c774bdc22759b1602341da70a0b3fec7b8830f2c
[sfrench/cifs-2.6.git] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  64-bit file support on 64-bit platforms by Jakub Jelinek
16  *      (jj@sunsite.ms.mff.cuni.cz)
17  *
18  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
19  */
20
21 #include <linux/fs.h>
22 #include <linux/time.h>
23 #include <linux/highuid.h>
24 #include <linux/pagemap.h>
25 #include <linux/dax.h>
26 #include <linux/quotaops.h>
27 #include <linux/string.h>
28 #include <linux/buffer_head.h>
29 #include <linux/writeback.h>
30 #include <linux/pagevec.h>
31 #include <linux/mpage.h>
32 #include <linux/namei.h>
33 #include <linux/uio.h>
34 #include <linux/bio.h>
35 #include <linux/workqueue.h>
36 #include <linux/kernel.h>
37 #include <linux/printk.h>
38 #include <linux/slab.h>
39 #include <linux/bitops.h>
40 #include <linux/iomap.h>
41
42 #include "ext4_jbd2.h"
43 #include "xattr.h"
44 #include "acl.h"
45 #include "truncate.h"
46
47 #include <trace/events/ext4.h>
48
49 #define MPAGE_DA_EXTENT_TAIL 0x01
50
51 static __u32 ext4_inode_csum(struct inode *inode, struct ext4_inode *raw,
52                               struct ext4_inode_info *ei)
53 {
54         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
55         __u32 csum;
56         __u16 dummy_csum = 0;
57         int offset = offsetof(struct ext4_inode, i_checksum_lo);
58         unsigned int csum_size = sizeof(dummy_csum);
59
60         csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)raw, offset);
61         csum = ext4_chksum(sbi, csum, (__u8 *)&dummy_csum, csum_size);
62         offset += csum_size;
63         csum = ext4_chksum(sbi, csum, (__u8 *)raw + offset,
64                            EXT4_GOOD_OLD_INODE_SIZE - offset);
65
66         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
67                 offset = offsetof(struct ext4_inode, i_checksum_hi);
68                 csum = ext4_chksum(sbi, csum, (__u8 *)raw +
69                                    EXT4_GOOD_OLD_INODE_SIZE,
70                                    offset - EXT4_GOOD_OLD_INODE_SIZE);
71                 if (EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi)) {
72                         csum = ext4_chksum(sbi, csum, (__u8 *)&dummy_csum,
73                                            csum_size);
74                         offset += csum_size;
75                 }
76                 csum = ext4_chksum(sbi, csum, (__u8 *)raw + offset,
77                                    EXT4_INODE_SIZE(inode->i_sb) - offset);
78         }
79
80         return csum;
81 }
82
83 static int ext4_inode_csum_verify(struct inode *inode, struct ext4_inode *raw,
84                                   struct ext4_inode_info *ei)
85 {
86         __u32 provided, calculated;
87
88         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
89             cpu_to_le32(EXT4_OS_LINUX) ||
90             !ext4_has_metadata_csum(inode->i_sb))
91                 return 1;
92
93         provided = le16_to_cpu(raw->i_checksum_lo);
94         calculated = ext4_inode_csum(inode, raw, ei);
95         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
96             EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
97                 provided |= ((__u32)le16_to_cpu(raw->i_checksum_hi)) << 16;
98         else
99                 calculated &= 0xFFFF;
100
101         return provided == calculated;
102 }
103
104 static void ext4_inode_csum_set(struct inode *inode, struct ext4_inode *raw,
105                                 struct ext4_inode_info *ei)
106 {
107         __u32 csum;
108
109         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
110             cpu_to_le32(EXT4_OS_LINUX) ||
111             !ext4_has_metadata_csum(inode->i_sb))
112                 return;
113
114         csum = ext4_inode_csum(inode, raw, ei);
115         raw->i_checksum_lo = cpu_to_le16(csum & 0xFFFF);
116         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
117             EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
118                 raw->i_checksum_hi = cpu_to_le16(csum >> 16);
119 }
120
121 static inline int ext4_begin_ordered_truncate(struct inode *inode,
122                                               loff_t new_size)
123 {
124         trace_ext4_begin_ordered_truncate(inode, new_size);
125         /*
126          * If jinode is zero, then we never opened the file for
127          * writing, so there's no need to call
128          * jbd2_journal_begin_ordered_truncate() since there's no
129          * outstanding writes we need to flush.
130          */
131         if (!EXT4_I(inode)->jinode)
132                 return 0;
133         return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
134                                                    EXT4_I(inode)->jinode,
135                                                    new_size);
136 }
137
138 static void ext4_invalidatepage(struct page *page, unsigned int offset,
139                                 unsigned int length);
140 static int __ext4_journalled_writepage(struct page *page, unsigned int len);
141 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
142 static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
143                                   int pextents);
144
145 /*
146  * Test whether an inode is a fast symlink.
147  * A fast symlink has its symlink data stored in ext4_inode_info->i_data.
148  */
149 int ext4_inode_is_fast_symlink(struct inode *inode)
150 {
151         return S_ISLNK(inode->i_mode) && inode->i_size &&
152                (inode->i_size < EXT4_N_BLOCKS * 4);
153 }
154
155 /*
156  * Restart the transaction associated with *handle.  This does a commit,
157  * so before we call here everything must be consistently dirtied against
158  * this transaction.
159  */
160 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
161                                  int nblocks)
162 {
163         int ret;
164
165         /*
166          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
167          * moment, get_block can be called only for blocks inside i_size since
168          * page cache has been already dropped and writes are blocked by
169          * i_mutex. So we can safely drop the i_data_sem here.
170          */
171         BUG_ON(EXT4_JOURNAL(inode) == NULL);
172         jbd_debug(2, "restarting handle %p\n", handle);
173         up_write(&EXT4_I(inode)->i_data_sem);
174         ret = ext4_journal_restart(handle, nblocks);
175         down_write(&EXT4_I(inode)->i_data_sem);
176         ext4_discard_preallocations(inode);
177
178         return ret;
179 }
180
181 /*
182  * Called at the last iput() if i_nlink is zero.
183  */
184 void ext4_evict_inode(struct inode *inode)
185 {
186         handle_t *handle;
187         int err;
188         int extra_credits = 3;
189         struct ext4_xattr_inode_array *ea_inode_array = NULL;
190
191         trace_ext4_evict_inode(inode);
192
193         if (inode->i_nlink) {
194                 /*
195                  * When journalling data dirty buffers are tracked only in the
196                  * journal. So although mm thinks everything is clean and
197                  * ready for reaping the inode might still have some pages to
198                  * write in the running transaction or waiting to be
199                  * checkpointed. Thus calling jbd2_journal_invalidatepage()
200                  * (via truncate_inode_pages()) to discard these buffers can
201                  * cause data loss. Also even if we did not discard these
202                  * buffers, we would have no way to find them after the inode
203                  * is reaped and thus user could see stale data if he tries to
204                  * read them before the transaction is checkpointed. So be
205                  * careful and force everything to disk here... We use
206                  * ei->i_datasync_tid to store the newest transaction
207                  * containing inode's data.
208                  *
209                  * Note that directories do not have this problem because they
210                  * don't use page cache.
211                  */
212                 if (inode->i_ino != EXT4_JOURNAL_INO &&
213                     ext4_should_journal_data(inode) &&
214                     (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode)) &&
215                     inode->i_data.nrpages) {
216                         journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
217                         tid_t commit_tid = EXT4_I(inode)->i_datasync_tid;
218
219                         jbd2_complete_transaction(journal, commit_tid);
220                         filemap_write_and_wait(&inode->i_data);
221                 }
222                 truncate_inode_pages_final(&inode->i_data);
223
224                 goto no_delete;
225         }
226
227         if (is_bad_inode(inode))
228                 goto no_delete;
229         dquot_initialize(inode);
230
231         if (ext4_should_order_data(inode))
232                 ext4_begin_ordered_truncate(inode, 0);
233         truncate_inode_pages_final(&inode->i_data);
234
235         /*
236          * Protect us against freezing - iput() caller didn't have to have any
237          * protection against it
238          */
239         sb_start_intwrite(inode->i_sb);
240
241         if (!IS_NOQUOTA(inode))
242                 extra_credits += EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb);
243
244         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE,
245                                  ext4_blocks_for_truncate(inode)+extra_credits);
246         if (IS_ERR(handle)) {
247                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
248                 /*
249                  * If we're going to skip the normal cleanup, we still need to
250                  * make sure that the in-core orphan linked list is properly
251                  * cleaned up.
252                  */
253                 ext4_orphan_del(NULL, inode);
254                 sb_end_intwrite(inode->i_sb);
255                 goto no_delete;
256         }
257
258         if (IS_SYNC(inode))
259                 ext4_handle_sync(handle);
260
261         /*
262          * Set inode->i_size to 0 before calling ext4_truncate(). We need
263          * special handling of symlinks here because i_size is used to
264          * determine whether ext4_inode_info->i_data contains symlink data or
265          * block mappings. Setting i_size to 0 will remove its fast symlink
266          * status. Erase i_data so that it becomes a valid empty block map.
267          */
268         if (ext4_inode_is_fast_symlink(inode))
269                 memset(EXT4_I(inode)->i_data, 0, sizeof(EXT4_I(inode)->i_data));
270         inode->i_size = 0;
271         err = ext4_mark_inode_dirty(handle, inode);
272         if (err) {
273                 ext4_warning(inode->i_sb,
274                              "couldn't mark inode dirty (err %d)", err);
275                 goto stop_handle;
276         }
277         if (inode->i_blocks) {
278                 err = ext4_truncate(inode);
279                 if (err) {
280                         ext4_error(inode->i_sb,
281                                    "couldn't truncate inode %lu (err %d)",
282                                    inode->i_ino, err);
283                         goto stop_handle;
284                 }
285         }
286
287         /* Remove xattr references. */
288         err = ext4_xattr_delete_inode(handle, inode, &ea_inode_array,
289                                       extra_credits);
290         if (err) {
291                 ext4_warning(inode->i_sb, "xattr delete (err %d)", err);
292 stop_handle:
293                 ext4_journal_stop(handle);
294                 ext4_orphan_del(NULL, inode);
295                 sb_end_intwrite(inode->i_sb);
296                 ext4_xattr_inode_array_free(ea_inode_array);
297                 goto no_delete;
298         }
299
300         /*
301          * Kill off the orphan record which ext4_truncate created.
302          * AKPM: I think this can be inside the above `if'.
303          * Note that ext4_orphan_del() has to be able to cope with the
304          * deletion of a non-existent orphan - this is because we don't
305          * know if ext4_truncate() actually created an orphan record.
306          * (Well, we could do this if we need to, but heck - it works)
307          */
308         ext4_orphan_del(handle, inode);
309         EXT4_I(inode)->i_dtime  = get_seconds();
310
311         /*
312          * One subtle ordering requirement: if anything has gone wrong
313          * (transaction abort, IO errors, whatever), then we can still
314          * do these next steps (the fs will already have been marked as
315          * having errors), but we can't free the inode if the mark_dirty
316          * fails.
317          */
318         if (ext4_mark_inode_dirty(handle, inode))
319                 /* If that failed, just do the required in-core inode clear. */
320                 ext4_clear_inode(inode);
321         else
322                 ext4_free_inode(handle, inode);
323         ext4_journal_stop(handle);
324         sb_end_intwrite(inode->i_sb);
325         ext4_xattr_inode_array_free(ea_inode_array);
326         return;
327 no_delete:
328         ext4_clear_inode(inode);        /* We must guarantee clearing of inode... */
329 }
330
331 #ifdef CONFIG_QUOTA
332 qsize_t *ext4_get_reserved_space(struct inode *inode)
333 {
334         return &EXT4_I(inode)->i_reserved_quota;
335 }
336 #endif
337
338 /*
339  * Called with i_data_sem down, which is important since we can call
340  * ext4_discard_preallocations() from here.
341  */
342 void ext4_da_update_reserve_space(struct inode *inode,
343                                         int used, int quota_claim)
344 {
345         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
346         struct ext4_inode_info *ei = EXT4_I(inode);
347
348         spin_lock(&ei->i_block_reservation_lock);
349         trace_ext4_da_update_reserve_space(inode, used, quota_claim);
350         if (unlikely(used > ei->i_reserved_data_blocks)) {
351                 ext4_warning(inode->i_sb, "%s: ino %lu, used %d "
352                          "with only %d reserved data blocks",
353                          __func__, inode->i_ino, used,
354                          ei->i_reserved_data_blocks);
355                 WARN_ON(1);
356                 used = ei->i_reserved_data_blocks;
357         }
358
359         /* Update per-inode reservations */
360         ei->i_reserved_data_blocks -= used;
361         percpu_counter_sub(&sbi->s_dirtyclusters_counter, used);
362
363         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
364
365         /* Update quota subsystem for data blocks */
366         if (quota_claim)
367                 dquot_claim_block(inode, EXT4_C2B(sbi, used));
368         else {
369                 /*
370                  * We did fallocate with an offset that is already delayed
371                  * allocated. So on delayed allocated writeback we should
372                  * not re-claim the quota for fallocated blocks.
373                  */
374                 dquot_release_reservation_block(inode, EXT4_C2B(sbi, used));
375         }
376
377         /*
378          * If we have done all the pending block allocations and if
379          * there aren't any writers on the inode, we can discard the
380          * inode's preallocations.
381          */
382         if ((ei->i_reserved_data_blocks == 0) &&
383             (atomic_read(&inode->i_writecount) == 0))
384                 ext4_discard_preallocations(inode);
385 }
386
387 static int __check_block_validity(struct inode *inode, const char *func,
388                                 unsigned int line,
389                                 struct ext4_map_blocks *map)
390 {
391         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
392                                    map->m_len)) {
393                 ext4_error_inode(inode, func, line, map->m_pblk,
394                                  "lblock %lu mapped to illegal pblock "
395                                  "(length %d)", (unsigned long) map->m_lblk,
396                                  map->m_len);
397                 return -EFSCORRUPTED;
398         }
399         return 0;
400 }
401
402 int ext4_issue_zeroout(struct inode *inode, ext4_lblk_t lblk, ext4_fsblk_t pblk,
403                        ext4_lblk_t len)
404 {
405         int ret;
406
407         if (ext4_encrypted_inode(inode))
408                 return fscrypt_zeroout_range(inode, lblk, pblk, len);
409
410         ret = sb_issue_zeroout(inode->i_sb, pblk, len, GFP_NOFS);
411         if (ret > 0)
412                 ret = 0;
413
414         return ret;
415 }
416
417 #define check_block_validity(inode, map)        \
418         __check_block_validity((inode), __func__, __LINE__, (map))
419
420 #ifdef ES_AGGRESSIVE_TEST
421 static void ext4_map_blocks_es_recheck(handle_t *handle,
422                                        struct inode *inode,
423                                        struct ext4_map_blocks *es_map,
424                                        struct ext4_map_blocks *map,
425                                        int flags)
426 {
427         int retval;
428
429         map->m_flags = 0;
430         /*
431          * There is a race window that the result is not the same.
432          * e.g. xfstests #223 when dioread_nolock enables.  The reason
433          * is that we lookup a block mapping in extent status tree with
434          * out taking i_data_sem.  So at the time the unwritten extent
435          * could be converted.
436          */
437         down_read(&EXT4_I(inode)->i_data_sem);
438         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
439                 retval = ext4_ext_map_blocks(handle, inode, map, flags &
440                                              EXT4_GET_BLOCKS_KEEP_SIZE);
441         } else {
442                 retval = ext4_ind_map_blocks(handle, inode, map, flags &
443                                              EXT4_GET_BLOCKS_KEEP_SIZE);
444         }
445         up_read((&EXT4_I(inode)->i_data_sem));
446
447         /*
448          * We don't check m_len because extent will be collpased in status
449          * tree.  So the m_len might not equal.
450          */
451         if (es_map->m_lblk != map->m_lblk ||
452             es_map->m_flags != map->m_flags ||
453             es_map->m_pblk != map->m_pblk) {
454                 printk("ES cache assertion failed for inode: %lu "
455                        "es_cached ex [%d/%d/%llu/%x] != "
456                        "found ex [%d/%d/%llu/%x] retval %d flags %x\n",
457                        inode->i_ino, es_map->m_lblk, es_map->m_len,
458                        es_map->m_pblk, es_map->m_flags, map->m_lblk,
459                        map->m_len, map->m_pblk, map->m_flags,
460                        retval, flags);
461         }
462 }
463 #endif /* ES_AGGRESSIVE_TEST */
464
465 /*
466  * The ext4_map_blocks() function tries to look up the requested blocks,
467  * and returns if the blocks are already mapped.
