Merge branch 'work.misc' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs
[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  = get_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 "
406                                  "(length %d)", (unsigned long) map->m_lblk,
407                                  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
1393         trace_ext4_write_end(inode, pos, len, copied);
1394         if (ext4_has_inline_data(inode)) {
1395                 ret = ext4_write_inline_data_end(inode, pos, len,
1396                                                  copied, page);
1397                 if (ret < 0) {
1398                         unlock_page(page);
1399                         put_page(page);
1400                         goto errout;
1401                 }
1402                 copied = ret;
1403         } else
1404                 copied = block_write_end(file, mapping, pos,
1405                                          len, copied, page, fsdata);
1406         /*
1407          * it's important to update i_size while still holding page lock:
1408          * page writeout could otherwise come in and zero beyond i_size.
1409          */
1410         i_size_changed = ext4_update_inode_size(inode, pos + copied);
1411         unlock_page(page);
1412         put_page(page);
1413
1414         if (old_size < pos)
1415                 pagecache_isize_extended(inode, old_size, pos);
1416         /*
1417          * Don't mark the inode dirty under page lock. First, it unnecessarily
1418          * makes the holding time of page lock longer. Second, it forces lock
1419          * ordering of page lock and transaction start for journaling
1420          * filesystems.
1421          */
1422         if (i_size_changed)
1423                 ext4_mark_inode_dirty(handle, inode);
1424
1425         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1426                 /* if we have allocated more blocks and copied
1427                  * less. We will have blocks allocated outside
1428                  * inode->i_size. So truncate them
1429                  */
1430                 ext4_orphan_add(handle, inode);
1431 errout:
1432         ret2 = ext4_journal_stop(handle);
1433         if (!ret)
1434                 ret = ret2;
1435
1436         if (pos + len > inode->i_size) {
1437                 ext4_truncate_failed_write(inode);
1438                 /*
1439                  * If truncate failed early the inode might still be
1440                  * on the orphan list; we need to make sure the inode
1441                  * is removed from the orphan list in that case.
1442                  */
1443                 if (inode->i_nlink)
1444                         ext4_orphan_del(NULL, inode);
1445         }
1446
1447         return ret ? ret : copied;
1448 }
1449
1450 /*
1451  * This is a private version of page_zero_new_buffers() which doesn't
1452  * set the buffer to be dirty, since in data=journalled mode we need
1453  * to call ext4_handle_dirty_metadata() instead.
1454  */
1455 static void ext4_journalled_zero_new_buffers(handle_t *handle,
1456                                             struct page *page,
1457                                             unsigned from, unsigned to)
1458 {
1459         unsigned int block_start = 0, block_end;
1460         struct buffer_head *head, *bh;
1461
1462         bh = head = page_buffers(page);
1463         do {
1464                 block_end = block_start + bh->b_size;
1465                 if (buffer_new(bh)) {
1466                         if (block_end > from && block_start < to) {
1467                                 if (!PageUptodate(page)) {
1468                                         unsigned start, size;
1469
1470                                         start = max(from, block_start);
1471                                         size = min(to, block_end) - start;
1472
1473                                         zero_user(page, start, size);
1474                                         write_end_fn(handle, bh);
1475                                 }
1476                                 clear_buffer_new(bh);
1477                         }
1478                 }
1479                 block_start = block_end;
1480                 bh = bh->b_this_page;
1481         } while (bh != head);
1482 }
1483
1484 static int ext4_journalled_write_end(struct file *file,
1485                                      struct address_space *mapping,
1486                                      loff_t pos, unsigned len, unsigned copied,
1487                                      struct page *page, void *fsdata)
1488 {
1489         handle_t *handle = ext4_journal_current_handle();
1490         struct inode *inode = mapping->host;
1491         loff_t old_size = inode->i_size;
1492         int ret = 0, ret2;
1493         int partial = 0;
1494         unsigned from, to;
1495         int size_changed = 0;
1496
1497         trace_ext4_journalled_write_end(inode, pos, len, copied);
1498         from = pos & (PAGE_SIZE - 1);
1499         to = from + len;
1500
1501         BUG_ON(!ext4_handle_valid(handle));
1502
1503         if (ext4_has_inline_data(inode)) {
1504                 ret = ext4_write_inline_data_end(inode, pos, len,
1505                                                  copied, page);
1506                 if (ret < 0) {
1507                         unlock_page(page);
1508                         put_page(page);
1509                         goto errout;
1510                 }
1511                 copied = ret;
1512         } else if (unlikely(copied < len) && !PageUptodate(page)) {
1513                 copied = 0;
1514                 ext4_journalled_zero_new_buffers(handle, page, from, to);
1515         } else {
1516                 if (unlikely(copied < len))
1517                         ext4_journalled_zero_new_buffers(handle, page,
1518                                                          from + copied, to);
1519                 ret = ext4_walk_page_buffers(handle, page_buffers(page), from,
1520                                              from + copied, &partial,
1521                                              write_end_fn);
1522                 if (!partial)
1523                         SetPageUptodate(page);
1524         }
1525         size_changed = ext4_update_inode_size(inode, pos + copied);
1526         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1527         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1528         unlock_page(page);
1529         put_page(page);
1530
1531         if (old_size < pos)
1532                 pagecache_isize_extended(inode, old_size, pos);
1533
1534         if (size_changed) {
1535                 ret2 = ext4_mark_inode_dirty(handle, inode);
1536                 if (!ret)
1537                         ret = ret2;
1538         }
1539
1540         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1541                 /* if we have allocated more blocks and copied
1542                  * less. We will have blocks allocated outside
1543                  * inode->i_size. So truncate them
1544                  */
1545                 ext4_orphan_add(handle, inode);
1546
1547 errout:
1548         ret2 = ext4_journal_stop(handle);
1549         if (!ret)
1550                 ret = ret2;
1551         if (pos + len > inode->i_size) {
1552                 ext4_truncate_failed_write(inode);
1553                 /*
1554                  * If truncate failed early the inode might still be
1555                  * on the orphan list; we need to make sure the inode
1556                  * is removed from the orphan list in that case.
1557                  */
1558                 if (inode->i_nlink)
1559                         ext4_orphan_del(NULL, inode);
1560         }
1561
1562         return ret ? ret : copied;
1563 }
1564
1565 /*
1566  * Reserve space for a single cluster
1567  */
1568 static int ext4_da_reserve_space(struct inode *inode)
1569 {
1570         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1571         struct ext4_inode_info *ei = EXT4_I(inode);
1572         int ret;
1573
1574         /*
1575          * We will charge metadata quota at writeout time; this saves
1576          * us from metadata over-estimation, though we may go over by
1577          * a small amount in the end.  Here we just reserve for data.
1578          */
1579         ret = dquot_reserve_block(inode, EXT4_C2B(sbi, 1));
1580         if (ret)
1581                 return ret;
1582
1583         spin_lock(&ei->i_block_reservation_lock);
1584         if (ext4_claim_free_clusters(sbi, 1, 0)) {
1585                 spin_unlock(&ei->i_block_reservation_lock);
1586                 dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
1587                 return -ENOSPC;
1588         }
1589         ei->i_reserved_data_blocks++;
1590         trace_ext4_da_reserve_space(inode);
1591         spin_unlock(&ei->i_block_reservation_lock);
1592
1593         return 0;       /* success */
1594 }
1595
1596 static void ext4_da_release_space(struct inode *inode, int to_free)
1597 {
1598         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1599         struct ext4_inode_info *ei = EXT4_I(inode);
1600
1601         if (!to_free)
1602                 return;         /* Nothing to release, exit */
1603
1604         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1605
1606         trace_ext4_da_release_space(inode, to_free);
1607         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1608                 /*
1609                  * if there aren't enough reserved blocks, then the
1610                  * counter is messed up somewhere.  Since this
1611                  * function is called from invalidate page, it's
1612                  * harmless to return without any action.
