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
  12 #define EXT3FS_DEBUG
  13
  14 #include <linux/ext3_jbd.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 ext3_new_group_data *input)
  25 {
  26         struct ext3_sb_info *sbi = EXT3_SB(sb);
  27         struct ext3_super_block *es = sbi->s_es;
  28         ext3_fsblk_t start = le32_to_cpu(es->s_blocks_count);
  29         ext3_fsblk_t end = start + input->blocks_count;
  30         unsigned group = input->group;
  31         ext3_fsblk_t itend = input->inode_table + sbi->s_itb_per_group;
  32         unsigned overhead = ext3_bg_has_super(sb, group) ?
  33                 (1 + ext3_bg_num_gdb(sb, group) +
  34                  le16_to_cpu(es->s_reserved_gdt_blocks)) : 0;
  35         ext3_fsblk_t metaend = start + overhead;
  36         struct buffer_head *bh = NULL;
  37         ext3_grpblk_t free_blocks_count;
  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 "EXT3-fs: adding %s group %u: %u blocks "
  45                        "(%d free, %u reserved)\n",
  46                        ext3_bg_has_super(sb, input->group) ? "normal" :
  47                        "no-super", input->group, input->blocks_count,
  48                        free_blocks_count, input->reserved_blocks);
  49
  50         if (group != sbi->s_groups_count)
  51                 ext3_warning(sb, __func__,
  52                              "Cannot add at group %u (only %lu groups)",
  53                              input->group, sbi->s_groups_count);
  54         else if ((start - le32_to_cpu(es->s_first_data_block)) %
  55                  EXT3_BLOCKS_PER_GROUP(sb))
  56                 ext3_warning(sb, __func__, "Last group not full");
  57         else if (input->reserved_blocks > input->blocks_count / 5)
  58                 ext3_warning(sb, __func__, "Reserved blocks too high (%u)",
  59                              input->reserved_blocks);
  60         else if (free_blocks_count < 0)
  61                 ext3_warning(sb, __func__, "Bad blocks count %u",
  62                              input->blocks_count);
  63         else if (!(bh = sb_bread(sb, end - 1)))
  64                 ext3_warning(sb, __func__,
  65                              "Cannot read last block ("E3FSBLK")",
  66                              end - 1);
  67         else if (outside(input->block_bitmap, start, end))
  68                 ext3_warning(sb, __func__,
  69                              "Block bitmap not in group (block %u)",
  70                              input->block_bitmap);
  71         else if (outside(input->inode_bitmap, start, end))
  72                 ext3_warning(sb, __func__,
  73                              "Inode bitmap not in group (block %u)",
  74                              input->inode_bitmap);
  75         else if (outside(input->inode_table, start, end) ||
  76                  outside(itend - 1, start, end))
  77                 ext3_warning(sb, __func__,
  78                              "Inode table not in group (blocks %u-"E3FSBLK")",
  79                              input->inode_table, itend - 1);
  80         else if (input->inode_bitmap == input->block_bitmap)
  81                 ext3_warning(sb, __func__,
  82                              "Block bitmap same as inode bitmap (%u)",
  83                              input->block_bitmap);
  84         else if (inside(input->block_bitmap, input->inode_table, itend))
  85                 ext3_warning(sb, __func__,
  86                              "Block bitmap (%u) in inode table (%u-"E3FSBLK")",
  87                              input->block_bitmap, input->inode_table, itend-1);
  88         else if (inside(input->inode_bitmap, input->inode_table, itend))
  89                 ext3_warning(sb, __func__,
  90                              "Inode bitmap (%u) in inode table (%u-"E3FSBLK")",
  91                              input->inode_bitmap, input->inode_table, itend-1);
  92         else if (inside(input->block_bitmap, start, metaend))
  93                 ext3_warning(sb, __func__,
  94                              "Block bitmap (%u) in GDT table"
  95                              " ("E3FSBLK"-"E3FSBLK")",
  96                              input->block_bitmap, start, metaend - 1);
  97         else if (inside(input->inode_bitmap, start, metaend))
  98                 ext3_warning(sb, __func__,
  99                              "Inode bitmap (%u) in GDT table"
 100                              " ("E3FSBLK"-"E3FSBLK")",
 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, __func__,
 105                              "Inode table (%u-"E3FSBLK") overlaps"
 106                              "GDT table ("E3FSBLK"-"E3FSBLK")",
 107                              input->inode_table, itend - 1, start, metaend - 1);
 108         else
 109                 err = 0;
 110         brelse(bh);
 111
 112         return err;
 113 }
 114
 115 static struct buffer_head *bclean(handle_t *handle, struct super_block *sb,
 116                                   ext3_fsblk_t blk)
 117 {
 118         struct buffer_head *bh;
 119         int err;
 120
 121         bh = sb_getblk(sb, blk);
 122         if (!bh)
 123                 return ERR_PTR(-EIO);
 124         if ((err = ext3_journal_get_write_access(handle, bh))) {
 125                 brelse(bh);
 126                 bh = ERR_PTR(err);
 127         } else {
 128                 lock_buffer(bh);
 129                 memset(bh->b_data, 0, sb->s_blocksize);
 130                 set_buffer_uptodate(bh);
 131                 unlock_buffer(bh);
 132         }
 133
 134         return bh;
 135 }
 136
 137 /*
 138  * To avoid calling the atomic setbit hundreds or thousands of times, we only
 139  * need to use it within a single byte (to ensure we get endianness right).
 140  * We can use memset for the rest of the bitmap as there are no other users.
 141  */
 142 static void mark_bitmap_end(int start_bit, int end_bit, char *bitmap)
 143 {
 144         int i;
 145
 146         if (start_bit >= end_bit)
 147                 return;
 148
 149         ext3_debug("mark end bits +%d through +%d used\n", start_bit, end_bit);
 150         for (i = start_bit; i < ((start_bit + 7) & ~7UL); i++)
 151                 ext3_set_bit(i, bitmap);
 152         if (i < end_bit)
 153                 memset(bitmap + (i >> 3), 0xff, (end_bit - i) >> 3);
 154 }
 155
 156 /*
 157  * If we have fewer than thresh credits, extend by EXT3_MAX_TRANS_DATA.
