d9733aa955e9d12fcd9d7492d4b6fb67d69f783c
[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 #include <linux/iomap.h>
41
42 #include "ext4_jbd2.h"
43 #include "xattr.h"
44 #include "acl.h"
45 #include "truncate.h"
46
47 #include <trace/events/ext4.h>
48
49 #define MPAGE_DA_EXTENT_TAIL 0x01
50
51 static __u32 ext4_inode_csum(struct inode *inode, struct ext4_inode *raw,
52                               struct ext4_inode_info *ei)
53 {
54         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
55         __u32 csum;
56         __u16 dummy_csum = 0;
57         int offset = offsetof(struct ext4_inode, i_checksum_lo);
58         unsigned int csum_size = sizeof(dummy_csum);
59
60         csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)raw, offset);
61         csum = ext4_chksum(sbi, csum, (__u8 *)&dummy_csum, csum_size);
62         offset += csum_size;
63         csum = ext4_chksum(sbi, csum, (__u8 *)raw + offset,
64                            EXT4_GOOD_OLD_INODE_SIZE - offset);
65
66         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
67                 offset = offsetof(struct ext4_inode, i_checksum_hi);
68                 csum = ext4_chksum(sbi, csum, (__u8 *)raw +
69                                    EXT4_GOOD_OLD_INODE_SIZE,
70                                    offset - EXT4_GOOD_OLD_INODE_SIZE);
71                 if (EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi)) {
72                         csum = ext4_chksum(sbi, csum, (__u8 *)&dummy_csum,
73                                            csum_size);
74                         offset += csum_size;
75                 }
76                 csum = ext4_chksum(sbi, csum, (__u8 *)raw + offset,
77                                    EXT4_INODE_SIZE(inode->i_sb) - offset);
78         }
79
80         return csum;
81 }
82
83 static int ext4_inode_csum_verify(struct inode *inode, struct ext4_inode *raw,
84                                   struct ext4_inode_info *ei)
85 {
86         __u32 provided, calculated;
87
88         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
89             cpu_to_le32(EXT4_OS_LINUX) ||
90             !ext4_has_metadata_csum(inode->i_sb))
91                 return 1;
92
93         provided = le16_to_cpu(raw->i_checksum_lo);
94         calculated = ext4_inode_csum(inode, raw, ei);
95         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
96             EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
97                 provided |= ((__u32)le16_to_cpu(raw->i_checksum_hi)) << 16;
98         else
99                 calculated &= 0xFFFF;
100
101         return provided == calculated;
102 }
103
104 static void ext4_inode_csum_set(struct inode *inode, struct ext4_inode *raw,
105                                 struct ext4_inode_info *ei)
106 {
107         __u32 csum;
108
109         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
110             cpu_to_le32(EXT4_OS_LINUX) ||
111             !ext4_has_metadata_csum(inode->i_sb))
112                 return;
113
114         csum = ext4_inode_csum(inode, raw, ei);
115         raw->i_checksum_lo = cpu_to_le16(csum & 0xFFFF);
116         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
117             EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
118                 raw->i_checksum_hi = cpu_to_le16(csum >> 16);
119 }
120
121 static inline int ext4_begin_ordered_truncate(struct inode *inode,
122                                               loff_t new_size)
123 {
124         trace_ext4_begin_ordered_truncate(inode, new_size);
125         /*
126          * If jinode is zero, then we never opened the file for
127          * writing, so there's no need to call
128          * jbd2_journal_begin_ordered_truncate() since there's no
129          * outstanding writes we need to flush.
130          */
131         if (!EXT4_I(inode)->jinode)
132                 return 0;
133         return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
134                                                    EXT4_I(inode)->jinode,
135                                                    new_size);
136 }
137
138 static void ext4_invalidatepage(struct page *page, unsigned int offset,
139                                 unsigned int length);
140 static int __ext4_journalled_writepage(struct page *page, unsigned int len);
141 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
142 static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
143                                   int pextents);
144
145 /*
146  * Test whether an inode is a fast symlink.
147  */
148 int ext4_inode_is_fast_symlink(struct inode *inode)
149 {
150         int ea_blocks = EXT4_I(inode)->i_file_acl ?
151                 EXT4_CLUSTER_SIZE(inode->i_sb) >> 9 : 0;
152
153         if (ext4_has_inline_data(inode))
154                 return 0;
155
156         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
157 }
158
159 /*
160  * Restart the transaction associated with *handle.  This does a commit,
161  * so before we call here everything must be consistently dirtied against
162  * this transaction.
163  */
164 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
165                                  int nblocks)
166 {
167         int ret;
168
169         /*
170          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
171          * moment, get_block can be called only for blocks inside i_size since
172          * page cache has been already dropped and writes are blocked by
173          * i_mutex. So we can safely drop the i_data_sem here.
174          */
175         BUG_ON(EXT4_JOURNAL(inode) == NULL);
176         jbd_debug(2, "restarting handle %p\n", handle);
177         up_write(&EXT4_I(inode)->i_data_sem);
178         ret = ext4_journal_restart(handle, nblocks);
179         down_write(&EXT4_I(inode)->i_data_sem);
180         ext4_discard_preallocations(inode);
181
182         return ret;
183 }
184
185 /*
186  * Called at the last iput() if i_nlink is zero.
187  */
188 void ext4_evict_inode(struct inode *inode)
189 {
190         handle_t *handle;
191         int err;
192         int extra_credits = 3;
193         struct ext4_xattr_inode_array *ea_inode_array = NULL;
194
195         trace_ext4_evict_inode(inode);
196
197         if (inode->i_nlink) {
198                 /*
199                  * When journalling data dirty buffers are tracked only in the
200                  * journal. So although mm thinks everything is clean and
201                  * ready for reaping the inode might still have some pages to
202                  * write in the running transaction or waiting to be
203                  * checkpointed. Thus calling jbd2_journal_invalidatepage()
204                  * (via truncate_inode_pages()) to discard these buffers can
205                  * cause data loss. Also even if we did not discard these
206                  * buffers, we would have no way to find them after the inode
207                  * is reaped and thus user could see stale data if he tries to
208                  * read them before the transaction is checkpointed. So be
209                  * careful and force everything to disk here... We use
210                  * ei->i_datasync_tid to store the newest transaction
211                  * containing inode's data.
212                  *
213                  * Note that directories do not have this problem because they
214                  * don't use page cache.
215                  */
216                 if (inode->i_ino != EXT4_JOURNAL_INO &&
217                     ext4_should_journal_data(inode) &&
218                     (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode))) {
219                         journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
220                         tid_t commit_tid = EXT4_I(inode)->i_datasync_tid;
221
222                         jbd2_complete_transaction(journal, commit_tid);
223                         filemap_write_and_wait(&inode->i_data);
224                 }
225                 truncate_inode_pages_final(&inode->i_data);
226
227                 goto no_delete;
228         }
229
230         if (is_bad_inode(inode))
231                 goto no_delete;
232         dquot_initialize(inode);
233
234         if (ext4_should_order_data(inode))
235                 ext4_begin_ordered_truncate(inode, 0);
236         truncate_inode_pages_final(&inode->i_data);
237
238         /*
239          * Protect us against freezing - iput() caller didn't have to have any
240          * protection against it
241          */
242         sb_start_intwrite(inode->i_sb);
243
244         if (!IS_NOQUOTA(inode))
245                 extra_credits += EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb);
246
247         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE,
248                                  ext4_blocks_for_truncate(inode)+extra_credits);
249         if (IS_ERR(handle)) {
250                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
251                 /*
252                  * If we're going to skip the normal cleanup, we still need to
253                  * make sure that the in-core orphan linked list is properly
254                  * cleaned up.
255                  */
256                 ext4_orphan_del(NULL, inode);
257                 sb_end_intwrite(inode->i_sb);
258                 goto no_delete;
259         }
260
261         if (IS_SYNC(inode))
262                 ext4_handle_sync(handle);
263         inode->i_size = 0;
264         err = ext4_mark_inode_dirty(handle, inode);
265         if (err) {
266                 ext4_warning(inode->i_sb,
267                              "couldn't mark inode dirty (err %d)", err);
268                 goto stop_handle;
269         }
270         if (inode->i_blocks) {
271                 err = ext4_truncate(inode);
272                 if (err) {
273                         ext4_error(inode->i_sb,
274                                    "couldn't truncate inode %lu (err %d)",
275                                    inode->i_ino, err);
276                         goto stop_handle;
277                 }
278         }
279
280         /* Remove xattr references. */
281         err = ext4_xattr_delete_inode(handle, inode, &ea_inode_array,
282                                       extra_credits);
283         if (err) {
284                 ext4_warning(inode->i_sb, "xattr delete (err %d)", err);
285 stop_handle:
286                 ext4_journal_stop(handle);
287                 ext4_orphan_del(NULL, inode);
288                 sb_end_intwrite(inode->i_sb);
289                 ext4_xattr_inode_array_free(ea_inode_array);
290                 goto no_delete;
291         }
292
293         /*
294          * Kill off the orphan record which ext4_truncate created.
295          * AKPM: I think this can be inside the above `if'.
296          * Note that ext4_orphan_del() has to be able to cope with the
297          * deletion of a non-existent orphan - this is because we don't
298          * know if ext4_truncate() actually created an orphan record.
299          * (Well, we could do this if we need to, but heck - it works)
300          */
301         ext4_orphan_del(handle, inode);
302         EXT4_I(inode)->i_dtime  = get_seconds();
303
304         /*
305          * One subtle ordering requirement: if anything has gone wrong
306          * (transaction abort, IO errors, whatever), then we can still
307          * do these next steps (the fs will already have been marked as
308          * having errors), but we can't free the inode if the mark_dirty
309          * fails.
310          */
311         if (ext4_mark_inode_dirty(handle, inode))
312                 /* If that failed, just do the required in-core inode clear. */
313                 ext4_clear_inode(inode);
314         else
315                 ext4_free_inode(handle, inode);
316         ext4_journal_stop(handle);
317         sb_end_intwrite(inode->i_sb);
318         ext4_xattr_inode_array_free(ea_inode_array);
319         return;
320 no_delete:
321         ext4_clear_inode(inode);        /* We must guarantee clearing of inode... */
322 }
323
324 #ifdef CONFIG_QUOTA
325 qsize_t *ext4_get_reserved_space(struct inode *inode)
326 {
327         return &EXT4_I(inode)->i_reserved_quota;
328 }
329 #endif
330
331 /*
332  * Called with i_data_sem down, which is important since we can call
333  * ext4_discard_preallocations() from here.
334  */
335 void ext4_da_update_reserve_space(struct inode *inode,
336                                         int used, int quota_claim)
337 {
338         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
339         struct ext4_inode_info *ei = EXT4_I(inode);
340
341         spin_lock(&ei->i_block_reservation_lock);
342         trace_ext4_da_update_reserve_space(inode, used, quota_claim);
343         if (unlikely(used > ei->i_reserved_data_blocks)) {
344                 ext4_warning(inode->i_sb, "%s: ino %lu, used %d "
345                          "with only %d reserved data blocks",
346                          __func__, inode->i_ino, used,
347                          ei->i_reserved_data_blocks);
348                 WARN_ON(1);
349                 used = ei->i_reserved_data_blocks;
350         }
351
352         /* Update per-inode reservations */
353         ei->i_reserved_data_blocks -= used;
354         percpu_counter_sub(&sbi->s_dirtyclusters_counter, used);
355
356         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
357
358         /* Update quota subsystem for data blocks */
359         if (quota_claim)
360                 dquot_claim_block(inode, EXT4_C2B(sbi, used));
361         else {
362                 /*
363                  * We did fallocate with an offset that is already delayed
364                  * allocated. So on delayed allocated writeback we should
365                  * not re-claim the quota for fallocated blocks.
366                  */
367                 dquot_release_reservation_block(inode, EXT4_C2B(sbi, used));
368         }
369
370         /*
371          * If we have done all the pending block allocations and if
372          * there aren't any writers on the inode, we can discard the
373          * inode's preallocations.
374          */
375         if ((ei->i_reserved_data_blocks == 0) &&
376             (atomic_read(&inode->i_writecount) == 0))
377                 ext4_discard_preallocations(inode);
378 }
379
380 static int __check_block_validity(struct inode *inode, const char *func,
381                                 unsigned int line,
382                                 struct ext4_map_blocks *map)
383 {
384         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
385                                    map->m_len)) {
386                 ext4_error_inode(inode, func, line, map->m_pblk,
387                                  "lblock %lu mapped to illegal pblock "
388                                  "(length %d)", (unsigned long) map->m_lblk,
389                                  map->m_len);
390                 return -EFSCORRUPTED;
391         }
392         return 0;
393 }
394
395 int ext4_issue_zeroout(struct inode *inode, ext4_lblk_t lblk, ext4_fsblk_t pblk,
396                        ext4_lblk_t len)
397 {
398         int ret;
399
400         if (ext4_encrypted_inode(inode))
401                 return fscrypt_zeroout_range(inode, lblk, pblk, len);
402
403         ret = sb_issue_zeroout(inode->i_sb, pblk, len, GFP_NOFS);
404         if (ret > 0)
405                 ret = 0;
406
407         return ret;
408 }
409
410 #define check_block_validity(inode, map)        \
411         __check_block_validity((inode), __func__, __LINE__, (map))
412
413 #ifdef ES_AGGRESSIVE_TEST
414 static void ext4_map_blocks_es_recheck(handle_t *handle,
415                                        struct inode *inode,
416                                        struct ext4_map_blocks *es_map,
417                                        struct ext4_map_blocks *map,
418                                        int flags)
419 {
420         int retval;
421
422         map->m_flags = 0;
423         /*
424          * There is a race window that the result is not the same.
425          * e.g. xfstests #223 when dioread_nolock enables.  The reason
426          * is that we lookup a block mapping in extent status tree with
427          * out taking i_data_sem.  So at the time the unwritten extent
428          * could be converted.
429          */
430         down_read(&EXT4_I(inode)->i_data_sem);
431         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
432                 retval = ext4_ext_map_blocks(handle, inode, map, flags &
433                                              EXT4_GET_BLOCKS_KEEP_SIZE);
434         } else {
435                 retval = ext4_ind_map_blocks(handle, inode, map, flags &
436                                              EXT4_GET_BLOCKS_KEEP_SIZE);
437         }
438         up_read((&EXT4_I(inode)->i_data_sem));
439
440         /*
441          * We don't check m_len because extent will be collpased in status
442          * tree.  So the m_len might not equal.
443          */
444         if (es_map->m_lblk != map->m_lblk ||
445             es_map->m_flags != map->m_flags ||
446             es_map->m_pblk != map->m_pblk) {
447                 printk("ES cache assertion failed for inode: %lu "
448                        "es_cached ex [%d/%d/%llu/%x] != "
449                        "found ex [%d/%d/%llu/%x] retval %d flags %x\n",
450                        inode->i_ino, es_map->m_lblk, es_map->m_len,
451                        es_map->m_pblk, es_map->m_flags, map->m_lblk,
452                        map->m_len, map->m_pblk, map->m_flags,
453                        retval, flags);
454         }
455 }
456 #endif /* ES_AGGRESSIVE_TEST */
457
458 /*
459  * The ext4_map_blocks() function tries to look up the requested blocks,
460  * and returns if the blocks are already mapped.
461  *
462  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
463  * and store the allocated blocks in the result buffer head and mark it
464  * mapped.
465  *
466  * If file type is extents based, it will call ext4_ext_map_blocks(),
467  * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
468  * based files
469  *
470  * On success, it returns the number of blocks being mapped or allocated.  if
471  * create==0 and the blocks are pre-allocated and unwritten, the resulting @map
472  * is marked as unwritten. If the create == 1, it will mark @map as mapped.
473  *
474  * It returns 0 if plain look up failed (blocks have not been allocated), in
475  * that case, @map is returned as unmapped but we still do fill map->m_len to
476  * indicate the length of a hole starting at map->m_lblk.
477  *
478  * It returns the error in case of allocation failure.
