block: separate priority boosting from REQ_META
[linux-2.6.git] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  64-bit file support on 64-bit platforms by Jakub Jelinek
16  *      (jj@sunsite.ms.mff.cuni.cz)
17  *
18  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
19  */
20
21 #include <linux/module.h>
22 #include <linux/fs.h>
23 #include <linux/time.h>
24 #include <linux/jbd2.h>
25 #include <linux/highuid.h>
26 #include <linux/pagemap.h>
27 #include <linux/quotaops.h>
28 #include <linux/string.h>
29 #include <linux/buffer_head.h>
30 #include <linux/writeback.h>
31 #include <linux/pagevec.h>
32 #include <linux/mpage.h>
33 #include <linux/namei.h>
34 #include <linux/uio.h>
35 #include <linux/bio.h>
36 #include <linux/workqueue.h>
37 #include <linux/kernel.h>
38 #include <linux/printk.h>
39 #include <linux/slab.h>
40 #include <linux/ratelimit.h>
41
42 #include "ext4_jbd2.h"
43 #include "xattr.h"
44 #include "acl.h"
45 #include "ext4_extents.h"
46 #include "truncate.h"
47
48 #include <trace/events/ext4.h>
49
50 #define MPAGE_DA_EXTENT_TAIL 0x01
51
52 static inline int ext4_begin_ordered_truncate(struct inode *inode,
53                                               loff_t new_size)
54 {
55         trace_ext4_begin_ordered_truncate(inode, new_size);
56         /*
57          * If jinode is zero, then we never opened the file for
58          * writing, so there's no need to call
59          * jbd2_journal_begin_ordered_truncate() since there's no
60          * outstanding writes we need to flush.
61          */
62         if (!EXT4_I(inode)->jinode)
63                 return 0;
64         return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
65                                                    EXT4_I(inode)->jinode,
66                                                    new_size);
67 }
68
69 static void ext4_invalidatepage(struct page *page, unsigned long offset);
70 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
71                                    struct buffer_head *bh_result, int create);
72 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
73 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
74 static int __ext4_journalled_writepage(struct page *page, unsigned int len);
75 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
76
77 /*
78  * Test whether an inode is a fast symlink.
79  */
80 static int ext4_inode_is_fast_symlink(struct inode *inode)
81 {
82         int ea_blocks = EXT4_I(inode)->i_file_acl ?
83                 (inode->i_sb->s_blocksize >> 9) : 0;
84
85         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
86 }
87
88 /*
89  * Restart the transaction associated with *handle.  This does a commit,
90  * so before we call here everything must be consistently dirtied against
91  * this transaction.
92  */
93 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
94                                  int nblocks)
95 {
96         int ret;
97
98         /*
99          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
100          * moment, get_block can be called only for blocks inside i_size since
101          * page cache has been already dropped and writes are blocked by
102          * i_mutex. So we can safely drop the i_data_sem here.
103          */
104         BUG_ON(EXT4_JOURNAL(inode) == NULL);
105         jbd_debug(2, "restarting handle %p\n", handle);
106         up_write(&EXT4_I(inode)->i_data_sem);
107         ret = ext4_journal_restart(handle, nblocks);
108         down_write(&EXT4_I(inode)->i_data_sem);
109         ext4_discard_preallocations(inode);
110
111         return ret;
112 }
113
114 /*
115  * Called at the last iput() if i_nlink is zero.
116  */
117 void ext4_evict_inode(struct inode *inode)
118 {
119         handle_t *handle;
120         int err;
121
122         trace_ext4_evict_inode(inode);
123
124         mutex_lock(&inode->i_mutex);
125         ext4_flush_completed_IO(inode);
126         mutex_unlock(&inode->i_mutex);
127         ext4_ioend_wait(inode);
128
129         if (inode->i_nlink) {
130                 /*
131                  * When journalling data dirty buffers are tracked only in the
132                  * journal. So although mm thinks everything is clean and
133                  * ready for reaping the inode might still have some pages to
134                  * write in the running transaction or waiting to be
135                  * checkpointed. Thus calling jbd2_journal_invalidatepage()
136                  * (via truncate_inode_pages()) to discard these buffers can
137                  * cause data loss. Also even if we did not discard these
138                  * buffers, we would have no way to find them after the inode
139                  * is reaped and thus user could see stale data if he tries to
140                  * read them before the transaction is checkpointed. So be
141                  * careful and force everything to disk here... We use
142                  * ei->i_datasync_tid to store the newest transaction
143                  * containing inode's data.
144                  *
145                  * Note that directories do not have this problem because they
146                  * don't use page cache.
147                  */
148                 if (ext4_should_journal_data(inode) &&
149                     (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode))) {
150                         journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
151                         tid_t commit_tid = EXT4_I(inode)->i_datasync_tid;
152
153                         jbd2_log_start_commit(journal, commit_tid);
154                         jbd2_log_wait_commit(journal, commit_tid);
155                         filemap_write_and_wait(&inode->i_data);
156                 }
157                 truncate_inode_pages(&inode->i_data, 0);
158                 goto no_delete;
159         }
160
161         if (!is_bad_inode(inode))
162                 dquot_initialize(inode);
163
164         if (ext4_should_order_data(inode))
165                 ext4_begin_ordered_truncate(inode, 0);
166         truncate_inode_pages(&inode->i_data, 0);
167
168         if (is_bad_inode(inode))
169                 goto no_delete;
170
171         handle = ext4_journal_start(inode, ext4_blocks_for_truncate(inode)+3);
172         if (IS_ERR(handle)) {
173                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
174                 /*
175                  * If we're going to skip the normal cleanup, we still need to
176                  * make sure that the in-core orphan linked list is properly
177                  * cleaned up.
178                  */
179                 ext4_orphan_del(NULL, inode);
180                 goto no_delete;
181         }
182
183         if (IS_SYNC(inode))
184                 ext4_handle_sync(handle);
185         inode->i_size = 0;
186         err = ext4_mark_inode_dirty(handle, inode);
187         if (err) {
188                 ext4_warning(inode->i_sb,
189                              "couldn't mark inode dirty (err %d)", err);
190                 goto stop_handle;
191         }
192         if (inode->i_blocks)
193                 ext4_truncate(inode);
194
195         /*
196          * ext4_ext_truncate() doesn't reserve any slop when it
197          * restarts journal transactions; therefore there may not be
198          * enough credits left in the handle to remove the inode from
199          * the orphan list and set the dtime field.
200          */
201         if (!ext4_handle_has_enough_credits(handle, 3)) {
202                 err = ext4_journal_extend(handle, 3);
203                 if (err > 0)
204                         err = ext4_journal_restart(handle, 3);
205                 if (err != 0) {
206                         ext4_warning(inode->i_sb,
207                                      "couldn't extend journal (err %d)", err);
208                 stop_handle:
209                         ext4_journal_stop(handle);
210                         ext4_orphan_del(NULL, inode);
211                         goto no_delete;
212                 }
213         }
214
215         /*
216          * Kill off the orphan record which ext4_truncate created.
217          * AKPM: I think this can be inside the above `if'.
218          * Note that ext4_orphan_del() has to be able to cope with the
219          * deletion of a non-existent orphan - this is because we don't
220          * know if ext4_truncate() actually created an orphan record.
221          * (Well, we could do this if we need to, but heck - it works)
222          */
223         ext4_orphan_del(handle, inode);
224         EXT4_I(inode)->i_dtime  = get_seconds();
225
226         /*
227          * One subtle ordering requirement: if anything has gone wrong
228          * (transaction abort, IO errors, whatever), then we can still
229          * do these next steps (the fs will already have been marked as
230          * having errors), but we can't free the inode if the mark_dirty
231          * fails.
232          */
233         if (ext4_mark_inode_dirty(handle, inode))
234                 /* If that failed, just do the required in-core inode clear. */
235                 ext4_clear_inode(inode);
236         else
237                 ext4_free_inode(handle, inode);
238         ext4_journal_stop(handle);
239         return;
240 no_delete:
241         ext4_clear_inode(inode);        /* We must guarantee clearing of inode... */
242 }
243
244 #ifdef CONFIG_QUOTA
245 qsize_t *ext4_get_reserved_space(struct inode *inode)
246 {
247         return &EXT4_I(inode)->i_reserved_quota;
248 }
249 #endif
250
251 /*
252  * Calculate the number of metadata blocks need to reserve
253  * to allocate a block located at @lblock
254  */
255 static int ext4_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
256 {
257         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
258                 return ext4_ext_calc_metadata_amount(inode, lblock);
259
260         return ext4_ind_calc_metadata_amount(inode, lblock);
261 }
262
263 /*
264  * Called with i_data_sem down, which is important since we can call
265  * ext4_discard_preallocations() from here.
266  */
267 void ext4_da_update_reserve_space(struct inode *inode,
268                                         int used, int quota_claim)
269 {
270         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
271         struct ext4_inode_info *ei = EXT4_I(inode);
272
273         spin_lock(&ei->i_block_reservation_lock);
274         trace_ext4_da_update_reserve_space(inode, used);
275         if (unlikely(used > ei->i_reserved_data_blocks)) {
276                 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
277                          "with only %d reserved data blocks\n",
278                          __func__, inode->i_ino, used,
279                          ei->i_reserved_data_blocks);
280                 WARN_ON(1);
281                 used = ei->i_reserved_data_blocks;
282         }
283
284         /* Update per-inode reservations */
285         ei->i_reserved_data_blocks -= used;
286         ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
287         percpu_counter_sub(&sbi->s_dirtyblocks_counter,
288                            used + ei->i_allocated_meta_blocks);
289         ei->i_allocated_meta_blocks = 0;
290
291         if (ei->i_reserved_data_blocks == 0) {
292                 /*
293                  * We can release all of the reserved metadata blocks
294                  * only when we have written all of the delayed
295                  * allocation blocks.
296                  */
297                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
298                                    ei->i_reserved_meta_blocks);
299                 ei->i_reserved_meta_blocks = 0;
300                 ei->i_da_metadata_calc_len = 0;
301         }
302         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
303
304         /* Update quota subsystem for data blocks */
305         if (quota_claim)
306                 dquot_claim_block(inode, used);
307         else {
308                 /*
309                  * We did fallocate with an offset that is already delayed
310                  * allocated. So on delayed allocated writeback we should
311                  * not re-claim the quota for fallocated blocks.
312                  */
313                 dquot_release_reservation_block(inode, used);
314         }
315
316         /*
317          * If we have done all the pending block allocations and if
318          * there aren't any writers on the inode, we can discard the
319          * inode's preallocations.
320          */
321         if ((ei->i_reserved_data_blocks == 0) &&
322             (atomic_read(&inode->i_writecount) == 0))
323                 ext4_discard_preallocations(inode);
324 }
325
326 static int __check_block_validity(struct inode *inode, const char *func,
327                                 unsigned int line,
328                                 struct ext4_map_blocks *map)
329 {
330         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
331                                    map->m_len)) {
332                 ext4_error_inode(inode, func, line, map->m_pblk,
333                                  "lblock %lu mapped to illegal pblock "
334                                  "(length %d)", (unsigned long) map->m_lblk,
335                                  map->m_len);
336                 return -EIO;
337         }
338         return 0;
339 }
340
341 #define check_block_validity(inode, map)        \
342         __check_block_validity((inode), __func__, __LINE__, (map))
343
344 /*
345  * Return the number of contiguous dirty pages in a given inode
346  * starting at page frame idx.
347  */
348 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
349                                     unsigned int max_pages)
350 {
351         struct address_space *mapping = inode->i_mapping;
352         pgoff_t index;
353         struct pagevec pvec;
354         pgoff_t num = 0;
355         int i, nr_pages, done = 0;
356
357         if (max_pages == 0)
358                 return 0;
359         pagevec_init(&pvec, 0);
360         while (!done) {
361                 index = idx;
362                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
363                                               PAGECACHE_TAG_DIRTY,
364                                               (pgoff_t)PAGEVEC_SIZE);
365                 if (nr_pages == 0)
366                         break;
367                 for (i = 0; i < nr_pages; i++) {
368                         struct page *page = pvec.pages[i];
369                         struct buffer_head *bh, *head;
370
371                         lock_page(page);
372                         if (unlikely(page->mapping != mapping) ||
373                             !PageDirty(page) ||
374                             PageWriteback(page) ||
375                             page->index != idx) {
376                                 done = 1;
377                                 unlock_page(page);
378                                 break;
379                         }
380                         if (page_has_buffers(page)) {
381                                 bh = head = page_buffers(page);
382                                 do {
383                                         if (!buffer_delay(bh) &&
384                                             !buffer_unwritten(bh))
385                                                 done = 1;
386                                         bh = bh->b_this_page;
387                                 } while (!done && (bh != head));
388                         }
389                         unlock_page(page);
390                         if (done)
391                                 break;
392                         idx++;
393                         num++;
394                         if (num >= max_pages) {
395                                 done = 1;
396                                 break;
397                         }
398                 }
399                 pagevec_release(&pvec);
400         }
401         return num;
402 }
403
404 /*
405  * The ext4_map_blocks() function tries to look up the requested blocks,
406  * and returns if the blocks are already mapped.
407  *
408  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
409  * and store the allocated blocks in the result buffer head and mark it
410  * mapped.
411  *
412  * If file type is extents based, it will call ext4_ext_map_blocks(),
413  * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
414  * based files
415  *
416  * On success, it returns the number of blocks being mapped or allocate.
417  * if create==0 and the blocks are pre-allocated and uninitialized block,
418  * the result buffer head is unmapped. If the create ==1, it will make sure
419  * the buffer head is mapped.
420  *
421  * It returns 0 if plain look up failed (blocks have not been allocated), in
422  * that casem, buffer head is unmapped
423  *
424  * It returns the error in case of allocation failure.
425  */
426 int ext4_map_blocks(handle_t *handle, struct inode *inode,
427                     struct ext4_map_blocks *map, int flags)
428 {
429         int retval;
430
431         map->m_flags = 0;
432         ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
433                   "logical block %lu\n", inode->i_ino, flags, map->m_len,
434                   (unsigned long) map->m_lblk);
435         /*
436          * Try to see if we can get the block without requesting a new
437          * file system block.
438          */
439         down_read((&EXT4_I(inode)->i_data_sem));
440         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
441                 retval = ext4_ext_map_blocks(handle, inode, map, 0);
442         } else {
443                 retval = ext4_ind_map_blocks(handle, inode, map, 0);
444         }
445         up_read((&EXT4_I(inode)->i_data_sem));
446
447         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
448                 int ret = check_block_validity(inode, map);
449                 if (ret != 0)
450                         return ret;
451         }
452
453         /* If it is only a block(s) look up */
454         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
455                 return retval;
456
457         /*
458          * Returns if the blocks have already allocated
459          *
460          * Note that if blocks have been preallocated
461          * ext4_ext_get_block() returns th create = 0
462          * with buffer head unmapped.
463          */
464         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
465                 return retval;
466
467         /*
468          * When we call get_blocks without the create flag, the
469          * BH_Unwritten flag could have gotten set if the blocks
470          * requested were part of a uninitialized extent.  We need to
471          * clear this flag now that we are committed to convert all or
472          * part of the uninitialized extent to be an initialized
473          * extent.  This is because we need to avoid the combination
474          * of BH_Unwritten and BH_Mapped flags being simultaneously
475          * set on the buffer_head.
476          */
477         map->m_flags &= ~EXT4_MAP_UNWRITTEN;
478
479         /*
480          * New blocks allocate and/or writing to uninitialized extent
481          * will possibly result in updating i_data, so we take
482          * the write lock of i_data_sem, and call get_blocks()
483          * with create == 1 flag.
484          */
485         down_write((&EXT4_I(inode)->i_data_sem));
486
487         /*
488          * if the caller is from delayed allocation writeout path
489          * we have already reserved fs blocks for allocation
490          * let the underlying get_block() function know to
491          * avoid double accounting
492          */
493         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
494                 ext4_set_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
495         /*
496          * We need to check for EXT4 here because migrate
497          * could have changed the inode type in between
498          */
499         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
500                 retval = ext4_ext_map_blocks(handle, inode, map, flags);
501         } else {
502                 retval = ext4_ind_map_blocks(handle, inode, map, flags);
503
504                 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
505                         /*
506                          * We allocated new blocks which will result in
507                          * i_data's format changing.  Force the migrate
508                          * to fail by clearing migrate flags
509                          */
510                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
511                 }
512
513                 /*
514                  * Update reserved blocks/metadata blocks after successful
515                  * block allocation which had been deferred till now. We don't
516                  * support fallocate for non extent files. So we can update
517                  * reserve space here.
518                  */
519                 if ((retval > 0) &&
520                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
521                         ext4_da_update_reserve_space(inode, retval, 1);
522         }
523         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
524                 ext4_clear_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
525
526         up_write((&EXT4_I(inode)->i_data_sem));
527         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
528                 int ret = check_block_validity(inode, map);
529                 if (ret != 0)
530                         return ret;
531         }
532         return retval;
533 }
534
535 /* Maximum number of blocks we map for direct IO at once. */
536 #define DIO_MAX_BLOCKS 4096
537
538 static int _ext4_get_block(struct inode *inode, sector_t iblock,
539                            struct buffer_head *bh, int flags)
540 {
541         handle_t *handle = ext4_journal_current_handle();
542         struct ext4_map_blocks map;
543         int ret = 0, started = 0;
544         int dio_credits;
545
546         map.m_lblk = iblock;
547         map.m_len = bh->b_size >> inode->i_blkbits;
548
549         if (flags && !handle) {
550                 /* Direct IO write... */
551                 if (map.m_len > DIO_MAX_BLOCKS)
552                         map.m_len = DIO_MAX_BLOCKS;
553                 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
554                 handle = ext4_journal_start(inode, dio_credits);
555                 if (IS_ERR(handle)) {
556                         ret = PTR_ERR(handle);
557                         return ret;
558                 }
559                 started = 1;
560         }
561
562         ret = ext4_map_blocks(handle, inode, &map, flags);
563         if (ret > 0) {
564                 map_bh(bh, inode->i_sb, map.m_pblk);
565                 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
566                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
567                 ret = 0;
568         }
569         if (started)
570                 ext4_journal_stop(handle);
571         return ret;
572 }
573
574 int ext4_get_block(struct inode *inode, sector_t iblock,
575                    struct buffer_head *bh, int create)
576 {
577         return _ext4_get_block(inode, iblock, bh,
578                                create ? EXT4_GET_BLOCKS_CREATE : 0);
579 }
580
581 /*
582  * `handle' can be NULL if create is zero
583  */
584 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
585                                 ext4_lblk_t block, int create, int *errp)
586 {
587         struct ext4_map_blocks map;
588         struct buffer_head *bh;
589         int fatal = 0, err;
590
591         J_ASSERT(handle != NULL || create == 0);
592
593         map.m_lblk = block;
594         map.m_len = 1;
595         err = ext4_map_blocks(handle, inode, &map,
596                               create ? EXT4_GET_BLOCKS_CREATE : 0);
597
598         if (err < 0)
599                 *errp = err;
600         if (err <= 0)
601                 return NULL;
602         *errp = 0;
603
604         bh = sb_getblk(inode->i_sb, map.m_pblk);
605         if (!bh) {
606                 *errp = -EIO;
607                 return NULL;
608         }
609         if (map.m_flags & EXT4_MAP_NEW) {
610                 J_ASSERT(create != 0);
611                 J_ASSERT(handle != NULL);
612
613                 /*
614                  * Now that we do not always journal data, we should
615                  * keep in mind whether this should always journal the
616                  * new buffer as metadata.  For now, regular file
617                  * writes use ext4_get_block instead, so it's not a
618                  * problem.
