Merge tag 'ext4_for_linus_stable' of git://git.kernel.org/pub/scm/linux/kernel/git...
[muen/linux.git] / fs / ext4 / inode.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  linux/fs/ext4/inode.c
4  *
5  * Copyright (C) 1992, 1993, 1994, 1995
6  * Remy Card (card@masi.ibp.fr)
7  * Laboratoire MASI - Institut Blaise Pascal
8  * Universite Pierre et Marie Curie (Paris VI)
9  *
10  *  from
11  *
12  *  linux/fs/minix/inode.c
13  *
14  *  Copyright (C) 1991, 1992  Linus Torvalds
15  *
16  *  64-bit file support on 64-bit platforms by Jakub Jelinek
17  *      (jj@sunsite.ms.mff.cuni.cz)
18  *
19  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
20  */
21
22 #include <linux/fs.h>
23 #include <linux/time.h>
24 #include <linux/highuid.h>
25 #include <linux/pagemap.h>
26 #include <linux/dax.h>
27 #include <linux/quotaops.h>
28 #include <linux/string.h>
29 #include <linux/buffer_head.h>
30 #include <linux/writeback.h>
31 #include <linux/pagevec.h>
32 #include <linux/mpage.h>
33 #include <linux/namei.h>
34 #include <linux/uio.h>
35 #include <linux/bio.h>
36 #include <linux/workqueue.h>
37 #include <linux/kernel.h>
38 #include <linux/printk.h>
39 #include <linux/slab.h>
40 #include <linux/bitops.h>
41 #include <linux/iomap.h>
42 #include <linux/iversion.h>
43
44 #include "ext4_jbd2.h"
45 #include "xattr.h"
46 #include "acl.h"
47 #include "truncate.h"
48
49 #include <trace/events/ext4.h>
50
51 #define MPAGE_DA_EXTENT_TAIL 0x01
52
53 static __u32 ext4_inode_csum(struct inode *inode, struct ext4_inode *raw,
54                               struct ext4_inode_info *ei)
55 {
56         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
57         __u32 csum;
58         __u16 dummy_csum = 0;
59         int offset = offsetof(struct ext4_inode, i_checksum_lo);
60         unsigned int csum_size = sizeof(dummy_csum);
61
62         csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)raw, offset);
63         csum = ext4_chksum(sbi, csum, (__u8 *)&dummy_csum, csum_size);
64         offset += csum_size;
65         csum = ext4_chksum(sbi, csum, (__u8 *)raw + offset,
66                            EXT4_GOOD_OLD_INODE_SIZE - offset);
67
68         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
69                 offset = offsetof(struct ext4_inode, i_checksum_hi);
70                 csum = ext4_chksum(sbi, csum, (__u8 *)raw +
71                                    EXT4_GOOD_OLD_INODE_SIZE,
72                                    offset - EXT4_GOOD_OLD_INODE_SIZE);
73                 if (EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi)) {
74                         csum = ext4_chksum(sbi, csum, (__u8 *)&dummy_csum,
75                                            csum_size);
76                         offset += csum_size;
77                 }
78                 csum = ext4_chksum(sbi, csum, (__u8 *)raw + offset,
79                                    EXT4_INODE_SIZE(inode->i_sb) - offset);
80         }
81
82         return csum;
83 }
84
85 static int ext4_inode_csum_verify(struct inode *inode, struct ext4_inode *raw,
86                                   struct ext4_inode_info *ei)
87 {
88         __u32 provided, calculated;
89
90         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
91             cpu_to_le32(EXT4_OS_LINUX) ||
92             !ext4_has_metadata_csum(inode->i_sb))
93                 return 1;
94
95         provided = le16_to_cpu(raw->i_checksum_lo);
96         calculated = ext4_inode_csum(inode, raw, ei);
97         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
98             EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
99                 provided |= ((__u32)le16_to_cpu(raw->i_checksum_hi)) << 16;
100         else
101                 calculated &= 0xFFFF;
102
103         return provided == calculated;
104 }
105
106 static void ext4_inode_csum_set(struct inode *inode, struct ext4_inode *raw,
107                                 struct ext4_inode_info *ei)
108 {
109         __u32 csum;
110
111         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
112             cpu_to_le32(EXT4_OS_LINUX) ||
113             !ext4_has_metadata_csum(inode->i_sb))
114                 return;
115
116         csum = ext4_inode_csum(inode, raw, ei);
117         raw->i_checksum_lo = cpu_to_le16(csum & 0xFFFF);
118         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
119             EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
120                 raw->i_checksum_hi = cpu_to_le16(csum >> 16);
121 }
122
123 static inline int ext4_begin_ordered_truncate(struct inode *inode,
124                                               loff_t new_size)
125 {
126         trace_ext4_begin_ordered_truncate(inode, new_size);
127         /*
128          * If jinode is zero, then we never opened the file for
129          * writing, so there's no need to call
130          * jbd2_journal_begin_ordered_truncate() since there's no
131          * outstanding writes we need to flush.
132          */
133         if (!EXT4_I(inode)->jinode)
134                 return 0;
135         return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
136                                                    EXT4_I(inode)->jinode,
137                                                    new_size);
138 }
139
140 static void ext4_invalidatepage(struct page *page, unsigned int offset,
141                                 unsigned int length);
142 static int __ext4_journalled_writepage(struct page *page, unsigned int len);
143 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
144 static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
145                                   int pextents);
146
147 /*
148  * Test whether an inode is a fast symlink.
149  * A fast symlink has its symlink data stored in ext4_inode_info->i_data.
150  */
151 int ext4_inode_is_fast_symlink(struct inode *inode)
152 {
153         if (!(EXT4_I(inode)->i_flags & EXT4_EA_INODE_FL)) {
154                 int ea_blocks = EXT4_I(inode)->i_file_acl ?
155                                 EXT4_CLUSTER_SIZE(inode->i_sb) >> 9 : 0;
156
157                 if (ext4_has_inline_data(inode))
158                         return 0;
159
160                 return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
161         }
162         return S_ISLNK(inode->i_mode) && inode->i_size &&
163                (inode->i_size < EXT4_N_BLOCKS * 4);
164 }
165
166 /*
167  * Restart the transaction associated with *handle.  This does a commit,
168  * so before we call here everything must be consistently dirtied against
169  * this transaction.
170  */
171 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
172                                  int nblocks)
173 {
174         int ret;
175
176         /*
177          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
178          * moment, get_block can be called only for blocks inside i_size since
179          * page cache has been already dropped and writes are blocked by
180          * i_mutex. So we can safely drop the i_data_sem here.
181          */
182         BUG_ON(EXT4_JOURNAL(inode) == NULL);
183         jbd_debug(2, "restarting handle %p\n", handle);
184         up_write(&EXT4_I(inode)->i_data_sem);
185         ret = ext4_journal_restart(handle, nblocks);
186         down_write(&EXT4_I(inode)->i_data_sem);
187         ext4_discard_preallocations(inode);
188
189         return ret;
190 }
191
192 /*
193  * Called at the last iput() if i_nlink is zero.
194  */
195 void ext4_evict_inode(struct inode *inode)
196 {
197         handle_t *handle;
198         int err;
199         int extra_credits = 3;
200         struct ext4_xattr_inode_array *ea_inode_array = NULL;
201
202         trace_ext4_evict_inode(inode);
203
204         if (inode->i_nlink) {
205                 /*
206                  * When journalling data dirty buffers are tracked only in the
207                  * journal. So although mm thinks everything is clean and
208                  * ready for reaping the inode might still have some pages to
209                  * write in the running transaction or waiting to be
210                  * checkpointed. Thus calling jbd2_journal_invalidatepage()
211                  * (via truncate_inode_pages()) to discard these buffers can
212                  * cause data loss. Also even if we did not discard these
213                  * buffers, we would have no way to find them after the inode
214                  * is reaped and thus user could see stale data if he tries to
215                  * read them before the transaction is checkpointed. So be
216                  * careful and force everything to disk here... We use
217                  * ei->i_datasync_tid to store the newest transaction
218                  * containing inode's data.
219                  *
220                  * Note that directories do not have this problem because they
221                  * don't use page cache.
222                  */
223                 if (inode->i_ino != EXT4_JOURNAL_INO &&
224                     ext4_should_journal_data(inode) &&
225                     (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode)) &&
226                     inode->i_data.nrpages) {
227                         journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
228                         tid_t commit_tid = EXT4_I(inode)->i_datasync_tid;
229
230                         jbd2_complete_transaction(journal, commit_tid);
231                         filemap_write_and_wait(&inode->i_data);
232                 }
233                 truncate_inode_pages_final(&inode->i_data);
234
235                 goto no_delete;
236         }
237
238         if (is_bad_inode(inode))
239                 goto no_delete;
240         dquot_initialize(inode);
241
242         if (ext4_should_order_data(inode))
243                 ext4_begin_ordered_truncate(inode, 0);
244         truncate_inode_pages_final(&inode->i_data);
245
246         /*
247          * Protect us against freezing - iput() caller didn't have to have any
248          * protection against it
249          */
250         sb_start_intwrite(inode->i_sb);
251
252         if (!IS_NOQUOTA(inode))
253                 extra_credits += EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb);
254
255         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE,
256                                  ext4_blocks_for_truncate(inode)+extra_credits);
257         if (IS_ERR(handle)) {
258                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
259                 /*
260                  * If we're going to skip the normal cleanup, we still need to
261                  * make sure that the in-core orphan linked list is properly
262                  * cleaned up.
263                  */
264                 ext4_orphan_del(NULL, inode);
265                 sb_end_intwrite(inode->i_sb);
266                 goto no_delete;
267         }
268
269         if (IS_SYNC(inode))
270                 ext4_handle_sync(handle);
271
272         /*
273          * Set inode->i_size to 0 before calling ext4_truncate(). We need
274          * special handling of symlinks here because i_size is used to
275          * determine whether ext4_inode_info->i_data contains symlink data or
276          * block mappings. Setting i_size to 0 will remove its fast symlink
277          * status. Erase i_data so that it becomes a valid empty block map.
278          */
279         if (ext4_inode_is_fast_symlink(inode))
280                 memset(EXT4_I(inode)->i_data, 0, sizeof(EXT4_I(inode)->i_data));
281         inode->i_size = 0;
282         err = ext4_mark_inode_dirty(handle, inode);
283         if (err) {
284                 ext4_warning(inode->i_sb,
285                              "couldn't mark inode dirty (err %d)", err);
286                 goto stop_handle;
287         }
288         if (inode->i_blocks) {
289                 err = ext4_truncate(inode);
290                 if (err) {
291                         ext4_error(inode->i_sb,
292                                    "couldn't truncate inode %lu (err %d)",
293                                    inode->i_ino, err);
294                         goto stop_handle;
295                 }
296         }
297
298         /* Remove xattr references. */
299         err = ext4_xattr_delete_inode(handle, inode, &ea_inode_array,
300                                       extra_credits);
301         if (err) {
302                 ext4_warning(inode->i_sb, "xattr delete (err %d)", err);
303 stop_handle:
304                 ext4_journal_stop(handle);
305                 ext4_orphan_del(NULL, inode);
306                 sb_end_intwrite(inode->i_sb);
307                 ext4_xattr_inode_array_free(ea_inode_array);
308                 goto no_delete;
309         }
310
311         /*
312          * Kill off the orphan record which ext4_truncate created.
313          * AKPM: I think this can be inside the above `if'.
314          * Note that ext4_orphan_del() has to be able to cope with the
315          * deletion of a non-existent orphan - this is because we don't
316          * know if ext4_truncate() actually created an orphan record.
317          * (Well, we could do this if we need to, but heck - it works)
318          */
319         ext4_orphan_del(handle, inode);
320         EXT4_I(inode)->i_dtime  = (__u32)ktime_get_real_seconds();
321
322         /*
323          * One subtle ordering requirement: if anything has gone wrong
324          * (transaction abort, IO errors, whatever), then we can still
325          * do these next steps (the fs will already have been marked as
326          * having errors), but we can't free the inode if the mark_dirty
327          * fails.
328          */
329         if (ext4_mark_inode_dirty(handle, inode))
330                 /* If that failed, just do the required in-core inode clear. */
331                 ext4_clear_inode(inode);
332         else
333                 ext4_free_inode(handle, inode);
334         ext4_journal_stop(handle);
335         sb_end_intwrite(inode->i_sb);
336         ext4_xattr_inode_array_free(ea_inode_array);
337         return;
338 no_delete:
339         ext4_clear_inode(inode);        /* We must guarantee clearing of inode... */
340 }
341
342 #ifdef CONFIG_QUOTA
343 qsize_t *ext4_get_reserved_space(struct inode *inode)
344 {
345         return &EXT4_I(inode)->i_reserved_quota;
346 }
347 #endif
348
349 /*
350  * Called with i_data_sem down, which is important since we can call
351  * ext4_discard_preallocations() from here.
352  */
353 void ext4_da_update_reserve_space(struct inode *inode,
354                                         int used, int quota_claim)
355 {
356         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
357         struct ext4_inode_info *ei = EXT4_I(inode);
358
359         spin_lock(&ei->i_block_reservation_lock);
360         trace_ext4_da_update_reserve_space(inode, used, quota_claim);
361         if (unlikely(used > ei->i_reserved_data_blocks)) {
362                 ext4_warning(inode->i_sb, "%s: ino %lu, used %d "
363                          "with only %d reserved data blocks",
364                          __func__, inode->i_ino, used,
365                          ei->i_reserved_data_blocks);
366                 WARN_ON(1);
367                 used = ei->i_reserved_data_blocks;
368         }
369
370         /* Update per-inode reservations */
371         ei->i_reserved_data_blocks -= used;
372         percpu_counter_sub(&sbi->s_dirtyclusters_counter, used);
373
374         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
375
376         /* Update quota subsystem for data blocks */
377         if (quota_claim)
378                 dquot_claim_block(inode, EXT4_C2B(sbi, used));
379         else {
380                 /*
381                  * We did fallocate with an offset that is already delayed
382                  * allocated. So on delayed allocated writeback we should
383                  * not re-claim the quota for fallocated blocks.
384                  */
385                 dquot_release_reservation_block(inode, EXT4_C2B(sbi, used));
386         }
387
388         /*
389          * If we have done all the pending block allocations and if
390          * there aren't any writers on the inode, we can discard the
391          * inode's preallocations.