468  *
469  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
470  * and store the allocated blocks in the result buffer head and mark it
471  * mapped.
472  *
473  * If file type is extents based, it will call ext4_ext_map_blocks(),
474  * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
475  * based files
476  *
477  * On success, it returns the number of blocks being mapped or allocated.  if
478  * create==0 and the blocks are pre-allocated and unwritten, the resulting @map
479  * is marked as unwritten. If the create == 1, it will mark @map as mapped.
480  *
481  * It returns 0 if plain look up failed (blocks have not been allocated), in
482  * that case, @map is returned as unmapped but we still do fill map->m_len to
483  * indicate the length of a hole starting at map->m_lblk.
484  *
485  * It returns the error in case of allocation failure.
486  */
487 int ext4_map_blocks(handle_t *handle, struct inode *inode,
488                     struct ext4_map_blocks *map, int flags)
489 {
490         struct extent_status es;
491         int retval;
492         int ret = 0;
493 #ifdef ES_AGGRESSIVE_TEST
494         struct ext4_map_blocks orig_map;
495
496         memcpy(&orig_map, map, sizeof(*map));
497 #endif
498
499         map->m_flags = 0;
500         ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
501                   "logical block %lu\n", inode->i_ino, flags, map->m_len,
502                   (unsigned long) map->m_lblk);
503
504         /*
505          * ext4_map_blocks returns an int, and m_len is an unsigned int
506          */
507         if (unlikely(map->m_len > INT_MAX))
508                 map->m_len = INT_MAX;
509
510         /* We can handle the block number less than EXT_MAX_BLOCKS */
511         if (unlikely(map->m_lblk >= EXT_MAX_BLOCKS))
512                 return -EFSCORRUPTED;
513
514         /* Lookup extent status tree firstly */
515         if (ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
516                 if (ext4_es_is_written(&es) || ext4_es_is_unwritten(&es)) {
517                         map->m_pblk = ext4_es_pblock(&es) +
518                                         map->m_lblk - es.es_lblk;
519                         map->m_flags |= ext4_es_is_written(&es) ?
520                                         EXT4_MAP_MAPPED : EXT4_MAP_UNWRITTEN;
521                         retval = es.es_len - (map->m_lblk - es.es_lblk);
522                         if (retval > map->m_len)
523                                 retval = map->m_len;
524                         map->m_len = retval;
525                 } else if (ext4_es_is_delayed(&es) || ext4_es_is_hole(&es)) {
526                         map->m_pblk = 0;
527                         retval = es.es_len - (map->m_lblk - es.es_lblk);
528                         if (retval > map->m_len)
529                                 retval = map->m_len;
530                         map->m_len = retval;
531                         retval = 0;
532                 } else {
533                         BUG_ON(1);
534                 }
535 #ifdef ES_AGGRESSIVE_TEST
536                 ext4_map_blocks_es_recheck(handle, inode, map,
537                                            &orig_map, flags);
538 #endif
539                 goto found;
540         }
541
542         /*
543          * Try to see if we can get the block without requesting a new
544          * file system block.
545          */
546         down_read(&EXT4_I(inode)->i_data_sem);
547         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
548                 retval = ext4_ext_map_blocks(handle, inode, map, flags &
549                                              EXT4_GET_BLOCKS_KEEP_SIZE);
550         } else {
551                 retval = ext4_ind_map_blocks(handle, inode, map, flags &
552                                              EXT4_GET_BLOCKS_KEEP_SIZE);
553         }
554         if (retval > 0) {
555                 unsigned int status;
556
557                 if (unlikely(retval != map->m_len)) {
558                         ext4_warning(inode->i_sb,
559                                      "ES len assertion failed for inode "
560                                      "%lu: retval %d != map->m_len %d",
561                                      inode->i_ino, retval, map->m_len);
562                         WARN_ON(1);
563                 }
564
565                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
566                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
567                 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
568                     !(status & EXTENT_STATUS_WRITTEN) &&
569                     ext4_find_delalloc_range(inode, map->m_lblk,
570                                              map->m_lblk + map->m_len - 1))
571                         status |= EXTENT_STATUS_DELAYED;
572                 ret = ext4_es_insert_extent(inode, map->m_lblk,
573                                             map->m_len, map->m_pblk, status);
574                 if (ret < 0)
575                         retval = ret;
576         }
577         up_read((&EXT4_I(inode)->i_data_sem));
578
579 found:
580         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
581                 ret = check_block_validity(inode, map);
582                 if (ret != 0)
583                         return ret;
584         }
585
586         /* If it is only a block(s) look up */
587         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
588                 return retval;
589
590         /*
591          * Returns if the blocks have already allocated
592          *
593          * Note that if blocks have been preallocated
594          * ext4_ext_get_block() returns the create = 0
595          * with buffer head unmapped.
596          */
597         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
598                 /*
599                  * If we need to convert extent to unwritten
600                  * we continue and do the actual work in
601                  * ext4_ext_map_blocks()
602                  */
603                 if (!(flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN))
604                         return retval;
605
606         /*
607          * Here we clear m_flags because after allocating an new extent,
608          * it will be set again.
609          */
610         map->m_flags &= ~EXT4_MAP_FLAGS;
611
612         /*
613          * New blocks allocate and/or writing to unwritten extent
614          * will possibly result in updating i_data, so we take
615          * the write lock of i_data_sem, and call get_block()
616          * with create == 1 flag.
617          */
618         down_write(&EXT4_I(inode)->i_data_sem);
619
620         /*
621          * We need to check for EXT4 here because migrate
622          * could have changed the inode type in between
623          */
624         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
625                 retval = ext4_ext_map_blocks(handle, inode, map, flags);
626         } else {
627                 retval = ext4_ind_map_blocks(handle, inode, map, flags);
628
629                 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
630                         /*
631                          * We allocated new blocks which will result in
632                          * i_data's format changing.  Force the migrate
633                          * to fail by clearing migrate flags
634                          */
635                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
636                 }
637
638                 /*
639                  * Update reserved blocks/metadata blocks after successful
640                  * block allocation which had been deferred till now. We don't
641                  * support fallocate for non extent files. So we can update
642                  * reserve space here.
643                  */
644                 if ((retval > 0) &&
645                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
646                         ext4_da_update_reserve_space(inode, retval, 1);
647         }
648
649         if (retval > 0) {
650                 unsigned int status;
651
652                 if (unlikely(retval != map->m_len)) {
653                         ext4_warning(inode->i_sb,
654                                      "ES len assertion failed for inode "
655                                      "%lu: retval %d != map->m_len %d",
656                                      inode->i_ino, retval, map->m_len);
657                         WARN_ON(1);
658                 }
659
660                 /*
661                  * We have to zeroout blocks before inserting them into extent
662                  * status tree. Otherwise someone could look them up there and
663                  * use them before they are really zeroed. We also have to
664                  * unmap metadata before zeroing as otherwise writeback can
665                  * overwrite zeros with stale data from block device.
666                  */
667                 if (flags & EXT4_GET_BLOCKS_ZERO &&
668                     map->m_flags & EXT4_MAP_MAPPED &&
669                     map->m_flags & EXT4_MAP_NEW) {
670                         clean_bdev_aliases(inode->i_sb->s_bdev, map->m_pblk,
671                                            map->m_len);
672                         ret = ext4_issue_zeroout(inode, map->m_lblk,
673                                                  map->m_pblk, map->m_len);
674                         if (ret) {
675                                 retval = ret;
676                                 goto out_sem;
677                         }
678                 }
679
680                 /*
681                  * If the extent has been zeroed out, we don't need to update
682                  * extent status tree.
683                  */
684                 if ((flags & EXT4_GET_BLOCKS_PRE_IO) &&
685                     ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
686                         if (ext4_es_is_written(&es))
687                                 goto out_sem;
688                 }
689                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
690                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
691                 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
692                     !(status & EXTENT_STATUS_WRITTEN) &&
693                     ext4_find_delalloc_range(inode, map->m_lblk,
694                                              map->m_lblk + map->m_len - 1))
695                         status |= EXTENT_STATUS_DELAYED;
696                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
697                                             map->m_pblk, status);
698                 if (ret < 0) {
699                         retval = ret;
700                         goto out_sem;
701                 }
702         }
703
704 out_sem:
705         up_write((&EXT4_I(inode)->i_data_sem));
706         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
707                 ret = check_block_validity(inode, map);
708                 if (ret != 0)
709                         return ret;
710
711                 /*
712                  * Inodes with freshly allocated blocks where contents will be
713                  * visible after transaction commit must be on transaction's
714                  * ordered data list.
715                  */
716                 if (map->m_flags & EXT4_MAP_NEW &&
717                     !(map->m_flags & EXT4_MAP_UNWRITTEN) &&
718                     !(flags & EXT4_GET_BLOCKS_ZERO) &&
719                     !ext4_is_quota_file(inode) &&
720                     ext4_should_order_data(inode)) {
721                         if (flags & EXT4_GET_BLOCKS_IO_SUBMIT)
722                                 ret = ext4_jbd2_inode_add_wait(handle, inode);
723                         else
724                                 ret = ext4_jbd2_inode_add_write(handle, inode);
725                         if (ret)
726                                 return ret;
727                 }
728         }
729         return retval;
730 }
731
732 /*
733  * Update EXT4_MAP_FLAGS in bh->b_state. For buffer heads attached to pages
734  * we have to be careful as someone else may be manipulating b_state as well.
735  */
736 static void ext4_update_bh_state(struct buffer_head *bh, unsigned long flags)
737 {
738         unsigned long old_state;
739         unsigned long new_state;
740
741         flags &= EXT4_MAP_FLAGS;
742
743         /* Dummy buffer_head? Set non-atomically. */
744         if (!bh->b_page) {
745                 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | flags;
746                 return;
747         }
748         /*
749          * Someone else may be modifying b_state. Be careful! This is ugly but
750          * once we get rid of using bh as a container for mapping information
751          * to pass to / from get_block functions, this can go away.
752          */
753         do {
754                 old_state = READ_ONCE(bh->b_state);
755                 new_state = (old_state & ~EXT4_MAP_FLAGS) | flags;
756         } while (unlikely(
757                  cmpxchg(&bh->b_state, old_state, new_state) != old_state));
758 }
759
760 static int _ext4_get_block(struct inode *inode, sector_t iblock,
761                            struct buffer_head *bh, int flags)
762 {
763         struct ext4_map_blocks map;
764         int ret = 0;
765
766         if (ext4_has_inline_data(inode))
767                 return -ERANGE;
768
769         map.m_lblk = iblock;
770         map.m_len = bh->b_size >> inode->i_blkbits;
771
772         ret = ext4_map_blocks(ext4_journal_current_handle(), inode, &map,
773                               flags);
774         if (ret > 0) {
775                 map_bh(bh, inode->i_sb, map.m_pblk);
776                 ext4_update_bh_state(bh, map.m_flags);
777                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
778                 ret = 0;
779         } else if (ret == 0) {
780                 /* hole case, need to fill in bh->b_size */
781                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
782         }
783         return ret;
784 }
785
786 int ext4_get_block(struct inode *inode, sector_t iblock,
787                    struct buffer_head *bh, int create)
788 {
789         return _ext4_get_block(inode, iblock, bh,
790                                create ? EXT4_GET_BLOCKS_CREATE : 0);
791 }
792
793 /*
794  * Get block function used when preparing for buffered write if we require
795  * creating an unwritten extent if blocks haven't been allocated.  The extent
796  * will be converted to written after the IO is complete.
797  */
798 int ext4_get_block_unwritten(struct inode *inode, sector_t iblock,
799                              struct buffer_head *bh_result, int create)
800 {
801         ext4_debug("ext4_get_block_unwritten: inode %lu, create flag %d\n",
802                    inode->i_ino, create);
803         return _ext4_get_block(inode, iblock, bh_result,
804                                EXT4_GET_BLOCKS_IO_CREATE_EXT);
805 }
806
807 /* Maximum number of blocks we map for direct IO at once. */
808 #define DIO_MAX_BLOCKS 4096
809
810 /*
811  * Get blocks function for the cases that need to start a transaction -
812  * generally difference cases of direct IO and DAX IO. It also handles retries
813  * in case of ENOSPC.
814  */
815 static int ext4_get_block_trans(struct inode *inode, sector_t iblock,
816                                 struct buffer_head *bh_result, int flags)
817 {
818         int dio_credits;
819         handle_t *handle;
820         int retries = 0;
821         int ret;
822
823         /* Trim mapping request to maximum we can map at once for DIO */
824         if (bh_result->b_size >> inode->i_blkbits > DIO_MAX_BLOCKS)
825                 bh_result->b_size = DIO_MAX_BLOCKS << inode->i_blkbits;
826         dio_credits = ext4_chunk_trans_blocks(inode,
827                                       bh_result->b_size >> inode->i_blkbits);
828 retry:
829         handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS, dio_credits);
830         if (IS_ERR(handle))
831                 return PTR_ERR(handle);
832
833         ret = _ext4_get_block(inode, iblock, bh_result, flags);
834         ext4_journal_stop(handle);
835
836         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
837                 goto retry;
838         return ret;
839 }
840
841 /* Get block function for DIO reads and writes to inodes without extents */
842 int ext4_dio_get_block(struct inode *inode, sector_t iblock,
843                        struct buffer_head *bh, int create)
844 {
845         /* We don't expect handle for direct IO */
846         WARN_ON_ONCE(ext4_journal_current_handle());
847
848         if (!create)
849                 return _ext4_get_block(inode, iblock, bh, 0);
850         return ext4_get_block_trans(inode, iblock, bh, EXT4_GET_BLOCKS_CREATE);
851 }
852
853 /*
854  * Get block function for AIO DIO writes when we create unwritten extent if
855  * blocks are not allocated yet. The extent will be converted to written
856  * after IO is complete.
857  */
858 static int ext4_dio_get_block_unwritten_async(struct inode *inode,
859                 sector_t iblock, struct buffer_head *bh_result, int create)
860 {
861         int ret;
862
863         /* We don't expect handle for direct IO */
864         WARN_ON_ONCE(ext4_journal_current_handle());
865
866         ret = ext4_get_block_trans(inode, iblock, bh_result,
867                                    EXT4_GET_BLOCKS_IO_CREATE_EXT);
868
869         /*
870          * When doing DIO using unwritten extents, we need io_end to convert
871          * unwritten extents to written on IO completion. We allocate io_end
872          * once we spot unwritten extent and store it in b_private. Generic
873          * DIO code keeps b_private set and furthermore passes the value to
874          * our completion callback in 'private' argument.
875          */
876         if (!ret && buffer_unwritten(bh_result)) {
877                 if (!bh_result->b_private) {
878                         ext4_io_end_t *io_end;
879
880                         io_end = ext4_init_io_end(inode, GFP_KERNEL);
881                         if (!io_end)
882                                 return -ENOMEM;
883                         bh_result->b_private = io_end;
884                         ext4_set_io_unwritten_flag(inode, io_end);
885                 }
886                 set_buffer_defer_completion(bh_result);
887         }
888
889         return ret;
890 }
891
892 /*
893  * Get block function for non-AIO DIO writes when we create unwritten extent if
894  * blocks are not allocated yet. The extent will be converted to written
895  * after IO is complete by ext4_direct_IO_write().
896  */
897 static int ext4_dio_get_block_unwritten_sync(struct inode *inode,
898                 sector_t iblock, struct buffer_head *bh_result, int create)
899 {
900         int ret;
901
902         /* We don't expect handle for direct IO */
903         WARN_ON_ONCE(ext4_journal_current_handle());
904
905         ret = ext4_get_block_trans(inode, iblock, bh_result,
906                                    EXT4_GET_BLOCKS_IO_CREATE_EXT);
907
908         /*
909          * Mark inode as having pending DIO writes to unwritten extents.
910          * ext4_direct_IO_write() checks this flag and converts extents to
911          * written.
912          */
913         if (!ret && buffer_unwritten(bh_result))
914                 ext4_set_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
915
916         return ret;
917 }
918
919 static int ext4_dio_get_block_overwrite(struct inode *inode, sector_t iblock,
920                    struct buffer_head *bh_result, int create)
921 {
922         int ret;
923
924         ext4_debug("ext4_dio_get_block_overwrite: inode %lu, create flag %d\n",
925                    inode->i_ino, create);
926         /* We don't expect handle for direct IO */
927         WARN_ON_ONCE(ext4_journal_current_handle());
928
929         ret = _ext4_get_block(inode, iblock, bh_result, 0);
930         /*
931          * Blocks should have been preallocated! ext4_file_write_iter() checks
932          * that.