619                  */
620                 lock_buffer(bh);
621                 BUFFER_TRACE(bh, "call get_create_access");
622                 fatal = ext4_journal_get_create_access(handle, bh);
623                 if (!fatal && !buffer_uptodate(bh)) {
624                         memset(bh->b_data, 0, inode->i_sb->s_blocksize);
625                         set_buffer_uptodate(bh);
626                 }
627                 unlock_buffer(bh);
628                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
629                 err = ext4_handle_dirty_metadata(handle, inode, bh);
630                 if (!fatal)
631                         fatal = err;
632         } else {
633                 BUFFER_TRACE(bh, "not a new buffer");
634         }
635         if (fatal) {
636                 *errp = fatal;
637                 brelse(bh);
638                 bh = NULL;
639         }
640         return bh;
641 }
642
643 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
644                                ext4_lblk_t block, int create, int *err)
645 {
646         struct buffer_head *bh;
647
648         bh = ext4_getblk(handle, inode, block, create, err);
649         if (!bh)
650                 return bh;
651         if (buffer_uptodate(bh))
652                 return bh;
653         ll_rw_block(READ | REQ_META | REQ_PRIO, 1, &bh);
654         wait_on_buffer(bh);
655         if (buffer_uptodate(bh))
656                 return bh;
657         put_bh(bh);
658         *err = -EIO;
659         return NULL;
660 }
661
662 static int walk_page_buffers(handle_t *handle,
663                              struct buffer_head *head,
664                              unsigned from,
665                              unsigned to,
666                              int *partial,
667                              int (*fn)(handle_t *handle,
668                                        struct buffer_head *bh))
669 {
670         struct buffer_head *bh;
671         unsigned block_start, block_end;
672         unsigned blocksize = head->b_size;
673         int err, ret = 0;
674         struct buffer_head *next;
675
676         for (bh = head, block_start = 0;
677              ret == 0 && (bh != head || !block_start);
678              block_start = block_end, bh = next) {
679                 next = bh->b_this_page;
680                 block_end = block_start + blocksize;
681                 if (block_end <= from || block_start >= to) {
682                         if (partial && !buffer_uptodate(bh))
683                                 *partial = 1;
684                         continue;
685                 }
686                 err = (*fn)(handle, bh);
687                 if (!ret)
688                         ret = err;
689         }
690         return ret;
691 }
692
693 /*
694  * To preserve ordering, it is essential that the hole instantiation and
695  * the data write be encapsulated in a single transaction.  We cannot
696  * close off a transaction and start a new one between the ext4_get_block()
697  * and the commit_write().  So doing the jbd2_journal_start at the start of
698  * prepare_write() is the right place.
699  *
700  * Also, this function can nest inside ext4_writepage() ->
701  * block_write_full_page(). In that case, we *know* that ext4_writepage()
702  * has generated enough buffer credits to do the whole page.  So we won't
703  * block on the journal in that case, which is good, because the caller may
704  * be PF_MEMALLOC.
705  *
706  * By accident, ext4 can be reentered when a transaction is open via
707  * quota file writes.  If we were to commit the transaction while thus
708  * reentered, there can be a deadlock - we would be holding a quota
709  * lock, and the commit would never complete if another thread had a
710  * transaction open and was blocking on the quota lock - a ranking
711  * violation.
712  *
713  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
714  * will _not_ run commit under these circumstances because handle->h_ref
715  * is elevated.  We'll still have enough credits for the tiny quotafile
716  * write.
717  */
718 static int do_journal_get_write_access(handle_t *handle,
719                                        struct buffer_head *bh)
720 {
721         int dirty = buffer_dirty(bh);
722         int ret;
723
724         if (!buffer_mapped(bh) || buffer_freed(bh))
725                 return 0;
726         /*
727          * __block_write_begin() could have dirtied some buffers. Clean
728          * the dirty bit as jbd2_journal_get_write_access() could complain
729          * otherwise about fs integrity issues. Setting of the dirty bit
730          * by __block_write_begin() isn't a real problem here as we clear
731          * the bit before releasing a page lock and thus writeback cannot
732          * ever write the buffer.
733          */
734         if (dirty)
735                 clear_buffer_dirty(bh);
736         ret = ext4_journal_get_write_access(handle, bh);
737         if (!ret && dirty)
738                 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
739         return ret;
740 }
741
742 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
743                    struct buffer_head *bh_result, int create);
744 static int ext4_write_begin(struct file *file, struct address_space *mapping,
745                             loff_t pos, unsigned len, unsigned flags,
746                             struct page **pagep, void **fsdata)
747 {
748         struct inode *inode = mapping->host;
749         int ret, needed_blocks;
750         handle_t *handle;
751         int retries = 0;
752         struct page *page;
753         pgoff_t index;
754         unsigned from, to;
755
756         trace_ext4_write_begin(inode, pos, len, flags);
757         /*
758          * Reserve one block more for addition to orphan list in case
759          * we allocate blocks but write fails for some reason
760          */
761         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
762         index = pos >> PAGE_CACHE_SHIFT;
763         from = pos & (PAGE_CACHE_SIZE - 1);
764         to = from + len;
765
766 retry:
767         handle = ext4_journal_start(inode, needed_blocks);
768         if (IS_ERR(handle)) {
769                 ret = PTR_ERR(handle);
770                 goto out;
771         }
772
773         /* We cannot recurse into the filesystem as the transaction is already
774          * started */
775         flags |= AOP_FLAG_NOFS;
776
777         page = grab_cache_page_write_begin(mapping, index, flags);
778         if (!page) {
779                 ext4_journal_stop(handle);
780                 ret = -ENOMEM;
781                 goto out;
782         }
783         *pagep = page;
784
785         if (ext4_should_dioread_nolock(inode))
786                 ret = __block_write_begin(page, pos, len, ext4_get_block_write);
787         else
788                 ret = __block_write_begin(page, pos, len, ext4_get_block);
789
790         if (!ret && ext4_should_journal_data(inode)) {
791                 ret = walk_page_buffers(handle, page_buffers(page),
792                                 from, to, NULL, do_journal_get_write_access);
793         }
794
795         if (ret) {
796                 unlock_page(page);
797                 page_cache_release(page);
798                 /*
799                  * __block_write_begin may have instantiated a few blocks
800                  * outside i_size.  Trim these off again. Don't need
801                  * i_size_read because we hold i_mutex.
802                  *
803                  * Add inode to orphan list in case we crash before
804                  * truncate finishes
805                  */
806                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
807                         ext4_orphan_add(handle, inode);
808
809                 ext4_journal_stop(handle);
810                 if (pos + len > inode->i_size) {
811                         ext4_truncate_failed_write(inode);
812                         /*
813                          * If truncate failed early the inode might
814                          * still be on the orphan list; we need to
815                          * make sure the inode is removed from the
816                          * orphan list in that case.
817                          */
818                         if (inode->i_nlink)
819                                 ext4_orphan_del(NULL, inode);
820                 }
821         }
822
823         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
824                 goto retry;
825 out:
826         return ret;
827 }
828
829 /* For write_end() in data=journal mode */
830 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
831 {
832         if (!buffer_mapped(bh) || buffer_freed(bh))
833                 return 0;
834         set_buffer_uptodate(bh);
835         return ext4_handle_dirty_metadata(handle, NULL, bh);
836 }
837
838 static int ext4_generic_write_end(struct file *file,
839                                   struct address_space *mapping,
840                                   loff_t pos, unsigned len, unsigned copied,
841                                   struct page *page, void *fsdata)
842 {
843         int i_size_changed = 0;
844         struct inode *inode = mapping->host;
845         handle_t *handle = ext4_journal_current_handle();
846
847         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
848
849         /*
850          * No need to use i_size_read() here, the i_size
851          * cannot change under us because we hold i_mutex.
852          *
853          * But it's important to update i_size while still holding page lock:
854          * page writeout could otherwise come in and zero beyond i_size.
855          */
856         if (pos + copied > inode->i_size) {
857                 i_size_write(inode, pos + copied);
858                 i_size_changed = 1;
859         }
860
861         if (pos + copied >  EXT4_I(inode)->i_disksize) {
862                 /* We need to mark inode dirty even if
863                  * new_i_size is less that inode->i_size
864                  * bu greater than i_disksize.(hint delalloc)
865                  */
866                 ext4_update_i_disksize(inode, (pos + copied));
867                 i_size_changed = 1;
868         }
869         unlock_page(page);
870         page_cache_release(page);
871
872         /*
873          * Don't mark the inode dirty under page lock. First, it unnecessarily
874          * makes the holding time of page lock longer. Second, it forces lock
875          * ordering of page lock and transaction start for journaling
876          * filesystems.
877          */
878         if (i_size_changed)
879                 ext4_mark_inode_dirty(handle, inode);
880
881         return copied;
882 }
883
884 /*
885  * We need to pick up the new inode size which generic_commit_write gave us
886  * `file' can be NULL - eg, when called from page_symlink().
887  *
888  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
889  * buffers are managed internally.
890  */
891 static int ext4_ordered_write_end(struct file *file,
892                                   struct address_space *mapping,
893                                   loff_t pos, unsigned len, unsigned copied,
894                                   struct page *page, void *fsdata)
895 {
896         handle_t *handle = ext4_journal_current_handle();
897         struct inode *inode = mapping->host;
898         int ret = 0, ret2;
899
900         trace_ext4_ordered_write_end(inode, pos, len, copied);
901         ret = ext4_jbd2_file_inode(handle, inode);
902
903         if (ret == 0) {
904                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
905                                                         page, fsdata);
906                 copied = ret2;
907                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
908                         /* if we have allocated more blocks and copied
909                          * less. We will have blocks allocated outside
910                          * inode->i_size. So truncate them
911                          */
912                         ext4_orphan_add(handle, inode);
913                 if (ret2 < 0)
914                         ret = ret2;
915         }
916         ret2 = ext4_journal_stop(handle);
917         if (!ret)
918                 ret = ret2;
919
920         if (pos + len > inode->i_size) {
921                 ext4_truncate_failed_write(inode);
922                 /*
923                  * If truncate failed early the inode might still be
924                  * on the orphan list; we need to make sure the inode
925                  * is removed from the orphan list in that case.
926                  */
927                 if (inode->i_nlink)
928                         ext4_orphan_del(NULL, inode);
929         }
930
931
932         return ret ? ret : copied;
933 }
934
935 static int ext4_writeback_write_end(struct file *file,
936                                     struct address_space *mapping,
937                                     loff_t pos, unsigned len, unsigned copied,
938                                     struct page *page, void *fsdata)
939 {
940         handle_t *handle = ext4_journal_current_handle();
941         struct inode *inode = mapping->host;
942         int ret = 0, ret2;
943
944         trace_ext4_writeback_write_end(inode, pos, len, copied);
945         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
946                                                         page, fsdata);
947         copied = ret2;
948         if (pos + len > inode->i_size && ext4_can_truncate(inode))
949                 /* if we have allocated more blocks and copied
950                  * less. We will have blocks allocated outside
951                  * inode->i_size. So truncate them
952                  */
953                 ext4_orphan_add(handle, inode);
954
955         if (ret2 < 0)
956                 ret = ret2;
957
958         ret2 = ext4_journal_stop(handle);
959         if (!ret)
960                 ret = ret2;
961
962         if (pos + len > inode->i_size) {
963                 ext4_truncate_failed_write(inode);
964                 /*
965                  * If truncate failed early the inode might still be
966                  * on the orphan list; we need to make sure the inode
967                  * is removed from the orphan list in that case.
968                  */
969                 if (inode->i_nlink)
970                         ext4_orphan_del(NULL, inode);
971         }
972
973         return ret ? ret : copied;
974 }
975
976 static int ext4_journalled_write_end(struct file *file,
977                                      struct address_space *mapping,
978                                      loff_t pos, unsigned len, unsigned copied,
979                                      struct page *page, void *fsdata)
980 {
981         handle_t *handle = ext4_journal_current_handle();
982         struct inode *inode = mapping->host;
983         int ret = 0, ret2;
984         int partial = 0;
985         unsigned from, to;
986         loff_t new_i_size;
987
988         trace_ext4_journalled_write_end(inode, pos, len, copied);
989         from = pos & (PAGE_CACHE_SIZE - 1);
990         to = from + len;
991
992         BUG_ON(!ext4_handle_valid(handle));
993
994         if (copied < len) {
995                 if (!PageUptodate(page))
996                         copied = 0;
997                 page_zero_new_buffers(page, from+copied, to);
998         }
999
1000         ret = walk_page_buffers(handle, page_buffers(page), from,
1001                                 to, &partial, write_end_fn);
1002         if (!partial)
1003                 SetPageUptodate(page);
1004         new_i_size = pos + copied;
1005         if (new_i_size > inode->i_size)
1006                 i_size_write(inode, pos+copied);
1007         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1008         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1009         if (new_i_size > EXT4_I(inode)->i_disksize) {
1010                 ext4_update_i_disksize(inode, new_i_size);
1011                 ret2 = ext4_mark_inode_dirty(handle, inode);
1012                 if (!ret)
1013                         ret = ret2;
1014         }
1015
1016         unlock_page(page);
1017         page_cache_release(page);
1018         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1019                 /* if we have allocated more blocks and copied
1020                  * less. We will have blocks allocated outside
1021                  * inode->i_size. So truncate them
1022                  */
1023                 ext4_orphan_add(handle, inode);
1024
1025         ret2 = ext4_journal_stop(handle);
1026         if (!ret)
1027                 ret = ret2;
1028         if (pos + len > inode->i_size) {
1029                 ext4_truncate_failed_write(inode);
1030                 /*
1031                  * If truncate failed early the inode might still be
1032                  * on the orphan list; we need to make sure the inode
1033                  * is removed from the orphan list in that case.
1034                  */
1035                 if (inode->i_nlink)
1036                         ext4_orphan_del(NULL, inode);
1037         }
1038
1039         return ret ? ret : copied;
1040 }
1041
1042 /*
1043  * Reserve a single block located at lblock
1044  */
1045 static int ext4_da_reserve_space(struct inode *inode, ext4_lblk_t lblock)
1046 {
1047         int retries = 0;
1048         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1049         struct ext4_inode_info *ei = EXT4_I(inode);
1050         unsigned long md_needed;
1051         int ret;
1052
1053         /*
1054          * recalculate the amount of metadata blocks to reserve
1055          * in order to allocate nrblocks
1056          * worse case is one extent per block
1057          */
1058 repeat:
1059         spin_lock(&ei->i_block_reservation_lock);
1060         md_needed = ext4_calc_metadata_amount(inode, lblock);
1061         trace_ext4_da_reserve_space(inode, md_needed);
1062         spin_unlock(&ei->i_block_reservation_lock);
1063
1064         /*
1065          * We will charge metadata quota at writeout time; this saves
1066          * us from metadata over-estimation, though we may go over by
1067          * a small amount in the end.  Here we just reserve for data.
1068          */
1069         ret = dquot_reserve_block(inode, 1);
1070         if (ret)
1071                 return ret;
1072         /*
1073          * We do still charge estimated metadata to the sb though;
1074          * we cannot afford to run out of free blocks.
1075          */
1076         if (ext4_claim_free_blocks(sbi, md_needed + 1, 0)) {
1077                 dquot_release_reservation_block(inode, 1);
1078                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1079                         yield();
1080                         goto repeat;
1081                 }
1082                 return -ENOSPC;
1083         }
1084         spin_lock(&ei->i_block_reservation_lock);
1085         ei->i_reserved_data_blocks++;
1086         ei->i_reserved_meta_blocks += md_needed;
1087         spin_unlock(&ei->i_block_reservation_lock);
1088
1089         return 0;       /* success */
1090 }
1091
1092 static void ext4_da_release_space(struct inode *inode, int to_free)
1093 {
1094         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1095         struct ext4_inode_info *ei = EXT4_I(inode);
1096
1097         if (!to_free)
1098                 return;         /* Nothing to release, exit */
1099
1100         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1101
1102         trace_ext4_da_release_space(inode, to_free);
1103         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1104                 /*
1105                  * if there aren't enough reserved blocks, then the
1106                  * counter is messed up somewhere.  Since this
1107                  * function is called from invalidate page, it's
1108                  * harmless to return without any action.
1109                  */
1110                 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1111                          "ino %lu, to_free %d with only %d reserved "
1112                          "data blocks\n", inode->i_ino, to_free,
1113                          ei->i_reserved_data_blocks);
1114                 WARN_ON(1);
1115                 to_free = ei->i_reserved_data_blocks;
1116         }
1117         ei->i_reserved_data_blocks -= to_free;
1118
1119         if (ei->i_reserved_data_blocks == 0) {
1120                 /*
1121                  * We can release all of the reserved metadata blocks
1122                  * only when we have written all of the delayed
1123                  * allocation blocks.
1124                  */
1125                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1126                                    ei->i_reserved_meta_blocks);
1127                 ei->i_reserved_meta_blocks = 0;
1128                 ei->i_da_metadata_calc_len = 0;
1129         }
1130
1131         /* update fs dirty data blocks counter */
1132         percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
1133
1134         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1135
1136         dquot_release_reservation_block(inode, to_free);
1137 }
1138
1139 static void ext4_da_page_release_reservation(struct page *page,
1140                                              unsigned long offset)
1141 {
1142         int to_release = 0;
1143         struct buffer_head *head, *bh;
1144         unsigned int curr_off = 0;
1145
1146         head = page_buffers(page);
1147         bh = head;
1148         do {
1149                 unsigned int next_off = curr_off + bh->b_size;
1150
1151                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1152                         to_release++;
1153                         clear_buffer_delay(bh);
1154                 }
1155                 curr_off = next_off;
1156         } while ((bh = bh->b_this_page) != head);
1157         ext4_da_release_space(page->mapping->host, to_release);
1158 }
1159
1160 /*
1161  * Delayed allocation stuff
1162  */
1163
1164 /*
1165  * mpage_da_submit_io - walks through extent of pages and try to write
1166  * them with writepage() call back
1167  *
1168  * @mpd->inode: inode
1169  * @mpd->first_page: first page of the extent
1170  * @mpd->next_page: page after the last page of the extent
1171  *
1172  * By the time mpage_da_submit_io() is called we expect all blocks
1173  * to be allocated. this may be wrong if allocation failed.
