Merge branches 'topic/fix/misc' and 'topic/fix/hda' into for-linus
[linux-2.6] / fs / ext4 / resize.c
1 /*
2  *  linux/fs/ext4/resize.c
3  *
4  * Support for resizing an ext4 filesystem while it is mounted.
5  *
6  * Copyright (C) 2001, 2002 Andreas Dilger <adilger@clusterfs.com>
7  *
8  * This could probably be made into a module, because it is not often in use.
9  */
10
11
12 #define EXT4FS_DEBUG
13
14 #include <linux/errno.h>
15 #include <linux/slab.h>
16
17 #include "ext4_jbd2.h"
18 #include "group.h"
19
20 #define outside(b, first, last) ((b) < (first) || (b) >= (last))
21 #define inside(b, first, last)  ((b) >= (first) && (b) < (last))
22
23 static int verify_group_input(struct super_block *sb,
24                               struct ext4_new_group_data *input)
25 {
26         struct ext4_sb_info *sbi = EXT4_SB(sb);
27         struct ext4_super_block *es = sbi->s_es;
28         ext4_fsblk_t start = ext4_blocks_count(es);
29         ext4_fsblk_t end = start + input->blocks_count;
30         ext4_group_t group = input->group;
31         ext4_fsblk_t itend = input->inode_table + sbi->s_itb_per_group;
32         unsigned overhead = ext4_bg_has_super(sb, group) ?
33                 (1 + ext4_bg_num_gdb(sb, group) +
34                  le16_to_cpu(es->s_reserved_gdt_blocks)) : 0;
35         ext4_fsblk_t metaend = start + overhead;
36         struct buffer_head *bh = NULL;
37         ext4_grpblk_t free_blocks_count, offset;
38         int err = -EINVAL;
39
40         input->free_blocks_count = free_blocks_count =
41                 input->blocks_count - 2 - overhead - sbi->s_itb_per_group;
42
43         if (test_opt(sb, DEBUG))
44                 printk(KERN_DEBUG "EXT4-fs: adding %s group %u: %u blocks "
45                        "(%d free, %u reserved)\n",
46                        ext4_bg_has_super(sb, input->group) ? "normal" :
47                        "no-super", input->group, input->blocks_count,
48                        free_blocks_count, input->reserved_blocks);
49
50         ext4_get_group_no_and_offset(sb, start, NULL, &offset);
51         if (group != sbi->s_groups_count)
52                 ext4_warning(sb, __func__,
53                              "Cannot add at group %u (only %lu groups)",
54                              input->group, sbi->s_groups_count);
55         else if (offset != 0)
56                         ext4_warning(sb, __func__, "Last group not full");
57         else if (input->reserved_blocks > input->blocks_count / 5)
58                 ext4_warning(sb, __func__, "Reserved blocks too high (%u)",
59                              input->reserved_blocks);
60         else if (free_blocks_count < 0)
61                 ext4_warning(sb, __func__, "Bad blocks count %u",
62                              input->blocks_count);
63         else if (!(bh = sb_bread(sb, end - 1)))
64                 ext4_warning(sb, __func__,
65                              "Cannot read last block (%llu)",
66                              end - 1);
67         else if (outside(input->block_bitmap, start, end))
68                 ext4_warning(sb, __func__,
69                              "Block bitmap not in group (block %llu)",
70                              (unsigned long long)input->block_bitmap);
71         else if (outside(input->inode_bitmap, start, end))
72                 ext4_warning(sb, __func__,
73                              "Inode bitmap not in group (block %llu)",
74                              (unsigned long long)input->inode_bitmap);
75         else if (outside(input->inode_table, start, end) ||
76                  outside(itend - 1, start, end))
77                 ext4_warning(sb, __func__,
78                              "Inode table not in group (blocks %llu-%llu)",
79                              (unsigned long long)input->inode_table, itend - 1);
80         else if (input->inode_bitmap == input->block_bitmap)
81                 ext4_warning(sb, __func__,
82                              "Block bitmap same as inode bitmap (%llu)",
83                              (unsigned long long)input->block_bitmap);
84         else if (inside(input->block_bitmap, input->inode_table, itend))
85                 ext4_warning(sb, __func__,
86                              "Block bitmap (%llu) in inode table (%llu-%llu)",
87                              (unsigned long long)input->block_bitmap,
88                              (unsigned long long)input->inode_table, itend - 1);
89         else if (inside(input->inode_bitmap, input->inode_table, itend))
90                 ext4_warning(sb, __func__,
91                              "Inode bitmap (%llu) in inode table (%llu-%llu)",
92                              (unsigned long long)input->inode_bitmap,
93                              (unsigned long long)input->inode_table, itend - 1);
94         else if (inside(input->block_bitmap, start, metaend))
95                 ext4_warning(sb, __func__,
96                              "Block bitmap (%llu) in GDT table"
97                              " (%llu-%llu)",
98                              (unsigned long long)input->block_bitmap,
99                              start, metaend - 1);
100         else if (inside(input->inode_bitmap, start, metaend))
101                 ext4_warning(sb, __func__,
102                              "Inode bitmap (%llu) in GDT table"
103                              " (%llu-%llu)",
104                              (unsigned long long)input->inode_bitmap,
105                              start, metaend - 1);
106         else if (inside(input->inode_table, start, metaend) ||
107                  inside(itend - 1, start, metaend))
108                 ext4_warning(sb, __func__,
109                              "Inode table (%llu-%llu) overlaps"
110                              "GDT table (%llu-%llu)",
111                              (unsigned long long)input->inode_table,
112                              itend - 1, start, metaend - 1);
113         else
114                 err = 0;
115         brelse(bh);
116
117         return err;
118 }
119
120 static struct buffer_head *bclean(handle_t *handle, struct super_block *sb,
121                                   ext4_fsblk_t blk)
122 {
123         struct buffer_head *bh;
124         int err;
125
126         bh = sb_getblk(sb, blk);
127         if (!bh)
128                 return ERR_PTR(-EIO);
129         if ((err = ext4_journal_get_write_access(handle, bh))) {
130                 brelse(bh);
131                 bh = ERR_PTR(err);
132         } else {
133                 lock_buffer(bh);
134                 memset(bh->b_data, 0, sb->s_blocksize);
135                 set_buffer_uptodate(bh);
136                 unlock_buffer(bh);
137         }
138
139         return bh;
140 }
141
142 /*
143  * If we have fewer than thresh credits, extend by EXT4_MAX_TRANS_DATA.
