Merge git://git.kernel.org/pub/scm/linux/kernel/git/herbert/crypto-2.6
[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 %u 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 (ext4_handle_has_enough_credits(handle, 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_handle_dirty_metadata(handle, NULL, 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_handle_dirty_metadata(handle, NULL, 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_handle_dirty_metadata(handle, NULL, 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, sb->s_blocksize * 8, bh->b_data);
288         ext4_handle_dirty_metadata(handle, NULL, bh);
289         brelse(bh);
290         /* Mark unused entries in inode bitmap used */
291         ext4_debug("clear inode bitmap %#04llx (+%llu)\n",
292                    input->inode_bitmap, input->inode_bitmap - start);
293         if (IS_ERR(bh = bclean(handle, sb, input->inode_bitmap))) {
294                 err = PTR_ERR(bh);
295                 goto exit_journal;
296         }
297
298         mark_bitmap_end(EXT4_INODES_PER_GROUP(sb), sb->s_blocksize * 8,
299                         bh->b_data);
300         ext4_handle_dirty_metadata(handle, NULL, bh);
301 exit_bh:
302         brelse(bh);
303
304 exit_journal:
305         unlock_super(sb);
306         if ((err2 = ext4_journal_stop(handle)) && !err)
307                 err = err2;
308
309         return err;
310 }
311
312 /*
313  * Iterate through the groups which hold BACKUP superblock/GDT copies in an
314  * ext4 filesystem.  The counters should be initialized to 1, 5, and 7 before
315  * calling this for the first time.  In a sparse filesystem it will be the
316  * sequence of powers of 3, 5, and 7: 1, 3, 5, 7, 9, 25, 27, 49, 81, ...
317  * For a non-sparse filesystem it will be every group: 1, 2, 3, 4, ...
318  */
319 static unsigned ext4_list_backups(struct super_block *sb, unsigned *three,
320                                   unsigned *five, unsigned *seven)
321 {
322         unsigned *min = three;
323         int mult = 3;
324         unsigned ret;
325
326         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
327                                         EXT4_FEATURE_RO_COMPAT_SPARSE_SUPER)) {
328                 ret = *min;
329                 *min += 1;
330                 return ret;
331         }
332
333         if (*five < *min) {
334                 min = five;
335                 mult = 5;
336         }
337         if (*seven < *min) {
338                 min = seven;
339                 mult = 7;
340         }
341
342         ret = *min;
343         *min *= mult;
344
345         return ret;
346 }
347
348 /*
349  * Check that all of the backup GDT blocks are held in the primary GDT block.
350  * It is assumed that they are stored in group order.  Returns the number of
351  * groups in current filesystem that have BACKUPS, or -ve error code.
352  */
353 static int verify_reserved_gdb(struct super_block *sb,
354                                struct buffer_head *primary)
355 {
356         const ext4_fsblk_t blk = primary->b_blocknr;
357         const ext4_group_t end = EXT4_SB(sb)->s_groups_count;
358         unsigned three = 1;
359         unsigned five = 5;
360         unsigned seven = 7;
361         unsigned grp;
362         __le32 *p = (__le32 *)primary->b_data;
363         int gdbackups = 0;
364
365         while ((grp = ext4_list_backups(sb, &three, &five, &seven)) < end) {
366                 if (le32_to_cpu(*p++) !=
367                     grp * EXT4_BLOCKS_PER_GROUP(sb) + blk){
368                         ext4_warning(sb, __func__,
369                                      "reserved GDT %llu"
370                                      " missing grp %d (%llu)",
371                                      blk, grp,
372                                      grp *
373                                      (ext4_fsblk_t)EXT4_BLOCKS_PER_GROUP(sb) +
374                                      blk);
375                         return -EINVAL;
376                 }
377                 if (++gdbackups > EXT4_ADDR_PER_BLOCK(sb))
378                         return -EFBIG;
379         }
380
381         return gdbackups;
382 }
383
384 /*
385  * Called when we need to bring a reserved group descriptor table block into
386  * use from the resize inode.  The primary copy of the new GDT block currently
387  * is an indirect block (under the double indirect block in the resize inode).
