fs/ext3/resize.c

   1 /*
   2  *  linux/fs/ext3/resize.c
   3  *
   4  * Support for resizing an ext3 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 #include <linux/config.h>
  12
  13 #define EXT3FS_DEBUG
  14
  15 #include <linux/sched.h>
  16 #include <linux/smp_lock.h>
  17 #include <linux/ext3_jbd.h>
  18
  19 #include <linux/errno.h>
  20 #include <linux/slab.h>
  21
  22
  23 #define outside(b, first, last) ((b) < (first) || (b) >= (last))
  24 #define inside(b, first, last)  ((b) >= (first) && (b) < (last))
  25
  26 static int verify_group_input(struct super_block *sb,
  27                               struct ext3_new_group_data *input)
  28 {
  29         struct ext3_sb_info *sbi = EXT3_SB(sb);
  30         struct ext3_super_block *es = sbi->s_es;
  31         unsigned start = le32_to_cpu(es->s_blocks_count);
  32         unsigned end = start + input->blocks_count;
  33         unsigned group = input->group;
  34         unsigned itend = input->inode_table + EXT3_SB(sb)->s_itb_per_group;
  35         unsigned overhead = ext3_bg_has_super(sb, group) ?
  36                 (1 + ext3_bg_num_gdb(sb, group) +
  37                  le16_to_cpu(es->s_reserved_gdt_blocks)) : 0;
  38         unsigned metaend = start + overhead;
  39         struct buffer_head *bh = NULL;
  40         int free_blocks_count;
  41         int err = -EINVAL;
  42
  43         input->free_blocks_count = free_blocks_count =
  44                 input->blocks_count - 2 - overhead - sbi->s_itb_per_group;
  45
  46         if (test_opt(sb, DEBUG))
  47                 printk(KERN_DEBUG "EXT3-fs: adding %s group %u: %u blocks "
  48                        "(%d free, %u reserved)\n",
  49                        ext3_bg_has_super(sb, input->group) ? "normal" :
  50                        "no-super", input->group, input->blocks_count,
  51                        free_blocks_count, input->reserved_blocks);
  52
  53         if (group != sbi->s_groups_count)
  54                 ext3_warning(sb, __FUNCTION__,
  55                              "Cannot add at group %u (only %lu groups)",
  56                              input->group, sbi->s_groups_count);
  57         else if ((start - le32_to_cpu(es->s_first_data_block)) %
  58                  EXT3_BLOCKS_PER_GROUP(sb))
  59                 ext3_warning(sb, __FUNCTION__, "Last group not full");
  60         else if (input->reserved_blocks > input->blocks_count / 5)
  61                 ext3_warning(sb, __FUNCTION__, "Reserved blocks too high (%u)",
  62                              input->reserved_blocks);
  63         else if (free_blocks_count < 0)
  64                 ext3_warning(sb, __FUNCTION__, "Bad blocks count %u",
  65                              input->blocks_count);
  66         else if (!(bh = sb_bread(sb, end - 1)))
  67                 ext3_warning(sb, __FUNCTION__, "Cannot read last block (%u)",
  68                              end - 1);
  69         else if (outside(input->block_bitmap, start, end))
  70                 ext3_warning(sb, __FUNCTION__,
  71                              "Block bitmap not in group (block %u)",
  72                              input->block_bitmap);
  73         else if (outside(input->inode_bitmap, start, end))
  74                 ext3_warning(sb, __FUNCTION__,
  75                              "Inode bitmap not in group (block %u)",
  76                              input->inode_bitmap);
  77         else if (outside(input->inode_table, start, end) ||
  78                  outside(itend - 1, start, end))
  79                 ext3_warning(sb, __FUNCTION__,
  80                              "Inode table not in group (blocks %u-%u)",
  81                              input->inode_table, itend - 1);
  82         else if (input->inode_bitmap == input->block_bitmap)
  83                 ext3_warning(sb, __FUNCTION__,
  84                              "Block bitmap same as inode bitmap (%u)",
  85                              input->block_bitmap);
  86         else if (inside(input->block_bitmap, input->inode_table, itend))
  87                 ext3_warning(sb, __FUNCTION__,
  88                              "Block bitmap (%u) in inode table (%u-%u)",
  89                              input->block_bitmap, input->inode_table, itend-1);
  90         else if (inside(input->inode_bitmap, input->inode_table, itend))
  91                 ext3_warning(sb, __FUNCTION__,
  92                              "Inode bitmap (%u) in inode table (%u-%u)",
  93                              input->inode_bitmap, input->inode_table, itend-1);
  94         else if (inside(input->block_bitmap, start, metaend))
  95                 ext3_warning(sb, __FUNCTION__,
  96                              "Block bitmap (%u) in GDT table (%u-%u)",
  97                              input->block_bitmap, start, metaend - 1);
  98         else if (inside(input->inode_bitmap, start, metaend))
  99                 ext3_warning(sb, __FUNCTION__,
 100                              "Inode bitmap (%u) in GDT table (%u-%u)",
 101                              input->inode_bitmap, start, metaend - 1);
 102         else if (inside(input->inode_table, start, metaend) ||
 103                  inside(itend - 1, start, metaend))
 104                 ext3_warning(sb, __FUNCTION__,
 105                              "Inode table (%u-%u) overlaps GDT table (%u-%u)",
 106                              input->inode_table, itend - 1, start, metaend - 1);
 107         else
 108                 err = 0;
 109         brelse(bh);
 110
 111         return err;
 112 }
 113
 114 static struct buffer_head *bclean(handle_t *handle, struct super_block *sb,
 115                                   unsigned long blk)
 116 {
 117         struct buffer_head *bh;
 118         int err;
 119
 120         bh = sb_getblk(sb, blk);
 121         if ((err = ext3_journal_get_write_access(handle, bh))) {
 122                 brelse(bh);
 123                 bh = ERR_PTR(err);
 124         } else {
 125                 lock_buffer(bh);
 126                 memset(bh->b_data, 0, sb->s_blocksize);
 127                 set_buffer_uptodate(bh);
 128                 unlock_buffer(bh);
 129         }
 130
 131         return bh;
 132 }
 133
 134 /*
 135  * To avoid calling the atomic setbit hundreds or thousands of times, we only
 136  * need to use it within a single byte (to ensure we get endianness right).
