Merge tag 'ext4_for_linus_stable' of git://git.kernel.org/pub/scm/linux/kernel/git...
[sfrench/cifs-2.6.git] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  64-bit file support on 64-bit platforms by Jakub Jelinek
16  *      (jj@sunsite.ms.mff.cuni.cz)
17  *
18  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
19  */
20
21 #include <linux/fs.h>
22 #include <linux/time.h>
23 #include <linux/highuid.h>
24 #include <linux/pagemap.h>
25 #include <linux/dax.h>
26 #include <linux/quotaops.h>
27 #include <linux/string.h>
28 #include <linux/buffer_head.h>
29 #include <linux/writeback.h>
30 #include <linux/pagevec.h>
31 #include <linux/mpage.h>
32 #include <linux/namei.h>
33 #include <linux/uio.h>
34 #include <linux/bio.h>
35 #include <linux/workqueue.h>
36 #include <linux/kernel.h>
37 #include <linux/printk.h>
38 #include <linux/slab.h>
39 #include <linux/bitops.h>
40
41 #include "ext4_jbd2.h"
42 #include "xattr.h"
43 #include "acl.h"
44 #include "truncate.h"
45
46 #include <trace/events/ext4.h>
47
48 #define MPAGE_DA_EXTENT_TAIL 0x01
49
50 static __u32 ext4_inode_csum(struct inode *inode, struct ext4_inode *raw,
51                               struct ext4_inode_info *ei)
52 {
53         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
54         __u32 csum;
55         __u16 dummy_csum = 0;
56         int offset = offsetof(struct ext4_inode, i_checksum_lo);
57         unsigned int csum_size = sizeof(dummy_csum);
58
59         csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)raw, offset);
60         csum = ext4_chksum(sbi, csum, (__u8 *)&dummy_csum, csum_size);
61         offset += csum_size;
62         csum = ext4_chksum(sbi, csum, (__u8 *)raw + offset,
63                            EXT4_GOOD_OLD_INODE_SIZE - offset);
64
65         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
66                 offset = offsetof(struct ext4_inode, i_checksum_hi);
67                 csum = ext4_chksum(sbi, csum, (__u8 *)raw +
68                                    EXT4_GOOD_OLD_INODE_SIZE,
69                                    offset - EXT4_GOOD_OLD_INODE_SIZE);
70                 if (EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi)) {
71                         csum = ext4_chksum(sbi, csum, (__u8 *)&dummy_csum,
72                                            csum_size);
73                         offset += csum_size;
74                         csum = ext4_chksum(sbi, csum, (__u8 *)raw + offset,
75                                            EXT4_INODE_SIZE(inode->i_sb) -
76                                            offset);
77                 }
78         }
79
80         return csum;
81 }
82
83 static int ext4_inode_csum_verify(struct inode *inode, struct ext4_inode *raw,
84                                   struct ext4_inode_info *ei)
85 {
86         __u32 provided, calculated;
87
88         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
89             cpu_to_le32(EXT4_OS_LINUX) ||
90             !ext4_has_metadata_csum(inode->i_sb))
91                 return 1;
92
93         provided = le16_to_cpu(raw->i_checksum_lo);
94         calculated = ext4_inode_csum(inode, raw, ei);
95         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
96             EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
97                 provided |= ((__u32)le16_to_cpu(raw->i_checksum_hi)) << 16;
98         else
99                 calculated &= 0xFFFF;
100
101         return provided == calculated;
102 }
103
104 static void ext4_inode_csum_set(struct inode *inode, struct ext4_inode *raw,
105                                 struct ext4_inode_info *ei)
106 {
107         __u32 csum;
108
109         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
110             cpu_to_le32(EXT4_OS_LINUX) ||
111             !ext4_has_metadata_csum(inode->i_sb))
112                 return;
113
114         csum = ext4_inode_csum(inode, raw, ei);
115         raw->i_checksum_lo = cpu_to_le16(csum & 0xFFFF);
116         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
117             EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
118                 raw->i_checksum_hi = cpu_to_le16(csum >> 16);
119 }
120
121 static inline int ext4_begin_ordered_truncate(struct inode *inode,
122                                               loff_t new_size)
123 {
124         trace_ext4_begin_ordered_truncate(inode, new_size);
125         /*
126          * If jinode is zero, then we never opened the file for
127          * writing, so there's no need to call
128          * jbd2_journal_begin_ordered_truncate() since there's no
129          * outstanding writes we need to flush.
130          */
131         if (!EXT4_I(inode)->jinode)
132                 return 0;
133         return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
134                                                    EXT4_I(inode)->jinode,
135                                                    new_size);
136 }
137
138 static void ext4_invalidatepage(struct page *page, unsigned int offset,
139                                 unsigned int length);
140 static int __ext4_journalled_writepage(struct page *page, unsigned int len);
141 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
142 static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
143                                   int pextents);
144
145 /*
146  * Test whether an inode is a fast symlink.
147  */
148 int ext4_inode_is_fast_symlink(struct inode *inode)
149 {
150         int ea_blocks = EXT4_I(inode)->i_file_acl ?
151                 EXT4_CLUSTER_SIZE(inode->i_sb) >> 9 : 0;
152
153         if (ext4_has_inline_data(inode))
154                 return 0;
155
156         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
157 }
158
159 /*
160  * Restart the transaction associated with *handle.  This does a commit,
161  * so before we call here everything must be consistently dirtied against
162  * this transaction.
163  */
164 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
165                                  int nblocks)
166 {
167         int ret;
168
169         /*
170          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
171          * moment, get_block can be called only for blocks inside i_size since
172          * page cache has been already dropped and writes are blocked by
173          * i_mutex. So we can safely drop the i_data_sem here.
174          */
175         BUG_ON(EXT4_JOURNAL(inode) == NULL);
176         jbd_debug(2, "restarting handle %p\n", handle);
177         up_write(&EXT4_I(inode)->i_data_sem);
178         ret = ext4_journal_restart(handle, nblocks);
179         down_write(&EXT4_I(inode)->i_data_sem);
180         ext4_discard_preallocations(inode);
181
182         return ret;
183 }
184
185 /*
186  * Called at the last iput() if i_nlink is zero.
187  */
188 void ext4_evict_inode(struct inode *inode)
189 {
190         handle_t *handle;
191         int err;
192
193         trace_ext4_evict_inode(inode);
194
195         if (inode->i_nlink) {
196                 /*
197                  * When journalling data dirty buffers are tracked only in the
198                  * journal. So although mm thinks everything is clean and
199                  * ready for reaping the inode might still have some pages to
200                  * write in the running transaction or waiting to be
201                  * checkpointed. Thus calling jbd2_journal_invalidatepage()
202                  * (via truncate_inode_pages()) to discard these buffers can
203                  * cause data loss. Also even if we did not discard these
204                  * buffers, we would have no way to find them after the inode
205                  * is reaped and thus user could see stale data if he tries to
206                  * read them before the transaction is checkpointed. So be
207                  * careful and force everything to disk here... We use
208                  * ei->i_datasync_tid to store the newest transaction
209                  * containing inode's data.
210                  *
211                  * Note that directories do not have this problem because they
212                  * don't use page cache.
213                  */
214                 if (inode->i_ino != EXT4_JOURNAL_INO &&
215                     ext4_should_journal_data(inode) &&
216                     (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode))) {
217                         journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
218                         tid_t commit_tid = EXT4_I(inode)->i_datasync_tid;
219
220                         jbd2_complete_transaction(journal, commit_tid);
221                         filemap_write_and_wait(&inode->i_data);
222                 }
223                 truncate_inode_pages_final(&inode->i_data);
224
225                 goto no_delete;
226         }
227
228         if (is_bad_inode(inode))
229                 goto no_delete;
230         dquot_initialize(inode);
231
232         if (ext4_should_order_data(inode))
233                 ext4_begin_ordered_truncate(inode, 0);
234         truncate_inode_pages_final(&inode->i_data);
235
236         /*
237          * Protect us against freezing - iput() caller didn't have to have any
238          * protection against it
239          */
240         sb_start_intwrite(inode->i_sb);
241         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE,
242                                     ext4_blocks_for_truncate(inode)+3);
243         if (IS_ERR(handle)) {
244                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
245                 /*
246                  * If we're going to skip the normal cleanup, we still need to
247                  * make sure that the in-core orphan linked list is properly
248                  * cleaned up.
249                  */
250                 ext4_orphan_del(NULL, inode);
251                 sb_end_intwrite(inode->i_sb);
252                 goto no_delete;
253         }
254
255         if (IS_SYNC(inode))
256                 ext4_handle_sync(handle);
257         inode->i_size = 0;
258         err = ext4_mark_inode_dirty(handle, inode);
259         if (err) {
260                 ext4_warning(inode->i_sb,
261                              "couldn't mark inode dirty (err %d)", err);
262                 goto stop_handle;
263         }
264         if (inode->i_blocks)
265                 ext4_truncate(inode);
266
267         /*
268          * ext4_ext_truncate() doesn't reserve any slop when it
269          * restarts journal transactions; therefore there may not be
270          * enough credits left in the handle to remove the inode from
271          * the orphan list and set the dtime field.
272          */
273         if (!ext4_handle_has_enough_credits(handle, 3)) {
274                 err = ext4_journal_extend(handle, 3);
275                 if (err > 0)
276                         err = ext4_journal_restart(handle, 3);
277                 if (err != 0) {
278                         ext4_warning(inode->i_sb,
279                                      "couldn't extend journal (err %d)", err);
280                 stop_handle:
281                         ext4_journal_stop(handle);
282                         ext4_orphan_del(NULL, inode);
283                         sb_end_intwrite(inode->i_sb);
284                         goto no_delete;
285                 }
286         }
287
288         /*
289          * Kill off the orphan record which ext4_truncate created.
290          * AKPM: I think this can be inside the above `if'.
291          * Note that ext4_orphan_del() has to be able to cope with the
292          * deletion of a non-existent orphan - this is because we don't
293          * know if ext4_truncate() actually created an orphan record.
294          * (Well, we could do this if we need to, but heck - it works)
295          */
296         ext4_orphan_del(handle, inode);
297         EXT4_I(inode)->i_dtime  = get_seconds();
298
299         /*
300          * One subtle ordering requirement: if anything has gone wrong
301          * (transaction abort, IO errors, whatever), then we can still
302          * do these next steps (the fs will already have been marked as
303          * having errors), but we can't free the inode if the mark_dirty
304          * fails.
305          */
306         if (ext4_mark_inode_dirty(handle, inode))
307                 /* If that failed, just do the required in-core inode clear. */
308                 ext4_clear_inode(inode);
309         else
310                 ext4_free_inode(handle, inode);
311         ext4_journal_stop(handle);
312         sb_end_intwrite(inode->i_sb);
313         return;
314 no_delete:
315         ext4_clear_inode(inode);        /* We must guarantee clearing of inode... */
316 }
317
318 #ifdef CONFIG_QUOTA
319 qsize_t *ext4_get_reserved_space(struct inode *inode)
320 {
321         return &EXT4_I(inode)->i_reserved_quota;
322 }
323 #endif
324
325 /*
326  * Called with i_data_sem down, which is important since we can call
327  * ext4_discard_preallocations() from here.
328  */
329 void ext4_da_update_reserve_space(struct inode *inode,
330                                         int used, int quota_claim)
331 {
332         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
333         struct ext4_inode_info *ei = EXT4_I(inode);
334
335         spin_lock(&ei->i_block_reservation_lock);
336         trace_ext4_da_update_reserve_space(inode, used, quota_claim);
337         if (unlikely(used > ei->i_reserved_data_blocks)) {
338                 ext4_warning(inode->i_sb, "%s: ino %lu, used %d "
339                          "with only %d reserved data blocks",
340                          __func__, inode->i_ino, used,
341                          ei->i_reserved_data_blocks);
342                 WARN_ON(1);
343                 used = ei->i_reserved_data_blocks;
344         }
345
346         /* Update per-inode reservations */
347         ei->i_reserved_data_blocks -= used;
348         percpu_counter_sub(&sbi->s_dirtyclusters_counter, used);
349
350         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
351
352         /* Update quota subsystem for data blocks */
353         if (quota_claim)
354                 dquot_claim_block(inode, EXT4_C2B(sbi, used));
355         else {
356                 /*
357                  * We did fallocate with an offset that is already delayed
358                  * allocated. So on delayed allocated writeback we should
359                  * not re-claim the quota for fallocated blocks.
360                  */
361                 dquot_release_reservation_block(inode, EXT4_C2B(sbi, used));
362         }
363
364         /*
365          * If we have done all the pending block allocations and if
366          * there aren't any writers on the inode, we can discard the
367          * inode's preallocations.
368          */
369         if ((ei->i_reserved_data_blocks == 0) &&
370             (atomic_read(&inode->i_writecount) == 0))
371                 ext4_discard_preallocations(inode);
372 }
373
374 static int __check_block_validity(struct inode *inode, const char *func,
375                                 unsigned int line,
376                                 struct ext4_map_blocks *map)
377 {
378         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
379                                    map->m_len)) {
380                 ext4_error_inode(inode, func, line, map->m_pblk,
381                                  "lblock %lu mapped to illegal pblock "
382                                  "(length %d)", (unsigned long) map->m_lblk,
383                                  map->m_len);
384                 return -EFSCORRUPTED;
385         }
386         return 0;
387 }
388
389 int ext4_issue_zeroout(struct inode *inode, ext4_lblk_t lblk, ext4_fsblk_t pblk,
390                        ext4_lblk_t len)
391 {
392         int ret;
393
394         if (ext4_encrypted_inode(inode))
395                 return fscrypt_zeroout_range(inode, lblk, pblk, len);
396
397         ret = sb_issue_zeroout(inode->i_sb, pblk, len, GFP_NOFS);
398         if (ret > 0)
399                 ret = 0;
400
401         return ret;
402 }
403
404 #define check_block_validity(inode, map)        \
405         __check_block_validity((inode), __func__, __LINE__, (map))
406
407 #ifdef ES_AGGRESSIVE_TEST
408 static void ext4_map_blocks_es_recheck(handle_t *handle,
409                                        struct inode *inode,
410                                        struct ext4_map_blocks *es_map,
411                                        struct ext4_map_blocks *map,
412                                        int flags)
413 {
414         int retval;
415
416         map->m_flags = 0;
417         /*
418          * There is a race window that the result is not the same.