933          */
934         WARN_ON_ONCE(!buffer_mapped(bh_result) || buffer_unwritten(bh_result));
935
936         return ret;
937 }
938
939
940 /*
941  * `handle' can be NULL if create is zero
942  */
943 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
944                                 ext4_lblk_t block, int map_flags)
945 {
946         struct ext4_map_blocks map;
947         struct buffer_head *bh;
948         int create = map_flags & EXT4_GET_BLOCKS_CREATE;
949         int err;
950
951         J_ASSERT(handle != NULL || create == 0);
952
953         map.m_lblk = block;
954         map.m_len = 1;
955         err = ext4_map_blocks(handle, inode, &map, map_flags);
956
957         if (err == 0)
958                 return create ? ERR_PTR(-ENOSPC) : NULL;
959         if (err < 0)
960                 return ERR_PTR(err);
961
962         bh = sb_getblk(inode->i_sb, map.m_pblk);
963         if (unlikely(!bh))
964                 return ERR_PTR(-ENOMEM);
965         if (map.m_flags & EXT4_MAP_NEW) {
966                 J_ASSERT(create != 0);
967                 J_ASSERT(handle != NULL);
968
969                 /*
970                  * Now that we do not always journal data, we should
971                  * keep in mind whether this should always journal the
972                  * new buffer as metadata.  For now, regular file
973                  * writes use ext4_get_block instead, so it's not a
974                  * problem.
975                  */
976                 lock_buffer(bh);
977                 BUFFER_TRACE(bh, "call get_create_access");
978                 err = ext4_journal_get_create_access(handle, bh);
979                 if (unlikely(err)) {
980                         unlock_buffer(bh);
981                         goto errout;
982                 }
983                 if (!buffer_uptodate(bh)) {
984                         memset(bh->b_data, 0, inode->i_sb->s_blocksize);
985                         set_buffer_uptodate(bh);
986                 }
987                 unlock_buffer(bh);
988                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
989                 err = ext4_handle_dirty_metadata(handle, inode, bh);
990                 if (unlikely(err))
991                         goto errout;
992         } else
993                 BUFFER_TRACE(bh, "not a new buffer");
994         return bh;
995 errout:
996         brelse(bh);
997         return ERR_PTR(err);
998 }
999
1000 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1001                                ext4_lblk_t block, int map_flags)
1002 {
1003         struct buffer_head *bh;
1004
1005         bh = ext4_getblk(handle, inode, block, map_flags);
1006         if (IS_ERR(bh))
1007                 return bh;
1008         if (!bh || buffer_uptodate(bh))
1009                 return bh;
1010         ll_rw_block(REQ_OP_READ, REQ_META | REQ_PRIO, 1, &bh);
1011         wait_on_buffer(bh);
1012         if (buffer_uptodate(bh))
1013                 return bh;
1014         put_bh(bh);
1015         return ERR_PTR(-EIO);
1016 }
1017
1018 /* Read a contiguous batch of blocks. */
1019 int ext4_bread_batch(struct inode *inode, ext4_lblk_t block, int bh_count,
1020                      bool wait, struct buffer_head **bhs)
1021 {
1022         int i, err;
1023
1024         for (i = 0; i < bh_count; i++) {
1025                 bhs[i] = ext4_getblk(NULL, inode, block + i, 0 /* map_flags */);
1026                 if (IS_ERR(bhs[i])) {
1027                         err = PTR_ERR(bhs[i]);
1028                         bh_count = i;
1029                         goto out_brelse;
1030                 }
1031         }
1032
1033         for (i = 0; i < bh_count; i++)
1034                 /* Note that NULL bhs[i] is valid because of holes. */
1035                 if (bhs[i] && !buffer_uptodate(bhs[i]))
1036                         ll_rw_block(REQ_OP_READ, REQ_META | REQ_PRIO, 1,
1037                                     &bhs[i]);
1038
1039         if (!wait)
1040                 return 0;
1041
1042         for (i = 0; i < bh_count; i++)
1043                 if (bhs[i])
1044                         wait_on_buffer(bhs[i]);
1045
1046         for (i = 0; i < bh_count; i++) {
1047                 if (bhs[i] && !buffer_uptodate(bhs[i])) {
1048                         err = -EIO;
1049                         goto out_brelse;
1050                 }
1051         }
1052         return 0;
1053
1054 out_brelse:
1055         for (i = 0; i < bh_count; i++) {
1056                 brelse(bhs[i]);
1057                 bhs[i] = NULL;
1058         }
1059         return err;
1060 }
1061
1062 int ext4_walk_page_buffers(handle_t *handle,
1063                            struct buffer_head *head,
1064                            unsigned from,
1065                            unsigned to,
1066                            int *partial,
1067                            int (*fn)(handle_t *handle,
1068                                      struct buffer_head *bh))
1069 {
1070         struct buffer_head *bh;
1071         unsigned block_start, block_end;
1072         unsigned blocksize = head->b_size;
1073         int err, ret = 0;
1074         struct buffer_head *next;
1075
1076         for (bh = head, block_start = 0;
1077              ret == 0 && (bh != head || !block_start);
1078              block_start = block_end, bh = next) {
1079                 next = bh->b_this_page;
1080                 block_end = block_start + blocksize;
1081                 if (block_end <= from || block_start >= to) {
1082                         if (partial && !buffer_uptodate(bh))
1083                                 *partial = 1;
1084                         continue;
1085                 }
1086                 err = (*fn)(handle, bh);
1087                 if (!ret)
1088                         ret = err;
1089         }
1090         return ret;
1091 }
1092
1093 /*
1094  * To preserve ordering, it is essential that the hole instantiation and
1095  * the data write be encapsulated in a single transaction.  We cannot
1096  * close off a transaction and start a new one between the ext4_get_block()
1097  * and the commit_write().  So doing the jbd2_journal_start at the start of
1098  * prepare_write() is the right place.
1099  *
1100  * Also, this function can nest inside ext4_writepage().  In that case, we
1101  * *know* that ext4_writepage() has generated enough buffer credits to do the
1102  * whole page.  So we won't block on the journal in that case, which is good,
1103  * because the caller may be PF_MEMALLOC.
1104  *
1105  * By accident, ext4 can be reentered when a transaction is open via
1106  * quota file writes.  If we were to commit the transaction while thus
1107  * reentered, there can be a deadlock - we would be holding a quota
1108  * lock, and the commit would never complete if another thread had a
1109  * transaction open and was blocking on the quota lock - a ranking
1110  * violation.
1111  *
1112  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1113  * will _not_ run commit under these circumstances because handle->h_ref
1114  * is elevated.  We'll still have enough credits for the tiny quotafile
1115  * write.
1116  */
1117 int do_journal_get_write_access(handle_t *handle,
1118                                 struct buffer_head *bh)
1119 {
1120         int dirty = buffer_dirty(bh);
1121         int ret;
1122
1123         if (!buffer_mapped(bh) || buffer_freed(bh))
1124                 return 0;
1125         /*
1126          * __block_write_begin() could have dirtied some buffers. Clean
1127          * the dirty bit as jbd2_journal_get_write_access() could complain
1128          * otherwise about fs integrity issues. Setting of the dirty bit
1129          * by __block_write_begin() isn't a real problem here as we clear
1130          * the bit before releasing a page lock and thus writeback cannot
1131          * ever write the buffer.
1132          */
1133         if (dirty)
1134                 clear_buffer_dirty(bh);
1135         BUFFER_TRACE(bh, "get write access");
1136         ret = ext4_journal_get_write_access(handle, bh);
1137         if (!ret && dirty)
1138                 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1139         return ret;
1140 }
1141
1142 #ifdef CONFIG_EXT4_FS_ENCRYPTION
1143 static int ext4_block_write_begin(struct page *page, loff_t pos, unsigned len,
1144                                   get_block_t *get_block)
1145 {
1146         unsigned from = pos & (PAGE_SIZE - 1);
1147         unsigned to = from + len;
1148         struct inode *inode = page->mapping->host;
1149         unsigned block_start, block_end;
1150         sector_t block;
1151         int err = 0;
1152         unsigned blocksize = inode->i_sb->s_blocksize;
1153         unsigned bbits;
1154         struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
1155         bool decrypt = false;
1156
1157         BUG_ON(!PageLocked(page));
1158         BUG_ON(from > PAGE_SIZE);
1159         BUG_ON(to > PAGE_SIZE);
1160         BUG_ON(from > to);
1161
1162         if (!page_has_buffers(page))
1163                 create_empty_buffers(page, blocksize, 0);
1164         head = page_buffers(page);
1165         bbits = ilog2(blocksize);
1166         block = (sector_t)page->index << (PAGE_SHIFT - bbits);
1167
1168         for (bh = head, block_start = 0; bh != head || !block_start;
1169             block++, block_start = block_end, bh = bh->b_this_page) {
1170                 block_end = block_start + blocksize;
1171                 if (block_end <= from || block_start >= to) {
1172                         if (PageUptodate(page)) {
1173                                 if (!buffer_uptodate(bh))
1174                                         set_buffer_uptodate(bh);
1175                         }
1176                         continue;
1177                 }
1178                 if (buffer_new(bh))
1179                         clear_buffer_new(bh);
1180                 if (!buffer_mapped(bh)) {
1181                         WARN_ON(bh->b_size != blocksize);
1182                         err = get_block(inode, block, bh, 1);
1183                         if (err)
1184                                 break;
1185                         if (buffer_new(bh)) {
1186                                 clean_bdev_bh_alias(bh);
1187                                 if (PageUptodate(page)) {
1188                                         clear_buffer_new(bh);
1189                                         set_buffer_uptodate(bh);
1190                                         mark_buffer_dirty(bh);
1191                                         continue;
1192                                 }
1193                                 if (block_end > to || block_start < from)
1194                                         zero_user_segments(page, to, block_end,
1195                                                            block_start, from);
1196                                 continue;
1197                         }
1198                 }
1199                 if (PageUptodate(page)) {
1200                         if (!buffer_uptodate(bh))
1201                                 set_buffer_uptodate(bh);
1202                         continue;
1203                 }
1204                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1205                     !buffer_unwritten(bh) &&
1206                     (block_start < from || block_end > to)) {
1207                         ll_rw_block(REQ_OP_READ, 0, 1, &bh);
1208                         *wait_bh++ = bh;
1209                         decrypt = ext4_encrypted_inode(inode) &&
1210                                 S_ISREG(inode->i_mode);
1211                 }
1212         }
1213         /*
1214          * If we issued read requests, let them complete.
1215          */
1216         while (wait_bh > wait) {
1217                 wait_on_buffer(*--wait_bh);
1218                 if (!buffer_uptodate(*wait_bh))
1219                         err = -EIO;
1220         }
1221         if (unlikely(err))
1222                 page_zero_new_buffers(page, from, to);
1223         else if (decrypt)
1224                 err = fscrypt_decrypt_page(page->mapping->host, page,
1225                                 PAGE_SIZE, 0, page->index);
1226         return err;
1227 }
1228 #endif
1229
1230 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1231                             loff_t pos, unsigned len, unsigned flags,
1232                             struct page **pagep, void **fsdata)
1233 {
1234         struct inode *inode = mapping->host;
1235         int ret, needed_blocks;
1236         handle_t *handle;
1237         int retries = 0;
1238         struct page *page;
1239         pgoff_t index;
1240         unsigned from, to;
1241
1242         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
1243                 return -EIO;
1244
1245         trace_ext4_write_begin(inode, pos, len, flags);
1246         /*
1247          * Reserve one block more for addition to orphan list in case
1248          * we allocate blocks but write fails for some reason
1249          */
1250         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1251         index = pos >> PAGE_SHIFT;
1252         from = pos & (PAGE_SIZE - 1);
1253         to = from + len;
1254
1255         if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
1256                 ret = ext4_try_to_write_inline_data(mapping, inode, pos, len,
1257                                                     flags, pagep);
1258                 if (ret < 0)
1259                         return ret;
1260                 if (ret == 1)
1261                         return 0;
1262         }
1263
1264         /*
1265          * grab_cache_page_write_begin() can take a long time if the
1266          * system is thrashing due to memory pressure, or if the page
1267          * is being written back.  So grab it first before we start
1268          * the transaction handle.  This also allows us to allocate
1269          * the page (if needed) without using GFP_NOFS.
1270          */
1271 retry_grab:
1272         page = grab_cache_page_write_begin(mapping, index, flags);
1273         if (!page)
1274                 return -ENOMEM;
1275         unlock_page(page);
1276
1277 retry_journal:
1278         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE, needed_blocks);
1279         if (IS_ERR(handle)) {
1280                 put_page(page);
1281                 return PTR_ERR(handle);
1282         }
1283
1284         lock_page(page);
1285         if (page->mapping != mapping) {
1286                 /* The page got truncated from under us */
1287                 unlock_page(page);
1288                 put_page(page);
1289                 ext4_journal_stop(handle);
1290                 goto retry_grab;
1291         }
1292         /* In case writeback began while the page was unlocked */
1293         wait_for_stable_page(page);
1294
1295 #ifdef CONFIG_EXT4_FS_ENCRYPTION
1296         if (ext4_should_dioread_nolock(inode))
1297                 ret = ext4_block_write_begin(page, pos, len,
1298                                              ext4_get_block_unwritten);
1299         else
1300                 ret = ext4_block_write_begin(page, pos, len,
1301                                              ext4_get_block);
1302 #else
1303         if (ext4_should_dioread_nolock(inode))
1304                 ret = __block_write_begin(page, pos, len,
1305                                           ext4_get_block_unwritten);
1306         else
1307                 ret = __block_write_begin(page, pos, len, ext4_get_block);
1308 #endif
1309         if (!ret && ext4_should_journal_data(inode)) {
1310                 ret = ext4_walk_page_buffers(handle, page_buffers(page),
1311                                              from, to, NULL,
1312                                              do_journal_get_write_access);
1313         }
1314
1315         if (ret) {
1316                 unlock_page(page);
1317                 /*
1318                  * __block_write_begin may have instantiated a few blocks
1319                  * outside i_size.  Trim these off again. Don't need
1320                  * i_size_read because we hold i_mutex.
1321                  *
1322                  * Add inode to orphan list in case we crash before
1323                  * truncate finishes
1324                  */
1325                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1326                         ext4_orphan_add(handle, inode);
1327
1328                 ext4_journal_stop(handle);
1329                 if (pos + len > inode->i_size) {
1330                         ext4_truncate_failed_write(inode);
1331                         /*
1332                          * If truncate failed early the inode might
1333                          * still be on the orphan list; we need to
1334                          * make sure the inode is removed from the
1335                          * orphan list in that case.
1336                          */
1337                         if (inode->i_nlink)
1338                                 ext4_orphan_del(NULL, inode);
1339                 }
1340
1341                 if (ret == -ENOSPC &&
1342                     ext4_should_retry_alloc(inode->i_sb, &retries))
1343                         goto retry_journal;
1344                 put_page(page);
1345                 return ret;
1346         }
1347         *pagep = page;
1348         return ret;
1349 }
1350
1351 /* For write_end() in data=journal mode */
1352 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1353 {
1354         int ret;
1355         if (!buffer_mapped(bh) || buffer_freed(bh))
1356                 return 0;
1357         set_buffer_uptodate(bh);
1358         ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1359         clear_buffer_meta(bh);
1360         clear_buffer_prio(bh);
1361         return ret;
1362 }
1363
1364 /*
1365  * We need to pick up the new inode size which generic_commit_write gave us
1366  * `file' can be NULL - eg, when called from page_symlink().
1367  *
1368  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1369  * buffers are managed internally.
1370  */
1371 static int ext4_write_end(struct file *file,
1372                           struct address_space *mapping,
1373                           loff_t pos, unsigned len, unsigned copied,
1374                           struct page *page, void *fsdata)
1375 {
1376         handle_t *handle = ext4_journal_current_handle();
1377         struct inode *inode = mapping->host;
1378         loff_t old_size = inode->i_size;
1379         int ret = 0, ret2;
1380         int i_size_changed = 0;
1381
1382         trace_ext4_write_end(inode, pos, len, copied);
1383         if (ext4_has_inline_data(inode)) {
1384                 ret = ext4_write_inline_data_end(inode, pos, len,
1385                                                  copied, page);
1386                 if (ret < 0) {
1387                         unlock_page(page);
1388                         put_page(page);
1389                         goto errout;
1390                 }
1391                 copied = ret;
1392         } else
1393                 copied = block_write_end(file, mapping, pos,
1394                                          len, copied, page, fsdata);
1395         /*
1396          * it's important to update i_size while still holding page lock:
1397          * page writeout could otherwise come in and zero beyond i_size.
