]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - fs/ext4/inode.c
Merge branch 'next/defconfig' into for-next
[karo-tx-linux.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
41 #include "ext4_jbd2.h"
42 #include "xattr.h"
43 #include "acl.h"
44 #include "truncate.h"
45
46 #include <trace/events/ext4.h>
47
48 #define MPAGE_DA_EXTENT_TAIL 0x01
49
50 static __u32 ext4_inode_csum(struct inode *inode, struct ext4_inode *raw,
51                               struct ext4_inode_info *ei)
52 {
53         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
54         __u16 csum_lo;
55         __u16 csum_hi = 0;
56         __u32 csum;
57
58         csum_lo = le16_to_cpu(raw->i_checksum_lo);
59         raw->i_checksum_lo = 0;
60         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
61             EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi)) {
62                 csum_hi = le16_to_cpu(raw->i_checksum_hi);
63                 raw->i_checksum_hi = 0;
64         }
65
66         csum = ext4_chksum(sbi, ei->i_csum_seed, (__u8 *)raw,
67                            EXT4_INODE_SIZE(inode->i_sb));
68
69         raw->i_checksum_lo = cpu_to_le16(csum_lo);
70         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
71             EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
72                 raw->i_checksum_hi = cpu_to_le16(csum_hi);
73
74         return csum;
75 }
76
77 static int ext4_inode_csum_verify(struct inode *inode, struct ext4_inode *raw,
78                                   struct ext4_inode_info *ei)
79 {
80         __u32 provided, calculated;
81
82         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
83             cpu_to_le32(EXT4_OS_LINUX) ||
84             !ext4_has_metadata_csum(inode->i_sb))
85                 return 1;
86
87         provided = le16_to_cpu(raw->i_checksum_lo);
88         calculated = ext4_inode_csum(inode, raw, ei);
89         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
90             EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
91                 provided |= ((__u32)le16_to_cpu(raw->i_checksum_hi)) << 16;
92         else
93                 calculated &= 0xFFFF;
94
95         return provided == calculated;
96 }
97
98 static void ext4_inode_csum_set(struct inode *inode, struct ext4_inode *raw,
99                                 struct ext4_inode_info *ei)
100 {
101         __u32 csum;
102
103         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
104             cpu_to_le32(EXT4_OS_LINUX) ||
105             !ext4_has_metadata_csum(inode->i_sb))
106                 return;
107
108         csum = ext4_inode_csum(inode, raw, ei);
109         raw->i_checksum_lo = cpu_to_le16(csum & 0xFFFF);
110         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE &&
111             EXT4_FITS_IN_INODE(raw, ei, i_checksum_hi))
112                 raw->i_checksum_hi = cpu_to_le16(csum >> 16);
113 }
114
115 static inline int ext4_begin_ordered_truncate(struct inode *inode,
116                                               loff_t new_size)
117 {
118         trace_ext4_begin_ordered_truncate(inode, new_size);
119         /*
120          * If jinode is zero, then we never opened the file for
121          * writing, so there's no need to call
122          * jbd2_journal_begin_ordered_truncate() since there's no
123          * outstanding writes we need to flush.
124          */
125         if (!EXT4_I(inode)->jinode)
126                 return 0;
127         return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
128                                                    EXT4_I(inode)->jinode,
129                                                    new_size);
130 }
131
132 static void ext4_invalidatepage(struct page *page, unsigned int offset,
133                                 unsigned int length);
134 static int __ext4_journalled_writepage(struct page *page, unsigned int len);
135 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
136 static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
137                                   int pextents);
138
139 /*
140  * Test whether an inode is a fast symlink.
141  */
142 int ext4_inode_is_fast_symlink(struct inode *inode)
143 {
144         int ea_blocks = EXT4_I(inode)->i_file_acl ?
145                 EXT4_CLUSTER_SIZE(inode->i_sb) >> 9 : 0;
146
147         if (ext4_has_inline_data(inode))
148                 return 0;
149
150         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
151 }
152
153 /*
154  * Restart the transaction associated with *handle.  This does a commit,
155  * so before we call here everything must be consistently dirtied against
156  * this transaction.
157  */
158 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
159                                  int nblocks)
160 {
161         int ret;
162
163         /*
164          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
165          * moment, get_block can be called only for blocks inside i_size since
166          * page cache has been already dropped and writes are blocked by
167          * i_mutex. So we can safely drop the i_data_sem here.
168          */
169         BUG_ON(EXT4_JOURNAL(inode) == NULL);
170         jbd_debug(2, "restarting handle %p\n", handle);
171         up_write(&EXT4_I(inode)->i_data_sem);
172         ret = ext4_journal_restart(handle, nblocks);
173         down_write(&EXT4_I(inode)->i_data_sem);
174         ext4_discard_preallocations(inode);
175
176         return ret;
177 }
178
179 /*
180  * Called at the last iput() if i_nlink is zero.
181  */
182 void ext4_evict_inode(struct inode *inode)
183 {
184         handle_t *handle;
185         int err;
186
187         trace_ext4_evict_inode(inode);
188
189         if (inode->i_nlink) {
190                 /*
191                  * When journalling data dirty buffers are tracked only in the
192                  * journal. So although mm thinks everything is clean and
193                  * ready for reaping the inode might still have some pages to
194                  * write in the running transaction or waiting to be
195                  * checkpointed. Thus calling jbd2_journal_invalidatepage()
196                  * (via truncate_inode_pages()) to discard these buffers can
197                  * cause data loss. Also even if we did not discard these
198                  * buffers, we would have no way to find them after the inode
199                  * is reaped and thus user could see stale data if he tries to
200                  * read them before the transaction is checkpointed. So be
201                  * careful and force everything to disk here... We use
202                  * ei->i_datasync_tid to store the newest transaction
203                  * containing inode's data.
204                  *
205                  * Note that directories do not have this problem because they
206                  * don't use page cache.
207                  */
208                 if (ext4_should_journal_data(inode) &&
209                     (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode)) &&
210                     inode->i_ino != EXT4_JOURNAL_INO) {
211                         journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
212                         tid_t commit_tid = EXT4_I(inode)->i_datasync_tid;
213
214                         jbd2_complete_transaction(journal, commit_tid);
215                         filemap_write_and_wait(&inode->i_data);
216                 }
217                 truncate_inode_pages_final(&inode->i_data);
218
219                 WARN_ON(atomic_read(&EXT4_I(inode)->i_ioend_count));
220                 goto no_delete;
221         }
222
223         if (is_bad_inode(inode))
224                 goto no_delete;
225         dquot_initialize(inode);
226
227         if (ext4_should_order_data(inode))
228                 ext4_begin_ordered_truncate(inode, 0);
229         truncate_inode_pages_final(&inode->i_data);
230
231         WARN_ON(atomic_read(&EXT4_I(inode)->i_ioend_count));
232
233         /*
234          * Protect us against freezing - iput() caller didn't have to have any
235          * protection against it
236          */
237         sb_start_intwrite(inode->i_sb);
238         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE,
239                                     ext4_blocks_for_truncate(inode)+3);
240         if (IS_ERR(handle)) {
241                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
242                 /*
243                  * If we're going to skip the normal cleanup, we still need to
244                  * make sure that the in-core orphan linked list is properly
245                  * cleaned up.
246                  */
247                 ext4_orphan_del(NULL, inode);
248                 sb_end_intwrite(inode->i_sb);
249                 goto no_delete;
250         }
251
252         if (IS_SYNC(inode))
253                 ext4_handle_sync(handle);
254         inode->i_size = 0;
255         err = ext4_mark_inode_dirty(handle, inode);
256         if (err) {
257                 ext4_warning(inode->i_sb,
258                              "couldn't mark inode dirty (err %d)", err);
259                 goto stop_handle;
260         }
261         if (inode->i_blocks)
262                 ext4_truncate(inode);
263
264         /*
265          * ext4_ext_truncate() doesn't reserve any slop when it
266          * restarts journal transactions; therefore there may not be
267          * enough credits left in the handle to remove the inode from
268          * the orphan list and set the dtime field.
269          */
270         if (!ext4_handle_has_enough_credits(handle, 3)) {
271                 err = ext4_journal_extend(handle, 3);
272                 if (err > 0)
273                         err = ext4_journal_restart(handle, 3);
274                 if (err != 0) {
275                         ext4_warning(inode->i_sb,
276                                      "couldn't extend journal (err %d)", err);
277                 stop_handle:
278                         ext4_journal_stop(handle);
279                         ext4_orphan_del(NULL, inode);
280                         sb_end_intwrite(inode->i_sb);
281                         goto no_delete;
282                 }
283         }
284
285         /*
286          * Kill off the orphan record which ext4_truncate created.
287          * AKPM: I think this can be inside the above `if'.
288          * Note that ext4_orphan_del() has to be able to cope with the
289          * deletion of a non-existent orphan - this is because we don't
290          * know if ext4_truncate() actually created an orphan record.
291          * (Well, we could do this if we need to, but heck - it works)
292          */
293         ext4_orphan_del(handle, inode);
294         EXT4_I(inode)->i_dtime  = get_seconds();
295
296         /*
297          * One subtle ordering requirement: if anything has gone wrong
298          * (transaction abort, IO errors, whatever), then we can still
299          * do these next steps (the fs will already have been marked as
300          * having errors), but we can't free the inode if the mark_dirty
301          * fails.
302          */
303         if (ext4_mark_inode_dirty(handle, inode))
304                 /* If that failed, just do the required in-core inode clear. */
305                 ext4_clear_inode(inode);
306         else
307                 ext4_free_inode(handle, inode);
308         ext4_journal_stop(handle);
309         sb_end_intwrite(inode->i_sb);
310         return;
311 no_delete:
312         ext4_clear_inode(inode);        /* We must guarantee clearing of inode... */
313 }
314
315 #ifdef CONFIG_QUOTA
316 qsize_t *ext4_get_reserved_space(struct inode *inode)
317 {
318         return &EXT4_I(inode)->i_reserved_quota;
319 }
320 #endif
321
322 /*
323  * Called with i_data_sem down, which is important since we can call
324  * ext4_discard_preallocations() from here.
325  */
326 void ext4_da_update_reserve_space(struct inode *inode,
327                                         int used, int quota_claim)
328 {
329         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
330         struct ext4_inode_info *ei = EXT4_I(inode);
331
332         spin_lock(&ei->i_block_reservation_lock);
333         trace_ext4_da_update_reserve_space(inode, used, quota_claim);
334         if (unlikely(used > ei->i_reserved_data_blocks)) {
335                 ext4_warning(inode->i_sb, "%s: ino %lu, used %d "
336                          "with only %d reserved data blocks",
337                          __func__, inode->i_ino, used,
338                          ei->i_reserved_data_blocks);
339                 WARN_ON(1);
340                 used = ei->i_reserved_data_blocks;
341         }
342
343         /* Update per-inode reservations */
344         ei->i_reserved_data_blocks -= used;
345         percpu_counter_sub(&sbi->s_dirtyclusters_counter, used);
346
347         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
348
349         /* Update quota subsystem for data blocks */
350         if (quota_claim)
351                 dquot_claim_block(inode, EXT4_C2B(sbi, used));
352         else {
353                 /*
354                  * We did fallocate with an offset that is already delayed
355                  * allocated. So on delayed allocated writeback we should
356                  * not re-claim the quota for fallocated blocks.
357                  */
358                 dquot_release_reservation_block(inode, EXT4_C2B(sbi, used));
359         }
360
361         /*
362          * If we have done all the pending block allocations and if
363          * there aren't any writers on the inode, we can discard the
364          * inode's preallocations.
365          */
366         if ((ei->i_reserved_data_blocks == 0) &&
367             (atomic_read(&inode->i_writecount) == 0))
368                 ext4_discard_preallocations(inode);
369 }
370
371 static int __check_block_validity(struct inode *inode, const char *func,
372                                 unsigned int line,
373                                 struct ext4_map_blocks *map)
374 {
375         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
376                                    map->m_len)) {
377                 ext4_error_inode(inode, func, line, map->m_pblk,
378                                  "lblock %lu mapped to illegal pblock "
379                                  "(length %d)", (unsigned long) map->m_lblk,
380                                  map->m_len);
381                 return -EIO;
382         }
383         return 0;
384 }
385
386 #define check_block_validity(inode, map)        \
387         __check_block_validity((inode), __func__, __LINE__, (map))
388
389 #ifdef ES_AGGRESSIVE_TEST
390 static void ext4_map_blocks_es_recheck(handle_t *handle,
391                                        struct inode *inode,
392                                        struct ext4_map_blocks *es_map,
393                                        struct ext4_map_blocks *map,
394                                        int flags)
395 {
396         int retval;
397
398         map->m_flags = 0;
399         /*
400          * There is a race window that the result is not the same.
401          * e.g. xfstests #223 when dioread_nolock enables.  The reason
402          * is that we lookup a block mapping in extent status tree with
403          * out taking i_data_sem.  So at the time the unwritten extent
404          * could be converted.
405          */
406         if (!(flags & EXT4_GET_BLOCKS_NO_LOCK))
407                 down_read(&EXT4_I(inode)->i_data_sem);
408         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
409                 retval = ext4_ext_map_blocks(handle, inode, map, flags &
410                                              EXT4_GET_BLOCKS_KEEP_SIZE);
411         } else {
412                 retval = ext4_ind_map_blocks(handle, inode, map, flags &
413                                              EXT4_GET_BLOCKS_KEEP_SIZE);
414         }
415         if (!(flags & EXT4_GET_BLOCKS_NO_LOCK))
416                 up_read((&EXT4_I(inode)->i_data_sem));
417
418         /*
419          * We don't check m_len because extent will be collpased in status
420          * tree.  So the m_len might not equal.
421          */
422         if (es_map->m_lblk != map->m_lblk ||
423             es_map->m_flags != map->m_flags ||
424             es_map->m_pblk != map->m_pblk) {
425                 printk("ES cache assertion failed for inode: %lu "
426                        "es_cached ex [%d/%d/%llu/%x] != "
427                        "found ex [%d/%d/%llu/%x] retval %d flags %x\n",
428                        inode->i_ino, es_map->m_lblk, es_map->m_len,
429                        es_map->m_pblk, es_map->m_flags, map->m_lblk,
430                        map->m_len, map->m_pblk, map->m_flags,
431                        retval, flags);
432         }
433 }
434 #endif /* ES_AGGRESSIVE_TEST */
435
436 /*
437  * The ext4_map_blocks() function tries to look up the requested blocks,
438  * and returns if the blocks are already mapped.
439  *
440  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
441  * and store the allocated blocks in the result buffer head and mark it
442  * mapped.
443  *
444  * If file type is extents based, it will call ext4_ext_map_blocks(),
445  * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
446  * based files
447  *
448  * On success, it returns the number of blocks being mapped or allocated.
449  * if create==0 and the blocks are pre-allocated and unwritten block,
450  * the result buffer head is unmapped. If the create ==1, it will make sure
451  * the buffer head is mapped.
452  *
453  * It returns 0 if plain look up failed (blocks have not been allocated), in
454  * that case, buffer head is unmapped
455  *
456  * It returns the error in case of allocation failure.
457  */
458 int ext4_map_blocks(handle_t *handle, struct inode *inode,
459                     struct ext4_map_blocks *map, int flags)
460 {
461         struct extent_status es;
462         int retval;
463         int ret = 0;
464 #ifdef ES_AGGRESSIVE_TEST
465         struct ext4_map_blocks orig_map;
466
467         memcpy(&orig_map, map, sizeof(*map));
468 #endif
469
470         map->m_flags = 0;
471         ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
472                   "logical block %lu\n", inode->i_ino, flags, map->m_len,
473                   (unsigned long) map->m_lblk);
474
475         /*
476          * ext4_map_blocks returns an int, and m_len is an unsigned int
477          */
478         if (unlikely(map->m_len > INT_MAX))
479                 map->m_len = INT_MAX;
480
481         /* We can handle the block number less than EXT_MAX_BLOCKS */
482         if (unlikely(map->m_lblk >= EXT_MAX_BLOCKS))
483                 return -EIO;
484
485         /* Lookup extent status tree firstly */
486         if (ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
487                 if (ext4_es_is_written(&es) || ext4_es_is_unwritten(&es)) {
488                         map->m_pblk = ext4_es_pblock(&es) +
489                                         map->m_lblk - es.es_lblk;
490                         map->m_flags |= ext4_es_is_written(&es) ?
491                                         EXT4_MAP_MAPPED : EXT4_MAP_UNWRITTEN;
492                         retval = es.es_len - (map->m_lblk - es.es_lblk);
493                         if (retval > map->m_len)
494                                 retval = map->m_len;
495                         map->m_len = retval;
496                 } else if (ext4_es_is_delayed(&es) || ext4_es_is_hole(&es)) {
497                         retval = 0;
498                 } else {
499                         BUG_ON(1);
500                 }
501 #ifdef ES_AGGRESSIVE_TEST
502                 ext4_map_blocks_es_recheck(handle, inode, map,
503                                            &orig_map, flags);
504 #endif
505                 goto found;
506         }
507
508         /*
509          * Try to see if we can get the block without requesting a new
510          * file system block.
511          */
512         if (!(flags & EXT4_GET_BLOCKS_NO_LOCK))
513                 down_read(&EXT4_I(inode)->i_data_sem);
514         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
515                 retval = ext4_ext_map_blocks(handle, inode, map, flags &
516                                              EXT4_GET_BLOCKS_KEEP_SIZE);
517         } else {
518                 retval = ext4_ind_map_blocks(handle, inode, map, flags &
519                                              EXT4_GET_BLOCKS_KEEP_SIZE);
520         }
521         if (retval > 0) {
522                 unsigned int status;
523
524                 if (unlikely(retval != map->m_len)) {
525                         ext4_warning(inode->i_sb,
526                                      "ES len assertion failed for inode "
527                                      "%lu: retval %d != map->m_len %d",
528                                      inode->i_ino, retval, map->m_len);
529                         WARN_ON(1);
530                 }
531
532                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
533                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
534                 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
535                     !(status & EXTENT_STATUS_WRITTEN) &&
536                     ext4_find_delalloc_range(inode, map->m_lblk,
537                                              map->m_lblk + map->m_len - 1))
538                         status |= EXTENT_STATUS_DELAYED;
539                 ret = ext4_es_insert_extent(inode, map->m_lblk,
540                                             map->m_len, map->m_pblk, status);
541                 if (ret < 0)
542                         retval = ret;
543         }
544         if (!(flags & EXT4_GET_BLOCKS_NO_LOCK))
545                 up_read((&EXT4_I(inode)->i_data_sem));
546
547 found:
548         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
549                 ret = check_block_validity(inode, map);
550                 if (ret != 0)
551                         return ret;
552         }
553
554         /* If it is only a block(s) look up */
555         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
556                 return retval;
557
558         /*
559          * Returns if the blocks have already allocated
560          *
561          * Note that if blocks have been preallocated
562          * ext4_ext_get_block() returns the create = 0
563          * with buffer head unmapped.
564          */
565         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
566                 /*
567                  * If we need to convert extent to unwritten
568                  * we continue and do the actual work in
569                  * ext4_ext_map_blocks()
570                  */
571                 if (!(flags & EXT4_GET_BLOCKS_CONVERT_UNWRITTEN))
572                         return retval;
573
574         /*
575          * Here we clear m_flags because after allocating an new extent,
576          * it will be set again.
577          */
578         map->m_flags &= ~EXT4_MAP_FLAGS;
579
580         /*
581          * New blocks allocate and/or writing to unwritten extent
582          * will possibly result in updating i_data, so we take
583          * the write lock of i_data_sem, and call get_block()
584          * with create == 1 flag.
585          */
586         down_write(&EXT4_I(inode)->i_data_sem);
587
588         /*
589          * We need to check for EXT4 here because migrate
590          * could have changed the inode type in between
591          */
592         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
593                 retval = ext4_ext_map_blocks(handle, inode, map, flags);
594         } else {
595                 retval = ext4_ind_map_blocks(handle, inode, map, flags);
596
597                 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
598                         /*
599                          * We allocated new blocks which will result in
600                          * i_data's format changing.  Force the migrate
601                          * to fail by clearing migrate flags
602                          */
603                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
604                 }
605
606                 /*
607                  * Update reserved blocks/metadata blocks after successful
608                  * block allocation which had been deferred till now. We don't
609                  * support fallocate for non extent files. So we can update
610                  * reserve space here.
611                  */
612                 if ((retval > 0) &&
613                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
614                         ext4_da_update_reserve_space(inode, retval, 1);
615         }
616
617         if (retval > 0) {
618                 unsigned int status;
619
620                 if (unlikely(retval != map->m_len)) {
621                         ext4_warning(inode->i_sb,
622                                      "ES len assertion failed for inode "
623                                      "%lu: retval %d != map->m_len %d",
624                                      inode->i_ino, retval, map->m_len);
625                         WARN_ON(1);
626                 }
627
628                 /*
629                  * If the extent has been zeroed out, we don't need to update
630                  * extent status tree.
631                  */
632                 if ((flags & EXT4_GET_BLOCKS_PRE_IO) &&
633                     ext4_es_lookup_extent(inode, map->m_lblk, &es)) {
634                         if (ext4_es_is_written(&es))
635                                 goto has_zeroout;
636                 }
637                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
638                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
639                 if (!(flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) &&
640                     !(status & EXTENT_STATUS_WRITTEN) &&
641                     ext4_find_delalloc_range(inode, map->m_lblk,
642                                              map->m_lblk + map->m_len - 1))
643                         status |= EXTENT_STATUS_DELAYED;
644                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
645                                             map->m_pblk, status);
646                 if (ret < 0)
647                         retval = ret;
648         }
649
650 has_zeroout:
651         up_write((&EXT4_I(inode)->i_data_sem));
652         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
653                 ret = check_block_validity(inode, map);
654                 if (ret != 0)
655                         return ret;
656         }
657         return retval;
658 }
659
660 /* Maximum number of blocks we map for direct IO at once. */
661 #define DIO_MAX_BLOCKS 4096
662
663 static int _ext4_get_block(struct inode *inode, sector_t iblock,
664                            struct buffer_head *bh, int flags)
665 {
666         handle_t *handle = ext4_journal_current_handle();
667         struct ext4_map_blocks map;
668         int ret = 0, started = 0;
669         int dio_credits;
670
671         if (ext4_has_inline_data(inode))
672                 return -ERANGE;
673
674         map.m_lblk = iblock;
675         map.m_len = bh->b_size >> inode->i_blkbits;
676
677         if (flags && !(flags & EXT4_GET_BLOCKS_NO_LOCK) && !handle) {
678                 /* Direct IO write... */
679                 if (map.m_len > DIO_MAX_BLOCKS)
680                         map.m_len = DIO_MAX_BLOCKS;
681                 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
682                 handle = ext4_journal_start(inode, EXT4_HT_MAP_BLOCKS,
683                                             dio_credits);
684                 if (IS_ERR(handle)) {
685                         ret = PTR_ERR(handle);
686                         return ret;
687                 }
688                 started = 1;
689         }
690
691         ret = ext4_map_blocks(handle, inode, &map, flags);
692         if (ret > 0) {
693                 ext4_io_end_t *io_end = ext4_inode_aio(inode);
694
695                 map_bh(bh, inode->i_sb, map.m_pblk);
696                 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
697                 if (IS_DAX(inode) && buffer_unwritten(bh)) {
698                         /*
699                          * dgc: I suspect unwritten conversion on ext4+DAX is
700                          * fundamentally broken here when there are concurrent
701                          * read/write in progress on this inode.