419          * e.g. xfstests #223 when dioread_nolock enables.  The reason
420          * is that we lookup a block mapping in extent status tree with
421          * out taking i_data_sem.  So at the time the unwritten extent
422          * could be converted.
423          */
424         down_read(&EXT4_I(inode)->i_data_sem);
425         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
426                 retval = ext4_ext_map_blocks(handle, inode, map, flags &
427                                              EXT4_GET_BLOCKS_KEEP_SIZE);
428         } else {
429                 retval = ext4_ind_map_blocks(handle, inode, map, flags &
430                                              EXT4_GET_BLOCKS_KEEP_SIZE);
431         }
432         up_read((&EXT4_I(inode)->i_data_sem));
433
434         /*
435          * We don't check m_len because extent will be collpased in status
436          * tree.  So the m_len might not equal.
437          */
438         if (es_map->m_lblk != map->m_lblk ||
439             es_map->m_flags != map->m_flags ||
440             es_map->m_pblk != map->m_pblk) {
441                 printk("ES cache assertion failed for inode: %lu "
442                        "es_cached ex [%d/%d/%llu/%x] != "
443                        "found ex [%d/%d/%llu/%x] retval %d flags %x\n",
444                        inode->i_ino, es_map->m_lblk, es_map->m_len,
445                        es_map->m_pblk, es_map->m_flags, map->m_lblk,
446                        map->m_len, map->m_pblk, map->m_flags,
447                        retval, flags);
448         }
449 }
450 #endif /* ES_AGGRESSIVE_TEST */
451
452 /*
453  * The ext4_map_blocks() function tries to look up the requested blocks,
454  * and returns if the blocks are already mapped.
455  *
456  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
457  * and store the allocated blocks in the result buffer head and mark it
458  * mapped.
459  *
460  * If file type is extents based, it will call ext4_ext_map_blocks(),
461  * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
462  * based files
463  *
464  * On success, it returns the number of blocks being mapped or allocated.  if
465  * create==0 and the blocks are pre-allocated and unwritten, the resulting @map
466  * is marked as unwritten. If the create == 1, it will mark @map as mapped.
467  *
468  * It returns 0 if plain look up failed (blocks have not been allocated), in
469  * that case, @map is returned as unmapped but we still do fill map->m_len to
470  * indicate the length of a hole starting at map->m_lblk.
471  *
472  * It returns the error in case of allocation failure.
473  */
474 int ext4_map_blocks(handle_t *handle, struct inode *inode,
475                     struct ext4_map_blocks *map, int flags)
476 {
477         struct extent_status es;
478         int retval;
479         int ret = 0;
480 #ifdef ES_AGGRESSIVE_TEST
481         struct ext4_map_blocks orig_map;
482
483         memcpy(&orig_map, map, sizeof(*map));
484 #endif
485
486         map->m_flags = 0;
487         ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
488                   "logical block %lu\n", inode->i_ino, flags, map->m_len,
489                   (unsigned long) map->m_lblk);
490
491         /*
492          * ext4_map_blocks returns an int, and m_len is an unsigned int
493          */
494         if (unlikely(map->m_len > INT_MAX))
495                 map->m_len = INT_MAX;
496
497         /* We can handle the block number less than EXT_MAX_BLOCKS */
498         if (unlikely(map->m_lblk >= EXT_MAX_BLOCKS))
499                 return -EFSCORRUPTED;
500
501         /* Lookup extent status tree firstly */
502         if (ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
503                 if (ext4_es_is_written(&es) || ext4_es_is_unwritten(&es)) {
504                         map->m_pblk = ext4_es_pblock(&es) +
505                                         map->m_lblk - es.es_lblk;
506                         map->m_flags |= ext4_es_is_written(&es) ?
507                                         EXT4_MAP_MAPPED : EXT4_MAP_UNWRITTEN;
508                         retval = es.es_len - (map->m_lblk - es.es_lblk);
509                         if (retval > map->m_len)
510                                 retval = map->m_len;
511                         map->m_len = retval;
512                 } else if (ext4_es_is_delayed(&es) || ext4_es_is_hole(&es)) {
513                         map->m_pblk = 0;
514                         retval = es.es_len - (map->m_lblk - es.es_lblk);
515                         if (retval > map->m_len)
516                                 retval = map->m_len;
517                         map->m_len = retval;
518                         retval = 0;
519                 } else {
520                         BUG_ON(1);
521                 }
522 #ifdef ES_AGGRESSIVE_TEST
523                 ext4_map_blocks_es_recheck(handle, inode, map,
524                                            &orig_map, flags);
525 #endif
526                 goto found;
527         }
528
529         /*
530          * Try to see if we can get the block without requesting a new
531          * file system block.
532          */
533         down_read(&EXT4_I(inode)->i_data_sem);
534         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
535                 retval = ext4_ext_map_blocks(handle, inode, map, flags &
536                                              EXT4_GET_BLOCKS_KEEP_SIZE);
537         } else {
538                 retval = ext4_ind_map_blocks(handle, inode, map, flags &
539                                              EXT4_GET_BLOCKS_KEEP_SIZE);
540         }
541         if (retval > 0) {
542                 unsigned int status;
543
544                 if (unlikely(retval != map->m_len)) {
545                         ext4_warning(inode->i_sb,
546                                      "ES len assertion failed for inode "
547                                      "%lu: retval %d != map->m_len %d",
548                                      inode->i_ino, retval, map->m_len);
549                         WARN_ON(1);
550                 }
551
552                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
553                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
554                 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
555                     !(status & EXTENT_STATUS_WRITTEN) &&
556                     ext4_find_delalloc_range(inode, map->m_lblk,
557                                              map->m_lblk + map->m_len - 1))
558                         status |= EXTENT_STATUS_DELAYED;
559                 ret = ext4_es_insert_extent(inode, map->m_lblk,
560                                             map->m_len, map->m_pblk, status);
561                 if (ret < 0)
562                         retval = ret;
563         }
564         up_read((&EXT4_I(inode)->i_data_sem));
565
566 found:
567         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
568                 ret = check_block_validity(inode, map);
569                 if (ret != 0)
570                         return ret;
571         }
572
573         /* If it is only a block(s) look up */
574         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
575                 return retval;
576
577         /*
578          * Returns if the blocks have already allocated
579          *
580          * Note that if blocks have been preallocated
581          * ext4_ext_get_block() returns the create = 0
582          * with buffer head unmapped.
583          */
584         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
585                 /*
586                  * If we need to convert extent to unwritten
587                  * we continue and do the actual work in
588                  * ext4_ext_map_blocks()
589                  */
590                 if (!(flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN))
591                         return retval;
592
593         /*
594          * Here we clear m_flags because after allocating an new extent,
595          * it will be set again.
596          */
597         map->m_flags &= ~EXT4_MAP_FLAGS;
598
599         /*
600          * New blocks allocate and/or writing to unwritten extent
601          * will possibly result in updating i_data, so we take
602          * the write lock of i_data_sem, and call get_block()
603          * with create == 1 flag.
604          */
605         down_write(&EXT4_I(inode)->i_data_sem);
606
607         /*
608          * We need to check for EXT4 here because migrate
609          * could have changed the inode type in between
610          */
611         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
612                 retval = ext4_ext_map_blocks(handle, inode, map, flags);
613         } else {
614                 retval = ext4_ind_map_blocks(handle, inode, map, flags);
615
616                 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
617                         /*
618                          * We allocated new blocks which will result in
619                          * i_data's format changing.  Force the migrate
620                          * to fail by clearing migrate flags
621                          */
622                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
623                 }
624
625                 /*
626                  * Update reserved blocks/metadata blocks after successful
627                  * block allocation which had been deferred till now. We don't
628                  * support fallocate for non extent files. So we can update
629                  * reserve space here.
630                  */
631                 if ((retval > 0) &&
632                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
633                         ext4_da_update_reserve_space(inode, retval, 1);
634         }
635
636         if (retval > 0) {
637                 unsigned int status;
638
639                 if (unlikely(retval != map->m_len)) {
640                         ext4_warning(inode->i_sb,
641                                      "ES len assertion failed for inode "
642                                      "%lu: retval %d != map->m_len %d",
643                                      inode->i_ino, retval, map->m_len);
644                         WARN_ON(1);
645                 }
646
647                 /*
648                  * We have to zeroout blocks before inserting them into extent
649                  * status tree. Otherwise someone could look them up there and
650                  * use them before they are really zeroed.
651                  */
652                 if (flags & EXT4_GET_BLOCKS_ZERO &&
653                     map->m_flags & EXT4_MAP_MAPPED &&
654                     map->m_flags & EXT4_MAP_NEW) {
655                         ret = ext4_issue_zeroout(inode, map->m_lblk,
656                                                  map->m_pblk, map->m_len);
657                         if (ret) {
658                                 retval = ret;
659                                 goto out_sem;
660                         }
661                 }
662
663                 /*
664                  * If the extent has been zeroed out, we don't need to update
665                  * extent status tree.
666                  */
667                 if ((flags & EXT4_GET_BLOCKS_PRE_IO) &&
668                     ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
669                         if (ext4_es_is_written(&es))
670                                 goto out_sem;
671                 }
672                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
673                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
674                 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
675                     !(status & EXTENT_STATUS_WRITTEN) &&
676                     ext4_find_delalloc_range(inode, map->m_lblk,
677                                              map->m_lblk + map->m_len - 1))
678                         status |= EXTENT_STATUS_DELAYED;
679                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
680                                             map->m_pblk, status);
681                 if (ret < 0) {
682                         retval = ret;
683                         goto out_sem;
684                 }
685         }
686
687 out_sem:
688         up_write((&EXT4_I(inode)->i_data_sem));
689         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
690                 ret = check_block_validity(inode, map);
691                 if (ret != 0)
692                         return ret;
693
694                 /*
695                  * Inodes with freshly allocated blocks where contents will be
696                  * visible after transaction commit must be on transaction's
697                  * ordered data list.
698                  */
699                 if (map->m_flags & EXT4_MAP_NEW &&
700                     !(map->m_flags & EXT4_MAP_UNWRITTEN) &&
701                     !(flags & EXT4_GET_BLOCKS_ZERO) &&
702                     !IS_NOQUOTA(inode) &&
703                     ext4_should_order_data(inode)) {
704                         if (flags & EXT4_GET_BLOCKS_IO_SUBMIT)
705                                 ret = ext4_jbd2_inode_add_wait(handle, inode);
706                         else
707                                 ret = ext4_jbd2_inode_add_write(handle, inode);
708                         if (ret)
709                                 return ret;
710                 }
711         }
712         return retval;
713 }
714
715 /*
716  * Update EXT4_MAP_FLAGS in bh->b_state. For buffer heads attached to pages
717  * we have to be careful as someone else may be manipulating b_state as well.
718  */
719 static void ext4_update_bh_state(struct buffer_head *bh, unsigned long flags)
720 {
721         unsigned long old_state;
722         unsigned long new_state;
723
724         flags &= EXT4_MAP_FLAGS;
725
726         /* Dummy buffer_head? Set non-atomically. */
727         if (!bh->b_page) {
728                 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | flags;
729                 return;
730         }
731         /*
732          * Someone else may be modifying b_state. Be careful! This is ugly but
733          * once we get rid of using bh as a container for mapping information
734          * to pass to / from get_block functions, this can go away.
735          */
736         do {
737                 old_state = READ_ONCE(bh->b_state);
738                 new_state = (old_state & ~EXT4_MAP_FLAGS) | flags;
739         } while (unlikely(
740                  cmpxchg(&bh->b_state, old_state, new_state) != old_state));
741 }
742
743 static int _ext4_get_block(struct inode *inode, sector_t iblock,
744                            struct buffer_head *bh, int flags)
745 {
746         struct ext4_map_blocks map;
747         int ret = 0;
748
749         if (ext4_has_inline_data(inode))
750                 return -ERANGE;
751
752         map.m_lblk = iblock;
753         map.m_len = bh->b_size >> inode->i_blkbits;
754
755         ret = ext4_map_blocks(ext4_journal_current_handle(), inode, &map,
756                               flags);
757         if (ret > 0) {
758                 map_bh(bh, inode->i_sb, map.m_pblk);
759                 ext4_update_bh_state(bh, map.m_flags);
760                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
761                 ret = 0;
762         }
763         return ret;
764 }
765
766 int ext4_get_block(struct inode *inode, sector_t iblock,
767                    struct buffer_head *bh, int create)
768 {
769         return _ext4_get_block(inode, iblock, bh,
770                                create ? EXT4_GET_BLOCKS_CREATE : 0);
771 }
772
773 /*
774  * Get block function used when preparing for buffered write if we require
775  * creating an unwritten extent if blocks haven't been allocated.  The extent
776  * will be converted to written after the IO is complete.
777  */
778 int ext4_get_block_unwritten(struct inode *inode, sector_t iblock,
779                              struct buffer_head *bh_result, int create)
780 {
781         ext4_debug("ext4_get_block_unwritten: inode %lu, create flag %d\n",
782                    inode->i_ino, create);
783         return _ext4_get_block(inode, iblock, bh_result,
784                                EXT4_GET_BLOCKS_IO_CREATE_EXT);
785 }
786
787 /* Maximum number of blocks we map for direct IO at once. */
788 #define DIO_MAX_BLOCKS 4096
789
790 /*
791  * Get blocks function for the cases that need to start a transaction -
792  * generally difference cases of direct IO and DAX IO. It also handles retries
793  * in case of ENOSPC.