1398          */
1399         i_size_changed = ext4_update_inode_size(inode, pos + copied);
1400         unlock_page(page);
1401         put_page(page);
1402
1403         if (old_size < pos)
1404                 pagecache_isize_extended(inode, old_size, pos);
1405         /*
1406          * Don't mark the inode dirty under page lock. First, it unnecessarily
1407          * makes the holding time of page lock longer. Second, it forces lock
1408          * ordering of page lock and transaction start for journaling
1409          * filesystems.
1410          */
1411         if (i_size_changed)
1412                 ext4_mark_inode_dirty(handle, inode);
1413
1414         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1415                 /* if we have allocated more blocks and copied
1416                  * less. We will have blocks allocated outside
1417                  * inode->i_size. So truncate them
1418                  */
1419                 ext4_orphan_add(handle, inode);
1420 errout:
1421         ret2 = ext4_journal_stop(handle);
1422         if (!ret)
1423                 ret = ret2;
1424
1425         if (pos + len > inode->i_size) {
1426                 ext4_truncate_failed_write(inode);
1427                 /*
1428                  * If truncate failed early the inode might still be
1429                  * on the orphan list; we need to make sure the inode
1430                  * is removed from the orphan list in that case.
1431                  */
1432                 if (inode->i_nlink)
1433                         ext4_orphan_del(NULL, inode);
1434         }
1435
1436         return ret ? ret : copied;
1437 }
1438
1439 /*
1440  * This is a private version of page_zero_new_buffers() which doesn't
1441  * set the buffer to be dirty, since in data=journalled mode we need
1442  * to call ext4_handle_dirty_metadata() instead.
1443  */
1444 static void ext4_journalled_zero_new_buffers(handle_t *handle,
1445                                             struct page *page,
1446                                             unsigned from, unsigned to)
1447 {
1448         unsigned int block_start = 0, block_end;
1449         struct buffer_head *head, *bh;
1450
1451         bh = head = page_buffers(page);
1452         do {
1453                 block_end = block_start + bh->b_size;
1454                 if (buffer_new(bh)) {
1455                         if (block_end > from && block_start < to) {
1456                                 if (!PageUptodate(page)) {
1457                                         unsigned start, size;
1458
1459                                         start = max(from, block_start);
1460                                         size = min(to, block_end) - start;
1461
1462                                         zero_user(page, start, size);
1463                                         write_end_fn(handle, bh);
1464                                 }
1465                                 clear_buffer_new(bh);
1466                         }
1467                 }
1468                 block_start = block_end;
1469                 bh = bh->b_this_page;
1470         } while (bh != head);
1471 }
1472
1473 static int ext4_journalled_write_end(struct file *file,
1474                                      struct address_space *mapping,
1475                                      loff_t pos, unsigned len, unsigned copied,
1476                                      struct page *page, void *fsdata)
1477 {
1478         handle_t *handle = ext4_journal_current_handle();
1479         struct inode *inode = mapping->host;
1480         loff_t old_size = inode->i_size;
1481         int ret = 0, ret2;
1482         int partial = 0;
1483         unsigned from, to;
1484         int size_changed = 0;
1485
1486         trace_ext4_journalled_write_end(inode, pos, len, copied);
1487         from = pos & (PAGE_SIZE - 1);
1488         to = from + len;
1489
1490         BUG_ON(!ext4_handle_valid(handle));
1491
1492         if (ext4_has_inline_data(inode)) {
1493                 ret = ext4_write_inline_data_end(inode, pos, len,
1494                                                  copied, page);
1495                 if (ret < 0) {
1496                         unlock_page(page);
1497                         put_page(page);
1498                         goto errout;
1499                 }
1500                 copied = ret;
1501         } else if (unlikely(copied < len) && !PageUptodate(page)) {
1502                 copied = 0;
1503                 ext4_journalled_zero_new_buffers(handle, page, from, to);
1504         } else {
1505                 if (unlikely(copied < len))
1506                         ext4_journalled_zero_new_buffers(handle, page,
1507                                                          from + copied, to);
1508                 ret = ext4_walk_page_buffers(handle, page_buffers(page), from,
1509                                              from + copied, &partial,
1510                                              write_end_fn);
1511                 if (!partial)
1512                         SetPageUptodate(page);
1513         }
1514         size_changed = ext4_update_inode_size(inode, pos + copied);
1515         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1516         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1517         unlock_page(page);
1518         put_page(page);
1519
1520         if (old_size < pos)
1521                 pagecache_isize_extended(inode, old_size, pos);
1522
1523         if (size_changed) {
1524                 ret2 = ext4_mark_inode_dirty(handle, inode);
1525                 if (!ret)
1526                         ret = ret2;
1527         }
1528
1529         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1530                 /* if we have allocated more blocks and copied
1531                  * less. We will have blocks allocated outside
1532                  * inode->i_size. So truncate them
1533                  */
1534                 ext4_orphan_add(handle, inode);
1535
1536 errout:
1537         ret2 = ext4_journal_stop(handle);
1538         if (!ret)
1539                 ret = ret2;
1540         if (pos + len > inode->i_size) {
1541                 ext4_truncate_failed_write(inode);
1542                 /*
1543                  * If truncate failed early the inode might still be
1544                  * on the orphan list; we need to make sure the inode
1545                  * is removed from the orphan list in that case.
1546                  */
1547                 if (inode->i_nlink)
1548                         ext4_orphan_del(NULL, inode);
1549         }
1550
1551         return ret ? ret : copied;
1552 }
1553
1554 /*
1555  * Reserve space for a single cluster
1556  */
1557 static int ext4_da_reserve_space(struct inode *inode)
1558 {
1559         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1560         struct ext4_inode_info *ei = EXT4_I(inode);
1561         int ret;
1562
1563         /*
1564          * We will charge metadata quota at writeout time; this saves
1565          * us from metadata over-estimation, though we may go over by
1566          * a small amount in the end.  Here we just reserve for data.
1567          */
1568         ret = dquot_reserve_block(inode, EXT4_C2B(sbi, 1));
1569         if (ret)
1570                 return ret;
1571
1572         spin_lock(&ei->i_block_reservation_lock);
1573         if (ext4_claim_free_clusters(sbi, 1, 0)) {
1574                 spin_unlock(&ei->i_block_reservation_lock);
1575                 dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
1576                 return -ENOSPC;
1577         }
1578         ei->i_reserved_data_blocks++;
1579         trace_ext4_da_reserve_space(inode);
1580         spin_unlock(&ei->i_block_reservation_lock);
1581
1582         return 0;       /* success */
1583 }
1584
1585 static void ext4_da_release_space(struct inode *inode, int to_free)
1586 {
1587         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1588         struct ext4_inode_info *ei = EXT4_I(inode);
1589
1590         if (!to_free)
1591                 return;         /* Nothing to release, exit */
1592
1593         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1594
1595         trace_ext4_da_release_space(inode, to_free);
1596         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1597                 /*
1598                  * if there aren't enough reserved blocks, then the
1599                  * counter is messed up somewhere.  Since this
1600                  * function is called from invalidate page, it's
1601                  * harmless to return without any action.
1602                  */
1603                 ext4_warning(inode->i_sb, "ext4_da_release_space: "
1604                          "ino %lu, to_free %d with only %d reserved "
1605                          "data blocks", inode->i_ino, to_free,
1606                          ei->i_reserved_data_blocks);
1607                 WARN_ON(1);
1608                 to_free = ei->i_reserved_data_blocks;
1609         }
1610         ei->i_reserved_data_blocks -= to_free;
1611
1612         /* update fs dirty data blocks counter */
1613         percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1614
1615         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1616
1617         dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1618 }
1619
1620 static void ext4_da_page_release_reservation(struct page *page,
1621                                              unsigned int offset,
1622                                              unsigned int length)
1623 {
1624         int to_release = 0, contiguous_blks = 0;
1625         struct buffer_head *head, *bh;
1626         unsigned int curr_off = 0;
1627         struct inode *inode = page->mapping->host;
1628         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1629         unsigned int stop = offset + length;
1630         int num_clusters;
1631         ext4_fsblk_t lblk;
1632
1633         BUG_ON(stop > PAGE_SIZE || stop < length);
1634
1635         head = page_buffers(page);
1636         bh = head;
1637         do {
1638                 unsigned int next_off = curr_off + bh->b_size;
1639
1640                 if (next_off > stop)
1641                         break;
1642
1643                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1644                         to_release++;
1645                         contiguous_blks++;
1646                         clear_buffer_delay(bh);
1647                 } else if (contiguous_blks) {
1648                         lblk = page->index <<
1649                                (PAGE_SHIFT - inode->i_blkbits);
1650                         lblk += (curr_off >> inode->i_blkbits) -
1651                                 contiguous_blks;
1652                         ext4_es_remove_extent(inode, lblk, contiguous_blks);
1653                         contiguous_blks = 0;
1654                 }
1655                 curr_off = next_off;
1656         } while ((bh = bh->b_this_page) != head);
1657
1658         if (contiguous_blks) {
1659                 lblk = page->index << (PAGE_SHIFT - inode->i_blkbits);
1660                 lblk += (curr_off >> inode->i_blkbits) - contiguous_blks;
1661                 ext4_es_remove_extent(inode, lblk, contiguous_blks);
1662         }
1663
1664         /* If we have released all the blocks belonging to a cluster, then we
1665          * need to release the reserved space for that cluster. */
1666         num_clusters = EXT4_NUM_B2C(sbi, to_release);
1667         while (num_clusters > 0) {
1668                 lblk = (page->index << (PAGE_SHIFT - inode->i_blkbits)) +
1669                         ((num_clusters - 1) << sbi->s_cluster_bits);
1670                 if (sbi->s_cluster_ratio == 1 ||
1671                     !ext4_find_delalloc_cluster(inode, lblk))
1672                         ext4_da_release_space(inode, 1);
1673
1674                 num_clusters--;
1675         }
1676 }
1677
1678 /*
1679  * Delayed allocation stuff
1680  */
1681
1682 struct mpage_da_data {
1683         struct inode *inode;
1684         struct writeback_control *wbc;
1685
1686         pgoff_t first_page;     /* The first page to write */
1687         pgoff_t next_page;      /* Current page to examine */
1688         pgoff_t last_page;      /* Last page to examine */
1689         /*
1690          * Extent to map - this can be after first_page because that can be
1691          * fully mapped. We somewhat abuse m_flags to store whether the extent
1692          * is delalloc or unwritten.
1693          */
1694         struct ext4_map_blocks map;
1695         struct ext4_io_submit io_submit;        /* IO submission data */
1696         unsigned int do_map:1;
1697 };
1698
1699 static void mpage_release_unused_pages(struct mpage_da_data *mpd,
1700                                        bool invalidate)
1701 {
1702         int nr_pages, i;
1703         pgoff_t index, end;
1704         struct pagevec pvec;
1705         struct inode *inode = mpd->inode;
1706         struct address_space *mapping = inode->i_mapping;
1707
1708         /* This is necessary when next_page == 0. */
1709         if (mpd->first_page >= mpd->next_page)
1710                 return;
1711
1712         index = mpd->first_page;
1713         end   = mpd->next_page - 1;
1714         if (invalidate) {
1715                 ext4_lblk_t start, last;
1716                 start = index << (PAGE_SHIFT - inode->i_blkbits);
1717                 last = end << (PAGE_SHIFT - inode->i_blkbits);
1718                 ext4_es_remove_extent(inode, start, last - start + 1);
1719         }
1720
1721         pagevec_init(&pvec, 0);
1722         while (index <= end) {
1723                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1724                 if (nr_pages == 0)
1725                         break;
1726                 for (i = 0; i < nr_pages; i++) {
1727                         struct page *page = pvec.pages[i];
1728                         if (page->index > end)
1729                                 break;
1730                         BUG_ON(!PageLocked(page));
1731                         BUG_ON(PageWriteback(page));
1732                         if (invalidate) {
1733                                 if (page_mapped(page))
1734                                         clear_page_dirty_for_io(page);
1735                                 block_invalidatepage(page, 0, PAGE_SIZE);
1736                                 ClearPageUptodate(page);
1737                         }
1738                         unlock_page(page);
1739                 }
1740                 index = pvec.pages[nr_pages - 1]->index + 1;
1741                 pagevec_release(&pvec);
1742         }
1743 }
1744
1745 static void ext4_print_free_blocks(struct inode *inode)
1746 {
1747         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1748         struct super_block *sb = inode->i_sb;
1749         struct ext4_inode_info *ei = EXT4_I(inode);
1750
1751         ext4_msg(sb, KERN_CRIT, "Total free blocks count %lld",
1752                EXT4_C2B(EXT4_SB(inode->i_sb),
1753                         ext4_count_free_clusters(sb)));
1754         ext4_msg(sb, KERN_CRIT, "Free/Dirty block details");
1755         ext4_msg(sb, KERN_CRIT, "free_blocks=%lld",
1756                (long long) EXT4_C2B(EXT4_SB(sb),
1757                 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1758         ext4_msg(sb, KERN_CRIT, "dirty_blocks=%lld",
1759                (long long) EXT4_C2B(EXT4_SB(sb),
1760                 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1761         ext4_msg(sb, KERN_CRIT, "Block reservation details");
1762         ext4_msg(sb, KERN_CRIT, "i_reserved_data_blocks=%u",
1763                  ei->i_reserved_data_blocks);
1764         return;
1765 }
1766
1767 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1768 {
1769         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1770 }
1771
1772 /*
1773  * This function is grabs code from the very beginning of
1774  * ext4_map_blocks, but assumes that the caller is from delayed write
1775  * time. This function looks up the requested blocks and sets the
1776  * buffer delay bit under the protection of i_data_sem.
1777  */
1778 static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1779                               struct ext4_map_blocks *map,
1780                               struct buffer_head *bh)
1781 {
1782         struct extent_status es;
1783         int retval;
1784         sector_t invalid_block = ~((sector_t) 0xffff);
1785 #ifdef ES_AGGRESSIVE_TEST
1786         struct ext4_map_blocks orig_map;
1787
1788         memcpy(&orig_map, map, sizeof(*map));
1789 #endif
1790
1791         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1792                 invalid_block = ~0;
1793
1794         map->m_flags = 0;
1795         ext_debug("ext4_da_map_blocks(): inode %lu, max_blocks %u,"
1796                   "logical block %lu\n", inode->i_ino, map->m_len,
1797                   (unsigned long) map->m_lblk);
1798
1799         /* Lookup extent status tree firstly */
1800         if (ext4_es_lookup_extent(inode, iblock, &es)) {
1801                 if (ext4_es_is_hole(&es)) {
1802                         retval = 0;
1803                         down_read(&EXT4_I(inode)->i_data_sem);
1804                         goto add_delayed;
1805                 }
1806
1807                 /*
1808                  * Delayed extent could be allocated by fallocate.
1809                  * So we need to check it.
1810                  */
1811                 if (ext4_es_is_delayed(&es) && !ext4_es_is_unwritten(&es)) {
1812                         map_bh(bh, inode->i_sb, invalid_block);
1813                         set_buffer_new(bh);
1814                         set_buffer_delay(bh);
1815                         return 0;
1816                 }
1817
1818                 map->m_pblk = ext4_es_pblock(&es) + iblock - es.es_lblk;
1819                 retval = es.es_len - (iblock - es.es_lblk);
1820                 if (retval > map->m_len)
1821                         retval = map->m_len;
1822                 map->m_len = retval;
1823                 if (ext4_es_is_written(&es))
1824                         map->m_flags |= EXT4_MAP_MAPPED;
1825                 else if (ext4_es_is_unwritten(&es))
1826                         map->m_flags |= EXT4_MAP_UNWRITTEN;
1827                 else
1828                         BUG_ON(1);
1829
1830 #ifdef ES_AGGRESSIVE_TEST
1831                 ext4_map_blocks_es_recheck(NULL, inode, map, &orig_map, 0);
1832 #endif
1833                 return retval;
1834         }
1835
1836         /*
1837          * Try to see if we can get the block without requesting a new
1838          * file system block.
1839          */
1840         down_read(&EXT4_I(inode)->i_data_sem);
1841         if (ext4_has_inline_data(inode))
1842                 retval = 0;
1843         else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1844                 retval = ext4_ext_map_blocks(NULL, inode, map, 0);
1845         else
1846                 retval = ext4_ind_map_blocks(NULL, inode, map, 0);
1847
1848 add_delayed:
1849         if (retval == 0) {
1850                 int ret;
1851                 /*
1852                  * XXX: __block_prepare_write() unmaps passed block,
1853                  * is it OK?