392          */
393         if ((ei->i_reserved_data_blocks == 0) &&
394             (atomic_read(&inode->i_writecount) == 0))
395                 ext4_discard_preallocations(inode);
396 }
397
398 static int __check_block_validity(struct inode *inode, const char *func,
399                                 unsigned int line,
400                                 struct ext4_map_blocks *map)
401 {
402         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
403                                    map->m_len)) {
404                 ext4_error_inode(inode, func, line, map->m_pblk,
405                                  "lblock %lu mapped to illegal pblock %llu "
406                                  "(length %d)", (unsigned long) map->m_lblk,
407                                  map->m_pblk, map->m_len);
408                 return -EFSCORRUPTED;
409         }
410         return 0;
411 }
412
413 int ext4_issue_zeroout(struct inode *inode, ext4_lblk_t lblk, ext4_fsblk_t pblk,
414                        ext4_lblk_t len)
415 {
416         int ret;
417
418         if (ext4_encrypted_inode(inode))
419                 return fscrypt_zeroout_range(inode, lblk, pblk, len);
420
421         ret = sb_issue_zeroout(inode->i_sb, pblk, len, GFP_NOFS);
422         if (ret > 0)
423                 ret = 0;
424
425         return ret;
426 }
427
428 #define check_block_validity(inode, map)        \
429         __check_block_validity((inode), __func__, __LINE__, (map))
430
431 #ifdef ES_AGGRESSIVE_TEST
432 static void ext4_map_blocks_es_recheck(handle_t *handle,
433                                        struct inode *inode,
434                                        struct ext4_map_blocks *es_map,
435                                        struct ext4_map_blocks *map,
436                                        int flags)
437 {
438         int retval;
439
440         map->m_flags = 0;
441         /*
442          * There is a race window that the result is not the same.
443          * e.g. xfstests #223 when dioread_nolock enables.  The reason
444          * is that we lookup a block mapping in extent status tree with
445          * out taking i_data_sem.  So at the time the unwritten extent
446          * could be converted.
447          */
448         down_read(&EXT4_I(inode)->i_data_sem);
449         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
450                 retval = ext4_ext_map_blocks(handle, inode, map, flags &
451                                              EXT4_GET_BLOCKS_KEEP_SIZE);
452         } else {
453                 retval = ext4_ind_map_blocks(handle, inode, map, flags &
454                                              EXT4_GET_BLOCKS_KEEP_SIZE);
455         }
456         up_read((&EXT4_I(inode)->i_data_sem));
457
458         /*
459          * We don't check m_len because extent will be collpased in status
460          * tree.  So the m_len might not equal.
461          */
462         if (es_map->m_lblk != map->m_lblk ||
463             es_map->m_flags != map->m_flags ||
464             es_map->m_pblk != map->m_pblk) {
465                 printk("ES cache assertion failed for inode: %lu "
466                        "es_cached ex [%d/%d/%llu/%x] != "
467                        "found ex [%d/%d/%llu/%x] retval %d flags %x\n",
468                        inode->i_ino, es_map->m_lblk, es_map->m_len,
469                        es_map->m_pblk, es_map->m_flags, map->m_lblk,
470                        map->m_len, map->m_pblk, map->m_flags,
471                        retval, flags);
472         }
473 }
474 #endif /* ES_AGGRESSIVE_TEST */
475
476 /*
477  * The ext4_map_blocks() function tries to look up the requested blocks,
478  * and returns if the blocks are already mapped.
479  *
480  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
481  * and store the allocated blocks in the result buffer head and mark it
482  * mapped.
483  *
484  * If file type is extents based, it will call ext4_ext_map_blocks(),
485  * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
486  * based files
487  *
488  * On success, it returns the number of blocks being mapped or allocated.  if
489  * create==0 and the blocks are pre-allocated and unwritten, the resulting @map
490  * is marked as unwritten. If the create == 1, it will mark @map as mapped.
491  *
492  * It returns 0 if plain look up failed (blocks have not been allocated), in
493  * that case, @map is returned as unmapped but we still do fill map->m_len to
494  * indicate the length of a hole starting at map->m_lblk.
495  *
496  * It returns the error in case of allocation failure.
497  */
498 int ext4_map_blocks(handle_t *handle, struct inode *inode,
499                     struct ext4_map_blocks *map, int flags)
500 {
501         struct extent_status es;
502         int retval;
503         int ret = 0;
504 #ifdef ES_AGGRESSIVE_TEST
505         struct ext4_map_blocks orig_map;
506
507         memcpy(&orig_map, map, sizeof(*map));
508 #endif
509
510         map->m_flags = 0;
511         ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
512                   "logical block %lu\n", inode->i_ino, flags, map->m_len,
513                   (unsigned long) map->m_lblk);
514
515         /*
516          * ext4_map_blocks returns an int, and m_len is an unsigned int
517          */
518         if (unlikely(map->m_len > INT_MAX))
519                 map->m_len = INT_MAX;
520
521         /* We can handle the block number less than EXT_MAX_BLOCKS */
522         if (unlikely(map->m_lblk >= EXT_MAX_BLOCKS))
523                 return -EFSCORRUPTED;
524
525         /* Lookup extent status tree firstly */
526         if (ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
527                 if (ext4_es_is_written(&es) || ext4_es_is_unwritten(&es)) {
528                         map->m_pblk = ext4_es_pblock(&es) +
529                                         map->m_lblk - es.es_lblk;
530                         map->m_flags |= ext4_es_is_written(&es) ?
531                                         EXT4_MAP_MAPPED : EXT4_MAP_UNWRITTEN;
532                         retval = es.es_len - (map->m_lblk - es.es_lblk);
533                         if (retval > map->m_len)
534                                 retval = map->m_len;
535                         map->m_len = retval;
536                 } else if (ext4_es_is_delayed(&es) || ext4_es_is_hole(&es)) {
537                         map->m_pblk = 0;
538                         retval = es.es_len - (map->m_lblk - es.es_lblk);
539                         if (retval > map->m_len)
540                                 retval = map->m_len;
541                         map->m_len = retval;
542                         retval = 0;
543                 } else {
544                         BUG_ON(1);
545                 }
546 #ifdef ES_AGGRESSIVE_TEST
547                 ext4_map_blocks_es_recheck(handle, inode, map,
548                                            &orig_map, flags);
549 #endif
550                 goto found;
551         }
552
553         /*
554          * Try to see if we can get the block without requesting a new
555          * file system block.
556          */
557         down_read(&EXT4_I(inode)->i_data_sem);
558         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
559                 retval = ext4_ext_map_blocks(handle, inode, map, flags &
560                                              EXT4_GET_BLOCKS_KEEP_SIZE);
561         } else {
562                 retval = ext4_ind_map_blocks(handle, inode, map, flags &
563                                              EXT4_GET_BLOCKS_KEEP_SIZE);
564         }
565         if (retval > 0) {
566                 unsigned int status;
567
568                 if (unlikely(retval != map->m_len)) {
569                         ext4_warning(inode->i_sb,
570                                      "ES len assertion failed for inode "
571                                      "%lu: retval %d != map->m_len %d",
572                                      inode->i_ino, retval, map->m_len);
573                         WARN_ON(1);
574                 }
575
576                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
577                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
578                 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
579                     !(status & EXTENT_STATUS_WRITTEN) &&
580                     ext4_es_scan_range(inode, &ext4_es_is_delayed, map->m_lblk,
581                                        map->m_lblk + map->m_len - 1))
582                         status |= EXTENT_STATUS_DELAYED;
583                 ret = ext4_es_insert_extent(inode, map->m_lblk,
584                                             map->m_len, map->m_pblk, status);
585                 if (ret < 0)
586                         retval = ret;
587         }
588         up_read((&EXT4_I(inode)->i_data_sem));
589
590 found:
591         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
592                 ret = check_block_validity(inode, map);
593                 if (ret != 0)
594                         return ret;
595         }
596
597         /* If it is only a block(s) look up */
598         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
599                 return retval;
600
601         /*
602          * Returns if the blocks have already allocated
603          *
604          * Note that if blocks have been preallocated
605          * ext4_ext_get_block() returns the create = 0
606          * with buffer head unmapped.
607          */
608         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
609                 /*
610                  * If we need to convert extent to unwritten
611                  * we continue and do the actual work in
612                  * ext4_ext_map_blocks()
613                  */
614                 if (!(flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN))
615                         return retval;
616
617         /*
618          * Here we clear m_flags because after allocating an new extent,
619          * it will be set again.
620          */
621         map->m_flags &= ~EXT4_MAP_FLAGS;
622
623         /*
624          * New blocks allocate and/or writing to unwritten extent
625          * will possibly result in updating i_data, so we take
626          * the write lock of i_data_sem, and call get_block()
627          * with create == 1 flag.
628          */
629         down_write(&EXT4_I(inode)->i_data_sem);
630
631         /*
632          * We need to check for EXT4 here because migrate
633          * could have changed the inode type in between
634          */
635         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
636                 retval = ext4_ext_map_blocks(handle, inode, map, flags);
637         } else {
638                 retval = ext4_ind_map_blocks(handle, inode, map, flags);
639
640                 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
641                         /*
642                          * We allocated new blocks which will result in
643                          * i_data's format changing.  Force the migrate
644                          * to fail by clearing migrate flags
645                          */
646                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
647                 }
648
649                 /*
650                  * Update reserved blocks/metadata blocks after successful
651                  * block allocation which had been deferred till now. We don't
652                  * support fallocate for non extent files. So we can update
653                  * reserve space here.
654                  */
655                 if ((retval > 0) &&
656                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
657                         ext4_da_update_reserve_space(inode, retval, 1);
658         }
659
660         if (retval > 0) {
661                 unsigned int status;
662
663                 if (unlikely(retval != map->m_len)) {
664                         ext4_warning(inode->i_sb,
665                                      "ES len assertion failed for inode "
666                                      "%lu: retval %d != map->m_len %d",
667                                      inode->i_ino, retval, map->m_len);
668                         WARN_ON(1);
669                 }
670
671                 /*
672                  * We have to zeroout blocks before inserting them into extent
673                  * status tree. Otherwise someone could look them up there and
674                  * use them before they are really zeroed. We also have to
675                  * unmap metadata before zeroing as otherwise writeback can
676                  * overwrite zeros with stale data from block device.
677                  */
678                 if (flags & EXT4_GET_BLOCKS_ZERO &&
679                     map->m_flags & EXT4_MAP_MAPPED &&
680                     map->m_flags & EXT4_MAP_NEW) {
681                         clean_bdev_aliases(inode->i_sb->s_bdev, map->m_pblk,
682                                            map->m_len);
683                         ret = ext4_issue_zeroout(inode, map->m_lblk,
684                                                  map->m_pblk, map->m_len);
685                         if (ret) {
686                                 retval = ret;
687                                 goto out_sem;
688                         }
689                 }
690
691                 /*
692                  * If the extent has been zeroed out, we don't need to update
693                  * extent status tree.
694                  */
695                 if ((flags & EXT4_GET_BLOCKS_PRE_IO) &&
696                     ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
697                         if (ext4_es_is_written(&es))
698                                 goto out_sem;
699                 }
700                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
701                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
702                 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
703                     !(status & EXTENT_STATUS_WRITTEN) &&
704                     ext4_es_scan_range(inode, &ext4_es_is_delayed, map->m_lblk,
705                                        map->m_lblk + map->m_len - 1))
706                         status |= EXTENT_STATUS_DELAYED;
707                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
708                                             map->m_pblk, status);
709                 if (ret < 0) {
710                         retval = ret;
711                         goto out_sem;
712                 }
713         }
714
715 out_sem:
716         up_write((&EXT4_I(inode)->i_data_sem));
717         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
718                 ret = check_block_validity(inode, map);
719                 if (ret != 0)
720                         return ret;
721
722                 /*
723                  * Inodes with freshly allocated blocks where contents will be
724                  * visible after transaction commit must be on transaction's
725                  * ordered data list.
726                  */
727                 if (map->m_flags & EXT4_MAP_NEW &&
728                     !(map->m_flags & EXT4_MAP_UNWRITTEN) &&
729                     !(flags & EXT4_GET_BLOCKS_ZERO) &&
730                     !ext4_is_quota_file(inode) &&
731                     ext4_should_order_data(inode)) {
732                         if (flags & EXT4_GET_BLOCKS_IO_SUBMIT)
733                                 ret = ext4_jbd2_inode_add_wait(handle, inode);
734                         else
735                                 ret = ext4_jbd2_inode_add_write(handle, inode);
736                         if (ret)
737                                 return ret;
738                 }
739         }
740         return retval;
741 }
742
743 /*
744  * Update EXT4_MAP_FLAGS in bh->b_state. For buffer heads attached to pages
745  * we have to be careful as someone else may be manipulating b_state as well.
746  */
747 static void ext4_update_bh_state(struct buffer_head *bh, unsigned long flags)
748 {
749         unsigned long old_state;
750         unsigned long new_state;
751
752         flags &= EXT4_MAP_FLAGS;
753
754         /* Dummy buffer_head? Set non-atomically. */
755         if (!bh->b_page) {
756                 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | flags;
757                 return;
758         }
759         /*
760          * Someone else may be modifying b_state. Be careful! This is ugly but
761          * once we get rid of using bh as a container for mapping information
762          * to pass to / from get_block functions, this can go away.
763          */
764         do {
765                 old_state = READ_ONCE(bh->b_state);
766                 new_state = (old_state & ~EXT4_MAP_FLAGS) | flags;
767         } while (unlikely(
768                  cmpxchg(&bh->b_state, old_state, new_state) != old_state));
769 }
770
771 static int _ext4_get_block(struct inode *inode, sector_t iblock,
772                            struct buffer_head *bh, int flags)
773 {
774         struct ext4_map_blocks map;
775         int ret = 0;
776
777         if (ext4_has_inline_data(inode))
778                 return -ERANGE;
779
780         map.m_lblk = iblock;
781         map.m_len = bh->b_size >> inode->i_blkbits;
782
783         ret = ext4_map_blocks(ext4_journal_current_handle(), inode, &map,
784                               flags);
785         if (ret > 0) {
786                 map_bh(bh, inode->i_sb, map.m_pblk);
787                 ext4_update_bh_state(bh, map.m_flags);
788                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
789                 ret = 0;
790         } else if (ret == 0) {
791                 /* hole case, need to fill in bh->b_size */
792                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
793         }
794         return ret;
795 }
796
797 int ext4_get_block(struct inode *inode, sector_t iblock,
798                    struct buffer_head *bh, int create)
799 {
800         return _ext4_get_block(inode, iblock, bh,
801                                create ? EXT4_GET_BLOCKS_CREATE : 0);
802 }
803
804 /*
805  * Get block function used when preparing for buffered write if we require
806  * creating an unwritten extent if blocks haven't been allocated.  The extent
807  * will be converted to written after the IO is complete.
808  */
809 int ext4_get_block_unwritten(struct inode *inode, sector_t iblock,
810                              struct buffer_head *bh_result, int create)
811 {
812         ext4_debug("ext4_get_block_unwritten: inode %lu, create flag %d\n",
813                    inode->i_ino, create);
814         return _ext4_get_block(inode, iblock, bh_result,
815                                EXT4_GET_BLOCKS_IO_CREATE_EXT);
816 }
817
818 /* Maximum number of blocks we map for direct IO at once. */
819 #define DIO_MAX_BLOCKS 4096
820
821 /*
822  * Get blocks function for the cases that need to start a transaction -
823  * generally difference cases of direct IO and DAX IO. It also handles retries
824  * in case of ENOSPC.
