ext4: Fix race between migration and mmap write
[linux-2.6] / fs / ext4 / migrate.c
1 /*
2  * Copyright IBM Corporation, 2007
3  * Author Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
4  *
5  * This program is free software; you can redistribute it and/or modify it
6  * under the terms of version 2.1 of the GNU Lesser General Public License
7  * as published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it would be useful, but
10  * WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
12  *
13  */
14
15 #include <linux/module.h>
16 #include <linux/ext4_jbd2.h>
17 #include <linux/ext4_fs_extents.h>
18
19 /*
20  * The contiguous blocks details which can be
21  * represented by a single extent
22  */
23 struct list_blocks_struct {
24         ext4_lblk_t first_block, last_block;
25         ext4_fsblk_t first_pblock, last_pblock;
26 };
27
28 static int finish_range(handle_t *handle, struct inode *inode,
29                                 struct list_blocks_struct *lb)
30
31 {
32         int retval = 0, needed;
33         struct ext4_extent newext;
34         struct ext4_ext_path *path;
35         if (lb->first_pblock == 0)
36                 return 0;
37
38         /* Add the extent to temp inode*/
39         newext.ee_block = cpu_to_le32(lb->first_block);
40         newext.ee_len   = cpu_to_le16(lb->last_block - lb->first_block + 1);
41         ext4_ext_store_pblock(&newext, lb->first_pblock);
42         path = ext4_ext_find_extent(inode, lb->first_block, NULL);
43
44         if (IS_ERR(path)) {
45                 retval = PTR_ERR(path);
46                 path = NULL;
47                 goto err_out;
48         }
49
50         /*
51          * Calculate the credit needed to inserting this extent
52          * Since we are doing this in loop we may accumalate extra
53          * credit. But below we try to not accumalate too much
54          * of them by restarting the journal.
55          */
56         needed = ext4_ext_calc_credits_for_insert(inode, path);
57
58         /*
59          * Make sure the credit we accumalated is not really high
60          */
61         if (needed && handle->h_buffer_credits >= EXT4_RESERVE_TRANS_BLOCKS) {
62                 retval = ext4_journal_restart(handle, needed);
63                 if (retval)
64                         goto err_out;
65         } else if (needed) {
66                 retval = ext4_journal_extend(handle, needed);
67                 if (retval) {
68                         /*
69                          * IF not able to extend the journal restart the journal
70                          */
71                         retval = ext4_journal_restart(handle, needed);
72                         if (retval)
73                                 goto err_out;
74                 }
75         }
76         retval = ext4_ext_insert_extent(handle, inode, path, &newext);
77 err_out:
78         if (path) {
79                 ext4_ext_drop_refs(path);
80                 kfree(path);
81         }
82         lb->first_pblock = 0;
83         return retval;
84 }
85
86 static int update_extent_range(handle_t *handle, struct inode *inode,
87                                 ext4_fsblk_t pblock, ext4_lblk_t blk_num,
88                                 struct list_blocks_struct *lb)
89 {
90         int retval;
91         /*
92          * See if we can add on to the existing range (if it exists)
93          */
94         if (lb->first_pblock &&
95                 (lb->last_pblock+1 == pblock) &&
96                 (lb->last_block+1 == blk_num)) {
97                 lb->last_pblock = pblock;
98                 lb->last_block = blk_num;
99                 return 0;
100         }
101         /*
102          * Start a new range.
