fs/ext4/resize.c

   1 /*
   2  *  linux/fs/ext4/resize.c
   3  *
   4  * Support for resizing an ext4 filesystem while it is mounted.
   5  *
   6  * Copyright (C) 2001, 2002 Andreas Dilger <adilger@clusterfs.com>
   7  *
   8  * This could probably be made into a module, because it is not often in use.
   9  */
  10
  11
  12 #define EXT4FS_DEBUG
  13
  14 #include <linux/ext4_jbd2.h>
  15
  16 #include <linux/errno.h>
  17 #include <linux/slab.h>
  18
  19
  20 #define outside(b, first, last) ((b) < (first) || (b) >= (last))
  21 #define inside(b, first, last)  ((b) >= (first) && (b) < (last))
  22
  23 static int verify_group_input(struct super_block *sb,
  24                               struct ext4_new_group_data *input)
  25 {
  26         struct ext4_sb_info *sbi = EXT4_SB(sb);
  27         struct ext4_super_block *es = sbi->s_es;
  28         ext4_fsblk_t start = ext4_blocks_count(es);
  29         ext4_fsblk_t end = start + input->blocks_count;
  30         unsigned group = input->group;
  31         ext4_fsblk_t itend = input->inode_table + sbi->s_itb_per_group;
  32         unsigned overhead = ext4_bg_has_super(sb, group) ?
  33                 (1 + ext4_bg_num_gdb(sb, group) +
  34                  le16_to_cpu(es->s_reserved_gdt_blocks)) : 0;
  35         ext4_fsblk_t metaend = start + overhead;
  36         struct buffer_head *bh = NULL;
  37         ext4_grpblk_t free_blocks_count, offset;
  38         int err = -EINVAL;
  39
  40         input->free_blocks_count = free_blocks_count =
  41                 input->blocks_count - 2 - overhead - sbi->s_itb_per_group;
  42
  43         if (test_opt(sb, DEBUG))
  44                 printk(KERN_DEBUG "EXT4-fs: adding %s group %u: %u blocks "
  45                        "(%d free, %u reserved)\n",
  46                        ext4_bg_has_super(sb, input->group) ? "normal" :
  47                        "no-super", input->group, input->blocks_count,
  48                        free_blocks_count, input->reserved_blocks);
  49
  50         ext4_get_group_no_and_offset(sb, start, NULL, &offset);
  51         if (group != sbi->s_groups_count)
  52                 ext4_warning(sb, __FUNCTION__,
  53                              "Cannot add at group %u (only %lu groups)",
  54                              input->group, sbi->s_groups_count);
  55         else if (offset != 0)
  56                         ext4_warning(sb, __FUNCTION__, "Last group not full");
  57         else if (input->reserved_blocks > input->blocks_count / 5)
  58                 ext4_warning(sb, __FUNCTION__, "Reserved blocks too high (%u)",
  59                              input->reserved_blocks);
  60         else if (free_blocks_count < 0)
  61                 ext4_warning(sb, __FUNCTION__, "Bad blocks count %u",
  62                              input->blocks_count);
  63         else if (!(bh = sb_bread(sb, end - 1)))
  64                 ext4_warning(sb, __FUNCTION__,
  65                              "Cannot read last block (%llu)",
  66                              end - 1);
  67         else if (outside(input->block_bitmap, start, end))
  68                 ext4_warning(sb, __FUNCTION__,
  69                              "Block bitmap not in group (block %llu)",
  70                              (unsigned long long)input->block_bitmap);
  71         else if (outside(input->inode_bitmap, start, end))
  72                 ext4_warning(sb, __FUNCTION__,
  73                              "Inode bitmap not in group (block %llu)",
  74                              (unsigned long long)input->inode_bitmap);
  75         else if (outside(input->inode_table, start, end) ||
  76                  outside(itend - 1, start, end))
  77                 ext4_warning(sb, __FUNCTION__,
  78                              "Inode table not in group (blocks %llu-%llu)",
  79                              (unsigned long long)input->inode_table, itend - 1);
  80         else if (input->inode_bitmap == input->block_bitmap)
  81                 ext4_warning(sb, __FUNCTION__,
  82                              "Block bitmap same as inode bitmap (%llu)",
  83                              (unsigned long long)input->block_bitmap);
  84         else if (inside(input->block_bitmap, input->inode_table, itend))
  85                 ext4_warning(sb, __FUNCTION__,
  86                              "Block bitmap (%llu) in inode table (%llu-%llu)",
  87                              (unsigned long long)input->block_bitmap,
  88                              (unsigned long long)input->inode_table, itend - 1);
  89         else if (inside(input->inode_bitmap, input->inode_table, itend))
  90                 ext4_warning(sb, __FUNCTION__,
  91                              "Inode bitmap (%llu) in inode table (%llu-%llu)",
  92                              (unsigned long long)input->inode_bitmap,
  93                              (unsigned long long)input->inode_table, itend - 1);
  94         else if (inside(input->block_bitmap, start, metaend))
  95                 ext4_warning(sb, __FUNCTION__,
  96                              "Block bitmap (%llu) in GDT table"
  97                              " (%llu-%llu)",
  98                              (unsigned long long)input->block_bitmap,
  99                              start, metaend - 1);
 100         else if (inside(input->inode_bitmap, start, metaend))
 101                 ext4_warning(sb, __FUNCTION__,
 102                              "Inode bitmap (%llu) in GDT table"
 103                              " (%llu-%llu)",
 104                              (unsigned long long)input->inode_bitmap,
 105                              start, metaend - 1);
 106         else if (inside(input->inode_table, start, metaend) ||
 107                  inside(itend - 1, start, metaend))
 108                 ext4_warning(sb, __FUNCTION__,
 109                              "Inode table (%llu-%llu) overlaps"
 110                              "GDT table (%llu-%llu)",
 111                              (unsigned long long)input->inode_table,
 112                              itend - 1, start, metaend - 1);
 113         else
 114                 err = 0;
 115         brelse(bh);
 116
 117         return err;
 118 }
 119
 120 static struct buffer_head *bclean(handle_t *handle, struct super_block *sb,
 121                                   ext4_fsblk_t blk)
 122 {
 123         struct buffer_head *bh;
 124         int err;
 125
 126         bh = sb_getblk(sb, blk);
 127         if (!bh)
 128                 return ERR_PTR(-EIO);
 129         if ((err = ext4_journal_get_write_access(handle, bh))) {
 130                 brelse(bh);
 131                 bh = ERR_PTR(err);
 132         } else {
 133                 lock_buffer(bh);
 134                 memset(bh->b_data, 0, sb->s_blocksize);
 135                 set_buffer_uptodate(bh);
 136                 unlock_buffer(bh);
 137         }
 138
 139         return bh;
 140 }
 141
 142 /*
 143  * To avoid calling the atomic setbit hundreds or thousands of times, we only
 144  * need to use it within a single byte (to ensure we get endianness right).