144  * If that fails, restart the transaction & regain write access for the
145  * buffer head which is used for block_bitmap modifications.
146  */
147 static int extend_or_restart_transaction(handle_t *handle, int thresh,
148                                          struct buffer_head *bh)
149 {
150         int err;
151
152         if (handle->h_buffer_credits >= thresh)
153                 return 0;
154
155         err = ext4_journal_extend(handle, EXT4_MAX_TRANS_DATA);
156         if (err < 0)
157                 return err;
158         if (err) {
159                 if ((err = ext4_journal_restart(handle, EXT4_MAX_TRANS_DATA)))
160                         return err;
161                 if ((err = ext4_journal_get_write_access(handle, bh)))
162                         return err;
163         }
164
165         return 0;
166 }
167
168 /*
169  * Set up the block and inode bitmaps, and the inode table for the new group.
170  * This doesn't need to be part of the main transaction, since we are only
171  * changing blocks outside the actual filesystem.  We still do journaling to
172  * ensure the recovery is correct in case of a failure just after resize.
173  * If any part of this fails, we simply abort the resize.
174  */
175 static int setup_new_group_blocks(struct super_block *sb,
176                                   struct ext4_new_group_data *input)
177 {
178         struct ext4_sb_info *sbi = EXT4_SB(sb);
179         ext4_fsblk_t start = ext4_group_first_block_no(sb, input->group);
180         int reserved_gdb = ext4_bg_has_super(sb, input->group) ?
181                 le16_to_cpu(sbi->s_es->s_reserved_gdt_blocks) : 0;
182         unsigned long gdblocks = ext4_bg_num_gdb(sb, input->group);
183         struct buffer_head *bh;
184         handle_t *handle;
185         ext4_fsblk_t block;
186         ext4_grpblk_t bit;
187         int i;
188         int err = 0, err2;
189
190         /* This transaction may be extended/restarted along the way */
191         handle = ext4_journal_start_sb(sb, EXT4_MAX_TRANS_DATA);
192
193         if (IS_ERR(handle))
194                 return PTR_ERR(handle);
195
196         lock_super(sb);
197         if (input->group != sbi->s_groups_count) {
198                 err = -EBUSY;
199                 goto exit_journal;
200         }
201
202         if (IS_ERR(bh = bclean(handle, sb, input->block_bitmap))) {
203                 err = PTR_ERR(bh);
204                 goto exit_journal;
205         }
206
207         if (ext4_bg_has_super(sb, input->group)) {
208                 ext4_debug("mark backup superblock %#04llx (+0)\n", start);
209                 ext4_set_bit(0, bh->b_data);
210         }
211
212         /* Copy all of the GDT blocks into the backup in this group */
213         for (i = 0, bit = 1, block = start + 1;
214              i < gdblocks; i++, block++, bit++) {
215                 struct buffer_head *gdb;
216
217                 ext4_debug("update backup group %#04llx (+%d)\n", block, bit);
218
219                 if ((err = extend_or_restart_transaction(handle, 1, bh)))
220                         goto exit_bh;
221
222                 gdb = sb_getblk(sb, block);
223                 if (!gdb) {
224                         err = -EIO;
225                         goto exit_bh;
226                 }
227                 if ((err = ext4_journal_get_write_access(handle, gdb))) {
228                         brelse(gdb);
229                         goto exit_bh;
230                 }
231                 lock_buffer(gdb);
232                 memcpy(gdb->b_data, sbi->s_group_desc[i]->b_data, gdb->b_size);
233                 set_buffer_uptodate(gdb);
234                 unlock_buffer(gdb);
235                 ext4_journal_dirty_metadata(handle, gdb);
236                 ext4_set_bit(bit, bh->b_data);
237                 brelse(gdb);
238         }
239
240         /* Zero out all of the reserved backup group descriptor table blocks */
241         for (i = 0, bit = gdblocks + 1, block = start + bit;
242              i < reserved_gdb; i++, block++, bit++) {
243                 struct buffer_head *gdb;
244
245                 ext4_debug("clear reserved block %#04llx (+%d)\n", block, bit);
246
247                 if ((err = extend_or_restart_transaction(handle, 1, bh)))
248                         goto exit_bh;
249
250                 if (IS_ERR(gdb = bclean(handle, sb, block))) {
251                         err = PTR_ERR(bh);
252                         goto exit_bh;
253                 }
254                 ext4_journal_dirty_metadata(handle, gdb);
255                 ext4_set_bit(bit, bh->b_data);
256                 brelse(gdb);
257         }
258         ext4_debug("mark block bitmap %#04llx (+%llu)\n", input->block_bitmap,
259                    input->block_bitmap - start);
260         ext4_set_bit(input->block_bitmap - start, bh->b_data);
261         ext4_debug("mark inode bitmap %#04llx (+%llu)\n", input->inode_bitmap,
262                    input->inode_bitmap - start);
263         ext4_set_bit(input->inode_bitmap - start, bh->b_data);
264
265         /* Zero out all of the inode table blocks */
266         for (i = 0, block = input->inode_table, bit = block - start;
267              i < sbi->s_itb_per_group; i++, bit++, block++) {
268                 struct buffer_head *it;
269
270                 ext4_debug("clear inode block %#04llx (+%d)\n", block, bit);
271
272                 if ((err = extend_or_restart_transaction(handle, 1, bh)))
273                         goto exit_bh;
274
275                 if (IS_ERR(it = bclean(handle, sb, block))) {
276                         err = PTR_ERR(it);
277                         goto exit_bh;
278                 }
279                 ext4_journal_dirty_metadata(handle, it);