103          */
104         retval = finish_range(handle, inode, lb);
105         lb->first_pblock = lb->last_pblock = pblock;
106         lb->first_block = lb->last_block = blk_num;
107
108         return retval;
109 }
110
111 static int update_ind_extent_range(handle_t *handle, struct inode *inode,
112                                    ext4_fsblk_t pblock, ext4_lblk_t *blk_nump,
113                                    struct list_blocks_struct *lb)
114 {
115         struct buffer_head *bh;
116         __le32 *i_data;
117         int i, retval = 0;
118         ext4_lblk_t blk_count = *blk_nump;
119         unsigned long max_entries = inode->i_sb->s_blocksize >> 2;
120
121         if (!pblock) {
122                 /* Only update the file block number */
123                 *blk_nump += max_entries;
124                 return 0;
125         }
126
127         bh = sb_bread(inode->i_sb, pblock);
128         if (!bh)
129                 return -EIO;
130
131         i_data = (__le32 *)bh->b_data;
132         for (i = 0; i < max_entries; i++, blk_count++) {
133                 if (i_data[i]) {
134                         retval = update_extent_range(handle, inode,
135                                                 le32_to_cpu(i_data[i]),
136                                                 blk_count, lb);
137                         if (retval)
138                                 break;
139                 }
140         }
141
142         /* Update the file block number */
143         *blk_nump = blk_count;
144         put_bh(bh);
145         return retval;
146
147 }
148
149 static int update_dind_extent_range(handle_t *handle, struct inode *inode,
150                                     ext4_fsblk_t pblock, ext4_lblk_t *blk_nump,
151                                     struct list_blocks_struct *lb)
152 {
153         struct buffer_head *bh;
154         __le32 *i_data;
155         int i, retval = 0;
156         ext4_lblk_t blk_count = *blk_nump;
157         unsigned long max_entries = inode->i_sb->s_blocksize >> 2;
158
159         if (!pblock) {
160                 /* Only update the file block number */
161                 *blk_nump += max_entries * max_entries;
162                 return 0;
163         }
164         bh = sb_bread(inode->i_sb, pblock);
165         if (!bh)
166                 return -EIO;
167
168         i_data = (__le32 *)bh->b_data;
169         for (i = 0; i < max_entries; i++) {
170                 if (i_data[i]) {
171                         retval = update_ind_extent_range(handle, inode,
172                                                 le32_to_cpu(i_data[i]),
173                                                 &blk_count, lb);
174                         if (retval)
175                                 break;
176                 } else {
177                         /* Only update the file block number */
178                         blk_count += max_entries;
179                 }
180         }
181
182         /* Update the file block number */
183         *blk_nump = blk_count;
184         put_bh(bh);
185         return retval;
186
187 }
188
189 static int update_tind_extent_range(handle_t *handle, struct inode *inode,
190                                      ext4_fsblk_t pblock, ext4_lblk_t *blk_nump,
191                                      struct list_blocks_struct *lb)
192 {
193         struct buffer_head *bh;
194         __le32 *i_data;
195         int i, retval = 0;
196         ext4_lblk_t blk_count = *blk_nump;
197         unsigned long max_entries = inode->i_sb->s_blocksize >> 2;
198
199         if (!pblock) {
200                 /* Only update the file block number */
201                 *blk_nump += max_entries * max_entries * max_entries;
202                 return 0;
203         }
204         bh = sb_bread(inode->i_sb, pblock);
205         if (!bh)
206                 return -EIO;
207
208         i_data = (__le32 *)bh->b_data;
209         for (i = 0; i < max_entries; i++) {
210                 if (i_data[i]) {
211                         retval = update_dind_extent_range(handle, inode,
212                                                 le32_to_cpu(i_data[i]),
213                                                 &blk_count, lb);
214                         if (retval)
215                                 break;
216                 } else
217                         /* Only update the file block number */
218                         blk_count += max_entries * max_entries;
219         }
220         /* Update the file block number */
221         *blk_nump = blk_count;
222         put_bh(bh);
223         return retval;
224
225 }
226
227 static int extend_credit_for_blkdel(handle_t *handle, struct inode *inode)
228 {
229         int retval = 0, needed;
230
231         if (handle->h_buffer_credits > EXT4_RESERVE_TRANS_BLOCKS)
232                 return 0;
233         /*
234          * We are freeing a blocks. During this we touch
235          * superblock, group descriptor and block bitmap.
236          * So allocate a credit of 3. We may update
237          * quota (user and group).