479  */
480 int ext4_map_blocks(handle_t *handle, struct inode *inode,
481                     struct ext4_map_blocks *map, int flags)
482 {
483         struct extent_status es;
484         int retval;
485         int ret = 0;
486 #ifdef ES_AGGRESSIVE_TEST
487         struct ext4_map_blocks orig_map;
488
489         memcpy(&orig_map, map, sizeof(*map));
490 #endif
491
492         map->m_flags = 0;
493         ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
494                   "logical block %lu\n", inode->i_ino, flags, map->m_len,
495                   (unsigned long) map->m_lblk);
496
497         /*
498          * ext4_map_blocks returns an int, and m_len is an unsigned int
499          */
500         if (unlikely(map->m_len > INT_MAX))
501                 map->m_len = INT_MAX;
502
503         /* We can handle the block number less than EXT_MAX_BLOCKS */
504         if (unlikely(map->m_lblk >= EXT_MAX_BLOCKS))
505                 return -EFSCORRUPTED;
506
507         /* Lookup extent status tree firstly */
508         if (ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
509                 if (ext4_es_is_written(&es) || ext4_es_is_unwritten(&es)) {
510                         map->m_pblk = ext4_es_pblock(&es) +
511                                         map->m_lblk - es.es_lblk;
512                         map->m_flags |= ext4_es_is_written(&es) ?
513                                         EXT4_MAP_MAPPED : EXT4_MAP_UNWRITTEN;
514                         retval = es.es_len - (map->m_lblk - es.es_lblk);
515                         if (retval > map->m_len)
516                                 retval = map->m_len;
517                         map->m_len = retval;
518                 } else if (ext4_es_is_delayed(&es) || ext4_es_is_hole(&es)) {
519                         map->m_pblk = 0;
520                         retval = es.es_len - (map->m_lblk - es.es_lblk);
521                         if (retval > map->m_len)
522                                 retval = map->m_len;
523                         map->m_len = retval;
524                         retval = 0;
525                 } else {
526                         BUG_ON(1);
527                 }
528 #ifdef ES_AGGRESSIVE_TEST
529                 ext4_map_blocks_es_recheck(handle, inode, map,
530                                            &orig_map, flags);
531 #endif
532                 goto found;
533         }
534
535         /*
536          * Try to see if we can get the block without requesting a new
537          * file system block.
538          */
539         down_read(&EXT4_I(inode)->i_data_sem);
540         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
541                 retval = ext4_ext_map_blocks(handle, inode, map, flags &
542                                              EXT4_GET_BLOCKS_KEEP_SIZE);
543         } else {
544                 retval = ext4_ind_map_blocks(handle, inode, map, flags &
545                                              EXT4_GET_BLOCKS_KEEP_SIZE);
546         }
547         if (retval > 0) {
548                 unsigned int status;
549
550                 if (unlikely(retval != map->m_len)) {
551                         ext4_warning(inode->i_sb,
552                                      "ES len assertion failed for inode "
553                                      "%lu: retval %d != map->m_len %d",
554                                      inode->i_ino, retval, map->m_len);
555                         WARN_ON(1);
556                 }
557
558                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
559                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
560                 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
561                     !(status & EXTENT_STATUS_WRITTEN) &&
562                     ext4_find_delalloc_range(inode, map->m_lblk,
563                                              map->m_lblk + map->m_len - 1))
564                         status |= EXTENT_STATUS_DELAYED;
565                 ret = ext4_es_insert_extent(inode, map->m_lblk,
566                                             map->m_len, map->m_pblk, status);
567                 if (ret < 0)
568                         retval = ret;
569         }
570         up_read((&EXT4_I(inode)->i_data_sem));
571
572 found:
573         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
574                 ret = check_block_validity(inode, map);
575                 if (ret != 0)
576                         return ret;
577         }
578
579         /* If it is only a block(s) look up */
580         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
581                 return retval;
582
583         /*
584          * Returns if the blocks have already allocated
585          *
586          * Note that if blocks have been preallocated
587          * ext4_ext_get_block() returns the create = 0
588          * with buffer head unmapped.
589          */
590         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
591                 /*
592                  * If we need to convert extent to unwritten
593                  * we continue and do the actual work in
594                  * ext4_ext_map_blocks()
595                  */
596                 if (!(flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN))
597                         return retval;
598
599         /*
600          * Here we clear m_flags because after allocating an new extent,
601          * it will be set again.
602          */
603         map->m_flags &= ~EXT4_MAP_FLAGS;
604
605         /*
606          * New blocks allocate and/or writing to unwritten extent
607          * will possibly result in updating i_data, so we take
608          * the write lock of i_data_sem, and call get_block()
609          * with create == 1 flag.
610          */
611         down_write(&EXT4_I(inode)->i_data_sem);
612
613         /*
614          * We need to check for EXT4 here because migrate
615          * could have changed the inode type in between
616          */
617         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
618                 retval = ext4_ext_map_blocks(handle, inode, map, flags);
619         } else {
620                 retval = ext4_ind_map_blocks(handle, inode, map, flags);
621
622                 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
623                         /*
624                          * We allocated new blocks which will result in
625                          * i_data's format changing.  Force the migrate
626                          * to fail by clearing migrate flags
627                          */
628                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
629                 }
630
631                 /*
632                  * Update reserved blocks/metadata blocks after successful
633                  * block allocation which had been deferred till now. We don't
634                  * support fallocate for non extent files. So we can update
635                  * reserve space here.
636                  */
637                 if ((retval > 0) &&
638                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
639                         ext4_da_update_reserve_space(inode, retval, 1);
640         }
641
642         if (retval > 0) {
643                 unsigned int status;
644
645                 if (unlikely(retval != map->m_len)) {
646                         ext4_warning(inode->i_sb,
647                                      "ES len assertion failed for inode "
648                                      "%lu: retval %d != map->m_len %d",
649                                      inode->i_ino, retval, map->m_len);
650                         WARN_ON(1);
651                 }
652
653                 /*
654                  * We have to zeroout blocks before inserting them into extent
655                  * status tree. Otherwise someone could look them up there and
656                  * use them before they are really zeroed. We also have to
657                  * unmap metadata before zeroing as otherwise writeback can
658                  * overwrite zeros with stale data from block device.
659                  */
660                 if (flags & EXT4_GET_BLOCKS_ZERO &&
661                     map->m_flags & EXT4_MAP_MAPPED &&
662                     map->m_flags & EXT4_MAP_NEW) {
663                         clean_bdev_aliases(inode->i_sb->s_bdev, map->m_pblk,
664                                            map->m_len);
665                         ret = ext4_issue_zeroout(inode, map->m_lblk,
666                                                  map->m_pblk, map->m_len);
667                         if (ret) {
668                                 retval = ret;
669                                 goto out_sem;
670                         }
671                 }
672
673                 /*
674                  * If the extent has been zeroed out, we don't need to update
675                  * extent status tree.
676                  */
677                 if ((flags & EXT4_GET_BLOCKS_PRE_IO) &&
678                     ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
679                         if (ext4_es_is_written(&es))
680                                 goto out_sem;
681                 }
682                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
683                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
684                 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
685                     !(status & EXTENT_STATUS_WRITTEN) &&
686                     ext4_find_delalloc_range(inode, map->m_lblk,
687                                              map->m_lblk + map->m_len - 1))
688                         status |= EXTENT_STATUS_DELAYED;
689                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
690                                             map->m_pblk, status);
691                 if (ret < 0) {
692                         retval = ret;
693                         goto out_sem;
694                 }
695         }
696
697 out_sem:
698         up_write((&EXT4_I(inode)->i_data_sem));
699         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
700                 ret = check_block_validity(inode, map);
701                 if (ret != 0)
702                         return ret;
703
704                 /*
705                  * Inodes with freshly allocated blocks where contents will be
706                  * visible after transaction commit must be on transaction's
707                  * ordered data list.
708                  */
709                 if (map->m_flags & EXT4_MAP_NEW &&
710                     !(map->m_flags & EXT4_MAP_UNWRITTEN) &&
711                     !(flags & EXT4_GET_BLOCKS_ZERO) &&
712                     !ext4_is_quota_file(inode) &&
713                     ext4_should_order_data(inode)) {
714                         if (flags & EXT4_GET_BLOCKS_IO_SUBMIT)
715                                 ret = ext4_jbd2_inode_add_wait(handle, inode);
716                         else
717                                 ret = ext4_jbd2_inode_add_write(handle, inode);
718                         if (ret)
719                                 return ret;
720                 }
721         }
722         return retval;
723 }
724
725 /*
726  * Update EXT4_MAP_FLAGS in bh->b_state. For buffer heads attached to pages
727  * we have to be careful as someone else may be manipulating b_state as well.
728  */
729 static void ext4_update_bh_state(struct buffer_head *bh, unsigned long flags)
730 {
731         unsigned long old_state;
732         unsigned long new_state;
733
734         flags &= EXT4_MAP_FLAGS;
735
736         /* Dummy buffer_head? Set non-atomically. */
737         if (!bh->b_page) {
738                 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | flags;
739                 return;
740         }
741         /*
742          * Someone else may be modifying b_state. Be careful! This is ugly but
743          * once we get rid of using bh as a container for mapping information
744          * to pass to / from get_block functions, this can go away.
745          */
746         do {
747                 old_state = READ_ONCE(bh->b_state);
748                 new_state = (old_state & ~EXT4_MAP_FLAGS) | flags;
749         } while (unlikely(
750                  cmpxchg(&bh->b_state, old_state, new_state) != old_state));
751 }
752
753 static int _ext4_get_block(struct inode *inode, sector_t iblock,
754                            struct buffer_head *bh, int flags)
755 {
756         struct ext4_map_blocks map;
757         int ret = 0;
758
759         if (ext4_has_inline_data(inode))
760                 return -ERANGE;
761
762         map.m_lblk = iblock;
763         map.m_len = bh->b_size >> inode->i_blkbits;
764
765         ret = ext4_map_blocks(ext4_journal_current_handle(), inode, &map,
766                               flags);
767         if (ret > 0) {
768                 map_bh(bh, inode->i_sb, map.m_pblk);
769                 ext4_update_bh_state(bh, map.m_flags);
770                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
771                 ret = 0;
772         } else if (ret == 0) {
773                 /* hole case, need to fill in bh->b_size */
774                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
775         }
776         return ret;
777 }
778
779 int ext4_get_block(struct inode *inode, sector_t iblock,
780                    struct buffer_head *bh, int create)
781 {
782         return _ext4_get_block(inode, iblock, bh,
783                                create ? EXT4_GET_BLOCKS_CREATE : 0);
784 }
785
786 /*
787  * Get block function used when preparing for buffered write if we require
788  * creating an unwritten extent if blocks haven't been allocated.  The extent
789  * will be converted to written after the IO is complete.
790  */
791 int ext4_get_block_unwritten(struct inode *inode, sector_t iblock,
792                              struct buffer_head *bh_result, int create)
793 {
794         ext4_debug("ext4_get_block_unwritten: inode %lu, create flag %d\n",
795                    inode->i_ino, create);
796         return _ext4_get_block(inode, iblock, bh_result,
797                                EXT4_GET_BLOCKS_IO_CREATE_EXT);
798 }
799
800 /* Maximum number of blocks we map for direct IO at once. */
801 #define DIO_MAX_BLOCKS 4096
802
803 /*
804  * Get blocks function for the cases that need to start a transaction -
805  * generally difference cases of direct IO and DAX IO. It also handles retries
806  * in case of ENOSPC.
807  */
808 static int ext4_get_block_trans(struct inode *inode, sector_t iblock,
809                                 struct buffer_head *bh_result, int flags)
810 {
811         int dio_credits;
812         handle_t *handle;
813         int retries = 0;
814         int ret;
815
816         /* Trim mapping request to maximum we can map at once for DIO */
817         if (bh_result->b_size >> inode->i_blkbits > DIO_MAX_BLOCKS)
818                 bh_result->b_size = DIO_MAX_BLOCKS << inode->i_blkbits;
819         dio_credits = ext4_chunk_trans_blocks(inode,
820                                       bh_result->b_size >> inode->i_blkbits);
821 retry:
822         handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, dio_credits);
823         if (IS_ERR(handle))
824                 return PTR_ERR(handle);
825
826         ret = _ext4_get_block(inode, iblock, bh_result, flags);
827         ext4_journal_stop(handle);
828
829         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
830                 goto retry;
831         return ret;
832 }
833
834 /* Get block function for DIO reads and writes to inodes without extents */
835 int ext4_dio_get_block(struct inode *inode, sector_t iblock,
836                        struct buffer_head *bh, int create)
837 {
838         /* We don't expect handle for direct IO */
839         WARN_ON_ONCE(ext4_journal_current_handle());
840
841         if (!create)
842                 return _ext4_get_block(inode, iblock, bh, 0);
843         return ext4_get_block_trans(inode, iblock, bh, EXT4_GET_BLOCKS_CREATE);
844 }
845
846 /*
847  * Get block function for AIO DIO writes when we create unwritten extent if
848  * blocks are not allocated yet. The extent will be converted to written
849  * after IO is complete.
850  */
851 static int ext4_dio_get_block_unwritten_async(struct inode *inode,
852                 sector_t iblock, struct buffer_head *bh_result, int create)
853 {
854         int ret;
855
856         /* We don't expect handle for direct IO */
857         WARN_ON_ONCE(ext4_journal_current_handle());
858
859         ret = ext4_get_block_trans(inode, iblock, bh_result,
860                                    EXT4_GET_BLOCKS_IO_CREATE_EXT);
861
862         /*
863          * When doing DIO using unwritten extents, we need io_end to convert
864          * unwritten extents to written on IO completion. We allocate io_end
865          * once we spot unwritten extent and store it in b_private. Generic
866          * DIO code keeps b_private set and furthermore passes the value to
867          * our completion callback in 'private' argument.
868          */
869         if (!ret && buffer_unwritten(bh_result)) {
870                 if (!bh_result->b_private) {
871                         ext4_io_end_t *io_end;
872
873                         io_end = ext4_init_io_end(inode, GFP_KERNEL);
874                         if (!io_end)
875                                 return -ENOMEM;
876                         bh_result->b_private = io_end;
877                         ext4_set_io_unwritten_flag(inode, io_end);
878                 }
879                 set_buffer_defer_completion(bh_result);
880         }
881
882         return ret;
883 }
884
885 /*
886  * Get block function for non-AIO DIO writes when we create unwritten extent if
887  * blocks are not allocated yet. The extent will be converted to written
888  * after IO is complete from ext4_ext_direct_IO() function.
889  */
890 static int ext4_dio_get_block_unwritten_sync(struct inode *inode,
891                 sector_t iblock, struct buffer_head *bh_result, int create)
892 {
893         int ret;
894
895         /* We don't expect handle for direct IO */
896         WARN_ON_ONCE(ext4_journal_current_handle());
897
898         ret = ext4_get_block_trans(inode, iblock, bh_result,
899                                    EXT4_GET_BLOCKS_IO_CREATE_EXT);
900
901         /*
902          * Mark inode as having pending DIO writes to unwritten extents.
903          * ext4_ext_direct_IO() checks this flag and converts extents to
904          * written.
905          */
906         if (!ret && buffer_unwritten(bh_result))
907                 ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
908
909         return ret;
910 }
911
912 static int ext4_dio_get_block_overwrite(struct inode *inode, sector_t iblock,
913                    struct buffer_head *bh_result, int create)
914 {
915         int ret;
916
917         ext4_debug("ext4_dio_get_block_overwrite: inode %lu, create flag %d\n",
918                    inode->i_ino, create);
919         /* We don't expect handle for direct IO */
920         WARN_ON_ONCE(ext4_journal_current_handle());
921
922         ret = _ext4_get_block(inode, iblock, bh_result, 0);
923         /*
924          * Blocks should have been preallocated! ext4_file_write_iter() checks
925          * that.
926          */
927         WARN_ON_ONCE(!buffer_mapped(bh_result) || buffer_unwritten(bh_result));
928
929         return ret;
930 }
931
932
933 /*
934  * `handle' can be NULL if create is zero
935  */
936 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
937                                 ext4_lblk_t block, int map_flags)
938 {
939         struct ext4_map_blocks map;
940         struct buffer_head *bh;
941         int create = map_flags & EXT4_GET_BLOCKS_CREATE;
942         int err;
943
944         J_ASSERT(handle != NULL || create == 0);
945
946         map.m_lblk = block;
947         map.m_len = 1;
948         err = ext4_map_blocks(handle, inode, &map, map_flags);
949
950         if (err == 0)
951                 return create ? ERR_PTR(-ENOSPC) : NULL;
952         if (err < 0)
953                 return ERR_PTR(err);
954
955         bh = sb_getblk(inode->i_sb, map.m_pblk);
956         if (unlikely(!bh))
957                 return ERR_PTR(-ENOMEM);
958         if (map.m_flags & EXT4_MAP_NEW) {
959                 J_ASSERT(create != 0);
960                 J_ASSERT(handle != NULL);
961
962                 /*
963                  * Now that we do not always journal data, we should
964                  * keep in mind whether this should always journal the
965                  * new buffer as metadata.  For now, regular file
966                  * writes use ext4_get_block instead, so it's not a
967                  * problem.