825  */
826 static int ext4_get_block_trans(struct inode *inode, sector_t iblock,
827                                 struct buffer_head *bh_result, int flags)
828 {
829         int dio_credits;
830         handle_t *handle;
831         int retries = 0;
832         int ret;
833
834         /* Trim mapping request to maximum we can map at once for DIO */
835         if (bh_result->b_size >> inode->i_blkbits > DIO_MAX_BLOCKS)
836                 bh_result->b_size = DIO_MAX_BLOCKS << inode->i_blkbits;
837         dio_credits = ext4_chunk_trans_blocks(inode,
838                                       bh_result->b_size >> inode->i_blkbits);
839 retry:
840         handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, dio_credits);
841         if (IS_ERR(handle))
842                 return PTR_ERR(handle);
843
844         ret = _ext4_get_block(inode, iblock, bh_result, flags);
845         ext4_journal_stop(handle);
846
847         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
848                 goto retry;
849         return ret;
850 }
851
852 /* Get block function for DIO reads and writes to inodes without extents */
853 int ext4_dio_get_block(struct inode *inode, sector_t iblock,
854                        struct buffer_head *bh, int create)
855 {
856         /* We don't expect handle for direct IO */
857         WARN_ON_ONCE(ext4_journal_current_handle());
858
859         if (!create)
860                 return _ext4_get_block(inode, iblock, bh, 0);
861         return ext4_get_block_trans(inode, iblock, bh, EXT4_GET_BLOCKS_CREATE);
862 }
863
864 /*
865  * Get block function for AIO DIO writes when we create unwritten extent if
866  * blocks are not allocated yet. The extent will be converted to written
867  * after IO is complete.
868  */
869 static int ext4_dio_get_block_unwritten_async(struct inode *inode,
870                 sector_t iblock, struct buffer_head *bh_result, int create)
871 {
872         int ret;
873
874         /* We don't expect handle for direct IO */
875         WARN_ON_ONCE(ext4_journal_current_handle());
876
877         ret = ext4_get_block_trans(inode, iblock, bh_result,
878                                    EXT4_GET_BLOCKS_IO_CREATE_EXT);
879
880         /*
881          * When doing DIO using unwritten extents, we need io_end to convert
882          * unwritten extents to written on IO completion. We allocate io_end
883          * once we spot unwritten extent and store it in b_private. Generic
884          * DIO code keeps b_private set and furthermore passes the value to
885          * our completion callback in 'private' argument.
886          */
887         if (!ret && buffer_unwritten(bh_result)) {
888                 if (!bh_result->b_private) {
889                         ext4_io_end_t *io_end;
890
891                         io_end = ext4_init_io_end(inode, GFP_KERNEL);
892                         if (!io_end)
893                                 return -ENOMEM;
894                         bh_result->b_private = io_end;
895                         ext4_set_io_unwritten_flag(inode, io_end);
896                 }
897                 set_buffer_defer_completion(bh_result);
898         }
899
900         return ret;
901 }
902
903 /*
904  * Get block function for non-AIO DIO writes when we create unwritten extent if
905  * blocks are not allocated yet. The extent will be converted to written
906  * after IO is complete by ext4_direct_IO_write().
907  */
908 static int ext4_dio_get_block_unwritten_sync(struct inode *inode,
909                 sector_t iblock, struct buffer_head *bh_result, int create)
910 {
911         int ret;
912
913         /* We don't expect handle for direct IO */
914         WARN_ON_ONCE(ext4_journal_current_handle());
915
916         ret = ext4_get_block_trans(inode, iblock, bh_result,
917                                    EXT4_GET_BLOCKS_IO_CREATE_EXT);
918
919         /*
920          * Mark inode as having pending DIO writes to unwritten extents.
921          * ext4_direct_IO_write() checks this flag and converts extents to
922          * written.
923          */
924         if (!ret && buffer_unwritten(bh_result))
925                 ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
926
927         return ret;
928 }
929
930 static int ext4_dio_get_block_overwrite(struct inode *inode, sector_t iblock,
931                    struct buffer_head *bh_result, int create)
932 {
933         int ret;
934
935         ext4_debug("ext4_dio_get_block_overwrite: inode %lu, create flag %d\n",
936                    inode->i_ino, create);
937         /* We don't expect handle for direct IO */
938         WARN_ON_ONCE(ext4_journal_current_handle());
939
940         ret = _ext4_get_block(inode, iblock, bh_result, 0);
941         /*
942          * Blocks should have been preallocated! ext4_file_write_iter() checks
943          * that.
944          */
945         WARN_ON_ONCE(!buffer_mapped(bh_result) || buffer_unwritten(bh_result));
946
947         return ret;
948 }
949
950
951 /*
952  * `handle' can be NULL if create is zero
953  */
954 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
955                                 ext4_lblk_t block, int map_flags)
956 {
957         struct ext4_map_blocks map;
958         struct buffer_head *bh;
959         int create = map_flags & EXT4_GET_BLOCKS_CREATE;
960         int err;
961
962         J_ASSERT(handle != NULL || create == 0);
963
964         map.m_lblk = block;
965         map.m_len = 1;
966         err = ext4_map_blocks(handle, inode, &map, map_flags);
967
968         if (err == 0)
969                 return create ? ERR_PTR(-ENOSPC) : NULL;
970         if (err < 0)
971                 return ERR_PTR(err);
972
973         bh = sb_getblk(inode->i_sb, map.m_pblk);
974         if (unlikely(!bh))
975                 return ERR_PTR(-ENOMEM);
976         if (map.m_flags & EXT4_MAP_NEW) {
977                 J_ASSERT(create != 0);
978                 J_ASSERT(handle != NULL);
979
980                 /*
981                  * Now that we do not always journal data, we should
982                  * keep in mind whether this should always journal the
983                  * new buffer as metadata.  For now, regular file
984                  * writes use ext4_get_block instead, so it's not a
985                  * problem.
986                  */
987                 lock_buffer(bh);
988                 BUFFER_TRACE(bh, "call get_create_access");
989                 err = ext4_journal_get_create_access(handle, bh);
990                 if (unlikely(err)) {
991                         unlock_buffer(bh);
992                         goto errout;
993                 }
994                 if (!buffer_uptodate(bh)) {
995                         memset(bh->b_data, 0, inode->i_sb->s_blocksize);
996                         set_buffer_uptodate(bh);
997                 }
998                 unlock_buffer(bh);
999                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1000                 err = ext4_handle_dirty_metadata(handle, inode, bh);
1001                 if (unlikely(err))
1002                         goto errout;
1003         } else
1004                 BUFFER_TRACE(bh, "not a new buffer");
1005         return bh;
1006 errout:
1007         brelse(bh);
1008         return ERR_PTR(err);
1009 }
1010
1011 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1012                                ext4_lblk_t block, int map_flags)
1013 {
1014         struct buffer_head *bh;
1015
1016         bh = ext4_getblk(handle, inode, block, map_flags);
1017         if (IS_ERR(bh))
1018                 return bh;
1019         if (!bh || buffer_uptodate(bh))
1020                 return bh;
1021         ll_rw_block(REQ_OP_READ, REQ_META | REQ_PRIO, 1, &bh);
1022         wait_on_buffer(bh);
1023         if (buffer_uptodate(bh))
1024                 return bh;
1025         put_bh(bh);
1026         return ERR_PTR(-EIO);
1027 }
1028
1029 /* Read a contiguous batch of blocks. */
1030 int ext4_bread_batch(struct inode *inode, ext4_lblk_t block, int bh_count,
1031                      bool wait, struct buffer_head **bhs)
1032 {
1033         int i, err;
1034
1035         for (i = 0; i < bh_count; i++) {
1036                 bhs[i] = ext4_getblk(NULL, inode, block + i, 0 /* map_flags */);
1037                 if (IS_ERR(bhs[i])) {
1038                         err = PTR_ERR(bhs[i]);
1039                         bh_count = i;
1040                         goto out_brelse;
1041                 }
1042         }
1043
1044         for (i = 0; i < bh_count; i++)
1045                 /* Note that NULL bhs[i] is valid because of holes. */
1046                 if (bhs[i] && !buffer_uptodate(bhs[i]))
1047                         ll_rw_block(REQ_OP_READ, REQ_META | REQ_PRIO, 1,
1048                                     &bhs[i]);
1049
1050         if (!wait)
1051                 return 0;
1052
1053         for (i = 0; i < bh_count; i++)
1054                 if (bhs[i])
1055                         wait_on_buffer(bhs[i]);
1056
1057         for (i = 0; i < bh_count; i++) {
1058                 if (bhs[i] && !buffer_uptodate(bhs[i])) {
1059                         err = -EIO;
1060                         goto out_brelse;
1061                 }
1062         }
1063         return 0;
1064
1065 out_brelse:
1066         for (i = 0; i < bh_count; i++) {
1067                 brelse(bhs[i]);
1068                 bhs[i] = NULL;
1069         }
1070         return err;
1071 }
1072
1073 int ext4_walk_page_buffers(handle_t *handle,
1074                            struct buffer_head *head,
1075                            unsigned from,
1076                            unsigned to,
1077                            int *partial,
1078                            int (*fn)(handle_t *handle,
1079                                      struct buffer_head *bh))
1080 {
1081         struct buffer_head *bh;
1082         unsigned block_start, block_end;
1083         unsigned blocksize = head->b_size;
1084         int err, ret = 0;
1085         struct buffer_head *next;
1086
1087         for (bh = head, block_start = 0;
1088              ret == 0 && (bh != head || !block_start);
1089              block_start = block_end, bh = next) {
1090                 next = bh->b_this_page;
1091                 block_end = block_start + blocksize;
1092                 if (block_end <= from || block_start >= to) {
1093                         if (partial && !buffer_uptodate(bh))
1094                                 *partial = 1;
1095                         continue;
1096                 }
1097                 err = (*fn)(handle, bh);
1098                 if (!ret)
1099                         ret = err;
1100         }
1101         return ret;
1102 }
1103
1104 /*
1105  * To preserve ordering, it is essential that the hole instantiation and
1106  * the data write be encapsulated in a single transaction.  We cannot
1107  * close off a transaction and start a new one between the ext4_get_block()
1108  * and the commit_write().  So doing the jbd2_journal_start at the start of
1109  * prepare_write() is the right place.
1110  *
1111  * Also, this function can nest inside ext4_writepage().  In that case, we
1112  * *know* that ext4_writepage() has generated enough buffer credits to do the
1113  * whole page.  So we won't block on the journal in that case, which is good,
1114  * because the caller may be PF_MEMALLOC.
1115  *
1116  * By accident, ext4 can be reentered when a transaction is open via
1117  * quota file writes.  If we were to commit the transaction while thus
1118  * reentered, there can be a deadlock - we would be holding a quota
1119  * lock, and the commit would never complete if another thread had a
1120  * transaction open and was blocking on the quota lock - a ranking
1121  * violation.
1122  *
1123  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1124  * will _not_ run commit under these circumstances because handle->h_ref
1125  * is elevated.  We'll still have enough credits for the tiny quotafile
1126  * write.
1127  */
1128 int do_journal_get_write_access(handle_t *handle,
1129                                 struct buffer_head *bh)
1130 {
1131         int dirty = buffer_dirty(bh);
1132         int ret;
1133
1134         if (!buffer_mapped(bh) || buffer_freed(bh))
1135                 return 0;
1136         /*
1137          * __block_write_begin() could have dirtied some buffers. Clean
1138          * the dirty bit as jbd2_journal_get_write_access() could complain
1139          * otherwise about fs integrity issues. Setting of the dirty bit
1140          * by __block_write_begin() isn't a real problem here as we clear
1141          * the bit before releasing a page lock and thus writeback cannot
1142          * ever write the buffer.
1143          */
1144         if (dirty)
1145                 clear_buffer_dirty(bh);
1146         BUFFER_TRACE(bh, "get write access");
1147         ret = ext4_journal_get_write_access(handle, bh);
1148         if (!ret && dirty)
1149                 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1150         return ret;
1151 }
1152
1153 #ifdef CONFIG_EXT4_FS_ENCRYPTION
1154 static int ext4_block_write_begin(struct page *page, loff_t pos, unsigned len,
1155                                   get_block_t *get_block)
1156 {
1157         unsigned from = pos & (PAGE_SIZE - 1);
1158         unsigned to = from + len;
1159         struct inode *inode = page->mapping->host;
1160         unsigned block_start, block_end;
1161         sector_t block;
1162         int err = 0;
1163         unsigned blocksize = inode->i_sb->s_blocksize;
1164         unsigned bbits;
1165         struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
1166         bool decrypt = false;
1167
1168         BUG_ON(!PageLocked(page));
1169         BUG_ON(from > PAGE_SIZE);
1170         BUG_ON(to > PAGE_SIZE);
1171         BUG_ON(from > to);
1172
1173         if (!page_has_buffers(page))
1174                 create_empty_buffers(page, blocksize, 0);
1175         head = page_buffers(page);
1176         bbits = ilog2(blocksize);
1177         block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1178
1179         for (bh = head, block_start = 0; bh != head || !block_start;
1180             block++, block_start = block_end, bh = bh->b_this_page) {
1181                 block_end = block_start + blocksize;
1182                 if (block_end <= from || block_start >= to) {
1183                         if (PageUptodate(page)) {
1184                                 if (!buffer_uptodate(bh))
1185                                         set_buffer_uptodate(bh);
1186                         }
1187                         continue;
1188                 }
1189                 if (buffer_new(bh))
1190                         clear_buffer_new(bh);
1191                 if (!buffer_mapped(bh)) {
1192                         WARN_ON(bh->b_size != blocksize);
1193                         err = get_block(inode, block, bh, 1);
1194                         if (err)
1195                                 break;
1196                         if (buffer_new(bh)) {
1197                                 clean_bdev_bh_alias(bh);
1198                                 if (PageUptodate(page)) {
1199                                         clear_buffer_new(bh);
1200                                         set_buffer_uptodate(bh);
1201                                         mark_buffer_dirty(bh);
1202                                         continue;
1203                                 }
1204                                 if (block_end > to || block_start < from)
1205                                         zero_user_segments(page, to, block_end,
1206                                                            block_start, from);
1207                                 continue;
1208                         }
1209                 }
1210                 if (PageUptodate(page)) {
1211                         if (!buffer_uptodate(bh))
1212                                 set_buffer_uptodate(bh);
1213                         continue;
1214                 }
1215                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1216                     !buffer_unwritten(bh) &&
1217                     (block_start < from || block_end > to)) {
1218                         ll_rw_block(REQ_OP_READ, 0, 1, &bh);
1219                         *wait_bh++ = bh;
1220                         decrypt = ext4_encrypted_inode(inode) &&
1221                                 S_ISREG(inode->i_mode);
1222                 }
1223         }
1224         /*
1225          * If we issued read requests, let them complete.