280                 brelse(it);
281                 ext4_set_bit(bit, bh->b_data);
282         }
283
284         if ((err = extend_or_restart_transaction(handle, 2, bh)))
285                 goto exit_bh;
286
287         mark_bitmap_end(input->blocks_count, EXT4_BLOCKS_PER_GROUP(sb),
288                         bh->b_data);
289         ext4_journal_dirty_metadata(handle, bh);
290         brelse(bh);
291
292         /* Mark unused entries in inode bitmap used */
293         ext4_debug("clear inode bitmap %#04llx (+%llu)\n",
294                    input->inode_bitmap, input->inode_bitmap - start);
295         if (IS_ERR(bh = bclean(handle, sb, input->inode_bitmap))) {
296                 err = PTR_ERR(bh);
297                 goto exit_journal;
298         }
299
300         mark_bitmap_end(EXT4_INODES_PER_GROUP(sb), EXT4_BLOCKS_PER_GROUP(sb),
301                         bh->b_data);
302         ext4_journal_dirty_metadata(handle, bh);
303 exit_bh:
304         brelse(bh);
305
306 exit_journal:
307         unlock_super(sb);
308         if ((err2 = ext4_journal_stop(handle)) && !err)
309                 err = err2;
310
311         return err;
312 }
313
314 /*
315  * Iterate through the groups which hold BACKUP superblock/GDT copies in an
316  * ext4 filesystem.  The counters should be initialized to 1, 5, and 7 before
317  * calling this for the first time.  In a sparse filesystem it will be the
318  * sequence of powers of 3, 5, and 7: 1, 3, 5, 7, 9, 25, 27, 49, 81, ...
319  * For a non-sparse filesystem it will be every group: 1, 2, 3, 4, ...
320  */
321 static unsigned ext4_list_backups(struct super_block *sb, unsigned *three,
322                                   unsigned *five, unsigned *seven)
323 {
324         unsigned *min = three;
325         int mult = 3;
326         unsigned ret;
327
328         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
329                                         EXT4_FEATURE_RO_COMPAT_SPARSE_SUPER)) {
330                 ret = *min;
331                 *min += 1;
332                 return ret;
333         }
334
335         if (*five < *min) {
336                 min = five;
337                 mult = 5;
338         }
339         if (*seven < *min) {
340                 min = seven;
341                 mult = 7;
342         }
343
344         ret = *min;
345         *min *= mult;
346
347         return ret;
348 }
349
350 /*
351  * Check that all of the backup GDT blocks are held in the primary GDT block.
352  * It is assumed that they are stored in group order.  Returns the number of
353  * groups in current filesystem that have BACKUPS, or -ve error code.
354  */
355 static int verify_reserved_gdb(struct super_block *sb,
356                                struct buffer_head *primary)
357 {
358         const ext4_fsblk_t blk = primary->b_blocknr;
359         const ext4_group_t end = EXT4_SB(sb)->s_groups_count;
360         unsigned three = 1;
361         unsigned five = 5;
362         unsigned seven = 7;
363         unsigned grp;
364         __le32 *p = (__le32 *)primary->b_data;
365         int gdbackups = 0;
366
367         while ((grp = ext4_list_backups(sb, &three, &five, &seven)) < end) {
368                 if (le32_to_cpu(*p++) !=
369                     grp * EXT4_BLOCKS_PER_GROUP(sb) + blk){
370                         ext4_warning(sb, __func__,
371                                      "reserved GDT %llu"
372                                      " missing grp %d (%llu)",
373                                      blk, grp,
374                                      grp *
375                                      (ext4_fsblk_t)EXT4_BLOCKS_PER_GROUP(sb) +
376                                      blk);
377                         return -EINVAL;
378                 }
379                 if (++gdbackups > EXT4_ADDR_PER_BLOCK(sb))
380                         return -EFBIG;
381         }
382
383         return gdbackups;
384 }
385
386 /*
387  * Called when we need to bring a reserved group descriptor table block into
388  * use from the resize inode.  The primary copy of the new GDT block currently
389  * is an indirect block (under the double indirect block in the resize inode).
390  * The new backup GDT blocks will be stored as leaf blocks in this indirect
391  * block, in group order.  Even though we know all the block numbers we need,
392  * we check to ensure that the resize inode has actually reserved these blocks.
393  *
394  * Don't need to update the block bitmaps because the blocks are still in use.
395  *
396  * We get all of the error cases out of the way, so that we are sure to not
397  * fail once we start modifying the data on disk, because JBD has no rollback.
398  */
399 static int add_new_gdb(handle_t *handle, struct inode *inode,
400                        struct ext4_new_group_data *input,
401                        struct buffer_head **primary)
402 {
403         struct super_block *sb = inode->i_sb;
404         struct ext4_super_block *es = EXT4_SB(sb)->s_es;
405         unsigned long gdb_num = input->group / EXT4_DESC_PER_BLOCK(sb);
406         ext4_fsblk_t gdblock = EXT4_SB(sb)->s_sbh->b_blocknr + 1 + gdb_num;
407         struct buffer_head **o_group_desc, **n_group_desc;
408         struct buffer_head *dind;
409         int gdbackups;
410         struct ext4_iloc iloc;
411         __le32 *data;
412         int err;
413
414         if (test_opt(sb, DEBUG))
415                 printk(KERN_DEBUG
416                        "EXT4-fs: ext4_add_new_gdb: adding group block %lu\n",
417                        gdb_num);
418
419         /*
420          * If we are not using the primary superblock/GDT copy don't resize,
421          * because the user tools have no way of handling this.  Probably a
422          * bad time to do it anyways.