388  * The new backup GDT blocks will be stored as leaf blocks in this indirect
389  * block, in group order.  Even though we know all the block numbers we need,
390  * we check to ensure that the resize inode has actually reserved these blocks.
391  *
392  * Don't need to update the block bitmaps because the blocks are still in use.
393  *
394  * We get all of the error cases out of the way, so that we are sure to not
395  * fail once we start modifying the data on disk, because JBD has no rollback.
396  */
397 static int add_new_gdb(handle_t *handle, struct inode *inode,
398                        struct ext4_new_group_data *input,
399                        struct buffer_head **primary)
400 {
401         struct super_block *sb = inode->i_sb;
402         struct ext4_super_block *es = EXT4_SB(sb)->s_es;
403         unsigned long gdb_num = input->group / EXT4_DESC_PER_BLOCK(sb);
404         ext4_fsblk_t gdblock = EXT4_SB(sb)->s_sbh->b_blocknr + 1 + gdb_num;
405         struct buffer_head **o_group_desc, **n_group_desc;
406         struct buffer_head *dind;
407         int gdbackups;
408         struct ext4_iloc iloc;
409         __le32 *data;
410         int err;
411
412         if (test_opt(sb, DEBUG))
413                 printk(KERN_DEBUG
414                        "EXT4-fs: ext4_add_new_gdb: adding group block %lu\n",
415                        gdb_num);
416
417         /*
418          * If we are not using the primary superblock/GDT copy don't resize,
419          * because the user tools have no way of handling this.  Probably a
420          * bad time to do it anyways.
421          */
422         if (EXT4_SB(sb)->s_sbh->b_blocknr !=
423             le32_to_cpu(EXT4_SB(sb)->s_es->s_first_data_block)) {
424                 ext4_warning(sb, __func__,
425                         "won't resize using backup superblock at %llu",
426                         (unsigned long long)EXT4_SB(sb)->s_sbh->b_blocknr);
427                 return -EPERM;
428         }
429
430         *primary = sb_bread(sb, gdblock);
431         if (!*primary)
432                 return -EIO;
433
434         if ((gdbackups = verify_reserved_gdb(sb, *primary)) < 0) {
435                 err = gdbackups;
436                 goto exit_bh;
437         }
438
439         data = EXT4_I(inode)->i_data + EXT4_DIND_BLOCK;
440         dind = sb_bread(sb, le32_to_cpu(*data));
441         if (!dind) {
442                 err = -EIO;
443                 goto exit_bh;
444         }
445
446         data = (__le32 *)dind->b_data;
447         if (le32_to_cpu(data[gdb_num % EXT4_ADDR_PER_BLOCK(sb)]) != gdblock) {
448                 ext4_warning(sb, __func__,
449                              "new group %u GDT block %llu not reserved",
450                              input->group, gdblock);
451                 err = -EINVAL;
452                 goto exit_dind;
453         }
454
455         if ((err = ext4_journal_get_write_access(handle, EXT4_SB(sb)->s_sbh)))
456                 goto exit_dind;
457
458         if ((err = ext4_journal_get_write_access(handle, *primary)))
459                 goto exit_sbh;
460
461         if ((err = ext4_journal_get_write_access(handle, dind)))
462                 goto exit_primary;
463
464         /* ext4_reserve_inode_write() gets a reference on the iloc */
465         if ((err = ext4_reserve_inode_write(handle, inode, &iloc)))
466                 goto exit_dindj;
467
468         n_group_desc = kmalloc((gdb_num + 1) * sizeof(struct buffer_head *),
469                         GFP_NOFS);
470         if (!n_group_desc) {
471                 err = -ENOMEM;
472                 ext4_warning(sb, __func__,
473                               "not enough memory for %lu groups", gdb_num + 1);
474                 goto exit_inode;
475         }
476
477         /*
478          * Finally, we have all of the possible failures behind us...
479          *
480          * Remove new GDT block from inode double-indirect block and clear out
481          * the new GDT block for use (which also "frees" the backup GDT blocks
482          * from the reserved inode).  We don't need to change the bitmaps for
483          * these blocks, because they are marked as in-use from being in the
484          * reserved inode, and will become GDT blocks (primary and backup).