1613                  */
1614                 ext4_warning(inode->i_sb, "ext4_da_release_space: "
1615                          "ino %lu, to_free %d with only %d reserved "
1616                          "data blocks", inode->i_ino, to_free,
1617                          ei->i_reserved_data_blocks);
1618                 WARN_ON(1);
1619                 to_free = ei->i_reserved_data_blocks;
1620         }
1621         ei->i_reserved_data_blocks -= to_free;
1622
1623         /* update fs dirty data blocks counter */
1624         percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1625
1626         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1627
1628         dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1629 }
1630
1631 static void ext4_da_page_release_reservation(struct page *page,
1632                                              unsigned int offset,
1633                                              unsigned int length)
1634 {
1635         int to_release = 0, contiguous_blks = 0;
1636         struct buffer_head *head, *bh;
1637         unsigned int curr_off = 0;
1638         struct inode *inode = page->mapping->host;
1639         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1640         unsigned int stop = offset + length;
1641         int num_clusters;
1642         ext4_fsblk_t lblk;
1643
1644         BUG_ON(stop > PAGE_SIZE || stop < length);
1645
1646         head = page_buffers(page);
1647         bh = head;
1648         do {
1649                 unsigned int next_off = curr_off + bh->b_size;
1650
1651                 if (next_off > stop)
1652                         break;
1653
1654                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1655                         to_release++;
1656                         contiguous_blks++;
1657                         clear_buffer_delay(bh);
1658                 } else if (contiguous_blks) {
1659                         lblk = page->index <<
1660                                (PAGE_SHIFT - inode->i_blkbits);
1661                         lblk += (curr_off >> inode->i_blkbits) -
1662                                 contiguous_blks;
1663                         ext4_es_remove_extent(inode, lblk, contiguous_blks);
1664                         contiguous_blks = 0;
1665                 }
1666                 curr_off = next_off;
1667         } while ((bh = bh->b_this_page) != head);
1668
1669         if (contiguous_blks) {
1670                 lblk = page->index << (PAGE_SHIFT - inode->i_blkbits);
1671                 lblk += (curr_off >> inode->i_blkbits) - contiguous_blks;
1672                 ext4_es_remove_extent(inode, lblk, contiguous_blks);
1673         }
1674
1675         /* If we have released all the blocks belonging to a cluster, then we
1676          * need to release the reserved space for that cluster. */
1677         num_clusters = EXT4_NUM_B2C(sbi, to_release);
1678         while (num_clusters > 0) {
1679                 lblk = (page->index << (PAGE_SHIFT - inode->i_blkbits)) +
1680                         ((num_clusters - 1) << sbi->s_cluster_bits);
1681                 if (sbi->s_cluster_ratio == 1 ||
1682                     !ext4_find_delalloc_cluster(inode, lblk))
1683                         ext4_da_release_space(inode, 1);
1684
1685                 num_clusters--;
1686         }
1687 }
1688
1689 /*
1690  * Delayed allocation stuff
1691  */
1692
1693 struct mpage_da_data {
1694         struct inode *inode;
1695         struct writeback_control *wbc;
1696
1697         pgoff_t first_page;     /* The first page to write */
1698         pgoff_t next_page;      /* Current page to examine */
1699         pgoff_t last_page;      /* Last page to examine */
1700         /*
1701          * Extent to map - this can be after first_page because that can be
1702          * fully mapped. We somewhat abuse m_flags to store whether the extent
1703          * is delalloc or unwritten.
1704          */
1705         struct ext4_map_blocks map;
1706         struct ext4_io_submit io_submit;        /* IO submission data */
1707         unsigned int do_map:1;
1708 };
1709
1710 static void mpage_release_unused_pages(struct mpage_da_data *mpd,
1711                                        bool invalidate)
1712 {
1713         int nr_pages, i;
1714         pgoff_t index, end;
1715         struct pagevec pvec;
1716         struct inode *inode = mpd->inode;
1717         struct address_space *mapping = inode->i_mapping;
1718
1719         /* This is necessary when next_page == 0. */
1720         if (mpd->first_page >= mpd->next_page)
1721                 return;
1722
1723         index = mpd->first_page;
1724         end   = mpd->next_page - 1;
1725         if (invalidate) {
1726                 ext4_lblk_t start, last;
1727                 start = index << (PAGE_SHIFT - inode->i_blkbits);
1728                 last = end << (PAGE_SHIFT - inode->i_blkbits);
1729                 ext4_es_remove_extent(inode, start, last - start + 1);
1730         }
1731
1732         pagevec_init(&pvec);
1733         while (index <= end) {
1734                 nr_pages = pagevec_lookup_range(&pvec, mapping, &index, end);
1735                 if (nr_pages == 0)
1736                         break;
1737                 for (i = 0; i < nr_pages; i++) {
1738                         struct page *page = pvec.pages[i];
1739
1740                         BUG_ON(!PageLocked(page));
1741                         BUG_ON(PageWriteback(page));
1742                         if (invalidate) {
1743                                 if (page_mapped(page))
1744                                         clear_page_dirty_for_io(page);
1745                                 block_invalidatepage(page, 0, PAGE_SIZE);
1746                                 ClearPageUptodate(page);
1747                         }
1748                         unlock_page(page);
1749                 }
1750                 pagevec_release(&pvec);
1751         }
1752 }
1753
1754 static void ext4_print_free_blocks(struct inode *inode)
1755 {
1756         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1757         struct super_block *sb = inode->i_sb;
1758         struct ext4_inode_info *ei = EXT4_I(inode);
1759
1760         ext4_msg(sb, KERN_CRIT, "Total free blocks count %lld",
1761                EXT4_C2B(EXT4_SB(inode->i_sb),
1762                         ext4_count_free_clusters(sb)));
1763         ext4_msg(sb, KERN_CRIT, "Free/Dirty block details");
1764         ext4_msg(sb, KERN_CRIT, "free_blocks=%lld",
1765                (long long) EXT4_C2B(EXT4_SB(sb),
1766                 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1767         ext4_msg(sb, KERN_CRIT, "dirty_blocks=%lld",
1768                (long long) EXT4_C2B(EXT4_SB(sb),
1769                 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1770         ext4_msg(sb, KERN_CRIT, "Block reservation details");
1771         ext4_msg(sb, KERN_CRIT, "i_reserved_data_blocks=%u",
1772                  ei->i_reserved_data_blocks);
1773         return;
1774 }
1775
1776 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1777 {
1778         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1779 }
1780
1781 /*
1782  * This function is grabs code from the very beginning of
1783  * ext4_map_blocks, but assumes that the caller is from delayed write
1784  * time. This function looks up the requested blocks and sets the
1785  * buffer delay bit under the protection of i_data_sem.
1786  */
1787 static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1788                               struct ext4_map_blocks *map,
1789                               struct buffer_head *bh)
1790 {
1791         struct extent_status es;
1792         int retval;
1793         sector_t invalid_block = ~((sector_t) 0xffff);
1794 #ifdef ES_AGGRESSIVE_TEST
1795         struct ext4_map_blocks orig_map;
1796
1797         memcpy(&orig_map, map, sizeof(*map));
1798 #endif
1799
1800         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1801                 invalid_block = ~0;
1802
1803         map->m_flags = 0;
1804         ext_debug("ext4_da_map_blocks(): inode %lu, max_blocks %u,"
1805                   "logical block %lu\n", inode->i_ino, map->m_len,
1806                   (unsigned long) map->m_lblk);
1807
1808         /* Lookup extent status tree firstly */
1809         if (ext4_es_lookup_extent(inode, iblock, &es)) {
1810                 if (ext4_es_is_hole(&es)) {
1811                         retval = 0;
1812                         down_read(&EXT4_I(inode)->i_data_sem);
1813                         goto add_delayed;
1814                 }
1815
1816                 /*
1817                  * Delayed extent could be allocated by fallocate.
1818                  * So we need to check it.
1819                  */
1820                 if (ext4_es_is_delayed(&es) && !ext4_es_is_unwritten(&es)) {
1821                         map_bh(bh, inode->i_sb, invalid_block);
1822                         set_buffer_new(bh);
1823                         set_buffer_delay(bh);
1824                         return 0;
1825                 }
1826
1827                 map->m_pblk = ext4_es_pblock(&es) + iblock - es.es_lblk;
1828                 retval = es.es_len - (iblock - es.es_lblk);
1829                 if (retval > map->m_len)
1830                         retval = map->m_len;
1831                 map->m_len = retval;
1832                 if (ext4_es_is_written(&es))
1833                         map->m_flags |= EXT4_MAP_MAPPED;
1834                 else if (ext4_es_is_unwritten(&es))
1835                         map->m_flags |= EXT4_MAP_UNWRITTEN;
1836                 else
1837                         BUG_ON(1);
1838
1839 #ifdef ES_AGGRESSIVE_TEST
1840                 ext4_map_blocks_es_recheck(NULL, inode, map, &orig_map, 0);
1841 #endif
1842                 return retval;
1843         }
1844
1845         /*
1846          * Try to see if we can get the block without requesting a new
1847          * file system block.
1848          */
1849         down_read(&EXT4_I(inode)->i_data_sem);
1850         if (ext4_has_inline_data(inode))
1851                 retval = 0;
1852         else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1853                 retval = ext4_ext_map_blocks(NULL, inode, map, 0);
1854         else
1855                 retval = ext4_ind_map_blocks(NULL, inode, map, 0);
1856
1857 add_delayed:
1858         if (retval == 0) {
1859                 int ret;
1860                 /*
1861                  * XXX: __block_prepare_write() unmaps passed block,
1862                  * is it OK?
1863                  */
1864                 /*
1865                  * If the block was allocated from previously allocated cluster,
1866                  * then we don't need to reserve it again. However we still need
1867                  * to reserve metadata for every block we're going to write.