423          */
424         if (EXT4_SB(sb)->s_sbh->b_blocknr !=
425             le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block)) {
426                 ext4_warning(sb, __func__,
427                         "won't resize using backup superblock at %llu",
428                         (unsigned long long)EXT4_SB(sb)->s_sbh->b_blocknr);
429                 return -EPERM;
430         }
431
432         *primary = sb_bread(sb, gdblock);
433         if (!*primary)
434                 return -EIO;
435
436         if ((gdbackups = verify_reserved_gdb(sb, *primary)) < 0) {
437                 err = gdbackups;
438                 goto exit_bh;
439         }
440
441         data = EXT4_I(inode)->i_data + EXT4_DIND_BLOCK;
442         dind = sb_bread(sb, le32_to_cpu(*data));
443         if (!dind) {
444                 err = -EIO;
445                 goto exit_bh;
446         }
447
448         data = (__le32 *)dind->b_data;
449         if (le32_to_cpu(data[gdb_num % EXT4_ADDR_PER_BLOCK(sb)]) != gdblock) {
450                 ext4_warning(sb, __func__,
451                              "new group %u GDT block %llu not reserved",
452                              input->group, gdblock);
453                 err = -EINVAL;
454                 goto exit_dind;
455         }
456
457         if ((err = ext4_journal_get_write_access(handle, EXT4_SB(sb)->s_sbh)))
458                 goto exit_dind;
459
460         if ((err = ext4_journal_get_write_access(handle, *primary)))
461                 goto exit_sbh;
462
463         if ((err = ext4_journal_get_write_access(handle, dind)))
464                 goto exit_primary;
465
466         /* ext4_reserve_inode_write() gets a reference on the iloc */
467         if ((err = ext4_reserve_inode_write(handle, inode, &iloc)))
468                 goto exit_dindj;
469
470         n_group_desc = kmalloc((gdb_num + 1) * sizeof(struct buffer_head *),
471                         GFP_NOFS);
472         if (!n_group_desc) {
473                 err = -ENOMEM;
474                 ext4_warning(sb, __func__,
475                               "not enough memory for %lu groups", gdb_num + 1);
476                 goto exit_inode;
477         }
478
479         /*
480          * Finally, we have all of the possible failures behind us...
481          *
482          * Remove new GDT block from inode double-indirect block and clear out
483          * the new GDT block for use (which also "frees" the backup GDT blocks
484          * from the reserved inode).  We don't need to change the bitmaps for
485          * these blocks, because they are marked as in-use from being in the
486          * reserved inode, and will become GDT blocks (primary and backup).
487          */
488         data[gdb_num % EXT4_ADDR_PER_BLOCK(sb)] = 0;
489         ext4_journal_dirty_metadata(handle, dind);
490         brelse(dind);
491         inode->i_blocks -= (gdbackups + 1) * sb->s_blocksize >> 9;
492         ext4_mark_iloc_dirty(handle, inode, &iloc);
493         memset((*primary)->b_data, 0, sb->s_blocksize);
494         ext4_journal_dirty_metadata(handle, *primary);
495
496         o_group_desc = EXT4_SB(sb)->s_group_desc;
497         memcpy(n_group_desc, o_group_desc,
498                EXT4_SB(sb)->s_gdb_count * sizeof(struct buffer_head *));
499         n_group_desc[gdb_num] = *primary;
500         EXT4_SB(sb)->s_group_desc = n_group_desc;
501         EXT4_SB(sb)->s_gdb_count++;
502         kfree(o_group_desc);
503
504         le16_add_cpu(&es->s_reserved_gdt_blocks, -1);
505         ext4_journal_dirty_metadata(handle, EXT4_SB(sb)->s_sbh);
506
507         return 0;
508
509 exit_inode:
510         /* ext4_journal_release_buffer(handle, iloc.bh); */
511         brelse(iloc.bh);
512 exit_dindj:
513         /* ext4_journal_release_buffer(handle, dind); */
514 exit_primary:
515         /* ext4_journal_release_buffer(handle, *primary); */
516 exit_sbh:
517         /* ext4_journal_release_buffer(handle, *primary); */
518 exit_dind:
519         brelse(dind);
520 exit_bh:
521         brelse(*primary);
522
523         ext4_debug("leaving with error %d\n", err);
524         return err;
525 }
526
527 /*
528  * Called when we are adding a new group which has a backup copy of each of
529  * the GDT blocks (i.e. sparse group) and there are reserved GDT blocks.
530  * We need to add these reserved backup GDT blocks to the resize inode, so
531  * that they are kept for future resizing and not allocated to files.
532  *
533  * Each reserved backup GDT block will go into a different indirect block.
534  * The indirect blocks are actually the primary reserved GDT blocks,
535  * so we know in advance what their block numbers are.  We only get the
536  * double-indirect block to verify it is pointing to the primary reserved
537  * GDT blocks so we don't overwrite a data block by accident.  The reserved
538  * backup GDT blocks are stored in their reserved primary GDT block.