485          */
486         data[gdb_num % EXT4_ADDR_PER_BLOCK(sb)] = 0;
487         ext4_handle_dirty_metadata(handle, NULL, dind);
488         brelse(dind);
489         inode->i_blocks -= (gdbackups + 1) * sb->s_blocksize >> 9;
490         ext4_mark_iloc_dirty(handle, inode, &iloc);
491         memset((*primary)->b_data, 0, sb->s_blocksize);
492         ext4_handle_dirty_metadata(handle, NULL, *primary);
493
494         o_group_desc = EXT4_SB(sb)->s_group_desc;
495         memcpy(n_group_desc, o_group_desc,
496                EXT4_SB(sb)->s_gdb_count * sizeof(struct buffer_head *));
497         n_group_desc[gdb_num] = *primary;
498         EXT4_SB(sb)->s_group_desc = n_group_desc;
499         EXT4_SB(sb)->s_gdb_count++;
500         kfree(o_group_desc);
501
502         le16_add_cpu(&es->s_reserved_gdt_blocks, -1);
503         ext4_handle_dirty_metadata(handle, NULL, EXT4_SB(sb)->s_sbh);
504
505         return 0;
506
507 exit_inode:
508         /* ext4_journal_release_buffer(handle, iloc.bh); */
509         brelse(iloc.bh);
510 exit_dindj:
511         /* ext4_journal_release_buffer(handle, dind); */
512 exit_primary:
513         /* ext4_journal_release_buffer(handle, *primary); */
514 exit_sbh:
515         /* ext4_journal_release_buffer(handle, *primary); */
516 exit_dind:
517         brelse(dind);
518 exit_bh:
519         brelse(*primary);
520
521         ext4_debug("leaving with error %d\n", err);
522         return err;
523 }
524
525 /*
526  * Called when we are adding a new group which has a backup copy of each of
527  * the GDT blocks (i.e. sparse group) and there are reserved GDT blocks.
528  * We need to add these reserved backup GDT blocks to the resize inode, so
529  * that they are kept for future resizing and not allocated to files.
530  *
531  * Each reserved backup GDT block will go into a different indirect block.
532  * The indirect blocks are actually the primary reserved GDT blocks,
533  * so we know in advance what their block numbers are.  We only get the
534  * double-indirect block to verify it is pointing to the primary reserved
535  * GDT blocks so we don't overwrite a data block by accident.  The reserved
536  * backup GDT blocks are stored in their reserved primary GDT block.
537  */
538 static int reserve_backup_gdb(handle_t *handle, struct inode *inode,
539                               struct ext4_new_group_data *input)
540 {
541         struct super_block *sb = inode->i_sb;
542         int reserved_gdb =le16_to_cpu(EXT4_SB(sb)->s_es->s_reserved_gdt_blocks);
543         struct buffer_head **primary;
544         struct buffer_head *dind;
545         struct ext4_iloc iloc;
546         ext4_fsblk_t blk;
547         __le32 *data, *end;
548         int gdbackups = 0;
549         int res, i;
550         int err;
551
552         primary = kmalloc(reserved_gdb * sizeof(*primary), GFP_NOFS);
553         if (!primary)
554                 return -ENOMEM;
555
556         data = EXT4_I(inode)->i_data + EXT4_DIND_BLOCK;
557         dind = sb_bread(sb, le32_to_cpu(*data));
558         if (!dind) {
559                 err = -EIO;
560                 goto exit_free;
561         }
562
563         blk = EXT4_SB(sb)->s_sbh->b_blocknr + 1 + EXT4_SB(sb)->s_gdb_count;
564         data = (__le32 *)dind->b_data + (EXT4_SB(sb)->s_gdb_count %
565                                          EXT4_ADDR_PER_BLOCK(sb));
566         end = (__le32 *)dind->b_data + EXT4_ADDR_PER_BLOCK(sb);
567
568         /* Get each reserved primary GDT block and verify it holds backups */
569         for (res = 0; res < reserved_gdb; res++, blk++) {
570                 if (le32_to_cpu(*data) != blk) {
