d0dd585add6a005684c569d6edaecc481f926f2d
[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_find_delalloc_range(inode, 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_find_delalloc_range(inode, 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 static 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 to_release = 0, contiguous_blks = 0;
1638         struct buffer_head *head, *bh;
1639         unsigned int curr_off = 0;
1640         struct inode *inode = page->mapping->host;
1641         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1642         unsigned int stop = offset + length;
1643         int num_clusters;
1644         ext4_fsblk_t lblk;
1645
1646         BUG_ON(stop > PAGE_SIZE || stop < length);
1647
1648         head = page_buffers(page);
1649         bh = head;
1650         do {
1651                 unsigned int next_off = curr_off + bh->b_size;
1652
1653                 if (next_off > stop)
1654                         break;
1655
1656                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1657                         to_release++;
1658                         contiguous_blks++;
1659                         clear_buffer_delay(bh);
1660                 } else if (contiguous_blks) {
1661                         lblk = page->index <<
1662                                (PAGE_SHIFT - inode->i_blkbits);
1663                         lblk += (curr_off >> inode->i_blkbits) -
1664                                 contiguous_blks;
1665                         ext4_es_remove_extent(inode, lblk, contiguous_blks);
1666                         contiguous_blks = 0;
1667                 }
1668                 curr_off = next_off;
1669         } while ((bh = bh->b_this_page) != head);
1670
1671         if (contiguous_blks) {
1672                 lblk = page->index << (PAGE_SHIFT - inode->i_blkbits);
1673                 lblk += (curr_off >> inode->i_blkbits) - contiguous_blks;
1674                 ext4_es_remove_extent(inode, lblk, contiguous_blks);
1675         }
1676
1677         /* If we have released all the blocks belonging to a cluster, then we
1678          * need to release the reserved space for that cluster. */
1679         num_clusters = EXT4_NUM_B2C(sbi, to_release);
1680         while (num_clusters > 0) {
1681                 lblk = (page->index << (PAGE_SHIFT - inode->i_blkbits)) +
1682                         ((num_clusters - 1) << sbi->s_cluster_bits);
1683                 if (sbi->s_cluster_ratio == 1 ||
1684                     !ext4_find_delalloc_cluster(inode, lblk))
1685                         ext4_da_release_space(inode, 1);
1686
1687                 num_clusters--;
1688         }
1689 }
1690
1691 /*
1692  * Delayed allocation stuff
1693  */
1694
1695 struct mpage_da_data {
1696         struct inode *inode;
1697         struct writeback_control *wbc;
1698
1699         pgoff_t first_page;     /* The first page to write */
1700         pgoff_t next_page;      /* Current page to examine */
1701         pgoff_t last_page;      /* Last page to examine */
1702         /*
1703          * Extent to map - this can be after first_page because that can be
1704          * fully mapped. We somewhat abuse m_flags to store whether the extent
1705          * is delalloc or unwritten.
1706          */
1707         struct ext4_map_blocks map;
1708         struct ext4_io_submit io_submit;        /* IO submission data */
1709         unsigned int do_map:1;
1710 };
1711
1712 static void mpage_release_unused_pages(struct mpage_da_data *mpd,
1713                                        bool invalidate)
1714 {
1715         int nr_pages, i;
1716         pgoff_t index, end;
1717         struct pagevec pvec;
1718         struct inode *inode = mpd->inode;
1719         struct address_space *mapping = inode->i_mapping;
1720
1721         /* This is necessary when next_page == 0. */
1722         if (mpd->first_page >= mpd->next_page)
1723                 return;
1724
1725         index = mpd->first_page;
1726         end   = mpd->next_page - 1;
1727         if (invalidate) {
1728                 ext4_lblk_t start, last;
1729                 start = index << (PAGE_SHIFT - inode->i_blkbits);
1730                 last = end << (PAGE_SHIFT - inode->i_blkbits);
1731                 ext4_es_remove_extent(inode, start, last - start + 1);
1732         }
1733
1734         pagevec_init(&pvec);
1735         while (index <= end) {
1736                 nr_pages = pagevec_lookup_range(&pvec, mapping, &index, end);
1737                 if (nr_pages == 0)
1738                         break;
1739                 for (i = 0; i < nr_pages; i++) {
1740                         struct page *page = pvec.pages[i];
1741
1742                         BUG_ON(!PageLocked(page));
1743                         BUG_ON(PageWriteback(page));
1744                         if (invalidate) {
1745                                 if (page_mapped(page))
1746                                         clear_page_dirty_for_io(page);
1747                                 block_invalidatepage(page, 0, PAGE_SIZE);
1748                                 ClearPageUptodate(page);
1749                         }
1750                         unlock_page(page);
1751                 }
1752                 pagevec_release(&pvec);
1753         }
1754 }
1755
1756 static void ext4_print_free_blocks(struct inode *inode)
1757 {
1758         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1759         struct super_block *sb = inode->i_sb;
1760         struct ext4_inode_info *ei = EXT4_I(inode);
1761
1762         ext4_msg(sb, KERN_CRIT, "Total free blocks count %lld",
1763                EXT4_C2B(EXT4_SB(inode->i_sb),
1764                         ext4_count_free_clusters(sb)));
1765         ext4_msg(sb, KERN_CRIT, "Free/Dirty block details");
1766         ext4_msg(sb, KERN_CRIT, "free_blocks=%lld",
1767                (long long) EXT4_C2B(EXT4_SB(sb),
1768                 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1769         ext4_msg(sb, KERN_CRIT, "dirty_blocks=%lld",
1770                (long long) EXT4_C2B(EXT4_SB(sb),
1771                 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1772         ext4_msg(sb, KERN_CRIT, "Block reservation details");
1773         ext4_msg(sb, KERN_CRIT, "i_reserved_data_blocks=%u",
1774                  ei->i_reserved_data_blocks);
1775         return;
1776 }
1777
1778 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1779 {
1780         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1781 }
1782
1783 /*
1784  * This function is grabs code from the very beginning of
1785  * ext4_map_blocks, but assumes that the caller is from delayed write
1786  * time. This function looks up the requested blocks and sets the
1787  * buffer delay bit under the protection of i_data_sem.
1788  */
1789 static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1790                               struct ext4_map_blocks *map,
1791                               struct buffer_head *bh)
1792 {
1793         struct extent_status es;
1794         int retval;
1795         sector_t invalid_block = ~((sector_t) 0xffff);
1796 #ifdef ES_AGGRESSIVE_TEST
1797         struct ext4_map_blocks orig_map;
1798
1799         memcpy(&orig_map, map, sizeof(*map));
1800 #endif
1801
1802         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1803                 invalid_block = ~0;
1804
1805         map->m_flags = 0;
1806         ext_debug("ext4_da_map_blocks(): inode %lu, max_blocks %u,"
1807                   "logical block %lu\n", inode->i_ino, map->m_len,
1808                   (unsigned long) map->m_lblk);
1809
1810         /* Lookup extent status tree firstly */
1811         if (ext4_es_lookup_extent(inode, iblock, &es)) {
1812                 if (ext4_es_is_hole(&es)) {
1813                         retval = 0;
1814                         down_read(&EXT4_I(inode)->i_data_sem);
1815                         goto add_delayed;
1816                 }
1817
1818                 /*
1819                  * Delayed extent could be allocated by fallocate.
1820                  * So we need to check it.
1821                  */
1822                 if (ext4_es_is_delayed(&es) && !ext4_es_is_unwritten(&es)) {
1823                         map_bh(bh, inode->i_sb, invalid_block);
1824                         set_buffer_new(bh);
1825                         set_buffer_delay(bh);
1826                         return 0;
1827                 }
1828
1829                 map->m_pblk = ext4_es_pblock(&es) + iblock - es.es_lblk;
1830                 retval = es.es_len - (iblock - es.es_lblk);
1831                 if (retval > map->m_len)
1832                         retval = map->m_len;
1833                 map->m_len = retval;
1834                 if (ext4_es_is_written(&es))
1835                         map->m_flags |= EXT4_MAP_MAPPED;
1836                 else if (ext4_es_is_unwritten(&es))
1837                         map->m_flags |= EXT4_MAP_UNWRITTEN;
1838                 else
1839                         BUG_ON(1);
1840
1841 #ifdef ES_AGGRESSIVE_TEST
1842                 ext4_map_blocks_es_recheck(NULL, inode, map, &orig_map, 0);
1843 #endif
1844                 return retval;
1845         }
1846
1847         /*
1848          * Try to see if we can get the block without requesting a new
1849          * file system block.