1854                  */
1855                 /*
1856                  * If the block was allocated from previously allocated cluster,
1857                  * then we don't need to reserve it again. However we still need
1858                  * to reserve metadata for every block we're going to write.
1859                  */
1860                 if (EXT4_SB(inode->i_sb)->s_cluster_ratio == 1 ||
1861                     !ext4_find_delalloc_cluster(inode, map->m_lblk)) {
1862                         ret = ext4_da_reserve_space(inode);
1863                         if (ret) {
1864                                 /* not enough space to reserve */
1865                                 retval = ret;
1866                                 goto out_unlock;
1867                         }
1868                 }
1869
1870                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1871                                             ~0, EXTENT_STATUS_DELAYED);
1872                 if (ret) {
1873                         retval = ret;
1874                         goto out_unlock;
1875                 }
1876
1877                 map_bh(bh, inode->i_sb, invalid_block);
1878                 set_buffer_new(bh);
1879                 set_buffer_delay(bh);
1880         } else if (retval > 0) {
1881                 int ret;
1882                 unsigned int status;
1883
1884                 if (unlikely(retval != map->m_len)) {
1885                         ext4_warning(inode->i_sb,
1886                                      "ES len assertion failed for inode "
1887                                      "%lu: retval %d != map->m_len %d",
1888                                      inode->i_ino, retval, map->m_len);
1889                         WARN_ON(1);
1890                 }
1891
1892                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
1893                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
1894                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1895                                             map->m_pblk, status);
1896                 if (ret != 0)
1897                         retval = ret;
1898         }
1899
1900 out_unlock:
1901         up_read((&EXT4_I(inode)->i_data_sem));
1902
1903         return retval;
1904 }
1905
1906 /*
1907  * This is a special get_block_t callback which is used by
1908  * ext4_da_write_begin().  It will either return mapped block or
1909  * reserve space for a single block.
1910  *
1911  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1912  * We also have b_blocknr = -1 and b_bdev initialized properly
1913  *
1914  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1915  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1916  * initialized properly.
1917  */
1918 int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1919                            struct buffer_head *bh, int create)
1920 {
1921         struct ext4_map_blocks map;
1922         int ret = 0;
1923
1924         BUG_ON(create == 0);
1925         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1926
1927         map.m_lblk = iblock;
1928         map.m_len = 1;
1929
1930         /*
1931          * first, we need to know whether the block is allocated already
1932          * preallocated blocks are unmapped but should treated
1933          * the same as allocated blocks.
1934          */
1935         ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1936         if (ret <= 0)
1937                 return ret;
1938
1939         map_bh(bh, inode->i_sb, map.m_pblk);
1940         ext4_update_bh_state(bh, map.m_flags);
1941
1942         if (buffer_unwritten(bh)) {
1943                 /* A delayed write to unwritten bh should be marked
1944                  * new and mapped.  Mapped ensures that we don't do
1945                  * get_block multiple times when we write to the same
1946                  * offset and new ensures that we do proper zero out
1947                  * for partial write.
1948                  */
1949                 set_buffer_new(bh);
1950                 set_buffer_mapped(bh);
1951         }
1952         return 0;
1953 }
1954
1955 static int bget_one(handle_t *handle, struct buffer_head *bh)
1956 {
1957         get_bh(bh);
1958         return 0;
1959 }
1960
1961 static int bput_one(handle_t *handle, struct buffer_head *bh)
1962 {
1963         put_bh(bh);
1964         return 0;
1965 }
1966
1967 static int __ext4_journalled_writepage(struct page *page,
1968                                        unsigned int len)
1969 {
1970         struct address_space *mapping = page->mapping;
1971         struct inode *inode = mapping->host;
1972         struct buffer_head *page_bufs = NULL;
1973         handle_t *handle = NULL;
1974         int ret = 0, err = 0;
1975         int inline_data = ext4_has_inline_data(inode);
1976         struct buffer_head *inode_bh = NULL;
1977
1978         ClearPageChecked(page);
1979
1980         if (inline_data) {
1981                 BUG_ON(page->index != 0);
1982                 BUG_ON(len > ext4_get_max_inline_size(inode));
1983                 inode_bh = ext4_journalled_write_inline_data(inode, len, page);
1984                 if (inode_bh == NULL)
1985                         goto out;
1986         } else {
1987                 page_bufs = page_buffers(page);
1988                 if (!page_bufs) {
1989                         BUG();
1990                         goto out;
1991                 }
1992                 ext4_walk_page_buffers(handle, page_bufs, 0, len,
1993                                        NULL, bget_one);
1994         }
1995         /*
1996          * We need to release the page lock before we start the
1997          * journal, so grab a reference so the page won't disappear
1998          * out from under us.
1999          */
2000         get_page(page);
2001         unlock_page(page);
2002
2003         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
2004                                     ext4_writepage_trans_blocks(inode));
2005         if (IS_ERR(handle)) {
2006                 ret = PTR_ERR(handle);
2007                 put_page(page);
2008                 goto out_no_pagelock;
2009         }
2010         BUG_ON(!ext4_handle_valid(handle));
2011
2012         lock_page(page);
2013         put_page(page);
2014         if (page->mapping != mapping) {
2015                 /* The page got truncated from under us */
2016                 ext4_journal_stop(handle);
2017                 ret = 0;
2018                 goto out;
2019         }
2020
2021         if (inline_data) {
2022                 BUFFER_TRACE(inode_bh, "get write access");
2023                 ret = ext4_journal_get_write_access(handle, inode_bh);
2024
2025                 err = ext4_handle_dirty_metadata(handle, inode, inode_bh);
2026
2027         } else {
2028                 ret = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
2029                                              do_journal_get_write_access);
2030
2031                 err = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
2032                                              write_end_fn);
2033         }
2034         if (ret == 0)
2035                 ret = err;
2036         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
2037         err = ext4_journal_stop(handle);
2038         if (!ret)
2039                 ret = err;
2040
2041         if (!ext4_has_inline_data(inode))
2042                 ext4_walk_page_buffers(NULL, page_bufs, 0, len,
2043                                        NULL, bput_one);
2044         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
2045 out:
2046         unlock_page(page);
2047 out_no_pagelock:
2048         brelse(inode_bh);
2049         return ret;
2050 }
2051
2052 /*
2053  * Note that we don't need to start a transaction unless we're journaling data
2054  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2055  * need to file the inode to the transaction's list in ordered mode because if
2056  * we are writing back data added by write(), the inode is already there and if
2057  * we are writing back data modified via mmap(), no one guarantees in which
2058  * transaction the data will hit the disk. In case we are journaling data, we
2059  * cannot start transaction directly because transaction start ranks above page
2060  * lock so we have to do some magic.
2061  *
2062  * This function can get called via...
2063  *   - ext4_writepages after taking page lock (have journal handle)
2064  *   - journal_submit_inode_data_buffers (no journal handle)
2065  *   - shrink_page_list via the kswapd/direct reclaim (no journal handle)
2066  *   - grab_page_cache when doing write_begin (have journal handle)
2067  *
2068  * We don't do any block allocation in this function. If we have page with
2069  * multiple blocks we need to write those buffer_heads that are mapped. This
2070  * is important for mmaped based write. So if we do with blocksize 1K
2071  * truncate(f, 1024);
2072  * a = mmap(f, 0, 4096);
2073  * a[0] = 'a';
2074  * truncate(f, 4096);
2075  * we have in the page first buffer_head mapped via page_mkwrite call back
2076  * but other buffer_heads would be unmapped but dirty (dirty done via the
2077  * do_wp_page). So writepage should write the first block. If we modify
2078  * the mmap area beyond 1024 we will again get a page_fault and the
2079  * page_mkwrite callback will do the block allocation and mark the
2080  * buffer_heads mapped.
2081  *
2082  * We redirty the page if we have any buffer_heads that is either delay or
2083  * unwritten in the page.
2084  *
2085  * We can get recursively called as show below.
2086  *
2087  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2088  *              ext4_writepage()
2089  *
2090  * But since we don't do any block allocation we should not deadlock.
2091  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2092  */
2093 static int ext4_writepage(struct page *page,
2094                           struct writeback_control *wbc)
2095 {
2096         int ret = 0;
2097         loff_t size;
2098         unsigned int len;
2099         struct buffer_head *page_bufs = NULL;
2100         struct inode *inode = page->mapping->host;
2101         struct ext4_io_submit io_submit;
2102         bool keep_towrite = false;
2103
2104         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb)))) {
2105                 ext4_invalidatepage(page, 0, PAGE_SIZE);
2106                 unlock_page(page);
2107                 return -EIO;
2108         }
2109
2110         trace_ext4_writepage(page);
2111         size = i_size_read(inode);
2112         if (page->index == size >> PAGE_SHIFT)
2113                 len = size & ~PAGE_MASK;
2114         else
2115                 len = PAGE_SIZE;
2116
2117         page_bufs = page_buffers(page);
2118         /*
2119          * We cannot do block allocation or other extent handling in this
2120          * function. If there are buffers needing that, we have to redirty
2121          * the page. But we may reach here when we do a journal commit via
2122          * journal_submit_inode_data_buffers() and in that case we must write
2123          * allocated buffers to achieve data=ordered mode guarantees.
2124          *
2125          * Also, if there is only one buffer per page (the fs block
2126          * size == the page size), if one buffer needs block
2127          * allocation or needs to modify the extent tree to clear the
2128          * unwritten flag, we know that the page can't be written at
2129          * all, so we might as well refuse the write immediately.
2130          * Unfortunately if the block size != page size, we can't as
2131          * easily detect this case using ext4_walk_page_buffers(), but
2132          * for the extremely common case, this is an optimization that
2133          * skips a useless round trip through ext4_bio_write_page().
2134          */
2135         if (ext4_walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2136                                    ext4_bh_delay_or_unwritten)) {
2137                 redirty_page_for_writepage(wbc, page);
2138                 if ((current->flags & PF_MEMALLOC) ||
2139                     (inode->i_sb->s_blocksize == PAGE_SIZE)) {
2140                         /*
2141                          * For memory cleaning there's no point in writing only
2142                          * some buffers. So just bail out. Warn if we came here
2143                          * from direct reclaim.
2144                          */
2145                         WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD))
2146                                                         == PF_MEMALLOC);
2147                         unlock_page(page);
2148                         return 0;
2149                 }
2150                 keep_towrite = true;
2151         }
2152
2153         if (PageChecked(page) && ext4_should_journal_data(inode))
2154                 /*
2155                  * It's mmapped pagecache.  Add buffers and journal it.  There
2156                  * doesn't seem much point in redirtying the page here.
2157                  */
2158                 return __ext4_journalled_writepage(page, len);
2159
2160         ext4_io_submit_init(&io_submit, wbc);
2161         io_submit.io_end = ext4_init_io_end(inode, GFP_NOFS);
2162         if (!io_submit.io_end) {
2163                 redirty_page_for_writepage(wbc, page);
2164                 unlock_page(page);
2165                 return -ENOMEM;
2166         }
2167         ret = ext4_bio_write_page(&io_submit, page, len, wbc, keep_towrite);
2168         ext4_io_submit(&io_submit);
2169         /* Drop io_end reference we got from init */
2170         ext4_put_io_end_defer(io_submit.io_end);
2171         return ret;
2172 }
2173
2174 static int mpage_submit_page(struct mpage_da_data *mpd, struct page *page)
2175 {
2176         int len;
2177         loff_t size;
2178         int err;
2179
2180         BUG_ON(page->index != mpd->first_page);
2181         clear_page_dirty_for_io(page);
2182         /*
2183          * We have to be very careful here!  Nothing protects writeback path
2184          * against i_size changes and the page can be writeably mapped into
2185          * page tables. So an application can be growing i_size and writing
2186          * data through mmap while writeback runs. clear_page_dirty_for_io()
2187          * write-protects our page in page tables and the page cannot get
2188          * written to again until we release page lock. So only after
2189          * clear_page_dirty_for_io() we are safe to sample i_size for
2190          * ext4_bio_write_page() to zero-out tail of the written page. We rely
2191          * on the barrier provided by TestClearPageDirty in
2192          * clear_page_dirty_for_io() to make sure i_size is really sampled only
2193          * after page tables are updated.
2194          */
2195         size = i_size_read(mpd->inode);
2196         if (page->index == size >> PAGE_SHIFT)
2197                 len = size & ~PAGE_MASK;
2198         else
2199                 len = PAGE_SIZE;
2200         err = ext4_bio_write_page(&mpd->io_submit, page, len, mpd->wbc, false);
2201         if (!err)
2202                 mpd->wbc->nr_to_write--;
2203         mpd->first_page++;
2204
2205         return err;
2206 }
2207
2208 #define BH_FLAGS ((1 << BH_Unwritten) | (1 << BH_Delay))
2209
2210 /*
2211  * mballoc gives us at most this number of blocks...
2212  * XXX: That seems to be only a limitation of ext4_mb_normalize_request().
2213  * The rest of mballoc seems to handle chunks up to full group size.
2214  */
2215 #define MAX_WRITEPAGES_EXTENT_LEN 2048
2216
2217 /*
2218  * mpage_add_bh_to_extent - try to add bh to extent of blocks to map
2219  *
2220  * @mpd - extent of blocks
2221  * @lblk - logical number of the block in the file
2222  * @bh - buffer head we want to add to the extent
2223  *
2224  * The function is used to collect contig. blocks in the same state. If the
2225  * buffer doesn't require mapping for writeback and we haven't started the
2226  * extent of buffers to map yet, the function returns 'true' immediately - the
2227  * caller can write the buffer right away. Otherwise the function returns true
2228  * if the block has been added to the extent, false if the block couldn't be
2229  * added.
2230  */
2231 static bool mpage_add_bh_to_extent(struct mpage_da_data *mpd, ext4_lblk_t lblk,
2232                                    struct buffer_head *bh)
2233 {
2234         struct ext4_map_blocks *map = &mpd->map;
2235
2236         /* Buffer that doesn't need mapping for writeback? */
2237         if (!buffer_dirty(bh) || !buffer_mapped(bh) ||
2238             (!buffer_delay(bh) && !buffer_unwritten(bh))) {
2239                 /* So far no extent to map => we write the buffer right away */
2240                 if (map->m_len == 0)
2241                         return true;
2242                 return false;
2243         }
2244
2245         /* First block in the extent? */
2246         if (map->m_len == 0) {
2247                 /* We cannot map unless handle is started... */
2248                 if (!mpd->do_map)
2249                         return false;
2250                 map->m_lblk = lblk;
2251                 map->m_len = 1;
2252                 map->m_flags = bh->b_state & BH_FLAGS;
2253                 return true;
2254         }
2255
2256         /* Don't go larger than mballoc is willing to allocate */
2257         if (map->m_len >= MAX_WRITEPAGES_EXTENT_LEN)
2258                 return false;
2259
2260         /* Can we merge the block to our big extent? */
2261         if (lblk == map->m_lblk + map->m_len &&
2262             (bh->b_state & BH_FLAGS) == map->m_flags) {
2263                 map->m_len++;
2264                 return true;
2265         }
2266         return false;
2267 }
2268
2269 /*
2270  * mpage_process_page_bufs - submit page buffers for IO or add them to extent
2271  *
2272  * @mpd - extent of blocks for mapping
2273  * @head - the first buffer in the page
2274  * @bh - buffer we should start processing from
2275  * @lblk - logical number of the block in the file corresponding to @bh
2276  *
2277  * Walk through page buffers from @bh upto @head (exclusive) and either submit
2278  * the page for IO if all buffers in this page were mapped and there's no
2279  * accumulated extent of buffers to map or add buffers in the page to the
2280  * extent of buffers to map. The function returns 1 if the caller can continue
2281  * by processing the next page, 0 if it should stop adding buffers to the
2282  * extent to map because we cannot extend it anymore. It can also return value
2283  * < 0 in case of error during IO submission.