1226          */
1227         while (wait_bh > wait) {
1228                 wait_on_buffer(*--wait_bh);
1229                 if (!buffer_uptodate(*wait_bh))
1230                         err = -EIO;
1231         }
1232         if (unlikely(err))
1233                 page_zero_new_buffers(page, from, to);
1234         else if (decrypt)
1235                 err = fscrypt_decrypt_page(page->mapping->host, page,
1236                                 PAGE_SIZE, 0, page->index);
1237         return err;
1238 }
1239 #endif
1240
1241 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1242                             loff_t pos, unsigned len, unsigned flags,
1243                             struct page **pagep, void **fsdata)
1244 {
1245         struct inode *inode = mapping->host;
1246         int ret, needed_blocks;
1247         handle_t *handle;
1248         int retries = 0;
1249         struct page *page;
1250         pgoff_t index;
1251         unsigned from, to;
1252
1253         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
1254                 return -EIO;
1255
1256         trace_ext4_write_begin(inode, pos, len, flags);
1257         /*
1258          * Reserve one block more for addition to orphan list in case
1259          * we allocate blocks but write fails for some reason
1260          */
1261         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1262         index = pos >> PAGE_SHIFT;
1263         from = pos & (PAGE_SIZE - 1);
1264         to = from + len;
1265
1266         if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
1267                 ret = ext4_try_to_write_inline_data(mapping, inode, pos, len,
1268                                                     flags, pagep);
1269                 if (ret < 0)
1270                         return ret;
1271                 if (ret == 1)
1272                         return 0;
1273         }
1274
1275         /*
1276          * grab_cache_page_write_begin() can take a long time if the
1277          * system is thrashing due to memory pressure, or if the page
1278          * is being written back.  So grab it first before we start
1279          * the transaction handle.  This also allows us to allocate
1280          * the page (if needed) without using GFP_NOFS.
1281          */
1282 retry_grab:
1283         page = grab_cache_page_write_begin(mapping, index, flags);
1284         if (!page)
1285                 return -ENOMEM;
1286         unlock_page(page);
1287
1288 retry_journal:
1289         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE, needed_blocks);
1290         if (IS_ERR(handle)) {
1291                 put_page(page);
1292                 return PTR_ERR(handle);
1293         }
1294
1295         lock_page(page);
1296         if (page->mapping != mapping) {
1297                 /* The page got truncated from under us */
1298                 unlock_page(page);
1299                 put_page(page);
1300                 ext4_journal_stop(handle);
1301                 goto retry_grab;
1302         }
1303         /* In case writeback began while the page was unlocked */
1304         wait_for_stable_page(page);
1305
1306 #ifdef CONFIG_EXT4_FS_ENCRYPTION
1307         if (ext4_should_dioread_nolock(inode))
1308                 ret = ext4_block_write_begin(page, pos, len,
1309                                              ext4_get_block_unwritten);
1310         else
1311                 ret = ext4_block_write_begin(page, pos, len,
1312                                              ext4_get_block);
1313 #else
1314         if (ext4_should_dioread_nolock(inode))
1315                 ret = __block_write_begin(page, pos, len,
1316                                           ext4_get_block_unwritten);
1317         else
1318                 ret = __block_write_begin(page, pos, len, ext4_get_block);
1319 #endif
1320         if (!ret && ext4_should_journal_data(inode)) {
1321                 ret = ext4_walk_page_buffers(handle, page_buffers(page),
1322                                              from, to, NULL,
1323                                              do_journal_get_write_access);
1324         }
1325
1326         if (ret) {
1327                 unlock_page(page);
1328                 /*
1329                  * __block_write_begin may have instantiated a few blocks
1330                  * outside i_size.  Trim these off again. Don't need
1331                  * i_size_read because we hold i_mutex.
1332                  *
1333                  * Add inode to orphan list in case we crash before
1334                  * truncate finishes
1335                  */
1336                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1337                         ext4_orphan_add(handle, inode);
1338
1339                 ext4_journal_stop(handle);
1340                 if (pos + len > inode->i_size) {
1341                         ext4_truncate_failed_write(inode);
1342                         /*
1343                          * If truncate failed early the inode might
1344                          * still be on the orphan list; we need to
1345                          * make sure the inode is removed from the
1346                          * orphan list in that case.
1347                          */
1348                         if (inode->i_nlink)
1349                                 ext4_orphan_del(NULL, inode);
1350                 }
1351
1352                 if (ret == -ENOSPC &&
1353                     ext4_should_retry_alloc(inode->i_sb, &retries))
1354                         goto retry_journal;
1355                 put_page(page);
1356                 return ret;
1357         }
1358         *pagep = page;
1359         return ret;
1360 }
1361
1362 /* For write_end() in data=journal mode */
1363 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1364 {
1365         int ret;
1366         if (!buffer_mapped(bh) || buffer_freed(bh))
1367                 return 0;
1368         set_buffer_uptodate(bh);
1369         ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1370         clear_buffer_meta(bh);
1371         clear_buffer_prio(bh);
1372         return ret;
1373 }
1374
1375 /*
1376  * We need to pick up the new inode size which generic_commit_write gave us
1377  * `file' can be NULL - eg, when called from page_symlink().
1378  *
1379  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1380  * buffers are managed internally.
1381  */
1382 static int ext4_write_end(struct file *file,
1383                           struct address_space *mapping,
1384                           loff_t pos, unsigned len, unsigned copied,
1385                           struct page *page, void *fsdata)
1386 {
1387         handle_t *handle = ext4_journal_current_handle();
1388         struct inode *inode = mapping->host;
1389         loff_t old_size = inode->i_size;
1390         int ret = 0, ret2;
1391         int i_size_changed = 0;
1392         int inline_data = ext4_has_inline_data(inode);
1393
1394         trace_ext4_write_end(inode, pos, len, copied);
1395         if (inline_data) {
1396                 ret = ext4_write_inline_data_end(inode, pos, len,
1397                                                  copied, page);
1398                 if (ret < 0) {
1399                         unlock_page(page);
1400                         put_page(page);
1401                         goto errout;
1402                 }
1403                 copied = ret;
1404         } else
1405                 copied = block_write_end(file, mapping, pos,
1406                                          len, copied, page, fsdata);
1407         /*
1408          * it's important to update i_size while still holding page lock:
1409          * page writeout could otherwise come in and zero beyond i_size.
1410          */
1411         i_size_changed = ext4_update_inode_size(inode, pos + copied);
1412         unlock_page(page);
1413         put_page(page);
1414
1415         if (old_size < pos)
1416                 pagecache_isize_extended(inode, old_size, pos);
1417         /*
1418          * Don't mark the inode dirty under page lock. First, it unnecessarily
1419          * makes the holding time of page lock longer. Second, it forces lock
1420          * ordering of page lock and transaction start for journaling
1421          * filesystems.
1422          */
1423         if (i_size_changed || inline_data)
1424                 ext4_mark_inode_dirty(handle, inode);
1425
1426         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1427                 /* if we have allocated more blocks and copied
1428                  * less. We will have blocks allocated outside
1429                  * inode->i_size. So truncate them
1430                  */
1431                 ext4_orphan_add(handle, inode);
1432 errout:
1433         ret2 = ext4_journal_stop(handle);
1434         if (!ret)
1435                 ret = ret2;
1436
1437         if (pos + len > inode->i_size) {
1438                 ext4_truncate_failed_write(inode);
1439                 /*
1440                  * If truncate failed early the inode might still be
1441                  * on the orphan list; we need to make sure the inode
1442                  * is removed from the orphan list in that case.
1443                  */
1444                 if (inode->i_nlink)
1445                         ext4_orphan_del(NULL, inode);
1446         }
1447
1448         return ret ? ret : copied;
1449 }
1450
1451 /*
1452  * This is a private version of page_zero_new_buffers() which doesn't
1453  * set the buffer to be dirty, since in data=journalled mode we need
1454  * to call ext4_handle_dirty_metadata() instead.
1455  */
1456 static void ext4_journalled_zero_new_buffers(handle_t *handle,
1457                                             struct page *page,
1458                                             unsigned from, unsigned to)
1459 {
1460         unsigned int block_start = 0, block_end;
1461         struct buffer_head *head, *bh;
1462
1463         bh = head = page_buffers(page);
1464         do {
1465                 block_end = block_start + bh->b_size;
1466                 if (buffer_new(bh)) {
1467                         if (block_end > from && block_start < to) {
1468                                 if (!PageUptodate(page)) {
1469                                         unsigned start, size;
1470
1471                                         start = max(from, block_start);
1472                                         size = min(to, block_end) - start;
1473
1474                                         zero_user(page, start, size);
1475                                         write_end_fn(handle, bh);
1476                                 }
1477                                 clear_buffer_new(bh);
1478                         }
1479                 }
1480                 block_start = block_end;
1481                 bh = bh->b_this_page;
1482         } while (bh != head);
1483 }
1484
1485 static int ext4_journalled_write_end(struct file *file,
1486                                      struct address_space *mapping,
1487                                      loff_t pos, unsigned len, unsigned copied,
1488                                      struct page *page, void *fsdata)
1489 {
1490         handle_t *handle = ext4_journal_current_handle();
1491         struct inode *inode = mapping->host;
1492         loff_t old_size = inode->i_size;
1493         int ret = 0, ret2;
1494         int partial = 0;
1495         unsigned from, to;
1496         int size_changed = 0;
1497         int inline_data = ext4_has_inline_data(inode);
1498
1499         trace_ext4_journalled_write_end(inode, pos, len, copied);
1500         from = pos & (PAGE_SIZE - 1);
1501         to = from + len;
1502
1503         BUG_ON(!ext4_handle_valid(handle));
1504
1505         if (inline_data) {
1506                 ret = ext4_write_inline_data_end(inode, pos, len,
1507                                                  copied, page);
1508                 if (ret < 0) {
1509                         unlock_page(page);
1510                         put_page(page);
1511                         goto errout;
1512                 }
1513                 copied = ret;
1514         } else if (unlikely(copied < len) && !PageUptodate(page)) {
1515                 copied = 0;
1516                 ext4_journalled_zero_new_buffers(handle, page, from, to);
1517         } else {
1518                 if (unlikely(copied < len))
1519                         ext4_journalled_zero_new_buffers(handle, page,
1520                                                          from + copied, to);
1521                 ret = ext4_walk_page_buffers(handle, page_buffers(page), from,
1522                                              from + copied, &partial,
1523                                              write_end_fn);
1524                 if (!partial)
1525                         SetPageUptodate(page);
1526         }
1527         size_changed = ext4_update_inode_size(inode, pos + copied);
1528         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1529         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1530         unlock_page(page);
1531         put_page(page);
1532
1533         if (old_size < pos)
1534                 pagecache_isize_extended(inode, old_size, pos);
1535
1536         if (size_changed || inline_data) {
1537                 ret2 = ext4_mark_inode_dirty(handle, inode);
1538                 if (!ret)
1539                         ret = ret2;
1540         }
1541
1542         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1543                 /* if we have allocated more blocks and copied
1544                  * less. We will have blocks allocated outside
1545                  * inode->i_size. So truncate them
1546                  */
1547                 ext4_orphan_add(handle, inode);
1548
1549 errout:
1550         ret2 = ext4_journal_stop(handle);
1551         if (!ret)
1552                 ret = ret2;
1553         if (pos + len > inode->i_size) {
1554                 ext4_truncate_failed_write(inode);
1555                 /*
1556                  * If truncate failed early the inode might still be
1557                  * on the orphan list; we need to make sure the inode
1558                  * is removed from the orphan list in that case.
1559                  */
1560                 if (inode->i_nlink)
1561                         ext4_orphan_del(NULL, inode);
1562         }
1563
1564         return ret ? ret : copied;
1565 }
1566
1567 /*
1568  * Reserve space for a single cluster
1569  */
1570 static int ext4_da_reserve_space(struct inode *inode)
1571 {
1572         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1573         struct ext4_inode_info *ei = EXT4_I(inode);
1574         int ret;
1575
1576         /*
1577          * We will charge metadata quota at writeout time; this saves
1578          * us from metadata over-estimation, though we may go over by
1579          * a small amount in the end.  Here we just reserve for data.
1580          */
1581         ret = dquot_reserve_block(inode, EXT4_C2B(sbi, 1));
1582         if (ret)
1583                 return ret;
1584
1585         spin_lock(&ei->i_block_reservation_lock);
1586         if (ext4_claim_free_clusters(sbi, 1, 0)) {
1587                 spin_unlock(&ei->i_block_reservation_lock);
1588                 dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
1589                 return -ENOSPC;
1590         }
1591         ei->i_reserved_data_blocks++;
1592         trace_ext4_da_reserve_space(inode);
1593         spin_unlock(&ei->i_block_reservation_lock);
1594
1595         return 0;       /* success */
1596 }
1597
1598 void ext4_da_release_space(struct inode *inode, int to_free)
1599 {
1600         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1601         struct ext4_inode_info *ei = EXT4_I(inode);
1602
1603         if (!to_free)
1604                 return;         /* Nothing to release, exit */
1605
1606         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1607
1608         trace_ext4_da_release_space(inode, to_free);
1609         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1610                 /*
1611                  * if there aren't enough reserved blocks, then the
1612                  * counter is messed up somewhere.  Since this
1613                  * function is called from invalidate page, it's
1614                  * harmless to return without any action.
1615                  */
1616                 ext4_warning(inode->i_sb, "ext4_da_release_space: "
1617                          "ino %lu, to_free %d with only %d reserved "
1618                          "data blocks", inode->i_ino, to_free,
1619                          ei->i_reserved_data_blocks);
1620                 WARN_ON(1);
1621                 to_free = ei->i_reserved_data_blocks;
1622         }
1623         ei->i_reserved_data_blocks -= to_free;
1624
1625         /* update fs dirty data blocks counter */
1626         percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1627
1628         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1629
1630         dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1631 }
1632
1633 static void ext4_da_page_release_reservation(struct page *page,
1634                                              unsigned int offset,
1635                                              unsigned int length)
1636 {
1637         int contiguous_blks = 0;
1638         struct buffer_head *head, *bh;
1639         unsigned int curr_off = 0;
1640         struct inode *inode = page->mapping->host;
1641         unsigned int stop = offset + length;
1642         ext4_fsblk_t lblk;
1643
1644         BUG_ON(stop > PAGE_SIZE || stop < length);
1645
1646         head = page_buffers(page);
1647         bh = head;
1648         do {
1649                 unsigned int next_off = curr_off + bh->b_size;
1650
1651                 if (next_off > stop)
1652                         break;
1653
1654                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1655                         contiguous_blks++;
1656                         clear_buffer_delay(bh);
1657                 } else if (contiguous_blks) {
1658                         lblk = page->index <<
1659                                (PAGE_SHIFT - inode->i_blkbits);
1660                         lblk += (curr_off >> inode->i_blkbits) -
1661                                 contiguous_blks;
1662                         ext4_es_remove_blks(inode, lblk, contiguous_blks);
1663                         contiguous_blks = 0;
1664                 }
1665                 curr_off = next_off;
1666         } while ((bh = bh->b_this_page) != head);
1667
1668         if (contiguous_blks) {
1669                 lblk = page->index << (PAGE_SHIFT - inode->i_blkbits);
1670                 lblk += (curr_off >> inode->i_blkbits) - contiguous_blks;
1671                 ext4_es_remove_blks(inode, lblk, contiguous_blks);
1672         }
1673
1674 }
1675
1676 /*
1677  * Delayed allocation stuff
1678  */
1679
1680 struct mpage_da_data {
1681         struct inode *inode;
1682         struct writeback_control *wbc;
1683
1684         pgoff_t first_page;     /* The first page to write */
1685         pgoff_t next_page;      /* Current page to examine */
1686         pgoff_t last_page;      /* Last page to examine */
1687         /*
1688          * Extent to map - this can be after first_page because that can be
1689          * fully mapped. We somewhat abuse m_flags to store whether the extent
1690          * is delalloc or unwritten.