539  */
540 static int reserve_backup_gdb(handle_t *handle, struct inode *inode,
541                               struct ext4_new_group_data *input)
542 {
543         struct super_block *sb = inode->i_sb;
544         int reserved_gdb =le16_to_cpu(EXT4_SB(sb)->s_es->s_reserved_gdt_blocks);
545         struct buffer_head **primary;
546         struct buffer_head *dind;
547         struct ext4_iloc iloc;
548         ext4_fsblk_t blk;
549         __le32 *data, *end;
550         int gdbackups = 0;
551         int res, i;
552         int err;
553
554         primary = kmalloc(reserved_gdb * sizeof(*primary), GFP_NOFS);
555         if (!primary)
556                 return -ENOMEM;
557
558         data = EXT4_I(inode)->i_data + EXT4_DIND_BLOCK;
559         dind = sb_bread(sb, le32_to_cpu(*data));
560         if (!dind) {
561                 err = -EIO;
562                 goto exit_free;
563         }
564
565         blk = EXT4_SB(sb)->s_sbh->b_blocknr + 1 + EXT4_SB(sb)->s_gdb_count;
566         data = (__le32 *)dind->b_data + (EXT4_SB(sb)->s_gdb_count %
567                                          EXT4_ADDR_PER_BLOCK(sb));
568         end = (__le32 *)dind->b_data + EXT4_ADDR_PER_BLOCK(sb);
569
570         /* Get each reserved primary GDT block and verify it holds backups */
571         for (res = 0; res < reserved_gdb; res++, blk++) {
572                 if (le32_to_cpu(*data) != blk) {
573                         ext4_warning(sb, __func__,
574                                      "reserved block %llu"
575                                      " not at offset %ld",
576                                      blk,
577                                      (long)(data - (__le32 *)dind->b_data));
578                         err = -EINVAL;
579                         goto exit_bh;
580                 }
581                 primary[res] = sb_bread(sb, blk);
582                 if (!primary[res]) {
583                         err = -EIO;
584                         goto exit_bh;
585                 }
586                 if ((gdbackups = verify_reserved_gdb(sb, primary[res])) < 0) {
587                         brelse(primary[res]);
588                         err = gdbackups;
589                         goto exit_bh;
590                 }
591                 if (++data >= end)
592                         data = (__le32 *)dind->b_data;
593         }
594
595         for (i = 0; i < reserved_gdb; i++) {
596                 if ((err = ext4_journal_get_write_access(handle, primary[i]))) {
597                         /*
598                         int j;
599                         for (j = 0; j < i; j++)
600                                 ext4_journal_release_buffer(handle, primary[j]);
601                          */
602                         goto exit_bh;
603                 }
604         }
605
606         if ((err = ext4_reserve_inode_write(handle, inode, &iloc)))
607                 goto exit_bh;
608
609         /*
610          * Finally we can add each of the reserved backup GDT blocks from
611          * the new group to its reserved primary GDT block.
612          */
613         blk = input->group * EXT4_BLOCKS_PER_GROUP(sb);
614         for (i = 0; i < reserved_gdb; i++) {
615                 int err2;
616                 data = (__le32 *)primary[i]->b_data;
617                 /* printk("reserving backup %lu[%u] = %lu\n",
618                        primary[i]->b_blocknr, gdbackups,
619                        blk + primary[i]->b_blocknr); */
620                 data[gdbackups] = cpu_to_le32(blk + primary[i]->b_blocknr);
621                 err2 = ext4_journal_dirty_metadata(handle, primary[i]);
622                 if (!err)
623                         err = err2;
624         }
625         inode->i_blocks += reserved_gdb * sb->s_blocksize >> 9;
626         ext4_mark_iloc_dirty(handle, inode, &iloc);
627
628 exit_bh:
629         while (--res >= 0)
630                 brelse(primary[res]);
631         brelse(dind);
632
633 exit_free:
634         kfree(primary);
635
636         return err;
637 }
638
639 /*
640  * Update the backup copies of the ext4 metadata.  These don't need to be part
641  * of the main resize transaction, because e2fsck will re-write them if there
642  * is a problem (basically only OOM will cause a problem).  However, we
643  * _should_ update the backups if possible, in case the primary gets trashed
644  * for some reason and we need to run e2fsck from a backup superblock.  The
645  * important part is that the new block and inode counts are in the backup
646  * superblocks, and the location of the new group metadata in the GDT backups.
647  *
648  * We do not need lock_super() for this, because these blocks are not
649  * otherwise touched by the filesystem code when it is mounted.  We don't
650  * need to worry about last changing from sbi->s_groups_count, because the
651  * worst that can happen is that we do not copy the full number of backups
652  * at this time.  The resize which changed s_groups_count will backup again.
653  */
654 static void update_backups(struct super_block *sb,
655                            int blk_off, char *data, int size)
656 {
657         struct ext4_sb_info *sbi = EXT4_SB(sb);
658         const ext4_group_t last = sbi->s_groups_count;
659         const int bpg = EXT4_BLOCKS_PER_GROUP(sb);
660         unsigned three = 1;
661         unsigned five = 5;
662         unsigned seven = 7;
663         ext4_group_t group;
664         int rest = sb->s_blocksize - size;
665         handle_t *handle;
666         int err = 0, err2;
667
668         handle = ext4_journal_start_sb(sb, EXT4_MAX_TRANS_DATA);
669         if (IS_ERR(handle)) {
670                 group = 1;
671                 err = PTR_ERR(handle);
672                 goto exit_err;
673         }
674
675         while ((group = ext4_list_backups(sb, &three, &five, &seven)) < last) {
676                 struct buffer_head *bh;
677
678                 /* Out of journal space, and can't get more - abort - so sad */
679                 if (handle->h_buffer_credits == 0 &&
680                     ext4_journal_extend(handle, EXT4_MAX_TRANS_DATA) &&
681                     (err = ext4_journal_restart(handle, EXT4_MAX_TRANS_DATA)))
682                         break;
683
684                 bh = sb_getblk(sb, group * bpg + blk_off);
685                 if (!bh) {
686                         err = -EIO;
687                         break;
688                 }
689                 ext4_debug("update metadata backup %#04lx\n",
690                           (unsigned long)bh->b_blocknr);
691                 if ((err = ext4_journal_get_write_access(handle, bh)))
692                         break;
693                 lock_buffer(bh);
694                 memcpy(bh->b_data, data, size);
695                 if (rest)
696                         memset(bh->b_data + size, 0, rest);
697                 set_buffer_uptodate(bh);
698                 unlock_buffer(bh);
699                 ext4_journal_dirty_metadata(handle, bh);
700                 brelse(bh);
701         }
702         if ((err2 = ext4_journal_stop(handle)) && !err)
703                 err = err2;
704
705         /*
706          * Ugh! Need to have e2fsck write the backup copies.  It is too
707          * late to revert the resize, we shouldn't fail just because of
708          * the backup copies (they are only needed in case of corruption).
709          *
710          * However, if we got here we have a journal problem too, so we
711          * can't really start a transaction to mark the superblock.
