NFSv4: Eliminate nfs4_path_walk()
[linux-2.6.git] / fs / ext3 / resize.c
1 /*
2  *  linux/fs/ext3/resize.c
3  *
4  * Support for resizing an ext3 filesystem while it is mounted.
5  *
6  * Copyright (C) 2001, 2002 Andreas Dilger <adilger@clusterfs.com>
7  *
8  * This could probably be made into a module, because it is not often in use.
9  */
10
11
12 #define EXT3FS_DEBUG
13
14 #include <linux/ext3_jbd.h>
15
16 #include <linux/errno.h>
17 #include <linux/slab.h>
18
19
20 #define outside(b, first, last) ((b) < (first) || (b) >= (last))
21 #define inside(b, first, last)  ((b) >= (first) && (b) < (last))
22
23 static int verify_group_input(struct super_block *sb,
24                               struct ext3_new_group_data *input)
25 {
26         struct ext3_sb_info *sbi = EXT3_SB(sb);
27         struct ext3_super_block *es = sbi->s_es;
28         ext3_fsblk_t start = le32_to_cpu(es->s_blocks_count);
29         ext3_fsblk_t end = start + input->blocks_count;
30         unsigned group = input->group;
31         ext3_fsblk_t itend = input->inode_table + sbi->s_itb_per_group;
32         unsigned overhead = ext3_bg_has_super(sb, group) ?
33                 (1 + ext3_bg_num_gdb(sb, group) +
34                  le16_to_cpu(es->s_reserved_gdt_blocks)) : 0;
35         ext3_fsblk_t metaend = start + overhead;
36         struct buffer_head *bh = NULL;
37         ext3_grpblk_t free_blocks_count;
38         int err = -EINVAL;
39
40         input->free_blocks_count = free_blocks_count =
41                 input->blocks_count - 2 - overhead - sbi->s_itb_per_group;
42
43         if (test_opt(sb, DEBUG))
44                 printk(KERN_DEBUG "EXT3-fs: adding %s group %u: %u blocks "
45                        "(%d free, %u reserved)\n",
46                        ext3_bg_has_super(sb, input->group) ? "normal" :
47                        "no-super", input->group, input->blocks_count,
48                        free_blocks_count, input->reserved_blocks);
49
50         if (group != sbi->s_groups_count)
51                 ext3_warning(sb, __func__,
52                              "Cannot add at group %u (only %lu groups)",
53                              input->group, sbi->s_groups_count);
54         else if ((start - le32_to_cpu(es->s_first_data_block)) %
55                  EXT3_BLOCKS_PER_GROUP(sb))
56                 ext3_warning(sb, __func__, "Last group not full");
57         else if (input->reserved_blocks > input->blocks_count / 5)
58                 ext3_warning(sb, __func__, "Reserved blocks too high (%u)",
59                              input->reserved_blocks);
60         else if (free_blocks_count < 0)
61                 ext3_warning(sb, __func__, "Bad blocks count %u",
62                              input->blocks_count);
63         else if (!(bh = sb_bread(sb, end - 1)))
64                 ext3_warning(sb, __func__,
65                              "Cannot read last block ("E3FSBLK")",
66                              end - 1);
67         else if (outside(input->block_bitmap, start, end))
68                 ext3_warning(sb, __func__,
69                              "Block bitmap not in group (block %u)",
70                              input->block_bitmap);
71         else if (outside(input->inode_bitmap, start, end))
72                 ext3_warning(sb, __func__,
73                              "Inode bitmap not in group (block %u)",
74                              input->inode_bitmap);
75         else if (outside(input->inode_table, start, end) ||
76                  outside(itend - 1, start, end))
77                 ext3_warning(sb, __func__,
78                              "Inode table not in group (blocks %u-"E3FSBLK")",
79                              input->inode_table, itend - 1);
80         else if (input->inode_bitmap == input->block_bitmap)
81                 ext3_warning(sb, __func__,
82                              "Block bitmap same as inode bitmap (%u)",
83                              input->block_bitmap);
84         else if (inside(input->block_bitmap, input->inode_table, itend))
85                 ext3_warning(sb, __func__,
86                              "Block bitmap (%u) in inode table (%u-"E3FSBLK")",
87                              input->block_bitmap, input->inode_table, itend-1);
88         else if (inside(input->inode_bitmap, input->inode_table, itend))
89                 ext3_warning(sb, __func__,
90                              "Inode bitmap (%u) in inode table (%u-"E3FSBLK")",
91                              input->inode_bitmap, input->inode_table, itend-1);
92         else if (inside(input->block_bitmap, start, metaend))
93                 ext3_warning(sb, __func__,
94                              "Block bitmap (%u) in GDT table"
95                              " ("E3FSBLK"-"E3FSBLK")",
96                              input->block_bitmap, start, metaend - 1);
97         else if (inside(input->inode_bitmap, start, metaend))
98                 ext3_warning(sb, __func__,
99                              "Inode bitmap (%u) in GDT table"
100                              " ("E3FSBLK"-"E3FSBLK")",
101                              input->inode_bitmap, start, metaend - 1);
102         else if (inside(input->inode_table, start, metaend) ||
103                  inside(itend - 1, start, metaend))
104                 ext3_warning(sb, __func__,
105                              "Inode table (%u-"E3FSBLK") overlaps"
106                              "GDT table ("E3FSBLK"-"E3FSBLK")",
107                              input->inode_table, itend - 1, start, metaend - 1);
108         else
109                 err = 0;
110         brelse(bh);
111
112         return err;
113 }
114
115 static struct buffer_head *bclean(handle_t *handle, struct super_block *sb,
116                                   ext3_fsblk_t blk)
117 {
118         struct buffer_head *bh;
119         int err;
120
121         bh = sb_getblk(sb, blk);
122         if (!bh)
123                 return ERR_PTR(-EIO);
124         if ((err = ext3_journal_get_write_access(handle, bh))) {
125                 brelse(bh);
126                 bh = ERR_PTR(err);
127         } else {
128                 lock_buffer(bh);
129                 memset(bh->b_data, 0, sb->s_blocksize);
130                 set_buffer_uptodate(bh);
131                 unlock_buffer(bh);
132         }
133
134         return bh;
135 }
136
137 /*
138  * To avoid calling the atomic setbit hundreds or thousands of times, we only
139  * need to use it within a single byte (to ensure we get endianness right).