794  */
795 static int ext4_get_block_trans(struct inode *inode, sector_t iblock,
796                                 struct buffer_head *bh_result, int flags)
797 {
798         int dio_credits;
799         handle_t *handle;
800         int retries = 0;
801         int ret;
802
803         /* Trim mapping request to maximum we can map at once for DIO */
804         if (bh_result->b_size >> inode->i_blkbits > DIO_MAX_BLOCKS)
805                 bh_result->b_size = DIO_MAX_BLOCKS << inode->i_blkbits;
806         dio_credits = ext4_chunk_trans_blocks(inode,
807                                       bh_result->b_size >> inode->i_blkbits);
808 retry:
809         handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, dio_credits);
810         if (IS_ERR(handle))
811                 return PTR_ERR(handle);
812
813         ret = _ext4_get_block(inode, iblock, bh_result, flags);
814         ext4_journal_stop(handle);
815
816         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
817                 goto retry;
818         return ret;
819 }
820
821 /* Get block function for DIO reads and writes to inodes without extents */
822 int ext4_dio_get_block(struct inode *inode, sector_t iblock,
823                        struct buffer_head *bh, int create)
824 {
825         /* We don't expect handle for direct IO */
826         WARN_ON_ONCE(ext4_journal_current_handle());
827
828         if (!create)
829                 return _ext4_get_block(inode, iblock, bh, 0);
830         return ext4_get_block_trans(inode, iblock, bh, EXT4_GET_BLOCKS_CREATE);
831 }
832
833 /*
834  * Get block function for AIO DIO writes when we create unwritten extent if
835  * blocks are not allocated yet. The extent will be converted to written
836  * after IO is complete.
837  */
838 static int ext4_dio_get_block_unwritten_async(struct inode *inode,
839                 sector_t iblock, struct buffer_head *bh_result, int create)
840 {
841         int ret;
842
843         /* We don't expect handle for direct IO */
844         WARN_ON_ONCE(ext4_journal_current_handle());
845
846         ret = ext4_get_block_trans(inode, iblock, bh_result,
847                                    EXT4_GET_BLOCKS_IO_CREATE_EXT);
848
849         /*
850          * When doing DIO using unwritten extents, we need io_end to convert
851          * unwritten extents to written on IO completion. We allocate io_end
852          * once we spot unwritten extent and store it in b_private. Generic
853          * DIO code keeps b_private set and furthermore passes the value to
854          * our completion callback in 'private' argument.
855          */
856         if (!ret && buffer_unwritten(bh_result)) {
857                 if (!bh_result->b_private) {
858                         ext4_io_end_t *io_end;
859
860                         io_end = ext4_init_io_end(inode, GFP_KERNEL);
861                         if (!io_end)
862                                 return -ENOMEM;
863                         bh_result->b_private = io_end;
864                         ext4_set_io_unwritten_flag(inode, io_end);
865                 }
866                 set_buffer_defer_completion(bh_result);
867         }
868
869         return ret;
870 }
871
872 /*
873  * Get block function for non-AIO DIO writes when we create unwritten extent if
874  * blocks are not allocated yet. The extent will be converted to written
875  * after IO is complete from ext4_ext_direct_IO() function.
876  */
877 static int ext4_dio_get_block_unwritten_sync(struct inode *inode,
878                 sector_t iblock, struct buffer_head *bh_result, int create)
879 {
880         int ret;
881
882         /* We don't expect handle for direct IO */
883         WARN_ON_ONCE(ext4_journal_current_handle());
884
885         ret = ext4_get_block_trans(inode, iblock, bh_result,
886                                    EXT4_GET_BLOCKS_IO_CREATE_EXT);
887
888         /*
889          * Mark inode as having pending DIO writes to unwritten extents.
890          * ext4_ext_direct_IO() checks this flag and converts extents to
891          * written.
892          */
893         if (!ret && buffer_unwritten(bh_result))
894                 ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
895
896         return ret;
897 }
898
899 static int ext4_dio_get_block_overwrite(struct inode *inode, sector_t iblock,
900                    struct buffer_head *bh_result, int create)
901 {
902         int ret;
903
904         ext4_debug("ext4_dio_get_block_overwrite: inode %lu, create flag %d\n",
905                    inode->i_ino, create);
906         /* We don't expect handle for direct IO */
907         WARN_ON_ONCE(ext4_journal_current_handle());
908
909         ret = _ext4_get_block(inode, iblock, bh_result, 0);
910         /*
911          * Blocks should have been preallocated! ext4_file_write_iter() checks
912          * that.
913          */
914         WARN_ON_ONCE(!buffer_mapped(bh_result) || buffer_unwritten(bh_result));
915
916         return ret;
917 }
918
919
920 /*
921  * `handle' can be NULL if create is zero
922  */
923 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
924                                 ext4_lblk_t block, int map_flags)
925 {
926         struct ext4_map_blocks map;
927         struct buffer_head *bh;
928         int create = map_flags & EXT4_GET_BLOCKS_CREATE;
929         int err;
930
931         J_ASSERT(handle != NULL || create == 0);
932
933         map.m_lblk = block;
934         map.m_len = 1;
935         err = ext4_map_blocks(handle, inode, &map, map_flags);
936
937         if (err == 0)
938                 return create ? ERR_PTR(-ENOSPC) : NULL;
939         if (err < 0)
940                 return ERR_PTR(err);
941
942         bh = sb_getblk(inode->i_sb, map.m_pblk);
943         if (unlikely(!bh))
944                 return ERR_PTR(-ENOMEM);
945         if (map.m_flags & EXT4_MAP_NEW) {
946                 J_ASSERT(create != 0);
947                 J_ASSERT(handle != NULL);
948
949                 /*
950                  * Now that we do not always journal data, we should
951                  * keep in mind whether this should always journal the
952                  * new buffer as metadata.  For now, regular file
953                  * writes use ext4_get_block instead, so it's not a
954                  * problem.
955                  */
956                 lock_buffer(bh);
957                 BUFFER_TRACE(bh, "call get_create_access");
958                 err = ext4_journal_get_create_access(handle, bh);
959                 if (unlikely(err)) {
960                         unlock_buffer(bh);
961                         goto errout;
962                 }
963                 if (!buffer_uptodate(bh)) {
964                         memset(bh->b_data, 0, inode->i_sb->s_blocksize);
965                         set_buffer_uptodate(bh);
966                 }
967                 unlock_buffer(bh);
968                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
969                 err = ext4_handle_dirty_metadata(handle, inode, bh);
970                 if (unlikely(err))
971                         goto errout;
972         } else
973                 BUFFER_TRACE(bh, "not a new buffer");
974         return bh;
975 errout:
976         brelse(bh);
977         return ERR_PTR(err);
978 }
979
980 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
981                                ext4_lblk_t block, int map_flags)
982 {
983         struct buffer_head *bh;
984
985         bh = ext4_getblk(handle, inode, block, map_flags);
986         if (IS_ERR(bh))
987                 return bh;
988         if (!bh || buffer_uptodate(bh))
989                 return bh;
990         ll_rw_block(REQ_OP_READ, REQ_META | REQ_PRIO, 1, &bh);
991         wait_on_buffer(bh);
992         if (buffer_uptodate(bh))
993                 return bh;
994         put_bh(bh);
995         return ERR_PTR(-EIO);
996 }
997
998 int ext4_walk_page_buffers(handle_t *handle,
999                            struct buffer_head *head,
1000                            unsigned from,
1001                            unsigned to,
1002                            int *partial,
1003                            int (*fn)(handle_t *handle,
1004                                      struct buffer_head *bh))
1005 {
1006         struct buffer_head *bh;
1007         unsigned block_start, block_end;
1008         unsigned blocksize = head->b_size;
1009         int err, ret = 0;
1010         struct buffer_head *next;
1011
1012         for (bh = head, block_start = 0;
1013              ret == 0 && (bh != head || !block_start);
1014              block_start = block_end, bh = next) {
1015                 next = bh->b_this_page;
1016                 block_end = block_start + blocksize;
1017                 if (block_end <= from || block_start >= to) {
1018                         if (partial && !buffer_uptodate(bh))
1019                                 *partial = 1;
1020                         continue;
1021                 }
1022                 err = (*fn)(handle, bh);
1023                 if (!ret)
1024                         ret = err;
1025         }
1026         return ret;
1027 }
1028
1029 /*
1030  * To preserve ordering, it is essential that the hole instantiation and
1031  * the data write be encapsulated in a single transaction.  We cannot
1032  * close off a transaction and start a new one between the ext4_get_block()
1033  * and the commit_write().  So doing the jbd2_journal_start at the start of
1034  * prepare_write() is the right place.
1035  *
1036  * Also, this function can nest inside ext4_writepage().  In that case, we
1037  * *know* that ext4_writepage() has generated enough buffer credits to do the
1038  * whole page.  So we won't block on the journal in that case, which is good,
1039  * because the caller may be PF_MEMALLOC.
1040  *
1041  * By accident, ext4 can be reentered when a transaction is open via
1042  * quota file writes.  If we were to commit the transaction while thus
1043  * reentered, there can be a deadlock - we would be holding a quota
1044  * lock, and the commit would never complete if another thread had a
1045  * transaction open and was blocking on the quota lock - a ranking
1046  * violation.
1047  *
1048  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1049  * will _not_ run commit under these circumstances because handle->h_ref
1050  * is elevated.  We'll still have enough credits for the tiny quotafile
1051  * write.
1052  */
1053 int do_journal_get_write_access(handle_t *handle,
1054                                 struct buffer_head *bh)
1055 {
1056         int dirty = buffer_dirty(bh);
1057         int ret;
1058
1059         if (!buffer_mapped(bh) || buffer_freed(bh))
1060                 return 0;
1061         /*
1062          * __block_write_begin() could have dirtied some buffers. Clean
1063          * the dirty bit as jbd2_journal_get_write_access() could complain
1064          * otherwise about fs integrity issues. Setting of the dirty bit
1065          * by __block_write_begin() isn't a real problem here as we clear
1066          * the bit before releasing a page lock and thus writeback cannot
1067          * ever write the buffer.
1068          */
1069         if (dirty)
1070                 clear_buffer_dirty(bh);
1071         BUFFER_TRACE(bh, "get write access");
1072         ret = ext4_journal_get_write_access(handle, bh);
1073         if (!ret && dirty)
1074                 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1075         return ret;
1076 }
1077
1078 #ifdef CONFIG_EXT4_FS_ENCRYPTION
1079 static int ext4_block_write_begin(struct page *page, loff_t pos, unsigned len,
1080                                   get_block_t *get_block)
1081 {
1082         unsigned from = pos & (PAGE_SIZE - 1);
1083         unsigned to = from + len;
1084         struct inode *inode = page->mapping->host;
1085         unsigned block_start, block_end;
1086         sector_t block;
1087         int err = 0;
1088         unsigned blocksize = inode->i_sb->s_blocksize;
1089         unsigned bbits;
1090         struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
1091         bool decrypt = false;
1092
1093         BUG_ON(!PageLocked(page));
1094         BUG_ON(from > PAGE_SIZE);
1095         BUG_ON(to > PAGE_SIZE);
1096         BUG_ON(from > to);
1097
1098         if (!page_has_buffers(page))
1099                 create_empty_buffers(page, blocksize, 0);
1100         head = page_buffers(page);
1101         bbits = ilog2(blocksize);
1102         block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1103
1104         for (bh = head, block_start = 0; bh != head || !block_start;
1105             block++, block_start = block_end, bh = bh->b_this_page) {
1106                 block_end = block_start + blocksize;
1107                 if (block_end <= from || block_start >= to) {
1108                         if (PageUptodate(page)) {
1109                                 if (!buffer_uptodate(bh))
1110                                         set_buffer_uptodate(bh);
1111                         }
1112                         continue;
1113                 }
1114                 if (buffer_new(bh))
1115                         clear_buffer_new(bh);
1116                 if (!buffer_mapped(bh)) {
1117                         WARN_ON(bh->b_size != blocksize);
1118                         err = get_block(inode, block, bh, 1);
1119                         if (err)
1120                                 break;
1121                         if (buffer_new(bh)) {
1122                                 unmap_underlying_metadata(bh->b_bdev,
1123                                                           bh->b_blocknr);
1124                                 if (PageUptodate(page)) {
1125                                         clear_buffer_new(bh);
1126                                         set_buffer_uptodate(bh);
1127                                         mark_buffer_dirty(bh);
1128                                         continue;
1129                                 }
1130                                 if (block_end > to || block_start < from)
1131                                         zero_user_segments(page, to, block_end,
1132                                                            block_start, from);
1133                                 continue;
1134                         }
1135                 }
1136                 if (PageUptodate(page)) {
1137                         if (!buffer_uptodate(bh))
1138                                 set_buffer_uptodate(bh);
1139                         continue;
1140                 }
1141                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1142                     !buffer_unwritten(bh) &&
1143                     (block_start < from || block_end > to)) {
1144                         ll_rw_block(REQ_OP_READ, 0, 1, &bh);
1145                         *wait_bh++ = bh;
1146                         decrypt = ext4_encrypted_inode(inode) &&
1147                                 S_ISREG(inode->i_mode);
1148                 }
1149         }
1150         /*
1151          * If we issued read requests, let them complete.
1152          */
1153         while (wait_bh > wait) {
1154                 wait_on_buffer(*--wait_bh);
1155                 if (!buffer_uptodate(*wait_bh))
1156                         err = -EIO;
1157         }
1158         if (unlikely(err))
1159                 page_zero_new_buffers(page, from, to);
1160         else if (decrypt)
1161                 err = fscrypt_decrypt_page(page);
1162         return err;
1163 }
1164 #endif
1165
1166 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1167                             loff_t pos, unsigned len, unsigned flags,
1168                             struct page **pagep, void **fsdata)
1169 {
1170         struct inode *inode = mapping->host;
1171         int ret, needed_blocks;
1172         handle_t *handle;
1173         int retries = 0;
1174         struct page *page;
1175         pgoff_t index;
1176         unsigned from, to;
1177
1178         trace_ext4_write_begin(inode, pos, len, flags);
1179         /*
1180          * Reserve one block more for addition to orphan list in case
1181          * we allocate blocks but write fails for some reason
1182          */
1183         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1184         index = pos >> PAGE_SHIFT;
1185         from = pos & (PAGE_SIZE - 1);
1186         to = from + len;
1187
1188         if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
1189                 ret = ext4_try_to_write_inline_data(mapping, inode, pos, len,
1190                                                     flags, pagep);
1191                 if (ret < 0)
1192                         return ret;
1193                 if (ret == 1)
1194                         return 0;
1195         }
1196
1197         /*
1198          * grab_cache_page_write_begin() can take a long time if the
1199          * system is thrashing due to memory pressure, or if the page
1200          * is being written back.  So grab it first before we start
1201          * the transaction handle.  This also allows us to allocate
1202          * the page (if needed) without using GFP_NOFS.