1850          */
1851         down_read(&EXT4_I(inode)->i_data_sem);
1852         if (ext4_has_inline_data(inode))
1853                 retval = 0;
1854         else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1855                 retval = ext4_ext_map_blocks(NULL, inode, map, 0);
1856         else
1857                 retval = ext4_ind_map_blocks(NULL, inode, map, 0);
1858
1859 add_delayed:
1860         if (retval == 0) {
1861                 int ret;
1862                 /*
1863                  * XXX: __block_prepare_write() unmaps passed block,
1864                  * is it OK?
1865                  */
1866                 /*
1867                  * If the block was allocated from previously allocated cluster,
1868                  * then we don't need to reserve it again. However we still need
1869                  * to reserve metadata for every block we're going to write.
1870                  */
1871                 if (EXT4_SB(inode->i_sb)->s_cluster_ratio == 1 ||
1872                     !ext4_find_delalloc_cluster(inode, map->m_lblk)) {
1873                         ret = ext4_da_reserve_space(inode);
1874                         if (ret) {
1875                                 /* not enough space to reserve */
1876                                 retval = ret;
1877                                 goto out_unlock;
1878                         }
1879                 }
1880
1881                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1882                                             ~0, EXTENT_STATUS_DELAYED);
1883                 if (ret) {
1884                         retval = ret;
1885                         goto out_unlock;
1886                 }
1887
1888                 map_bh(bh, inode->i_sb, invalid_block);
1889                 set_buffer_new(bh);
1890                 set_buffer_delay(bh);
1891         } else if (retval > 0) {
1892                 int ret;
1893                 unsigned int status;
1894
1895                 if (unlikely(retval != map->m_len)) {
1896                         ext4_warning(inode->i_sb,
1897                                      "ES len assertion failed for inode "
1898                                      "%lu: retval %d != map->m_len %d",
1899                                      inode->i_ino, retval, map->m_len);
1900                         WARN_ON(1);
1901                 }
1902
1903                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
1904                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
1905                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1906                                             map->m_pblk, status);
1907                 if (ret != 0)
1908                         retval = ret;
1909         }
1910
1911 out_unlock:
1912         up_read((&EXT4_I(inode)->i_data_sem));
1913
1914         return retval;
1915 }
1916
1917 /*
1918  * This is a special get_block_t callback which is used by
1919  * ext4_da_write_begin().  It will either return mapped block or
1920  * reserve space for a single block.
1921  *
1922  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1923  * We also have b_blocknr = -1 and b_bdev initialized properly
1924  *
1925  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1926  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1927  * initialized properly.
1928  */
1929 int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1930                            struct buffer_head *bh, int create)
1931 {
1932         struct ext4_map_blocks map;
1933         int ret = 0;
1934
1935         BUG_ON(create == 0);
1936         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1937
1938         map.m_lblk = iblock;
1939         map.m_len = 1;
1940
1941         /*
1942          * first, we need to know whether the block is allocated already
1943          * preallocated blocks are unmapped but should treated
1944          * the same as allocated blocks.
1945          */
1946         ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1947         if (ret <= 0)
1948                 return ret;
1949
1950         map_bh(bh, inode->i_sb, map.m_pblk);
1951         ext4_update_bh_state(bh, map.m_flags);
1952
1953         if (buffer_unwritten(bh)) {
1954                 /* A delayed write to unwritten bh should be marked
1955                  * new and mapped.  Mapped ensures that we don't do
1956                  * get_block multiple times when we write to the same
1957                  * offset and new ensures that we do proper zero out
1958                  * for partial write.
1959                  */
1960                 set_buffer_new(bh);
1961                 set_buffer_mapped(bh);
1962         }
1963         return 0;
1964 }
1965
1966 static int bget_one(handle_t *handle, struct buffer_head *bh)
1967 {
1968         get_bh(bh);
1969         return 0;
1970 }
1971
1972 static int bput_one(handle_t *handle, struct buffer_head *bh)
1973 {
1974         put_bh(bh);
1975         return 0;
1976 }
1977
1978 static int __ext4_journalled_writepage(struct page *page,
1979                                        unsigned int len)
1980 {
1981         struct address_space *mapping = page->mapping;
1982         struct inode *inode = mapping->host;
1983         struct buffer_head *page_bufs = NULL;
1984         handle_t *handle = NULL;
1985         int ret = 0, err = 0;
1986         int inline_data = ext4_has_inline_data(inode);
1987         struct buffer_head *inode_bh = NULL;
1988
1989         ClearPageChecked(page);
1990
1991         if (inline_data) {
1992                 BUG_ON(page->index != 0);
1993                 BUG_ON(len > ext4_get_max_inline_size(inode));
1994                 inode_bh = ext4_journalled_write_inline_data(inode, len, page);
1995                 if (inode_bh == NULL)
1996                         goto out;
1997         } else {
1998                 page_bufs = page_buffers(page);
1999                 if (!page_bufs) {
2000                         BUG();
2001                         goto out;
2002                 }
2003                 ext4_walk_page_buffers(handle, page_bufs, 0, len,
2004                                        NULL, bget_one);
2005         }
2006         /*
2007          * We need to release the page lock before we start the
2008          * journal, so grab a reference so the page won't disappear
2009          * out from under us.
2010          */
2011         get_page(page);
2012         unlock_page(page);
2013
2014         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
2015                                     ext4_writepage_trans_blocks(inode));
2016         if (IS_ERR(handle)) {
2017                 ret = PTR_ERR(handle);
2018                 put_page(page);
2019                 goto out_no_pagelock;
2020         }
2021         BUG_ON(!ext4_handle_valid(handle));
2022
2023         lock_page(page);
2024         put_page(page);
2025         if (page->mapping != mapping) {
2026                 /* The page got truncated from under us */
2027                 ext4_journal_stop(handle);
2028                 ret = 0;
2029                 goto out;
2030         }
2031
2032         if (inline_data) {
2033                 ret = ext4_mark_inode_dirty(handle, inode);
2034         } else {
2035                 ret = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
2036                                              do_journal_get_write_access);
2037
2038                 err = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
2039                                              write_end_fn);
2040         }
2041         if (ret == 0)
2042                 ret = err;
2043         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
2044         err = ext4_journal_stop(handle);
2045         if (!ret)
2046                 ret = err;
2047
2048         if (!ext4_has_inline_data(inode))
2049                 ext4_walk_page_buffers(NULL, page_bufs, 0, len,
2050                                        NULL, bput_one);
2051         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2052 out:
2053         unlock_page(page);
2054 out_no_pagelock:
2055         brelse(inode_bh);
2056         return ret;
2057 }
2058
2059 /*
2060  * Note that we don't need to start a transaction unless we're journaling data
2061  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2062  * need to file the inode to the transaction's list in ordered mode because if
2063  * we are writing back data added by write(), the inode is already there and if
2064  * we are writing back data modified via mmap(), no one guarantees in which
2065  * transaction the data will hit the disk. In case we are journaling data, we
2066  * cannot start transaction directly because transaction start ranks above page
2067  * lock so we have to do some magic.
2068  *
2069  * This function can get called via...