140  * We can use memset for the rest of the bitmap as there are no other users.
141  */
142 static void mark_bitmap_end(int start_bit, int end_bit, char *bitmap)
143 {
144         int i;
145
146         if (start_bit >= end_bit)
147                 return;
148
149         ext3_debug("mark end bits +%d through +%d used\n", start_bit, end_bit);
150         for (i = start_bit; i < ((start_bit + 7) & ~7UL); i++)
151                 ext3_set_bit(i, bitmap);
152         if (i < end_bit)
153                 memset(bitmap + (i >> 3), 0xff, (end_bit - i) >> 3);
154 }
155
156 /*
157  * If we have fewer than thresh credits, extend by EXT3_MAX_TRANS_DATA.
158  * If that fails, restart the transaction & regain write access for the
159  * buffer head which is used for block_bitmap modifications.
160  */
161 static int extend_or_restart_transaction(handle_t *handle, int thresh,
162                                          struct buffer_head *bh)
163 {
164         int err;
165
166         if (handle->h_buffer_credits >= thresh)
167                 return 0;
168
169         err = ext3_journal_extend(handle, EXT3_MAX_TRANS_DATA);
170         if (err < 0)
171                 return err;
172         if (err) {
173                 err = ext3_journal_restart(handle, EXT3_MAX_TRANS_DATA);
174                 if (err)
175                         return err;
176                 err = ext3_journal_get_write_access(handle, bh);
177                 if (err)
178                         return err;
179         }
180
181         return 0;
182 }
183
184 /*
185  * Set up the block and inode bitmaps, and the inode table for the new group.
186  * This doesn't need to be part of the main transaction, since we are only
187  * changing blocks outside the actual filesystem.  We still do journaling to
188  * ensure the recovery is correct in case of a failure just after resize.
189  * If any part of this fails, we simply abort the resize.
190  */
191 static int setup_new_group_blocks(struct super_block *sb,
192                                   struct ext3_new_group_data *input)
193 {
194         struct ext3_sb_info *sbi = EXT3_SB(sb);
195         ext3_fsblk_t start = ext3_group_first_block_no(sb, input->group);
196         int reserved_gdb = ext3_bg_has_super(sb, input->group) ?
197                 le16_to_cpu(sbi->s_es->s_reserved_gdt_blocks) : 0;
198         unsigned long gdblocks = ext3_bg_num_gdb(sb, input->group);
199         struct buffer_head *bh;
200         handle_t *handle;
201         ext3_fsblk_t block;
202         ext3_grpblk_t bit;
203         int i;
204         int err = 0, err2;
205
206         /* This transaction may be extended/restarted along the way */
207         handle = ext3_journal_start_sb(sb, EXT3_MAX_TRANS_DATA);
208
209         if (IS_ERR(handle))
210                 return PTR_ERR(handle);
211
212         mutex_lock(&sbi->s_resize_lock);
213         if (input->group != sbi->s_groups_count) {
214                 err = -EBUSY;
215                 goto exit_journal;
216         }
217
218         if (IS_ERR(bh = bclean(handle, sb, input->block_bitmap))) {
219                 err = PTR_ERR(bh);
220                 goto exit_journal;
221         }
222
223         if (ext3_bg_has_super(sb, input->group)) {
224                 ext3_debug("mark backup superblock %#04lx (+0)\n", start);
225                 ext3_set_bit(0, bh->b_data);
226         }
227
228         /* Copy all of the GDT blocks into the backup in this group */
229         for (i = 0, bit = 1, block = start + 1;
230              i < gdblocks; i++, block++, bit++) {
231                 struct buffer_head *gdb;
232
233                 ext3_debug("update backup group %#04lx (+%d)\n", block, bit);
234
235                 err = extend_or_restart_transaction(handle, 1, bh);
236                 if (err)
237                         goto exit_bh;
238
239                 gdb = sb_getblk(sb, block);
240                 if (!gdb) {
241                         err = -EIO;
242                         goto exit_bh;
243                 }
244                 if ((err = ext3_journal_get_write_access(handle, gdb))) {
245                         brelse(gdb);
246                         goto exit_bh;
247                 }
248                 lock_buffer(gdb);
249                 memcpy(gdb->b_data, sbi->s_group_desc[i]->b_data, gdb->b_size);
250                 set_buffer_uptodate(gdb);
251                 unlock_buffer(gdb);
252                 ext3_journal_dirty_metadata(handle, gdb);
253                 ext3_set_bit(bit, bh->b_data);
254                 brelse(gdb);
255         }
256
257         /* Zero out all of the reserved backup group descriptor table blocks */
258         for (i = 0, bit = gdblocks + 1, block = start + bit;
259              i < reserved_gdb; i++, block++, bit++) {
260                 struct buffer_head *gdb;
261
262                 ext3_debug("clear reserved block %#04lx (+%d)\n", block, bit);
263
264                 err = extend_or_restart_transaction(handle, 1, bh);
265                 if (err)
266                         goto exit_bh;
267
268                 if (IS_ERR(gdb = bclean(handle, sb, block))) {
269                         err = PTR_ERR(gdb);
270                         goto exit_bh;
271                 }
272                 ext3_journal_dirty_metadata(handle, gdb);
273                 ext3_set_bit(bit, bh->b_data);
274                 brelse(gdb);
275         }
276         ext3_debug("mark block bitmap %#04x (+%ld)\n", input->block_bitmap,
277                    input->block_bitmap - start);