 158  * If that fails, restart the transaction & regain write access for the
 159  * buffer head which is used for block_bitmap modifications.
 160  */
 161 static int extend_or_restart_transaction(handle_t *handle, int thresh,
 162                                          struct buffer_head *bh)
 163 {
 164         int err;
 165
 166         if (handle->h_buffer_credits >= thresh)
 167                 return 0;
 168
 169         err = ext3_journal_extend(handle, EXT3_MAX_TRANS_DATA);
 170         if (err < 0)
 171                 return err;
 172         if (err) {
 173                 err = ext3_journal_restart(handle, EXT3_MAX_TRANS_DATA);
 174                 if (err)
 175                         return err;
 176                 err = ext3_journal_get_write_access(handle, bh);
 177                 if (err)
 178                         return err;
 179         }
 180
 181         return 0;
 182 }
 183
 184 /*
 185  * Set up the block and inode bitmaps, and the inode table for the new group.
 186  * This doesn't need to be part of the main transaction, since we are only
 187  * changing blocks outside the actual filesystem.  We still do journaling to
 188  * ensure the recovery is correct in case of a failure just after resize.
 189  * If any part of this fails, we simply abort the resize.
 190  */
 191 static int setup_new_group_blocks(struct super_block *sb,
 192                                   struct ext3_new_group_data *input)
 193 {
 194         struct ext3_sb_info *sbi = EXT3_SB(sb);
 195         ext3_fsblk_t start = ext3_group_first_block_no(sb, input->group);
 196         int reserved_gdb = ext3_bg_has_super(sb, input->group) ?
 197                 le16_to_cpu(sbi->s_es->s_reserved_gdt_blocks) : 0;
 198         unsigned long gdblocks = ext3_bg_num_gdb(sb, input->group);
 199         struct buffer_head *bh;
 200         handle_t *handle;
 201         ext3_fsblk_t block;
 202         ext3_grpblk_t bit;
 203         int i;
 204         int err = 0, err2;
 205
 206         /* This transaction may be extended/restarted along the way */
 207         handle = ext3_journal_start_sb(sb, EXT3_MAX_TRANS_DATA);
 208
 209         if (IS_ERR(handle))
 210                 return PTR_ERR(handle);
 211
 212         mutex_lock(&sbi->s_resize_lock);
 213         if (input->group != sbi->s_groups_count) {
 214                 err = -EBUSY;
 215                 goto exit_journal;
 216         }
 217
 218         if (IS_ERR(bh = bclean(handle, sb, input->block_bitmap))) {
 219                 err = PTR_ERR(bh);
 220                 goto exit_journal;
 221         }
 222
 223         if (ext3_bg_has_super(sb, input->group)) {
 224                 ext3_debug("mark backup superblock %#04lx (+0)\n", start);
 225                 ext3_set_bit(0, bh->b_data);
 226         }
 227
 228         /* Copy all of the GDT blocks into the backup in this group */
 229         for (i = 0, bit = 1, block = start + 1;
 230              i < gdblocks; i++, block++, bit++) {
 231                 struct buffer_head *gdb;
 232
 233                 ext3_debug("update backup group %#04lx (+%d)\n", block, bit);
 234
 235                 err = extend_or_restart_transaction(handle, 1, bh);
 236                 if (err)
 237                         goto exit_bh;
 238
 239                 gdb = sb_getblk(sb, block);
 240                 if (!gdb) {
 241                         err = -EIO;
 242                         goto exit_bh;
 243                 }
 244                 if ((err = ext3_journal_get_write_access(handle, gdb))) {
 245                         brelse(gdb);
 246                         goto exit_bh;
 247                 }
 248                 lock_buffer(gdb);
 249                 memcpy(gdb->b_data, sbi->s_group_desc[i]->b_data, gdb->b_size);
 250                 set_buffer_uptodate(gdb);
 251                 unlock_buffer(gdb);
 252                 err = ext3_journal_dirty_metadata(handle, gdb);
 253                 if (err) {
 254                         brelse(gdb);
 255                         goto exit_bh;
 256                 }
 257                 ext3_set_bit(bit, bh->b_data);
 258                 brelse(gdb);
 259         }
 260
 261         /* Zero out all of the reserved backup group descriptor table blocks */
 262         for (i = 0, bit = gdblocks + 1, block = start + bit;
 263              i < reserved_gdb; i++, block++, bit++) {
 264                 struct buffer_head *gdb;
 265
 266                 ext3_debug("clear reserved block %#04lx (+%d)\n", block, bit);
 267
 268                 err = extend_or_restart_transaction(handle, 1, bh);
 269                 if (err)
 270                         goto exit_bh;
 271
 272                 if (IS_ERR(gdb = bclean(handle, sb, block))) {
 273                         err = PTR_ERR(gdb);
 274                         goto exit_bh;
 275                 }
 276                 err = ext3_journal_dirty_metadata(handle, gdb);
 277                 if (err) {
 278                         brelse(gdb);
 279                         goto exit_bh;
 280                 }
 281                 ext3_set_bit(bit, bh->b_data);
 282                 brelse(gdb);
 283         }
 284         ext3_debug("mark block bitmap %#04x (+%ld)\n", input->block_bitmap,
 285                    input->block_bitmap - start);
 286         ext3_set_bit(input->block_bitmap - start, bh->b_data);
 287         ext3_debug("mark inode bitmap %#04x (+%ld)\n", input->inode_bitmap,
 288                    input->inode_bitmap - start);
 289         ext3_set_bit(input->inode_bitmap - start, bh->b_data);
 290
 291         /* Zero out all of the inode table blocks */
 292         for (i = 0, block = input->inode_table, bit = block - start;
 293              i < sbi->s_itb_per_group; i++, bit++, block++) {
 294                 struct buffer_head *it;
 295
 296                 ext3_debug("clear inode block %#04lx (+%d)\n", block, bit);
 297
 298                 err = extend_or_restart_transaction(handle, 1, bh);
 299                 if (err)
 300                         goto exit_bh;
 301
 302                 if (IS_ERR(it = bclean(handle, sb, block))) {
 303                         err = PTR_ERR(it);