2070  *   - ext4_writepages after taking page lock (have journal handle)
2071  *   - journal_submit_inode_data_buffers (no journal handle)
2072  *   - shrink_page_list via the kswapd/direct reclaim (no journal handle)
2073  *   - grab_page_cache when doing write_begin (have journal handle)
2074  *
2075  * We don't do any block allocation in this function. If we have page with
2076  * multiple blocks we need to write those buffer_heads that are mapped. This
2077  * is important for mmaped based write. So if we do with blocksize 1K
2078  * truncate(f, 1024);
2079  * a = mmap(f, 0, 4096);
2080  * a[0] = 'a';
2081  * truncate(f, 4096);
2082  * we have in the page first buffer_head mapped via page_mkwrite call back
2083  * but other buffer_heads would be unmapped but dirty (dirty done via the
2084  * do_wp_page). So writepage should write the first block. If we modify
2085  * the mmap area beyond 1024 we will again get a page_fault and the
2086  * page_mkwrite callback will do the block allocation and mark the
2087  * buffer_heads mapped.
2088  *
2089  * We redirty the page if we have any buffer_heads that is either delay or
2090  * unwritten in the page.
2091  *
2092  * We can get recursively called as show below.
2093  *
2094  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2095  *              ext4_writepage()
2096  *
2097  * But since we don't do any block allocation we should not deadlock.
2098  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2099  */
2100 static int ext4_writepage(struct page *page,
2101                           struct writeback_control *wbc)
2102 {
2103         int ret = 0;
2104         loff_t size;
2105         unsigned int len;
2106         struct buffer_head *page_bufs = NULL;
2107         struct inode *inode = page->mapping->host;
2108         struct ext4_io_submit io_submit;
2109         bool keep_towrite = false;
2110
2111         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb)))) {
2112                 ext4_invalidatepage(page, 0, PAGE_SIZE);
2113                 unlock_page(page);
2114                 return -EIO;
2115         }
2116
2117         trace_ext4_writepage(page);
2118         size = i_size_read(inode);
2119         if (page->index == size >> PAGE_SHIFT)
2120                 len = size & ~PAGE_MASK;
2121         else
2122                 len = PAGE_SIZE;
2123
2124         page_bufs = page_buffers(page);
2125         /*
2126          * We cannot do block allocation or other extent handling in this
2127          * function. If there are buffers needing that, we have to redirty
2128          * the page. But we may reach here when we do a journal commit via
2129          * journal_submit_inode_data_buffers() and in that case we must write
2130          * allocated buffers to achieve data=ordered mode guarantees.
2131          *
2132          * Also, if there is only one buffer per page (the fs block
2133          * size == the page size), if one buffer needs block
2134          * allocation or needs to modify the extent tree to clear the
2135          * unwritten flag, we know that the page can't be written at
2136          * all, so we might as well refuse the write immediately.
2137          * Unfortunately if the block size != page size, we can't as
2138          * easily detect this case using ext4_walk_page_buffers(), but
2139          * for the extremely common case, this is an optimization that
2140          * skips a useless round trip through ext4_bio_write_page().
2141          */
2142         if (ext4_walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2143                                    ext4_bh_delay_or_unwritten)) {
2144                 redirty_page_for_writepage(wbc, page);
2145                 if ((current->flags & PF_MEMALLOC) ||
2146                     (inode->i_sb->s_blocksize == PAGE_SIZE)) {
2147                         /*
2148                          * For memory cleaning there's no point in writing only
2149                          * some buffers. So just bail out. Warn if we came here
2150                          * from direct reclaim.
2151                          */
2152                         WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD))
2153                                                         == PF_MEMALLOC);
2154                         unlock_page(page);
2155                         return 0;
2156                 }
2157                 keep_towrite = true;
2158         }
2159
2160         if (PageChecked(page) && ext4_should_journal_data(inode))
2161                 /*
2162                  * It's mmapped pagecache.  Add buffers and journal it.  There
2163                  * doesn't seem much point in redirtying the page here.
2164                  */
2165                 return __ext4_journalled_writepage(page, len);
2166
2167         ext4_io_submit_init(&io_submit, wbc);
2168         io_submit.io_end = ext4_init_io_end(inode, GFP_NOFS);
2169         if (!io_submit.io_end) {
2170                 redirty_page_for_writepage(wbc, page);
2171                 unlock_page(page);
2172                 return -ENOMEM;
2173         }
2174         ret = ext4_bio_write_page(&io_submit, page, len, wbc, keep_towrite);
2175         ext4_io_submit(&io_submit);
2176         /* Drop io_end reference we got from init */
2177         ext4_put_io_end_defer(io_submit.io_end);
2178         return ret;
2179 }
2180
2181 static int mpage_submit_page(struct mpage_da_data *mpd, struct page *page)
2182 {
2183         int len;
2184         loff_t size;
2185         int err;
2186
2187         BUG_ON(page->index != mpd->first_page);
2188         clear_page_dirty_for_io(page);
2189         /*
2190          * We have to be very careful here!  Nothing protects writeback path
2191          * against i_size changes and the page can be writeably mapped into
2192          * page tables. So an application can be growing i_size and writing
2193          * data through mmap while writeback runs. clear_page_dirty_for_io()
2194          * write-protects our page in page tables and the page cannot get
2195          * written to again until we release page lock. So only after
2196          * clear_page_dirty_for_io() we are safe to sample i_size for
2197          * ext4_bio_write_page() to zero-out tail of the written page. We rely
2198          * on the barrier provided by TestClearPageDirty in
2199          * clear_page_dirty_for_io() to make sure i_size is really sampled only
2200          * after page tables are updated.
2201          */
2202         size = i_size_read(mpd->inode);
2203         if (page->index == size >> PAGE_SHIFT)
2204                 len = size & ~PAGE_MASK;
2205         else
2206                 len = PAGE_SIZE;
2207         err = ext4_bio_write_page(&mpd->io_submit, page, len, mpd->wbc, false);
2208         if (!err)
2209                 mpd->wbc->nr_to_write--;
2210         mpd->first_page++;
2211
2212         return err;
2213 }
2214
2215 #define BH_FLAGS ((1 << BH_Unwritten) | (1 << BH_Delay))
2216
2217 /*
2218  * mballoc gives us at most this number of blocks...
2219  * XXX: That seems to be only a limitation of ext4_mb_normalize_request().
2220  * The rest of mballoc seems to handle chunks up to full group size.
2221  */
2222 #define MAX_WRITEPAGES_EXTENT_LEN 2048
2223
2224 /*
2225  * mpage_add_bh_to_extent - try to add bh to extent of blocks to map
2226  *
2227  * @mpd - extent of blocks
2228  * @lblk - logical number of the block in the file
2229  * @bh - buffer head we want to add to the extent
2230  *
2231  * The function is used to collect contig. blocks in the same state. If the
2232  * buffer doesn't require mapping for writeback and we haven't started the
2233  * extent of buffers to map yet, the function returns 'true' immediately - the
2234  * caller can write the buffer right away. Otherwise the function returns true
2235  * if the block has been added to the extent, false if the block couldn't be
2236  * added.
2237  */
2238 static bool mpage_add_bh_to_extent(struct mpage_da_data *mpd, ext4_lblk_t lblk,
2239                                    struct buffer_head *bh)
2240 {
2241         struct ext4_map_blocks *map = &mpd->map;
2242
2243         /* Buffer that doesn't need mapping for writeback? */
2244         if (!buffer_dirty(bh) || !buffer_mapped(bh) ||
2245             (!buffer_delay(bh) && !buffer_unwritten(bh))) {
2246                 /* So far no extent to map => we write the buffer right away */
2247                 if (map->m_len == 0)
2248                         return true;
2249                 return false;
2250         }
2251
2252         /* First block in the extent? */
2253         if (map->m_len == 0) {
2254                 /* We cannot map unless handle is started... */
2255                 if (!mpd->do_map)
2256                         return false;
2257                 map->m_lblk = lblk;
2258                 map->m_len = 1;
2259                 map->m_flags = bh->b_state & BH_FLAGS;
2260                 return true;
2261         }
2262
2263         /* Don't go larger than mballoc is willing to allocate */
2264         if (map->m_len >= MAX_WRITEPAGES_EXTENT_LEN)
2265                 return false;
2266
2267         /* Can we merge the block to our big extent? */
2268         if (lblk == map->m_lblk + map->m_len &&
2269             (bh->b_state & BH_FLAGS) == map->m_flags) {
2270                 map->m_len++;
2271                 return true;
2272         }
2273         return false;
2274 }
2275
2276 /*
2277  * mpage_process_page_bufs - submit page buffers for IO or add them to extent
2278  *
2279  * @mpd - extent of blocks for mapping
2280  * @head - the first buffer in the page
2281  * @bh - buffer we should start processing from
2282  * @lblk - logical number of the block in the file corresponding to @bh
2283  *
2284  * Walk through page buffers from @bh upto @head (exclusive) and either submit
2285  * the page for IO if all buffers in this page were mapped and there's no
2286  * accumulated extent of buffers to map or add buffers in the page to the
2287  * extent of buffers to map. The function returns 1 if the caller can continue
2288  * by processing the next page, 0 if it should stop adding buffers to the
2289  * extent to map because we cannot extend it anymore. It can also return value
2290  * < 0 in case of error during IO submission.