278         ext3_set_bit(input->block_bitmap - start, bh->b_data);
279         ext3_debug("mark inode bitmap %#04x (+%ld)\n", input->inode_bitmap,
280                    input->inode_bitmap - start);
281         ext3_set_bit(input->inode_bitmap - start, bh->b_data);
282
283         /* Zero out all of the inode table blocks */
284         for (i = 0, block = input->inode_table, bit = block - start;
285              i < sbi->s_itb_per_group; i++, bit++, block++) {
286                 struct buffer_head *it;
287
288                 ext3_debug("clear inode block %#04lx (+%d)\n", block, bit);
289
290                 err = extend_or_restart_transaction(handle, 1, bh);
291                 if (err)
292                         goto exit_bh;
293
294                 if (IS_ERR(it = bclean(handle, sb, block))) {
295                         err = PTR_ERR(it);
296                         goto exit_bh;
297                 }
298                 ext3_journal_dirty_metadata(handle, it);
299                 brelse(it);
300                 ext3_set_bit(bit, bh->b_data);
301         }
302
303         err = extend_or_restart_transaction(handle, 2, bh);
304         if (err)
305                 goto exit_bh;
306
307         mark_bitmap_end(input->blocks_count, EXT3_BLOCKS_PER_GROUP(sb),
308                         bh->b_data);
309         ext3_journal_dirty_metadata(handle, bh);
310         brelse(bh);
311
312         /* Mark unused entries in inode bitmap used */
313         ext3_debug("clear inode bitmap %#04x (+%ld)\n",
314                    input->inode_bitmap, input->inode_bitmap - start);
315         if (IS_ERR(bh = bclean(handle, sb, input->inode_bitmap))) {
316                 err = PTR_ERR(bh);
317                 goto exit_journal;
318         }
319
320         mark_bitmap_end(EXT3_INODES_PER_GROUP(sb), EXT3_BLOCKS_PER_GROUP(sb),
321                         bh->b_data);
322         ext3_journal_dirty_metadata(handle, bh);
323 exit_bh:
324         brelse(bh);
325
326 exit_journal:
327         mutex_unlock(&sbi->s_resize_lock);
328         if ((err2 = ext3_journal_stop(handle)) && !err)
329                 err = err2;
330
331         return err;
332 }
333
334 /*
335  * Iterate through the groups which hold BACKUP superblock/GDT copies in an
336  * ext3 filesystem.  The counters should be initialized to 1, 5, and 7 before
337  * calling this for the first time.  In a sparse filesystem it will be the
338  * sequence of powers of 3, 5, and 7: 1, 3, 5, 7, 9, 25, 27, 49, 81, ...
339  * For a non-sparse filesystem it will be every group: 1, 2, 3, 4, ...
340  */
341 static unsigned ext3_list_backups(struct super_block *sb, unsigned *three,
342                                   unsigned *five, unsigned *seven)
343 {
344         unsigned *min = three;
345         int mult = 3;
346         unsigned ret;
347
348         if (!EXT3_HAS_RO_COMPAT_FEATURE(sb,
349                                         EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER)) {
350                 ret = *min;
351                 *min += 1;
352                 return ret;
353         }
354
355         if (*five < *min) {
356                 min = five;
357                 mult = 5;
358         }
359         if (*seven < *min) {
360                 min = seven;
361                 mult = 7;
362         }
363
364         ret = *min;
365         *min *= mult;
366
367         return ret;
368 }
369
370 /*
371  * Check that all of the backup GDT blocks are held in the primary GDT block.
372  * It is assumed that they are stored in group order.  Returns the number of
373  * groups in current filesystem that have BACKUPS, or -ve error code.
374  */
375 static int verify_reserved_gdb(struct super_block *sb,
376                                struct buffer_head *primary)
377 {
378         const ext3_fsblk_t blk = primary->b_blocknr;
379         const unsigned long end = EXT3_SB(sb)->s_groups_count;
380         unsigned three = 1;
381         unsigned five = 5;
382         unsigned seven = 7;
383         unsigned grp;
384         __le32 *p = (__le32 *)primary->b_data;
385         int gdbackups = 0;
386
387         while ((grp = ext3_list_backups(sb, &three, &five, &seven)) < end) {
388                 if (le32_to_cpu(*p++) != grp * EXT3_BLOCKS_PER_GROUP(sb) + blk){
389                         ext3_warning(sb, __func__,
390                                      "reserved GDT "E3FSBLK
391                                      " missing grp %d ("E3FSBLK")",
392                                      blk, grp,
393                                      grp * EXT3_BLOCKS_PER_GROUP(sb) + blk);
394                         return -EINVAL;
395                 }
396                 if (++gdbackups > EXT3_ADDR_PER_BLOCK(sb))
397                         return -EFBIG;
398         }
399
400         return gdbackups;
401 }
402
403 /*
404  * Called when we need to bring a reserved group descriptor table block into
405  * use from the resize inode.  The primary copy of the new GDT block currently
406  * is an indirect block (under the double indirect block in the resize inode).
407  * The new backup GDT blocks will be stored as leaf blocks in this indirect
408  * block, in group order.  Even though we know all the block numbers we need,
409  * we check to ensure that the resize inode has actually reserved these blocks.
410  *
411  * Don't need to update the block bitmaps because the blocks are still in use.
412  *
413  * We get all of the error cases out of the way, so that we are sure to not
414  * fail once we start modifying the data on disk, because JBD has no rollback.