238          */
239         needed = 3 + 2*EXT4_QUOTA_TRANS_BLOCKS(inode->i_sb);
240
241         if (ext4_journal_extend(handle, needed) != 0)
242                 retval = ext4_journal_restart(handle, needed);
243
244         return retval;
245 }
246
247 static int free_dind_blocks(handle_t *handle,
248                                 struct inode *inode, __le32 i_data)
249 {
250         int i;
251         __le32 *tmp_idata;
252         struct buffer_head *bh;
253         unsigned long max_entries = inode->i_sb->s_blocksize >> 2;
254
255         bh = sb_bread(inode->i_sb, le32_to_cpu(i_data));
256         if (!bh)
257                 return -EIO;
258
259         tmp_idata = (__le32 *)bh->b_data;
260         for (i = 0; i < max_entries; i++) {
261                 if (tmp_idata[i]) {
262                         extend_credit_for_blkdel(handle, inode);
263                         ext4_free_blocks(handle, inode,
264                                         le32_to_cpu(tmp_idata[i]), 1, 1);
265                 }
266         }
267         put_bh(bh);
268         extend_credit_for_blkdel(handle, inode);
269         ext4_free_blocks(handle, inode, le32_to_cpu(i_data), 1, 1);
270         return 0;
271 }
272
273 static int free_tind_blocks(handle_t *handle,
274                                 struct inode *inode, __le32 i_data)
275 {
276         int i, retval = 0;
277         __le32 *tmp_idata;
278         struct buffer_head *bh;
279         unsigned long max_entries = inode->i_sb->s_blocksize >> 2;
280
281         bh = sb_bread(inode->i_sb, le32_to_cpu(i_data));
282         if (!bh)
283                 return -EIO;
284
285         tmp_idata = (__le32 *)bh->b_data;
286         for (i = 0; i < max_entries; i++) {
287                 if (tmp_idata[i]) {
288                         retval = free_dind_blocks(handle,
289                                         inode, tmp_idata[i]);
290                         if (retval) {
291                                 put_bh(bh);
292                                 return retval;
293                         }
294                 }
295         }
296         put_bh(bh);
297         extend_credit_for_blkdel(handle, inode);
298         ext4_free_blocks(handle, inode, le32_to_cpu(i_data), 1, 1);
299         return 0;
300 }
301
302 static int free_ind_block(handle_t *handle, struct inode *inode, __le32 *i_data)
303 {
304         int retval;
305
306         /* ei->i_data[EXT4_IND_BLOCK] */
307         if (i_data[0]) {
308                 extend_credit_for_blkdel(handle, inode);
309                 ext4_free_blocks(handle, inode,
310                                 le32_to_cpu(i_data[0]), 1, 1);
311         }
312
313         /* ei->i_data[EXT4_DIND_BLOCK] */
314         if (i_data[1]) {
315                 retval = free_dind_blocks(handle, inode, i_data[1]);
316                 if (retval)
317                         return retval;
318         }
319
320         /* ei->i_data[EXT4_TIND_BLOCK] */
321         if (i_data[2]) {
322                 retval = free_tind_blocks(handle, inode, i_data[2]);
323                 if (retval)
324                         return retval;
325         }
326         return 0;
327 }
328
329 static int ext4_ext_swap_inode_data(handle_t *handle, struct inode *inode,
330                                                 struct inode *tmp_inode)
331 {
332         int retval;
333         __le32  i_data[3];
334         struct ext4_inode_info *ei = EXT4_I(inode);
335         struct ext4_inode_info *tmp_ei = EXT4_I(tmp_inode);
336
337         /*
338          * One credit accounted for writing the
339          * i_data field of the original inode
340          */
341         retval = ext4_journal_extend(handle, 1);
342         if (retval) {
343                 retval = ext4_journal_restart(handle, 1);
344                 if (retval)
345                         goto err_out;
346         }
347
348         i_data[0] = ei->i_data[EXT4_IND_BLOCK];
349         i_data[1] = ei->i_data[EXT4_DIND_BLOCK];
350         i_data[2] = ei->i_data[EXT4_TIND_BLOCK];
351
352         down_write(&EXT4_I(inode)->i_data_sem);
353         /*
354          * if EXT4_EXT_MIGRATE is cleared a block allocation
355          * happened after we started the migrate. We need to
356          * fail the migrate
357          */
358         if (!(EXT4_I(inode)->i_flags & EXT4_EXT_MIGRATE)) {
359                 retval = -EAGAIN;
360                 up_write(&EXT4_I(inode)->i_data_sem);
361                 goto err_out;
362         } else
363                 EXT4_I(inode)->i_flags = EXT4_I(inode)->i_flags &
364                                                         ~EXT4_EXT_MIGRATE;
365         /*
366          * We have the extent map build with the tmp inode.
367          * Now copy the i_data across
368          */
369         ei->i_flags |= EXT4_EXTENTS_FL;
370         memcpy(ei->i_data, tmp_ei->i_data, sizeof(ei->i_data));
371
372         /*
373          * Update i_blocks with the new blocks that got
374          * allocated while adding extents for extent index
375          * blocks.
376          *
377          * While converting to extents we need not
378          * update the orignal inode i_blocks for extent blocks
379          * via quota APIs. The quota update happened via tmp_inode already.