 145  * We can use memset for the rest of the bitmap as there are no other users.
 146  */
 147 static void mark_bitmap_end(int start_bit, int end_bit, char *bitmap)
 148 {
 149         int i;
 150
 151         if (start_bit >= end_bit)
 152                 return;
 153
 154         ext4_debug("mark end bits +%d through +%d used\n", start_bit, end_bit);
 155         for (i = start_bit; i < ((start_bit + 7) & ~7UL); i++)
 156                 ext4_set_bit(i, bitmap);
 157         if (i < end_bit)
 158                 memset(bitmap + (i >> 3), 0xff, (end_bit - i) >> 3);
 159 }
 160
 161 /*
 162  * Set up the block and inode bitmaps, and the inode table for the new group.
 163  * This doesn't need to be part of the main transaction, since we are only
 164  * changing blocks outside the actual filesystem.  We still do journaling to
 165  * ensure the recovery is correct in case of a failure just after resize.
 166  * If any part of this fails, we simply abort the resize.
 167  */
 168 static int setup_new_group_blocks(struct super_block *sb,
 169                                   struct ext4_new_group_data *input)
 170 {
 171         struct ext4_sb_info *sbi = EXT4_SB(sb);
 172         ext4_fsblk_t start = ext4_group_first_block_no(sb, input->group);
 173         int reserved_gdb = ext4_bg_has_super(sb, input->group) ?
 174                 le16_to_cpu(sbi->s_es->s_reserved_gdt_blocks) : 0;
 175         unsigned long gdblocks = ext4_bg_num_gdb(sb, input->group);
 176         struct buffer_head *bh;
 177         handle_t *handle;
 178         ext4_fsblk_t block;
 179         ext4_grpblk_t bit;
 180         int i;
 181         int err = 0, err2;
 182
 183         handle = ext4_journal_start_sb(sb, reserved_gdb + gdblocks +
 184                                        2 + sbi->s_itb_per_group);
 185         if (IS_ERR(handle))
 186                 return PTR_ERR(handle);
 187
 188         lock_super(sb);
 189         if (input->group != sbi->s_groups_count) {
 190                 err = -EBUSY;
 191                 goto exit_journal;
 192         }
 193
 194         if (IS_ERR(bh = bclean(handle, sb, input->block_bitmap))) {
 195                 err = PTR_ERR(bh);
 196                 goto exit_journal;
 197         }
 198
 199         if (ext4_bg_has_super(sb, input->group)) {
 200                 ext4_debug("mark backup superblock %#04lx (+0)\n", start);
 201                 ext4_set_bit(0, bh->b_data);
 202         }
 203
 204         /* Copy all of the GDT blocks into the backup in this group */
 205         for (i = 0, bit = 1, block = start + 1;
 206              i < gdblocks; i++, block++, bit++) {
 207                 struct buffer_head *gdb;
 208
 209                 ext4_debug("update backup group %#04lx (+%d)\n", block, bit);
 210
 211                 gdb = sb_getblk(sb, block);
 212                 if (!gdb) {
 213                         err = -EIO;
 214                         goto exit_bh;
 215                 }
 216                 if ((err = ext4_journal_get_write_access(handle, gdb))) {
 217                         brelse(gdb);
 218                         goto exit_bh;
 219                 }
 220                 lock_buffer(bh);
 221                 memcpy(gdb->b_data, sbi->s_group_desc[i]->b_data, bh->b_size);
 222                 set_buffer_uptodate(gdb);
 223                 unlock_buffer(bh);
 224                 ext4_journal_dirty_metadata(handle, gdb);
 225                 ext4_set_bit(bit, bh->b_data);
 226                 brelse(gdb);
 227         }
 228
 229         /* Zero out all of the reserved backup group descriptor table blocks */
 230         for (i = 0, bit = gdblocks + 1, block = start + bit;
 231              i < reserved_gdb; i++, block++, bit++) {
 232                 struct buffer_head *gdb;
 233
 234                 ext4_debug("clear reserved block %#04lx (+%d)\n", block, bit);
 235
 236                 if (IS_ERR(gdb = bclean(handle, sb, block))) {
 237                         err = PTR_ERR(bh);
 238                         goto exit_bh;
 239                 }
 240                 ext4_journal_dirty_metadata(handle, gdb);
 241                 ext4_set_bit(bit, bh->b_data);
 242                 brelse(gdb);
 243         }
 244         ext4_debug("mark block bitmap %#04x (+%ld)\n", input->block_bitmap,
 245                    input->block_bitmap - start);
 246         ext4_set_bit(input->block_bitmap - start, bh->b_data);
 247         ext4_debug("mark inode bitmap %#04x (+%ld)\n", input->inode_bitmap,
 248                    input->inode_bitmap - start);
 249         ext4_set_bit(input->inode_bitmap - start, bh->b_data);
 250
 251         /* Zero out all of the inode table blocks */
 252         for (i = 0, block = input->inode_table, bit = block - start;
 253              i < sbi->s_itb_per_group; i++, bit++, block++) {
 254                 struct buffer_head *it;
 255
 256                 ext4_debug("clear inode block %#04lx (+%d)\n", block, bit);
 257                 if (IS_ERR(it = bclean(handle, sb, block))) {
 258                         err = PTR_ERR(it);
 259                         goto exit_bh;
 260                 }
 261                 ext4_journal_dirty_metadata(handle, it);
 262                 brelse(it);
 263                 ext4_set_bit(bit, bh->b_data);
 264         }
 265         mark_bitmap_end(input->blocks_count, EXT4_BLOCKS_PER_GROUP(sb),
 266                         bh->b_data);
 267         ext4_journal_dirty_metadata(handle, bh);
 268         brelse(bh);
 269