2284  */
2285 static int mpage_process_page_bufs(struct mpage_da_data *mpd,
2286                                    struct buffer_head *head,
2287                                    struct buffer_head *bh,
2288                                    ext4_lblk_t lblk)
2289 {
2290         struct inode *inode = mpd->inode;
2291         int err;
2292         ext4_lblk_t blocks = (i_size_read(inode) + i_blocksize(inode) - 1)
2293                                                         >> inode->i_blkbits;
2294
2295         do {
2296                 BUG_ON(buffer_locked(bh));
2297
2298                 if (lblk >= blocks || !mpage_add_bh_to_extent(mpd, lblk, bh)) {
2299                         /* Found extent to map? */
2300                         if (mpd->map.m_len)
2301                                 return 0;
2302                         /* Buffer needs mapping and handle is not started? */
2303                         if (!mpd->do_map)
2304                                 return 0;
2305                         /* Everything mapped so far and we hit EOF */
2306                         break;
2307                 }
2308         } while (lblk++, (bh = bh->b_this_page) != head);
2309         /* So far everything mapped? Submit the page for IO. */
2310         if (mpd->map.m_len == 0) {
2311                 err = mpage_submit_page(mpd, head->b_page);
2312                 if (err < 0)
2313                         return err;
2314         }
2315         return lblk < blocks;
2316 }
2317
2318 /*
2319  * mpage_map_buffers - update buffers corresponding to changed extent and
2320  *                     submit fully mapped pages for IO
2321  *
2322  * @mpd - description of extent to map, on return next extent to map
2323  *
2324  * Scan buffers corresponding to changed extent (we expect corresponding pages
2325  * to be already locked) and update buffer state according to new extent state.
2326  * We map delalloc buffers to their physical location, clear unwritten bits,
2327  * and mark buffers as uninit when we perform writes to unwritten extents
2328  * and do extent conversion after IO is finished. If the last page is not fully
2329  * mapped, we update @map to the next extent in the last page that needs
2330  * mapping. Otherwise we submit the page for IO.
2331  */
2332 static int mpage_map_and_submit_buffers(struct mpage_da_data *mpd)
2333 {
2334         struct pagevec pvec;
2335         int nr_pages, i;
2336         struct inode *inode = mpd->inode;
2337         struct buffer_head *head, *bh;
2338         int bpp_bits = PAGE_SHIFT - inode->i_blkbits;
2339         pgoff_t start, end;
2340         ext4_lblk_t lblk;
2341         sector_t pblock;
2342         int err;
2343
2344         start = mpd->map.m_lblk >> bpp_bits;
2345         end = (mpd->map.m_lblk + mpd->map.m_len - 1) >> bpp_bits;
2346         lblk = start << bpp_bits;
2347         pblock = mpd->map.m_pblk;
2348
2349         pagevec_init(&pvec, 0);
2350         while (start <= end) {
2351                 nr_pages = pagevec_lookup(&pvec, inode->i_mapping, start,
2352                                           PAGEVEC_SIZE);
2353                 if (nr_pages == 0)
2354                         break;
2355                 for (i = 0; i < nr_pages; i++) {
2356                         struct page *page = pvec.pages[i];
2357
2358                         if (page->index > end)
2359                                 break;
2360                         /* Up to 'end' pages must be contiguous */
2361                         BUG_ON(page->index != start);
2362                         bh = head = page_buffers(page);
2363                         do {
2364                                 if (lblk < mpd->map.m_lblk)
2365                                         continue;
2366                                 if (lblk >= mpd->map.m_lblk + mpd->map.m_len) {
2367                                         /*
2368                                          * Buffer after end of mapped extent.
2369                                          * Find next buffer in the page to map.
2370                                          */
2371                                         mpd->map.m_len = 0;
2372                                         mpd->map.m_flags = 0;
2373                                         /*
2374                                          * FIXME: If dioread_nolock supports
2375                                          * blocksize < pagesize, we need to make
2376                                          * sure we add size mapped so far to
2377                                          * io_end->size as the following call
2378                                          * can submit the page for IO.
2379                                          */
2380                                         err = mpage_process_page_bufs(mpd, head,
2381                                                                       bh, lblk);
2382                                         pagevec_release(&pvec);
2383                                         if (err > 0)
2384                                                 err = 0;
2385                                         return err;
2386                                 }
2387                                 if (buffer_delay(bh)) {
2388                                         clear_buffer_delay(bh);
2389                                         bh->b_blocknr = pblock++;
2390                                 }
2391                                 clear_buffer_unwritten(bh);
2392                         } while (lblk++, (bh = bh->b_this_page) != head);
2393
2394                         /*
2395                          * FIXME: This is going to break if dioread_nolock
2396                          * supports blocksize < pagesize as we will try to
2397                          * convert potentially unmapped parts of inode.
2398                          */
2399                         mpd->io_submit.io_end->size += PAGE_SIZE;
2400                         /* Page fully mapped - let IO run! */
2401                         err = mpage_submit_page(mpd, page);
2402                         if (err < 0) {
2403                                 pagevec_release(&pvec);
2404                                 return err;
2405                         }
2406                         start++;
2407                 }
2408                 pagevec_release(&pvec);
2409         }
2410         /* Extent fully mapped and matches with page boundary. We are done. */
2411         mpd->map.m_len = 0;
2412         mpd->map.m_flags = 0;
2413         return 0;
2414 }
2415
2416 static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
2417 {
2418         struct inode *inode = mpd->inode;
2419         struct ext4_map_blocks *map = &mpd->map;
2420         int get_blocks_flags;
2421         int err, dioread_nolock;
2422
2423         trace_ext4_da_write_pages_extent(inode, map);
2424         /*
2425          * Call ext4_map_blocks() to allocate any delayed allocation blocks, or
2426          * to convert an unwritten extent to be initialized (in the case
2427          * where we have written into one or more preallocated blocks).  It is
2428          * possible that we're going to need more metadata blocks than
2429          * previously reserved. However we must not fail because we're in
2430          * writeback and there is nothing we can do about it so it might result
2431          * in data loss.  So use reserved blocks to allocate metadata if
2432          * possible.
2433          *
2434          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE if
2435          * the blocks in question are delalloc blocks.  This indicates
2436          * that the blocks and quotas has already been checked when
2437          * the data was copied into the page cache.
2438          */
2439         get_blocks_flags = EXT4_GET_BLOCKS_CREATE |
2440                            EXT4_GET_BLOCKS_METADATA_NOFAIL |
2441                            EXT4_GET_BLOCKS_IO_SUBMIT;
2442         dioread_nolock = ext4_should_dioread_nolock(inode);
2443         if (dioread_nolock)
2444                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2445         if (map->m_flags & (1 << BH_Delay))
2446                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2447
2448         err = ext4_map_blocks(handle, inode, map, get_blocks_flags);
2449         if (err < 0)
2450                 return err;
2451         if (dioread_nolock && (map->m_flags & EXT4_MAP_UNWRITTEN)) {
2452                 if (!mpd->io_submit.io_end->handle &&
2453                     ext4_handle_valid(handle)) {
2454                         mpd->io_submit.io_end->handle = handle->h_rsv_handle;
2455                         handle->h_rsv_handle = NULL;
2456                 }
2457                 ext4_set_io_unwritten_flag(inode, mpd->io_submit.io_end);
2458         }
2459
2460         BUG_ON(map->m_len == 0);
2461         if (map->m_flags & EXT4_MAP_NEW) {
2462                 clean_bdev_aliases(inode->i_sb->s_bdev, map->m_pblk,
2463                                    map->m_len);
2464         }
2465         return 0;
2466 }
2467
2468 /*
2469  * mpage_map_and_submit_extent - map extent starting at mpd->lblk of length
2470  *                               mpd->len and submit pages underlying it for IO
2471  *
2472  * @handle - handle for journal operations
2473  * @mpd - extent to map
2474  * @give_up_on_write - we set this to true iff there is a fatal error and there
2475  *                     is no hope of writing the data. The caller should discard
2476  *                     dirty pages to avoid infinite loops.
2477  *
2478  * The function maps extent starting at mpd->lblk of length mpd->len. If it is
2479  * delayed, blocks are allocated, if it is unwritten, we may need to convert
2480  * them to initialized or split the described range from larger unwritten
2481  * extent. Note that we need not map all the described range since allocation
2482  * can return less blocks or the range is covered by more unwritten extents. We
2483  * cannot map more because we are limited by reserved transaction credits. On
2484  * the other hand we always make sure that the last touched page is fully
2485  * mapped so that it can be written out (and thus forward progress is
2486  * guaranteed). After mapping we submit all mapped pages for IO.
2487  */
2488 static int mpage_map_and_submit_extent(handle_t *handle,
2489                                        struct mpage_da_data *mpd,
2490                                        bool *give_up_on_write)
2491 {
2492         struct inode *inode = mpd->inode;
2493         struct ext4_map_blocks *map = &mpd->map;
2494         int err;
2495         loff_t disksize;
2496         int progress = 0;
2497
2498         mpd->io_submit.io_end->offset =
2499                                 ((loff_t)map->m_lblk) << inode->i_blkbits;
2500         do {
2501                 err = mpage_map_one_extent(handle, mpd);
2502                 if (err < 0) {
2503                         struct super_block *sb = inode->i_sb;
2504
2505                         if (ext4_forced_shutdown(EXT4_SB(sb)) ||
2506                             EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
2507                                 goto invalidate_dirty_pages;
2508                         /*
2509                          * Let the uper layers retry transient errors.
2510                          * In the case of ENOSPC, if ext4_count_free_blocks()
2511                          * is non-zero, a commit should free up blocks.
2512                          */
2513                         if ((err == -ENOMEM) ||
2514                             (err == -ENOSPC && ext4_count_free_clusters(sb))) {
2515                                 if (progress)
2516                                         goto update_disksize;
2517                                 return err;
2518                         }
2519                         ext4_msg(sb, KERN_CRIT,
2520                                  "Delayed block allocation failed for "
2521                                  "inode %lu at logical offset %llu with"
2522                                  " max blocks %u with error %d",
2523                                  inode->i_ino,
2524                                  (unsigned long long)map->m_lblk,
2525                                  (unsigned)map->m_len, -err);
2526                         ext4_msg(sb, KERN_CRIT,
2527                                  "This should not happen!! Data will "
2528                                  "be lost\n");
2529                         if (err == -ENOSPC)
2530                                 ext4_print_free_blocks(inode);
2531                 invalidate_dirty_pages:
2532                         *give_up_on_write = true;
2533                         return err;
2534                 }
2535                 progress = 1;
2536                 /*
2537                  * Update buffer state, submit mapped pages, and get us new
2538                  * extent to map
2539                  */
2540                 err = mpage_map_and_submit_buffers(mpd);
2541                 if (err < 0)
2542                         goto update_disksize;
2543         } while (map->m_len);
2544
2545 update_disksize:
2546         /*
2547          * Update on-disk size after IO is submitted.  Races with
2548          * truncate are avoided by checking i_size under i_data_sem.
2549          */
2550         disksize = ((loff_t)mpd->first_page) << PAGE_SHIFT;
2551         if (disksize > EXT4_I(inode)->i_disksize) {
2552                 int err2;
2553                 loff_t i_size;
2554
2555                 down_write(&EXT4_I(inode)->i_data_sem);
2556                 i_size = i_size_read(inode);
2557                 if (disksize > i_size)
2558                         disksize = i_size;
2559                 if (disksize > EXT4_I(inode)->i_disksize)
2560                         EXT4_I(inode)->i_disksize = disksize;
2561                 up_write(&EXT4_I(inode)->i_data_sem);
2562                 err2 = ext4_mark_inode_dirty(handle, inode);
2563                 if (err2)
2564                         ext4_error(inode->i_sb,
2565                                    "Failed to mark inode %lu dirty",
2566                                    inode->i_ino);
2567                 if (!err)
2568                         err = err2;
2569         }
2570         return err;
2571 }
2572
2573 /*
2574  * Calculate the total number of credits to reserve for one writepages
2575  * iteration. This is called from ext4_writepages(). We map an extent of
2576  * up to MAX_WRITEPAGES_EXTENT_LEN blocks and then we go on and finish mapping
2577  * the last partial page. So in total we can map MAX_WRITEPAGES_EXTENT_LEN +
2578  * bpp - 1 blocks in bpp different extents.
2579  */
2580 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2581 {
2582         int bpp = ext4_journal_blocks_per_page(inode);
2583
2584         return ext4_meta_trans_blocks(inode,
2585                                 MAX_WRITEPAGES_EXTENT_LEN + bpp - 1, bpp);
2586 }
2587
2588 /*
2589  * mpage_prepare_extent_to_map - find & lock contiguous range of dirty pages
2590  *                               and underlying extent to map
2591  *
2592  * @mpd - where to look for pages
2593  *
2594  * Walk dirty pages in the mapping. If they are fully mapped, submit them for
2595  * IO immediately. When we find a page which isn't mapped we start accumulating
2596  * extent of buffers underlying these pages that needs mapping (formed by
2597  * either delayed or unwritten buffers). We also lock the pages containing
2598  * these buffers. The extent found is returned in @mpd structure (starting at
2599  * mpd->lblk with length mpd->len blocks).
2600  *
2601  * Note that this function can attach bios to one io_end structure which are
2602  * neither logically nor physically contiguous. Although it may seem as an
2603  * unnecessary complication, it is actually inevitable in blocksize < pagesize
2604  * case as we need to track IO to all buffers underlying a page in one io_end.
2605  */
2606 static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
2607 {
2608         struct address_space *mapping = mpd->inode->i_mapping;
2609         struct pagevec pvec;
2610         unsigned int nr_pages;
2611         long left = mpd->wbc->nr_to_write;
2612         pgoff_t index = mpd->first_page;
2613         pgoff_t end = mpd->last_page;
2614         int tag;
2615         int i, err = 0;
2616         int blkbits = mpd->inode->i_blkbits;
2617         ext4_lblk_t lblk;
2618         struct buffer_head *head;
2619
2620         if (mpd->wbc->sync_mode == WB_SYNC_ALL || mpd->wbc->tagged_writepages)
2621                 tag = PAGECACHE_TAG_TOWRITE;
2622         else
2623                 tag = PAGECACHE_TAG_DIRTY;
2624
2625         pagevec_init(&pvec, 0);
2626         mpd->map.m_len = 0;
2627         mpd->next_page = index;
2628         while (index <= end) {
2629                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2630                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2631                 if (nr_pages == 0)
2632                         goto out;
2633
2634                 for (i = 0; i < nr_pages; i++) {
2635                         struct page *page = pvec.pages[i];
2636
2637                         /*
2638                          * At this point, the page may be truncated or
2639                          * invalidated (changing page->mapping to NULL), or
2640                          * even swizzled back from swapper_space to tmpfs file
2641                          * mapping. However, page->index will not change
2642                          * because we have a reference on the page.
2643                          */
2644                         if (page->index > end)
2645                                 goto out;
2646
2647                         /*
2648                          * Accumulated enough dirty pages? This doesn't apply
2649                          * to WB_SYNC_ALL mode. For integrity sync we have to
2650                          * keep going because someone may be concurrently
2651                          * dirtying pages, and we might have synced a lot of
2652                          * newly appeared dirty pages, but have not synced all
2653                          * of the old dirty pages.
2654                          */
2655                         if (mpd->wbc->sync_mode == WB_SYNC_NONE && left <= 0)
2656                                 goto out;
2657
2658                         /* If we can't merge this page, we are done. */
2659                         if (mpd->map.m_len > 0 && mpd->next_page != page->index)
2660                                 goto out;
2661
2662                         lock_page(page);
2663                         /*
2664                          * If the page is no longer dirty, or its mapping no
2665                          * longer corresponds to inode we are writing (which
2666                          * means it has been truncated or invalidated), or the
2667                          * page is already under writeback and we are not doing
2668                          * a data integrity writeback, skip the page
2669                          */
2670                         if (!PageDirty(page) ||
2671                             (PageWriteback(page) &&
2672                              (mpd->wbc->sync_mode == WB_SYNC_NONE)) ||
2673                             unlikely(page->mapping != mapping)) {
2674                                 unlock_page(page);
2675                                 continue;
2676                         }
2677
2678                         wait_on_page_writeback(page);
2679                         BUG_ON(PageWriteback(page));
2680
2681                         if (mpd->map.m_len == 0)
2682                                 mpd->first_page = page->index;
2683                         mpd->next_page = page->index + 1;
2684                         /* Add all dirty buffers to mpd */
2685                         lblk = ((ext4_lblk_t)page->index) <<
2686                                 (PAGE_SHIFT - blkbits);
2687                         head = page_buffers(page);
2688                         err = mpage_process_page_bufs(mpd, head, head, lblk);
2689                         if (err <= 0)
2690                                 goto out;
2691                         err = 0;
2692                         left--;
2693                 }
2694                 pagevec_release(&pvec);
2695                 cond_resched();
2696         }
2697         return 0;
2698 out:
2699         pagevec_release(&pvec);
2700         return err;
2701 }
2702
2703 static int __writepage(struct page *page, struct writeback_control *wbc,
2704                        void *data)
2705 {
2706         struct address_space *mapping = data;
2707         int ret = ext4_writepage(page, wbc);
2708         mapping_set_error(mapping, ret);
2709         return ret;
2710 }
2711
2712 static int ext4_writepages(struct address_space *mapping,
2713                            struct writeback_control *wbc)
2714 {
2715         pgoff_t writeback_index = 0;
2716         long nr_to_write = wbc->nr_to_write;
2717         int range_whole = 0;
2718         int cycled = 1;
2719         handle_t *handle = NULL;
2720         struct mpage_da_data mpd;
2721         struct inode *inode = mapping->host;
2722         int needed_blocks, rsv_blocks = 0, ret = 0;
2723         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2724         bool done;
2725         struct blk_plug plug;
2726         bool give_up_on_write = false;
2727
2728         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
2729                 return -EIO;
2730
2731         percpu_down_read(&sbi->s_journal_flag_rwsem);
2732         trace_ext4_writepages(inode, wbc);
2733
2734         if (dax_mapping(mapping)) {
2735                 ret = dax_writeback_mapping_range(mapping, inode->i_sb->s_bdev,
2736                                                   wbc);
2737                 goto out_writepages;
2738         }
2739
2740         /*
2741          * No pages to write? This is mainly a kludge to avoid starting
2742          * a transaction for special inodes like journal inode on last iput()
2743          * because that could violate lock ordering on umount
2744          */
2745         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2746                 goto out_writepages;
2747
2748         if (ext4_should_journal_data(inode)) {
2749                 struct blk_plug plug;
2750
2751                 blk_start_plug(&plug);
2752                 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2753                 blk_finish_plug(&plug);
2754                 goto out_writepages;
2755         }
2756
2757         /*
2758          * If the filesystem has aborted, it is read-only, so return
2759          * right away instead of dumping stack traces later on that
2760          * will obscure the real source of the problem.  We test
2761          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2762          * the latter could be true if the filesystem is mounted
2763          * read-only, and in that case, ext4_writepages should
2764          * *never* be called, so if that ever happens, we would want
2765          * the stack trace.