968                  */
969                 lock_buffer(bh);
970                 BUFFER_TRACE(bh, "call get_create_access");
971                 err = ext4_journal_get_create_access(handle, bh);
972                 if (unlikely(err)) {
973                         unlock_buffer(bh);
974                         goto errout;
975                 }
976                 if (!buffer_uptodate(bh)) {
977                         memset(bh->b_data, 0, inode->i_sb->s_blocksize);
978                         set_buffer_uptodate(bh);
979                 }
980                 unlock_buffer(bh);
981                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
982                 err = ext4_handle_dirty_metadata(handle, inode, bh);
983                 if (unlikely(err))
984                         goto errout;
985         } else
986                 BUFFER_TRACE(bh, "not a new buffer");
987         return bh;
988 errout:
989         brelse(bh);
990         return ERR_PTR(err);
991 }
992
993 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
994                                ext4_lblk_t block, int map_flags)
995 {
996         struct buffer_head *bh;
997
998         bh = ext4_getblk(handle, inode, block, map_flags);
999         if (IS_ERR(bh))
1000                 return bh;
1001         if (!bh || buffer_uptodate(bh))
1002                 return bh;
1003         ll_rw_block(REQ_OP_READ, REQ_META | REQ_PRIO, 1, &bh);
1004         wait_on_buffer(bh);
1005         if (buffer_uptodate(bh))
1006                 return bh;
1007         put_bh(bh);
1008         return ERR_PTR(-EIO);
1009 }
1010
1011 int ext4_walk_page_buffers(handle_t *handle,
1012                            struct buffer_head *head,
1013                            unsigned from,
1014                            unsigned to,
1015                            int *partial,
1016                            int (*fn)(handle_t *handle,
1017                                      struct buffer_head *bh))
1018 {
1019         struct buffer_head *bh;
1020         unsigned block_start, block_end;
1021         unsigned blocksize = head->b_size;
1022         int err, ret = 0;
1023         struct buffer_head *next;
1024
1025         for (bh = head, block_start = 0;
1026              ret == 0 && (bh != head || !block_start);
1027              block_start = block_end, bh = next) {
1028                 next = bh->b_this_page;
1029                 block_end = block_start + blocksize;
1030                 if (block_end <= from || block_start >= to) {
1031                         if (partial && !buffer_uptodate(bh))
1032                                 *partial = 1;
1033                         continue;
1034                 }
1035                 err = (*fn)(handle, bh);
1036                 if (!ret)
1037                         ret = err;
1038         }
1039         return ret;
1040 }
1041
1042 /*
1043  * To preserve ordering, it is essential that the hole instantiation and
1044  * the data write be encapsulated in a single transaction.  We cannot
1045  * close off a transaction and start a new one between the ext4_get_block()
1046  * and the commit_write().  So doing the jbd2_journal_start at the start of
1047  * prepare_write() is the right place.
1048  *
1049  * Also, this function can nest inside ext4_writepage().  In that case, we
1050  * *know* that ext4_writepage() has generated enough buffer credits to do the
1051  * whole page.  So we won't block on the journal in that case, which is good,
1052  * because the caller may be PF_MEMALLOC.
1053  *
1054  * By accident, ext4 can be reentered when a transaction is open via
1055  * quota file writes.  If we were to commit the transaction while thus
1056  * reentered, there can be a deadlock - we would be holding a quota
1057  * lock, and the commit would never complete if another thread had a
1058  * transaction open and was blocking on the quota lock - a ranking
1059  * violation.
1060  *
1061  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1062  * will _not_ run commit under these circumstances because handle->h_ref
1063  * is elevated.  We'll still have enough credits for the tiny quotafile
1064  * write.
1065  */
1066 int do_journal_get_write_access(handle_t *handle,
1067                                 struct buffer_head *bh)
1068 {
1069         int dirty = buffer_dirty(bh);
1070         int ret;
1071
1072         if (!buffer_mapped(bh) || buffer_freed(bh))
1073                 return 0;
1074         /*
1075          * __block_write_begin() could have dirtied some buffers. Clean
1076          * the dirty bit as jbd2_journal_get_write_access() could complain
1077          * otherwise about fs integrity issues. Setting of the dirty bit
1078          * by __block_write_begin() isn't a real problem here as we clear
1079          * the bit before releasing a page lock and thus writeback cannot
1080          * ever write the buffer.
1081          */
1082         if (dirty)
1083                 clear_buffer_dirty(bh);
1084         BUFFER_TRACE(bh, "get write access");
1085         ret = ext4_journal_get_write_access(handle, bh);
1086         if (!ret && dirty)
1087                 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1088         return ret;
1089 }
1090
1091 #ifdef CONFIG_EXT4_FS_ENCRYPTION
1092 static int ext4_block_write_begin(struct page *page, loff_t pos, unsigned len,
1093                                   get_block_t *get_block)
1094 {
1095         unsigned from = pos & (PAGE_SIZE - 1);
1096         unsigned to = from + len;
1097         struct inode *inode = page->mapping->host;
1098         unsigned block_start, block_end;
1099         sector_t block;
1100         int err = 0;
1101         unsigned blocksize = inode->i_sb->s_blocksize;
1102         unsigned bbits;
1103         struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
1104         bool decrypt = false;
1105
1106         BUG_ON(!PageLocked(page));
1107         BUG_ON(from > PAGE_SIZE);
1108         BUG_ON(to > PAGE_SIZE);
1109         BUG_ON(from > to);
1110
1111         if (!page_has_buffers(page))
1112                 create_empty_buffers(page, blocksize, 0);
1113         head = page_buffers(page);
1114         bbits = ilog2(blocksize);
1115         block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1116
1117         for (bh = head, block_start = 0; bh != head || !block_start;
1118             block++, block_start = block_end, bh = bh->b_this_page) {
1119                 block_end = block_start + blocksize;
1120                 if (block_end <= from || block_start >= to) {
1121                         if (PageUptodate(page)) {
1122                                 if (!buffer_uptodate(bh))
1123                                         set_buffer_uptodate(bh);
1124                         }
1125                         continue;
1126                 }
1127                 if (buffer_new(bh))
1128                         clear_buffer_new(bh);
1129                 if (!buffer_mapped(bh)) {
1130                         WARN_ON(bh->b_size != blocksize);
1131                         err = get_block(inode, block, bh, 1);
1132                         if (err)
1133                                 break;
1134                         if (buffer_new(bh)) {
1135                                 clean_bdev_bh_alias(bh);
1136                                 if (PageUptodate(page)) {
1137                                         clear_buffer_new(bh);
1138                                         set_buffer_uptodate(bh);
1139                                         mark_buffer_dirty(bh);
1140                                         continue;
1141                                 }
1142                                 if (block_end > to || block_start < from)
1143                                         zero_user_segments(page, to, block_end,
1144                                                            block_start, from);
1145                                 continue;
1146                         }
1147                 }
1148                 if (PageUptodate(page)) {
1149                         if (!buffer_uptodate(bh))
1150                                 set_buffer_uptodate(bh);
1151                         continue;
1152                 }
1153                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1154                     !buffer_unwritten(bh) &&
1155                     (block_start < from || block_end > to)) {
1156                         ll_rw_block(REQ_OP_READ, 0, 1, &bh);
1157                         *wait_bh++ = bh;
1158                         decrypt = ext4_encrypted_inode(inode) &&
1159                                 S_ISREG(inode->i_mode);
1160                 }
1161         }
1162         /*
1163          * If we issued read requests, let them complete.
1164          */
1165         while (wait_bh > wait) {
1166                 wait_on_buffer(*--wait_bh);
1167                 if (!buffer_uptodate(*wait_bh))
1168                         err = -EIO;
1169         }
1170         if (unlikely(err))
1171                 page_zero_new_buffers(page, from, to);
1172         else if (decrypt)
1173                 err = fscrypt_decrypt_page(page->mapping->host, page,
1174                                 PAGE_SIZE, 0, page->index);
1175         return err;
1176 }
1177 #endif
1178
1179 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1180                             loff_t pos, unsigned len, unsigned flags,
1181                             struct page **pagep, void **fsdata)
1182 {
1183         struct inode *inode = mapping->host;
1184         int ret, needed_blocks;
1185         handle_t *handle;
1186         int retries = 0;
1187         struct page *page;
1188         pgoff_t index;
1189         unsigned from, to;
1190
1191         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
1192                 return -EIO;
1193
1194         trace_ext4_write_begin(inode, pos, len, flags);
1195         /*
1196          * Reserve one block more for addition to orphan list in case
1197          * we allocate blocks but write fails for some reason
1198          */
1199         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1200         index = pos >> PAGE_SHIFT;
1201         from = pos & (PAGE_SIZE - 1);
1202         to = from + len;
1203
1204         if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
1205                 ret = ext4_try_to_write_inline_data(mapping, inode, pos, len,
1206                                                     flags, pagep);
1207                 if (ret < 0)
1208                         return ret;
1209                 if (ret == 1)
1210                         return 0;
1211         }
1212
1213         /*
1214          * grab_cache_page_write_begin() can take a long time if the
1215          * system is thrashing due to memory pressure, or if the page
1216          * is being written back.  So grab it first before we start
1217          * the transaction handle.  This also allows us to allocate
1218          * the page (if needed) without using GFP_NOFS.
1219          */
1220 retry_grab:
1221         page = grab_cache_page_write_begin(mapping, index, flags);
1222         if (!page)
1223                 return -ENOMEM;
1224         unlock_page(page);
1225
1226 retry_journal:
1227         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE, needed_blocks);
1228         if (IS_ERR(handle)) {
1229                 put_page(page);
1230                 return PTR_ERR(handle);
1231         }
1232
1233         lock_page(page);
1234         if (page->mapping != mapping) {
1235                 /* The page got truncated from under us */
1236                 unlock_page(page);
1237                 put_page(page);
1238                 ext4_journal_stop(handle);
1239                 goto retry_grab;
1240         }
1241         /* In case writeback began while the page was unlocked */
1242         wait_for_stable_page(page);
1243
1244 #ifdef CONFIG_EXT4_FS_ENCRYPTION
1245         if (ext4_should_dioread_nolock(inode))
1246                 ret = ext4_block_write_begin(page, pos, len,
1247                                              ext4_get_block_unwritten);
1248         else
1249                 ret = ext4_block_write_begin(page, pos, len,
1250                                              ext4_get_block);
1251 #else
1252         if (ext4_should_dioread_nolock(inode))
1253                 ret = __block_write_begin(page, pos, len,
1254                                           ext4_get_block_unwritten);
1255         else
1256                 ret = __block_write_begin(page, pos, len, ext4_get_block);
1257 #endif
1258         if (!ret && ext4_should_journal_data(inode)) {
1259                 ret = ext4_walk_page_buffers(handle, page_buffers(page),
1260                                              from, to, NULL,
1261                                              do_journal_get_write_access);
1262         }
1263
1264         if (ret) {
1265                 unlock_page(page);
1266                 /*
1267                  * __block_write_begin may have instantiated a few blocks
1268                  * outside i_size.  Trim these off again. Don't need
1269                  * i_size_read because we hold i_mutex.
1270                  *
1271                  * Add inode to orphan list in case we crash before
1272                  * truncate finishes
1273                  */
1274                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1275                         ext4_orphan_add(handle, inode);
1276
1277                 ext4_journal_stop(handle);
1278                 if (pos + len > inode->i_size) {
1279                         ext4_truncate_failed_write(inode);
1280                         /*
1281                          * If truncate failed early the inode might
1282                          * still be on the orphan list; we need to
1283                          * make sure the inode is removed from the
1284                          * orphan list in that case.
1285                          */
1286                         if (inode->i_nlink)
1287                                 ext4_orphan_del(NULL, inode);
1288                 }
1289
1290                 if (ret == -ENOSPC &&
1291                     ext4_should_retry_alloc(inode->i_sb, &retries))
1292                         goto retry_journal;
1293                 put_page(page);
1294                 return ret;
1295         }
1296         *pagep = page;
1297         return ret;
1298 }
1299
1300 /* For write_end() in data=journal mode */
1301 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1302 {
1303         int ret;
1304         if (!buffer_mapped(bh) || buffer_freed(bh))
1305                 return 0;
1306         set_buffer_uptodate(bh);
1307         ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1308         clear_buffer_meta(bh);
1309         clear_buffer_prio(bh);
1310         return ret;
1311 }
1312
1313 /*
1314  * We need to pick up the new inode size which generic_commit_write gave us
1315  * `file' can be NULL - eg, when called from page_symlink().
1316  *
1317  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1318  * buffers are managed internally.
1319  */
1320 static int ext4_write_end(struct file *file,
1321                           struct address_space *mapping,
1322                           loff_t pos, unsigned len, unsigned copied,
1323                           struct page *page, void *fsdata)
1324 {
1325         handle_t *handle = ext4_journal_current_handle();
1326         struct inode *inode = mapping->host;
1327         loff_t old_size = inode->i_size;
1328         int ret = 0, ret2;
1329         int i_size_changed = 0;
1330
1331         trace_ext4_write_end(inode, pos, len, copied);
1332         if (ext4_has_inline_data(inode)) {
1333                 ret = ext4_write_inline_data_end(inode, pos, len,
1334                                                  copied, page);
1335                 if (ret < 0) {
1336                         unlock_page(page);
1337                         put_page(page);
1338                         goto errout;
1339                 }
1340                 copied = ret;
1341         } else
1342                 copied = block_write_end(file, mapping, pos,
1343                                          len, copied, page, fsdata);
1344         /*
1345          * it's important to update i_size while still holding page lock:
1346          * page writeout could otherwise come in and zero beyond i_size.
1347          */
1348         i_size_changed = ext4_update_inode_size(inode, pos + copied);
1349         unlock_page(page);
1350         put_page(page);
1351
1352         if (old_size < pos)
1353                 pagecache_isize_extended(inode, old_size, pos);
1354         /*
1355          * Don't mark the inode dirty under page lock. First, it unnecessarily
1356          * makes the holding time of page lock longer. Second, it forces lock
1357          * ordering of page lock and transaction start for journaling
1358          * filesystems.
1359          */
1360         if (i_size_changed)
1361                 ext4_mark_inode_dirty(handle, inode);
1362
1363         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1364                 /* if we have allocated more blocks and copied
1365                  * less. We will have blocks allocated outside
1366                  * inode->i_size. So truncate them
1367                  */
1368                 ext4_orphan_add(handle, inode);
1369 errout:
1370         ret2 = ext4_journal_stop(handle);
1371         if (!ret)
1372                 ret = ret2;
1373
1374         if (pos + len > inode->i_size) {
1375                 ext4_truncate_failed_write(inode);
1376                 /*
1377                  * If truncate failed early the inode might still be
1378                  * on the orphan list; we need to make sure the inode
1379                  * is removed from the orphan list in that case.
1380                  */
1381                 if (inode->i_nlink)
1382                         ext4_orphan_del(NULL, inode);
1383         }
1384
1385         return ret ? ret : copied;
1386 }
1387
1388 /*
1389  * This is a private version of page_zero_new_buffers() which doesn't
1390  * set the buffer to be dirty, since in data=journalled mode we need
1391  * to call ext4_handle_dirty_metadata() instead.