1868                  */
1869                 if (EXT4_SB(inode->i_sb)->s_cluster_ratio == 1 ||
1870                     !ext4_find_delalloc_cluster(inode, map->m_lblk)) {
1871                         ret = ext4_da_reserve_space(inode);
1872                         if (ret) {
1873                                 /* not enough space to reserve */
1874                                 retval = ret;
1875                                 goto out_unlock;
1876                         }
1877                 }
1878
1879                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1880                                             ~0, EXTENT_STATUS_DELAYED);
1881                 if (ret) {
1882                         retval = ret;
1883                         goto out_unlock;
1884                 }
1885
1886                 map_bh(bh, inode->i_sb, invalid_block);
1887                 set_buffer_new(bh);
1888                 set_buffer_delay(bh);
1889         } else if (retval > 0) {
1890                 int ret;
1891                 unsigned int status;
1892
1893                 if (unlikely(retval != map->m_len)) {
1894                         ext4_warning(inode->i_sb,
1895                                      "ES len assertion failed for inode "
1896                                      "%lu: retval %d != map->m_len %d",
1897                                      inode->i_ino, retval, map->m_len);
1898                         WARN_ON(1);
1899                 }
1900
1901                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
1902                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
1903                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1904                                             map->m_pblk, status);
1905                 if (ret != 0)
1906                         retval = ret;
1907         }
1908
1909 out_unlock:
1910         up_read((&EXT4_I(inode)->i_data_sem));
1911
1912         return retval;
1913 }
1914
1915 /*
1916  * This is a special get_block_t callback which is used by
1917  * ext4_da_write_begin().  It will either return mapped block or
1918  * reserve space for a single block.
1919  *
1920  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1921  * We also have b_blocknr = -1 and b_bdev initialized properly
1922  *
1923  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1924  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1925  * initialized properly.
1926  */
1927 int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1928                            struct buffer_head *bh, int create)
1929 {
1930         struct ext4_map_blocks map;
1931         int ret = 0;
1932
1933         BUG_ON(create == 0);
1934         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1935
1936         map.m_lblk = iblock;
1937         map.m_len = 1;
1938
1939         /*
1940          * first, we need to know whether the block is allocated already
1941          * preallocated blocks are unmapped but should treated
1942          * the same as allocated blocks.
1943          */
1944         ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1945         if (ret <= 0)
1946                 return ret;
1947
1948         map_bh(bh, inode->i_sb, map.m_pblk);
1949         ext4_update_bh_state(bh, map.m_flags);
1950
1951         if (buffer_unwritten(bh)) {
1952                 /* A delayed write to unwritten bh should be marked
1953                  * new and mapped.  Mapped ensures that we don't do
1954                  * get_block multiple times when we write to the same
1955                  * offset and new ensures that we do proper zero out
1956                  * for partial write.
1957                  */
1958                 set_buffer_new(bh);
1959                 set_buffer_mapped(bh);
1960         }
1961         return 0;
1962 }
1963
1964 static int bget_one(handle_t *handle, struct buffer_head *bh)
1965 {
1966         get_bh(bh);
1967         return 0;
1968 }
1969
1970 static int bput_one(handle_t *handle, struct buffer_head *bh)
1971 {
1972         put_bh(bh);
1973         return 0;
1974 }
1975
1976 static int __ext4_journalled_writepage(struct page *page,
1977                                        unsigned int len)
1978 {
1979         struct address_space *mapping = page->mapping;
1980         struct inode *inode = mapping->host;
1981         struct buffer_head *page_bufs = NULL;
1982         handle_t *handle = NULL;
1983         int ret = 0, err = 0;
1984         int inline_data = ext4_has_inline_data(inode);
1985         struct buffer_head *inode_bh = NULL;
1986
1987         ClearPageChecked(page);
1988
1989         if (inline_data) {
1990                 BUG_ON(page->index != 0);
1991                 BUG_ON(len > ext4_get_max_inline_size(inode));
1992                 inode_bh = ext4_journalled_write_inline_data(inode, len, page);
1993                 if (inode_bh == NULL)
1994                         goto out;
1995         } else {
1996                 page_bufs = page_buffers(page);
1997                 if (!page_bufs) {
1998                         BUG();
1999                         goto out;
2000                 }
2001                 ext4_walk_page_buffers(handle, page_bufs, 0, len,
2002                                        NULL, bget_one);
2003         }
2004         /*
2005          * We need to release the page lock before we start the
2006          * journal, so grab a reference so the page won't disappear
2007          * out from under us.
2008          */
2009         get_page(page);
2010         unlock_page(page);
2011
2012         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
2013                                     ext4_writepage_trans_blocks(inode));
2014         if (IS_ERR(handle)) {
2015                 ret = PTR_ERR(handle);
2016                 put_page(page);
2017                 goto out_no_pagelock;
2018         }
2019         BUG_ON(!ext4_handle_valid(handle));
2020
2021         lock_page(page);
2022         put_page(page);
2023         if (page->mapping != mapping) {
2024                 /* The page got truncated from under us */
2025                 ext4_journal_stop(handle);
2026                 ret = 0;
2027                 goto out;
2028         }
2029
2030         if (inline_data) {
2031                 BUFFER_TRACE(inode_bh, "get write access");
2032                 ret = ext4_journal_get_write_access(handle, inode_bh);
2033
2034                 err = ext4_handle_dirty_metadata(handle, inode, inode_bh);
2035
2036         } else {
2037                 ret = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
2038                                              do_journal_get_write_access);
2039
2040                 err = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
2041                                              write_end_fn);
2042         }
2043         if (ret == 0)
2044                 ret = err;
2045         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
2046         err = ext4_journal_stop(handle);
2047         if (!ret)
2048                 ret = err;
2049
2050         if (!ext4_has_inline_data(inode))
2051                 ext4_walk_page_buffers(NULL, page_bufs, 0, len,
2052                                        NULL, bput_one);
2053         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2054 out:
2055         unlock_page(page);
2056 out_no_pagelock:
2057         brelse(inode_bh);
2058         return ret;
2059 }
2060
2061 /*
2062  * Note that we don't need to start a transaction unless we're journaling data
2063  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2064  * need to file the inode to the transaction's list in ordered mode because if
2065  * we are writing back data added by write(), the inode is already there and if
2066  * we are writing back data modified via mmap(), no one guarantees in which
2067  * transaction the data will hit the disk. In case we are journaling data, we
2068  * cannot start transaction directly because transaction start ranks above page
2069  * lock so we have to do some magic.
2070  *
2071  * This function can get called via...
2072  *   - ext4_writepages after taking page lock (have journal handle)
2073  *   - journal_submit_inode_data_buffers (no journal handle)
2074  *   - shrink_page_list via the kswapd/direct reclaim (no journal handle)
2075  *   - grab_page_cache when doing write_begin (have journal handle)
2076  *
2077  * We don't do any block allocation in this function. If we have page with
2078  * multiple blocks we need to write those buffer_heads that are mapped. This
2079  * is important for mmaped based write. So if we do with blocksize 1K
2080  * truncate(f, 1024);
2081  * a = mmap(f, 0, 4096);
2082  * a[0] = 'a';
2083  * truncate(f, 4096);
2084  * we have in the page first buffer_head mapped via page_mkwrite call back
2085  * but other buffer_heads would be unmapped but dirty (dirty done via the
2086  * do_wp_page). So writepage should write the first block. If we modify
2087  * the mmap area beyond 1024 we will again get a page_fault and the
2088  * page_mkwrite callback will do the block allocation and mark the
2089  * buffer_heads mapped.
2090  *
2091  * We redirty the page if we have any buffer_heads that is either delay or
2092  * unwritten in the page.
2093  *
2094  * We can get recursively called as show below.
2095  *
2096  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2097  *              ext4_writepage()
2098  *
2099  * But since we don't do any block allocation we should not deadlock.
2100  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2101  */
2102 static int ext4_writepage(struct page *page,
2103                           struct writeback_control *wbc)
2104 {
2105         int ret = 0;
2106         loff_t size;
2107         unsigned int len;
2108         struct buffer_head *page_bufs = NULL;
2109         struct inode *inode = page->mapping->host;
2110         struct ext4_io_submit io_submit;
2111         bool keep_towrite = false;
2112
2113         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb)))) {
2114                 ext4_invalidatepage(page, 0, PAGE_SIZE);
2115                 unlock_page(page);
2116                 return -EIO;
2117         }
2118
2119         trace_ext4_writepage(page);
2120         size = i_size_read(inode);
2121         if (page->index == size >> PAGE_SHIFT)
2122                 len = size & ~PAGE_MASK;
2123         else
2124                 len = PAGE_SIZE;
2125
2126         page_bufs = page_buffers(page);
2127         /*
2128          * We cannot do block allocation or other extent handling in this
2129          * function. If there are buffers needing that, we have to redirty
2130          * the page. But we may reach here when we do a journal commit via
2131          * journal_submit_inode_data_buffers() and in that case we must write
2132          * allocated buffers to achieve data=ordered mode guarantees.
2133          *
2134          * Also, if there is only one buffer per page (the fs block
2135          * size == the page size), if one buffer needs block
2136          * allocation or needs to modify the extent tree to clear the
2137          * unwritten flag, we know that the page can't be written at
2138          * all, so we might as well refuse the write immediately.