 270         /* Mark unused entries in inode bitmap used */
 271         ext4_debug("clear inode bitmap %#04x (+%ld)\n",
 272                    input->inode_bitmap, input->inode_bitmap - start);
 273         if (IS_ERR(bh = bclean(handle, sb, input->inode_bitmap))) {
 274                 err = PTR_ERR(bh);
 275                 goto exit_journal;
 276         }
 277
 278         mark_bitmap_end(EXT4_INODES_PER_GROUP(sb), EXT4_BLOCKS_PER_GROUP(sb),
 279                         bh->b_data);
 280         ext4_journal_dirty_metadata(handle, bh);
 281 exit_bh:
 282         brelse(bh);
 283
 284 exit_journal:
 285         unlock_super(sb);
 286         if ((err2 = ext4_journal_stop(handle)) && !err)
 287                 err = err2;
 288
 289         return err;
 290 }
 291
 292
 293 /*
 294  * Iterate through the groups which hold BACKUP superblock/GDT copies in an
 295  * ext4 filesystem.  The counters should be initialized to 1, 5, and 7 before
 296  * calling this for the first time.  In a sparse filesystem it will be the
 297  * sequence of powers of 3, 5, and 7: 1, 3, 5, 7, 9, 25, 27, 49, 81, ...
 298  * For a non-sparse filesystem it will be every group: 1, 2, 3, 4, ...
 299  */
 300 static unsigned ext4_list_backups(struct super_block *sb, unsigned *three,
 301                                   unsigned *five, unsigned *seven)
 302 {
 303         unsigned *min = three;
 304         int mult = 3;
 305         unsigned ret;
 306
 307         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
 308                                         EXT4_FEATURE_RO_COMPAT_SPARSE_SUPER)) {
 309                 ret = *min;
 310                 *min += 1;
 311                 return ret;
 312         }
 313
 314         if (*five < *min) {
 315                 min = five;
 316                 mult = 5;
 317         }
 318         if (*seven < *min) {
 319                 min = seven;
 320                 mult = 7;
 321         }
 322
 323         ret = *min;
 324         *min *= mult;
 325
 326         return ret;
 327 }
 328
 329 /*
 330  * Check that all of the backup GDT blocks are held in the primary GDT block.
 331  * It is assumed that they are stored in group order.  Returns the number of
 332  * groups in current filesystem that have BACKUPS, or -ve error code.
 333  */
 334 static int verify_reserved_gdb(struct super_block *sb,
 335                                struct buffer_head *primary)
 336 {
 337         const ext4_fsblk_t blk = primary->b_blocknr;
 338         const unsigned long end = EXT4_SB(sb)->s_groups_count;
 339         unsigned three = 1;
 340         unsigned five = 5;
 341         unsigned seven = 7;
 342         unsigned grp;
 343         __le32 *p = (__le32 *)primary->b_data;
 344         int gdbackups = 0;
 345
 346         while ((grp = ext4_list_backups(sb, &three, &five, &seven)) < end) {
 347                 if (le32_to_cpu(*p++) !=
 348                     grp * EXT4_BLOCKS_PER_GROUP(sb) + blk){
 349                         ext4_warning(sb, __FUNCTION__,
 350                                      "reserved GDT %llu"
 351                                      " missing grp %d (%llu)",
 352                                      blk, grp,
 353                                      grp *
 354                                      (ext4_fsblk_t)EXT4_BLOCKS_PER_GROUP(sb) +
 355                                      blk);
 356                         return -EINVAL;
 357                 }
 358                 if (++gdbackups > EXT4_ADDR_PER_BLOCK(sb))
 359                         return -EFBIG;
 360         }
 361
 362         return gdbackups;
 363 }
 364
 365 /*
 366  * Called when we need to bring a reserved group descriptor table block into
 367  * use from the resize inode.  The primary copy of the new GDT block currently
 368  * is an indirect block (under the double indirect block in the resize inode).
 369  * The new backup GDT blocks will be stored as leaf blocks in this indirect
 370  * block, in group order.  Even though we know all the block numbers we need,
 371  * we check to ensure that the resize inode has actually reserved these blocks.
 372  *
 373  * Don't need to update the block bitmaps because the blocks are still in use.
 374  *
 375  * We get all of the error cases out of the way, so that we are sure to not
 376  * fail once we start modifying the data on disk, because JBD has no rollback.
 377  */
 378 static int add_new_gdb(handle_t *handle, struct inode *inode,
 379                        struct ext4_new_group_data *input,
 380                        struct buffer_head **primary)
 381 {
 382         struct super_block *sb = inode->i_sb;
 383         struct ext4_super_block *es = EXT4_SB(sb)->s_es;
 384         unsigned long gdb_num = input->group / EXT4_DESC_PER_BLOCK(sb);
 385         ext4_fsblk_t gdblock = EXT4_SB(sb)->s_sbh->b_blocknr + 1 + gdb_num;
 386         struct buffer_head **o_group_desc, **n_group_desc;
 387         struct buffer_head *dind;
 388         int gdbackups;
 389         struct ext4_iloc iloc;
 390         __le32 *data;
 391         int err;
 392
 393         if (test_opt(sb, DEBUG))
 394                 printk(KERN_DEBUG
 395                        "EXT4-fs: ext4_add_new_gdb: adding group block %lu\n",
 396                        gdb_num);
 397
 398         /*
 399          * If we are not using the primary superblock/GDT copy don't resize,
 400          * because the user tools have no way of handling this.  Probably a
 401          * bad time to do it anyways.