 137  * We can use memset for the rest of the bitmap as there are no other users.
 138  */
 139 static void mark_bitmap_end(int start_bit, int end_bit, char *bitmap)
 140 {
 141         int i;
 142
 143         if (start_bit >= end_bit)
 144                 return;
 145
 146         ext3_debug("mark end bits +%d through +%d used\n", start_bit, end_bit);
 147         for (i = start_bit; i < ((start_bit + 7) & ~7UL); i++)
 148                 ext3_set_bit(i, bitmap);
 149         if (i < end_bit)
 150                 memset(bitmap + (i >> 3), 0xff, (end_bit - i) >> 3);
 151 }
 152
 153 /*
 154  * Set up the block and inode bitmaps, and the inode table for the new group.
 155  * This doesn't need to be part of the main transaction, since we are only
 156  * changing blocks outside the actual filesystem.  We still do journaling to
 157  * ensure the recovery is correct in case of a failure just after resize.
 158  * If any part of this fails, we simply abort the resize.
 159  */
 160 static int setup_new_group_blocks(struct super_block *sb,
 161                                   struct ext3_new_group_data *input)
 162 {
 163         struct ext3_sb_info *sbi = EXT3_SB(sb);
 164         unsigned long start = input->group * sbi->s_blocks_per_group +
 165                 le32_to_cpu(sbi->s_es->s_first_data_block);
 166         int reserved_gdb = ext3_bg_has_super(sb, input->group) ?
 167                 le16_to_cpu(sbi->s_es->s_reserved_gdt_blocks) : 0;
 168         unsigned long gdblocks = ext3_bg_num_gdb(sb, input->group);
 169         struct buffer_head *bh;
 170         handle_t *handle;
 171         unsigned long block;
 172         int bit;
 173         int i;
 174         int err = 0, err2;
 175
 176         handle = ext3_journal_start_sb(sb, reserved_gdb + gdblocks +
 177                                        2 + sbi->s_itb_per_group);
 178         if (IS_ERR(handle))
 179                 return PTR_ERR(handle);
 180
 181         lock_super(sb);
 182         if (input->group != sbi->s_groups_count) {
 183                 err = -EBUSY;
 184                 goto exit_journal;
 185         }
 186
 187         if (IS_ERR(bh = bclean(handle, sb, input->block_bitmap))) {
 188                 err = PTR_ERR(bh);
 189                 goto exit_journal;
 190         }
 191
 192         if (ext3_bg_has_super(sb, input->group)) {
 193                 ext3_debug("mark backup superblock %#04lx (+0)\n", start);
 194                 ext3_set_bit(0, bh->b_data);
 195         }
 196
 197         /* Copy all of the GDT blocks into the backup in this group */
 198         for (i = 0, bit = 1, block = start + 1;
 199              i < gdblocks; i++, block++, bit++) {
 200                 struct buffer_head *gdb;
 201
 202                 ext3_debug("update backup group %#04lx (+%d)\n", block, bit);
 203
 204                 gdb = sb_getblk(sb, block);
 205                 if ((err = ext3_journal_get_write_access(handle, gdb))) {
 206                         brelse(gdb);
 207                         goto exit_bh;
 208                 }
 209                 lock_buffer(bh);
 210                 memcpy(gdb->b_data, sbi->s_group_desc[i], bh->b_size);
 211                 set_buffer_uptodate(gdb);
 212                 unlock_buffer(bh);
 213                 ext3_journal_dirty_metadata(handle, gdb);
 214                 ext3_set_bit(bit, bh->b_data);
 215                 brelse(gdb);
 216         }
 217
 218         /* Zero out all of the reserved backup group descriptor table blocks */
 219         for (i = 0, bit = gdblocks + 1, block = start + bit;
 220              i < reserved_gdb; i++, block++, bit++) {
 221                 struct buffer_head *gdb;
 222
 223                 ext3_debug("clear reserved block %#04lx (+%d)\n", block, bit);
 224
 225                 if (IS_ERR(gdb = bclean(handle, sb, block))) {
 226                         err = PTR_ERR(bh);
 227                         goto exit_bh;
 228                 }
 229                 ext3_journal_dirty_metadata(handle, gdb);
 230                 ext3_set_bit(bit, bh->b_data);
 231                 brelse(gdb);
 232         }
 233         ext3_debug("mark block bitmap %#04x (+%ld)\n", input->block_bitmap,
 234                    input->block_bitmap - start);
 235         ext3_set_bit(input->block_bitmap - start, bh->b_data);
 236         ext3_debug("mark inode bitmap %#04x (+%ld)\n", input->inode_bitmap,
 237                    input->inode_bitmap - start);
 238         ext3_set_bit(input->inode_bitmap - start, bh->b_data);
 239
 240         /* Zero out all of the inode table blocks */
 241         for (i = 0, block = input->inode_table, bit = block - start;
 242              i < sbi->s_itb_per_group; i++, bit++, block++) {
 243                 struct buffer_head *it;
 244
 245                 ext3_debug("clear inode block %#04x (+%ld)\n", block, bit);
 246                 if (IS_ERR(it = bclean(handle, sb, block))) {
 247                         err = PTR_ERR(it);
 248                         goto exit_bh;
 249                 }
 250                 ext3_journal_dirty_metadata(handle, it);
 251                 brelse(it);
 252                 ext3_set_bit(bit, bh->b_data);
 253         }
 254         mark_bitmap_end(input->blocks_count, EXT3_BLOCKS_PER_GROUP(sb),
 255                         bh->b_data);
 256         ext3_journal_dirty_metadata(handle, bh);