2139          * Unfortunately if the block size != page size, we can't as
2140          * easily detect this case using ext4_walk_page_buffers(), but
2141          * for the extremely common case, this is an optimization that
2142          * skips a useless round trip through ext4_bio_write_page().
2143          */
2144         if (ext4_walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2145                                    ext4_bh_delay_or_unwritten)) {
2146                 redirty_page_for_writepage(wbc, page);
2147                 if ((current->flags & PF_MEMALLOC) ||
2148                     (inode->i_sb->s_blocksize == PAGE_SIZE)) {
2149                         /*
2150                          * For memory cleaning there's no point in writing only
2151                          * some buffers. So just bail out. Warn if we came here
2152                          * from direct reclaim.
2153                          */
2154                         WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD))
2155                                                         == PF_MEMALLOC);
2156                         unlock_page(page);
2157                         return 0;
2158                 }
2159                 keep_towrite = true;
2160         }
2161
2162         if (PageChecked(page) && ext4_should_journal_data(inode))
2163                 /*
2164                  * It's mmapped pagecache.  Add buffers and journal it.  There
2165                  * doesn't seem much point in redirtying the page here.
2166                  */
2167                 return __ext4_journalled_writepage(page, len);
2168
2169         ext4_io_submit_init(&io_submit, wbc);
2170         io_submit.io_end = ext4_init_io_end(inode, GFP_NOFS);
2171         if (!io_submit.io_end) {
2172                 redirty_page_for_writepage(wbc, page);
2173                 unlock_page(page);
2174                 return -ENOMEM;
2175         }
2176         ret = ext4_bio_write_page(&io_submit, page, len, wbc, keep_towrite);
2177         ext4_io_submit(&io_submit);
2178         /* Drop io_end reference we got from init */
2179         ext4_put_io_end_defer(io_submit.io_end);
2180         return ret;
2181 }
2182
2183 static int mpage_submit_page(struct mpage_da_data *mpd, struct page *page)
2184 {
2185         int len;
2186         loff_t size;
2187         int err;
2188
2189         BUG_ON(page->index != mpd->first_page);
2190         clear_page_dirty_for_io(page);
2191         /*
2192          * We have to be very careful here!  Nothing protects writeback path
2193          * against i_size changes and the page can be writeably mapped into
2194          * page tables. So an application can be growing i_size and writing
2195          * data through mmap while writeback runs. clear_page_dirty_for_io()
2196          * write-protects our page in page tables and the page cannot get
2197          * written to again until we release page lock. So only after
2198          * clear_page_dirty_for_io() we are safe to sample i_size for
2199          * ext4_bio_write_page() to zero-out tail of the written page. We rely
2200          * on the barrier provided by TestClearPageDirty in
2201          * clear_page_dirty_for_io() to make sure i_size is really sampled only
2202          * after page tables are updated.
2203          */
2204         size = i_size_read(mpd->inode);
2205         if (page->index == size >> PAGE_SHIFT)
2206                 len = size & ~PAGE_MASK;
2207         else
2208                 len = PAGE_SIZE;
2209         err = ext4_bio_write_page(&mpd->io_submit, page, len, mpd->wbc, false);
2210         if (!err)
2211                 mpd->wbc->nr_to_write--;
2212         mpd->first_page++;
2213
2214         return err;
2215 }
2216
2217 #define BH_FLAGS ((1 << BH_Unwritten) | (1 << BH_Delay))
2218
2219 /*
2220  * mballoc gives us at most this number of blocks...
2221  * XXX: That seems to be only a limitation of ext4_mb_normalize_request().
2222  * The rest of mballoc seems to handle chunks up to full group size.
2223  */
2224 #define MAX_WRITEPAGES_EXTENT_LEN 2048
2225
2226 /*
2227  * mpage_add_bh_to_extent - try to add bh to extent of blocks to map
2228  *
2229  * @mpd - extent of blocks
2230  * @lblk - logical number of the block in the file
2231  * @bh - buffer head we want to add to the extent
2232  *
2233  * The function is used to collect contig. blocks in the same state. If the
2234  * buffer doesn't require mapping for writeback and we haven't started the
2235  * extent of buffers to map yet, the function returns 'true' immediately - the
2236  * caller can write the buffer right away. Otherwise the function returns true
2237  * if the block has been added to the extent, false if the block couldn't be
2238  * added.
2239  */
2240 static bool mpage_add_bh_to_extent(struct mpage_da_data *mpd, ext4_lblk_t lblk,
2241                                    struct buffer_head *bh)
2242 {
2243         struct ext4_map_blocks *map = &mpd->map;
2244
2245         /* Buffer that doesn't need mapping for writeback? */
2246         if (!buffer_dirty(bh) || !buffer_mapped(bh) ||
2247             (!buffer_delay(bh) && !buffer_unwritten(bh))) {
2248                 /* So far no extent to map => we write the buffer right away */
2249                 if (map->m_len == 0)
2250                         return true;
2251                 return false;
2252         }
2253
2254         /* First block in the extent? */
2255         if (map->m_len == 0) {
2256                 /* We cannot map unless handle is started... */
2257                 if (!mpd->do_map)
2258                         return false;
2259                 map->m_lblk = lblk;
2260                 map->m_len = 1;
2261                 map->m_flags = bh->b_state & BH_FLAGS;
2262                 return true;
2263         }
2264
2265         /* Don't go larger than mballoc is willing to allocate */
2266         if (map->m_len >= MAX_WRITEPAGES_EXTENT_LEN)
2267                 return false;
2268
2269         /* Can we merge the block to our big extent? */
2270         if (lblk == map->m_lblk + map->m_len &&
2271             (bh->b_state & BH_FLAGS) == map->m_flags) {
2272                 map->m_len++;
2273                 return true;
2274         }
2275         return false;
2276 }
2277
2278 /*
2279  * mpage_process_page_bufs - submit page buffers for IO or add them to extent
2280  *
2281  * @mpd - extent of blocks for mapping
2282  * @head - the first buffer in the page
2283  * @bh - buffer we should start processing from
2284  * @lblk - logical number of the block in the file corresponding to @bh
2285  *
2286  * Walk through page buffers from @bh upto @head (exclusive) and either submit
2287  * the page for IO if all buffers in this page were mapped and there's no
2288  * accumulated extent of buffers to map or add buffers in the page to the
2289  * extent of buffers to map. The function returns 1 if the caller can continue
2290  * by processing the next page, 0 if it should stop adding buffers to the
2291  * extent to map because we cannot extend it anymore. It can also return value
2292  * < 0 in case of error during IO submission.
2293  */
2294 static int mpage_process_page_bufs(struct mpage_da_data *mpd,
2295                                    struct buffer_head *head,
2296                                    struct buffer_head *bh,
2297                                    ext4_lblk_t lblk)
2298 {
2299         struct inode *inode = mpd->inode;
2300         int err;
2301         ext4_lblk_t blocks = (i_size_read(inode) + i_blocksize(inode) - 1)
2302                                                         >> inode->i_blkbits;
2303
2304         do {
2305                 BUG_ON(buffer_locked(bh));
2306
2307                 if (lblk >= blocks || !mpage_add_bh_to_extent(mpd, lblk, bh)) {
2308                         /* Found extent to map? */
2309                         if (mpd->map.m_len)
2310                                 return 0;
2311                         /* Buffer needs mapping and handle is not started? */
2312                         if (!mpd->do_map)
2313                                 return 0;
2314                         /* Everything mapped so far and we hit EOF */
2315                         break;
2316                 }
2317         } while (lblk++, (bh = bh->b_this_page) != head);
2318         /* So far everything mapped? Submit the page for IO. */
2319         if (mpd->map.m_len == 0) {
2320                 err = mpage_submit_page(mpd, head->b_page);
2321                 if (err < 0)
2322                         return err;
2323         }
2324         return lblk < blocks;
2325 }
2326
2327 /*
2328  * mpage_map_buffers - update buffers corresponding to changed extent and
2329  *                     submit fully mapped pages for IO
2330  *
2331  * @mpd - description of extent to map, on return next extent to map
2332  *
2333  * Scan buffers corresponding to changed extent (we expect corresponding pages
2334  * to be already locked) and update buffer state according to new extent state.
2335  * We map delalloc buffers to their physical location, clear unwritten bits,
2336  * and mark buffers as uninit when we perform writes to unwritten extents
2337  * and do extent conversion after IO is finished. If the last page is not fully
2338  * mapped, we update @map to the next extent in the last page that needs
2339  * mapping. Otherwise we submit the page for IO.