1174  *
1175  * As pages are already locked by write_cache_pages(), we can't use it
1176  */
1177 static int mpage_da_submit_io(struct mpage_da_data *mpd,
1178                               struct ext4_map_blocks *map)
1179 {
1180         struct pagevec pvec;
1181         unsigned long index, end;
1182         int ret = 0, err, nr_pages, i;
1183         struct inode *inode = mpd->inode;
1184         struct address_space *mapping = inode->i_mapping;
1185         loff_t size = i_size_read(inode);
1186         unsigned int len, block_start;
1187         struct buffer_head *bh, *page_bufs = NULL;
1188         int journal_data = ext4_should_journal_data(inode);
1189         sector_t pblock = 0, cur_logical = 0;
1190         struct ext4_io_submit io_submit;
1191
1192         BUG_ON(mpd->next_page <= mpd->first_page);
1193         memset(&io_submit, 0, sizeof(io_submit));
1194         /*
1195          * We need to start from the first_page to the next_page - 1
1196          * to make sure we also write the mapped dirty buffer_heads.
1197          * If we look at mpd->b_blocknr we would only be looking
1198          * at the currently mapped buffer_heads.
1199          */
1200         index = mpd->first_page;
1201         end = mpd->next_page - 1;
1202
1203         pagevec_init(&pvec, 0);
1204         while (index <= end) {
1205                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1206                 if (nr_pages == 0)
1207                         break;
1208                 for (i = 0; i < nr_pages; i++) {
1209                         int commit_write = 0, skip_page = 0;
1210                         struct page *page = pvec.pages[i];
1211
1212                         index = page->index;
1213                         if (index > end)
1214                                 break;
1215
1216                         if (index == size >> PAGE_CACHE_SHIFT)
1217                                 len = size & ~PAGE_CACHE_MASK;
1218                         else
1219                                 len = PAGE_CACHE_SIZE;
1220                         if (map) {
1221                                 cur_logical = index << (PAGE_CACHE_SHIFT -
1222                                                         inode->i_blkbits);
1223                                 pblock = map->m_pblk + (cur_logical -
1224                                                         map->m_lblk);
1225                         }
1226                         index++;
1227
1228                         BUG_ON(!PageLocked(page));
1229                         BUG_ON(PageWriteback(page));
1230
1231                         /*
1232                          * If the page does not have buffers (for
1233                          * whatever reason), try to create them using
1234                          * __block_write_begin.  If this fails,
1235                          * skip the page and move on.
1236                          */
1237                         if (!page_has_buffers(page)) {
1238                                 if (__block_write_begin(page, 0, len,
1239                                                 noalloc_get_block_write)) {
1240                                 skip_page:
1241                                         unlock_page(page);
1242                                         continue;
1243                                 }
1244                                 commit_write = 1;
1245                         }
1246
1247                         bh = page_bufs = page_buffers(page);
1248                         block_start = 0;
1249                         do {
1250                                 if (!bh)
1251                                         goto skip_page;
1252                                 if (map && (cur_logical >= map->m_lblk) &&
1253                                     (cur_logical <= (map->m_lblk +
1254                                                      (map->m_len - 1)))) {
1255                                         if (buffer_delay(bh)) {
1256                                                 clear_buffer_delay(bh);
1257                                                 bh->b_blocknr = pblock;
1258                                         }
1259                                         if (buffer_unwritten(bh) ||
1260                                             buffer_mapped(bh))
1261                                                 BUG_ON(bh->b_blocknr != pblock);
1262                                         if (map->m_flags & EXT4_MAP_UNINIT)
1263                                                 set_buffer_uninit(bh);
1264                                         clear_buffer_unwritten(bh);
1265                                 }
1266
1267                                 /* skip page if block allocation undone */
1268                                 if (buffer_delay(bh) || buffer_unwritten(bh))
1269                                         skip_page = 1;
1270                                 bh = bh->b_this_page;
1271                                 block_start += bh->b_size;
1272                                 cur_logical++;
1273                                 pblock++;
1274                         } while (bh != page_bufs);
1275
1276                         if (skip_page)
1277                                 goto skip_page;
1278
1279                         if (commit_write)
1280                                 /* mark the buffer_heads as dirty & uptodate */
1281                                 block_commit_write(page, 0, len);
1282
1283                         clear_page_dirty_for_io(page);
1284                         /*
1285                          * Delalloc doesn't support data journalling,
1286                          * but eventually maybe we'll lift this
1287                          * restriction.
1288                          */
1289                         if (unlikely(journal_data && PageChecked(page)))
1290                                 err = __ext4_journalled_writepage(page, len);
1291                         else if (test_opt(inode->i_sb, MBLK_IO_SUBMIT))
1292                                 err = ext4_bio_write_page(&io_submit, page,
1293                                                           len, mpd->wbc);
1294                         else if (buffer_uninit(page_bufs)) {
1295                                 ext4_set_bh_endio(page_bufs, inode);
1296                                 err = block_write_full_page_endio(page,
1297                                         noalloc_get_block_write,
1298                                         mpd->wbc, ext4_end_io_buffer_write);
1299                         } else
1300                                 err = block_write_full_page(page,
1301                                         noalloc_get_block_write, mpd->wbc);
1302
1303                         if (!err)
1304                                 mpd->pages_written++;
1305                         /*
1306                          * In error case, we have to continue because
1307                          * remaining pages are still locked
1308                          */
1309                         if (ret == 0)
1310                                 ret = err;
1311                 }
1312                 pagevec_release(&pvec);
1313         }
1314         ext4_io_submit(&io_submit);
1315         return ret;
1316 }
1317
1318 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd)
1319 {
1320         int nr_pages, i;
1321         pgoff_t index, end;
1322         struct pagevec pvec;
1323         struct inode *inode = mpd->inode;
1324         struct address_space *mapping = inode->i_mapping;
1325
1326         index = mpd->first_page;
1327         end   = mpd->next_page - 1;
1328         while (index <= end) {
1329                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1330                 if (nr_pages == 0)
1331                         break;
1332                 for (i = 0; i < nr_pages; i++) {
1333                         struct page *page = pvec.pages[i];
1334                         if (page->index > end)
1335                                 break;
1336                         BUG_ON(!PageLocked(page));
1337                         BUG_ON(PageWriteback(page));
1338                         block_invalidatepage(page, 0);
1339                         ClearPageUptodate(page);
1340                         unlock_page(page);
1341                 }
1342                 index = pvec.pages[nr_pages - 1]->index + 1;
1343                 pagevec_release(&pvec);
1344         }
1345         return;
1346 }
1347
1348 static void ext4_print_free_blocks(struct inode *inode)
1349 {
1350         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1351         printk(KERN_CRIT "Total free blocks count %lld\n",
1352                ext4_count_free_blocks(inode->i_sb));
1353         printk(KERN_CRIT "Free/Dirty block details\n");
1354         printk(KERN_CRIT "free_blocks=%lld\n",
1355                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
1356         printk(KERN_CRIT "dirty_blocks=%lld\n",
1357                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
1358         printk(KERN_CRIT "Block reservation details\n");
1359         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
1360                EXT4_I(inode)->i_reserved_data_blocks);
1361         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
1362                EXT4_I(inode)->i_reserved_meta_blocks);
1363         return;
1364 }
1365
1366 /*
1367  * mpage_da_map_and_submit - go through given space, map them
1368  *       if necessary, and then submit them for I/O
1369  *
1370  * @mpd - bh describing space
1371  *
1372  * The function skips space we know is already mapped to disk blocks.
1373  *
1374  */
1375 static void mpage_da_map_and_submit(struct mpage_da_data *mpd)
1376 {
1377         int err, blks, get_blocks_flags;
1378         struct ext4_map_blocks map, *mapp = NULL;
1379         sector_t next = mpd->b_blocknr;
1380         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
1381         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
1382         handle_t *handle = NULL;
1383
1384         /*
1385          * If the blocks are mapped already, or we couldn't accumulate
1386          * any blocks, then proceed immediately to the submission stage.
1387          */
1388         if ((mpd->b_size == 0) ||
1389             ((mpd->b_state  & (1 << BH_Mapped)) &&
1390              !(mpd->b_state & (1 << BH_Delay)) &&
1391              !(mpd->b_state & (1 << BH_Unwritten))))
1392                 goto submit_io;
1393
1394         handle = ext4_journal_current_handle();
1395         BUG_ON(!handle);
1396
1397         /*
1398          * Call ext4_map_blocks() to allocate any delayed allocation
1399          * blocks, or to convert an uninitialized extent to be
1400          * initialized (in the case where we have written into
1401          * one or more preallocated blocks).
1402          *
1403          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
1404          * indicate that we are on the delayed allocation path.  This
1405          * affects functions in many different parts of the allocation
1406          * call path.  This flag exists primarily because we don't
1407          * want to change *many* call functions, so ext4_map_blocks()
1408          * will set the EXT4_STATE_DELALLOC_RESERVED flag once the
1409          * inode's allocation semaphore is taken.
1410          *
1411          * If the blocks in questions were delalloc blocks, set
1412          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
1413          * variables are updated after the blocks have been allocated.
1414          */
1415         map.m_lblk = next;
1416         map.m_len = max_blocks;
1417         get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
1418         if (ext4_should_dioread_nolock(mpd->inode))
1419                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
1420         if (mpd->b_state & (1 << BH_Delay))
1421                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
1422
1423         blks = ext4_map_blocks(handle, mpd->inode, &map, get_blocks_flags);
1424         if (blks < 0) {
1425                 struct super_block *sb = mpd->inode->i_sb;
1426
1427                 err = blks;
1428                 /*
1429                  * If get block returns EAGAIN or ENOSPC and there
1430                  * appears to be free blocks we will just let
1431                  * mpage_da_submit_io() unlock all of the pages.
1432                  */
1433                 if (err == -EAGAIN)
1434                         goto submit_io;
1435
1436                 if (err == -ENOSPC &&
1437                     ext4_count_free_blocks(sb)) {
1438                         mpd->retval = err;
1439                         goto submit_io;
1440                 }
1441
1442                 /*
1443                  * get block failure will cause us to loop in
1444                  * writepages, because a_ops->writepage won't be able
1445                  * to make progress. The page will be redirtied by
1446                  * writepage and writepages will again try to write
1447                  * the same.
1448                  */
1449                 if (!(EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) {
1450                         ext4_msg(sb, KERN_CRIT,
1451                                  "delayed block allocation failed for inode %lu "
1452                                  "at logical offset %llu with max blocks %zd "
1453                                  "with error %d", mpd->inode->i_ino,
1454                                  (unsigned long long) next,
1455                                  mpd->b_size >> mpd->inode->i_blkbits, err);
1456                         ext4_msg(sb, KERN_CRIT,
1457                                 "This should not happen!! Data will be lost\n");
1458                         if (err == -ENOSPC)
1459                                 ext4_print_free_blocks(mpd->inode);
1460                 }
1461                 /* invalidate all the pages */
1462                 ext4_da_block_invalidatepages(mpd);
1463
1464                 /* Mark this page range as having been completed */
1465                 mpd->io_done = 1;
1466                 return;
1467         }
1468         BUG_ON(blks == 0);
1469
1470         mapp = &map;
1471         if (map.m_flags & EXT4_MAP_NEW) {
1472                 struct block_device *bdev = mpd->inode->i_sb->s_bdev;
1473                 int i;
1474
1475                 for (i = 0; i < map.m_len; i++)
1476                         unmap_underlying_metadata(bdev, map.m_pblk + i);
1477         }
1478
1479         if (ext4_should_order_data(mpd->inode)) {
1480                 err = ext4_jbd2_file_inode(handle, mpd->inode);
1481                 if (err)
1482                         /* This only happens if the journal is aborted */
1483                         return;
1484         }
1485
1486         /*
1487          * Update on-disk size along with block allocation.
1488          */
1489         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
1490         if (disksize > i_size_read(mpd->inode))
1491                 disksize = i_size_read(mpd->inode);
1492         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
1493                 ext4_update_i_disksize(mpd->inode, disksize);
1494                 err = ext4_mark_inode_dirty(handle, mpd->inode);
1495                 if (err)
1496                         ext4_error(mpd->inode->i_sb,
1497                                    "Failed to mark inode %lu dirty",
1498                                    mpd->inode->i_ino);
1499         }
1500
1501 submit_io:
1502         mpage_da_submit_io(mpd, mapp);
1503         mpd->io_done = 1;
1504 }
1505
1506 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
1507                 (1 << BH_Delay) | (1 << BH_Unwritten))
1508
1509 /*
1510  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
1511  *
1512  * @mpd->lbh - extent of blocks
1513  * @logical - logical number of the block in the file
1514  * @bh - bh of the block (used to access block's state)
1515  *
1516  * the function is used to collect contig. blocks in same state
1517  */
1518 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
1519                                    sector_t logical, size_t b_size,
1520                                    unsigned long b_state)
1521 {
1522         sector_t next;
1523         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
1524
1525         /*
1526          * XXX Don't go larger than mballoc is willing to allocate
1527          * This is a stopgap solution.  We eventually need to fold
1528          * mpage_da_submit_io() into this function and then call
1529          * ext4_map_blocks() multiple times in a loop
1530          */
1531         if (nrblocks >= 8*1024*1024/mpd->inode->i_sb->s_blocksize)
1532                 goto flush_it;
1533
1534         /* check if thereserved journal credits might overflow */
1535         if (!(ext4_test_inode_flag(mpd->inode, EXT4_INODE_EXTENTS))) {
1536                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
1537                         /*
1538                          * With non-extent format we are limited by the journal
1539                          * credit available.  Total credit needed to insert
1540                          * nrblocks contiguous blocks is dependent on the
1541                          * nrblocks.  So limit nrblocks.
1542                          */
1543                         goto flush_it;
1544                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
1545                                 EXT4_MAX_TRANS_DATA) {
1546                         /*
1547                          * Adding the new buffer_head would make it cross the
1548                          * allowed limit for which we have journal credit
1549                          * reserved. So limit the new bh->b_size
1550                          */
1551                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
1552                                                 mpd->inode->i_blkbits;
1553                         /* we will do mpage_da_submit_io in the next loop */
1554                 }
1555         }
1556         /*
1557          * First block in the extent
1558          */
1559         if (mpd->b_size == 0) {
1560                 mpd->b_blocknr = logical;
1561                 mpd->b_size = b_size;
1562                 mpd->b_state = b_state & BH_FLAGS;
1563                 return;
1564         }
1565
1566         next = mpd->b_blocknr + nrblocks;
1567         /*
1568          * Can we merge the block to our big extent?
1569          */
1570         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
1571                 mpd->b_size += b_size;
1572                 return;
1573         }
1574
1575 flush_it:
1576         /*
1577          * We couldn't merge the block to our extent, so we
1578          * need to flush current  extent and start new one
1579          */
1580         mpage_da_map_and_submit(mpd);
1581         return;
1582 }
1583
1584 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1585 {
1586         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1587 }
1588
1589 /*
1590  * This is a special get_blocks_t callback which is used by
1591  * ext4_da_write_begin().  It will either return mapped block or
1592  * reserve space for a single block.
1593  *
1594  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1595  * We also have b_blocknr = -1 and b_bdev initialized properly
1596  *
1597  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1598  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1599  * initialized properly.
1600  */
1601 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1602                                   struct buffer_head *bh, int create)
1603 {
1604         struct ext4_map_blocks map;
1605         int ret = 0;
1606         sector_t invalid_block = ~((sector_t) 0xffff);
1607
1608         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1609                 invalid_block = ~0;
1610
1611         BUG_ON(create == 0);
1612         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1613
1614         map.m_lblk = iblock;
1615         map.m_len = 1;
1616
1617         /*
1618          * first, we need to know whether the block is allocated already
1619          * preallocated blocks are unmapped but should treated
1620          * the same as allocated blocks.
1621          */
1622         ret = ext4_map_blocks(NULL, inode, &map, 0);
1623         if (ret < 0)
1624                 return ret;
1625         if (ret == 0) {
1626                 if (buffer_delay(bh))
1627                         return 0; /* Not sure this could or should happen */
1628                 /*
1629                  * XXX: __block_write_begin() unmaps passed block, is it OK?
1630                  */
1631                 ret = ext4_da_reserve_space(inode, iblock);
1632                 if (ret)
1633                         /* not enough space to reserve */
1634                         return ret;
1635
1636                 map_bh(bh, inode->i_sb, invalid_block);
1637                 set_buffer_new(bh);
1638                 set_buffer_delay(bh);
1639                 return 0;
1640         }
1641
1642         map_bh(bh, inode->i_sb, map.m_pblk);
1643         bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
1644
1645         if (buffer_unwritten(bh)) {
1646                 /* A delayed write to unwritten bh should be marked
1647                  * new and mapped.  Mapped ensures that we don't do
1648                  * get_block multiple times when we write to the same
1649                  * offset and new ensures that we do proper zero out
1650                  * for partial write.
1651                  */
1652                 set_buffer_new(bh);
1653                 set_buffer_mapped(bh);
1654         }
1655         return 0;
1656 }
1657
1658 /*
1659  * This function is used as a standard get_block_t calback function
1660  * when there is no desire to allocate any blocks.  It is used as a
1661  * callback function for block_write_begin() and block_write_full_page().
1662  * These functions should only try to map a single block at a time.
1663  *
1664  * Since this function doesn't do block allocations even if the caller
1665  * requests it by passing in create=1, it is critically important that
1666  * any caller checks to make sure that any buffer heads are returned
1667  * by this function are either all already mapped or marked for
1668  * delayed allocation before calling  block_write_full_page().  Otherwise,
1669  * b_blocknr could be left unitialized, and the page write functions will
1670  * be taken by surprise.
1671  */
1672 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
1673                                    struct buffer_head *bh_result, int create)
1674 {
1675         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
1676         return _ext4_get_block(inode, iblock, bh_result, 0);
1677 }
1678
1679 static int bget_one(handle_t *handle, struct buffer_head *bh)
1680 {
1681         get_bh(bh);
1682         return 0;
1683 }
1684
1685 static int bput_one(handle_t *handle, struct buffer_head *bh)
1686 {
1687         put_bh(bh);
1688         return 0;
1689 }
1690
1691 static int __ext4_journalled_writepage(struct page *page,
1692                                        unsigned int len)
1693 {
1694         struct address_space *mapping = page->mapping;
1695         struct inode *inode = mapping->host;
1696         struct buffer_head *page_bufs;
1697         handle_t *handle = NULL;
1698         int ret = 0;
1699         int err;
1700
1701         ClearPageChecked(page);
1702         page_bufs = page_buffers(page);
1703         BUG_ON(!page_bufs);
1704         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
1705         /* As soon as we unlock the page, it can go away, but we have
1706          * references to buffers so we are safe */
1707         unlock_page(page);
1708
1709         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
1710         if (IS_ERR(handle)) {
1711                 ret = PTR_ERR(handle);
1712                 goto out;
1713         }
1714
1715         BUG_ON(!ext4_handle_valid(handle));
1716
1717         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
1718                                 do_journal_get_write_access);
1719
1720         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
1721                                 write_end_fn);
1722         if (ret == 0)
1723                 ret = err;
1724         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1725         err = ext4_journal_stop(handle);
1726         if (!ret)
1727                 ret = err;
1728
1729         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
1730         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1731 out:
1732         return ret;
1733 }
1734
1735 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
1736 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
1737
1738 /*
1739  * Note that we don't need to start a transaction unless we're journaling data
1740  * because we should have holes filled from ext4_page_mkwrite(). We even don't
1741  * need to file the inode to the transaction's list in ordered mode because if
1742  * we are writing back data added by write(), the inode is already there and if
1743  * we are writing back data modified via mmap(), no one guarantees in which
1744  * transaction the data will hit the disk. In case we are journaling data, we
1745  * cannot start transaction directly because transaction start ranks above page
1746  * lock so we have to do some magic.