 304                         goto exit_bh;
 305                 }
 306                 err = ext3_journal_dirty_metadata(handle, it);
 307                 if (err) {
 308                         brelse(it);
 309                         goto exit_bh;
 310                 }
 311                 brelse(it);
 312                 ext3_set_bit(bit, bh->b_data);
 313         }
 314
 315         err = extend_or_restart_transaction(handle, 2, bh);
 316         if (err)
 317                 goto exit_bh;
 318
 319         mark_bitmap_end(input->blocks_count, EXT3_BLOCKS_PER_GROUP(sb),
 320                         bh->b_data);
 321         err = ext3_journal_dirty_metadata(handle, bh);
 322         if (err)
 323                 goto exit_bh;
 324         brelse(bh);
 325
 326         /* Mark unused entries in inode bitmap used */
 327         ext3_debug("clear inode bitmap %#04x (+%ld)\n",
 328                    input->inode_bitmap, input->inode_bitmap - start);
 329         if (IS_ERR(bh = bclean(handle, sb, input->inode_bitmap))) {
 330                 err = PTR_ERR(bh);
 331                 goto exit_journal;
 332         }
 333
 334         mark_bitmap_end(EXT3_INODES_PER_GROUP(sb), EXT3_BLOCKS_PER_GROUP(sb),
 335                         bh->b_data);
 336         err = ext3_journal_dirty_metadata(handle, bh);
 337 exit_bh:
 338         brelse(bh);
 339
 340 exit_journal:
 341         mutex_unlock(&sbi->s_resize_lock);
 342         if ((err2 = ext3_journal_stop(handle)) && !err)
 343                 err = err2;
 344
 345         return err;
 346 }
 347
 348 /*
 349  * Iterate through the groups which hold BACKUP superblock/GDT copies in an
 350  * ext3 filesystem.  The counters should be initialized to 1, 5, and 7 before
 351  * calling this for the first time.  In a sparse filesystem it will be the
 352  * sequence of powers of 3, 5, and 7: 1, 3, 5, 7, 9, 25, 27, 49, 81, ...
 353  * For a non-sparse filesystem it will be every group: 1, 2, 3, 4, ...
 354  */
 355 static unsigned ext3_list_backups(struct super_block *sb, unsigned *three,
 356                                   unsigned *five, unsigned *seven)
 357 {
 358         unsigned *min = three;
 359         int mult = 3;
 360         unsigned ret;
 361
 362         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
 363                                         EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER)) {
 364                 ret = *min;
 365                 *min += 1;
 366                 return ret;
 367         }
 368
 369         if (*five < *min) {
 370                 min = five;
 371                 mult = 5;
 372         }
 373         if (*seven < *min) {
 374                 min = seven;
 375                 mult = 7;
 376         }
 377
 378         ret = *min;
 379         *min *= mult;
 380
 381         return ret;
 382 }
 383
 384 /*
 385  * Check that all of the backup GDT blocks are held in the primary GDT block.
 386  * It is assumed that they are stored in group order.  Returns the number of
 387  * groups in current filesystem that have BACKUPS, or -ve error code.
 388  */
 389 static int verify_reserved_gdb(struct super_block *sb,
 390                                struct buffer_head *primary)
 391 {
 392         const ext3_fsblk_t blk = primary->b_blocknr;
 393         const unsigned long end = EXT3_SB(sb)->s_groups_count;
 394         unsigned three = 1;
 395         unsigned five = 5;
 396         unsigned seven = 7;
 397         unsigned grp;
 398         __le32 *p = (__le32 *)primary->b_data;
 399         int gdbackups = 0;
 400
 401         while ((grp = ext3_list_backups(sb, &three, &five, &seven)) < end) {
 402                 if (le32_to_cpu(*p++) != grp * EXT3_BLOCKS_PER_GROUP(sb) + blk){
 403                         ext3_warning(sb, __func__,
 404                                      "reserved GDT "E3FSBLK
 405                                      " missing grp %d ("E3FSBLK")",
 406                                      blk, grp,
 407                                      grp * EXT3_BLOCKS_PER_GROUP(sb) + blk);
 408                         return -EINVAL;
 409                 }
 410                 if (++gdbackups > EXT3_ADDR_PER_BLOCK(sb))
 411                         return -EFBIG;
 412         }
 413
 414         return gdbackups;
 415 }
 416
 417 /*
 418  * Called when we need to bring a reserved group descriptor table block into
 419  * use from the resize inode.  The primary copy of the new GDT block currently
 420  * is an indirect block (under the double indirect block in the resize inode).
 421  * The new backup GDT blocks will be stored as leaf blocks in this indirect
 422  * block, in group order.  Even though we know all the block numbers we need,
 423  * we check to ensure that the resize inode has actually reserved these blocks.
 424  *
 425  * Don't need to update the block bitmaps because the blocks are still in use.
 426  *
 427  * We get all of the error cases out of the way, so that we are sure to not
 428  * fail once we start modifying the data on disk, because JBD has no rollback.
 429  */
 430 static int add_new_gdb(handle_t *handle, struct inode *inode,
 431                        struct ext3_new_group_data *input,
 432                        struct buffer_head **primary)
 433 {
 434         struct super_block *sb = inode->i_sb;
 435         struct ext3_super_block *es = EXT3_SB(sb)->s_es;
 436         unsigned long gdb_num = input->group / EXT3_DESC_PER_BLOCK(sb);
 437         ext3_fsblk_t gdblock = EXT3_SB(sb)->s_sbh->b_blocknr + 1 + gdb_num;
 438         struct buffer_head **o_group_desc, **n_group_desc;
 439         struct buffer_head *dind;
 440         int gdbackups;
 441         struct ext3_iloc iloc;
 442         __le32 *data;
 443         int err;
 444
 445         if (test_opt(sb, DEBUG))
 446                 printk(KERN_DEBUG
 447                        "EXT3-fs: ext3_add_new_gdb: adding group block %lu\n",
 448                        gdb_num);
 449
 450         /*
 451          * If we are not using the primary superblock/GDT copy don't resize,
 452          * because the user tools have no way of handling this.  Probably a
 453          * bad time to do it anyways.