380          */
381         spin_lock(&inode->i_lock);
382         inode->i_blocks += tmp_inode->i_blocks;
383         spin_unlock(&inode->i_lock);
384         up_write(&EXT4_I(inode)->i_data_sem);
385
386         /*
387          * We mark the inode dirty after, because we decrement the
388          * i_blocks when freeing the indirect meta-data blocks
389          */
390         retval = free_ind_block(handle, inode, i_data);
391         ext4_mark_inode_dirty(handle, inode);
392
393 err_out:
394         return retval;
395 }
396
397 static int free_ext_idx(handle_t *handle, struct inode *inode,
398                                         struct ext4_extent_idx *ix)
399 {
400         int i, retval = 0;
401         ext4_fsblk_t block;
402         struct buffer_head *bh;
403         struct ext4_extent_header *eh;
404
405         block = idx_pblock(ix);
406         bh = sb_bread(inode->i_sb, block);
407         if (!bh)
408                 return -EIO;
409
410         eh = (struct ext4_extent_header *)bh->b_data;
411         if (eh->eh_depth != 0) {
412                 ix = EXT_FIRST_INDEX(eh);
413                 for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ix++) {
414                         retval = free_ext_idx(handle, inode, ix);
415                         if (retval)
416                                 break;
417                 }
418         }
419         put_bh(bh);
420         extend_credit_for_blkdel(handle, inode);
421         ext4_free_blocks(handle, inode, block, 1, 1);
422         return retval;
423 }
424
425 /*
426  * Free the extent meta data blocks only
427  */
428 static int free_ext_block(handle_t *handle, struct inode *inode)
429 {
430         int i, retval = 0;
431         struct ext4_inode_info *ei = EXT4_I(inode);
432         struct ext4_extent_header *eh = (struct ext4_extent_header *)ei->i_data;
433         struct ext4_extent_idx *ix;
434         if (eh->eh_depth == 0)
435                 /*
436                  * No extra blocks allocated for extent meta data
437                  */
438                 return 0;
439         ix = EXT_FIRST_INDEX(eh);
440         for (i = 0; i < le16_to_cpu(eh->eh_entries); i++, ix++) {
441                 retval = free_ext_idx(handle, inode, ix);
442                 if (retval)
443                         return retval;
444         }
445         return retval;
446
447 }
448
449 int ext4_ext_migrate(struct inode *inode, struct file *filp,
450                                 unsigned int cmd, unsigned long arg)
451 {
452         handle_t *handle;
453         int retval = 0, i;
454         __le32 *i_data;
455         ext4_lblk_t blk_count = 0;
456         struct ext4_inode_info *ei;
457         struct inode *tmp_inode = NULL;
458         struct list_blocks_struct lb;
459         unsigned long max_entries;
460
461         if (!test_opt(inode->i_sb, EXTENTS))
462                 /*
463                  * if mounted with noextents we don't allow the migrate
464                  */
465                 return -EINVAL;
466
467         if ((EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL))
468                 return -EINVAL;
469
470         if (S_ISLNK(inode->i_mode) && inode->i_blocks == 0)
471                 /*
472                  * don't migrate fast symlink
473                  */
474                 return retval;
475
476         handle = ext4_journal_start(inode,
477                                         EXT4_DATA_TRANS_BLOCKS(inode->i_sb) +
478                                         EXT4_INDEX_EXTRA_TRANS_BLOCKS + 3 +
479                                         2 * EXT4_QUOTA_INIT_BLOCKS(inode->i_sb)
480                                         + 1);
481         if (IS_ERR(handle)) {
482                 retval = PTR_ERR(handle);
483                 goto err_out;
484         }
485         tmp_inode = ext4_new_inode(handle,
486                                 inode->i_sb->s_root->d_inode,
487                                 S_IFREG);
488         if (IS_ERR(tmp_inode)) {
489                 retval = -ENOMEM;
490                 ext4_journal_stop(handle);
491                 tmp_inode = NULL;
492                 goto err_out;
493         }
494         i_size_write(tmp_inode, i_size_read(inode));
495         /*
496          * We don't want the inode to be reclaimed
497          * if we got interrupted in between. We have
498          * this tmp inode carrying reference to the
499          * data blocks of the original file. We set
500          * the i_nlink to zero at the last stage after
501          * switching the original file to extent format
502          */
503         tmp_inode->i_nlink = 1;
504
505         ext4_ext_tree_init(handle, tmp_inode);
506         ext4_orphan_add(handle, tmp_inode);
507         ext4_journal_stop(handle);
508
509         /*
510          * start with one credit accounted for
511          * superblock modification.