702                          */
703                         WARN_ON_ONCE(io_end);
704                         bh->b_assoc_map = inode->i_mapping;
705                         bh->b_private = (void *)(unsigned long)iblock;
706                 }
707                 if (io_end && io_end->flag & EXT4_IO_END_UNWRITTEN)
708                         set_buffer_defer_completion(bh);
709                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
710                 ret = 0;
711         }
712         if (started)
713                 ext4_journal_stop(handle);
714         return ret;
715 }
716
717 int ext4_get_block(struct inode *inode, sector_t iblock,
718                    struct buffer_head *bh, int create)
719 {
720         return _ext4_get_block(inode, iblock, bh,
721                                create ? EXT4_GET_BLOCKS_CREATE : 0);
722 }
723
724 /*
725  * `handle' can be NULL if create is zero
726  */
727 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
728                                 ext4_lblk_t block, int map_flags)
729 {
730         struct ext4_map_blocks map;
731         struct buffer_head *bh;
732         int create = map_flags & EXT4_GET_BLOCKS_CREATE;
733         int err;
734
735         J_ASSERT(handle != NULL || create == 0);
736
737         map.m_lblk = block;
738         map.m_len = 1;
739         err = ext4_map_blocks(handle, inode, &map, map_flags);
740
741         if (err == 0)
742                 return create ? ERR_PTR(-ENOSPC) : NULL;
743         if (err < 0)
744                 return ERR_PTR(err);
745
746         bh = sb_getblk(inode->i_sb, map.m_pblk);
747         if (unlikely(!bh))
748                 return ERR_PTR(-ENOMEM);
749         if (map.m_flags & EXT4_MAP_NEW) {
750                 J_ASSERT(create != 0);
751                 J_ASSERT(handle != NULL);
752
753                 /*
754                  * Now that we do not always journal data, we should
755                  * keep in mind whether this should always journal the
756                  * new buffer as metadata.  For now, regular file
757                  * writes use ext4_get_block instead, so it's not a
758                  * problem.
759                  */
760                 lock_buffer(bh);
761                 BUFFER_TRACE(bh, "call get_create_access");
762                 err = ext4_journal_get_create_access(handle, bh);
763                 if (unlikely(err)) {
764                         unlock_buffer(bh);
765                         goto errout;
766                 }
767                 if (!buffer_uptodate(bh)) {
768                         memset(bh->b_data, 0, inode->i_sb->s_blocksize);
769                         set_buffer_uptodate(bh);
770                 }
771                 unlock_buffer(bh);
772                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
773                 err = ext4_handle_dirty_metadata(handle, inode, bh);
774                 if (unlikely(err))
775                         goto errout;
776         } else
777                 BUFFER_TRACE(bh, "not a new buffer");
778         return bh;
779 errout:
780         brelse(bh);
781         return ERR_PTR(err);
782 }
783
784 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
785                                ext4_lblk_t block, int map_flags)
786 {
787         struct buffer_head *bh;
788
789         bh = ext4_getblk(handle, inode, block, map_flags);
790         if (IS_ERR(bh))
791                 return bh;
792         if (!bh || buffer_uptodate(bh))
793                 return bh;
794         ll_rw_block(READ | REQ_META | REQ_PRIO, 1, &bh);
795         wait_on_buffer(bh);
796         if (buffer_uptodate(bh))
797                 return bh;
798         put_bh(bh);
799         return ERR_PTR(-EIO);
800 }
801
802 int ext4_walk_page_buffers(handle_t *handle,
803                            struct buffer_head *head,
804                            unsigned from,
805                            unsigned to,
806                            int *partial,
807                            int (*fn)(handle_t *handle,
808                                      struct buffer_head *bh))
809 {
810         struct buffer_head *bh;
811         unsigned block_start, block_end;
812         unsigned blocksize = head->b_size;
813         int err, ret = 0;
814         struct buffer_head *next;
815
816         for (bh = head, block_start = 0;
817              ret == 0 && (bh != head || !block_start);
818              block_start = block_end, bh = next) {
819                 next = bh->b_this_page;
820                 block_end = block_start + blocksize;
821                 if (block_end <= from || block_start >= to) {
822                         if (partial && !buffer_uptodate(bh))
823                                 *partial = 1;
824                         continue;
825                 }
826                 err = (*fn)(handle, bh);
827                 if (!ret)
828                         ret = err;
829         }
830         return ret;
831 }
832
833 /*
834  * To preserve ordering, it is essential that the hole instantiation and
835  * the data write be encapsulated in a single transaction.  We cannot
836  * close off a transaction and start a new one between the ext4_get_block()
837  * and the commit_write().  So doing the jbd2_journal_start at the start of
838  * prepare_write() is the right place.
839  *
840  * Also, this function can nest inside ext4_writepage().  In that case, we
841  * *know* that ext4_writepage() has generated enough buffer credits to do the
842  * whole page.  So we won't block on the journal in that case, which is good,
843  * because the caller may be PF_MEMALLOC.
844  *
845  * By accident, ext4 can be reentered when a transaction is open via
846  * quota file writes.  If we were to commit the transaction while thus
847  * reentered, there can be a deadlock - we would be holding a quota
848  * lock, and the commit would never complete if another thread had a
849  * transaction open and was blocking on the quota lock - a ranking
850  * violation.
851  *
852  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
853  * will _not_ run commit under these circumstances because handle->h_ref
854  * is elevated.  We'll still have enough credits for the tiny quotafile
855  * write.
856  */
857 int do_journal_get_write_access(handle_t *handle,
858                                 struct buffer_head *bh)
859 {
860         int dirty = buffer_dirty(bh);
861         int ret;
862
863         if (!buffer_mapped(bh) || buffer_freed(bh))
864                 return 0;
865         /*
866          * __block_write_begin() could have dirtied some buffers. Clean
867          * the dirty bit as jbd2_journal_get_write_access() could complain
868          * otherwise about fs integrity issues. Setting of the dirty bit
869          * by __block_write_begin() isn't a real problem here as we clear
870          * the bit before releasing a page lock and thus writeback cannot
871          * ever write the buffer.
872          */
873         if (dirty)
874                 clear_buffer_dirty(bh);
875         BUFFER_TRACE(bh, "get write access");
876         ret = ext4_journal_get_write_access(handle, bh);
877         if (!ret && dirty)
878                 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
879         return ret;
880 }
881
882 static int ext4_get_block_write_nolock(struct inode *inode, sector_t iblock,
883                    struct buffer_head *bh_result, int create);
884
885 #ifdef CONFIG_EXT4_FS_ENCRYPTION
886 static int ext4_block_write_begin(struct page *page, loff_t pos, unsigned len,
887                                   get_block_t *get_block)
888 {
889         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
890         unsigned to = from + len;
891         struct inode *inode = page->mapping->host;
892         unsigned block_start, block_end;
893         sector_t block;
894         int err = 0;
895         unsigned blocksize = inode->i_sb->s_blocksize;
896         unsigned bbits;
897         struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
898         bool decrypt = false;
899
900         BUG_ON(!PageLocked(page));
901         BUG_ON(from > PAGE_CACHE_SIZE);
902         BUG_ON(to > PAGE_CACHE_SIZE);
903         BUG_ON(from > to);
904
905         if (!page_has_buffers(page))
906                 create_empty_buffers(page, blocksize, 0);
907         head = page_buffers(page);
908         bbits = ilog2(blocksize);
909         block = (sector_t)page->index << (PAGE_CACHE_SHIFT - bbits);
910
911         for (bh = head, block_start = 0; bh != head || !block_start;
912             block++, block_start = block_end, bh = bh->b_this_page) {
913                 block_end = block_start + blocksize;
914                 if (block_end <= from || block_start >= to) {
915                         if (PageUptodate(page)) {
916                                 if (!buffer_uptodate(bh))
917                                         set_buffer_uptodate(bh);
918                         }
919                         continue;
920                 }
921                 if (buffer_new(bh))
922                         clear_buffer_new(bh);
923                 if (!buffer_mapped(bh)) {
924                         WARN_ON(bh->b_size != blocksize);
925                         err = get_block(inode, block, bh, 1);
926                         if (err)
927                                 break;
928                         if (buffer_new(bh)) {
929                                 unmap_underlying_metadata(bh->b_bdev,
930                                                           bh->b_blocknr);
931                                 if (PageUptodate(page)) {
932                                         clear_buffer_new(bh);
933                                         set_buffer_uptodate(bh);
934                                         mark_buffer_dirty(bh);
935                                         continue;
936                                 }
937                                 if (block_end > to || block_start < from)
938                                         zero_user_segments(page, to, block_end,
939                                                            block_start, from);
940                                 continue;
941                         }
942                 }
943                 if (PageUptodate(page)) {
944                         if (!buffer_uptodate(bh))
945                                 set_buffer_uptodate(bh);
946                         continue;
947                 }
948                 if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
949                     !buffer_unwritten(bh) &&
950                     (block_start < from || block_end > to)) {
951                         ll_rw_block(READ, 1, &bh);
952                         *wait_bh++ = bh;
953                         decrypt = ext4_encrypted_inode(inode) &&
954                                 S_ISREG(inode->i_mode);
955                 }
956         }
957         /*
958          * If we issued read requests, let them complete.
959          */
960         while (wait_bh > wait) {
961                 wait_on_buffer(*--wait_bh);
962                 if (!buffer_uptodate(*wait_bh))
963                         err = -EIO;
964         }
965         if (unlikely(err))
966                 page_zero_new_buffers(page, from, to);
967         else if (decrypt)
968                 err = ext4_decrypt_one(inode, page);
969         return err;
970 }
971 #endif
972
973 static int ext4_write_begin(struct file *file, struct address_space *mapping,
974                             loff_t pos, unsigned len, unsigned flags,
975                             struct page **pagep, void **fsdata)
976 {
977         struct inode *inode = mapping->host;
978         int ret, needed_blocks;
979         handle_t *handle;
980         int retries = 0;
981         struct page *page;
982         pgoff_t index;
983         unsigned from, to;
984
985         trace_ext4_write_begin(inode, pos, len, flags);
986         /*
987          * Reserve one block more for addition to orphan list in case
988          * we allocate blocks but write fails for some reason
989          */
990         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
991         index = pos >> PAGE_CACHE_SHIFT;
992         from = pos & (PAGE_CACHE_SIZE - 1);
993         to = from + len;
994
995         if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
996                 ret = ext4_try_to_write_inline_data(mapping, inode, pos, len,
997                                                     flags, pagep);
998                 if (ret < 0)
999                         return ret;
1000                 if (ret == 1)
1001                         return 0;
1002         }
1003
1004         /*
1005          * grab_cache_page_write_begin() can take a long time if the
1006          * system is thrashing due to memory pressure, or if the page
1007          * is being written back.  So grab it first before we start
1008          * the transaction handle.  This also allows us to allocate
1009          * the page (if needed) without using GFP_NOFS.
1010          */
1011 retry_grab:
1012         page = grab_cache_page_write_begin(mapping, index, flags);
1013         if (!page)
1014                 return -ENOMEM;
1015         unlock_page(page);
1016
1017 retry_journal:
1018         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE, needed_blocks);
1019         if (IS_ERR(handle)) {
1020                 page_cache_release(page);
1021                 return PTR_ERR(handle);
1022         }
1023
1024         lock_page(page);
1025         if (page->mapping != mapping) {
1026                 /* The page got truncated from under us */
1027                 unlock_page(page);
1028                 page_cache_release(page);
1029                 ext4_journal_stop(handle);
1030                 goto retry_grab;
1031         }
1032         /* In case writeback began while the page was unlocked */
1033         wait_for_stable_page(page);
1034
1035 #ifdef CONFIG_EXT4_FS_ENCRYPTION
1036         if (ext4_should_dioread_nolock(inode))
1037                 ret = ext4_block_write_begin(page, pos, len,
1038                                              ext4_get_block_write);
1039         else
1040                 ret = ext4_block_write_begin(page, pos, len,
1041                                              ext4_get_block);
1042 #else
1043         if (ext4_should_dioread_nolock(inode))
1044                 ret = __block_write_begin(page, pos, len, ext4_get_block_write);
1045         else
1046                 ret = __block_write_begin(page, pos, len, ext4_get_block);
1047 #endif
1048         if (!ret && ext4_should_journal_data(inode)) {
1049                 ret = ext4_walk_page_buffers(handle, page_buffers(page),
1050                                              from, to, NULL,
1051                                              do_journal_get_write_access);
1052         }
1053
1054         if (ret) {
1055                 unlock_page(page);
1056                 /*
1057                  * __block_write_begin may have instantiated a few blocks
1058                  * outside i_size.  Trim these off again. Don't need
1059                  * i_size_read because we hold i_mutex.
1060                  *
1061                  * Add inode to orphan list in case we crash before
1062                  * truncate finishes
1063                  */
1064                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1065                         ext4_orphan_add(handle, inode);
1066
1067                 ext4_journal_stop(handle);
1068                 if (pos + len > inode->i_size) {
1069                         ext4_truncate_failed_write(inode);
1070                         /*
1071                          * If truncate failed early the inode might
1072                          * still be on the orphan list; we need to
1073                          * make sure the inode is removed from the
1074                          * orphan list in that case.
1075                          */
1076                         if (inode->i_nlink)
1077                                 ext4_orphan_del(NULL, inode);
1078                 }
1079
1080                 if (ret == -ENOSPC &&
1081                     ext4_should_retry_alloc(inode->i_sb, &retries))
1082                         goto retry_journal;
1083                 page_cache_release(page);
1084                 return ret;
1085         }
1086         *pagep = page;
1087         return ret;
1088 }
1089
1090 /* For write_end() in data=journal mode */
1091 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1092 {
1093         int ret;
1094         if (!buffer_mapped(bh) || buffer_freed(bh))
1095                 return 0;
1096         set_buffer_uptodate(bh);
1097         ret = ext4_handle_dirty_metadata(handle, NULL, bh);
1098         clear_buffer_meta(bh);
1099         clear_buffer_prio(bh);
1100         return ret;
1101 }
1102
1103 /*
1104  * We need to pick up the new inode size which generic_commit_write gave us
1105  * `file' can be NULL - eg, when called from page_symlink().
1106  *
1107  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1108  * buffers are managed internally.
1109  */
1110 static int ext4_write_end(struct file *file,
1111                           struct address_space *mapping,
1112                           loff_t pos, unsigned len, unsigned copied,
1113                           struct page *page, void *fsdata)
1114 {
1115         handle_t *handle = ext4_journal_current_handle();
1116         struct inode *inode = mapping->host;
1117         loff_t old_size = inode->i_size;
1118         int ret = 0, ret2;
1119         int i_size_changed = 0;
1120
1121         trace_ext4_write_end(inode, pos, len, copied);
1122         if (ext4_test_inode_state(inode, EXT4_STATE_ORDERED_MODE)) {
1123                 ret = ext4_jbd2_file_inode(handle, inode);
1124                 if (ret) {
1125                         unlock_page(page);
1126                         page_cache_release(page);
1127                         goto errout;
1128                 }
1129         }
1130
1131         if (ext4_has_inline_data(inode)) {
1132                 ret = ext4_write_inline_data_end(inode, pos, len,
1133                                                  copied, page);
1134                 if (ret < 0)
1135                         goto errout;
1136                 copied = ret;
1137         } else
1138                 copied = block_write_end(file, mapping, pos,
1139                                          len, copied, page, fsdata);
1140         /*
1141          * it's important to update i_size while still holding page lock:
1142          * page writeout could otherwise come in and zero beyond i_size.
1143          */
1144         i_size_changed = ext4_update_inode_size(inode, pos + copied);
1145         unlock_page(page);
1146         page_cache_release(page);
1147
1148         if (old_size < pos)
1149                 pagecache_isize_extended(inode, old_size, pos);
1150         /*
1151          * Don't mark the inode dirty under page lock. First, it unnecessarily
1152          * makes the holding time of page lock longer. Second, it forces lock
1153          * ordering of page lock and transaction start for journaling
1154          * filesystems.
1155          */
1156         if (i_size_changed)
1157                 ext4_mark_inode_dirty(handle, inode);
1158
1159         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1160                 /* if we have allocated more blocks and copied
1161                  * less. We will have blocks allocated outside
1162                  * inode->i_size. So truncate them
1163                  */
1164                 ext4_orphan_add(handle, inode);
1165 errout:
1166         ret2 = ext4_journal_stop(handle);
1167         if (!ret)
1168                 ret = ret2;
1169
1170         if (pos + len > inode->i_size) {
1171                 ext4_truncate_failed_write(inode);
1172                 /*
1173                  * If truncate failed early the inode might still be
1174                  * on the orphan list; we need to make sure the inode
1175                  * is removed from the orphan list in that case.
1176                  */
1177                 if (inode->i_nlink)
1178                         ext4_orphan_del(NULL, inode);
1179         }
1180
1181         return ret ? ret : copied;
1182 }
1183
1184 static int ext4_journalled_write_end(struct file *file,
1185                                      struct address_space *mapping,
1186                                      loff_t pos, unsigned len, unsigned copied,
1187                                      struct page *page, void *fsdata)
1188 {
1189         handle_t *handle = ext4_journal_current_handle();
1190         struct inode *inode = mapping->host;
1191         loff_t old_size = inode->i_size;
1192         int ret = 0, ret2;
1193         int partial = 0;
1194         unsigned from, to;
1195         int size_changed = 0;
1196
1197         trace_ext4_journalled_write_end(inode, pos, len, copied);
1198         from = pos & (PAGE_CACHE_SIZE - 1);
1199         to = from + len;
1200
1201         BUG_ON(!ext4_handle_valid(handle));
1202
1203         if (ext4_has_inline_data(inode))
1204                 copied = ext4_write_inline_data_end(inode, pos, len,
1205                                                     copied, page);
1206         else {
1207                 if (copied < len) {
1208                         if (!PageUptodate(page))
1209                                 copied = 0;
1210                         page_zero_new_buffers(page, from+copied, to);
1211                 }
1212
1213                 ret = ext4_walk_page_buffers(handle, page_buffers(page), from,
1214                                              to, &partial, write_end_fn);
1215                 if (!partial)
1216                         SetPageUptodate(page);
1217         }
1218         size_changed = ext4_update_inode_size(inode, pos + copied);
1219         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1220         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1221         unlock_page(page);
1222         page_cache_release(page);
1223
1224         if (old_size < pos)
1225                 pagecache_isize_extended(inode, old_size, pos);
1226
1227         if (size_changed) {
1228                 ret2 = ext4_mark_inode_dirty(handle, inode);
1229                 if (!ret)
1230                         ret = ret2;
1231         }
1232
1233         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1234                 /* if we have allocated more blocks and copied
1235                  * less. We will have blocks allocated outside
1236                  * inode->i_size. So truncate them
1237                  */
1238                 ext4_orphan_add(handle, inode);
1239
1240         ret2 = ext4_journal_stop(handle);
1241         if (!ret)
1242                 ret = ret2;
1243         if (pos + len > inode->i_size) {
1244                 ext4_truncate_failed_write(inode);
1245                 /*
1246                  * If truncate failed early the inode might still be
1247                  * on the orphan list; we need to make sure the inode
1248                  * is removed from the orphan list in that case.
1249                  */
1250                 if (inode->i_nlink)
1251                         ext4_orphan_del(NULL, inode);
1252         }
1253
1254         return ret ? ret : copied;
1255 }
1256
1257 /*
1258  * Reserve space for a single cluster
1259  */
1260 static int ext4_da_reserve_space(struct inode *inode)
1261 {
1262         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1263         struct ext4_inode_info *ei = EXT4_I(inode);
1264         int ret;
1265
1266         /*
1267          * We will charge metadata quota at writeout time; this saves
1268          * us from metadata over-estimation, though we may go over by
1269          * a small amount in the end.  Here we just reserve for data.
1270          */
1271         ret = dquot_reserve_block(inode, EXT4_C2B(sbi, 1));
1272         if (ret)
1273                 return ret;
1274
1275         spin_lock(&ei->i_block_reservation_lock);
1276         if (ext4_claim_free_clusters(sbi, 1, 0)) {
1277                 spin_unlock(&ei->i_block_reservation_lock);
1278                 dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
1279                 return -ENOSPC;
1280         }
1281         ei->i_reserved_data_blocks++;
1282         trace_ext4_da_reserve_space(inode);
1283         spin_unlock(&ei->i_block_reservation_lock);
1284
1285         return 0;       /* success */
1286 }
1287
1288 static void ext4_da_release_space(struct inode *inode, int to_free)
1289 {
1290         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1291         struct ext4_inode_info *ei = EXT4_I(inode);
1292
1293         if (!to_free)
1294                 return;         /* Nothing to release, exit */
1295
1296         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1297
1298         trace_ext4_da_release_space(inode, to_free);
1299         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1300                 /*
1301                  * if there aren't enough reserved blocks, then the
1302                  * counter is messed up somewhere.  Since this
1303                  * function is called from invalidate page, it's
1304                  * harmless to return without any action.
1305                  */
1306                 ext4_warning(inode->i_sb, "ext4_da_release_space: "
1307                          "ino %lu, to_free %d with only %d reserved "
1308                          "data blocks", inode->i_ino, to_free,
1309                          ei->i_reserved_data_blocks);
1310                 WARN_ON(1);
1311                 to_free = ei->i_reserved_data_blocks;
1312         }
1313         ei->i_reserved_data_blocks -= to_free;
1314
1315         /* update fs dirty data blocks counter */
1316         percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1317
1318         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1319
1320         dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1321 }
1322
1323 static void ext4_da_page_release_reservation(struct page *page,
1324                                              unsigned int offset,
1325                                              unsigned int length)
1326 {
1327         int to_release = 0, contiguous_blks = 0;
1328         struct buffer_head *head, *bh;
1329         unsigned int curr_off = 0;
1330         struct inode *inode = page->mapping->host;
1331         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1332         unsigned int stop = offset + length;
1333         int num_clusters;
1334         ext4_fsblk_t lblk;
1335
1336         BUG_ON(stop > PAGE_CACHE_SIZE || stop < length);
1337
1338         head = page_buffers(page);
1339         bh = head;
1340         do {
1341                 unsigned int next_off = curr_off + bh->b_size;
1342
1343                 if (next_off > stop)
1344                         break;
1345
1346                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1347                         to_release++;
1348                         contiguous_blks++;
1349                         clear_buffer_delay(bh);
1350                 } else if (contiguous_blks) {
1351                         lblk = page->index <<
1352                                (PAGE_CACHE_SHIFT - inode->i_blkbits);
1353                         lblk += (curr_off >> inode->i_blkbits) -
1354                                 contiguous_blks;
1355                         ext4_es_remove_extent(inode, lblk, contiguous_blks);
1356                         contiguous_blks = 0;
1357                 }
1358                 curr_off = next_off;
1359         } while ((bh = bh->b_this_page) != head);
1360
1361         if (contiguous_blks) {
1362                 lblk = page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1363                 lblk += (curr_off >> inode->i_blkbits) - contiguous_blks;
1364                 ext4_es_remove_extent(inode, lblk, contiguous_blks);
1365         }
1366
1367         /* If we have released all the blocks belonging to a cluster, then we
1368          * need to release the reserved space for that cluster. */
1369         num_clusters = EXT4_NUM_B2C(sbi, to_release);
1370         while (num_clusters > 0) {
1371                 lblk = (page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits)) +
1372                         ((num_clusters - 1) << sbi->s_cluster_bits);
1373                 if (sbi->s_cluster_ratio == 1 ||
1374                     !ext4_find_delalloc_cluster(inode, lblk))
1375                         ext4_da_release_space(inode, 1);
1376
1377                 num_clusters--;
1378         }
1379 }
1380
1381 /*
1382  * Delayed allocation stuff
1383  */
1384
1385 struct mpage_da_data {
1386         struct inode *inode;
1387         struct writeback_control *wbc;
1388
1389         pgoff_t first_page;     /* The first page to write */
1390         pgoff_t next_page;      /* Current page to examine */
1391         pgoff_t last_page;      /* Last page to examine */
1392         /*
1393          * Extent to map - this can be after first_page because that can be
1394          * fully mapped. We somewhat abuse m_flags to store whether the extent
1395          * is delalloc or unwritten.