571                         ext4_warning(sb, __func__,
572                                      "reserved block %llu"
573                                      " not at offset %ld",
574                                      blk,
575                                      (long)(data - (__le32 *)dind->b_data));
576                         err = -EINVAL;
577                         goto exit_bh;
578                 }
579                 primary[res] = sb_bread(sb, blk);
580                 if (!primary[res]) {
581                         err = -EIO;
582                         goto exit_bh;
583                 }
584                 if ((gdbackups = verify_reserved_gdb(sb, primary[res])) < 0) {
585                         brelse(primary[res]);
586                         err = gdbackups;
587                         goto exit_bh;
588                 }
589                 if (++data >= end)
590                         data = (__le32 *)dind->b_data;
591         }
592
593         for (i = 0; i < reserved_gdb; i++) {
594                 if ((err = ext4_journal_get_write_access(handle, primary[i]))) {
595                         /*
596                         int j;
597                         for (j = 0; j < i; j++)
598                                 ext4_journal_release_buffer(handle, primary[j]);
599                          */
600                         goto exit_bh;
601                 }
602         }
603
604         if ((err = ext4_reserve_inode_write(handle, inode, &iloc)))
605                 goto exit_bh;
606
607         /*
608          * Finally we can add each of the reserved backup GDT blocks from
609          * the new group to its reserved primary GDT block.
610          */
611         blk = input->group * EXT4_BLOCKS_PER_GROUP(sb);
612         for (i = 0; i < reserved_gdb; i++) {
613                 int err2;
614                 data = (__le32 *)primary[i]->b_data;
615                 /* printk("reserving backup %lu[%u] = %lu\n",
616                        primary[i]->b_blocknr, gdbackups,
617                        blk + primary[i]->b_blocknr); */
618                 data[gdbackups] = cpu_to_le32(blk + primary[i]->b_blocknr);
619                 err2 = ext4_handle_dirty_metadata(handle, NULL, primary[i]);
620                 if (!err)
621                         err = err2;
622         }
623         inode->i_blocks += reserved_gdb * sb->s_blocksize >> 9;
624         ext4_mark_iloc_dirty(handle, inode, &iloc);
625
626 exit_bh:
627         while (--res >= 0)
628                 brelse(primary[res]);
629         brelse(dind);
630
631 exit_free:
632         kfree(primary);
633
634         return err;
635 }
636
637 /*
638  * Update the backup copies of the ext4 metadata.  These don't need to be part
639  * of the main resize transaction, because e2fsck will re-write them if there
640  * is a problem (basically only OOM will cause a problem).  However, we
641  * _should_ update the backups if possible, in case the primary gets trashed
642  * for some reason and we need to run e2fsck from a backup superblock.  The
643  * important part is that the new block and inode counts are in the backup
644  * superblocks, and the location of the new group metadata in the GDT backups.
645  *
646  * We do not need lock_super() for this, because these blocks are not
647  * otherwise touched by the filesystem code when it is mounted.  We don't
648  * need to worry about last changing from sbi->s_groups_count, because the
649  * worst that can happen is that we do not copy the full number of backups
650  * at this time.  The resize which changed s_groups_count will backup again.