2291  */
2292 static int mpage_process_page_bufs(struct mpage_da_data *mpd,
2293                                    struct buffer_head *head,
2294                                    struct buffer_head *bh,
2295                                    ext4_lblk_t lblk)
2296 {
2297         struct inode *inode = mpd->inode;
2298         int err;
2299         ext4_lblk_t blocks = (i_size_read(inode) + i_blocksize(inode) - 1)
2300                                                         >> inode->i_blkbits;
2301
2302         do {
2303                 BUG_ON(buffer_locked(bh));
2304
2305                 if (lblk >= blocks || !mpage_add_bh_to_extent(mpd, lblk, bh)) {
2306                         /* Found extent to map? */
2307                         if (mpd->map.m_len)
2308                                 return 0;
2309                         /* Buffer needs mapping and handle is not started? */
2310                         if (!mpd->do_map)
2311                                 return 0;
2312                         /* Everything mapped so far and we hit EOF */
2313                         break;
2314                 }
2315         } while (lblk++, (bh = bh->b_this_page) != head);
2316         /* So far everything mapped? Submit the page for IO. */
2317         if (mpd->map.m_len == 0) {
2318                 err = mpage_submit_page(mpd, head->b_page);
2319                 if (err < 0)
2320                         return err;
2321         }
2322         return lblk < blocks;
2323 }
2324
2325 /*
2326  * mpage_map_buffers - update buffers corresponding to changed extent and
2327  *                     submit fully mapped pages for IO
2328  *
2329  * @mpd - description of extent to map, on return next extent to map
2330  *
2331  * Scan buffers corresponding to changed extent (we expect corresponding pages
2332  * to be already locked) and update buffer state according to new extent state.
2333  * We map delalloc buffers to their physical location, clear unwritten bits,
2334  * and mark buffers as uninit when we perform writes to unwritten extents
2335  * and do extent conversion after IO is finished. If the last page is not fully
2336  * mapped, we update @map to the next extent in the last page that needs
2337  * mapping. Otherwise we submit the page for IO.
2338  */
2339 static int mpage_map_and_submit_buffers(struct mpage_da_data *mpd)
2340 {
2341         struct pagevec pvec;
2342         int nr_pages, i;
2343         struct inode *inode = mpd->inode;
2344         struct buffer_head *head, *bh;
2345         int bpp_bits = PAGE_SHIFT - inode->i_blkbits;
2346         pgoff_t start, end;
2347         ext4_lblk_t lblk;
2348         sector_t pblock;
2349         int err;
2350
2351         start = mpd->map.m_lblk >> bpp_bits;
2352         end = (mpd->map.m_lblk + mpd->map.m_len - 1) >> bpp_bits;
2353         lblk = start << bpp_bits;
2354         pblock = mpd->map.m_pblk;
2355
2356         pagevec_init(&pvec);
2357         while (start <= end) {
2358                 nr_pages = pagevec_lookup_range(&pvec, inode->i_mapping,
2359                                                 &start, end);
2360                 if (nr_pages == 0)
2361                         break;
2362                 for (i = 0; i < nr_pages; i++) {
2363                         struct page *page = pvec.pages[i];
2364
2365                         bh = head = page_buffers(page);
2366                         do {
2367                                 if (lblk < mpd->map.m_lblk)
2368                                         continue;
2369                                 if (lblk >= mpd->map.m_lblk + mpd->map.m_len) {
2370                                         /*
2371                                          * Buffer after end of mapped extent.
2372                                          * Find next buffer in the page to map.
2373                                          */
2374                                         mpd->map.m_len = 0;
2375                                         mpd->map.m_flags = 0;
2376                                         /*
2377                                          * FIXME: If dioread_nolock supports
2378                                          * blocksize < pagesize, we need to make
2379                                          * sure we add size mapped so far to
2380                                          * io_end->size as the following call
2381                                          * can submit the page for IO.
2382                                          */
2383                                         err = mpage_process_page_bufs(mpd, head,
2384                                                                       bh, lblk);
2385                                         pagevec_release(&pvec);
2386                                         if (err > 0)
2387                                                 err = 0;
2388                                         return err;
2389                                 }
2390                                 if (buffer_delay(bh)) {
2391                                         clear_buffer_delay(bh);
2392                                         bh->b_blocknr = pblock++;
2393                                 }
2394                                 clear_buffer_unwritten(bh);
2395                         } while (lblk++, (bh = bh->b_this_page) != head);
2396
2397                         /*
2398                          * FIXME: This is going to break if dioread_nolock
2399                          * supports blocksize < pagesize as we will try to
2400                          * convert potentially unmapped parts of inode.
2401                          */
2402                         mpd->io_submit.io_end->size += PAGE_SIZE;
2403                         /* Page fully mapped - let IO run! */
2404                         err = mpage_submit_page(mpd, page);
2405                         if (err < 0) {
2406                                 pagevec_release(&pvec);
2407                                 return err;
2408                         }
2409                 }
2410                 pagevec_release(&pvec);
2411         }
2412         /* Extent fully mapped and matches with page boundary. We are done. */
2413         mpd->map.m_len = 0;
2414         mpd->map.m_flags = 0;
2415         return 0;
2416 }
2417
2418 static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
2419 {
2420         struct inode *inode = mpd->inode;
2421         struct ext4_map_blocks *map = &mpd->map;
2422         int get_blocks_flags;
2423         int err, dioread_nolock;
2424
2425         trace_ext4_da_write_pages_extent(inode, map);
2426         /*
2427          * Call ext4_map_blocks() to allocate any delayed allocation blocks, or
2428          * to convert an unwritten extent to be initialized (in the case
2429          * where we have written into one or more preallocated blocks).  It is
2430          * possible that we're going to need more metadata blocks than
2431          * previously reserved. However we must not fail because we're in
2432          * writeback and there is nothing we can do about it so it might result
2433          * in data loss.  So use reserved blocks to allocate metadata if
2434          * possible.
2435          *
2436          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE if
2437          * the blocks in question are delalloc blocks.  This indicates
2438          * that the blocks and quotas has already been checked when
2439          * the data was copied into the page cache.