1396          */
1397         struct ext4_map_blocks map;
1398         struct ext4_io_submit io_submit;        /* IO submission data */
1399 };
1400
1401 static void mpage_release_unused_pages(struct mpage_da_data *mpd,
1402                                        bool invalidate)
1403 {
1404         int nr_pages, i;
1405         pgoff_t index, end;
1406         struct pagevec pvec;
1407         struct inode *inode = mpd->inode;
1408         struct address_space *mapping = inode->i_mapping;
1409
1410         /* This is necessary when next_page == 0. */
1411         if (mpd->first_page >= mpd->next_page)
1412                 return;
1413
1414         index = mpd->first_page;
1415         end   = mpd->next_page - 1;
1416         if (invalidate) {
1417                 ext4_lblk_t start, last;
1418                 start = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1419                 last = end << (PAGE_CACHE_SHIFT - inode->i_blkbits);
1420                 ext4_es_remove_extent(inode, start, last - start + 1);
1421         }
1422
1423         pagevec_init(&pvec, 0);
1424         while (index <= end) {
1425                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1426                 if (nr_pages == 0)
1427                         break;
1428                 for (i = 0; i < nr_pages; i++) {
1429                         struct page *page = pvec.pages[i];
1430                         if (page->index > end)
1431                                 break;
1432                         BUG_ON(!PageLocked(page));
1433                         BUG_ON(PageWriteback(page));
1434                         if (invalidate) {
1435                                 block_invalidatepage(page, 0, PAGE_CACHE_SIZE);
1436                                 ClearPageUptodate(page);
1437                         }
1438                         unlock_page(page);
1439                 }
1440                 index = pvec.pages[nr_pages - 1]->index + 1;
1441                 pagevec_release(&pvec);
1442         }
1443 }
1444
1445 static void ext4_print_free_blocks(struct inode *inode)
1446 {
1447         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1448         struct super_block *sb = inode->i_sb;
1449         struct ext4_inode_info *ei = EXT4_I(inode);
1450
1451         ext4_msg(sb, KERN_CRIT, "Total free blocks count %lld",
1452                EXT4_C2B(EXT4_SB(inode->i_sb),
1453                         ext4_count_free_clusters(sb)));
1454         ext4_msg(sb, KERN_CRIT, "Free/Dirty block details");
1455         ext4_msg(sb, KERN_CRIT, "free_blocks=%lld",
1456                (long long) EXT4_C2B(EXT4_SB(sb),
1457                 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1458         ext4_msg(sb, KERN_CRIT, "dirty_blocks=%lld",
1459                (long long) EXT4_C2B(EXT4_SB(sb),
1460                 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1461         ext4_msg(sb, KERN_CRIT, "Block reservation details");
1462         ext4_msg(sb, KERN_CRIT, "i_reserved_data_blocks=%u",
1463                  ei->i_reserved_data_blocks);
1464         return;
1465 }
1466
1467 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1468 {
1469         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1470 }
1471
1472 /*
1473  * This function is grabs code from the very beginning of
1474  * ext4_map_blocks, but assumes that the caller is from delayed write
1475  * time. This function looks up the requested blocks and sets the
1476  * buffer delay bit under the protection of i_data_sem.
1477  */
1478 static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1479                               struct ext4_map_blocks *map,
1480                               struct buffer_head *bh)
1481 {
1482         struct extent_status es;
1483         int retval;
1484         sector_t invalid_block = ~((sector_t) 0xffff);
1485 #ifdef ES_AGGRESSIVE_TEST
1486         struct ext4_map_blocks orig_map;
1487
1488         memcpy(&orig_map, map, sizeof(*map));
1489 #endif
1490
1491         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1492                 invalid_block = ~0;
1493
1494         map->m_flags = 0;
1495         ext_debug("ext4_da_map_blocks(): inode %lu, max_blocks %u,"
1496                   "logical block %lu\n", inode->i_ino, map->m_len,
1497                   (unsigned long) map->m_lblk);
1498
1499         /* Lookup extent status tree firstly */
1500         if (ext4_es_lookup_extent(inode, iblock, &es)) {
1501                 if (ext4_es_is_hole(&es)) {
1502                         retval = 0;
1503                         down_read(&EXT4_I(inode)->i_data_sem);
1504                         goto add_delayed;
1505                 }
1506
1507                 /*
1508                  * Delayed extent could be allocated by fallocate.
1509                  * So we need to check it.
1510                  */
1511                 if (ext4_es_is_delayed(&es) && !ext4_es_is_unwritten(&es)) {
1512                         map_bh(bh, inode->i_sb, invalid_block);
1513                         set_buffer_new(bh);
1514                         set_buffer_delay(bh);
1515                         return 0;
1516                 }
1517
1518                 map->m_pblk = ext4_es_pblock(&es) + iblock - es.es_lblk;
1519                 retval = es.es_len - (iblock - es.es_lblk);
1520                 if (retval > map->m_len)
1521                         retval = map->m_len;
1522                 map->m_len = retval;
1523                 if (ext4_es_is_written(&es))
1524                         map->m_flags |= EXT4_MAP_MAPPED;
1525                 else if (ext4_es_is_unwritten(&es))
1526                         map->m_flags |= EXT4_MAP_UNWRITTEN;
1527                 else
1528                         BUG_ON(1);
1529
1530 #ifdef ES_AGGRESSIVE_TEST
1531                 ext4_map_blocks_es_recheck(NULL, inode, map, &orig_map, 0);
1532 #endif
1533                 return retval;
1534         }
1535
1536         /*
1537          * Try to see if we can get the block without requesting a new
1538          * file system block.
1539          */
1540         down_read(&EXT4_I(inode)->i_data_sem);
1541         if (ext4_has_inline_data(inode))
1542                 retval = 0;
1543         else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1544                 retval = ext4_ext_map_blocks(NULL, inode, map, 0);
1545         else
1546                 retval = ext4_ind_map_blocks(NULL, inode, map, 0);
1547
1548 add_delayed:
1549         if (retval == 0) {
1550                 int ret;
1551                 /*
1552                  * XXX: __block_prepare_write() unmaps passed block,
1553                  * is it OK?
1554                  */
1555                 /*
1556                  * If the block was allocated from previously allocated cluster,
1557                  * then we don't need to reserve it again. However we still need
1558                  * to reserve metadata for every block we're going to write.
1559                  */
1560                 if (EXT4_SB(inode->i_sb)->s_cluster_ratio == 1 ||
1561                     !ext4_find_delalloc_cluster(inode, map->m_lblk)) {
1562                         ret = ext4_da_reserve_space(inode);
1563                         if (ret) {
1564                                 /* not enough space to reserve */
1565                                 retval = ret;
1566                                 goto out_unlock;
1567                         }
1568                 }
1569
1570                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1571                                             ~0, EXTENT_STATUS_DELAYED);
1572                 if (ret) {
1573                         retval = ret;
1574                         goto out_unlock;
1575                 }
1576
1577                 map_bh(bh, inode->i_sb, invalid_block);
1578                 set_buffer_new(bh);
1579                 set_buffer_delay(bh);
1580         } else if (retval > 0) {
1581                 int ret;
1582                 unsigned int status;
1583
1584                 if (unlikely(retval != map->m_len)) {
1585                         ext4_warning(inode->i_sb,
1586                                      "ES len assertion failed for inode "
1587                                      "%lu: retval %d != map->m_len %d",
1588                                      inode->i_ino, retval, map->m_len);
1589                         WARN_ON(1);
1590                 }
1591
1592                 status = map->m_flags & EXT4_MAP_UNWRITTEN ?
1593                                 EXTENT_STATUS_UNWRITTEN : EXTENT_STATUS_WRITTEN;
1594                 ret = ext4_es_insert_extent(inode, map->m_lblk, map->m_len,
1595                                             map->m_pblk, status);
1596                 if (ret != 0)
1597                         retval = ret;
1598         }
1599
1600 out_unlock:
1601         up_read((&EXT4_I(inode)->i_data_sem));
1602
1603         return retval;
1604 }
1605
1606 /*
1607  * This is a special get_block_t callback which is used by
1608  * ext4_da_write_begin().  It will either return mapped block or
1609  * reserve space for a single block.
1610  *
1611  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1612  * We also have b_blocknr = -1 and b_bdev initialized properly
1613  *
1614  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1615  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1616  * initialized properly.
1617  */
1618 int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1619                            struct buffer_head *bh, int create)
1620 {
1621         struct ext4_map_blocks map;
1622         int ret = 0;
1623
1624         BUG_ON(create == 0);
1625         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1626
1627         map.m_lblk = iblock;
1628         map.m_len = 1;
1629
1630         /*
1631          * first, we need to know whether the block is allocated already
1632          * preallocated blocks are unmapped but should treated
1633          * the same as allocated blocks.
1634          */
1635         ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1636         if (ret <= 0)
1637                 return ret;
1638
1639         map_bh(bh, inode->i_sb, map.m_pblk);
1640         bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
1641
1642         if (buffer_unwritten(bh)) {
1643                 /* A delayed write to unwritten bh should be marked
1644                  * new and mapped.  Mapped ensures that we don't do
1645                  * get_block multiple times when we write to the same
1646                  * offset and new ensures that we do proper zero out
1647                  * for partial write.
1648                  */
1649                 set_buffer_new(bh);
1650                 set_buffer_mapped(bh);
1651         }
1652         return 0;
1653 }
1654
1655 static int bget_one(handle_t *handle, struct buffer_head *bh)
1656 {
1657         get_bh(bh);
1658         return 0;
1659 }
1660
1661 static int bput_one(handle_t *handle, struct buffer_head *bh)
1662 {
1663         put_bh(bh);
1664         return 0;
1665 }
1666
1667 static int __ext4_journalled_writepage(struct page *page,
1668                                        unsigned int len)
1669 {
1670         struct address_space *mapping = page->mapping;
1671         struct inode *inode = mapping->host;
1672         struct buffer_head *page_bufs = NULL;
1673         handle_t *handle = NULL;
1674         int ret = 0, err = 0;
1675         int inline_data = ext4_has_inline_data(inode);
1676         struct buffer_head *inode_bh = NULL;
1677
1678         ClearPageChecked(page);
1679
1680         if (inline_data) {
1681                 BUG_ON(page->index != 0);
1682                 BUG_ON(len > ext4_get_max_inline_size(inode));
1683                 inode_bh = ext4_journalled_write_inline_data(inode, len, page);
1684                 if (inode_bh == NULL)
1685                         goto out;
1686         } else {
1687                 page_bufs = page_buffers(page);
1688                 if (!page_bufs) {
1689                         BUG();
1690                         goto out;
1691                 }
1692                 ext4_walk_page_buffers(handle, page_bufs, 0, len,
1693                                        NULL, bget_one);
1694         }
1695         /*
1696          * We need to release the page lock before we start the
1697          * journal, so grab a reference so the page won't disappear
1698          * out from under us.
1699          */
1700         get_page(page);
1701         unlock_page(page);
1702
1703         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
1704                                     ext4_writepage_trans_blocks(inode));
1705         if (IS_ERR(handle)) {
1706                 ret = PTR_ERR(handle);
1707                 put_page(page);
1708                 goto out_no_pagelock;
1709         }
1710         BUG_ON(!ext4_handle_valid(handle));
1711
1712         lock_page(page);
1713         put_page(page);
1714         if (page->mapping != mapping) {
1715                 /* The page got truncated from under us */
1716                 ext4_journal_stop(handle);
1717                 ret = 0;
1718                 goto out;
1719         }
1720
1721         if (inline_data) {
1722                 BUFFER_TRACE(inode_bh, "get write access");
1723                 ret = ext4_journal_get_write_access(handle, inode_bh);
1724
1725                 err = ext4_handle_dirty_metadata(handle, inode, inode_bh);
1726
1727         } else {
1728                 ret = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
1729                                              do_journal_get_write_access);
1730
1731                 err = ext4_walk_page_buffers(handle, page_bufs, 0, len, NULL,
1732                                              write_end_fn);
1733         }
1734         if (ret == 0)
1735                 ret = err;
1736         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1737         err = ext4_journal_stop(handle);
1738         if (!ret)
1739                 ret = err;
1740
1741         if (!ext4_has_inline_data(inode))
1742                 ext4_walk_page_buffers(NULL, page_bufs, 0, len,
1743                                        NULL, bput_one);
1744         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1745 out:
1746         unlock_page(page);
1747 out_no_pagelock:
1748         brelse(inode_bh);
1749         return ret;
1750 }
1751
1752 /*
1753  * Note that we don't need to start a transaction unless we're journaling data
1754  * because we should have holes filled from ext4_page_mkwrite(). We even don't
1755  * need to file the inode to the transaction's list in ordered mode because if
1756  * we are writing back data added by write(), the inode is already there and if
1757  * we are writing back data modified via mmap(), no one guarantees in which
1758  * transaction the data will hit the disk. In case we are journaling data, we
1759  * cannot start transaction directly because transaction start ranks above page
1760  * lock so we have to do some magic.
1761  *
1762  * This function can get called via...
1763  *   - ext4_writepages after taking page lock (have journal handle)
1764  *   - journal_submit_inode_data_buffers (no journal handle)
1765  *   - shrink_page_list via the kswapd/direct reclaim (no journal handle)
1766  *   - grab_page_cache when doing write_begin (have journal handle)
1767  *
1768  * We don't do any block allocation in this function. If we have page with
1769  * multiple blocks we need to write those buffer_heads that are mapped. This
1770  * is important for mmaped based write. So if we do with blocksize 1K
1771  * truncate(f, 1024);
1772  * a = mmap(f, 0, 4096);
1773  * a[0] = 'a';
1774  * truncate(f, 4096);
1775  * we have in the page first buffer_head mapped via page_mkwrite call back
1776  * but other buffer_heads would be unmapped but dirty (dirty done via the
1777  * do_wp_page). So writepage should write the first block. If we modify
1778  * the mmap area beyond 1024 we will again get a page_fault and the
1779  * page_mkwrite callback will do the block allocation and mark the
1780  * buffer_heads mapped.
1781  *
1782  * We redirty the page if we have any buffer_heads that is either delay or
1783  * unwritten in the page.
1784  *
1785  * We can get recursively called as show below.
1786  *
1787  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1788  *              ext4_writepage()
1789  *
1790  * But since we don't do any block allocation we should not deadlock.
1791  * Page also have the dirty flag cleared so we don't get recurive page_lock.
1792  */
1793 static int ext4_writepage(struct page *page,
1794                           struct writeback_control *wbc)
1795 {
1796         int ret = 0;
1797         loff_t size;
1798         unsigned int len;
1799         struct buffer_head *page_bufs = NULL;
1800         struct inode *inode = page->mapping->host;
1801         struct ext4_io_submit io_submit;
1802         bool keep_towrite = false;
1803
1804         trace_ext4_writepage(page);
1805         size = i_size_read(inode);
1806         if (page->index == size >> PAGE_CACHE_SHIFT)
1807                 len = size & ~PAGE_CACHE_MASK;
1808         else
1809                 len = PAGE_CACHE_SIZE;
1810
1811         page_bufs = page_buffers(page);
1812         /*
1813          * We cannot do block allocation or other extent handling in this
1814          * function. If there are buffers needing that, we have to redirty
1815          * the page. But we may reach here when we do a journal commit via
1816          * journal_submit_inode_data_buffers() and in that case we must write
1817          * allocated buffers to achieve data=ordered mode guarantees.
1818          */
1819         if (ext4_walk_page_buffers(NULL, page_bufs, 0, len, NULL,
1820                                    ext4_bh_delay_or_unwritten)) {
1821                 redirty_page_for_writepage(wbc, page);
1822                 if (current->flags & PF_MEMALLOC) {
1823                         /*
1824                          * For memory cleaning there's no point in writing only
1825                          * some buffers. So just bail out. Warn if we came here
1826                          * from direct reclaim.
1827                          */
1828                         WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD))
1829                                                         == PF_MEMALLOC);
1830                         unlock_page(page);
1831                         return 0;
1832                 }
1833                 keep_towrite = true;
1834         }
1835
1836         if (PageChecked(page) && ext4_should_journal_data(inode))
1837                 /*
1838                  * It's mmapped pagecache.  Add buffers and journal it.  There
1839                  * doesn't seem much point in redirtying the page here.
1840                  */
1841                 return __ext4_journalled_writepage(page, len);
1842
1843         ext4_io_submit_init(&io_submit, wbc);
1844         io_submit.io_end = ext4_init_io_end(inode, GFP_NOFS);
1845         if (!io_submit.io_end) {
1846                 redirty_page_for_writepage(wbc, page);
1847                 unlock_page(page);
1848                 return -ENOMEM;
1849         }
1850         ret = ext4_bio_write_page(&io_submit, page, len, wbc, keep_towrite);
1851         ext4_io_submit(&io_submit);
1852         /* Drop io_end reference we got from init */
1853         ext4_put_io_end_defer(io_submit.io_end);
1854         return ret;
1855 }
1856
1857 static int mpage_submit_page(struct mpage_da_data *mpd, struct page *page)
1858 {
1859         int len;
1860         loff_t size = i_size_read(mpd->inode);
1861         int err;
1862
1863         BUG_ON(page->index != mpd->first_page);
1864         if (page->index == size >> PAGE_CACHE_SHIFT)
1865                 len = size & ~PAGE_CACHE_MASK;
1866         else
1867                 len = PAGE_CACHE_SIZE;
1868         clear_page_dirty_for_io(page);
1869         err = ext4_bio_write_page(&mpd->io_submit, page, len, mpd->wbc, false);
1870         if (!err)
1871                 mpd->wbc->nr_to_write--;
1872         mpd->first_page++;
1873
1874         return err;
1875 }
1876
1877 #define BH_FLAGS ((1 << BH_Unwritten) | (1 << BH_Delay))
1878
1879 /*
1880  * mballoc gives us at most this number of blocks...
1881  * XXX: That seems to be only a limitation of ext4_mb_normalize_request().
1882  * The rest of mballoc seems to handle chunks up to full group size.
1883  */
1884 #define MAX_WRITEPAGES_EXTENT_LEN 2048
1885
1886 /*
1887  * mpage_add_bh_to_extent - try to add bh to extent of blocks to map
1888  *
1889  * @mpd - extent of blocks
1890  * @lblk - logical number of the block in the file
1891  * @bh - buffer head we want to add to the extent
1892  *
1893  * The function is used to collect contig. blocks in the same state. If the
1894  * buffer doesn't require mapping for writeback and we haven't started the
1895  * extent of buffers to map yet, the function returns 'true' immediately - the
1896  * caller can write the buffer right away. Otherwise the function returns true
1897  * if the block has been added to the extent, false if the block couldn't be
1898  * added.
1899  */
1900 static bool mpage_add_bh_to_extent(struct mpage_da_data *mpd, ext4_lblk_t lblk,
1901                                    struct buffer_head *bh)
1902 {
1903         struct ext4_map_blocks *map = &mpd->map;
1904
1905         /* Buffer that doesn't need mapping for writeback? */
1906         if (!buffer_dirty(bh) || !buffer_mapped(bh) ||
1907             (!buffer_delay(bh) && !buffer_unwritten(bh))) {
1908                 /* So far no extent to map => we write the buffer right away */
1909                 if (map->m_len == 0)
1910                         return true;
1911                 return false;
1912         }
1913
1914         /* First block in the extent? */
1915         if (map->m_len == 0) {
1916                 map->m_lblk = lblk;
1917                 map->m_len = 1;
1918                 map->m_flags = bh->b_state & BH_FLAGS;
1919                 return true;
1920         }
1921
1922         /* Don't go larger than mballoc is willing to allocate */
1923         if (map->m_len >= MAX_WRITEPAGES_EXTENT_LEN)
1924                 return false;
1925
1926         /* Can we merge the block to our big extent? */
1927         if (lblk == map->m_lblk + map->m_len &&
1928             (bh->b_state & BH_FLAGS) == map->m_flags) {
1929                 map->m_len++;
1930                 return true;
1931         }
1932         return false;
1933 }
1934
1935 /*
1936  * mpage_process_page_bufs - submit page buffers for IO or add them to extent
1937  *
1938  * @mpd - extent of blocks for mapping
1939  * @head - the first buffer in the page
1940  * @bh - buffer we should start processing from
1941  * @lblk - logical number of the block in the file corresponding to @bh
1942  *
1943  * Walk through page buffers from @bh upto @head (exclusive) and either submit
1944  * the page for IO if all buffers in this page were mapped and there's no
1945  * accumulated extent of buffers to map or add buffers in the page to the
1946  * extent of buffers to map. The function returns 1 if the caller can continue
1947  * by processing the next page, 0 if it should stop adding buffers to the
1948  * extent to map because we cannot extend it anymore. It can also return value
1949  * < 0 in case of error during IO submission.
1950  */
1951 static int mpage_process_page_bufs(struct mpage_da_data *mpd,
1952                                    struct buffer_head *head,
1953                                    struct buffer_head *bh,
1954                                    ext4_lblk_t lblk)
1955 {
1956         struct inode *inode = mpd->inode;
1957         int err;
1958         ext4_lblk_t blocks = (i_size_read(inode) + (1 << inode->i_blkbits) - 1)
1959                                                         >> inode->i_blkbits;
1960
1961         do {
1962                 BUG_ON(buffer_locked(bh));
1963
1964                 if (lblk >= blocks || !mpage_add_bh_to_extent(mpd, lblk, bh)) {
1965                         /* Found extent to map? */
1966                         if (mpd->map.m_len)
1967                                 return 0;
1968                         /* Everything mapped so far and we hit EOF */
1969                         break;
1970                 }
1971         } while (lblk++, (bh = bh->b_this_page) != head);
1972         /* So far everything mapped? Submit the page for IO. */
1973         if (mpd->map.m_len == 0) {
1974                 err = mpage_submit_page(mpd, head->b_page);
1975                 if (err < 0)
1976                         return err;
1977         }
1978         return lblk < blocks;
1979 }
1980
1981 /*
1982  * mpage_map_buffers - update buffers corresponding to changed extent and
1983  *                     submit fully mapped pages for IO
1984  *
1985  * @mpd - description of extent to map, on return next extent to map
1986  *
1987  * Scan buffers corresponding to changed extent (we expect corresponding pages
1988  * to be already locked) and update buffer state according to new extent state.
1989  * We map delalloc buffers to their physical location, clear unwritten bits,
1990  * and mark buffers as uninit when we perform writes to unwritten extents
1991  * and do extent conversion after IO is finished. If the last page is not fully
1992  * mapped, we update @map to the next extent in the last page that needs
1993  * mapping. Otherwise we submit the page for IO.
1994  */
1995 static int mpage_map_and_submit_buffers(struct mpage_da_data *mpd)
1996 {
1997         struct pagevec pvec;
1998         int nr_pages, i;
1999         struct inode *inode = mpd->inode;
2000         struct buffer_head *head, *bh;
2001         int bpp_bits = PAGE_CACHE_SHIFT - inode->i_blkbits;
2002         pgoff_t start, end;
2003         ext4_lblk_t lblk;
2004         sector_t pblock;
2005         int err;
2006
2007         start = mpd->map.m_lblk >> bpp_bits;
2008         end = (mpd->map.m_lblk + mpd->map.m_len - 1) >> bpp_bits;
2009         lblk = start << bpp_bits;
2010         pblock = mpd->map.m_pblk;
2011
2012         pagevec_init(&pvec, 0);
2013         while (start <= end) {
2014                 nr_pages = pagevec_lookup(&pvec, inode->i_mapping, start,
2015                                           PAGEVEC_SIZE);
2016                 if (nr_pages == 0)
2017                         break;
2018                 for (i = 0; i < nr_pages; i++) {
2019                         struct page *page = pvec.pages[i];
2020
2021                         if (page->index > end)
2022                                 break;
2023                         /* Up to 'end' pages must be contiguous */
2024                         BUG_ON(page->index != start);
2025                         bh = head = page_buffers(page);
2026                         do {
2027                                 if (lblk < mpd->map.m_lblk)
2028                                         continue;
2029                                 if (lblk >= mpd->map.m_lblk + mpd->map.m_len) {
2030                                         /*
2031                                          * Buffer after end of mapped extent.
2032                                          * Find next buffer in the page to map.
2033                                          */
2034                                         mpd->map.m_len = 0;
2035                                         mpd->map.m_flags = 0;
2036                                         /*
2037                                          * FIXME: If dioread_nolock supports
2038                                          * blocksize < pagesize, we need to make
2039                                          * sure we add size mapped so far to
2040                                          * io_end->size as the following call
2041                                          * can submit the page for IO.
2042                                          */
2043                                         err = mpage_process_page_bufs(mpd, head,
2044                                                                       bh, lblk);
2045                                         pagevec_release(&pvec);
2046                                         if (err > 0)
2047                                                 err = 0;
2048                                         return err;
2049                                 }
2050                                 if (buffer_delay(bh)) {
2051                                         clear_buffer_delay(bh);
2052                                         bh->b_blocknr = pblock++;
2053                                 }
2054                                 clear_buffer_unwritten(bh);
2055                         } while (lblk++, (bh = bh->b_this_page) != head);
2056
2057                         /*
2058                          * FIXME: This is going to break if dioread_nolock
2059                          * supports blocksize < pagesize as we will try to
2060                          * convert potentially unmapped parts of inode.