1203          */
1204 retry_grab:
1205         page = grab_cache_page_write_begin(mapping, index, flags);
1206         if (!page)
1207                 return -ENOMEM;
1208         unlock_page(page);
1209
1210 retry_journal:
1211         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE, needed_blocks);
1212         if (IS_ERR(handle)) {
1213                 put_page(page);
1214                 return PTR_ERR(handle);
1215         }
1216
1217         lock_page(page);
1218         if (page->mapping != mapping) {
1219                 /* The page got truncated from under us */
1220                 unlock_page(page);
1221                 put_page(page);
1222                 ext4_journal_stop(handle);
1223                 goto retry_grab;
1224         }
1225         /* In case writeback began while the page was unlocked */
1226         wait_for_stable_page(page);
1227
1228 #ifdef CONFIG_EXT4_FS_ENCRYPTION
1229         if (ext4_should_dioread_nolock(inode))
1230                 ret = ext4_block_write_begin(page, pos, len,
1231                                              ext4_get_block_unwritten);
1232         else
1233                 ret = ext4_block_write_begin(page, pos, len,
1234                                              ext4_get_block);
1235 #else
1236         if (ext4_should_dioread_nolock(inode))
1237                 ret = __block_write_begin(page, pos, len,
1238                                           ext4_get_block_unwritten);
1239         else
1240                 ret = __block_write_begin(page, pos, len, ext4_get_block);
1241 #endif
1242         if (!ret && ext4_should_journal_data(inode)) {
1243                 ret = ext4_walk_page_buffers(handle, page_buffers(page),
1244                                              from, to, NULL,
1245                                              do_journal_get_write_access);
1246         }
1247
1248         if (ret) {
1249                 unlock_page(page);
1250                 /*
1251                  * __block_write_begin may have instantiated a few blocks
1252                  * outside i_size.  Trim these off again. Don't need
1253                  * i_size_read because we hold i_mutex.
1254                  *
1255                  * Add inode to orphan list in case we crash before
1256                  * truncate finishes
1257                  */
1258                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1259                         ext4_orphan_add(handle, inode);
1260
1261                 ext4_journal_stop(handle);
1262                 if (pos + len > inode->i_size) {
1263                         ext4_truncate_failed_write(inode);
1264                         /*
1265                          * If truncate failed early the inode might
1266                          * still be on the orphan list; we need to
1267                          * make sure the inode is removed from the
1268                          * orphan list in that case.
1269                          */
1270                         if (inode->i_nlink)
1271                                 ext4_orphan_del(NULL, inode);
1272                 }
1273
1274                 if (ret == -ENOSPC &&
1275                     ext4_should_retry_alloc(inode->i_sb, &retries))
1276                         goto retry_journal;
1277                 put_page(page);
1278                 return ret;
1279         }
1280         *pagep = page;
1281         return ret;
1282 }
1283
1284 /* For write_end() in data=journal mode */
1285 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1286 {
1287         int ret;
1288         if (!buffer_mapped(bh) || buffer_freed(bh))
1289                 return 0;
1290         set_buffer_uptodate(bh);
1291         ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1292         clear_buffer_meta(bh);
1293         clear_buffer_prio(bh);
1294         return ret;
1295 }
1296
1297 /*
1298  * We need to pick up the new inode size which generic_commit_write gave us
1299  * `file' can be NULL - eg, when called from page_symlink().
1300  *
1301  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1302  * buffers are managed internally.
1303  */
1304 static int ext4_write_end(struct file *file,
1305                           struct address_space *mapping,
1306                           loff_t pos, unsigned len, unsigned copied,
1307                           struct page *page, void *fsdata)
1308 {
1309         handle_t *handle = ext4_journal_current_handle();
1310         struct inode *inode = mapping->host;
1311         loff_t old_size = inode->i_size;
1312         int ret = 0, ret2;
1313         int i_size_changed = 0;
1314
1315         trace_ext4_write_end(inode, pos, len, copied);
1316         if (ext4_has_inline_data(inode)) {
1317                 ret = ext4_write_inline_data_end(inode, pos, len,
1318                                                  copied, page);
1319                 if (ret < 0)
1320                         goto errout;
1321                 copied = ret;
1322         } else
1323                 copied = block_write_end(file, mapping, pos,
1324                                          len, copied, page, fsdata);
1325         /*
1326          * it's important to update i_size while still holding page lock:
1327          * page writeout could otherwise come in and zero beyond i_size.
1328          */
1329         i_size_changed = ext4_update_inode_size(inode, pos + copied);
1330         unlock_page(page);
1331         put_page(page);
1332
1333         if (old_size < pos)
1334                 pagecache_isize_extended(inode, old_size, pos);
1335         /*
1336          * Don't mark the inode dirty under page lock. First, it unnecessarily
1337          * makes the holding time of page lock longer. Second, it forces lock
1338          * ordering of page lock and transaction start for journaling
1339          * filesystems.
1340          */
1341         if (i_size_changed)
1342                 ext4_mark_inode_dirty(handle, inode);
1343
1344         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1345                 /* if we have allocated more blocks and copied
1346                  * less. We will have blocks allocated outside
1347                  * inode->i_size. So truncate them
1348                  */
1349                 ext4_orphan_add(handle, inode);
1350 errout:
1351         ret2 = ext4_journal_stop(handle);
1352         if (!ret)
1353                 ret = ret2;
1354
1355         if (pos + len > inode->i_size) {
1356                 ext4_truncate_failed_write(inode);
1357                 /*
1358                  * If truncate failed early the inode might still be
1359                  * on the orphan list; we need to make sure the inode
1360                  * is removed from the orphan list in that case.
1361                  */
1362                 if (inode->i_nlink)
1363                         ext4_orphan_del(NULL, inode);
1364         }
1365
1366         return ret ? ret : copied;
1367 }
1368
1369 /*
1370  * This is a private version of page_zero_new_buffers() which doesn't
1371  * set the buffer to be dirty, since in data=journalled mode we need
1372  * to call ext4_handle_dirty_metadata() instead.
1373  */
1374 static void zero_new_buffers(struct page *page, unsigned from, unsigned to)
1375 {
1376         unsigned int block_start = 0, block_end;
1377         struct buffer_head *head, *bh;
1378
1379         bh = head = page_buffers(page);
1380         do {
1381                 block_end = block_start + bh->b_size;
1382                 if (buffer_new(bh)) {
1383                         if (block_end > from && block_start < to) {
1384                                 if (!PageUptodate(page)) {
1385                                         unsigned start, size;
1386
1387                                         start = max(from, block_start);
1388                                         size = min(to, block_end) - start;
1389
1390                                         zero_user(page, start, size);
1391                                         set_buffer_uptodate(bh);
1392                                 }
1393                                 clear_buffer_new(bh);
1394                         }
1395                 }
1396                 block_start = block_end;
1397                 bh = bh->b_this_page;
1398         } while (bh != head);
1399 }
1400
1401 static int ext4_journalled_write_end(struct file *file,
1402                                      struct address_space *mapping,
1403                                      loff_t pos, unsigned len, unsigned copied,
1404                                      struct page *page, void *fsdata)
1405 {
1406         handle_t *handle = ext4_journal_current_handle();
1407         struct inode *inode = mapping->host;
1408         loff_t old_size = inode->i_size;
1409         int ret = 0, ret2;
1410         int partial = 0;
1411         unsigned from, to;
1412         int size_changed = 0;
1413
1414         trace_ext4_journalled_write_end(inode, pos, len, copied);
1415         from = pos & (PAGE_SIZE - 1);
1416         to = from + len;
1417
1418         BUG_ON(!ext4_handle_valid(handle));
1419
1420         if (ext4_has_inline_data(inode))
1421                 copied = ext4_write_inline_data_end(inode, pos, len,
1422                                                     copied, page);
1423         else {
1424                 if (copied < len) {
1425                         if (!PageUptodate(page))
1426                                 copied = 0;
1427                         zero_new_buffers(page, from+copied, to);
1428                 }
1429
1430                 ret = ext4_walk_page_buffers(handle, page_buffers(page), from,
1431                                              to, &partial, write_end_fn);
1432                 if (!partial)
1433                         SetPageUptodate(page);
1434         }
1435         size_changed = ext4_update_inode_size(inode, pos + copied);
1436         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1437         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1438         unlock_page(page);
1439         put_page(page);
1440
1441         if (old_size < pos)
1442                 pagecache_isize_extended(inode, old_size, pos);
1443
1444         if (size_changed) {
1445                 ret2 = ext4_mark_inode_dirty(handle, inode);
1446                 if (!ret)
1447                         ret = ret2;
1448         }
1449
1450         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1451                 /* if we have allocated more blocks and copied
1452                  * less. We will have blocks allocated outside
1453                  * inode->i_size. So truncate them
1454                  */
1455                 ext4_orphan_add(handle, inode);
1456
1457         ret2 = ext4_journal_stop(handle);
1458         if (!ret)
1459                 ret = ret2;
1460         if (pos + len > inode->i_size) {
1461                 ext4_truncate_failed_write(inode);
1462                 /*
1463                  * If truncate failed early the inode might still be
1464                  * on the orphan list; we need to make sure the inode
1465                  * is removed from the orphan list in that case.
1466                  */
1467                 if (inode->i_nlink)
1468                         ext4_orphan_del(NULL, inode);
1469         }
1470
1471         return ret ? ret : copied;
1472 }
1473
1474 /*
1475  * Reserve space for a single cluster
1476  */
1477 static int ext4_da_reserve_space(struct inode *inode)
1478 {
1479         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1480         struct ext4_inode_info *ei = EXT4_I(inode);
1481         int ret;
1482
1483         /*
1484          * We will charge metadata quota at writeout time; this saves
1485          * us from metadata over-estimation, though we may go over by
1486          * a small amount in the end.  Here we just reserve for data.
1487          */
1488         ret = dquot_reserve_block(inode, EXT4_C2B(sbi, 1));
1489         if (ret)
1490                 return ret;
1491
1492         spin_lock(&ei->i_block_reservation_lock);
1493         if (ext4_claim_free_clusters(sbi, 1, 0)) {
1494                 spin_unlock(&ei->i_block_reservation_lock);
1495                 dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
1496                 return -ENOSPC;
1497         }
1498         ei->i_reserved_data_blocks++;
1499         trace_ext4_da_reserve_space(inode);
1500         spin_unlock(&ei->i_block_reservation_lock);
1501
1502         return 0;       /* success */
1503 }
1504
1505 static void ext4_da_release_space(struct inode *inode, int to_free)
1506 {
1507         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1508         struct ext4_inode_info *ei = EXT4_I(inode);
1509
1510         if (!to_free)
1511                 return;         /* Nothing to release, exit */
1512
1513         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1514
1515         trace_ext4_da_release_space(inode, to_free);
1516         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1517                 /*
1518                  * if there aren't enough reserved blocks, then the
1519                  * counter is messed up somewhere.  Since this
1520                  * function is called from invalidate page, it's
1521                  * harmless to return without any action.
1522                  */
1523                 ext4_warning(inode->i_sb, "ext4_da_release_space: "
1524                          "ino %lu, to_free %d with only %d reserved "
1525                          "data blocks", inode->i_ino, to_free,
1526                          ei->i_reserved_data_blocks);
1527                 WARN_ON(1);
1528                 to_free = ei->i_reserved_data_blocks;
1529         }
1530         ei->i_reserved_data_blocks -= to_free;
1531
1532         /* update fs dirty data blocks counter */
1533         percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1534
1535         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1536
1537         dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1538 }
1539
1540 static void ext4_da_page_release_reservation(struct page *page,
1541                                              unsigned int offset,
1542                                              unsigned int length)
1543 {
1544         int to_release = 0, contiguous_blks = 0;
1545         struct buffer_head *head, *bh;
1546         unsigned int curr_off = 0;
1547         struct inode *inode = page->mapping->host;
1548         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1549         unsigned int stop = offset + length;
1550         int num_clusters;
1551         ext4_fsblk_t lblk;
1552
1553         BUG_ON(stop > PAGE_SIZE || stop < length);
1554
1555         head = page_buffers(page);
1556         bh = head;
1557         do {
1558                 unsigned int next_off = curr_off + bh->b_size;
1559
1560                 if (next_off > stop)
1561                         break;
1562
1563                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1564                         to_release++;
1565                         contiguous_blks++;
1566                         clear_buffer_delay(bh);
1567                 } else if (contiguous_blks) {
1568                         lblk = page->index <<
1569                                (PAGE_SHIFT - inode->i_blkbits);
1570                         lblk += (curr_off >> inode->i_blkbits) -
1571                                 contiguous_blks;
1572                         ext4_es_remove_extent(inode, lblk, contiguous_blks);
1573                         contiguous_blks = 0;
1574                 }
1575                 curr_off = next_off;
1576         } while ((bh = bh->b_this_page) != head);
1577
1578         if (contiguous_blks) {
1579                 lblk = page->index << (PAGE_SHIFT - inode->i_blkbits);
1580                 lblk += (curr_off >> inode->i_blkbits) - contiguous_blks;
1581                 ext4_es_remove_extent(inode, lblk, contiguous_blks);
1582         }
1583
1584         /* If we have released all the blocks belonging to a cluster, then we
1585          * need to release the reserved space for that cluster. */
1586         num_clusters = EXT4_NUM_B2C(sbi, to_release);
1587         while (num_clusters > 0) {
1588                 lblk = (page->index << (PAGE_SHIFT - inode->i_blkbits)) +
1589                         ((num_clusters - 1) << sbi->s_cluster_bits);
1590                 if (sbi->s_cluster_ratio == 1 ||
1591                     !ext4_find_delalloc_cluster(inode, lblk))
1592                         ext4_da_release_space(inode, 1);
1593
1594                 num_clusters--;
1595         }
1596 }
1597
1598 /*
1599  * Delayed allocation stuff
1600  */
1601
1602 struct mpage_da_data {
1603         struct inode *inode;
1604         struct writeback_control *wbc;
1605
1606         pgoff_t first_page;     /* The first page to write */
1607         pgoff_t next_page;      /* Current page to examine */
1608         pgoff_t last_page;      /* Last page to examine */
1609         /*
1610          * Extent to map - this can be after first_page because that can be
1611          * fully mapped. We somewhat abuse m_flags to store whether the extent
1612          * is delalloc or unwritten.