1392  */
1393 static void ext4_journalled_zero_new_buffers(handle_t *handle,
1394                                             struct page *page,
1395                                             unsigned from, unsigned to)
1396 {
1397         unsigned int block_start = 0, block_end;
1398         struct buffer_head *head, *bh;
1399
1400         bh = head = page_buffers(page);
1401         do {
1402                 block_end = block_start + bh->b_size;
1403                 if (buffer_new(bh)) {
1404                         if (block_end > from && block_start < to) {
1405                                 if (!PageUptodate(page)) {
1406                                         unsigned start, size;
1407
1408                                         start = max(from, block_start);
1409                                         size = min(to, block_end) - start;
1410
1411                                         zero_user(page, start, size);
1412                                         write_end_fn(handle, bh);
1413                                 }
1414                                 clear_buffer_new(bh);
1415                         }
1416                 }
1417                 block_start = block_end;
1418                 bh = bh->b_this_page;
1419         } while (bh != head);
1420 }
1421
1422 static int ext4_journalled_write_end(struct file *file,
1423                                      struct address_space *mapping,
1424                                      loff_t pos, unsigned len, unsigned copied,
1425                                      struct page *page, void *fsdata)
1426 {
1427         handle_t *handle = ext4_journal_current_handle();
1428         struct inode *inode = mapping->host;
1429         loff_t old_size = inode->i_size;
1430         int ret = 0, ret2;
1431         int partial = 0;
1432         unsigned from, to;
1433         int size_changed = 0;
1434
1435         trace_ext4_journalled_write_end(inode, pos, len, copied);
1436         from = pos & (PAGE_SIZE - 1);
1437         to = from + len;
1438
1439         BUG_ON(!ext4_handle_valid(handle));
1440
1441         if (ext4_has_inline_data(inode)) {
1442                 ret = ext4_write_inline_data_end(inode, pos, len,
1443                                                  copied, page);
1444                 if (ret < 0) {
1445                         unlock_page(page);
1446                         put_page(page);
1447                         goto errout;
1448                 }
1449                 copied = ret;
1450         } else if (unlikely(copied < len) && !PageUptodate(page)) {
1451                 copied = 0;
1452                 ext4_journalled_zero_new_buffers(handle, page, from, to);
1453         } else {
1454                 if (unlikely(copied < len))
1455                         ext4_journalled_zero_new_buffers(handle, page,
1456                                                          from + copied, to);
1457                 ret = ext4_walk_page_buffers(handle, page_buffers(page), from,
1458                                              from + copied, &partial,
1459                                              write_end_fn);
1460                 if (!partial)
1461                         SetPageUptodate(page);
1462         }
1463         size_changed = ext4_update_inode_size(inode, pos + copied);
1464         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1465         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1466         unlock_page(page);
1467         put_page(page);
1468
1469         if (old_size < pos)
1470                 pagecache_isize_extended(inode, old_size, pos);
1471
1472         if (size_changed) {
1473                 ret2 = ext4_mark_inode_dirty(handle, inode);
1474                 if (!ret)
1475                         ret = ret2;
1476         }
1477
1478         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1479                 /* if we have allocated more blocks and copied
1480                  * less. We will have blocks allocated outside
1481                  * inode->i_size. So truncate them
1482                  */
1483                 ext4_orphan_add(handle, inode);
1484
1485 errout:
1486         ret2 = ext4_journal_stop(handle);
1487         if (!ret)
1488                 ret = ret2;
1489         if (pos + len > inode->i_size) {
1490                 ext4_truncate_failed_write(inode);
1491                 /*
1492                  * If truncate failed early the inode might still be
1493                  * on the orphan list; we need to make sure the inode
1494                  * is removed from the orphan list in that case.
1495                  */
1496                 if (inode->i_nlink)
1497                         ext4_orphan_del(NULL, inode);
1498         }
1499
1500         return ret ? ret : copied;
1501 }
1502
1503 /*
1504  * Reserve space for a single cluster
1505  */
1506 static int ext4_da_reserve_space(struct inode *inode)
1507 {
1508         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1509         struct ext4_inode_info *ei = EXT4_I(inode);
1510         int ret;
1511
1512         /*
1513          * We will charge metadata quota at writeout time; this saves
1514          * us from metadata over-estimation, though we may go over by
1515          * a small amount in the end.  Here we just reserve for data.
1516          */
1517         ret = dquot_reserve_block(inode, EXT4_C2B(sbi, 1));
1518         if (ret)
1519                 return ret;
1520
1521         spin_lock(&ei->i_block_reservation_lock);
1522         if (ext4_claim_free_clusters(sbi, 1, 0)) {
1523                 spin_unlock(&ei->i_block_reservation_lock);
1524                 dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
1525                 return -ENOSPC;
1526         }
1527         ei->i_reserved_data_blocks++;
1528         trace_ext4_da_reserve_space(inode);
1529         spin_unlock(&ei->i_block_reservation_lock);
1530
1531         return 0;       /* success */
1532 }
1533
1534 static void ext4_da_release_space(struct inode *inode, int to_free)
1535 {
1536         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1537         struct ext4_inode_info *ei = EXT4_I(inode);
1538
1539         if (!to_free)
1540                 return;         /* Nothing to release, exit */
1541
1542         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1543
1544         trace_ext4_da_release_space(inode, to_free);
1545         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1546                 /*
1547                  * if there aren't enough reserved blocks, then the
1548                  * counter is messed up somewhere.  Since this
1549                  * function is called from invalidate page, it's
1550                  * harmless to return without any action.
1551                  */
1552                 ext4_warning(inode->i_sb, "ext4_da_release_space: "
1553                          "ino %lu, to_free %d with only %d reserved "
1554                          "data blocks", inode->i_ino, to_free,
1555                          ei->i_reserved_data_blocks);
1556                 WARN_ON(1);
1557                 to_free = ei->i_reserved_data_blocks;
1558         }
1559         ei->i_reserved_data_blocks -= to_free;
1560
1561         /* update fs dirty data blocks counter */
1562         percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1563
1564         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1565
1566         dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1567 }
1568
1569 static void ext4_da_page_release_reservation(struct page *page,
1570                                              unsigned int offset,
1571                                              unsigned int length)
1572 {
1573         int to_release = 0, contiguous_blks = 0;
1574         struct buffer_head *head, *bh;
1575         unsigned int curr_off = 0;
1576         struct inode *inode = page->mapping->host;
1577         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1578         unsigned int stop = offset + length;
1579         int num_clusters;
1580         ext4_fsblk_t lblk;
1581
1582         BUG_ON(stop > PAGE_SIZE || stop < length);
1583
1584         head = page_buffers(page);
1585         bh = head;
1586         do {
1587                 unsigned int next_off = curr_off + bh->b_size;
1588
1589                 if (next_off > stop)
1590                         break;
1591
1592                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1593                         to_release++;
1594                         contiguous_blks++;
1595                         clear_buffer_delay(bh);
1596                 } else if (contiguous_blks) {
1597                         lblk = page->index <<
1598                                (PAGE_SHIFT - inode->i_blkbits);
1599                         lblk += (curr_off >> inode->i_blkbits) -
1600                                 contiguous_blks;
1601                         ext4_es_remove_extent(inode, lblk, contiguous_blks);
1602                         contiguous_blks = 0;
1603                 }
1604                 curr_off = next_off;
1605         } while ((bh = bh->b_this_page) != head);
1606
1607         if (contiguous_blks) {
1608                 lblk = page->index << (PAGE_SHIFT - inode->i_blkbits);
1609                 lblk += (curr_off >> inode->i_blkbits) - contiguous_blks;
1610                 ext4_es_remove_extent(inode, lblk, contiguous_blks);
1611         }
1612
1613         /* If we have released all the blocks belonging to a cluster, then we
1614          * need to release the reserved space for that cluster. */
1615         num_clusters = EXT4_NUM_B2C(sbi, to_release);
1616         while (num_clusters > 0) {
1617                 lblk = (page->index << (PAGE_SHIFT - inode->i_blkbits)) +
1618                         ((num_clusters - 1) << sbi->s_cluster_bits);
1619                 if (sbi->s_cluster_ratio == 1 ||
1620                     !ext4_find_delalloc_cluster(inode, lblk))
1621                         ext4_da_release_space(inode, 1);
1622
1623                 num_clusters--;
1624         }
1625 }
1626
1627 /*
1628  * Delayed allocation stuff
1629  */
1630
1631 struct mpage_da_data {
1632         struct inode *inode;
1633         struct writeback_control *wbc;
1634
1635         pgoff_t first_page;     /* The first page to write */
1636         pgoff_t next_page;      /* Current page to examine */
1637         pgoff_t last_page;      /* Last page to examine */
1638         /*
1639          * Extent to map - this can be after first_page because that can be
1640          * fully mapped. We somewhat abuse m_flags to store whether the extent
1641          * is delalloc or unwritten.
1642          */
1643         struct ext4_map_blocks map;
1644         struct ext4_io_submit io_submit;        /* IO submission data */
1645         unsigned int do_map:1;
1646 };
1647
1648 static void mpage_release_unused_pages(struct mpage_da_data *mpd,
1649                                        bool invalidate)
1650 {
1651         int nr_pages, i;
1652         pgoff_t index, end;
1653         struct pagevec pvec;
1654         struct inode *inode = mpd->inode;
1655         struct address_space *mapping = inode->i_mapping;
1656
1657         /* This is necessary when next_page == 0. */
1658         if (mpd->first_page >= mpd->next_page)
1659                 return;
1660
1661         index = mpd->first_page;
1662         end   = mpd->next_page - 1;
1663         if (invalidate) {
1664                 ext4_lblk_t start, last;
1665                 start = index << (PAGE_SHIFT - inode->i_blkbits);
1666                 last = end << (PAGE_SHIFT - inode->i_blkbits);
1667                 ext4_es_remove_extent(inode, start, last - start + 1);
1668         }
1669
1670         pagevec_init(&pvec, 0);
1671         while (index <= end) {
1672                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1673                 if (nr_pages == 0)
1674                         break;
1675                 for (i = 0; i < nr_pages; i++) {
1676                         struct page *page = pvec.pages[i];
1677                         if (page->index > end)
1678                                 break;
1679                         BUG_ON(!PageLocked(page));
1680                         BUG_ON(PageWriteback(page));
1681                         if (invalidate) {
1682                                 if (page_mapped(page))
1683                                         clear_page_dirty_for_io(page);
1684                                 block_invalidatepage(page, 0, PAGE_SIZE);
1685                                 ClearPageUptodate(page);
1686                         }
1687                         unlock_page(page);
1688                 }
1689                 index = pvec.pages[nr_pages - 1]->index + 1;
1690                 pagevec_release(&pvec);
1691         }
1692 }
1693
1694 static void ext4_print_free_blocks(struct inode *inode)
1695 {
1696         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1697         struct super_block *sb = inode->i_sb;
1698         struct ext4_inode_info *ei = EXT4_I(inode);
1699
1700         ext4_msg(sb, KERN_CRIT, "Total free blocks count %lld",
1701                EXT4_C2B(EXT4_SB(inode->i_sb),
1702                         ext4_count_free_clusters(sb)));
1703         ext4_msg(sb, KERN_CRIT, "Free/Dirty block details");
1704         ext4_msg(sb, KERN_CRIT, "free_blocks=%lld",
1705                (long long) EXT4_C2B(EXT4_SB(sb),
1706                 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1707         ext4_msg(sb, KERN_CRIT, "dirty_blocks=%lld",
1708                (long long) EXT4_C2B(EXT4_SB(sb),
1709                 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1710         ext4_msg(sb, KERN_CRIT, "Block reservation details");
1711         ext4_msg(sb, KERN_CRIT, "i_reserved_data_blocks=%u",
1712                  ei->i_reserved_data_blocks);
1713         return;
1714 }
1715
1716 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1717 {
1718         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1719 }
1720
1721 /*
1722  * This function is grabs code from the very beginning of
1723  * ext4_map_blocks, but assumes that the caller is from delayed write
1724  * time. This function looks up the requested blocks and sets the
1725  * buffer delay bit under the protection of i_data_sem.
1726  */
1727 static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1728                               struct ext4_map_blocks *map,
1729                               struct buffer_head *bh)
1730 {
1731         struct extent_status es;
1732         int retval;
1733         sector_t invalid_block = ~((sector_t) 0xffff);
1734 #ifdef ES_AGGRESSIVE_TEST
1735         struct ext4_map_blocks orig_map;
1736
1737         memcpy(&orig_map, map, sizeof(*map));
1738 #endif
1739
1740         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1741                 invalid_block = ~0;
1742
1743         map->m_flags = 0;
1744         ext_debug("ext4_da_map_blocks(): inode %lu, max_blocks %u,"
1745                   "logical block %lu\n", inode->i_ino, map->m_len,
1746                   (unsigned long) map->m_lblk);
1747
1748         /* Lookup extent status tree firstly */
1749         if (ext4_es_lookup_extent(inode, iblock, &es)) {
1750                 if (ext4_es_is_hole(&es)) {
1751                         retval = 0;
1752                         down_read(&EXT4_I(inode)->i_data_sem);
1753                         goto add_delayed;
1754                 }
1755
1756                 /*
1757                  * Delayed extent could be allocated by fallocate.
1758                  * So we need to check it.
1759                  */
1760                 if (ext4_es_is_delayed(&es) && !ext4_es_is_unwritten(&es)) {
1761                         map_bh(bh, inode->i_sb, invalid_block);
1762                         set_buffer_new(bh);
1763                         set_buffer_delay(bh);
1764                         return 0;
1765                 }
1766
1767                 map->m_pblk = ext4_es_pblock(&es) + iblock - es.es_lblk;
1768                 retval = es.es_len - (iblock - es.es_lblk);
1769                 if (retval > map->m_len)
1770                         retval = map->m_len;
1771                 map->m_len = retval;
1772                 if (ext4_es_is_written(&es))
1773                         map->m_flags |= EXT4_MAP_MAPPED;
1774                 else if (ext4_es_is_unwritten(&es))
1775                         map->m_flags |= EXT4_MAP_UNWRITTEN;
1776                 else
1777                         BUG_ON(1);
1778
1779 #ifdef ES_AGGRESSIVE_TEST
1780                 ext4_map_blocks_es_recheck(NULL, inode, map, &orig_map, 0);
1781 #endif
1782                 return retval;
1783         }
1784
1785         /*
1786          * Try to see if we can get the block without requesting a new
1787          * file system block.
1788          */
1789         down_read(&EXT4_I(inode)->i_data_sem);
1790         if (ext4_has_inline_data(inode))
1791                 retval = 0;
1792         else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1793                 retval = ext4_ext_map_blocks(NULL, inode, map, 0);
1794         else
1795                 retval = ext4_ind_map_blocks(NULL, inode, map, 0);
1796
1797 add_delayed:
1798         if (retval == 0) {
1799                 int ret;
1800                 /*
1801                  * XXX: __block_prepare_write() unmaps passed block,
1802                  * is it OK?
1803                  */
1804                 /*
1805                  * If the block was allocated from previously allocated cluster,
1806                  * then we don't need to reserve it again. However we still need
1807                  * to reserve metadata for every block we're going to write.
1808                  */
1809                 if (EXT4_SB(inode->i_sb)->s_cluster_ratio == 1 ||
1810                     !ext4_find_delalloc_cluster(inode, map->m_lblk)) {
1811                         ret = ext4_da_reserve_space(inode);
1812                         if (ret) {
1813                                 /* not enough space to reserve */
1814                                 retval = ret;
1815                                 goto out_unlock;
1816                         }
1817                 }
1818
1819                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1820                                             ~0, EXTENT_STATUS_DELAYED);
1821                 if (ret) {
1822                         retval = ret;
1823                         goto out_unlock;
1824                 }
1825
1826                 map_bh(bh, inode->i_sb, invalid_block);
1827                 set_buffer_new(bh);
1828                 set_buffer_delay(bh);
1829         } else if (retval > 0) {
1830                 int ret;
1831                 unsigned int status;
1832
1833                 if (unlikely(retval != map->m_len)) {
1834                         ext4_warning(inode->i_sb,
1835                                      "ES len assertion failed for inode "
1836                                      "%lu: retval %d != map->m_len %d",
1837                                      inode->i_ino, retval, map->m_len);
1838                         WARN_ON(1);
1839                 }
1840
1841                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
1842                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
1843                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1844                                             map->m_pblk, status);
1845                 if (ret != 0)
1846                         retval = ret;
1847         }
1848
1849 out_unlock:
1850         up_read((&EXT4_I(inode)->i_data_sem));
1851
1852         return retval;
1853 }
1854
1855 /*
1856  * This is a special get_block_t callback which is used by
1857  * ext4_da_write_begin().  It will either return mapped block or
1858  * reserve space for a single block.
1859  *
1860  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1861  * We also have b_blocknr = -1 and b_bdev initialized properly
1862  *
1863  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1864  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1865  * initialized properly.
1866  */
1867 int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1868                            struct buffer_head *bh, int create)
1869 {
1870         struct ext4_map_blocks map;
1871         int ret = 0;
1872
1873         BUG_ON(create == 0);
1874         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1875
1876         map.m_lblk = iblock;
1877         map.m_len = 1;
1878
1879         /*
1880          * first, we need to know whether the block is allocated already
1881          * preallocated blocks are unmapped but should treated
1882          * the same as allocated blocks.