2340  */
2341 static int mpage_map_and_submit_buffers(struct mpage_da_data *mpd)
2342 {
2343         struct pagevec pvec;
2344         int nr_pages, i;
2345         struct inode *inode = mpd->inode;
2346         struct buffer_head *head, *bh;
2347         int bpp_bits = PAGE_SHIFT - inode->i_blkbits;
2348         pgoff_t start, end;
2349         ext4_lblk_t lblk;
2350         sector_t pblock;
2351         int err;
2352
2353         start = mpd->map.m_lblk >> bpp_bits;
2354         end = (mpd->map.m_lblk + mpd->map.m_len - 1) >> bpp_bits;
2355         lblk = start << bpp_bits;
2356         pblock = mpd->map.m_pblk;
2357
2358         pagevec_init(&pvec);
2359         while (start <= end) {
2360                 nr_pages = pagevec_lookup_range(&pvec, inode->i_mapping,
2361                                                 &start, end);
2362                 if (nr_pages == 0)
2363                         break;
2364                 for (i = 0; i < nr_pages; i++) {
2365                         struct page *page = pvec.pages[i];
2366
2367                         bh = head = page_buffers(page);
2368                         do {
2369                                 if (lblk < mpd->map.m_lblk)
2370                                         continue;
2371                                 if (lblk >= mpd->map.m_lblk + mpd->map.m_len) {
2372                                         /*
2373                                          * Buffer after end of mapped extent.
2374                                          * Find next buffer in the page to map.
2375                                          */
2376                                         mpd->map.m_len = 0;
2377                                         mpd->map.m_flags = 0;
2378                                         /*
2379                                          * FIXME: If dioread_nolock supports
2380                                          * blocksize < pagesize, we need to make
2381                                          * sure we add size mapped so far to
2382                                          * io_end->size as the following call
2383                                          * can submit the page for IO.
2384                                          */
2385                                         err = mpage_process_page_bufs(mpd, head,
2386                                                                       bh, lblk);
2387                                         pagevec_release(&pvec);
2388                                         if (err > 0)
2389                                                 err = 0;
2390                                         return err;
2391                                 }
2392                                 if (buffer_delay(bh)) {
2393                                         clear_buffer_delay(bh);
2394                                         bh->b_blocknr = pblock++;
2395                                 }
2396                                 clear_buffer_unwritten(bh);
2397                         } while (lblk++, (bh = bh->b_this_page) != head);
2398
2399                         /*
2400                          * FIXME: This is going to break if dioread_nolock
2401                          * supports blocksize < pagesize as we will try to
2402                          * convert potentially unmapped parts of inode.
2403                          */
2404                         mpd->io_submit.io_end->size += PAGE_SIZE;
2405                         /* Page fully mapped - let IO run! */
2406                         err = mpage_submit_page(mpd, page);
2407                         if (err < 0) {
2408                                 pagevec_release(&pvec);
2409                                 return err;
2410                         }
2411                 }
2412                 pagevec_release(&pvec);
2413         }
2414         /* Extent fully mapped and matches with page boundary. We are done. */
2415         mpd->map.m_len = 0;
2416         mpd->map.m_flags = 0;
2417         return 0;
2418 }
2419
2420 static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
2421 {
2422         struct inode *inode = mpd->inode;
2423         struct ext4_map_blocks *map = &mpd->map;
2424         int get_blocks_flags;
2425         int err, dioread_nolock;
2426
2427         trace_ext4_da_write_pages_extent(inode, map);
2428         /*
2429          * Call ext4_map_blocks() to allocate any delayed allocation blocks, or
2430          * to convert an unwritten extent to be initialized (in the case
2431          * where we have written into one or more preallocated blocks).  It is
2432          * possible that we're going to need more metadata blocks than
2433          * previously reserved. However we must not fail because we're in
2434          * writeback and there is nothing we can do about it so it might result
2435          * in data loss.  So use reserved blocks to allocate metadata if
2436          * possible.
2437          *
2438          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE if
2439          * the blocks in question are delalloc blocks.  This indicates
2440          * that the blocks and quotas has already been checked when
2441          * the data was copied into the page cache.
2442          */
2443         get_blocks_flags = EXT4_GET_BLOCKS_CREATE |
2444                            EXT4_GET_BLOCKS_METADATA_NOFAIL |
2445                            EXT4_GET_BLOCKS_IO_SUBMIT;
2446         dioread_nolock = ext4_should_dioread_nolock(inode);
2447         if (dioread_nolock)
2448                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2449         if (map->m_flags & (1 << BH_Delay))
2450                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2451
2452         err = ext4_map_blocks(handle, inode, map, get_blocks_flags);
2453         if (err < 0)
2454                 return err;
2455         if (dioread_nolock && (map->m_flags & EXT4_MAP_UNWRITTEN)) {
2456                 if (!mpd->io_submit.io_end->handle &&
2457                     ext4_handle_valid(handle)) {
2458                         mpd->io_submit.io_end->handle = handle->h_rsv_handle;
2459                         handle->h_rsv_handle = NULL;
2460                 }
2461                 ext4_set_io_unwritten_flag(inode, mpd->io_submit.io_end);
2462         }
2463
2464         BUG_ON(map->m_len == 0);
2465         if (map->m_flags & EXT4_MAP_NEW) {
2466                 clean_bdev_aliases(inode->i_sb->s_bdev, map->m_pblk,
2467                                    map->m_len);
2468         }
2469         return 0;
2470 }
2471
2472 /*
2473  * mpage_map_and_submit_extent - map extent starting at mpd->lblk of length
2474  *                               mpd->len and submit pages underlying it for IO
2475  *
2476  * @handle - handle for journal operations
2477  * @mpd - extent to map
2478  * @give_up_on_write - we set this to true iff there is a fatal error and there
2479  *                     is no hope of writing the data. The caller should discard
2480  *                     dirty pages to avoid infinite loops.
2481  *
2482  * The function maps extent starting at mpd->lblk of length mpd->len. If it is
2483  * delayed, blocks are allocated, if it is unwritten, we may need to convert
2484  * them to initialized or split the described range from larger unwritten
2485  * extent. Note that we need not map all the described range since allocation
2486  * can return less blocks or the range is covered by more unwritten extents. We
2487  * cannot map more because we are limited by reserved transaction credits. On
2488  * the other hand we always make sure that the last touched page is fully
2489  * mapped so that it can be written out (and thus forward progress is
2490  * guaranteed). After mapping we submit all mapped pages for IO.
2491  */
2492 static int mpage_map_and_submit_extent(handle_t *handle,
2493                                        struct mpage_da_data *mpd,
2494                                        bool *give_up_on_write)
2495 {
2496         struct inode *inode = mpd->inode;
2497         struct ext4_map_blocks *map = &mpd->map;
2498         int err;
2499         loff_t disksize;
2500         int progress = 0;
2501
2502         mpd->io_submit.io_end->offset =
2503                                 ((loff_t)map->m_lblk) << inode->i_blkbits;
2504         do {
2505                 err = mpage_map_one_extent(handle, mpd);
2506                 if (err < 0) {
2507                         struct super_block *sb = inode->i_sb;
2508
2509                         if (ext4_forced_shutdown(EXT4_SB(sb)) ||
2510                             EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
2511                                 goto invalidate_dirty_pages;
2512                         /*
2513                          * Let the uper layers retry transient errors.
2514                          * In the case of ENOSPC, if ext4_count_free_blocks()
2515                          * is non-zero, a commit should free up blocks.
2516                          */
2517                         if ((err == -ENOMEM) ||
2518                             (err == -ENOSPC && ext4_count_free_clusters(sb))) {
2519                                 if (progress)
2520                                         goto update_disksize;
2521                                 return err;
2522                         }
2523                         ext4_msg(sb, KERN_CRIT,
2524                                  "Delayed block allocation failed for "
2525                                  "inode %lu at logical offset %llu with"
2526                                  " max blocks %u with error %d",
2527                                  inode->i_ino,
2528                                  (unsigned long long)map->m_lblk,
2529                                  (unsigned)map->m_len, -err);
2530                         ext4_msg(sb, KERN_CRIT,
2531                                  "This should not happen!! Data will "
2532                                  "be lost\n");
2533                         if (err == -ENOSPC)
2534                                 ext4_print_free_blocks(inode);
2535                 invalidate_dirty_pages:
2536                         *give_up_on_write = true;
2537                         return err;
2538                 }
2539                 progress = 1;
2540                 /*
2541                  * Update buffer state, submit mapped pages, and get us new
2542                  * extent to map
2543                  */
2544                 err = mpage_map_and_submit_buffers(mpd);
2545                 if (err < 0)
2546                         goto update_disksize;
2547         } while (map->m_len);
2548
2549 update_disksize:
2550         /*
2551          * Update on-disk size after IO is submitted.  Races with
2552          * truncate are avoided by checking i_size under i_data_sem.