 402          */
 403         if (EXT4_SB(sb)->s_sbh->b_blocknr !=
 404             le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block)) {
 405                 ext4_warning(sb, __FUNCTION__,
 406                         "won't resize using backup superblock at %llu",
 407                         (unsigned long long)EXT4_SB(sb)->s_sbh->b_blocknr);
 408                 return -EPERM;
 409         }
 410
 411         *primary = sb_bread(sb, gdblock);
 412         if (!*primary)
 413                 return -EIO;
 414
 415         if ((gdbackups = verify_reserved_gdb(sb, *primary)) < 0) {
 416                 err = gdbackups;
 417                 goto exit_bh;
 418         }
 419
 420         data = EXT4_I(inode)->i_data + EXT4_DIND_BLOCK;
 421         dind = sb_bread(sb, le32_to_cpu(*data));
 422         if (!dind) {
 423                 err = -EIO;
 424                 goto exit_bh;
 425         }
 426
 427         data = (__le32 *)dind->b_data;
 428         if (le32_to_cpu(data[gdb_num % EXT4_ADDR_PER_BLOCK(sb)]) != gdblock) {
 429                 ext4_warning(sb, __FUNCTION__,
 430                              "new group %u GDT block %llu not reserved",
 431                              input->group, gdblock);
 432                 err = -EINVAL;
 433                 goto exit_dind;
 434         }
 435
 436         if ((err = ext4_journal_get_write_access(handle, EXT4_SB(sb)->s_sbh)))
 437                 goto exit_dind;
 438
 439         if ((err = ext4_journal_get_write_access(handle, *primary)))
 440                 goto exit_sbh;
 441
 442         if ((err = ext4_journal_get_write_access(handle, dind)))
 443                 goto exit_primary;
 444
 445         /* ext4_reserve_inode_write() gets a reference on the iloc */
 446         if ((err = ext4_reserve_inode_write(handle, inode, &iloc)))
 447                 goto exit_dindj;
 448
 449         n_group_desc = kmalloc((gdb_num + 1) * sizeof(struct buffer_head *),
 450                         GFP_KERNEL);
 451         if (!n_group_desc) {
 452                 err = -ENOMEM;
 453                 ext4_warning (sb, __FUNCTION__,
 454                               "not enough memory for %lu groups", gdb_num + 1);
 455                 goto exit_inode;
 456         }
 457
 458         /*
 459          * Finally, we have all of the possible failures behind us...
 460          *
 461          * Remove new GDT block from inode double-indirect block and clear out
 462          * the new GDT block for use (which also "frees" the backup GDT blocks
 463          * from the reserved inode).  We don't need to change the bitmaps for
 464          * these blocks, because they are marked as in-use from being in the
 465          * reserved inode, and will become GDT blocks (primary and backup).
 466          */
 467         data[gdb_num % EXT4_ADDR_PER_BLOCK(sb)] = 0;
 468         ext4_journal_dirty_metadata(handle, dind);
 469         brelse(dind);
 470         inode->i_blocks -= (gdbackups + 1) * sb->s_blocksize >> 9;
 471         ext4_mark_iloc_dirty(handle, inode, &iloc);
 472         memset((*primary)->b_data, 0, sb->s_blocksize);
 473         ext4_journal_dirty_metadata(handle, *primary);
 474
 475         o_group_desc = EXT4_SB(sb)->s_group_desc;
 476         memcpy(n_group_desc, o_group_desc,
 477                EXT4_SB(sb)->s_gdb_count * sizeof(struct buffer_head *));
 478         n_group_desc[gdb_num] = *primary;
 479         EXT4_SB(sb)->s_group_desc = n_group_desc;
 480         EXT4_SB(sb)->s_gdb_count++;
 481         kfree(o_group_desc);
 482
 483         es->s_reserved_gdt_blocks =
 484                 cpu_to_le16(le16_to_cpu(es->s_reserved_gdt_blocks) - 1);
 485         ext4_journal_dirty_metadata(handle, EXT4_SB(sb)->s_sbh);
 486
 487         return 0;
 488
 489 exit_inode:
 490         //ext4_journal_release_buffer(handle, iloc.bh);
 491         brelse(iloc.bh);
 492 exit_dindj:
 493         //ext4_journal_release_buffer(handle, dind);
 494 exit_primary:
 495         //ext4_journal_release_buffer(handle, *primary);
 496 exit_sbh:
 497         //ext4_journal_release_buffer(handle, *primary);
 498 exit_dind:
 499         brelse(dind);
 500 exit_bh:
 501         brelse(*primary);
 502
 503         ext4_debug("leaving with error %d\n", err);
 504         return err;
 505 }
 506
 507 /*
 508  * Called when we are adding a new group which has a backup copy of each of
 509  * the GDT blocks (i.e. sparse group) and there are reserved GDT blocks.
 510  * We need to add these reserved backup GDT blocks to the resize inode, so
 511  * that they are kept for future resizing and not allocated to files.
 512  *
 513  * Each reserved backup GDT block will go into a different indirect block.
 514  * The indirect blocks are actually the primary reserved GDT blocks,
 515  * so we know in advance what their block numbers are.  We only get the
 516  * double-indirect block to verify it is pointing to the primary reserved
 517  * GDT blocks so we don't overwrite a data block by accident.  The reserved
 518  * backup GDT blocks are stored in their reserved primary GDT block.