1691          */
1692         struct ext4_map_blocks map;
1693         struct ext4_io_submit io_submit;        /* IO submission data */
1694         unsigned int do_map:1;
1695 };
1696
1697 static void mpage_release_unused_pages(struct mpage_da_data *mpd,
1698                                        bool invalidate)
1699 {
1700         int nr_pages, i;
1701         pgoff_t index, end;
1702         struct pagevec pvec;
1703         struct inode *inode = mpd->inode;
1704         struct address_space *mapping = inode->i_mapping;
1705
1706         /* This is necessary when next_page == 0. */
1707         if (mpd->first_page >= mpd->next_page)
1708                 return;
1709
1710         index = mpd->first_page;
1711         end   = mpd->next_page - 1;
1712         if (invalidate) {
1713                 ext4_lblk_t start, last;
1714                 start = index << (PAGE_SHIFT - inode->i_blkbits);
1715                 last = end << (PAGE_SHIFT - inode->i_blkbits);
1716                 ext4_es_remove_extent(inode, start, last - start + 1);
1717         }
1718
1719         pagevec_init(&pvec);
1720         while (index <= end) {
1721                 nr_pages = pagevec_lookup_range(&pvec, mapping, &index, end);
1722                 if (nr_pages == 0)
1723                         break;
1724                 for (i = 0; i < nr_pages; i++) {
1725                         struct page *page = pvec.pages[i];
1726
1727                         BUG_ON(!PageLocked(page));
1728                         BUG_ON(PageWriteback(page));
1729                         if (invalidate) {
1730                                 if (page_mapped(page))
1731                                         clear_page_dirty_for_io(page);
1732                                 block_invalidatepage(page, 0, PAGE_SIZE);
1733                                 ClearPageUptodate(page);
1734                         }
1735                         unlock_page(page);
1736                 }
1737                 pagevec_release(&pvec);
1738         }
1739 }
1740
1741 static void ext4_print_free_blocks(struct inode *inode)
1742 {
1743         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1744         struct super_block *sb = inode->i_sb;
1745         struct ext4_inode_info *ei = EXT4_I(inode);
1746
1747         ext4_msg(sb, KERN_CRIT, "Total free blocks count %lld",
1748                EXT4_C2B(EXT4_SB(inode->i_sb),
1749                         ext4_count_free_clusters(sb)));
1750         ext4_msg(sb, KERN_CRIT, "Free/Dirty block details");
1751         ext4_msg(sb, KERN_CRIT, "free_blocks=%lld",
1752                (long long) EXT4_C2B(EXT4_SB(sb),
1753                 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1754         ext4_msg(sb, KERN_CRIT, "dirty_blocks=%lld",
1755                (long long) EXT4_C2B(EXT4_SB(sb),
1756                 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1757         ext4_msg(sb, KERN_CRIT, "Block reservation details");
1758         ext4_msg(sb, KERN_CRIT, "i_reserved_data_blocks=%u",
1759                  ei->i_reserved_data_blocks);
1760         return;
1761 }
1762
1763 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1764 {
1765         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1766 }
1767
1768 /*
1769  * ext4_insert_delayed_block - adds a delayed block to the extents status
1770  *                             tree, incrementing the reserved cluster/block
1771  *                             count or making a pending reservation
1772  *                             where needed
1773  *
1774  * @inode - file containing the newly added block
1775  * @lblk - logical block to be added
1776  *
1777  * Returns 0 on success, negative error code on failure.
1778  */
1779 static int ext4_insert_delayed_block(struct inode *inode, ext4_lblk_t lblk)
1780 {
1781         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1782         int ret;
1783         bool allocated = false;
1784
1785         /*
1786          * If the cluster containing lblk is shared with a delayed,
1787          * written, or unwritten extent in a bigalloc file system, it's
1788          * already been accounted for and does not need to be reserved.
1789          * A pending reservation must be made for the cluster if it's
1790          * shared with a written or unwritten extent and doesn't already
1791          * have one.  Written and unwritten extents can be purged from the
1792          * extents status tree if the system is under memory pressure, so
1793          * it's necessary to examine the extent tree if a search of the
1794          * extents status tree doesn't get a match.
1795          */
1796         if (sbi->s_cluster_ratio == 1) {
1797                 ret = ext4_da_reserve_space(inode);
1798                 if (ret != 0)   /* ENOSPC */
1799                         goto errout;
1800         } else {   /* bigalloc */
1801                 if (!ext4_es_scan_clu(inode, &ext4_es_is_delonly, lblk)) {
1802                         if (!ext4_es_scan_clu(inode,
1803                                               &ext4_es_is_mapped, lblk)) {
1804                                 ret = ext4_clu_mapped(inode,
1805                                                       EXT4_B2C(sbi, lblk));
1806                                 if (ret < 0)
1807                                         goto errout;
1808                                 if (ret == 0) {
1809                                         ret = ext4_da_reserve_space(inode);
1810                                         if (ret != 0)   /* ENOSPC */
1811                                                 goto errout;
1812                                 } else {
1813                                         allocated = true;
1814                                 }
1815                         } else {
1816                                 allocated = true;
1817                         }
1818                 }
1819         }
1820
1821         ret = ext4_es_insert_delayed_block(inode, lblk, allocated);
1822
1823 errout:
1824         return ret;
1825 }
1826
1827 /*
1828  * This function is grabs code from the very beginning of
1829  * ext4_map_blocks, but assumes that the caller is from delayed write
1830  * time. This function looks up the requested blocks and sets the
1831  * buffer delay bit under the protection of i_data_sem.
1832  */
1833 static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1834                               struct ext4_map_blocks *map,
1835                               struct buffer_head *bh)
1836 {
1837         struct extent_status es;
1838         int retval;
1839         sector_t invalid_block = ~((sector_t) 0xffff);
1840 #ifdef ES_AGGRESSIVE_TEST
1841         struct ext4_map_blocks orig_map;
1842
1843         memcpy(&orig_map, map, sizeof(*map));
1844 #endif
1845
1846         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1847                 invalid_block = ~0;
1848
1849         map->m_flags = 0;
1850         ext_debug("ext4_da_map_blocks(): inode %lu, max_blocks %u,"
1851                   "logical block %lu\n", inode->i_ino, map->m_len,
1852                   (unsigned long) map->m_lblk);
1853
1854         /* Lookup extent status tree firstly */
1855         if (ext4_es_lookup_extent(inode, iblock, &es)) {
1856                 if (ext4_es_is_hole(&es)) {
1857                         retval = 0;
1858                         down_read(&EXT4_I(inode)->i_data_sem);
1859                         goto add_delayed;
1860                 }
1861
1862                 /*
1863                  * Delayed extent could be allocated by fallocate.
1864                  * So we need to check it.
1865                  */
1866                 if (ext4_es_is_delayed(&es) && !ext4_es_is_unwritten(&es)) {
1867                         map_bh(bh, inode->i_sb, invalid_block);
1868                         set_buffer_new(bh);
1869                         set_buffer_delay(bh);
1870                         return 0;
1871                 }
1872
1873                 map->m_pblk = ext4_es_pblock(&es) + iblock - es.es_lblk;
1874                 retval = es.es_len - (iblock - es.es_lblk);
1875                 if (retval > map->m_len)
1876                         retval = map->m_len;
1877                 map->m_len = retval;
1878                 if (ext4_es_is_written(&es))
1879                         map->m_flags |= EXT4_MAP_MAPPED;
1880                 else if (ext4_es_is_unwritten(&es))
1881                         map->m_flags |= EXT4_MAP_UNWRITTEN;
1882                 else
1883                         BUG_ON(1);
1884
1885 #ifdef ES_AGGRESSIVE_TEST
1886                 ext4_map_blocks_es_recheck(NULL, inode, map, &orig_map, 0);
1887 #endif
1888                 return retval;
1889         }
1890
1891         /*
1892          * Try to see if we can get the block without requesting a new
1893          * file system block.
1894          */
1895         down_read(&EXT4_I(inode)->i_data_sem);
1896         if (ext4_has_inline_data(inode))
1897                 retval = 0;
1898         else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1899                 retval = ext4_ext_map_blocks(NULL, inode, map, 0);
1900         else
1901                 retval = ext4_ind_map_blocks(NULL, inode, map, 0);
1902
1903 add_delayed:
1904         if (retval == 0) {
1905                 int ret;
1906
1907                 /*
1908                  * XXX: __block_prepare_write() unmaps passed block,
1909                  * is it OK?
1910                  */
1911
1912                 ret = ext4_insert_delayed_block(inode, map->m_lblk);
1913                 if (ret != 0) {
1914                         retval = ret;
1915                         goto out_unlock;
1916                 }
1917
1918                 map_bh(bh, inode->i_sb, invalid_block);
1919                 set_buffer_new(bh);
1920                 set_buffer_delay(bh);
1921         } else if (retval > 0) {
1922                 int ret;
1923                 unsigned int status;
1924
1925                 if (unlikely(retval != map->m_len)) {
1926                         ext4_warning(inode->i_sb,
1927                                      "ES len assertion failed for inode "
1928                                      "%lu: retval %d != map->m_len %d",
1929                                      inode->i_ino, retval, map->m_len);
1930                         WARN_ON(1);
1931                 }
1932
1933                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
1934                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
1935                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1936                                             map->m_pblk, status);
1937                 if (ret != 0)
1938                         retval = ret;
1939         }
1940
1941 out_unlock:
1942         up_read((&EXT4_I(inode)->i_data_sem));
1943
1944         return retval;
1945 }
1946
1947 /*
1948  * This is a special get_block_t callback which is used by
1949  * ext4_da_write_begin().  It will either return mapped block or
1950  * reserve space for a single block.
1951  *
1952  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1953  * We also have b_blocknr = -1 and b_bdev initialized properly
1954  *
1955  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1956  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1957  * initialized properly.
1958  */
1959 int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1960                            struct buffer_head *bh, int create)
1961 {
1962         struct ext4_map_blocks map;
1963         int ret = 0;
1964
1965         BUG_ON(create == 0);
1966         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1967
1968         map.m_lblk = iblock;
1969         map.m_len = 1;
1970
1971         /*
1972          * first, we need to know whether the block is allocated already
1973          * preallocated blocks are unmapped but should treated
1974          * the same as allocated blocks.
1975          */
1976         ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1977         if (ret <= 0)
1978                 return ret;
1979
1980         map_bh(bh, inode->i_sb, map.m_pblk);
1981         ext4_update_bh_state(bh, map.m_flags);
1982
1983         if (buffer_unwritten(bh)) {
1984                 /* A delayed write to unwritten bh should be marked
1985                  * new and mapped.  Mapped ensures that we don't do
1986                  * get_block multiple times when we write to the same
1987                  * offset and new ensures that we do proper zero out
1988                  * for partial write.
1989                  */
1990                 set_buffer_new(bh);
1991                 set_buffer_mapped(bh);
1992         }
1993         return 0;
1994 }
1995
1996 static int bget_one(handle_t *handle, struct buffer_head *bh)
1997 {
1998         get_bh(bh);
1999         return 0;
2000 }
2001
2002 static int bput_one(handle_t *handle, struct buffer_head *bh)
2003 {
2004         put_bh(bh);
2005         return 0;
2006 }
2007
2008 static int __ext4_journalled_writepage(struct page *page,
2009                                        unsigned int len)
2010 {
2011         struct address_space *mapping = page->mapping;
2012         struct inode *inode = mapping->host;
2013         struct buffer_head *page_bufs = NULL;
2014         handle_t *handle = NULL;
2015         int ret = 0, err = 0;
2016         int inline_data = ext4_has_inline_data(inode);
2017         struct buffer_head *inode_bh = NULL;
2018
2019         ClearPageChecked(page);
2020
2021         if (inline_data) {
2022                 BUG_ON(page->index != 0);
2023                 BUG_ON(len > ext4_get_max_inline_size(inode));
2024                 inode_bh = ext4_journalled_write_inline_data(inode, len, page);
2025                 if (inode_bh == NULL)
2026                         goto out;
2027         } else {
2028                 page_bufs = page_buffers(page);
2029                 if (!page_bufs) {
2030                         BUG();
2031                         goto out;
2032                 }
2033                 ext4_walk_page_buffers(handle, page_bufs, 0, len,
2034                                        NULL, bget_one);
2035         }
2036         /*
2037          * We need to release the page lock before we start the
2038          * journal, so grab a reference so the page won't disappear
2039          * out from under us.
2040          */
2041         get_page(page);
2042         unlock_page(page);
2043
2044         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
2045                                     ext4_writepage_trans_blocks(inode));
2046         if (IS_ERR(handle)) {
2047                 ret = PTR_ERR(handle);
2048                 put_page(page);
2049                 goto out_no_pagelock;
2050         }
2051         BUG_ON(!ext4_handle_valid(handle));
2052
2053         lock_page(page);
2054         put_page(page);
2055         if (page->mapping != mapping) {
2056                 /* The page got truncated from under us */
2057                 ext4_journal_stop(handle);
2058                 ret = 0;
2059                 goto out;
2060         }
2061
2062         if (inline_data) {
2063                 ret = ext4_mark_inode_dirty(handle, inode);
2064         } else {
2065                 ret = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
2066                                              do_journal_get_write_access);
2067
2068                 err = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
2069                                              write_end_fn);
2070         }
2071         if (ret == 0)
2072                 ret = err;
2073         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
2074         err = ext4_journal_stop(handle);
2075         if (!ret)
2076                 ret = err;
2077
2078         if (!ext4_has_inline_data(inode))
2079                 ext4_walk_page_buffers(NULL, page_bufs, 0, len,
2080                                        NULL, bput_one);
2081         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2082 out:
2083         unlock_page(page);
2084 out_no_pagelock:
2085         brelse(inode_bh);
2086         return ret;
2087 }
2088
2089 /*
2090  * Note that we don't need to start a transaction unless we're journaling data
2091  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2092  * need to file the inode to the transaction's list in ordered mode because if
2093  * we are writing back data added by write(), the inode is already there and if
2094  * we are writing back data modified via mmap(), no one guarantees in which
2095  * transaction the data will hit the disk. In case we are journaling data, we
2096  * cannot start transaction directly because transaction start ranks above page
2097  * lock so we have to do some magic.
2098  *
2099  * This function can get called via...
2100  *   - ext4_writepages after taking page lock (have journal handle)
2101  *   - journal_submit_inode_data_buffers (no journal handle)
2102  *   - shrink_page_list via the kswapd/direct reclaim (no journal handle)
2103  *   - grab_page_cache when doing write_begin (have journal handle)
2104  *
2105  * We don't do any block allocation in this function. If we have page with
2106  * multiple blocks we need to write those buffer_heads that are mapped. This
2107  * is important for mmaped based write. So if we do with blocksize 1K
2108  * truncate(f, 1024);
2109  * a = mmap(f, 0, 4096);
2110  * a[0] = 'a';
2111  * truncate(f, 4096);
2112  * we have in the page first buffer_head mapped via page_mkwrite call back
2113  * but other buffer_heads would be unmapped but dirty (dirty done via the
2114  * do_wp_page). So writepage should write the first block. If we modify
2115  * the mmap area beyond 1024 we will again get a page_fault and the
2116  * page_mkwrite callback will do the block allocation and mark the
2117  * buffer_heads mapped.