 257         brelse(bh);
 258
 259         /* Mark unused entries in inode bitmap used */
 260         ext3_debug("clear inode bitmap %#04x (+%ld)\n",
 261                    input->inode_bitmap, input->inode_bitmap - start);
 262         if (IS_ERR(bh = bclean(handle, sb, input->inode_bitmap))) {
 263                 err = PTR_ERR(bh);
 264                 goto exit_journal;
 265         }
 266
 267         mark_bitmap_end(EXT3_INODES_PER_GROUP(sb), EXT3_BLOCKS_PER_GROUP(sb),
 268                         bh->b_data);
 269         ext3_journal_dirty_metadata(handle, bh);
 270 exit_bh:
 271         brelse(bh);
 272
 273 exit_journal:
 274         unlock_super(sb);
 275         if ((err2 = ext3_journal_stop(handle)) && !err)
 276                 err = err2;
 277
 278         return err;
 279 }
 280
 281 /*
 282  * Iterate through the groups which hold BACKUP superblock/GDT copies in an
 283  * ext3 filesystem.  The counters should be initialized to 1, 5, and 7 before
 284  * calling this for the first time.  In a sparse filesystem it will be the
 285  * sequence of powers of 3, 5, and 7: 1, 3, 5, 7, 9, 25, 27, 49, 81, ...
 286  * For a non-sparse filesystem it will be every group: 1, 2, 3, 4, ...
 287  */
 288 static unsigned ext3_list_backups(struct super_block *sb, unsigned *three,
 289                                   unsigned *five, unsigned *seven)
 290 {
 291         unsigned *min = three;
 292         int mult = 3;
 293         unsigned ret;
 294
 295         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
 296                                         EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER)) {
 297                 ret = *min;
 298                 *min += 1;
 299                 return ret;
 300         }
 301
 302         if (*five < *min) {
 303                 min = five;
 304                 mult = 5;
 305         }
 306         if (*seven < *min) {
 307                 min = seven;
 308                 mult = 7;
 309         }
 310
 311         ret = *min;
 312         *min *= mult;
 313
 314         return ret;
 315 }
 316
 317 /*
 318  * Check that all of the backup GDT blocks are held in the primary GDT block.
 319  * It is assumed that they are stored in group order.  Returns the number of
 320  * groups in current filesystem that have BACKUPS, or -ve error code.
 321  */
 322 static int verify_reserved_gdb(struct super_block *sb,
 323                                struct buffer_head *primary)
 324 {
 325         const unsigned long blk = primary->b_blocknr;
 326         const unsigned long end = EXT3_SB(sb)->s_groups_count;
 327         unsigned three = 1;
 328         unsigned five = 5;
 329         unsigned seven = 7;
 330         unsigned grp;
 331         __u32 *p = (__u32 *)primary->b_data;
 332         int gdbackups = 0;
 333
 334         while ((grp = ext3_list_backups(sb, &three, &five, &seven)) < end) {
 335                 if (le32_to_cpu(*p++) != grp * EXT3_BLOCKS_PER_GROUP(sb) + blk){
 336                         ext3_warning(sb, __FUNCTION__,
 337                                      "reserved GDT %ld missing grp %d (%ld)\n",
 338                                      blk, grp,
 339                                      grp * EXT3_BLOCKS_PER_GROUP(sb) + blk);
 340                         return -EINVAL;
 341                 }
 342                 if (++gdbackups > EXT3_ADDR_PER_BLOCK(sb))
 343                         return -EFBIG;
 344         }
 345
 346         return gdbackups;
 347 }
 348
 349 /*
 350  * Called when we need to bring a reserved group descriptor table block into
 351  * use from the resize inode.  The primary copy of the new GDT block currently
 352  * is an indirect block (under the double indirect block in the resize inode).
 353  * The new backup GDT blocks will be stored as leaf blocks in this indirect
 354  * block, in group order.  Even though we know all the block numbers we need,
 355  * we check to ensure that the resize inode has actually reserved these blocks.
 356  *
 357  * Don't need to update the block bitmaps because the blocks are still in use.
 358  *
 359  * We get all of the error cases out of the way, so that we are sure to not
 360  * fail once we start modifying the data on disk, because JBD has no rollback.
 361  */
 362 static int add_new_gdb(handle_t *handle, struct inode *inode,
 363                        struct ext3_new_group_data *input,
 364                        struct buffer_head **primary)
 365 {
 366         struct super_block *sb = inode->i_sb;
 367         struct ext3_super_block *es = EXT3_SB(sb)->s_es;
 368         unsigned long gdb_num = input->group / EXT3_DESC_PER_BLOCK(sb);
 369         unsigned long gdblock = EXT3_SB(sb)->s_sbh->b_blocknr + 1 + gdb_num;
 370         struct buffer_head **o_group_desc, **n_group_desc;
 371         struct buffer_head *dind;
 372         int gdbackups;
 373         struct ext3_iloc iloc;
 374         __u32 *data;
 375         int err;
 376
 377         if (test_opt(sb, DEBUG))
 378                 printk(KERN_DEBUG
 379                        "EXT3-fs: ext3_add_new_gdb: adding group block %lu\n",
 380                        gdb_num);
 381
 382         /*
 383          * If we are not using the primary superblock/GDT copy don't resize,
 384          * because the user tools have no way of handling this.  Probably a
 385          * bad time to do it anyways.