 519  */
 520 static int reserve_backup_gdb(handle_t *handle, struct inode *inode,
 521                               struct ext4_new_group_data *input)
 522 {
 523         struct super_block *sb = inode->i_sb;
 524         int reserved_gdb =le16_to_cpu(EXT4_SB(sb)->s_es->s_reserved_gdt_blocks);
 525         struct buffer_head **primary;
 526         struct buffer_head *dind;
 527         struct ext4_iloc iloc;
 528         ext4_fsblk_t blk;
 529         __le32 *data, *end;
 530         int gdbackups = 0;
 531         int res, i;
 532         int err;
 533
 534         primary = kmalloc(reserved_gdb * sizeof(*primary), GFP_KERNEL);
 535         if (!primary)
 536                 return -ENOMEM;
 537
 538         data = EXT4_I(inode)->i_data + EXT4_DIND_BLOCK;
 539         dind = sb_bread(sb, le32_to_cpu(*data));
 540         if (!dind) {
 541                 err = -EIO;
 542                 goto exit_free;
 543         }
 544
 545         blk = EXT4_SB(sb)->s_sbh->b_blocknr + 1 + EXT4_SB(sb)->s_gdb_count;
 546         data = (__le32 *)dind->b_data + EXT4_SB(sb)->s_gdb_count;
 547         end = (__le32 *)dind->b_data + EXT4_ADDR_PER_BLOCK(sb);
 548
 549         /* Get each reserved primary GDT block and verify it holds backups */
 550         for (res = 0; res < reserved_gdb; res++, blk++) {
 551                 if (le32_to_cpu(*data) != blk) {
 552                         ext4_warning(sb, __FUNCTION__,
 553                                      "reserved block %llu"
 554                                      " not at offset %ld",
 555                                      blk,
 556                                      (long)(data - (__le32 *)dind->b_data));
 557                         err = -EINVAL;
 558                         goto exit_bh;
 559                 }
 560                 primary[res] = sb_bread(sb, blk);
 561                 if (!primary[res]) {
 562                         err = -EIO;
 563                         goto exit_bh;
 564                 }
 565                 if ((gdbackups = verify_reserved_gdb(sb, primary[res])) < 0) {
 566                         brelse(primary[res]);
 567                         err = gdbackups;
 568                         goto exit_bh;
 569                 }
 570                 if (++data >= end)
 571                         data = (__le32 *)dind->b_data;
 572         }
 573
 574         for (i = 0; i < reserved_gdb; i++) {
 575                 if ((err = ext4_journal_get_write_access(handle, primary[i]))) {
 576                         /*
 577                         int j;
 578                         for (j = 0; j < i; j++)
 579                                 ext4_journal_release_buffer(handle, primary[j]);
 580                          */
 581                         goto exit_bh;
 582                 }
 583         }
 584
 585         if ((err = ext4_reserve_inode_write(handle, inode, &iloc)))
 586                 goto exit_bh;
 587
 588         /*
 589          * Finally we can add each of the reserved backup GDT blocks from
 590          * the new group to its reserved primary GDT block.
 591          */
 592         blk = input->group * EXT4_BLOCKS_PER_GROUP(sb);
 593         for (i = 0; i < reserved_gdb; i++) {
 594                 int err2;
 595                 data = (__le32 *)primary[i]->b_data;
 596                 /* printk("reserving backup %lu[%u] = %lu\n",
 597                        primary[i]->b_blocknr, gdbackups,
 598                        blk + primary[i]->b_blocknr); */
 599                 data[gdbackups] = cpu_to_le32(blk + primary[i]->b_blocknr);
 600                 err2 = ext4_journal_dirty_metadata(handle, primary[i]);
 601                 if (!err)
 602                         err = err2;
 603         }
 604         inode->i_blocks += reserved_gdb * sb->s_blocksize >> 9;
 605         ext4_mark_iloc_dirty(handle, inode, &iloc);
 606
 607 exit_bh:
 608         while (--res >= 0)
 609                 brelse(primary[res]);
 610         brelse(dind);
 611
 612 exit_free:
 613         kfree(primary);
 614
 615         return err;
 616 }
 617
 618 /*
 619  * Update the backup copies of the ext4 metadata.  These don't need to be part
 620  * of the main resize transaction, because e2fsck will re-write them if there
 621  * is a problem (basically only OOM will cause a problem).  However, we
 622  * _should_ update the backups if possible, in case the primary gets trashed
 623  * for some reason and we need to run e2fsck from a backup superblock.  The
 624  * important part is that the new block and inode counts are in the backup
 625  * superblocks, and the location of the new group metadata in the GDT backups.
 626  *
 627  * We do not need lock_super() for this, because these blocks are not
 628  * otherwise touched by the filesystem code when it is mounted.  We don't
 629  * need to worry about last changing from sbi->s_groups_count, because the
 630  * worst that can happen is that we do not copy the full number of backups
 631  * at this time.  The resize which changed s_groups_count will backup again.