1747  *
1748  * This function can get called via...
1749  *   - ext4_da_writepages after taking page lock (have journal handle)
1750  *   - journal_submit_inode_data_buffers (no journal handle)
1751  *   - shrink_page_list via pdflush (no journal handle)
1752  *   - grab_page_cache when doing write_begin (have journal handle)
1753  *
1754  * We don't do any block allocation in this function. If we have page with
1755  * multiple blocks we need to write those buffer_heads that are mapped. This
1756  * is important for mmaped based write. So if we do with blocksize 1K
1757  * truncate(f, 1024);
1758  * a = mmap(f, 0, 4096);
1759  * a[0] = 'a';
1760  * truncate(f, 4096);
1761  * we have in the page first buffer_head mapped via page_mkwrite call back
1762  * but other bufer_heads would be unmapped but dirty(dirty done via the
1763  * do_wp_page). So writepage should write the first block. If we modify
1764  * the mmap area beyond 1024 we will again get a page_fault and the
1765  * page_mkwrite callback will do the block allocation and mark the
1766  * buffer_heads mapped.
1767  *
1768  * We redirty the page if we have any buffer_heads that is either delay or
1769  * unwritten in the page.
1770  *
1771  * We can get recursively called as show below.
1772  *
1773  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1774  *              ext4_writepage()
1775  *
1776  * But since we don't do any block allocation we should not deadlock.
1777  * Page also have the dirty flag cleared so we don't get recurive page_lock.
1778  */
1779 static int ext4_writepage(struct page *page,
1780                           struct writeback_control *wbc)
1781 {
1782         int ret = 0, commit_write = 0;
1783         loff_t size;
1784         unsigned int len;
1785         struct buffer_head *page_bufs = NULL;
1786         struct inode *inode = page->mapping->host;
1787
1788         trace_ext4_writepage(page);
1789         size = i_size_read(inode);
1790         if (page->index == size >> PAGE_CACHE_SHIFT)
1791                 len = size & ~PAGE_CACHE_MASK;
1792         else
1793                 len = PAGE_CACHE_SIZE;
1794
1795         /*
1796          * If the page does not have buffers (for whatever reason),
1797          * try to create them using __block_write_begin.  If this
1798          * fails, redirty the page and move on.
1799          */
1800         if (!page_has_buffers(page)) {
1801                 if (__block_write_begin(page, 0, len,
1802                                         noalloc_get_block_write)) {
1803                 redirty_page:
1804                         redirty_page_for_writepage(wbc, page);
1805                         unlock_page(page);
1806                         return 0;
1807                 }
1808                 commit_write = 1;
1809         }
1810         page_bufs = page_buffers(page);
1811         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
1812                               ext4_bh_delay_or_unwritten)) {
1813                 /*
1814                  * We don't want to do block allocation, so redirty
1815                  * the page and return.  We may reach here when we do
1816                  * a journal commit via journal_submit_inode_data_buffers.
1817                  * We can also reach here via shrink_page_list
1818                  */
1819                 goto redirty_page;
1820         }
1821         if (commit_write)
1822                 /* now mark the buffer_heads as dirty and uptodate */
1823                 block_commit_write(page, 0, len);
1824
1825         if (PageChecked(page) && ext4_should_journal_data(inode))
1826                 /*
1827                  * It's mmapped pagecache.  Add buffers and journal it.  There
1828                  * doesn't seem much point in redirtying the page here.
1829                  */
1830                 return __ext4_journalled_writepage(page, len);
1831
1832         if (buffer_uninit(page_bufs)) {
1833                 ext4_set_bh_endio(page_bufs, inode);
1834                 ret = block_write_full_page_endio(page, noalloc_get_block_write,
1835                                             wbc, ext4_end_io_buffer_write);
1836         } else
1837                 ret = block_write_full_page(page, noalloc_get_block_write,
1838                                             wbc);
1839
1840         return ret;
1841 }
1842
1843 /*
1844  * This is called via ext4_da_writepages() to
1845  * calculate the total number of credits to reserve to fit
1846  * a single extent allocation into a single transaction,
1847  * ext4_da_writpeages() will loop calling this before
1848  * the block allocation.
1849  */
1850
1851 static int ext4_da_writepages_trans_blocks(struct inode *inode)
1852 {
1853         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
1854
1855         /*
1856          * With non-extent format the journal credit needed to
1857          * insert nrblocks contiguous block is dependent on
1858          * number of contiguous block. So we will limit
1859          * number of contiguous block to a sane value
1860          */
1861         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) &&
1862             (max_blocks > EXT4_MAX_TRANS_DATA))
1863                 max_blocks = EXT4_MAX_TRANS_DATA;
1864
1865         return ext4_chunk_trans_blocks(inode, max_blocks);
1866 }
1867
1868 /*
1869  * write_cache_pages_da - walk the list of dirty pages of the given
1870  * address space and accumulate pages that need writing, and call
1871  * mpage_da_map_and_submit to map a single contiguous memory region
1872  * and then write them.
1873  */
1874 static int write_cache_pages_da(struct address_space *mapping,
1875                                 struct writeback_control *wbc,
1876                                 struct mpage_da_data *mpd,
1877                                 pgoff_t *done_index)
1878 {
1879         struct buffer_head      *bh, *head;
1880         struct inode            *inode = mapping->host;
1881         struct pagevec          pvec;
1882         unsigned int            nr_pages;
1883         sector_t                logical;
1884         pgoff_t                 index, end;
1885         long                    nr_to_write = wbc->nr_to_write;
1886         int                     i, tag, ret = 0;
1887
1888         memset(mpd, 0, sizeof(struct mpage_da_data));
1889         mpd->wbc = wbc;
1890         mpd->inode = inode;
1891         pagevec_init(&pvec, 0);
1892         index = wbc->range_start >> PAGE_CACHE_SHIFT;
1893         end = wbc->range_end >> PAGE_CACHE_SHIFT;
1894
1895         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1896                 tag = PAGECACHE_TAG_TOWRITE;
1897         else
1898                 tag = PAGECACHE_TAG_DIRTY;
1899
1900         *done_index = index;
1901         while (index <= end) {
1902                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1903                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1904                 if (nr_pages == 0)
1905                         return 0;
1906
1907                 for (i = 0; i < nr_pages; i++) {
1908                         struct page *page = pvec.pages[i];
1909
1910                         /*
1911                          * At this point, the page may be truncated or
1912                          * invalidated (changing page->mapping to NULL), or
1913                          * even swizzled back from swapper_space to tmpfs file
1914                          * mapping. However, page->index will not change
1915                          * because we have a reference on the page.
1916                          */
1917                         if (page->index > end)
1918                                 goto out;
1919
1920                         *done_index = page->index + 1;
1921
1922                         /*
1923                          * If we can't merge this page, and we have
1924                          * accumulated an contiguous region, write it
1925                          */
1926                         if ((mpd->next_page != page->index) &&
1927                             (mpd->next_page != mpd->first_page)) {
1928                                 mpage_da_map_and_submit(mpd);
1929                                 goto ret_extent_tail;
1930                         }
1931
1932                         lock_page(page);
1933
1934                         /*
1935                          * If the page is no longer dirty, or its
1936                          * mapping no longer corresponds to inode we
1937                          * are writing (which means it has been
1938                          * truncated or invalidated), or the page is
1939                          * already under writeback and we are not
1940                          * doing a data integrity writeback, skip the page
1941                          */
1942                         if (!PageDirty(page) ||
1943                             (PageWriteback(page) &&
1944                              (wbc->sync_mode == WB_SYNC_NONE)) ||
1945                             unlikely(page->mapping != mapping)) {
1946                                 unlock_page(page);
1947                                 continue;
1948                         }
1949
1950                         wait_on_page_writeback(page);
1951                         BUG_ON(PageWriteback(page));
1952
1953                         if (mpd->next_page != page->index)
1954                                 mpd->first_page = page->index;
1955                         mpd->next_page = page->index + 1;
1956                         logical = (sector_t) page->index <<
1957                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
1958
1959                         if (!page_has_buffers(page)) {
1960                                 mpage_add_bh_to_extent(mpd, logical,
1961                                                        PAGE_CACHE_SIZE,
1962                                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
1963                                 if (mpd->io_done)
1964                                         goto ret_extent_tail;
1965                         } else {
1966                                 /*
1967                                  * Page with regular buffer heads,
1968                                  * just add all dirty ones
1969                                  */
1970                                 head = page_buffers(page);
1971                                 bh = head;
1972                                 do {
1973                                         BUG_ON(buffer_locked(bh));
1974                                         /*
1975                                          * We need to try to allocate
1976                                          * unmapped blocks in the same page.
1977                                          * Otherwise we won't make progress
1978                                          * with the page in ext4_writepage
1979                                          */
1980                                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
1981                                                 mpage_add_bh_to_extent(mpd, logical,
1982                                                                        bh->b_size,
1983                                                                        bh->b_state);
1984                                                 if (mpd->io_done)
1985                                                         goto ret_extent_tail;
1986                                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
1987                                                 /*
1988                                                  * mapped dirty buffer. We need
1989                                                  * to update the b_state
1990                                                  * because we look at b_state
1991                                                  * in mpage_da_map_blocks.  We
1992                                                  * don't update b_size because
1993                                                  * if we find an unmapped
1994                                                  * buffer_head later we need to
1995                                                  * use the b_state flag of that
1996                                                  * buffer_head.
1997                                                  */
1998                                                 if (mpd->b_size == 0)
1999                                                         mpd->b_state = bh->b_state & BH_FLAGS;
2000                                         }
2001                                         logical++;
2002                                 } while ((bh = bh->b_this_page) != head);
2003                         }
2004
2005                         if (nr_to_write > 0) {
2006                                 nr_to_write--;
2007                                 if (nr_to_write == 0 &&
2008                                     wbc->sync_mode == WB_SYNC_NONE)
2009                                         /*
2010                                          * We stop writing back only if we are
2011                                          * not doing integrity sync. In case of
2012                                          * integrity sync we have to keep going
2013                                          * because someone may be concurrently
2014                                          * dirtying pages, and we might have
2015                                          * synced a lot of newly appeared dirty
2016                                          * pages, but have not synced all of the
2017                                          * old dirty pages.
2018                                          */
2019                                         goto out;
2020                         }
2021                 }
2022                 pagevec_release(&pvec);
2023                 cond_resched();
2024         }
2025         return 0;
2026 ret_extent_tail:
2027         ret = MPAGE_DA_EXTENT_TAIL;
2028 out:
2029         pagevec_release(&pvec);
2030         cond_resched();
2031         return ret;
2032 }
2033
2034
2035 static int ext4_da_writepages(struct address_space *mapping,
2036                               struct writeback_control *wbc)
2037 {
2038         pgoff_t index;
2039         int range_whole = 0;
2040         handle_t *handle = NULL;
2041         struct mpage_da_data mpd;
2042         struct inode *inode = mapping->host;
2043         int pages_written = 0;
2044         unsigned int max_pages;
2045         int range_cyclic, cycled = 1, io_done = 0;
2046         int needed_blocks, ret = 0;
2047         long desired_nr_to_write, nr_to_writebump = 0;
2048         loff_t range_start = wbc->range_start;
2049         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2050         pgoff_t done_index = 0;
2051         pgoff_t end;
2052
2053         trace_ext4_da_writepages(inode, wbc);
2054
2055         /*
2056          * No pages to write? This is mainly a kludge to avoid starting
2057          * a transaction for special inodes like journal inode on last iput()
2058          * because that could violate lock ordering on umount
2059          */
2060         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2061                 return 0;
2062
2063         /*
2064          * If the filesystem has aborted, it is read-only, so return
2065          * right away instead of dumping stack traces later on that
2066          * will obscure the real source of the problem.  We test
2067          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2068          * the latter could be true if the filesystem is mounted
2069          * read-only, and in that case, ext4_da_writepages should
2070          * *never* be called, so if that ever happens, we would want
2071          * the stack trace.
2072          */
2073         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2074                 return -EROFS;
2075
2076         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2077                 range_whole = 1;
2078
2079         range_cyclic = wbc->range_cyclic;
2080         if (wbc->range_cyclic) {
2081                 index = mapping->writeback_index;
2082                 if (index)
2083                         cycled = 0;
2084                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2085                 wbc->range_end  = LLONG_MAX;
2086                 wbc->range_cyclic = 0;
2087                 end = -1;
2088         } else {
2089                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2090                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2091         }
2092
2093         /*
2094          * This works around two forms of stupidity.  The first is in
2095          * the writeback code, which caps the maximum number of pages
2096          * written to be 1024 pages.  This is wrong on multiple
2097          * levels; different architectues have a different page size,
2098          * which changes the maximum amount of data which gets
2099          * written.  Secondly, 4 megabytes is way too small.  XFS
2100          * forces this value to be 16 megabytes by multiplying
2101          * nr_to_write parameter by four, and then relies on its
2102          * allocator to allocate larger extents to make them
2103          * contiguous.  Unfortunately this brings us to the second
2104          * stupidity, which is that ext4's mballoc code only allocates
2105          * at most 2048 blocks.  So we force contiguous writes up to
2106          * the number of dirty blocks in the inode, or
2107          * sbi->max_writeback_mb_bump whichever is smaller.
2108          */
2109         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2110         if (!range_cyclic && range_whole) {
2111                 if (wbc->nr_to_write == LONG_MAX)
2112                         desired_nr_to_write = wbc->nr_to_write;
2113                 else
2114                         desired_nr_to_write = wbc->nr_to_write * 8;
2115         } else
2116                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2117                                                            max_pages);
2118         if (desired_nr_to_write > max_pages)
2119                 desired_nr_to_write = max_pages;
2120
2121         if (wbc->nr_to_write < desired_nr_to_write) {
2122                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2123                 wbc->nr_to_write = desired_nr_to_write;
2124         }
2125
2126 retry:
2127         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2128                 tag_pages_for_writeback(mapping, index, end);
2129
2130         while (!ret && wbc->nr_to_write > 0) {
2131
2132                 /*
2133                  * we  insert one extent at a time. So we need
2134                  * credit needed for single extent allocation.
2135                  * journalled mode is currently not supported
2136                  * by delalloc
2137                  */
2138                 BUG_ON(ext4_should_journal_data(inode));
2139                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2140
2141                 /* start a new transaction*/
2142                 handle = ext4_journal_start(inode, needed_blocks);
2143                 if (IS_ERR(handle)) {
2144                         ret = PTR_ERR(handle);
2145                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2146                                "%ld pages, ino %lu; err %d", __func__,
2147                                 wbc->nr_to_write, inode->i_ino, ret);
2148                         goto out_writepages;
2149                 }
2150
2151                 /*
2152                  * Now call write_cache_pages_da() to find the next
2153                  * contiguous region of logical blocks that need
2154                  * blocks to be allocated by ext4 and submit them.
2155                  */
2156                 ret = write_cache_pages_da(mapping, wbc, &mpd, &done_index);
2157                 /*
2158                  * If we have a contiguous extent of pages and we
2159                  * haven't done the I/O yet, map the blocks and submit
2160                  * them for I/O.
2161                  */
2162                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2163                         mpage_da_map_and_submit(&mpd);
2164                         ret = MPAGE_DA_EXTENT_TAIL;
2165                 }
2166                 trace_ext4_da_write_pages(inode, &mpd);
2167                 wbc->nr_to_write -= mpd.pages_written;
2168
2169                 ext4_journal_stop(handle);
2170
2171                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2172                         /* commit the transaction which would
2173                          * free blocks released in the transaction
2174                          * and try again
2175                          */
2176                         jbd2_journal_force_commit_nested(sbi->s_journal);
2177                         ret = 0;
2178                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2179                         /*
2180                          * got one extent now try with
2181                          * rest of the pages
2182                          */
2183                         pages_written += mpd.pages_written;
2184                         ret = 0;
2185                         io_done = 1;
2186                 } else if (wbc->nr_to_write)
2187                         /*
2188                          * There is no more writeout needed
2189                          * or we requested for a noblocking writeout
2190                          * and we found the device congested
2191                          */
2192                         break;
2193         }
2194         if (!io_done && !cycled) {
2195                 cycled = 1;
2196                 index = 0;
2197                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2198                 wbc->range_end  = mapping->writeback_index - 1;
2199                 goto retry;
2200         }
2201
2202         /* Update index */
2203         wbc->range_cyclic = range_cyclic;
2204         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2205                 /*
2206                  * set the writeback_index so that range_cyclic
2207                  * mode will write it back later
2208                  */
2209                 mapping->writeback_index = done_index;
2210
2211 out_writepages:
2212         wbc->nr_to_write -= nr_to_writebump;
2213         wbc->range_start = range_start;
2214         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
2215         return ret;
2216 }
2217
2218 #define FALL_BACK_TO_NONDELALLOC 1
2219 static int ext4_nonda_switch(struct super_block *sb)
2220 {
2221         s64 free_blocks, dirty_blocks;
2222         struct ext4_sb_info *sbi = EXT4_SB(sb);
2223
2224         /*
2225          * switch to non delalloc mode if we are running low
2226          * on free block. The free block accounting via percpu
2227          * counters can get slightly wrong with percpu_counter_batch getting
2228          * accumulated on each CPU without updating global counters
2229          * Delalloc need an accurate free block accounting. So switch
2230          * to non delalloc when we are near to error range.
2231          */
2232         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
2233         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
2234         if (2 * free_blocks < 3 * dirty_blocks ||
2235                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
2236                 /*
2237                  * free block count is less than 150% of dirty blocks
2238                  * or free blocks is less than watermark
2239                  */
2240                 return 1;
2241         }
2242         /*
2243          * Even if we don't switch but are nearing capacity,
2244          * start pushing delalloc when 1/2 of free blocks are dirty.