 386          */
 387         if (EXT3_SB(sb)->s_sbh->b_blocknr !=
 388             le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block)) {
 389                 ext3_warning(sb, __FUNCTION__,
 390                         "won't resize using backup superblock at %llu\n",
 391                         (unsigned long long)EXT3_SB(sb)->s_sbh->b_blocknr);
 392                 return -EPERM;
 393         }
 394
 395         *primary = sb_bread(sb, gdblock);
 396         if (!*primary)
 397                 return -EIO;
 398
 399         if ((gdbackups = verify_reserved_gdb(sb, *primary)) < 0) {
 400                 err = gdbackups;
 401                 goto exit_bh;
 402         }
 403
 404         data = EXT3_I(inode)->i_data + EXT3_DIND_BLOCK;
 405         dind = sb_bread(sb, le32_to_cpu(*data));
 406         if (!dind) {
 407                 err = -EIO;
 408                 goto exit_bh;
 409         }
 410
 411         data = (__u32 *)dind->b_data;
 412         if (le32_to_cpu(data[gdb_num % EXT3_ADDR_PER_BLOCK(sb)]) != gdblock) {
 413                 ext3_warning(sb, __FUNCTION__,
 414                              "new group %u GDT block %lu not reserved\n",
 415                              input->group, gdblock);
 416                 err = -EINVAL;
 417                 goto exit_dind;
 418         }
 419
 420         if ((err = ext3_journal_get_write_access(handle, EXT3_SB(sb)->s_sbh)))
 421                 goto exit_dind;
 422
 423         if ((err = ext3_journal_get_write_access(handle, *primary)))
 424                 goto exit_sbh;
 425
 426         if ((err = ext3_journal_get_write_access(handle, dind)))
 427                 goto exit_primary;
 428
 429         /* ext3_reserve_inode_write() gets a reference on the iloc */
 430         if ((err = ext3_reserve_inode_write(handle, inode, &iloc)))
 431                 goto exit_dindj;
 432
 433         n_group_desc = (struct buffer_head **)kmalloc((gdb_num + 1) *
 434                                 sizeof(struct buffer_head *), GFP_KERNEL);
 435         if (!n_group_desc) {
 436                 err = -ENOMEM;
 437                 ext3_warning (sb, __FUNCTION__,
 438                               "not enough memory for %lu groups", gdb_num + 1);
 439                 goto exit_inode;
 440         }
 441
 442         /*
 443          * Finally, we have all of the possible failures behind us...
 444          *
 445          * Remove new GDT block from inode double-indirect block and clear out
 446          * the new GDT block for use (which also "frees" the backup GDT blocks
 447          * from the reserved inode).  We don't need to change the bitmaps for
 448          * these blocks, because they are marked as in-use from being in the
 449          * reserved inode, and will become GDT blocks (primary and backup).
 450          */
 451         data[gdb_num % EXT3_ADDR_PER_BLOCK(sb)] = 0;
 452         ext3_journal_dirty_metadata(handle, dind);
 453         brelse(dind);
 454         inode->i_blocks -= (gdbackups + 1) * sb->s_blocksize >> 9;
 455         ext3_mark_iloc_dirty(handle, inode, &iloc);
 456         memset((*primary)->b_data, 0, sb->s_blocksize);
 457         ext3_journal_dirty_metadata(handle, *primary);
 458
 459         o_group_desc = EXT3_SB(sb)->s_group_desc;
 460         memcpy(n_group_desc, o_group_desc,
 461                EXT3_SB(sb)->s_gdb_count * sizeof(struct buffer_head *));
 462         n_group_desc[gdb_num] = *primary;
 463         EXT3_SB(sb)->s_group_desc = n_group_desc;
 464         EXT3_SB(sb)->s_gdb_count++;
 465         kfree(o_group_desc);
 466
 467         es->s_reserved_gdt_blocks =
 468                 cpu_to_le16(le16_to_cpu(es->s_reserved_gdt_blocks) - 1);
 469         ext3_journal_dirty_metadata(handle, EXT3_SB(sb)->s_sbh);
 470
 471         return 0;
 472
 473 exit_inode:
 474         //ext3_journal_release_buffer(handle, iloc.bh);
 475         brelse(iloc.bh);
 476 exit_dindj:
 477         //ext3_journal_release_buffer(handle, dind);
 478 exit_primary:
 479         //ext3_journal_release_buffer(handle, *primary);
 480 exit_sbh:
 481         //ext3_journal_release_buffer(handle, *primary);
 482 exit_dind:
 483         brelse(dind);
 484 exit_bh:
 485         brelse(*primary);
 486
 487         ext3_debug("leaving with error %d\n", err);
 488         return err;
 489 }
 490
 491 /*
 492  * Called when we are adding a new group which has a backup copy of each of
 493  * the GDT blocks (i.e. sparse group) and there are reserved GDT blocks.
 494  * We need to add these reserved backup GDT blocks to the resize inode, so
 495  * that they are kept for future resizing and not allocated to files.
 496  *
 497  * Each reserved backup GDT block will go into a different indirect block.