712          * Chicken out and just set the flag on the hope it will be written
713          * to disk, and if not - we will simply wait until next fsck.
714          */
715 exit_err:
716         if (err) {
717                 ext4_warning(sb, __func__,
718                              "can't update backup for group %lu (err %d), "
719                              "forcing fsck on next reboot", group, err);
720                 sbi->s_mount_state &= ~EXT4_VALID_FS;
721                 sbi->s_es->s_state &= cpu_to_le16(~EXT4_VALID_FS);
722                 mark_buffer_dirty(sbi->s_sbh);
723         }
724 }
725
726 /* Add group descriptor data to an existing or new group descriptor block.
727  * Ensure we handle all possible error conditions _before_ we start modifying
728  * the filesystem, because we cannot abort the transaction and not have it
729  * write the data to disk.
730  *
731  * If we are on a GDT block boundary, we need to get the reserved GDT block.
732  * Otherwise, we may need to add backup GDT blocks for a sparse group.
733  *
734  * We only need to hold the superblock lock while we are actually adding
735  * in the new group's counts to the superblock.  Prior to that we have
736  * not really "added" the group at all.  We re-check that we are still
737  * adding in the last group in case things have changed since verifying.
738  */
739 int ext4_group_add(struct super_block *sb, struct ext4_new_group_data *input)
740 {
741         struct ext4_sb_info *sbi = EXT4_SB(sb);
742         struct ext4_super_block *es = sbi->s_es;
743         int reserved_gdb = ext4_bg_has_super(sb, input->group) ?
744                 le16_to_cpu(es->s_reserved_gdt_blocks) : 0;
745         struct buffer_head *primary = NULL;
746         struct ext4_group_desc *gdp;
747         struct inode *inode = NULL;
748         handle_t *handle;
749         int gdb_off, gdb_num;
750         int err, err2;
751
752         gdb_num = input->group / EXT4_DESC_PER_BLOCK(sb);
753         gdb_off = input->group % EXT4_DESC_PER_BLOCK(sb);
754
755         if (gdb_off == 0 && !EXT4_HAS_RO_COMPAT_FEATURE(sb,
756                                         EXT4_FEATURE_RO_COMPAT_SPARSE_SUPER)) {
757                 ext4_warning(sb, __func__,
758                              "Can't resize non-sparse filesystem further");
759                 return -EPERM;
760         }
761
762         if (ext4_blocks_count(es) + input->blocks_count <
763             ext4_blocks_count(es)) {
764                 ext4_warning(sb, __func__, "blocks_count overflow\n");
765                 return -EINVAL;
766         }
767
768         if (le32_to_cpu(es->s_inodes_count) + EXT4_INODES_PER_GROUP(sb) <
769             le32_to_cpu(es->s_inodes_count)) {
770                 ext4_warning(sb, __func__, "inodes_count overflow\n");
771                 return -EINVAL;
772         }
773
774         if (reserved_gdb || gdb_off == 0) {
775                 if (!EXT4_HAS_COMPAT_FEATURE(sb,
776                                              EXT4_FEATURE_COMPAT_RESIZE_INODE)
777                     || !le16_to_cpu(es->s_reserved_gdt_blocks)) {
778                         ext4_warning(sb, __func__,
779                                      "No reserved GDT blocks, can't resize");
780                         return -EPERM;
781                 }
782                 inode = ext4_iget(sb, EXT4_RESIZE_INO);
783                 if (IS_ERR(inode)) {
784                         ext4_warning(sb, __func__,
785                                      "Error opening resize inode");
786                         return PTR_ERR(inode);
787                 }
788         }
789
790         if ((err = verify_group_input(sb, input)))
791                 goto exit_put;
792
793         if ((err = setup_new_group_blocks(sb, input)))
794                 goto exit_put;
795
796         /*
797          * We will always be modifying at least the superblock and a GDT
798          * block.  If we are adding a group past the last current GDT block,
799          * we will also modify the inode and the dindirect block.  If we
800          * are adding a group with superblock/GDT backups  we will also
801          * modify each of the reserved GDT dindirect blocks.
802          */
803         handle = ext4_journal_start_sb(sb,
804                                        ext4_bg_has_super(sb, input->group) ?
805                                        3 + reserved_gdb : 4);
806         if (IS_ERR(handle)) {
807                 err = PTR_ERR(handle);
808                 goto exit_put;
809         }
810
811         lock_super(sb);
812         if (input->group != sbi->s_groups_count) {
813                 ext4_warning(sb, __func__,
814                              "multiple resizers run on filesystem!");
815                 err = -EBUSY;
816                 goto exit_journal;
817         }
818
819         if ((err = ext4_journal_get_write_access(handle, sbi->s_sbh)))
820                 goto exit_journal;
821
822         /*
823          * We will only either add reserved group blocks to a backup group
824          * or remove reserved blocks for the first group in a new group block.
825          * Doing both would be mean more complex code, and sane people don't
826          * use non-sparse filesystems anymore.  This is already checked above.
827          */
828         if (gdb_off) {
829                 primary = sbi->s_group_desc[gdb_num];
830                 if ((err = ext4_journal_get_write_access(handle, primary)))
831                         goto exit_journal;
832
833                 if (reserved_gdb && ext4_bg_num_gdb(sb, input->group) &&
834                     (err = reserve_backup_gdb(handle, inode, input)))
835                         goto exit_journal;
836         } else if ((err = add_new_gdb(handle, inode, input, &primary)))
837                 goto exit_journal;
838
839         /*
840          * OK, now we've set up the new group.  Time to make it active.
841          *
842          * Current kernels don't lock all allocations via lock_super(),
843          * so we have to be safe wrt. concurrent accesses the group
844          * data.  So we need to be careful to set all of the relevant
845          * group descriptor data etc. *before* we enable the group.
846          *
847          * The key field here is sbi->s_groups_count: as long as
848          * that retains its old value, nobody is going to access the new
849          * group.