2440          */
2441         get_blocks_flags = EXT4_GET_BLOCKS_CREATE |
2442                            EXT4_GET_BLOCKS_METADATA_NOFAIL |
2443                            EXT4_GET_BLOCKS_IO_SUBMIT;
2444         dioread_nolock = ext4_should_dioread_nolock(inode);
2445         if (dioread_nolock)
2446                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2447         if (map->m_flags & (1 << BH_Delay))
2448                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2449
2450         err = ext4_map_blocks(handle, inode, map, get_blocks_flags);
2451         if (err < 0)
2452                 return err;
2453         if (dioread_nolock && (map->m_flags & EXT4_MAP_UNWRITTEN)) {
2454                 if (!mpd->io_submit.io_end->handle &&
2455                     ext4_handle_valid(handle)) {
2456                         mpd->io_submit.io_end->handle = handle->h_rsv_handle;
2457                         handle->h_rsv_handle = NULL;
2458                 }
2459                 ext4_set_io_unwritten_flag(inode, mpd->io_submit.io_end);
2460         }
2461
2462         BUG_ON(map->m_len == 0);
2463         if (map->m_flags & EXT4_MAP_NEW) {
2464                 clean_bdev_aliases(inode->i_sb->s_bdev, map->m_pblk,
2465                                    map->m_len);
2466         }
2467         return 0;
2468 }
2469
2470 /*
2471  * mpage_map_and_submit_extent - map extent starting at mpd->lblk of length
2472  *                               mpd->len and submit pages underlying it for IO
2473  *
2474  * @handle - handle for journal operations
2475  * @mpd - extent to map
2476  * @give_up_on_write - we set this to true iff there is a fatal error and there
2477  *                     is no hope of writing the data. The caller should discard
2478  *                     dirty pages to avoid infinite loops.
2479  *
2480  * The function maps extent starting at mpd->lblk of length mpd->len. If it is
2481  * delayed, blocks are allocated, if it is unwritten, we may need to convert
2482  * them to initialized or split the described range from larger unwritten
2483  * extent. Note that we need not map all the described range since allocation
2484  * can return less blocks or the range is covered by more unwritten extents. We
2485  * cannot map more because we are limited by reserved transaction credits. On
2486  * the other hand we always make sure that the last touched page is fully
2487  * mapped so that it can be written out (and thus forward progress is
2488  * guaranteed). After mapping we submit all mapped pages for IO.
2489  */
2490 static int mpage_map_and_submit_extent(handle_t *handle,
2491                                        struct mpage_da_data *mpd,
2492                                        bool *give_up_on_write)
2493 {
2494         struct inode *inode = mpd->inode;
2495         struct ext4_map_blocks *map = &mpd->map;
2496         int err;
2497         loff_t disksize;
2498         int progress = 0;
2499
2500         mpd->io_submit.io_end->offset =
2501                                 ((loff_t)map->m_lblk) << inode->i_blkbits;
2502         do {
2503                 err = mpage_map_one_extent(handle, mpd);
2504                 if (err < 0) {
2505                         struct super_block *sb = inode->i_sb;
2506
2507                         if (ext4_forced_shutdown(EXT4_SB(sb)) ||
2508                             EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
2509                                 goto invalidate_dirty_pages;
2510                         /*
2511                          * Let the uper layers retry transient errors.
2512                          * In the case of ENOSPC, if ext4_count_free_blocks()
2513                          * is non-zero, a commit should free up blocks.
2514                          */
2515                         if ((err == -ENOMEM) ||
2516                             (err == -ENOSPC && ext4_count_free_clusters(sb))) {
2517                                 if (progress)
2518                                         goto update_disksize;
2519                                 return err;
2520                         }
2521                         ext4_msg(sb, KERN_CRIT,
2522                                  "Delayed block allocation failed for "
2523                                  "inode %lu at logical offset %llu with"
2524                                  " max blocks %u with error %d",
2525                                  inode->i_ino,
2526                                  (unsigned long long)map->m_lblk,
2527                                  (unsigned)map->m_len, -err);
2528                         ext4_msg(sb, KERN_CRIT,
2529                                  "This should not happen!! Data will "
2530                                  "be lost\n");
2531                         if (err == -ENOSPC)
2532                                 ext4_print_free_blocks(inode);
2533                 invalidate_dirty_pages:
2534                         *give_up_on_write = true;
2535                         return err;
2536                 }
2537                 progress = 1;
2538                 /*
2539                  * Update buffer state, submit mapped pages, and get us new
2540                  * extent to map
2541                  */
2542                 err = mpage_map_and_submit_buffers(mpd);
2543                 if (err < 0)
2544                         goto update_disksize;
2545         } while (map->m_len);
2546
2547 update_disksize:
2548         /*
2549          * Update on-disk size after IO is submitted.  Races with
2550          * truncate are avoided by checking i_size under i_data_sem.
2551          */
2552         disksize = ((loff_t)mpd->first_page) << PAGE_SHIFT;
2553         if (disksize > EXT4_I(inode)->i_disksize) {
2554                 int err2;
2555                 loff_t i_size;
2556
2557                 down_write(&EXT4_I(inode)->i_data_sem);
2558                 i_size = i_size_read(inode);
2559                 if (disksize > i_size)
2560                         disksize = i_size;
2561                 if (disksize > EXT4_I(inode)->i_disksize)
2562                         EXT4_I(inode)->i_disksize = disksize;
2563                 up_write(&EXT4_I(inode)->i_data_sem);
2564                 err2 = ext4_mark_inode_dirty(handle, inode);
2565                 if (err2)
2566                         ext4_error(inode->i_sb,
2567                                    "Failed to mark inode %lu dirty",
2568                                    inode->i_ino);
2569                 if (!err)
2570                         err = err2;
2571         }
2572         return err;
2573 }
2574
2575 /*
2576  * Calculate the total number of credits to reserve for one writepages
2577  * iteration. This is called from ext4_writepages(). We map an extent of
2578  * up to MAX_WRITEPAGES_EXTENT_LEN blocks and then we go on and finish mapping
2579  * the last partial page. So in total we can map MAX_WRITEPAGES_EXTENT_LEN +
2580  * bpp - 1 blocks in bpp different extents.
2581  */
2582 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2583 {
2584         int bpp = ext4_journal_blocks_per_page(inode);
2585
2586         return ext4_meta_trans_blocks(inode,
2587                                 MAX_WRITEPAGES_EXTENT_LEN + bpp - 1, bpp);
2588 }
2589
2590 /*
2591  * mpage_prepare_extent_to_map - find & lock contiguous range of dirty pages
2592  *                               and underlying extent to map
2593  *
2594  * @mpd - where to look for pages
2595  *
2596  * Walk dirty pages in the mapping. If they are fully mapped, submit them for
2597  * IO immediately. When we find a page which isn't mapped we start accumulating
2598  * extent of buffers underlying these pages that needs mapping (formed by
2599  * either delayed or unwritten buffers). We also lock the pages containing
2600  * these buffers. The extent found is returned in @mpd structure (starting at
2601  * mpd->lblk with length mpd->len blocks).
2602  *
2603  * Note that this function can attach bios to one io_end structure which are
2604  * neither logically nor physically contiguous. Although it may seem as an
2605  * unnecessary complication, it is actually inevitable in blocksize < pagesize
2606  * case as we need to track IO to all buffers underlying a page in one io_end.
2607  */
2608 static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
2609 {
2610         struct address_space *mapping = mpd->inode->i_mapping;
2611         struct pagevec pvec;
2612         unsigned int nr_pages;
2613         long left = mpd->wbc->nr_to_write;
2614         pgoff_t index = mpd->first_page;
2615         pgoff_t end = mpd->last_page;
2616         int tag;
2617         int i, err = 0;
2618         int blkbits = mpd->inode->i_blkbits;
2619         ext4_lblk_t lblk;
2620         struct buffer_head *head;
2621
2622         if (mpd->wbc->sync_mode == WB_SYNC_ALL || mpd->wbc->tagged_writepages)
2623                 tag = PAGECACHE_TAG_TOWRITE;
2624         else
2625                 tag = PAGECACHE_TAG_DIRTY;
2626
2627         pagevec_init(&pvec);
2628         mpd->map.m_len = 0;
2629         mpd->next_page = index;
2630         while (index <= end) {
2631                 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2632                                 tag);
2633                 if (nr_pages == 0)
2634                         goto out;
2635
2636                 for (i = 0; i < nr_pages; i++) {
2637                         struct page *page = pvec.pages[i];
2638
2639                         /*
2640                          * Accumulated enough dirty pages? This doesn't apply
2641                          * to WB_SYNC_ALL mode. For integrity sync we have to
2642                          * keep going because someone may be concurrently
2643                          * dirtying pages, and we might have synced a lot of
2644                          * newly appeared dirty pages, but have not synced all
2645                          * of the old dirty pages.