651  */
652 static void update_backups(struct super_block *sb,
653                            int blk_off, char *data, int size)
654 {
655         struct ext4_sb_info *sbi = EXT4_SB(sb);
656         const ext4_group_t last = sbi->s_groups_count;
657         const int bpg = EXT4_BLOCKS_PER_GROUP(sb);
658         unsigned three = 1;
659         unsigned five = 5;
660         unsigned seven = 7;
661         ext4_group_t group;
662         int rest = sb->s_blocksize - size;
663         handle_t *handle;
664         int err = 0, err2;
665
666         handle = ext4_journal_start_sb(sb, EXT4_MAX_TRANS_DATA);
667         if (IS_ERR(handle)) {
668                 group = 1;
669                 err = PTR_ERR(handle);
670                 goto exit_err;
671         }
672
673         while ((group = ext4_list_backups(sb, &three, &five, &seven)) < last) {
674                 struct buffer_head *bh;
675
676                 /* Out of journal space, and can't get more - abort - so sad */
677                 if (ext4_handle_valid(handle) &&
678                     handle->h_buffer_credits == 0 &&
679                     ext4_journal_extend(handle, EXT4_MAX_TRANS_DATA) &&
680                     (err = ext4_journal_restart(handle, EXT4_MAX_TRANS_DATA)))
681                         break;
682
683                 bh = sb_getblk(sb, group * bpg + blk_off);
684                 if (!bh) {
685                         err = -EIO;
686                         break;
687                 }
688                 ext4_debug("update metadata backup %#04lx\n",
689                           (unsigned long)bh->b_blocknr);
690                 if ((err = ext4_journal_get_write_access(handle, bh)))
691                         break;
692                 lock_buffer(bh);
693                 memcpy(bh->b_data, data, size);
694                 if (rest)
695                         memset(bh->b_data + size, 0, rest);
696                 set_buffer_uptodate(bh);
697                 unlock_buffer(bh);
698                 ext4_handle_dirty_metadata(handle, NULL, bh);
699                 brelse(bh);
700         }
701         if ((err2 = ext4_journal_stop(handle)) && !err)
702                 err = err2;
703
704         /*
705          * Ugh! Need to have e2fsck write the backup copies.  It is too
706          * late to revert the resize, we shouldn't fail just because of
707          * the backup copies (they are only needed in case of corruption).
708          *
709          * However, if we got here we have a journal problem too, so we
710          * can't really start a transaction to mark the superblock.
711          * Chicken out and just set the flag on the hope it will be written
712          * to disk, and if not - we will simply wait until next fsck.
713          */
714 exit_err:
715         if (err) {
716                 ext4_warning(sb, __func__,
717                              "can't update backup for group %u (err %d), "
718                              "forcing fsck on next reboot", group, err);
719                 sbi->s_mount_state &= ~EXT4_VALID_FS;
720                 sbi->s_es->s_state &= cpu_to_le16(~EXT4_VALID_FS);
721                 mark_buffer_dirty(sbi->s_sbh);
722         }
723 }
724
725 /* Add group descriptor data to an existing or new group descriptor block.
726  * Ensure we handle all possible error conditions _before_ we start modifying
727  * the filesystem, because we cannot abort the transaction and not have it
728  * write the data to disk.
729  *
730  * If we are on a GDT block boundary, we need to get the reserved GDT block.
731  * Otherwise, we may need to add backup GDT blocks for a sparse group.
732  *
733  * We only need to hold the superblock lock while we are actually adding
734  * in the new group's counts to the superblock.  Prior to that we have
735  * not really "added" the group at all.  We re-check that we are still
736  * adding in the last group in case things have changed since verifying.
737  */
738 int ext4_group_add(struct super_block *sb, struct ext4_new_group_data *input)
739 {
740         struct ext4_sb_info *sbi = EXT4_SB(sb);
741         struct ext4_super_block *es = sbi->s_es;
742         int reserved_gdb = ext4_bg_has_super(sb, input->group) ?
743                 le16_to_cpu(es->s_reserved_gdt_blocks) : 0;
744         struct buffer_head *primary = NULL;
745         struct ext4_group_desc *gdp;
746         struct inode *inode = NULL;
747         handle_t *handle;
748         int gdb_off, gdb_num;
749         int num_grp_locked = 0;
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");
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");
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
791         if ((err = verify_group_input(sb, input)))
792                 goto exit_put;
793
794         if ((err = setup_new_group_blocks(sb, input)))
795                 goto exit_put;
796
797         /*
798          * We will always be modifying at least the superblock and a GDT
799          * block.  If we are adding a group past the last current GDT block,
800          * we will also modify the inode and the dindirect block.  If we
801          * are adding a group with superblock/GDT backups  we will also
802          * modify each of the reserved GDT dindirect blocks.
803          */
804         handle = ext4_journal_start_sb(sb,
805                                        ext4_bg_has_super(sb, input->group) ?