2118  *
2119  * We redirty the page if we have any buffer_heads that is either delay or
2120  * unwritten in the page.
2121  *
2122  * We can get recursively called as show below.
2123  *
2124  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2125  *              ext4_writepage()
2126  *
2127  * But since we don't do any block allocation we should not deadlock.
2128  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2129  */
2130 static int ext4_writepage(struct page *page,
2131                           struct writeback_control *wbc)
2132 {
2133         int ret = 0;
2134         loff_t size;
2135         unsigned int len;
2136         struct buffer_head *page_bufs = NULL;
2137         struct inode *inode = page->mapping->host;
2138         struct ext4_io_submit io_submit;
2139         bool keep_towrite = false;
2140
2141         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb)))) {
2142                 ext4_invalidatepage(page, 0, PAGE_SIZE);
2143                 unlock_page(page);
2144                 return -EIO;
2145         }
2146
2147         trace_ext4_writepage(page);
2148         size = i_size_read(inode);
2149         if (page->index == size >> PAGE_SHIFT)
2150                 len = size & ~PAGE_MASK;
2151         else
2152                 len = PAGE_SIZE;
2153
2154         page_bufs = page_buffers(page);
2155         /*
2156          * We cannot do block allocation or other extent handling in this
2157          * function. If there are buffers needing that, we have to redirty
2158          * the page. But we may reach here when we do a journal commit via
2159          * journal_submit_inode_data_buffers() and in that case we must write
2160          * allocated buffers to achieve data=ordered mode guarantees.
2161          *
2162          * Also, if there is only one buffer per page (the fs block
2163          * size == the page size), if one buffer needs block
2164          * allocation or needs to modify the extent tree to clear the
2165          * unwritten flag, we know that the page can't be written at
2166          * all, so we might as well refuse the write immediately.
2167          * Unfortunately if the block size != page size, we can't as
2168          * easily detect this case using ext4_walk_page_buffers(), but
2169          * for the extremely common case, this is an optimization that
2170          * skips a useless round trip through ext4_bio_write_page().
2171          */
2172         if (ext4_walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2173                                    ext4_bh_delay_or_unwritten)) {
2174                 redirty_page_for_writepage(wbc, page);
2175                 if ((current->flags & PF_MEMALLOC) ||
2176                     (inode->i_sb->s_blocksize == PAGE_SIZE)) {
2177                         /*
2178                          * For memory cleaning there's no point in writing only
2179                          * some buffers. So just bail out. Warn if we came here
2180                          * from direct reclaim.
2181                          */
2182                         WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD))
2183                                                         == PF_MEMALLOC);
2184                         unlock_page(page);
2185                         return 0;
2186                 }
2187                 keep_towrite = true;
2188         }
2189
2190         if (PageChecked(page) && ext4_should_journal_data(inode))
2191                 /*
2192                  * It's mmapped pagecache.  Add buffers and journal it.  There
2193                  * doesn't seem much point in redirtying the page here.
2194                  */
2195                 return __ext4_journalled_writepage(page, len);
2196
2197         ext4_io_submit_init(&io_submit, wbc);
2198         io_submit.io_end = ext4_init_io_end(inode, GFP_NOFS);
2199         if (!io_submit.io_end) {
2200                 redirty_page_for_writepage(wbc, page);
2201                 unlock_page(page);
2202                 return -ENOMEM;
2203         }
2204         ret = ext4_bio_write_page(&io_submit, page, len, wbc, keep_towrite);
2205         ext4_io_submit(&io_submit);
2206         /* Drop io_end reference we got from init */
2207         ext4_put_io_end_defer(io_submit.io_end);
2208         return ret;
2209 }
2210
2211 static int mpage_submit_page(struct mpage_da_data *mpd, struct page *page)
2212 {
2213         int len;
2214         loff_t size;
2215         int err;
2216
2217         BUG_ON(page->index != mpd->first_page);
2218         clear_page_dirty_for_io(page);
2219         /*
2220          * We have to be very careful here!  Nothing protects writeback path
2221          * against i_size changes and the page can be writeably mapped into
2222          * page tables. So an application can be growing i_size and writing
2223          * data through mmap while writeback runs. clear_page_dirty_for_io()
2224          * write-protects our page in page tables and the page cannot get
2225          * written to again until we release page lock. So only after
2226          * clear_page_dirty_for_io() we are safe to sample i_size for
2227          * ext4_bio_write_page() to zero-out tail of the written page. We rely
2228          * on the barrier provided by TestClearPageDirty in
2229          * clear_page_dirty_for_io() to make sure i_size is really sampled only
2230          * after page tables are updated.
2231          */
2232         size = i_size_read(mpd->inode);
2233         if (page->index == size >> PAGE_SHIFT)
2234                 len = size & ~PAGE_MASK;
2235         else
2236                 len = PAGE_SIZE;
2237         err = ext4_bio_write_page(&mpd->io_submit, page, len, mpd->wbc, false);
2238         if (!err)
2239                 mpd->wbc->nr_to_write--;
2240         mpd->first_page++;
2241
2242         return err;
2243 }
2244
2245 #define BH_FLAGS ((1 << BH_Unwritten) | (1 << BH_Delay))
2246
2247 /*
2248  * mballoc gives us at most this number of blocks...
2249  * XXX: That seems to be only a limitation of ext4_mb_normalize_request().
2250  * The rest of mballoc seems to handle chunks up to full group size.
2251  */
2252 #define MAX_WRITEPAGES_EXTENT_LEN 2048
2253
2254 /*
2255  * mpage_add_bh_to_extent - try to add bh to extent of blocks to map
2256  *
2257  * @mpd - extent of blocks
2258  * @lblk - logical number of the block in the file
2259  * @bh - buffer head we want to add to the extent
2260  *
2261  * The function is used to collect contig. blocks in the same state. If the
2262  * buffer doesn't require mapping for writeback and we haven't started the
2263  * extent of buffers to map yet, the function returns 'true' immediately - the
2264  * caller can write the buffer right away. Otherwise the function returns true
2265  * if the block has been added to the extent, false if the block couldn't be
2266  * added.
2267  */
2268 static bool mpage_add_bh_to_extent(struct mpage_da_data *mpd, ext4_lblk_t lblk,
2269                                    struct buffer_head *bh)
2270 {
2271         struct ext4_map_blocks *map = &mpd->map;
2272
2273         /* Buffer that doesn't need mapping for writeback? */
2274         if (!buffer_dirty(bh) || !buffer_mapped(bh) ||
2275             (!buffer_delay(bh) && !buffer_unwritten(bh))) {
2276                 /* So far no extent to map => we write the buffer right away */
2277                 if (map->m_len == 0)
2278                         return true;
2279                 return false;
2280         }
2281
2282         /* First block in the extent? */
2283         if (map->m_len == 0) {
2284                 /* We cannot map unless handle is started... */
2285                 if (!mpd->do_map)
2286                         return false;
2287                 map->m_lblk = lblk;
2288                 map->m_len = 1;
2289                 map->m_flags = bh->b_state & BH_FLAGS;
2290                 return true;
2291         }
2292
2293         /* Don't go larger than mballoc is willing to allocate */
2294         if (map->m_len >= MAX_WRITEPAGES_EXTENT_LEN)
2295                 return false;
2296
2297         /* Can we merge the block to our big extent? */
2298         if (lblk == map->m_lblk + map->m_len &&
2299             (bh->b_state & BH_FLAGS) == map->m_flags) {
2300                 map->m_len++;
2301                 return true;
2302         }
2303         return false;
2304 }
2305
2306 /*
2307  * mpage_process_page_bufs - submit page buffers for IO or add them to extent
2308  *
2309  * @mpd - extent of blocks for mapping
2310  * @head - the first buffer in the page
2311  * @bh - buffer we should start processing from
2312  * @lblk - logical number of the block in the file corresponding to @bh
2313  *
2314  * Walk through page buffers from @bh upto @head (exclusive) and either submit
2315  * the page for IO if all buffers in this page were mapped and there's no
2316  * accumulated extent of buffers to map or add buffers in the page to the
2317  * extent of buffers to map. The function returns 1 if the caller can continue
2318  * by processing the next page, 0 if it should stop adding buffers to the
2319  * extent to map because we cannot extend it anymore. It can also return value
2320  * < 0 in case of error during IO submission.
2321  */
2322 static int mpage_process_page_bufs(struct mpage_da_data *mpd,
2323                                    struct buffer_head *head,
2324                                    struct buffer_head *bh,
2325                                    ext4_lblk_t lblk)
2326 {
2327         struct inode *inode = mpd->inode;
2328         int err;
2329         ext4_lblk_t blocks = (i_size_read(inode) + i_blocksize(inode) - 1)
2330                                                         >> inode->i_blkbits;
2331
2332         do {
2333                 BUG_ON(buffer_locked(bh));
2334
2335                 if (lblk >= blocks || !mpage_add_bh_to_extent(mpd, lblk, bh)) {
2336                         /* Found extent to map? */
2337                         if (mpd->map.m_len)
2338                                 return 0;
2339                         /* Buffer needs mapping and handle is not started? */
2340                         if (!mpd->do_map)
2341                                 return 0;
2342                         /* Everything mapped so far and we hit EOF */
2343                         break;
2344                 }
2345         } while (lblk++, (bh = bh->b_this_page) != head);
2346         /* So far everything mapped? Submit the page for IO. */
2347         if (mpd->map.m_len == 0) {
2348                 err = mpage_submit_page(mpd, head->b_page);
2349                 if (err < 0)
2350                         return err;
2351         }
2352         return lblk < blocks;
2353 }
2354
2355 /*
2356  * mpage_map_buffers - update buffers corresponding to changed extent and
2357  *                     submit fully mapped pages for IO
2358  *
2359  * @mpd - description of extent to map, on return next extent to map
2360  *
2361  * Scan buffers corresponding to changed extent (we expect corresponding pages
2362  * to be already locked) and update buffer state according to new extent state.
2363  * We map delalloc buffers to their physical location, clear unwritten bits,
2364  * and mark buffers as uninit when we perform writes to unwritten extents
2365  * and do extent conversion after IO is finished. If the last page is not fully
2366  * mapped, we update @map to the next extent in the last page that needs
2367  * mapping. Otherwise we submit the page for IO.
2368  */
2369 static int mpage_map_and_submit_buffers(struct mpage_da_data *mpd)
2370 {
2371         struct pagevec pvec;
2372         int nr_pages, i;
2373         struct inode *inode = mpd->inode;
2374         struct buffer_head *head, *bh;
2375         int bpp_bits = PAGE_SHIFT - inode->i_blkbits;
2376         pgoff_t start, end;
2377         ext4_lblk_t lblk;
2378         sector_t pblock;
2379         int err;
2380
2381         start = mpd->map.m_lblk >> bpp_bits;
2382         end = (mpd->map.m_lblk + mpd->map.m_len - 1) >> bpp_bits;
2383         lblk = start << bpp_bits;
2384         pblock = mpd->map.m_pblk;
2385
2386         pagevec_init(&pvec);
2387         while (start <= end) {
2388                 nr_pages = pagevec_lookup_range(&pvec, inode->i_mapping,
2389                                                 &start, end);
2390                 if (nr_pages == 0)
2391                         break;
2392                 for (i = 0; i < nr_pages; i++) {
2393                         struct page *page = pvec.pages[i];
2394
2395                         bh = head = page_buffers(page);
2396                         do {
2397                                 if (lblk < mpd->map.m_lblk)
2398                                         continue;
2399                                 if (lblk >= mpd->map.m_lblk + mpd->map.m_len) {
2400                                         /*
2401                                          * Buffer after end of mapped extent.
2402                                          * Find next buffer in the page to map.
2403                                          */
2404                                         mpd->map.m_len = 0;
2405                                         mpd->map.m_flags = 0;
2406                                         /*
2407                                          * FIXME: If dioread_nolock supports
2408                                          * blocksize < pagesize, we need to make
2409                                          * sure we add size mapped so far to
2410                                          * io_end->size as the following call
2411                                          * can submit the page for IO.
2412                                          */
2413                                         err = mpage_process_page_bufs(mpd, head,
2414                                                                       bh, lblk);
2415                                         pagevec_release(&pvec);
2416                                         if (err > 0)
2417                                                 err = 0;
2418                                         return err;
2419                                 }
2420                                 if (buffer_delay(bh)) {
2421                                         clear_buffer_delay(bh);
2422                                         bh->b_blocknr = pblock++;
2423                                 }
2424                                 clear_buffer_unwritten(bh);
2425                         } while (lblk++, (bh = bh->b_this_page) != head);
2426
2427                         /*
2428                          * FIXME: This is going to break if dioread_nolock
2429                          * supports blocksize < pagesize as we will try to
2430                          * convert potentially unmapped parts of inode.
2431                          */
2432                         mpd->io_submit.io_end->size += PAGE_SIZE;
2433                         /* Page fully mapped - let IO run! */
2434                         err = mpage_submit_page(mpd, page);
2435                         if (err < 0) {
2436                                 pagevec_release(&pvec);
2437                                 return err;
2438                         }
2439                 }
2440                 pagevec_release(&pvec);
2441         }
2442         /* Extent fully mapped and matches with page boundary. We are done. */
2443         mpd->map.m_len = 0;
2444         mpd->map.m_flags = 0;
2445         return 0;
2446 }
2447
2448 static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
2449 {
2450         struct inode *inode = mpd->inode;
2451         struct ext4_map_blocks *map = &mpd->map;
2452         int get_blocks_flags;
2453         int err, dioread_nolock;
2454
2455         trace_ext4_da_write_pages_extent(inode, map);
2456         /*
2457          * Call ext4_map_blocks() to allocate any delayed allocation blocks, or
2458          * to convert an unwritten extent to be initialized (in the case
2459          * where we have written into one or more preallocated blocks).  It is
2460          * possible that we're going to need more metadata blocks than
2461          * previously reserved. However we must not fail because we're in
2462          * writeback and there is nothing we can do about it so it might result
2463          * in data loss.  So use reserved blocks to allocate metadata if
2464          * possible.
2465          *
2466          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE if
2467          * the blocks in question are delalloc blocks.  This indicates
2468          * that the blocks and quotas has already been checked when
2469          * the data was copied into the page cache.
2470          */
2471         get_blocks_flags = EXT4_GET_BLOCKS_CREATE |
2472                            EXT4_GET_BLOCKS_METADATA_NOFAIL |
2473                            EXT4_GET_BLOCKS_IO_SUBMIT;
2474         dioread_nolock = ext4_should_dioread_nolock(inode);
2475         if (dioread_nolock)
2476                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2477         if (map->m_flags & (1 << BH_Delay))
2478                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2479
2480         err = ext4_map_blocks(handle, inode, map, get_blocks_flags);
2481         if (err < 0)
2482                 return err;
2483         if (dioread_nolock && (map->m_flags & EXT4_MAP_UNWRITTEN)) {
2484                 if (!mpd->io_submit.io_end->handle &&
2485                     ext4_handle_valid(handle)) {
2486                         mpd->io_submit.io_end->handle = handle->h_rsv_handle;
2487                         handle->h_rsv_handle = NULL;
2488                 }
2489                 ext4_set_io_unwritten_flag(inode, mpd->io_submit.io_end);
2490         }
2491
2492         BUG_ON(map->m_len == 0);
2493         if (map->m_flags & EXT4_MAP_NEW) {
2494                 clean_bdev_aliases(inode->i_sb->s_bdev, map->m_pblk,
2495                                    map->m_len);
2496         }
2497         return 0;
2498 }
2499
2500 /*
2501  * mpage_map_and_submit_extent - map extent starting at mpd->lblk of length
2502  *                               mpd->len and submit pages underlying it for IO
2503  *
2504  * @handle - handle for journal operations
2505  * @mpd - extent to map
2506  * @give_up_on_write - we set this to true iff there is a fatal error and there
2507  *                     is no hope of writing the data. The caller should discard
2508  *                     dirty pages to avoid infinite loops.