415  */
416 static int add_new_gdb(handle_t *handle, struct inode *inode,
417                        struct ext3_new_group_data *input,
418                        struct buffer_head **primary)
419 {
420         struct super_block *sb = inode->i_sb;
421         struct ext3_super_block *es = EXT3_SB(sb)->s_es;
422         unsigned long gdb_num = input->group / EXT3_DESC_PER_BLOCK(sb);
423         ext3_fsblk_t gdblock = EXT3_SB(sb)->s_sbh->b_blocknr + 1 + gdb_num;
424         struct buffer_head **o_group_desc, **n_group_desc;
425         struct buffer_head *dind;
426         int gdbackups;
427         struct ext3_iloc iloc;
428         __le32 *data;
429         int err;
430
431         if (test_opt(sb, DEBUG))
432                 printk(KERN_DEBUG
433                        "EXT3-fs: ext3_add_new_gdb: adding group block %lu\n",
434                        gdb_num);
435
436         /*
437          * If we are not using the primary superblock/GDT copy don't resize,
438          * because the user tools have no way of handling this.  Probably a
439          * bad time to do it anyways.
440          */
441         if (EXT3_SB(sb)->s_sbh->b_blocknr !=
442             le32_to_cpu(EXT3_SB(sb)->s_es->s_first_data_block)) {
443                 ext3_warning(sb, __func__,
444                         "won't resize using backup superblock at %llu",
445                         (unsigned long long)EXT3_SB(sb)->s_sbh->b_blocknr);
446                 return -EPERM;
447         }
448
449         *primary = sb_bread(sb, gdblock);
450         if (!*primary)
451                 return -EIO;
452
453         if ((gdbackups = verify_reserved_gdb(sb, *primary)) < 0) {
454                 err = gdbackups;
455                 goto exit_bh;
456         }
457
458         data = EXT3_I(inode)->i_data + EXT3_DIND_BLOCK;
459         dind = sb_bread(sb, le32_to_cpu(*data));
460         if (!dind) {
461                 err = -EIO;
462                 goto exit_bh;
463         }
464
465         data = (__le32 *)dind->b_data;
466         if (le32_to_cpu(data[gdb_num % EXT3_ADDR_PER_BLOCK(sb)]) != gdblock) {
467                 ext3_warning(sb, __func__,
468                              "new group %u GDT block "E3FSBLK" not reserved",
469                              input->group, gdblock);
470                 err = -EINVAL;
471                 goto exit_dind;
472         }
473
474         if ((err = ext3_journal_get_write_access(handle, EXT3_SB(sb)->s_sbh)))
475                 goto exit_dind;
476
477         if ((err = ext3_journal_get_write_access(handle, *primary)))
478                 goto exit_sbh;
479
480         if ((err = ext3_journal_get_write_access(handle, dind)))
481                 goto exit_primary;
482
483         /* ext3_reserve_inode_write() gets a reference on the iloc */
484         if ((err = ext3_reserve_inode_write(handle, inode, &iloc)))
485                 goto exit_dindj;
486
487         n_group_desc = kmalloc((gdb_num + 1) * sizeof(struct buffer_head *),
488                         GFP_NOFS);
489         if (!n_group_desc) {
490                 err = -ENOMEM;
491                 ext3_warning (sb, __func__,
492                               "not enough memory for %lu groups", gdb_num + 1);
493                 goto exit_inode;
494         }
495
496         /*
497          * Finally, we have all of the possible failures behind us...
498          *
499          * Remove new GDT block from inode double-indirect block and clear out
500          * the new GDT block for use (which also "frees" the backup GDT blocks
501          * from the reserved inode).  We don't need to change the bitmaps for
502          * these blocks, because they are marked as in-use from being in the
503          * reserved inode, and will become GDT blocks (primary and backup).
504          */
505         data[gdb_num % EXT3_ADDR_PER_BLOCK(sb)] = 0;
506         ext3_journal_dirty_metadata(handle, dind);
507         brelse(dind);
508         inode->i_blocks -= (gdbackups + 1) * sb->s_blocksize >> 9;
509         ext3_mark_iloc_dirty(handle, inode, &iloc);
510         memset((*primary)->b_data, 0, sb->s_blocksize);
511         ext3_journal_dirty_metadata(handle, *primary);
512
513         o_group_desc = EXT3_SB(sb)->s_group_desc;
514         memcpy(n_group_desc, o_group_desc,
515                EXT3_SB(sb)->s_gdb_count * sizeof(struct buffer_head *));
516         n_group_desc[gdb_num] = *primary;
517         EXT3_SB(sb)->s_group_desc = n_group_desc;
518         EXT3_SB(sb)->s_gdb_count++;
519         kfree(o_group_desc);
520
521         le16_add_cpu(&es->s_reserved_gdt_blocks, -1);
522         ext3_journal_dirty_metadata(handle, EXT3_SB(sb)->s_sbh);
523
524         return 0;
525
526 exit_inode:
527         //ext3_journal_release_buffer(handle, iloc.bh);
528         brelse(iloc.bh);
529 exit_dindj:
530         //ext3_journal_release_buffer(handle, dind);
531 exit_primary:
532         //ext3_journal_release_buffer(handle, *primary);
533 exit_sbh:
534         //ext3_journal_release_buffer(handle, *primary);
535 exit_dind:
536         brelse(dind);
537 exit_bh:
538         brelse(*primary);
539
540         ext3_debug("leaving with error %d\n", err);
541         return err;
542 }
543
544 /*
545  * Called when we are adding a new group which has a backup copy of each of
546  * the GDT blocks (i.e. sparse group) and there are reserved GDT blocks.
547  * We need to add these reserved backup GDT blocks to the resize inode, so
548  * that they are kept for future resizing and not allocated to files.