 632  */
 633 static void update_backups(struct super_block *sb,
 634                            int blk_off, char *data, int size)
 635 {
 636         struct ext4_sb_info *sbi = EXT4_SB(sb);
 637         const unsigned long last = sbi->s_groups_count;
 638         const int bpg = EXT4_BLOCKS_PER_GROUP(sb);
 639         unsigned three = 1;
 640         unsigned five = 5;
 641         unsigned seven = 7;
 642         unsigned group;
 643         int rest = sb->s_blocksize - size;
 644         handle_t *handle;
 645         int err = 0, err2;
 646
 647         handle = ext4_journal_start_sb(sb, EXT4_MAX_TRANS_DATA);
 648         if (IS_ERR(handle)) {
 649                 group = 1;
 650                 err = PTR_ERR(handle);
 651                 goto exit_err;
 652         }
 653
 654         while ((group = ext4_list_backups(sb, &three, &five, &seven)) < last) {
 655                 struct buffer_head *bh;
 656
 657                 /* Out of journal space, and can't get more - abort - so sad */
 658                 if (handle->h_buffer_credits == 0 &&
 659                     ext4_journal_extend(handle, EXT4_MAX_TRANS_DATA) &&
 660                     (err = ext4_journal_restart(handle, EXT4_MAX_TRANS_DATA)))
 661                         break;
 662
 663                 bh = sb_getblk(sb, group * bpg + blk_off);
 664                 if (!bh) {
 665                         err = -EIO;
 666                         break;
 667                 }
 668                 ext4_debug("update metadata backup %#04lx\n",
 669                           (unsigned long)bh->b_blocknr);
 670                 if ((err = ext4_journal_get_write_access(handle, bh)))
 671                         break;
 672                 lock_buffer(bh);
 673                 memcpy(bh->b_data, data, size);
 674                 if (rest)
 675                         memset(bh->b_data + size, 0, rest);
 676                 set_buffer_uptodate(bh);
 677                 unlock_buffer(bh);
 678                 ext4_journal_dirty_metadata(handle, bh);
 679                 brelse(bh);
 680         }
 681         if ((err2 = ext4_journal_stop(handle)) && !err)
 682                 err = err2;
 683
 684         /*
 685          * Ugh! Need to have e2fsck write the backup copies.  It is too
 686          * late to revert the resize, we shouldn't fail just because of
 687          * the backup copies (they are only needed in case of corruption).
 688          *
 689          * However, if we got here we have a journal problem too, so we
 690          * can't really start a transaction to mark the superblock.
 691          * Chicken out and just set the flag on the hope it will be written
 692          * to disk, and if not - we will simply wait until next fsck.
 693          */
 694 exit_err:
 695         if (err) {
 696                 ext4_warning(sb, __FUNCTION__,
 697                              "can't update backup for group %d (err %d), "
 698                              "forcing fsck on next reboot", group, err);
 699                 sbi->s_mount_state &= ~EXT4_VALID_FS;
 700                 sbi->s_es->s_state &= cpu_to_le16(~EXT4_VALID_FS);
 701                 mark_buffer_dirty(sbi->s_sbh);
 702         }
 703 }
 704
 705 /* Add group descriptor data to an existing or new group descriptor block.
 706  * Ensure we handle all possible error conditions _before_ we start modifying
 707  * the filesystem, because we cannot abort the transaction and not have it
 708  * write the data to disk.
 709  *
 710  * If we are on a GDT block boundary, we need to get the reserved GDT block.
 711  * Otherwise, we may need to add backup GDT blocks for a sparse group.
 712  *
 713  * We only need to hold the superblock lock while we are actually adding
 714  * in the new group's counts to the superblock.  Prior to that we have
 715  * not really "added" the group at all.  We re-check that we are still
 716  * adding in the last group in case things have changed since verifying.
 717  */
 718 int ext4_group_add(struct super_block *sb, struct ext4_new_group_data *input)
 719 {
 720         struct ext4_sb_info *sbi = EXT4_SB(sb);
 721         struct ext4_super_block *es = sbi->s_es;
 722         int reserved_gdb = ext4_bg_has_super(sb, input->group) ?
 723                 le16_to_cpu(es->s_reserved_gdt_blocks) : 0;
 724         struct buffer_head *primary = NULL;
 725         struct ext4_group_desc *gdp;
 726         struct inode *inode = NULL;
 727         handle_t *handle;
 728         int gdb_off, gdb_num;
 729         int err, err2;
 730
 731         gdb_num = input->group / EXT4_DESC_PER_BLOCK(sb);
 732         gdb_off = input->group % EXT4_DESC_PER_BLOCK(sb);
 733
 734         if (gdb_off == 0 && !EXT4_HAS_RO_COMPAT_FEATURE(sb,
 735                                         EXT4_FEATURE_RO_COMPAT_SPARSE_SUPER)) {
 736                 ext4_warning(sb, __FUNCTION__,
 737                              "Can't resize non-sparse filesystem further");
 738                 return -EPERM;
 739         }
 740
 741         if (ext4_blocks_count(es) + input->blocks_count <
 742             ext4_blocks_count(es)) {
 743                 ext4_warning(sb, __FUNCTION__, "blocks_count overflow\n");
 744                 return -EINVAL;
 745         }
 746
 747         if (le32_to_cpu(es->s_inodes_count) + EXT4_INODES_PER_GROUP(sb) <
 748             le32_to_cpu(es->s_inodes_count)) {
 749                 ext4_warning(sb, __FUNCTION__, "inodes_count overflow\n");
 750                 return -EINVAL;
 751         }
 752
 753         if (reserved_gdb || gdb_off == 0) {
 754                 if (!EXT4_HAS_COMPAT_FEATURE(sb,
 755                                              EXT4_FEATURE_COMPAT_RESIZE_INODE)){
 756                         ext4_warning(sb, __FUNCTION__,
 757                                      "No reserved GDT blocks, can't resize");
 758                         return -EPERM;
 759                 }
 760                 inode = iget(sb, EXT4_RESIZE_INO);
 761                 if (!inode || is_bad_inode(inode)) {
 762                         ext4_warning(sb, __FUNCTION__,
 763                                      "Error opening resize inode");
 764                         iput(inode);
 765                         return -ENOENT;
 766                 }
 767         }
 768
 769         if ((err = verify_group_input(sb, input)))
 770                 goto exit_put;
 771
 772         if ((err = setup_new_group_blocks(sb, input)))
 773                 goto exit_put;
 774
 775         /*
 776          * We will always be modifying at least the superblock and a GDT
 777          * block.  If we are adding a group past the last current GDT block,
 778          * we will also modify the inode and the dindirect block.  If we
 779          * are adding a group with superblock/GDT backups  we will also
 780          * modify each of the reserved GDT dindirect blocks.