2245          */
2246         if (free_blocks < 2 * dirty_blocks)
2247                 writeback_inodes_sb_if_idle(sb);
2248
2249         return 0;
2250 }
2251
2252 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2253                                loff_t pos, unsigned len, unsigned flags,
2254                                struct page **pagep, void **fsdata)
2255 {
2256         int ret, retries = 0;
2257         struct page *page;
2258         pgoff_t index;
2259         struct inode *inode = mapping->host;
2260         handle_t *handle;
2261
2262         index = pos >> PAGE_CACHE_SHIFT;
2263
2264         if (ext4_nonda_switch(inode->i_sb)) {
2265                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2266                 return ext4_write_begin(file, mapping, pos,
2267                                         len, flags, pagep, fsdata);
2268         }
2269         *fsdata = (void *)0;
2270         trace_ext4_da_write_begin(inode, pos, len, flags);
2271 retry:
2272         /*
2273          * With delayed allocation, we don't log the i_disksize update
2274          * if there is delayed block allocation. But we still need
2275          * to journalling the i_disksize update if writes to the end
2276          * of file which has an already mapped buffer.
2277          */
2278         handle = ext4_journal_start(inode, 1);
2279         if (IS_ERR(handle)) {
2280                 ret = PTR_ERR(handle);
2281                 goto out;
2282         }
2283         /* We cannot recurse into the filesystem as the transaction is already
2284          * started */
2285         flags |= AOP_FLAG_NOFS;
2286
2287         page = grab_cache_page_write_begin(mapping, index, flags);
2288         if (!page) {
2289                 ext4_journal_stop(handle);
2290                 ret = -ENOMEM;
2291                 goto out;
2292         }
2293         *pagep = page;
2294
2295         ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
2296         if (ret < 0) {
2297                 unlock_page(page);
2298                 ext4_journal_stop(handle);
2299                 page_cache_release(page);
2300                 /*
2301                  * block_write_begin may have instantiated a few blocks
2302                  * outside i_size.  Trim these off again. Don't need
2303                  * i_size_read because we hold i_mutex.
2304                  */
2305                 if (pos + len > inode->i_size)
2306                         ext4_truncate_failed_write(inode);
2307         }
2308
2309         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
2310                 goto retry;
2311 out:
2312         return ret;
2313 }
2314
2315 /*
2316  * Check if we should update i_disksize
2317  * when write to the end of file but not require block allocation
2318  */
2319 static int ext4_da_should_update_i_disksize(struct page *page,
2320                                             unsigned long offset)
2321 {
2322         struct buffer_head *bh;
2323         struct inode *inode = page->mapping->host;
2324         unsigned int idx;
2325         int i;
2326
2327         bh = page_buffers(page);
2328         idx = offset >> inode->i_blkbits;
2329
2330         for (i = 0; i < idx; i++)
2331                 bh = bh->b_this_page;
2332
2333         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
2334                 return 0;
2335         return 1;
2336 }
2337
2338 static int ext4_da_write_end(struct file *file,
2339                              struct address_space *mapping,
2340                              loff_t pos, unsigned len, unsigned copied,
2341                              struct page *page, void *fsdata)
2342 {
2343         struct inode *inode = mapping->host;
2344         int ret = 0, ret2;
2345         handle_t *handle = ext4_journal_current_handle();
2346         loff_t new_i_size;
2347         unsigned long start, end;
2348         int write_mode = (int)(unsigned long)fsdata;
2349
2350         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
2351                 if (ext4_should_order_data(inode)) {
2352                         return ext4_ordered_write_end(file, mapping, pos,
2353                                         len, copied, page, fsdata);
2354                 } else if (ext4_should_writeback_data(inode)) {
2355                         return ext4_writeback_write_end(file, mapping, pos,
2356                                         len, copied, page, fsdata);
2357                 } else {
2358                         BUG();
2359                 }
2360         }
2361
2362         trace_ext4_da_write_end(inode, pos, len, copied);
2363         start = pos & (PAGE_CACHE_SIZE - 1);
2364         end = start + copied - 1;
2365
2366         /*
2367          * generic_write_end() will run mark_inode_dirty() if i_size
2368          * changes.  So let's piggyback the i_disksize mark_inode_dirty
2369          * into that.
2370          */
2371
2372         new_i_size = pos + copied;
2373         if (new_i_size > EXT4_I(inode)->i_disksize) {
2374                 if (ext4_da_should_update_i_disksize(page, end)) {
2375                         down_write(&EXT4_I(inode)->i_data_sem);
2376                         if (new_i_size > EXT4_I(inode)->i_disksize) {
2377                                 /*
2378                                  * Updating i_disksize when extending file
2379                                  * without needing block allocation
2380                                  */
2381                                 if (ext4_should_order_data(inode))
2382                                         ret = ext4_jbd2_file_inode(handle,
2383                                                                    inode);
2384
2385                                 EXT4_I(inode)->i_disksize = new_i_size;
2386                         }
2387                         up_write(&EXT4_I(inode)->i_data_sem);
2388                         /* We need to mark inode dirty even if
2389                          * new_i_size is less that inode->i_size
2390                          * bu greater than i_disksize.(hint delalloc)
2391                          */
2392                         ext4_mark_inode_dirty(handle, inode);
2393                 }
2394         }
2395         ret2 = generic_write_end(file, mapping, pos, len, copied,
2396                                                         page, fsdata);
2397         copied = ret2;
2398         if (ret2 < 0)
2399                 ret = ret2;
2400         ret2 = ext4_journal_stop(handle);
2401         if (!ret)
2402                 ret = ret2;
2403
2404         return ret ? ret : copied;
2405 }
2406
2407 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
2408 {
2409         /*
2410          * Drop reserved blocks
2411          */
2412         BUG_ON(!PageLocked(page));
2413         if (!page_has_buffers(page))
2414                 goto out;
2415
2416         ext4_da_page_release_reservation(page, offset);
2417
2418 out:
2419         ext4_invalidatepage(page, offset);
2420
2421         return;
2422 }
2423
2424 /*
2425  * Force all delayed allocation blocks to be allocated for a given inode.
2426  */
2427 int ext4_alloc_da_blocks(struct inode *inode)
2428 {
2429         trace_ext4_alloc_da_blocks(inode);
2430
2431         if (!EXT4_I(inode)->i_reserved_data_blocks &&
2432             !EXT4_I(inode)->i_reserved_meta_blocks)
2433                 return 0;
2434
2435         /*
2436          * We do something simple for now.  The filemap_flush() will
2437          * also start triggering a write of the data blocks, which is
2438          * not strictly speaking necessary (and for users of
2439          * laptop_mode, not even desirable).  However, to do otherwise
2440          * would require replicating code paths in:
2441          *
2442          * ext4_da_writepages() ->
2443          *    write_cache_pages() ---> (via passed in callback function)
2444          *        __mpage_da_writepage() -->
2445          *           mpage_add_bh_to_extent()
2446          *           mpage_da_map_blocks()
2447          *
2448          * The problem is that write_cache_pages(), located in
2449          * mm/page-writeback.c, marks pages clean in preparation for
2450          * doing I/O, which is not desirable if we're not planning on
2451          * doing I/O at all.
2452          *
2453          * We could call write_cache_pages(), and then redirty all of
2454          * the pages by calling redirty_page_for_writepage() but that
2455          * would be ugly in the extreme.  So instead we would need to
2456          * replicate parts of the code in the above functions,
2457          * simplifying them because we wouldn't actually intend to
2458          * write out the pages, but rather only collect contiguous
2459          * logical block extents, call the multi-block allocator, and
2460          * then update the buffer heads with the block allocations.
2461          *
2462          * For now, though, we'll cheat by calling filemap_flush(),
2463          * which will map the blocks, and start the I/O, but not
2464          * actually wait for the I/O to complete.
2465          */
2466         return filemap_flush(inode->i_mapping);
2467 }
2468
2469 /*
2470  * bmap() is special.  It gets used by applications such as lilo and by
2471  * the swapper to find the on-disk block of a specific piece of data.
2472  *
2473  * Naturally, this is dangerous if the block concerned is still in the
2474  * journal.  If somebody makes a swapfile on an ext4 data-journaling
2475  * filesystem and enables swap, then they may get a nasty shock when the
2476  * data getting swapped to that swapfile suddenly gets overwritten by
2477  * the original zero's written out previously to the journal and
2478  * awaiting writeback in the kernel's buffer cache.
2479  *
2480  * So, if we see any bmap calls here on a modified, data-journaled file,
2481  * take extra steps to flush any blocks which might be in the cache.
2482  */
2483 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
2484 {
2485         struct inode *inode = mapping->host;
2486         journal_t *journal;
2487         int err;
2488
2489         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
2490                         test_opt(inode->i_sb, DELALLOC)) {
2491                 /*
2492                  * With delalloc we want to sync the file
2493                  * so that we can make sure we allocate
2494                  * blocks for file
2495                  */
2496                 filemap_write_and_wait(mapping);
2497         }
2498
2499         if (EXT4_JOURNAL(inode) &&
2500             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
2501                 /*
2502                  * This is a REALLY heavyweight approach, but the use of
2503                  * bmap on dirty files is expected to be extremely rare:
2504                  * only if we run lilo or swapon on a freshly made file
2505                  * do we expect this to happen.
2506                  *
2507                  * (bmap requires CAP_SYS_RAWIO so this does not
2508                  * represent an unprivileged user DOS attack --- we'd be
2509                  * in trouble if mortal users could trigger this path at
2510                  * will.)
2511                  *
2512                  * NB. EXT4_STATE_JDATA is not set on files other than
2513                  * regular files.  If somebody wants to bmap a directory
2514                  * or symlink and gets confused because the buffer
2515                  * hasn't yet been flushed to disk, they deserve
2516                  * everything they get.
2517                  */
2518
2519                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
2520                 journal = EXT4_JOURNAL(inode);
2521                 jbd2_journal_lock_updates(journal);
2522                 err = jbd2_journal_flush(journal);
2523                 jbd2_journal_unlock_updates(journal);
2524
2525                 if (err)
2526                         return 0;
2527         }
2528
2529         return generic_block_bmap(mapping, block, ext4_get_block);
2530 }
2531
2532 static int ext4_readpage(struct file *file, struct page *page)
2533 {
2534         trace_ext4_readpage(page);
2535         return mpage_readpage(page, ext4_get_block);
2536 }
2537
2538 static int
2539 ext4_readpages(struct file *file, struct address_space *mapping,
2540                 struct list_head *pages, unsigned nr_pages)
2541 {
2542         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
2543 }
2544
2545 static void ext4_invalidatepage_free_endio(struct page *page, unsigned long offset)
2546 {
2547         struct buffer_head *head, *bh;
2548         unsigned int curr_off = 0;
2549
2550         if (!page_has_buffers(page))
2551                 return;
2552         head = bh = page_buffers(page);
2553         do {
2554                 if (offset <= curr_off && test_clear_buffer_uninit(bh)
2555                                         && bh->b_private) {
2556                         ext4_free_io_end(bh->b_private);
2557                         bh->b_private = NULL;
2558                         bh->b_end_io = NULL;
2559                 }
2560                 curr_off = curr_off + bh->b_size;
2561                 bh = bh->b_this_page;
2562         } while (bh != head);
2563 }
2564
2565 static void ext4_invalidatepage(struct page *page, unsigned long offset)
2566 {
2567         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2568
2569         trace_ext4_invalidatepage(page, offset);
2570
2571         /*
2572          * free any io_end structure allocated for buffers to be discarded
2573          */
2574         if (ext4_should_dioread_nolock(page->mapping->host))
2575                 ext4_invalidatepage_free_endio(page, offset);
2576         /*
2577          * If it's a full truncate we just forget about the pending dirtying
2578          */
2579         if (offset == 0)
2580                 ClearPageChecked(page);
2581
2582         if (journal)
2583                 jbd2_journal_invalidatepage(journal, page, offset);
2584         else
2585                 block_invalidatepage(page, offset);
2586 }
2587
2588 static int ext4_releasepage(struct page *page, gfp_t wait)
2589 {
2590         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2591
2592         trace_ext4_releasepage(page);
2593
2594         WARN_ON(PageChecked(page));
2595         if (!page_has_buffers(page))
2596                 return 0;
2597         if (journal)
2598                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
2599         else
2600                 return try_to_free_buffers(page);
2601 }
2602
2603 /*
2604  * ext4_get_block used when preparing for a DIO write or buffer write.
2605  * We allocate an uinitialized extent if blocks haven't been allocated.
2606  * The extent will be converted to initialized after the IO is complete.
2607  */
2608 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
2609                    struct buffer_head *bh_result, int create)
2610 {
2611         ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
2612                    inode->i_ino, create);
2613         return _ext4_get_block(inode, iblock, bh_result,
2614                                EXT4_GET_BLOCKS_IO_CREATE_EXT);
2615 }
2616
2617 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
2618                             ssize_t size, void *private, int ret,
2619                             bool is_async)
2620 {
2621         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
2622         ext4_io_end_t *io_end = iocb->private;
2623         struct workqueue_struct *wq;
2624         unsigned long flags;
2625         struct ext4_inode_info *ei;
2626
2627         /* if not async direct IO or dio with 0 bytes write, just return */
2628         if (!io_end || !size)
2629                 goto out;
2630
2631         ext_debug("ext4_end_io_dio(): io_end 0x%p"
2632                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
2633                   iocb->private, io_end->inode->i_ino, iocb, offset,
2634                   size);
2635
2636         /* if not aio dio with unwritten extents, just free io and return */
2637         if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
2638                 ext4_free_io_end(io_end);
2639                 iocb->private = NULL;
2640 out:
2641                 if (is_async)
2642                         aio_complete(iocb, ret, 0);
2643                 inode_dio_done(inode);
2644                 return;
2645         }
2646
2647         io_end->offset = offset;
2648         io_end->size = size;
2649         if (is_async) {
2650                 io_end->iocb = iocb;
2651                 io_end->result = ret;
2652         }
2653         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
2654
2655         /* Add the io_end to per-inode completed aio dio list*/
2656         ei = EXT4_I(io_end->inode);
2657         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
2658         list_add_tail(&io_end->list, &ei->i_completed_io_list);
2659         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
2660
2661         /* queue the work to convert unwritten extents to written */
2662         queue_work(wq, &io_end->work);
2663         iocb->private = NULL;
2664
2665         /* XXX: probably should move into the real I/O completion handler */
2666         inode_dio_done(inode);
2667 }
2668
2669 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
2670 {
2671         ext4_io_end_t *io_end = bh->b_private;
2672         struct workqueue_struct *wq;
2673         struct inode *inode;
2674         unsigned long flags;
2675
2676         if (!test_clear_buffer_uninit(bh) || !io_end)
2677                 goto out;
2678
2679         if (!(io_end->inode->i_sb->s_flags & MS_ACTIVE)) {
2680                 printk("sb umounted, discard end_io request for inode %lu\n",
2681                         io_end->inode->i_ino);
2682                 ext4_free_io_end(io_end);
2683                 goto out;
2684         }
2685
2686         /*
2687          * It may be over-defensive here to check EXT4_IO_END_UNWRITTEN now,
2688          * but being more careful is always safe for the future change.
2689          */
2690         inode = io_end->inode;
2691         if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
2692                 io_end->flag |= EXT4_IO_END_UNWRITTEN;
2693                 atomic_inc(&EXT4_I(inode)->i_aiodio_unwritten);
2694         }
2695
2696         /* Add the io_end to per-inode completed io list*/
2697         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
2698         list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
2699         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
2700
2701         wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
2702         /* queue the work to convert unwritten extents to written */
2703         queue_work(wq, &io_end->work);
2704 out:
2705         bh->b_private = NULL;
2706         bh->b_end_io = NULL;
2707         clear_buffer_uninit(bh);
2708         end_buffer_async_write(bh, uptodate);
2709 }
2710
2711 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode)
2712 {
2713         ext4_io_end_t *io_end;
2714         struct page *page = bh->b_page;
2715         loff_t offset = (sector_t)page->index << PAGE_CACHE_SHIFT;
2716         size_t size = bh->b_size;
2717
2718 retry:
2719         io_end = ext4_init_io_end(inode, GFP_ATOMIC);
2720         if (!io_end) {
2721                 pr_warn_ratelimited("%s: allocation fail\n", __func__);
2722                 schedule();
2723                 goto retry;
2724         }
2725         io_end->offset = offset;
2726         io_end->size = size;
2727         /*
2728          * We need to hold a reference to the page to make sure it
2729          * doesn't get evicted before ext4_end_io_work() has a chance
2730          * to convert the extent from written to unwritten.
2731          */
2732         io_end->page = page;
2733         get_page(io_end->page);
2734
2735         bh->b_private = io_end;
2736         bh->b_end_io = ext4_end_io_buffer_write;
2737         return 0;
2738 }
2739
2740 /*
2741  * For ext4 extent files, ext4 will do direct-io write to holes,
2742  * preallocated extents, and those write extend the file, no need to
2743  * fall back to buffered IO.
2744  *
2745  * For holes, we fallocate those blocks, mark them as uninitialized
2746  * If those blocks were preallocated, we mark sure they are splited, but
2747  * still keep the range to write as uninitialized.
2748  *
2749  * The unwrritten extents will be converted to written when DIO is completed.
2750  * For async direct IO, since the IO may still pending when return, we
2751  * set up an end_io call back function, which will do the conversion
2752  * when async direct IO completed.
2753  *
2754  * If the O_DIRECT write will extend the file then add this inode to the
2755  * orphan list.  So recovery will truncate it back to the original size
2756  * if the machine crashes during the write.
2757  *
2758  */
2759 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
2760                               const struct iovec *iov, loff_t offset,
2761                               unsigned long nr_segs)
2762 {
2763         struct file *file = iocb->ki_filp;
2764         struct inode *inode = file->f_mapping->host;
2765         ssize_t ret;
2766         size_t count = iov_length(iov, nr_segs);
2767
2768         loff_t final_size = offset + count;
2769         if (rw == WRITE && final_size <= inode->i_size) {
2770                 /*
2771                  * We could direct write to holes and fallocate.
2772                  *
2773                  * Allocated blocks to fill the hole are marked as uninitialized
2774                  * to prevent parallel buffered read to expose the stale data
2775                  * before DIO complete the data IO.
2776                  *
2777                  * As to previously fallocated extents, ext4 get_block
2778                  * will just simply mark the buffer mapped but still
2779                  * keep the extents uninitialized.
2780                  *
2781                  * for non AIO case, we will convert those unwritten extents
2782                  * to written after return back from blockdev_direct_IO.
2783                  *
2784                  * for async DIO, the conversion needs to be defered when
2785                  * the IO is completed. The ext4 end_io callback function
2786                  * will be called to take care of the conversion work.
2787                  * Here for async case, we allocate an io_end structure to
2788                  * hook to the iocb.
2789                  */
2790                 iocb->private = NULL;
2791                 EXT4_I(inode)->cur_aio_dio = NULL;
2792                 if (!is_sync_kiocb(iocb)) {
2793                         iocb->private = ext4_init_io_end(inode, GFP_NOFS);
2794                         if (!iocb->private)
2795                                 return -ENOMEM;
2796                         /*
2797                          * we save the io structure for current async
2798                          * direct IO, so that later ext4_map_blocks()
2799                          * could flag the io structure whether there
2800                          * is a unwritten extents needs to be converted
2801                          * when IO is completed.
