ext4: do not ask jbd2 to write data for delalloc buffers
[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         __u16 csum_lo;
55         __u16 csum_hi = 0;
56         __u32 csum;
57
58         csum_lo = le16_to_cpu(raw->i_checksum_lo);
59         raw->i_checksum_lo = 0;
60         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
61             EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi)) {
62                 csum_hi = le16_to_cpu(raw->i_checksum_hi);
63                 raw->i_checksum_hi = 0;
64         }
65
66         csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)raw,
67                            EXT4_INODE_SIZE(inode->i_sb));
68
69         raw->i_checksum_lo = cpu_to_le16(csum_lo);
70         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
71             EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
72                 raw->i_checksum_hi = cpu_to_le16(csum_hi);
73
74         return csum;
75 }
76
77 static int ext4_inode_csum_verify(struct inode *inode, struct ext4_inode *raw,
78                                   struct ext4_inode_info *ei)
79 {
80         __u32 provided, calculated;
81
82         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
83             cpu_to_le32(EXT4_OS_LINUX) ||
84             !ext4_has_metadata_csum(inode->i_sb))
85                 return 1;
86
87         provided = le16_to_cpu(raw->i_checksum_lo);
88         calculated = ext4_inode_csum(inode, raw, ei);
89         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
90             EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
91                 provided |= ((__u32)le16_to_cpu(raw->i_checksum_hi)) << 16;
92         else
93                 calculated &= 0xFFFF;
94
95         return provided == calculated;
96 }
97
98 static void ext4_inode_csum_set(struct inode *inode, struct ext4_inode *raw,
99                                 struct ext4_inode_info *ei)
100 {
101         __u32 csum;
102
103         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
104             cpu_to_le32(EXT4_OS_LINUX) ||
105             !ext4_has_metadata_csum(inode->i_sb))
106                 return;
107
108         csum = ext4_inode_csum(inode, raw, ei);
109         raw->i_checksum_lo = cpu_to_le16(csum & 0xFFFF);
110         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
111             EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
112                 raw->i_checksum_hi = cpu_to_le16(csum >> 16);
113 }
114
115 static inline int ext4_begin_ordered_truncate(struct inode *inode,
116                                               loff_t new_size)
117 {
118         trace_ext4_begin_ordered_truncate(inode, new_size);
119         /*
120          * If jinode is zero, then we never opened the file for
121          * writing, so there's no need to call
122          * jbd2_journal_begin_ordered_truncate() since there's no
123          * outstanding writes we need to flush.
124          */
125         if (!EXT4_I(inode)->jinode)
126                 return 0;
127         return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
128                                                    EXT4_I(inode)->jinode,
129                                                    new_size);
130 }
131
132 static void ext4_invalidatepage(struct page *page, unsigned int offset,
133                                 unsigned int length);
134 static int __ext4_journalled_writepage(struct page *page, unsigned int len);
135 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
136 static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
137                                   int pextents);
138
139 /*
140  * Test whether an inode is a fast symlink.
141  */
142 int ext4_inode_is_fast_symlink(struct inode *inode)
143 {
144         int ea_blocks = EXT4_I(inode)->i_file_acl ?
145                 EXT4_CLUSTER_SIZE(inode->i_sb) >> 9 : 0;
146
147         if (ext4_has_inline_data(inode))
148                 return 0;
149
150         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
151 }
152
153 /*
154  * Restart the transaction associated with *handle.  This does a commit,
155  * so before we call here everything must be consistently dirtied against
156  * this transaction.
157  */
158 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
159                                  int nblocks)
160 {
161         int ret;
162
163         /*
164          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
165          * moment, get_block can be called only for blocks inside i_size since
166          * page cache has been already dropped and writes are blocked by
167          * i_mutex. So we can safely drop the i_data_sem here.
168          */
169         BUG_ON(EXT4_JOURNAL(inode) == NULL);
170         jbd_debug(2, "restarting handle %p\n", handle);
171         up_write(&EXT4_I(inode)->i_data_sem);
172         ret = ext4_journal_restart(handle, nblocks);
173         down_write(&EXT4_I(inode)->i_data_sem);
174         ext4_discard_preallocations(inode);
175
176         return ret;
177 }
178
179 /*
180  * Called at the last iput() if i_nlink is zero.
181  */
182 void ext4_evict_inode(struct inode *inode)
183 {
184         handle_t *handle;
185         int err;
186
187         trace_ext4_evict_inode(inode);
188
189         if (inode->i_nlink) {
190                 /*
191                  * When journalling data dirty buffers are tracked only in the
192                  * journal. So although mm thinks everything is clean and
193                  * ready for reaping the inode might still have some pages to
194                  * write in the running transaction or waiting to be
195                  * checkpointed. Thus calling jbd2_journal_invalidatepage()
196                  * (via truncate_inode_pages()) to discard these buffers can
197                  * cause data loss. Also even if we did not discard these
198                  * buffers, we would have no way to find them after the inode
199                  * is reaped and thus user could see stale data if he tries to
200                  * read them before the transaction is checkpointed. So be
201                  * careful and force everything to disk here... We use
202                  * ei->i_datasync_tid to store the newest transaction
203                  * containing inode's data.
204                  *
205                  * Note that directories do not have this problem because they
206                  * don't use page cache.
207                  */
208                 if (ext4_should_journal_data(inode) &&
209                     (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode)) &&
210                     inode->i_ino != EXT4_JOURNAL_INO) {
211                         journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
212                         tid_t commit_tid = EXT4_I(inode)->i_datasync_tid;
213
214                         jbd2_complete_transaction(journal, commit_tid);
215                         filemap_write_and_wait(&inode->i_data);
216                 }
217                 truncate_inode_pages_final(&inode->i_data);
218
219                 goto no_delete;
220         }
221
222         if (is_bad_inode(inode))
223                 goto no_delete;
224         dquot_initialize(inode);
225
226         if (ext4_should_order_data(inode))
227                 ext4_begin_ordered_truncate(inode, 0);
228         truncate_inode_pages_final(&inode->i_data);
229
230         /*
231          * Protect us against freezing - iput() caller didn't have to have any
232          * protection against it
233          */
234         sb_start_intwrite(inode->i_sb);
235         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE,
236                                     ext4_blocks_for_truncate(inode)+3);
237         if (IS_ERR(handle)) {
238                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
239                 /*
240                  * If we're going to skip the normal cleanup, we still need to
241                  * make sure that the in-core orphan linked list is properly
242                  * cleaned up.
243                  */
244                 ext4_orphan_del(NULL, inode);
245                 sb_end_intwrite(inode->i_sb);
246                 goto no_delete;
247         }
248
249         if (IS_SYNC(inode))
250                 ext4_handle_sync(handle);
251         inode->i_size = 0;
252         err = ext4_mark_inode_dirty(handle, inode);
253         if (err) {
254                 ext4_warning(inode->i_sb,
255                              "couldn't mark inode dirty (err %d)", err);
256                 goto stop_handle;
257         }
258         if (inode->i_blocks)
259                 ext4_truncate(inode);
260
261         /*
262          * ext4_ext_truncate() doesn't reserve any slop when it
263          * restarts journal transactions; therefore there may not be
264          * enough credits left in the handle to remove the inode from
265          * the orphan list and set the dtime field.
266          */
267         if (!ext4_handle_has_enough_credits(handle, 3)) {
268                 err = ext4_journal_extend(handle, 3);
269                 if (err > 0)
270                         err = ext4_journal_restart(handle, 3);
271                 if (err != 0) {
272                         ext4_warning(inode->i_sb,
273                                      "couldn't extend journal (err %d)", err);
274                 stop_handle:
275                         ext4_journal_stop(handle);
276                         ext4_orphan_del(NULL, inode);
277                         sb_end_intwrite(inode->i_sb);
278                         goto no_delete;
279                 }
280         }
281
282         /*
283          * Kill off the orphan record which ext4_truncate created.
284          * AKPM: I think this can be inside the above `if'.
285          * Note that ext4_orphan_del() has to be able to cope with the
286          * deletion of a non-existent orphan - this is because we don't
287          * know if ext4_truncate() actually created an orphan record.
288          * (Well, we could do this if we need to, but heck - it works)
289          */
290         ext4_orphan_del(handle, inode);
291         EXT4_I(inode)->i_dtime  = get_seconds();
292
293         /*
294          * One subtle ordering requirement: if anything has gone wrong
295          * (transaction abort, IO errors, whatever), then we can still
296          * do these next steps (the fs will already have been marked as
297          * having errors), but we can't free the inode if the mark_dirty
298          * fails.
299          */
300         if (ext4_mark_inode_dirty(handle, inode))
301                 /* If that failed, just do the required in-core inode clear. */
302                 ext4_clear_inode(inode);
303         else
304                 ext4_free_inode(handle, inode);
305         ext4_journal_stop(handle);
306         sb_end_intwrite(inode->i_sb);
307         return;
308 no_delete:
309         ext4_clear_inode(inode);        /* We must guarantee clearing of inode... */
310 }
311
312 #ifdef CONFIG_QUOTA
313 qsize_t *ext4_get_reserved_space(struct inode *inode)
314 {
315         return &EXT4_I(inode)->i_reserved_quota;
316 }
317 #endif
318
319 /*
320  * Called with i_data_sem down, which is important since we can call
321  * ext4_discard_preallocations() from here.
322  */
323 void ext4_da_update_reserve_space(struct inode *inode,
324                                         int used, int quota_claim)
325 {
326         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
327         struct ext4_inode_info *ei = EXT4_I(inode);
328
329         spin_lock(&ei->i_block_reservation_lock);
330         trace_ext4_da_update_reserve_space(inode, used, quota_claim);
331         if (unlikely(used > ei->i_reserved_data_blocks)) {
332                 ext4_warning(inode->i_sb, "%s: ino %lu, used %d "
333                          "with only %d reserved data blocks",
334                          __func__, inode->i_ino, used,
335                          ei->i_reserved_data_blocks);
336                 WARN_ON(1);
337                 used = ei->i_reserved_data_blocks;
338         }
339
340         /* Update per-inode reservations */
341         ei->i_reserved_data_blocks -= used;
342         percpu_counter_sub(&sbi->s_dirtyclusters_counter, used);
343
344         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
345
346         /* Update quota subsystem for data blocks */
347         if (quota_claim)
348                 dquot_claim_block(inode, EXT4_C2B(sbi, used));
349         else {
350                 /*
351                  * We did fallocate with an offset that is already delayed
352                  * allocated. So on delayed allocated writeback we should
353                  * not re-claim the quota for fallocated blocks.
354                  */
355                 dquot_release_reservation_block(inode, EXT4_C2B(sbi, used));
356         }
357
358         /*
359          * If we have done all the pending block allocations and if
360          * there aren't any writers on the inode, we can discard the
361          * inode's preallocations.
362          */
363         if ((ei->i_reserved_data_blocks == 0) &&
364             (atomic_read(&inode->i_writecount) == 0))
365                 ext4_discard_preallocations(inode);
366 }
367
368 static int __check_block_validity(struct inode *inode, const char *func,
369                                 unsigned int line,
370                                 struct ext4_map_blocks *map)
371 {
372         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
373                                    map->m_len)) {
374                 ext4_error_inode(inode, func, line, map->m_pblk,
375                                  "lblock %lu mapped to illegal pblock "
376                                  "(length %d)", (unsigned long) map->m_lblk,
377                                  map->m_len);
378                 return -EFSCORRUPTED;
379         }
380         return 0;
381 }
382
383 int ext4_issue_zeroout(struct inode *inode, ext4_lblk_t lblk, ext4_fsblk_t pblk,
384                        ext4_lblk_t len)
385 {
386         int ret;
387
388         if (ext4_encrypted_inode(inode))
389                 return ext4_encrypted_zeroout(inode, lblk, pblk, len);
390
391         ret = sb_issue_zeroout(inode->i_sb, pblk, len, GFP_NOFS);
392         if (ret > 0)
393                 ret = 0;
394
395         return ret;
396 }
397
398 #define check_block_validity(inode, map)        \
399         __check_block_validity((inode), __func__, __LINE__, (map))
400
401 #ifdef ES_AGGRESSIVE_TEST
402 static void ext4_map_blocks_es_recheck(handle_t *handle,
403                                        struct inode *inode,
404                                        struct ext4_map_blocks *es_map,
405                                        struct ext4_map_blocks *map,
406                                        int flags)
407 {
408         int retval;
409
410         map->m_flags = 0;
411         /*
412          * There is a race window that the result is not the same.
413          * e.g. xfstests #223 when dioread_nolock enables.  The reason
414          * is that we lookup a block mapping in extent status tree with
415          * out taking i_data_sem.  So at the time the unwritten extent
416          * could be converted.
417          */
418         down_read(&EXT4_I(inode)->i_data_sem);
419         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
420                 retval = ext4_ext_map_blocks(handle, inode, map, flags &
421                                              EXT4_GET_BLOCKS_KEEP_SIZE);
422         } else {
423                 retval = ext4_ind_map_blocks(handle, inode, map, flags &
424                                              EXT4_GET_BLOCKS_KEEP_SIZE);
425         }
426         up_read((&EXT4_I(inode)->i_data_sem));
427
428         /*
429          * We don't check m_len because extent will be collpased in status
430          * tree.  So the m_len might not equal.
431          */
432         if (es_map->m_lblk != map->m_lblk ||
433             es_map->m_flags != map->m_flags ||
434             es_map->m_pblk != map->m_pblk) {
435                 printk("ES cache assertion failed for inode: %lu "
436                        "es_cached ex [%d/%d/%llu/%x] != "
437                        "found ex [%d/%d/%llu/%x] retval %d flags %x\n",
438                        inode->i_ino, es_map->m_lblk, es_map->m_len,
439                        es_map->m_pblk, es_map->m_flags, map->m_lblk,
440                        map->m_len, map->m_pblk, map->m_flags,
441                        retval, flags);
442         }
443 }
444 #endif /* ES_AGGRESSIVE_TEST */
445
446 /*
447  * The ext4_map_blocks() function tries to look up the requested blocks,
448  * and returns if the blocks are already mapped.