2766          */
2767         if (unlikely(ext4_forced_shutdown(EXT4_SB(mapping->host->i_sb)) ||
2768                      sbi->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2769                 ret = -EROFS;
2770                 goto out_writepages;
2771         }
2772
2773         if (ext4_should_dioread_nolock(inode)) {
2774                 /*
2775                  * We may need to convert up to one extent per block in
2776                  * the page and we may dirty the inode.
2777                  */
2778                 rsv_blocks = 1 + (PAGE_SIZE >> inode->i_blkbits);
2779         }
2780
2781         /*
2782          * If we have inline data and arrive here, it means that
2783          * we will soon create the block for the 1st page, so
2784          * we'd better clear the inline data here.
2785          */
2786         if (ext4_has_inline_data(inode)) {
2787                 /* Just inode will be modified... */
2788                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
2789                 if (IS_ERR(handle)) {
2790                         ret = PTR_ERR(handle);
2791                         goto out_writepages;
2792                 }
2793                 BUG_ON(ext4_test_inode_state(inode,
2794                                 EXT4_STATE_MAY_INLINE_DATA));
2795                 ext4_destroy_inline_data(handle, inode);
2796                 ext4_journal_stop(handle);
2797         }
2798
2799         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2800                 range_whole = 1;
2801
2802         if (wbc->range_cyclic) {
2803                 writeback_index = mapping->writeback_index;
2804                 if (writeback_index)
2805                         cycled = 0;
2806                 mpd.first_page = writeback_index;
2807                 mpd.last_page = -1;
2808         } else {
2809                 mpd.first_page = wbc->range_start >> PAGE_SHIFT;
2810                 mpd.last_page = wbc->range_end >> PAGE_SHIFT;
2811         }
2812
2813         mpd.inode = inode;
2814         mpd.wbc = wbc;
2815         ext4_io_submit_init(&mpd.io_submit, wbc);
2816 retry:
2817         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2818                 tag_pages_for_writeback(mapping, mpd.first_page, mpd.last_page);
2819         done = false;
2820         blk_start_plug(&plug);
2821
2822         /*
2823          * First writeback pages that don't need mapping - we can avoid
2824          * starting a transaction unnecessarily and also avoid being blocked
2825          * in the block layer on device congestion while having transaction
2826          * started.
2827          */
2828         mpd.do_map = 0;
2829         mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2830         if (!mpd.io_submit.io_end) {
2831                 ret = -ENOMEM;
2832                 goto unplug;
2833         }
2834         ret = mpage_prepare_extent_to_map(&mpd);
2835         /* Submit prepared bio */
2836         ext4_io_submit(&mpd.io_submit);
2837         ext4_put_io_end_defer(mpd.io_submit.io_end);
2838         mpd.io_submit.io_end = NULL;
2839         /* Unlock pages we didn't use */
2840         mpage_release_unused_pages(&mpd, false);
2841         if (ret < 0)
2842                 goto unplug;
2843
2844         while (!done && mpd.first_page <= mpd.last_page) {
2845                 /* For each extent of pages we use new io_end */
2846                 mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2847                 if (!mpd.io_submit.io_end) {
2848                         ret = -ENOMEM;
2849                         break;
2850                 }
2851
2852                 /*
2853                  * We have two constraints: We find one extent to map and we
2854                  * must always write out whole page (makes a difference when
2855                  * blocksize < pagesize) so that we don't block on IO when we
2856                  * try to write out the rest of the page. Journalled mode is
2857                  * not supported by delalloc.
2858                  */
2859                 BUG_ON(ext4_should_journal_data(inode));
2860                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2861
2862                 /* start a new transaction */
2863                 handle = ext4_journal_start_with_reserve(inode,
2864                                 EXT4_HT_WRITE_PAGE, needed_blocks, rsv_blocks);
2865                 if (IS_ERR(handle)) {
2866                         ret = PTR_ERR(handle);
2867                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2868                                "%ld pages, ino %lu; err %d", __func__,
2869                                 wbc->nr_to_write, inode->i_ino, ret);
2870                         /* Release allocated io_end */
2871                         ext4_put_io_end(mpd.io_submit.io_end);
2872                         mpd.io_submit.io_end = NULL;
2873                         break;
2874                 }
2875                 mpd.do_map = 1;
2876
2877                 trace_ext4_da_write_pages(inode, mpd.first_page, mpd.wbc);
2878                 ret = mpage_prepare_extent_to_map(&mpd);
2879                 if (!ret) {
2880                         if (mpd.map.m_len)
2881                                 ret = mpage_map_and_submit_extent(handle, &mpd,
2882                                         &give_up_on_write);
2883                         else {
2884                                 /*
2885                                  * We scanned the whole range (or exhausted
2886                                  * nr_to_write), submitted what was mapped and
2887                                  * didn't find anything needing mapping. We are
2888                                  * done.
2889                                  */
2890                                 done = true;
2891                         }
2892                 }
2893                 /*
2894                  * Caution: If the handle is synchronous,
2895                  * ext4_journal_stop() can wait for transaction commit
2896                  * to finish which may depend on writeback of pages to
2897                  * complete or on page lock to be released.  In that
2898                  * case, we have to wait until after after we have
2899                  * submitted all the IO, released page locks we hold,
2900                  * and dropped io_end reference (for extent conversion
2901                  * to be able to complete) before stopping the handle.
2902                  */
2903                 if (!ext4_handle_valid(handle) || handle->h_sync == 0) {
2904                         ext4_journal_stop(handle);
2905                         handle = NULL;
2906                         mpd.do_map = 0;
2907                 }
2908                 /* Submit prepared bio */
2909                 ext4_io_submit(&mpd.io_submit);
2910                 /* Unlock pages we didn't use */
2911                 mpage_release_unused_pages(&mpd, give_up_on_write);
2912                 /*
2913                  * Drop our io_end reference we got from init. We have
2914                  * to be careful and use deferred io_end finishing if
2915                  * we are still holding the transaction as we can
2916                  * release the last reference to io_end which may end
2917                  * up doing unwritten extent conversion.
2918                  */
2919                 if (handle) {
2920                         ext4_put_io_end_defer(mpd.io_submit.io_end);
2921                         ext4_journal_stop(handle);
2922                 } else
2923                         ext4_put_io_end(mpd.io_submit.io_end);
2924                 mpd.io_submit.io_end = NULL;
2925
2926                 if (ret == -ENOSPC && sbi->s_journal) {
2927                         /*
2928                          * Commit the transaction which would
2929                          * free blocks released in the transaction
2930                          * and try again
2931                          */
2932                         jbd2_journal_force_commit_nested(sbi->s_journal);
2933                         ret = 0;
2934                         continue;
2935                 }
2936                 /* Fatal error - ENOMEM, EIO... */
2937                 if (ret)
2938                         break;
2939         }
2940 unplug:
2941         blk_finish_plug(&plug);
2942         if (!ret && !cycled && wbc->nr_to_write > 0) {
2943                 cycled = 1;
2944                 mpd.last_page = writeback_index - 1;
2945                 mpd.first_page = 0;
2946                 goto retry;
2947         }
2948
2949         /* Update index */
2950         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2951                 /*
2952                  * Set the writeback_index so that range_cyclic
2953                  * mode will write it back later
2954                  */
2955                 mapping->writeback_index = mpd.first_page;
2956
2957 out_writepages:
2958         trace_ext4_writepages_result(inode, wbc, ret,
2959                                      nr_to_write - wbc->nr_to_write);
2960         percpu_up_read(&sbi->s_journal_flag_rwsem);
2961         return ret;
2962 }
2963
2964 static int ext4_nonda_switch(struct super_block *sb)
2965 {
2966         s64 free_clusters, dirty_clusters;
2967         struct ext4_sb_info *sbi = EXT4_SB(sb);
2968
2969         /*
2970          * switch to non delalloc mode if we are running low
2971          * on free block. The free block accounting via percpu
2972          * counters can get slightly wrong with percpu_counter_batch getting
2973          * accumulated on each CPU without updating global counters
2974          * Delalloc need an accurate free block accounting. So switch
2975          * to non delalloc when we are near to error range.
2976          */
2977         free_clusters =
2978                 percpu_counter_read_positive(&sbi->s_freeclusters_counter);
2979         dirty_clusters =
2980                 percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
2981         /*
2982          * Start pushing delalloc when 1/2 of free blocks are dirty.
2983          */
2984         if (dirty_clusters && (free_clusters < 2 * dirty_clusters))
2985                 try_to_writeback_inodes_sb(sb, WB_REASON_FS_FREE_SPACE);
2986
2987         if (2 * free_clusters < 3 * dirty_clusters ||
2988             free_clusters < (dirty_clusters + EXT4_FREECLUSTERS_WATERMARK)) {
2989                 /*
2990                  * free block count is less than 150% of dirty blocks
2991                  * or free blocks is less than watermark
2992                  */
2993                 return 1;
2994         }
2995         return 0;
2996 }
2997
2998 /* We always reserve for an inode update; the superblock could be there too */
2999 static int ext4_da_write_credits(struct inode *inode, loff_t pos, unsigned len)
3000 {
3001         if (likely(ext4_has_feature_large_file(inode->i_sb)))
3002                 return 1;
3003
3004         if (pos + len <= 0x7fffffffULL)
3005                 return 1;
3006
3007         /* We might need to update the superblock to set LARGE_FILE */
3008         return 2;
3009 }
3010
3011 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3012                                loff_t pos, unsigned len, unsigned flags,
3013                                struct page **pagep, void **fsdata)
3014 {
3015         int ret, retries = 0;
3016         struct page *page;
3017         pgoff_t index;
3018         struct inode *inode = mapping->host;
3019         handle_t *handle;
3020
3021         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
3022                 return -EIO;
3023
3024         index = pos >> PAGE_SHIFT;
3025
3026         if (ext4_nonda_switch(inode->i_sb) ||
3027             S_ISLNK(inode->i_mode)) {
3028                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3029                 return ext4_write_begin(file, mapping, pos,
3030                                         len, flags, pagep, fsdata);
3031         }
3032         *fsdata = (void *)0;
3033         trace_ext4_da_write_begin(inode, pos, len, flags);
3034
3035         if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
3036                 ret = ext4_da_write_inline_data_begin(mapping, inode,
3037                                                       pos, len, flags,
3038                                                       pagep, fsdata);
3039                 if (ret < 0)
3040                         return ret;
3041                 if (ret == 1)
3042                         return 0;
3043         }
3044
3045         /*
3046          * grab_cache_page_write_begin() can take a long time if the
3047          * system is thrashing due to memory pressure, or if the page
3048          * is being written back.  So grab it first before we start
3049          * the transaction handle.  This also allows us to allocate
3050          * the page (if needed) without using GFP_NOFS.
3051          */
3052 retry_grab:
3053         page = grab_cache_page_write_begin(mapping, index, flags);
3054         if (!page)
3055                 return -ENOMEM;
3056         unlock_page(page);
3057
3058         /*
3059          * With delayed allocation, we don't log the i_disksize update
3060          * if there is delayed block allocation. But we still need
3061          * to journalling the i_disksize update if writes to the end
3062          * of file which has an already mapped buffer.
3063          */
3064 retry_journal:
3065         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
3066                                 ext4_da_write_credits(inode, pos, len));
3067         if (IS_ERR(handle)) {
3068                 put_page(page);
3069                 return PTR_ERR(handle);
3070         }
3071
3072         lock_page(page);
3073         if (page->mapping != mapping) {
3074                 /* The page got truncated from under us */
3075                 unlock_page(page);
3076                 put_page(page);
3077                 ext4_journal_stop(handle);
3078                 goto retry_grab;
3079         }
3080         /* In case writeback began while the page was unlocked */
3081         wait_for_stable_page(page);
3082
3083 #ifdef CONFIG_EXT4_FS_ENCRYPTION
3084         ret = ext4_block_write_begin(page, pos, len,
3085                                      ext4_da_get_block_prep);
3086 #else
3087         ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
3088 #endif
3089         if (ret < 0) {
3090                 unlock_page(page);
3091                 ext4_journal_stop(handle);
3092                 /*
3093                  * block_write_begin may have instantiated a few blocks
3094                  * outside i_size.  Trim these off again. Don't need
3095                  * i_size_read because we hold i_mutex.
3096                  */
3097                 if (pos + len > inode->i_size)
3098                         ext4_truncate_failed_write(inode);
3099
3100                 if (ret == -ENOSPC &&
3101                     ext4_should_retry_alloc(inode->i_sb, &retries))
3102                         goto retry_journal;
3103
3104                 put_page(page);
3105                 return ret;
3106         }
3107
3108         *pagep = page;
3109         return ret;
3110 }
3111
3112 /*
3113  * Check if we should update i_disksize
3114  * when write to the end of file but not require block allocation
3115  */
3116 static int ext4_da_should_update_i_disksize(struct page *page,
3117                                             unsigned long offset)
3118 {
3119         struct buffer_head *bh;
3120         struct inode *inode = page->mapping->host;
3121         unsigned int idx;
3122         int i;
3123
3124         bh = page_buffers(page);
3125         idx = offset >> inode->i_blkbits;
3126
3127         for (i = 0; i < idx; i++)
3128                 bh = bh->b_this_page;
3129
3130         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3131                 return 0;
3132         return 1;
3133 }
3134
3135 static int ext4_da_write_end(struct file *file,
3136                              struct address_space *mapping,
3137                              loff_t pos, unsigned len, unsigned copied,
3138                              struct page *page, void *fsdata)
3139 {
3140         struct inode *inode = mapping->host;
3141         int ret = 0, ret2;
3142         handle_t *handle = ext4_journal_current_handle();
3143         loff_t new_i_size;
3144         unsigned long start, end;
3145         int write_mode = (int)(unsigned long)fsdata;
3146
3147         if (write_mode == FALL_BACK_TO_NONDELALLOC)
3148                 return ext4_write_end(file, mapping, pos,
3149                                       len, copied, page, fsdata);
3150
3151         trace_ext4_da_write_end(inode, pos, len, copied);
3152         start = pos & (PAGE_SIZE - 1);
3153         end = start + copied - 1;
3154
3155         /*
3156          * generic_write_end() will run mark_inode_dirty() if i_size
3157          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3158          * into that.