 498  * The indirect blocks are actually the primary reserved GDT blocks,
 499  * so we know in advance what their block numbers are.  We only get the
 500  * double-indirect block to verify it is pointing to the primary reserved
 501  * GDT blocks so we don't overwrite a data block by accident.  The reserved
 502  * backup GDT blocks are stored in their reserved primary GDT block.
 503  */
 504 static int reserve_backup_gdb(handle_t *handle, struct inode *inode,
 505                               struct ext3_new_group_data *input)
 506 {
 507         struct super_block *sb = inode->i_sb;
 508         int reserved_gdb =le16_to_cpu(EXT3_SB(sb)->s_es->s_reserved_gdt_blocks);
 509         struct buffer_head **primary;
 510         struct buffer_head *dind;
 511         struct ext3_iloc iloc;
 512         unsigned long blk;
 513         __u32 *data, *end;
 514         int gdbackups = 0;
 515         int res, i;
 516         int err;
 517
 518         primary = kmalloc(reserved_gdb * sizeof(*primary), GFP_KERNEL);
 519         if (!primary)
 520                 return -ENOMEM;
 521
 522         data = EXT3_I(inode)->i_data + EXT3_DIND_BLOCK;
 523         dind = sb_bread(sb, le32_to_cpu(*data));
 524         if (!dind) {
 525                 err = -EIO;
 526                 goto exit_free;
 527         }
 528
 529         blk = EXT3_SB(sb)->s_sbh->b_blocknr + 1 + EXT3_SB(sb)->s_gdb_count;
 530         data = (__u32 *)dind->b_data + EXT3_SB(sb)->s_gdb_count;
 531         end = (__u32 *)dind->b_data + EXT3_ADDR_PER_BLOCK(sb);
 532
 533         /* Get each reserved primary GDT block and verify it holds backups */
 534         for (res = 0; res < reserved_gdb; res++, blk++) {
 535                 if (le32_to_cpu(*data) != blk) {
 536                         ext3_warning(sb, __FUNCTION__,
 537                                      "reserved block %lu not at offset %ld\n",
 538                                      blk, (long)(data - (__u32 *)dind->b_data));
 539                         err = -EINVAL;
 540                         goto exit_bh;
 541                 }
 542                 primary[res] = sb_bread(sb, blk);
 543                 if (!primary[res]) {
 544                         err = -EIO;
 545                         goto exit_bh;
 546                 }
 547                 if ((gdbackups = verify_reserved_gdb(sb, primary[res])) < 0) {
 548                         brelse(primary[res]);
 549                         err = gdbackups;
 550                         goto exit_bh;
 551                 }
 552                 if (++data >= end)
 553                         data = (__u32 *)dind->b_data;
 554         }
 555
 556         for (i = 0; i < reserved_gdb; i++) {
 557                 if ((err = ext3_journal_get_write_access(handle, primary[i]))) {
 558                         /*
 559                         int j;
 560                         for (j = 0; j < i; j++)
 561                                 ext3_journal_release_buffer(handle, primary[j]);
 562                          */
 563                         goto exit_bh;
 564                 }
 565         }
 566
 567         if ((err = ext3_reserve_inode_write(handle, inode, &iloc)))
 568                 goto exit_bh;
 569
 570         /*
 571          * Finally we can add each of the reserved backup GDT blocks from
 572          * the new group to its reserved primary GDT block.
 573          */
 574         blk = input->group * EXT3_BLOCKS_PER_GROUP(sb);
 575         for (i = 0; i < reserved_gdb; i++) {
 576                 int err2;
 577                 data = (__u32 *)primary[i]->b_data;
 578                 /* printk("reserving backup %lu[%u] = %lu\n",
 579                        primary[i]->b_blocknr, gdbackups,
 580                        blk + primary[i]->b_blocknr); */
 581                 data[gdbackups] = cpu_to_le32(blk + primary[i]->b_blocknr);
 582                 err2 = ext3_journal_dirty_metadata(handle, primary[i]);
 583                 if (!err)
 584                         err = err2;
 585         }
 586         inode->i_blocks += reserved_gdb * sb->s_blocksize >> 9;
 587         ext3_mark_iloc_dirty(handle, inode, &iloc);
 588
 589 exit_bh:
 590         while (--res >= 0)
 591                 brelse(primary[res]);
 592         brelse(dind);
 593
 594 exit_free:
 595         kfree(primary);
 596
 597         return err;
 598 }
 599
 600 /*
 601  * Update the backup copies of the ext3 metadata.  These don't need to be part
 602  * of the main resize transaction, because e2fsck will re-write them if there
 603  * is a problem (basically only OOM will cause a problem).  However, we
 604  * _should_ update the backups if possible, in case the primary gets trashed
 605  * for some reason and we need to run e2fsck from a backup superblock.  The
 606  * important part is that the new block and inode counts are in the backup
 607  * superblocks, and the location of the new group metadata in the GDT backups.
 608  *
 609  * We do not need lock_super() for this, because these blocks are not
 610  * otherwise touched by the filesystem code when it is mounted.  We don't
 611  * need to worry about last changing from sbi->s_groups_count, because the
 612  * worst that can happen is that we do not copy the full number of backups
 613  * at this time.  The resize which changed s_groups_count will backup again.