449  *
450  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
451  * and store the allocated blocks in the result buffer head and mark it
452  * mapped.
453  *
454  * If file type is extents based, it will call ext4_ext_map_blocks(),
455  * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
456  * based files
457  *
458  * On success, it returns the number of blocks being mapped or allocated.  if
459  * create==0 and the blocks are pre-allocated and unwritten, the resulting @map
460  * is marked as unwritten. If the create == 1, it will mark @map as mapped.
461  *
462  * It returns 0 if plain look up failed (blocks have not been allocated), in
463  * that case, @map is returned as unmapped but we still do fill map->m_len to
464  * indicate the length of a hole starting at map->m_lblk.
465  *
466  * It returns the error in case of allocation failure.
467  */
468 int ext4_map_blocks(handle_t *handle, struct inode *inode,
469                     struct ext4_map_blocks *map, int flags)
470 {
471         struct extent_status es;
472         int retval;
473         int ret = 0;
474 #ifdef ES_AGGRESSIVE_TEST
475         struct ext4_map_blocks orig_map;
476
477         memcpy(&orig_map, map, sizeof(*map));
478 #endif
479
480         map->m_flags = 0;
481         ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
482                   "logical block %lu\n", inode->i_ino, flags, map->m_len,
483                   (unsigned long) map->m_lblk);
484
485         /*
486          * ext4_map_blocks returns an int, and m_len is an unsigned int
487          */
488         if (unlikely(map->m_len > INT_MAX))
489                 map->m_len = INT_MAX;
490
491         /* We can handle the block number less than EXT_MAX_BLOCKS */
492         if (unlikely(map->m_lblk >= EXT_MAX_BLOCKS))
493                 return -EFSCORRUPTED;
494
495         /* Lookup extent status tree firstly */
496         if (ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
497                 if (ext4_es_is_written(&es) || ext4_es_is_unwritten(&es)) {
498                         map->m_pblk = ext4_es_pblock(&es) +
499                                         map->m_lblk - es.es_lblk;
500                         map->m_flags |= ext4_es_is_written(&es) ?
501                                         EXT4_MAP_MAPPED : EXT4_MAP_UNWRITTEN;
502                         retval = es.es_len - (map->m_lblk - es.es_lblk);
503                         if (retval > map->m_len)
504                                 retval = map->m_len;
505                         map->m_len = retval;
506                 } else if (ext4_es_is_delayed(&es) || ext4_es_is_hole(&es)) {
507                         map->m_pblk = 0;
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                         retval = 0;
513                 } else {
514                         BUG_ON(1);
515                 }
516 #ifdef ES_AGGRESSIVE_TEST
517                 ext4_map_blocks_es_recheck(handle, inode, map,
518                                            &orig_map, flags);
519 #endif
520                 goto found;
521         }
522
523         /*
524          * Try to see if we can get the block without requesting a new
525          * file system block.
526          */
527         down_read(&EXT4_I(inode)->i_data_sem);
528         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
529                 retval = ext4_ext_map_blocks(handle, inode, map, flags &
530                                              EXT4_GET_BLOCKS_KEEP_SIZE);
531         } else {
532                 retval = ext4_ind_map_blocks(handle, inode, map, flags &
533                                              EXT4_GET_BLOCKS_KEEP_SIZE);
534         }
535         if (retval > 0) {
536                 unsigned int status;
537
538                 if (unlikely(retval != map->m_len)) {
539                         ext4_warning(inode->i_sb,
540                                      "ES len assertion failed for inode "
541                                      "%lu: retval %d != map->m_len %d",
542                                      inode->i_ino, retval, map->m_len);
543                         WARN_ON(1);
544                 }
545
546                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
547                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
548                 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
549                     !(status & EXTENT_STATUS_WRITTEN) &&
550                     ext4_find_delalloc_range(inode, map->m_lblk,
551                                              map->m_lblk + map->m_len - 1))
552                         status |= EXTENT_STATUS_DELAYED;
553                 ret = ext4_es_insert_extent(inode, map->m_lblk,
554                                             map->m_len, map->m_pblk, status);
555                 if (ret < 0)
556                         retval = ret;
557         }
558         up_read((&EXT4_I(inode)->i_data_sem));
559
560 found:
561         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
562                 ret = check_block_validity(inode, map);
563                 if (ret != 0)
564                         return ret;
565         }
566
567         /* If it is only a block(s) look up */
568         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
569                 return retval;
570
571         /*
572          * Returns if the blocks have already allocated
573          *
574          * Note that if blocks have been preallocated
575          * ext4_ext_get_block() returns the create = 0
576          * with buffer head unmapped.
577          */
578         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
579                 /*
580                  * If we need to convert extent to unwritten
581                  * we continue and do the actual work in
582                  * ext4_ext_map_blocks()
583                  */
584                 if (!(flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN))
585                         return retval;
586
587         /*
588          * Here we clear m_flags because after allocating an new extent,
589          * it will be set again.
590          */
591         map->m_flags &= ~EXT4_MAP_FLAGS;
592
593         /*
594          * New blocks allocate and/or writing to unwritten extent
595          * will possibly result in updating i_data, so we take
596          * the write lock of i_data_sem, and call get_block()
597          * with create == 1 flag.
598          */
599         down_write(&EXT4_I(inode)->i_data_sem);
600
601         /*
602          * We need to check for EXT4 here because migrate
603          * could have changed the inode type in between
604          */
605         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
606                 retval = ext4_ext_map_blocks(handle, inode, map, flags);
607         } else {
608                 retval = ext4_ind_map_blocks(handle, inode, map, flags);
609
610                 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
611                         /*
612                          * We allocated new blocks which will result in
613                          * i_data's format changing.  Force the migrate
614                          * to fail by clearing migrate flags
615                          */
616                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
617                 }
618
619                 /*
620                  * Update reserved blocks/metadata blocks after successful
621                  * block allocation which had been deferred till now. We don't
622                  * support fallocate for non extent files. So we can update
623                  * reserve space here.
624                  */
625                 if ((retval > 0) &&
626                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
627                         ext4_da_update_reserve_space(inode, retval, 1);
628         }
629
630         if (retval > 0) {
631                 unsigned int status;
632
633                 if (unlikely(retval != map->m_len)) {
634                         ext4_warning(inode->i_sb,
635                                      "ES len assertion failed for inode "
636                                      "%lu: retval %d != map->m_len %d",
637                                      inode->i_ino, retval, map->m_len);
638                         WARN_ON(1);
639                 }
640
641                 /*
642                  * We have to zeroout blocks before inserting them into extent
643                  * status tree. Otherwise someone could look them up there and
644                  * use them before they are really zeroed.
645                  */
646                 if (flags & EXT4_GET_BLOCKS_ZERO &&
647                     map->m_flags & EXT4_MAP_MAPPED &&
648                     map->m_flags & EXT4_MAP_NEW) {
649                         ret = ext4_issue_zeroout(inode, map->m_lblk,
650                                                  map->m_pblk, map->m_len);
651                         if (ret) {
652                                 retval = ret;
653                                 goto out_sem;
654                         }
655                 }
656
657                 /*
658                  * If the extent has been zeroed out, we don't need to update
659                  * extent status tree.
660                  */
661                 if ((flags & EXT4_GET_BLOCKS_PRE_IO) &&
662                     ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
663                         if (ext4_es_is_written(&es))
664                                 goto out_sem;
665                 }
666                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
667                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
668                 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
669                     !(status & EXTENT_STATUS_WRITTEN) &&
670                     ext4_find_delalloc_range(inode, map->m_lblk,
671                                              map->m_lblk + map->m_len - 1))
672                         status |= EXTENT_STATUS_DELAYED;
673                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
674                                             map->m_pblk, status);
675                 if (ret < 0) {
676                         retval = ret;
677                         goto out_sem;
678                 }
679         }
680
681 out_sem:
682         up_write((&EXT4_I(inode)->i_data_sem));
683         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
684                 ret = check_block_validity(inode, map);
685                 if (ret != 0)
686                         return ret;
687
688                 /*
689                  * Inodes with freshly allocated blocks where contents will be
690                  * visible after transaction commit must be on transaction's
691                  * ordered data list.
692                  */
693                 if (map->m_flags & EXT4_MAP_NEW &&
694                     !(map->m_flags & EXT4_MAP_UNWRITTEN) &&
695                     !(flags & EXT4_GET_BLOCKS_ZERO) &&
696                     !IS_NOQUOTA(inode) &&
697                     ext4_should_order_data(inode)) {
698                         if (flags & EXT4_GET_BLOCKS_IO_SUBMIT)
699                                 ret = ext4_jbd2_inode_add_wait(handle, inode);
700                         else
701                                 ret = ext4_jbd2_inode_add_write(handle, inode);
702                         if (ret)
703                                 return ret;
704                 }
705         }
706         return retval;
707 }
708
709 /*
710  * Update EXT4_MAP_FLAGS in bh->b_state. For buffer heads attached to pages
711  * we have to be careful as someone else may be manipulating b_state as well.
712  */
713 static void ext4_update_bh_state(struct buffer_head *bh, unsigned long flags)
714 {
715         unsigned long old_state;
716         unsigned long new_state;
717
718         flags &= EXT4_MAP_FLAGS;
719
720         /* Dummy buffer_head? Set non-atomically. */
721         if (!bh->b_page) {
722                 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | flags;
723                 return;
724         }
725         /*
726          * Someone else may be modifying b_state. Be careful! This is ugly but
727          * once we get rid of using bh as a container for mapping information
728          * to pass to / from get_block functions, this can go away.
729          */
730         do {
731                 old_state = READ_ONCE(bh->b_state);
732                 new_state = (old_state & ~EXT4_MAP_FLAGS) | flags;
733         } while (unlikely(
734                  cmpxchg(&bh->b_state, old_state, new_state) != old_state));
735 }
736
737 static int _ext4_get_block(struct inode *inode, sector_t iblock,
738                            struct buffer_head *bh, int flags)
739 {
740         struct ext4_map_blocks map;
741         int ret = 0;
742
743         if (ext4_has_inline_data(inode))
744                 return -ERANGE;
745
746         map.m_lblk = iblock;
747         map.m_len = bh->b_size >> inode->i_blkbits;
748
749         ret = ext4_map_blocks(ext4_journal_current_handle(), inode, &map,
750                               flags);
751         if (ret > 0) {
752                 map_bh(bh, inode->i_sb, map.m_pblk);
753                 ext4_update_bh_state(bh, map.m_flags);
754                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
755                 ret = 0;
756         }
757         return ret;
758 }
759
760 int ext4_get_block(struct inode *inode, sector_t iblock,
761                    struct buffer_head *bh, int create)
762 {
763         return _ext4_get_block(inode, iblock, bh,
764                                create ? EXT4_GET_BLOCKS_CREATE : 0);
765 }
766
767 /*
768  * Get block function used when preparing for buffered write if we require
769  * creating an unwritten extent if blocks haven't been allocated.  The extent
770  * will be converted to written after the IO is complete.
771  */
772 int ext4_get_block_unwritten(struct inode *inode, sector_t iblock,
773                              struct buffer_head *bh_result, int create)
774 {
775         ext4_debug("ext4_get_block_unwritten: inode %lu, create flag %d\n",
776                    inode->i_ino, create);
777         return _ext4_get_block(inode, iblock, bh_result,
778                                EXT4_GET_BLOCKS_IO_CREATE_EXT);
779 }
780
781 /* Maximum number of blocks we map for direct IO at once. */
782 #define DIO_MAX_BLOCKS 4096
783
784 /*
785  * Get blocks function for the cases that need to start a transaction -
786  * generally difference cases of direct IO and DAX IO. It also handles retries
787  * in case of ENOSPC.
788  */
789 static int ext4_get_block_trans(struct inode *inode, sector_t iblock,
790                                 struct buffer_head *bh_result, int flags)
791 {
792         int dio_credits;
793         handle_t *handle;
794         int retries = 0;
795         int ret;
796
797         /* Trim mapping request to maximum we can map at once for DIO */
798         if (bh_result->b_size >> inode->i_blkbits > DIO_MAX_BLOCKS)
799                 bh_result->b_size = DIO_MAX_BLOCKS << inode->i_blkbits;
800         dio_credits = ext4_chunk_trans_blocks(inode,
801                                       bh_result->b_size >> inode->i_blkbits);
802 retry:
803         handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, dio_credits);
804         if (IS_ERR(handle))
805                 return PTR_ERR(handle);
806
807         ret = _ext4_get_block(inode, iblock, bh_result, flags);
808         ext4_journal_stop(handle);
809
810         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
811                 goto retry;
812         return ret;
813 }
814
815 /* Get block function for DIO reads and writes to inodes without extents */
816 int ext4_dio_get_block(struct inode *inode, sector_t iblock,
817                        struct buffer_head *bh, int create)
818 {
819         /* We don't expect handle for direct IO */
820         WARN_ON_ONCE(ext4_journal_current_handle());
821
822         if (!create)
823                 return _ext4_get_block(inode, iblock, bh, 0);
824         return ext4_get_block_trans(inode, iblock, bh, EXT4_GET_BLOCKS_CREATE);
825 }
826
827 /*
828  * Get block function for AIO DIO writes when we create unwritten extent if
829  * blocks are not allocated yet. The extent will be converted to written
830  * after IO is complete.
831  */
832 static int ext4_dio_get_block_unwritten_async(struct inode *inode,
833                 sector_t iblock, struct buffer_head *bh_result, int create)
834 {
835         int ret;
836
837         /* We don't expect handle for direct IO */
838         WARN_ON_ONCE(ext4_journal_current_handle());
839
840         ret = ext4_get_block_trans(inode, iblock, bh_result,
841                                    EXT4_GET_BLOCKS_IO_CREATE_EXT);
842
843         /*
844          * When doing DIO using unwritten extents, we need io_end to convert
845          * unwritten extents to written on IO completion. We allocate io_end
846          * once we spot unwritten extent and store it in b_private. Generic
847          * DIO code keeps b_private set and furthermore passes the value to
848          * our completion callback in 'private' argument.