549  *
550  * Each reserved backup GDT block will go into a different indirect block.
551  * The indirect blocks are actually the primary reserved GDT blocks,
552  * so we know in advance what their block numbers are.  We only get the
553  * double-indirect block to verify it is pointing to the primary reserved
554  * GDT blocks so we don't overwrite a data block by accident.  The reserved
555  * backup GDT blocks are stored in their reserved primary GDT block.
556  */
557 static int reserve_backup_gdb(handle_t *handle, struct inode *inode,
558                               struct ext3_new_group_data *input)
559 {
560         struct super_block *sb = inode->i_sb;
561         int reserved_gdb =le16_to_cpu(EXT3_SB(sb)->s_es->s_reserved_gdt_blocks);
562         struct buffer_head **primary;
563         struct buffer_head *dind;
564         struct ext3_iloc iloc;
565         ext3_fsblk_t blk;
566         __le32 *data, *end;
567         int gdbackups = 0;
568         int res, i;
569         int err;
570
571         primary = kmalloc(reserved_gdb * sizeof(*primary), GFP_NOFS);
572         if (!primary)
573                 return -ENOMEM;
574
575         data = EXT3_I(inode)->i_data + EXT3_DIND_BLOCK;
576         dind = sb_bread(sb, le32_to_cpu(*data));
577         if (!dind) {
578                 err = -EIO;
579                 goto exit_free;
580         }
581
582         blk = EXT3_SB(sb)->s_sbh->b_blocknr + 1 + EXT3_SB(sb)->s_gdb_count;
583         data = (__le32 *)dind->b_data + (EXT3_SB(sb)->s_gdb_count %
584                                          EXT3_ADDR_PER_BLOCK(sb));
585         end = (__le32 *)dind->b_data + EXT3_ADDR_PER_BLOCK(sb);
586
587         /* Get each reserved primary GDT block and verify it holds backups */
588         for (res = 0; res < reserved_gdb; res++, blk++) {
589                 if (le32_to_cpu(*data) != blk) {
590                         ext3_warning(sb, __func__,
591                                      "reserved block "E3FSBLK
592                                      " not at offset %ld",
593                                      blk,
594                                      (long)(data - (__le32 *)dind->b_data));
595                         err = -EINVAL;
596                         goto exit_bh;
597                 }
598                 primary[res] = sb_bread(sb, blk);
599                 if (!primary[res]) {
600                         err = -EIO;
601                         goto exit_bh;
602                 }
603                 if ((gdbackups = verify_reserved_gdb(sb, primary[res])) < 0) {
604                         brelse(primary[res]);
605                         err = gdbackups;
606                         goto exit_bh;
607                 }
608                 if (++data >= end)
609                         data = (__le32 *)dind->b_data;
610         }
611
612         for (i = 0; i < reserved_gdb; i++) {
613                 if ((err = ext3_journal_get_write_access(handle, primary[i]))) {
614                         /*
615                         int j;
616                         for (j = 0; j < i; j++)
617                                 ext3_journal_release_buffer(handle, primary[j]);
618                          */
619                         goto exit_bh;
620                 }
621         }
622
623         if ((err = ext3_reserve_inode_write(handle, inode, &iloc)))
624                 goto exit_bh;
625
626         /*
627          * Finally we can add each of the reserved backup GDT blocks from
628          * the new group to its reserved primary GDT block.
629          */
630         blk = input->group * EXT3_BLOCKS_PER_GROUP(sb);
631         for (i = 0; i < reserved_gdb; i++) {
632                 int err2;
633                 data = (__le32 *)primary[i]->b_data;
634                 /* printk("reserving backup %lu[%u] = %lu\n",
635                        primary[i]->b_blocknr, gdbackups,
636                        blk + primary[i]->b_blocknr); */
637                 data[gdbackups] = cpu_to_le32(blk + primary[i]->b_blocknr);
638                 err2 = ext3_journal_dirty_metadata(handle, primary[i]);
639                 if (!err)
640                         err = err2;
641         }
642         inode->i_blocks += reserved_gdb * sb->s_blocksize >> 9;
643         ext3_mark_iloc_dirty(handle, inode, &iloc);
644
645 exit_bh:
646         while (--res >= 0)
647                 brelse(primary[res]);
648         brelse(dind);
649
650 exit_free:
651         kfree(primary);
652
653         return err;
654 }
655
656 /*
657  * Update the backup copies of the ext3 metadata.  These don't need to be part
658  * of the main resize transaction, because e2fsck will re-write them if there
659  * is a problem (basically only OOM will cause a problem).  However, we
660  * _should_ update the backups if possible, in case the primary gets trashed
661  * for some reason and we need to run e2fsck from a backup superblock.  The
662  * important part is that the new block and inode counts are in the backup
663  * superblocks, and the location of the new group metadata in the GDT backups.
664  *
665  * We do not need take the s_resize_lock for this, because these
666  * blocks are not otherwise touched by the filesystem code when it is
667  * mounted.  We don't need to worry about last changing from
668  * sbi->s_groups_count, because the worst that can happen is that we
669  * do not copy the full number of backups at this time.  The resize
670  * which changed s_groups_count will backup again.