1883          */
1884         ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1885         if (ret <= 0)
1886                 return ret;
1887
1888         map_bh(bh, inode->i_sb, map.m_pblk);
1889         ext4_update_bh_state(bh, map.m_flags);
1890
1891         if (buffer_unwritten(bh)) {
1892                 /* A delayed write to unwritten bh should be marked
1893                  * new and mapped.  Mapped ensures that we don't do
1894                  * get_block multiple times when we write to the same
1895                  * offset and new ensures that we do proper zero out
1896                  * for partial write.
1897                  */
1898                 set_buffer_new(bh);
1899                 set_buffer_mapped(bh);
1900         }
1901         return 0;
1902 }
1903
1904 static int bget_one(handle_t *handle, struct buffer_head *bh)
1905 {
1906         get_bh(bh);
1907         return 0;
1908 }
1909
1910 static int bput_one(handle_t *handle, struct buffer_head *bh)
1911 {
1912         put_bh(bh);
1913         return 0;
1914 }
1915
1916 static int __ext4_journalled_writepage(struct page *page,
1917                                        unsigned int len)
1918 {
1919         struct address_space *mapping = page->mapping;
1920         struct inode *inode = mapping->host;
1921         struct buffer_head *page_bufs = NULL;
1922         handle_t *handle = NULL;
1923         int ret = 0, err = 0;
1924         int inline_data = ext4_has_inline_data(inode);
1925         struct buffer_head *inode_bh = NULL;
1926
1927         ClearPageChecked(page);
1928
1929         if (inline_data) {
1930                 BUG_ON(page->index != 0);
1931                 BUG_ON(len > ext4_get_max_inline_size(inode));
1932                 inode_bh = ext4_journalled_write_inline_data(inode, len, page);
1933                 if (inode_bh == NULL)
1934                         goto out;
1935         } else {
1936                 page_bufs = page_buffers(page);
1937                 if (!page_bufs) {
1938                         BUG();
1939                         goto out;
1940                 }
1941                 ext4_walk_page_buffers(handle, page_bufs, 0, len,
1942                                        NULL, bget_one);
1943         }
1944         /*
1945          * We need to release the page lock before we start the
1946          * journal, so grab a reference so the page won't disappear
1947          * out from under us.
1948          */
1949         get_page(page);
1950         unlock_page(page);
1951
1952         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
1953                                     ext4_writepage_trans_blocks(inode));
1954         if (IS_ERR(handle)) {
1955                 ret = PTR_ERR(handle);
1956                 put_page(page);
1957                 goto out_no_pagelock;
1958         }
1959         BUG_ON(!ext4_handle_valid(handle));
1960
1961         lock_page(page);
1962         put_page(page);
1963         if (page->mapping != mapping) {
1964                 /* The page got truncated from under us */
1965                 ext4_journal_stop(handle);
1966                 ret = 0;
1967                 goto out;
1968         }
1969
1970         if (inline_data) {
1971                 BUFFER_TRACE(inode_bh, "get write access");
1972                 ret = ext4_journal_get_write_access(handle, inode_bh);
1973
1974                 err = ext4_handle_dirty_metadata(handle, inode, inode_bh);
1975
1976         } else {
1977                 ret = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
1978                                              do_journal_get_write_access);
1979
1980                 err = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
1981                                              write_end_fn);
1982         }
1983         if (ret == 0)
1984                 ret = err;
1985         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1986         err = ext4_journal_stop(handle);
1987         if (!ret)
1988                 ret = err;
1989
1990         if (!ext4_has_inline_data(inode))
1991                 ext4_walk_page_buffers(NULL, page_bufs, 0, len,
1992                                        NULL, bput_one);
1993         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1994 out:
1995         unlock_page(page);
1996 out_no_pagelock:
1997         brelse(inode_bh);
1998         return ret;
1999 }
2000
2001 /*
2002  * Note that we don't need to start a transaction unless we're journaling data
2003  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2004  * need to file the inode to the transaction's list in ordered mode because if
2005  * we are writing back data added by write(), the inode is already there and if
2006  * we are writing back data modified via mmap(), no one guarantees in which
2007  * transaction the data will hit the disk. In case we are journaling data, we
2008  * cannot start transaction directly because transaction start ranks above page
2009  * lock so we have to do some magic.
2010  *
2011  * This function can get called via...
2012  *   - ext4_writepages after taking page lock (have journal handle)
2013  *   - journal_submit_inode_data_buffers (no journal handle)
2014  *   - shrink_page_list via the kswapd/direct reclaim (no journal handle)
2015  *   - grab_page_cache when doing write_begin (have journal handle)
2016  *
2017  * We don't do any block allocation in this function. If we have page with
2018  * multiple blocks we need to write those buffer_heads that are mapped. This
2019  * is important for mmaped based write. So if we do with blocksize 1K
2020  * truncate(f, 1024);
2021  * a = mmap(f, 0, 4096);
2022  * a[0] = 'a';
2023  * truncate(f, 4096);
2024  * we have in the page first buffer_head mapped via page_mkwrite call back
2025  * but other buffer_heads would be unmapped but dirty (dirty done via the
2026  * do_wp_page). So writepage should write the first block. If we modify
2027  * the mmap area beyond 1024 we will again get a page_fault and the
2028  * page_mkwrite callback will do the block allocation and mark the
2029  * buffer_heads mapped.
2030  *
2031  * We redirty the page if we have any buffer_heads that is either delay or
2032  * unwritten in the page.
2033  *
2034  * We can get recursively called as show below.
2035  *
2036  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2037  *              ext4_writepage()
2038  *
2039  * But since we don't do any block allocation we should not deadlock.
2040  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2041  */
2042 static int ext4_writepage(struct page *page,
2043                           struct writeback_control *wbc)
2044 {
2045         int ret = 0;
2046         loff_t size;
2047         unsigned int len;
2048         struct buffer_head *page_bufs = NULL;
2049         struct inode *inode = page->mapping->host;
2050         struct ext4_io_submit io_submit;
2051         bool keep_towrite = false;
2052
2053         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb)))) {
2054                 ext4_invalidatepage(page, 0, PAGE_SIZE);
2055                 unlock_page(page);
2056                 return -EIO;
2057         }
2058
2059         trace_ext4_writepage(page);
2060         size = i_size_read(inode);
2061         if (page->index == size >> PAGE_SHIFT)
2062                 len = size & ~PAGE_MASK;
2063         else
2064                 len = PAGE_SIZE;
2065
2066         page_bufs = page_buffers(page);
2067         /*
2068          * We cannot do block allocation or other extent handling in this
2069          * function. If there are buffers needing that, we have to redirty
2070          * the page. But we may reach here when we do a journal commit via
2071          * journal_submit_inode_data_buffers() and in that case we must write
2072          * allocated buffers to achieve data=ordered mode guarantees.
2073          *
2074          * Also, if there is only one buffer per page (the fs block
2075          * size == the page size), if one buffer needs block
2076          * allocation or needs to modify the extent tree to clear the
2077          * unwritten flag, we know that the page can't be written at
2078          * all, so we might as well refuse the write immediately.
2079          * Unfortunately if the block size != page size, we can't as
2080          * easily detect this case using ext4_walk_page_buffers(), but
2081          * for the extremely common case, this is an optimization that
2082          * skips a useless round trip through ext4_bio_write_page().
2083          */
2084         if (ext4_walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2085                                    ext4_bh_delay_or_unwritten)) {
2086                 redirty_page_for_writepage(wbc, page);
2087                 if ((current->flags & PF_MEMALLOC) ||
2088                     (inode->i_sb->s_blocksize == PAGE_SIZE)) {
2089                         /*
2090                          * For memory cleaning there's no point in writing only
2091                          * some buffers. So just bail out. Warn if we came here
2092                          * from direct reclaim.
2093                          */
2094                         WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD))
2095                                                         == PF_MEMALLOC);
2096                         unlock_page(page);
2097                         return 0;
2098                 }
2099                 keep_towrite = true;
2100         }
2101
2102         if (PageChecked(page) && ext4_should_journal_data(inode))
2103                 /*
2104                  * It's mmapped pagecache.  Add buffers and journal it.  There
2105                  * doesn't seem much point in redirtying the page here.
2106                  */
2107                 return __ext4_journalled_writepage(page, len);
2108
2109         ext4_io_submit_init(&io_submit, wbc);
2110         io_submit.io_end = ext4_init_io_end(inode, GFP_NOFS);
2111         if (!io_submit.io_end) {
2112                 redirty_page_for_writepage(wbc, page);
2113                 unlock_page(page);
2114                 return -ENOMEM;
2115         }
2116         ret = ext4_bio_write_page(&io_submit, page, len, wbc, keep_towrite);
2117         ext4_io_submit(&io_submit);
2118         /* Drop io_end reference we got from init */
2119         ext4_put_io_end_defer(io_submit.io_end);
2120         return ret;
2121 }
2122
2123 static int mpage_submit_page(struct mpage_da_data *mpd, struct page *page)
2124 {
2125         int len;
2126         loff_t size;
2127         int err;
2128
2129         BUG_ON(page->index != mpd->first_page);
2130         clear_page_dirty_for_io(page);
2131         /*
2132          * We have to be very careful here!  Nothing protects writeback path
2133          * against i_size changes and the page can be writeably mapped into
2134          * page tables. So an application can be growing i_size and writing
2135          * data through mmap while writeback runs. clear_page_dirty_for_io()
2136          * write-protects our page in page tables and the page cannot get
2137          * written to again until we release page lock. So only after
2138          * clear_page_dirty_for_io() we are safe to sample i_size for
2139          * ext4_bio_write_page() to zero-out tail of the written page. We rely
2140          * on the barrier provided by TestClearPageDirty in
2141          * clear_page_dirty_for_io() to make sure i_size is really sampled only
2142          * after page tables are updated.
2143          */
2144         size = i_size_read(mpd->inode);
2145         if (page->index == size >> PAGE_SHIFT)
2146                 len = size & ~PAGE_MASK;
2147         else
2148                 len = PAGE_SIZE;
2149         err = ext4_bio_write_page(&mpd->io_submit, page, len, mpd->wbc, false);
2150         if (!err)
2151                 mpd->wbc->nr_to_write--;
2152         mpd->first_page++;
2153
2154         return err;
2155 }
2156
2157 #define BH_FLAGS ((1 << BH_Unwritten) | (1 << BH_Delay))
2158
2159 /*
2160  * mballoc gives us at most this number of blocks...
2161  * XXX: That seems to be only a limitation of ext4_mb_normalize_request().
2162  * The rest of mballoc seems to handle chunks up to full group size.
2163  */
2164 #define MAX_WRITEPAGES_EXTENT_LEN 2048
2165
2166 /*
2167  * mpage_add_bh_to_extent - try to add bh to extent of blocks to map
2168  *
2169  * @mpd - extent of blocks
2170  * @lblk - logical number of the block in the file
2171  * @bh - buffer head we want to add to the extent
2172  *
2173  * The function is used to collect contig. blocks in the same state. If the
2174  * buffer doesn't require mapping for writeback and we haven't started the
2175  * extent of buffers to map yet, the function returns 'true' immediately - the
2176  * caller can write the buffer right away. Otherwise the function returns true
2177  * if the block has been added to the extent, false if the block couldn't be
2178  * added.
2179  */
2180 static bool mpage_add_bh_to_extent(struct mpage_da_data *mpd, ext4_lblk_t lblk,
2181                                    struct buffer_head *bh)
2182 {
2183         struct ext4_map_blocks *map = &mpd->map;
2184
2185         /* Buffer that doesn't need mapping for writeback? */
2186         if (!buffer_dirty(bh) || !buffer_mapped(bh) ||
2187             (!buffer_delay(bh) && !buffer_unwritten(bh))) {
2188                 /* So far no extent to map => we write the buffer right away */
2189                 if (map->m_len == 0)
2190                         return true;
2191                 return false;
2192         }
2193
2194         /* First block in the extent? */
2195         if (map->m_len == 0) {
2196                 /* We cannot map unless handle is started... */
2197                 if (!mpd->do_map)
2198                         return false;
2199                 map->m_lblk = lblk;
2200                 map->m_len = 1;
2201                 map->m_flags = bh->b_state & BH_FLAGS;
2202                 return true;
2203         }
2204
2205         /* Don't go larger than mballoc is willing to allocate */
2206         if (map->m_len >= MAX_WRITEPAGES_EXTENT_LEN)
2207                 return false;
2208
2209         /* Can we merge the block to our big extent? */
2210         if (lblk == map->m_lblk + map->m_len &&
2211             (bh->b_state & BH_FLAGS) == map->m_flags) {
2212                 map->m_len++;
2213                 return true;
2214         }
2215         return false;
2216 }
2217
2218 /*
2219  * mpage_process_page_bufs - submit page buffers for IO or add them to extent
2220  *
2221  * @mpd - extent of blocks for mapping
2222  * @head - the first buffer in the page
2223  * @bh - buffer we should start processing from
2224  * @lblk - logical number of the block in the file corresponding to @bh
2225  *
2226  * Walk through page buffers from @bh upto @head (exclusive) and either submit
2227  * the page for IO if all buffers in this page were mapped and there's no
2228  * accumulated extent of buffers to map or add buffers in the page to the
2229  * extent of buffers to map. The function returns 1 if the caller can continue
2230  * by processing the next page, 0 if it should stop adding buffers to the
2231  * extent to map because we cannot extend it anymore. It can also return value
2232  * < 0 in case of error during IO submission.
2233  */
2234 static int mpage_process_page_bufs(struct mpage_da_data *mpd,
2235                                    struct buffer_head *head,
2236                                    struct buffer_head *bh,
2237                                    ext4_lblk_t lblk)
2238 {
2239         struct inode *inode = mpd->inode;
2240         int err;
2241         ext4_lblk_t blocks = (i_size_read(inode) + i_blocksize(inode) - 1)
2242                                                         >> inode->i_blkbits;
2243
2244         do {
2245                 BUG_ON(buffer_locked(bh));
2246
2247                 if (lblk >= blocks || !mpage_add_bh_to_extent(mpd, lblk, bh)) {
2248                         /* Found extent to map? */
2249                         if (mpd->map.m_len)
2250                                 return 0;
2251                         /* Buffer needs mapping and handle is not started? */
2252                         if (!mpd->do_map)
2253                                 return 0;
2254                         /* Everything mapped so far and we hit EOF */
2255                         break;
2256                 }
2257         } while (lblk++, (bh = bh->b_this_page) != head);
2258         /* So far everything mapped? Submit the page for IO. */
2259         if (mpd->map.m_len == 0) {
2260                 err = mpage_submit_page(mpd, head->b_page);
2261                 if (err < 0)
2262                         return err;
2263         }
2264         return lblk < blocks;
2265 }
2266
2267 /*
2268  * mpage_map_buffers - update buffers corresponding to changed extent and
2269  *                     submit fully mapped pages for IO
2270  *
2271  * @mpd - description of extent to map, on return next extent to map
2272  *
2273  * Scan buffers corresponding to changed extent (we expect corresponding pages
2274  * to be already locked) and update buffer state according to new extent state.
2275  * We map delalloc buffers to their physical location, clear unwritten bits,
2276  * and mark buffers as uninit when we perform writes to unwritten extents
2277  * and do extent conversion after IO is finished. If the last page is not fully
2278  * mapped, we update @map to the next extent in the last page that needs
2279  * mapping. Otherwise we submit the page for IO.
2280  */
2281 static int mpage_map_and_submit_buffers(struct mpage_da_data *mpd)
2282 {
2283         struct pagevec pvec;
2284         int nr_pages, i;
2285         struct inode *inode = mpd->inode;
2286         struct buffer_head *head, *bh;
2287         int bpp_bits = PAGE_SHIFT - inode->i_blkbits;
2288         pgoff_t start, end;
2289         ext4_lblk_t lblk;
2290         sector_t pblock;
2291         int err;
2292
2293         start = mpd->map.m_lblk >> bpp_bits;
2294         end = (mpd->map.m_lblk + mpd->map.m_len - 1) >> bpp_bits;
2295         lblk = start << bpp_bits;
2296         pblock = mpd->map.m_pblk;
2297
2298         pagevec_init(&pvec, 0);
2299         while (start <= end) {
2300                 nr_pages = pagevec_lookup(&pvec, inode->i_mapping, start,
2301                                           PAGEVEC_SIZE);
2302                 if (nr_pages == 0)
2303                         break;
2304                 for (i = 0; i < nr_pages; i++) {
2305                         struct page *page = pvec.pages[i];
2306
2307                         if (page->index > end)
2308                                 break;
2309                         /* Up to 'end' pages must be contiguous */
2310                         BUG_ON(page->index != start);
2311                         bh = head = page_buffers(page);
2312                         do {
2313                                 if (lblk < mpd->map.m_lblk)
2314                                         continue;
2315                                 if (lblk >= mpd->map.m_lblk + mpd->map.m_len) {
2316                                         /*
2317                                          * Buffer after end of mapped extent.