1613          */
1614         struct ext4_map_blocks map;
1615         struct ext4_io_submit io_submit;        /* IO submission data */
1616 };
1617
1618 static void mpage_release_unused_pages(struct mpage_da_data *mpd,
1619                                        bool invalidate)
1620 {
1621         int nr_pages, i;
1622         pgoff_t index, end;
1623         struct pagevec pvec;
1624         struct inode *inode = mpd->inode;
1625         struct address_space *mapping = inode->i_mapping;
1626
1627         /* This is necessary when next_page == 0. */
1628         if (mpd->first_page >= mpd->next_page)
1629                 return;
1630
1631         index = mpd->first_page;
1632         end   = mpd->next_page - 1;
1633         if (invalidate) {
1634                 ext4_lblk_t start, last;
1635                 start = index << (PAGE_SHIFT - inode->i_blkbits);
1636                 last = end << (PAGE_SHIFT - inode->i_blkbits);
1637                 ext4_es_remove_extent(inode, start, last - start + 1);
1638         }
1639
1640         pagevec_init(&pvec, 0);
1641         while (index <= end) {
1642                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1643                 if (nr_pages == 0)
1644                         break;
1645                 for (i = 0; i < nr_pages; i++) {
1646                         struct page *page = pvec.pages[i];
1647                         if (page->index > end)
1648                                 break;
1649                         BUG_ON(!PageLocked(page));
1650                         BUG_ON(PageWriteback(page));
1651                         if (invalidate) {
1652                                 block_invalidatepage(page, 0, PAGE_SIZE);
1653                                 ClearPageUptodate(page);
1654                         }
1655                         unlock_page(page);
1656                 }
1657                 index = pvec.pages[nr_pages - 1]->index + 1;
1658                 pagevec_release(&pvec);
1659         }
1660 }
1661
1662 static void ext4_print_free_blocks(struct inode *inode)
1663 {
1664         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1665         struct super_block *sb = inode->i_sb;
1666         struct ext4_inode_info *ei = EXT4_I(inode);
1667
1668         ext4_msg(sb, KERN_CRIT, "Total free blocks count %lld",
1669                EXT4_C2B(EXT4_SB(inode->i_sb),
1670                         ext4_count_free_clusters(sb)));
1671         ext4_msg(sb, KERN_CRIT, "Free/Dirty block details");
1672         ext4_msg(sb, KERN_CRIT, "free_blocks=%lld",
1673                (long long) EXT4_C2B(EXT4_SB(sb),
1674                 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1675         ext4_msg(sb, KERN_CRIT, "dirty_blocks=%lld",
1676                (long long) EXT4_C2B(EXT4_SB(sb),
1677                 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1678         ext4_msg(sb, KERN_CRIT, "Block reservation details");
1679         ext4_msg(sb, KERN_CRIT, "i_reserved_data_blocks=%u",
1680                  ei->i_reserved_data_blocks);
1681         return;
1682 }
1683
1684 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1685 {
1686         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1687 }
1688
1689 /*
1690  * This function is grabs code from the very beginning of
1691  * ext4_map_blocks, but assumes that the caller is from delayed write
1692  * time. This function looks up the requested blocks and sets the
1693  * buffer delay bit under the protection of i_data_sem.
1694  */
1695 static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1696                               struct ext4_map_blocks *map,
1697                               struct buffer_head *bh)
1698 {
1699         struct extent_status es;
1700         int retval;
1701         sector_t invalid_block = ~((sector_t) 0xffff);
1702 #ifdef ES_AGGRESSIVE_TEST
1703         struct ext4_map_blocks orig_map;
1704
1705         memcpy(&orig_map, map, sizeof(*map));
1706 #endif
1707
1708         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1709                 invalid_block = ~0;
1710
1711         map->m_flags = 0;
1712         ext_debug("ext4_da_map_blocks(): inode %lu, max_blocks %u,"
1713                   "logical block %lu\n", inode->i_ino, map->m_len,
1714                   (unsigned long) map->m_lblk);
1715
1716         /* Lookup extent status tree firstly */
1717         if (ext4_es_lookup_extent(inode, iblock, &es)) {
1718                 if (ext4_es_is_hole(&es)) {
1719                         retval = 0;
1720                         down_read(&EXT4_I(inode)->i_data_sem);
1721                         goto add_delayed;
1722                 }
1723
1724                 /*
1725                  * Delayed extent could be allocated by fallocate.
1726                  * So we need to check it.
1727                  */
1728                 if (ext4_es_is_delayed(&es) && !ext4_es_is_unwritten(&es)) {
1729                         map_bh(bh, inode->i_sb, invalid_block);
1730                         set_buffer_new(bh);
1731                         set_buffer_delay(bh);
1732                         return 0;
1733                 }
1734
1735                 map->m_pblk = ext4_es_pblock(&es) + iblock - es.es_lblk;
1736                 retval = es.es_len - (iblock - es.es_lblk);
1737                 if (retval > map->m_len)
1738                         retval = map->m_len;
1739                 map->m_len = retval;
1740                 if (ext4_es_is_written(&es))
1741                         map->m_flags |= EXT4_MAP_MAPPED;
1742                 else if (ext4_es_is_unwritten(&es))
1743                         map->m_flags |= EXT4_MAP_UNWRITTEN;
1744                 else
1745                         BUG_ON(1);
1746
1747 #ifdef ES_AGGRESSIVE_TEST
1748                 ext4_map_blocks_es_recheck(NULL, inode, map, &orig_map, 0);
1749 #endif
1750                 return retval;
1751         }
1752
1753         /*
1754          * Try to see if we can get the block without requesting a new
1755          * file system block.
1756          */
1757         down_read(&EXT4_I(inode)->i_data_sem);
1758         if (ext4_has_inline_data(inode))
1759                 retval = 0;
1760         else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1761                 retval = ext4_ext_map_blocks(NULL, inode, map, 0);
1762         else
1763                 retval = ext4_ind_map_blocks(NULL, inode, map, 0);
1764
1765 add_delayed:
1766         if (retval == 0) {
1767                 int ret;
1768                 /*
1769                  * XXX: __block_prepare_write() unmaps passed block,
1770                  * is it OK?
1771                  */
1772                 /*
1773                  * If the block was allocated from previously allocated cluster,
1774                  * then we don't need to reserve it again. However we still need
1775                  * to reserve metadata for every block we're going to write.
1776                  */
1777                 if (EXT4_SB(inode->i_sb)->s_cluster_ratio == 1 ||
1778                     !ext4_find_delalloc_cluster(inode, map->m_lblk)) {
1779                         ret = ext4_da_reserve_space(inode);
1780                         if (ret) {
1781                                 /* not enough space to reserve */
1782                                 retval = ret;
1783                                 goto out_unlock;
1784                         }
1785                 }
1786
1787                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1788                                             ~0, EXTENT_STATUS_DELAYED);
1789                 if (ret) {
1790                         retval = ret;
1791                         goto out_unlock;
1792                 }
1793
1794                 map_bh(bh, inode->i_sb, invalid_block);
1795                 set_buffer_new(bh);
1796                 set_buffer_delay(bh);
1797         } else if (retval > 0) {
1798                 int ret;
1799                 unsigned int status;
1800
1801                 if (unlikely(retval != map->m_len)) {
1802                         ext4_warning(inode->i_sb,
1803                                      "ES len assertion failed for inode "
1804                                      "%lu: retval %d != map->m_len %d",
1805                                      inode->i_ino, retval, map->m_len);
1806                         WARN_ON(1);
1807                 }
1808
1809                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
1810                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
1811                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1812                                             map->m_pblk, status);
1813                 if (ret != 0)
1814                         retval = ret;
1815         }
1816
1817 out_unlock:
1818         up_read((&EXT4_I(inode)->i_data_sem));
1819
1820         return retval;
1821 }
1822
1823 /*
1824  * This is a special get_block_t callback which is used by
1825  * ext4_da_write_begin().  It will either return mapped block or
1826  * reserve space for a single block.
1827  *
1828  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1829  * We also have b_blocknr = -1 and b_bdev initialized properly
1830  *
1831  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1832  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1833  * initialized properly.
1834  */
1835 int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1836                            struct buffer_head *bh, int create)
1837 {
1838         struct ext4_map_blocks map;
1839         int ret = 0;
1840
1841         BUG_ON(create == 0);
1842         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1843
1844         map.m_lblk = iblock;
1845         map.m_len = 1;
1846
1847         /*
1848          * first, we need to know whether the block is allocated already
1849          * preallocated blocks are unmapped but should treated
1850          * the same as allocated blocks.
1851          */
1852         ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1853         if (ret <= 0)
1854                 return ret;
1855
1856         map_bh(bh, inode->i_sb, map.m_pblk);
1857         ext4_update_bh_state(bh, map.m_flags);
1858
1859         if (buffer_unwritten(bh)) {
1860                 /* A delayed write to unwritten bh should be marked
1861                  * new and mapped.  Mapped ensures that we don't do
1862                  * get_block multiple times when we write to the same
1863                  * offset and new ensures that we do proper zero out
1864                  * for partial write.
1865                  */
1866                 set_buffer_new(bh);
1867                 set_buffer_mapped(bh);
1868         }
1869         return 0;
1870 }
1871
1872 static int bget_one(handle_t *handle, struct buffer_head *bh)
1873 {
1874         get_bh(bh);
1875         return 0;
1876 }
1877
1878 static int bput_one(handle_t *handle, struct buffer_head *bh)
1879 {
1880         put_bh(bh);
1881         return 0;
1882 }
1883
1884 static int __ext4_journalled_writepage(struct page *page,
1885                                        unsigned int len)
1886 {
1887         struct address_space *mapping = page->mapping;
1888         struct inode *inode = mapping->host;
1889         struct buffer_head *page_bufs = NULL;
1890         handle_t *handle = NULL;
1891         int ret = 0, err = 0;
1892         int inline_data = ext4_has_inline_data(inode);
1893         struct buffer_head *inode_bh = NULL;
1894
1895         ClearPageChecked(page);
1896
1897         if (inline_data) {
1898                 BUG_ON(page->index != 0);
1899                 BUG_ON(len > ext4_get_max_inline_size(inode));
1900                 inode_bh = ext4_journalled_write_inline_data(inode, len, page);
1901                 if (inode_bh == NULL)
1902                         goto out;
1903         } else {
1904                 page_bufs = page_buffers(page);
1905                 if (!page_bufs) {
1906                         BUG();
1907                         goto out;
1908                 }
1909                 ext4_walk_page_buffers(handle, page_bufs, 0, len,
1910                                        NULL, bget_one);
1911         }
1912         /*
1913          * We need to release the page lock before we start the
1914          * journal, so grab a reference so the page won't disappear
1915          * out from under us.
1916          */
1917         get_page(page);
1918         unlock_page(page);
1919
1920         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
1921                                     ext4_writepage_trans_blocks(inode));
1922         if (IS_ERR(handle)) {
1923                 ret = PTR_ERR(handle);
1924                 put_page(page);
1925                 goto out_no_pagelock;
1926         }
1927         BUG_ON(!ext4_handle_valid(handle));
1928
1929         lock_page(page);
1930         put_page(page);
1931         if (page->mapping != mapping) {
1932                 /* The page got truncated from under us */
1933                 ext4_journal_stop(handle);
1934                 ret = 0;
1935                 goto out;
1936         }
1937
1938         if (inline_data) {
1939                 BUFFER_TRACE(inode_bh, "get write access");
1940                 ret = ext4_journal_get_write_access(handle, inode_bh);
1941
1942                 err = ext4_handle_dirty_metadata(handle, inode, inode_bh);
1943
1944         } else {
1945                 ret = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
1946                                              do_journal_get_write_access);
1947
1948                 err = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
1949                                              write_end_fn);
1950         }
1951         if (ret == 0)
1952                 ret = err;
1953         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1954         err = ext4_journal_stop(handle);
1955         if (!ret)
1956                 ret = err;
1957
1958         if (!ext4_has_inline_data(inode))
1959                 ext4_walk_page_buffers(NULL, page_bufs, 0, len,
1960                                        NULL, bput_one);
1961         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1962 out:
1963         unlock_page(page);
1964 out_no_pagelock:
1965         brelse(inode_bh);
1966         return ret;
1967 }
1968
1969 /*
1970  * Note that we don't need to start a transaction unless we're journaling data
1971  * because we should have holes filled from ext4_page_mkwrite(). We even don't
1972  * need to file the inode to the transaction's list in ordered mode because if
1973  * we are writing back data added by write(), the inode is already there and if
1974  * we are writing back data modified via mmap(), no one guarantees in which
1975  * transaction the data will hit the disk. In case we are journaling data, we
1976  * cannot start transaction directly because transaction start ranks above page
1977  * lock so we have to do some magic.
1978  *
1979  * This function can get called via...
1980  *   - ext4_writepages after taking page lock (have journal handle)
1981  *   - journal_submit_inode_data_buffers (no journal handle)
1982  *   - shrink_page_list via the kswapd/direct reclaim (no journal handle)
1983  *   - grab_page_cache when doing write_begin (have journal handle)
1984  *
1985  * We don't do any block allocation in this function. If we have page with
1986  * multiple blocks we need to write those buffer_heads that are mapped. This
1987  * is important for mmaped based write. So if we do with blocksize 1K
1988  * truncate(f, 1024);
1989  * a = mmap(f, 0, 4096);
1990  * a[0] = 'a';
1991  * truncate(f, 4096);
1992  * we have in the page first buffer_head mapped via page_mkwrite call back
1993  * but other buffer_heads would be unmapped but dirty (dirty done via the
1994  * do_wp_page). So writepage should write the first block. If we modify
1995  * the mmap area beyond 1024 we will again get a page_fault and the
1996  * page_mkwrite callback will do the block allocation and mark the
1997  * buffer_heads mapped.
1998  *
1999  * We redirty the page if we have any buffer_heads that is either delay or
2000  * unwritten in the page.
2001  *
2002  * We can get recursively called as show below.
2003  *
2004  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2005  *              ext4_writepage()
2006  *
2007  * But since we don't do any block allocation we should not deadlock.