671  */
672 static void update_backups(struct super_block *sb,
673                            int blk_off, char *data, int size)
674 {
675         struct ext3_sb_info *sbi = EXT3_SB(sb);
676         const unsigned long last = sbi->s_groups_count;
677         const int bpg = EXT3_BLOCKS_PER_GROUP(sb);
678         unsigned three = 1;
679         unsigned five = 5;
680         unsigned seven = 7;
681         unsigned group;
682         int rest = sb->s_blocksize - size;
683         handle_t *handle;
684         int err = 0, err2;
685
686         handle = ext3_journal_start_sb(sb, EXT3_MAX_TRANS_DATA);
687         if (IS_ERR(handle)) {
688                 group = 1;
689                 err = PTR_ERR(handle);
690                 goto exit_err;
691         }
692
693         while ((group = ext3_list_backups(sb, &three, &five, &seven)) < last) {
694                 struct buffer_head *bh;
695
696                 /* Out of journal space, and can't get more - abort - so sad */
697                 if (handle->h_buffer_credits == 0 &&
698                     ext3_journal_extend(handle, EXT3_MAX_TRANS_DATA) &&
699                     (err = ext3_journal_restart(handle, EXT3_MAX_TRANS_DATA)))
700                         break;
701
702                 bh = sb_getblk(sb, group * bpg + blk_off);
703                 if (!bh) {
704                         err = -EIO;
705                         break;
706                 }
707                 ext3_debug("update metadata backup %#04lx\n",
708                           (unsigned long)bh->b_blocknr);
709                 if ((err = ext3_journal_get_write_access(handle, bh)))
710                         break;
711                 lock_buffer(bh);
712                 memcpy(bh->b_data, data, size);
713                 if (rest)
714                         memset(bh->b_data + size, 0, rest);
715                 set_buffer_uptodate(bh);
716                 unlock_buffer(bh);
717                 ext3_journal_dirty_metadata(handle, bh);
718                 brelse(bh);
719         }
720         if ((err2 = ext3_journal_stop(handle)) && !err)
721                 err = err2;
722
723         /*
724          * Ugh! Need to have e2fsck write the backup copies.  It is too
725          * late to revert the resize, we shouldn't fail just because of
726          * the backup copies (they are only needed in case of corruption).
727          *
728          * However, if we got here we have a journal problem too, so we
729          * can't really start a transaction to mark the superblock.
730          * Chicken out and just set the flag on the hope it will be written
731          * to disk, and if not - we will simply wait until next fsck.
732          */
733 exit_err:
734         if (err) {
735                 ext3_warning(sb, __func__,
736                              "can't update backup for group %d (err %d), "
737                              "forcing fsck on next reboot", group, err);
738                 sbi->s_mount_state &= ~EXT3_VALID_FS;
739                 sbi->s_es->s_state &= cpu_to_le16(~EXT3_VALID_FS);
740                 mark_buffer_dirty(sbi->s_sbh);
741         }
742 }
743
744 /* Add group descriptor data to an existing or new group descriptor block.
745  * Ensure we handle all possible error conditions _before_ we start modifying
746  * the filesystem, because we cannot abort the transaction and not have it
747  * write the data to disk.
748  *
749  * If we are on a GDT block boundary, we need to get the reserved GDT block.
750  * Otherwise, we may need to add backup GDT blocks for a sparse group.
751  *
752  * We only need to hold the superblock lock while we are actually adding
753  * in the new group's counts to the superblock.  Prior to that we have
754  * not really "added" the group at all.  We re-check that we are still
755  * adding in the last group in case things have changed since verifying.
756  */
757 int ext3_group_add(struct super_block *sb, struct ext3_new_group_data *input)
758 {
759         struct ext3_sb_info *sbi = EXT3_SB(sb);
760         struct ext3_super_block *es = sbi->s_es;
761         int reserved_gdb = ext3_bg_has_super(sb, input->group) ?
762                 le16_to_cpu(es->s_reserved_gdt_blocks) : 0;
763         struct buffer_head *primary = NULL;
764         struct ext3_group_desc *gdp;
765         struct inode *inode = NULL;
766         handle_t *handle;
767         int gdb_off, gdb_num;
768         int err, err2;
769
770         gdb_num = input->group / EXT3_DESC_PER_BLOCK(sb);
771         gdb_off = input->group % EXT3_DESC_PER_BLOCK(sb);
772
773         if (gdb_off == 0 && !EXT3_HAS_RO_COMPAT_FEATURE(sb,
774                                         EXT3_FEATURE_RO_COMPAT_SPARSE_SUPER)) {
775                 ext3_warning(sb, __func__,
776                              "Can't resize non-sparse filesystem further");
777                 return -EPERM;
778         }
779
780         if (le32_to_cpu(es->s_blocks_count) + input->blocks_count <
781             le32_to_cpu(es->s_blocks_count)) {
782                 ext3_warning(sb, __func__, "blocks_count overflow\n");
783                 return -EINVAL;
784         }
785
786         if (le32_to_cpu(es->s_inodes_count) + EXT3_INODES_PER_GROUP(sb) <
787             le32_to_cpu(es->s_inodes_count)) {
788                 ext3_warning(sb, __func__, "inodes_count overflow\n");
789                 return -EINVAL;
790         }
791
792         if (reserved_gdb || gdb_off == 0) {
793                 if (!EXT3_HAS_COMPAT_FEATURE(sb,
794                                              EXT3_FEATURE_COMPAT_RESIZE_INODE)
795                     || !le16_to_cpu(es->s_reserved_gdt_blocks)) {
796                         ext3_warning(sb, __func__,
797                                      "No reserved GDT blocks, can't resize");
798                         return -EPERM;
799                 }
800                 inode = ext3_iget(sb, EXT3_RESIZE_INO);
801                 if (IS_ERR(inode)) {
802                         ext3_warning(sb, __func__,
803                                      "Error opening resize inode");
804                         return PTR_ERR(inode);
805                 }
806         }
807
808         if ((err = verify_group_input(sb, input)))
809                 goto exit_put;
810
811         if ((err = setup_new_group_blocks(sb, input)))
812                 goto exit_put;
813
814         /*
815          * We will always be modifying at least the superblock and a GDT
816          * block.  If we are adding a group past the last current GDT block,
817          * we will also modify the inode and the dindirect block.  If we
818          * are adding a group with superblock/GDT backups  we will also
819          * modify each of the reserved GDT dindirect blocks.