849          */
850         if (!ret && buffer_unwritten(bh_result)) {
851                 if (!bh_result->b_private) {
852                         ext4_io_end_t *io_end;
853
854                         io_end = ext4_init_io_end(inode, GFP_KERNEL);
855                         if (!io_end)
856                                 return -ENOMEM;
857                         bh_result->b_private = io_end;
858                         ext4_set_io_unwritten_flag(inode, io_end);
859                 }
860                 set_buffer_defer_completion(bh_result);
861         }
862
863         return ret;
864 }
865
866 /*
867  * Get block function for non-AIO DIO writes when we create unwritten extent if
868  * blocks are not allocated yet. The extent will be converted to written
869  * after IO is complete from ext4_ext_direct_IO() function.
870  */
871 static int ext4_dio_get_block_unwritten_sync(struct inode *inode,
872                 sector_t iblock, struct buffer_head *bh_result, int create)
873 {
874         int ret;
875
876         /* We don't expect handle for direct IO */
877         WARN_ON_ONCE(ext4_journal_current_handle());
878
879         ret = ext4_get_block_trans(inode, iblock, bh_result,
880                                    EXT4_GET_BLOCKS_IO_CREATE_EXT);
881
882         /*
883          * Mark inode as having pending DIO writes to unwritten extents.
884          * ext4_ext_direct_IO() checks this flag and converts extents to
885          * written.
886          */
887         if (!ret && buffer_unwritten(bh_result))
888                 ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
889
890         return ret;
891 }
892
893 static int ext4_dio_get_block_overwrite(struct inode *inode, sector_t iblock,
894                    struct buffer_head *bh_result, int create)
895 {
896         int ret;
897
898         ext4_debug("ext4_dio_get_block_overwrite: inode %lu, create flag %d\n",
899                    inode->i_ino, create);
900         /* We don't expect handle for direct IO */
901         WARN_ON_ONCE(ext4_journal_current_handle());
902
903         ret = _ext4_get_block(inode, iblock, bh_result, 0);
904         /*
905          * Blocks should have been preallocated! ext4_file_write_iter() checks
906          * that.
907          */
908         WARN_ON_ONCE(!buffer_mapped(bh_result) || buffer_unwritten(bh_result));
909
910         return ret;
911 }
912
913
914 /*
915  * `handle' can be NULL if create is zero
916  */
917 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
918                                 ext4_lblk_t block, int map_flags)
919 {
920         struct ext4_map_blocks map;
921         struct buffer_head *bh;
922         int create = map_flags & EXT4_GET_BLOCKS_CREATE;
923         int err;
924
925         J_ASSERT(handle != NULL || create == 0);
926
927         map.m_lblk = block;
928         map.m_len = 1;
929         err = ext4_map_blocks(handle, inode, &map, map_flags);
930
931         if (err == 0)
932                 return create ? ERR_PTR(-ENOSPC) : NULL;
933         if (err < 0)
934                 return ERR_PTR(err);
935
936         bh = sb_getblk(inode->i_sb, map.m_pblk);
937         if (unlikely(!bh))
938                 return ERR_PTR(-ENOMEM);
939         if (map.m_flags & EXT4_MAP_NEW) {
940                 J_ASSERT(create != 0);
941                 J_ASSERT(handle != NULL);
942
943                 /*
944                  * Now that we do not always journal data, we should
945                  * keep in mind whether this should always journal the
946                  * new buffer as metadata.  For now, regular file
947                  * writes use ext4_get_block instead, so it's not a
948                  * problem.
949                  */
950                 lock_buffer(bh);
951                 BUFFER_TRACE(bh, "call get_create_access");
952                 err = ext4_journal_get_create_access(handle, bh);
953                 if (unlikely(err)) {
954                         unlock_buffer(bh);
955                         goto errout;
956                 }
957                 if (!buffer_uptodate(bh)) {
958                         memset(bh->b_data, 0, inode->i_sb->s_blocksize);
959                         set_buffer_uptodate(bh);
960                 }
961                 unlock_buffer(bh);
962                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
963                 err = ext4_handle_dirty_metadata(handle, inode, bh);
964                 if (unlikely(err))
965                         goto errout;
966         } else
967                 BUFFER_TRACE(bh, "not a new buffer");
968         return bh;
969 errout:
970         brelse(bh);
971         return ERR_PTR(err);
972 }
973
974 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
975                                ext4_lblk_t block, int map_flags)
976 {
977         struct buffer_head *bh;
978
979         bh = ext4_getblk(handle, inode, block, map_flags);
980         if (IS_ERR(bh))
981                 return bh;
982         if (!bh || buffer_uptodate(bh))
983                 return bh;
984         ll_rw_block(READ | REQ_META | REQ_PRIO, 1, &bh);
985         wait_on_buffer(bh);
986         if (buffer_uptodate(bh))
987                 return bh;
988         put_bh(bh);
989         return ERR_PTR(-EIO);
990 }
991
992 int ext4_walk_page_buffers(handle_t *handle,
993                            struct buffer_head *head,
994                            unsigned from,
995                            unsigned to,
996                            int *partial,
997                            int (*fn)(handle_t *handle,
998                                      struct buffer_head *bh))
999 {
1000         struct buffer_head *bh;
1001         unsigned block_start, block_end;
1002         unsigned blocksize = head->b_size;
1003         int err, ret = 0;
1004         struct buffer_head *next;
1005
1006         for (bh = head, block_start = 0;
1007              ret == 0 && (bh != head || !block_start);
1008              block_start = block_end, bh = next) {
1009                 next = bh->b_this_page;
1010                 block_end = block_start + blocksize;
1011                 if (block_end <= from || block_start >= to) {
1012                         if (partial && !buffer_uptodate(bh))
1013                                 *partial = 1;
1014                         continue;
1015                 }
1016                 err = (*fn)(handle, bh);
1017                 if (!ret)
1018                         ret = err;
1019         }
1020         return ret;
1021 }
1022
1023 /*
1024  * To preserve ordering, it is essential that the hole instantiation and
1025  * the data write be encapsulated in a single transaction.  We cannot
1026  * close off a transaction and start a new one between the ext4_get_block()
1027  * and the commit_write().  So doing the jbd2_journal_start at the start of
1028  * prepare_write() is the right place.
1029  *
1030  * Also, this function can nest inside ext4_writepage().  In that case, we
1031  * *know* that ext4_writepage() has generated enough buffer credits to do the
1032  * whole page.  So we won't block on the journal in that case, which is good,
1033  * because the caller may be PF_MEMALLOC.
1034  *
1035  * By accident, ext4 can be reentered when a transaction is open via
1036  * quota file writes.  If we were to commit the transaction while thus
1037  * reentered, there can be a deadlock - we would be holding a quota
1038  * lock, and the commit would never complete if another thread had a
1039  * transaction open and was blocking on the quota lock - a ranking
1040  * violation.
1041  *
1042  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1043  * will _not_ run commit under these circumstances because handle->h_ref
1044  * is elevated.  We'll still have enough credits for the tiny quotafile
1045  * write.
1046  */
1047 int do_journal_get_write_access(handle_t *handle,
1048                                 struct buffer_head *bh)
1049 {
1050         int dirty = buffer_dirty(bh);
1051         int ret;
1052
1053         if (!buffer_mapped(bh) || buffer_freed(bh))
1054                 return 0;
1055         /*
1056          * __block_write_begin() could have dirtied some buffers. Clean
1057          * the dirty bit as jbd2_journal_get_write_access() could complain
1058          * otherwise about fs integrity issues. Setting of the dirty bit
1059          * by __block_write_begin() isn't a real problem here as we clear
1060          * the bit before releasing a page lock and thus writeback cannot
1061          * ever write the buffer.
1062          */
1063         if (dirty)
1064                 clear_buffer_dirty(bh);
1065         BUFFER_TRACE(bh, "get write access");
1066         ret = ext4_journal_get_write_access(handle, bh);
1067         if (!ret && dirty)
1068                 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1069         return ret;
1070 }
1071
1072 #ifdef CONFIG_EXT4_FS_ENCRYPTION
1073 static int ext4_block_write_begin(struct page *page, loff_t pos, unsigned len,
1074                                   get_block_t *get_block)
1075 {
1076         unsigned from = pos & (PAGE_SIZE - 1);
1077         unsigned to = from + len;
1078         struct inode *inode = page->mapping->host;
1079         unsigned block_start, block_end;
1080         sector_t block;
1081         int err = 0;
1082         unsigned blocksize = inode->i_sb->s_blocksize;
1083         unsigned bbits;
1084         struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
1085         bool decrypt = false;
1086
1087         BUG_ON(!PageLocked(page));
1088         BUG_ON(from > PAGE_SIZE);
1089         BUG_ON(to > PAGE_SIZE);
1090         BUG_ON(from > to);
1091
1092         if (!page_has_buffers(page))
1093                 create_empty_buffers(page, blocksize, 0);
1094         head = page_buffers(page);
1095         bbits = ilog2(blocksize);
1096         block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1097
1098         for (bh = head, block_start = 0; bh != head || !block_start;
1099             block++, block_start = block_end, bh = bh->b_this_page) {
1100                 block_end = block_start + blocksize;
1101                 if (block_end <= from || block_start >= to) {
1102                         if (PageUptodate(page)) {
1103                                 if (!buffer_uptodate(bh))
1104                                         set_buffer_uptodate(bh);
1105                         }
1106                         continue;
1107                 }
1108                 if (buffer_new(bh))
1109                         clear_buffer_new(bh);
1110                 if (!buffer_mapped(bh)) {
1111                         WARN_ON(bh->b_size != blocksize);
1112                         err = get_block(inode, block, bh, 1);
1113                         if (err)
1114                                 break;
1115                         if (buffer_new(bh)) {
1116                                 unmap_underlying_metadata(bh->b_bdev,
1117                                                           bh->b_blocknr);
1118                                 if (PageUptodate(page)) {
1119                                         clear_buffer_new(bh);
1120                                         set_buffer_uptodate(bh);
1121                                         mark_buffer_dirty(bh);
1122                                         continue;
1123                                 }
1124                                 if (block_end > to || block_start < from)
1125                                         zero_user_segments(page, to, block_end,
1126                                                            block_start, from);
1127                                 continue;
1128                         }
1129                 }
1130                 if (PageUptodate(page)) {
1131                         if (!buffer_uptodate(bh))
1132                                 set_buffer_uptodate(bh);
1133                         continue;
1134                 }
1135                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1136                     !buffer_unwritten(bh) &&
1137                     (block_start < from || block_end > to)) {
1138                         ll_rw_block(READ, 1, &bh);
1139                         *wait_bh++ = bh;
1140                         decrypt = ext4_encrypted_inode(inode) &&
1141                                 S_ISREG(inode->i_mode);
1142                 }
1143         }
1144         /*
1145          * If we issued read requests, let them complete.
1146          */
1147         while (wait_bh > wait) {
1148                 wait_on_buffer(*--wait_bh);
1149                 if (!buffer_uptodate(*wait_bh))
1150                         err = -EIO;
1151         }
1152         if (unlikely(err))
1153                 page_zero_new_buffers(page, from, to);
1154         else if (decrypt)
1155                 err = ext4_decrypt(page);
1156         return err;
1157 }
1158 #endif
1159
1160 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1161                             loff_t pos, unsigned len, unsigned flags,
1162                             struct page **pagep, void **fsdata)
1163 {
1164         struct inode *inode = mapping->host;
1165         int ret, needed_blocks;
1166         handle_t *handle;
1167         int retries = 0;
1168         struct page *page;
1169         pgoff_t index;
1170         unsigned from, to;
1171
1172         trace_ext4_write_begin(inode, pos, len, flags);
1173         /*
1174          * Reserve one block more for addition to orphan list in case
1175          * we allocate blocks but write fails for some reason
1176          */
1177         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1178         index = pos >> PAGE_SHIFT;
1179         from = pos & (PAGE_SIZE - 1);
1180         to = from + len;
1181
1182         if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
1183                 ret = ext4_try_to_write_inline_data(mapping, inode, pos, len,
1184                                                     flags, pagep);
1185                 if (ret < 0)
1186                         return ret;
1187                 if (ret == 1)
1188                         return 0;
1189         }
1190
1191         /*
1192          * grab_cache_page_write_begin() can take a long time if the
1193          * system is thrashing due to memory pressure, or if the page
1194          * is being written back.  So grab it first before we start
1195          * the transaction handle.  This also allows us to allocate
1196          * the page (if needed) without using GFP_NOFS.
1197          */
1198 retry_grab:
1199         page = grab_cache_page_write_begin(mapping, index, flags);
1200         if (!page)
1201                 return -ENOMEM;
1202         unlock_page(page);
1203
1204 retry_journal:
1205         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE, needed_blocks);
1206         if (IS_ERR(handle)) {
1207                 put_page(page);
1208                 return PTR_ERR(handle);
1209         }
1210
1211         lock_page(page);
1212         if (page->mapping != mapping) {
1213                 /* The page got truncated from under us */
1214                 unlock_page(page);
1215                 put_page(page);
1216                 ext4_journal_stop(handle);
1217                 goto retry_grab;
1218         }
1219         /* In case writeback began while the page was unlocked */
1220         wait_for_stable_page(page);
1221
1222 #ifdef CONFIG_EXT4_FS_ENCRYPTION
1223         if (ext4_should_dioread_nolock(inode))
1224                 ret = ext4_block_write_begin(page, pos, len,
1225                                              ext4_get_block_unwritten);
1226         else
1227                 ret = ext4_block_write_begin(page, pos, len,
1228                                              ext4_get_block);
1229 #else
1230         if (ext4_should_dioread_nolock(inode))
1231                 ret = __block_write_begin(page, pos, len,
1232                                           ext4_get_block_unwritten);
1233         else
1234                 ret = __block_write_begin(page, pos, len, ext4_get_block);
1235 #endif
1236         if (!ret && ext4_should_journal_data(inode)) {
1237                 ret = ext4_walk_page_buffers(handle, page_buffers(page),
1238                                              from, to, NULL,
1239                                              do_journal_get_write_access);
1240         }
1241
1242         if (ret) {
1243                 unlock_page(page);
1244                 /*
1245                  * __block_write_begin may have instantiated a few blocks
1246                  * outside i_size.  Trim these off again. Don't need
1247                  * i_size_read because we hold i_mutex.