2553          */
2554         disksize = ((loff_t)mpd->first_page) << PAGE_SHIFT;
2555         if (disksize > EXT4_I(inode)->i_disksize) {
2556                 int err2;
2557                 loff_t i_size;
2558
2559                 down_write(&EXT4_I(inode)->i_data_sem);
2560                 i_size = i_size_read(inode);
2561                 if (disksize > i_size)
2562                         disksize = i_size;
2563                 if (disksize > EXT4_I(inode)->i_disksize)
2564                         EXT4_I(inode)->i_disksize = disksize;
2565                 up_write(&EXT4_I(inode)->i_data_sem);
2566                 err2 = ext4_mark_inode_dirty(handle, inode);
2567                 if (err2)
2568                         ext4_error(inode->i_sb,
2569                                    "Failed to mark inode %lu dirty",
2570                                    inode->i_ino);
2571                 if (!err)
2572                         err = err2;
2573         }
2574         return err;
2575 }
2576
2577 /*
2578  * Calculate the total number of credits to reserve for one writepages
2579  * iteration. This is called from ext4_writepages(). We map an extent of
2580  * up to MAX_WRITEPAGES_EXTENT_LEN blocks and then we go on and finish mapping
2581  * the last partial page. So in total we can map MAX_WRITEPAGES_EXTENT_LEN +
2582  * bpp - 1 blocks in bpp different extents.
2583  */
2584 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2585 {
2586         int bpp = ext4_journal_blocks_per_page(inode);
2587
2588         return ext4_meta_trans_blocks(inode,
2589                                 MAX_WRITEPAGES_EXTENT_LEN + bpp - 1, bpp);
2590 }
2591
2592 /*
2593  * mpage_prepare_extent_to_map - find & lock contiguous range of dirty pages
2594  *                               and underlying extent to map
2595  *
2596  * @mpd - where to look for pages
2597  *
2598  * Walk dirty pages in the mapping. If they are fully mapped, submit them for
2599  * IO immediately. When we find a page which isn't mapped we start accumulating
2600  * extent of buffers underlying these pages that needs mapping (formed by
2601  * either delayed or unwritten buffers). We also lock the pages containing
2602  * these buffers. The extent found is returned in @mpd structure (starting at
2603  * mpd->lblk with length mpd->len blocks).
2604  *
2605  * Note that this function can attach bios to one io_end structure which are
2606  * neither logically nor physically contiguous. Although it may seem as an
2607  * unnecessary complication, it is actually inevitable in blocksize < pagesize
2608  * case as we need to track IO to all buffers underlying a page in one io_end.
2609  */
2610 static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
2611 {
2612         struct address_space *mapping = mpd->inode->i_mapping;
2613         struct pagevec pvec;
2614         unsigned int nr_pages;
2615         long left = mpd->wbc->nr_to_write;
2616         pgoff_t index = mpd->first_page;
2617         pgoff_t end = mpd->last_page;
2618         int tag;
2619         int i, err = 0;
2620         int blkbits = mpd->inode->i_blkbits;
2621         ext4_lblk_t lblk;
2622         struct buffer_head *head;
2623
2624         if (mpd->wbc->sync_mode == WB_SYNC_ALL || mpd->wbc->tagged_writepages)
2625                 tag = PAGECACHE_TAG_TOWRITE;
2626         else
2627                 tag = PAGECACHE_TAG_DIRTY;
2628
2629         pagevec_init(&pvec);
2630         mpd->map.m_len = 0;
2631         mpd->next_page = index;
2632         while (index <= end) {
2633                 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
2634                                 tag);
2635                 if (nr_pages == 0)
2636                         goto out;
2637
2638                 for (i = 0; i < nr_pages; i++) {
2639                         struct page *page = pvec.pages[i];
2640
2641                         /*
2642                          * Accumulated enough dirty pages? This doesn't apply
2643                          * to WB_SYNC_ALL mode. For integrity sync we have to
2644                          * keep going because someone may be concurrently
2645                          * dirtying pages, and we might have synced a lot of
2646                          * newly appeared dirty pages, but have not synced all
2647                          * of the old dirty pages.
2648                          */
2649                         if (mpd->wbc->sync_mode == WB_SYNC_NONE && left <= 0)
2650                                 goto out;
2651
2652                         /* If we can't merge this page, we are done. */
2653                         if (mpd->map.m_len > 0 && mpd->next_page != page->index)
2654                                 goto out;
2655
2656                         lock_page(page);
2657                         /*
2658                          * If the page is no longer dirty, or its mapping no
2659                          * longer corresponds to inode we are writing (which
2660                          * means it has been truncated or invalidated), or the
2661                          * page is already under writeback and we are not doing
2662                          * a data integrity writeback, skip the page
2663                          */
2664                         if (!PageDirty(page) ||
2665                             (PageWriteback(page) &&
2666                              (mpd->wbc->sync_mode == WB_SYNC_NONE)) ||
2667                             unlikely(page->mapping != mapping)) {
2668                                 unlock_page(page);
2669                                 continue;
2670                         }
2671
2672                         wait_on_page_writeback(page);
2673                         BUG_ON(PageWriteback(page));
2674
2675                         if (mpd->map.m_len == 0)
2676                                 mpd->first_page = page->index;
2677                         mpd->next_page = page->index + 1;
2678                         /* Add all dirty buffers to mpd */
2679                         lblk = ((ext4_lblk_t)page->index) <<
2680                                 (PAGE_SHIFT - blkbits);
2681                         head = page_buffers(page);
2682                         err = mpage_process_page_bufs(mpd, head, head, lblk);
2683                         if (err <= 0)
2684                                 goto out;
2685                         err = 0;
2686                         left--;
2687                 }
2688                 pagevec_release(&pvec);
2689                 cond_resched();
2690         }
2691         return 0;
2692 out:
2693         pagevec_release(&pvec);
2694         return err;
2695 }
2696
2697 static int ext4_writepages(struct address_space *mapping,
2698                            struct writeback_control *wbc)
2699 {
2700         pgoff_t writeback_index = 0;
2701         long nr_to_write = wbc->nr_to_write;
2702         int range_whole = 0;
2703         int cycled = 1;
2704         handle_t *handle = NULL;
2705         struct mpage_da_data mpd;
2706         struct inode *inode = mapping->host;
2707         int needed_blocks, rsv_blocks = 0, ret = 0;
2708         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2709         bool done;
2710         struct blk_plug plug;
2711         bool give_up_on_write = false;
2712
2713         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
2714                 return -EIO;
2715
2716         percpu_down_read(&sbi->s_journal_flag_rwsem);
2717         trace_ext4_writepages(inode, wbc);
2718
2719         if (dax_mapping(mapping)) {
2720                 ret = dax_writeback_mapping_range(mapping, inode->i_sb->s_bdev,
2721                                                   wbc);
2722                 goto out_writepages;
2723         }
2724
2725         /*
2726          * No pages to write? This is mainly a kludge to avoid starting
2727          * a transaction for special inodes like journal inode on last iput()
2728          * because that could violate lock ordering on umount
2729          */
2730         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2731                 goto out_writepages;
2732
2733         if (ext4_should_journal_data(inode)) {
2734                 ret = generic_writepages(mapping, wbc);
2735                 goto out_writepages;
2736         }
2737
2738         /*
2739          * If the filesystem has aborted, it is read-only, so return
2740          * right away instead of dumping stack traces later on that
2741          * will obscure the real source of the problem.  We test
2742          * EXT4_MF_FS_ABORTED instead of sb->s_flag's SB_RDONLY because
2743          * the latter could be true if the filesystem is mounted
2744          * read-only, and in that case, ext4_writepages should
2745          * *never* be called, so if that ever happens, we would want
2746          * the stack trace.
2747          */
2748         if (unlikely(ext4_forced_shutdown(EXT4_SB(mapping->host->i_sb)) ||
2749                      sbi->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2750                 ret = -EROFS;
2751                 goto out_writepages;
2752         }
2753
2754         if (ext4_should_dioread_nolock(inode)) {
2755                 /*
2756                  * We may need to convert up to one extent per block in
2757                  * the page and we may dirty the inode.
2758                  */
2759                 rsv_blocks = 1 + (PAGE_SIZE >> inode->i_blkbits);
2760         }
2761
2762         /*
2763          * If we have inline data and arrive here, it means that
2764          * we will soon create the block for the 1st page, so
2765          * we'd better clear the inline data here.
2766          */
2767         if (ext4_has_inline_data(inode)) {
2768                 /* Just inode will be modified... */
2769                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
2770                 if (IS_ERR(handle)) {
2771                         ret = PTR_ERR(handle);
2772                         goto out_writepages;
2773                 }
2774                 BUG_ON(ext4_test_inode_state(inode,
2775                                 EXT4_STATE_MAY_INLINE_DATA));
2776                 ext4_destroy_inline_data(handle, inode);
2777                 ext4_journal_stop(handle);
2778         }
2779
2780         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2781                 range_whole = 1;
2782
2783         if (wbc->range_cyclic) {
2784                 writeback_index = mapping->writeback_index;
2785                 if (writeback_index)
2786                         cycled = 0;
2787                 mpd.first_page = writeback_index;
2788                 mpd.last_page = -1;
2789         } else {
2790                 mpd.first_page = wbc->range_start >> PAGE_SHIFT;
2791                 mpd.last_page = wbc->range_end >> PAGE_SHIFT;
2792         }
2793
2794         mpd.inode = inode;
2795         mpd.wbc = wbc;
2796         ext4_io_submit_init(&mpd.io_submit, wbc);
2797 retry:
2798         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2799                 tag_pages_for_writeback(mapping, mpd.first_page, mpd.last_page);
2800         done = false;
2801         blk_start_plug(&plug);
2802
2803         /*
2804          * First writeback pages that don't need mapping - we can avoid
2805          * starting a transaction unnecessarily and also avoid being blocked
2806          * in the block layer on device congestion while having transaction
2807          * started.