2509  *
2510  * The function maps extent starting at mpd->lblk of length mpd->len. If it is
2511  * delayed, blocks are allocated, if it is unwritten, we may need to convert
2512  * them to initialized or split the described range from larger unwritten
2513  * extent. Note that we need not map all the described range since allocation
2514  * can return less blocks or the range is covered by more unwritten extents. We
2515  * cannot map more because we are limited by reserved transaction credits. On
2516  * the other hand we always make sure that the last touched page is fully
2517  * mapped so that it can be written out (and thus forward progress is
2518  * guaranteed). After mapping we submit all mapped pages for IO.
2519  */
2520 static int mpage_map_and_submit_extent(handle_t *handle,
2521                                        struct mpage_da_data *mpd,
2522                                        bool *give_up_on_write)
2523 {
2524         struct inode *inode = mpd->inode;
2525         struct ext4_map_blocks *map = &mpd->map;
2526         int err;
2527         loff_t disksize;
2528         int progress = 0;
2529
2530         mpd->io_submit.io_end->offset =
2531                                 ((loff_t)map->m_lblk) << inode->i_blkbits;
2532         do {
2533                 err = mpage_map_one_extent(handle, mpd);
2534                 if (err < 0) {
2535                         struct super_block *sb = inode->i_sb;
2536
2537                         if (ext4_forced_shutdown(EXT4_SB(sb)) ||
2538                             EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
2539                                 goto invalidate_dirty_pages;
2540                         /*
2541                          * Let the uper layers retry transient errors.
2542                          * In the case of ENOSPC, if ext4_count_free_blocks()
2543                          * is non-zero, a commit should free up blocks.
2544                          */
2545                         if ((err == -ENOMEM) ||
2546                             (err == -ENOSPC && ext4_count_free_clusters(sb))) {
2547                                 if (progress)
2548                                         goto update_disksize;
2549                                 return err;
2550                         }
2551                         ext4_msg(sb, KERN_CRIT,
2552                                  "Delayed block allocation failed for "
2553                                  "inode %lu at logical offset %llu with"
2554                                  " max blocks %u with error %d",
2555                                  inode->i_ino,
2556                                  (unsigned long long)map->m_lblk,
2557                                  (unsigned)map->m_len, -err);
2558                         ext4_msg(sb, KERN_CRIT,
2559                                  "This should not happen!! Data will "
2560                                  "be lost\n");
2561                         if (err == -ENOSPC)
2562                                 ext4_print_free_blocks(inode);
2563                 invalidate_dirty_pages:
2564                         *give_up_on_write = true;
2565                         return err;
2566                 }
2567                 progress = 1;
2568                 /*
2569                  * Update buffer state, submit mapped pages, and get us new
2570                  * extent to map
2571                  */
2572                 err = mpage_map_and_submit_buffers(mpd);
2573                 if (err < 0)
2574                         goto update_disksize;
2575         } while (map->m_len);
2576
2577 update_disksize:
2578         /*
2579          * Update on-disk size after IO is submitted.  Races with
2580          * truncate are avoided by checking i_size under i_data_sem.
2581          */
2582         disksize = ((loff_t)mpd->first_page) << PAGE_SHIFT;
2583         if (disksize > EXT4_I(inode)->i_disksize) {
2584                 int err2;
2585                 loff_t i_size;
2586
2587                 down_write(&EXT4_I(inode)->i_data_sem);
2588                 i_size = i_size_read(inode);
2589                 if (disksize > i_size)
2590                         disksize = i_size;
2591                 if (disksize > EXT4_I(inode)->i_disksize)
2592                         EXT4_I(inode)->i_disksize = disksize;
2593                 up_write(&EXT4_I(inode)->i_data_sem);
2594                 err2 = ext4_mark_inode_dirty(handle, inode);
2595                 if (err2)
2596                         ext4_error(inode->i_sb,
2597                                    "Failed to mark inode %lu dirty",
2598                                    inode->i_ino);
2599                 if (!err)
2600                         err = err2;
2601         }
2602         return err;
2603 }
2604
2605 /*
2606  * Calculate the total number of credits to reserve for one writepages
2607  * iteration. This is called from ext4_writepages(). We map an extent of
2608  * up to MAX_WRITEPAGES_EXTENT_LEN blocks and then we go on and finish mapping
2609  * the last partial page. So in total we can map MAX_WRITEPAGES_EXTENT_LEN +
2610  * bpp - 1 blocks in bpp different extents.
2611  */
2612 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2613 {
2614         int bpp = ext4_journal_blocks_per_page(inode);
2615
2616         return ext4_meta_trans_blocks(inode,
2617                                 MAX_WRITEPAGES_EXTENT_LEN + bpp - 1, bpp);
2618 }
2619
2620 /*
2621  * mpage_prepare_extent_to_map - find & lock contiguous range of dirty pages
2622  *                               and underlying extent to map
2623  *
2624  * @mpd - where to look for pages
2625  *
2626  * Walk dirty pages in the mapping. If they are fully mapped, submit them for
2627  * IO immediately. When we find a page which isn't mapped we start accumulating
2628  * extent of buffers underlying these pages that needs mapping (formed by
2629  * either delayed or unwritten buffers). We also lock the pages containing
2630  * these buffers. The extent found is returned in @mpd structure (starting at
2631  * mpd->lblk with length mpd->len blocks).
2632  *
2633  * Note that this function can attach bios to one io_end structure which are
2634  * neither logically nor physically contiguous. Although it may seem as an
2635  * unnecessary complication, it is actually inevitable in blocksize < pagesize
2636  * case as we need to track IO to all buffers underlying a page in one io_end.
2637  */
2638 static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
2639 {
2640         struct address_space *mapping = mpd->inode->i_mapping;
2641         struct pagevec pvec;
2642         unsigned int nr_pages;
2643         long left = mpd->wbc->nr_to_write;
2644         pgoff_t index = mpd->first_page;
2645         pgoff_t end = mpd->last_page;
2646         xa_mark_t tag;
2647         int i, err = 0;
2648         int blkbits = mpd->inode->i_blkbits;
2649         ext4_lblk_t lblk;
2650         struct buffer_head *head;
2651
2652         if (mpd->wbc->sync_mode == WB_SYNC_ALL || mpd->wbc->tagged_writepages)
2653                 tag = PAGECACHE_TAG_TOWRITE;
2654         else
2655                 tag = PAGECACHE_TAG_DIRTY;
2656
2657         pagevec_init(&pvec);
2658         mpd->map.m_len = 0;
2659         mpd->next_page = index;
2660         while (index <= end) {
2661                 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2662                                 tag);
2663                 if (nr_pages == 0)
2664                         goto out;
2665
2666                 for (i = 0; i < nr_pages; i++) {
2667                         struct page *page = pvec.pages[i];
2668
2669                         /*
2670                          * Accumulated enough dirty pages? This doesn't apply
2671                          * to WB_SYNC_ALL mode. For integrity sync we have to
2672                          * keep going because someone may be concurrently
2673                          * dirtying pages, and we might have synced a lot of
2674                          * newly appeared dirty pages, but have not synced all
2675                          * of the old dirty pages.
2676                          */
2677                         if (mpd->wbc->sync_mode == WB_SYNC_NONE && left <= 0)
2678                                 goto out;
2679
2680                         /* If we can't merge this page, we are done. */
2681                         if (mpd->map.m_len > 0 && mpd->next_page != page->index)
2682                                 goto out;
2683
2684                         lock_page(page);
2685                         /*
2686                          * If the page is no longer dirty, or its mapping no
2687                          * longer corresponds to inode we are writing (which
2688                          * means it has been truncated or invalidated), or the
2689                          * page is already under writeback and we are not doing
2690                          * a data integrity writeback, skip the page
2691                          */
2692                         if (!PageDirty(page) ||
2693                             (PageWriteback(page) &&
2694                              (mpd->wbc->sync_mode == WB_SYNC_NONE)) ||
2695                             unlikely(page->mapping != mapping)) {
2696                                 unlock_page(page);
2697                                 continue;
2698                         }
2699
2700                         wait_on_page_writeback(page);
2701                         BUG_ON(PageWriteback(page));
2702
2703                         if (mpd->map.m_len == 0)
2704                                 mpd->first_page = page->index;
2705                         mpd->next_page = page->index + 1;
2706                         /* Add all dirty buffers to mpd */
2707                         lblk = ((ext4_lblk_t)page->index) <<
2708                                 (PAGE_SHIFT - blkbits);
2709                         head = page_buffers(page);
2710                         err = mpage_process_page_bufs(mpd, head, head, lblk);
2711                         if (err <= 0)
2712                                 goto out;
2713                         err = 0;
2714                         left--;
2715                 }
2716                 pagevec_release(&pvec);
2717                 cond_resched();
2718         }
2719         return 0;
2720 out:
2721         pagevec_release(&pvec);
2722         return err;
2723 }
2724
2725 static int ext4_writepages(struct address_space *mapping,
2726                            struct writeback_control *wbc)
2727 {
2728         pgoff_t writeback_index = 0;
2729         long nr_to_write = wbc->nr_to_write;
2730         int range_whole = 0;
2731         int cycled = 1;
2732         handle_t *handle = NULL;
2733         struct mpage_da_data mpd;
2734         struct inode *inode = mapping->host;
2735         int needed_blocks, rsv_blocks = 0, ret = 0;
2736         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2737         bool done;
2738         struct blk_plug plug;
2739         bool give_up_on_write = false;
2740
2741         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
2742                 return -EIO;
2743
2744         percpu_down_read(&sbi->s_journal_flag_rwsem);
2745         trace_ext4_writepages(inode, wbc);
2746
2747         /*
2748          * No pages to write? This is mainly a kludge to avoid starting
2749          * a transaction for special inodes like journal inode on last iput()
2750          * because that could violate lock ordering on umount
2751          */
2752         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2753                 goto out_writepages;
2754
2755         if (ext4_should_journal_data(inode)) {
2756                 ret = generic_writepages(mapping, wbc);
2757                 goto out_writepages;
2758         }
2759
2760         /*
2761          * If the filesystem has aborted, it is read-only, so return
2762          * right away instead of dumping stack traces later on that
2763          * will obscure the real source of the problem.  We test
2764          * EXT4_MF_FS_ABORTED instead of sb->s_flag's SB_RDONLY because
2765          * the latter could be true if the filesystem is mounted
2766          * read-only, and in that case, ext4_writepages should
2767          * *never* be called, so if that ever happens, we would want
2768          * the stack trace.
2769          */
2770         if (unlikely(ext4_forced_shutdown(EXT4_SB(mapping->host->i_sb)) ||
2771                      sbi->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2772                 ret = -EROFS;
2773                 goto out_writepages;
2774         }
2775
2776         if (ext4_should_dioread_nolock(inode)) {
2777                 /*
2778                  * We may need to convert up to one extent per block in
2779                  * the page and we may dirty the inode.
2780                  */
2781                 rsv_blocks = 1 + (PAGE_SIZE >> inode->i_blkbits);
2782         }
2783
2784         /*
2785          * If we have inline data and arrive here, it means that
2786          * we will soon create the block for the 1st page, so
2787          * we'd better clear the inline data here.
2788          */
2789         if (ext4_has_inline_data(inode)) {
2790                 /* Just inode will be modified... */
2791                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
2792                 if (IS_ERR(handle)) {
2793                         ret = PTR_ERR(handle);
2794                         goto out_writepages;
2795                 }
2796                 BUG_ON(ext4_test_inode_state(inode,
2797                                 EXT4_STATE_MAY_INLINE_DATA));
2798                 ext4_destroy_inline_data(handle, inode);
2799                 ext4_journal_stop(handle);
2800         }
2801
2802         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2803                 range_whole = 1;
2804
2805         if (wbc->range_cyclic) {
2806                 writeback_index = mapping->writeback_index;
2807                 if (writeback_index)
2808                         cycled = 0;
2809                 mpd.first_page = writeback_index;
2810                 mpd.last_page = -1;
2811         } else {
2812                 mpd.first_page = wbc->range_start >> PAGE_SHIFT;
2813                 mpd.last_page = wbc->range_end >> PAGE_SHIFT;
2814         }
2815
2816         mpd.inode = inode;
2817         mpd.wbc = wbc;
2818         ext4_io_submit_init(&mpd.io_submit, wbc);
2819 retry:
2820         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2821                 tag_pages_for_writeback(mapping, mpd.first_page, mpd.last_page);
2822         done = false;
2823         blk_start_plug(&plug);
2824
2825         /*
2826          * First writeback pages that don't need mapping - we can avoid
2827          * starting a transaction unnecessarily and also avoid being blocked
2828          * in the block layer on device congestion while having transaction
2829          * started.
2830          */
2831         mpd.do_map = 0;
2832         mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2833         if (!mpd.io_submit.io_end) {
2834                 ret = -ENOMEM;
2835                 goto unplug;
2836         }
2837         ret = mpage_prepare_extent_to_map(&mpd);
2838         /* Submit prepared bio */
2839         ext4_io_submit(&mpd.io_submit);
2840         ext4_put_io_end_defer(mpd.io_submit.io_end);
2841         mpd.io_submit.io_end = NULL;
2842         /* Unlock pages we didn't use */
2843         mpage_release_unused_pages(&mpd, false);
2844         if (ret < 0)
2845                 goto unplug;
2846
2847         while (!done && mpd.first_page <= mpd.last_page) {
2848                 /* For each extent of pages we use new io_end */
2849                 mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2850                 if (!mpd.io_submit.io_end) {
2851                         ret = -ENOMEM;
2852                         break;
2853                 }
2854
2855                 /*
2856                  * We have two constraints: We find one extent to map and we
2857                  * must always write out whole page (makes a difference when
2858                  * blocksize < pagesize) so that we don't block on IO when we
2859                  * try to write out the rest of the page. Journalled mode is
2860                  * not supported by delalloc.