 454          */
 455         if (EXT3_SB(sb)->s_sbh->b_blocknr !=
 456             le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block)) {
 457                 ext3_warning(sb, __func__,
 458                         "won't resize using backup superblock at %llu",
 459                         (unsigned long long)EXT3_SB(sb)->s_sbh->b_blocknr);
 460                 return -EPERM;
 461         }
 462
 463         *primary = sb_bread(sb, gdblock);
 464         if (!*primary)
 465                 return -EIO;
 466
 467         if ((gdbackups = verify_reserved_gdb(sb, *primary)) < 0) {
 468                 err = gdbackups;
 469                 goto exit_bh;
 470         }
 471
 472         data = EXT3_I(inode)->i_data + EXT3_DIND_BLOCK;
 473         dind = sb_bread(sb, le32_to_cpu(*data));
 474         if (!dind) {
 475                 err = -EIO;
 476                 goto exit_bh;
 477         }
 478
 479         data = (__le32 *)dind->b_data;
 480         if (le32_to_cpu(data[gdb_num % EXT3_ADDR_PER_BLOCK(sb)]) != gdblock) {
 481                 ext3_warning(sb, __func__,
 482                              "new group %u GDT block "E3FSBLK" not reserved",
 483                              input->group, gdblock);
 484                 err = -EINVAL;
 485                 goto exit_dind;
 486         }
 487
 488         if ((err = ext3_journal_get_write_access(handle, EXT3_SB(sb)->s_sbh)))
 489                 goto exit_dind;
 490
 491         if ((err = ext3_journal_get_write_access(handle, *primary)))
 492                 goto exit_sbh;
 493
 494         if ((err = ext3_journal_get_write_access(handle, dind)))
 495                 goto exit_primary;
 496
 497         /* ext3_reserve_inode_write() gets a reference on the iloc */
 498         if ((err = ext3_reserve_inode_write(handle, inode, &iloc)))
 499                 goto exit_dindj;
 500
 501         n_group_desc = kmalloc((gdb_num + 1) * sizeof(struct buffer_head *),
 502                         GFP_NOFS);
 503         if (!n_group_desc) {
 504                 err = -ENOMEM;
 505                 ext3_warning (sb, __func__,
 506                               "not enough memory for %lu groups", gdb_num + 1);
 507                 goto exit_inode;
 508         }
 509
 510         /*
 511          * Finally, we have all of the possible failures behind us...
 512          *
 513          * Remove new GDT block from inode double-indirect block and clear out
 514          * the new GDT block for use (which also "frees" the backup GDT blocks
 515          * from the reserved inode).  We don't need to change the bitmaps for
 516          * these blocks, because they are marked as in-use from being in the
 517          * reserved inode, and will become GDT blocks (primary and backup).
 518          */
 519         data[gdb_num % EXT3_ADDR_PER_BLOCK(sb)] = 0;
 520         err = ext3_journal_dirty_metadata(handle, dind);
 521         if (err)
 522                 goto exit_group_desc;
 523         brelse(dind);
 524         dind = NULL;
 525         inode->i_blocks -= (gdbackups + 1) * sb->s_blocksize >> 9;
 526         err = ext3_mark_iloc_dirty(handle, inode, &iloc);
 527         if (err)
 528                 goto exit_group_desc;
 529         memset((*primary)->b_data, 0, sb->s_blocksize);
 530         err = ext3_journal_dirty_metadata(handle, *primary);
 531         if (err)
 532                 goto exit_group_desc;
 533
 534         o_group_desc = EXT3_SB(sb)->s_group_desc;
 535         memcpy(n_group_desc, o_group_desc,
 536                EXT3_SB(sb)->s_gdb_count * sizeof(struct buffer_head *));
 537         n_group_desc[gdb_num] = *primary;
 538         EXT3_SB(sb)->s_group_desc = n_group_desc;
 539         EXT3_SB(sb)->s_gdb_count++;
 540         kfree(o_group_desc);
 541
 542         le16_add_cpu(&es->s_reserved_gdt_blocks, -1);
 543         err = ext3_journal_dirty_metadata(handle, EXT3_SB(sb)->s_sbh);
 544         if (err)
 545                 goto exit_inode;
 546
 547         return 0;
 548
 549 exit_group_desc:
 550         kfree(n_group_desc);
 551 exit_inode:
 552         //ext3_journal_release_buffer(handle, iloc.bh);
 553         brelse(iloc.bh);
 554 exit_dindj:
 555         //ext3_journal_release_buffer(handle, dind);
 556 exit_primary:
 557         //ext3_journal_release_buffer(handle, *primary);
 558 exit_sbh:
 559         //ext3_journal_release_buffer(handle, *primary);
 560 exit_dind:
 561         brelse(dind);
 562 exit_bh:
 563         brelse(*primary);
 564
 565         ext3_debug("leaving with error %d\n", err);
 566         return err;
 567 }
 568
 569 /*
 570  * Called when we are adding a new group which has a backup copy of each of
 571  * the GDT blocks (i.e. sparse group) and there are reserved GDT blocks.
 572  * We need to add these reserved backup GDT blocks to the resize inode, so
 573  * that they are kept for future resizing and not allocated to files.
 574  *
 575  * Each reserved backup GDT block will go into a different indirect block.
 576  * The indirect blocks are actually the primary reserved GDT blocks,
 577  * so we know in advance what their block numbers are.  We only get the
 578  * double-indirect block to verify it is pointing to the primary reserved
 579  * GDT blocks so we don't overwrite a data block by accident.  The reserved
 580  * backup GDT blocks are stored in their reserved primary GDT block.