 614  */
 615 static void update_backups(struct super_block *sb,
 616                            int blk_off, char *data, int size)
 617 {
 618         struct ext3_sb_info *sbi = EXT3_SB(sb);
 619         const unsigned long last = sbi->s_groups_count;
 620         const int bpg = EXT3_BLOCKS_PER_GROUP(sb);
 621         unsigned three = 1;
 622         unsigned five = 5;
 623         unsigned seven = 7;
 624         unsigned group;
 625         int rest = sb->s_blocksize - size;
 626         handle_t *handle;
 627         int err = 0, err2;
 628
 629         handle = ext3_journal_start_sb(sb, EXT3_MAX_TRANS_DATA);
 630         if (IS_ERR(handle)) {
 631                 group = 1;
 632                 err = PTR_ERR(handle);
 633                 goto exit_err;
 634         }
 635
 636         while ((group = ext3_list_backups(sb, &three, &five, &seven)) < last) {
 637                 struct buffer_head *bh;
 638
 639                 /* Out of journal space, and can't get more - abort - so sad */
 640                 if (handle->h_buffer_credits == 0 &&
 641                     ext3_journal_extend(handle, EXT3_MAX_TRANS_DATA) &&
 642                     (err = ext3_journal_restart(handle, EXT3_MAX_TRANS_DATA)))
 643                         break;
 644
 645                 bh = sb_getblk(sb, group * bpg + blk_off);
 646                 ext3_debug(sb, __FUNCTION__, "update metadata backup %#04lx\n",
 647                            bh->b_blocknr);
 648                 if ((err = ext3_journal_get_write_access(handle, bh)))
 649                         break;
 650                 lock_buffer(bh);
 651                 memcpy(bh->b_data, data, size);
 652                 if (rest)
 653                         memset(bh->b_data + size, 0, rest);
 654                 set_buffer_uptodate(bh);
 655                 unlock_buffer(bh);
 656                 ext3_journal_dirty_metadata(handle, bh);
 657                 brelse(bh);
 658         }
 659         if ((err2 = ext3_journal_stop(handle)) && !err)
 660                 err = err2;
 661
 662         /*
 663          * Ugh! Need to have e2fsck write the backup copies.  It is too
 664          * late to revert the resize, we shouldn't fail just because of
 665          * the backup copies (they are only needed in case of corruption).
 666          *
 667          * However, if we got here we have a journal problem too, so we
 668          * can't really start a transaction to mark the superblock.
 669          * Chicken out and just set the flag on the hope it will be written
 670          * to disk, and if not - we will simply wait until next fsck.
 671          */
 672 exit_err:
 673         if (err) {
 674                 ext3_warning(sb, __FUNCTION__,
 675                              "can't update backup for group %d (err %d), "
 676                              "forcing fsck on next reboot\n", group, err);
 677                 sbi->s_mount_state &= ~EXT3_VALID_FS;
 678                 sbi->s_es->s_state &= ~cpu_to_le16(EXT3_VALID_FS);
 679                 mark_buffer_dirty(sbi->s_sbh);
 680         }
 681 }
 682
 683 /* Add group descriptor data to an existing or new group descriptor block.
 684  * Ensure we handle all possible error conditions _before_ we start modifying
 685  * the filesystem, because we cannot abort the transaction and not have it
 686  * write the data to disk.
 687  *
 688  * If we are on a GDT block boundary, we need to get the reserved GDT block.
 689  * Otherwise, we may need to add backup GDT blocks for a sparse group.
 690  *
 691  * We only need to hold the superblock lock while we are actually adding
 692  * in the new group's counts to the superblock.  Prior to that we have
 693  * not really "added" the group at all.  We re-check that we are still
 694  * adding in the last group in case things have changed since verifying.
 695  */
 696 int ext3_group_add(struct super_block *sb, struct ext3_new_group_data *input)
 697 {
 698         struct ext3_sb_info *sbi = EXT3_SB(sb);
 699         struct ext3_super_block *es = sbi->s_es;
 700         int reserved_gdb = ext3_bg_has_super(sb, input->group) ?
 701                 le16_to_cpu(es->s_reserved_gdt_blocks) : 0;
 702         struct buffer_head *primary = NULL;
 703         struct ext3_group_desc *gdp;
 704         struct inode *inode = NULL;
 705         handle_t *handle;
 706         int gdb_off, gdb_num;
 707         int err, err2;
 708
 709         gdb_num = input->group / EXT3_DESC_PER_BLOCK(sb);
 710         gdb_off = input->group % EXT3_DESC_PER_BLOCK(sb);
 711
 712         if (gdb_off == 0 && !EXT3_HAS_RO_COMPAT_FEATURE(sb,
 713                                         EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER)) {
 714                 ext3_warning(sb, __FUNCTION__,
 715                              "Can't resize non-sparse filesystem further\n");
 716                 return -EPERM;
 717         }
 718
 719         if (reserved_gdb || gdb_off == 0) {
 720                 if (!EXT3_HAS_COMPAT_FEATURE(sb,
 721                                              EXT3_FEATURE_COMPAT_RESIZE_INODE)){
 722                         ext3_warning(sb, __FUNCTION__,
 723                                      "No reserved GDT blocks, can't resize\n");
 724                         return -EPERM;
 725                 }
 726                 inode = iget(sb, EXT3_RESIZE_INO);
 727                 if (!inode || is_bad_inode(inode)) {
 728                         ext3_warning(sb, __FUNCTION__,
 729                                      "Error opening resize inode\n");
 730                         iput(inode);
 731                         return -ENOENT;
 732                 }
 733         }
 734
 735         if ((err = verify_group_input(sb, input)))
 736                 goto exit_put;
 737
 738         if ((err = setup_new_group_blocks(sb, input)))
 739                 goto exit_put;
 740
 741         /*
 742          * We will always be modifying at least the superblock and a GDT
 743          * block.  If we are adding a group past the last current GDT block,
 744          * we will also modify the inode and the dindirect block.  If we
 745          * are adding a group with superblock/GDT backups  we will also
 746          * modify each of the reserved GDT dindirect blocks.