2008  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2009  */
2010 static int ext4_writepage(struct page *page,
2011                           struct writeback_control *wbc)
2012 {
2013         int ret = 0;
2014         loff_t size;
2015         unsigned int len;
2016         struct buffer_head *page_bufs = NULL;
2017         struct inode *inode = page->mapping->host;
2018         struct ext4_io_submit io_submit;
2019         bool keep_towrite = false;
2020
2021         trace_ext4_writepage(page);
2022         size = i_size_read(inode);
2023         if (page->index == size >> PAGE_SHIFT)
2024                 len = size & ~PAGE_MASK;
2025         else
2026                 len = PAGE_SIZE;
2027
2028         page_bufs = page_buffers(page);
2029         /*
2030          * We cannot do block allocation or other extent handling in this
2031          * function. If there are buffers needing that, we have to redirty
2032          * the page. But we may reach here when we do a journal commit via
2033          * journal_submit_inode_data_buffers() and in that case we must write
2034          * allocated buffers to achieve data=ordered mode guarantees.
2035          *
2036          * Also, if there is only one buffer per page (the fs block
2037          * size == the page size), if one buffer needs block
2038          * allocation or needs to modify the extent tree to clear the
2039          * unwritten flag, we know that the page can't be written at
2040          * all, so we might as well refuse the write immediately.
2041          * Unfortunately if the block size != page size, we can't as
2042          * easily detect this case using ext4_walk_page_buffers(), but
2043          * for the extremely common case, this is an optimization that
2044          * skips a useless round trip through ext4_bio_write_page().
2045          */
2046         if (ext4_walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2047                                    ext4_bh_delay_or_unwritten)) {
2048                 redirty_page_for_writepage(wbc, page);
2049                 if ((current->flags & PF_MEMALLOC) ||
2050                     (inode->i_sb->s_blocksize == PAGE_SIZE)) {
2051                         /*
2052                          * For memory cleaning there's no point in writing only
2053                          * some buffers. So just bail out. Warn if we came here
2054                          * from direct reclaim.
2055                          */
2056                         WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD))
2057                                                         == PF_MEMALLOC);
2058                         unlock_page(page);
2059                         return 0;
2060                 }
2061                 keep_towrite = true;
2062         }
2063
2064         if (PageChecked(page) && ext4_should_journal_data(inode))
2065                 /*
2066                  * It's mmapped pagecache.  Add buffers and journal it.  There
2067                  * doesn't seem much point in redirtying the page here.
2068                  */
2069                 return __ext4_journalled_writepage(page, len);
2070
2071         ext4_io_submit_init(&io_submit, wbc);
2072         io_submit.io_end = ext4_init_io_end(inode, GFP_NOFS);
2073         if (!io_submit.io_end) {
2074                 redirty_page_for_writepage(wbc, page);
2075                 unlock_page(page);
2076                 return -ENOMEM;
2077         }
2078         ret = ext4_bio_write_page(&io_submit, page, len, wbc, keep_towrite);
2079         ext4_io_submit(&io_submit);
2080         /* Drop io_end reference we got from init */
2081         ext4_put_io_end_defer(io_submit.io_end);
2082         return ret;
2083 }
2084
2085 static int mpage_submit_page(struct mpage_da_data *mpd, struct page *page)
2086 {
2087         int len;
2088         loff_t size = i_size_read(mpd->inode);
2089         int err;
2090
2091         BUG_ON(page->index != mpd->first_page);
2092         if (page->index == size >> PAGE_SHIFT)
2093                 len = size & ~PAGE_MASK;
2094         else
2095                 len = PAGE_SIZE;
2096         clear_page_dirty_for_io(page);
2097         err = ext4_bio_write_page(&mpd->io_submit, page, len, mpd->wbc, false);
2098         if (!err)
2099                 mpd->wbc->nr_to_write--;
2100         mpd->first_page++;
2101
2102         return err;
2103 }
2104
2105 #define BH_FLAGS ((1 << BH_Unwritten) | (1 << BH_Delay))
2106
2107 /*
2108  * mballoc gives us at most this number of blocks...
2109  * XXX: That seems to be only a limitation of ext4_mb_normalize_request().
2110  * The rest of mballoc seems to handle chunks up to full group size.
2111  */
2112 #define MAX_WRITEPAGES_EXTENT_LEN 2048
2113
2114 /*
2115  * mpage_add_bh_to_extent - try to add bh to extent of blocks to map
2116  *
2117  * @mpd - extent of blocks
2118  * @lblk - logical number of the block in the file
2119  * @bh - buffer head we want to add to the extent
2120  *
2121  * The function is used to collect contig. blocks in the same state. If the
2122  * buffer doesn't require mapping for writeback and we haven't started the
2123  * extent of buffers to map yet, the function returns 'true' immediately - the
2124  * caller can write the buffer right away. Otherwise the function returns true
2125  * if the block has been added to the extent, false if the block couldn't be
2126  * added.
2127  */
2128 static bool mpage_add_bh_to_extent(struct mpage_da_data *mpd, ext4_lblk_t lblk,
2129                                    struct buffer_head *bh)
2130 {
2131         struct ext4_map_blocks *map = &mpd->map;
2132
2133         /* Buffer that doesn't need mapping for writeback? */
2134         if (!buffer_dirty(bh) || !buffer_mapped(bh) ||
2135             (!buffer_delay(bh) && !buffer_unwritten(bh))) {
2136                 /* So far no extent to map => we write the buffer right away */
2137                 if (map->m_len == 0)
2138                         return true;
2139                 return false;
2140         }
2141
2142         /* First block in the extent? */
2143         if (map->m_len == 0) {
2144                 map->m_lblk = lblk;
2145                 map->m_len = 1;
2146                 map->m_flags = bh->b_state & BH_FLAGS;
2147                 return true;
2148         }
2149
2150         /* Don't go larger than mballoc is willing to allocate */
2151         if (map->m_len >= MAX_WRITEPAGES_EXTENT_LEN)
2152                 return false;
2153
2154         /* Can we merge the block to our big extent? */
2155         if (lblk == map->m_lblk + map->m_len &&
2156             (bh->b_state & BH_FLAGS) == map->m_flags) {
2157                 map->m_len++;
2158                 return true;
2159         }
2160         return false;
2161 }
2162
2163 /*
2164  * mpage_process_page_bufs - submit page buffers for IO or add them to extent
2165  *
2166  * @mpd - extent of blocks for mapping
2167  * @head - the first buffer in the page
2168  * @bh - buffer we should start processing from
2169  * @lblk - logical number of the block in the file corresponding to @bh
2170  *
2171  * Walk through page buffers from @bh upto @head (exclusive) and either submit
2172  * the page for IO if all buffers in this page were mapped and there's no
2173  * accumulated extent of buffers to map or add buffers in the page to the
2174  * extent of buffers to map. The function returns 1 if the caller can continue
2175  * by processing the next page, 0 if it should stop adding buffers to the
2176  * extent to map because we cannot extend it anymore. It can also return value
2177  * < 0 in case of error during IO submission.
2178  */
2179 static int mpage_process_page_bufs(struct mpage_da_data *mpd,
2180                                    struct buffer_head *head,
2181                                    struct buffer_head *bh,
2182                                    ext4_lblk_t lblk)
2183 {
2184         struct inode *inode = mpd->inode;
2185         int err;
2186         ext4_lblk_t blocks = (i_size_read(inode) + (1 << inode->i_blkbits) - 1)
2187                                                         >> inode->i_blkbits;
2188
2189         do {
2190                 BUG_ON(buffer_locked(bh));
2191
2192                 if (lblk >= blocks || !mpage_add_bh_to_extent(mpd, lblk, bh)) {
2193                         /* Found extent to map? */
2194                         if (mpd->map.m_len)
2195                                 return 0;
2196                         /* Everything mapped so far and we hit EOF */
2197                         break;
2198                 }
2199         } while (lblk++, (bh = bh->b_this_page) != head);
2200         /* So far everything mapped? Submit the page for IO. */
2201         if (mpd->map.m_len == 0) {
2202                 err = mpage_submit_page(mpd, head->b_page);
2203                 if (err < 0)
2204                         return err;
2205         }
2206         return lblk < blocks;
2207 }
2208
2209 /*
2210  * mpage_map_buffers - update buffers corresponding to changed extent and
2211  *                     submit fully mapped pages for IO
2212  *
2213  * @mpd - description of extent to map, on return next extent to map
2214  *
2215  * Scan buffers corresponding to changed extent (we expect corresponding pages
2216  * to be already locked) and update buffer state according to new extent state.
2217  * We map delalloc buffers to their physical location, clear unwritten bits,
2218  * and mark buffers as uninit when we perform writes to unwritten extents
2219  * and do extent conversion after IO is finished. If the last page is not fully
2220  * mapped, we update @map to the next extent in the last page that needs
2221  * mapping. Otherwise we submit the page for IO.
2222  */
2223 static int mpage_map_and_submit_buffers(struct mpage_da_data *mpd)
2224 {
2225         struct pagevec pvec;
2226         int nr_pages, i;
2227         struct inode *inode = mpd->inode;
2228         struct buffer_head *head, *bh;
2229         int bpp_bits = PAGE_SHIFT - inode->i_blkbits;
2230         pgoff_t start, end;
2231         ext4_lblk_t lblk;
2232         sector_t pblock;
2233         int err;
2234
2235         start = mpd->map.m_lblk >> bpp_bits;
2236         end = (mpd->map.m_lblk + mpd->map.m_len - 1) >> bpp_bits;
2237         lblk = start << bpp_bits;
2238         pblock = mpd->map.m_pblk;
2239
2240         pagevec_init(&pvec, 0);
2241         while (start <= end) {
2242                 nr_pages = pagevec_lookup(&pvec, inode->i_mapping, start,
2243                                           PAGEVEC_SIZE);
2244                 if (nr_pages == 0)
2245                         break;
2246                 for (i = 0; i < nr_pages; i++) {
2247                         struct page *page = pvec.pages[i];
2248
2249                         if (page->index > end)
2250                                 break;
2251                         /* Up to 'end' pages must be contiguous */
2252                         BUG_ON(page->index != start);
2253                         bh = head = page_buffers(page);
2254                         do {
2255                                 if (lblk < mpd->map.m_lblk)
2256                                         continue;
2257                                 if (lblk >= mpd->map.m_lblk + mpd->map.m_len) {
2258                                         /*
2259                                          * Buffer after end of mapped extent.
2260                                          * Find next buffer in the page to map.
2261                                          */
2262                                         mpd->map.m_len = 0;
2263                                         mpd->map.m_flags = 0;
2264                                         /*
2265                                          * FIXME: If dioread_nolock supports
2266                                          * blocksize < pagesize, we need to make
2267                                          * sure we add size mapped so far to
2268                                          * io_end->size as the following call
2269                                          * can submit the page for IO.
2270                                          */
2271                                         err = mpage_process_page_bufs(mpd, head,
2272                                                                       bh, lblk);
2273                                         pagevec_release(&pvec);
2274                                         if (err > 0)
2275                                                 err = 0;
2276                                         return err;
2277                                 }
2278                                 if (buffer_delay(bh)) {
2279                                         clear_buffer_delay(bh);
2280                                         bh->b_blocknr = pblock++;
2281                                 }
2282                                 clear_buffer_unwritten(bh);
2283                         } while (lblk++, (bh = bh->b_this_page) != head);
2284
2285                         /*
2286                          * FIXME: This is going to break if dioread_nolock
2287                          * supports blocksize < pagesize as we will try to
2288                          * convert potentially unmapped parts of inode.
2289                          */
2290                         mpd->io_submit.io_end->size += PAGE_SIZE;
2291                         /* Page fully mapped - let IO run! */
2292                         err = mpage_submit_page(mpd, page);
2293                         if (err < 0) {
2294                                 pagevec_release(&pvec);
2295                                 return err;
2296                         }
2297                         start++;
2298                 }
2299                 pagevec_release(&pvec);
2300         }
2301         /* Extent fully mapped and matches with page boundary. We are done. */
2302         mpd->map.m_len = 0;
2303         mpd->map.m_flags = 0;
2304         return 0;
2305 }
2306
2307 static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
2308 {
2309         struct inode *inode = mpd->inode;
2310         struct ext4_map_blocks *map = &mpd->map;
2311         int get_blocks_flags;
2312         int err, dioread_nolock;
2313
2314         trace_ext4_da_write_pages_extent(inode, map);
2315         /*
2316          * Call ext4_map_blocks() to allocate any delayed allocation blocks, or
2317          * to convert an unwritten extent to be initialized (in the case
2318          * where we have written into one or more preallocated blocks).  It is
2319          * possible that we're going to need more metadata blocks than
2320          * previously reserved. However we must not fail because we're in
2321          * writeback and there is nothing we can do about it so it might result
2322          * in data loss.  So use reserved blocks to allocate metadata if
2323          * possible.
2324          *
2325          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE if
2326          * the blocks in question are delalloc blocks.  This indicates
2327          * that the blocks and quotas has already been checked when
2328          * the data was copied into the page cache.
2329          */
2330         get_blocks_flags = EXT4_GET_BLOCKS_CREATE |
2331                            EXT4_GET_BLOCKS_METADATA_NOFAIL |
2332                            EXT4_GET_BLOCKS_IO_SUBMIT;
2333         dioread_nolock = ext4_should_dioread_nolock(inode);
2334         if (dioread_nolock)
2335                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2336         if (map->m_flags & (1 << BH_Delay))
2337                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2338
2339         err = ext4_map_blocks(handle, inode, map, get_blocks_flags);
2340         if (err < 0)
2341                 return err;
2342         if (dioread_nolock && (map->m_flags & EXT4_MAP_UNWRITTEN)) {
2343                 if (!mpd->io_submit.io_end->handle &&
2344                     ext4_handle_valid(handle)) {
2345                         mpd->io_submit.io_end->handle = handle->h_rsv_handle;
2346                         handle->h_rsv_handle = NULL;
2347                 }
2348                 ext4_set_io_unwritten_flag(inode, mpd->io_submit.io_end);
2349         }
2350
2351         BUG_ON(map->m_len == 0);
2352         if (map->m_flags & EXT4_MAP_NEW) {
2353                 struct block_device *bdev = inode->i_sb->s_bdev;
2354                 int i;
2355
2356                 for (i = 0; i < map->m_len; i++)
2357                         unmap_underlying_metadata(bdev, map->m_pblk + i);
2358         }
2359         return 0;
2360 }
2361
2362 /*
2363  * mpage_map_and_submit_extent - map extent starting at mpd->lblk of length
2364  *                               mpd->len and submit pages underlying it for IO
2365  *
2366  * @handle - handle for journal operations
2367  * @mpd - extent to map
2368  * @give_up_on_write - we set this to true iff there is a fatal error and there
2369  *                     is no hope of writing the data. The caller should discard
2370  *                     dirty pages to avoid infinite loops.