 581  */
 582 static int reserve_backup_gdb(handle_t *handle, struct inode *inode,
 583                               struct ext3_new_group_data *input)
 584 {
 585         struct super_block *sb = inode->i_sb;
 586         int reserved_gdb =le16_to_cpu(EXT3_SB(sb)->s_es->s_reserved_gdt_blocks);
 587         struct buffer_head **primary;
 588         struct buffer_head *dind;
 589         struct ext3_iloc iloc;
 590         ext3_fsblk_t blk;
 591         __le32 *data, *end;
 592         int gdbackups = 0;
 593         int res, i;
 594         int err;
 595
 596         primary = kmalloc(reserved_gdb * sizeof(*primary), GFP_NOFS);
 597         if (!primary)
 598                 return -ENOMEM;
 599
 600         data = EXT3_I(inode)->i_data + EXT3_DIND_BLOCK;
 601         dind = sb_bread(sb, le32_to_cpu(*data));
 602         if (!dind) {
 603                 err = -EIO;
 604                 goto exit_free;
 605         }
 606
 607         blk = EXT3_SB(sb)->s_sbh->b_blocknr + 1 + EXT3_SB(sb)->s_gdb_count;
 608         data = (__le32 *)dind->b_data + (EXT3_SB(sb)->s_gdb_count %
 609                                          EXT3_ADDR_PER_BLOCK(sb));
 610         end = (__le32 *)dind->b_data + EXT3_ADDR_PER_BLOCK(sb);
 611
 612         /* Get each reserved primary GDT block and verify it holds backups */
 613         for (res = 0; res < reserved_gdb; res++, blk++) {
 614                 if (le32_to_cpu(*data) != blk) {
 615                         ext3_warning(sb, __func__,
 616                                      "reserved block "E3FSBLK
 617                                      " not at offset %ld",
 618                                      blk,
 619                                      (long)(data - (__le32 *)dind->b_data));
 620                         err = -EINVAL;
 621                         goto exit_bh;
 622                 }
 623                 primary[res] = sb_bread(sb, blk);
 624                 if (!primary[res]) {
 625                         err = -EIO;
 626                         goto exit_bh;
 627                 }
 628                 if ((gdbackups = verify_reserved_gdb(sb, primary[res])) < 0) {
 629                         brelse(primary[res]);
 630                         err = gdbackups;
 631                         goto exit_bh;
 632                 }
 633                 if (++data >= end)
 634                         data = (__le32 *)dind->b_data;
 635         }
 636
 637         for (i = 0; i < reserved_gdb; i++) {
 638                 if ((err = ext3_journal_get_write_access(handle, primary[i]))) {
 639                         /*
 640                         int j;
 641                         for (j = 0; j < i; j++)
 642                                 ext3_journal_release_buffer(handle, primary[j]);
 643                          */
 644                         goto exit_bh;
 645                 }
 646         }
 647
 648         if ((err = ext3_reserve_inode_write(handle, inode, &iloc)))
 649                 goto exit_bh;
 650
 651         /*
 652          * Finally we can add each of the reserved backup GDT blocks from
 653          * the new group to its reserved primary GDT block.
 654          */
 655         blk = input->group * EXT3_BLOCKS_PER_GROUP(sb);
 656         for (i = 0; i < reserved_gdb; i++) {
 657                 int err2;
 658                 data = (__le32 *)primary[i]->b_data;
 659                 /* printk("reserving backup %lu[%u] = %lu\n",
 660                        primary[i]->b_blocknr, gdbackups,
 661                        blk + primary[i]->b_blocknr); */
 662                 data[gdbackups] = cpu_to_le32(blk + primary[i]->b_blocknr);
 663                 err2 = ext3_journal_dirty_metadata(handle, primary[i]);
 664                 if (!err)
 665                         err = err2;
 666         }
 667         inode->i_blocks += reserved_gdb * sb->s_blocksize >> 9;
 668         ext3_mark_iloc_dirty(handle, inode, &iloc);
 669
 670 exit_bh:
 671         while (--res >= 0)
 672                 brelse(primary[res]);
 673         brelse(dind);
 674
 675 exit_free:
 676         kfree(primary);
 677
 678         return err;
 679 }
 680
 681 /*
 682  * Update the backup copies of the ext3 metadata.  These don't need to be part
 683  * of the main resize transaction, because e2fsck will re-write them if there
 684  * is a problem (basically only OOM will cause a problem).  However, we
 685  * _should_ update the backups if possible, in case the primary gets trashed
 686  * for some reason and we need to run e2fsck from a backup superblock.  The
 687  * important part is that the new block and inode counts are in the backup
 688  * superblocks, and the location of the new group metadata in the GDT backups.
 689  *
 690  * We do not need take the s_resize_lock for this, because these
 691  * blocks are not otherwise touched by the filesystem code when it is
 692  * mounted.  We don't need to worry about last changing from
 693  * sbi->s_groups_count, because the worst that can happen is that we
 694  * do not copy the full number of backups at this time.  The resize
 695  * which changed s_groups_count will backup again.