2646                          */
2647                         if (mpd->wbc->sync_mode == WB_SYNC_NONE && left <= 0)
2648                                 goto out;
2649
2650                         /* If we can't merge this page, we are done. */
2651                         if (mpd->map.m_len > 0 && mpd->next_page != page->index)
2652                                 goto out;
2653
2654                         lock_page(page);
2655                         /*
2656                          * If the page is no longer dirty, or its mapping no
2657                          * longer corresponds to inode we are writing (which
2658                          * means it has been truncated or invalidated), or the
2659                          * page is already under writeback and we are not doing
2660                          * a data integrity writeback, skip the page
2661                          */
2662                         if (!PageDirty(page) ||
2663                             (PageWriteback(page) &&
2664                              (mpd->wbc->sync_mode == WB_SYNC_NONE)) ||
2665                             unlikely(page->mapping != mapping)) {
2666                                 unlock_page(page);
2667                                 continue;
2668                         }
2669
2670                         wait_on_page_writeback(page);
2671                         BUG_ON(PageWriteback(page));
2672
2673                         if (mpd->map.m_len == 0)
2674                                 mpd->first_page = page->index;
2675                         mpd->next_page = page->index + 1;
2676                         /* Add all dirty buffers to mpd */
2677                         lblk = ((ext4_lblk_t)page->index) <<
2678                                 (PAGE_SHIFT - blkbits);
2679                         head = page_buffers(page);
2680                         err = mpage_process_page_bufs(mpd, head, head, lblk);
2681                         if (err <= 0)
2682                                 goto out;
2683                         err = 0;
2684                         left--;
2685                 }
2686                 pagevec_release(&pvec);
2687                 cond_resched();
2688         }
2689         return 0;
2690 out:
2691         pagevec_release(&pvec);
2692         return err;
2693 }
2694
2695 static int ext4_writepages(struct address_space *mapping,
2696                            struct writeback_control *wbc)
2697 {
2698         pgoff_t writeback_index = 0;
2699         long nr_to_write = wbc->nr_to_write;
2700         int range_whole = 0;
2701         int cycled = 1;
2702         handle_t *handle = NULL;
2703         struct mpage_da_data mpd;
2704         struct inode *inode = mapping->host;
2705         int needed_blocks, rsv_blocks = 0, ret = 0;
2706         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2707         bool done;
2708         struct blk_plug plug;
2709         bool give_up_on_write = false;
2710
2711         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
2712                 return -EIO;
2713
2714         percpu_down_read(&sbi->s_journal_flag_rwsem);
2715         trace_ext4_writepages(inode, wbc);
2716
2717         /*
2718          * No pages to write? This is mainly a kludge to avoid starting
2719          * a transaction for special inodes like journal inode on last iput()
2720          * because that could violate lock ordering on umount
2721          */
2722         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2723                 goto out_writepages;
2724
2725         if (ext4_should_journal_data(inode)) {
2726                 ret = generic_writepages(mapping, wbc);
2727                 goto out_writepages;
2728         }
2729
2730         /*
2731          * If the filesystem has aborted, it is read-only, so return
2732          * right away instead of dumping stack traces later on that
2733          * will obscure the real source of the problem.  We test
2734          * EXT4_MF_FS_ABORTED instead of sb->s_flag's SB_RDONLY because
2735          * the latter could be true if the filesystem is mounted
2736          * read-only, and in that case, ext4_writepages should
2737          * *never* be called, so if that ever happens, we would want
2738          * the stack trace.
2739          */
2740         if (unlikely(ext4_forced_shutdown(EXT4_SB(mapping->host->i_sb)) ||
2741                      sbi->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2742                 ret = -EROFS;
2743                 goto out_writepages;
2744         }
2745
2746         if (ext4_should_dioread_nolock(inode)) {
2747                 /*
2748                  * We may need to convert up to one extent per block in
2749                  * the page and we may dirty the inode.
2750                  */
2751                 rsv_blocks = 1 + (PAGE_SIZE >> inode->i_blkbits);
2752         }
2753
2754         /*
2755          * If we have inline data and arrive here, it means that
2756          * we will soon create the block for the 1st page, so
2757          * we'd better clear the inline data here.
2758          */
2759         if (ext4_has_inline_data(inode)) {
2760                 /* Just inode will be modified... */
2761                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
2762                 if (IS_ERR(handle)) {
2763                         ret = PTR_ERR(handle);
2764                         goto out_writepages;
2765                 }
2766                 BUG_ON(ext4_test_inode_state(inode,
2767                                 EXT4_STATE_MAY_INLINE_DATA));
2768                 ext4_destroy_inline_data(handle, inode);
2769                 ext4_journal_stop(handle);
2770         }
2771
2772         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2773                 range_whole = 1;
2774
2775         if (wbc->range_cyclic) {
2776                 writeback_index = mapping->writeback_index;
2777                 if (writeback_index)
2778                         cycled = 0;
2779                 mpd.first_page = writeback_index;
2780                 mpd.last_page = -1;
2781         } else {
2782                 mpd.first_page = wbc->range_start >> PAGE_SHIFT;
2783                 mpd.last_page = wbc->range_end >> PAGE_SHIFT;
2784         }
2785
2786         mpd.inode = inode;
2787         mpd.wbc = wbc;
2788         ext4_io_submit_init(&mpd.io_submit, wbc);
2789 retry:
2790         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2791                 tag_pages_for_writeback(mapping, mpd.first_page, mpd.last_page);
2792         done = false;
2793         blk_start_plug(&plug);
2794
2795         /*
2796          * First writeback pages that don't need mapping - we can avoid
2797          * starting a transaction unnecessarily and also avoid being blocked
2798          * in the block layer on device congestion while having transaction
2799          * started.
2800          */
2801         mpd.do_map = 0;
2802         mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2803         if (!mpd.io_submit.io_end) {
2804                 ret = -ENOMEM;
2805                 goto unplug;
2806         }
2807         ret = mpage_prepare_extent_to_map(&mpd);
2808         /* Submit prepared bio */
2809         ext4_io_submit(&mpd.io_submit);
2810         ext4_put_io_end_defer(mpd.io_submit.io_end);
2811         mpd.io_submit.io_end = NULL;
2812         /* Unlock pages we didn't use */
2813         mpage_release_unused_pages(&mpd, false);
2814         if (ret < 0)
2815                 goto unplug;
2816
2817         while (!done && mpd.first_page <= mpd.last_page) {
2818                 /* For each extent of pages we use new io_end */
2819                 mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2820                 if (!mpd.io_submit.io_end) {
2821                         ret = -ENOMEM;
2822                         break;
2823                 }
2824
2825                 /*
2826                  * We have two constraints: We find one extent to map and we
2827                  * must always write out whole page (makes a difference when
2828                  * blocksize < pagesize) so that we don't block on IO when we
2829                  * try to write out the rest of the page. Journalled mode is
2830                  * not supported by delalloc.
2831                  */
2832                 BUG_ON(ext4_should_journal_data(inode));
2833                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2834
2835                 /* start a new transaction */
2836                 handle = ext4_journal_start_with_reserve(inode,
2837                                 EXT4_HT_WRITE_PAGE, needed_blocks, rsv_blocks);
2838                 if (IS_ERR(handle)) {
2839                         ret = PTR_ERR(handle);
2840                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2841                                "%ld pages, ino %lu; err %d", __func__,
2842                                 wbc->nr_to_write, inode->i_ino, ret);
2843                         /* Release allocated io_end */
2844                         ext4_put_io_end(mpd.io_submit.io_end);
2845                         mpd.io_submit.io_end = NULL;
2846                         break;
2847                 }
2848                 mpd.do_map = 1;
2849
2850                 trace_ext4_da_write_pages(inode, mpd.first_page, mpd.wbc);
2851                 ret = mpage_prepare_extent_to_map(&mpd);
2852                 if (!ret) {
2853                         if (mpd.map.m_len)
2854                                 ret = mpage_map_and_submit_extent(handle, &mpd,
2855                                         &give_up_on_write);
2856                         else {
2857                                 /*
2858                                  * We scanned the whole range (or exhausted
2859                                  * nr_to_write), submitted what was mapped and
2860                                  * didn't find anything needing mapping. We are
2861                                  * done.