3159          */
3160         new_i_size = pos + copied;
3161         if (copied && new_i_size > EXT4_I(inode)->i_disksize) {
3162                 if (ext4_has_inline_data(inode) ||
3163                     ext4_da_should_update_i_disksize(page, end)) {
3164                         ext4_update_i_disksize(inode, new_i_size);
3165                         /* We need to mark inode dirty even if
3166                          * new_i_size is less that inode->i_size
3167                          * bu greater than i_disksize.(hint delalloc)
3168                          */
3169                         ext4_mark_inode_dirty(handle, inode);
3170                 }
3171         }
3172
3173         if (write_mode != CONVERT_INLINE_DATA &&
3174             ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA) &&
3175             ext4_has_inline_data(inode))
3176                 ret2 = ext4_da_write_inline_data_end(inode, pos, len, copied,
3177                                                      page);
3178         else
3179                 ret2 = generic_write_end(file, mapping, pos, len, copied,
3180                                                         page, fsdata);
3181
3182         copied = ret2;
3183         if (ret2 < 0)
3184                 ret = ret2;
3185         ret2 = ext4_journal_stop(handle);
3186         if (!ret)
3187                 ret = ret2;
3188
3189         return ret ? ret : copied;
3190 }
3191
3192 static void ext4_da_invalidatepage(struct page *page, unsigned int offset,
3193                                    unsigned int length)
3194 {
3195         /*
3196          * Drop reserved blocks
3197          */
3198         BUG_ON(!PageLocked(page));
3199         if (!page_has_buffers(page))
3200                 goto out;
3201
3202         ext4_da_page_release_reservation(page, offset, length);
3203
3204 out:
3205         ext4_invalidatepage(page, offset, length);
3206
3207         return;
3208 }
3209
3210 /*
3211  * Force all delayed allocation blocks to be allocated for a given inode.
3212  */
3213 int ext4_alloc_da_blocks(struct inode *inode)
3214 {
3215         trace_ext4_alloc_da_blocks(inode);
3216
3217         if (!EXT4_I(inode)->i_reserved_data_blocks)
3218                 return 0;
3219
3220         /*
3221          * We do something simple for now.  The filemap_flush() will
3222          * also start triggering a write of the data blocks, which is
3223          * not strictly speaking necessary (and for users of
3224          * laptop_mode, not even desirable).  However, to do otherwise
3225          * would require replicating code paths in:
3226          *
3227          * ext4_writepages() ->
3228          *    write_cache_pages() ---> (via passed in callback function)
3229          *        __mpage_da_writepage() -->
3230          *           mpage_add_bh_to_extent()
3231          *           mpage_da_map_blocks()
3232          *
3233          * The problem is that write_cache_pages(), located in
3234          * mm/page-writeback.c, marks pages clean in preparation for
3235          * doing I/O, which is not desirable if we're not planning on
3236          * doing I/O at all.
3237          *
3238          * We could call write_cache_pages(), and then redirty all of
3239          * the pages by calling redirty_page_for_writepage() but that
3240          * would be ugly in the extreme.  So instead we would need to
3241          * replicate parts of the code in the above functions,
3242          * simplifying them because we wouldn't actually intend to
3243          * write out the pages, but rather only collect contiguous
3244          * logical block extents, call the multi-block allocator, and
3245          * then update the buffer heads with the block allocations.
3246          *
3247          * For now, though, we'll cheat by calling filemap_flush(),
3248          * which will map the blocks, and start the I/O, but not
3249          * actually wait for the I/O to complete.
3250          */
3251         return filemap_flush(inode->i_mapping);
3252 }
3253
3254 /*
3255  * bmap() is special.  It gets used by applications such as lilo and by
3256  * the swapper to find the on-disk block of a specific piece of data.
3257  *
3258  * Naturally, this is dangerous if the block concerned is still in the
3259  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3260  * filesystem and enables swap, then they may get a nasty shock when the
3261  * data getting swapped to that swapfile suddenly gets overwritten by
3262  * the original zero's written out previously to the journal and
3263  * awaiting writeback in the kernel's buffer cache.
3264  *
3265  * So, if we see any bmap calls here on a modified, data-journaled file,
3266  * take extra steps to flush any blocks which might be in the cache.
3267  */
3268 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3269 {
3270         struct inode *inode = mapping->host;
3271         journal_t *journal;
3272         int err;
3273
3274         /*
3275          * We can get here for an inline file via the FIBMAP ioctl
3276          */
3277         if (ext4_has_inline_data(inode))
3278                 return 0;
3279
3280         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3281                         test_opt(inode->i_sb, DELALLOC)) {
3282                 /*
3283                  * With delalloc we want to sync the file
3284                  * so that we can make sure we allocate
3285                  * blocks for file
3286                  */
3287                 filemap_write_and_wait(mapping);
3288         }
3289
3290         if (EXT4_JOURNAL(inode) &&
3291             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
3292                 /*
3293                  * This is a REALLY heavyweight approach, but the use of
3294                  * bmap on dirty files is expected to be extremely rare:
3295                  * only if we run lilo or swapon on a freshly made file
3296                  * do we expect this to happen.
3297                  *
3298                  * (bmap requires CAP_SYS_RAWIO so this does not
3299                  * represent an unprivileged user DOS attack --- we'd be
3300                  * in trouble if mortal users could trigger this path at
3301                  * will.)
3302                  *
3303                  * NB. EXT4_STATE_JDATA is not set on files other than
3304                  * regular files.  If somebody wants to bmap a directory
3305                  * or symlink and gets confused because the buffer
3306                  * hasn't yet been flushed to disk, they deserve
3307                  * everything they get.
3308                  */
3309
3310                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
3311                 journal = EXT4_JOURNAL(inode);
3312                 jbd2_journal_lock_updates(journal);
3313                 err = jbd2_journal_flush(journal);
3314                 jbd2_journal_unlock_updates(journal);
3315
3316                 if (err)
3317                         return 0;
3318         }
3319
3320         return generic_block_bmap(mapping, block, ext4_get_block);
3321 }
3322
3323 static int ext4_readpage(struct file *file, struct page *page)
3324 {
3325         int ret = -EAGAIN;
3326         struct inode *inode = page->mapping->host;
3327
3328         trace_ext4_readpage(page);
3329
3330         if (ext4_has_inline_data(inode))
3331                 ret = ext4_readpage_inline(inode, page);
3332
3333         if (ret == -EAGAIN)
3334                 return ext4_mpage_readpages(page->mapping, NULL, page, 1);
3335
3336         return ret;
3337 }
3338
3339 static int
3340 ext4_readpages(struct file *file, struct address_space *mapping,
3341                 struct list_head *pages, unsigned nr_pages)
3342 {
3343         struct inode *inode = mapping->host;
3344
3345         /* If the file has inline data, no need to do readpages. */
3346         if (ext4_has_inline_data(inode))
3347                 return 0;
3348
3349         return ext4_mpage_readpages(mapping, pages, NULL, nr_pages);
3350 }
3351
3352 static void ext4_invalidatepage(struct page *page, unsigned int offset,
3353                                 unsigned int length)
3354 {
3355         trace_ext4_invalidatepage(page, offset, length);
3356
3357         /* No journalling happens on data buffers when this function is used */
3358         WARN_ON(page_has_buffers(page) && buffer_jbd(page_buffers(page)));
3359
3360         block_invalidatepage(page, offset, length);
3361 }
3362
3363 static int __ext4_journalled_invalidatepage(struct page *page,
3364                                             unsigned int offset,
3365                                             unsigned int length)
3366 {
3367         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3368
3369         trace_ext4_journalled_invalidatepage(page, offset, length);
3370
3371         /*
3372          * If it's a full truncate we just forget about the pending dirtying
3373          */
3374         if (offset == 0 && length == PAGE_SIZE)
3375                 ClearPageChecked(page);
3376
3377         return jbd2_journal_invalidatepage(journal, page, offset, length);
3378 }
3379
3380 /* Wrapper for aops... */
3381 static void ext4_journalled_invalidatepage(struct page *page,
3382                                            unsigned int offset,
3383                                            unsigned int length)
3384 {
3385         WARN_ON(__ext4_journalled_invalidatepage(page, offset, length) < 0);
3386 }
3387
3388 static int ext4_releasepage(struct page *page, gfp_t wait)
3389 {
3390         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3391
3392         trace_ext4_releasepage(page);
3393
3394         /* Page has dirty journalled data -> cannot release */
3395         if (PageChecked(page))
3396                 return 0;
3397         if (journal)
3398                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3399         else
3400                 return try_to_free_buffers(page);
3401 }
3402
3403 #ifdef CONFIG_FS_DAX
3404 static int ext4_iomap_begin(struct inode *inode, loff_t offset, loff_t length,
3405                             unsigned flags, struct iomap *iomap)
3406 {
3407         struct block_device *bdev;
3408         unsigned int blkbits = inode->i_blkbits;
3409         unsigned long first_block = offset >> blkbits;
3410         unsigned long last_block = (offset + length - 1) >> blkbits;
3411         struct ext4_map_blocks map;
3412         int ret;
3413
3414         if (WARN_ON_ONCE(ext4_has_inline_data(inode)))
3415                 return -ERANGE;
3416
3417         map.m_lblk = first_block;
3418         map.m_len = last_block - first_block + 1;
3419
3420         if (!(flags & IOMAP_WRITE)) {
3421                 ret = ext4_map_blocks(NULL, inode, &map, 0);
3422         } else {
3423                 int dio_credits;
3424                 handle_t *handle;
3425                 int retries = 0;
3426
3427                 /* Trim mapping request to maximum we can map at once for DIO */
3428                 if (map.m_len > DIO_MAX_BLOCKS)
3429                         map.m_len = DIO_MAX_BLOCKS;
3430                 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
3431 retry:
3432                 /*
3433                  * Either we allocate blocks and then we don't get unwritten
3434                  * extent so we have reserved enough credits, or the blocks
3435                  * are already allocated and unwritten and in that case
3436                  * extent conversion fits in the credits as well.
3437                  */
3438                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
3439                                             dio_credits);
3440                 if (IS_ERR(handle))
3441                         return PTR_ERR(handle);
3442
3443                 ret = ext4_map_blocks(handle, inode, &map,
3444                                       EXT4_GET_BLOCKS_CREATE_ZERO);
3445                 if (ret < 0) {
3446                         ext4_journal_stop(handle);
3447                         if (ret == -ENOSPC &&
3448                             ext4_should_retry_alloc(inode->i_sb, &retries))
3449                                 goto retry;
3450                         return ret;
3451                 }
3452
3453                 /*
3454                  * If we added blocks beyond i_size, we need to make sure they
3455                  * will get truncated if we crash before updating i_size in
3456                  * ext4_iomap_end(). For faults we don't need to do that (and
3457                  * even cannot because for orphan list operations inode_lock is
3458                  * required) - if we happen to instantiate block beyond i_size,
3459                  * it is because we race with truncate which has already added
3460                  * the inode to the orphan list.
3461                  */
3462                 if (!(flags & IOMAP_FAULT) && first_block + map.m_len >
3463                     (i_size_read(inode) + (1 << blkbits) - 1) >> blkbits) {
3464                         int err;
3465
3466                         err = ext4_orphan_add(handle, inode);
3467                         if (err < 0) {
3468                                 ext4_journal_stop(handle);
3469                                 return err;
3470                         }
3471                 }
3472                 ext4_journal_stop(handle);
3473         }
3474
3475         iomap->flags = 0;
3476         bdev = inode->i_sb->s_bdev;
3477         iomap->bdev = bdev;
3478         if (blk_queue_dax(bdev->bd_queue))
3479                 iomap->dax_dev = fs_dax_get_by_host(bdev->bd_disk->disk_name);
3480         else
3481                 iomap->dax_dev = NULL;
3482         iomap->offset = first_block << blkbits;
3483
3484         if (ret == 0) {
3485                 iomap->type = IOMAP_HOLE;
3486                 iomap->blkno = IOMAP_NULL_BLOCK;
3487                 iomap->length = (u64)map.m_len << blkbits;
3488         } else {
3489                 if (map.m_flags & EXT4_MAP_MAPPED) {
3490                         iomap->type = IOMAP_MAPPED;
3491                 } else if (map.m_flags & EXT4_MAP_UNWRITTEN) {
3492                         iomap->type = IOMAP_UNWRITTEN;
3493                 } else {
3494                         WARN_ON_ONCE(1);
3495                         return -EIO;
3496                 }
3497                 iomap->blkno = (sector_t)map.m_pblk << (blkbits - 9);
3498                 iomap->length = (u64)map.m_len << blkbits;
3499         }
3500
3501         if (map.m_flags & EXT4_MAP_NEW)
3502                 iomap->flags |= IOMAP_F_NEW;
3503         return 0;
3504 }
3505
3506 static int ext4_iomap_end(struct inode *inode, loff_t offset, loff_t length,
3507                           ssize_t written, unsigned flags, struct iomap *iomap)
3508 {
3509         int ret = 0;
3510         handle_t *handle;
3511         int blkbits = inode->i_blkbits;
3512         bool truncate = false;
3513
3514         fs_put_dax(iomap->dax_dev);
3515         if (!(flags & IOMAP_WRITE) || (flags & IOMAP_FAULT))
3516                 return 0;
3517
3518         handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
3519         if (IS_ERR(handle)) {
3520                 ret = PTR_ERR(handle);
3521                 goto orphan_del;
3522         }
3523         if (ext4_update_inode_size(inode, offset + written))
3524                 ext4_mark_inode_dirty(handle, inode);
3525         /*
3526          * We may need to truncate allocated but not written blocks beyond EOF.
3527          */
3528         if (iomap->offset + iomap->length > 
3529             ALIGN(inode->i_size, 1 << blkbits)) {
3530                 ext4_lblk_t written_blk, end_blk;
3531
3532                 written_blk = (offset + written) >> blkbits;
3533                 end_blk = (offset + length) >> blkbits;
3534                 if (written_blk < end_blk && ext4_can_truncate(inode))
3535                         truncate = true;
3536         }
3537         /*
3538          * Remove inode from orphan list if we were extending a inode and
3539          * everything went fine.
3540          */
3541         if (!truncate && inode->i_nlink &&
3542             !list_empty(&EXT4_I(inode)->i_orphan))
3543                 ext4_orphan_del(handle, inode);
3544         ext4_journal_stop(handle);
3545         if (truncate) {
3546                 ext4_truncate_failed_write(inode);
3547 orphan_del:
3548                 /*
3549                  * If truncate failed early the inode might still be on the
3550                  * orphan list; we need to make sure the inode is removed from
3551                  * the orphan list in that case.
3552                  */
3553                 if (inode->i_nlink)
3554                         ext4_orphan_del(NULL, inode);
3555         }
3556         return ret;
3557 }
3558
3559 const struct iomap_ops ext4_iomap_ops = {
3560         .iomap_begin            = ext4_iomap_begin,
3561         .iomap_end              = ext4_iomap_end,
3562 };
3563
3564 #endif
3565
3566 static int ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3567                             ssize_t size, void *private)
3568 {
3569         ext4_io_end_t *io_end = private;
3570
3571         /* if not async direct IO just return */
3572         if (!io_end)
3573                 return 0;
3574
3575         ext_debug("ext4_end_io_dio(): io_end 0x%p "
3576                   "for inode %lu, iocb 0x%p, offset %llu, size %zd\n",
3577                   io_end, io_end->inode->i_ino, iocb, offset, size);
3578
3579         /*
3580          * Error during AIO DIO. We cannot convert unwritten extents as the
3581          * data was not written. Just clear the unwritten flag and drop io_end.
3582          */
3583         if (size <= 0) {
3584                 ext4_clear_io_unwritten_flag(io_end);
3585                 size = 0;
3586         }
3587         io_end->offset = offset;
3588         io_end->size = size;
3589         ext4_put_io_end(io_end);
3590
3591         return 0;
3592 }
3593
3594 /*
3595  * Handling of direct IO writes.
3596  *
3597  * For ext4 extent files, ext4 will do direct-io write even to holes,
3598  * preallocated extents, and those write extend the file, no need to
3599  * fall back to buffered IO.
3600  *
3601  * For holes, we fallocate those blocks, mark them as unwritten
3602  * If those blocks were preallocated, we mark sure they are split, but
3603  * still keep the range to write as unwritten.
3604  *
3605  * The unwritten extents will be converted to written when DIO is completed.
3606  * For async direct IO, since the IO may still pending when return, we
3607  * set up an end_io call back function, which will do the conversion
3608  * when async direct IO completed.
3609  *
3610  * If the O_DIRECT write will extend the file then add this inode to the
3611  * orphan list.  So recovery will truncate it back to the original size
3612  * if the machine crashes during the write.
3613  *
3614  */
3615 static ssize_t ext4_direct_IO_write(struct kiocb *iocb, struct iov_iter *iter)
3616 {
3617         struct file *file = iocb->ki_filp;
3618         struct inode *inode = file->f_mapping->host;
3619         struct ext4_inode_info *ei = EXT4_I(inode);
3620         ssize_t ret;
<