2802                          */
2803                         EXT4_I(inode)->cur_aio_dio = iocb->private;
2804                 }
2805
2806                 ret = __blockdev_direct_IO(rw, iocb, inode,
2807                                          inode->i_sb->s_bdev, iov,
2808                                          offset, nr_segs,
2809                                          ext4_get_block_write,
2810                                          ext4_end_io_dio,
2811                                          NULL,
2812                                          DIO_LOCKING | DIO_SKIP_HOLES);
2813                 if (iocb->private)
2814                         EXT4_I(inode)->cur_aio_dio = NULL;
2815                 /*
2816                  * The io_end structure takes a reference to the inode,
2817                  * that structure needs to be destroyed and the
2818                  * reference to the inode need to be dropped, when IO is
2819                  * complete, even with 0 byte write, or failed.
2820                  *
2821                  * In the successful AIO DIO case, the io_end structure will be
2822                  * desctroyed and the reference to the inode will be dropped
2823                  * after the end_io call back function is called.
2824                  *
2825                  * In the case there is 0 byte write, or error case, since
2826                  * VFS direct IO won't invoke the end_io call back function,
2827                  * we need to free the end_io structure here.
2828                  */
2829                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
2830                         ext4_free_io_end(iocb->private);
2831                         iocb->private = NULL;
2832                 } else if (ret > 0 && ext4_test_inode_state(inode,
2833                                                 EXT4_STATE_DIO_UNWRITTEN)) {
2834                         int err;
2835                         /*
2836                          * for non AIO case, since the IO is already
2837                          * completed, we could do the conversion right here
2838                          */
2839                         err = ext4_convert_unwritten_extents(inode,
2840                                                              offset, ret);
2841                         if (err < 0)
2842                                 ret = err;
2843                         ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
2844                 }
2845                 return ret;
2846         }
2847
2848         /* for write the the end of file case, we fall back to old way */
2849         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
2850 }
2851
2852 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
2853                               const struct iovec *iov, loff_t offset,
2854                               unsigned long nr_segs)
2855 {
2856         struct file *file = iocb->ki_filp;
2857         struct inode *inode = file->f_mapping->host;
2858         ssize_t ret;
2859
2860         trace_ext4_direct_IO_enter(inode, offset, iov_length(iov, nr_segs), rw);
2861         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
2862                 ret = ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
2863         else
2864                 ret = ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
2865         trace_ext4_direct_IO_exit(inode, offset,
2866                                 iov_length(iov, nr_segs), rw, ret);
2867         return ret;
2868 }
2869
2870 /*
2871  * Pages can be marked dirty completely asynchronously from ext4's journalling
2872  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
2873  * much here because ->set_page_dirty is called under VFS locks.  The page is
2874  * not necessarily locked.
2875  *
2876  * We cannot just dirty the page and leave attached buffers clean, because the
2877  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
2878  * or jbddirty because all the journalling code will explode.
2879  *
2880  * So what we do is to mark the page "pending dirty" and next time writepage
2881  * is called, propagate that into the buffers appropriately.
2882  */
2883 static int ext4_journalled_set_page_dirty(struct page *page)
2884 {
2885         SetPageChecked(page);
2886         return __set_page_dirty_nobuffers(page);
2887 }
2888
2889 static const struct address_space_operations ext4_ordered_aops = {
2890         .readpage               = ext4_readpage,
2891         .readpages              = ext4_readpages,
2892         .writepage              = ext4_writepage,
2893         .write_begin            = ext4_write_begin,
2894         .write_end              = ext4_ordered_write_end,
2895         .bmap                   = ext4_bmap,
2896         .invalidatepage         = ext4_invalidatepage,
2897         .releasepage            = ext4_releasepage,
2898         .direct_IO              = ext4_direct_IO,
2899         .migratepage            = buffer_migrate_page,
2900         .is_partially_uptodate  = block_is_partially_uptodate,
2901         .error_remove_page      = generic_error_remove_page,
2902 };
2903
2904 static const struct address_space_operations ext4_writeback_aops = {
2905         .readpage               = ext4_readpage,
2906         .readpages              = ext4_readpages,
2907         .writepage              = ext4_writepage,
2908         .write_begin            = ext4_write_begin,
2909         .write_end              = ext4_writeback_write_end,
2910         .bmap                   = ext4_bmap,
2911         .invalidatepage         = ext4_invalidatepage,
2912         .releasepage            = ext4_releasepage,
2913         .direct_IO              = ext4_direct_IO,
2914         .migratepage            = buffer_migrate_page,
2915         .is_partially_uptodate  = block_is_partially_uptodate,
2916         .error_remove_page      = generic_error_remove_page,
2917 };
2918
2919 static const struct address_space_operations ext4_journalled_aops = {
2920         .readpage               = ext4_readpage,
2921         .readpages              = ext4_readpages,
2922         .writepage              = ext4_writepage,
2923         .write_begin            = ext4_write_begin,
2924         .write_end              = ext4_journalled_write_end,
2925         .set_page_dirty         = ext4_journalled_set_page_dirty,
2926         .bmap                   = ext4_bmap,
2927         .invalidatepage         = ext4_invalidatepage,
2928         .releasepage            = ext4_releasepage,
2929         .is_partially_uptodate  = block_is_partially_uptodate,
2930         .error_remove_page      = generic_error_remove_page,
2931 };
2932
2933 static const struct address_space_operations ext4_da_aops = {
2934         .readpage               = ext4_readpage,
2935         .readpages              = ext4_readpages,
2936         .writepage              = ext4_writepage,
2937         .writepages             = ext4_da_writepages,
2938         .write_begin            = ext4_da_write_begin,
2939         .write_end              = ext4_da_write_end,
2940         .bmap                   = ext4_bmap,
2941         .invalidatepage         = ext4_da_invalidatepage,
2942         .releasepage            = ext4_releasepage,
2943         .direct_IO              = ext4_direct_IO,
2944         .migratepage            = buffer_migrate_page,
2945         .is_partially_uptodate  = block_is_partially_uptodate,
2946         .error_remove_page      = generic_error_remove_page,
2947 };
2948
2949 void ext4_set_aops(struct inode *inode)
2950 {
2951         if (ext4_should_order_data(inode) &&
2952                 test_opt(inode->i_sb, DELALLOC))
2953                 inode->i_mapping->a_ops = &ext4_da_aops;
2954         else if (ext4_should_order_data(inode))
2955                 inode->i_mapping->a_ops = &ext4_ordered_aops;
2956         else if (ext4_should_writeback_data(inode) &&
2957                  test_opt(inode->i_sb, DELALLOC))
2958                 inode->i_mapping->a_ops = &ext4_da_aops;
2959         else if (ext4_should_writeback_data(inode))
2960                 inode->i_mapping->a_ops = &ext4_writeback_aops;
2961         else
2962                 inode->i_mapping->a_ops = &ext4_journalled_aops;
2963 }
2964
2965 /*
2966  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
2967  * up to the end of the block which corresponds to `from'.
2968  * This required during truncate. We need to physically zero the tail end
2969  * of that block so it doesn't yield old data if the file is later grown.
2970  */
2971 int ext4_block_truncate_page(handle_t *handle,
2972                 struct address_space *mapping, loff_t from)
2973 {
2974         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2975         unsigned length;
2976         unsigned blocksize;
2977         struct inode *inode = mapping->host;
2978
2979         blocksize = inode->i_sb->s_blocksize;
2980         length = blocksize - (offset & (blocksize - 1));
2981
2982         return ext4_block_zero_page_range(handle, mapping, from, length);
2983 }
2984
2985 /*
2986  * ext4_block_zero_page_range() zeros out a mapping of length 'length'
2987  * starting from file offset 'from'.  The range to be zero'd must
2988  * be contained with in one block.  If the specified range exceeds
2989  * the end of the block it will be shortened to end of the block
2990  * that cooresponds to 'from'
2991  */
2992 int ext4_block_zero_page_range(handle_t *handle,
2993                 struct address_space *mapping, loff_t from, loff_t length)
2994 {
2995         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
2996         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2997         unsigned blocksize, max, pos;
2998         ext4_lblk_t iblock;
2999         struct inode *inode = mapping->host;
3000         struct buffer_head *bh;
3001         struct page *page;
3002         int err = 0;
3003
3004         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
3005                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
3006         if (!page)
3007                 return -EINVAL;
3008
3009         blocksize = inode->i_sb->s_blocksize;
3010         max = blocksize - (offset & (blocksize - 1));
3011
3012         /*
3013          * correct length if it does not fall between
3014          * 'from' and the end of the block
3015          */
3016         if (length > max || length < 0)
3017                 length = max;
3018
3019         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3020
3021         if (!page_has_buffers(page))
3022                 create_empty_buffers(page, blocksize, 0);
3023
3024         /* Find the buffer that contains "offset" */
3025         bh = page_buffers(page);
3026         pos = blocksize;
3027         while (offset >= pos) {
3028                 bh = bh->b_this_page;
3029                 iblock++;
3030                 pos += blocksize;
3031         }
3032
3033         err = 0;
3034         if (buffer_freed(bh)) {
3035                 BUFFER_TRACE(bh, "freed: skip");
3036                 goto unlock;
3037         }
3038
3039         if (!buffer_mapped(bh)) {
3040                 BUFFER_TRACE(bh, "unmapped");
3041                 ext4_get_block(inode, iblock, bh, 0);
3042                 /* unmapped? It's a hole - nothing to do */
3043                 if (!buffer_mapped(bh)) {
3044                         BUFFER_TRACE(bh, "still unmapped");
3045                         goto unlock;
3046                 }
3047         }
3048
3049         /* Ok, it's mapped. Make sure it's up-to-date */
3050         if (PageUptodate(page))
3051                 set_buffer_uptodate(bh);
3052
3053         if (!buffer_uptodate(bh)) {
3054                 err = -EIO;
3055                 ll_rw_block(READ, 1, &bh);
3056                 wait_on_buffer(bh);
3057                 /* Uhhuh. Read error. Complain and punt. */
3058                 if (!buffer_uptodate(bh))
3059                         goto unlock;
3060         }
3061
3062         if (ext4_should_journal_data(inode)) {
3063                 BUFFER_TRACE(bh, "get write access");
3064                 err = ext4_journal_get_write_access(handle, bh);
3065                 if (err)
3066                         goto unlock;
3067         }
3068
3069         zero_user(page, offset, length);
3070
3071         BUFFER_TRACE(bh, "zeroed end of block");
3072
3073         err = 0;
3074         if (ext4_should_journal_data(inode)) {
3075                 err = ext4_handle_dirty_metadata(handle, inode, bh);
3076         } else {
3077                 if (ext4_should_order_data(inode) && EXT4_I(inode)->jinode)
3078                         err = ext4_jbd2_file_inode(handle, inode);
3079                 mark_buffer_dirty(bh);
3080         }
3081
3082 unlock:
3083         unlock_page(page);
3084         page_cache_release(page);
3085         return err;
3086 }
3087
3088 int ext4_can_truncate(struct inode *inode)
3089 {
3090         if (S_ISREG(inode->i_mode))
3091                 return 1;
3092         if (S_ISDIR(inode->i_mode))
3093                 return 1;
3094         if (S_ISLNK(inode->i_mode))
3095                 return !ext4_inode_is_fast_symlink(inode);
3096         return 0;
3097 }
3098
3099 /*
3100  * ext4_punch_hole: punches a hole in a file by releaseing the blocks
3101  * associated with the given offset and length
3102  *
3103  * @inode:  File inode
3104  * @offset: The offset where the hole will begin
3105  * @len:    The length of the hole
3106  *
3107  * Returns: 0 on sucess or negative on failure
3108  */
3109
3110 int ext4_punch_hole(struct file *file, loff_t offset, loff_t length)
3111 {
3112         struct inode *inode = file->f_path.dentry->d_inode;
3113         if (!S_ISREG(inode->i_mode))
3114                 return -ENOTSUPP;
3115
3116         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
3117                 /* TODO: Add support for non extent hole punching */
3118                 return -ENOTSUPP;
3119         }
3120
3121         return ext4_ext_punch_hole(file, offset, length);
3122 }
3123
3124 /*
3125  * ext4_truncate()
3126  *
3127  * We block out ext4_get_block() block instantiations across the entire
3128  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
3129  * simultaneously on behalf of the same inode.
3130  *
3131  * As we work through the truncate and commmit bits of it to the journal there
3132  * is one core, guiding principle: the file's tree must always be consistent on
3133  * disk.  We must be able to restart the truncate after a crash.
3134  *
3135  * The file's tree may be transiently inconsistent in memory (although it
3136  * probably isn't), but whenever we close off and commit a journal transaction,
3137  * the contents of (the filesystem + the journal) must be consistent and
3138  * restartable.  It's pretty simple, really: bottom up, right to left (although
3139  * left-to-right works OK too).
3140  *
3141  * Note that at recovery time, journal replay occurs *before* the restart of
3142  * truncate against the orphan inode list.
3143  *
3144  * The committed inode has the new, desired i_size (which is the same as
3145  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
3146  * that this inode's truncate did not complete and it will again call
3147  * ext4_truncate() to have another go.  So there will be instantiated blocks
3148  * to the right of the truncation point in a crashed ext4 filesystem.  But
3149  * that's fine - as long as they are linked from the inode, the post-crash
3150  * ext4_truncate() run will find them and release them.
3151  */
3152 void ext4_truncate(struct inode *inode)
3153 {
3154         trace_ext4_truncate_enter(inode);
3155
3156         if (!ext4_can_truncate(inode))
3157                 return;
3158
3159         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3160
3161         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
3162                 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
3163
3164         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3165                 ext4_ext_truncate(inode);
3166         else
3167                 ext4_ind_truncate(inode);
3168
3169         trace_ext4_truncate_exit(inode);
3170 }
3171
3172 /*
3173  * ext4_get_inode_loc returns with an extra refcount against the inode's
3174  * underlying buffer_head on success. If 'in_mem' is true, we have all
3175  * data in memory that is needed to recreate the on-disk version of this
3176  * inode.
3177  */
3178 static int __ext4_get_inode_loc(struct inode *inode,
3179                                 struct ext4_iloc *iloc, int in_mem)
3180 {
3181         struct ext4_group_desc  *gdp;
3182         struct buffer_head      *bh;
3183         struct super_block      *sb = inode->i_sb;
3184         ext4_fsblk_t            block;
3185         int                     inodes_per_block, inode_offset;
3186
3187         iloc->bh = NULL;
3188         if (!ext4_valid_inum(sb, inode->i_ino))
3189                 return -EIO;
3190
3191         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
3192         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
3193         if (!gdp)
3194                 return -EIO;
3195
3196         /*
3197          * Figure out the offset within the block group inode table
3198          */
3199         inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
3200         inode_offset = ((inode->i_ino - 1) %
3201                         EXT4_INODES_PER_GROUP(sb));
3202         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
3203         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
3204
3205         bh = sb_getblk(sb, block);
3206         if (!bh) {
3207                 EXT4_ERROR_INODE_BLOCK(inode, block,
3208                                        "unable to read itable block");
3209                 return -EIO;
3210         }
3211         if (!buffer_uptodate(bh)) {
3212                 lock_buffer(bh);
3213
3214                 /*
3215                  * If the buffer has the write error flag, we have failed
3216                  * to write out another inode in the same block.  In this
3217                  * case, we don't have to read the block because we may
3218                  * read the old inode data successfully.
3219                  */
3220                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
3221                         set_buffer_uptodate(bh);
3222
3223                 if (buffer_uptodate(bh)) {
3224                         /* someone brought it uptodate while we waited */
3225                         unlock_buffer(bh);
3226                         goto has_buffer;
3227                 }
3228
3229                 /*
3230                  * If we have all information of the inode in memory and this
3231                  * is the only valid inode in the block, we need not read the
3232                  * block.
3233                  */
3234                 if (in_mem) {
3235                         struct buffer_head *bitmap_bh;
3236                         int i, start;
3237
3238                         start = inode_offset & ~(inodes_per_block - 1);
3239
3240                         /* Is the inode bitmap in cache? */
3241                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
3242                         if (!bitmap_bh)
3243                                 goto make_io;
3244
3245                         /*
3246                          * If the inode bitmap isn't in cache then the
3247                          * optimisation may end up performing two reads instead
3248                          * of one, so skip it.
3249                          */
3250                         if (!buffer_uptodate(bitmap_bh)) {
3251                                 brelse(bitmap_bh);
3252                                 goto make_io;
3253                         }
3254                         for (i = start; i < start + inodes_per_block; i++) {
3255                                 if (i == inode_offset)
3256                                         continue;
3257                                 if (ext4_test_bit(i, bitmap_bh->b_data))
3258                                         break;
3259                         }
3260                         brelse(bitmap_bh);
3261                         if (i == start + inodes_per_block) {
3262                                 /* all other inodes are free, so skip I/O */
3263                                 memset(bh->b_data, 0, bh->b_size);
3264                                 set_buffer_uptodate(bh);
3265                                 unlock_buffer(bh);
3266                                 goto has_buffer;
3267                         }
3268                 }
3269
3270 make_io:
3271                 /*
3272                  * If we need to do any I/O, try to pre-readahead extra
3273                  * blocks from the inode table.
3274                  */
3275                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
3276                         ext4_fsblk_t b, end, table;
3277                         unsigned num;
3278
3279                         table = ext4_inode_table(sb, gdp);
3280                         /* s_inode_readahead_blks is always a power of 2 */
3281                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
3282                         if (table > b)
3283                                 b = table;
3284                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
3285                         num = EXT4_INODES_PER_GROUP(sb);
3286                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
3287                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
3288                                 num -= ext4_itable_unused_count(sb, gdp);
3289                         table += num / inodes_per_block;
3290                         if (end > table)
3291                                 end = table;
3292                         while (b <= end)
3293                                 sb_breadahead(sb, b++);
3294                 }
3295
3296                 /*
3297                  * There are other valid inodes in the buffer, this inode
3298                  * has in-inode xattrs, or we don't have this inode in memory.
3299                  * Read the block from disk.