 696  */
 697 static void update_backups(struct super_block *sb,
 698                            int blk_off, char *data, int size)
 699 {
 700         struct ext3_sb_info *sbi = EXT3_SB(sb);
 701         const unsigned long last = sbi->s_groups_count;
 702         const int bpg = EXT3_BLOCKS_PER_GROUP(sb);
 703         unsigned three = 1;
 704         unsigned five = 5;
 705         unsigned seven = 7;
 706         unsigned group;
 707         int rest = sb->s_blocksize - size;
 708         handle_t *handle;
 709         int err = 0, err2;
 710
 711         handle = ext3_journal_start_sb(sb, EXT3_MAX_TRANS_DATA);
 712         if (IS_ERR(handle)) {
 713                 group = 1;
 714                 err = PTR_ERR(handle);
 715                 goto exit_err;
 716         }
 717
 718         while ((group = ext3_list_backups(sb, &three, &five, &seven)) < last) {
 719                 struct buffer_head *bh;
 720
 721                 /* Out of journal space, and can't get more - abort - so sad */
 722                 if (handle->h_buffer_credits == 0 &&
 723                     ext3_journal_extend(handle, EXT3_MAX_TRANS_DATA) &&
 724                     (err = ext3_journal_restart(handle, EXT3_MAX_TRANS_DATA)))
 725                         break;
 726
 727                 bh = sb_getblk(sb, group * bpg + blk_off);
 728                 if (!bh) {
 729                         err = -EIO;
 730                         break;
 731                 }
 732                 ext3_debug("update metadata backup %#04lx\n",
 733                           (unsigned long)bh->b_blocknr);
 734                 if ((err = ext3_journal_get_write_access(handle, bh))) {
 735                         brelse(bh);
 736                         break;
 737                 }
 738                 lock_buffer(bh);
 739                 memcpy(bh->b_data, data, size);
 740                 if (rest)
 741                         memset(bh->b_data + size, 0, rest);
 742                 set_buffer_uptodate(bh);
 743                 unlock_buffer(bh);
 744                 err = ext3_journal_dirty_metadata(handle, bh);
 745                 brelse(bh);
 746                 if (err)
 747                         break;
 748         }
 749         if ((err2 = ext3_journal_stop(handle)) && !err)
 750                 err = err2;
 751
 752         /*
 753          * Ugh! Need to have e2fsck write the backup copies.  It is too
 754          * late to revert the resize, we shouldn't fail just because of
 755          * the backup copies (they are only needed in case of corruption).
 756          *
 757          * However, if we got here we have a journal problem too, so we
 758          * can't really start a transaction to mark the superblock.
 759          * Chicken out and just set the flag on the hope it will be written
 760          * to disk, and if not - we will simply wait until next fsck.
 761          */
 762 exit_err:
 763         if (err) {
 764                 ext3_warning(sb, __func__,
 765                              "can't update backup for group %d (err %d), "
 766                              "forcing fsck on next reboot", group, err);
 767                 sbi->s_mount_state &= ~EXT3_VALID_FS;
 768                 sbi->s_es->s_state &= cpu_to_le16(~EXT3_VALID_FS);
 769                 mark_buffer_dirty(sbi->s_sbh);
 770         }
 771 }
 772
 773 /* Add group descriptor data to an existing or new group descriptor block.
 774  * Ensure we handle all possible error conditions _before_ we start modifying
 775  * the filesystem, because we cannot abort the transaction and not have it
 776  * write the data to disk.
 777  *
 778  * If we are on a GDT block boundary, we need to get the reserved GDT block.
 779  * Otherwise, we may need to add backup GDT blocks for a sparse group.
 780  *
 781  * We only need to hold the superblock lock while we are actually adding
 782  * in the new group's counts to the superblock.  Prior to that we have
 783  * not really "added" the group at all.  We re-check that we are still
 784  * adding in the last group in case things have changed since verifying.
 785  */
 786 int ext3_group_add(struct super_block *sb, struct ext3_new_group_data *input)
 787 {
 788         struct ext3_sb_info *sbi = EXT3_SB(sb);
 789         struct ext3_super_block *es = sbi->s_es;
 790         int reserved_gdb = ext3_bg_has_super(sb, input->group) ?
 791                 le16_to_cpu(es->s_reserved_gdt_blocks) : 0;
 792         struct buffer_head *primary = NULL;
 793         struct ext3_group_desc *gdp;
 794         struct inode *inode = NULL;
 795         handle_t *handle;
 796         int gdb_off, gdb_num;
 797         int err, err2;
 798
 799         gdb_num = input->group / EXT3_DESC_PER_BLOCK(sb);
 800         gdb_off = input->group % EXT3_DESC_PER_BLOCK(sb);
 801
 802         if (gdb_off == 0 && !EXT3_HAS_RO_COMPAT_FEATURE(sb,
 803                                         EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER)) {
 804                 ext3_warning(sb, __func__,
 805                              "Can't resize non-sparse filesystem further");
 806                 return -EPERM;
 807         }
 808
 809         if (le32_to_cpu(es->s_blocks_count) + input->blocks_count <
 810             le32_to_cpu(es->s_blocks_count)) {
 811                 ext3_warning(sb, __func__, "blocks_count overflow\n");
 812                 return -EINVAL;
 813         }
 814
 815         if (le32_to_cpu(es->s_inodes_count) + EXT3_INODES_PER_GROUP(sb) <
 816             le32_to_cpu(es->s_inodes_count)) {
 817                 ext3_warning(sb, __func__, "inodes_count overflow\n");
 818                 return -EINVAL;
 819         }
 820
 821         if (reserved_gdb || gdb_off == 0) {
 822                 if (!EXT3_HAS_COMPAT_FEATURE(sb,
 823                                              EXT3_FEATURE_COMPAT_RESIZE_INODE)
 824                     || !le16_to_cpu(es->s_reserved_gdt_blocks)) {
 825                         ext3_warning(sb, __func__,
 826                                      "No reserved GDT blocks, can't resize");
 827                         return -EPERM;
 828                 }
 829                 inode = ext3_iget(sb, EXT3_RESIZE_INO);
 830                 if (IS_ERR(inode)) {
 831                         ext3_warning(sb, __func__,
 832                                      "Error opening resize inode");
 833                         return PTR_ERR(inode);
 834                 }
 835         }
 836
 837         if ((err = verify_group_input(sb, input)))
 838                 goto exit_put;
 839
 840         if ((err = setup_new_group_blocks(sb, input)))
 841                 goto exit_put;
 842
 843         /*
 844          * We will always be modifying at least the superblock and a GDT
 845          * block.  If we are adding a group past the last current GDT block,
 846          * we will also modify the inode and the dindirect block.  If we
 847          * are adding a group with superblock/GDT backups  we will also
 848          * modify each of the reserved GDT dindirect blocks.
 849          */
 850         handle = ext3_journal_start_sb(sb,
 851                                        ext3_bg_has_super(sb, input->group) ?