2061                          */
2062                         mpd->io_submit.io_end->size += PAGE_CACHE_SIZE;
2063                         /* Page fully mapped - let IO run! */
2064                         err = mpage_submit_page(mpd, page);
2065                         if (err < 0) {
2066                                 pagevec_release(&pvec);
2067                                 return err;
2068                         }
2069                         start++;
2070                 }
2071                 pagevec_release(&pvec);
2072         }
2073         /* Extent fully mapped and matches with page boundary. We are done. */
2074         mpd->map.m_len = 0;
2075         mpd->map.m_flags = 0;
2076         return 0;
2077 }
2078
2079 static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
2080 {
2081         struct inode *inode = mpd->inode;
2082         struct ext4_map_blocks *map = &mpd->map;
2083         int get_blocks_flags;
2084         int err, dioread_nolock;
2085
2086         trace_ext4_da_write_pages_extent(inode, map);
2087         /*
2088          * Call ext4_map_blocks() to allocate any delayed allocation blocks, or
2089          * to convert an unwritten extent to be initialized (in the case
2090          * where we have written into one or more preallocated blocks).  It is
2091          * possible that we're going to need more metadata blocks than
2092          * previously reserved. However we must not fail because we're in
2093          * writeback and there is nothing we can do about it so it might result
2094          * in data loss.  So use reserved blocks to allocate metadata if
2095          * possible.
2096          *
2097          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE if
2098          * the blocks in question are delalloc blocks.  This indicates
2099          * that the blocks and quotas has already been checked when
2100          * the data was copied into the page cache.
2101          */
2102         get_blocks_flags = EXT4_GET_BLOCKS_CREATE |
2103                            EXT4_GET_BLOCKS_METADATA_NOFAIL;
2104         dioread_nolock = ext4_should_dioread_nolock(inode);
2105         if (dioread_nolock)
2106                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
2107         if (map->m_flags & (1 << BH_Delay))
2108                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
2109
2110         err = ext4_map_blocks(handle, inode, map, get_blocks_flags);
2111         if (err < 0)
2112                 return err;
2113         if (dioread_nolock && (map->m_flags & EXT4_MAP_UNWRITTEN)) {
2114                 if (!mpd->io_submit.io_end->handle &&
2115                     ext4_handle_valid(handle)) {
2116                         mpd->io_submit.io_end->handle = handle->h_rsv_handle;
2117                         handle->h_rsv_handle = NULL;
2118                 }
2119                 ext4_set_io_unwritten_flag(inode, mpd->io_submit.io_end);
2120         }
2121
2122         BUG_ON(map->m_len == 0);
2123         if (map->m_flags & EXT4_MAP_NEW) {
2124                 struct block_device *bdev = inode->i_sb->s_bdev;
2125                 int i;
2126
2127                 for (i = 0; i < map->m_len; i++)
2128                         unmap_underlying_metadata(bdev, map->m_pblk + i);
2129         }
2130         return 0;
2131 }
2132
2133 /*
2134  * mpage_map_and_submit_extent - map extent starting at mpd->lblk of length
2135  *                               mpd->len and submit pages underlying it for IO
2136  *
2137  * @handle - handle for journal operations
2138  * @mpd - extent to map
2139  * @give_up_on_write - we set this to true iff there is a fatal error and there
2140  *                     is no hope of writing the data. The caller should discard
2141  *                     dirty pages to avoid infinite loops.
2142  *
2143  * The function maps extent starting at mpd->lblk of length mpd->len. If it is
2144  * delayed, blocks are allocated, if it is unwritten, we may need to convert
2145  * them to initialized or split the described range from larger unwritten
2146  * extent. Note that we need not map all the described range since allocation
2147  * can return less blocks or the range is covered by more unwritten extents. We
2148  * cannot map more because we are limited by reserved transaction credits. On
2149  * the other hand we always make sure that the last touched page is fully
2150  * mapped so that it can be written out (and thus forward progress is
2151  * guaranteed). After mapping we submit all mapped pages for IO.
2152  */
2153 static int mpage_map_and_submit_extent(handle_t *handle,
2154                                        struct mpage_da_data *mpd,
2155                                        bool *give_up_on_write)
2156 {
2157         struct inode *inode = mpd->inode;
2158         struct ext4_map_blocks *map = &mpd->map;
2159         int err;
2160         loff_t disksize;
2161         int progress = 0;
2162
2163         mpd->io_submit.io_end->offset =
2164                                 ((loff_t)map->m_lblk) << inode->i_blkbits;
2165         do {
2166                 err = mpage_map_one_extent(handle, mpd);
2167                 if (err < 0) {
2168                         struct super_block *sb = inode->i_sb;
2169
2170                         if (EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)
2171                                 goto invalidate_dirty_pages;
2172                         /*
2173                          * Let the uper layers retry transient errors.
2174                          * In the case of ENOSPC, if ext4_count_free_blocks()
2175                          * is non-zero, a commit should free up blocks.
2176                          */
2177                         if ((err == -ENOMEM) ||
2178                             (err == -ENOSPC && ext4_count_free_clusters(sb))) {
2179                                 if (progress)
2180                                         goto update_disksize;
2181                                 return err;
2182                         }
2183                         ext4_msg(sb, KERN_CRIT,
2184                                  "Delayed block allocation failed for "
2185                                  "inode %lu at logical offset %llu with"
2186                                  " max blocks %u with error %d",
2187                                  inode->i_ino,
2188                                  (unsigned long long)map->m_lblk,
2189                                  (unsigned)map->m_len, -err);
2190                         ext4_msg(sb, KERN_CRIT,
2191                                  "This should not happen!! Data will "
2192                                  "be lost\n");
2193                         if (err == -ENOSPC)
2194                                 ext4_print_free_blocks(inode);
2195                 invalidate_dirty_pages:
2196                         *give_up_on_write = true;
2197                         return err;
2198                 }
2199                 progress = 1;
2200                 /*
2201                  * Update buffer state, submit mapped pages, and get us new
2202                  * extent to map
2203                  */
2204                 err = mpage_map_and_submit_buffers(mpd);
2205                 if (err < 0)
2206                         goto update_disksize;
2207         } while (map->m_len);
2208
2209 update_disksize:
2210         /*
2211          * Update on-disk size after IO is submitted.  Races with
2212          * truncate are avoided by checking i_size under i_data_sem.
2213          */
2214         disksize = ((loff_t)mpd->first_page) << PAGE_CACHE_SHIFT;
2215         if (disksize > EXT4_I(inode)->i_disksize) {
2216                 int err2;
2217                 loff_t i_size;
2218
2219                 down_write(&EXT4_I(inode)->i_data_sem);
2220                 i_size = i_size_read(inode);
2221                 if (disksize > i_size)
2222                         disksize = i_size;
2223                 if (disksize > EXT4_I(inode)->i_disksize)
2224                         EXT4_I(inode)->i_disksize = disksize;
2225                 err2 = ext4_mark_inode_dirty(handle, inode);
2226                 up_write(&EXT4_I(inode)->i_data_sem);
2227                 if (err2)
2228                         ext4_error(inode->i_sb,
2229                                    "Failed to mark inode %lu dirty",
2230                                    inode->i_ino);
2231                 if (!err)
2232                         err = err2;
2233         }
2234         return err;
2235 }
2236
2237 /*
2238  * Calculate the total number of credits to reserve for one writepages
2239  * iteration. This is called from ext4_writepages(). We map an extent of
2240  * up to MAX_WRITEPAGES_EXTENT_LEN blocks and then we go on and finish mapping
2241  * the last partial page. So in total we can map MAX_WRITEPAGES_EXTENT_LEN +
2242  * bpp - 1 blocks in bpp different extents.
2243  */
2244 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2245 {
2246         int bpp = ext4_journal_blocks_per_page(inode);
2247
2248         return ext4_meta_trans_blocks(inode,
2249                                 MAX_WRITEPAGES_EXTENT_LEN + bpp - 1, bpp);
2250 }
2251
2252 /*
2253  * mpage_prepare_extent_to_map - find & lock contiguous range of dirty pages
2254  *                               and underlying extent to map
2255  *
2256  * @mpd - where to look for pages
2257  *
2258  * Walk dirty pages in the mapping. If they are fully mapped, submit them for
2259  * IO immediately. When we find a page which isn't mapped we start accumulating
2260  * extent of buffers underlying these pages that needs mapping (formed by
2261  * either delayed or unwritten buffers). We also lock the pages containing
2262  * these buffers. The extent found is returned in @mpd structure (starting at
2263  * mpd->lblk with length mpd->len blocks).
2264  *
2265  * Note that this function can attach bios to one io_end structure which are
2266  * neither logically nor physically contiguous. Although it may seem as an
2267  * unnecessary complication, it is actually inevitable in blocksize < pagesize
2268  * case as we need to track IO to all buffers underlying a page in one io_end.
2269  */
2270 static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
2271 {
2272         struct address_space *mapping = mpd->inode->i_mapping;
2273         struct pagevec pvec;
2274         unsigned int nr_pages;
2275         long left = mpd->wbc->nr_to_write;
2276         pgoff_t index = mpd->first_page;
2277         pgoff_t end = mpd->last_page;
2278         int tag;
2279         int i, err = 0;
2280         int blkbits = mpd->inode->i_blkbits;
2281         ext4_lblk_t lblk;
2282         struct buffer_head *head;
2283
2284         if (mpd->wbc->sync_mode == WB_SYNC_ALL || mpd->wbc->tagged_writepages)
2285                 tag = PAGECACHE_TAG_TOWRITE;
2286         else
2287                 tag = PAGECACHE_TAG_DIRTY;
2288
2289         pagevec_init(&pvec, 0);
2290         mpd->map.m_len = 0;
2291         mpd->next_page = index;
2292         while (index <= end) {
2293                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2294                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2295                 if (nr_pages == 0)
2296                         goto out;
2297
2298                 for (i = 0; i < nr_pages; i++) {
2299                         struct page *page = pvec.pages[i];
2300
2301                         /*
2302                          * At this point, the page may be truncated or
2303                          * invalidated (changing page->mapping to NULL), or
2304                          * even swizzled back from swapper_space to tmpfs file
2305                          * mapping. However, page->index will not change
2306                          * because we have a reference on the page.
2307                          */
2308                         if (page->index > end)
2309                                 goto out;
2310
2311                         /*
2312                          * Accumulated enough dirty pages? This doesn't apply
2313                          * to WB_SYNC_ALL mode. For integrity sync we have to
2314                          * keep going because someone may be concurrently
2315                          * dirtying pages, and we might have synced a lot of
2316                          * newly appeared dirty pages, but have not synced all
2317                          * of the old dirty pages.
2318                          */
2319                         if (mpd->wbc->sync_mode == WB_SYNC_NONE && left <= 0)
2320                                 goto out;
2321
2322                         /* If we can't merge this page, we are done. */
2323                         if (mpd->map.m_len > 0 && mpd->next_page != page->index)
2324                                 goto out;
2325
2326                         lock_page(page);
2327                         /*
2328                          * If the page is no longer dirty, or its mapping no
2329                          * longer corresponds to inode we are writing (which
2330                          * means it has been truncated or invalidated), or the
2331                          * page is already under writeback and we are not doing
2332                          * a data integrity writeback, skip the page
2333                          */
2334                         if (!PageDirty(page) ||
2335                             (PageWriteback(page) &&
2336                              (mpd->wbc->sync_mode == WB_SYNC_NONE)) ||
2337                             unlikely(page->mapping != mapping)) {
2338                                 unlock_page(page);
2339                                 continue;
2340                         }
2341
2342                         wait_on_page_writeback(page);
2343                         BUG_ON(PageWriteback(page));
2344
2345                         if (mpd->map.m_len == 0)
2346                                 mpd->first_page = page->index;
2347                         mpd->next_page = page->index + 1;
2348                         /* Add all dirty buffers to mpd */
2349                         lblk = ((ext4_lblk_t)page->index) <<
2350                                 (PAGE_CACHE_SHIFT - blkbits);
2351                         head = page_buffers(page);
2352                         err = mpage_process_page_bufs(mpd, head, head, lblk);
2353                         if (err <= 0)
2354                                 goto out;
2355                         err = 0;
2356                         left--;
2357                 }
2358                 pagevec_release(&pvec);
2359                 cond_resched();
2360         }
2361         return 0;
2362 out:
2363         pagevec_release(&pvec);
2364         return err;
2365 }
2366
2367 static int __writepage(struct page *page, struct writeback_control *wbc,
2368                        void *data)
2369 {
2370         struct address_space *mapping = data;
2371         int ret = ext4_writepage(page, wbc);
2372         mapping_set_error(mapping, ret);
2373         return ret;
2374 }
2375
2376 static int ext4_writepages(struct address_space *mapping,
2377                            struct writeback_control *wbc)
2378 {
2379         pgoff_t writeback_index = 0;
2380         long nr_to_write = wbc->nr_to_write;
2381         int range_whole = 0;
2382         int cycled = 1;
2383         handle_t *handle = NULL;
2384         struct mpage_da_data mpd;
2385         struct inode *inode = mapping->host;
2386         int needed_blocks, rsv_blocks = 0, ret = 0;
2387         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2388         bool done;
2389         struct blk_plug plug;
2390         bool give_up_on_write = false;
2391
2392         trace_ext4_writepages(inode, wbc);
2393
2394         /*
2395          * No pages to write? This is mainly a kludge to avoid starting
2396          * a transaction for special inodes like journal inode on last iput()
2397          * because that could violate lock ordering on umount
2398          */
2399         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2400                 goto out_writepages;
2401
2402         if (ext4_should_journal_data(inode)) {
2403                 struct blk_plug plug;
2404
2405                 blk_start_plug(&plug);
2406                 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2407                 blk_finish_plug(&plug);
2408                 goto out_writepages;
2409         }
2410
2411         /*
2412          * If the filesystem has aborted, it is read-only, so return
2413          * right away instead of dumping stack traces later on that
2414          * will obscure the real source of the problem.  We test
2415          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2416          * the latter could be true if the filesystem is mounted
2417          * read-only, and in that case, ext4_writepages should
2418          * *never* be called, so if that ever happens, we would want
2419          * the stack trace.
2420          */
2421         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED)) {
2422                 ret = -EROFS;
2423                 goto out_writepages;
2424         }
2425
2426         if (ext4_should_dioread_nolock(inode)) {
2427                 /*
2428                  * We may need to convert up to one extent per block in
2429                  * the page and we may dirty the inode.
2430                  */
2431                 rsv_blocks = 1 + (PAGE_CACHE_SIZE >> inode->i_blkbits);
2432         }
2433
2434         /*
2435          * If we have inline data and arrive here, it means that
2436          * we will soon create the block for the 1st page, so
2437          * we'd better clear the inline data here.
2438          */
2439         if (ext4_has_inline_data(inode)) {
2440                 /* Just inode will be modified... */
2441                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
2442                 if (IS_ERR(handle)) {
2443                         ret = PTR_ERR(handle);
2444                         goto out_writepages;
2445                 }
2446                 BUG_ON(ext4_test_inode_state(inode,
2447                                 EXT4_STATE_MAY_INLINE_DATA));
2448                 ext4_destroy_inline_data(handle, inode);
2449                 ext4_journal_stop(handle);
2450         }
2451
2452         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2453                 range_whole = 1;
2454
2455         if (wbc->range_cyclic) {
2456                 writeback_index = mapping->writeback_index;
2457                 if (writeback_index)
2458                         cycled = 0;
2459                 mpd.first_page = writeback_index;
2460                 mpd.last_page = -1;
2461         } else {
2462                 mpd.first_page = wbc->range_start >> PAGE_CACHE_SHIFT;
2463                 mpd.last_page = wbc->range_end >> PAGE_CACHE_SHIFT;
2464         }
2465
2466         mpd.inode = inode;
2467         mpd.wbc = wbc;
2468         ext4_io_submit_init(&mpd.io_submit, wbc);
2469 retry:
2470         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2471                 tag_pages_for_writeback(mapping, mpd.first_page, mpd.last_page);
2472         done = false;
2473         blk_start_plug(&plug);
2474         while (!done && mpd.first_page <= mpd.last_page) {
2475                 /* For each extent of pages we use new io_end */
2476                 mpd.io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
2477                 if (!mpd.io_submit.io_end) {
2478                         ret = -ENOMEM;
2479                         break;
2480                 }
2481
2482                 /*
2483                  * We have two constraints: We find one extent to map and we
2484                  * must always write out whole page (makes a difference when
2485                  * blocksize < pagesize) so that we don't block on IO when we
2486                  * try to write out the rest of the page. Journalled mode is
2487                  * not supported by delalloc.
2488                  */
2489                 BUG_ON(ext4_should_journal_data(inode));
2490                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2491
2492                 /* start a new transaction */
2493                 handle = ext4_journal_start_with_reserve(inode,
2494                                 EXT4_HT_WRITE_PAGE, needed_blocks, rsv_blocks);
2495                 if (IS_ERR(handle)) {
2496                         ret = PTR_ERR(handle);
2497                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2498                                "%ld pages, ino %lu; err %d", __func__,
2499                                 wbc->nr_to_write, inode->i_ino, ret);
2500                         /* Release allocated io_end */
2501                         ext4_put_io_end(mpd.io_submit.io_end);
2502                         break;
2503                 }
2504
2505                 trace_ext4_da_write_pages(inode, mpd.first_page, mpd.wbc);
2506                 ret = mpage_prepare_extent_to_map(&mpd);
2507                 if (!ret) {
2508                         if (mpd.map.m_len)
2509                                 ret = mpage_map_and_submit_extent(handle, &mpd,
2510                                         &give_up_on_write);
2511                         else {
2512                                 /*
2513                                  * We scanned the whole range (or exhausted
2514                                  * nr_to_write), submitted what was mapped and
2515                                  * didn't find anything needing mapping. We are
2516                                  * done.
2517                                  */
2518                                 done = true;
2519                         }
2520                 }
2521                 ext4_journal_stop(handle);
2522                 /* Submit prepared bio */
2523                 ext4_io_submit(&mpd.io_submit);
2524                 /* Unlock pages we didn't use */
2525                 mpage_release_unused_pages(&mpd, give_up_on_write);
2526                 /* Drop our io_end reference we got from init */
2527                 ext4_put_io_end(mpd.io_submit.io_end);
2528
2529                 if (ret == -ENOSPC && sbi->s_journal) {
2530                         /*
2531                          * Commit the transaction which would
2532                          * free blocks released in the transaction
2533                          * and try again
2534                          */
2535                         jbd2_journal_force_commit_nested(sbi->s_journal);
2536                         ret = 0;
2537                         continue;
2538                 }
2539                 /* Fatal error - ENOMEM, EIO... */
2540                 if (ret)
2541                         break;
2542         }
2543         blk_finish_plug(&plug);
2544         if (!ret && !cycled && wbc->nr_to_write > 0) {
2545                 cycled = 1;
2546                 mpd.last_page = writeback_index - 1;
2547                 mpd.first_page = 0;
2548                 goto retry;
2549         }
2550
2551         /* Update index */
2552         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2553                 /*
2554                  * Set the writeback_index so that range_cyclic
2555                  * mode will write it back later
2556                  */
2557                 mapping->writeback_index = mpd.first_page;
2558
2559 out_writepages:
2560         trace_ext4_writepages_result(inode, wbc, ret,
2561                                      nr_to_write - wbc->nr_to_write);
2562         return ret;
2563 }
2564
2565 static int ext4_nonda_switch(struct super_block *sb)
2566 {
2567         s64 free_clusters, dirty_clusters;
2568         struct ext4_sb_info *sbi = EXT4_SB(sb);
2569
2570         /*
2571          * switch to non delalloc mode if we are running low
2572          * on free block. The free block accounting via percpu
2573          * counters can get slightly wrong with percpu_counter_batch getting
2574          * accumulated on each CPU without updating global counters
2575          * Delalloc need an accurate free block accounting. So switch
2576          * to non delalloc when we are near to error range.
2577          */
2578         free_clusters =
2579                 percpu_counter_read_positive(&sbi->s_freeclusters_counter);
2580         dirty_clusters =
2581                 percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
2582         /*
2583          * Start pushing delalloc when 1/2 of free blocks are dirty.
2584          */
2585         if (dirty_clusters && (free_clusters < 2 * dirty_clusters))
2586                 try_to_writeback_inodes_sb(sb, WB_REASON_FS_FREE_SPACE);
2587
2588         if (2 * free_clusters < 3 * dirty_clusters ||
2589             free_clusters < (dirty_clusters + EXT4_FREECLUSTERS_WATERMARK)) {
2590                 /*
2591                  * free block count is less than 150% of dirty blocks
2592                  * or free blocks is less than watermark
2593                  */
2594                 return 1;
2595         }
2596         return 0;
2597 }
2598
2599 /* We always reserve for an inode update; the superblock could be there too */
2600 static int ext4_da_write_credits(struct inode *inode, loff_t pos, unsigned len)
2601 {
2602         if (likely(EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
2603                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE)))
2604                 return 1;
2605
2606         if (pos + len <= 0x7fffffffULL)
2607                 return 1;
2608
2609         /* We might need to update the superblock to set LARGE_FILE */
2610         return 2;
2611 }
2612
2613 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2614                                loff_t pos, unsigned len, unsigned flags,
2615                                struct page **pagep, void **fsdata)
2616 {
2617         int ret, retries = 0;
2618         struct page *page;
2619         pgoff_t index;
2620         struct inode *inode = mapping->host;
2621         handle_t *handle;
2622
2623         index = pos >> PAGE_CACHE_SHIFT;
2624
2625         if (ext4_nonda_switch(inode->i_sb)) {
2626                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2627                 return ext4_write_begin(file, mapping, pos,
2628                                         len, flags, pagep, fsdata);
2629         }
2630         *fsdata = (void *)0;
2631         trace_ext4_da_write_begin(inode, pos, len, flags);
2632
2633         if (ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA)) {
2634                 ret = ext4_da_write_inline_data_begin(mapping, inode,
2635                                                       pos, len, flags,
2636                                                       pagep, fsdata);
2637                 if (ret < 0)
2638                         return ret;
2639                 if (ret == 1)
2640                         return 0;
2641         }
2642
2643         /*
2644          * grab_cache_page_write_begin() can take a long time if the
2645          * system is thrashing due to memory pressure, or if the page
2646          * is being written back.  So grab it first before we start
2647          * the transaction handle.  This also allows us to allocate
2648          * the page (if needed) without using GFP_NOFS.
2649          */
2650 retry_grab:
2651         page = grab_cache_page_write_begin(mapping, index, flags);
2652         if (!page)
2653                 return -ENOMEM;
2654         unlock_page(page);
2655
2656         /*
2657          * With delayed allocation, we don't log the i_disksize update
2658          * if there is delayed block allocation. But we still need
2659          * to journalling the i_disksize update if writes to the end
2660          * of file which has an already mapped buffer.
2661          */
2662 retry_journal:
2663         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
2664                                 ext4_da_write_credits(inode, pos, len));
2665         if (IS_ERR(handle)) {
2666                 page_cache_release(page);
2667                 return PTR_ERR(handle);
2668         }
2669
2670         lock_page(page);
2671         if (page->mapping != mapping) {
2672                 /* The page got truncated from under us */
2673                 unlock_page(page);
2674                 page_cache_release(page);
2675                 ext4_journal_stop(handle);
2676                 goto retry_grab;
2677         }
2678         /* In case writeback began while the page was unlocked */
2679         wait_for_stable_page(page);
2680
2681 #ifdef CONFIG_EXT4_FS_ENCRYPTION
2682         ret = ext4_block_write_begin(page, pos, len,
2683                                      ext4_da_get_block_prep);
2684 #else
2685         ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
2686 #endif
2687         if (ret < 0) {
2688                 unlock_page(page);
2689                 ext4_journal_stop(handle);
2690                 /*
2691                  * block_write_begin may have instantiated a few blocks
2692                  * outside i_size.  Trim these off again. Don't need
2693                  * i_size_read because we hold i_mutex.