806                                        3 + reserved_gdb : 4);
807         if (IS_ERR(handle)) {
808                 err = PTR_ERR(handle);
809                 goto exit_put;
810         }
811
812         lock_super(sb);
813         if (input->group != sbi->s_groups_count) {
814                 ext4_warning(sb, __func__,
815                              "multiple resizers run on filesystem!");
816                 err = -EBUSY;
817                 goto exit_journal;
818         }
819
820         if ((err = ext4_journal_get_write_access(handle, sbi->s_sbh)))
821                 goto exit_journal;
822
823         /*
824          * We will only either add reserved group blocks to a backup group
825          * or remove reserved blocks for the first group in a new group block.
826          * Doing both would be mean more complex code, and sane people don't
827          * use non-sparse filesystems anymore.  This is already checked above.
828          */
829         if (gdb_off) {
830                 primary = sbi->s_group_desc[gdb_num];
831                 if ((err = ext4_journal_get_write_access(handle, primary)))
832                         goto exit_journal;
833
834                 if (reserved_gdb && ext4_bg_num_gdb(sb, input->group) &&
835                     (err = reserve_backup_gdb(handle, inode, input)))
836                         goto exit_journal;
837         } else if ((err = add_new_gdb(handle, inode, input, &primary)))
838                 goto exit_journal;
839
840         /*
841          * OK, now we've set up the new group.  Time to make it active.
842          *
843          * Current kernels don't lock all allocations via lock_super(),
844          * so we have to be safe wrt. concurrent accesses the group
845          * data.  So we need to be careful to set all of the relevant
846          * group descriptor data etc. *before* we enable the group.
847          *
848          * The key field here is sbi->s_groups_count: as long as
849          * that retains its old value, nobody is going to access the new
850          * group.
851          *
852          * So first we update all the descriptor metadata for the new
853          * group; then we update the total disk blocks count; then we
854          * update the groups count to enable the group; then finally we
855          * update the free space counts so that the system can start
856          * using the new disk blocks.
857          */
858
859         num_grp_locked = ext4_mb_get_buddy_cache_lock(sb, input->group);
860         /* Update group descriptor block for new group */
861         gdp = (struct ext4_group_desc *)((char *)primary->b_data +
862                                          gdb_off * EXT4_DESC_SIZE(sb));
863
864         memset(gdp, 0, EXT4_DESC_SIZE(sb));
865         ext4_block_bitmap_set(sb, gdp, input->block_bitmap); /* LV FIXME */
866         ext4_inode_bitmap_set(sb, gdp, input->inode_bitmap); /* LV FIXME */
867         ext4_inode_table_set(sb, gdp, input->inode_table); /* LV FIXME */
868         ext4_free_blks_set(sb, gdp, input->free_blocks_count);
869         ext4_free_inodes_set(sb, gdp, EXT4_INODES_PER_GROUP(sb));
870         gdp->bg_flags = cpu_to_le16(EXT4_BG_INODE_ZEROED);
871         gdp->bg_checksum = ext4_group_desc_csum(sbi, input->group, gdp);
872
873         /*
874          * We can allocate memory for mb_alloc based on the new group
875          * descriptor
876          */
877         err = ext4_mb_add_groupinfo(sb, input->group, gdp);
878         if (err) {
879                 ext4_mb_put_buddy_cache_lock(sb, input->group, num_grp_locked);
880                 goto exit_journal;
881         }
882
883         /*
884          * Make the new blocks and inodes valid next.  We do this before
885          * increasing the group count so that once the group is enabled,
886          * all of its blocks and inodes are already valid.
887          *
888          * We always allocate group-by-group, then block-by-block or
889          * inode-by-inode within a group, so enabling these
890          * blocks/inodes before the group is live won't actually let us
891          * allocate the new space yet.
892          */
893         ext4_blocks_count_set(es, ext4_blocks_count(es) +
894                 input->blocks_count);
895         le32_add_cpu(&es->s_inodes_count, EXT4_INODES_PER_GROUP(sb));
896
897         /*
898          * We need to protect s_groups_count against other CPUs seeing
899          * inconsistent state in the superblock.