850          *
851          * So first we update all the descriptor metadata for the new
852          * group; then we update the total disk blocks count; then we
853          * update the groups count to enable the group; then finally we
854          * update the free space counts so that the system can start
855          * using the new disk blocks.
856          */
857
858         /* Update group descriptor block for new group */
859         gdp = (struct ext4_group_desc *)((char *)primary->b_data +
860                                          gdb_off * EXT4_DESC_SIZE(sb));
861
862         ext4_block_bitmap_set(sb, gdp, input->block_bitmap); /* LV FIXME */
863         ext4_inode_bitmap_set(sb, gdp, input->inode_bitmap); /* LV FIXME */
864         ext4_inode_table_set(sb, gdp, input->inode_table); /* LV FIXME */
865         gdp->bg_free_blocks_count = cpu_to_le16(input->free_blocks_count);
866         gdp->bg_free_inodes_count = cpu_to_le16(EXT4_INODES_PER_GROUP(sb));
867         gdp->bg_checksum = ext4_group_desc_csum(sbi, input->group, gdp);
868
869         /*
870          * We can allocate memory for mb_alloc based on the new group
871          * descriptor
872          */
873         err = ext4_mb_add_more_groupinfo(sb, input->group, gdp);
874         if (err)
875                 goto exit_journal;
876
877         /*
878          * Make the new blocks and inodes valid next.  We do this before
879          * increasing the group count so that once the group is enabled,
880          * all of its blocks and inodes are already valid.
881          *
882          * We always allocate group-by-group, then block-by-block or
883          * inode-by-inode within a group, so enabling these
884          * blocks/inodes before the group is live won't actually let us
885          * allocate the new space yet.
886          */
887         ext4_blocks_count_set(es, ext4_blocks_count(es) +
888                 input->blocks_count);
889         le32_add_cpu(&es->s_inodes_count, EXT4_INODES_PER_GROUP(sb));
890
891         /*
892          * We need to protect s_groups_count against other CPUs seeing
893          * inconsistent state in the superblock.
894          *
895          * The precise rules we use are:
896          *
897          * * Writers of s_groups_count *must* hold lock_super
898          * AND
899          * * Writers must perform a smp_wmb() after updating all dependent
900          *   data and before modifying the groups count
901          *
902          * * Readers must hold lock_super() over the access
903          * OR
904          * * Readers must perform an smp_rmb() after reading the groups count
905          *   and before reading any dependent data.
906          *
907          * NB. These rules can be relaxed when checking the group count
908          * while freeing data, as we can only allocate from a block
909          * group after serialising against the group count, and we can
910          * only then free after serialising in turn against that
911          * allocation.
912          */
913         smp_wmb();
914
915         /* Update the global fs size fields */
916         sbi->s_groups_count++;
917
918         ext4_journal_dirty_metadata(handle, primary);
919
920         /* Update the reserved block counts only once the new group is
921          * active. */
922         ext4_r_blocks_count_set(es, ext4_r_blocks_count(es) +
923                 input->reserved_blocks);
924
925         /* Update the free space counts */
926         percpu_counter_add(&sbi->s_freeblocks_counter,
927                            input->free_blocks_count);
928         percpu_counter_add(&sbi->s_freeinodes_counter,
929                            EXT4_INODES_PER_GROUP(sb));
930
931         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_FLEX_BG)) {
932                 ext4_group_t flex_group;
933                 flex_group = ext4_flex_group(sbi, input->group);
934                 sbi->s_flex_groups[flex_group].free_blocks +=
935                         input->free_blocks_count;
936                 sbi->s_flex_groups[flex_group].free_inodes +=
937                         EXT4_INODES_PER_GROUP(sb);
938         }
939
940         ext4_journal_dirty_metadata(handle, sbi->s_sbh);
941         sb->s_dirt = 1;
942
943 exit_journal:
944         unlock_super(sb);
945         if ((err2 = ext4_journal_stop(handle)) && !err)
946                 err = err2;
947         if (!err) {
948                 update_backups(sb, sbi->s_sbh->b_blocknr, (char *)es,
949                                sizeof(struct ext4_super_block));
950                 update_backups(sb, primary->b_blocknr, primary->b_data,
951                                primary->b_size);
952         }
953 exit_put:
954         iput(inode);
955         return err;
956 } /* ext4_group_add */
957
958 /*
959  * Extend the filesystem to the new number of blocks specified.  This entry
960  * point is only used to extend the current filesystem to the end of the last
961  * existing group.  It can be accessed via ioctl, or by "remount,resize=<size>"
962  * for emergencies (because it has no dependencies on reserved blocks).
963  *
964  * If we _really_ wanted, we could use default values to call ext4_group_add()
965  * allow the "remount" trick to work for arbitrary resizing, assuming enough
966  * GDT blocks are reserved to grow to the desired size.