2862                                  */
2863                                 done = true;
2864                         }
2865                 }
2866                 /*
2867                  * Caution: If the handle is synchronous,
2868                  * ext4_journal_stop() can wait for transaction commit
2869                  * to finish which may depend on writeback of pages to
2870                  * complete or on page lock to be released.  In that
2871                  * case, we have to wait until after after we have
2872                  * submitted all the IO, released page locks we hold,
2873                  * and dropped io_end reference (for extent conversion
2874                  * to be able to complete) before stopping the handle.
2875                  */
2876                 if (!ext4_handle_valid(handle) || handle->h_sync == 0) {
2877                         ext4_journal_stop(handle);
2878                         handle = NULL;
2879                         mpd.do_map = 0;
2880                 }
2881                 /* Submit prepared bio */
2882                 ext4_io_submit(&mpd.io_submit);
2883                 /* Unlock pages we didn't use */
2884                 mpage_release_unused_pages(&mpd, give_up_on_write);
2885                 /*
2886                  * Drop our io_end reference we got from init. We have
2887                  * to be careful and use deferred io_end finishing if
2888                  * we are still holding the transaction as we can
2889                  * release the last reference to io_end which may end
2890                  * up doing unwritten extent conversion.
2891                  */
2892                 if (handle) {
2893                         ext4_put_io_end_defer(mpd.io_submit.io_end);
2894                         ext4_journal_stop(handle);
2895                 } else
2896                         ext4_put_io_end(mpd.io_submit.io_end);
2897                 mpd.io_submit.io_end = NULL;
2898
2899                 if (ret == -ENOSPC && sbi->s_journal) {
2900                         /*
2901                          * Commit the transaction which would
2902                          * free blocks released in the transaction
2903                          * and try again
2904                          */
2905                         jbd2_journal_force_commit_nested(sbi->s_journal);
2906                         ret = 0;
2907                         continue;
2908                 }
2909                 /* Fatal error - ENOMEM, EIO... */
2910                 if (ret)
2911                         break;
2912         }
2913 unplug:
2914         blk_finish_plug(&plug);
2915         if (!ret && !cycled && wbc->nr_to_write > 0) {
2916                 cycled = 1;
2917                 mpd.last_page = writeback_index - 1;
2918                 mpd.first_page = 0;
2919                 goto retry;
2920         }
2921
2922         /* Update index */
2923         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2924                 /*
2925                  * Set the writeback_index so that range_cyclic
2926                  * mode will write it back later
2927                  */
2928                 mapping->writeback_index = mpd.first_page;
2929
2930 out_writepages:
2931         trace_ext4_writepages_result(inode, wbc, ret,
2932                                      nr_to_write - wbc->nr_to_write);
2933         percpu_up_read(&sbi->s_journal_flag_rwsem);
2934         return ret;
2935 }
2936
2937 static int ext4_dax_writepages(struct address_space *mapping,
2938                                struct writeback_control *wbc)
2939 {
2940         int ret;
2941         long nr_to_write = wbc->nr_to_write;
2942         struct inode *inode = mapping->host;
2943         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2944
2945         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
2946                 return -EIO;
2947
2948         percpu_down_read(&sbi->s_journal_flag_rwsem);
2949         trace_ext4_writepages(inode, wbc);
2950
2951         ret = dax_writeback_mapping_range(mapping, inode->i_sb->s_bdev, wbc);
2952         trace_ext4_writepages_result(inode, wbc, ret,
2953                                      nr_to_write - wbc->nr_to_write);
2954         percpu_up_read(&sbi->s_journal_flag_rwsem);
2955         return ret;
2956 }
2957
2958 static int ext4_nonda_switch(struct super_block *sb)
2959 {
2960         s64 free_clusters, dirty_clusters;
2961         struct ext4_sb_info *sbi = EXT4_SB(sb);
2962
2963         /*
2964          * switch to non delalloc mode if we are running low
2965          * on free block. The free block accounting via percpu
2966          * counters can get slightly wrong with percpu_counter_batch getting
2967          * accumulated on each CPU without updating global counters
2968          * Delalloc need an accurate free block accounting. So switch
2969          * to non delalloc when we are near to error range.
2970          */
2971         free_clusters =
2972                 percpu_counter_read_positive(&sbi->s_freeclusters_counter);
2973         dirty_clusters =
2974                 percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
2975         /*
2976          * Start pushing delalloc when 1/2 of free blocks are dirty.
2977          */
2978         if (dirty_clusters && (free_clusters < 2 * dirty_clusters))
2979                 try_to_writeback_inodes_sb(sb, WB_REASON_FS_FREE_SPACE);
2980
2981         if (2 * free_clusters < 3 * dirty_clusters ||
2982             free_clusters < (dirty_clusters + EXT4_FREECLUSTERS_WATERMARK)) {
2983                 /*
2984                  * free block count is less than 150% of dirty blocks
2985                  * or free blocks is less than watermark
2986                  */
2987                 return 1;
2988         }
2989         return 0;
2990 }
2991
2992 /* We always reserve for an inode update; the superblock could be there too */
2993 static int ext4_da_write_credits(struct inode *inode, loff_t pos, unsigned len)
2994 {
2995         if (likely(ext4_has_feature_large_file(inode->i_sb)))
2996                 return 1;
2997
2998         if (pos + len <= 0x7fffffffULL)
2999                 return 1;
3000
3001         /* We might need to update the superblock to set LARGE_FILE */
3002         return 2;
3003 }
3004
3005 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3006                                loff_t pos, unsigned len, unsigned flags,
3007                                struct page **pagep, void **fsdata)
3008 {
3009         int ret, retries = 0;
3010         struct page *page;
3011         pgoff_t index;
3012         struct inode *inode = mapping->host;
3013         handle_t *handle;
3014
3015         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
3016                 return -EIO;
3017
3018         index = pos >> PAGE_SHIFT;
3019
3020         if (ext4_nonda_switch(inode->i_sb) ||
3021             S_ISLNK(inode->i_mode)) {
3022                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3023                 return ext4_write_begin(file, mapping, pos,
3024                                         len, flags, pagep, fsdata);
3025         }
3026         *fsdata = (void *)0;
3027         trace_ext4_da_write_begin(inode, pos, len, flags);
3028
3029         if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
3030                 ret = ext4_da_write_inline_data_begin(mapping, inode,
3031                                                       pos, len, flags,
3032                                                       pagep, fsdata);
3033                 if (ret < 0)
3034                         return ret;
3035                 if (ret == 1)
3036                         return 0;
3037         }
3038
3039         /*
3040          * grab_cache_page_write_begin() can take a long time if the
3041          * system is thrashing due to memory pressure, or if the page
3042          * is being written back.  So grab it first before we start
3043          * the transaction handle.  This also allows us to allocate
3044          * the page (if needed) without using GFP_NOFS.
3045          */
3046 retry_grab:
3047         page = grab_cache_page_write_begin(mapping, index, flags);
3048         if (!page)
3049                 return -ENOMEM;
3050         unlock_page(page);
3051
3052         /*
3053          * With delayed allocation, we don't log the i_disksize update
3054          * if there is delayed block allocation. But we still need
3055          * to journalling the i_disksize update if writes to the end
3056          * of file which has an already mapped buffer.