2861                  */
2862                 BUG_ON(ext4_should_journal_data(inode));
2863                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2864
2865                 /* start a new transaction */
2866                 handle = ext4_journal_start_with_reserve(inode,
2867                                 EXT4_HT_WRITE_PAGE, needed_blocks, rsv_blocks);
2868                 if (IS_ERR(handle)) {
2869                         ret = PTR_ERR(handle);
2870                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2871                                "%ld pages, ino %lu; err %d", __func__,
2872                                 wbc->nr_to_write, inode->i_ino, ret);
2873                         /* Release allocated io_end */
2874                         ext4_put_io_end(mpd.io_submit.io_end);
2875                         mpd.io_submit.io_end = NULL;
2876                         break;
2877                 }
2878                 mpd.do_map = 1;
2879
2880                 trace_ext4_da_write_pages(inode, mpd.first_page, mpd.wbc);
2881                 ret = mpage_prepare_extent_to_map(&mpd);
2882                 if (!ret) {
2883                         if (mpd.map.m_len)
2884                                 ret = mpage_map_and_submit_extent(handle, &mpd,
2885                                         &give_up_on_write);
2886                         else {
2887                                 /*
2888                                  * We scanned the whole range (or exhausted
2889                                  * nr_to_write), submitted what was mapped and
2890                                  * didn't find anything needing mapping. We are
2891                                  * done.
2892                                  */
2893                                 done = true;
2894                         }
2895                 }
2896                 /*
2897                  * Caution: If the handle is synchronous,
2898                  * ext4_journal_stop() can wait for transaction commit
2899                  * to finish which may depend on writeback of pages to
2900                  * complete or on page lock to be released.  In that
2901                  * case, we have to wait until after after we have
2902                  * submitted all the IO, released page locks we hold,
2903                  * and dropped io_end reference (for extent conversion
2904                  * to be able to complete) before stopping the handle.
2905                  */
2906                 if (!ext4_handle_valid(handle) || handle->h_sync == 0) {
2907                         ext4_journal_stop(handle);
2908                         handle = NULL;
2909                         mpd.do_map = 0;
2910                 }
2911                 /* Submit prepared bio */
2912                 ext4_io_submit(&mpd.io_submit);
2913                 /* Unlock pages we didn't use */
2914                 mpage_release_unused_pages(&mpd, give_up_on_write);
2915                 /*
2916                  * Drop our io_end reference we got from init. We have
2917                  * to be careful and use deferred io_end finishing if
2918                  * we are still holding the transaction as we can
2919                  * release the last reference to io_end which may end
2920                  * up doing unwritten extent conversion.
2921                  */
2922                 if (handle) {
2923                         ext4_put_io_end_defer(mpd.io_submit.io_end);
2924                         ext4_journal_stop(handle);
2925                 } else
2926                         ext4_put_io_end(mpd.io_submit.io_end);
2927                 mpd.io_submit.io_end = NULL;
2928
2929                 if (ret == -ENOSPC && sbi->s_journal) {
2930                         /*
2931                          * Commit the transaction which would
2932                          * free blocks released in the transaction
2933                          * and try again
2934                          */
2935                         jbd2_journal_force_commit_nested(sbi->s_journal);
2936                         ret = 0;
2937                         continue;
2938                 }
2939                 /* Fatal error - ENOMEM, EIO... */
2940                 if (ret)
2941                         break;
2942         }
2943 unplug:
2944         blk_finish_plug(&plug);
2945         if (!ret && !cycled && wbc->nr_to_write > 0) {
2946                 cycled = 1;
2947                 mpd.last_page = writeback_index - 1;
2948                 mpd.first_page = 0;
2949                 goto retry;
2950         }
2951
2952         /* Update index */
2953         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2954                 /*
2955                  * Set the writeback_index so that range_cyclic
2956                  * mode will write it back later
2957                  */
2958                 mapping->writeback_index = mpd.first_page;
2959
2960 out_writepages:
2961         trace_ext4_writepages_result(inode, wbc, ret,
2962                                      nr_to_write - wbc->nr_to_write);
2963         percpu_up_read(&sbi->s_journal_flag_rwsem);
2964         return ret;
2965 }
2966
2967 static int ext4_dax_writepages(struct address_space *mapping,
2968                                struct writeback_control *wbc)
2969 {
2970         int ret;
2971         long nr_to_write = wbc->nr_to_write;
2972         struct inode *inode = mapping->host;
2973         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2974
2975         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
2976                 return -EIO;
2977
2978         percpu_down_read(&sbi->s_journal_flag_rwsem);
2979         trace_ext4_writepages(inode, wbc);
2980
2981         ret = dax_writeback_mapping_range(mapping, inode->i_sb->s_bdev, wbc);
2982         trace_ext4_writepages_result(inode, wbc, ret,
2983                                      nr_to_write - wbc->nr_to_write);
2984         percpu_up_read(&sbi->s_journal_flag_rwsem);
2985         return ret;
2986 }
2987
2988 static int ext4_nonda_switch(struct super_block *sb)
2989 {
2990         s64 free_clusters, dirty_clusters;
2991         struct ext4_sb_info *sbi = EXT4_SB(sb);
2992
2993         /*
2994          * switch to non delalloc mode if we are running low
2995          * on free block. The free block accounting via percpu
2996          * counters can get slightly wrong with percpu_counter_batch getting
2997          * accumulated on each CPU without updating global counters
2998          * Delalloc need an accurate free block accounting. So switch
2999          * to non delalloc when we are near to error range.
3000          */
3001         free_clusters =
3002                 percpu_counter_read_positive(&sbi->s_freeclusters_counter);
3003         dirty_clusters =
3004                 percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
3005         /*
3006          * Start pushing delalloc when 1/2 of free blocks are dirty.
3007          */
3008         if (dirty_clusters && (free_clusters < 2 * dirty_clusters))
3009                 try_to_writeback_inodes_sb(sb, WB_REASON_FS_FREE_SPACE);
3010
3011         if (2 * free_clusters < 3 * dirty_clusters ||
3012             free_clusters < (dirty_clusters + EXT4_FREECLUSTERS_WATERMARK)) {
3013                 /*
3014                  * free block count is less than 150% of dirty blocks
3015                  * or free blocks is less than watermark
3016                  */
3017                 return 1;
3018         }
3019         return 0;
3020 }
3021
3022 /* We always reserve for an inode update; the superblock could be there too */
3023 static int ext4_da_write_credits(struct inode *inode, loff_t pos, unsigned len)
3024 {
3025         if (likely(ext4_has_feature_large_file(inode->i_sb)))
3026                 return 1;
3027
3028         if (pos + len <= 0x7fffffffULL)
3029                 return 1;
3030
3031         /* We might need to update the superblock to set LARGE_FILE */
3032         return 2;
3033 }
3034
3035 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3036                                loff_t pos, unsigned len, unsigned flags,
3037                                struct page **pagep, void **fsdata)
3038 {
3039         int ret, retries = 0;
3040         struct page *page;
3041         pgoff_t index;
3042         struct inode *inode = mapping->host;
3043         handle_t *handle;
3044
3045         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
3046                 return -EIO;
3047
3048         index = pos >> PAGE_SHIFT;
3049
3050         if (ext4_nonda_switch(inode->i_sb) ||
3051             S_ISLNK(inode->i_mode)) {
3052                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3053                 return ext4_write_begin(file, mapping, pos,
3054                                         len, flags, pagep, fsdata);
3055         }
3056         *fsdata = (void *)0;
3057         trace_ext4_da_write_begin(inode, pos, len, flags);
3058
3059         if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
3060                 ret = ext4_da_write_inline_data_begin(mapping, inode,
3061                                                       pos, len, flags,
3062                                                       pagep, fsdata);
3063                 if (ret < 0)
3064                         return ret;
3065                 if (ret == 1)
3066                         return 0;
3067         }
3068
3069         /*
3070          * grab_cache_page_write_begin() can take a long time if the
3071          * system is thrashing due to memory pressure, or if the page
3072          * is being written back.  So grab it first before we start
3073          * the transaction handle.  This also allows us to allocate
3074          * the page (if needed) without using GFP_NOFS.
3075          */
3076 retry_grab:
3077         page = grab_cache_page_write_begin(mapping, index, flags);
3078         if (!page)
3079                 return -ENOMEM;
3080         unlock_page(page);
3081
3082         /*
3083          * With delayed allocation, we don't log the i_disksize update
3084          * if there is delayed block allocation. But we still need
3085          * to journalling the i_disksize update if writes to the end
3086          * of file which has an already mapped buffer.
3087          */
3088 retry_journal:
3089         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
3090                                 ext4_da_write_credits(inode, pos, len));
3091         if (IS_ERR(handle)) {
3092                 put_page(page);
3093                 return PTR_ERR(handle);
3094         }
3095
3096         lock_page(page);
3097         if (page->mapping != mapping) {
3098                 /* The page got truncated from under us */
3099                 unlock_page(page);
3100                 put_page(page);
3101                 ext4_journal_stop(handle);
3102                 goto retry_grab;
3103         }
3104         /* In case writeback began while the page was unlocked */
3105         wait_for_stable_page(page);
3106
3107 #ifdef CONFIG_EXT4_FS_ENCRYPTION
3108         ret = ext4_block_write_begin(page, pos, len,
3109                                      ext4_da_get_block_prep);
3110 #else
3111         ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
3112 #endif
3113         if (ret < 0) {
3114                 unlock_page(page);
3115                 ext4_journal_stop(handle);
3116                 /*
3117                  * block_write_begin may have instantiated a few blocks
3118                  * outside i_size.  Trim these off again. Don't need
3119                  * i_size_read because we hold i_mutex.
3120                  */
3121                 if (pos + len > inode->i_size)
3122                         ext4_truncate_failed_write(inode);
3123
3124                 if (ret == -ENOSPC &&
3125                     ext4_should_retry_alloc(inode->i_sb, &retries))
3126                         goto retry_journal;
3127
3128                 put_page(page);
3129                 return ret;
3130         }
3131
3132         *pagep = page;
3133         return ret;
3134 }
3135
3136 /*
3137  * Check if we should update i_disksize
3138  * when write to the end of file but not require block allocation
3139  */
3140 static int ext4_da_should_update_i_disksize(struct page *page,
3141                                             unsigned long offset)
3142 {
3143         struct buffer_head *bh;
3144         struct inode *inode = page->mapping->host;
3145         unsigned int idx;
3146         int i;
3147
3148         bh = page_buffers(page);
3149         idx = offset >> inode->i_blkbits;
3150
3151         for (i = 0; i < idx; i++)
3152                 bh = bh->b_this_page;
3153
3154         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3155                 return 0;
3156         return 1;
3157 }
3158
3159 static int ext4_da_write_end(struct file *file,
3160                              struct address_space *mapping,
3161                              loff_t pos, unsigned len, unsigned copied,
3162                              struct page *page, void *fsdata)
3163 {
3164         struct inode *inode = mapping->host;
3165         int ret = 0, ret2;
3166         handle_t *handle = ext4_journal_current_handle();
3167         loff_t new_i_size;
3168         unsigned long start, end;
3169         int write_mode = (int)(unsigned long)fsdata;
3170
3171         if (write_mode == FALL_BACK_TO_NONDELALLOC)
3172                 return ext4_write_end(file, mapping, pos,
3173                                       len, copied, page, fsdata);
3174
3175         trace_ext4_da_write_end(inode, pos, len, copied);
3176         start = pos & (PAGE_SIZE - 1);
3177         end = start + copied - 1;
3178
3179         /*
3180          * generic_write_end() will run mark_inode_dirty() if i_size
3181          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3182          * into that.
3183          */
3184         new_i_size = pos + copied;
3185         if (copied && new_i_size > EXT4_I(inode)->i_disksize) {
3186                 if (ext4_has_inline_data(inode) ||
3187                     ext4_da_should_update_i_disksize(page, end)) {
3188                         ext4_update_i_disksize(inode, new_i_size);
3189                         /* We need to mark inode dirty even if
3190                          * new_i_size is less that inode->i_size
3191                          * bu greater than i_disksize.(hint delalloc)
3192                          */
3193                         ext4_mark_inode_dirty(handle, inode);
3194                 }
3195         }
3196
3197         if (write_mode != CONVERT_INLINE_DATA &&
3198             ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA) &&
3199             ext4_has_inline_data(inode))
3200                 ret2 = ext4_da_write_inline_data_end(inode, pos, len, copied,
3201                                                      page);
3202         else
3203                 ret2 = generic_write_end(file, mapping, pos, len, copied,
3204                                                         page, fsdata);
3205
3206         copied = ret2;
3207         if (ret2 < 0)
3208                 ret = ret2;
3209         ret2 = ext4_journal_stop(handle);
3210         if (!ret)
3211                 ret = ret2;
3212
3213         return ret ? ret : copied;
3214 }
3215
3216 static void ext4_da_invalidatepage(struct page *page, unsigned int offset,
3217                                    unsigned int length)
3218 {
3219         /*
3220          * Drop reserved blocks
3221          */
3222         BUG_ON(!PageLocked(page));
3223         if (!page_has_buffers(page))
3224                 goto out;
3225
3226         ext4_da_page_release_reservation(page, offset, length);
3227
3228 out:
3229         ext4_invalidatepage(page, offset, length);
3230
3231         return;
3232 }
3233
3234 /*
3235  * Force all delayed allocation blocks to be allocated for a given inode.
3236  */
3237 int ext4_alloc_da_blocks(struct inode *inode)
3238 {
3239         trace_ext4_alloc_da_blocks(inode);
3240
3241         if (!EXT4_I(inode)->i_reserved_data_blocks)
3242                 return 0;
3243
3244         /*
3245          * We do something simple for now.  The filemap_flush() will
3246          * also start triggering a write of the data blocks, which is
3247          * not strictly speaking necessary (and for users of
3248          * laptop_mode, not even desirable).  However, to do otherwise
3249          * would require replicating code paths in:
3250          *
3251          * ext4_writepages() ->
3252          *    write_cache_pages() ---> (via passed in callback function)
3253          *        __mpage_da_writepage() -->
3254          *           mpage_add_bh_to_extent()
3255          *           mpage_da_map_blocks()
3256          *
3257          * The problem is that write_cache_pages(), located in
3258          * mm/page-writeback.c, marks pages clean in preparation for
3259          * doing I/O, which is not desirable if we're not planning on
3260          * doing I/O at all.
3261          *
3262          * We could call write_cache_pages(), and then redirty all of
3263          * the pages by calling redirty_page_for_writepage() but that
3264          * would be ugly in the extreme.  So instead we would need to
3265          * replicate parts of the code in the above functions,
3266          * simplifying them because we wouldn't actually intend to
3267          * write out the pages, but rather only collect contiguous
3268          * logical block extents, call the multi-block allocator, and
3269          * then update the buffer heads with the block allocations.
3270          *
3271          * For now, though, we'll cheat by calling filemap_flush(),
3272          * which will map the blocks, and start the I/O, but not
3273          * actually wait for the I/O to complete.
3274          */
3275         return filemap_flush(inode->i_mapping);
3276 }
3277
3278 /*
3279  * bmap() is special.  It gets used by applications such as lilo and by
3280  * the swapper to find the on-disk block of a specific piece of data.
3281  *
3282  * Naturally, this is dangerous if the block concerned is still in the
3283  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3284  * filesystem and enables swap, then they may get a nasty shock when the
3285  * data getting swapped to that swapfile suddenly gets overwritten by
3286  * the original zero's written out previously to the journal and
3287  * awaiting writeback in the kernel's buffer cache.
3288  *
3289  * So, if we see any bmap calls here on a modified, data-journaled file,
3290  * take extra steps to flush any blocks which might be in the cache.
3291  */