820          */
821         handle = ext3_journal_start_sb(sb,
822                                        ext3_bg_has_super(sb, input->group) ?
823                                        3 + reserved_gdb : 4);
824         if (IS_ERR(handle)) {
825                 err = PTR_ERR(handle);
826                 goto exit_put;
827         }
828
829         mutex_lock(&sbi->s_resize_lock);
830         if (input->group != sbi->s_groups_count) {
831                 ext3_warning(sb, __func__,
832                              "multiple resizers run on filesystem!");
833                 err = -EBUSY;
834                 goto exit_journal;
835         }
836
837         if ((err = ext3_journal_get_write_access(handle, sbi->s_sbh)))
838                 goto exit_journal;
839
840         /*
841          * We will only either add reserved group blocks to a backup group
842          * or remove reserved blocks for the first group in a new group block.
843          * Doing both would be mean more complex code, and sane people don't
844          * use non-sparse filesystems anymore.  This is already checked above.
845          */
846         if (gdb_off) {
847                 primary = sbi->s_group_desc[gdb_num];
848                 if ((err = ext3_journal_get_write_access(handle, primary)))
849                         goto exit_journal;
850
851                 if (reserved_gdb && ext3_bg_num_gdb(sb, input->group) &&
852                     (err = reserve_backup_gdb(handle, inode, input)))
853                         goto exit_journal;
854         } else if ((err = add_new_gdb(handle, inode, input, &primary)))
855                 goto exit_journal;
856
857         /*
858          * OK, now we've set up the new group.  Time to make it active.
859          *
860          * We do not lock all allocations via s_resize_lock
861          * so we have to be safe wrt. concurrent accesses the group
862          * data.  So we need to be careful to set all of the relevant
863          * group descriptor data etc. *before* we enable the group.
864          *
865          * The key field here is sbi->s_groups_count: as long as
866          * that retains its old value, nobody is going to access the new
867          * group.
868          *
869          * So first we update all the descriptor metadata for the new
870          * group; then we update the total disk blocks count; then we
871          * update the groups count to enable the group; then finally we
872          * update the free space counts so that the system can start
873          * using the new disk blocks.
874          */
875
876         /* Update group descriptor block for new group */
877         gdp = (struct ext3_group_desc *)primary->b_data + gdb_off;
878
879         gdp->bg_block_bitmap = cpu_to_le32(input->block_bitmap);
880         gdp->bg_inode_bitmap = cpu_to_le32(input->inode_bitmap);
881         gdp->bg_inode_table = cpu_to_le32(input->inode_table);
882         gdp->bg_free_blocks_count = cpu_to_le16(input->free_blocks_count);
883         gdp->bg_free_inodes_count = cpu_to_le16(EXT3_INODES_PER_GROUP(sb));
884
885         /*
886          * Make the new blocks and inodes valid next.  We do this before
887          * increasing the group count so that once the group is enabled,
888          * all of its blocks and inodes are already valid.
889          *
890          * We always allocate group-by-group, then block-by-block or
891          * inode-by-inode within a group, so enabling these
892          * blocks/inodes before the group is live won't actually let us
893          * allocate the new space yet.
894          */
895         le32_add_cpu(&es->s_blocks_count, input->blocks_count);
896         le32_add_cpu(&es->s_inodes_count, EXT3_INODES_PER_GROUP(sb));
897
898         /*
899          * We need to protect s_groups_count against other CPUs seeing
900          * inconsistent state in the superblock.
901          *
902          * The precise rules we use are:
903          *
904          * * Writers of s_groups_count *must* hold s_resize_lock
905          * AND
906          * * Writers must perform a smp_wmb() after updating all dependent
907          *   data and before modifying the groups count
908          *
909          * * Readers must hold s_resize_lock over the access
910          * OR
911          * * Readers must perform an smp_rmb() after reading the groups count
912          *   and before reading any dependent data.
913          *
914          * NB. These rules can be relaxed when checking the group count
915          * while freeing data, as we can only allocate from a block
916          * group after serialising against the group count, and we can
917          * only then free after serialising in turn against that
918          * allocation.
919          */
920         smp_wmb();
921
922         /* Update the global fs size fields */
923         sbi->s_groups_count++;
924
925         ext3_journal_dirty_metadata(handle, primary);
926
927         /* Update the reserved block counts only once the new group is
928          * active. */
929         le32_add_cpu(&es->s_r_blocks_count, input->reserved_blocks);
930
931         /* Update the free space counts */
932         percpu_counter_add(&sbi->s_freeblocks_counter,
933                            input->free_blocks_count);
934         percpu_counter_add(&sbi->s_freeinodes_counter,
935                            EXT3_INODES_PER_GROUP(sb));
936
937         ext3_journal_dirty_metadata(handle, sbi->s_sbh);
938
939 exit_journal:
940         mutex_unlock(&sbi->s_resize_lock);
941         if ((err2 = ext3_journal_stop(handle)) && !err)
942                 err = err2;
943         if (!err) {
944                 update_backups(sb, sbi->s_sbh->b_blocknr, (char *)es,
945                                sizeof(struct ext3_super_block));
946                 update_backups(sb, primary->b_blocknr, primary->b_data,
947                                primary->b_size);
948         }
949 exit_put:
950         iput(inode);
951         return err;
952 } /* ext3_group_add */
953
954 /* Extend the filesystem to the new number of blocks specified.  This entry
955  * point is only used to extend the current filesystem to the end of the last
956  * existing group.  It can be accessed via ioctl, or by "remount,resize=<size>"
957  * for emergencies (because it has no dependencies on reserved blocks).