2371  *
2372  * The function maps extent starting at mpd->lblk of length mpd->len. If it is
2373  * delayed, blocks are allocated, if it is unwritten, we may need to convert
2374  * them to initialized or split the described range from larger unwritten
2375  * extent. Note that we need not map all the described range since allocation
2376  * can return less blocks or the range is covered by more unwritten extents. We
2377  * cannot map more because we are limited by reserved transaction credits. On
2378  * the other hand we always make sure that the last touched page is fully
2379  * mapped so that it can be written out (and thus forward progress is
2380  * guaranteed). After mapping we submit all mapped pages for IO.
2381  */
2382 static int mpage_map_and_submit_extent(handle_t *handle,
2383                                        struct mpage_da_data *mpd,
2384                                        bool *give_up_on_write)
2385 {
2386         struct inode *inode = mpd->inode;
2387         struct ext4_map_blocks *map = &mpd->map;
2388         int err;
2389         loff_t disksize;
2390         int progress = 0;
2391
2392         mpd->io_submit.io_end->offset =
2393                                 ((loff_t)map->m_lblk) << inode->i_blkbits;
2394         do {
2395                 err = mpage_map_one_extent(handle, mpd);
2396                 if (err < 0) {
2397                         struct super_block *sb = inode->i_sb;
2398
2399                         if (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
2400                                 goto invalidate_dirty_pages;
2401                         /*
2402                          * Let the uper layers retry transient errors.
2403                          * In the case of ENOSPC, if ext4_count_free_blocks()
2404                          * is non-zero, a commit should free up blocks.
2405                          */
2406                         if ((err == -ENOMEM) ||
2407                             (err == -ENOSPC && ext4_count_free_clusters(sb))) {
2408                                 if (progress)
2409                                         goto update_disksize;
2410                                 return err;
2411                         }
2412                         ext4_msg(sb, KERN_CRIT,
2413                                  "Delayed block allocation failed for "
2414                                  "inode %lu at logical offset %llu with"
2415                                  " max blocks %u with error %d",
2416                                  inode->i_ino,
2417                                  (unsigned long long)map->m_lblk,
2418                                  (unsigned)map->m_len, -err);
2419                         ext4_msg(sb, KERN_CRIT,
2420                                  "This should not happen!! Data will "
2421                                  "be lost\n");
2422                         if (err == -ENOSPC)
2423                                 ext4_print_free_blocks(inode);
2424                 invalidate_dirty_pages:
2425                         *give_up_on_write = true;
2426                         return err;
2427                 }
2428                 progress = 1;
2429                 /*
2430                  * Update buffer state, submit mapped pages, and get us new
2431                  * extent to map
2432                  */
2433                 err = mpage_map_and_submit_buffers(mpd);
2434                 if (err < 0)
2435                         goto update_disksize;
2436         } while (map->m_len);
2437
2438 update_disksize:
2439         /*
2440          * Update on-disk size after IO is submitted.  Races with
2441          * truncate are avoided by checking i_size under i_data_sem.
2442          */
2443         disksize = ((loff_t)mpd->first_page) << PAGE_SHIFT;
2444         if (disksize > EXT4_I(inode)->i_disksize) {
2445                 int err2;
2446                 loff_t i_size;
2447
2448                 down_write(&EXT4_I(inode)->i_data_sem);
2449                 i_size = i_size_read(inode);
2450                 if (disksize > i_size)
2451                         disksize = i_size;
2452                 if (disksize > EXT4_I(inode)->i_disksize)
2453                         EXT4_I(inode)->i_disksize = disksize;
2454                 err2 = ext4_mark_inode_dirty(handle, inode);
2455                 up_write(&EXT4_I(inode)->i_data_sem);
2456                 if (err2)
2457                         ext4_error(inode->i_sb,
2458                                    "Failed to mark inode %lu dirty",
2459                                    inode->i_ino);
2460                 if (!err)
2461                         err = err2;
2462         }
2463         return err;
2464 }
2465
2466 /*
2467  * Calculate the total number of credits to reserve for one writepages
2468  * iteration. This is called from ext4_writepages(). We map an extent of
2469  * up to MAX_WRITEPAGES_EXTENT_LEN blocks and then we go on and finish mapping
2470  * the last partial page. So in total we can map MAX_WRITEPAGES_EXTENT_LEN +
2471  * bpp - 1 blocks in bpp different extents.
2472  */
2473 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2474 {
2475         int bpp = ext4_journal_blocks_per_page(inode);
2476
2477         return ext4_meta_trans_blocks(inode,
2478                                 MAX_WRITEPAGES_EXTENT_LEN + bpp - 1, bpp);
2479 }
2480
2481 /*
2482  * mpage_prepare_extent_to_map - find & lock contiguous range of dirty pages
2483  *                               and underlying extent to map
2484  *
2485  * @mpd - where to look for pages
2486  *
2487  * Walk dirty pages in the mapping. If they are fully mapped, submit them for
2488  * IO immediately. When we find a page which isn't mapped we start accumulating
2489  * extent of buffers underlying these pages that needs mapping (formed by
2490  * either delayed or unwritten buffers). We also lock the pages containing
2491  * these buffers. The extent found is returned in @mpd structure (starting at
2492  * mpd->lblk with length mpd->len blocks).
2493  *
2494  * Note that this function can attach bios to one io_end structure which are
2495  * neither logically nor physically contiguous. Although it may seem as an
2496  * unnecessary complication, it is actually inevitable in blocksize < pagesize
2497  * case as we need to track IO to all buffers underlying a page in one io_end.
2498  */
2499 static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
2500 {
2501         struct address_space *mapping = mpd->inode->i_mapping;
2502         struct pagevec pvec;
2503         unsigned int nr_pages;
2504         long left = mpd->wbc->nr_to_write;
2505         pgoff_t index = mpd->first_page;
2506         pgoff_t end = mpd->last_page;
2507         int tag;
2508         int i, err = 0;
2509         int blkbits = mpd->inode->i_blkbits;
2510         ext4_lblk_t lblk;
2511         struct buffer_head *head;
2512
2513         if (mpd->wbc->sync_mode == WB_SYNC_ALL || mpd->wbc->tagged_writepages)
2514                 tag = PAGECACHE_TAG_TOWRITE;
2515         else
2516                 tag = PAGECACHE_TAG_DIRTY;
2517
2518         pagevec_init(&pvec, 0);
2519         mpd->map.m_len = 0;
2520         mpd->next_page = index;
2521         while (index <= end) {
2522                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2523                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2524                 if (nr_pages == 0)
2525                         goto out;
2526
2527                 for (i = 0; i < nr_pages; i++) {
2528                         struct page *page = pvec.pages[i];
2529
2530                         /*
2531                          * At this point, the page may be truncated or
2532                          * invalidated (changing page->mapping to NULL), or
2533                          * even swizzled back from swapper_space to tmpfs file
2534                          * mapping. However, page->index will not change
2535                          * because we have a reference on the page.
2536                          */
2537                         if (page->index > end)
2538                                 goto out;
2539
2540                         /*
2541                          * Accumulated enough dirty pages? This doesn't apply
2542                          * to WB_SYNC_ALL mode. For integrity sync we have to
2543                          * keep going because someone may be concurrently
2544                          * dirtying pages, and we might have synced a lot of
2545                          * newly appeared dirty pages, but have not synced all
2546                          * of the old dirty pages.
2547                          */
2548                         if (mpd->wbc->sync_mode == WB_SYNC_NONE && left <= 0)
2549                                 goto out;
2550
2551                         /* If we can't merge this page, we are done. */
2552                         if (mpd->map.m_len > 0 && mpd->next_page != page->index)
2553                                 goto out;
2554
2555                         lock_page(page);
2556                         /*
2557                          * If the page is no longer dirty, or its mapping no
2558                          * longer corresponds to inode we are writing (which
2559                          * means it has been truncated or invalidated), or the
2560                          * page is already under writeback and we are not doing
2561                          * a data integrity writeback, skip the page
2562                          */
2563                         if (!PageDirty(page) ||
2564                             (PageWriteback(page) &&
2565                              (mpd->wbc->sync_mode == WB_SYNC_NONE)) ||
2566                             unlikely(page->mapping != mapping)) {
2567                                 unlock_page(page);
2568                                 continue;
2569                         }
2570
2571                         wait_on_page_writeback(page);
2572                         BUG_ON(PageWriteback(page));
2573
2574                         if (mpd->map.m_len == 0)
2575                                 mpd->first_page = page->index;
2576                         mpd->next_page = page->index + 1;
2577                         /* Add all dirty buffers to mpd */
2578                         lblk = ((ext4_lblk_t)page->index) <<
2579                                 (PAGE_SHIFT - blkbits);
2580                         head = page_buffers(page);
2581                         err = mpage_process_page_bufs(mpd, head, head, lblk);
2582                         if (err <= 0)
2583                                 goto out;
2584                         err = 0;
2585                         left--;
2586                 }
2587                 pagevec_release(&pvec);
2588                 cond_resched();
2589         }
2590         return 0;
2591 out:
2592         pagevec_release(&pvec);
2593         return err;
2594 }
2595
2596 static int __writepage(struct page *page, struct writeback_control *wbc,
2597                        void *data)
2598 {
2599         struct address_space *mapping = data;
2600         int ret = ext4_writepage(page, wbc);
2601         mapping_set_error(mapping, ret);
2602         return ret;
2603 }
2604
2605 static int ext4_writepages(struct address_space *mapping,
2606                            struct writeback_control *wbc)
2607 {
2608         pgoff_t writeback_index = 0;
2609         long nr_to_write = wbc->nr_to_write;
2610         int range_whole = 0;
2611         int cycled = 1;
2612         handle_t *handle = NULL;
2613         struct mpage_da_data mpd;
2614         struct inode *inode = mapping->host;
2615         int needed_blocks, rsv_blocks = 0, ret = 0;
2616         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2617         bool done;
2618         struct blk_plug plug;
2619         bool give_up_on_write = false;
2620
2621         percpu_down_read(&sbi->s_journal_flag_rwsem);
2622         trace_ext4_writepages(inode, wbc);
2623
2624         if (dax_mapping(mapping)) {
2625                 ret = dax_writeback_mapping_range(mapping, inode->i_sb->s_bdev,
2626                                                   wbc);
2627                 goto out_writepages;
2628         }
2629
2630         /*
2631          * No pages to write? This is mainly a kludge to avoid starting
2632          * a transaction for special inodes like journal inode on last iput()
2633          * because that could violate lock ordering on umount
2634          */
2635         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2636                 goto out_writepages;
2637
2638         if (ext4_should_journal_data(inode)) {
2639                 struct blk_plug plug;
2640
2641                 blk_start_plug(&plug);
2642                 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2643                 blk_finish_plug(&plug);
2644                 goto out_writepages;
2645         }
2646
2647         /*
2648          * If the filesystem has aborted, it is read-only, so return
2649          * right away instead of dumping stack traces later on that
2650          * will obscure the real source of the problem.  We test
2651          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2652          * the latter could be true if the filesystem is mounted
2653          * read-only, and in that case, ext4_writepages should
2654          * *never* be called, so if that ever happens, we would want
2655          * the stack trace.
2656          */
2657         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2658                 ret = -EROFS;
2659                 goto out_writepages;
2660         }
2661
2662         if (ext4_should_dioread_nolock(inode)) {
2663                 /*
2664                  * We may need to convert up to one extent per block in
2665                  * the page and we may dirty the inode.
2666                  */
2667                 rsv_blocks = 1 + (PAGE_SIZE >> inode->i_blkbits);
2668         }
2669
2670         /*
2671          * If we have inline data and arrive here, it means that
2672          * we will soon create the block for the 1st page, so
2673          * we'd better clear the inline data here.
2674          */
2675         if (ext4_has_inline_data(inode)) {
2676                 /* Just inode will be modified... */
2677                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
2678                 if (IS_ERR(handle)) {
2679                         ret = PTR_ERR(handle);
2680                         goto out_writepages;
2681                 }
2682                 BUG_ON(ext4_test_inode_state(inode,
2683                                 EXT4_STATE_MAY_INLINE_DATA));
2684                 ext4_destroy_inline_data(handle, inode);
2685                 ext4_journal_stop(handle);
2686         }
2687
2688         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2689                 range_whole = 1;
2690
2691         if (wbc->range_cyclic) {
2692                 writeback_index = mapping->writeback_index;
2693                 if (writeback_index)
2694                         cycled = 0;
2695                 mpd.first_page = writeback_index;
2696                 mpd.last_page = -1;
2697         } else {
2698                 mpd.first_page = wbc->range_start >> PAGE_SHIFT;
2699                 mpd.last_page = wbc->range_end >> PAGE_SHIFT;
2700         }
2701
2702         mpd.inode = inode;
2703         mpd.wbc = wbc;
2704         ext4_io_submit_init(&mpd.io_submit, wbc);
2705 retry:
2706         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2707                 tag_pages_for_writeback(mapping, mpd.first_page, mpd.last_page);
2708         done = false;
2709         blk_start_plug(&plug);
2710         while (!done && mpd.first_page <= mpd.last_page) {
2711                 /* For each extent of pages we use new io_end */
2712                 mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2713                 if (!mpd.io_submit.io_end) {
2714                         ret = -ENOMEM;
2715                         break;
2716                 }
2717
2718                 /*
2719                  * We have two constraints: We find one extent to map and we
2720                  * must always write out whole page (makes a difference when
2721                  * blocksize < pagesize) so that we don't block on IO when we
2722                  * try to write out the rest of the page. Journalled mode is
2723                  * not supported by delalloc.