1248                  *
1249                  * Add inode to orphan list in case we crash before
1250                  * truncate finishes
1251                  */
1252                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1253                         ext4_orphan_add(handle, inode);
1254
1255                 ext4_journal_stop(handle);
1256                 if (pos + len > inode->i_size) {
1257                         ext4_truncate_failed_write(inode);
1258                         /*
1259                          * If truncate failed early the inode might
1260                          * still be on the orphan list; we need to
1261                          * make sure the inode is removed from the
1262                          * orphan list in that case.
1263                          */
1264                         if (inode->i_nlink)
1265                                 ext4_orphan_del(NULL, inode);
1266                 }
1267
1268                 if (ret == -ENOSPC &&
1269                     ext4_should_retry_alloc(inode->i_sb, &retries))
1270                         goto retry_journal;
1271                 put_page(page);
1272                 return ret;
1273         }
1274         *pagep = page;
1275         return ret;
1276 }
1277
1278 /* For write_end() in data=journal mode */
1279 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1280 {
1281         int ret;
1282         if (!buffer_mapped(bh) || buffer_freed(bh))
1283                 return 0;
1284         set_buffer_uptodate(bh);
1285         ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1286         clear_buffer_meta(bh);
1287         clear_buffer_prio(bh);
1288         return ret;
1289 }
1290
1291 /*
1292  * We need to pick up the new inode size which generic_commit_write gave us
1293  * `file' can be NULL - eg, when called from page_symlink().
1294  *
1295  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1296  * buffers are managed internally.
1297  */
1298 static int ext4_write_end(struct file *file,
1299                           struct address_space *mapping,
1300                           loff_t pos, unsigned len, unsigned copied,
1301                           struct page *page, void *fsdata)
1302 {
1303         handle_t *handle = ext4_journal_current_handle();
1304         struct inode *inode = mapping->host;
1305         loff_t old_size = inode->i_size;
1306         int ret = 0, ret2;
1307         int i_size_changed = 0;
1308
1309         trace_ext4_write_end(inode, pos, len, copied);
1310         if (ext4_has_inline_data(inode)) {
1311                 ret = ext4_write_inline_data_end(inode, pos, len,
1312                                                  copied, page);
1313                 if (ret < 0)
1314                         goto errout;
1315                 copied = ret;
1316         } else
1317                 copied = block_write_end(file, mapping, pos,
1318                                          len, copied, page, fsdata);
1319         /*
1320          * it's important to update i_size while still holding page lock:
1321          * page writeout could otherwise come in and zero beyond i_size.
1322          */
1323         i_size_changed = ext4_update_inode_size(inode, pos + copied);
1324         unlock_page(page);
1325         put_page(page);
1326
1327         if (old_size < pos)
1328                 pagecache_isize_extended(inode, old_size, pos);
1329         /*
1330          * Don't mark the inode dirty under page lock. First, it unnecessarily
1331          * makes the holding time of page lock longer. Second, it forces lock
1332          * ordering of page lock and transaction start for journaling
1333          * filesystems.
1334          */
1335         if (i_size_changed)
1336                 ext4_mark_inode_dirty(handle, inode);
1337
1338         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1339                 /* if we have allocated more blocks and copied
1340                  * less. We will have blocks allocated outside
1341                  * inode->i_size. So truncate them
1342                  */
1343                 ext4_orphan_add(handle, inode);
1344 errout:
1345         ret2 = ext4_journal_stop(handle);
1346         if (!ret)
1347                 ret = ret2;
1348
1349         if (pos + len > inode->i_size) {
1350                 ext4_truncate_failed_write(inode);
1351                 /*
1352                  * If truncate failed early the inode might still be
1353                  * on the orphan list; we need to make sure the inode
1354                  * is removed from the orphan list in that case.
1355                  */
1356                 if (inode->i_nlink)
1357                         ext4_orphan_del(NULL, inode);
1358         }
1359
1360         return ret ? ret : copied;
1361 }
1362
1363 /*
1364  * This is a private version of page_zero_new_buffers() which doesn't
1365  * set the buffer to be dirty, since in data=journalled mode we need
1366  * to call ext4_handle_dirty_metadata() instead.
1367  */
1368 static void zero_new_buffers(struct page *page, unsigned from, unsigned to)
1369 {
1370         unsigned int block_start = 0, block_end;
1371         struct buffer_head *head, *bh;
1372
1373         bh = head = page_buffers(page);
1374         do {
1375                 block_end = block_start + bh->b_size;
1376                 if (buffer_new(bh)) {
1377                         if (block_end > from && block_start < to) {
1378                                 if (!PageUptodate(page)) {
1379                                         unsigned start, size;
1380
1381                                         start = max(from, block_start);
1382                                         size = min(to, block_end) - start;
1383
1384                                         zero_user(page, start, size);
1385                                         set_buffer_uptodate(bh);
1386                                 }
1387                                 clear_buffer_new(bh);
1388                         }
1389                 }
1390                 block_start = block_end;
1391                 bh = bh->b_this_page;
1392         } while (bh != head);
1393 }
1394
1395 static int ext4_journalled_write_end(struct file *file,
1396                                      struct address_space *mapping,
1397                                      loff_t pos, unsigned len, unsigned copied,
1398                                      struct page *page, void *fsdata)
1399 {
1400         handle_t *handle = ext4_journal_current_handle();
1401         struct inode *inode = mapping->host;
1402         loff_t old_size = inode->i_size;
1403         int ret = 0, ret2;
1404         int partial = 0;
1405         unsigned from, to;
1406         int size_changed = 0;
1407
1408         trace_ext4_journalled_write_end(inode, pos, len, copied);
1409         from = pos & (PAGE_SIZE - 1);
1410         to = from + len;
1411
1412         BUG_ON(!ext4_handle_valid(handle));
1413
1414         if (ext4_has_inline_data(inode))
1415                 copied = ext4_write_inline_data_end(inode, pos, len,
1416                                                     copied, page);
1417         else {
1418                 if (copied < len) {
1419                         if (!PageUptodate(page))
1420                                 copied = 0;
1421                         zero_new_buffers(page, from+copied, to);
1422                 }
1423
1424                 ret = ext4_walk_page_buffers(handle, page_buffers(page), from,
1425                                              to, &partial, write_end_fn);
1426                 if (!partial)
1427                         SetPageUptodate(page);
1428         }
1429         size_changed = ext4_update_inode_size(inode, pos + copied);
1430         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1431         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1432         unlock_page(page);
1433         put_page(page);
1434
1435         if (old_size < pos)
1436                 pagecache_isize_extended(inode, old_size, pos);
1437
1438         if (size_changed) {
1439                 ret2 = ext4_mark_inode_dirty(handle, inode);
1440                 if (!ret)
1441                         ret = ret2;
1442         }
1443
1444         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1445                 /* if we have allocated more blocks and copied
1446                  * less. We will have blocks allocated outside
1447                  * inode->i_size. So truncate them
1448                  */
1449                 ext4_orphan_add(handle, inode);
1450
1451         ret2 = ext4_journal_stop(handle);
1452         if (!ret)
1453                 ret = ret2;
1454         if (pos + len > inode->i_size) {
1455                 ext4_truncate_failed_write(inode);
1456                 /*
1457                  * If truncate failed early the inode might still be
1458                  * on the orphan list; we need to make sure the inode
1459                  * is removed from the orphan list in that case.
1460                  */
1461                 if (inode->i_nlink)
1462                         ext4_orphan_del(NULL, inode);
1463         }
1464
1465         return ret ? ret : copied;
1466 }
1467
1468 /*
1469  * Reserve space for a single cluster
1470  */
1471 static int ext4_da_reserve_space(struct inode *inode)
1472 {
1473         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1474         struct ext4_inode_info *ei = EXT4_I(inode);
1475         int ret;
1476
1477         /*
1478          * We will charge metadata quota at writeout time; this saves
1479          * us from metadata over-estimation, though we may go over by
1480          * a small amount in the end.  Here we just reserve for data.
1481          */
1482         ret = dquot_reserve_block(inode, EXT4_C2B(sbi, 1));
1483         if (ret)
1484                 return ret;
1485
1486         spin_lock(&ei->i_block_reservation_lock);
1487         if (ext4_claim_free_clusters(sbi, 1, 0)) {
1488                 spin_unlock(&ei->i_block_reservation_lock);
1489                 dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
1490                 return -ENOSPC;
1491         }
1492         ei->i_reserved_data_blocks++;
1493         trace_ext4_da_reserve_space(inode);
1494         spin_unlock(&ei->i_block_reservation_lock);
1495
1496         return 0;       /* success */
1497 }
1498
1499 static void ext4_da_release_space(struct inode *inode, int to_free)
1500 {
1501         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1502         struct ext4_inode_info *ei = EXT4_I(inode);
1503
1504         if (!to_free)
1505                 return;         /* Nothing to release, exit */
1506
1507         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1508
1509         trace_ext4_da_release_space(inode, to_free);
1510         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1511                 /*
1512                  * if there aren't enough reserved blocks, then the
1513                  * counter is messed up somewhere.  Since this
1514                  * function is called from invalidate page, it's
1515                  * harmless to return without any action.
1516                  */
1517                 ext4_warning(inode->i_sb, "ext4_da_release_space: "
1518                          "ino %lu, to_free %d with only %d reserved "
1519                          "data blocks", inode->i_ino, to_free,
1520                          ei->i_reserved_data_blocks);
1521                 WARN_ON(1);
1522                 to_free = ei->i_reserved_data_blocks;
1523         }
1524         ei->i_reserved_data_blocks -= to_free;
1525
1526         /* update fs dirty data blocks counter */
1527         percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1528
1529         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1530
1531         dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1532 }
1533
1534 static void ext4_da_page_release_reservation(struct page *page,
1535                                              unsigned int offset,
1536                                              unsigned int length)
1537 {
1538         int to_release = 0, contiguous_blks = 0;
1539         struct buffer_head *head, *bh;
1540         unsigned int curr_off = 0;
1541         struct inode *inode = page->mapping->host;
1542         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1543         unsigned int stop = offset + length;
1544         int num_clusters;
1545         ext4_fsblk_t lblk;
1546
1547         BUG_ON(stop > PAGE_SIZE || stop < length);
1548
1549         head = page_buffers(page);
1550         bh = head;
1551         do {
1552                 unsigned int next_off = curr_off + bh->b_size;
1553
1554                 if (next_off > stop)
1555                         break;
1556
1557                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1558                         to_release++;
1559                         contiguous_blks++;
1560                         clear_buffer_delay(bh);
1561                 } else if (contiguous_blks) {
1562                         lblk = page->index <<
1563                                (PAGE_SHIFT - inode->i_blkbits);
1564                         lblk += (curr_off >> inode->i_blkbits) -
1565                                 contiguous_blks;
1566                         ext4_es_remove_extent(inode, lblk, contiguous_blks);
1567                         contiguous_blks = 0;
1568                 }
1569                 curr_off = next_off;
1570         } while ((bh = bh->b_this_page) != head);
1571
1572         if (contiguous_blks) {
1573                 lblk = page->index << (PAGE_SHIFT - inode->i_blkbits);
1574                 lblk += (curr_off >> inode->i_blkbits) - contiguous_blks;
1575                 ext4_es_remove_extent(inode, lblk, contiguous_blks);
1576         }
1577
1578         /* If we have released all the blocks belonging to a cluster, then we
1579          * need to release the reserved space for that cluster. */
1580         num_clusters = EXT4_NUM_B2C(sbi, to_release);
1581         while (num_clusters > 0) {
1582                 lblk = (page->index << (PAGE_SHIFT - inode->i_blkbits)) +
1583                         ((num_clusters - 1) << sbi->s_cluster_bits);
1584                 if (sbi->s_cluster_ratio == 1 ||
1585                     !ext4_find_delalloc_cluster(inode, lblk))
1586                         ext4_da_release_space(inode, 1);
1587
1588                 num_clusters--;
1589         }
1590 }
1591
1592 /*
1593  * Delayed allocation stuff
1594  */
1595
1596 struct mpage_da_data {
1597         struct inode *inode;
1598         struct writeback_control *wbc;
1599
1600         pgoff_t first_page;     /* The first page to write */
1601         pgoff_t next_page;      /* Current page to examine */
1602         pgoff_t last_page;      /* Last page to examine */
1603         /*
1604          * Extent to map - this can be after first_page because that can be
1605          * fully mapped. We somewhat abuse m_flags to store whether the extent
1606          * is delalloc or unwritten.