2318                                          * Find next buffer in the page to map.
2319                                          */
2320                                         mpd->map.m_len = 0;
2321                                         mpd->map.m_flags = 0;
2322                                         /*
2323                                          * FIXME: If dioread_nolock supports
2324                                          * blocksize < pagesize, we need to make
2325                                          * sure we add size mapped so far to
2326                                          * io_end->size as the following call
2327                                          * can submit the page for IO.
2328                                          */
2329                                         err = mpage_process_page_bufs(mpd, head,
2330                                                                       bh, lblk);
2331                                         pagevec_release(&pvec);
2332                                         if (err > 0)
2333                                                 err = 0;
2334                                         return err;
2335                                 }
2336                                 if (buffer_delay(bh)) {
2337                                         clear_buffer_delay(bh);
2338                                         bh->b_blocknr = pblock++;
2339                                 }
2340                                 clear_buffer_unwritten(bh);
2341                         } while (lblk++, (bh = bh->b_this_page) != head);
2342
2343                         /*
2344                          * FIXME: This is going to break if dioread_nolock
2345                          * supports blocksize < pagesize as we will try to
2346                          * convert potentially unmapped parts of inode.
2347                          */
2348                         mpd->io_submit.io_end->size += PAGE_SIZE;
2349                         /* Page fully mapped - let IO run! */
2350                         err = mpage_submit_page(mpd, page);
2351                         if (err < 0) {
2352                                 pagevec_release(&pvec);
2353                                 return err;
2354                         }
2355                         start++;
2356                 }
2357                 pagevec_release(&pvec);
2358         }
2359         /* Extent fully mapped and matches with page boundary. We are done. */
2360         mpd->map.m_len = 0;
2361         mpd->map.m_flags = 0;
2362         return 0;
2363 }
2364
2365 static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
2366 {
2367         struct inode *inode = mpd->inode;
2368         struct ext4_map_blocks *map = &mpd->map;
2369         int get_blocks_flags;
2370         int err, dioread_nolock;
2371
2372         trace_ext4_da_write_pages_extent(inode, map);
2373         /*
2374          * Call ext4_map_blocks() to allocate any delayed allocation blocks, or
2375          * to convert an unwritten extent to be initialized (in the case
2376          * where we have written into one or more preallocated blocks).  It is
2377          * possible that we're going to need more metadata blocks than
2378          * previously reserved. However we must not fail because we're in
2379          * writeback and there is nothing we can do about it so it might result
2380          * in data loss.  So use reserved blocks to allocate metadata if
2381          * possible.
2382          *
2383          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE if
2384          * the blocks in question are delalloc blocks.  This indicates
2385          * that the blocks and quotas has already been checked when
2386          * the data was copied into the page cache.
2387          */
2388         get_blocks_flags = EXT4_GET_BLOCKS_CREATE |
2389                            EXT4_GET_BLOCKS_METADATA_NOFAIL |
2390                            EXT4_GET_BLOCKS_IO_SUBMIT;
2391         dioread_nolock = ext4_should_dioread_nolock(inode);
2392         if (dioread_nolock)
2393                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2394         if (map->m_flags & (1 << BH_Delay))
2395                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2396
2397         err = ext4_map_blocks(handle, inode, map, get_blocks_flags);
2398         if (err < 0)
2399                 return err;
2400         if (dioread_nolock && (map->m_flags & EXT4_MAP_UNWRITTEN)) {
2401                 if (!mpd->io_submit.io_end->handle &&
2402                     ext4_handle_valid(handle)) {
2403                         mpd->io_submit.io_end->handle = handle->h_rsv_handle;
2404                         handle->h_rsv_handle = NULL;
2405                 }
2406                 ext4_set_io_unwritten_flag(inode, mpd->io_submit.io_end);
2407         }
2408
2409         BUG_ON(map->m_len == 0);
2410         if (map->m_flags & EXT4_MAP_NEW) {
2411                 clean_bdev_aliases(inode->i_sb->s_bdev, map->m_pblk,
2412                                    map->m_len);
2413         }
2414         return 0;
2415 }
2416
2417 /*
2418  * mpage_map_and_submit_extent - map extent starting at mpd->lblk of length
2419  *                               mpd->len and submit pages underlying it for IO
2420  *
2421  * @handle - handle for journal operations
2422  * @mpd - extent to map
2423  * @give_up_on_write - we set this to true iff there is a fatal error and there
2424  *                     is no hope of writing the data. The caller should discard
2425  *                     dirty pages to avoid infinite loops.
2426  *
2427  * The function maps extent starting at mpd->lblk of length mpd->len. If it is
2428  * delayed, blocks are allocated, if it is unwritten, we may need to convert
2429  * them to initialized or split the described range from larger unwritten
2430  * extent. Note that we need not map all the described range since allocation
2431  * can return less blocks or the range is covered by more unwritten extents. We
2432  * cannot map more because we are limited by reserved transaction credits. On
2433  * the other hand we always make sure that the last touched page is fully
2434  * mapped so that it can be written out (and thus forward progress is
2435  * guaranteed). After mapping we submit all mapped pages for IO.
2436  */
2437 static int mpage_map_and_submit_extent(handle_t *handle,
2438                                        struct mpage_da_data *mpd,
2439                                        bool *give_up_on_write)
2440 {
2441         struct inode *inode = mpd->inode;
2442         struct ext4_map_blocks *map = &mpd->map;
2443         int err;
2444         loff_t disksize;
2445         int progress = 0;
2446
2447         mpd->io_submit.io_end->offset =
2448                                 ((loff_t)map->m_lblk) << inode->i_blkbits;
2449         do {
2450                 err = mpage_map_one_extent(handle, mpd);
2451                 if (err < 0) {
2452                         struct super_block *sb = inode->i_sb;
2453
2454                         if (ext4_forced_shutdown(EXT4_SB(sb)) ||
2455                             EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
2456                                 goto invalidate_dirty_pages;
2457                         /*
2458                          * Let the uper layers retry transient errors.
2459                          * In the case of ENOSPC, if ext4_count_free_blocks()
2460                          * is non-zero, a commit should free up blocks.
2461                          */
2462                         if ((err == -ENOMEM) ||
2463                             (err == -ENOSPC && ext4_count_free_clusters(sb))) {
2464                                 if (progress)
2465                                         goto update_disksize;
2466                                 return err;
2467                         }
2468                         ext4_msg(sb, KERN_CRIT,
2469                                  "Delayed block allocation failed for "
2470                                  "inode %lu at logical offset %llu with"
2471                                  " max blocks %u with error %d",
2472                                  inode->i_ino,
2473                                  (unsigned long long)map->m_lblk,
2474                                  (unsigned)map->m_len, -err);
2475                         ext4_msg(sb, KERN_CRIT,
2476                                  "This should not happen!! Data will "
2477                                  "be lost\n");
2478                         if (err == -ENOSPC)
2479                                 ext4_print_free_blocks(inode);
2480                 invalidate_dirty_pages:
2481                         *give_up_on_write = true;
2482                         return err;
2483                 }
2484                 progress = 1;
2485                 /*
2486                  * Update buffer state, submit mapped pages, and get us new
2487                  * extent to map
2488                  */
2489                 err = mpage_map_and_submit_buffers(mpd);
2490                 if (err < 0)
2491                         goto update_disksize;
2492         } while (map->m_len);
2493
2494 update_disksize:
2495         /*
2496          * Update on-disk size after IO is submitted.  Races with
2497          * truncate are avoided by checking i_size under i_data_sem.
2498          */
2499         disksize = ((loff_t)mpd->first_page) << PAGE_SHIFT;
2500         if (disksize > EXT4_I(inode)->i_disksize) {
2501                 int err2;
2502                 loff_t i_size;
2503
2504                 down_write(&EXT4_I(inode)->i_data_sem);
2505                 i_size = i_size_read(inode);
2506                 if (disksize > i_size)
2507                         disksize = i_size;
2508                 if (disksize > EXT4_I(inode)->i_disksize)
2509                         EXT4_I(inode)->i_disksize = disksize;
2510                 up_write(&EXT4_I(inode)->i_data_sem);
2511                 err2 = ext4_mark_inode_dirty(handle, inode);
2512                 if (err2)
2513                         ext4_error(inode->i_sb,
2514                                    "Failed to mark inode %lu dirty",
2515                                    inode->i_ino);
2516                 if (!err)
2517                         err = err2;
2518         }
2519         return err;
2520 }
2521
2522 /*
2523  * Calculate the total number of credits to reserve for one writepages
2524  * iteration. This is called from ext4_writepages(). We map an extent of
2525  * up to MAX_WRITEPAGES_EXTENT_LEN blocks and then we go on and finish mapping
2526  * the last partial page. So in total we can map MAX_WRITEPAGES_EXTENT_LEN +
2527  * bpp - 1 blocks in bpp different extents.
2528  */
2529 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2530 {
2531         int bpp = ext4_journal_blocks_per_page(inode);
2532
2533         return ext4_meta_trans_blocks(inode,
2534                                 MAX_WRITEPAGES_EXTENT_LEN + bpp - 1, bpp);
2535 }
2536
2537 /*
2538  * mpage_prepare_extent_to_map - find & lock contiguous range of dirty pages
2539  *                               and underlying extent to map
2540  *
2541  * @mpd - where to look for pages
2542  *
2543  * Walk dirty pages in the mapping. If they are fully mapped, submit them for
2544  * IO immediately. When we find a page which isn't mapped we start accumulating
2545  * extent of buffers underlying these pages that needs mapping (formed by
2546  * either delayed or unwritten buffers). We also lock the pages containing
2547  * these buffers. The extent found is returned in @mpd structure (starting at
2548  * mpd->lblk with length mpd->len blocks).
2549  *
2550  * Note that this function can attach bios to one io_end structure which are
2551  * neither logically nor physically contiguous. Although it may seem as an
2552  * unnecessary complication, it is actually inevitable in blocksize < pagesize
2553  * case as we need to track IO to all buffers underlying a page in one io_end.
2554  */
2555 static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
2556 {
2557         struct address_space *mapping = mpd->inode->i_mapping;
2558         struct pagevec pvec;
2559         unsigned int nr_pages;
2560         long left = mpd->wbc->nr_to_write;
2561         pgoff_t index = mpd->first_page;
2562         pgoff_t end = mpd->last_page;
2563         int tag;
2564         int i, err = 0;
2565         int blkbits = mpd->inode->i_blkbits;
2566         ext4_lblk_t lblk;
2567         struct buffer_head *head;
2568
2569         if (mpd->wbc->sync_mode == WB_SYNC_ALL || mpd->wbc->tagged_writepages)
2570                 tag = PAGECACHE_TAG_TOWRITE;
2571         else
2572                 tag = PAGECACHE_TAG_DIRTY;
2573
2574         pagevec_init(&pvec, 0);
2575         mpd->map.m_len = 0;
2576         mpd->next_page = index;
2577         while (index <= end) {
2578                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2579                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2580                 if (nr_pages == 0)
2581                         goto out;
2582
2583                 for (i = 0; i < nr_pages; i++) {
2584                         struct page *page = pvec.pages[i];
2585
2586                         /*
2587                          * At this point, the page may be truncated or
2588                          * invalidated (changing page->mapping to NULL), or
2589                          * even swizzled back from swapper_space to tmpfs file
2590                          * mapping. However, page->index will not change
2591                          * because we have a reference on the page.
2592                          */
2593                         if (page->index > end)
2594                                 goto out;
2595
2596                         /*
2597                          * Accumulated enough dirty pages? This doesn't apply
2598                          * to WB_SYNC_ALL mode. For integrity sync we have to
2599                          * keep going because someone may be concurrently
2600                          * dirtying pages, and we might have synced a lot of
2601                          * newly appeared dirty pages, but have not synced all
2602                          * of the old dirty pages.
2603                          */
2604                         if (mpd->wbc->sync_mode == WB_SYNC_NONE && left <= 0)
2605                                 goto out;
2606
2607                         /* If we can't merge this page, we are done. */
2608                         if (mpd->map.m_len > 0 && mpd->next_page != page->index)
2609                                 goto out;
2610
2611                         lock_page(page);
2612                         /*
2613                          * If the page is no longer dirty, or its mapping no
2614                          * longer corresponds to inode we are writing (which
2615                          * means it has been truncated or invalidated), or the
2616                          * page is already under writeback and we are not doing
2617                          * a data integrity writeback, skip the page
2618                          */
2619                         if (!PageDirty(page) ||
2620                             (PageWriteback(page) &&
2621                              (mpd->wbc->sync_mode == WB_SYNC_NONE)) ||
2622                             unlikely(page->mapping != mapping)) {
2623                                 unlock_page(page);
2624                                 continue;
2625                         }
2626
2627                         wait_on_page_writeback(page);
2628                         BUG_ON(PageWriteback(page));
2629
2630                         if (mpd->map.m_len == 0)
2631                                 mpd->first_page = page->index;
2632                         mpd->next_page = page->index + 1;
2633                         /* Add all dirty buffers to mpd */
2634                         lblk = ((ext4_lblk_t)page->index) <<
2635                                 (PAGE_SHIFT - blkbits);
2636                         head = page_buffers(page);
2637                         err = mpage_process_page_bufs(mpd, head, head, lblk);
2638                         if (err <= 0)
2639                                 goto out;
2640                         err = 0;
2641                         left--;
2642                 }
2643                 pagevec_release(&pvec);
2644                 cond_resched();
2645         }
2646         return 0;
2647 out:
2648         pagevec_release(&pvec);
2649         return err;
2650 }
2651
2652 static int __writepage(struct page *page, struct writeback_control *wbc,
2653                        void *data)
2654 {
2655         struct address_space *mapping = data;
2656         int ret = ext4_writepage(page, wbc);
2657         mapping_set_error(mapping, ret);
2658         return ret;
2659 }
2660
2661 static int ext4_writepages(struct address_space *mapping,
2662                            struct writeback_control *wbc)
2663 {
2664         pgoff_t writeback_index = 0;
2665         long nr_to_write = wbc->nr_to_write;
2666         int range_whole = 0;
2667         int cycled = 1;
2668         handle_t *handle = NULL;
2669         struct mpage_da_data mpd;
2670         struct inode *inode = mapping->host;
2671         int needed_blocks, rsv_blocks = 0, ret = 0;
2672         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2673         bool done;
2674         struct blk_plug plug;
2675         bool give_up_on_write = false;
2676
2677         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
2678                 return -EIO;
2679
2680         percpu_down_read(&sbi->s_journal_flag_rwsem);
2681         trace_ext4_writepages(inode, wbc);
2682
2683         if (dax_mapping(mapping)) {
2684                 ret = dax_writeback_mapping_range(mapping, inode->i_sb->s_bdev,
2685                                                   wbc);
2686                 goto out_writepages;
2687         }
2688
2689         /*
2690          * No pages to write? This is mainly a kludge to avoid starting
2691          * a transaction for special inodes like journal inode on last iput()
2692          * because that could violate lock ordering on umount
2693          */
2694         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2695                 goto out_writepages;
2696
2697         if (ext4_should_journal_data(inode)) {
2698                 struct blk_plug plug;
2699
2700                 blk_start_plug(&plug);
2701                 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2702                 blk_finish_plug(&plug);
2703                 goto out_writepages;
2704         }
2705
2706         /*
2707          * If the filesystem has aborted, it is read-only, so return
2708          * right away instead of dumping stack traces later on that
2709          * will obscure the real source of the problem.  We test
2710          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2711          * the latter could be true if the filesystem is mounted
2712          * read-only, and in that case, ext4_writepages should
2713          * *never* be called, so if that ever happens, we would want
2714          * the stack trace.