900          *
901          * The precise rules we use are:
902          *
903          * * Writers of s_groups_count *must* hold lock_super
904          * AND
905          * * Writers must perform a smp_wmb() after updating all dependent
906          *   data and before modifying the groups count
907          *
908          * * Readers must hold lock_super() over the access
909          * OR
910          * * Readers must perform an smp_rmb() after reading the groups count
911          *   and before reading any dependent data.
912          *
913          * NB. These rules can be relaxed when checking the group count
914          * while freeing data, as we can only allocate from a block
915          * group after serialising against the group count, and we can
916          * only then free after serialising in turn against that
917          * allocation.
918          */
919         smp_wmb();
920
921         /* Update the global fs size fields */
922         sbi->s_groups_count++;
923         ext4_mb_put_buddy_cache_lock(sb, input->group, num_grp_locked);
924
925         ext4_handle_dirty_metadata(handle, NULL, primary);
926
927         /* Update the reserved block counts only once the new group is
928          * active. */
929         ext4_r_blocks_count_set(es, ext4_r_blocks_count(es) +
930                 input->reserved_blocks);
931
932         /* Update the free space counts */
933         percpu_counter_add(&sbi->s_freeblocks_counter,
934                            input->free_blocks_count);
935         percpu_counter_add(&sbi->s_freeinodes_counter,
936                            EXT4_INODES_PER_GROUP(sb));
937
938         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_FLEX_BG)) {
939                 ext4_group_t flex_group;
940                 flex_group = ext4_flex_group(sbi, input->group);
941                 atomic_add(input->free_blocks_count,
942                            &sbi->s_flex_groups[flex_group].free_blocks);
943                 atomic_add(EXT4_INODES_PER_GROUP(sb),
944                            &sbi->s_flex_groups[flex_group].free_inodes);
945         }
946
947         ext4_handle_dirty_metadata(handle, NULL, sbi->s_sbh);
948         sb->s_dirt = 1;
949
950 exit_journal:
951         unlock_super(sb);
952         if ((err2 = ext4_journal_stop(handle)) && !err)
953                 err = err2;
954         if (!err) {
955                 update_backups(sb, sbi->s_sbh->b_blocknr, (char *)es,
956                                sizeof(struct ext4_super_block));
957                 update_backups(sb, primary->b_blocknr, primary->b_data,
958                                primary->b_size);
959         }
960 exit_put:
961         iput(inode);
962         return err;
963 } /* ext4_group_add */
964
965 /*
966  * Extend the filesystem to the new number of blocks specified.  This entry
967  * point is only used to extend the current filesystem to the end of the last
968  * existing group.  It can be accessed via ioctl, or by "remount,resize=<size>"
969  * for emergencies (because it has no dependencies on reserved blocks).
970  *
971  * If we _really_ wanted, we could use default values to call ext4_group_add()
972  * allow the "remount" trick to work for arbitrary resizing, assuming enough
973  * GDT blocks are reserved to grow to the desired size.
974  */
975 int ext4_group_extend(struct super_block *sb, struct ext4_super_block *es,
976                       ext4_fsblk_t n_blocks_count)
977 {
978         ext4_fsblk_t o_blocks_count;
979         ext4_group_t o_groups_count;
980         ext4_grpblk_t last;
981         ext4_grpblk_t add;
982         struct buffer_head *bh;
983         handle_t *handle;
984         int err;
985         ext4_group_t group;
986
987         /* We don't need to worry about locking wrt other resizers just
988          * yet: we're going to revalidate es->s_blocks_count after
989          * taking lock_super() below. */
990         o_blocks_count = ext4_blocks_count(es);
991         o_groups_count = EXT4_SB(sb)->s_groups_count;
992
993         if (test_opt(sb, DEBUG))
994                 printk(KERN_DEBUG "EXT4-fs: extending last group from %llu uto %llu blocks\n",
995                        o_blocks_count, n_blocks_count);
996
997         if (n_blocks_count == 0 || n_blocks_count == o_blocks_count)
998                 return 0;
999