2694                  */
2695                 if (pos + len > inode->i_size)
2696                         ext4_truncate_failed_write(inode);
2697
2698                 if (ret == -ENOSPC &&
2699                     ext4_should_retry_alloc(inode->i_sb, &retries))
2700                         goto retry_journal;
2701
2702                 page_cache_release(page);
2703                 return ret;
2704         }
2705
2706         *pagep = page;
2707         return ret;
2708 }
2709
2710 /*
2711  * Check if we should update i_disksize
2712  * when write to the end of file but not require block allocation
2713  */
2714 static int ext4_da_should_update_i_disksize(struct page *page,
2715                                             unsigned long offset)
2716 {
2717         struct buffer_head *bh;
2718         struct inode *inode = page->mapping->host;
2719         unsigned int idx;
2720         int i;
2721
2722         bh = page_buffers(page);
2723         idx = offset >> inode->i_blkbits;
2724
2725         for (i = 0; i < idx; i++)
2726                 bh = bh->b_this_page;
2727
2728         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
2729                 return 0;
2730         return 1;
2731 }
2732
2733 static int ext4_da_write_end(struct file *file,
2734                              struct address_space *mapping,
2735                              loff_t pos, unsigned len, unsigned copied,
2736                              struct page *page, void *fsdata)
2737 {
2738         struct inode *inode = mapping->host;
2739         int ret = 0, ret2;
2740         handle_t *handle = ext4_journal_current_handle();
2741         loff_t new_i_size;
2742         unsigned long start, end;
2743         int write_mode = (int)(unsigned long)fsdata;
2744
2745         if (write_mode == FALL_BACK_TO_NONDELALLOC)
2746                 return ext4_write_end(file, mapping, pos,
2747                                       len, copied, page, fsdata);
2748
2749         trace_ext4_da_write_end(inode, pos, len, copied);
2750         start = pos & (PAGE_CACHE_SIZE - 1);
2751         end = start + copied - 1;
2752
2753         /*
2754          * generic_write_end() will run mark_inode_dirty() if i_size
2755          * changes.  So let's piggyback the i_disksize mark_inode_dirty
2756          * into that.
2757          */
2758         new_i_size = pos + copied;
2759         if (copied && new_i_size > EXT4_I(inode)->i_disksize) {
2760                 if (ext4_has_inline_data(inode) ||
2761                     ext4_da_should_update_i_disksize(page, end)) {
2762                         ext4_update_i_disksize(inode, new_i_size);
2763                         /* We need to mark inode dirty even if
2764                          * new_i_size is less that inode->i_size
2765                          * bu greater than i_disksize.(hint delalloc)
2766                          */
2767                         ext4_mark_inode_dirty(handle, inode);
2768                 }
2769         }
2770
2771         if (write_mode != CONVERT_INLINE_DATA &&
2772             ext4_test_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA) &&
2773             ext4_has_inline_data(inode))
2774                 ret2 = ext4_da_write_inline_data_end(inode, pos, len, copied,
2775                                                      page);
2776         else
2777                 ret2 = generic_write_end(file, mapping, pos, len, copied,
2778                                                         page, fsdata);
2779
2780         copied = ret2;
2781         if (ret2 < 0)
2782                 ret = ret2;
2783         ret2 = ext4_journal_stop(handle);
2784         if (!ret)
2785                 ret = ret2;
2786
2787         return ret ? ret : copied;
2788 }
2789
2790 static void ext4_da_invalidatepage(struct page *page, unsigned int offset,
2791                                    unsigned int length)
2792 {
2793         /*
2794          * Drop reserved blocks
2795          */
2796         BUG_ON(!PageLocked(page));
2797         if (!page_has_buffers(page))
2798                 goto out;
2799
2800         ext4_da_page_release_reservation(page, offset, length);
2801
2802 out:
2803         ext4_invalidatepage(page, offset, length);
2804
2805         return;
2806 }
2807
2808 /*
2809  * Force all delayed allocation blocks to be allocated for a given inode.
2810  */
2811 int ext4_alloc_da_blocks(struct inode *inode)
2812 {
2813         trace_ext4_alloc_da_blocks(inode);
2814
2815         if (!EXT4_I(inode)->i_reserved_data_blocks)
2816                 return 0;
2817
2818         /*
2819          * We do something simple for now.  The filemap_flush() will
2820          * also start triggering a write of the data blocks, which is
2821          * not strictly speaking necessary (and for users of
2822          * laptop_mode, not even desirable).  However, to do otherwise
2823          * would require replicating code paths in:
2824          *
2825          * ext4_writepages() ->
2826          *    write_cache_pages() ---> (via passed in callback function)
2827          *        __mpage_da_writepage() -->
2828          *           mpage_add_bh_to_extent()
2829          *           mpage_da_map_blocks()
2830          *
2831          * The problem is that write_cache_pages(), located in
2832          * mm/page-writeback.c, marks pages clean in preparation for
2833          * doing I/O, which is not desirable if we're not planning on
2834          * doing I/O at all.
2835          *
2836          * We could call write_cache_pages(), and then redirty all of
2837          * the pages by calling redirty_page_for_writepage() but that
2838          * would be ugly in the extreme.  So instead we would need to
2839          * replicate parts of the code in the above functions,
2840          * simplifying them because we wouldn't actually intend to
2841          * write out the pages, but rather only collect contiguous
2842          * logical block extents, call the multi-block allocator, and
2843          * then update the buffer heads with the block allocations.
2844          *
2845          * For now, though, we'll cheat by calling filemap_flush(),
2846          * which will map the blocks, and start the I/O, but not
2847          * actually wait for the I/O to complete.
2848          */
2849         return filemap_flush(inode->i_mapping);
2850 }
2851
2852 /*
2853  * bmap() is special.  It gets used by applications such as lilo and by
2854  * the swapper to find the on-disk block of a specific piece of data.
2855  *
2856  * Naturally, this is dangerous if the block concerned is still in the
2857  * journal.  If somebody makes a swapfile on an ext4 data-journaling
2858  * filesystem and enables swap, then they may get a nasty shock when the
2859  * data getting swapped to that swapfile suddenly gets overwritten by
2860  * the original zero's written out previously to the journal and
2861  * awaiting writeback in the kernel's buffer cache.
2862  *
2863  * So, if we see any bmap calls here on a modified, data-journaled file,
2864  * take extra steps to flush any blocks which might be in the cache.
2865  */
2866 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
2867 {
2868         struct inode *inode = mapping->host;
2869         journal_t *journal;
2870         int err;
2871
2872         /*
2873          * We can get here for an inline file via the FIBMAP ioctl
2874          */
2875         if (ext4_has_inline_data(inode))
2876                 return 0;
2877
2878         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
2879                         test_opt(inode->i_sb, DELALLOC)) {
2880                 /*
2881                  * With delalloc we want to sync the file
2882                  * so that we can make sure we allocate
2883                  * blocks for file
2884                  */
2885                 filemap_write_and_wait(mapping);
2886         }
2887
2888         if (EXT4_JOURNAL(inode) &&
2889             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
2890                 /*
2891                  * This is a REALLY heavyweight approach, but the use of
2892                  * bmap on dirty files is expected to be extremely rare:
2893                  * only if we run lilo or swapon on a freshly made file
2894                  * do we expect this to happen.
2895                  *
2896                  * (bmap requires CAP_SYS_RAWIO so this does not
2897                  * represent an unprivileged user DOS attack --- we'd be
2898                  * in trouble if mortal users could trigger this path at
2899                  * will.)
2900                  *
2901                  * NB. EXT4_STATE_JDATA is not set on files other than
2902                  * regular files.  If somebody wants to bmap a directory
2903                  * or symlink and gets confused because the buffer
2904                  * hasn't yet been flushed to disk, they deserve
2905                  * everything they get.
2906                  */
2907
2908                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
2909                 journal = EXT4_JOURNAL(inode);
2910                 jbd2_journal_lock_updates(journal);
2911                 err = jbd2_journal_flush(journal);
2912                 jbd2_journal_unlock_updates(journal);
2913
2914                 if (err)
2915                         return 0;
2916         }
2917
2918         return generic_block_bmap(mapping, block, ext4_get_block);
2919 }
2920
2921 static int ext4_readpage(struct file *file, struct page *page)
2922 {
2923         int ret = -EAGAIN;
2924         struct inode *inode = page->mapping->host;
2925
2926         trace_ext4_readpage(page);
2927
2928         if (ext4_has_inline_data(inode))
2929                 ret = ext4_readpage_inline(inode, page);
2930
2931         if (ret == -EAGAIN)
2932                 return ext4_mpage_readpages(page->mapping, NULL, page, 1);
2933
2934         return ret;
2935 }
2936
2937 static int
2938 ext4_readpages(struct file *file, struct address_space *mapping,
2939                 struct list_head *pages, unsigned nr_pages)
2940 {
2941         struct inode *inode = mapping->host;
2942
2943         /* If the file has inline data, no need to do readpages. */
2944         if (ext4_has_inline_data(inode))
2945                 return 0;
2946
2947         return ext4_mpage_readpages(mapping, pages, NULL, nr_pages);
2948 }
2949
2950 static void ext4_invalidatepage(struct page *page, unsigned int offset,
2951                                 unsigned int length)
2952 {
2953         trace_ext4_invalidatepage(page, offset, length);
2954
2955         /* No journalling happens on data buffers when this function is used */
2956         WARN_ON(page_has_buffers(page) && buffer_jbd(page_buffers(page)));
2957
2958         block_invalidatepage(page, offset, length);
2959 }
2960
2961 static int __ext4_journalled_invalidatepage(struct page *page,
2962                                             unsigned int offset,
2963                                             unsigned int length)
2964 {
2965         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2966
2967         trace_ext4_journalled_invalidatepage(page, offset, length);
2968
2969         /*
2970          * If it's a full truncate we just forget about the pending dirtying
2971          */
2972         if (offset == 0 && length == PAGE_CACHE_SIZE)
2973                 ClearPageChecked(page);
2974
2975         return jbd2_journal_invalidatepage(journal, page, offset, length);
2976 }
2977
2978 /* Wrapper for aops... */
2979 static void ext4_journalled_invalidatepage(struct page *page,
2980                                            unsigned int offset,
2981                                            unsigned int length)
2982 {
2983         WARN_ON(__ext4_journalled_invalidatepage(page, offset, length) < 0);
2984 }
2985
2986 static int ext4_releasepage(struct page *page, gfp_t wait)
2987 {
2988         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2989
2990         trace_ext4_releasepage(page);
2991
2992         /* Page has dirty journalled data -> cannot release */
2993         if (PageChecked(page))
2994                 return 0;
2995         if (journal)
2996                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
2997         else
2998                 return try_to_free_buffers(page);
2999 }
3000
3001 /*
3002  * ext4_get_block used when preparing for a DIO write or buffer write.
3003  * We allocate an uinitialized extent if blocks haven't been allocated.
3004  * The extent will be converted to initialized after the IO is complete.
3005  */
3006 int ext4_get_block_write(struct inode *inode, sector_t iblock,
3007                    struct buffer_head *bh_result, int create)
3008 {
3009         ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
3010                    inode->i_ino, create);
3011         return _ext4_get_block(inode, iblock, bh_result,
3012                                EXT4_GET_BLOCKS_IO_CREATE_EXT);
3013 }
3014
3015 static int ext4_get_block_write_nolock(struct inode *inode, sector_t iblock,
3016                    struct buffer_head *bh_result, int create)
3017 {
3018         ext4_debug("ext4_get_block_write_nolock: inode %lu, create flag %d\n",
3019                    inode->i_ino, create);
3020         return _ext4_get_block(inode, iblock, bh_result,
3021                                EXT4_GET_BLOCKS_NO_LOCK);
3022 }
3023
3024 int ext4_get_block_dax(struct inode *inode, sector_t iblock,
3025                    struct buffer_head *bh_result, int create)
3026 {
3027         int flags = EXT4_GET_BLOCKS_PRE_IO | EXT4_GET_BLOCKS_UNWRIT_EXT;
3028         if (create)
3029                 flags |= EXT4_GET_BLOCKS_CREATE;
3030         ext4_debug("ext4_get_block_dax: inode %lu, create flag %d\n",
3031                    inode->i_ino, create);
3032         return _ext4_get_block(inode, iblock, bh_result, flags);
3033 }
3034
3035 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3036                             ssize_t size, void *private)
3037 {
3038         ext4_io_end_t *io_end = iocb->private;
3039
3040         /* if not async direct IO just return */
3041         if (!io_end)
3042                 return;
3043
3044         ext_debug("ext4_end_io_dio(): io_end 0x%p "
3045                   "for inode %lu, iocb 0x%p, offset %llu, size %zd\n",
3046                   iocb->private, io_end->inode->i_ino, iocb, offset,
3047                   size);
3048
3049         iocb->private = NULL;
3050         io_end->offset = offset;
3051         io_end->size = size;
3052         ext4_put_io_end(io_end);
3053 }
3054
3055 /*
3056  * For ext4 extent files, ext4 will do direct-io write to holes,
3057  * preallocated extents, and those write extend the file, no need to
3058  * fall back to buffered IO.
3059  *
3060  * For holes, we fallocate those blocks, mark them as unwritten
3061  * If those blocks were preallocated, we mark sure they are split, but
3062  * still keep the range to write as unwritten.
3063  *
3064  * The unwritten extents will be converted to written when DIO is completed.
3065  * For async direct IO, since the IO may still pending when return, we
3066  * set up an end_io call back function, which will do the conversion
3067  * when async direct IO completed.
3068  *
3069  * If the O_DIRECT write will extend the file then add this inode to the
3070  * orphan list.  So recovery will truncate it back to the original size
3071  * if the machine crashes during the write.
3072  *
3073  */
3074 static ssize_t ext4_ext_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
3075                                   loff_t offset)
3076 {
3077         struct file *file = iocb->ki_filp;
3078         struct inode *inode = file->f_mapping->host;
3079         ssize_t ret;
3080         size_t count = iov_iter_count(iter);
3081         int overwrite = 0;
3082         get_block_t *get_block_func = NULL;
3083         int dio_flags = 0;
3084         loff_t final_size = offset + count;
3085         ext4_io_end_t *io_end = NULL;
3086
3087         /* Use the old path for reads and writes beyond i_size. */
3088         if (iov_iter_rw(iter) != WRITE || final_size > inode->i_size)
3089                 return ext4_ind_direct_IO(iocb, iter, offset);
3090
3091         BUG_ON(iocb->private == NULL);
3092
3093         /*
3094          * Make all waiters for direct IO properly wait also for extent
3095          * conversion. This also disallows race between truncate() and
3096          * overwrite DIO as i_dio_count needs to be incremented under i_mutex.
3097          */
3098         if (iov_iter_rw(iter) == WRITE)
3099                 inode_dio_begin(inode);
3100
3101         /* If we do a overwrite dio, i_mutex locking can be released */
3102         overwrite = *((int *)iocb->private);
3103
3104         if (overwrite) {
3105                 down_read(&EXT4_I(inode)->i_data_sem);
3106                 mutex_unlock(&inode->i_mutex);
3107         }
3108
3109         /*
3110          * We could direct write to holes and fallocate.
3111          *
3112          * Allocated blocks to fill the hole are marked as
3113          * unwritten to prevent parallel buffered read to expose
3114          * the stale data before DIO complete the data IO.
3115          *
3116          * As to previously fallocated extents, ext4 get_block will
3117          * just simply mark the buffer mapped but still keep the
3118          * extents unwritten.
3119          *
3120          * For non AIO case, we will convert those unwritten extents
3121          * to written after return back from blockdev_direct_IO.
3122          *
3123          * For async DIO, the conversion needs to be deferred when the
3124          * IO is completed. The ext4 end_io callback function will be
3125          * called to take care of the conversion work.  Here for async
3126          * case, we allocate an io_end structure to hook to the iocb.
3127          */
3128         iocb->private = NULL;
3129         ext4_inode_aio_set(inode, NULL);
3130         if (!is_sync_kiocb(iocb)) {
3131                 io_end = ext4_init_io_end(inode, GFP_NOFS);
3132                 if (!io_end) {
3133                         ret = -ENOMEM;
3134                         goto retake_lock;
3135                 }
3136                 /*
3137                  * Grab reference for DIO. Will be dropped in ext4_end_io_dio()
3138                  */
3139                 iocb->private = ext4_get_io_end(io_end);
3140                 /*
3141                  * we save the io structure for current async direct
3142                  * IO, so that later ext4_map_blocks() could flag the
3143                  * io structure whether there is a unwritten extents
3144                  * needs to be converted when IO is completed.
3145                  */
3146                 ext4_inode_aio_set(inode, io_end);
3147         }
3148
3149         if (overwrite) {
3150                 get_block_func = ext4_get_block_write_nolock;
3151         } else {
3152                 get_block_func = ext4_get_block_write;
3153                 dio_flags = DIO_LOCKING;
3154         }
3155 #ifdef CONFIG_EXT4_FS_ENCRYPTION
3156         BUG_ON(ext4_encrypted_inode(inode) && S_ISREG(inode->i_mode));
3157 #endif
3158         if (IS_DAX(inode))
3159                 ret = dax_do_io(iocb, inode, iter, offset, get_block_func,
3160                                 ext4_end_io_dio, dio_flags);
3161         else
3162                 ret = __blockdev_direct_IO(iocb, inode,
3163                                            inode->i_sb->s_bdev, iter, offset,
3164                                            get_block_func,
3165                                            ext4_end_io_dio, NULL, dio_flags);
3166
3167         /*
3168          * Put our reference to io_end. This can free the io_end structure e.g.
3169          * in sync IO case or in case of error. It can even perform extent
3170          * conversion if all bios we submitted finished before we got here.
3171          * Note that in that case iocb->private can be already set to NULL
3172          * here.
3173          */
3174         if (io_end) {
3175                 ext4_inode_aio_set(inode, NULL);
3176                 ext4_put_io_end(io_end);
3177                 /*
3178                  * When no IO was submitted ext4_end_io_dio() was not
3179                  * called so we have to put iocb's reference.
3180                  */
3181                 if (ret <= 0 && ret != -EIOCBQUEUED && iocb->private) {
3182                         WARN_ON(iocb->private != io_end);
3183                         WARN_ON(io_end->flag & EXT4_IO_END_UNWRITTEN);
3184                         ext4_put_io_end(io_end);
3185                         iocb->private = NULL;
3186                 }
3187         }
3188         if (ret > 0 && !overwrite && ext4_test_inode_state(inode,
3189                                                 EXT4_STATE_DIO_UNWRITTEN)) {
3190                 int err;
3191                 /*
3192                  * for non AIO case, since the IO is already
3193                  * completed, we could do the conversion right here
3194                  */
3195                 err = ext4_convert_unwritten_extents(NULL, inode,
3196                                                      offset, ret);
3197                 if (err < 0)
3198                         ret = err;
3199                 ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
3200         }
3201
3202 retake_lock:
3203         if (iov_iter_rw(iter) == WRITE)
3204                 inode_dio_end(inode);
3205         /* take i_mutex locking again if we do a ovewrite dio */
3206         if (overwrite) {
3207                 up_read(&EXT4_I(inode)->i_data_sem);
3208                 mutex_lock(&inode->i_mutex);
3209         }
3210
3211         return ret;
3212 }
3213
3214 static ssize_t ext4_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
3215                               loff_t offset)
3216 {
3217         struct file *file = iocb->ki_filp;
3218         struct inode *inode = file->f_mapping->host;
3219         size_t count = iov_iter_count(iter);
3220         ssize_t ret;
3221
3222 #ifdef CONFIG_EXT4_FS_ENCRYPTION
3223         if (ext4_encrypted_inode(inode) && S_ISREG(inode->i_mode))
3224                 return 0;
3225 #endif
3226
3227         /*
3228          * If we are doing data journalling we don't support O_DIRECT
3229          */
3230         if (ext4_should_journal_data(inode))
3231                 return 0;
3232
3233         /* Let buffer I/O handle the inline data case. */
3234         if (ext4_has_inline_data(inode))
3235                 return 0;
3236
3237         trace_ext4_direct_IO_enter(inode, offset, count, iov_iter_rw(iter));
3238         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3239                 ret = ext4_ext_direct_IO(iocb, iter, offset);
3240         else
3241                 ret = ext4_ind_direct_IO(iocb, iter, offset);
3242         trace_ext4_direct_IO_exit(inode, offset, count, iov_iter_rw(iter), ret);
3243         return ret;
3244 }
3245
3246 /*
3247  * Pages can be marked dirty completely asynchronously from ext4's journalling
3248  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3249  * much here because ->set_page_dirty is called under VFS locks.  The page is
3250  * not necessarily locked.
3251  *
3252  * We cannot just dirty the page and leave attached buffers clean, because the
3253  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3254  * or jbddirty because all the journalling code will explode.
3255  *
3256  * So what we do is to mark the page "pending dirty" and next time writepage
3257  * is called, propagate that into the buffers appropriately.