 781          */
 782         handle = ext4_journal_start_sb(sb,
 783                                        ext4_bg_has_super(sb, input->group) ?
 784                                        3 + reserved_gdb : 4);
 785         if (IS_ERR(handle)) {
 786                 err = PTR_ERR(handle);
 787                 goto exit_put;
 788         }
 789
 790         lock_super(sb);
 791         if (input->group != sbi->s_groups_count) {
 792                 ext4_warning(sb, __FUNCTION__,
 793                              "multiple resizers run on filesystem!");
 794                 err = -EBUSY;
 795                 goto exit_journal;
 796         }
 797
 798         if ((err = ext4_journal_get_write_access(handle, sbi->s_sbh)))
 799                 goto exit_journal;
 800
 801         /*
 802          * We will only either add reserved group blocks to a backup group
 803          * or remove reserved blocks for the first group in a new group block.
 804          * Doing both would be mean more complex code, and sane people don't
 805          * use non-sparse filesystems anymore.  This is already checked above.
 806          */
 807         if (gdb_off) {
 808                 primary = sbi->s_group_desc[gdb_num];
 809                 if ((err = ext4_journal_get_write_access(handle, primary)))
 810                         goto exit_journal;
 811
 812                 if (reserved_gdb && ext4_bg_num_gdb(sb, input->group) &&
 813                     (err = reserve_backup_gdb(handle, inode, input)))
 814                         goto exit_journal;
 815         } else if ((err = add_new_gdb(handle, inode, input, &primary)))
 816                 goto exit_journal;
 817
 818         /*
 819          * OK, now we've set up the new group.  Time to make it active.
 820          *
 821          * Current kernels don't lock all allocations via lock_super(),
 822          * so we have to be safe wrt. concurrent accesses the group
 823          * data.  So we need to be careful to set all of the relevant
 824          * group descriptor data etc. *before* we enable the group.
 825          *
 826          * The key field here is sbi->s_groups_count: as long as
 827          * that retains its old value, nobody is going to access the new
 828          * group.
 829          *
 830          * So first we update all the descriptor metadata for the new
 831          * group; then we update the total disk blocks count; then we
 832          * update the groups count to enable the group; then finally we
 833          * update the free space counts so that the system can start
 834          * using the new disk blocks.
 835          */
 836
 837         /* Update group descriptor block for new group */
 838         gdp = (struct ext4_group_desc *)primary->b_data + gdb_off;
 839
 840         ext4_block_bitmap_set(sb, gdp, input->block_bitmap); /* LV FIXME */
 841         ext4_inode_bitmap_set(sb, gdp, input->inode_bitmap); /* LV FIXME */
 842         ext4_inode_table_set(sb, gdp, input->inode_table); /* LV FIXME */
 843         gdp->bg_free_blocks_count = cpu_to_le16(input->free_blocks_count);
 844         gdp->bg_free_inodes_count = cpu_to_le16(EXT4_INODES_PER_GROUP(sb));
 845
 846         /*
 847          * Make the new blocks and inodes valid next.  We do this before
 848          * increasing the group count so that once the group is enabled,
 849          * all of its blocks and inodes are already valid.
 850          *
 851          * We always allocate group-by-group, then block-by-block or
 852          * inode-by-inode within a group, so enabling these
 853          * blocks/inodes before the group is live won't actually let us
 854          * allocate the new space yet.
 855          */
 856         ext4_blocks_count_set(es, ext4_blocks_count(es) +
 857                 input->blocks_count);
 858         es->s_inodes_count = cpu_to_le32(le32_to_cpu(es->s_inodes_count) +
 859                 EXT4_INODES_PER_GROUP(sb));
 860
 861         /*
 862          * We need to protect s_groups_count against other CPUs seeing
 863          * inconsistent state in the superblock.
 864          *
 865          * The precise rules we use are:
 866          *
 867          * * Writers of s_groups_count *must* hold lock_super
 868          * AND
 869          * * Writers must perform a smp_wmb() after updating all dependent
 870          *   data and before modifying the groups count
 871          *
 872          * * Readers must hold lock_super() over the access
 873          * OR
 874          * * Readers must perform an smp_rmb() after reading the groups count
 875          *   and before reading any dependent data.
 876          *
 877          * NB. These rules can be relaxed when checking the group count
 878          * while freeing data, as we can only allocate from a block
 879          * group after serialising against the group count, and we can
 880          * only then free after serialising in turn against that
 881          * allocation.
 882          */
 883         smp_wmb();
 884
 885         /* Update the global fs size fields */
 886         sbi->s_groups_count++;
 887
 888         ext4_journal_dirty_metadata(handle, primary);
 889
 890         /* Update the reserved block counts only once the new group is
 891          * active. */
 892         ext4_r_blocks_count_set(es, ext4_r_blocks_count(es) +
 893                 input->reserved_blocks);
 894
 895         /* Update the free space counts */
 896         percpu_counter_mod(&sbi->s_freeblocks_counter,
 897                            input->free_blocks_count);
 898         percpu_counter_mod(&sbi->s_freeinodes_counter,
 899                            EXT4_INODES_PER_GROUP(sb));
 900
 901         ext4_journal_dirty_metadata(handle, sbi->s_sbh);
 902         sb->s_dirt = 1;
 903
 904 exit_journal:
 905         unlock_super(sb);
 906         if ((err2 = ext4_journal_stop(handle)) && !err)
 907                 err = err2;
 908         if (!err) {
 909                 update_backups(sb, sbi->s_sbh->b_blocknr, (char *)es,
 910                                sizeof(struct ext4_super_block));
 911                 update_backups(sb, primary->b_blocknr, primary->b_data,
 912                                primary->b_size);
 913         }
 914 exit_put:
 915         iput(inode);
 916         return err;
 917 } /* ext4_group_add */
 918
 919 /* Extend the filesystem to the new number of blocks specified.  This entry
 920  * point is only used to extend the current filesystem to the end of the last
 921  * existing group.  It can be accessed via ioctl, or by "remount,resize=<size>"
 922  * for emergencies (because it has no dependencies on reserved blocks).