1607          */
1608         struct ext4_map_blocks map;
1609         struct ext4_io_submit io_submit;        /* IO submission data */
1610 };
1611
1612 static void mpage_release_unused_pages(struct mpage_da_data *mpd,
1613                                        bool invalidate)
1614 {
1615         int nr_pages, i;
1616         pgoff_t index, end;
1617         struct pagevec pvec;
1618         struct inode *inode = mpd->inode;
1619         struct address_space *mapping = inode->i_mapping;
1620
1621         /* This is necessary when next_page == 0. */
1622         if (mpd->first_page >= mpd->next_page)
1623                 return;
1624
1625         index = mpd->first_page;
1626         end   = mpd->next_page - 1;
1627         if (invalidate) {
1628                 ext4_lblk_t start, last;
1629                 start = index << (PAGE_SHIFT - inode->i_blkbits);
1630                 last = end << (PAGE_SHIFT - inode->i_blkbits);
1631                 ext4_es_remove_extent(inode, start, last - start + 1);
1632         }
1633
1634         pagevec_init(&pvec, 0);
1635         while (index <= end) {
1636                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1637                 if (nr_pages == 0)
1638                         break;
1639                 for (i = 0; i < nr_pages; i++) {
1640                         struct page *page = pvec.pages[i];
1641                         if (page->index > end)
1642                                 break;
1643                         BUG_ON(!PageLocked(page));
1644                         BUG_ON(PageWriteback(page));
1645                         if (invalidate) {
1646                                 block_invalidatepage(page, 0, PAGE_SIZE);
1647                                 ClearPageUptodate(page);
1648                         }
1649                         unlock_page(page);
1650                 }
1651                 index = pvec.pages[nr_pages - 1]->index + 1;
1652                 pagevec_release(&pvec);
1653         }
1654 }
1655
1656 static void ext4_print_free_blocks(struct inode *inode)
1657 {
1658         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1659         struct super_block *sb = inode->i_sb;
1660         struct ext4_inode_info *ei = EXT4_I(inode);
1661
1662         ext4_msg(sb, KERN_CRIT, "Total free blocks count %lld",
1663                EXT4_C2B(EXT4_SB(inode->i_sb),
1664                         ext4_count_free_clusters(sb)));
1665         ext4_msg(sb, KERN_CRIT, "Free/Dirty block details");
1666         ext4_msg(sb, KERN_CRIT, "free_blocks=%lld",
1667                (long long) EXT4_C2B(EXT4_SB(sb),
1668                 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1669         ext4_msg(sb, KERN_CRIT, "dirty_blocks=%lld",
1670                (long long) EXT4_C2B(EXT4_SB(sb),
1671                 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1672         ext4_msg(sb, KERN_CRIT, "Block reservation details");
1673         ext4_msg(sb, KERN_CRIT, "i_reserved_data_blocks=%u",
1674                  ei->i_reserved_data_blocks);
1675         return;
1676 }
1677
1678 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1679 {
1680         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1681 }
1682
1683 /*
1684  * This function is grabs code from the very beginning of
1685  * ext4_map_blocks, but assumes that the caller is from delayed write
1686  * time. This function looks up the requested blocks and sets the
1687  * buffer delay bit under the protection of i_data_sem.
1688  */
1689 static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1690                               struct ext4_map_blocks *map,
1691                               struct buffer_head *bh)
1692 {
1693         struct extent_status es;
1694         int retval;
1695         sector_t invalid_block = ~((sector_t) 0xffff);
1696 #ifdef ES_AGGRESSIVE_TEST
1697         struct ext4_map_blocks orig_map;
1698
1699         memcpy(&orig_map, map, sizeof(*map));
1700 #endif
1701
1702         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1703                 invalid_block = ~0;
1704
1705         map->m_flags = 0;
1706         ext_debug("ext4_da_map_blocks(): inode %lu, max_blocks %u,"
1707                   "logical block %lu\n", inode->i_ino, map->m_len,
1708                   (unsigned long) map->m_lblk);
1709
1710         /* Lookup extent status tree firstly */
1711         if (ext4_es_lookup_extent(inode, iblock, &es)) {
1712                 if (ext4_es_is_hole(&es)) {
1713                         retval = 0;
1714                         down_read(&EXT4_I(inode)->i_data_sem);
1715                         goto add_delayed;
1716                 }
1717
1718                 /*
1719                  * Delayed extent could be allocated by fallocate.
1720                  * So we need to check it.
1721                  */
1722                 if (ext4_es_is_delayed(&es) && !ext4_es_is_unwritten(&es)) {
1723                         map_bh(bh, inode->i_sb, invalid_block);
1724                         set_buffer_new(bh);
1725                         set_buffer_delay(bh);
1726                         return 0;
1727                 }
1728
1729                 map->m_pblk = ext4_es_pblock(&es) + iblock - es.es_lblk;
1730                 retval = es.es_len - (iblock - es.es_lblk);
1731                 if (retval > map->m_len)
1732                         retval = map->m_len;
1733                 map->m_len = retval;
1734                 if (ext4_es_is_written(&es))
1735                         map->m_flags |= EXT4_MAP_MAPPED;
1736                 else if (ext4_es_is_unwritten(&es))
1737                         map->m_flags |= EXT4_MAP_UNWRITTEN;
1738                 else
1739                         BUG_ON(1);
1740
1741 #ifdef ES_AGGRESSIVE_TEST
1742                 ext4_map_blocks_es_recheck(NULL, inode, map, &orig_map, 0);
1743 #endif
1744                 return retval;
1745         }
1746
1747         /*
1748          * Try to see if we can get the block without requesting a new
1749          * file system block.
1750          */
1751         down_read(&EXT4_I(inode)->i_data_sem);
1752         if (ext4_has_inline_data(inode))
1753                 retval = 0;
1754         else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1755                 retval = ext4_ext_map_blocks(NULL, inode, map, 0);
1756         else
1757                 retval = ext4_ind_map_blocks(NULL, inode, map, 0);
1758
1759 add_delayed:
1760         if (retval == 0) {
1761                 int ret;
1762                 /*
1763                  * XXX: __block_prepare_write() unmaps passed block,
1764                  * is it OK?
1765                  */
1766                 /*
1767                  * If the block was allocated from previously allocated cluster,
1768                  * then we don't need to reserve it again. However we still need
1769                  * to reserve metadata for every block we're going to write.
1770                  */
1771                 if (EXT4_SB(inode->i_sb)->s_cluster_ratio == 1 ||
1772                     !ext4_find_delalloc_cluster(inode, map->m_lblk)) {
1773                         ret = ext4_da_reserve_space(inode);
1774                         if (ret) {
1775                                 /* not enough space to reserve */
1776                                 retval = ret;
1777                                 goto out_unlock;
1778                         }
1779                 }
1780
1781                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1782                                             ~0, EXTENT_STATUS_DELAYED);
1783                 if (ret) {
1784                         retval = ret;
1785                         goto out_unlock;
1786                 }
1787
1788                 map_bh(bh, inode->i_sb, invalid_block);
1789                 set_buffer_new(bh);
1790                 set_buffer_delay(bh);
1791         } else if (retval > 0) {
1792                 int ret;
1793                 unsigned int status;
1794
1795                 if (unlikely(retval != map->m_len)) {
1796                         ext4_warning(inode->i_sb,
1797                                      "ES len assertion failed for inode "
1798                                      "%lu: retval %d != map->m_len %d",
1799                                      inode->i_ino, retval, map->m_len);
1800                         WARN_ON(1);
1801                 }
1802
1803                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
1804                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
1805                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1806                                             map->m_pblk, status);
1807                 if (ret != 0)
1808                         retval = ret;
1809         }
1810
1811 out_unlock:
1812         up_read((&EXT4_I(inode)->i_data_sem));
1813
1814         return retval;
1815 }
1816
1817 /*
1818  * This is a special get_block_t callback which is used by
1819  * ext4_da_write_begin().  It will either return mapped block or
1820  * reserve space for a single block.
1821  *
1822  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1823  * We also have b_blocknr = -1 and b_bdev initialized properly
1824  *
1825  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1826  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1827  * initialized properly.
1828  */
1829 int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1830                            struct buffer_head *bh, int create)
1831 {
1832         struct ext4_map_blocks map;
1833         int ret = 0;
1834
1835         BUG_ON(create == 0);
1836         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1837
1838         map.m_lblk = iblock;
1839         map.m_len = 1;
1840
1841         /*
1842          * first, we need to know whether the block is allocated already
1843          * preallocated blocks are unmapped but should treated
1844          * the same as allocated blocks.
1845          */
1846         ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1847         if (ret <= 0)
1848                 return ret;
1849
1850         map_bh(bh, inode->i_sb, map.m_pblk);
1851         ext4_update_bh_state(bh, map.m_flags);
1852
1853         if (buffer_unwritten(bh)) {
1854                 /* A delayed write to unwritten bh should be marked
1855                  * new and mapped.  Mapped ensures that we don't do
1856                  * get_block multiple times when we write to the same
1857                  * offset and new ensures that we do proper zero out
1858                  * for partial write.
1859                  */
1860                 set_buffer_new(bh);
1861                 set_buffer_mapped(bh);
1862         }
1863         return 0;
1864 }
1865
1866 static int bget_one(handle_t *handle, struct buffer_head *bh)
1867 {
1868         get_bh(bh);
1869         return 0;
1870 }
1871
1872 static int bput_one(handle_t *handle, struct buffer_head *bh)
1873 {
1874         put_bh(bh);
1875         return 0;
1876 }
1877
1878 static int __ext4_journalled_writepage(struct page *page,
1879                                        unsigned int len)
1880 {
1881         struct address_space *mapping = page->mapping;
1882         struct inode *inode = mapping->host;
1883         struct buffer_head *page_bufs = NULL;
1884         handle_t *handle = NULL;
1885         int ret = 0, err = 0;
1886         int inline_data = ext4_has_inline_data(inode);
1887         struct buffer_head *inode_bh = NULL;
1888
1889         ClearPageChecked(page);
1890
1891         if (inline_data) {
1892                 BUG_ON(page->index != 0);
1893                 BUG_ON(len > ext4_get_max_inline_size(inode));
1894                 inode_bh = ext4_journalled_write_inline_data(inode, len, page);
1895                 if (inode_bh == NULL)
1896                         goto out;
1897         } else {
1898                 page_bufs = page_buffers(page);
1899                 if (!page_bufs) {
1900                         BUG();
1901                         goto out;
1902                 }
1903                 ext4_walk_page_buffers(handle, page_bufs, 0, len,
1904                                        NULL, bget_one);
1905         }
1906         /*
1907          * We need to release the page lock before we start the
1908          * journal, so grab a reference so the page won't disappear
1909          * out from under us.
1910          */
1911         get_page(page);
1912         unlock_page(page);
1913
1914         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
1915                                     ext4_writepage_trans_blocks(inode));
1916         if (IS_ERR(handle)) {
1917                 ret = PTR_ERR(handle);
1918                 put_page(page);
1919                 goto out_no_pagelock;
1920         }
1921         BUG_ON(!ext4_handle_valid(handle));
1922
1923         lock_page(page);
1924         put_page(page);
1925         if (page->mapping != mapping) {
1926                 /* The page got truncated from under us */
1927                 ext4_journal_stop(handle);
1928                 ret = 0;
1929                 goto out;
1930         }
1931
1932         if (inline_data) {
1933                 BUFFER_TRACE(inode_bh, "get write access");
1934                 ret = ext4_journal_get_write_access(handle, inode_bh);
1935
1936                 err = ext4_handle_dirty_metadata(handle, inode, inode_bh);
1937
1938         } else {
1939                 ret = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
1940                                              do_journal_get_write_access);
1941
1942                 err = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
1943                                              write_end_fn);
1944         }
1945         if (ret == 0)
1946                 ret = err;
1947         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1948         err = ext4_journal_stop(handle);
1949         if (!ret)
1950                 ret = err;
1951
1952         if (!ext4_has_inline_data(inode))
1953                 ext4_walk_page_buffers(NULL, page_bufs, 0, len,
1954                                        NULL, bput_one);
1955         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1956 out:
1957         unlock_page(page);
1958 out_no_pagelock:
1959         brelse(inode_bh);
1960         return ret;
1961 }
1962
1963 /*
1964  * Note that we don't need to start a transaction unless we're journaling data
1965  * because we should have holes filled from ext4_page_mkwrite(). We even don't
1966  * need to file the inode to the transaction's list in ordered mode because if
1967  * we are writing back data added by write(), the inode is already there and if
1968  * we are writing back data modified via mmap(), no one guarantees in which
1969  * transaction the data will hit the disk. In case we are journaling data, we
1970  * cannot start transaction directly because transaction start ranks above page
1971  * lock so we have to do some magic.
1972  *
1973  * This function can get called via...
1974  *   - ext4_writepages after taking page lock (have journal handle)
1975  *   - journal_submit_inode_data_buffers (no journal handle)
1976  *   - shrink_page_list via the kswapd/direct reclaim (no journal handle)
1977  *   - grab_page_cache when doing write_begin (have journal handle)
1978  *
1979  * We don't do any block allocation in this function. If we have page with
1980  * multiple blocks we need to write those buffer_heads that are mapped. This
1981  * is important for mmaped based write. So if we do with blocksize 1K
1982  * truncate(f, 1024);
1983  * a = mmap(f, 0, 4096);
1984  * a[0] = 'a';
1985  * truncate(f, 4096);
1986  * we have in the page first buffer_head mapped via page_mkwrite call back
1987  * but other buffer_heads would be unmapped but dirty (dirty done via the
1988  * do_wp_page). So writepage should write the first block. If we modify
1989  * the mmap area beyond 1024 we will again get a page_fault and the
1990  * page_mkwrite callback will do the block allocation and mark the
1991  * buffer_heads mapped.
1992  *
1993  * We redirty the page if we have any buffer_heads that is either delay or
1994  * unwritten in the page.
1995  *
1996  * We can get recursively called as show below.
1997  *
1998  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1999  *              ext4_writepage()
2000  *
2001  * But since we don't do any block allocation we should not deadlock.
2002  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2003  */
2004 static int ext4_writepage(struct page *page,
2005                           struct writeback_control *wbc)
2006 {
2007         int ret = 0;
2008         loff_t size;
2009         unsigned int len;
2010         struct buffer_head *page_bufs = NULL;
2011         struct inode *inode = page->mapping->host;
2012         struct ext4_io_submit io_submit;
2013         bool keep_towrite = false;
2014
2015         trace_ext4_writepage(page);
2016         size = i_size_read(inode);
2017         if (page->index == size >> PAGE_SHIFT)
2018                 len = size & ~PAGE_MASK;
2019         else
2020                 len = PAGE_SIZE;
2021
2022         page_bufs = page_buffers(page);
2023         /*
2024          * We cannot do block allocation or other extent handling in this
2025          * function. If there are buffers needing that, we have to redirty
2026          * the page. But we may reach here when we do a journal commit via
2027          * journal_submit_inode_data_buffers() and in that case we must write
2028          * allocated buffers to achieve data=ordered mode guarantees.
2029          *
2030          * Also, if there is only one buffer per page (the fs block
2031          * size == the page size), if one buffer needs block
2032          * allocation or needs to modify the extent tree to clear the
2033          * unwritten flag, we know that the page can't be written at
2034          * all, so we might as well refuse the write immediately.