2715          */
2716         if (unlikely(ext4_forced_shutdown(EXT4_SB(mapping->host->i_sb)) ||
2717                      sbi->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2718                 ret = -EROFS;
2719                 goto out_writepages;
2720         }
2721
2722         if (ext4_should_dioread_nolock(inode)) {
2723                 /*
2724                  * We may need to convert up to one extent per block in
2725                  * the page and we may dirty the inode.
2726                  */
2727                 rsv_blocks = 1 + (PAGE_SIZE >> inode->i_blkbits);
2728         }
2729
2730         /*
2731          * If we have inline data and arrive here, it means that
2732          * we will soon create the block for the 1st page, so
2733          * we'd better clear the inline data here.
2734          */
2735         if (ext4_has_inline_data(inode)) {
2736                 /* Just inode will be modified... */
2737                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
2738                 if (IS_ERR(handle)) {
2739                         ret = PTR_ERR(handle);
2740                         goto out_writepages;
2741                 }
2742                 BUG_ON(ext4_test_inode_state(inode,
2743                                 EXT4_STATE_MAY_INLINE_DATA));
2744                 ext4_destroy_inline_data(handle, inode);
2745                 ext4_journal_stop(handle);
2746         }
2747
2748         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2749                 range_whole = 1;
2750
2751         if (wbc->range_cyclic) {
2752                 writeback_index = mapping->writeback_index;
2753                 if (writeback_index)
2754                         cycled = 0;
2755                 mpd.first_page = writeback_index;
2756                 mpd.last_page = -1;
2757         } else {
2758                 mpd.first_page = wbc->range_start >> PAGE_SHIFT;
2759                 mpd.last_page = wbc->range_end >> PAGE_SHIFT;
2760         }
2761
2762         mpd.inode = inode;
2763         mpd.wbc = wbc;
2764         ext4_io_submit_init(&mpd.io_submit, wbc);
2765 retry:
2766         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2767                 tag_pages_for_writeback(mapping, mpd.first_page, mpd.last_page);
2768         done = false;
2769         blk_start_plug(&plug);
2770
2771         /*
2772          * First writeback pages that don't need mapping - we can avoid
2773          * starting a transaction unnecessarily and also avoid being blocked
2774          * in the block layer on device congestion while having transaction
2775          * started.
2776          */
2777         mpd.do_map = 0;
2778         mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2779         if (!mpd.io_submit.io_end) {
2780                 ret = -ENOMEM;
2781                 goto unplug;
2782         }
2783         ret = mpage_prepare_extent_to_map(&mpd);
2784         /* Submit prepared bio */
2785         ext4_io_submit(&mpd.io_submit);
2786         ext4_put_io_end_defer(mpd.io_submit.io_end);
2787         mpd.io_submit.io_end = NULL;
2788         /* Unlock pages we didn't use */
2789         mpage_release_unused_pages(&mpd, false);
2790         if (ret < 0)
2791                 goto unplug;
2792
2793         while (!done && mpd.first_page <= mpd.last_page) {
2794                 /* For each extent of pages we use new io_end */
2795                 mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2796                 if (!mpd.io_submit.io_end) {
2797                         ret = -ENOMEM;
2798                         break;
2799                 }
2800
2801                 /*
2802                  * We have two constraints: We find one extent to map and we
2803                  * must always write out whole page (makes a difference when
2804                  * blocksize < pagesize) so that we don't block on IO when we
2805                  * try to write out the rest of the page. Journalled mode is
2806                  * not supported by delalloc.
2807                  */
2808                 BUG_ON(ext4_should_journal_data(inode));
2809                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2810
2811                 /* start a new transaction */
2812                 handle = ext4_journal_start_with_reserve(inode,
2813                                 EXT4_HT_WRITE_PAGE, needed_blocks, rsv_blocks);
2814                 if (IS_ERR(handle)) {
2815                         ret = PTR_ERR(handle);
2816                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2817                                "%ld pages, ino %lu; err %d", __func__,
2818                                 wbc->nr_to_write, inode->i_ino, ret);
2819                         /* Release allocated io_end */
2820                         ext4_put_io_end(mpd.io_submit.io_end);
2821                         mpd.io_submit.io_end = NULL;
2822                         break;
2823                 }
2824                 mpd.do_map = 1;
2825
2826                 trace_ext4_da_write_pages(inode, mpd.first_page, mpd.wbc);
2827                 ret = mpage_prepare_extent_to_map(&mpd);
2828                 if (!ret) {
2829                         if (mpd.map.m_len)
2830                                 ret = mpage_map_and_submit_extent(handle, &mpd,
2831                                         &give_up_on_write);
2832                         else {
2833                                 /*
2834                                  * We scanned the whole range (or exhausted
2835                                  * nr_to_write), submitted what was mapped and
2836                                  * didn't find anything needing mapping. We are
2837                                  * done.
2838                                  */
2839                                 done = true;
2840                         }
2841                 }
2842                 /*
2843                  * Caution: If the handle is synchronous,
2844                  * ext4_journal_stop() can wait for transaction commit
2845                  * to finish which may depend on writeback of pages to
2846                  * complete or on page lock to be released.  In that
2847                  * case, we have to wait until after after we have
2848                  * submitted all the IO, released page locks we hold,
2849                  * and dropped io_end reference (for extent conversion
2850                  * to be able to complete) before stopping the handle.
2851                  */
2852                 if (!ext4_handle_valid(handle) || handle->h_sync == 0) {
2853                         ext4_journal_stop(handle);
2854                         handle = NULL;
2855                         mpd.do_map = 0;
2856                 }
2857                 /* Submit prepared bio */
2858                 ext4_io_submit(&mpd.io_submit);
2859                 /* Unlock pages we didn't use */
2860                 mpage_release_unused_pages(&mpd, give_up_on_write);
2861                 /*
2862                  * Drop our io_end reference we got from init. We have
2863                  * to be careful and use deferred io_end finishing if
2864                  * we are still holding the transaction as we can
2865                  * release the last reference to io_end which may end
2866                  * up doing unwritten extent conversion.
2867                  */
2868                 if (handle) {
2869                         ext4_put_io_end_defer(mpd.io_submit.io_end);
2870                         ext4_journal_stop(handle);
2871                 } else
2872                         ext4_put_io_end(mpd.io_submit.io_end);
2873                 mpd.io_submit.io_end = NULL;
2874
2875                 if (ret == -ENOSPC && sbi->s_journal) {
2876                         /*
2877                          * Commit the transaction which would
2878                          * free blocks released in the transaction
2879                          * and try again
2880                          */
2881                         jbd2_journal_force_commit_nested(sbi->s_journal);
2882                         ret = 0;
2883                         continue;
2884                 }
2885                 /* Fatal error - ENOMEM, EIO... */
2886                 if (ret)
2887                         break;
2888         }
2889 unplug:
2890         blk_finish_plug(&plug);
2891         if (!ret && !cycled && wbc->nr_to_write > 0) {
2892                 cycled = 1;
2893                 mpd.last_page = writeback_index - 1;
2894                 mpd.first_page = 0;
2895                 goto retry;
2896         }
2897
2898         /* Update index */
2899         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2900                 /*
2901                  * Set the writeback_index so that range_cyclic
2902                  * mode will write it back later
2903                  */
2904                 mapping->writeback_index = mpd.first_page;
2905
2906 out_writepages:
2907         trace_ext4_writepages_result(inode, wbc, ret,
2908                                      nr_to_write - wbc->nr_to_write);
2909         percpu_up_read(&sbi->s_journal_flag_rwsem);
2910         return ret;
2911 }
2912
2913 static int ext4_nonda_switch(struct super_block *sb)
2914 {
2915         s64 free_clusters, dirty_clusters;
2916         struct ext4_sb_info *sbi = EXT4_SB(sb);
2917
2918         /*
2919          * switch to non delalloc mode if we are running low
2920          * on free block. The free block accounting via percpu
2921          * counters can get slightly wrong with percpu_counter_batch getting
2922          * accumulated on each CPU without updating global counters
2923          * Delalloc need an accurate free block accounting. So switch
2924          * to non delalloc when we are near to error range.
2925          */
2926         free_clusters =
2927                 percpu_counter_read_positive(&sbi->s_freeclusters_counter);
2928         dirty_clusters =
2929                 percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
2930         /*
2931          * Start pushing delalloc when 1/2 of free blocks are dirty.
2932          */
2933         if (dirty_clusters && (free_clusters < 2 * dirty_clusters))
2934                 try_to_writeback_inodes_sb(sb, WB_REASON_FS_FREE_SPACE);
2935
2936         if (2 * free_clusters < 3 * dirty_clusters ||
2937             free_clusters < (dirty_clusters + EXT4_FREECLUSTERS_WATERMARK)) {
2938                 /*
2939                  * free block count is less than 150% of dirty blocks
2940                  * or free blocks is less than watermark
2941                  */
2942                 return 1;
2943         }
2944         return 0;
2945 }
2946
2947 /* We always reserve for an inode update; the superblock could be there too */
2948 static int ext4_da_write_credits(struct inode *inode, loff_t pos, unsigned len)
2949 {
2950         if (likely(ext4_has_feature_large_file(inode->i_sb)))
2951                 return 1;
2952
2953         if (pos + len <= 0x7fffffffULL)
2954                 return 1;
2955
2956         /* We might need to update the superblock to set LARGE_FILE */
2957         return 2;
2958 }
2959
2960 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2961                                loff_t pos, unsigned len, unsigned flags,
2962                                struct page **pagep, void **fsdata)
2963 {
2964         int ret, retries = 0;
2965         struct page *page;
2966         pgoff_t index;
2967         struct inode *inode = mapping->host;
2968         handle_t *handle;
2969
2970         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
2971                 return -EIO;
2972
2973         index = pos >> PAGE_SHIFT;
2974
2975         if (ext4_nonda_switch(inode->i_sb) ||
2976             S_ISLNK(inode->i_mode)) {
2977                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2978                 return ext4_write_begin(file, mapping, pos,
2979                                         len, flags, pagep, fsdata);
2980         }
2981         *fsdata = (void *)0;
2982         trace_ext4_da_write_begin(inode, pos, len, flags);
2983
2984         if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
2985                 ret = ext4_da_write_inline_data_begin(mapping, inode,
2986                                                       pos, len, flags,
2987                                                       pagep, fsdata);
2988                 if (ret < 0)
2989                         return ret;
2990                 if (ret == 1)
2991                         return 0;
2992         }
2993
2994         /*
2995          * grab_cache_page_write_begin() can take a long time if the
2996          * system is thrashing due to memory pressure, or if the page
2997          * is being written back.  So grab it first before we start
2998          * the transaction handle.  This also allows us to allocate
2999          * the page (if needed) without using GFP_NOFS.
3000          */
3001 retry_grab:
3002         page = grab_cache_page_write_begin(mapping, index, flags);
3003         if (!page)
3004                 return -ENOMEM;
3005         unlock_page(page);
3006
3007         /*
3008          * With delayed allocation, we don't log the i_disksize update
3009          * if there is delayed block allocation. But we still need
3010          * to journalling the i_disksize update if writes to the end
3011          * of file which has an already mapped buffer.
3012          */
3013 retry_journal:
3014         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
3015                                 ext4_da_write_credits(inode, pos, len));
3016         if (IS_ERR(handle)) {
3017                 put_page(page);
3018                 return PTR_ERR(handle);
3019         }
3020
3021         lock_page(page);
3022         if (page->mapping != mapping) {
3023                 /* The page got truncated from under us */
3024                 unlock_page(page);
3025                 put_page(page);
3026                 ext4_journal_stop(handle);
3027                 goto retry_grab;
3028         }
3029         /* In case writeback began while the page was unlocked */
3030         wait_for_stable_page(page);
3031
3032 #ifdef CONFIG_EXT4_FS_ENCRYPTION
3033         ret = ext4_block_write_begin(page, pos, len,
3034                                      ext4_da_get_block_prep);
3035 #else
3036         ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
3037 #endif
3038         if (ret < 0) {
3039                 unlock_page(page);
3040                 ext4_journal_stop(handle);
3041                 /*
3042                  * block_write_begin may have instantiated a few blocks
3043                  * outside i_size.  Trim these off again. Don't need
3044                  * i_size_read because we hold i_mutex.
3045                  */
3046                 if (pos + len > inode->i_size)
3047                         ext4_truncate_failed_write(inode);
3048
3049                 if (ret == -ENOSPC &&
3050                     ext4_should_retry_alloc(inode->i_sb, &retries))
3051                         goto retry_journal;
3052
3053                 put_page(page);
3054                 return ret;
3055         }
3056
3057         *pagep = page;
3058         return ret;
3059 }
3060
3061 /*
3062  * Check if we should update i_disksize
3063  * when write to the end of file but not require block allocation
3064  */
3065 static int ext4_da_should_update_i_disksize(struct page *page,
3066                                             unsigned long offset)
3067 {
3068         struct buffer_head *bh;
3069         struct inode *inode = page->mapping->host;
3070         unsigned int idx;
3071         int i;
3072
3073         bh = page_buffers(page);
3074         idx = offset >> inode->i_blkbits;
3075
3076         for (i = 0; i < idx; i++)
3077                 bh = bh->b_this_page;
3078
3079         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3080                 return 0;
3081         return 1;
3082 }
3083
3084 static int ext4_da_write_end(struct file *file,
3085                              struct address_space *mapping,
3086                              loff_t pos, unsigned len, unsigned copied,
3087                              struct page *page, void *fsdata)
3088 {
3089         struct inode *inode = mapping->host;
3090         int ret = 0, ret2;
3091         handle_t *handle = ext4_journal_current_handle();
3092         loff_t new_i_size;
3093         unsigned long start, end;
3094         int write_mode = (int)(unsigned long)fsdata;
3095
3096         if (write_mode == FALL_BACK_TO_NONDELALLOC)
3097                 return ext4_write_end(file, mapping, pos,
3098                                       len, copied, page, fsdata);
3099
3100         trace_ext4_da_write_end(inode, pos, len, copied);
3101         start = pos & (PAGE_SIZE - 1);
3102         end = start + copied - 1;
3103
3104         /*
3105          * generic_write_end() will run mark_inode_dirty() if i_size
3106          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3107          * into that.
3108          */
3109         new_i_size = pos + copied;
3110         if (copied && new_i_size > EXT4_I(inode)->i_disksize) {
3111                 if (ext4_has_inline_data(inode) ||
3112                     ext4_da_should_update_i_disksize(page, end)) {
3113                         ext4_update_i_disksize(inode, new_i_size);
3114                         /* We need to mark inode dirty even if
3115                          * new_i_size is less that inode->i_size
3116                          * bu greater than i_disksize.(hint delalloc)
3117                          */
3118                         ext4_mark_inode_dirty(handle, inode);
3119                 }
3120         }
3121
3122         if (write_mode != CONVERT_INLINE_DATA &&
3123             ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA) &&
3124             ext4_has_inline_data(inode))
3125                 ret2 = ext4_da_write_inline_data_end(inode, pos, len, copied,
3126                                                      page);
3127         else
3128                 ret2 = generic_write_end(file, mapping, pos, len, copied,
3129                                                         page, fsdata);
3130
3131         copied = ret2;
3132         if (ret2 < 0)
3133                 ret = ret2;
3134         ret2 = ext4_journal_stop(handle);
3135         if (!ret)
3136                 ret = ret2;
3137
3138         return ret ? ret : copied;
3139 }
3140
3141 static void ext4_da_invalidatepage(struct page *page, unsigned int offset,
3142                                    unsigned int length)
3143 {
3144         /*
3145          * Drop reserved blocks
3146          */
3147         BUG_ON(!PageLocked(page));
3148         if (!page_has_buffers(page))
3149                 goto out;
3150
3151         ext4_da_page_release_reservation(page, offset, length);
3152
3153 out:
3154         ext4_invalidatepage(page, offset, length);
3155
3156         return;
3157 }
3158
3159 /*
3160  * Force all delayed allocation blocks to be allocated for a given inode.
3161  */
3162 int ext4_alloc_da_blocks(struct inode *inode)
3163 {
3164         trace_ext4_alloc_da_blocks(inode);
3165
3166         if (!EXT4_I(inode)->i_reserved_data_blocks)
3167                 return 0;
3168
3169         /*
3170          * We do something simple for now.  The filemap_flush() will
3171          * also start triggering a write of the data blocks, which is
3172          * not strictly speaking necessary (and for users of
3173          * laptop_mode, not even desirable).  However, to do otherwise
3174          * would require replicating code paths in:
3175          *
3176          * ext4_writepages() ->
3177          *    write_cache_pages() ---> (via passed in callback function)
3178          *        __mpage_da_writepage() -->
3179          *           mpage_add_bh_to_extent()
3180          *           mpage_da_map_blocks()
3181          *
3182          * The problem is that write_cache_pages(), located in
3183          * mm/page-writeback.c, marks pages clean in preparation for
3184          * doing I/O, which is not desirable if we're not planning on
3185          * doing I/O at all.