 747          */
 748         handle = ext3_journal_start_sb(sb,
 749                                        ext3_bg_has_super(sb, input->group) ?
 750                                        3 + reserved_gdb : 4);
 751         if (IS_ERR(handle)) {
 752                 err = PTR_ERR(handle);
 753                 goto exit_put;
 754         }
 755
 756         lock_super(sb);
 757         if (input->group != EXT3_SB(sb)->s_groups_count) {
 758                 ext3_warning(sb, __FUNCTION__,
 759                              "multiple resizers run on filesystem!\n");
 760                 goto exit_journal;
 761         }
 762
 763         if ((err = ext3_journal_get_write_access(handle, sbi->s_sbh)))
 764                 goto exit_journal;
 765
 766         /*
 767          * We will only either add reserved group blocks to a backup group
 768          * or remove reserved blocks for the first group in a new group block.
 769          * Doing both would be mean more complex code, and sane people don't
 770          * use non-sparse filesystems anymore.  This is already checked above.
 771          */
 772         if (gdb_off) {
 773                 primary = sbi->s_group_desc[gdb_num];
 774                 if ((err = ext3_journal_get_write_access(handle, primary)))
 775                         goto exit_journal;
 776
 777                 if (reserved_gdb && ext3_bg_num_gdb(sb, input->group) &&
 778                     (err = reserve_backup_gdb(handle, inode, input)))
 779                         goto exit_journal;
 780         } else if ((err = add_new_gdb(handle, inode, input, &primary)))
 781                 goto exit_journal;
 782
 783         /*
 784          * OK, now we've set up the new group.  Time to make it active.
 785          *
 786          * Current kernels don't lock all allocations via lock_super(),
 787          * so we have to be safe wrt. concurrent accesses the group
 788          * data.  So we need to be careful to set all of the relevant
 789          * group descriptor data etc. *before* we enable the group.
 790          *
 791          * The key field here is EXT3_SB(sb)->s_groups_count: as long as
 792          * that retains its old value, nobody is going to access the new
 793          * group.
 794          *
 795          * So first we update all the descriptor metadata for the new
 796          * group; then we update the total disk blocks count; then we
 797          * update the groups count to enable the group; then finally we
 798          * update the free space counts so that the system can start
 799          * using the new disk blocks.
 800          */
 801
 802         /* Update group descriptor block for new group */
 803         gdp = (struct ext3_group_desc *)primary->b_data + gdb_off;
 804
 805         gdp->bg_block_bitmap = cpu_to_le32(input->block_bitmap);
 806         gdp->bg_inode_bitmap = cpu_to_le32(input->inode_bitmap);
 807         gdp->bg_inode_table = cpu_to_le32(input->inode_table);
 808         gdp->bg_free_blocks_count = cpu_to_le16(input->free_blocks_count);
 809         gdp->bg_free_inodes_count = cpu_to_le16(EXT3_INODES_PER_GROUP(sb));
 810
 811         /*
 812          * Make the new blocks and inodes valid next.  We do this before
 813          * increasing the group count so that once the group is enabled,
 814          * all of its blocks and inodes are already valid.
 815          *
 816          * We always allocate group-by-group, then block-by-block or
 817          * inode-by-inode within a group, so enabling these
 818          * blocks/inodes before the group is live won't actually let us
 819          * allocate the new space yet.
 820          */
 821         es->s_blocks_count = cpu_to_le32(le32_to_cpu(es->s_blocks_count) +
 822                 input->blocks_count);
 823         es->s_inodes_count = cpu_to_le32(le32_to_cpu(es->s_inodes_count) +
 824                 EXT3_INODES_PER_GROUP(sb));
 825
 826         /*
 827          * We need to protect s_groups_count against other CPUs seeing
 828          * inconsistent state in the superblock.
 829          *
 830          * The precise rules we use are:
 831          *
 832          * * Writers of s_groups_count *must* hold lock_super
 833          * AND
 834          * * Writers must perform a smp_wmb() after updating all dependent
 835          *   data and before modifying the groups count
 836          *
 837          * * Readers must hold lock_super() over the access
 838          * OR
 839          * * Readers must perform an smp_rmb() after reading the groups count
 840          *   and before reading any dependent data.
 841          *
 842          * NB. These rules can be relaxed when checking the group count
 843          * while freeing data, as we can only allocate from a block
 844          * group after serialising against the group count, and we can
 845          * only then free after serialising in turn against that
 846          * allocation.
 847          */
 848         smp_wmb();
 849
 850         /* Update the global fs size fields */
 851         EXT3_SB(sb)->s_groups_count++;
 852
 853         ext3_journal_dirty_metadata(handle, primary);
 854
 855         /* Update the reserved block counts only once the new group is
 856          * active. */
 857         es->s_r_blocks_count = cpu_to_le32(le32_to_cpu(es->s_r_blocks_count) +
 858                 input->reserved_blocks);
 859
 860         /* Update the free space counts */
 861         percpu_counter_mod(&sbi->s_freeblocks_counter,
 862                            input->free_blocks_count);
 863         percpu_counter_mod(&sbi->s_freeinodes_counter,
 864                            EXT3_INODES_PER_GROUP(sb));
 865
 866         ext3_journal_dirty_metadata(handle, EXT3_SB(sb)->s_sbh);
 867         sb->s_dirt = 1;
 868
 869 exit_journal:
 870         unlock_super(sb);
 871         if ((err2 = ext3_journal_stop(handle)) && !err)
 872                 err = err2;
 873         if (!err) {
 874                 update_backups(sb, sbi->s_sbh->b_blocknr, (char *)es,
 875                                sizeof(struct ext3_super_block));
 876                 update_backups(sb, primary->b_blocknr, primary->b_data,
 877                                primary->b_size);
 878         }
 879 exit_put:
 880         iput(inode);
 881         return err;
 882 } /* ext3_group_add */
 883
 884 /* Extend the filesystem to the new number of blocks specified.  This entry
 885  * point is only used to extend the current filesystem to the end of the last
 886  * existing group.  It can be accessed via ioctl, or by "remount,resize=<size>"
 887  * for emergencies (because it has no dependencies on reserved blocks).