2035          * Unfortunately if the block size != page size, we can't as
2036          * easily detect this case using ext4_walk_page_buffers(), but
2037          * for the extremely common case, this is an optimization that
2038          * skips a useless round trip through ext4_bio_write_page().
2039          */
2040         if (ext4_walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2041                                    ext4_bh_delay_or_unwritten)) {
2042                 redirty_page_for_writepage(wbc, page);
2043                 if ((current->flags & PF_MEMALLOC) ||
2044                     (inode->i_sb->s_blocksize == PAGE_SIZE)) {
2045                         /*
2046                          * For memory cleaning there's no point in writing only
2047                          * some buffers. So just bail out. Warn if we came here
2048                          * from direct reclaim.
2049                          */
2050                         WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD))
2051                                                         == PF_MEMALLOC);
2052                         unlock_page(page);
2053                         return 0;
2054                 }
2055                 keep_towrite = true;
2056         }
2057
2058         if (PageChecked(page) && ext4_should_journal_data(inode))
2059                 /*
2060                  * It's mmapped pagecache.  Add buffers and journal it.  There
2061                  * doesn't seem much point in redirtying the page here.
2062                  */
2063                 return __ext4_journalled_writepage(page, len);
2064
2065         ext4_io_submit_init(&io_submit, wbc);
2066         io_submit.io_end = ext4_init_io_end(inode, GFP_NOFS);
2067         if (!io_submit.io_end) {
2068                 redirty_page_for_writepage(wbc, page);
2069                 unlock_page(page);
2070                 return -ENOMEM;
2071         }
2072         ret = ext4_bio_write_page(&io_submit, page, len, wbc, keep_towrite);
2073         ext4_io_submit(&io_submit);
2074         /* Drop io_end reference we got from init */
2075         ext4_put_io_end_defer(io_submit.io_end);
2076         return ret;
2077 }
2078
2079 static int mpage_submit_page(struct mpage_da_data *mpd, struct page *page)
2080 {
2081         int len;
2082         loff_t size = i_size_read(mpd->inode);
2083         int err;
2084
2085         BUG_ON(page->index != mpd->first_page);
2086         if (page->index == size >> PAGE_SHIFT)
2087                 len = size & ~PAGE_MASK;
2088         else
2089                 len = PAGE_SIZE;
2090         clear_page_dirty_for_io(page);
2091         err = ext4_bio_write_page(&mpd->io_submit, page, len, mpd->wbc, false);
2092         if (!err)
2093                 mpd->wbc->nr_to_write--;
2094         mpd->first_page++;
2095
2096         return err;
2097 }
2098
2099 #define BH_FLAGS ((1 << BH_Unwritten) | (1 << BH_Delay))
2100
2101 /*
2102  * mballoc gives us at most this number of blocks...
2103  * XXX: That seems to be only a limitation of ext4_mb_normalize_request().
2104  * The rest of mballoc seems to handle chunks up to full group size.
2105  */
2106 #define MAX_WRITEPAGES_EXTENT_LEN 2048
2107
2108 /*
2109  * mpage_add_bh_to_extent - try to add bh to extent of blocks to map
2110  *
2111  * @mpd - extent of blocks
2112  * @lblk - logical number of the block in the file
2113  * @bh - buffer head we want to add to the extent
2114  *
2115  * The function is used to collect contig. blocks in the same state. If the
2116  * buffer doesn't require mapping for writeback and we haven't started the
2117  * extent of buffers to map yet, the function returns 'true' immediately - the
2118  * caller can write the buffer right away. Otherwise the function returns true
2119  * if the block has been added to the extent, false if the block couldn't be
2120  * added.
2121  */
2122 static bool mpage_add_bh_to_extent(struct mpage_da_data *mpd, ext4_lblk_t lblk,
2123                                    struct buffer_head *bh)
2124 {
2125         struct ext4_map_blocks *map = &mpd->map;
2126
2127         /* Buffer that doesn't need mapping for writeback? */
2128         if (!buffer_dirty(bh) || !buffer_mapped(bh) ||
2129             (!buffer_delay(bh) && !buffer_unwritten(bh))) {
2130                 /* So far no extent to map => we write the buffer right away */
2131                 if (map->m_len == 0)
2132                         return true;
2133                 return false;
2134         }
2135
2136         /* First block in the extent? */
2137         if (map->m_len == 0) {
2138                 map->m_lblk = lblk;
2139                 map->m_len = 1;
2140                 map->m_flags = bh->b_state & BH_FLAGS;
2141                 return true;
2142         }
2143
2144         /* Don't go larger than mballoc is willing to allocate */
2145         if (map->m_len >= MAX_WRITEPAGES_EXTENT_LEN)
2146                 return false;
2147
2148         /* Can we merge the block to our big extent? */
2149         if (lblk == map->m_lblk + map->m_len &&
2150             (bh->b_state & BH_FLAGS) == map->m_flags) {
2151                 map->m_len++;
2152                 return true;
2153         }
2154         return false;
2155 }
2156
2157 /*
2158  * mpage_process_page_bufs - submit page buffers for IO or add them to extent
2159  *
2160  * @mpd - extent of blocks for mapping
2161  * @head - the first buffer in the page
2162  * @bh - buffer we should start processing from
2163  * @lblk - logical number of the block in the file corresponding to @bh
2164  *
2165  * Walk through page buffers from @bh upto @head (exclusive) and either submit
2166  * the page for IO if all buffers in this page were mapped and there's no
2167  * accumulated extent of buffers to map or add buffers in the page to the
2168  * extent of buffers to map. The function returns 1 if the caller can continue
2169  * by processing the next page, 0 if it should stop adding buffers to the
2170  * extent to map because we cannot extend it anymore. It can also return value
2171  * < 0 in case of error during IO submission.
2172  */
2173 static int mpage_process_page_bufs(struct mpage_da_data *mpd,
2174                                    struct buffer_head *head,
2175                                    struct buffer_head *bh,
2176                                    ext4_lblk_t lblk)
2177 {
2178         struct inode *inode = mpd->inode;
2179         int err;
2180         ext4_lblk_t blocks = (i_size_read(inode) + (1 << inode->i_blkbits) - 1)
2181                                                         >> inode->i_blkbits;
2182
2183         do {
2184                 BUG_ON(buffer_locked(bh));
2185
2186                 if (lblk >= blocks || !mpage_add_bh_to_extent(mpd, lblk, bh)) {
2187                         /* Found extent to map? */
2188                         if (mpd->map.m_len)
2189                                 return 0;
2190                         /* Everything mapped so far and we hit EOF */
2191                         break;
2192                 }
2193         } while (lblk++, (bh = bh->b_this_page) != head);
2194         /* So far everything mapped? Submit the page for IO. */
2195         if (mpd->map.m_len == 0) {
2196                 err = mpage_submit_page(mpd, head->b_page);
2197                 if (err < 0)
2198                         return err;
2199         }
2200         return lblk < blocks;
2201 }
2202
2203 /*
2204  * mpage_map_buffers - update buffers corresponding to changed extent and
2205  *                     submit fully mapped pages for IO
2206  *
2207  * @mpd - description of extent to map, on return next extent to map
2208  *
2209  * Scan buffers corresponding to changed extent (we expect corresponding pages
2210  * to be already locked) and update buffer state according to new extent state.
2211  * We map delalloc buffers to their physical location, clear unwritten bits,
2212  * and mark buffers as uninit when we perform writes to unwritten extents
2213  * and do extent conversion after IO is finished. If the last page is not fully
2214  * mapped, we update @map to the next extent in the last page that needs
2215  * mapping. Otherwise we submit the page for IO.
2216  */
2217 static int mpage_map_and_submit_buffers(struct mpage_da_data *mpd)
2218 {
2219         struct pagevec pvec;
2220         int nr_pages, i;
2221         struct inode *inode = mpd->inode;
2222         struct buffer_head *head, *bh;
2223         int bpp_bits = PAGE_SHIFT - inode->i_blkbits;
2224         pgoff_t start, end;
2225         ext4_lblk_t lblk;
2226         sector_t pblock;
2227         int err;
2228
2229         start = mpd->map.m_lblk >> bpp_bits;
2230         end = (mpd->map.m_lblk + mpd->map.m_len - 1) >> bpp_bits;
2231         lblk = start << bpp_bits;
2232         pblock = mpd->map.m_pblk;
2233
2234         pagevec_init(&pvec, 0);
2235         while (start <= end) {
2236                 nr_pages = pagevec_lookup(&pvec, inode->i_mapping, start,
2237                                           PAGEVEC_SIZE);
2238                 if (nr_pages == 0)
2239                         break;
2240                 for (i = 0; i < nr_pages; i++) {
2241                         struct page *page = pvec.pages[i];
2242
2243                         if (page->index > end)
2244                                 break;
2245                         /* Up to 'end' pages must be contiguous */
2246                         BUG_ON(page->index != start);
2247                         bh = head = page_buffers(page);
2248                         do {
2249                                 if (lblk < mpd->map.m_lblk)
2250                                         continue;
2251                                 if (lblk >= mpd->map.m_lblk + mpd->map.m_len) {
2252                                         /*
2253                                          * Buffer after end of mapped extent.
2254                                          * Find next buffer in the page to map.
2255                                          */
2256                                         mpd->map.m_len = 0;
2257                                         mpd->map.m_flags = 0;
2258                                         /*
2259                                          * FIXME: If dioread_nolock supports
2260                                          * blocksize < pagesize, we need to make
2261                                          * sure we add size mapped so far to
2262                                          * io_end->size as the following call
2263                                          * can submit the page for IO.
2264                                          */
2265                                         err = mpage_process_page_bufs(mpd, head,
2266                                                                       bh, lblk);
2267                                         pagevec_release(&pvec);
2268                                         if (err > 0)
2269                                                 err = 0;
2270                                         return err;
2271                                 }
2272                                 if (buffer_delay(bh)) {
2273                                         clear_buffer_delay(bh);
2274                                         bh->b_blocknr = pblock++;
2275                                 }
2276                                 clear_buffer_unwritten(bh);
2277                         } while (lblk++, (bh = bh->b_this_page) != head);
2278
2279                         /*
2280                          * FIXME: This is going to break if dioread_nolock
2281                          * supports blocksize < pagesize as we will try to
2282                          * convert potentially unmapped parts of inode.
2283                          */
2284                         mpd->io_submit.io_end->size += PAGE_SIZE;
2285                         /* Page fully mapped - let IO run! */
2286                         err = mpage_submit_page(mpd, page);
2287                         if (err < 0) {
2288                                 pagevec_release(&pvec);
2289                                 return err;
2290                         }
2291                         start++;
2292                 }
2293                 pagevec_release(&pvec);
2294         }
2295         /* Extent fully mapped and matches with page boundary. We are done. */
2296         mpd->map.m_len = 0;
2297         mpd->map.m_flags = 0;
2298         return 0;
2299 }
2300
2301 static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
2302 {
2303         struct inode *inode = mpd->inode;
2304         struct ext4_map_blocks *map = &mpd->map;
2305         int get_blocks_flags;
2306         int err, dioread_nolock;
2307
2308         trace_ext4_da_write_pages_extent(inode, map);
2309         /*
2310          * Call ext4_map_blocks() to allocate any delayed allocation blocks, or
2311          * to convert an unwritten extent to be initialized (in the case
2312          * where we have written into one or more preallocated blocks).  It is
2313          * possible that we're going to need more metadata blocks than
2314          * previously reserved. However we must not fail because we're in
2315          * writeback and there is nothing we can do about it so it might result
2316          * in data loss.  So use reserved blocks to allocate metadata if
2317          * possible.
2318          *
2319          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE if
2320          * the blocks in question are delalloc blocks.  This indicates
2321          * that the blocks and quotas has already been checked when
2322          * the data was copied into the page cache.
2323          */
2324         get_blocks_flags = EXT4_GET_BLOCKS_CREATE |
2325                            EXT4_GET_BLOCKS_METADATA_NOFAIL |
2326                            EXT4_GET_BLOCKS_IO_SUBMIT;
2327         dioread_nolock = ext4_should_dioread_nolock(inode);
2328         if (dioread_nolock)
2329                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2330         if (map->m_flags & (1 << BH_Delay))
2331                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2332
2333         err = ext4_map_blocks(handle, inode, map, get_blocks_flags);
2334         if (err < 0)
2335                 return err;
2336         if (dioread_nolock && (map->m_flags & EXT4_MAP_UNWRITTEN)) {
2337                 if (!mpd->io_submit.io_end->handle &&
2338                     ext4_handle_valid(handle)) {
2339                         mpd->io_submit.io_end->handle = handle->h_rsv_handle;
2340                         handle->h_rsv_handle = NULL;
2341                 }
2342                 ext4_set_io_unwritten_flag(inode, mpd->io_submit.io_end);
2343         }
2344
2345         BUG_ON(map->m_len == 0);
2346         if (map->m_flags & EXT4_MAP_NEW) {
2347                 struct block_device *bdev = inode->i_sb->s_bdev;
2348                 int i;
2349
2350                 for (i = 0; i < map->m_len; i++)
2351                         unmap_underlying_metadata(bdev, map->m_pblk + i);
2352         }
2353         return 0;
2354 }
2355
2356 /*
2357  * mpage_map_and_submit_extent - map extent starting at mpd->lblk of length
2358  *                               mpd->len and submit pages underlying it for IO
2359  *
2360  * @handle - handle for journal operations
2361  * @mpd - extent to map
2362  * @give_up_on_write - we set this to true iff there is a fatal error and there
2363  *                     is no hope of writing the data. The caller should discard
2364  *                     dirty pages to avoid infinite loops.
2365  *
2366  * The function maps extent starting at mpd->lblk of length mpd->len. If it is
2367  * delayed, blocks are allocated, if it is unwritten, we may need to convert
2368  * them to initialized or split the described range from larger unwritten
2369  * extent. Note that we need not map all the described range since allocation
2370  * can return less blocks or the range is covered by more unwritten extents. We
2371  * cannot map more because we are limited by reserved transaction credits. On
2372  * the other hand we always make sure that the last touched page is fully
2373  * mapped so that it can be written out (and thus forward progress is
2374  * guaranteed). After mapping we submit all mapped pages for IO.