3057          */
3058 retry_journal:
3059         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
3060                                 ext4_da_write_credits(inode, pos, len));
3061         if (IS_ERR(handle)) {
3062                 put_page(page);
3063                 return PTR_ERR(handle);
3064         }
3065
3066         lock_page(page);
3067         if (page->mapping != mapping) {
3068                 /* The page got truncated from under us */
3069                 unlock_page(page);
3070                 put_page(page);
3071                 ext4_journal_stop(handle);
3072                 goto retry_grab;
3073         }
3074         /* In case writeback began while the page was unlocked */
3075         wait_for_stable_page(page);
3076
3077 #ifdef CONFIG_EXT4_FS_ENCRYPTION
3078         ret = ext4_block_write_begin(page, pos, len,
3079                                      ext4_da_get_block_prep);
3080 #else
3081         ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
3082 #endif
3083         if (ret < 0) {
3084                 unlock_page(page);
3085                 ext4_journal_stop(handle);
3086                 /*
3087                  * block_write_begin may have instantiated a few blocks
3088                  * outside i_size.  Trim these off again. Don't need
3089                  * i_size_read because we hold i_mutex.
3090                  */
3091                 if (pos + len > inode->i_size)
3092                         ext4_truncate_failed_write(inode);
3093
3094                 if (ret == -ENOSPC &&
3095                     ext4_should_retry_alloc(inode->i_sb, &retries))
3096                         goto retry_journal;
3097
3098                 put_page(page);
3099                 return ret;
3100         }
3101
3102         *pagep = page;
3103         return ret;
3104 }
3105
3106 /*
3107  * Check if we should update i_disksize
3108  * when write to the end of file but not require block allocation
3109  */
3110 static int ext4_da_should_update_i_disksize(struct page *page,
3111                                             unsigned long offset)
3112 {
3113         struct buffer_head *bh;
3114         struct inode *inode = page->mapping->host;
3115         unsigned int idx;
3116         int i;
3117
3118         bh = page_buffers(page);
3119         idx = offset >> inode->i_blkbits;
3120
3121         for (i = 0; i < idx; i++)
3122                 bh = bh->b_this_page;
3123
3124         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3125                 return 0;
3126         return 1;
3127 }
3128
3129 static int ext4_da_write_end(struct file *file,
3130                              struct address_space *mapping,
3131                              loff_t pos, unsigned len, unsigned copied,
3132                              struct page *page, void *fsdata)
3133 {
3134         struct inode *inode = mapping->host;
3135         int ret = 0, ret2;
3136         handle_t *handle = ext4_journal_current_handle();
3137         loff_t new_i_size;
3138         unsigned long start, end;
3139         int write_mode = (int)(unsigned long)fsdata;
3140
3141         if (write_mode == FALL_BACK_TO_NONDELALLOC)
3142                 return ext4_write_end(file, mapping, pos,
3143                                       len, copied, page, fsdata);
3144
3145         trace_ext4_da_write_end(inode, pos, len, copied);
3146         start = pos & (PAGE_SIZE - 1);
3147         end = start + copied - 1;
3148
3149         /*
3150          * generic_write_end() will run mark_inode_dirty() if i_size
3151          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3152          * into that.
3153          */
3154         new_i_size = pos + copied;
3155         if (copied && new_i_size > EXT4_I(inode)->i_disksize) {
3156                 if (ext4_has_inline_data(inode) ||
3157                     ext4_da_should_update_i_disksize(page, end)) {
3158                         ext4_update_i_disksize(inode, new_i_size);
3159                         /* We need to mark inode dirty even if
3160                          * new_i_size is less that inode->i_size
3161                          * bu greater than i_disksize.(hint delalloc)
3162                          */
3163                         ext4_mark_inode_dirty(handle, inode);
3164                 }
3165         }
3166
3167         if (write_mode != CONVERT_INLINE_DATA &&
3168             ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA) &&
3169             ext4_has_inline_data(inode))
3170                 ret2 = ext4_da_write_inline_data_end(inode, pos, len, copied,
3171                                                      page);
3172         else
3173                 ret2 = generic_write_end(file, mapping, pos, len, copied,
3174                                                         page, fsdata);
3175
3176         copied = ret2;
3177         if (ret2 < 0)
3178                 ret = ret2;
3179         ret2 = ext4_journal_stop(handle);
3180         if (!ret)
3181                 ret = ret2;
3182
3183         return ret ? ret : copied;
3184 }
3185
3186 static void ext4_da_invalidatepage(struct page *page, unsigned int offset,
3187                                    unsigned int length)
3188 {
3189         /*
3190          * Drop reserved blocks
3191          */
3192         BUG_ON(!PageLocked(page));
3193         if (!page_has_buffers(page))
3194                 goto out;
3195
3196         ext4_da_page_release_reservation(page, offset, length);
3197
3198 out:
3199         ext4_invalidatepage(page, offset, length);
3200
3201         return;
3202 }
3203
3204 /*
3205  * Force all delayed allocation blocks to be allocated for a given inode.
3206  */
3207 int ext4_alloc_da_blocks(struct inode *inode)
3208 {
3209         trace_ext4_alloc_da_blocks(inode);
3210
3211         if (!EXT4_I(inode)->i_reserved_data_blocks)
3212                 return 0;
3213
3214         /*
3215          * We do something simple for now.  The filemap_flush() will
3216          * also start triggering a write of the data blocks, which is
3217          * not strictly speaking necessary (and for users of
3218          * laptop_mode, not even desirable).  However, to do otherwise
3219          * would require replicating code paths in:
3220          *
3221          * ext4_writepages() ->
3222          *    write_cache_pages() ---> (via passed in callback function)
3223          *        __mpage_da_writepage() -->
3224          *           mpage_add_bh_to_extent()
3225          *           mpage_da_map_blocks()
3226          *
3227          * The problem is that write_cache_pages(), located in
3228          * mm/page-writeback.c, marks pages clean in preparation for
3229          * doing I/O, which is not desirable if we're not planning on
3230          * doing I/O at all.
3231          *
3232          * We could call write_cache_pages(), and then redirty all of
3233          * the pages by calling redirty_page_for_writepage() but that
3234          * would be ugly in the extreme.  So instead we would need to
3235          * replicate parts of the code in the above functions,
3236          * simplifying them because we wouldn't actually intend to
3237          * write out the pages, but rather only collect contiguous
3238          * logical block extents, call the multi-block allocator, and
3239          * then update the buffer heads with the block allocations.
3240          *
3241          * For now, though, we'll cheat by calling filemap_flush(),
3242          * which will map the blocks, and start the I/O, but not
3243          * actually wait for the I/O to complete.
3244          */
3245         return filemap_flush(inode->i_mapping);
3246 }
3247
3248 /*
3249  * bmap() is special.  It gets used by applications such as lilo and by
3250  * the swapper to find the on-disk block of a specific piece of data.
3251  *
3252  * Naturally, this is dangerous if the block concerned is still in the
3253  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3254  * filesystem and enables swap, then they may get a nasty shock when the
3255  * data getting swapped to that swapfile suddenly gets overwritten by
3256  * the original zero's written out previously to the journal and
3257  * awaiting writeback in the kernel's buffer cache.
3258  *
3259  * So, if we see any bmap calls here on a modified, data-journaled file,
3260  * take extra steps to flush any blocks which might be in the cache.
3261  */
3262 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3263 {
3264         struct inode *inode = mapping->host;
3265         journal_t *journal;
3266         int err;
3267
3268         /*
3269          * We can get here for an inline file via the FIBMAP ioctl
3270          */
3271         if (ext4_has_inline_data(inode))
3272                 return 0;
3273
3274         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3275                         test_opt(inode->i_sb, DELALLOC)) {
3276                 /*
3277                  * With delalloc we want to sync the file
3278                  * so that we can make sure we allocate
3279                  * blocks for file
3280                  */
3281                 filemap_write_and_wait(mapping);
3282         }
3283
3284         if (EXT4_JOURNAL(inode) &&
3285             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3286                 /*
3287                  * This is a REALLY heavyweight approach, but the use of
3288                  * bmap on dirty files is expected to be extremely rare:
3289                  * only if we run lilo or swapon on a freshly made file
3290                  * do we expect this to happen.
3291                  *
3292                  * (bmap requires CAP_SYS_RAWIO so this does not
3293                  * represent an unprivileged user DOS attack --- we'd be
3294                  * in trouble if mortal users could trigger this path at