3258  */
3259 static int ext4_journalled_set_page_dirty(struct page *page)
3260 {
3261         SetPageChecked(page);
3262         return __set_page_dirty_nobuffers(page);
3263 }
3264
3265 static const struct address_space_operations ext4_aops = {
3266         .readpage               = ext4_readpage,
3267         .readpages              = ext4_readpages,
3268         .writepage              = ext4_writepage,
3269         .writepages             = ext4_writepages,
3270         .write_begin            = ext4_write_begin,
3271         .write_end              = ext4_write_end,
3272         .bmap                   = ext4_bmap,
3273         .invalidatepage         = ext4_invalidatepage,
3274         .releasepage            = ext4_releasepage,
3275         .direct_IO              = ext4_direct_IO,
3276         .migratepage            = buffer_migrate_page,
3277         .is_partially_uptodate  = block_is_partially_uptodate,
3278         .error_remove_page      = generic_error_remove_page,
3279 };
3280
3281 static const struct address_space_operations ext4_journalled_aops = {
3282         .readpage               = ext4_readpage,
3283         .readpages              = ext4_readpages,
3284         .writepage              = ext4_writepage,
3285         .writepages             = ext4_writepages,
3286         .write_begin            = ext4_write_begin,
3287         .write_end              = ext4_journalled_write_end,
3288         .set_page_dirty         = ext4_journalled_set_page_dirty,
3289         .bmap                   = ext4_bmap,
3290         .invalidatepage         = ext4_journalled_invalidatepage,
3291         .releasepage            = ext4_releasepage,
3292         .direct_IO              = ext4_direct_IO,
3293         .is_partially_uptodate  = block_is_partially_uptodate,
3294         .error_remove_page      = generic_error_remove_page,
3295 };
3296
3297 static const struct address_space_operations ext4_da_aops = {
3298         .readpage               = ext4_readpage,
3299         .readpages              = ext4_readpages,
3300         .writepage              = ext4_writepage,
3301         .writepages             = ext4_writepages,
3302         .write_begin            = ext4_da_write_begin,
3303         .write_end              = ext4_da_write_end,
3304         .bmap                   = ext4_bmap,
3305         .invalidatepage         = ext4_da_invalidatepage,
3306         .releasepage            = ext4_releasepage,
3307         .direct_IO              = ext4_direct_IO,
3308         .migratepage            = buffer_migrate_page,
3309         .is_partially_uptodate  = block_is_partially_uptodate,
3310         .error_remove_page      = generic_error_remove_page,
3311 };
3312
3313 void ext4_set_aops(struct inode *inode)
3314 {
3315         switch (ext4_inode_journal_mode(inode)) {
3316         case EXT4_INODE_ORDERED_DATA_MODE:
3317                 ext4_set_inode_state(inode, EXT4_STATE_ORDERED_MODE);
3318                 break;
3319         case EXT4_INODE_WRITEBACK_DATA_MODE:
3320                 ext4_clear_inode_state(inode, EXT4_STATE_ORDERED_MODE);
3321                 break;
3322         case EXT4_INODE_JOURNAL_DATA_MODE:
3323                 inode->i_mapping->a_ops = &ext4_journalled_aops;
3324                 return;
3325         default:
3326                 BUG();
3327         }
3328         if (test_opt(inode->i_sb, DELALLOC))
3329                 inode->i_mapping->a_ops = &ext4_da_aops;
3330         else
3331                 inode->i_mapping->a_ops = &ext4_aops;
3332 }
3333
3334 static int __ext4_block_zero_page_range(handle_t *handle,
3335                 struct address_space *mapping, loff_t from, loff_t length)
3336 {
3337         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3338         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3339         unsigned blocksize, pos;
3340         ext4_lblk_t iblock;
3341         struct inode *inode = mapping->host;
3342         struct buffer_head *bh;
3343         struct page *page;
3344         int err = 0;
3345
3346         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
3347                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
3348         if (!page)
3349                 return -ENOMEM;
3350
3351         blocksize = inode->i_sb->s_blocksize;
3352
3353         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3354
3355         if (!page_has_buffers(page))
3356                 create_empty_buffers(page, blocksize, 0);
3357
3358         /* Find the buffer that contains "offset" */
3359         bh = page_buffers(page);
3360         pos = blocksize;
3361         while (offset >= pos) {
3362                 bh = bh->b_this_page;
3363                 iblock++;
3364                 pos += blocksize;
3365         }
3366         if (buffer_freed(bh)) {
3367                 BUFFER_TRACE(bh, "freed: skip");
3368                 goto unlock;
3369         }
3370         if (!buffer_mapped(bh)) {
3371                 BUFFER_TRACE(bh, "unmapped");
3372                 ext4_get_block(inode, iblock, bh, 0);
3373                 /* unmapped? It's a hole - nothing to do */
3374                 if (!buffer_mapped(bh)) {
3375                         BUFFER_TRACE(bh, "still unmapped");
3376                         goto unlock;
3377                 }
3378         }
3379
3380         /* Ok, it's mapped. Make sure it's up-to-date */
3381         if (PageUptodate(page))
3382                 set_buffer_uptodate(bh);
3383
3384         if (!buffer_uptodate(bh)) {
3385                 err = -EIO;
3386                 ll_rw_block(READ, 1, &bh);
3387                 wait_on_buffer(bh);
3388                 /* Uhhuh. Read error. Complain and punt. */
3389                 if (!buffer_uptodate(bh))
3390                         goto unlock;
3391                 if (S_ISREG(inode->i_mode) &&
3392                     ext4_encrypted_inode(inode)) {
3393                         /* We expect the key to be set. */
3394                         BUG_ON(!ext4_has_encryption_key(inode));
3395                         BUG_ON(blocksize != PAGE_CACHE_SIZE);
3396                         WARN_ON_ONCE(ext4_decrypt_one(inode, page));
3397                 }
3398         }
3399         if (ext4_should_journal_data(inode)) {
3400                 BUFFER_TRACE(bh, "get write access");
3401                 err = ext4_journal_get_write_access(handle, bh);
3402                 if (err)
3403                         goto unlock;
3404         }
3405         zero_user(page, offset, length);
3406         BUFFER_TRACE(bh, "zeroed end of block");
3407
3408         if (ext4_should_journal_data(inode)) {
3409                 err = ext4_handle_dirty_metadata(handle, inode, bh);
3410         } else {
3411                 err = 0;
3412                 mark_buffer_dirty(bh);
3413                 if (ext4_test_inode_state(inode, EXT4_STATE_ORDERED_MODE))
3414                         err = ext4_jbd2_file_inode(handle, inode);
3415         }
3416
3417 unlock:
3418         unlock_page(page);
3419         page_cache_release(page);
3420         return err;
3421 }
3422
3423 /*
3424  * ext4_block_zero_page_range() zeros out a mapping of length 'length'
3425  * starting from file offset 'from'.  The range to be zero'd must
3426  * be contained with in one block.  If the specified range exceeds
3427  * the end of the block it will be shortened to end of the block
3428  * that cooresponds to 'from'
3429  */
3430 static int ext4_block_zero_page_range(handle_t *handle,
3431                 struct address_space *mapping, loff_t from, loff_t length)
3432 {
3433         struct inode *inode = mapping->host;
3434         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3435         unsigned blocksize = inode->i_sb->s_blocksize;
3436         unsigned max = blocksize - (offset & (blocksize - 1));
3437
3438         /*
3439          * correct length if it does not fall between
3440          * 'from' and the end of the block
3441          */
3442         if (length > max || length < 0)
3443                 length = max;
3444
3445         if (IS_DAX(inode))
3446                 return dax_zero_page_range(inode, from, length, ext4_get_block);
3447         return __ext4_block_zero_page_range(handle, mapping, from, length);
3448 }
3449
3450 /*
3451  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3452  * up to the end of the block which corresponds to `from'.
3453  * This required during truncate. We need to physically zero the tail end
3454  * of that block so it doesn't yield old data if the file is later grown.
3455  */
3456 static int ext4_block_truncate_page(handle_t *handle,
3457                 struct address_space *mapping, loff_t from)
3458 {
3459         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3460         unsigned length;
3461         unsigned blocksize;
3462         struct inode *inode = mapping->host;
3463
3464         blocksize = inode->i_sb->s_blocksize;
3465         length = blocksize - (offset & (blocksize - 1));
3466
3467         return ext4_block_zero_page_range(handle, mapping, from, length);
3468 }
3469
3470 int ext4_zero_partial_blocks(handle_t *handle, struct inode *inode,
3471                              loff_t lstart, loff_t length)
3472 {
3473         struct super_block *sb = inode->i_sb;
3474         struct address_space *mapping = inode->i_mapping;
3475         unsigned partial_start, partial_end;
3476         ext4_fsblk_t start, end;
3477         loff_t byte_end = (lstart + length - 1);
3478         int err = 0;
3479
3480         partial_start = lstart & (sb->s_blocksize - 1);
3481         partial_end = byte_end & (sb->s_blocksize - 1);
3482
3483         start = lstart >> sb->s_blocksize_bits;
3484         end = byte_end >> sb->s_blocksize_bits;
3485
3486         /* Handle partial zero within the single block */
3487         if (start == end &&
3488             (partial_start || (partial_end != sb->s_blocksize - 1))) {
3489                 err = ext4_block_zero_page_range(handle, mapping,
3490                                                  lstart, length);
3491                 return err;
3492         }
3493         /* Handle partial zero out on the start of the range */
3494         if (partial_start) {
3495                 err = ext4_block_zero_page_range(handle, mapping,
3496                                                  lstart, sb->s_blocksize);
3497                 if (err)
3498                         return err;
3499         }
3500         /* Handle partial zero out on the end of the range */
3501         if (partial_end != sb->s_blocksize - 1)
3502                 err = ext4_block_zero_page_range(handle, mapping,
3503                                                  byte_end - partial_end,
3504                                                  partial_end + 1);
3505         return err;
3506 }
3507
3508 int ext4_can_truncate(struct inode *inode)
3509 {
3510         if (S_ISREG(inode->i_mode))
3511                 return 1;
3512         if (S_ISDIR(inode->i_mode))
3513                 return 1;
3514         if (S_ISLNK(inode->i_mode))
3515                 return !ext4_inode_is_fast_symlink(inode);
3516         return 0;
3517 }
3518
3519 /*
3520  * ext4_punch_hole: punches a hole in a file by releaseing the blocks
3521  * associated with the given offset and length
3522  *
3523  * @inode:  File inode
3524  * @offset: The offset where the hole will begin
3525  * @len:    The length of the hole
3526  *
3527  * Returns: 0 on success or negative on failure
3528  */
3529
3530 int ext4_punch_hole(struct inode *inode, loff_t offset, loff_t length)
3531 {
3532         struct super_block *sb = inode->i_sb;
3533         ext4_lblk_t first_block, stop_block;
3534         struct address_space *mapping = inode->i_mapping;
3535         loff_t first_block_offset, last_block_offset;
3536         handle_t *handle;
3537         unsigned int credits;
3538         int ret = 0;
3539
3540         if (!S_ISREG(inode->i_mode))
3541                 return -EOPNOTSUPP;
3542
3543         trace_ext4_punch_hole(inode, offset, length, 0);
3544
3545         /*
3546          * Write out all dirty pages to avoid race conditions
3547          * Then release them.
3548          */
3549         if (mapping->nrpages && mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
3550                 ret = filemap_write_and_wait_range(mapping, offset,
3551                                                    offset + length - 1);
3552                 if (ret)
3553                         return ret;
3554         }
3555
3556         mutex_lock(&inode->i_mutex);
3557
3558         /* No need to punch hole beyond i_size */
3559         if (offset >= inode->i_size)
3560                 goto out_mutex;
3561
3562         /*
3563          * If the hole extends beyond i_size, set the hole
3564          * to end after the page that contains i_size
3565          */
3566         if (offset + length > inode->i_size) {
3567                 length = inode->i_size +
3568                    PAGE_CACHE_SIZE - (inode->i_size & (PAGE_CACHE_SIZE - 1)) -
3569                    offset;
3570         }
3571
3572         if (offset & (sb->s_blocksize - 1) ||
3573             (offset + length) & (sb->s_blocksize - 1)) {
3574                 /*
3575                  * Attach jinode to inode for jbd2 if we do any zeroing of
3576                  * partial block
3577                  */
3578                 ret = ext4_inode_attach_jinode(inode);
3579                 if (ret < 0)
3580                         goto out_mutex;
3581
3582         }
3583
3584         first_block_offset = round_up(offset, sb->s_blocksize);
3585         last_block_offset = round_down((offset + length), sb->s_blocksize) - 1;
3586
3587         /* Now release the pages and zero block aligned part of pages*/
3588         if (last_block_offset > first_block_offset)
3589                 truncate_pagecache_range(inode, first_block_offset,
3590                                          last_block_offset);
3591
3592         /* Wait all existing dio workers, newcomers will block on i_mutex */
3593         ext4_inode_block_unlocked_dio(inode);
3594         inode_dio_wait(inode);
3595
3596         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3597                 credits = ext4_writepage_trans_blocks(inode);
3598         else
3599                 credits = ext4_blocks_for_truncate(inode);
3600         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
3601         if (IS_ERR(handle)) {
3602                 ret = PTR_ERR(handle);
3603                 ext4_std_error(sb, ret);
3604                 goto out_dio;
3605         }
3606
3607         ret = ext4_zero_partial_blocks(handle, inode, offset,
3608                                        length);
3609         if (ret)
3610                 goto out_stop;
3611
3612         first_block = (offset + sb->s_blocksize - 1) >>
3613                 EXT4_BLOCK_SIZE_BITS(sb);
3614         stop_block = (offset + length) >> EXT4_BLOCK_SIZE_BITS(sb);
3615
3616         /* If there are no blocks to remove, return now */
3617         if (first_block >= stop_block)
3618                 goto out_stop;
3619
3620         down_write(&EXT4_I(inode)->i_data_sem);
3621         ext4_discard_preallocations(inode);
3622
3623         ret = ext4_es_remove_extent(inode, first_block,
3624                                     stop_block - first_block);
3625         if (ret) {
3626                 up_write(&EXT4_I(inode)->i_data_sem);
3627                 goto out_stop;
3628         }
3629
3630         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3631                 ret = ext4_ext_remove_space(inode, first_block,
3632                                             stop_block - 1);
3633         else
3634                 ret = ext4_ind_remove_space(handle, inode, first_block,
3635                                             stop_block);
3636
3637         up_write(&EXT4_I(inode)->i_data_sem);
3638         if (IS_SYNC(inode))
3639                 ext4_handle_sync(handle);
3640
3641         /* Now release the pages again to reduce race window */
3642         if (last_block_offset > first_block_offset)
3643                 truncate_pagecache_range(inode, first_block_offset,
3644                                          last_block_offset);
3645
3646         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
3647         ext4_mark_inode_dirty(handle, inode);
3648 out_stop:
3649         ext4_journal_stop(handle);
3650 out_dio:
3651         ext4_inode_resume_unlocked_dio(inode);
3652 out_mutex:
3653         mutex_unlock(&inode->i_mutex);
3654         return ret;
3655 }
3656
3657 int ext4_inode_attach_jinode(struct inode *inode)
3658 {
3659         struct ext4_inode_info *ei = EXT4_I(inode);
3660         struct jbd2_inode *jinode;
3661
3662         if (ei->jinode || !EXT4_SB(inode->i_sb)->s_journal)
3663                 return 0;
3664
3665         jinode = jbd2_alloc_inode(GFP_KERNEL);
3666         spin_lock(&inode->i_lock);
3667         if (!ei->jinode) {
3668                 if (!jinode) {
3669                         spin_unlock(&inode->i_lock);
3670                         return -ENOMEM;
3671                 }
3672                 ei->jinode = jinode;
3673                 jbd2_journal_init_jbd_inode(ei->jinode, inode);
3674                 jinode = NULL;
3675         }
3676         spin_unlock(&inode->i_lock);
3677         if (unlikely(jinode != NULL))
3678                 jbd2_free_inode(jinode);
3679         return 0;
3680 }
3681
3682 /*
3683  * ext4_truncate()
3684  *
3685  * We block out ext4_get_block() block instantiations across the entire
3686  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
3687  * simultaneously on behalf of the same inode.
3688  *
3689  * As we work through the truncate and commit bits of it to the journal there
3690  * is one core, guiding principle: the file's tree must always be consistent on
3691  * disk.  We must be able to restart the truncate after a crash.
3692  *
3693  * The file's tree may be transiently inconsistent in memory (although it
3694  * probably isn't), but whenever we close off and commit a journal transaction,
3695  * the contents of (the filesystem + the journal) must be consistent and
3696  * restartable.  It's pretty simple, really: bottom up, right to left (although
3697  * left-to-right works OK too).
3698  *
3699  * Note that at recovery time, journal replay occurs *before* the restart of
3700  * truncate against the orphan inode list.
3701  *
3702  * The committed inode has the new, desired i_size (which is the same as
3703  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
3704  * that this inode's truncate did not complete and it will again call
3705  * ext4_truncate() to have another go.  So there will be instantiated blocks
3706  * to the right of the truncation point in a crashed ext4 filesystem.  But
3707  * that's fine - as long as they are linked from the inode, the post-crash
3708  * ext4_truncate() run will find them and release them.
3709  */
3710 void ext4_truncate(struct inode *inode)
3711 {
3712         struct ext4_inode_info *ei = EXT4_I(inode);
3713         unsigned int credits;
3714         handle_t *handle;
3715         struct address_space *mapping = inode->i_mapping;
3716
3717         /*
3718          * There is a possibility that we're either freeing the inode
3719          * or it's a completely new inode. In those cases we might not
3720          * have i_mutex locked because it's not necessary.
3721          */
3722         if (!(inode->i_state & (I_NEW|I_FREEING)))
3723                 WARN_ON(!mutex_is_locked(&inode->i_mutex));
3724         trace_ext4_truncate_enter(inode);
3725
3726         if (!ext4_can_truncate(inode))
3727                 return;
3728
3729         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3730
3731         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
3732                 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
3733
3734         if (ext4_has_inline_data(inode)) {
3735                 int has_inline = 1;
3736
3737                 ext4_inline_data_truncate(inode, &has_inline);
3738                 if (has_inline)
3739                         return;
3740         }
3741
3742         /* If we zero-out tail of the page, we have to create jinode for jbd2 */
3743         if (inode->i_size & (inode->i_sb->s_blocksize - 1)) {
3744                 if (ext4_inode_attach_jinode(inode) < 0)
3745                         return;
3746         }
3747
3748         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3749                 credits = ext4_writepage_trans_blocks(inode);
3750         else
3751                 credits = ext4_blocks_for_truncate(inode);
3752
3753         handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, credits);
3754         if (IS_ERR(handle)) {
3755                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
3756                 return;
3757         }
3758
3759         if (inode->i_size & (inode->i_sb->s_blocksize - 1))
3760                 ext4_block_truncate_page(handle, mapping, inode->i_size);
3761
3762         /*
3763          * We add the inode to the orphan list, so that if this
3764          * truncate spans multiple transactions, and we crash, we will
3765          * resume the truncate when the filesystem recovers.  It also
3766          * marks the inode dirty, to catch the new size.
3767          *
3768          * Implication: the file must always be in a sane, consistent
3769          * truncatable state while each transaction commits.
3770          */
3771         if (ext4_orphan_add(handle, inode))
3772                 goto out_stop;
3773
3774         down_write(&EXT4_I(inode)->i_data_sem);
3775
3776         ext4_discard_preallocations(inode);
3777
3778         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3779                 ext4_ext_truncate(handle, inode);
3780         else
3781                 ext4_ind_truncate(handle, inode);
3782
3783         up_write(&ei->i_data_sem);
3784
3785         if (IS_SYNC(inode))
3786                 ext4_handle_sync(handle);
3787
3788 out_stop:
3789         /*
3790          * If this was a simple ftruncate() and the file will remain alive,
3791          * then we need to clear up the orphan record which we created above.
3792          * However, if this was a real unlink then we were called by
3793          * ext4_evict_inode(), and we allow that function to clean up the
3794          * orphan info for us.
3795          */
3796         if (inode->i_nlink)
3797                 ext4_orphan_del(handle, inode);
3798
3799         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
3800         ext4_mark_inode_dirty(handle, inode);
3801         ext4_journal_stop(handle);
3802
3803         trace_ext4_truncate_exit(inode);
3804 }
3805
3806 /*
3807  * ext4_get_inode_loc returns with an extra refcount against the inode's
3808  * underlying buffer_head on success. If 'in_mem' is true, we have all
3809  * data in memory that is needed to recreate the on-disk version of this
3810  * inode.
3811  */
3812 static int __ext4_get_inode_loc(struct inode *inode,
3813                                 struct ext4_iloc *iloc, int in_mem)
3814 {
3815         struct ext4_group_desc  *gdp;
3816         struct buffer_head      *bh;
3817         struct super_block      *sb = inode->i_sb;
3818         ext4_fsblk_t            block;
3819         int                     inodes_per_block, inode_offset;
3820
3821         iloc->bh = NULL;
3822         if (!ext4_valid_inum(sb, inode->i_ino))
3823                 return -EIO;
3824
3825         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
3826         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
3827         if (!gdp)
3828                 return -EIO;
3829
3830         /*
3831          * Figure out the offset within the block group inode table
3832          */
3833         inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
3834         inode_offset = ((inode->i_ino - 1) %
3835                         EXT4_INODES_PER_GROUP(sb));
3836         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
3837         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
3838
3839         bh = sb_getblk(sb, block);
3840         if (unlikely(!bh))
3841                 return -ENOMEM;
3842         if (!buffer_uptodate(bh)) {
3843                 lock_buffer(bh);
3844
3845                 /*
3846                  * If the buffer has the write error flag, we have failed
3847                  * to write out another inode in the same block.  In this
3848                  * case, we don't have to read the block because we may
3849                  * read the old inode data successfully.
3850                  */
3851                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
3852                         set_buffer_uptodate(bh);
3853
3854                 if (buffer_uptodate(bh)) {
3855                         /* someone brought it uptodate while we waited */
3856                         unlock_buffer(bh);
3857                         goto has_buffer;
3858                 }
3859
3860                 /*
3861                  * If we have all information of the inode in memory and this
3862                  * is the only valid inode in the block, we need not read the
3863                  * block.
3864                  */
3865                 if (in_mem) {
3866                         struct buffer_head *bitmap_bh;
3867                         int i, start;
3868
3869                         start = inode_offset & ~(inodes_per_block - 1);
3870
3871                         /* Is the inode bitmap in cache? */
3872                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
3873                         if (unlikely(!bitmap_bh))
3874                                 goto make_io;
3875
3876                         /*
3877                          * If the inode bitmap isn't in cache then the
3878                          * optimisation may end up performing two reads instead
3879                          * of one, so skip it.
3880                          */
3881                         if (!buffer_uptodate(bitmap_bh)) {
3882                                 brelse(bitmap_bh);
3883                                 goto make_io;
3884                         }
3885                         for (i = start; i < start + inodes_per_block; i++) {
3886                                 if (i == inode_offset)
3887                                         continue;
3888                                 if (ext4_test_bit(i, bitmap_bh->b_data))
3889                                         break;
3890                         }
3891                         brelse(bitmap_bh);
3892                         if (i == start + inodes_per_block) {
3893                                 /* all other inodes are free, so skip I/O */
3894                                 memset(bh->b_data, 0, bh->b_size);
3895                                 set_buffer_uptodate(bh);
3896                                 unlock_buffer(bh);
3897                                 goto has_buffer;
3898                         }
3899                 }
3900
3901 make_io:
3902                 /*
3903                  * If we need to do any I/O, try to pre-readahead extra
3904                  * blocks from the inode table.
3905                  */
3906                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
3907                         ext4_fsblk_t b, end, table;
3908                         unsigned num;
3909                         __u32 ra_blks = EXT4_SB(sb)->s_inode_readahead_blks;
3910
3911                         table = ext4_inode_table(sb, gdp);
3912                         /* s_inode_readahead_blks is always a power of 2 */
3913                         b = block & ~((ext4_fsblk_t) ra_blks - 1);
3914                         if (table > b)
3915                                 b = table;
3916                         end = b + ra_blks;
3917                         num = EXT4_INODES_PER_GROUP(sb);
3918                         if (ext4_has_group_desc_csum(sb))
3919                                 num -= ext4_itable_unused_count(sb, gdp);
3920                         table += num / inodes_per_block;
3921                         if (end > table)
3922                                 end = table;
3923                         while (b <= end)
3924                                 sb_breadahead(sb, b++);
3925                 }
3926
3927                 /*
3928                  * There are other valid inodes in the buffer, this inode
3929                  * has in-inode xattrs, or we don't have this inode in memory.
3930                  * Read the block from disk.