 923  *
 924  * If we _really_ wanted, we could use default values to call ext4_group_add()
 925  * allow the "remount" trick to work for arbitrary resizing, assuming enough
 926  * GDT blocks are reserved to grow to the desired size.
 927  */
 928 int ext4_group_extend(struct super_block *sb, struct ext4_super_block *es,
 929                       ext4_fsblk_t n_blocks_count)
 930 {
 931         ext4_fsblk_t o_blocks_count;
 932         unsigned long o_groups_count;
 933         ext4_grpblk_t last;
 934         ext4_grpblk_t add;
 935         struct buffer_head * bh;
 936         handle_t *handle;
 937         int err;
 938         unsigned long freed_blocks;
 939
 940         /* We don't need to worry about locking wrt other resizers just
 941          * yet: we're going to revalidate es->s_blocks_count after
 942          * taking lock_super() below. */
 943         o_blocks_count = ext4_blocks_count(es);
 944         o_groups_count = EXT4_SB(sb)->s_groups_count;
 945
 946         if (test_opt(sb, DEBUG))
 947                 printk(KERN_DEBUG "EXT4-fs: extending last group from %llu uto %llu blocks\n",
 948                        o_blocks_count, n_blocks_count);
 949
 950         if (n_blocks_count == 0 || n_blocks_count == o_blocks_count)
 951                 return 0;
 952
 953         if (n_blocks_count > (sector_t)(~0ULL) >> (sb->s_blocksize_bits - 9)) {
 954                 printk(KERN_ERR "EXT4-fs: filesystem on %s:"
 955                         " too large to resize to %llu blocks safely\n",
 956                         sb->s_id, n_blocks_count);
 957                 if (sizeof(sector_t) < 8)
 958                         ext4_warning(sb, __FUNCTION__,
 959                         "CONFIG_LBD not enabled\n");
 960                 return -EINVAL;
 961         }
 962
 963         if (n_blocks_count < o_blocks_count) {
 964                 ext4_warning(sb, __FUNCTION__,
 965                              "can't shrink FS - resize aborted");
 966                 return -EBUSY;
 967         }
 968
 969         /* Handle the remaining blocks in the last group only. */
 970         ext4_get_group_no_and_offset(sb, o_blocks_count, NULL, &last);
 971
 972         if (last == 0) {
 973                 ext4_warning(sb, __FUNCTION__,
 974                              "need to use ext2online to resize further");
 975                 return -EPERM;
 976         }
 977
 978         add = EXT4_BLOCKS_PER_GROUP(sb) - last;
 979
 980         if (o_blocks_count + add < o_blocks_count) {
 981                 ext4_warning(sb, __FUNCTION__, "blocks_count overflow");
 982                 return -EINVAL;
 983         }
 984
 985         if (o_blocks_count + add > n_blocks_count)
 986                 add = n_blocks_count - o_blocks_count;
 987
 988         if (o_blocks_count + add < n_blocks_count)
 989                 ext4_warning(sb, __FUNCTION__,
 990                              "will only finish group (%llu"
 991                              " blocks, %u new)",
 992                              o_blocks_count + add, add);
 993
 994         /* See if the device is actually as big as what was requested */
 995         bh = sb_bread(sb, o_blocks_count + add -1);
 996         if (!bh) {
 997                 ext4_warning(sb, __FUNCTION__,
 998                              "can't read last block, resize aborted");
 999                 return -ENOSPC;
1000         }
1001         brelse(bh);
1002
1003         /* We will update the superblock, one block bitmap, and
1004          * one group descriptor via ext4_free_blocks().
1005          */
1006         handle = ext4_journal_start_sb(sb, 3);
1007         if (IS_ERR(handle)) {
1008                 err = PTR_ERR(handle);
1009                 ext4_warning(sb, __FUNCTION__, "error %d on journal start",err);
1010                 goto exit_put;
1011         }
1012
1013         lock_super(sb);
1014         if (o_blocks_count != ext4_blocks_count(es)) {
1015                 ext4_warning(sb, __FUNCTION__,
1016                              "multiple resizers run on filesystem!");
1017                 unlock_super(sb);
1018                 err = -EBUSY;
1019                 goto exit_put;
1020         }
1021
1022         if ((err = ext4_journal_get_write_access(handle,
1023                                                  EXT4_SB(sb)->s_sbh))) {
1024                 ext4_warning(sb, __FUNCTION__,
1025                              "error %d on journal write access", err);
1026                 unlock_super(sb);
1027                 ext4_journal_stop(handle);
1028                 goto exit_put;
1029         }
1030         ext4_blocks_count_set(es, o_blocks_count + add);
1031         ext4_journal_dirty_metadata(handle, EXT4_SB(sb)->s_sbh);
1032         sb->s_dirt = 1;
1033         unlock_super(sb);
1034         ext4_debug("freeing blocks %lu through %llu\n", o_blocks_count,
1035                    o_blocks_count + add);
1036         ext4_free_blocks_sb(handle, sb, o_blocks_count, add, &freed_blocks);
1037         ext4_debug("freed blocks %llu through %llu\n", o_blocks_count,
1038                    o_blocks_count + add);
1039         if ((err = ext4_journal_stop(handle)))
1040                 goto exit_put;
1041         if (test_opt(sb, DEBUG))
1042                 printk(KERN_DEBUG "EXT4-fs: extended group to %llu blocks\n",
1043                        ext4_blocks_count(es));
1044         update_backups(sb, EXT4_SB(sb)->s_sbh->b_blocknr, (char *)es,
1045                        sizeof(struct ext4_super_block));
1046 exit_put:
1047         return err;
1048 } /* ext4_group_extend */