512          *
513          * For the tmp_inode we already have commited the
514          * trascation that created the inode. Later as and
515          * when we add extents we extent the journal
516          */
517         /*
518          * inode_mutex prevent write and truncate on the file. Read still goes
519          * through. We take i_data_sem in ext4_ext_swap_inode_data before we
520          * switch the inode format to prevent read.
521          */
522         mutex_lock(&(inode->i_mutex));
523         /*
524          * Even though we take i_mutex we can still cause block allocation
525          * via mmap write to holes. If we have allocated new blocks we fail
526          * migrate.  New block allocation will clear EXT4_EXT_MIGRATE flag.
527          * The flag is updated with i_data_sem held to prevent racing with
528          * block allocation.
529          */
530         down_read((&EXT4_I(inode)->i_data_sem));
531         EXT4_I(inode)->i_flags = EXT4_I(inode)->i_flags | EXT4_EXT_MIGRATE;
532         up_read((&EXT4_I(inode)->i_data_sem));
533
534         handle = ext4_journal_start(inode, 1);
535
536         ei = EXT4_I(inode);
537         i_data = ei->i_data;
538         memset(&lb, 0, sizeof(lb));
539
540         /* 32 bit block address 4 bytes */
541         max_entries = inode->i_sb->s_blocksize >> 2;
542         for (i = 0; i < EXT4_NDIR_BLOCKS; i++, blk_count++) {
543                 if (i_data[i]) {
544                         retval = update_extent_range(handle, tmp_inode,
545                                                 le32_to_cpu(i_data[i]),
546                                                 blk_count, &lb);
547                         if (retval)
548                                 goto err_out;
549                 }
550         }
551         if (i_data[EXT4_IND_BLOCK]) {
552                 retval = update_ind_extent_range(handle, tmp_inode,
553                                         le32_to_cpu(i_data[EXT4_IND_BLOCK]),
554                                         &blk_count, &lb);
555                         if (retval)
556                                 goto err_out;
557         } else
558                 blk_count +=  max_entries;
559         if (i_data[EXT4_DIND_BLOCK]) {
560                 retval = update_dind_extent_range(handle, tmp_inode,
561                                         le32_to_cpu(i_data[EXT4_DIND_BLOCK]),
562                                         &blk_count, &lb);
563                         if (retval)
564                                 goto err_out;
565         } else
566                 blk_count += max_entries * max_entries;
567         if (i_data[EXT4_TIND_BLOCK]) {
568                 retval = update_tind_extent_range(handle, tmp_inode,
569                                         le32_to_cpu(i_data[EXT4_TIND_BLOCK]),
570                                         &blk_count, &lb);
571                         if (retval)
572                                 goto err_out;
573         }
574         /*
575          * Build the last extent
576          */
577         retval = finish_range(handle, tmp_inode, &lb);
578 err_out:
579         if (retval)
580                 /*
581                  * Failure case delete the extent information with the
582                  * tmp_inode
583                  */
584                 free_ext_block(handle, tmp_inode);
585         else {
586                 retval = ext4_ext_swap_inode_data(handle, inode, tmp_inode);
587                 if (retval)
588                         /*
589                          * if we fail to swap inode data free the extent
590                          * details of the tmp inode
591                          */
592                         free_ext_block(handle, tmp_inode);
593         }
594
595         /* We mark the tmp_inode dirty via ext4_ext_tree_init. */
596         if (ext4_journal_extend(handle, 1) != 0)
597                 ext4_journal_restart(handle, 1);
598
599         /*
600          * Mark the tmp_inode as of size zero
601          */
602         i_size_write(tmp_inode, 0);
603
604         /*
605          * set the  i_blocks count to zero
606          * so that the ext4_delete_inode does the
607          * right job
608          *
609          * We don't need to take the i_lock because
610          * the inode is not visible to user space.
611          */
612         tmp_inode->i_blocks = 0;
613
614         /* Reset the extent details */
615         ext4_ext_tree_init(handle, tmp_inode);
616
617         /*
618          * Set the i_nlink to zero so that
619          * generic_drop_inode really deletes the
620          * inode
621          */
622         tmp_inode->i_nlink = 0;
623
624         ext4_journal_stop(handle);
625         mutex_unlock(&(inode->i_mutex));
626
627         if (tmp_inode)
628                 iput(tmp_inode);
629
630         return retval;
631 }