2808          */
2809         mpd.do_map = 0;
2810         mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2811         if (!mpd.io_submit.io_end) {
2812                 ret = -ENOMEM;
2813                 goto unplug;
2814         }
2815         ret = mpage_prepare_extent_to_map(&mpd);
2816         /* Submit prepared bio */
2817         ext4_io_submit(&mpd.io_submit);
2818         ext4_put_io_end_defer(mpd.io_submit.io_end);
2819         mpd.io_submit.io_end = NULL;
2820         /* Unlock pages we didn't use */
2821         mpage_release_unused_pages(&mpd, false);
2822         if (ret < 0)
2823                 goto unplug;
2824
2825         while (!done && mpd.first_page <= mpd.last_page) {
2826                 /* For each extent of pages we use new io_end */
2827                 mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2828                 if (!mpd.io_submit.io_end) {
2829                         ret = -ENOMEM;
2830                         break;
2831                 }
2832
2833                 /*
2834                  * We have two constraints: We find one extent to map and we
2835                  * must always write out whole page (makes a difference when
2836                  * blocksize < pagesize) so that we don't block on IO when we
2837                  * try to write out the rest of the page. Journalled mode is
2838                  * not supported by delalloc.
2839                  */
2840                 BUG_ON(ext4_should_journal_data(inode));
2841                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2842
2843                 /* start a new transaction */
2844                 handle = ext4_journal_start_with_reserve(inode,
2845                                 EXT4_HT_WRITE_PAGE, needed_blocks, rsv_blocks);
2846                 if (IS_ERR(handle)) {
2847                         ret = PTR_ERR(handle);
2848                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2849                                "%ld pages, ino %lu; err %d", __func__,
2850                                 wbc->nr_to_write, inode->i_ino, ret);
2851                         /* Release allocated io_end */
2852                         ext4_put_io_end(mpd.io_submit.io_end);
2853                         mpd.io_submit.io_end = NULL;
2854                         break;
2855                 }
2856                 mpd.do_map = 1;
2857
2858                 trace_ext4_da_write_pages(inode, mpd.first_page, mpd.wbc);
2859                 ret = mpage_prepare_extent_to_map(&mpd);
2860                 if (!ret) {
2861                         if (mpd.map.m_len)
2862                                 ret = mpage_map_and_submit_extent(handle, &mpd,
2863                                         &give_up_on_write);
2864                         else {
2865                                 /*
2866                                  * We scanned the whole range (or exhausted
2867                                  * nr_to_write), submitted what was mapped and
2868                                  * didn't find anything needing mapping. We are
2869                                  * done.
2870                                  */
2871                                 done = true;
2872                         }
2873                 }
2874                 /*
2875                  * Caution: If the handle is synchronous,
2876                  * ext4_journal_stop() can wait for transaction commit
2877                  * to finish which may depend on writeback of pages to
2878                  * complete or on page lock to be released.  In that
2879                  * case, we have to wait until after after we have
2880                  * submitted all the IO, released page locks we hold,
2881                  * and dropped io_end reference (for extent conversion
2882                  * to be able to complete) before stopping the handle.
2883                  */
2884                 if (!ext4_handle_valid(handle) || handle->h_sync == 0) {
2885                         ext4_journal_stop(handle);
2886                         handle = NULL;
2887                         mpd.do_map = 0;
2888                 }
2889                 /* Submit prepared bio */
2890                 ext4_io_submit(&mpd.io_submit);
2891                 /* Unlock pages we didn't use */
2892                 mpage_release_unused_pages(&mpd, give_up_on_write);
2893                 /*
2894                  * Drop our io_end reference we got from init. We have
2895                  * to be careful and use deferred io_end finishing if
2896                  * we are still holding the transaction as we can
2897                  * release the last reference to io_end which may end
2898                  * up doing unwritten extent conversion.
2899                  */
2900                 if (handle) {
2901                         ext4_put_io_end_defer(mpd.io_submit.io_end);
2902                         ext4_journal_stop(handle);
2903                 } else
2904                         ext4_put_io_end(mpd.io_submit.io_end);
2905                 mpd.io_submit.io_end = NULL;
2906
2907                 if (ret == -ENOSPC && sbi->s_journal) {
2908                         /*
2909                          * Commit the transaction which would
2910                          * free blocks released in the transaction
2911                          * and try again
2912                          */
2913                         jbd2_journal_force_commit_nested(sbi->s_journal);
2914                         ret = 0;
2915                         continue;
2916                 }
2917                 /* Fatal error - ENOMEM, EIO... */
2918                 if (ret)
2919                         break;
2920         }
2921 unplug:
2922         blk_finish_plug(&plug);
2923         if (!ret && !cycled && wbc->nr_to_write > 0) {
2924                 cycled = 1;
2925                 mpd.last_page = writeback_index - 1;
2926                 mpd.first_page = 0;
2927                 goto retry;
2928         }
2929
2930         /* Update index */
2931         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2932                 /*
2933                  * Set the writeback_index so that range_cyclic
2934                  * mode will write it back later
2935                  */
2936                 mapping->writeback_index = mpd.first_page;
2937
2938 out_writepages:
2939         trace_ext4_writepages_result(inode, wbc, ret,
2940                                      nr_to_write - wbc->nr_to_write);
2941         percpu_up_read(&sbi->s_journal_flag_rwsem);
2942         return ret;
2943 }
2944
2945 static int ext4_nonda_switch(struct super_block *sb)
2946 {
2947         s64 free_clusters, dirty_clusters;
2948         struct ext4_sb_info *sbi = EXT4_SB(sb);
2949
2950         /*
2951          * switch to non delalloc mode if we are running low
2952          * on free block. The free block accounting via percpu
2953          * counters can get slightly wrong with percpu_counter_batch getting
2954          * accumulated on each CPU without updating global counters
2955          * Delalloc need an accurate free block accounting. So switch
2956          * to non delalloc when we are near to error range.
2957          */
2958         free_clusters =
2959                 percpu_counter_read_positive(&sbi->s_freeclusters_counter);
2960         dirty_clusters =
2961                 percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
2962         /*
2963          * Start pushing delalloc when 1/2 of free blocks are dirty.
2964          */
2965         if (dirty_clusters && (free_clusters < 2 * dirty_clusters))
2966                 try_to_writeback_inodes_sb(sb, WB_REASON_FS_FREE_SPACE);
2967
2968         if (2 * free_clusters < 3 * dirty_clusters ||
2969             free_clusters < (dirty_clusters + EXT4_FREECLUSTERS_WATERMARK)) {
2970                 /*
2971                  * free block count is less than 150% of dirty blocks
2972                  * or free blocks is less than watermark
2973                  */
2974                 return 1;
2975         }
2976         return 0;
2977 }
2978
2979 /* We always reserve for an inode update; the superblock could be there too */
2980 static int ext4_da_write_credits(struct inode *inode, loff_t pos, unsigned len)
2981 {
2982         if (likely(ext4_has_feature_large_file(inode->i_sb)))
2983                 return 1;
2984
2985         if (pos + len <= 0x7fffffffULL)
2986                 return 1;
2987
2988         /* We might need to update the superblock to set LARGE_FILE */
2989         return 2;
2990 }
2991
2992 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2993                                loff_t pos, unsigned len, unsigned flags,
2994                                struct page **pagep, void **fsdata)
2995 {
2996         int ret, retries = 0;
2997         struct page *page;
2998         pgoff_t index;
2999         struct inode *inode = mapping->host;
3000         handle_t *handle;
3001
3002         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
3003                 return -EIO;
3004
3005         index = pos >> PAGE_SHIFT;
3006
3007         if (ext4_nonda_switch(inode->i_sb) ||
3008             S_ISLNK(inode->i_mode)) {
3009                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3010                 return ext4_write_begin(file, mapping, pos,
3011                                         len, flags, pagep, fsdata);
3012         }
3013         *fsdata = (void *)0;
3014         trace_ext4_da_write_begin(inode, pos, len, flags);
3015
3016         if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
3017                 ret = ext4_da_write_inline_data_begin(mapping, inode,
3018                                                       pos, len, flags,
3019                                                       pagep, fsdata);
3020                 if (ret < 0)
3021                         return ret;
3022                 if (ret == 1)
3023                         return 0;
3024         }
3025
3026         /*
3027          * grab_cache_page_write_begin() can take a long time if the
3028          * system is thrashing due to memory pressure, or if the page
3029          * is being written back.  So grab it first before we start
3030          * the transaction handle.  This also allows us to allocate
3031          * the page (if needed) without using GFP_NOFS.