958  *
959  * If we _really_ wanted, we could use default values to call ext3_group_add()
960  * allow the "remount" trick to work for arbitrary resizing, assuming enough
961  * GDT blocks are reserved to grow to the desired size.
962  */
963 int ext3_group_extend(struct super_block *sb, struct ext3_super_block *es,
964                       ext3_fsblk_t n_blocks_count)
965 {
966         ext3_fsblk_t o_blocks_count;
967         unsigned long o_groups_count;
968         ext3_grpblk_t last;
969         ext3_grpblk_t add;
970         struct buffer_head * bh;
971         handle_t *handle;
972         int err;
973         unsigned long freed_blocks;
974
975         /* We don't need to worry about locking wrt other resizers just
976          * yet: we're going to revalidate es->s_blocks_count after
977          * taking the s_resize_lock below. */
978         o_blocks_count = le32_to_cpu(es->s_blocks_count);
979         o_groups_count = EXT3_SB(sb)->s_groups_count;
980
981         if (test_opt(sb, DEBUG))
982                 printk(KERN_DEBUG "EXT3-fs: extending last group from "E3FSBLK" uto "E3FSBLK" blocks\n",
983                        o_blocks_count, n_blocks_count);
984
985         if (n_blocks_count == 0 || n_blocks_count == o_blocks_count)
986                 return 0;
987
988         if (n_blocks_count > (sector_t)(~0ULL) >> (sb->s_blocksize_bits - 9)) {
989                 printk(KERN_ERR "EXT3-fs: filesystem on %s:"
990                         " too large to resize to %lu blocks safely\n",
991                         sb->s_id, n_blocks_count);
992                 if (sizeof(sector_t) < 8)
993                         ext3_warning(sb, __func__,
994                         "CONFIG_LBDAF not enabled\n");
995                 return -EINVAL;
996         }
997
998         if (n_blocks_count < o_blocks_count) {
999                 ext3_warning(sb, __func__,
1000                              "can't shrink FS - resize aborted");
1001                 return -EBUSY;
1002         }
1003
1004         /* Handle the remaining blocks in the last group only. */
1005         last = (o_blocks_count - le32_to_cpu(es->s_first_data_block)) %
1006                 EXT3_BLOCKS_PER_GROUP(sb);
1007
1008         if (last == 0) {
1009                 ext3_warning(sb, __func__,
1010                              "need to use ext2online to resize further");
1011                 return -EPERM;
1012         }
1013
1014         add = EXT3_BLOCKS_PER_GROUP(sb) - last;
1015
1016         if (o_blocks_count + add < o_blocks_count) {
1017                 ext3_warning(sb, __func__, "blocks_count overflow");
1018                 return -EINVAL;
1019         }
1020
1021         if (o_blocks_count + add > n_blocks_count)
1022                 add = n_blocks_count - o_blocks_count;
1023
1024         if (o_blocks_count + add < n_blocks_count)
1025                 ext3_warning(sb, __func__,
1026                              "will only finish group ("E3FSBLK
1027                              " blocks, %u new)",
1028                              o_blocks_count + add, add);
1029
1030         /* See if the device is actually as big as what was requested */
1031         bh = sb_bread(sb, o_blocks_count + add -1);
1032         if (!bh) {
1033                 ext3_warning(sb, __func__,
1034                              "can't read last block, resize aborted");
1035                 return -ENOSPC;
1036         }
1037         brelse(bh);
1038
1039         /* We will update the superblock, one block bitmap, and
1040          * one group descriptor via ext3_free_blocks().
1041          */
1042         handle = ext3_journal_start_sb(sb, 3);
1043         if (IS_ERR(handle)) {
1044                 err = PTR_ERR(handle);
1045                 ext3_warning(sb, __func__, "error %d on journal start",err);
1046                 goto exit_put;
1047         }
1048
1049         mutex_lock(&EXT3_SB(sb)->s_resize_lock);
1050         if (o_blocks_count != le32_to_cpu(es->s_blocks_count)) {
1051                 ext3_warning(sb, __func__,
1052                              "multiple resizers run on filesystem!");
1053                 mutex_unlock(&EXT3_SB(sb)->s_resize_lock);
1054                 ext3_journal_stop(handle);
1055                 err = -EBUSY;
1056                 goto exit_put;
1057         }
1058
1059         if ((err = ext3_journal_get_write_access(handle,
1060                                                  EXT3_SB(sb)->s_sbh))) {
1061                 ext3_warning(sb, __func__,
1062                              "error %d on journal write access", err);
1063                 mutex_unlock(&EXT3_SB(sb)->s_resize_lock);
1064                 ext3_journal_stop(handle);
1065                 goto exit_put;
1066         }
1067         es->s_blocks_count = cpu_to_le32(o_blocks_count + add);
1068         ext3_journal_dirty_metadata(handle, EXT3_SB(sb)->s_sbh);
1069         mutex_unlock(&EXT3_SB(sb)->s_resize_lock);
1070         ext3_debug("freeing blocks %lu through "E3FSBLK"\n", o_blocks_count,
1071                    o_blocks_count + add);
1072         ext3_free_blocks_sb(handle, sb, o_blocks_count, add, &freed_blocks);
1073         ext3_debug("freed blocks "E3FSBLK" through "E3FSBLK"\n", o_blocks_count,
1074                    o_blocks_count + add);
1075         if ((err = ext3_journal_stop(handle)))
1076                 goto exit_put;
1077         if (test_opt(sb, DEBUG))
1078                 printk(KERN_DEBUG "EXT3-fs: extended group to %u blocks\n",
1079                        le32_to_cpu(es->s_blocks_count));
1080         update_backups(sb, EXT3_SB(sb)->s_sbh->b_blocknr, (char *)es,
1081                        sizeof(struct ext3_super_block));
1082 exit_put:
1083         return err;
1084 } /* ext3_group_extend */