3931                  */
3932                 trace_ext4_load_inode(inode);
3933                 get_bh(bh);
3934                 bh->b_end_io = end_buffer_read_sync;
3935                 submit_bh(READ | REQ_META | REQ_PRIO, bh);
3936                 wait_on_buffer(bh);
3937                 if (!buffer_uptodate(bh)) {
3938                         EXT4_ERROR_INODE_BLOCK(inode, block,
3939                                                "unable to read itable block");
3940                         brelse(bh);
3941                         return -EIO;
3942                 }
3943         }
3944 has_buffer:
3945         iloc->bh = bh;
3946         return 0;
3947 }
3948
3949 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
3950 {
3951         /* We have all inode data except xattrs in memory here. */
3952         return __ext4_get_inode_loc(inode, iloc,
3953                 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
3954 }
3955
3956 void ext4_set_inode_flags(struct inode *inode)
3957 {
3958         unsigned int flags = EXT4_I(inode)->i_flags;
3959         unsigned int new_fl = 0;
3960
3961         if (flags & EXT4_SYNC_FL)
3962                 new_fl |= S_SYNC;
3963         if (flags & EXT4_APPEND_FL)
3964                 new_fl |= S_APPEND;
3965         if (flags & EXT4_IMMUTABLE_FL)
3966                 new_fl |= S_IMMUTABLE;
3967         if (flags & EXT4_NOATIME_FL)
3968                 new_fl |= S_NOATIME;
3969         if (flags & EXT4_DIRSYNC_FL)
3970                 new_fl |= S_DIRSYNC;
3971         if (test_opt(inode->i_sb, DAX))
3972                 new_fl |= S_DAX;
3973         inode_set_flags(inode, new_fl,
3974                         S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC|S_DAX);
3975 }
3976
3977 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
3978 void ext4_get_inode_flags(struct ext4_inode_info *ei)
3979 {
3980         unsigned int vfs_fl;
3981         unsigned long old_fl, new_fl;
3982
3983         do {
3984                 vfs_fl = ei->vfs_inode.i_flags;
3985                 old_fl = ei->i_flags;
3986                 new_fl = old_fl & ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
3987                                 EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|
3988                                 EXT4_DIRSYNC_FL);
3989                 if (vfs_fl & S_SYNC)
3990                         new_fl |= EXT4_SYNC_FL;
3991                 if (vfs_fl & S_APPEND)
3992                         new_fl |= EXT4_APPEND_FL;
3993                 if (vfs_fl & S_IMMUTABLE)
3994                         new_fl |= EXT4_IMMUTABLE_FL;
3995                 if (vfs_fl & S_NOATIME)
3996                         new_fl |= EXT4_NOATIME_FL;
3997                 if (vfs_fl & S_DIRSYNC)
3998                         new_fl |= EXT4_DIRSYNC_FL;
3999         } while (cmpxchg(&ei->i_flags, old_fl, new_fl) != old_fl);
4000 }
4001
4002 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4003                                   struct ext4_inode_info *ei)
4004 {
4005         blkcnt_t i_blocks ;
4006         struct inode *inode = &(ei->vfs_inode);
4007         struct super_block *sb = inode->i_sb;
4008
4009         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4010                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4011                 /* we are using combined 48 bit field */
4012                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4013                                         le32_to_cpu(raw_inode->i_blocks_lo);
4014                 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
4015                         /* i_blocks represent file system block size */
4016                         return i_blocks  << (inode->i_blkbits - 9);
4017                 } else {
4018                         return i_blocks;
4019                 }
4020         } else {
4021                 return le32_to_cpu(raw_inode->i_blocks_lo);
4022         }
4023 }
4024
4025 static inline void ext4_iget_extra_inode(struct inode *inode,
4026                                          struct ext4_inode *raw_inode,
4027                                          struct ext4_inode_info *ei)
4028 {
4029         __le32 *magic = (void *)raw_inode +
4030                         EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize;
4031         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC)) {
4032                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
4033                 ext4_find_inline_data_nolock(inode);
4034         } else
4035                 EXT4_I(inode)->i_inline_off = 0;
4036 }
4037
4038 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4039 {
4040         struct ext4_iloc iloc;
4041         struct ext4_inode *raw_inode;
4042         struct ext4_inode_info *ei;
4043         struct inode *inode;
4044         journal_t *journal = EXT4_SB(sb)->s_journal;
4045         long ret;
4046         int block;
4047         uid_t i_uid;
4048         gid_t i_gid;
4049
4050         inode = iget_locked(sb, ino);
4051         if (!inode)
4052                 return ERR_PTR(-ENOMEM);
4053         if (!(inode->i_state & I_NEW))
4054                 return inode;
4055
4056         ei = EXT4_I(inode);
4057         iloc.bh = NULL;
4058
4059         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4060         if (ret < 0)
4061                 goto bad_inode;
4062         raw_inode = ext4_raw_inode(&iloc);
4063
4064         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4065                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4066                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4067                     EXT4_INODE_SIZE(inode->i_sb)) {
4068                         EXT4_ERROR_INODE(inode, "bad extra_isize (%u != %u)",
4069                                 EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize,
4070                                 EXT4_INODE_SIZE(inode->i_sb));
4071                         ret = -EIO;
4072                         goto bad_inode;
4073                 }
4074         } else
4075                 ei->i_extra_isize = 0;
4076
4077         /* Precompute checksum seed for inode metadata */
4078         if (ext4_has_metadata_csum(sb)) {
4079                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4080                 __u32 csum;
4081                 __le32 inum = cpu_to_le32(inode->i_ino);
4082                 __le32 gen = raw_inode->i_generation;
4083                 csum = ext4_chksum(sbi, sbi->s_csum_seed, (__u8 *)&inum,
4084                                    sizeof(inum));
4085                 ei->i_csum_seed = ext4_chksum(sbi, csum, (__u8 *)&gen,
4086                                               sizeof(gen));
4087         }
4088
4089         if (!ext4_inode_csum_verify(inode, raw_inode, ei)) {
4090                 EXT4_ERROR_INODE(inode, "checksum invalid");
4091                 ret = -EIO;
4092                 goto bad_inode;
4093         }
4094
4095         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4096         i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4097         i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4098         if (!(test_opt(inode->i_sb, NO_UID32))) {
4099                 i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4100                 i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4101         }
4102         i_uid_write(inode, i_uid);
4103         i_gid_write(inode, i_gid);
4104         set_nlink(inode, le16_to_cpu(raw_inode->i_links_count));
4105
4106         ext4_clear_state_flags(ei);     /* Only relevant on 32-bit archs */
4107         ei->i_inline_off = 0;
4108         ei->i_dir_start_lookup = 0;
4109         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4110         /* We now have enough fields to check if the inode was active or not.
4111          * This is needed because nfsd might try to access dead inodes
4112          * the test is that same one that e2fsck uses
4113          * NeilBrown 1999oct15
4114          */
4115         if (inode->i_nlink == 0) {
4116                 if ((inode->i_mode == 0 ||
4117                      !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) &&
4118                     ino != EXT4_BOOT_LOADER_INO) {
4119                         /* this inode is deleted */
4120                         ret = -ESTALE;
4121                         goto bad_inode;
4122                 }
4123                 /* The only unlinked inodes we let through here have
4124                  * valid i_mode and are being read by the orphan
4125                  * recovery code: that's fine, we're about to complete
4126                  * the process of deleting those.
4127                  * OR it is the EXT4_BOOT_LOADER_INO which is
4128                  * not initialized on a new filesystem. */
4129         }
4130         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4131         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4132         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4133         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
4134                 ei->i_file_acl |=
4135                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4136         inode->i_size = ext4_isize(raw_inode);
4137         ei->i_disksize = inode->i_size;
4138 #ifdef CONFIG_QUOTA
4139         ei->i_reserved_quota = 0;
4140 #endif
4141         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4142         ei->i_block_group = iloc.block_group;
4143         ei->i_last_alloc_group = ~0;
4144         /*
4145          * NOTE! The in-memory inode i_data array is in little-endian order
4146          * even on big-endian machines: we do NOT byteswap the block numbers!
4147          */
4148         for (block = 0; block < EXT4_N_BLOCKS; block++)
4149                 ei->i_data[block] = raw_inode->i_block[block];
4150         INIT_LIST_HEAD(&ei->i_orphan);
4151
4152         /*
4153          * Set transaction id's of transactions that have to be committed
4154          * to finish f[data]sync. We set them to currently running transaction
4155          * as we cannot be sure that the inode or some of its metadata isn't
4156          * part of the transaction - the inode could have been reclaimed and
4157          * now it is reread from disk.
4158          */
4159         if (journal) {
4160                 transaction_t *transaction;
4161                 tid_t tid;
4162
4163                 read_lock(&journal->j_state_lock);
4164                 if (journal->j_running_transaction)
4165                         transaction = journal->j_running_transaction;
4166                 else
4167                         transaction = journal->j_committing_transaction;
4168                 if (transaction)
4169                         tid = transaction->t_tid;
4170                 else
4171                         tid = journal->j_commit_sequence;
4172                 read_unlock(&journal->j_state_lock);
4173                 ei->i_sync_tid = tid;
4174                 ei->i_datasync_tid = tid;
4175         }
4176
4177         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4178                 if (ei->i_extra_isize == 0) {
4179                         /* The extra space is currently unused. Use it. */
4180                         ei->i_extra_isize = sizeof(struct ext4_inode) -
4181                                             EXT4_GOOD_OLD_INODE_SIZE;
4182                 } else {
4183                         ext4_iget_extra_inode(inode, raw_inode, ei);
4184                 }
4185         }
4186
4187         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4188         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4189         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4190         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4191
4192         if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
4193                 inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
4194                 if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4195                         if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4196                                 inode->i_version |=
4197                     (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4198                 }
4199         }
4200
4201         ret = 0;
4202         if (ei->i_file_acl &&
4203             !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
4204                 EXT4_ERROR_INODE(inode, "bad extended attribute block %llu",
4205                                  ei->i_file_acl);
4206                 ret = -EIO;
4207                 goto bad_inode;
4208         } else if (!ext4_has_inline_data(inode)) {
4209                 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
4210                         if ((S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4211                             (S_ISLNK(inode->i_mode) &&
4212                              !ext4_inode_is_fast_symlink(inode))))
4213                                 /* Validate extent which is part of inode */
4214                                 ret = ext4_ext_check_inode(inode);
4215                 } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4216                            (S_ISLNK(inode->i_mode) &&
4217                             !ext4_inode_is_fast_symlink(inode))) {
4218                         /* Validate block references which are part of inode */
4219                         ret = ext4_ind_check_inode(inode);
4220                 }
4221         }
4222         if (ret)
4223                 goto bad_inode;
4224
4225         if (S_ISREG(inode->i_mode)) {
4226                 inode->i_op = &ext4_file_inode_operations;
4227                 inode->i_fop = &ext4_file_operations;
4228                 ext4_set_aops(inode);
4229         } else if (S_ISDIR(inode->i_mode)) {
4230                 inode->i_op = &ext4_dir_inode_operations;
4231                 inode->i_fop = &ext4_dir_operations;
4232         } else if (S_ISLNK(inode->i_mode)) {
4233                 if (ext4_encrypted_inode(inode)) {
4234                         inode->i_op = &ext4_encrypted_symlink_inode_operations;
4235                         ext4_set_aops(inode);
4236                 } else if (ext4_inode_is_fast_symlink(inode)) {
4237                         inode->i_link = (char *)ei->i_data;
4238                         inode->i_op = &ext4_fast_symlink_inode_operations;
4239                         nd_terminate_link(ei->i_data, inode->i_size,
4240                                 sizeof(ei->i_data) - 1);
4241                 } else {
4242                         inode->i_op = &ext4_symlink_inode_operations;
4243                         ext4_set_aops(inode);
4244                 }
4245         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
4246               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
4247                 inode->i_op = &ext4_special_inode_operations;
4248                 if (raw_inode->i_block[0])
4249                         init_special_inode(inode, inode->i_mode,
4250                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4251                 else
4252                         init_special_inode(inode, inode->i_mode,
4253                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4254         } else if (ino == EXT4_BOOT_LOADER_INO) {
4255                 make_bad_inode(inode);
4256         } else {
4257                 ret = -EIO;
4258                 EXT4_ERROR_INODE(inode, "bogus i_mode (%o)", inode->i_mode);
4259                 goto bad_inode;
4260         }
4261         brelse(iloc.bh);
4262         ext4_set_inode_flags(inode);
4263         unlock_new_inode(inode);
4264         return inode;
4265
4266 bad_inode:
4267         brelse(iloc.bh);
4268         iget_failed(inode);
4269         return ERR_PTR(ret);
4270 }
4271
4272 struct inode *ext4_iget_normal(struct super_block *sb, unsigned long ino)
4273 {
4274         if (ino < EXT4_FIRST_INO(sb) && ino != EXT4_ROOT_INO)
4275                 return ERR_PTR(-EIO);
4276         return ext4_iget(sb, ino);
4277 }
4278
4279 static int ext4_inode_blocks_set(handle_t *handle,
4280                                 struct ext4_inode *raw_inode,
4281                                 struct ext4_inode_info *ei)
4282 {
4283         struct inode *inode = &(ei->vfs_inode);
4284         u64 i_blocks = inode->i_blocks;
4285         struct super_block *sb = inode->i_sb;
4286
4287         if (i_blocks <= ~0U) {
4288                 /*
4289                  * i_blocks can be represented in a 32 bit variable
4290                  * as multiple of 512 bytes
4291                  */
4292                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4293                 raw_inode->i_blocks_high = 0;
4294                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4295                 return 0;
4296         }
4297         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
4298                 return -EFBIG;
4299
4300         if (i_blocks <= 0xffffffffffffULL) {
4301                 /*
4302                  * i_blocks can be represented in a 48 bit variable
4303                  * as multiple of 512 bytes
4304                  */
4305                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4306                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4307                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4308         } else {
4309                 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
4310                 /* i_block is stored in file system block size */
4311                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
4312                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4313                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4314         }
4315         return 0;
4316 }
4317
4318 struct other_inode {
4319         unsigned long           orig_ino;
4320         struct ext4_inode       *raw_inode;
4321 };
4322
4323 static int other_inode_match(struct inode * inode, unsigned long ino,
4324                              void *data)
4325 {
4326         struct other_inode *oi = (struct other_inode *) data;
4327
4328         if ((inode->i_ino != ino) ||
4329             (inode->i_state & (I_FREEING | I_WILL_FREE | I_NEW |
4330                                I_DIRTY_SYNC | I_DIRTY_DATASYNC)) ||
4331             ((inode->i_state & I_DIRTY_TIME) == 0))
4332                 return 0;
4333         spin_lock(&inode->i_lock);
4334         if (((inode->i_state & (I_FREEING | I_WILL_FREE | I_NEW |
4335                                 I_DIRTY_SYNC | I_DIRTY_DATASYNC)) == 0) &&
4336             (inode->i_state & I_DIRTY_TIME)) {
4337                 struct ext4_inode_info  *ei = EXT4_I(inode);
4338
4339                 inode->i_state &= ~(I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED);
4340                 spin_unlock(&inode->i_lock);
4341
4342                 spin_lock(&ei->i_raw_lock);
4343                 EXT4_INODE_SET_XTIME(i_ctime, inode, oi->raw_inode);
4344                 EXT4_INODE_SET_XTIME(i_mtime, inode, oi->raw_inode);
4345                 EXT4_INODE_SET_XTIME(i_atime, inode, oi->raw_inode);
4346                 ext4_inode_csum_set(inode, oi->raw_inode, ei);
4347                 spin_unlock(&ei->i_raw_lock);
4348                 trace_ext4_other_inode_update_time(inode, oi->orig_ino);
4349                 return -1;
4350         }
4351         spin_unlock(&inode->i_lock);
4352         return -1;
4353 }
4354
4355 /*
4356  * Opportunistically update the other time fields for other inodes in
4357  * the same inode table block.
4358  */
4359 static void ext4_update_other_inodes_time(struct super_block *sb,
4360                                           unsigned long orig_ino, char *buf)
4361 {
4362         struct other_inode oi;
4363         unsigned long ino;
4364         int i, inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
4365         int inode_size = EXT4_INODE_SIZE(sb);
4366
4367         oi.orig_ino = orig_ino;
4368         /*
4369          * Calculate the first inode in the inode table block.  Inode
4370          * numbers are one-based.  That is, the first inode in a block
4371          * (assuming 4k blocks and 256 byte inodes) is (n*16 + 1).
4372          */
4373         ino = ((orig_ino - 1) & ~(inodes_per_block - 1)) + 1;
4374         for (i = 0; i < inodes_per_block; i++, ino++, buf += inode_size) {
4375                 if (ino == orig_ino)
4376                         continue;
4377                 oi.raw_inode = (struct ext4_inode *) buf;
4378                 (void) find_inode_nowait(sb, ino, other_inode_match, &oi);
4379         }
4380 }
4381
4382 /*
4383  * Post the struct inode info into an on-disk inode location in the
4384  * buffer-cache.  This gobbles the caller's reference to the
4385  * buffer_head in the inode location struct.
4386  *
4387  * The caller must have write access to iloc->bh.
4388  */
4389 static int ext4_do_update_inode(handle_t *handle,
4390                                 struct inode *inode,
4391                                 struct ext4_iloc *iloc)
4392 {
4393         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
4394         struct ext4_inode_info *ei = EXT4_I(inode);
4395         struct buffer_head *bh = iloc->bh;
4396         struct super_block *sb = inode->i_sb;
4397         int err = 0, rc, block;
4398         int need_datasync = 0, set_large_file = 0;
4399         uid_t i_uid;
4400         gid_t i_gid;
4401
4402         spin_lock(&ei->i_raw_lock);
4403
4404         /* For fields not tracked in the in-memory inode,
4405          * initialise them to zero for new inodes. */
4406         if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
4407                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
4408
4409         ext4_get_inode_flags(ei);
4410         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
4411         i_uid = i_uid_read(inode);
4412         i_gid = i_gid_read(inode);
4413         if (!(test_opt(inode->i_sb, NO_UID32))) {
4414                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(i_uid));
4415                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(i_gid));
4416 /*
4417  * Fix up interoperability with old kernels. Otherwise, old inodes get
4418  * re-used with the upper 16 bits of the uid/gid intact
4419  */
4420                 if (!ei->i_dtime) {
4421                         raw_inode->i_uid_high =
4422                                 cpu_to_le16(high_16_bits(i_uid));
4423                         raw_inode->i_gid_high =
4424                                 cpu_to_le16(high_16_bits(i_gid));
4425                 } else {
4426                         raw_inode->i_uid_high = 0;
4427                         raw_inode->i_gid_high = 0;
4428                 }
4429         } else {
4430                 raw_inode->i_uid_low = cpu_to_le16(fs_high2lowuid(i_uid));
4431                 raw_inode->i_gid_low = cpu_to_le16(fs_high2lowgid(i_gid));
4432                 raw_inode->i_uid_high = 0;
4433                 raw_inode->i_gid_high = 0;
4434         }
4435         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
4436
4437         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
4438         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
4439         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
4440         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
4441
4442         err = ext4_inode_blocks_set(handle, raw_inode, ei);
4443         if (err) {
4444                 spin_unlock(&ei->i_raw_lock);
4445                 goto out_brelse;
4446         }
4447         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
4448         raw_inode->i_flags = cpu_to_le32(ei->i_flags & 0xFFFFFFFF);
4449         if (likely(!test_opt2(inode->i_sb, HURD_COMPAT)))
4450                 raw_inode->i_file_acl_high =
4451                         cpu_to_le16(ei->i_file_acl >> 32);
4452         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
4453         if (ei->i_disksize != ext4_isize(raw_inode)) {
4454                 ext4_isize_set(raw_inode, ei->i_disksize);
4455                 need_datasync = 1;
4456         }
4457         if (ei->i_disksize > 0x7fffffffULL) {
4458                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
4459                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
4460                                 EXT4_SB(sb)->s_es->s_rev_level ==
4461                     cpu_to_le32(EXT4_GOOD_OLD_REV))
4462                         set_large_file = 1;
4463         }
4464         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
4465         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
4466                 if (old_valid_dev(inode->i_rdev)) {
4467                         raw_inode->i_block[0] =
4468                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
4469                         raw_inode->i_block[1] = 0;
4470                 } else {
4471                         raw_inode->i_block[0] = 0;
4472                         raw_inode->i_block[1] =
4473                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
4474                         raw_inode->i_block[2] = 0;
4475                 }
4476         } else if (!ext4_has_inline_data(inode)) {
4477                 for (block = 0; block < EXT4_N_BLOCKS; block++)
4478                         raw_inode->i_block[block] = ei->i_data[block];
4479         }
4480
4481         if (likely(!test_opt2(inode->i_sb, HURD_COMPAT))) {
4482                 raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
4483                 if (ei->i_extra_isize) {
4484                         if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4485                                 raw_inode->i_version_hi =
4486                                         cpu_to_le32(inode->i_version >> 32);
4487                         raw_inode->i_extra_isize =
4488                                 cpu_to_le16(ei->i_extra_isize);
4489                 }
4490         }
4491         ext4_inode_csum_set(inode, raw_inode, ei);
4492         spin_unlock(&ei->i_raw_lock);
4493         if (inode->i_sb->s_flags & MS_LAZYTIME)
4494                 ext4_update_other_inodes_time(inode->i_sb, inode->i_ino,
4495                                               bh->b_data);
4496
4497         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4498         rc = ext4_handle_dirty_metadata(handle, NULL, bh);
4499         if (!err)
4500                 err = rc;
4501         ext4_clear_inode_state(inode, EXT4_STATE_NEW);
4502         if (set_large_file) {
4503                 BUFFER_TRACE(EXT4_SB(sb)->s_sbh, "get write access");
4504                 err = ext4_journal_get_write_access(handle, EXT4_SB(sb)->s_sbh);
4505                 if (err)
4506                         goto out_brelse;
4507                 ext4_update_dynamic_rev(sb);
4508                 EXT4_SET_RO_COMPAT_FEATURE(sb,
4509                                            EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
4510                 ext4_handle_sync(handle);
4511                 err = ext4_handle_dirty_super(handle, sb);
4512         }
4513         ext4_update_inode_fsync_trans(handle, inode, need_datasync);
4514 out_brelse:
4515         brelse(bh);
4516         ext4_std_error(inode->i_sb, err);
4517         return err;
4518 }
4519
4520 /*
4521  * ext4_write_inode()
4522  *
4523  * We are called from a few places:
4524  *
4525  * - Within generic_file_aio_write() -> generic_write_sync() for O_SYNC files.
4526  *   Here, there will be no transaction running. We wait for any running
4527  *   transaction to commit.
4528  *
4529  * - Within flush work (sys_sync(), kupdate and such).
4530  *   We wait on commit, if told to.
4531  *
4532  * - Within iput_final() -> write_inode_now()
4533  *   We wait on commit, if told to.
4534  *
4535  * In all cases it is actually safe for us to return without doing anything,
4536  * because the inode has been copied into a raw inode buffer in
4537  * ext4_mark_inode_dirty().  This is a correctness thing for WB_SYNC_ALL
4538  * writeback.
4539  *
4540  * Note that we are absolutely dependent upon all inode dirtiers doing the
4541  * right thing: they *must* call mark_inode_dirty() after dirtying info in
4542  * which we are interested.
4543  *
4544  * It would be a bug for them to not do this.  The code:
4545  *
4546  *      mark_inode_dirty(inode)
4547  *      stuff();
4548  *      inode->i_size = expr;
4549  *
4550  * is in error because write_inode() could occur while `stuff()' is running,
4551  * and the new i_size will be lost.  Plus the inode will no longer be on the
4552  * superblock's dirty inode list.
4553  */
4554 int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
4555 {
4556         int err;
4557
4558         if (WARN_ON_ONCE(current->flags & PF_MEMALLOC))
4559                 return 0;
4560
4561         if (EXT4_SB(inode->i_sb)->s_journal) {
4562                 if (ext4_journal_current_handle()) {
4563                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
4564                         dump_stack();
4565                         return -EIO;
4566                 }
4567
4568                 /*
4569                  * No need to force transaction in WB_SYNC_NONE mode. Also
4570                  * ext4_sync_fs() will force the commit after everything is
4571                  * written.
4572                  */
4573                 if (wbc->sync_mode != WB_SYNC_ALL || wbc->for_sync)
4574                         return 0;
4575
4576                 err = ext4_force_commit(inode->i_sb);
4577         } else {
4578                 struct ext4_iloc iloc;
4579
4580                 err = __ext4_get_inode_loc(inode, &iloc, 0);
4581                 if (err)
4582                         return err;
4583                 /*
4584                  * sync(2) will flush the whole buffer cache. No need to do
4585                  * it here separately for each inode.
4586                  */
4587                 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync)
4588                         sync_dirty_buffer(iloc.bh);
4589                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
4590                         EXT4_ERROR_INODE_BLOCK(inode, iloc.bh->b_blocknr,
4591                                          "IO error syncing inode");
4592                         err = -EIO;
4593                 }
4594                 brelse(iloc.bh);
4595         }
4596         return err;
4597 }
4598
4599 /*
4600  * In data=journal mode ext4_journalled_invalidatepage() may fail to invalidate
4601  * buffers that are attached to a page stradding i_size and are undergoing
4602  * commit. In that case we have to wait for commit to finish and try again.
4603  */
4604 static void ext4_wait_for_tail_page_commit(struct inode *inode)
4605 {
4606         struct page *page;
4607         unsigned offset;
4608         journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
4609         tid_t commit_tid = 0;
4610         int ret;
4611
4612         offset = inode->i_size & (PAGE_CACHE_SIZE - 1);
4613         /*
4614          * All buffers in the last page remain valid? Then there's nothing to
4615          * do. We do the check mainly to optimize the common PAGE_CACHE_SIZE ==
4616          * blocksize case
4617          */
4618         if (offset > PAGE_CACHE_SIZE - (1 << inode->i_blkbits))
4619                 return;
4620         while (1) {
4621                 page = find_lock_page(inode->i_mapping,
4622                                       inode->i_size >> PAGE_CACHE_SHIFT);
4623                 if (!page)
4624                         return;
4625                 ret = __ext4_journalled_invalidatepage(page, offset,
4626                                                 PAGE_CACHE_SIZE - offset);
4627                 unlock_page(page);
4628                 page_cache_release(page);
4629                 if (ret != -EBUSY)
4630                         return;
4631                 commit_tid = 0;
4632                 read_lock(&journal->j_state_lock);
4633                 if (journal->j_committing_transaction)
4634                         commit_tid = journal->j_committing_transaction->t_tid;
4635                 read_unlock(&journal->j_state_lock);
4636                 if (commit_tid)
4637                         jbd2_log_wait_commit(journal, commit_tid);
4638         }
4639 }
4640
4641 /*
4642  * ext4_setattr()
4643  *
4644  * Called from notify_change.