3186          *
3187          * We could call write_cache_pages(), and then redirty all of
3188          * the pages by calling redirty_page_for_writepage() but that
3189          * would be ugly in the extreme.  So instead we would need to
3190          * replicate parts of the code in the above functions,
3191          * simplifying them because we wouldn't actually intend to
3192          * write out the pages, but rather only collect contiguous
3193          * logical block extents, call the multi-block allocator, and
3194          * then update the buffer heads with the block allocations.
3195          *
3196          * For now, though, we'll cheat by calling filemap_flush(),
3197          * which will map the blocks, and start the I/O, but not
3198          * actually wait for the I/O to complete.
3199          */
3200         return filemap_flush(inode->i_mapping);
3201 }
3202
3203 /*
3204  * bmap() is special.  It gets used by applications such as lilo and by
3205  * the swapper to find the on-disk block of a specific piece of data.
3206  *
3207  * Naturally, this is dangerous if the block concerned is still in the
3208  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3209  * filesystem and enables swap, then they may get a nasty shock when the
3210  * data getting swapped to that swapfile suddenly gets overwritten by
3211  * the original zero's written out previously to the journal and
3212  * awaiting writeback in the kernel's buffer cache.
3213  *
3214  * So, if we see any bmap calls here on a modified, data-journaled file,
3215  * take extra steps to flush any blocks which might be in the cache.
3216  */
3217 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3218 {
3219         struct inode *inode = mapping->host;
3220         journal_t *journal;
3221         int err;
3222
3223         /*
3224          * We can get here for an inline file via the FIBMAP ioctl
3225          */
3226         if (ext4_has_inline_data(inode))
3227                 return 0;
3228
3229         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3230                         test_opt(inode->i_sb, DELALLOC)) {
3231                 /*
3232                  * With delalloc we want to sync the file
3233                  * so that we can make sure we allocate
3234                  * blocks for file
3235                  */
3236                 filemap_write_and_wait(mapping);
3237         }
3238
3239         if (EXT4_JOURNAL(inode) &&
3240             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3241                 /*
3242                  * This is a REALLY heavyweight approach, but the use of
3243                  * bmap on dirty files is expected to be extremely rare:
3244                  * only if we run lilo or swapon on a freshly made file
3245                  * do we expect this to happen.
3246                  *
3247                  * (bmap requires CAP_SYS_RAWIO so this does not
3248                  * represent an unprivileged user DOS attack --- we'd be
3249                  * in trouble if mortal users could trigger this path at
3250                  * will.)
3251                  *
3252                  * NB. EXT4_STATE_JDATA is not set on files other than
3253                  * regular files.  If somebody wants to bmap a directory
3254                  * or symlink and gets confused because the buffer
3255                  * hasn't yet been flushed to disk, they deserve
3256                  * everything they get.
3257                  */
3258
3259                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3260                 journal = EXT4_JOURNAL(inode);
3261                 jbd2_journal_lock_updates(journal);
3262                 err = jbd2_journal_flush(journal);
3263                 jbd2_journal_unlock_updates(journal);
3264
3265                 if (err)
3266                         return 0;
3267         }
3268
3269         return generic_block_bmap(mapping, block, ext4_get_block);
3270 }
3271
3272 static int ext4_readpage(struct file *file, struct page *page)
3273 {
3274         int ret = -EAGAIN;
3275         struct inode *inode = page->mapping->host;
3276
3277         trace_ext4_readpage(page);
3278
3279         if (ext4_has_inline_data(inode))
3280                 ret = ext4_readpage_inline(inode, page);
3281
3282         if (ret == -EAGAIN)
3283                 return ext4_mpage_readpages(page->mapping, NULL, page, 1);
3284
3285         return ret;
3286 }
3287
3288 static int
3289 ext4_readpages(struct file *file, struct address_space *mapping,
3290                 struct list_head *pages, unsigned nr_pages)
3291 {
3292         struct inode *inode = mapping->host;
3293
3294         /* If the file has inline data, no need to do readpages. */
3295         if (ext4_has_inline_data(inode))
3296                 return 0;
3297
3298         return ext4_mpage_readpages(mapping, pages, NULL, nr_pages);
3299 }
3300
3301 static void ext4_invalidatepage(struct page *page, unsigned int offset,
3302                                 unsigned int length)
3303 {
3304         trace_ext4_invalidatepage(page, offset, length);
3305
3306         /* No journalling happens on data buffers when this function is used */
3307         WARN_ON(page_has_buffers(page) && buffer_jbd(page_buffers(page)));
3308
3309         block_invalidatepage(page, offset, length);
3310 }
3311
3312 static int __ext4_journalled_invalidatepage(struct page *page,
3313                                             unsigned int offset,
3314                                             unsigned int length)
3315 {
3316         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3317
3318         trace_ext4_journalled_invalidatepage(page, offset, length);
3319
3320         /*
3321          * If it's a full truncate we just forget about the pending dirtying
3322          */
3323         if (offset == 0 && length == PAGE_SIZE)
3324                 ClearPageChecked(page);
3325
3326         return jbd2_journal_invalidatepage(journal, page, offset, length);
3327 }
3328
3329 /* Wrapper for aops... */
3330 static void ext4_journalled_invalidatepage(struct page *page,
3331                                            unsigned int offset,
3332                                            unsigned int length)
3333 {
3334         WARN_ON(__ext4_journalled_invalidatepage(page, offset, length) < 0);
3335 }
3336
3337 static int ext4_releasepage(struct page *page, gfp_t wait)
3338 {
3339         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3340
3341         trace_ext4_releasepage(page);
3342
3343         /* Page has dirty journalled data -> cannot release */
3344         if (PageChecked(page))
3345                 return 0;
3346         if (journal)
3347                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3348         else
3349                 return try_to_free_buffers(page);
3350 }
3351
3352 #ifdef CONFIG_FS_DAX
3353 static int ext4_iomap_begin(struct inode *inode, loff_t offset, loff_t length,
3354                             unsigned flags, struct iomap *iomap)
3355 {
3356         struct block_device *bdev;
3357         unsigned int blkbits = inode->i_blkbits;
3358         unsigned long first_block = offset >> blkbits;
3359         unsigned long last_block = (offset + length - 1) >> blkbits;
3360         struct ext4_map_blocks map;
3361         int ret;
3362
3363         if (WARN_ON_ONCE(ext4_has_inline_data(inode)))
3364                 return -ERANGE;
3365
3366         map.m_lblk = first_block;
3367         map.m_len = last_block - first_block + 1;
3368
3369         if (!(flags & IOMAP_WRITE)) {
3370                 ret = ext4_map_blocks(NULL, inode, &map, 0);
3371         } else {
3372                 int dio_credits;
3373                 handle_t *handle;
3374                 int retries = 0;
3375
3376                 /* Trim mapping request to maximum we can map at once for DIO */
3377                 if (map.m_len > DIO_MAX_BLOCKS)
3378                         map.m_len = DIO_MAX_BLOCKS;
3379                 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
3380 retry:
3381                 /*
3382                  * Either we allocate blocks and then we don't get unwritten
3383                  * extent so we have reserved enough credits, or the blocks
3384                  * are already allocated and unwritten and in that case
3385                  * extent conversion fits in the credits as well.
3386                  */
3387                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
3388                                             dio_credits);
3389                 if (IS_ERR(handle))
3390                         return PTR_ERR(handle);
3391
3392                 ret = ext4_map_blocks(handle, inode, &map,
3393                                       EXT4_GET_BLOCKS_CREATE_ZERO);
3394                 if (ret < 0) {
3395                         ext4_journal_stop(handle);
3396                         if (ret == -ENOSPC &&
3397                             ext4_should_retry_alloc(inode->i_sb, &retries))
3398                                 goto retry;
3399                         return ret;
3400                 }
3401
3402                 /*
3403                  * If we added blocks beyond i_size, we need to make sure they
3404                  * will get truncated if we crash before updating i_size in
3405                  * ext4_iomap_end(). For faults we don't need to do that (and
3406                  * even cannot because for orphan list operations inode_lock is
3407                  * required) - if we happen to instantiate block beyond i_size,
3408                  * it is because we race with truncate which has already added
3409                  * the inode to the orphan list.
3410                  */
3411                 if (!(flags & IOMAP_FAULT) && first_block + map.m_len >
3412                     (i_size_read(inode) + (1 << blkbits) - 1) >> blkbits) {
3413                         int err;
3414
3415                         err = ext4_orphan_add(handle, inode);
3416                         if (err < 0) {
3417                                 ext4_journal_stop(handle);
3418                                 return err;
3419                         }
3420                 }
3421                 ext4_journal_stop(handle);
3422         }
3423
3424         iomap->flags = 0;
3425         bdev = inode->i_sb->s_bdev;
3426         iomap->bdev = bdev;
3427         if (blk_queue_dax(bdev->bd_queue))
3428                 iomap->dax_dev = fs_dax_get_by_host(bdev->bd_disk->disk_name);
3429         else
3430                 iomap->dax_dev = NULL;
3431         iomap->offset = first_block << blkbits;
3432
3433         if (ret == 0) {
3434                 iomap->type = IOMAP_HOLE;
3435                 iomap->blkno = IOMAP_NULL_BLOCK;
3436                 iomap->length = (u64)map.m_len << blkbits;
3437         } else {
3438                 if (map.m_flags & EXT4_MAP_MAPPED) {
3439                         iomap->type = IOMAP_MAPPED;
3440                 } else if (map.m_flags & EXT4_MAP_UNWRITTEN) {
3441                         iomap->type = IOMAP_UNWRITTEN;
3442                 } else {
3443                         WARN_ON_ONCE(1);
3444                         return -EIO;
3445                 }
3446                 iomap->blkno = (sector_t)map.m_pblk << (blkbits - 9);
3447                 iomap->length = (u64)map.m_len << blkbits;
3448         }
3449
3450         if (map.m_flags & EXT4_MAP_NEW)
3451                 iomap->flags |= IOMAP_F_NEW;
3452         return 0;
3453 }
3454
3455 static int ext4_iomap_end(struct inode *inode, loff_t offset, loff_t length,
3456                           ssize_t written, unsigned flags, struct iomap *iomap)
3457 {
3458         int ret = 0;
3459         handle_t *handle;
3460         int blkbits = inode->i_blkbits;
3461         bool truncate = false;
3462
3463         fs_put_dax(iomap->dax_dev);
3464         if (!(flags & IOMAP_WRITE) || (flags & IOMAP_FAULT))
3465                 return 0;
3466
3467         handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
3468         if (IS_ERR(handle)) {
3469                 ret = PTR_ERR(handle);
3470                 goto orphan_del;
3471         }
3472         if (ext4_update_inode_size(inode, offset + written))
3473                 ext4_mark_inode_dirty(handle, inode);
3474         /*
3475          * We may need to truncate allocated but not written blocks beyond EOF.
3476          */
3477         if (iomap->offset + iomap->length > 
3478             ALIGN(inode->i_size, 1 << blkbits)) {
3479                 ext4_lblk_t written_blk, end_blk;
3480
3481                 written_blk = (offset + written) >> blkbits;
3482                 end_blk = (offset + length) >> blkbits;
3483                 if (written_blk < end_blk && ext4_can_truncate(inode))
3484                         truncate = true;
3485         }
3486         /*
3487          * Remove inode from orphan list if we were extending a inode and
3488          * everything went fine.
3489          */
3490         if (!truncate && inode->i_nlink &&
3491             !list_empty(&EXT4_I(inode)->i_orphan))
3492                 ext4_orphan_del(handle, inode);
3493         ext4_journal_stop(handle);
3494         if (truncate) {
3495                 ext4_truncate_failed_write(inode);
3496 orphan_del:
3497                 /*
3498                  * If truncate failed early the inode might still be on the
3499                  * orphan list; we need to make sure the inode is removed from
3500                  * the orphan list in that case.
3501                  */
3502                 if (inode->i_nlink)
3503                         ext4_orphan_del(NULL, inode);
3504         }
3505         return ret;
3506 }
3507
3508 const struct iomap_ops ext4_iomap_ops = {
3509         .iomap_begin            = ext4_iomap_begin,
3510         .iomap_end              = ext4_iomap_end,
3511 };
3512
3513 #endif
3514
3515 static int ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3516                             ssize_t size, void *private)
3517 {
3518         ext4_io_end_t *io_end = private;
3519
3520         /* if not async direct IO just return */
3521         if (!io_end)
3522                 return 0;
3523
3524         ext_debug("ext4_end_io_dio(): io_end 0x%p "
3525                   "for inode %lu, iocb 0x%p, offset %llu, size %zd\n",
3526                   io_end, io_end->inode->i_ino, iocb, offset, size);
3527
3528         /*
3529          * Error during AIO DIO. We cannot convert unwritten extents as the
3530          * data was not written. Just clear the unwritten flag and drop io_end.
3531          */
3532         if (size <= 0) {
3533                 ext4_clear_io_unwritten_flag(io_end);
3534                 size = 0;
3535         }
3536         io_end->offset = offset;
3537         io_end->size = size;
3538         ext4_put_io_end(io_end);
3539
3540         return 0;
3541 }
3542
3543 /*
3544  * Handling of direct IO writes.
3545  *
3546  * For ext4 extent files, ext4 will do direct-io write even to holes,
3547  * preallocated extents, and those write extend the file, no need to
3548  * fall back to buffered IO.
3549  *
3550  * For holes, we fallocate those blocks, mark them as unwritten
3551  * If those blocks were preallocated, we mark sure they are split, but
3552  * still keep the range to write as unwritten.
3553  *
3554  * The unwritten extents will be converted to written when DIO is completed.
3555  * For async direct IO, since the IO may still pending when return, we
3556  * set up an end_io call back function, which will do the conversion
3557  * when async direct IO completed.
3558  *
3559  * If the O_DIRECT write will extend the file then add this inode to the
3560  * orphan list.  So recovery will truncate it back to the original size
3561  * if the machine crashes during the write.
3562  *
3563  */
3564 static ssize_t ext4_direct_IO_write(struct kiocb *iocb, struct iov_iter *iter)
3565 {
3566         struct file *file = iocb->ki_filp;
3567         struct inode *inode = file->f_mapping->host;
3568         struct ext4_inode_info *ei = EXT4_I(inode);
3569         ssize_t ret;
3570         loff_t offset = iocb->ki_pos;
3571         size_t count = iov_iter_count(iter);
3572         int overwrite = 0;
3573         get_block_t *get_block_func = NULL;
3574         int dio_flags = 0;
3575         loff_t final_size = offset + count;
3576         int orphan = 0;
3577         handle_t *handle;
3578
3579         if (final_size > inode->i_size) {
3580                 /* Credits for sb + inode write */
3581                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
3582                 if (IS_ERR(handle)) {
3583                         ret = PTR_ERR(handle);
3584                         goto out;
3585                 }
3586                 ret = ext4_orphan_add(handle, inode);
3587                 if (ret) {
3588                         ext4_journal_stop(handle);
3589                         goto out;
3590                 }
3591                 orphan = 1;
3592                 ei->i_disksize = inode->i_size;
3593                 ext4_journal_stop(handle);
3594         }
3595
3596         BUG_ON(iocb->private == NULL);
3597
3598         /*
3599          * Make all waiters for direct IO properly wait also for extent
3600          * conversion. This also disallows race between truncate() and
3601          * overwrite DIO as i_dio_count needs to be incremented under i_mutex.
3602          */
3603         inode_dio_begin(inode);
3604
3605         /* If we do a overwrite dio, i_mutex locking can be released */
3606         overwrite = *((int *)iocb->private);
3607
3608         if (overwrite)
3609                 inode_unlock(inode);
3610
3611         /*
3612          * For extent mapped files we could direct write to holes and fallocate.
3613          *
3614          * Allocated blocks to fill the hole are marked as unwritten to prevent
3615          * parallel buffered read to expose the stale data before DIO complete
3616          * the data IO.
3617          *
3618          * As to previously fallocated extents, ext4 get_block will just simply
3619          * mark the buffer mapped but still keep the extents unwritten.
3620          *
3621          * For non AI