3300                  */
3301                 trace_ext4_load_inode(inode);
3302                 get_bh(bh);
3303                 bh->b_end_io = end_buffer_read_sync;
3304                 submit_bh(READ | REQ_META | REQ_PRIO, bh);
3305                 wait_on_buffer(bh);
3306                 if (!buffer_uptodate(bh)) {
3307                         EXT4_ERROR_INODE_BLOCK(inode, block,
3308                                                "unable to read itable block");
3309                         brelse(bh);
3310                         return -EIO;
3311                 }
3312         }
3313 has_buffer:
3314         iloc->bh = bh;
3315         return 0;
3316 }
3317
3318 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
3319 {
3320         /* We have all inode data except xattrs in memory here. */
3321         return __ext4_get_inode_loc(inode, iloc,
3322                 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
3323 }
3324
3325 void ext4_set_inode_flags(struct inode *inode)
3326 {
3327         unsigned int flags = EXT4_I(inode)->i_flags;
3328
3329         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
3330         if (flags & EXT4_SYNC_FL)
3331                 inode->i_flags |= S_SYNC;
3332         if (flags & EXT4_APPEND_FL)
3333                 inode->i_flags |= S_APPEND;
3334         if (flags & EXT4_IMMUTABLE_FL)
3335                 inode->i_flags |= S_IMMUTABLE;
3336         if (flags & EXT4_NOATIME_FL)
3337                 inode->i_flags |= S_NOATIME;
3338         if (flags & EXT4_DIRSYNC_FL)
3339                 inode->i_flags |= S_DIRSYNC;
3340 }
3341
3342 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
3343 void ext4_get_inode_flags(struct ext4_inode_info *ei)
3344 {
3345         unsigned int vfs_fl;
3346         unsigned long old_fl, new_fl;
3347
3348         do {
3349                 vfs_fl = ei->vfs_inode.i_flags;
3350                 old_fl = ei->i_flags;
3351                 new_fl = old_fl & ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
3352                                 EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|
3353                                 EXT4_DIRSYNC_FL);
3354                 if (vfs_fl & S_SYNC)
3355                         new_fl |= EXT4_SYNC_FL;
3356                 if (vfs_fl & S_APPEND)
3357                         new_fl |= EXT4_APPEND_FL;
3358                 if (vfs_fl & S_IMMUTABLE)
3359                         new_fl |= EXT4_IMMUTABLE_FL;
3360                 if (vfs_fl & S_NOATIME)
3361                         new_fl |= EXT4_NOATIME_FL;
3362                 if (vfs_fl & S_DIRSYNC)
3363                         new_fl |= EXT4_DIRSYNC_FL;
3364         } while (cmpxchg(&ei->i_flags, old_fl, new_fl) != old_fl);
3365 }
3366
3367 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
3368                                   struct ext4_inode_info *ei)
3369 {
3370         blkcnt_t i_blocks ;
3371         struct inode *inode = &(ei->vfs_inode);
3372         struct super_block *sb = inode->i_sb;
3373
3374         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
3375                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
3376                 /* we are using combined 48 bit field */
3377                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
3378                                         le32_to_cpu(raw_inode->i_blocks_lo);
3379                 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
3380                         /* i_blocks represent file system block size */
3381                         return i_blocks  << (inode->i_blkbits - 9);
3382                 } else {
3383                         return i_blocks;
3384                 }
3385         } else {
3386                 return le32_to_cpu(raw_inode->i_blocks_lo);
3387         }
3388 }
3389
3390 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
3391 {
3392         struct ext4_iloc iloc;
3393         struct ext4_inode *raw_inode;
3394         struct ext4_inode_info *ei;
3395         struct inode *inode;
3396         journal_t *journal = EXT4_SB(sb)->s_journal;
3397         long ret;
3398         int block;
3399
3400         inode = iget_locked(sb, ino);
3401         if (!inode)
3402                 return ERR_PTR(-ENOMEM);
3403         if (!(inode->i_state & I_NEW))
3404                 return inode;
3405
3406         ei = EXT4_I(inode);
3407         iloc.bh = NULL;
3408
3409         ret = __ext4_get_inode_loc(inode, &iloc, 0);
3410         if (ret < 0)
3411                 goto bad_inode;
3412         raw_inode = ext4_raw_inode(&iloc);
3413         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
3414         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
3415         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
3416         if (!(test_opt(inode->i_sb, NO_UID32))) {
3417                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
3418                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
3419         }
3420         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
3421
3422         ext4_clear_state_flags(ei);     /* Only relevant on 32-bit archs */
3423         ei->i_dir_start_lookup = 0;
3424         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
3425         /* We now have enough fields to check if the inode was active or not.
3426          * This is needed because nfsd might try to access dead inodes
3427          * the test is that same one that e2fsck uses
3428          * NeilBrown 1999oct15
3429          */
3430         if (inode->i_nlink == 0) {
3431                 if (inode->i_mode == 0 ||
3432                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
3433                         /* this inode is deleted */
3434                         ret = -ESTALE;
3435                         goto bad_inode;
3436                 }
3437                 /* The only unlinked inodes we let through here have
3438                  * valid i_mode and are being read by the orphan
3439                  * recovery code: that's fine, we're about to complete
3440                  * the process of deleting those. */
3441         }
3442         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
3443         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
3444         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
3445         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
3446                 ei->i_file_acl |=
3447                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
3448         inode->i_size = ext4_isize(raw_inode);
3449         ei->i_disksize = inode->i_size;
3450 #ifdef CONFIG_QUOTA
3451         ei->i_reserved_quota = 0;
3452 #endif
3453         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
3454         ei->i_block_group = iloc.block_group;
3455         ei->i_last_alloc_group = ~0;
3456         /*
3457          * NOTE! The in-memory inode i_data array is in little-endian order
3458          * even on big-endian machines: we do NOT byteswap the block numbers!
3459          */
3460         for (block = 0; block < EXT4_N_BLOCKS; block++)
3461                 ei->i_data[block] = raw_inode->i_block[block];
3462         INIT_LIST_HEAD(&ei->i_orphan);
3463
3464         /*
3465          * Set transaction id's of transactions that have to be committed
3466          * to finish f[data]sync. We set them to currently running transaction
3467          * as we cannot be sure that the inode or some of its metadata isn't
3468          * part of the transaction - the inode could have been reclaimed and
3469          * now it is reread from disk.
3470          */
3471         if (journal) {
3472                 transaction_t *transaction;
3473                 tid_t tid;
3474
3475                 read_lock(&journal->j_state_lock);
3476                 if (journal->j_running_transaction)
3477                         transaction = journal->j_running_transaction;
3478                 else
3479                         transaction = journal->j_committing_transaction;
3480                 if (transaction)
3481                         tid = transaction->t_tid;
3482                 else
3483                         tid = journal->j_commit_sequence;
3484                 read_unlock(&journal->j_state_lock);
3485                 ei->i_sync_tid = tid;
3486                 ei->i_datasync_tid = tid;
3487         }
3488
3489         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
3490                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
3491                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
3492                     EXT4_INODE_SIZE(inode->i_sb)) {
3493                         ret = -EIO;
3494                         goto bad_inode;
3495                 }
3496                 if (ei->i_extra_isize == 0) {
3497                         /* The extra space is currently unused. Use it. */
3498                         ei->i_extra_isize = sizeof(struct ext4_inode) -
3499                                             EXT4_GOOD_OLD_INODE_SIZE;
3500                 } else {
3501                         __le32 *magic = (void *)raw_inode +
3502                                         EXT4_GOOD_OLD_INODE_SIZE +
3503                                         ei->i_extra_isize;
3504                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
3505                                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
3506                 }
3507         } else
3508                 ei->i_extra_isize = 0;
3509
3510         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
3511         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
3512         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
3513         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
3514
3515         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
3516         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
3517                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
3518                         inode->i_version |=
3519                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
3520         }
3521
3522         ret = 0;
3523         if (ei->i_file_acl &&
3524             !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
3525                 EXT4_ERROR_INODE(inode, "bad extended attribute block %llu",
3526                                  ei->i_file_acl);
3527                 ret = -EIO;
3528                 goto bad_inode;
3529         } else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
3530                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
3531                     (S_ISLNK(inode->i_mode) &&
3532                      !ext4_inode_is_fast_symlink(inode)))
3533                         /* Validate extent which is part of inode */
3534                         ret = ext4_ext_check_inode(inode);
3535         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
3536                    (S_ISLNK(inode->i_mode) &&
3537                     !ext4_inode_is_fast_symlink(inode))) {
3538                 /* Validate block references which are part of inode */
3539                 ret = ext4_ind_check_inode(inode);
3540         }
3541         if (ret)
3542                 goto bad_inode;
3543
3544         if (S_ISREG(inode->i_mode)) {
3545                 inode->i_op = &ext4_file_inode_operations;
3546                 inode->i_fop = &ext4_file_operations;
3547                 ext4_set_aops(inode);
3548         } else if (S_ISDIR(inode->i_mode)) {
3549                 inode->i_op = &ext4_dir_inode_operations;
3550                 inode->i_fop = &ext4_dir_operations;
3551         } else if (S_ISLNK(inode->i_mode)) {
3552                 if (ext4_inode_is_fast_symlink(inode)) {
3553                         inode->i_op = &ext4_fast_symlink_inode_operations;
3554                         nd_terminate_link(ei->i_data, inode->i_size,
3555                                 sizeof(ei->i_data) - 1);
3556                 } else {
3557                         inode->i_op = &ext4_symlink_inode_operations;
3558                         ext4_set_aops(inode);
3559                 }
3560         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
3561               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
3562                 inode->i_op = &ext4_special_inode_operations;
3563                 if (raw_inode->i_block[0])
3564                         init_special_inode(inode, inode->i_mode,
3565                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
3566                 else
3567                         init_special_inode(inode, inode->i_mode,
3568                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
3569         } else {
3570                 ret = -EIO;
3571                 EXT4_ERROR_INODE(inode, "bogus i_mode (%o)", inode->i_mode);
3572                 goto bad_inode;
3573         }
3574         brelse(iloc.bh);
3575         ext4_set_inode_flags(inode);
3576         unlock_new_inode(inode);
3577         return inode;
3578
3579 bad_inode:
3580         brelse(iloc.bh);
3581         iget_failed(inode);
3582         return ERR_PTR(ret);
3583 }
3584
3585 static int ext4_inode_blocks_set(handle_t *handle,
3586                                 struct ext4_inode *raw_inode,
3587                                 struct ext4_inode_info *ei)
3588 {
3589         struct inode *inode = &(ei->vfs_inode);
3590         u64 i_blocks = inode->i_blocks;
3591         struct super_block *sb = inode->i_sb;
3592
3593         if (i_blocks <= ~0U) {
3594                 /*
3595                  * i_blocks can be represnted in a 32 bit variable
3596                  * as multiple of 512 bytes
3597                  */
3598                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
3599                 raw_inode->i_blocks_high = 0;
3600                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
3601                 return 0;
3602         }
3603         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
3604                 return -EFBIG;
3605
3606         if (i_blocks <= 0xffffffffffffULL) {
3607                 /*
3608                  * i_blocks can be represented in a 48 bit variable
3609                  * as multiple of 512 bytes
3610                  */
3611                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
3612                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
3613                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
3614         } else {
3615                 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
3616                 /* i_block is stored in file system block size */
3617                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
3618                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
3619                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
3620         }
3621         return 0;
3622 }
3623
3624 /*
3625  * Post the struct inode info into an on-disk inode location in the
3626  * buffer-cache.  This gobbles the caller's reference to the
3627  * buffer_head in the inode location struct.
3628  *
3629  * The caller must have write access to iloc->bh.
3630  */
3631 static int ext4_do_update_inode(handle_t *handle,
3632                                 struct inode *inode,
3633                                 struct ext4_iloc *iloc)
3634 {
3635         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
3636         struct ext4_inode_info *ei = EXT4_I(inode);
3637         struct buffer_head *bh = iloc->bh;
3638         int err = 0, rc, block;
3639
3640         /* For fields not not tracking in the in-memory inode,
3641          * initialise them to zero for new inodes. */
3642         if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
3643                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
3644
3645         ext4_get_inode_flags(ei);
3646         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
3647         if (!(test_opt(inode->i_sb, NO_UID32))) {
3648                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
3649                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
3650 /*
3651  * Fix up interoperability with old kernels. Otherwise, old inodes get
3652  * re-used with the upper 16 bits of the uid/gid intact
3653  */
3654                 if (!ei->i_dtime) {
3655                         raw_inode->i_uid_high =
3656                                 cpu_to_le16(high_16_bits(inode->i_uid));
3657                         raw_inode->i_gid_high =
3658                                 cpu_to_le16(high_16_bits(inode->i_gid));
3659                 } else {
3660                         raw_inode->i_uid_high = 0;
3661                         raw_inode->i_gid_high = 0;
3662                 }
3663         } else {
3664                 raw_inode->i_uid_low =
3665                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
3666                 raw_inode->i_gid_low =
3667                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
3668                 raw_inode->i_uid_high = 0;
3669                 raw_inode->i_gid_high = 0;
3670         }
3671         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
3672
3673         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
3674         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
3675         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
3676         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
3677
3678         if (ext4_inode_blocks_set(handle, raw_inode, ei))
3679                 goto out_brelse;
3680         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
3681         raw_inode->i_flags = cpu_to_le32(ei->i_flags & 0xFFFFFFFF);
3682         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
3683             cpu_to_le32(EXT4_OS_HURD))
3684                 raw_inode->i_file_acl_high =
3685                         cpu_to_le16(ei->i_file_acl >> 32);
3686         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
3687         ext4_isize_set(raw_inode, ei->i_disksize);
3688         if (ei->i_disksize > 0x7fffffffULL) {
3689                 struct super_block *sb = inode->i_sb;
3690                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
3691                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
3692                                 EXT4_SB(sb)->s_es->s_rev_level ==
3693                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
3694                         /* If this is the first large file
3695                          * created, add a flag to the superblock.
3696                          */
3697                         err = ext4_journal_get_write_access(handle,
3698                                         EXT4_SB(sb)->s_sbh);
3699                         if (err)
3700                                 goto out_brelse;
3701                         ext4_update_dynamic_rev(sb);
3702                         EXT4_SET_RO_COMPAT_FEATURE(sb,
3703                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
3704                         sb->s_dirt = 1;
3705                         ext4_handle_sync(handle);
3706                         err = ext4_handle_dirty_metadata(handle, NULL,
3707                                         EXT4_SB(sb)->s_sbh);
3708                 }
3709         }
3710         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
3711         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
3712                 if (old_valid_dev(inode->i_rdev)) {
3713                         raw_inode->i_block[0] =
3714                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
3715                         raw_inode->i_block[1] = 0;
3716                 } else {
3717                         raw_inode->i_block[0] = 0;
3718                         raw_inode->i_block[1] =
3719                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
3720                         raw_inode->i_block[2] = 0;
3721                 }
3722         } else
3723                 for (block = 0; block < EXT4_N_BLOCKS; block++)
3724                         raw_inode->i_block[block] = ei->i_data[block];
3725
3726         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
3727         if (ei->i_extra_isize) {
3728                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
3729                         raw_inode->i_version_hi =
3730                         cpu_to_le32(inode->i_version >> 32);
3731                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
3732         }
3733
3734         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
3735         rc = ext4_handle_dirty_metadata(handle, NULL, bh);
3736         if (!err)
3737                 err = rc;
3738         ext4_clear_inode_state(inode, EXT4_STATE_NEW);
3739
3740         ext4_update_inode_fsync_trans(handle, inode, 0);
3741 out_brelse:
3742         brelse(bh);
3743         ext4_std_error(inode->i_sb, err);
3744         return err;
3745 }
3746
3747 /*
3748  * ext4_write_inode()
3749  *
3750  * We are called from a few places:
3751  *
3752  * - Within generic_file_write() for O_SYNC files.
3753  *   Here, there will be no transaction running. We wait for any running
3754  *   trasnaction to commit.
3755  *
3756  * - Within sys_sync(), kupdate and such.
3757  *   We wait on commit, if tol to.
3758  *
3759  * - Within prune_icache() (PF_MEMALLOC == true)
3760  *   Here we simply return.  We can't afford to block kswapd on the
3761  *   journal commit.
3762  *
3763  * In all cases it is actually safe for us to return without doing anything,
3764  * because the inode has been copied into a raw inode buffer in
3765  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
3766  * knfsd.
3767  *
3768  * Note that we are absolutely dependent upon all inode dirtiers doing the
3769  * right thing: they *must* call mark_inode_dirty() after dirtying info in
3770  * which we are interested.
3771  *
3772  * It would be a bug for them to not do this.  The code:
3773  *
3774  *      mark_inode_dirty(inode)
3775  *      stuff();
3776  *      inode->i_size = expr;
3777  *
3778  * is in error because a kswapd-driven write_inode() could occur while
3779  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
3780  * will no longer be on the superblock's dirty inode list.
3781  */
3782 int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
3783 {
3784         int err;
3785
3786         if (current->flags & PF_MEMALLOC)
3787                 return 0;
3788
3789         if (EXT4_SB(inode->i_sb)->s_journal) {
3790                 if (ext4_journal_current_handle()) {
3791                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
3792                         dump_stack();
3793                         return -EIO;
3794                 }
3795
3796                 if (wbc->sync_mode != WB_SYNC_ALL)
3797                         return 0;
3798
3799                 err = ext4_force_commit(inode->i_sb);
3800         } else {
3801                 struct ext4_iloc iloc;
3802
3803                 err = __ext4_get_inode_loc(inode, &iloc, 0);
3804                 if (err)
3805                         return err;
3806                 if (wbc->sync_mode == WB_SYNC_ALL)
3807                         sync_dirty_buffer(iloc.bh);
3808                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
3809                         EXT4_ERROR_INODE_BLOCK(inode, iloc.bh->b_blocknr,
3810                                          "IO error syncing inode");
3811                         err = -EIO;
3812                 }
3813                 brelse(iloc.bh);
3814         }
3815         return err;
3816 }
3817
3818 /*
3819  * ext4_setattr()
3820  *
3821  * Called from notify_change.
3822  *
3823  * We want to trap VFS attempts to truncate the file as soon as
3824  * possible.  In particular, we want to make sure that when the VFS
3825  * shrinks i_size, we put the inode on the orphan list and modify
3826  * i_disksize immediately, so that during the subsequent flushing of
3827  * dirty pages and freeing of disk blocks, we can guarantee that any
3828  * commit will leave the blocks being flushed in an unused state on
3829  * disk.  (On recovery, the inode will get truncated and the blocks will
3830  * be freed, so we have a strong guarantee that no future commit will
3831  * leave these blocks visible to the user.)
3832  *
3833  * Another thing we have to assure is that if we are in ordered mode
3834  * and inode is still attached to the committing transaction, we must
3835  * we start writeout of all the dirty pages which are being truncated.
3836  * This way we are sure that all the data written in the previous
3837  * transaction are already on disk (truncate waits for pages under
3838  * writeback).
3839  *
3840  * Called with inode->i_mutex down.