3032          */
3033 retry_grab:
3034         page = grab_cache_page_write_begin(mapping, index, flags);
3035         if (!page)
3036                 return -ENOMEM;
3037         unlock_page(page);
3038
3039         /*
3040          * With delayed allocation, we don't log the i_disksize update
3041          * if there is delayed block allocation. But we still need
3042          * to journalling the i_disksize update if writes to the end
3043          * of file which has an already mapped buffer.
3044          */
3045 retry_journal:
3046         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
3047                                 ext4_da_write_credits(inode, pos, len));
3048         if (IS_ERR(handle)) {
3049                 put_page(page);
3050                 return PTR_ERR(handle);
3051         }
3052
3053         lock_page(page);
3054         if (page->mapping != mapping) {
3055                 /* The page got truncated from under us */
3056                 unlock_page(page);
3057                 put_page(page);
3058                 ext4_journal_stop(handle);
3059                 goto retry_grab;
3060         }
3061         /* In case writeback began while the page was unlocked */
3062         wait_for_stable_page(page);
3063
3064 #ifdef CONFIG_EXT4_FS_ENCRYPTION
3065         ret = ext4_block_write_begin(page, pos, len,
3066                                      ext4_da_get_block_prep);
3067 #else
3068         ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
3069 #endif
3070         if (ret < 0) {
3071                 unlock_page(page);
3072                 ext4_journal_stop(handle);
3073                 /*
3074                  * block_write_begin may have instantiated a few blocks
3075                  * outside i_size.  Trim these off again. Don't need
3076                  * i_size_read because we hold i_mutex.
3077                  */
3078                 if (pos + len > inode->i_size)
3079                         ext4_truncate_failed_write(inode);
3080
3081                 if (ret == -ENOSPC &&
3082                     ext4_should_retry_alloc(inode->i_sb, &retries))
3083                         goto retry_journal;
3084
3085                 put_page(page);
3086                 return ret;
3087         }
3088
3089         *pagep = page;
3090         return ret;
3091 }
3092
3093 /*
3094  * Check if we should update i_disksize
3095  * when write to the end of file but not require block allocation
3096  */
3097 static int ext4_da_should_update_i_disksize(struct page *page,
3098                                             unsigned long offset)
3099 {
3100         struct buffer_head *bh;
3101         struct inode *inode = page->mapping->host;
3102         unsigned int idx;
3103         int i;
3104
3105         bh = page_buffers(page);
3106         idx = offset >> inode->i_blkbits;
3107
3108         for (i = 0; i < idx; i++)
3109                 bh = bh->b_this_page;
3110
3111         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3112                 return 0;
3113         return 1;
3114 }
3115
3116 static int ext4_da_write_end(struct file *file,
3117                              struct address_space *mapping,
3118                              loff_t pos, unsigned len, unsigned copied,
3119                              struct page *page, void *fsdata)
3120 {
3121         struct inode *inode = mapping->host;
3122         int ret = 0, ret2;
3123         handle_t *handle = ext4_journal_current_handle();
3124         loff_t new_i_size;
3125         unsigned long start, end;
3126         int write_mode = (int)(unsigned long)fsdata;
3127
3128         if (write_mode == FALL_BACK_TO_NONDELALLOC)
3129                 return ext4_write_end(file, mapping, pos,
3130                                       len, copied, page, fsdata);
3131
3132         trace_ext4_da_write_end(inode, pos, len, copied);
3133         start = pos & (PAGE_SIZE - 1);
3134         end = start + copied - 1;
3135
3136         /*
3137          * generic_write_end() will run mark_inode_dirty() if i_size
3138          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3139          * into that.
3140          */
3141         new_i_size = pos + copied;
3142         if (copied && new_i_size > EXT4_I(inode)->i_disksize) {
3143                 if (ext4_has_inline_data(inode) ||
3144                     ext4_da_should_update_i_disksize(page, end)) {
3145                         ext4_update_i_disksize(inode, new_i_size);
3146                         /* We need to mark inode dirty even if
3147                          * new_i_size is less that inode->i_size
3148                          * bu greater than i_disksize.(hint delalloc)
3149                          */
3150                         ext4_mark_inode_dirty(handle, inode);
3151                 }
3152         }
3153
3154         if (write_mode != CONVERT_INLINE_DATA &&
3155             ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA) &&
3156             ext4_has_inline_data(inode))
3157                 ret2 = ext4_da_write_inline_data_end(inode, pos, len, copied,
3158                                                      page);
3159         else
3160                 ret2 = generic_write_end(file, mapping, pos, len, copied,
3161                                                         page, fsdata);
3162
3163         copied = ret2;
3164         if (ret2 < 0)
3165                 ret = ret2;
3166         ret2 = ext4_journal_stop(handle);
3167         if (!ret)
3168                 ret = ret2;
3169
3170         return ret ? ret : copied;
3171 }
3172
3173 static void ext4_da_invalidatepage(struct page *page, unsigned int offset,
3174                                    unsigned int length)
3175 {
3176         /*
3177          * Drop reserved blocks
3178          */
3179         BUG_ON(!PageLocked(page));
3180         if (!page_has_buffers(page))
3181                 goto out;
3182
3183         ext4_da_page_release_reservation(page, offset, length);
3184
3185 out:
3186         ext4_invalidatepage(page, offset, length);
3187
3188         return;
3189 }
3190
3191 /*
3192  * Force all delayed allocation blocks to be allocated for a given inode.
3193  */
3194 int ext4_alloc_da_blocks(struct inode *inode)
3195 {
3196         trace_ext4_alloc_da_blocks(inode);
3197
3198         if (!EXT4_I(inode)->i_reserved_data_blocks)
3199                 return 0;
3200
3201         /*
3202          * We do something simple for now.  The filemap_flush() will
3203          * also start triggering a write of the data blocks, which is
3204          * not strictly speaking necessary (and for users of
3205          * laptop_mode, not even desirable).  However, to do otherwise
3206          * would require replicating code paths in:
3207          *
3208          * ext4_writepages() ->
3209          *    write_cache_pages() ---> (via passed in callback function)
3210          *        __mpage_da_writepage() -->
3211          *           mpage_add_bh_to_extent()
3212          *           mpage_da_map_blocks()
3213          *
3214          * The problem is that write_cache_pages(), located in
3215          * mm/page-writeback.c, marks pages clean in preparation for
3216          * doing I/O, which is not desirable if we're not planning on
3217          * doing I/O at all.
3218          *
3219          * We could call write_cache_pages(), and then redirty all of
3220          * the pages by calling redirty_page_for_writepage() but that
3221          * would be ugly in the extreme.  So instead we would need to
3222          * replicate parts of the code in the above functions,
3223          * simplifying them because we wouldn't actually intend to
3224          * write out the pages, but rather only collect contiguous
3225          * logical block extents, call the multi-block allocator, and
3226          * then update the buffer heads with the block allocations.
3227          *
3228          * For now, though, we'll cheat by calling filemap_flush(),
3229          * which will map the blocks, and start the I/O, but not
3230          * actually wait for the I/O to complete.
3231          */
3232         return filemap_flush(inode->i_mapping);
3233 }
3234
3235 /*
3236  * bmap() is special.  It gets used by applications such as lilo and by
3237  * the swapper to find the on-disk block of a specific piece of data.
3238  *
3239  * Naturally, this is dangerous if the block concerned is still in the
3240  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3241  * filesystem and enables swap, then they may get a nasty shock when the
3242  * data getting swapped to that swapfile suddenly gets overwritten by
3243  * the original zero's written out previously to the journal and
3244  * awaiting writeback in the kernel's buffer cache.
3245  *
3246  * So, if we see any bmap calls here on a modified, data-journaled file,
3247  * take extra steps to flush any blocks which might be in the cache.
3248  */
3249 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3250 {
3251         struct inode *inode = mapping->host;
3252         journal_t *journal;
3253         int err;
3254
3255         /*
3256          * We can get here for an inline file via the FIBMAP ioctl
3257          */
3258         if (ext4_has_inline_data(inode))
3259                 return 0;
3260
3261         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3262                         test_opt(inode->i_sb, DELALLOC)) {
3263                 /*
3264                  * With delalloc we want to sync the file
3265                  * so that we can make sure we allocate
3266                  * blocks for file
3267                  */
3268                 filemap_write_and_wait(mapping);
3269         }
3270
3271         if (EXT4_JOURNAL(inode) &&
3272             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3273                 /*
3274                  * This is a REALLY heavyweight approach, but the use of
3275                  * bmap on dirty files is expected to be extremely rare:
3276                  * only if we run lilo or swapon on a freshly made file
3277                  * do we expect this to happen.
3278                  *
3279                  * (bmap requires CAP_SYS_RAWIO so this does not
3280                  * represent an unprivileged user DOS attack --- we'd be
3281                  * in trouble if mortal users could trigger this path at
3282                  * will.)
3283                  *
3284                  * NB. EXT4_STATE_JDATA is not set on files other than
3285                  * regular files.  If somebody wants to bmap a directory
3286                  * or symlink and gets confused because the buffer
3287                  * hasn't yet been flushed to disk, they deserve
3288                  * everything they get.