2375  */
2376 static int mpage_map_and_submit_extent(handle_t *handle,
2377                                        struct mpage_da_data *mpd,
2378                                        bool *give_up_on_write)
2379 {
2380         struct inode *inode = mpd->inode;
2381         struct ext4_map_blocks *map = &mpd->map;
2382         int err;
2383         loff_t disksize;
2384         int progress = 0;
2385
2386         mpd->io_submit.io_end->offset =
2387                                 ((loff_t)map->m_lblk) << inode->i_blkbits;
2388         do {
2389                 err = mpage_map_one_extent(handle, mpd);
2390                 if (err < 0) {
2391                         struct super_block *sb = inode->i_sb;
2392
2393                         if (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
2394                                 goto invalidate_dirty_pages;
2395                         /*
2396                          * Let the uper layers retry transient errors.
2397                          * In the case of ENOSPC, if ext4_count_free_blocks()
2398                          * is non-zero, a commit should free up blocks.
2399                          */
2400                         if ((err == -ENOMEM) ||
2401                             (err == -ENOSPC && ext4_count_free_clusters(sb))) {
2402                                 if (progress)
2403                                         goto update_disksize;
2404                                 return err;
2405                         }
2406                         ext4_msg(sb, KERN_CRIT,
2407                                  "Delayed block allocation failed for "
2408                                  "inode %lu at logical offset %llu with"
2409                                  " max blocks %u with error %d",
2410                                  inode->i_ino,
2411                                  (unsigned long long)map->m_lblk,
2412                                  (unsigned)map->m_len, -err);
2413                         ext4_msg(sb, KERN_CRIT,
2414                                  "This should not happen!! Data will "
2415                                  "be lost\n");
2416                         if (err == -ENOSPC)
2417                                 ext4_print_free_blocks(inode);
2418                 invalidate_dirty_pages:
2419                         *give_up_on_write = true;
2420                         return err;
2421                 }
2422                 progress = 1;
2423                 /*
2424                  * Update buffer state, submit mapped pages, and get us new
2425                  * extent to map
2426                  */
2427                 err = mpage_map_and_submit_buffers(mpd);
2428                 if (err < 0)
2429                         goto update_disksize;
2430         } while (map->m_len);
2431
2432 update_disksize:
2433         /*
2434          * Update on-disk size after IO is submitted.  Races with
2435          * truncate are avoided by checking i_size under i_data_sem.
2436          */
2437         disksize = ((loff_t)mpd->first_page) << PAGE_SHIFT;
2438         if (disksize > EXT4_I(inode)->i_disksize) {
2439                 int err2;
2440                 loff_t i_size;
2441
2442                 down_write(&EXT4_I(inode)->i_data_sem);
2443                 i_size = i_size_read(inode);
2444                 if (disksize > i_size)
2445                         disksize = i_size;
2446                 if (disksize > EXT4_I(inode)->i_disksize)
2447                         EXT4_I(inode)->i_disksize = disksize;
2448                 err2 = ext4_mark_inode_dirty(handle, inode);
2449                 up_write(&EXT4_I(inode)->i_data_sem);
2450                 if (err2)
2451                         ext4_error(inode->i_sb,
2452                                    "Failed to mark inode %lu dirty",
2453                                    inode->i_ino);
2454                 if (!err)
2455                         err = err2;
2456         }
2457         return err;
2458 }
2459
2460 /*
2461  * Calculate the total number of credits to reserve for one writepages
2462  * iteration. This is called from ext4_writepages(). We map an extent of
2463  * up to MAX_WRITEPAGES_EXTENT_LEN blocks and then we go on and finish mapping
2464  * the last partial page. So in total we can map MAX_WRITEPAGES_EXTENT_LEN +
2465  * bpp - 1 blocks in bpp different extents.
2466  */
2467 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2468 {
2469         int bpp = ext4_journal_blocks_per_page(inode);
2470
2471         return ext4_meta_trans_blocks(inode,
2472                                 MAX_WRITEPAGES_EXTENT_LEN + bpp - 1, bpp);
2473 }
2474
2475 /*
2476  * mpage_prepare_extent_to_map - find & lock contiguous range of dirty pages
2477  *                               and underlying extent to map
2478  *
2479  * @mpd - where to look for pages
2480  *
2481  * Walk dirty pages in the mapping. If they are fully mapped, submit them for
2482  * IO immediately. When we find a page which isn't mapped we start accumulating
2483  * extent of buffers underlying these pages that needs mapping (formed by
2484  * either delayed or unwritten buffers). We also lock the pages containing
2485  * these buffers. The extent found is returned in @mpd structure (starting at
2486  * mpd->lblk with length mpd->len blocks).
2487  *
2488  * Note that this function can attach bios to one io_end structure which are
2489  * neither logically nor physically contiguous. Although it may seem as an
2490  * unnecessary complication, it is actually inevitable in blocksize < pagesize
2491  * case as we need to track IO to all buffers underlying a page in one io_end.
2492  */
2493 static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
2494 {
2495         struct address_space *mapping = mpd->inode->i_mapping;
2496         struct pagevec pvec;
2497         unsigned int nr_pages;
2498         long left = mpd->wbc->nr_to_write;
2499         pgoff_t index = mpd->first_page;
2500         pgoff_t end = mpd->last_page;
2501         int tag;
2502         int i, err = 0;
2503         int blkbits = mpd->inode->i_blkbits;
2504         ext4_lblk_t lblk;
2505         struct buffer_head *head;
2506
2507         if (mpd->wbc->sync_mode == WB_SYNC_ALL || mpd->wbc->tagged_writepages)
2508                 tag = PAGECACHE_TAG_TOWRITE;
2509         else
2510                 tag = PAGECACHE_TAG_DIRTY;
2511
2512         pagevec_init(&pvec, 0);
2513         mpd->map.m_len = 0;
2514         mpd->next_page = index;
2515         while (index <= end) {
2516                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2517                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2518                 if (nr_pages == 0)
2519                         goto out;
2520
2521                 for (i = 0; i < nr_pages; i++) {
2522                         struct page *page = pvec.pages[i];
2523
2524                         /*
2525                          * At this point, the page may be truncated or
2526                          * invalidated (changing page->mapping to NULL), or
2527                          * even swizzled back from swapper_space to tmpfs file
2528                          * mapping. However, page->index will not change
2529                          * because we have a reference on the page.
2530                          */
2531                         if (page->index > end)
2532                                 goto out;
2533
2534                         /*
2535                          * Accumulated enough dirty pages? This doesn't apply
2536                          * to WB_SYNC_ALL mode. For integrity sync we have to
2537                          * keep going because someone may be concurrently
2538                          * dirtying pages, and we might have synced a lot of
2539                          * newly appeared dirty pages, but have not synced all
2540                          * of the old dirty pages.
2541                          */
2542                         if (mpd->wbc->sync_mode == WB_SYNC_NONE && left <= 0)
2543                                 goto out;
2544
2545                         /* If we can't merge this page, we are done. */
2546                         if (mpd->map.m_len > 0 && mpd->next_page != page->index)
2547                                 goto out;
2548
2549                         lock_page(page);
2550                         /*
2551                          * If the page is no longer dirty, or its mapping no
2552                          * longer corresponds to inode we are writing (which
2553                          * means it has been truncated or invalidated), or the
2554                          * page is already under writeback and we are not doing
2555                          * a data integrity writeback, skip the page
2556                          */
2557                         if (!PageDirty(page) ||
2558                             (PageWriteback(page) &&
2559                              (mpd->wbc->sync_mode == WB_SYNC_NONE)) ||
2560                             unlikely(page->mapping != mapping)) {
2561                                 unlock_page(page);
2562                                 continue;
2563                         }
2564
2565                         wait_on_page_writeback(page);
2566                         BUG_ON(PageWriteback(page));
2567
2568                         if (mpd->map.m_len == 0)
2569                                 mpd->first_page = page->index;
2570                         mpd->next_page = page->index + 1;
2571                         /* Add all dirty buffers to mpd */
2572                         lblk = ((ext4_lblk_t)page->index) <<
2573                                 (PAGE_SHIFT - blkbits);
2574                         head = page_buffers(page);
2575                         err = mpage_process_page_bufs(mpd, head, head, lblk);
2576                         if (err <= 0)
2577                                 goto out;
2578                         err = 0;
2579                         left--;
2580                 }
2581                 pagevec_release(&pvec);
2582                 cond_resched();
2583         }
2584         return 0;
2585 out:
2586         pagevec_release(&pvec);
2587         return err;
2588 }
2589
2590 static int __writepage(struct page *page, struct writeback_control *wbc,
2591                        void *data)
2592 {
2593         struct address_space *mapping = data;
2594         int ret = ext4_writepage(page, wbc);
2595         mapping_set_error(mapping, ret);
2596         return ret;
2597 }
2598
2599 static int ext4_writepages(struct address_space *mapping,
2600                            struct writeback_control *wbc)
2601 {
2602         pgoff_t writeback_index = 0;
2603         long nr_to_write = wbc->nr_to_write;
2604         int range_whole = 0;
2605         int cycled = 1;
2606         handle_t *handle = NULL;
2607         struct mpage_da_data mpd;
2608         struct inode *inode = mapping->host;
2609         int needed_blocks, rsv_blocks = 0, ret = 0;
2610         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2611         bool done;
2612         struct blk_plug plug;
2613         bool give_up_on_write = false;
2614
2615         trace_ext4_writepages(inode, wbc);
2616
2617         if (dax_mapping(mapping))
2618                 return dax_writeback_mapping_range(mapping, inode->i_sb->s_bdev,
2619                                                    wbc);
2620
2621         /*
2622          * No pages to write? This is mainly a kludge to avoid starting
2623          * a transaction for special inodes like journal inode on last iput()
2624          * because that could violate lock ordering on umount
2625          */
2626         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2627                 goto out_writepages;
2628
2629         if (ext4_should_journal_data(inode)) {
2630                 struct blk_plug plug;
2631
2632                 blk_start_plug(&plug);
2633                 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2634                 blk_finish_plug(&plug);
2635                 goto out_writepages;
2636         }
2637
2638         /*
2639          * If the filesystem has aborted, it is read-only, so return
2640          * right away instead of dumping stack traces later on that
2641          * will obscure the real source of the problem.  We test
2642          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2643          * the latter could be true if the filesystem is mounted
2644          * read-only, and in that case, ext4_writepages should
2645          * *never* be called, so if that ever happens, we would want
2646          * the stack trace.
2647          */
2648         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2649                 ret = -EROFS;
2650                 goto out_writepages;
2651         }
2652
2653         if (ext4_should_dioread_nolock(inode)) {
2654                 /*
2655                  * We may need to convert up to one extent per block in
2656                  * the page and we may dirty the inode.
2657                  */
2658                 rsv_blocks = 1 + (PAGE_SIZE >> inode->i_blkbits);
2659         }
2660
2661         /*
2662          * If we have inline data and arrive here, it means that
2663          * we will soon create the block for the 1st page, so
2664          * we'd better clear the inline data here.
2665          */
2666         if (ext4_has_inline_data(inode)) {
2667                 /* Just inode will be modified... */
2668                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
2669                 if (IS_ERR(handle)) {
2670                         ret = PTR_ERR(handle);
2671                         goto out_writepages;
2672                 }
2673                 BUG_ON(ext4_test_inode_state(inode,
2674                                 EXT4_STATE_MAY_INLINE_DATA));
2675                 ext4_destroy_inline_data(handle, inode);
2676                 ext4_journal_stop(handle);
2677         }
2678
2679         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2680                 range_whole = 1;
2681
2682         if (wbc->range_cyclic) {
2683                 writeback_index = mapping->writeback_index;
2684                 if (writeback_index)
2685                         cycled = 0;
2686                 mpd.first_page = writeback_index;
2687                 mpd.last_page = -1;
2688         } else {
2689                 mpd.first_page = wbc->range_start >> PAGE_SHIFT;
2690                 mpd.last_page = wbc->range_end >> PAGE_SHIFT;
2691         }
2692
2693         mpd.inode = inode;
2694         mpd.wbc = wbc;
2695         ext4_io_submit_init(&mpd.io_submit, wbc);
2696 retry:
2697         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2698                 tag_pages_for_writeback(mapping, mpd.first_page, mpd.last_page);
2699         done = false;
2700         blk_start_plug(&plug);
2701         while (!done && mpd.first_page <= mpd.last_page) {
2702                 /* For each extent of pages we use new io_end */
2703                 mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2704                 if (!mpd.io_submit.io_end) {
2705                         ret = -ENOMEM;
2706                         break;
2707                 }
2708
2709                 /*
2710                  * We have two constraints: We find one extent to map and we
2711                  * must always write out whole page (makes a difference when
2712                  * blocksize < pagesize) so that we don't block on IO when we
2713                  * try to write out the rest of the page. Journalled mode is
2714                  * not supported by delalloc.
2715                  */
2716                 BUG_ON(ext4_should_journal_data(inode));
2717                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2718
2719                 /* start a new transaction */
2720                 handle = ext4_journal_start_with_reserve(inode,
2721                                 EXT4_HT_WRITE_PAGE, needed_blocks, rsv_blocks);
2722                 if (IS_ERR(handle)) {
2723                         ret = PTR_ERR(handle);
2724                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2725                                "%ld pages, ino %lu; err %d", __func__,
2726                                 wbc->nr_to_write, inode->i_ino, ret);
2727                         /* Release allocated io_end */
2728                         ext4_put_io_end(mpd.io_submit.io_end);
2729                         break;
2730                 }
2731
2732                 trace_ext4_da_write_pages(inode, mpd.first_page, mpd.wbc);
2733                 ret = mpage_prepare_extent_to_map(&mpd);
2734                 if (!ret) {
2735                         if (mpd.map.m_len)
2736                                 ret = mpage_map_and_submit_extent(handle, &mpd,
2737                                         &give_up_on_write);
2738                         else {
2739                                 /*
2740                                  * We scanned the whole range (or exhausted
2741                                  * nr_to_write), submitted what was mapped and
2742                                  * didn't find anything needing mapping. We are
2743                                  * done.