4645  *
4646  * We want to trap VFS attempts to truncate the file as soon as
4647  * possible.  In particular, we want to make sure that when the VFS
4648  * shrinks i_size, we put the inode on the orphan list and modify
4649  * i_disksize immediately, so that during the subsequent flushing of
4650  * dirty pages and freeing of disk blocks, we can guarantee that any
4651  * commit will leave the blocks being flushed in an unused state on
4652  * disk.  (On recovery, the inode will get truncated and the blocks will
4653  * be freed, so we have a strong guarantee that no future commit will
4654  * leave these blocks visible to the user.)
4655  *
4656  * Another thing we have to assure is that if we are in ordered mode
4657  * and inode is still attached to the committing transaction, we must
4658  * we start writeout of all the dirty pages which are being truncated.
4659  * This way we are sure that all the data written in the previous
4660  * transaction are already on disk (truncate waits for pages under
4661  * writeback).
4662  *
4663  * Called with inode->i_mutex down.
4664  */
4665 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
4666 {
4667         struct inode *inode = d_inode(dentry);
4668         int error, rc = 0;
4669         int orphan = 0;
4670         const unsigned int ia_valid = attr->ia_valid;
4671
4672         error = inode_change_ok(inode, attr);
4673         if (error)
4674                 return error;
4675
4676         if (is_quota_modification(inode, attr)) {
4677                 error = dquot_initialize(inode);
4678                 if (error)
4679                         return error;
4680         }
4681         if ((ia_valid & ATTR_UID && !uid_eq(attr->ia_uid, inode->i_uid)) ||
4682             (ia_valid & ATTR_GID && !gid_eq(attr->ia_gid, inode->i_gid))) {
4683                 handle_t *handle;
4684
4685                 /* (user+group)*(old+new) structure, inode write (sb,
4686                  * inode block, ? - but truncate inode update has it) */
4687                 handle = ext4_journal_start(inode, EXT4_HT_QUOTA,
4688                         (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb) +
4689                          EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb)) + 3);
4690                 if (IS_ERR(handle)) {
4691                         error = PTR_ERR(handle);
4692                         goto err_out;
4693                 }
4694                 error = dquot_transfer(inode, attr);
4695                 if (error) {
4696                         ext4_journal_stop(handle);
4697                         return error;
4698                 }
4699                 /* Update corresponding info in inode so that everything is in
4700                  * one transaction */
4701                 if (attr->ia_valid & ATTR_UID)
4702                         inode->i_uid = attr->ia_uid;
4703                 if (attr->ia_valid & ATTR_GID)
4704                         inode->i_gid = attr->ia_gid;
4705                 error = ext4_mark_inode_dirty(handle, inode);
4706                 ext4_journal_stop(handle);
4707         }
4708
4709         if (attr->ia_valid & ATTR_SIZE) {
4710                 handle_t *handle;
4711                 loff_t oldsize = inode->i_size;
4712                 int shrink = (attr->ia_size <= inode->i_size);
4713
4714                 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4715                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4716
4717                         if (attr->ia_size > sbi->s_bitmap_maxbytes)
4718                                 return -EFBIG;
4719                 }
4720                 if (!S_ISREG(inode->i_mode))
4721                         return -EINVAL;
4722
4723                 if (IS_I_VERSION(inode) && attr->ia_size != inode->i_size)
4724                         inode_inc_iversion(inode);
4725
4726                 if (ext4_should_order_data(inode) &&
4727                     (attr->ia_size < inode->i_size)) {
4728                         error = ext4_begin_ordered_truncate(inode,
4729                                                             attr->ia_size);
4730                         if (error)
4731                                 goto err_out;
4732                 }
4733                 if (attr->ia_size != inode->i_size) {
4734                         handle = ext4_journal_start(inode, EXT4_HT_INODE, 3);
4735                         if (IS_ERR(handle)) {
4736                                 error = PTR_ERR(handle);
4737                                 goto err_out;
4738                         }
4739                         if (ext4_handle_valid(handle) && shrink) {
4740                                 error = ext4_orphan_add(handle, inode);
4741                                 orphan = 1;
4742                         }
4743                         /*
4744                          * Update c/mtime on truncate up, ext4_truncate() will
4745                          * update c/mtime in shrink case below
4746                          */
4747                         if (!shrink) {
4748                                 inode->i_mtime = ext4_current_time(inode);
4749                                 inode->i_ctime = inode->i_mtime;
4750                         }
4751                         down_write(&EXT4_I(inode)->i_data_sem);
4752                         EXT4_I(inode)->i_disksize = attr->ia_size;
4753                         rc = ext4_mark_inode_dirty(handle, inode);
4754                         if (!error)
4755                                 error = rc;
4756                         /*
4757                          * We have to update i_size under i_data_sem together
4758                          * with i_disksize to avoid races with writeback code
4759                          * running ext4_wb_update_i_disksize().
4760                          */
4761                         if (!error)
4762                                 i_size_write(inode, attr->ia_size);
4763                         up_write(&EXT4_I(inode)->i_data_sem);
4764                         ext4_journal_stop(handle);
4765                         if (error) {
4766                                 if (orphan)
4767                                         ext4_orphan_del(NULL, inode);
4768                                 goto err_out;
4769                         }
4770                 }
4771                 if (!shrink)
4772                         pagecache_isize_extended(inode, oldsize, inode->i_size);
4773
4774                 /*
4775                  * Blocks are going to be removed from the inode. Wait
4776                  * for dio in flight.  Temporarily disable
4777                  * dioread_nolock to prevent livelock.
4778                  */
4779                 if (orphan) {
4780                         if (!ext4_should_journal_data(inode)) {
4781                                 ext4_inode_block_unlocked_dio(inode);
4782                                 inode_dio_wait(inode);
4783                                 ext4_inode_resume_unlocked_dio(inode);
4784                         } else
4785                                 ext4_wait_for_tail_page_commit(inode);
4786                 }
4787                 /*
4788                  * Truncate pagecache after we've waited for commit
4789                  * in data=journal mode to make pages freeable.
4790                  */
4791                 truncate_pagecache(inode, inode->i_size);
4792                 if (shrink)
4793                         ext4_truncate(inode);
4794         }
4795
4796         if (!rc) {
4797                 setattr_copy(inode, attr);
4798                 mark_inode_dirty(inode);
4799         }
4800
4801         /*
4802          * If the call to ext4_truncate failed to get a transaction handle at
4803          * all, we need to clean up the in-core orphan list manually.
4804          */
4805         if (orphan && inode->i_nlink)
4806                 ext4_orphan_del(NULL, inode);
4807
4808         if (!rc && (ia_valid & ATTR_MODE))
4809                 rc = posix_acl_chmod(inode, inode->i_mode);
4810
4811 err_out:
4812         ext4_std_error(inode->i_sb, error);
4813         if (!error)
4814                 error = rc;
4815         return error;
4816 }
4817
4818 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
4819                  struct kstat *stat)
4820 {
4821         struct inode *inode;
4822         unsigned long long delalloc_blocks;
4823
4824         inode = d_inode(dentry);
4825         generic_fillattr(inode, stat);
4826
4827         /*
4828          * If there is inline data in the inode, the inode will normally not
4829          * have data blocks allocated (it may have an external xattr block).
4830          * Report at least one sector for such files, so tools like tar, rsync,
4831          * others doen't incorrectly think the file is completely sparse.
4832          */
4833         if (unlikely(ext4_has_inline_data(inode)))
4834                 stat->blocks += (stat->size + 511) >> 9;
4835
4836         /*
4837          * We can't update i_blocks if the block allocation is delayed
4838          * otherwise in the case of system crash before the real block
4839          * allocation is done, we will have i_blocks inconsistent with
4840          * on-disk file blocks.
4841          * We always keep i_blocks updated together with real
4842          * allocation. But to not confuse with user, stat
4843          * will return the blocks that include the delayed allocation
4844          * blocks for this file.
4845          */
4846         delalloc_blocks = EXT4_C2B(EXT4_SB(inode->i_sb),
4847                                    EXT4_I(inode)->i_reserved_data_blocks);
4848         stat->blocks += delalloc_blocks << (inode->i_sb->s_blocksize_bits - 9);
4849         return 0;
4850 }
4851
4852 static int ext4_index_trans_blocks(struct inode *inode, int lblocks,
4853                                    int pextents)
4854 {
4855         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4856                 return ext4_ind_trans_blocks(inode, lblocks);
4857         return ext4_ext_index_trans_blocks(inode, pextents);
4858 }
4859
4860 /*
4861  * Account for index blocks, block groups bitmaps and block group
4862  * descriptor blocks if modify datablocks and index blocks
4863  * worse case, the indexs blocks spread over different block groups
4864  *
4865  * If datablocks are discontiguous, they are possible to spread over
4866  * different block groups too. If they are contiguous, with flexbg,
4867  * they could still across block group boundary.
4868  *
4869  * Also account for superblock, inode, quota and xattr blocks
4870  */
4871 static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
4872                                   int pextents)
4873 {
4874         ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
4875         int gdpblocks;
4876         int idxblocks;
4877         int ret = 0;
4878
4879         /*
4880          * How many index blocks need to touch to map @lblocks logical blocks
4881          * to @pextents physical extents?
4882          */
4883         idxblocks = ext4_index_trans_blocks(inode, lblocks, pextents);
4884
4885         ret = idxblocks;
4886
4887         /*
4888          * Now let's see how many group bitmaps and group descriptors need
4889          * to account
4890          */
4891         groups = idxblocks + pextents;
4892         gdpblocks = groups;
4893         if (groups > ngroups)
4894                 groups = ngroups;
4895         if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
4896                 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
4897
4898         /* bitmaps and block group descriptor blocks */
4899         ret += groups + gdpblocks;
4900
4901         /* Blocks for super block, inode, quota and xattr blocks */
4902         ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
4903
4904         return ret;
4905 }
4906
4907 /*
4908  * Calculate the total number of credits to reserve to fit
4909  * the modification of a single pages into a single transaction,
4910  * which may include multiple chunks of block allocations.
4911  *
4912  * This could be called via ext4_write_begin()
4913  *
4914  * We need to consider the worse case, when
4915  * one new block per extent.
4916  */
4917 int ext4_writepage_trans_blocks(struct inode *inode)
4918 {
4919         int bpp = ext4_journal_blocks_per_page(inode);
4920         int ret;
4921
4922         ret = ext4_meta_trans_blocks(inode, bpp, bpp);
4923
4924         /* Account for data blocks for journalled mode */
4925         if (ext4_should_journal_data(inode))
4926                 ret += bpp;
4927         return ret;
4928 }
4929
4930 /*
4931  * Calculate the journal credits for a chunk of data modification.
4932  *
4933  * This is called from DIO, fallocate or whoever calling
4934  * ext4_map_blocks() to map/allocate a chunk of contiguous disk blocks.
4935  *
4936  * journal buffers for data blocks are not included here, as DIO
4937  * and fallocate do no need to journal data buffers.
4938  */
4939 int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
4940 {
4941         return ext4_meta_trans_blocks(inode, nrblocks, 1);
4942 }
4943
4944 /*
4945  * The caller must have previously called ext4_reserve_inode_write().
4946  * Give this, we know that the caller already has write access to iloc->bh.
4947  */
4948 int ext4_mark_iloc_dirty(handle_t *handle,
4949                          struct inode *inode, struct ext4_iloc *iloc)
4950 {
4951         int err = 0;
4952
4953         if (IS_I_VERSION(inode))
4954                 inode_inc_iversion(inode);
4955
4956         /* the do_update_inode consumes one bh->b_count */
4957         get_bh(iloc->bh);
4958
4959         /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
4960         err = ext4_do_update_inode(handle, inode, iloc);
4961         put_bh(iloc->bh);
4962         return err;
4963 }
4964
4965 /*
4966  * On success, We end up with an outstanding reference count against
4967  * iloc->bh.  This _must_ be cleaned up later.
4968  */
4969
4970 int
4971 ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
4972                          struct ext4_iloc *iloc)
4973 {
4974         int err;
4975
4976         err = ext4_get_inode_loc(inode, iloc);
4977         if (!err) {
4978                 BUFFER_TRACE(iloc->bh, "get_write_access");
4979                 err = ext4_journal_get_write_access(handle, iloc->bh);
4980                 if (err) {
4981                         brelse(iloc->bh);
4982                         iloc->bh = NULL;
4983                 }
4984         }
4985         ext4_std_error(inode->i_sb, err);
4986         return err;
4987 }
4988
4989 /*
4990  * Expand an inode by new_extra_isize bytes.
4991  * Returns 0 on success or negative error number on failure.
4992  */
4993 static int ext4_expand_extra_isize(struct inode *inode,
4994                                    unsigned int new_extra_isize,
4995                                    struct ext4_iloc iloc,
4996                                    handle_t *handle)
4997 {
4998         struct ext4_inode *raw_inode;
4999         struct ext4_xattr_ibody_header *header;
5000
5001         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
5002                 return 0;
5003
5004         raw_inode = ext4_raw_inode(&iloc);
5005
5006         header = IHDR(inode, raw_inode);
5007
5008         /* No extended attributes present */
5009         if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR) ||
5010             header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
5011                 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
5012                         new_extra_isize);
5013                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
5014                 return 0;
5015         }
5016
5017         /* try to expand with EAs present */
5018         return ext4_expand_extra_isize_ea(inode, new_extra_isize,
5019                                           raw_inode, handle);
5020 }
5021
5022 /*
5023  * What we do here is to mark the in-core inode as clean with respect to inode
5024  * dirtiness (it may still be data-dirty).
5025  * This means that the in-core inode may be reaped by prune_icache
5026  * without having to perform any I/O.  This is a very good thing,
5027  * because *any* task may call prune_icache - even ones which
5028  * have a transaction open against a different journal.
5029  *
5030  * Is this cheating?  Not really.  Sure, we haven't written the
5031  * inode out, but prune_icache isn't a user-visible syncing function.
5032  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
5033  * we start and wait on commits.
5034  */
5035 int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
5036 {
5037         struct ext4_iloc iloc;
5038         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5039         static unsigned int mnt_count;
5040         int err, ret;
5041
5042         might_sleep();
5043         trace_ext4_mark_inode_dirty(inode, _RET_IP_);
5044         err = ext4_reserve_inode_write(handle, inode, &iloc);
5045         if (ext4_handle_valid(handle) &&
5046             EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
5047             !ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND)) {
5048                 /*
5049                  * We need extra buffer credits since we may write into EA block
5050                  * with this same handle. If journal_extend fails, then it will
5051                  * only result in a minor loss of functionality for that inode.
5052                  * If this is felt to be critical, then e2fsck should be run to
5053                  * force a large enough s_min_extra_isize.
5054                  */
5055                 if ((jbd2_journal_extend(handle,
5056                              EXT4_DATA_TRANS_BLOCKS(inode->i_sb))) == 0) {
5057                         ret = ext4_expand_extra_isize(inode,
5058                                                       sbi->s_want_extra_isize,
5059                                                       iloc, handle);
5060                         if (ret) {
5061                                 ext4_set_inode_state(inode,
5062                                                      EXT4_STATE_NO_EXPAND);
5063                                 if (mnt_count !=
5064                                         le16_to_cpu(sbi->s_es->s_mnt_count)) {
5065                                         ext4_warning(inode->i_sb,
5066                                         "Unable to expand inode %lu. Delete"
5067                                         " some EAs or run e2fsck.",
5068                                         inode->i_ino);
5069                                         mnt_count =
5070                                           le16_to_cpu(sbi->s_es->s_mnt_count);
5071                                 }
5072                         }
5073                 }
5074         }
5075         if (!err)
5076                 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
5077         return err;
5078 }
5079
5080 /*
5081  * ext4_dirty_inode() is called from __mark_inode_dirty()
5082  *
5083  * We're really interested in the case where a file is being extended.
5084  * i_size has been changed by generic_commit_write() and we thus need
5085  * to include the updated inode in the current transaction.
5086  *
5087  * Also, dquot_alloc_block() will always dirty the inode when blocks
5088  * are allocated to the file.
5089  *
5090  * If the inode is marked synchronous, we don't honour that here - doing
5091  * so would cause a commit on atime updates, which we don't bother doing.
5092  * We handle synchronous inodes at the highest possible level.
5093  *
5094  * If only the I_DIRTY_TIME flag is set, we can skip everything.  If
5095  * I_DIRTY_TIME and I_DIRTY_SYNC is set, the only inode fields we need
5096  * to copy into the on-disk inode structure are the timestamp files.
5097  */
5098 void ext4_dirty_inode(struct inode *inode, int flags)
5099 {
5100         handle_t *handle;
5101
5102         if (flags == I_DIRTY_TIME)
5103                 return;
5104         handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
5105         if (IS_ERR(handle))
5106                 goto out;
5107
5108         ext4_mark_inode_dirty(handle, inode);
5109
5110         ext4_journal_stop(handle);
5111 out:
5112         return;
5113 }
5114
5115 #if 0
5116 /*
5117  * Bind an inode's backing buffer_head into this transaction, to prevent
5118  * it from being flushed to disk early.  Unlike
5119  * ext4_reserve_inode_write, this leaves behind no bh reference and
5120  * returns no iloc structure, so the caller needs to repeat the iloc
5121  * lookup to mark the inode dirty later.
5122  */
5123 static int ext4_pin_inode(handle_t *handle, struct inode *inode)
5124 {
5125         struct ext4_iloc iloc;
5126
5127         int err = 0;
5128         if (handle) {
5129                 err = ext4_get_inode_loc(inode, &iloc);
5130                 if (!err) {
5131                         BUFFER_TRACE(iloc.bh, "get_write_access");
5132                         err = jbd2_journal_get_write_access(handle, iloc.bh);
5133                         if (!err)
5134                                 err = ext4_handle_dirty_metadata(handle,
5135                                                                  NULL,
5136                                                                  iloc.bh);
5137                         brelse(iloc.bh);
5138                 }
5139         }
5140         ext4_std_error(inode->i_sb, err);
5141         return err;
5142 }
5143 #endif
5144
5145 int ext4_change_inode_journal_flag(struct inode *inode, int val)
5146 {
5147         journal_t *journal;
5148         handle_t *handle;
5149         int err;
5150
5151         /*
5152          * We have to be very careful here: changing a data block's
5153          * journaling status dynamically is dangerous.  If we write a
5154          * data block to the journal, change the status and then delete
5155          * that block, we risk forgetting to revoke the old log record
5156          * from the journal and so a subsequent replay can corrupt data.
5157          * So, first we make sure that the journal is empty and that
5158          * nobody is changing anything.
5159          */
5160
5161         journal = EXT4_JOURNAL(inode);
5162         if (!journal)
5163                 return 0;
5164         if (is_journal_aborted(journal))
5165                 return -EROFS;
5166         /* We have to allocate physical blocks for delalloc blocks
5167          * before flushing journal. otherwise delalloc blocks can not
5168          * be allocated any more. even more truncate on delalloc blocks
5169          * could trigger BUG by flushing delalloc blocks in journal.
5170          * There is no delalloc block in non-journal data mode.
5171          */
5172         if (val && test_opt(inode->i_sb, DELALLOC)) {
5173                 err = ext4_alloc_da_blocks(inode);
5174                 if (err < 0)
5175                         return err;
5176         }
5177
5178         /* Wait for all existing dio workers */
5179         ext4_inode_block_unlocked_dio(inode);
5180         inode_dio_wait(inode);
5181
5182         jbd2_journal_lock_updates(journal);
5183
5184         /*
5185          * OK, there are no updates running now, and all cached data is
5186          * synced to disk.  We are now in a completely consistent state
5187          * which doesn't have anything in the journal, and we know that
5188          * no filesystem updates are running, so it is safe to modify
5189          * the inode's in-core data-journaling state flag now.
5190          */
5191
5192         if (val)
5193                 ext4_set_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
5194         else {
5195                 err = jbd2_journal_flush(journal);
5196                 if (err < 0) {
5197                         jbd2_journal_unlock_updates(journal);
5198                         ext4_inode_resume_unlocked_dio(inode);
5199                         return err;
5200                 }
5201                 ext4_clear_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
5202         }
5203         ext4_set_aops(inode);
5204
5205         jbd2_journal_unlock_updates(journal);
5206         ext4_inode_resume_unlocked_dio(inode);
5207
5208         /* Finally we can mark the inode as dirty. */
5209
5210         handle = ext4_journal_start(inode, EXT4_HT_INODE, 1);
5211         if (IS_ERR(handle))
5212                 return PTR_ERR(handle);
5213
5214         err = ext4_mark_inode_dirty(handle, inode);
5215         ext4_handle_sync(handle);
5216         ext4_journal_stop(handle);
5217         ext4_std_error(inode->i_sb, err);
5218
5219         return err;
5220 }
5221
5222 static int ext4_bh_unmapped(handle_t *handle, struct buffer_head *bh)
5223 {
5224         return !buffer_mapped(bh);
5225 }
5226
5227 int ext4_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
5228 {
5229         struct page *page = vmf->page;
5230         loff_t size;
5231         unsigned long len;
5232         int ret;
5233         struct file *file = vma->vm_file;
5234         struct inode *inode = file_inode(file);
5235         struct address_space *mapping = inode->i_mapping;
5236         handle_t *handle;
5237         get_block_t *get_block;
5238         int retries = 0;
5239
5240         sb_start_pagefault(inode->i_sb);
5241         file_update_time(vma->vm_file);
5242         /* Delalloc case is easy... */
5243         if (test_opt(inode->i_sb, DELALLOC) &&
5244             !ext4_should_journal_data(inode) &&
5245             !ext4_nonda_switch(inode->i_sb)) {
5246                 do {
5247                         ret = __block_page_mkwrite(vma, vmf,
5248                                                    ext4_da_get_block_prep);
5249                 } while (ret == -ENOSPC &&
5250                        ext4_should_retry_alloc(inode->i_sb, &retries));
5251                 goto out_ret;
5252         }
5253
5254         lock_page(page);
5255         size = i_size_read(inode);
5256         /* Page got truncated from under us? */
5257         if (page->mapping != mapping || page_offset(page) > size) {
5258                 unlock_page(page);
5259                 ret = VM_FAULT_NOPAGE;
5260                 goto out;
5261         }
5262
5263         if (page->index == size >> PAGE_CACHE_SHIFT)
5264                 len = size & ~PAGE_CACHE_MASK;
5265         else
5266                 len = PAGE_CACHE_SIZE;
5267         /*
5268          * Return if we have all the buffers mapped. This avoids the need to do
5269          * journal_start/journal_stop which can block and take a long time
5270          */
5271         if (page_has_buffers(page)) {
5272                 if (!ext4_walk_page_buffers(NULL, page_buffers(page),
5273                                             0, len, NULL,
5274                                             ext4_bh_unmapped)) {
5275                         /* Wait so that we don't change page under IO */
5276                         wait_for_stable_page(page);
5277                         ret = VM_FAULT_LOCKED;
5278                         goto out;
5279                 }
5280         }
5281         unlock_page(page);
5282         /* OK, we need to fill the hole... */
5283         if (ext4_should_dioread_nolock(inode))
5284                 get_block = ext4_get_block_write;
5285         else
5286                 get_block = ext4_get_block;
5287 retry_alloc:
5288         handle = ext4_journal_start(inode, EXT4_HT_WRITE_PAGE,
5289                                     ext4_writepage_trans_blocks(inode));
5290         if (IS_ERR(handle)) {
5291                 ret = VM_FAULT_SIGBUS;
5292                 goto out;
5293         }
5294         ret = __block_page_mkwrite(vma, vmf, get_block);
5295         if (!ret && ext4_should_journal_data(inode)) {
5296                 if (ext4_walk_page_buffers(handle, page_buffers(page), 0,
5297                           PAGE_CACHE_SIZE, NULL, do_journal_get_write_access)) {
5298                         unlock_page(page);
5299                         ret = VM_FAULT_SIGBUS;
5300                         ext4_journal_stop(handle);
5301                         goto out;
5302                 }
5303                 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
5304         }
5305         ext4_journal_stop(handle);
5306         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
5307                 goto retry_alloc;
5308 out_ret:
5309         ret = block_page_mkwrite_return(ret);
5310 out:
5311         sb_end_pagefault(inode->i_sb);
5312         return ret;
5313 }