2724                  */
2725                 BUG_ON(ext4_should_journal_data(inode));
2726                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2727
2728                 /* start a new transaction */
2729                 handle = ext4_journal_start_with_reserve(inode,
2730                                 EXT4_HT_WRITE_PAGE, needed_blocks, rsv_blocks);
2731                 if (IS_ERR(handle)) {
2732                         ret = PTR_ERR(handle);
2733                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2734                                "%ld pages, ino %lu; err %d", __func__,
2735                                 wbc->nr_to_write, inode->i_ino, ret);
2736                         /* Release allocated io_end */
2737                         ext4_put_io_end(mpd.io_submit.io_end);
2738                         break;
2739                 }
2740
2741                 trace_ext4_da_write_pages(inode, mpd.first_page, mpd.wbc);
2742                 ret = mpage_prepare_extent_to_map(&mpd);
2743                 if (!ret) {
2744                         if (mpd.map.m_len)
2745                                 ret = mpage_map_and_submit_extent(handle, &mpd,
2746                                         &give_up_on_write);
2747                         else {
2748                                 /*
2749                                  * We scanned the whole range (or exhausted
2750                                  * nr_to_write), submitted what was mapped and
2751                                  * didn't find anything needing mapping. We are
2752                                  * done.
2753                                  */
2754                                 done = true;
2755                         }
2756                 }
2757                 /*
2758                  * Caution: If the handle is synchronous,
2759                  * ext4_journal_stop() can wait for transaction commit
2760                  * to finish which may depend on writeback of pages to
2761                  * complete or on page lock to be released.  In that
2762                  * case, we have to wait until after after we have
2763                  * submitted all the IO, released page locks we hold,
2764                  * and dropped io_end reference (for extent conversion
2765                  * to be able to complete) before stopping the handle.
2766                  */
2767                 if (!ext4_handle_valid(handle) || handle->h_sync == 0) {
2768                         ext4_journal_stop(handle);
2769                         handle = NULL;
2770                 }
2771                 /* Submit prepared bio */
2772                 ext4_io_submit(&mpd.io_submit);
2773                 /* Unlock pages we didn't use */
2774                 mpage_release_unused_pages(&mpd, give_up_on_write);
2775                 /*
2776                  * Drop our io_end reference we got from init. We have
2777                  * to be careful and use deferred io_end finishing if
2778                  * we are still holding the transaction as we can
2779                  * release the last reference to io_end which may end
2780                  * up doing unwritten extent conversion.
2781                  */
2782                 if (handle) {
2783                         ext4_put_io_end_defer(mpd.io_submit.io_end);
2784                         ext4_journal_stop(handle);
2785                 } else
2786                         ext4_put_io_end(mpd.io_submit.io_end);
2787
2788                 if (ret == -ENOSPC && sbi->s_journal) {
2789                         /*
2790                          * Commit the transaction which would
2791                          * free blocks released in the transaction
2792                          * and try again
2793                          */
2794                         jbd2_journal_force_commit_nested(sbi->s_journal);
2795                         ret = 0;
2796                         continue;
2797                 }
2798                 /* Fatal error - ENOMEM, EIO... */
2799                 if (ret)
2800                         break;
2801         }
2802         blk_finish_plug(&plug);
2803         if (!ret && !cycled && wbc->nr_to_write > 0) {
2804                 cycled = 1;
2805                 mpd.last_page = writeback_index - 1;
2806                 mpd.first_page = 0;
2807                 goto retry;
2808         }
2809
2810         /* Update index */
2811         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2812                 /*
2813                  * Set the writeback_index so that range_cyclic
2814                  * mode will write it back later
2815                  */
2816                 mapping->writeback_index = mpd.first_page;
2817
2818 out_writepages:
2819         trace_ext4_writepages_result(inode, wbc, ret,
2820                                      nr_to_write - wbc->nr_to_write);
2821         percpu_up_read(&sbi->s_journal_flag_rwsem);
2822         return ret;
2823 }
2824
2825 static int ext4_nonda_switch(struct super_block *sb)
2826 {
2827         s64 free_clusters, dirty_clusters;
2828         struct ext4_sb_info *sbi = EXT4_SB(sb);
2829
2830         /*
2831          * switch to non delalloc mode if we are running low
2832          * on free block. The free block accounting via percpu
2833          * counters can get slightly wrong with percpu_counter_batch getting
2834          * accumulated on each CPU without updating global counters
2835          * Delalloc need an accurate free block accounting. So switch
2836          * to non delalloc when we are near to error range.
2837          */
2838         free_clusters =
2839                 percpu_counter_read_positive(&sbi->s_freeclusters_counter);
2840         dirty_clusters =
2841                 percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
2842         /*
2843          * Start pushing delalloc when 1/2 of free blocks are dirty.
2844          */
2845         if (dirty_clusters && (free_clusters < 2 * dirty_clusters))
2846                 try_to_writeback_inodes_sb(sb, WB_REASON_FS_FREE_SPACE);
2847
2848         if (2 * free_clusters < 3 * dirty_clusters ||
2849             free_clusters < (dirty_clusters + EXT4_FREECLUSTERS_WATERMARK)) {
2850                 /*
2851                  * free block count is less than 150% of dirty blocks
2852                  * or free blocks is less than watermark
2853                  */
2854                 return 1;
2855         }
2856         return 0;
2857 }
2858
2859 /* We always reserve for an inode update; the superblock could be there too */
2860 static int ext4_da_write_credits(struct inode *inode, loff_t pos, unsigned len)
2861 {
2862         if (likely(ext4_has_feature_large_file(inode->i_sb)))
2863                 return 1;
2864
2865         if (pos + len <= 0x7fffffffULL)
2866                 return 1;
2867
2868         /* We might need to update the superblock to set LARGE_FILE */
2869         return 2;
2870 }
2871
2872 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2873                                loff_t pos, unsigned len, unsigned flags,
2874                                struct page **pagep, void **fsdata)
2875 {
2876         int ret, retries = 0;
2877         struct page *page;
2878         pgoff_t index;
2879         struct inode *inode = mapping->host;
2880         handle_t *handle;
2881
2882         index = pos >> PAGE_SHIFT;
2883
2884         if (ext4_nonda_switch(inode->i_sb)) {
2885                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2886                 return ext4_write_begin(file, mapping, pos,
2887                                         len, flags, pagep, fsdata);
2888         }
2889         *fsdata = (void *)0;
2890         trace_ext4_da_write_begin(inode, pos, len, flags);
2891
2892         if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
2893                 ret = ext4_da_write_inline_data_begin(mapping, inode,
2894                                                       pos, len, flags,
2895                                                       pagep, fsdata);
2896                 if (ret < 0)
2897                         return ret;
2898                 if (ret == 1)
2899                         return 0;
2900         }
2901
2902         /*
2903          * grab_cache_page_write_begin() can take a long time if the
2904          * system is thrashing due to memory pressure, or if the page
2905          * is being written back.  So grab it first before we start
2906          * the transaction handle.  This also allows us to allocate
2907          * the page (if needed) without using GFP_NOFS.
2908          */
2909 retry_grab:
2910         page = grab_cache_page_write_begin(mapping, index, flags);
2911         if (!page)
2912                 return -ENOMEM;
2913         unlock_page(page);
2914
2915         /*
2916          * With delayed allocation, we don't log the i_disksize update
2917          * if there is delayed block allocation. But we still need
2918          * to journalling the i_disksize update if writes to the end
2919          * of file which has an already mapped buffer.
2920          */
2921 retry_journal:
2922         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
2923                                 ext4_da_write_credits(inode, pos, len));
2924         if (IS_ERR(handle)) {
2925                 put_page(page);
2926                 return PTR_ERR(handle);
2927         }
2928
2929         lock_page(page);
2930         if (page->mapping != mapping) {
2931                 /* The page got truncated from under us */
2932                 unlock_page(page);
2933                 put_page(page);
2934                 ext4_journal_stop(handle);
2935                 goto retry_grab;
2936         }
2937         /* In case writeback began while the page was unlocked */
2938         wait_for_stable_page(page);
2939
2940 #ifdef CONFIG_EXT4_FS_ENCRYPTION
2941         ret = ext4_block_write_begin(page, pos, len,
2942                                      ext4_da_get_block_prep);
2943 #else
2944         ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
2945 #endif
2946         if (ret < 0) {
2947                 unlock_page(page);
2948                 ext4_journal_stop(handle);
2949                 /*
2950                  * block_write_begin may have instantiated a few blocks
2951                  * outside i_size.  Trim these off again. Don't need
2952                  * i_size_read because we hold i_mutex.
2953                  */
2954                 if (pos + len > inode->i_size)
2955                         ext4_truncate_failed_write(inode);
2956
2957                 if (ret == -ENOSPC &&
2958                     ext4_should_retry_alloc(inode->i_sb, &retries))
2959                         goto retry_journal;
2960
2961                 put_page(page);
2962                 return ret;
2963         }
2964
2965         *pagep = page;
2966         return ret;
2967 }
2968
2969 /*
2970  * Check if we should update i_disksize
2971  * when write to the end of file but not require block allocation
2972  */
2973 static int ext4_da_should_update_i_disksize(struct page *page,
2974                                             unsigned long offset)
2975 {
2976         struct buffer_head *bh;
2977         struct inode *inode = page->mapping->host;
2978         unsigned int idx;
2979         int i;
2980
2981         bh = page_buffers(page);
2982         idx = offset >> inode->i_blkbits;
2983
2984         for (i = 0; i < idx; i++)
2985                 bh = bh->b_this_page;
2986
2987         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
2988                 return 0;
2989         return 1;
2990 }
2991
2992 static int ext4_da_write_end(struct file *file,
2993                              struct address_space *mapping,
2994                              loff_t pos, unsigned len, unsigned copied,
2995                              struct page *page, void *fsdata)
2996 {
2997         struct inode *inode = mapping->host;
2998         int ret = 0, ret2;
2999         handle_t *handle = ext4_journal_current_handle();
3000         loff_t new_i_size;
3001         unsigned long start, end;
3002         int write_mode = (int)(unsigned long)fsdata;
3003
3004         if (write_mode == FALL_BACK_TO_NONDELALLOC)
3005                 return ext4_write_end(file, mapping, pos,
3006                                       len, copied, page, fsdata);
3007
3008         trace_ext4_da_write_end(inode, pos, len, copied);
3009         start = pos & (PAGE_SIZE - 1);
3010         end = start + copied - 1;
3011
3012         /*
3013          * generic_write_end() will run mark_inode_dirty() if i_size
3014          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3015          * into that.
3016          */
3017         new_i_size = pos + copied;
3018         if (copied && new_i_size > EXT4_I(inode)->i_disksize) {
3019                 if (ext4_has_inline_data(inode) ||
3020                     ext4_da_should_update_i_disksize(page, end)) {
3021                         ext4_update_i_disksize(inode, new_i_size);
3022                         /* We need to mark inode dirty even if
3023                          * new_i_size is less that inode->i_size
3024                          * bu greater than i_disksize.(hint delalloc)
3025                          */
3026                         ext4_mark_inode_dirty(handle, inode);
3027                 }
3028         }
3029
3030         if (write_mode != CONVERT_INLINE_DATA &&
3031             ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA) &&
3032             ext4_has_inline_data(inode))
3033                 ret2 = ext4_da_write_inline_data_end(inode, pos, len, copied,
3034                                                      page);
3035         else
3036                 ret2 = generic_write_end(file, mapping, pos, len, copied,
3037                                                         page, fsdata);
3038
3039         copied = ret2;
3040         if (ret2 < 0)
3041                 ret = ret2;
3042         ret2 = ext4_journal_stop(handle);
3043         if (!ret)
3044                 ret = ret2;
3045
3046         return ret ? ret : copied;
3047 }
3048
3049 static void ext4_da_invalidatepage(struct page *page, unsigned int offset,
3050                                    unsigned int length)
3051 {
3052         /*
3053          * Drop reserved blocks
3054          */
3055         BUG_ON(!PageLocked(page));
3056         if (!page_has_buffers(page))
3057                 goto out;
3058
3059         ext4_da_page_release_reservation(page, offset, length);
3060
3061 out:
3062         ext4_invalidatepage(page, offset, length);
3063
3064         return;
3065 }
3066
3067 /*
3068  * Force all delayed allocation blocks to be allocated for a given inode.
3069  */
3070 int ext4_alloc_da_blocks(struct inode *inode)
3071 {
3072         trace_ext4_alloc_da_blocks(inode);
3073
3074         if (!EXT4_I(inode)->i_reserved_data_blocks)
3075                 return 0;
3076
3077         /*
3078          * We do something simple for now.  The filemap_flush() will
3079          * also start triggering a write of the data blocks, which is
3080          * not strictly speaking necessary (and for users of
3081          * laptop_mode, not even desirable).  However, to do otherwise
3082          * would require replicating code paths in:
3083          *
3084          * ext4_writepages() ->
3085          *    write_cache_pages() ---> (via passed in callback function)
3086          *        __mpage_da_writepage() -->
3087          *           mpage_add_bh_to_extent()
3088          *           mpage_da_map_blocks()
3089          *
3090          * The problem is that write_cache_pages(), located in
3091          * mm/page-writeback.c, marks pages clean in preparation for
3092          * doing I/O, which is not desirable if we're not planning on
3093          * doing I/O at all.
3094          *
3095          * We could call write_cache_pages(), and then redirty all of
3096          * the pages by calling redirty_page_for_writepage() but that
3097          * would be ugly in the extreme.  So instead we would need to
3098          * replicate parts of the code in the above functions,
3099          * simplifying them because we wouldn't actually intend to
3100          * write out the pages, but rather only collect contiguous
3101          * logical block extents, call the multi-block allocator, and
3102          * then update the buffer heads with the block allocations.
3103          *
3104          * For now, though, we'll cheat by calling filemap_flush(),
3105          * which will map the blocks, and start the I/O, but not
3106          * actually wait for the I/O to complete.
3107          */
3108         return filemap_flush(inode->i_mapping);
3109 }
3110
3111 /*
3112  * bmap() is special.  It gets used by applications such as lilo and by
3113  * the swapper to find the on-disk block of a specific piece of data.
3114  *
3115  * Naturally, this is dangerous if the block concerned is still in the
3116  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3117  * filesystem and enables swap, then they may get a nasty shock when the
3118  * data getting swapped to that swapfile suddenly gets overwritten by
3119  * the original zero's written out previously to the journal and
3120  * awaiting writeback in the kernel's buffer cache.
3121  *
3122  * So, if we see any bmap calls here on a modified, data-journaled file,
3123  * take extra steps to flush any blocks which might be in the cache.
3124  */
3125 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3126 {
3127         struct inode *inode = mapping->host;
3128         journal_t *journal;
3129         int err;
3130
3131         /*