 852                                        3 + reserved_gdb : 4);
 853         if (IS_ERR(handle)) {
 854                 err = PTR_ERR(handle);
 855                 goto exit_put;
 856         }
 857
 858         mutex_lock(&sbi->s_resize_lock);
 859         if (input->group != sbi->s_groups_count) {
 860                 ext3_warning(sb, __func__,
 861                              "multiple resizers run on filesystem!");
 862                 err = -EBUSY;
 863                 goto exit_journal;
 864         }
 865
 866         if ((err = ext3_journal_get_write_access(handle, sbi->s_sbh)))
 867                 goto exit_journal;
 868
 869         /*
 870          * We will only either add reserved group blocks to a backup group
 871          * or remove reserved blocks for the first group in a new group block.
 872          * Doing both would be mean more complex code, and sane people don't
 873          * use non-sparse filesystems anymore.  This is already checked above.
 874          */
 875         if (gdb_off) {
 876                 primary = sbi->s_group_desc[gdb_num];
 877                 if ((err = ext3_journal_get_write_access(handle, primary)))
 878                         goto exit_journal;
 879
 880                 if (reserved_gdb && ext3_bg_num_gdb(sb, input->group) &&
 881                     (err = reserve_backup_gdb(handle, inode, input)))
 882                         goto exit_journal;
 883         } else if ((err = add_new_gdb(handle, inode, input, &primary)))
 884                 goto exit_journal;
 885
 886         /*
 887          * OK, now we've set up the new group.  Time to make it active.
 888          *
 889          * We do not lock all allocations via s_resize_lock
 890          * so we have to be safe wrt. concurrent accesses the group
 891          * data.  So we need to be careful to set all of the relevant
 892          * group descriptor data etc. *before* we enable the group.
 893          *
 894          * The key field here is sbi->s_groups_count: as long as
 895          * that retains its old value, nobody is going to access the new
 896          * group.
 897          *
 898          * So first we update all the descriptor metadata for the new
 899          * group; then we update the total disk blocks count; then we
 900          * update the groups count to enable the group; then finally we
 901          * update the free space counts so that the system can start
 902          * using the new disk blocks.
 903          */
 904
 905         /* Update group descriptor block for new group */
 906         gdp = (struct ext3_group_desc *)primary->b_data + gdb_off;
 907
 908         gdp->bg_block_bitmap = cpu_to_le32(input->block_bitmap);
 909         gdp->bg_inode_bitmap = cpu_to_le32(input->inode_bitmap);
 910         gdp->bg_inode_table = cpu_to_le32(input->inode_table);
 911         gdp->bg_free_blocks_count = cpu_to_le16(input->free_blocks_count);
 912         gdp->bg_free_inodes_count = cpu_to_le16(EXT3_INODES_PER_GROUP(sb));
 913
 914         /*
 915          * Make the new blocks and inodes valid next.  We do this before
 916          * increasing the group count so that once the group is enabled,
 917          * all of its blocks and inodes are already valid.
 918          *
 919          * We always allocate group-by-group, then block-by-block or
 920          * inode-by-inode within a group, so enabling these
 921          * blocks/inodes before the group is live won't actually let us
 922          * allocate the new space yet.
 923          */
 924         le32_add_cpu(&es->s_blocks_count, input->blocks_count);
 925         le32_add_cpu(&es->s_inodes_count, EXT3_INODES_PER_GROUP(sb));
 926
 927         /*
 928          * We need to protect s_groups_count against other CPUs seeing
 929          * inconsistent state in the superblock.
 930          *
 931          * The precise rules we use are:
 932          *
 933          * * Writers of s_groups_count *must* hold s_resize_lock
 934          * AND
 935          * * Writers must perform a smp_wmb() after updating all dependent
 936          *   data and before modifying the groups count
 937          *
 938          * * Readers must hold s_resize_lock over the access
 939          * OR
 940          * * Readers must perform an smp_rmb() after reading the groups count
 941          *   and before reading any dependent data.
 942          *
 943          * NB. These rules can be relaxed when checking the group count
 944          * while freeing data, as we can only allocate from a block
 945          * group after serialising against the group count, and we can
 946          * only then free after serialising in turn against that
 947          * allocation.
 948          */
 949         smp_wmb();
 950
 951         /* Update the global fs size fields */
 952         sbi->s_groups_count++;
 953
 954         err = ext3_journal_dirty_metadata(handle, primary);
 955         if (err)
 956                 goto exit_journal;
 957
 958         /* Update the reserved block counts only once the new group is
 959          * active. */
 960         le32_add_cpu(&es->s_r_blocks_count, input->reserved_blocks);
 961
 962         /* Update the free space counts */
 963         percpu_counter_add(&sbi->s_freeblocks_counter,
 964                            input->free_blocks_count);
 965         percpu_counter_add(&sbi->s_freeinodes_counter,
 966                            EXT3_INODES_PER_GROUP(sb));
 967
 968         err = ext3_journal_dirty_metadata(handle, sbi->s_sbh);
 969
 970 exit_journal:
 971         mutex_unlock(&sbi->s_resize_lock);
 972         if ((err2 = ext3_journal_stop(handle)) && !err)
 973                 err = err2;
 974         if (!err) {
 975                 update_backups(sb, sbi->s_sbh->b_blocknr, (char *)es,
 976                                sizeof(struct ext3_super_block));
 977                 update_backups(sb, primary->b_blocknr, primary->b_data,
 978                                primary->b_size);
 979         }
 980 exit_put:
 981         iput(inode);
 982         return err;
 983 } /* ext3_group_add */
 984
 985 /* Extend the filesystem to the new number of blocks specified.  This entry
 986  * point is only used to extend the current filesystem to the end of the last
 987  * existing group.  It can be accessed via ioctl, or by "remount,resize=<size>"
 988  * for emergencies (because it has no dependencies on reserved blocks).
