb)
1077 {
1078         unsigned long count = 0;
1079         ext4_group_t i;
1080
1081         for (i = 0; i < EXT4_SB(sb)->s_groups_count; i++) {
1082                 struct ext4_group_desc *gdp = ext4_get_group_desc(sb, i, NULL);
1083                 if (!gdp)
1084                         continue;
1085                 count += ext4_used_dirs_count(sb, gdp);
1086         }
1087         return count;
1088 }