967  */
968 int ext4_group_extend(struct super_block *sb, struct ext4_super_block *es,
969                       ext4_fsblk_t n_blocks_count)
970 {
971         ext4_fsblk_t o_blocks_count;
972         ext4_group_t o_groups_count;
973         ext4_grpblk_t last;
974         ext4_grpblk_t add;
975         struct buffer_head *bh;
976         handle_t *handle;
977         int err;
978         unsigned long freed_blocks;
979         ext4_group_t group;
980         struct ext4_group_info *grp;
981
982         /* We don't need to worry about locking wrt other resizers just
983          * yet: we're going to revalidate es->s_blocks_count after
984          * taking lock_super() below. */
985         o_blocks_count = ext4_blocks_count(es);
986         o_groups_count = EXT4_SB(sb)->s_groups_count;
987
988         if (test_opt(sb, DEBUG))
989                 printk(KERN_DEBUG "EXT4-fs: extending last group from %llu uto %llu blocks\n",
990                        o_blocks_count, n_blocks_count);
991
992         if (n_blocks_count == 0 || n_blocks_count == o_blocks_count)
993                 return 0;
994
995         if (n_blocks_count > (sector_t)(~0ULL) >> (sb->s_blocksize_bits - 9)) {
996                 printk(KERN_ERR "EXT4-fs: filesystem on %s:"
997                         " too large to resize to %llu blocks safely\n",
998                         sb->s_id, n_blocks_count);
999                 if (sizeof(sector_t) < 8)
1000                         ext4_warning(sb, __func__,
1001                         "CONFIG_LBD not enabled\n");
1002                 return -EINVAL;
1003         }
1004
1005         if (n_blocks_count < o_blocks_count) {
1006                 ext4_warning(sb, __func__,
1007                              "can't shrink FS - resize aborted");
1008                 return -EBUSY;
1009         }
1010
1011         /* Handle the remaining blocks in the last group only. */
1012         ext4_get_group_no_and_offset(sb, o_blocks_count, &group, &last);
1013
1014         if (last == 0) {
1015                 ext4_warning(sb, __func__,
1016                              "need to use ext2online to resize further");
1017                 return -EPERM;
1018         }
1019
1020         add = EXT4_BLOCKS_PER_GROUP(sb) - last;
1021
1022         if (o_blocks_count + add < o_blocks_count) {
1023                 ext4_warning(sb, __func__, "blocks_count overflow");
1024                 return -EINVAL;
1025         }
1026
1027         if (o_blocks_count + add > n_blocks_count)
1028                 add = n_blocks_count - o_blocks_count;
1029
1030         if (o_blocks_count + add < n_blocks_count)
1031                 ext4_warning(sb, __func__,
1032                              "will only finish group (%llu"
1033                              " blocks, %u new)",
1034                              o_blocks_count + add, add);
1035
1036         /* See if the device is actually as big as what was requested */
1037         bh = sb_bread(sb, o_blocks_count + add - 1);
1038         if (!bh) {
1039                 ext4_warning(sb, __func__,
1040                              "can't read last block, resize aborted");
1041                 return -ENOSPC;
1042         }
1043         brelse(bh);
1044
1045         /* We will update the superblock, one block bitmap, and
1046          * one group descriptor via ext4_free_blocks().
1047          */
1048         handle = ext4_journal_start_sb(sb, 3);
1049         if (IS_ERR(handle)) {
1050                 err = PTR_ERR(handle);
1051                 ext4_warning(sb, __func__, "error %d on journal start", err);
1052                 goto exit_put;
1053         }
1054
1055         lock_super(sb);
1056         if (o_blocks_count != ext4_blocks_count(es)) {
1057                 ext4_warning(sb, __func__,
1058                              "multiple resizers run on filesystem!");
1059                 unlock_super(sb);
1060                 ext4_journal_stop(handle);
1061                 err = -EBUSY;
1062                 goto exit_put;
1063         }
1064
1065         if ((err = ext4_journal_get_write_access(handle,
1066                                                  EXT4_SB(sb)->s_sbh))) {
1067                 ext4_warning(sb, __func__,
1068                              "error %d on journal write access", err);
1069                 unlock_super(sb);
1070                 ext4_journal_stop(handle);
1071                 goto exit_put;
1072         }
1073         ext4_blocks_count_set(es, o_blocks_count + add);
1074         ext4_journal_dirty_metadata(handle, EXT4_SB(sb)->s_sbh);
1075         sb->s_dirt = 1;
1076         unlock_super(sb);
1077         ext4_debug("freeing blocks %llu through %llu\n", o_blocks_count,
1078                    o_blocks_count + add);
1079         ext4_free_blocks_sb(handle, sb, o_blocks_count, add, &freed_blocks);
1080         ext4_debug("freed blocks %llu through %llu\n", o_blocks_count,
1081                    o_blocks_count + add);
1082         if ((err = ext4_journal_stop(handle)))
1083                 goto exit_put;
1084
1085         /*
1086          * Mark mballoc pages as not up to date so that they will be updated
1087          * next time they are loaded by ext4_mb_load_buddy.
1088          *
1089          * XXX Bad, Bad, BAD!!!  We should not be overloading the
1090          * Uptodate flag, particularly on thte bitmap bh, as way of
1091          * hinting to ext4_mb_load_buddy() that it needs to be
1092          * overloaded.  A user could take a LVM snapshot, then do an
1093          * on-line fsck, and clear the uptodate flag, and this would
1094          * not be a bug in userspace, but a bug in the kernel.  FIXME!!!
1095          */
1096         {
1097                 struct ext4_sb_info *sbi = EXT4_SB(sb);
1098                 struct inode *inode = sbi->s_buddy_cache;
1099                 int blocks_per_page;
1100                 int block;
1101                 int pnum;
1102                 struct page *page;
1103
1104                 /* Set buddy page as not up to date */
1105                 blocks_per_page = PAGE_CACHE_SIZE / sb->s_blocksize;
1106                 block = group * 2;
1107                 pnum = block / blocks_per_page;
1108                 page = find_get_page(inode->i_mapping, pnum);
1109                 if (page != NULL) {
1110                         ClearPageUptodate(page);
1111                         page_cache_release(page);
1112                 }
1113
1114                 /* Set bitmap page as not up to date */
1115                 block++;
1116                 pnum = block / blocks_per_page;
1117                 page = find_get_page(inode->i_mapping, pnum);
1118                 if (page != NULL) {
1119                         ClearPageUptodate(page);
1120                         page_cache_release(page);
1121                 }
1122
1123                 /* Get the info on the last group */
1124                 grp = ext4_get_group_info(sb, group);
1125
1126                 /* Update free blocks in group info */
1127                 ext4_mb_update_group_info(grp, add);
1128         }
1129
1130         if (test_opt(sb, DEBUG))
1131                 printk(KERN_DEBUG "EXT4-fs: extended group to %llu blocks\n",
1132                        ext4_blocks_count(es));
1133         update_backups(sb, EXT4_SB(sb)->s_sbh->b_blocknr, (char *)es,
1134                        sizeof(struct ext4_super_block));
1135 exit_put:
1136         return err;
1137 } /* ext4_group_extend */