1000         if (n_blocks_count > (sector_t)(~0ULL) >> (sb->s_blocksize_bits - 9)) {
1001                 printk(KERN_ERR "EXT4-fs: filesystem on %s:"
1002                         " too large to resize to %llu blocks safely\n",
1003                         sb->s_id, n_blocks_count);
1004                 if (sizeof(sector_t) < 8)
1005                         ext4_warning(sb, __func__, "CONFIG_LBD not enabled");
1006                 return -EINVAL;
1007         }
1008
1009         if (n_blocks_count < o_blocks_count) {
1010                 ext4_warning(sb, __func__,
1011                              "can't shrink FS - resize aborted");
1012                 return -EBUSY;
1013         }
1014
1015         /* Handle the remaining blocks in the last group only. */
1016         ext4_get_group_no_and_offset(sb, o_blocks_count, &group, &last);
1017
1018         if (last == 0) {
1019                 ext4_warning(sb, __func__,
1020                              "need to use ext2online to resize further");
1021                 return -EPERM;
1022         }
1023
1024         add = EXT4_BLOCKS_PER_GROUP(sb) - last;
1025
1026         if (o_blocks_count + add < o_blocks_count) {
1027                 ext4_warning(sb, __func__, "blocks_count overflow");
1028                 return -EINVAL;
1029         }
1030
1031         if (o_blocks_count + add > n_blocks_count)
1032                 add = n_blocks_count - o_blocks_count;
1033
1034         if (o_blocks_count + add < n_blocks_count)
1035                 ext4_warning(sb, __func__,
1036                              "will only finish group (%llu"
1037                              " blocks, %u new)",
1038                              o_blocks_count + add, add);
1039
1040         /* See if the device is actually as big as what was requested */
1041         bh = sb_bread(sb, o_blocks_count + add - 1);
1042         if (!bh) {
1043                 ext4_warning(sb, __func__,
1044                              "can't read last block, resize aborted");
1045                 return -ENOSPC;
1046         }
1047         brelse(bh);
1048
1049         /* We will update the superblock, one block bitmap, and
1050          * one group descriptor via ext4_free_blocks().
1051          */
1052         handle = ext4_journal_start_sb(sb, 3);
1053         if (IS_ERR(handle)) {
1054                 err = PTR_ERR(handle);
1055                 ext4_warning(sb, __func__, "error %d on journal start", err);
1056                 goto exit_put;
1057         }
1058
1059         lock_super(sb);
1060         if (o_blocks_count != ext4_blocks_count(es)) {
1061                 ext4_warning(sb, __func__,
1062                              "multiple resizers run on filesystem!");
1063                 unlock_super(sb);
1064                 ext4_journal_stop(handle);
1065                 err = -EBUSY;
1066                 goto exit_put;
1067         }
1068
1069         if ((err = ext4_journal_get_write_access(handle,
1070                                                  EXT4_SB(sb)->s_sbh))) {
1071                 ext4_warning(sb, __func__,
1072                              "error %d on journal write access", err);
1073                 unlock_super(sb);
1074                 ext4_journal_stop(handle);
1075                 goto exit_put;
1076         }
1077         ext4_blocks_count_set(es, o_blocks_count + add);
1078         ext4_handle_dirty_metadata(handle, NULL, EXT4_SB(sb)->s_sbh);
1079         sb->s_dirt = 1;
1080         unlock_super(sb);
1081         ext4_debug("freeing blocks %llu through %llu\n", o_blocks_count,
1082                    o_blocks_count + add);
1083         /* We add the blocks to the bitmap and set the group need init bit */
1084         ext4_add_groupblocks(handle, sb, o_blocks_count, add);
1085         ext4_debug("freed blocks %llu through %llu\n", o_blocks_count,
1086                    o_blocks_count + add);
1087         if ((err = ext4_journal_stop(handle)))
1088                 goto exit_put;
1089
1090         if (test_opt(sb, DEBUG))
1091                 printk(KERN_DEBUG "EXT4-fs: extended group to %llu blocks\n",
1092                        ext4_blocks_count(es));
1093         update_backups(sb, EXT4_SB(sb)->s_sbh->b_blocknr, (char *)es,
1094                        sizeof(struct ext4_super_block));
1095 exit_put:
1096         return err;
1097 } /* ext4_group_extend */