 989  *
 990  * If we _really_ wanted, we could use default values to call ext3_group_add()
 991  * allow the "remount" trick to work for arbitrary resizing, assuming enough
 992  * GDT blocks are reserved to grow to the desired size.
 993  */
 994 int ext3_group_extend(struct super_block *sb, struct ext3_super_block *es,
 995                       ext3_fsblk_t n_blocks_count)
 996 {
 997         ext3_fsblk_t o_blocks_count;
 998         ext3_grpblk_t last;
 999         ext3_grpblk_t add;
1000         struct buffer_head * bh;
1001         handle_t *handle;
1002         int err;
1003         unsigned long freed_blocks;
1004
1005         /* We don't need to worry about locking wrt other resizers just
1006          * yet: we're going to revalidate es->s_blocks_count after
1007          * taking the s_resize_lock below. */
1008         o_blocks_count = le32_to_cpu(es->s_blocks_count);
1009
1010         if (test_opt(sb, DEBUG))
1011                 printk(KERN_DEBUG "EXT3-fs: extending last group from "E3FSBLK
1012                        " upto "E3FSBLK" blocks\n",
1013                        o_blocks_count, n_blocks_count);
1014
1015         if (n_blocks_count == 0 || n_blocks_count == o_blocks_count)
1016                 return 0;
1017
1018         if (n_blocks_count > (sector_t)(~0ULL) >> (sb->s_blocksize_bits - 9)) {
1019                 printk(KERN_ERR "EXT3-fs: filesystem on %s:"
1020                         " too large to resize to "E3FSBLK" blocks safely\n",
1021                         sb->s_id, n_blocks_count);
1022                 if (sizeof(sector_t) < 8)
1023                         ext3_warning(sb, __func__,
1024                         "CONFIG_LBDAF not enabled\n");
1025                 return -EINVAL;
1026         }
1027
1028         if (n_blocks_count < o_blocks_count) {
1029                 ext3_warning(sb, __func__,
1030                              "can't shrink FS - resize aborted");
1031                 return -EBUSY;
1032         }
1033
1034         /* Handle the remaining blocks in the last group only. */
1035         last = (o_blocks_count - le32_to_cpu(es->s_first_data_block)) %
1036                 EXT3_BLOCKS_PER_GROUP(sb);
1037
1038         if (last == 0) {
1039                 ext3_warning(sb, __func__,
1040                              "need to use ext2online to resize further");
1041                 return -EPERM;
1042         }
1043
1044         add = EXT3_BLOCKS_PER_GROUP(sb) - last;
1045
1046         if (o_blocks_count + add < o_blocks_count) {
1047                 ext3_warning(sb, __func__, "blocks_count overflow");
1048                 return -EINVAL;
1049         }
1050
1051         if (o_blocks_count + add > n_blocks_count)
1052                 add = n_blocks_count - o_blocks_count;
1053
1054         if (o_blocks_count + add < n_blocks_count)
1055                 ext3_warning(sb, __func__,
1056                              "will only finish group ("E3FSBLK
1057                              " blocks, %u new)",
1058                              o_blocks_count + add, add);
1059
1060         /* See if the device is actually as big as what was requested */
1061         bh = sb_bread(sb, o_blocks_count + add -1);
1062         if (!bh) {
1063                 ext3_warning(sb, __func__,
1064                              "can't read last block, resize aborted");
1065                 return -ENOSPC;
1066         }
1067         brelse(bh);
1068
1069         /* We will update the superblock, one block bitmap, and
1070          * one group descriptor via ext3_free_blocks().
1071          */
1072         handle = ext3_journal_start_sb(sb, 3);
1073         if (IS_ERR(handle)) {
1074                 err = PTR_ERR(handle);
1075                 ext3_warning(sb, __func__, "error %d on journal start",err);
1076                 goto exit_put;
1077         }
1078
1079         mutex_lock(&EXT3_SB(sb)->s_resize_lock);
1080         if (o_blocks_count != le32_to_cpu(es->s_blocks_count)) {
1081                 ext3_warning(sb, __func__,
1082                              "multiple resizers run on filesystem!");
1083                 mutex_unlock(&EXT3_SB(sb)->s_resize_lock);
1084                 ext3_journal_stop(handle);
1085                 err = -EBUSY;
1086                 goto exit_put;
1087         }
1088
1089         if ((err = ext3_journal_get_write_access(handle,
1090                                                  EXT3_SB(sb)->s_sbh))) {
1091                 ext3_warning(sb, __func__,
1092                              "error %d on journal write access", err);
1093                 mutex_unlock(&EXT3_SB(sb)->s_resize_lock);
1094                 ext3_journal_stop(handle);
1095                 goto exit_put;
1096         }
1097         es->s_blocks_count = cpu_to_le32(o_blocks_count + add);
1098         err = ext3_journal_dirty_metadata(handle, EXT3_SB(sb)->s_sbh);
1099         mutex_unlock(&EXT3_SB(sb)->s_resize_lock);
1100         if (err) {
1101                 ext3_warning(sb, __func__,
1102                              "error %d on journal dirty metadata", err);
1103                 ext3_journal_stop(handle);
1104                 goto exit_put;
1105         }
1106         ext3_debug("freeing blocks "E3FSBLK" through "E3FSBLK"\n",
1107                    o_blocks_count, o_blocks_count + add);
1108         ext3_free_blocks_sb(handle, sb, o_blocks_count, add, &freed_blocks);
1109         ext3_debug("freed blocks "E3FSBLK" through "E3FSBLK"\n",
1110                    o_blocks_count, o_blocks_count + add);
1111         if ((err = ext3_journal_stop(handle)))
1112                 goto exit_put;
1113         if (test_opt(sb, DEBUG))
1114                 printk(KERN_DEBUG "EXT3-fs: extended group to %u blocks\n",
1115                        le32_to_cpu(es->s_blocks_count));
1116         update_backups(sb, EXT3_SB(sb)->s_sbh->b_blocknr, (char *)es,
1117                        sizeof(struct ext3_super_block));
1118 exit_put:
1119         return err;
1120 } /* ext3_group_extend */