2744                                  */
2745                                 done = true;
2746                         }
2747                 }
2748                 ext4_journal_stop(handle);
2749                 /* Submit prepared bio */
2750                 ext4_io_submit(&mpd.io_submit);
2751                 /* Unlock pages we didn't use */
2752                 mpage_release_unused_pages(&mpd, give_up_on_write);
2753                 /* Drop our io_end reference we got from init */
2754                 ext4_put_io_end(mpd.io_submit.io_end);
2755
2756                 if (ret == -ENOSPC && sbi->s_journal) {
2757                         /*
2758                          * Commit the transaction which would
2759                          * free blocks released in the transaction
2760                          * and try again
2761                          */
2762                         jbd2_journal_force_commit_nested(sbi->s_journal);
2763                         ret = 0;
2764                         continue;
2765                 }
2766                 /* Fatal error - ENOMEM, EIO... */
2767                 if (ret)
2768                         break;
2769         }
2770         blk_finish_plug(&plug);
2771         if (!ret && !cycled && wbc->nr_to_write > 0) {
2772                 cycled = 1;
2773                 mpd.last_page = writeback_index - 1;
2774                 mpd.first_page = 0;
2775                 goto retry;
2776         }
2777
2778         /* Update index */
2779         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2780                 /*
2781                  * Set the writeback_index so that range_cyclic
2782                  * mode will write it back later
2783                  */
2784                 mapping->writeback_index = mpd.first_page;
2785
2786 out_writepages:
2787         trace_ext4_writepages_result(inode, wbc, ret,
2788                                      nr_to_write - wbc->nr_to_write);
2789         return ret;
2790 }
2791
2792 static int ext4_nonda_switch(struct super_block *sb)
2793 {
2794         s64 free_clusters, dirty_clusters;
2795         struct ext4_sb_info *sbi = EXT4_SB(sb);
2796
2797         /*
2798          * switch to non delalloc mode if we are running low
2799          * on free block. The free block accounting via percpu
2800          * counters can get slightly wrong with percpu_counter_batch getting
2801          * accumulated on each CPU without updating global counters
2802          * Delalloc need an accurate free block accounting. So switch
2803          * to non delalloc when we are near to error range.
2804          */
2805         free_clusters =
2806                 percpu_counter_read_positive(&sbi->s_freeclusters_counter);
2807         dirty_clusters =
2808                 percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
2809         /*
2810          * Start pushing delalloc when 1/2 of free blocks are dirty.
2811          */
2812         if (dirty_clusters && (free_clusters < 2 * dirty_clusters))
2813                 try_to_writeback_inodes_sb(sb, WB_REASON_FS_FREE_SPACE);
2814
2815         if (2 * free_clusters < 3 * dirty_clusters ||
2816             free_clusters < (dirty_clusters + EXT4_FREECLUSTERS_WATERMARK)) {
2817                 /*
2818                  * free block count is less than 150% of dirty blocks
2819                  * or free blocks is less than watermark
2820                  */
2821                 return 1;
2822         }
2823         return 0;
2824 }
2825
2826 /* We always reserve for an inode update; the superblock could be there too */
2827 static int ext4_da_write_credits(struct inode *inode, loff_t pos, unsigned len)
2828 {
2829         if (likely(ext4_has_feature_large_file(inode->i_sb)))
2830                 return 1;
2831
2832         if (pos + len <= 0x7fffffffULL)
2833                 return 1;
2834
2835         /* We might need to update the superblock to set LARGE_FILE */
2836         return 2;
2837 }
2838
2839 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2840                                loff_t pos, unsigned len, unsigned flags,
2841                                struct page **pagep, void **fsdata)
2842 {
2843         int ret, retries = 0;
2844         struct page *page;
2845         pgoff_t index;
2846         struct inode *inode = mapping->host;
2847         handle_t *handle;
2848
2849         index = pos >> PAGE_SHIFT;
2850
2851         if (ext4_nonda_switch(inode->i_sb)) {
2852                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2853                 return ext4_write_begin(file, mapping, pos,
2854                                         len, flags, pagep, fsdata);
2855         }
2856         *fsdata = (void *)0;
2857         trace_ext4_da_write_begin(inode, pos, len, flags);
2858
2859         if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
2860                 ret = ext4_da_write_inline_data_begin(mapping, inode,
2861                                                       pos, len, flags,
2862                                                       pagep, fsdata);
2863                 if (ret < 0)
2864                         return ret;
2865                 if (ret == 1)
2866                         return 0;
2867         }
2868
2869         /*
2870          * grab_cache_page_write_begin() can take a long time if the
2871          * system is thrashing due to memory pressure, or if the page
2872          * is being written back.  So grab it first before we start
2873          * the transaction handle.  This also allows us to allocate
2874          * the page (if needed) without using GFP_NOFS.
2875          */
2876 retry_grab:
2877         page = grab_cache_page_write_begin(mapping, index, flags);
2878         if (!page)
2879                 return -ENOMEM;
2880         unlock_page(page);
2881
2882         /*
2883          * With delayed allocation, we don't log the i_disksize update
2884          * if there is delayed block allocation. But we still need
2885          * to journalling the i_disksize update if writes to the end
2886          * of file which has an already mapped buffer.
2887          */
2888 retry_journal:
2889         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
2890                                 ext4_da_write_credits(inode, pos, len));
2891         if (IS_ERR(handle)) {
2892                 put_page(page);
2893                 return PTR_ERR(handle);
2894         }
2895
2896         lock_page(page);
2897         if (page->mapping != mapping) {
2898                 /* The page got truncated from under us */
2899                 unlock_page(page);
2900                 put_page(page);
2901                 ext4_journal_stop(handle);
2902                 goto retry_grab;
2903         }
2904         /* In case writeback began while the page was unlocked */
2905         wait_for_stable_page(page);
2906
2907 #ifdef CONFIG_EXT4_FS_ENCRYPTION
2908         ret = ext4_block_write_begin(page, pos, len,
2909                                      ext4_da_get_block_prep);
2910 #else
2911         ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
2912 #endif
2913         if (ret < 0) {
2914                 unlock_page(page);
2915                 ext4_journal_stop(handle);
2916                 /*
2917                  * block_write_begin may have instantiated a few blocks
2918                  * outside i_size.  Trim these off again. Don't need
2919                  * i_size_read because we hold i_mutex.
2920                  */
2921                 if (pos + len > inode->i_size)
2922                         ext4_truncate_failed_write(inode);
2923
2924                 if (ret == -ENOSPC &&
2925                     ext4_should_retry_alloc(inode->i_sb, &retries))
2926                         goto retry_journal;
2927
2928                 put_page(page);
2929                 return ret;
2930         }
2931
2932         *pagep = page;
2933         return ret;
2934 }
2935
2936 /*
2937  * Check if we should update i_disksize
2938  * when write to the end of file but not require block allocation
2939  */
2940 static int ext4_da_should_update_i_disksize(struct page *page,
2941                                             unsigned long offset)
2942 {
2943         struct buffer_head *bh;
2944         struct inode *inode = page->mapping->host;
2945         unsigned int idx;
2946         int i;
2947
2948         bh = page_buffers(page);
2949         idx = offset >> inode->i_blkbits;
2950
2951         for (i = 0; i < idx; i++)
2952                 bh = bh->b_this_page;
2953
2954         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
2955                 return 0;
2956         return 1;
2957 }
2958
2959 static int ext4_da_write_end(struct file *file,
2960                              struct address_space *mapping,
2961                              loff_t pos, unsigned len, unsigned copied,
2962                              struct page *page, void *fsdata)
2963 {
2964         struct inode *inode = mapping->host;
2965         int ret = 0, ret2;
2966         handle_t *handle = ext4_journal_current_handle();
2967         loff_t new_i_size;
2968         unsigned long start, end;
2969         int write_mode = (int)(unsigned long)fsdata;
2970
2971         if (write_mode == FALL_BACK_TO_NONDELALLOC)
2972                 return ext4_write_end(file, mapping, pos,
2973                                       len, copied, page, fsdata);
2974
2975         trace_ext4_da_write_end(inode, pos, len, copied);
2976         start = pos & (PAGE_SIZE - 1);
2977         end = start + copied - 1;
2978
2979         /*
2980          * generic_write_end() will run mark_inode_dirty() if i_size
2981          * changes.  So let's piggyback the i_disksize mark_inode_dirty
2982          * into that.
2983          */
2984         new_i_size = pos + copied;
2985         if (copied && new_i_size > EXT4_I(inode)->i_disksize) {
2986                 if (ext4_has_inline_data(inode) ||
2987                     ext4_da_should_update_i_disksize(page, end)) {
2988                         ext4_update_i_disksize(inode, new_i_size);
2989                         /* We need to mark inode dirty even if
2990                          * new_i_size is less that inode->i_size
2991                          * bu greater than i_disksize.(hint delalloc)
2992                          */
2993                         ext4_mark_inode_dirty(handle, inode);
2994                 }
2995         }
2996
2997         if (write_mode != CONVERT_INLINE_DATA &&
2998             ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA) &&
2999             ext4_has_inline_data(inode))
3000                 ret2 = ext4_da_write_inline_data_end(inode, pos, len, copied,
3001                                                      page);
3002         else
3003                 ret2 = generic_write_end(file, mapping, pos, len, copied,
3004                                                         page, fsdata);
3005
3006         copied = ret2;
3007         if (ret2 < 0)
3008                 ret = ret2;
3009         ret2 = ext4_journal_stop(handle);
3010         if (!ret)
3011                 ret = ret2;
3012
3013         return ret ? ret : copied;
3014 }
3015
3016 static void ext4_da_invalidatepage(struct page *page, unsigned int offset,
3017                                    unsigned int length)
3018 {
3019         /*
3020          * Drop reserved blocks
3021          */
3022         BUG_ON(!PageLocked(page));
3023         if (!page_has_buffers(page))
3024                 goto out;
3025
3026         ext4_da_page_release_reservation(page, offset, length);
3027
3028 out:
3029         ext4_invalidatepage(page, offset, length);
3030
3031         return;
3032 }
3033
3034 /*
3035  * Force all delayed allocation blocks to be allocated for a given inode.
3036  */
3037 int ext4_alloc_da_blocks(struct inode *inode)
3038 {
3039         trace_ext4_alloc_da_blocks(inode);
3040
3041         if (!EXT4_I(inode)->i_reserved_data_blocks)
3042                 return 0;
3043
3044         /*
3045          * We do something simple for now.  The filemap_flush() will
3046          * also start triggering a write of the data blocks, which is
3047          * not strictly speaking necessary (and for users of
3048          * laptop_mode, not even desirable).  However, to do otherwise
3049          * would require replicating code paths in:
3050          *
3051          * ext4_writepages() ->
3052          *    write_cache_pages() ---> (via passed in callback function)
3053          *        __mpage_da_writepage() -->
3054          *           mpage_add_bh_to_extent()
3055          *           mpage_da_map_blocks()
3056          *
3057          * The problem is that write_cache_pages(), located in
3058          * mm/page-writeback.c, marks pages clean in preparation for
3059          * doing I/O, which is not desirable if we're not planning on
3060          * doing I/O at all.
3061          *
3062          * We could call write_cache_pages(), and then redirty all of
3063          * the pages by calling redirty_page_for_writepage() but that
3064          * would be ugly in the extreme.  So instead we would need to
3065          * replicate parts of the code in the above functions,
3066          * simplifying them because we wouldn't actually intend to
3067          * write out the pages, but rather only collect contiguous
3068          * logical block extents, call the multi-block allocator, and
3069          * then update the buffer heads with the block allocations.
3070          *
3071          * For now, though, we'll cheat by calling filemap_flush(),
3072          * which will map the blocks, and start the I/O, but not
3073          * actually wait for the I/O to complete.
3074          */
3075         return filemap_flush(inode->i_mapping);
3076 }
3077
3078 /*
3079  * bmap() is special.  It gets used by applications such as lilo and by
3080  * the swapper to find the on-disk block of a specific piece of data.
3081  *
3082  * Naturally, this is dangerous if the block concerned is still in the
3083  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3084  * filesystem and enables swap, then they may get a nasty shock when the
3085  * data getting swapped to that swapfile suddenly gets overwritten by
3086  * the original zero's written out previously to the journal and
3087  * awaiting writeback in the kernel's buffer cache.
3088  *
3089  * So, if we see any bmap calls here on a modified, data-journaled file,
3090  * take extra steps to flush any blocks which might be in the cache.
3091  */
3092 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3093 {
3094         struct inode *inode = mapping->host;
3095         journal_t *journal;
3096         int err;
3097
3098         /*
3099          * We can get here for an inline file via the FIBMAP ioctl
3100          */
3101         if (ext4_has_inline_data(inode))
3102                 return 0;
3103
3104         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3105                         test_opt(inode->i_sb, DELALLOC)) {
3106                 /*
3107                  * With delalloc we want to sync the file
3108                  * so that we can make sure we allocate
3109                  * blocks for file
3110                  */
3111                 filemap_write_and_wait(mapping);
3112         }
3113
3114         if (EXT4_JOURNAL(inode) &&
3115             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3116                 /*
3117                  * This is a REALLY heavyweight approach, but the use of
3118                  * bmap on dirty files is expected to be extremely rare:
3119                  * only if we run lilo or swapon on a freshly made file
3120                  * do we expect this to happen.
3121                  *
3122                  * (bmap requires CAP_SYS_RAWIO so this does not
3123                  * represent an unprivileged user DOS attack --- we'd be
3124                  * in trouble if mortal users could trigger this path at
3125                  * will.)
3126                  *
3127                  * NB. EXT4_STATE_JDATA is not set on files other than
3128                  * regular files.  If somebody wants to bmap a directory
3129                  * or symlink and gets confused because the buffer
3130                  * hasn't yet been flushed to disk, they deserve
3131                  * everything they get.
3132                  */
3133
3134                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3135                 journal = EXT4_JOURNAL(inode);
3136                 jbd2_journal_lock_updates(journal);