3841  */
3842 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
3843 {
3844         struct inode *inode = dentry->d_inode;
3845         int error, rc = 0;
3846         int orphan = 0;
3847         const unsigned int ia_valid = attr->ia_valid;
3848
3849         error = inode_change_ok(inode, attr);
3850         if (error)
3851                 return error;
3852
3853         if (is_quota_modification(inode, attr))
3854                 dquot_initialize(inode);
3855         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
3856                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
3857                 handle_t *handle;
3858
3859                 /* (user+group)*(old+new) structure, inode write (sb,
3860                  * inode block, ? - but truncate inode update has it) */
3861                 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
3862                                         EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
3863                 if (IS_ERR(handle)) {
3864                         error = PTR_ERR(handle);
3865                         goto err_out;
3866                 }
3867                 error = dquot_transfer(inode, attr);
3868                 if (error) {
3869                         ext4_journal_stop(handle);
3870                         return error;
3871                 }
3872                 /* Update corresponding info in inode so that everything is in
3873                  * one transaction */
3874                 if (attr->ia_valid & ATTR_UID)
3875                         inode->i_uid = attr->ia_uid;
3876                 if (attr->ia_valid & ATTR_GID)
3877                         inode->i_gid = attr->ia_gid;
3878                 error = ext4_mark_inode_dirty(handle, inode);
3879                 ext4_journal_stop(handle);
3880         }
3881
3882         if (attr->ia_valid & ATTR_SIZE) {
3883                 inode_dio_wait(inode);
3884
3885                 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
3886                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
3887
3888                         if (attr->ia_size > sbi->s_bitmap_maxbytes)
3889                                 return -EFBIG;
3890                 }
3891         }
3892
3893         if (S_ISREG(inode->i_mode) &&
3894             attr->ia_valid & ATTR_SIZE &&
3895             (attr->ia_size < inode->i_size)) {
3896                 handle_t *handle;
3897
3898                 handle = ext4_journal_start(inode, 3);
3899                 if (IS_ERR(handle)) {
3900                         error = PTR_ERR(handle);
3901                         goto err_out;
3902                 }
3903                 if (ext4_handle_valid(handle)) {
3904                         error = ext4_orphan_add(handle, inode);
3905                         orphan = 1;
3906                 }
3907                 EXT4_I(inode)->i_disksize = attr->ia_size;
3908                 rc = ext4_mark_inode_dirty(handle, inode);
3909                 if (!error)
3910                         error = rc;
3911                 ext4_journal_stop(handle);
3912
3913                 if (ext4_should_order_data(inode)) {
3914                         error = ext4_begin_ordered_truncate(inode,
3915                                                             attr->ia_size);
3916                         if (error) {
3917                                 /* Do as much error cleanup as possible */
3918                                 handle = ext4_journal_start(inode, 3);
3919                                 if (IS_ERR(handle)) {
3920                                         ext4_orphan_del(NULL, inode);
3921                                         goto err_out;
3922                                 }
3923                                 ext4_orphan_del(handle, inode);
3924                                 orphan = 0;
3925                                 ext4_journal_stop(handle);
3926                                 goto err_out;
3927                         }
3928                 }
3929         }
3930
3931         if (attr->ia_valid & ATTR_SIZE) {
3932                 if (attr->ia_size != i_size_read(inode)) {
3933                         truncate_setsize(inode, attr->ia_size);
3934                         ext4_truncate(inode);
3935                 } else if (ext4_test_inode_flag(inode, EXT4_INODE_EOFBLOCKS))
3936                         ext4_truncate(inode);
3937         }
3938
3939         if (!rc) {
3940                 setattr_copy(inode, attr);
3941                 mark_inode_dirty(inode);
3942         }
3943
3944         /*
3945          * If the call to ext4_truncate failed to get a transaction handle at
3946          * all, we need to clean up the in-core orphan list manually.
3947          */
3948         if (orphan && inode->i_nlink)
3949                 ext4_orphan_del(NULL, inode);
3950
3951         if (!rc && (ia_valid & ATTR_MODE))
3952                 rc = ext4_acl_chmod(inode);
3953
3954 err_out:
3955         ext4_std_error(inode->i_sb, error);
3956         if (!error)
3957                 error = rc;
3958         return error;
3959 }
3960
3961 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
3962                  struct kstat *stat)
3963 {
3964         struct inode *inode;
3965         unsigned long delalloc_blocks;
3966
3967         inode = dentry->d_inode;
3968         generic_fillattr(inode, stat);
3969
3970         /*
3971          * We can't update i_blocks if the block allocation is delayed
3972          * otherwise in the case of system crash before the real block
3973          * allocation is done, we will have i_blocks inconsistent with
3974          * on-disk file blocks.
3975          * We always keep i_blocks updated together with real
3976          * allocation. But to not confuse with user, stat
3977          * will return the blocks that include the delayed allocation
3978          * blocks for this file.
3979          */
3980         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
3981
3982         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
3983         return 0;
3984 }
3985
3986 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
3987 {
3988         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
3989                 return ext4_ind_trans_blocks(inode, nrblocks, chunk);
3990         return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
3991 }
3992
3993 /*
3994  * Account for index blocks, block groups bitmaps and block group
3995  * descriptor blocks if modify datablocks and index blocks
3996  * worse case, the indexs blocks spread over different block groups
3997  *
3998  * If datablocks are discontiguous, they are possible to spread over
3999  * different block groups too. If they are contiuguous, with flexbg,
4000  * they could still across block group boundary.
4001  *
4002  * Also account for superblock, inode, quota and xattr blocks
4003  */
4004 static int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4005 {
4006         ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
4007         int gdpblocks;
4008         int idxblocks;
4009         int ret = 0;
4010
4011         /*
4012          * How many index blocks need to touch to modify nrblocks?
4013          * The "Chunk" flag indicating whether the nrblocks is
4014          * physically contiguous on disk
4015          *
4016          * For Direct IO and fallocate, they calls get_block to allocate
4017          * one single extent at a time, so they could set the "Chunk" flag
4018          */
4019         idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
4020
4021         ret = idxblocks;
4022
4023         /*
4024          * Now let's see how many group bitmaps and group descriptors need
4025          * to account
4026          */
4027         groups = idxblocks;
4028         if (chunk)
4029                 groups += 1;
4030         else
4031                 groups += nrblocks;
4032
4033         gdpblocks = groups;
4034         if (groups > ngroups)
4035                 groups = ngroups;
4036         if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
4037                 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
4038
4039         /* bitmaps and block group descriptor blocks */
4040         ret += groups + gdpblocks;
4041
4042         /* Blocks for super block, inode, quota and xattr blocks */
4043         ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
4044
4045         return ret;
4046 }
4047
4048 /*
4049  * Calculate the total number of credits to reserve to fit
4050  * the modification of a single pages into a single transaction,
4051  * which may include multiple chunks of block allocations.
4052  *
4053  * This could be called via ext4_write_begin()
4054  *
4055  * We need to consider the worse case, when
4056  * one new block per extent.
4057  */
4058 int ext4_writepage_trans_blocks(struct inode *inode)
4059 {
4060         int bpp = ext4_journal_blocks_per_page(inode);
4061         int ret;
4062
4063         ret = ext4_meta_trans_blocks(inode, bpp, 0);
4064
4065         /* Account for data blocks for journalled mode */
4066         if (ext4_should_journal_data(inode))
4067                 ret += bpp;
4068         return ret;
4069 }
4070
4071 /*
4072  * Calculate the journal credits for a chunk of data modification.
4073  *
4074  * This is called from DIO, fallocate or whoever calling
4075  * ext4_map_blocks() to map/allocate a chunk of contiguous disk blocks.
4076  *
4077  * journal buffers for data blocks are not included here, as DIO
4078  * and fallocate do no need to journal data buffers.
4079  */
4080 int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
4081 {
4082         return ext4_meta_trans_blocks(inode, nrblocks, 1);
4083 }
4084
4085 /*
4086  * The caller must have previously called ext4_reserve_inode_write().
4087  * Give this, we know that the caller already has write access to iloc->bh.
4088  */
4089 int ext4_mark_iloc_dirty(handle_t *handle,
4090                          struct inode *inode, struct ext4_iloc *iloc)
4091 {
4092         int err = 0;
4093
4094         if (test_opt(inode->i_sb, I_VERSION))
4095                 inode_inc_iversion(inode);
4096
4097         /* the do_update_inode consumes one bh->b_count */
4098         get_bh(iloc->bh);
4099
4100         /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
4101         err = ext4_do_update_inode(handle, inode, iloc);
4102         put_bh(iloc->bh);
4103         return err;
4104 }
4105
4106 /*
4107  * On success, We end up with an outstanding reference count against
4108  * iloc->bh.  This _must_ be cleaned up later.
4109  */
4110
4111 int
4112 ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
4113                          struct ext4_iloc *iloc)
4114 {
4115         int err;
4116
4117         err = ext4_get_inode_loc(inode, iloc);
4118         if (!err) {
4119                 BUFFER_TRACE(iloc->bh, "get_write_access");
4120                 err = ext4_journal_get_write_access(handle, iloc->bh);
4121                 if (err) {
4122                         brelse(iloc->bh);
4123                         iloc->bh = NULL;
4124                 }
4125         }
4126         ext4_std_error(inode->i_sb, err);
4127         return err;
4128 }
4129
4130 /*
4131  * Expand an inode by new_extra_isize bytes.
4132  * Returns 0 on success or negative error number on failure.
4133  */
4134 static int ext4_expand_extra_isize(struct inode *inode,
4135                                    unsigned int new_extra_isize,
4136                                    struct ext4_iloc iloc,
4137                                    handle_t *handle)
4138 {
4139         struct ext4_inode *raw_inode;
4140         struct ext4_xattr_ibody_header *header;
4141
4142         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
4143                 return 0;
4144
4145         raw_inode = ext4_raw_inode(&iloc);
4146
4147         header = IHDR(inode, raw_inode);
4148
4149         /* No extended attributes present */
4150         if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR) ||
4151             header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
4152                 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
4153                         new_extra_isize);
4154                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
4155                 return 0;
4156         }
4157
4158         /* try to expand with EAs present */
4159         return ext4_expand_extra_isize_ea(inode, new_extra_isize,
4160                                           raw_inode, handle);
4161 }
4162
4163 /*
4164  * What we do here is to mark the in-core inode as clean with respect to inode
4165  * dirtiness (it may still be data-dirty).
4166  * This means that the in-core inode may be reaped by prune_icache
4167  * without having to perform any I/O.  This is a very good thing,
4168  * because *any* task may call prune_icache - even ones which
4169  * have a transaction open against a different journal.
4170  *
4171  * Is this cheating?  Not really.  Sure, we haven't written the
4172  * inode out, but prune_icache isn't a user-visible syncing function.
4173  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
4174  * we start and wait on commits.
4175  *
4176  * Is this efficient/effective?  Well, we're being nice to the system
4177  * by cleaning up our inodes proactively so they can be reaped
4178  * without I/O.  But we are potentially leaving up to five seconds'
4179  * worth of inodes floating about which prune_icache wants us to
4180  * write out.  One way to fix that would be to get prune_icache()
4181  * to do a write_super() to free up some memory.  It has the desired
4182  * effect.
4183  */
4184 int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
4185 {
4186         struct ext4_iloc iloc;
4187         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4188         static unsigned int mnt_count;
4189         int err, ret;
4190
4191         might_sleep();
4192         trace_ext4_mark_inode_dirty(inode, _RET_IP_);
4193         err = ext4_reserve_inode_write(handle, inode, &iloc);
4194         if (ext4_handle_valid(handle) &&
4195             EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
4196             !ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND)) {
4197                 /*
4198                  * We need extra buffer credits since we may write into EA block
4199                  * with this same handle. If journal_extend fails, then it will
4200                  * only result in a minor loss of functionality for that inode.
4201                  * If this is felt to be critical, then e2fsck should be run to
4202                  * force a large enough s_min_extra_isize.
4203                  */
4204                 if ((jbd2_journal_extend(handle,
4205                              EXT4_DATA_TRANS_BLOCKS(inode->i_sb))) == 0) {
4206                         ret = ext4_expand_extra_isize(inode,
4207                                                       sbi->s_want_extra_isize,
4208                                                       iloc, handle);
4209                         if (ret) {
4210                                 ext4_set_inode_state(inode,
4211                                                      EXT4_STATE_NO_EXPAND);
4212                                 if (mnt_count !=
4213                                         le16_to_cpu(sbi->s_es->s_mnt_count)) {
4214                                         ext4_warning(inode->i_sb,
4215                                         "Unable to expand inode %lu. Delete"
4216                                         " some EAs or run e2fsck.",
4217                                         inode->i_ino);
4218                                         mnt_count =
4219                                           le16_to_cpu(sbi->s_es->s_mnt_count);
4220                                 }
4221                         }
4222                 }
4223         }
4224         if (!err)
4225                 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
4226         return err;
4227 }
4228
4229 /*
4230  * ext4_dirty_inode() is called from __mark_inode_dirty()
4231  *
4232  * We're really interested in the case where a file is being extended.
4233  * i_size has been changed by generic_commit_write() and we thus need
4234  * to include the updated inode in the current transaction.
4235  *
4236  * Also, dquot_alloc_block() will always dirty the inode when blocks
4237  * are allocated to the file.
4238  *
4239  * If the inode is marked synchronous, we don't honour that here - doing
4240  * so would cause a commit on atime updates, which we don't bother doing.
4241  * We handle synchronous inodes at the highest possible level.
4242  */
4243 void ext4_dirty_inode(struct inode *inode, int flags)
4244 {
4245         handle_t *handle;
4246
4247         handle = ext4_journal_start(inode, 2);
4248         if (IS_ERR(handle))
4249                 goto out;
4250
4251         ext4_mark_inode_dirty(handle, inode);
4252
4253         ext4_journal_stop(handle);
4254 out:
4255         return;
4256 }
4257
4258 #if 0
4259 /*
4260  * Bind an inode's backing buffer_head into this transaction, to prevent
4261  * it from being flushed to disk early.  Unlike
4262  * ext4_reserve_inode_write, this leaves behind no bh reference and
4263  * returns no iloc structure, so the caller needs to repeat the iloc
4264  * lookup to mark the inode dirty later.
4265  */
4266 static int ext4_pin_inode(handle_t *handle, struct inode *inode)
4267 {
4268         struct ext4_iloc iloc;
4269
4270         int err = 0;
4271         if (handle) {
4272                 err = ext4_get_inode_loc(inode, &iloc);
4273                 if (!err) {
4274                         BUFFER_TRACE(iloc.bh, "get_write_access");
4275                         err = jbd2_journal_get_write_access(handle, iloc.bh);
4276                         if (!err)
4277                                 err = ext4_handle_dirty_metadata(handle,
4278                                                                  NULL,
4279                                                                  iloc.bh);
4280                         brelse(iloc.bh);
4281                 }
4282         }
4283         ext4_std_error(inode->i_sb, err);
4284         return err;
4285 }
4286 #endif
4287
4288 int ext4_change_inode_journal_flag(struct inode *inode, int val)
4289 {
4290         journal_t *journal;
4291         handle_t *handle;
4292         int err;
4293
4294         /*
4295          * We have to be very careful here: changing a data block's
4296          * journaling status dynamically is dangerous.  If we write a
4297          * data block to the journal, change the status and then delete
4298          * that block, we risk forgetting to revoke the old log record
4299          * from the journal and so a subsequent replay can corrupt data.
4300          * So, first we make sure that the journal is empty and that
4301          * nobody is changing anything.
4302          */
4303
4304         journal = EXT4_JOURNAL(inode);
4305         if (!journal)
4306                 return 0;
4307         if (is_journal_aborted(journal))
4308                 return -EROFS;
4309
4310         jbd2_journal_lock_updates(journal);
4311         jbd2_journal_flush(journal);
4312
4313         /*
4314          * OK, there are no updates running now, and all cached data is
4315          * synced to disk.  We are now in a completely consistent state
4316          * which doesn't have anything in the journal, and we know that
4317          * no filesystem updates are running, so it is safe to modify
4318          * the inode's in-core data-journaling state flag now.
4319          */
4320
4321         if (val)
4322                 ext4_set_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
4323         else
4324                 ext4_clear_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
4325         ext4_set_aops(inode);
4326
4327         jbd2_journal_unlock_updates(journal);
4328
4329         /* Finally we can mark the inode as dirty. */
4330
4331         handle = ext4_journal_start(inode, 1);
4332         if (IS_ERR(handle))
4333                 return PTR_ERR(handle);
4334
4335         err = ext4_mark_inode_dirty(handle, inode);
4336         ext4_handle_sync(handle);
4337         ext4_journal_stop(handle);
4338         ext4_std_error(inode->i_sb, err);
4339
4340         return err;
4341 }
4342
4343 static int ext4_bh_unmapped(handle_t *handle, struct buffer_head *bh)
4344 {
4345         return !buffer_mapped(bh);
4346 }
4347
4348 int ext4_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
4349 {
4350         struct page *page = vmf->page;
4351         loff_t size;
4352         unsigned long len;
4353         int ret;
4354         struct file *file = vma->vm_file;
4355         struct inode *inode = file->f_path.dentry->d_inode;
4356         struct address_space *mapping = inode->i_mapping;
4357         handle_t *handle;
4358         get_block_t *get_block;
4359         int retries = 0;
4360
4361         /*
4362          * This check is racy but catches the common case. We rely on
4363          * __block_page_mkwrite() to do a reliable check.
4364          */
4365         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
4366         /* Delalloc case is easy... */
4367         if (test_opt(inode->i_sb, DELALLOC) &&
4368             !ext4_should_journal_data(inode) &&
4369             !ext4_nonda_switch(inode->i_sb)) {
4370                 do {
4371                         ret = __block_page_mkwrite(vma, vmf,
4372                                                    ext4_da_get_block_prep);
4373                 } while (ret == -ENOSPC &&
4374                        ext4_should_retry_alloc(inode->i_sb, &retries));
4375                 goto out_ret;
4376         }
4377
4378         lock_page(page);
4379         size = i_size_read(inode);
4380         /* Page got truncated from under us? */
4381         if (page->mapping != mapping || page_offset(page) > size) {
4382                 unlock_page(page);
4383                 ret = VM_FAULT_NOPAGE;
4384                 goto out;
4385         }
4386
4387         if (page->index == size >> PAGE_CACHE_SHIFT)
4388                 len = size & ~PAGE_CACHE_MASK;
4389         else
4390                 len = PAGE_CACHE_SIZE;
4391         /*
4392          * Return if we have all the buffers mapped. This avoids the need to do
4393          * journal_start/journal_stop which can block and take a long time
4394          */
4395         if (page_has_buffers(page)) {
4396                 if (!walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
4397                                         ext4_bh_unmapped)) {
4398                         /* Wait so that we don't change page under IO */
4399                         wait_on_page_writeback(page);
4400                         ret = VM_FAULT_LOCKED;
4401                         goto out;
4402                 }
4403         }
4404         unlock_page(page);
4405         /* OK, we need to fill the hole... */
4406         if (ext4_should_dioread_nolock(inode))
4407                 get_block = ext4_get_block_write;
4408         else
4409                 get_block = ext4_get_block;
4410 retry_alloc:
4411         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
4412         if (IS_ERR(handle)) {
4413                 ret = VM_FAULT_SIGBUS;
4414                 goto out;
4415         }
4416         ret = __block_page_mkwrite(vma, vmf, get_block);
4417         if (!ret && ext4_should_journal_data(inode)) {
4418                 if (walk_page_buffers(handle, page_buffers(page), 0,
4419                           PAGE_CACHE_SIZE, NULL, do_journal_get_write_access)) {
4420                         unlock_page(page);
4421                         ret = VM_FAULT_SIGBUS;
4422                         goto out;
4423                 }
4424                 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
4425         }
4426         ext4_journal_stop(handle);
4427         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
4428                 goto retry_alloc;
4429 out_ret:
4430         ret = block_page_mkwrite_return(ret);
4431 out:
4432         return ret;
4433 }