 888  *
 889  * If we _really_ wanted, we could use default values to call ext3_group_add()
 890  * allow the "remount" trick to work for arbitrary resizing, assuming enough
 891  * GDT blocks are reserved to grow to the desired size.
 892  */
 893 int ext3_group_extend(struct super_block *sb, struct ext3_super_block *es,
 894                       unsigned long n_blocks_count)
 895 {
 896         unsigned long o_blocks_count;
 897         unsigned long o_groups_count;
 898         unsigned long last;
 899         int add;
 900         struct buffer_head * bh;
 901         handle_t *handle;
 902         int err, freed_blocks;
 903
 904         /* We don't need to worry about locking wrt other resizers just
 905          * yet: we're going to revalidate es->s_blocks_count after
 906          * taking lock_super() below. */
 907         o_blocks_count = le32_to_cpu(es->s_blocks_count);
 908         o_groups_count = EXT3_SB(sb)->s_groups_count;
 909
 910         if (test_opt(sb, DEBUG))
 911                 printk(KERN_DEBUG "EXT3-fs: extending last group from %lu to %lu blocks\n",
 912                        o_blocks_count, n_blocks_count);
 913
 914         if (n_blocks_count == 0 || n_blocks_count == o_blocks_count)
 915                 return 0;
 916
 917         if (n_blocks_count < o_blocks_count) {
 918                 ext3_warning(sb, __FUNCTION__,
 919                              "can't shrink FS - resize aborted");
 920                 return -EBUSY;
 921         }
 922
 923         /* Handle the remaining blocks in the last group only. */
 924         last = (o_blocks_count - le32_to_cpu(es->s_first_data_block)) %
 925                 EXT3_BLOCKS_PER_GROUP(sb);
 926
 927         if (last == 0) {
 928                 ext3_warning(sb, __FUNCTION__,
 929                              "need to use ext2online to resize further\n");
 930                 return -EPERM;
 931         }
 932
 933         add = EXT3_BLOCKS_PER_GROUP(sb) - last;
 934
 935         if (o_blocks_count + add > n_blocks_count)
 936                 add = n_blocks_count - o_blocks_count;
 937
 938         if (o_blocks_count + add < n_blocks_count)
 939                 ext3_warning(sb, __FUNCTION__,
 940                              "will only finish group (%lu blocks, %u new)",
 941                              o_blocks_count + add, add);
 942
 943         /* See if the device is actually as big as what was requested */
 944         bh = sb_bread(sb, o_blocks_count + add -1);
 945         if (!bh) {
 946                 ext3_warning(sb, __FUNCTION__,
 947                              "can't read last block, resize aborted");
 948                 return -ENOSPC;
 949         }
 950         brelse(bh);
 951
 952         /* We will update the superblock, one block bitmap, and
 953          * one group descriptor via ext3_free_blocks().
 954          */
 955         handle = ext3_journal_start_sb(sb, 3);
 956         if (IS_ERR(handle)) {
 957                 err = PTR_ERR(handle);
 958                 ext3_warning(sb, __FUNCTION__, "error %d on journal start",err);
 959                 goto exit_put;
 960         }
 961
 962         lock_super(sb);
 963         if (o_blocks_count != le32_to_cpu(es->s_blocks_count)) {
 964                 ext3_warning(sb, __FUNCTION__,
 965                              "multiple resizers run on filesystem!\n");
 966                 err = -EBUSY;
 967                 goto exit_put;
 968         }
 969
 970         if ((err = ext3_journal_get_write_access(handle,
 971                                                  EXT3_SB(sb)->s_sbh))) {
 972                 ext3_warning(sb, __FUNCTION__,
 973                              "error %d on journal write access", err);
 974                 unlock_super(sb);
 975                 ext3_journal_stop(handle);
 976                 goto exit_put;
 977         }
 978         es->s_blocks_count = cpu_to_le32(o_blocks_count + add);
 979         ext3_journal_dirty_metadata(handle, EXT3_SB(sb)->s_sbh);
 980         sb->s_dirt = 1;
 981         unlock_super(sb);
 982         ext3_debug("freeing blocks %ld through %ld\n", o_blocks_count,
 983                    o_blocks_count + add);
 984         ext3_free_blocks_sb(handle, sb, o_blocks_count, add, &freed_blocks);
 985         ext3_debug("freed blocks %ld through %ld\n", o_blocks_count,
 986                    o_blocks_count + add);
 987         if ((err = ext3_journal_stop(handle)))
 988                 goto exit_put;
 989         if (test_opt(sb, DEBUG))
 990                 printk(KERN_DEBUG "EXT3-fs: extended group to %u blocks\n",
 991                        le32_to_cpu(es->s_blocks_count));
 992         update_backups(sb, EXT3_SB(sb)->s_sbh->b_blocknr, (char *)es,
 993                        sizeof(struct ext3_super_block));
 994 exit_put:
 995         return err;
 996 } /* ext3_group_extend */